id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2404.03301-zh-1
|
2404.03301-zh-1.jpg
|
2. Related works
2.1. Scalar Adjective Lexical Semantics
自从 Kamp (1975) 的开创性工作以来，SA 在形式语义和语言哲学方面受到了很多关注。对于 NLP 研究，情况并非如此，SA 直到最近才引起人们的关注。。其中一些研究与 SA 的词汇语义有关， 尤其是关注其规模成员资格和强度 (de Melo and Bansal, 2013; Kim and de Marneffe, 2013; Shivade et al., 2015; Wilkinson and Tim, 2016; Cocos et al., 2018) 。这项研究的一个特点是使用来自LLM 的静态而非上下文化的单词嵌人。这很重要因为 Garí Soler and Apidianaki (2020) 发现基于BERT 的 (Devlin et al., 2019) 上下文化单词嵌人在向量空间中编码了有关标量强度的更丰富信息。尽管如此，Liu et al.(2023b) 报告说，即使在 MNLI (Williams et al.,2018) 上进行了微调之后，许多最先进的模型在文本推理中的形容词程度估计也不令人满意
# 2.2. Scalar Implicature and Scalar Diversity Pragmatics
在对 SIs 的研究中，Schuster et al.(2020) 发现可以训练 LSTM (Hochreiter and Schmidhuber, 1997) 来推断某些含义，而不是全部含义(例如。$\mathfrak{g}$。
些学生作弊”暗示并非所有学生都作弊)。Jeretic et al.(2020)报告说，对 MNLI 进行微调的 BERT 几乎总是预测确定者并不是全部，而是将 SA 视为同义词，无论其强度如何。这意味着该模型可能缺乏关于形容词的相关语用知识。相比之下，最近的两项研究 (Hu et al.,2022,2023) 通过 GPT2 得出的令人惊讶的措施获得了更有利的结果。基于字符串的惊奇，它认为给定强词的惊奇程度在上下文中所有可能的字符串中都是弱词的替代品，它是通过GPT-2 模型 (Radford et al., 2019) 估计强词出现在其位置的可能性来估计的。基于概念的惊奇将替代项视为具有类似概念的字符串集的成员。备选方案的惊讶率被认为是这样一组中所有字符串的平均值。使用这两种方法，他们发现标量多样性与基于字符串和基于概念的惊奇相关。Ruis et al.(2024) 测试了不同的 LLM 对一般会话含义的理解。他们发现 GPT-4 是性能最好的模型，它具有 30 个思路链提示 (Wei et al.,2022) ,实现了88。66% 在广义蕴涵计算中的应用。这是我们形容词触发的 SI 属于的类别
## 2.3. Direct and Indirect Probing
直接和间接探测方法被广泛用于了解 LLM 的知识。直接探测用于分析 LLM (Garí Soler and Apidianaki,2020) 中编码的隐藏表示。这仅在 LLM 是开源的情况下才有可能。间接探测用于分析闭源模型的表示。这涉及测试各种语义任务的性能。典型任务使用文本提示来评估提示答案 (Ettinger,2020) 。Petroni et al. (2019) 认为间接探测只能揭示模型能力的下限。后续工作提供了各种方法来增加绑定的(Zhong et al., 2021; White et al., 2023)
我们建立在这些结果如下。由于直接探测可以很好地估计模型的能力，但闭源 LLM 只允许间接探测，我们对一些开源模型使用这两种方法。然后， 我们将这些模型的结果用于其他模型的间接探测结果。我们做出以下假设：如果两种探测方法在开源模型的模型性能方面显示出相似的趋势，我们认为我们的间接探测方法在比较不同闭源模型的相对能力方面是有效的
## 3. Probing Lexical Semantics
最近对 SA 的词汇语义的研究主要集中在从开源模型 (Garí Soler and Apidianaki, 2020; Liu et al. 2023b) 中获取其强度信息。在本节中，我们将描述如何针对 SA 词汇语义的两个方面探查三个模型族：(i)规模成员资格(e。g。可能和确定在可能性尺度上),以及(li)形容词强度(e。g。在同一规模内，确定比可能更强烈)
我们评估了三个模型族，总共包含八个模型，以了解不同的模型体系结构，训练目标和大小如何影响词汇语义知识。除 GPT-4 之外的所有模型都通过 huggingface 库 (Wolf et al., 2020) 访问 , GPT-4通过 OpenAIAPI 访问。3
编码器型号 BERT-base/large(b/l)(110M/340M)
(Devlin
et
al.,
2019)
,RoBERTa-
base/large(b/l)(123M/354M) (Liu et al., 2019)
解 码 器 型 号 Falcon-7B-instruct(Falcon) (Al-
mazrouei et al., 2023) ,GPT-4 $^{4}$
编码器-解码器型号 Flan-T5-xl/xxl(3。5B/11B)
(Chung et al.,2022)
## 3.1. Datasets for Lexical-Semantid Probing
3.1.1. Scalar Adiective Datasets
在 Garí Soler and Apidianaki (2020) 之后，我们使用三个 SA 数据集：DEMELO (DM) (de Melo and Bansal, 2013) , CROWD ( CD) (Cocos et al. 2018) 和 WILKINSON (WK) (Wilkinson and Tim, 2016),它们总共包含 185 个半标度(表$\mathbb{E}$1)。例如，质量等级的积极方面从好到棒不等。每个包含不同类别的 SA。有些包含描述外表的形容词，而另一些则没有
<table>
<tbody>
<tr>
<th> </th>
<th>DM</th>
<th>CD</th>
<th>$WK$</th>
</tr>
<tr>
<td>Half-scales</td>
<td>$\overline{87}$</td>
<td>$\overline{77}$</td>
<td>$\overline{21}$</td>
</tr>
<tr>
<td>Distinct adjective pairs</td>
<td>548</td>
<td>330</td>
<td>61</td>
</tr>
</tbody>
</table>
Table 1: Overview of half-scale counts and distinct
adjective pairs in SA datasets.
首先从 WordNet3 收集 DM 半标度。0 (Gross, Derek and Miller, Katherine J, 1990) ,然后由两个以英语为母语的人注释强度
See appendix 11.6 for more implementation details We do not use LLaMA (Touvron et al., 2023) due
to our institutional requirement. We use gpt-4-0613 for
$GPT-4.$
|
[
[
0.208251953125,
0.07244873046875,
0.39208984375,
0.07244873046875,
0.39208984375,
0.09027099609375,
0.208251953125,
0.09027099609375
],
[
0.5322265625,
0.07598876953125,
0.8818359375,
0.07598876953125,
0.8818359375,
0.08905029296875,
0.5322265625,
0.08905029296875
],
[
0.5146484375,
0.0885009765625,
0.88330078125,
0.08905029296875,
0.88330078125,
0.10394287109375,
0.5146484375,
0.10272216796875
],
[
0.118408203125,
0.10394287109375,
0.483642578125,
0.10272216796875,
0.483642578125,
0.1175537109375,
0.118408203125,
0.1181640625
],
[
0.5146484375,
0.10394287109375,
0.880859375,
0.10394287109375,
0.880859375,
0.1175537109375,
0.5146484375,
0.1175537109375
],
[
0.5146484375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.5146484375,
0.1318359375
],
[
0.12176513671875,
0.127685546875,
0.489501953125,
0.127685546875,
0.489501953125,
0.141357421875,
0.12176513671875,
0.141357421875
],
[
0.5146484375,
0.133056640625,
0.8818359375,
0.133056640625,
0.8818359375,
0.146728515625,
0.5146484375,
0.146728515625
],
[
0.118408203125,
0.141357421875,
0.48779296875,
0.14013671875,
0.48779296875,
0.15380859375,
0.118408203125,
0.155029296875
],
[
0.51318359375,
0.1455078125,
0.880859375,
0.146728515625,
0.880859375,
0.160888671875,
0.51318359375,
0.1602783203125
],
[
0.12176513671875,
0.157958984375,
0.48779296875,
0.157958984375,
0.48779296875,
0.168701171875,
0.12176513671875,
0.168701171875
],
[
0.515625,
0.160888671875,
0.8818359375,
0.160888671875,
0.8818359375,
0.174560546875,
0.515625,
0.174560546875
],
[
0.118408203125,
0.169189453125,
0.486083984375,
0.169189453125,
0.486083984375,
0.182861328125,
0.118408203125,
0.182861328125
],
[
0.5146484375,
0.17578125,
0.66650390625,
0.17578125,
0.66650390625,
0.189453125,
0.5146484375,
0.189453125
],
[
0.12176513671875,
0.18408203125,
0.48779296875,
0.18408203125,
0.48779296875,
0.19775390625,
0.12176513671875,
0.19775390625
],
[
0.12005615234375,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.12005615234375,
0.2120361328125
],
[
0.5439453125,
0.2078857421875,
0.8447265625,
0.2078857421875,
0.8447265625,
0.221435546875,
0.5439453125,
0.221435546875
],
[
0.12176513671875,
0.21435546875,
0.486083984375,
0.21435546875,
0.486083984375,
0.2244873046875,
0.12176513671875,
0.2244873046875
],
[
0.12005615234375,
0.226806640625,
0.48779296875,
0.226806640625,
0.48779296875,
0.240478515625,
0.12005615234375,
0.240478515625
],
[
0.5146484375,
0.235107421875,
0.880859375,
0.236328125,
0.880859375,
0.2493896484375,
0.5146484375,
0.248779296875
],
[
0.12005615234375,
0.2410888671875,
0.48779296875,
0.2410888671875,
0.48779296875,
0.254638671875,
0.12005615234375,
0.254638671875
],
[
0.12176513671875,
0.254638671875,
0.48779296875,
0.254638671875,
0.48779296875,
0.268310546875,
0.12176513671875,
0.268310546875
],
[
0.5146484375,
0.2493896484375,
0.8818359375,
0.2493896484375,
0.8818359375,
0.26318359375,
0.5146484375,
0.26318359375
],
[
0.118408203125,
0.268310546875,
0.489501953125,
0.26953125,
0.489501953125,
0.283203125,
0.118408203125,
0.281982421875
],
[
0.5146484375,
0.26416015625,
0.88330078125,
0.26416015625,
0.88330078125,
0.27783203125,
0.5146484375,
0.27783203125
],
[
0.51318359375,
0.276611328125,
0.8818359375,
0.27783203125,
0.8818359375,
0.292236328125,
0.51318359375,
0.29150390625
],
[
0.12176513671875,
0.283935546875,
0.489501953125,
0.283935546875,
0.489501953125,
0.297607421875,
0.12176513671875,
0.297607421875
],
[
0.118408203125,
0.297607421875,
0.48779296875,
0.29638671875,
0.48779296875,
0.311279296875,
0.118408203125,
0.312255859375
],
[
0.5146484375,
0.292236328125,
0.88330078125,
0.292236328125,
0.88330078125,
0.305908203125,
0.5146484375,
0.305908203125
],
[
0.118408203125,
0.311279296875,
0.48779296875,
0.3134765625,
0.48779296875,
0.32666015625,
0.118408203125,
0.32470703125
],
[
0.5146484375,
0.30712890625,
0.88330078125,
0.30712890625,
0.88330078125,
0.320556640625,
0.5146484375,
0.320556640625
],
[
0.5146484375,
0.320556640625,
0.7060546875,
0.320556640625,
0.7060546875,
0.334228515625,
0.5146484375,
0.334228515625
],
[
0.12176513671875,
0.32666015625,
0.48779296875,
0.32666015625,
0.48779296875,
0.34033203125,
0.12176513671875,
0.34033203125
],
[
0.53076171875,
0.3349609375,
0.88330078125,
0.3349609375,
0.88330078125,
0.3486328125,
0.53076171875,
0.3486328125
],
[
0.12176513671875,
0.342529296875,
0.187255859375,
0.342529296875,
0.187255859375,
0.35400390625,
0.12176513671875,
0.35400390625
],
[
0.51708984375,
0.349853515625,
0.8818359375,
0.349853515625,
0.8818359375,
0.363525390625,
0.51708984375,
0.363525390625
],
[
0.5146484375,
0.3623046875,
0.880859375,
0.3623046875,
0.880859375,
0.3759765625,
0.5146484375,
0.3759765625
],
[
0.118408203125,
0.37353515625,
0.17041015625,
0.37353515625,
0.17041015625,
0.388427734375,
0.118408203125,
0.388427734375
],
[
0.172119140625,
0.372314453125,
0.432373046875,
0.37353515625,
0.432373046875,
0.388427734375,
0.172119140625,
0.38720703125
],
[
0.5146484375,
0.3759765625,
0.88330078125,
0.3759765625,
0.88330078125,
0.39013671875,
0.5146484375,
0.39013671875
],
[
0.16796875,
0.388427734375,
0.353515625,
0.3896484375,
0.353515625,
0.40380859375,
0.16796875,
0.402587890625
],
[
0.51318359375,
0.392578125,
0.69873046875,
0.39013671875,
0.69873046875,
0.40380859375,
0.51318359375,
0.40625
],
[
0.53076171875,
0.40625,
0.8818359375,
0.40625,
0.8818359375,
0.419189453125,
0.53076171875,
0.419189453125
],
[
0.12176513671875,
0.412109375,
0.484375,
0.412109375,
0.484375,
0.42578125,
0.12176513671875,
0.42578125
],
[
0.515625,
0.421630859375,
0.57861328125,
0.421630859375,
0.57861328125,
0.431640625,
0.515625,
0.431640625
],
[
0.71728515625,
0.421630859375,
0.767578125,
0.421630859375,
0.767578125,
0.432861328125,
0.71728515625,
0.432861328125
],
[
0.8017578125,
0.42041015625,
0.8818359375,
0.421630859375,
0.8818359375,
0.432861328125,
0.8017578125,
0.431640625
],
[
0.6572265625,
0.4228515625,
0.68408203125,
0.4228515625,
0.68408203125,
0.43115234375,
0.6572265625,
0.43115234375
],
[
0.5751953125,
0.42578125,
0.587890625,
0.42578125,
0.587890625,
0.4287109375,
0.5751953125,
0.4287109375
],
[
0.59375,
0.42578125,
0.60107421875,
0.42578125,
0.60107421875,
0.4287109375,
0.59375,
0.4287109375
],
[
0.60205078125,
0.4228515625,
0.62890625,
0.4228515625,
0.62890625,
0.43115234375,
0.60205078125,
0.43115234375
],
[
0.51318359375,
0.43408203125,
0.84619140625,
0.43408203125,
0.84619140625,
0.44775390625,
0.51318359375,
0.44775390625
],
[
0.12005615234375,
0.427490234375,
0.483642578125,
0.427490234375,
0.483642578125,
0.438232421875,
0.12005615234375,
0.438232421875
],
[
0.12005615234375,
0.441162109375,
0.455078125,
0.441162109375,
0.455078125,
0.454833984375,
0.12005615234375,
0.454833984375
],
[
0.52880859375,
0.44775390625,
0.84814453125,
0.448974609375,
0.84814453125,
0.4619140625,
0.52880859375,
0.46142578125
],
[
0.12005615234375,
0.454833984375,
0.48779296875,
0.454833984375,
0.48779296875,
0.468505859375,
0.12005615234375,
0.468505859375
],
[
0.84033203125,
0.449462890625,
0.8818359375,
0.449462890625,
0.8818359375,
0.460205078125,
0.84033203125,
0.460205078125
],
[
0.5146484375,
0.4619140625,
0.748046875,
0.4619140625,
0.748046875,
0.4755859375,
0.5146484375,
0.4755859375
],
[
0.12005615234375,
0.4697265625,
0.48779296875,
0.4697265625,
0.48779296875,
0.482666015625,
0.12005615234375,
0.482666015625
],
[
0.5322265625,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.490478515625,
0.5322265625,
0.490478515625
],
[
0.12176513671875,
0.48388671875,
0.48779296875,
0.48388671875,
0.48779296875,
0.49755859375,
0.12176513671875,
0.49755859375
],
[
0.5146484375,
0.4912109375,
0.66162109375,
0.4912109375,
0.66162109375,
0.5048828125,
0.5146484375,
0.5048828125
],
[
0.12005615234375,
0.49755859375,
0.48779296875,
0.49755859375,
0.48779296875,
0.51123046875,
0.12005615234375,
0.51123046875
],
[
0.12005615234375,
0.51123046875,
0.486083984375,
0.5126953125,
0.486083984375,
0.525390625,
0.12005615234375,
0.52490234375
],
[
0.5146484375,
0.5224609375,
0.5556640625,
0.5224609375,
0.5556640625,
0.53369140625,
0.5146484375,
0.53369140625
],
[
0.568359375,
0.521484375,
0.8310546875,
0.521484375,
0.8310546875,
0.53515625,
0.568359375,
0.53515625
],
[
0.5615234375,
0.53515625,
0.63623046875,
0.53759765625,
0.63623046875,
0.5517578125,
0.5615234375,
0.5498046875
],
[
0.118408203125,
0.525390625,
0.486083984375,
0.525390625,
0.486083984375,
0.5390625,
0.118408203125,
0.5390625
],
[
0.12176513671875,
0.5419921875,
0.486083984375,
0.5419921875,
0.486083984375,
0.552734375,
0.12176513671875,
0.552734375
],
[
0.12005615234375,
0.5546875,
0.486083984375,
0.5546875,
0.486083984375,
0.568359375,
0.12005615234375,
0.568359375
],
[
0.5146484375,
0.56103515625,
0.580078125,
0.56103515625,
0.580078125,
0.572265625,
0.5146484375,
0.572265625
],
[
0.57421875,
0.56005859375,
0.78271484375,
0.56005859375,
0.78271484375,
0.57373046875,
0.57421875,
0.57373046875
],
[
0.12176513671875,
0.57080078125,
0.48779296875,
0.57080078125,
0.48779296875,
0.58056640625,
0.12176513671875,
0.58056640625
],
[
0.12176513671875,
0.5830078125,
0.48779296875,
0.58447265625,
0.48779296875,
0.595703125,
0.12176513671875,
0.59423828125
],
[
0.51318359375,
0.58056640625,
0.8818359375,
0.58203125,
0.8818359375,
0.5966796875,
0.51318359375,
0.595703125
],
[
0.12005615234375,
0.59716796875,
0.489501953125,
0.59716796875,
0.489501953125,
0.61083984375,
0.12005615234375,
0.61083984375
],
[
0.5146484375,
0.5966796875,
0.8818359375,
0.5966796875,
0.8818359375,
0.60986328125,
0.5146484375,
0.60986328125
],
[
0.12005615234375,
0.6123046875,
0.489501953125,
0.6123046875,
0.489501953125,
0.6259765625,
0.12005615234375,
0.6259765625
],
[
0.5146484375,
0.61083984375,
0.87744140625,
0.61083984375,
0.87744140625,
0.62158203125,
0.5146484375,
0.62158203125
],
[
0.12176513671875,
0.62744140625,
0.486083984375,
0.62744140625,
0.486083984375,
0.63818359375,
0.12176513671875,
0.63818359375
],
[
0.515625,
0.626953125,
0.87890625,
0.626953125,
0.87890625,
0.63720703125,
0.515625,
0.63720703125
],
[
0.12176513671875,
0.642578125,
0.486083984375,
0.642578125,
0.486083984375,
0.65283203125,
0.12176513671875,
0.65283203125
],
[
0.51318359375,
0.6396484375,
0.8818359375,
0.6396484375,
0.8818359375,
0.65283203125,
0.51318359375,
0.65283203125
],
[
0.12005615234375,
0.65478515625,
0.489501953125,
0.65478515625,
0.489501953125,
0.66845703125,
0.12005615234375,
0.66845703125
],
[
0.5146484375,
0.65380859375,
0.87890625,
0.65380859375,
0.87890625,
0.66455078125,
0.5146484375,
0.66455078125
],
[
0.12005615234375,
0.66845703125,
0.484375,
0.66845703125,
0.484375,
0.6787109375,
0.12005615234375,
0.6787109375
],
[
0.5146484375,
0.66748046875,
0.88330078125,
0.66748046875,
0.88330078125,
0.68115234375,
0.5146484375,
0.68115234375
],
[
0.1234130859375,
0.68408203125,
0.486083984375,
0.68408203125,
0.486083984375,
0.69482421875,
0.1234130859375,
0.69482421875
],
[
0.515625,
0.68310546875,
0.61376953125,
0.68310546875,
0.61376953125,
0.69482421875,
0.515625,
0.69482421875
],
[
0.12176513671875,
0.69677734375,
0.48193359375,
0.69677734375,
0.48193359375,
0.71044921875,
0.12176513671875,
0.71044921875
],
[
0.72314453125,
0.705078125,
0.81201171875,
0.705078125,
0.81201171875,
0.7197265625,
0.72314453125,
0.7197265625
],
[
0.12176513671875,
0.71142578125,
0.48779296875,
0.71142578125,
0.48779296875,
0.724609375,
0.12176513671875,
0.724609375
],
[
0.81689453125,
0.7060546875,
0.85791015625,
0.7060546875,
0.85791015625,
0.7197265625,
0.81689453125,
0.7197265625
],
[
0.5810546875,
0.71875,
0.6708984375,
0.72021484375,
0.6708984375,
0.7353515625,
0.5810546875,
0.73291015625
],
[
0.7314453125,
0.7197265625,
0.755859375,
0.7197265625,
0.755859375,
0.73388671875,
0.7314453125,
0.73388671875
],
[
0.779296875,
0.7197265625,
0.80419921875,
0.7197265625,
0.80419921875,
0.73388671875,
0.779296875,
0.73388671875
],
[
0.8251953125,
0.71630859375,
0.85546875,
0.7197265625,
0.8505859375,
0.736328125,
0.8212890625,
0.73291015625
],
[
0.12176513671875,
0.72705078125,
0.20654296875,
0.72705078125,
0.20654296875,
0.73828125,
0.12176513671875,
0.73828125
],
[
0.54150390625,
0.7353515625,
0.71728515625,
0.7353515625,
0.71728515625,
0.7490234375,
0.54150390625,
0.7490234375
],
[
0.71728515625,
0.73388671875,
0.82861328125,
0.73388671875,
0.82861328125,
0.7490234375,
0.71728515625,
0.7490234375
],
[
0.7119140625,
0.74072265625,
0.7197265625,
0.74072265625,
0.7197265625,
0.74462890625,
0.7119140625,
0.74462890625
],
[
0.8212890625,
0.7353515625,
0.8505859375,
0.7353515625,
0.8505859375,
0.74755859375,
0.8212890625,
0.74755859375
],
[
0.12005615234375,
0.76025390625,
0.177978515625,
0.76025390625,
0.177978515625,
0.771484375,
0.12005615234375,
0.771484375
],
[
0.1695556640625,
0.75732421875,
0.411376953125,
0.7587890625,
0.411376953125,
0.77392578125,
0.1695556640625,
0.771484375
],
[
0.5146484375,
0.76123046875,
0.88330078125,
0.76123046875,
0.88330078125,
0.77490234375,
0.5146484375,
0.77490234375
],
[
0.12176513671875,
0.78076171875,
0.48779296875,
0.78076171875,
0.48779296875,
0.79443359375,
0.12176513671875,
0.79443359375
],
[
0.51318359375,
0.775390625,
0.73779296875,
0.775390625,
0.73779296875,
0.7890625,
0.51318359375,
0.7890625
],
[
0.12005615234375,
0.79638671875,
0.489501953125,
0.79638671875,
0.489501953125,
0.81005859375,
0.12005615234375,
0.81005859375
],
[
0.53076171875,
0.802734375,
0.880859375,
0.80419921875,
0.880859375,
0.81787109375,
0.53076171875,
0.81640625
],
[
0.12005615234375,
0.81005859375,
0.486083984375,
0.81005859375,
0.486083984375,
0.82373046875,
0.12005615234375,
0.82373046875
],
[
0.51318359375,
0.81640625,
0.880859375,
0.81787109375,
0.880859375,
0.83056640625,
0.51318359375,
0.830078125
],
[
0.12176513671875,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.12176513671875,
0.83837890625
],
[
0.515625,
0.83203125,
0.71728515625,
0.83203125,
0.71728515625,
0.845703125,
0.515625,
0.845703125
],
[
0.118408203125,
0.83837890625,
0.48779296875,
0.83740234375,
0.48779296875,
0.8515625,
0.118408203125,
0.8525390625
],
[
0.12176513671875,
0.8525390625,
0.489501953125,
0.8525390625,
0.489501953125,
0.8662109375,
0.12176513671875,
0.8662109375
],
[
0.5341796875,
0.85693359375,
0.880859375,
0.85791015625,
0.880859375,
0.87255859375,
0.5341796875,
0.87158203125
],
[
0.12005615234375,
0.86767578125,
0.48779296875,
0.86767578125,
0.48779296875,
0.88134765625,
0.12005615234375,
0.88134765625
],
[
0.5341796875,
0.87060546875,
0.8818359375,
0.87158203125,
0.8818359375,
0.8857421875,
0.5341796875,
0.88525390625
],
[
0.1234130859375,
0.8828125,
0.484375,
0.8828125,
0.484375,
0.8935546875,
0.1234130859375,
0.8935546875
],
[
0.51318359375,
0.88525390625,
0.8818359375,
0.88525390625,
0.8818359375,
0.8984375,
0.51318359375,
0.8984375
],
[
0.118408203125,
0.8955078125,
0.40478515625,
0.89453125,
0.40478515625,
0.9091796875,
0.118408203125,
0.91015625
],
[
0.5146484375,
0.8984375,
0.5615234375,
0.8984375,
0.5615234375,
0.91015625,
0.5146484375,
0.91015625
]
] |
[
"2.Related works",
"我们建立在这些结果如下。由于直接探测可以很",
"好地估计模型的能力，但闭源LLM只允许间接探",
"2.1.Scalar AdjectiveLexicalSemantics",
"测，我们对一些开源模型使用这两种方法。然后",
"我们将这些模型的结果用于其他模型的间接探测",
"自从Kamp（1975）的开创性工作以来，SA在形式",
"结果。我们做出以下假设：如果两种探测方法在开",
"语义和语言哲学方面受到了很多关注。对于NLP",
"源模型的模型性能方面显示出相似的趋势，我们认",
"研究，情况并非如此，SA直到最近才引起人们",
"为我们的间接探测方法在比较不同闭源模型的相",
"的关注。。其中一些研究与SA的词汇语义有关",
"对能力方面是有效的",
"尤其是关注其规模成员资格和强度（deMeloand",
"Bansal, 2013; Kim and de Marneffe, 2013; Shiv-",
"3.ProbingLexical Semantics",
"adeetal..2015:WilkinsonandTim,2016:Cocos",
"etal.，2018）。这项研究的一个特点是使用来自",
"最近对SA的词汇语义的研究主要集中在从开源",
"LLM的静态而非上下文化的单词嵌入。这很重要，",
"因为 Gari Soler and Apidianaki（2020）发现基于",
"模型 (Gari Soler and Apidianaki, 2020; Liu et al.,",
"BERT的(Devlinetal.,2019上下文化单词嵌人在",
"2023b）中获取其强度信息。在本节中，我们将描",
"述如何针对SA词汇语义的两个方面探查三个模型",
"向量空间中编码了有关标量强度的更丰富信息。尽",
"管如此，Liuetal.（2023b）报告说，即使在MNLI",
"族：（i）规模成员资格（e。g。可能和确定在可能",
"(Williamsetal.,2018）上进行了微调之后，许多最",
"性尺度上)，以及(li)形容词强度(e。g。在同一规",
"模内，确定比可能更强烈)",
"先进的模型在文本推理中的形容词程度估计也不",
"我们评估了三个模型族，总共包含八个模型，以",
"令人满意",
"了解不同的模型体系结构，训练目标和大小如何影",
"响词汇语义知识。除GPT-4之外的所有模型都通",
"2.2.",
"ScalarImplicature and Scalar",
"过huggingface库（Wolf etal.,2020)访问，GPT-4",
"Diversity Pragmatics",
"通过OpenAIAPI访问。3",
"编码器型号BERT-base/large(b/l)(110M/340M)",
"在对Sls的研究中，Schusteretal.(2020）发现可",
"(Devlin",
"2019)",
",RoBERTa",
"al..",
"",
"",
"et",
"base/large(b/l)(123M/354M)(Liuetal.,2019)",
"以训练LSTM（HochreiterandSchmidhuber，1997",
"来推断某些含义，而不是全部含义（例如。9。",
"解码器型号Falcon-7B-instruct(Falcon)",
"些学生作弊”暗示并非所有学生都作弊）。Jeretic",
")(Al",
"mazrouei et al., 2023),GPT-4 4",
"etal.2020）报告说，对MNLI进行微调的BERT",
"编码器-解码器型号Flan-T5-x/xxl(3。5B/11B)",
"几乎总是预测确定者并不是全部，而是将SA视为",
"(Chung et al., 2022)",
"同义词，无论其强度如何。这意味着该模型可能缺",
"乏关于形容词的相关语用知识。相比之下，最近的",
"3.1.",
"DatasetsforLexical-Semantic",
"Probing",
"两项研究（Huetal.,2022,2023）通过GPT2得出",
"的令人惊的措施获得了更有利的结果。基于字符",
"串的惊奇，它认为给定强词的惊奇程度在上下文中",
"3.1.1.",
"Scalar Adjective Datasets",
"所有可能的字符串中都是弱词的替代品，它是通过",
"GPT-2模型（Radfordetal.2019）估计强词出现在",
"在Gari SolerandApidianaki（2020）之后，我们",
"其位置的可能性来估计的。基于概念的惊奇将替代",
"使用三个SA数据集：DEMELO（DM）（deMelo",
"项视为具有类似概念的字符串集的成员。备选方案",
"andBansal.2o13)CROWD(CD)(Cocosetal.",
"的惊牙率被认为是这样一组中所有字符串的平均",
"2018）和WILKINSON（WK）（WilkinsonandTim",
"值。使用这两种方法，他们发现标量多样性与基于",
"2016），它们总共包含185个半标度（表E1）。例",
"字符串和基于概念的惊奇相关。Ruis etal.(2024)",
"如，质量等级的积极方面从好到棒不等。每个包含",
"测试了不同的LM对一般会话含义的理解。他们",
"不同类别的SA。有些包含描述外表的形容词，而",
"发现GPT-4是性能最好的模型，它具有30个思路",
"另一些则没有",
"链提示（Weietal.,2022），实现了88。66%在广",
"DMCD",
"义蕴涵计算中的应用。这是我们形容词触发的SI",
"WK",
"Half-scales",
"87",
"77",
"21",
"属于的类别",
"Distinct adjective pairs",
"548330",
"",
"61",
"2.3.",
"Directand IndirectProbing",
"Table 1: Overview of half-scale counts and distinct",
"直接和间接探测方法被广泛用于了解LLM的知",
"adjective pairs in SA datasets.",
"识。直接探测用于分析LLM(Gari SolerandApidi-",
"首先从WordNet3收集DM半标度。O（Gross",
"anaki,2020）中编码的隐藏表示。这仅在LLM是开",
"DerekandMiller，KatherineJ，1990），然后由两个",
"源的情况下才有可能。间接探测用于分析闭源模型",
"以英语为母语的人注释强度",
"的表示。这涉及测试各种语义任务的性能。典型任",
"务使用文本提示来评估提示答案（Ettinger，2020）。",
"3See appendix 11.6 for more implementation details.",
"Petronietal.(2019）认为间接探测只能揭示模型能",
"4We do not use LLaMA (Touvron et al., 2023) due",
"力的下限。后续工作提供了各种方法来增加绑定的",
"to our institutional requirement. We use gpt-4-0613 for",
"(Zhong et al., 2021; White et al., 2023)",
"GPT-4."
] |
{
"arxiv_id": "2404.03301",
"language": "zh",
"page_idx": 1
}
|
||
2404.03301-zh-2
|
2404.03301-zh-2.jpg
|
CD 形容词首先从释义数据库 (Ganitkevitch et al., 2013; Pavlick et al., 2015) 中收集，然后由人群工人注释。注释者被要求识别给定的形容词在多轮中是否在相同的尺度上，然后注释形容词的强度
WK 形容词通过众包收集。首先向人群工作者提供属于完整范围的提示词，并要求他们列出其他语义相关的形容词。然后，根据工人的能力自动清理量表，并手动注释形容词强度。本研究在 Cocos et al.(2018) 中使用具有半标度的分区 WK,该 WK 源自释义数据库 (Ganitkevitch et al., 2013; Pavlick et al.,2015)
3.1.2. Context Sentence Datasets
要计算形容词的上下文化单词嵌人，我们使用Garí Soler and Apidianaki (2020) 中的上下文句子数据集 ukWac。ukWac 为 DM, CD 和 WK 中的每个半音阶提供 10 个上下文句子。s 上的所有形容词共享 10 个不同句子的相同上下文集。每个候选句子在 s 上都包含一个形容词，并允许对其比例伴侣进行词汇替换。通过使用 context2vec(Melamud et al., 2016) 计算其可接受性得分来确保替换后的可接受性。下面给出一个例子，其中斜体字是 s上的形容词，出现在原始上下文句子中。它的其他比例替代品在括号内。
They are extremely
catchy and are not only great to listen to, but they are also thrilling (interesting/moving/exciting)
to sing .
# 3.2. Probing Scale Membership
在本小节中，我们评估 LLM 是否编码了规模成员资格的概念。理想情况下，我们希望模型能够在上下文中将形容词分配给它们相应的尺度 (例如。$\mathfrak{g}$。温暖属于温标)
# 3.2.1. Scale Membership Direct Probing Method
我们通过获取 SA 和半标度的上下文化表示并计算它们的相似性来直接探测标度成员资格。我们使用ukWac 检索单词嵌人，方法是在半标度$s$上对每个形容词$a$的不同句子进行采样，以获得其上下文表示$\vec{a}$ 。我们通过在$s$上添加最弱和最强形容词的表示形式，得出每个$s$的比例向量$\vec{s}$。例如，对于适当-精细-拟合-良好的量表，其中适当是最不强烈的，而良好是最强烈的形容词$,\vec{s}$是$adequate+good$ 。然后，计算 SA 数据集$D_a$中的每个形容词$a$与数据集中所有其他$\vec{s}$之间的余弦相似度$cos(\vec{a},\vec{s})$ 。每个形容词的音阶按$cos(\vec{a},\vec{s})$排序$;cos(\vec{a},\vec{s})$最高的音阶被认为是$a$最可能属于的音阶，依此类推。为了进行评估，考虑了$a$ 属于($rank_s$)的规模 $s$ 的排名。我们将平均倒数秩(MRR)作为评估指标； 参见方程(1)
(1)
$$MRR=\frac{1}{|D_a|}\sum\limits_{s\in D_a}\frac{1}{rank_s}$$
从直觉上讲，形容词所属的等级越高，MRR 就越接近 1。例如，如果所有形容词在其相应的尺度上对齐，则 MRR 将为 1。如果模型未能将正确的音阶排序为最接近，则意味着它们不会对单词多义和不同音阶之间的细粒度区别进行编码
# 3.2.2. Scale Membership Direct Probing Experiment and Results
当目标形容词被标记为多个片段时，标记表示被平均为形容词的最终表示。我们使用逐层形容词词嵌人来评估每一层中的信息(例如。g。$\vec{a}$和$\vec{s}$ 使用 12 层模型有 12 种表示形式)。用不同的随机种子对上下文句子重复实验十次。我们使用在600BCommonCrawl 数据上训练的快速文本静态单词嵌人$^{5}$(Mikolov et al.,2018)作为基线。对于基线，我们将 s 上最强和最弱形容词的表示形式添加为$\vec{s}$ ,然后还为每个 a 对$cos(\vec{a},\vec{s})$进行排名以计算 MRR。在这里，我们仅报告表 2 中非基线模型的最佳执行层的结果。6
<table>
<tbody>
<tr>
<th>Models</th>
<th>$DM$</th>
<th>$CD$</th>
<th>$WK$</th>
</tr>
<tr>
<td>fast-text</td>
<td>0.842</td>
<td>0.716</td>
<td>0.983</td>
</tr>
<tr>
<td>BERT-b</td>
<td>0.829 $\pm0.010$</td>
<td>0.797 $\pm0.010$</td>
<td>0.997 $\pm0.00.$</td>
</tr>
<tr>
<td>BERT-I</td>
<td>0.853 $\pm0.007$</td>
<td>0.805 $\pm0.011$</td>
<td>0.997 $\pm0.006$</td>
</tr>
<tr>
<td>$RoBERTa-b$</td>
<td>0.668 $\pm0.014$</td>
<td>0.705 $\pm0.007$</td>
<td>0.906 $\pm0.018$</td>
</tr>
<tr>
<td>RoBERTa-l</td>
<td>0.777 $\pm0.011$</td>
<td>0.757 $\pm0.008$</td>
<td>0.977 $\pm0.010$</td>
</tr>
</tbody>
</table>
Table 2: Direct scale membership probing results, subscripted numbers are standard deviation in ten runs. The best results per dataset are in bold.
考虑到其训练数据大小，快速文本基线相当强。所有 LLM 都编码了丰富的规模信息，因为它们的表现大多优于 (有些表现与基线相当)。模型方面， BERT 模型比 RoBERTa 编码更多信息。在同一架构中，模型越大，它们就越好。在整个数据集中， WK 相对容易，因为它(i)在不同尺度上具有相同形式的多义形容词更少，并且(ii)与其他两个数据集相比，用于分类的尺度更少
# 3.2.3. Scale Membership Indirect Probing Method
受Lorge and Pierrehumbert (2023) 和 de Melo and Bansal (2013)的启发，我们使用四个模板，包括$\mathbf{ADJ}_{weak}$ ,如果不是 ADJ$_strong$ ,则间接评估 LLM 有关规模成员资格的知识。相同比例的形容词在这些结构中更有可能同时出现，而不是(e。g。温暖， 如果不热，应该比温暖，如果不薄)更容易出现。$^{7}$
$^{5}$https://github.com/facebookresearch/
fastText/
We also tried other alternative methods to compute membership, which does not generally work as well as the method we report here. See Appendix 11.1 .
$^{7}$See Appendix 11.3 for all templates.
|
[
[
0.1368408203125,
0.07598876953125,
0.489501953125,
0.07598876953125,
0.489501953125,
0.08905029296875,
0.1368408203125,
0.08905029296875
],
[
0.6826171875,
0.08489990234375,
0.70947265625,
0.0872802734375,
0.7080078125,
0.097412109375,
0.68115234375,
0.0955810546875
],
[
0.74462890625,
0.08612060546875,
0.77490234375,
0.08612060546875,
0.77490234375,
0.097412109375,
0.74462890625,
0.097412109375
],
[
0.7060546875,
0.0914306640625,
0.744140625,
0.0914306640625,
0.744140625,
0.1009521484375,
0.7060546875,
0.1009521484375
],
[
0.118408203125,
0.09027099609375,
0.48779296875,
0.09027099609375,
0.48779296875,
0.10394287109375,
0.118408203125,
0.10394287109375
],
[
0.60546875,
0.09320068359375,
0.66650390625,
0.09320068359375,
0.66650390625,
0.1051025390625,
0.60546875,
0.1051025390625
],
[
0.67236328125,
0.0968017578125,
0.75,
0.1092529296875,
0.74462890625,
0.12469482421875,
0.66748046875,
0.11224365234375
],
[
0.85986328125,
0.0926513671875,
0.884765625,
0.0926513671875,
0.884765625,
0.105712890625,
0.85986328125,
0.105712890625
],
[
0.12176513671875,
0.10394287109375,
0.48779296875,
0.10394287109375,
0.48779296875,
0.1175537109375,
0.12176513671875,
0.1175537109375
],
[
0.7314453125,
0.09857177734375,
0.79248046875,
0.1009521484375,
0.791015625,
0.11517333984375,
0.73046875,
0.1134033203125
],
[
0.12005615234375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.5322265625,
0.127685546875,
0.88330078125,
0.127685546875,
0.88330078125,
0.141357421875,
0.5322265625,
0.141357421875
],
[
0.12176513671875,
0.1341552734375,
0.1368408203125,
0.1341552734375,
0.1368408203125,
0.144287109375,
0.12176513671875,
0.144287109375
],
[
0.133544921875,
0.1455078125,
0.48779296875,
0.146728515625,
0.48779296875,
0.160888671875,
0.133544921875,
0.1602783203125
],
[
0.5146484375,
0.141357421875,
0.8818359375,
0.141357421875,
0.8818359375,
0.155029296875,
0.5146484375,
0.155029296875
],
[
0.515625,
0.155029296875,
0.8818359375,
0.155029296875,
0.8818359375,
0.168701171875,
0.515625,
0.168701171875
],
[
0.12005615234375,
0.162109375,
0.489501953125,
0.162109375,
0.489501953125,
0.17578125,
0.12005615234375,
0.17578125
],
[
0.5146484375,
0.169189453125,
0.8818359375,
0.169189453125,
0.8818359375,
0.182861328125,
0.5146484375,
0.182861328125
],
[
0.12005615234375,
0.177001953125,
0.484375,
0.177001953125,
0.484375,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.118408203125,
0.189453125,
0.48779296875,
0.190673828125,
0.48779296875,
0.2037353515625,
0.118408203125,
0.203125
],
[
0.5146484375,
0.18408203125,
0.8017578125,
0.18408203125,
0.8017578125,
0.19775390625,
0.5146484375,
0.19775390625
],
[
0.12005615234375,
0.204833984375,
0.489501953125,
0.204833984375,
0.489501953125,
0.218505859375,
0.12005615234375,
0.218505859375
],
[
0.5146484375,
0.215576171875,
0.55908203125,
0.215576171875,
0.55908203125,
0.2257080078125,
0.5146484375,
0.2257080078125
],
[
0.57421875,
0.213134765625,
0.84619140625,
0.21435546875,
0.84619140625,
0.2286376953125,
0.57421875,
0.22802734375
],
[
0.12176513671875,
0.2197265625,
0.48779296875,
0.2197265625,
0.48779296875,
0.2298583984375,
0.12176513671875,
0.2298583984375
],
[
0.12176513671875,
0.2327880859375,
0.2091064453125,
0.2327880859375,
0.2091064453125,
0.24462890625,
0.12176513671875,
0.24462890625
],
[
0.57275390625,
0.2286376953125,
0.765625,
0.2286376953125,
0.765625,
0.2423095703125,
0.57275390625,
0.2423095703125
],
[
0.515625,
0.25048828125,
0.88330078125,
0.25048828125,
0.88330078125,
0.26416015625,
0.515625,
0.26416015625
],
[
0.1796875,
0.2666015625,
0.399658203125,
0.2666015625,
0.399658203125,
0.27978515625,
0.1796875,
0.27978515625
],
[
0.5146484375,
0.26318359375,
0.88330078125,
0.26416015625,
0.88330078125,
0.279052734375,
0.5146484375,
0.27783203125
],
[
0.5146484375,
0.27978515625,
0.88671875,
0.27978515625,
0.88671875,
0.29345703125,
0.5146484375,
0.29345703125
],
[
0.12005615234375,
0.267333984375,
0.181396484375,
0.267333984375,
0.181396484375,
0.27783203125,
0.12005615234375,
0.27783203125
],
[
0.12005615234375,
0.289306640625,
0.489501953125,
0.289306640625,
0.489501953125,
0.302734375,
0.12005615234375,
0.302734375
],
[
0.5146484375,
0.29345703125,
0.88330078125,
0.29345703125,
0.88330078125,
0.30712890625,
0.5146484375,
0.30712890625
],
[
0.12005615234375,
0.302734375,
0.489501953125,
0.302734375,
0.489501953125,
0.31640625,
0.12005615234375,
0.31640625
],
[
0.51123046875,
0.30712890625,
0.88330078125,
0.30810546875,
0.88330078125,
0.32177734375,
0.51123046875,
0.320556640625
],
[
0.12176513671875,
0.317626953125,
0.489501953125,
0.317626953125,
0.489501953125,
0.330810546875,
0.12176513671875,
0.330810546875
],
[
0.51123046875,
0.320556640625,
0.88330078125,
0.32177734375,
0.88330078125,
0.336181640625,
0.51123046875,
0.3349609375
],
[
0.118408203125,
0.330810546875,
0.48779296875,
0.33203125,
0.48779296875,
0.345703125,
0.118408203125,
0.344482421875
],
[
0.5146484375,
0.336181640625,
0.8818359375,
0.336181640625,
0.8818359375,
0.349853515625,
0.5146484375,
0.349853515625
],
[
0.118408203125,
0.344482421875,
0.489501953125,
0.3466796875,
0.489501953125,
0.35986328125,
0.118408203125,
0.358154296875
],
[
0.5146484375,
0.350830078125,
0.88330078125,
0.350830078125,
0.88330078125,
0.364013671875,
0.5146484375,
0.364013671875
],
[
0.12176513671875,
0.361083984375,
0.486083984375,
0.361083984375,
0.486083984375,
0.371826171875,
0.12176513671875,
0.371826171875
],
[
0.5146484375,
0.364013671875,
0.8818359375,
0.364013671875,
0.8818359375,
0.377685546875,
0.5146484375,
0.377685546875
],
[
0.12176513671875,
0.3759765625,
0.48779296875,
0.3759765625,
0.48779296875,
0.385986328125,
0.12176513671875,
0.385986328125
],
[
0.5146484375,
0.37890625,
0.8818359375,
0.37890625,
0.8818359375,
0.392578125,
0.5146484375,
0.392578125
],
[
0.12005615234375,
0.388427734375,
0.489501953125,
0.388427734375,
0.489501953125,
0.402099609375,
0.12005615234375,
0.402099609375
],
[
0.51318359375,
0.392578125,
0.70947265625,
0.39013671875,
0.70947265625,
0.405029296875,
0.51318359375,
0.40673828125
],
[
0.12005615234375,
0.402099609375,
0.48779296875,
0.402099609375,
0.48779296875,
0.4150390625,
0.12005615234375,
0.4150390625
],
[
0.5380859375,
0.4150390625,
0.59033203125,
0.41748046875,
0.58837890625,
0.431640625,
0.53662109375,
0.429931640625
],
[
0.73291015625,
0.41748046875,
0.75732421875,
0.41748046875,
0.75732421875,
0.429931640625,
0.73291015625,
0.429931640625
],
[
0.82275390625,
0.41748046875,
0.8505859375,
0.41748046875,
0.8505859375,
0.43115234375,
0.82275390625,
0.43115234375
],
[
0.64208984375,
0.418701171875,
0.66748046875,
0.418701171875,
0.66748046875,
0.43115234375,
0.64208984375,
0.43115234375
],
[
0.12176513671875,
0.418701171875,
0.486083984375,
0.418701171875,
0.486083984375,
0.4287109375,
0.12176513671875,
0.4287109375
],
[
0.5380859375,
0.427490234375,
0.59033203125,
0.429931640625,
0.58837890625,
0.443603515625,
0.53662109375,
0.441162109375
],
[
0.634765625,
0.4287109375,
0.67529296875,
0.43115234375,
0.67333984375,
0.443603515625,
0.6337890625,
0.441162109375
],
[
0.724609375,
0.4287109375,
0.765625,
0.43115234375,
0.76416015625,
0.443603515625,
0.72314453125,
0.441162109375
],
[
0.8154296875,
0.4287109375,
0.8564453125,
0.43115234375,
0.85546875,
0.443603515625,
0.8134765625,
0.441162109375
],
[
0.60791015625,
0.440673828125,
0.69873046875,
0.44482421875,
0.69775390625,
0.457763671875,
0.60693359375,
0.45361328125
],
[
0.118408203125,
0.43115234375,
0.26025390625,
0.43115234375,
0.26025390625,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.7001953125,
0.440673828125,
0.7900390625,
0.443603515625,
0.78857421875,
0.457763671875,
0.69873046875,
0.454833984375
],
[
0.79248046875,
0.440673828125,
0.87890625,
0.44482421875,
0.87744140625,
0.458984375,
0.791015625,
0.454833984375
],
[
0.7900390625,
0.44775390625,
0.796875,
0.44775390625,
0.796875,
0.451904296875,
0.7900390625,
0.451904296875
],
[
0.611328125,
0.453125,
0.69775390625,
0.4560546875,
0.6962890625,
0.47021484375,
0.609375,
0.46728515625
],
[
0.70361328125,
0.453125,
0.78662109375,
0.4560546875,
0.78515625,
0.47021484375,
0.7021484375,
0.46728515625
],
[
0.79443359375,
0.453125,
0.87744140625,
0.4560546875,
0.8759765625,
0.47021484375,
0.79248046875,
0.46728515625
],
[
0.54150390625,
0.4560546875,
0.59033203125,
0.4560546875,
0.59033203125,
0.46728515625,
0.54150390625,
0.46728515625
],
[
0.5380859375,
0.443603515625,
0.591796875,
0.443603515625,
0.591796875,
0.454833984375,
0.5380859375,
0.454833984375
],
[
0.15283203125,
0.457275390625,
0.334228515625,
0.457275390625,
0.334228515625,
0.47021484375,
0.15283203125,
0.47021484375
],
[
0.51904296875,
0.465576171875,
0.69775390625,
0.46728515625,
0.69775390625,
0.482177734375,
0.51904296875,
0.479736328125
],
[
0.6943359375,
0.465576171875,
0.78662109375,
0.468505859375,
0.78662109375,
0.482177734375,
0.6943359375,
0.478515625
],
[
0.7958984375,
0.465576171875,
0.87744140625,
0.468505859375,
0.8759765625,
0.482177734375,
0.79443359375,
0.478515625
],
[
0.5205078125,
0.47802734375,
0.8759765625,
0.479736328125,
0.8759765625,
0.49462890625,
0.5205078125,
0.4921875
],
[
0.15283203125,
0.47021484375,
0.38623046875,
0.47021484375,
0.38623046875,
0.48388671875,
0.15283203125,
0.48388671875
],
[
0.15283203125,
0.485107421875,
0.3837890625,
0.485107421875,
0.3837890625,
0.498779296875,
0.15283203125,
0.498779296875
],
[
0.15283203125,
0.49951171875,
0.38623046875,
0.49951171875,
0.38623046875,
0.51318359375,
0.15283203125,
0.51318359375
],
[
0.5146484375,
0.5029296875,
0.880859375,
0.50439453125,
0.880859375,
0.51708984375,
0.5146484375,
0.5166015625
],
[
0.154541015625,
0.51318359375,
0.441650390625,
0.51318359375,
0.441650390625,
0.52685546875,
0.154541015625,
0.52685546875
],
[
0.515625,
0.51953125,
0.8818359375,
0.51953125,
0.8818359375,
0.52978515625,
0.515625,
0.52978515625
],
[
0.151123046875,
0.52783203125,
0.2333984375,
0.52978515625,
0.2333984375,
0.54150390625,
0.151123046875,
0.5390625
],
[
0.5146484375,
0.5322265625,
0.85205078125,
0.5322265625,
0.85205078125,
0.5458984375,
0.5146484375,
0.5458984375
],
[
0.118408203125,
0.56103515625,
0.411376953125,
0.56103515625,
0.411376953125,
0.57470703125,
0.118408203125,
0.57470703125
],
[
0.5322265625,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.5322265625,
0.57666015625
],
[
0.5146484375,
0.57763671875,
0.8818359375,
0.57763671875,
0.8818359375,
0.59130859375,
0.5146484375,
0.59130859375
],
[
0.12005615234375,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.5146484375,
0.59130859375,
0.8818359375,
0.59130859375,
0.8818359375,
0.60498046875,
0.5146484375,
0.60498046875
],
[
0.12005615234375,
0.5986328125,
0.48779296875,
0.5986328125,
0.48779296875,
0.6123046875,
0.12005615234375,
0.6123046875
],
[
0.5146484375,
0.60498046875,
0.88330078125,
0.60498046875,
0.88330078125,
0.6181640625,
0.5146484375,
0.6181640625
],
[
0.12176513671875,
0.61328125,
0.490234375,
0.61328125,
0.490234375,
0.626953125,
0.12176513671875,
0.626953125
],
[
0.5146484375,
0.619140625,
0.8818359375,
0.619140625,
0.8818359375,
0.6328125,
0.5146484375,
0.6328125
],
[
0.12005615234375,
0.62744140625,
0.225830078125,
0.62744140625,
0.225830078125,
0.6396484375,
0.12005615234375,
0.6396484375
],
[
0.51318359375,
0.6328125,
0.880859375,
0.6328125,
0.880859375,
0.646484375,
0.51318359375,
0.646484375
],
[
0.51708984375,
0.6494140625,
0.8818359375,
0.6494140625,
0.8818359375,
0.66015625,
0.51708984375,
0.66015625
],
[
0.12005615234375,
0.66015625,
0.16796875,
0.66015625,
0.16796875,
0.67138671875,
0.12005615234375,
0.67138671875
],
[
0.16455078125,
0.66455078125,
0.177978515625,
0.66455078125,
0.177978515625,
0.66845703125,
0.16455078125,
0.66845703125
],
[
0.177978515625,
0.6591796875,
0.45166015625,
0.66015625,
0.45166015625,
0.67333984375,
0.177978515625,
0.6728515625
],
[
0.5146484375,
0.66455078125,
0.75,
0.66455078125,
0.75,
0.6748046875,
0.5146484375,
0.6748046875
],
[
0.177978515625,
0.6748046875,
0.2392578125,
0.6748046875,
0.2392578125,
0.68603515625,
0.177978515625,
0.68603515625
],
[
0.51318359375,
0.693359375,
0.55908203125,
0.693359375,
0.55908203125,
0.705078125,
0.51318359375,
0.705078125
],
[
0.57421875,
0.69140625,
0.85986328125,
0.6923828125,
0.85986328125,
0.70703125,
0.57421875,
0.7060546875
],
[
0.5576171875,
0.69775390625,
0.57080078125,
0.69775390625,
0.57080078125,
0.70166015625,
0.5576171875,
0.70166015625
],
[
0.57275390625,
0.7080078125,
0.6337890625,
0.7080078125,
0.6337890625,
0.7197265625,
0.57275390625,
0.7197265625
],
[
0.12005615234375,
0.69677734375,
0.48779296875,
0.69677734375,
0.48779296875,
0.71044921875,
0.12005615234375,
0.71044921875
],
[
0.12176513671875,
0.71142578125,
0.48779296875,
0.71142578125,
0.48779296875,
0.724609375,
0.12176513671875,
0.724609375
],
[
0.118408203125,
0.7255859375,
0.489501953125,
0.7255859375,
0.489501953125,
0.7392578125,
0.118408203125,
0.7392578125
],
[
0.515625,
0.73095703125,
0.8818359375,
0.73095703125,
0.8818359375,
0.7412109375,
0.515625,
0.7412109375
],
[
0.118408203125,
0.73828125,
0.489501953125,
0.7392578125,
0.489501953125,
0.7529296875,
0.118408203125,
0.751953125
],
[
0.5146484375,
0.7421875,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.5146484375,
0.755859375
],
[
0.12176513671875,
0.755859375,
0.48779296875,
0.755859375,
0.48779296875,
0.7666015625,
0.12176513671875,
0.7666015625
],
[
0.509765625,
0.7548828125,
0.884765625,
0.755859375,
0.884765625,
0.77392578125,
0.509765625,
0.7724609375
],
[
0.12005615234375,
0.7685546875,
0.48779296875,
0.7685546875,
0.48779296875,
0.7822265625,
0.12005615234375,
0.7822265625
],
[
0.5146484375,
0.7724609375,
0.8818359375,
0.7724609375,
0.8818359375,
0.7861328125,
0.5146484375,
0.7861328125
],
[
0.12005615234375,
0.783203125,
0.48779296875,
0.783203125,
0.48779296875,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.51123046875,
0.78515625,
0.880859375,
0.7861328125,
0.880859375,
0.80029296875,
0.51123046875,
0.7998046875
],
[
0.118408203125,
0.79638671875,
0.48779296875,
0.79638671875,
0.48779296875,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.5146484375,
0.80029296875,
0.869140625,
0.80029296875,
0.869140625,
0.81396484375,
0.5146484375,
0.81396484375
],
[
0.1351318359375,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.1351318359375,
0.82373046875
],
[
0.118408203125,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.5322265625,
0.82763671875,
0.8671875,
0.830078125,
0.8671875,
0.84423828125,
0.5322265625,
0.8427734375
],
[
0.118408203125,
0.8388671875,
0.48779296875,
0.8388671875,
0.48779296875,
0.8525390625,
0.118408203125,
0.8525390625
],
[
0.515625,
0.84423828125,
0.59619140625,
0.84423828125,
0.59619140625,
0.85498046875,
0.515625,
0.85498046875
],
[
0.1175537109375,
0.8525390625,
0.489501953125,
0.8525390625,
0.489501953125,
0.8662109375,
0.1175537109375,
0.8662109375
],
[
0.53076171875,
0.85498046875,
0.8818359375,
0.85546875,
0.8818359375,
0.87060546875,
0.53076171875,
0.869140625
],
[
0.118408203125,
0.86767578125,
0.48779296875,
0.86767578125,
0.48779296875,
0.88134765625,
0.118408203125,
0.88134765625
],
[
0.51318359375,
0.87060546875,
0.88330078125,
0.87060546875,
0.88330078125,
0.88427734375,
0.51318359375,
0.88427734375
],
[
0.12005615234375,
0.8818359375,
0.491943359375,
0.8818359375,
0.491943359375,
0.8955078125,
0.12005615234375,
0.8955078125
],
[
0.51123046875,
0.8828125,
0.8427734375,
0.8828125,
0.8427734375,
0.896484375,
0.51123046875,
0.896484375
],
[
0.12005615234375,
0.89794921875,
0.2216796875,
0.89794921875,
0.2216796875,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.5380859375,
0.896484375,
0.78515625,
0.896484375,
0.78515625,
0.91015625,
0.5380859375,
0.91015625
]
] |
[
"CD形容词首先从释义数据库（Ganitkevitch",
"1",
"¥1",
"1",
"etal.,2013Pavlicketal.,2015）中收集，然后由",
"MRR-",
"DalseDa",
"(1)",
"人群工人注释。注释者被要求识别给定的形容词在",
"ranks",
"多轮中是否在相同的尺度上，然后注释形容词的强",
"从直觉上讲，形容词所属的等级越高，MRR就",
"度",
"WK形容词通过众包收集。首先向人群工作者提",
"越接近1。例如，如果所有形容词在其相应的尺度",
"上对齐，则MRR将为1。如果模型未能将正确的",
"供属于完整范围的提示词，并要求他们列出其他",
"音阶排序为最接近，则意味着它们不会对单词多义",
"语义相关的形容词。然后：根据工人的能力自动清",
"理量表，并手动注释形容词强度。本研究在Cocos",
"和不同音阶之间的细粒度区别进行编码",
"etal.(2018）中使用具有半标度的分区WK，该WK",
"3.2.2.",
"ScaleMembershipDirectProbing",
"源自释义数据库（Ganitkevitchetal.,2013；Pavlick",
"etal.,2015)",
"ExperimentandResults",
"当目标形容词被标记为多个片段时，标记表示被",
"Context Sentence Datasets",
"平均为形容词的最终表示。我们使用逐层形容词",
"词嵌人来评估每一层中的信息（例如。g。a和",
"3.1.2.",
"要计算形容词的上下文化单词嵌入，我们使用",
"使用12层模型有12种表示形式)。用不同的随",
"Gari Soler and Apidianaki(2020）中的上下文句",
"机种子对上下文句子重复实验十次。我们使用在",
"子数据集ukWac。ukWac为DM，CD和WK中的",
"600BCommonCrawl数据上训练的快速文本静态",
"每个半音阶提供10个上下文句子。s上的所有形容",
"单词嵌入5(Mikolovetal.,2018）作为基线。对于",
"词共享10个不同句子的相同上下文集。每个候选",
"基线，我们将s上最强和最弱形容词的表示形式添",
"句子在s上都包含一个形容词，并允许对其比例伴",
"加为，然后还为每个a对 cos(a,)进行排名以",
"侣进行词汇替换。通过使用context2vec（Melamud",
"计算MRR。在这里，我们仅报告表2中非基线模",
"etal.,2016）计算其可接受性得分来确保替换后的",
"型的最佳执行层的结果。6",
"可接受性。下面给出一个例子，其中斜体字是s上",
"Models",
"CD",
"WK",
"DM",
"的形容词，出现在原始上下文句子中。它的其他比",
"fast-text",
"0.842",
"0.716",
"0.983",
"0.829±0.010",
"例替代品在括号内。",
"0.797±0.010",
"0.997 ±0.004",
"",
"0.853 ±0.007",
"0.805 ±0.011",
"0.997 ±0.006",
"BERT-I",
"BERT-b",
"They are extremely",
"RoBERTa-b0.668±0.014",
"0.705±0.007",
"0.906±0.018",
"RoBERTa-0.777 ±0.0110.757 ±0.0080.977±0.010",
"catchy and are not only",
"great to listen to, but",
"they are also thrilling",
"Table2:Direct scalemembershipprobing results,",
"(interesting/moving/exciting)",
"subscriptednumbersarestandarddeviationinten",
"to sing.",
"runs. The best results per dataset are in bold.",
"3.2.ProbingScaleMembership",
"考虑到其训练数据大小，快速文本基线相当强。",
"所有LLM都编码了丰富的规模信息，因为它们的",
"在本小节中，我们评估LLM是否编码了规模成员",
"表现大多优于(有些表现与基线相当)。模型方面：",
"资格的概念。理想情况下，我们希望模型能够在上",
"BERT模型比RoBERTa编码更多信息。在同一架",
"下文中将形容词分配给它们相应的尺度（例如。9。",
"构中，模型越大，它们就越好。在整个数据集中，",
"温暖属于温标）",
"WK相对容易，因为它（i）在不同尺度上具有相同",
"形式的多义形容词更少，并且（ii）与其他两个数",
"3.2.1.",
"",
"ScaleMembershipDirectProbing",
"据集相比，用于分类的尺度更少",
"Method",
"3.2.3.",
"Scale Membership Indirect Probing",
"",
"Method",
"我们通过获取SA和半标度的上下文化表示并计算",
"它们的相似性来直接探测标度成员资格。我们使用",
"ukWac检索单词嵌入，方法是在半标度s上对每",
"受LorgeandPierrehumbert(2023)和deMeloand",
"个形容词α的不同句子进行采样，以获得其上下文",
"Bansal（2013）的启发，我们使用四个模板，包括",
"表示a。我们通过在s上添加最弱和最强形容词的",
"ADJweak，如果不是ADJ strong，则间接评估LLM",
"表示形式，得出每个s的比例向量。例如，对于",
"有关规模成员资格的知识。相同比例的形容词在这",
"适当-精细-拟合-良好的量表，其中适当是最不强烈",
"些结构中更有可能同时出现，而不是(e。9。温暖",
"的，而良好是最强烈的形容词，s是adequate+good",
"如果不热，应该比温暖，如果不薄）更容易出现。",
"然后，计算 SA数据集D。中的每个形容词α与",
"数据集中所有其他3之间的余弦相似度cos(a,s)。",
"5https://github.com/facebookresearch/",
"每个形容词的音阶按cos(a,s)排序；cos(a,s)最高",
"fastText/",
"的音阶被认为是α最可能属于的音阶，依此类推。",
"6We also tried other alternative methods to compute",
"为了进行评估，考虑了α属于（ranks）的规模s",
"membership, which does not generally work as well as",
"的排名。我们将平均倒数秩（MRR）作为评估指标；",
"the method we report here. See Appendix 11.1 .",
"参见方程（1）",
"SeeAppendix 11.3for all templates."
] |
{
"arxiv_id": "2404.03301",
"language": "zh",
"page_idx": 2
}
|
||
2404.03301-zh-3
|
2404.03301-zh-3.jpg
|
对于不是最强项目的半标度上的每个形容词， 我们将其用作模板中的 ADJ $_{weak}$,并获得 5 个最可能的单词作为 ADJ $_strong$ 。例如，对于比例适当-精细-拟合良好，我们提示模型回答适当/精细/拟合中可能是 ADJ $_strong$的内容，如果不是 $ADJ_{strong}$。如果 ADJ$_strong$中的任何前5 个单词与 ADJ $_{weak}$的比例相同，并且重叠不是微不足道的，我们认为这种情况是正确的(例如。g。adequateorevenadequateisatrimalcompletion,goodorevenadequateisconsideredtobecorrectdesp蘇efhéwå暖gintensityranking)
# 3.2.4. Scale Membership Indirect Probing Experiment and Results
对于编码器型号，我们将[MASK]和一个逗号放在$\mathbf{ADJ}_{strong}$的位置(e。g。ADJ$_weak$,如果不是[MASK],)作为初步实验结果表明，BERT 主要将标点符号预测为[MASK]的前 5 个预测。之所以使用逗号，是因为它不限制其前面的形容词是否是定语 (i。e。可以紧跟名词)或谓语(i。e。不能紧跟名词)。对于除 GPT-4 之外的非基线模型，我们在给出上述提示的情况下直接生成 ADJ $_{strong}$。我们为 GPT-4 使用类似的生成目标，并在其他模型中使用最佳模板(请参阅附录 11.7 )。我们使用2019 (Lin et al., 2012) 的 GoogleNgram 语料库切片作为具有类似设置的间接探测基线，以获得每个构造中 ADJ $_{strong}$的 5 个最有可能的完成。在表 3中，我们报告了评估数据集中所有模板中的最佳结果 (我们根据在表$\overline{\mathbb{E}}$13 中选择具有保留数据集的最佳模板来报告结果)。
<table>
<tbody>
<tr>
<th>Models</th>
<th>$DM$</th>
<th>$CD$</th>
<th>$WK$</th>
</tr>
<tr>
<td>Gooqle Ngram 11</td>
<td>0.287 1</td>
<td>0.075 1</td>
<td>0.368</td>
</tr>
<tr>
<td>BERT-b</td>
<td>0.425 1</td>
<td>0.066 1</td>
<td>0.167 1</td>
</tr>
<tr>
<td>BERT-1</td>
<td>0.504 1</td>
<td>0.109 3</td>
<td>0.250</td>
</tr>
<tr>
<td>RoBERTa- b</td>
<td>0.271 4</td>
<td>0.064 3</td>
<td>0.114 一</td>
</tr>
<tr>
<td>RoBERTa-1</td>
<td>0.463</td>
<td>0.100 1</td>
<td>0.286</td>
</tr>
<tr>
<td>Falcon</td>
<td>0.265 1</td>
<td>0.045 1</td>
<td>0.167 1</td>
</tr>
<tr>
<td>GPT-4</td>
<td>0.540</td>
<td>0.273</td>
<td>0.500</td>
</tr>
<tr>
<td>Flan-T5-x1</td>
<td>0.544 2</td>
<td>0.177 3</td>
<td>0.500</td>
</tr>
<tr>
<td>$Flan-T5-xx$</td>
<td>0.515 3</td>
<td>0.156 3</td>
<td>0.364</td>
</tr>
</tbody>
</table>
Table 3: Indirect probing results for scale membership. The best results per dataset are in bold. Subscripts refer to the indices of the best templates.
间接探测始终不如所有模型的直接探测。此结果与 Petroni et al.(2019) 一致。不同车型的相对排名基本一致。给定相同的大小等级，BERT 优于RoBERTa,并且较大的编码器型号优于具有相同架构的较小的编码器型号。这些观察结果支持了我们方法的有效性
我们发现，除 Falcon 之外的所有模型都对有关规模成员资格的丰富信息进行了编码，因为它们的行为与基线相当或优于基线。GPT-4 和 Flan-T5-xl 在两个基准测试中均取得了最佳结果。虽然 GPT- 4在 DM 和 WK 中具有与 Flan-T5-xl 竞争的结果， 但对于 CD 来说要好得多。因此，我们得出结论， GPT-4 是这项任务中表现最好的模型。在同一模型
族中，比例定律倾向于成立(i。e。,越大越好)。然而 ,缩放并不能完全解释结果的模式。例如，Falcon 的性能远远低于较小的编码器模型，甚至低于基线，而 Flan-T5-xxl 的性能低于 Flan-T5-xl
# 3.3. Probing Scalar Intensity
在本节中，我们将描述如何探究 LLM 是否了解相同规模的 SA 具有不同的强度 (例如。g。热表示温
### 3.3.1. Scalar Intensity Direct Probing Method
与以前的方法不同，这些方法将 SA 置于相同的上下文中并在单独的运行中进行计算(例如。$\mathfrak{g}$。这是一部很好/很棒/很棒/很棒的电影)(Garí Soler and Apidianaki, 2020) (G&A), 我们使用了一种新颖的方法，该方法涉及将所有 SA 绑定在与单个输入相同的比例上，以获取其上下文化表示(e。g。goodgreatwonderfulawesome),asillustratedinFigure 2 。我们的假设是，当标量词在同一上下文中突出时，人们更有可能注意到标量词具有不同的强度(Ronai and Xiang, 2023a)
对于每个半标度$s$,我们在$s$上对所有 SA $a$进行十次混洗和绑定，以获得 10 个不同的输人 (e。$\mathbf{g}$。好极了，好极了，好极了，好极了，好极了，等等。)。对于每个$a$ ,我们将其在 10 个输入中的表示平均， 以获得其最终表示。这是为了避免词序和标量强度之间的潜在启发式关联
在获得所有$a$的表示形式之后，我们利用Garí Soler and Apidianaki (2020) (G&A) 中的最佳方法(DiffVec)来获取它们的强度排名。具体来说，数据集$D_a$中所有形容词$a$的强度是通过它们与从另一个 SA 数据集$D_{vec}$ $( D_{vec}\neq D_a$ )得出的外部全局强度向量$\vec d_{Vec}$的余弦相似性来衡量的。8
对于$D_{vec}$中的每个半标度$s,s$上最温和和极端的形容词表示为$a_ms$和$a_es$ 。我们用方程(2) 计算$d_Vec$
(2)
$$d\vec{Vec}=\frac{1}{|D_{vec}\setminus D_a|}\sum_{s\in D_{vec}\setminus D_a}\vec{a_{es}}-\vec{a_{ms}}$$
对于所有$s\in D_a$ , $s$上的形容词均按其与$\vec d_{Vec}$ 的余弦相似性进行排名：较高的相似性意味着较高的强度。如果两个形容词具有相同的相似性，则将它们视为同等强度
# 3.3.2. Scalar Intensity Direct Probing Experiment and Results
我们在BERT-b/l,RoBERTa-b//中探测逐层词嵌人并报告在不同数据集中使用$\vec{d_{Vec}}$跨层的成对准确性的最佳结果。我们再次在这里提供快速文本嵌入作为基线。对于基线，我们对所有$s\in D_{vec}\setminus D_a$
We don't use the global intensity vector in $D_a$ due to
the concern of generalisability.
|
[
[
0.1368408203125,
0.07598876953125,
0.495361328125,
0.07598876953125,
0.495361328125,
0.08905029296875,
0.1368408203125,
0.08905029296875
],
[
0.51123046875,
0.0748291015625,
0.8818359375,
0.07598876953125,
0.8818359375,
0.09027099609375,
0.51123046875,
0.08905029296875
],
[
0.12005615234375,
0.08905029296875,
0.489501953125,
0.08905029296875,
0.489501953125,
0.10394287109375,
0.12005615234375,
0.10394287109375
],
[
0.515625,
0.0914306640625,
0.880859375,
0.0914306640625,
0.880859375,
0.1015625,
0.515625,
0.1015625
],
[
0.12005615234375,
0.10394287109375,
0.489501953125,
0.10394287109375,
0.489501953125,
0.1181640625,
0.12005615234375,
0.1181640625
],
[
0.51318359375,
0.1051025390625,
0.88330078125,
0.1051025390625,
0.88330078125,
0.1181640625,
0.51318359375,
0.1181640625
],
[
0.12005615234375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.51318359375,
0.1181640625,
0.8212890625,
0.1181640625,
0.8212890625,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.118408203125,
0.12890625,
0.489501953125,
0.1312255859375,
0.489501953125,
0.1484375,
0.118408203125,
0.146728515625
],
[
0.118408203125,
0.1455078125,
0.489501953125,
0.1455078125,
0.489501953125,
0.163330078125,
0.118408203125,
0.163330078125
],
[
0.51318359375,
0.149658203125,
0.78076171875,
0.149658203125,
0.78076171875,
0.163330078125,
0.51318359375,
0.163330078125
],
[
0.12005615234375,
0.160888671875,
0.48779296875,
0.160888671875,
0.48779296875,
0.174560546875,
0.12005615234375,
0.174560546875
],
[
0.12005615234375,
0.17578125,
0.489501953125,
0.17578125,
0.489501953125,
0.189453125,
0.12005615234375,
0.189453125
],
[
0.51318359375,
0.171630859375,
0.8818359375,
0.171630859375,
0.8818359375,
0.185302734375,
0.51318359375,
0.185302734375
],
[
0.118408203125,
0.189453125,
0.490234375,
0.188232421875,
0.490234375,
0.203125,
0.118408203125,
0.2037353515625
],
[
0.5146484375,
0.185302734375,
0.8818359375,
0.185302734375,
0.8818359375,
0.1995849609375,
0.5146484375,
0.1995849609375
],
[
0.1175537109375,
0.20068359375,
0.7314453125,
0.20068359375,
0.7314453125,
0.218505859375,
0.1175537109375,
0.218505859375
],
[
0.5146484375,
0.23095703125,
0.580078125,
0.23095703125,
0.580078125,
0.2423095703125,
0.5146484375,
0.2423095703125
],
[
0.57080078125,
0.2286376953125,
0.82275390625,
0.2298583984375,
0.82275390625,
0.24462890625,
0.57080078125,
0.243408203125
],
[
0.12005615234375,
0.235107421875,
0.181396484375,
0.235107421875,
0.181396484375,
0.2452392578125,
0.12005615234375,
0.2452392578125
],
[
0.177978515625,
0.2327880859375,
0.465087890625,
0.2340087890625,
0.465087890625,
0.2476806640625,
0.177978515625,
0.2464599609375
],
[
0.177978515625,
0.248779296875,
0.3720703125,
0.248779296875,
0.3720703125,
0.262451171875,
0.177978515625,
0.262451171875
],
[
0.57275390625,
0.2452392578125,
0.634765625,
0.2452392578125,
0.634765625,
0.257080078125,
0.57275390625,
0.257080078125
],
[
0.12005615234375,
0.270751953125,
0.489501953125,
0.270751953125,
0.489501953125,
0.283935546875,
0.12005615234375,
0.283935546875
],
[
0.515625,
0.267333984375,
0.87890625,
0.267333984375,
0.87890625,
0.27783203125,
0.515625,
0.27783203125
],
[
0.118408203125,
0.283203125,
0.489501953125,
0.283203125,
0.489501953125,
0.300537109375,
0.118408203125,
0.300537109375
],
[
0.515625,
0.28076171875,
0.8818359375,
0.28076171875,
0.8818359375,
0.29443359375,
0.515625,
0.29443359375
],
[
0.118408203125,
0.298583984375,
0.48779296875,
0.298583984375,
0.48779296875,
0.312255859375,
0.118408203125,
0.312255859375
],
[
0.5224609375,
0.298583984375,
0.52880859375,
0.298583984375,
0.52880859375,
0.302734375,
0.5224609375,
0.302734375
],
[
0.5283203125,
0.29443359375,
0.88330078125,
0.29443359375,
0.88330078125,
0.30810546875,
0.5283203125,
0.30810546875
],
[
0.12005615234375,
0.3134765625,
0.489501953125,
0.3134765625,
0.489501953125,
0.32666015625,
0.12005615234375,
0.32666015625
],
[
0.5146484375,
0.31005859375,
0.880859375,
0.31005859375,
0.880859375,
0.320556640625,
0.5146484375,
0.320556640625
],
[
0.12005615234375,
0.32666015625,
0.48779296875,
0.32666015625,
0.48779296875,
0.34033203125,
0.12005615234375,
0.34033203125
],
[
0.5146484375,
0.32373046875,
0.88330078125,
0.32373046875,
0.88330078125,
0.33740234375,
0.5146484375,
0.33740234375
],
[
0.12005615234375,
0.341552734375,
0.489501953125,
0.341552734375,
0.489501953125,
0.355224609375,
0.12005615234375,
0.355224609375
],
[
0.5146484375,
0.33837890625,
0.88330078125,
0.33837890625,
0.88330078125,
0.3515625,
0.5146484375,
0.3515625
],
[
0.118408203125,
0.35400390625,
0.48779296875,
0.35400390625,
0.48779296875,
0.36767578125,
0.118408203125,
0.36767578125
],
[
0.51318359375,
0.3515625,
0.87744140625,
0.3515625,
0.87744140625,
0.365234375,
0.51318359375,
0.365234375
],
[
0.118408203125,
0.3681640625,
0.489501953125,
0.369384765625,
0.489501953125,
0.38427734375,
0.118408203125,
0.383056640625
],
[
0.5146484375,
0.36767578125,
0.880859375,
0.36767578125,
0.880859375,
0.377685546875,
0.5146484375,
0.377685546875
],
[
0.12176513671875,
0.385498046875,
0.48779296875,
0.385498046875,
0.48779296875,
0.3955078125,
0.12176513671875,
0.3955078125
],
[
0.51318359375,
0.38134765625,
0.884765625,
0.38134765625,
0.884765625,
0.394287109375,
0.51318359375,
0.394287109375
],
[
0.12005615234375,
0.39794921875,
0.489501953125,
0.39794921875,
0.489501953125,
0.410888671875,
0.12005615234375,
0.410888671875
],
[
0.51318359375,
0.394287109375,
0.705078125,
0.394287109375,
0.705078125,
0.407958984375,
0.51318359375,
0.407958984375
],
[
0.12005615234375,
0.412109375,
0.489501953125,
0.412109375,
0.489501953125,
0.42578125,
0.12005615234375,
0.42578125
],
[
0.52880859375,
0.4091796875,
0.88330078125,
0.4091796875,
0.88330078125,
0.4228515625,
0.52880859375,
0.4228515625
],
[
0.1234130859375,
0.427490234375,
0.48779296875,
0.427490234375,
0.48779296875,
0.438232421875,
0.1234130859375,
0.438232421875
],
[
0.51123046875,
0.4228515625,
0.8818359375,
0.4228515625,
0.8818359375,
0.435791015625,
0.51123046875,
0.435791015625
],
[
0.118408203125,
0.439453125,
0.489501953125,
0.440673828125,
0.489501953125,
0.454833984375,
0.118408203125,
0.45361328125
],
[
0.51318359375,
0.43701171875,
0.8818359375,
0.43701171875,
0.8818359375,
0.45068359375,
0.51318359375,
0.45068359375
],
[
0.12005615234375,
0.454833984375,
0.489501953125,
0.454833984375,
0.489501953125,
0.468505859375,
0.12005615234375,
0.468505859375
],
[
0.5146484375,
0.45068359375,
0.8818359375,
0.45068359375,
0.8818359375,
0.46435546875,
0.5146484375,
0.46435546875
],
[
0.12005615234375,
0.4697265625,
0.48779296875,
0.4697265625,
0.48779296875,
0.482666015625,
0.12005615234375,
0.482666015625
],
[
0.51318359375,
0.465576171875,
0.8818359375,
0.465576171875,
0.8818359375,
0.478515625,
0.51318359375,
0.478515625
],
[
0.12005615234375,
0.482666015625,
0.285400390625,
0.482666015625,
0.285400390625,
0.496337890625,
0.12005615234375,
0.496337890625
],
[
0.515625,
0.479736328125,
0.68408203125,
0.479736328125,
0.68408203125,
0.493408203125,
0.515625,
0.493408203125
],
[
0.53076171875,
0.493408203125,
0.8818359375,
0.493408203125,
0.8818359375,
0.50732421875,
0.53076171875,
0.50732421875
],
[
0.162109375,
0.50830078125,
0.2216796875,
0.50830078125,
0.2216796875,
0.5224609375,
0.162109375,
0.5224609375
],
[
0.355224609375,
0.50830078125,
0.382080078125,
0.50830078125,
0.382080078125,
0.521484375,
0.355224609375,
0.521484375
],
[
0.51318359375,
0.50732421875,
0.88330078125,
0.50830078125,
0.88330078125,
0.521484375,
0.51318359375,
0.52099609375
],
[
0.426513671875,
0.5087890625,
0.459228515625,
0.5087890625,
0.459228515625,
0.521484375,
0.426513671875,
0.521484375
],
[
0.27783203125,
0.5087890625,
0.30908203125,
0.5087890625,
0.30908203125,
0.521484375,
0.27783203125,
0.521484375
],
[
0.411376953125,
0.521484375,
0.476806640625,
0.52392578125,
0.47607421875,
0.5380859375,
0.41064453125,
0.5361328125
],
[
0.5146484375,
0.521484375,
0.88330078125,
0.521484375,
0.88330078125,
0.53515625,
0.5146484375,
0.53515625
],
[
0.26025390625,
0.5224609375,
0.32421875,
0.5224609375,
0.32421875,
0.53759765625,
0.26025390625,
0.53759765625
],
[
0.1318359375,
0.5224609375,
0.2509765625,
0.52392578125,
0.2509765625,
0.5380859375,
0.1318359375,
0.53759765625
],
[
0.312255859375,
0.52685546875,
0.318115234375,
0.52978515625,
0.312255859375,
0.5361328125,
0.306396484375,
0.533203125
],
[
0.334228515625,
0.5224609375,
0.39794921875,
0.5224609375,
0.39794921875,
0.53759765625,
0.334228515625,
0.53759765625
],
[
0.160400390625,
0.53515625,
0.22412109375,
0.53759765625,
0.2225341796875,
0.5517578125,
0.15869140625,
0.5498046875
],
[
0.411376953125,
0.5361328125,
0.47607421875,
0.5380859375,
0.474365234375,
0.552734375,
0.41064453125,
0.55029296875
],
[
0.515625,
0.5361328125,
0.8818359375,
0.5361328125,
0.8818359375,
0.5498046875,
0.515625,
0.5498046875
],
[
0.3359375,
0.5390625,
0.4013671875,
0.5390625,
0.4013671875,
0.55029296875,
0.3359375,
0.55029296875
],
[
0.26123046875,
0.5390625,
0.32568359375,
0.5390625,
0.32568359375,
0.55029296875,
0.26123046875,
0.55029296875
],
[
0.3359375,
0.55029296875,
0.399658203125,
0.552734375,
0.39794921875,
0.56787109375,
0.334228515625,
0.5654296875
],
[
0.411376953125,
0.55029296875,
0.474365234375,
0.552734375,
0.47265625,
0.56787109375,
0.41064453125,
0.5654296875
],
[
0.515625,
0.55029296875,
0.8818359375,
0.55029296875,
0.8818359375,
0.56396484375,
0.515625,
0.56396484375
],
[
0.2626953125,
0.5517578125,
0.322509765625,
0.5517578125,
0.322509765625,
0.56640625,
0.2626953125,
0.56640625
],
[
0.162841796875,
0.5517578125,
0.2216796875,
0.5517578125,
0.2216796875,
0.56640625,
0.162841796875,
0.56640625
],
[
0.258544921875,
0.5654296875,
0.328369140625,
0.56787109375,
0.328369140625,
0.58203125,
0.258544921875,
0.580078125
],
[
0.330078125,
0.5654296875,
0.4013671875,
0.56787109375,
0.4013671875,
0.58203125,
0.330078125,
0.580078125
],
[
0.41064453125,
0.56787109375,
0.474365234375,
0.56787109375,
0.474365234375,
0.58203125,
0.41064453125,
0.58203125
],
[
0.5146484375,
0.5634765625,
0.884765625,
0.5654296875,
0.884765625,
0.5791015625,
0.5146484375,
0.57666015625
],
[
0.14697265625,
0.580078125,
0.235107421875,
0.58203125,
0.235107421875,
0.5966796875,
0.14697265625,
0.59423828125
],
[
0.145263671875,
0.568359375,
0.2359619140625,
0.568359375,
0.2359619140625,
0.580078125,
0.145263671875,
0.580078125
],
[
0.26025390625,
0.58203125,
0.32568359375,
0.58203125,
0.32568359375,
0.5966796875,
0.26025390625,
0.5966796875
],
[
0.334228515625,
0.58203125,
0.399658203125,
0.58203125,
0.399658203125,
0.5966796875,
0.334228515625,
0.5966796875
],
[
0.411376953125,
0.58203125,
0.474365234375,
0.58203125,
0.474365234375,
0.5966796875,
0.411376953125,
0.5966796875
],
[
0.411376953125,
0.59423828125,
0.474365234375,
0.5966796875,
0.47265625,
0.61083984375,
0.41064453125,
0.609375
],
[
0.53076171875,
0.59326171875,
0.8818359375,
0.59326171875,
0.8818359375,
0.60693359375,
0.53076171875,
0.60693359375
],
[
0.26025390625,
0.595703125,
0.33154296875,
0.595703125,
0.33154296875,
0.60986328125,
0.26025390625,
0.60986328125
],
[
0.166259765625,
0.59716796875,
0.21826171875,
0.59716796875,
0.21826171875,
0.609375,
0.166259765625,
0.609375
],
[
0.333251953125,
0.59716796875,
0.4013671875,
0.59716796875,
0.4013671875,
0.609375,
0.333251953125,
0.609375
],
[
0.166259765625,
0.61083984375,
0.21826171875,
0.61083984375,
0.21826171875,
0.6259765625,
0.166259765625,
0.6259765625
],
[
0.268798828125,
0.61083984375,
0.316650390625,
0.61083984375,
0.316650390625,
0.6259765625,
0.268798828125,
0.6259765625
],
[
0.343505859375,
0.61083984375,
0.39208984375,
0.61083984375,
0.39208984375,
0.6259765625,
0.343505859375,
0.6259765625
],
[
0.4189453125,
0.61083984375,
0.468505859375,
0.61083984375,
0.468505859375,
0.6259765625,
0.4189453125,
0.6259765625
],
[
0.5146484375,
0.60546875,
0.884765625,
0.60693359375,
0.884765625,
0.62158203125,
0.5146484375,
0.62060546875
],
[
0.411376953125,
0.62353515625,
0.47607421875,
0.6259765625,
0.474365234375,
0.64111328125,
0.41064453125,
0.6396484375
],
[
0.5146484375,
0.619140625,
0.587890625,
0.62255859375,
0.5859375,
0.63720703125,
0.51318359375,
0.63427734375
],
[
0.330078125,
0.62451171875,
0.4013671875,
0.626953125,
0.4013671875,
0.64111328125,
0.330078125,
0.6396484375
],
[
0.151123046875,
0.6259765625,
0.2301025390625,
0.6259765625,
0.2301025390625,
0.64013671875,
0.151123046875,
0.64013671875
],
[
0.26123046875,
0.62451171875,
0.337646484375,
0.626953125,
0.337646484375,
0.64111328125,
0.26123046875,
0.6396484375
],
[
0.26025390625,
0.64013671875,
0.418212890625,
0.64013671875,
0.418212890625,
0.65478515625,
0.26025390625,
0.65478515625
],
[
0.405517578125,
0.63818359375,
0.474365234375,
0.64013671875,
0.474365234375,
0.65478515625,
0.405517578125,
0.65283203125
],
[
0.6396484375,
0.65380859375,
0.6533203125,
0.65380859375,
0.6533203125,
0.66650390625,
0.6396484375,
0.66650390625
],
[
0.1502685546875,
0.64111328125,
0.2333984375,
0.64111328125,
0.2333984375,
0.65283203125,
0.1502685546875,
0.65283203125
],
[
0.54248046875,
0.6591796875,
0.60107421875,
0.662109375,
0.599609375,
0.67578125,
0.54150390625,
0.6728515625
],
[
0.7060546875,
0.65478515625,
0.86572265625,
0.66015625,
0.86376953125,
0.68505859375,
0.705078125,
0.6796875
],
[
0.60205078125,
0.662109375,
0.765625,
0.6748046875,
0.76318359375,
0.6953125,
0.599609375,
0.68310546875
],
[
0.85986328125,
0.662109375,
0.8818359375,
0.662109375,
0.8818359375,
0.6748046875,
0.85986328125,
0.6748046875
],
[
0.118408203125,
0.6796875,
0.486083984375,
0.6796875,
0.486083984375,
0.693359375,
0.118408203125,
0.693359375
],
[
0.118408203125,
0.66748046875,
0.489501953125,
0.66748046875,
0.489501953125,
0.68115234375,
0.118408203125,
0.68115234375
],
[
0.1175537109375,
0.69482421875,
0.47607421875,
0.69677734375,
0.47607421875,
0.71044921875,
0.1175537109375,
0.7080078125
],
[
0.52880859375,
0.70068359375,
0.8818359375,
0.703125,
0.8818359375,
0.71728515625,
0.52880859375,
0.71533203125
],
[
0.5146484375,
0.71533203125,
0.880859375,
0.71533203125,
0.880859375,
0.728515625,
0.5146484375,
0.728515625
],
[
0.1351318359375,
0.7255859375,
0.484375,
0.7255859375,
0.484375,
0.736328125,
0.1351318359375,
0.736328125
],
[
0.5146484375,
0.73095703125,
0.8818359375,
0.73095703125,
0.8818359375,
0.74462890625,
0.5146484375,
0.74462890625
],
[
0.118408203125,
0.73828125,
0.486083984375,
0.73828125,
0.486083984375,
0.751953125,
0.118408203125,
0.751953125
],
[
0.5146484375,
0.74658203125,
0.6484375,
0.74658203125,
0.6484375,
0.75732421875,
0.5146484375,
0.75732421875
],
[
0.12005615234375,
0.75439453125,
0.486083984375,
0.75439453125,
0.486083984375,
0.76708984375,
0.12005615234375,
0.76708984375
],
[
0.118408203125,
0.7666015625,
0.48779296875,
0.76708984375,
0.48779296875,
0.78076171875,
0.118408203125,
0.77978515625
],
[
0.5146484375,
0.775390625,
0.55908203125,
0.775390625,
0.55908203125,
0.78759765625,
0.5146484375,
0.78759765625
],
[
0.5751953125,
0.77392578125,
0.82275390625,
0.77490234375,
0.82275390625,
0.7890625,
0.5751953125,
0.7880859375
],
[
0.12005615234375,
0.7822265625,
0.489501953125,
0.7822265625,
0.489501953125,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.57275390625,
0.7890625,
0.765625,
0.7890625,
0.765625,
0.802734375,
0.57275390625,
0.802734375
],
[
0.12005615234375,
0.79638671875,
0.2376708984375,
0.79638671875,
0.2376708984375,
0.81005859375,
0.12005615234375,
0.81005859375
],
[
0.1351318359375,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.1351318359375,
0.82373046875
],
[
0.5146484375,
0.81103515625,
0.880859375,
0.81103515625,
0.880859375,
0.82470703125,
0.5146484375,
0.82470703125
],
[
0.12005615234375,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.12005615234375,
0.83837890625
],
[
0.5146484375,
0.826171875,
0.8818359375,
0.826171875,
0.8818359375,
0.8388671875,
0.5146484375,
0.8388671875
],
[
0.118408203125,
0.83837890625,
0.489501953125,
0.83837890625,
0.489501953125,
0.8515625,
0.118408203125,
0.8515625
],
[
0.51123046875,
0.8388671875,
0.880859375,
0.8388671875,
0.880859375,
0.8525390625,
0.51123046875,
0.8525390625
],
[
0.12005615234375,
0.8525390625,
0.490234375,
0.8525390625,
0.490234375,
0.8662109375,
0.12005615234375,
0.8662109375
],
[
0.5146484375,
0.85400390625,
0.8818359375,
0.85400390625,
0.8818359375,
0.86767578125,
0.5146484375,
0.86767578125
],
[
0.12005615234375,
0.8681640625,
0.486083984375,
0.8681640625,
0.486083984375,
0.87890625,
0.12005615234375,
0.87890625
],
[
0.12005615234375,
0.8818359375,
0.484375,
0.8818359375,
0.484375,
0.8955078125,
0.12005615234375,
0.8955078125
],
[
0.5322265625,
0.88134765625,
0.88330078125,
0.8828125,
0.88330078125,
0.896484375,
0.5322265625,
0.89453125
],
[
0.118408203125,
0.89453125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.118408203125,
0.90771484375
],
[
0.51318359375,
0.896484375,
0.71875,
0.896484375,
0.71875,
0.91015625,
0.51318359375,
0.91015625
]
] |
[
"对于不是最强项目的半标度上的每个形容词，",
"族中，比例定律倾向于成立(i。e。，越大越好)。然",
"我们将其用作模板中的ADJweak，并获得5个",
"而，缩放并不能完全解释结果的模式。例如，Falcon",
"最可能的单词作为ADJ strong。例如，对于比",
"的性能远远低于较小的编码器模型，甚至低于基",
"例适当-精细-拟合良好，我们提示模型回答适",
"线，而Flan-T5-xxl的性能低于Flan-T5-xl",
"当/精细/拟合中可能是ADJstrong的内容，如果",
"不是ADJ strong，。如果ADJstrong中的任何前",
"3.3.ProbingScalar Intensity",
"5个单词与ADJweak的比例相同，并且重叠不",
"是微不足道的，我们认为这种情况是正确的（例",
"在本节中，我们将描述如何探究LLM是否了解相",
"如。g。adequateorevenadequateisatrimalcomple-",
"同规模的SA具有不同的强度（例如。9。热表示温",
"tion,goodorevenadequateisconsideredtobecorrectdesphgintensityranking)",
"3.3.1.",
"Scalar Intensity Direct Probing",
"3.2.4.",
"Scale Membership Indirect Probing",
"Experimentand Results",
"Method",
"对于编码器型号，我们将[MASK]和一个逗号放在",
"与以前的方法不同，这些方法将SA置于相同的上",
"ADJ strong的位置（e。g。ADJ weak，如果不是",
"下文中并在单独的运行中进行计算（例如。9。这是",
"[MASK]，）作为初步实验结果表明，BERT主要将",
"-",
"部很好/很棒/很棒/很棒的电影）（GariSolerand",
"标点符号预测为[MASK]的前5个预测。之所以使",
"Apidianaki,2020）（G&A），我们使用了一种新颖的",
"用逗号，是因为它不限制其前面的形容词是否是定",
"方法，该方法涉及将所有SA绑定在与单个输入相",
"语(i。e。可以紧跟名词)或谓语(i。e。不能紧跟",
"同的比例上，以获取其上下文化表示（e。g。good-",
"名词)。对于除GPT-4之外的非基线模型，我们在",
"greatwonderfulawesome),asillustratedinFigure2",
"给出上述提示的情况下直接生成ADJstrong。我",
"我们的假设是，当标量词在同一上下文中突出",
"们为GPT-4使用类似的生成目标，并在其他模型",
"时，人们更有可能注意到标量词具有不同的强度",
"中使用最佳模板（请参阅附录11.7）。我们使用",
"(Ronai and Xiang, 2023a)",
"2019(Lin et al.,2012)的GoogleNgram语料库切",
"对于每个半标度s，我们在s上对所有SAa进",
"片作为具有类似设置的间接探测基线，以获得每个",
"行十次混洗和绑定，以获得10个不同的输人（e",
"构造中ADJstrong的5个最有可能的完成。在表3",
"9。好极了，好极了，好极了，好极了，等等。）。对",
"中，我们报告了评估数据集中所有模板中的最佳结",
"于每个a，我们将其在10个输入中的表示平均，",
"果（我们根据在表E13中选择具有保留数据集的",
"以获得其最终表示。这是为了避免词序和标量强度",
"最佳模板来报告结果)。",
"之间的潜在启发式关联",
"在获得所有α的表示形式之后，我们利用",
"Models",
"CD",
"Gari SolerandApidianaki(2020）（G&A）中的最",
"WK",
"DM",
"0.3681",
"佳方法（DiffVec）来获取它们的强度排名。具体来",
"0.2871",
"GoogleNgram",
"",
"0.0751",
"BERT-b",
"0.1671",
"说，数据集D。中所有形容词α的强度是通过它们",
"0.0661",
"0.4251",
"0.1093",
"0.2501",
"与从另一个SA数据集Duec（Duec≠Da）得出的",
"0.5041",
"BERT-I",
"0.2714",
"0.0643",
"0.1141",
"外部全局强度向量dvec的余弦相似性来衡量的。。",
"RoBERTa-l",
"RoBERTa-b",
"0.4631",
"0.1001",
"0.2861",
"0.1671",
"对于Duec中的每个半标度s，S上最温和和极",
"0.2651",
"Falcon",
"0.0451",
"GPT-4",
"0.540",
"0.273",
"0.500",
"端的形容词表示为ams和aes。我们用方程（2）",
"0.5003",
"计算dvec",
"0.1773",
"Flan-T5-xl",
"0.5443",
"0.51530.1563",
"0.3643",
"1",
"Flan-T5-xxl",
"dVec=",
"_aes -ams",
"Dvee|DalseDeeD.",
"(2)",
"ship. The best results per dataset are in bold. Sub",
"Table 3: Indirect probing results for scale member-",
"scripts referto the indicesofthebesttemplates.",
"对于所有sEDa，s上的形容词均按其与dvec",
"的余弦相似性进行排名：较高的相似性意味着较高",
"间接探测始终不如所有模型的直接探测。此结",
"的强度。如果两个形容词具有相同的相似性，则将",
"果与 Petroni et al.(2019)一致。不同车型的相对",
"它们视为同等强度",
"排名基本一致。给定相同的大小等级，BERT优于",
"RoBERTa，并且较大的编码器型号优于具有相同",
"3.3.2.",
"ScalarIntensity Direct Probing",
"架构的较小的编码器型号。这些观察结果支持了我",
"ExperimentandResults",
"们方法的有效性",
"我们发现，除Falcon之外的所有模型都对有关",
"我们在BERT-b/l,RoBERTa-b/中探测逐层词嵌人",
"规模成员资格的丰富信息进行了编码，因为它们的",
"并报告在不同数据集中使用dvec跨层的成对准确",
"行为与基线相当或优于基线。GPT-4和Flan-T5-xl",
"性的最佳结果。我们再次在这里提供快速文本嵌入",
"在两个基准测试中均取得了最佳结果。虽然 GPT-",
"作为基线。对于基线，我们对所有sEDuecDa",
"4在DM和WK中具有与Flan-T5-xl竞争的结果",
"但对于CD来说要好得多。因此，我们得出结论",
"8We don't use the global intensity vector in Da due to",
"GPT-4是这项任务中表现最好的模型。在同一模型",
"the concern of generalisability."
] |
{
"arxiv_id": "2404.03301",
"language": "zh",
"page_idx": 3
}
|
||
2404.03301-zh-4
|
2404.03301-zh-4.jpg
|
<FigureHere>
Figure 2: The process to derive intensity vector $\vec{d_{vec}}.$ First, an adjective half-scale is randomly shuffled ten times for the order of adjectives as inputs to a language model. Then the encoded word vectors for the same word in different inputs are conducted with the Hadamard mean to derive the final representation of the word. After that, intensity vector $\vec{d_{vec}}$ is calculated by subtracting layer-wise representation of the weakest adjective from the strongest adjective $(awe\vec{s}ome-good$ in this case) then averaging over all relevant half-scale subtractions in a dataset. Then layer-wise $\vec{d_{vec}}$ is used to probe language models' knowledge for adjective intensities.
为它不包含关系
使用$\vec a_{es}$和$\vec a_ms$的静态嵌人来导出$\vec d\vec Vec$,如方程
2 所示
结果是对形容词对的平均值，如表 4 所示
## 3.3.3. Scalar Intensity Indirect Probing Method
<table>
<tbody>
<tr>
<th>Model</th>
<th>Method</th>
<th>$DM$</th>
<th>$CD$</th>
<th>$WK$</th>
</tr>
<tr>
<td>fast-text</td>
<td> </td>
<td>0.637 $WK$</td>
<td>0.685 $DM$</td>
<td>0.836 $CD$</td>
</tr>
<tr>
<td>BERT-b</td>
<td>$G A$ Ours</td>
<td>0.646 $\underline{CD}$ 0.639 $CD$</td>
<td>0.735 $\underline{DM}$ 0.706 $DM$</td>
<td>0.902 $DM$ 0.967 $\underline{DM}$</td>
</tr>
<tr>
<td>BERT-I</td>
<td>P $\overline{G 8A}$ Ours</td>
<td>$\textbf{1111111}^{***}VD$ 0.695 $\underline{CD}$ 0.673 $CD$</td>
<td>$\because DNE$ 0.731 $\underline{DM}$ 1 0.727 $DM$</td>
<td>$\textbf{Trr}+ Di$ 0.918 $DM$ 0.902</td>
</tr>
<tr>
<td>RoBERTa-b</td>
<td>$\overline{G 8A}$ Ours</td>
<td>0.557 $WK$ 0.648 $CD$</td>
<td>0.645 $DM$ 0.748 $DM$</td>
<td>0.820 $DM$ 0.934 $DM$</td>
</tr>
<tr>
<td>RoBERTa-1</td>
<td>$G A$ Ours</td>
<td>0.595 $CD$ 0.664 $CD$</td>
<td>0.682 $DM$ 0.752 $DM$</td>
<td>0.836 $DM$ 0.934 $DM$</td>
</tr>
</tbody>
</table>
为了补充闭源模型，我们使用与 Lorge and Pierrehumbert (2023) 类似的物镜间接探测强度。具体来说，对于每对用于比较强度的 SA,我们比较了包含这些形容词的最小对提示的困惑，例如“好但不好”和“很棒但不好”,前者的标量形容词顺序正确，后者不正确。理想情况下，模型应该总是为第一个提示计算较低的困惑，因为后者是愚蠢的，甚至是不合语法的
我们采用 de Melo and Bansal (2013) 数据收集中使用的所有 34 个模板。$^{10}$对于每个模板，我们比较正确和不正确结构的困惑。我们认为，如果正确的形容词对具有较低的困惑度，则 LLM 会对相应形容词对的标量强度进行编码。如果在模板中切换位置会产生同样不可能的短语，我们认为两个形容词的强度相同
Table 4: Direct scalar intensity probing with our method and G& A on different models. Subscripted letters refer to the source dataset of $\vec{d_{Vec}}$ . The best results per dataset are in bold. The best results using the same model are underlined.
### 3.3.4. Scalar Intensity Indirect Probing Experiment and Results
我们发现我们的简单方法通常比 G&A 更好。$^{9}$ 在全球范围内，我们的方法在三个数据集中的两个中实现了最佳排名结果。尽管 G&A 在 BERT 中的某些数据集上略微优于我们的方法，但我们的方法在 RoBERTa 中的表现始终优于 G&A,这表明它在不同的模型和数据集上更加健壮
所有模型都通过优于基线来编码有关 SA 强度的丰富信息。在所有模型中，BERT 模型再次在三分之二的数据集中占主导地位，表明它们比 RoBERTa 编码更多的强度信息，这表明更大的模型并不总是更好。在整个数据集中，WK 还是相对容易的，因
对于编码器模型，我们通过逐个掩蔽令牌并乘以原始令牌出现在掩码位置的可能性来计算序列的伪似然 (Salazar et al., 2020) 来估计伪困惑。对于除
$^{10}\mathrm{Except~ADJ~}_{strong~(,)~or~very~ADJ~}_{weak}$because some adjectives examined in this study are extreme ones (Paradis, 1998) which cannot be modified by very (e.g. very terrific ). See Appendix 11.4 for the full template list.
Our method is also superior to Garí Soler and Apid-
ianaki (2020) in other metrics. See appendix 11.2 .
|
[
[
0.364501953125,
0.09320068359375,
0.4072265625,
0.09320068359375,
0.4072265625,
0.10394287109375,
0.364501953125,
0.10394287109375
],
[
0.429931640625,
0.0968017578125,
0.44921875,
0.0968017578125,
0.44921875,
0.1009521484375,
0.429931640625,
0.1009521484375
],
[
0.330078125,
0.12469482421875,
0.4130859375,
0.12469482421875,
0.4130859375,
0.135986328125,
0.330078125,
0.135986328125
],
[
0.219970703125,
0.141357421875,
0.2509765625,
0.141357421875,
0.2509765625,
0.149658203125,
0.219970703125,
0.149658203125
],
[
0.47607421875,
0.141357421875,
0.55517578125,
0.141357421875,
0.55517578125,
0.1519775390625,
0.47607421875,
0.1519775390625
],
[
0.219970703125,
0.149658203125,
0.2476806640625,
0.149658203125,
0.2476806640625,
0.157958984375,
0.219970703125,
0.157958984375
],
[
0.57421875,
0.1455078125,
0.62646484375,
0.1455078125,
0.62646484375,
0.1561279296875,
0.57421875,
0.1561279296875
],
[
0.429931640625,
0.155029296875,
0.441650390625,
0.162109375,
0.43408203125,
0.168701171875,
0.424072265625,
0.160888671875
],
[
0.50146484375,
0.1519775390625,
0.5341796875,
0.1519775390625,
0.5341796875,
0.1591796875,
0.50146484375,
0.1591796875
],
[
0.5751953125,
0.1561279296875,
0.62744140625,
0.1561279296875,
0.62744140625,
0.163330078125,
0.5751953125,
0.163330078125
],
[
0.378662109375,
0.15673828125,
0.385498046875,
0.160888671875,
0.38037109375,
0.1644287109375,
0.37451171875,
0.1602783203125
],
[
0.21826171875,
0.15673828125,
0.258544921875,
0.15673828125,
0.258544921875,
0.16748046875,
0.21826171875,
0.16748046875
],
[
0.391357421875,
0.1591796875,
0.39794921875,
0.1591796875,
0.39794921875,
0.162109375,
0.391357421875,
0.162109375
],
[
0.509765625,
0.160888671875,
0.52880859375,
0.160888671875,
0.52880859375,
0.1650390625,
0.509765625,
0.1650390625
],
[
0.58251953125,
0.1650390625,
0.6171875,
0.1650390625,
0.6171875,
0.1776123046875,
0.58251953125,
0.1776123046875
],
[
0.64208984375,
0.162109375,
0.6943359375,
0.162109375,
0.6943359375,
0.1728515625,
0.64208984375,
0.1728515625
],
[
0.6484375,
0.17333984375,
0.6591796875,
0.174560546875,
0.6572265625,
0.1787109375,
0.646484375,
0.1776123046875
],
[
0.6591796875,
0.1728515625,
0.69287109375,
0.1728515625,
0.69287109375,
0.18115234375,
0.6591796875,
0.18115234375
],
[
0.587890625,
0.1787109375,
0.59375,
0.1817626953125,
0.587890625,
0.188232421875,
0.5810546875,
0.185302734375
],
[
0.6650390625,
0.1817626953125,
0.67529296875,
0.1817626953125,
0.67529296875,
0.18701171875,
0.6650390625,
0.18701171875
],
[
0.64404296875,
0.188232421875,
0.6943359375,
0.188232421875,
0.6943359375,
0.198974609375,
0.64404296875,
0.198974609375
],
[
0.73876953125,
0.1859130859375,
0.7958984375,
0.1859130859375,
0.7958984375,
0.193603515625,
0.73876953125,
0.193603515625
],
[
0.28369140625,
0.193603515625,
0.297119140625,
0.193603515625,
0.297119140625,
0.19775390625,
0.28369140625,
0.19775390625
],
[
0.32568359375,
0.193603515625,
0.339111328125,
0.193603515625,
0.339111328125,
0.19775390625,
0.32568359375,
0.19775390625
],
[
0.366943359375,
0.193603515625,
0.39794921875,
0.193603515625,
0.39794921875,
0.19775390625,
0.366943359375,
0.19775390625
],
[
0.405517578125,
0.189453125,
0.456787109375,
0.189453125,
0.456787109375,
0.1995849609375,
0.405517578125,
0.1995849609375
],
[
0.74072265625,
0.196533203125,
0.7958984375,
0.196533203125,
0.7958984375,
0.2037353515625,
0.74072265625,
0.2037353515625
],
[
0.32666015625,
0.221435546875,
0.41064453125,
0.221435546875,
0.41064453125,
0.2327880859375,
0.32666015625,
0.2327880859375
],
[
0.57861328125,
0.232177734375,
0.62890625,
0.232177734375,
0.62890625,
0.2423095703125,
0.57861328125,
0.2423095703125
],
[
0.28369140625,
0.251708984375,
0.31494140625,
0.251708984375,
0.31494140625,
0.26416015625,
0.28369140625,
0.26416015625
],
[
0.31982421875,
0.254150390625,
0.333251953125,
0.254150390625,
0.333251953125,
0.25830078125,
0.31982421875,
0.25830078125
],
[
0.4248046875,
0.25048828125,
0.455078125,
0.25048828125,
0.455078125,
0.26318359375,
0.4248046875,
0.26318359375
],
[
0.118408203125,
0.32666015625,
0.884765625,
0.32666015625,
0.884765625,
0.34033203125,
0.118408203125,
0.34033203125
],
[
0.118408203125,
0.34033203125,
0.8818359375,
0.34033203125,
0.8818359375,
0.35400390625,
0.118408203125,
0.35400390625
],
[
0.118408203125,
0.355712890625,
0.884765625,
0.355712890625,
0.884765625,
0.369384765625,
0.118408203125,
0.369384765625
],
[
0.12005615234375,
0.37060546875,
0.884765625,
0.37060546875,
0.884765625,
0.38427734375,
0.12005615234375,
0.38427734375
],
[
0.118408203125,
0.385498046875,
0.884765625,
0.385498046875,
0.884765625,
0.3984375,
0.118408203125,
0.3984375
],
[
0.118408203125,
0.3984375,
0.8818359375,
0.3984375,
0.8818359375,
0.412109375,
0.118408203125,
0.412109375
],
[
0.118408203125,
0.413330078125,
0.378662109375,
0.413330078125,
0.378662109375,
0.427001953125,
0.118408203125,
0.427001953125
],
[
0.12005615234375,
0.451904296875,
0.490234375,
0.451904296875,
0.490234375,
0.466064453125,
0.12005615234375,
0.466064453125
],
[
0.5146484375,
0.453125,
0.6337890625,
0.453125,
0.6337890625,
0.466064453125,
0.5146484375,
0.466064453125
],
[
0.118408203125,
0.466064453125,
0.1695556640625,
0.466064453125,
0.1695556640625,
0.48095703125,
0.118408203125,
0.48095703125
],
[
0.1351318359375,
0.48095703125,
0.443359375,
0.48095703125,
0.443359375,
0.49462890625,
0.1351318359375,
0.49462890625
],
[
0.5146484375,
0.48388671875,
0.56103515625,
0.48388671875,
0.56103515625,
0.495361328125,
0.5146484375,
0.495361328125
],
[
0.57421875,
0.482177734375,
0.8369140625,
0.482666015625,
0.8369140625,
0.49755859375,
0.57421875,
0.496337890625
],
[
0.57275390625,
0.49755859375,
0.63623046875,
0.49755859375,
0.63623046875,
0.5087890625,
0.57275390625,
0.5087890625
],
[
0.14697265625,
0.5048828125,
0.187255859375,
0.5048828125,
0.187255859375,
0.5166015625,
0.14697265625,
0.5166015625
],
[
0.365234375,
0.5048828125,
0.3896484375,
0.5048828125,
0.3896484375,
0.51708984375,
0.365234375,
0.51708984375
],
[
0.4375,
0.5048828125,
0.46337890625,
0.5048828125,
0.46337890625,
0.51708984375,
0.4375,
0.51708984375
],
[
0.2939453125,
0.505859375,
0.316650390625,
0.505859375,
0.316650390625,
0.5166015625,
0.2939453125,
0.5166015625
],
[
0.2149658203125,
0.505859375,
0.267822265625,
0.505859375,
0.267822265625,
0.5166015625,
0.2149658203125,
0.5166015625
],
[
0.27197265625,
0.51416015625,
0.334228515625,
0.5166015625,
0.334228515625,
0.52978515625,
0.27197265625,
0.52783203125
],
[
0.344970703125,
0.51416015625,
0.4072265625,
0.51708984375,
0.405517578125,
0.52978515625,
0.343505859375,
0.52685546875
],
[
0.418212890625,
0.51416015625,
0.484375,
0.51708984375,
0.483642578125,
0.529296875,
0.41650390625,
0.525390625
],
[
0.51708984375,
0.51953125,
0.88330078125,
0.51953125,
0.88330078125,
0.533203125,
0.51708984375,
0.533203125
],
[
0.344970703125,
0.525390625,
0.405517578125,
0.529296875,
0.40478515625,
0.54052734375,
0.343505859375,
0.53759765625
],
[
0.1435546875,
0.5166015625,
0.19140625,
0.5166015625,
0.19140625,
0.52783203125,
0.1435546875,
0.52783203125
],
[
0.27197265625,
0.52685546875,
0.333251953125,
0.529296875,
0.333251953125,
0.5390625,
0.27197265625,
0.53759765625
],
[
0.418212890625,
0.525390625,
0.480224609375,
0.52783203125,
0.480224609375,
0.5390625,
0.418212890625,
0.53759765625
],
[
0.1435546875,
0.533203125,
0.19140625,
0.533203125,
0.19140625,
0.54443359375,
0.1435546875,
0.54443359375
],
[
0.219970703125,
0.52783203125,
0.26123046875,
0.52783203125,
0.26123046875,
0.5390625,
0.219970703125,
0.5390625
],
[
0.345947265625,
0.53515625,
0.4072265625,
0.5380859375,
0.405517578125,
0.552734375,
0.344970703125,
0.5498046875
],
[
0.4189453125,
0.5361328125,
0.478515625,
0.54052734375,
0.476806640625,
0.552734375,
0.418212890625,
0.548828125
],
[
0.5146484375,
0.53369140625,
0.87890625,
0.53369140625,
0.87890625,
0.54443359375,
0.5146484375,
0.54443359375
],
[
0.273681640625,
0.53759765625,
0.334228515625,
0.5390625,
0.334228515625,
0.55029296875,
0.273681640625,
0.548828125
],
[
0.2225341796875,
0.53759765625,
0.258544921875,
0.5390625,
0.2568359375,
0.55029296875,
0.2216796875,
0.548828125
],
[
0.27197265625,
0.54736328125,
0.333251953125,
0.5498046875,
0.333251953125,
0.56103515625,
0.27197265625,
0.55859375
],
[
0.345947265625,
0.54736328125,
0.405517578125,
0.55029296875,
0.40478515625,
0.5625,
0.344970703125,
0.55859375
],
[
0.418212890625,
0.54736328125,
0.480224609375,
0.5498046875,
0.480224609375,
0.56103515625,
0.418212890625,
0.55859375
],
[
0.515625,
0.54736328125,
0.8818359375,
0.54736328125,
0.8818359375,
0.56103515625,
0.515625,
0.56103515625
],
[
0.145263671875,
0.55419921875,
0.189697265625,
0.55419921875,
0.189697265625,
0.5654296875,
0.145263671875,
0.5654296875
],
[
0.219970703125,
0.5498046875,
0.26123046875,
0.5498046875,
0.26123046875,
0.56103515625,
0.219970703125,
0.56103515625
],
[
0.273681640625,
0.55810546875,
0.333251953125,
0.56005859375,
0.333251953125,
0.57177734375,
0.273681640625,
0.5693359375
],
[
0.345947265625,
0.55810546875,
0.405517578125,
0.56103515625,
0.40478515625,
0.572265625,
0.344970703125,
0.5693359375
],
[
0.4189453125,
0.55810546875,
0.480224609375,
0.56103515625,
0.478515625,
0.572265625,
0.418212890625,
0.5693359375
],
[
0.515625,
0.5625,
0.88330078125,
0.5625,
0.88330078125,
0.576171875,
0.515625,
0.576171875
],
[
0.418212890625,
0.56640625,
0.484375,
0.57080078125,
0.48193359375,
0.5849609375,
0.41650390625,
0.58056640625
],
[
0.22412109375,
0.56103515625,
0.2568359375,
0.56103515625,
0.2568359375,
0.5693359375,
0.22412109375,
0.5693359375
],
[
0.27197265625,
0.568359375,
0.334228515625,
0.572265625,
0.333251953125,
0.58447265625,
0.270263671875,
0.580078125
],
[
0.343505859375,
0.568359375,
0.405517578125,
0.57177734375,
0.40478515625,
0.5830078125,
0.341796875,
0.580078125
],
[
0.5146484375,
0.57470703125,
0.88330078125,
0.576171875,
0.88330078125,
0.59033203125,
0.5146484375,
0.5888671875
],
[
0.219970703125,
0.57080078125,
0.26123046875,
0.57080078125,
0.26123046875,
0.58203125,
0.219970703125,
0.58203125
],
[
0.1318359375,
0.576171875,
0.2032470703125,
0.576171875,
0.2032470703125,
0.58740234375,
0.1318359375,
0.58740234375
],
[
0.2225341796875,
0.58056640625,
0.258544921875,
0.58056640625,
0.258544921875,
0.5927734375,
0.2225341796875,
0.5927734375
],
[
0.273681640625,
0.58056640625,
0.33154296875,
0.58056640625,
0.33154296875,
0.5927734375,
0.273681640625,
0.5927734375
],
[
0.344970703125,
0.58056640625,
0.403076171875,
0.58056640625,
0.403076171875,
0.5927734375,
0.344970703125,
0.5927734375
],
[
0.4189453125,
0.58056640625,
0.47607421875,
0.58056640625,
0.47607421875,
0.5927734375,
0.4189453125,
0.5927734375
],
[
0.274658203125,
0.59033203125,
0.333251953125,
0.59326171875,
0.33154296875,
0.60546875,
0.273681640625,
0.6025390625
],
[
0.345947265625,
0.5888671875,
0.40478515625,
0.5927734375,
0.403076171875,
0.60693359375,
0.344970703125,
0.60400390625
],
[
0.418212890625,
0.59130859375,
0.48193359375,
0.59326171875,
0.48193359375,
0.60498046875,
0.418212890625,
0.6025390625
],
[
0.5146484375,
0.59033203125,
0.88330078125,
0.59033203125,
0.88330078125,
0.60400390625,
0.5146484375,
0.60400390625
],
[
0.133544921875,
0.59716796875,
0.20068359375,
0.59716796875,
0.20068359375,
0.609375,
0.133544921875,
0.609375
],
[
0.2216796875,
0.5927734375,
0.26025390625,
0.5927734375,
0.26025390625,
0.60400390625,
0.2216796875,
0.60400390625
],
[
0.270263671875,
0.6015625,
0.333251953125,
0.60400390625,
0.333251953125,
0.615234375,
0.270263671875,
0.61328125
],
[
0.344970703125,
0.60400390625,
0.4072265625,
0.60400390625,
0.4072265625,
0.615234375,
0.344970703125,
0.615234375
],
[
0.418212890625,
0.6015625,
0.478515625,
0.60498046875,
0.476806640625,
0.6162109375,
0.41650390625,
0.61328125
],
[
0.5263671875,
0.60400390625,
0.88330078125,
0.60400390625,
0.88330078125,
0.61767578125,
0.5263671875,
0.61767578125
],
[
0.515625,
0.619140625,
0.634765625,
0.619140625,
0.634765625,
0.6328125,
0.515625,
0.6328125
],
[
0.22412109375,
0.60498046875,
0.2568359375,
0.60498046875,
0.2568359375,
0.61328125,
0.22412109375,
0.61328125
],
[
0.118408203125,
0.6259765625,
0.489501953125,
0.6259765625,
0.489501953125,
0.64013671875,
0.118408203125,
0.64013671875
],
[
0.53076171875,
0.6328125,
0.88330078125,
0.6328125,
0.88330078125,
0.646484375,
0.53076171875,
0.646484375
],
[
0.118408203125,
0.64013671875,
0.456787109375,
0.64013671875,
0.456787109375,
0.65380859375,
0.118408203125,
0.65380859375
],
[
0.44921875,
0.6396484375,
0.486083984375,
0.64111328125,
0.484375,
0.65283203125,
0.447509765625,
0.65087890625
],
[
0.5146484375,
0.6435546875,
0.884765625,
0.64453125,
0.884765625,
0.662109375,
0.5146484375,
0.6611328125
],
[
0.1175537109375,
0.65380859375,
0.476806640625,
0.65478515625,
0.476806640625,
0.66845703125,
0.1175537109375,
0.66748046875
],
[
0.118408203125,
0.66748046875,
0.48779296875,
0.66748046875,
0.48779296875,
0.68115234375,
0.118408203125,
0.68115234375
],
[
0.515625,
0.662109375,
0.880859375,
0.662109375,
0.880859375,
0.6728515625,
0.515625,
0.6728515625
],
[
0.5146484375,
0.6748046875,
0.8818359375,
0.6748046875,
0.8818359375,
0.6884765625,
0.5146484375,
0.6884765625
],
[
0.1175537109375,
0.681640625,
0.45166015625,
0.681640625,
0.45166015625,
0.6953125,
0.1175537109375,
0.6953125
],
[
0.515625,
0.6904296875,
0.88330078125,
0.6904296875,
0.88330078125,
0.70361328125,
0.515625,
0.70361328125
],
[
0.51318359375,
0.703125,
0.884765625,
0.70361328125,
0.884765625,
0.71875,
0.51318359375,
0.71728515625
],
[
0.133544921875,
0.71142578125,
0.489501953125,
0.71044921875,
0.489501953125,
0.724609375,
0.133544921875,
0.7255859375
],
[
0.5146484375,
0.71875,
0.634765625,
0.71875,
0.634765625,
0.732421875,
0.5146484375,
0.732421875
],
[
0.12005615234375,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.74072265625,
0.12005615234375,
0.74072265625
],
[
0.118408203125,
0.7392578125,
0.486083984375,
0.7392578125,
0.486083984375,
0.7529296875,
0.118408203125,
0.7529296875
],
[
0.5146484375,
0.75,
0.56103515625,
0.75,
0.56103515625,
0.76123046875,
0.5146484375,
0.76123046875
],
[
0.5751953125,
0.75,
0.8369140625,
0.75,
0.8369140625,
0.76318359375,
0.5751953125,
0.76318359375
],
[
0.57275390625,
0.76318359375,
0.767578125,
0.76318359375,
0.767578125,
0.77685546875,
0.57275390625,
0.77685546875
],
[
0.12005615234375,
0.7548828125,
0.48779296875,
0.7548828125,
0.48779296875,
0.7685546875,
0.12005615234375,
0.7685546875
],
[
0.118408203125,
0.76708984375,
0.486083984375,
0.76708984375,
0.486083984375,
0.78076171875,
0.118408203125,
0.78076171875
],
[
0.12005615234375,
0.783203125,
0.372802734375,
0.783203125,
0.372802734375,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.51708984375,
0.78759765625,
0.8818359375,
0.78759765625,
0.8818359375,
0.79736328125,
0.51708984375,
0.79736328125
],
[
0.1368408203125,
0.79736328125,
0.489501953125,
0.79736328125,
0.489501953125,
0.81103515625,
0.1368408203125,
0.81103515625
],
[
0.515625,
0.7998046875,
0.8818359375,
0.7998046875,
0.8818359375,
0.8134765625,
0.515625,
0.8134765625
],
[
0.118408203125,
0.81103515625,
0.484375,
0.81103515625,
0.484375,
0.82470703125,
0.118408203125,
0.82470703125
],
[
0.51318359375,
0.8134765625,
0.8818359375,
0.8134765625,
0.8818359375,
0.82666015625,
0.51318359375,
0.82666015625
],
[
0.1234130859375,
0.826171875,
0.489501953125,
0.826171875,
0.489501953125,
0.8388671875,
0.1234130859375,
0.8388671875
],
[
0.12005615234375,
0.84033203125,
0.48779296875,
0.84033203125,
0.48779296875,
0.85400390625,
0.12005615234375,
0.85400390625
],
[
0.5283203125,
0.8388671875,
0.88671875,
0.84130859375,
0.88671875,
0.85986328125,
0.5283203125,
0.85791015625
],
[
0.12005615234375,
0.85498046875,
0.489501953125,
0.85498046875,
0.489501953125,
0.8681640625,
0.12005615234375,
0.8681640625
],
[
0.5146484375,
0.85791015625,
0.88330078125,
0.85791015625,
0.88330078125,
0.8681640625,
0.5146484375,
0.8681640625
],
[
0.5146484375,
0.869140625,
0.88330078125,
0.869140625,
0.88330078125,
0.8828125,
0.5146484375,
0.8828125
],
[
0.139404296875,
0.8818359375,
0.489501953125,
0.8818359375,
0.489501953125,
0.8955078125,
0.139404296875,
0.8955078125
],
[
0.51123046875,
0.8828125,
0.88330078125,
0.8828125,
0.88330078125,
0.896484375,
0.51123046875,
0.896484375
],
[
0.12005615234375,
0.89794921875,
0.455078125,
0.89794921875,
0.455078125,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.5146484375,
0.896484375,
0.5751953125,
0.896484375,
0.5751953125,
0.90771484375,
0.5146484375,
0.90771484375
]
] |
[
"onderful",
"veso",
"Language Model",
"Shuffle",
"Hadamard mean",
"input",
"Layer-wise",
"X",
"overall",
"subtraction",
"",
"10 times",
"",
"onto",
"good",
"Hadamard",
"",
"an over",
"C",
"all",
"half-scales",
"Laverdveo",
"rer",
"000",
"esoms",
"wonderful",
"Laver2dveo",
"Language Model",
"awesome",
"",
"",
"",
"Figure 2: The process to derive intensity vector duec . First, an adjective half-scale is randomly shuffled",
"ten times for the order of adjectives as inputs to a language model. Then the encoded word vectors for the",
"same word in different inputs are conducted with the Hadamard mean to derive the final representation",
"of the word. After that, intensity vector duec is calculated by subtracting layer-wise representation of the",
"weakest adjective from the strongest adjective ( awesome -good in this case) then averaging over all",
"relevant half-scale subtractions in a dataset. Then layer-wise dwec is used to probe language models",
"knowledge for adjective intensities.",
"使用aes和ams的静态嵌入来导出dVec，如方程",
"为它不包含关系",
"2所示",
"结果是对形容词对的平均值，如表4所示",
"3.3.3.",
"ScalarIntensity Indirect Probing",
"Method",
"Model",
"CD",
"WK",
"DM",
"Method",
"0.637WK",
"0.685DM",
"0.836cD",
"为了补充闭源模型，我们使用与LorgeandPierre-",
"0.735DM",
"fast-text",
"0.646CD",
"0.902DM",
"BERT-b",
"G&A",
"0.706DM",
"0.967DM",
"humbert（2023）类似的物镜间接探测强度。具体来",
"0.639cD",
"Ours",
"0.695cD",
"0.731DM",
"0.918DM",
"说，对于每对用于比较强度的SA，我们比较了包",
"BERT-I",
"G&A",
"0.673cD",
"0.727DM",
"0.902DM",
"含这些形容词的最小对提示的困惑，例如“好但不",
"0.820DM",
"Ours",
"0.557wK",
"0.645DM",
"好”和“很棒但不好”，前者的标量形容词顺序正",
"G&A",
"RoBERTa-b",
"Ours",
"0.648cD",
"0.748 DM",
"0.934 DM",
"0.595CD",
"0.682DM",
"0.836DM",
"确，后者不正确。理想情况下，模型应该总是为第",
"RoBERTa-I",
"G&A",
"0.664CD",
"0.752 DM",
"0.934DM",
"个提示计算较低的困惑，因为后者是愚蠢的，甚",
"至是不合语法的",
"Ours",
"Table 4: Direct scalar intensity probing with our",
"我们采用deMeloandBansal（2013）数据收集",
"method and G & A on different models.",
"Sub",
"中使用的所有34个模板。10对于每个模板，我们",
"scripted letters refer to the source dataset of dvec",
"The best results per dataset are in bold. The best",
"比较正确和不正确结构的困惑。我们认为，如果正",
"确的形容词对具有较低的困惑度，则LLM会对相",
"results using the same model are underlined",
"应形容词对的标量强度进行编码。如果在模板中切",
"换位置会产生同样不可能的短语，我们认为两个形",
"我们发现我们的简单方法通常比G&A更好。9",
"容词的强度相同",
"在全球范围内，我们的方法在三个数据集中的两个",
"中实现了最佳排名结果。尽管G&A在BERT中的",
"3.3.4.",
"ScalarIntensity IndirectProbing",
"Experimentand Results",
"某些数据集上略微优于我们的方法，但我们的方法",
"在RoBERTa中的表现始终优于G&A，这表明它",
"在不同的模型和数据集上更加健壮",
"对于编码器模型，我们通过逐个掩蔽令牌并乘以原",
"所有模型都通过优于基线来编码有关SA强度的",
"始令牌出现在掩码位置的可能性来计算序列的伪",
"丰富信息。在所有模型中，BERT模型再次在三分之",
"似然（Salazaretal.，2020）来估计伪困惑。对于除",
"二的数据集中占主导地位，表明它们比RoBERTa",
"编码更多的强度信息，这表明更大的模型并不总是",
"10Except ADJ strong () or very ADJ weak because",
"更好。在整个数据集中，WK还是相对容易的，因",
"someadjectives examinedinthisstudyare extreme",
"ones (Paradis, 1998) which cannot be modified by very",
"'Our method is also superior to Gari Soler and Apid-",
"(e.g. very terrific ). See Appendix 11.4 for the full tem-",
"ianaki(2020)inothermetrics.Seeappendix11.2",
"plate list."
] |
{
"arxiv_id": "2404.03301",
"language": "zh",
"page_idx": 4
}
|
||
2205.13913-zh-1
|
2205.13913-zh-1.jpg
|
# 动态域泛化
Zhishu Sun$^1$, Zhifeng Shen$^1$, Luojun Lin$^1,\dagger$, Yuanlong Yu$^1,\dagger$, Zhifeng Yang$^1$
Shicai Yang$^2$ and Weijie Chen$^2$
$^{1}$College of Computer and Data Science, Fuzhou University, Fuzhou, China
$^{2}$Hikvision Research Institute, Hangzhou, China $\operatorname*{linluojun2009@126.com,~yu.yuanlong@fzu.edu.cn}$
$\operatorname{Abstract}$
Domain generalization (DG) is a fundamental yet very challenging research topic in machine learning. The existing arts mainly focus on $\operatorname*{learning}\overset{\circ}{\operatorname*{\operatorname*{domain-invariant~features~with~limited}}}$ source domains in a static model. Unfortunately, there is a lack of training-free mechanism to adjust the model when generalized to the agnostic target domains. To tackle this problem,we develop a brand-new DG variant, namely Dynamic Domain Generalization ( DDG) , in which the model learns to twist the network parameters to adapt the data from different domains. Specifically, we leverage a meta-adjuster to twist the network parameters based on the static model with respect to different data from different domains. In this way, the static model is optimized to learn domainshared features, while the meta-adjuster is designed to learn domain-specific features. To enable this process, DomainMix is exploited to simulate ${\mathrm{~data~from~diverse~domains~during}}$ teaching the meta-adjuster to adapt the upcoming agnostic target domains. This learning mechanism urges the model to generalize to different agnostic target domains via adjusting the model without training. Extensive experiments demonstrate the effectiveness of our proposed method.
Code is available:
https://github.com/MetaVisionLab/DDG
## 1 Introduction
深度神经网络在各种视觉任务上取得了巨大成功，例如图像识别和物体检测$\boxed{\mathrm{E}}[$He et al., 2016]。大多数深度模型都是在封闭的感知环境中学习的，在这种环境中，训练和测试数据应该在。i。d(独立且相同的分布)假设。然而，现实是开放的，复合的，不受控制的，与i不相容的。i。d 假设。这种复杂的场景会使传统模型遭受巨大的性能下降，由于缺乏泛化能力来处理域转移，这无疑会降低智能系统的可靠性和安全性，例如。g。,自动驾驶系统
${^{\dagger}\text{Correspon}}$ding authors: Luojun Lin and Yuanlong Yu
广泛的研究旨在通过域泛化(DG)解决此问题，其目的是通过利用多个源域E[ Ganin and Lempitsky, 2015; Zhou et al., 2021] 学习域不变表示来获得鲁棒的静态模型。然而，静态模型的一个严重瓶颈限制了它推广到具有不同分布的不可知目标域的自由。相反，探索了域自适应(DA$)\boxtimes$E[Ganin et al., 2016] 和测试时间 DG$\boxed{\mathrm{F}}[$Iwasawa and Matsuo, 2021] , 以利用从目标域采样的未标记数据进行进一步的无监督微调，从而增强了面向目标的泛化能力。但是，在测试过程中需要额外的培训工作和延迟， 而制造的芯片和边缘设备大多不具备培训资源。考虑到这两个角度，当遇到来自新目标域的测试数据时，缺乏无训练机制来调整模型
为了补充这个空白，我们提出了一个新的 DG 变体， 动态域泛化 (DDG)。如图 1 所示，与 DA, DG 以及测试时 DG 方法不同，所提出的 DDG 附带了一个元调整器， 该元调整器在预训练期间学习适应新领域，并在测试期间扭曲模型，而无需任何微调。为了推动学习适应的过程，我们需要从众多新颖的领域收集训练数据，而现实是我们通常受到有限的源域的阻碍。一个简单而友好的解决方案是，我们通过不同源域之间的随机凸插值来模拟来自有限源域的新样本，在本文中称为 DomainMix。这样，每个模拟样本都被限制在一个唯一的域中。因此域感知元调整器可以放宽为实例感知元调整器，它可以调整每个实例的网络参数。此外，通过在优化元调整器期间丢弃域标签的利用，它使模型更容易从源域适应具有不可定义域区别的不可知目标域
在 DDG 的优化流水线之外，它也是如何设计元调整器的关键因素。为了分离域共享特征和域特定特征，网络参数被解耦为静态部分和动态部分，其中后者由每个实例的元调整器生成。为了简化元调整器的优化，将动态参数分解为具有低维动态系数的多个精心设计的内核模板的线性组合，其中系数由元调整器通过实例嵌人动态生成作为输人。如图 3 所示，为了在空间和通道维度上加强内核骨架，内核模板被设计为四个非对称内核，这明确地将元调整器正则化，以生成不同的动态参数，以适应各种各样的新样本
在三个流行的 DG 基准上进行了广泛的实验，包括PACS[Li et al., 2017] , Office-Home [Venkateswara et al., 2017] 和 DomainNet [Peng et al., 2019] 。这些数据集包含从小到大的各种情况，从简单到复杂的场景，从小到大的领域转移，足以证明我们提出的方法的有效性。综上所述，DDG 开辟了研究 DG 问题的新范式
|
[
[
0.441162109375,
0.11297607421875,
0.55810546875,
0.11297607421875,
0.55810546875,
0.1307373046875,
0.441162109375,
0.1307373046875
],
[
0.1748046875,
0.1514892578125,
0.8203125,
0.1514892578125,
0.8203125,
0.1666259765625,
0.1748046875,
0.1666259765625
],
[
0.363525390625,
0.1685791015625,
0.63134765625,
0.169189453125,
0.63134765625,
0.184326171875,
0.363525390625,
0.18310546875
],
[
0.1846923828125,
0.1900634765625,
0.814453125,
0.1925048828125,
0.814453125,
0.20703125,
0.1846923828125,
0.2052001953125
],
[
0.29736328125,
0.2083740234375,
0.701171875,
0.208984375,
0.701171875,
0.2235107421875,
0.29736328125,
0.222900390625
],
[
0.291748046875,
0.2259521484375,
0.70361328125,
0.2266845703125,
0.70361328125,
0.2442626953125,
0.291748046875,
0.24365234375
],
[
0.2484130859375,
0.271484375,
0.324462890625,
0.2734375,
0.324462890625,
0.2890625,
0.2484130859375,
0.287353515625
],
[
0.53173828125,
0.274658203125,
0.90869140625,
0.274658203125,
0.90869140625,
0.28857421875,
0.53173828125,
0.28857421875
],
[
0.515625,
0.28857421875,
0.91259765625,
0.28857421875,
0.91259765625,
0.3017578125,
0.515625,
0.3017578125
],
[
0.1192626953125,
0.296630859375,
0.453369140625,
0.296630859375,
0.453369140625,
0.311279296875,
0.1192626953125,
0.311279296875
],
[
0.5146484375,
0.3017578125,
0.9111328125,
0.3017578125,
0.9111328125,
0.31494140625,
0.5146484375,
0.31494140625
],
[
0.1192626953125,
0.311279296875,
0.451904296875,
0.311279296875,
0.451904296875,
0.324462890625,
0.1192626953125,
0.324462890625
],
[
0.5146484375,
0.31640625,
0.91015625,
0.31640625,
0.91015625,
0.330078125,
0.5146484375,
0.330078125
],
[
0.12091064453125,
0.32568359375,
0.451904296875,
0.32568359375,
0.451904296875,
0.339111328125,
0.12091064453125,
0.339111328125
],
[
0.515625,
0.330078125,
0.91259765625,
0.330078125,
0.91259765625,
0.343505859375,
0.515625,
0.343505859375
],
[
0.11767578125,
0.33837890625,
0.453369140625,
0.337158203125,
0.453369140625,
0.3515625,
0.11767578125,
0.352783203125
],
[
0.5146484375,
0.343505859375,
0.9111328125,
0.343505859375,
0.9111328125,
0.356689453125,
0.5146484375,
0.356689453125
],
[
0.1192626953125,
0.352783203125,
0.450927734375,
0.352783203125,
0.450927734375,
0.3662109375,
0.1192626953125,
0.3662109375
],
[
0.51318359375,
0.35595703125,
0.91259765625,
0.356689453125,
0.91259765625,
0.371826171875,
0.51318359375,
0.37060546875
],
[
0.1192626953125,
0.367431640625,
0.451904296875,
0.367431640625,
0.451904296875,
0.380615234375,
0.1192626953125,
0.380615234375
],
[
0.5146484375,
0.371826171875,
0.9111328125,
0.371826171875,
0.9111328125,
0.385009765625,
0.5146484375,
0.385009765625
],
[
0.1192626953125,
0.380615234375,
0.453369140625,
0.380615234375,
0.453369140625,
0.39453125,
0.1192626953125,
0.39453125
],
[
0.5146484375,
0.385009765625,
0.91015625,
0.385009765625,
0.91015625,
0.3984375,
0.5146484375,
0.3984375
],
[
0.11767578125,
0.393310546875,
0.450927734375,
0.393310546875,
0.450927734375,
0.4072265625,
0.11767578125,
0.4072265625
],
[
0.5146484375,
0.399658203125,
0.9111328125,
0.399658203125,
0.9111328125,
0.41357421875,
0.5146484375,
0.41357421875
],
[
0.11767578125,
0.40771484375,
0.450927734375,
0.4072265625,
0.450927734375,
0.421630859375,
0.11767578125,
0.42236328125
],
[
0.5146484375,
0.41357421875,
0.91015625,
0.41357421875,
0.91015625,
0.4267578125,
0.5146484375,
0.4267578125
],
[
0.1192626953125,
0.42236328125,
0.451904296875,
0.42236328125,
0.451904296875,
0.436279296875,
0.1192626953125,
0.436279296875
],
[
0.5146484375,
0.4267578125,
0.6787109375,
0.4267578125,
0.6787109375,
0.43994140625,
0.5146484375,
0.43994140625
],
[
0.12091064453125,
0.436279296875,
0.451904296875,
0.436279296875,
0.451904296875,
0.449462890625,
0.12091064453125,
0.449462890625
],
[
0.1192626953125,
0.449462890625,
0.450927734375,
0.449462890625,
0.450927734375,
0.46337890625,
0.1192626953125,
0.46337890625
],
[
0.53173828125,
0.451904296875,
0.90869140625,
0.451904296875,
0.90869140625,
0.46533203125,
0.53173828125,
0.46533203125
],
[
0.11767578125,
0.4609375,
0.453369140625,
0.46337890625,
0.453369140625,
0.477783203125,
0.11767578125,
0.47607421875
],
[
0.5146484375,
0.464111328125,
0.91015625,
0.46533203125,
0.91015625,
0.479736328125,
0.5146484375,
0.479248046875
],
[
0.12091064453125,
0.4765625,
0.451904296875,
0.477783203125,
0.451904296875,
0.4912109375,
0.12091064453125,
0.490478515625
],
[
0.51318359375,
0.479736328125,
0.90673828125,
0.479736328125,
0.90673828125,
0.49365234375,
0.51318359375,
0.49365234375
],
[
0.12091064453125,
0.4912109375,
0.451904296875,
0.4912109375,
0.451904296875,
0.5048828125,
0.12091064453125,
0.5048828125
],
[
0.5146484375,
0.49365234375,
0.9111328125,
0.49365234375,
0.9111328125,
0.5068359375,
0.5146484375,
0.5068359375
],
[
0.11767578125,
0.50244140625,
0.451904296875,
0.5048828125,
0.451904296875,
0.51953125,
0.11767578125,
0.517578125
],
[
0.515625,
0.5068359375,
0.9111328125,
0.5068359375,
0.9111328125,
0.52099609375,
0.515625,
0.52099609375
],
[
0.12091064453125,
0.52099609375,
0.449462890625,
0.52099609375,
0.449462890625,
0.53076171875,
0.12091064453125,
0.53076171875
],
[
0.51318359375,
0.51953125,
0.91259765625,
0.52099609375,
0.91259765625,
0.5341796875,
0.51318359375,
0.53271484375
],
[
0.1192626953125,
0.53271484375,
0.450927734375,
0.53271484375,
0.450927734375,
0.546875,
0.1192626953125,
0.546875
],
[
0.515625,
0.5341796875,
0.9111328125,
0.5341796875,
0.9111328125,
0.54736328125,
0.515625,
0.54736328125
],
[
0.1192626953125,
0.546875,
0.451904296875,
0.546875,
0.451904296875,
0.56005859375,
0.1192626953125,
0.56005859375
],
[
0.5146484375,
0.548828125,
0.9111328125,
0.548828125,
0.9111328125,
0.5625,
0.5146484375,
0.5625
],
[
0.12091064453125,
0.5625,
0.450927734375,
0.5625,
0.450927734375,
0.57275390625,
0.12091064453125,
0.57275390625
],
[
0.5146484375,
0.5625,
0.9111328125,
0.5625,
0.9111328125,
0.57568359375,
0.5146484375,
0.57568359375
],
[
0.1192626953125,
0.57373046875,
0.450927734375,
0.57470703125,
0.450927734375,
0.58837890625,
0.1192626953125,
0.5869140625
],
[
0.5146484375,
0.5771484375,
0.91015625,
0.5771484375,
0.91015625,
0.59033203125,
0.5146484375,
0.59033203125
],
[
0.11767578125,
0.5869140625,
0.450927734375,
0.58837890625,
0.450927734375,
0.60302734375,
0.11767578125,
0.6015625
],
[
0.51318359375,
0.5888671875,
0.9111328125,
0.59033203125,
0.9111328125,
0.60400390625,
0.51318359375,
0.60302734375
],
[
0.11767578125,
0.6015625,
0.450927734375,
0.60107421875,
0.450927734375,
0.61572265625,
0.11767578125,
0.6162109375
],
[
0.5146484375,
0.60302734375,
0.9111328125,
0.60400390625,
0.9111328125,
0.6171875,
0.5146484375,
0.6162109375
],
[
0.12091064453125,
0.6171875,
0.450927734375,
0.6171875,
0.450927734375,
0.62744140625,
0.12091064453125,
0.62744140625
],
[
0.51318359375,
0.6162109375,
0.91015625,
0.6162109375,
0.91015625,
0.6298828125,
0.51318359375,
0.6298828125
],
[
0.1192626953125,
0.6298828125,
0.451904296875,
0.6298828125,
0.451904296875,
0.64306640625,
0.1192626953125,
0.64306640625
],
[
0.515625,
0.6318359375,
0.7587890625,
0.6318359375,
0.7587890625,
0.64599609375,
0.515625,
0.64599609375
],
[
0.11767578125,
0.64306640625,
0.390625,
0.64306640625,
0.390625,
0.6572265625,
0.11767578125,
0.6572265625
],
[
0.400390625,
0.64453125,
0.449462890625,
0.64453125,
0.449462890625,
0.65576171875,
0.400390625,
0.65576171875
],
[
0.11767578125,
0.65771484375,
0.449462890625,
0.65771484375,
0.449462890625,
0.66845703125,
0.11767578125,
0.66845703125
],
[
0.53125,
0.65576171875,
0.91259765625,
0.65576171875,
0.91259765625,
0.66943359375,
0.53125,
0.66943359375
],
[
0.1192626953125,
0.671875,
0.306396484375,
0.671875,
0.306396484375,
0.68505859375,
0.1192626953125,
0.68505859375
],
[
0.324462890625,
0.67236328125,
0.451904296875,
0.67236328125,
0.451904296875,
0.68310546875,
0.324462890625,
0.68310546875
],
[
0.5146484375,
0.67041015625,
0.9111328125,
0.67041015625,
0.9111328125,
0.6845703125,
0.5146484375,
0.6845703125
],
[
0.1192626953125,
0.68505859375,
0.420654296875,
0.68505859375,
0.420654296875,
0.69970703125,
0.1192626953125,
0.69970703125
],
[
0.5146484375,
0.6845703125,
0.9111328125,
0.6845703125,
0.9111328125,
0.69775390625,
0.5146484375,
0.69775390625
],
[
0.515625,
0.69970703125,
0.91015625,
0.69970703125,
0.91015625,
0.71044921875,
0.515625,
0.71044921875
],
[
0.515625,
0.7119140625,
0.91015625,
0.7119140625,
0.91015625,
0.72607421875,
0.515625,
0.72607421875
],
[
0.0882568359375,
0.72265625,
0.10540771484375,
0.72265625,
0.10540771484375,
0.73095703125,
0.0882568359375,
0.73095703125
],
[
0.11767578125,
0.72021484375,
0.2296142578125,
0.72021484375,
0.2296142578125,
0.734375,
0.11767578125,
0.734375
],
[
0.5146484375,
0.72607421875,
0.91259765625,
0.72607421875,
0.91259765625,
0.7392578125,
0.5146484375,
0.7392578125
],
[
0.0882568359375,
0.7412109375,
0.486083984375,
0.7412109375,
0.486083984375,
0.7548828125,
0.0882568359375,
0.7548828125
],
[
0.5146484375,
0.74072265625,
0.9111328125,
0.74072265625,
0.9111328125,
0.75390625,
0.5146484375,
0.75390625
],
[
0.08905029296875,
0.7548828125,
0.484375,
0.7548828125,
0.484375,
0.76806640625,
0.08905029296875,
0.76806640625
],
[
0.5146484375,
0.75390625,
0.9111328125,
0.75390625,
0.9111328125,
0.767578125,
0.5146484375,
0.767578125
],
[
0.0882568359375,
0.767578125,
0.483642578125,
0.767578125,
0.483642578125,
0.78076171875,
0.0882568359375,
0.78076171875
],
[
0.5146484375,
0.767578125,
0.91259765625,
0.767578125,
0.91259765625,
0.78076171875,
0.5146484375,
0.78076171875
],
[
0.0882568359375,
0.7822265625,
0.484375,
0.7822265625,
0.484375,
0.79541015625,
0.0882568359375,
0.79541015625
],
[
0.51123046875,
0.7822265625,
0.67724609375,
0.7822265625,
0.67724609375,
0.79541015625,
0.51123046875,
0.79541015625
],
[
0.08660888671875,
0.79443359375,
0.484375,
0.796875,
0.484375,
0.81005859375,
0.08660888671875,
0.80810546875
],
[
0.0882568359375,
0.81005859375,
0.484375,
0.81005859375,
0.484375,
0.82373046875,
0.0882568359375,
0.82373046875
],
[
0.53125,
0.80615234375,
0.9111328125,
0.80615234375,
0.9111328125,
0.8193359375,
0.53125,
0.8193359375
],
[
0.08660888671875,
0.82373046875,
0.484375,
0.82275390625,
0.484375,
0.8369140625,
0.08660888671875,
0.83837890625
],
[
0.51123046875,
0.8193359375,
0.91259765625,
0.8193359375,
0.91259765625,
0.83349609375,
0.51123046875,
0.83349609375
],
[
0.0882568359375,
0.83837890625,
0.484375,
0.83837890625,
0.484375,
0.85302734375,
0.0882568359375,
0.85302734375
],
[
0.51318359375,
0.833984375,
0.91015625,
0.833984375,
0.91015625,
0.84765625,
0.51318359375,
0.84765625
],
[
0.08905029296875,
0.85302734375,
0.136474609375,
0.85302734375,
0.136474609375,
0.8642578125,
0.08905029296875,
0.8642578125
],
[
0.5146484375,
0.84765625,
0.91015625,
0.84765625,
0.91015625,
0.861328125,
0.5146484375,
0.861328125
],
[
0.515625,
0.8623046875,
0.9111328125,
0.8623046875,
0.9111328125,
0.87548828125,
0.515625,
0.87548828125
],
[
0.101318359375,
0.87353515625,
0.4697265625,
0.87548828125,
0.4697265625,
0.89013671875,
0.101318359375,
0.88818359375
],
[
0.51123046875,
0.87548828125,
0.8544921875,
0.87548828125,
0.8544921875,
0.8896484375,
0.51123046875,
0.8896484375
]
] |
[
"动态域泛化",
"Zhishu Sun,Zhifeng Shen',Luojun Linlt,Yuanlong Yui.t,Zhifeng Yang",
"Shicai Yang2 and Weijie Chen?",
"1College of Computer and Data Science, Fuzhou University, Fuzhou, China",
"2Hikvision Research Institute, Hangzhou, China",
"linluojun2009@126.com, yu.yuanlong@fzu.edu.cn",
"Abstract",
"广泛的研究旨在通过域泛化（DG）解决此问题，其目",
"的是通过利用多个源域E[GaninandLempitsky,2015;",
"Domain generalization (DG) is a fundamental",
"Zhouetal.,2021]学习域不变表示来获得鲁棒的静态模",
"yet very challenging research topic in machine",
"型。然而，静态模型的一个严重瓶颈限制了它推广到具有",
"learning.The existing arts mainly focus on",
"不同分布的不可知目标域的自由。相反，探索了域自适应",
"learning domain-invariant features with limited",
"(DA)E[Ganinetal.,2016]和测试时间DGE[Iwasawa",
"source domains in a static model.Unfortu-",
"andMatsuo，2021]，以利用从目标域采样的未标记数据",
"nately, there is a lack of training-free mecha-",
"进行进一步的无监督微调，从而增强了面向目标的泛化",
"nism to adjust the model when generalized to",
"能力。但是，在测试过程中需要额外的培训工作和延迟",
"the agnostic target domains.To tackle this",
"而制造的芯片和边缘设备大多不具备培训资源。考虑到",
"problem, we develop a brand-new DG vari-",
"这两个角度，当遇到来自新目标域的测试数据时，缺乏",
"ant, namely Dynamic Domain Generalization",
"无训练机制来调整模型",
"(DDG), in which the model learns to twist the",
"network parameters to adapt the data from",
"为了补充这个空白，我们提出了一个新的DG变体",
"different domains. Specifically, we leverage a",
"动态域泛化（DDG)。如图1所示，与DA，DG以及测试",
"meta-adjuster to twist the network parameters",
"时DG方法不同，所提出的DDG附带了一个元调整器",
"based on the static model with respect to dif-",
"该元调整器在预训练期间学习适应新领域，并在测试期",
"ferent data from different domains. In this way,",
"间扭曲模型，而无需任何微调。为了推动学习适应的过",
"the staticmodel is optimized to learn domain",
"程，我们需要从众多新颖的领域收集训练数据，而现实",
"shared features, while the meta-adjuster is de-",
"是我们通常受到有限的源域的阻碍。一个简单而友好的",
"signed to learn domain-specific features. To",
"解决方案是，我们通过不同源域之间的随机凸插值来模",
"enable thisprocess,DomainMixisexploited",
"拟来自有限源域的新样本，在本文中称为DomainMix",
"to simulate data from diverse domains during",
"这样，每个模拟样本都被限制在一个唯一的域中。因此",
"teaching the meta-adjuster to adapt the up",
"域感知元调整器可以放宽为实例感知元调整器，它可以",
"coming agnostic target domains. This learn",
"调整每个实例的网络参数。此外，通过在优化元调整器",
"ing mechanism urges the model to generalize",
"期间丢弃域标签的利用，它使模型更容易从源域适应具",
"to different agnostic target domains via ad-",
"有不可定义域区别的不可知目标域",
"justing the model without training.",
"Exten",
"siveexperiments demonstratetheeffectiveness",
"在DDG的优化流水线之外，它也是如何设计元调整",
"of our proposed method.",
"Code is available:",
"器的关键因素。为了分离域共享特征和域特定特征，网",
"https://github.com/MetaVisionLab/DDG",
"络参数被解耦为静态部分和动态部分，其中后者由每个",
"实例的元调整器生成。为了简化元调整器的优化，将动",
"态参数分解为具有低维动态系数的多个精心设计的内核",
"",
"Introduction",
"模板的线性组合，其中系数由元调整器通过实例嵌人动",
"深度神经网络在各种视觉任务上取得了巨大成功，例如",
"态生成作为输人。如图3所示，为了在空间和通道维度",
"图像识别和物体检测E[Heetal.,2016]。大多数深度模",
"上加强内核骨架，内核模板被设计为四个非对称内核，这",
"型都是在封闭的感知环境中学习的，在这种环境中，训",
"明确地将元调整器正则化，以生成不同的动态参数，以",
"练和测试数据应该在。i。d(独立且相同的分布）假设。然",
"适应各种各样的新样本",
"而，现实是开放的，复合的，不受控制的，与i不相容",
"的。i。d假设。这种复杂的场景会使传统模型遭受巨大",
"在三个流行的DG基准上进行了广泛的实验，包括",
"的性能下降，由于缺乏泛化能力来处理域转移，这无疑",
"PACS [Li et al., 2017] , Office-Home [Venkateswara et",
"会降低智能系统的可靠性和安全性，例如。g。，自动驾",
"al.,2017]和DomainNet[Pengetal.,2019]。这些数据",
"驶系统",
"集包含从小到大的各种情况，从简单到复杂的场景，从",
"小到大的领域转移，足以证明我们提出的方法的有效性。",
" Corresponding authors: Luojun Lin and Yuanlong Yu",
"综上所述，DDG开辟了研究DG问题的新范式"
] |
{
"arxiv_id": "2205.13913",
"language": "zh",
"page_idx": 1
}
|
||
2205.13913-zh-2
|
2205.13913-zh-2.jpg
|
<FigureHere>
Figure 1: Comparison among DA, DG, Test-time DG and DDG methods. DA and test-time DG are able to adjust the model when encountering novel target domains via finetuning, while DG freezes a static model to adapt any agnostic target domain. Different from the existing methods, DDG is able to adjust the model without any training effort for better domain generalization.
# 2 Related Work
Domain Adaptation.
常规 DA 旨在将从标记的源域
中学到的知识转移到可能包含一些标记样本或根本不包含的目标域。基于差异的方法是早期开发的，它通过最小化定义明确的距离损失[Haeusser et al.,2017] 来减轻两个域之间的分布差异。DA 方法的另一行基于对抗性学习，其中经常使用域鉴别器与特征提取器进行对抗性训练，以便通过混淆域鉴别器[Ganin et al.,2016] 来强制特征提取器学习域不变特征。与调整模型以明确适应目标域的 DA 方法相比，我们的方法可以被视为一种通过仅在源域中进行预训练的无训练 DA
Domain Generalization.
早期的 DG 相关研究集中
于学习域不变表示，如果它可以抑制多个源域之间的域偏移，则被认为是健壮的$\mathbb{E}[$Motiian et al.,2017]。另一种 DG 方法是基于数据增强，增强图像风格的多样性，以优化模型的泛化能力[Zhou et al.,2020a; Zhou et al., 2021] 。自监督学习也被应用于学习任务无关表示的 DG, 其过拟合概率较小 [Carlucci et al., 2019; Kim et al., 2021]。元学习还涉及解决 DG$\underline{\mathrm{E}}[$Balaji et al., 2018] , 其中通过模拟源域中的元训练和元测试任务来训练深度模型，以增强其泛化能力。我们的方法可以看作是一个非常简单的元学习范式，其中元调整器学习为每个单独的“域”学习参数泛化。此外，一些新兴的作品为 DG 展示了一个新的见解，即深度模型可以在测试时适应目标域 [Iwasawa and Matsuo, 2021] 。然而，
Train Time
<FigureHere>
Figure 2: Pipeline of the proposed DDG method. In the training phase, two stages are required, namely novel samples simulation and learning to adjust, which aims to teach the meta-adjuster to adapt novel samples. In the testing phase, the network can be adjusted without training when encountering agnostic target domain.
在这种方法中，计算成本是不可避免的。相反，我们的方法可以直接更新参数以适应不可知的目标域，而无需任何反向传播
3 Methods
## 3.1 Overview
DG 旨在训练一个通用模型，该模型可以很好地推广到任何不可知的目标域。特别是，我们在本文中只考虑$K$ -way 分类任务。对于香草 DG 任务，我们从多源域$\mathcal{X}_S=\{\mathcal{X}_{S_1},\mathcal{X}_{S_2},...,\mathcal{X}_{S_M}\}$中提取$n_s$标记的样本$\{ x_{s}^{i}, y_{s}^{i}, d_{s}^{i}\} _{i= 1}^{n_{s}}$ , 其中$x_s^i\in \mathcal{X} _S$ , $y_s^{i}\in \mathcal{Y} _S$和域标签$d_s^i\in$ $[1,M]$。DG 的目标是学习映射$f_\Theta:\{\mathcal{X}_{S_1},...,\mathcal{X}_{S_M}\}\to$ ${\mathcal{Y} }_{S}$ , 该映射也可以在任何不可知的目标域上进行精确预测
Vanilla Domain Generalization. 对于香草 DG, 在训练阶段获得的学习映射$f_\Theta$将在推理中保持静态。由于源域和目标域之间不可避免的数据分布差异，它通常会导致不可知目标域的性能下降
Dynamic Domain Generalization. 为了克服香草 DG 的弱点，我们提出了 DG 的新变体，即动态域泛化(DDG)。DDG 主要侧重于学习扭曲网络参数以适应来自不同新域的实例。在本文中，每个实例都被假设为
|
[
[
0.19775390625,
0.0625,
0.2156982421875,
0.0625,
0.2156982421875,
0.0706787109375,
0.19775390625,
0.0706787109375
],
[
0.35302734375,
0.061248779296875,
0.373291015625,
0.061248779296875,
0.373291015625,
0.07135009765625,
0.35302734375,
0.07135009765625
],
[
0.52392578125,
0.06817626953125,
0.5849609375,
0.06817626953125,
0.5849609375,
0.07891845703125,
0.52392578125,
0.07891845703125
],
[
0.56396484375,
0.07891845703125,
0.67333984375,
0.07891845703125,
0.67333984375,
0.08587646484375,
0.56396484375,
0.08587646484375
],
[
0.7685546875,
0.07891845703125,
0.81396484375,
0.07763671875,
0.81396484375,
0.085205078125,
0.7685546875,
0.08587646484375
],
[
0.3349609375,
0.10162353515625,
0.3505859375,
0.10162353515625,
0.3505859375,
0.1060791015625,
0.3349609375,
0.1060791015625
],
[
0.78515625,
0.10986328125,
0.7958984375,
0.10986328125,
0.7958984375,
0.1287841796875,
0.78515625,
0.1287841796875
],
[
0.85693359375,
0.10986328125,
0.87158203125,
0.10986328125,
0.87158203125,
0.1351318359375,
0.85693359375,
0.1351318359375
],
[
0.53759765625,
0.11932373046875,
0.54736328125,
0.11932373046875,
0.54736328125,
0.125,
0.53759765625,
0.125
],
[
0.400390625,
0.111083984375,
0.4248046875,
0.111083984375,
0.4248046875,
0.11553955078125,
0.400390625,
0.11553955078125
],
[
0.31787109375,
0.121826171875,
0.3447265625,
0.121826171875,
0.3447265625,
0.1300048828125,
0.31787109375,
0.1300048828125
],
[
0.81884765625,
0.1331787109375,
0.830078125,
0.1363525390625,
0.82666015625,
0.1395263671875,
0.81787109375,
0.1363525390625
],
[
0.348876953125,
0.14453125,
0.386474609375,
0.14453125,
0.386474609375,
0.1546630859375,
0.348876953125,
0.1546630859375
],
[
0.3896484375,
0.14453125,
0.406005859375,
0.14453125,
0.406005859375,
0.1483154296875,
0.3896484375,
0.1483154296875
],
[
0.85693359375,
0.1427001953125,
0.87158203125,
0.1427001953125,
0.87158203125,
0.169189453125,
0.85693359375,
0.169189453125
],
[
0.53515625,
0.1558837890625,
0.54833984375,
0.1558837890625,
0.54833984375,
0.1641845703125,
0.53515625,
0.1641845703125
],
[
0.1993408203125,
0.147705078125,
0.221435546875,
0.147705078125,
0.221435546875,
0.1514892578125,
0.1993408203125,
0.1514892578125
],
[
0.7861328125,
0.1558837890625,
0.7958984375,
0.1558837890625,
0.7958984375,
0.1673583984375,
0.7861328125,
0.1673583984375
],
[
0.6435546875,
0.1634521484375,
0.65869140625,
0.1634521484375,
0.65869140625,
0.17236328125,
0.6435546875,
0.17236328125
],
[
0.76708984375,
0.17236328125,
0.8203125,
0.17236328125,
0.8203125,
0.184326171875,
0.76708984375,
0.184326171875
],
[
0.8486328125,
0.17236328125,
0.88134765625,
0.17236328125,
0.88134765625,
0.184326171875,
0.8486328125,
0.184326171875
],
[
0.580078125,
0.189453125,
0.5908203125,
0.189453125,
0.5908203125,
0.203857421875,
0.580078125,
0.203857421875
],
[
0.71337890625,
0.189453125,
0.7236328125,
0.189453125,
0.7236328125,
0.20263671875,
0.71337890625,
0.20263671875
],
[
0.328369140625,
0.222900390625,
0.39453125,
0.222900390625,
0.39453125,
0.23291015625,
0.328369140625,
0.23291015625
],
[
0.78759765625,
0.222900390625,
0.79345703125,
0.222900390625,
0.79345703125,
0.2279052734375,
0.78759765625,
0.2279052734375
],
[
0.861328125,
0.2215576171875,
0.86767578125,
0.2215576171875,
0.86767578125,
0.2279052734375,
0.861328125,
0.2279052734375
],
[
0.53759765625,
0.231689453125,
0.54736328125,
0.231689453125,
0.54736328125,
0.2374267578125,
0.53759765625,
0.2374267578125
],
[
0.1748046875,
0.2235107421875,
0.2467041015625,
0.2235107421875,
0.2467041015625,
0.2342529296875,
0.1748046875,
0.2342529296875
],
[
0.303955078125,
0.25244140625,
0.32177734375,
0.25244140625,
0.32177734375,
0.261474609375,
0.303955078125,
0.261474609375
],
[
0.611328125,
0.246826171875,
0.6259765625,
0.246826171875,
0.6259765625,
0.25390625,
0.611328125,
0.25390625
],
[
0.859375,
0.260009765625,
0.869140625,
0.260009765625,
0.869140625,
0.271484375,
0.859375,
0.271484375
],
[
0.193603515625,
0.276611328125,
0.2017822265625,
0.276611328125,
0.2017822265625,
0.27978515625,
0.193603515625,
0.27978515625
],
[
0.64208984375,
0.270751953125,
0.66015625,
0.270751953125,
0.66015625,
0.285400390625,
0.64208984375,
0.285400390625
],
[
0.7607421875,
0.2822265625,
0.814453125,
0.2822265625,
0.814453125,
0.292236328125,
0.7607421875,
0.292236328125
],
[
0.8486328125,
0.281005859375,
0.88134765625,
0.281005859375,
0.88134765625,
0.292236328125,
0.8486328125,
0.292236328125
],
[
0.17236328125,
0.29345703125,
0.188720703125,
0.29345703125,
0.188720703125,
0.29736328125,
0.17236328125,
0.29736328125
],
[
0.330078125,
0.285888671875,
0.340576171875,
0.285888671875,
0.340576171875,
0.2890625,
0.330078125,
0.2890625
],
[
0.35546875,
0.3037109375,
0.375732421875,
0.3037109375,
0.375732421875,
0.311279296875,
0.35546875,
0.311279296875
],
[
0.52783203125,
0.306884765625,
0.580078125,
0.306884765625,
0.580078125,
0.317626953125,
0.52783203125,
0.317626953125
],
[
0.200927734375,
0.324462890625,
0.2115478515625,
0.324462890625,
0.2115478515625,
0.32763671875,
0.200927734375,
0.32763671875
],
[
0.81787109375,
0.322509765625,
0.82861328125,
0.324462890625,
0.82421875,
0.333984375,
0.81396484375,
0.33203125
],
[
0.85546875,
0.33837890625,
0.88232421875,
0.33837890625,
0.88232421875,
0.3466796875,
0.85546875,
0.3466796875
],
[
0.57421875,
0.375732421875,
0.59228515625,
0.375732421875,
0.59228515625,
0.389404296875,
0.57421875,
0.389404296875
],
[
0.79150390625,
0.38330078125,
0.8447265625,
0.38330078125,
0.8447265625,
0.39453125,
0.79150390625,
0.39453125
],
[
0.1380615234375,
0.410400390625,
0.21728515625,
0.410400390625,
0.21728515625,
0.417236328125,
0.1380615234375,
0.417236328125
],
[
0.238525390625,
0.410400390625,
0.31787109375,
0.410400390625,
0.31787109375,
0.417236328125,
0.238525390625,
0.417236328125
],
[
0.35302734375,
0.410400390625,
0.4306640625,
0.410400390625,
0.4306640625,
0.417236328125,
0.35302734375,
0.417236328125
],
[
0.576171875,
0.417236328125,
0.650390625,
0.417236328125,
0.650390625,
0.431884765625,
0.576171875,
0.431884765625
],
[
0.65869140625,
0.41845703125,
0.744140625,
0.41845703125,
0.744140625,
0.429931640625,
0.65869140625,
0.429931640625
],
[
0.08660888671875,
0.4248046875,
0.48779296875,
0.4248046875,
0.48779296875,
0.438232421875,
0.08660888671875,
0.438232421875
],
[
0.77783203125,
0.419189453125,
0.85107421875,
0.419189453125,
0.85107421875,
0.429931640625,
0.77783203125,
0.429931640625
],
[
0.08660888671875,
0.4375,
0.48779296875,
0.4375,
0.48779296875,
0.45068359375,
0.08660888671875,
0.45068359375
],
[
0.5498046875,
0.4375,
0.6435546875,
0.4375,
0.6435546875,
0.447509765625,
0.5498046875,
0.447509765625
],
[
0.67724609375,
0.4375,
0.76318359375,
0.4375,
0.76318359375,
0.447509765625,
0.67724609375,
0.447509765625
],
[
0.79736328125,
0.4375,
0.8935546875,
0.4375,
0.8935546875,
0.447509765625,
0.79736328125,
0.447509765625
],
[
0.08660888671875,
0.449462890625,
0.486083984375,
0.449462890625,
0.486083984375,
0.46337890625,
0.08660888671875,
0.46337890625
],
[
0.08660888671875,
0.462158203125,
0.48779296875,
0.462158203125,
0.48779296875,
0.47607421875,
0.08660888671875,
0.47607421875
],
[
0.51318359375,
0.46728515625,
0.9140625,
0.46728515625,
0.9140625,
0.48095703125,
0.51318359375,
0.48095703125
],
[
0.0849609375,
0.474853515625,
0.488525390625,
0.474853515625,
0.488525390625,
0.488037109375,
0.0849609375,
0.488037109375
],
[
0.51123046875,
0.48095703125,
0.91259765625,
0.48095703125,
0.91259765625,
0.494384765625,
0.51123046875,
0.494384765625
],
[
0.08660888671875,
0.4892578125,
0.484375,
0.4892578125,
0.484375,
0.5,
0.08660888671875,
0.5
],
[
0.51318359375,
0.49365234375,
0.9140625,
0.49365234375,
0.9140625,
0.5068359375,
0.51318359375,
0.5068359375
],
[
0.0849609375,
0.50048828125,
0.238525390625,
0.50048828125,
0.238525390625,
0.5146484375,
0.0849609375,
0.5146484375
],
[
0.51318359375,
0.505859375,
0.9140625,
0.505859375,
0.9140625,
0.51953125,
0.51318359375,
0.51953125
],
[
0.51318359375,
0.51806640625,
0.9140625,
0.51806640625,
0.9140625,
0.5322265625,
0.51318359375,
0.5322265625
],
[
0.08660888671875,
0.52978515625,
0.098876953125,
0.52978515625,
0.098876953125,
0.54052734375,
0.08660888671875,
0.54052734375
],
[
0.12091064453125,
0.52880859375,
0.240234375,
0.52880859375,
0.240234375,
0.54248046875,
0.12091064453125,
0.54248046875
],
[
0.5107421875,
0.53076171875,
0.7607421875,
0.52978515625,
0.7607421875,
0.54345703125,
0.5107421875,
0.54443359375
],
[
0.09234619140625,
0.5498046875,
0.2442626953125,
0.5498046875,
0.2442626953125,
0.56298828125,
0.09234619140625,
0.56298828125
],
[
0.26318359375,
0.548828125,
0.48779296875,
0.5498046875,
0.48779296875,
0.564453125,
0.26318359375,
0.56298828125
],
[
0.08905029296875,
0.56298828125,
0.484375,
0.56298828125,
0.484375,
0.57763671875,
0.08905029296875,
0.57763671875
],
[
0.51123046875,
0.55810546875,
0.9111328125,
0.5595703125,
0.9111328125,
0.57373046875,
0.51123046875,
0.57275390625
],
[
0.08905029296875,
0.57763671875,
0.484375,
0.57763671875,
0.484375,
0.59130859375,
0.08905029296875,
0.59130859375
],
[
0.515625,
0.57373046875,
0.91259765625,
0.57373046875,
0.91259765625,
0.5869140625,
0.515625,
0.5869140625
],
[
0.51318359375,
0.58544921875,
0.61376953125,
0.5869140625,
0.61376953125,
0.6015625,
0.51318359375,
0.599609375
],
[
0.0882568359375,
0.59228515625,
0.484375,
0.59228515625,
0.484375,
0.60595703125,
0.0882568359375,
0.60595703125
],
[
0.08660888671875,
0.60400390625,
0.484375,
0.60498046875,
0.484375,
0.619140625,
0.08660888671875,
0.61865234375
],
[
0.0882568359375,
0.619140625,
0.484375,
0.619140625,
0.484375,
0.63330078125,
0.0882568359375,
0.63330078125
],
[
0.5107421875,
0.62060546875,
0.52294921875,
0.6171875,
0.52685546875,
0.626953125,
0.5146484375,
0.6298828125
],
[
0.54833984375,
0.61865234375,
0.62353515625,
0.61865234375,
0.62353515625,
0.6298828125,
0.54833984375,
0.6298828125
],
[
0.0882568359375,
0.6318359375,
0.484375,
0.6318359375,
0.484375,
0.64599609375,
0.0882568359375,
0.64599609375
],
[
0.51318359375,
0.63818359375,
0.5390625,
0.63818359375,
0.5390625,
0.65087890625,
0.51318359375,
0.65087890625
],
[
0.55615234375,
0.63623046875,
0.6357421875,
0.63818359375,
0.6357421875,
0.65283203125,
0.55615234375,
0.65087890625
],
[
0.08905029296875,
0.646484375,
0.484375,
0.646484375,
0.484375,
0.6611328125,
0.08905029296875,
0.6611328125
],
[
0.51123046875,
0.65478515625,
0.90869140625,
0.65576171875,
0.90869140625,
0.67041015625,
0.51123046875,
0.66943359375
],
[
0.08905029296875,
0.6611328125,
0.486083984375,
0.6611328125,
0.486083984375,
0.6748046875,
0.08905029296875,
0.6748046875
],
[
0.0882568359375,
0.67529296875,
0.36279296875,
0.67529296875,
0.36279296875,
0.689453125,
0.0882568359375,
0.689453125
],
[
0.5146484375,
0.67041015625,
0.9111328125,
0.67041015625,
0.9111328125,
0.6845703125,
0.5146484375,
0.6845703125
],
[
0.5146484375,
0.6845703125,
0.9111328125,
0.6845703125,
0.9111328125,
0.69873046875,
0.5146484375,
0.69873046875
],
[
0.09234619140625,
0.693359375,
0.26953125,
0.693359375,
0.26953125,
0.70703125,
0.09234619140625,
0.70703125
],
[
0.289306640625,
0.693359375,
0.484375,
0.6943359375,
0.484375,
0.708984375,
0.289306640625,
0.70849609375
],
[
0.5146484375,
0.69873046875,
0.9111328125,
0.69873046875,
0.9111328125,
0.71337890625,
0.5146484375,
0.71337890625
],
[
0.08905029296875,
0.70849609375,
0.486083984375,
0.70849609375,
0.486083984375,
0.7216796875,
0.08905029296875,
0.7216796875
],
[
0.5146484375,
0.71044921875,
0.91259765625,
0.71044921875,
0.91259765625,
0.72802734375,
0.5146484375,
0.72802734375
],
[
0.0882568359375,
0.72265625,
0.484375,
0.72265625,
0.484375,
0.7373046875,
0.0882568359375,
0.7373046875
],
[
0.51318359375,
0.72265625,
0.89453125,
0.724609375,
0.89453125,
0.7421875,
0.51318359375,
0.74072265625
],
[
0.88818359375,
0.72998046875,
0.91015625,
0.72998046875,
0.91015625,
0.7373046875,
0.88818359375,
0.7373046875
],
[
0.101318359375,
0.7373046875,
0.484375,
0.7373046875,
0.484375,
0.75048828125,
0.101318359375,
0.75048828125
],
[
0.51123046875,
0.73779296875,
0.91259765625,
0.74072265625,
0.91259765625,
0.7548828125,
0.51123046875,
0.75244140625
],
[
0.08660888671875,
0.75048828125,
0.488525390625,
0.75048828125,
0.488525390625,
0.7646484375,
0.08660888671875,
0.7646484375
],
[
0.515625,
0.755859375,
0.53125,
0.755859375,
0.53125,
0.76513671875,
0.515625,
0.76513671875
],
[
0.08660888671875,
0.76513671875,
0.484375,
0.76513671875,
0.484375,
0.77880859375,
0.08660888671875,
0.77880859375
],
[
0.5185546875,
0.77294921875,
0.7490234375,
0.77294921875,
0.7490234375,
0.7861328125,
0.5185546875,
0.7861328125
],
[
0.76318359375,
0.77294921875,
0.91259765625,
0.77294921875,
0.91259765625,
0.7861328125,
0.76318359375,
0.7861328125
],
[
0.5146484375,
0.7861328125,
0.9111328125,
0.7861328125,
0.9111328125,
0.80029296875,
0.5146484375,
0.80029296875
],
[
0.0882568359375,
0.77783203125,
0.483642578125,
0.77783203125,
0.483642578125,
0.79150390625,
0.0882568359375,
0.79150390625
],
[
0.08660888671875,
0.79150390625,
0.486083984375,
0.79150390625,
0.486083984375,
0.80615234375,
0.08660888671875,
0.80615234375
],
[
0.08416748046875,
0.8037109375,
0.484375,
0.8046875,
0.484375,
0.8193359375,
0.08416748046875,
0.818359375
],
[
0.5146484375,
0.80029296875,
0.9111328125,
0.80029296875,
0.9111328125,
0.814453125,
0.5146484375,
0.814453125
],
[
0.51123046875,
0.81298828125,
0.72607421875,
0.814453125,
0.72607421875,
0.8291015625,
0.51123046875,
0.82763671875
],
[
0.0882568359375,
0.8193359375,
0.484375,
0.8193359375,
0.484375,
0.83349609375,
0.0882568359375,
0.83349609375
],
[
0.0882568359375,
0.83349609375,
0.483642578125,
0.83349609375,
0.483642578125,
0.84765625,
0.0882568359375,
0.84765625
],
[
0.5185546875,
0.833984375,
0.7646484375,
0.833984375,
0.7646484375,
0.84765625,
0.5185546875,
0.84765625
],
[
0.779296875,
0.83349609375,
0.91259765625,
0.83349609375,
0.91259765625,
0.8466796875,
0.779296875,
0.8466796875
],
[
0.08905029296875,
0.84765625,
0.484375,
0.84765625,
0.484375,
0.861328125,
0.08905029296875,
0.861328125
],
[
0.5146484375,
0.8466796875,
0.91259765625,
0.8466796875,
0.91259765625,
0.861328125,
0.5146484375,
0.861328125
],
[
0.08905029296875,
0.861328125,
0.484375,
0.861328125,
0.484375,
0.87548828125,
0.08905029296875,
0.87548828125
],
[
0.5107421875,
0.85986328125,
0.91259765625,
0.861328125,
0.91259765625,
0.87548828125,
0.5107421875,
0.875
],
[
0.0882568359375,
0.87548828125,
0.486083984375,
0.87548828125,
0.486083984375,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.5146484375,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.5146484375,
0.8896484375
]
] |
[
"DA",
"DG",
"Train Time",
"NovelSamplesSimulatior",
"toAdiust",
"",
"C",
"",
"",
"Sketch",
"Train",
"0",
"Model",
"Teast",
"",
"t:",
"do",
"图",
"T.",
"Prediction",
"Label",
"...",
"..",
"DDG (Ours)",
"3",
"编",
"T",
"Test-Time DG",
"A",
"X",
"",
"he",
"a",
"Prediction",
"Label",
"LE",
"a",
"Meta",
"Test Time",
"A",
"7",
"Dog",
"ci",
"Prediction",
"SourceSamples",
"TargetSamples",
"-Decision Boundary",
"CDomainMix",
"Meta-Adjuster",
"Figure l: Comparison among DA, DG, Test-time DG and",
"Static Component",
"DDG methods.DA and test-time DG are able to adjust",
"Dynamic Component",
"Forward Propagation",
"Backward Propagation",
"the model when encountering novel target domains via fine-",
"tuning, while DG freezes a static model to adapt any agnostic",
"Figure 2: Pipeline of the proposed DDG method. In the",
"target domain. Different from the existing methods, DDG is",
"training phase, two stages are required, namely novel sam-",
"able to adjust the model without any training effort for better",
"ples simulation and learning to adjust, which aims to teach",
"domain generalization.",
"the meta-adjuster to adapt novel samples. In the testing",
"phase, the network can be adjusted without training when",
"2",
"Related Work",
"encountering agnostictarget domain.",
"Domain Adaptation.",
"常规DA旨在将从标记的源域",
"中学到的知识转移到可能包含一些标记样本或根本不包",
"在这种方法中，计算成本是不可避免的。。相反，我们的",
"含的目标域。基于差异的方法是早期开发的，它通过最",
"方法可以直接更新参数以适应不可知的目标域，而无需",
"任何反向传播",
"小化定义明确的距离损失[Haeusseretal，2017]来减轻",
"两个域之间的分布差异。DA方法的另一行基于对抗性",
"学习，其中经常使用域鉴别器与特征提取器进行对抗性",
"3",
"Methods",
"训练，以便通过混滑域鉴别器[Ganinetal.2016]来强",
"3.1",
"Overview",
"制特征提取器学习域不变特征。。与调整模型以明确适应",
"DG旨在训练一个通用模型，该模型可以很好地推广",
"目标域的DA方法相比，我们的方法可以被视为一种通",
"过仅在源域中进行预训练的无训练DA",
"到任何不可知的目标域。特别是，我们在本文中只考",
"虑K-way分类任务。对于香草DG任务，我们从多",
"Domain Generalization.",
"早期的DG相关研究集中",
"源域Xs=(asi,Xs2,Xsm)中提取ns标记的样本",
"于学习域不变表示，如果它可以抑制多个源域之间的域",
"(a,y,d'ni，其中aeXs，y.eys和域标签d,e",
"偏移，则被认为是健壮的E[Motianet al.，2017]。另",
"[1,M]。BG的目标是学习映射f：（Xst,sm]",
"→",
"种DG方法是基于数据增强，增强图像风格的多样",
"s，该映射也可以在任何不可知的目标域上进行精确预",
"性，以优化模型的泛化能力[Zhou etal.，2020a；Zhou",
"测",
"etal.，2021]。自监督学习也被应用于学习任务无关表",
"Vanilla Domain Generalization.",
"对于香草DG，在训",
"练阶段获得的学习映射fe将在推理中保持静态。由于",
"示的DG，其过拟合概率较小[Carluccietal.，2019",
"Kim et al.,2021] 。元学习还涉及解决DGE[Balaji et",
"al.20181，其中通过模拟源域中的元训练和元测试任务",
"源域和目标域之间不可避免的数据分布差异，它通常会",
"导致不可知目标域的性能下降",
"来训练深度模型，以增强其泛化能力。我们的方法可以",
"看作是一个非常简单的元学习范式，其中元调整器学习",
"Dynamic Domain Generalization.",
"为了克服香草DG",
"为每个单独的“域”学习参数泛化。。此外，一些新兴的",
"的弱点，我们提出了DG的新变体，即动态域泛化",
"作品为DG展示了一个新的见解，即深度模型可以在测",
"(DDG）。DDG主要侧重于学习扭曲网络参数以适应",
"试时适应目标域[IwasawaandMatsuo，2021]。然而，",
"来自不同新域的实例。在本文中，每个实例都被假设为"
] |
{
"arxiv_id": "2205.13913",
"language": "zh",
"page_idx": 2
}
|
||
2205.13913-zh-3
|
2205.13913-zh-3.jpg
|
<FigureHere>
Figure 3: Structure illustration of the basic block for DDG, $\bar{\mathrm{which~contains~a~static~and~a~dynamic~component.~The~latter}}$ one is a learnable linear combination of four asymmetric kernel templates and corresponding dynamic coefficients which are generated by the meta-adjuster. For simplicity, $C_out$ is omitted.
一个独立的潜在域。在此假设下，可以将模型参数$\Theta$建模为输入实例$x:\Theta=\Theta(x)$的函数。为此，设计了一个元调整器来生成网络参数，以实现多样化的网络，从而适应多样化的新颖样本。在这种情况下，映射$f_\Theta(x)$更有可能在实例感知参数$\Theta(x)$的反射下将不同的实例对齐到相同的特征空间中。与静态映射难以对齐来自不同数据分布的实例相比，DDG 更容易实现不同新域之间的特征对齐
我们的方法的管道在图 2 中进行了说明，其中 DDG 框架由静态组件和由元调整器根据每个输人实例更新的动态组件组成。前者负责学习域共享功能，而后者旨在学习特定于域的功能。为了教导元调整器适应不同的新领域，我们通过有限源域之间的随机凸插值来模拟许多新的训练样本
3.2 Novel Samples Simulation
需要大量新颖的样本来驱动学习的优化以进行调整。通过利用有限的源域，我们从 MixUp [Zhang et al.,2018] 开发了一种 DomainMix 方法来合成具有不同风格的新域。给定从两个不同源域随机采样的两个实例$(x_s^i,y_s^i)$ 和$( x_{\mathrm{s} }^{j}, y_{\mathrm{s} }^{j})$ ,将新颖样本模拟的过程表述为：
$$\bar{x}_s^i=\alpha x_s^i+(1-\alpha)x_s^j,\quad\bar{y}_s^i=\alpha y_s^i+(1-\alpha)y_s^j,$$
,其中 $\alpha$ 是从每个前馈中的 Beta 分布中随机采样的。每个 DomainMix 样本$(\bar{x}_s^i,\bar{y}_s^i)$都可以被视为一个唯一的域，并被用来驱动以下学习调整过程
3.3 Learn to Adjust
在本节中，我们将网络参数解耦为静态部分和动态部分。它们隐式地分离了域共享和域特定的功能。这里的动态参数是由一个称为元调整器的浅生成器生成的，它将实例嵌入作为输人。为了更好地说明，我们采用最流行的卷积神经网络之一，即 ResNet$\underline{\mathrm{E}}[$He et al.,2016],作为动态网络的骨干。如图 3 所示，元调整器嵌人在残差块中的每个中间卷积层中(请注意，每个残差块由$1\times1$, $3\times3$和$1\times1$卷积层的顺序组成)。在下面，我们将介绍如何有效地设计元调整器
Meta-Adjuster. 元调整器采用实例嵌人$F_in(x)$ (i。e。最后一个块的输出)作为输入，因此元调整器实际上可以被视为输入实例$\boldsymbol{\Delta\Theta}(x)$的函数。因此，网络参数调整过程可以表示为：
$$\mathbf{\Theta}(x)=\mathbf{\Theta}_{s}+\mathbf{\Delta}\mathbf{\Theta}(x),$$
(2), (2)
其中$\Theta_s$和$\Theta(x)$表示静态参数及其相应的最新参数$\boldsymbol{\Delta\Theta}(x)$表示其动态项。从理论上讲，动态项$\boldsymbol{\Delta\Theta}(x)$应该与静态参数$\Theta_\mathrm{s}$具有相同的维度。然而，它使网络参数的数量呈二次增长，然后明显增加了过拟合的风险。因此，根据矩阵分解，我们建议将动态项分解为几个内核模板$\{V_n\}_{n=1}^N$的可学习线性组合，可以表示为：
$$\boldsymbol{\Delta\Theta}(x)=\sum_{n=1}^N\lambda_n(x)\boldsymbol{V}_n,$$
(3)
其中$N$是表示内核模板数量的超参数。$\{\lambda_n(x)\}_n=1^N$表示动态系数的集合，它是由可学习的浅层元调整器$\lambda(\cdot)$ 通过将实例$x$ (实际上是嵌人$F_in(x)$的实例)作为输入而生成的一组实例感知标量。这样，动态组件被限制在有限的空间内，这有利于整个模型的优化
Asymmetric Kernel Templates. 将元调整器分解为内核模板和少数动态系数的线性组合的一个明显副作用是由动态参数带来的调制很容易变得微小和僵化。为了避免这种趋势，我们设计了四个非对称内核模板来探索不同维度的动态参数之间的关系，并强制元调整器为各种新颖的样本生成更多不同的动态参数。另一个发现是，卷积内核的学习知识并不是均匀分布的，因为内核的骨架(i。e。中央十字交叉位置)通常具有较大的权重值[Ding et al., 2019]。因此，我们建议通过生成基于骨架的动态参数来增强内核的骨架，该参数是从这些基于非对称骨架的内核模板的线性组合中得出的
假设静态内核$\Theta_s$的维度表示为$c_in\times c_{out}\times k\times k$, 其中$k$ , $c_{in}$和$c_out$表示内核大小，输入通道号以及输出通道号。如图 3 所示，基于非对称骨架的内核模板被设计为四个不同形状的矩阵，分别为$c_{in}\times1\times k\times k$ , $c_{in}\times c_{out}\times 1\times 1$ , $c_{in}\times c_{out}\times k\times 1$ 和 $c_{in}\times c_{out}\times1\times k$ 。请注意，大小为$c_in\times1\times k\times k$的内核模板是作为深度卷积实现的，因为它只关注空间关系。通过这种方式， 可以鼓励针对不同的输入样本改变动态参数。此外，网络参数的骨架可以在空间和通道维度上放大
## 4 Experiments
4.1 Datasets and Experimental Setup
? ? ? ) PACS$\underline{\mathrm{E}}[$Li et al.,2017] 包含超过 9$k$的图像，从四个领域收集了七个类别，其中包括：照片，素描，卡通和艺术画。2) Office-Home$\underline{\mathrm{E}}[$Venkateswara et al.,2017] 包含 15,500 图像与 65 类别，每个类别几个图像。所有图像均从四个领域采样，包括：艺术，剪贴画，产品和现实世界。3)DomainNet [Peng et al., 2019] 是一个包含 345 类别的 600$k$图像的大型基准测试。这些图像是$\underline{\text{从六个不同的领域收集的，分别是 Clipart, Inforgraph}}$ Painting, Quickdraw, Real 和 Sketch
Experimental Setup. 对于每个数据集，我们执行leave-one-out 策略，该策略随机选择一个域进行评估， 并将其余域作为源域进行训练。每个实验用不同的随机种子重复三次，这些结果的均值和方差在本文中报道。
|
[
[
0.2320556640625,
0.07257080078125,
0.35693359375,
0.07257080078125,
0.35693359375,
0.0802001953125,
0.2320556640625,
0.0802001953125
],
[
0.5185546875,
0.0706787109375,
0.90869140625,
0.0706787109375,
0.90869140625,
0.083984375,
0.5185546875,
0.083984375
],
[
0.1519775390625,
0.08209228515625,
0.2017822265625,
0.08209228515625,
0.2017822265625,
0.0889892578125,
0.1519775390625,
0.0889892578125
],
[
0.2181396484375,
0.07452392578125,
0.233642578125,
0.07452392578125,
0.233642578125,
0.07891845703125,
0.2181396484375,
0.07891845703125
],
[
0.1519775390625,
0.0889892578125,
0.2034912109375,
0.0889892578125,
0.2034912109375,
0.10101318359375,
0.1519775390625,
0.10101318359375
],
[
0.2418212890625,
0.0889892578125,
0.289306640625,
0.0889892578125,
0.289306640625,
0.0997314453125,
0.2418212890625,
0.0997314453125
],
[
0.31787109375,
0.0889892578125,
0.366943359375,
0.0889892578125,
0.366943359375,
0.0997314453125,
0.31787109375,
0.0997314453125
],
[
0.400390625,
0.0889892578125,
0.449462890625,
0.0889892578125,
0.449462890625,
0.10101318359375,
0.400390625,
0.10101318359375
],
[
0.5146484375,
0.083984375,
0.90869140625,
0.083984375,
0.90869140625,
0.09783935546875,
0.5146484375,
0.09783935546875
],
[
0.515625,
0.09783935546875,
0.91259765625,
0.09783935546875,
0.91259765625,
0.111083984375,
0.515625,
0.111083984375
],
[
0.12091064453125,
0.10479736328125,
0.16015625,
0.10479736328125,
0.16015625,
0.11614990234375,
0.12091064453125,
0.11614990234375
],
[
0.1993408203125,
0.10479736328125,
0.328369140625,
0.10479736328125,
0.328369140625,
0.11553955078125,
0.1993408203125,
0.11553955078125
],
[
0.351318359375,
0.10479736328125,
0.410888671875,
0.10479736328125,
0.410888671875,
0.11553955078125,
0.351318359375,
0.11553955078125
],
[
0.4150390625,
0.1060791015625,
0.43701171875,
0.1060791015625,
0.43701171875,
0.1142578125,
0.4150390625,
0.1142578125
],
[
0.5146484375,
0.11236572265625,
0.654296875,
0.11236572265625,
0.654296875,
0.1256103515625,
0.5146484375,
0.1256103515625
],
[
0.25732421875,
0.12249755859375,
0.26953125,
0.12249755859375,
0.26953125,
0.1319580078125,
0.25732421875,
0.1319580078125
],
[
0.334228515625,
0.12249755859375,
0.346435546875,
0.12249755859375,
0.346435546875,
0.1319580078125,
0.334228515625,
0.1319580078125
],
[
0.6337890625,
0.1337890625,
0.7900390625,
0.1337890625,
0.7900390625,
0.147705078125,
0.6337890625,
0.147705078125
],
[
0.1690673828125,
0.12371826171875,
0.1846923828125,
0.12371826171875,
0.1846923828125,
0.1331787109375,
0.1690673828125,
0.1331787109375
],
[
0.890625,
0.1351318359375,
0.91259765625,
0.1351318359375,
0.91259765625,
0.147705078125,
0.890625,
0.147705078125
],
[
0.1282958984375,
0.154052734375,
0.2059326171875,
0.154052734375,
0.2059326171875,
0.1641845703125,
0.1282958984375,
0.1641845703125
],
[
0.310546875,
0.1546630859375,
0.38232421875,
0.15283203125,
0.38232421875,
0.1634521484375,
0.310546875,
0.1654052734375
],
[
0.5146484375,
0.1572265625,
0.9111328125,
0.1572265625,
0.9111328125,
0.17041015625,
0.5146484375,
0.17041015625
],
[
0.371826171875,
0.173583984375,
0.3857421875,
0.173583984375,
0.3857421875,
0.1817626953125,
0.371826171875,
0.1817626953125
],
[
0.515625,
0.17041015625,
0.91259765625,
0.17041015625,
0.91259765625,
0.184326171875,
0.515625,
0.184326171875
],
[
0.393798828125,
0.186279296875,
0.44775390625,
0.186279296875,
0.44775390625,
0.1962890625,
0.393798828125,
0.1962890625
],
[
0.515625,
0.184326171875,
0.9111328125,
0.184326171875,
0.9111328125,
0.1976318359375,
0.515625,
0.1976318359375
],
[
0.44775390625,
0.1932373046875,
0.476318359375,
0.1932373046875,
0.476318359375,
0.2052001953125,
0.44775390625,
0.2052001953125
],
[
0.515625,
0.1976318359375,
0.9111328125,
0.1976318359375,
0.9111328125,
0.2108154296875,
0.515625,
0.2108154296875
],
[
0.2646484375,
0.1900634765625,
0.28759765625,
0.1900634765625,
0.28759765625,
0.1944580078125,
0.2646484375,
0.1944580078125
],
[
0.324462890625,
0.2020263671875,
0.3798828125,
0.2020263671875,
0.3798828125,
0.21337890625,
0.324462890625,
0.21337890625
],
[
0.384033203125,
0.2020263671875,
0.396240234375,
0.2020263671875,
0.396240234375,
0.210205078125,
0.384033203125,
0.210205078125
],
[
0.515625,
0.212158203125,
0.9111328125,
0.212158203125,
0.9111328125,
0.2259521484375,
0.515625,
0.2259521484375
],
[
0.36767578125,
0.2298583984375,
0.3798828125,
0.2298583984375,
0.3798828125,
0.23291015625,
0.36767578125,
0.23291015625
],
[
0.5146484375,
0.2259521484375,
0.85693359375,
0.2259521484375,
0.85693359375,
0.240478515625,
0.5146484375,
0.240478515625
],
[
0.09722900390625,
0.2374267578125,
0.10540771484375,
0.2374267578125,
0.10540771484375,
0.242431640625,
0.09722900390625,
0.242431640625
],
[
0.7060546875,
0.248779296875,
0.72119140625,
0.248779296875,
0.72119140625,
0.2587890625,
0.7060546875,
0.2587890625
],
[
0.08660888671875,
0.261474609375,
0.483642578125,
0.261474609375,
0.483642578125,
0.274658203125,
0.08660888671875,
0.274658203125
],
[
0.6259765625,
0.25830078125,
0.80126953125,
0.25830078125,
0.80126953125,
0.27978515625,
0.6259765625,
0.27978515625
],
[
0.8896484375,
0.261962890625,
0.91259765625,
0.261962890625,
0.91259765625,
0.276611328125,
0.8896484375,
0.276611328125
],
[
0.0882568359375,
0.27587890625,
0.483642578125,
0.27587890625,
0.483642578125,
0.285888671875,
0.0882568359375,
0.285888671875
],
[
0.701171875,
0.27783203125,
0.72607421875,
0.27783203125,
0.72607421875,
0.2890625,
0.701171875,
0.2890625
],
[
0.08660888671875,
0.285888671875,
0.483642578125,
0.287353515625,
0.483642578125,
0.300537109375,
0.08660888671875,
0.2998046875
],
[
0.08660888671875,
0.2998046875,
0.486083984375,
0.2998046875,
0.486083984375,
0.313232421875,
0.08660888671875,
0.313232421875
],
[
0.51318359375,
0.29736328125,
0.91259765625,
0.298583984375,
0.91259765625,
0.313232421875,
0.51318359375,
0.3125
],
[
0.0882568359375,
0.314453125,
0.48779296875,
0.314453125,
0.48779296875,
0.324462890625,
0.0882568359375,
0.324462890625
],
[
0.51318359375,
0.313232421875,
0.9111328125,
0.313232421875,
0.9111328125,
0.326904296875,
0.51318359375,
0.326904296875
],
[
0.08660888671875,
0.326904296875,
0.1422119140625,
0.326904296875,
0.1422119140625,
0.337158203125,
0.08660888671875,
0.337158203125
],
[
0.51318359375,
0.32568359375,
0.9111328125,
0.326904296875,
0.9111328125,
0.34033203125,
0.51318359375,
0.339111328125
],
[
0.515625,
0.34033203125,
0.91259765625,
0.34033203125,
0.91259765625,
0.354248046875,
0.515625,
0.354248046875
],
[
0.515625,
0.35595703125,
0.80615234375,
0.35595703125,
0.80615234375,
0.3662109375,
0.515625,
0.3662109375
],
[
0.1004638671875,
0.363037109375,
0.484375,
0.363037109375,
0.484375,
0.373779296875,
0.1004638671875,
0.373779296875
],
[
0.08660888671875,
0.375,
0.486083984375,
0.373779296875,
0.486083984375,
0.38818359375,
0.08660888671875,
0.389404296875
],
[
0.51953125,
0.37255859375,
0.7509765625,
0.37255859375,
0.7509765625,
0.386474609375,
0.51953125,
0.386474609375
],
[
0.75244140625,
0.375,
0.90869140625,
0.375,
0.90869140625,
0.385009765625,
0.75244140625,
0.385009765625
],
[
0.08905029296875,
0.389404296875,
0.486083984375,
0.389404296875,
0.486083984375,
0.40283203125,
0.08905029296875,
0.40283203125
],
[
0.515625,
0.38818359375,
0.90673828125,
0.38818359375,
0.90673828125,
0.3984375,
0.515625,
0.3984375
],
[
0.0882568359375,
0.40283203125,
0.486083984375,
0.40283203125,
0.486083984375,
0.417236328125,
0.0882568359375,
0.417236328125
],
[
0.515625,
0.40087890625,
0.9111328125,
0.40087890625,
0.9111328125,
0.4140625,
0.515625,
0.4140625
],
[
0.0906982421875,
0.419189453125,
0.48193359375,
0.419189453125,
0.48193359375,
0.429931640625,
0.0906982421875,
0.429931640625
],
[
0.5146484375,
0.4140625,
0.9111328125,
0.4140625,
0.9111328125,
0.427978515625,
0.5146484375,
0.427978515625
],
[
0.08905029296875,
0.431884765625,
0.483642578125,
0.431884765625,
0.483642578125,
0.44580078125,
0.08905029296875,
0.44580078125
],
[
0.515625,
0.4287109375,
0.9111328125,
0.4287109375,
0.9111328125,
0.442626953125,
0.515625,
0.442626953125
],
[
0.0882568359375,
0.4462890625,
0.484375,
0.4462890625,
0.484375,
0.460205078125,
0.0882568359375,
0.460205078125
],
[
0.515625,
0.443115234375,
0.9111328125,
0.443115234375,
0.9111328125,
0.453857421875,
0.515625,
0.453857421875
],
[
0.08905029296875,
0.4609375,
0.1519775390625,
0.4609375,
0.1519775390625,
0.472900390625,
0.08905029296875,
0.472900390625
],
[
0.5146484375,
0.455810546875,
0.9111328125,
0.455810546875,
0.9111328125,
0.4697265625,
0.5146484375,
0.4697265625
],
[
0.10296630859375,
0.473388671875,
0.486083984375,
0.472900390625,
0.486083984375,
0.486083984375,
0.10296630859375,
0.4873046875
],
[
0.51318359375,
0.47021484375,
0.91259765625,
0.47021484375,
0.91259765625,
0.484130859375,
0.51318359375,
0.484130859375
],
[
0.51123046875,
0.481689453125,
0.9111328125,
0.48291015625,
0.9111328125,
0.49755859375,
0.51123046875,
0.496826171875
],
[
0.08660888671875,
0.488037109375,
0.484375,
0.488037109375,
0.484375,
0.501953125,
0.08660888671875,
0.501953125
],
[
0.0882568359375,
0.501953125,
0.484375,
0.501953125,
0.484375,
0.51513671875,
0.0882568359375,
0.51513671875
],
[
0.515625,
0.49755859375,
0.9111328125,
0.49755859375,
0.9111328125,
0.51123046875,
0.515625,
0.51123046875
],
[
0.0882568359375,
0.51513671875,
0.486083984375,
0.51513671875,
0.486083984375,
0.52880859375,
0.0882568359375,
0.52880859375
],
[
0.51318359375,
0.51220703125,
0.7587890625,
0.51220703125,
0.7587890625,
0.52587890625,
0.51318359375,
0.52587890625
],
[
0.0882568359375,
0.52783203125,
0.484375,
0.52880859375,
0.484375,
0.54345703125,
0.0882568359375,
0.54248046875
],
[
0.529296875,
0.52587890625,
0.9111328125,
0.52587890625,
0.9111328125,
0.5390625,
0.529296875,
0.5390625
],
[
0.51318359375,
0.5361328125,
0.9111328125,
0.537109375,
0.9111328125,
0.5546875,
0.51318359375,
0.5537109375
],
[
0.08905029296875,
0.54345703125,
0.1846923828125,
0.54345703125,
0.1846923828125,
0.5546875,
0.08905029296875,
0.5546875
],
[
0.515625,
0.55322265625,
0.9111328125,
0.55322265625,
0.9111328125,
0.56640625,
0.515625,
0.56640625
],
[
0.0849609375,
0.5693359375,
0.336669921875,
0.5693359375,
0.336669921875,
0.58349609375,
0.0849609375,
0.58349609375
],
[
0.5146484375,
0.56494140625,
0.90283203125,
0.56640625,
0.90283203125,
0.58056640625,
0.5146484375,
0.580078125
],
[
0.5146484375,
0.58056640625,
0.81884765625,
0.58056640625,
0.81884765625,
0.59521484375,
0.5146484375,
0.59521484375
],
[
0.82421875,
0.5869140625,
0.85546875,
0.5869140625,
0.85546875,
0.59130859375,
0.82421875,
0.59130859375
],
[
0.85302734375,
0.58203125,
0.9111328125,
0.58203125,
0.9111328125,
0.59228515625,
0.85302734375,
0.59228515625
],
[
0.08905029296875,
0.58837890625,
0.484375,
0.58837890625,
0.484375,
0.6015625,
0.08905029296875,
0.6015625
],
[
0.515625,
0.5947265625,
0.9111328125,
0.5947265625,
0.9111328125,
0.60791015625,
0.515625,
0.60791015625
],
[
0.08905029296875,
0.6015625,
0.483642578125,
0.6015625,
0.483642578125,
0.61572265625,
0.08905029296875,
0.61572265625
],
[
0.51318359375,
0.60693359375,
0.91015625,
0.609375,
0.91015625,
0.62255859375,
0.51318359375,
0.62060546875
],
[
0.0882568359375,
0.6162109375,
0.484375,
0.6162109375,
0.484375,
0.6298828125,
0.0882568359375,
0.6298828125
],
[
0.51318359375,
0.62060546875,
0.9111328125,
0.6220703125,
0.9111328125,
0.63623046875,
0.51318359375,
0.63525390625
],
[
0.08660888671875,
0.62890625,
0.484375,
0.6298828125,
0.484375,
0.64453125,
0.08660888671875,
0.64306640625
],
[
0.5146484375,
0.63623046875,
0.82421875,
0.63623046875,
0.82421875,
0.6494140625,
0.5146484375,
0.6494140625
],
[
0.0882568359375,
0.64306640625,
0.400390625,
0.64306640625,
0.400390625,
0.65771484375,
0.0882568359375,
0.65771484375
],
[
0.10296630859375,
0.66845703125,
0.484375,
0.66845703125,
0.484375,
0.681640625,
0.10296630859375,
0.681640625
],
[
0.51123046875,
0.66650390625,
0.65673828125,
0.66650390625,
0.65673828125,
0.6796875,
0.51123046875,
0.6796875
],
[
0.51123046875,
0.6845703125,
0.82421875,
0.68603515625,
0.82421875,
0.70068359375,
0.51123046875,
0.69873046875
],
[
0.098876953125,
0.6923828125,
0.486083984375,
0.6923828125,
0.486083984375,
0.70556640625,
0.098876953125,
0.70556640625
],
[
0.0849609375,
0.7041015625,
0.484375,
0.705078125,
0.484375,
0.7197265625,
0.0849609375,
0.71826171875
],
[
0.51318359375,
0.7041015625,
0.9111328125,
0.7041015625,
0.9111328125,
0.71728515625,
0.51318359375,
0.71728515625
],
[
0.08660888671875,
0.7197265625,
0.3505859375,
0.7197265625,
0.3505859375,
0.73291015625,
0.08660888671875,
0.73291015625
],
[
0.515625,
0.71630859375,
0.90869140625,
0.71728515625,
0.90869140625,
0.73193359375,
0.515625,
0.73095703125
],
[
0.515625,
0.73193359375,
0.91015625,
0.73193359375,
0.91015625,
0.74560546875,
0.515625,
0.74560546875
],
[
0.0849609375,
0.74560546875,
0.259033203125,
0.74560546875,
0.259033203125,
0.7587890625,
0.0849609375,
0.7587890625
],
[
0.51708984375,
0.74755859375,
0.90869140625,
0.74755859375,
0.90869140625,
0.75830078125,
0.51708984375,
0.75830078125
],
[
0.0882568359375,
0.76513671875,
0.483642578125,
0.76513671875,
0.483642578125,
0.77880859375,
0.0882568359375,
0.77880859375
],
[
0.515625,
0.76025390625,
0.9111328125,
0.76025390625,
0.9111328125,
0.7734375,
0.515625,
0.7734375
],
[
0.51318359375,
0.771484375,
0.9111328125,
0.77294921875,
0.9111328125,
0.787109375,
0.51318359375,
0.7861328125
],
[
0.08905029296875,
0.77880859375,
0.486083984375,
0.77880859375,
0.486083984375,
0.79248046875,
0.08905029296875,
0.79248046875
],
[
0.5146484375,
0.7861328125,
0.9111328125,
0.7861328125,
0.9111328125,
0.80029296875,
0.5146484375,
0.80029296875
],
[
0.08905029296875,
0.79248046875,
0.486083984375,
0.79248046875,
0.486083984375,
0.80615234375,
0.08905029296875,
0.80615234375
],
[
0.51123046875,
0.7998046875,
0.91259765625,
0.80029296875,
0.91259765625,
0.81494140625,
0.51123046875,
0.814453125
],
[
0.0882568359375,
0.80615234375,
0.484375,
0.80615234375,
0.484375,
0.8193359375,
0.0882568359375,
0.8193359375
],
[
0.08905029296875,
0.8193359375,
0.484375,
0.8193359375,
0.484375,
0.83349609375,
0.08905029296875,
0.83349609375
],
[
0.51318359375,
0.81494140625,
0.7900390625,
0.81494140625,
0.7900390625,
0.8291015625,
0.51318359375,
0.8291015625
],
[
0.0882568359375,
0.833984375,
0.486083984375,
0.833984375,
0.486083984375,
0.84765625,
0.0882568359375,
0.84765625
],
[
0.521484375,
0.833984375,
0.9111328125,
0.833984375,
0.9111328125,
0.84765625,
0.521484375,
0.84765625
],
[
0.0882568359375,
0.84765625,
0.4794921875,
0.84765625,
0.4794921875,
0.861328125,
0.0882568359375,
0.861328125
],
[
0.5146484375,
0.84765625,
0.91015625,
0.84765625,
0.91015625,
0.861328125,
0.5146484375,
0.861328125
],
[
0.08660888671875,
0.8623046875,
0.484375,
0.8623046875,
0.484375,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.5146484375,
0.85986328125,
0.9111328125,
0.861328125,
0.9111328125,
0.87548828125,
0.5146484375,
0.875
],
[
0.08660888671875,
0.87548828125,
0.28515625,
0.87548828125,
0.28515625,
0.8896484375,
0.08660888671875,
0.8896484375
],
[
0.515625,
0.87548828125,
0.896484375,
0.87548828125,
0.896484375,
0.8896484375,
0.515625,
0.8896484375
]
] |
[
"mmetricKernelTemplates",
"Meta-Adjuster.元调整器采用实例嵌人Fin（c）（i。e。",
"(Depth Wise",
"ASVI",
"Caxkxk",
"Cm×1x1",
"CmXkx1",
"Cmxixk",
"最后一个块的输出）作为输人，因此元调整器实际上可",
"以被视为输人实例△①(）的函数。因此，网络参数调",
"(a) ×",
"()+()x",
"+入(a)×",
"",
"整过程可以表示为：",
"V",
"V",
"0(a) =0 +0(r)",
"Vi",
"(2)",
"Meta-Adjuster",
"Residual Block",
"其中①。和(α）表示静态参数及其相应的最新参数",
"1e",
"A(α）表示其动态项。从理论上讲，动态项△①(r）应",
"CX1x1",
"该与静态参数?s具有相同的维度。然而，它使网络参数",
"Fout",
"的数量呈二次增长，然后明显增加了过拟合的风险。因",
"v1",
"CnXkxk",
"?",
"此，根据矩阵分解，我们建议将动态项分解为几个内核",
"",
"模板【Vn)n=1的可学习线性组合，可以表示为",
"2",
"N",
"Figure 3: Structure illustration of the basic block for DDG",
"40(a) =an(a)V n;",
"(3)",
"which containsastatic and a dynamic component. Thelatter",
"n=1",
"one is a learnable linear combination of four asymmetric ker-",
"nel templates and corresponding dynamic coefficients which",
"其中N是表示内核模板数量的超参数。(入n(a))n=1表",
"are generated by the meta-adjuster.For simplicity,Cout is",
"示动态系数的集合，它是由可学习的浅层元调整器入()）",
"omitted.",
"通过将实例（实际上是嵌入Fin（α）的实例）作为输人",
"而生成的一组实例感知标量。这样，动态组件被限制在",
"有限的空间内，这有利于整个模型的优化",
"个独立的潜在域。在此假设下，可以将模型参数④建",
"模为输人实例a：=①()的函数。为此，设计了一个",
"Asymmetric Kernel Templates.",
"将元调整器分解为内",
"元调整器来生成网络参数，以实现多样化的网络，从而",
"核模板和少数动态系数的线性组合的一个明显副作用是",
"适应多样化的新颖样本。在这种情况下，映射fe(）更",
"由动态参数带来的调制很容易变得微小和化。为了避",
"有可能在实例感知参数?（）的反射下将不同的实例对",
"免这种趋势，我们设计了四个非对称内核模板来探索不",
"齐到相同的特征空间中。与静态映射难以对齐来自不同",
"同维度的动态参数之间的关系，并强制元调整器为各种",
"数据分布的实例相比，DDG更容易实现不同新域之间的",
"新颖的样本生成更多不同的动态参数。另一个发现是，卷",
"特征对齐",
"积内核的学习知识并不是均匀分布的，因为内核的骨架",
"我们的方法的管道在图2中进行了说明，其中DDG",
"(i。e。中央十字交叉位置)通常具有较大的权重值[Ding",
"etal.,2019。因此，我们建议通过生成基于骨架的动态",
"框架由静态组件和由元调整器根据每个输人实例更新的",
"动态组件组成。前者负责学习域共享功能，而后者旨在",
"参数来增强内核的骨架，该参数是从这些基于非对称骨",
"学习特定于域的功能。为了教导元调整器适应不同的新",
"架的内核模板的线性组合中得出的",
"领域，我们通过有限源域之间的随机凸插值来模拟许多",
"假设静态内核O。的维度表示为Cin×Cout×k×k",
"其中k，Cin和cout表示内核大小，输人通道号以及输",
"新的训练样本",
"出通道号。如图3所示，基于非对称骨架的内核模板被",
"3.2Novel Samples Simulation",
"设计为四个不同形状的矩阵，分别为Cin×1×k×k",
"Cin × Cout ×1×1，Cin × Cout × k ×1和 Cin ",
"Cna",
"tx1xk",
"需要大量新颖的样本来驱动学习的优化以进行调整。通",
"。请注意，大小为cCin×1×k×k的内核模板是作为深",
"过利用有限的源域，我们从MixUp[Zhangetal.,2018]",
"度卷积实现的，因为它只关注空间关系。通过这种方式",
"开发了一种DomainMix方法来合成具有不同风格的新",
"可以鼓励针对不同的输人样本改变动态参数。此外，网",
"域。给定从两个不同源域随机采样的两个实例（cs，y）",
"络参数的骨架可以在空间和通道维度上放大",
"和（r,y），将新颖样本模拟的过程表述为：",
"=ar+ (1-α)rs, =αy,+(-a)ys,(1)",
"4Experiments",
"4.1Datasets and Experimental Setup",
"其中α是从每个前馈中的Beta分布中随机采样的。每",
"个DomainMix样本（，y）都可以被视为一个唯一的",
"???）PACSE[Lietal.，2017]包含超过9k的图像，从四",
"域，并被用来驱动以下学习调整过程",
"个领域收集了七个类别，其中包括：照片，素描，卡通",
"和艺术画。2)Office-HomeE[Venkateswaraetal.,2017]",
"3.3Learn to Adjust",
"包含15.500图像与65类别，每个类别几个图像。所有",
"在本节中，我们将网络参数解耦为静态部分和动态部分",
"图像均从四个领域采样，包括：艺术，剪贴画，产品和",
"现实世界。3)DomainNet[Pengetal.，2019]是一个包",
"它们隐式地分离了域共享和域特定的功能。这里的动态",
"含345类别的600k图像的大型基准测试。这些图像是",
"参数是由一个称为元调整器的浅生成器生成的，它将实",
"从六个不同的领域收集的，分别是Clipart，Inforgraph，",
"例嵌人作为输人。为了更好地说明，我们采用最流行的",
"卷积神经网络之一，即ResNetE[Heetal.,2016]，作为",
"Painting,Quickdraw,Real 和 Sketch",
"动态网络的骨干。如图3所示，元调整器嵌入在残差块",
"ExperimentalSetup.对于每个数据集，我们执行",
"中的每个中间卷积层中（请注意，每个残差块由1×1",
"leave-one-out策略，该策略随机选择一个域进行评估",
"3×3和1×1卷积层的顺序组成）。在下面，我们将介",
"并将其余域作为源域进行训练。每个实验用不同的随机",
"绍如何有效地设计元调整器",
"种子重复三次，这些结果的均值和方差在本文中报道。"
] |
{
"arxiv_id": "2205.13913",
"language": "zh",
"page_idx": 3
}
|
||
2205.13913-zh-4
|
2205.13913-zh-4.jpg
|
4.2 Implementation Details
所有实验均在带有 PyTorch1 的 RTX3090GPU 上进行。10。0。我们使用 ResNet-50 作为动态网络的骨干，通过将元调整器附加到原始残差块(如图 3 所示),其中内核模板的空间大小设置为$k=3$ 。请注意，在每个块中元调整器由全局平均池化层，夹在两个全连接层之间的ReLU 层和 SoftMax 层的顺序实现，以输出实例感知系数
在训练 DG 任务之前，我们在 ImageNet 上预训练我们的网络，以便与现有作品进行公平的比较。然后，我们通过使用 SGD 微调预训练模型来训练 DG 任务。训练目标是交叉熵损失函数，与香草 DG 完全相同。在训练过程中，非对称内核模板和元调整器与主干网同时优化。对于 PACS 和 Office-Home, 网络优化设置为批处理大小为 64, 训练时期为 50, 初始学习速率为 1e-3, 由余弦调度器衰减。在 DomainNet 上训练时，大多数超参数与 PACS 相同，只是初始学习率和最大 epoch 为 2e-3和 15 , 并且使用随机域采样器策略$\underline{\mathrm{E}}[$Zhou et al., 2021] 获取小批量，以确保每个域都被均匀采样
4.3 Comparisons with the State-of-the-Art
我们将我们的方法与 PACS, Office-Home 和 DomainNet 数据集上的最新方法进行了比较，其中结果分别在表 1 , 表 2 和表 3 中报告。从这些表中，我们可以观察到：
1) Our method achieves an average accuracy of 86.21
72.13 and 46.89 percent on PACS, Office-Home and DomainNet benchmark datasets, respectively. Compared to other DG methods, our method achieves the state-of-the-art results on most of DG benchmarks, which shows the superiority of our DDG.
2) Our method is flexible to combine with other meth-
ods to further improve the performance. Taking Mixstyle [ Zhou et al. , 2021] as an example, the per- formance of our model can be further improved if we insert MixStyle layers into our model. The results can reach 87.87,72.31 and 46.93 percent on PACS, Office-Home and DomainNet benchmarks, respectively. It shows that our method is rather robust even on the very strong baselines, i.e. MixStyle.
## 4.4 Ablation Studies
Effectiveness of Asymmetric Kernel Templates. 我们通过使用由四个空间尺寸为$1\times1$或$3\times3$的相同内核组
我们对 PACS 数据集进行消融研究，以分别评估非对称内核模板和 DomainMix 的有效性，其中结果如表 4 所示。
成的通用内核模板，研究了内核模板非对称设计的有效性。此外，我们还比较了一种渠道关注方法，即 SENet$\boxed{\mathrm{F}}$ [Hu et al., 2018] , 它也是 CNN 社区中流行且功能强大的体系结构。如表 4 所示 , 我们可以看到：
$1){\text{Our method achieves superior performance to the}}$
vanilla ResNet and SENet by a large margin, which suggests the effectiveness of meta-adjusters on DG $tasks.$
2) The result of asymmetric kernel templates is obvi-
ously better than that of four itentical $1\times1$ kernel templates and $3\times3$ kernel templates, which confirms
<FigureHere>
Figure 4: Comparison among static kernel of vanilla ResNet, and dynamic kernel of our network in terms of parameter magnitude. The kernels in sub- figures are acquired by aver- aging all the kernels, and being normalized in layer dimension.
the effectiveness of the asymmetric design of kernel templates. In-depth analysis of asymmetric kernel templates will be introduced in the visualization section.
为了验证
Effectiveness of Novel Sample Simulation.
新颖的样本模拟方法的有效性，我们还对 DomainMix 进行了消融研究。我们不使用 DomainMix,而是使用有限的源域训练网络。表 4 显示了比较结果，从中我们可以观察到，在没有 DomainMix 策略的情况下进行训练时，我们的方法会遭受明显的性能下降。该结果证明了新颖的样本模拟在教导元调整器适应来自不可知目标域的新颖样本中的重要性。
## 4.5 Visualization
为了更好地理解我们的方法，我们以不同的方式可视化内核模板和动态系数。请注意，我们使用在多源域上训练的模型 (i。e。, PACS 的艺术绘画，卡通和素描)
Visualization of Asymmetric Kernel Templates. 为了进一步探索不对称内核模板的机制，我们通过可视化内核幅度矩阵的方法，按照文献[Ding et al.,2019] 对香草ResNet 中的静态内核和动态网络中的动态内核进行了比较。对于 ResNet, 我们沿着通道和层维度对$3\times3$内核进行求和，以获得$3\times3$内核幅度矩阵。对于动态网络，我们从验证集中随机选择一个图像输入模型，并将所有内核模板与相应的动态系数聚合在一起，沿着通道和层维度求和。比较结果如图 4 所示，从中我们可以看到：
1) The central criss-cross positions have larger values
than the corners of kernel, no matter whether static network or dynamic network. It confirms the kernel skeletons are more crucial than the corners when learning knowledge.
2) The kernel skeletons are further strengthened in dy-
namic network (as shown in Figure 4 (b)). The reason is that the design of dynamic combination among kernel templates can adjust or enlarge the weights of kernel skeletons in instance-aware manner. Consequently, it can promote the model for learning advanced knowledge, which further improves the ability of training-free adapt in inference.
t-SNE Visualization of Dynamic Coefficients and Fea-
为了研究动态网络对不同分布数据的影响，我
$tures.$
们选择了几个块来可视化相应的动态系数和特征。图 5
|
[
[
0.0849609375,
0.06817626953125,
0.316162109375,
0.06817626953125,
0.316162109375,
0.08209228515625,
0.0849609375,
0.08209228515625
],
[
0.7509765625,
0.08209228515625,
0.77783203125,
0.08209228515625,
0.77783203125,
0.0889892578125,
0.7509765625,
0.0889892578125
],
[
0.78515625,
0.083984375,
0.80224609375,
0.083984375,
0.80224609375,
0.08709716796875,
0.78515625,
0.08709716796875
],
[
0.82177734375,
0.080810546875,
0.8486328125,
0.08331298828125,
0.84716796875,
0.09027099609375,
0.8203125,
0.08837890625
],
[
0.08905029296875,
0.08709716796875,
0.48193359375,
0.08709716796875,
0.48193359375,
0.10101318359375,
0.08905029296875,
0.10101318359375
],
[
0.57421875,
0.0947265625,
0.63623046875,
0.0947265625,
0.63623046875,
0.10162353515625,
0.57421875,
0.10162353515625
],
[
0.63818359375,
0.0947265625,
0.66650390625,
0.0947265625,
0.66650390625,
0.10162353515625,
0.63818359375,
0.10162353515625
],
[
0.0882568359375,
0.10162353515625,
0.484375,
0.10162353515625,
0.484375,
0.11553955078125,
0.0882568359375,
0.11553955078125
],
[
0.08660888671875,
0.11553955078125,
0.484375,
0.11553955078125,
0.484375,
0.1287841796875,
0.08660888671875,
0.1287841796875
],
[
0.8203125,
0.10986328125,
0.8447265625,
0.10986328125,
0.8447265625,
0.117431640625,
0.8203125,
0.117431640625
],
[
0.57275390625,
0.12249755859375,
0.6396484375,
0.12249755859375,
0.6396484375,
0.1300048828125,
0.57275390625,
0.1300048828125
],
[
0.63623046875,
0.12249755859375,
0.6650390625,
0.12249755859375,
0.6650390625,
0.1300048828125,
0.63623046875,
0.1300048828125
],
[
0.0882568359375,
0.1287841796875,
0.484375,
0.1287841796875,
0.484375,
0.1427001953125,
0.0882568359375,
0.1427001953125
],
[
0.74658203125,
0.1351318359375,
0.85107421875,
0.1363525390625,
0.85107421875,
0.147705078125,
0.74658203125,
0.146484375
],
[
0.0882568359375,
0.1427001953125,
0.484375,
0.1427001953125,
0.484375,
0.1558837890625,
0.0882568359375,
0.1558837890625
],
[
0.5703125,
0.149658203125,
0.67236328125,
0.149658203125,
0.67236328125,
0.160400390625,
0.5703125,
0.160400390625
],
[
0.08660888671875,
0.1558837890625,
0.484375,
0.1558837890625,
0.484375,
0.169189453125,
0.08660888671875,
0.169189453125
],
[
0.73779296875,
0.1578369140625,
0.859375,
0.1590576171875,
0.859375,
0.173583984375,
0.73779296875,
0.17236328125
],
[
0.08905029296875,
0.17236328125,
0.10296630859375,
0.17236328125,
0.10296630859375,
0.1817626953125,
0.08905029296875,
0.1817626953125
],
[
0.56640625,
0.17236328125,
0.67626953125,
0.17236328125,
0.67626953125,
0.186279296875,
0.56640625,
0.186279296875
],
[
0.73876953125,
0.173583984375,
0.76123046875,
0.173583984375,
0.76123046875,
0.1849365234375,
0.73876953125,
0.1849365234375
],
[
0.10296630859375,
0.184326171875,
0.486083984375,
0.184326171875,
0.486083984375,
0.1976318359375,
0.10296630859375,
0.1976318359375
],
[
0.51123046875,
0.1900634765625,
0.90869140625,
0.1900634765625,
0.90869140625,
0.203857421875,
0.51123046875,
0.203857421875
],
[
0.0882568359375,
0.1988525390625,
0.486083984375,
0.1988525390625,
0.486083984375,
0.212158203125,
0.0882568359375,
0.212158203125
],
[
0.51318359375,
0.2052001953125,
0.91015625,
0.2052001953125,
0.91015625,
0.21533203125,
0.51318359375,
0.21533203125
],
[
0.0882568359375,
0.212158203125,
0.484375,
0.212158203125,
0.484375,
0.2259521484375,
0.0882568359375,
0.2259521484375
],
[
0.5107421875,
0.21533203125,
0.90673828125,
0.21533203125,
0.90673828125,
0.2291259765625,
0.5107421875,
0.2291259765625
],
[
0.0882568359375,
0.2259521484375,
0.484375,
0.2259521484375,
0.484375,
0.2392578125,
0.0882568359375,
0.2392578125
],
[
0.5107421875,
0.2291259765625,
0.90869140625,
0.2279052734375,
0.90869140625,
0.2412109375,
0.5107421875,
0.242431640625
],
[
0.08660888671875,
0.238037109375,
0.486083984375,
0.2392578125,
0.486083984375,
0.25244140625,
0.08660888671875,
0.251953125
],
[
0.5146484375,
0.24365234375,
0.54345703125,
0.24365234375,
0.54345703125,
0.251953125,
0.5146484375,
0.251953125
],
[
0.08905029296875,
0.255126953125,
0.48193359375,
0.255126953125,
0.48193359375,
0.26513671875,
0.08905029296875,
0.26513671875
],
[
0.08660888671875,
0.267578125,
0.484375,
0.267578125,
0.484375,
0.281005859375,
0.08660888671875,
0.281005859375
],
[
0.5361328125,
0.26953125,
0.9111328125,
0.26953125,
0.9111328125,
0.28271484375,
0.5361328125,
0.28271484375
],
[
0.08660888671875,
0.281005859375,
0.484375,
0.27978515625,
0.484375,
0.294189453125,
0.08660888671875,
0.29541015625
],
[
0.5361328125,
0.28271484375,
0.9111328125,
0.2822265625,
0.9111328125,
0.296630859375,
0.5361328125,
0.29736328125
],
[
0.0882568359375,
0.29541015625,
0.486083984375,
0.29541015625,
0.486083984375,
0.309326171875,
0.0882568359375,
0.309326171875
],
[
0.53759765625,
0.296630859375,
0.9111328125,
0.296630859375,
0.9111328125,
0.31005859375,
0.53759765625,
0.31005859375
],
[
0.0882568359375,
0.309326171875,
0.484375,
0.309326171875,
0.484375,
0.322509765625,
0.0882568359375,
0.322509765625
],
[
0.5361328125,
0.31005859375,
0.56884765625,
0.3125,
0.56787109375,
0.324462890625,
0.53515625,
0.322509765625
],
[
0.0882568359375,
0.322509765625,
0.384033203125,
0.322509765625,
0.384033203125,
0.3359375,
0.0882568359375,
0.3359375
],
[
0.51953125,
0.328857421875,
0.82861328125,
0.328857421875,
0.82861328125,
0.34228515625,
0.51953125,
0.34228515625
],
[
0.830078125,
0.333984375,
0.84228515625,
0.333984375,
0.84228515625,
0.337158203125,
0.830078125,
0.337158203125
],
[
0.84326171875,
0.330810546875,
0.9111328125,
0.330810546875,
0.9111328125,
0.34228515625,
0.84326171875,
0.34228515625
],
[
0.1298828125,
0.3466796875,
0.43701171875,
0.34521484375,
0.43701171875,
0.35986328125,
0.1298828125,
0.361083984375
],
[
0.5146484375,
0.343505859375,
0.91259765625,
0.343505859375,
0.91259765625,
0.356689453125,
0.5146484375,
0.356689453125
],
[
0.51318359375,
0.35595703125,
0.91015625,
0.35595703125,
0.91015625,
0.369384765625,
0.51318359375,
0.369384765625
],
[
0.0882568359375,
0.347900390625,
0.13232421875,
0.347900390625,
0.13232421875,
0.359130859375,
0.0882568359375,
0.359130859375
],
[
0.0882568359375,
0.365478515625,
0.486083984375,
0.365478515625,
0.486083984375,
0.37890625,
0.0882568359375,
0.37890625
],
[
0.51318359375,
0.369384765625,
0.9111328125,
0.37060546875,
0.9111328125,
0.385009765625,
0.51318359375,
0.3837890625
],
[
0.0882568359375,
0.37890625,
0.488525390625,
0.37890625,
0.488525390625,
0.392578125,
0.0882568359375,
0.392578125
],
[
0.51318359375,
0.38330078125,
0.91015625,
0.38330078125,
0.91015625,
0.396484375,
0.51318359375,
0.396484375
],
[
0.0882568359375,
0.393310546875,
0.4599609375,
0.393310546875,
0.4599609375,
0.4072265625,
0.0882568359375,
0.4072265625
],
[
0.51318359375,
0.397705078125,
0.90869140625,
0.397705078125,
0.90869140625,
0.412353515625,
0.51318359375,
0.412353515625
],
[
0.08660888671875,
0.410888671875,
0.4755859375,
0.410400390625,
0.4755859375,
0.423583984375,
0.08660888671875,
0.4248046875
],
[
0.5146484375,
0.412353515625,
0.91259765625,
0.412353515625,
0.91259765625,
0.425537109375,
0.5146484375,
0.425537109375
],
[
0.111083984375,
0.425537109375,
0.486083984375,
0.425537109375,
0.486083984375,
0.439453125,
0.111083984375,
0.439453125
],
[
0.5146484375,
0.425537109375,
0.685546875,
0.425537109375,
0.685546875,
0.439453125,
0.5146484375,
0.439453125
],
[
0.111083984375,
0.439453125,
0.483642578125,
0.439453125,
0.483642578125,
0.45263671875,
0.111083984375,
0.45263671875
],
[
0.108642578125,
0.453857421875,
0.486083984375,
0.45263671875,
0.486083984375,
0.466552734375,
0.108642578125,
0.46728515625
],
[
0.5146484375,
0.45068359375,
0.66259765625,
0.45068359375,
0.66259765625,
0.464111328125,
0.5146484375,
0.464111328125
],
[
0.1094970703125,
0.46728515625,
0.484375,
0.46728515625,
0.484375,
0.48095703125,
0.1094970703125,
0.48095703125
],
[
0.5146484375,
0.46728515625,
0.9111328125,
0.468505859375,
0.9111328125,
0.481689453125,
0.5146484375,
0.48095703125
],
[
0.1094970703125,
0.48095703125,
0.40185546875,
0.48095703125,
0.40185546875,
0.494384765625,
0.1094970703125,
0.494384765625
],
[
0.515625,
0.481689453125,
0.9111328125,
0.481689453125,
0.9111328125,
0.49560546875,
0.515625,
0.49560546875
],
[
0.0882568359375,
0.501953125,
0.483642578125,
0.501953125,
0.483642578125,
0.51220703125,
0.0882568359375,
0.51220703125
],
[
0.5146484375,
0.496826171875,
0.87841796875,
0.496826171875,
0.87841796875,
0.51025390625,
0.5146484375,
0.51025390625
],
[
0.1094970703125,
0.5146484375,
0.423095703125,
0.5146484375,
0.423095703125,
0.52783203125,
0.1094970703125,
0.52783203125
],
[
0.4248046875,
0.51513671875,
0.486083984375,
0.51513671875,
0.486083984375,
0.5263671875,
0.4248046875,
0.5263671875
],
[
0.51708984375,
0.51318359375,
0.87841796875,
0.5146484375,
0.87841796875,
0.52880859375,
0.51708984375,
0.52783203125
],
[
0.87744140625,
0.5166015625,
0.91015625,
0.5166015625,
0.91015625,
0.52783203125,
0.87744140625,
0.52783203125
],
[
0.1094970703125,
0.52783203125,
0.483642578125,
0.52783203125,
0.483642578125,
0.54150390625,
0.1094970703125,
0.54150390625
],
[
0.5146484375,
0.52880859375,
0.9111328125,
0.52880859375,
0.9111328125,
0.54248046875,
0.5146484375,
0.54248046875
],
[
0.111083984375,
0.54345703125,
0.48779296875,
0.54345703125,
0.48779296875,
0.5537109375,
0.111083984375,
0.5537109375
],
[
0.5146484375,
0.54345703125,
0.91015625,
0.54345703125,
0.91015625,
0.5537109375,
0.5146484375,
0.5537109375
],
[
0.1094970703125,
0.5546875,
0.486083984375,
0.5546875,
0.486083984375,
0.568359375,
0.1094970703125,
0.568359375
],
[
0.51123046875,
0.55615234375,
0.91015625,
0.556640625,
0.91015625,
0.57080078125,
0.51123046875,
0.5693359375
],
[
0.1094970703125,
0.568359375,
0.486083984375,
0.568359375,
0.486083984375,
0.58203125,
0.1094970703125,
0.58203125
],
[
0.5146484375,
0.57080078125,
0.9111328125,
0.57080078125,
0.9111328125,
0.583984375,
0.5146484375,
0.583984375
],
[
0.1094970703125,
0.58349609375,
0.484375,
0.58349609375,
0.484375,
0.5966796875,
0.1094970703125,
0.5966796875
],
[
0.5146484375,
0.583984375,
0.9111328125,
0.583984375,
0.9111328125,
0.59765625,
0.5146484375,
0.59765625
],
[
0.108642578125,
0.5966796875,
0.486083984375,
0.59521484375,
0.486083984375,
0.60986328125,
0.108642578125,
0.611328125
],
[
0.51318359375,
0.5986328125,
0.9111328125,
0.5986328125,
0.9111328125,
0.6123046875,
0.51318359375,
0.6123046875
],
[
0.1094970703125,
0.611328125,
0.451904296875,
0.611328125,
0.451904296875,
0.625,
0.1094970703125,
0.625
],
[
0.51318359375,
0.60986328125,
0.9111328125,
0.6123046875,
0.9111328125,
0.626953125,
0.51318359375,
0.625
],
[
0.5146484375,
0.62548828125,
0.9111328125,
0.62548828125,
0.9111328125,
0.6396484375,
0.5146484375,
0.6396484375
],
[
0.08660888671875,
0.63525390625,
0.26318359375,
0.63525390625,
0.26318359375,
0.64599609375,
0.08660888671875,
0.64599609375
],
[
0.5146484375,
0.64013671875,
0.53759765625,
0.64013671875,
0.53759765625,
0.6513671875,
0.5146484375,
0.6513671875
],
[
0.0882568359375,
0.65283203125,
0.484375,
0.65283203125,
0.484375,
0.66650390625,
0.0882568359375,
0.66650390625
],
[
0.51123046875,
0.6572265625,
0.91259765625,
0.65771484375,
0.91259765625,
0.67236328125,
0.51123046875,
0.671875
],
[
0.08905029296875,
0.66748046875,
0.486083984375,
0.66748046875,
0.486083984375,
0.68115234375,
0.08905029296875,
0.68115234375
],
[
0.53515625,
0.671875,
0.91259765625,
0.671875,
0.91259765625,
0.68505859375,
0.53515625,
0.68505859375
],
[
0.0882568359375,
0.681640625,
0.111083984375,
0.681640625,
0.111083984375,
0.6943359375,
0.0882568359375,
0.6943359375
],
[
0.5361328125,
0.6875,
0.90869140625,
0.68603515625,
0.90869140625,
0.69775390625,
0.5361328125,
0.69873046875
],
[
0.09234619140625,
0.69970703125,
0.484375,
0.69970703125,
0.484375,
0.71337890625,
0.09234619140625,
0.71337890625
],
[
0.53759765625,
0.7021484375,
0.9111328125,
0.7021484375,
0.9111328125,
0.7119140625,
0.53759765625,
0.7119140625
],
[
0.0882568359375,
0.71337890625,
0.484375,
0.71337890625,
0.484375,
0.7265625,
0.0882568359375,
0.7265625
],
[
0.53515625,
0.7138671875,
0.67724609375,
0.7138671875,
0.67724609375,
0.72802734375,
0.53515625,
0.72802734375
],
[
0.08905029296875,
0.7265625,
0.484375,
0.7265625,
0.484375,
0.74072265625,
0.08905029296875,
0.74072265625
],
[
0.51123046875,
0.73193359375,
0.91015625,
0.73193359375,
0.91015625,
0.74560546875,
0.51123046875,
0.74560546875
],
[
0.0882568359375,
0.7412109375,
0.484375,
0.7412109375,
0.484375,
0.7548828125,
0.0882568359375,
0.7548828125
],
[
0.5361328125,
0.74560546875,
0.91015625,
0.74560546875,
0.91015625,
0.7587890625,
0.5361328125,
0.7587890625
],
[
0.08905029296875,
0.7548828125,
0.486083984375,
0.7548828125,
0.486083984375,
0.76806640625,
0.08905029296875,
0.76806640625
],
[
0.5361328125,
0.76025390625,
0.91259765625,
0.76025390625,
0.91259765625,
0.7734375,
0.5361328125,
0.7734375
],
[
0.08660888671875,
0.76806640625,
0.3896484375,
0.76806640625,
0.3896484375,
0.7822265625,
0.08660888671875,
0.7822265625
],
[
0.53759765625,
0.77587890625,
0.9111328125,
0.77587890625,
0.9111328125,
0.7861328125,
0.53759765625,
0.7861328125
],
[
0.08905029296875,
0.78857421875,
0.48193359375,
0.78857421875,
0.48193359375,
0.798828125,
0.08905029296875,
0.798828125
],
[
0.5361328125,
0.78857421875,
0.91015625,
0.78857421875,
0.91015625,
0.8017578125,
0.5361328125,
0.8017578125
],
[
0.111083984375,
0.8017578125,
0.484375,
0.8017578125,
0.484375,
0.81494140625,
0.111083984375,
0.81494140625
],
[
0.5361328125,
0.8017578125,
0.91015625,
0.8017578125,
0.91015625,
0.81494140625,
0.5361328125,
0.81494140625
],
[
0.1094970703125,
0.81494140625,
0.484375,
0.81494140625,
0.484375,
0.8291015625,
0.1094970703125,
0.8291015625
],
[
0.53759765625,
0.81640625,
0.9111328125,
0.81640625,
0.9111328125,
0.830078125,
0.53759765625,
0.830078125
],
[
0.111083984375,
0.83056640625,
0.15283203125,
0.83056640625,
0.15283203125,
0.84228515625,
0.111083984375,
0.84228515625
],
[
0.5361328125,
0.830078125,
0.78173828125,
0.830078125,
0.78173828125,
0.84326171875,
0.5361328125,
0.84326171875
],
[
0.0849609375,
0.84765625,
0.484375,
0.84765625,
0.484375,
0.861328125,
0.0849609375,
0.861328125
],
[
0.51953125,
0.84765625,
0.91015625,
0.84765625,
0.91015625,
0.861328125,
0.51953125,
0.861328125
],
[
0.108642578125,
0.8623046875,
0.484375,
0.861328125,
0.484375,
0.875,
0.108642578125,
0.87548828125
],
[
0.5146484375,
0.86279296875,
0.57275390625,
0.86279296875,
0.57275390625,
0.875,
0.5146484375,
0.875
],
[
0.576171875,
0.85986328125,
0.9111328125,
0.861328125,
0.9111328125,
0.87548828125,
0.576171875,
0.875
],
[
0.1094970703125,
0.87548828125,
0.486083984375,
0.87548828125,
0.486083984375,
0.8896484375,
0.1094970703125,
0.8896484375
],
[
0.5146484375,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.5146484375,
0.8896484375
]
] |
[
"4.2 Implementation Details",
"0.124",
"",
"0.134",
"所有实验均在带有PyTorch1的RTX3090GPU上进行",
".7560.860",
"0810",
"10。0。我们使用ResNet-50作为动态网络的骨干，通过",
"将元调整器附加到原始残差块（如图3所示），其中内",
"0.45",
"0.8350952",
"10.91",
"核模板的空间大小设置为k=3。请注意，在每个块中",
"0.1320.5800.147",
"元调整器由全局平均池化层，夹在两个全连接层之间的",
"0.7990.9350.875",
"ReLU层和SoftMax层的顺序实现，以输出实例感知系",
"(b) Dynamic ker-",
"数",
"(a) Static kernel",
"nel",
"在训练DG任务之前，我们在ImageNet上预训练我",
"Figure 4: Comparison among static kernel of vanilla ResNet",
"们的网络，以便与现有作品进行公平的比较。然后，我",
"and dynamic kernel of our network in terms of parameter",
"们通过使用SGD微调预训练模型来训练DG任务。训",
" magnitude. The kernels in sub-figures are acquired by aver",
"练目标是交叉熵损失函数，与香草DG完全相同。在训",
"aging all the kernels, and being normalized in layer dimen",
"练过程中，非对称内核模板和元调整器与主干网同时优",
"sion.",
"化。对于PACS和Ofice-Home，网络优化设置为批处",
"理大小为64，训练时期为50，初始学习速率为1e-3，由",
"the effectiveness of the asymmetric design of kernel",
"余弦调度器衰减。在DomainNet上训练时，大多数超参",
"templates..In-depth analysis of asymmetric kernel",
"数与PACS相同，只是初始学习率和最大epoch为2e-3",
"templates will be introduced in the visualization sec-",
"和15，并且使用随机域采样器策略E[Zhouetal.,2021]",
"tion",
"获取小批量，以确保每个域都被均匀采样",
"Effectiveness of Novel Sample Simulation.",
"",
"为了验证",
"Comparisons with the State-of-the-Art",
"新颖的样本模拟方法的有效性，我们还对DomainMix",
"进行了消融研究。我们不使用DomainMix，而是使用有",
"4.3",
"我们将我们的方法与PACS，Office-Home和DomainNet",
"限的源域训练网络。表4显示了比较结果，从中我们可",
"数据集上的最新方法进行了比较，其中结果分别在表1，",
"以观察到，在没有DomainMix策略的情况下进行训练",
"表2和表3中报告。从这些表中，我们可以观察到：",
"时，我们的方法会遭受明显的性能下降。该结果证明了",
"1) Our method achieves an average accuracy of 86.21",
"新颖的样本模拟在教导元调整器适应来自不可知目标域",
"72.13 and 46.89 percent on PACS, Office-Home and",
"的新颖样本中的重要性。",
"DomainNet benchmark datasets, respectively. Com-",
"pared to other DG methods, our method achieves the",
"4.5Visualization",
"state-of-the-art results on most of DG benchmarks,",
"为了更好地理解我们的方法，我们以不同的方式可视化",
"which shows the superiority of our DDG.",
"内核模板和动态系数。请注意，我们使用在多源域上训",
"2)Ourmethodisflexibletocombinewithothermeth-",
"练的模型(i。e。，PACS的艺术绘画，卡通和素描）",
"ods to further improve the performance.",
"Taking",
"Visualization of Asymmetric Kernel Templates.",
"为了",
"Mixstyle [Zhou et al., 2021] as an example, the per-",
"进一步探索不对称内核模板的机制，我们通过可视化内",
"formanceof ourmodel canbefurtherimproved if",
"核幅度矩阵的方法按照文献Dingetal，2019l对香草",
"we insert MixStyle layers into our model.The re-",
"ResNet中的静态内核和动态网络中的动态内核进行了",
"sults can reach 87.87 , 72.31 and 46.93 percent on",
"比较。对于ResNet，我们沿着通道和层维度对3×3内",
"PACS, Ofice-Home and DomainNet benchmarks, re-",
"核进行求和，以获得3×3内核幅度矩阵。对于动态网",
"spectively. It shows that our method is rather robust",
"络，我们从验证集中随机选择一个图像输人模型，并将",
"even on the very strong baselines, i.e. MixStyle.",
"所有内核模板与相应的动态系数聚合在一起，沿着通道",
"和层维度求和。比较结果如图4所示，从中我们可以看",
"4.4Ablation Studies",
"到：",
"我们对PACS数据集进行消融研究，以分别评估非对称",
"1) The central criss-cross positions have larger values",
"内核模板和DomainMix的有效性，其中结果如表4所",
"than the corners of kernel, no matter whether static",
"示。",
"network or dynamic network.It confirms the ker",
"Effectiveness ofAsymmetricKernelTemplates.我们",
"nel skeletons are more crucial thanthe corners when",
"通过使用由四个空间尺寸为1×1或3×3的相同内核组",
"learning knowledge.",
"成的通用内核模板，研究了内核模板非对称设计的有效",
"2) The kernel skeletons are further strengthened in dy-",
"性。此外，我们还比较了一种渠道关注方法，即SENetF",
"namic network (as shown in Figure 4 (b)). The rea-",
"[Huetal.,2018]，它也是CNN社区中流行且功能强大",
"son is that the design of dynamic combination among",
"的体系结构。如表4所示，我们可以看到：",
"kernel templates can adjust or enlargetheweights",
"D)Ourmethod achievessuperiorperformancetothe",
"of kernel skeletons in instance-aware manner Con-",
"vanilla ResNet and SENet by a large margin, which",
"sequently, it can promote the model for learning ad-",
"suggests the effectiveness of meta-adjusters on DG",
"vanced knowledge, which further improves the ability",
"tasks.",
"of training-free adapt in inference.",
"2) The result of asymmetric kernel templates is obvi-",
"t-SNE Visualization of Dynamic Coefficients and Fea-",
"ously better than that of four itentical 1 × 1 kernel",
"tures.",
"为了研究动态网络对不同分布数据的影响，我",
"templates and 3 × 3 kernel templates, which confirms",
"们选择了几个块来可视化相应的动态系数和特征。图5"
] |
{
"arxiv_id": "2205.13913",
"language": "zh",
"page_idx": 4
}
|
||
2210.03053-zh-1
|
2210.03053-zh-1.jpg
|
用于神经机器翻译的具有大动作空间的强化学习
# Abstract
Applying Reinforcement learning (RL) following maximum likelihood estimation (MLE) pre-training is a versatile method for enhancing neural machine translation (NMT) performance. However, recent work has argued that the gains produced by RL for NMT are mostly due to promoting tokens that have already received a fairly high probability in pre-training. We hypothesize that the large action space is a main obstacle to RL's effectiveness in MT, and conduct two sets of experiments that lend support to our hypothesis. First, we find that reducing the size of the vocabulary improves RL's effectiveness. Second, we find that effectively reducing the dimension of the action space without changing the vocabulary also yields notable improvement as evaluated by ${\mathrm{BLEU,~semantic~similarity,~and~human~evalua-}}$ tion. Indeed, by initializing the network's final fully connected layer (that maps the network's internal dimension to the vocabulary dimension), with a layer that generalizes over similar actions, we obtain a substantial improvement in RL performance: 1.5 BLEU points on average.$^{\mathrm{l}}$
## 1 $\textbf{Introduction}$
序列到序列任务的标准训练方法，特别是对于NMT,是在给定黄金标准前缀(以下称为最大似然估计或 MLE)的情况下，最大化目标句子中令牌的可能性。然而，尽管 MLE 训练的模型显示了强大的性能，但这种令牌级目标函数在惩罚序列级错误的能力方面受到限制，并且与它旨在改进的序列级评估指标不一致。解决这一差距的一种有吸引力的方法是应用策略梯度方法，该方法允许合并不可微分的奖励函数，例如经常用于 MT 评估的函数(?,请参阅$\S$2)。为简洁起见，我们将这些方法简称为 RL
RL 训练过程包括几个步骤：(1)使用预训练的 MLE 模型生成翻译，(2)计算一些序列
$^{1}$https://github.com/AsafYehudai/Reinforcement-
Learning-with-Large-Action-Spaces-for-Neural-Machine-
Translation
级奖励函数，通常是评估生成的翻译和参考的相似性的函数，以及(3)更新模型，以使其未来的输出获得更高的奖励。该方法的灵活性以及解决曝光偏差 (Ranzato et al., 2016; Wang and Sennrich, 2020) 的能力使 RL 成为提高 NMT 性能的诱人途径。但是，最近的一项研究(C19; ?)表明，只有在 MLE 模型已经为该令牌分配了相当高的概率的情况下，当前的 RL 实践才有可能改善目标令牌的预测
在这项工作中，我们观察到 NMT 与 RL 方法擅长的其他任务之间的一个主要区别是动作空间的大小。通常，NMT 中动作空间的大小包括词汇表中的所有令牌，通常为数万个。相比之下，常见的 RL 设置具有较小的离散动作空间 (例如。g。, 雅达利游戏 (Mnih et al., 2013) ), 或低维度的连续动作空间 (e。g。、MuJoCo (Todorov et al., 2012) 和类似的控制问题)。直观地，RL 采取 (样本)行动并评估其结果，这与监督学习 (MLE) 不同，后者直接接收所有行动的分数。因此，较大的操作空间将降低 RL 的效率，因为必须对单个操作进行采样以评估其质量。因此，我们尝试了两种减小动作空间大小的方法，并评估了它们对 RL 有效性的影响
我们首先减少词汇量(或等效地减少操作数量),在低资源环境中进行将四种语言翻译成英语的实验，同时使用 BLEU 作为奖励函数和评估指标。我们的结果表明，与标准词汇量的RL相比，RL 比 MLE 训练产生了更大的性能提高 (平均约为1 BLEU 点)。此外，我们的研究结果表明，即使在不接近正确的情况下，减少词汇量也可以改善 MLE 模型。请参见 §4
但是，在某些情况下，更改词汇表可能是不希望的或不可行的。因此，我们尝试了两种方法，可以有效地降低动作空间的维数而不改变词汇量。我们注意到，通常在 NMT 架构中，解码器的内部层(此后称为$d$)的维数明显小于目标词汇表大小(此后称为$\left|V_T\right|$),即动作空间的大小。全连接层通常用于将内部表示映射到合适的输出。因此，我们可以将此层的矩阵(参数)的行称为 target embeddings,将网络的
|
[
[
0.2493896484375,
0.0968017578125,
0.75,
0.0968017578125,
0.75,
0.10986328125,
0.2493896484375,
0.10986328125
],
[
0.2626953125,
0.254150390625,
0.340087890625,
0.254150390625,
0.340087890625,
0.268310546875,
0.2626953125,
0.268310546875
],
[
0.51318359375,
0.254638671875,
0.880859375,
0.254638671875,
0.880859375,
0.268310546875,
0.51318359375,
0.268310546875
],
[
0.51318359375,
0.26953125,
0.87890625,
0.26953125,
0.87890625,
0.283203125,
0.51318359375,
0.283203125
],
[
0.145263671875,
0.27783203125,
0.45751953125,
0.276611328125,
0.45751953125,
0.29150390625,
0.145263671875,
0.292236328125
],
[
0.51123046875,
0.2861328125,
0.87890625,
0.287353515625,
0.87890625,
0.300537109375,
0.51123046875,
0.2998046875
],
[
0.14697265625,
0.29345703125,
0.456787109375,
0.292236328125,
0.456787109375,
0.303955078125,
0.14697265625,
0.3046875
],
[
0.145263671875,
0.30712890625,
0.45751953125,
0.305908203125,
0.45751953125,
0.319580078125,
0.145263671875,
0.320556640625
],
[
0.5146484375,
0.302734375,
0.880859375,
0.302734375,
0.880859375,
0.31640625,
0.5146484375,
0.31640625
],
[
0.14697265625,
0.32177734375,
0.45751953125,
0.32177734375,
0.45751953125,
0.3349609375,
0.14697265625,
0.3349609375
],
[
0.51318359375,
0.319580078125,
0.880859375,
0.319580078125,
0.880859375,
0.333251953125,
0.51318359375,
0.333251953125
],
[
0.14697265625,
0.3349609375,
0.45751953125,
0.3349609375,
0.45751953125,
0.3486328125,
0.14697265625,
0.3486328125
],
[
0.5146484375,
0.33740234375,
0.87890625,
0.33740234375,
0.87890625,
0.347412109375,
0.5146484375,
0.347412109375
],
[
0.145263671875,
0.347412109375,
0.455078125,
0.347412109375,
0.455078125,
0.361083984375,
0.145263671875,
0.361083984375
],
[
0.51318359375,
0.3515625,
0.880859375,
0.3515625,
0.880859375,
0.365234375,
0.51318359375,
0.365234375
],
[
0.145263671875,
0.363525390625,
0.45751953125,
0.363525390625,
0.45751953125,
0.377197265625,
0.145263671875,
0.377197265625
],
[
0.51708984375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.51708984375,
0.3818359375
],
[
0.145263671875,
0.377197265625,
0.455078125,
0.377685546875,
0.455078125,
0.391357421875,
0.145263671875,
0.39013671875
],
[
0.51318359375,
0.38427734375,
0.73291015625,
0.38427734375,
0.73291015625,
0.39794921875,
0.51318359375,
0.39794921875
],
[
0.1435546875,
0.39013671875,
0.456787109375,
0.391357421875,
0.456787109375,
0.405029296875,
0.1435546875,
0.40380859375
],
[
0.1435546875,
0.405029296875,
0.456787109375,
0.40380859375,
0.456787109375,
0.418701171875,
0.1435546875,
0.419189453125
],
[
0.5322265625,
0.402099609375,
0.880859375,
0.402099609375,
0.880859375,
0.4150390625,
0.5322265625,
0.4150390625
],
[
0.1485595703125,
0.421630859375,
0.456787109375,
0.421630859375,
0.456787109375,
0.431640625,
0.1485595703125,
0.431640625
],
[
0.51318359375,
0.418701171875,
0.880859375,
0.418701171875,
0.880859375,
0.431640625,
0.51318359375,
0.431640625
],
[
0.145263671875,
0.435302734375,
0.456787109375,
0.43408203125,
0.456787109375,
0.44775390625,
0.145263671875,
0.448974609375
],
[
0.51318359375,
0.43408203125,
0.880859375,
0.432861328125,
0.880859375,
0.44775390625,
0.51318359375,
0.448974609375
],
[
0.14697265625,
0.448974609375,
0.453369140625,
0.448974609375,
0.453369140625,
0.4619140625,
0.14697265625,
0.4619140625
],
[
0.5146484375,
0.45068359375,
0.880859375,
0.45068359375,
0.880859375,
0.46435546875,
0.5146484375,
0.46435546875
],
[
0.1485595703125,
0.46435546875,
0.455078125,
0.46435546875,
0.455078125,
0.474365234375,
0.1485595703125,
0.474365234375
],
[
0.5146484375,
0.466064453125,
0.880859375,
0.466064453125,
0.880859375,
0.479736328125,
0.5146484375,
0.479736328125
],
[
0.145263671875,
0.476806640625,
0.45751953125,
0.476806640625,
0.45751953125,
0.490478515625,
0.145263671875,
0.490478515625
],
[
0.5146484375,
0.482666015625,
0.880859375,
0.482666015625,
0.880859375,
0.496337890625,
0.5146484375,
0.496337890625
],
[
0.1485595703125,
0.4912109375,
0.456787109375,
0.4912109375,
0.456787109375,
0.5048828125,
0.1485595703125,
0.5048828125
],
[
0.51318359375,
0.498779296875,
0.880859375,
0.498779296875,
0.880859375,
0.51318359375,
0.51318359375,
0.51318359375
],
[
0.145263671875,
0.5048828125,
0.45751953125,
0.5048828125,
0.45751953125,
0.5185546875,
0.145263671875,
0.5185546875
],
[
0.1435546875,
0.5185546875,
0.459228515625,
0.51953125,
0.459228515625,
0.533203125,
0.1435546875,
0.5322265625
],
[
0.51123046875,
0.51416015625,
0.87890625,
0.51416015625,
0.87890625,
0.52783203125,
0.51123046875,
0.52783203125
],
[
0.14697265625,
0.53369140625,
0.45751953125,
0.53369140625,
0.45751953125,
0.54736328125,
0.14697265625,
0.54736328125
],
[
0.51123046875,
0.529296875,
0.880859375,
0.52978515625,
0.880859375,
0.54443359375,
0.51123046875,
0.54345703125
],
[
0.14697265625,
0.54736328125,
0.459228515625,
0.54736328125,
0.459228515625,
0.56103515625,
0.14697265625,
0.56103515625
],
[
0.51318359375,
0.5458984375,
0.880859375,
0.54638671875,
0.880859375,
0.56103515625,
0.51318359375,
0.56005859375
],
[
0.14697265625,
0.5625,
0.45751953125,
0.5625,
0.45751953125,
0.576171875,
0.14697265625,
0.576171875
],
[
0.5146484375,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.5146484375,
0.57666015625
],
[
0.1485595703125,
0.57763671875,
0.455078125,
0.57763671875,
0.455078125,
0.58837890625,
0.1485595703125,
0.58837890625
],
[
0.5146484375,
0.5791015625,
0.880859375,
0.5791015625,
0.880859375,
0.5927734375,
0.5146484375,
0.5927734375
],
[
0.1485595703125,
0.5927734375,
0.456787109375,
0.5927734375,
0.456787109375,
0.6025390625,
0.1485595703125,
0.6025390625
],
[
0.51123046875,
0.59326171875,
0.880859375,
0.59423828125,
0.880859375,
0.609375,
0.51123046875,
0.60791015625
],
[
0.145263671875,
0.60400390625,
0.45751953125,
0.60400390625,
0.45751953125,
0.61767578125,
0.145263671875,
0.61767578125
],
[
0.51318359375,
0.61083984375,
0.880859375,
0.61083984375,
0.880859375,
0.62451171875,
0.51318359375,
0.62451171875
],
[
0.14697265625,
0.62060546875,
0.185546875,
0.62060546875,
0.185546875,
0.6318359375,
0.14697265625,
0.6318359375
],
[
0.51318359375,
0.62744140625,
0.53564453125,
0.62744140625,
0.53564453125,
0.64013671875,
0.51318359375,
0.64013671875
],
[
0.1175537109375,
0.642578125,
0.145263671875,
0.642578125,
0.145263671875,
0.65478515625,
0.1175537109375,
0.65478515625
],
[
0.1502685546875,
0.64111328125,
0.2626953125,
0.64111328125,
0.2626953125,
0.65625,
0.1502685546875,
0.65625
],
[
0.5322265625,
0.6455078125,
0.880859375,
0.6455078125,
0.880859375,
0.6591796875,
0.5322265625,
0.6591796875
],
[
0.118408203125,
0.66748046875,
0.48779296875,
0.66748046875,
0.48779296875,
0.68115234375,
0.118408203125,
0.68115234375
],
[
0.5146484375,
0.6630859375,
0.880859375,
0.6630859375,
0.880859375,
0.67333984375,
0.5146484375,
0.67333984375
],
[
0.1175537109375,
0.68310546875,
0.48779296875,
0.68310546875,
0.48779296875,
0.69677734375,
0.1175537109375,
0.69677734375
],
[
0.51318359375,
0.677734375,
0.880859375,
0.677734375,
0.880859375,
0.69140625,
0.51318359375,
0.69140625
],
[
0.118408203125,
0.69970703125,
0.486083984375,
0.69970703125,
0.486083984375,
0.71337890625,
0.118408203125,
0.71337890625
],
[
0.5146484375,
0.6953125,
0.87890625,
0.6953125,
0.87890625,
0.7060546875,
0.5146484375,
0.7060546875
],
[
0.12005615234375,
0.71533203125,
0.484375,
0.71533203125,
0.484375,
0.728515625,
0.12005615234375,
0.728515625
],
[
0.51318359375,
0.71044921875,
0.880859375,
0.71044921875,
0.880859375,
0.7236328125,
0.51318359375,
0.7236328125
],
[
0.51318359375,
0.7255859375,
0.880859375,
0.7255859375,
0.880859375,
0.7392578125,
0.51318359375,
0.7392578125
],
[
0.118408203125,
0.732421875,
0.484375,
0.732421875,
0.484375,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.51123046875,
0.7412109375,
0.880859375,
0.7412109375,
0.880859375,
0.7548828125,
0.51123046875,
0.7548828125
],
[
0.1175537109375,
0.74755859375,
0.486083984375,
0.74658203125,
0.486083984375,
0.76123046875,
0.1175537109375,
0.7626953125
],
[
0.51123046875,
0.75732421875,
0.85986328125,
0.75732421875,
0.85986328125,
0.77099609375,
0.51123046875,
0.77099609375
],
[
0.12005615234375,
0.76416015625,
0.486083984375,
0.76416015625,
0.486083984375,
0.77783203125,
0.12005615234375,
0.77783203125
],
[
0.118408203125,
0.77978515625,
0.484375,
0.77978515625,
0.484375,
0.79345703125,
0.118408203125,
0.79345703125
],
[
0.52880859375,
0.77490234375,
0.880859375,
0.775390625,
0.880859375,
0.79052734375,
0.52880859375,
0.7890625
],
[
0.12005615234375,
0.7958984375,
0.484375,
0.7958984375,
0.484375,
0.80859375,
0.12005615234375,
0.80859375
],
[
0.51318359375,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.118408203125,
0.8134765625,
0.48193359375,
0.8134765625,
0.48193359375,
0.82373046875,
0.118408203125,
0.82373046875
],
[
0.51318359375,
0.80859375,
0.880859375,
0.80859375,
0.880859375,
0.822265625,
0.51318359375,
0.822265625
],
[
0.1175537109375,
0.82763671875,
0.484375,
0.82763671875,
0.484375,
0.84130859375,
0.1175537109375,
0.84130859375
],
[
0.51318359375,
0.82470703125,
0.880859375,
0.82373046875,
0.880859375,
0.83837890625,
0.51318359375,
0.8388671875
],
[
0.1368408203125,
0.84326171875,
0.486083984375,
0.84326171875,
0.486083984375,
0.85693359375,
0.1368408203125,
0.85693359375
],
[
0.51123046875,
0.84130859375,
0.87744140625,
0.84130859375,
0.87744140625,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.118408203125,
0.85986328125,
0.486083984375,
0.85986328125,
0.486083984375,
0.87353515625,
0.118408203125,
0.87353515625
],
[
0.51318359375,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.51318359375,
0.87353515625,
0.880859375,
0.87353515625,
0.880859375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.139404296875,
0.8818359375,
0.43408203125,
0.8818359375,
0.43408203125,
0.8955078125,
0.139404296875,
0.8955078125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.1175537109375,
0.8955078125,
0.465087890625,
0.89453125,
0.465087890625,
0.90625,
0.1175537109375,
0.90673828125
],
[
0.118408203125,
0.90673828125,
0.185546875,
0.90673828125,
0.185546875,
0.91748046875,
0.118408203125,
0.91748046875
],
[
0.509765625,
0.90625,
0.87890625,
0.90625,
0.87890625,
0.91943359375,
0.509765625,
0.91943359375
]
] |
[
"用于神经机器翻译的具有大动作空间的强化学习",
"Abstract",
"级奖励函数，通常是评估生成的翻译和参考的",
"相似性的函数，以及（3）更新模型，以使其未",
"ApplyingReinforcement learning (RL)follow",
"来的输出获得更高的奖励。该方法的灵活性以",
"ingmaximumlikelihoodestimation(MLE)",
"pre-training is a versatile method for enhanc-",
"及解决曝光偏差（Ranzatoetal.2016;Wangand",
"ing neural machine translation (NMT) perfor-",
"Sennrich，2020）的能力使RL成为提高NMT性",
"mance. However, recent work has argued that",
"能的诱人途径。但是，最近的一项研究（C19",
"the gains produced by RL for NMT are mostly",
"?）表明，只有在MLE模型已经为该令牌分配",
"due to promoting tokens that have already re-",
"了相当高的概率的情况下，当前的RL实践才",
"ceived a fairly high probability in pre-training",
"有可能改善目标令牌的预测",
"We hypothesize that the large action space is",
"a main obstacle to RL's effectiveness in MT,",
"在这项工作中，我们观察到NMT与RL方",
"andconducttwosetsofexperimentsthatlend",
"法擅长的其他任务之间的一个主要区别是动作",
"support to our hypothesis. First, we find that",
"空间的大小。通常，NMT中动作空间的大小",
"reducing the size of the vocabulary improves",
"包括词汇表中的所有令牌，通常为数万个。相",
"RL's effectiveness.Second,we find that ef",
"比之下，常见的RL设置具有较小的离散动作",
"fectively reducing the dimension of the action",
"空间（例如。g。，雅达利游戏（Mnihetal.,2013)",
"space without changing the vocabulary also",
")，或低维度的连续动作空间(e。g。、MuJoCo",
"yields notable improvement as evaluated by",
"BLEU, semantic similarity, and human evalua-",
"(Todorovetal.2012）和类似的控制问题）。直观",
"tion. Indeed, by initializing the network's final",
"地，RL采取（样本）行动并评估其结果，这与",
"fully connected layer (that maps the network's",
"监督学习（MLE）不同，后者直接接收所有行",
"internal dimension to the vocabulary dimen-",
"动的分数。因此，较大的操作空间将降低RL",
"sion),with alayerthatgeneralizes over similar",
"的效率，因为必须对单个操作进行采样以评估",
"actions,weobtainasubstantialimprovement",
"其质量。因此，我们尝试了两种减小动作空间",
"in RL performance: 1.5 BLEU points on aver-",
"大小的方法，并评估了它们对RL有效性的影",
"age.",
"响",
"1",
"Introduction",
"我们首先减少词汇量（或等效地减少操作数",
"序列到序列任务的标准训练方法，特别是对于",
"量），在低资源环境中进行将四种语言翻译成",
"NMT，是在给定黄金标准前缀（以下称为最大",
"英语的实验，同时使用BLEU作为奖励函数和",
"似然估计或MLE）的情况下，最大化目标句",
"评估指标。我们的结果表明，与标准词汇量的",
"子中令牌的可能性。然而，尽管MLE训练的",
"RL相比，RL比MLE训练产生了更大的性能",
"提高（平均约为1BLEU点）。此外，我们的研",
"模型显示了强大的性能，但这种令牌级目标函",
"究结果表明，即使在不接近正确的情况下，减",
"数在惩罚序列级错误的能力方面受到限制，并",
"少词汇量也可以改善MLE模型。请参见84",
"且与它旨在改进的序列级评估指标不一致。解",
"决这一差距的一种有吸引力的方法是应用策略",
"但是，在某些情况下，更改词汇表可能是不",
"梯度方法，该方法允许合并不可微分的奖励函",
"希望的或不可行的。因此，我们尝试了两种方",
"数，例如经常用于MT评估的函数（？，请参阅",
"法，可以有效地降低动作空间的维数而不改变",
"82）。为简洁起见，我们将这些方法简称为RL",
"词汇量。我们注意到，通常在NMT架构中，解",
"RL训练过程包括几个步骤：（1）使用预训",
"码器的内部层（此后称为d）的维数明显小于",
"练的MLE模型生成翻译，（2）计算一些序列",
"目标词汇表大小（此后称为|VTl），即动作空",
"间的大小。全连接层通常用于将内部表示映射",
"'https://github.com/AsafYehudai/Reinforcement",
"到合适的输出。因此，我们可以将此层的矩阵",
"Learning-with-Large-Action-Spaces-for-Neural-Machine",
"Translation",
"（参数）的行称为targetembeddings，将网络的"
] |
{
"arxiv_id": "2210.03053",
"language": "zh",
"page_idx": 1
}
|
||
2210.03053-zh-2
|
2210.03053-zh-2.jpg
|
使用这个术语来强调网络的第一个嵌入层之间的类比，将维度$\left|V_T\right|$的向量映射到维度$d$ 的向量，以及以相反方式工作的目标嵌入。事实上， (e。g。,在 BERT, ?)中，源嵌人和目标嵌入的权重在训练过程中是共享的，强调了两者之间的关系。使用这个术语，我们在模拟(§5.1 )中表明，当类似的动作共享目标嵌入时，RL 更有效此外，当基于高质量嵌入(在我们的例子中是 BERT) 初始化目标嵌入时，在 RL 期间冻结它们仍然会产生进一步的改进。在 NMT 上进行实验时，我们获得了类似的结果。实际上，将 BERT 的嵌入用于目标嵌入可以提高四个语言对的性能，并且冻结它们会对 MLE 和RL产生额外的改进，如自动指标和人工评估所报告的。在 NMT 的 RL 训练中，初始化和冻结都是新颖的。此外，当使用 BERT 的嵌入时，RL 提高预训练的 MLE 模型没有分配高概率的目标令牌的性能的能力得到了增强(§5.2 )。
# 2 Background
2.1 RL in Machine Translation
RL 用于文本生成 (TG),因为它能够合并不可微分信号 ,解决曝光偏差并引入序列级约束。后两者是 TG 系统开发中的持续挑战，也已通过
这些特性引起了对一般(Shah et al., 2018) 和NMT 尤其是(Wu et al., 2018a) 中的 TG 的 RL 的极大兴趣。通常使用许多策略梯度方法，特别是 REINFORCE (Williams, 1992) 和最小风险培训 (MRT;e。g。,Och,2003;?)。然而，尽管人们的兴趣越来越大，结果也越来越好，但只有少数作品研究了 NLP 中 RL 观察到的性能提高的来源及其训练动力学，其中一些表明RL 的提高部分归因于工件 (Caccia et al., 2018; Choshen et al., 2019)
在最近的一篇论文中，C19 表明，现有的MTRL 训练协议(REINFORCE 和 MRT)需要很长的时间才能收敛。他们的结果表明，只有在 MLE 参数已经接近产生正确翻译的情况下， MT 中的 RL 实践才有可能提高性能。他们进一步表明，观察到的增益可能是由于与训练信号无关的影响，而是由于分布曲线形状的变化。这些结果可能表明 RL 的缺点之一是异常大的动作空间，在 TG 中包括词汇表中的所有标记，通常是成千上万的动作或更多
据我们所知，以前的工作都没有考虑到 TG 中大型动作空间的挑战，而在不同情况下考虑它的研究相对较少。一行工作假定有关问题的
内部低维表示映射回词汇表大小，操作。我们 先验领域知识，并将操作划分为子组 (Sharma
et al., 2017) , 或类似于我们的方法，将操作嵌入连续空间，其中该空间上的某些度量允许对类似操作 (Dulac-Arnold et al., 2016)进行泛化。当动作空间的基础结构先验未知时，建议使用专家演示 (Tennenholtz and Mannor, 2019; Chandak et al., 2019) 来学习目标嵌人的最新工作
本文确定了较大的动作空间是 NMTRL 应用的限制因素，并提出了应对这一挑战的方法。我们的技术通过使用底层连续表示来显式或隐式地限制嵌入空间的大小。
2.2 $\textbf{Technical Background and Notatior}$
$\textbf{Notation. 我 们 用 }X= ( x_1, . . . , x_S)$表示源句子，用$Y=(y_1,...,y_T)$ 表示参考句子。给定 $X$ ,网络以目标语言$Y^\prime=(y_1^{\prime},...,y_M^{\prime})$生成一个句子。目标标记取自词汇表$V_T$。在推理过程中，在每个步骤$i$中，生成令牌$y_i^\prime\in V_T$的概率取决于句子和预测的令牌，i。e。,$P_\theta(y_i^{\prime}|X,y_{<i}^{\prime})$ ,其中$\theta$是模型参数。我们假设只有一个有效的目标令牌，即引用令牌，因为在实践中，训练是针对单个引用 (Schulz et al.,2018)完成的$\mathbf{NMTwithRL.}$在RL 术语中，可以将 NMT 模型视为与环境交互的代理。在这种情况下，
阿 自 足 10 求 完 八 及 门 的 对 实 现 以 , 也 已 通 过 环 境 状 态 由 前 面 的 单 词 $y_{< i}^{\prime }$和 源 句 $X$ 组 成 。在 非 RL 方 法 ( e。g。, Zhang et al, 2019; ? ) 。Inaddi- 环 境 状 态 由 前 面 的 单 词 $y_{< i}^{\prime }$和 源 句 $X$ 组 成 。在 tion, RLisgroundedwithinabroadtheoreticalandempirica$\text{\^{a} r\'{e} arrchichideradebresap略 p选 择 一 个 操 作 , }$ 主 中 握 化 是 今 牌 第 略 由 横 邦 的 参 数
其中操作是令牌。策略由模型的参数定义，即。e。,条件概率$P_\theta(y_i^{\prime}|y_{<i}^{\prime},X)$ 。仅在代理生成完整序列$Y^{\prime}$时给予奖励。MT 的标准奖励是句子级别 BLEU 度量 (Papineni et al.,2002), 与评估度量匹配。我们的目标是找到将最大化预期奖励的参数。在这项工作中，我们使用 MRT (Och,2003; Shen et al.,2015) , 这是一种适用于MT 的策略梯度方法。该方法的关键思想是在每个步骤中优化重新标准化的风险，仅在抽样批次中定义。具体而言，预期风险定义为：
$$L_{risk}=\sum_{u\in U(X)}R(Y,u)\frac{P(u|X)^\beta}{\sum_{u^{\prime}\in U(X)}P(u^{\prime}|X)^\beta}$$
(1)
其中$u$是候选假设句子，$U(x)$是$k$候选假设的样本，$Y$是参考$,P$ 是模型在给定源句子$X$的情况下分配候选假设$u$的条件概率，$\beta$是平滑参数，$R$是
### $\textbf{3}$ $\textbf{Methodology}$
Architecture. 我们使用与 Wieting et al. (2019) 类似的设置，使其基于 fairSeq 的 (Ott
|
[
[
0.12176513671875,
0.0872802734375,
0.48779296875,
0.0872802734375,
0.48779296875,
0.1009521484375,
0.12176513671875,
0.1009521484375
],
[
0.51123046875,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51123046875,
0.1009521484375
],
[
0.12005615234375,
0.10272216796875,
0.484375,
0.10272216796875,
0.484375,
0.11639404296875,
0.12005615234375,
0.11639404296875
],
[
0.51123046875,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51123046875,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.483642578125,
0.1181640625,
0.483642578125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51123046875,
0.1181640625,
0.87890625,
0.1181640625,
0.87890625,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.12005615234375,
0.1353759765625,
0.489501953125,
0.1353759765625,
0.489501953125,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.87890625,
0.1353759765625,
0.87890625,
0.1484375,
0.51318359375,
0.1484375
],
[
0.118408203125,
0.15087890625,
0.484375,
0.15087890625,
0.484375,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.5146484375,
0.152587890625,
0.87890625,
0.152587890625,
0.87890625,
0.163330078125,
0.5146484375,
0.163330078125
],
[
0.118408203125,
0.16748046875,
0.484375,
0.16748046875,
0.484375,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.880859375,
0.16748046875,
0.880859375,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.118408203125,
0.182861328125,
0.484375,
0.182861328125,
0.484375,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.51318359375,
0.182861328125,
0.87890625,
0.182861328125,
0.87890625,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.118408203125,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.118408203125,
0.2120361328125
],
[
0.5146484375,
0.20068359375,
0.5322265625,
0.20068359375,
0.5322265625,
0.210205078125,
0.5146484375,
0.210205078125
],
[
0.12005615234375,
0.215576171875,
0.486083984375,
0.215576171875,
0.486083984375,
0.2286376953125,
0.12005615234375,
0.2286376953125
],
[
0.52880859375,
0.21728515625,
0.87890625,
0.21728515625,
0.87890625,
0.23095703125,
0.52880859375,
0.23095703125
],
[
0.118408203125,
0.2298583984375,
0.484375,
0.23095703125,
0.484375,
0.2452392578125,
0.118408203125,
0.24462890625
],
[
0.51123046875,
0.2327880859375,
0.87890625,
0.2327880859375,
0.87890625,
0.2464599609375,
0.51123046875,
0.2464599609375
],
[
0.118408203125,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.51123046875,
0.248779296875,
0.880859375,
0.248779296875,
0.880859375,
0.262451171875,
0.51123046875,
0.262451171875
],
[
0.118408203125,
0.26416015625,
0.484375,
0.26416015625,
0.484375,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.51123046875,
0.265380859375,
0.72314453125,
0.265380859375,
0.72314453125,
0.279052734375,
0.51123046875,
0.279052734375
],
[
0.12005615234375,
0.27978515625,
0.484375,
0.27978515625,
0.484375,
0.29345703125,
0.12005615234375,
0.29345703125
],
[
0.118408203125,
0.295654296875,
0.486083984375,
0.295654296875,
0.486083984375,
0.308837890625,
0.118408203125,
0.308837890625
],
[
0.51318359375,
0.29638671875,
0.84228515625,
0.29638671875,
0.84228515625,
0.31005859375,
0.51318359375,
0.31005859375
],
[
0.118408203125,
0.312255859375,
0.484375,
0.312255859375,
0.484375,
0.325927734375,
0.118408203125,
0.325927734375
],
[
0.515625,
0.318359375,
0.880859375,
0.320556640625,
0.880859375,
0.3349609375,
0.515625,
0.333251953125
],
[
0.118408203125,
0.327880859375,
0.484375,
0.327880859375,
0.484375,
0.341552734375,
0.118408203125,
0.341552734375
],
[
0.51318359375,
0.336181640625,
0.87890625,
0.336181640625,
0.87890625,
0.350830078125,
0.51318359375,
0.350830078125
],
[
0.118408203125,
0.342529296875,
0.486083984375,
0.342529296875,
0.486083984375,
0.355712890625,
0.118408203125,
0.355712890625
],
[
0.5283203125,
0.3515625,
0.880859375,
0.352783203125,
0.880859375,
0.36767578125,
0.5283203125,
0.366455078125
],
[
0.118408203125,
0.35986328125,
0.48779296875,
0.35986328125,
0.48779296875,
0.37353515625,
0.118408203125,
0.37353515625
],
[
0.5205078125,
0.358154296875,
0.52880859375,
0.358154296875,
0.52880859375,
0.3623046875,
0.5205078125,
0.3623046875
],
[
0.1175537109375,
0.374755859375,
0.48779296875,
0.377197265625,
0.48779296875,
0.39013671875,
0.1175537109375,
0.388427734375
],
[
0.51318359375,
0.369384765625,
0.880859375,
0.369384765625,
0.880859375,
0.383056640625,
0.51318359375,
0.383056640625
],
[
0.51123046875,
0.38427734375,
0.880859375,
0.385498046875,
0.880859375,
0.399658203125,
0.51123046875,
0.3984375
],
[
0.11419677734375,
0.393798828125,
0.1351318359375,
0.393798828125,
0.1351318359375,
0.407958984375,
0.11419677734375,
0.407958984375
],
[
0.51123046875,
0.399658203125,
0.880859375,
0.402099609375,
0.880859375,
0.416259765625,
0.51123046875,
0.41455078125
],
[
0.11083984375,
0.418701171875,
0.26123046875,
0.42041015625,
0.26123046875,
0.438232421875,
0.11083984375,
0.435791015625
],
[
0.52880859375,
0.41748046875,
0.880859375,
0.41748046875,
0.880859375,
0.43115234375,
0.52880859375,
0.43115234375
],
[
0.5224609375,
0.42333984375,
0.52880859375,
0.42333984375,
0.52880859375,
0.427490234375,
0.5224609375,
0.427490234375
],
[
0.5146484375,
0.43408203125,
0.87890625,
0.43408203125,
0.87890625,
0.44775390625,
0.5146484375,
0.44775390625
],
[
0.11419677734375,
0.44775390625,
0.376220703125,
0.44775390625,
0.376220703125,
0.46142578125,
0.11419677734375,
0.46142578125
],
[
0.51318359375,
0.449462890625,
0.87744140625,
0.449462890625,
0.87744140625,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.1175537109375,
0.47021484375,
0.480224609375,
0.47021484375,
0.480224609375,
0.48095703125,
0.1175537109375,
0.48095703125
],
[
0.515625,
0.47802734375,
0.8818359375,
0.478515625,
0.8818359375,
0.493408203125,
0.515625,
0.4921875
],
[
0.118408203125,
0.485107421875,
0.484375,
0.485107421875,
0.484375,
0.498779296875,
0.118408203125,
0.498779296875
],
[
0.51318359375,
0.495361328125,
0.8818359375,
0.495361328125,
0.8818359375,
0.5087890625,
0.51318359375,
0.5087890625
],
[
0.118408203125,
0.501953125,
0.48779296875,
0.501953125,
0.48779296875,
0.515625,
0.118408203125,
0.515625
],
[
0.11419677734375,
0.5166015625,
0.491943359375,
0.515625,
0.491943359375,
0.533203125,
0.11419677734375,
0.53369140625
],
[
0.51318359375,
0.5126953125,
0.8818359375,
0.5126953125,
0.8818359375,
0.525390625,
0.51318359375,
0.525390625
],
[
0.11248779296875,
0.53076171875,
0.884765625,
0.52392578125,
0.884765625,
0.5458984375,
0.11248779296875,
0.552734375
],
[
0.1351318359375,
0.548828125,
0.48779296875,
0.5498046875,
0.48779296875,
0.5634765625,
0.1351318359375,
0.5625
],
[
0.5087890625,
0.54443359375,
0.8818359375,
0.54345703125,
0.8818359375,
0.55810546875,
0.5087890625,
0.55859375
],
[
0.118408203125,
0.56640625,
0.489501953125,
0.56640625,
0.489501953125,
0.580078125,
0.118408203125,
0.580078125
],
[
0.51318359375,
0.56103515625,
0.880859375,
0.56103515625,
0.880859375,
0.57470703125,
0.51318359375,
0.57470703125
],
[
0.51123046875,
0.57666015625,
0.880859375,
0.57666015625,
0.880859375,
0.59033203125,
0.51123046875,
0.59033203125
],
[
0.118408203125,
0.5830078125,
0.48779296875,
0.5830078125,
0.48779296875,
0.5966796875,
0.118408203125,
0.5966796875
],
[
0.51123046875,
0.59130859375,
0.87890625,
0.59130859375,
0.87890625,
0.60498046875,
0.51123046875,
0.60498046875
],
[
0.118408203125,
0.5986328125,
0.484375,
0.5986328125,
0.484375,
0.6123046875,
0.118408203125,
0.6123046875
],
[
0.118408203125,
0.6142578125,
0.484375,
0.6142578125,
0.484375,
0.62890625,
0.118408203125,
0.62890625
],
[
0.51318359375,
0.609375,
0.880859375,
0.609375,
0.880859375,
0.62255859375,
0.51318359375,
0.62255859375
],
[
0.51318359375,
0.62451171875,
0.880859375,
0.62451171875,
0.880859375,
0.63818359375,
0.51318359375,
0.63818359375
],
[
0.118408203125,
0.63134765625,
0.486083984375,
0.63134765625,
0.486083984375,
0.64453125,
0.118408203125,
0.64453125
],
[
0.12005615234375,
0.646484375,
0.486083984375,
0.646484375,
0.486083984375,
0.66015625,
0.12005615234375,
0.66015625
],
[
0.51123046875,
0.64111328125,
0.880859375,
0.64111328125,
0.880859375,
0.65478515625,
0.51123046875,
0.65478515625
],
[
0.118408203125,
0.6611328125,
0.484375,
0.6611328125,
0.484375,
0.6748046875,
0.118408203125,
0.6748046875
],
[
0.51123046875,
0.65673828125,
0.880859375,
0.65673828125,
0.880859375,
0.67041015625,
0.51123046875,
0.67041015625
],
[
0.1175537109375,
0.677734375,
0.483642578125,
0.677734375,
0.483642578125,
0.69140625,
0.1175537109375,
0.69140625
],
[
0.51318359375,
0.6728515625,
0.880859375,
0.6728515625,
0.880859375,
0.68603515625,
0.51318359375,
0.68603515625
],
[
0.1175537109375,
0.69482421875,
0.274658203125,
0.69482421875,
0.274658203125,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.51318359375,
0.689453125,
0.8486328125,
0.689453125,
0.8486328125,
0.703125,
0.51318359375,
0.703125
],
[
0.1351318359375,
0.71044921875,
0.484375,
0.71142578125,
0.484375,
0.7255859375,
0.1351318359375,
0.724609375
],
[
0.1175537109375,
0.7255859375,
0.484375,
0.72705078125,
0.484375,
0.7412109375,
0.1175537109375,
0.74072265625
],
[
0.75,
0.7216796875,
0.8193359375,
0.7197265625,
0.8212890625,
0.7353515625,
0.75146484375,
0.73681640625
],
[
0.52490234375,
0.72802734375,
0.71728515625,
0.72802734375,
0.71728515625,
0.7451171875,
0.52490234375,
0.7451171875
],
[
0.1175537109375,
0.7421875,
0.483642578125,
0.7421875,
0.483642578125,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.705078125,
0.74072265625,
0.86376953125,
0.73681640625,
0.86376953125,
0.7548828125,
0.705078125,
0.75830078125
],
[
0.591796875,
0.75,
0.6455078125,
0.75,
0.6455078125,
0.76123046875,
0.591796875,
0.76123046875
],
[
0.118408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.85791015625,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.77490234375,
0.85791015625,
0.77490234375
],
[
0.1175537109375,
0.77685546875,
0.48779296875,
0.77685546875,
0.48779296875,
0.79052734375,
0.1175537109375,
0.79052734375
],
[
0.53076171875,
0.779296875,
0.880859375,
0.779296875,
0.880859375,
0.7919921875,
0.53076171875,
0.7919921875
],
[
0.1234130859375,
0.7919921875,
0.48779296875,
0.7919921875,
0.48779296875,
0.8056640625,
0.1234130859375,
0.8056640625
],
[
0.51123046875,
0.79345703125,
0.87890625,
0.79345703125,
0.87890625,
0.80712890625,
0.51123046875,
0.80712890625
],
[
0.118408203125,
0.81005859375,
0.48193359375,
0.81005859375,
0.48193359375,
0.82080078125,
0.118408203125,
0.82080078125
],
[
0.5146484375,
0.8125,
0.87890625,
0.8125,
0.87890625,
0.822265625,
0.5146484375,
0.822265625
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51318359375,
0.82666015625,
0.64208984375,
0.82666015625,
0.64208984375,
0.84033203125,
0.51318359375,
0.84033203125
],
[
0.118408203125,
0.84033203125,
0.483642578125,
0.84033203125,
0.483642578125,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.118408203125,
0.85693359375,
0.4130859375,
0.85693359375,
0.4130859375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.5087890625,
0.85498046875,
0.66162109375,
0.859375,
0.66162109375,
0.876953125,
0.5087890625,
0.87255859375
],
[
0.1376953125,
0.87451171875,
0.484375,
0.87451171875,
0.484375,
0.88525390625,
0.1376953125,
0.88525390625
],
[
0.12005615234375,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.12005615234375,
0.90283203125
],
[
0.5224609375,
0.88720703125,
0.6533203125,
0.88818359375,
0.6533203125,
0.90283203125,
0.5224609375,
0.90185546875
],
[
0.64404296875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.64404296875,
0.90283203125
],
[
0.12005615234375,
0.90673828125,
0.486083984375,
0.90673828125,
0.486083984375,
0.91748046875,
0.12005615234375,
0.91748046875
],
[
0.509765625,
0.90380859375,
0.880859375,
0.90478515625,
0.880859375,
0.91943359375,
0.509765625,
0.91845703125
]
] |
[
"内部低维表示映射回词汇表大小，操作。我们",
"先验领域知识，并将操作划分为子组（Sharma",
"使用这个术语来强调网络的第一个嵌入层之间",
"etal.，2017），或类似于我们的方法，将操作",
"的类比，将维度[VT|的向量映射到维度d的向",
"嵌入连续空间，其中该空间上的某些度量允许",
"量，以及以相反方式工作的目标嵌入。事实上，",
"对类似操作（Dulac-Arnoldetal.，2016）进行泛",
"(e。g。，在BERT，？）中，源嵌人和目标嵌入",
"化。当动作空间的基础结构先验未知时，建议",
"的权重在训练过程中是共享的，强调了两者之",
"使用专家演示（TennenholtzandMannor，2019;",
"间的关系。使用这个术语，我们在模拟（85.1",
"Chandaketal.,2019）来学习目标嵌人的最新工",
"）中表明，当类似的动作共享目标嵌人时，RL",
"作",
"更有效此外，当基于高质量嵌入（在我们的例",
"本文确定了较大的动作空间是NMTRL应用",
"子中是BERT）初始化目标嵌入时，在RL期间",
"的限制因素，并提出了应对这一挑战的方法。",
"冻结它们仍然会产生进一步的改进。在NMT",
"我们的技术通过使用底层连续表示来显式或隐",
"上进行实验时，我们获得了类似的结果。实际",
"式地限制嵌人空间的大小。",
"上，将BERT的嵌入用于目标嵌人可以提高四",
"个语言对的性能，并且冻结它们会对MLE和",
"2.2Technical Background and Notation",
"RL产生额外的改进，如自动指标和人工评估",
"Notation。我们用X=（ci，,s）表示源句",
"所报告的。在NMT的RL训练中，初始化和",
"子，用Y=（y1，..,UT）表示参考句子。给定X",
"冻结都是新颖的。此外，当使用BERT的嵌入",
"网络以目标语言Y\"=（yi,yM）生成一个",
"时，RL提高预训练的MLE模型没有分配高概",
"",
"率的目标令牌的性能的能力得到了增强（85.2",
"句子。目标标记取自词汇表VT。在推理过程",
"中，在每个步骤i中，生成令牌y,EVT的概率",
")。",
"取决于句子和预测的令牌，i。e。，Pe(yiX,y<i）",
"2Background",
"其中θ是模型参数。我们假设只有一个有效",
"",
"的目标令牌，即引用令牌，因为在实践中，训",
"2.1RL in Machine Translation",
"练是针对单个引用（Schulzetal.,2018）完成的",
"RL用于文本生成（TG），因为它能够合并不可",
"NMTwithRL．在RL术语中，可以将NMT",
"微分信号，解决曝光偏差并引人序列级约束。后",
"模型视为与环境交互的代理。在这种情况下，",
"两者是TG系统开发中的持续挑战，也已通过",
"非RL方法(e。g。,Zhang et al.,2019;?)。Inaddi-",
"环境状态由前面的单词y，和源句X组成。在",
"tion,RLisgroundedwithinabroadtheoreticalandempiric舞企骤出hic绒竭据基第选择一个操作，",
"这些特性引起了对一般（Shahetal.,2018)和",
"其中操作是令牌。策略由模型的参数定义，即。",
"NMT尤其是（Wuetal.,2018a）中的TG的RL",
"e。，条件概率Pe(yily<i，X）。仅在代理生成完",
"整序列Y'时给予奖励。MT的标准奖励是句",
"的极大兴趣。通常使用许多策略梯度方法，特",
"子级别BLEU度量(Papinenietal.，2002），与",
"别是REINFORCE（Williams，1992）和最小风险",
"培训(MRT;e。g。，Och,2003；？）。然而，尽",
"评估度量匹配。我们的目标是找到将最大化预",
"期奖励的参数。在这项工作中，我们使用MRT",
"管人们的兴趣越来越大，结果也越来越好，但",
"只有少数作品研究了NLP中RL观察到的性",
"(Och,2003;Shenetal.,2015），这是一种适用于",
"能提高的来源及其训练动力学，其中一些表明",
"MT的策略梯度方法。该方法的关键思想是在",
"RL的提高部分归因于工件（Cacciaetal.,2018",
"每个步骤中优化重新标准化的风险，仅在抽样",
"Choshen et al., 2019)",
"批次中定义。具体而言，预期风险定义为",
"在最近的一篇论文中，C19表明，现有的",
"MTRL训练协议（REINFORCE和MRT）需要",
"P(u|X)B",
"Lrisk = R(Y,u)",
"很长的时间才能收敛。他们的结果表明，只有",
"Zu'eu(x) P(u|X)β",
"EU(X)",
"在MLE参数已经接近产生正确翻译的情况下，",
"(1)",
"MT中的RL实践才有可能提高性能。他们进",
"其中u是候选假设句子，U(α）是k候选假",
"一步表明，观察到的增益可能是由于与训练",
"设的样本，Y是参考，P是模型在给定源句子",
"信号无关的影响，而是由于分布曲线形状的变",
"X的情况下分配候选假设u的条件概率，βB是",
"化。这些结果可能表明RL的缺点之一是异常",
"平滑参数，R是",
"大的动作空间，在TG中包括词汇表中的所有",
"标记，通常是成千上万的动作或更多",
"3Methodology",
"据我们所知，以前的工作都没有考虑到TG",
"中大型动作空间的挑战，而在不同情况下考虑",
"Architecture.手",
"我们使用与Wietingetal.",
"它的研究相对较少。一行工作假定有关问题的",
"(2019类似的设置，使其基于fairSeq的（Ott"
] |
{
"arxiv_id": "2210.03053",
"language": "zh",
"page_idx": 2
}
|
||
2210.03053-zh-3
|
2210.03053-zh-3.jpg
|
<table>
<tbody>
<tr>
<th>$\mathbf{Model}$</th>
<th>DE- EN</th>
<th>CS- EN</th>
<th>RU- EN</th>
<th>TR- EN</th>
</tr>
<tr>
<td>$LTV$</td>
<td>25.07</td>
<td>15.16</td>
<td>16.67</td>
<td>12.76</td>
</tr>
<tr>
<td>LTV+ RL</td>
<td>25.67</td>
<td>15.33</td>
<td>16.9</td>
<td>12.98</td>
</tr>
<tr>
<td>$Diff.$</td>
<td>0.6</td>
<td>0.17</td>
<td>0.23</td>
<td>0.22</td>
</tr>
<tr>
<td>$STV$</td>
<td>21.83</td>
<td>13.79</td>
<td>14.63</td>
<td>10.37</td>
</tr>
<tr>
<td>STV+ RL</td>
<td>23.23</td>
<td>14.62</td>
<td>15.73</td>
<td>11.96</td>
</tr>
<tr>
<td>$Diff.$</td>
<td>1.4</td>
<td>0.83</td>
<td>1.1</td>
<td>1.59</td>
</tr>
</tbody>
</table>
Table 1: BLEU scores for translating four languages to English using MLE pretraining followed by RL, and comparing a model with a large vocabulary (LTV) to a small one (STV). The top (bottom) block presents results for LTV (STV) with and without RL, and the difference between them (Diff.). RL with STV gains more than 1 point more (on average) over the pre-trained MLE model, than RL with LTV.
et al., 2019) 代码库适应我们的目的。$^{2}$与其Transformer 架构类似，我们使用门控卷积编码器和解码器 (Gehring et al.,2017)。我们为编码器使用 4 层，为解码器使用 3 层，所有层的隐藏状态大小为 768, 内核的过滤器宽度为 3。此外，BPE 嵌人的维度设置为 768
Data Prepossessing. 我们使用 BPE (Sennrich et al., 2016) 进行标记化。Wieting et al. (2019) 将组合的源和目标词汇表的词汇表大小设置为$40\mathbf{K}$。对于小目标词汇实验，我们将目标词汇大小更改为 1K, 并保持源词汇不变
Objective Functions. 在 Edunov et al.(2018) 之后，我们使用带有标签平滑大小为 0 的(Szegedy et al., 2016; Pereyra et al., 2017)的MLE 训练模型。1。对于 RL,我们使用 MRT $L_{risk}$和令牌级别损失$L_mle$的加权平均值对模型进行微调
因此，我们的微调目标变为：
$$L_{Average}=\alpha\cdot L_{mle}+(1-\alpha)\cdot L_{risk}$$
(2)
我们将$\alpha$设置为0。3 通过 Wu et al.(2018b) 显示效果最佳。我们将$\beta$设置为 1。我们使用波束搜索为每个 MRT 步骤(k=8)生成八个假设。我们使用 Moses 实现的平滑 BLEU (Lin and Och, 2004)进行训练。$^{3}$此外，我们使用此指标报告结果并验证它们是否与 sacrebleu (Post,2018)匹配。“
$\mathbf{O}$ptimization. 我们训练超过 200 个时期的MLE 目标和超过 15 个时期的组合 RL 目标。我们通过选择验证损失最小的模型来执行提前停止。我们使用 Nesterov 的加速梯度法 (Sutskever et al.,2013) 进行优化，学习率为0。25, 动量为 0。99, 并将梯度重新归一化为0。1norm (Pascanu et al., 2012)
Data. 我们使用四种语言进行实验：德语(De),捷克语(Cs),俄语(Ru)和土耳其语 (Tr),将每种语言翻译成英语(En)。对于 cs-en、de-en 和 ru-en 的训练数据，我们使用 WMT 新闻评论 v13$^{5}$(Bojar et al., 2017)。对于 tr-en 训练数据，我们使用 WMT2018 并行数据，该数据由 SETIMES2 语料库 (Tiede-) mann, 2012) 组成。验证集是为 WMT18 发布的 newsdev2016 和 2017 的串联。测试集是官
$^2$https://github.com/jwieting/
beyond-bleu
$^{3}$https://github.com/jwieting/
beyond-bleu/blob/master/multi-bleu.perl
$^{4}$https://github.com/mjpost/sacrebleu
$^{5}$http://data.statmt.org/wmt18/
translation-task
方的 WMT18 测试集。这些实验侧重于低资源环境。我们选择此设置，因为 RL 实验在计算上要求很高，并且此设置在像我们的 Wieting et al.(2019) 这样的 RL 实验的文献中很常见。(请参阅 Supp 中的数据统计。§A )
### 4 $\textbf{Reducing the Vocabularv Size}$
我们首先直接测试我们的假设，即动作空间的大小是 NMTRL 收敛时间长的原因。为此，我们使用目标侧 BPE 训练模型，该模型取自比通常使用的词汇表小得多的词汇表
我们首先训练两个 MLE 模型，一个具有大的 (17K-31K) 目标词汇量 (LTV),另一个具有大小为 1K 的目标词汇量 (STV)。源词汇表保持不变。我们从 MLE 预训练开始，然后用RL 训练两个模型中的每一个
结果(表1)表明，使用 STV 的 RL 训练比使用 LTV 的 RL 训练获得的 1 BLEU 点更多。$^{6}$ 有关模型熵的比较，请参见 Supp。§B 。为了验证改进不是源于混合 RL 和 MLE 的$\alpha$的选择(见方程 2),我们用$\alpha\in\{0,1\}$重复对 De-En 的训练，我们发现$\alpha=0.3$优于两者。此外，当仅使用 RL 目标进行训练时，RL 比 LTV 提高了 STV $(\alpha=1$ )。这表明 RL 训练有助于观察到的改进。
接下来，我们将分析哪些令牌负责观察到的性能增益。具体来说，我们检查减少词汇量是否导致 RL 能够提升通过预训练的 MLE 模型获得低排名的目标令牌。对于每个模型，对于700K 试验，我们计算模型为上下文$y_<i^{\prime}$和源句$X$分配给黄金令牌$y_i$的等级。正式，$\forall r\in|V_T|$
$^{6}$Preliminary experiments showed that altering the ran-
dom seed changes the BLEU score by $\pm0.01$ points.
|
[
[
0.118408203125,
0.08489990234375,
0.48779296875,
0.08489990234375,
0.48779296875,
0.09979248046875,
0.118408203125,
0.09979248046875
],
[
0.51123046875,
0.0885009765625,
0.5634765625,
0.0885009765625,
0.5634765625,
0.10272216796875,
0.51123046875,
0.10272216796875
],
[
0.60107421875,
0.09027099609375,
0.66015625,
0.09027099609375,
0.66015625,
0.1015625,
0.60107421875,
0.1015625
],
[
0.67529296875,
0.09027099609375,
0.7314453125,
0.09027099609375,
0.7314453125,
0.1015625,
0.67529296875,
0.1015625
],
[
0.75,
0.09027099609375,
0.8076171875,
0.09027099609375,
0.8076171875,
0.1015625,
0.75,
0.1015625
],
[
0.82373046875,
0.09027099609375,
0.8818359375,
0.09027099609375,
0.8818359375,
0.1015625,
0.82373046875,
0.1015625
],
[
0.1175537109375,
0.1015625,
0.484375,
0.1015625,
0.484375,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.509765625,
0.10986328125,
0.54833984375,
0.10986328125,
0.54833984375,
0.1258544921875,
0.509765625,
0.1258544921875
],
[
0.61279296875,
0.10986328125,
0.6591796875,
0.10986328125,
0.6591796875,
0.1258544921875,
0.61279296875,
0.1258544921875
],
[
0.68701171875,
0.10986328125,
0.73291015625,
0.10986328125,
0.73291015625,
0.1258544921875,
0.68701171875,
0.1258544921875
],
[
0.76318359375,
0.10986328125,
0.8076171875,
0.10986328125,
0.8076171875,
0.1258544921875,
0.76318359375,
0.1258544921875
],
[
0.84814453125,
0.10986328125,
0.892578125,
0.10986328125,
0.892578125,
0.1258544921875,
0.84814453125,
0.1258544921875
],
[
0.12005615234375,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.51123046875,
0.1270751953125,
0.58251953125,
0.1270751953125,
0.58251953125,
0.141357421875,
0.51123046875,
0.141357421875
],
[
0.61279296875,
0.1270751953125,
0.66015625,
0.1270751953125,
0.66015625,
0.141357421875,
0.61279296875,
0.141357421875
],
[
0.68701171875,
0.1270751953125,
0.73291015625,
0.1270751953125,
0.73291015625,
0.141357421875,
0.68701171875,
0.141357421875
],
[
0.771484375,
0.127685546875,
0.8076171875,
0.127685546875,
0.8076171875,
0.14013671875,
0.771484375,
0.14013671875
],
[
0.84814453125,
0.1270751953125,
0.892578125,
0.1270751953125,
0.892578125,
0.142578125,
0.84814453125,
0.142578125
],
[
0.12005615234375,
0.135986328125,
0.483642578125,
0.135986328125,
0.483642578125,
0.146728515625,
0.12005615234375,
0.146728515625
],
[
0.509765625,
0.142578125,
0.54833984375,
0.142578125,
0.54833984375,
0.157958984375,
0.509765625,
0.157958984375
],
[
0.69287109375,
0.142578125,
0.73291015625,
0.142578125,
0.73291015625,
0.157958984375,
0.69287109375,
0.157958984375
],
[
0.77099609375,
0.142578125,
0.8095703125,
0.142578125,
0.8095703125,
0.157958984375,
0.77099609375,
0.157958984375
],
[
0.85400390625,
0.14013671875,
0.892578125,
0.142578125,
0.890625,
0.1591796875,
0.8505859375,
0.15673828125
],
[
0.118408203125,
0.15087890625,
0.484375,
0.15087890625,
0.484375,
0.1644287109375,
0.118408203125,
0.1644287109375
],
[
0.63037109375,
0.1436767578125,
0.66162109375,
0.1436767578125,
0.66162109375,
0.157958984375,
0.63037109375,
0.157958984375
],
[
0.51123046875,
0.1650390625,
0.5498046875,
0.1650390625,
0.5498046875,
0.18115234375,
0.51123046875,
0.18115234375
],
[
0.61279296875,
0.1650390625,
0.66015625,
0.1650390625,
0.66015625,
0.18115234375,
0.61279296875,
0.18115234375
],
[
0.68701171875,
0.1650390625,
0.7314453125,
0.1650390625,
0.7314453125,
0.18115234375,
0.68701171875,
0.18115234375
],
[
0.84814453125,
0.1650390625,
0.890625,
0.1650390625,
0.890625,
0.18115234375,
0.84814453125,
0.18115234375
],
[
0.76318359375,
0.166259765625,
0.8076171875,
0.166259765625,
0.8076171875,
0.18115234375,
0.76318359375,
0.18115234375
],
[
0.118408203125,
0.166259765625,
0.372802734375,
0.166259765625,
0.372802734375,
0.179931640625,
0.118408203125,
0.179931640625
],
[
0.61279296875,
0.18115234375,
0.6591796875,
0.18115234375,
0.6591796875,
0.196533203125,
0.61279296875,
0.196533203125
],
[
0.68701171875,
0.1817626953125,
0.73291015625,
0.1817626953125,
0.73291015625,
0.196533203125,
0.68701171875,
0.196533203125
],
[
0.76318359375,
0.18115234375,
0.8076171875,
0.18115234375,
0.8076171875,
0.196533203125,
0.76318359375,
0.196533203125
],
[
0.84619140625,
0.1817626953125,
0.892578125,
0.1817626953125,
0.892578125,
0.196533203125,
0.84619140625,
0.196533203125
],
[
0.1234130859375,
0.191162109375,
0.484375,
0.191162109375,
0.484375,
0.2060546875,
0.1234130859375,
0.2060546875
],
[
0.51318359375,
0.182861328125,
0.58251953125,
0.182861328125,
0.58251953125,
0.19482421875,
0.51318359375,
0.19482421875
],
[
0.509765625,
0.19775390625,
0.5498046875,
0.19775390625,
0.5498046875,
0.213134765625,
0.509765625,
0.213134765625
],
[
0.6337890625,
0.19775390625,
0.66015625,
0.19775390625,
0.66015625,
0.21142578125,
0.6337890625,
0.21142578125
],
[
0.6943359375,
0.198974609375,
0.7314453125,
0.198974609375,
0.7314453125,
0.21142578125,
0.6943359375,
0.21142578125
],
[
0.78076171875,
0.19775390625,
0.8076171875,
0.19775390625,
0.8076171875,
0.21142578125,
0.78076171875,
0.21142578125
],
[
0.85546875,
0.198974609375,
0.890625,
0.198974609375,
0.890625,
0.21142578125,
0.85546875,
0.21142578125
],
[
0.1175537109375,
0.207275390625,
0.484375,
0.2060546875,
0.484375,
0.2203369140625,
0.1175537109375,
0.221435546875
],
[
0.118408203125,
0.223876953125,
0.484375,
0.223876953125,
0.484375,
0.2369384765625,
0.118408203125,
0.2369384765625
],
[
0.509765625,
0.226806640625,
0.8818359375,
0.2286376953125,
0.8818359375,
0.2423095703125,
0.509765625,
0.240478515625
],
[
0.1175537109375,
0.240478515625,
0.484375,
0.240478515625,
0.484375,
0.254150390625,
0.1175537109375,
0.254150390625
],
[
0.51123046875,
0.24462890625,
0.87890625,
0.24462890625,
0.87890625,
0.254638671875,
0.51123046875,
0.254638671875
],
[
0.1175537109375,
0.255859375,
0.403076171875,
0.255859375,
0.403076171875,
0.26953125,
0.1175537109375,
0.26953125
],
[
0.51318359375,
0.257080078125,
0.880859375,
0.257080078125,
0.880859375,
0.270751953125,
0.51318359375,
0.270751953125
],
[
0.51123046875,
0.271484375,
0.8818359375,
0.271484375,
0.8818359375,
0.28515625,
0.51123046875,
0.28515625
],
[
0.124267578125,
0.28076171875,
0.486083984375,
0.28076171875,
0.486083984375,
0.29443359375,
0.124267578125,
0.29443359375
],
[
0.51123046875,
0.2861328125,
0.87890625,
0.2861328125,
0.87890625,
0.29638671875,
0.51123046875,
0.29638671875
],
[
0.118408203125,
0.297607421875,
0.48779296875,
0.297607421875,
0.48779296875,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.51123046875,
0.2998046875,
0.87890625,
0.300537109375,
0.87890625,
0.312255859375,
0.51123046875,
0.311279296875
],
[
0.118408203125,
0.3134765625,
0.486083984375,
0.3134765625,
0.486083984375,
0.32666015625,
0.118408203125,
0.32666015625
],
[
0.51123046875,
0.314208984375,
0.8818359375,
0.314208984375,
0.8818359375,
0.327880859375,
0.51123046875,
0.327880859375
],
[
0.1175537109375,
0.328857421875,
0.486083984375,
0.328857421875,
0.486083984375,
0.342529296875,
0.1175537109375,
0.342529296875
],
[
0.51318359375,
0.330078125,
0.73046875,
0.330078125,
0.73046875,
0.34033203125,
0.51318359375,
0.34033203125
],
[
0.118408203125,
0.34326171875,
0.486083984375,
0.344482421875,
0.486083984375,
0.359375,
0.118408203125,
0.358154296875
],
[
0.1175537109375,
0.359375,
0.210693359375,
0.361083984375,
0.210693359375,
0.3759765625,
0.1175537109375,
0.37353515625
],
[
0.5146484375,
0.369384765625,
0.87890625,
0.369384765625,
0.87890625,
0.383056640625,
0.5146484375,
0.383056640625
],
[
0.1368408203125,
0.377197265625,
0.364501953125,
0.377197265625,
0.364501953125,
0.391357421875,
0.1368408203125,
0.391357421875
],
[
0.51318359375,
0.385986328125,
0.880859375,
0.385986328125,
0.880859375,
0.399658203125,
0.51318359375,
0.399658203125
],
[
0.1485595703125,
0.407958984375,
0.443359375,
0.407958984375,
0.443359375,
0.4228515625,
0.1485595703125,
0.4228515625
],
[
0.5146484375,
0.402587890625,
0.880859375,
0.402587890625,
0.880859375,
0.416259765625,
0.5146484375,
0.416259765625
],
[
0.46337890625,
0.4091796875,
0.486083984375,
0.4091796875,
0.486083984375,
0.419189453125,
0.46337890625,
0.419189453125
],
[
0.51123046875,
0.418701171875,
0.880859375,
0.418701171875,
0.880859375,
0.431640625,
0.51123046875,
0.431640625
],
[
0.1368408203125,
0.435791015625,
0.484375,
0.435791015625,
0.484375,
0.449462890625,
0.1368408203125,
0.449462890625
],
[
0.51123046875,
0.435302734375,
0.7841796875,
0.435302734375,
0.7841796875,
0.448974609375,
0.51123046875,
0.448974609375
],
[
0.1175537109375,
0.451904296875,
0.483642578125,
0.451904296875,
0.483642578125,
0.465576171875,
0.1175537109375,
0.465576171875
],
[
0.12005615234375,
0.4697265625,
0.483642578125,
0.4697265625,
0.483642578125,
0.479736328125,
0.12005615234375,
0.479736328125
],
[
0.51123046875,
0.46435546875,
0.80615234375,
0.465576171875,
0.80615234375,
0.479736328125,
0.51123046875,
0.478515625
],
[
0.12005615234375,
0.48388671875,
0.486083984375,
0.48388671875,
0.486083984375,
0.49755859375,
0.12005615234375,
0.49755859375
],
[
0.51123046875,
0.493408203125,
0.87890625,
0.493408203125,
0.87890625,
0.50732421875,
0.51123046875,
0.50732421875
],
[
0.1175537109375,
0.50048828125,
0.484375,
0.50048828125,
0.484375,
0.51416015625,
0.1175537109375,
0.51416015625
],
[
0.5146484375,
0.5087890625,
0.8818359375,
0.5087890625,
0.8818359375,
0.5224609375,
0.5146484375,
0.5224609375
],
[
0.118408203125,
0.5166015625,
0.484375,
0.5166015625,
0.484375,
0.52978515625,
0.118408203125,
0.52978515625
],
[
0.5146484375,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.5390625,
0.5146484375,
0.5390625
],
[
0.1175537109375,
0.5322265625,
0.276123046875,
0.5322265625,
0.276123046875,
0.54638671875,
0.1175537109375,
0.54638671875
],
[
0.51318359375,
0.5419921875,
0.7900390625,
0.5419921875,
0.7900390625,
0.5556640625,
0.51318359375,
0.5556640625
],
[
0.1234130859375,
0.55712890625,
0.484375,
0.55712890625,
0.484375,
0.57080078125,
0.1234130859375,
0.57080078125
],
[
0.52880859375,
0.55810546875,
0.880859375,
0.55810546875,
0.880859375,
0.572265625,
0.52880859375,
0.572265625
],
[
0.1175537109375,
0.57373046875,
0.486083984375,
0.57373046875,
0.486083984375,
0.58740234375,
0.1175537109375,
0.58740234375
],
[
0.51123046875,
0.57373046875,
0.880859375,
0.57373046875,
0.880859375,
0.58837890625,
0.51123046875,
0.58837890625
],
[
0.118408203125,
0.59033203125,
0.486083984375,
0.59033203125,
0.486083984375,
0.60400390625,
0.118408203125,
0.60400390625
],
[
0.51123046875,
0.5888671875,
0.880859375,
0.5888671875,
0.880859375,
0.6025390625,
0.51123046875,
0.6025390625
],
[
0.118408203125,
0.60546875,
0.48779296875,
0.60546875,
0.48779296875,
0.619140625,
0.118408203125,
0.619140625
],
[
0.51318359375,
0.60693359375,
0.880859375,
0.60693359375,
0.880859375,
0.62060546875,
0.51318359375,
0.62060546875
],
[
0.1158447265625,
0.62060546875,
0.48779296875,
0.62158203125,
0.48779296875,
0.6357421875,
0.1158447265625,
0.63525390625
],
[
0.51123046875,
0.62255859375,
0.744140625,
0.62255859375,
0.744140625,
0.6357421875,
0.51123046875,
0.6357421875
],
[
0.118408203125,
0.63818359375,
0.48779296875,
0.63818359375,
0.48779296875,
0.65185546875,
0.118408203125,
0.65185546875
],
[
0.5283203125,
0.63818359375,
0.8818359375,
0.6396484375,
0.8818359375,
0.65380859375,
0.5283203125,
0.65283203125
],
[
0.1175537109375,
0.65380859375,
0.27783203125,
0.65380859375,
0.27783203125,
0.66748046875,
0.1175537109375,
0.66748046875
],
[
0.51123046875,
0.65380859375,
0.88330078125,
0.65185546875,
0.88330078125,
0.66943359375,
0.51123046875,
0.67138671875
],
[
0.51318359375,
0.6728515625,
0.880859375,
0.6728515625,
0.880859375,
0.68603515625,
0.51318359375,
0.68603515625
],
[
0.124267578125,
0.6787109375,
0.486083984375,
0.6787109375,
0.486083984375,
0.6923828125,
0.124267578125,
0.6923828125
],
[
0.51318359375,
0.6884765625,
0.8818359375,
0.6884765625,
0.8818359375,
0.70166015625,
0.51318359375,
0.70166015625
],
[
0.1175537109375,
0.69482421875,
0.486083984375,
0.69482421875,
0.486083984375,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.51123046875,
0.705078125,
0.8818359375,
0.705078125,
0.8818359375,
0.71875,
0.51123046875,
0.71875
],
[
0.118408203125,
0.71142578125,
0.484375,
0.71142578125,
0.484375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.51123046875,
0.7197265625,
0.87890625,
0.7197265625,
0.87890625,
0.73291015625,
0.51123046875,
0.73291015625
],
[
0.12005615234375,
0.72802734375,
0.486083984375,
0.72802734375,
0.486083984375,
0.7412109375,
0.12005615234375,
0.7412109375
],
[
0.51318359375,
0.7353515625,
0.880859375,
0.736328125,
0.880859375,
0.75048828125,
0.51318359375,
0.75
],
[
0.118408203125,
0.74365234375,
0.478515625,
0.74365234375,
0.478515625,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.5146484375,
0.75048828125,
0.880859375,
0.751953125,
0.880859375,
0.7666015625,
0.5146484375,
0.765625
],
[
0.1175537109375,
0.7587890625,
0.484375,
0.75830078125,
0.484375,
0.7724609375,
0.1175537109375,
0.77392578125
],
[
0.51318359375,
0.76708984375,
0.59423828125,
0.76708984375,
0.59423828125,
0.7822265625,
0.51318359375,
0.7822265625
],
[
0.118408203125,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.7890625,
0.118408203125,
0.7890625
],
[
0.52880859375,
0.78515625,
0.880859375,
0.78515625,
0.880859375,
0.7998046875,
0.52880859375,
0.7998046875
],
[
0.118408203125,
0.79150390625,
0.486083984375,
0.79150390625,
0.486083984375,
0.8046875,
0.118408203125,
0.8046875
],
[
0.51318359375,
0.8017578125,
0.880859375,
0.8017578125,
0.880859375,
0.8154296875,
0.51318359375,
0.8154296875
],
[
0.1175537109375,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.5146484375,
0.8193359375,
0.87744140625,
0.8193359375,
0.87744140625,
0.830078125,
0.5146484375,
0.830078125
],
[
0.1351318359375,
0.8291015625,
0.39794921875,
0.83056640625,
0.39794921875,
0.845703125,
0.1351318359375,
0.84326171875
],
[
0.5146484375,
0.83447265625,
0.87890625,
0.83447265625,
0.87890625,
0.84716796875,
0.5146484375,
0.84716796875
],
[
0.1175537109375,
0.84423828125,
0.219970703125,
0.84423828125,
0.219970703125,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.1376953125,
0.85400390625,
0.39794921875,
0.85546875,
0.39794921875,
0.87060546875,
0.1376953125,
0.8681640625
],
[
0.51123046875,
0.8486328125,
0.880859375,
0.849609375,
0.880859375,
0.8642578125,
0.51123046875,
0.86328125
],
[
0.1158447265625,
0.8681640625,
0.470947265625,
0.8681640625,
0.470947265625,
0.8818359375,
0.1158447265625,
0.8818359375
],
[
0.51123046875,
0.8662109375,
0.87744140625,
0.8662109375,
0.87744140625,
0.8798828125,
0.51123046875,
0.8798828125
],
[
0.1351318359375,
0.87744140625,
0.46337890625,
0.8798828125,
0.46337890625,
0.89453125,
0.1351318359375,
0.892578125
],
[
0.1368408203125,
0.89111328125,
0.4072265625,
0.8935546875,
0.4072265625,
0.90771484375,
0.1368408203125,
0.90625
],
[
0.5322265625,
0.89111328125,
0.8818359375,
0.8935546875,
0.8818359375,
0.90771484375,
0.5322265625,
0.90625
],
[
0.118408203125,
0.90673828125,
0.264404296875,
0.90673828125,
0.264404296875,
0.91748046875,
0.118408203125,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.82958984375,
0.90478515625,
0.82958984375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"etal.,2019）代码库适应我们的目的。2与其",
"Model",
"DE-EN",
"CS-EN",
"RU-EN",
"TR-EN",
"Transformer架构类似，我们使用门控卷积编码",
"LTV",
"25.07",
"15.16",
"16.67",
"12.76",
"器和解码器（Gehringetal.，2017）。我们为编码",
"LTV+RL",
"25.67",
"15.33",
"16.9",
"12.98",
"器使用4层，为解码器使用3层，所有层的隐",
"Diff.",
"0.17",
"0.23",
"0.22",
"藏状态大小为768，内核的过滤器宽度为3。此",
"0.6",
"STV",
"21.83",
"13.79",
"10.37",
"14.63",
"外，BPE嵌人的维度设置为768",
"23.23",
"14.62",
"15.73",
"11.96",
"DataPrepossessing.我们使用BPE（Sennrich",
"STV+RL",
"Diff.",
"1.4",
"0.83",
"1.1",
"1.59",
"et al.,2016）进行标记化。Wieting etal.（2019)",
"将组合的源和目标词汇表的词汇表大小设置为",
"Table 1:BLEU scores for translating four languages",
"40K。对于小目标词汇实验，我们将目标词汇",
"toEnglishusingMLEpretrainingfollowedbyRL,and",
"大小更改为1K，并保持源词汇不变",
"comparing a model with a large vocabulary (LTV) to a",
"small one (STV). The top (bottom) block presents re-",
"Objective Functions.在Edunov et al.(2018)",
"sultsforLTV (STV)with and withoutRL.and the dif",
"之后，我们使用带有标签平滑大小为0的",
"ferencebetweenthem(Diff.).RLwithSTVgainsmore",
"(Szegedy et al.,2016;Pereyra et al.,2017)的",
"than 1 point more (on average) over the pre-trained",
"MLE训练模型。1。对于RL，我们使用MRT",
"MLEmodel,thanRLwithLTV.",
"Lrish和令牌级别损失Lmle的加权平均值对模",
"型进行微调",
"方的WMT18测试集。这些实验侧重于低资源",
"因此，我们的微调目标变为：",
"环境。我们选择此设置，因为RL实验在计算",
"LAverage = α· Lmle + (1 - α) · Lrisk",
"上要求很高，并且此设置在像我们的Wieting",
"(2)",
"etal.（2019)这样的RL实验的文献中很常见。",
"我们将α设置为0。3通过Wuetal.(2018b)",
"(请参阅 Supp中的数据统计。sA）",
"显示效果最佳。我们将β设置为1。我们使用",
"波束搜索为每个MRT步骤（k=8）生成八个",
"4Reducing the Vocabulary Size",
"假设。我们使用Moses实现的平滑BLEU(Lin",
"我们首先直接测试我们的假设，即动作空间的",
"andOch，2004）进行训练。3此外，我们使用",
"大小是NMTRL收敛时间长的原因。为此，我",
"此指标报告结果并验证它们是否与sacrebleu",
"们使用目标侧BPE训练模型，该模型取自比",
"(Post,2018)匹配。4",
"通常使用的词汇表小得多的词汇表",
"Optimization。我们训练超过200个时期的",
"我们首先训练两个MLE模型，一个具有大",
"MLE目标和超过15个时期的组合RL目标。我",
"的（17K-31K）目标词汇量（LTV），另一个具",
"们通过选择验证损失最小的模型来执行提前停",
"有大小为1K的目标词汇量（STV）。源词汇表",
"止。我们使用Nesterov的加速梯度法（Sutskever",
"保持不变。我们从MLE预训练开始，然后用",
"etal.,2013）进行优化，学习率为0。25，动量",
"RL训练两个模型中的每一个",
"为0。99，并将梯度重新归一化为0。1norm",
"结果（表1）表明，使用STV的RL训练比",
"(Pascanu et al., 2012)",
"使用 LTV的 RL训练获得的1BLEU点更多。6",
"有关模型摘的比较，请参见Supp。SB。为了验",
"Data。我们使用四种语言进行实验：德语",
"证改进不是源于混合RL和MLE的α的选择",
"(De），捷克语（Cs），俄语（Ru）和土耳其",
"(见方程2），我们用αE[0,1)重复对De-En",
"语（Tr），将每种语言翻译成英语（En）。对",
"的训练，我们发现α=0.3优于两者。此外，当",
"于cs-en、de-en和ru-en的训练数据，我们使",
"仅使用RL目标进行训练时，RL比LTV提高",
"用WMT新闻评论v135(Bojaretal.,2017)",
"了STV（α=1）。这表明RL训练有助于观察",
"对于tr-en训练数据，我们使用WMT2018并",
"到的改进。",
"行数据，该数据由SETIMES2语料库（Tiede-",
"接下来，我们将分析哪些令牌负责观察到的",
"mann，2012）组成。验证集是为WMT18发布",
"性能增益。具体来说，我们检查减少词汇量是",
"的newsdev2016和2017的串联。测试集是官",
"否导致RL能够提升通过预训练的MLE模型",
"2https://github.com/jwieting/",
"获得低排名的目标令牌。对于每个模型，对于",
"beyond-bleu",
"\"https://github.com/jwieting/",
"700K试验，我们计算模型为上下文y和源句",
"beyond-bleu/blob/master/multi-bleu.perl",
"X分配给黄金令牌yi的等级。正式，VrE[VT",
"4https://github.com/mjpost/sacrebleu",
"Shttp://data.statmt.org/wmt18/",
"6preliminary experiments showed that altering the ran-",
"translation-task",
"dom seed changes the BLEU score by ±0.01 points."
] |
{
"arxiv_id": "2210.03053",
"language": "zh",
"page_idx": 3
}
|
||
2210.03053-zh-4
|
2210.03053-zh-4.jpg
|
<FigureHere>
Figure 1: Comparison of probability shift due to RL training of assign $y_{best}$ for ten first words for both LTV and STV. in blue, you can see the results with BPE of size 1,000, STV. in red are the results with BPE of size 30,000, LTV. a clear improvement of assigning $y_{best}$ in first place for STV.
$,P_{model}^{r}=\frac{\#\{gold~token~assigns~to~the~r~rank\}}{\#\{all~trials\}}$。然
后，我们通过减去这两个分布来比较 MLE 模型和 RL 模型的秩分布。在我们的表示法中， 对于每个秩 r, $\Delta P^r=P_{RL}^r-P_{MLE}^r$。此减法表示 RL 如何影响模型为每个等级分配正确令牌$y_i$的能力。RL 的积极作用越大，第一等级的概率为正的可能性就越大，而较低等级的概率为负(由于概率转移到了第一位)
图1显示了LTV 和 STV 的每秩概率差。我们可以看到，对于第一秩，由于使用 STV 进行 RL 训练而导致的概率偏移是使用 LTV 进行RL 训练所导致的偏移的两倍以上。因此，对于小词汇量设置，以下排名的概率偏移通常更负。该图表明，当我们使用较小的动作空间应用 RL 时，概率质量向较高位置的移动实际上发生得更多。此外，与LTV 相比，STV 训练能够将概率质量从较低的等级上移。对此的指示是，在前一百个等级中，STV 降低了它们的 83的概率，而仅 2 的 LTV。
### 5 Reducing the Effective Dimensionality of the Action Space
发现减少动作数量可以提高 RL 的性能，我们提出了一种减少有效动作数量的方法，而不会改变实际输出。词汇大小可能是静态的，如在预训练模型 (Devlin et al.,2019)中，减少它可能有助于 RL,但对于 MLE (Gowda and May, 2020) 是最佳的 , 或者引人域外单词 (Koehn and Knowles, 2017)。我们建议通过使用在类似上下文中出现的令牌上泛化的目标嵌入来做到这一点。我们探索了这个想法的两个实现，一个是我们用高质量的嵌入初始化目标嵌入，另一
个是我们在 RL 期间冻结学习的目标嵌人。我们还探索了两种方法的结合。冻结目标嵌入(解码器的最后一层)可以解释为训练网络以输出倒数第二层的激活，然后固定函数将其映射到词汇表的维度
我们注意到，尽管冻结是一种常见的过程(Zoph et al., 2016; Thompson et al., 2018; Lee et al., 2019; Coster et al., 2021) , 但据我们所知，它从未应用于将 RL 用于序列到序列模型。表示网络用$f_\theta$计算的函数。$f_\theta$可以写成$h_{\theta _2}\circ g_{\theta _1}$ , 其中$\theta = ( \theta _1, \theta _2)$ $、 g$将输入——源句$X$和模型翻译前缀$y_< i^{\prime }— — 映 射 到 \mathbb{R} ^d$ , $h$ 将$g$的输出映射到$\mathbb{R}^|V_T|$
使用此表示法，我们可以将方法表述为将预训练的 MLE 目标嵌入加载到$h_{\theta_2}$或冻结它(或两者)。至于许多编码器-解码器架构(包括 Transformer), 它认为$d\ll|V_T|$ ,这可以被认为是约束代理选择$d$维连续动作，其中$h_{\theta_2}$ 是由环境执行的已知转换。目标嵌入的重要性在于，它们允许更好地概括操作。直觉如下。假设两个令牌具有相同的嵌入和相似的语义， 即。e。, 它们适用于相同的上下文(同义词)。由于它们具有相同的目标嵌人，因此在训练期间，当遇到它们中的任何一个时，网络将执行相同的梯度更新，除了在最后一层(因为它们仍然被认为是不同的输出)。如果目标嵌入未冻结，则在训练过程中遇到其中任何一个都会导致非常相似的更新(因为它们具有相同的目标嵌人),但是它们的目标嵌入可能会稍微偏离，这将导致较低层的后续漂移。如果目标嵌人被冻结，它们将产生的梯度更新将保持不变并加快学习速度。我们假设在训练期间有类似的效果，其中具有相似(但不相同)嵌入的令牌和相似(但不相同)的分布将在训练中彼此受益。这促使我们探索信息初始化和参数冻结的组合。(请参阅 Supp 中的正式证明。§E )
### 5.1 Motivating Simulation through Policy Parameterization in Large Action Spaces
为了在受控设置中检查上面概述的直觉，我们考虑一个合成 RL 问题，其中动作空间被表面放大。该任务是一个(上下文)多臂强盗，具有$K$操作。在每个步骤中，从环境中采样输入状态(“上下文”; 从多元高斯分布采样的随机向量)。随机，固定，非线性二进制分类器根据给定的上下文确定动作 #1 或动作 #2 是否有回报(动作 3- $K$永远不会有回报),并且每个动作的奖励为$r+z$ ,其中$r=1$用于奖励动作， 0用于所有其他动作，以及$z\sim\mathcal{N}(0,0.1)$ 。至关重要的是，我们将每个操作$a$次复制，从而在策略级别上总共产生$K\times a$个操作-而对于
|
[
[
0.197265625,
0.09027099609375,
0.2359619140625,
0.08612060546875,
0.2392578125,
0.0968017578125,
0.20068359375,
0.1009521484375
],
[
0.51904296875,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51904296875,
0.1009521484375
],
[
0.197265625,
0.097412109375,
0.21240234375,
0.097412109375,
0.21240234375,
0.1134033203125,
0.197265625,
0.1134033203125
],
[
0.405517578125,
0.1051025390625,
0.44921875,
0.1051025390625,
0.44921875,
0.11639404296875,
0.405517578125,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.162109375,
0.1092529296875,
0.18896484375,
0.1092529296875,
0.18896484375,
0.1217041015625,
0.162109375,
0.1217041015625
],
[
0.1275634765625,
0.11517333984375,
0.1502685546875,
0.11517333984375,
0.1485595703125,
0.2244873046875,
0.1259765625,
0.2244873046875
],
[
0.3955078125,
0.1175537109375,
0.459228515625,
0.1175537109375,
0.459228515625,
0.12890625,
0.3955078125,
0.12890625
],
[
0.51123046875,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.133056640625,
0.51123046875,
0.133056640625
],
[
0.16796875,
0.12469482421875,
0.189697265625,
0.12469482421875,
0.189697265625,
0.1484375,
0.16796875,
0.1484375
],
[
0.405517578125,
0.1312255859375,
0.45751953125,
0.1312255859375,
0.45751953125,
0.141357421875,
0.405517578125,
0.141357421875
],
[
0.51318359375,
0.1353759765625,
0.8818359375,
0.1353759765625,
0.8818359375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.181396484375,
0.155029296875,
0.189697265625,
0.155029296875,
0.189697265625,
0.1644287109375,
0.181396484375,
0.1644287109375
],
[
0.5146484375,
0.152587890625,
0.6591796875,
0.152587890625,
0.6591796875,
0.163330078125,
0.5146484375,
0.163330078125
],
[
0.2418212890625,
0.1650390625,
0.2493896484375,
0.1650390625,
0.2493896484375,
0.169189453125,
0.2418212890625,
0.169189453125
],
[
0.246826171875,
0.160888671875,
0.28955078125,
0.157958984375,
0.291259765625,
0.169189453125,
0.2476806640625,
0.1728515625
],
[
0.225830078125,
0.174560546875,
0.2359619140625,
0.174560546875,
0.2359619140625,
0.1787109375,
0.225830078125,
0.1787109375
],
[
0.5322265625,
0.169189453125,
0.880859375,
0.169189453125,
0.880859375,
0.182861328125,
0.5322265625,
0.182861328125
],
[
0.160400390625,
0.17578125,
0.18896484375,
0.17578125,
0.18896484375,
0.188232421875,
0.160400390625,
0.188232421875
],
[
0.509765625,
0.18408203125,
0.87890625,
0.18408203125,
0.87890625,
0.19775390625,
0.509765625,
0.19775390625
],
[
0.2359619140625,
0.190673828125,
0.246826171875,
0.190673828125,
0.246826171875,
0.19482421875,
0.2359619140625,
0.19482421875
],
[
0.154541015625,
0.198974609375,
0.189697265625,
0.198974609375,
0.189697265625,
0.21142578125,
0.154541015625,
0.21142578125
],
[
0.5146484375,
0.203125,
0.87890625,
0.203125,
0.87890625,
0.213134765625,
0.5146484375,
0.213134765625
],
[
0.154541015625,
0.2203369140625,
0.189697265625,
0.22265625,
0.18896484375,
0.2340087890625,
0.15283203125,
0.232177734375
],
[
0.5146484375,
0.21728515625,
0.880859375,
0.21728515625,
0.880859375,
0.23095703125,
0.5146484375,
0.23095703125
],
[
0.20654296875,
0.236328125,
0.2225341796875,
0.236328125,
0.2225341796875,
0.2392578125,
0.20654296875,
0.2392578125
],
[
0.5146484375,
0.2340087890625,
0.880859375,
0.2340087890625,
0.880859375,
0.2476806640625,
0.5146484375,
0.2476806640625
],
[
0.2301025390625,
0.24462890625,
0.333251953125,
0.24462890625,
0.333251953125,
0.251708984375,
0.2301025390625,
0.251708984375
],
[
0.393798828125,
0.243408203125,
0.466796875,
0.243408203125,
0.466796875,
0.254150390625,
0.393798828125,
0.254150390625
],
[
0.33154296875,
0.2464599609375,
0.355224609375,
0.2464599609375,
0.355224609375,
0.25048828125,
0.33154296875,
0.25048828125
],
[
0.208251953125,
0.2464599609375,
0.231689453125,
0.2464599609375,
0.231689453125,
0.25048828125,
0.208251953125,
0.25048828125
],
[
0.361083984375,
0.2476806640625,
0.36865234375,
0.2476806640625,
0.36865234375,
0.25048828125,
0.361083984375,
0.25048828125
],
[
0.51123046875,
0.2493896484375,
0.880859375,
0.248779296875,
0.880859375,
0.26318359375,
0.51123046875,
0.26416015625
],
[
0.5146484375,
0.265380859375,
0.880859375,
0.265380859375,
0.880859375,
0.27978515625,
0.5146484375,
0.27978515625
],
[
0.198974609375,
0.2529296875,
0.466796875,
0.254150390625,
0.466796875,
0.267333984375,
0.198974609375,
0.2666015625
],
[
0.1175537109375,
0.283935546875,
0.48779296875,
0.283935546875,
0.48779296875,
0.297607421875,
0.1175537109375,
0.297607421875
],
[
0.509765625,
0.28076171875,
0.70947265625,
0.276611328125,
0.70947265625,
0.29443359375,
0.509765625,
0.298583984375
],
[
0.1158447265625,
0.297607421875,
0.486083984375,
0.29638671875,
0.486083984375,
0.311279296875,
0.1158447265625,
0.312255859375
],
[
0.53076171875,
0.298583984375,
0.880859375,
0.2998046875,
0.880859375,
0.314208984375,
0.53076171875,
0.3134765625
],
[
0.118408203125,
0.312255859375,
0.489501953125,
0.311279296875,
0.489501953125,
0.32470703125,
0.118408203125,
0.325927734375
],
[
0.51318359375,
0.31640625,
0.87890625,
0.31640625,
0.87890625,
0.330078125,
0.51318359375,
0.330078125
],
[
0.118408203125,
0.32666015625,
0.484375,
0.32666015625,
0.484375,
0.33740234375,
0.118408203125,
0.33740234375
],
[
0.51123046875,
0.33203125,
0.880859375,
0.33203125,
0.880859375,
0.345703125,
0.51123046875,
0.345703125
],
[
0.118408203125,
0.339111328125,
0.486083984375,
0.341552734375,
0.486083984375,
0.355224609375,
0.118408203125,
0.352783203125
],
[
0.51123046875,
0.3466796875,
0.8818359375,
0.347412109375,
0.8818359375,
0.3623046875,
0.51123046875,
0.361083984375
],
[
0.118408203125,
0.35693359375,
0.246826171875,
0.35693359375,
0.246826171875,
0.36767578125,
0.118408203125,
0.36767578125
],
[
0.51318359375,
0.363525390625,
0.87744140625,
0.364013671875,
0.87744140625,
0.37890625,
0.51318359375,
0.377685546875
],
[
0.185546875,
0.385498046875,
0.490234375,
0.39013671875,
0.490234375,
0.412109375,
0.185546875,
0.40673828125
],
[
0.5146484375,
0.380126953125,
0.880859375,
0.380126953125,
0.880859375,
0.393798828125,
0.5146484375,
0.393798828125
],
[
0.118408203125,
0.3955078125,
0.187255859375,
0.399658203125,
0.185546875,
0.4150390625,
0.1175537109375,
0.410888671875
],
[
0.5146484375,
0.39794921875,
0.880859375,
0.39794921875,
0.880859375,
0.407958984375,
0.5146484375,
0.407958984375
],
[
0.280517578125,
0.40380859375,
0.364501953125,
0.40380859375,
0.364501953125,
0.41455078125,
0.280517578125,
0.41455078125
],
[
0.118408203125,
0.41455078125,
0.48779296875,
0.41455078125,
0.48779296875,
0.4287109375,
0.118408203125,
0.4287109375
],
[
0.51123046875,
0.412109375,
0.880859375,
0.412109375,
0.880859375,
0.42578125,
0.51123046875,
0.42578125
],
[
0.118408203125,
0.43115234375,
0.489501953125,
0.43115234375,
0.489501953125,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.51318359375,
0.427490234375,
0.880859375,
0.427490234375,
0.880859375,
0.441162109375,
0.51318359375,
0.441162109375
],
[
0.1158447265625,
0.44482421875,
0.489501953125,
0.446533203125,
0.489501953125,
0.46435546875,
0.1158447265625,
0.4619140625
],
[
0.51318359375,
0.4453125,
0.87890625,
0.4453125,
0.87890625,
0.458984375,
0.51318359375,
0.458984375
],
[
0.118408203125,
0.463134765625,
0.486083984375,
0.463134765625,
0.486083984375,
0.476806640625,
0.118408203125,
0.476806640625
],
[
0.51123046875,
0.46142578125,
0.87890625,
0.46142578125,
0.87890625,
0.474365234375,
0.51123046875,
0.474365234375
],
[
0.1175537109375,
0.478515625,
0.484375,
0.478515625,
0.484375,
0.4921875,
0.1175537109375,
0.4921875
],
[
0.51123046875,
0.4755859375,
0.880859375,
0.476806640625,
0.880859375,
0.4912109375,
0.51123046875,
0.490478515625
],
[
0.118408203125,
0.495361328125,
0.48779296875,
0.495361328125,
0.48779296875,
0.5087890625,
0.118408203125,
0.5087890625
],
[
0.51318359375,
0.493408203125,
0.880859375,
0.493408203125,
0.880859375,
0.50732421875,
0.51318359375,
0.50732421875
],
[
0.118408203125,
0.51123046875,
0.4013671875,
0.51123046875,
0.4013671875,
0.52490234375,
0.118408203125,
0.52490234375
],
[
0.51123046875,
0.50830078125,
0.87890625,
0.5087890625,
0.87890625,
0.52392578125,
0.51123046875,
0.5224609375
],
[
0.1368408203125,
0.52783203125,
0.486083984375,
0.52783203125,
0.486083984375,
0.54150390625,
0.1368408203125,
0.54150390625
],
[
0.5146484375,
0.525390625,
0.87744140625,
0.525390625,
0.87744140625,
0.5390625,
0.5146484375,
0.5390625
],
[
0.1175537109375,
0.54345703125,
0.48779296875,
0.5419921875,
0.48779296875,
0.55712890625,
0.1175537109375,
0.55810546875
],
[
0.51318359375,
0.54150390625,
0.880859375,
0.54150390625,
0.880859375,
0.5546875,
0.51318359375,
0.5546875
],
[
0.118408203125,
0.56005859375,
0.486083984375,
0.56005859375,
0.486083984375,
0.57373046875,
0.118408203125,
0.57373046875
],
[
0.51318359375,
0.55810546875,
0.880859375,
0.55810546875,
0.880859375,
0.57177734375,
0.51318359375,
0.57177734375
],
[
0.1175537109375,
0.57470703125,
0.484375,
0.576171875,
0.484375,
0.59033203125,
0.1175537109375,
0.5888671875
],
[
0.5146484375,
0.57373046875,
0.880859375,
0.57373046875,
0.880859375,
0.58740234375,
0.5146484375,
0.58740234375
],
[
0.118408203125,
0.59130859375,
0.484375,
0.59130859375,
0.484375,
0.60498046875,
0.118408203125,
0.60498046875
],
[
0.51318359375,
0.5888671875,
0.880859375,
0.5888671875,
0.880859375,
0.6025390625,
0.51318359375,
0.6025390625
],
[
0.12005615234375,
0.60791015625,
0.486083984375,
0.60791015625,
0.486083984375,
0.62158203125,
0.12005615234375,
0.62158203125
],
[
0.51123046875,
0.60546875,
0.880859375,
0.60546875,
0.880859375,
0.619140625,
0.51123046875,
0.619140625
],
[
0.118408203125,
0.62451171875,
0.486083984375,
0.62451171875,
0.486083984375,
0.63818359375,
0.118408203125,
0.63818359375
],
[
0.51123046875,
0.62060546875,
0.87890625,
0.62060546875,
0.87890625,
0.63427734375,
0.51123046875,
0.63427734375
],
[
0.12005615234375,
0.64013671875,
0.486083984375,
0.64013671875,
0.486083984375,
0.65380859375,
0.12005615234375,
0.65380859375
],
[
0.51318359375,
0.63818359375,
0.880859375,
0.63818359375,
0.880859375,
0.65185546875,
0.51318359375,
0.65185546875
],
[
0.1175537109375,
0.65478515625,
0.484375,
0.65478515625,
0.484375,
0.66845703125,
0.1175537109375,
0.66845703125
],
[
0.51318359375,
0.65380859375,
0.85546875,
0.65380859375,
0.85546875,
0.66845703125,
0.51318359375,
0.66845703125
],
[
0.1175537109375,
0.6728515625,
0.48779296875,
0.67138671875,
0.48779296875,
0.68603515625,
0.1175537109375,
0.68701171875
],
[
0.5146484375,
0.6884765625,
0.5615234375,
0.6884765625,
0.5615234375,
0.69873046875,
0.5146484375,
0.69873046875
],
[
0.55322265625,
0.68603515625,
0.85546875,
0.68701171875,
0.85546875,
0.70166015625,
0.55322265625,
0.70068359375
],
[
0.55322265625,
0.70361328125,
0.87744140625,
0.70361328125,
0.87744140625,
0.71728515625,
0.55322265625,
0.71728515625
],
[
0.118408203125,
0.689453125,
0.313232421875,
0.689453125,
0.313232421875,
0.703125,
0.118408203125,
0.703125
],
[
0.1158447265625,
0.71875,
0.151123046875,
0.71875,
0.151123046875,
0.72998046875,
0.1158447265625,
0.72998046875
],
[
0.1485595703125,
0.71630859375,
0.48193359375,
0.71728515625,
0.48193359375,
0.732421875,
0.1485595703125,
0.73095703125
],
[
0.51318359375,
0.72705078125,
0.880859375,
0.72802734375,
0.880859375,
0.7421875,
0.51318359375,
0.7412109375
],
[
0.1485595703125,
0.73388671875,
0.31494140625,
0.73388671875,
0.31494140625,
0.74755859375,
0.1485595703125,
0.74755859375
],
[
0.51318359375,
0.7421875,
0.87890625,
0.74365234375,
0.87890625,
0.75830078125,
0.51318359375,
0.75732421875
],
[
0.118408203125,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.118408203125,
0.7724609375
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.118408203125,
0.77685546875,
0.486083984375,
0.77685546875,
0.486083984375,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.51318359375,
0.77685546875,
0.880859375,
0.77685546875,
0.880859375,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51318359375,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.118408203125,
0.80810546875,
0.483642578125,
0.80810546875,
0.483642578125,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.51318359375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.1175537109375,
0.82470703125,
0.48193359375,
0.82470703125,
0.48193359375,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.51123046875,
0.82470703125,
0.87744140625,
0.82470703125,
0.87744140625,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1175537109375,
0.8388671875,
0.486083984375,
0.84033203125,
0.486083984375,
0.85498046875,
0.1175537109375,
0.85400390625
],
[
0.51123046875,
0.84130859375,
0.87890625,
0.84130859375,
0.87890625,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.12176513671875,
0.87451171875,
0.486083984375,
0.87451171875,
0.486083984375,
0.88525390625,
0.12176513671875,
0.88525390625
],
[
0.509765625,
0.87255859375,
0.87890625,
0.87255859375,
0.87890625,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1318359375,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1318359375,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.12176513671875,
0.90673828125,
0.48779296875,
0.90673828125,
0.48779296875,
0.91748046875,
0.12176513671875,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.87744140625,
0.90478515625,
0.87744140625,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"X10-4",
"个是我们在RL期间冻结学习的目标嵌人。我",
"2",
"Model",
"们还探索了两种方法的结合。冻结目标嵌入",
"100",
"o",
"STV",
"（解码器的最后一层）可以解释为训练网络以",
"0",
"LTV",
"输出倒数第二层的激活，然后固定函数将其映",
"0",
"射到词汇表的维度",
"",
"",
"2",
"我们注意到，尽管冻结是一种常见的过程",
"50",
"(Zoph et al., 2016; Thompson et al., 2018; Lee",
"-6",
"-100",
"etal.2019：Costeretal.，2021），但据我们所",
"-150",
"知，它从未应用于将RL用于序列到序列模",
"",
"型。表示网络用fe计算的函数。fe可以写成",
"2345",
"89.10",
"",
"",
"",
"hg20901，其中=（01,02）、g将输入源",
"句X和模型翻译前缀y<映射到Rd，h",
"Gold token rank by model (De-En)",
"Figure 1: Comparison of probability shift due to RL",
"将g的输出映射到 RIVrl",
"training of assign ybest for ten first words for both LTV",
"使用此表示法，我们可以将方法表述为将",
"and STV. in blue, you can see the results with BPE of",
"预训练的MLE目标嵌入加载到he或冻结它",
"size 1.O00. STV. in red are the results with BPE of size",
"（或两者）。至于许多编码器-解码器架构（包",
"30,000, LTV. a clear improvement of assigning ybest in",
"括Transformer），它认为d<|Vrl，这可以被",
"first place for STV.",
"认为是约束代理选择d维连续动作，其中he2",
"=#gold tokenassignstotherrank)。然",
"是由环境执行的已知转换。目标嵌人的重要性",
", Pmodel",
"在于，它们充许更好地概括操作。直觉如下。",
"#(all trials)",
"后，我们通过减去这两个分布来比较MLE模",
"假设两个令牌具有相同的嵌入和相似的语义，",
"型和RL模型的秩分布。在我们的表示法中，",
"即。e。？它们适用于相同的上下文(同义词)。",
"对于每个秩r，△Pr=PRL-PMLE。此减法",
"由于它们具有相同的目标嵌入，因此在训练期",
"表示RL如何影响模型为每个等级分配正确令",
"间，当遇到它们中的任何一个时，网络将执行",
"牌yi的能力。RL的积极作用越大，第一等级",
"相同的梯度更新，除了在最后一层（因为它们",
"的概率为正的可能性就越大，而较低等级的概",
"仍然被认为是不同的输出）。如果目标嵌人未",
"率为负（由于概率转移到了第一位）",
"冻结，则在训练过程中遇到其中任何一个都会",
"图1显示了LTV和STV的每秩概率差。我",
"导致非常相似的更新（因为它们具有相同的目",
"们可以看到，对于第一秩，由于使用STV进",
"标嵌入），但是它们的目标嵌入可能会稍微偏",
"行RL训练而导致的概率偏移是使用LTV进行",
"离，这将导致较低层的后续漂移。如果目标嵌",
"RL训练所导致的偏移的两倍以上。因此，对",
"入被冻结，它们将产生的梯度更新将保持不变",
"于小词汇量设置，以下排名的概率偏移通常更",
"并加快学习速度。我们假设在训练期间有类似",
"负。该图表明，当我们使用较小的动作空间应",
"的效果，其中具有相似（但不相同）嵌入的令",
"用RL时，概率质量向较高位置的移动实际上",
"牌和相似（但不相同）的分布将在训练中彼此",
"发生得更多。此外，与LTV相比，STV训练能",
"受益。这促使我们探索信息初始化和参数冻结",
"够将概率质量从较低的等级上移。对此的指示",
"的组合。（请参阅Supp中的正式证明。SE）",
"是，在前一百个等级中，STV降低了它们的83",
"5.1",
" Motivating Simulation through Policy",
"Parameterization in Large Action Spaces",
"的概率，而仅2的LTV。",
"5",
"Reducing the Effective Dimensionality",
"为了在受控设置中检查上面概述的直觉，我们",
"of the Action Space",
"考虑一个合成RL问题，其中动作空间被表面",
"发现减少动作数量可以提高RL的性能，我们",
"放大。该任务是一个（上下文）多臂强盗，具",
"提出了一种减少有效动作数量的方法，而不会",
"有K操作。在每个步骤中，从环境中采样输入",
"改变实际输出。词汇大小可能是静态的，如在",
"状态（“上下文”；从多元高斯分布采样的随机",
"预训练模型(Devlinetal.,2019）中，减少它可",
"向量）。随机，固定，非线性二进制分类器根据",
"能有助于RL，但对于MLE（Gowdaand May",
"给定的上下文确定动作#1或动作#2是否有回",
"2020)是最佳的，或者引人域外单词（Koehnand",
"报（动作3-K永远不会有回报），并且每个动",
"Knowles，2017）。我们建议通过使用在类似上",
"作的奖励为r+，其中r=1用于奖励动作",
"下文中出现的令牌上泛化的自标嵌人来做到这",
"0用于所有其他动作，以及z～（0,0.1）。至",
"点。我们探索了这个想法的两个实现，一个",
"关重要的是，我们将每个操作α次复制，从而",
"是我们用高质量的嵌入初始化目标嵌入，另一",
"在策略级别上总共产生K×a个操作-而对于"
] |
{
"arxiv_id": "2210.03053",
"language": "zh",
"page_idx": 4
}
|
||
2108.02126-zh-0
|
2108.02126-zh-0.jpg
|
I WILL HAVE ORDER! OPTIMIZING ORDERS FOR FAIR REVIEWER ASSIGNMENT
# I Will Have Order! Optimizing Orders for Fair Reviewer Assignment
Justin Payan
$\mathbf{Yair}$ Zick
JPAYAN$\textcircled{\mathrm{Q}}$UMASS.EDU YZICK$\textcircled{\mathrm{Q}}$UMASS.EDU
$UniversityofMassachusetts$, College $of$ Information and Computer $Sciences$,
$140\textit{ Governors Drive, Amherst, MA 01002 USA}$
$\mathbf{Abstract}$
We present fast, fair, flexible, and welfare efficient algorithms for assigning reviewers to submitted conference papers. Our approaches extend picking sequence mechanisms, standard tools from the fair allocation literature to ensure approximate envy-freeness (typically envy-freeness up to one item, or EF1). However, fairness often comes at the cost of decreased efficiency. To overcome this challenge, we carefully select approximately optimal picking sequence orders. Applying a relaxation of submodularity, $\gamma$ -weak submodularity, we show our Greedy Reviewer Round Robin ( GRRR ) approach is EF1 and yields a $(1+\gamma)$ -approximation to the maximum welfare attainable by a round-robin picking sequence mechanism under any order. We present a weighted picking sequence mechanism called FairSequence that targets the Weighted EF1 criterion to offer fairness in a more general setting. Using data from three conferences, we show that FairSequence runs an order of magnitude faster and provides approximate envy-freeness guarantees that are violated by existing approaches. Its simple design also makes it very flexible to new assignment constraints. FairSequence is available in the OpenReview $^1$ conference management platform, giving conference organizers access to faster reviewer assignment with high welfare and envy-freeness guarantees.
## 1. Introduction
同行评议在学术界的几乎所有方面都发挥着重要作用。它具有许多功能-选择最佳手稿，评估原创性，提供反馈等等 (Mulligan et al.,2013) 。鉴于同行评议的广泛应用及其重要的把关作用，这一过程必须尽可能保持客观
一个重要的参数是论文是否分配了能够适当审查其工作的专家。因此，为提交的论文选择审稿人是任何审查过程中至关重要的第一步。在诸如 NeurIPS, ICML, AAAI 和 IJCAI 之类的大型会议中，审阅者的分配在很大程度上是通过诸如多伦多论文匹配系统(TPMS)[E] (Charlin&Zemel, 2013) , MicrosoftCMT $^2$或 OpenReview $^3$之类的系统实现的。。分配不当的审阅者可能会导致失败：错误的决定，审阅者不感兴趣以及对同行评审过程的普遍不信任
我们寻求快速，公平，灵活(针对新约束)且准确的审阅者分配算法。准确性和公平性是审阅者分配$\overline{\mathrm{E}}($Shah, 2019) 中两个非常重要的标准，而一般 (Bouveret et al., 2016) 中的公平划分。总体分配准确性保持了学术出版物的质量标准：我们希望审稿人与论文的整体匹配。然而，当务之急是我们不要牺牲一些论文的复习质量来获得更好的整体匹配。即使
1. https://github.com/openreview/openreview-matcher
2. https://cmt3.research.microsoft.com/
3. https://openreview.net/
1
|
[
[
0.2017822265625,
0.04986572265625,
0.7900390625,
0.04986572265625,
0.7900390625,
0.0631103515625,
0.2017822265625,
0.0631103515625
],
[
0.38818359375,
0.1300048828125,
0.60888671875,
0.1300048828125,
0.60888671875,
0.14453125,
0.38818359375,
0.14453125
],
[
0.2181396484375,
0.15283203125,
0.78173828125,
0.15283203125,
0.78173828125,
0.1673583984375,
0.2181396484375,
0.1673583984375
],
[
0.1446533203125,
0.1956787109375,
0.26318359375,
0.1956787109375,
0.26318359375,
0.210205078125,
0.1446533203125,
0.210205078125
],
[
0.705078125,
0.1962890625,
0.8544921875,
0.1962890625,
0.8544921875,
0.210205078125,
0.705078125,
0.210205078125
],
[
0.146240234375,
0.212158203125,
0.2303466796875,
0.212158203125,
0.2303466796875,
0.2266845703125,
0.146240234375,
0.2266845703125
],
[
0.7138671875,
0.21533203125,
0.85302734375,
0.21533203125,
0.85302734375,
0.2259521484375,
0.7138671875,
0.2259521484375
],
[
0.146240234375,
0.2291259765625,
0.6953125,
0.2298583984375,
0.6953125,
0.2442626953125,
0.146240234375,
0.24365234375
],
[
0.148681640625,
0.248779296875,
0.48779296875,
0.248779296875,
0.48779296875,
0.2587890625,
0.148681640625,
0.2587890625
],
[
0.459228515625,
0.29345703125,
0.54150390625,
0.29345703125,
0.54150390625,
0.30810546875,
0.459228515625,
0.30810546875
],
[
0.2059326171875,
0.31494140625,
0.8203125,
0.31494140625,
0.8203125,
0.328857421875,
0.2059326171875,
0.328857421875
],
[
0.178955078125,
0.330078125,
0.81884765625,
0.330078125,
0.81884765625,
0.3447265625,
0.178955078125,
0.3447265625
],
[
0.178955078125,
0.34521484375,
0.8203125,
0.34521484375,
0.8203125,
0.35986328125,
0.178955078125,
0.35986328125
],
[
0.1805419921875,
0.3623046875,
0.81640625,
0.3623046875,
0.81640625,
0.37255859375,
0.1805419921875,
0.37255859375
],
[
0.178955078125,
0.375732421875,
0.8203125,
0.375732421875,
0.8203125,
0.389404296875,
0.178955078125,
0.389404296875
],
[
0.178955078125,
0.39013671875,
0.81787109375,
0.39013671875,
0.81787109375,
0.40478515625,
0.178955078125,
0.40478515625
],
[
0.17724609375,
0.406005859375,
0.8203125,
0.406005859375,
0.8203125,
0.419189453125,
0.17724609375,
0.419189453125
],
[
0.17724609375,
0.42041015625,
0.8203125,
0.419189453125,
0.8203125,
0.434326171875,
0.17724609375,
0.43505859375
],
[
0.178955078125,
0.436279296875,
0.81884765625,
0.436279296875,
0.81884765625,
0.449462890625,
0.178955078125,
0.449462890625
],
[
0.178955078125,
0.451904296875,
0.8203125,
0.451904296875,
0.8203125,
0.46533203125,
0.178955078125,
0.46533203125
],
[
0.178955078125,
0.46728515625,
0.8203125,
0.46728515625,
0.8203125,
0.48095703125,
0.178955078125,
0.48095703125
],
[
0.178955078125,
0.481689453125,
0.81884765625,
0.481689453125,
0.81884765625,
0.49560546875,
0.178955078125,
0.49560546875
],
[
0.17724609375,
0.496826171875,
0.8203125,
0.496826171875,
0.8203125,
0.51025390625,
0.17724609375,
0.51025390625
],
[
0.178955078125,
0.51123046875,
0.81884765625,
0.51123046875,
0.81884765625,
0.5244140625,
0.178955078125,
0.5244140625
],
[
0.178955078125,
0.5263671875,
0.82177734375,
0.5263671875,
0.82177734375,
0.54052734375,
0.178955078125,
0.54052734375
],
[
0.176513671875,
0.54150390625,
0.390625,
0.54248046875,
0.390625,
0.55615234375,
0.176513671875,
0.5546875
],
[
0.143798828125,
0.580078125,
0.30224609375,
0.58056640625,
0.30224609375,
0.59521484375,
0.143798828125,
0.5947265625
],
[
0.1478271484375,
0.60986328125,
0.85302734375,
0.60986328125,
0.85302734375,
0.62353515625,
0.1478271484375,
0.62353515625
],
[
0.146240234375,
0.626953125,
0.85302734375,
0.626953125,
0.85302734375,
0.64013671875,
0.146240234375,
0.64013671875
],
[
0.148681640625,
0.64599609375,
0.44775390625,
0.64599609375,
0.44775390625,
0.65576171875,
0.148681640625,
0.65576171875
],
[
0.1854248046875,
0.662109375,
0.85107421875,
0.662109375,
0.85107421875,
0.67529296875,
0.1854248046875,
0.67529296875
],
[
0.146240234375,
0.6796875,
0.85302734375,
0.6796875,
0.85302734375,
0.693359375,
0.146240234375,
0.693359375
],
[
0.146240234375,
0.6962890625,
0.85302734375,
0.6962890625,
0.85302734375,
0.71044921875,
0.146240234375,
0.71044921875
],
[
0.1446533203125,
0.7119140625,
0.85302734375,
0.71337890625,
0.85302734375,
0.72802734375,
0.1446533203125,
0.7265625
],
[
0.1478271484375,
0.73095703125,
0.85302734375,
0.73095703125,
0.85302734375,
0.744140625,
0.1478271484375,
0.744140625
],
[
0.146240234375,
0.74755859375,
0.1846923828125,
0.74755859375,
0.1846923828125,
0.76318359375,
0.146240234375,
0.76318359375
],
[
0.1732177734375,
0.7666015625,
0.85302734375,
0.7666015625,
0.85302734375,
0.77978515625,
0.1732177734375,
0.77978515625
],
[
0.146240234375,
0.7841796875,
0.85302734375,
0.7841796875,
0.85302734375,
0.79736328125,
0.146240234375,
0.79736328125
],
[
0.1478271484375,
0.8017578125,
0.85107421875,
0.8017578125,
0.85107421875,
0.81494140625,
0.1478271484375,
0.81494140625
],
[
0.146240234375,
0.818359375,
0.85302734375,
0.818359375,
0.85302734375,
0.83203125,
0.146240234375,
0.83203125
],
[
0.1502685546875,
0.8486328125,
0.5400390625,
0.849609375,
0.5400390625,
0.8642578125,
0.1502685546875,
0.86279296875
],
[
0.1502685546875,
0.8623046875,
0.44677734375,
0.86279296875,
0.44677734375,
0.87744140625,
0.1502685546875,
0.876953125
],
[
0.1502685546875,
0.87744140625,
0.34716796875,
0.87744140625,
0.34716796875,
0.8916015625,
0.1502685546875,
0.8916015625
],
[
0.49267578125,
0.91552734375,
0.50634765625,
0.91552734375,
0.50634765625,
0.9267578125,
0.49267578125,
0.9267578125
]
] |
[
"I WILL HAVE ORDER! OPTIMIZING ORDERS FOR FAIR REVIEWER ASSIGNMENT",
"I Will Have Order!",
"Optimizing Orders for Fair Reviewer Assignment",
"Justin Payan",
"JPAYAN@UMASS.EDU",
"Yair Zick",
"YZICK@UMASS.EDU",
"University of Massachusetts, College of Information and Computer Sciences,",
"140GovernorsDrive,Amherst,MA01002USA",
"Abstract",
"We present fast, fair, flexible, and welfare efficient algorithms for assigning reviewers to",
"submitted conference papers. Our approaches extend picking sequence mechanisms, stan-",
"dard tools from the fair allocation literature to ensure approximate envy-freeness (typically",
"envy-freenessuptooneitem,or EFl)However,fairnessoften comes at thecostof de",
"creased efficiency. To overcome this challenge, we carefully select approximately optimal",
"picking sequence orders. Applying a relaxation of submodularity, -weak submodular-",
"ity, we show our Greedy Reviewer Round Robin ( GRRR) approach is EF1 and yields a",
"(1+)-approximation to the maximum welfare attainable by a round-robin picking se-",
"quence mechanism under any order: We present a weighted picking sequence mechanism",
"called FairSequence that targets the Weighted EFl criterion to offer fairness in a more",
"general setting. Using data from three conferences, we show that FairSequence runs an",
"order of magnitude faster and provides approximate envy-freeness guarantees that are vio",
"lated by existing approaches. Its simple design also makes it very flexible to new assignment",
"constraints. FairSequence is available in the OpenReview 1 conference management plat-",
"form, giving conference organizers access to faster reviewer assignment with high welfare",
"and envy-freeness guarantees.",
"1.Introduction",
"同行评议在学术界的几乎所有方面都发挥着重要作用。它具有许多功能-选择最佳手稿，评",
"估原创性，提供反馈等等（Mulliganetal.，2013）。鉴于同行评议的广泛应用及其重要的把",
"关作用，这一过程必须尽可能保持客观",
"个重要的参数是论文是否分配了能够适当审查其工作的专家。因此，为提交的论文选",
"择审稿人是任何审查过程中至关重要的第一步。在诸如NeurIPS，ICML，AAAI和IJCAI",
"之类的大型会议中，审阅者的分配在很大程度上是通过诸如多伦多论文匹配系统（TPMS）E",
"（Charlin&Zemel，2013），MicrosoftCMT2或OpenReview3之类的系统实现的。。分配不",
"当的审阅者可能会导致失败：错误的决定，审阅者不感兴趣以及对同行评审过程的普遍不",
"信任",
"我们寻求快速，公平，灵活（针对新约束）且准确的审阅者分配算法。准确性和公平性",
"是审阅者分配E(Shah，2019）中两个非常重要的标准，而一般（Bouveretetal，2016）中的",
"公平划分。总体分配准确性保持了学术出版物的质量标准：我们希望审稿人与论文的整体",
"匹配。然而，当务之急是我们不要牺性一些论文的复习质量来获得更好的整体匹配。即使",
"l.https://github.com/openreview/openreview-matcher",
"2.https://cmt3.research.microsoft.com/",
"3.https://openreview.net/",
"1"
] |
{
"arxiv_id": "2108.02126",
"language": "zh",
"page_idx": 0
}
|
||
2108.02126-zh-1
|
2108.02126-zh-1.jpg
|
PAYAN$\&$ZICK
总体任务很好，但审稿人匹配不佳的论文可能会被不公平地拒绝或收到无益的反馈，从而给作者造成真正的伤害。因此，我们希望算法是全球准确和公平的论文。此外，我们需要快速运行的算法，因为会议组织者经常计算多个潜在的分配，每次更改约束或冻结一些匹配项。我们还需要我们的算法简单并能适应新的约束；实际上，大型会议的组织者通常会有详细的约束，例如禁止将同一机构的两名审稿人分配给同一篇论文，或者每篇论文至少需要一名高级审稿人 (Leyton-Brown et al., 2022) 。为了实现这些目标，我们从公平分配的角度考虑公平审稿人的分配问题。
我们的主要公平标准令人羡慕：如果另一篇论文的指定审阅者比自己的指定审阅者更匹配，则一篇论文会羡慕另一篇论文。嫉妒的自由及其放松可以用来排除审稿人论文亲和力得分的显著差异。最先进的公平审稿人分配算法要么最大化最小论文亲和力得分，要么确保所有亲和力得分超过某个阈值 (Stelmakh et al.,2019; Kobren et al., 2019) 。尽管这种方法可确保所有论文的审稿人都通过了某些最低质量门槛，但仍有许多论文的审稿人排位不佳，可以从少量低成本交易中受益。基于嫉妒的标准，如最多一个项目的嫉妒自由 (EF1), 确保不存在这样的“明显错误”,在本地平衡任务
对于不可分割的项目E$( Bouveret et al. , 2016) , $通常不可能获得无嫉妒的分配；因此， 我们将重点放在最多一个项目 (EF1) (Budish, 2011; Lipton et al., 2004) 的无嫉妒的宽松标准上。EF1 分配具有以下特性：每当论文$p$对论文$p^\prime$的审阅者具有较高的亲和力时，这是由于单个高亲和力的审阅者而不是结果完全不平衡。EF1 条件下的福利最大化是 NP 难的，在多项式时间内是不可逼近的 (Barman et al.,2019) 。因此，我们探索使用挑选序列机制产生具有良好经验绩效和部分福利保证的 EF1 分配的方法。在标准的公平分配设置中， 众所周知的循环( RR ) 机制通过设置代理的顺序并让他们一次选择一个项目来产生 EF1 分配。由于审阅者选择的限制，轮询不是用于审阅者分配的 EF1。因此，我们提出了经典 RR 的变体，我们称之为 ReviewerRoundRobin (RRR)
循环机制为每篇论文分配相同数量的审阅者，但有时会议需要可变大小的分配。例如， 当某些审阅者在第一阶段未做出响应时，两阶段审阅过程通常需要在第二阶段分配可变数量的审阅者。因此，我们还研究了满足加权 EF1 约束 (WEF1)$\operatorname{|E}($Chakraborty et al., 2021a)的加权采摘序列族。此约束将 EF1 推广到代理具有不等项目需求的设置，确保代理在按其需求标准化后为 EF1。我们的算法，加权评审员挑选(wRP)和 FairSequence,遵循 RRRR , 通过应用标准加权挑选序列并添加一些步骤来适应评审员分配的约束
虽然已知拣选顺序机制可以满足公平约束，但其效率保证在很大程度上取决于玩家拣选物品的顺序。例如，考虑一个风格化的设置，其中有两篇论文($p_i$和$p_j$ )和两个审阅者$(r_{1}$和$r_2):$论文$p_i$的两个审阅者的亲和力得分为 5,而论文$p_j$的$r_1$得分为 10, $r_2$得分为0。让$p_i$首先选择的循环机制冒着让$p_i$选择$r_1$的风险，而将$p_j$留给$r_2$ 。让$p_j$首先选择一个更好的结果，而不影响公平性。通常很难确定最佳采摘序列 (Bouveret &Lang, 2011; Kalinowski et al., 2013; Aziz et al., 2015, 2016)。我们间这个问题：我们能确定近似最优的纸张订单吗？
# 1.1 Our Contributions
我们对论文的顺序进行组合搜索，从而为 RRR 和 WRP 等拾取顺序机制产生高效分配。为此， 我们研究了通过子模块优化的镜头找到最佳纸张订单的问题。。我们优化了部分纸张序列上的函数，该函数根据拾取序列产生的分配的福利而变化。此函数通常不是子模块化的，但是我们可以通过变量$\gamma$捕获其与子模块化的距离。我们的主要理论结果(定理 5.1)是公平划分社区的独立利益，表明一种简单的贪婪方法使该函数最大化到$1+\gamma$的因子。我们
2
|
[
[
0.441162109375,
0.050506591796875,
0.55224609375,
0.050506591796875,
0.55224609375,
0.06439208984375,
0.441162109375,
0.06439208984375
],
[
0.1478271484375,
0.11871337890625,
0.85107421875,
0.11871337890625,
0.85107421875,
0.1319580078125,
0.1478271484375,
0.1319580078125
],
[
0.1478271484375,
0.1363525390625,
0.463134765625,
0.1363525390625,
0.463134765625,
0.146484375,
0.1478271484375,
0.146484375
],
[
0.4599609375,
0.1383056640625,
0.4794921875,
0.1383056640625,
0.4794921875,
0.1458740234375,
0.4599609375,
0.1458740234375
],
[
0.5556640625,
0.136962890625,
0.85107421875,
0.136962890625,
0.85107421875,
0.147705078125,
0.5556640625,
0.147705078125
],
[
0.1478271484375,
0.15283203125,
0.438720703125,
0.15283203125,
0.438720703125,
0.1666259765625,
0.1478271484375,
0.1666259765625
],
[
0.441162109375,
0.1578369140625,
0.455810546875,
0.1578369140625,
0.455810546875,
0.1610107421875,
0.441162109375,
0.1610107421875
],
[
0.5048828125,
0.15283203125,
0.85302734375,
0.15283203125,
0.85302734375,
0.1666259765625,
0.5048828125,
0.1666259765625
],
[
0.146240234375,
0.169189453125,
0.85107421875,
0.169189453125,
0.85107421875,
0.18310546875,
0.146240234375,
0.18310546875
],
[
0.1478271484375,
0.186279296875,
0.85302734375,
0.186279296875,
0.85302734375,
0.199462890625,
0.1478271484375,
0.199462890625
],
[
0.1478271484375,
0.203857421875,
0.8486328125,
0.203857421875,
0.8486328125,
0.2171630859375,
0.1478271484375,
0.2171630859375
],
[
0.1478271484375,
0.2203369140625,
0.3857421875,
0.2203369140625,
0.3857421875,
0.2342529296875,
0.1478271484375,
0.2342529296875
],
[
0.1748046875,
0.24365234375,
0.85302734375,
0.24365234375,
0.85302734375,
0.2568359375,
0.1748046875,
0.2568359375
],
[
0.146240234375,
0.260009765625,
0.85302734375,
0.260009765625,
0.85302734375,
0.2734375,
0.146240234375,
0.2734375
],
[
0.146240234375,
0.27783203125,
0.85302734375,
0.27783203125,
0.85302734375,
0.291748046875,
0.146240234375,
0.291748046875
],
[
0.1478271484375,
0.294189453125,
0.85107421875,
0.294189453125,
0.85107421875,
0.30810546875,
0.1478271484375,
0.30810546875
],
[
0.146240234375,
0.311279296875,
0.85302734375,
0.311279296875,
0.85302734375,
0.324462890625,
0.146240234375,
0.324462890625
],
[
0.1478271484375,
0.328857421875,
0.8544921875,
0.328857421875,
0.8544921875,
0.34228515625,
0.1478271484375,
0.34228515625
],
[
0.1478271484375,
0.34521484375,
0.52783203125,
0.34521484375,
0.52783203125,
0.359130859375,
0.1478271484375,
0.359130859375
],
[
0.1748046875,
0.36865234375,
0.8544921875,
0.36865234375,
0.8544921875,
0.3818359375,
0.1748046875,
0.3818359375
],
[
0.146240234375,
0.3837890625,
0.85302734375,
0.385009765625,
0.85302734375,
0.399658203125,
0.146240234375,
0.3984375
],
[
0.146240234375,
0.40283203125,
0.410888671875,
0.40283203125,
0.410888671875,
0.416748046875,
0.146240234375,
0.416748046875
],
[
0.41259765625,
0.40478515625,
0.441162109375,
0.40478515625,
0.441162109375,
0.412353515625,
0.41259765625,
0.412353515625
],
[
0.45751953125,
0.40771484375,
0.47216796875,
0.40771484375,
0.47216796875,
0.410888671875,
0.45751953125,
0.410888671875
],
[
0.496826171875,
0.40283203125,
0.85302734375,
0.40283203125,
0.85302734375,
0.416748046875,
0.496826171875,
0.416748046875
],
[
0.1478271484375,
0.419189453125,
0.438720703125,
0.419189453125,
0.438720703125,
0.43310546875,
0.1478271484375,
0.43310546875
],
[
0.560546875,
0.419189453125,
0.85302734375,
0.419189453125,
0.85302734375,
0.43310546875,
0.560546875,
0.43310546875
],
[
0.146240234375,
0.436279296875,
0.419189453125,
0.436279296875,
0.419189453125,
0.449462890625,
0.146240234375,
0.449462890625
],
[
0.5537109375,
0.436279296875,
0.85302734375,
0.436279296875,
0.85302734375,
0.449462890625,
0.5537109375,
0.449462890625
],
[
0.1478271484375,
0.453857421875,
0.85302734375,
0.453857421875,
0.85302734375,
0.46728515625,
0.1478271484375,
0.46728515625
],
[
0.1478271484375,
0.47021484375,
0.85107421875,
0.47021484375,
0.85107421875,
0.48486328125,
0.1478271484375,
0.48486328125
],
[
0.146240234375,
0.488037109375,
0.8544921875,
0.488037109375,
0.8544921875,
0.501953125,
0.146240234375,
0.501953125
],
[
0.146240234375,
0.5048828125,
0.54345703125,
0.5048828125,
0.54345703125,
0.51806640625,
0.146240234375,
0.51806640625
],
[
0.1748046875,
0.5263671875,
0.8544921875,
0.5263671875,
0.8544921875,
0.54150390625,
0.1748046875,
0.54150390625
],
[
0.1478271484375,
0.54443359375,
0.85302734375,
0.54443359375,
0.85302734375,
0.55810546875,
0.1478271484375,
0.55810546875
],
[
0.146240234375,
0.56103515625,
0.85107421875,
0.56103515625,
0.85107421875,
0.57470703125,
0.146240234375,
0.57470703125
],
[
0.146240234375,
0.57763671875,
0.8505859375,
0.57763671875,
0.8505859375,
0.59130859375,
0.146240234375,
0.59130859375
],
[
0.146240234375,
0.59326171875,
0.8544921875,
0.59521484375,
0.8544921875,
0.6123046875,
0.146240234375,
0.60986328125
],
[
0.1478271484375,
0.6123046875,
0.7548828125,
0.6123046875,
0.7548828125,
0.62548828125,
0.1478271484375,
0.62548828125
],
[
0.1748046875,
0.63525390625,
0.85302734375,
0.63525390625,
0.85302734375,
0.6484375,
0.1748046875,
0.6484375
],
[
0.1446533203125,
0.65087890625,
0.85302734375,
0.6513671875,
0.85302734375,
0.66650390625,
0.1446533203125,
0.66552734375
],
[
0.146240234375,
0.67041015625,
0.85107421875,
0.67041015625,
0.85107421875,
0.6845703125,
0.146240234375,
0.6845703125
],
[
0.1478271484375,
0.6875,
0.85302734375,
0.6875,
0.85302734375,
0.7021484375,
0.1478271484375,
0.7021484375
],
[
0.143798828125,
0.70068359375,
0.85107421875,
0.70263671875,
0.85107421875,
0.72021484375,
0.143798828125,
0.71826171875
],
[
0.146240234375,
0.7197265625,
0.85302734375,
0.72021484375,
0.85302734375,
0.73486328125,
0.146240234375,
0.734375
],
[
0.146240234375,
0.73779296875,
0.29736328125,
0.73779296875,
0.29736328125,
0.751953125,
0.146240234375,
0.751953125
],
[
0.143798828125,
0.77783203125,
0.3544921875,
0.77880859375,
0.3544921875,
0.79345703125,
0.143798828125,
0.79248046875
],
[
0.146240234375,
0.80810546875,
0.8544921875,
0.80810546875,
0.8544921875,
0.82275390625,
0.146240234375,
0.82275390625
],
[
0.146240234375,
0.82568359375,
0.85302734375,
0.82568359375,
0.85302734375,
0.8388671875,
0.146240234375,
0.8388671875
],
[
0.146240234375,
0.84326171875,
0.85302734375,
0.84326171875,
0.85302734375,
0.85693359375,
0.146240234375,
0.85693359375
],
[
0.1478271484375,
0.85986328125,
0.85302734375,
0.85986328125,
0.85302734375,
0.87353515625,
0.1478271484375,
0.87353515625
],
[
0.1478271484375,
0.87744140625,
0.85302734375,
0.87744140625,
0.85302734375,
0.8916015625,
0.1478271484375,
0.8916015625
],
[
0.491943359375,
0.91552734375,
0.50732421875,
0.91552734375,
0.50732421875,
0.9296875,
0.491943359375,
0.9296875
]
] |
[
"PAYAN & ZICK",
"总体任务很好，但审稿人匹配不佳的论文可能会被不公平地拒绝或收到无益的反馈，从而",
"给作者造成真正的伤害。因此：我们希望",
"算",
"准确和公平的论文。此外，我们需要快",
"速运行的算法，因为会议组织者经常",
"",
"潜在的分配，每次更改约束或冻结一些匹配",
"项。我们还需要我们的算法简单并能适应新的约束；实际上，大型会议的组织者通常会有",
"详细的约束，例如禁止将同一机构的两名审稿人分配给同一篇论文，或者每篇论文至少需",
"要一名高级审稿人（Leyton-Brownetal.，2022）。为了实现这些目标，我们从公平分配的角",
"度考虑公平审稿人的分配问题。",
"我们的主要公平标准令人羡慕：如果另一篇论文的指定审阅者比自己的指定审阅者更匹",
"配，则一篇论文会羡慕另一篇论文。嫉的自由及其放松可以用来排除审稿人论文亲和力",
"得分的显著差异。最先进的公平审稿人分配算法要么最大化最小论文亲和力得分，要么确",
"保所有亲和力得分超过某个阐值（Stelmakhetal.，2019；Kobrenetal.，2019）。尽管这种方",
"法可确保所有论文的审稿人都通过了某些最低质量门槛，但仍有许多论文的审稿人排位不",
"佳，可以从少量低成本交易中受益。基于嫉妒的标准，如最多一个项目的嫉妒自由（EF1），",
"确保不存在这样的“明显错误”，在本地平衡任务",
"对于不可分割的项目E(Bouveretetal.，2016），通常不可能获得无嫉妒的分配；因此，",
"我们将重点放在最多一个项目（EF1）（Budish，2011；Liptonetal.，2004）的无嫉妒的宽松",
"标准上。EF1分配具有以下特性：",
"舟兰",
"",
"对论文p’的审阅者具有较高的亲和力时，这",
"是由于单个高亲和力的审阅者而不是",
"。EF1条件下的福利最大化是NP难",
"的，在多项式时间内是不可逼近的",
"2019）。因此，我们探索使用挑选序列",
"机制产生具有良好经验绩效和部分福利保证的EF1分配的方法。在标准的公平分配设置中",
"众所周知的循环（RR）机制通过设置代理的顺序并让他们一次选择一个项目来产生EF1分",
"配。由于审阅者选择的限制，轮询不是用于审阅者分配的EF1。因此，我们提出了经典RR",
"的变体，我们称之为ReviewerRoundRobin（RRR）",
"循环机制为每篇论文分配相同数量的审阅者，但有时会议需要可变大小的分配。例如，",
"当某些审阅者在第一阶段未做出响应时，两阶段审阅过程通常需要在第二阶段分配可变",
"数量的审阅者。因此，我们还研究了满足加权EF1约束（WEF1)）E(Chakrabortyetal.",
"2021a）的加权采摘序列族。此约束将EF1推广到代理具有不等项目需求的设置，确保代理",
"在按其需求标准化后为EF1。我们的算法，加权评审员挑选（WRP）和FairSequence，遵",
"循RRR，通过应用标准加权挑选序列并添加一些步骤来适应评审员分配的约束",
"虽然已知抹选顺序机制可以满足公平约束，但其效率保证在很大程度上取决于玩家选",
"物品的顺序。例如，考虑一个风格化的设置，其中有两篇论文（p和Pi）和两个审阅者",
"(ri和r2）：论文pi的两个审阅者的亲和力得分为5，而论文pi的ri得分为10，r2得分",
"为0。让pi首先选择的循环机制冒着让pi选择r1的风险，而将pi留给r2。让pi首先",
"选择一个更好的结果，而不影响公平性。通常很难确定最佳采摘序列（Bouveret&Lang，",
"2011；Kalinowskietal.，2013；Azizetal.，2015，2016）。我们问这个问题：我们能确定近似",
"最优的纸张订单吗？",
"l.1 OurContributions",
"我们对论文的顺序进行组合搜索，从而为RRR和WRP等拾取顺序机制产生高效分配。为此，",
"我们研究了通过子模块优化的镜头找到最佳纸张订单的问题。。我们优化了部分纸张序列上",
"的函数，该函数根据拾取序列产生的分配的福利而变化。此函数通常不是子模块化的，但",
"是我们可以通过变量捕获其与子模块化的距离。我们的主要理论结果（定理5.1）是公",
"平划分社区的独立利益，表明一种简单的贪婪方法使该函数最大化到1十的因子。我们",
"2"
] |
{
"arxiv_id": "2108.02126",
"language": "zh",
"page_idx": 1
}
|
||
2108.02126-zh-2
|
2108.02126-zh-2.jpg
|
I WILL HAVE ORDER! OPTIMIZING ORDERS FOR FAIR REVIEWER ASSIGNMENT
称这种方法为贪婪审阅者循环(GRR )。通过优化优先级关系被打破的顺序，该方法也可以
应用于像 WRP 这样的加权拣选序列
尽管我们没有为 FairSequence 算法提供理论上的福利保证，但我们提出了一种启发式方法，该方法针对高福利加权采摘序列进行了优化。FairSequence 通过将审稿人分配给可以获得最高福利增长的任何具有最高优先级的论文，从而自适应地打破优先级顺序。这种方法是快速和简单的实现。FairSequence 因此实现了我们的四个目标：它是一种快速，公平，灵活和福利高效的审阅者分配机制
我们将 GRRR 和 FairSequence 算法与三个真实世界会议数据集上的其他三个最先进的
论文分配框架进行了比较
GRRR 和 FairSequence 不仅是唯一可证明的(W) EF1 方法，FairSequence 比所有数据集上的其他方法快一个数量级。最重要的是，我们表明，在几乎所有公平性指标上， FairSequence 通常比公平分配协议 (Kobren et al., 2019) FairFlow 公平得多。
这项工作扩展了以前的会议记录文件 (Payan & Zick, 2022) 。我们添加了满足可变纸张
需求并大幅改善运行时的算法(WRP 和 FairSequence ),以及更详细的实验分析
# 1.2 Related work
审阅者分配问题已被广泛研究。大多数作品将问题建模为混合整数线性程序，以最大程度地提高审稿人和论文之间的亲和力；多伦多论文匹配系统(TPMS)是该公式最著名的工作$\mathsf E($Charlin et al., 2011; Charlin & Zemel, 2013) 。亲和力通常模拟审稿人专业知识和论文主题之间的一致性，但可以结合其他相关概念，如审稿人出价，利益冲突和作者建议； 许多作品研究这些值是如何生成的，并与我们的作品 (Leyton-Brown et al., 2022; Mysore et al., 2023; Rozenzweig et al., 2023) 正交。亲和力通常被认为是个人和集体层面价值的良好代表，因为亲和力较高的审稿人通常更有资格和对论文感兴趣，从而获得更详细和准确的评论。亲和分数在 TPMS , MicrosoftCMT 或 OpenReview 等系统中普遍可用，标准做法是使用这些亲和分数来计算福利和公平措施 (Stelmakh et al., 2019; Kobren et al., 2019; Jecmen et al., 2020; Leyton-Brown et al., 2022; Shah, 2022; Cousins et al., 2023a)
许多先前的作品都在同行评审中考虑了公平目标，尽管没有一个作品考虑到一项令人羡慕的自由。Hartvigsen et al. (1999)确保每篇论文至少分配一名合格的审阅者。最近的一些方法近似最大化最小纸张分数，或最大化受最小个人分数阈值E(O'Dell et al., 2005; Kobren et al., 2019; Stelmakh et al., 2019) 的分数总和。Aziz et al. (2023) 提出了一种满足核心的审阅者分配算法，确保没有一组论文可以在不减少其总福利的情况下偏离。许多作品研究将论文公平分配给审稿人，使审稿人可以通过竞标来表达对论文的偏好$\overline{\mathrm{E}}($Garg et al., 2010; Lian et al., 2018; Aziz et al., 2019; Tan et al., 2021)。这种设置旨在公平对待审稿人，而不是论文，就像我们的工作一样。$^{4}$其他作品的目标是评审员分配，除了公平或效率之外，还具有其他属性；Long et al.(2013)避免了利益冲突，而 Kou et al.(2015) 和Ahmed et al. (2017)则分配了具有不同兴趣并完全涵盖论文主题的评审员。
公平分配中的一些作品与公平的审稿人分配有关，但是它们在我们的设置中具有重要的局限性。Aziz et al.(2019) 提出了一种算法，该算法尝试输出$W$ -满足约束$W$的 EF1 分配。当$W$ 包括福利的最低阈值时，他们的方法与我们的方法有些相似。但是，只要仍然可以实现$W$ ,他们就不会通过让本地最佳代理进行选择来贪婪地最大化福利，而是让任何代理进行选择。Biswas and Barman (2018)对循环机制进行了修改，该机制在将项目划分为类别并且代理可以从每个类别接收有限数量的项目时分配完整的 EF1 分配，但是每个代理的项目总数是无限的。Dror et al. (2021) 研究拟阵约束下的公平分配，但仅适用于相同或
$4.{\mathrm{~We~briefly~discuss~fairness~to~reviewers~in~Appendix~D~}}.$
3
|
[
[
0.2017822265625,
0.050506591796875,
0.78759765625,
0.050506591796875,
0.78759765625,
0.06439208984375,
0.2017822265625,
0.06439208984375
],
[
0.1478271484375,
0.11871337890625,
0.85107421875,
0.11871337890625,
0.85107421875,
0.1319580078125,
0.1478271484375,
0.1319580078125
],
[
0.1478271484375,
0.1351318359375,
0.416748046875,
0.1351318359375,
0.416748046875,
0.1483154296875,
0.1478271484375,
0.1483154296875
],
[
0.1732177734375,
0.1514892578125,
0.85107421875,
0.1514892578125,
0.85107421875,
0.1654052734375,
0.1732177734375,
0.1654052734375
],
[
0.146240234375,
0.17041015625,
0.8505859375,
0.17041015625,
0.8505859375,
0.184326171875,
0.146240234375,
0.184326171875
],
[
0.1478271484375,
0.1875,
0.85107421875,
0.1875,
0.85107421875,
0.2008056640625,
0.1478271484375,
0.2008056640625
],
[
0.1478271484375,
0.203857421875,
0.85107421875,
0.203857421875,
0.85107421875,
0.2171630859375,
0.1478271484375,
0.2171630859375
],
[
0.1478271484375,
0.2215576171875,
0.450927734375,
0.2215576171875,
0.450927734375,
0.23486328125,
0.1478271484375,
0.23486328125
],
[
0.1732177734375,
0.2392578125,
0.85107421875,
0.2392578125,
0.85107421875,
0.25244140625,
0.1732177734375,
0.25244140625
],
[
0.146240234375,
0.255615234375,
0.343017578125,
0.255615234375,
0.343017578125,
0.26953125,
0.146240234375,
0.26953125
],
[
0.1732177734375,
0.2734375,
0.85107421875,
0.2734375,
0.85107421875,
0.28857421875,
0.1732177734375,
0.28857421875
],
[
0.1478271484375,
0.290283203125,
0.8544921875,
0.290283203125,
0.8544921875,
0.304931640625,
0.1478271484375,
0.304931640625
],
[
0.146240234375,
0.306884765625,
0.76123046875,
0.306884765625,
0.76123046875,
0.3212890625,
0.146240234375,
0.3212890625
],
[
0.1732177734375,
0.324462890625,
0.85107421875,
0.324462890625,
0.85107421875,
0.339111328125,
0.1732177734375,
0.339111328125
],
[
0.1478271484375,
0.34228515625,
0.77685546875,
0.34228515625,
0.77685546875,
0.35595703125,
0.1478271484375,
0.35595703125
],
[
0.143798828125,
0.376953125,
0.3095703125,
0.37744140625,
0.3095703125,
0.392578125,
0.143798828125,
0.391357421875
],
[
0.1478271484375,
0.40283203125,
0.85302734375,
0.40283203125,
0.85302734375,
0.416748046875,
0.1478271484375,
0.416748046875
],
[
0.1478271484375,
0.42041015625,
0.85107421875,
0.42041015625,
0.85107421875,
0.434326171875,
0.1478271484375,
0.434326171875
],
[
0.1478271484375,
0.4375,
0.85302734375,
0.4375,
0.85302734375,
0.45068359375,
0.1478271484375,
0.45068359375
],
[
0.1478271484375,
0.455078125,
0.85302734375,
0.455078125,
0.85302734375,
0.468505859375,
0.1478271484375,
0.468505859375
],
[
0.146240234375,
0.4716796875,
0.85302734375,
0.4716796875,
0.85302734375,
0.48486328125,
0.146240234375,
0.48486328125
],
[
0.1478271484375,
0.4892578125,
0.85107421875,
0.4892578125,
0.85107421875,
0.50244140625,
0.1478271484375,
0.50244140625
],
[
0.1478271484375,
0.505859375,
0.85107421875,
0.505859375,
0.85107421875,
0.51953125,
0.1478271484375,
0.51953125
],
[
0.1446533203125,
0.52294921875,
0.8544921875,
0.521484375,
0.8544921875,
0.5361328125,
0.1446533203125,
0.537109375
],
[
0.146240234375,
0.5390625,
0.8544921875,
0.5390625,
0.8544921875,
0.5537109375,
0.146240234375,
0.5537109375
],
[
0.146240234375,
0.556640625,
0.78759765625,
0.556640625,
0.78759765625,
0.57080078125,
0.146240234375,
0.57080078125
],
[
0.1748046875,
0.57470703125,
0.85302734375,
0.57470703125,
0.85302734375,
0.58837890625,
0.1748046875,
0.58837890625
],
[
0.146240234375,
0.59228515625,
0.85302734375,
0.59228515625,
0.85302734375,
0.60595703125,
0.146240234375,
0.60595703125
],
[
0.160888671875,
0.609375,
0.85302734375,
0.609375,
0.85302734375,
0.62255859375,
0.160888671875,
0.62255859375
],
[
0.1446533203125,
0.625,
0.85302734375,
0.62548828125,
0.85302734375,
0.64013671875,
0.1446533203125,
0.6396484375
],
[
0.1478271484375,
0.64306640625,
0.85302734375,
0.64306640625,
0.85302734375,
0.6572265625,
0.1478271484375,
0.6572265625
],
[
0.146240234375,
0.6591796875,
0.85107421875,
0.6591796875,
0.85107421875,
0.673828125,
0.146240234375,
0.673828125
],
[
0.146240234375,
0.67822265625,
0.85302734375,
0.67822265625,
0.85302734375,
0.69140625,
0.146240234375,
0.69140625
],
[
0.1478271484375,
0.693359375,
0.85107421875,
0.693359375,
0.85107421875,
0.70703125,
0.1478271484375,
0.70703125
],
[
0.146240234375,
0.7109375,
0.85302734375,
0.7109375,
0.85302734375,
0.72607421875,
0.146240234375,
0.72607421875
],
[
0.1478271484375,
0.72900390625,
0.72802734375,
0.72900390625,
0.72802734375,
0.7421875,
0.1478271484375,
0.7421875
],
[
0.176513671875,
0.7470703125,
0.85302734375,
0.7470703125,
0.85302734375,
0.76025390625,
0.176513671875,
0.76025390625
],
[
0.146240234375,
0.76220703125,
0.8544921875,
0.76318359375,
0.8544921875,
0.77783203125,
0.146240234375,
0.7763671875
],
[
0.1446533203125,
0.77880859375,
0.85107421875,
0.77978515625,
0.85107421875,
0.79443359375,
0.1446533203125,
0.79345703125
],
[
0.1478271484375,
0.79736328125,
0.8544921875,
0.79736328125,
0.8544921875,
0.81103515625,
0.1478271484375,
0.81103515625
],
[
0.1446533203125,
0.81298828125,
0.85302734375,
0.814453125,
0.85302734375,
0.8291015625,
0.1446533203125,
0.82763671875
],
[
0.146240234375,
0.83203125,
0.85302734375,
0.83203125,
0.85302734375,
0.84521484375,
0.146240234375,
0.84521484375
],
[
0.146240234375,
0.849609375,
0.85302734375,
0.849609375,
0.85302734375,
0.86279296875,
0.146240234375,
0.86279296875
],
[
0.1502685546875,
0.87548828125,
0.53515625,
0.876953125,
0.53515625,
0.8916015625,
0.1502685546875,
0.89013671875
],
[
0.491943359375,
0.91552734375,
0.5087890625,
0.91552734375,
0.5087890625,
0.9296875,
0.491943359375,
0.9296875
]
] |
[
"IWILL HAVE ORDER! OPTIMIZING ORDERS FOR FAIR REVIEWER ASSIGNMENT",
"称这种方法为贪婪审阅者循环（GRRR）。通过优化优先级关系被打破的顺序，该方法也可以",
"应用于像WRP这样的加权栋选序列",
"尽管我们没有为FairSequence算法提供理论上的福利保证，但我们提出了一种启发式",
"方法，该方法针对高福利加权采摘序列进行了优化。FairSequence通过将审稿人分配给可",
"以获得最高福利增长的任何具有最高优先级的论文，从而自适应地打破优先级顺序。这种",
"方法是快速和简单的实现。FairSequence因此实现了我们的四个目标：它是一种快速，公",
"平，灵活和福利高效的审阅者分配机制",
"我们将GRRR和FairSequence算法与三个真实世界会议数据集上的其他三个最先进的",
"论文分配框架进行了比较",
"GRRR和FairSequence不仅是唯一可证明的（W）EF1方法，FairSequence比所有",
"数据集上的其他方法快一个数量级。最重要的是，我们表明，在几乎所有公平性指标上，",
"FairSequence通常比公平分配协议（Kobrenetal.，2019）FairFlow公平得多。",
"这项工作扩展了以前的会议记录文件（Payan&Zick，2022）。我们添加了满足可变纸张",
"需求并大幅改善运行时的算法（WRP和FairSequence），以及更详细的实验分析",
"1.2 Related work",
"审阅者分配问题已被广泛研究。大多数作品将问题建模为混合整数线性程序，以最大程度",
"地提高审稿人和论文之间的亲和力；多伦多论文匹配系统（TPMS）是该公式最著名的工",
"作E（Charlinetal.，2011；Charlin&Zemel，2013）。亲和力通常模拟审稿人专业知识和论",
"文主题之间的一致性，但可以结合其他相关概念，如审稿人出价，利益冲突和作者建议；",
"许多作品研究这些值是如何生成的，并与我们的作品（Leyton-Brownetal.，2022；Mysore",
"etal.，2023；Rozenzweigetal.，2023）正交。亲和力通常被认为是个人和集体层面价值的良",
"好代表，因为亲和力较高的审稿人通常更有资格和对论文感兴趣，从而获得更详细和准确",
"的评论。亲和分数在TPMS，MicrosoftCMT或OpenReview等系统中普遍可用，标准做法",
"是使用这些亲和分数来计算福利和公平措施（Stelmakhetal.，2019；Kobrenetal.，2019;",
"Jecmen et al., 2020; Leyton-Brown et al., 2022; Shah, 2022; Cousins et al., 2023a)",
"许多先前的作品都在同行评审中考虑了公平目标，尽管没有一个作品考虑到一项令人",
"羡慕的自由。Hartvigsenetal.（1999）确保每篇论文至少分配一名合格的审阅者。最近的",
"-些方法近似最大化最小纸张分数，或最大化受最小个人分数阈值E(O'Delletal.，2005;",
"Kobrenetal.，2019；Stelmakhetal.，2019）的分数总和。Azizetal.（2023）提出了一种满",
"足核心的审阅者分配算法，确保没有一组论文可以在不减少其总福利的情况下偏离。许多",
"作品研究将论文公平分配给审稿人，使审稿人可以通过竞标来表达对论文的偏好E(Garg",
"etal.,2010；Lianetal.,2018；Azizetal.2019；Tanetal.,2021）。这种设置旨在公平对待",
"审稿人，而不是论文，就像我们的工作一样。4其他作品的目标是评审员分配，除了公平或",
"效率之外，还具有其他属性；Longetal.（2013）避免了利益冲突，而Kouetal.（2015）和",
"Ahmedetal.（2017）则分配了具有不同兴趣并完全涵盖论文主题的评审员。",
"公平分配中的一些作品与公平的审稿人分配有关，但是它们在我们的设置中具有重要的",
"局限性。Azizetal.（2019）提出了一种算法，该算法尝试输出W-满足约束W的EF1分",
"配。当W包括福利的最低阈值时，他们的方法与我们的方法有些相似。但是，只要仍然可",
"以实现W，他们就不会通过让本地最佳代理进行选择来贪婪地最大化福利，而是让任何代",
"理进行选择。BiswasandBarman（2018）对循环机制进行了修改，该机制在将项目划分为",
"类别并且代理可以从每个类别接收有限数量的项目时分配完整的EF1分配，但是每个代理",
"的项目总数是无限的。Droretal.（2021）研究拟阵约束下的公平分配，但仅适用于相同或",
"4. We briefly discuss fairness to reviewers in Appendix D",
"3"
] |
{
"arxiv_id": "2108.02126",
"language": "zh",
"page_idx": 2
}
|
||
2108.02126-zh-3
|
2108.02126-zh-3.jpg
|
PAYAN$\&$ZICK
二进制估值，少于四个代理或单个均匀拟阵约束。Caragiannis and Rathi (2023) 表明，当
代理最多可以选择一个项目时，贪婪地选择代理是最大福利挑选序列的 2 近似
最多一个项目的加权嫉妒自由是为具有不平等权利的代理研究的第一个公平概念之一(Benabbou et al., 2020) 。最近在公平分配方面的许多工作都集中在对具有不平等权利的代理的此保证和其他保证上$\operatorname{E}($Babaioff et al., 2021b; Chakraborty et al., 2022; Suksompong $\&$ Teh, 2022,2022,2023; Montanari et al., 2024) 。这项工作的大部分重点是二进制估值或二进制子模块估值的设置。特别相关的是最近的工作，建议使用加权公平划分 (Chakraborty et al., 2021b) 的拾取序列。我们将子模块优化应用于优化循环订单，其灵感来自先前关于子模块估值的公平分配工作E(Benabbou et al., 2020; Babaioff et al., 2021a; Barman & Verma, 2021; Viswanathan & Zick, 2023b) 。先前的工作还研究了近似子模函数的最大化， 尽管没有一个将拟阵约束与类似于我们的E(Das &Kempe, 2011; Gölz & Procaccia, 2019) 的近似子模定义结合在一起。
现有工作显示了最大化 EF1 福利和采摘顺序分配的难度。Aziz et al.(2019) 表明 EF1的福利最大化是 NP 难的，而 Barman et al. (2019) 表明同一问题甚至不是多项式时间近似的。Aziz et al.(2016)表明，对于某些类别的采摘序列(但不是循环法),确定在某个类别的采摘序列下是否可能获得给定福利的问题是 NP 完全的。
最后，我们的 FairSequence 算法使用了一个受扬基交换算法 (Viswanathan & Zick,
2023a, 2023b; Cousins et al., 2023b, 2023c) 的交换图例程启发的子例程。
# 2. Preliminaries
我们将审阅者分配表示为分配不可分割的商品的问题，以论文为代理，以审阅者为商品。我们得到了一组$n$论文$P=\{p_1,\ldots,p_n\}$和一组$m$评论者$R=\{r_1,r_2,\ldots,r_m\}$ 。每篇论文$p$都具有对审阅者$v_p:R\to\mathbb{R}_{\geq0}$的亲和力功能 ,该功能定义了审阅者对论文的适用性
论文通常会收到多个审阅者，因此我们在审阅者集合上定义关联函数。我们假设加法函数，其中对于纸张$p\in P$和子集$S\subseteq R$ , $v_p( S) = \sum _{r\in S}v_p( r)$ 。将审阅者分配给论文是一个有序元组$A=(A_p_1,A_{p_2},\ldots A_{p_n})$ ,其中每个$A_p\subseteq R$是一组分配给论文$p$的不同审阅者。每篇论文$p$都需要$k_p$审阅者。通常我们有一个固定的$k$,这样所有$p$的$k_p=k$ 。如果为每篇论文$p$分配了$k_p$个不同的审阅者 ,则分配完成 (否则不完整)。我们将$A_p$称为纸张$p$的捆绑包，将$v_p(A_p)$称为纸张$p$在$A$下的估值。给定论文$p_i$ ,我们有时会简化符号并编写$v_i$和$A_i$,而不是$v_p_i$和$A_p_i$
每个审阅者$r\in R$在他们可以审阅的论文数量上都有一个上限$u_r$ ,并且可能有下限$l_r$ ,这有助于确保更均匀地分配工作。给定一个分配$A$ ,我们表示将审阅者$r$分配为$c_r^A=\sum_{p\in P}|\{r\}\cap A_p|$的论文数，或者当从上下文中清除$A$时仅为$c_r$
通常，对审阅者分配有其他约束。我们将这些附加约束建模为函数$C$ ,其中$C(p,r,A)\in$ $\{True,False\}$表示在不完全分配$A$的情况下，是否可以将审阅者$r$分配给论文$p$ 。我们将要求这些约束是单调的；如果我们有两个分配$A=(A_1,A_2,\ldots,A_n)$和$A^\prime=(A_1^{\prime},A_2^{\prime},\ldots,A_n^{\prime})$ ,所有$p$都有$A_p\subseteq A_p^\prime$,那么$C(p,r,A)$意味着$C(p,r,A^\prime)$ 。利益冲突表示最简单的此类约束；当论文$p$和审阅者$r$之间存在利益冲突时，$C(p,r,A)=True$。由于$C(p,r,A)$不依赖于$A$来解决利益冲突，因此当没有其他约束应用时，我们将简单地编写$C(p,r)$ 。我们有时会在纸捆中应用更详细的约束 (例如。g。, 任何论文都不能接收来自同一机构的两名审稿人)或跨多个捆绑包(e。g。,一个反共谋约束，要求没有对$p$和$p^\prime$以及审阅者$r$和$r^{\prime}$分配给$p^\prime$,并要求将$p^\prime$ 、$r^\prime$和$r$分配给$r$ 。这些复杂的约束越来越多地用于大型会议(Leyton-Brown et al.,2022)
4
|
[
[
0.441162109375,
0.050506591796875,
0.55224609375,
0.050506591796875,
0.55224609375,
0.06439208984375,
0.441162109375,
0.06439208984375
],
[
0.148681640625,
0.11871337890625,
0.85107421875,
0.11871337890625,
0.85107421875,
0.1319580078125,
0.148681640625,
0.1319580078125
],
[
0.1478271484375,
0.1337890625,
0.75341796875,
0.1337890625,
0.75341796875,
0.147705078125,
0.1478271484375,
0.147705078125
],
[
0.1748046875,
0.1558837890625,
0.8505859375,
0.1558837890625,
0.8505859375,
0.169189453125,
0.1748046875,
0.169189453125
],
[
0.1478271484375,
0.1717529296875,
0.85107421875,
0.1717529296875,
0.85107421875,
0.1849365234375,
0.1478271484375,
0.1849365234375
],
[
0.143798828125,
0.186279296875,
0.8544921875,
0.1881103515625,
0.8544921875,
0.205810546875,
0.143798828125,
0.203857421875
],
[
0.146240234375,
0.20703125,
0.85302734375,
0.20703125,
0.85302734375,
0.2203369140625,
0.146240234375,
0.2203369140625
],
[
0.146240234375,
0.222900390625,
0.85302734375,
0.2235107421875,
0.85302734375,
0.238037109375,
0.146240234375,
0.2374267578125
],
[
0.1446533203125,
0.2392578125,
0.85302734375,
0.240478515625,
0.85302734375,
0.255126953125,
0.1446533203125,
0.25390625
],
[
0.1478271484375,
0.25830078125,
0.85302734375,
0.25830078125,
0.85302734375,
0.271484375,
0.1478271484375,
0.271484375
],
[
0.146240234375,
0.2734375,
0.85302734375,
0.27587890625,
0.85302734375,
0.290283203125,
0.146240234375,
0.28857421875
],
[
0.1478271484375,
0.292236328125,
0.85107421875,
0.292236328125,
0.85107421875,
0.306884765625,
0.1478271484375,
0.306884765625
],
[
0.146240234375,
0.309326171875,
0.36767578125,
0.309326171875,
0.36767578125,
0.322509765625,
0.146240234375,
0.322509765625
],
[
0.1748046875,
0.330078125,
0.85107421875,
0.330078125,
0.85107421875,
0.343505859375,
0.1748046875,
0.343505859375
],
[
0.146240234375,
0.3466796875,
0.85302734375,
0.3466796875,
0.85302734375,
0.35986328125,
0.146240234375,
0.35986328125
],
[
0.1446533203125,
0.3623046875,
0.85302734375,
0.363037109375,
0.85302734375,
0.37744140625,
0.1446533203125,
0.376953125
],
[
0.1446533203125,
0.380126953125,
0.6005859375,
0.380615234375,
0.6005859375,
0.395751953125,
0.1446533203125,
0.39453125
],
[
0.1748046875,
0.401611328125,
0.85107421875,
0.401611328125,
0.85107421875,
0.415283203125,
0.1748046875,
0.415283203125
],
[
0.146240234375,
0.41845703125,
0.7099609375,
0.41845703125,
0.7099609375,
0.431884765625,
0.146240234375,
0.431884765625
],
[
0.1446533203125,
0.460205078125,
0.30810546875,
0.460205078125,
0.30810546875,
0.474853515625,
0.1446533203125,
0.474853515625
],
[
0.146240234375,
0.490478515625,
0.85302734375,
0.490478515625,
0.85302734375,
0.5048828125,
0.146240234375,
0.5048828125
],
[
0.146240234375,
0.50830078125,
0.85302734375,
0.50830078125,
0.85302734375,
0.52294921875,
0.146240234375,
0.52294921875
],
[
0.143798828125,
0.5244140625,
0.81201171875,
0.5234375,
0.81201171875,
0.53857421875,
0.143798828125,
0.5390625
],
[
0.1748046875,
0.54541015625,
0.85302734375,
0.54541015625,
0.85302734375,
0.5595703125,
0.1748046875,
0.5595703125
],
[
0.143798828125,
0.5595703125,
0.8544921875,
0.56103515625,
0.8544921875,
0.58056640625,
0.143798828125,
0.5791015625
],
[
0.1446533203125,
0.5771484375,
0.8544921875,
0.57763671875,
0.8544921875,
0.59521484375,
0.1446533203125,
0.5947265625
],
[
0.1478271484375,
0.5966796875,
0.85302734375,
0.5966796875,
0.85302734375,
0.611328125,
0.1478271484375,
0.611328125
],
[
0.1478271484375,
0.61279296875,
0.85107421875,
0.61279296875,
0.85107421875,
0.62744140625,
0.1478271484375,
0.62744140625
],
[
0.1478271484375,
0.630859375,
0.85302734375,
0.630859375,
0.85302734375,
0.64599609375,
0.1478271484375,
0.64599609375
],
[
0.146240234375,
0.64599609375,
0.4453125,
0.646484375,
0.4453125,
0.6640625,
0.146240234375,
0.66357421875
],
[
0.1748046875,
0.66845703125,
0.85302734375,
0.66845703125,
0.85302734375,
0.681640625,
0.1748046875,
0.681640625
],
[
0.1446533203125,
0.68505859375,
0.85107421875,
0.6845703125,
0.85107421875,
0.69873046875,
0.1446533203125,
0.69970703125
],
[
0.146240234375,
0.7021484375,
0.69140625,
0.7021484375,
0.69140625,
0.7197265625,
0.146240234375,
0.7197265625
],
[
0.1732177734375,
0.7216796875,
0.85302734375,
0.72265625,
0.85302734375,
0.7373046875,
0.1732177734375,
0.736328125
],
[
0.148681640625,
0.7392578125,
0.85107421875,
0.7392578125,
0.85107421875,
0.75244140625,
0.148681640625,
0.75244140625
],
[
0.1478271484375,
0.7568359375,
0.85107421875,
0.7568359375,
0.85107421875,
0.771484375,
0.1478271484375,
0.771484375
],
[
0.1502685546875,
0.7734375,
0.85302734375,
0.77294921875,
0.85302734375,
0.787109375,
0.1502685546875,
0.78857421875
],
[
0.1446533203125,
0.78857421875,
0.85302734375,
0.7890625,
0.85302734375,
0.806640625,
0.1446533203125,
0.80615234375
],
[
0.146240234375,
0.806640625,
0.85302734375,
0.80810546875,
0.85302734375,
0.82275390625,
0.146240234375,
0.8212890625
],
[
0.1478271484375,
0.82568359375,
0.85107421875,
0.82568359375,
0.85107421875,
0.84033203125,
0.1478271484375,
0.84033203125
],
[
0.146240234375,
0.84228515625,
0.8544921875,
0.84228515625,
0.8544921875,
0.85986328125,
0.146240234375,
0.85986328125
],
[
0.146240234375,
0.85986328125,
0.85302734375,
0.85986328125,
0.85302734375,
0.875,
0.146240234375,
0.875
],
[
0.146240234375,
0.87744140625,
0.363525390625,
0.87744140625,
0.363525390625,
0.8916015625,
0.146240234375,
0.8916015625
],
[
0.491943359375,
0.91552734375,
0.5087890625,
0.91552734375,
0.5087890625,
0.9296875,
0.491943359375,
0.9296875
]
] |
[
"PAYAN&ZICK",
"二进制估值，少于四个代理或单个均匀拟阵约束。CaragiannisandRathi（2023）表明，当",
"代理最多可以选择一个项目时，贪地选择代理是最大福利挑选序列的2近似",
"最多一个项目的加权嫉妒自由是为具有不平等权利的代理研究的第一个公平概念之一",
"(Benabbouetal.，2020）。最近在公平分配方面的许多工作都集中在对具有不平等权利的代",
"理的此保证和其他保证上E(Babaioff etal.，2021b；Chakrabortyetal.，2022；Suksompong",
"&Teh，2022，2022，2023；Montanarietal.，2024）。这项工作的大部分重点是二进制估值或",
"二进制子模块估值的设置。特别相关的是最近的工作，建议使用加权公平划分（Chakraborty",
"etal.，2021b）的拾取序列。我们将子模块优化应用于优化循环订单，其灵感来自先前关",
"于子模块估值的公平分配工作E(Benabbouetal.，2020;Babaioffetal.，202la；Barman&",
"Verma，2021；Viswanathan&Zick，2023b）。先前的工作还研究了近似子模函数的最大化，",
"尽管没有一个将拟阵约束与类似于我们的E(Das&Kempe,2011；Golz&Procaccia，2019)",
"的近似子模定义结合在一起。",
"现有工作显示了最大化EF1福利和采摘顺序分配的难度。Azizetal.（2019）表明EF1",
"的福利最大化是NP难的，而Barmanetal.（2019）表明同一问题甚至不是多项式时间近似",
"的。Azizetal.（2016）表明，对于某些类别的采摘序列（但不是循环法），确定在某个类别",
"的采摘序列下是否可能获得给定福利的问题是NP完全的。",
"最后，我们的FairSequence算法使用了一个受扬基交换算法（Viswanathan&Zick，",
"2023a，2023b；Cousinsetal.，2023b，2023c）的交换图例程启发的子例程。",
"2. Preliminaries",
"我们将审阅者分配表示为分配不可分割的商品的问题，以论文为代理，以审阅者为商品。我",
"们得到了一组n论文P={pi,..·,Pn}和一组m评论者R={ri,r2，.·,rm}。每篇论文",
"p都具有对审阅者Up：R→R>0的亲和力功能，该功能定义了审阅者对论文的适用性",
"论文通常会收到多个审阅者，因此我们在审阅者集合上定义关联函数。我们假设加法函",
"数，其中对于纸张pEP和子集 SR，Up(S)=resUp(r）。将审阅者分配给论文是一",
"个有序元组A=（Ap1，Ap2，..Apn），其中每个ApR是一组分配给论文p的不同审阅",
"者。每篇论文p都需要kp审阅者。通常我们有一个固定的k，这样所有p的kpk。如",
"果为每篇论文P分配了kp个不同的审阅者，则分配完成（否则不完整）。我们将Ap称为",
"纸张p的捆绑包，将Up(Ap）称为纸张p在A下的估值。给定论文pi，我们有时会简化符",
"号并编写Ui和Ai，而不是Up和Ap",
"每个审阅者rER在他们可以审阅的论文数量上都有一个上限ur，并且可能有下限",
"l，这有助于确保更均匀地分配工作。给定一个分配A，我们表示将审阅者r分配为",
"cA=pEPKr)nApl的论文数，或者当从上下文中清除A时仅为Cr",
"通常，对审阅者分配有其他约束。我们将这些附加约束建模为函数C，其中C(p,r,A)E",
"True,False}表示在不完全分配A的情况下，是否可以将审阅者r分配给论文p。我们将要",
"求这些约束是单调的；如果我们有两个分配A=（A1,A2，..，An）和A'=（Ai,A2，...，An）",
"，所有p都有ApCA，那么C(p,r,A）意味着C(p,r,A）。利益冲突表示最简单的此类",
"约束；当论文 p和审阅者r之间存在利益冲突时，C(p,r,A)=True。由于 C(p,r,A)不",
"依赖于A来解决利益冲突，因此当没有其他约束应用时，我们将简单地编写C(p,r）。我",
"们有时会在纸捆中应用更详细的约束（例如。g。，任何论文都不能接收来自同一机构的两名",
"审稿人）或跨多个捆绑包（e。g。，一个反共谋约束，要求没有对p和p以及审阅者r和",
"r分配给p，并要求将p、r和r分配给r。这些复杂的约束越来越多地用于大型会议",
"(Leyton-Brown et al., 2022)",
"4"
] |
{
"arxiv_id": "2108.02126",
"language": "zh",
"page_idx": 3
}
|
||
2108.02126-zh-4
|
2108.02126-zh-4.jpg
|
I WILL HAVE ORDER! OPTIMIZING ORDERS FOR FAIR REVIEWER ASSIGNMENT
为了简化表示法，给定一个集合$X$和一个元素$y$ ,我们通常编写$X+y$和$X- y$ , 而
不是$X\cup\{y\}$和$X\setminus\{y\}$
我们现在讨论公平的概念。如果对于所有成对的论文$p$和$p^{\prime }, v_p( A_p) \geq v_p( A_{p^{\prime }})$ ,则分配$A$被认为是令人羡慕的。该标准通常是无法实现的(考虑两篇论文和一个审阅者$r$ 的示例，其上限为$u_r=1$ ),因此我们放宽了标准。如果对于所有成对的论文$p$和$p^{\prime}$, $v_p(A_p)\geq v_p(A_{p^{\prime}})$或$\exists r\in A_{p^{\prime}}$使得$v_p(A_p)\geq v_p(A_{p^{\prime}}-r)$ 。EF1 应该只在所有论文都有相同的要求和$k_p=k$的所有$p$ 。当论文有不同的需求$k_p$时，我们应用加权嫉妒的公平概念 (Chakraborty et al., 2021a)。如果对于所有成对的论文$p$和$p^\prime,\frac{v_{p}(A_{p})}{k_{p}}\geq\frac{v_{p}(A_{p^{\prime}})}{k_{p^{\prime}}}$,则分配$A$被认为是加权的。。类似地，如果对于所有成对的论文$p$和$p^{\prime},\frac{v_{p}(A_{p})}{k_{p}}\geq\frac{\dot{v}_{p}(A_{p^{\prime}})}{k_{p^{\prime}}}$或有一个审阅者$r\in A_{p^{\prime}}$使得$\frac{v_p(A_p)}{k_p}\geq\frac{v_p(A_{p^{\prime}}-r)}{k_{p^{\prime}}}$
分配的功利主义社会福利 (也称为功利主义福利或 USW) $A$是该分配下论文估值的总和。USW 是审阅者分配上下文中的自然目标，并且已在有关该主题的许多先前作品中使用E(Conry et al., 2009; Charlin & Zemel, 2013; Kobren et al., 2019; Stelmakh et al., 2019; Payan & Zick, 2022; Cousins et al., 2023a)
对于拾取序列，我们将纸张$p\in P$上的顺序定义为元组$\mathcal{O}=(S,o)$ ,其中$S\subseteq P$是顺序中的纸张集，$o:S\to\{1,2,\ldots,|S|\}$是$S$上的排列，将纸张映射到位置。令$\Omega(P)$表示$P$ 子集上所有阶数的集合。我们还将引用$\Omega(R)$ ,即$R$子集上的所有订单集，其定义类似于$\Omega(P)$ 。我们稍微滥用表示法，并说如果$p\in S$ 。对于任何$p,p^\prime\in\mathcal{O}$ ,我们说$p\prec_\mathcal{O}p^\prime$当且仅当$o(p)<o(p^\prime)$ 。当$\mathcal{O}$从上下文中清除时，我们编写$p\prec p^\prime$ 。我们可以将订单$\mathcal{O}=(S,o)$ 视为有序列表$[o_1,o_2,\ldots o_{|S|}]$ ,使得$o_l=o^-1(l)$用于所有位置$l$ 。我们使用符号$\mathcal{O}+p$来表示将$p$附加到$\mathcal{O}$末尾的顺序$(S^\prime,o^{\prime})$ 。形式上，$S^\prime = S+ p$ , $o^{\prime }( p^{\prime }) = o( p^{\prime })$用于$p^\prime\in S$ 和$o^{\prime}(p)=|S^{\prime}|$ 。我们用任何函数$o$将空订单写为$\mathcal{O}_\emptyset=(\emptyset,o)$
3. Fair and Efficient Assignment with Reviewer Round Robin
我们首先展示当所有论文都具有相等的需求$k_p= k$ , 并且审阅者没有下限$l_r$时，如何获得 EF1 审阅者分配；我们处理 Section 4 中更一般的情况。我们的算法利用了简单而著名的循环机制。给定一份有序的文件清单，循环赛将轮流进行。在每一轮中，我们按照提供的顺序遍历论文，为每篇论文分配其最高价值的剩余审阅者。对于没有唯一性约束的加性估值，此分配是 EF1,由一个简单的参数 (Caragiannis et al.,2019)。对于任何论文$p$ ,我们将作业分为几轮。与任何论文$p^\prime\succ p$的审稿人相比，论文$p$更喜欢自己的审稿人，如果我们忽略第一轮给予$p^\prime$的审稿人，它更喜欢自己的捆绑包而不是任何代理$p^{\prime}\prec p$的捆绑包
最多必须分配论文的限制$k$不同的审阅者构成了挑战。对 round-robin 的简单修改使我们能够满足基数约束一继续进行$k$轮，然后停止。我们可能会天真地更新循环以满足独特性约束，方法是为每篇论文分配他们还没有的最佳审阅者。但是，Caragiannis et al. (2019)中的参数失败。为了了解原因，假设在一轮中为论文$p$分配了审阅者$r$ 。在下一轮中，$p$可能仍然比任何其他审阅者更喜欢$r$ ,但我们无法分配它。我们将被迫给$p$分配一个更糟糕的审阅者，给另一篇需要$r$“第二副本”的论文。更详细的反例如表 1 所示
我们提出了一种对循环的修改，该修改采用论文顺序$\mathcal{O}$ ,并按顺序$\mathcal{O}$将审阅者分配给论文，从而满足所有约束，并且分配为 EF1。审阅者 RoundRobin 或 RRRE(算法$\boxed{\mathrm{E}}1$ )禁止任何违反证明 EF1 的关键不变量的赋值。此不变量源自加法情况下 EF1 的证明。任何时候我们都会分配一个审阅者，这样就会违反 EF1,我们禁止分配，而是分配一个不同的审阅者。只有当另一篇论文更喜欢该审稿人但该作业被禁止时，才会出现 EF1 违规，要么
5
|
[
[
0.2034912109375,
0.050506591796875,
0.7890625,
0.050506591796875,
0.7890625,
0.06439208984375,
0.2034912109375,
0.06439208984375
],
[
0.1748046875,
0.11871337890625,
0.85302734375,
0.11871337890625,
0.85302734375,
0.1331787109375,
0.1748046875,
0.1331787109375
],
[
0.146240234375,
0.1337890625,
0.340576171875,
0.1351318359375,
0.340576171875,
0.149658203125,
0.146240234375,
0.1483154296875
],
[
0.1748046875,
0.154052734375,
0.85302734375,
0.154052734375,
0.85302734375,
0.1685791015625,
0.1748046875,
0.1685791015625
],
[
0.1478271484375,
0.1717529296875,
0.85302734375,
0.1717529296875,
0.85302734375,
0.1849365234375,
0.1478271484375,
0.1849365234375
],
[
0.146240234375,
0.189453125,
0.85546875,
0.189453125,
0.85546875,
0.20263671875,
0.146240234375,
0.20263671875
],
[
0.146240234375,
0.205810546875,
0.85302734375,
0.205810546875,
0.85302734375,
0.2203369140625,
0.146240234375,
0.2203369140625
],
[
0.146240234375,
0.2215576171875,
0.8544921875,
0.222900390625,
0.8544921875,
0.2374267578125,
0.146240234375,
0.236083984375
],
[
0.1396484375,
0.2374267578125,
0.85888671875,
0.23291015625,
0.85888671875,
0.261474609375,
0.1396484375,
0.26513671875
],
[
0.134033203125,
0.24365234375,
0.86669921875,
0.25390625,
0.865234375,
0.3037109375,
0.13232421875,
0.29345703125
],
[
0.69384765625,
0.25244140625,
0.701171875,
0.25244140625,
0.701171875,
0.255615234375,
0.69384765625,
0.255615234375
],
[
0.75634765625,
0.2474365234375,
0.8037109375,
0.25390625,
0.80126953125,
0.26513671875,
0.75341796875,
0.2587890625
],
[
0.705078125,
0.255126953125,
0.7197265625,
0.255126953125,
0.7197265625,
0.2587890625,
0.705078125,
0.2587890625
],
[
0.72119140625,
0.281005859375,
0.7587890625,
0.281005859375,
0.7587890625,
0.28857421875,
0.72119140625,
0.28857421875
],
[
0.1422119140625,
0.294189453125,
0.50244140625,
0.285888671875,
0.50341796875,
0.311279296875,
0.143798828125,
0.3193359375
],
[
0.358642578125,
0.3037109375,
0.4755859375,
0.3037109375,
0.4755859375,
0.317626953125,
0.358642578125,
0.317626953125
],
[
0.1748046875,
0.32080078125,
0.85107421875,
0.32080078125,
0.85107421875,
0.333984375,
0.1748046875,
0.333984375
],
[
0.146240234375,
0.3359375,
0.85302734375,
0.337158203125,
0.85302734375,
0.3515625,
0.146240234375,
0.35107421875
],
[
0.148681640625,
0.354736328125,
0.85302734375,
0.354736328125,
0.85302734375,
0.36865234375,
0.148681640625,
0.36865234375
],
[
0.146240234375,
0.37255859375,
0.47802734375,
0.37255859375,
0.47802734375,
0.386474609375,
0.146240234375,
0.386474609375
],
[
0.1732177734375,
0.391357421875,
0.85302734375,
0.391357421875,
0.85302734375,
0.406005859375,
0.1732177734375,
0.406005859375
],
[
0.1478271484375,
0.4091796875,
0.85302734375,
0.4091796875,
0.85302734375,
0.42236328125,
0.1478271484375,
0.42236328125
],
[
0.1478271484375,
0.425537109375,
0.85302734375,
0.425537109375,
0.85302734375,
0.439453125,
0.1478271484375,
0.439453125
],
[
0.146240234375,
0.443115234375,
0.85107421875,
0.443115234375,
0.85107421875,
0.458251953125,
0.146240234375,
0.458251953125
],
[
0.1478271484375,
0.460205078125,
0.85107421875,
0.460205078125,
0.85107421875,
0.474853515625,
0.1478271484375,
0.474853515625
],
[
0.146240234375,
0.477783203125,
0.85302734375,
0.477783203125,
0.85302734375,
0.492431640625,
0.146240234375,
0.492431640625
],
[
0.1478271484375,
0.494384765625,
0.85302734375,
0.494384765625,
0.85302734375,
0.5087890625,
0.1478271484375,
0.5087890625
],
[
0.1446533203125,
0.51025390625,
0.6259765625,
0.51123046875,
0.6259765625,
0.52587890625,
0.1446533203125,
0.5244140625
],
[
0.1446533203125,
0.5517578125,
0.74609375,
0.5517578125,
0.74609375,
0.56494140625,
0.1446533203125,
0.56494140625
],
[
0.146240234375,
0.58203125,
0.85302734375,
0.58203125,
0.85302734375,
0.59521484375,
0.146240234375,
0.59521484375
],
[
0.146240234375,
0.5986328125,
0.85107421875,
0.5986328125,
0.85107421875,
0.6123046875,
0.146240234375,
0.6123046875
],
[
0.146240234375,
0.61572265625,
0.85107421875,
0.61572265625,
0.85107421875,
0.62890625,
0.146240234375,
0.62890625
],
[
0.1478271484375,
0.63330078125,
0.85302734375,
0.63330078125,
0.85302734375,
0.646484375,
0.1478271484375,
0.646484375
],
[
0.146240234375,
0.6484375,
0.85302734375,
0.6494140625,
0.85302734375,
0.6640625,
0.146240234375,
0.66357421875
],
[
0.1478271484375,
0.66748046875,
0.85302734375,
0.66748046875,
0.85302734375,
0.68115234375,
0.1478271484375,
0.68115234375
],
[
0.1478271484375,
0.6845703125,
0.42236328125,
0.6845703125,
0.42236328125,
0.69775390625,
0.1478271484375,
0.69775390625
],
[
0.4150390625,
0.68603515625,
0.85107421875,
0.68603515625,
0.85107421875,
0.6962890625,
0.4150390625,
0.6962890625
],
[
0.176513671875,
0.705078125,
0.398681640625,
0.705078125,
0.398681640625,
0.71533203125,
0.176513671875,
0.71533203125
],
[
0.464111328125,
0.70263671875,
0.85302734375,
0.70263671875,
0.85302734375,
0.71630859375,
0.464111328125,
0.71630859375
],
[
0.1478271484375,
0.72021484375,
0.85302734375,
0.72021484375,
0.85302734375,
0.734375,
0.1478271484375,
0.734375
],
[
0.1478271484375,
0.7373046875,
0.8505859375,
0.7373046875,
0.8505859375,
0.751953125,
0.1478271484375,
0.751953125
],
[
0.1478271484375,
0.75390625,
0.85107421875,
0.75390625,
0.85107421875,
0.767578125,
0.1478271484375,
0.767578125
],
[
0.1478271484375,
0.77294921875,
0.84716796875,
0.77294921875,
0.84716796875,
0.7861328125,
0.1478271484375,
0.7861328125
],
[
0.1478271484375,
0.78857421875,
0.81640625,
0.78857421875,
0.81640625,
0.8017578125,
0.1478271484375,
0.8017578125
],
[
0.1748046875,
0.80810546875,
0.85107421875,
0.80810546875,
0.85107421875,
0.82275390625,
0.1748046875,
0.82275390625
],
[
0.146240234375,
0.82470703125,
0.85302734375,
0.82373046875,
0.85302734375,
0.83837890625,
0.146240234375,
0.8388671875
],
[
0.1478271484375,
0.84228515625,
0.85302734375,
0.84228515625,
0.85302734375,
0.85693359375,
0.1478271484375,
0.85693359375
],
[
0.146240234375,
0.85986328125,
0.85302734375,
0.85986328125,
0.85302734375,
0.87353515625,
0.146240234375,
0.87353515625
],
[
0.1478271484375,
0.87744140625,
0.85302734375,
0.87744140625,
0.85302734375,
0.8916015625,
0.1478271484375,
0.8916015625
],
[
0.49267578125,
0.91650390625,
0.50634765625,
0.91650390625,
0.50634765625,
0.92822265625,
0.49267578125,
0.92822265625
]
] |
[
"I WILL HAVE ORDER! OPTIMIZING ORDERS FOR FAIR REVIEWER ASSIGNMENT",
"为了简化表示法，给定一个集合X和一个元素y，我们通常编写X+和X一y，而",
"不是XU(y)和X/(y)",
"我们现在讨论公平的概念。如果对于所有成对的论文p和p，Up(Ap）≥Up(Ap），则",
"分配A被认为是令人羡慕的。该标准通常是无法实现的（考虑两篇论文和一个审阅者r",
"的示例，其上限为ur=1），因此我们放宽了标准。如果对于所有成对的论文p和p，",
"Up(Ap）≥Up(Ap）或rEAp使得Up(Ap）≥Up(Ap-r）。EF1应该只在所有论文都有相",
"同的要求和kp=k的所有p。当论文有不同的需求kp时，我们应用加权嫉妒的公平概",
"念（Chakrabortyetal.，2021a）。如果对于所有成对的论文p和p，\"≥\"，则",
"分配 最认为是加段的。 美似地。如果对于所有成对的论交节和,、1≥会个 成",
"",
"kp",
"",
"kp",
"有一个审阅者EA使得“(≥\"(A-)",
"kp=kp",
"分配的功利主义社会福利（也称为功利主义福利或USW）A是该分配下论文估值的总",
"和。USW是审阅者分配上下文中的自然目标，并且已在有关该主题的许多先前作品中使",
"用E(Conry et al.,2009;Charlin & Zemel,2013;Kobren et al.,2019; Stelmakh et al.,2019;",
"Payan & Zick, 2022; Cousins et al., 2023a)",
"对于拾取序列，我们将纸张pEP上的顺序定义为元组O=（S,o），其中SCP是顺",
"序中的纸张集，o：S→1,2,.，S}是S上的排列，将纸张映射到位置。令2(P）表示P",
"子集上所有阶数的集合。我们还将引用2(R），即R子集上的所有订单集，其定义类似于",
"2(P）。我们稍微滥用表示法，并说如果pES。对于任何p,pEO，我们说p<op当且",
"仅当o(p)<o(p）。当O从上下文中清除时，我们编写p<p。我们可以将订单=(S,o)",
"视为有序列表[o1,02，...0s]，使得01=0-1(U）用于所有位置1。我们使用符号①+p来",
"表示将p附加到O末尾的顺序（S,o）。形式上，S=S+，o(p）=o(p）用于pES",
"和o(p=Sl。我们用任何函数0将空订单写为Oα=(の,o)",
"3. Fair and Eficient Assignment with Reviewer Round Robin",
"我们首先展示当所有论文都具有相等的需求kp=k，并且审阅者没有下限l时，如何获",
"得EF1审阅者分配；我们处理Section4中更一般的情况。我们的算法利用了简单而著名",
"的循环机制。给定一份有序的文件清单，循环赛将轮流进行。在每一轮中，我们按照提供的",
"顺序遍历论文，为每篇论文分配其最高价值的剩余审阅者。对于没有唯一性约束的加性估",
"值，此分配是EF1，由一个简单的参数（Caragiannisetal.，2019）。对于任何论文p，我",
"们将作业分为几轮。与任何论文p>p的审稿人相比，论文p更喜欢自已的审稿人，如果",
"我们忽略第一轮给予p的审稿人，",
"它更喜欢自已的捆绑包而不是任何代理P人P的捆绑包",
"最多必须分配论文的限制",
"盲阅者构成了挑战。对round-robin的简单修改使",
"我们能够满足基数约束一继续进行k轮，然后停止。我们可能会天真地更新循环以满足独",
"特性约束，方法是为每篇论文分配他们还没有的最佳审阅者。但是，Caragiannisetal.",
"(2019）中的参数失败。为了了解原因，假设在一轮中为论文P分配了审阅者r。在下一轮",
"中，p可能仍然比任何其他审阅者更喜欢r，但我们无法分配它。我们将被迫给p分配",
"个更糟糕的审阅者，给另一篇需要r“第二副本”的论文。更详细的反例如表1所示",
"我们提出了一种对循环的修改，该修改采用论文顺序○，并按顺序○将审阅者分配给",
"论文，从而满足所有约束，并且分配为EF1。审阅者RoundRobin或RRRE（算法E1）禁",
"止任何违反证明EF1的关键不变量的赋值。此不变量源自加法情况下EF1的证明。任何",
"时候我们都会分配一个审阅者，这样就会违反EF1，我们禁止分配，而是分配一个不同的",
"审阅者。只有当另一篇论文更喜欢该审稿人但该作业被禁止时，才会出现EF1违规，要么",
"5"
] |
{
"arxiv_id": "2108.02126",
"language": "zh",
"page_idx": 4
}
|
||
2201.08024-zh-0
|
2201.08024-zh-0.jpg
|
# UKD: Debiasing Conversion Rate Estimation via Uncertainty-regularized Knowledge Distillation
Zixuan Xu$^{\frac\pi4}$, Penghui Wei$^{\frac\pi4}$, Weimin Zhang, Shaoguo Liu, Liang Wang and Bo Zheng *
Alibaba Group
$\{$xuzixuan.xzx,wph242967,dutan.zwm,shaoguo.lsg,liangbo.wl,bozheng$\}@$alibaba-inc.com
$ABSTRACT$
<FigureHere>
Figure 1: User click and conversion behaviors in advertising.
In online advertising, conventional post-click conversion rate (CVR) estimation models are trained using clicked samples. However, during online serving the models need to estimate for all impression ads, leading to the sample selection bias (SSB) issue. Intuitively, providing reliable supervision signals for unclicked ads is a feasible way to alleviate the SSB issue. This paper proposes an uncertaintyregularized knowledge distillation (UKD) framework to debias CVR estimation via distilling knowledge from unclicked ads. A teacher model learns click-adaptive representations and produces pseudoconversion labels on unclicked ads as supervision signals. Then a student model is trained on both clicked and unclicked ads with knowledge distillation, performing uncertainty modeling to alleviate the inherent noise in pseudo-labels. Experiments on billionscale datasets show that UKD outperforms previous debiasing methods. Online results verify that UKD achieves significant improvements.
# $\csc\csc$coNCEPTS $\cdot$Information systems$\to0$nline advertising.
$\mathbf{KEYWORDS}$
CVR Estimation, Debiasing, Distillation, Uncertainty Modeling
ACM Reference Format:
Zixuan Xu $^{\frac{1}{4}}$,Penghui Wei$^{\frac{1}{4}}$,Weimin Zhang, Shaoguo Liu, Liang Wang and Bo Zheng " .2022. UKD: Debiasing Conversion Rate Estimation via Uncertainty-regularized Knowledge Distillation . In Proceedings ofthe ACM Web Conference 2022 (WWW'22), April 25-29, 2022, Virtual Event, Lyon, $France.$ ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3485447. 3512081
图 1 显示了在线广告中的用户点击和转化行为。如果用户点击印象广告，他们将到达显示广告详细信息的着陆页，然后用户可能会采取转化操作或不。显然，只有点击的广告才有点击后转化的标签，而对于用户没有点击的广告，我们永远不会知道是否会发生点击后转化的动作。由于未点击广告缺乏地面实况标签，传统的 CVR 估计模型通常仅使用点击广告进行训练，但这些模型需要在在线投放期间预测整个印象广告(包括点击和未点击广告)的 CVR。训练空间之间存在差距的问题(i。e。, 单击空间) 和推理空间(i。e。, 整个印象空间) 称为样本选择偏差 (SSB) [30] 。这些模型可能偏向于未点击的广告， 因为他们的培训程序没有从未点击的广告中学到很多知识
探讨 SSB 问题的代表性方法可分为两类：1) 基于$\overline{\mathbb{E}}[16,27]$ 的辅助任务学习，以及2) 基于反事实学习的方法$\mathbb{E}[9,31]$ 。Ma et al.[16] 建议合并两个可以在整个印象空间中训练的辅助任务，以间接学习整个空间 CVR 估计器。但是，对于未点击的广告，该模型倾向于将预测的 CVR 分数优化为零(有关证明， 请参见 §$\underline{\mathrm{E}}$2.3.3 ), 而其实际转换标签为“未知”。$\underline{\mathrm{E}}$Zhang et al. [31]采用反事实学习来产生理论上无偏的 CVR 估计器，但是CVR 任务的训练过程并未明确考虑未点击的广告。总之，目前的方法基本上不能消除未点击广告的 CVR 模型，CVR 估计中的 SSB 问题仍然是一个悬而未决的问题
要学习可以准确估计所有印象广告的 CVR 的整个空间 CVR 模型，一种可行的方法是为未点击的广告提供可靠的伪转换标签作为监督信号。通过这种方式，CVR 模型的训练过程可以明确地利用点击广告(带有来自日志的标签)和未点击广告(带有伪标签)。因此，与仅在点击广告上训练的模型相比，这些模型可以受益于未点击广告的学习。为了实现这一目标，关键是当我们只能访问点击广告的真实标签时，如何为未点击广告生成可靠的伪转换标签，以及如何学习具有真实标签和伪标签的准确的整个空间 CVR 估计器。对于前者，考虑到点击和未点击广告的数据分布之间存在差异(这会导致 SSB 问题)$\boxed{\mathbb{E} }[ 31]$ ,我们建议将伪转换标签生成表述为无监督域自适应问题。点击/取消点击空间被视为源/目标域，我们的目标是根据标记的点击广告(源域)为未标记的未点击广告(目标域) 获取伪标签。对于后者，考虑到未点击广告的伪标签的置信度不如点击广告的真实标签，我们建议通过在训练过程中对伪标签的不确定性进行建模来减少伪标签中存在的固有噪声的负面影响
出于上述考虑，在本文中，我们提出了U 不确定性正则化的 KnowledgeD 蒸馏 (UKD), 其目的是通过从未点击的广告中提取知识来消除 CVR 估计的偏差。UKD 的整体工作流程包
# 1 INTRODUCTION
在在线广告系统中，点击后转化率(CVR)估计是预测广告点击事件后的转化概率，而预测的 CVR 得分是许多应用程序(例如排名程序和智能出价)中的关键因素。在许多营销场景中，转化是广告主的最终目的，因而 CVR 估算起着重要作用
$^{\ddagger}$ Co-first authorship. * Corresponding author.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org.
$WWW^{\prime}22,April~25-29,~2022,~Virtual~Event,~Lyon,~France$
$\odot$ 2022 Association for Computing Machinery. ACM ISBN 978-1-4503-9096-5/22/04....815.00 https://doi.org/10.1145/3485447.3512081
|
[
[
0.1895751953125,
0.1029052734375,
0.80810546875,
0.10418701171875,
0.80810546875,
0.125,
0.1895751953125,
0.12371826171875
],
[
0.1953125,
0.1307373046875,
0.8056640625,
0.1307373046875,
0.8056640625,
0.1483154296875,
0.1953125,
0.1483154296875
],
[
0.148681640625,
0.1641845703125,
0.84716796875,
0.1654052734375,
0.84716796875,
0.18310546875,
0.148681640625,
0.1817626953125
],
[
0.449462890625,
0.18115234375,
0.54833984375,
0.18310546875,
0.54833984375,
0.1976318359375,
0.449462890625,
0.1956787109375
],
[
0.2017822265625,
0.1988525390625,
0.79345703125,
0.1988525390625,
0.79345703125,
0.212158203125,
0.2017822265625,
0.212158203125
],
[
0.0882568359375,
0.2235107421875,
0.1829833984375,
0.2235107421875,
0.1829833984375,
0.23486328125,
0.0882568359375,
0.23486328125
],
[
0.71826171875,
0.2266845703125,
0.75244140625,
0.2266845703125,
0.75244140625,
0.236083984375,
0.71826171875,
0.236083984375
],
[
0.78515625,
0.2266845703125,
0.82275390625,
0.2197265625,
0.82861328125,
0.2342529296875,
0.7900390625,
0.2412109375
],
[
0.87744140625,
0.2247314453125,
0.88818359375,
0.2247314453125,
0.88818359375,
0.236083984375,
0.87744140625,
0.236083984375
],
[
0.5947265625,
0.240478515625,
0.615234375,
0.240478515625,
0.615234375,
0.2474365234375,
0.5947265625,
0.2474365234375
],
[
0.70751953125,
0.238037109375,
0.76318359375,
0.24560546875,
0.7607421875,
0.2568359375,
0.705078125,
0.25
],
[
0.78759765625,
0.246826171875,
0.82861328125,
0.238037109375,
0.833984375,
0.25390625,
0.79345703125,
0.26318359375
],
[
0.8759765625,
0.246826171875,
0.88818359375,
0.246826171875,
0.88818359375,
0.25830078125,
0.8759765625,
0.25830078125
],
[
0.0849609375,
0.255126953125,
0.48193359375,
0.255126953125,
0.48193359375,
0.268310546875,
0.0849609375,
0.268310546875
],
[
0.08416748046875,
0.2412109375,
0.486083984375,
0.2412109375,
0.486083984375,
0.255126953125,
0.08416748046875,
0.255126953125
],
[
0.58251953125,
0.264404296875,
0.62158203125,
0.264404296875,
0.62158203125,
0.274658203125,
0.58251953125,
0.274658203125
],
[
0.654296875,
0.268310546875,
0.67236328125,
0.268310546875,
0.67236328125,
0.2822265625,
0.654296875,
0.2822265625
],
[
0.701171875,
0.264404296875,
0.76123046875,
0.266357421875,
0.76123046875,
0.27587890625,
0.701171875,
0.2734375
],
[
0.8544921875,
0.270751953125,
0.90869140625,
0.270751953125,
0.90869140625,
0.281005859375,
0.8544921875,
0.281005859375
],
[
0.7861328125,
0.271484375,
0.826171875,
0.26318359375,
0.8310546875,
0.279052734375,
0.79150390625,
0.287353515625
],
[
0.08660888671875,
0.26953125,
0.4794921875,
0.26953125,
0.4794921875,
0.27978515625,
0.08660888671875,
0.27978515625
],
[
0.69140625,
0.27587890625,
0.77294921875,
0.27587890625,
0.77294921875,
0.28271484375,
0.69140625,
0.28271484375
],
[
0.0849609375,
0.28271484375,
0.4794921875,
0.28271484375,
0.4794921875,
0.296630859375,
0.0849609375,
0.296630859375
],
[
0.51708984375,
0.290283203125,
0.576171875,
0.290283203125,
0.576171875,
0.300537109375,
0.51708984375,
0.300537109375
],
[
0.62841796875,
0.290283203125,
0.69677734375,
0.290283203125,
0.69677734375,
0.300537109375,
0.62841796875,
0.300537109375
],
[
0.76318359375,
0.290283203125,
0.84716796875,
0.290283203125,
0.84716796875,
0.29736328125,
0.76318359375,
0.29736328125
],
[
0.0849609375,
0.296630859375,
0.48046875,
0.296630859375,
0.48046875,
0.31005859375,
0.0849609375,
0.31005859375
],
[
0.87158203125,
0.291748046875,
0.890625,
0.291748046875,
0.890625,
0.298583984375,
0.87158203125,
0.298583984375
],
[
0.51708984375,
0.3037109375,
0.91259765625,
0.304931640625,
0.91259765625,
0.318115234375,
0.51708984375,
0.317626953125
],
[
0.0849609375,
0.311279296875,
0.48046875,
0.311279296875,
0.48046875,
0.324462890625,
0.0849609375,
0.324462890625
],
[
0.08416748046875,
0.324462890625,
0.484375,
0.323974609375,
0.484375,
0.337158203125,
0.08416748046875,
0.33837890625
],
[
0.53515625,
0.32763671875,
0.91259765625,
0.32763671875,
0.91259765625,
0.341552734375,
0.53515625,
0.341552734375
],
[
0.0849609375,
0.33837890625,
0.48193359375,
0.33837890625,
0.48193359375,
0.3515625,
0.0849609375,
0.3515625
],
[
0.51953125,
0.343505859375,
0.91259765625,
0.343505859375,
0.91259765625,
0.356689453125,
0.51953125,
0.356689453125
],
[
0.0849609375,
0.3515625,
0.48046875,
0.3515625,
0.48046875,
0.365478515625,
0.0849609375,
0.365478515625
],
[
0.521484375,
0.359130859375,
0.9111328125,
0.359130859375,
0.9111328125,
0.369384765625,
0.521484375,
0.369384765625
],
[
0.08660888671875,
0.365478515625,
0.48193359375,
0.365478515625,
0.48193359375,
0.37890625,
0.08660888671875,
0.37890625
],
[
0.51953125,
0.37255859375,
0.9111328125,
0.37255859375,
0.9111328125,
0.38330078125,
0.51953125,
0.38330078125
],
[
0.0882568359375,
0.380615234375,
0.48046875,
0.380615234375,
0.48046875,
0.391357421875,
0.0882568359375,
0.391357421875
],
[
0.5185546875,
0.385009765625,
0.91259765625,
0.385009765625,
0.91259765625,
0.3984375,
0.5185546875,
0.3984375
],
[
0.0849609375,
0.393310546875,
0.483642578125,
0.393310546875,
0.483642578125,
0.4072265625,
0.0849609375,
0.4072265625
],
[
0.5185546875,
0.3984375,
0.91259765625,
0.3984375,
0.91259765625,
0.412353515625,
0.5185546875,
0.412353515625
],
[
0.08660888671875,
0.4091796875,
0.48046875,
0.4091796875,
0.48046875,
0.419189453125,
0.08660888671875,
0.419189453125
],
[
0.51953125,
0.412353515625,
0.9111328125,
0.41357421875,
0.9111328125,
0.4248046875,
0.51953125,
0.423583984375
],
[
0.0849609375,
0.421630859375,
0.486083984375,
0.421630859375,
0.486083984375,
0.43505859375,
0.0849609375,
0.43505859375
],
[
0.5185546875,
0.425537109375,
0.91259765625,
0.425537109375,
0.91259765625,
0.439453125,
0.5185546875,
0.439453125
],
[
0.0849609375,
0.43505859375,
0.48193359375,
0.43505859375,
0.48193359375,
0.448974609375,
0.0849609375,
0.448974609375
],
[
0.5185546875,
0.43994140625,
0.91259765625,
0.43994140625,
0.91259765625,
0.453857421875,
0.5185546875,
0.453857421875
],
[
0.0849609375,
0.45068359375,
0.12744140625,
0.45068359375,
0.12744140625,
0.462158203125,
0.0849609375,
0.462158203125
],
[
0.5185546875,
0.453857421875,
0.91259765625,
0.453857421875,
0.91259765625,
0.46728515625,
0.5185546875,
0.46728515625
],
[
0.5185546875,
0.46728515625,
0.88623046875,
0.46728515625,
0.88623046875,
0.48095703125,
0.5185546875,
0.48095703125
],
[
0.08660888671875,
0.47607421875,
0.221435546875,
0.47607421875,
0.221435546875,
0.4892578125,
0.08660888671875,
0.4892578125
],
[
0.53515625,
0.481689453125,
0.91259765625,
0.481689453125,
0.91259765625,
0.49560546875,
0.53515625,
0.49560546875
],
[
0.08416748046875,
0.492431640625,
0.38232421875,
0.494384765625,
0.38232421875,
0.50830078125,
0.08416748046875,
0.505859375
],
[
0.5185546875,
0.49560546875,
0.91259765625,
0.49560546875,
0.91259765625,
0.5087890625,
0.5185546875,
0.5087890625
],
[
0.51708984375,
0.5087890625,
0.91259765625,
0.5087890625,
0.91259765625,
0.52294921875,
0.51708984375,
0.52294921875
],
[
0.08660888671875,
0.52294921875,
0.188720703125,
0.52294921875,
0.188720703125,
0.5341796875,
0.08660888671875,
0.5341796875
],
[
0.5185546875,
0.5234375,
0.91259765625,
0.5234375,
0.91259765625,
0.537109375,
0.5185546875,
0.537109375
],
[
0.08416748046875,
0.53857421875,
0.463134765625,
0.5390625,
0.463134765625,
0.5537109375,
0.08416748046875,
0.55322265625
],
[
0.5185546875,
0.537109375,
0.9111328125,
0.537109375,
0.9111328125,
0.55029296875,
0.5185546875,
0.55029296875
],
[
0.51953125,
0.55029296875,
0.91259765625,
0.55029296875,
0.91259765625,
0.564453125,
0.51953125,
0.564453125
],
[
0.08660888671875,
0.56103515625,
0.2296142578125,
0.56103515625,
0.2296142578125,
0.5712890625,
0.08660888671875,
0.5712890625
],
[
0.5185546875,
0.56494140625,
0.91259765625,
0.56494140625,
0.91259765625,
0.5791015625,
0.5185546875,
0.5791015625
],
[
0.0849609375,
0.57275390625,
0.48193359375,
0.57275390625,
0.48193359375,
0.5859375,
0.0849609375,
0.5859375
],
[
0.5185546875,
0.5791015625,
0.91259765625,
0.5791015625,
0.91259765625,
0.59228515625,
0.5185546875,
0.59228515625
],
[
0.0849609375,
0.58544921875,
0.48046875,
0.58544921875,
0.48046875,
0.5986328125,
0.0849609375,
0.5986328125
],
[
0.51708984375,
0.59130859375,
0.91259765625,
0.59033203125,
0.91259765625,
0.60498046875,
0.51708984375,
0.60595703125
],
[
0.0849609375,
0.5986328125,
0.48193359375,
0.5986328125,
0.48193359375,
0.6123046875,
0.0849609375,
0.6123046875
],
[
0.08660888671875,
0.611328125,
0.48046875,
0.611328125,
0.48046875,
0.625,
0.08660888671875,
0.625
],
[
0.5185546875,
0.60595703125,
0.76953125,
0.60595703125,
0.76953125,
0.619140625,
0.5185546875,
0.619140625
],
[
0.0849609375,
0.62353515625,
0.48193359375,
0.62353515625,
0.48193359375,
0.63720703125,
0.0849609375,
0.63720703125
],
[
0.53515625,
0.619140625,
0.9140625,
0.619140625,
0.9140625,
0.63330078125,
0.53515625,
0.63330078125
],
[
0.0849609375,
0.63720703125,
0.131591796875,
0.63720703125,
0.131591796875,
0.6484375,
0.0849609375,
0.6484375
],
[
0.51708984375,
0.6318359375,
0.91259765625,
0.63330078125,
0.91259765625,
0.64794921875,
0.51708984375,
0.646484375
],
[
0.51953125,
0.64794921875,
0.91259765625,
0.64794921875,
0.91259765625,
0.6611328125,
0.51953125,
0.6611328125
],
[
0.11517333984375,
0.66650390625,
0.2607421875,
0.66650390625,
0.2607421875,
0.6796875,
0.11517333984375,
0.6796875
],
[
0.521484375,
0.66357421875,
0.9111328125,
0.66357421875,
0.9111328125,
0.673828125,
0.521484375,
0.673828125
],
[
0.5185546875,
0.67529296875,
0.9140625,
0.67529296875,
0.9140625,
0.689453125,
0.5185546875,
0.689453125
],
[
0.0882568359375,
0.66748046875,
0.1168212890625,
0.66748046875,
0.1168212890625,
0.6787109375,
0.0882568359375,
0.6787109375
],
[
0.08660888671875,
0.68310546875,
0.4794921875,
0.68310546875,
0.4794921875,
0.6962890625,
0.08660888671875,
0.6962890625
],
[
0.5185546875,
0.68798828125,
0.9140625,
0.68798828125,
0.9140625,
0.7021484375,
0.5185546875,
0.7021484375
],
[
0.08660888671875,
0.69775390625,
0.48046875,
0.69775390625,
0.48046875,
0.7109375,
0.08660888671875,
0.7109375
],
[
0.51708984375,
0.7021484375,
0.91259765625,
0.70263671875,
0.91259765625,
0.71630859375,
0.51708984375,
0.71533203125
],
[
0.08660888671875,
0.7119140625,
0.48046875,
0.7119140625,
0.48046875,
0.72607421875,
0.08660888671875,
0.72607421875
],
[
0.521484375,
0.71826171875,
0.9111328125,
0.71826171875,
0.9111328125,
0.72900390625,
0.521484375,
0.72900390625
],
[
0.0849609375,
0.724609375,
0.4755859375,
0.72607421875,
0.4755859375,
0.7392578125,
0.0849609375,
0.73779296875
],
[
0.51953125,
0.73095703125,
0.91259765625,
0.73095703125,
0.91259765625,
0.744140625,
0.51953125,
0.744140625
],
[
0.0882568359375,
0.751953125,
0.3056640625,
0.75390625,
0.3056640625,
0.767578125,
0.0882568359375,
0.76513671875
],
[
0.51953125,
0.7470703125,
0.91259765625,
0.7470703125,
0.91259765625,
0.7568359375,
0.51953125,
0.7568359375
],
[
0.51953125,
0.76025390625,
0.9111328125,
0.76025390625,
0.9111328125,
0.77099609375,
0.51953125,
0.77099609375
],
[
0.51708984375,
0.77099609375,
0.9140625,
0.771484375,
0.9140625,
0.78515625,
0.51708984375,
0.7841796875
],
[
0.08660888671875,
0.7822265625,
0.48046875,
0.7822265625,
0.48046875,
0.79248046875,
0.08660888671875,
0.79248046875
],
[
0.0849609375,
0.79150390625,
0.48193359375,
0.79150390625,
0.48193359375,
0.8046875,
0.0849609375,
0.8046875
],
[
0.51953125,
0.787109375,
0.91259765625,
0.787109375,
0.91259765625,
0.79736328125,
0.51953125,
0.79736328125
],
[
0.08660888671875,
0.80322265625,
0.48046875,
0.80322265625,
0.48046875,
0.81298828125,
0.08660888671875,
0.81298828125
],
[
0.51708984375,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.51708984375,
0.81298828125
],
[
0.0849609375,
0.81103515625,
0.48193359375,
0.81103515625,
0.48193359375,
0.82470703125,
0.0849609375,
0.82470703125
],
[
0.5185546875,
0.81298828125,
0.91259765625,
0.81298828125,
0.91259765625,
0.8271484375,
0.5185546875,
0.8271484375
],
[
0.08416748046875,
0.8212890625,
0.483642578125,
0.82080078125,
0.483642578125,
0.833984375,
0.08416748046875,
0.83544921875
],
[
0.08660888671875,
0.83349609375,
0.48046875,
0.83349609375,
0.48046875,
0.84326171875,
0.08660888671875,
0.84326171875
],
[
0.5185546875,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.5185546875,
0.8388671875
],
[
0.0882568359375,
0.84326171875,
0.378173828125,
0.84326171875,
0.378173828125,
0.85400390625,
0.0882568359375,
0.85400390625
],
[
0.51953125,
0.84228515625,
0.60693359375,
0.84228515625,
0.60693359375,
0.85400390625,
0.51953125,
0.85400390625
],
[
0.0882568359375,
0.85400390625,
0.34716796875,
0.85400390625,
0.34716796875,
0.8642578125,
0.0882568359375,
0.8642578125
],
[
0.53515625,
0.85498046875,
0.91259765625,
0.85498046875,
0.91259765625,
0.86865234375,
0.53515625,
0.86865234375
],
[
0.08660888671875,
0.86572265625,
0.30224609375,
0.86572265625,
0.30224609375,
0.87548828125,
0.08660888671875,
0.87548828125
],
[
0.08660888671875,
0.875,
0.293212890625,
0.875,
0.293212890625,
0.88525390625,
0.08660888671875,
0.88525390625
],
[
0.5185546875,
0.869140625,
0.91259765625,
0.869140625,
0.91259765625,
0.88330078125,
0.5185546875,
0.88330078125
],
[
0.0849609375,
0.88525390625,
0.27294921875,
0.8837890625,
0.27294921875,
0.89599609375,
0.0849609375,
0.896484375
],
[
0.51953125,
0.88330078125,
0.91259765625,
0.88330078125,
0.91259765625,
0.896484375,
0.51953125,
0.896484375
]
] |
[
"UKD: Debiasing Conversion Rate Estimation via",
"Uncertainty-regularized Knowledge Distillation",
"Zixuan Xu , Penghui Wei * , Weimin Zhang, Shaoguo Liu, Liang Wang and Bo Zheng ",
"Alibaba Group",
"(xuzixuan.xzx,wph242967,dutan.zwm,shaoguo.lsg,liangbo.wl,bozheng)@alibaba-inc.com",
"ABSTRACT",
"Convert",
"ST",
"1",
"ClicK",
"Not Convert",
"X",
"0",
"estimation models are trained using clicked samples.However, dur-",
"In online advertising, conventional post-click conversion rate (CVR)",
"Unclick",
"X",
"Convertornot",
"\"unknown\"",
"",
"ingonlineservingthemodelsneedtoestimateforall impression",
"iftheuserclickson?",
"ads, leading to the sample selection bias (SSB) issue. Intuitively",
" Impression Ad",
"Ad Landing Page",
"Post-click Conversior",
"providing reliable supervision signals for unclicked ads is a feasible",
"Labe",
"Figure l:User click and conversion behaviors in advertising.",
"way to alleviate the SSB issue. This paper proposes an uncertainty",
"regularizedknowledge distillation (UKD)framework todebias CVR",
"图1显示了在线广告中的用户点击和转化行为。如果用户",
"estimation via distilling knowledge from unclicked ads. A teacher",
"点击印象广告，他们将到达显示广告详细信息的着陆页，然后",
"model learns click-adaptive representations and produces pseudo",
"用户可能会采取转化操作或不。显然，只有点击的广告才有点",
"conversion labels on unclicked ads as supervision signals. Then a",
"击后转化的标签，而对于用户没有点击的广告，我们永远不会",
"studentmodel is trained on both clicked and unclicked adswith",
"知道是否会发生点击后转化的动作。由于未点击广告缺乏地",
"knowledge distillation, performing uncertainty modeling to alle-",
"面实况标签，传统的CVR估计模型通常仅使用点击广告进行",
"viatethe inherentnoise inpseudo-labels.Experiments onbillion-",
"训练，但这些模型需要在在线投放期间预测整个印象广告（包",
"scale datasets showthatUKD outperforms previous debiasingmeth-",
"括点击和未点击广告）的CVR。训练空间之间存在差距的问",
"ods. Online results verify that UKD achieves significant improve-",
"题(i。e。，单击空间)和推理空间(i。e。，整个印象空间)称为",
"ments.",
"样本选择偏差(SSB)[30]。这些模型可能偏向于未点击的广告，",
"因为他们的培训程序没有从未点击的广告中学到很多知识",
"CCS CONCEPTS",
"探讨SSB问题的代表性方法可分为两类：1基于E[16,27]",
".Information systems→Onlineadvertising",
"的辅助任务学习，以及2）基于反事实学习的方法国[9,31]。Ma",
"etal.[16]建议合并两个可以在整个印象空间中训练的辅助任",
"KEYWORDS",
"务，以间接学习整个空间CVR估计器。但是，对于未点击的",
"CVR Estimation, Debiasing, Distillation, Uncertainty Modeling",
"广告，该模型倾向于将预测的CVR分数优化为零（有关证明",
"请参见sE2.3.3），而其实际转换标签为“未知\"。EZhangetal.",
"ACMReferenceFormat:",
"[31]采用反事实学习来产生理论上无偏的CVR估计器，但是",
"Zixuan Xu +, Penghui Wei +, Weimin Zhang, Shaoguo Liu, Liang Wang",
"CVR任务的训练过程并未明确考虑未点击的广告。总之，目前",
"and Bo Zheng 2022. UKD: Debiasing Conversion Rate Estimation via",
"的方法基本上不能消除未点击广告的CVR模型，CVR估计中",
"Uncertainty-regularized Knowledge Distillation.In Proceedings ofthe ACM",
"Web Conference 2022 (wwW 22), April 25-29, 2022, Virtual Event, Lyon",
"的SSB问题仍然是一个悬而未决的问题",
"France.ACM,New York,NY,USA, 9 pages. https://doi.org/10.1145/3485447",
"要学习可以准确估计所有印象广告的CVR的整个空间CVR",
"3512081",
"模型，一种可行的方法是为未点击的广告提供可靠的伪转换",
"标签作为监督信号。通过这种方式，CVR模型的训练过程可",
"INTRODUCTION",
"以明确地利用点击广告（带有来自日志的标签）和未点击广告",
"（带有伪标签）。因此，与仅在点击广告上训练的模型相比，这",
"1",
"在在线广告系统中，点击后转化率（CVR）估计是预测广告",
"些模型可以受益于未点击广告的学习。为了实现这一目标，关",
"点击事件后的转化概率，而预测的CVR得分是许多应用程序",
"键是当我们只能访问点击广告的真实标签时，如何为未点击",
"（例如排名程序和智能出价）中的关键因素。在许多营销场景",
"广告生成可靠的伪转换标签，以及如何学习具有真实标签和",
"中，转化是广告主的最终目的，因而CVR估算起着重要作用",
"伪标签的准确的整个空间CVR估计器。对于前者，考虑到点",
" Co-first authorship.Corresponding author",
"击和未点击广告的数据分布之间存在差异（这会导致SSB问",
"题E[31]，我们建议将伪转换标签生成表述为无监督域自适",
"应问题。点击/取消点击空间被视为源/目标域，我们的目标是",
"Permission to make digital or hard copies of all or part of this work for personal or",
"classroom use is granted without fee provided that copies are not made or distributed",
"根据标记的点击广告（源域）为未标记的未点击广告（目标",
"for profit or commercial advantage and that copies bear this notice and the full cita",
"域）获取伪标签。对于后者，考虑到未点击广告的伪标签的置",
"tion on the first page. Copyrights for components of this work owned by others than",
"信度不如点击广告的真实标签，我们建议通过在训练过程中",
"ACM must be honored.Abstracting with credit ispermitted.Tocopyotherwise,orre",
"publish, to post on servers or to redistribute to lists, requires prior specific permission",
"对伪标签的不确定性进行建模来减少伪标签中存在的固有噪",
"and/ora fee.Requestpermissions from permissions@acm.org.",
"声的负面影响",
"Www22,April 25-29,2022,Virtual Event,Lyon,France",
"出于上述考虑，在本文中，我们提出了U不确定性正则化",
"2022Associationfor Computing Machinery",
"ACMISBN978-1-4503-9096-5/22/04..$15.00",
"的KnowledgeD蒸馏（UKD），其目的是通过从未点击的广告",
"https://doi.org/10.1145/3485447.3512081",
"中提取知识来消除CVR估计的偏差。UKD的整体工作流程包"
] |
{
"arxiv_id": "2201.08024",
"language": "zh",
"page_idx": 0
}
|
||
2201.08024-zh-1
|
2201.08024-zh-1.jpg
|
含一个点击自适应教师模型，该模型为未点击的广告生成伪转换标签，以及一个不确定性正则化的学生模型，该模型可以有效地提炼教师学到的未点击广告中的知识。具体来说，为了产生未点击广告的监督信号，教师使用领域自适应学习印象广告的点击自适应表示，其对未点击广告的预测 CVR 得分被作为其伪转换标签。然后，学生可以从点击广告(带有真实标签)和未点击广告(带有老师的伪标签)中学习，并且还可以对伪标签进行不确定性估计，以减轻其中的固有噪声。对于每个未点击的广告，我们的学生估计其伪标签的不确定性，并在训练期间动态调整其 CVR 损失的权重，以削弱其负面影响。在十亿级数据集上的实验结果表明，UKD 优于以前的最新方法。我们已经在阿里巴巴广告平台部署了UKD ,在线结果验证了 UKD 实现了显著的改进。这项工作的主要贡献是：
$\bullet$ We propose uncertainty-regularized knowledge distillation
(UKD) to debias CVR models via learning from unclicked ads. It employs a click-adaptive teacher to generate pseudoconversion labels for unclicked ads, and then trains a student model that takes both clicked and unclicked ads into account.
$\bullet$ Our student model performs uncertainty estimation to pseudo-
labels generated by the teacher, alleviating the inherent noise
to reduce the negative impact during distillation.
$\bullet$ Experimental results on public and large-scale production
datasets show that UKD outperforms the state-of-the-art methods. Online experiments further verify that it achieves significantly improvements on core metrics
# 2 РREREQUISITES 2.1 $\textbf{Problem Definition}$
在在线广告系统中，我们可以记录用户对印象广告的反馈，以训练用于估计 CTR (点击率), CVR 和 CTCVR (点击率)的模型。令$\mathcal{D}=\left\{(x,y_{click},y_{conv},y_{pv-conv})\right\}$表示所收集的印象广告数据集。对于每个印象样本，$x$表示特征信息，通常是一个高维向量，由来自用户，广告和上下文字段的一热编码组成。$y_{click}$ 、$y_{conv}$和$y_{pv-conv}$分别是点击事件、点击后转换事件和查看后转换事件的二进制标签
根据点击标签的值，我们将$\mathcal{D}$中的所有样本分为两个子集： $\begin{array}{l}{\text{点击样本 }\mathcal{D}_{click}=\{\mathcal{D}\mid y_{click}=1\}}&{(\text{观察其转换标签 }y_{conv})}\\{\text{和未点击样本 }\mathcal{D}_{unclick}=\{\mathcal{D}\mid y_{click}=0\}}&{(\text{所有转换标签均为}}\end{array}$ “未知”)
CTR估计是为了预测点击事件的概率，i。e。$p_CTR=p(y_{click}=$ $1\mid x)$。CVR 估计是预测用户点击广告时的转化概率，即。e。$, p_{CVR}= p( y_{pv- conv}= 1\mid y_{click}= 1, x) = p( y_{conv}= 1\mid x)$ 。对于 CTCVR 估计，我们有$p_{CTCVR}=p(y_{pv-conv}=1\mid x)=$ $pcTR\cdot p_{CVR}$
传统的 CVR 估计模型采用与 CTR 估计任务类似的技术，例如逻辑回归$\mathbb{E}[18]$ ,因式分解机$\mathbb{E}[12,17]$ 和深度神经网络(DNN)$\boxed{\mathrm{E}}[4,8]$。接下来，我们介绍传统模型和整个空间模型
2.2 Base CVR Models Trained in Click Space
2.2.1 Single-Task CVR Model.常规 CVR 估计模型在点击数据$\mathcal{D}_{click}$上训练。令$\hat{p}_CVR=F_{v}\left(x\right)$表示单任务 CVR 模型，其中$\hat{p}_{CVR}\in(0,1)$是印象广告$x$的预测 CVR 得分。$F_v(\cdot)$表示由一个特征嵌入层和几个密集层组成的网络。目标是基于交叉熵
损失 $\ell(\cdot,\cdot)$ 制定的：
(1)
$$\min\limits_{F_{v}}\frac{1}{|\mathcal{D}_{click}|}\sum\limits_{\mathcal{D}_{click}}\ell\left(y_{conv},F_{v}\left(x\right)\right)\:.$$
2.2.2 Joint Estimation of CVR and CTR. 为了缓解 CVR 任务中的数据稀疏问题，联合优化 CVR 和 CTR 估计任务是一种常用
的方法，因为 CTR 任务是在印象广告$\mathcal{D}$上训练的，并且比CVR 任务具有更丰富的样本$\mathbb{E}[16,25]$ 。联合模型包含一个共享的特征嵌入层，以及两个单独的密集块，分别预测 CVR 和CTR 分数。令$\hat{p}_{CVR}=F_{v}\left(x\right)$和$\hat{p}_{CTR}=F_{c}\left(x\right)$表示联合模型的预测 CVR 和 CTR 得分 (注意$F_v(\cdot)$和$F_c(\cdot)$共享特征嵌人层), 联合模型的目标是：
$$_{c}\frac{1}{|\mathcal{D}_{click}|}\sum_{\mathcal{D}_{click}}\ell\:(y_{conv},F_{v}\:(x))+\gamma\frac{1}{|\mathcal{D}|}\sum_{\mathcal{D}}\ell\:(y_{click},F_{c}\:(x))$$
(2)
其中$\gamma$是权衡超参数
2.2.3 Limitations.单任务 CVR 模型的训练过程不会从未点击的广告中学习，并且联合模型仅通过来自 CTR 任务的共享嵌人来合并此类信息。因此，他们在未点击空间中的预测 CVR 得分可能具有不可忽略的偏差，因为点击和取消点击广告的数据分布之间存在差异
2.3 Entire Space CVR Estimation Models
2.3.1 Auxiliary Task Learning based Models. [EMa et al.[16] 包含两个辅助任务，即点击率(CTR)和点击率(CTCVR),可以在整个印象空间中进行训练，以间接学习整个空间 CVR 估计器。该模型与联合模型具有相同的体系结构(i。e。, $\hat{p}_{CVR}=F_{v}\left(x\right)$ 和$\hat{p}_{CTR}=F_{c}\left(x\right))$,而目标是最小化 CTCVR 和 CTR 估计上的交叉熵损失：
$$\operatorname*{min}_{F_{v},F_{c}}{\frac{1}{|\mathcal{D}|}}\sum_{\mathcal{D}}\:\left(\ell\left(y_{pv-conv},F_{c}(x)\cdot F_{v}(x)\right)+\gamma\ell\left(y_{click},F_{c}(x)\right)\right)$$
(3,
其中$F_{c}(x)\cdot F_{v}(x)=\hat{p}_{CTR}\cdot\hat{p}_{CVR}$是预测的 CTCVR 分数。在学习 CTCVR 估计的帮助下，用于 CVR 的网络$F_v(x)$可以从未点击的广告中学习，从而缓解 SSB 问题
2.3.2 Counterfactual Learning based Models.反事实学习提供了一种解决非随机缺失问题的方法[1,2,5,15,21,23,26,29] 。最近的几项研究$\mathbb{E}[9,31]$ 采用反事实学习，例如逆倾向得分(IPS)和双鲁棒(DR)估计器，以消除CVR 估计。基于IPS 的方法利用预测的 CTR 得分作为点击广告上 CVR 损失的倾向， 以实现理论上无偏的 CVR 估计器。优化目标为：
$$\frac{1}{|\mathcal{D}_{click}|}\sum_{\mathcal{D}_{click}}\frac{1}{F_{c}(x)}\ell\:(y_{conv},F_{v}(x))+\gamma\frac{1}{|\mathcal{D}|}\sum_{\mathcal{D}}\ell\:(y_{click},F_{c}(x))$$
$\underline{\mathrm{min}}$
(4)
基于 DR 的方法进一步学习估算每个未点击广告的 CVR 损
失的估算模型$F_i(\cdot)$ 。CVR 任务和插补任务交替训练，其中$F_i(\cdot)$在点击数据$\mathcal{D}_click$上训练。有关详细信息，请参考[9,31] 2.3.3 Limitations .尽管基于辅助任务学习的模型可以通过CTCVR 估计任务的学习从未点击的广告中学习，但它们有两个主要限制
$\bullet$ For a clicked ad (i.e., $y_click=1)$, if its post-view conversion
label $y_{pv-conv}=0$, this ad is a positive sample of CTR task as well as an negative sample of CTCVR task, which may result in gradient conflict to the two learning tasks.
|
[
[
0.0882568359375,
0.10797119140625,
0.48046875,
0.10797119140625,
0.48046875,
0.121826171875,
0.0882568359375,
0.121826171875
],
[
0.51708984375,
0.10797119140625,
0.63818359375,
0.10797119140625,
0.63818359375,
0.121826171875,
0.51708984375,
0.121826171875
],
[
0.08660888671875,
0.121826171875,
0.48046875,
0.121826171875,
0.48046875,
0.1351318359375,
0.08660888671875,
0.1351318359375
],
[
0.60107421875,
0.126953125,
0.72607421875,
0.1395263671875,
0.72216796875,
0.1654052734375,
0.59716796875,
0.15283203125
],
[
0.654296875,
0.126953125,
0.6611328125,
0.126953125,
0.6611328125,
0.1331787109375,
0.654296875,
0.1331787109375
],
[
0.69775390625,
0.1287841796875,
0.71337890625,
0.1319580078125,
0.708984375,
0.1395263671875,
0.69384765625,
0.1363525390625
],
[
0.71728515625,
0.1300048828125,
0.82861328125,
0.1319580078125,
0.82861328125,
0.146484375,
0.71728515625,
0.14453125
],
[
0.71826171875,
0.1351318359375,
0.72216796875,
0.1351318359375,
0.72216796875,
0.14013671875,
0.71826171875,
0.14013671875
],
[
0.89453125,
0.1319580078125,
0.9140625,
0.1319580078125,
0.9140625,
0.1458740234375,
0.89453125,
0.1458740234375
],
[
0.08660888671875,
0.1483154296875,
0.47802734375,
0.1483154296875,
0.47802734375,
0.1622314453125,
0.08660888671875,
0.1622314453125
],
[
0.0882568359375,
0.136962890625,
0.47802734375,
0.136962890625,
0.47802734375,
0.147705078125,
0.0882568359375,
0.147705078125
],
[
0.08660888671875,
0.1634521484375,
0.48193359375,
0.1634521484375,
0.48193359375,
0.1767578125,
0.08660888671875,
0.1767578125
],
[
0.51708984375,
0.1641845703125,
0.91259765625,
0.1641845703125,
0.91259765625,
0.177978515625,
0.51708984375,
0.177978515625
],
[
0.08905029296875,
0.1787109375,
0.4794921875,
0.1787109375,
0.4794921875,
0.189453125,
0.08905029296875,
0.189453125
],
[
0.51953125,
0.179931640625,
0.9111328125,
0.179931640625,
0.9111328125,
0.1900634765625,
0.51953125,
0.1900634765625
],
[
0.08660888671875,
0.1912841796875,
0.48046875,
0.1912841796875,
0.48046875,
0.2052001953125,
0.08660888671875,
0.2052001953125
],
[
0.5185546875,
0.1925048828125,
0.9140625,
0.1925048828125,
0.9140625,
0.205810546875,
0.5185546875,
0.205810546875
],
[
0.08660888671875,
0.2052001953125,
0.48046875,
0.2052001953125,
0.48046875,
0.2183837890625,
0.08660888671875,
0.2183837890625
],
[
0.51708984375,
0.205810546875,
0.9140625,
0.205810546875,
0.9140625,
0.2197265625,
0.51708984375,
0.2197265625
],
[
0.08660888671875,
0.2183837890625,
0.48046875,
0.2183837890625,
0.48046875,
0.231689453125,
0.08660888671875,
0.231689453125
],
[
0.521484375,
0.2215576171875,
0.91259765625,
0.2215576171875,
0.91259765625,
0.231689453125,
0.521484375,
0.231689453125
],
[
0.08660888671875,
0.23291015625,
0.48046875,
0.23291015625,
0.48046875,
0.246826171875,
0.08660888671875,
0.246826171875
],
[
0.51708984375,
0.2342529296875,
0.9140625,
0.2342529296875,
0.9140625,
0.2474365234375,
0.51708984375,
0.2474365234375
],
[
0.08660888671875,
0.246826171875,
0.4794921875,
0.246826171875,
0.4794921875,
0.260009765625,
0.08660888671875,
0.260009765625
],
[
0.5185546875,
0.2474365234375,
0.9140625,
0.2474365234375,
0.9140625,
0.261474609375,
0.5185546875,
0.261474609375
],
[
0.08660888671875,
0.260009765625,
0.48046875,
0.260009765625,
0.48046875,
0.2734375,
0.08660888671875,
0.2734375
],
[
0.515625,
0.260009765625,
0.64208984375,
0.261474609375,
0.64208984375,
0.27587890625,
0.515625,
0.274658203125
],
[
0.08660888671875,
0.2734375,
0.43701171875,
0.2734375,
0.43701171875,
0.287353515625,
0.08660888671875,
0.287353515625
],
[
0.61279296875,
0.2734375,
0.80126953125,
0.285888671875,
0.79833984375,
0.31005859375,
0.609375,
0.29736328125
],
[
0.78369140625,
0.276611328125,
0.90869140625,
0.28271484375,
0.90673828125,
0.304931640625,
0.78173828125,
0.298583984375
],
[
0.7685546875,
0.279052734375,
0.7900390625,
0.281005859375,
0.7890625,
0.2890625,
0.76708984375,
0.287353515625
],
[
0.52392578125,
0.2841796875,
0.64794921875,
0.29541015625,
0.6435546875,
0.317626953125,
0.51953125,
0.30615234375
],
[
0.57421875,
0.279052734375,
0.611328125,
0.279052734375,
0.611328125,
0.2890625,
0.57421875,
0.2890625
],
[
0.52783203125,
0.28857421875,
0.55615234375,
0.28857421875,
0.55615234375,
0.296630859375,
0.52783203125,
0.296630859375
],
[
0.11273193359375,
0.2998046875,
0.48046875,
0.2998046875,
0.48046875,
0.313232421875,
0.11273193359375,
0.313232421875
],
[
0.78759765625,
0.294189453125,
0.80810546875,
0.2998046875,
0.80224609375,
0.3125,
0.78173828125,
0.306884765625
],
[
0.1258544921875,
0.313232421875,
0.48193359375,
0.313232421875,
0.48193359375,
0.326904296875,
0.1258544921875,
0.326904296875
],
[
0.89453125,
0.311279296875,
0.91259765625,
0.311279296875,
0.91259765625,
0.324462890625,
0.89453125,
0.324462890625
],
[
0.1258544921875,
0.326904296875,
0.48046875,
0.326904296875,
0.48046875,
0.34033203125,
0.1258544921875,
0.34033203125
],
[
0.5185546875,
0.324462890625,
0.65185546875,
0.324462890625,
0.65185546875,
0.33837890625,
0.5185546875,
0.33837890625
],
[
0.12744140625,
0.34228515625,
0.4794921875,
0.34228515625,
0.4794921875,
0.352783203125,
0.12744140625,
0.352783203125
],
[
0.5185546875,
0.3447265625,
0.91259765625,
0.3447265625,
0.91259765625,
0.35791015625,
0.5185546875,
0.35791015625
],
[
0.125,
0.354248046875,
0.48046875,
0.354248046875,
0.48046875,
0.367431640625,
0.125,
0.367431640625
],
[
0.51953125,
0.35986328125,
0.9111328125,
0.35986328125,
0.9111328125,
0.37060546875,
0.51953125,
0.37060546875
],
[
0.1258544921875,
0.37060546875,
0.17724609375,
0.37060546875,
0.17724609375,
0.3818359375,
0.1258544921875,
0.3818359375
],
[
0.521484375,
0.373779296875,
0.91259765625,
0.373779296875,
0.91259765625,
0.3837890625,
0.521484375,
0.3837890625
],
[
0.11358642578125,
0.3837890625,
0.494384765625,
0.3837890625,
0.494384765625,
0.39453125,
0.11358642578125,
0.39453125
],
[
0.5185546875,
0.386474609375,
0.91259765625,
0.386474609375,
0.91259765625,
0.399658203125,
0.5185546875,
0.399658203125
],
[
0.125,
0.396484375,
0.486083984375,
0.396484375,
0.486083984375,
0.410400390625,
0.125,
0.410400390625
],
[
0.5185546875,
0.401611328125,
0.6650390625,
0.401611328125,
0.6650390625,
0.412353515625,
0.5185546875,
0.412353515625
],
[
0.125,
0.410400390625,
0.420654296875,
0.410400390625,
0.420654296875,
0.423583984375,
0.125,
0.423583984375
],
[
0.11273193359375,
0.423583984375,
0.48046875,
0.423583984375,
0.48046875,
0.4375,
0.11273193359375,
0.4375
],
[
0.51708984375,
0.4267578125,
0.861328125,
0.4267578125,
0.861328125,
0.43994140625,
0.51708984375,
0.43994140625
],
[
0.12744140625,
0.439453125,
0.4794921875,
0.439453125,
0.4794921875,
0.449462890625,
0.12744140625,
0.449462890625
],
[
0.5185546875,
0.443115234375,
0.91259765625,
0.443115234375,
0.91259765625,
0.45703125,
0.5185546875,
0.45703125
],
[
0.1258544921875,
0.45068359375,
0.48193359375,
0.45068359375,
0.48193359375,
0.464111328125,
0.1258544921875,
0.464111328125
],
[
0.51953125,
0.458251953125,
0.91259765625,
0.458251953125,
0.91259765625,
0.4716796875,
0.51953125,
0.4716796875
],
[
0.12744140625,
0.46728515625,
0.3896484375,
0.46728515625,
0.3896484375,
0.477783203125,
0.12744140625,
0.477783203125
],
[
0.5185546875,
0.4716796875,
0.91259765625,
0.4716796875,
0.91259765625,
0.48486328125,
0.5185546875,
0.48486328125
],
[
0.5185546875,
0.486083984375,
0.91259765625,
0.486083984375,
0.91259765625,
0.5,
0.5185546875,
0.5
],
[
0.111083984375,
0.501953125,
0.256591796875,
0.501953125,
0.256591796875,
0.51220703125,
0.111083984375,
0.51220703125
],
[
0.5185546875,
0.50048828125,
0.91259765625,
0.50048828125,
0.91259765625,
0.5146484375,
0.5185546875,
0.5146484375
],
[
0.51953125,
0.51513671875,
0.59716796875,
0.51513671875,
0.59716796875,
0.52587890625,
0.51953125,
0.52587890625
],
[
0.08660888671875,
0.50244140625,
0.11517333984375,
0.50244140625,
0.11517333984375,
0.51123046875,
0.08660888671875,
0.51123046875
],
[
0.08660888671875,
0.521484375,
0.13232421875,
0.521484375,
0.13232421875,
0.5322265625,
0.08660888671875,
0.5322265625
],
[
0.1258544921875,
0.52099609375,
0.289306640625,
0.52099609375,
0.289306640625,
0.5341796875,
0.1258544921875,
0.5341796875
],
[
0.57177734375,
0.53271484375,
0.57861328125,
0.53857421875,
0.57275390625,
0.54248046875,
0.56640625,
0.537109375
],
[
0.5810546875,
0.52783203125,
0.8857421875,
0.5341796875,
0.8837890625,
0.56298828125,
0.580078125,
0.556640625
],
[
0.08660888671875,
0.53857421875,
0.48046875,
0.53857421875,
0.48046875,
0.5517578125,
0.08660888671875,
0.5517578125
],
[
0.52783203125,
0.5390625,
0.61376953125,
0.546875,
0.609375,
0.568359375,
0.52392578125,
0.56103515625
],
[
0.53173828125,
0.54150390625,
0.57275390625,
0.54150390625,
0.57275390625,
0.5517578125,
0.53173828125,
0.5517578125
],
[
0.57861328125,
0.5361328125,
0.58837890625,
0.5361328125,
0.58837890625,
0.54345703125,
0.57861328125,
0.54345703125
],
[
0.900390625,
0.54248046875,
0.90673828125,
0.54248046875,
0.90673828125,
0.548828125,
0.900390625,
0.548828125
],
[
0.0882568359375,
0.5517578125,
0.48193359375,
0.5517578125,
0.48193359375,
0.56494140625,
0.0882568359375,
0.56494140625
],
[
0.08660888671875,
0.56640625,
0.47802734375,
0.56640625,
0.47802734375,
0.58056640625,
0.08660888671875,
0.58056640625
],
[
0.5185546875,
0.568359375,
0.91259765625,
0.568359375,
0.91259765625,
0.58349609375,
0.5185546875,
0.58349609375
],
[
0.08660888671875,
0.580078125,
0.48046875,
0.580078125,
0.48046875,
0.59326171875,
0.08660888671875,
0.59326171875
],
[
0.5185546875,
0.58349609375,
0.91259765625,
0.58349609375,
0.91259765625,
0.5966796875,
0.5185546875,
0.5966796875
],
[
0.08660888671875,
0.59326171875,
0.48046875,
0.59326171875,
0.48046875,
0.60693359375,
0.08660888671875,
0.60693359375
],
[
0.5185546875,
0.5966796875,
0.7548828125,
0.5966796875,
0.7548828125,
0.60986328125,
0.5185546875,
0.60986328125
],
[
0.08416748046875,
0.609375,
0.48193359375,
0.60595703125,
0.48193359375,
0.62060546875,
0.08416748046875,
0.62353515625
],
[
0.08660888671875,
0.6220703125,
0.281005859375,
0.6220703125,
0.281005859375,
0.63525390625,
0.08660888671875,
0.63525390625
],
[
0.5185546875,
0.6162109375,
0.91259765625,
0.6162109375,
0.91259765625,
0.6298828125,
0.5185546875,
0.6298828125
],
[
0.521484375,
0.630859375,
0.91259765625,
0.630859375,
0.91259765625,
0.64453125,
0.521484375,
0.64453125
],
[
0.10455322265625,
0.63720703125,
0.48046875,
0.63720703125,
0.48046875,
0.64794921875,
0.10455322265625,
0.64794921875
],
[
0.0882568359375,
0.6494140625,
0.48193359375,
0.6494140625,
0.48193359375,
0.6640625,
0.0882568359375,
0.6640625
],
[
0.51953125,
0.646484375,
0.9111328125,
0.646484375,
0.9111328125,
0.6572265625,
0.51953125,
0.6572265625
],
[
0.0849609375,
0.662109375,
0.48193359375,
0.66357421875,
0.48193359375,
0.67822265625,
0.0849609375,
0.6767578125
],
[
0.515625,
0.65771484375,
0.91259765625,
0.6591796875,
0.91259765625,
0.67236328125,
0.515625,
0.671875
],
[
0.51708984375,
0.67041015625,
0.91259765625,
0.671875,
0.91259765625,
0.68603515625,
0.51708984375,
0.68505859375
],
[
0.08660888671875,
0.67822265625,
0.1380615234375,
0.67822265625,
0.1380615234375,
0.689453125,
0.08660888671875,
0.689453125
],
[
0.1004638671875,
0.689453125,
0.484375,
0.6923828125,
0.484375,
0.70703125,
0.1004638671875,
0.7041015625
],
[
0.51708984375,
0.68505859375,
0.82421875,
0.68603515625,
0.82421875,
0.69970703125,
0.51708984375,
0.69873046875
],
[
0.564453125,
0.7021484375,
0.6640625,
0.7021484375,
0.6640625,
0.7197265625,
0.564453125,
0.7197265625
],
[
0.6689453125,
0.70263671875,
0.82275390625,
0.7119140625,
0.8203125,
0.73291015625,
0.66650390625,
0.7236328125
],
[
0.7900390625,
0.7041015625,
0.810546875,
0.70556640625,
0.8095703125,
0.7138671875,
0.7890625,
0.7119140625
],
[
0.80224609375,
0.7021484375,
0.92724609375,
0.70849609375,
0.92626953125,
0.72998046875,
0.80126953125,
0.7236328125
],
[
0.5185546875,
0.7119140625,
0.64208984375,
0.7216796875,
0.6396484375,
0.7392578125,
0.515625,
0.72998046875
],
[
0.52294921875,
0.71337890625,
0.5458984375,
0.71337890625,
0.5458984375,
0.7216796875,
0.52294921875,
0.7216796875
],
[
0.08660888671875,
0.70703125,
0.48046875,
0.70703125,
0.48046875,
0.71728515625,
0.08660888671875,
0.71728515625
],
[
0.0849609375,
0.7197265625,
0.483642578125,
0.7197265625,
0.483642578125,
0.734375,
0.0849609375,
0.734375
],
[
0.61279296875,
0.7138671875,
0.67724609375,
0.71630859375,
0.67724609375,
0.73095703125,
0.61279296875,
0.72900390625
],
[
0.814453125,
0.72021484375,
0.830078125,
0.7265625,
0.82275390625,
0.7373046875,
0.80810546875,
0.73095703125
],
[
0.0849609375,
0.73193359375,
0.48193359375,
0.73291015625,
0.48193359375,
0.74755859375,
0.0849609375,
0.7470703125
],
[
0.896484375,
0.73779296875,
0.9111328125,
0.73779296875,
0.9111328125,
0.74755859375,
0.896484375,
0.74755859375
],
[
0.08660888671875,
0.75,
0.16259765625,
0.75,
0.16259765625,
0.76025390625,
0.08660888671875,
0.76025390625
],
[
0.53515625,
0.74853515625,
0.9140625,
0.74853515625,
0.9140625,
0.76220703125,
0.53515625,
0.76220703125
],
[
0.10455322265625,
0.76220703125,
0.4794921875,
0.76220703125,
0.4794921875,
0.77294921875,
0.10455322265625,
0.77294921875
],
[
0.5185546875,
0.76220703125,
0.91259765625,
0.76220703125,
0.91259765625,
0.77587890625,
0.5185546875,
0.77587890625
],
[
0.08660888671875,
0.7734375,
0.4794921875,
0.7734375,
0.4794921875,
0.787109375,
0.08660888671875,
0.787109375
],
[
0.51708984375,
0.7763671875,
0.91015625,
0.7763671875,
0.91015625,
0.79052734375,
0.51708984375,
0.79052734375
],
[
0.08660888671875,
0.78857421875,
0.473876953125,
0.78857421875,
0.473876953125,
0.8017578125,
0.08660888671875,
0.8017578125
],
[
0.5185546875,
0.796875,
0.91259765625,
0.796875,
0.91259765625,
0.81005859375,
0.5185546875,
0.81005859375
],
[
0.51708984375,
0.81103515625,
0.9140625,
0.81103515625,
0.9140625,
0.82470703125,
0.51708984375,
0.82470703125
],
[
0.08660888671875,
0.82470703125,
0.13232421875,
0.82470703125,
0.13232421875,
0.8359375,
0.08660888671875,
0.8359375
],
[
0.12744140625,
0.82373046875,
0.461669921875,
0.82373046875,
0.461669921875,
0.8369140625,
0.12744140625,
0.8369140625
],
[
0.521484375,
0.8271484375,
0.6005859375,
0.8271484375,
0.6005859375,
0.8369140625,
0.521484375,
0.8369140625
],
[
0.0849609375,
0.84130859375,
0.48046875,
0.84130859375,
0.48046875,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.54345703125,
0.84228515625,
0.9140625,
0.84228515625,
0.9140625,
0.85693359375,
0.54345703125,
0.85693359375
],
[
0.0849609375,
0.85498046875,
0.48046875,
0.85400390625,
0.48046875,
0.86865234375,
0.0849609375,
0.869140625
],
[
0.5556640625,
0.85400390625,
0.9150390625,
0.85400390625,
0.9150390625,
0.8720703125,
0.5556640625,
0.8720703125
],
[
0.08660888671875,
0.869140625,
0.46728515625,
0.869140625,
0.46728515625,
0.88330078125,
0.08660888671875,
0.88330078125
],
[
0.55615234375,
0.869140625,
0.91259765625,
0.869140625,
0.91259765625,
0.88330078125,
0.55615234375,
0.88330078125
],
[
0.0882568359375,
0.88525390625,
0.48046875,
0.88525390625,
0.48046875,
0.89599609375,
0.0882568359375,
0.89599609375
],
[
0.55810546875,
0.8837890625,
0.865234375,
0.8837890625,
0.865234375,
0.89794921875,
0.55810546875,
0.89794921875
]
] |
[
"含一个点击自适应教师模型，该模型为未点击的广告生成伪",
"损失（，）制定的：",
"转换标签，以及一个不确定性正则化的学生模型，该模型可以",
"mnin Delicl Dane",
"1",
"",
"L(yconv, Fu (x)",
"",
"(1)",
"产生未点击广告的监督信号，教师使用领域自适应学习印象",
"有效地提炼教师学到的未点击广告中的知识。具体来说，为了",
"广告的点击自适应表示，其对未点击广告的预测CVR得分被",
"2.2.2JointEstimationofCVRandCTR.为了缓解CVR任务中",
"作为其伪转换标签。然后，学生可以从点击广告（带有真实标",
"的数据稀疏问题，联合优化CVR和CTR估计任务是一种常用",
"签）和未点击广告（带有老师的伪标签）中学习，并且还可以",
"的方法，因为CTR任务是在印象广告D上训练的，并且比",
"对伪标签进行不确定性估计，以减轻其中的固有噪声。对于每",
"CVR任务具有更丰富的样本E[16,25]。联合模型包含一个共",
"个未点击的广告，我们的学生估计其伪标签的不确定性，并",
"享的特征入层，以及两个单独的密集块，分别预测CVR和",
"在训练期间动态调整其CVR损失的权重，以削弱其负面影响。",
"CTR分数。令PCVR=Fu(x）和pCTR=F(x）表示联合模型的",
"在十亿级数据集上的实验结果表明，UKD优于以前的最新方",
"预测CVR和CTR得分（注意Fo）和Fc（）共享特征嵌人层），",
"法。我们已经在阿里巴巴广告平台部署了UKD，在线结果验",
"联合模型的目标是：",
"证了UKD实现了显著的改进。这项工作的主要贡献是：",
"t(ycono,Fe (x)+",
"Ee (yelick, Fe (x)",
"1",
"if, Delickl Delick",
"1",
"min",
".We propose uncertainty-regularized knowledge distillation",
"D",
"(UKD) to debias CVR models via learning from unclicked",
"(2)",
"ads. It employs a click-adaptive teacher to generate pseudo",
"其中y是权衡超参数",
"conversion labels for unclicked ads, and then trains a stu",
"2.2.3Limitations，单任务CVR模型的训练过程不会从未点击",
"dent model that takes both clicked and unclicked ads into",
"的广告中学习，并且联合模型仅通过来自CTR任务的共享嵌",
"account.",
"入来合并此类信息。因此，他们在未点击空间中的预测CVR",
".Ourstudentmodelperforms uncertaintyestimationtopseudo-",
"得分可能具有不可忽略的偏差，因为点击和取消点击广告的",
"labels generated by the teacher, alleviating the inherent noise",
"数据分布之间存在差异",
"to reduce the negative impact during distillation.",
". Experimental results on public and large-scale production",
"2.3EntireSpaceCVREstimationModels",
"datasets show that UKD outperforms the state-of-the-art",
"2.3.1Auxiliary Task Learning basedModels.EMa et al. [16]包",
"methods. Online experiments further verify that it achieves",
"含两个辅助任务，即点击率（CTR）和点击率（CTCVR），可以在",
"significantlyimprovements oncoremetrics.",
"整个印象空间中进行训练，以间接学习整个空间CVR估计器。",
"该模型与联合模型具有相同的体系结构(i。e。，pcVR=F(x)",
"PREREOUISITES",
"和PCTR=Fc（x）），而目标是最小化CTCVR和CTR估计上的",
"交叉损失：",
"2",
"2.1",
"ProblemDefinition",
"",
"( (up-cono () F()+ (yciek, e(x)",
"在在线广告系统中，我们可以记录用户对印象广告的反馈，以",
"D",
"min-a",
"",
"3",
"训练用于估计CTR（点击率），CVR和CTCVR（点击率）的",
"模型。令D=【(x,yclick,yconu,pu-cono)】表示所收集的印象广",
"其中Fc(x)·F(x）=pCTR·pCVR是预测的CTCVR分数。在学",
"告数据集。对于每个印象样本，x表示特征信息，通常是一个",
"习CTCVR估计的帮助下，用于CVR的网络F（x）可以从未点",
"高维向量，由来自用户，广告和上下文字段的一热编码组成。",
"击的广告中学习，从而缓解SSB问题",
"YclickYconu和ypu-conu分别是点击事件、点击后转换事件和",
"查看后转换事件的二进制标签",
"2.3.2Counterfactual LearningbasedModels.反事实学习提供",
"了一种解决非随机缺失问题的方法[1,2,5,15,21,23,26,29]。",
"根据点击标签的值，我们将①中的所有样本分为两个子集：",
"点击样本Dclick=（D|yclick=1)（观察其转换标签yconu）",
"最近的几项研究E[9，31]采用反事实学习，例如逆倾向得分",
"和未点击样本Dunclick=(D「yclick=0)（所有转换标签均为",
"（IPS）和双鲁棒（DR）估计器，以消除CVR估计。基于IPS的",
"方法利用预测的CTR得分作为点击广告上CVR损失的倾向，",
"“未知”）",
"CTR估计是为了预测点击事件的概率，i。e。PCTR=p(yelick",
"以实现理论上无偏的CVR估计器。优化目标为：",
"17",
" (ycono,Fo(x)+回",
"1",
"Ze(yelick, Fe(x)",
"Delickl Deliek",
"min",
"1x）。CVR估计是预测用户点击广告时的转化概率，即。e。",
" PCVR = p(ypu-conu = 1 I Yelick = 1,x) = p(yconu = 1 I x) 。",
"ZFe(x)",
"D",
"对于CTCVR估计，我们有pCTCVR=p(ypu-conu=1|x）=",
"(4)",
"PCTRPCVR",
"基于DR的方法进一步学习估算每个未点击广告的CVR损",
"传统的CVR估计模型采用与CTR估计任务类似的技术，例",
"失的估算模型F:(）。CVR任务和插补任务交替训练，其中",
"如逻辑回归E[18]，因式分解机E[12,17]和深度神经网络",
"F;(在点击数据Dclick上训练。有关详细信息，请参考[9,31]",
"(DNN)E[4,8]。接下来，我们介绍传统模型和整个空间模型",
"2.3.3Limitations.尽管基于辅助任务学习的模型可以通过",
"CTCVR估计任务的学习从未点击的广告中学习，但它们有两",
"2.2",
"Base CVR Models Trained in Click Space",
"个主要限制：",
"2.2.1Single-TaskCVRModel.常规CVR估计模型在点击数据",
". For a clicked ad (i.e., yclick = 1), if its post-view conversion",
"Dclick上训练。令pcVR=F（x）表示单任务CVR模型，其中",
"label ypo-conu = O, this ad is a positive sample of CTR task",
"PCVRE（0,1）是印象广告x的预测CVR得分。Fo(）表示由",
"as well as an negative sample of CTCVR task, which may",
"个特征嵌人层和几个密集层组成的网络。自标是基于交叉熵",
"result in gradient conflict to the two learning tasks"
] |
{
"arxiv_id": "2201.08024",
"language": "zh",
"page_idx": 1
}
|
||
2201.08024-zh-2
|
2201.08024-zh-2.jpg
|
$\bullet$ Foran unclicked ad(i.e.,$y_click=0$ and $y_conv$ is "unknown"),
the models tend to optimize the predicted CVR scores to zeros. The proof is given here: The loss of CTCVR estimation is $\ell=-y_{pv-conv}\log(\hat{p}_{CTR}\cdot\hat{p}_{CVR})-(1-y_{pv-conv})\log(1-\hat{p}_{CTR}\cdot$ $\hat{p} _{CVR}) \cdot$For an unclicked ad whose $( y_{click}, y_{conv}, y_{pv- conv}) =$ (0,unknown, 0),the gradient to $\hat{p} _{CVR}$ is $\frac {\partial \ell }{\partial \hat{p} _{CVR}}= \frac {\hat{p} _{CTR}}{1- \hat{p} _{CTR}\cdot \hat{p} _{CVR}}$ Note that $\hat{n} _{CTB}\in ( 0, 1)$ and $\hat{p} _{CVB}\in ( 0, 1) $,thus the gr
.Note that $\hat{p}_{CTR}\in(0,1)$ and $\hat{p}_{CVR}\in(0,1)$ ,thus the gradient is always positive, which means that it tends to optimize $\hat{p}_{CVR}$ of unclicked ads to 0, but the actual label is “unknown” .
基于反事实学习的模型已经实现了最先进的性能。但是，它
们的限制包含：
$\bullet$ For IPS-based models, the training procedure of the CVR task is on clicked data and does not explicitly take unclicked ads into account.
$\bullet$ For DR-based models, although the imputation task is used to estimate CVR loss of each unclicked ad, its learning procedure still utilizes clicked data only, and thus the imputation task is lack of accurate supervision.
# 3 PROPOSED METHOD
我们提出了一个U不确定性正则化的 KnowledgeD 蒸馏(UKD) 框架，该框架旨在通过从未点击的广告中提取知识来消除 CVR 估计的偏差。基本思想是，我们通过为未点击的广告产生可靠的伪转换标签作为监督信号来构建整个空间 CVR 估计模型
图。2 说明了 UKD 的整体工作流程，它包括一个为未点击的广告产生伪转化标签的点击自适应教师模型，以及一个提炼教师学到的有价值的知识以进行整个空间 CVR 估计的不确定性正则化学生模型。接下来，我们分别从老师和学生的细节来阐述我们的 UKD
## 3.1 Click-Adaptive Teacher Model
UKD 中老师的目标是在我们只能访问点击广告$\mathcal{D}_{click}$的真实转换标签的情况下，为未点击广告$\mathcal{D}_unclick$生成伪转换标签， 从而促进 CVR 估计任务的整个空间训练
点击和未点击样本的特征分布之间存在差异。为了具有对未点击的广告进行准确推断的能力，教师模型可能不会学习特定于点击样本的特征表示，而是学习点击自适应表示。我们建议从无监督域自适应的角度处理伪转换标签生成，其中源/目标域是点击/未点击的空间，灵感来自[3]。通过这种方式，问题被公式化为在给定标记的点击广告($\mathcal{D}_click$作为源域)的情况下，为未标记的未点击广告($\mathcal{D}_unclick$作为目标域)产生可靠的伪转换标签
3.1.1 Click-Adaptive Representation Learning.具体来说，我们的点击自适应教师模型采用对抗性学习[7,24] ,它引人了点击鉴别器来减轻训练期间点击/未点击样本的不一致特征分布Model Architecture 如图所示。2 , 点击自适应教师模型由
特征表示学习器$T_f(\cdot)$ 、CVR 预测器$T_p(\cdot)$和点击鉴别器$T_d(\cdot)$ 组成。形式上，$T_f(\cdot)$将每个样本的特征$x$作为输人来学习其密集表示$h^{(T)}$ ,其中$T_f(\cdot)$包含一个特征嵌入层和几个密集层。$T_{p}(\cdot)$拟估计样本的 CVR 得分$\hat{p}_{CVR}^{(T)}$。它由几个致密层和顶部的 softmax 函数组成，以产生概率分布
为了使特征表示$h^{(T)}$更具点击适应性，以便于在未点击的广告上生成伪转换标签，教师模型引人了点击鉴别器$T_d(\cdot)$来
对每个样本的域进行分类(i。e。,单击或取消单击)基于示例的表示形式$h^{(T)}$。直觉是，如果强点击鉴别器不能正确预测样本的域标签，则其表示$h^{(T)}$是点击自适应的
总体而言，教师模型的正向过程为：
(5)
$$\begin{aligned}&\boldsymbol{h}^{(T)}=T_{f}(x)\\&\boldsymbol{p}_{conv}^{(T)}=\mathrm{softmax}\left(T_{p}\left(\boldsymbol{h}^{(T)}\right)\right)=\left(\hat{p}_{CVR}^{(T)},1-\hat{p}_{CVR}^{(T)}\right)\\&\boldsymbol{p}_{d}=\mathrm{softmax}\left(T_{d}\left(\boldsymbol{h}^{(T)}\right)\right)\end{aligned}$$
其中$p_{conv}^{(T)}$是预测的 CVR 分布$(\hat{p}_{CVR}^{(T)}$是预测的 CVR 分数)。
$p_d$是预测的域分布
要学习点击自适应表示，给定印象广告$x,T_f(\cdot)$学习的其表示$h^{(T)}$旨在混淆点击鉴别器域并最大化域分类损失，而点击鉴别器$T_d(\cdot)$本身旨在最大程度地减少分类损失。教师通过对抗性学习过程进行优化：
(6)
$$\min_{T_{f},T_{p}}\mathcal{L}_{CVR}^{(T)}=\frac{1}{|\mathcal{D}_{click}|}\sum_{\mathcal{D}_{click}}\ell(y_{conv},\boldsymbol{p}_{conv}^{(T)})$$
$$\max\limits_{T_f}\min\limits_{T_d}\mathcal{L}_d^{(T)}=\frac{1}{|\mathcal{D}|}\sum\limits_{\mathcal{D}}\ell(y_{click},\boldsymbol{p}_d)$$
(7)
第一个方程使 CVR 估计的损失最小化，以优化学习者$T_f(\cdot)$和预测器$T_p(\cdot)$ 。第二个意味着学习者$T_f(\cdot)$使点击和未点击广告的表示无法区分，而点击鉴别器$T_d(\cdot)$经过优化，可以更好地将点击广告与未点击广告区分开。在实践中，我们通过梯度反转$\mathbb{E}[7]$实现它
当训练有素的鉴别器无法区分它们时，来自两个域的学习表示会有效地对齐。因此，基于点击自适应表示，教师能够对未点击的广告进行可靠的 CVR 估计
3.1.2 Produce Pseudo-Conversion Labels for Unclicked Ads. 经过训练的教师模型对$\mathcal{D}_unclick$中每个未点击的广告进行推理，以产生预测的 CVR 分布 $p_{conv}^{(T)}$作为伪转换标签，其中推理的正向过程仅包括$T_f(\cdot)$和$T_p(\cdot)$,而不需要$T_d(\cdot)$
我们使用$\widetilde{\mathcal{D}}_{unclick}=\{(x,\boldsymbol{p}_{conv}^{(T)})\}$来表示与伪转换标签耦合
的未单击样本，这些样本将用于训练整个空间 CVR 模型
3.2 Uncertainty-Regularized Student Model
基于点击自适应教师模型学习的未点击广告的伪转换标签，我们的 UKD 框架进一步构建了基于知识提炼的学生模型$\boxed{\mathrm{E} }[ 11]$ , 该模型从点击广告 (带有地面实况标签)和未点击广告 (带有伪标签)中学习，以执行整个空间 CVR 估计。与仅使用点击样本训练的模型相比，我们的模型通过在训练期间明确考虑未点击的样本来缓解 SSB 问题
我们阐述了蒸馏策略，可以指导学生模型挖掘老师学到的有价值的知识。由于教师预测中存在固有的噪音，未点击广告的伪标签的置信度不如点击广告的真实转化标签。为了解决这个问题，我们提出了一个不确定性正则化的学生，通过对蒸馏过程中伪标签的不确定性进行建模来减少噪声的负面影响。接下来，我们详细介绍了学生模型的蒸馏策略及其两个关键模块：标签蒸馏和不确定性正则化
3.2.1 Base Student Model: Label Distillation. 我们从引人基础学生模型开始，该模型是与 CVR 和 CTR 估计任务共同学习的， 作为 §$\underline{\mathrm{E}}$2.2.2 中的联合模型
|
[
[
0.11273193359375,
0.106689453125,
0.48046875,
0.106689453125,
0.48046875,
0.125,
0.11273193359375,
0.125
],
[
0.51708984375,
0.106689453125,
0.9111328125,
0.10797119140625,
0.9111328125,
0.121826171875,
0.51708984375,
0.12060546875
],
[
0.1258544921875,
0.121826171875,
0.48046875,
0.121826171875,
0.48046875,
0.1351318359375,
0.1258544921875,
0.1351318359375
],
[
0.5185546875,
0.12060546875,
0.91259765625,
0.12060546875,
0.91259765625,
0.1351318359375,
0.5185546875,
0.1351318359375
],
[
0.125,
0.1363525390625,
0.48193359375,
0.1351318359375,
0.48193359375,
0.1483154296875,
0.125,
0.149658203125
],
[
0.5185546875,
0.1351318359375,
0.81787109375,
0.1351318359375,
0.81787109375,
0.149658203125,
0.5185546875,
0.149658203125
],
[
0.1258544921875,
0.149658203125,
0.47802734375,
0.149658203125,
0.47802734375,
0.1634521484375,
0.1258544921875,
0.1634521484375
],
[
0.53515625,
0.15087890625,
0.76123046875,
0.15087890625,
0.76123046875,
0.1641845703125,
0.53515625,
0.1641845703125
],
[
0.125,
0.1641845703125,
0.48046875,
0.1641845703125,
0.48046875,
0.1787109375,
0.125,
0.1787109375
],
[
0.1217041015625,
0.173583984375,
0.410888671875,
0.18115234375,
0.410888671875,
0.20263671875,
0.1217041015625,
0.1956787109375
],
[
0.36767578125,
0.173583984375,
0.5009765625,
0.186279296875,
0.496826171875,
0.2083740234375,
0.363525390625,
0.1956787109375
],
[
0.5703125,
0.1685791015625,
0.654296875,
0.17236328125,
0.65283203125,
0.1900634765625,
0.56884765625,
0.186279296875
],
[
0.441162109375,
0.179931640625,
0.4755859375,
0.179931640625,
0.4755859375,
0.189453125,
0.441162109375,
0.189453125
],
[
0.560546875,
0.1881103515625,
0.86669921875,
0.1912841796875,
0.86669921875,
0.21533203125,
0.560546875,
0.212158203125
],
[
0.1258544921875,
0.1962890625,
0.48193359375,
0.1976318359375,
0.48193359375,
0.212158203125,
0.1258544921875,
0.2108154296875
],
[
0.83642578125,
0.1962890625,
0.85302734375,
0.1962890625,
0.85302734375,
0.199462890625,
0.83642578125,
0.199462890625
],
[
0.89208984375,
0.199462890625,
0.91015625,
0.1962890625,
0.9140625,
0.2083740234375,
0.896484375,
0.2108154296875
],
[
0.12744140625,
0.212158203125,
0.4794921875,
0.212158203125,
0.4794921875,
0.2259521484375,
0.12744140625,
0.2259521484375
],
[
0.580078125,
0.2183837890625,
0.73876953125,
0.2171630859375,
0.73876953125,
0.23486328125,
0.580078125,
0.236083984375
],
[
0.125,
0.2259521484375,
0.483642578125,
0.2259521484375,
0.483642578125,
0.2392578125,
0.125,
0.2392578125
],
[
0.1282958984375,
0.240478515625,
0.1993408203125,
0.240478515625,
0.1993408203125,
0.251953125,
0.1282958984375,
0.251953125
],
[
0.515625,
0.24560546875,
0.9189453125,
0.24560546875,
0.9189453125,
0.266357421875,
0.515625,
0.266357421875
],
[
0.101318359375,
0.255615234375,
0.48193359375,
0.2568359375,
0.48193359375,
0.270751953125,
0.101318359375,
0.26953125
],
[
0.51708984375,
0.264404296875,
0.6435546875,
0.26318359375,
0.6435546875,
0.27783203125,
0.51708984375,
0.279052734375
],
[
0.0882568359375,
0.272705078125,
0.181396484375,
0.272705078125,
0.181396484375,
0.28271484375,
0.0882568359375,
0.28271484375
],
[
0.54833984375,
0.27978515625,
0.91259765625,
0.27978515625,
0.91259765625,
0.29345703125,
0.54833984375,
0.29345703125
],
[
0.1094970703125,
0.287353515625,
0.48046875,
0.28857421875,
0.48046875,
0.302978515625,
0.1094970703125,
0.3017578125
],
[
0.5185546875,
0.29541015625,
0.91259765625,
0.29541015625,
0.91259765625,
0.31005859375,
0.5185546875,
0.31005859375
],
[
0.12744140625,
0.3037109375,
0.4794921875,
0.3037109375,
0.4794921875,
0.314453125,
0.12744140625,
0.314453125
],
[
0.5185546875,
0.31005859375,
0.91259765625,
0.31005859375,
0.91259765625,
0.323974609375,
0.5185546875,
0.323974609375
],
[
0.125,
0.317626953125,
0.2279052734375,
0.318115234375,
0.2279052734375,
0.330078125,
0.125,
0.328857421875
],
[
0.51708984375,
0.324462890625,
0.705078125,
0.324462890625,
0.705078125,
0.33837890625,
0.51708984375,
0.33837890625
],
[
0.111083984375,
0.330810546875,
0.483642578125,
0.330810546875,
0.483642578125,
0.3447265625,
0.111083984375,
0.3447265625
],
[
0.12335205078125,
0.343505859375,
0.48193359375,
0.3447265625,
0.48193359375,
0.35791015625,
0.12335205078125,
0.356689453125
],
[
0.583984375,
0.34521484375,
0.7587890625,
0.363037109375,
0.7548828125,
0.3818359375,
0.5810546875,
0.3642578125
],
[
0.62109375,
0.347900390625,
0.646484375,
0.34521484375,
0.64794921875,
0.354248046875,
0.62158203125,
0.35595703125
],
[
0.689453125,
0.343505859375,
0.701171875,
0.347900390625,
0.6953125,
0.35595703125,
0.6845703125,
0.3515625
],
[
0.73193359375,
0.348388671875,
0.74609375,
0.348388671875,
0.74609375,
0.356689453125,
0.73193359375,
0.356689453125
],
[
0.7490234375,
0.348388671875,
0.84619140625,
0.3515625,
0.84619140625,
0.3662109375,
0.7490234375,
0.363037109375
],
[
0.81884765625,
0.349853515625,
0.83251953125,
0.349853515625,
0.83251953125,
0.354248046875,
0.81884765625,
0.354248046875
],
[
0.8935546875,
0.35107421875,
0.91259765625,
0.35107421875,
0.91259765625,
0.3642578125,
0.8935546875,
0.3642578125
],
[
0.1258544921875,
0.35791015625,
0.48046875,
0.35791015625,
0.48046875,
0.371826171875,
0.1258544921875,
0.371826171875
],
[
0.744140625,
0.354248046875,
0.75634765625,
0.354248046875,
0.75634765625,
0.35986328125,
0.744140625,
0.35986328125
],
[
0.58837890625,
0.3642578125,
0.61279296875,
0.3642578125,
0.61279296875,
0.371826171875,
0.58837890625,
0.371826171875
],
[
0.73046875,
0.361083984375,
0.736328125,
0.356689453125,
0.740234375,
0.35986328125,
0.734375,
0.3642578125
],
[
0.1258544921875,
0.37255859375,
0.340576171875,
0.37255859375,
0.340576171875,
0.386474609375,
0.1258544921875,
0.386474609375
],
[
0.59326171875,
0.3818359375,
0.8349609375,
0.380615234375,
0.8349609375,
0.4140625,
0.59326171875,
0.415283203125
],
[
0.68701171875,
0.395751953125,
0.7548828125,
0.40478515625,
0.7509765625,
0.42041015625,
0.68310546875,
0.410888671875
],
[
0.89453125,
0.393310546875,
0.9140625,
0.393310546875,
0.9140625,
0.4072265625,
0.89453125,
0.4072265625
],
[
0.0882568359375,
0.40283203125,
0.11517333984375,
0.40283203125,
0.11517333984375,
0.410888671875,
0.0882568359375,
0.410888671875
],
[
0.11517333984375,
0.401611328125,
0.294189453125,
0.401611328125,
0.294189453125,
0.412353515625,
0.11517333984375,
0.412353515625
],
[
0.60693359375,
0.40478515625,
0.65673828125,
0.40478515625,
0.65673828125,
0.416748046875,
0.60693359375,
0.416748046875
],
[
0.65625,
0.4072265625,
0.66259765625,
0.4072265625,
0.66259765625,
0.410888671875,
0.65625,
0.410888671875
],
[
0.08660888671875,
0.417236328125,
0.48193359375,
0.417236328125,
0.48193359375,
0.43115234375,
0.08660888671875,
0.43115234375
],
[
0.515625,
0.42236328125,
0.9140625,
0.423583984375,
0.9140625,
0.438232421875,
0.515625,
0.4375
],
[
0.08660888671875,
0.431884765625,
0.48046875,
0.431884765625,
0.48046875,
0.44580078125,
0.08660888671875,
0.44580078125
],
[
0.5185546875,
0.439453125,
0.91259765625,
0.439453125,
0.91259765625,
0.45263671875,
0.5185546875,
0.45263671875
],
[
0.0882568359375,
0.447509765625,
0.4794921875,
0.447509765625,
0.4794921875,
0.458251953125,
0.0882568359375,
0.458251953125
],
[
0.51953125,
0.453857421875,
0.91259765625,
0.453857421875,
0.91259765625,
0.46728515625,
0.51953125,
0.46728515625
],
[
0.0882568359375,
0.4609375,
0.45751953125,
0.4609375,
0.45751953125,
0.4716796875,
0.0882568359375,
0.4716796875
],
[
0.5185546875,
0.46728515625,
0.9140625,
0.46728515625,
0.9140625,
0.48095703125,
0.5185546875,
0.48095703125
],
[
0.10296630859375,
0.473388671875,
0.4794921875,
0.474853515625,
0.4794921875,
0.486083984375,
0.10296630859375,
0.48486328125
],
[
0.08660888671875,
0.4873046875,
0.48046875,
0.4873046875,
0.48046875,
0.50048828125,
0.08660888671875,
0.50048828125
],
[
0.5185546875,
0.481689453125,
0.62841796875,
0.481689453125,
0.62841796875,
0.49560546875,
0.5185546875,
0.49560546875
],
[
0.08660888671875,
0.50048828125,
0.4794921875,
0.50048828125,
0.4794921875,
0.5146484375,
0.08660888671875,
0.5146484375
],
[
0.53515625,
0.49560546875,
0.9111328125,
0.49560546875,
0.9111328125,
0.5087890625,
0.53515625,
0.5087890625
],
[
0.0882568359375,
0.5146484375,
0.48193359375,
0.5146484375,
0.48193359375,
0.52783203125,
0.0882568359375,
0.52783203125
],
[
0.51953125,
0.51123046875,
0.9111328125,
0.51123046875,
0.9111328125,
0.521484375,
0.51953125,
0.521484375
],
[
0.5185546875,
0.52294921875,
0.74609375,
0.52294921875,
0.74609375,
0.5361328125,
0.5185546875,
0.5361328125
],
[
0.0882568359375,
0.52978515625,
0.193603515625,
0.52978515625,
0.193603515625,
0.54052734375,
0.0882568359375,
0.54052734375
],
[
0.5185546875,
0.54541015625,
0.91259765625,
0.54541015625,
0.91259765625,
0.5595703125,
0.5185546875,
0.5595703125
],
[
0.08660888671875,
0.55810546875,
0.13232421875,
0.55810546875,
0.13232421875,
0.568359375,
0.08660888671875,
0.568359375
],
[
0.1282958984375,
0.556640625,
0.378173828125,
0.556640625,
0.378173828125,
0.57080078125,
0.1282958984375,
0.57080078125
],
[
0.5185546875,
0.56005859375,
0.91259765625,
0.56005859375,
0.91259765625,
0.57470703125,
0.5185546875,
0.57470703125
],
[
0.0849609375,
0.57470703125,
0.48046875,
0.57470703125,
0.48046875,
0.5888671875,
0.0849609375,
0.5888671875
],
[
0.51953125,
0.57568359375,
0.91259765625,
0.57568359375,
0.91259765625,
0.59033203125,
0.51953125,
0.59033203125
],
[
0.0849609375,
0.5869140625,
0.48046875,
0.58837890625,
0.48046875,
0.60302734375,
0.0849609375,
0.6015625
],
[
0.5185546875,
0.5888671875,
0.80615234375,
0.5888671875,
0.80615234375,
0.60302734375,
0.5185546875,
0.60302734375
],
[
0.0849609375,
0.60107421875,
0.34228515625,
0.60107421875,
0.34228515625,
0.6142578125,
0.0849609375,
0.6142578125
],
[
0.53173828125,
0.60498046875,
0.91259765625,
0.60498046875,
0.91259765625,
0.62255859375,
0.53173828125,
0.62255859375
],
[
0.10296630859375,
0.6162109375,
0.48046875,
0.6162109375,
0.48046875,
0.6298828125,
0.10296630859375,
0.6298828125
],
[
0.5185546875,
0.62255859375,
0.87744140625,
0.62255859375,
0.87744140625,
0.63330078125,
0.5185546875,
0.63330078125
],
[
0.0882568359375,
0.6298828125,
0.48046875,
0.6298828125,
0.48046875,
0.64306640625,
0.0882568359375,
0.64306640625
],
[
0.0882568359375,
0.64599609375,
0.4794921875,
0.64599609375,
0.4794921875,
0.65576171875,
0.0882568359375,
0.65576171875
],
[
0.515625,
0.6484375,
0.88623046875,
0.64794921875,
0.88623046875,
0.662109375,
0.515625,
0.66357421875
],
[
0.08660888671875,
0.65771484375,
0.4794921875,
0.65771484375,
0.4794921875,
0.671875,
0.08660888671875,
0.671875
],
[
0.08660888671875,
0.67041015625,
0.4794921875,
0.67041015625,
0.4794921875,
0.6845703125,
0.08660888671875,
0.6845703125
],
[
0.51953125,
0.66845703125,
0.9111328125,
0.66845703125,
0.9111328125,
0.6787109375,
0.51953125,
0.6787109375
],
[
0.08660888671875,
0.68505859375,
0.48046875,
0.68505859375,
0.48046875,
0.69970703125,
0.08660888671875,
0.69970703125
],
[
0.5185546875,
0.6796875,
0.91259765625,
0.6796875,
0.91259765625,
0.693359375,
0.5185546875,
0.693359375
],
[
0.0849609375,
0.69873046875,
0.4794921875,
0.69775390625,
0.4794921875,
0.7119140625,
0.0849609375,
0.71337890625
],
[
0.51708984375,
0.693359375,
0.91259765625,
0.6943359375,
0.91259765625,
0.70849609375,
0.51708984375,
0.70703125
],
[
0.08660888671875,
0.7138671875,
0.191162109375,
0.7138671875,
0.191162109375,
0.7265625,
0.08660888671875,
0.7265625
],
[
0.521484375,
0.71044921875,
0.9111328125,
0.71044921875,
0.9111328125,
0.72021484375,
0.521484375,
0.72021484375
],
[
0.5185546875,
0.7216796875,
0.9140625,
0.7216796875,
0.9140625,
0.73486328125,
0.5185546875,
0.73486328125
],
[
0.08660888671875,
0.7373046875,
0.4794921875,
0.7373046875,
0.4794921875,
0.74755859375,
0.08660888671875,
0.74755859375
],
[
0.51953125,
0.7373046875,
0.7099609375,
0.7373046875,
0.7099609375,
0.74755859375,
0.51953125,
0.74755859375
],
[
0.0882568359375,
0.75048828125,
0.47802734375,
0.75048828125,
0.47802734375,
0.76123046875,
0.0882568359375,
0.76123046875
],
[
0.53515625,
0.751953125,
0.91259765625,
0.751953125,
0.91259765625,
0.76220703125,
0.53515625,
0.76220703125
],
[
0.08660888671875,
0.76318359375,
0.473876953125,
0.76318359375,
0.473876953125,
0.7763671875,
0.08660888671875,
0.7763671875
],
[
0.5185546875,
0.7646484375,
0.91259765625,
0.7646484375,
0.91259765625,
0.77783203125,
0.5185546875,
0.77783203125
],
[
0.23779296875,
0.7763671875,
0.48046875,
0.7763671875,
0.48046875,
0.79052734375,
0.23779296875,
0.79052734375
],
[
0.5185546875,
0.7763671875,
0.91259765625,
0.77783203125,
0.91259765625,
0.79150390625,
0.5185546875,
0.79052734375
],
[
0.2303466796875,
0.7822265625,
0.238525390625,
0.7822265625,
0.238525390625,
0.7861328125,
0.2303466796875,
0.7861328125
],
[
0.101318359375,
0.77880859375,
0.223876953125,
0.77880859375,
0.223876953125,
0.7890625,
0.101318359375,
0.7890625
],
[
0.08660888671875,
0.79150390625,
0.48193359375,
0.79150390625,
0.48193359375,
0.8046875,
0.08660888671875,
0.8046875
],
[
0.51953125,
0.79248046875,
0.91259765625,
0.79248046875,
0.91259765625,
0.80322265625,
0.51953125,
0.80322265625
],
[
0.0849609375,
0.80615234375,
0.48193359375,
0.80615234375,
0.48193359375,
0.8193359375,
0.0849609375,
0.8193359375
],
[
0.5185546875,
0.80615234375,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.5185546875,
0.8193359375
],
[
0.08660888671875,
0.8212890625,
0.48193359375,
0.8212890625,
0.48193359375,
0.83544921875,
0.08660888671875,
0.83544921875
],
[
0.51708984375,
0.8193359375,
0.91259765625,
0.8193359375,
0.91259765625,
0.83349609375,
0.51708984375,
0.83349609375
],
[
0.08251953125,
0.83544921875,
0.35546875,
0.83544921875,
0.35546875,
0.85986328125,
0.08251953125,
0.85986328125
],
[
0.51708984375,
0.83349609375,
0.740234375,
0.83349609375,
0.740234375,
0.8466796875,
0.51708984375,
0.8466796875
],
[
0.336669921875,
0.84033203125,
0.48193359375,
0.84033203125,
0.48193359375,
0.85400390625,
0.336669921875,
0.85400390625
],
[
0.08660888671875,
0.85498046875,
0.363525390625,
0.85498046875,
0.363525390625,
0.86865234375,
0.08660888671875,
0.86865234375
],
[
0.51953125,
0.85693359375,
0.9111328125,
0.85693359375,
0.9111328125,
0.8671875,
0.51953125,
0.8671875
],
[
0.10296630859375,
0.86865234375,
0.48046875,
0.86865234375,
0.48046875,
0.88330078125,
0.10296630859375,
0.88330078125
],
[
0.51953125,
0.87060546875,
0.9111328125,
0.87060546875,
0.9111328125,
0.880859375,
0.51953125,
0.880859375
],
[
0.08660888671875,
0.88330078125,
0.48046875,
0.88330078125,
0.48046875,
0.896484375,
0.08660888671875,
0.896484375
],
[
0.51953125,
0.88525390625,
0.69287109375,
0.88525390625,
0.69287109375,
0.89599609375,
0.51953125,
0.89599609375
]
] |
[
". For an unclicked ad (ie., Yelick = O and yeonu is \"unknown\"),",
"对每个样本的域进行分类(i。e。，单击或取消单击)基于示例",
"the models tend to optimize the predicted CVR scores to ze",
"的表示形式h(T）。直觉是，如果强点击鉴别器不能正确预测",
"ros. The proofis given here: The loss of CTCVR estimation is",
"样本的域标签，则其表示h(T）是点击自适应的",
" =-Ypu-conu log(PCTR·PCVR)-(1-ypu-conu) log(1-PCTR",
"总体而言，教师模型的正向过程为：",
"PcVR).For an unclicked ad whose (yelick, yconu, ypu-con) =",
"(unwn)hadnt",
"OPCVR1-PCTR-PCVR",
"h(T) = Tf(x)",
"PCTR",
"p(Tn=softmax(T (h(T)= (pCVR.1-PCVR)",
". Note that pcTR E (o, 1) and pCVR E (o, i), thus the gra-",
"",
"(5)",
"dient is always positive, which means that it tends to op",
"Pd = softmax (Ta (h(T))",
"timize pcvR of unclicked ads to O , but the actual label is",
"“unknown”\"",
"其中pCOnu是预测的CVR分布（PCVR是预测的CVR分数）。",
"基于反事实学习的模型已经实现了最先进的性能。但是，它",
"Pd是预测的域分布",
"们的限制包含：",
"要学习点击自适应表示，给定印象广告x，Tr(）学习的",
". For IPS-based models, the training procedure of the CVR",
"其表示h(T）旨在混淆点击鉴别器域并最大化域分类损失，而",
"taskisonclickeddataanddoesnotexplicitlytakeunclicked",
"点击鉴别器Ta(）本身旨在最大程度地减少分类损失。教师通",
"ads intoaccount.",
"过对抗性学习过程进行优化：",
"· For DR-based models, although the imputation task is used",
"to estimate CVRloss ofeach unclicked ad,its learning proce-",
"minLCVR=Delickl Delick",
"(T)",
"1",
"1",
"l(yconu,Pcong)",
"1",
"(6)",
"dure still utilizes clicked data only, and thus the imputation",
"",
"f.1p",
"-",
"task is lack of accurate supervision.",
"x (uelec a)",
"D",
"(7)",
"3",
"PROPOSEDMETHOD",
"TrTd",
"",
"我们提出了一个U不确定性正则化的KnowledgeD蒸馏（UKD）",
"第一个方程使CVR估计的损失最小化，以优化学习者Tr(）和",
"框架，该框架旨在通过从未点击的广告中提取知识来消除CVR",
"预测器Tp(）。第二个意味着学习者Tr（）使点击和未点击广",
"估计的偏差。基本思想是，我们通过为未点击的广告产生可靠",
"告的表示无法区分，而点击鉴别器Ta（）经过优化，可以更好",
"的伪转换标签作为监督信号来构建整个空间CVR估计模型",
"地将点击广告与未点击广告区分开。在实践中，我们通过梯度",
"图。2说明了UKD的整体工作流程，它包括一个为未点击的",
"广告产生伪转化标签的点击自适应教师模型，以及一个提炼",
"反转E[7] 实现它",
"教师学到的有价值的知识以进行整个空间CVR估计的不确定",
"当训练有素的鉴别器无法区分它们时，来自两个域的学习",
"性正则化学生模型。接下来，我们分别从老师和学生的细节来",
"表示会有效地对齐。因此，基于点击自适应表示，教师能够对",
"未点击的广告进行可靠的CVR估计",
"阐述我们的UKD",
"3.1.2Produce Pseudo-Conversion Labelsfor Unclicked Ads:经过",
"3.1",
"Click-Adaptive Teacher Model",
"训练的教师模型对Dunclick中每个未点击的广告进行推理，以",
"UKD中老师的目标是在我们只能访问点击广告Dclick的真实",
"产生预测的CVR分布Pcono作为伪转换标签，其中推理的正",
"转换标签的情况下，为未点击广告Dunclick生成伪转换标签，",
"向过程仅包括Tr（）和Tp（），而不需要Ta()",
"从而促进CVR估计任务的整个空间训练",
"我们使用Dunclick=（(x,pcona）)来表示与伪转换标签耦合",
"点击和未点击样本的特征分布之间存在差异。为了具有对",
"的未单击样本，这些样本将用于训练整个空间CVR模型",
"未点击的广告进行准确推断的能力，教师模型可能不会学习特",
"定于点击样本的特征表示，而是学习点击自适应表示。我们建",
"3.2Uncertainty-Regularized Student Model",
"议从无监督域自适应的角度处理伪转换标签生成，其中源/目",
"标域是点击/未点击的空间，灵感来自[3]。通过这种方式，问",
"基于点击自适应教师模型学习的未点击广告的伪转换标签，我",
"题被公式化为在给定标记的点击广告（Dclick作为源域）的情",
"们的UKD框架进一步构建了基于知识提炼的学生模型E[11]",
"况下，为未标记的未点击广告（Dunclick作为目标域）产生可",
"该模型从点击广告（带有地面实况标签）和未点击广告（带有",
"靠的伪转换标签",
"伪标签）中学习，以执行整个空间CVR估计。与仅使用点击",
"样本训练的模型相比，我们的模型通过在训练期间明确考虑",
"3.1.1Click-AdaptiveRepresentationLearning：具体来说，我们",
"未点击的样本来缓解SSB问题",
"的点击自适应教师模型采用对抗性学习[7.24]，它引入了点",
"我们阐述了蒸馏策略，可以指导学生模型挖掘老师学到的",
"击鉴别器来减轻训练期间点击/未点击样本的不一致特征分布",
"有价值的知识。由于教师预测中存在固有的噪音，未点击广告",
"如图所示。2，点击自适应教师模型由",
"的伪标签的置信度不如点击广告的真实转化标签。为了解决",
"",
"Model Architecture",
"特征表示学习器Tr(）、CVR预测器Tp()和点击鉴别器Ta()",
"这个问题，我们提出了一个不确定性正则化的学生，通过对蒸",
"组成。形式上，Tr()）将每个样本的特征x作为输人来学习其",
"馏过程中伪标签的不确定性进行建模来减少噪声的负面影响。",
"密集表示h(T），其中Tr（）包含一个特征嵌入层和几个密集",
"接下来，我们详细介绍了学生模型的蒸馏策略及其两个关键",
"层。To）拟估计样本的CVR得分PR：",
"模块：标签蒸馏和不确定性正则化",
"R。它由几个致密层和",
"顶部的softmax函数组成，以产生概率分布",
"3.2.1BaseStudentModel:LabelDistillation，我们从引人基础",
"为了使特征表示h(T）更具点击适应性，以便于在未点击的",
"学生模型开始，该模型是与CVR和CTR估计任务共同学习的",
"广告上生成伪转换标签，教师模型引入了点击鉴别器Ta(）来",
"作为sE2.2.2中的联合模型"
] |
{
"arxiv_id": "2201.08024",
"language": "zh",
"page_idx": 2
}
|
||
2201.08024-zh-4
|
2201.08024-zh-4.jpg
|
<FigureHere>
Figure 3: KL-divergence of two predictors' outputs $(\times10^{-3}).$
从噪声数据中学习的研究$\mathbb{E}[10,13,22,32]$ 表明，两个神经模型对噪声训练样本的预测的不一致性通常大于对干净样本的预测。基于这样的指导，我们使用嘈杂的数据集$\mathcal{D}_click^{\prime}$来训练 CVR 模型，该模型包含一个嵌人层，一个表示学习器和顶部的两个单独的 CVR 预测器 (目标是两个预测器的交叉熵损失的平均值)。训练后，我们分别在(1)噪声样本和(2)干净样本上观察到两个预测因子输出的平均 KL-散度。图 3 显示了不同$k$ 的结果，对于$k$ 的每个值，噪声样本上的 KL-散度大于干净样本上的 KL-散度。这种现象可以解释为，如果样本的标签是干净的，那么在训练过程中，两个预测器都很容易拟合标签，其中对应的两个损失值很小，两个预测结果相似。相反，如果标签有噪声，则两个预测因子的拟合过程将很难保持一致，并且它们往往会产生不一致的预测，从而导致较大的方差。实验结果验证了两个 CVR 预测器的不一致性可以用来识别有噪声的训练样本
不确定性建模 受上述实验的启发，我们的不确定性正则化学生模型包含两个 CVR 预测因子$S_p^v(\cdot)$和$S_{p^{\prime}}^v(\cdot)$,以同时估计 CVR 分数 (如图右图所示。2 ), 然后将不确定性建模为它们之间的不一致性。形式上，让$p_{conv}$和$p_{conv}^{\prime}$表示来自两个CVR 预测因子的预测分布。我们将不确定性表示为两个预测的 KL 发散：
$$\begin{aligned}&\boldsymbol{p}_{conv}=S_{p}^{v}\left(\mathrm{dropout}(\boldsymbol{h}_{conv})\right)\:,\:\boldsymbol{p}_{conv}^{\prime}=S_{p^{\prime}}^{v}\left(\mathrm{dropout}^{\prime}(\boldsymbol{h}_{conv})\right)\\&\mathsf{KL}\left(\boldsymbol{p}_{conv}||\boldsymbol{p}_{conv}^{\prime}\right)=\boldsymbol{p}_{conv}\log\frac{\boldsymbol{p}_{conv}}{\boldsymbol{p}_{conv}^{\prime}}\end{aligned}$$
$\begin{array}{cc}(11)&(11)\\\text{其中我们将独立的退出操作应用于学习表示}&\boldsymbol{h}_{conv}\end{array},\begin{array}{cc}(11)\\\text{以增加}\end{array}$
两个预测变量的差异
具有不确定性正则化的蒸馏 基于每个未点击样品的估计不确定性，我们通过将基于不确定性的权重动态调整为 CVR 损失来减少蒸馏过程中嘈杂的未点击样品的负面影响。与基本学生模型相比，现在未点击广告的蒸馏过程被伪标签的不确定性正则化，从而减轻了教师预测中存在的固有噪声
对于每个未单击的样本，我们以因子$\exp\left(-\lambda\cdot\mathrm{KL}\left(\boldsymbol{p}_{conv}||\boldsymbol{p}_{conv}^{\prime}\right)\right)\in$
(0,1]作为不确定性正则化对其原始 CVR 损失进行加权。该因素与不确定性成反比$\left(\lambda\text{控制其规模}\right)$。因此，损失$\mathcal{L}_CVR\_unclick$ 被重新表述为：
$$\sum_{\tilde{\mathcal{D}}_{unclick}}\exp\left(-\lambda\cdot\mathsf{KL}\left(\boldsymbol{p}_{conv}||\boldsymbol{p}_{conv}'\right)\right)\cdot\ell\left(\boldsymbol{p}_{conv}^{(T)},\boldsymbol{p}_{conv}\right)$$
(12)
如果样本具有很高的不确定性，则该因子返回一个较小的值以降低其 CVR 损失。如果不确定度接近0，则因子趋于 1,这样的学生就会下放到基础学生模型
我们还添加了一个损失项$\sum_{\tilde{\mathcal{D}}_{unclick}}$KL$\left(\boldsymbol p_{conv}||\boldsymbol{p}_{conv}^{\prime}\right)$,它充当不确定性估计的正则化。如果不最小化这样的术语，大的KL-散度会导致小的标签蒸馏损失，这将使模型的优化效果不佳
Table 1: Statistics of five datasets.
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>$\operatorname{Impression}$ 市 1</th>
<th>Click 拼！</th>
<th>带 Conversion : I </th>
<th>折 Time Interval</th>
</tr>
<tr>
<td>Ali- CCP</td>
<td>$84M$</td>
<td>$3.4M$</td>
<td>$18k$</td>
<td>$N/A$</td>
</tr>
<tr>
<td>EC- Small</td>
<td>$0.18B$</td>
<td>$26M$</td>
<td>$88k$</td>
<td>39 Days</td>
</tr>
<tr>
<td>EC- Large</td>
<td>$0.51B$</td>
<td>$95M$</td>
<td>0.3M</td>
<td>72 Days</td>
</tr>
<tr>
<td>LS- Small</td>
<td>$0.28B$</td>
<td>$15M$</td>
<td>$0.2M$</td>
<td>311 Days</td>
</tr>
<tr>
<td>LS- Large</td>
<td>$0.75B$</td>
<td>$37M$</td>
<td>$0.5M$</td>
<td>92 Days</td>
</tr>
</tbody>
</table>
## 4 EXPERIMENTS
在本节中，我们进行离线和在线实验，并打算回答以下研究问
题：
$\bullet$ RQ1 (offline performance): Does our proposed UKD outper-
form the state-of-the-art CVR estimation models?
$\bullet$ RQ2 (teacher's utility): Is the teacher model necessary for
debiasing CVR estimation models, and does the choice of
teacher model affect the performance of our UKD ?
$\bullet$ RQ3 (student's utility): Does the distillation strategy of un-
certainty regularization effectively help the student model? $\bullet$ RQ4 (model analysis): Does our UKD benefit from incorporating more unclicked samples during distillation?
$\bullet$ RQ5 (online performance): Does our UKD achieve improvements when we deploy it to industrial advertising system?
## 4.1 Experimental Setup
4.1.1 $Datasets.$离线实验是在公共可用数据集 Ali-CCP 和来自领先广告平台的四个大规模生产数据集上进行的。表 1 列出了统计数据
公共数据集 阿里-CCP 数据集[16]是 CVR 估计的基准，从
淘宝的用户流量日志中收集
生产数据集 我们进一步从阿里巴巴广告平台收集了四个大规模生产数据集。具体来说，我们从两个不同的营销场景中收集了连续 3个月的用户反馈日志：第一个场景名为EC，其中包含用于吸引潜在客户的电子商务广告。第二个场景名为LS, 其中的广告是关于本地生活服务的。我们将它们组织成四个数据集：EC-小型/大型和 LS-小型/大型。它们的统计信息见表 1
4.1.2 $Competitors ofCVREstimation.$我们将 UKD 与以下强基线进行比较。除第一个模型外，其余所有模型均使用印象数据集$\mathcal{D}$进行训练：
$\bullet$ SingleCVR $( \S 2. 2. 1)$ is a single-task network that esti-
mates $\hat{p}_{CVR}$ , and is trained on clicked samples $\mathcal{D}_{click}$
$\bullet$ Joint $( \S 2. 2. 2)$ is a multi-task model that estimates both CVR and CTR. The CTR task is trained on impression data $\mathcal{D}$ , and the CVR task is trained on clicked data $\mathcal{D}_{click}$ .
$\bullet$ESMM$( \S 2. 3. 1)$ is an entire space model that learns CTR and CTCVR estimation [16]. It is optimized in impression space $\mathcal{D}$ ,where the predicted CTCVR is equal to $\hat{p}_{CTR}$ $\hat{p} _{CVR}$ .
$\bullet$ Division isavariant ofESMM,which formulates the CVR
estimation task as $p_{CVR}=p_{CTCVR}/p_{CTR}\left[16\right].$ Compared to ESMM , its two dense blocks produce $\hat{p}_{CTR}$ and $\hat{p}_{CTCVR}$ $\bullet$ CFL(§ 2.3.2) employs counterfactual learning and achieves the state-of-the-art performance [9,31].Here we implement the model in [31] as our competitor because we find
|
[
[
0.1422119140625,
0.10919189453125,
0.1544189453125,
0.10418701171875,
0.160888671875,
0.11297607421875,
0.148681640625,
0.11871337890625
],
[
0.6044921875,
0.10479736328125,
0.82421875,
0.1060791015625,
0.82421875,
0.117431640625,
0.6044921875,
0.11614990234375
],
[
0.2254638671875,
0.11297607421875,
0.310546875,
0.11297607421875,
0.310546875,
0.12060546875,
0.2254638671875,
0.12060546875
],
[
0.348876953125,
0.111083984375,
0.414306640625,
0.111083984375,
0.414306640625,
0.121826171875,
0.348876953125,
0.121826171875
],
[
0.1446533203125,
0.117431640625,
0.16015625,
0.117431640625,
0.16015625,
0.1275634765625,
0.1446533203125,
0.1275634765625
],
[
0.72802734375,
0.125,
0.79736328125,
0.1256103515625,
0.79736328125,
0.136962890625,
0.72802734375,
0.1363525390625
],
[
0.8056640625,
0.125,
0.8837890625,
0.1256103515625,
0.8837890625,
0.136962890625,
0.8056640625,
0.1363525390625
],
[
0.6806640625,
0.1256103515625,
0.71728515625,
0.1256103515625,
0.71728515625,
0.1363525390625,
0.6806640625,
0.1363525390625
],
[
0.54833984375,
0.1256103515625,
0.5908203125,
0.1256103515625,
0.5908203125,
0.1363525390625,
0.54833984375,
0.1363525390625
],
[
0.6044921875,
0.126953125,
0.66845703125,
0.126953125,
0.66845703125,
0.136962890625,
0.6044921875,
0.136962890625
],
[
0.5458984375,
0.14013671875,
0.5908203125,
0.14013671875,
0.5908203125,
0.1514892578125,
0.5458984375,
0.1514892578125
],
[
0.146240234375,
0.1287841796875,
0.16015625,
0.1287841796875,
0.16015625,
0.1395263671875,
0.146240234375,
0.1395263671875
],
[
0.62353515625,
0.14013671875,
0.650390625,
0.14013671875,
0.650390625,
0.15283203125,
0.62353515625,
0.15283203125
],
[
0.685546875,
0.14013671875,
0.71337890625,
0.14013671875,
0.71337890625,
0.15283203125,
0.685546875,
0.15283203125
],
[
0.75244140625,
0.14013671875,
0.77294921875,
0.14013671875,
0.77294921875,
0.15283203125,
0.75244140625,
0.15283203125
],
[
0.83251953125,
0.14013671875,
0.85888671875,
0.14013671875,
0.85888671875,
0.15283203125,
0.83251953125,
0.15283203125
],
[
0.351318359375,
0.146484375,
0.363525390625,
0.146484375,
0.363525390625,
0.149658203125,
0.351318359375,
0.149658203125
],
[
0.24267578125,
0.149658203125,
0.2548828125,
0.149658203125,
0.2548828125,
0.154052734375,
0.24267578125,
0.154052734375
],
[
0.82421875,
0.15087890625,
0.86669921875,
0.15283203125,
0.865234375,
0.1641845703125,
0.82275390625,
0.1622314453125
],
[
0.5439453125,
0.1514892578125,
0.59326171875,
0.1514892578125,
0.59326171875,
0.1634521484375,
0.5439453125,
0.1634521484375
],
[
0.1478271484375,
0.141357421875,
0.158447265625,
0.141357421875,
0.158447265625,
0.15087890625,
0.1478271484375,
0.15087890625
],
[
0.62158203125,
0.1514892578125,
0.65283203125,
0.1514892578125,
0.65283203125,
0.1634521484375,
0.62158203125,
0.1634521484375
],
[
0.685546875,
0.1514892578125,
0.71337890625,
0.1514892578125,
0.71337890625,
0.1634521484375,
0.685546875,
0.1634521484375
],
[
0.7509765625,
0.1514892578125,
0.775390625,
0.1514892578125,
0.775390625,
0.1634521484375,
0.7509765625,
0.1634521484375
],
[
0.54345703125,
0.160400390625,
0.59228515625,
0.1622314453125,
0.59228515625,
0.173583984375,
0.54345703125,
0.1717529296875
],
[
0.62158203125,
0.1610107421875,
0.65283203125,
0.1610107421875,
0.65283203125,
0.17236328125,
0.62158203125,
0.17236328125
],
[
0.685546875,
0.1610107421875,
0.71337890625,
0.1610107421875,
0.71337890625,
0.173583984375,
0.685546875,
0.173583984375
],
[
0.74853515625,
0.1610107421875,
0.77685546875,
0.1610107421875,
0.77685546875,
0.17236328125,
0.74853515625,
0.17236328125
],
[
0.82421875,
0.1590576171875,
0.86669921875,
0.1622314453125,
0.865234375,
0.1749267578125,
0.82275390625,
0.1717529296875
],
[
0.2320556640625,
0.1717529296875,
0.2442626953125,
0.1717529296875,
0.2442626953125,
0.1749267578125,
0.2320556640625,
0.1749267578125
],
[
0.340576171875,
0.1717529296875,
0.34716796875,
0.1717529296875,
0.34716796875,
0.1749267578125,
0.340576171875,
0.1749267578125
],
[
0.5458984375,
0.1717529296875,
0.59228515625,
0.1717529296875,
0.59228515625,
0.18310546875,
0.5458984375,
0.18310546875
],
[
0.62158203125,
0.1717529296875,
0.65283203125,
0.1717529296875,
0.65283203125,
0.184326171875,
0.62158203125,
0.184326171875
],
[
0.68701171875,
0.1717529296875,
0.71337890625,
0.1717529296875,
0.71337890625,
0.184326171875,
0.68701171875,
0.184326171875
],
[
0.74853515625,
0.1717529296875,
0.77783203125,
0.1717529296875,
0.77783203125,
0.184326171875,
0.74853515625,
0.184326171875
],
[
0.82421875,
0.17041015625,
0.86669921875,
0.17236328125,
0.865234375,
0.184326171875,
0.82275390625,
0.1817626953125
],
[
0.5439453125,
0.179931640625,
0.59228515625,
0.18310546875,
0.5908203125,
0.1956787109375,
0.54345703125,
0.1925048828125
],
[
0.62109375,
0.1817626953125,
0.65283203125,
0.1817626953125,
0.65283203125,
0.1932373046875,
0.62109375,
0.1932373046875
],
[
0.685546875,
0.1817626953125,
0.71337890625,
0.1817626953125,
0.71337890625,
0.1944580078125,
0.685546875,
0.1944580078125
],
[
0.74658203125,
0.1817626953125,
0.77685546875,
0.1817626953125,
0.77685546875,
0.1932373046875,
0.74658203125,
0.1932373046875
],
[
0.82421875,
0.18115234375,
0.86669921875,
0.18310546875,
0.865234375,
0.1944580078125,
0.82275390625,
0.1925048828125
],
[
0.0882568359375,
0.1925048828125,
0.47802734375,
0.1925048828125,
0.47802734375,
0.20263671875,
0.0882568359375,
0.20263671875
],
[
0.51953125,
0.2203369140625,
0.53173828125,
0.2203369140625,
0.53173828125,
0.2291259765625,
0.51953125,
0.2291259765625
],
[
0.54736328125,
0.2171630859375,
0.67724609375,
0.2171630859375,
0.67724609375,
0.2310791015625,
0.54736328125,
0.2310791015625
],
[
0.10296630859375,
0.2266845703125,
0.48046875,
0.2266845703125,
0.48046875,
0.240478515625,
0.10296630859375,
0.240478515625
],
[
0.0849609375,
0.238037109375,
0.48193359375,
0.2392578125,
0.48193359375,
0.25390625,
0.0849609375,
0.25244140625
],
[
0.5185546875,
0.23486328125,
0.91259765625,
0.23486328125,
0.91259765625,
0.248779296875,
0.5185546875,
0.248779296875
],
[
0.08416748046875,
0.251953125,
0.484375,
0.251953125,
0.484375,
0.272705078125,
0.08416748046875,
0.272705078125
],
[
0.51708984375,
0.250732421875,
0.54150390625,
0.250732421875,
0.54150390625,
0.26318359375,
0.51708984375,
0.26318359375
],
[
0.0882568359375,
0.26953125,
0.48193359375,
0.26953125,
0.48193359375,
0.28271484375,
0.0882568359375,
0.28271484375
],
[
0.54345703125,
0.268310546875,
0.9140625,
0.268310546875,
0.9140625,
0.2822265625,
0.54345703125,
0.2822265625
],
[
0.08660888671875,
0.28271484375,
0.48046875,
0.28271484375,
0.48046875,
0.296630859375,
0.08660888671875,
0.296630859375
],
[
0.55615234375,
0.2822265625,
0.85888671875,
0.2822265625,
0.85888671875,
0.29541015625,
0.55615234375,
0.29541015625
],
[
0.08660888671875,
0.296630859375,
0.48046875,
0.296630859375,
0.48046875,
0.31005859375,
0.08660888671875,
0.31005859375
],
[
0.54150390625,
0.294189453125,
0.91259765625,
0.29541015625,
0.91259765625,
0.309326171875,
0.54150390625,
0.30810546875
],
[
0.08416748046875,
0.31005859375,
0.48046875,
0.311279296875,
0.48046875,
0.324462890625,
0.08416748046875,
0.323974609375
],
[
0.55615234375,
0.30810546875,
0.9140625,
0.30810546875,
0.9140625,
0.3212890625,
0.55615234375,
0.3212890625
],
[
0.0882568359375,
0.324462890625,
0.48046875,
0.324462890625,
0.48046875,
0.33837890625,
0.0882568359375,
0.33837890625
],
[
0.55615234375,
0.323974609375,
0.86767578125,
0.323974609375,
0.86767578125,
0.337158203125,
0.55615234375,
0.337158203125
],
[
0.0882568359375,
0.33837890625,
0.48046875,
0.33837890625,
0.48046875,
0.3515625,
0.0882568359375,
0.3515625
],
[
0.54150390625,
0.3359375,
0.9140625,
0.337158203125,
0.9140625,
0.3515625,
0.54150390625,
0.35107421875
],
[
0.0882568359375,
0.354248046875,
0.4794921875,
0.354248046875,
0.4794921875,
0.3642578125,
0.0882568359375,
0.3642578125
],
[
0.55810546875,
0.3515625,
0.9111328125,
0.3515625,
0.9111328125,
0.3623046875,
0.55810546875,
0.3623046875
],
[
0.0882568359375,
0.3662109375,
0.48046875,
0.3662109375,
0.48046875,
0.380126953125,
0.0882568359375,
0.380126953125
],
[
0.54345703125,
0.3642578125,
0.91259765625,
0.365478515625,
0.91259765625,
0.37890625,
0.54345703125,
0.37744140625
],
[
0.08660888671875,
0.380126953125,
0.48046875,
0.380126953125,
0.48046875,
0.393310546875,
0.08660888671875,
0.393310546875
],
[
0.55810546875,
0.380615234375,
0.861328125,
0.380615234375,
0.861328125,
0.391357421875,
0.55810546875,
0.391357421875
],
[
0.09722900390625,
0.393310546875,
0.4794921875,
0.393310546875,
0.4794921875,
0.4072265625,
0.09722900390625,
0.4072265625
],
[
0.54150390625,
0.391357421875,
0.9140625,
0.393310546875,
0.9140625,
0.4072265625,
0.54150390625,
0.40478515625
],
[
0.08660888671875,
0.4072265625,
0.48046875,
0.406005859375,
0.48046875,
0.42041015625,
0.08660888671875,
0.421630859375
],
[
0.55615234375,
0.4072265625,
0.91015625,
0.4072265625,
0.91015625,
0.42041015625,
0.55615234375,
0.42041015625
],
[
0.08660888671875,
0.421630859375,
0.22216796875,
0.421630859375,
0.22216796875,
0.43505859375,
0.08660888671875,
0.43505859375
],
[
0.5185546875,
0.436279296875,
0.56201171875,
0.436279296875,
0.56201171875,
0.447509765625,
0.5185546875,
0.447509765625
],
[
0.5595703125,
0.434326171875,
0.7255859375,
0.43505859375,
0.7255859375,
0.449462890625,
0.5595703125,
0.448974609375
],
[
0.08660888671875,
0.451904296875,
0.48046875,
0.451904296875,
0.48046875,
0.46533203125,
0.08660888671875,
0.46533203125
],
[
0.10296630859375,
0.4375,
0.48193359375,
0.4375,
0.48193359375,
0.45068359375,
0.10296630859375,
0.45068359375
],
[
0.51953125,
0.455078125,
0.56201171875,
0.455078125,
0.56201171875,
0.46533203125,
0.51953125,
0.46533203125
],
[
0.560546875,
0.45263671875,
0.91259765625,
0.45263671875,
0.91259765625,
0.466552734375,
0.560546875,
0.466552734375
],
[
0.08660888671875,
0.46728515625,
0.48046875,
0.46728515625,
0.48046875,
0.48095703125,
0.08660888671875,
0.48095703125
],
[
0.51953125,
0.466552734375,
0.91259765625,
0.466552734375,
0.91259765625,
0.479736328125,
0.51953125,
0.479736328125
],
[
0.08660888671875,
0.481689453125,
0.4794921875,
0.481689453125,
0.4794921875,
0.496826171875,
0.08660888671875,
0.496826171875
],
[
0.521484375,
0.481689453125,
0.59326171875,
0.481689453125,
0.59326171875,
0.49365234375,
0.521484375,
0.49365234375
],
[
0.0849609375,
0.494384765625,
0.48046875,
0.49560546875,
0.48046875,
0.5087890625,
0.0849609375,
0.50830078125
],
[
0.53515625,
0.494384765625,
0.91259765625,
0.494384765625,
0.91259765625,
0.50830078125,
0.53515625,
0.50830078125
],
[
0.08660888671875,
0.51025390625,
0.160888671875,
0.51025390625,
0.160888671875,
0.521484375,
0.08660888671875,
0.521484375
],
[
0.51953125,
0.51025390625,
0.69677734375,
0.51025390625,
0.69677734375,
0.52099609375,
0.51953125,
0.52099609375
],
[
0.53515625,
0.52294921875,
0.91259765625,
0.52294921875,
0.91259765625,
0.5361328125,
0.53515625,
0.5361328125
],
[
0.096435546875,
0.53271484375,
0.468017578125,
0.53271484375,
0.468017578125,
0.54736328125,
0.096435546875,
0.54736328125
],
[
0.51953125,
0.53515625,
0.91259765625,
0.53515625,
0.91259765625,
0.548828125,
0.51953125,
0.548828125
],
[
0.279296875,
0.548828125,
0.318603515625,
0.5517578125,
0.31787109375,
0.564453125,
0.27783203125,
0.56103515625
],
[
0.5185546875,
0.5498046875,
0.9140625,
0.5498046875,
0.9140625,
0.56298828125,
0.5185546875,
0.56298828125
],
[
0.279296875,
0.56298828125,
0.32177734375,
0.56494140625,
0.3203125,
0.58056640625,
0.27783203125,
0.5791015625
],
[
0.0931396484375,
0.5537109375,
0.28759765625,
0.5546875,
0.28759765625,
0.57275390625,
0.0931396484375,
0.5712890625
],
[
0.51953125,
0.564453125,
0.91259765625,
0.564453125,
0.91259765625,
0.57763671875,
0.51953125,
0.57763671875
],
[
0.455810546875,
0.5771484375,
0.48193359375,
0.5771484375,
0.48193359375,
0.58837890625,
0.455810546875,
0.58837890625
],
[
0.51708984375,
0.57763671875,
0.91259765625,
0.57763671875,
0.91259765625,
0.59130859375,
0.51708984375,
0.59130859375
],
[
0.0931396484375,
0.5888671875,
0.48046875,
0.58837890625,
0.48046875,
0.60302734375,
0.0931396484375,
0.60400390625
],
[
0.51953125,
0.59130859375,
0.91259765625,
0.59130859375,
0.91259765625,
0.60498046875,
0.51953125,
0.60498046875
],
[
0.0849609375,
0.60400390625,
0.221435546875,
0.60400390625,
0.221435546875,
0.6171875,
0.0849609375,
0.6171875
],
[
0.5185546875,
0.60693359375,
0.5458984375,
0.60693359375,
0.5458984375,
0.61865234375,
0.5185546875,
0.61865234375
],
[
0.10296630859375,
0.619140625,
0.4794921875,
0.619140625,
0.4794921875,
0.63330078125,
0.10296630859375,
0.63330078125
],
[
0.51953125,
0.6298828125,
0.56201171875,
0.6298828125,
0.56201171875,
0.64013671875,
0.51953125,
0.64013671875
],
[
0.560546875,
0.62890625,
0.91259765625,
0.62890625,
0.91259765625,
0.642578125,
0.560546875,
0.642578125
],
[
0.5185546875,
0.642578125,
0.91259765625,
0.642578125,
0.91259765625,
0.65576171875,
0.5185546875,
0.65576171875
],
[
0.0849609375,
0.63330078125,
0.48046875,
0.6318359375,
0.48046875,
0.64599609375,
0.0849609375,
0.646484375
],
[
0.0849609375,
0.646484375,
0.48193359375,
0.64794921875,
0.48193359375,
0.6611328125,
0.0849609375,
0.66015625
],
[
0.08660888671875,
0.6611328125,
0.48046875,
0.6611328125,
0.48046875,
0.6748046875,
0.08660888671875,
0.6748046875
],
[
0.5185546875,
0.6572265625,
0.62109375,
0.6572265625,
0.62109375,
0.66845703125,
0.5185546875,
0.66845703125
],
[
0.0882568359375,
0.6767578125,
0.439453125,
0.6767578125,
0.439453125,
0.6875,
0.0882568359375,
0.6875
],
[
0.54345703125,
0.6748046875,
0.9140625,
0.6748046875,
0.9140625,
0.68798828125,
0.54345703125,
0.68798828125
],
[
0.10296630859375,
0.68798828125,
0.51708984375,
0.68798828125,
0.51708984375,
0.70263671875,
0.10296630859375,
0.70263671875
],
[
0.5556640625,
0.68798828125,
0.8896484375,
0.689453125,
0.8896484375,
0.7041015625,
0.5556640625,
0.70263671875
],
[
0.08660888671875,
0.70263671875,
0.48193359375,
0.70263671875,
0.48193359375,
0.71630859375,
0.08660888671875,
0.71630859375
],
[
0.54345703125,
0.70263671875,
0.91259765625,
0.70263671875,
0.91259765625,
0.71630859375,
0.54345703125,
0.71630859375
],
[
0.0849609375,
0.71337890625,
0.48193359375,
0.71630859375,
0.48193359375,
0.734375,
0.0849609375,
0.73095703125
],
[
0.55615234375,
0.71630859375,
0.91259765625,
0.71728515625,
0.91259765625,
0.72900390625,
0.55615234375,
0.72802734375
],
[
0.0882568359375,
0.73193359375,
0.1805419921875,
0.73193359375,
0.1805419921875,
0.74365234375,
0.0882568359375,
0.74365234375
],
[
0.55615234375,
0.72900390625,
0.8935546875,
0.73095703125,
0.8935546875,
0.74560546875,
0.55615234375,
0.74365234375
],
[
0.54150390625,
0.74365234375,
0.91259765625,
0.744140625,
0.91259765625,
0.75830078125,
0.54150390625,
0.7568359375
],
[
0.1258544921875,
0.755859375,
0.136474609375,
0.755859375,
0.136474609375,
0.76513671875,
0.1258544921875,
0.76513671875
],
[
0.1356201171875,
0.7548828125,
0.35546875,
0.755859375,
0.35546875,
0.7734375,
0.1356201171875,
0.77294921875
],
[
0.455078125,
0.75830078125,
0.48046875,
0.75830078125,
0.48046875,
0.77099609375,
0.455078125,
0.77099609375
],
[
0.55615234375,
0.75830078125,
0.91259765625,
0.75830078125,
0.91259765625,
0.771484375,
0.55615234375,
0.771484375
],
[
0.12091064453125,
0.7646484375,
0.1446533203125,
0.75830078125,
0.1502685546875,
0.77099609375,
0.125,
0.7763671875
],
[
0.34716796875,
0.7587890625,
0.431396484375,
0.7587890625,
0.431396484375,
0.77099609375,
0.34716796875,
0.77099609375
],
[
0.1094970703125,
0.7734375,
0.158447265625,
0.77587890625,
0.1568603515625,
0.79052734375,
0.108642578125,
0.78857421875
],
[
0.5556640625,
0.76806640625,
0.9111328125,
0.77099609375,
0.9111328125,
0.78857421875,
0.5556640625,
0.7861328125
],
[
0.55810546875,
0.787109375,
0.6005859375,
0.787109375,
0.6005859375,
0.798828125,
0.55810546875,
0.798828125
],
[
0.08660888671875,
0.796875,
0.48046875,
0.796875,
0.48046875,
0.81005859375,
0.08660888671875,
0.81005859375
],
[
0.5439453125,
0.80029296875,
0.619140625,
0.80029296875,
0.619140625,
0.8125,
0.5439453125,
0.8125
],
[
0.61767578125,
0.8046875,
0.625,
0.8046875,
0.625,
0.8095703125,
0.61767578125,
0.8095703125
],
[
0.62158203125,
0.7998046875,
0.9140625,
0.7998046875,
0.9140625,
0.81298828125,
0.62158203125,
0.81298828125
],
[
0.08660888671875,
0.81103515625,
0.48193359375,
0.81103515625,
0.48193359375,
0.82470703125,
0.08660888671875,
0.82470703125
],
[
0.5556640625,
0.81298828125,
0.9140625,
0.8125,
0.9140625,
0.8271484375,
0.5556640625,
0.82763671875
],
[
0.08660888671875,
0.82470703125,
0.30810546875,
0.82470703125,
0.30810546875,
0.83837890625,
0.08660888671875,
0.83837890625
],
[
0.5556640625,
0.82568359375,
0.9140625,
0.82763671875,
0.9140625,
0.84228515625,
0.5556640625,
0.84033203125
],
[
0.1004638671875,
0.83203125,
0.330810546875,
0.83837890625,
0.330078125,
0.859375,
0.098876953125,
0.85302734375
],
[
0.330810546875,
0.8369140625,
0.483642578125,
0.8369140625,
0.483642578125,
0.85498046875,
0.330810546875,
0.85498046875
],
[
0.08660888671875,
0.85498046875,
0.4794921875,
0.85498046875,
0.4794921875,
0.86865234375,
0.08660888671875,
0.86865234375
],
[
0.54345703125,
0.85400390625,
0.9189453125,
0.85498046875,
0.9189453125,
0.869140625,
0.54345703125,
0.86865234375
],
[
0.0849609375,
0.86865234375,
0.48046875,
0.86865234375,
0.48046875,
0.8818359375,
0.0849609375,
0.8818359375
],
[
0.5556640625,
0.8671875,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.5556640625,
0.880859375
],
[
0.08660888671875,
0.88330078125,
0.10455322265625,
0.88330078125,
0.10455322265625,
0.896484375,
0.08660888671875,
0.896484375
],
[
0.5556640625,
0.88330078125,
0.9140625,
0.8818359375,
0.9140625,
0.89599609375,
0.5556640625,
0.896484375
]
] |
[
"22",
"Table 1:Statistics of five datasets.",
"NoisySamples",
"CleanSamples",
"2.0",
"#Conversion",
"#Time Interval",
"# Click",
"Dataset",
"#Impression",
"Ali-CCP",
"1.8",
"84M",
"3.4M",
"18k",
"N/A",
"",
"",
"39 Days",
"EC-Small",
"1.6",
"0.18B",
"26M",
"88k",
"EC-Large",
"0.51B",
"95M",
"0.3M",
"72 Days",
"2",
"4",
"LS-Small",
"0.28B",
"15M",
"0.2M",
"31Days",
"LS-Large",
"0.75B",
"37M",
"0.5M",
"92 Days",
"Figure3:KL-divergenceoftwopredictors'outputs(x1o-3)",
"4",
"EXPERIMENTS",
"从噪声数据中学习的研究国[10,13,22,32]表明，两个神经",
"模型对噪声训练样本的预测的不一致性通常大于对干净样本",
"在本节中，我们进行离线和在线实验，并打算回答以下研究问",
"的预测。基于这样的指导，我们使用噜杂的数据集 Dclick来",
"题：",
"训练CVR模型，该模型包含一个嵌入层，一个表示学习器和",
".RQ1 (offline performance): Does our proposed UKD outper-",
"顶部的两个单独的CVR预测器（目标是两个预测器的交叉摘",
"form the state-of-the-art CVR estimation models?",
"损失的平均值）。训练后，我们分别在（1）噪声样本和（2）干",
".RQ2 (teacher's utility): Is the teacher model necessary for",
"净样本上观察到两个预测因子输出的平均KL-散度。图3显示",
"debiasing CVR estimation models, and does the choice of",
"了不同k的结果，对于k的每个值，噪声样本上的KL-散度大",
"teacher model affect the performance of our UKD ?",
"于干净样本上的KL-散度。这种现象可以解释为，如果样本的",
". RQ3 (student's utility): Does the distillation strategy of un-",
"标签是于净的，那么在训练过程中，两个预测器都很容易拟",
"certaintyregularization effectivelyhelpthestudentmodel?",
"合标签，其中对应的两个损失值很小，两个预测结果相似。相",
".RQ4 (model analysis):Does our UKD benefitfrom incorpo-",
"反，如果标签有噪声，则两个预测因子的拟合过程将很难保持",
"rating more unclicked samples during distillation?",
"一致，并且它们往往会产生不一致的预测，从而导致较大的方",
"RO5 (online performance):Does ourUKD achieve improve-",
"差。实验结果验证了两个CVR预测器的不一致性可以用来识",
"ments when we deploy it to industrial advertising system?",
"别有噪声的训练样本",
"4.1",
"Experimental Setup",
"化学生模型包含两个CVR预测因子S（）和S（），以同时估",
"不确定性建模，受上述实验的启发，我们的不确定性正则",
"4.1.1.",
"Datasets：离线实验是在公共可用数据集Ali-CCP和来",
"计CVR分数（如图右图所示。2），然后将不确定性建模为它",
"自领先广告平台的四个大规模生产数据集上进行的。表1列出",
"们之间的不一致性。形式上，让Pconu和Pconu表示来自两个",
"了统计数据",
"CVR预测因子的预测分布。我们将不确定性表示为两个预测",
"公共数据集阿里-CCP数据集[16]是CVR估计的基准，从",
"的 KL 发散：",
"淘宝的用户流量日志中收集",
"生产数据集我们进一步从阿里巴巴广告平台收集了四个",
"Pconu = Sp (dropout(hcong)), Pconu = Sp, (dropout'(hconu))",
"大规模生产数据集。具体来说，我们从两个不同的营销场景中",
"Pconu",
"收集了连续3个月的用户反馈日志：第一个场景名为EC，其",
"Pconu",
"KL(pconullpcong) = Pconu log ",
"中包含用于吸引潜在客户的电子商务广告。第二个场景名为",
"(11)",
"LS，其中的广告是关于本地生活服务的。我们将它们组织成四",
"其中我们将独立的退出操作应用于学习表示hconu；以增加",
"个数据集：EC-小型/大型和LS-小型/大型。它们的统计信息见",
"两个预测变量的差异",
"表1",
"具有不确定性正则化的蒸馏基于每个未点击样品的估计",
"4.1.2",
"CompetitorsofCVREstimation，我们将UKD与以下强基",
"线进行比较。除第一个模型外，其余所有模型均使用印象数据",
"不确定性，我们通过将基于不确定性的权重动态调整为CVR",
"损失来减少蒸馏过程中噜杂的未点击样品的负面影响。与基",
"本学生模型相比，现在未点击广告的蒸馏过程被伪标签的不",
"集D进行训练：",
"确定性正则化，从而减轻了教师预测中存在的固有噪声",
". SingleCVR (s 2.2.1 ) is a single-task network that esti-",
"对于每个未单击的样本，我们以因子exp（-入·KL（pconullpconu））",
"mates pcvR , and is trained on clicked samples Delick .",
"(O,1)作为不确定性正则化对其原始CVR损失进行加权。该因",
".Joint (s 2.2.2)is a multi-task model that estimates both",
"素与不确定性成反比(入控制其规模)。因此，损失LcvR_unclick",
"CVRandCTR.TheCTRtaskistrainedonimpressiondata",
"被重新表述为：",
"D, and the CVR task is trained on clicked data Delick :",
"ESMM(S2.3.1)is anentirespacemodel that learnsCTR",
"1",
"C exp(-α-KL (Pconllpconu)) -(",
"(12)",
"and CTCVR estimation [16] .It is optimized in impression",
"2",
"(Pconu,Pconu",
"Dunclick",
"space D , where the predicted CTCVR is equal to PCTR ",
"PCVR.",
"如果样本具有很高的不确定性，则该因子返回一个较小的值",
".Division",
"",
"is a variant of ESMM, which formulates the CVR",
"以降低其CVR损失。如果不确定度接近0，则因子趋于1，这",
"estimation task as PCVR=PCTCVR/PCTR[16].Compared",
"样的学生就会下放到基础学生模型",
"to ESMM, its two dense blocks produce pCTR and pcTCVR",
"我们还添加了一个损失项ZDunclie",
"KL(pconollPcon)，它",
"充当不确定性估计的正则化。如果不最小化这样的术语，大的",
".CFL(S2.3.2)employs counterfactual learning and achieves",
"KL-散度会导致小的标签蒸馏损失，这将使模型的优化效果不",
"the state-of-the-art performance [9, 31] .Here we imple-",
"佳",
"ment the model in [31] as our competitor because we find"
] |
{
"arxiv_id": "2201.08024",
"language": "zh",
"page_idx": 4
}
|
||
2201.13176-zh-4
|
2201.13176-zh-4.jpg
|
a UCT formula:
$$a^*=\underset{a}{\mathrm{argmax}}(Q(a)+U(a)),\quad Q(a)=\frac{1}{|V_a|}\sum_{i\in V_a}S(i)$$
where $S(i)$ is the scaled score difference estimated by the value head for node $i,Q(a)$ is the mean action value over the set $V_a$ of the visited MCTS tree nodes that follow $a$ , and $U(a)$ is the same as in [1] ,with the $c_\mathrm{puct}$ constant raised from 1 to 1.5 to account for the larger variability of the scaled scores of siblings with respect to the winrate probabilities.
$\bullet$ Game ending. LZ employs
self-plays to avoid uselessl most games are allowed to en the winrate drops below $5\%$ ,and there are some hard-coded conditions for passing. These heuristics do not work anymore in our new setting: to maximize the score, the agent should keep playing to the very end, so that the computationally amenable Tromp-Taylor score becomes equivalent to the proper area score of the game of Go.
# 使用$\rho<1$将电路板的尺寸从$N=19$缩小到$\rho N=9$会带来一些好处
$\bullet$ Average number of legal moves at each position scales down by $\rho ^{2}$ . The number of good, meaningful moves might scale similarly.
$\bullet$ Average length of games scales down by $\rho^2$
might scale down by as much as $\rho^4$ , as can be inferred
from the previous two.
$\bullet$ The suggested number of layers in the residual con-
volutional neural network stack scales by $\rho.$
$\bullet$ The fully connected layers at the end of the neural
network stack are smaller.
总结一下性能方面的好处，可以估计自玩游戏的总速度提高大约为$\rho^{9}$ ,即大约为 10$^-3$ 。实际上，我们的资源使我们能够实施具有更多访问次数的管道。因此，我们选择了相同数量的 9×9 SAI 访问，以便进行公平的比较
$B.$ Reinforcement learning $p$
我们的培训由多个阶段组成，灵感来自原始的 LZ 培训过程以及 SAI 有据可查的管道。特别是，我们遵循了与$9\times9$ ESAIE[7]的第一次运行几乎相同的训练过程，唯一的例外是我们重新引人了门控 (i。e。最好的网络只有在更强大的网络出现时才会被替换)。从 AlphaGoZero 到AlphaZero 的段落中删除了 Gating,因为它被认为不是严格必要的：由于我们正在尝试基于分数的强化学习，因此我们认为将其保留在管道中更为保守。实际上，当我们训练网络以最大化其得分时，通过门控，我们正在评估其赢得比赛的能力。每个训练周期称为一代，它由以下步骤组成。
1) Self-play. LZS engine use
to play a total of 2,000 self-play games (against itself and sharing the tree search), with the following parameters: fixed number of visits $v$ depending on AlphaGo 类代理，并从经验上表明，尽管围棋游戏是确定
the generation (see the table below), random choice of the frst 15 moves (with probability proportional to the visits number), Dirichlet noise at the root. The agent is allowed to pass only if there is no other $\begin{array}{c}\text{move with an expected score larger than the Tromp-}\\\end{array}$ $\underline{\text{Taylor score.}}$
<table>
<tbody>
<tr>
<th>generations</th>
<th>$v$</th>
<th>$w$</th>
<th>generations</th>
<th>$v$</th>
<th>$w$</th>
</tr>
<tr>
<td>$0–15$</td>
<td>100</td>
<td>4</td>
<td>$16–31$</td>
<td>150</td>
<td>8</td>
</tr>
<tr>
<td>$32–47$</td>
<td>250</td>
<td>12</td>
<td>$48–63$</td>
<td>400</td>
<td>16 1</td>
</tr>
<tr>
<td>$64–79$</td>
<td>600</td>
<td>20</td>
<td>$80-\infty$</td>
<td>850</td>
<td>24</td>
</tr>
</tbody>
</table>
$2){\mathrm{~Training.~Starting~from~the~current~best~network}}$,
new training is started over a window of the most recent table above). The number of steps adapts to the number of games inside the window, with a typical number of 30,000 steps. The batch size is 512.
3) Gating. The newly trained network is matched
against the current best, for a total of 400 games, with the same number of visits for self-play games and no randomness in the move choice. The agent is allowed to pass when UCT selects this move. If at least $55\%$ of the 400 matches are won by the candidate network, it is promoted to be the new best network.
所有游戏均使用 komi 7.5 进行，也就是说，在白人玩家的Tromp-Taylor 得分上增加了 7.5 分，以补偿黑人扮演第一步的事实
$\bullet$ The suggested number of visits in the MCTS tree $C.Outcomeofthelearningpipeline$
学习管道由门匹配的中间结果引导。实际上，在 LZ 之后，在每一代中，我们都从门控比赛中的 400 场比赛中估算了新网络的 Elo 等级，将第一个随机网络任意锚定到Elo0。这个评级显然被高估了(例如，每个提升的网络都被限制为至少有 34 个。8 分 , -对应于一个 55 的%winrate-比前一个更多), 所以我们把它称为未校准的 Elo。尽管如此， 该估计足以评估学习管道何时停止：当没有新网络克服太多代 (经验选择)的门控时，我们扩大了网络规模或降低了训练优化器的学习率
从随机的$2\times64$网络(具有 2 个剩余的 64 卷积块)开始，我们在 2,000 场比赛中移至$4\times128$ ;在154,000场比赛中移至$8\times160$ ;在 748,000 场比赛中移至$10\times192$ ;在 982,000 场比赛中移至 12$\times256$ 。学习率从 0 开始。02 和减少一次到0。002 在 850,000 场比赛中。Fig。5 根据未校准的 Elo 等级显示管道结果
## VI. CONCLUSIONS
在本文中，我们讨论了在基于分数的赢/输游戏中最大化分数在胜率方面何时以及为什么可能是次优的。由于这在确定性游戏中不可能发生，因此我们从最简单的非确定性案例开始，即多臂匪徒。我们证明，在多臂匪徒中，次优行为与分数方差以及代理是赢还是输有关：当失败时， 基于结果的代理更喜欢导致更大分数方差的行为，反之亦然。然后，我们定义了一类对基于得分的输赢游戏进行建模的 MDP,并通过经验表明，次优游戏与得分方差之间的这种关系仍然有效。最后，我们训练了一个基于分数的
|
[
[
0.1094970703125,
0.0631103515625,
0.2279052734375,
0.0631103515625,
0.2279052734375,
0.0770263671875,
0.1094970703125,
0.0770263671875
],
[
0.54833984375,
0.0631103515625,
0.9189453125,
0.0631103515625,
0.9189453125,
0.0770263671875,
0.54833984375,
0.0770263671875
],
[
0.420654296875,
0.083984375,
0.484375,
0.09027099609375,
0.48193359375,
0.1060791015625,
0.418212890625,
0.0997314453125
],
[
0.54833984375,
0.07891845703125,
0.9189453125,
0.07891845703125,
0.9189453125,
0.0921630859375,
0.54833984375,
0.0921630859375
],
[
0.11767578125,
0.09027099609375,
0.3349609375,
0.09027099609375,
0.3349609375,
0.10418701171875,
0.11767578125,
0.10418701171875
],
[
0.40185546875,
0.08587646484375,
0.414306640625,
0.08587646484375,
0.414306640625,
0.0921630859375,
0.40185546875,
0.0921630859375
],
[
0.41259765625,
0.09027099609375,
0.423095703125,
0.09027099609375,
0.423095703125,
0.09344482421875,
0.41259765625,
0.09344482421875
],
[
0.39208984375,
0.091552734375,
0.455810546875,
0.1060791015625,
0.44775390625,
0.12371826171875,
0.384033203125,
0.10919189453125
],
[
0.328369140625,
0.091552734375,
0.375732421875,
0.091552734375,
0.375732421875,
0.1029052734375,
0.328369140625,
0.1029052734375
],
[
0.374267578125,
0.0947265625,
0.378173828125,
0.0985107421875,
0.374267578125,
0.10101318359375,
0.370849609375,
0.09661865234375
],
[
0.54833984375,
0.09344482421875,
0.9189453125,
0.09344482421875,
0.9189453125,
0.106689453125,
0.54833984375,
0.106689453125
],
[
0.54833984375,
0.10919189453125,
0.9189453125,
0.10919189453125,
0.9189453125,
0.12249755859375,
0.54833984375,
0.12249755859375
],
[
0.11273193359375,
0.1256103515625,
0.490234375,
0.1256103515625,
0.490234375,
0.1395263671875,
0.11273193359375,
0.1395263671875
],
[
0.54736328125,
0.125,
0.9189453125,
0.125,
0.9189453125,
0.1383056640625,
0.54736328125,
0.1383056640625
],
[
0.11273193359375,
0.141357421875,
0.490234375,
0.141357421875,
0.490234375,
0.1546630859375,
0.11273193359375,
0.1546630859375
],
[
0.54736328125,
0.136962890625,
0.640625,
0.1395263671875,
0.640625,
0.154052734375,
0.54736328125,
0.1514892578125
],
[
0.55810546875,
0.1546630859375,
0.64892578125,
0.1546630859375,
0.64892578125,
0.169189453125,
0.55810546875,
0.169189453125
],
[
0.73388671875,
0.1546630859375,
0.8310546875,
0.1546630859375,
0.8310546875,
0.169189453125,
0.73388671875,
0.169189453125
],
[
0.70166015625,
0.1572265625,
0.73388671875,
0.1572265625,
0.73388671875,
0.1685791015625,
0.70166015625,
0.1685791015625
],
[
0.111083984375,
0.1558837890625,
0.491943359375,
0.1558837890625,
0.491943359375,
0.169189453125,
0.111083984375,
0.169189453125
],
[
0.8447265625,
0.1572265625,
0.86767578125,
0.1572265625,
0.86767578125,
0.1685791015625,
0.8447265625,
0.1685791015625
],
[
0.88134765625,
0.1572265625,
0.90869140625,
0.1572265625,
0.90869140625,
0.169189453125,
0.88134765625,
0.169189453125
],
[
0.89794921875,
0.1673583984375,
0.91015625,
0.1767578125,
0.89794921875,
0.186279296875,
0.8857421875,
0.1767578125
],
[
0.6689453125,
0.1578369140625,
0.685546875,
0.1578369140625,
0.685546875,
0.1673583984375,
0.6689453125,
0.1673583984375
],
[
0.6640625,
0.17041015625,
0.69287109375,
0.17041015625,
0.69287109375,
0.1849365234375,
0.6640625,
0.1849365234375
],
[
0.76318359375,
0.17041015625,
0.80615234375,
0.17041015625,
0.80615234375,
0.1849365234375,
0.76318359375,
0.1849365234375
],
[
0.84228515625,
0.169189453125,
0.87353515625,
0.169189453125,
0.87353515625,
0.1849365234375,
0.84228515625,
0.1849365234375
],
[
0.58642578125,
0.1717529296875,
0.62353515625,
0.1717529296875,
0.62353515625,
0.18310546875,
0.58642578125,
0.18310546875
],
[
0.11273193359375,
0.1717529296875,
0.490234375,
0.1717529296875,
0.490234375,
0.1849365234375,
0.11273193359375,
0.1849365234375
],
[
0.7099609375,
0.17236328125,
0.7236328125,
0.17236328125,
0.7236328125,
0.18310546875,
0.7099609375,
0.18310546875
],
[
0.5810546875,
0.184326171875,
0.62744140625,
0.184326171875,
0.62744140625,
0.199462890625,
0.5810546875,
0.199462890625
],
[
0.6611328125,
0.184326171875,
0.69287109375,
0.184326171875,
0.69287109375,
0.199462890625,
0.6611328125,
0.199462890625
],
[
0.70751953125,
0.1849365234375,
0.72607421875,
0.1849365234375,
0.72607421875,
0.1988525390625,
0.70751953125,
0.1988525390625
],
[
0.76123046875,
0.1849365234375,
0.80810546875,
0.1849365234375,
0.80810546875,
0.199462890625,
0.76123046875,
0.199462890625
],
[
0.8408203125,
0.184326171875,
0.87353515625,
0.184326171875,
0.87353515625,
0.199462890625,
0.8408203125,
0.199462890625
],
[
0.88623046875,
0.1849365234375,
0.90673828125,
0.1849365234375,
0.90673828125,
0.199462890625,
0.88623046875,
0.199462890625
],
[
0.83837890625,
0.1976318359375,
0.8759765625,
0.199462890625,
0.875,
0.21533203125,
0.83642578125,
0.21337890625
],
[
0.111083984375,
0.186279296875,
0.490234375,
0.1849365234375,
0.490234375,
0.199462890625,
0.111083984375,
0.2008056640625
],
[
0.5810546875,
0.1988525390625,
0.62744140625,
0.2008056640625,
0.6259765625,
0.216552734375,
0.580078125,
0.2139892578125
],
[
0.6611328125,
0.2008056640625,
0.69287109375,
0.2008056640625,
0.69287109375,
0.21337890625,
0.6611328125,
0.21337890625
],
[
0.705078125,
0.2008056640625,
0.72607421875,
0.2008056640625,
0.72607421875,
0.21337890625,
0.705078125,
0.21337890625
],
[
0.7587890625,
0.199462890625,
0.810546875,
0.199462890625,
0.810546875,
0.21533203125,
0.7587890625,
0.21533203125
],
[
0.88134765625,
0.2020263671875,
0.90869140625,
0.2020263671875,
0.90869140625,
0.2139892578125,
0.88134765625,
0.2139892578125
],
[
0.11273193359375,
0.2183837890625,
0.404296875,
0.2183837890625,
0.404296875,
0.2291259765625,
0.11273193359375,
0.2291259765625
],
[
0.111083984375,
0.2020263671875,
0.490234375,
0.2020263671875,
0.490234375,
0.21533203125,
0.111083984375,
0.21533203125
],
[
0.52392578125,
0.2183837890625,
0.9189453125,
0.2183837890625,
0.9189453125,
0.231689453125,
0.52392578125,
0.231689453125
],
[
0.09478759765625,
0.2310791015625,
0.490234375,
0.231689453125,
0.490234375,
0.246826171875,
0.09478759765625,
0.24560546875
],
[
0.54736328125,
0.231689453125,
0.9208984375,
0.23291015625,
0.9208984375,
0.2474365234375,
0.54736328125,
0.246826171875
],
[
0.1094970703125,
0.246826171875,
0.490234375,
0.246826171875,
0.490234375,
0.261474609375,
0.1094970703125,
0.261474609375
],
[
0.54833984375,
0.2474365234375,
0.9208984375,
0.2474365234375,
0.9208984375,
0.261474609375,
0.54833984375,
0.261474609375
],
[
0.1094970703125,
0.261962890625,
0.490234375,
0.261474609375,
0.490234375,
0.27587890625,
0.1094970703125,
0.276611328125
],
[
0.54833984375,
0.261962890625,
0.9189453125,
0.261962890625,
0.9189453125,
0.276611328125,
0.54833984375,
0.276611328125
],
[
0.111083984375,
0.276611328125,
0.490234375,
0.276611328125,
0.490234375,
0.290283203125,
0.111083984375,
0.290283203125
],
[
0.54833984375,
0.27783203125,
0.9208984375,
0.27783203125,
0.9208984375,
0.291748046875,
0.54833984375,
0.291748046875
],
[
0.111083984375,
0.292236328125,
0.490234375,
0.292236328125,
0.490234375,
0.30615234375,
0.111083984375,
0.30615234375
],
[
0.54833984375,
0.29345703125,
0.8798828125,
0.29345703125,
0.8798828125,
0.306884765625,
0.54833984375,
0.306884765625
],
[
0.11273193359375,
0.309326171875,
0.488525390625,
0.309326171875,
0.488525390625,
0.3193359375,
0.11273193359375,
0.3193359375
],
[
0.52392578125,
0.30810546875,
0.9208984375,
0.30810546875,
0.9208984375,
0.3212890625,
0.52392578125,
0.3212890625
],
[
0.1094970703125,
0.322509765625,
0.490234375,
0.322509765625,
0.490234375,
0.3359375,
0.1094970703125,
0.3359375
],
[
0.54833984375,
0.323974609375,
0.9189453125,
0.323974609375,
0.9189453125,
0.337158203125,
0.54833984375,
0.337158203125
],
[
0.111083984375,
0.337158203125,
0.490234375,
0.337158203125,
0.490234375,
0.3515625,
0.111083984375,
0.3515625
],
[
0.54833984375,
0.339111328125,
0.9208984375,
0.339111328125,
0.9208984375,
0.352783203125,
0.54833984375,
0.352783203125
],
[
0.111083984375,
0.352783203125,
0.491943359375,
0.352783203125,
0.491943359375,
0.3662109375,
0.111083984375,
0.3662109375
],
[
0.54833984375,
0.354248046875,
0.9189453125,
0.354248046875,
0.9189453125,
0.367431640625,
0.54833984375,
0.367431640625
],
[
0.111083984375,
0.367431640625,
0.2279052734375,
0.367431640625,
0.2279052734375,
0.380615234375,
0.111083984375,
0.380615234375
],
[
0.54736328125,
0.36865234375,
0.9208984375,
0.36865234375,
0.9208984375,
0.3818359375,
0.54736328125,
0.3818359375
],
[
0.07843017578125,
0.3837890625,
0.490234375,
0.385009765625,
0.490234375,
0.399658203125,
0.07843017578125,
0.3984375
],
[
0.54833984375,
0.3837890625,
0.9189453125,
0.3837890625,
0.9189453125,
0.397705078125,
0.54833984375,
0.397705078125
],
[
0.08087158203125,
0.401611328125,
0.17724609375,
0.401611328125,
0.17724609375,
0.41357421875,
0.08087158203125,
0.41357421875
],
[
0.5498046875,
0.40087890625,
0.91845703125,
0.40087890625,
0.91845703125,
0.410888671875,
0.5498046875,
0.410888671875
],
[
0.0931396484375,
0.419189453125,
0.490234375,
0.419189453125,
0.490234375,
0.43310546875,
0.0931396484375,
0.43310546875
],
[
0.54833984375,
0.415283203125,
0.611328125,
0.415283203125,
0.611328125,
0.4267578125,
0.54833984375,
0.4267578125
],
[
0.111083984375,
0.434326171875,
0.49267578125,
0.434326171875,
0.49267578125,
0.448974609375,
0.111083984375,
0.448974609375
],
[
0.5107421875,
0.43115234375,
0.9189453125,
0.43115234375,
0.9189453125,
0.44580078125,
0.5107421875,
0.44580078125
],
[
0.111083984375,
0.449462890625,
0.26318359375,
0.449462890625,
0.26318359375,
0.46337890625,
0.111083984375,
0.46337890625
],
[
0.5107421875,
0.4462890625,
0.91650390625,
0.4462890625,
0.91650390625,
0.460205078125,
0.5107421875,
0.460205078125
],
[
0.09478759765625,
0.464111328125,
0.428955078125,
0.46337890625,
0.428955078125,
0.477783203125,
0.09478759765625,
0.479248046875
],
[
0.50732421875,
0.458984375,
0.57666015625,
0.4609375,
0.57666015625,
0.4765625,
0.50732421875,
0.474853515625
],
[
0.09478759765625,
0.479248046875,
0.491943359375,
0.479248046875,
0.491943359375,
0.49365234375,
0.09478759765625,
0.49365234375
],
[
0.50732421875,
0.48486328125,
0.77294921875,
0.48486328125,
0.77294921875,
0.5,
0.50732421875,
0.5
],
[
0.1094970703125,
0.494384765625,
0.491943359375,
0.494384765625,
0.491943359375,
0.50830078125,
0.1094970703125,
0.50830078125
],
[
0.525390625,
0.50390625,
0.91845703125,
0.50390625,
0.91845703125,
0.517578125,
0.525390625,
0.517578125
],
[
0.111083984375,
0.51025390625,
0.275390625,
0.51025390625,
0.275390625,
0.5234375,
0.111083984375,
0.5234375
],
[
0.09234619140625,
0.5234375,
0.491943359375,
0.5244140625,
0.491943359375,
0.5390625,
0.09234619140625,
0.53857421875
],
[
0.5087890625,
0.51953125,
0.91845703125,
0.51953125,
0.91845703125,
0.53271484375,
0.5087890625,
0.53271484375
],
[
0.108642578125,
0.5390625,
0.431396484375,
0.54052734375,
0.431396484375,
0.5537109375,
0.108642578125,
0.55322265625
],
[
0.5087890625,
0.5341796875,
0.9189453125,
0.5341796875,
0.9189453125,
0.54736328125,
0.5087890625,
0.54736328125
],
[
0.0931396484375,
0.5546875,
0.494384765625,
0.5546875,
0.494384765625,
0.568359375,
0.0931396484375,
0.568359375
],
[
0.50732421875,
0.5498046875,
0.9189453125,
0.5498046875,
0.9189453125,
0.56298828125,
0.50732421875,
0.56298828125
],
[
0.111083984375,
0.5712890625,
0.2998046875,
0.5712890625,
0.2998046875,
0.58203125,
0.111083984375,
0.58203125
],
[
0.51123046875,
0.56640625,
0.9189453125,
0.56640625,
0.9189453125,
0.5771484375,
0.51123046875,
0.5771484375
],
[
0.51123046875,
0.580078125,
0.9189453125,
0.580078125,
0.9189453125,
0.59326171875,
0.51123046875,
0.59326171875
],
[
0.09722900390625,
0.58837890625,
0.490234375,
0.58837890625,
0.490234375,
0.6015625,
0.09722900390625,
0.6015625
],
[
0.50732421875,
0.59228515625,
0.9189453125,
0.5947265625,
0.9189453125,
0.609375,
0.50732421875,
0.60693359375
],
[
0.080078125,
0.60302734375,
0.313720703125,
0.60302734375,
0.313720703125,
0.6162109375,
0.080078125,
0.6162109375
],
[
0.310546875,
0.60498046875,
0.488525390625,
0.60498046875,
0.488525390625,
0.61572265625,
0.310546875,
0.61572265625
],
[
0.5107421875,
0.60986328125,
0.9208984375,
0.60986328125,
0.9208984375,
0.62353515625,
0.5107421875,
0.62353515625
],
[
0.080078125,
0.61865234375,
0.491943359375,
0.61865234375,
0.491943359375,
0.6318359375,
0.080078125,
0.6318359375
],
[
0.51318359375,
0.625,
0.6748046875,
0.625,
0.6748046875,
0.63818359375,
0.51318359375,
0.63818359375
],
[
0.08251953125,
0.63330078125,
0.4453125,
0.63330078125,
0.4453125,
0.646484375,
0.08251953125,
0.646484375
],
[
0.52685546875,
0.64013671875,
0.9189453125,
0.64013671875,
0.9189453125,
0.65380859375,
0.52685546875,
0.65380859375
],
[
0.0767822265625,
0.65771484375,
0.336669921875,
0.6591796875,
0.336669921875,
0.673828125,
0.0767822265625,
0.67236328125
],
[
0.5107421875,
0.65478515625,
0.9208984375,
0.65478515625,
0.9208984375,
0.66845703125,
0.5107421875,
0.66845703125
],
[
0.5087890625,
0.66943359375,
0.9208984375,
0.66943359375,
0.9208984375,
0.68310546875,
0.5087890625,
0.68310546875
],
[
0.09478759765625,
0.6787109375,
0.490234375,
0.6787109375,
0.490234375,
0.6923828125,
0.09478759765625,
0.6923828125
],
[
0.5087890625,
0.68505859375,
0.9189453125,
0.68505859375,
0.9189453125,
0.69873046875,
0.5087890625,
0.69873046875
],
[
0.08087158203125,
0.69580078125,
0.488525390625,
0.69580078125,
0.488525390625,
0.70556640625,
0.08087158203125,
0.70556640625
],
[
0.50634765625,
0.69873046875,
0.9189453125,
0.69970703125,
0.9189453125,
0.7138671875,
0.50634765625,
0.71337890625
],
[
0.07843017578125,
0.708984375,
0.488525390625,
0.708984375,
0.488525390625,
0.72265625,
0.07843017578125,
0.72265625
],
[
0.50732421875,
0.71533203125,
0.87109375,
0.71533203125,
0.87109375,
0.72998046875,
0.50732421875,
0.72998046875
],
[
0.080078125,
0.7236328125,
0.488525390625,
0.7236328125,
0.488525390625,
0.73779296875,
0.080078125,
0.73779296875
],
[
0.080078125,
0.7392578125,
0.490234375,
0.7392578125,
0.490234375,
0.75244140625,
0.080078125,
0.75244140625
],
[
0.64208984375,
0.7373046875,
0.7861328125,
0.7392578125,
0.7861328125,
0.75244140625,
0.64208984375,
0.75048828125
],
[
0.080078125,
0.75390625,
0.490234375,
0.75244140625,
0.490234375,
0.767578125,
0.080078125,
0.76806640625
],
[
0.52392578125,
0.75830078125,
0.9189453125,
0.75830078125,
0.9189453125,
0.771484375,
0.52392578125,
0.771484375
],
[
0.080078125,
0.76953125,
0.490234375,
0.76953125,
0.490234375,
0.78271484375,
0.080078125,
0.78271484375
],
[
0.5107421875,
0.7734375,
0.9189453125,
0.7734375,
0.9189453125,
0.787109375,
0.5107421875,
0.787109375
],
[
0.080078125,
0.7841796875,
0.490234375,
0.78271484375,
0.490234375,
0.79736328125,
0.080078125,
0.798828125
],
[
0.5087890625,
0.78857421875,
0.9189453125,
0.78857421875,
0.9189453125,
0.8017578125,
0.5087890625,
0.8017578125
],
[
0.07843017578125,
0.798828125,
0.491943359375,
0.7998046875,
0.491943359375,
0.814453125,
0.07843017578125,
0.81298828125
],
[
0.50634765625,
0.8017578125,
0.9189453125,
0.8037109375,
0.9189453125,
0.8173828125,
0.50634765625,
0.81494140625
],
[
0.080078125,
0.81494140625,
0.490234375,
0.81494140625,
0.490234375,
0.8291015625,
0.080078125,
0.8291015625
],
[
0.5087890625,
0.818359375,
0.9189453125,
0.818359375,
0.9189453125,
0.83203125,
0.5087890625,
0.83203125
],
[
0.080078125,
0.83056640625,
0.10296630859375,
0.83056640625,
0.10296630859375,
0.8447265625,
0.080078125,
0.8447265625
],
[
0.5107421875,
0.83349609375,
0.9189453125,
0.83349609375,
0.9189453125,
0.8466796875,
0.5107421875,
0.8466796875
],
[
0.0931396484375,
0.84765625,
0.495849609375,
0.8466796875,
0.495849609375,
0.861328125,
0.0931396484375,
0.8623046875
],
[
0.5087890625,
0.8486328125,
0.9189453125,
0.8486328125,
0.9189453125,
0.8623046875,
0.5087890625,
0.8623046875
],
[
0.1192626953125,
0.86279296875,
0.495849609375,
0.86279296875,
0.495849609375,
0.87744140625,
0.1192626953125,
0.87744140625
],
[
0.5087890625,
0.86279296875,
0.9189453125,
0.86279296875,
0.9189453125,
0.876953125,
0.5087890625,
0.876953125
],
[
0.11767578125,
0.87744140625,
0.495849609375,
0.87744140625,
0.495849609375,
0.892578125,
0.11767578125,
0.892578125
],
[
0.50732421875,
0.87890625,
0.91845703125,
0.87890625,
0.91845703125,
0.892578125,
0.50732421875,
0.892578125
],
[
0.1168212890625,
0.89453125,
0.494384765625,
0.89306640625,
0.494384765625,
0.9072265625,
0.1168212890625,
0.90771484375
],
[
0.50634765625,
0.89306640625,
0.9189453125,
0.89306640625,
0.9189453125,
0.9072265625,
0.50634765625,
0.9072265625
]
] |
[
"a UCT formula:",
"the generation (see the table below), random choice",
" s(i)",
"of the first 15 moves (with probability proportional",
"a* = argmax(Q(a) + U(a)),",
"+",
"",
"Valiev",
"Q(a) :",
"",
"to the visits number), Dirichlet noise at the root.",
"The agent is allowed to pass only if there is no other",
"where S(i) is the scaled score difference estimated by",
"move with an expected score larger than the Tromp-",
"the value head for node i, Q(a) is the mean action",
"Taylor score.",
"generations",
"generations",
"w",
"value over the set V. of the visited MCTS tree nodes",
"",
"w",
"",
"U",
"100",
"16-31",
"150",
"0-15",
"that follow a , and U(a) is the same as in [i] , with",
"4",
"32-47",
"250",
"12",
"48-63",
"400",
"16",
"850",
"the Cpuct constant raised from 1 to 1.5 to account for",
"6479",
"600",
"20",
"80-00",
"24",
"withrespecttothewinrateprobabilities.",
"the larger variability of the scaled scores of siblings",
"2) Training. Starting from the current best network.",
". Game ending. LZ employs a set of heuristics during",
"new training is started over the self-play games of",
"self-plays to avoid uselessly long games. For example,",
"a window of the most recent w generations (see the",
"most games are allowed to end by resignation, when",
"table above). The number of steps adapts to the",
"the winrate drops below 5 %, and there are some",
"number of games inside the window, with a typical",
"hard-coded conditions for passing. These heuristics do",
"number of 30,000 steps. The batch size is 512.",
"notworkanymoreinournewsetting:tomaximizethe",
"3) Gating.The newly trained network is matched",
"score, the agent should keep playing to the very end,",
"against the current best, for a total of 400 games.",
"so that the computationally amenable Tromp-Taylor",
"with the same number of visits for self-play games",
"score becomes equivalent to the proper area score of",
"and no randomness in the move choice. The agent",
"the game of Go.",
"is allowed to pass when UCT selects this move. If",
"使用p<1将电路板的尺寸从N=19缩小到pN=9会",
"at least 55% of the 400 matches are won by the",
"带来一些好处",
"candidatenetwork.itispromotedtobethenewbest",
": Average number of legal moves at each position scales",
"network.",
"down by p? . The number of good, meaningful moves",
"所有游戏均使用komi7.5进行，也就是说，在白人玩家的",
"might scale similarly.",
"Tromp-Taylor得分上增加了7.5分，以补偿黑人扮演第一",
": Average length of games scales down by p?",
"步的事实",
": The suggested number of visits in the MCTS tree",
"C. Outcome of the learning pipeline",
"might scale down by as much as p4 , as can be inferred",
"学习管道由门匹配的中间结果引导。实际上，在LZ之",
"from the previous two.",
". The suggested number of layers in the residual con-",
"后，在每一代中，我们都从门控比赛中的400场比赛中估",
"volutional neural network stack scales by p.",
"算了新网络的Elo等级，将第一个随机网络任意锚定到",
". The fully connected layers at the end of the neural",
"Elo0。这个评级显然被高估了（例如，每个提升的网络都被",
"network stack are smaller.",
"限制为至少有34个。8分，一对应于一个55的%winrate-比",
"前一个更多），所以我们把它称为未校准的Elo。尽管如此",
"总结一下性能方面的好处，可以估计自玩游戏的总速度",
"该估计足以评估学习管道何时停止：当没有新网络克服太",
"提高大约为p9，即大约为10-3",
"。实际上，我们的资源使",
"多代（经验选择）的门控时，我们扩大了网络规模或降低",
"我们能够实施具有更多访问次数的管道。因此，我们选择",
"了训练优化器的学习率",
"了相同数量的9×9SAI访问，以便进行公平的比较",
"从随机的2×64网络（具有2个剩余的64个过滤器的",
"B. Reinforcement learning pipeline",
"卷积块）开始，我们在2,000场比赛中移至4×128；在",
"154,000场比赛中移至8×160；在748.000场比赛中移",
"我们的培训由多个阶段组成，灵感来自原始的LZ培训",
"至10×192；在982,000场比赛中移至12×256。学习",
"过程以及SAI有据可香的管道。特别是，我们遵循了与",
"率从0开始。02和减少一次到0。002在850,000场比赛",
"9×9ESAIE[7]的第一次运行儿乎相同的训练过程，唯一",
"中。Fig。5根据未校准的Elo等级显示管道结果",
"的例外是我们重新引人了门控（i。e。最好的网络只有",
"在更强大的网络出现时才会被替换)。从AlphaGoZero到",
"VI.CONCLUSIONS",
"AlphaZero的段落中删除了Gating，因为它被认为不是严",
"在本文中，我们讨论了在基于分数的赢/输游戏中最大",
"格必要的：由于我们正在尝试基于分数的强化学习，因此",
"化分数在胜率方面何时以及为什么可能是次优的。由于这",
"我们认为将其保留在管道中更为保守。实际上，当我们训",
"在确定性游戏中不可能发生，因此我们从最简单的非确定",
"练网络以最大化其得分时，通过门控，我们正在评估其赢",
"性案例开始，即多臂匪徒。我们证明，在多臂匪徒中，次",
"得比赛的能力。每个训练周期称为一代，它由以下步骤组",
"优行为与分数方差以及代理是赢还是输有关：当失败时，",
"成。",
"基于结果的代理更喜欢导致更大分数方差的行为，反之亦",
"1) Self-play. LzS engine uses the current best network",
"然。然后，我们定义了一类对基于得分的输赢游戏进行建",
"to play a total of 2,000 self-play games (against",
"模的MDP，并通过经验表明，次优游戏与得分方差之间",
"itself and sharing the tree search), with the following",
"的这种关系仍然有效。最后，我们训练了一个基于分数的",
"parameters:fixed number of visits depending on",
"AlphaGo类代理，并从经验上表明，尽管围棋游戏是确定"
] |
{
"arxiv_id": "2201.13176",
"language": "zh",
"page_idx": 4
}
|
||
2207.13117-zh-1
|
2207.13117-zh-1.jpg
|
Users and Contemporary SERPs: A (Re-)Investigation
Examining User Interactions and Experiences
David Maxwell
Delft University of Technology
Delft, The Netherlands d.m.maxwell@tudelft.nl
Nirmal Roy
Delft University of Technology
Delft, The Netherlands
n.roy@tudelft.nl
# $\mathbf{ABSTRACT}$
# 1 INTRODUCTION
The Search Engine Results Page (SERP) has evolved significantly over the last two decades, moving away from the simple ten blue links paradigm to considerably more complex presentations that contain results from multiple verticals and granularities of textual information. Prior works have investigated how user interactions on the SERP are influenced by the presence or absence of heterogeneous content (e.g., images, videos, or news content), the layout of the SERP ( list vs. grid layout), and task complexity . In this paper, we reproduce the user studies conducted in prior worksspecifically those of Arguello et al.[4] and Siu and Chaparro [31] -to explore to what extent the findings from research conducted five to ten years ago still hold today as the average web user has become accustomed to SERPs with ever-increasing presentational complexity. To this end, we designed and ran a user study with four different SERP interfaces: (i) a heterogeneous grid ; (ii) a heterogeneous list ; (iii) a simple grid ; and (iv) a simple list . We collected the interactions of 41 study participants over 12 search tasks for our analyses. We observed that SERP types and task complexity affect user interactions with search results. We also find evidence to support most (6 out of 8) observations from [4,31] indicating that user interactions with different interfaces and to solve tasks of different complexity have remained mostly similar over time.
在过去的二十年中，搜索引擎结果页面(SERP)已从十个蓝色链接范式发展为更为复杂的演示文稿，其中包含来自多个垂直领域和多个文本信息粒度的结果(片段，直接答案，实体卡等。)一全部在一页内交错。在 SERP 中合并异构内容已被证明可以改变用户与 Web 结果的交互方式[2,3,7,10,18,22,29,38] 。SERP 中显示内容的方式(以及位置)也会影响用户交互[4,30,32]。虽然当代的 WebSERP 保持了按相关性降序排列的项目列表的原始想法，但诸如网格布局之类的替代演示文稿-最近(再次) 由 You.com 推广-也已被探索$\boxed{\mathbb{E}}[14,21,31,39]$
此外，过去的研究$\overline{\mathbb{E}}[4,31,32,38]$ 表明，SERP 上的用户行为不仅取决于信息的呈现，还取决于手头的搜索任务。对于诸如“查找并访问 SIGIR2022 的主页”之类的导航任务，用户-在理想情况下-需要一次查询和一次单击。将此与信息任务进行对比，例如“SIGIR2022 场地附近的好餐馆”。这需要扫描多个结果，并可能导致进一步的查询重构，以了解有关特定建议的更多信息
随着商业网络搜索引擎 SERP 随着时间的推移而发展 (因此最终用户已经习惯了不同类型的 SERP), 我们在本文中探讨了5-10年前的用户研究结果在多大程度上仍然适用于今天。具体来说，我们将注意力集中在再现两个先前研究的实验设置上：$\boxed{\mathrm{E}}$Arguello et al.[4] (2012 年发布)以及$\boxed{\mathrm{E}}$Siu and Chaparro [31] (2014年发布) -这些都研究了 SERP 上的用户交互如何受到异构内容的存在或不存在，SERP 的布局 (列表与。网格) 和任务复杂性。受我们复制的两篇论文的启发，我们的研究受到以下研究问题的指导
RQ1 How does a user's interactions with a SERP differ when re-
$\csc$coNCEPTS
## $\cdot$Information systems$\to$Search interfaces;$•$Human-centered computing $\to$Empirical studies in HCI.
sults are presented in a list and grid layout?
RQ2 How does task complexity affect user interactions with a
$\mathbf{KEYWORDS}$
SERP?
Human Computer Interaction; Interactive Information Retrieval;
Search Interfaces; Search Tasks; Reproducibility
RQ3 What is the interplay between task complexity and SERP
layout on user interactions?
ACM Reference Format:
Nirmal Roy and David Maxwell. 2022. Users and Contemporary SERPs: A
RQ4 How do users perceive the different SERP layouts?
为此，我们对$n=41$参与者进行了用户研究，每个参与者
都被赋予了不同复杂性的 12 搜索任务 (范围从 Remember 到
(Re-)Investigation : Examining User Interactions and Experiences. In Proceedings of the 45th International ACM SIGIR Conference on Research and DevelopmentinInformation Retrieval (SIGIR'22), july 11-15, 2022, Madrid, $Spain.$ ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/3477495. 3531719
Analyse类型的搜索任务),以使用以下四个不同的 SERP 接口
之一进行求解：(i)$\operatorname{E\text{异构网格；(ii)}}\operatorname{E\text{异构列表；(ii}}$)$\operatorname{E\text{简单}}$
网格；(iv)。我们探讨了[4,31] 中关于用户及其与列表的交互
的以下八个观察结果是否与。网格布局——和异构 vs。简单的
结果-跨不同的任务复杂性今天举行
$\frac{\text{This research has been supported by NWO VIDI project SearchX (639.022.722) and}}$ NWO project Aspasia ( 015. 013. 027) .
O1 Users fxated significantly more on the grid layout SERP
compared to the list layout SERP for completing more com-
$\textcircled{6}$
This work is licensed under a Creative Commons Attribu-
plex tasks [31].
$(\mathbf{cc})$
tion International 4.0 License
O2 On the grid layout SERP, users fixated on search results
significantly more for completing more complex tasks com-
$BY$
${\overline{SIGIR~'22,~fuly~11-15,~2022,~Madrid,~Spain}}$
$\odot$2022 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8732-3/22/07. https://doi.org/10.1145/3477495.3531719
pared to simple tasks. A similar observation was found for
the list layout SERP [31] .
|
[
[
0.158447265625,
0.1029052734375,
0.8388671875,
0.10418701171875,
0.8388671875,
0.125,
0.158447265625,
0.12371826171875
],
[
0.31787109375,
0.1351318359375,
0.68310546875,
0.1351318359375,
0.68310546875,
0.1483154296875,
0.31787109375,
0.1483154296875
],
[
0.2484130859375,
0.1590576171875,
0.34716796875,
0.1610107421875,
0.34716796875,
0.1787109375,
0.2484130859375,
0.1767578125
],
[
0.6396484375,
0.1622314453125,
0.7646484375,
0.1622314453125,
0.7646484375,
0.1767578125,
0.6396484375,
0.1767578125
],
[
0.19287109375,
0.175537109375,
0.40283203125,
0.177978515625,
0.40283203125,
0.1956787109375,
0.19287109375,
0.1932373046875
],
[
0.59619140625,
0.177978515625,
0.8056640625,
0.179931640625,
0.8056640625,
0.1944580078125,
0.59619140625,
0.1925048828125
],
[
0.22216796875,
0.1956787109375,
0.374267578125,
0.1956787109375,
0.374267578125,
0.205810546875,
0.22216796875,
0.205810546875
],
[
0.625,
0.1944580078125,
0.77685546875,
0.1944580078125,
0.77685546875,
0.2083740234375,
0.625,
0.2083740234375
],
[
0.240234375,
0.208984375,
0.35693359375,
0.2083740234375,
0.35693359375,
0.222900390625,
0.240234375,
0.2235107421875
],
[
0.62109375,
0.208984375,
0.78369140625,
0.208984375,
0.78369140625,
0.2235107421875,
0.62109375,
0.2235107421875
],
[
0.0882568359375,
0.23486328125,
0.1829833984375,
0.23486328125,
0.1829833984375,
0.246826171875,
0.0882568359375,
0.246826171875
],
[
0.51953125,
0.2374267578125,
0.5439453125,
0.2374267578125,
0.5439453125,
0.24560546875,
0.51953125,
0.24560546875
],
[
0.54345703125,
0.238037109375,
0.54833984375,
0.238037109375,
0.54833984375,
0.24365234375,
0.54345703125,
0.24365234375
],
[
0.5439453125,
0.2342529296875,
0.69287109375,
0.2342529296875,
0.69287109375,
0.2474365234375,
0.5439453125,
0.2474365234375
],
[
0.0849609375,
0.251953125,
0.48046875,
0.251953125,
0.48046875,
0.26513671875,
0.0849609375,
0.26513671875
],
[
0.5185546875,
0.251953125,
0.91259765625,
0.251953125,
0.91259765625,
0.26513671875,
0.5185546875,
0.26513671875
],
[
0.0849609375,
0.266357421875,
0.48193359375,
0.266357421875,
0.48193359375,
0.27978515625,
0.0849609375,
0.27978515625
],
[
0.5185546875,
0.266357421875,
0.91259765625,
0.266357421875,
0.91259765625,
0.27978515625,
0.5185546875,
0.27978515625
],
[
0.0849609375,
0.27978515625,
0.48193359375,
0.27978515625,
0.48193359375,
0.29345703125,
0.0849609375,
0.29345703125
],
[
0.51953125,
0.27978515625,
0.91259765625,
0.27978515625,
0.91259765625,
0.29345703125,
0.51953125,
0.29345703125
],
[
0.0849609375,
0.294189453125,
0.48193359375,
0.294189453125,
0.48193359375,
0.30810546875,
0.0849609375,
0.30810546875
],
[
0.51953125,
0.29541015625,
0.9111328125,
0.29541015625,
0.9111328125,
0.30615234375,
0.51953125,
0.30615234375
],
[
0.0849609375,
0.30810546875,
0.48193359375,
0.30810546875,
0.48193359375,
0.3212890625,
0.0849609375,
0.3212890625
],
[
0.5185546875,
0.30810546875,
0.91259765625,
0.30810546875,
0.91259765625,
0.3212890625,
0.5185546875,
0.3212890625
],
[
0.08416748046875,
0.3212890625,
0.48193359375,
0.3212890625,
0.48193359375,
0.335205078125,
0.08416748046875,
0.335205078125
],
[
0.5185546875,
0.3212890625,
0.91259765625,
0.3212890625,
0.91259765625,
0.335205078125,
0.5185546875,
0.335205078125
],
[
0.0849609375,
0.335205078125,
0.48193359375,
0.335205078125,
0.48193359375,
0.348388671875,
0.0849609375,
0.348388671875
],
[
0.5185546875,
0.335205078125,
0.91259765625,
0.335205078125,
0.91259765625,
0.348388671875,
0.5185546875,
0.348388671875
],
[
0.0849609375,
0.348388671875,
0.48193359375,
0.348388671875,
0.48193359375,
0.3623046875,
0.0849609375,
0.3623046875
],
[
0.5185546875,
0.348388671875,
0.9140625,
0.348388671875,
0.9140625,
0.3623046875,
0.5185546875,
0.3623046875
],
[
0.08416748046875,
0.363037109375,
0.4794921875,
0.361083984375,
0.4794921875,
0.375,
0.08416748046875,
0.376953125
],
[
0.51953125,
0.3642578125,
0.80615234375,
0.3642578125,
0.80615234375,
0.375,
0.51953125,
0.375
],
[
0.0849609375,
0.376953125,
0.48046875,
0.376953125,
0.48046875,
0.39013671875,
0.0849609375,
0.39013671875
],
[
0.53369140625,
0.376953125,
0.91259765625,
0.376953125,
0.91259765625,
0.39013671875,
0.53369140625,
0.39013671875
],
[
0.08905029296875,
0.392578125,
0.48046875,
0.392578125,
0.48046875,
0.40283203125,
0.08905029296875,
0.40283203125
],
[
0.51953125,
0.39013671875,
0.91259765625,
0.39013671875,
0.91259765625,
0.404052734375,
0.51953125,
0.404052734375
],
[
0.0849609375,
0.40478515625,
0.48193359375,
0.40478515625,
0.48193359375,
0.41845703125,
0.0849609375,
0.41845703125
],
[
0.51708984375,
0.40283203125,
0.91259765625,
0.404052734375,
0.91259765625,
0.417236328125,
0.51708984375,
0.416748046875
],
[
0.0849609375,
0.41845703125,
0.48193359375,
0.41845703125,
0.48193359375,
0.431884765625,
0.0849609375,
0.431884765625
],
[
0.5185546875,
0.41845703125,
0.91259765625,
0.41845703125,
0.91259765625,
0.431884765625,
0.5185546875,
0.431884765625
],
[
0.08660888671875,
0.434326171875,
0.4794921875,
0.434326171875,
0.4794921875,
0.4443359375,
0.08660888671875,
0.4443359375
],
[
0.5185546875,
0.431884765625,
0.91259765625,
0.431884765625,
0.91259765625,
0.44580078125,
0.5185546875,
0.44580078125
],
[
0.08660888671875,
0.447509765625,
0.48193359375,
0.447509765625,
0.48193359375,
0.458251953125,
0.08660888671875,
0.458251953125
],
[
0.5185546875,
0.4462890625,
0.91259765625,
0.4462890625,
0.91259765625,
0.460205078125,
0.5185546875,
0.460205078125
],
[
0.0849609375,
0.460205078125,
0.48193359375,
0.460205078125,
0.48193359375,
0.473388671875,
0.0849609375,
0.473388671875
],
[
0.5185546875,
0.458984375,
0.56396484375,
0.458984375,
0.56396484375,
0.473388671875,
0.5185546875,
0.473388671875
],
[
0.08660888671875,
0.47607421875,
0.48046875,
0.47607421875,
0.48046875,
0.486083984375,
0.08660888671875,
0.486083984375
],
[
0.53173828125,
0.473388671875,
0.91259765625,
0.472900390625,
0.91259765625,
0.4873046875,
0.53173828125,
0.488037109375
],
[
0.08416748046875,
0.4873046875,
0.48046875,
0.488037109375,
0.48046875,
0.501953125,
0.08416748046875,
0.50048828125
],
[
0.5185546875,
0.4873046875,
0.9111328125,
0.4873046875,
0.9111328125,
0.50048828125,
0.5185546875,
0.50048828125
],
[
0.0849609375,
0.501953125,
0.48046875,
0.501953125,
0.48046875,
0.51513671875,
0.0849609375,
0.51513671875
],
[
0.5185546875,
0.50244140625,
0.9111328125,
0.50244140625,
0.9111328125,
0.51318359375,
0.5185546875,
0.51318359375
],
[
0.0849609375,
0.51513671875,
0.48193359375,
0.51513671875,
0.48193359375,
0.52880859375,
0.0849609375,
0.52880859375
],
[
0.51953125,
0.5166015625,
0.9111328125,
0.5166015625,
0.9111328125,
0.5263671875,
0.51953125,
0.5263671875
],
[
0.0849609375,
0.52880859375,
0.48193359375,
0.52880859375,
0.48193359375,
0.54248046875,
0.0849609375,
0.54248046875
],
[
0.521484375,
0.52880859375,
0.9111328125,
0.52978515625,
0.9111328125,
0.54150390625,
0.521484375,
0.54052734375
],
[
0.0849609375,
0.54443359375,
0.46728515625,
0.54443359375,
0.46728515625,
0.5546875,
0.0849609375,
0.5546875
],
[
0.5185546875,
0.54248046875,
0.91259765625,
0.54248046875,
0.91259765625,
0.55615234375,
0.5185546875,
0.55615234375
],
[
0.5185546875,
0.556640625,
0.9140625,
0.556640625,
0.9140625,
0.57080078125,
0.5185546875,
0.57080078125
],
[
0.0882568359375,
0.5712890625,
0.2197265625,
0.5712890625,
0.2197265625,
0.58203125,
0.0882568359375,
0.58203125
],
[
0.5185546875,
0.57080078125,
0.91259765625,
0.57080078125,
0.91259765625,
0.583984375,
0.5185546875,
0.583984375
],
[
0.0849609375,
0.58837890625,
0.49267578125,
0.58837890625,
0.49267578125,
0.6015625,
0.0849609375,
0.6015625
],
[
0.51953125,
0.5859375,
0.6650390625,
0.5859375,
0.6650390625,
0.5966796875,
0.51953125,
0.5966796875
],
[
0.0849609375,
0.60302734375,
0.3505859375,
0.60302734375,
0.3505859375,
0.6162109375,
0.0849609375,
0.6162109375
],
[
0.5185546875,
0.60107421875,
0.9140625,
0.60107421875,
0.9140625,
0.6142578125,
0.5185546875,
0.6142578125
],
[
0.55810546875,
0.6162109375,
0.8203125,
0.6162109375,
0.8203125,
0.626953125,
0.55810546875,
0.626953125
],
[
0.08660888671875,
0.6298828125,
0.188720703125,
0.6298828125,
0.188720703125,
0.6416015625,
0.08660888671875,
0.6416015625
],
[
0.51708984375,
0.626953125,
0.9140625,
0.62890625,
0.9140625,
0.64306640625,
0.51708984375,
0.6416015625
],
[
0.08416748046875,
0.64794921875,
0.48193359375,
0.64794921875,
0.48193359375,
0.6611328125,
0.08416748046875,
0.6611328125
],
[
0.55810546875,
0.64306640625,
0.6005859375,
0.64306640625,
0.6005859375,
0.65380859375,
0.55810546875,
0.65380859375
],
[
0.5185546875,
0.6572265625,
0.56201171875,
0.6572265625,
0.56201171875,
0.66845703125,
0.5185546875,
0.66845703125
],
[
0.5556640625,
0.65576171875,
0.91259765625,
0.65576171875,
0.91259765625,
0.66943359375,
0.5556640625,
0.66943359375
],
[
0.08416748046875,
0.662109375,
0.374267578125,
0.662109375,
0.374267578125,
0.67529296875,
0.08416748046875,
0.67529296875
],
[
0.55810546875,
0.671875,
0.72802734375,
0.671875,
0.72802734375,
0.681640625,
0.55810546875,
0.681640625
],
[
0.0849609375,
0.681640625,
0.2303466796875,
0.681640625,
0.2303466796875,
0.69580078125,
0.0849609375,
0.69580078125
],
[
0.5185546875,
0.6845703125,
0.86376953125,
0.6845703125,
0.86376953125,
0.69775390625,
0.5185546875,
0.69775390625
],
[
0.08660888671875,
0.6962890625,
0.48046875,
0.6962890625,
0.48046875,
0.70703125,
0.08660888671875,
0.70703125
],
[
0.53515625,
0.70068359375,
0.91259765625,
0.70068359375,
0.91259765625,
0.7138671875,
0.53515625,
0.7138671875
],
[
0.08660888671875,
0.70849609375,
0.4794921875,
0.70849609375,
0.4794921875,
0.71826171875,
0.08660888671875,
0.71826171875
],
[
0.5185546875,
0.7138671875,
0.91259765625,
0.7138671875,
0.91259765625,
0.72802734375,
0.5185546875,
0.72802734375
],
[
0.08416748046875,
0.72021484375,
0.48193359375,
0.7197265625,
0.48193359375,
0.73291015625,
0.08416748046875,
0.734375
],
[
0.08660888671875,
0.73193359375,
0.4794921875,
0.73193359375,
0.4794921875,
0.74560546875,
0.08660888671875,
0.74560546875
],
[
0.51708984375,
0.72802734375,
0.91259765625,
0.7265625,
0.91259765625,
0.74072265625,
0.51708984375,
0.7412109375
],
[
0.51953125,
0.7412109375,
0.91259765625,
0.7412109375,
0.91259765625,
0.7548828125,
0.51953125,
0.7548828125
],
[
0.08660888671875,
0.74755859375,
0.48046875,
0.74755859375,
0.48046875,
0.75830078125,
0.08660888671875,
0.75830078125
],
[
0.08660888671875,
0.76025390625,
0.131591796875,
0.76025390625,
0.131591796875,
0.77099609375,
0.08660888671875,
0.77099609375
],
[
0.51953125,
0.7548828125,
0.91259765625,
0.7548828125,
0.91259765625,
0.76806640625,
0.51953125,
0.76806640625
],
[
0.5185546875,
0.76953125,
0.77294921875,
0.76953125,
0.77294921875,
0.78271484375,
0.5185546875,
0.78271484375
],
[
0.7900390625,
0.76953125,
0.9140625,
0.76953125,
0.9140625,
0.78271484375,
0.7900390625,
0.78271484375
],
[
0.08660888671875,
0.7861328125,
0.48046875,
0.7861328125,
0.48046875,
0.796875,
0.08660888671875,
0.796875
],
[
0.51953125,
0.78515625,
0.744140625,
0.78515625,
0.744140625,
0.79541015625,
0.51953125,
0.79541015625
],
[
0.08416748046875,
0.79443359375,
0.256591796875,
0.79443359375,
0.256591796875,
0.80810546875,
0.08416748046875,
0.80810546875
],
[
0.529296875,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.529296875,
0.81298828125
],
[
0.5556640625,
0.81298828125,
0.9140625,
0.8125,
0.9140625,
0.8271484375,
0.5556640625,
0.82763671875
],
[
0.1422119140625,
0.82763671875,
0.1650390625,
0.82763671875,
0.1650390625,
0.84130859375,
0.1422119140625,
0.84130859375
],
[
0.2034912109375,
0.8291015625,
0.48193359375,
0.8291015625,
0.48193359375,
0.8388671875,
0.2034912109375,
0.8388671875
],
[
0.5595703125,
0.8291015625,
0.65185546875,
0.8291015625,
0.65185546875,
0.8388671875,
0.5595703125,
0.8388671875
],
[
0.09722900390625,
0.83544921875,
0.11273193359375,
0.83544921875,
0.11273193359375,
0.8388671875,
0.09722900390625,
0.8388671875
],
[
0.205078125,
0.8388671875,
0.3447265625,
0.8388671875,
0.3447265625,
0.849609375,
0.205078125,
0.849609375
],
[
0.53125,
0.84228515625,
0.56201171875,
0.84228515625,
0.56201171875,
0.85400390625,
0.53125,
0.85400390625
],
[
0.5556640625,
0.84130859375,
0.9140625,
0.84130859375,
0.9140625,
0.85498046875,
0.5556640625,
0.85498046875
],
[
0.08251953125,
0.8515625,
0.279296875,
0.8515625,
0.279296875,
0.86572265625,
0.08251953125,
0.86572265625
],
[
0.1446533203125,
0.8466796875,
0.16259765625,
0.8466796875,
0.16259765625,
0.85107421875,
0.1446533203125,
0.85107421875
],
[
0.5556640625,
0.85400390625,
0.9140625,
0.85498046875,
0.9140625,
0.869140625,
0.5556640625,
0.86865234375
],
[
0.08416748046875,
0.86279296875,
0.306396484375,
0.86279296875,
0.306396484375,
0.876953125,
0.08416748046875,
0.876953125
],
[
0.0849609375,
0.875,
0.25732421875,
0.875,
0.25732421875,
0.88525390625,
0.0849609375,
0.88525390625
],
[
0.5556640625,
0.869140625,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.5556640625,
0.88330078125
],
[
0.08416748046875,
0.8837890625,
0.275390625,
0.8837890625,
0.275390625,
0.89794921875,
0.08416748046875,
0.89794921875
],
[
0.55810546875,
0.88330078125,
0.71337890625,
0.88330078125,
0.71337890625,
0.896484375,
0.55810546875,
0.896484375
]
] |
[
"Users and Contemporary SERPs: A (Re-)Investigation",
"Examining UserInteractions and Experiences",
" Nirmal Roy",
"David Maxwell",
"Delft University of Technology",
"Delft University of Technology",
"Delft.TheNetherlands",
"Delft, The Netherlands",
"n.roy@tudelft.nl",
"d.m.maxwell@tudelft.nl",
"ABSTRACT",
"1",
"",
"INTRODUCTION",
"The Search Engine Results Page (SERP) has evolved significantly",
"在过去的二十年中，搜索引擎结果页面（SERP）已从十个蓝色",
"over the last two decades, moving away from the simple ten blue",
"链接范式发展为更为复杂的演示文稿，其中包含来自多个垂直",
"links paradigm to considerably more complex presentations that",
"领域和多个文本信息粒度的结果（片段，直接答案，实体卡等。",
"contain results from multiple verticals and granularities of textual",
")一全部在一页内交错。在SERP中合并异构内容已被证明可以",
"information. Prior works have investigated how user interactions",
"改变用户与Web结果的交互方式[2,3,7,10,18,22,29,38]。SERP",
"on the SERP are influenced by the presence or absence of hetero-",
"中显示内容的方式（以及位置）也会影响用户交互[4,30,32]。",
"geneous content (e.g., images, videos, or news content), the layout",
"虽然当代的WebSERP保持了按相关性降序排列的项目列表的",
"of the SERP (list vs. grid layout), and task complexity.In this",
"原始想法，但诸如网格布局之类的替代演示文稿-最近（再次）",
"paper, we reproduce the user studies conducted in prior works-",
"由You.com推广-也已被探索E[14,21,31,39]",
"specifically those of Arguello et al. [4] and Siu and Chaparro [3]]",
"此外，过去的研究E[4,31,32,38]表明，SERP上的用户行为",
"-toexploretowhatextentthefindingsfromresearchconducted",
"不仅取决于信息的呈现，还取决于手头的搜索任务。对于诸如",
"five to ten years ago still hold today as the average web user has",
"“查找并访问SIGIR2022的主页”之类的导航任务，用户一在理",
"become accustomed to SERPs with ever-increasing presentational",
"想情况下一需要一次查询和一次单击。将此与信息任务进行对",
"complexity.Tothis end, we designedandran a user studywithfour",
"比，例如“SIGIR2022场地附近的好餐馆”。这需要扫描多个结",
"different SERP interfaces:(i)a heterogeneousgrid:(i)aheteroge-",
"果，并可能导致进一步的查询重构，以了解有关特定建议的更",
"neous list : (ii) a simple grid ; and (iv) a simple list . We collected",
"多信息",
"the interactions of41 study participants over 12search tasksfor",
"随着商业网络搜索引擎SERP随着时间的推移而发展（因此",
"our analyses.Weobserved that SERP types and task complexity",
"最终用户已经习惯了不同类型的SERP），我们在本文中探讨了",
"affect user interactions with search results. We also find evidence",
"5-10年前的用户研究结果在多大程度上仍然适用于今天。具",
"to support most (6 out of 8) observations from [4, 31] indicating",
"体来说，我们将注意力集中在再现两个先前研究的实验设置",
"that user interactions with different interfaces and to solve tasks",
"上：EArguelloetal.[4]（2012年发布）以及ESiuandChaparro",
"of differentcomplexityhaveremainedmostlysimilar overtime",
"[31]（2014年发布）-这些都研究了SERP上的用户交互如何",
"受到异构内容的存在或不存在，SERP的布局（列表与。网格)",
"CCS CONCEPTS",
"和任务复杂性。受我们复制的两篇论文的启发，我们的研究受",
".Information systems- Search interfaces;Human-centered",
"到以下研究问题的指导",
"computing Empirical studies in HCI.",
"RQ1 How does a user's interactions with a SERP differ when re-",
"sults are presented in a list and grid layout?",
"KEYWORDS",
"RQ2 How does task complexity affect user interactions with a",
"Human Computer Interaction; Interactive Information Retrieval;",
"SERP?",
"RQ3",
"What is the interplay between task complexity and SERP",
"Search Interfaces; Search Tasks; Reproducibility",
"layout on userinteractions?",
"ACM Reference Format:",
"RQ4 How do users perceive the different SERP layouts?",
"Nirmal Roy and David Maxwell.2022.Users and Contemporary SERPs:A",
"为此，我们对n=41参与者进行了用户研究，每个参与者",
"(Re-)Investigation :Examining User Interactions and Experiences.In Pro",
"都被赋予了不同复杂性的12搜索任务（范围从Remember到",
"ceedings of the 45th International ACM SIGIR Conference on Research and",
"Development in Information Retrieval (SIGIR '22), July 11-15, 2022, Madrid",
"Analyse类型的搜索任务），以使用以下四个不同的SERP接口",
"之一进行求解：(i)E异构网格；(ii)E异构列表；(ii)国简单",
"Spain.ACM,NewYork,NY,USA,11pages.https:/doi.org/10.1145/3477495",
"3531719",
"网格；（iv)。我们探讨了[4,31]中关于用户及其与列表的交互",
"的以下八个观察结果是否与。网格布局一",
"和异构vs。简单的",
"This research has been supported by NWO VIDI project SearchX (639.022.722) and",
"结果-跨不同的任务复杂性今天举行",
"NWO project Aspasia (015.013.027).",
"O1 Users fixated significantly more on the grid layout SERP",
"compared to the list layout SERP for completing more com-",
"O",
"This work is licensed under a Creative Commons Attribu-",
"plex tasks[31]",
"",
"tion International 4.0 License.",
"02",
"On the grid layout SERP, users fixated on search results",
"SIGIR22,July115,222,Madrid, Spain",
"BY",
"significantly more for completing more complex tasks com-",
"@ 2022 Copyright held by the owner/author(s)",
"ACMISBN978-1-4503-8732-3/22/07.",
"pared to simple tasks.A similar observation was found for",
"https://doi.org/10.1145/3477495.3531719",
"the list layout SERP [31] ."
] |
{
"arxiv_id": "2207.13117",
"language": "zh",
"page_idx": 1
}
|
||
2207.13117-zh-2
|
2207.13117-zh-2.jpg
|
O3 On the list layout SERP, users fixated significantly longer for
completing more complex tasks compared to simple tasks. For the grid SERP, there were no significant differences in fixation duration between varying task complexities [31] .
$\mathbf{O}4$ In the list layout SERP, more complex tasks required signifi-
cantly greater levels of search interaction: longer search sessions, more clicks on SERP, and more web pages visited [4]
05 In a SERP where web results are arranged in a list layout,
users clicked on significantly more vertical results when they were present on the main page of the SERP (blended, heterogeneous display) compared to when they were only present as tabs (non-blended, simple display) [4]
O6 Task complexity did not have a significant effect on user
interaction with vertical results in the list layout SERP [4]. O7 The interplay between task complexity and display of verti-
cals (blended, heterogeneous display vs. non-blended, simple display) did not have a significant effect on user interaction with vertical results in the list layout SERP [4] .
O8 Neither study [4,31] found significant differences in user
evaluation of the different SERP types, list vs grid layout for the former and blended vs non-blended display for the latter, in their experiments.
在我们的用户研究中，我们观察到 SERP 类型和任务复杂性会影响用户与搜索结果的交互。我们还发现了支持$\mathbb{E}[4,31]$ 的 8个观测值中的大多数-6 个的证据
# 2 RELATED WORK
2.1 Task Complexity and User Interactions
许多工作都集中在任务类型对 SERP 上用户交互的影响上。Buscher
et al.[8]使用查询日志进行了大规模分析，以了解个人和任务差异如何影响搜索行为。他们的发现表明，有一群用户以类似的方式检查搜索结果。他们还表明，任务类型对用户与 SERP 的互动方式有显着影响。Arguello et al.[4]观察到任务越复杂， 与 SERP 上的各种组件交互的用户就越多，而 Thomas et al.[34] 发现用户在面对复杂任务时倾向于更深入、更快速地检查结果列表。Jiang et al.[12] 比较了四种不同类型的搜索任务在相对较长的搜索会话(10 分钟；关于5 查询) 中的用户交互。$\mathbb{E}$ Wu et al.[38] 还观察到用户与 SERP 的交互差异，这取决于他们是否必须寻找事实问题或非事实问题的答案。在这些研究中，SERP 由事实上的列表格式的网络搜索结果组成
2.2 $\textbf{SERP Presentation and User Interactions}$
Sushmita et al.[32] 观察到，当垂直 (新闻，图像和视频)以混合方式呈现时，SERP 上不同垂直的位置(顶部，中间，底部) 会影响用户的点击率。Arguello et al.[4] 还研究了任务复杂性如何影响用户交互和聚合垂直结果的使用，当它们与 Web 结果交错时，与它们作为选项卡呈现时相比。类似地，他们观察到，对于更复杂的任务，用户在与网络搜索结果交错时点击了更垂直的结果。Bota et al.[7] 进行了一项众包在线用户研究，以调查给定模糊搜索主题的实体卡的影响。他们发现，实体卡的存在对用户与搜索结果的交互方式及其感知的任务工作量都有很大影响。此外$,\boxed{\mathrm{ELevi~et~al.[16]}}$根据商业网络搜索引擎的日志，对来自七个不同垂直领域(包括垂直社区问题回答)的结果的呈现进行了全面分析。他们观察到，与其他垂直领域相比，社区问题回答垂直领域平均获得的点击次数最高。
Wu et al.[38] 研究了 SERP 上答案模块的存在如何影响用户行为，以及这是否随问题类型而变化 (factoidvs。non-factoid)。他们发现答案模块可以帮助用户更快地完成搜索任务，并减少用户的工作量。在存在答案模块的情况下，用户在回答事实问题时对网络搜索结果的点击显着减少 Shao et al.[30] 进行了一项用户研究，以了解异构 SERP 的右轨道中存在的结果以及左轨道中的传统 Web 结果如何影响用户交互。他们发现，用户与 SERP 有更多的互动，似乎更挣扎，如果他们检查右轨结果， 就会感到不太满意。总体而言，研究结果观察到垂直和其他异构形式的结果及其在 SERP 上的位置会影响用户交互。在这些研究中，结果也以事实上的列表格式呈现
Kammerer and Gerjets [14] 观察到，当网络搜索结果以网格布局呈现时，与更传统的列表方法相比，搜索结果定位对选择可信来源的影响大大降低。用户在扫描列表时通常遵循自顶向下的方法，如果列表中的位置较高，则更容易选择不可信的来源。对于基于网格的演示文稿，此效果会降低。但是，作者没有比较不同类型的任务，也没有探索当搜索引擎的各种垂直特征的结果出现在 SERP 上时的用户行为。Siu and Chaparro [31]将网格和列表 SERP 布局的眼动追踪数据与两种类型的任务进行了比较(信息性与。导航),并调查了凝视模式的潜在差异。与列表布局相比，“F形”图案在网格上不太突出。这两项研究探讨了当网络结果以网格形式呈现时，用户交互是如何变化的。列表布局。但是，他们在研究中不包括垂直结果
???
# 3 метнорособт
条件之一 (界面：受试者之间),并完成了 12 个搜索任务 (任
务：受试者内部)
为了解决 §1 中概述的四个总体研究问题，我们对$n=41$参与者进行了用户研究。每个参与者被分配到四个实验性搜索界面我们的四个实验性搜索接口条件考虑了布局类型(list-vs。基于网格的布局) 和 SERP 上存在的垂直行业 (异构内容与。均匀含量)。这些组合导致下面概述的接口条件，图1中提供了两种布局类型的示例，§3.1 中提供了更多详细信息
SL Simple List Considered as our baseline interface condition
(the standard and widely used ten blue links [11]), this interface presents results in a list, with each result presented one under the other. All results are web results , and as such are homogeneous in terms of presented content.
SG Simple Grid The same homogeneous approach to content is
taken as for SL, but with results presented in a grid-based approach. Instead of scrolling along the vertical, participants subjected to this interface scroll along the horizontal
HL Heterogeneous List Similar in approach to SL, HL presents
results in a list. However, different verticals are mixed in with the standard web results. Beyond web results, heterogeneous content used in this study includes image and video results.
HG Heterogeneous Grid Similar to HL but now the content is
displayed in grid form, with web-based results appearing in
a grid, before additional image and video content.
3.1 Search Interface Design and System
鉴于上述情况，我们的目标是找出$\overline{\mathrm{E}}$Arguello et al.[4] 和$\overline{\mathrm{FSiu}}$ and Chaparro [31] 先前研究的观察结果在经过近十年的 SERP 设计演变(和添加)后在多大程度上有效。为了实现我们的四
|
[
[
0.098876953125,
0.10797119140625,
0.48046875,
0.10797119140625,
0.48046875,
0.121826171875,
0.098876953125,
0.121826171875
],
[
0.51708984375,
0.10797119140625,
0.91259765625,
0.10797119140625,
0.91259765625,
0.121826171875,
0.51708984375,
0.121826171875
],
[
0.1258544921875,
0.12249755859375,
0.48046875,
0.12249755859375,
0.48046875,
0.1363525390625,
0.1258544921875,
0.1363525390625
],
[
0.51953125,
0.121826171875,
0.91259765625,
0.121826171875,
0.91259765625,
0.1351318359375,
0.51953125,
0.1351318359375
],
[
0.1258544921875,
0.1363525390625,
0.48193359375,
0.1363525390625,
0.48193359375,
0.149658203125,
0.1258544921875,
0.149658203125
],
[
0.5185546875,
0.1351318359375,
0.91259765625,
0.1351318359375,
0.91259765625,
0.1483154296875,
0.5185546875,
0.1483154296875
],
[
0.125,
0.1483154296875,
0.473876953125,
0.149658203125,
0.473876953125,
0.1634521484375,
0.125,
0.1622314453125
],
[
0.521484375,
0.1514892578125,
0.9111328125,
0.1514892578125,
0.9111328125,
0.1622314453125,
0.521484375,
0.1622314453125
],
[
0.09722900390625,
0.1641845703125,
0.483642578125,
0.1641845703125,
0.483642578125,
0.177978515625,
0.09722900390625,
0.177978515625
],
[
0.51953125,
0.1654052734375,
0.90283203125,
0.1654052734375,
0.90283203125,
0.175537109375,
0.51953125,
0.175537109375
],
[
0.1258544921875,
0.177978515625,
0.48193359375,
0.177978515625,
0.48193359375,
0.1912841796875,
0.1258544921875,
0.1912841796875
],
[
0.51953125,
0.1787109375,
0.9111328125,
0.1787109375,
0.9111328125,
0.189453125,
0.51953125,
0.189453125
],
[
0.125,
0.1912841796875,
0.48046875,
0.1900634765625,
0.48046875,
0.2052001953125,
0.125,
0.205810546875
],
[
0.5185546875,
0.1912841796875,
0.91259765625,
0.1912841796875,
0.91259765625,
0.2052001953125,
0.5185546875,
0.2052001953125
],
[
0.51953125,
0.2052001953125,
0.91259765625,
0.2052001953125,
0.91259765625,
0.2183837890625,
0.51953125,
0.2183837890625
],
[
0.096435546875,
0.2183837890625,
0.48046875,
0.2197265625,
0.48046875,
0.23291015625,
0.096435546875,
0.231689453125
],
[
0.51953125,
0.2203369140625,
0.9111328125,
0.2203369140625,
0.9111328125,
0.2310791015625,
0.51953125,
0.2310791015625
],
[
0.12744140625,
0.23486328125,
0.47802734375,
0.23486328125,
0.47802734375,
0.24560546875,
0.12744140625,
0.24560546875
],
[
0.5185546875,
0.23291015625,
0.91259765625,
0.23291015625,
0.91259765625,
0.246826171875,
0.5185546875,
0.246826171875
],
[
0.1258544921875,
0.246826171875,
0.48046875,
0.246826171875,
0.48046875,
0.260009765625,
0.1258544921875,
0.260009765625
],
[
0.5185546875,
0.246826171875,
0.78369140625,
0.246826171875,
0.78369140625,
0.260009765625,
0.5185546875,
0.260009765625
],
[
0.12744140625,
0.261962890625,
0.4794921875,
0.261962890625,
0.4794921875,
0.272705078125,
0.12744140625,
0.272705078125
],
[
0.53369140625,
0.260009765625,
0.91259765625,
0.260009765625,
0.91259765625,
0.2734375,
0.53369140625,
0.2734375
],
[
0.1258544921875,
0.274658203125,
0.4248046875,
0.274658203125,
0.4248046875,
0.28857421875,
0.1258544921875,
0.28857421875
],
[
0.51953125,
0.27587890625,
0.9111328125,
0.27587890625,
0.9111328125,
0.285888671875,
0.51953125,
0.285888671875
],
[
0.09722900390625,
0.28857421875,
0.48046875,
0.28857421875,
0.48046875,
0.3017578125,
0.09722900390625,
0.3017578125
],
[
0.51953125,
0.28857421875,
0.9140625,
0.28857421875,
0.9140625,
0.3017578125,
0.51953125,
0.3017578125
],
[
0.1258544921875,
0.3017578125,
0.4794921875,
0.3017578125,
0.4794921875,
0.31494140625,
0.1258544921875,
0.31494140625
],
[
0.5185546875,
0.3017578125,
0.91259765625,
0.3017578125,
0.91259765625,
0.31494140625,
0.5185546875,
0.31494140625
],
[
0.12091064453125,
0.31494140625,
0.48046875,
0.31494140625,
0.48046875,
0.328857421875,
0.12091064453125,
0.328857421875
],
[
0.5185546875,
0.31640625,
0.91259765625,
0.31640625,
0.91259765625,
0.330078125,
0.5185546875,
0.330078125
],
[
0.1258544921875,
0.330078125,
0.483642578125,
0.330078125,
0.483642578125,
0.343505859375,
0.1258544921875,
0.343505859375
],
[
0.1004638671875,
0.318115234375,
0.1258544921875,
0.318115234375,
0.1258544921875,
0.326904296875,
0.1004638671875,
0.326904296875
],
[
0.5185546875,
0.330078125,
0.91259765625,
0.330078125,
0.91259765625,
0.343505859375,
0.5185546875,
0.343505859375
],
[
0.1258544921875,
0.343505859375,
0.48193359375,
0.343505859375,
0.48193359375,
0.356689453125,
0.1258544921875,
0.356689453125
],
[
0.51953125,
0.343505859375,
0.91259765625,
0.343505859375,
0.91259765625,
0.356689453125,
0.51953125,
0.356689453125
],
[
0.12744140625,
0.359130859375,
0.439453125,
0.359130859375,
0.439453125,
0.369384765625,
0.12744140625,
0.369384765625
],
[
0.521484375,
0.359130859375,
0.9111328125,
0.359130859375,
0.9111328125,
0.369384765625,
0.521484375,
0.369384765625
],
[
0.09722900390625,
0.371826171875,
0.48193359375,
0.371826171875,
0.48193359375,
0.385009765625,
0.09722900390625,
0.385009765625
],
[
0.51953125,
0.37060546875,
0.9140625,
0.37060546875,
0.9140625,
0.3837890625,
0.51953125,
0.3837890625
],
[
0.1258544921875,
0.3837890625,
0.48046875,
0.3837890625,
0.48046875,
0.397705078125,
0.1258544921875,
0.397705078125
],
[
0.51708984375,
0.3837890625,
0.91259765625,
0.385009765625,
0.91259765625,
0.3984375,
0.51708984375,
0.397705078125
],
[
0.1258544921875,
0.399658203125,
0.48046875,
0.399658203125,
0.48046875,
0.41357421875,
0.1258544921875,
0.41357421875
],
[
0.5185546875,
0.399658203125,
0.91259765625,
0.399658203125,
0.91259765625,
0.41357421875,
0.5185546875,
0.41357421875
],
[
0.12744140625,
0.4140625,
0.289306640625,
0.4140625,
0.289306640625,
0.4248046875,
0.12744140625,
0.4248046875
],
[
0.51953125,
0.4140625,
0.8857421875,
0.4140625,
0.8857421875,
0.4248046875,
0.51953125,
0.4248046875
],
[
0.0882568359375,
0.429931640625,
0.47802734375,
0.429931640625,
0.47802734375,
0.43994140625,
0.0882568359375,
0.43994140625
],
[
0.53369140625,
0.4267578125,
0.5556640625,
0.4267578125,
0.5556640625,
0.439453125,
0.53369140625,
0.439453125
],
[
0.08660888671875,
0.442626953125,
0.48193359375,
0.442626953125,
0.48193359375,
0.455810546875,
0.08660888671875,
0.455810546875
],
[
0.0882568359375,
0.458251953125,
0.293212890625,
0.458251953125,
0.293212890625,
0.468505859375,
0.0882568359375,
0.468505859375
],
[
0.51953125,
0.458984375,
0.53125,
0.458984375,
0.53125,
0.468505859375,
0.51953125,
0.468505859375
],
[
0.5361328125,
0.4609375,
0.54345703125,
0.4609375,
0.54345703125,
0.46533203125,
0.5361328125,
0.46533203125
],
[
0.54833984375,
0.458251953125,
0.68896484375,
0.458251953125,
0.68896484375,
0.468505859375,
0.54833984375,
0.468505859375
],
[
0.51953125,
0.474853515625,
0.91259765625,
0.474853515625,
0.91259765625,
0.488037109375,
0.51953125,
0.488037109375
],
[
0.0849609375,
0.484130859375,
0.1168212890625,
0.484130859375,
0.1168212890625,
0.49560546875,
0.0849609375,
0.49560546875
],
[
0.1094970703125,
0.48291015625,
0.26318359375,
0.48291015625,
0.26318359375,
0.496826171875,
0.1094970703125,
0.496826171875
],
[
0.51708984375,
0.4873046875,
0.91259765625,
0.488037109375,
0.91259765625,
0.50244140625,
0.51708984375,
0.501953125
],
[
0.0849609375,
0.50244140625,
0.449462890625,
0.50244140625,
0.449462890625,
0.5166015625,
0.0849609375,
0.5166015625
],
[
0.5185546875,
0.50244140625,
0.9140625,
0.50244140625,
0.9140625,
0.5166015625,
0.5185546875,
0.5166015625
],
[
0.0882568359375,
0.52099609375,
0.50341796875,
0.52099609375,
0.50341796875,
0.5341796875,
0.0882568359375,
0.5341796875
],
[
0.51708984375,
0.5166015625,
0.62841796875,
0.5166015625,
0.62841796875,
0.52978515625,
0.51708984375,
0.52978515625
],
[
0.53369140625,
0.52978515625,
0.91259765625,
0.52978515625,
0.91259765625,
0.54345703125,
0.53369140625,
0.54345703125
],
[
0.0882568359375,
0.5361328125,
0.47802734375,
0.5361328125,
0.47802734375,
0.546875,
0.0882568359375,
0.546875
],
[
0.08660888671875,
0.54736328125,
0.48046875,
0.54736328125,
0.48046875,
0.56103515625,
0.08660888671875,
0.56103515625
],
[
0.51708984375,
0.54345703125,
0.91259765625,
0.54443359375,
0.91259765625,
0.55810546875,
0.51708984375,
0.556640625
],
[
0.0882568359375,
0.56298828125,
0.4794921875,
0.56298828125,
0.4794921875,
0.57373046875,
0.0882568359375,
0.57373046875
],
[
0.51953125,
0.56005859375,
0.91259765625,
0.56005859375,
0.91259765625,
0.57080078125,
0.51953125,
0.57080078125
],
[
0.08660888671875,
0.57568359375,
0.4794921875,
0.57568359375,
0.4794921875,
0.5888671875,
0.08660888671875,
0.5888671875
],
[
0.5185546875,
0.57275390625,
0.826171875,
0.57275390625,
0.826171875,
0.5859375,
0.5185546875,
0.5859375
],
[
0.08660888671875,
0.58837890625,
0.48046875,
0.5888671875,
0.48046875,
0.60302734375,
0.08660888671875,
0.6015625
],
[
0.52685546875,
0.59130859375,
0.9111328125,
0.59228515625,
0.9111328125,
0.60595703125,
0.52685546875,
0.60498046875
],
[
0.0882568359375,
0.60400390625,
0.48046875,
0.60400390625,
0.48046875,
0.6171875,
0.0882568359375,
0.6171875
],
[
0.55126953125,
0.60791015625,
0.9111328125,
0.60791015625,
0.9111328125,
0.61865234375,
0.55126953125,
0.61865234375
],
[
0.08660888671875,
0.6171875,
0.48046875,
0.6171875,
0.48046875,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.55126953125,
0.62060546875,
0.9140625,
0.62060546875,
0.9140625,
0.6337890625,
0.55126953125,
0.6337890625
],
[
0.0849609375,
0.630859375,
0.48193359375,
0.6298828125,
0.48193359375,
0.64306640625,
0.0849609375,
0.64453125
],
[
0.5498046875,
0.63330078125,
0.91259765625,
0.63330078125,
0.91259765625,
0.646484375,
0.5498046875,
0.646484375
],
[
0.0882568359375,
0.646484375,
0.4794921875,
0.646484375,
0.4794921875,
0.6572265625,
0.0882568359375,
0.6572265625
],
[
0.5498046875,
0.6484375,
0.8408203125,
0.6484375,
0.8408203125,
0.662109375,
0.5498046875,
0.662109375
],
[
0.08660888671875,
0.6591796875,
0.48193359375,
0.6591796875,
0.48193359375,
0.67236328125,
0.08660888671875,
0.67236328125
],
[
0.52392578125,
0.66015625,
0.91259765625,
0.6611328125,
0.91259765625,
0.67529296875,
0.52392578125,
0.6748046875
],
[
0.0882568359375,
0.6748046875,
0.4150390625,
0.6748046875,
0.4150390625,
0.68505859375,
0.0882568359375,
0.68505859375
],
[
0.54833984375,
0.67529296875,
0.91259765625,
0.67529296875,
0.91259765625,
0.689453125,
0.54833984375,
0.689453125
],
[
0.54833984375,
0.68994140625,
0.91259765625,
0.689453125,
0.91259765625,
0.70263671875,
0.54833984375,
0.7041015625
],
[
0.08660888671875,
0.69970703125,
0.13232421875,
0.69970703125,
0.13232421875,
0.7109375,
0.08660888671875,
0.7109375
],
[
0.1258544921875,
0.69873046875,
0.461669921875,
0.69873046875,
0.461669921875,
0.7119140625,
0.1258544921875,
0.7119140625
],
[
0.5498046875,
0.7041015625,
0.87158203125,
0.7041015625,
0.87158203125,
0.71728515625,
0.5498046875,
0.71728515625
],
[
0.0849609375,
0.71630859375,
0.48046875,
0.71630859375,
0.48046875,
0.72998046875,
0.0849609375,
0.72998046875
],
[
0.521484375,
0.71630859375,
0.91259765625,
0.71728515625,
0.91259765625,
0.73095703125,
0.521484375,
0.72998046875
],
[
0.08660888671875,
0.73095703125,
0.48193359375,
0.73095703125,
0.48193359375,
0.744140625,
0.08660888671875,
0.744140625
],
[
0.55126953125,
0.73291015625,
0.91259765625,
0.73291015625,
0.91259765625,
0.74365234375,
0.55126953125,
0.74365234375
],
[
0.0882568359375,
0.7470703125,
0.4794921875,
0.7470703125,
0.4794921875,
0.7568359375,
0.0882568359375,
0.7568359375
],
[
0.5498046875,
0.744140625,
0.9111328125,
0.74560546875,
0.9111328125,
0.7587890625,
0.5498046875,
0.75830078125
],
[
0.0882568359375,
0.76025390625,
0.48046875,
0.76025390625,
0.48046875,
0.77099609375,
0.0882568359375,
0.77099609375
],
[
0.5498046875,
0.7587890625,
0.90869140625,
0.7587890625,
0.90869140625,
0.77294921875,
0.5498046875,
0.77294921875
],
[
0.0882568359375,
0.7734375,
0.48046875,
0.7734375,
0.48046875,
0.7841796875,
0.0882568359375,
0.7841796875
],
[
0.51953125,
0.77099609375,
0.91259765625,
0.771484375,
0.91259765625,
0.7861328125,
0.51953125,
0.78515625
],
[
0.08660888671875,
0.7861328125,
0.48046875,
0.7861328125,
0.48046875,
0.7998046875,
0.08660888671875,
0.7998046875
],
[
0.55126953125,
0.7861328125,
0.9140625,
0.7861328125,
0.9140625,
0.7998046875,
0.55126953125,
0.7998046875
],
[
0.0882568359375,
0.8017578125,
0.4794921875,
0.8017578125,
0.4794921875,
0.8125,
0.0882568359375,
0.8125
],
[
0.5498046875,
0.80029296875,
0.8486328125,
0.80029296875,
0.8486328125,
0.814453125,
0.5498046875,
0.814453125
],
[
0.0882568359375,
0.81494140625,
0.48046875,
0.81494140625,
0.48046875,
0.82568359375,
0.0882568359375,
0.82568359375
],
[
0.0882568359375,
0.8291015625,
0.4794921875,
0.8291015625,
0.4794921875,
0.8388671875,
0.0882568359375,
0.8388671875
],
[
0.51953125,
0.83837890625,
0.56201171875,
0.83837890625,
0.56201171875,
0.8486328125,
0.51953125,
0.8486328125
],
[
0.560546875,
0.8369140625,
0.85107421875,
0.8369140625,
0.85107421875,
0.85107421875,
0.560546875,
0.85107421875
],
[
0.0882568359375,
0.84326171875,
0.4794921875,
0.84326171875,
0.4794921875,
0.85400390625,
0.0882568359375,
0.85400390625
],
[
0.0882568359375,
0.85693359375,
0.4794921875,
0.85693359375,
0.4794921875,
0.8671875,
0.0882568359375,
0.8671875
],
[
0.51953125,
0.85693359375,
0.9111328125,
0.85693359375,
0.9111328125,
0.8671875,
0.51953125,
0.8671875
],
[
0.08660888671875,
0.869140625,
0.48046875,
0.869140625,
0.48046875,
0.88330078125,
0.08660888671875,
0.88330078125
],
[
0.5185546875,
0.869140625,
0.91259765625,
0.869140625,
0.91259765625,
0.88330078125,
0.5185546875,
0.88330078125
],
[
0.0882568359375,
0.88525390625,
0.4794921875,
0.88525390625,
0.4794921875,
0.89599609375,
0.0882568359375,
0.89599609375
],
[
0.51953125,
0.88525390625,
0.9111328125,
0.88525390625,
0.9111328125,
0.89599609375,
0.51953125,
0.89599609375
]
] |
[
"O3On the list layout SERP,users fixated significantlylongerfor",
"Wuetal.[38]研究了SERP上答案模块的存在如何影响用户行",
"completing more complex tasks compared to simple tasks.",
"为，以及这是否随问题类型而变化(factoidvs。non-factoid)。他",
"For the grid SERP, there were no significant differences in",
"们发现答案模块可以帮助用户更快地完成搜索任务，并减少",
"fixation duration between varying task complexities [31]",
"用户的工作量。在存在答案模块的情况下，用户在回答事实问",
"O4 In the list layout SERP, more complex tasks required signifi-",
"题时对网络搜索结果的点击显着减少Shaoetal.[30]进行了-",
"cantly greater levels of search interaction: longer search ses",
"项用户研究，以了解异构SERP的右轨道中存在的结果以及左",
"sions, more clicks on SERP, and more web pages visited [4]",
"轨道中的传统Web结果如何影响用户交互。他们发现，用户",
"与SERP有更多的互动，似乎更挣扎，如果他们检查右轨结果，",
"O5 In a SERP where web results are arranged in a list layout,",
"就会感到不太满意。总体而言，研究结果观察到垂直和其他异",
"usersclicked on significantly more vertical resultswher",
"构形式的结果及其在SERP上的位置会影响用户交互。在这些",
"they were present on the main page of the SERP (blended,",
"研究中，结果也以事实上的列表格式呈现",
"heterogeneous display)compared to whenthey were only",
"KammererandGerjets[14]观察到，当网络搜索结果以网格",
"present as tabs (non-blended, simple display) [4] .",
"布局呈现时，与更传统的列表方法相比，搜索结果定位对选择",
"O6 Task complexity did not have a significant effect on user",
"可信来源的影响大大降低。用户在扫描列表时通常遵循自顶",
"interaction with vertical results in the list layout SERP [4] .",
"向下的方法，如果列表中的位置较高，则更容易选择不可信的",
"The interplay between task complexity and display of verti",
"来源。对于基于网格的演示文稿，此效果会降低。但是，作者",
"cals (blended, heterogeneous display vs. non-blended, sim-",
"07",
"没有比较不同类型的任务，也没有探索当搜索引擎的各种垂",
"ple display) did not have a significant effect on user interac-",
"直特征的结果出现在SERP上时的用户行为。SiuandChaparro",
"tionwithvertical resultsinthelistlayoutSERP[4].",
"[31]将网格和列表SERP布局的眼动追踪数据与两种类型的任",
"O8 Neither study [4, 31] found significant differences in user",
"务进行了比较（信息性与。导航)，并调查了凝视模式的潜在差",
"evaluation of the different SERP types, list vs grid layout",
"异。与列表布局相比，“F形”图案在网格上不太突出。这两项",
"for the former and blended vs non-blended display for the",
"研究探讨了当网络结果以网格形式呈现时，用户交互是如何",
"latter,intheirexperiments.",
"变化的。列表布局。但是，他们在研究中不包括垂直结果",
"在我们的用户研究中，我们观察到SERP类型和任务复杂性会",
"???",
"影响用户与搜索结果的交互。我们还发现了支持国[4,31]的8",
"个观测值中的大多数-6个的证据",
"3",
"",
"METHODOLOGY",
"为了解决s1中概述的四个总体研究问题，我们对n=41参与",
"2",
"RELATEDWORK",
"者进行了用户研究。每个参与者被分配到四个实验性搜索界面",
"2.1Task Complexity and User Interactions",
"条件之一（界面：受试者之间），并完成了12个搜索任务（任",
"许多工作都集中在任务类型对SERP上用户交互的影响上。Buscher",
"务：受试者内部）",
"我们的四个实验性搜索接口条件考虑了布局类型(list-vs。基",
"etal.[8]使用查询日志进行了大规模分析，以了解个人和任务",
"差异如何影响搜索行为。他们的发现表明，有一群用户以类似",
"于网格的布局）和SERP上存在的垂直行业（异构内容与。均匀",
"的方式检查搜索结果。他们还表明，任务类型对用户与SERP",
"含量）。这些组合导致下面概述的接口条件，图1中提供了两",
"的互动方式有显着影响。Arguelloetal.[4]观察到任务越复杂，",
"种布局类型的示例，S3.1中提供了更多详细信息",
"与SERP上的各种组件交互的用户就越多，而Thomasetal.[34]",
"SL Simple List Considered as our baseline interface condition",
"发现用户在面对复杂任务时倾向于更深入、更快速地检查结",
"(the standard and widely used ten blue links[11]),this in",
"果列表。Jiang etal.[12]比较了四种不同类型的搜索任务在相",
"terface presents results in a list, with each result presented",
"对较长的搜索会话（10分钟；关于5查询）中的用户交互。E",
"one under the other. All results are web results , and as such",
"Wuetal.[38]还观察到用户与SERP的交互差异，这取决于他",
"are homogeneous in terms of presented content.",
"们是否必须寻找事实问题或非事实问题的答案。在这些研究",
"SG Simple Grid The same homogeneous approach to content is",
"中，SERP由事实上的列表格式的网络搜索结果组成",
"taken as for SL, but with results presented in a grid-based",
"approach.Insteadof scrolling along thevertical,participants",
"2.2",
"SERPPresentation and UserInteractions",
"subjected to this interface scroll along the horizontal",
"Sushmitaetal.[32]观察到，当垂直（新闻，图像和视频）以混",
"HL Heterogeneous List Similar in approach to SL,HL presents",
"合方式呈现时，SERP上不同垂直的位置（顶部，中间，底部）",
"results in a list.However, different verticals aremixed in with",
"会影响用户的点击率。Arguelloetal.[4]还研究了任务复杂性",
"the standard web results. Beyond web results, heterogeneous",
"如何影响用户交互和聚合垂直结果的使用，当它们与Web结",
"content used in this study includes image and video results.",
"果交错时，与它们作为选项卡呈现时相比。类似地，他们观",
"HG Heterogeneous Grid Similar to HL but now the content is",
"察到，对于更复杂的任务，用户在与网络搜索结果交错时点",
"displayed in grid form, with web-based results appearing in",
"击了更垂直的结果。Botaetal.[7]进行了一项众包在线用户研",
"a grid, before additional image and video content.",
"究，以调查给定模糊搜索主题的实体卡的影响。他们发现，实",
"体卡的存在对用户与搜索结果的交互方式及其感知的任务工",
"3.1",
"Search Interface Design and System",
"作量都有很大影响。此外，ELevietal.[16]根据商业网络搜索",
"引擎的日志，对来自七个不同垂直领域（包括垂直社区问题回",
"鉴于上述情况，我们的目标是找出FArguelloetal.【4】和FSiu",
"答）的结果的呈现进行了全面分析。他们观察到，与其他垂直",
"and Chaparro[31]先前研究的观察结果在经过近十年的SERP",
"领域相比，社区问题回答垂直领域平均获得的点击次数最高。",
"设计演变（和添加）后在多天程度上有效。为了实现我们的四"
] |
{
"arxiv_id": "2207.13117",
"language": "zh",
"page_idx": 2
}
|
||
2207.13117-zh-4
|
2207.13117-zh-4.jpg
|
<FigureHere>
Figure 1: Examples of both the (a) list-based and (b) grid-based interfaces trialled. Note the inclusion of links for the separate $\tilde{\mathbf{Q}}$ All(as shown),[2]mages, and WVideos result pages. Heterogeneous content is displayed in red boxes, and is not present in the two homogeneous content interface conditions ( SG and SL ). Circled numbers correspond to the narrative of Section 3.1
通过许多 Qualtrics $^{3}$调查捕获了经验数据；除了参与者必须编写的小的任务后摘要外，每个参与者还完成了实验前和实验后的调查。有关这些问题和实验流程的更多详细信息，请参见 §3.4 。我们的设置可确保参与者在需要时在 Qualtrics 调查和 SERP 之间跳转
静态 SERP 正如禁用的查询框0所暗示的那样，我们的实验设置没有可编程的后端或搜索功能。这意味着没有额外的查询功能。我们提供了手动策划的 SERP,我们为我们要求参与者进行的每个 12 搜索任务生成了先验。此设置确保所有研究参与者查看相同的结果(先前研究中也选择了设置，例如$\mathbb{E}$ Sushmita et al.[32] 和$\boxed{\mathrm{E}}$Wu et al.[38])。虽然使搜索体验不太现实，但它确实为我们提供了不必处理提交各种查询的参与者的好处。该设计还删除了一个混淆变量，并允许我们通过计算一组固定的 Web 结果，图像和视频的用户交互度量来解决我们的四个 RQ
# 3.2 Search Tasks
在本研究中，我们使用了四种不同类型的信息需求：一种导航信息类型(个人寻求找到特定网站),以及三种不同的信息类别，属于 Remember (涉及检索，识别和召回相关知识的部
$\overline{^3{\mathrm{https://www.qu}}}$altrics.com/
分), Understand (构造含义)。对于每个类别，我们产生了三个独特的信息需求。这导致了总共 12 信息需求，这些信息需求在表 2 中列出。特别注意设计任务，诱使参与者不仅查看网络结果，还查看图像和视频搜索结果
我们的信息需求的选择基于$\overline{\mathrm{E}}$Arguello et al.[4] 和$\overline{\mathrm{E}}$Siu and Chaparro [31] 使用的研究设计。更具体地说，EArguello et al. [4]设计了一系列任务，这些任务需要完成不同级别的信息多样性以及不同数量的搜索工作。任务以修订后的 Bloom 分类法为基础，如$\overline\text{E}$Anderson and Krathwohl $[1]\text{ 所概述。}^{4}$ 搜索任务是信息性的，属于 Remember、Understand 和 Analyse 类别。此外，Siu and Chaparro [31] 在他们的研究中只使用了两类任务
一导航和信息。因此，我们的设计结合了我们希望检查的两
个先前作品的设置
3.3 Query Selection and SERP Curatior
为了确保参与者收到有帮助的搜索结果，我们需要为每个参与者提供一个通用的搜索查询。为此，我们在多产平台$^{5}$上进行了一项小型众包试点研究。该飞行员有$n=25$工人，其设计在很大程度上受到了$\overline{\mathrm{E}}$Bailey et al.[6]报告的研究的启发。
$\overline{{^4\text{The Bloom taxo}}}$nomy is typically used to create educational materials.
$^5$https://www.prolific.co/
|
[
[
0.101318359375,
0.126953125,
0.1356201171875,
0.126953125,
0.1356201171875,
0.1307373046875,
0.101318359375,
0.1307373046875
],
[
0.529296875,
0.126953125,
0.560546875,
0.126953125,
0.560546875,
0.1307373046875,
0.529296875,
0.1307373046875
],
[
0.865234375,
0.1256103515625,
0.8798828125,
0.1256103515625,
0.8798828125,
0.1287841796875,
0.865234375,
0.1287841796875
],
[
0.8857421875,
0.126953125,
0.89453125,
0.126953125,
0.89453125,
0.1300048828125,
0.8857421875,
0.1300048828125
],
[
0.25,
0.15087890625,
0.2998046875,
0.15087890625,
0.2998046875,
0.1610107421875,
0.25,
0.1610107421875
],
[
0.29736328125,
0.154052734375,
0.314453125,
0.154052734375,
0.314453125,
0.1578369140625,
0.29736328125,
0.1578369140625
],
[
0.334228515625,
0.1514892578125,
0.375732421875,
0.1514892578125,
0.375732421875,
0.1590576171875,
0.334228515625,
0.1590576171875
],
[
0.54736328125,
0.1673583984375,
0.56396484375,
0.1673583984375,
0.56396484375,
0.17041015625,
0.54736328125,
0.17041015625
],
[
0.396240234375,
0.1875,
0.423095703125,
0.1881103515625,
0.423095703125,
0.1925048828125,
0.396240234375,
0.1912841796875
],
[
0.43701171875,
0.18310546875,
0.51123046875,
0.18310546875,
0.51123046875,
0.1976318359375,
0.43701171875,
0.1976318359375
],
[
0.11767578125,
0.2392578125,
0.1298828125,
0.2392578125,
0.1298828125,
0.242431640625,
0.11767578125,
0.242431640625
],
[
0.68896484375,
0.27783203125,
0.7138671875,
0.27783203125,
0.7138671875,
0.281005859375,
0.68896484375,
0.281005859375
],
[
0.67236328125,
0.285400390625,
0.68310546875,
0.285400390625,
0.68310546875,
0.28857421875,
0.67236328125,
0.28857421875
],
[
0.80615234375,
0.2890625,
0.85302734375,
0.2890625,
0.85302734375,
0.29345703125,
0.80615234375,
0.29345703125
],
[
0.09478759765625,
0.313232421875,
0.11517333984375,
0.313232421875,
0.11517333984375,
0.32080078125,
0.09478759765625,
0.32080078125
],
[
0.53759765625,
0.40087890625,
0.5595703125,
0.40087890625,
0.5595703125,
0.40478515625,
0.53759765625,
0.40478515625
],
[
0.1953125,
0.4873046875,
0.2459716796875,
0.4873046875,
0.2459716796875,
0.501953125,
0.1953125,
0.501953125
],
[
0.238525390625,
0.488037109375,
0.373291015625,
0.488037109375,
0.373291015625,
0.501953125,
0.238525390625,
0.501953125
],
[
0.6259765625,
0.4892578125,
0.67236328125,
0.4892578125,
0.67236328125,
0.50048828125,
0.6259765625,
0.50048828125
],
[
0.66845703125,
0.488037109375,
0.8037109375,
0.488037109375,
0.8037109375,
0.501953125,
0.66845703125,
0.501953125
],
[
0.08660888671875,
0.5166015625,
0.9140625,
0.5166015625,
0.9140625,
0.52978515625,
0.08660888671875,
0.52978515625
],
[
0.08660888671875,
0.52978515625,
0.91259765625,
0.52978515625,
0.91259765625,
0.54443359375,
0.08660888671875,
0.54443359375
],
[
0.0849609375,
0.54345703125,
0.9140625,
0.54345703125,
0.9140625,
0.556640625,
0.0849609375,
0.556640625
],
[
0.10296630859375,
0.59228515625,
0.48193359375,
0.59228515625,
0.48193359375,
0.60595703125,
0.10296630859375,
0.60595703125
],
[
0.51953125,
0.59228515625,
0.91259765625,
0.59228515625,
0.91259765625,
0.60595703125,
0.51953125,
0.60595703125
],
[
0.08660888671875,
0.60595703125,
0.4794921875,
0.60595703125,
0.4794921875,
0.619140625,
0.08660888671875,
0.619140625
],
[
0.51953125,
0.60693359375,
0.91259765625,
0.60693359375,
0.91259765625,
0.62060546875,
0.51953125,
0.62060546875
],
[
0.08660888671875,
0.62060546875,
0.48046875,
0.62060546875,
0.48046875,
0.6337890625,
0.08660888671875,
0.6337890625
],
[
0.5185546875,
0.62060546875,
0.91259765625,
0.62060546875,
0.91259765625,
0.6337890625,
0.5185546875,
0.6337890625
],
[
0.0882568359375,
0.63623046875,
0.48046875,
0.63623046875,
0.48046875,
0.646484375,
0.0882568359375,
0.646484375
],
[
0.51953125,
0.63623046875,
0.75341796875,
0.63623046875,
0.75341796875,
0.646484375,
0.51953125,
0.646484375
],
[
0.08660888671875,
0.6484375,
0.200927734375,
0.6484375,
0.200927734375,
0.662109375,
0.08660888671875,
0.662109375
],
[
0.53515625,
0.6484375,
0.9140625,
0.6484375,
0.9140625,
0.662109375,
0.53515625,
0.662109375
],
[
0.5185546875,
0.662109375,
0.9111328125,
0.662109375,
0.9111328125,
0.67529296875,
0.5185546875,
0.67529296875
],
[
0.0882568359375,
0.67041015625,
0.4794921875,
0.67041015625,
0.4794921875,
0.68115234375,
0.0882568359375,
0.68115234375
],
[
0.0882568359375,
0.68310546875,
0.48046875,
0.68310546875,
0.48046875,
0.6962890625,
0.0882568359375,
0.6962890625
],
[
0.521484375,
0.67822265625,
0.9111328125,
0.67822265625,
0.9111328125,
0.68798828125,
0.521484375,
0.68798828125
],
[
0.5185546875,
0.68994140625,
0.91259765625,
0.68994140625,
0.91259765625,
0.7041015625,
0.5185546875,
0.7041015625
],
[
0.08905029296875,
0.69873046875,
0.48046875,
0.69873046875,
0.48046875,
0.708984375,
0.08905029296875,
0.708984375
],
[
0.5185546875,
0.7041015625,
0.91259765625,
0.7041015625,
0.91259765625,
0.71728515625,
0.5185546875,
0.71728515625
],
[
0.0882568359375,
0.7109375,
0.48193359375,
0.7109375,
0.48193359375,
0.724609375,
0.0882568359375,
0.724609375
],
[
0.51953125,
0.71728515625,
0.91259765625,
0.71728515625,
0.91259765625,
0.73095703125,
0.51953125,
0.73095703125
],
[
0.0882568359375,
0.724609375,
0.48193359375,
0.724609375,
0.48193359375,
0.73779296875,
0.0882568359375,
0.73779296875
],
[
0.5185546875,
0.73193359375,
0.9111328125,
0.73193359375,
0.9111328125,
0.74560546875,
0.5185546875,
0.74560546875
],
[
0.08660888671875,
0.73779296875,
0.48046875,
0.73779296875,
0.48046875,
0.751953125,
0.08660888671875,
0.751953125
],
[
0.5498046875,
0.74560546875,
0.91259765625,
0.74560546875,
0.91259765625,
0.7587890625,
0.5498046875,
0.7587890625
],
[
0.0882568359375,
0.75244140625,
0.48046875,
0.75244140625,
0.48046875,
0.7666015625,
0.0882568359375,
0.7666015625
],
[
0.521484375,
0.7587890625,
0.6396484375,
0.7587890625,
0.6396484375,
0.77294921875,
0.521484375,
0.77294921875
],
[
0.08905029296875,
0.76806640625,
0.47802734375,
0.76806640625,
0.47802734375,
0.77880859375,
0.08905029296875,
0.77880859375
],
[
0.0882568359375,
0.7822265625,
0.48046875,
0.7822265625,
0.48046875,
0.79248046875,
0.0882568359375,
0.79248046875
],
[
0.5185546875,
0.787109375,
0.55224609375,
0.787109375,
0.55224609375,
0.798828125,
0.5185546875,
0.798828125
],
[
0.560546875,
0.7861328125,
0.85302734375,
0.78515625,
0.85302734375,
0.7998046875,
0.560546875,
0.80029296875
],
[
0.08660888671875,
0.79443359375,
0.1829833984375,
0.79443359375,
0.1829833984375,
0.80615234375,
0.08660888671875,
0.80615234375
],
[
0.51953125,
0.8037109375,
0.9111328125,
0.8037109375,
0.9111328125,
0.8173828125,
0.51953125,
0.8173828125
],
[
0.08416748046875,
0.8125,
0.23779296875,
0.81298828125,
0.23779296875,
0.82763671875,
0.08416748046875,
0.8271484375
],
[
0.521484375,
0.8173828125,
0.91259765625,
0.8173828125,
0.91259765625,
0.83203125,
0.521484375,
0.83203125
],
[
0.0849609375,
0.830078125,
0.48046875,
0.83056640625,
0.48046875,
0.84521484375,
0.0849609375,
0.8447265625
],
[
0.51708984375,
0.83203125,
0.9111328125,
0.83056640625,
0.9111328125,
0.84521484375,
0.51708984375,
0.8466796875
],
[
0.08660888671875,
0.84521484375,
0.4794921875,
0.84521484375,
0.4794921875,
0.859375,
0.08660888671875,
0.859375
],
[
0.5185546875,
0.8466796875,
0.91259765625,
0.8466796875,
0.91259765625,
0.85986328125,
0.5185546875,
0.85986328125
],
[
0.08660888671875,
0.859375,
0.48193359375,
0.859375,
0.48193359375,
0.87255859375,
0.08660888671875,
0.87255859375
],
[
0.515625,
0.87060546875,
0.8505859375,
0.87255859375,
0.8505859375,
0.88720703125,
0.515625,
0.88525390625
],
[
0.08660888671875,
0.88330078125,
0.22216796875,
0.88330078125,
0.22216796875,
0.896484375,
0.08660888671875,
0.896484375
],
[
0.51708984375,
0.8818359375,
0.63818359375,
0.8837890625,
0.63818359375,
0.89794921875,
0.51708984375,
0.89599609375
]
] |
[
"Senrch",
"Searc",
"A",
"Va",
"QAlI四I",
"mar",
"Videc",
"shD",
"din",
"les results page",
"101",
"Sa",
"Mo",
"Ivdroelectrie",
"",
"Videc",
"(a) HL",
"Heterogeneous List",
"(b) HG",
"Heterogeneous Grid",
"Figure l: Examples of both the (a) list-based and (b) grid-based interfaces trialled. Note the inclusion of links for the separate",
"Q All (as shown), Images , and Videos result pages. Heterogeneous content is displayed in red boxes, and is not present",
"in the two homogeneous content interface conditions (SG and SL). Circled numbers correspond to the narrative of Section 3.1",
"通过许多Qualtrics3调查捕获了经验数据；除了参与者必须",
"分），Understand（构造含义）。对于每个类别，我们产生了三",
"编写的小的任务后摘要外，每个参与者还完成了实验前和实",
"个独特的信息需求。这导致了总共12信息需求，这些信息需",
"验后的调查。有关这些问题和实验流程的更多详细信息，请参",
"求在表2中列出。特别注意设计任务，诱使参与者不仅查看网",
"见S3.4。我们的设置可确保参与者在需要时在Qualtrics调查",
"络结果，还查看图像和视频搜索结果",
"和SERP之间跳转",
"我们的信息需求的选择基于国Arguelloetal.[4]和ESiuand",
"Chaparro[31]使用的研究设计。更具体地说，EArguelloetal",
"静态SERP正如禁用的查询框①所暗示的那样，我们的实验",
"设置没有可编程的后端或搜索功能。这意味着没有额外的查",
"4]设计了一系列任务，这些任务需要完成不同级别的信息多",
"样性以及不同数量的搜索工作。任务以修订后的Bloom分类",
"询功能。我们提供了手动策划的SERP，我们为我们要求参与",
"法为基础，如国AndersonandKrathwohl[1]所概述。4搜索任",
"者进行的每个12搜索任务生成了先验。此设置确保所有研究",
"务是信息性的，属于Remember、Understand和Analyse类别。",
"参与者查看相同的结果（先前研究中也选择了设置，例如E",
"此外，SiuandChaparro[31]在他们的研究中只使用了两类任务",
"Sushmitaetal.[32]和EWuetal.[38]）。虽然使搜索体验不太",
"导航和信息。因此，我们的设计结合了我们希望检查的两",
"现实，但它确实为我们提供了不必处理提交各种查询的参与",
"个先前作品的设置",
"者的好处。该设计还删除了一个混变量，并充允许我们通过计",
"算一组固定的Web结果，图像和视频的用户交互度量来解决",
"3.3",
"Query Selection and SERP Curation",
"我们的四个RO",
"为了确保参与者收到有帮助的搜索结果，我们需要为每个参",
"3.2SearchTasks",
"与者提供一个通用的搜索查询。为此，我们在多产平台5上进",
"在本研究中，我们使用了四种不同类型的信息需求：一种导",
"行了一项小型众包试点研究。该飞行员有n=25工人，其设",
"航信息类型（个人寻求找到特定网站），以及三种不同的信息",
"计在很大程度上受到了EBaileyet al.[6]报告的研究的启发。。",
"类别，属于Remember（涉及检索，识别和召回相关知识的部",
"4The Bloom taxonomy is typically used to create educational materials.",
"3https:/www.qualtrics.com/",
"5https://www.prolific.co/"
] |
{
"arxiv_id": "2207.13117",
"language": "zh",
"page_idx": 4
}
|
||
2209.10951-zh-1
|
2209.10951-zh-1.jpg
|
基于信息最小化的无监督句子嵌人对比学习模型
Shaobin Chen$^1,\textbf{Jie Zhou}^2,\textbf{Yuling Sun}^{1*}$and Liang He$^1$
$^{1}$School of Computer Science and Technology, East China Normal University Shanghai, China
$^{2}$School of Computer Science, Fudan Univerisity Shanghai, China
shaobin\_chen@stu.ecnu.edu.cn, jie\_zhou@fudan.edu.cn,{ ylsun, lhe{ }@cs.ecnu.edu.cn
### $\mathbf{Abstract}$
Unsupervised sentence embeddings learning has been recently dominated by contrastive learning methods (e.g., SimCSE), which keep positive pairs similar and push negative pairs apart. The contrast operation aims to keep as much information as possible by maximizing the mutual information between positive instances, which leads to redundant information in sentence embedding. To address this problem, we present an information minimization based contrastive learning( InforMin-CL) model to retain the useful information and discard the redundant information by maximizing the mutual information and minimizing the information entropy between positive instances meanwhile for unsupervised sentence representation learning. Specifically, we find that information minimization can be achieved by simple contrast and reconstruction objectives. The reconstruction operation reconstitutes the positive instance via the other positive instance to minimize the information entropy between positive instances. We evaluate our model on fourteen downstream tasks, including both supervised and unsupervised (semantic textual similarity) tasks. Extensive experimental results show that our InforMin-CL obtains a state-of-the-art performance. Code is made available. $^{\mathrm{l}}$
### 1 $\textbf{Introduction}$
如何通过大型预训练模型(??)(例如 BERT) 学习通用句子嵌人，已在文献(??)中进行了广泛的研究。最近，对比学习已被广泛用于学习更好的句子嵌人(??)。通常，对比学习使用各种数据增强方法来生成输入句子的不同视图，并从视图中采样正面实例和负面实例。对比学习旨在通过将正面实例放在一起并将正面和负面实例分开来学习有效的嵌入。此操作着重于最大化正实例之间的互信息，以保留尽可能多的信息，其中包括有用和无用的信息。像
$^*$ Yuling Sun is the corresponding authors of this paper.
$^{1}$https://github.com/Bin199/InforMin-CL
Table 1: Training texts may contain various kinds of redundant information, such as stop words, restatement, capitalization, and hyphen.
<table>
<tbody>
<tr>
<th>Original</th>
<th>Where is the party, it sounds great.</th>
</tr>
<tr>
<td>Stop words</td>
<td>Where is the $party$, $it$ sounds s great</td>
</tr>
<tr>
<td>Restatement</td>
<td>The party sounds great, where is it.</td>
</tr>
<tr>
<td>Capitalization</td>
<td>Where Is The Party, It Sounds Great.</td>
</tr>
<tr>
<td>Hyphen</td>
<td>Where- is- the- party, , it-sounds-great.</td>
</tr>
</tbody>
</table>
(??)这样的先前研究忽略了视图中存储的冗余信息，这对下游任务的性能产生了不良影响， 正如(??)所证明的那样
表 1 给出了存储在训练文本中的冗余信息的示例。在训练文本中，句子包含许多冗余信息， 不利于下游任务。冗余信息可能是停用词和句子的风格(e。g。重述、大写和连字符)。现有的研究 XMTHX\_1 还表明，丢弃视图中的冗余信息有助于提高下游任务的性能。因此，我们提出了以下问题：如何通过选择最佳视图来丢弃这些冗余信息？
通过信息瓶颈(IB)解决上述问题是很自然的，它已被用作一种有效且简单的方法，通过在各种任务中保留重要信息并忘记冗余信息来学习良好的嵌入(???)。在此提示下，我们通过从信息最小化原则(IB 理论中的一个想法) (?)中汲取灵感来解决问题：一组好的视图共享在下游任务中表现出色所需的最少信息。此方法旨在保留有用的信息并忘记冗余信息。在本文中，我们探讨了如何解决(i。e。,忽略视图中存储的冗余信息),通过信息最小化原则对无监督句子表示进行先前的工作
我们提出了一种基于信息最小化的对比学习 (InforMin-CL)模型，用于无监督句子嵌入学习。InforMin-CL 将信息最小化原则融人到对比学习中，不仅可以学习重要信息， 还可以丢弃冗余信息。我们从两个角度优化了InforMin-CL 模型：对比和重建。首先，我们通过对比任务学习有用的信息，以最大化互信息。此任务设法吸引正对和排斥负对。吸引正对代表最大化正实例之间的互信息。其次，
|
[
[
0.246826171875,
0.09442138671875,
0.75,
0.0955810546875,
0.75,
0.10986328125,
0.246826171875,
0.1092529296875
],
[
0.264404296875,
0.1395263671875,
0.736328125,
0.1395263671875,
0.736328125,
0.15380859375,
0.264404296875,
0.15380859375
],
[
0.12005615234375,
0.15380859375,
0.88330078125,
0.155029296875,
0.88330078125,
0.1728515625,
0.12005615234375,
0.171630859375
],
[
0.2359619140625,
0.17333984375,
0.765625,
0.17333984375,
0.765625,
0.188232421875,
0.2359619140625,
0.188232421875
],
[
0.1435546875,
0.191162109375,
0.85791015625,
0.191162109375,
0.85791015625,
0.204833984375,
0.1435546875,
0.204833984375
],
[
0.2626953125,
0.254150390625,
0.341796875,
0.254150390625,
0.341796875,
0.268310546875,
0.2626953125,
0.268310546875
],
[
0.51123046875,
0.2529296875,
0.880859375,
0.2529296875,
0.880859375,
0.2666015625,
0.51123046875,
0.2666015625
],
[
0.51318359375,
0.267333984375,
0.880859375,
0.267333984375,
0.880859375,
0.28076171875,
0.51318359375,
0.28076171875
],
[
0.14697265625,
0.27978515625,
0.45751953125,
0.27978515625,
0.45751953125,
0.29345703125,
0.14697265625,
0.29345703125
],
[
0.51123046875,
0.28076171875,
0.6943359375,
0.28076171875,
0.6943359375,
0.29443359375,
0.51123046875,
0.29443359375
],
[
0.145263671875,
0.29345703125,
0.45751953125,
0.29345703125,
0.45751953125,
0.30712890625,
0.145263671875,
0.30712890625
],
[
0.145263671875,
0.30712890625,
0.456787109375,
0.30712890625,
0.456787109375,
0.320556640625,
0.145263671875,
0.320556640625
],
[
0.5263671875,
0.311279296875,
0.5810546875,
0.311279296875,
0.5810546875,
0.325927734375,
0.5263671875,
0.325927734375
],
[
0.63232421875,
0.311279296875,
0.84228515625,
0.312255859375,
0.84228515625,
0.325927734375,
0.63232421875,
0.32470703125
],
[
0.145263671875,
0.322509765625,
0.45751953125,
0.322509765625,
0.45751953125,
0.336181640625,
0.145263671875,
0.336181640625
],
[
0.634765625,
0.334228515625,
0.6767578125,
0.334228515625,
0.6767578125,
0.345703125,
0.634765625,
0.345703125
],
[
0.67236328125,
0.333251953125,
0.86376953125,
0.333251953125,
0.86376953125,
0.3466796875,
0.67236328125,
0.3466796875
],
[
0.5283203125,
0.3349609375,
0.599609375,
0.3349609375,
0.599609375,
0.3466796875,
0.5283203125,
0.3466796875
],
[
0.145263671875,
0.336181640625,
0.45751953125,
0.336181640625,
0.45751953125,
0.349853515625,
0.145263671875,
0.349853515625
],
[
0.779296875,
0.33837890625,
0.7900390625,
0.33837890625,
0.7900390625,
0.342529296875,
0.779296875,
0.342529296875
],
[
0.5283203125,
0.347412109375,
0.60693359375,
0.347412109375,
0.60693359375,
0.359375,
0.5283203125,
0.359375
],
[
0.63037109375,
0.344482421875,
0.84619140625,
0.345703125,
0.84619140625,
0.35986328125,
0.63037109375,
0.359375
],
[
0.14697265625,
0.350830078125,
0.45751953125,
0.350830078125,
0.45751953125,
0.364013671875,
0.14697265625,
0.364013671875
],
[
0.5283203125,
0.359375,
0.615234375,
0.359375,
0.615234375,
0.37060546875,
0.5283203125,
0.37060546875
],
[
0.63037109375,
0.355712890625,
0.85546875,
0.358154296875,
0.85546875,
0.371826171875,
0.63037109375,
0.369384765625
],
[
0.145263671875,
0.365234375,
0.459228515625,
0.365234375,
0.459228515625,
0.37890625,
0.145263671875,
0.37890625
],
[
0.5263671875,
0.369384765625,
0.5810546875,
0.369384765625,
0.5810546875,
0.38427734375,
0.5263671875,
0.38427734375
],
[
0.63232421875,
0.3681640625,
0.8505859375,
0.37060546875,
0.8505859375,
0.38427734375,
0.63232421875,
0.3818359375
],
[
0.14697265625,
0.37890625,
0.45751953125,
0.37890625,
0.45751953125,
0.392578125,
0.14697265625,
0.392578125
],
[
0.14697265625,
0.392578125,
0.45751953125,
0.392578125,
0.45751953125,
0.40625,
0.14697265625,
0.40625
],
[
0.145263671875,
0.407958984375,
0.45751953125,
0.407958984375,
0.45751953125,
0.421630859375,
0.145263671875,
0.421630859375
],
[
0.145263671875,
0.421630859375,
0.45751953125,
0.421630859375,
0.45751953125,
0.435302734375,
0.145263671875,
0.435302734375
],
[
0.51318359375,
0.41748046875,
0.880859375,
0.41748046875,
0.880859375,
0.43115234375,
0.51318359375,
0.43115234375
],
[
0.14697265625,
0.43701171875,
0.456787109375,
0.43701171875,
0.456787109375,
0.44775390625,
0.14697265625,
0.44775390625
],
[
0.5146484375,
0.432861328125,
0.8818359375,
0.432861328125,
0.8818359375,
0.446533203125,
0.5146484375,
0.446533203125
],
[
0.145263671875,
0.449462890625,
0.45751953125,
0.449462890625,
0.45751953125,
0.463134765625,
0.145263671875,
0.463134765625
],
[
0.51318359375,
0.448974609375,
0.7021484375,
0.448974609375,
0.7021484375,
0.4619140625,
0.51318359375,
0.4619140625
],
[
0.14697265625,
0.46435546875,
0.45751953125,
0.46435546875,
0.45751953125,
0.47802734375,
0.14697265625,
0.47802734375
],
[
0.53076171875,
0.46728515625,
0.880859375,
0.46728515625,
0.880859375,
0.48095703125,
0.53076171875,
0.48095703125
],
[
0.145263671875,
0.478515625,
0.45751953125,
0.478515625,
0.45751953125,
0.4921875,
0.145263671875,
0.4921875
],
[
0.51318359375,
0.482666015625,
0.87744140625,
0.482666015625,
0.87744140625,
0.496337890625,
0.51318359375,
0.496337890625
],
[
0.14697265625,
0.4921875,
0.45751953125,
0.4921875,
0.45751953125,
0.505859375,
0.14697265625,
0.505859375
],
[
0.515625,
0.50048828125,
0.87890625,
0.50048828125,
0.87890625,
0.51123046875,
0.515625,
0.51123046875
],
[
0.14697265625,
0.50732421875,
0.45751953125,
0.50732421875,
0.45751953125,
0.52099609375,
0.14697265625,
0.52099609375
],
[
0.145263671875,
0.52099609375,
0.456787109375,
0.521484375,
0.456787109375,
0.53515625,
0.145263671875,
0.53369140625
],
[
0.515625,
0.515625,
0.880859375,
0.515625,
0.880859375,
0.52978515625,
0.515625,
0.52978515625
],
[
0.145263671875,
0.53515625,
0.456787109375,
0.53515625,
0.456787109375,
0.548828125,
0.145263671875,
0.548828125
],
[
0.51123046875,
0.52978515625,
0.87890625,
0.52978515625,
0.87890625,
0.54345703125,
0.51123046875,
0.54345703125
],
[
0.145263671875,
0.5498046875,
0.456787109375,
0.5498046875,
0.456787109375,
0.5634765625,
0.145263671875,
0.5634765625
],
[
0.5146484375,
0.54736328125,
0.880859375,
0.54736328125,
0.880859375,
0.56103515625,
0.5146484375,
0.56103515625
],
[
0.145263671875,
0.56396484375,
0.45751953125,
0.56396484375,
0.45751953125,
0.57763671875,
0.145263671875,
0.57763671875
],
[
0.51318359375,
0.56396484375,
0.880859375,
0.56396484375,
0.880859375,
0.57763671875,
0.51318359375,
0.57763671875
],
[
0.14697265625,
0.580078125,
0.456787109375,
0.580078125,
0.456787109375,
0.59033203125,
0.14697265625,
0.59033203125
],
[
0.51318359375,
0.580078125,
0.6533203125,
0.580078125,
0.6533203125,
0.59423828125,
0.51318359375,
0.59423828125
],
[
0.14697265625,
0.59326171875,
0.456787109375,
0.59326171875,
0.456787109375,
0.60400390625,
0.14697265625,
0.60400390625
],
[
0.53076171875,
0.59716796875,
0.880859375,
0.59716796875,
0.880859375,
0.61083984375,
0.53076171875,
0.61083984375
],
[
0.145263671875,
0.60546875,
0.45751953125,
0.60546875,
0.45751953125,
0.619140625,
0.145263671875,
0.619140625
],
[
0.51318359375,
0.6142578125,
0.87744140625,
0.6142578125,
0.87744140625,
0.62451171875,
0.51318359375,
0.62451171875
],
[
0.14697265625,
0.62060546875,
0.45751953125,
0.62060546875,
0.45751953125,
0.63427734375,
0.14697265625,
0.63427734375
],
[
0.14697265625,
0.63525390625,
0.45751953125,
0.63525390625,
0.45751953125,
0.6484375,
0.14697265625,
0.6484375
],
[
0.51318359375,
0.6298828125,
0.880859375,
0.6298828125,
0.880859375,
0.6435546875,
0.51318359375,
0.6435546875
],
[
0.1485595703125,
0.6494140625,
0.455078125,
0.6494140625,
0.455078125,
0.66015625,
0.1485595703125,
0.66015625
],
[
0.51318359375,
0.6455078125,
0.880859375,
0.6455078125,
0.880859375,
0.6591796875,
0.51318359375,
0.6591796875
],
[
0.145263671875,
0.6630859375,
0.45751953125,
0.6630859375,
0.45751953125,
0.6767578125,
0.145263671875,
0.6767578125
],
[
0.51123046875,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.51123046875,
0.67333984375
],
[
0.14697265625,
0.677734375,
0.2225341796875,
0.677734375,
0.2225341796875,
0.689453125,
0.14697265625,
0.689453125
],
[
0.51123046875,
0.677734375,
0.87744140625,
0.677734375,
0.87744140625,
0.6884765625,
0.51123046875,
0.6884765625
],
[
0.5146484375,
0.69482421875,
0.8818359375,
0.69482421875,
0.8818359375,
0.7080078125,
0.5146484375,
0.7080078125
],
[
0.1158447265625,
0.703125,
0.145263671875,
0.703125,
0.145263671875,
0.71435546875,
0.1158447265625,
0.71435546875
],
[
0.151123046875,
0.703125,
0.26123046875,
0.703125,
0.26123046875,
0.71435546875,
0.151123046875,
0.71435546875
],
[
0.51318359375,
0.71142578125,
0.87744140625,
0.71142578125,
0.87744140625,
0.7216796875,
0.51318359375,
0.7216796875
],
[
0.118408203125,
0.72705078125,
0.48779296875,
0.72705078125,
0.48779296875,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.51123046875,
0.7255859375,
0.87890625,
0.7255859375,
0.87890625,
0.7392578125,
0.51123046875,
0.7392578125
],
[
0.118408203125,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.5146484375,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.75732421875,
0.5146484375,
0.75732421875
],
[
0.118408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51318359375,
0.7587890625,
0.78857421875,
0.7587890625,
0.78857421875,
0.7724609375,
0.51318359375,
0.7724609375
],
[
0.118408203125,
0.775390625,
0.486083984375,
0.775390625,
0.486083984375,
0.7890625,
0.118408203125,
0.7890625
],
[
0.52880859375,
0.77685546875,
0.8818359375,
0.77685546875,
0.8818359375,
0.79052734375,
0.52880859375,
0.79052734375
],
[
0.12176513671875,
0.79345703125,
0.484375,
0.79345703125,
0.484375,
0.80419921875,
0.12176513671875,
0.80419921875
],
[
0.51123046875,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.51123046875,
0.8056640625
],
[
0.118408203125,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.51123046875,
0.80810546875,
0.87890625,
0.80810546875,
0.87890625,
0.82177734375,
0.51123046875,
0.82177734375
],
[
0.12005615234375,
0.826171875,
0.483642578125,
0.826171875,
0.483642578125,
0.8359375,
0.12005615234375,
0.8359375
],
[
0.51123046875,
0.82373046875,
0.87744140625,
0.82470703125,
0.87744140625,
0.8388671875,
0.51123046875,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.8515625,
0.118408203125,
0.8515625
],
[
0.51318359375,
0.84130859375,
0.87744140625,
0.84130859375,
0.87744140625,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.118408203125,
0.85546875,
0.484375,
0.85546875,
0.484375,
0.869140625,
0.118408203125,
0.869140625
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51318359375,
0.87451171875,
0.87744140625,
0.87451171875,
0.87744140625,
0.88525390625,
0.51318359375,
0.88525390625
],
[
0.1561279296875,
0.8935546875,
0.483642578125,
0.8935546875,
0.483642578125,
0.90673828125,
0.1561279296875,
0.90673828125
],
[
0.51318359375,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.15283203125,
0.896484375,
0.160400390625,
0.89794921875,
0.15869140625,
0.90185546875,
0.151123046875,
0.8994140625
],
[
0.139404296875,
0.90625,
0.3896484375,
0.90625,
0.3896484375,
0.91943359375,
0.139404296875,
0.91943359375
],
[
0.5146484375,
0.90625,
0.8818359375,
0.90625,
0.8818359375,
0.91943359375,
0.5146484375,
0.91943359375
]
] |
[
"基于信息最小化的无监督句子嵌人对比学习模型",
"Shaobin Chen', Jie Zhou?, Yuling Sunl* and Liang He",
"1School of Computer Science and Technology, East China Normal University Shanghai, China",
"2School of Computer Science, Fudan Univerisity Shanghai, China",
"shaobin_chen@stu.ecnu.edu.cn, jie_zhou@fudan.edu.cn, ( ylsun, lhe @cs.ecnu.edu.cn",
"Abstract",
"Table I: Training texts may contain various kinds of re-",
"dundant information, such as stop words, restatement,",
"Unsupervised sentence embeddings learning",
"capitalization, and hyphen.",
"has been recently dominated by contrastive",
"learning methods (e.g., SimCSE), which keep",
"Original",
"Where is the party, it sounds great",
"positive pairs similar and push negative pairs",
"Where",
"is the party,it sounds great.",
"Stop words",
"apart. The contrast operation aims to keep as",
"S",
"Restatement",
"The party sounds great, where is it.",
"much information as possible by maximizing",
"Capitalization",
"Where Is The Party, It Sounds Great.",
"the mutual information between positive in-",
"Hyphen",
"Where-is-the-party, it-sounds-great.",
"stances, which leads to redundant information",
"in sentence embedding. To address this prob-",
"lem, we present an information minimization",
"based contrastive learning ( InforMin-CL)",
"（??）这样的先前研究忽略了视图中存储的穴余",
"model to retain the useful information and dis-",
"信息，这对下游任务的性能产生了不良影响，",
"card the redundant information by maximizing",
"正如（??）所证明的那样",
"the mutual information and minimizing the in-",
"表1给出了存储在训练文本中的余信息的",
"formation entropy between positive instances",
"示例。在训练文本中，句子包含许多穴余信息",
"meanwhile for unsupervised sentence repre-",
"不利于下游任务。元余信息可能是停用词和句",
"sentation learning. Specifically, we find that",
"information minimization can be achieved by",
"子的风格(e。g。、重述、大写和连字符)。现有",
"simple contrast and reconstruction objectives.",
"的研究XMTHX_1还表明，丢弃视图中的余",
"The reconstruction operation reconstitutes the",
"信息有助于提高下游任务的性能。因此，我们",
"positive instance via the other positive instance",
"提出了以下问题：如何通过选择最佳视图来丢",
"tominimizetheinformationentropybetween",
"弃这些穴余信息？",
"positiveinstances.Weevaluate our model on",
"通过信息瓶颈（IB）解决上述问题是很自然",
"fourteen downstream tasks, including both su-",
"的，它已被用作一种有效月简单的方法，通过",
"pervised and unsupervised (semantic textual",
"similarity) tasks. Extensive experimental re-",
"在各种任务中保留重要信息并忘记穴余信息来",
"sults showthat our InforMin-CLobtains",
"学习良好的嵌人（?？？）。在此提示下，我们通",
"a state-of-the-art performance. Code is made",
"过从信息最小化原则（IB理论中的一个想法）",
"available.1",
"（?）中汲取灵感来解决问题：一组好的视图共",
"享在下游任务中表现出色所需的最少信息。此",
"1",
"Introduction",
"方法旨在保留有用的信息并忘记余信息。在",
"如何通过大型预训练模型（??）（例如BERT)",
"本文中，我们探讨了如何解决i。e。，忽略视",
"学习通用句子嵌人，已在文献（??）中进行了",
"图中存储的穴余信息），通过信息最小化原则",
"广泛的研究。最近，对比学习已被广泛用于学",
"对无监督句子表示进行先前的工作",
"习更好的句子嵌入（?？）。通常，对比学习使",
"我们提出了一种基于信息最小化的对比学",
"用各种数据增强方法来生成输入句子的不同视",
"习（InforMin-CL）模型，用于无监督句子",
"图，并从视图中采样正面实例和负面实例。对",
"嵌入学习。InforMin-cL将信息最小化原则",
"比学习盲在通过将止面实例放在一起并将止面",
"融人到对比学习中，不仅可以学习重要信息",
"和负面实例分开来学习有效的嵌人。此操作着",
"还可以丢弃穴余信息。我们从两个角度优化了",
"重于最大化正实例之间的互信息，以保留尽可",
"InforMin-cL模型：对比和重建。首先，我",
"能多的信息，其中包括有用和无用的信息。像",
"们通过对比任务学习有用的信息，以最大化互",
"Yuling Sun is the corresponding authors of this paper.",
"信息。此任务设法吸引正对和排斥负对。吸引",
"",
"\"https://github.com/Bin199/nforMin-CL",
"正对代表最大化正实例之间的互信息。其次，"
] |
{
"arxiv_id": "2209.10951",
"language": "zh",
"page_idx": 1
}
|
||
2209.10951-zh-2
|
2209.10951-zh-2.jpg
|
对中的另一个重建正实例的表示
InforMin-CL 的算法易于理解，只需几行代码即可实现。此外，我们的方法不会改变主要的网络结构，因此它是模型不可知的，可以应用于任何基于对比学习的表示学习神经网络。4节中的实验表明，InforMin-CL 可以帮助模型学习有效的表示形式，从而提高下游任务性能。我们的主要贡献总结如下：
$\cdot$ We propose an InforMin-CL model to
learn good sentence embeddings by keeping
useful information and getting rid of redun-
dant information between positive instances.
$\cdot$ We achieve our model via simple contrast and
reconstruction tasks and prove that the recon-
struction task can drop redundant information
by minimizing the information entropy.
$\cdot$ We achieve new state-of-the-art results on seven supervised tasks and seven unsupervised tasks, which indicates the great advantages of our proposed model.
## 2 $\textbf{Related Work}$
2.1 Sentence Representation Learning
学习句子嵌入作为 NLP 中的一个重要问题已经得到了广泛的研究。最近的工作重点是利用BERT(?)的功能来学习有效的句子嵌人，这些句子嵌人不受人为监督的信号的影响。BERTflow(?)通过归一化流将各向异性句子嵌人分布转换为光滑的各向同性高斯分布。BERT-白化(?)进一步提出了一种白化操作来增强句子嵌入的各向同性，并取得了更好的效果
然后，将对比学习方法应用于句子嵌人学习。IS-BERT(?)提出了一个模型，该模型在BERT 之上具有特征提取器，并且目标是最大化全局句子嵌人和局部句子嵌入之间的互信息。CLEAR(?)采用多种句子级增强策略来学习句子表示。Coco-LM(?)采用辅助语言模型来破坏文本序列，在此基础上构造一个令牌级任务和一个序列级任务，用于预训练主模型。Gao 等人。(?)提出了一种无监督的方法，可以预测带有丢失噪声的输入本身，以及一种利用自然语言推理数据集的监督方法。SCD(?) 利用了通过退出获得的增强样本的自我对比， 从而消除了对负对的依赖。然而，所有这些研究都没有考虑丢弃存储在视图中的冗余信息。在我们的工作中，我们在信息最小化原则的框架内考虑和解决问题
我们提出了一个重建任务，鼓励模型通过同一2.2 Information Minimization Principle
信息最小化原则(?)已被提出来保留必要的最小信息。近年来，研究人员利用信息最小化原理来改进图像表示(??) 。此外，信息瓶颈被用来提高基于注意力的模型(?)的可解释性。Tian 等人。(?)在对比表示学习框架中显示了给定任务的良好视图，应保留与任务相关的信息，同时最大程度地减少无关的麻烦。然而， 它侧重于通过下游数据集消除与任务无关的信息。Tsai 等人。(?)专注于输入和自监督信号之间的多视图设置，并采用自监督信号来重建学习表示，以丢弃与任务无关的信息。我们的工作在两个方面与(?)不同：1)我们丢弃文本中的冗余信息，而(?)丢弃图像中存储的噪声信息，并且在两项工作中使用了不同的自监督信号；2)(?)通过学习由监督任务评估的视觉特征来验证其方法，而句子嵌入学习不仅通过监督任务而且通过无监督任务来评估嵌人
### $\textbf{3}$ $\textbf{Method}$
我们提出了一种用于无监督句子表示学习的InforMin-CL 模型(图1)。有两个主要步骤，对比和重建。我们首先提出对比目标来学习有用的信息：我们将正面实例和负面实例分开，同时将正面实例拉在一起，这意味着最大化正面实例之间的互信息。稍后，我们提出重建任务以丢弃无用的信息：在给定另一个正实例的情况下，我们最小化一个正实例的条件信息熵。算法 1 提供了 InforMin-CL 的伪代码
Input: batch size N, temperature $\tau$, structure of $f$ and $\Gamma.$
${\mathrm{Output:~encoder~network~}f\left(\cdot\right)}.$
$\begin{array}{c}{\textbf{for sampled minibatch}\left\{x_{k}\right\}_{k=1}^{N}}&{\textbf{do}}\\{\textbf{for all }k\in\{1,\ldots,N\}}&{\textbf{do}}\end{array}$
draw two augmentation functions $t\sim\Gamma,t^\prime\sim\Gamma$
$\overline{\mathbf{Algorithm}\:\mathbf{1}\:\text{Pseudocode of InforMin-CL in a}}$
PyTorch-like style.
$\tilde{x}_{2k-1}=t\left(\bar{x}_{k}\right)$ $z_{2k-1}=f\left(\tilde{x}_{2k-1}\right)$ $\tilde{x}_{2k}=t^{\prime}\stackrel{\sim}{\left(x_{k}\right)}$ $z_{2k}=f\left(\tilde{x}_{2k}\right)$
end for
$\textbf{for all}$ $i\in \{ 1, \ldots , 2N\}$and$j\in \{ 1, \ldots , 2N\} \textbf{do}$
$\begin{array}{c}{s_{i,j}=z_{i}^{\mathrm{T}}z_{j}/\:(\|z_{i}\|\:\|z_{j}\|)}\\{\textbf{end for}}\end{array}$
$L_{C}=\operatorname*{max}$E$\left|\frac1N\sum_{i=1}^{N}\log\frac{\exp\left(s_{2i-1,2i}/\tau\right)}{\frac1N\sum_{i}^{N}\exp\left(s_{2i-1,2k}/\tau\right)}\right.$
$\begin{array}{l}{L_R=\operatorname{E}\left[-\|z_{2k-1}-z_{2k}\|_2^2\right]}\\{L=L_C+L_R}\end{array}$
update $f$ to minimize $L$
end for
|
[
[
0.118408203125,
0.0872802734375,
0.515625,
0.0872802734375,
0.515625,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.50537109375,
0.08489990234375,
0.84033203125,
0.0872802734375,
0.84033203125,
0.1009521484375,
0.50537109375,
0.09857177734375
],
[
0.118408203125,
0.10272216796875,
0.37451171875,
0.10272216796875,
0.37451171875,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.1376953125,
0.119384765625,
0.484375,
0.119384765625,
0.484375,
0.133056640625,
0.1376953125,
0.133056640625
],
[
0.51123046875,
0.11517333984375,
0.87890625,
0.11517333984375,
0.87890625,
0.12890625,
0.51123046875,
0.12890625
],
[
0.1175537109375,
0.1353759765625,
0.486083984375,
0.1341552734375,
0.486083984375,
0.1478271484375,
0.1175537109375,
0.1484375
],
[
0.5146484375,
0.133056640625,
0.87890625,
0.133056640625,
0.87890625,
0.1436767578125,
0.5146484375,
0.1436767578125
],
[
0.118408203125,
0.1519775390625,
0.486083984375,
0.1519775390625,
0.486083984375,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.51123046875,
0.146728515625,
0.880859375,
0.146728515625,
0.880859375,
0.1602783203125,
0.51123046875,
0.1602783203125
],
[
0.51123046875,
0.162109375,
0.87890625,
0.162109375,
0.87890625,
0.17578125,
0.51123046875,
0.17578125
],
[
0.12005615234375,
0.169189453125,
0.483642578125,
0.169189453125,
0.483642578125,
0.179931640625,
0.12005615234375,
0.179931640625
],
[
0.118408203125,
0.18408203125,
0.48193359375,
0.18408203125,
0.48193359375,
0.19482421875,
0.118408203125,
0.19482421875
],
[
0.51123046875,
0.179931640625,
0.87890625,
0.179931640625,
0.87890625,
0.193603515625,
0.51123046875,
0.193603515625
],
[
0.118408203125,
0.1995849609375,
0.48779296875,
0.1995849609375,
0.48779296875,
0.213134765625,
0.118408203125,
0.213134765625
],
[
0.51123046875,
0.19482421875,
0.87890625,
0.19482421875,
0.87890625,
0.2078857421875,
0.51123046875,
0.2078857421875
],
[
0.12005615234375,
0.21728515625,
0.4189453125,
0.21728515625,
0.4189453125,
0.22802734375,
0.12005615234375,
0.22802734375
],
[
0.5146484375,
0.2120361328125,
0.8818359375,
0.2120361328125,
0.8818359375,
0.2257080078125,
0.5146484375,
0.2257080078125
],
[
0.51318359375,
0.22802734375,
0.87890625,
0.22802734375,
0.87890625,
0.2410888671875,
0.51318359375,
0.2410888671875
],
[
0.1376953125,
0.24462890625,
0.48779296875,
0.24462890625,
0.48779296875,
0.25830078125,
0.1376953125,
0.25830078125
],
[
0.51318359375,
0.24462890625,
0.87890625,
0.24462890625,
0.87890625,
0.25830078125,
0.51318359375,
0.25830078125
],
[
0.151123046875,
0.260009765625,
0.48779296875,
0.26123046875,
0.48779296875,
0.27490234375,
0.151123046875,
0.273681640625
],
[
0.51318359375,
0.260009765625,
0.880859375,
0.260009765625,
0.880859375,
0.273681640625,
0.51318359375,
0.273681640625
],
[
0.154541015625,
0.275634765625,
0.489501953125,
0.275634765625,
0.489501953125,
0.289306640625,
0.154541015625,
0.289306640625
],
[
0.51318359375,
0.276611328125,
0.880859375,
0.276611328125,
0.880859375,
0.290283203125,
0.51318359375,
0.290283203125
],
[
0.154541015625,
0.292236328125,
0.483642578125,
0.292236328125,
0.483642578125,
0.305908203125,
0.154541015625,
0.305908203125
],
[
0.51318359375,
0.292236328125,
0.880859375,
0.292236328125,
0.880859375,
0.305908203125,
0.51318359375,
0.305908203125
],
[
0.51318359375,
0.30712890625,
0.880859375,
0.30810546875,
0.880859375,
0.322509765625,
0.51318359375,
0.32177734375
],
[
0.1376953125,
0.320556640625,
0.48779296875,
0.320556640625,
0.48779296875,
0.334228515625,
0.1376953125,
0.334228515625
],
[
0.51123046875,
0.32373046875,
0.87890625,
0.32373046875,
0.87890625,
0.33740234375,
0.51123046875,
0.33740234375
],
[
0.15283203125,
0.33740234375,
0.490234375,
0.33740234375,
0.490234375,
0.350830078125,
0.15283203125,
0.350830078125
],
[
0.51318359375,
0.34033203125,
0.880859375,
0.34033203125,
0.880859375,
0.35400390625,
0.51318359375,
0.35400390625
],
[
0.154541015625,
0.352783203125,
0.48779296875,
0.352783203125,
0.48779296875,
0.366455078125,
0.154541015625,
0.366455078125
],
[
0.509765625,
0.355712890625,
0.87890625,
0.355712890625,
0.87890625,
0.369384765625,
0.509765625,
0.369384765625
],
[
0.151123046875,
0.36767578125,
0.44580078125,
0.369384765625,
0.44580078125,
0.38427734375,
0.151123046875,
0.3818359375
],
[
0.51318359375,
0.37353515625,
0.8447265625,
0.37353515625,
0.8447265625,
0.38720703125,
0.51318359375,
0.38720703125
],
[
0.1351318359375,
0.3955078125,
0.486083984375,
0.396728515625,
0.486083984375,
0.410888671875,
0.1351318359375,
0.410400390625
],
[
0.15283203125,
0.413330078125,
0.496337890625,
0.413330078125,
0.496337890625,
0.427490234375,
0.15283203125,
0.427490234375
],
[
0.491943359375,
0.418701171875,
0.5087890625,
0.418701171875,
0.5087890625,
0.4228515625,
0.491943359375,
0.4228515625
],
[
0.509765625,
0.4150390625,
0.615234375,
0.4150390625,
0.615234375,
0.427001953125,
0.509765625,
0.427001953125
],
[
0.154541015625,
0.429931640625,
0.489501953125,
0.429931640625,
0.489501953125,
0.443603515625,
0.154541015625,
0.443603515625
],
[
0.15283203125,
0.4453125,
0.36865234375,
0.44482421875,
0.36865234375,
0.458984375,
0.15283203125,
0.460205078125
],
[
0.5146484375,
0.449462890625,
0.87890625,
0.449462890625,
0.87890625,
0.463134765625,
0.5146484375,
0.463134765625
],
[
0.5146484375,
0.465576171875,
0.87890625,
0.465576171875,
0.87890625,
0.478515625,
0.5146484375,
0.478515625
],
[
0.11419677734375,
0.486328125,
0.274658203125,
0.486328125,
0.274658203125,
0.50048828125,
0.11419677734375,
0.50048828125
],
[
0.515625,
0.48388671875,
0.87890625,
0.48388671875,
0.87890625,
0.49462890625,
0.515625,
0.49462890625
],
[
0.5146484375,
0.498779296875,
0.87890625,
0.498779296875,
0.87890625,
0.5126953125,
0.5146484375,
0.5126953125
],
[
0.1158447265625,
0.51123046875,
0.435791015625,
0.51318359375,
0.435791015625,
0.52783203125,
0.1158447265625,
0.525390625
],
[
0.51318359375,
0.51416015625,
0.880859375,
0.51416015625,
0.880859375,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.1175537109375,
0.53369140625,
0.484375,
0.533203125,
0.484375,
0.54736328125,
0.1175537109375,
0.548828125
],
[
0.51318359375,
0.53076171875,
0.880859375,
0.53076171875,
0.880859375,
0.54443359375,
0.51318359375,
0.54443359375
],
[
0.1175537109375,
0.5498046875,
0.483642578125,
0.5498046875,
0.483642578125,
0.5634765625,
0.1175537109375,
0.5634765625
],
[
0.51318359375,
0.54638671875,
0.8818359375,
0.54638671875,
0.8818359375,
0.56005859375,
0.51318359375,
0.56005859375
],
[
0.1175537109375,
0.56640625,
0.486083984375,
0.5654296875,
0.486083984375,
0.5791015625,
0.1175537109375,
0.580078125
],
[
0.51318359375,
0.5625,
0.880859375,
0.5625,
0.880859375,
0.576171875,
0.51318359375,
0.576171875
],
[
0.1175537109375,
0.58203125,
0.489501953125,
0.58056640625,
0.489501953125,
0.595703125,
0.1175537109375,
0.5966796875
],
[
0.5146484375,
0.5791015625,
0.85791015625,
0.5791015625,
0.85791015625,
0.5927734375,
0.5146484375,
0.5927734375
],
[
0.1175537109375,
0.5986328125,
0.484375,
0.5986328125,
0.484375,
0.6123046875,
0.1175537109375,
0.6123046875
],
[
0.118408203125,
0.615234375,
0.486083984375,
0.615234375,
0.486083984375,
0.62890625,
0.118408203125,
0.62890625
],
[
0.51318359375,
0.615234375,
0.880859375,
0.615234375,
0.880859375,
0.62890625,
0.51318359375,
0.62890625
],
[
0.12005615234375,
0.6328125,
0.486083984375,
0.6328125,
0.486083984375,
0.6435546875,
0.12005615234375,
0.6435546875
],
[
0.51123046875,
0.6318359375,
0.6513671875,
0.6318359375,
0.6513671875,
0.6455078125,
0.51123046875,
0.6455078125
],
[
0.118408203125,
0.64794921875,
0.429931640625,
0.64794921875,
0.429931640625,
0.6611328125,
0.118408203125,
0.6611328125
],
[
0.515625,
0.6494140625,
0.875,
0.6494140625,
0.875,
0.6630859375,
0.515625,
0.6630859375
],
[
0.1368408203125,
0.66455078125,
0.486083984375,
0.66455078125,
0.486083984375,
0.677734375,
0.1368408203125,
0.677734375
],
[
0.51123046875,
0.66015625,
0.7119140625,
0.662109375,
0.7119140625,
0.6767578125,
0.51123046875,
0.6748046875
],
[
0.52490234375,
0.67138671875,
0.71875,
0.6748046875,
0.71875,
0.69140625,
0.52490234375,
0.6884765625
],
[
0.118408203125,
0.6796875,
0.486083984375,
0.6796875,
0.486083984375,
0.693359375,
0.118408203125,
0.693359375
],
[
0.7119140625,
0.67578125,
0.748046875,
0.67578125,
0.748046875,
0.6884765625,
0.7119140625,
0.6884765625
],
[
0.5439453125,
0.68701171875,
0.7197265625,
0.68701171875,
0.7197265625,
0.70068359375,
0.5439453125,
0.70068359375
],
[
0.118408203125,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.5634765625,
0.69873046875,
0.85791015625,
0.69775390625,
0.85791015625,
0.71142578125,
0.5634765625,
0.7119140625
],
[
0.12005615234375,
0.7119140625,
0.483642578125,
0.7119140625,
0.483642578125,
0.7255859375,
0.12005615234375,
0.7255859375
],
[
0.56689453125,
0.71337890625,
0.66015625,
0.71337890625,
0.66015625,
0.7236328125,
0.56689453125,
0.7236328125
],
[
0.12005615234375,
0.72705078125,
0.483642578125,
0.72705078125,
0.483642578125,
0.74072265625,
0.12005615234375,
0.74072265625
],
[
0.56494140625,
0.7236328125,
0.68603515625,
0.7236328125,
0.68603515625,
0.73681640625,
0.56494140625,
0.73681640625
],
[
0.56103515625,
0.73095703125,
0.64990234375,
0.73095703125,
0.64990234375,
0.7490234375,
0.56103515625,
0.7490234375
],
[
0.118408203125,
0.74462890625,
0.484375,
0.74462890625,
0.484375,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.5634765625,
0.74755859375,
0.654296875,
0.7451171875,
0.654296875,
0.76025390625,
0.5634765625,
0.7626953125
],
[
0.118408203125,
0.76025390625,
0.484375,
0.76025390625,
0.484375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.54833984375,
0.76025390625,
0.59375,
0.76025390625,
0.59375,
0.77099609375,
0.54833984375,
0.77099609375
],
[
0.5458984375,
0.77099609375,
0.85546875,
0.77099609375,
0.85546875,
0.78369140625,
0.5458984375,
0.78369140625
],
[
0.118408203125,
0.77685546875,
0.484375,
0.77685546875,
0.484375,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.57666015625,
0.78369140625,
0.72314453125,
0.78369140625,
0.72314453125,
0.79736328125,
0.57666015625,
0.79736328125
],
[
0.118408203125,
0.7919921875,
0.483642578125,
0.7919921875,
0.483642578125,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.54736328125,
0.7958984375,
0.59375,
0.7958984375,
0.59375,
0.80712890625,
0.54736328125,
0.80712890625
],
[
0.12005615234375,
0.81005859375,
0.483642578125,
0.81005859375,
0.483642578125,
0.82080078125,
0.12005615234375,
0.82080078125
],
[
0.64208984375,
0.8125,
0.7197265625,
0.830078125,
0.7080078125,
0.85400390625,
0.63037109375,
0.8359375
],
[
0.66162109375,
0.81005859375,
0.71728515625,
0.822265625,
0.7080078125,
0.8388671875,
0.6533203125,
0.82666015625
],
[
0.7255859375,
0.8125,
0.82373046875,
0.8125,
0.82373046875,
0.82666015625,
0.7255859375,
0.82666015625
],
[
0.54736328125,
0.8193359375,
0.6337890625,
0.8193359375,
0.6337890625,
0.83056640625,
0.54736328125,
0.83056640625
],
[
0.66015625,
0.82177734375,
0.66748046875,
0.826171875,
0.66162109375,
0.83056640625,
0.654296875,
0.82666015625
],
[
0.7021484375,
0.82666015625,
0.84619140625,
0.82666015625,
0.84619140625,
0.84423828125,
0.7021484375,
0.84423828125
],
[
0.118408203125,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.118408203125,
0.82470703125,
0.483642578125,
0.82470703125,
0.483642578125,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.5458984375,
0.845703125,
0.72900390625,
0.845703125,
0.72900390625,
0.86328125,
0.5458984375,
0.86328125
],
[
0.1175537109375,
0.85693359375,
0.483642578125,
0.85693359375,
0.483642578125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.54736328125,
0.859375,
0.634765625,
0.85986328125,
0.634765625,
0.87158203125,
0.54736328125,
0.87060546875
],
[
0.118408203125,
0.87353515625,
0.486083984375,
0.87353515625,
0.486083984375,
0.88720703125,
0.118408203125,
0.88720703125
],
[
0.5458984375,
0.87060546875,
0.69189453125,
0.87060546875,
0.69189453125,
0.88427734375,
0.5458984375,
0.88427734375
],
[
0.1175537109375,
0.8896484375,
0.484375,
0.8896484375,
0.484375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.5283203125,
0.88427734375,
0.57421875,
0.88427734375,
0.57421875,
0.8955078125,
0.5283203125,
0.8955078125
],
[
0.118408203125,
0.90478515625,
0.28369140625,
0.90478515625,
0.28369140625,
0.91845703125,
0.118408203125,
0.91845703125
]
] |
[
"我们提出了一个重建任务，鼓励模型通过同一，",
"2.2Information Minimization Principle",
"对中的另一个重建正实例的表示",
"InforMin-cL的算法易于理解，只需几行",
"信息最小化原则（？）已被提出来保留必要的最",
"代码即可实现。此外，我们的方法不会改变主",
"小信息。近年来，研究人员利用信息最小化原",
"要的网络结构，因此它是模型不可知的，可以",
"理来改进图像表示（?？）。此外，信息瓶颈被",
"用来提高基于注意力的模型（？）的可解释性。",
"应用于任何基于对比学习的表示学习神经网",
"络。4节中的实验表明，InforMin-CL可以",
"Tian等人。（?）在对比表示学习框架中显示了",
"帮助模型学习有效的表示形式，从而提高下游",
"给定任务的良好视图，应保留与任务相关的信",
"任务性能。我们的主要贡献总结如下：",
"息，同时最大程度地减少无关的麻烦。然而，",
"它侧重于通过下游数据集消除与任务无关的信",
":We propose an InforMin-CL model to",
"息。Tsai等人。（?）专注于输人和自监督信号",
"learn good sentence embeddings by keeping",
"之间的多视图设置，并采用自监督信号来重建",
"useful information and getting rid of redun-",
"学习表示，以丢弃与任务无关的信息。我们的",
"dant information between positive instances.",
"工作在两个方面与（？）不同：1）我们丢弃文本",
"中的余信息，而（？）丢弃图像中存储的噪声",
": We achieve our model via simple contrast and",
"信息，并且在两项工作中使用了不同的自监督",
"reconstruction tasks and prove that the recon-",
"信号；2）（？）通过学习由监督任务评估的视觉",
"struction task can drop redundant information",
"特征来验证其方法，而句子嵌人学习不仅通过",
"by minimizing the information entropy.",
"监督任务而且通过无监督任务来评估嵌入",
": We achieve new state-of-the-art results on",
"seven supervised tasks and seven unsuper-",
"",
"3Method",
"vised tasks, which indicates the great advan-",
"tages of our proposed model.",
"我们提出了一种用于无监督句子表示学习的",
"InforMin-cL模型（图1）。有两个主要步",
"2.Related Work",
"骤，对比和重建。我们首先提出对比目标来学",
"习有用的信息：我们将正面实例和负面实例分",
"2.1Sentence Representation Learning",
"开，同时将正面实例拉在一起，这意味着最大",
"学习句子嵌人作为NLP中的一个重要问题已",
"化正面实例之间的互信息。稍后，我们提出重",
"经得到了广泛的研究。最近的工作重点是利用",
"建任务以丢弃无用的信息：在给定另一个正实",
"BERT（?）的功能来学习有效的句子嵌入，这些",
"例的情况下，我们最小化一个正实例的条件信",
"句子嵌入不受人为监督的信号的影响。BERT-",
"息摘。算法1提供了InforMin-CL的伪代码",
"flow（?）通过归一化流将各向异性句子嵌入分",
"布转换为光滑的各向同性高斯分布。BERT-白",
"Algorithm 1 Pseudocode of InforMin-CL in a",
"化（?）进一步提出了一种白化操作来增强句子",
"PyTorch-like style",
"嵌入的各向同性，并取得了更好的效果",
"Input: batch size N, temperature T , structure of f and .",
"然后，将对比学习方法应用于句子嵌入学",
"Output: encodernetwork f (),",
"for sampled minibatch ()A-",
"习。IS-BERT（?）提出了一个模型，该模型在",
"do",
"for all k e (1,..., N) do",
"BERT之上具有特征提取器，并且目标是最大",
"draw two augmentation functionst~I,t~1",
"化全局句子嵌入和局部句子嵌入之间的互信",
"2k-1=t(Tk)",
"息。CLEAR（?）采用多种句子级增强策略来学",
"22h-1 = f (F2k-1)",
"i2k=t (rk)",
"习句子表示。Coco-LM（?）采用辅助语言模型",
"22h =f (2k)",
"来破坏文本序列，在此基础上构造一个令牌级",
"end for",
"for all i e [1, ..., 2N) and j e{1, ..., 2N) do",
"任务和一个序列级任务，用于预训练主模型。",
"=(zl)",
"Gao等人。（?)提出了一种无监督的方法，可",
"end for",
"以预测带有丢失噪声的输人本身，以及一种利",
"210",
"Ylog",
"exp(s2i-1.2: /-)",
"Lc =maxE",
"",
"±≥, exp(s2i-1,2% / )",
"利用了通过退出获得的增强样本的自我对比，",
"用自然语言推理数据集的监督方法。SCD（?）",
"LR = E[-[22-1 - 22k/[2]",
"从而消除了对负对的依赖。然而，所有这些研",
"L=Lc+LI",
"究都没有考虑丢弃存储在视图中的穴余信息。",
"update f to minimize L",
"在我们的工作中，我们在信息最小化原则的框",
"endfor",
"架内考虑和解决问题"
] |
{
"arxiv_id": "2209.10951",
"language": "zh",
"page_idx": 2
}
|
||
2209.10951-zh-3
|
2209.10951-zh-3.jpg
|
<FigureHere>
Figure 1: The architecture of our proposed framework. Contrast: Negative pairs are pushed apart while positive pairs are pulled together, which suggests maximizing the mutual information between positive instances. Reconstruction: We drop redundant information in $z^1$ not shared with $z^2$ (marked by the green point) by reconstructing one positive instance via the other positive instance.
## 3.1 Contrast
小且足够的自监督表示：
$$\begin{aligned}&Z^{ssl}=\arg\max_{Z}I\left(Z;S\right)\\&Z^{ssl_{\mathrm{min}}}=\arg\min_{Z}H\left(Z|S\right)\\&s.t.\:I\left(Z;S\right)is\:\mathrm{maximized}\end{aligned}$$
(2)
然后，我们给出定理 1。有关证明，请参阅
(?)。
定理 1 监督学习表示包含输入中的所有关键信息(i。e。$I(X;T)$)。自监督学习表示包含输入中的所有关键信息，可能会丢失$\varepsilon:$
对比学习致力于通过将正样本对拉到一起并推开负样本对来学习有效的表示。我们建立在无监督的 SimCSE(?)最近的成功基础上，并将从具有独立不同的退出掩码的同一句子中得出的嵌入作为积极的实例。我们采用退出掩码(默认退出概率 p=0。1)作为增强技能，被证明优于其他技能(?),获得正对。正对采用相同的句子，并且它们的嵌人仅在退出掩码方面有所不同。同一小批量中的其他句子被视为否定实例
我们将输入，实例和自监督信号表示为 X, Z 和 S。我们通过应用不同的 dropout 掩码将相同的输入 x 馈送到编码器两次，然后获得正实例$z^1$和$z^2$ 。在我们的工作中，我们将正对中的一个实例$z^2$作为自监督信号。下游任务所需的信息称为“关键信息”:T。$I$和$H$表示互信息和信息熵
令$Z^{sup}$为充分监督表示$,Z^{sup_\mathrm{min}}$为最小和
充分监督表示：
(3)
$$\begin{aligned}I\left(X;T\right)&=I\left(Z^{\mathrm{sup}};T\right)=I\left(Z^{\mathrm{sup}_{\mathrm{min}}};T\right)\\&\geq I\left(Z^{ssl};T\right)\\&\geq I\left(Z^{ssl_{\mathrm{min}}};T\right)\\&\geq I\left(X;T\right)-\varepsilon\end{aligned}$$
$$\begin{aligned}&Z^{\mathrm{sup}}=\arg\max_{Z}I\left(Z;T\right)\\&Z^{sup_{\mathrm{min}}}=\arg\min_{Z}H\left(Z|T\right)\\\end{aligned}$$
(1)
对比学习目标最大化了正实例$z^{1}$和自监督信号$z^2$之间的依赖性，这表明最大化了互信息$I\left(z^1;z^2\right)$。定理 1 表明，最大化$I\left(z^1;z^2\right)$ 会导致$z_1$包含来自输人 x 的下游任务所需的几乎所有信息。请注意，T 仅用于描述我们的方法，实际上，在预训练阶段不使用下游数据集
我们使用类似于(?)中的对比学习目标，它
令$Z^{ssl}$为足够的自监督表示$,Z^ssl_{\mathrm{min}}$为最
|
[
[
0.438232421875,
0.08074951171875,
0.5556640625,
0.08074951171875,
0.5556640625,
0.09442138671875,
0.438232421875,
0.09442138671875
],
[
0.64990234375,
0.0831298828125,
0.83642578125,
0.0831298828125,
0.83642578125,
0.09320068359375,
0.64990234375,
0.09320068359375
],
[
0.22412109375,
0.1181640625,
0.27783203125,
0.1181640625,
0.27783203125,
0.122314453125,
0.22412109375,
0.122314453125
],
[
0.476806640625,
0.11639404296875,
0.49365234375,
0.11639404296875,
0.49365234375,
0.12054443359375,
0.476806640625,
0.12054443359375
],
[
0.483642578125,
0.12054443359375,
0.49365234375,
0.12054443359375,
0.49365234375,
0.12469482421875,
0.483642578125,
0.12469482421875
],
[
0.66162109375,
0.1312255859375,
0.8310546875,
0.1312255859375,
0.8310546875,
0.144287109375,
0.66162109375,
0.144287109375
],
[
0.14697265625,
0.1436767578125,
0.306396484375,
0.1436767578125,
0.306396484375,
0.1478271484375,
0.14697265625,
0.1478271484375
],
[
0.474365234375,
0.1484375,
0.5029296875,
0.1484375,
0.5029296875,
0.1519775390625,
0.474365234375,
0.1519775390625
],
[
0.4609375,
0.160888671875,
0.5615234375,
0.160888671875,
0.5615234375,
0.168701171875,
0.4609375,
0.168701171875
],
[
0.372802734375,
0.17578125,
0.391357421875,
0.17578125,
0.391357421875,
0.193603515625,
0.372802734375,
0.193603515625
],
[
0.66162109375,
0.1787109375,
0.67236328125,
0.1787109375,
0.67236328125,
0.182861328125,
0.66162109375,
0.182861328125
],
[
0.372802734375,
0.189453125,
0.391357421875,
0.189453125,
0.391357421875,
0.2037353515625,
0.372802734375,
0.2037353515625
],
[
0.453369140625,
0.1953125,
0.51904296875,
0.1953125,
0.51904296875,
0.21435546875,
0.453369140625,
0.21435546875
],
[
0.67529296875,
0.19482421875,
0.81787109375,
0.19482421875,
0.81787109375,
0.2078857421875,
0.67529296875,
0.2078857421875
],
[
0.3720703125,
0.201904296875,
0.39208984375,
0.201904296875,
0.39208984375,
0.236328125,
0.3720703125,
0.236328125
],
[
0.73486328125,
0.208984375,
0.81005859375,
0.215576171875,
0.8076171875,
0.232177734375,
0.7314453125,
0.2257080078125
],
[
0.16796875,
0.198974609375,
0.301513671875,
0.1995849609375,
0.301513671875,
0.21142578125,
0.16796875,
0.210205078125
],
[
0.444091796875,
0.21728515625,
0.56689453125,
0.21728515625,
0.56689453125,
0.22802734375,
0.444091796875,
0.22802734375
],
[
0.66162109375,
0.2161865234375,
0.72900390625,
0.213134765625,
0.73046875,
0.22802734375,
0.6630859375,
0.23095703125
],
[
0.1435546875,
0.2476806640625,
0.30322265625,
0.2476806640625,
0.30322265625,
0.25830078125,
0.1435546875,
0.25830078125
],
[
0.453369140625,
0.2423095703125,
0.52880859375,
0.2423095703125,
0.52880859375,
0.26123046875,
0.453369140625,
0.26123046875
],
[
0.44580078125,
0.2666015625,
0.568359375,
0.2666015625,
0.568359375,
0.276611328125,
0.44580078125,
0.276611328125
],
[
0.63623046875,
0.26953125,
0.64208984375,
0.26953125,
0.64208984375,
0.27490234375,
0.63623046875,
0.27490234375
],
[
0.6591796875,
0.27490234375,
0.67919921875,
0.27490234375,
0.67919921875,
0.281982421875,
0.6591796875,
0.281982421875
],
[
0.468505859375,
0.29443359375,
0.53662109375,
0.29443359375,
0.53662109375,
0.305908203125,
0.468505859375,
0.305908203125
],
[
0.62451171875,
0.29150390625,
0.8193359375,
0.29150390625,
0.8193359375,
0.3046875,
0.62451171875,
0.3046875
],
[
0.298828125,
0.2998046875,
0.37451171875,
0.3017578125,
0.37451171875,
0.31640625,
0.298828125,
0.314208984375
],
[
0.63232421875,
0.3046875,
0.86962890625,
0.3046875,
0.86962890625,
0.318359375,
0.63232421875,
0.318359375
],
[
0.1502685546875,
0.318359375,
0.17626953125,
0.318359375,
0.17626953125,
0.32666015625,
0.1502685546875,
0.32666015625
],
[
0.173828125,
0.318359375,
0.2149658203125,
0.318359375,
0.2149658203125,
0.325927734375,
0.173828125,
0.325927734375
],
[
0.330078125,
0.32373046875,
0.3837890625,
0.32373046875,
0.3837890625,
0.334228515625,
0.330078125,
0.334228515625
],
[
0.405517578125,
0.32470703125,
0.60107421875,
0.32470703125,
0.60107421875,
0.3349609375,
0.405517578125,
0.3349609375
],
[
0.64990234375,
0.320556640625,
0.8017578125,
0.320556640625,
0.8017578125,
0.334228515625,
0.64990234375,
0.334228515625
],
[
0.64990234375,
0.333251953125,
0.8759765625,
0.333251953125,
0.8759765625,
0.3466796875,
0.64990234375,
0.3466796875
],
[
0.6513671875,
0.345703125,
0.8486328125,
0.345703125,
0.8486328125,
0.355712890625,
0.6513671875,
0.355712890625
],
[
0.1175537109375,
0.377685546875,
0.8818359375,
0.377685546875,
0.8818359375,
0.391357421875,
0.1175537109375,
0.391357421875
],
[
0.1175537109375,
0.39013671875,
0.880859375,
0.39013671875,
0.880859375,
0.405029296875,
0.1175537109375,
0.405029296875
],
[
0.1175537109375,
0.40625,
0.8818359375,
0.40625,
0.8818359375,
0.419189453125,
0.1175537109375,
0.419189453125
],
[
0.1158447265625,
0.42041015625,
0.468505859375,
0.42041015625,
0.468505859375,
0.43408203125,
0.1158447265625,
0.43408203125
],
[
0.1175537109375,
0.46142578125,
0.1435546875,
0.46142578125,
0.1435546875,
0.47265625,
0.1175537109375,
0.47265625
],
[
0.1502685546875,
0.46142578125,
0.235107421875,
0.46142578125,
0.235107421875,
0.47265625,
0.1502685546875,
0.47265625
],
[
0.5146484375,
0.460205078125,
0.705078125,
0.460205078125,
0.705078125,
0.473876953125,
0.5146484375,
0.473876953125
],
[
0.1175537109375,
0.48388671875,
0.483642578125,
0.48388671875,
0.483642578125,
0.49755859375,
0.1175537109375,
0.49755859375
],
[
0.591796875,
0.493408203125,
0.62646484375,
0.48681640625,
0.63037109375,
0.498779296875,
0.59619140625,
0.5048828125
],
[
0.62890625,
0.4921875,
0.77734375,
0.4912109375,
0.77734375,
0.505859375,
0.62890625,
0.50732421875
],
[
0.1175537109375,
0.50048828125,
0.483642578125,
0.50048828125,
0.483642578125,
0.51416015625,
0.1175537109375,
0.51416015625
],
[
0.67529296875,
0.50732421875,
0.68994140625,
0.50732421875,
0.68994140625,
0.515625,
0.67529296875,
0.515625
],
[
0.118408203125,
0.51708984375,
0.484375,
0.51708984375,
0.484375,
0.53076171875,
0.118408203125,
0.53076171875
],
[
0.591796875,
0.51953125,
0.64990234375,
0.5166015625,
0.6513671875,
0.53076171875,
0.59375,
0.53369140625
],
[
0.6484375,
0.521484375,
0.80029296875,
0.521484375,
0.80029296875,
0.53515625,
0.6484375,
0.53515625
],
[
0.8564453125,
0.521484375,
0.8818359375,
0.521484375,
0.8818359375,
0.53515625,
0.8564453125,
0.53515625
],
[
0.118408203125,
0.533203125,
0.484375,
0.533203125,
0.484375,
0.54638671875,
0.118408203125,
0.54638671875
],
[
0.7001953125,
0.533203125,
0.7109375,
0.54052734375,
0.7021484375,
0.5458984375,
0.69287109375,
0.5390625
],
[
0.118408203125,
0.548828125,
0.484375,
0.548828125,
0.484375,
0.5625,
0.118408203125,
0.5625
],
[
0.59375,
0.5517578125,
0.796875,
0.5517578125,
0.796875,
0.5625,
0.59375,
0.5625
],
[
0.1175537109375,
0.56396484375,
0.484375,
0.5634765625,
0.484375,
0.57763671875,
0.1175537109375,
0.5791015625
],
[
0.12005615234375,
0.58056640625,
0.484375,
0.58056640625,
0.484375,
0.59423828125,
0.12005615234375,
0.59423828125
],
[
0.53076171875,
0.58447265625,
0.8818359375,
0.58447265625,
0.8818359375,
0.59716796875,
0.53076171875,
0.59716796875
],
[
0.1175537109375,
0.595703125,
0.484375,
0.5966796875,
0.484375,
0.61083984375,
0.1175537109375,
0.60986328125
],
[
0.51318359375,
0.60107421875,
0.54833984375,
0.60107421875,
0.54833984375,
0.61328125,
0.51318359375,
0.61328125
],
[
0.1175537109375,
0.6123046875,
0.484375,
0.6123046875,
0.484375,
0.626953125,
0.1175537109375,
0.626953125
],
[
0.51318359375,
0.6162109375,
0.880859375,
0.6162109375,
0.880859375,
0.6298828125,
0.51318359375,
0.6298828125
],
[
0.1175537109375,
0.62890625,
0.156982421875,
0.62890625,
0.156982421875,
0.64453125,
0.1175537109375,
0.64453125
],
[
0.51318359375,
0.63134765625,
0.87890625,
0.63134765625,
0.87890625,
0.64453125,
0.51318359375,
0.64453125
],
[
0.1351318359375,
0.64794921875,
0.483642578125,
0.64794921875,
0.483642578125,
0.6611328125,
0.1351318359375,
0.6611328125
],
[
0.51318359375,
0.64794921875,
0.796875,
0.64794921875,
0.796875,
0.662109375,
0.51318359375,
0.662109375
],
[
0.1158447265625,
0.662109375,
0.484375,
0.6630859375,
0.484375,
0.677734375,
0.1158447265625,
0.6767578125
],
[
0.118408203125,
0.6796875,
0.484375,
0.6796875,
0.484375,
0.693359375,
0.118408203125,
0.693359375
],
[
0.53564453125,
0.6787109375,
0.8388671875,
0.6787109375,
0.8388671875,
0.693359375,
0.53564453125,
0.693359375
],
[
0.12005615234375,
0.69482421875,
0.483642578125,
0.69482421875,
0.483642578125,
0.708984375,
0.12005615234375,
0.708984375
],
[
0.603515625,
0.703125,
0.705078125,
0.70068359375,
0.705078125,
0.71630859375,
0.603515625,
0.71875
],
[
0.118408203125,
0.71044921875,
0.484375,
0.71044921875,
0.484375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.6455078125,
0.7060546875,
0.65576171875,
0.7060546875,
0.65576171875,
0.71044921875,
0.6455078125,
0.71044921875
],
[
0.8564453125,
0.71875,
0.88330078125,
0.71875,
0.88330078125,
0.732421875,
0.8564453125,
0.732421875
],
[
0.118408203125,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.60791015625,
0.73291015625,
0.6396484375,
0.73291015625,
0.6396484375,
0.74462890625,
0.60791015625,
0.74462890625
],
[
0.6689453125,
0.73095703125,
0.7080078125,
0.73291015625,
0.7080078125,
0.74072265625,
0.6689453125,
0.73828125
],
[
0.118408203125,
0.74365234375,
0.2476806640625,
0.74365234375,
0.2476806640625,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.1368408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77490234375,
0.1368408203125,
0.77490234375
],
[
0.603515625,
0.755859375,
0.72119140625,
0.755859375,
0.72119140625,
0.77099609375,
0.603515625,
0.77099609375
],
[
0.118408203125,
0.77783203125,
0.2376708984375,
0.77783203125,
0.2376708984375,
0.79150390625,
0.118408203125,
0.79150390625
],
[
0.5322265625,
0.79052734375,
0.880859375,
0.79052734375,
0.880859375,
0.80419921875,
0.5322265625,
0.80419921875
],
[
0.19140625,
0.8056640625,
0.38623046875,
0.80859375,
0.38623046875,
0.82763671875,
0.19140625,
0.82470703125
],
[
0.51318359375,
0.8056640625,
0.87890625,
0.8056640625,
0.87890625,
0.82080078125,
0.51318359375,
0.82080078125
],
[
0.2822265625,
0.826171875,
0.295654296875,
0.826171875,
0.295654296875,
0.83447265625,
0.2822265625,
0.83447265625
],
[
0.5146484375,
0.822265625,
0.87890625,
0.822265625,
0.87890625,
0.83740234375,
0.5146484375,
0.83740234375
],
[
0.19482421875,
0.8388671875,
0.255126953125,
0.8359375,
0.2568359375,
0.84716796875,
0.1956787109375,
0.85107421875
],
[
0.25439453125,
0.8388671875,
0.4072265625,
0.83837890625,
0.4072265625,
0.8525390625,
0.25439453125,
0.85400390625
],
[
0.462646484375,
0.84033203125,
0.491943359375,
0.84033203125,
0.491943359375,
0.8515625,
0.462646484375,
0.8515625
],
[
0.51123046875,
0.8388671875,
0.87744140625,
0.8388671875,
0.87744140625,
0.8525390625,
0.51123046875,
0.8525390625
],
[
0.307373046875,
0.8486328125,
0.31982421875,
0.85693359375,
0.307373046875,
0.865234375,
0.295654296875,
0.85693359375
],
[
0.5146484375,
0.85546875,
0.880859375,
0.85546875,
0.880859375,
0.869140625,
0.5146484375,
0.869140625
],
[
0.19482421875,
0.86767578125,
0.403076171875,
0.86767578125,
0.403076171875,
0.88134765625,
0.19482421875,
0.88134765625
],
[
0.5146484375,
0.87158203125,
0.8818359375,
0.87158203125,
0.8818359375,
0.88525390625,
0.5146484375,
0.88525390625
],
[
0.5146484375,
0.88720703125,
0.5341796875,
0.88720703125,
0.5341796875,
0.90087890625,
0.5146484375,
0.90087890625
],
[
0.1368408203125,
0.90185546875,
0.48779296875,
0.90185546875,
0.48779296875,
0.91943359375,
0.1368408203125,
0.91943359375
],
[
0.52880859375,
0.90380859375,
0.880859375,
0.90380859375,
0.880859375,
0.91845703125,
0.52880859375,
0.91845703125
]
] |
[
"Reconstruction",
"DetailsofReconstruction",
"stonetap",
"",
"",
"Share some information",
"easeGecKeOvesgD",
"",
"ositiveinstance",
"3",
"",
"0",
"8",
"(Shared information)",
"DEL",
"1121-2212",
"What is party all night",
"Negative instance",
"LR = E[-",
"Play somemusicusing sla",
"8",
"Negativeinstance",
"Z",
"二",
"Contrast",
"(\"InfoMin Principle\":Discard",
"sim(zi.zf)/,",
"redundant information not shared)",
"Lc=",
"maxE",
"sim(ziz)",
"Mutual InformationMaximization",
"Key information shared",
"Redundant information not shared",
"Redundant information shared",
"Figure 1: The architecture of our proposed framework. Contrast: Negative pairs are pushed apart while positive",
"pairs are pulled together, which suggests maximizing the mutual information between positive instances. Recon",
"struction: We drop redundant information in z not shared with z2 (marked by the green point) by reconstructing",
"one positive instance via the other positive instance.",
"3.1",
"Contrast",
"小且足够的自监督表示：",
"对比学习致力于通过将正样本对拉到一起并",
"Zssl",
"= arg max I (Z; S)",
"推开负样本对来学习有效的表示。我们建立在",
"Z",
"无监督的SimCSE（?）最近的成功基础上，并",
"Zsslmin",
"= argmin H (Z|S)",
"(2)",
"将从具有独立不同的退出掩码的同一句子中得",
"",
"出的嵌入作为积极的实例。我们采用退出掩码",
"s.t.I (Z;S)ismaximized",
"(默认退出概率p=0。1)作为增强技能，被证明",
"优于其他技能（？），获得正对。正对采用相同",
"然后，我们给出定理1。有关证明，请参阅",
"的句子，并且它们的嵌入仅在退出掩码方面有",
"(?)。",
"所不同。同一小批量中的其他句子被视为否定",
"定理1监督学习表示包含输人中的所有关键信",
"实例",
"息（i。e。I（X：T））。自监督学习表示包含输入",
"我们将输入，实例和自监督信号表示为X",
"中的所有关键信息，可能会丢失e：",
"Z和S。我们通过应用不同的dropout掩码将相",
"同的输入x馈送到编码器两次，然后获得正实",
"I(X; T) = I (ZsuP; T) = I(ZsuPmin; T)",
"例z1和22。在我们的工作中，我们将正对中",
"≥ I(Zssl;T",
"的一个实例22作为自监督信号。下游任务所",
"F",
"(3)",
"需的信息称为“关键信息”：T。I和H表示",
">I",
"slmin.",
"互信息和信息摘",
"令Zsup为充分监督表示，Zsupmin为最小和",
"≥ I(X;T) -E",
"充分监督表示：",
"对比学习目标最大化了正实例z1和自监督",
"Zsup = arg max I (Z; T)",
"信号22之间的依赖性，这表明最大化了互信",
"Z",
"息I（zl;22）。定理1表明，最大化I（2l;z2)",
"Zsupmin",
"= arg min H (Z|T)",
"(1)",
"会导致21包含来自输人x的下游任务所需的",
"A",
"几乎所有信息。请注意，T仅用于描述我们的",
"s.t. I (Z;T) is maximized",
"方法，实际上，在预训练阶段不使用下游数据",
"集",
"令Zssl为足够的自监督表示，Zsslmin为最",
"我们使用类似于（？）中的对比学习目标，它"
] |
{
"arxiv_id": "2209.10951",
"language": "zh",
"page_idx": 3
}
|
||
2209.10951-zh-4
|
2209.10951-zh-4.jpg
|
是具有低方差的互信息下限：
(4)
其中$\left ( z_1^1, z_1^2\right ) , \ldots , \left ( z_N^1, z_N^2\right ) \sim P^N\left ( Z^1, Z^2\right ) $, 估。在$\left(??\right)$之后，我们将无监督任务作为句子
$z_i^1,z_i^2$是$i$的两个正实例。N 指批量大小，P
指$(Z^1,Z^2)$的统计分布
# 3.2 Reconstruction
重建的细节如图 1 的右侧所示。正实例$z^{1}$和$z^2$包含独立信息，同时共享一些信息。$z^1$中的噪声信息(标记为绿点)预计将被重建任务丢弃。我们证明了该任务可以通过最小化信息熵来丢弃无用的信息。重建任务鼓励自监督信号$z^2$重建学习表示$z^1$,这建议最大化对数条件似然$\mathbb{E}_{P_{Z^1,Z^2}}\left[\log P\left(Z^1|Z^2\right)\right]$。我们知道
$$-H\left(Z^1|Z^2\right)=\mathbb{E}_{P_{Z^1,Z^2}}\begin{bmatrix}\log P\left(Z^1|Z^2\right)\end{bmatrix}$$
$\begin{array}{c|c}&&(5)\end{array}$
因此，此重建也意味着最小化$H\left(Z^1|Z^2\right)$ 。定理 2 足够的自监督表示在输入中包含比足够和最小的自监督表示更多的冗余信息。后者包含大量的信息，$I\left(X;S|T\right)$,不能从输入丢弃：
$$\begin{aligned}I\left(Z^{ssl};X|T\right)&=I\left(X;S|T\right)+I\left(Z^{ssl};X|S,T\right)\\&\geq I\left(Z^{ssl_{\min}};X|T\right)=I\left(X;S|T\right)\\&\geq I\left(Z^{\mathrm{sup}_{\mathrm{min}}};X|T\right)=0\end{aligned}$$
定理 2(请参考(?)进行证明)表示$Z^{ssl}$包
含两部分冗余信息，而$Z^{ssl_\mathrm{min}}$包含一部分冗
余信息，丢弃$I\left(Z^{ssl};X|S,T\right)$
因此，如果$z^2$可以在$I\left(z^1;z^2\right)$最大化的约
束下完美地重构任何
$$\begin{pmatrix}z^1,z^2\end{pmatrix}\sim P_{Z^1,Z^2}$$
(7)
的$z^1$,则根据等式 2 得到$z^1_{ssl_{\mathrm{min}}}$。然后$z^1$丢弃冗余信息，不包括$I\left(z^1;z^2|t\right)$(i。e。,两个正实例$z^1$和$z^2$之间的共享信息中的冗余信息量)。为了更容易优化，我们使用$\mathbb{E}_P_{Z^1,Z^2}\left[\log Q_\Phi\left(Z^1|Z^2\right)\right]$作为$\mathbb{E}_{P_{Z^1,Z^2}}\left[\log P\left(Z^1|Z^2\right)\right]$的下界。在我们的部署中，我们利用方程 4 中的设计，让$Q_\Phi\left(Z^1|Z^2\right)$ 是高斯$N\left(Z^1|Z^2,\sigma\mathrm{I}\right),\sigma$I 作为对角矩阵。因此，我们得到的重建目标如下：
(8)
$$\mathcal{L}_R=\mathbb{E}_{z^1,z^2\sim P_{Z^1,Z^2}}\left[-\left\|z^1-z^2\right\|_2^2\right]$$
我们将两个目标合并为一个总目标：
$$\mathcal{L}=\mathcal{L}_C+\lambda*\mathcal{L}_R$$
(9)
,其中$\lambda$是一个超参数。具有总损失的训练模
型使我们能够丢弃视图中的冗余信息
### 4 $\textbf{Experiment}$
$\mathcal{L} _C= \max \mathbb{E} \left [ \frac 1N\sum _{i= 1}^N\log \frac {e^{sim\left ( z_i^1, z_i^2\right ) }/ \tau }{\frac 1N\sum _{k= 1}^Ne^{sim\left ( z_i^1, z_k^2\right ) }/ \tau }\right ] _{\begin{array} { c } 4. 1& \text{Evaluation Setup}\\ \text{我 们 对 七 个 标 准 监 督 任 务 和 七 个 无 监 督 任 务 进 }\\ \text{行 实 验 。我 们 使 用 SentEval 工 具 包 ( ? ) 进 行 评 }\end{array} }$
行实验。我们使用 SentEval 工具包(?)进行评
嵌入方法和监督结果的主要比较，以供参考
无监督任务我们评估七个语义文本相似性(STS) 任务的表示：STS2012-2016 (?????) , STSBenchmark(?)和 SICK 相关性(?),并计算句子嵌人之间的余弦相似性。所有无监督实验都是完全无监督的，这意味着不使用 STS 训练数据集，并且所有嵌入在训练后都是固定的。为了具有可比性，我们遵循(?)的评估协议，在所有主题子集上采用 Spearman 秩相关和聚合
受监督的任务我们评估七个受监督任务的表示：MR(?)、CR(?)、SUBJ(?)、MPQA(?) 、 SST- 2 ( ? ) , TREC ( ? ) 和 MRPC ( ? ) 。在 由 不同方法产生的(冻结)句子嵌入的顶部训练逻辑回归分类器。我们遵循 SentEval 的默认配置，并使用准确性作为指标
培训详细信息我们从预训练的 BERT(?) (无套管)或 RoBERTa(?)(套管)开始。与(?)类似，我们以无监督的方式在来自英语维基百科的 10$^{6}$随机抽样句子上训练我们的InforMin-CL 。在训练过程中，我们在[CLS] 表示的顶部添加一个 MLP 层作为句子嵌人，并在测试时直接将[CLS] 表示作为句子嵌入。掩码语言建模 (MLM)目标 (?)作为可选辅助损失添加到方程 9$:\mathcal{L}+\beta*\mathcal{L}_{MLM}(\beta$是超参数)。对于所有结果，我们使用以下超参数：epoch:1, 温度$\tau:0$。05,优化器：Adam (?))。我们在 STS- $\operatorname{B}$开发集上对批量大小$\in\{64,128,256\}$ 和学习速率$\in\{1e-5,3e-5,5e-5\}$ 进行网格搜索。在训练过程中，我们将得分最高的检查点保存在 STS-B 开发集上，以找到最佳的超参数。我们采用表 3 中列出的超参数设置。对于所有结果，我们使用带有 GeForceRTX3090GPU 的 PC (CUDA11, PyTorch1。7。1)
## 4.2 Main Results
基线我们将 InforMin-CL 与以前的典型句子嵌入方法进行比较，其中包括平均 GloVe 嵌入(?),跳过思想(?)和平均 BERT 或 RoBERTa 嵌入。我们还比较了后处理方法和使用对比目标的方法。后处理方法包括 BERT-flow(?)和BERT-白化(?)。使用对比物镜的方法包括 ISBERT(?)、DeCLUTR(?)、CT(?)、SimCSE (?)和 SCD(?)。IS-BERT 最大化全局和局部特征之间的一致性。DeCLUTR 从同一文档中获取不同的跨度作为正对。CT 对齐来自两个
|
[
[
0.118408203125,
0.0872802734375,
0.344970703125,
0.0872802734375,
0.344970703125,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.5087890625,
0.0831298828125,
0.6533203125,
0.08489990234375,
0.6533203125,
0.10272216796875,
0.5087890625,
0.1009521484375
],
[
0.2568359375,
0.10687255859375,
0.279541015625,
0.10687255859375,
0.279541015625,
0.119384765625,
0.2568359375,
0.119384765625
],
[
0.51123046875,
0.10986328125,
0.5634765625,
0.10986328125,
0.5634765625,
0.12469482421875,
0.51123046875,
0.12469482421875
],
[
0.55517578125,
0.10986328125,
0.6962890625,
0.11224365234375,
0.6962890625,
0.1270751953125,
0.55517578125,
0.12469482421875
],
[
0.2359619140625,
0.11639404296875,
0.258544921875,
0.11639404296875,
0.258544921875,
0.1270751953125,
0.2359619140625,
0.1270751953125
],
[
0.3525390625,
0.11102294921875,
0.411376953125,
0.10687255859375,
0.4130859375,
0.1181640625,
0.353515625,
0.122314453125
],
[
0.12005615234375,
0.12353515625,
0.219970703125,
0.12353515625,
0.219970703125,
0.1353759765625,
0.12005615234375,
0.1353759765625
],
[
0.252685546875,
0.1181640625,
0.318115234375,
0.12353515625,
0.31494140625,
0.13720703125,
0.2493896484375,
0.1318359375
],
[
0.235107421875,
0.1300048828125,
0.264404296875,
0.1300048828125,
0.264404296875,
0.142578125,
0.235107421875,
0.142578125
],
[
0.307373046875,
0.1312255859375,
0.48779296875,
0.1258544921875,
0.489501953125,
0.15087890625,
0.30908203125,
0.1561279296875
],
[
0.252685546875,
0.1383056640625,
0.287109375,
0.1383056640625,
0.287109375,
0.149658203125,
0.252685546875,
0.149658203125
],
[
0.5146484375,
0.133056640625,
0.880859375,
0.133056640625,
0.880859375,
0.146728515625,
0.5146484375,
0.146728515625
],
[
0.225830078125,
0.144287109375,
0.255126953125,
0.141357421875,
0.2568359375,
0.1484375,
0.2275390625,
0.1519775390625
],
[
0.280517578125,
0.1395263671875,
0.318115234375,
0.1395263671875,
0.318115234375,
0.15087890625,
0.280517578125,
0.15087890625
],
[
0.46337890625,
0.152587890625,
0.496337890625,
0.152587890625,
0.496337890625,
0.1650390625,
0.46337890625,
0.1650390625
],
[
0.51318359375,
0.1484375,
0.87890625,
0.1484375,
0.87890625,
0.162109375,
0.51318359375,
0.162109375
],
[
0.1158447265625,
0.166259765625,
0.476806640625,
0.1650390625,
0.476806640625,
0.182861328125,
0.1158447265625,
0.18408203125
],
[
0.51318359375,
0.1650390625,
0.87744140625,
0.1650390625,
0.87744140625,
0.17578125,
0.51318359375,
0.17578125
],
[
0.474365234375,
0.17333984375,
0.486083984375,
0.17333984375,
0.486083984375,
0.1776123046875,
0.474365234375,
0.1776123046875
],
[
0.1175537109375,
0.18408203125,
0.48779296875,
0.182861328125,
0.48779296875,
0.19775390625,
0.1175537109375,
0.198974609375
],
[
0.5146484375,
0.182861328125,
0.85986328125,
0.182861328125,
0.85986328125,
0.193603515625,
0.5146484375,
0.193603515625
],
[
0.118408203125,
0.20068359375,
0.306396484375,
0.20068359375,
0.306396484375,
0.21435546875,
0.118408203125,
0.21435546875
],
[
0.5322265625,
0.19775390625,
0.8818359375,
0.19775390625,
0.8818359375,
0.21142578125,
0.5322265625,
0.21142578125
],
[
0.51318359375,
0.213134765625,
0.86572265625,
0.213134765625,
0.86572265625,
0.226806640625,
0.51318359375,
0.226806640625
],
[
0.1175537109375,
0.2257080078125,
0.28369140625,
0.2257080078125,
0.28369140625,
0.2392578125,
0.1175537109375,
0.2392578125
],
[
0.51318359375,
0.2286376953125,
0.87890625,
0.2286376953125,
0.87890625,
0.2423095703125,
0.51318359375,
0.2423095703125
],
[
0.1175537109375,
0.2464599609375,
0.486083984375,
0.2452392578125,
0.486083984375,
0.260009765625,
0.1175537109375,
0.26123046875
],
[
0.51318359375,
0.2452392578125,
0.880859375,
0.2452392578125,
0.880859375,
0.2587890625,
0.51318359375,
0.2587890625
],
[
0.118408203125,
0.262451171875,
0.486083984375,
0.262451171875,
0.486083984375,
0.276611328125,
0.118408203125,
0.276611328125
],
[
0.51318359375,
0.26123046875,
0.880859375,
0.26123046875,
0.880859375,
0.27490234375,
0.51318359375,
0.27490234375
],
[
0.118408203125,
0.279052734375,
0.484375,
0.279052734375,
0.484375,
0.292236328125,
0.118408203125,
0.292236328125
],
[
0.51318359375,
0.276611328125,
0.880859375,
0.276611328125,
0.880859375,
0.290283203125,
0.51318359375,
0.290283203125
],
[
0.1175537109375,
0.29345703125,
0.484375,
0.29345703125,
0.484375,
0.30712890625,
0.1175537109375,
0.30712890625
],
[
0.51318359375,
0.29345703125,
0.880859375,
0.29345703125,
0.880859375,
0.30712890625,
0.51318359375,
0.30712890625
],
[
0.118408203125,
0.312255859375,
0.484375,
0.312255859375,
0.484375,
0.325927734375,
0.118408203125,
0.325927734375
],
[
0.51318359375,
0.31005859375,
0.8818359375,
0.31005859375,
0.8818359375,
0.32373046875,
0.51318359375,
0.32373046875
],
[
0.1175537109375,
0.32666015625,
0.483642578125,
0.32666015625,
0.483642578125,
0.34033203125,
0.1175537109375,
0.34033203125
],
[
0.51123046875,
0.325927734375,
0.5693359375,
0.325927734375,
0.5693359375,
0.34033203125,
0.51123046875,
0.34033203125
],
[
0.1158447265625,
0.341552734375,
0.443359375,
0.339111328125,
0.443359375,
0.35986328125,
0.1158447265625,
0.3623046875
],
[
0.53076171875,
0.342529296875,
0.880859375,
0.342529296875,
0.880859375,
0.355712890625,
0.53076171875,
0.355712890625
],
[
0.51123046875,
0.35693359375,
0.87890625,
0.35693359375,
0.87890625,
0.37060546875,
0.51123046875,
0.37060546875
],
[
0.1368408203125,
0.3681640625,
0.486083984375,
0.3681640625,
0.486083984375,
0.385986328125,
0.1368408203125,
0.385986328125
],
[
0.5224609375,
0.377685546875,
0.52880859375,
0.377685546875,
0.52880859375,
0.38427734375,
0.5224609375,
0.38427734375
],
[
0.52880859375,
0.374755859375,
0.87890625,
0.374755859375,
0.87890625,
0.388427734375,
0.52880859375,
0.388427734375
],
[
0.1351318359375,
0.393798828125,
0.466796875,
0.392578125,
0.466796875,
0.40673828125,
0.1351318359375,
0.407958984375
],
[
0.5146484375,
0.39013671875,
0.880859375,
0.39013671875,
0.880859375,
0.40380859375,
0.5146484375,
0.40380859375
],
[
0.462646484375,
0.3984375,
0.470947265625,
0.3984375,
0.470947265625,
0.402587890625,
0.462646484375,
0.402587890625
],
[
0.118408203125,
0.410400390625,
0.484375,
0.410400390625,
0.484375,
0.42333984375,
0.118408203125,
0.42333984375
],
[
0.51318359375,
0.40625,
0.880859375,
0.40625,
0.880859375,
0.419189453125,
0.51318359375,
0.419189453125
],
[
0.1175537109375,
0.427001953125,
0.483642578125,
0.427001953125,
0.483642578125,
0.440673828125,
0.1175537109375,
0.440673828125
],
[
0.51318359375,
0.4228515625,
0.736328125,
0.4228515625,
0.736328125,
0.435791015625,
0.51318359375,
0.435791015625
],
[
0.12005615234375,
0.443603515625,
0.48779296875,
0.443603515625,
0.48779296875,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.52880859375,
0.439453125,
0.880859375,
0.439453125,
0.880859375,
0.453125,
0.52880859375,
0.453125
],
[
0.509765625,
0.454833984375,
0.8818359375,
0.45361328125,
0.8818359375,
0.468505859375,
0.509765625,
0.4697265625
],
[
0.1234130859375,
0.46728515625,
0.480224609375,
0.46728515625,
0.480224609375,
0.482177734375,
0.1234130859375,
0.482177734375
],
[
0.51123046875,
0.47021484375,
0.880859375,
0.47021484375,
0.880859375,
0.485107421875,
0.51123046875,
0.485107421875
],
[
0.2301025390625,
0.4912109375,
0.491943359375,
0.490478515625,
0.491943359375,
0.5126953125,
0.2301025390625,
0.51318359375
],
[
0.51318359375,
0.486328125,
0.87890625,
0.485107421875,
0.87890625,
0.49951171875,
0.51318359375,
0.50048828125
],
[
0.5146484375,
0.5029296875,
0.880859375,
0.5029296875,
0.880859375,
0.5166015625,
0.5146484375,
0.5166015625
],
[
0.231689453125,
0.51953125,
0.4130859375,
0.51953125,
0.4130859375,
0.53369140625,
0.231689453125,
0.53369140625
],
[
0.5146484375,
0.52099609375,
0.87744140625,
0.52099609375,
0.87744140625,
0.53076171875,
0.5146484375,
0.53076171875
],
[
0.46337890625,
0.53759765625,
0.486083984375,
0.53759765625,
0.486083984375,
0.54736328125,
0.46337890625,
0.54736328125
],
[
0.51123046875,
0.53369140625,
0.87890625,
0.53369140625,
0.87890625,
0.54736328125,
0.51123046875,
0.54736328125
],
[
0.1368408203125,
0.55029296875,
0.48779296875,
0.55029296875,
0.48779296875,
0.56396484375,
0.1368408203125,
0.56396484375
],
[
0.51123046875,
0.5498046875,
0.880859375,
0.55029296875,
0.880859375,
0.5654296875,
0.51123046875,
0.56396484375
],
[
0.12005615234375,
0.56640625,
0.48779296875,
0.56640625,
0.48779296875,
0.58056640625,
0.12005615234375,
0.58056640625
],
[
0.51318359375,
0.568359375,
0.880859375,
0.568359375,
0.880859375,
0.58203125,
0.51318359375,
0.58203125
],
[
0.1175537109375,
0.58056640625,
0.35693359375,
0.580078125,
0.35693359375,
0.59716796875,
0.1175537109375,
0.5986328125
],
[
0.5146484375,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.5146484375,
0.59716796875
],
[
0.1368408203125,
0.59716796875,
0.484375,
0.59716796875,
0.484375,
0.6123046875,
0.1368408203125,
0.6123046875
],
[
0.51123046875,
0.5986328125,
0.87890625,
0.5986328125,
0.87890625,
0.6123046875,
0.51123046875,
0.6123046875
],
[
0.118408203125,
0.615234375,
0.28369140625,
0.615234375,
0.28369140625,
0.62890625,
0.118408203125,
0.62890625
],
[
0.51318359375,
0.6162109375,
0.8759765625,
0.6162109375,
0.8759765625,
0.626953125,
0.51318359375,
0.626953125
],
[
0.231689453125,
0.6328125,
0.36865234375,
0.6357421875,
0.36865234375,
0.65478515625,
0.231689453125,
0.65185546875
],
[
0.51123046875,
0.63134765625,
0.87890625,
0.63134765625,
0.87890625,
0.64453125,
0.51123046875,
0.64453125
],
[
0.46337890625,
0.63818359375,
0.489501953125,
0.63818359375,
0.489501953125,
0.65087890625,
0.46337890625,
0.65087890625
],
[
0.51318359375,
0.64794921875,
0.880859375,
0.64794921875,
0.880859375,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.1175537109375,
0.65673828125,
0.490234375,
0.65673828125,
0.490234375,
0.6748046875,
0.1175537109375,
0.6748046875
],
[
0.51123046875,
0.662109375,
0.8818359375,
0.6630859375,
0.8818359375,
0.677734375,
0.51123046875,
0.6767578125
],
[
0.118408203125,
0.6748046875,
0.489501953125,
0.6748046875,
0.489501953125,
0.689453125,
0.118408203125,
0.689453125
],
[
0.51123046875,
0.6787109375,
0.880859375,
0.6796875,
0.880859375,
0.69482421875,
0.51123046875,
0.693359375
],
[
0.154541015625,
0.69140625,
0.48779296875,
0.69140625,
0.48779296875,
0.705078125,
0.154541015625,
0.705078125
],
[
0.12176513671875,
0.69775390625,
0.151123046875,
0.69775390625,
0.151123046875,
0.70166015625,
0.12176513671875,
0.70166015625
],
[
0.51318359375,
0.6953125,
0.880859375,
0.6953125,
0.880859375,
0.708984375,
0.51318359375,
0.708984375
],
[
0.118408203125,
0.7080078125,
0.48779296875,
0.7080078125,
0.48779296875,
0.7216796875,
0.118408203125,
0.7216796875
],
[
0.51123046875,
0.7119140625,
0.7421875,
0.7119140625,
0.7421875,
0.7255859375,
0.51123046875,
0.7255859375
],
[
0.1158447265625,
0.7216796875,
0.486083984375,
0.72021484375,
0.486083984375,
0.7412109375,
0.1158447265625,
0.7421875
],
[
0.1158447265625,
0.7412109375,
0.486083984375,
0.73828125,
0.486083984375,
0.75732421875,
0.1158447265625,
0.76025390625
],
[
0.51123046875,
0.7392578125,
0.6630859375,
0.7392578125,
0.6630859375,
0.7529296875,
0.51123046875,
0.7529296875
],
[
0.12005615234375,
0.76123046875,
0.483642578125,
0.76123046875,
0.483642578125,
0.775390625,
0.12005615234375,
0.775390625
],
[
0.51123046875,
0.7587890625,
0.87890625,
0.7587890625,
0.87890625,
0.7724609375,
0.51123046875,
0.7724609375
],
[
0.118408203125,
0.77783203125,
0.484375,
0.77783203125,
0.484375,
0.79150390625,
0.118408203125,
0.79150390625
],
[
0.51318359375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.118408203125,
0.79345703125,
0.364501953125,
0.79345703125,
0.364501953125,
0.80810546875,
0.118408203125,
0.80810546875
],
[
0.51318359375,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.5146484375,
0.81005859375,
0.8759765625,
0.81005859375,
0.8759765625,
0.82080078125,
0.5146484375,
0.82080078125
],
[
0.15869140625,
0.81640625,
0.435791015625,
0.81396484375,
0.435791015625,
0.8388671875,
0.15869140625,
0.84130859375
],
[
0.46337890625,
0.82177734375,
0.486083984375,
0.82177734375,
0.486083984375,
0.83203125,
0.46337890625,
0.83203125
],
[
0.51318359375,
0.82470703125,
0.87890625,
0.82470703125,
0.87890625,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.1368408203125,
0.84423828125,
0.418212890625,
0.84423828125,
0.418212890625,
0.85791015625,
0.1368408203125,
0.85791015625
],
[
0.5146484375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.5146484375,
0.85498046875
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.231689453125,
0.865234375,
0.370361328125,
0.8662109375,
0.370361328125,
0.88134765625,
0.231689453125,
0.8798828125
],
[
0.462646484375,
0.8681640625,
0.48779296875,
0.8681640625,
0.48779296875,
0.88134765625,
0.462646484375,
0.88134765625
],
[
0.509765625,
0.87255859375,
0.87890625,
0.87255859375,
0.87890625,
0.8857421875,
0.509765625,
0.8857421875
],
[
0.1318359375,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1318359375,
0.90283203125
],
[
0.51318359375,
0.88818359375,
0.87890625,
0.88818359375,
0.87890625,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.1292724609375,
0.89453125,
0.1368408203125,
0.89453125,
0.1368408203125,
0.8984375,
0.1292724609375,
0.8984375
],
[
0.118408203125,
0.90380859375,
0.411376953125,
0.90478515625,
0.411376953125,
0.91943359375,
0.118408203125,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.87744140625,
0.90478515625,
0.87744140625,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"是具有低方差的互信息下限：",
"4Experiment",
"N",
"4.1J",
"Evaluation Setup",
"1",
"esim(cl.s",
"Lc = max E",
"log",
"N",
"DN, () /",
"=1",
"我们对七个标准监督任务和七个无监督任务进",
"",
"",
"(4)",
"行实验。我们使用SentEval工具包（?）进行评",
"其中 (2,2),...,(2,2) ～ pN (z1,z2)",
"估。在（？？）之后：我们将无监督任务作为句子",
"",
"z,2是i的两个正实例。N指批量大小，P",
"嵌入方法和监督结果的主要比较，以供参考",
"指（Z1,Z2）的统计分布",
"无监督任务我们评估七个语义文本相似性",
"（STS）任务的表示：STS2012-2016（?????）",
"3.2Reconstruction",
"STSBenchmark（?）和SICK相关性（?），并计",
"重建的细节如图1的右侧所示。正实例21和",
"算句子嵌入之间的余弦相似性。所有无监督",
"22包含独立信息，同时共享一些信息。21中的",
"实验都是完全无监督的，这意味着不使用STS",
"噪声信息（标记为绿点）预计将被重建任务丢",
"训练数据集，并且所有嵌入在训练后都是固定",
"弃。我们证明了该任务可以通过最小化信息熵",
"的。为了具有可比性，我们遵循（？）的评估协",
"来丢弃无用的信息。重建任务鼓励自监督信号",
"议，在所有主题子集上采用Spearman秩相关",
"22重建学习表示z1，这建议最大化对数条件",
"和聚合",
"似然EPz1,z2[1ogP(z1Z2)]。我们知道",
"受监督的任务我们评估七个受监督任务的表",
"示：MR(?）、CR(?)、SUBJ(?)、MPQA(?)",
"-H (Z′Z2) = Epz1,z2 [1og P (Z'|Z2)] (5)",
"",
"SST-2（?）、TREC（?）和MRPC（?）。在由",
"因此，此重建也意味着最小化H（Z1Z2）",
"不同方法产生的（冻结）句子嵌人的顶部训练",
"",
"定理2足够的自监督表示在输入中包含比足够",
"逻辑回归分类器。我们遵循SentEval的默认配",
"和最小的自监督表示更多的穴余信息。后者包",
"置，并使用准确性作为指标",
"含大量的信息，I（X；ST），不能从输入丢弃：",
"培训详细信息我们从预训练的BERT（?)",
"（无套管）或RoBERTa（?）（套管）开始。与",
"(Zsst; X|T) = I (X; SIT) + I (Zsst;XIS,T",
"（？）类似，我们以无监督的方式在来自英语",
"≥I (zslmin;X|T) = I(X; SIT)",
"维基百科的106随机抽样句子上训练我们的",
"InforMin-CL。在训练过程中，我们在[CLS]",
"≥ I (ZsuPmin; X|T)= 0",
"表示的顶部添加一个MLP层作为句子嵌入，并",
"(6)",
"在测试时直接将[CLS]表示作为句子嵌入。掩",
"定理2（请参考（?）进行证明）表示Zssl包",
"码语言建模（MLM）目标（?）作为可选辅助损",
"含两部分允余信息，而Zsstmin包含一部分",
"失添加到方程9：L+β*LMLM（β是超参数）。",
"余信息，丢弃I(Zssl;XS,T)",
"对于所有结果，我们使用以下超参数：epoch：1，",
"因此，如果22可以在I（2；22）最大化的约",
"温度T：0。05，优化器：Adam（?））。我们在STS",
"束下完美地重构任何",
"B开发集上对批量大小E364.128.2561和学立",
"(21, 22) ~ Pz1,z2",
"速率E[1e-5,3e-5,5e-5}进行网格搜索。",
"(7)",
"在训练过程中，我们将得分最高的检查点保存",
"的21，则根据等式2得到2llmin。然",
"在STS-B开发集上，以找到最佳的超参数。我",
"后z1丢弃余信息，不包括I（z1;22t）（i。",
"们采用表3中列出的超参数设置。对于所有结",
"，两个正实例1和z2之间的共享",
"",
"果，我们使用带有GeForceRTX3090GPU的PC",
"信息中的余信息量）。为了更容易优",
"(CUDA11, PyTorchl。 7。 1)",
"化，我们使用EPz,z2[1ogQ（z1/z2)]作为",
"EPz1,z2[1ogP（z1|z2)]的下界。在我们的部署",
"4.2Main Results",
"中，我们利用方程4中的设计，让Q（Z1Z2)",
"基线我们将InforMin-cL与以前的典型句子",
"是高斯N（Z1Z2,αI），αI作为对角矩阵。因",
"嵌入方法进行比较，其中包括平均GloVe嵌人",
"此，我们得到的重建目标如下：",
"(?），跳过思想（?）和平均BERT或RoBERTa",
"嵌入。我们还比较了后处理方法和使用对比自",
"CR = E21,2~ P212 [-[21 - 222]",
"(8)",
"标的方法。后处理方法包括BERT-floW（?）和",
"我们将两个目标合并为一个总目标：",
"BERT-白化（?）。使用对比物镜的方法包括IS",
"BERT(?)DeCLUTR(?)、CT(?)、SimCSE",
"C= Lc+入* LR",
"(9)",
"（?）和SCD（?）。IS-BERT最大化全局和局部",
"其中入是一个超参数。具有总损失的训练模",
"特征之间的一致性。DeCLUTR从同一文档中",
"",
"型使我们能够丢弃视图中的余信息",
"获取不同的跨度作为正对。CT对齐来自两个"
] |
{
"arxiv_id": "2209.10951",
"language": "zh",
"page_idx": 4
}
|
||
2211.06127-zh-2
|
2211.06127-zh-2.jpg
|
在我们的实验中，如图$\boxed{\mathrm{E} }1$ (b)所示，我们惊讶地发现，纯英语数据上的对比学习似乎能够对齐跨语言表示。跨语言语义上相似的句子显然更接近。与以前的无监督或弱监督方法相比，我们的无监督方法显着提高了跨语言 STS 和句子检索任务的性能。在检索低资源语言和STS 任务时，我们的方法甚至与在数十亿并行数据上训练的完全监督方法不相上下。我们的结果表明，使用对比学习来学习句子关系比使用大规模并行数据来学习通用句子嵌入更有效。据我们所知，我们是第一个证明仅使用英语数据可以有效地学习通用句子嵌人的人
## 2 $\textbf{Related Work}$
# 2.1 Cross-lingual Zero-shot Transfer Learning
由于不同语言之间的相似性，例如单词，语法和语义，多语言模型$\overline{\mathrm{E}}($Devlin et al., 2019; Conneau and Lample, 2019; Conneau et al., 2020; Wei et al., 2021; Chi et al., 2022) 已被证明可以在各种任务中推广到看不见的语言。在机器翻译中，多语言模型表现出执行零转换的能力，因为它们可以在看不见的语言对上进行翻译 (Zoph et al., 2016; Johnson et al., 2017)。在自监督学习的帮助下，模型可以更好地获得语言不变的表示，从而改善零拍机器翻译(Siddhant et al., 2020; Liu et al., 2020)
要学习语言不变的表示跨语言任务，以前的工作应用对抗性网络$\overline{\mathrm{E}}$(Keung et al., 2019; Chen et al.,2019)或通过并行语料库$\overline{\mathrm{E}}($Cao et al., 2020) 对齐表示。Pires et al.(2019) 揭示了 mBERT 擅长零镜头跨语言转移，对特定的单语任务进行微调可以推广到其他语言。最近，X-Mixup (Yang et al., 2022) 执行源语言和目标语言的流形混合，以学习一般表示。我们的方法与通过使用英语下游任务来微调多语言语言模型$\overline{\mathbb{F}}($Phang et al., 2020) 来增强跨语言迁移的工作最相关，因为这两种方法都利用了英语 NLI 监督。我们的方法与他们的方法不同， 因为我们将对比学习应用于表示学习，并且我们只关注句子嵌入
# 2.2 English Sentence Embeddings
句子嵌入旨在将句子映射到共享的嵌入空间中，在该空间中，具有相似含义的句子可以彼此靠近。以前的工作通过以生成的$\overline{\mathrm{E}}($Kiros et al., 2015) 或判别的E(Logeswaran and Lee, 2018) 方式预测输入句子的周围句子来学习句子嵌入。最近，随着对比学习在学习视觉表示 (Chen et al.,2020b)上的成功，越来越多的工作探索了句子嵌入的对比学习。对比学习的
训练信号可以通过数据增强 (Fang et al.,2020; Yan et al., 2021; Meng et al., 2021; Carlsson et al., 2021)或自引导架构(Kim et al.,2021; Carlsson et al., 2021) 获得。其中，SimCSE$\underline{\mathrm{E}}($Gao et al., 2021)是最著名的一个，它采用 NLI 监督来定义相关句子或无监督退出作为数据增强，并提高了最先进的结果。由于其令人印象深刻的性能和简单性，我们选择将 SimCSE 扩展到多语言
2.3 Cross-lingual Sentence Embeddings
通用跨语言句子嵌入将语义相似的跨语言句子对齐到共享嵌入空间中。为了学习跨语言嵌入，以前的工作使用大量并行数据来训练神经网络，以使并行句子的嵌入更接近。LASER (Artetxe and Schwenk, 2019b) 用 93 种语言的平行句子训练 Bi-LSTM,以鼓励跨语言句子嵌入之间的一致性。MUSE (Yang et al., 2020) 通过基于翻译的桥接任务学习 16 种语言的通用句子嵌人，包括多特征问答预测，翻译排名和 NLI。LaBSE (Feng et al.,2020) 利用双编码器框架在 109 种语言的 60 亿个并行句子对上微调 mBERT。Reimers and Gurevych (2020) 通过师生培训将单语句子嵌入扩展到多语言，其中目标语言学生模型模仿教师模型中的源语言嵌入。句子编码器(SP)$\operatorname{E}($Wieting et al., 2021) 只需使用并行数据学习句子嵌入， 并优于 BERT 基模型
也可以以弱监督甚至无监督的方式学习跨语言句子嵌人。CRISS (Tran et al.,2020) 使用mBART 编码器的表示形式作为初始句子嵌人， 用于从单语语料库中挖掘并行文本。它们迭代地提取 bitexts 并使用 bitexts 在自训练循环中更新模型。请注意，我们的方法可以很容易地与 CRISS 结合使用 mSimCSE 作为初始模型。我们的方法与 DuEAM$\underline{\mathrm{E}}($Goswami et al., 2021) 密切相关，DuEAME(Goswami et al., 2021) 是一种双编码器模型，它利用单词移动器距离和跨语言并行 NLI 句子来构造正训练对。主要区别在于我们使用对比学习，它比双编码器更有效，并且我们的方法更简单，只需要英语数据3 Method
## 3.1 SimCSE
SimCSE (Gao et al.,2021)应用批量对比学习 (Chen et al., 2020b) 学习英语数据上的句子嵌人。批量对比学习将正训练对和负训练对放人同一批次，增加了对比任务的难度
具体来说，正训练对表示该对中的句子在语义上相似，它们的嵌入应该拉得更近，而负训练对表示句子在语义上不同。给定一批正训练对$B=\{(x_i,x_i^+)\}_{i=1}^N$, SimCSE 将第 i 对的批
|
[
[
0.1368408203125,
0.0872802734375,
0.489501953125,
0.0872802734375,
0.489501953125,
0.1009521484375,
0.1368408203125,
0.1009521484375
],
[
0.51318359375,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.1175537109375,
0.10272216796875,
0.486083984375,
0.10272216796875,
0.486083984375,
0.11639404296875,
0.1175537109375,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.118408203125,
0.119384765625,
0.48779296875,
0.119384765625,
0.48779296875,
0.133056640625,
0.118408203125,
0.133056640625
],
[
0.5146484375,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.12005615234375,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.51123046875,
0.1341552734375,
0.87890625,
0.1341552734375,
0.87890625,
0.1478271484375,
0.51123046875,
0.1478271484375
],
[
0.12005615234375,
0.1519775390625,
0.486083984375,
0.1519775390625,
0.486083984375,
0.1650390625,
0.12005615234375,
0.1650390625
],
[
0.51318359375,
0.15087890625,
0.87890625,
0.15087890625,
0.87890625,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.12005615234375,
0.16748046875,
0.484375,
0.16748046875,
0.484375,
0.18115234375,
0.12005615234375,
0.18115234375
],
[
0.515625,
0.16748046875,
0.880859375,
0.16748046875,
0.880859375,
0.18115234375,
0.515625,
0.18115234375
],
[
0.118408203125,
0.182861328125,
0.484375,
0.182861328125,
0.484375,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.51123046875,
0.1817626953125,
0.87890625,
0.1817626953125,
0.87890625,
0.1953125,
0.51123046875,
0.1953125
],
[
0.118408203125,
0.20068359375,
0.48193359375,
0.20068359375,
0.48193359375,
0.21142578125,
0.118408203125,
0.21142578125
],
[
0.51318359375,
0.20068359375,
0.87890625,
0.20068359375,
0.87890625,
0.21142578125,
0.51318359375,
0.21142578125
],
[
0.118408203125,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.50390625,
0.2161865234375,
0.5322265625,
0.2161865234375,
0.5322265625,
0.2286376953125,
0.50390625,
0.2286376953125
],
[
0.118408203125,
0.232177734375,
0.484375,
0.232177734375,
0.484375,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.5146484375,
0.243408203125,
0.5615234375,
0.243408203125,
0.5615234375,
0.254150390625,
0.5146484375,
0.254150390625
],
[
0.55517578125,
0.240478515625,
0.84033203125,
0.2410888671875,
0.84033203125,
0.255859375,
0.55517578125,
0.254638671875
],
[
0.118408203125,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.1175537109375,
0.26318359375,
0.447509765625,
0.26318359375,
0.447509765625,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.51123046875,
0.262451171875,
0.87890625,
0.262451171875,
0.87890625,
0.275634765625,
0.51123046875,
0.275634765625
],
[
0.51318359375,
0.27783203125,
0.880859375,
0.279052734375,
0.880859375,
0.292236328125,
0.51318359375,
0.29150390625
],
[
0.1175537109375,
0.292236328125,
0.154541015625,
0.292236328125,
0.154541015625,
0.3046875,
0.1175537109375,
0.3046875
],
[
0.1502685546875,
0.292236328125,
0.274658203125,
0.292236328125,
0.274658203125,
0.305908203125,
0.1502685546875,
0.305908203125
],
[
0.5146484375,
0.29443359375,
0.880859375,
0.29443359375,
0.880859375,
0.30810546875,
0.5146484375,
0.30810546875
],
[
0.51123046875,
0.308837890625,
0.87890625,
0.308837890625,
0.87890625,
0.322509765625,
0.51123046875,
0.322509765625
],
[
0.1158447265625,
0.318359375,
0.42236328125,
0.318359375,
0.42236328125,
0.333251953125,
0.1158447265625,
0.333251953125
],
[
0.51318359375,
0.32666015625,
0.880859375,
0.32666015625,
0.880859375,
0.34033203125,
0.51318359375,
0.34033203125
],
[
0.15869140625,
0.333251953125,
0.235107421875,
0.336181640625,
0.2333984375,
0.350830078125,
0.156982421875,
0.347412109375
],
[
0.51123046875,
0.341552734375,
0.8818359375,
0.342529296875,
0.8818359375,
0.35693359375,
0.51123046875,
0.355712890625
],
[
0.118408203125,
0.355712890625,
0.486083984375,
0.355712890625,
0.486083984375,
0.369384765625,
0.118408203125,
0.369384765625
],
[
0.51318359375,
0.358154296875,
0.87890625,
0.358154296875,
0.87890625,
0.371826171875,
0.51318359375,
0.371826171875
],
[
0.118408203125,
0.372314453125,
0.48779296875,
0.372314453125,
0.48779296875,
0.385986328125,
0.118408203125,
0.385986328125
],
[
0.51318359375,
0.3759765625,
0.880859375,
0.3759765625,
0.880859375,
0.3896484375,
0.51318359375,
0.3896484375
],
[
0.118408203125,
0.388427734375,
0.486083984375,
0.388427734375,
0.486083984375,
0.402099609375,
0.118408203125,
0.402099609375
],
[
0.5146484375,
0.391357421875,
0.880859375,
0.391357421875,
0.880859375,
0.405029296875,
0.5146484375,
0.405029296875
],
[
0.1175537109375,
0.40380859375,
0.484375,
0.40380859375,
0.484375,
0.41748046875,
0.1175537109375,
0.41748046875
],
[
0.51123046875,
0.407958984375,
0.87744140625,
0.407958984375,
0.87744140625,
0.421630859375,
0.51123046875,
0.421630859375
],
[
0.118408203125,
0.4228515625,
0.483642578125,
0.4228515625,
0.483642578125,
0.432861328125,
0.118408203125,
0.432861328125
],
[
0.51318359375,
0.42333984375,
0.880859375,
0.42333984375,
0.880859375,
0.43701171875,
0.51318359375,
0.43701171875
],
[
0.118408203125,
0.43701171875,
0.486083984375,
0.43701171875,
0.486083984375,
0.45068359375,
0.118408203125,
0.45068359375
],
[
0.51318359375,
0.439453125,
0.8818359375,
0.439453125,
0.8818359375,
0.453125,
0.51318359375,
0.453125
],
[
0.1175537109375,
0.453125,
0.483642578125,
0.453125,
0.483642578125,
0.466064453125,
0.1175537109375,
0.466064453125
],
[
0.51123046875,
0.4560546875,
0.880859375,
0.4560546875,
0.880859375,
0.4697265625,
0.51123046875,
0.4697265625
],
[
0.118408203125,
0.4697265625,
0.483642578125,
0.4697265625,
0.483642578125,
0.482666015625,
0.118408203125,
0.482666015625
],
[
0.51318359375,
0.471435546875,
0.880859375,
0.471435546875,
0.880859375,
0.485107421875,
0.51318359375,
0.485107421875
],
[
0.118408203125,
0.48681640625,
0.48193359375,
0.48681640625,
0.48193359375,
0.49755859375,
0.118408203125,
0.49755859375
],
[
0.51318359375,
0.48681640625,
0.8818359375,
0.48681640625,
0.8818359375,
0.501953125,
0.51318359375,
0.501953125
],
[
0.118408203125,
0.501953125,
0.484375,
0.501953125,
0.484375,
0.515625,
0.118408203125,
0.515625
],
[
0.5146484375,
0.50439453125,
0.8818359375,
0.50439453125,
0.8818359375,
0.51708984375,
0.5146484375,
0.51708984375
],
[
0.118408203125,
0.51708984375,
0.32666015625,
0.51708984375,
0.32666015625,
0.53076171875,
0.118408203125,
0.53076171875
],
[
0.51318359375,
0.51953125,
0.67822265625,
0.51953125,
0.67822265625,
0.533203125,
0.51318359375,
0.533203125
],
[
0.1368408203125,
0.53369140625,
0.484375,
0.53369140625,
0.484375,
0.54736328125,
0.1368408203125,
0.54736328125
],
[
0.5322265625,
0.5361328125,
0.880859375,
0.5361328125,
0.880859375,
0.5498046875,
0.5322265625,
0.5498046875
],
[
0.118408203125,
0.55029296875,
0.48779296875,
0.55029296875,
0.48779296875,
0.56396484375,
0.118408203125,
0.56396484375
],
[
0.51123046875,
0.552734375,
0.880859375,
0.552734375,
0.880859375,
0.56640625,
0.51123046875,
0.56640625
],
[
0.118408203125,
0.56640625,
0.484375,
0.56640625,
0.484375,
0.580078125,
0.118408203125,
0.580078125
],
[
0.51123046875,
0.56787109375,
0.880859375,
0.568359375,
0.880859375,
0.58203125,
0.51123046875,
0.58056640625
],
[
0.118408203125,
0.58203125,
0.486083984375,
0.58203125,
0.486083984375,
0.595703125,
0.118408203125,
0.595703125
],
[
0.51318359375,
0.5849609375,
0.8818359375,
0.5849609375,
0.8818359375,
0.5986328125,
0.51318359375,
0.5986328125
],
[
0.118408203125,
0.5986328125,
0.48779296875,
0.5986328125,
0.48779296875,
0.6123046875,
0.118408203125,
0.6123046875
],
[
0.51123046875,
0.599609375,
0.87744140625,
0.599609375,
0.87744140625,
0.61328125,
0.51123046875,
0.61328125
],
[
0.118408203125,
0.6142578125,
0.48779296875,
0.6142578125,
0.48779296875,
0.62744140625,
0.118408203125,
0.62744140625
],
[
0.5146484375,
0.6181640625,
0.87890625,
0.6181640625,
0.87890625,
0.62890625,
0.5146484375,
0.62890625
],
[
0.118408203125,
0.62890625,
0.486083984375,
0.62890625,
0.486083984375,
0.642578125,
0.118408203125,
0.642578125
],
[
0.509765625,
0.63134765625,
0.880859375,
0.6318359375,
0.880859375,
0.646484375,
0.509765625,
0.6455078125
],
[
0.118408203125,
0.6455078125,
0.48779296875,
0.646484375,
0.48779296875,
0.6611328125,
0.118408203125,
0.66015625
],
[
0.51318359375,
0.6484375,
0.880859375,
0.6484375,
0.880859375,
0.662109375,
0.51318359375,
0.662109375
],
[
0.118408203125,
0.6630859375,
0.486083984375,
0.6630859375,
0.486083984375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.509765625,
0.6630859375,
0.87890625,
0.6630859375,
0.87890625,
0.6767578125,
0.509765625,
0.6767578125
],
[
0.118408203125,
0.6787109375,
0.484375,
0.6787109375,
0.484375,
0.6923828125,
0.118408203125,
0.6923828125
],
[
0.515625,
0.68408203125,
0.52880859375,
0.68408203125,
0.52880859375,
0.6884765625,
0.515625,
0.6884765625
],
[
0.52490234375,
0.68115234375,
0.880859375,
0.68115234375,
0.880859375,
0.69482421875,
0.52490234375,
0.69482421875
],
[
0.118408203125,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.118408203125,
0.7080078125
],
[
0.51123046875,
0.69677734375,
0.880859375,
0.69677734375,
0.880859375,
0.71142578125,
0.51123046875,
0.71142578125
],
[
0.118408203125,
0.71142578125,
0.489501953125,
0.71142578125,
0.489501953125,
0.724609375,
0.118408203125,
0.724609375
],
[
0.51318359375,
0.71337890625,
0.87890625,
0.71337890625,
0.87890625,
0.72705078125,
0.51318359375,
0.72705078125
],
[
0.118408203125,
0.72705078125,
0.489501953125,
0.72705078125,
0.489501953125,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.51123046875,
0.72998046875,
0.880859375,
0.72998046875,
0.880859375,
0.74365234375,
0.51123046875,
0.74365234375
],
[
0.118408203125,
0.74365234375,
0.26611328125,
0.74365234375,
0.26611328125,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.509765625,
0.7529296875,
0.6171875,
0.7529296875,
0.6171875,
0.76708984375,
0.509765625,
0.76708984375
],
[
0.1158447265625,
0.76953125,
0.399658203125,
0.77099609375,
0.399658203125,
0.78515625,
0.1158447265625,
0.78369140625
],
[
0.51123046875,
0.77099609375,
0.62744140625,
0.77099609375,
0.62744140625,
0.78515625,
0.51123046875,
0.78515625
],
[
0.118408203125,
0.7919921875,
0.484375,
0.7919921875,
0.484375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51318359375,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.118408203125,
0.81005859375,
0.483642578125,
0.81005859375,
0.483642578125,
0.82080078125,
0.118408203125,
0.82080078125
],
[
0.51318359375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.118408203125,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.5146484375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.82275390625,
0.84130859375,
0.82275390625,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.53076171875,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.53076171875,
0.87060546875
],
[
0.12005615234375,
0.87353515625,
0.48779296875,
0.87353515625,
0.48779296875,
0.88720703125,
0.12005615234375,
0.88720703125
],
[
0.515625,
0.87451171875,
0.87890625,
0.87451171875,
0.87890625,
0.88525390625,
0.515625,
0.88525390625
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.12005615234375,
0.90673828125,
0.486083984375,
0.90673828125,
0.486083984375,
0.91748046875,
0.12005615234375,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"在我们的实验中，如图E1（b）所示，我们",
"训练信号可以通过数据增强（Fangetal.，2020;",
"惊牙地发现，纯英语数据上的对比学习似乎能",
"Yan et al., 2021; Meng et al., 2021; Carlsson et al..",
"够对齐跨语言表示。跨语言语义上相似的句子",
"2021）或自引/导架构（Kimetal.2021；Carlsson",
"显然更接近。与以前的无监督或弱监督方法相",
"etal.,2021)获得。其中，SimCSEE(Gaoetal.",
"比，我们的无监督方法显着提高了跨语言STS",
"2021)是最著名的一个，它采用NLI监督来定",
"和句子检索任务的性能。在检索低资源语言和",
"义相关句子或无监督退出作为数据增强，并提",
"STS任务时，我们的方法甚至与在数十亿并行",
"高了最先进的结果。由于其令人印象深刻的性",
"数据上训练的完全监督方法不相上下。我们的",
"能和简单性，我们选择将SimCSE扩展到多语",
"结果表明，使用对比学习来学习句子关系比使",
"言",
"用大规模并行数据来学习通用句子嵌人更有",
"2.3",
"Cross-lingual Sentence Embeddings",
"效。据我们所知，我们是第一个证明仅使用英",
"语数据可以有效地学习通用句子嵌人的人",
"通用跨语言句子嵌入将语义相似的跨语言",
"句子对齐到共享嵌人空间中。为了学习跨语",
"2",
"Related Work",
"言嵌入，以前的工作使用大量并行数据来训",
"练神经网络，以使并行句子的嵌入更接近。",
"2.1 Cross-lingual Zero-shot Transfer",
"LASER (Artetxe and Schwenk, 2019b)用 93种",
"Learning",
"语言的平行句子训练Bi-LSTM，以鼓励跨语",
"由于不同语言之间的相似性，例如单词，语",
"言句子嵌人之间的一致性。MUSE（Yangetal.，",
"法和语义，多语言模型E(Devlinetal，2019;",
"2020通过基于翻译的桥接任务学习16种语",
"Conneau and Lample, 2019; Conneau et al., 2020;",
"言的通用句子嵌入，包括多特征问答预测，翻",
"Weietal.,2021;Chi etal.,2022)已被证明可以",
"译排名和NLI。LaBSE(Feng et al.，2020）利",
"在各种任务中推广到看不见的语言。在机器",
"用双编码器框架在109种语言的60亿个并行",
"翻译中，多语言模型表现出执行零转换的能",
"句子对上微调mBERT。ReimersandGurevych",
"力，因为它们可以在看不见的语言对上进行翻",
"(2020)通过师生培训将单语句子嵌入扩展到多",
"译(Zoph et al.,2016;Johnson et al.,2017)。在自",
"语言，其中目标语言学生模型模仿教师模型中",
"监督学习的帮助下，模型可以更好地获得语言",
"的源语言嵌入。句子编码器（SP）E(Wieting",
"不变的表示，从而改善零拍机器翻译（Siddhant",
"etal.,2021）只需使用并行数据学习句子嵌入，",
"et al., 2020; Liu et al., 2020)",
"并优于BERT基模型",
"要学习语言不变的表示跨语言任务，以前",
"也可以以弱监督甚至无监督的方式学习跨",
"的工作应用对抗性网络E(Keungetal,2019;",
"语言句子嵌人。CRISS（Tranetal.，2020）使用",
"Chenetal.，2019）或通过并行语料库E(Cao",
"mBART编码器的表示形式作为初始句子嵌入，",
"et al.,2020）对齐表示。Pires etal.(2019）揭",
"用于从单语语料库中挖掘并行文本。它们选代",
"示了mBERT擅长零镜头跨语言转移，对特定",
"地提取bitexts并使用bitexts在自训练循环中",
"的单语任务进行微调可以推广到其他语言。最",
"更新模型。请注意，我们的方法可以很容易地",
"近，X-Mixup（Yangetal.,2022）执行源语言和",
"与CRISS结合使用mSimCSE作为初始模型。",
"目标语言的流形混合，以学习一般表示。我们",
"我们的方法与DuEAME(Goswamietal.，2021)",
"的方法与通过使用英语下游任务来微调多语言",
"密切相关，DuEAME(Goswamietal.,2021）是",
"语言模型E(Phangetal.,2020)来增强跨语言迁",
"",
"-种双编码器模型，它利用单词移动器距离和",
"移的工作最相关，因为这两种方法都利用了英",
"跨语言并行NLI句子来构造正训练对。主要区",
"语NLI监督。我们的方法与他们的方法不同，",
"别在于我们使用对比学习，它比双编码器更有",
"因为我们将对比学习应用于表示学习，并且我",
"效，并且我们的方法更简单，只需要英语数据",
"们只关注句子嵌入",
"3Method",
"2.2English Sentence Embeddings",
"3.1SimCSE",
"句子嵌入旨在将句子映射到共享的嵌入空间",
"SimCSE（Gaoetal.,2021）应用批量对比学",
"中，在该空间中，具有相似含义的句子可以",
"习（Chenetal.，2020b）学习英语数据上的句",
"彼此靠近。以前的工作通过以生成的E(Kiros",
"子嵌入。批量对比学习将正训练对和负训练对",
"et al.，2015）或判别的E(Logeswaran andLee,",
"放人同一批次，增加了对比任务的难度",
"2018）方式预测输入句子的周围句子来学习句",
"具体来说，正训练对表示该对中的句子在语",
"子嵌入。最近，随着对比学习在学习视觉表",
"义上相似，它们的嵌入应该拉得更近，而负训",
"示（Chenetal.，2020b）上的成功，越来越多的",
"练对表示句子在语义上不同。给定一批正训练",
"工作探索了句子嵌人的对比学习。对比学习的",
"对B={(ci，a)}-1，SimCSE将第i对的批"
] |
{
"arxiv_id": "2211.06127",
"language": "zh",
"page_idx": 2
}
|
||
2211.06127-zh-4
|
2211.06127-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>$\overline{\mathrm{Models}}$</th>
<th>$\overline{{\mathbf{BUCC}}}$</th>
<th>${\mathrm{Tatoeba-14}}$</th>
<th>Tatoeba- 36</th>
</tr>
<tr>
<td> </td>
<td>${\overline{{Unsupervis}}}$</td>
<td>$ed$</td>
<td> </td>
</tr>
<tr>
<td>XLM- R</td>
<td>$\overline{66.0}$</td>
<td>57.6</td>
<td>53.4</td>
</tr>
<tr>
<td>INFOXLM</td>
<td> </td>
<td>77.8</td>
<td>67.3</td>
</tr>
<tr>
<td>DuEAM</td>
<td>77.2</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>XLM- E</td>
<td> </td>
<td>72.3</td>
<td>62.3</td>
</tr>
<tr>
<td>HiCTL</td>
<td>68.4</td>
<td> </td>
<td>59.7</td>
</tr>
<tr>
<td>$mSimCSE_{en}$</td>
<td>87.5</td>
<td>82.0</td>
<td>78.0</td>
</tr>
<tr>
<td>$\overline{\mathrm{Engl}}$</td>
<td>ish $1NLI$ sup</td>
<td>ervised</td>
<td> </td>
</tr>
<tr>
<td>(Phang et al., 2020)</td>
<td>71.9</td>
<td> </td>
<td>81.2</td>
</tr>
<tr>
<td>$mSimCSE_{en}$</td>
<td>93.6</td>
<td>89.9</td>
<td>87.7</td>
</tr>
<tr>
<td>Cross- li</td>
<td>$NLI$ ngual</td>
<td>supervised</td>
<td> </td>
</tr>
<tr>
<td>${\overline{mSimCSE_{en,fr}}}$</td>
<td>$\overline{94.2}$</td>
<td>$\overline{90.8}$</td>
<td>88.8</td>
</tr>
<tr>
<td>$mSimCSE_{en,fr,sw}$</td>
<td>94.3</td>
<td>93.3</td>
<td>90.3</td>
</tr>
<tr>
<td>$mSimCSE_{all}$</td>
<td>95.2</td>
<td>93.2</td>
<td>91.4</td>
</tr>
<tr>
<td>DuEAM</td>
<td>81.7</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$F$</td>
<td>ully Superv</td>
<td>ised</td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{LASER}$</td>
<td>92.9</td>
<td>95.3</td>
<td>84.4</td>
</tr>
<tr>
<td>LaBSE</td>
<td>93.5</td>
<td>95.3</td>
<td>95.0</td>
</tr>
<tr>
<td>$mSimCSE_{sw}$</td>
<td>86.8</td>
<td>87.7</td>
<td>86.3</td>
</tr>
<tr>
<td>$mSimCSE_{fr}$</td>
<td>87.1</td>
<td>87.9</td>
<td>85.9</td>
</tr>
<tr>
<td>mSimCSE $sw,fr$</td>
<td>88.8</td>
<td>90.2</td>
<td>88.3</td>
</tr>
<tr>
<td>mSimCSE ${\mathrm{e}}_{w},{\mathrm{f}}_{r}+{\mathrm{NLI}}$</td>
<td>93.6</td>
<td>91.9</td>
<td>90.0</td>
</tr>
</tbody>
</table>
Table 1: Results of sentence retrieval task on Xtreme benchmark. We report Fl-scores for BUCC and accuracy for
Tatoeba.
large(XLM-R) (Conneau et al.,2020)。我们在
附录 A 中检查了不同超参数的性能
Training Data for Different mSimCSE Strategies 在无监督的 mSimCSE 和英语 NLI 监督的 mSimCSE 中，我们分别使用从 SimCSE 代码库下载的预处理的英语 Wikipedia 和英语 NLI 训练元组。在本文的所有表中，mSimCSE的下标表示我们用来训练模型的语言。在跨语言 NLI 监督中，$mSimCSE_{en,fr}$表示我们使用英语和翻译的法语 NLI 数据来训练我们的模型，而$mSimCSE_{all}$表示我们使用 XNLIE (Conneau et al., 2018) 数据集中的所有语言
在监督微调中，$mSimCSE_{sw}$表示我们使用英语和斯瓦希里语的翻译对。对于每种语言， 我们通过 OPUS 语料库集合$\mathbb{E}^3$ 从 ParaCrawl 项目$\mathbb{E}^4$ 中随机选择 100k 个平行句子。在监督微调中，“mSimCSE_{sw} NLI” 表 示 我 们 将 英 语 NLI 句子对与英语-斯瓦希里语翻译对混合在一起。请注意，由于平行句子没有硬否定句为了将它们与 NLI 数据混合，我们还删除了“$mSimCSE_{sw}+$NLI”中 NLI 的硬否定句
## 4.2 Baselines
首先，我们将我们的方法与包括 XLM-R 和 MBERT 在内的无监督预训练语言模型进行比较而无需微调，以表明无监督对比学习可以学习更广泛的跨语言句子嵌人。在某些任务中，我
$^3$https://opus.nlpl.eu/ $^{4}$http://paracrawl.eu
们还将我们的方法与 Xtreme$\underline{\mathrm{E}}($Hu et al., 2020) 基准中更具竞争力的语言模型进行了比较，例如 XLM-E$\boxed{\mathrm{F}}($Chi et al., 2022) , HICTL$\boxed{\mathrm{E}}($Wei et al., 2021) 和 INFOXLM$\boxed{\mathrm{E}}$(Chi et al., 2021) 。CRISSE(Tran et al., 2020) 是一种无监督的句子检索方法，它使用自我训练从多个单语语料库中挖掘平行句子
我们的主要基线是也利用 NLI 监管的其他方法。Phang et al.(2020) 在各种英语任务上微调多语言语言模型，包括英语 NLI 任务。DuEAM (Goswami et al., 2021) 还利用多语言NLI 监督来学习通用句子嵌入
我们还将我们的方法与利用并行句子的完全监督方法进行了比较，包括 LASERE (Artetxe and Schwenk, 2019b) , LaBSEE(Feng et al., 2020) 和 SP$\overline{\mathrm{E}}$(Wieting et al., 2021)。请注意，在所有方法中，$mSimCSE_{en}$是唯一仅使用英文数据的方法
### 4.3 Sentence Retrieval
继先前的跨语言句子嵌入学习$\operatorname{E}($Goswami et al.,2021)的工作之后，我们对多语言句子检索模型进行了评估，包括 Tatoeba$\boxed{\mathrm{F}}($Artetxe and Schwenk, 2019b) 和 BUCCE(Zweigenbaum et al., 2018)。Tatoeba 需要模型来匹配源语言和目标语言句子池中的平行句子。BUCC 是一个 bitext 挖掘任务，其中模型需要对所有可能的句子对进行排名，并预测分数高于优化阈值的句子对
在表 1 中，我们遵循 Xtreme 基准测试中的
|
[
[
0.35693359375,
0.0831298828125,
0.4072265625,
0.0831298828125,
0.4072265625,
0.09442138671875,
0.35693359375,
0.09442138671875
],
[
0.47265625,
0.08197021484375,
0.51904296875,
0.08197021484375,
0.51904296875,
0.09320068359375,
0.47265625,
0.09320068359375
],
[
0.53564453125,
0.0831298828125,
0.60693359375,
0.0831298828125,
0.60693359375,
0.09320068359375,
0.53564453125,
0.09320068359375
],
[
0.62109375,
0.0831298828125,
0.6962890625,
0.0831298828125,
0.6962890625,
0.09442138671875,
0.62109375,
0.09442138671875
],
[
0.456787109375,
0.0968017578125,
0.54248046875,
0.0968017578125,
0.54248046875,
0.10687255859375,
0.456787109375,
0.10687255859375
],
[
0.355224609375,
0.10809326171875,
0.4072265625,
0.10809326171875,
0.4072265625,
0.119384765625,
0.355224609375,
0.119384765625
],
[
0.480224609375,
0.10809326171875,
0.51123046875,
0.10809326171875,
0.51123046875,
0.119384765625,
0.480224609375,
0.119384765625
],
[
0.55322265625,
0.10809326171875,
0.5859375,
0.10809326171875,
0.5859375,
0.12054443359375,
0.55322265625,
0.12054443359375
],
[
0.640625,
0.10809326171875,
0.67236328125,
0.10809326171875,
0.67236328125,
0.12054443359375,
0.640625,
0.12054443359375
],
[
0.345947265625,
0.119384765625,
0.418212890625,
0.119384765625,
0.418212890625,
0.1312255859375,
0.345947265625,
0.1312255859375
],
[
0.55517578125,
0.119384765625,
0.5859375,
0.119384765625,
0.5859375,
0.1318359375,
0.55517578125,
0.1318359375
],
[
0.640625,
0.119384765625,
0.67333984375,
0.119384765625,
0.67333984375,
0.1318359375,
0.640625,
0.1318359375
],
[
0.353515625,
0.1318359375,
0.41064453125,
0.1318359375,
0.41064453125,
0.1436767578125,
0.353515625,
0.1436767578125
],
[
0.478515625,
0.1300048828125,
0.51318359375,
0.1300048828125,
0.51318359375,
0.1455078125,
0.478515625,
0.1455078125
],
[
0.55322265625,
0.141357421875,
0.587890625,
0.141357421875,
0.587890625,
0.15673828125,
0.55322265625,
0.15673828125
],
[
0.640625,
0.142578125,
0.67333984375,
0.142578125,
0.67333984375,
0.155029296875,
0.640625,
0.155029296875
],
[
0.640625,
0.1519775390625,
0.6767578125,
0.155029296875,
0.67333984375,
0.169189453125,
0.63818359375,
0.166259765625
],
[
0.355224609375,
0.1436767578125,
0.4072265625,
0.1436767578125,
0.4072265625,
0.155029296875,
0.355224609375,
0.155029296875
],
[
0.35693359375,
0.15380859375,
0.4072265625,
0.15380859375,
0.4072265625,
0.168701171875,
0.35693359375,
0.168701171875
],
[
0.480224609375,
0.155029296875,
0.51123046875,
0.155029296875,
0.51123046875,
0.16748046875,
0.480224609375,
0.16748046875
],
[
0.478515625,
0.166259765625,
0.509765625,
0.166259765625,
0.509765625,
0.1787109375,
0.478515625,
0.1787109375
],
[
0.55517578125,
0.166259765625,
0.5859375,
0.166259765625,
0.5859375,
0.1787109375,
0.55517578125,
0.1787109375
],
[
0.6396484375,
0.1650390625,
0.67529296875,
0.1650390625,
0.67529296875,
0.18115234375,
0.6396484375,
0.18115234375
],
[
0.4248046875,
0.1787109375,
0.5751953125,
0.1787109375,
0.5751953125,
0.1923828125,
0.4248046875,
0.1923828125
],
[
0.334228515625,
0.16748046875,
0.426513671875,
0.16748046875,
0.426513671875,
0.1787109375,
0.334228515625,
0.1787109375
],
[
0.32080078125,
0.190673828125,
0.44580078125,
0.190673828125,
0.44580078125,
0.2037353515625,
0.32080078125,
0.2037353515625
],
[
0.48193359375,
0.191162109375,
0.51123046875,
0.191162109375,
0.51123046875,
0.203125,
0.48193359375,
0.203125
],
[
0.640625,
0.191162109375,
0.67333984375,
0.191162109375,
0.67333984375,
0.2037353515625,
0.640625,
0.2037353515625
],
[
0.480224609375,
0.203125,
0.509765625,
0.203125,
0.509765625,
0.215576171875,
0.480224609375,
0.215576171875
],
[
0.55322265625,
0.203125,
0.5859375,
0.203125,
0.5859375,
0.215576171875,
0.55322265625,
0.215576171875
],
[
0.640625,
0.201904296875,
0.67333984375,
0.2037353515625,
0.67236328125,
0.21728515625,
0.6396484375,
0.215576171875
],
[
0.4072265625,
0.215576171875,
0.59033203125,
0.215576171875,
0.59033203125,
0.2286376953125,
0.4072265625,
0.2286376953125
],
[
0.3359375,
0.2037353515625,
0.424072265625,
0.2037353515625,
0.424072265625,
0.215576171875,
0.3359375,
0.215576171875
],
[
0.32421875,
0.223876953125,
0.435791015625,
0.22802734375,
0.43408203125,
0.2423095703125,
0.322509765625,
0.2381591796875
],
[
0.478515625,
0.22802734375,
0.51123046875,
0.22802734375,
0.51123046875,
0.2392578125,
0.478515625,
0.2392578125
],
[
0.55517578125,
0.22802734375,
0.58447265625,
0.22802734375,
0.58447265625,
0.2392578125,
0.55517578125,
0.2392578125
],
[
0.640625,
0.226806640625,
0.67333984375,
0.226806640625,
0.67333984375,
0.2392578125,
0.640625,
0.2392578125
],
[
0.313232421875,
0.2340087890625,
0.44580078125,
0.2381591796875,
0.444091796875,
0.255859375,
0.312255859375,
0.251708984375
],
[
0.55517578125,
0.235107421875,
0.5859375,
0.2369384765625,
0.58251953125,
0.2529296875,
0.5517578125,
0.25048828125
],
[
0.640625,
0.2369384765625,
0.67333984375,
0.2392578125,
0.67236328125,
0.2529296875,
0.6396484375,
0.25048828125
],
[
0.33154296875,
0.248779296875,
0.4306640625,
0.25048828125,
0.4306640625,
0.265380859375,
0.33154296875,
0.26318359375
],
[
0.480224609375,
0.2381591796875,
0.51123046875,
0.2381591796875,
0.51123046875,
0.25048828125,
0.480224609375,
0.25048828125
],
[
0.5517578125,
0.2493896484375,
0.587890625,
0.2493896484375,
0.587890625,
0.265380859375,
0.5517578125,
0.265380859375
],
[
0.640625,
0.248779296875,
0.67333984375,
0.25048828125,
0.67236328125,
0.26416015625,
0.6396484375,
0.262451171875
],
[
0.478515625,
0.260009765625,
0.51318359375,
0.262451171875,
0.509765625,
0.27783203125,
0.47607421875,
0.27490234375
],
[
0.480224609375,
0.25048828125,
0.51123046875,
0.25048828125,
0.51123046875,
0.26318359375,
0.480224609375,
0.26318359375
],
[
0.353515625,
0.26318359375,
0.408935546875,
0.26318359375,
0.408935546875,
0.27490234375,
0.353515625,
0.27490234375
],
[
0.654296875,
0.267333984375,
0.66015625,
0.267333984375,
0.66015625,
0.273681640625,
0.654296875,
0.273681640625
],
[
0.44921875,
0.275634765625,
0.55322265625,
0.275634765625,
0.55322265625,
0.287353515625,
0.44921875,
0.287353515625
],
[
0.355224609375,
0.2880859375,
0.4072265625,
0.2880859375,
0.4072265625,
0.2998046875,
0.355224609375,
0.2998046875
],
[
0.480224609375,
0.287353515625,
0.51123046875,
0.287353515625,
0.51123046875,
0.2998046875,
0.480224609375,
0.2998046875
],
[
0.55517578125,
0.2880859375,
0.58447265625,
0.2880859375,
0.58447265625,
0.2998046875,
0.55517578125,
0.2998046875
],
[
0.640625,
0.2880859375,
0.67333984375,
0.2880859375,
0.67333984375,
0.300537109375,
0.640625,
0.300537109375
],
[
0.355224609375,
0.298583984375,
0.408935546875,
0.298583984375,
0.408935546875,
0.3134765625,
0.355224609375,
0.3134765625
],
[
0.480224609375,
0.297607421875,
0.509765625,
0.2998046875,
0.5087890625,
0.312255859375,
0.476806640625,
0.308837890625
],
[
0.55517578125,
0.2998046875,
0.58447265625,
0.2998046875,
0.58447265625,
0.312255859375,
0.55517578125,
0.312255859375
],
[
0.640625,
0.2998046875,
0.67333984375,
0.2998046875,
0.67333984375,
0.312255859375,
0.640625,
0.312255859375
],
[
0.333251953125,
0.311279296875,
0.42822265625,
0.312255859375,
0.42822265625,
0.32373046875,
0.333251953125,
0.322509765625
],
[
0.480224609375,
0.311279296875,
0.51123046875,
0.311279296875,
0.51123046875,
0.32373046875,
0.480224609375,
0.32373046875
],
[
0.55322265625,
0.311279296875,
0.58447265625,
0.311279296875,
0.58447265625,
0.32373046875,
0.55322265625,
0.32373046875
],
[
0.640625,
0.311279296875,
0.67333984375,
0.311279296875,
0.67333984375,
0.32373046875,
0.640625,
0.32373046875
],
[
0.33154296875,
0.319580078125,
0.42822265625,
0.322509765625,
0.42822265625,
0.33740234375,
0.33154296875,
0.334228515625
],
[
0.480224609375,
0.322509765625,
0.509765625,
0.322509765625,
0.509765625,
0.3349609375,
0.480224609375,
0.3349609375
],
[
0.55517578125,
0.322509765625,
0.5859375,
0.322509765625,
0.5859375,
0.3349609375,
0.55517578125,
0.3349609375
],
[
0.640625,
0.322509765625,
0.67333984375,
0.322509765625,
0.67333984375,
0.3349609375,
0.640625,
0.3349609375
],
[
0.322509765625,
0.33203125,
0.43408203125,
0.3349609375,
0.43408203125,
0.349853515625,
0.322509765625,
0.3466796875
],
[
0.480224609375,
0.3349609375,
0.51123046875,
0.3349609375,
0.51123046875,
0.347412109375,
0.480224609375,
0.347412109375
],
[
0.55517578125,
0.333251953125,
0.5859375,
0.3349609375,
0.58447265625,
0.347412109375,
0.55322265625,
0.345703125
],
[
0.640625,
0.3349609375,
0.67333984375,
0.3349609375,
0.67333984375,
0.3466796875,
0.640625,
0.3466796875
],
[
0.30322265625,
0.3466796875,
0.45751953125,
0.3466796875,
0.45751953125,
0.361083984375,
0.30322265625,
0.361083984375
],
[
0.478515625,
0.3466796875,
0.51123046875,
0.3466796875,
0.51123046875,
0.359375,
0.478515625,
0.359375
],
[
0.55322265625,
0.3466796875,
0.5859375,
0.3466796875,
0.5859375,
0.359375,
0.55322265625,
0.359375
],
[
0.640625,
0.3466796875,
0.67333984375,
0.3466796875,
0.67333984375,
0.359375,
0.640625,
0.359375
],
[
0.118408203125,
0.37890625,
0.87890625,
0.37890625,
0.87890625,
0.3896484375,
0.118408203125,
0.3896484375
],
[
0.1175537109375,
0.393798828125,
0.175537109375,
0.393798828125,
0.175537109375,
0.405029296875,
0.1175537109375,
0.405029296875
],
[
0.1158447265625,
0.431640625,
0.48779296875,
0.43115234375,
0.48779296875,
0.4453125,
0.1158447265625,
0.446533203125
],
[
0.51123046875,
0.431640625,
0.880859375,
0.43115234375,
0.880859375,
0.4453125,
0.51123046875,
0.446533203125
],
[
0.1175537109375,
0.448974609375,
0.403076171875,
0.448974609375,
0.403076171875,
0.4619140625,
0.1175537109375,
0.4619140625
],
[
0.51318359375,
0.44775390625,
0.880859375,
0.448974609375,
0.880859375,
0.463134765625,
0.51318359375,
0.4619140625
],
[
0.51123046875,
0.46435546875,
0.8818359375,
0.463134765625,
0.8818359375,
0.47802734375,
0.51123046875,
0.478515625
],
[
0.1234130859375,
0.4755859375,
0.48779296875,
0.4755859375,
0.48779296875,
0.4892578125,
0.1234130859375,
0.4892578125
],
[
0.51123046875,
0.479736328125,
0.88330078125,
0.48095703125,
0.88330078125,
0.495361328125,
0.51123046875,
0.49462890625
],
[
0.1175537109375,
0.4921875,
0.486083984375,
0.4921875,
0.486083984375,
0.505859375,
0.1175537109375,
0.505859375
],
[
0.509765625,
0.495361328125,
0.880859375,
0.496337890625,
0.880859375,
0.51123046875,
0.509765625,
0.51025390625
],
[
0.12005615234375,
0.51025390625,
0.484375,
0.51025390625,
0.484375,
0.52099609375,
0.12005615234375,
0.52099609375
],
[
0.5146484375,
0.51318359375,
0.8818359375,
0.51318359375,
0.8818359375,
0.52783203125,
0.5146484375,
0.52783203125
],
[
0.118408203125,
0.52490234375,
0.486083984375,
0.52490234375,
0.486083984375,
0.5380859375,
0.118408203125,
0.5380859375
],
[
0.51318359375,
0.52978515625,
0.66015625,
0.52978515625,
0.66015625,
0.54345703125,
0.51318359375,
0.54345703125
],
[
0.118408203125,
0.54052734375,
0.486083984375,
0.54052734375,
0.486083984375,
0.55419921875,
0.118408203125,
0.55419921875
],
[
0.53076171875,
0.5458984375,
0.880859375,
0.5458984375,
0.880859375,
0.55859375,
0.53076171875,
0.55859375
],
[
0.12005615234375,
0.55712890625,
0.486083984375,
0.55712890625,
0.486083984375,
0.57080078125,
0.12005615234375,
0.57080078125
],
[
0.51123046875,
0.56005859375,
0.87890625,
0.56005859375,
0.87890625,
0.57470703125,
0.51123046875,
0.57470703125
],
[
0.118408203125,
0.572265625,
0.486083984375,
0.572265625,
0.486083984375,
0.58740234375,
0.118408203125,
0.58740234375
],
[
0.51318359375,
0.57763671875,
0.87744140625,
0.57763671875,
0.87744140625,
0.5927734375,
0.51318359375,
0.5927734375
],
[
0.118408203125,
0.58837890625,
0.484375,
0.58837890625,
0.484375,
0.6025390625,
0.118408203125,
0.6025390625
],
[
0.51318359375,
0.59423828125,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.51318359375,
0.60791015625
],
[
0.1175537109375,
0.60400390625,
0.48779296875,
0.6025390625,
0.48779296875,
0.61767578125,
0.1175537109375,
0.6181640625
],
[
0.51318359375,
0.60986328125,
0.75,
0.60986328125,
0.75,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.1158447265625,
0.6181640625,
0.450927734375,
0.62060546875,
0.450927734375,
0.63720703125,
0.1158447265625,
0.63525390625
],
[
0.52880859375,
0.6259765625,
0.880859375,
0.626953125,
0.880859375,
0.64111328125,
0.52880859375,
0.64013671875
],
[
0.1351318359375,
0.6357421875,
0.484375,
0.63720703125,
0.484375,
0.65185546875,
0.1351318359375,
0.65087890625
],
[
0.51318359375,
0.6435546875,
0.8818359375,
0.6435546875,
0.8818359375,
0.65673828125,
0.51318359375,
0.65673828125
],
[
0.1175537109375,
0.65283203125,
0.48779296875,
0.65380859375,
0.48779296875,
0.66845703125,
0.1175537109375,
0.66748046875
],
[
0.509765625,
0.65771484375,
0.880859375,
0.6591796875,
0.880859375,
0.67333984375,
0.509765625,
0.6728515625
],
[
0.1175537109375,
0.66845703125,
0.484375,
0.66845703125,
0.484375,
0.681640625,
0.1175537109375,
0.681640625
],
[
0.51123046875,
0.67333984375,
0.880859375,
0.6748046875,
0.880859375,
0.689453125,
0.51123046875,
0.6884765625
],
[
0.1175537109375,
0.68408203125,
0.484375,
0.68505859375,
0.484375,
0.69970703125,
0.1175537109375,
0.69873046875
],
[
0.5146484375,
0.69140625,
0.8818359375,
0.69140625,
0.8818359375,
0.705078125,
0.5146484375,
0.705078125
],
[
0.118408203125,
0.70166015625,
0.486083984375,
0.70166015625,
0.486083984375,
0.71533203125,
0.118408203125,
0.71533203125
],
[
0.51123046875,
0.7060546875,
0.66162109375,
0.7080078125,
0.66162109375,
0.72265625,
0.51123046875,
0.72021484375
],
[
0.118408203125,
0.71875,
0.484375,
0.71875,
0.484375,
0.732421875,
0.118408203125,
0.732421875
],
[
0.118408203125,
0.7353515625,
0.484375,
0.7353515625,
0.484375,
0.7490234375,
0.118408203125,
0.7490234375
],
[
0.51318359375,
0.73828125,
0.5556640625,
0.73828125,
0.5556640625,
0.7490234375,
0.51318359375,
0.7490234375
],
[
0.55517578125,
0.73681640625,
0.7060546875,
0.73681640625,
0.7060546875,
0.75048828125,
0.55517578125,
0.75048828125
],
[
0.118408203125,
0.75048828125,
0.483642578125,
0.75048828125,
0.483642578125,
0.765625,
0.118408203125,
0.765625
],
[
0.51123046875,
0.75830078125,
0.880859375,
0.75830078125,
0.880859375,
0.771484375,
0.51123046875,
0.771484375
],
[
0.1158447265625,
0.765625,
0.44921875,
0.7666015625,
0.44921875,
0.78076171875,
0.1158447265625,
0.77978515625
],
[
0.51123046875,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1175537109375,
0.7958984375,
0.2392578125,
0.7958984375,
0.2392578125,
0.80859375,
0.1175537109375,
0.80859375
],
[
0.5146484375,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.5146484375,
0.8046875
],
[
0.51318359375,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.51318359375,
0.82080078125
],
[
0.12005615234375,
0.818359375,
0.489501953125,
0.818359375,
0.489501953125,
0.83203125,
0.12005615234375,
0.83203125
],
[
0.51123046875,
0.82373046875,
0.880859375,
0.82373046875,
0.880859375,
0.83740234375,
0.51123046875,
0.83740234375
],
[
0.1175537109375,
0.83447265625,
0.484375,
0.83447265625,
0.484375,
0.84716796875,
0.1175537109375,
0.84716796875
],
[
0.51318359375,
0.8388671875,
0.86572265625,
0.8388671875,
0.86572265625,
0.8525390625,
0.51318359375,
0.8525390625
],
[
0.118408203125,
0.85107421875,
0.484375,
0.85107421875,
0.484375,
0.8642578125,
0.118408203125,
0.8642578125
],
[
0.51123046875,
0.85546875,
0.87890625,
0.85546875,
0.87890625,
0.869140625,
0.51123046875,
0.869140625
],
[
0.118408203125,
0.8662109375,
0.486083984375,
0.8662109375,
0.486083984375,
0.8798828125,
0.118408203125,
0.8798828125
],
[
0.51318359375,
0.87158203125,
0.8818359375,
0.87158203125,
0.8818359375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.1368408203125,
0.89013671875,
0.334228515625,
0.89111328125,
0.334228515625,
0.90625,
0.1368408203125,
0.90478515625
],
[
0.5146484375,
0.8896484375,
0.587890625,
0.8896484375,
0.587890625,
0.90087890625,
0.5146484375,
0.90087890625
],
[
0.1368408203125,
0.90380859375,
0.32080078125,
0.90478515625,
0.32080078125,
0.91943359375,
0.1368408203125,
0.91845703125
],
[
0.5322265625,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.5322265625,
0.91845703125
]
] |
[
"Models",
"BUCC",
"Tatoeba-14",
"Tatoeba-36",
"Unsupervised",
"XLM-R",
"66.0",
"57.6",
"53.4",
"INFOXLM",
"77.8",
"67.3",
"DuEAM",
"77.2",
"72.3",
"62.3",
"59.7",
"XLM-E",
"HiCTL",
"68.4",
"87.5",
"82.0",
"78.0",
"English NLI supervised",
"mSimCSEen",
"(Phang et al., 2020)",
"71.9",
"81.2",
"93.6",
"89.9",
"87.7",
"Cross-lingual NLI supervised",
"mSimCSEer",
"mSimCSEen.fr",
"94.2",
"90.8",
"88.8",
"mSimCSEen.fr,sw",
"93.3",
"90.3",
"mSimCSEall",
"94.3",
"93.2",
"91.4",
"81.7",
"95.2",
"DuEAM",
"",
"Fully Supervised",
"LASER",
"92.9",
"95.3",
"84.4",
"LaBSE",
"93.5",
"95.3",
"95.0",
"mSimCSEsw",
"86.8",
"87.7",
"86.3",
"mSimCSEfr",
"87.1",
"87.9",
"85.9",
"mSimCSEsw.fr",
"88.8",
"90.2",
"88.3",
"mSimCSEsw.fr +NLI",
"93.6",
"91.9",
"90.0",
"Table I:Results of sentence retrieval task on Xtreme benchmark.We report F1-scores for BUCC and accuracyfor",
"Tatoeba.",
"large(XLM-R）（Conneauet al.,2020)。我们在",
"们还将我们的方法与XtremeE(Huetal.,2020)",
"附录A中检查了不同超参数的性能",
"基准中更具竞争力的语言模型进行了比较，例",
"如 XLM-EE(Chi et al., 2022)，HICTLE(Wei",
"Training Data for Different mSimCSE Strate-",
"et al., 2021) 和 INFOXLME(Chi et al., 2021)。",
"gies在无监督的 mSimCSE和英语NLI监督",
"CRISSE(Tranetal.，2020）是一种无监督的句",
"的mSimCSE中，我们分别使用从SimCSE代码",
"子检索方法，它使用自我训练从多个单语语料",
"库下载的预处理的英语Wikipedia和英语NLI",
"库中挖掘平行句子",
"训练元组。在本文的所有表中，mSimCSE的",
"我们的主要基线是也利用NLI监管的其他",
"下标表示我们用来训练模型的语言。在跨语",
"方法。Phangetal.(2020)在各种英语任务上",
"言NLI监督中，mSimCSEen.fr表示我们使",
"微调多语言语言模型，包括英语NLI任务",
"用英语和翻译的法语NLI数据来训练我们的",
"DuEAM（Goswamietal.，2021）还利用多语言",
"模型，而mSimCSEal表示我们使用XNLIE",
"NLI监督来学习通用句子嵌入",
"(Conneauetal.,2018)数据集中的所有语言",
"我们还将我们的方法与利用并行句子的",
"在监督微调中，mSimCSEsw表示我们使用",
"完全监督方法进行了比较，包括LASERE",
"英语和斯瓦希里语的翻译对。对于每种语言，",
"(Artetxe and Schwenk,2019b)，LaBSEE(Feng",
"我们通过OPUS语料库集合F3从ParaCrawl",
"etal.,2020)和SPE(Wieting et al.,2021）。请注",
"项目E4中随机选择100k个平行句子。在监督",
"意，在所有方法中，mSimCSEen是唯一仅使",
"微调中，“mSimCSEswNLI”表示我们将英",
"用英文数据的方法",
"语NLI句子对与英语-斯瓦希里语翻译对混合",
"在一起。请注意，由于平行句子没有硬否定句，",
"4.3",
"SentenceRetrieval",
"为了将它们与NLI数据混合，我们还删除了",
"继先前的跨语言句子嵌入学习E(Goswami",
"“mSimCSEsw+NLI”中NLI的硬否定句",
"etal.，2021）的工作之后，我们对多语言句子",
"4.2Baselines",
"检索模型进行了评估，包括TatoebaE(Artetxe",
"and Schwenk,2019b)和 BUCCE(Zweigenbaum",
"首先，我们将我们的方法与包括XLM-R和M-",
"etal.，2018）。Tatoeba需要模型来匹配源语言",
"BERT在内的无监督预训练语言模型进行比较",
"和目标语言句子池中的平行句子。BUCC是",
"而无需微调，以表明无监督对比学习可以学习",
"个bitext挖掘任务，其中模型需要对所有可能",
"更广泛的跨语言句子嵌入。在某些任务中，我",
"的句子对进行排名，并预测分数高于优化阈值",
"3https://opus.nlpl.eu/",
"的句子对",
"thttp://paracrawl.eu",
"在表1中，我们遵循Xtreme基准测试中的"
] |
{
"arxiv_id": "2211.06127",
"language": "zh",
"page_idx": 4
}
|
||
2109.07095-zh-2
|
2109.07095-zh-2.jpg
|
数据集中的释义不涉及句子间的多样性，但我们的模型可以通过 ALBERT (Lan et al., 2020) 生成的连贯性关系图学习句子之间的连贯性关系，并利用学习到的句子连贯性感知表示对它们进行重新排序，同时保持生成文档的良好连贯性
我们的模型由三部分组成：句子编码器，图形 GRU 和解码器。句子编码器仅对文档中的每个句子进行单独编码。我们提出了图 GRU, 它结合了图注意力 (Velickovic et al., 2018) 和GRU,以捕捉一致性关系信息。最后，将图形GRU 和句子编码器的输出连接在一起，并用作解码器的输入，以生成释义。进行了广泛的评估，我们的模型在自动评估和人工评估中获得了大多数指标的最佳分数
我们的工作贡献总结如下：
1)据我们所知，我们是第一个探索文档级释义生成问题的人，并指出了文档级释义与句子级释义之间的区别
2) 我们提出了一种新的模型 CoRPG,以解决句子重写和重新排序的文档级释义生成。我们的模型可以利用图 GRU 来学习句子的连贯感知表示，并重新排列输入句子，以改善生成的文档释义的句子间多样性
3) 自动评估和人类评估均表明，我们的模型可以生成具有高度多样性，语义相关性和连贯性的文档释义。我们的代码在 https:// github.com/L-Zhe/CoRPG 上公开可用
# 2 Related Work
随着深度学习的发展，大多数释义模型都基于seq2seq 模型。Prakash et al.(2016) 利用堆叠残差 LSTM 网络生成释义。Gupta et al.(2018) 发现诸如变分自动编码器之类的深度生成模型可以显着提高释义的质量。Li et al.(2019) 提出了 DNPG 将句子分解为句子级模式和短语级模式，以使神经释义生成更加可控。Goyal and Durrett (2020) 使用句法转换来轻柔地“重新排序”源句子，并引导神经模型生成更多样化的释义。Kazemnejad et al. (2020) 探索通过编辑原始句子来生成释义
此外，还有另一种生成释义的方法，称为“旋转”,它利用反向翻译来引入多样性。最近，Mallinson et al. (2017) 用神经机器翻译重新审视了这种方法，以提高释义质量。Wieting and Gimpel (2018) 利用双向翻译模型构建paraNMT, 这是一个非常大的句子级释义数据集
以上所有工作都集中在句子级释义生成上， 据我们所知，没有关于文档级释义生成的研究
$\textbf{3}$ $\textbf{Our CoRPG Model}$
3.1 Overview and Notations
## 3.1.1 Model Overview
图 1 显示了我们模型的概述，该模型由句子编码器，图形 GRU 和解码器组成。给定一个输入文档，我们使用句子编码器来获取文档中每个句子的表示形式，而忽略句子的位置信息。我们为句子构建一个连贯关系图，并使用图 GRU 来获得每个句子的连贯感知表示。句子编码器和图形 GRU 的输出由解码器获取，用于生成连贯的释义文档。请注意，我们没有具有句子间多样性的文档释义数据集。相反，我们使用句子释义模型来构建仅具有句子内多样性的伪文档释义数据集(i。e。, 每个句子内的词汇或句法多样性)。我们的模型在此数据集上进行了训练，以便能够通过修改和排列多个输入句子来重建连贯的文档，而无需使用句子的原始位置信息$^{1}$。换句话说，我们模型的输入可以看作只是一组没有顺序的句子。实现这一目标的关键是图形 GRU 学习的句子的一致性感知表示。在测试过程中，我们的模型可以根据学习到的句子的连贯感知和语义表示来重新组织文本。与输入文档相比，输出文档很可能具有不同的句子顺序和排列方式，因为除了原始句子顺序之外，通常还有不同的合理方式来排列一组句子。数据集和模型模块的详细信息将在下一节中给出
# $\begin{array}{cc}3.1.2&\text{Pseudo Document-Level Paraphrase}\\&\text{Dataset}\end{array}$
我们的模型将释义视为单语翻译任务。给定文$\begin{array}{l}{\text{档 }D=\{S^{1},S^{2},\cdots,S^{N}\}~,~\text{其中 }N~\text{是文档中}}\\{\text{的句子总数，}S^{i}~\text{是文档中的第 }i~\text{个句子。由}}\end{array}$ 于缺乏黄金文档释义数据集，我们利用现成的句子释义模型逐句生成伪文档释义$D_p=$ $\{ S_{p}^{1}, S_{p}^{2}\cdots , S_{p}^{N}\}$ ,其中$S_p^{i}$是通过将$S^i$与句子释义模型进行释义而获得的。然后，我们将$D_p$设置为输入，将$D$设置为目标来训练我们的模型
3.13 Coherence Relationship Graph
有许多作品专注于文本连贯性，例如 NCOH (Moon et al., 2019) 和 CohEval (Mohiuddin et al., 2020)。许多预先训练的模型还将文本一致性作为子任务引入，以提高泛化能力。例如， Devlin et al. (2019) 使用下一个句子预测 (NSP) 任务来训练 BERT; Lan et al.(2020)提出了
$^{1}$If we use the original positional information of input sentences for training, the model can simply output a document with the same positions of these sentences. During testing, the model cannot generate document paraphrase with sentence reordering and rearrangement.
|
[
[
0.12176513671875,
0.0885009765625,
0.486083984375,
0.0885009765625,
0.486083984375,
0.09857177734375,
0.12176513671875,
0.09857177734375
],
[
0.48779296875,
0.08612060546875,
0.71728515625,
0.08612060546875,
0.71728515625,
0.09979248046875,
0.48779296875,
0.09979248046875
],
[
0.118408203125,
0.10272216796875,
0.484375,
0.10272216796875,
0.484375,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51123046875,
0.11224365234375,
0.5615234375,
0.11224365234375,
0.5615234375,
0.12353515625,
0.51123046875,
0.12353515625
],
[
0.55322265625,
0.11224365234375,
0.748046875,
0.11224365234375,
0.748046875,
0.1258544921875,
0.55322265625,
0.1258544921875
],
[
0.118408203125,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.51123046875,
0.133056640625,
0.57421875,
0.133056640625,
0.57421875,
0.144287109375,
0.51123046875,
0.144287109375
],
[
0.568359375,
0.133056640625,
0.7021484375,
0.133056640625,
0.7021484375,
0.146728515625,
0.568359375,
0.146728515625
],
[
0.51123046875,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51123046875,
0.1650390625
],
[
0.12005615234375,
0.1353759765625,
0.486083984375,
0.1353759765625,
0.486083984375,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.12005615234375,
0.152587890625,
0.484375,
0.152587890625,
0.484375,
0.163330078125,
0.12005615234375,
0.163330078125
],
[
0.12005615234375,
0.168701171875,
0.1561279296875,
0.168701171875,
0.1561279296875,
0.179931640625,
0.12005615234375,
0.179931640625
],
[
0.51318359375,
0.166259765625,
0.880859375,
0.16748046875,
0.880859375,
0.1817626953125,
0.51318359375,
0.18115234375
],
[
0.1376953125,
0.18701171875,
0.484375,
0.18701171875,
0.484375,
0.19775390625,
0.1376953125,
0.19775390625
],
[
0.5146484375,
0.18408203125,
0.880859375,
0.18408203125,
0.880859375,
0.19775390625,
0.5146484375,
0.19775390625
],
[
0.118408203125,
0.201904296875,
0.484375,
0.201904296875,
0.484375,
0.215576171875,
0.118408203125,
0.215576171875
],
[
0.51123046875,
0.1995849609375,
0.87890625,
0.1995849609375,
0.87890625,
0.213134765625,
0.51123046875,
0.213134765625
],
[
0.118408203125,
0.21728515625,
0.48779296875,
0.21728515625,
0.48779296875,
0.23095703125,
0.118408203125,
0.23095703125
],
[
0.51123046875,
0.215576171875,
0.87890625,
0.215576171875,
0.87890625,
0.2286376953125,
0.51123046875,
0.2286376953125
],
[
0.118408203125,
0.2340087890625,
0.486083984375,
0.2340087890625,
0.486083984375,
0.2476806640625,
0.118408203125,
0.2476806640625
],
[
0.51123046875,
0.2327880859375,
0.880859375,
0.2327880859375,
0.880859375,
0.2464599609375,
0.51123046875,
0.2464599609375
],
[
0.118408203125,
0.2493896484375,
0.48779296875,
0.2493896484375,
0.48779296875,
0.26318359375,
0.118408203125,
0.26318359375
],
[
0.51123046875,
0.248779296875,
0.8818359375,
0.248779296875,
0.8818359375,
0.262451171875,
0.51123046875,
0.262451171875
],
[
0.1175537109375,
0.26416015625,
0.484375,
0.265380859375,
0.484375,
0.279052734375,
0.1175537109375,
0.27783203125
],
[
0.51123046875,
0.26318359375,
0.87890625,
0.26318359375,
0.87890625,
0.276611328125,
0.51123046875,
0.276611328125
],
[
0.118408203125,
0.281982421875,
0.486083984375,
0.281982421875,
0.486083984375,
0.295654296875,
0.118408203125,
0.295654296875
],
[
0.51318359375,
0.28076171875,
0.880859375,
0.28076171875,
0.880859375,
0.29443359375,
0.51318359375,
0.29443359375
],
[
0.118408203125,
0.297607421875,
0.48779296875,
0.297607421875,
0.48779296875,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.51318359375,
0.29638671875,
0.880859375,
0.29638671875,
0.880859375,
0.31005859375,
0.51318359375,
0.31005859375
],
[
0.118408203125,
0.314208984375,
0.340087890625,
0.314208984375,
0.340087890625,
0.327880859375,
0.118408203125,
0.327880859375
],
[
0.5146484375,
0.312255859375,
0.880859375,
0.312255859375,
0.880859375,
0.325927734375,
0.5146484375,
0.325927734375
],
[
0.1368408203125,
0.33203125,
0.34765625,
0.33203125,
0.34765625,
0.345703125,
0.1368408203125,
0.345703125
],
[
0.5146484375,
0.328857421875,
0.87744140625,
0.328857421875,
0.87744140625,
0.339111328125,
0.5146484375,
0.339111328125
],
[
0.51318359375,
0.34326171875,
0.87890625,
0.34326171875,
0.87890625,
0.35693359375,
0.51318359375,
0.35693359375
],
[
0.1368408203125,
0.349853515625,
0.486083984375,
0.349853515625,
0.486083984375,
0.363525390625,
0.1368408203125,
0.363525390625
],
[
0.118408203125,
0.366455078125,
0.486083984375,
0.366455078125,
0.486083984375,
0.380126953125,
0.118408203125,
0.380126953125
],
[
0.515625,
0.3623046875,
0.87890625,
0.3623046875,
0.87890625,
0.372314453125,
0.515625,
0.372314453125
],
[
0.118408203125,
0.3818359375,
0.28369140625,
0.3818359375,
0.28369140625,
0.3955078125,
0.118408203125,
0.3955078125
],
[
0.5146484375,
0.377197265625,
0.880859375,
0.377197265625,
0.880859375,
0.39013671875,
0.5146484375,
0.39013671875
],
[
0.51318359375,
0.393798828125,
0.87744140625,
0.393798828125,
0.87744140625,
0.40380859375,
0.51318359375,
0.40380859375
],
[
0.1351318359375,
0.40087890625,
0.48779296875,
0.40087890625,
0.48779296875,
0.41455078125,
0.1351318359375,
0.41455078125
],
[
0.118408203125,
0.416259765625,
0.486083984375,
0.416259765625,
0.486083984375,
0.429931640625,
0.118408203125,
0.429931640625
],
[
0.51318359375,
0.410888671875,
0.87744140625,
0.410888671875,
0.87744140625,
0.421630859375,
0.51318359375,
0.421630859375
],
[
0.51318359375,
0.42578125,
0.880859375,
0.42578125,
0.880859375,
0.439453125,
0.51318359375,
0.439453125
],
[
0.118408203125,
0.431640625,
0.486083984375,
0.431640625,
0.486083984375,
0.4453125,
0.118408203125,
0.4453125
],
[
0.51123046875,
0.440673828125,
0.880859375,
0.441162109375,
0.880859375,
0.454833984375,
0.51123046875,
0.45361328125
],
[
0.118408203125,
0.448974609375,
0.48779296875,
0.448974609375,
0.48779296875,
0.4619140625,
0.118408203125,
0.4619140625
],
[
0.5146484375,
0.457763671875,
0.880859375,
0.457763671875,
0.880859375,
0.471435546875,
0.5146484375,
0.471435546875
],
[
0.1175537109375,
0.46435546875,
0.339111328125,
0.46435546875,
0.339111328125,
0.47802734375,
0.1175537109375,
0.47802734375
],
[
0.5146484375,
0.473876953125,
0.880859375,
0.473876953125,
0.880859375,
0.48681640625,
0.5146484375,
0.48681640625
],
[
0.1351318359375,
0.482177734375,
0.486083984375,
0.482177734375,
0.486083984375,
0.495361328125,
0.1351318359375,
0.495361328125
],
[
0.5146484375,
0.4912109375,
0.87890625,
0.4912109375,
0.87890625,
0.501953125,
0.5146484375,
0.501953125
],
[
0.118408203125,
0.498779296875,
0.486083984375,
0.498779296875,
0.486083984375,
0.5126953125,
0.118408203125,
0.5126953125
],
[
0.515625,
0.505859375,
0.880859375,
0.505859375,
0.880859375,
0.51953125,
0.515625,
0.51953125
],
[
0.1175537109375,
0.51416015625,
0.48779296875,
0.515625,
0.48779296875,
0.529296875,
0.1175537109375,
0.52783203125
],
[
0.5146484375,
0.5224609375,
0.623046875,
0.5224609375,
0.623046875,
0.53369140625,
0.5146484375,
0.53369140625
],
[
0.1158447265625,
0.5322265625,
0.45751953125,
0.53076171875,
0.45751953125,
0.54443359375,
0.1158447265625,
0.5458984375
],
[
0.51123046875,
0.54443359375,
0.85791015625,
0.54638671875,
0.85791015625,
0.56103515625,
0.51123046875,
0.55859375
],
[
0.11419677734375,
0.5654296875,
0.274658203125,
0.5654296875,
0.274658203125,
0.5791015625,
0.11419677734375,
0.5791015625
],
[
0.568359375,
0.56396484375,
0.62744140625,
0.56396484375,
0.62744140625,
0.576171875,
0.568359375,
0.576171875
],
[
0.51318359375,
0.58203125,
0.880859375,
0.58203125,
0.880859375,
0.595703125,
0.51318359375,
0.595703125
],
[
0.118408203125,
0.595703125,
0.48779296875,
0.595703125,
0.48779296875,
0.609375,
0.118408203125,
0.609375
],
[
0.51123046875,
0.5966796875,
0.87890625,
0.59716796875,
0.87890625,
0.61083984375,
0.51123046875,
0.60986328125
],
[
0.1175537109375,
0.6123046875,
0.48779296875,
0.61083984375,
0.48779296875,
0.62451171875,
0.1175537109375,
0.6259765625
],
[
0.51123046875,
0.61328125,
0.880859375,
0.61328125,
0.880859375,
0.62744140625,
0.51123046875,
0.62744140625
],
[
0.118408203125,
0.62744140625,
0.486083984375,
0.62744140625,
0.486083984375,
0.64111328125,
0.118408203125,
0.64111328125
],
[
0.509765625,
0.63134765625,
0.87890625,
0.63134765625,
0.87890625,
0.64111328125,
0.509765625,
0.64111328125
],
[
0.118408203125,
0.6435546875,
0.484375,
0.642578125,
0.484375,
0.65625,
0.118408203125,
0.65673828125
],
[
0.51123046875,
0.6455078125,
0.87890625,
0.646484375,
0.87890625,
0.66015625,
0.51123046875,
0.6591796875
],
[
0.118408203125,
0.66015625,
0.484375,
0.66015625,
0.484375,
0.67333984375,
0.118408203125,
0.67333984375
],
[
0.5146484375,
0.6611328125,
0.880859375,
0.6611328125,
0.880859375,
0.67578125,
0.5146484375,
0.67578125
],
[
0.12005615234375,
0.67578125,
0.486083984375,
0.67578125,
0.486083984375,
0.689453125,
0.12005615234375,
0.689453125
],
[
0.5146484375,
0.677734375,
0.880859375,
0.677734375,
0.880859375,
0.69140625,
0.5146484375,
0.69140625
],
[
0.12176513671875,
0.6923828125,
0.484375,
0.6923828125,
0.484375,
0.703125,
0.12176513671875,
0.703125
],
[
0.51123046875,
0.693359375,
0.87890625,
0.693359375,
0.87890625,
0.70703125,
0.51123046875,
0.70703125
],
[
0.1175537109375,
0.70703125,
0.486083984375,
0.70703125,
0.486083984375,
0.72021484375,
0.1175537109375,
0.72021484375
],
[
0.51318359375,
0.71142578125,
0.568359375,
0.71142578125,
0.568359375,
0.72265625,
0.51318359375,
0.72265625
],
[
0.118408203125,
0.7236328125,
0.484375,
0.7236328125,
0.484375,
0.73681640625,
0.118408203125,
0.73681640625
],
[
0.118408203125,
0.74072265625,
0.486083984375,
0.74072265625,
0.486083984375,
0.75439453125,
0.118408203125,
0.75439453125
],
[
0.509765625,
0.736328125,
0.81689453125,
0.736328125,
0.81689453125,
0.75,
0.509765625,
0.75
],
[
0.118408203125,
0.75732421875,
0.267822265625,
0.75732421875,
0.267822265625,
0.77099609375,
0.118408203125,
0.77099609375
],
[
0.51318359375,
0.7548828125,
0.880859375,
0.7548828125,
0.880859375,
0.7685546875,
0.51318359375,
0.7685546875
],
[
0.1351318359375,
0.77490234375,
0.486083984375,
0.77490234375,
0.486083984375,
0.7880859375,
0.1351318359375,
0.7880859375
],
[
0.51123046875,
0.77099609375,
0.87890625,
0.77099609375,
0.87890625,
0.78369140625,
0.51123046875,
0.78369140625
],
[
0.1175537109375,
0.79052734375,
0.48779296875,
0.79052734375,
0.48779296875,
0.80419921875,
0.1175537109375,
0.80419921875
],
[
0.51123046875,
0.7861328125,
0.880859375,
0.7861328125,
0.880859375,
0.7998046875,
0.51123046875,
0.7998046875
],
[
0.118408203125,
0.80712890625,
0.48779296875,
0.80712890625,
0.48779296875,
0.82080078125,
0.118408203125,
0.82080078125
],
[
0.51318359375,
0.802734375,
0.880859375,
0.802734375,
0.880859375,
0.81640625,
0.51318359375,
0.81640625
],
[
0.118408203125,
0.822265625,
0.490234375,
0.822265625,
0.490234375,
0.8359375,
0.118408203125,
0.8359375
],
[
0.51318359375,
0.82080078125,
0.87744140625,
0.82080078125,
0.87744140625,
0.83056640625,
0.51318359375,
0.83056640625
],
[
0.1158447265625,
0.83837890625,
0.486083984375,
0.83740234375,
0.486083984375,
0.8515625,
0.1158447265625,
0.8525390625
],
[
0.51318359375,
0.8349609375,
0.87744140625,
0.8349609375,
0.87744140625,
0.8486328125,
0.51318359375,
0.8486328125
],
[
0.1175537109375,
0.85498046875,
0.486083984375,
0.85498046875,
0.486083984375,
0.8681640625,
0.1175537109375,
0.8681640625
],
[
0.53076171875,
0.85693359375,
0.8818359375,
0.859375,
0.8818359375,
0.87255859375,
0.53076171875,
0.87060546875
],
[
0.1175537109375,
0.87158203125,
0.1376953125,
0.87158203125,
0.1376953125,
0.88525390625,
0.1175537109375,
0.88525390625
],
[
0.51318359375,
0.87255859375,
0.87744140625,
0.87255859375,
0.87744140625,
0.8828125,
0.51318359375,
0.8828125
],
[
0.1368408203125,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.1368408203125,
0.90283203125
],
[
0.5146484375,
0.88427734375,
0.87890625,
0.88427734375,
0.87890625,
0.89453125,
0.5146484375,
0.89453125
],
[
0.51123046875,
0.89453125,
0.8818359375,
0.89453125,
0.8818359375,
0.90771484375,
0.51123046875,
0.90771484375
],
[
0.118408203125,
0.90625,
0.48779296875,
0.90625,
0.48779296875,
0.91943359375,
0.118408203125,
0.91943359375
],
[
0.5087890625,
0.90478515625,
0.7314453125,
0.90673828125,
0.7314453125,
0.92041015625,
0.5087890625,
0.91845703125
]
] |
[
"数据集中的释义不涉及句子间的多样性，但我",
"3OurCoRPGModel",
"们的模型可以通过ALBERT(Lanetal.,2020)",
"3.1",
"Overview and Notations",
"生成的连贯性关系图学习句子之间的连贯性关",
"3.1.1",
"Model Overview",
"图1显示了我们模型的概述，该模型由句子编",
"系，并利用学习到的句子连贯性感知表示对它",
"们进行重新排序，同时保持生成文档的良好连",
"贯性",
"码器，图形GRU和解码器组成。给定一个输入",
"我们的模型由三部分组成：句子编码器，图",
"文档，我们使用句子编码器来获取文档中每个",
"形GRU和解码器。句子编码器仅对文档中的",
"句子的表示形式，而忽略句子的位置信息。我",
"每个句子进行单独编码。我们提出了图GRU，",
"们为句子构建一个连贯关系图，并使用图GRU",
"它结合了图注意力（Velickovicetal.，2018）和",
"来获得每个句子的连贯感知表示。句子编码器",
"GRU，以捕捉一致性关系信息。最后，将图形",
"和图形GRU的输出由解码器获取，用于生成",
"GRU和句子编码器的输出连接在一起，并用",
"连贯的释义文档。请注意，我们没有具有句子",
"作解码器的输人，以生成释义。进行了广泛的",
"间多样性的文档释义数据集。相反，我们使用",
"评估，我们的模型在自动评估和人工评估中获",
"句子释义模型来构建仅具有句子内多样性的伪",
"得了大多数指标的最佳分数",
"文档释义数据集（i。e。，每个句子内的词汇或",
"我们的工作贡献总结如下：",
"句法多样性）。我们的模型在此数据集上进行",
"了训练，以便能够通过修改和排列多个输入句",
"1）据我们所知，我们是第一个探索文档级",
"释义生成问题的人，并指出了文档级释义与句",
"子来重建连贯的文档，而无需使用句子的原始",
"子级释义之间的区别",
"位置信息1。换句话说，我们模型的输入可以",
"看作只是一组没有顺序的句子。实现这一目标",
"2）我们提出了一种新的模型CoRPG，以解",
"决句子重写和重新排序的文档级释义生成。我",
"的关键是图形GRU学习的句子的一致性感知",
"表示。在测试过程中，我们的模型可以根据学",
"们的模型可以利用图GRU来学习句子的连贯",
"习到的句子的连贯感知和语义表示来重新组织",
"感知表示，并重新排列输人句子，以改善生成",
"文本。与输入文档相比，输出文档很可能具有",
"的文档释义的句子间多样性",
"不同的句子顺序和排列方式，因为除了原始句",
"3）自动评估和人类评估均表明，我们的模",
"子顺序之外，通常还有不同的合理方式来排列",
"型可以生成具有高度多样性，语义相关性和",
"一组句子。数据集和模型模块的详细信息将在",
"连贯性的文档释义。我们的代码在https：//",
"下一节中给出",
"github.com/L-Zhe/CoRPG上公开可用",
"3.1.2Pseudo Document-Level Paraphrase",
"2Related Work",
"Dataset",
"我们的模型将释义视为单语翻译任务。给定文",
"随着深度学习的发展，大多数释义模型都基于",
"档D=SI,S?,SN，其中N是文档中",
"seq2seq模型。Prakashetal.（2016）利用堆叠残",
"的句子总数，Si是文档中的第i个句子。由",
"差LSTM网络生成释义。Guptaetal.(2018）发",
"干缺之黄金文档释义数据集，我们利用现成",
"现诸如变分自动编码器之类的深度生成模型可",
"的句子释义模型逐句生成伪文档释义Dp=",
"以显着提高释义的质量。Li etal.(2019)提出",
"[S,,S?..·,SN]，其中S,是通过将Si与句",
"了DNPG将句子分解为句子级模式和短语级",
"子释义模型进行释义而获得的。然后，我们将",
"模式，以使神经释义生成更加可控。Goyaland",
"D设置为输人，将D设置为目标来训练我们",
"Durrett(2020）使用句法转换来轻柔地“重新排",
"的模型",
"序”源句子，并引导神经模型生成更多样化的",
"释义。Kazemnejadetal.(2020）探索通过编辑原",
"3.1.3Coherence Relationship Graph",
"始句子来生成释义",
"有许多作品专注于文本连贯性，例如NCOH",
"此外，还有另一种生成释义的方法，称为",
"(Moon et al., 2019)和 CohEval (Mohiuddin et al.",
"“旋转”，它利用反向翻译来引入多样性。最",
"2020）。许多预先训练的模型还将文本一致",
"近，Mallinsonetal.（2017）用神经机器翻译重",
"性作为子任务引人，以提高泛化能力。例如",
"新审视了这种方法，以提高释义质量。Wiet-",
"Devlinetal.（2019）使用下二个句子预测（NSP",
"ingandGimpel（2018）利用双向翻译模型构建",
"任务来训练BERT；Lanetal.（2020）提出了",
"paraNMT，这是一个非常大的句子级释义数据",
"'If we use the original positional information of input sen-",
"集",
"tences for training, the model can simply output a documen",
"以上所有工作都集中在句子级释义生成上，",
"with the samepositions of these sentences.During testing",
"the model cannot generate document paraphrase with sen-",
"据我们所知，没有关于文档级释义生成的研究",
"tence reordering and rearrangement."
] |
{
"arxiv_id": "2109.07095",
"language": "zh",
"page_idx": 2
}
|
||
2109.07095-zh-4
|
2109.07095-zh-4.jpg
|
<FigureHere>
Figure 2: The structure of graph GRU, which is a stack of $L_{g}$ identical layers. Each layer includes a multi-head graph attentior
block, a GRU cell and a layer normalization.
(3)
$$s_{ij}=\left(r_{i}\boldsymbol{W}_{Q}\right)\left(g_{j}\boldsymbol{W}_{K}\right)^{\top}\\\alpha_{ij}=\frac{\exp\left(s_{ij}\right)}{\sum_{\mathbb{G}(i,k)=1}\exp\left(s_{ik}\right)}$$
性信息进行编码。每个新层将增加一个编码序列的长度。因此，图GRU层$L_g$的数量不是固定的数字，而是等于文档中的句子数。我们将最后一层的输出作为其最终输出
遵循双向 RNN 的思想，我们采用两个不共享参数的图形 GRU 来聚合两个方向的相干性信息。我们将$\mathbb{G}$发送到正向图 GRU 中，将$\mathbb{G}^\top$ 发送到反向图 GRU 中，并分别获得它们的输出$\vec{G}$ 和$\vec{G}$。最后，我们将两个方向的输出组合为双图 GRU 的最终输出
, 其中$W_Q,W_K\in\mathbb{R}^d_{model}\times d_u$ 是可学习的参数。请注意，在一致性关系图中存在 sink节点。如果第$i$个节点是接收器节点，则$\mathbb{G}(i,\cdot)=0$ 。对于所有 sink 节点，我们将其所有注意力权重设置为$\alpha_i.=0$
为了获得更好的性能，我们在图注意中引入
了多头操作
$$G_o=\stackrel{\rightarrow}{G}+\stackrel{\leftarrow}{G}$$
(7)
其中$G_o=\left\{g_1,g_2,\cdots,g_N\right\},g_i\in\mathbb{R}^{d_{model}}$是包
含连贯关系信息的句子向量
## 3.4 Decoder
(4)
$$\begin{aligned}\hat{g}_{i}&=\mathrm{GAT}(r_{i},G)\quad r_{i}\in D,\\\mathrm{Head}_{j}&=(\hat{g}_{1},\hat{g}_{2},\cdots,\hat{g}_{N})\\\mathrm{MHGAT}(D_{r},G)&=\left(\prod_{j=1}^{H}\mathrm{Head}_{j}\right)\boldsymbol{W}_{o}\end{aligned}$$
我们利用 Transformer 解码器作为我们的解码器。首先，我们结合句子编码器和图形 GRU 的输出
其中$H$是头号，$\parallel$是连接操作$,W_o\in$
$\mathbb{R}^{H*d_u\times d_{model}}$
$G_{l-1}$包含长度为$l-1$的节点的一致性信息我们采用多头图注意力来聚合每个节点的前驱节点信息，并将聚合向量作为隐藏信息发送到GRU 单元中。详情如下：
(8)
$$\begin{aligned}&S_{c}^{i}=\left[h_{j}^{i}\left|\right|g_{i}\right]_{j=1}^{n}\\&\tilde{d}_{c}=\left[S_{c}^{1},S_{c}^{2},\cdots,S_{c}^{N}\right]\boldsymbol{W}_{c}+\boldsymbol{b}_{c}\\&d_{c}=\mathrm{LayerNorm}\left(\mathrm{ReLU}\left(\tilde{d}_{c}\right)\right)\end{aligned}$$
$$\begin{aligned}&\bar{G}_{l}=\mathrm{MHGAT}\left(D_{r},G_{l-1}\right)\\&z_{t}=\sigma\left(\left[\bar{G}_{l}\left\|D_{r}\right]\boldsymbol{W}_{z}\right)\\&r_{t}=\sigma\left(\left[\bar{G}_{l}\left\|D_{r}\right]\boldsymbol{W}_{r}\right)\right)\\&\tilde{G}_{l}=\mathrm{tanh}\left(\left[r_{t}\otimes\bar{G}_{l}\left\|D_{r}\right]\boldsymbol{W}_{m}\right)\right)\\&\hat{G}_{l}=(1-z_{t})\otimes\bar{G}_{l}+z_{t}\otimes\tilde{G}_{l}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\hat{G}_{l}=(1-z_{t})\otimes\bar{G}_{l}+z_{t}\otimes\tilde{G}_{l}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}\\&\text{一}$$
(5)
其中$W_c\in\mathbb{R}^{2d_{model}\times d_{model}}$、$b_c\in\mathbb{R}^{d_{model}}$、$g_i\in$ $G_o$ 。为了避免过拟合，我们在 ReLU 函数后添加 dropout。上面的组合操作可以看作是将句子间的连贯关系信息引入到每个句子嵌人矩阵中
然后，我们将$d_c$发送到解码器中以指导生成。我们添加了复制机制(See et al.,2017)。我们利用最后一个解码器层中所有头部的平均注意力权重作为复制概率来计算最终输出的概率
### 3.5 Diversity Coefficient
,其中$\sigma$是 S 形激活函数，$\otimes$是矩阵之间的元素乘积，而$W_z,W_r,W_m\in\mathbb{R}^{2d_{model}\times d_{model}}$
最后，我们利用层规范化(Ba et al.,2016) 来
规范化$\hat{G}_l$。第$l$图 GRU 层的输出如下：
在实验过程中，我们发现释义模型倾向于复制原始句子。因此，与原始句子释义数据集相比， 我们由句子级释义模型创建的伪文档释义数据集在词汇和句法上的多样性都较小。为了解决这个问题，我们引入多样性系数来更加关注N-gram 短语的多样性
$$G_l=\text{LayerNorm}(\hat{G}_l)$$
(6)
与循环遍历序列中每个令牌的传统 RNN 模型不同，我们的图 GRU 通过多层传播对一致
|
[
[
0.496337890625,
0.0843505859375,
0.5263671875,
0.0843505859375,
0.5263671875,
0.0968017578125,
0.496337890625,
0.0968017578125
],
[
0.466796875,
0.11102294921875,
0.6572265625,
0.105712890625,
0.6591796875,
0.1353759765625,
0.468505859375,
0.14013671875
],
[
0.359375,
0.122314453125,
0.399658203125,
0.122314453125,
0.399658203125,
0.1300048828125,
0.359375,
0.1300048828125
],
[
0.59619140625,
0.122314453125,
0.64990234375,
0.122314453125,
0.64990234375,
0.1300048828125,
0.59619140625,
0.1300048828125
],
[
0.74072265625,
0.122314453125,
0.765625,
0.122314453125,
0.765625,
0.1300048828125,
0.74072265625,
0.1300048828125
],
[
0.52490234375,
0.157958984375,
0.5498046875,
0.157958984375,
0.5498046875,
0.1728515625,
0.52490234375,
0.1728515625
],
[
0.118408203125,
0.18701171875,
0.87890625,
0.18701171875,
0.87890625,
0.19775390625,
0.118408203125,
0.19775390625
],
[
0.1158447265625,
0.19775390625,
0.385498046875,
0.1995849609375,
0.385498046875,
0.210205078125,
0.1158447265625,
0.2078857421875
],
[
0.5146484375,
0.2369384765625,
0.880859375,
0.2369384765625,
0.880859375,
0.25048828125,
0.5146484375,
0.25048828125
],
[
0.202392578125,
0.25048828125,
0.387939453125,
0.2476806640625,
0.387939453125,
0.265380859375,
0.202392578125,
0.268310546875
],
[
0.51318359375,
0.2529296875,
0.880859375,
0.2529296875,
0.880859375,
0.2666015625,
0.51318359375,
0.2666015625
],
[
0.46337890625,
0.26953125,
0.489501953125,
0.26953125,
0.489501953125,
0.281982421875,
0.46337890625,
0.281982421875
],
[
0.51318359375,
0.26953125,
0.880859375,
0.26953125,
0.880859375,
0.283203125,
0.51318359375,
0.283203125
],
[
0.486083984375,
0.27490234375,
0.5029296875,
0.27490234375,
0.5029296875,
0.27783203125,
0.486083984375,
0.27783203125
],
[
0.291259765625,
0.270751953125,
0.361083984375,
0.270751953125,
0.361083984375,
0.28515625,
0.291259765625,
0.28515625
],
[
0.202392578125,
0.28076171875,
0.252685546875,
0.28515625,
0.2509765625,
0.297607421875,
0.20068359375,
0.292236328125
],
[
0.2476806640625,
0.2861328125,
0.403076171875,
0.2861328125,
0.403076171875,
0.30712890625,
0.2476806640625,
0.30712890625
],
[
0.5146484375,
0.28515625,
0.77099609375,
0.28515625,
0.77099609375,
0.298583984375,
0.5146484375,
0.298583984375
],
[
0.53076171875,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.53076171875,
0.3154296875
],
[
0.1318359375,
0.30810546875,
0.48779296875,
0.30810546875,
0.48779296875,
0.328857421875,
0.1318359375,
0.328857421875
],
[
0.5146484375,
0.317626953125,
0.880859375,
0.317626953125,
0.880859375,
0.330810546875,
0.5146484375,
0.330810546875
],
[
0.118408203125,
0.328857421875,
0.489501953125,
0.328857421875,
0.489501953125,
0.342529296875,
0.118408203125,
0.342529296875
],
[
0.51318359375,
0.333251953125,
0.8759765625,
0.33203125,
0.8759765625,
0.3466796875,
0.51318359375,
0.347412109375
],
[
0.118408203125,
0.344482421875,
0.48779296875,
0.344482421875,
0.48779296875,
0.359375,
0.118408203125,
0.359375
],
[
0.5146484375,
0.349853515625,
0.87890625,
0.349853515625,
0.87890625,
0.364013671875,
0.5146484375,
0.364013671875
],
[
0.12176513671875,
0.364013671875,
0.1376953125,
0.364013671875,
0.1376953125,
0.372314453125,
0.12176513671875,
0.372314453125
],
[
0.1318359375,
0.361083984375,
0.484375,
0.361083984375,
0.484375,
0.374755859375,
0.1318359375,
0.374755859375
],
[
0.51318359375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.51318359375,
0.3818359375
],
[
0.118408203125,
0.377685546875,
0.25439453125,
0.377685546875,
0.25439453125,
0.391357421875,
0.118408203125,
0.391357421875
],
[
0.5146484375,
0.38427734375,
0.7255859375,
0.38427734375,
0.7255859375,
0.3984375,
0.5146484375,
0.3984375
],
[
0.1368408203125,
0.393798828125,
0.483642578125,
0.393798828125,
0.483642578125,
0.407958984375,
0.1368408203125,
0.407958984375
],
[
0.12176513671875,
0.410888671875,
0.2091064453125,
0.410888671875,
0.2091064453125,
0.4228515625,
0.12176513671875,
0.4228515625
],
[
0.6455078125,
0.413330078125,
0.74658203125,
0.410888671875,
0.74658203125,
0.429931640625,
0.6455078125,
0.431640625
],
[
0.85791015625,
0.41748046875,
0.8818359375,
0.41748046875,
0.8818359375,
0.43115234375,
0.85791015625,
0.43115234375
],
[
0.509765625,
0.438232421875,
0.8818359375,
0.43701171875,
0.8818359375,
0.454833984375,
0.509765625,
0.4560546875
],
[
0.2418212890625,
0.449462890625,
0.453369140625,
0.448974609375,
0.453369140625,
0.463134765625,
0.2418212890625,
0.46435546875
],
[
0.5146484375,
0.4560546875,
0.73486328125,
0.4560546875,
0.73486328125,
0.4697265625,
0.5146484375,
0.4697265625
],
[
0.20654296875,
0.466064453125,
0.4072265625,
0.46728515625,
0.4072265625,
0.485107421875,
0.20654296875,
0.48388671875
],
[
0.462646484375,
0.485107421875,
0.48779296875,
0.485107421875,
0.48779296875,
0.49755859375,
0.462646484375,
0.49755859375
],
[
0.51123046875,
0.482177734375,
0.62646484375,
0.482177734375,
0.62646484375,
0.496337890625,
0.51123046875,
0.496337890625
],
[
0.298828125,
0.488037109375,
0.32080078125,
0.488037109375,
0.32080078125,
0.501953125,
0.298828125,
0.501953125
],
[
0.3046875,
0.50439453125,
0.38037109375,
0.50439453125,
0.38037109375,
0.52099609375,
0.3046875,
0.52099609375
],
[
0.51123046875,
0.501953125,
0.87890625,
0.5029296875,
0.87890625,
0.51708984375,
0.51123046875,
0.5166015625
],
[
0.391357421875,
0.505859375,
0.424072265625,
0.505859375,
0.424072265625,
0.5185546875,
0.391357421875,
0.5185546875
],
[
0.1275634765625,
0.5048828125,
0.2822265625,
0.5048828125,
0.2822265625,
0.5185546875,
0.1275634765625,
0.5185546875
],
[
0.29296875,
0.5224609375,
0.32666015625,
0.5224609375,
0.32666015625,
0.53515625,
0.29296875,
0.53515625
],
[
0.51318359375,
0.5185546875,
0.87890625,
0.5185546875,
0.87890625,
0.533203125,
0.51318359375,
0.533203125
],
[
0.51318359375,
0.5361328125,
0.57080078125,
0.5361328125,
0.57080078125,
0.55029296875,
0.51318359375,
0.55029296875
],
[
0.1175537109375,
0.54638671875,
0.484375,
0.54638671875,
0.484375,
0.56005859375,
0.1175537109375,
0.56005859375
],
[
0.5693359375,
0.55419921875,
0.7119140625,
0.56103515625,
0.7080078125,
0.59423828125,
0.56494140625,
0.58740234375
],
[
0.1158447265625,
0.56103515625,
0.2166748046875,
0.55810546875,
0.2166748046875,
0.572265625,
0.1158447265625,
0.576171875
],
[
0.1368408203125,
0.580078125,
0.483642578125,
0.580078125,
0.483642578125,
0.59326171875,
0.1368408203125,
0.59326171875
],
[
0.1175537109375,
0.59423828125,
0.484375,
0.595703125,
0.484375,
0.60986328125,
0.1175537109375,
0.609375
],
[
0.57421875,
0.59326171875,
0.82373046875,
0.59326171875,
0.82373046875,
0.61083984375,
0.57421875,
0.61083984375
],
[
0.82373046875,
0.60107421875,
0.84228515625,
0.599609375,
0.84228515625,
0.60498046875,
0.82373046875,
0.60546875
],
[
0.8564453125,
0.595703125,
0.880859375,
0.595703125,
0.880859375,
0.60791015625,
0.8564453125,
0.60791015625
],
[
0.12005615234375,
0.6123046875,
0.484375,
0.6123046875,
0.484375,
0.6259765625,
0.12005615234375,
0.6259765625
],
[
0.5693359375,
0.61328125,
0.81787109375,
0.6162109375,
0.81787109375,
0.63818359375,
0.5693359375,
0.63525390625
],
[
0.1175537109375,
0.62744140625,
0.316650390625,
0.62744140625,
0.316650390625,
0.642578125,
0.1175537109375,
0.642578125
],
[
0.509765625,
0.64111328125,
0.88330078125,
0.64453125,
0.88330078125,
0.6650390625,
0.509765625,
0.662109375
],
[
0.172119140625,
0.65380859375,
0.3779296875,
0.65478515625,
0.3779296875,
0.6728515625,
0.172119140625,
0.67138671875
],
[
0.51318359375,
0.66455078125,
0.880859375,
0.66455078125,
0.880859375,
0.677734375,
0.51318359375,
0.677734375
],
[
0.177978515625,
0.677734375,
0.359375,
0.677734375,
0.359375,
0.69873046875,
0.177978515625,
0.69873046875
],
[
0.51318359375,
0.68115234375,
0.87890625,
0.68115234375,
0.87890625,
0.69482421875,
0.51318359375,
0.69482421875
],
[
0.51318359375,
0.69677734375,
0.880859375,
0.6953125,
0.880859375,
0.71044921875,
0.51318359375,
0.71142578125
],
[
0.177978515625,
0.705078125,
0.358642578125,
0.705078125,
0.358642578125,
0.72705078125,
0.177978515625,
0.72705078125
],
[
0.46337890625,
0.71044921875,
0.489501953125,
0.71044921875,
0.489501953125,
0.72265625,
0.46337890625,
0.72265625
],
[
0.5146484375,
0.71435546875,
0.53076171875,
0.71435546875,
0.53076171875,
0.724609375,
0.5146484375,
0.724609375
],
[
0.172119140625,
0.73388671875,
0.426513671875,
0.73388671875,
0.426513671875,
0.7548828125,
0.172119140625,
0.7548828125
],
[
0.5322265625,
0.728515625,
0.87890625,
0.728515625,
0.87890625,
0.7421875,
0.5322265625,
0.7421875
],
[
0.515625,
0.74462890625,
0.880859375,
0.74462890625,
0.880859375,
0.75830078125,
0.515625,
0.75830078125
],
[
0.172119140625,
0.7587890625,
0.399658203125,
0.76025390625,
0.399658203125,
0.77783203125,
0.172119140625,
0.77685546875
],
[
0.5146484375,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.77490234375,
0.5146484375,
0.77490234375
],
[
0.51123046875,
0.77685546875,
0.880859375,
0.775390625,
0.880859375,
0.79052734375,
0.51123046875,
0.79150390625
],
[
0.1259765625,
0.79052734375,
0.1368408203125,
0.79052734375,
0.1368408203125,
0.79443359375,
0.1259765625,
0.79443359375
],
[
0.1318359375,
0.78515625,
0.486083984375,
0.78515625,
0.486083984375,
0.798828125,
0.1318359375,
0.798828125
],
[
0.1158447265625,
0.80029296875,
0.456787109375,
0.79638671875,
0.456787109375,
0.81396484375,
0.1158447265625,
0.818359375
],
[
0.51123046875,
0.8017578125,
0.72119140625,
0.802734375,
0.72119140625,
0.81787109375,
0.51123046875,
0.81640625
],
[
0.1368408203125,
0.81787109375,
0.484375,
0.81787109375,
0.484375,
0.83056640625,
0.1368408203125,
0.83056640625
],
[
0.5146484375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.118408203125,
0.83447265625,
0.4375,
0.83447265625,
0.4375,
0.84716796875,
0.118408203125,
0.84716796875
],
[
0.51318359375,
0.8388671875,
0.87744140625,
0.84033203125,
0.87744140625,
0.85498046875,
0.51318359375,
0.85400390625
],
[
0.51318359375,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.2166748046875,
0.865234375,
0.385498046875,
0.865234375,
0.385498046875,
0.8798828125,
0.2166748046875,
0.8798828125
],
[
0.462646484375,
0.8662109375,
0.48779296875,
0.8662109375,
0.48779296875,
0.8798828125,
0.462646484375,
0.8798828125
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.1368408203125,
0.88818359375,
0.48779296875,
0.88720703125,
0.48779296875,
0.90185546875,
0.1368408203125,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.68701171875,
0.90478515625,
0.68701171875,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"G.t",
"",
"GAT",
"GAT",
"rmal",
"D,Y",
"Figure 2:The structure of graph GRU, which is a stack of Lgidentical layers.Each layer includes a multi-head graph attention",
"block,aGRU cell andalayernormalization",
"性信息进行编码。每个新层将增加一个编码序",
"Sij = (r; Wo) (g; Wk) T",
"列的长度。因此，图GRU层L。的数量不是固",
"(3)",
"定的数字，而是等于文档中的句子数。我们将",
"",
"exp (sii)",
"Qi=",
"ZG(i,b)=1 exp (sik)",
"最后一层的输出作为其最终输出",
"遵循双向RNN的思想，我们采用两个不共",
"其中Wo,WkERd.modelxdu是可学习的参",
"享参数的图形GRU来聚合两个方向的相干性",
"数。请注意，在一致性关系图中存在sink节点。",
"信息。我们将G发送到正向图GRU中，将GT",
"如果第i个节点是接收器节点，则G(i，）=0",
"发送到反向图GRU中，并分别获得它们的输",
"",
"对于所有sink节点，我们将其所有注意力权",
"出G和G。最后，我们将两个方向的输出组",
"重设置为α.=0",
"合为双图GRU的最终输出",
"为了获得更好的性能，我们在图注意中引人",
"了多头操作",
"G。=G+G",
"(7)",
"其中G。={g1,92,·*·,9N}，9iE Rdmodel是包",
"gi=GAT(ri,G) riED,",
"含连贯关系信息的句子向量",
"Head, = (91, 92,.. , 9N)",
"(4)",
"3.4Decoder",
"H",
"I Head,",
"我们利用Transformer解码器作为我们的解码",
"W.",
"MHGAT(Dr,G) =",
"(i=1",
"器。首先，我们结合句子编码器和图形GRU",
"的输出",
"其中H是头号，Ⅱ是连接操作，W。E",
"s = [5l1] -]",
"RH*duxdmode",
"Gl-1包含长度为-1的节点的一致性信息。",
"我们采用多头图注意力来聚合每个节点的前驱",
"de = [st, se,..., sn] We+be",
"",
"(8)",
"节点信息，并将聚合向量作为隐藏信息发送到",
"de = LayerNorm (ReLU (ae))",
"GRU单元中。详情如下：",
"其中 WeE R2dmodel xdmodel 、be E Rd'model 、gi E",
"G = MHGAT(Dr,Gl-1)",
"G。。为了避免过拟合，我们在ReLU函数后添",
"zt = o ([GillD,W.)",
"加dropout。上面的组合操作可以看作是将句",
"子间的连贯关系信息引入到每个句子嵌人矩阵",
"rt = ([GlD.|w.)",
"(5)",
"中",
"Gi= tanh ([rt @GillD,Wm)",
"然后，我们将d发送到解码器中以指导生",
"成。我们添加了复制机制（Seeetal.,2017）。我",
"G = (1 - zt) @GI + zt Gi",
"们利用最后一个解码器层中所有头部的平均注",
"意力权重作为复制概率来计算最终输出的概率",
"",
"其中是S形激活函数，是矩阵之间的元",
"素乘积，而Wz,W,WmER2dmodelxdmodel",
"3.5Diversity Coefficient",
"最后，我们利用层规范化(Baetal.,2016)来",
"在实验过程中，我们发现释义模型倾向于复制",
"规范化Gi。第1图GRU层的输出如下：",
"原始句子。因此，与原始句子释义数据集相比",
"我们由句子级释义模型创建的伪文档释义数",
"Gi = LayerNorm(G)",
"(6)",
"据集在词汇和句法上的多样性都较小。为了解",
"与循环遍历序列中每个令牌的传统RNN模",
"决这个问题，我们引入多样性系数来更加关注",
"型不同，我们的图GRU通过多层传播对一致",
"N-gram短语的多样性"
] |
{
"arxiv_id": "2109.07095",
"language": "zh",
"page_idx": 4
}
|
||
1904.10898-zh-0
|
1904.10898-zh-0.jpg
|
ViDeNN: Deep Blind Video Denoising
Jan van Gemert
Computer Vision lab
Delft University of Technology
http://jvgemert.github.io/
Michele Claus
Computer Vision lab
Delft University of Technology
claus.michele@hotmail.it
<FigureHere>
Figure 1: ViDeNN approach to Video Denoising: combining two networks performing first Single Frame Spatial Denoising and subsequently Temporal Denoising over a window of three frames, all in a single feedforward process.
# $\operatorname{Abstract}$
We propose ViDeNN: a CNN for Video Denoising without prior knowledge on the noise distribution (blind denoising). The CNN architecture uses a combination of spatial and temporal filtering, learning to spatially denoise the frames first and at the same time how to combine their temporal information, handling ob- ${\text{jects motion, brightness changes, low-light conditions}}$ and temporal inconsistencies. We demonstrate the importance of the data used for CNNs training, creating for this purpose a specific dataset for low-light conditions. $\hat{\text{We test ViDeNN on common benchmarks and}}$ on self-collected data, achieving good results comparable with the state-of-the-art.
## 1. Introduction
图像和视频去噪旨在从可用的噪声观测$Y$中获得原始信号 $X$ 。噪声影响感知的视觉质量，但也分割$\mathbb{E}[1]$ 和压缩$\mathbb{E}[2]$使去噪成为重要的一步。以$X$为原始信号，$N$为噪声，$Y$为可用的噪声观测值，对于加性类型的噪声，噪声退化模型可以描述为$Y=X+N$ 。在弱光条件下，噪声与信号有关，在黑暗区域更敏感，建模为$Y=H(X)+N$ ,以$H$为退化函数
成像噪声是由于热效应，传感器缺陷或弱光引起的。手动调整多个滤波器参数是为每个增益级别优化新相机的质量和带宽的基础，需要花费大量时间和精力。在这里，我们使用 CNN 自动执行去噪过程，以实现灵活高效的视频去噪，能够盲目去除噪声。在“盲”条件下工作的噪声去除算法在现实世界中至关重要，在这种情况下，颜色和光线条件可能会突然变化，从而为每个帧产生不同的噪声分布
基于统计模型的解决方案包括马尔可夫随机场模型[3] ,梯度模型[4] ,稀疏模型[5]和当前在 BM3D [6] , LSSC[ 7] , NCSR [ 8] 和 WNNMAX6 等最新技术中使用的非局部自相似性(NSS)。尽管它们实现了可观的去噪性能，但大多数算法都有一些缺点。首先，它们通常用于处理特定的噪声模型和水平，限制了它们在盲降噪中的使用。其次，它们涉及耗时的手动优化程序
在图像去噪方面已经做了很多工作，而专门为视频设计的算法却很少。视频去噪的关键假设是视频帧是强相关的。最基本的视频去噪技术包括各种后续帧的时间平均值。虽然为稳定的场景提供了出色的效果，但它会模糊运动，导致伪影。VBM4D 方法[10] 是视频去噪的最新技术。它通过搜索相似的补丁来扩展 BM3DE [6]单幅图像去噪，不仅在空间域，而且在时间域。在更多帧中搜索类似的补丁会大大增加处理时间
在本文中，我们提出了 ViDeNN,如图E1 所示：用于盲视频去噪的卷积神经网络，能够在没有噪声模型和视频内容的先验知识的情况下对视频进行去噪。为了进行比较，已经在公开可用的视频和自我捕获的视频上进行了实验。。视频，弱光数据集和代码将发布在项目的 GitHub 页面上。
我们工作的主要贡献是：(i)一种新颖的 CNN 架构， 能够对视频进行盲降噪，将多个帧的空间和时间信息与一个单一的前馈过程相结合；(ii)在弱光条件下对加性高斯白噪声和真实数据进行灵活性测试；(iii)在具有挑战性的情况下对运动的鲁棒性；(iv)用于特定博世安全摄像机的新的低光数据集，具有无噪音和
用于图像去噪的
### 2. Related Work
CNN。从2008年Jain 和 Seung [11]的 CNN 图像去噪工作中，由于更多的计算能力和高质量的数据集，
1
|
[
[
0.293212890625,
0.1319580078125,
0.67626953125,
0.1337890625,
0.67626953125,
0.1514892578125,
0.293212890625,
0.149658203125
],
[
0.26953125,
0.184326171875,
0.38818359375,
0.184326171875,
0.38818359375,
0.1988525390625,
0.26953125,
0.1988525390625
],
[
0.57275390625,
0.1817626953125,
0.71337890625,
0.1817626953125,
0.71337890625,
0.199462890625,
0.57275390625,
0.199462890625
],
[
0.2418212890625,
0.2008056640625,
0.416748046875,
0.2008056640625,
0.416748046875,
0.21533203125,
0.2418212890625,
0.21533203125
],
[
0.5537109375,
0.2008056640625,
0.73046875,
0.2008056640625,
0.73046875,
0.21533203125,
0.5537109375,
0.21533203125
],
[
0.19775390625,
0.21533203125,
0.455810546875,
0.2171630859375,
0.455810546875,
0.23486328125,
0.19775390625,
0.23291015625
],
[
0.51123046875,
0.21533203125,
0.76953125,
0.2171630859375,
0.76953125,
0.23486328125,
0.51123046875,
0.23291015625
],
[
0.2484130859375,
0.2392578125,
0.406982421875,
0.2392578125,
0.406982421875,
0.25,
0.2484130859375,
0.25
],
[
0.5556640625,
0.238037109375,
0.7294921875,
0.238037109375,
0.7294921875,
0.251953125,
0.5556640625,
0.251953125
],
[
0.250732421875,
0.285888671875,
0.30224609375,
0.285888671875,
0.30224609375,
0.296630859375,
0.250732421875,
0.296630859375
],
[
0.340576171875,
0.290283203125,
0.408447265625,
0.291748046875,
0.408447265625,
0.302978515625,
0.340576171875,
0.3017578125
],
[
0.52392578125,
0.2890625,
0.88818359375,
0.2890625,
0.88818359375,
0.302978515625,
0.52392578125,
0.302978515625
],
[
0.226318359375,
0.2998046875,
0.2442626953125,
0.2998046875,
0.2442626953125,
0.30810546875,
0.226318359375,
0.30810546875
],
[
0.146240234375,
0.292236328125,
0.1968994140625,
0.292236328125,
0.1968994140625,
0.2998046875,
0.146240234375,
0.2998046875
],
[
0.2418212890625,
0.2998046875,
0.2548828125,
0.2998046875,
0.2548828125,
0.30810546875,
0.2418212890625,
0.30810546875
],
[
0.410888671875,
0.304931640625,
0.4150390625,
0.3125,
0.41259765625,
0.313232421875,
0.406982421875,
0.306884765625
],
[
0.5048828125,
0.3037109375,
0.8896484375,
0.3037109375,
0.8896484375,
0.317626953125,
0.5048828125,
0.317626953125
],
[
0.273681640625,
0.318115234375,
0.3095703125,
0.318115234375,
0.3095703125,
0.330810546875,
0.273681640625,
0.330810546875
],
[
0.50341796875,
0.318115234375,
0.890625,
0.3193359375,
0.890625,
0.333251953125,
0.50341796875,
0.33203125
],
[
0.5048828125,
0.335205078125,
0.890625,
0.335205078125,
0.890625,
0.348388671875,
0.5048828125,
0.348388671875
],
[
0.50634765625,
0.3515625,
0.88818359375,
0.3515625,
0.88818359375,
0.3623046875,
0.50634765625,
0.3623046875
],
[
0.50634765625,
0.3662109375,
0.88818359375,
0.3662109375,
0.88818359375,
0.376953125,
0.50634765625,
0.376953125
],
[
0.5048828125,
0.380126953125,
0.890625,
0.380126953125,
0.890625,
0.393310546875,
0.5048828125,
0.393310546875
],
[
0.50341796875,
0.395751953125,
0.66650390625,
0.395751953125,
0.66650390625,
0.4091796875,
0.50341796875,
0.4091796875
],
[
0.52392578125,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.52392578125,
0.423583984375
],
[
0.08087158203125,
0.42041015625,
0.46728515625,
0.42041015625,
0.46728515625,
0.434326171875,
0.08087158203125,
0.434326171875
],
[
0.5048828125,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.439453125,
0.5048828125,
0.439453125
],
[
0.08251953125,
0.4375,
0.464111328125,
0.4375,
0.464111328125,
0.447509765625,
0.08251953125,
0.447509765625
],
[
0.50244140625,
0.438232421875,
0.890625,
0.439453125,
0.890625,
0.453857421875,
0.50244140625,
0.45263671875
],
[
0.080078125,
0.448974609375,
0.468017578125,
0.45068359375,
0.468017578125,
0.46533203125,
0.080078125,
0.46337890625
],
[
0.50244140625,
0.453857421875,
0.890625,
0.455078125,
0.890625,
0.4697265625,
0.50244140625,
0.468505859375
],
[
0.08251953125,
0.466552734375,
0.46728515625,
0.466552734375,
0.46728515625,
0.479736328125,
0.08251953125,
0.479736328125
],
[
0.50244140625,
0.468505859375,
0.890625,
0.4697265625,
0.890625,
0.484130859375,
0.50244140625,
0.48291015625
],
[
0.07843017578125,
0.479736328125,
0.1968994140625,
0.48095703125,
0.1968994140625,
0.49560546875,
0.07843017578125,
0.494384765625
],
[
0.5048828125,
0.48486328125,
0.8896484375,
0.48486328125,
0.8896484375,
0.498779296875,
0.5048828125,
0.498779296875
],
[
0.50341796875,
0.50048828125,
0.87841796875,
0.50048828125,
0.87841796875,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.236083984375,
0.5087890625,
0.313720703125,
0.51123046875,
0.313720703125,
0.5263671875,
0.236083984375,
0.5244140625
],
[
0.52294921875,
0.51513671875,
0.890625,
0.5166015625,
0.890625,
0.52978515625,
0.52294921875,
0.52880859375
],
[
0.5048828125,
0.53076171875,
0.8896484375,
0.53076171875,
0.8896484375,
0.54443359375,
0.5048828125,
0.54443359375
],
[
0.1004638671875,
0.54248046875,
0.46728515625,
0.54248046875,
0.46728515625,
0.556640625,
0.1004638671875,
0.556640625
],
[
0.50634765625,
0.54736328125,
0.8896484375,
0.54736328125,
0.8896484375,
0.55810546875,
0.50634765625,
0.55810546875
],
[
0.08087158203125,
0.55810546875,
0.468017578125,
0.55810546875,
0.468017578125,
0.5712890625,
0.08087158203125,
0.5712890625
],
[
0.5048828125,
0.56103515625,
0.8896484375,
0.56103515625,
0.8896484375,
0.57470703125,
0.5048828125,
0.57470703125
],
[
0.08251953125,
0.57470703125,
0.465576171875,
0.57470703125,
0.465576171875,
0.58544921875,
0.08251953125,
0.58544921875
],
[
0.5048828125,
0.5771484375,
0.890625,
0.5771484375,
0.890625,
0.59033203125,
0.5048828125,
0.59033203125
],
[
0.08087158203125,
0.58837890625,
0.46728515625,
0.58837890625,
0.46728515625,
0.60302734375,
0.08087158203125,
0.60302734375
],
[
0.5048828125,
0.59130859375,
0.890625,
0.59130859375,
0.890625,
0.60498046875,
0.5048828125,
0.60498046875
],
[
0.08087158203125,
0.60400390625,
0.46728515625,
0.60400390625,
0.46728515625,
0.6171875,
0.08087158203125,
0.6171875
],
[
0.50341796875,
0.60595703125,
0.890625,
0.60595703125,
0.890625,
0.619140625,
0.50341796875,
0.619140625
],
[
0.08251953125,
0.62060546875,
0.464111328125,
0.62060546875,
0.464111328125,
0.630859375,
0.08251953125,
0.630859375
],
[
0.50341796875,
0.619140625,
0.8310546875,
0.62060546875,
0.8310546875,
0.63525390625,
0.50341796875,
0.6337890625
],
[
0.080078125,
0.63330078125,
0.46728515625,
0.63330078125,
0.46728515625,
0.646484375,
0.080078125,
0.646484375
],
[
0.52392578125,
0.63623046875,
0.8896484375,
0.63623046875,
0.8896484375,
0.6494140625,
0.52392578125,
0.6494140625
],
[
0.08251953125,
0.65087890625,
0.465576171875,
0.65087890625,
0.465576171875,
0.6611328125,
0.08251953125,
0.6611328125
],
[
0.50341796875,
0.65087890625,
0.8896484375,
0.65087890625,
0.8896484375,
0.6640625,
0.50341796875,
0.6640625
],
[
0.08087158203125,
0.6640625,
0.468017578125,
0.6640625,
0.468017578125,
0.67822265625,
0.08087158203125,
0.67822265625
],
[
0.50634765625,
0.66845703125,
0.8896484375,
0.66845703125,
0.8896484375,
0.6787109375,
0.50634765625,
0.6787109375
],
[
0.08087158203125,
0.6796875,
0.46728515625,
0.6796875,
0.46728515625,
0.693359375,
0.08087158203125,
0.693359375
],
[
0.50341796875,
0.681640625,
0.890625,
0.68115234375,
0.890625,
0.6943359375,
0.50341796875,
0.69580078125
],
[
0.080078125,
0.693359375,
0.468017578125,
0.693359375,
0.468017578125,
0.70703125,
0.080078125,
0.70703125
],
[
0.50341796875,
0.69580078125,
0.89208984375,
0.6962890625,
0.89208984375,
0.7109375,
0.50341796875,
0.71044921875
],
[
0.080078125,
0.70849609375,
0.46728515625,
0.708984375,
0.46728515625,
0.72265625,
0.080078125,
0.7216796875
],
[
0.50341796875,
0.7119140625,
0.67333984375,
0.7119140625,
0.67333984375,
0.72607421875,
0.50341796875,
0.72607421875
],
[
0.080078125,
0.7236328125,
0.28515625,
0.7236328125,
0.28515625,
0.7373046875,
0.080078125,
0.7373046875
],
[
0.50341796875,
0.7265625,
0.890625,
0.72802734375,
0.890625,
0.7412109375,
0.50341796875,
0.74072265625
],
[
0.50341796875,
0.7412109375,
0.8896484375,
0.7421875,
0.8896484375,
0.7568359375,
0.50341796875,
0.755859375
],
[
0.5048828125,
0.75830078125,
0.890625,
0.75830078125,
0.890625,
0.771484375,
0.5048828125,
0.771484375
],
[
0.08087158203125,
0.771484375,
0.21728515625,
0.771484375,
0.21728515625,
0.78515625,
0.08087158203125,
0.78515625
],
[
0.5048828125,
0.77294921875,
0.890625,
0.77294921875,
0.890625,
0.7861328125,
0.5048828125,
0.7861328125
],
[
0.50341796875,
0.7861328125,
0.890625,
0.787109375,
0.890625,
0.8017578125,
0.50341796875,
0.80029296875
],
[
0.10296630859375,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.10296630859375,
0.81005859375
],
[
0.50341796875,
0.80322265625,
0.86376953125,
0.80322265625,
0.86376953125,
0.81640625,
0.50341796875,
0.81640625
],
[
0.08251953125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.08251953125,
0.82470703125
],
[
0.52294921875,
0.818359375,
0.63818359375,
0.818359375,
0.63818359375,
0.83203125,
0.52294921875,
0.83203125
],
[
0.08251953125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.08251953125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08251953125,
0.85498046875
],
[
0.50244140625,
0.8466796875,
0.650390625,
0.8466796875,
0.650390625,
0.85986328125,
0.50244140625,
0.85986328125
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.52294921875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.52294921875,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.37744140625,
0.88720703125,
0.37744140625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.5048828125,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5048828125,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.9375,
0.48046875,
0.9375
]
] |
[
"ViDeNN: Deep Blind Video Denoising",
"Michele Claus",
"Jan van Gemert",
"Computer Vision lab",
"Computer Vision lab",
"Delft University of Technology",
"Delft University of Technology",
"claus.michele@hotmail.i",
"http://jvgemert.github.io/",
"ViDeNN",
"Denoised Vide",
"成像噪声是由于热效应，传感器缺陷或弱光引起的，",
"Denol",
"VoisyVide",
"isin",
"",
"手动调整多个滤波器参数是为每个增益级别优化新相",
"Dempsra!",
"机的质量和带宽的基础，需要花费大量时间和精力。在",
"这里，我们使用CNN自动执行去噪过程，以实现灵活",
"高效的视频去噪，能够盲自去除噪声。在“盲”条件下",
"工作的噪声去除算法在现实世界中至关重要，在这种",
"情况下，颜色和光线条件可能会突然变化，从而为每个",
"顿产生不同的噪声分布",
"基于统计模型的解决方案包括马尔可夫随机场模型",
"Figure l: ViDeNN approach to Video Denoising: com-",
"[3]，梯度模型[4]，稀疏模型[5]和当前在BM3D[6]，",
"biningtwonetworksperformingfirstSingleFrameSpa",
"LSSC7]，NCSR[8]和WNNMAX6等最新技术中使",
"tial Denoising and subsequently Temporal Denoising",
"用的非局部自相似性（NSS）。尽管它们实现了可观的",
"over a window of three frames, all in a single feed-",
"去噪性能，但大多数算法都有一些缺点。首先，它们通",
"forward process.",
"常用于处理特定的噪声模型和水平，限制了它们在盲",
"降噪中的使用。其次，它们涉及耗时的手动优化程序",
"Abstract",
"在图像去噪方面已经做了很多工作，而专门为视频",
"设计的算法却很少。视频去噪的关键假设是视频顿是",
"We propose ViDeNN: a CNN for Video Denois-",
"强相关的。最基本的视频去噪技术包括各种后续顿的时",
"ing without prior knowledge on the noise distribution",
"间平均值。虽然为稳定的场景提供了出色的效果，但它",
"(blind denoising). The CNN architecture uses a combi",
"会模糊运动，导致伪影。VBM4D方法「10】是视频去噪",
"nation of spatial and temporal filtering, learning to spa-",
"的最新技术。它通过搜索相似的补丁来扩展BM3DE",
"tially denoise the frames first and at the same time how",
"[6]单幅图像去噪，不仅在空间域，而且在时间域。在",
"tocombinetheirtemporal information,handlingob",
"更多顿中搜索类似的补丁会大大增加处理时间",
"jects motion, brightness changes, low-light conditions",
"在本文中，我们提出了ViDeNN，如图E1所示：用",
"and temporal inconsistencies.Wedemonstratetheim-",
"于盲视频去噪的卷积神经网络，能够在没有噪声模型",
"portance of the data used for CNNs training, creating",
"和视频内容的先验知识的情况下对视频进行去噪。为",
"for this purpose a specific dataset for low-light condi-",
"了进行比较，已经在公开可用的视频和自我捕获的视",
"tions. We test ViDeNN on common benchmarks and",
"频上进行了实验。。视频，弱光数据集和代码将发布在",
"on self-collected data, achieving good results compara-",
"项目的GitHub页面上。",
"ble with the state-of-the-art.",
"我们工作的主要贡献是：（i）一种新颖的CNN架构，",
"能够对视频进行盲降噪，将多个帧的空间和时间信息",
"与一个单一的前馈过程相结合；（i）在弱光条件下对",
"1.Introduction",
"加性高斯白噪声和真实数据进行灵活性测试；（i）在",
"具有挑战性的情况下对运动的鲁棒性；（iv）用于特定",
"图像和视频去噪旨在从可用的噪声观测Y中获得原",
"博世安全摄像机的新的低光数据集，具有无噪音和",
"始信号X。噪声影响感知的视觉质量，但也分割E[1]",
"用于图像去噪的",
"和压缩E2]使去噪成为重要的一步。以X为原始信",
"号，N为噪声，Y为可用的噪声观测值，对于加性类",
"2. Related Work",
"型的噪声，噪声退化模型可以描述为Y=X+N。在",
"弱光条件下，噪声与信号有关，在黑暗区域更敏感，建",
"CNN。从2008年Jain和Seung「1Il的CNN图像",
"模为Y=H(X)+N，以H为退化函数",
"去噪工作中，由于更多的计算能力和高质量的数据集，",
"1"
] |
{
"arxiv_id": "1904.10898",
"language": "zh",
"page_idx": 0
}
|
||
1904.10898-zh-1
|
1904.10898-zh-1.jpg
|
已经有了巨大的改进。2012 年，Burger et al. [12]展示了即使是一个简单的多层感知器也能获得与 BM3D [6]相当的结果，尽管训练[13]需要一个巨大的数据集。最近，在 2016 年，张 et al. $\operatorname{E}[14]$在其 DnCNN 架构中使用残差学习和批量归一化 [15] 进行图像去噪。凭借其简单而有效的架构，它已被证明是灵活的任务，如盲高斯去噪，JPEG 去块和图像修复。FFDNet[16] 通过处理扩展范围的噪声水平来扩展 DnCNN,并具有去除空间变化噪声的能力。Ulyanov et al.[17] 展示了如何通过其 DeepPrior 增强给定图像，而无需图像本身以外的先验训练数据，这可以看作是“盲”去噪。也有一些直接受 BM3D 启发的 CNN 作品，如[18,19]。在[20]中，Ying et al. 提出了一个名为 MemNet 的深度持久内存网络，该网络获得有效的结果，引人了一个内存块，其动机是人类的思想是持久的。然而，网络结构仍然很复杂，不容易重现。U-Net 变体已成功用于 et al. $\operatorname{E}[21]$的工作中的图像去噪，并在最近的郭 et al. $\operatorname{E}$ [22]的图像去噪工作中称为 CBDNet。通过他们的新颖方法，CBDNet 在现实世界的盲图像去噪中达到了非凡的效果。最近提出的 Noise2Noise [23] 模型基于编码器-解码器结构，仅使用噪声图像进行训练即可获得几乎相同的结果，而不是干净噪声对，这对于无法获得基本事实的情况特别有用
视频和深度神经网络。使用深度学习的视频去噪仍然是一个未被探索的研究领域。新元 et al. [24] 的开创性工作，是目前唯一使用神经网络(递归神经网络)解决视频去噪的。我们的不同之处在于解决彩色视频去噪问题，并提供与最新技术相当的结果。使用 CNN 解决的其他基于视频的任务包括视频帧增强，插值，去模糊和超分辨率，其中关键组成部分是如何处理运动和时间变化。对于帧插值，Niklaus et al. $\textcircled{\mathrm{E}}[25]$使用预先计算的光流将运动信息馈送到帧插值 CNN。Meyer et al. [26] 使用基于相位的特征来描述运动。Caballero et al. $\mathsf{E}[27]$开发了一个网络，该网络可以自行估计视频超分辨率的运动。同样，在多帧质量增强(MFQE)中， Yang et al. $\mathbb{E}[28]$使用运动补偿网络和质量增强网络， 考虑到 H265 压缩视频的三个非连续帧。专门针对视频去模糊，Su et al. $\operatorname{E}[29]$开发了一个名为 DeBlurNet 的网络：U-NetCNN, 它将三个帧堆叠在一起作为输人。同样，我们在 ViDeNN 中也使用了三个堆叠帧。此外， 我们还研究了输入帧数量的变化
真实世界数据集。图像或视频去噪算法必须对现实世界的数据有效才能成功。然而，很难获得真实图片的真实情况，因为不存在完美的传感器和通道。2014 年， Anaya 和 Barbu 创建了一个名为 RENOIR [30] 的低光条件数据集：他们使用不同的曝光时间和 ISO 级别来获取同一静态场景的嘈杂和干净的图像。同样，在2017 年，Plotz 和 Roth 创建了一个名为 DND [31] 的数据集。在这种情况下，只有嘈杂的样本已被释放，而无噪声的样本则保持不公开，以进行基准测试。最近， 另外两篇相关论文已经发表。第一个由 Abdelhamed et al. $\mathbb{E}[32]$撰写，涉及创建嘈杂和干净图像的智能手机图像数据集，在撰写本文时尚未公开。由 Chen et al.
<FigureHere>
Figure 2: The architecture of the proposed ViDeNN network. Every frame will go through a spatial denoising CNN. The temporal CNN takes as input three spatially denoised frames and outputs the final estimate of the central frame. Both CNNs estimate first the noise residual, i.e. the unwanted values noise adds to an image, and then subtracts them from the noisy input $(\oplus$ means addition of the two signals, and "-" the negation). ViDeNN is composed only by Convolutional Layers. The number of feature maps is written at the bottom of each layer.
$\mathbb{E}[33]$撰写的第二篇文章提出了一种新的基于 CNN 的算法，该算法能够增强低光原始图像的质量。他们创建了一个类似于 [30] 的两种相机类型的专用数据集
## 3. ViDeNN: Video DeNoising Net
ViDeNN 有两个子网：时空去噪 CNN,如图 2 所
示
## $3. 1. \textbf{Spatial Denoising CNN}$
对于空间去噪，我们建立在$\mathbb{E}[14]$ 上，它显示出同时处理多种退化类型的极大灵活性，并尝试使用相同的架构进行盲空间去噪。结果表明，该架构可以实现高斯去噪的最新结果。当网络必须同时处理不同的噪声模型时，深度为 128 的第一层会有所帮助。网络深度设置为 20, 使用 BatchNormalization(BN)[15]。激活函数为 ReLU (整流线性单元)。我们还研究了使用LeakyReLU 作为激活函数，该函数可以更有效地 [34] .而无需对 ReLU 讲行改讲。补充材料中提供了比较结果。我们的 Spatial-CNN 使用残差学习，它是在 [14] 中首次引入的，用于解决图像去噪问题。残差架构不是强制网络直接输出去噪帧，而是预测残差图像，残差图像包括原始干净图像和嘈杂观察之间的差异。损失函数$L$是 L2 范数，也称为最小二乘误差(LSE),是目标值$Y$与估计值$Y_est$之间的差$S$的平方和。在这种情况下，差$S$表示由 Spatial-CNN 估计的噪声残差图像，并由$L=\sum_x\sum_y\left(\underbrace{Y(x,y)-Y_{est}(x,y)}_{\mathrm{Noise~Besidual}}\right)^{2}$给出
|
[
[
0.08087158203125,
0.091552734375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.106689453125,
0.08087158203125,
0.1060791015625
],
[
0.5107421875,
0.1029052734375,
0.54150390625,
0.1029052734375,
0.54150390625,
0.106689453125,
0.5107421875,
0.106689453125
],
[
0.55810546875,
0.10479736328125,
0.69140625,
0.1029052734375,
0.69287109375,
0.141357421875,
0.5595703125,
0.143310546875
],
[
0.08416748046875,
0.125,
0.465576171875,
0.125,
0.465576171875,
0.1351318359375,
0.08416748046875,
0.1351318359375
],
[
0.08251953125,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.121826171875,
0.08251953125,
0.121826171875
],
[
0.08087158203125,
0.136962890625,
0.468017578125,
0.136962890625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.861328125,
0.1395263671875,
0.8896484375,
0.1395263671875,
0.8896484375,
0.15087890625,
0.861328125,
0.15087890625
],
[
0.08251953125,
0.1514892578125,
0.46728515625,
0.1514892578125,
0.46728515625,
0.1654052734375,
0.08251953125,
0.1654052734375
],
[
0.5595703125,
0.149658203125,
0.85888671875,
0.1514892578125,
0.85888671875,
0.1875,
0.5595703125,
0.1849365234375
],
[
0.84619140625,
0.1634521484375,
0.85107421875,
0.1634521484375,
0.85107421875,
0.1717529296875,
0.84619140625,
0.1717529296875
],
[
0.84716796875,
0.1590576171875,
0.8857421875,
0.1590576171875,
0.8857421875,
0.1749267578125,
0.84716796875,
0.1749267578125
],
[
0.08251953125,
0.1673583984375,
0.46728515625,
0.1673583984375,
0.46728515625,
0.18115234375,
0.08251953125,
0.18115234375
],
[
0.08416748046875,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.08416748046875,
0.1962890625
],
[
0.5146484375,
0.1900634765625,
0.53125,
0.1900634765625,
0.53125,
0.1932373046875,
0.5146484375,
0.1932373046875
],
[
0.08087158203125,
0.1962890625,
0.468017578125,
0.1976318359375,
0.468017578125,
0.212158203125,
0.08087158203125,
0.2108154296875
],
[
0.54833984375,
0.1912841796875,
0.69384765625,
0.1932373046875,
0.69287109375,
0.2342529296875,
0.54736328125,
0.231689453125
],
[
0.525390625,
0.205810546875,
0.53173828125,
0.205810546875,
0.53173828125,
0.21337890625,
0.525390625,
0.21337890625
],
[
0.529296875,
0.203857421875,
0.55224609375,
0.203857421875,
0.55224609375,
0.21533203125,
0.529296875,
0.21533203125
],
[
0.08087158203125,
0.21337890625,
0.46728515625,
0.21337890625,
0.46728515625,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.08251953125,
0.2298583984375,
0.465576171875,
0.2298583984375,
0.465576171875,
0.240478515625,
0.08251953125,
0.240478515625
],
[
0.60107421875,
0.23291015625,
0.66015625,
0.23291015625,
0.66015625,
0.24365234375,
0.60107421875,
0.24365234375
],
[
0.736328125,
0.236083984375,
0.79736328125,
0.236083984375,
0.79736328125,
0.240478515625,
0.736328125,
0.240478515625
],
[
0.08251953125,
0.24365234375,
0.46728515625,
0.24365234375,
0.46728515625,
0.2568359375,
0.08251953125,
0.2568359375
],
[
0.09722900390625,
0.2587890625,
0.46728515625,
0.2587890625,
0.46728515625,
0.272705078125,
0.09722900390625,
0.272705078125
],
[
0.08087158203125,
0.2734375,
0.46728515625,
0.2734375,
0.46728515625,
0.287353515625,
0.08087158203125,
0.287353515625
],
[
0.50341796875,
0.267578125,
0.88818359375,
0.267578125,
0.88818359375,
0.2822265625,
0.50341796875,
0.2822265625
],
[
0.50244140625,
0.2822265625,
0.88623046875,
0.28271484375,
0.88623046875,
0.29736328125,
0.50244140625,
0.296630859375
],
[
0.08087158203125,
0.2890625,
0.46728515625,
0.2890625,
0.46728515625,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.50244140625,
0.296630859375,
0.8896484375,
0.298583984375,
0.8896484375,
0.3125,
0.50244140625,
0.31005859375
],
[
0.08251953125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.08251953125,
0.317626953125
],
[
0.50341796875,
0.313232421875,
0.8896484375,
0.313232421875,
0.8896484375,
0.32763671875,
0.50341796875,
0.32763671875
],
[
0.08416748046875,
0.32080078125,
0.46728515625,
0.32080078125,
0.46728515625,
0.330810546875,
0.08416748046875,
0.330810546875
],
[
0.08087158203125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.08087158203125,
0.347900390625
],
[
0.50341796875,
0.328857421875,
0.890625,
0.328857421875,
0.890625,
0.34228515625,
0.50341796875,
0.34228515625
],
[
0.08087158203125,
0.348388671875,
0.468017578125,
0.348388671875,
0.468017578125,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.50341796875,
0.3447265625,
0.890625,
0.3447265625,
0.890625,
0.35791015625,
0.50341796875,
0.35791015625
],
[
0.08251953125,
0.363037109375,
0.468017578125,
0.363037109375,
0.468017578125,
0.37744140625,
0.08251953125,
0.37744140625
],
[
0.50341796875,
0.359130859375,
0.890625,
0.359130859375,
0.890625,
0.37255859375,
0.50341796875,
0.37255859375
],
[
0.5009765625,
0.371826171875,
0.890625,
0.369384765625,
0.890625,
0.386962890625,
0.5009765625,
0.389404296875
],
[
0.08416748046875,
0.37890625,
0.468017578125,
0.37890625,
0.468017578125,
0.392578125,
0.08416748046875,
0.392578125
],
[
0.08087158203125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.08087158203125,
0.40771484375
],
[
0.50341796875,
0.389404296875,
0.890625,
0.389404296875,
0.890625,
0.40283203125,
0.50341796875,
0.40283203125
],
[
0.50244140625,
0.40283203125,
0.890625,
0.404052734375,
0.890625,
0.41845703125,
0.50244140625,
0.417236328125
],
[
0.08251953125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.08251953125,
0.423583984375
],
[
0.50341796875,
0.41845703125,
0.70361328125,
0.419189453125,
0.70361328125,
0.43310546875,
0.50341796875,
0.431884765625
],
[
0.08251953125,
0.4248046875,
0.2459716796875,
0.4248046875,
0.2459716796875,
0.438232421875,
0.08251953125,
0.438232421875
],
[
0.101318359375,
0.4443359375,
0.46728515625,
0.4443359375,
0.46728515625,
0.458251953125,
0.101318359375,
0.458251953125
],
[
0.08087158203125,
0.458984375,
0.46728515625,
0.458984375,
0.46728515625,
0.473388671875,
0.08087158203125,
0.473388671875
],
[
0.50244140625,
0.453857421875,
0.890625,
0.455810546875,
0.890625,
0.47021484375,
0.50244140625,
0.468505859375
],
[
0.08251953125,
0.474853515625,
0.46728515625,
0.474853515625,
0.46728515625,
0.488037109375,
0.08251953125,
0.488037109375
],
[
0.50341796875,
0.47021484375,
0.890625,
0.47021484375,
0.890625,
0.484130859375,
0.50341796875,
0.484130859375
],
[
0.08251953125,
0.4912109375,
0.46728515625,
0.4912109375,
0.46728515625,
0.501953125,
0.08251953125,
0.501953125
],
[
0.5048828125,
0.486083984375,
0.8486328125,
0.486083984375,
0.8486328125,
0.5,
0.5048828125,
0.5
],
[
0.08087158203125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51806640625,
0.08087158203125,
0.51806640625
],
[
0.08416748046875,
0.521484375,
0.465576171875,
0.521484375,
0.465576171875,
0.5322265625,
0.08416748046875,
0.5322265625
],
[
0.5009765625,
0.521484375,
0.7939453125,
0.52294921875,
0.7939453125,
0.537109375,
0.5009765625,
0.5361328125
],
[
0.08251953125,
0.53515625,
0.46728515625,
0.53515625,
0.46728515625,
0.548828125,
0.08251953125,
0.548828125
],
[
0.08251953125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.08251953125,
0.564453125
],
[
0.52294921875,
0.548828125,
0.890625,
0.5498046875,
0.890625,
0.564453125,
0.52294921875,
0.56298828125
],
[
0.08251953125,
0.56494140625,
0.46728515625,
0.56494140625,
0.46728515625,
0.5791015625,
0.08251953125,
0.5791015625
],
[
0.5048828125,
0.568359375,
0.521484375,
0.568359375,
0.521484375,
0.57763671875,
0.5048828125,
0.57763671875
],
[
0.08087158203125,
0.580078125,
0.46728515625,
0.580078125,
0.46728515625,
0.59326171875,
0.08087158203125,
0.59326171875
],
[
0.080078125,
0.59326171875,
0.468017578125,
0.5947265625,
0.468017578125,
0.609375,
0.080078125,
0.60791015625
],
[
0.50341796875,
0.5986328125,
0.7197265625,
0.5986328125,
0.7197265625,
0.6123046875,
0.50341796875,
0.6123046875
],
[
0.08087158203125,
0.609375,
0.465576171875,
0.609375,
0.465576171875,
0.62255859375,
0.08087158203125,
0.62255859375
],
[
0.08251953125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.08251953125,
0.6396484375
],
[
0.52294921875,
0.62353515625,
0.88818359375,
0.62353515625,
0.88818359375,
0.63720703125,
0.52294921875,
0.63720703125
],
[
0.08251953125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.08251953125,
0.65380859375
],
[
0.50341796875,
0.6396484375,
0.890625,
0.6396484375,
0.890625,
0.65283203125,
0.50341796875,
0.65283203125
],
[
0.08251953125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66943359375,
0.08251953125,
0.66943359375
],
[
0.50244140625,
0.65380859375,
0.8896484375,
0.65380859375,
0.8896484375,
0.66748046875,
0.50244140625,
0.66748046875
],
[
0.08087158203125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.08087158203125,
0.6845703125
],
[
0.5048828125,
0.66943359375,
0.890625,
0.66943359375,
0.890625,
0.68310546875,
0.5048828125,
0.68310546875
],
[
0.08251953125,
0.68505859375,
0.468017578125,
0.68505859375,
0.468017578125,
0.69970703125,
0.08251953125,
0.69970703125
],
[
0.50341796875,
0.68310546875,
0.890625,
0.6845703125,
0.890625,
0.69873046875,
0.50341796875,
0.69775390625
],
[
0.08087158203125,
0.7021484375,
0.3095703125,
0.7021484375,
0.3095703125,
0.71533203125,
0.08087158203125,
0.71533203125
],
[
0.50634765625,
0.69970703125,
0.88818359375,
0.69970703125,
0.88818359375,
0.71337890625,
0.50634765625,
0.71337890625
],
[
0.5048828125,
0.7138671875,
0.88818359375,
0.7138671875,
0.88818359375,
0.72802734375,
0.5048828125,
0.72802734375
],
[
0.10296630859375,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.10296630859375,
0.734375
],
[
0.50341796875,
0.72998046875,
0.88818359375,
0.72998046875,
0.88818359375,
0.74365234375,
0.50341796875,
0.74365234375
],
[
0.08251953125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.08251953125,
0.75
],
[
0.08416748046875,
0.75048828125,
0.465576171875,
0.75048828125,
0.465576171875,
0.7646484375,
0.08416748046875,
0.7646484375
],
[
0.51953125,
0.74560546875,
0.8896484375,
0.74560546875,
0.8896484375,
0.7587890625,
0.51953125,
0.7587890625
],
[
0.50244140625,
0.75830078125,
0.8896484375,
0.76025390625,
0.8896484375,
0.7744140625,
0.50244140625,
0.77294921875
],
[
0.08251953125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.08251953125,
0.77978515625
],
[
0.5048828125,
0.7744140625,
0.8896484375,
0.7744140625,
0.8896484375,
0.7890625,
0.5048828125,
0.7890625
],
[
0.08251953125,
0.78076171875,
0.46728515625,
0.78076171875,
0.46728515625,
0.79443359375,
0.08251953125,
0.79443359375
],
[
0.50244140625,
0.7890625,
0.88818359375,
0.79052734375,
0.88818359375,
0.8037109375,
0.50244140625,
0.80322265625
],
[
0.08251953125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.08251953125,
0.81005859375
],
[
0.08087158203125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.5048828125,
0.80615234375,
0.88818359375,
0.80615234375,
0.88818359375,
0.8193359375,
0.5048828125,
0.8193359375
],
[
0.5048828125,
0.82080078125,
0.88818359375,
0.82080078125,
0.88818359375,
0.833984375,
0.5048828125,
0.833984375
],
[
0.08251953125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.50341796875,
0.83544921875,
0.88818359375,
0.83544921875,
0.88818359375,
0.849609375,
0.50341796875,
0.849609375
],
[
0.08251953125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08251953125,
0.85498046875
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.50244140625,
0.85107421875,
0.88818359375,
0.849609375,
0.88818359375,
0.8642578125,
0.50244140625,
0.86572265625
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.50244140625,
0.87060546875,
0.8095703125,
0.87060546875,
0.8095703125,
0.88818359375,
0.50244140625,
0.88818359375
],
[
0.80810546875,
0.87255859375,
0.85107421875,
0.87255859375,
0.85107421875,
0.8837890625,
0.80810546875,
0.8837890625
],
[
0.08087158203125,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.61376953125,
0.88818359375,
0.62353515625,
0.88818359375,
0.62353515625,
0.8916015625,
0.61376953125,
0.8916015625
],
[
0.62353515625,
0.88818359375,
0.6396484375,
0.88818359375,
0.6396484375,
0.8916015625,
0.62353515625,
0.8916015625
],
[
0.685546875,
0.89306640625,
0.76953125,
0.89306640625,
0.76953125,
0.90478515625,
0.685546875,
0.90478515625
]
] |
[
"已经有了巨大的改进。2012年，Burgeretal.[12展",
"-rameT.",
"",
"6相当的结果，尽管训练「13需要一个巨大的数据集。",
"示了即使是一个简单的多层感知器也能获得与BM3D",
"最近，在2016年，张etal.E[14]在其DnCNN架构",
"Denotsec",
"中使用残差学习和批量归一化「151进行图像去噪。凭",
"",
"",
"",
"借其简单而有效的架构，它已被证明是灵活的任务，如",
"盲高斯去噪，JPEG去块和图像修复。FFDNet「16]通",
"Fam",
"过处理扩展范围的噪声水平来扩展DnCNN，并具有去",
"",
"",
"",
"除空间变化噪声的能力。Ulyanovetal．[17l展示了如",
"何通过其DeepPrior增强给定图像，而无需图像本身",
"Spatial Denoisin",
"emporalDenoisin",
"以外的先验训练数据，这可以看作是“盲”去噪。也有",
"些直接受BM3D启发的CNN作品，如【18，19]。在",
"[20]中，Yingetal.提出了一个名为MemNet的深度",
"Figure 2:The architecture of the proposed Vi",
"DeNN network.Every frame will go through a spa",
"持久内存网络，该网络获得有效的结果，引人了一个内",
"tial denoising CNN.The temporal CNN takes as input",
"存块，其动机是人类的思想是持久的。然而，网络结构",
"three spatially denoised frames and outputs the final es-",
"仍然很复杂：不容易重现。U-Net变体已成功用于et",
"al.E[21]的工作中的图像去噪，并在最近的郭etal.E",
"timate of the central frame. Both CNNs estimate first",
"[22]的图像去噪工作中称为CBDNet。通过他们的新颖",
"the noise residual, i.e. the unwanted values noise adds",
"方法，CBDNet在现实世界的盲图像去噪中达到了非",
"to an image, and then subtracts them from the noisy",
"input ( @ means addition of the two signals, and \",",
"凡的效果。最近提出的Noise2Noise[23]模型基于编码",
"器-解码器结构，仅使用噪声图像进行训练即可获得几",
"the negation). ViDeNN is composed only by Convolu-",
"tional Layers. The number of feature maps is written",
"乎相同的结果，而不是干净噪声对，这对于无法获得基",
"atthebottom ofeachlayer.",
"本事实的情况特别有用",
"视频和深度神经网络。使用深度学习的视频去噪仍",
"然是一个未被探索的研究领域。新元etal.[24]的开创",
"E[331撰写的第二篇文章提出了一种新的基于CNN的",
"性工作，是目前唯一使用神经网络（递归神经网络）解",
"算法，该算法能够增强低光原始图像的质量。他们创建",
"决视频去噪的。我们的不同之处在于解决彩色视频去",
"了一个类似于【30]的两种相机类型的专用数据集",
"噪问题，并提供与最新技术相当的结果。使用CNN解",
"决的其他基于视频的任务包括视频顿增强，插值，去模",
"3.ViDeNN: Video DeNoising Net",
"糊和超分辨率，其中关键组成部分是如何处理运动和",
"时间变化。对于顿插值，Niklausetal.E[25]使用预先",
"ViDeNN有两个子网：时空去噪CNN，如图2所",
"计算的光流将运动信息馈送到顿插值CNN。Meyeret",
"示",
"al.261使用基于相位的特征来描述运动。Caballeroet",
"al.E[27]开发了一个网络，该网络可以自行估计视频",
"3.1. Spatial Denoising CNN",
"超分辨率的运动。同样，在多顿质量增强（MFQE）中",
"Yangetal.E[28]使用运动补偿网络和质量增强网络,",
"对于空间去噪，我们建立在E14上，它显示出同",
"考虑到H265压缩视频的三个非连续顿。专门针对视频",
"时处理多种退化类型的极大灵活性，并尝试使用相同",
"去模糊，Suet al.E[29]开发了一个名为DeBlurNet 的",
"的架构进行盲空间去噪。结果表明，该架构可以实现",
"网络：U-NetCNN，它将三个顿堆叠在一起作为输人。",
"高斯去噪的最新结果。当网络必须同时处理不同的噪",
"同样，我们在ViDeNN中也使用了三个堆叠顿。此外，",
"声模型时，深度为128的第一层会有所帮助。网络深",
"我们还研究了输人顿数量的变化",
"度设置为20，使用BatchNormalization(BN[15]。激",
"活函数为ReLU（整流线性单元）。我们还研究了使用",
"真实世界数据集。图像或视频去噪算法必须对现实",
"LeakyReLU作为激活函数，该函数可以更有效地[34]",
"世界的数据有效才能成功。然而，很难获得真实图片的",
"真实情况，因为不存在完美的传感器和通道。2014年",
"而无需对ReLU进行改进。补充材料中提供了比较",
"结果。我们的Spatial-CNN使用残差学习，它是在[14]",
"Anaya和Barbu创建了一个名为RENOIR[30]的低",
"中首次引人的，用于解决图像去噪问题。残差架构不是",
"光条件数据集：他们使用不同的曝光时间和ISO级别",
"强制网络直接输出去噪顿，而是预测残差图像，残差图",
"来获取同一静态场景的噜杂和干净的图像。同样，在",
"2017年，Plotz和Roth创建了一个名为DND[31]的",
"像包括原始干净图像和噜杂观察之间的差异。损失函",
"数L是L2范数，也称为最小二乘误差（LSE），是目",
"数据集。在这种情况下，只有噜杂的样本已被释放，而",
"标值Y与估计值Yest之间的差S的平方和。在这种",
"无噪声的样本则保持不公开，以进行基准测试。最近，",
"另外两篇相关论文已经发表。第一个由Abdelhamed et",
"情况下，差S表示由Spatial-CNN估计的噪声残差图",
"al.E[32]撰写，涉及创建噜杂和干净图像的智能手机",
"像，并由L=Z(Y(a,y)-Yest(r,y))",
"给出",
"图像数据集，在撰写本文时尚未公开。由Chenetal.",
"",
"71",
"Noise Residual"
] |
{
"arxiv_id": "1904.10898",
"language": "zh",
"page_idx": 1
}
|
||
1904.10898-zh-3
|
1904.10898-zh-3.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>=5 $\sigma:$</th>
<th>10 $\sigma$ = </th>
<th>15 $\sigma:$ = </th>
<th>25 $\sigma=$ =</th>
<th>35 $\sigma=$</th>
<th>$5C$ $\sigma=$</th>
</tr>
<tr>
<td>$Spatial-CNN^{*}$</td>
<td>39.73</td>
<td>35.92</td>
<td>33.66</td>
<td>30.99</td>
<td>29.34</td>
<td>27.63</td>
</tr>
<tr>
<td>DnCNN- $B^{* }$ 14</td>
<td>39.79</td>
<td>35.87</td>
<td>33.57</td>
<td>30.69</td>
<td>28.74</td>
<td>26.53</td>
</tr>
<tr>
<td>DnCNN-B $14^{'}$</td>
<td>40.62</td>
<td>36.14</td>
<td>33.88</td>
<td>31.22</td>
<td>29.57</td>
<td>27.91</td>
</tr>
</tbody>
</table>
Table 1: Comparison of blind Gaussian denoising on the CBSD68 dataset. Our modified version of DnCNN for spatial denoising has comparable results with the original one. The values represent PSNR[dB] , the higher the better. DnCNN results obtained with the provided Matlab implementation [38] . CBSD68 avail- ${\mathrm{able~here~[ 39] }}$ .
$^{*}$Noisy images clipped in range [0,255].
# 4.2. Spatial CNN Training
训练分为两部分：(i)我们训练空间去噪 CNN,以
及(2)在收敛后，我们训练时间去噪 CNN
我们的理想模型必须同时处理多种退化类型，例如AWGN 和真实噪声模型 $\color{red}{2}$,包括弱光条件。在训练阶段，我们的神经网络将学习如何使用干净的图像作为参考来估计输入噪声图像的残余噪声含量。因此，我们需要一对干净而嘈杂的图像。它们很容易为 AWGN 和方程 2 中的真实噪声模型创建
我们使用滑铁卢探索数据集 [36] ,其中包含 4,744张干净的图像。可用图像的数量大大有助于概括，并允许我们保留其中的大部分用于测试。数据集随机分为两部分，70 个% 用于训练，30 个% 用于测试。一半的图像正在添加 AWGN 与$\sigma=[0,55]$。后半部分使用公式 2 进行处理，该公式是真实的噪声模型，模拟增益Ag= [ 0, 64] , 数 字 增 益 Dg= [ 0, 32]
在 [14] 之后，使用$50\times50\times3$补丁训练网络。我们从滑铁卢训练集中获得了 120,000 个补丁，其中包含 AWGN 和真实噪声类型，使用旋转，翻转和镜像等数据增强。对于低光照条件，我们从自己的训练数据集中为每个光照水平使用了五幅噪声图像，获得了60 对噪声干净的图像进行训练。提取的补丁为 80,000。从 Renoir 数据集中，我们使用子集 T3 并随机裁剪了40,000个补丁。对于低光测试，我们将使用来自不同场景的相机的 5 张图像，这些图像不存在于训练集中， 并且是雷诺阿 T3 集的一部分。我们训练了 100 个时期，使用了一批 128 个和 AdamOptimizer [37] ,前 20个时期的学习率为$10^-3$,最新的 80 个时期为$10^{-4}$。
4.3. Validation of static Image Denoising
我们将盲高斯去噪与我们所基于的 DnCNN 的原始实现进行了比较。从我们在 BSD68 测试集上的表 1 中的测试中，我们注意到我们的盲模型的结果与[14]论文提出的结果具有可比性
为了进一步验证真实世界的图像，我们评估sRGBDND 数据集 [31] 并提交评估。结果 [40] 令人鼓舞，因为我们的训练模型(在 DND 网页上称为 128- DnCNNTensorflow)的 PSNR 平均得分为 37.0343dB
<table>
<tbody>
<tr>
<th> </th>
<th>$PSNR$ [dB]</th>
<th>$SSIM$</th>
</tr>
<tr>
<td>Spatial-CNN</td>
<td>37.0343</td>
<td>0.9324</td>
</tr>
<tr>
<td>CBDNet $[222^{'}$</td>
<td>38.0564</td>
<td>0.9421</td>
</tr>
<tr>
<td>DnCNN+ $14^{\prime}$</td>
<td>37.9018</td>
<td>0.943</td>
</tr>
<tr>
<td>FFDNet+ 16</td>
<td>37.6107</td>
<td>0.9415</td>
</tr>
<tr>
<td>BM3D $[35^{\prime}$</td>
<td>34.51</td>
<td>0.8507</td>
</tr>
</tbody>
</table>
Table 2: Results of the DND benchmark [31] on realworld noisy images. It shows that our dataset, containing different noise models, is valid for real-world image denoising, placing our Spatial-CNN in the top 10 for sRGB denoising.
SSIM 的 0.9324 , 将其置于前 10 个位置。有趣的是， DnCNN 的作者在一周后提交了一个名为 DnCNN+ 的微调模型的结果，获得了 SSIM 的总体最高分，这进一步验证了其灵活性，请参见表 2
4.4. Temp3-CNN: Temporal CNN Training
对于视频评估，我们需要成对的干净和嘈杂的视频。对于人为添加的噪声，如加性白高斯噪声(AWGN)或方程 2 中的真实噪声模型，很容易创建这样的耦合。然而，对于现实世界和弱光条件下的视频，这几乎是不可能的。因此，这种提供成对的噪声和无噪声序列的视频数据集不可用。因此，我们决定按照以下顺序进行操作：
1. Select 31 publicly available videos from [41].
2. Divide videos in sequences of 3 frames.
3. Added either Gaussian noise with $\sigma=[0,55]$ or real
noise 2 with Ag=[0,64] and Dg=[0,32].
$4.{\mathrm{~Apply~Spatial-CNN}}$
5. Train on pairs of spatially-denoised and clean
$video.$
我们遵循与 Spatial-CNN 相同的训练过程，即使现在网络将使用$50\times50\times9$维的补丁进行训练，其中包含来自三个后续帧的三个补丁。
选定的 31 个视频包含 8922 帧，这意味着 2974 个三帧序列和 300032 的最终补丁数。我们运行了 60 个epoch 的训练，批处理大小为 128, Adam 优化器以$10^{-4}$的学习速率和 LeakyReLU 作为激活函数。它显示了 LeakyReLU 可以胜过 ReLU[34]。但是，我们没有在空间 CNN 中使用 LeakyRelu, 因为 ReLU 的性能更好。我们在补充材料中介绍了比较结果。在 Temp3- CNN 的最终版本中，未使用批量归一化(BN):实验表明它减慢了训练和去噪过程。BN 在 PSNR 方面没有改善最终结果。此外，没有 BN 的去噪需要大约 5% 更少的时间。图 5 表示 Temp3-CNN 的 L2 损失的演变：避免归一化步骤会使损失立即从低值开始。我们还使用 LeakyReLU, LeakyReLUBN 和 ReLUBN 进行了培训，并在补充材料中介绍了结果
|
[
[
0.37744140625,
0.09344482421875,
0.419189453125,
0.09344482421875,
0.419189453125,
0.10479736328125,
0.37744140625,
0.10479736328125
],
[
0.4248046875,
0.09344482421875,
0.4697265625,
0.09344482421875,
0.4697265625,
0.10479736328125,
0.4248046875,
0.10479736328125
],
[
0.689453125,
0.09344482421875,
0.7490234375,
0.09344482421875,
0.7490234375,
0.10479736328125,
0.689453125,
0.10479736328125
],
[
0.75634765625,
0.09344482421875,
0.79736328125,
0.09344482421875,
0.79736328125,
0.10418701171875,
0.75634765625,
0.10418701171875
],
[
0.2197265625,
0.0953369140625,
0.2646484375,
0.0953369140625,
0.2646484375,
0.1029052734375,
0.2197265625,
0.1029052734375
],
[
0.1805419921875,
0.0947265625,
0.21728515625,
0.0947265625,
0.21728515625,
0.10479736328125,
0.1805419921875,
0.10479736328125
],
[
0.295654296875,
0.0953369140625,
0.314453125,
0.0953369140625,
0.314453125,
0.1029052734375,
0.295654296875,
0.1029052734375
],
[
0.328369140625,
0.0953369140625,
0.343017578125,
0.0953369140625,
0.343017578125,
0.1029052734375,
0.328369140625,
0.1029052734375
],
[
0.3447265625,
0.0953369140625,
0.366943359375,
0.0953369140625,
0.366943359375,
0.10418701171875,
0.3447265625,
0.10418701171875
],
[
0.080078125,
0.10919189453125,
0.15283203125,
0.10919189453125,
0.15283203125,
0.12060546875,
0.080078125,
0.12060546875
],
[
0.275390625,
0.09783935546875,
0.28759765625,
0.09783935546875,
0.28759765625,
0.10162353515625,
0.275390625,
0.10162353515625
],
[
0.2296142578125,
0.10919189453125,
0.2607421875,
0.10919189453125,
0.2607421875,
0.121826171875,
0.2296142578125,
0.121826171875
],
[
0.281005859375,
0.10919189453125,
0.31201171875,
0.10919189453125,
0.31201171875,
0.121826171875,
0.281005859375,
0.121826171875
],
[
0.330810546875,
0.10919189453125,
0.365234375,
0.10919189453125,
0.365234375,
0.121826171875,
0.330810546875,
0.121826171875
],
[
0.38232421875,
0.10919189453125,
0.4150390625,
0.10919189453125,
0.4150390625,
0.121826171875,
0.38232421875,
0.121826171875
],
[
0.4345703125,
0.10919189453125,
0.46728515625,
0.10919189453125,
0.46728515625,
0.121826171875,
0.4345703125,
0.121826171875
],
[
0.59619140625,
0.10919189453125,
0.66650390625,
0.10919189453125,
0.66650390625,
0.12060546875,
0.59619140625,
0.12060546875
],
[
0.69677734375,
0.10986328125,
0.74267578125,
0.10986328125,
0.74267578125,
0.12060546875,
0.69677734375,
0.12060546875
],
[
0.76123046875,
0.10919189453125,
0.7998046875,
0.10919189453125,
0.7998046875,
0.12060546875,
0.76123046875,
0.12060546875
],
[
0.330810546875,
0.11871337890625,
0.365234375,
0.12060546875,
0.363525390625,
0.1331787109375,
0.330078125,
0.1307373046875
],
[
0.181396484375,
0.111083984375,
0.2132568359375,
0.111083984375,
0.2132568359375,
0.11932373046875,
0.181396484375,
0.11932373046875
],
[
0.2296142578125,
0.11932373046875,
0.2607421875,
0.121826171875,
0.259033203125,
0.1337890625,
0.2279052734375,
0.1319580078125
],
[
0.281005859375,
0.11932373046875,
0.31201171875,
0.121826171875,
0.310546875,
0.1331787109375,
0.279296875,
0.1307373046875
],
[
0.4345703125,
0.11932373046875,
0.46728515625,
0.121826171875,
0.465576171875,
0.1331787109375,
0.43310546875,
0.1307373046875
],
[
0.38232421875,
0.12060546875,
0.4150390625,
0.12060546875,
0.4150390625,
0.1331787109375,
0.38232421875,
0.1331787109375
],
[
0.080078125,
0.12060546875,
0.1683349609375,
0.12060546875,
0.1683349609375,
0.1319580078125,
0.080078125,
0.1319580078125
],
[
0.5947265625,
0.12060546875,
0.66650390625,
0.12060546875,
0.66650390625,
0.1319580078125,
0.5947265625,
0.1319580078125
],
[
0.76123046875,
0.12060546875,
0.79833984375,
0.12060546875,
0.79833984375,
0.1319580078125,
0.76123046875,
0.1319580078125
],
[
0.69677734375,
0.121826171875,
0.74169921875,
0.121826171875,
0.74169921875,
0.1319580078125,
0.69677734375,
0.1319580078125
],
[
0.1805419921875,
0.121826171875,
0.2132568359375,
0.121826171875,
0.2132568359375,
0.1331787109375,
0.1805419921875,
0.1331787109375
],
[
0.07843017578125,
0.1307373046875,
0.16015625,
0.1331787109375,
0.16015625,
0.14453125,
0.07843017578125,
0.1427001953125
],
[
0.2296142578125,
0.1307373046875,
0.261474609375,
0.1331787109375,
0.2607421875,
0.14453125,
0.2279052734375,
0.1427001953125
],
[
0.330810546875,
0.1319580078125,
0.363525390625,
0.1319580078125,
0.363525390625,
0.14453125,
0.330810546875,
0.14453125
],
[
0.38232421875,
0.1319580078125,
0.4150390625,
0.1319580078125,
0.4150390625,
0.14453125,
0.38232421875,
0.14453125
],
[
0.59619140625,
0.1319580078125,
0.67236328125,
0.1319580078125,
0.67236328125,
0.143310546875,
0.59619140625,
0.143310546875
],
[
0.7646484375,
0.1319580078125,
0.79736328125,
0.1319580078125,
0.79736328125,
0.143310546875,
0.7646484375,
0.143310546875
],
[
0.281005859375,
0.1331787109375,
0.31201171875,
0.1331787109375,
0.31201171875,
0.14453125,
0.281005859375,
0.14453125
],
[
0.181396484375,
0.1331787109375,
0.214111328125,
0.1331787109375,
0.214111328125,
0.14453125,
0.181396484375,
0.14453125
],
[
0.4345703125,
0.1331787109375,
0.465576171875,
0.1331787109375,
0.465576171875,
0.14453125,
0.4345703125,
0.14453125
],
[
0.69677734375,
0.1331787109375,
0.740234375,
0.1331787109375,
0.740234375,
0.143310546875,
0.69677734375,
0.143310546875
],
[
0.5947265625,
0.1427001953125,
0.67333984375,
0.143310546875,
0.67333984375,
0.1546630859375,
0.5947265625,
0.154052734375
],
[
0.69677734375,
0.143310546875,
0.740234375,
0.143310546875,
0.740234375,
0.1546630859375,
0.69677734375,
0.1546630859375
],
[
0.76123046875,
0.143310546875,
0.7998046875,
0.143310546875,
0.7998046875,
0.1546630859375,
0.76123046875,
0.1546630859375
],
[
0.59619140625,
0.1546630859375,
0.65673828125,
0.1546630859375,
0.65673828125,
0.1666259765625,
0.59619140625,
0.1666259765625
],
[
0.70166015625,
0.1546630859375,
0.734375,
0.1546630859375,
0.734375,
0.1666259765625,
0.70166015625,
0.1666259765625
],
[
0.7587890625,
0.154052734375,
0.7998046875,
0.154052734375,
0.7998046875,
0.1685791015625,
0.7587890625,
0.1685791015625
],
[
0.08087158203125,
0.1717529296875,
0.46728515625,
0.1717529296875,
0.46728515625,
0.186279296875,
0.08087158203125,
0.186279296875
],
[
0.08087158203125,
0.1849365234375,
0.468017578125,
0.1849365234375,
0.468017578125,
0.1988525390625,
0.08087158203125,
0.1988525390625
],
[
0.50341796875,
0.1817626953125,
0.890625,
0.1817626953125,
0.890625,
0.1956787109375,
0.50341796875,
0.1956787109375
],
[
0.08087158203125,
0.2008056640625,
0.468017578125,
0.2008056640625,
0.468017578125,
0.2139892578125,
0.08087158203125,
0.2139892578125
],
[
0.50244140625,
0.1976318359375,
0.8896484375,
0.1976318359375,
0.8896484375,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.08087158203125,
0.2139892578125,
0.468017578125,
0.2139892578125,
0.468017578125,
0.2279052734375,
0.08087158203125,
0.2279052734375
],
[
0.50341796875,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.50341796875,
0.2266845703125
],
[
0.08087158203125,
0.2279052734375,
0.468017578125,
0.2279052734375,
0.468017578125,
0.2412109375,
0.08087158203125,
0.2412109375
],
[
0.50341796875,
0.2279052734375,
0.890625,
0.2279052734375,
0.890625,
0.2412109375,
0.50341796875,
0.2412109375
],
[
0.08251953125,
0.24365234375,
0.465576171875,
0.24365234375,
0.465576171875,
0.25390625,
0.08251953125,
0.25390625
],
[
0.5009765625,
0.240478515625,
0.62353515625,
0.24365234375,
0.62353515625,
0.25830078125,
0.5009765625,
0.255126953125
],
[
0.07843017578125,
0.25390625,
0.1871337890625,
0.255126953125,
0.1871337890625,
0.26953125,
0.07843017578125,
0.268310546875
],
[
0.08087158203125,
0.26953125,
0.336669921875,
0.26953125,
0.336669921875,
0.2841796875,
0.08087158203125,
0.2841796875
],
[
0.5048828125,
0.28271484375,
0.890625,
0.28271484375,
0.890625,
0.296630859375,
0.5048828125,
0.296630859375
],
[
0.50244140625,
0.296630859375,
0.8896484375,
0.29736328125,
0.8896484375,
0.3125,
0.50244140625,
0.311279296875
],
[
0.080078125,
0.31005859375,
0.289306640625,
0.3125,
0.289306640625,
0.326904296875,
0.080078125,
0.324462890625
],
[
0.50341796875,
0.3125,
0.87744140625,
0.313232421875,
0.87744140625,
0.32763671875,
0.50341796875,
0.326904296875
],
[
0.50341796875,
0.328857421875,
0.73876953125,
0.328857421875,
0.73876953125,
0.34228515625,
0.50341796875,
0.34228515625
],
[
0.101318359375,
0.3359375,
0.468017578125,
0.3359375,
0.468017578125,
0.349853515625,
0.101318359375,
0.349853515625
],
[
0.08087158203125,
0.35107421875,
0.39208984375,
0.35107421875,
0.39208984375,
0.3642578125,
0.08087158203125,
0.3642578125
],
[
0.50244140625,
0.35107421875,
0.83837890625,
0.3515625,
0.83837890625,
0.3662109375,
0.50244140625,
0.365478515625
],
[
0.101318359375,
0.3662109375,
0.46728515625,
0.3662109375,
0.46728515625,
0.380126953125,
0.101318359375,
0.380126953125
],
[
0.080078125,
0.380126953125,
0.46728515625,
0.380615234375,
0.46728515625,
0.395751953125,
0.080078125,
0.39453125
],
[
0.52294921875,
0.375732421875,
0.890625,
0.375732421875,
0.890625,
0.389404296875,
0.52294921875,
0.389404296875
],
[
0.50244140625,
0.389404296875,
0.8896484375,
0.389404296875,
0.8896484375,
0.40283203125,
0.50244140625,
0.40283203125
],
[
0.08251953125,
0.3984375,
0.46728515625,
0.3984375,
0.46728515625,
0.4091796875,
0.08251953125,
0.4091796875
],
[
0.50244140625,
0.404052734375,
0.89208984375,
0.40478515625,
0.89208984375,
0.419189453125,
0.50244140625,
0.41845703125
],
[
0.08087158203125,
0.412353515625,
0.468017578125,
0.412353515625,
0.468017578125,
0.425537109375,
0.08087158203125,
0.425537109375
],
[
0.50341796875,
0.42041015625,
0.890625,
0.42041015625,
0.890625,
0.434326171875,
0.50341796875,
0.434326171875
],
[
0.08087158203125,
0.427978515625,
0.468017578125,
0.427978515625,
0.468017578125,
0.44140625,
0.08087158203125,
0.44140625
],
[
0.5048828125,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.448974609375,
0.5048828125,
0.448974609375
],
[
0.08251953125,
0.44140625,
0.29736328125,
0.44140625,
0.29736328125,
0.455078125,
0.08251953125,
0.455078125
],
[
0.1004638671875,
0.458251953125,
0.468017578125,
0.458251953125,
0.468017578125,
0.4716796875,
0.1004638671875,
0.4716796875
],
[
0.50634765625,
0.45263671875,
0.8896484375,
0.45263671875,
0.8896484375,
0.46337890625,
0.50634765625,
0.46337890625
],
[
0.50341796875,
0.466552734375,
0.52783203125,
0.466552734375,
0.52783203125,
0.479248046875,
0.50341796875,
0.479248046875
],
[
0.08087158203125,
0.472900390625,
0.468017578125,
0.472900390625,
0.468017578125,
0.486083984375,
0.08087158203125,
0.486083984375
],
[
0.08087158203125,
0.488037109375,
0.468017578125,
0.488037109375,
0.468017578125,
0.501953125,
0.08087158203125,
0.501953125
],
[
0.51318359375,
0.492431640625,
0.8544921875,
0.492431640625,
0.8544921875,
0.505859375,
0.51318359375,
0.505859375
],
[
0.08087158203125,
0.50244140625,
0.46728515625,
0.50244140625,
0.46728515625,
0.5166015625,
0.08087158203125,
0.5166015625
],
[
0.08251953125,
0.51806640625,
0.46728515625,
0.51806640625,
0.46728515625,
0.5322265625,
0.08251953125,
0.5322265625
],
[
0.5146484375,
0.5166015625,
0.81396484375,
0.5166015625,
0.81396484375,
0.52978515625,
0.5146484375,
0.52978515625
],
[
0.080078125,
0.5322265625,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.080078125,
0.54541015625
],
[
0.51318359375,
0.54052734375,
0.890625,
0.54052734375,
0.890625,
0.5537109375,
0.51318359375,
0.5537109375
],
[
0.08087158203125,
0.548828125,
0.3095703125,
0.548828125,
0.3095703125,
0.5625,
0.08087158203125,
0.5625
],
[
0.53515625,
0.55615234375,
0.81201171875,
0.55615234375,
0.81201171875,
0.5693359375,
0.53515625,
0.5693359375
],
[
0.098876953125,
0.56298828125,
0.468017578125,
0.56298828125,
0.468017578125,
0.57763671875,
0.098876953125,
0.57763671875
],
[
0.08251953125,
0.5791015625,
0.468017578125,
0.5791015625,
0.468017578125,
0.59228515625,
0.08251953125,
0.59228515625
],
[
0.5146484375,
0.5791015625,
0.67724609375,
0.5791015625,
0.67724609375,
0.59326171875,
0.5146484375,
0.59326171875
],
[
0.08251953125,
0.5947265625,
0.46728515625,
0.5947265625,
0.46728515625,
0.60791015625,
0.08251953125,
0.60791015625
],
[
0.08087158203125,
0.60791015625,
0.468017578125,
0.609375,
0.468017578125,
0.62353515625,
0.08087158203125,
0.62255859375
],
[
0.51123046875,
0.60302734375,
0.890625,
0.60400390625,
0.890625,
0.61865234375,
0.51123046875,
0.6171875
],
[
0.08087158203125,
0.625,
0.465576171875,
0.625,
0.465576171875,
0.63818359375,
0.08087158203125,
0.63818359375
],
[
0.5361328125,
0.619140625,
0.580078125,
0.619140625,
0.580078125,
0.630859375,
0.5361328125,
0.630859375
],
[
0.08087158203125,
0.6396484375,
0.468017578125,
0.6396484375,
0.468017578125,
0.65283203125,
0.08087158203125,
0.65283203125
],
[
0.50341796875,
0.64453125,
0.890625,
0.64306640625,
0.890625,
0.65771484375,
0.50341796875,
0.6591796875
],
[
0.08087158203125,
0.65380859375,
0.46728515625,
0.65380859375,
0.46728515625,
0.66845703125,
0.08087158203125,
0.66845703125
],
[
0.5048828125,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.5048828125,
0.673828125
],
[
0.08087158203125,
0.66943359375,
0.468017578125,
0.66943359375,
0.468017578125,
0.68310546875,
0.08087158203125,
0.68310546875
],
[
0.50244140625,
0.6748046875,
0.70361328125,
0.6748046875,
0.70361328125,
0.68798828125,
0.50244140625,
0.68798828125
],
[
0.08087158203125,
0.68505859375,
0.468017578125,
0.68505859375,
0.468017578125,
0.69873046875,
0.08087158203125,
0.69873046875
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.08251953125,
0.69970703125,
0.46728515625,
0.69970703125,
0.46728515625,
0.71337890625,
0.08251953125,
0.71337890625
],
[
0.50341796875,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.08087158203125,
0.71533203125,
0.468017578125,
0.71533203125,
0.468017578125,
0.72900390625,
0.08087158203125,
0.72900390625
],
[
0.50341796875,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.50341796875,
0.734375
],
[
0.08087158203125,
0.72998046875,
0.45751953125,
0.72998046875,
0.45751953125,
0.744140625,
0.08087158203125,
0.744140625
],
[
0.50244140625,
0.734375,
0.890625,
0.73486328125,
0.890625,
0.75,
0.50244140625,
0.74853515625
],
[
0.08087158203125,
0.7568359375,
0.396240234375,
0.7568359375,
0.396240234375,
0.771484375,
0.08087158203125,
0.771484375
],
[
0.5048828125,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.5048828125,
0.76513671875
],
[
0.5048828125,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.5048828125,
0.77978515625
],
[
0.101318359375,
0.77978515625,
0.468017578125,
0.77978515625,
0.468017578125,
0.79443359375,
0.101318359375,
0.79443359375
],
[
0.5048828125,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.5048828125,
0.79345703125
],
[
0.08087158203125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.50244140625,
0.79443359375,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.50244140625,
0.8095703125
],
[
0.08087158203125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.08087158203125,
0.82373046875
],
[
0.5048828125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5048828125,
0.82470703125
],
[
0.08087158203125,
0.82568359375,
0.2607421875,
0.82568359375,
0.2607421875,
0.8388671875,
0.08087158203125,
0.8388671875
],
[
0.5048828125,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.101318359375,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.101318359375,
0.85498046875
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.50244140625,
0.85595703125,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.869140625
],
[
0.08087158203125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50341796875,
0.87060546875,
0.88818359375,
0.8720703125,
0.88818359375,
0.88623046875,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.74853515625,
0.88720703125,
0.74853515625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"=35",
"g=50",
"PSNR [dB]",
"SSIM",
"g=10",
"a=5",
"15",
"q:",
"=25",
"Spatial-CNN",
"",
"35.92",
"33.66",
"30.99",
"29.34",
"27.63",
"Spatial-CNN",
"37.0343",
"0.9324",
"30.69",
"39.73",
"35.87",
"33.57",
"26.53",
"28.74",
"DnCNN-B*[14]",
"CBDNet[22]",
"0.9421",
"38.0564",
"39.79",
"DnCNN-B [14]",
"36.14",
"31.22",
"29.57",
"DnCNN+[14]",
"0.943",
"33.88",
"40.62",
"27.91",
"37.9018",
"FFDNet+[16",
"37.6107",
"0.9415",
"BM3D[3]]",
"34.51",
"0.8507",
"Table l: Comparison of blind Gaussian denoising on",
"the CBSD68 dataset. Our modified version of DnCNN",
"Table 2: Results of the DND benchmark [31] on real-",
"for spatial denoising has comparable results with the",
"world noisy images. It shows that our dataset, contain-",
"original one.The values represent PSNR[dB], the",
"ing different noise models, is valid for real-world image",
"higher the better. DnCNN results obtained with the",
"denoising, placing our Spatial-CNN in the top 10 for",
"providedMatlabimplementation38CBSD68avail-",
"sRGB denoising.",
"able here [39] .",
"*Noisy images clipped in range [0,255]",
"，SSIM的0.9324，将其置于前10个位置。有趣的是，",
"DnCNN的作者在一周后提交了一个名为DnCNN+的",
"4.2. Spatial CNN Training",
"微调模型的结果，获得了SSIM的总体最高分，这进",
"步验证了其灵活性，请参见表2",
"训练分为两部分：（i）我们训练空间去噪CNN，以",
"及（2）在收敛后，我们训练时间去噪CNN",
"4.4. Temp3-CNN: Temporal CNN Training",
"我们的理想模型必须同时处理多种退化类型，例如",
"AWGN和真实噪声模型2，包括弱光条件。在训练阶",
"对于视频评估，我们需要成对的干净和噜杂的视频。",
"对于人为添加的噪声，如加性白高斯噪声（AWGN）或",
"段：我们的神经网络将学习如何使用干净的图像作为",
"方程2中的真实噪声模型，很容易创建这样的耦合。然",
"参考来估计输入噪声图像的残余噪声含量。因此，我们",
"而，对于现实世界和弱光条件下的视频，这几乎是不可",
"需要一对干净而噜杂的图像。它们很容易为AWGN和",
"能的。因此，这种提供成对的噪声和无噪声序列的视频",
"方程2中的真实噪声模型创建",
"我们使用滑铁卢探索数据集[36】，其中包含4,744",
"数据集不可用。因此，我们决定按照以下顺序进行操",
"作：",
"张干净的图像。可用图像的数量大大有助于概括，并允",
"许我们保留其中的大部分用于测试。数据集随机分为",
"1. Select 31 publicly available videos from [41]",
"两部分，70个%用于训练，30个%用于测试。一半的",
"图像正在添加AWGN与α=[0,55]。后半部分使用公",
"2. Divide videos in sequences of 3 frames.",
"式2进行处理，该公式是真实的噪声模型，模拟增益",
"3. Added either Gaussian noise with =[0,55] or real",
"Ag-[0,64]，数字增益Dg=[0,32]",
"noise 2 with Ag[0,64] and Dg=[0,32].",
"在「14]之后，使用50×50×3补丁训练网络。我",
"们从滑铁卢训练集中获得了120.000个补丁，其中包",
"4. Apply Spatial-CNN",
"含AWGN和真实噪声类型，使用旋转，翻转和镜像",
"等数据增强。对于低光照条件，我们从自已的训练数",
"5. Train on pairs of spatially-denoised and clean",
"据集中为每个光照水平使用了五幅噪声图像，获得了",
"video.",
"60对噪声干净的图像进行训练。提取的补丁为80,000。",
"我们遵循与Spatial-CNN相同的训练过程，即使现在",
"从Renoir数据集中，我们使用子集T3并随机裁剪了",
"网络将使用50×50×9维的补丁进行训练，其中包含",
"40,000个补丁。对于低光测试，我们将使用来自不同",
"来自三个后续顿的三个补丁",
"场景的相机的5张图像，这些图像不存在于训练集中，",
"选定的31个视频包含8922顿，这意味着2974个三",
"并且是雷诺阿T3集的一部分。我们训练了100个时",
"顿序列和300032的最终补丁数。我们运行了60个",
"期，使用了一批128个和AdamOptimizer[37]，前20",
"epoch的训练，批处理大小为128，Adam优化器以",
"个时期的学习率为10-3，最新的80个时期为10-4",
"10-4的学习速率和LeakyReLU作为激活函数。它显",
"4.3. Validation of static Image Denoising",
"示了LeakyReLU可以胜过ReLU[34]。但是，我们没",
"有在空间CNN中使用LeakyRelu，因为ReLU的性能",
"我们将盲高斯去噪与我们所基于的DnCNN的原始",
"更好。我们在补充材料中介绍了比较结果。在Temp3-",
"实现进行了比较。从我们在BSD68测试集上的表1中",
"CNN的最终版本中，未使用批量归一化（BN)：实验",
"的测试中，我们注意到我们的盲模型的结果与[14］论",
"表明它减慢了训练和去噪过程。BN在PSNR方面没",
"文提出的结果具有可比性",
"有改善最终结果。此外，没有BN的去噪需要大约5%",
"为了进一步验证真实世界的图像，我们评估",
"更少的时间。图5表示Temp3-CNN的L2损失的演",
"sRGBDND数据集[31]并提交评估。结果【40】令人",
"变：避免归一化步骤会使损失立即从低值开始。我们还",
"鼓舞，因为我们的训练模型（在DND网页上称为128-",
"使用LeakyReLU，LeakyReLUBN和ReLUBN进行了",
"DnCNNTensorflow）的PSNR平均得分为37.0343dB",
"培训，并在补充材料中介绍了结果"
] |
{
"arxiv_id": "1904.10898",
"language": "zh",
"page_idx": 3
}
|
||
1903.02351-zh-0
|
1903.02351-zh-0.jpg
|
CANet: Class-Agnostic Segmentation Networks with Iterative Refinement and
Attentive Few-Shot Learning
$\begin{array}{ll}{\mathrm{Chi~Zhang~}^{1},}&{\mathrm{Guosheng~Lin~}^{1};}^{*}&{\mathrm{Fayao~Liu~}^{2},}&{\mathrm{Rui~Yao~}^{3},}\\{{}^{1}\text{Nanyang Technological University,Singapore}}\\{{}^{2}\text{The University of Adelaide. Australia}}\end{array}$
$^{2}$The University of Adelaide, Australia
$^{3}$China University of Mining and Technology, China
chi$007@$e.ntu.edu.sg gslin@ntu.edu.sg
## Abstract
Recent progress in semantic segmentation is driven by deep Convolutional Neural Wetworks and largescale labeled image datasets. However, data labeling for pixel-wise segmentation is tedious and costly. $\bar{\text{Moreover, a trained model can only make predictions}}$ within a set of pre-defined classes. In this paper, we present CANet, a class-agnostic segmentation network that performs few-shot segmentation on new classes with only a few annotated images available. Our network consists of a two-branch dense comparison module which performs multi-level feature comparison between the support image and the query image, and an iterative optimization module which iteratively refines the predicted results. Furthermore, we introduce an attention mechanism to effectively fuse information from multiple support examples under the setting of $k$ -shot learning. Experiments on PASCAL VOC 2012 show that our method achieves a mean Intersectionover-Union score of $55.4\%$ for 1-shot segmentation and $57.1\%$ for 5-shot segmentation, outperforming stateof-the-art methods by a large margin of 14.6 % and $13.2\%$,respectively.
### 1. Introduction
深度卷积神经网络在许多视觉理解任务中取得了重大突破，包括图像分类E[13,9,30],对象检测回[27,8,26] 和语义分割$\mathbb{E}[16,2,20]$ 。一个关键的原因是大规模数据集的可用性，如 ImageNet [4],可以训练深度模型。然而，数据标记是昂贵的，特别是对于密集的预测任务，e.g. , 语义分割和实例分割。除此之外， 在训练模型之后，很难将模型应用于预测新类。与机器学习算法相比，当只看到几个例子时，人类能够很容易地从图像中分割出新的概念。人类和机器学习算法之
${^*\text{Guosheng Lin is the corresponding author}.}$
<FigureHere>
Figure 1-Overview of our proposed network for l-shot segmentation. Our framework consists of a dense comparison module (DCM) and an iterative optimization module (IOM). Given only one annotated training image, our network is able to segment test images with new classes and iteratively optimize the results.
间的差距激发了对少射学习的研究，该研究旨在学习-种模型，该模型可以很好地推广到具有稀缺标签训练数据的新类。
在本文中，我们承担了少量语义分割的任务，该任务仅使用少量带注释的训练图像对新类进行分割。关于此任务的先前工作[29,24,5]遵循双分支结构的设计，该结构包括一个支持分支和一个查询分支。支持分支旨在从支持集中提取信息，以指导查询分支中的分割。我们还在框架中采用了两分支设计来解决少拍分割问题
我们的网络包括一个两分支密集比较模块，其中共享的特征提取器从查询集和支持集中提取表示以进行比较。密集比较模块的设计灵感来自于图像分类任务上的度量学习[37,31] ,其中距离函数评估图像之间的相似性。但是，与每个图像都有一个标签的图像分类不同，图像分割需要对具有结构化表示的数据进行预测。很难将度量学习直接应用于密集预测问题。为了解决这个问题，一种简单的方法是在所有像素对之间进行比较。但是，图像中有数百万个像素，所有像素对的比较需要花费大量的计算成本。相反，我们的目标是从支持图像中获取全局表示以进行比较。全局图像特征被证明在分割任务中很有用[19,40,3] ,这可以通过全局平均池化轻松实现。在这里，为了只关注分配的类
1
|
[
[
0.09478759765625,
0.1337890625,
0.87841796875,
0.1337890625,
0.87841796875,
0.1483154296875,
0.09478759765625,
0.1483154296875
],
[
0.339111328125,
0.1546630859375,
0.63232421875,
0.1572265625,
0.63232421875,
0.1749267578125,
0.339111328125,
0.17236328125
],
[
0.1356201171875,
0.203857421875,
0.259033203125,
0.205810546875,
0.259033203125,
0.2215576171875,
0.1356201171875,
0.2197265625
],
[
0.273681640625,
0.2052001953125,
0.69921875,
0.2052001953125,
0.69921875,
0.2197265625,
0.273681640625,
0.2197265625
],
[
0.69287109375,
0.2052001953125,
0.8349609375,
0.2052001953125,
0.8349609375,
0.2197265625,
0.69287109375,
0.2197265625
],
[
0.289306640625,
0.2235107421875,
0.6806640625,
0.2235107421875,
0.6806640625,
0.238037109375,
0.289306640625,
0.238037109375
],
[
0.25732421875,
0.208984375,
0.27294921875,
0.208984375,
0.27294921875,
0.216552734375,
0.25732421875,
0.216552734375
],
[
0.314453125,
0.23486328125,
0.654296875,
0.238037109375,
0.654296875,
0.2587890625,
0.314453125,
0.255615234375
],
[
0.2607421875,
0.255126953125,
0.708984375,
0.2568359375,
0.708984375,
0.274658203125,
0.2607421875,
0.272705078125
],
[
0.316162109375,
0.2734375,
0.654296875,
0.274658203125,
0.654296875,
0.292236328125,
0.316162109375,
0.291748046875
],
[
0.23779296875,
0.32568359375,
0.313720703125,
0.32763671875,
0.313720703125,
0.343505859375,
0.23779296875,
0.341552734375
],
[
0.68896484375,
0.34033203125,
0.73876953125,
0.34033203125,
0.73876953125,
0.359130859375,
0.68896484375,
0.359130859375
],
[
0.73193359375,
0.341552734375,
0.7646484375,
0.341552734375,
0.7646484375,
0.356689453125,
0.73193359375,
0.356689453125
],
[
0.09722900390625,
0.35791015625,
0.468017578125,
0.359130859375,
0.468017578125,
0.373779296875,
0.09722900390625,
0.37255859375
],
[
0.08087158203125,
0.375,
0.46728515625,
0.375,
0.46728515625,
0.38818359375,
0.08087158203125,
0.38818359375
],
[
0.080078125,
0.38818359375,
0.46728515625,
0.389404296875,
0.46728515625,
0.404052734375,
0.080078125,
0.40283203125
],
[
0.515625,
0.3984375,
0.52392578125,
0.3984375,
0.52392578125,
0.401611328125,
0.515625,
0.401611328125
],
[
0.08087158203125,
0.40478515625,
0.468017578125,
0.40478515625,
0.468017578125,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.08087158203125,
0.42041015625,
0.468017578125,
0.42041015625,
0.468017578125,
0.434326171875,
0.08087158203125,
0.434326171875
],
[
0.67724609375,
0.421630859375,
0.689453125,
0.421630859375,
0.689453125,
0.4248046875,
0.67724609375,
0.4248046875
],
[
0.08087158203125,
0.43505859375,
0.468017578125,
0.43505859375,
0.468017578125,
0.448974609375,
0.08087158203125,
0.448974609375
],
[
0.53125,
0.43505859375,
0.54736328125,
0.43505859375,
0.54736328125,
0.439453125,
0.53125,
0.439453125
],
[
0.08087158203125,
0.45068359375,
0.468017578125,
0.45068359375,
0.468017578125,
0.464111328125,
0.08087158203125,
0.464111328125
],
[
0.5087890625,
0.455078125,
0.57275390625,
0.455078125,
0.57275390625,
0.46533203125,
0.5087890625,
0.46533203125
],
[
0.5703125,
0.453857421875,
0.7099609375,
0.455078125,
0.7099609375,
0.46533203125,
0.5703125,
0.464111328125
],
[
0.7236328125,
0.455078125,
0.8798828125,
0.455078125,
0.8798828125,
0.46533203125,
0.7236328125,
0.46533203125
],
[
0.080078125,
0.46533203125,
0.468017578125,
0.46533203125,
0.468017578125,
0.479248046875,
0.080078125,
0.479248046875
],
[
0.51123046875,
0.468505859375,
0.5810546875,
0.468505859375,
0.5810546875,
0.47607421875,
0.51123046875,
0.47607421875
],
[
0.57666015625,
0.46728515625,
0.88232421875,
0.46728515625,
0.88232421875,
0.477783203125,
0.57666015625,
0.477783203125
],
[
0.08087158203125,
0.479736328125,
0.46728515625,
0.479736328125,
0.46728515625,
0.494384765625,
0.08087158203125,
0.494384765625
],
[
0.5107421875,
0.4765625,
0.8857421875,
0.4765625,
0.8857421875,
0.490478515625,
0.5107421875,
0.490478515625
],
[
0.50732421875,
0.4873046875,
0.88623046875,
0.488037109375,
0.88623046875,
0.50244140625,
0.50732421875,
0.501953125
],
[
0.08087158203125,
0.49560546875,
0.46728515625,
0.49560546875,
0.46728515625,
0.5087890625,
0.08087158203125,
0.5087890625
],
[
0.51123046875,
0.50244140625,
0.8837890625,
0.50244140625,
0.8837890625,
0.51318359375,
0.51123046875,
0.51318359375
],
[
0.080078125,
0.5087890625,
0.464111328125,
0.51025390625,
0.464111328125,
0.5244140625,
0.080078125,
0.5234375
],
[
0.51123046875,
0.5146484375,
0.611328125,
0.5146484375,
0.611328125,
0.5244140625,
0.51123046875,
0.5244140625
],
[
0.080078125,
0.52587890625,
0.468017578125,
0.52587890625,
0.468017578125,
0.54052734375,
0.080078125,
0.54052734375
],
[
0.08087158203125,
0.54052734375,
0.464111328125,
0.54052734375,
0.464111328125,
0.5537109375,
0.08087158203125,
0.5537109375
],
[
0.5048828125,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.5048828125,
0.56103515625
],
[
0.080078125,
0.55615234375,
0.468017578125,
0.55615234375,
0.468017578125,
0.5693359375,
0.080078125,
0.5693359375
],
[
0.51708984375,
0.56103515625,
0.890625,
0.5625,
0.890625,
0.5771484375,
0.51708984375,
0.57568359375
],
[
0.08251953125,
0.57275390625,
0.46728515625,
0.57275390625,
0.46728515625,
0.58349609375,
0.08251953125,
0.58349609375
],
[
0.5048828125,
0.5791015625,
0.609375,
0.5791015625,
0.609375,
0.59130859375,
0.5048828125,
0.59130859375
],
[
0.08087158203125,
0.58837890625,
0.46728515625,
0.58837890625,
0.46728515625,
0.5986328125,
0.08087158203125,
0.5986328125
],
[
0.52392578125,
0.59521484375,
0.890625,
0.59521484375,
0.890625,
0.609375,
0.52392578125,
0.609375
],
[
0.08251953125,
0.6015625,
0.468017578125,
0.6015625,
0.468017578125,
0.61572265625,
0.08251953125,
0.61572265625
],
[
0.08087158203125,
0.6162109375,
0.464111328125,
0.6162109375,
0.464111328125,
0.6298828125,
0.08087158203125,
0.6298828125
],
[
0.5048828125,
0.611328125,
0.890625,
0.611328125,
0.890625,
0.625,
0.5048828125,
0.625
],
[
0.080078125,
0.6298828125,
0.468017578125,
0.630859375,
0.468017578125,
0.64599609375,
0.080078125,
0.64453125
],
[
0.5048828125,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.64013671875,
0.5048828125,
0.64013671875
],
[
0.08087158203125,
0.646484375,
0.46728515625,
0.646484375,
0.46728515625,
0.66015625,
0.08087158203125,
0.66015625
],
[
0.50341796875,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65576171875,
0.50341796875,
0.65576171875
],
[
0.08087158203125,
0.6611328125,
0.468017578125,
0.6611328125,
0.468017578125,
0.67529296875,
0.08087158203125,
0.67529296875
],
[
0.5048828125,
0.6572265625,
0.890625,
0.6572265625,
0.890625,
0.67041015625,
0.5048828125,
0.67041015625
],
[
0.08087158203125,
0.6767578125,
0.233642578125,
0.6767578125,
0.233642578125,
0.68994140625,
0.08087158203125,
0.68994140625
],
[
0.50341796875,
0.671875,
0.890625,
0.671875,
0.890625,
0.68603515625,
0.50341796875,
0.68603515625
],
[
0.50341796875,
0.68798828125,
0.5537109375,
0.68798828125,
0.5537109375,
0.69970703125,
0.50341796875,
0.69970703125
],
[
0.52294921875,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.52294921875,
0.7197265625
],
[
0.080078125,
0.72021484375,
0.2156982421875,
0.72021484375,
0.2156982421875,
0.734375,
0.080078125,
0.734375
],
[
0.5048828125,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.5048828125,
0.734375
],
[
0.50341796875,
0.736328125,
0.8896484375,
0.736328125,
0.8896484375,
0.75,
0.50341796875,
0.75
],
[
0.101318359375,
0.744140625,
0.46728515625,
0.744140625,
0.46728515625,
0.75830078125,
0.101318359375,
0.75830078125
],
[
0.5048828125,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.5048828125,
0.76513671875
],
[
0.08087158203125,
0.76025390625,
0.468017578125,
0.76025390625,
0.468017578125,
0.7734375,
0.08087158203125,
0.7734375
],
[
0.50341796875,
0.7666015625,
0.8896484375,
0.7666015625,
0.8896484375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08251953125,
0.7763671875,
0.346435546875,
0.7763671875,
0.346435546875,
0.787109375,
0.08251953125,
0.787109375
],
[
0.363525390625,
0.7744140625,
0.468017578125,
0.7744140625,
0.468017578125,
0.7890625,
0.363525390625,
0.7890625
],
[
0.5048828125,
0.78271484375,
0.88818359375,
0.78271484375,
0.88818359375,
0.79345703125,
0.5048828125,
0.79345703125
],
[
0.08087158203125,
0.79052734375,
0.468017578125,
0.79052734375,
0.468017578125,
0.8037109375,
0.08087158203125,
0.8037109375
],
[
0.5048828125,
0.796875,
0.8896484375,
0.796875,
0.8896484375,
0.81005859375,
0.5048828125,
0.81005859375
],
[
0.08087158203125,
0.80615234375,
0.468017578125,
0.80615234375,
0.468017578125,
0.8193359375,
0.08087158203125,
0.8193359375
],
[
0.50341796875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08087158203125,
0.82080078125,
0.46728515625,
0.82080078125,
0.46728515625,
0.833984375,
0.08087158203125,
0.833984375
],
[
0.50341796875,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.08087158203125,
0.83544921875,
0.46728515625,
0.83544921875,
0.46728515625,
0.8486328125,
0.08087158203125,
0.8486328125
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.08251953125,
0.85107421875,
0.46728515625,
0.85107421875,
0.46728515625,
0.8642578125,
0.08251953125,
0.8642578125
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.08087158203125,
0.86572265625,
0.46728515625,
0.86572265625,
0.46728515625,
0.87890625,
0.08087158203125,
0.87890625
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.09722900390625,
0.88623046875,
0.363525390625,
0.88720703125,
0.363525390625,
0.90087890625,
0.09722900390625,
0.8994140625
],
[
0.50341796875,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.491943359375,
0.9248046875,
0.491943359375,
0.9375,
0.48046875,
0.9375
]
] |
[
"CANet: Class-Agnostic Segmentation Networks with Iterative Refinement and",
"Attentive Few-Shot Learning",
"Chi Zhang 1,",
"Guosheng Lin 1*Fayao Liu2，Rui Yao 3，",
"Chunhua Shen 2",
"1 Nanyang Technological University, Singapore",
"",
"2 The University of Adelaide, Australia",
"3 China University of Mining and Technology, China",
"chi007@e.ntu.edu.sggslin@ntu.edu.sg",
"Abstract",
"3",
"福",
"Recent progress in semantic segmentation is driven",
"by deep Convolutional Neural Networks and large-",
"scale labeled image datasets.However, data label-",
"",
"ing for pixel-wise segmentation is tedious and costly.",
"Moreover, a trained model can only make predictions",
"Taer",
"within a set of pre-defined classes. In this paper, we",
"Duen",
"present CANet, a class-agnostic segmentation network",
"Figure 1",
"Overviewofourprop",
"ed networkfor 1-shotseg",
"that performs few-shot segmentation on new classes",
"mentation.",
"Our framework consists of a dense comparison",
"with only a few annotated images available. Our net-",
"module (DCM) and an iterative optimization module (IOM).",
"Given only one annotated training image, our network is able",
"work consists of a two-branch dense comparison mod-",
"tosegmenttest imageswithnewclassesanditerativelyopti-",
"ule which performs multi-level feature comparison be",
"mize the results.",
"tween the support image and the query image, and",
"an iterative optimization module which iteratively re",
"间的差距激发了对少射学习的研究，该研究旨在学习",
"fines the predicted results. Furthermore, we introduce",
"种模型，该模型可以很好地推广到具有稀缺标签训",
"anattentionmechanismtoeffectivelyfuseinformation",
"练数据的新类。",
"frommultiplesupportexamplesunderthesettingofk",
"在本文中，我们承担了少量语义分割的任务，该任",
"-shot learning.Experiments on PASCAL VOC 2012",
"show that our method achieves a mean Intersection",
"务仅使用少量带注释的训练图像对新类进行分割。关",
"over-Union score of 55.4 %for 1-shot segmentation and",
"于此任务的先前工作29，24，5】遵循双分支结构的设",
"57.1 % for 5-shot segmentation, outperforming state-",
"计，该结构包括一个支持分支和一个查询分支。支持分",
"of-the-art methods by a large margin of 14.6 % and",
"支旨在从支持集中提取信息，以指导查询分支中的分",
"13.2 % , respectively.",
"割。我们还在框架中采用了两分支设计来解决少拍分",
"割问题",
"我们的网络包括一个两分支密集比较模块，其中共",
"1.Introduction",
"享的特征提取器从查询集和支持集中提取表示以进行",
"比较。密集比较模块的设计灵感来自于图像分类任务",
"深度卷积神经网络在许多视觉理解任务中取得了",
"上的度量学习【37，31]，其中距离函数评估图像之间",
"重大突破，包括图像分类E13，9，301，对象检测E",
"的相似性。但是，与每个图像都有一个标签的图像分类",
"27.8.26和语义分割E16.2.20。",
"-个关键的原因",
"不同，图像分割需要对具有结构化表示的数据进行预",
"是大规模数据集的可用性，如ImageNet「4，可以训",
"测。很难将度量学习直接应用于密集预测问题。为了解",
"练深度模型。然而，数据标记是昂贵的，特别是对于密",
"决这个问题，一种简单的方法是在所有像素对之间进",
"集的预测任务，e.g：，语义分割和实例分割。除此之外",
"行比较。但是，图像中有数百万个像素，所有像素对的",
"在训练模型之后，很难将模型应用于预测新类。与机器",
"比较需要花费大量的计算成本。相反，我们的目标是从",
"学习算法相比，当只看到几个例子时，人类能够很容易",
"支持图像中获取全局表示以进行比较。全局图像特征",
"地从图像中分割出新的概念。人类和机器学习算法之",
"被证明在分割任务中很有用[19，40,31，这可以通过",
"*Guosheng Lin is the corresponding author.",
"全局平均池化轻松实现。在这里，为了只关注分配的类",
"1"
] |
{
"arxiv_id": "1903.02351",
"language": "zh",
"page_idx": 0
}
|
||
1903.02351-zh-1
|
1903.02351-zh-1.jpg
|
别，我们使用前景区域的全局平均池化来过滤掉不相关的信息。然后将全局特征与查询分支中的每个位置进行比较，这可以看作是度量学习方法的一种密集形式
在 few-shot 设置下，网络应该能够处理在训练过程中从未见过的新类。因此，我们的目标是从 CNN 中挖掘可转移的表示形式以进行比较。正如在特征可视化文献$\mathbb{E}[39,38]$ 中观察到的那样，较低层中的特征与低级提示，e.g. , 边缘和颜色有关，而较高层中的特征与对象级概念有关，例如类别。我们专注于可能构成由看不见的类共享的对象部分的中间层特征。例如，如果CNN 在类车上训练模型时学习了与车轮相关的特征， 则该特征也可能对新车辆类别 e.g. , 卡车和公共汽车的特征比较有用。我们在 CNN 中提取多个级别的表示以进行密集比较
由于同一类别中的外观存在差异，因此来自同一类的对象可能仅共享一些相似的功能。密集的特征比较不足以指导整个对象区域的分割。然而，这给出了一个重要的线索，即物体在哪里。在半自动分割文献中， 为类不可知分割，e.g. ,带有单击或涂鸦注释的交互式分割$\mathbb{E}[36,14]$ 以及带有边界框或极端点先验的实例分割$\mathbb{E}[10,21]$ 给出了弱注释。在训练过程中学习可转移的知识来定位对象区域。受半自动分割任务的启发，我们希望根据先验的密集比较结果逐步将对象与背景区分开。我们提出了一个迭代优化模块 (IOM),它学习迭代地优化预测结果。细化以循环形式执行，将密集比较结果和预测掩码发送到 IOM 进行优化，并将输出循环发送到下一个 IOM。经过几次细化迭代后，我们的密集比较模块能够生成细粒度的分割图。在每个 IOM 内部，我们采用剩余连接，以便在最后一个迭代步骤中有效地合并预测的掩码。Fig。1 显示了我们的一次性分割网络的概述
以前的$k$ -shot 分割方法基于单镜头模型。他们使用不可学习的融合方法来融合单个单发结果，e.g. , 平均单发预测或中间特征。相反，我们采用注意力机制来有效地融合来自多个支持示例的信息
为了进一步减少少拍分割的标记工作，我们探索了一种新的测试设置：我们的模型使用边界框注释支持集在查询图像中执行分割。我们对 PASCALVOC2012数据集和 COCO 数据集进行了全面的实验，以验证我们网络的有效性。本文的主要贡献总结如下。
$\cdot{\mathrm{~We~develop~a~novel~two-branch~dense~comparison}}$
module which effectively exploits multiple levels of feature representations from CNNs to make dense feature comparison.
$\bullet$ We propose an iterative optimization module to
refine predicted results in an iterative manner. The ability of iterative refinement can be generalized to unseen classes with few-shot learning for generating fine-grained maps.
$\bullet$ We adopt an attention mechanism to effectively
fuse information from multiple support examples
in the $k$ -shot setting, which outperforms non-
learnable fusion methods of 1-shot results.
$\bullet$ We demonstrate that given support set with weak
annotations, i.e., bounding boxes, our model can still achieve comparable performance to the result with expensive pixel-level annotated support set, which further reduces the labeling efforts of new classes for few-shot segmentation significantly.
$\bullet$Experiments on the PASCALVOC 2012
dataset show that our method achieves a mean Intersection-over-Union score of 55.4 % for 1-shot segmentation and 57.1 % for 5-shot segmentation, which significantly outperform state-of-the-art results by 14.6% and 13.2%, respectively.
## 2. Related Work
语义分割。语义分割是将图像中的每个像素分类为一组预定义类别的任务[16,2,20,15,17] 。最先进的方法基于完全卷积网络(FCN),它通常采用预先训练的卷积神经网络(CNN) 作为骨干架构进行分类。为了适应密集预测的任务，完全连接的层被一个卷积层代替，该卷积层预测每个像素的标签。为了捕获抽象的特征表示，CNN 采用连续的池化操作或卷积步幅来降低特征图的空间分辨率。然而，这与密集的预测任务冲突，其中输出应该是高分辨率的。为了平衡网络的输出分辨率和感受野，在密集的预测任务中经常使用扩张卷积[2]。扩张去除了最后几层中的下采样操作，并在卷积滤波器中插人孔以扩大感受野。在我们的模型中， 我们还采用扩张卷积来保持空间分辨率。在完全监督分割中，训练 FCN 模型需要大量昂贵的像素级注释图像， 并且一旦训练了模型，就无法对新类别进行分割。Our model, on the other hand, can be generalized to any new categories with only a few annotated examples.
少拍学习。很少的学习旨在学习可转移的知识，这些知识可以推广到具有稀缺标记的训练数据的新类。关于少射分类存在许多公式，包括带有记忆的递归神经网络$\mathbb{E}[28,23]$ ,学习微调模型$\mathbb{E}[6,25]$ ,网络参数预测$\boxed{\mathrm{F}}[1,35]$ 和度量学习$\boxed{\mathrm{E}}[31,37,11]$ 。基于度量学习的方法在少量分类任务中实现了最先进的性能，并且具有快速和以前馈方式进行预测的特点。我们的工作是最相关的关系网络 [37] 。关系网络元学习深度距离度量以比较图像并计算相似性得分以进行分类。该网络由一个生成图像表示的嵌入模块和一个比较嵌人并输出相似性得分的关系模块组成。这两个模块都是卷积运算的形式。我们网络中的密集比较模块可以看作是关系网络的扩展，以密集的形式处理分割任务
少拍语义分割。先前关于少射语义分割的工作采用了两个分支结构。Shaban et al. [29] 首次采用语义分割的少射学习。支持分支直接预测查询分支中最后一层的权重进行分割。在$\mathbb{E}[24]$ 中，支持分支生成嵌入， 该嵌人作为附加功能融合到查询分支中。我们的网络也遵循两分支设计。但是，与以前的两个分支具有不同
2
|
[
[
0.08251953125,
0.09344482421875,
0.46728515625,
0.09344482421875,
0.46728515625,
0.106689453125,
0.08251953125,
0.106689453125
],
[
0.53369140625,
0.091552734375,
0.8896484375,
0.0921630859375,
0.8896484375,
0.106689453125,
0.53369140625,
0.1060791015625
],
[
0.08251953125,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.121826171875,
0.08251953125,
0.121826171875
],
[
0.53515625,
0.10797119140625,
0.83837890625,
0.10797119140625,
0.83837890625,
0.121826171875,
0.53515625,
0.121826171875
],
[
0.080078125,
0.121826171875,
0.468017578125,
0.12249755859375,
0.468017578125,
0.136962890625,
0.080078125,
0.1363525390625
],
[
0.5185546875,
0.125,
0.890625,
0.125,
0.890625,
0.1383056640625,
0.5185546875,
0.1383056640625
],
[
0.08416748046875,
0.1383056640625,
0.1004638671875,
0.14013671875,
0.09722900390625,
0.15087890625,
0.08087158203125,
0.1483154296875
],
[
0.5361328125,
0.1395263671875,
0.890625,
0.1395263671875,
0.890625,
0.15283203125,
0.5361328125,
0.15283203125
],
[
0.098876953125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.098876953125,
0.1673583984375
],
[
0.53759765625,
0.1558837890625,
0.890625,
0.1558837890625,
0.890625,
0.1666259765625,
0.53759765625,
0.1666259765625
],
[
0.08251953125,
0.169189453125,
0.468017578125,
0.169189453125,
0.468017578125,
0.18310546875,
0.08251953125,
0.18310546875
],
[
0.5361328125,
0.169189453125,
0.890625,
0.169189453125,
0.890625,
0.18310546875,
0.5361328125,
0.18310546875
],
[
0.08087158203125,
0.1849365234375,
0.468017578125,
0.1849365234375,
0.468017578125,
0.1988525390625,
0.08087158203125,
0.1988525390625
],
[
0.5390625,
0.186279296875,
0.88818359375,
0.186279296875,
0.88818359375,
0.1962890625,
0.5390625,
0.1962890625
],
[
0.08251953125,
0.2008056640625,
0.468017578125,
0.2008056640625,
0.468017578125,
0.2139892578125,
0.08251953125,
0.2139892578125
],
[
0.5361328125,
0.199462890625,
0.86279296875,
0.2008056640625,
0.86279296875,
0.212158203125,
0.5361328125,
0.2108154296875
],
[
0.08087158203125,
0.21533203125,
0.46728515625,
0.21533203125,
0.46728515625,
0.2291259765625,
0.08087158203125,
0.2291259765625
],
[
0.51953125,
0.2171630859375,
0.66259765625,
0.2171630859375,
0.66259765625,
0.2279052734375,
0.51953125,
0.2279052734375
],
[
0.65869140625,
0.216552734375,
0.8896484375,
0.216552734375,
0.8896484375,
0.2266845703125,
0.65869140625,
0.2266845703125
],
[
0.080078125,
0.2291259765625,
0.46728515625,
0.2298583984375,
0.46728515625,
0.24365234375,
0.080078125,
0.242431640625
],
[
0.5361328125,
0.231689453125,
0.87109375,
0.231689453125,
0.87109375,
0.242431640625,
0.5361328125,
0.242431640625
],
[
0.87353515625,
0.236083984375,
0.88623046875,
0.236083984375,
0.88623046875,
0.240478515625,
0.87353515625,
0.240478515625
],
[
0.08416748046875,
0.246826171875,
0.46728515625,
0.246826171875,
0.46728515625,
0.2568359375,
0.08416748046875,
0.2568359375
],
[
0.5361328125,
0.24560546875,
0.89208984375,
0.2442626953125,
0.89208984375,
0.25830078125,
0.5361328125,
0.2587890625
],
[
0.08087158203125,
0.260009765625,
0.465576171875,
0.261474609375,
0.465576171875,
0.272705078125,
0.08087158203125,
0.271484375
],
[
0.5361328125,
0.261474609375,
0.89208984375,
0.261474609375,
0.89208984375,
0.274658203125,
0.5361328125,
0.274658203125
],
[
0.08251953125,
0.274658203125,
0.46728515625,
0.274658203125,
0.46728515625,
0.28857421875,
0.08251953125,
0.28857421875
],
[
0.5361328125,
0.27783203125,
0.64208984375,
0.276611328125,
0.64208984375,
0.28857421875,
0.5361328125,
0.2890625
],
[
0.650390625,
0.274658203125,
0.8896484375,
0.274658203125,
0.8896484375,
0.28857421875,
0.650390625,
0.28857421875
],
[
0.64453125,
0.281005859375,
0.65185546875,
0.281005859375,
0.65185546875,
0.285400390625,
0.64453125,
0.285400390625
],
[
0.08087158203125,
0.290283203125,
0.468017578125,
0.290283203125,
0.468017578125,
0.3037109375,
0.08087158203125,
0.3037109375
],
[
0.5361328125,
0.290283203125,
0.8935546875,
0.290283203125,
0.8935546875,
0.3037109375,
0.5361328125,
0.3037109375
],
[
0.08087158203125,
0.304931640625,
0.1953125,
0.304931640625,
0.1953125,
0.318115234375,
0.08087158203125,
0.318115234375
],
[
0.53515625,
0.30615234375,
0.62353515625,
0.30615234375,
0.62353515625,
0.32080078125,
0.53515625,
0.32080078125
],
[
0.101318359375,
0.3212890625,
0.468017578125,
0.3212890625,
0.468017578125,
0.335205078125,
0.101318359375,
0.335205078125
],
[
0.08087158203125,
0.3359375,
0.468017578125,
0.3359375,
0.468017578125,
0.349853515625,
0.08087158203125,
0.349853515625
],
[
0.50244140625,
0.333984375,
0.650390625,
0.333984375,
0.650390625,
0.347900390625,
0.50244140625,
0.347900390625
],
[
0.08251953125,
0.3515625,
0.455810546875,
0.3515625,
0.455810546875,
0.365478515625,
0.08251953125,
0.365478515625
],
[
0.52392578125,
0.359130859375,
0.890625,
0.359130859375,
0.890625,
0.37255859375,
0.52392578125,
0.37255859375
],
[
0.08251953125,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.380615234375,
0.08251953125,
0.380615234375
],
[
0.51318359375,
0.375732421875,
0.88818359375,
0.375732421875,
0.88818359375,
0.386474609375,
0.51318359375,
0.386474609375
],
[
0.08251953125,
0.38330078125,
0.468017578125,
0.38330078125,
0.468017578125,
0.396484375,
0.08251953125,
0.396484375
],
[
0.50341796875,
0.389404296875,
0.89208984375,
0.389404296875,
0.89208984375,
0.40283203125,
0.50341796875,
0.40283203125
],
[
0.08251953125,
0.396484375,
0.46728515625,
0.396484375,
0.46728515625,
0.410400390625,
0.08251953125,
0.410400390625
],
[
0.50341796875,
0.404052734375,
0.89208984375,
0.404052734375,
0.89208984375,
0.417236328125,
0.50341796875,
0.417236328125
],
[
0.08416748046875,
0.41357421875,
0.46728515625,
0.41357421875,
0.46728515625,
0.423583984375,
0.08416748046875,
0.423583984375
],
[
0.50634765625,
0.42041015625,
0.8896484375,
0.42041015625,
0.8896484375,
0.43115234375,
0.50634765625,
0.43115234375
],
[
0.08087158203125,
0.4267578125,
0.468017578125,
0.4267578125,
0.468017578125,
0.43994140625,
0.08087158203125,
0.43994140625
],
[
0.50244140625,
0.43310546875,
0.890625,
0.434326171875,
0.890625,
0.448974609375,
0.50244140625,
0.447509765625
],
[
0.08251953125,
0.442626953125,
0.468017578125,
0.442626953125,
0.468017578125,
0.455810546875,
0.08251953125,
0.455810546875
],
[
0.50244140625,
0.449462890625,
0.890625,
0.448974609375,
0.890625,
0.46337890625,
0.50244140625,
0.464111328125
],
[
0.08251953125,
0.458251953125,
0.46728515625,
0.458251953125,
0.46728515625,
0.4716796875,
0.08251953125,
0.4716796875
],
[
0.5048828125,
0.466552734375,
0.88818359375,
0.466552734375,
0.88818359375,
0.4765625,
0.5048828125,
0.4765625
],
[
0.08087158203125,
0.472900390625,
0.468017578125,
0.472900390625,
0.468017578125,
0.486083984375,
0.08087158203125,
0.486083984375
],
[
0.5048828125,
0.48095703125,
0.88818359375,
0.48095703125,
0.88818359375,
0.4912109375,
0.5048828125,
0.4912109375
],
[
0.08087158203125,
0.4873046875,
0.468017578125,
0.4873046875,
0.468017578125,
0.50048828125,
0.08087158203125,
0.50048828125
],
[
0.08087158203125,
0.50244140625,
0.46728515625,
0.50244140625,
0.46728515625,
0.5166015625,
0.08087158203125,
0.5166015625
],
[
0.50634765625,
0.496826171875,
0.8896484375,
0.496826171875,
0.8896484375,
0.5068359375,
0.50634765625,
0.5068359375
],
[
0.5048828125,
0.51025390625,
0.890625,
0.51025390625,
0.890625,
0.5234375,
0.5048828125,
0.5234375
],
[
0.080078125,
0.517578125,
0.468017578125,
0.5166015625,
0.468017578125,
0.53076171875,
0.080078125,
0.5322265625
],
[
0.5048828125,
0.5244140625,
0.890625,
0.5244140625,
0.890625,
0.53857421875,
0.5048828125,
0.53857421875
],
[
0.08251953125,
0.5322265625,
0.465576171875,
0.5322265625,
0.465576171875,
0.54541015625,
0.08251953125,
0.54541015625
],
[
0.50341796875,
0.5390625,
0.8896484375,
0.54052734375,
0.8896484375,
0.5537109375,
0.50341796875,
0.55322265625
],
[
0.08087158203125,
0.548828125,
0.468017578125,
0.548828125,
0.468017578125,
0.5625,
0.08087158203125,
0.5625
],
[
0.50341796875,
0.5537109375,
0.890625,
0.5546875,
0.890625,
0.5693359375,
0.50341796875,
0.568359375
],
[
0.08087158203125,
0.56298828125,
0.1968994140625,
0.56298828125,
0.1968994140625,
0.5771484375,
0.08087158203125,
0.5771484375
],
[
0.5048828125,
0.57080078125,
0.890625,
0.57080078125,
0.890625,
0.583984375,
0.5048828125,
0.583984375
],
[
0.101318359375,
0.5791015625,
0.46728515625,
0.5791015625,
0.46728515625,
0.59228515625,
0.101318359375,
0.59228515625
],
[
0.50244140625,
0.58544921875,
0.88818359375,
0.5859375,
0.88818359375,
0.599609375,
0.50244140625,
0.5986328125
],
[
0.08087158203125,
0.5947265625,
0.468017578125,
0.5947265625,
0.468017578125,
0.60791015625,
0.08087158203125,
0.60791015625
],
[
0.5048828125,
0.60302734375,
0.87158203125,
0.60302734375,
0.87158203125,
0.61279296875,
0.5048828125,
0.61279296875
],
[
0.08251953125,
0.60986328125,
0.46728515625,
0.60986328125,
0.46728515625,
0.62353515625,
0.08251953125,
0.62353515625
],
[
0.52294921875,
0.6142578125,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.52294921875,
0.62744140625
],
[
0.08087158203125,
0.625,
0.325927734375,
0.625,
0.325927734375,
0.63818359375,
0.08087158203125,
0.63818359375
],
[
0.5048828125,
0.630859375,
0.87841796875,
0.630859375,
0.87841796875,
0.64453125,
0.5048828125,
0.64453125
],
[
0.87353515625,
0.6337890625,
0.88818359375,
0.6337890625,
0.88818359375,
0.642578125,
0.87353515625,
0.642578125
],
[
0.10296630859375,
0.642578125,
0.465576171875,
0.642578125,
0.465576171875,
0.65283203125,
0.10296630859375,
0.65283203125
],
[
0.5048828125,
0.64599609375,
0.890625,
0.64599609375,
0.890625,
0.6591796875,
0.5048828125,
0.6591796875
],
[
0.0849609375,
0.65380859375,
0.468017578125,
0.65478515625,
0.468017578125,
0.66943359375,
0.0849609375,
0.66845703125
],
[
0.5048828125,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.5048828125,
0.673828125
],
[
0.08087158203125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.08087158203125,
0.6845703125
],
[
0.5048828125,
0.67529296875,
0.88818359375,
0.67529296875,
0.88818359375,
0.689453125,
0.5048828125,
0.689453125
],
[
0.08087158203125,
0.68603515625,
0.468017578125,
0.68603515625,
0.468017578125,
0.69970703125,
0.08087158203125,
0.69970703125
],
[
0.5048828125,
0.6923828125,
0.8896484375,
0.6923828125,
0.8896484375,
0.70263671875,
0.5048828125,
0.70263671875
],
[
0.08251953125,
0.7021484375,
0.398681640625,
0.7021484375,
0.398681640625,
0.71533203125,
0.08251953125,
0.71533203125
],
[
0.5048828125,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.5048828125,
0.7197265625
],
[
0.50341796875,
0.7197265625,
0.890625,
0.7216796875,
0.890625,
0.73486328125,
0.50341796875,
0.73291015625
],
[
0.0931396484375,
0.72900390625,
0.46728515625,
0.72900390625,
0.46728515625,
0.74365234375,
0.0931396484375,
0.74365234375
],
[
0.50634765625,
0.73779296875,
0.88623046875,
0.73779296875,
0.88623046875,
0.74853515625,
0.50634765625,
0.74853515625
],
[
0.11273193359375,
0.744140625,
0.4697265625,
0.744140625,
0.4697265625,
0.75830078125,
0.11273193359375,
0.75830078125
],
[
0.50341796875,
0.751953125,
0.8896484375,
0.751953125,
0.8896484375,
0.76513671875,
0.50341796875,
0.76513671875
],
[
0.11273193359375,
0.76025390625,
0.46728515625,
0.76025390625,
0.46728515625,
0.7734375,
0.11273193359375,
0.7734375
],
[
0.5048828125,
0.767578125,
0.890625,
0.767578125,
0.890625,
0.78076171875,
0.5048828125,
0.78076171875
],
[
0.11273193359375,
0.7744140625,
0.253173828125,
0.77587890625,
0.253173828125,
0.787109375,
0.11273193359375,
0.7861328125
],
[
0.50634765625,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.50634765625,
0.79345703125
],
[
0.09478759765625,
0.79345703125,
0.46728515625,
0.79345703125,
0.46728515625,
0.806640625,
0.09478759765625,
0.806640625
],
[
0.5048828125,
0.798828125,
0.84326171875,
0.798828125,
0.84326171875,
0.8095703125,
0.5048828125,
0.8095703125
],
[
0.11273193359375,
0.80810546875,
0.465576171875,
0.80810546875,
0.465576171875,
0.8212890625,
0.11273193359375,
0.8212890625
],
[
0.52392578125,
0.8125,
0.8896484375,
0.8125,
0.8896484375,
0.82568359375,
0.52392578125,
0.82568359375
],
[
0.11358642578125,
0.82373046875,
0.465576171875,
0.82373046875,
0.465576171875,
0.8369140625,
0.11358642578125,
0.8369140625
],
[
0.5048828125,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.11358642578125,
0.8388671875,
0.46728515625,
0.8388671875,
0.46728515625,
0.85302734375,
0.11358642578125,
0.85302734375
],
[
0.5048828125,
0.84228515625,
0.8798828125,
0.84228515625,
0.8798828125,
0.85595703125,
0.5048828125,
0.85595703125
],
[
0.111083984375,
0.85302734375,
0.330078125,
0.85400390625,
0.330078125,
0.86865234375,
0.111083984375,
0.8671875
],
[
0.5048828125,
0.85693359375,
0.88818359375,
0.85693359375,
0.88818359375,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.0931396484375,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.0931396484375,
0.88623046875
],
[
0.5048828125,
0.87255859375,
0.890625,
0.87255859375,
0.890625,
0.88623046875,
0.5048828125,
0.88623046875
],
[
0.111083984375,
0.88818359375,
0.468017578125,
0.88818359375,
0.468017578125,
0.9013671875,
0.111083984375,
0.9013671875
],
[
0.5048828125,
0.8896484375,
0.88818359375,
0.8896484375,
0.88818359375,
0.8994140625,
0.5048828125,
0.8994140625
],
[
0.48046875,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.93798828125,
0.48046875,
0.93798828125
]
] |
[
"别，我们使用前景区域的全局平均池化来过滤掉不相",
"in the k -shot setting, which outperforms non-",
"关的信息。然后将全局特征与查询分支中的每个位置",
"learnable fusion methods of 1-shot results.",
"进行比较，这可以看作是度量学习方法的一种密集形",
". We demonstrate that given support set with weak",
"式",
"annotations, i.e., bounding boxes, our model can",
"在few-shot设置下，网络应该能够处理在训练过程",
"stillachieve comparableperformancetotheresult",
"中从未见过的新类。因此，我们的目标是从CNN中挖",
"with expensive pixel-level annotated support set,",
"掘可转移的表示形式以进行比较。正如在特征可视化",
"whichfurtherreducesthelabeling effortsofnew",
"文献E[39，38]中观察到的那样，较低层中的特征与低",
"classesfor few-shot segmentation significantly",
"级提示，e.g.，边缘和颜色有关，而较高层中的特征与",
".Experimentson",
"thePASCALVOC2012",
"对象级概念有关，例如类别。我们专注于可能构成由看",
"dataset showthatourmethodachieves",
"",
"不见的类共享的对象部分的中间层特征。例如，如果",
"mean Intersection-over-Union score of 55.4 %",
"CNN在类车上训练模型时学习了与车轮相关的特征",
"for 1-shot segmentation and 57.1 % for 5-shot",
"则该特征也可能对新车辆类别e.g.，卡车和公共汽车",
"segmentation,",
"which significantly outperform",
"",
"的特征比较有用。我们在CNN中提取多个级别的表示",
"state-of-the-art results by 14.6 % and 13.2 % ,",
"以进行密集比较",
"respectively.",
"由于同一类别中的外观存在差异，因此来自同一类",
"的对象可能仅共享一些相似的功能。密集的特征比较",
"2. Related Work",
"不足以指导整个对象区域的分割。然而，这给出了",
"语义分割。语义分割是将图像中的每个像素分类为",
"个重要的线索，即物体在哪里。在半自动分割文献中，",
"组预定义类别的任务：16，2.20，15，17。最先进的",
"为类不可知分割，e.g.，带有单击或涂鸦注释的交互式",
"方法基于完全卷积网络（FCN），它通常采用预先训练",
"分割E[36，14]以及带有边界框或极端点先验的实例分",
"的卷积神经网络（CNN）作为骨干架构进行分类。为",
"割E10，21给出了弱注释。在训练过程中学习可转移",
"了适应密集预测的任务，完全连接的层被一个卷积层",
"的知识来定位对象区域。受半自动分割任务的启发，我",
"代替，该卷积层预测每个像素的标签。为了捕获抽象的",
"们希望根据先验的密集比较结果逐步将对象与背景区",
"特征表示，CNN采用连续的池化操作或卷积步幅来降",
"分开。我们提出了一个选代优化模块（IOM)，它学习",
"低特征图的空间分辨率。然而，这与密集的预测任务冲",
"迭代地优化预测结果。细化以循环形式执行，将密集比",
"突，其中输出应该是高分辨率的。为了平衡网络的输出",
"较结果和预测掩码发送到IOM进行优化，并将输出循",
"环发送到下一个IOM。经过几次细化迭代后，我们的",
"分辨率和感受野，在密集的预测任务中经常使用扩张",
"卷积[2。扩张去除了最后几层中的下采样操作，并在",
"密集比较模块能够生成细粒度的分割图。在每个IOM",
"卷积滤波器中插人孔以扩大感受野。在我们的模型中，",
"内部，我们采用剩余连接，以便在最后一个迭代步骤中",
"我们还采用扩张卷积来保持空间分辨率。在完全监督分",
"有效地合并预测的掩码。Fig。1显示了我们的一次性",
"割中，训练FCN模型需要大量昂贵的像素级注释图像，",
"分割网络的概述",
"并且一旦训练了模型，就无法对新类别进行分割。Our",
"以前的k-shot分割方法基于单镜头模型。他们使用",
"model, on the other hand, can be generalized to any",
"不可学习的融合方法来融合单个单发结果，e.g.，平均",
"newcategorieswithonlyafewannotatedexamples",
"单发预测或中间特征。相反，我们采用注意力机制来有",
"少拍学习。很少的学习旨在学习可转移的知识，这些",
"效地融合来自多个支持示例的信息",
"知识可以推广到具有稀缺标记的训练数据的新类。",
"关",
"为了进一步减少少拍分割的标记工作，我们探索了",
"于少射分类存在许多公式，包括带有记忆的递归神经",
"一种新的测试设置：我们的模型使用边界框注释支持",
"网络E28，23]，学习微调模型E[6,25]，网络参数预",
"集在查询图像中执行分割。我们对PASCALVOC2012",
"测E[1,35]和度量学习E[31,37,11]。基于度量学习",
"数据集和COCO数据集进行了全面的实验，以验证我",
"的方法在少量分类任务中实现了最先进的性能，并且",
"们网络的有效性。本文的主要贡献总结如下。",
"具有快速和以前馈方式进行预测的特点。我们的工作",
"是最相关的关系网络37。关系网络元学习深度距离",
": We develop a novel two-branch dense comparison",
"度量以比较图像并计算相似性得分以进行分类。该网",
"module which effectively exploits multiple levels of",
"络由一个生成图像表示的嵌人模块和一个比较嵌人并",
"feature representations from CNNs to make dense",
"输出相似性得分的关系模块组成。这两个模块都是卷",
"feature comparison.",
"积运算的形式。我们网络中的密集比较模块可以看作",
". We propose an iterative optimization module to",
"是关系网络的扩展，以密集的形式处理分割任务",
"refine predicted results in an iterative manner.",
"少拍语义分割。先前关于少射语义分割的工作采用",
"The ability of iterative refinement can be gener-",
"了两个分支结构。Shabanetal.[29]首次采用语义分",
"alized to unseen classes with few-shot learning for",
"割的少射学习。支持分支直接预测查询分支中最后",
"generating fine-grained maps.",
"层的权重进行分割。在E[24]中，支持分支生成嵌入",
": We adopt an attention mechanism to effectively",
"该嵌人作为附加功能融合到查询分支中。我们的网络",
"fuse information from multiple support examples",
"也遵循两分支设计。但是，与以前的两个分支具有不同",
"2"
] |
{
"arxiv_id": "1903.02351",
"language": "zh",
"page_idx": 1
}
|
||
1903.02351-zh-2
|
1903.02351-zh-2.jpg
|
结构的工作不同，我们网络中的两个分支共享同一骨干网。以前的方法中的模型集中在 1-shot 设置上，当将1-shot 扩展到$k$ -shot 时，它们将 1-shot 方法独立地应用于每个支持示例，并使用不可学习的融合方法在图像级别或特征级别融合单个预测结果。例如，Shaban et al. $\mathbb{E}[29]$建议使用逻辑 OR 运算来融合单个预测掩码，而 Rakelly$\underline{\mathrm{Eet~al.~}}\underline{{\mathrm{E[24]~}}}$则对不同支持示例生成的支持分支中的嵌入进行平均。相反，我们通过注意力机制采用可学习的方法来有效地融合来自多个支持示例的信息
# 3. Task Description
假设我们的模型是在类集为$\mathcal{C}_train$的数据集上训练的，我们的目标是使用训练后的模型对具有新类$\mathcal{C}_{test}$ 的不同数据集进行预测，其中只有几个注释示例可用。直观地，我们训练模型具有以下能力：对于新类$c\not \in \mathcal{C} _{train}$ ,我们的模型能够在仅看到该类的几张图片时从图像中分割该类。一旦模型被训练，参数是固定的，并且在新数据集上测试时不需要优化
我们将训练和测试与情节范式$\mathbb{E}[33]$ 保持一致，以处理少数场景。具体来说，给定一个$k$ -shot 学习任务， 每个情节都是通过以下方式构建的：1)支持(训练)集$\begin{array}{l}{\mathcal{S}=\{(x_s^i,y_s^i(c))\}_{i=1}^k,\text{其中}x_s^i\in\mathbb{R}^{H_i\times W_i\times3}\text{是 RGB}}\\{\text{图像，而}y_s^i(c)\in\mathbb{R}^{H_i\times W_i}\text{是支持图像中}c\text{类的二进制}}\end{array}$ 掩码；2)查询(测试)集$\mathcal{Q}=\{x_q,y_q(c)\}$,其中$x_q$ 是查询类的真值。模型的输入是支持集$S$和查询图像$x_q$ ,输出是查询图像中$c$类的预测掩码$\hat{y}_q(c)$ 。由于一个查询图像$x_q$中可能有多个类，因此当分配不同的标签$c$时，地面实况查询掩码是不同的。Fig。1 显示了当$k=1$。
# 4. Method
我们提出了一个新的框架来解决少射语义分割问题。我们首先在不失一般性的情况下，在单发设置中说明我们的模型。我们的网络由两个模块组成：密集比较模块(DCM)和迭代优化模块(IOM)。DCM 在支持示例和查询示例之间执行密集的功能比较，而 IOM 对预测结果执行迭代细化。Fig。2 (a) 显示了我们框架的概述。为了将我们的网络从一次学习推广到 $k$ 次学习， 我们采用了一种注意力机制来融合来自不同支持示例的信息。此外，我们提出了一种新的测试设置，该设置使用带有边界框注释的支持图像进行少拍分割，随后将对此进行描述
# 4.1. Dense Comparison Module
我们开发了一个两分支密集比较模块，该模块将查询图像中的每个位置与支持示例进行密集比较，如图。$\color{red}2$ (b)。该模块由两个子模块组成：提取表示的特征提取器和执行特征比较的比较模块
特征提取器。特征提取器旨在从 CNN 中获取不同级别的表示以进行特征匹配。我们使用 ResNet-50 [9] 作为特征提取器的骨干。与以前的几次分割工作一样，
骨干模型在 Imagenet$\underline{\mathbb{E}}[4]$上进行了预训练。正如在CNN 特征可视化文献$\overline{\mathbb{E}}[39,38]$ 中观察到的那样，较低层中的特征通常与低级提示，e.g. , 边缘和颜色有关， 而较高层中的特征与对象级别的概念有关，例如对象类别。在少数场景中，我们的模型应该适应任何看不见的类。因此，我们不能假设在训练过程中学习了与看不见的类别相对应的特征。相反，我们专注于可能构成由看不见的类共享的对象部分的中间层特征。ResNet 中的层根据空间分辨率分为 4 个块，这些块自然对应于4 个不同的表示级别。我们选择$block2$和$block3$生成的特征进行特征比较，并在$block3$之后放弃层。我们在$block2$之后的层中使用扩张卷积[2]来保持特征图的空间分辨率。$block2$之后的所有特征图的固定大小为输入图像的 1/8。$block2$和$block3$之后的特征被连接并通过$3\times3$卷积编码为 256 维。我们研究了5.1.3节中用于比较的特征的选择。支持分支和查询分支都使用相同的特征提取器。我们在训练期间保持 ResNet 中的权重固定
密度比较。由于支持图像中可能存在多个对象类别和杂乱的背景，因此我们希望获取仅与目标类别相对应的嵌人以进行比较。在这里，我们使用前景区域的全局平均池化将特征图压缩为特征向量。全局图像特征在分割任务中非常有用$\mathbb{E}[19,40,3]$,可以通过全局平均池化轻松实现。在我们的网络中，我们只对前景区域的特征进行平均，以过滤掉不相关的区域。从支持集获得全局特征向量后，我们将该向量与查询分支生成的特征图中的所有空间位置连接在一起。此操作旨在将查询分支中的所有空间位置与来自支持分支的全局特征向量进行比较。然后，级联的特征图通过另一个具有256个 3$\times3$卷积滤波器的卷积块进行比较
为了有效地实现，我们首先将二进制支持掩码双线性下采样到特征图的相同空间大小，然后将元素乘法应用于特征图。结果，属于背景区域的特征变为零。然后采用全局和池，将得到的向量除以前景区域，得到平均特征向量。我们将向量上采样为相同的查询特征空间大小，并将它们连接起来进行密集比较
# 4.2. Iterative Optimization Module
由于同一类别中存在外观差异，因此密集比较只能匹配对象的一部分，而该部分可能不足以精确地分割图像中的整个对象。我们观察到初始预测是有关对象粗略位置的重要线索。我们提出了一个迭代优化模块来迭代优化预测结果。结构如图所示。2(c)。该模块的输入是由密集比较模块生成的特征图和最后一次迭代的预测掩码。将特征图与预测掩码直接连接作为附加通道会导致特征分布不匹配，因为第一次前向传递没有预测掩码。相反，我们建议将预测掩码合并为残差形式：
$$M_t=x+F(x,y_{t-1}),$$
, (1)
其中$x$是密集比较模块的输出功能；$y_t-1$是最后一个迭代步骤的预测掩码，$M_t$是残差块的输出。函数$F(\cdot)$ 是特征$x$和预测掩码$y_{t-1}$的级联，后面是两个具有256个滤波器的$3\times3$卷积块。然后我们用相同数量的
3
|
[
[
0.08251953125,
0.0947265625,
0.46728515625,
0.0947265625,
0.46728515625,
0.10479736328125,
0.08251953125,
0.10479736328125
],
[
0.50244140625,
0.091552734375,
0.890625,
0.0921630859375,
0.890625,
0.106689453125,
0.50244140625,
0.1060791015625
],
[
0.08251953125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08251953125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08251953125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08251953125,
0.1351318359375
],
[
0.50341796875,
0.12371826171875,
0.88818359375,
0.12371826171875,
0.88818359375,
0.136962890625,
0.50341796875,
0.136962890625
],
[
0.08416748046875,
0.14013671875,
0.46728515625,
0.14013671875,
0.46728515625,
0.15087890625,
0.08416748046875,
0.15087890625
],
[
0.5048828125,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.08251953125,
0.1546630859375,
0.465576171875,
0.1546630859375,
0.465576171875,
0.1654052734375,
0.08251953125,
0.1654052734375
],
[
0.5048828125,
0.154052734375,
0.8896484375,
0.154052734375,
0.8896484375,
0.1673583984375,
0.5048828125,
0.1673583984375
],
[
0.08251953125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08251953125,
0.179931640625
],
[
0.5048828125,
0.169189453125,
0.88818359375,
0.169189453125,
0.88818359375,
0.179931640625,
0.5048828125,
0.179931640625
],
[
0.08416748046875,
0.1849365234375,
0.465576171875,
0.1849365234375,
0.465576171875,
0.1956787109375,
0.08416748046875,
0.1956787109375
],
[
0.50634765625,
0.184326171875,
0.8896484375,
0.184326171875,
0.8896484375,
0.1976318359375,
0.50634765625,
0.1976318359375
],
[
0.0849609375,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.0849609375,
0.210205078125
],
[
0.5048828125,
0.199462890625,
0.88818359375,
0.199462890625,
0.88818359375,
0.210205078125,
0.5048828125,
0.210205078125
],
[
0.08251953125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.08251953125,
0.2266845703125
],
[
0.50341796875,
0.21337890625,
0.8896484375,
0.21337890625,
0.8896484375,
0.2266845703125,
0.50341796875,
0.2266845703125
],
[
0.080078125,
0.2291259765625,
0.131591796875,
0.2291259765625,
0.131591796875,
0.24365234375,
0.080078125,
0.24365234375
],
[
0.5048828125,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.5048828125,
0.242431640625
],
[
0.50634765625,
0.24560546875,
0.88818359375,
0.24560546875,
0.88818359375,
0.255615234375,
0.50634765625,
0.255615234375
],
[
0.07843017578125,
0.25830078125,
0.25244140625,
0.2587890625,
0.25244140625,
0.2734375,
0.07843017578125,
0.272705078125
],
[
0.5048828125,
0.2587890625,
0.8896484375,
0.2587890625,
0.8896484375,
0.272705078125,
0.5048828125,
0.272705078125
],
[
0.50634765625,
0.27587890625,
0.8896484375,
0.27587890625,
0.8896484375,
0.285888671875,
0.50634765625,
0.285888671875
],
[
0.101318359375,
0.2841796875,
0.468017578125,
0.2841796875,
0.468017578125,
0.298583984375,
0.101318359375,
0.298583984375
],
[
0.5048828125,
0.2890625,
0.890625,
0.2890625,
0.890625,
0.302978515625,
0.5048828125,
0.302978515625
],
[
0.08087158203125,
0.2998046875,
0.46728515625,
0.2998046875,
0.46728515625,
0.313232421875,
0.08087158203125,
0.313232421875
],
[
0.50341796875,
0.3037109375,
0.89208984375,
0.3037109375,
0.89208984375,
0.317626953125,
0.50341796875,
0.317626953125
],
[
0.08087158203125,
0.313232421875,
0.465576171875,
0.313232421875,
0.465576171875,
0.326904296875,
0.08087158203125,
0.326904296875
],
[
0.50341796875,
0.318115234375,
0.890625,
0.3193359375,
0.890625,
0.333251953125,
0.50341796875,
0.33203125
],
[
0.08251953125,
0.328857421875,
0.468017578125,
0.328857421875,
0.468017578125,
0.34228515625,
0.08251953125,
0.34228515625
],
[
0.5048828125,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.347900390625,
0.5048828125,
0.347900390625
],
[
0.080078125,
0.3447265625,
0.46728515625,
0.343505859375,
0.46728515625,
0.356689453125,
0.080078125,
0.35791015625
],
[
0.5048828125,
0.35107421875,
0.6005859375,
0.35107421875,
0.6005859375,
0.3623046875,
0.5048828125,
0.3623046875
],
[
0.08087158203125,
0.359130859375,
0.46728515625,
0.359130859375,
0.46728515625,
0.37255859375,
0.08087158203125,
0.37255859375
],
[
0.52392578125,
0.3642578125,
0.890625,
0.3642578125,
0.890625,
0.37744140625,
0.52392578125,
0.37744140625
],
[
0.08087158203125,
0.375,
0.374267578125,
0.375,
0.374267578125,
0.38818359375,
0.08087158203125,
0.38818359375
],
[
0.5048828125,
0.380126953125,
0.890625,
0.380126953125,
0.890625,
0.393310546875,
0.5048828125,
0.393310546875
],
[
0.101318359375,
0.391357421875,
0.465576171875,
0.391357421875,
0.465576171875,
0.401611328125,
0.101318359375,
0.401611328125
],
[
0.5048828125,
0.39453125,
0.890625,
0.39453125,
0.890625,
0.40771484375,
0.5048828125,
0.40771484375
],
[
0.08251953125,
0.40478515625,
0.468017578125,
0.40478515625,
0.468017578125,
0.41845703125,
0.08251953125,
0.41845703125
],
[
0.50341796875,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.50341796875,
0.423583984375
],
[
0.08251953125,
0.419189453125,
0.468017578125,
0.419189453125,
0.468017578125,
0.43310546875,
0.08251953125,
0.43310546875
],
[
0.0767822265625,
0.431884765625,
0.471435546875,
0.4287109375,
0.471435546875,
0.449462890625,
0.0767822265625,
0.45263671875
],
[
0.50634765625,
0.4267578125,
0.8896484375,
0.4267578125,
0.8896484375,
0.4375,
0.50634765625,
0.4375
],
[
0.07598876953125,
0.439453125,
0.47216796875,
0.4375,
0.47216796875,
0.466552734375,
0.07598876953125,
0.468505859375
],
[
0.50341796875,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.50341796875,
0.453857421875
],
[
0.50634765625,
0.45703125,
0.8896484375,
0.45703125,
0.8896484375,
0.46728515625,
0.50634765625,
0.46728515625
],
[
0.080078125,
0.464111328125,
0.465576171875,
0.46533203125,
0.465576171875,
0.479736328125,
0.080078125,
0.479248046875
],
[
0.5048828125,
0.47021484375,
0.890625,
0.47021484375,
0.890625,
0.484130859375,
0.5048828125,
0.484130859375
],
[
0.08251953125,
0.48095703125,
0.46728515625,
0.48095703125,
0.46728515625,
0.494384765625,
0.08251953125,
0.494384765625
],
[
0.5048828125,
0.4873046875,
0.88818359375,
0.4873046875,
0.88818359375,
0.49755859375,
0.5048828125,
0.49755859375
],
[
0.080078125,
0.496826171875,
0.46728515625,
0.49365234375,
0.46728515625,
0.50830078125,
0.080078125,
0.51123046875
],
[
0.50341796875,
0.5,
0.8896484375,
0.50048828125,
0.8896484375,
0.5146484375,
0.50341796875,
0.51318359375
],
[
0.0882568359375,
0.5087890625,
0.46728515625,
0.51025390625,
0.46728515625,
0.5244140625,
0.0882568359375,
0.5234375
],
[
0.5048828125,
0.517578125,
0.88818359375,
0.517578125,
0.88818359375,
0.52783203125,
0.5048828125,
0.52783203125
],
[
0.08087158203125,
0.5263671875,
0.468017578125,
0.5263671875,
0.468017578125,
0.54052734375,
0.08087158203125,
0.54052734375
],
[
0.50244140625,
0.52880859375,
0.80615234375,
0.52880859375,
0.80615234375,
0.54248046875,
0.50244140625,
0.54248046875
],
[
0.08416748046875,
0.54150390625,
0.1707763671875,
0.54150390625,
0.1707763671875,
0.55322265625,
0.08416748046875,
0.55322265625
],
[
0.525390625,
0.54541015625,
0.890625,
0.54541015625,
0.890625,
0.5595703125,
0.525390625,
0.5595703125
],
[
0.50634765625,
0.5625,
0.8896484375,
0.5625,
0.8896484375,
0.57275390625,
0.50634765625,
0.57275390625
],
[
0.08087158203125,
0.57275390625,
0.176513671875,
0.57275390625,
0.176513671875,
0.583984375,
0.08087158203125,
0.583984375
],
[
0.50634765625,
0.57763671875,
0.8896484375,
0.57763671875,
0.8896484375,
0.58837890625,
0.50634765625,
0.58837890625
],
[
0.101318359375,
0.59521484375,
0.468017578125,
0.59521484375,
0.468017578125,
0.609375,
0.101318359375,
0.609375
],
[
0.5048828125,
0.59130859375,
0.8896484375,
0.59130859375,
0.8896484375,
0.60498046875,
0.5048828125,
0.60498046875
],
[
0.08087158203125,
0.611328125,
0.468017578125,
0.611328125,
0.468017578125,
0.625,
0.08087158203125,
0.625
],
[
0.50341796875,
0.60595703125,
0.890625,
0.60595703125,
0.890625,
0.619140625,
0.50341796875,
0.619140625
],
[
0.50341796875,
0.62060546875,
0.79833984375,
0.6220703125,
0.79833984375,
0.63623046875,
0.50341796875,
0.63525390625
],
[
0.08251953125,
0.62744140625,
0.46728515625,
0.62744140625,
0.46728515625,
0.63818359375,
0.08251953125,
0.63818359375
],
[
0.08087158203125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.08087158203125,
0.65478515625
],
[
0.5048828125,
0.64599609375,
0.77685546875,
0.64599609375,
0.77685546875,
0.6591796875,
0.5048828125,
0.6591796875
],
[
0.08087158203125,
0.6572265625,
0.46728515625,
0.6572265625,
0.46728515625,
0.67041015625,
0.08087158203125,
0.67041015625
],
[
0.08251953125,
0.673828125,
0.46728515625,
0.673828125,
0.46728515625,
0.6845703125,
0.08251953125,
0.6845703125
],
[
0.525390625,
0.67041015625,
0.8896484375,
0.67041015625,
0.8896484375,
0.68115234375,
0.525390625,
0.68115234375
],
[
0.08251953125,
0.68603515625,
0.46728515625,
0.68603515625,
0.46728515625,
0.69970703125,
0.08251953125,
0.69970703125
],
[
0.5048828125,
0.6845703125,
0.890625,
0.6845703125,
0.890625,
0.69775390625,
0.5048828125,
0.69775390625
],
[
0.08416748046875,
0.7041015625,
0.465576171875,
0.7041015625,
0.465576171875,
0.7138671875,
0.08416748046875,
0.7138671875
],
[
0.50634765625,
0.70068359375,
0.88818359375,
0.70068359375,
0.88818359375,
0.7109375,
0.50634765625,
0.7109375
],
[
0.08251953125,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.73095703125,
0.08251953125,
0.73095703125
],
[
0.50341796875,
0.7138671875,
0.890625,
0.7138671875,
0.890625,
0.72802734375,
0.50341796875,
0.72802734375
],
[
0.08251953125,
0.73193359375,
0.468017578125,
0.73193359375,
0.468017578125,
0.74560546875,
0.08251953125,
0.74560546875
],
[
0.50341796875,
0.72900390625,
0.890625,
0.72900390625,
0.890625,
0.7421875,
0.50341796875,
0.7421875
],
[
0.08087158203125,
0.74755859375,
0.1968994140625,
0.74755859375,
0.1968994140625,
0.76123046875,
0.08087158203125,
0.76123046875
],
[
0.5048828125,
0.744140625,
0.8896484375,
0.744140625,
0.8896484375,
0.75830078125,
0.5048828125,
0.75830078125
],
[
0.50732421875,
0.76123046875,
0.88818359375,
0.76123046875,
0.88818359375,
0.771484375,
0.50732421875,
0.771484375
],
[
0.08087158203125,
0.7734375,
0.326904296875,
0.7734375,
0.326904296875,
0.787109375,
0.08087158203125,
0.787109375
],
[
0.50634765625,
0.77587890625,
0.8896484375,
0.77587890625,
0.8896484375,
0.7861328125,
0.50634765625,
0.7861328125
],
[
0.1004638671875,
0.79541015625,
0.468017578125,
0.79443359375,
0.468017578125,
0.8095703125,
0.1004638671875,
0.81005859375
],
[
0.5048828125,
0.79150390625,
0.8896484375,
0.79150390625,
0.8896484375,
0.8017578125,
0.5048828125,
0.8017578125
],
[
0.08087158203125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.5048828125,
0.80615234375,
0.54150390625,
0.80615234375,
0.54150390625,
0.8173828125,
0.5048828125,
0.8173828125
],
[
0.62109375,
0.8173828125,
0.77294921875,
0.8193359375,
0.77294921875,
0.833984375,
0.62109375,
0.83203125
],
[
0.86767578125,
0.82080078125,
0.89208984375,
0.82080078125,
0.89208984375,
0.83349609375,
0.86767578125,
0.83349609375
],
[
0.08087158203125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.08251953125,
0.84326171875,
0.3095703125,
0.84326171875,
0.3095703125,
0.85400390625,
0.08251953125,
0.85400390625
],
[
0.50341796875,
0.84033203125,
0.890625,
0.84130859375,
0.890625,
0.85595703125,
0.50341796875,
0.85498046875
],
[
0.098876953125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.098876953125,
0.87060546875
],
[
0.5048828125,
0.85693359375,
0.8896484375,
0.85693359375,
0.8896484375,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50341796875,
0.87255859375,
0.8896484375,
0.87255859375,
0.8896484375,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.08251953125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.08251953125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.4794921875,
0.92529296875,
0.491943359375,
0.92529296875,
0.491943359375,
0.9375,
0.4794921875,
0.9375
]
] |
[
"结构的工作不同，我们网络中的两个分支共享同一骨",
"骨干模型在ImagenetE[4]上进行了预训练。正如在",
"干网。以前的方法中的模型集中在1-shot设置上，当将",
"CNN特征可视化文献E[39，38]中观察到的那样，较低",
"1-shot扩展到k-shot时，它们将1-shot方法独立地应",
"层中的特征通常与低级提示，e.g.，边缘和颜色有关",
"用于每个支持示例，并使用不可学习的融合方法在图",
"而较高层中的特征与对象级别的概念有关，例如对象",
"像级别或特征级别融合单个预测结果。例如，Shaban",
"类别。在少数场景中，我们的模型应该适应任何看不见",
"etal.E29建议使用逻辑OR运算来融合单个预测掩",
"的类。因此，我们不能假设在训练过程中学习了与看不",
"码，而RakellyEletal.E24则对不同支持示例生成的",
"见的类别相对应的特征。相反，我们专注于可能构成由",
"支持分支中的嵌入进行平均。相反，我们通过注意力机",
"看不见的类共享的对象部分的中间层特征。ResNet中",
"制采用可学习的方法来有效地融合来自多个支持示例",
"的层根据空间分辨率分为4个块，这些块自然对应于",
"的信息",
"4个不同的表示级别。我们选择block2和block3生成",
"的特征进行特征比较，并在6lock3之后放弃层。我们",
"3.Task Description",
"在block2之后的层中使用扩张卷积[2]来保持特征图",
"的空间分辨率。block2之后的所有特征图的固定大小",
"假设我们的模型是在类集为Ctrain的数据集上训练",
"为输入图像的1/8。block2和block3之后的特征被连",
"的，我们的目标是使用训练后的模型对具有新类Ctest",
"接并通过3×3卷积编码为256维。我们研究了5.1.3",
"的不同数据集进行预测，其中只有几个注释示例可",
"节中用于比较的特征的选择。支持分支和查询分支都",
"用。直观地，我们训练模型具有以下能力：对于新类",
"使用相同的特征提取器。我们在训练期间保持ResNet",
"CCtrain，我们的模型能够在仅看到该类的几张图片",
"中的权重固定",
"时从图像中分割该类。一旦模型被训练，参数是固定",
"密度比较。由于支持图像中可能存在多个对象类别",
"的，并且在新数据集上测试时不需要优化",
"和杂乱的背景，因此我们希望获取仅与目标类别相对",
"我们将训练和测试与情节范式F33保持一致，以",
"应的嵌入以进行比较。在这里，我们使用前景区域的全",
"处理少数场景。具体来说，给定一个k-shot学习任务，",
"局平均池化将特征图压缩为特征向量。全局图像特征",
"每个情节都是通过以下方式构建的：1）支持（训练）集",
"S=(c,y(C)，其中E RH.xWx3 是 RGB",
"在分割任务中非常有用E19，40，3，可以通过全局平",
"图像，需hw是支持图像中。类的二进制",
"均池化轻松实现。在我们的网络中，我们只对前景区域",
"的特征进行平均，以过滤掉不相关的区域。从支持集获",
"掩码；2）查询（测试）集Q=（cg,9g(c))，其中g",
"得全局特征向量后，我们将该向量与查询分支生成的",
"是查询类的真值。模型的输人是支持集S和查询图像",
"特征图中的所有空间位置连接在一起。此操作旨在将",
"Cq，输出是查询图像中c类的预测掩码9a(c）。由于",
"查询分支中的所有空间位置与来自支持分支的全局特",
"一个查询图像。中可能有多个类，因此当分配不同的",
"征向量进行比较。然后，级联的特征图通过另一个具有",
"标签c时，地面实况查询掩码是不同的。Fig。1显示",
"256个3×3卷积滤波器的卷积块进行比较",
"了当k=1。",
"为了有效地实现，我们首先将二进制支持掩码双线",
"性下采样到特征图的相同空间大小，然后将元素乘法",
"4.Method",
"应用于特征图。结果，属于背景区域的特征变为零。然",
"我们提出了一个新的框架来解决少射语义分割问题。",
"后采用全局和池，将得到的向量除以前景区域，得到平",
"我们首先在不失一般性的情况下，在单发设置中说明",
"均特征向量。我们将向量上采样为相同的查询特征空",
"间大小，并将它们连接起来进行密集比较",
"我们的模型。我们的网络由两个模块组成：密集比较模",
"块（DCM）和迭代优化模块（IOM）。DCM在支持示",
"4.2. Iterative Optimization Module",
"例和查询示例之间执行密集的功能比较，而IOM对预",
"测结果执行送代细化。Fig。2（a）显示了我们框架的",
"由于同一类别中存在外观差异，因此密集比较只能",
"概述。为了将我们的网络从一次学习推广到K次学习，",
"匹配对象的一部分，而该部分可能不足以精确地分割",
"我们采用了一种注意力机制来融合来自不同支持示例",
"图像中的整个对象。我们观察到初始预测是有关对象",
"的信息。此外，我们提出了一种新的测试设置，该设置",
"粗略位置的重要线索。我们提出了一个送代优化模块",
"使用带有边界框注释的支持图像进行少拍分割，随后",
"来迭代优化预测结果。结构如图所示。2(c)。该模块",
"将对此进行描述",
"的输人是由密集比较模块生成的特征图和最后一次迭",
"代的预测掩码。将特征图与预测掩码直接连接作为附",
"4.1. Dense Comparison Module",
"加通道会导致特征分布不匹配，因为第一次前向传递",
"我们开发了一个两分支密集比较模块，该模块将查",
"没有预测掩码。相反，我们建议将预测掩码合并为残差",
"询图像中的每个位置与支持示例进行密集比较，如图。",
"形式",
"Mt=+F(r,yt-1),",
"(1)",
"2（b)。该模块由两个子模块组成：提取表示的特征提",
"取器和执行特征比较的比较模块",
"其中是密集比较模块的输出功能；yt-1是最后一个",
"特征提取器。特征提取器旨在从CNN中获取不同",
"迭代步骤的预测掩码，Mt是残差块的输出。函数F()",
"级别的表示以进行特征匹配。我们使用ResNet-50[9]",
"是特征和预测掩码yt-1的级联，后面是两个具有",
"作为特征提取器的骨干。与以前的几次分割工作一样，",
"256个滤波器的3×3卷积块。然后我们用相同数量的",
"3"
] |
{
"arxiv_id": "1903.02351",
"language": "zh",
"page_idx": 2
}
|
||
1903.02351-zh-3
|
1903.02351-zh-3.jpg
|
<FigureHere>
Figure 2- CANet for 1-shot semantic segmentation. (a) Overview of our network structure. (b) Dense Comparison Module. (c)
Iterative Optimization Module.
版，其中包括背景蒙版和前景蒙版。我们使用 softmax 函数对每个位置的分数进行归一化，从而输出前景和背景的置信度图。然后将置信度图提供给下一个 IOM 进行优化。我们的最终结果是通过将置信度图双线性上采样到查询图像的相同空间大小并根据置信度图对每个位置进行分类来实现的。在训练时，为了避免迭代优化模块过度拟合预测掩码，我们或者使用最后一个时期的预测掩码和空掩码作为 IOM 的输人。预测掩码$y_{t-1}$被重置为空掩码，概率为$p_r$ 。这可以看作是整个掩码的丢失，是标准丢失的扩展$\overline{\mathbb{E}}[32]$ 。与分割文献$\overline{\mathbb{E}}$ [14,34,22]中以前的迭代细化方法相比，我们的方法将细化方案集成到具有残差连接的模型中，以便整个模型可以以前馈方式运行并进行端到端训练
<FigureHere>
4.3. Attention Mechanism for k-shot Segmentation
Figure 3-Attention mechanism for $k$ -shot semantic segmentation. We use the softmax function to normalize the outputs of the attention module from different support examples.
为了有效地合并$k$ -shot 设置中的信息，我们使用注意机制来融合由不同支持示例生成的比较结果。具体来说，我们在 DCM 中添加了一个与密集比较卷积并行的注意模块(见图。3)。注意力分支由两个卷积块组成。第一个有 256 个$3\times3$过滤器，其次是$3\times3$最大池化。第二个有一个$3\times3$卷积，后跟一个全局平均池。注意力分支的结果作为权重$\lambda$ 。然后，来自所有支持示例的权重由 softmax 函数归一化：
卷积滤波器添加两个香草残差块。最重要的是，我们使用 DeeplabV3$\mathbb{E}[3]$中提出的 Atrous 空间金字塔池模块 (ASPP) 来捕获多尺度信息。该模块由四个并行分支组成，其中包括三个$3\times3$卷积，分别具有 6、12 和18的速率，以及一个$1\times1$卷积。$1\times1$卷积是在通过全局平均池化实现的图像级特征上进行的。然后，将生成的向量进行双线性上采样到原始空间大小。来自 4个分支的输出特征由另一个$1\times1$卷积与 256 个滤波器连接并融合最后，我们使用$1\times1$卷积生成最终的蒙
(2)
$$\hat{\lambda}_i=\frac{e^{\lambda_i}}{\sum_{j=1}^ke^{\lambda_j}}.$$
最终输出是由不同支持样本生成的特征的加权和
4
|
[
[
0.2115478515625,
0.09661865234375,
0.36279296875,
0.09661865234375,
0.36279296875,
0.106689453125,
0.2115478515625,
0.106689453125
],
[
0.29736328125,
0.11553955078125,
0.334228515625,
0.11553955078125,
0.334228515625,
0.1256103515625,
0.29736328125,
0.1256103515625
],
[
0.32177734375,
0.1275634765625,
0.358642578125,
0.1275634765625,
0.358642578125,
0.1383056640625,
0.32177734375,
0.1383056640625
],
[
0.3896484375,
0.1307373046875,
0.406982421875,
0.1307373046875,
0.406982421875,
0.1351318359375,
0.3896484375,
0.1351318359375
],
[
0.16259765625,
0.136962890625,
0.1829833984375,
0.136962890625,
0.1829833984375,
0.15087890625,
0.16259765625,
0.15087890625
],
[
0.2279052734375,
0.15283203125,
0.279296875,
0.15283203125,
0.279296875,
0.1572265625,
0.2279052734375,
0.1572265625
],
[
0.346435546875,
0.15283203125,
0.428955078125,
0.154052734375,
0.428955078125,
0.1654052734375,
0.346435546875,
0.1641845703125
],
[
0.298095703125,
0.169189453125,
0.33251953125,
0.169189453125,
0.33251953125,
0.1767578125,
0.298095703125,
0.1767578125
],
[
0.3203125,
0.18115234375,
0.361083984375,
0.18115234375,
0.361083984375,
0.1912841796875,
0.3203125,
0.1912841796875
],
[
0.38232421875,
0.18310546875,
0.4150390625,
0.18310546875,
0.4150390625,
0.1900634765625,
0.38232421875,
0.1900634765625
],
[
0.226318359375,
0.1962890625,
0.27685546875,
0.1962890625,
0.27685546875,
0.203857421875,
0.226318359375,
0.203857421875
],
[
0.1993408203125,
0.2442626953125,
0.268798828125,
0.2442626953125,
0.268798828125,
0.255615234375,
0.1993408203125,
0.255615234375
],
[
0.1641845703125,
0.255615234375,
0.1871337890625,
0.255615234375,
0.1871337890625,
0.268310546875,
0.1641845703125,
0.268310546875
],
[
0.205078125,
0.2841796875,
0.261474609375,
0.2841796875,
0.261474609375,
0.29541015625,
0.205078125,
0.29541015625
],
[
0.43701171875,
0.300537109375,
0.468017578125,
0.300537109375,
0.468017578125,
0.309326171875,
0.43701171875,
0.309326171875
],
[
0.47802734375,
0.3017578125,
0.50732421875,
0.3017578125,
0.50732421875,
0.309326171875,
0.47802734375,
0.309326171875
],
[
0.56201171875,
0.300537109375,
0.59619140625,
0.300537109375,
0.59619140625,
0.31005859375,
0.56201171875,
0.31005859375
],
[
0.1650390625,
0.36865234375,
0.1871337890625,
0.36865234375,
0.1871337890625,
0.38330078125,
0.1650390625,
0.38330078125
],
[
0.62109375,
0.36865234375,
0.6435546875,
0.36865234375,
0.6435546875,
0.380126953125,
0.62109375,
0.380126953125
],
[
0.70751953125,
0.399658203125,
0.71728515625,
0.399658203125,
0.71728515625,
0.404052734375,
0.70751953125,
0.404052734375
],
[
0.7158203125,
0.3984375,
0.740234375,
0.3984375,
0.740234375,
0.406005859375,
0.7158203125,
0.406005859375
],
[
0.2467041015625,
0.4091796875,
0.420654296875,
0.4091796875,
0.420654296875,
0.419189453125,
0.2467041015625,
0.419189453125
],
[
0.50244140625,
0.4091796875,
0.51708984375,
0.4091796875,
0.51708984375,
0.416748046875,
0.50244140625,
0.416748046875
],
[
0.08660888671875,
0.43310546875,
0.8857421875,
0.43310546875,
0.8857421875,
0.4462890625,
0.08660888671875,
0.4462890625
],
[
0.08660888671875,
0.4462890625,
0.27685546875,
0.4462890625,
0.27685546875,
0.45703125,
0.08660888671875,
0.45703125
],
[
0.50341796875,
0.472900390625,
0.89208984375,
0.472900390625,
0.89208984375,
0.486083984375,
0.50341796875,
0.486083984375
],
[
0.0882568359375,
0.4892578125,
0.1094970703125,
0.4892578125,
0.1094970703125,
0.496826171875,
0.0882568359375,
0.496826171875
],
[
0.50244140625,
0.486083984375,
0.890625,
0.4873046875,
0.890625,
0.501953125,
0.50244140625,
0.50048828125
],
[
0.298095703125,
0.494384765625,
0.310546875,
0.494384765625,
0.310546875,
0.50048828125,
0.298095703125,
0.50048828125
],
[
0.5048828125,
0.50244140625,
0.890625,
0.50244140625,
0.890625,
0.5166015625,
0.5048828125,
0.5166015625
],
[
0.5048828125,
0.51806640625,
0.890625,
0.51806640625,
0.890625,
0.5322265625,
0.5048828125,
0.5322265625
],
[
0.50732421875,
0.53271484375,
0.890625,
0.53271484375,
0.890625,
0.546875,
0.50732421875,
0.546875
],
[
0.0882568359375,
0.54052734375,
0.1544189453125,
0.54052734375,
0.1544189453125,
0.54736328125,
0.0882568359375,
0.54736328125
],
[
0.298095703125,
0.54443359375,
0.30810546875,
0.548828125,
0.3056640625,
0.5517578125,
0.295654296875,
0.548828125
],
[
0.50244140625,
0.546875,
0.890625,
0.54736328125,
0.890625,
0.5625,
0.50244140625,
0.56103515625
],
[
0.50341796875,
0.56298828125,
0.8896484375,
0.56298828125,
0.8896484375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.11767578125,
0.57275390625,
0.1282958984375,
0.57275390625,
0.1282958984375,
0.58544921875,
0.11767578125,
0.58544921875
],
[
0.2484130859375,
0.5771484375,
0.2548828125,
0.5771484375,
0.2548828125,
0.58349609375,
0.2484130859375,
0.58349609375
],
[
0.50341796875,
0.5791015625,
0.890625,
0.5791015625,
0.890625,
0.59228515625,
0.50341796875,
0.59228515625
],
[
0.2181396484375,
0.59521484375,
0.259033203125,
0.59521484375,
0.259033203125,
0.60693359375,
0.2181396484375,
0.60693359375
],
[
0.259033203125,
0.5986328125,
0.26953125,
0.5986328125,
0.26953125,
0.60302734375,
0.259033203125,
0.60302734375
],
[
0.50341796875,
0.59326171875,
0.8896484375,
0.59326171875,
0.8896484375,
0.60791015625,
0.50341796875,
0.60791015625
],
[
0.0849609375,
0.60498046875,
0.1544189453125,
0.60498046875,
0.1544189453125,
0.61572265625,
0.0849609375,
0.61572265625
],
[
0.2998046875,
0.61279296875,
0.303955078125,
0.61279296875,
0.303955078125,
0.619140625,
0.2998046875,
0.619140625
],
[
0.50341796875,
0.609375,
0.890625,
0.609375,
0.890625,
0.62255859375,
0.50341796875,
0.62255859375
],
[
0.5048828125,
0.62353515625,
0.890625,
0.62353515625,
0.890625,
0.63720703125,
0.5048828125,
0.63720703125
],
[
0.50341796875,
0.63818359375,
0.890625,
0.63818359375,
0.890625,
0.65283203125,
0.50341796875,
0.65283203125
],
[
0.2099609375,
0.6572265625,
0.2998046875,
0.6572265625,
0.2998046875,
0.6787109375,
0.2099609375,
0.6787109375
],
[
0.50244140625,
0.65283203125,
0.81396484375,
0.65283203125,
0.81396484375,
0.66650390625,
0.50244140625,
0.66650390625
],
[
0.2132568359375,
0.6796875,
0.2900390625,
0.6796875,
0.2900390625,
0.68994140625,
0.2132568359375,
0.68994140625
],
[
0.50341796875,
0.6796875,
0.89208984375,
0.6796875,
0.89208984375,
0.693359375,
0.50341796875,
0.693359375
],
[
0.0849609375,
0.7021484375,
0.4599609375,
0.7021484375,
0.4599609375,
0.71533203125,
0.0849609375,
0.71533203125
],
[
0.52392578125,
0.70263671875,
0.8896484375,
0.70263671875,
0.8896484375,
0.71728515625,
0.52392578125,
0.71728515625
],
[
0.0849609375,
0.71337890625,
0.463134765625,
0.71337890625,
0.463134765625,
0.7265625,
0.0849609375,
0.7265625
],
[
0.5048828125,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.5048828125,
0.73193359375
],
[
0.08660888671875,
0.72607421875,
0.4345703125,
0.72607421875,
0.4345703125,
0.7392578125,
0.08660888671875,
0.7392578125
],
[
0.5048828125,
0.734375,
0.8896484375,
0.734375,
0.8896484375,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.5048828125,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.5048828125,
0.76318359375
],
[
0.08251953125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.08251953125,
0.77978515625
],
[
0.50341796875,
0.7646484375,
0.890625,
0.7646484375,
0.890625,
0.77783203125,
0.50341796875,
0.77783203125
],
[
0.08251953125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.08251953125,
0.79443359375
],
[
0.50341796875,
0.77783203125,
0.8896484375,
0.77880859375,
0.8896484375,
0.79345703125,
0.50341796875,
0.79248046875
],
[
0.080078125,
0.79443359375,
0.46728515625,
0.79541015625,
0.46728515625,
0.81005859375,
0.080078125,
0.8095703125
],
[
0.50341796875,
0.79443359375,
0.8896484375,
0.79443359375,
0.8896484375,
0.8095703125,
0.50341796875,
0.8095703125
],
[
0.08416748046875,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.08416748046875,
0.82470703125
],
[
0.5048828125,
0.8095703125,
0.79150390625,
0.8095703125,
0.79150390625,
0.82275390625,
0.5048828125,
0.82275390625
],
[
0.08087158203125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.08087158203125,
0.8388671875
],
[
0.701171875,
0.833984375,
0.7236328125,
0.833984375,
0.7236328125,
0.8486328125,
0.701171875,
0.8486328125
],
[
0.08251953125,
0.84130859375,
0.465576171875,
0.84130859375,
0.465576171875,
0.85498046875,
0.08251953125,
0.85498046875
],
[
0.64892578125,
0.84130859375,
0.6650390625,
0.85400390625,
0.65283203125,
0.8642578125,
0.6357421875,
0.85107421875
],
[
0.86669921875,
0.8447265625,
0.8935546875,
0.8447265625,
0.8935546875,
0.8623046875,
0.86669921875,
0.8623046875
],
[
0.08087158203125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.08087158203125,
0.869140625
],
[
0.7138671875,
0.85498046875,
0.74658203125,
0.85498046875,
0.74658203125,
0.8671875,
0.7138671875,
0.8671875
],
[
0.08251953125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08251953125,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.84716796875,
0.88720703125,
0.84716796875,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.4794921875,
0.923828125,
0.49267578125,
0.9248046875,
0.491943359375,
0.93603515625,
0.47802734375,
0.93505859375
]
] |
[
"DenseComparison Module",
"ResNet",
"Block-2",
"lock",
"(b)",
"supportimag",
"Weights sharing",
"ResNet",
"Block-2",
"Block-3",
"Queryimage",
"Support set",
"(a)",
"Query set",
"Itero",
"Iter1",
"Itern",
"(c)",
"3",
"imy",
"+ReL",
"IterativeOptimizationModule",
"ASP",
"Figure 2 - CANet for 1-shot semantic segmentation. (a) Overview of our network structure. (b) Dense Comparison Module. (c)",
"IterativeOptimization Module.",
"版，其中包括背景蒙版和前景蒙版。我们使用softmax",
"Suppo",
"函数对每个位置的分数进行归一化，从而输出前景和",
"X",
"背景的置信度图。然后将置信度图提供给下一个IOM",
"进行优化。我们的最终结果是通过将置信度图双线性",
"上采样到查询图像的相同空间大小并根据置信度图对",
"Supportsample2",
"",
"每个位置进行分类来实现的。在训练时，为了避免迭代",
"优化模块过度拟合预测掩码，我们或者使用最后一个",
"..",
":",
"时期的预测掩码和空掩码作为IOM的输人。预测掩码",
"",
"",
"yt-1被重置为空掩码，概率为Pr。这可以看作是整个",
"Support sample k",
"",
"掩码的丢失，是标准丢失的扩展E[32。与分割文献E",
"14，34，22中以前的迭代细化方法相比，我们的方法",
"将细化方案集成到具有残差连接的模型中，以便整个",
"D-k",
"模型可以以前馈方式运行并进行端到端训练",
"Attention Module",
"4.3.AttentionMechanismfork-shotSegmentation",
"Figure 3 - Attention mechanism for k -shot semantic segmen-",
"为了有效地合并k-shot设置中的信息，我们使用注",
"tation. We use the softmax function to normalize the outputs",
"意机制来融合由不同支持示例生成的比较结果。具体",
"of the attention module from different support examples.",
"来说，我们在DCM中添加了一个与密集比较卷积并",
"行的注意模块（见图。3）。注意力分支由两个卷积块",
"卷积滤波器添加两个香草残差块。最重要的是，我们使",
"组成。第一个有256个3×3过滤器，其次是3×3最",
"用DeeplabV3E[3]中提出的Atrous空间金字塔池模",
"大池化。第二个有一个3×3卷积，后跟一个全局平均",
"块（ASPP）来捕获多尺度信息。该模块由四个并行分",
"池。注意力分支的结果作为权重入。然后，来自所有",
"支组成，其中包括三个3×3卷积，分别具有6、12和",
"支持示例的权重由softmax函数归一化：",
"18的速率，以及一个1×1卷积。1×1卷积是在通过",
"ed",
"全局平均池化实现的图像级特征上进行的。然后，将",
"T",
"(2)",
"生成的向量进行双线性上采样到原始空间大小。来自4",
"led,",
"个分支的输出特征由另一个1×1卷积与256个滤波",
"器连接并融合最后，我们使用1×1卷积生成最终的蒙",
"最终输出是由不同支持样本生成的特征的加权和",
"4"
] |
{
"arxiv_id": "1903.02351",
"language": "zh",
"page_idx": 3
}
|
||
2403.00323-zh-0
|
2403.00323-zh-0.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
# 神经符号系统的软化符号接地
Zenan Li$^1$, Yuan Yao$^1$, Taolue Chen$^2$, Jingwei Xu$^1$, Chun Cao$^1$, Xiaoxing Ma$^1,\text{Jian L\"{u}}^1$
$^{1}$State Key Lab of Novel Software Technology and
Department of Computer Science and Technology, Nanjing University, China
$\bar{^2}$Department of $\bar\text{Computer Science, Birkbeck, University of London, UK}$
lizn@ smail. nju. edu. cn, t. chen@ bbk. ac. uk
$\{$ y.yao,jingweix,caochun,xxm,lj $\}$ @nju.edu.cn
## Abstract
Neuro-symbolic learning generally consists of two separated worlds, i.e., neural network training and symbolic constraint solving, whose success hinges on symbol grounding, a fundamental problem in AI. This paper presents a novel, softened symbol grounding process, bridging the gap between the two worlds, and resulting in an effective and efficient neurosymbolic learning framework. Technically, the framework features (1) modeling of symbol solution states as a Boltzmann distribution, which'avoids expensive state searching and facilitates mutually beneficial interactions between network training and symbolic reasoning; (2) a new MCMC tech- $\begin{array}{c}\text{nique leveraging projection and SMT solvers, which efficiently samples from}\\\end{array}$ $\bar{\text{disconnected symbol solution spaces;(3) an annealing mechanism that can}}$ escape from sub-optimal symbol groundings. Experiments with three rep- $\dot{\text{resentative neuro-symbolic learning tasks demonstrate that, owining to its}}$ superior symbol grounding capability, our framework successfully solves problems well beyond the frontier of the existing proposals.
## 1 Introduction
已经提出了神经符号系统来连接神经网络学习和符号约束满足E(Zhou, 2019; Garcez et al. 2019; Marra et al., 2021; Yu et al., 2021; Hitzler, 2022) 。在这些系统中，神经网络组件首先将 raw input 识别为符号，然后将其进一步馈送到符号组件中以产生 fnal output $\underline{\mathrm{E}}($Yi et al., 2018; Li et al., 2020; Liang et al., 2017) 。这种神经符号范式已显示出前所未有的能力，并在许多任务中取得了令人印象深刻的成果，包括视觉问答E(Yi et al., 2018; Vedantam et al., 2019; Amizadeh et al., 2020) , 视觉语言导航$\mathbb{E}($Anderson et al., 2018; Fried et al., 2018) 和数学单词问题解决$\overline{\mathbb{E}}($Hong et al., 2021; Qin et al., 2021) , 仅举几例
如图$\color{red}{\operatorname*{El}}$1 所示，为了最大程度地提高通用性，此类问题通常在弱监督设置中进行$\color{red}{\mathrm{E}}($Garcez et al., 2022) : 神经符号计算的最终输出是在训练过程中作为监督提供的，而不是中间符号的标签。缺乏直接监督标签的网络训练呼吁一种有效和高效的方法来解决符号接地问题我。e。, 建立从原始输人到潜在符号的可行且可推广的映射。请注意，旁路符号接地 (通过， e。g。, 将问题视为有逻辑约束的学习) 是可能的，但不能达到令人满意的性能 (Manhaeve et al., 2018; Xu et al., 2018; Pryor et al., 2022) 。在网络学习中结合符号接地的现有方法严重依赖于良好的初始模型，并且在从头开始时表现不佳 (Dai et al.,2019; Li et al., 2020; Huang et al., 2021)
符号接地的一个关键挑战在于随机和连续的神经学习与确定性和离散的符号推理之间的语义鸿沟。为了弥合这一差距，我们建议 soften 将符号接地。也就是说，我们不是直接搜索确定性的输入符号映射，而是使用退火策略来优化其玻尔兹曼分布，以逐渐收敛到确定性的。直观上，软化的玻尔兹曼分布提供了一个游乐场，在那里输入符号映射的搜索可以由神经网络引导，网络训练可以通过从分布中采样来监督。博弈论确实为这种策略提供了理论支持 (Conitzer, 2016): 软化使学习过程在退火过程中成为一系列混合策略游戏，这鼓励了神经世界和符号世界之间更强的相互作用
剩下的挑战是如何有效地采样可行的输人符号映射。具体而言，可行解在整个符号空间中极为稀疏，不同解之间的连接较差，这使得马尔可夫链蒙特卡罗(MCMC)抽样无法有效
1
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.176513671875,
0.10162353515625,
0.5400390625,
0.1029052734375,
0.5400390625,
0.125,
0.176513671875,
0.12371826171875
],
[
0.191162109375,
0.15283203125,
0.82177734375,
0.15283203125,
0.82177734375,
0.1666259765625,
0.191162109375,
0.1666259765625
],
[
0.32275390625,
0.1787109375,
0.685546875,
0.179931640625,
0.685546875,
0.1944580078125,
0.32275390625,
0.1932373046875
],
[
0.2296142578125,
0.1944580078125,
0.779296875,
0.1944580078125,
0.779296875,
0.2083740234375,
0.2296142578125,
0.2083740234375
],
[
0.2459716796875,
0.20703125,
0.76123046875,
0.2083740234375,
0.76123046875,
0.222900390625,
0.2459716796875,
0.2215576171875
],
[
0.35693359375,
0.222900390625,
0.650390625,
0.222900390625,
0.650390625,
0.236083984375,
0.35693359375,
0.236083984375
],
[
0.34228515625,
0.236083984375,
0.67236328125,
0.236083984375,
0.67236328125,
0.25,
0.34228515625,
0.25
],
[
0.463134765625,
0.285888671875,
0.5361328125,
0.285888671875,
0.5361328125,
0.300537109375,
0.463134765625,
0.300537109375
],
[
0.233642578125,
0.317626953125,
0.7646484375,
0.317626953125,
0.7646484375,
0.330810546875,
0.233642578125,
0.330810546875
],
[
0.233642578125,
0.330078125,
0.7646484375,
0.330078125,
0.7646484375,
0.3447265625,
0.233642578125,
0.3447265625
],
[
0.233642578125,
0.343505859375,
0.765625,
0.343505859375,
0.765625,
0.35791015625,
0.233642578125,
0.35791015625
],
[
0.233642578125,
0.35791015625,
0.76318359375,
0.35791015625,
0.76318359375,
0.37255859375,
0.233642578125,
0.37255859375
],
[
0.233642578125,
0.371826171875,
0.7646484375,
0.371826171875,
0.7646484375,
0.385009765625,
0.233642578125,
0.385009765625
],
[
0.233642578125,
0.385009765625,
0.76318359375,
0.385009765625,
0.76318359375,
0.3984375,
0.233642578125,
0.3984375
],
[
0.2320556640625,
0.397705078125,
0.7646484375,
0.397705078125,
0.7646484375,
0.410888671875,
0.2320556640625,
0.410888671875
],
[
0.233642578125,
0.412353515625,
0.7646484375,
0.412353515625,
0.7646484375,
0.425537109375,
0.233642578125,
0.425537109375
],
[
0.233642578125,
0.425537109375,
0.7646484375,
0.425537109375,
0.7646484375,
0.439453125,
0.233642578125,
0.439453125
],
[
0.2320556640625,
0.439453125,
0.7646484375,
0.438232421875,
0.7646484375,
0.45263671875,
0.2320556640625,
0.453857421875
],
[
0.233642578125,
0.45263671875,
0.7646484375,
0.45263671875,
0.7646484375,
0.466552734375,
0.233642578125,
0.466552734375
],
[
0.2320556640625,
0.466552734375,
0.76123046875,
0.46728515625,
0.76123046875,
0.48095703125,
0.2320556640625,
0.479736328125
],
[
0.233642578125,
0.479736328125,
0.7646484375,
0.479736328125,
0.7646484375,
0.49365234375,
0.233642578125,
0.49365234375
],
[
0.233642578125,
0.49365234375,
0.765625,
0.49365234375,
0.765625,
0.5068359375,
0.233642578125,
0.5068359375
],
[
0.2320556640625,
0.505859375,
0.65869140625,
0.5068359375,
0.65869140625,
0.521484375,
0.2320556640625,
0.52099609375
],
[
0.17236328125,
0.54443359375,
0.31787109375,
0.546875,
0.31787109375,
0.56103515625,
0.17236328125,
0.5595703125
],
[
0.17724609375,
0.57568359375,
0.8203125,
0.57568359375,
0.8203125,
0.5888671875,
0.17724609375,
0.5888671875
],
[
0.1748046875,
0.5888671875,
0.82177734375,
0.5888671875,
0.82177734375,
0.60400390625,
0.1748046875,
0.60400390625
],
[
0.176513671875,
0.60400390625,
0.82421875,
0.60400390625,
0.82421875,
0.6171875,
0.176513671875,
0.6171875
],
[
0.1748046875,
0.6171875,
0.82421875,
0.6171875,
0.82421875,
0.630859375,
0.1748046875,
0.630859375
],
[
0.176513671875,
0.630859375,
0.82177734375,
0.630859375,
0.82177734375,
0.64453125,
0.176513671875,
0.64453125
],
[
0.1748046875,
0.64453125,
0.82177734375,
0.64453125,
0.82177734375,
0.65771484375,
0.1748046875,
0.65771484375
],
[
0.176513671875,
0.65771484375,
0.72607421875,
0.65771484375,
0.72607421875,
0.671875,
0.176513671875,
0.671875
],
[
0.1748046875,
0.6767578125,
0.82421875,
0.67822265625,
0.82421875,
0.6923828125,
0.1748046875,
0.69140625
],
[
0.176513671875,
0.6923828125,
0.82275390625,
0.6923828125,
0.82275390625,
0.70556640625,
0.176513671875,
0.70556640625
],
[
0.176513671875,
0.70556640625,
0.82177734375,
0.70556640625,
0.82177734375,
0.7197265625,
0.176513671875,
0.7197265625
],
[
0.176513671875,
0.72021484375,
0.82177734375,
0.72021484375,
0.82177734375,
0.734375,
0.176513671875,
0.734375
],
[
0.1732177734375,
0.73291015625,
0.82275390625,
0.73193359375,
0.82275390625,
0.7470703125,
0.1732177734375,
0.74755859375
],
[
0.1732177734375,
0.74560546875,
0.82275390625,
0.7470703125,
0.82275390625,
0.76123046875,
0.1732177734375,
0.76025390625
],
[
0.176513671875,
0.76025390625,
0.82421875,
0.76025390625,
0.82421875,
0.7734375,
0.176513671875,
0.7734375
],
[
0.1748046875,
0.7734375,
0.314453125,
0.7734375,
0.314453125,
0.787109375,
0.1748046875,
0.787109375
],
[
0.176513671875,
0.79443359375,
0.82421875,
0.79443359375,
0.82421875,
0.80810546875,
0.176513671875,
0.80810546875
],
[
0.17724609375,
0.80810546875,
0.82275390625,
0.80810546875,
0.82275390625,
0.8212890625,
0.17724609375,
0.8212890625
],
[
0.176513671875,
0.82275390625,
0.82275390625,
0.82275390625,
0.82275390625,
0.8359375,
0.176513671875,
0.8359375
],
[
0.176513671875,
0.83544921875,
0.82275390625,
0.83544921875,
0.82275390625,
0.8486328125,
0.176513671875,
0.8486328125
],
[
0.176513671875,
0.849609375,
0.82275390625,
0.849609375,
0.82275390625,
0.86279296875,
0.176513671875,
0.86279296875
],
[
0.176513671875,
0.86279296875,
0.82275390625,
0.86279296875,
0.82275390625,
0.876953125,
0.176513671875,
0.876953125
],
[
0.178955078125,
0.87744140625,
0.483642578125,
0.87744140625,
0.483642578125,
0.88818359375,
0.178955078125,
0.88818359375
],
[
0.1748046875,
0.89599609375,
0.82177734375,
0.896484375,
0.82177734375,
0.9111328125,
0.1748046875,
0.91015625
],
[
0.1748046875,
0.9111328125,
0.82275390625,
0.9111328125,
0.82275390625,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2023",
"神经符号系统的软化符号接地",
"Zenan Li', Yuan Yaol, Taolue Chen?, Jingwei Xul, Chun Caol, Xiaoxing Ma', Jian Luil",
"1State Key Lab of Novel Software Technology and",
"Department of Computer Science and Technology, Nanjing University, China",
"2Department of Computer Science, Birkbeck, University of London, UK",
"lizn@smail.nju.edu.cn, t.chen@bbk.ac.uk.",
" y.yao,jingweix,caochun,xxm,lj ] @nju.edu.cn",
"Abstract",
"Neuro-symbolic learning generally consists of two separated worlds, i.e.,",
"neural network training and symbolic constraint solving, whose success",
"hinges on symbol grounding, a fundamental problem in AI. This paper",
"presents a novel, softened symbol grounding process, bridging the gap be-",
"tween the two worlds, and resulting in an effective and efficient neuro-",
"symbolic learning framework. Technically, the framework features (1) mod-",
"eling of symbol solution states as a Boltzmann distribution, which avoids",
"expensive state searching and facilitates mutually beneficial interactions",
"between network training and symbolic reasoning; (2) a new MCMC tech-",
"nique leveraging projection and SMT solvers, which efficiently samples from",
"disconnected symbol solution spaces; (3) an annealing mechanism that can",
"escapefrom sub-optimal symbolgroundings.Experiments with threerep",
"resentative neuro-symbolic learning tasks demonstrate that, owining to its",
"superior symbol grounding capability, our framework successfully solves",
"problems well beyond the frontier of the existing proposals.",
"1Introduction",
"已经提出了神经符号系统来连接神经网络学习和符号约束满足E(Zhou，2019:Garcezetal.",
"2019；Marraetal.，2021；Yuetal.，2021；Hitzler，2022）。在这些系统中，神经网络组件首",
"先将rawinput识别为符号，然后将其进一步馈送到符号组件中以产生finaloutputE（Yi",
"etal.，2018;Lietal.，2020；Liangetal.，2017）。这种神经符号范式已显示出前所未有的能",
"力，并在许多任务中取得了令人印象深刻的成果，包括视觉问答E(Yietal.，2018；Vedantam",
"et al.,2019;Amizadeh et al.,2020)，视觉语言导航E(Anderson et al.,2018;Fried et al.",
"2018）和数学单词问题解决E(Hongetal.，2021Qinetal.，2021），仅举几例",
"如图E1所示，为了最大程度地提高通用性，此类问题通常在弱监督设置中进行E(Garcez",
"etal.，2022）：神经符号计算的最终输出是在训练过程中作为监督提供的，而不是中间符",
"号的标签。缺乏直接监督标签的网络训练呼吁一种有效和高效的方法来解决符号接地问题",
"我。e。，建立从原始输人到潜在符号的可行且可推广的映射。请注意，旁路符号接地（通过",
"e。g。，将问题视为有逻辑约束的学习）是可能的，但不能达到令人满意的性能（Manhaeve",
"etal.，2018；Xuetal.,2018；Pryoretal.,2022）。在网络学习中结合符号接地的现有方法",
"严重依赖于良好的初始模型，并且在从头开始时表现不佳（Daietal.，2019；Lietal.2020;",
"Huang et al., 2021)",
"符号接地的一个关键挑战在于随机和连续的神经学习与确定性和离散的符号推理之间的语",
"义鸿沟。为了弥合这一差距，我们建议soften将符号接地。也就是说，我们不是直接搜索",
"确定性的输人符号映射，而是使用退火策略来优化其玻尔兹曼分布，以逐渐收敛到确定性",
"的。直观上，软化的玻尔兹曼分布提供了一个游乐场，在那里输入符号映射的搜索可以由神",
"经网络引导，网络训练可以通过从分布中采样来监督。博奔论确实为这种策略提供了理论",
"支持（Conitzer，2016）：软化使学习过程在退火过程中成为一系列混合策略游戏，这鼓励",
"了神经世界和符号世界之间更强的相互作用",
"剩下的挑战是如何有效地采样可行的输入符号映射。具体而言，可行解在整个符号空间中",
"极为稀疏，不同解之间的连接较差，这使得马尔可夫链蒙特卡罗（MCMC）抽样无法有效",
"1"
] |
{
"arxiv_id": "2403.00323",
"language": "zh",
"page_idx": 0
}
|
||
2403.00323-zh-1
|
2403.00323-zh-1.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
<FigureHere>
Figure 1: An example neural-symbolic system for handwritten formula evaluation. It takes a handwritten arithmetic expression $x$ as input and evaluate the expression to output $y.$ The neural network component $\mathcal{M}_{\boldsymbol{\theta}}$ recognizes the symbols $\boldsymbol z$ (i.e., digits and operators) in the expression, and the symbolic component evaluates the recognized formula by, e.g., the Python function 'eval' The challenge in training $\mathcal{M}_{\theta}$ comes from the lack of explicit $z$ to bridge the gap between the neural world $(x$ to $z)$ and the symbol world $(z$ to $\bar{y}).$ Through softened symbol grounding, the model training and the constraint satisfaction join force to resolve the latent $z$ to ft both the given $x$ and $y.$
地探索解空间。为了克服这一缺陷，我们利用投影技术来加速采样$\overline{\mathrm{E}}($Feng et al., 2021b) 的随机游走，并借助可满足性模理论(SMT)求解器$\mathbb{E}($Nieuwenhuis et al., 2006; Moura & Bjørner, 2008) 。直觉是，当将高维空间中的断开连接的解决方案投影到低维空间中时，它们可能会变得连接，从而导致 MCMC 采样$\overline{\mathrm{E}}($Feng et al., 2021a) 的快速混合时间。按需调用的 SMT 求解器用作计算逆投影的通用方法。尽管 MCMC 采样和 SMT 求解器可能会引人偏差，但理论结果证实，所提出的随机梯度下降算法可以令人满意地抵消偏差
对各种典型的神经符号任务的实验评估，包括手写表情评估，视觉数独分类和最短路径搜
索，证明了所提出的方法优于最先进的方法
### 2 Softening Svmbol Grounding
在整篇文章中，我们将$\mathcal{X}$称为神经符号系统的输入空间，将$\mathcal{Z}$称为其符号空间或状态空间(e。g。, HWF 任务中的所有合法和非法算术表达式)。我们考虑神经符号计算任务，该任务首先训练神经网络(由$\boldsymbol{\theta}$参数化),将原始输入$x\in\mathcal{X}$映射到具有 (变分)概率分布$P_{\boldsymbol{\theta}}(z|x)$ 的某些潜在状态$z\in Z$ 。状态$z$被进一步馈送到预定义的符号推理过程中，以产生最终输出$y$。训练数据仅包含输入$x$和相应的$y$ ,这会将问题转换为所谓的弱监督设置。通常， 我们将预定义的符号推理过程和输出$y$表示为符号空间上的一组 symbolic constraint $S_{y}$ 例如，在图 1 中，约束指定算术表达式必须计算为 42。我们说状态$z$是 feasible ,或者如果$z\in\mathcal{S}_y$满足符号约束
这种神经符号学习范式的主要挑战在于 symbol grounding problem , i。e。, 建立从原始输人到满足符号约束的可行状态的映射$h{:}\mathcal{X}\to\mathcal{Z}$ 。具体来说，有效的映射$h$应使模型能够解释尽可能多的观察结果。结果，给定数据集$\{(x^i,y^i)\}_{i=1,\ldots,N}$上的符号接地问题可以表示为
$$\min_{h}\left\{\min_{\boldsymbol{\theta}}\ell(\boldsymbol{\theta}):=-\sum_{i=1}^{N}\log P_{\boldsymbol{\theta}}(\boldsymbol{z}^{i}|\boldsymbol{x}^{i})\right\}\quad\mathrm{s.t.}\quad\boldsymbol{z}^{i}=h(\boldsymbol{x}^{i})\in\mathcal{S}_{\boldsymbol{y}^{i}},i=1,\ldots,N.$$
(1)
上述公式的直接解决方案将首先为每个可行映射训练一个网络，然后选择一个以最大似然实现最终输出$y$的网络。然而，这个解决方案是不切实际的，因为可行映射的数量呈指数增长
上述解决方案的一个明显缺点是，神经网络学习过程没有利用符号约束中体现的知识。反之亦然，搜索最佳映射不是由网络引导的。为了克服这个缺点，可以切换问题$\color{red}{\mathbb{F}}(1)$ 中的最小化顺序，并获得一个新的但在数值上等效的问题：
$$\min\limits_{\boldsymbol{\theta}}\left\{\min\limits_{h}\ell(\boldsymbol{\theta}):=-\sum\limits_{i=1}^{N}\log P_{\boldsymbol{\theta}}(\boldsymbol{z}^{i}|\boldsymbol{x}^{i})\quad\mathrm{s.t.}\quad\boldsymbol{z}^{i}=h(\boldsymbol{x}^{i})\in\mathcal{S}_{\boldsymbol{y}^{i}},i=1,\ldots,N\right\}.$$
(2)
2
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.259033203125,
0.10797119140625,
0.325927734375,
0.10797119140625,
0.325927734375,
0.11932373046875,
0.259033203125,
0.11932373046875
],
[
0.496826171875,
0.111083984375,
0.5849609375,
0.111083984375,
0.5849609375,
0.121826171875,
0.496826171875,
0.121826171875
],
[
0.7138671875,
0.10986328125,
0.76318359375,
0.10986328125,
0.76318359375,
0.121826171875,
0.7138671875,
0.121826171875
],
[
0.2320556640625,
0.125,
0.381591796875,
0.125,
0.381591796875,
0.1427001953125,
0.2320556640625,
0.1427001953125
],
[
0.61767578125,
0.12060546875,
0.6611328125,
0.117431640625,
0.66650390625,
0.154052734375,
0.62353515625,
0.1572265625
],
[
0.40283203125,
0.1275634765625,
0.4453125,
0.121826171875,
0.450927734375,
0.1427001953125,
0.408447265625,
0.1483154296875
],
[
0.5009765625,
0.1287841796875,
0.56884765625,
0.1287841796875,
0.56884765625,
0.14013671875,
0.5009765625,
0.14013671875
],
[
0.71728515625,
0.126953125,
0.75732421875,
0.126953125,
0.75732421875,
0.1427001953125,
0.71728515625,
0.1427001953125
],
[
0.386474609375,
0.15283203125,
0.459228515625,
0.15283203125,
0.459228515625,
0.1641845703125,
0.386474609375,
0.1641845703125
],
[
0.63134765625,
0.1546630859375,
0.66015625,
0.1546630859375,
0.66015625,
0.1666259765625,
0.63134765625,
0.1666259765625
],
[
0.47216796875,
0.1610107421875,
0.55224609375,
0.1610107421875,
0.55224609375,
0.17236328125,
0.47216796875,
0.17236328125
],
[
0.328369140625,
0.17236328125,
0.400390625,
0.1749267578125,
0.400390625,
0.189453125,
0.328369140625,
0.1875
],
[
0.2303466796875,
0.1925048828125,
0.306396484375,
0.1944580078125,
0.306396484375,
0.20703125,
0.2303466796875,
0.2052001953125
],
[
0.449462890625,
0.1956787109375,
0.53173828125,
0.1962890625,
0.53173828125,
0.2083740234375,
0.449462890625,
0.20703125
],
[
0.5556640625,
0.1956787109375,
0.61376953125,
0.1956787109375,
0.61376953125,
0.20703125,
0.5556640625,
0.20703125
],
[
0.65625,
0.1932373046875,
0.689453125,
0.1932373046875,
0.689453125,
0.20263671875,
0.65625,
0.20263671875
],
[
0.250732421875,
0.2083740234375,
0.30224609375,
0.2083740234375,
0.30224609375,
0.2197265625,
0.250732421875,
0.2197265625
],
[
0.44775390625,
0.203857421875,
0.60888671875,
0.2108154296875,
0.60693359375,
0.2259521484375,
0.44677734375,
0.2183837890625
],
[
0.25,
0.2215576171875,
0.301513671875,
0.2235107421875,
0.301513671875,
0.23486328125,
0.25,
0.23291015625
],
[
0.650390625,
0.2197265625,
0.7158203125,
0.2197265625,
0.7158203125,
0.2310791015625,
0.650390625,
0.2310791015625
],
[
0.3447265625,
0.2291259765625,
0.619140625,
0.2298583984375,
0.619140625,
0.2442626953125,
0.3447265625,
0.24365234375
],
[
0.64453125,
0.2291259765625,
0.71826171875,
0.2310791015625,
0.71826171875,
0.24560546875,
0.64453125,
0.24365234375
],
[
0.176513671875,
0.261962890625,
0.82421875,
0.261962890625,
0.82421875,
0.27587890625,
0.176513671875,
0.27587890625
],
[
0.1748046875,
0.2734375,
0.81640625,
0.27587890625,
0.81640625,
0.290283203125,
0.1748046875,
0.28857421875
],
[
0.1748046875,
0.28857421875,
0.82177734375,
0.290283203125,
0.82177734375,
0.3037109375,
0.1748046875,
0.3017578125
],
[
0.1732177734375,
0.3017578125,
0.82177734375,
0.302978515625,
0.82177734375,
0.317626953125,
0.1732177734375,
0.31640625
],
[
0.1732177734375,
0.31494140625,
0.82275390625,
0.31640625,
0.82275390625,
0.330810546875,
0.1732177734375,
0.330078125
],
[
0.17724609375,
0.330078125,
0.82177734375,
0.330078125,
0.82177734375,
0.3447265625,
0.17724609375,
0.3447265625
],
[
0.176513671875,
0.343505859375,
0.82421875,
0.343505859375,
0.82421875,
0.356689453125,
0.176513671875,
0.356689453125
],
[
0.1732177734375,
0.35595703125,
0.6005859375,
0.356689453125,
0.6005859375,
0.371826171875,
0.1732177734375,
0.37060546875
],
[
0.176513671875,
0.392578125,
0.82275390625,
0.392578125,
0.82275390625,
0.406005859375,
0.176513671875,
0.406005859375
],
[
0.176513671875,
0.406005859375,
0.82275390625,
0.406005859375,
0.82275390625,
0.419189453125,
0.176513671875,
0.419189453125
],
[
0.176513671875,
0.419189453125,
0.82275390625,
0.419189453125,
0.82275390625,
0.43310546875,
0.176513671875,
0.43310546875
],
[
0.17724609375,
0.434326171875,
0.82275390625,
0.434326171875,
0.82275390625,
0.447509765625,
0.17724609375,
0.447509765625
],
[
0.176513671875,
0.4462890625,
0.82177734375,
0.4462890625,
0.82177734375,
0.460205078125,
0.176513671875,
0.460205078125
],
[
0.176513671875,
0.4609375,
0.74853515625,
0.4609375,
0.74853515625,
0.474853515625,
0.176513671875,
0.474853515625
],
[
0.176513671875,
0.481689453125,
0.82275390625,
0.481689453125,
0.82275390625,
0.49560546875,
0.176513671875,
0.49560546875
],
[
0.1732177734375,
0.49560546875,
0.484375,
0.49560546875,
0.484375,
0.5087890625,
0.1732177734375,
0.5087890625
],
[
0.17236328125,
0.53076171875,
0.455078125,
0.53271484375,
0.455078125,
0.54736328125,
0.17236328125,
0.54541015625
],
[
0.176513671875,
0.564453125,
0.82275390625,
0.564453125,
0.82275390625,
0.57763671875,
0.176513671875,
0.57763671875
],
[
0.1748046875,
0.57763671875,
0.82177734375,
0.5771484375,
0.82177734375,
0.59130859375,
0.1748046875,
0.59228515625
],
[
0.17724609375,
0.59130859375,
0.82177734375,
0.59130859375,
0.82177734375,
0.60498046875,
0.17724609375,
0.60498046875
],
[
0.176513671875,
0.60498046875,
0.82177734375,
0.60498046875,
0.82177734375,
0.61865234375,
0.176513671875,
0.61865234375
],
[
0.176513671875,
0.61865234375,
0.82275390625,
0.61865234375,
0.82275390625,
0.63330078125,
0.176513671875,
0.63330078125
],
[
0.1732177734375,
0.6298828125,
0.82275390625,
0.630859375,
0.82275390625,
0.6484375,
0.1732177734375,
0.64794921875
],
[
0.1748046875,
0.64453125,
0.82177734375,
0.64599609375,
0.82177734375,
0.66015625,
0.1748046875,
0.6591796875
],
[
0.176513671875,
0.66015625,
0.36767578125,
0.66015625,
0.36767578125,
0.673828125,
0.176513671875,
0.673828125
],
[
0.1748046875,
0.6787109375,
0.82177734375,
0.6796875,
0.82177734375,
0.6943359375,
0.1748046875,
0.693359375
],
[
0.17724609375,
0.6943359375,
0.82177734375,
0.6943359375,
0.82177734375,
0.70849609375,
0.17724609375,
0.70849609375
],
[
0.1732177734375,
0.705078125,
0.82275390625,
0.70556640625,
0.82275390625,
0.7236328125,
0.1732177734375,
0.72265625
],
[
0.1748046875,
0.7216796875,
0.193603515625,
0.7216796875,
0.193603515625,
0.73486328125,
0.1748046875,
0.73486328125
],
[
0.2254638671875,
0.75,
0.488525390625,
0.75,
0.488525390625,
0.767578125,
0.2254638671875,
0.767578125
],
[
0.50634765625,
0.74853515625,
0.826171875,
0.75,
0.826171875,
0.767578125,
0.50634765625,
0.7666015625
],
[
0.2320556640625,
0.76318359375,
0.291748046875,
0.76318359375,
0.291748046875,
0.7734375,
0.2320556640625,
0.7734375
],
[
0.371826171875,
0.767578125,
0.390625,
0.767578125,
0.390625,
0.77587890625,
0.371826171875,
0.77587890625
],
[
0.178955078125,
0.78857421875,
0.82177734375,
0.78857421875,
0.82177734375,
0.8017578125,
0.178955078125,
0.8017578125
],
[
0.176513671875,
0.80029296875,
0.82275390625,
0.80029296875,
0.82275390625,
0.81494140625,
0.176513671875,
0.81494140625
],
[
0.1748046875,
0.81494140625,
0.2091064453125,
0.81494140625,
0.2091064453125,
0.8271484375,
0.1748046875,
0.8271484375
],
[
0.181396484375,
0.8369140625,
0.82177734375,
0.8369140625,
0.82177734375,
0.84765625,
0.181396484375,
0.84765625
],
[
0.176513671875,
0.849609375,
0.82275390625,
0.849609375,
0.82275390625,
0.86279296875,
0.176513671875,
0.86279296875
],
[
0.176513671875,
0.86279296875,
0.55810546875,
0.86279296875,
0.55810546875,
0.876953125,
0.176513671875,
0.876953125
],
[
0.373291015625,
0.88818359375,
0.38818359375,
0.88818359375,
0.38818359375,
0.896484375,
0.373291015625,
0.896484375
],
[
0.226318359375,
0.90087890625,
0.35546875,
0.90087890625,
0.35546875,
0.9140625,
0.226318359375,
0.9140625
],
[
0.346435546875,
0.8984375,
0.7646484375,
0.8984375,
0.7646484375,
0.91650390625,
0.346435546875,
0.91650390625
],
[
0.79345703125,
0.90576171875,
0.79833984375,
0.90576171875,
0.79833984375,
0.9111328125,
0.79345703125,
0.9111328125
],
[
0.79736328125,
0.9013671875,
0.82177734375,
0.9013671875,
0.82177734375,
0.9140625,
0.79736328125,
0.9140625
],
[
0.26953125,
0.91552734375,
0.28515625,
0.91552734375,
0.28515625,
0.91845703125,
0.26953125,
0.91845703125
],
[
0.369384765625,
0.91650390625,
0.393798828125,
0.91650390625,
0.393798828125,
0.9287109375,
0.369384765625,
0.9287109375
],
[
0.49267578125,
0.94873046875,
0.50634765625,
0.94873046875,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2023",
"Rawinput x",
"Latent symbol z",
"Outputy",
"4×8-5-9",
"邮",
"容",
"7*8-5-9",
"42.0",
"Network Me",
"Eval",
"Supervising M",
"Improving Q",
"Information flow",
"Projection-based",
"4*8+3±7",
"Z3",
"：Inference",
"MCMCsampling7*8-5-9",
"Learning",
"On-demand",
"Boltzmann distribution Q over solution space Sy",
"SMT solving",
"Figure 1: An example neural-symbolic system for handwritten formula evaluation. It takes",
"a handwritten arithmetic expression α as input and evaluate the expression to output y",
"The neural network component Me recognizes the symbols z (i.e., digits and operators)",
"in the expression, and the symbolic component evaluates the recognized formula by, e.g.",
"the Python function 'eval': The challenge in training Me comes from the lack of explicit",
"z to bridge the gap between the neural world ( to z ) and the symbol world ( z to y ).",
"Through softened symbol grounding, the model training and the constraint satisfaction join",
"force to resolve the latent z to fit both the given a and y .",
"地探索解空间。为了克服这一缺陷，我们利用投影技术来加速采样E(Fengetal.，2021b）的",
"随机游走，并借助可满足性模理论（SMT）求解器E(Nieuwenhuisetal.，2006：Moura&",
"Bjorner，2008）。直觉是，当将高维空间中的断开连接的解决方案投影到低维空间中时，它",
"们可能会变得连接，从而导致MCMC采样E(Fengetal.，202la）的快速混合时间。按需调",
"用的SMT求解器用作计算逆投影的通用方法。尽管MCMC采样和SMT求解器可能会引",
"人偏差，但理论结果证实，所提出的随机梯度下降算法可以令人满意地抵消偏差",
"对各种典型的神经符号任务的实验评估，包括手写表情评估，视觉数独分类和最短路径搜",
"索，证明了所提出的方法优于最先进的方法",
"2Softening Symbol Grounding",
"在整篇文章中，我们将称为神经符号系统的输入空间，将2称为其符号空间或状态空间",
"(e。g。，HWF任务中的所有合法和非法算术表达式）。我们考虑神经符号计算任务，该任务",
"首先训练神经网络（由θ参数化），将原始输入aEX映射到具有（变分）概率分布Pe(zα）",
"的某些潜在状态zE2。状态z被进一步馈送到预定义的符号推理过程中，以产生最终输",
"出y。训练数据仅包含输人和相应的y，这会将问题转换为所谓的弱监督设置。通常，",
"我们将预定义的符号推理过程和输出y表示为符号空间上的一组symbolic constraintSu.",
"例如，在图1中，约束指定算术表达式必须计算为42。我们说状态z是feasible，或者",
"如果zESu满足符号约束",
"这种神经符号学习范式的主要挑战在于symbolgroundingproblem，i。e。，建立从原始输",
"人到满足符号约束的可行状态的映射h：→。具体来说，有效的映射h应使模型能够",
"解释尽奇能多的观察结果。结果，给定数据集（(α,y\")}i=1….N上的符号接地问题可以表示",
"为",
"min /minl(0) := -log Pe(z'ci)",
"s.t. z\"= h(α')eSy,i= l,..., N.(1)",
"h1e",
"-1",
"上述公式的直接解决方案将首先为每个可行映射训练一个网络，然后选择一个以最大似然",
"实现最终输出的网络。然而，这个解决方案是不切实际的，因为可行映射的数量呈指数",
"增长",
"上述解决方案的一个明显缺点是，神经网络学习过程没有利用符号约束中体现的知识。反",
"之亦然，搜索最佳映射不是由网络引导的。为了克服这个缺点，可以切换问题E（1）中的",
"最小化顺序，并获得一个新的但在数值上等效的问题：",
"N",
"min3 min l(0) :=",
"-log Po(za\") s.t. zi=h(ai) ESy.,i=1,...,N",
"",
"(2)",
"",
"i=1",
"2"
] |
{
"arxiv_id": "2403.00323",
"language": "zh",
"page_idx": 1
}
|
||
2403.00323-zh-2
|
2403.00323-zh-2.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
The optimization problem $(\begin{array}{c}2\end{array})$ first determines a“best”mapping based on the initial model, and then updates the model to fit this mapping. The two steps are iterated until no improvement can be made. 但是，这种接地策略可能很容易陷入局部最优。原因是，每当实现可行的映射$h$时，$h$往往会引导神经网络(过度)拟合自己。因为映射是确定性的和离散的，所以没有平滑的路由到替代的可行映射，这将进一步改善网络。网络训练和符号推理之间的信息交换不足，使得符号接地的成功高度依赖于初始模型的质量
在这项工作中，我们建议软化符号接地，以促进神经感知和符号推理之间的相互作用。精心地说，我们不是直接搜索确定性映射$h$ ,而是首先追求$h$的最佳概率分布，然后逐渐“锐化”该分布以获得最终的确定性$h$ 。形式上，对于每个输人$x$ ,我们在$S_y$上引人玻尔兹曼分布$Q_{\boldsymbol{\phi}}$,由$\phi$参数化，以指示满足符号约束的每个可行状态的概率。然后，可以将软化的符号接地问题表述如下：
$$\begin{aligned}&\min_{\boldsymbol{\theta},\boldsymbol{\phi}}\quad\ell(\boldsymbol{\theta},\boldsymbol{\phi}):=-\sum_{i=1}^{N}\sum_{\boldsymbol{z}^{i}\in\mathcal{S}_{\boldsymbol{y}^{i}}}Q_{\boldsymbol{\phi}}^{i}(\boldsymbol{z}^{i})\log P_{\boldsymbol{\theta}}(\boldsymbol{z}^{i}|\boldsymbol{x}^{i})+\gamma Q_{\boldsymbol{\phi}}^{i}(\boldsymbol{z}^{i})\log Q_{\boldsymbol{\phi}}^{i}(\boldsymbol{z}^{i})\\&\mathrm{s.t.}\quad\mathrm{supp}(Q_{\boldsymbol{\phi}}^{i})\subseteq\mathcal{S}_{\boldsymbol{y}^{i}},\quad i=1,\ldots,N,\end{aligned}$$
(P)
其中 supp$(Q_\phi)$表示$Q_\phi$的支持。引人熵项$\sum_{\boldsymbol{z}\in\mathcal{S}_y}Q_\phi(z)\log Q_\phi(z)$来控制$Q_\phi$ (MacKay 接将$P_{\boldsymbol{\theta}}$和$Q_{\boldsymbol{\phi}}$之间的 KL 发散减少到$h$的情况外，我们还研究了两种极端情况：(1)当$\gamma \to + \infty$ , $Q_\phi$被迫向均匀分布时，因此最小化仅旨在限制$P_{\boldsymbol{\theta}}$对 XMATHX\_
Advantages 。博弈论提供了一个视角来理解为什么我们的软化策略通过模型训练和符号推理之间更好的交互来改进问题(1)和(2)。问题$\mathbb{E}(1)$或$\mathbb{E}(2)$ 都可以视为纯策略Stackelberg 游戏。也就是说，模型训练和符号推理都被迫采取一定的行动(e。g。,在优化过程中选择确定性映射 $h$ )。相反，问题$\overline{\mathbb{E}}(\mathbb{P})$ 可以看作是具有混合策略的 Stackelberg 游戏，其中玩家以分布$Q_\mathrm{\phi}$采取随机行动。与纯策略相比，混合策略确实提供了更多的博弈信息。从而严格提高了模型训练的效用 (Letchford et al.,2014;Conitzer,2016)。1
此外，这种软化技术还避免了枚举，从而提高了效率。在问题$\mathbb{E}(1)$ 中，对于每个输人 $x$ , 其相应符号$z$中的最小化需要在整个$\mathcal{S}_y$ (i。e。, 枚举满足符号约束的所有可行状态)。这在理论上通常是一个棘手的 #P-完全问题 (Arora & Barak, 2009)。问题$\overline{\mathbb{E}}(\color{red}{P})$ 通过估计$Q_{\phi}$的期望值来规避这种昂贵的计算，可以使用下一节中讨论的定制采样策略来有效地计算
3 Markov chain Monte Carlo Estimate via Projection
为了便于演示，在本节中，我们考虑单个数据样本。具体来说，通过删除所有样本的总和并
删除上标，问题$\mathbb{E}($P)可以表示为
$$\min_{\theta,\phi}\:\ell(\theta,\phi):=\sum_{z\in S_{y}}Q_{\phi}(z)\log P_{\theta}(z|x)+\gamma Q_{\phi}(z)\log Q_{\phi}(z),\quad\mathrm{supp}(Q_{\phi})\subseteq\mathcal{S}_{y}.$$
该问题可以通过在$\theta$中的梯度下降步骤和$\phi$中的最小化步骤之间交替解决。$\theta$和$\phi$在第 k
次迭代时的更新为
(4)
$$\boldsymbol{\theta}_{k+1}=\boldsymbol{\theta}_{k}-\eta\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{\theta}_{k},\phi_{k}),\quad\phi_{k+1}=\underset{\boldsymbol{\phi}|\mathrm{supp}(Q_{\boldsymbol{\phi}})}{\arg\operatorname*{min}}\:\ell(\boldsymbol{\theta}_{k+1},\phi).$$
注意，当$P_{\boldsymbol{\theta}}$固定时，存在闭式解$Q_{\boldsymbol{\phi}^*}$ ,确保了$\boldsymbol{\theta}$ $\boxed{\mathrm{E}}($Jin et al., 2020, Theorem 31) 中梯度下降的收敛性。有关详细信息，较低级别的问题，我。e。,$\min_{\boldsymbol{\phi}}\ell(\boldsymbol{\theta},\phi)$ ,是严格凸的，因此包含唯一的最小值：
$$Q_{\phi^*}(\boldsymbol{z})=\begin{cases}P_{\boldsymbol{\theta}}(\boldsymbol{z}|\boldsymbol{x})^{\frac{1}{\gamma}}/\sum_{\boldsymbol{z}'\in\mathcal{S}_{\boldsymbol{y}}}P_{\boldsymbol{\theta}}(\boldsymbol{z}'|\boldsymbol{x})^{\frac{1}{\gamma}},&\mathrm{if}\:\boldsymbol{z}\in\mathcal{S}_{\boldsymbol{y}},\\0,&\mathrm{otherwise}.\end{cases}$$
(5)
给定封闭形式的解$Q_{\phi^*}$,可以通过$Q_{\phi^*}$上的蒙特卡洛采样来估计损失函数 $\ell(\boldsymbol{\theta},\boldsymbol{\phi}^*)$ 及其梯度$\nabla_{\boldsymbol{\theta}}\ell(\boldsymbol{\theta},\boldsymbol{\phi}^*)$。剩下的问题是如何对$^\prime Q_{\boldsymbol{\phi}^*}$ 进行采样，由于 $S_{\boldsymbol{y}}$ 的结构未知，因此具有挑战性。现有方法通常从整个符号/状态空间$^{\prime}z$ 中采样，然后要么拒绝状态 $z\notin\mathcal{S_y}$ (e。g。,策
Conitzer (2016) provides a nice illustrative example, which one may consult for more details
3
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.1829833984375,
0.1060791015625,
0.82421875,
0.1060791015625,
0.82421875,
0.11932373046875,
0.1829833984375,
0.11932373046875
],
[
0.176513671875,
0.11871337890625,
0.82275390625,
0.11871337890625,
0.82275390625,
0.1319580078125,
0.176513671875,
0.1319580078125
],
[
0.1748046875,
0.1331787109375,
0.82177734375,
0.1319580078125,
0.82177734375,
0.1458740234375,
0.1748046875,
0.146484375
],
[
0.178955078125,
0.147705078125,
0.8203125,
0.147705078125,
0.8203125,
0.1578369140625,
0.178955078125,
0.1578369140625
],
[
0.1748046875,
0.1590576171875,
0.82177734375,
0.1590576171875,
0.82177734375,
0.17236328125,
0.1748046875,
0.17236328125
],
[
0.17724609375,
0.17236328125,
0.68310546875,
0.17236328125,
0.68310546875,
0.186279296875,
0.17724609375,
0.186279296875
],
[
0.176513671875,
0.1932373046875,
0.82177734375,
0.1932373046875,
0.82177734375,
0.20703125,
0.176513671875,
0.20703125
],
[
0.176513671875,
0.2083740234375,
0.82177734375,
0.2083740234375,
0.82177734375,
0.2215576171875,
0.176513671875,
0.2215576171875
],
[
0.176513671875,
0.2197265625,
0.82177734375,
0.2203369140625,
0.82177734375,
0.23486328125,
0.176513671875,
0.2342529296875
],
[
0.176513671875,
0.23486328125,
0.82421875,
0.23486328125,
0.82421875,
0.248779296875,
0.176513671875,
0.248779296875
],
[
0.17724609375,
0.25,
0.37744140625,
0.25,
0.37744140625,
0.260009765625,
0.17724609375,
0.260009765625
],
[
0.386474609375,
0.270751953125,
0.40283203125,
0.270751953125,
0.40283203125,
0.27978515625,
0.386474609375,
0.27978515625
],
[
0.2442626953125,
0.285400390625,
0.3896484375,
0.276611328125,
0.390625,
0.294189453125,
0.2459716796875,
0.302978515625
],
[
0.3798828125,
0.27978515625,
0.7548828125,
0.27978515625,
0.7548828125,
0.29736328125,
0.3798828125,
0.29736328125
],
[
0.381591796875,
0.29541015625,
0.449462890625,
0.2998046875,
0.44775390625,
0.314453125,
0.3798828125,
0.31005859375
],
[
0.79833984375,
0.294189453125,
0.826171875,
0.294189453125,
0.826171875,
0.311279296875,
0.79833984375,
0.311279296875
],
[
0.253173828125,
0.3193359375,
0.521484375,
0.318115234375,
0.521484375,
0.333251953125,
0.253173828125,
0.333984375
],
[
0.176513671875,
0.343505859375,
0.82177734375,
0.343505859375,
0.82177734375,
0.35791015625,
0.176513671875,
0.35791015625
],
[
0.176513671875,
0.361083984375,
0.82275390625,
0.361083984375,
0.82275390625,
0.375732421875,
0.176513671875,
0.375732421875
],
[
0.176513671875,
0.375,
0.82275390625,
0.375,
0.82275390625,
0.38818359375,
0.176513671875,
0.38818359375
],
[
0.1732177734375,
0.389404296875,
0.73046875,
0.386962890625,
0.73046875,
0.40087890625,
0.1732177734375,
0.40283203125
],
[
0.176513671875,
0.4091796875,
0.82177734375,
0.4091796875,
0.82177734375,
0.42236328125,
0.176513671875,
0.42236328125
],
[
0.1748046875,
0.42236328125,
0.82275390625,
0.42236328125,
0.82275390625,
0.436279296875,
0.1748046875,
0.436279296875
],
[
0.1732177734375,
0.43505859375,
0.82177734375,
0.436279296875,
0.82177734375,
0.45068359375,
0.1732177734375,
0.449462890625
],
[
0.176513671875,
0.45068359375,
0.82275390625,
0.45068359375,
0.82275390625,
0.464111328125,
0.176513671875,
0.464111328125
],
[
0.17724609375,
0.46533203125,
0.8203125,
0.46533203125,
0.8203125,
0.47607421875,
0.17724609375,
0.47607421875
],
[
0.176513671875,
0.479248046875,
0.75341796875,
0.479248046875,
0.75341796875,
0.492431640625,
0.176513671875,
0.492431640625
],
[
0.176513671875,
0.498779296875,
0.81884765625,
0.5,
0.81884765625,
0.5146484375,
0.176513671875,
0.51318359375
],
[
0.176513671875,
0.51318359375,
0.82177734375,
0.51318359375,
0.82177734375,
0.52783203125,
0.176513671875,
0.52783203125
],
[
0.176513671875,
0.52880859375,
0.82177734375,
0.52880859375,
0.82177734375,
0.54248046875,
0.176513671875,
0.54248046875
],
[
0.176513671875,
0.54248046875,
0.82177734375,
0.54248046875,
0.82177734375,
0.55615234375,
0.176513671875,
0.55615234375
],
[
0.1732177734375,
0.5771484375,
0.640625,
0.5771484375,
0.640625,
0.59033203125,
0.1732177734375,
0.59033203125
],
[
0.17724609375,
0.60693359375,
0.82177734375,
0.60693359375,
0.82177734375,
0.62060546875,
0.17724609375,
0.62060546875
],
[
0.17724609375,
0.619140625,
0.4306640625,
0.619140625,
0.4306640625,
0.6337890625,
0.17724609375,
0.6337890625
],
[
0.21728515625,
0.6416015625,
0.82666015625,
0.63818359375,
0.82666015625,
0.66357421875,
0.21728515625,
0.66650390625
],
[
0.330078125,
0.662109375,
0.363525390625,
0.662109375,
0.363525390625,
0.673828125,
0.330078125,
0.673828125
],
[
0.176513671875,
0.68310546875,
0.82421875,
0.68310546875,
0.82421875,
0.6962890625,
0.176513671875,
0.6962890625
],
[
0.1748046875,
0.6962890625,
0.30810546875,
0.6962890625,
0.30810546875,
0.71044921875,
0.1748046875,
0.71044921875
],
[
0.291748046875,
0.71630859375,
0.7060546875,
0.71630859375,
0.7060546875,
0.73095703125,
0.291748046875,
0.73095703125
],
[
0.80224609375,
0.71826171875,
0.82421875,
0.71826171875,
0.82421875,
0.73095703125,
0.80224609375,
0.73095703125
],
[
0.560546875,
0.73095703125,
0.62744140625,
0.73095703125,
0.62744140625,
0.7421875,
0.560546875,
0.7421875
],
[
0.176513671875,
0.751953125,
0.82275390625,
0.751953125,
0.82275390625,
0.76513671875,
0.176513671875,
0.76513671875
],
[
0.178955078125,
0.76513671875,
0.82275390625,
0.76513671875,
0.82275390625,
0.77880859375,
0.178955078125,
0.77880859375
],
[
0.17724609375,
0.78076171875,
0.330078125,
0.78076171875,
0.330078125,
0.79150390625,
0.17724609375,
0.79150390625
],
[
0.38232421875,
0.80322265625,
0.69140625,
0.80322265625,
0.69140625,
0.82080078125,
0.38232421875,
0.82080078125
],
[
0.3056640625,
0.81298828125,
0.373291015625,
0.81298828125,
0.373291015625,
0.82763671875,
0.3056640625,
0.82763671875
],
[
0.80224609375,
0.8125,
0.82421875,
0.8125,
0.82421875,
0.8271484375,
0.80224609375,
0.8271484375
],
[
0.386474609375,
0.82470703125,
0.40283203125,
0.82470703125,
0.40283203125,
0.833984375,
0.386474609375,
0.833984375
],
[
0.61669921875,
0.82373046875,
0.68896484375,
0.82373046875,
0.68896484375,
0.833984375,
0.61669921875,
0.833984375
],
[
0.176513671875,
0.8466796875,
0.82177734375,
0.8466796875,
0.82177734375,
0.85986328125,
0.176513671875,
0.85986328125
],
[
0.176513671875,
0.85986328125,
0.82177734375,
0.85986328125,
0.82177734375,
0.875,
0.176513671875,
0.875
],
[
0.176513671875,
0.875,
0.82275390625,
0.875,
0.82275390625,
0.8896484375,
0.176513671875,
0.8896484375
],
[
0.1968994140625,
0.8994140625,
0.200927734375,
0.8994140625,
0.200927734375,
0.90478515625,
0.1968994140625,
0.90478515625
],
[
0.1895751953125,
0.912109375,
0.80126953125,
0.912109375,
0.80126953125,
0.92529296875,
0.1895751953125,
0.92529296875
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.9609375,
0.49267578125,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2023",
"The optimization problem (2) first determines a“best\" mapping based on the initial",
"model, and then updates the model to fit this mapping. The two steps are iterated until no",
"improvementcanbemade.但是，这种接地策略可能很容易陷人局部最优。原因是，每当",
"实现可行的映射h时，h往往会号导神经网络（过度）拟合自已。因为映射是确定性的和离",
"散的，所以没有平滑的路由到替代的可行映射，这将进一步改善网络。网络训练和符号推理",
"之间的信息交换不足，使得符号接地的成功高度依赖于初始模型的质量",
"在这项工作中，我们建议软化符号接地，以促进神经感知和符号推理之间的相互作用。精心",
"地说，我们不是直接搜索确定性映射h，而是首先追求h的最佳概率分布，然后逐渐“锐",
"化”该分布以获得最终的确定性h。形式上，对于每个输人，我们在Su上引人玻尔兹",
"曼分布QΦ，由Φ参数化，以指示满足符号约束的每个可行状态的概率。然后，可以将软",
"化的符号接地问题表述如下：",
"N",
"minl(o,d) := _",
" Qs(zi) log Pe(zi[ri) + Qs(zi) log Qi(zi)",
"i-lziESy",
"(P)",
"s.t.supp(Q) CSu,i=l,...,N",
"其中supp(QΦ）表示QΦ的支持。引人炳项乙EsQ(z)logQ(z）来控制Q（MacKay",
"etal.,2003）的锐度，其系数的减小确保接地可以收敛到确定性映射h。除了=1直",
"接将Pe和QΦ之间的KL发散减少到h的情况外，我们还研究了两种极端情况：（1）当",
"→+oo，QΦ被迫向均匀分布时，因此最小化仅旨在限制Pe对XMATHX",
"Advantages。博奔论提供了一个视角来理解为什么我们的软化策略通过模型训练和符号",
"推理之间更好的交互来改进问题（1）和（2）。问题E（1）或E（2）都可以视为纯策略",
"Stackelberg游戏。也就是说，模型训练和符号推理都被迫采取一定的行动（e。g。，在优化",
"过程中选择确定性映射h）。相反，问题E（P）可以看作是具有混合策略的Stackelberg游",
"戏，其中玩家以分布Q采取随机行动。与纯策略相比，混合策略确实提供了更多的博奔信",
"息，从而严格提高了模型训练的效用（Letchfordetal.，2014；Conitzer，2016）。1",
"此外，这种软化技术还避免了枚举，从而提高了效率。在问题E（1）中，对于每个输人",
"其相应符号z中的最小化需要在整个Sy(i。e。，枚举满足符号约束的所有可行状态)。这",
"在理论上通常是一个棘手的#P-完全问题（Arora&Barak，2009）。问题E(P）通过估计",
"QΦ的期望值来规避这种昂贵的计算，可以使用下一节中讨论的定制采样策略来有效地计算",
"3Markovchain Monte Carlo EstimateviaProjection",
"为了便于演示，在本节中，我们考虑单个数据样本。具体来说，通过删除所有样本的总和并",
"删除上标，问题E(P）可以表示为",
"min (0, d) :=Z Qs(z) og Po(2/a) + Qs(2)log Qs(2), supp(Qs) ≤ Sy. (3)",
"ZESy",
"该问题可以通过在θ中的梯度下降步骤和Φ中的最小化步骤之间交替解决。和Φ在第k",
"次送代时的更新为",
"0k+1 =0-nVel(0k,Φk),Φk+1 = argmin l(0k+1,Φ)",
"(4)",
"[supp(Q)",
"注意，当P。固定时，存在闭式解QΦ*，确保了θE(Jinetal.,2020,Theorem31）中梯度",
"下降的收敛性。有关详细信息，较低级别的问题，我。e。，minΦe(O,Φ），是严格凸的，因",
"此包含唯一的最小值：",
"[ Po(za)/z'esu Po(z'la), if zE Sy,",
"QΦ*(z) =",
"(5)",
"10",
"otherwise.",
"给定封闭形式的解QΦ*，可以通过Q*上的蒙特卡洛采样来估计损失函数l(0.Φ*）及其梯",
"度√el(O,Φ*）。剩下的问题是如何对Qs*进行采样，由于Su的结构未知，因此具有挑战",
"性。现有方法通常从整个符号/状态空间Z中采样，然后要么拒绝状态zSy（e。g。，策",
"-",
"Conitzer (2016) provides a nice illustrative example, which one may consult for more details.",
"3"
] |
{
"arxiv_id": "2403.00323",
"language": "zh",
"page_idx": 2
}
|
||
2403.00323-zh-3
|
2403.00323-zh-3.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
$\begin{array}{l}\text{略梯度法(Williams,1992)),或者将不可行状态 }z\text{ 投影到 }S_{\boldsymbol{y}}\mathrm{~(e。g。,~反向搜索方法~(Li}\\\mathrm{et~al..~2020)~)。不幸的是，~这些方決受到~sparsitv~problem~,~i.~e.~可行的~}z\text{ 在 }z\text{ 中非常}\end{array}$
稀疏，导致策略梯度消失，反向搜索失败
为了克服稀疏性问题，我们建议直接从符号约束$\mathcal{S_{y}}$ (i。e。,解空间)。通过应用 Metropolis 算法$\overline{\mathrm{E}}($Bhanot, 1988; Beichl & Sullivan, 2000) , 从一个可行状态 $z$ 跳到另一个可行状态$z^{\prime}$的接受率不会消失，并且可以计算为
$$\tau=\frac{Q_{\phi^*}(z')}{Q_{\phi^*}(z)}=\left(\frac{P_{\boldsymbol{\theta}}(\boldsymbol{z}'|\boldsymbol{x})}{P_{\boldsymbol{\theta}}(\boldsymbol{z}|\boldsymbol{x})}\right)^{\frac{1}{\gamma}}.$$
(6)
因此问题变为：(1)如何生成初始状态$z$ ,以及 (2)如何从 XMATHX4 跳转。对于前者一种自然的方法是利用 SMT 求解器$\overline{\mathrm{E}}($Moura & Bjørner, 2008)。$^2$ 对于后者，最常用的策略是通过随机游走实现新状态$\overline{\mathrm{E}}($Sherlock et al.,2010)。但是，在解空间中缺乏系统的随机游走方法，因为解空间很可能是未连接的E(Wigderson, 2019) ,从而产生了所谓的连接障碍。
<FigureHere>
Figure 2: Sampling unconnected solutions via projection. For our running example, $\begin{array}{l}\mathrm{we~use~the~projection~}\Pi([\boldsymbol{z}_1;\ldots;\boldsymbol{z}_7])&=\\[\boldsymbol{z}_2;\ldots;\boldsymbol{z}_6]&,\mathrm{i.e.,~dropping~the~first~and}\end{array}$ the last digits. The current state $z=$ $[4;\times;9;+;3;+;3]$is projected to $\Pi(\boldsymbol z)=$ $[ \times ; 9; + ; 3; + ; ]$ . We then randomly se $\operatorname*{lect~an~individual~component,~say,~}9$,and update it to‘8’. Next, a new feasible ${\mathrm{state~(e.g.,~4~\times~8~+~3~+~7~=~42)~is~de-}}$ rived by computing the inverse projection of$\Pi ( \boldsymbol{z}^{i})$ = [ $\bar{\times } ; \bar{8} ; + ; 3; + ; ]$ with an SMT solver.
受EFeng et al. ( 2021a) 的启发，我们建议通过投影技术来克服连接障碍。详细地，我们引$\widetilde {\text{人 了 一 个 投 影 算 子 }\Pi }( \cdot ) : \mathcal{Z} \to \Omega $,该算子将状态空间$z$映射到一个低维空间$\Omega$ ,然后在$\Omega$中应用单站点 Metropolis 算法。投影本质上压缩了状态空间，从而显著提高了解空间的连通性。图 2说明了关键思想。考虑图 1中的运行示例，其中$S_{\boldsymbol{y}}$要求所有表达式都计算为 42。SMT 求解器，以及标准的单站点 Metropolis(a。k。a。吉布斯大都会) (Metropolis et al., 1953; Bai, 2009) , 可以很容易地导出一个初始状态 (e。g。, 4??????????????) 满足符号约束，但由于连通性障碍而无法进一步探索其他可行状态$\overline{\mathrm{E}}(\bar{\mathrm{Ermon~et~al.,2012)}}$。$^3$相比之下，在低维空间$\Omega$中，跳转到另一个可行状态要容易得多。
# 4 Algorithm and Analysis
我们的神经符号学习的整体算法如 Alg 所示。1。简而言之，我们首先进行一些随机游走步骤，以在分布$Q_{\boldsymbol{\phi}^*}$上对新状态$z$进行采样；然后，我们根据$z$估计梯度，并进行一个随机梯度下降步骤。如$\operatorname{E}$Feng et al.(2021a)所示，在某些适当的假设下，Metropolis 算法在$\begin{array}{l}\text{投影空间}\mathbb{E}(\mathrm{Levin~\&~Peres,~2017)~上具有快速混合的特性。因此，我们可以有效地在}\\\text{上构建近似采样，而无需在 Metropolis 算法中采取太多步骤。此外，对于不同的示例，采}\end{array}$ 样方法和 SMT 求解器都可以并行，因此我们的框架很好地支持了批处理梯度下降
收敛。在理想情况下，梯度估计$\hat{\nabla}\ell(\boldsymbol{\theta})$是无偏的，并且$\boldsymbol{\theta}$中的梯度下降 (i。e。, 网络参数) 可以收敛。但是，由于以下原因而引入了偏差：1) Metropolis 算法$\overline{\mathrm{E}}($Jacob et al., 2017)的近似采样；2) SMT 求解器$\mathbb{E}($Moura & Bjørner, 2008) 实现的逆投影。对于前者，我们必
$^2$Current SMT solvers are mainly designed for the satisfaction problem, namely, they are efficient
in finding a solution, but underperform in generating all solutions.
$^{3}$To maintain stability, the single-site Metropolis conducts random walk in a component-wise way. That is, in each iteration, it randomly selects and updates an individual component in the current state $\boldsymbol{z}$ to generate a new state $\boldsymbol z^\prime.$ However, it can be observed that the update of any $\bar{\text{individual component (i.e.,“4’,“×’,“9”,‘+’,“3”,‘+’,‘3’) will result in no feasible new states.}}$
4
|
[
[
0.176513671875,
0.034088134765625,
0.5087890625,
0.034088134765625,
0.5087890625,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.1748046875,
0.10418701171875,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11932373046875,
0.1748046875,
0.11871337890625
],
[
0.17236328125,
0.11553955078125,
0.82421875,
0.11614990234375,
0.82421875,
0.1337890625,
0.17236328125,
0.1331787109375
],
[
0.176513671875,
0.1319580078125,
0.4697265625,
0.1319580078125,
0.4697265625,
0.1458740234375,
0.176513671875,
0.1458740234375
],
[
0.17724609375,
0.154052734375,
0.82421875,
0.154052734375,
0.82421875,
0.1673583984375,
0.17724609375,
0.1673583984375
],
[
0.176513671875,
0.1673583984375,
0.82275390625,
0.1673583984375,
0.82275390625,
0.1817626953125,
0.176513671875,
0.1817626953125
],
[
0.1748046875,
0.1817626953125,
0.45751953125,
0.1817626953125,
0.45751953125,
0.1956787109375,
0.1748046875,
0.1956787109375
],
[
0.416748046875,
0.205810546875,
0.60302734375,
0.2052001953125,
0.60302734375,
0.222900390625,
0.416748046875,
0.2235107421875
],
[
0.3896484375,
0.2197265625,
0.41259765625,
0.2197265625,
0.41259765625,
0.2279052734375,
0.3896484375,
0.2279052734375
],
[
0.80224609375,
0.2171630859375,
0.82421875,
0.2171630859375,
0.82421875,
0.2298583984375,
0.80224609375,
0.2298583984375
],
[
0.420654296875,
0.2247314453125,
0.5908203125,
0.2247314453125,
0.5908203125,
0.2392578125,
0.420654296875,
0.2392578125
],
[
0.1732177734375,
0.24365234375,
0.82275390625,
0.24560546875,
0.82275390625,
0.26318359375,
0.1732177734375,
0.261474609375
],
[
0.1895751953125,
0.261474609375,
0.82275390625,
0.261474609375,
0.82275390625,
0.274658203125,
0.1895751953125,
0.274658203125
],
[
0.176513671875,
0.274658203125,
0.82177734375,
0.274658203125,
0.82177734375,
0.28857421875,
0.176513671875,
0.28857421875
],
[
0.1748046875,
0.287353515625,
0.82275390625,
0.28857421875,
0.82275390625,
0.302978515625,
0.1748046875,
0.3017578125
],
[
0.176513671875,
0.302978515625,
0.1993408203125,
0.302978515625,
0.1993408203125,
0.314453125,
0.176513671875,
0.314453125
],
[
0.50732421875,
0.333984375,
0.8056640625,
0.333984375,
0.8056640625,
0.347900390625,
0.50732421875,
0.347900390625
],
[
0.301513671875,
0.343505859375,
0.343017578125,
0.343505859375,
0.343017578125,
0.35107421875,
0.301513671875,
0.35107421875
],
[
0.34228515625,
0.34521484375,
0.36767578125,
0.34521484375,
0.36767578125,
0.349853515625,
0.34228515625,
0.349853515625
],
[
0.298095703125,
0.3515625,
0.408447265625,
0.35107421875,
0.408447265625,
0.3623046875,
0.298095703125,
0.363037109375
],
[
0.50732421875,
0.347900390625,
0.8056640625,
0.347900390625,
0.8056640625,
0.361083984375,
0.50732421875,
0.361083984375
],
[
0.5087890625,
0.361083984375,
0.78369140625,
0.361083984375,
0.78369140625,
0.375,
0.5087890625,
0.375
],
[
0.240234375,
0.367431640625,
0.326904296875,
0.367431640625,
0.326904296875,
0.37890625,
0.240234375,
0.37890625
],
[
0.78173828125,
0.3662109375,
0.79150390625,
0.363037109375,
0.7958984375,
0.369384765625,
0.7861328125,
0.37255859375
],
[
0.25732421875,
0.380615234375,
0.26953125,
0.380615234375,
0.26953125,
0.39013671875,
0.25732421875,
0.39013671875
],
[
0.5087890625,
0.375,
0.80615234375,
0.373779296875,
0.80615234375,
0.38818359375,
0.5087890625,
0.389404296875
],
[
0.381591796875,
0.389404296875,
0.471435546875,
0.389404296875,
0.471435546875,
0.40087890625,
0.381591796875,
0.40087890625
],
[
0.50732421875,
0.386474609375,
0.7890625,
0.386962890625,
0.7890625,
0.401611328125,
0.50732421875,
0.40087890625
],
[
0.7861328125,
0.39453125,
0.79345703125,
0.39453125,
0.79345703125,
0.397705078125,
0.7861328125,
0.397705078125
],
[
0.416748046875,
0.40478515625,
0.4306640625,
0.40478515625,
0.4306640625,
0.4140625,
0.416748046875,
0.4140625
],
[
0.5107421875,
0.40087890625,
0.8056640625,
0.40087890625,
0.8056640625,
0.415283203125,
0.5107421875,
0.415283203125
],
[
0.51318359375,
0.417236328125,
0.63818359375,
0.417236328125,
0.63818359375,
0.427978515625,
0.51318359375,
0.427978515625
],
[
0.63232421875,
0.4140625,
0.80224609375,
0.415283203125,
0.80224609375,
0.4287109375,
0.63232421875,
0.427978515625
],
[
0.233642578125,
0.4287109375,
0.279296875,
0.434326171875,
0.27685546875,
0.4443359375,
0.2303466796875,
0.439453125
],
[
0.3896484375,
0.4267578125,
0.47216796875,
0.4267578125,
0.47216796875,
0.438232421875,
0.3896484375,
0.438232421875
],
[
0.50732421875,
0.4287109375,
0.80615234375,
0.4287109375,
0.80615234375,
0.442626953125,
0.50732421875,
0.442626953125
],
[
0.38818359375,
0.4375,
0.431396484375,
0.4375,
0.431396484375,
0.447509765625,
0.38818359375,
0.447509765625
],
[
0.50732421875,
0.442626953125,
0.80615234375,
0.442626953125,
0.80615234375,
0.455810546875,
0.50732421875,
0.455810546875
],
[
0.41015625,
0.460205078125,
0.42724609375,
0.460205078125,
0.42724609375,
0.468505859375,
0.41015625,
0.468505859375
],
[
0.5048828125,
0.455810546875,
0.8056640625,
0.455078125,
0.8056640625,
0.4697265625,
0.5048828125,
0.47021484375
],
[
0.404296875,
0.47021484375,
0.461669921875,
0.47021484375,
0.461669921875,
0.48095703125,
0.404296875,
0.48095703125
],
[
0.50732421875,
0.4697265625,
0.8056640625,
0.4697265625,
0.8056640625,
0.48291015625,
0.50732421875,
0.48291015625
],
[
0.2034912109375,
0.479248046875,
0.26708984375,
0.4765625,
0.26708984375,
0.4912109375,
0.2034912109375,
0.49365234375
],
[
0.298095703125,
0.48095703125,
0.32275390625,
0.48095703125,
0.32275390625,
0.48486328125,
0.298095703125,
0.48486328125
],
[
0.50634765625,
0.48291015625,
0.8056640625,
0.48291015625,
0.8056640625,
0.496826171875,
0.50634765625,
0.496826171875
],
[
0.298095703125,
0.4892578125,
0.36279296875,
0.4892578125,
0.36279296875,
0.496826171875,
0.298095703125,
0.496826171875
],
[
0.2115478515625,
0.496826171875,
0.21728515625,
0.496826171875,
0.21728515625,
0.501953125,
0.2115478515625,
0.501953125
],
[
0.29736328125,
0.496826171875,
0.39208984375,
0.496826171875,
0.39208984375,
0.50830078125,
0.29736328125,
0.50830078125
],
[
0.50634765625,
0.498779296875,
0.5537109375,
0.498779296875,
0.5537109375,
0.51025390625,
0.50634765625,
0.51025390625
],
[
0.1748046875,
0.54345703125,
0.8203125,
0.54345703125,
0.8203125,
0.556640625,
0.1748046875,
0.556640625
],
[
0.17724609375,
0.55810546875,
0.82177734375,
0.55810546875,
0.82177734375,
0.5712890625,
0.17724609375,
0.5712890625
],
[
0.1748046875,
0.5712890625,
0.82275390625,
0.5712890625,
0.82275390625,
0.5859375,
0.1748046875,
0.5859375
],
[
0.1748046875,
0.583984375,
0.8203125,
0.583984375,
0.8203125,
0.59765625,
0.1748046875,
0.59765625
],
[
0.17724609375,
0.599609375,
0.82421875,
0.599609375,
0.82421875,
0.61279296875,
0.17724609375,
0.61279296875
],
[
0.1748046875,
0.6123046875,
0.82275390625,
0.6123046875,
0.82275390625,
0.626953125,
0.1748046875,
0.626953125
],
[
0.176513671875,
0.62548828125,
0.82421875,
0.62548828125,
0.82421875,
0.6396484375,
0.176513671875,
0.6396484375
],
[
0.1748046875,
0.63818359375,
0.61279296875,
0.63818359375,
0.61279296875,
0.6513671875,
0.1748046875,
0.6513671875
],
[
0.1748046875,
0.6748046875,
0.41259765625,
0.6748046875,
0.41259765625,
0.68798828125,
0.1748046875,
0.68798828125
],
[
0.1748046875,
0.70263671875,
0.82275390625,
0.7041015625,
0.82275390625,
0.71826171875,
0.1748046875,
0.71728515625
],
[
0.176513671875,
0.71728515625,
0.82177734375,
0.71728515625,
0.82177734375,
0.73193359375,
0.176513671875,
0.73193359375
],
[
0.176513671875,
0.73291015625,
0.82275390625,
0.73291015625,
0.82275390625,
0.7470703125,
0.176513671875,
0.7470703125
],
[
0.1748046875,
0.74560546875,
0.81884765625,
0.7470703125,
0.81884765625,
0.76123046875,
0.1748046875,
0.76025390625
],
[
0.17724609375,
0.76025390625,
0.82275390625,
0.76025390625,
0.82275390625,
0.7734375,
0.17724609375,
0.7734375
],
[
0.176513671875,
0.7734375,
0.7607421875,
0.7734375,
0.7607421875,
0.787109375,
0.176513671875,
0.787109375
],
[
0.1748046875,
0.79443359375,
0.8203125,
0.79443359375,
0.8203125,
0.8095703125,
0.1748046875,
0.8095703125
],
[
0.1748046875,
0.81005859375,
0.82177734375,
0.81005859375,
0.82177734375,
0.82373046875,
0.1748046875,
0.82373046875
],
[
0.176513671875,
0.82373046875,
0.82275390625,
0.82373046875,
0.82275390625,
0.8369140625,
0.176513671875,
0.8369140625
],
[
0.19287109375,
0.84521484375,
0.82421875,
0.84765625,
0.82421875,
0.8623046875,
0.19287109375,
0.85986328125
],
[
0.1732177734375,
0.861328125,
0.61376953125,
0.861328125,
0.61376953125,
0.875,
0.1732177734375,
0.875
],
[
0.193603515625,
0.87255859375,
0.82421875,
0.875,
0.82421875,
0.88818359375,
0.193603515625,
0.88623046875
],
[
0.176513671875,
0.88818359375,
0.82177734375,
0.88818359375,
0.82177734375,
0.8984375,
0.176513671875,
0.8984375
],
[
0.176513671875,
0.8984375,
0.82177734375,
0.8984375,
0.82177734375,
0.912109375,
0.176513671875,
0.912109375
],
[
0.1748046875,
0.9111328125,
0.80810546875,
0.9111328125,
0.80810546875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.49267578125,
0.94970703125,
0.50732421875,
0.94970703125,
0.50732421875,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2023",
"略梯度法（Williams，1992）），或者将不可行状态投影到Su（e。g。，反向搜索方法（Li",
"et al，2020)）。不幸的是，这些方法受到sparsity problem，1。e。，可行的z在中非常",
"稀疏，导致策略梯度消失，反向搜索失败",
"为了克服稀疏性问题，我们建议直接从符号约束Sy（i。e。，解空间）。通过应用Metropolis",
"算法EBhanot，1988;Beichl&Sullivan，2000），从一个可行状态z跳到另一个可行状态",
"之的接受率不会消失，并且可以计算为",
"Qs*(z\")_(Pe(zlα))",
"T=",
"(6)",
"Q*(z)(Po(zα) )",
"因此问题变为：（1）如何生成初始状态，以及（2）如何从XMATHX4跳转。对于前者",
"种自然的方法是利用SMT求解器E(Moura&Bjorner，2008）。2对于后者，最常用的策",
"略是通过随机游走实现新状态E(Sherlocketal.，2010）。但是，在解空间中缺乏系统的随机",
"游走方法，因为解空间很可能是未连接的E（Wigderson，2019），从而产生了所谓的连接障",
"碍。",
"Figure 2: Sampling unconnected solutions",
"Unconnec",
"tofso",
"in high-dimension space Z",
"via projection. For our running example,",
"we use the projection II([zi;...;, z7])",
"2 = [4;x:9;+:3;+;3]",
"",
"★",
"[z2;...;z6] , i.e., dropping the first and",
"2 = [4;x:8:+;3+;7]",
"the last digits.The current state z",
"",
"★",
"[4; x;9; +;3; +;3] is projected to I1(z) =",
"X;9;+;3;+:.",
"Wethenrandomly se",
"Projection",
"Inversq projection",
"iect an individual component, say, '9', and",
"via SMT",
"update it to \"8\"Next, a new feasible",
"",
"state (e.g., 4 × 8 + 3 + 7= 42) is de-",
"[x:8;+;3;+:]",
"rived by computing the inverse projection",
"[x:9:+3;+:",
"",
"of II(z') =[x;8; +;3; +:] with an SMT",
"Randomwalkin",
"X",
"low-dimension space n",
"solver.",
"受EFengetal.（2021a）的启发，我们建议通过投影技术来克服连接障碍。详细地，我们引",
"人了一个投影算子II()：2→2，该算子将状态空间2映射到一个低维空间2，然后在",
"2中应用单站点Metropolis算法。投影本质上压缩了状态空间，从而显著提高了解空间的",
"连通性。图2说明了关键思想。考虑图1中的运行示例，其中S要求所有表达式都计算",
"为42。SMT求解器，以及标准的单站点Metropolis(a。k。a。吉布斯大都会）（Metropolis",
"etal.，1953；Bai，2009），可以很容易地导出一个初始状态（e。g。，4????????????）满足符",
"号约束，但由于连通性障碍而无法进一步探索其他可行状态E(Ermonetal.，2012）。3相比",
"之下，在低维空间中，跳转到另一个可行状态要容易得多。",
"4Algorithm and Analysis",
"我们的神经符号学习的整体算法如Alg所示。1。简而言之，我们首先进行一些随机游走",
"步骤，以在分布QΦ*上对新状态进行采样；然后，我们根据z估计梯度，并进行一个随",
"机梯度下降步骤。如EFengetal.（2021a）所示，在某些适当的假设下，Metropolis算法在",
"投影空间E(Levin&Peres，2017）上具有快速混合的特性。因此，我们可以有效地在Qp",
"上构建近似采样，而无需在Metropolis算法中采取太多步骤。此外，对于不同的示例，采",
"样方法和SMT求解器都可以并行，因此我们的框架很好地支持了批处理梯度下降",
"收敛。在理想情况下，梯度估计立e(の)是无偏的，并且θ中的梯度下降（i。e。，网络参数）",
"可以收敛。但是，由于以下原因而引人了偏差：1）Metropolis算法E(Jacobetal.,2017）的",
"近似采样；2）SMT求解器EMoura&Bjorner，2008）实现的逆投影。对于前者，我们必",
"2Current SMT solvers are mainly designed for the satisfaction problem, namely, they are efficient",
"in finding a solution, but underperform in generating all solutions.",
"3To maintain stability, the single-site Metropolis conducts random walk in a component-wise",
"way.That is, in each iteration, it randomly selects and updates an individual component in the",
"current state z to generate a new state z': However, it can be observed that the update of any",
"individual component (i.e., ‘4', × \", \"9', *+, ‘3',+, ‘3\") will result in no feasible new states.",
"4"
] |
{
"arxiv_id": "2403.00323",
"language": "zh",
"page_idx": 3
}
|
||
2403.00323-zh-4
|
2403.00323-zh-4.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
<table>
<tbody>
<tr>
<td>$\log C$</td>
<td>$m$ 11Neural Svmbolic Learning Procedure</td>
</tr>
</tbody>
</table>
须增加算法中的内部迭代次数，或者考虑 Metropolis 算法的自适应变体。对于后者，我们可以在训练过程中改变投影算子，或者增加投影空间的维度。然而，这些方法都不能完全避免梯度估计的偏差。为此，我们提供了具有有限偏差的随机梯度下降的收敛结果
Proposition 1. Assume the loss function $\ell(\boldsymbol{\theta})$ is $L$ -Lipschitz and $\ell$ -smooth, and let the actual sampling distribution be $\widehat{Q}.$ Then, if the total variation distance $d_\mathrm{tv}(\widehat{Q},Q^{*})$ is bounded by $\epsilon$ ,it holds after $K$ steps of the stochastic gradient descent with learning rate $\eta=\alpha/(\sqrt T+1)$
$$\dfrac{1}{K}\sum\limits_{k=1}^{K}\|\nabla\ell(\boldsymbol{\theta}_{k})\|^{2}\leq\mathcal{O}\left(\dfrac{\ell\sigma^{2}+\Delta_{0}}{\alpha\sqrt{T+1}}\right)+(n\epsilon L)^{2},$$
where $\Delta_0=\ell(\boldsymbol{\theta}_0)-\min\ell(\boldsymbol{\theta})$ , and $n$ is the cardinal number of supp$( Q^* )$ .
Remarks 。附录 A.1 中给出了证明。该命题指出，具有 MCMC 梯度估计的随机梯度下降收敛到近似的平稳点。此外，由于γ的减小缩小了$Q^*$ 的支持度，因此在训练过程中逐渐消除了偏差项，从而使梯度估计最终与真实的
现有方法的泛化。现有的神经符号学习框架，即。语义丢失$\mathbb{E}($Xu et al.,2018) ,deepproblog$\boxed{\mathrm{F}}$ (Manhaeve et al., 2018) 和神经语法符号学习E(Li et al., 2020) 可以理解为我们框架的特殊情况。
Proposition 2. All the three frameworks (semantic loss, deepproblog, and neural-grammar-
symbolic learning) share the same loss function
$$\hat\ell(\boldsymbol\theta):=-\sum_{i=1}^N\sum_{\boldsymbol z^i\in\mathcal S_y^i}\log P_{\boldsymbol\theta}(\boldsymbol z^i|\boldsymbol x^i),$$
and they are equivalent to Problem $(\begin{array}{c}\mathrm{P}\end{array})$with a fxed $\gamma(\begin{array}{c}\gamma=1\end{array}).\begin{array}{c}\mathrm{Here,~the~equivalence}\\\mathrm{Here,~the~equivalence}\end{array}$ means that the problems have the same optimal solution and gradient descent dynamics. Remarks 。证明见附录$\color{red}{\mathrm{A.2}}$ 。与最小化$\hat{\ell}$相比，我们的框架具有两个优点：(i)明确表示了 Boltzmann 分布$Q$ ,即使在状态空间很大的情况下，也可以使采样易于处理且易于实现；(ii) $\gamma$的退火策略在很大程度上减轻了对初始点的敏感性，从而引导了更好的最优解退火策略。接下来，我们讨论$\gamma$的递减策略。通过将$\phi_z^*=-\log P_\theta(z|x)$设置为每个状态$\boldsymbol{z}$ $\underline {{\mathrm{E} }}($Thomas & Joy, 2006) 的熵，我们可以得到
(7)
$$Q_{\phi^*}(\boldsymbol{z})=\frac{\exp(-\phi_{\boldsymbol{z}}^*/\gamma)}{\sum_{\boldsymbol{z}^{\prime}\in\mathcal{S}_{\boldsymbol{y}}}\exp(-\phi_{\boldsymbol{z}}^*/\gamma)}.$$
5
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.1748046875,
0.10418701171875,
0.53173828125,
0.10479736328125,
0.53173828125,
0.11932373046875,
0.1748046875,
0.11871337890625
],
[
0.191162109375,
0.1256103515625,
0.36279296875,
0.1256103515625,
0.36279296875,
0.1363525390625,
0.191162109375,
0.1363525390625
],
[
0.19287109375,
0.1383056640625,
0.8056640625,
0.1383056640625,
0.8056640625,
0.1514892578125,
0.19287109375,
0.1514892578125
],
[
0.188720703125,
0.15087890625,
0.35693359375,
0.15087890625,
0.35693359375,
0.1641845703125,
0.188720703125,
0.1641845703125
],
[
0.2091064453125,
0.1654052734375,
0.69921875,
0.1654052734375,
0.69921875,
0.179931640625,
0.2091064453125,
0.179931640625
],
[
0.20751953125,
0.1787109375,
0.369384765625,
0.1787109375,
0.369384765625,
0.1925048828125,
0.20751953125,
0.1925048828125
],
[
0.2279052734375,
0.1925048828125,
0.370849609375,
0.1925048828125,
0.370849609375,
0.205810546875,
0.2279052734375,
0.205810546875
],
[
0.22216796875,
0.203857421875,
0.5087890625,
0.2052001953125,
0.5087890625,
0.2197265625,
0.22216796875,
0.2183837890625
],
[
0.223876953125,
0.2197265625,
0.69287109375,
0.2197265625,
0.69287109375,
0.23291015625,
0.223876953125,
0.23291015625
],
[
0.223876953125,
0.23291015625,
0.80810546875,
0.23291015625,
0.80810546875,
0.246826171875,
0.223876953125,
0.246826171875
],
[
0.2296142578125,
0.248779296875,
0.40283203125,
0.248779296875,
0.40283203125,
0.2587890625,
0.2296142578125,
0.2587890625
],
[
0.223876953125,
0.26318359375,
0.64453125,
0.26318359375,
0.64453125,
0.27783203125,
0.223876953125,
0.27783203125
],
[
0.223876953125,
0.279052734375,
0.5537109375,
0.279052734375,
0.5537109375,
0.292236328125,
0.223876953125,
0.292236328125
],
[
0.22216796875,
0.290283203125,
0.471435546875,
0.292236328125,
0.471435546875,
0.306884765625,
0.22216796875,
0.304931640625
],
[
0.2091064453125,
0.30615234375,
0.259033203125,
0.30615234375,
0.259033203125,
0.317626953125,
0.2091064453125,
0.317626953125
],
[
0.2099609375,
0.318115234375,
0.346435546875,
0.318115234375,
0.346435546875,
0.33203125,
0.2099609375,
0.33203125
],
[
0.20751953125,
0.333251953125,
0.5556640625,
0.333251953125,
0.5556640625,
0.347900390625,
0.20751953125,
0.347900390625
],
[
0.20751953125,
0.348388671875,
0.66650390625,
0.348388671875,
0.66650390625,
0.3623046875,
0.20751953125,
0.3623046875
],
[
0.20751953125,
0.3623046875,
0.398681640625,
0.3642578125,
0.398681640625,
0.375,
0.20751953125,
0.37255859375
],
[
0.1871337890625,
0.376953125,
0.24267578125,
0.376953125,
0.24267578125,
0.38818359375,
0.1871337890625,
0.38818359375
],
[
0.1748046875,
0.40478515625,
0.82177734375,
0.406005859375,
0.82177734375,
0.42041015625,
0.1748046875,
0.419189453125
],
[
0.17724609375,
0.419189453125,
0.82275390625,
0.419189453125,
0.82275390625,
0.43310546875,
0.17724609375,
0.43310546875
],
[
0.176513671875,
0.434326171875,
0.74609375,
0.434326171875,
0.74609375,
0.447509765625,
0.176513671875,
0.447509765625
],
[
0.1748046875,
0.451904296875,
0.82421875,
0.451904296875,
0.82421875,
0.46533203125,
0.1748046875,
0.46533203125
],
[
0.1748046875,
0.46728515625,
0.82177734375,
0.468505859375,
0.82177734375,
0.48291015625,
0.1748046875,
0.481689453125
],
[
0.1748046875,
0.48291015625,
0.82275390625,
0.484130859375,
0.82275390625,
0.498779296875,
0.1748046875,
0.49755859375
],
[
0.338134765625,
0.51025390625,
0.5009765625,
0.517578125,
0.5,
0.54345703125,
0.336669921875,
0.5361328125
],
[
0.36279296875,
0.51513671875,
0.374267578125,
0.51513671875,
0.374267578125,
0.51806640625,
0.36279296875,
0.51806640625
],
[
0.50732421875,
0.51318359375,
0.5849609375,
0.51513671875,
0.5849609375,
0.52978515625,
0.50732421875,
0.52783203125
],
[
0.5888671875,
0.521484375,
0.66015625,
0.521484375,
0.66015625,
0.5361328125,
0.5888671875,
0.5361328125
],
[
0.339111328125,
0.53076171875,
0.365234375,
0.53076171875,
0.365234375,
0.54443359375,
0.339111328125,
0.54443359375
],
[
0.5185546875,
0.53271484375,
0.576171875,
0.53271484375,
0.576171875,
0.54345703125,
0.5185546875,
0.54345703125
],
[
0.176513671875,
0.55615234375,
0.70361328125,
0.55615234375,
0.70361328125,
0.5693359375,
0.176513671875,
0.5693359375
],
[
0.1748046875,
0.580078125,
0.8203125,
0.580078125,
0.8203125,
0.5947265625,
0.1748046875,
0.5947265625
],
[
0.176513671875,
0.59521484375,
0.82275390625,
0.59521484375,
0.82275390625,
0.609375,
0.176513671875,
0.609375
],
[
0.176513671875,
0.60986328125,
0.48779296875,
0.60986328125,
0.48779296875,
0.62353515625,
0.176513671875,
0.62353515625
],
[
0.176513671875,
0.62890625,
0.82421875,
0.62890625,
0.82421875,
0.64306640625,
0.176513671875,
0.64306640625
],
[
0.176513671875,
0.64306640625,
0.82275390625,
0.64306640625,
0.82275390625,
0.6572265625,
0.176513671875,
0.6572265625
],
[
0.17724609375,
0.65771484375,
0.2156982421875,
0.65771484375,
0.2156982421875,
0.66943359375,
0.17724609375,
0.66943359375
],
[
0.1732177734375,
0.6748046875,
0.82275390625,
0.67529296875,
0.82275390625,
0.68994140625,
0.1732177734375,
0.689453125
],
[
0.1748046875,
0.689453125,
0.515625,
0.689453125,
0.515625,
0.70263671875,
0.1748046875,
0.70263671875
],
[
0.3798828125,
0.72021484375,
0.61767578125,
0.72021484375,
0.61767578125,
0.7412109375,
0.3798828125,
0.7412109375
],
[
0.4599609375,
0.71533203125,
0.471435546875,
0.71533203125,
0.471435546875,
0.7197265625,
0.4599609375,
0.7197265625
],
[
0.451904296875,
0.74072265625,
0.51708984375,
0.7421875,
0.51708984375,
0.7568359375,
0.451904296875,
0.7548828125
],
[
0.1748046875,
0.7646484375,
0.82421875,
0.7666015625,
0.82421875,
0.77978515625,
0.1748046875,
0.77783203125
],
[
0.1748046875,
0.78076171875,
0.810546875,
0.78076171875,
0.810546875,
0.79443359375,
0.1748046875,
0.79443359375
],
[
0.1748046875,
0.806640625,
0.82421875,
0.80810546875,
0.82421875,
0.82275390625,
0.1748046875,
0.8212890625
],
[
0.1732177734375,
0.82080078125,
0.826171875,
0.818359375,
0.826171875,
0.8359375,
0.1732177734375,
0.83837890625
],
[
0.176513671875,
0.8359375,
0.810546875,
0.8359375,
0.810546875,
0.849609375,
0.176513671875,
0.849609375
],
[
0.176513671875,
0.85693359375,
0.82421875,
0.85693359375,
0.82421875,
0.87060546875,
0.176513671875,
0.87060546875
],
[
0.1748046875,
0.869140625,
0.521484375,
0.87060546875,
0.521484375,
0.8837890625,
0.1748046875,
0.88330078125
],
[
0.488525390625,
0.89599609375,
0.57666015625,
0.89599609375,
0.57666015625,
0.9072265625,
0.488525390625,
0.9072265625
],
[
0.3857421875,
0.90283203125,
0.459228515625,
0.90283203125,
0.459228515625,
0.91748046875,
0.3857421875,
0.91748046875
],
[
0.80126953125,
0.90087890625,
0.826171875,
0.90087890625,
0.826171875,
0.91845703125,
0.80126953125,
0.91845703125
],
[
0.4599609375,
0.9091796875,
0.609375,
0.9111328125,
0.609375,
0.9296875,
0.4599609375,
0.92822265625
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2023",
"Algorithm 1Neural Symbolic Learning Procedure",
"Set an initial value of ",
"Calculate an initial state z E Su for each input-output pair (α,y) via the SMT solver.",
"for k = 0,1,...,K do",
"Randomly draw an example (c, y) from training data [(ci, yi))-1 :",
"for t=O,l,...,T do",
"Generate new state",
"Compute the projection u =II(z) E .",
"Obtain u' by randomly selecting and updating a component of u",
"Calculate the inverse projection z' = II-1(u') via the SMT solver, as a new state.",
"Accept/rejectnewstate",
"Calculate the acceptance ratio T = (Pe(zα)/Pe(zα))1/",
"Generate a uniform random number v E [0, 1]",
"Update the state z to z' if v ≤T .",
"endfor",
"Train the network",
"Estimate the gradient Vl(0) =-Velog Pe(zα)",
"Update network parameters by the stochastic gradient decent.",
"Decreasethecoefficient",
"end for",
"须增加算法中的内部迭代次数，或者考虑Metropolis算法的自适应变体。对于后者，我们",
"可以在训练过程中改变投影算子，或者增加投影空间的维度。然而，这些方法都不能完全避",
"免梯度估计的偏差。为此，我们提供了具有有限偏差的随机梯度下降的收敛结果",
"Proposition 1. Assume the loss function l(O) is L -Lipschitz and -smooth, and let the actual",
"sampling distribution be Q. Then, if the total variation distance dtv(Q, Q*) is bounded by e",
", it holds after K steps of the stochastic gradient descent with learning rate n = α/(VT + 1)",
"V(0x)12 ≤0",
"11",
"lo?+△",
"+ (neL)?",
"K",
"VT+1",
"where △o =l(Oo)-min e(0), and n is the cardinal number of supp(Q*) :",
"Remarks。附录A.1中给出了证明。该命题指出，具有MCMC梯度估计的随机梯度下降",
"收敛到近似的平稳点。此外，由于的减小缩小了Q*的支持度，因此在训练过程中逐渐消",
"除了偏差项，从而使梯度估计最终与真实的",
"现有方法的泛化。现有的神经符号学习框架，即。语义丢失E(Xuetal.，2018），deepproblogE",
"（Manhaeveetal.，2018）和神经语法符号学习E(Lietal.，2020）可以理解为我们框架的特殊",
"情况。",
"Proposition 2. All the three frameworks (semantic loss, deepproblog, and neural-grammar-",
"symbolic learning) share the same loss function",
"é(0) := - log Pe(zia),",
"N",
"i=1zies",
"and they are equivalent to Problem (P)with a fixed (=1 ). Here, the equivalence",
"means that the problems have the same optimal solution and gradient descent dynamics.",
"Remarks。证明见附录A.2。与最小化i相比，我们的框架具有两个优点：（i）明确表",
"示了Boltzmann分布＆，即使在状态空间很大的情况下，也可以使采样易宇处理且易于实",
"现：（i）的退火策略在很大程度上减轻了对初始点的敏感性，从而引导了更好的最优解",
"退火策略。接下来，我们讨论的递减策略。通过将Φ=－logPe(zα）设置为每个状态",
"z（Thomas&Joy，2006）的摘，我们可以得到",
"exp(-Φ*/)",
"Qs*(z) =",
"(7)",
"Zzes, exp(-Φ*/)",
"5"
] |
{
"arxiv_id": "2403.00323",
"language": "zh",
"page_idx": 4
}
|
||
2311.08572-zh-1
|
2311.08572-zh-1.jpg
|
$\textbf{Chenxi Whitehouse}^{1,2,*}\textbf{ Fantine Huot}^2$ Jasmijn Bastings$^2$ $\textbf{Mostafa Dehghani}^2$ $\textbf{Chu- Cheng Lin}^3$ $\textbf{Mirella Lapata}^2$ $^{1}$University of Cambridge $^2$Google DeepMind$^3$Google
chenxi.whitehouse@cl.cam.ac.uk
$\{$ fantinehuot, bastings, dehghani, kitsing, lapata $\}$ @google.com
### $\mathbf{Abstract}$
Although the advancements of pre-trained Large Language Models have significantly accelerated recent progress in NLP, their everincreasing size poses significant challenges for conventional fine-tuning, especially in memory-intensive tasks. We investigate the potential of Parameter-Efficient Fine-Tuning, focusing on Low-Rank Adaptation (LoRA), in the domain of multilingual summarization, a task that is both challenging (due to typically long inputs), and relatively unexplored. We conduct an extensive study across different data availability scenarios, including highand low-data settings, and cross-lingual transfer, leveraging models of different sizes. Our findings reveal that LoRA is competitive with full fine-tuning when trained with high quantities of data, and excels in low-data scenarios and cross-lingual transfer. We also study different strategies for few-shot cross-lingual transfer, finding that continued LoRA tuning outperforms full fine-tuning and the dynamic composition of language-specific LoRA modules.
### 1 $\textbf{Introduction}$
LLMPaLM2 (Anil et al., 2023) LLaMA2 (Touvron et al., 2023)
OpenAIGPTNLPHowever,theexpansioninthesizeofLLMposessignificantchallengesfortraditionalfine-
tuning,particularlywhenfacingwithmanydownstreamtasksortasksortaskRtand gerneHibh yfootprint,eg
PEFTPEFT(Houlsby et al., 2019; Pfeiffer et al., 2021) LoRA; Hu et al. 2022 prefix-tuning (Li and Liang, 2021) prompt-tuning (Lester et al., 2021) LoRAPEFTg GLUE (Wang et al., 2018) SuperGLUE (Wang et al., 2019) eg (Li and Liang, 2021)
LoRA (Hasan et al., 2021) PEFT(g ) (Parida and Motlicek, 2019) PEFT (Kirkpatrick et al., 2017; Mitchell et al., 2022)
iLoRAiiLoRA high-data
low-data (g En-
glish) (Artetxe et al., 2020; Karthikeyan et al.,
$^{*}$Work conducted as Research Intern at Google DeepMind.
2020) LoRALoraHub (Huang et al., 2023) XLSum (Hasan et al., 2021) XWikis (PerezBeltrachini and Lapata, 2021) PaLM2100LLM (Anil et al., 2023)
iLoRAiiLoRAiiiLoRA
### 2 $\textbf{Related Work}$
Parameter Efficient Fine-Tuning methods LoRAPEFT (Hu et al.,2022; Chen et al.,2022) (Zhang et al., 2023b; Valipour et al., 2023) (He et al., 2022; Chavan et al., 2023) LoRAPEFTLoRA (Dettmers et al., 2023) LoRA
Cross-lingual Transfer requires
(Artetxe
et al., 2020; Karthikeyan et al., 2020; Lauscher et al., 2020; Whitehouse et al., 2022, 2023a) PEFT (Pfeiffer et al., 2020; Ansell et al., 2021) (Ansell et al., 2022) Vu et al.(2022) (Lester et al., 2021) Wikilingua (Ladhak et al., 2020) LoRA(g, fewshotsettings)
Model Composition and Weight Merging aim (Lv et al., 2023) (Zhang et al., 2023a) (Sun et al., 2023) (Jin et al., 2023) LoRA (Huang et al., 2023)
## 3 LoRA for Multilingual Summarization
LoRA (Hu et al., 2022) LoRA (Huang et al., 2023)
$\textbf{LoRA}$ $W_{0}\in \tilde{\mathbb{R} } ^{d\times k}$ LLM $d^{\mathrm{~}}k$ LoRA $W_{0}+$ $\Delta W= W_0+ BAW$ $B\in \mathbb{R} ^{d\times r}$ $A\in \mathbb{R} ^{r\times k}$ $r$ $\ll$ $\min ( d, k)$ $BA$ $W_{0}$ $W_{0}$ $B$ $A$ $2r/ d$ LoRA Hu et al. (2022) Transformer self-attention (ie, query,key,value,out)
$\textbf{LoraHub}_{- N}$ Huang et al.(2023) LoRALoRA $m_i$
$\hat{m} = \sum _{i= 1}^{N}w_{i}m_{i}$ $w_{i}$ (Sun et al.,2022)
### 3.2 Data Regimes
LoRA
High Data
|
[
[
0.2493896484375,
0.1341552734375,
0.755859375,
0.1341552734375,
0.755859375,
0.1519775390625,
0.2493896484375,
0.1519775390625
],
[
0.255126953125,
0.1519775390625,
0.74462890625,
0.152587890625,
0.74462890625,
0.17041015625,
0.255126953125,
0.169189453125
],
[
0.270263671875,
0.16748046875,
0.73876953125,
0.169189453125,
0.73876953125,
0.18701171875,
0.270263671875,
0.185302734375
],
[
0.35107421875,
0.188232421875,
0.6591796875,
0.188232421875,
0.6591796875,
0.201904296875,
0.35107421875,
0.201904296875
],
[
0.18896484375,
0.204833984375,
0.8330078125,
0.204833984375,
0.8330078125,
0.2197265625,
0.18896484375,
0.2197265625
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.51318359375,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.51318359375,
0.267333984375
],
[
0.51123046875,
0.26953125,
0.8818359375,
0.26953125,
0.8818359375,
0.283935546875,
0.51123046875,
0.283935546875
],
[
0.1435546875,
0.27783203125,
0.459228515625,
0.279052734375,
0.459228515625,
0.29345703125,
0.1435546875,
0.292236328125
],
[
0.51318359375,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.300537109375,
0.51318359375,
0.300537109375
],
[
0.14697265625,
0.295654296875,
0.45751953125,
0.295654296875,
0.45751953125,
0.305908203125,
0.14697265625,
0.305908203125
],
[
0.51123046875,
0.3017578125,
0.6455078125,
0.3017578125,
0.6455078125,
0.3154296875,
0.51123046875,
0.3154296875
],
[
0.145263671875,
0.30810546875,
0.459228515625,
0.30810546875,
0.459228515625,
0.32177734375,
0.145263671875,
0.32177734375
],
[
0.145263671875,
0.322509765625,
0.45751953125,
0.322509765625,
0.45751953125,
0.336181640625,
0.145263671875,
0.336181640625
],
[
0.52880859375,
0.318359375,
0.7021484375,
0.318359375,
0.7021484375,
0.33203125,
0.52880859375,
0.33203125
],
[
0.145263671875,
0.33740234375,
0.459228515625,
0.33740234375,
0.459228515625,
0.350830078125,
0.145263671875,
0.350830078125
],
[
0.5146484375,
0.3486328125,
0.5341796875,
0.3486328125,
0.5341796875,
0.358154296875,
0.5146484375,
0.358154296875
],
[
0.53076171875,
0.3466796875,
0.6708984375,
0.3466796875,
0.6708984375,
0.361083984375,
0.53076171875,
0.361083984375
],
[
0.145263671875,
0.365234375,
0.45751953125,
0.365234375,
0.45751953125,
0.37890625,
0.145263671875,
0.37890625
],
[
0.145263671875,
0.350830078125,
0.459228515625,
0.350830078125,
0.459228515625,
0.364013671875,
0.145263671875,
0.364013671875
],
[
0.515625,
0.37353515625,
0.88330078125,
0.37353515625,
0.88330078125,
0.38720703125,
0.515625,
0.38720703125
],
[
0.145263671875,
0.380126953125,
0.45751953125,
0.380126953125,
0.45751953125,
0.393798828125,
0.145263671875,
0.393798828125
],
[
0.145263671875,
0.393798828125,
0.45751953125,
0.393798828125,
0.45751953125,
0.40673828125,
0.145263671875,
0.40673828125
],
[
0.51123046875,
0.39013671875,
0.873046875,
0.39013671875,
0.873046875,
0.40380859375,
0.51123046875,
0.40380859375
],
[
0.145263671875,
0.407958984375,
0.459228515625,
0.407958984375,
0.459228515625,
0.421630859375,
0.145263671875,
0.421630859375
],
[
0.51123046875,
0.40625,
0.8818359375,
0.40625,
0.8818359375,
0.419189453125,
0.51123046875,
0.419189453125
],
[
0.1435546875,
0.4228515625,
0.45751953125,
0.4228515625,
0.45751953125,
0.435791015625,
0.1435546875,
0.435791015625
],
[
0.509765625,
0.421630859375,
0.88330078125,
0.42041015625,
0.88330078125,
0.435302734375,
0.509765625,
0.435791015625
],
[
0.145263671875,
0.43701171875,
0.45751953125,
0.43701171875,
0.45751953125,
0.44775390625,
0.145263671875,
0.44775390625
],
[
0.51123046875,
0.438232421875,
0.78271484375,
0.438232421875,
0.78271484375,
0.451904296875,
0.51123046875,
0.451904296875
],
[
0.145263671875,
0.449462890625,
0.45751953125,
0.449462890625,
0.45751953125,
0.463134765625,
0.145263671875,
0.463134765625
],
[
0.145263671875,
0.46435546875,
0.4609375,
0.46435546875,
0.4609375,
0.47802734375,
0.145263671875,
0.47802734375
],
[
0.5263671875,
0.46435546875,
0.791015625,
0.46435546875,
0.791015625,
0.47802734375,
0.5263671875,
0.47802734375
],
[
0.8134765625,
0.4619140625,
0.8818359375,
0.46435546875,
0.8818359375,
0.478515625,
0.8134765625,
0.476806640625
],
[
0.145263671875,
0.478515625,
0.459228515625,
0.478515625,
0.459228515625,
0.4921875,
0.145263671875,
0.4921875
],
[
0.51123046875,
0.479736328125,
0.88330078125,
0.479736328125,
0.88330078125,
0.49462890625,
0.51123046875,
0.49462890625
],
[
0.145263671875,
0.4921875,
0.459228515625,
0.4921875,
0.459228515625,
0.505859375,
0.145263671875,
0.505859375
],
[
0.51123046875,
0.496337890625,
0.8818359375,
0.496337890625,
0.8818359375,
0.51025390625,
0.51123046875,
0.51025390625
],
[
0.145263671875,
0.50732421875,
0.4609375,
0.50732421875,
0.4609375,
0.52099609375,
0.145263671875,
0.52099609375
],
[
0.51123046875,
0.5126953125,
0.8818359375,
0.5126953125,
0.8818359375,
0.525390625,
0.51123046875,
0.525390625
],
[
0.145263671875,
0.521484375,
0.459228515625,
0.521484375,
0.459228515625,
0.53515625,
0.145263671875,
0.53515625
],
[
0.51318359375,
0.52783203125,
0.8818359375,
0.52783203125,
0.8818359375,
0.54150390625,
0.51318359375,
0.54150390625
],
[
0.145263671875,
0.53515625,
0.456787109375,
0.53515625,
0.456787109375,
0.548828125,
0.145263671875,
0.548828125
],
[
0.145263671875,
0.5498046875,
0.45751953125,
0.5498046875,
0.45751953125,
0.5634765625,
0.145263671875,
0.5634765625
],
[
0.51123046875,
0.54443359375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55859375,
0.51123046875,
0.55859375
],
[
0.145263671875,
0.5634765625,
0.456787109375,
0.5634765625,
0.456787109375,
0.57666015625,
0.145263671875,
0.57666015625
],
[
0.51318359375,
0.5625,
0.60791015625,
0.5625,
0.60791015625,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.145263671875,
0.580078125,
0.456787109375,
0.580078125,
0.456787109375,
0.59033203125,
0.145263671875,
0.59033203125
],
[
0.51708984375,
0.5849609375,
0.8818359375,
0.5859375,
0.8818359375,
0.60107421875,
0.51708984375,
0.599609375
],
[
0.145263671875,
0.59130859375,
0.45751953125,
0.59130859375,
0.45751953125,
0.60498046875,
0.145263671875,
0.60498046875
],
[
0.145263671875,
0.60693359375,
0.1796875,
0.60693359375,
0.1796875,
0.6181640625,
0.145263671875,
0.6181640625
],
[
0.51123046875,
0.6025390625,
0.880859375,
0.6025390625,
0.880859375,
0.6162109375,
0.51123046875,
0.6162109375
],
[
0.51123046875,
0.6181640625,
0.880859375,
0.6181640625,
0.880859375,
0.6328125,
0.51123046875,
0.6328125
],
[
0.1175537109375,
0.63427734375,
0.1318359375,
0.63427734375,
0.1318359375,
0.64453125,
0.1175537109375,
0.64453125
],
[
0.1485595703125,
0.6318359375,
0.264404296875,
0.6328125,
0.264404296875,
0.64794921875,
0.1485595703125,
0.646484375
],
[
0.51123046875,
0.64794921875,
0.8818359375,
0.64794921875,
0.8818359375,
0.662109375,
0.51123046875,
0.662109375
],
[
0.2476806640625,
0.65673828125,
0.29296875,
0.65673828125,
0.29296875,
0.67138671875,
0.2476806640625,
0.67138671875
],
[
0.43994140625,
0.65673828125,
0.48779296875,
0.65673828125,
0.48779296875,
0.67138671875,
0.43994140625,
0.67138671875
],
[
0.32568359375,
0.6591796875,
0.345947265625,
0.6591796875,
0.345947265625,
0.67138671875,
0.32568359375,
0.67138671875
],
[
0.1175537109375,
0.6591796875,
0.214111328125,
0.6591796875,
0.214111328125,
0.67041015625,
0.1175537109375,
0.67041015625
],
[
0.376220703125,
0.6591796875,
0.405517578125,
0.6591796875,
0.405517578125,
0.67138671875,
0.376220703125,
0.67138671875
],
[
0.2216796875,
0.6728515625,
0.295654296875,
0.6748046875,
0.295654296875,
0.689453125,
0.2216796875,
0.68701171875
],
[
0.43994140625,
0.6728515625,
0.489501953125,
0.6728515625,
0.489501953125,
0.68701171875,
0.43994140625,
0.68701171875
],
[
0.509765625,
0.67333984375,
0.8759765625,
0.67333984375,
0.8759765625,
0.68701171875,
0.509765625,
0.68701171875
],
[
0.328369140625,
0.6748046875,
0.349365234375,
0.6748046875,
0.349365234375,
0.68701171875,
0.328369140625,
0.68701171875
],
[
0.1175537109375,
0.6748046875,
0.189697265625,
0.6748046875,
0.189697265625,
0.68603515625,
0.1175537109375,
0.68603515625
],
[
0.378662109375,
0.6748046875,
0.4072265625,
0.6748046875,
0.4072265625,
0.68701171875,
0.378662109375,
0.68701171875
],
[
0.1158447265625,
0.689453125,
0.8564453125,
0.6904296875,
0.8564453125,
0.705078125,
0.1158447265625,
0.70361328125
],
[
0.11419677734375,
0.70361328125,
0.8193359375,
0.70361328125,
0.8193359375,
0.7216796875,
0.11419677734375,
0.7216796875
],
[
0.1351318359375,
0.72265625,
0.489501953125,
0.72265625,
0.489501953125,
0.736328125,
0.1351318359375,
0.736328125
],
[
0.51708984375,
0.7236328125,
0.568359375,
0.7236328125,
0.568359375,
0.7353515625,
0.51708984375,
0.7353515625
],
[
0.59375,
0.7197265625,
0.884765625,
0.72021484375,
0.884765625,
0.73828125,
0.59375,
0.73681640625
],
[
0.1158447265625,
0.73681640625,
0.48779296875,
0.73828125,
0.48779296875,
0.7529296875,
0.1158447265625,
0.751953125
],
[
0.5087890625,
0.736328125,
0.87744140625,
0.7353515625,
0.87744140625,
0.7529296875,
0.5087890625,
0.75439453125
],
[
0.1158447265625,
0.7548828125,
0.489501953125,
0.7548828125,
0.489501953125,
0.76953125,
0.1158447265625,
0.76953125
],
[
0.50732421875,
0.7548828125,
0.8818359375,
0.75439453125,
0.8818359375,
0.7685546875,
0.50732421875,
0.76953125
],
[
0.1175537109375,
0.77099609375,
0.491943359375,
0.77099609375,
0.491943359375,
0.78515625,
0.1175537109375,
0.78515625
],
[
0.509765625,
0.771484375,
0.8818359375,
0.771484375,
0.8818359375,
0.78515625,
0.509765625,
0.78515625
],
[
0.11248779296875,
0.78369140625,
0.490234375,
0.78515625,
0.490234375,
0.802734375,
0.11248779296875,
0.8017578125
],
[
0.509765625,
0.7880859375,
0.69287109375,
0.7880859375,
0.69287109375,
0.8017578125,
0.509765625,
0.8017578125
],
[
0.1158447265625,
0.8017578125,
0.162841796875,
0.8017578125,
0.162841796875,
0.81640625,
0.1158447265625,
0.81640625
],
[
0.51708984375,
0.81396484375,
0.59619140625,
0.81396484375,
0.59619140625,
0.826171875,
0.51708984375,
0.826171875
],
[
0.59765625,
0.8125,
0.8818359375,
0.8134765625,
0.8818359375,
0.82763671875,
0.59765625,
0.82666015625
],
[
0.1351318359375,
0.818359375,
0.48779296875,
0.818359375,
0.48779296875,
0.83203125,
0.1351318359375,
0.83203125
],
[
0.515625,
0.83056640625,
0.5615234375,
0.83056640625,
0.5615234375,
0.8427734375,
0.515625,
0.8427734375
],
[
0.54736328125,
0.83447265625,
0.5810546875,
0.81640625,
0.59765625,
0.83203125,
0.5634765625,
0.849609375
],
[
0.58837890625,
0.830078125,
0.81689453125,
0.82763671875,
0.81689453125,
0.8427734375,
0.58837890625,
0.84423828125
],
[
0.1175537109375,
0.85107421875,
0.273681640625,
0.85107421875,
0.273681640625,
0.8642578125,
0.1175537109375,
0.8642578125
],
[
0.1175537109375,
0.8349609375,
0.48779296875,
0.8349609375,
0.48779296875,
0.8486328125,
0.1175537109375,
0.8486328125
],
[
0.51123046875,
0.85693359375,
0.54150390625,
0.85693359375,
0.54150390625,
0.869140625,
0.51123046875,
0.869140625
],
[
0.55517578125,
0.85546875,
0.6708984375,
0.85693359375,
0.6708984375,
0.87158203125,
0.55517578125,
0.87060546875
],
[
0.133544921875,
0.86767578125,
0.328369140625,
0.86767578125,
0.328369140625,
0.88134765625,
0.133544921875,
0.88134765625
],
[
0.3525390625,
0.86767578125,
0.490234375,
0.86767578125,
0.490234375,
0.8818359375,
0.3525390625,
0.8818359375
],
[
0.509765625,
0.87744140625,
0.5615234375,
0.87744140625,
0.5615234375,
0.892578125,
0.509765625,
0.892578125
],
[
0.118408203125,
0.88427734375,
0.489501953125,
0.88427734375,
0.489501953125,
0.89794921875,
0.118408203125,
0.89794921875
],
[
0.1427001953125,
0.90478515625,
0.483642578125,
0.90478515625,
0.483642578125,
0.91845703125,
0.1427001953125,
0.91845703125
],
[
0.515625,
0.90380859375,
0.60205078125,
0.90380859375,
0.60205078125,
0.91845703125,
0.515625,
0.91845703125
]
] |
[
"Chenxi Whitehousel,2,* Fantine Huot? Jasmijn Bastings",
"Mostafa Dehghani?Chu-Cheng Lin3 Mirella Lapata?",
"1University of Cambridge 2Google DeepMind 3Google",
"chenxi.whitehouse@cl.cam.ac.uk",
"( fantinehuot,bastings,dehghani,kitsing, lapata J @google.com",
"Abstract",
"2020) LoRALoraHub (Huang et al.,2023)",
"XLSum (Hasan et al., 2021) XWikis (Perez-",
"Although the advancements of pre-trained",
"Beltrachini and Lapata, 2021) PaLM2100LLM",
"LargeLanguageModelshavesignificantlyac-",
"(Anil et al., 2023)",
"celerated recent progress in NLP, their ever-",
"increasing size poses significant challenges",
"iLoRAiiLoRAiiLoRA",
"for conventional fine-tuning, especially in",
"2",
"Related Work",
"potential of Parameter-Efficient Fine-Tuning,",
"memory-intensive tasks.We investigate the",
"Parameter Efficient Fine-Tuning methods Lo-",
"focusing on Low-Rank Adaptation (LoRA),",
"in the domain of multilingual summarization,",
"RAPEFT (Hu et al., 2022; Chen et al., 2022)",
"a task that is both challenging (due to typi-",
"(Zhang et al., 2023b; Valipour et al., 2023) (He",
"cally long inputs), and relatively unexplored.",
"et al., 2022; Chavan et al., 2023) LoRAPEFT-",
"We conduct anextensive study acrossdiffer-",
"LoRA (Dettmers et al., 2023) LoRA",
"ent data availability scenarios, including high-",
"and low-data settings, and cross-lingual trans-",
"Cross-lingual Transfer requires",
"(Artetxe",
"fer, leveraging models of different sizes. Our",
"et al., 2020; Karthikeyan et al., 2020; Lauscher",
"findings reveal that LoRA is competitive with",
"et al., 2020; Whitehouse et al., 2022, 2023a) PEFT",
"full fine-tuning when trained with high quan-",
"(Pfeiffer et al., 2020; Ansell et al., 2021) (Ansell",
"tities of data, and excels in low-data scenar-",
"et al., 2022) Vu et al. (2022) (Lester et al., 2021)",
"ios and cross-lingual transfer. We also study",
"different strategies for few-shot cross-lingual",
"Wikilingua (Ladhak et al., 2020) LoRA(g, few-",
"transfer, finding that continued LoRA tuning",
"shotsettings)",
"outperformsfullfine-tuningand thedynamic",
"Model Composition and Weight Merging aim",
"composition of language-specific LoRA mod-",
"ules.",
"(Lv et al., 2023) (Zhang et al., 2023a) (Sun et al.,",
"2023) (Jin et al., 2023) LoRA (Huang et al., 2023)",
"1",
"Introduction",
"3LoRA for Multilingual Summarization",
"(Anil",
"2023)",
"et",
"LLMPaLM2",
"al.,",
"(Touvron",
"2023)",
"LoRA (Hu et al., 2022) LoRA (Huang et al., 2023)",
"et",
"LLaMA2",
"al.,",
"OpenAIGPTNLPHowever,theexpansioninthesizeofLLMposessignificantchallengesfortraditionalfine-",
"tuning,particularlywhenfacingwithmanydownstreamtaskSortasrRWiafargeaHibyfootprint,eg",
"PEFTPEFT (Houlsby et al., 2019; Pfeiffer et al.,",
"LoRA",
"Wo e Rdxk LLM d k LoRA Wo +",
"2021) LoRA; Hu et al. 2022 prefix-tuning (Li",
"△W = Wo+BA W B E Rdxr AE Rrxk",
"and Liang, 2021) prompt-tuning (Lester et al.,",
"r min(d,k) BA Wo Wo B A 2r/d",
"2021) LoRAPEFTg GLUE (Wang et al., 2018) Su-",
"LoRA Hu et al. (2022) Transformer self-attention",
"perGLUE (Wang et al., 2019) eg (Li and Liang,",
"(ie, query,key,value,out)",
"2021)",
"LoraHub",
"Huang et al. (2023) LoRALoRA mi",
"LoRA (Hasan et al., 2021) PEFT(g ) (Parida and",
"m=",
"",
"-i wimi wi (Sunet al., 2022)",
"Mitchell et al., 2022)",
"Motlicek, 2019) PEFT (Kirkpatrick et al., 2017;",
"3.2",
"Data Regimes",
"iLoRAiiLoRA high-data",
"low-data (g En-",
"LoRA",
"glish) (Artetxe et al., 2020; Karthikeyan et al.,",
"*Work conducted as Research Intern at Google DeepMind.",
"High Data"
] |
{
"arxiv_id": "2311.08572",
"language": "zh",
"page_idx": 1
}
|
||
2311.08572-zh-2
|
2311.08572-zh-2.jpg
|
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>XLSUM</th>
<th>$\mathbf{XWIKIS}$</th>
</tr>
<tr>
<td>Source</td>
<td>BBCNews</td>
<td>Wikipedia</td>
</tr>
<tr>
<td>Languages</td>
<td>44</td>
<td>5</td>
</tr>
<tr>
<td>Train/Val/Test Data</td>
<td>$1.12\mathbf{M}/114\mathbf{K}/114\mathbf{K}$</td>
<td>1.43M/40K/35K</td>
</tr>
<tr>
<td>Input/Output $\textbf{t Words}$ 1</td>
<td>470.2/22.1</td>
<td>1042.7/63.7</td>
</tr>
</tbody>
</table>
Table 1:
Summary statistics for the XLSum
and XWikis multilingual summarization datasets. Train/Val/Test shows the number of examples in each split. Input/Output shows the average number of words in the English input document and output summary. XWikis has long documents and multi-sentence summaries.
<table>
<tbody>
<tr>
<th> </th>
<th rowspan="2">Params</th>
<th colspan="3">$XLSUM$</th>
<th colspan="3">$XWIKIS$</th>
</tr>
<tr>
<th> </th>
<th>$R-L$</th>
<th>$NLI$</th>
<th>SH</th>
<th>$R-L$</th>
<th>$NLI$</th>
<th>SH</th>
</tr>
<tr>
<td>Reference</td>
<td> </td>
<td> </td>
<td>48.50</td>
<td>31.65</td>
<td> </td>
<td>39.20</td>
<td>25.19</td>
</tr>
<tr>
<td>Full FT</td>
<td>$100\%$</td>
<td>31.11</td>
<td>42.93</td>
<td>31.64</td>
<td>34.08</td>
<td>41.04</td>
<td>25.19</td>
</tr>
<tr>
<td>FT-Att</td>
<td>$20\%$</td>
<td>30.88</td>
<td>50.32</td>
<td>36.12</td>
<td>32.22</td>
<td>37.06</td>
<td>24.2C</td>
</tr>
<tr>
<td>LoRA- 512</td>
<td>13.3 $\%$</td>
<td>29.81</td>
<td>42.58</td>
<td>30.16</td>
<td>33.38</td>
<td>40.48</td>
<td>24.78</td>
</tr>
<tr>
<td>LoRA- 64</td>
<td>$1.7\%$</td>
<td>29.79</td>
<td>45.51</td>
<td>31.80</td>
<td>34.04</td>
<td>45.34</td>
<td>27.02</td>
</tr>
<tr>
<td>LoRA- 16</td>
<td>$0.4\%$</td>
<td>29.77</td>
<td>48.48</td>
<td>33.25</td>
<td>33.80</td>
<td>46.10</td>
<td>27.42</td>
</tr>
<tr>
<td>LoRA- 4</td>
<td>$0.1\%$</td>
<td>29.03</td>
<td>51.16</td>
<td>34.42</td>
<td>32.92</td>
<td>47.43</td>
<td>27.72</td>
</tr>
</tbody>
</table>
Table 2: Results on XLSum and XWikis with PaLM 2-XXS trained in the high-data regime: full fine-tuning on all layers (Full FT), on attention layers (FT-Att), and LoRA-$^*($with different ranks). Params denotes the proportion of trainable parameters. Best ROUGE-L (R-L), NLI, and SEAHORSE (SH) conciseness scores (area under the ROC curve) are in bold. Reference shows NLI and SH scores on the reference/target summaries.
### $\mathbf{LowData}$
## Cross-Lingual Transfer iiiii 4 Experimental Setup
4.1 Datasets
Table l
XLSum (Hasan et al., 2021) 45BBC300K-1K
Appendix A Table 7
XWikis (Perez-Beltrachini and Lapata, 2021)
WikipediaXWikise Appendix A Table 8
### 4.2 $\textbf{Modeling Details}$
PaLM2 (Anil et al., 2023) LLMPaLM (Chowdhery et al., 2023) XXSSFLAN-PaLM2 (Wei et al., 2022a) TPU 1$\{ 1e^{- 3}$ $2e^{- 4}$ $2e^{- 5}$ $\}$XLSum/ 2,048/128XWikis/ 2,048/256
## 4.3 Automatic Evaluation
ROUGE (Lin, 2004) n-gram Aharoni et al. (2023) SentencePiece (Kudo and Richardson, 2018) $\operatorname{ROUGE}$
(Falke et al., 2019; Kumar and Talukdar, 2020; Honovich et al., 2022; Whitehouse et al., 2023b; Huot et al., 2024) NLImT5-XXL (Xue et al., 2021) ANLI (Nie et al., 2020) XNLI (Conneau et al., 2018) (Aharoni et al., 2023; Huot et al., 2024)
SEAHORSE (Clark et al.,2023) mT5-XXL
(Xue et al., 2021) 2
### 5 $\textbf{Results and Analvsis}$
LoRA
$^{1}$See Appendix B for more details of the TPUs used.
$^{2}$https://huggingface.co/google/seahorse-large-q6
# 5.1 High-data Regime
XLSumXWikis Table 2 FullFTLoRA FT-At $r= \{$ 4, 16, 64, 512 $\}$ PaLM 2-XXS ROUGE- ${\mathrm{L(~p<0.01~)()()}}$NLI SEAHORSE
XLSum
XLSumXWikisROUGE-L
1XWikis86ROUGE-LLoRA4LoRA(208)
ROUGE-L XLSum 116XWikisXWikis64LoRA XLSum LoRA1ROUGE -L Chen et al.(2022) LoRA
LoRANLI
LoRAXL-
SEAHORSE
Sum5213SentencePieceLoRA44882 SEAHORSE
LoRANLI Appendix D Appendix C
Takeaways LoRALoRA
## 5.2 Low-data Regime
LoRA4LoRA
XLSumXWikis16642561,0244,096 3 ROUGE-
$L$
Figure 1 LoRAPaLM2-XXS16x log XLSum $\sim2^{14.6}$ XWikis $\sim2^18.1$ 256LoRANLI SEAHORSE ROUGE-LLoRA2561,024
LoRA
Takeaways LoRALoRA
## 5.3 Cross-lingual Transfer $LoRA4$
Zero-shot Transfer from English ie
Table 3 LoRA PaLM 2-XXS Table 3 5.1 XL-
SumNLISHLoRA Table 3 XLSumXWiki(ie )
3When the number of training samples is set to 4,096, three languages in XLSum already lack sufficient data, so we refrain from selecting more examples per language.
|
[
[
0.1234130859375,
0.08489990234375,
0.172119140625,
0.0872802734375,
0.17041015625,
0.09979248046875,
0.12176513671875,
0.097412109375
],
[
0.6689453125,
0.08612060546875,
0.71728515625,
0.08612060546875,
0.71728515625,
0.097412109375,
0.6689453125,
0.097412109375
],
[
0.7900390625,
0.08612060546875,
0.8369140625,
0.08612060546875,
0.8369140625,
0.097412109375,
0.7900390625,
0.097412109375
],
[
0.4013671875,
0.0872802734375,
0.45166015625,
0.0872802734375,
0.45166015625,
0.09857177734375,
0.4013671875,
0.09857177734375
],
[
0.280517578125,
0.0872802734375,
0.33154296875,
0.0872802734375,
0.33154296875,
0.09857177734375,
0.280517578125,
0.09857177734375
],
[
0.58447265625,
0.0968017578125,
0.62890625,
0.0968017578125,
0.62890625,
0.10809326171875,
0.58447265625,
0.10809326171875
],
[
0.63818359375,
0.0968017578125,
0.6650390625,
0.0968017578125,
0.6650390625,
0.10809326171875,
0.63818359375,
0.10809326171875
],
[
0.67919921875,
0.0968017578125,
0.7080078125,
0.0968017578125,
0.7080078125,
0.1092529296875,
0.67919921875,
0.1092529296875
],
[
0.76318359375,
0.0968017578125,
0.78857421875,
0.0968017578125,
0.78857421875,
0.10809326171875,
0.76318359375,
0.10809326171875
],
[
0.8037109375,
0.0968017578125,
0.8330078125,
0.0968017578125,
0.8330078125,
0.1092529296875,
0.8037109375,
0.1092529296875
],
[
0.84619140625,
0.0968017578125,
0.86962890625,
0.0968017578125,
0.86962890625,
0.1092529296875,
0.84619140625,
0.1092529296875
],
[
0.724609375,
0.097412109375,
0.74658203125,
0.097412109375,
0.74658203125,
0.10809326171875,
0.724609375,
0.10809326171875
],
[
0.3955078125,
0.1015625,
0.459228515625,
0.1015625,
0.459228515625,
0.11639404296875,
0.3955078125,
0.11639404296875
],
[
0.273681640625,
0.10394287109375,
0.339111328125,
0.10394287109375,
0.339111328125,
0.11517333984375,
0.273681640625,
0.11517333984375
],
[
0.124267578125,
0.10394287109375,
0.1695556640625,
0.10394287109375,
0.1695556640625,
0.11517333984375,
0.124267578125,
0.11517333984375
],
[
0.4248046875,
0.11102294921875,
0.4375,
0.119384765625,
0.4248046875,
0.127685546875,
0.4130859375,
0.119384765625
],
[
0.51708984375,
0.10986328125,
0.5751953125,
0.11224365234375,
0.5751953125,
0.12353515625,
0.51708984375,
0.1217041015625
],
[
0.71728515625,
0.11224365234375,
0.75390625,
0.11224365234375,
0.75390625,
0.12353515625,
0.71728515625,
0.12353515625
],
[
0.8017578125,
0.11224365234375,
0.83642578125,
0.11224365234375,
0.83642578125,
0.12353515625,
0.8017578125,
0.12353515625
],
[
0.84033203125,
0.11102294921875,
0.875,
0.11102294921875,
0.875,
0.122314453125,
0.84033203125,
0.122314453125
],
[
0.67822265625,
0.1134033203125,
0.70947265625,
0.1134033203125,
0.70947265625,
0.1217041015625,
0.67822265625,
0.1217041015625
],
[
0.124267578125,
0.11639404296875,
0.189697265625,
0.11639404296875,
0.189697265625,
0.127685546875,
0.124267578125,
0.127685546875
],
[
0.295654296875,
0.11639404296875,
0.318115234375,
0.11639404296875,
0.318115234375,
0.1270751953125,
0.295654296875,
0.1270751953125
],
[
0.2418212890625,
0.1270751953125,
0.36865234375,
0.1270751953125,
0.36865234375,
0.14013671875,
0.2418212890625,
0.14013671875
],
[
0.36279296875,
0.1270751953125,
0.48193359375,
0.1270751953125,
0.48193359375,
0.14013671875,
0.36279296875,
0.14013671875
],
[
0.51904296875,
0.1270751953125,
0.5615234375,
0.1270751953125,
0.5615234375,
0.1383056640625,
0.51904296875,
0.1383056640625
],
[
0.591796875,
0.1270751953125,
0.62890625,
0.1270751953125,
0.62890625,
0.1383056640625,
0.591796875,
0.1383056640625
],
[
0.623046875,
0.1270751953125,
0.6689453125,
0.1270751953125,
0.6689453125,
0.1383056640625,
0.623046875,
0.1383056640625
],
[
0.67529296875,
0.1270751953125,
0.7109375,
0.1270751953125,
0.7109375,
0.1383056640625,
0.67529296875,
0.1383056640625
],
[
0.71533203125,
0.1270751953125,
0.75390625,
0.1270751953125,
0.75390625,
0.1383056640625,
0.71533203125,
0.1383056640625
],
[
0.755859375,
0.1270751953125,
0.79248046875,
0.1270751953125,
0.79248046875,
0.1383056640625,
0.755859375,
0.1383056640625
],
[
0.80029296875,
0.1270751953125,
0.83642578125,
0.1270751953125,
0.83642578125,
0.1383056640625,
0.80029296875,
0.1383056640625
],
[
0.8388671875,
0.1270751953125,
0.8759765625,
0.1270751953125,
0.8759765625,
0.1383056640625,
0.8388671875,
0.1383056640625
],
[
0.80029296875,
0.1353759765625,
0.83447265625,
0.13720703125,
0.8330078125,
0.149658203125,
0.79833984375,
0.1478271484375
],
[
0.1259765625,
0.127685546875,
0.2376708984375,
0.127685546875,
0.2376708984375,
0.1395263671875,
0.1259765625,
0.1395263671875
],
[
0.75732421875,
0.13720703125,
0.79248046875,
0.13720703125,
0.79248046875,
0.149658203125,
0.75732421875,
0.149658203125
],
[
0.84033203125,
0.13720703125,
0.875,
0.13720703125,
0.875,
0.149658203125,
0.84033203125,
0.149658203125
],
[
0.59423828125,
0.1383056640625,
0.62890625,
0.1383056640625,
0.62890625,
0.149658203125,
0.59423828125,
0.149658203125
],
[
0.51904296875,
0.1383056640625,
0.5576171875,
0.1383056640625,
0.5576171875,
0.149658203125,
0.51904296875,
0.149658203125
],
[
0.6337890625,
0.1383056640625,
0.6689453125,
0.1383056640625,
0.6689453125,
0.149658203125,
0.6337890625,
0.149658203125
],
[
0.67529296875,
0.1383056640625,
0.7109375,
0.1383056640625,
0.7109375,
0.149658203125,
0.67529296875,
0.149658203125
],
[
0.71728515625,
0.1383056640625,
0.755859375,
0.1383056640625,
0.755859375,
0.149658203125,
0.71728515625,
0.149658203125
],
[
0.387939453125,
0.1395263671875,
0.466796875,
0.1395263671875,
0.466796875,
0.15087890625,
0.387939453125,
0.15087890625
],
[
0.124267578125,
0.14013671875,
0.2418212890625,
0.14013671875,
0.2418212890625,
0.15087890625,
0.124267578125,
0.15087890625
],
[
0.270263671875,
0.14013671875,
0.341796875,
0.14013671875,
0.341796875,
0.15087890625,
0.270263671875,
0.15087890625
],
[
0.51904296875,
0.152587890625,
0.57861328125,
0.152587890625,
0.57861328125,
0.1644287109375,
0.51904296875,
0.1644287109375
],
[
0.587890625,
0.152587890625,
0.62744140625,
0.152587890625,
0.62744140625,
0.1644287109375,
0.587890625,
0.1644287109375
],
[
0.6337890625,
0.152587890625,
0.66748046875,
0.152587890625,
0.66748046875,
0.1644287109375,
0.6337890625,
0.1644287109375
],
[
0.6767578125,
0.152587890625,
0.7109375,
0.152587890625,
0.7109375,
0.1644287109375,
0.6767578125,
0.1644287109375
],
[
0.75732421875,
0.152587890625,
0.79248046875,
0.152587890625,
0.79248046875,
0.1644287109375,
0.75732421875,
0.1644287109375
],
[
0.80029296875,
0.152587890625,
0.83447265625,
0.152587890625,
0.83447265625,
0.1644287109375,
0.80029296875,
0.1644287109375
],
[
0.84033203125,
0.152587890625,
0.875,
0.152587890625,
0.875,
0.1644287109375,
0.84033203125,
0.1644287109375
],
[
0.67529296875,
0.162109375,
0.7109375,
0.1644287109375,
0.70947265625,
0.17578125,
0.67333984375,
0.17333984375
],
[
0.71875,
0.155029296875,
0.75244140625,
0.155029296875,
0.75244140625,
0.163330078125,
0.71875,
0.163330078125
],
[
0.75732421875,
0.162109375,
0.79443359375,
0.1644287109375,
0.79248046875,
0.177001953125,
0.755859375,
0.174560546875
],
[
0.8388671875,
0.162109375,
0.875,
0.1644287109375,
0.873046875,
0.17578125,
0.8369140625,
0.17333984375
],
[
0.79833984375,
0.163330078125,
0.83642578125,
0.163330078125,
0.83642578125,
0.174560546875,
0.79833984375,
0.174560546875
],
[
0.51904296875,
0.1644287109375,
0.57421875,
0.1644287109375,
0.57421875,
0.17578125,
0.51904296875,
0.17578125
],
[
0.591796875,
0.163330078125,
0.62744140625,
0.163330078125,
0.62744140625,
0.17578125,
0.591796875,
0.17578125
],
[
0.6337890625,
0.1644287109375,
0.6708984375,
0.1644287109375,
0.6708984375,
0.17578125,
0.6337890625,
0.17578125
],
[
0.71728515625,
0.1644287109375,
0.75244140625,
0.1644287109375,
0.75244140625,
0.17578125,
0.71728515625,
0.17578125
],
[
0.20654296875,
0.1650390625,
0.355224609375,
0.166259765625,
0.355224609375,
0.179931640625,
0.20654296875,
0.1787109375
],
[
0.1158447265625,
0.16748046875,
0.187255859375,
0.1650390625,
0.187255859375,
0.177001953125,
0.1158447265625,
0.1787109375
],
[
0.39208984375,
0.16748046875,
0.42822265625,
0.16748046875,
0.42822265625,
0.1787109375,
0.39208984375,
0.1787109375
],
[
0.426513671875,
0.16748046875,
0.486083984375,
0.16748046875,
0.486083984375,
0.1787109375,
0.426513671875,
0.1787109375
],
[
0.6337890625,
0.17333984375,
0.6689453125,
0.17333984375,
0.6689453125,
0.185302734375,
0.6337890625,
0.185302734375
],
[
0.75830078125,
0.17333984375,
0.79443359375,
0.17333984375,
0.79443359375,
0.185302734375,
0.75830078125,
0.185302734375
],
[
0.8388671875,
0.17333984375,
0.875,
0.17333984375,
0.875,
0.185302734375,
0.8388671875,
0.185302734375
],
[
0.51904296875,
0.174560546875,
0.57275390625,
0.174560546875,
0.57275390625,
0.1859130859375,
0.51904296875,
0.1859130859375
],
[
0.59375,
0.17578125,
0.62646484375,
0.17578125,
0.62646484375,
0.18408203125,
0.59375,
0.18408203125
],
[
0.67822265625,
0.17578125,
0.70947265625,
0.17578125,
0.70947265625,
0.18408203125,
0.67822265625,
0.18408203125
],
[
0.71875,
0.17578125,
0.75,
0.17578125,
0.75,
0.18408203125,
0.71875,
0.18408203125
],
[
0.80029296875,
0.17578125,
0.8330078125,
0.17578125,
0.8330078125,
0.18408203125,
0.80029296875,
0.18408203125
],
[
0.1158447265625,
0.18115234375,
0.42236328125,
0.18115234375,
0.42236328125,
0.19482421875,
0.1158447265625,
0.19482421875
],
[
0.35693359375,
0.168701171875,
0.3896484375,
0.168701171875,
0.3896484375,
0.177001953125,
0.35693359375,
0.177001953125
],
[
0.414794921875,
0.1817626953125,
0.484375,
0.1817626953125,
0.484375,
0.193603515625,
0.414794921875,
0.193603515625
],
[
0.51904296875,
0.185302734375,
0.56689453125,
0.185302734375,
0.56689453125,
0.196533203125,
0.51904296875,
0.196533203125
],
[
0.591796875,
0.185302734375,
0.62744140625,
0.185302734375,
0.62744140625,
0.196533203125,
0.591796875,
0.196533203125
],
[
0.62646484375,
0.185302734375,
0.6689453125,
0.185302734375,
0.6689453125,
0.196533203125,
0.62646484375,
0.196533203125
],
[
0.67529296875,
0.185302734375,
0.7109375,
0.185302734375,
0.7109375,
0.196533203125,
0.67529296875,
0.196533203125
],
[
0.75732421875,
0.185302734375,
0.79248046875,
0.185302734375,
0.79248046875,
0.196533203125,
0.75732421875,
0.196533203125
],
[
0.80029296875,
0.185302734375,
0.83447265625,
0.185302734375,
0.83447265625,
0.196533203125,
0.80029296875,
0.196533203125
],
[
0.84033203125,
0.185302734375,
0.875,
0.185302734375,
0.875,
0.196533203125,
0.84033203125,
0.196533203125
],
[
0.1158447265625,
0.19482421875,
0.48779296875,
0.19482421875,
0.48779296875,
0.2078857421875,
0.1158447265625,
0.2078857421875
],
[
0.71875,
0.1859130859375,
0.75146484375,
0.1859130859375,
0.75146484375,
0.19482421875,
0.71875,
0.19482421875
],
[
0.1158447265625,
0.208984375,
0.486083984375,
0.208984375,
0.486083984375,
0.22265625,
0.1158447265625,
0.22265625
],
[
0.509765625,
0.2120361328125,
0.87890625,
0.2120361328125,
0.87890625,
0.22265625,
0.509765625,
0.22265625
],
[
0.11419677734375,
0.223876953125,
0.484375,
0.223876953125,
0.484375,
0.2369384765625,
0.11419677734375,
0.2369384765625
],
[
0.509765625,
0.223876953125,
0.8818359375,
0.2244873046875,
0.8818359375,
0.2392578125,
0.509765625,
0.2381591796875
],
[
0.1158447265625,
0.2381591796875,
0.491943359375,
0.2381591796875,
0.491943359375,
0.251708984375,
0.1158447265625,
0.251708984375
],
[
0.509765625,
0.2392578125,
0.880859375,
0.2392578125,
0.880859375,
0.2529296875,
0.509765625,
0.2529296875
],
[
0.11419677734375,
0.251708984375,
0.1695556640625,
0.254150390625,
0.1695556640625,
0.265380859375,
0.11419677734375,
0.26318359375
],
[
0.51123046875,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.51123046875,
0.267333984375
],
[
0.51123046875,
0.267333984375,
0.880859375,
0.267333984375,
0.880859375,
0.28076171875,
0.51123046875,
0.28076171875
],
[
0.509765625,
0.27978515625,
0.8818359375,
0.281982421875,
0.8818359375,
0.295654296875,
0.509765625,
0.29345703125
],
[
0.118408203125,
0.292236328125,
0.202392578125,
0.292236328125,
0.202392578125,
0.30712890625,
0.118408203125,
0.30712890625
],
[
0.51123046875,
0.29638671875,
0.87890625,
0.29638671875,
0.87890625,
0.30712890625,
0.51123046875,
0.30712890625
],
[
0.51123046875,
0.31005859375,
0.8447265625,
0.31005859375,
0.8447265625,
0.32373046875,
0.51123046875,
0.32373046875
],
[
0.12176513671875,
0.318359375,
0.312255859375,
0.318359375,
0.312255859375,
0.33203125,
0.12176513671875,
0.33203125
],
[
0.3046875,
0.319580078125,
0.35693359375,
0.319580078125,
0.35693359375,
0.330810546875,
0.3046875,
0.330810546875
],
[
0.1175537109375,
0.349853515625,
0.1318359375,
0.349853515625,
0.1318359375,
0.35986328125,
0.1175537109375,
0.35986328125
],
[
0.1502685546875,
0.347412109375,
0.32568359375,
0.3486328125,
0.32568359375,
0.363525390625,
0.1502685546875,
0.3623046875
],
[
0.51318359375,
0.352783203125,
0.5498046875,
0.352783203125,
0.5498046875,
0.364013671875,
0.51318359375,
0.364013671875
],
[
0.5556640625,
0.35400390625,
0.70361328125,
0.35400390625,
0.70361328125,
0.364013671875,
0.5556640625,
0.364013671875
],
[
0.509765625,
0.37353515625,
0.87158203125,
0.37353515625,
0.87158203125,
0.38720703125,
0.509765625,
0.38720703125
],
[
0.515625,
0.39013671875,
0.8818359375,
0.39013671875,
0.8818359375,
0.40380859375,
0.515625,
0.40380859375
],
[
0.1175537109375,
0.402587890625,
0.1435546875,
0.402587890625,
0.1435546875,
0.41455078125,
0.1175537109375,
0.41455078125
],
[
0.156982421875,
0.40087890625,
0.2333984375,
0.402587890625,
0.2333984375,
0.41748046875,
0.156982421875,
0.4150390625
],
[
0.51123046875,
0.40673828125,
0.75146484375,
0.40673828125,
0.75146484375,
0.42041015625,
0.51123046875,
0.42041015625
],
[
0.12005615234375,
0.4228515625,
0.1796875,
0.4228515625,
0.1796875,
0.43701171875,
0.12005615234375,
0.43701171875
],
[
0.52880859375,
0.42333984375,
0.72900390625,
0.42333984375,
0.72900390625,
0.43701171875,
0.52880859375,
0.43701171875
],
[
0.8193359375,
0.4228515625,
0.88330078125,
0.4228515625,
0.88330078125,
0.43701171875,
0.8193359375,
0.43701171875
],
[
0.51123046875,
0.439453125,
0.81201171875,
0.439453125,
0.81201171875,
0.453125,
0.51123046875,
0.453125
],
[
0.1234130859375,
0.449462890625,
0.486083984375,
0.449462890625,
0.486083984375,
0.463134765625,
0.1234130859375,
0.463134765625
],
[
0.51123046875,
0.454833984375,
0.880859375,
0.454833984375,
0.880859375,
0.468505859375,
0.51123046875,
0.468505859375
],
[
0.1158447265625,
0.465576171875,
0.273681640625,
0.46435546875,
0.273681640625,
0.478515625,
0.1158447265625,
0.479736328125
],
[
0.509765625,
0.471435546875,
0.8818359375,
0.471435546875,
0.8818359375,
0.485107421875,
0.509765625,
0.485107421875
],
[
0.509765625,
0.485107421875,
0.5634765625,
0.48681640625,
0.5615234375,
0.501953125,
0.5087890625,
0.49951171875
],
[
0.124267578125,
0.4921875,
0.48779296875,
0.4921875,
0.48779296875,
0.505859375,
0.124267578125,
0.505859375
],
[
0.52880859375,
0.505859375,
0.60791015625,
0.505859375,
0.60791015625,
0.51708984375,
0.52880859375,
0.51708984375
],
[
0.8037109375,
0.505859375,
0.880859375,
0.505859375,
0.880859375,
0.51708984375,
0.8037109375,
0.51708984375
],
[
0.6650390625,
0.50732421875,
0.75146484375,
0.50732421875,
0.75146484375,
0.51708984375,
0.6650390625,
0.51708984375
],
[
0.1158447265625,
0.50732421875,
0.42236328125,
0.50830078125,
0.42236328125,
0.5224609375,
0.1158447265625,
0.521484375
],
[
0.509765625,
0.52099609375,
0.87158203125,
0.521484375,
0.87158203125,
0.53515625,
0.509765625,
0.53369140625
],
[
0.156982421875,
0.5361328125,
0.298828125,
0.5361328125,
0.298828125,
0.55029296875,
0.156982421875,
0.55029296875
],
[
0.5283203125,
0.5361328125,
0.8037109375,
0.53759765625,
0.8037109375,
0.5517578125,
0.5283203125,
0.55029296875
],
[
0.154541015625,
0.54052734375,
0.160400390625,
0.54052734375,
0.160400390625,
0.5458984375,
0.154541015625,
0.5458984375
],
[
0.118408203125,
0.53759765625,
0.156982421875,
0.53759765625,
0.156982421875,
0.548828125,
0.118408203125,
0.548828125
],
[
0.118408203125,
0.55810546875,
0.177978515625,
0.55810546875,
0.177978515625,
0.5693359375,
0.118408203125,
0.5693359375
],
[
0.1796875,
0.55859375,
0.2392578125,
0.55859375,
0.2392578125,
0.57080078125,
0.1796875,
0.57080078125
],
[
0.2392578125,
0.55859375,
0.32568359375,
0.55859375,
0.32568359375,
0.5693359375,
0.2392578125,
0.5693359375
],
[
0.333251953125,
0.55810546875,
0.382080078125,
0.55810546875,
0.382080078125,
0.5693359375,
0.333251953125,
0.5693359375
],
[
0.3779296875,
0.56005859375,
0.399658203125,
0.56005859375,
0.399658203125,
0.56787109375,
0.3779296875,
0.56787109375
],
[
0.3955078125,
0.55859375,
0.486083984375,
0.55859375,
0.486083984375,
0.5693359375,
0.3955078125,
0.5693359375
],
[
0.515625,
0.56396484375,
0.623046875,
0.56640625,
0.623046875,
0.58056640625,
0.515625,
0.5791015625
],
[
0.615234375,
0.56640625,
0.7138671875,
0.56640625,
0.7138671875,
0.57763671875,
0.615234375,
0.57763671875
],
[
0.1175537109375,
0.57373046875,
0.48779296875,
0.57373046875,
0.48779296875,
0.58740234375,
0.1175537109375,
0.58740234375
],
[
0.1158447265625,
0.58837890625,
0.4375,
0.58837890625,
0.4375,
0.6015625,
0.1158447265625,
0.6015625
],
[
0.432373046875,
0.5888671875,
0.486083984375,
0.58740234375,
0.48779296875,
0.6015625,
0.43408203125,
0.60400390625
],
[
0.51318359375,
0.59423828125,
0.54150390625,
0.59423828125,
0.54150390625,
0.60693359375,
0.51318359375,
0.60693359375
],
[
0.55322265625,
0.59326171875,
0.7001953125,
0.59423828125,
0.7001953125,
0.609375,
0.55322265625,
0.60791015625
],
[
0.1175537109375,
0.60546875,
0.408935546875,
0.60546875,
0.408935546875,
0.619140625,
0.1175537109375,
0.619140625
],
[
0.51318359375,
0.61767578125,
0.61376953125,
0.61767578125,
0.61376953125,
0.62890625,
0.51318359375,
0.62890625
],
[
0.1175537109375,
0.63427734375,
0.1485595703125,
0.63427734375,
0.1485595703125,
0.6455078125,
0.1175537109375,
0.6455078125
],
[
0.1561279296875,
0.63427734375,
0.337646484375,
0.63427734375,
0.337646484375,
0.64794921875,
0.1561279296875,
0.64794921875
],
[
0.5283203125,
0.6318359375,
0.880859375,
0.6318359375,
0.880859375,
0.6455078125,
0.5283203125,
0.6455078125
],
[
0.15283203125,
0.63818359375,
0.160400390625,
0.63818359375,
0.160400390625,
0.642578125,
0.15283203125,
0.642578125
],
[
0.51904296875,
0.6484375,
0.53076171875,
0.65673828125,
0.51904296875,
0.6650390625,
0.50732421875,
0.65673828125
],
[
0.1175537109375,
0.65478515625,
0.486083984375,
0.65478515625,
0.486083984375,
0.66943359375,
0.1175537109375,
0.66943359375
],
[
0.5283203125,
0.6650390625,
0.8818359375,
0.66650390625,
0.8818359375,
0.68115234375,
0.5283203125,
0.6796875
],
[
0.1158447265625,
0.67138671875,
0.48779296875,
0.67138671875,
0.48779296875,
0.68505859375,
0.1158447265625,
0.68505859375
],
[
0.51318359375,
0.6787109375,
0.8818359375,
0.6787109375,
0.8818359375,
0.69677734375,
0.51318359375,
0.69677734375
],
[
0.1175537109375,
0.689453125,
0.1822509765625,
0.689453125,
0.1822509765625,
0.69970703125,
0.1175537109375,
0.69970703125
],
[
0.509765625,
0.69775390625,
0.7080078125,
0.69775390625,
0.7080078125,
0.71142578125,
0.509765625,
0.71142578125
],
[
0.14111328125,
0.70361328125,
0.486083984375,
0.70361328125,
0.486083984375,
0.71728515625,
0.14111328125,
0.71728515625
],
[
0.1158447265625,
0.71875,
0.484375,
0.71875,
0.484375,
0.732421875,
0.1158447265625,
0.732421875
],
[
0.5283203125,
0.71533203125,
0.580078125,
0.71533203125,
0.580078125,
0.72998046875,
0.5283203125,
0.72998046875
],
[
0.1158447265625,
0.73388671875,
0.48779296875,
0.736328125,
0.48779296875,
0.75,
0.1158447265625,
0.74755859375
],
[
0.5146484375,
0.7421875,
0.60791015625,
0.74462890625,
0.60791015625,
0.7587890625,
0.5146484375,
0.75732421875
],
[
0.60791015625,
0.74462890625,
0.7138671875,
0.74462890625,
0.7138671875,
0.755859375,
0.60791015625,
0.755859375
],
[
0.118408203125,
0.75439453125,
0.484375,
0.75439453125,
0.484375,
0.76416015625,
0.118408203125,
0.76416015625
],
[
0.1158447265625,
0.76953125,
0.2166748046875,
0.76708984375,
0.2166748046875,
0.78076171875,
0.1158447265625,
0.783203125
],
[
0.2091064453125,
0.76708984375,
0.489501953125,
0.7685546875,
0.489501953125,
0.783203125,
0.2091064453125,
0.7822265625
],
[
0.51318359375,
0.7724609375,
0.5400390625,
0.7724609375,
0.5400390625,
0.78515625,
0.51318359375,
0.78515625
],
[
0.55517578125,
0.771484375,
0.73779296875,
0.7724609375,
0.73779296875,
0.78759765625,
0.55517578125,
0.7861328125
],
[
0.1158447265625,
0.78369140625,
0.16455078125,
0.78369140625,
0.16455078125,
0.798828125,
0.1158447265625,
0.798828125
],
[
0.231689453125,
0.80029296875,
0.486083984375,
0.80029296875,
0.486083984375,
0.81396484375,
0.231689453125,
0.81396484375
],
[
0.51123046875,
0.7958984375,
0.57080078125,
0.7958984375,
0.57080078125,
0.80712890625,
0.51123046875,
0.80712890625
],
[
0.145263671875,
0.8017578125,
0.2333984375,
0.8017578125,
0.2333984375,
0.8134765625,
0.145263671875,
0.8134765625
],
[
0.1175537109375,
0.81640625,
0.26611328125,
0.81396484375,
0.26611328125,
0.8291015625,
0.1175537109375,
0.83056640625
],
[
0.515625,
0.822265625,
0.8076171875,
0.822265625,
0.8076171875,
0.8359375,
0.515625,
0.8359375
],
[
0.1502685546875,
0.84423828125,
0.334228515625,
0.845703125,
0.334228515625,
0.85986328125,
0.1502685546875,
0.859375
],
[
0.52880859375,
0.8388671875,
0.8818359375,
0.8388671875,
0.8818359375,
0.8525390625,
0.52880859375,
0.8525390625
],
[
0.118408203125,
0.8466796875,
0.1318359375,
0.8466796875,
0.1318359375,
0.85693359375,
0.118408203125,
0.85693359375
],
[
0.509765625,
0.85400390625,
0.8486328125,
0.85498046875,
0.8486328125,
0.869140625,
0.509765625,
0.8681640625
],
[
0.11419677734375,
0.87158203125,
0.16796875,
0.87158203125,
0.16796875,
0.8857421875,
0.11419677734375,
0.8857421875
],
[
0.53076171875,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.53076171875,
0.8935546875
],
[
0.1376953125,
0.8935546875,
0.438232421875,
0.8935546875,
0.438232421875,
0.90673828125,
0.1376953125,
0.90673828125
],
[
0.51123046875,
0.8935546875,
0.8818359375,
0.8935546875,
0.8818359375,
0.90673828125,
0.51123046875,
0.90673828125
],
[
0.1351318359375,
0.90380859375,
0.483642578125,
0.90478515625,
0.483642578125,
0.91943359375,
0.1351318359375,
0.91845703125
],
[
0.51123046875,
0.90625,
0.765625,
0.90625,
0.765625,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"Dataset",
"XLSUM",
"XWIKIS",
"XWIKIS",
"XLSUM",
"Params",
"R-L",
"NLI",
"R-L",
"NLI",
"SH",
"SH",
"Wikipedia",
"BBCNews",
"Source",
"5",
"Reference",
"31.65",
"39.20",
"25.19",
"48.50",
"Languages",
"44",
"1.12M/114K/114K",
"1.43M/40K/35K",
"FullFT",
"100%",
"31.11",
"42.93",
"31.64",
"34.08",
"41.04",
"25.19",
"37.06",
"Train/Val/Test Data",
"32.22",
"24.20",
"20%",
"FT-Att",
"30.88",
"50.32",
"36.12",
"1042.7/63.7",
"Input/OutputWords",
"470.2/22.1",
"LoRA-512",
"13.3 %",
"29.81",
"42.58",
"33.38",
"40.48",
"24.78",
"45.51",
"30.16",
"34.04",
"27.02",
"45.34",
"LoRA-64",
"1.7%",
"29.79",
"31.80",
"Summary statistics",
"Table1:",
"the",
"XLSum",
"29.77",
"33.80",
"27.42",
"LoRA-16",
"0.4%",
"48.48",
"33.25",
"46.10",
"and XWikis multilingual summarization",
"for",
"datasets.",
"LoRA-4",
"0.1 %",
"29.03",
"51.16",
"32.92",
"47.43",
"27.72",
"Train/Val/Test shows the number of examples in each",
"34.42",
"split. Input/Output shows the average number of words",
"Table 2: Results on XLSum and XWikis with PaLM",
"in the English input document and output summary.",
"2-XxS trained in the high-data regime: full fine-tuning",
"XWikis has long documents and multi-sentence sum-",
"on all layers (Full FT), on attention layers (FT-Att), and",
"maries.",
"LoRA-* (with different ranks). Params denotes the pro-",
"portion of trainable parameters. Best ROUGE-L (R-L),",
"NLI, and sEAHORSE (SH) conciseness scores (area un-",
"Low Data",
"dertheROC curve)are in bold.Reference shows NLI",
"and SH scores on the reference/target summaries.",
"Cross-Lingual Transfer",
"iiii",
"4",
"Experimental Setup",
"5.1",
"High-dataRegime",
"XLSumXWikis Table 2 FullFTLoRA FT-Att",
"r =[ 4, 16, 64, 5121 PaLM2-XXS ROUGE-",
"4.1",
"Datasets",
"L( p<0.01 )OONLI SEAHORSE",
"Table 1",
"XLSumXWikisROUGE-L",
"XLSum",
"1XWikis86ROUGE-LLoRA4LoRA(208",
"XLSum (Hasan et al., 2021) 45BBC300K-1K",
"ROUGE-L XLSum 116XWikisXWikis64LoRA",
"Appendix A Table 7",
"XLSum LoRA1ROUGE -L Chen et al. (2022)",
"LoRA",
"XWikis (Perez-Beltrachini and Lapata, 2021)",
"LoRANLI",
"LoRAXL",
"SEAHORSE",
"WikipediaXWikise Appendix A Table 8",
"Sum5213SentencePieceLoRA44882SEAHORSE",
" Modeling Details",
"LoRANLI Appendix D Appendix C",
"",
"4.2",
"PaLM2",
"(Anil",
"et al..",
"2023)",
"",
" LLMPaLM",
"Takeaways",
"LoRALoRA",
"(Chowdhery et al., 2023) XXSSFLAN-PaLM2",
"(Wei et al.. 2022a) TPU 1 ( 1e-3 2e-4",
"2e-5",
"5.2",
"Low-data Regime",
"jXLSum/2,048/128XWikis/2,048/256",
"LoRA4LoRA",
"4.3",
" Automatic Evaluation",
"XLSumXWikis16642561.0244.096 3 ROUGE",
"",
"V",
"ROUGE (Lin, 2004) n-gram Aharoni et al. (2023)",
"Figure 1 LoRAPaLM2-XXS16x log XLSum",
"SentencePiece (Kudo and Richardson, 2018)",
"~214.6 XWikis ~218.1 256LoRANLI SEAHORSE",
"ROUGE",
"ROUGE-LLoRA2561.024",
"(Falke et al., 2019; Kumar and Talukdar, 2020;",
"Honovich et al., 2022; Whitehouse et al., 2023b:",
"LoRA",
"Huot et al., 2024) NLImT5-XXL (Xue et al.,",
"Takeaways",
"LoRALoRA",
"2021)ANLI (Nie et al..2020) XNLI (Conneau",
"et al., 2018)",
"(Aharoni et al., 2023; Huot et al.,",
"5.3",
"Cross-lingual Transfer",
"2024)",
"(Clark et al., 2023) mT5-XXL",
"LoRA4",
"SEAHORSE",
"(Xue et al., 2021) 2",
"Zero-shot Transfer from English ie",
"Results and Analysis",
"Table 3 LoRA PaLM2-XXS Table 3 5.1 XL",
"5",
"SumNLISHLoRA Table 3 XLSumXWiki(ie)",
"LoRA",
"3When the number of training samples is set to 4,096, three",
"ISee Appendix B for more details of the TPUs used.",
"languages in XLSum already lack sufficient data, so we refrain",
"2https://huggingface.co/google/seahorse-large-q6",
"from selecting more examples per language."
] |
{
"arxiv_id": "2311.08572",
"language": "zh",
"page_idx": 2
}
|
||
2311.08572-zh-3
|
2311.08572-zh-3.jpg
|
<FigureHere>
Figure 1: Results on the XLSum and XWikis datasets with PaLM 2-XXS trained in the low $\to$ high-data regime: Full FT vs. LoRA-4. Results for up to 256 examples per language are averaged over three seeds, with standard deviation shown in shaded areas.
<table>
<tbody>
<tr>
<th> </th>
<th rowspan="2">Test Languages</th>
<th colspan="3">$XLSUM$</th>
<th colspan="3">XWIKIS</th>
</tr>
<tr>
<th> </th>
<th>$R-L$</th>
<th>$NLI$</th>
<th>SH</th>
<th>$R-L$</th>
<th>$NLI$</th>
<th>SH</th>
</tr>
<tr>
<td>Full FT</td>
<td rowspan="2">Non-English</td>
<td>5.20</td>
<td>4.49</td>
<td>6.88</td>
<td>17.51</td>
<td>35.95</td>
<td>22.43</td>
</tr>
<tr>
<td>LoRA- 4</td>
<td>21.13</td>
<td>39.07</td>
<td>23.08</td>
<td>23.86</td>
<td>45.54</td>
<td>25.96</td>
</tr>
<tr>
<td>Full FT</td>
<td rowspan="2">English</td>
<td>32.58</td>
<td>57.09</td>
<td>38.01</td>
<td>36.59</td>
<td>53.59</td>
<td>30.81</td>
</tr>
<tr>
<td>LoRA- 4</td>
<td>32.21</td>
<td>63.13</td>
<td>43.44</td>
<td>34.07</td>
<td>49.94</td>
<td>29.01</td>
</tr>
</tbody>
</table>
Table 3: Zero-shot cross-lingual transfer using full finetuning (Full FT) and LoRA (rank 4); PaLM 2-XXS models are trained and validated on English and tested on all other languages (Non-English) and English only. Best ROUGE-L (R-L), NLI, and SEAHORSE (SH) conciseness scores (area under the ROC curve) are in bold.
# Figure 2 HausaXLSum() Appendix C PaLM2 4 Vu et al. (2022)
# Zero-shot Transfer from Multiple Languages
XLSum10( AR )( ZH )( EN )( HA )( HI )( ID )( FA )( PT )(XMAT10 AZ BN JA RN KO NE 5 XWikis
LoRALoRALoRA2
$^{4}$See Table 21 in Anil et al. (2023) regarding the distribution
of languages used in the pre-training of PaLM 2.
$^{5}$See Appendix A for details of language families in XLSum.
<FigureHere>
Figure 2: XLSum output examples: zero-shot transfer from English using Full FT and LoRA with PaLM 2-XXS. Full FT fails to generate summaries in the target language and the content is off-topic.
<table>
<tbody>
<tr>
<th> </th>
<th> </th>
<th colspan="3">$XLSUM$</th>
<th colspan="3">$\mathbf{XWIKIS}$</th>
</tr>
<tr>
<th> </th>
<th> </th>
<th>$R-L$</th>
<th>$NLI$</th>
<th>SH</th>
<th>$R-L$</th>
<th>$NLI$</th>
<th>SH</th>
</tr>
<tr>
<td rowspan="3">ZERO- SHOT</td>
<td>Full FT</td>
<td>14.48</td>
<td>28.87</td>
<td>13.71</td>
<td>20.22</td>
<td>30.17</td>
<td>16.26</td>
</tr>
<tr>
<td>LoRA- 4</td>
<td>22.59</td>
<td>37.39</td>
<td>24.21</td>
<td>28.46</td>
<td>48.31</td>
<td>26.40</td>
</tr>
<tr>
<td>Avg. LoRA</td>
<td>22.74</td>
<td>49.14</td>
<td>32.44</td>
<td>26.93</td>
<td>49.29</td>
<td>26.86</td>
</tr>
<tr>
<td rowspan="2">16.</td>
<td>Full FT $^{\prime}($CL)</td>
<td>22.31</td>
<td>30.15</td>
<td>18.79</td>
<td>26.90</td>
<td>34.17</td>
<td>21.82</td>
</tr>
<tr>
<td>LoRA- 4 1(CL)</td>
<td>24.71</td>
<td>41.12</td>
<td>26.47</td>
<td>30.05</td>
<td>45.90</td>
<td>28.20</td>
</tr>
<tr>
<td rowspan="1">16- SHOT</td>
<td>LoraHub</td>
<td>23.37</td>
<td>38.95</td>
<td>26.07</td>
<td>27.59</td>
<td>47.45</td>
<td>25.84</td>
</tr>
<tr>
<td rowspan="1">64.</td>
<td>Full FT $^{\prime}($CL)</td>
<td>24.30</td>
<td>30.65</td>
<td>19.57</td>
<td>28.73</td>
<td>39.42</td>
<td>24.16</td>
</tr>
<tr>
<td rowspan="2">W SHOT</td>
<td>LoRA- 4 $\vdash($CL)</td>
<td>25.94</td>
<td>42.07</td>
<td>27.66</td>
<td>31.08</td>
<td>45.12</td>
<td>28.05</td>
</tr>
<tr>
<td>LoraHub</td>
<td>24.21</td>
<td>41.34</td>
<td>28.02</td>
<td>27.66</td>
<td>48.09</td>
<td>26.56</td>
</tr>
</tbody>
</table>
Table 4: Cross-lingual transfer on 10 XLSum languages and fve XWikis languages (using leave-oneout training) for PaLM 2-XXS model. 16- and 64-shot experiments show average results from three different seed runs. For continued learning( CL ),we use a 14/2 and 60/4 training/validation split. Best results are in bold. Results for individual languages are in Tables 10 and 13 in Appendix C .
Figure 3 ROUGE -L SEEN UNSEEN UNSEEN
LoRA( 4) LoRA
Figure 3 iLoRAiiLoRA Avg.LoRA LoRAXLSum(ieRNGDSOYO)LoRALoRA(ie )(iii) XLSumZHJASWRNXWikis
Few-shot Cross-lingual Transfer LoRA
LLM (Brown et al., 2020; Wei et al., 2022b) LoraHub 3.1 LoraHub (Huang et al., 2023) LoRA(ie ) $N$ LoRALoraHub $1/N$
1664LoraHub14/260/4LoraHub Nevergrad 6
ROUGE- LROUGE- L
Table
1664XL-
416)
SumXWikis(I)(eg
7XLSum68ROUGE-
LXWikis7
et
al.
Lauscher
(2020)
iiXLSumLoraHubROUGE-L16iiiLoRAROUGE-
$^{6}$https://facebookresearch.github.io/nevergrad
|
[
[
0.2032470703125,
0.08612060546875,
0.2435302734375,
0.08612060546875,
0.2435302734375,
0.097412109375,
0.2032470703125,
0.097412109375
],
[
0.382080078125,
0.08612060546875,
0.420654296875,
0.08612060546875,
0.420654296875,
0.097412109375,
0.382080078125,
0.097412109375
],
[
0.57666015625,
0.08612060546875,
0.62060546875,
0.08612060546875,
0.62060546875,
0.097412109375,
0.57666015625,
0.097412109375
],
[
0.75390625,
0.08612060546875,
0.8193359375,
0.08612060546875,
0.8193359375,
0.097412109375,
0.75390625,
0.097412109375
],
[
0.1368408203125,
0.0968017578125,
0.1561279296875,
0.0968017578125,
0.1561279296875,
0.10986328125,
0.1368408203125,
0.10986328125
],
[
0.312255859375,
0.097412109375,
0.32568359375,
0.097412109375,
0.32568359375,
0.105712890625,
0.312255859375,
0.105712890625
],
[
0.1368408203125,
0.10687255859375,
0.1561279296875,
0.10687255859375,
0.1561279296875,
0.122314453125,
0.1368408203125,
0.122314453125
],
[
0.307373046875,
0.105712890625,
0.32666015625,
0.105712890625,
0.32666015625,
0.119384765625,
0.307373046875,
0.119384765625
],
[
0.51904296875,
0.11224365234375,
0.6767578125,
0.11224365234375,
0.6767578125,
0.122314453125,
0.51904296875,
0.122314453125
],
[
0.69775390625,
0.10986328125,
0.8759765625,
0.10986328125,
0.8759765625,
0.12054443359375,
0.69775390625,
0.12054443359375
],
[
0.6962890625,
0.119384765625,
0.87890625,
0.119384765625,
0.87890625,
0.1300048828125,
0.6962890625,
0.1300048828125
],
[
0.28955078125,
0.12054443359375,
0.334228515625,
0.12353515625,
0.330078125,
0.166259765625,
0.285400390625,
0.163330078125
],
[
0.118408203125,
0.11517333984375,
0.160400390625,
0.11517333984375,
0.160400390625,
0.163330078125,
0.118408203125,
0.163330078125
],
[
0.3046875,
0.122314453125,
0.318115234375,
0.114013671875,
0.330078125,
0.122314453125,
0.316650390625,
0.1312255859375
],
[
0.51708984375,
0.122314453125,
0.6767578125,
0.122314453125,
0.6767578125,
0.133056640625,
0.51708984375,
0.133056640625
],
[
0.52294921875,
0.1318359375,
0.67236328125,
0.1318359375,
0.67236328125,
0.142578125,
0.52294921875,
0.142578125
],
[
0.7060546875,
0.127685546875,
0.869140625,
0.127685546875,
0.869140625,
0.1383056640625,
0.7060546875,
0.1383056640625
],
[
0.54248046875,
0.141357421875,
0.654296875,
0.141357421875,
0.654296875,
0.1484375,
0.54248046875,
0.1484375
],
[
0.6962890625,
0.135986328125,
0.87744140625,
0.13720703125,
0.87744140625,
0.1484375,
0.6962890625,
0.1478271484375
],
[
0.744140625,
0.1478271484375,
0.82861328125,
0.1478271484375,
0.82861328125,
0.155029296875,
0.744140625,
0.155029296875
],
[
0.1318359375,
0.1561279296875,
0.154541015625,
0.1561279296875,
0.154541015625,
0.169189453125,
0.1318359375,
0.169189453125
],
[
0.2451171875,
0.160888671875,
0.2880859375,
0.160888671875,
0.2880859375,
0.171630859375,
0.2451171875,
0.171630859375
],
[
0.310546875,
0.1602783203125,
0.32568359375,
0.1602783203125,
0.32568359375,
0.17041015625,
0.310546875,
0.17041015625
],
[
0.418212890625,
0.160888671875,
0.47265625,
0.160888671875,
0.47265625,
0.171630859375,
0.418212890625,
0.171630859375
],
[
0.515625,
0.163330078125,
0.6767578125,
0.163330078125,
0.6767578125,
0.17333984375,
0.515625,
0.17333984375
],
[
0.69775390625,
0.1650390625,
0.8759765625,
0.16748046875,
0.8759765625,
0.1817626953125,
0.69775390625,
0.179931640625
],
[
0.30908203125,
0.17041015625,
0.32568359375,
0.17041015625,
0.32568359375,
0.18115234375,
0.30908203125,
0.18115234375
],
[
0.1376953125,
0.16748046875,
0.1695556640625,
0.169189453125,
0.16796875,
0.1817626953125,
0.1368408203125,
0.179931640625
],
[
0.2476806640625,
0.1728515625,
0.285400390625,
0.1728515625,
0.285400390625,
0.179931640625,
0.2476806640625,
0.179931640625
],
[
0.424072265625,
0.171630859375,
0.474365234375,
0.171630859375,
0.474365234375,
0.1817626953125,
0.424072265625,
0.1817626953125
],
[
0.5146484375,
0.17333984375,
0.68115234375,
0.17333984375,
0.68115234375,
0.18408203125,
0.5146484375,
0.18408203125
],
[
0.70361328125,
0.1776123046875,
0.87158203125,
0.1776123046875,
0.87158203125,
0.185302734375,
0.70361328125,
0.185302734375
],
[
0.382080078125,
0.185302734375,
0.47607421875,
0.185302734375,
0.47607421875,
0.1953125,
0.382080078125,
0.1953125
],
[
0.5205078125,
0.18408203125,
0.67333984375,
0.18408203125,
0.67333984375,
0.191162109375,
0.5205078125,
0.191162109375
],
[
0.35107421875,
0.1859130859375,
0.358642578125,
0.191162109375,
0.3525390625,
0.1953125,
0.344970703125,
0.190673828125
],
[
0.2225341796875,
0.18701171875,
0.2392578125,
0.18701171875,
0.2392578125,
0.1953125,
0.2225341796875,
0.1953125
],
[
0.240966796875,
0.18701171875,
0.2626953125,
0.18701171875,
0.2626953125,
0.1953125,
0.240966796875,
0.1953125
],
[
0.264404296875,
0.18701171875,
0.285400390625,
0.18701171875,
0.285400390625,
0.1953125,
0.264404296875,
0.1953125
],
[
0.378662109375,
0.189453125,
0.38623046875,
0.189453125,
0.38623046875,
0.193603515625,
0.378662109375,
0.193603515625
],
[
0.51708984375,
0.1923828125,
0.67333984375,
0.1923828125,
0.67333984375,
0.1995849609375,
0.51708984375,
0.1995849609375
],
[
0.70361328125,
0.188232421875,
0.869140625,
0.188232421875,
0.869140625,
0.1953125,
0.70361328125,
0.1953125
],
[
0.2032470703125,
0.19775390625,
0.2435302734375,
0.19775390625,
0.2435302734375,
0.208984375,
0.2032470703125,
0.208984375
],
[
0.382080078125,
0.19775390625,
0.42236328125,
0.19775390625,
0.42236328125,
0.208984375,
0.382080078125,
0.208984375
],
[
0.69775390625,
0.1953125,
0.87744140625,
0.1953125,
0.87744140625,
0.2060546875,
0.69775390625,
0.2060546875
],
[
0.310546875,
0.203125,
0.32568359375,
0.203125,
0.32568359375,
0.213134765625,
0.310546875,
0.213134765625
],
[
0.5205078125,
0.201904296875,
0.67333984375,
0.201904296875,
0.67333984375,
0.208984375,
0.5205078125,
0.208984375
],
[
0.736328125,
0.2060546875,
0.83642578125,
0.2060546875,
0.83642578125,
0.2161865234375,
0.736328125,
0.2161865234375
],
[
0.53662109375,
0.208984375,
0.6591796875,
0.208984375,
0.6591796875,
0.2197265625,
0.53662109375,
0.2197265625
],
[
0.1368408203125,
0.207275390625,
0.15283203125,
0.207275390625,
0.15283203125,
0.21728515625,
0.1368408203125,
0.21728515625
],
[
0.133544921875,
0.21435546875,
0.154541015625,
0.21435546875,
0.154541015625,
0.2286376953125,
0.133544921875,
0.2286376953125
],
[
0.301513671875,
0.2120361328125,
0.328369140625,
0.208984375,
0.333251953125,
0.2340087890625,
0.306396484375,
0.2369384765625
],
[
0.1368408203125,
0.22265625,
0.162109375,
0.2298583984375,
0.15283203125,
0.24462890625,
0.1292724609375,
0.2369384765625
],
[
0.7021484375,
0.22265625,
0.875,
0.22265625,
0.875,
0.236328125,
0.7021484375,
0.236328125
],
[
0.12176513671875,
0.2298583984375,
0.162841796875,
0.235107421875,
0.1561279296875,
0.2666015625,
0.11419677734375,
0.26123046875
],
[
0.307373046875,
0.2298583984375,
0.340087890625,
0.2298583984375,
0.340087890625,
0.2464599609375,
0.307373046875,
0.2464599609375
],
[
0.515625,
0.2298583984375,
0.6767578125,
0.23095703125,
0.6767578125,
0.2423095703125,
0.515625,
0.2410888671875
],
[
0.7001953125,
0.2327880859375,
0.873046875,
0.232177734375,
0.873046875,
0.2452392578125,
0.7001953125,
0.2464599609375
],
[
0.291259765625,
0.2392578125,
0.33154296875,
0.2392578125,
0.33154296875,
0.267333984375,
0.291259765625,
0.267333984375
],
[
0.5205078125,
0.2410888671875,
0.67236328125,
0.2410888671875,
0.67236328125,
0.251708984375,
0.5205078125,
0.251708984375
],
[
0.7109375,
0.243408203125,
0.861328125,
0.243408203125,
0.861328125,
0.254150390625,
0.7109375,
0.254150390625
],
[
0.54833984375,
0.25048828125,
0.6455078125,
0.25048828125,
0.6455078125,
0.26123046875,
0.54833984375,
0.26123046875
],
[
0.70361328125,
0.2529296875,
0.86962890625,
0.2529296875,
0.86962890625,
0.26318359375,
0.70361328125,
0.26318359375
],
[
0.1351318359375,
0.2587890625,
0.154541015625,
0.2587890625,
0.154541015625,
0.2724609375,
0.1351318359375,
0.2724609375
],
[
0.318115234375,
0.260009765625,
0.33154296875,
0.26953125,
0.318115234375,
0.279052734375,
0.3046875,
0.26953125
],
[
0.1351318359375,
0.268310546875,
0.154541015625,
0.268310546875,
0.154541015625,
0.281982421875,
0.1351318359375,
0.281982421875
],
[
0.72900390625,
0.26318359375,
0.84619140625,
0.26318359375,
0.84619140625,
0.270751953125,
0.72900390625,
0.270751953125
],
[
0.240966796875,
0.2724609375,
0.2880859375,
0.2724609375,
0.2880859375,
0.283935546875,
0.240966796875,
0.283935546875
],
[
0.310546875,
0.273681640625,
0.32568359375,
0.273681640625,
0.32568359375,
0.283935546875,
0.310546875,
0.283935546875
],
[
0.418212890625,
0.2724609375,
0.47265625,
0.2724609375,
0.47265625,
0.283935546875,
0.418212890625,
0.283935546875
],
[
0.1368408203125,
0.28076171875,
0.15283203125,
0.28076171875,
0.15283203125,
0.29150390625,
0.1368408203125,
0.29150390625
],
[
0.2333984375,
0.281982421875,
0.2880859375,
0.281982421875,
0.2880859375,
0.292236328125,
0.2333984375,
0.292236328125
],
[
0.313232421875,
0.28515625,
0.322509765625,
0.28515625,
0.322509765625,
0.29443359375,
0.313232421875,
0.29443359375
],
[
0.424072265625,
0.281982421875,
0.47265625,
0.281982421875,
0.47265625,
0.292236328125,
0.424072265625,
0.292236328125
],
[
0.509765625,
0.28515625,
0.8818359375,
0.28515625,
0.8818359375,
0.298583984375,
0.509765625,
0.298583984375
],
[
0.240966796875,
0.297607421875,
0.2626953125,
0.297607421875,
0.2626953125,
0.305908203125,
0.240966796875,
0.305908203125
],
[
0.264404296875,
0.297607421875,
0.285400390625,
0.297607421875,
0.285400390625,
0.305908203125,
0.264404296875,
0.305908203125
],
[
0.37451171875,
0.29638671875,
0.476806640625,
0.29638671875,
0.476806640625,
0.30712890625,
0.37451171875,
0.30712890625
],
[
0.51123046875,
0.2998046875,
0.8818359375,
0.2998046875,
0.8818359375,
0.3134765625,
0.51123046875,
0.3134765625
],
[
0.382080078125,
0.30810546875,
0.420654296875,
0.30810546875,
0.420654296875,
0.319580078125,
0.382080078125,
0.319580078125
],
[
0.2275390625,
0.2998046875,
0.2376708984375,
0.2998046875,
0.2376708984375,
0.3046875,
0.2275390625,
0.3046875
],
[
0.204833984375,
0.308837890625,
0.2418212890625,
0.308837890625,
0.2418212890625,
0.318359375,
0.204833984375,
0.318359375
],
[
0.307373046875,
0.314208984375,
0.32666015625,
0.314208984375,
0.32666015625,
0.328857421875,
0.307373046875,
0.328857421875
],
[
0.509765625,
0.3134765625,
0.88330078125,
0.3154296875,
0.88330078125,
0.328857421875,
0.509765625,
0.32666015625
],
[
0.307373046875,
0.325927734375,
0.32666015625,
0.325927734375,
0.32666015625,
0.34033203125,
0.307373046875,
0.34033203125
],
[
0.139404296875,
0.3154296875,
0.15283203125,
0.3154296875,
0.15283203125,
0.32666015625,
0.139404296875,
0.32666015625
],
[
0.139404296875,
0.330810546875,
0.15283203125,
0.330810546875,
0.15283203125,
0.342529296875,
0.139404296875,
0.342529296875
],
[
0.2880859375,
0.334228515625,
0.334228515625,
0.334228515625,
0.334228515625,
0.392578125,
0.2880859375,
0.392578125
],
[
0.51123046875,
0.328857421875,
0.78857421875,
0.328857421875,
0.78857421875,
0.342529296875,
0.51123046875,
0.342529296875
],
[
0.1158447265625,
0.3349609375,
0.16796875,
0.3349609375,
0.16796875,
0.391357421875,
0.1158447265625,
0.391357421875
],
[
0.67236328125,
0.359375,
0.72119140625,
0.359375,
0.72119140625,
0.37353515625,
0.67236328125,
0.37353515625
],
[
0.79443359375,
0.359375,
0.84033203125,
0.361083984375,
0.8388671875,
0.372314453125,
0.79248046875,
0.37060546875
],
[
0.64208984375,
0.371826171875,
0.6708984375,
0.371826171875,
0.6708984375,
0.383056640625,
0.64208984375,
0.383056640625
],
[
0.68115234375,
0.371826171875,
0.7109375,
0.371826171875,
0.7109375,
0.383056640625,
0.68115234375,
0.383056640625
],
[
0.76318359375,
0.371826171875,
0.7900390625,
0.371826171875,
0.7900390625,
0.383056640625,
0.76318359375,
0.383056640625
],
[
0.8017578125,
0.371826171875,
0.8330078125,
0.371826171875,
0.8330078125,
0.383056640625,
0.8017578125,
0.383056640625
],
[
0.84814453125,
0.372314453125,
0.8671875,
0.372314453125,
0.8671875,
0.3818359375,
0.84814453125,
0.3818359375
],
[
0.139404296875,
0.38134765625,
0.151123046875,
0.38134765625,
0.151123046875,
0.391357421875,
0.139404296875,
0.391357421875
],
[
0.72705078125,
0.372314453125,
0.74658203125,
0.372314453125,
0.74658203125,
0.3818359375,
0.72705078125,
0.3818359375
],
[
0.2333984375,
0.383056640625,
0.285400390625,
0.383056640625,
0.285400390625,
0.393798828125,
0.2333984375,
0.393798828125
],
[
0.316650390625,
0.385498046875,
0.328369140625,
0.392578125,
0.316650390625,
0.402587890625,
0.3046875,
0.3955078125
],
[
0.418212890625,
0.383056640625,
0.470947265625,
0.383056640625,
0.470947265625,
0.393798828125,
0.418212890625,
0.393798828125
],
[
0.640625,
0.385498046875,
0.67333984375,
0.385498046875,
0.67333984375,
0.39794921875,
0.640625,
0.39794921875
],
[
0.67919921875,
0.385498046875,
0.7119140625,
0.385498046875,
0.7119140625,
0.396728515625,
0.67919921875,
0.396728515625
],
[
0.75830078125,
0.385498046875,
0.79248046875,
0.385498046875,
0.79248046875,
0.396728515625,
0.75830078125,
0.396728515625
],
[
0.8017578125,
0.385498046875,
0.83447265625,
0.385498046875,
0.83447265625,
0.396728515625,
0.8017578125,
0.396728515625
],
[
0.84033203125,
0.385498046875,
0.8759765625,
0.385498046875,
0.8759765625,
0.396728515625,
0.84033203125,
0.396728515625
],
[
0.72314453125,
0.38720703125,
0.75146484375,
0.38720703125,
0.75146484375,
0.3955078125,
0.72314453125,
0.3955078125
],
[
0.5634765625,
0.385986328125,
0.60546875,
0.385986328125,
0.60546875,
0.396728515625,
0.5634765625,
0.396728515625
],
[
0.2435302734375,
0.393798828125,
0.287109375,
0.391357421875,
0.2880859375,
0.402587890625,
0.2451171875,
0.405029296875
],
[
0.429931640625,
0.393798828125,
0.474365234375,
0.393798828125,
0.474365234375,
0.40380859375,
0.429931640625,
0.40380859375
],
[
0.5224609375,
0.392578125,
0.55517578125,
0.392578125,
0.55517578125,
0.40087890625,
0.5224609375,
0.40087890625
],
[
0.2359619140625,
0.3955078125,
0.246826171875,
0.3955078125,
0.246826171875,
0.40087890625,
0.2359619140625,
0.40087890625
],
[
0.1435546875,
0.394287109375,
0.15869140625,
0.40380859375,
0.1485595703125,
0.412109375,
0.133544921875,
0.40380859375
],
[
0.6396484375,
0.3955078125,
0.67529296875,
0.3955078125,
0.67529296875,
0.407958984375,
0.6396484375,
0.407958984375
],
[
0.67919921875,
0.3955078125,
0.7138671875,
0.3955078125,
0.7138671875,
0.40673828125,
0.67919921875,
0.40673828125
],
[
0.75830078125,
0.3955078125,
0.79443359375,
0.3955078125,
0.79443359375,
0.40673828125,
0.75830078125,
0.40673828125
],
[
0.72314453125,
0.39794921875,
0.75146484375,
0.39794921875,
0.75146484375,
0.40625,
0.72314453125,
0.40625
],
[
0.5634765625,
0.396728515625,
0.60791015625,
0.396728515625,
0.60791015625,
0.40673828125,
0.5634765625,
0.40673828125
],
[
0.8037109375,
0.39794921875,
0.8310546875,
0.39794921875,
0.8310546875,
0.40625,
0.8037109375,
0.40625
],
[
0.84228515625,
0.39794921875,
0.873046875,
0.39794921875,
0.873046875,
0.40625,
0.84228515625,
0.40625
],
[
0.5205078125,
0.402099609375,
0.55322265625,
0.402099609375,
0.55322265625,
0.413330078125,
0.5205078125,
0.413330078125
],
[
0.316650390625,
0.399658203125,
0.32568359375,
0.40625,
0.316650390625,
0.412109375,
0.307373046875,
0.40625
],
[
0.364501953125,
0.40625,
0.476806640625,
0.407958984375,
0.476806640625,
0.418701171875,
0.364501953125,
0.416259765625
],
[
0.5634765625,
0.40673828125,
0.62646484375,
0.40673828125,
0.62646484375,
0.418701171875,
0.5634765625,
0.418701171875
],
[
0.6396484375,
0.40673828125,
0.67333984375,
0.40673828125,
0.67333984375,
0.418701171875,
0.6396484375,
0.418701171875
],
[
0.67919921875,
0.40673828125,
0.7138671875,
0.40673828125,
0.7138671875,
0.418701171875,
0.67919921875,
0.418701171875
],
[
0.72119140625,
0.40673828125,
0.755859375,
0.40673828125,
0.755859375,
0.418701171875,
0.72119140625,
0.418701171875
],
[
0.75830078125,
0.40673828125,
0.79443359375,
0.40673828125,
0.79443359375,
0.418701171875,
0.75830078125,
0.418701171875
],
[
0.80029296875,
0.40673828125,
0.83447265625,
0.40673828125,
0.83447265625,
0.418701171875,
0.80029296875,
0.418701171875
],
[
0.8388671875,
0.40673828125,
0.875,
0.40673828125,
0.875,
0.418701171875,
0.8388671875,
0.418701171875
],
[
0.264404296875,
0.4091796875,
0.28369140625,
0.4091796875,
0.28369140625,
0.416259765625,
0.264404296875,
0.416259765625
],
[
0.225830078125,
0.410400390625,
0.2392578125,
0.410400390625,
0.2392578125,
0.416259765625,
0.225830078125,
0.416259765625
],
[
0.240966796875,
0.4091796875,
0.2626953125,
0.4091796875,
0.2626953125,
0.416259765625,
0.240966796875,
0.416259765625
],
[
0.349365234375,
0.410400390625,
0.35693359375,
0.410400390625,
0.35693359375,
0.41455078125,
0.349365234375,
0.41455078125
],
[
0.177978515625,
0.410888671875,
0.18896484375,
0.410888671875,
0.18896484375,
0.4150390625,
0.177978515625,
0.4150390625
],
[
0.15869140625,
0.410888671875,
0.166259765625,
0.410888671875,
0.166259765625,
0.4150390625,
0.15869140625,
0.4150390625
],
[
0.204833984375,
0.410888671875,
0.21240234375,
0.410888671875,
0.21240234375,
0.41455078125,
0.204833984375,
0.41455078125
],
[
0.219970703125,
0.419189453125,
0.3837890625,
0.419189453125,
0.3837890625,
0.429931640625,
0.219970703125,
0.429931640625
],
[
0.5634765625,
0.421630859375,
0.62890625,
0.421630859375,
0.62890625,
0.432861328125,
0.5634765625,
0.432861328125
],
[
0.6396484375,
0.421630859375,
0.67236328125,
0.421630859375,
0.67236328125,
0.432861328125,
0.6396484375,
0.432861328125
],
[
0.68115234375,
0.421630859375,
0.7119140625,
0.421630859375,
0.7119140625,
0.432861328125,
0.68115234375,
0.432861328125
],
[
0.72119140625,
0.421630859375,
0.75390625,
0.421630859375,
0.75390625,
0.432861328125,
0.72119140625,
0.432861328125
],
[
0.759765625,
0.421630859375,
0.79248046875,
0.421630859375,
0.79248046875,
0.432861328125,
0.759765625,
0.432861328125
],
[
0.80029296875,
0.421630859375,
0.8330078125,
0.421630859375,
0.8330078125,
0.432861328125,
0.80029296875,
0.432861328125
],
[
0.84033203125,
0.421630859375,
0.875,
0.421630859375,
0.875,
0.432861328125,
0.84033203125,
0.432861328125
],
[
0.5615234375,
0.431640625,
0.6337890625,
0.431640625,
0.6337890625,
0.4423828125,
0.5615234375,
0.4423828125
],
[
0.5205078125,
0.42578125,
0.5439453125,
0.42578125,
0.5439453125,
0.438232421875,
0.5205078125,
0.438232421875
],
[
0.6396484375,
0.431640625,
0.67333984375,
0.431640625,
0.67333984375,
0.443603515625,
0.6396484375,
0.443603515625
],
[
0.67919921875,
0.431640625,
0.7138671875,
0.431640625,
0.7138671875,
0.443603515625,
0.67919921875,
0.443603515625
],
[
0.72119140625,
0.432861328125,
0.75390625,
0.432861328125,
0.75390625,
0.441162109375,
0.72119140625,
0.441162109375
],
[
0.759765625,
0.432861328125,
0.79248046875,
0.432861328125,
0.79248046875,
0.441162109375,
0.759765625,
0.441162109375
],
[
0.8017578125,
0.432861328125,
0.83447265625,
0.432861328125,
0.83447265625,
0.441162109375,
0.8017578125,
0.441162109375
],
[
0.8388671875,
0.431640625,
0.875,
0.431640625,
0.875,
0.443603515625,
0.8388671875,
0.443603515625
],
[
0.5224609375,
0.439453125,
0.5517578125,
0.439453125,
0.5517578125,
0.44775390625,
0.5224609375,
0.44775390625
],
[
0.640625,
0.441162109375,
0.67333984375,
0.441162109375,
0.67333984375,
0.45361328125,
0.640625,
0.45361328125
],
[
0.759765625,
0.441162109375,
0.79248046875,
0.441162109375,
0.79248046875,
0.45361328125,
0.759765625,
0.45361328125
],
[
0.68115234375,
0.4423828125,
0.7138671875,
0.4423828125,
0.7138671875,
0.45361328125,
0.68115234375,
0.45361328125
],
[
0.5634765625,
0.4423828125,
0.61279296875,
0.4423828125,
0.61279296875,
0.45361328125,
0.5634765625,
0.45361328125
],
[
0.72119140625,
0.4423828125,
0.755859375,
0.4423828125,
0.755859375,
0.45361328125,
0.72119140625,
0.45361328125
],
[
0.8017578125,
0.4423828125,
0.8330078125,
0.4423828125,
0.8330078125,
0.45361328125,
0.8017578125,
0.45361328125
],
[
0.84033203125,
0.4423828125,
0.8759765625,
0.4423828125,
0.8759765625,
0.454833984375,
0.84033203125,
0.454833984375
],
[
0.63818359375,
0.4560546875,
0.67919921875,
0.4560546875,
0.67919921875,
0.4697265625,
0.63818359375,
0.4697265625
],
[
0.1158447265625,
0.446533203125,
0.486083984375,
0.446533203125,
0.486083984375,
0.460205078125,
0.1158447265625,
0.460205078125
],
[
0.5615234375,
0.457275390625,
0.63037109375,
0.457275390625,
0.63037109375,
0.46728515625,
0.5615234375,
0.46728515625
],
[
0.67919921875,
0.4560546875,
0.7138671875,
0.4560546875,
0.7138671875,
0.46728515625,
0.67919921875,
0.46728515625
],
[
0.72119140625,
0.457275390625,
0.79443359375,
0.457275390625,
0.79443359375,
0.468505859375,
0.72119140625,
0.468505859375
],
[
0.80029296875,
0.457763671875,
0.83447265625,
0.457763671875,
0.83447265625,
0.466064453125,
0.80029296875,
0.466064453125
],
[
0.8369140625,
0.457275390625,
0.8759765625,
0.457275390625,
0.8759765625,
0.468505859375,
0.8369140625,
0.468505859375
],
[
0.5224609375,
0.463134765625,
0.5439453125,
0.463134765625,
0.5439453125,
0.473876953125,
0.5224609375,
0.473876953125
],
[
0.6396484375,
0.465576171875,
0.67529296875,
0.46728515625,
0.67333984375,
0.479736328125,
0.63818359375,
0.47802734375
],
[
0.56103515625,
0.466064453125,
0.6337890625,
0.468505859375,
0.6337890625,
0.479736328125,
0.56103515625,
0.47802734375
],
[
0.1158447265625,
0.46142578125,
0.48779296875,
0.46142578125,
0.48779296875,
0.474365234375,
0.1158447265625,
0.474365234375
],
[
0.68115234375,
0.46728515625,
0.7138671875,
0.46728515625,
0.7138671875,
0.478515625,
0.68115234375,
0.478515625
],
[
0.72119140625,
0.46728515625,
0.75390625,
0.46728515625,
0.75390625,
0.478515625,
0.72119140625,
0.478515625
],
[
0.75830078125,
0.46728515625,
0.79443359375,
0.46728515625,
0.79443359375,
0.478515625,
0.75830078125,
0.478515625
],
[
0.80029296875,
0.46728515625,
0.83642578125,
0.46728515625,
0.83642578125,
0.478515625,
0.80029296875,
0.478515625
],
[
0.8369140625,
0.46728515625,
0.873046875,
0.46728515625,
0.873046875,
0.478515625,
0.8369140625,
0.478515625
],
[
0.11419677734375,
0.474365234375,
0.48779296875,
0.473876953125,
0.48779296875,
0.488037109375,
0.11419677734375,
0.4892578125
],
[
0.5224609375,
0.474365234375,
0.5517578125,
0.474365234375,
0.5517578125,
0.482666015625,
0.5224609375,
0.482666015625
],
[
0.6396484375,
0.47802734375,
0.67529296875,
0.47802734375,
0.67529296875,
0.4892578125,
0.6396484375,
0.4892578125
],
[
0.67919921875,
0.47802734375,
0.7138671875,
0.47802734375,
0.7138671875,
0.4892578125,
0.67919921875,
0.4892578125
],
[
0.72119140625,
0.47802734375,
0.79443359375,
0.47802734375,
0.79443359375,
0.4892578125,
0.72119140625,
0.4892578125
],
[
0.796875,
0.47802734375,
0.83642578125,
0.47802734375,
0.83642578125,
0.4892578125,
0.796875,
0.4892578125
],
[
0.8388671875,
0.47802734375,
0.8759765625,
0.47802734375,
0.8759765625,
0.4892578125,
0.8388671875,
0.4892578125
],
[
0.1158447265625,
0.490478515625,
0.486083984375,
0.490478515625,
0.486083984375,
0.50439453125,
0.1158447265625,
0.50439453125
],
[
0.5615234375,
0.478515625,
0.61279296875,
0.478515625,
0.61279296875,
0.4892578125,
0.5615234375,
0.4892578125
],
[
0.1158447265625,
0.50439453125,
0.435791015625,
0.50439453125,
0.435791015625,
0.51708984375,
0.1158447265625,
0.51708984375
],
[
0.51123046875,
0.50439453125,
0.58447265625,
0.50439453125,
0.58447265625,
0.51416015625,
0.51123046875,
0.51416015625
],
[
0.580078125,
0.5048828125,
0.8818359375,
0.5048828125,
0.8818359375,
0.515625,
0.580078125,
0.515625
],
[
0.51123046875,
0.51708984375,
0.88330078125,
0.51708984375,
0.88330078125,
0.53076171875,
0.51123046875,
0.53076171875
],
[
0.173828125,
0.53515625,
0.2032470703125,
0.53515625,
0.2032470703125,
0.54638671875,
0.173828125,
0.54638671875
],
[
0.28369140625,
0.53515625,
0.33154296875,
0.53515625,
0.33154296875,
0.54638671875,
0.28369140625,
0.54638671875
],
[
0.4013671875,
0.53515625,
0.44921875,
0.53515625,
0.44921875,
0.54638671875,
0.4013671875,
0.54638671875
],
[
0.509765625,
0.5322265625,
0.8818359375,
0.53076171875,
0.8818359375,
0.54443359375,
0.509765625,
0.5458984375
],
[
0.2509765625,
0.5458984375,
0.279541015625,
0.5458984375,
0.279541015625,
0.55712890625,
0.2509765625,
0.55712890625
],
[
0.337646484375,
0.54345703125,
0.359375,
0.54638671875,
0.355224609375,
0.55859375,
0.333251953125,
0.5556640625
],
[
0.37451171875,
0.5458984375,
0.4013671875,
0.5458984375,
0.4013671875,
0.55810546875,
0.37451171875,
0.55810546875
],
[
0.414794921875,
0.5458984375,
0.444091796875,
0.5458984375,
0.444091796875,
0.55810546875,
0.414794921875,
0.55810546875
],
[
0.29296875,
0.54638671875,
0.32080078125,
0.54638671875,
0.32080078125,
0.55810546875,
0.29296875,
0.55810546875
],
[
0.175537109375,
0.54638671875,
0.231689453125,
0.54638671875,
0.231689453125,
0.55810546875,
0.175537109375,
0.55810546875
],
[
0.459228515625,
0.54638671875,
0.478515625,
0.54638671875,
0.478515625,
0.5556640625,
0.459228515625,
0.5556640625
],
[
0.51318359375,
0.54736328125,
0.880859375,
0.54736328125,
0.880859375,
0.55810546875,
0.51318359375,
0.55810546875
],
[
0.12176513671875,
0.56005859375,
0.162841796875,
0.56005859375,
0.162841796875,
0.57177734375,
0.12176513671875,
0.57177734375
],
[
0.2509765625,
0.56005859375,
0.28369140625,
0.56005859375,
0.28369140625,
0.57177734375,
0.2509765625,
0.57177734375
],
[
0.29296875,
0.56005859375,
0.32421875,
0.56005859375,
0.32421875,
0.57177734375,
0.29296875,
0.57177734375
],
[
0.370361328125,
0.56005859375,
0.40478515625,
0.56005859375,
0.40478515625,
0.572265625,
0.370361328125,
0.572265625
],
[
0.411376953125,
0.56005859375,
0.450927734375,
0.56005859375,
0.450927734375,
0.57177734375,
0.411376953125,
0.57177734375
],
[
0.44921875,
0.56005859375,
0.484375,
0.56005859375,
0.484375,
0.57177734375,
0.44921875,
0.57177734375
],
[
0.509765625,
0.55859375,
0.880859375,
0.55859375,
0.880859375,
0.572265625,
0.509765625,
0.572265625
],
[
0.173828125,
0.56640625,
0.2418212890625,
0.56640625,
0.2418212890625,
0.57666015625,
0.173828125,
0.57666015625
],
[
0.3359375,
0.5625,
0.364501953125,
0.5625,
0.364501953125,
0.57080078125,
0.3359375,
0.57080078125
],
[
0.12176513671875,
0.57177734375,
0.16796875,
0.57177734375,
0.16796875,
0.5830078125,
0.12176513671875,
0.5830078125
],
[
0.2493896484375,
0.57177734375,
0.28369140625,
0.57177734375,
0.28369140625,
0.5830078125,
0.2493896484375,
0.5830078125
],
[
0.291259765625,
0.57177734375,
0.32421875,
0.57177734375,
0.32421875,
0.5830078125,
0.291259765625,
0.5830078125
],
[
0.330078125,
0.57177734375,
0.365234375,
0.57177734375,
0.365234375,
0.5830078125,
0.330078125,
0.5830078125
],
[
0.370361328125,
0.57177734375,
0.405517578125,
0.57177734375,
0.405517578125,
0.5830078125,
0.370361328125,
0.5830078125
],
[
0.4130859375,
0.57177734375,
0.447509765625,
0.57177734375,
0.447509765625,
0.5830078125,
0.4130859375,
0.5830078125
],
[
0.45166015625,
0.5693359375,
0.484375,
0.57177734375,
0.483642578125,
0.58447265625,
0.450927734375,
0.58203125
],
[
0.51123046875,
0.57470703125,
0.8818359375,
0.57470703125,
0.8818359375,
0.58837890625,
0.51123046875,
0.58837890625
],
[
0.12176513671875,
0.58447265625,
0.162841796875,
0.5859375,
0.162109375,
0.5986328125,
0.12005615234375,
0.5966796875
],
[
0.2476806640625,
0.5859375,
0.2822265625,
0.5859375,
0.2822265625,
0.59716796875,
0.2476806640625,
0.59716796875
],
[
0.2880859375,
0.5859375,
0.32421875,
0.5859375,
0.32421875,
0.59716796875,
0.2880859375,
0.59716796875
],
[
0.370361328125,
0.5859375,
0.405517578125,
0.5859375,
0.405517578125,
0.59716796875,
0.370361328125,
0.59716796875
],
[
0.453369140625,
0.5849609375,
0.484375,
0.5849609375,
0.484375,
0.5966796875,
0.453369140625,
0.5966796875
],
[
0.4130859375,
0.58740234375,
0.44580078125,
0.58740234375,
0.44580078125,
0.595703125,
0.4130859375,
0.595703125
],
[
0.333251953125,
0.58740234375,
0.364501953125,
0.58740234375,
0.364501953125,
0.595703125,
0.333251953125,
0.595703125
],
[
0.51123046875,
0.58837890625,
0.88330078125,
0.58837890625,
0.88330078125,
0.6025390625,
0.51123046875,
0.6025390625
],
[
0.12176513671875,
0.59716796875,
0.16796875,
0.59716796875,
0.16796875,
0.60791015625,
0.12176513671875,
0.60791015625
],
[
0.173828125,
0.59130859375,
0.2166748046875,
0.59130859375,
0.2166748046875,
0.6025390625,
0.173828125,
0.6025390625
],
[
0.246826171875,
0.59716796875,
0.32568359375,
0.59716796875,
0.32568359375,
0.609375,
0.246826171875,
0.609375
],
[
0.32080078125,
0.59716796875,
0.366943359375,
0.59716796875,
0.366943359375,
0.60791015625,
0.32080078125,
0.60791015625
],
[
0.366943359375,
0.59716796875,
0.4072265625,
0.59716796875,
0.4072265625,
0.609375,
0.366943359375,
0.609375
],
[
0.41064453125,
0.59716796875,
0.455078125,
0.59716796875,
0.455078125,
0.60791015625,
0.41064453125,
0.60791015625
],
[
0.45166015625,
0.59716796875,
0.483642578125,
0.59716796875,
0.483642578125,
0.609375,
0.45166015625,
0.609375
],
[
0.509765625,
0.60107421875,
0.67919921875,
0.6015625,
0.67919921875,
0.6162109375,
0.509765625,
0.615234375
],
[
0.11419677734375,
0.62158203125,
0.48779296875,
0.62158203125,
0.48779296875,
0.63525390625,
0.11419677734375,
0.63525390625
],
[
0.1175537109375,
0.63818359375,
0.484375,
0.63818359375,
0.484375,
0.6484375,
0.1175537109375,
0.6484375
],
[
0.53076171875,
0.64453125,
0.8818359375,
0.64453125,
0.8818359375,
0.65771484375,
0.53076171875,
0.65771484375
],
[
0.1158447265625,
0.65087890625,
0.486083984375,
0.65087890625,
0.486083984375,
0.66455078125,
0.1158447265625,
0.66455078125
],
[
0.11419677734375,
0.66455078125,
0.483642578125,
0.66455078125,
0.483642578125,
0.677734375,
0.11419677734375,
0.677734375
],
[
0.51123046875,
0.66015625,
0.62890625,
0.66015625,
0.62890625,
0.67333984375,
0.51123046875,
0.67333984375
],
[
0.118408203125,
0.68115234375,
0.48779296875,
0.68115234375,
0.48779296875,
0.69140625,
0.118408203125,
0.69140625
],
[
0.53564453125,
0.67578125,
0.88330078125,
0.6767578125,
0.88330078125,
0.69140625,
0.53564453125,
0.6904296875
],
[
0.1158447265625,
0.693359375,
0.48779296875,
0.693359375,
0.48779296875,
0.70703125,
0.1158447265625,
0.70703125
],
[
0.51123046875,
0.693359375,
0.88330078125,
0.693359375,
0.88330078125,
0.70703125,
0.51123046875,
0.70703125
],
[
0.51123046875,
0.708984375,
0.7021484375,
0.708984375,
0.7021484375,
0.72265625,
0.51123046875,
0.72265625
],
[
0.133544921875,
0.73291015625,
0.48779296875,
0.73095703125,
0.48779296875,
0.7490234375,
0.133544921875,
0.75048828125
],
[
0.515625,
0.7353515625,
0.83447265625,
0.736328125,
0.83447265625,
0.75048828125,
0.515625,
0.75
],
[
0.12005615234375,
0.751953125,
0.235107421875,
0.751953125,
0.235107421875,
0.76416015625,
0.12005615234375,
0.76416015625
],
[
0.52880859375,
0.7529296875,
0.880859375,
0.7529296875,
0.880859375,
0.7666015625,
0.52880859375,
0.7666015625
],
[
0.509765625,
0.7685546875,
0.8818359375,
0.76708984375,
0.8818359375,
0.7822265625,
0.509765625,
0.783203125
],
[
0.1351318359375,
0.779296875,
0.2149658203125,
0.779296875,
0.2149658203125,
0.79052734375,
0.1351318359375,
0.79052734375
],
[
0.21240234375,
0.77783203125,
0.489501953125,
0.77783203125,
0.489501953125,
0.79150390625,
0.21240234375,
0.79150390625
],
[
0.51123046875,
0.78515625,
0.765625,
0.78515625,
0.765625,
0.798828125,
0.51123046875,
0.798828125
],
[
0.1175537109375,
0.7958984375,
0.173828125,
0.7958984375,
0.173828125,
0.80712890625,
0.1175537109375,
0.80712890625
],
[
0.53076171875,
0.80029296875,
0.873046875,
0.7998046875,
0.873046875,
0.81396484375,
0.53076171875,
0.8154296875
],
[
0.875,
0.7998046875,
0.8818359375,
0.8046875,
0.875,
0.81005859375,
0.8671875,
0.8046875
],
[
0.1351318359375,
0.81005859375,
0.484375,
0.81005859375,
0.484375,
0.82373046875,
0.1351318359375,
0.82373046875
],
[
0.509765625,
0.81787109375,
0.67236328125,
0.81787109375,
0.67236328125,
0.83056640625,
0.509765625,
0.83056640625
],
[
0.1158447265625,
0.82666015625,
0.403076171875,
0.82666015625,
0.403076171875,
0.84033203125,
0.1158447265625,
0.84033203125
],
[
0.393798828125,
0.8291015625,
0.435791015625,
0.8291015625,
0.435791015625,
0.8388671875,
0.393798828125,
0.8388671875
],
[
0.426513671875,
0.82666015625,
0.484375,
0.82373046875,
0.486083984375,
0.83837890625,
0.42822265625,
0.84130859375
],
[
0.5283203125,
0.83447265625,
0.5751953125,
0.83447265625,
0.5751953125,
0.8486328125,
0.5283203125,
0.8486328125
],
[
0.68701171875,
0.8359375,
0.7001953125,
0.8359375,
0.7001953125,
0.84716796875,
0.68701171875,
0.84716796875
],
[
0.81689453125,
0.8359375,
0.88330078125,
0.8359375,
0.88330078125,
0.84716796875,
0.81689453125,
0.84716796875
],
[
0.1158447265625,
0.8427734375,
0.17626953125,
0.8427734375,
0.17626953125,
0.85693359375,
0.1158447265625,
0.85693359375
],
[
0.51123046875,
0.85107421875,
0.646484375,
0.85107421875,
0.646484375,
0.865234375,
0.51123046875,
0.865234375
],
[
0.67333984375,
0.8515625,
0.7021484375,
0.8515625,
0.7021484375,
0.8642578125,
0.67333984375,
0.8642578125
],
[
0.72705078125,
0.85107421875,
0.8818359375,
0.85107421875,
0.8818359375,
0.8642578125,
0.72705078125,
0.8642578125
],
[
0.133544921875,
0.859375,
0.28955078125,
0.859375,
0.28955078125,
0.87255859375,
0.133544921875,
0.87255859375
],
[
0.509765625,
0.8642578125,
0.59375,
0.8662109375,
0.59375,
0.88134765625,
0.509765625,
0.87890625
],
[
0.7734375,
0.86767578125,
0.79833984375,
0.86767578125,
0.79833984375,
0.8798828125,
0.7734375,
0.8798828125
],
[
0.8310546875,
0.8642578125,
0.8818359375,
0.8662109375,
0.880859375,
0.88134765625,
0.82958984375,
0.87890625
],
[
0.72705078125,
0.869140625,
0.74462890625,
0.869140625,
0.74462890625,
0.87890625,
0.72705078125,
0.87890625
],
[
0.62451171875,
0.8681640625,
0.6943359375,
0.8681640625,
0.6943359375,
0.8798828125,
0.62451171875,
0.8798828125
],
[
0.133544921875,
0.88134765625,
0.48779296875,
0.88134765625,
0.48779296875,
0.89453125,
0.133544921875,
0.89453125
],
[
0.509765625,
0.8828125,
0.8759765625,
0.8828125,
0.8759765625,
0.896484375,
0.509765625,
0.896484375
],
[
0.1158447265625,
0.8935546875,
0.3955078125,
0.8935546875,
0.3955078125,
0.90673828125,
0.1158447265625,
0.90673828125
],
[
0.133544921875,
0.90380859375,
0.48779296875,
0.90478515625,
0.48779296875,
0.91943359375,
0.133544921875,
0.91845703125
],
[
0.5322265625,
0.90673828125,
0.85400390625,
0.90673828125,
0.85400390625,
0.91748046875,
0.5322265625,
0.91748046875
]
] |
[
"XLSum",
"xWikis",
"Hausa",
"Indonesian",
"31",
"34",
"30",
"33",
"Target:Gwamnatin Najeriya ta",
"Target:Perempuan Vietnam yang",
"dituding terlibat dalam pembunuhan",
"",
"A6",
"4",
"ce yan kasar sun ga irin amfani",
"da rufe iyakokin kasar ya yi a",
"Kim Jong-nam, saudara tiri dari",
"fannintattalinarzikinta",
"pemimpin Korea Utara Kim Jong-un,",
"telahdibebaskan",
"26",
"FullFT",
"28",
"X-Full FT",
"Full FT:PresidentMuhammadu",
"Full FT:Kim Jong-namthewifeof",
"27",
"25X",
"LORA-4",
"LORA-4",
"Buhari has appointed his deputy.",
"NorthKoreanleaderKimJong-un.",
"1012141618",
"theBBCpresenterandformer",
"6",
"10",
"12",
"14",
"",
"ministerShugabaMuhammadi",
"hasdiedinafightwithMalaysia",
"XLSum",
"xwikis",
"Airlines flight MHi7. Here are the key",
"48",
"Buhari,asthenewchairmanof",
"points of the ruling:",
"the Presidential Council.",
"54",
"52",
"8",
"50",
"LoRA-4: Seorang wanita Vietnam",
"",
"42",
"LoRA-4: Gwamnatin Nijeriya ta",
"yang didakwa sebagai bagian dari",
"",
"yi tsokacin da shawarar da zai",
" pembunuhan Kim Jong-nam,",
" rufe iyakokin kasar.",
" saudara tiri dari pemimpin Korea",
"44",
"o",
"42",
"Utara,telahdibebaskan.",
"Full FT",
"34",
"Full FT",
"40",
"+LORA-4",
"32",
"LORA-4",
"Figure 2: XLSum output examples: zero-shot trans-",
"12",
"14",
"1012141618",
"fer from English using Full FT and LoRA with PaLM",
"xwikis",
"10",
"XLSum",
"28",
"2-XXS.Full FT fails to generate summaries in the tar-",
"27",
"36",
"34",
"B8MS",
"get language and the content is off-topic.",
"ON",
"XLSUM",
"XWIKIS",
"R-L",
"NLI",
"R-L",
"NLI",
"SH",
"28",
"SH",
"x-FullFT",
"2",
"*Full FT",
"14.48",
"28.87",
"20.22",
"30.17",
"16.26",
"13.71",
"FullFT",
"LORA-4",
"LORA-4",
"ZERO-",
"+",
"26",
"22.59",
"37.39",
"28.46",
"24.21",
"LoRA-4",
"48.31",
"26.40",
"SHOT",
"2",
"81012141618",
"Avg. LoRA",
"22.74",
"49.14",
"32.44",
"26.93",
"49.29",
"26.86",
"14",
"10",
"12",
"6",
"6",
"4",
"R",
"log2of Examples per Language",
"Full FT (CL)",
"22.31",
"30.15",
"18.79",
"26.90",
"34.17",
"21.82",
"LORA-4 (CL)",
"16-",
"24.71",
"41.12",
"26.47",
"30.05",
"45.90",
"28.20",
"SHOT",
"23.37",
"27.59",
"38.95",
"LoraHub",
"26.07",
"47.45",
"25.84",
"24.30",
"Figure l: Results on the XLSum and XWikis datasets",
"Full FT (CL)",
"30.65",
"19.5728.73",
"39.42",
"24.16",
"64-",
"25.94",
"LoRA-4 (CL)",
"with PaLM 2-XXS trained in the low -→ high-data",
"42.07",
"27.66",
"31.08",
"45.12",
"28.05",
"regime: Full FT vs. LoRA-4. Results for up to 256",
"SHOT",
"24.21",
"41.34",
"28.0227.66",
"48.09",
"26.56",
"examples per language are averaged over three seeds,",
"LoraHub",
"with standard deviation shown in shaded areas.",
"Table 4:",
"Cross-lingual transferon 10 XLSum lan-",
"guages and five XWikis languages (using leave-one-",
"Test",
"XLSUM",
"XWIKIS",
"out training) for PaLM 2-XXS model. 16- and 64-shot",
"R-L",
"SH",
"R-L",
"NLI",
"NLI",
"Languages",
"SH",
"experimentsshowaverageresultsfromthreedifferent",
"FullFT",
"5.20",
"4.49",
"17.51",
"35.95",
"22.43",
" seed runs. For continued learning (cL ), we use a 14/2",
"Non-English",
"6.88",
"LoRA-4",
"21.13",
"39.07",
"23.08",
"23.86",
"45.54",
"25.96",
"and 60/4 training/validation split.Best results are in",
"FullFT",
"32.58",
"57.09",
"36.59",
"30.81",
"53.59",
"38.01",
"bold. Results for individual languages are in Tables 10",
"LoRA-4",
"English",
"32.2163.13",
"43.44",
"34.07",
"49.94",
"29.01",
"and 13 in Appendix C.",
" Table 3: Zero-shot cross-lingual transfer using full fine-",
"tuning (Full FT) and LoRA (rank 4):PaLM 2-XXS",
"Figure 3ROUGE-L SEEN UNSEEN UNSEEN",
"models are trained and validated on English and tested",
"on all other languages (Non-English) and English only",
"LoRA(4)LoRA",
"BestROUGE-L (R-L),NLI, and SEAHORSE (SH)con-",
"Figure 3 iLoRAiLoRA Avg.LoRA LoRAXL-",
"ciseness scores (area under the ROC curve) are in bold.",
"Sum(ieRNGDSOYO)LoRALoRA(ie ） (iii) XL-",
"SumZHJASWRNXWikis",
"Figure 2 HausaXLSumO Appendix C PaLM2 4",
"Few-shot Cross-lingual Transfer LoRA",
"Vu et al. (2022)",
"LLM (Brown et al., 2020; Wei et al., 2022b)",
"LoraHub 3.1 LoraHub (Huang et al., 2023)",
"Zero-shot",
"Transfer from Multiple Lan-",
"LoRA(ie)N LoRALoraHub 1/N",
"guages",
"1664LoraHub14/260/4LoraHub Nevergrad",
"6",
"XLSumIO(AR )(ZH)(EN)( HA )( HI )(ID )",
"ROUGE-LROUGE-L",
"FA)(PT )(XMAT1O AZ BN JA RN",
"KO",
"NE5",
"Table",
"4",
"1664XL-",
"XWikis",
"SumXWikis()(eg",
"16)",
"7XLSum68ROUGE-",
"LoRALoRALoRA 2",
"LXWikis7",
"al.",
"(2020)",
"et",
"Lauscher",
"4See Table 21 in Anil et al. (2023) regarding the distribution",
"iXLSumLoraHubROUGE-L16iiLoRAROUGE-",
"of languages used in the pre-training of PaLM 2.",
"SSee Appendix A for details of language families in XLSum.",
"6https://facebookresearch.github.io/nevergrad"
] |
{
"arxiv_id": "2311.08572",
"language": "zh",
"page_idx": 3
}
|
||
1812.00456-zh-0
|
1812.00456-zh-0.jpg
|
# Revisiting the Softmax Bellman Operator: New Benefts and New Perspective
Zhao Song$^1{}^{* }$ RonaldE.Parr$^1$ Lawrence Carin$^1$
。3in
$\operatorname{Abstract}$
The impact of softmax on the value function itself in reinforcement learning (RL) is often viewed as problematic because it leads to sub-optimal value (or Q) functions and interferes with the contraction properties of the Bellman operator. Surprisingly, despite these concerns, and independent of its effect on exploration , the softmax Bellman operator when combined with Deep Q-learning, leads to Q-functions with superior policies in practice, even outperforming its double Qlearning counterpart. To better understand how and why this occurs, we revisit theoretical properties of the softmax Bellman operator, and prove that $(i)$ it converges to the standard Bellman operator exponentially fast in the inverse temperature parameter, and $(ii)$ the distance of its Q function from the optimal one can be bounded. These alone do not explain its superior performance, so we also show that the softmax operator can reduce the overestimation error, which may give some insight into why a sub-optimal operator leads to better performance in the presence of value function approximation. A comparison among different Bellman operators is then presented, showing the trade-offs when selecting them.
## 1. Introduction
Bellman 方程E$( Bellman, 1957)$ 一直是E$reinforce-$ ment learning (RL)中的基本工具，因为它为动态规划中的最优策略提供了充分条件。在 Bellman 方程中使用 max 函数进一步表明，最优策略应该是贪婪的
${^{1}\text{Duke University}^{*}\text{Work performed as a graduate stu-}}$ dent at Duke University; now at Baidu Research. Correspondence to: Zhao Song<zsong@alumni.duke.edu>, Ronald E. Parr<parr@cs.duke.edu>,Lawrence Carin< $\operatorname{lcarin@duke.edu}>.$
Proceedings of the 36$^{th}$ International Conference on Machine Learning, Long Beach, California, PMLR 97, 2019. Copyright 2019 by the author(s).
w。r。t。Q 值。另一方面，探索和利用(Thrun,1992) 之间的权衡促使在学习过程中使用探索性和潜在的次优行为，一种常用的策略是通过用 softmax 函数替换$\max$函数来增加随机性，如在玻尔兹曼探索 (Sutton $\&$ Barto, 1998) 中。此外，softmax 函数是 max 函数的可微分近似，因此可以促进分析 (Reverdy&Leonard, 2016)。
softmaxBellman 算子的有益特性与其对结果值或 Q 函数的准确性的潜在负面影响形成对比。例如，已经证明，对于某些温度参数$\overline{\mathrm{F}}($Littman, 1996, Page 205) , softmaxBellman 算子不是收缩。鉴于此，人们可能会期望 softmaxBellman 算子的便利属性将以结果值或 Q 函数的准确性或结果策略的质量为代价。在本文中，我们证明，在深度 Q 学习的情况下，这种期望是令人惊讶的不正确。在目标网络中，我们将 softmaxBellman 算子与Edeep Q- network ( DQN) E( Mnih et al. , 2015) 和$\boxed{\mathrm{F}}$double DQN (DDQN) E(van Hasselt et al., 2016a)算法相结合，方法是将其中的 max 函数替换为softmax 函数。然后，我们在 Arcade Learning Environment (ALE) $\operatorname{E}($Bellemare et al., 2013) 中的几个游戏上测试了这些变体，这是一个标准的大型深度$\boxed{\mathrm{ERL}}$ 测试平台。结果表明，使用 softmaxBellman 算子的变体可以获得更高的测试分数，并减少了 Q 值的高估以及其中大多数的梯度噪声。This effect is independent of exploration and is entirely attributable to the change in the Bellman operator.
这个令人惊讶的结果表明，有必要对 softmaxBellman 运算符有更深入的了解。为此，我们证明了从相同的初始 Q 值开始，我们可以上限和下限使用 softmax 算子计算的 Q 函数可以与使用常规 Bellman 算子计算的Q 函数偏离多少。我们进一步证明了 softmaxBellman 算子以指数速率 w 收敛到最优 Bellman 算子。r。t。逆温度参数。这让我们了解了为什么负收敛结果可能不像最初看起来那么令人沮丧，但它并不能解释在实践中观察到的卓越性能。由于最近的工作$\mathbb{E}($van Hasselt et al., 2016a; Anschel et al., 2017) 的目标偏差和原始E DQN E(Mnih et al.,2015)的不稳定性，我们进一步研究了 softmaxBellman 算子是否可以缓解这些问题。正如 van Hasselt et al. (2016a) 中所讨论的那样，在某些 Atari 游戏中，香草 DQN 性能不佳的一种可能解释是在计算目标网络时由于其中的最大运算符而高估了偏差。我们证明，给定与Evan Hasselt et al. (2016a) 相同的假设，对于任何逆温度参数，softmaxBellman
|
[
[
0.27294921875,
0.11236572265625,
0.69921875,
0.11297607421875,
0.69921875,
0.1307373046875,
0.27294921875,
0.1300048828125
],
[
0.313720703125,
0.136962890625,
0.6611328125,
0.136962890625,
0.6611328125,
0.1514892578125,
0.313720703125,
0.1514892578125
],
[
0.298095703125,
0.2008056640625,
0.6708984375,
0.199462890625,
0.6708984375,
0.2139892578125,
0.298095703125,
0.21533203125
],
[
0.08905029296875,
0.2183837890625,
0.131591796875,
0.2183837890625,
0.131591796875,
0.23291015625,
0.08905029296875,
0.23291015625
],
[
0.24267578125,
0.2291259765625,
0.3203125,
0.2310791015625,
0.3203125,
0.246826171875,
0.24267578125,
0.2442626953125
],
[
0.5,
0.23291015625,
0.8837890625,
0.2310791015625,
0.8837890625,
0.2442626953125,
0.5,
0.246826171875
],
[
0.50244140625,
0.2474365234375,
0.88232421875,
0.2474365234375,
0.88232421875,
0.261474609375,
0.50244140625,
0.261474609375
],
[
0.1217041015625,
0.25390625,
0.441162109375,
0.25390625,
0.441162109375,
0.268310546875,
0.1217041015625,
0.268310546875
],
[
0.50244140625,
0.261962890625,
0.8857421875,
0.261962890625,
0.8857421875,
0.276611328125,
0.50244140625,
0.276611328125
],
[
0.1217041015625,
0.270751953125,
0.441162109375,
0.270751953125,
0.441162109375,
0.2841796875,
0.1217041015625,
0.2841796875
],
[
0.5009765625,
0.27783203125,
0.88232421875,
0.27783203125,
0.88232421875,
0.291748046875,
0.5009765625,
0.291748046875
],
[
0.1217041015625,
0.285400390625,
0.441162109375,
0.285400390625,
0.441162109375,
0.298583984375,
0.1217041015625,
0.298583984375
],
[
0.5,
0.291748046875,
0.8837890625,
0.292236328125,
0.8837890625,
0.306884765625,
0.5,
0.30615234375
],
[
0.1217041015625,
0.300537109375,
0.439453125,
0.300537109375,
0.439453125,
0.314453125,
0.1217041015625,
0.314453125
],
[
0.50341796875,
0.30810546875,
0.8857421875,
0.30810546875,
0.8857421875,
0.3212890625,
0.50341796875,
0.3212890625
],
[
0.1217041015625,
0.31640625,
0.441162109375,
0.31640625,
0.441162109375,
0.330078125,
0.1217041015625,
0.330078125
],
[
0.5009765625,
0.322509765625,
0.5498046875,
0.324462890625,
0.5498046875,
0.3359375,
0.5009765625,
0.333984375
],
[
0.1217041015625,
0.330810546875,
0.441162109375,
0.330810546875,
0.441162109375,
0.3447265625,
0.1217041015625,
0.3447265625
],
[
0.1217041015625,
0.34521484375,
0.441162109375,
0.34521484375,
0.441162109375,
0.359130859375,
0.1217041015625,
0.359130859375
],
[
0.5009765625,
0.343505859375,
0.88623046875,
0.343505859375,
0.88623046875,
0.356689453125,
0.5009765625,
0.356689453125
],
[
0.12091064453125,
0.361083984375,
0.441162109375,
0.361083984375,
0.441162109375,
0.375,
0.12091064453125,
0.375
],
[
0.50244140625,
0.359130859375,
0.8857421875,
0.359130859375,
0.8857421875,
0.37255859375,
0.50244140625,
0.37255859375
],
[
0.12091064453125,
0.375,
0.439453125,
0.375732421875,
0.439453125,
0.39013671875,
0.12091064453125,
0.389404296875
],
[
0.50244140625,
0.375,
0.88232421875,
0.375,
0.88232421875,
0.38818359375,
0.50244140625,
0.38818359375
],
[
0.12091064453125,
0.39013671875,
0.439453125,
0.39013671875,
0.439453125,
0.404052734375,
0.12091064453125,
0.404052734375
],
[
0.515625,
0.389404296875,
0.8857421875,
0.389404296875,
0.8857421875,
0.40283203125,
0.515625,
0.40283203125
],
[
0.12091064453125,
0.4072265625,
0.441162109375,
0.4072265625,
0.441162109375,
0.42041015625,
0.12091064453125,
0.42041015625
],
[
0.50244140625,
0.401611328125,
0.8857421875,
0.404052734375,
0.8857421875,
0.417236328125,
0.50244140625,
0.415283203125
],
[
0.1217041015625,
0.421630859375,
0.441162109375,
0.421630859375,
0.441162109375,
0.43505859375,
0.1217041015625,
0.43505859375
],
[
0.50341796875,
0.42041015625,
0.8837890625,
0.42041015625,
0.8837890625,
0.43115234375,
0.50341796875,
0.43115234375
],
[
0.12091064453125,
0.436279296875,
0.439453125,
0.436279296875,
0.439453125,
0.449462890625,
0.12091064453125,
0.449462890625
],
[
0.50341796875,
0.434326171875,
0.88232421875,
0.434326171875,
0.88232421875,
0.448974609375,
0.50341796875,
0.448974609375
],
[
0.12091064453125,
0.451904296875,
0.439453125,
0.451904296875,
0.439453125,
0.46533203125,
0.12091064453125,
0.46533203125
],
[
0.50341796875,
0.449462890625,
0.8837890625,
0.449462890625,
0.8837890625,
0.46337890625,
0.50341796875,
0.46337890625
],
[
0.1217041015625,
0.466552734375,
0.439453125,
0.466552734375,
0.439453125,
0.479736328125,
0.1217041015625,
0.479736328125
],
[
0.50244140625,
0.464111328125,
0.8837890625,
0.464111328125,
0.8837890625,
0.477783203125,
0.50244140625,
0.477783203125
],
[
0.1217041015625,
0.481689453125,
0.439453125,
0.481689453125,
0.439453125,
0.49560546875,
0.1217041015625,
0.49560546875
],
[
0.50341796875,
0.479248046875,
0.88232421875,
0.479248046875,
0.88232421875,
0.492431640625,
0.50341796875,
0.492431640625
],
[
0.12091064453125,
0.496826171875,
0.439453125,
0.496826171875,
0.439453125,
0.51025390625,
0.12091064453125,
0.51025390625
],
[
0.5009765625,
0.494384765625,
0.8857421875,
0.494384765625,
0.8857421875,
0.50830078125,
0.5009765625,
0.50830078125
],
[
0.12091064453125,
0.51220703125,
0.441162109375,
0.51220703125,
0.441162109375,
0.52587890625,
0.12091064453125,
0.52587890625
],
[
0.5009765625,
0.51025390625,
0.8857421875,
0.51025390625,
0.8857421875,
0.5234375,
0.5009765625,
0.5234375
],
[
0.12091064453125,
0.5263671875,
0.439453125,
0.52587890625,
0.439453125,
0.54052734375,
0.12091064453125,
0.54150390625
],
[
0.50244140625,
0.5244140625,
0.8857421875,
0.5244140625,
0.8857421875,
0.53857421875,
0.50244140625,
0.53857421875
],
[
0.12091064453125,
0.54248046875,
0.441162109375,
0.54248046875,
0.441162109375,
0.55615234375,
0.12091064453125,
0.55615234375
],
[
0.50341796875,
0.54052734375,
0.8857421875,
0.54052734375,
0.8857421875,
0.5537109375,
0.50341796875,
0.5537109375
],
[
0.12091064453125,
0.55615234375,
0.439453125,
0.556640625,
0.439453125,
0.5712890625,
0.12091064453125,
0.57080078125
],
[
0.5009765625,
0.5546875,
0.8857421875,
0.5537109375,
0.8857421875,
0.568359375,
0.5009765625,
0.5693359375
],
[
0.1217041015625,
0.57275390625,
0.439453125,
0.57275390625,
0.439453125,
0.5859375,
0.1217041015625,
0.5859375
],
[
0.50341796875,
0.5712890625,
0.88134765625,
0.5712890625,
0.88134765625,
0.58203125,
0.50341796875,
0.58203125
],
[
0.1217041015625,
0.5869140625,
0.441162109375,
0.5869140625,
0.441162109375,
0.60107421875,
0.1217041015625,
0.60107421875
],
[
0.5009765625,
0.583984375,
0.8857421875,
0.583984375,
0.8857421875,
0.5986328125,
0.5009765625,
0.5986328125
],
[
0.12091064453125,
0.60107421875,
0.439453125,
0.6015625,
0.439453125,
0.6162109375,
0.12091064453125,
0.61572265625
],
[
0.50244140625,
0.60107421875,
0.8857421875,
0.60107421875,
0.8857421875,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.12091064453125,
0.6171875,
0.441162109375,
0.6171875,
0.441162109375,
0.630859375,
0.12091064453125,
0.630859375
],
[
0.5009765625,
0.61572265625,
0.67724609375,
0.6171875,
0.67724609375,
0.62744140625,
0.5009765625,
0.62548828125
],
[
0.12091064453125,
0.6318359375,
0.439453125,
0.6318359375,
0.439453125,
0.64599609375,
0.12091064453125,
0.64599609375
],
[
0.5,
0.63525390625,
0.8798828125,
0.63525390625,
0.8798828125,
0.6484375,
0.5,
0.6484375
],
[
0.1217041015625,
0.64794921875,
0.439453125,
0.64794921875,
0.439453125,
0.6611328125,
0.1217041015625,
0.6611328125
],
[
0.50244140625,
0.6513671875,
0.8837890625,
0.6513671875,
0.8837890625,
0.66552734375,
0.50244140625,
0.66552734375
],
[
0.12091064453125,
0.66357421875,
0.2303466796875,
0.66357421875,
0.2303466796875,
0.6767578125,
0.12091064453125,
0.6767578125
],
[
0.50244140625,
0.66650390625,
0.88232421875,
0.66650390625,
0.88232421875,
0.68115234375,
0.50244140625,
0.68115234375
],
[
0.50341796875,
0.681640625,
0.8837890625,
0.681640625,
0.8837890625,
0.69580078125,
0.50341796875,
0.69580078125
],
[
0.08905029296875,
0.6962890625,
0.22216796875,
0.6962890625,
0.22216796875,
0.7109375,
0.08905029296875,
0.7109375
],
[
0.50244140625,
0.6962890625,
0.88232421875,
0.6962890625,
0.88232421875,
0.71044921875,
0.50244140625,
0.71044921875
],
[
0.50341796875,
0.71337890625,
0.88134765625,
0.71337890625,
0.88134765625,
0.7236328125,
0.50341796875,
0.7236328125
],
[
0.08905029296875,
0.72265625,
0.4697265625,
0.72265625,
0.4697265625,
0.736328125,
0.08905029296875,
0.736328125
],
[
0.50244140625,
0.7265625,
0.8837890625,
0.7265625,
0.8837890625,
0.74072265625,
0.50244140625,
0.74072265625
],
[
0.0906982421875,
0.73779296875,
0.473876953125,
0.73779296875,
0.473876953125,
0.751953125,
0.0906982421875,
0.751953125
],
[
0.50244140625,
0.7421875,
0.8857421875,
0.7421875,
0.8857421875,
0.755859375,
0.50244140625,
0.755859375
],
[
0.09234619140625,
0.75390625,
0.471435546875,
0.75390625,
0.471435546875,
0.7646484375,
0.09234619140625,
0.7646484375
],
[
0.50244140625,
0.7568359375,
0.8837890625,
0.7568359375,
0.8837890625,
0.77099609375,
0.50244140625,
0.77099609375
],
[
0.0906982421875,
0.767578125,
0.47216796875,
0.767578125,
0.47216796875,
0.7822265625,
0.0906982421875,
0.7822265625
],
[
0.5009765625,
0.771484375,
0.8857421875,
0.771484375,
0.8857421875,
0.7861328125,
0.5009765625,
0.7861328125
],
[
0.5,
0.78515625,
0.88232421875,
0.7861328125,
0.88232421875,
0.80029296875,
0.5,
0.7998046875
],
[
0.108642578125,
0.79150390625,
0.473876953125,
0.79150390625,
0.473876953125,
0.8046875,
0.108642578125,
0.8046875
],
[
0.0882568359375,
0.80322265625,
0.473876953125,
0.8037109375,
0.473876953125,
0.818359375,
0.0882568359375,
0.8173828125
],
[
0.5009765625,
0.8017578125,
0.88232421875,
0.8017578125,
0.88232421875,
0.81494140625,
0.5009765625,
0.81494140625
],
[
0.0906982421875,
0.8173828125,
0.465576171875,
0.8173828125,
0.465576171875,
0.83056640625,
0.0906982421875,
0.83056640625
],
[
0.50244140625,
0.8173828125,
0.8857421875,
0.8173828125,
0.8857421875,
0.83056640625,
0.50244140625,
0.83056640625
],
[
0.0906982421875,
0.830078125,
0.47216796875,
0.830078125,
0.47216796875,
0.84326171875,
0.0906982421875,
0.84326171875
],
[
0.5009765625,
0.83056640625,
0.88232421875,
0.83203125,
0.88232421875,
0.8466796875,
0.5009765625,
0.84521484375
],
[
0.08905029296875,
0.84130859375,
0.2181396484375,
0.84228515625,
0.2181396484375,
0.85595703125,
0.08905029296875,
0.85498046875
],
[
0.50244140625,
0.84765625,
0.8857421875,
0.84765625,
0.8857421875,
0.861328125,
0.50244140625,
0.861328125
],
[
0.0882568359375,
0.8662109375,
0.473876953125,
0.86572265625,
0.473876953125,
0.8798828125,
0.0882568359375,
0.880859375
],
[
0.5048828125,
0.8623046875,
0.8857421875,
0.8623046875,
0.8857421875,
0.876953125,
0.5048828125,
0.876953125
],
[
0.08905029296875,
0.8798828125,
0.473876953125,
0.8798828125,
0.473876953125,
0.89306640625,
0.08905029296875,
0.89306640625
],
[
0.5009765625,
0.87744140625,
0.8857421875,
0.87744140625,
0.8857421875,
0.8916015625,
0.5009765625,
0.8916015625
],
[
0.08905029296875,
0.892578125,
0.3095703125,
0.892578125,
0.3095703125,
0.90576171875,
0.08905029296875,
0.90576171875
]
] |
[
"Revisiting the Softmax Bellman Operator:",
"NewBenefits and NewPerspective",
"Zhao Song1 *Ronald E. Parr1 Lawrence Carin 1",
"。3in",
"Abstract",
"W。r。t。Q值。另一方面，探索和利用（Thrun，1992)",
"之间的权衡促使在学习过程中使用探索性和潜在的次",
"The impact of softmax on the value func-",
"优行为，一种常用的策略是通过用softmax函数替换",
"tion itself in reinforcement learning (RL) is",
"max函数来增加随机性，如在玻尔兹曼探索（Sutton",
"often viewed as problematic because it leads",
"&Barto，1998）中。此外，softmax函数是max函数的",
"to sub-optimal value (or Q) functions and",
"可微分近似，因此可以促进分析（Reverdy&Leonard,",
"interferes with the contraction properties of",
"2016)",
"the Bellman operator. Surprisingly, despite",
"these concerns, and independent of its effect",
"softmaxBellman算子的有益特性与其对结果值或Q",
"on exploration , the softmax Bellman oper-",
"函数的准确性的潜在负面影响形成对比。例如，已经",
"ator when combined with Deep Q-learning,",
"证明，对于某些温度参数E(Littman，1996，Page205）",
"leads to Q-functions with superior policies",
"softmaxBellman算子不是收缩。鉴于此，人们可能",
"in practice, even outperforming its double Q-",
"会期望softmaxBellman算子的便利属性将以结果值",
"learning counterpart. To better understand",
"或Q函数的准确性或结果策略的质量为代价。在本文",
"how and why this occurs, we revisit theoret",
"中，我们证明，在深度Q学习的情况下，这种期望是令",
"ical properties of the softmax Bellman oper",
"人惊呀的不正确。在目标网络中，我们将softmaxBell",
"ator, and prove that (i) it converges to the",
"man算子与EdeepQ-network(DQN)E(Mnih etal.,",
"standard Bellman operator exponentially fast",
"2015)和EdoubleDQN (DDQN)E(vanHasselt etal..",
"in the inverse temperature parameter, and",
"2016a）算法相结合，方法是将其中的max函数替换为",
"(ii) the distance of its Q function from the",
"softmax函数。然后，我们在ArcadeLearningEnviron",
"optimal one can be bounded.These alone",
"ment（ALE）E(Bellemareetal.,2013）中的几个游戏",
"do not explain its superior performance, so",
"上测试了这些变体，这是一个标准的大型深度ERL测",
"we also show that the softmax operator can",
"试平台。结果表明，使用softmaxBellman算子的变体",
"reduce the overestimation error, which may",
"可以获得更高的测试分数，并减少了Q值的高估以及",
"give some insight into why a sub-optimal op-",
"其中大多数的梯度噪声。This effect is independent of",
"erator leads to better performance in the pres-",
"exploration and is entirely attributable to the change",
"ence of value function approximation. A com-",
"intheBellmanoperator.",
"parison among different Bellman operators is",
"这个令人惊守的结果表明，有必要对softmaxBellmal",
"then presented, showing the trade-offs when",
"运算符有更深入的了解。为此，我们证明了从相同的",
"selecting them.",
"初始Q值开始，我们可以上限和下限使用softmax算",
"子计算的Q函数可以与使用常规Bellman算子计算的",
"1.Introduction",
"Q函数偏离多少。我们进一步证明了softmaxBellman",
"算子以指数速率w收敛到最优Bellman算子。r。t。逆",
"Bellman方程E(Bellman，1957）一直是Ereinforce",
"温度参数。这让我们了解了为什么负收敛结果可能不",
"mentlearning（RL）中的基本工具，因为它为动态规",
"像最初看起来那么令人沮丧，但它并不能解释在实践",
"划中的最优策略提供了充分条件。在Bellman方程中",
"中观察到的卓越性能。由于最近的工作E(vanHasselt",
"使用max函数进一步表明，最优策略应该是贪的",
"etal,2016a；Anscheletal.,2017）的目标偏差和原始E",
"DQNE（Mnihetal.，2015）的不稳定性，我们进一步",
"1 Duke University * Work performed as a graduate stu-",
"dent at Duke University; now at Baidu Research . Cor-",
"研究了softmaxBellman算子是否可以缓解这些问题",
"respondence to: Zhao Song < zsong@alumni.duke.edu >",
"正如vanHasseltetal.（2016a）中所讨论的那样，在某",
"Ronald E. Parr < parr@cs.duke.edu >, Lawrence Carin <",
"些Atari游戏中，香草DQN性能不佳的一种可能解",
"Icarin@duke.edu>.",
"释是在计算目标网络时由于其中的最大运算符而高估",
"Proceedings of the 36th International Conference on Ma-",
"了偏差。我们证明，给定与EvanHasseltetal.（2016a）",
"chine Learning, Long Beach, California, PMLR 97, 2019.",
"相同的假设，对于任何逆温度参数，softmaxBellman",
"Copyright 2019 by the author(s)."
] |
{
"arxiv_id": "1812.00456",
"language": "zh",
"page_idx": 0
}
|
||
1812.00456-zh-1
|
1812.00456-zh-1.jpg
|
算子都可以减少高估偏差。我们还通过提供其下限和
上限来量化高估的减少
我们的结果是对现有工作的补充，并为探索软化贝尔曼算子的各种方式的现有工作增加了新的动力。例如， 熵正则化已被用于平滑策略。这种方法的动机包括计算方便，探索或鲁棒性E(Fox et al.,2016; Haarnoja et al., 2017; Schulman et al., 2017; Neu et al., 2017) 。关于值函数，EAsadi & Littman (2017)提出了一种替代的 mellowmax 算子，并证明了它是一种收缩。他们的实验结果表明，它可以改善勘探，但是没有考虑次优 Bellman 算子可以独立于勘探而导致更好的策略的可能性。。因此，我们的结果为进一步研究诸如mellowmax 之 类 的 运 算 符 提 供 了 额 外 的 动 力 。。虽 然 与 softmax 非常相似，但 mellowmax 运算符需要一些额外的计算来表示策略，如 Asadi&Littman (2017)中所述。我们的论文讨论了不同贝尔曼算子之间的这种权衡和其他权衡
本文的其余部分组织如下：我们在$\boxed{\mathrm{E}}$2 节中提供了必要的背景和符号。在E3 节中介绍了 softmaxBellman 算子，其中提供了其收敛属性和性能界限。尽管在 Q 函数上是次优的，但在$\overline{\mathrm{E}}$4 节中显示了 softmax 运算符，在几款 Atari 游戏中始终优于其最大对手。如此令人惊讶的结果进一步促使我们调查为什么在$\mathbb{E}$5 节中发生这种情况。6 节中对不同的 Bellman 算子进行了彻底的比较。第 7 节讨论了相关工作，第 8 节总结了本文
# 2. Background and Notation
A Markov decision process (MDP) 表示预期的MAMTHX<XNETHX\_1>中的 MAMTHX\_0 元素可以表示为 5 元组$\langle\mathcal{S},\mathcal{A},P,R,\gamma\rangle$ ,其中$\mathcal{S}$是状态空间，$\mathcal{A}$是动作空间，$P$是过渡内核，其元素 XMATHX。MDP 中的策略$\pi$可以用 probability mass function (PMF)表示，其中$\pi(s,a)\in[0,1]$表示在状态$s$ 和$\sum_{a\in\mathcal{A}}\pi(s,a)=1$中选择动作$a$的概率
对于给定的策略$\pi$ ,其状态-动作值函数$Q^\pi(s,a)$ 被定义为在状态$s$中采取动作$a$并随后遵循策略$\pi$时累积的，预期的折扣奖励，i。e。, $Q^\pi(s,a)=$ $\mathbb{E}_{a_t\sim\pi}\left[\sum_{t=0}^\infty\gamma^tr_t|s_0=s,a_0=a\right].$对于最优策略$\pi^*$, 其对应的 Q 函数满足以下 Bellman 方程：
$$Q^*(s,a)=R(s,a)+\gamma\sum_{s'}P(s'|s,a)\max_{a'}Q^*(s',a').$$
在EDQN E( Mnih et al. , 2015) 中 , 用神经网络将 Q 函数参数化为$Q_{\boldsymbol{\theta }}( s, a)$ ,该神经网络以状态$s$为输人， 并在最终的全连接线性层中输出相应的 Q 值，对于每个动作$a$ 。DQN 的训练目标可以表示为
$$\min\limits_{\boldsymbol{\theta}}\frac{1}{2}\left\|Q_{\boldsymbol{\theta}}(s,a)-[R(s,a)+\gamma\max\limits_{a'}Q_{\boldsymbol{\theta}^{-}}(s',a')]\right\|^{2},$$
(1)
,其中$\theta^-$对应于目标网络中的冻结权重，并以固定间隔更新。通过 RMSProp$\underline{\mathrm{F}}($Tieleman &Hinton, 2012) 对EEq. ( 1) 进行优化，并从重放缓冲区采样小批量
为了减少过高估计偏差，double DQN (DDQN) 算法
将$Q_{\boldsymbol{\theta}}(s,a)$旨在适应 Eq.(1)的目标修改为
$$R(s,a)+\gamma\:Q_{\boldsymbol{\theta}^{-}}\big(s',\arg\max_{a}Q_{\boldsymbol{\theta}_{t}}(s',a)\big).$$
请注意，基于最新估计$\boldsymbol{\theta}_t$的单独网络用于动作选择，
并且此策略的评估是由于
Notation softmax 函数定义为
$$f_\tau(\mathrm x)=\frac{[\exp(\tau x_1),\exp(\tau x_2),\dots,\exp(\tau x_m)]^T}{\sum_{i=1}^m\exp(\tau x_i)},$$
,其中上标$T$表示向量转置。随后，softmax 加权函数表示为 $g_\mathrm{x}(\tau)=f_{\tau}^{T}($x)x,作为 $\tau$ 的函数。此外，我们定义了向量$Q(s,)=[Q(s,a_{1}),Q(s,a_{2}),\ldots,Q(s,a_{m})]^{T}$。我们进一步将$m$设置为 MDP 中的动作集$\mathcal{A}$的大小。最后，$R_{\mathrm{min}}$和$R_\mathrm{max}$分别表示最小和最大即时奖励
$3.{\mathrm{~The~Softmax~Bellman~Operator}}$
我们首先提供以下标准 Bellman 运算符：
$$\mathcal{T}\:Q(s,a)=R(s,a)+\gamma\sum_{s'}P(s'|s,a)\:\max_{a'}Q(s',a').$$
(2)
我们建议使用 softmaxBellman 运算符，定义为
$$\begin{aligned}\mathcal{T}_{\mathrm{soft}}\:Q(s,a)=&R(s,a)+\gamma\sum_{s^{\prime}}P(s^{\prime}|s,a)\\&\times\underbrace{\sum_{a^{\prime}}\frac{\exp[\tau\:Q(s^{\prime},a^{\prime})]}{\sum_{\bar{a}}\exp[\tau\:Q(s^{\prime},\bar{a})]}\:Q(s^{\prime},a^{\prime})}_{sm_{\tau}(Q(s^{\prime},))},\end{aligned}$$
,其中$\tau\geq0$表示逆温度参数。请注意，当$\tau\to\infty$时，
$\mathcal{T}_\mathrm{soft}$将简化为$\mathcal{T}_\mathrm{soft}$将简化为$\mathcal{T}_\mathrm{soft}$
在$\boxed{\mathrm{E}}$Asadi & Littman (2017) 中引人了 mellowmax 运算符，作为EEq. ( 3) 中 softmax 运算符的替代方法， 定义为
$$mm_\omega\big(Q(s,)\big)=\frac{\log\big(\frac{1}{m}\sum_{a'}\exp[\omega\:Q(s,a')]\big)}{\omega},$$
(4)
其中$\omega>0$是调整参数。与 Eq.(3)中的 softmax 算子$sm_\tau(Q(s^{\prime},))$类似，当$\omega$分别接近 0 和$\infty$时， $mm_{\omega}$收敛于均值和 max 算子。但是，与 softmax 运算符不同，Eq.(4)中的 mellowmax 运算符不直接提供动作的概率分布，因此需要额外的步骤来获得相应的策略 (Asadi &Littman, 2017)
与其用于改进探索的典型用途(也是$\overline{\mathrm{E}}$Asadi & Littman (2017)中的 mellowmax 运算符的目的)相
|
[
[
0.0906982421875,
0.08837890625,
0.471435546875,
0.08837890625,
0.471435546875,
0.0985107421875,
0.0906982421875,
0.0985107421875
],
[
0.50341796875,
0.085205078125,
0.8837890625,
0.08709716796875,
0.8837890625,
0.10101318359375,
0.50341796875,
0.0985107421875
],
[
0.09234619140625,
0.1029052734375,
0.25244140625,
0.1029052734375,
0.25244140625,
0.11297607421875,
0.09234619140625,
0.11297607421875
],
[
0.50244140625,
0.10162353515625,
0.8837890625,
0.10162353515625,
0.8837890625,
0.11553955078125,
0.50244140625,
0.11553955078125
],
[
0.50244140625,
0.117431640625,
0.86767578125,
0.117431640625,
0.86767578125,
0.1307373046875,
0.50244140625,
0.1307373046875
],
[
0.0906982421875,
0.125,
0.473876953125,
0.125,
0.473876953125,
0.1383056640625,
0.0906982421875,
0.1383056640625
],
[
0.08905029296875,
0.1383056640625,
0.4755859375,
0.14013671875,
0.4755859375,
0.154052734375,
0.08905029296875,
0.1514892578125
],
[
0.50244140625,
0.1395263671875,
0.8857421875,
0.1395263671875,
0.8857421875,
0.15283203125,
0.50244140625,
0.15283203125
],
[
0.0906982421875,
0.154052734375,
0.47216796875,
0.154052734375,
0.47216796875,
0.1673583984375,
0.0906982421875,
0.1673583984375
],
[
0.5,
0.154052734375,
0.814453125,
0.154052734375,
0.814453125,
0.1673583984375,
0.5,
0.1673583984375
],
[
0.08905029296875,
0.1685791015625,
0.47216796875,
0.169189453125,
0.47216796875,
0.184326171875,
0.08905029296875,
0.18310546875
],
[
0.08905029296875,
0.184326171875,
0.47216796875,
0.184326171875,
0.47216796875,
0.1976318359375,
0.08905029296875,
0.1976318359375
],
[
0.5556640625,
0.18115234375,
0.82861328125,
0.18115234375,
0.82861328125,
0.1956787109375,
0.5556640625,
0.1956787109375
],
[
0.10455322265625,
0.199462890625,
0.45751953125,
0.199462890625,
0.45751953125,
0.21337890625,
0.10455322265625,
0.21337890625
],
[
0.0906982421875,
0.2139892578125,
0.47802734375,
0.2139892578125,
0.47802734375,
0.2279052734375,
0.0906982421875,
0.2279052734375
],
[
0.5009765625,
0.210205078125,
0.8857421875,
0.2108154296875,
0.8857421875,
0.2259521484375,
0.5009765625,
0.2247314453125
],
[
0.0906982421875,
0.2298583984375,
0.473876953125,
0.2298583984375,
0.473876953125,
0.24365234375,
0.0906982421875,
0.24365234375
],
[
0.5009765625,
0.2266845703125,
0.68115234375,
0.2266845703125,
0.68115234375,
0.240478515625,
0.5009765625,
0.240478515625
],
[
0.09234619140625,
0.2442626953125,
0.47216796875,
0.2442626953125,
0.47216796875,
0.25830078125,
0.09234619140625,
0.25830078125
],
[
0.08905029296875,
0.260009765625,
0.473876953125,
0.260009765625,
0.473876953125,
0.2734375,
0.08905029296875,
0.2734375
],
[
0.5,
0.2568359375,
0.7236328125,
0.2568359375,
0.7236328125,
0.270751953125,
0.5,
0.270751953125
],
[
0.0882568359375,
0.274658203125,
0.473876953125,
0.27587890625,
0.473876953125,
0.2890625,
0.0882568359375,
0.28857421875
],
[
0.5908203125,
0.281005859375,
0.8486328125,
0.281005859375,
0.8486328125,
0.29541015625,
0.5908203125,
0.29541015625
],
[
0.08905029296875,
0.290283203125,
0.473876953125,
0.290283203125,
0.473876953125,
0.3037109375,
0.08905029296875,
0.3037109375
],
[
0.529296875,
0.291748046875,
0.58837890625,
0.28857421875,
0.5888671875,
0.3037109375,
0.53125,
0.306884765625
],
[
0.08905029296875,
0.304931640625,
0.47216796875,
0.3037109375,
0.47216796875,
0.318115234375,
0.08905029296875,
0.3193359375
],
[
0.66845703125,
0.2998046875,
0.77294921875,
0.2998046875,
0.77294921875,
0.314453125,
0.66845703125,
0.314453125
],
[
0.0906982421875,
0.32080078125,
0.473876953125,
0.32080078125,
0.473876953125,
0.333984375,
0.0906982421875,
0.333984375
],
[
0.5048828125,
0.324462890625,
0.8857421875,
0.324462890625,
0.8857421875,
0.33837890625,
0.5048828125,
0.33837890625
],
[
0.08905029296875,
0.3359375,
0.205078125,
0.3359375,
0.205078125,
0.349853515625,
0.08905029296875,
0.349853515625
],
[
0.50244140625,
0.34033203125,
0.88232421875,
0.34033203125,
0.88232421875,
0.354248046875,
0.50244140625,
0.354248046875
],
[
0.0906982421875,
0.35791015625,
0.47216796875,
0.35791015625,
0.47216796875,
0.371826171875,
0.0906982421875,
0.371826171875
],
[
0.50341796875,
0.35595703125,
0.87109375,
0.35595703125,
0.87109375,
0.369384765625,
0.50341796875,
0.369384765625
],
[
0.0906982421875,
0.37255859375,
0.47216796875,
0.37255859375,
0.47216796875,
0.386474609375,
0.0906982421875,
0.386474609375
],
[
0.50244140625,
0.369384765625,
0.8857421875,
0.369384765625,
0.8857421875,
0.3837890625,
0.50244140625,
0.3837890625
],
[
0.0906982421875,
0.38818359375,
0.473876953125,
0.38818359375,
0.473876953125,
0.401611328125,
0.0906982421875,
0.401611328125
],
[
0.5009765625,
0.386474609375,
0.86767578125,
0.386474609375,
0.86767578125,
0.399658203125,
0.5009765625,
0.399658203125
],
[
0.0906982421875,
0.40283203125,
0.47216796875,
0.401611328125,
0.47216796875,
0.416748046875,
0.0906982421875,
0.417236328125
],
[
0.0906982421875,
0.41845703125,
0.473876953125,
0.41845703125,
0.473876953125,
0.431884765625,
0.0906982421875,
0.431884765625
],
[
0.5,
0.42041015625,
0.7939453125,
0.42041015625,
0.7939453125,
0.43505859375,
0.5,
0.43505859375
],
[
0.09234619140625,
0.434326171875,
0.47216796875,
0.434326171875,
0.47216796875,
0.447509765625,
0.09234619140625,
0.447509765625
],
[
0.09234619140625,
0.448974609375,
0.471435546875,
0.448974609375,
0.471435546875,
0.462158203125,
0.09234619140625,
0.462158203125
],
[
0.5009765625,
0.4462890625,
0.7861328125,
0.4462890625,
0.7861328125,
0.460205078125,
0.5009765625,
0.460205078125
],
[
0.09234619140625,
0.464111328125,
0.47216796875,
0.464111328125,
0.47216796875,
0.477783203125,
0.09234619140625,
0.477783203125
],
[
0.51123046875,
0.4716796875,
0.87158203125,
0.4716796875,
0.87158203125,
0.4892578125,
0.51123046875,
0.4892578125
],
[
0.08905029296875,
0.48095703125,
0.156005859375,
0.48095703125,
0.156005859375,
0.492431640625,
0.08905029296875,
0.492431640625
],
[
0.6806640625,
0.494384765625,
0.69384765625,
0.4892578125,
0.701171875,
0.5,
0.68701171875,
0.5048828125
],
[
0.86279296875,
0.501953125,
0.8857421875,
0.501953125,
0.8857421875,
0.51513671875,
0.86279296875,
0.51513671875
],
[
0.0882568359375,
0.5146484375,
0.3349609375,
0.5146484375,
0.3349609375,
0.52880859375,
0.0882568359375,
0.52880859375
],
[
0.5009765625,
0.5234375,
0.83837890625,
0.5234375,
0.83837890625,
0.537109375,
0.5009765625,
0.537109375
],
[
0.0882568359375,
0.54052734375,
0.47216796875,
0.54052734375,
0.47216796875,
0.5537109375,
0.0882568359375,
0.5537109375
],
[
0.08905029296875,
0.5537109375,
0.47216796875,
0.5537109375,
0.47216796875,
0.568359375,
0.08905029296875,
0.568359375
],
[
0.5087890625,
0.5498046875,
0.78369140625,
0.548828125,
0.78369140625,
0.56640625,
0.5087890625,
0.5673828125
],
[
0.0906982421875,
0.5693359375,
0.47216796875,
0.5693359375,
0.47216796875,
0.58349609375,
0.0906982421875,
0.58349609375
],
[
0.69384765625,
0.5673828125,
0.71142578125,
0.5673828125,
0.71142578125,
0.58056640625,
0.69384765625,
0.58056640625
],
[
0.09234619140625,
0.58544921875,
0.473876953125,
0.58544921875,
0.473876953125,
0.5986328125,
0.09234619140625,
0.5986328125
],
[
0.60888671875,
0.583984375,
0.77783203125,
0.57080078125,
0.78173828125,
0.599609375,
0.61279296875,
0.61279296875
],
[
0.7900390625,
0.59130859375,
0.8837890625,
0.59130859375,
0.8837890625,
0.60595703125,
0.7900390625,
0.60595703125
],
[
0.0906982421875,
0.599609375,
0.47216796875,
0.599609375,
0.47216796875,
0.6142578125,
0.0906982421875,
0.6142578125
],
[
0.65869140625,
0.599609375,
0.7861328125,
0.599609375,
0.7861328125,
0.6142578125,
0.65869140625,
0.6142578125
],
[
0.0906982421875,
0.61572265625,
0.47216796875,
0.61572265625,
0.47216796875,
0.62890625,
0.0906982421875,
0.62890625
],
[
0.63232421875,
0.6162109375,
0.6435546875,
0.60791015625,
0.650390625,
0.61279296875,
0.6396484375,
0.6220703125
],
[
0.0882568359375,
0.6298828125,
0.369384765625,
0.626953125,
0.369384765625,
0.64306640625,
0.0882568359375,
0.646484375
],
[
0.705078125,
0.62548828125,
0.77685546875,
0.62548828125,
0.77685546875,
0.64013671875,
0.705078125,
0.64013671875
],
[
0.0906982421875,
0.65283203125,
0.473876953125,
0.65283203125,
0.473876953125,
0.66650390625,
0.0906982421875,
0.66650390625
],
[
0.50244140625,
0.6494140625,
0.88623046875,
0.6494140625,
0.88623046875,
0.6640625,
0.50244140625,
0.6640625
],
[
0.08905029296875,
0.66650390625,
0.47216796875,
0.66748046875,
0.47216796875,
0.681640625,
0.08905029296875,
0.68115234375
],
[
0.50244140625,
0.66748046875,
0.619140625,
0.66748046875,
0.619140625,
0.67822265625,
0.50244140625,
0.67822265625
],
[
0.09234619140625,
0.68505859375,
0.4697265625,
0.68505859375,
0.4697265625,
0.69580078125,
0.09234619140625,
0.69580078125
],
[
0.5009765625,
0.68603515625,
0.88232421875,
0.68603515625,
0.88232421875,
0.70068359375,
0.5009765625,
0.70068359375
],
[
0.08905029296875,
0.69873046875,
0.473876953125,
0.69873046875,
0.473876953125,
0.7119140625,
0.08905029296875,
0.7119140625
],
[
0.5,
0.7021484375,
0.88232421875,
0.70263671875,
0.88232421875,
0.71728515625,
0.5,
0.71630859375
],
[
0.0882568359375,
0.7119140625,
0.3857421875,
0.71337890625,
0.3857421875,
0.72802734375,
0.0882568359375,
0.7265625
],
[
0.50244140625,
0.7197265625,
0.55126953125,
0.7197265625,
0.55126953125,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.101318359375,
0.7412109375,
0.4599609375,
0.7412109375,
0.4599609375,
0.7587890625,
0.101318359375,
0.7587890625
],
[
0.5185546875,
0.74365234375,
0.8505859375,
0.73486328125,
0.85107421875,
0.76318359375,
0.51953125,
0.771484375
],
[
0.8505859375,
0.75390625,
0.88232421875,
0.75390625,
0.88232421875,
0.76513671875,
0.8505859375,
0.76513671875
],
[
0.26318359375,
0.76123046875,
0.27783203125,
0.76123046875,
0.27783203125,
0.77099609375,
0.26318359375,
0.77099609375
],
[
0.08905029296875,
0.7822265625,
0.473876953125,
0.78271484375,
0.473876953125,
0.79736328125,
0.08905029296875,
0.796875
],
[
0.515625,
0.77880859375,
0.8857421875,
0.77880859375,
0.8857421875,
0.79248046875,
0.515625,
0.79248046875
],
[
0.5087890625,
0.7861328125,
0.515625,
0.7861328125,
0.515625,
0.7890625,
0.5087890625,
0.7890625
],
[
0.0906982421875,
0.79736328125,
0.473876953125,
0.798828125,
0.473876953125,
0.8125,
0.0906982421875,
0.81103515625
],
[
0.5009765625,
0.79443359375,
0.8857421875,
0.79443359375,
0.8857421875,
0.80810546875,
0.5009765625,
0.80810546875
],
[
0.0906982421875,
0.81298828125,
0.473876953125,
0.81298828125,
0.473876953125,
0.8271484375,
0.0906982421875,
0.8271484375
],
[
0.5009765625,
0.8095703125,
0.8837890625,
0.8095703125,
0.8837890625,
0.82275390625,
0.5009765625,
0.82275390625
],
[
0.0906982421875,
0.8291015625,
0.39208984375,
0.8291015625,
0.39208984375,
0.84228515625,
0.0906982421875,
0.84228515625
],
[
0.5009765625,
0.82470703125,
0.88232421875,
0.82470703125,
0.88232421875,
0.83837890625,
0.5009765625,
0.83837890625
],
[
0.50244140625,
0.8388671875,
0.8837890625,
0.8388671875,
0.8837890625,
0.85302734375,
0.50244140625,
0.85302734375
],
[
0.50244140625,
0.85498046875,
0.73876953125,
0.85498046875,
0.73876953125,
0.86865234375,
0.50244140625,
0.86865234375
],
[
0.1380615234375,
0.861328125,
0.1568603515625,
0.861328125,
0.1568603515625,
0.869140625,
0.1380615234375,
0.869140625
],
[
0.1544189453125,
0.8642578125,
0.441162109375,
0.8642578125,
0.441162109375,
0.8818359375,
0.1544189453125,
0.8818359375
],
[
0.50341796875,
0.876953125,
0.8857421875,
0.876953125,
0.8857421875,
0.89013671875,
0.50341796875,
0.89013671875
],
[
0.10540771484375,
0.8662109375,
0.1544189453125,
0.86865234375,
0.15283203125,
0.8896484375,
0.10455322265625,
0.88720703125
],
[
0.451904296875,
0.8896484375,
0.47216796875,
0.8896484375,
0.47216796875,
0.9013671875,
0.451904296875,
0.9013671875
],
[
0.5,
0.892578125,
0.8857421875,
0.892578125,
0.8857421875,
0.90576171875,
0.5,
0.90576171875
]
] |
[
"算子都可以减少高估偏差。我们还通过提供其下限和",
"，其中θ-对应于目标网络中的冻结权重，并以固定间",
"上限来量化高估的减少",
"隔更新。通过RMSPropE(Tieleman&Hinton，2012)",
"对EEq.（1）进行优化，并从重放缓冲区采样小批量",
"我们的结果是对现有工作的补充，并为探索软化贝尔",
"曼算子的各种方式的现有工作增加了新的动力。例如，",
"为了减少过高估计偏差，doubleDQN（DDQN）算法",
"正则化已被用于平滑策略。这种方法的动机包括计",
"将Qe(s,a)旨在适应Eq.（1)的目标修改为",
"算方便，探索或鲁棒性E(Foxetal.，2016;Haarnoja",
"et al., 2017; Schulman et al., 2017; Neu et al., 2017)",
"R(s, a) +Qe- (s,argmaxQe,(s',a))",
"关于值函数，EAsadi&Littman（2017）提出了",
"种替代的mellowmax算子，并证明了它是一种收缩。",
"请注意，基于最新估计0的单独网络用于动作选择，",
"他们的实验结果表明，它可以改善勘探，但是没有考",
"并且此策略的评估是由于",
"虑次优Bellman算子可以独立于勘探而导致更好的策",
"略的可能性。。因此，我们的结果为进一步研究诸如",
"Notationsoftmax函数定义为",
"mellowmax之类的运算符提供了额外的动力。。虽然与",
"[exp(TC1), exp(T2), ..., exp(TEm))T",
"softmax非常相似，但mellowmax运算符需要一些额",
"f-(x) =",
"外的计算来表示策略，如Asadi&Littman（2017）中",
"Emiexp(Tri)",
"所述。我们的论文讨论了不同贝尔曼算子之间的这种",
"，其中上标T表示向量转置。随后，softmax加权函数",
"权衡和其他权衡",
"表示为9x(）=fT(x)x，作为T的函数。此外，我们定",
"本文的其余部分组织如下：我们在E2节中提供了必",
"义了向量 Q(s,) =[Q(s,a1),Q(s,a2),...,Q(s,am)]T",
"要的背景和符号。在E3节中介绍了softmaxBellman",
"我们进一步将m设置为MDP中的动作集A的大小。",
"算子，其中提供了其收敛属性和性能界限。尽管在Q",
"最后，Rmin和Rmax分别表示最小和最大即时奖励",
"函数上是次优的，但在E4节中显示了softmax运算",
"符，在几款Atari游戏中始终优于其最大对手。如此",
"3.The SoftmaxBellman Operator",
"令人惊的结果进一步促使我们调查为什么在E5节",
"中发生这种情况。6节中对不同的Bellman算子进行",
"我们首先提供以下标准Bellman运算符",
"了彻底的比较。第7节讨论了相关工作，第8节总",
"TQ(s,a) = R(s,a) +P(s'ls,a) maxQ(s',a'),",
"结了本文",
"$",
"(2)",
"2. Background and Notation",
"我们建议使用softmaxBellman运算符，定义为",
"AMarkovdecision process（MDP）表示预期的",
"MAMTHX<XNETHX_1>中的 MAMTHX_0 元素",
"Tsoft Q(s, a) =R(s,a) +P(s'ls, a)",
"可以表示为5元组S,A,P,R,），其中S是状态空",
"s",
"间，A是动作空间，P是过渡内核，其元素XMATHX。",
"×E,explr Q(sa\")",
"Q(s,a'), (3)",
"MDP中的策略元可以用probabilitymassfunction",
"La exp[r Q(s', a)]",
"(PMF）表示，其中π(s,a）E[0,1]表示在状态s和",
"2",
"aeA元(s,a)=1中选择动作α的概率",
"sm+(Q(s*,)",
"对于给定的策略元，其状态-动作值函数Q\"(s,α)",
"，其中≥0表示逆温度参数。请注意，当T→8时，",
"被定义为在状态s中采取动作a并随后遵循策略",
"Tsoft将简化为T",
"元时累积的，预期的折扣奖励，i。e。，Q元（sa）=",
"在EAsadi&Littman（2017）中引人了mellowmax运",
"Eat~元[Zt-ortso=s,ao=al.对于最优策略元*，",
"算符，作为EEq：（3）中softmax运算符的替代方法",
"其对应的Q函数满足以下Bellman方程：",
"定义为",
"Q*(s, a) = R(s,a) +P(s'ls,a) maxQ*(s',a'),",
"mmw (Q(s,) = log(Za, explw Q(s, al),",
"(4)",
"s'",
"在EDQNE(Mnihetal.,2015）中，用神经网络将Q",
"其中w>0是调整参数。与Eq.(3）中的softmax",
"",
"函数参数化为Qe(s,a），该神经网络以状态s为输人，",
"算子sm-（Q(s）类似，当w分别接近0和8时，",
"并在最终的全连接线性层中输出相应的Q值，对于每",
"mmw收敛于均值和max算子。但是，与softmax运",
"个动作a。DQN的训练目标可以表示为",
"算符不同，Eq.（4）中的mellowmax运算符不直接提",
"供动作的概率分布，因此需要额外的步骤来获得相应",
"的策略（Asadi&Littman，2017)",
"11",
"Qo(s,a) -[R(s, a) +maxQ-(s,a')]",
"与其用于改进探索的典型用途（也是EAsadi&",
"min2",
"(1)",
"Littman（2017）中的mellowmax运算符的目的）相"
] |
{
"arxiv_id": "1812.00456",
"language": "zh",
"page_idx": 1
}
|
||
1812.00456-zh-2
|
1812.00456-zh-2.jpg
|
反，EEq.(3) 中的 softmaxBellman 运算符在本文中以非策略方式与$\operatorname{EDQN}$ 和$\operatorname{EDDQN}$ 算法结合在一起， 其中根据与$\epsilon$ -贪婪策略相同的 XMATHX\_TH
即使在$\boxed{\mathrm{E}}$Littman (1996) 中显示 softmaxBellman 运算符不是收缩，但有一个反例，我们不知道有任何已发表的工作显示在 Q 迭代中使用 softmaxBellman 运算符的性能限制。此外，我们在$\mathbb{E}$4 节中的$\mathbb{E}$令人惊讶的结果表明，仅通过用 softmax 运算符替换目标网络中的 max 运算符，就可以提高$\overline{\mathrm{EDQNs~}}$和$\overline{\mathrm{EDDQNs~}}$中的测试分数。这种改进与探索策略无关，因此激发了对 softmaxBellman 算子的理论性质的研究。
在介绍我们的主要理论结果之前，我们首先展示如何绑 定 softmax- weighted 和 max 运 算 符 之 间 的 距 离 , 随 后将其用于性能限制和高估降低的证明中
${\mathrm{Defnition~1.~}}{\widehat{\delta}}(s)\triangleq\sup_{Q}\max_{i,j}|Q(s,a_{i})-Q(s,a_{j})|$ denotes the largest distance between Q-functions w.r.t. actions for state $s$, by taking the supremum of Qfunctions and the maximum of all action pairs.
Lemma 2. By assuming $\widehat{\delta}(s)>0^1$, we have $\forall Q$ $\frac {\widehat{\delta } ( s) }{m\exp [ \tau \:\widehat{\delta } ( s) ] }$ $\leq$ $\max _{a}$ $Q( s, a)$ - $f_{\tau }^{T}\left ( Q( s, ) \right )$ $Q( s, \right )$ $\leq$
$\left(m-1\right)\operatorname*{max}\left\{\frac{1}{\tau+2},\frac{2Q_{\max}}{1+\exp(\tau)}\right\}^{2}$,where$Q_\max=\frac{R_{\max}}{1-\gamma}$ represents the maximum $Q$ -value in Q-iteration with $\mathcal{T}_{\mathrm{soft}}^{\mathrm{r}}.{}^{3}$
请注意，O'Donoghue et al. (2017) 中提供了此间隙的另一个上限。我们在这里的证明使用了不同的策略， 通过考虑具有不同动作的 Q 值之间的差异的可能值。我们进一步推导了这个差距的下限。3.1. Performance Bound for $\mathcal{T}_\mathrm{soft}$
已知最佳状态动作值函数$Q^*$是标准 Bellman 算子$\mathcal{T}$ $\mathbb{E}($Williams & Baird III, 1993) 的固定点，i。e。, $\mathcal{T}Q^{*}=Q^{*}$。由于$\mathcal{T}$是具有$\gamma$速率的收缩，因此我们还知道$\lim_k\to\infty\mathcal{T}^kQ_0=Q^*$,对于任意$Q_0$。鉴于这些事实，人们可能会想知道，作为$\tau$的函数，在$Q_0$上迭代应用$\mathcal{T}_\mathrm{soft}$代替$\mathcal{T}$将远离$Q^*$有多远
Theorem 3. Let $\mathcal{T}^kQ_0$ and $\mathcal{T}_\mathrm{soft}^kQ_0$ denote that the operators $\mathcal{T}$ and $\mathcal{T}_\mathrm{soft}$ are iteratively applied over an initial state-action value function $Q_0$ for $k$ times. Then,
$$\begin{aligned}&(I)\:\forall(s,a),\:\operatorname*{lim}_{k\to\infty}\:\mathcal{T}_{\mathrm{soft}}^{k}\:Q_{0}(s,a)\:\leq\:Q^{*}(s,a)\:\mathrm{and}\\&\operatorname*{lim}_{k\to\infty}\operatorname*{inf}_{k\to\infty}\mathcal{T}_{\mathrm{soft}}^{k}\:Q_{0}(s,a)\:\geq\\&Q^{*}(s,a)\:-\:\frac{\gamma(m-1)}{(1-\gamma)}\:\mathrm{max}\:\Big\{\frac{1}{\tau+2},\:\frac{2Q_{\mathrm{max}}}{1+\exp(\tau)}\Big\}.\end{aligned}$$
$(II)~\mathcal{T}_{\mathrm{soft}}~$converges to $\mathcal{T}~$with an exponential rate, in terms of $\tau$,i.e., the upper bound of $\mathcal{T}^kQ_0-\mathcal{T}_{\mathrm{soft}}^kQ_0$
$^{1}$Note that if $\widehat\delta(s)=0$, then all actions are equivalent
in state s , and softmax= max.
$^{2}$The sup$_Q$is over all$Q$-functions that occur during $Q$ - iteration, starting from $Q_{0}$ until the iteration terminates.
$^{3}$The Supplemental Material formally bounds $Q_\mathrm{max}.$
decays exponentially fast, as a function of $\tau$, the proof of which does not depend on the bound in part $( I)$ . 关于定理$\overline{\mathrm{E}}$3 的 $(I)$ 部分的一个值得注意的一点是，它与$\mathcal{T}_\mathrm{soft}$的负收敛结果并不矛盾，因为它的证明和结果不需要假设$\tau_\mathrm{soft}$的收敛。此外，此界限意味着$\tau_\mathrm{soft}$的非收敛性与任意范围的发散或振荡不同：即使$\mathcal{T}_\mathrm{soft}$在某些情况下可能不收敛，Q 值仍可能在合理范围内$\mathrm{E}^{4}$ 。与$\mathcal{T}$相比，$\mathcal{T}_\mathrm{soft}$还具有其他优点，如第 5节所示请注意，尽管在 Lee et al.(2018)中显示了熵正则化的 MDP 框架下 mellowmax 的界限具有更好的标量项$\log(m)$而不是$(m-1)$,但其收敛速度是线性的w。r。t。$\tau$。相反，我们的 softmaxBellman 算子结果具有指数速率，如第$(II)$部分所示，这可能会深人了解为什么在第 4 节中不需要非常大的$\tau$值。。稀疏Bellman 算子 (Lee et al., 2018) 中的误差界在$\log(m)$ 上得到了改进，但仍然是线性的。r。t。$\tau。我们进一$ 步从经验上说明了 6节的图 7 中 softmaxBellman 算子的更快收敛
# 4. Main Experiments
我们的理论结果适用于表格值函数表示和已知的下一状态分布的情况，它们限制了次优性，而不是暗示softmax 的优越性。我们主要实验的目标是表明，尽管是次优的。r。t。Q 值，通过改进 DQN 和 DDQN 算法，softmaxBellman 算子在实践中仍然有用。这种改进完全归因于将目标网络中的 max 运算符更改为softmax 运算符
我们将 softmax 运算符与$\operatorname{EDQN}$ 和$\operatorname{EDDQN}$ 结合在一起，方法是将其中的 max 函数替换为 softmax 函数，所有其他步骤都相同。相应的新算法分别称为 S-DQN 和 S-DDQN。他们的探索策略设置为$\epsilon$ - 贪婪，与 DQN 和 DDQN 相同。我们的代码提供在 https://github.com/zhao-song/Softmax-DQN
我们测试了六款雅达利游戏：Q*Bert, Ms。Pacman, CrazyClimber, Breakout, Asterix, andSeaquest。我 们的代码是建立在 Theano+Lasagne 实现从 https: //github.com/spragunr/deep\_q\_rl/ 。培训共包含 200个纪元。测试程序和所有超参数设置与EDQN 相同， 详细信息请参见$\boxed{\mathrm{FMnih~et~al.~(2015)~。基于~\{1,5,10\}}}$ 上的网格搜索选择逆温度参数$\tau$。我们还在模拟退火中实施了对数冷却方案 (Mitra et al.,1986),以逐渐增加$\tau$,但没有观察到更好的政策，与恒温相比。结果统计是通过运行五个独立的随机种子获得的
图 1显示了 Atari 游戏的平均测试分数的平均值和一个标准差，作为训练时期的函数：在前两行中，S-DQN
和 S-DDQN 通常比最大的对手获得更高的最大分数和更快的学习速度，说明了用 softmax 代替 max 的承
$^4$The lower bound can become loose when $\tau$ is extremely small, but this degenerate case has never been used in experiments.
|
[
[
0.0906982421875,
0.085205078125,
0.47216796875,
0.085205078125,
0.47216796875,
0.0985107421875,
0.0906982421875,
0.0985107421875
],
[
0.5,
0.08587646484375,
0.88623046875,
0.08587646484375,
0.88623046875,
0.0997314453125,
0.5,
0.0997314453125
],
[
0.0906982421875,
0.10162353515625,
0.48046875,
0.10162353515625,
0.48046875,
0.11553955078125,
0.0906982421875,
0.11553955078125
],
[
0.5,
0.10162353515625,
0.87744140625,
0.10162353515625,
0.87744140625,
0.11553955078125,
0.5,
0.11553955078125
],
[
0.08905029296875,
0.11614990234375,
0.4248046875,
0.11553955078125,
0.4248046875,
0.1287841796875,
0.08905029296875,
0.1300048828125
],
[
0.09234619140625,
0.1395263671875,
0.473876953125,
0.1395263671875,
0.473876953125,
0.15283203125,
0.09234619140625,
0.15283203125
],
[
0.50244140625,
0.1337890625,
0.8857421875,
0.1337890625,
0.8857421875,
0.147705078125,
0.50244140625,
0.147705078125
],
[
0.5009765625,
0.147705078125,
0.8837890625,
0.1483154296875,
0.8837890625,
0.1622314453125,
0.5009765625,
0.1610107421875
],
[
0.0906982421875,
0.1546630859375,
0.47216796875,
0.1546630859375,
0.47216796875,
0.1685791015625,
0.0906982421875,
0.1685791015625
],
[
0.0906982421875,
0.169189453125,
0.473876953125,
0.169189453125,
0.473876953125,
0.18310546875,
0.0906982421875,
0.18310546875
],
[
0.5009765625,
0.1641845703125,
0.8857421875,
0.1641845703125,
0.8857421875,
0.177978515625,
0.5009765625,
0.177978515625
],
[
0.5009765625,
0.177978515625,
0.8857421875,
0.1787109375,
0.8857421875,
0.1932373046875,
0.5009765625,
0.1925048828125
],
[
0.08905029296875,
0.184326171875,
0.47216796875,
0.184326171875,
0.47216796875,
0.1976318359375,
0.08905029296875,
0.1976318359375
],
[
0.50244140625,
0.1932373046875,
0.88232421875,
0.1932373046875,
0.88232421875,
0.20703125,
0.50244140625,
0.20703125
],
[
0.0906982421875,
0.199462890625,
0.471435546875,
0.199462890625,
0.471435546875,
0.21337890625,
0.0906982421875,
0.21337890625
],
[
0.0906982421875,
0.2139892578125,
0.47216796875,
0.2139892578125,
0.47216796875,
0.2279052734375,
0.0906982421875,
0.2279052734375
],
[
0.5009765625,
0.208984375,
0.8798828125,
0.208984375,
0.8798828125,
0.222900390625,
0.5009765625,
0.222900390625
],
[
0.0906982421875,
0.2298583984375,
0.47216796875,
0.2298583984375,
0.47216796875,
0.24365234375,
0.0906982421875,
0.24365234375
],
[
0.5009765625,
0.2310791015625,
0.88232421875,
0.2310791015625,
0.88232421875,
0.24560546875,
0.5009765625,
0.24560546875
],
[
0.0882568359375,
0.24365234375,
0.400390625,
0.2442626953125,
0.400390625,
0.2587890625,
0.0882568359375,
0.25830078125
],
[
0.50244140625,
0.246826171875,
0.8837890625,
0.246826171875,
0.8837890625,
0.260009765625,
0.50244140625,
0.260009765625
],
[
0.0906982421875,
0.267578125,
0.47216796875,
0.267578125,
0.47216796875,
0.281005859375,
0.0906982421875,
0.281005859375
],
[
0.50244140625,
0.261962890625,
0.8837890625,
0.261962890625,
0.8837890625,
0.27587890625,
0.50244140625,
0.27587890625
],
[
0.5009765625,
0.276611328125,
0.8857421875,
0.276611328125,
0.8857421875,
0.290283203125,
0.5009765625,
0.290283203125
],
[
0.0906982421875,
0.28271484375,
0.473876953125,
0.28271484375,
0.473876953125,
0.296630859375,
0.0906982421875,
0.296630859375
],
[
0.0906982421875,
0.29736328125,
0.38232421875,
0.29736328125,
0.38232421875,
0.311279296875,
0.0906982421875,
0.311279296875
],
[
0.50244140625,
0.292236328125,
0.8837890625,
0.292236328125,
0.8837890625,
0.30615234375,
0.50244140625,
0.30615234375
],
[
0.50341796875,
0.306884765625,
0.8857421875,
0.306884765625,
0.8857421875,
0.32080078125,
0.50341796875,
0.32080078125
],
[
0.08660888671875,
0.313232421875,
0.473876953125,
0.314453125,
0.473876953125,
0.33203125,
0.08660888671875,
0.330810546875
],
[
0.5,
0.32080078125,
0.8857421875,
0.322509765625,
0.8857421875,
0.337158203125,
0.5,
0.335205078125
],
[
0.08905029296875,
0.330810546875,
0.47216796875,
0.330810546875,
0.47216796875,
0.3447265625,
0.08905029296875,
0.3447265625
],
[
0.50244140625,
0.3359375,
0.87158203125,
0.337158203125,
0.87158203125,
0.3515625,
0.50244140625,
0.35107421875
],
[
0.08905029296875,
0.3466796875,
0.473876953125,
0.3466796875,
0.473876953125,
0.35986328125,
0.08905029296875,
0.35986328125
],
[
0.5,
0.35107421875,
0.8837890625,
0.3515625,
0.8837890625,
0.3662109375,
0.5,
0.365478515625
],
[
0.08905029296875,
0.361083984375,
0.4248046875,
0.361083984375,
0.4248046875,
0.375,
0.08905029296875,
0.375
],
[
0.5048828125,
0.36865234375,
0.5986328125,
0.36865234375,
0.5986328125,
0.380126953125,
0.5048828125,
0.380126953125
],
[
0.08905029296875,
0.37744140625,
0.47216796875,
0.37744140625,
0.47216796875,
0.392578125,
0.08905029296875,
0.392578125
],
[
0.09234619140625,
0.397705078125,
0.476318359375,
0.389404296875,
0.476318359375,
0.4140625,
0.09234619140625,
0.42236328125
],
[
0.131591796875,
0.393310546875,
0.16015625,
0.393310546875,
0.16015625,
0.40478515625,
0.131591796875,
0.40478515625
],
[
0.5,
0.401611328125,
0.6845703125,
0.40283203125,
0.6845703125,
0.417236328125,
0.5,
0.416748046875
],
[
0.2303466796875,
0.41845703125,
0.26708984375,
0.421630859375,
0.265625,
0.429931640625,
0.2296142578125,
0.4267578125
],
[
0.29736328125,
0.419189453125,
0.465576171875,
0.41845703125,
0.465576171875,
0.436279296875,
0.29736328125,
0.4375
],
[
0.294189453125,
0.42236328125,
0.303955078125,
0.423583984375,
0.30224609375,
0.427978515625,
0.293212890625,
0.4267578125
],
[
0.08905029296875,
0.423583984375,
0.1871337890625,
0.423583984375,
0.1871337890625,
0.43505859375,
0.08905029296875,
0.43505859375
],
[
0.178955078125,
0.427978515625,
0.289306640625,
0.4287109375,
0.289306640625,
0.43994140625,
0.178955078125,
0.439453125
],
[
0.188720703125,
0.42236328125,
0.2017822265625,
0.42236328125,
0.2017822265625,
0.4267578125,
0.188720703125,
0.4267578125
],
[
0.44287109375,
0.43115234375,
0.465576171875,
0.43115234375,
0.465576171875,
0.438232421875,
0.44287109375,
0.438232421875
],
[
0.5009765625,
0.427978515625,
0.8837890625,
0.427978515625,
0.8837890625,
0.44140625,
0.5009765625,
0.44140625
],
[
0.08905029296875,
0.43994140625,
0.471435546875,
0.43994140625,
0.471435546875,
0.455078125,
0.08905029296875,
0.455078125
],
[
0.5146484375,
0.443115234375,
0.8837890625,
0.443115234375,
0.8837890625,
0.45703125,
0.5146484375,
0.45703125
],
[
0.0882568359375,
0.455810546875,
0.148681640625,
0.453857421875,
0.1502685546875,
0.4697265625,
0.08905029296875,
0.4716796875
],
[
0.50244140625,
0.458251953125,
0.8837890625,
0.458251953125,
0.8837890625,
0.4716796875,
0.50244140625,
0.4716796875
],
[
0.08905029296875,
0.4716796875,
0.473876953125,
0.4716796875,
0.473876953125,
0.486083984375,
0.08905029296875,
0.486083984375
],
[
0.50244140625,
0.473388671875,
0.8857421875,
0.473388671875,
0.8857421875,
0.4873046875,
0.50244140625,
0.4873046875
],
[
0.0906982421875,
0.4873046875,
0.4755859375,
0.4873046875,
0.4755859375,
0.50048828125,
0.0906982421875,
0.50048828125
],
[
0.5009765625,
0.488037109375,
0.8857421875,
0.488037109375,
0.8857421875,
0.50244140625,
0.5009765625,
0.50244140625
],
[
0.09234619140625,
0.50390625,
0.473876953125,
0.50390625,
0.473876953125,
0.5146484375,
0.09234619140625,
0.5146484375
],
[
0.5009765625,
0.50390625,
0.8857421875,
0.50390625,
0.8857421875,
0.517578125,
0.5009765625,
0.517578125
],
[
0.08905029296875,
0.517578125,
0.34228515625,
0.517578125,
0.34228515625,
0.53076171875,
0.08905029296875,
0.53076171875
],
[
0.5009765625,
0.51953125,
0.61376953125,
0.51953125,
0.61376953125,
0.53271484375,
0.5009765625,
0.53271484375
],
[
0.08660888671875,
0.54248046875,
0.328369140625,
0.54443359375,
0.328369140625,
0.5595703125,
0.08660888671875,
0.556640625
],
[
0.50244140625,
0.54150390625,
0.8837890625,
0.54150390625,
0.8837890625,
0.5546875,
0.50244140625,
0.5546875
],
[
0.5009765625,
0.55615234375,
0.8837890625,
0.55615234375,
0.8837890625,
0.5693359375,
0.5009765625,
0.5693359375
],
[
0.09234619140625,
0.568359375,
0.473876953125,
0.568359375,
0.473876953125,
0.58203125,
0.09234619140625,
0.58203125
],
[
0.50244140625,
0.5712890625,
0.8857421875,
0.5712890625,
0.8857421875,
0.58544921875,
0.50244140625,
0.58544921875
],
[
0.0906982421875,
0.58349609375,
0.46728515625,
0.58349609375,
0.46728515625,
0.5966796875,
0.0906982421875,
0.5966796875
],
[
0.50244140625,
0.5859375,
0.8798828125,
0.5859375,
0.8798828125,
0.599609375,
0.50244140625,
0.599609375
],
[
0.0906982421875,
0.59765625,
0.471435546875,
0.59765625,
0.471435546875,
0.611328125,
0.0906982421875,
0.611328125
],
[
0.50244140625,
0.6015625,
0.8857421875,
0.6015625,
0.8857421875,
0.61572265625,
0.50244140625,
0.61572265625
],
[
0.0906982421875,
0.61279296875,
0.47216796875,
0.61279296875,
0.47216796875,
0.62744140625,
0.0906982421875,
0.62744140625
],
[
0.5009765625,
0.6171875,
0.85546875,
0.6171875,
0.85546875,
0.630859375,
0.5009765625,
0.630859375
],
[
0.08905029296875,
0.626953125,
0.47216796875,
0.62744140625,
0.47216796875,
0.642578125,
0.08905029296875,
0.6416015625
],
[
0.0906982421875,
0.64306640625,
0.38232421875,
0.64306640625,
0.38232421875,
0.6572265625,
0.0906982421875,
0.6572265625
],
[
0.50244140625,
0.6396484375,
0.8837890625,
0.6396484375,
0.8837890625,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.5009765625,
0.65283203125,
0.8857421875,
0.65380859375,
0.8857421875,
0.66845703125,
0.5009765625,
0.66748046875
],
[
0.0882568359375,
0.66015625,
0.47216796875,
0.6611328125,
0.47216796875,
0.67529296875,
0.0882568359375,
0.6748046875
],
[
0.50244140625,
0.66943359375,
0.8857421875,
0.66943359375,
0.8857421875,
0.68310546875,
0.50244140625,
0.68310546875
],
[
0.08905029296875,
0.6767578125,
0.473876953125,
0.6767578125,
0.473876953125,
0.68994140625,
0.08905029296875,
0.68994140625
],
[
0.0906982421875,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.0906982421875,
0.7041015625
],
[
0.50244140625,
0.68505859375,
0.8857421875,
0.68505859375,
0.8857421875,
0.69873046875,
0.50244140625,
0.69873046875
],
[
0.50341796875,
0.69970703125,
0.88232421875,
0.69970703125,
0.88232421875,
0.71337890625,
0.50341796875,
0.71337890625
],
[
0.08660888671875,
0.7109375,
0.426513671875,
0.7119140625,
0.426513671875,
0.72998046875,
0.08660888671875,
0.72900390625
],
[
0.50244140625,
0.71533203125,
0.88134765625,
0.71533203125,
0.88134765625,
0.72900390625,
0.50244140625,
0.72900390625
],
[
0.11517333984375,
0.73486328125,
0.273681640625,
0.73486328125,
0.273681640625,
0.75244140625,
0.11517333984375,
0.75244140625
],
[
0.5048828125,
0.72998046875,
0.88232421875,
0.72998046875,
0.88232421875,
0.74365234375,
0.5048828125,
0.74365234375
],
[
0.101318359375,
0.7470703125,
0.455810546875,
0.76123046875,
0.453369140625,
0.8037109375,
0.098876953125,
0.7890625
],
[
0.5009765625,
0.74365234375,
0.88232421875,
0.744140625,
0.88232421875,
0.75830078125,
0.5009765625,
0.7568359375
],
[
0.29736328125,
0.76123046875,
0.36767578125,
0.7587890625,
0.36767578125,
0.7734375,
0.29736328125,
0.77587890625
],
[
0.366943359375,
0.76123046875,
0.4306640625,
0.76123046875,
0.4306640625,
0.77587890625,
0.366943359375,
0.77587890625
],
[
0.50244140625,
0.76025390625,
0.8857421875,
0.76025390625,
0.8857421875,
0.7734375,
0.50244140625,
0.7734375
],
[
0.50244140625,
0.77587890625,
0.82861328125,
0.77587890625,
0.82861328125,
0.7890625,
0.50244140625,
0.7890625
],
[
0.363525390625,
0.7646484375,
0.369384765625,
0.7646484375,
0.369384765625,
0.77099609375,
0.363525390625,
0.77099609375
],
[
0.19775390625,
0.77783203125,
0.2607421875,
0.77783203125,
0.2607421875,
0.79248046875,
0.19775390625,
0.79248046875
],
[
0.08905029296875,
0.798828125,
0.473876953125,
0.798828125,
0.473876953125,
0.81298828125,
0.08905029296875,
0.81298828125
],
[
0.5009765625,
0.79736328125,
0.87744140625,
0.79736328125,
0.87744140625,
0.81103515625,
0.5009765625,
0.81103515625
],
[
0.0882568359375,
0.81103515625,
0.473876953125,
0.8125,
0.473876953125,
0.830078125,
0.0882568359375,
0.8291015625
],
[
0.50244140625,
0.81298828125,
0.8857421875,
0.81298828125,
0.8857421875,
0.8271484375,
0.50244140625,
0.8271484375
],
[
0.50244140625,
0.82763671875,
0.8857421875,
0.82763671875,
0.8857421875,
0.84130859375,
0.50244140625,
0.84130859375
],
[
0.1070556640625,
0.83837890625,
0.47216796875,
0.8388671875,
0.47216796875,
0.85400390625,
0.1070556640625,
0.85302734375
],
[
0.50244140625,
0.84326171875,
0.8857421875,
0.84326171875,
0.8857421875,
0.85693359375,
0.50244140625,
0.85693359375
],
[
0.0882568359375,
0.85107421875,
0.295654296875,
0.85302734375,
0.295654296875,
0.8662109375,
0.0882568359375,
0.8642578125
],
[
0.1094970703125,
0.8642578125,
0.473876953125,
0.8662109375,
0.473876953125,
0.880859375,
0.1094970703125,
0.87890625
],
[
0.5185546875,
0.86572265625,
0.8837890625,
0.8671875,
0.8837890625,
0.880859375,
0.5185546875,
0.87890625
],
[
0.08660888671875,
0.87890625,
0.468017578125,
0.8798828125,
0.468017578125,
0.89306640625,
0.08660888671875,
0.892578125
],
[
0.5009765625,
0.8798828125,
0.8857421875,
0.8798828125,
0.8857421875,
0.89306640625,
0.5009765625,
0.89306640625
],
[
0.108642578125,
0.8916015625,
0.45751953125,
0.89306640625,
0.45751953125,
0.9072265625,
0.108642578125,
0.90478515625
],
[
0.5009765625,
0.89453125,
0.5703125,
0.89453125,
0.5703125,
0.90478515625,
0.5009765625,
0.90478515625
]
] |
[
"反，EE.（3）中的softmaxBellman运算符在本文中",
"decays exponentially fast, as a function of T, the proof",
"以非策略方式与EDQN和EDDQN算法结合在一起，",
"of which does not depend on the bound in part ().",
"其中根据与E-贪婪策略相同的XMATHXTH",
"即使在ELittman（1996）中显示softmaxBellman运算",
"关于定理E3的（I）部分的一个值得注意的一点是，它",
"与Toft的负收敛结果并不矛盾，因为它的证明和结果",
"符不是收缩，但有一个反例，我们不知道有任何已发",
"表的工作显示在Q选代中使用softmaxBellman运算",
"不需要假设Tsoft的收敛。此外，此界限意味着Tsoft的",
"非收敛性与任意范围的发散或振荡不同：即使Tsoft在",
"符的性能限制。此外，我们在E4节中的E令人惊的",
"某些情况下可能不收敛，Q值仍可能在合理范围内E",
"结果表明，仅通过用softmax运算符替换目标网络中",
"的max运算符，就可以提高EDQNs和EDDQNs中",
"。与T相比，Tsoft还具有其他优点，如第5节所示",
"的测试分数。这种改进与探索策略无关，因此激发了",
"请注意，尽管在Leeetal.（2018）中显示了摘正则化",
"对softmaxBellman算子的理论性质的研究。",
"的MDP框架下mellowmax的界限具有更好的标量",
"在介绍我们的主要理论结果之前，我们首先展示如何",
"项log(m）而不是（m一1），但其收敛速度是线性的",
"w。r。t。T。相反，我们的softmaxBellman算子结",
"绑定softmax-weighted和max运算符之间的距离，随",
"后将其用于性能限制和高估降低的证明中",
"果具有指数速率，如第（II）部分所示，这可能会深人",
"了解为什么在第4节中不需要非常大的T值。。稀疏",
"Definition 1. (s) ≤ supQ maxi,; Q(s, as) -Q(s,aj)",
"Bellman算子（Leeetal.,2018）中的误差界在log(m)",
"denotes the largest distance between Q-functions w.r.t.",
"上得到了改进，但仍然是线性的。r。t。T。我们进",
"actions for state s,by taking the supremum of Q-",
"步从经验上说明了6节的图7中softmaxBellman算",
"functions and the maximum of all action pairs.",
"子的更快收敛",
"Lemma 2. By assuming (s) > 0 1 , we have VQ",
"mexps(o), ≤ maxa Q(s,a) - FT(Q(s,) Q(s,) ≤",
"8(s)",
"4.Main Experiments",
"2Qma",
"2, where Qmax = Rma",
"",
"(m-1)max3",
"T+21+exp()",
"",
"1-2",
"我们的理论结果适用于表格值函数表示和已知的下",
"represents the maximum Q-value in Q-iteration with",
"状态分布的情况，它们限制了次优性，而不是暗示",
"Tsoft3",
"softmax的优越性。我们主要实验的目标是表明，尽管",
"请注意，，O'Donoghueetal.（2017）中提供了此间隙",
"是次优的。r。t。Q值，通过改进DQN和DDQN",
"的另一个上限。我们在这里的证明使用了不同的策略，",
"算法，softmaxBellman算子在实践中仍然有用。这种",
"通过考虑具有不同动作的Q值之间的差异的可能值。",
"改进完全归因于将目标网络中的max运算符更改为",
"我们进一步推导了这个差距的下限。",
"softmax运算符",
"3.1.PerformanceBound for Tsoft",
"我们将softmax运算符与EDQN和EDDQN结合",
"在一起，方法是将其中的max函数替换为softmax",
"已知最佳状态动作值函数Q*是标准Bellman算子",
"函数，所有其他步骤都相同。相应的新算法分别称",
"TE(Williams&BairdIII,1993）的固定点，i。e。",
"为S-DQN和S-DDQN。他们的探索策略设置为E",
"TQ*=Q*。由于T是具有速率的收缩，因此我们",
"贪婪，与DQN和DDQN相同。我们的代码提供",
"还知道limk→αoTkQo=Q*，对于任意Qo。鉴于这",
"在https://github.com/zhao-song/Softmax-DQN",
"些事实，人们可能会想知道，作为↑的函数，在Q。上",
"送代应用Tsoft代替T将远离Q*有多远",
"我们测试了六款雅达利游戏：Q*Bert，Ms。Pac",
"man,CrazyClimber,Breakout,Asterix,andSeaquest。我",
"Theorem 3. Let TkQo and Toft Qo denote that the op",
"们的代码是建立在Theano+Lasagne实现从https:",
"erators T and Tsoft are iteratively applied over an ini-",
"tial state-action value function Qo for k times. Then.",
"//github.com/spragunr/deep_q_rl/。培训共包含200",
"个纪元。测试程序和所有超参数设置与EDQN相同",
"(I) V(s, a), lim sup Toft Qo(s,a) ≤ Q*(s, a) and",
"详细信息请参见EMnihetal.（2015）。基于【1,5,10】",
"lim inf Toft Qo(s, a) ≥",
"上的网格搜索选择逆温度参数。我们还在模拟退火",
"0() -(x(2 )",
"中实施了对数冷却方案（Mitraetal，1986），以逐渐",
"1",
"2Qmax",
"增加，但没有观察到更好的政策，与恒温相比。结",
"果统计是通过运行五个独立的随机种子获得的",
"",
"(1-)",
"(II) Tsoft converges to T with an exponential rate, in",
"图1显示了Atari游戏的平均测试分数的平均值和一",
"terms of T , i.e., the upper bound of ThQo - ToftQo",
"个标准差，作为训练时期的函数：在前两行中，S-DQN",
"和S-DDQN通常比最大的对手获得更高的最大分数",
"INote that if (s) = 0, then all actions are equivalent",
"和更快的学习速度，说明了用softmax代替max的承",
"in state s , and softmax = max",
"2The supo is over all Q -functions that occur during Q",
"4The lower bound can become loose when is extremely",
"-iteration, starting from Qo until the iteration terminates.",
"small, but this degenerate case has never been used in ex-",
"3The Supplemental Material formallybounds Qmax.",
"periments."
] |
{
"arxiv_id": "1812.00456",
"language": "zh",
"page_idx": 2
}
|
||
1812.00456-zh-3
|
1812.00456-zh-3.jpg
|
$Q^*Bert$
Ms. Pacman
Breakout
Crazy Climber
Asterix
<FigureHere>
Figure 1. Performance of (top row) S-DQN vs. DQN, (middle row) S-DDQN vs. DDQN, and (bottom row) S-DQN vs. DDQN on the Atari games. X-axis and Y-axis correspond to the training epoch and test score, respectively. All curves are averaged over five independent random seeds.
诺。对于游戏 Asterix , 通常用作EDQN E(van Hasselt et al., 2016a; Anschel et al., 2017) 中高估问题的示例， S-DQN 和 S-DDQN 的测试得分均远高于其最大对应物，这表明 softmax 运算符可以减轻高估偏差。尽管
不像 Asterix 那样引人注目，但图$\boxed{\mathrm{E}}1$ 的底行显示，对于整个测试套件，S-DQN 通常优于 DDQN, 具有更高的测试分数，这表明在存在函数逼近的情况下，可以使用 softmax 运算符而不是双 Q 学习$\overline{\mathrm{E}}($van Hasselt,
|
[
[
0.2017822265625,
0.080810546875,
0.23779296875,
0.080810546875,
0.23779296875,
0.0921630859375,
0.2017822265625,
0.0921630859375
],
[
0.324462890625,
0.080810546875,
0.38232421875,
0.080810546875,
0.38232421875,
0.0921630859375,
0.324462890625,
0.0921630859375
],
[
0.464111328125,
0.080810546875,
0.5107421875,
0.080810546875,
0.5107421875,
0.0921630859375,
0.464111328125,
0.0921630859375
],
[
0.58642578125,
0.08209228515625,
0.65625,
0.08209228515625,
0.65625,
0.0921630859375,
0.58642578125,
0.0921630859375
],
[
0.73779296875,
0.080810546875,
0.77392578125,
0.080810546875,
0.77392578125,
0.0921630859375,
0.73779296875,
0.0921630859375
],
[
0.293212890625,
0.0947265625,
0.310546875,
0.0947265625,
0.310546875,
0.0985107421875,
0.293212890625,
0.0985107421875
],
[
0.56201171875,
0.0947265625,
0.5849609375,
0.0947265625,
0.5849609375,
0.0985107421875,
0.56201171875,
0.0985107421875
],
[
0.69384765625,
0.0921630859375,
0.71337890625,
0.0889892578125,
0.7158203125,
0.09783935546875,
0.6953125,
0.0997314453125
],
[
0.69384765625,
0.10101318359375,
0.7197265625,
0.10101318359375,
0.7197265625,
0.11236572265625,
0.69384765625,
0.11236572265625
],
[
0.291748046875,
0.1029052734375,
0.31201171875,
0.1029052734375,
0.31201171875,
0.11297607421875,
0.291748046875,
0.11297607421875
],
[
0.156005859375,
0.09783935546875,
0.181396484375,
0.09783935546875,
0.181396484375,
0.10919189453125,
0.156005859375,
0.10919189453125
],
[
0.560546875,
0.1060791015625,
0.58251953125,
0.1060791015625,
0.58251953125,
0.1142578125,
0.560546875,
0.1142578125
],
[
0.156005859375,
0.11236572265625,
0.181396484375,
0.11236572265625,
0.181396484375,
0.12371826171875,
0.156005859375,
0.12371826171875
],
[
0.426513671875,
0.11236572265625,
0.441162109375,
0.11236572265625,
0.441162109375,
0.12060546875,
0.426513671875,
0.12060546875
],
[
0.69384765625,
0.11297607421875,
0.71826171875,
0.11297607421875,
0.71826171875,
0.1256103515625,
0.69384765625,
0.1256103515625
],
[
0.35546875,
0.11614990234375,
0.393798828125,
0.11614990234375,
0.393798828125,
0.126953125,
0.35546875,
0.126953125
],
[
0.291748046875,
0.1142578125,
0.31201171875,
0.1142578125,
0.31201171875,
0.125,
0.291748046875,
0.125
],
[
0.48193359375,
0.12060546875,
0.5361328125,
0.12060546875,
0.5361328125,
0.1319580078125,
0.48193359375,
0.1319580078125
],
[
0.56396484375,
0.11871337890625,
0.58251953125,
0.11871337890625,
0.58251953125,
0.126953125,
0.56396484375,
0.126953125
],
[
0.16015625,
0.125,
0.1829833984375,
0.125,
0.1829833984375,
0.1363525390625,
0.16015625,
0.1363525390625
],
[
0.2900390625,
0.126953125,
0.3095703125,
0.1256103515625,
0.3095703125,
0.1351318359375,
0.291748046875,
0.1363525390625
],
[
0.69677734375,
0.125,
0.7197265625,
0.125,
0.7197265625,
0.136962890625,
0.69677734375,
0.136962890625
],
[
0.25244140625,
0.1275634765625,
0.268798828125,
0.12060546875,
0.273681640625,
0.1275634765625,
0.259033203125,
0.1363525390625
],
[
0.2442626953125,
0.1275634765625,
0.2548828125,
0.1275634765625,
0.2548828125,
0.1307373046875,
0.2442626953125,
0.1307373046875
],
[
0.4306640625,
0.1319580078125,
0.435546875,
0.1319580078125,
0.435546875,
0.1383056640625,
0.4306640625,
0.1383056640625
],
[
0.5595703125,
0.1307373046875,
0.5849609375,
0.1275634765625,
0.58837890625,
0.1383056640625,
0.56201171875,
0.141357421875
],
[
0.291748046875,
0.1383056640625,
0.3095703125,
0.1383056640625,
0.3095703125,
0.147705078125,
0.291748046875,
0.147705078125
],
[
0.69677734375,
0.1363525390625,
0.7197265625,
0.1363525390625,
0.7197265625,
0.149658203125,
0.69677734375,
0.149658203125
],
[
0.56640625,
0.146484375,
0.5810546875,
0.146484375,
0.5810546875,
0.15087890625,
0.56640625,
0.15087890625
],
[
0.1641845703125,
0.14013671875,
0.17724609375,
0.14013671875,
0.17724609375,
0.1483154296875,
0.1641845703125,
0.1483154296875
],
[
0.2344970703125,
0.1483154296875,
0.271240234375,
0.1483154296875,
0.271240234375,
0.1590576171875,
0.2344970703125,
0.1590576171875
],
[
0.2900390625,
0.1483154296875,
0.313720703125,
0.1483154296875,
0.313720703125,
0.160400390625,
0.2900390625,
0.160400390625
],
[
0.69677734375,
0.147705078125,
0.71826171875,
0.147705078125,
0.71826171875,
0.160400390625,
0.69677734375,
0.160400390625
],
[
0.2344970703125,
0.1590576171875,
0.261474609375,
0.1590576171875,
0.261474609375,
0.17041015625,
0.2344970703125,
0.17041015625
],
[
0.29736328125,
0.1622314453125,
0.31787109375,
0.1622314453125,
0.31787109375,
0.169189453125,
0.29736328125,
0.169189453125
],
[
0.56396484375,
0.1578369140625,
0.58837890625,
0.1578369140625,
0.58837890625,
0.1666259765625,
0.56396484375,
0.1666259765625
],
[
0.4345703125,
0.169189453125,
0.44677734375,
0.169189453125,
0.44677734375,
0.173583984375,
0.4345703125,
0.173583984375
],
[
0.35693359375,
0.1767578125,
0.370849609375,
0.1767578125,
0.370849609375,
0.186279296875,
0.35693359375,
0.186279296875
],
[
0.400390625,
0.1767578125,
0.414306640625,
0.1767578125,
0.414306640625,
0.1849365234375,
0.400390625,
0.1849365234375
],
[
0.488525390625,
0.1767578125,
0.50244140625,
0.1767578125,
0.50244140625,
0.1849365234375,
0.488525390625,
0.1849365234375
],
[
0.53515625,
0.1767578125,
0.54736328125,
0.1767578125,
0.54736328125,
0.1849365234375,
0.53515625,
0.1849365234375
],
[
0.6298828125,
0.1767578125,
0.64208984375,
0.1767578125,
0.64208984375,
0.1849365234375,
0.6298828125,
0.1849365234375
],
[
0.26953125,
0.177978515625,
0.27783203125,
0.177978515625,
0.27783203125,
0.18310546875,
0.26953125,
0.18310546875
],
[
0.226318359375,
0.177978515625,
0.236083984375,
0.177978515625,
0.236083984375,
0.1849365234375,
0.226318359375,
0.1849365234375
],
[
0.67333984375,
0.1787109375,
0.6806640625,
0.1787109375,
0.6806640625,
0.18310546875,
0.67333984375,
0.18310546875
],
[
0.463134765625,
0.20703125,
0.5107421875,
0.20703125,
0.5107421875,
0.2171630859375,
0.463134765625,
0.2171630859375
],
[
0.4248046875,
0.2259521484375,
0.451904296875,
0.2259521484375,
0.451904296875,
0.2374267578125,
0.4248046875,
0.2374267578125
],
[
0.428955078125,
0.2392578125,
0.451904296875,
0.2392578125,
0.451904296875,
0.250732421875,
0.428955078125,
0.250732421875
],
[
0.5,
0.24560546875,
0.515625,
0.24560546875,
0.515625,
0.25,
0.5,
0.25
],
[
0.428955078125,
0.25244140625,
0.451904296875,
0.25244140625,
0.451904296875,
0.264404296875,
0.428955078125,
0.264404296875
],
[
0.428955078125,
0.264404296875,
0.450927734375,
0.264404296875,
0.450927734375,
0.276611328125,
0.428955078125,
0.276611328125
],
[
0.431396484375,
0.27978515625,
0.44677734375,
0.27978515625,
0.44677734375,
0.28857421875,
0.431396484375,
0.28857421875
],
[
0.29736328125,
0.323974609375,
0.303955078125,
0.323974609375,
0.303955078125,
0.326904296875,
0.29736328125,
0.326904296875
],
[
0.423095703125,
0.326904296875,
0.441162109375,
0.323974609375,
0.443603515625,
0.333251953125,
0.426513671875,
0.3359375
],
[
0.69384765625,
0.326904296875,
0.71728515625,
0.326904296875,
0.71728515625,
0.335205078125,
0.69384765625,
0.335205078125
],
[
0.2900390625,
0.333251953125,
0.313720703125,
0.333251953125,
0.313720703125,
0.3447265625,
0.2900390625,
0.3447265625
],
[
0.560546875,
0.335205078125,
0.583984375,
0.335205078125,
0.583984375,
0.343505859375,
0.560546875,
0.343505859375
],
[
0.15283203125,
0.32763671875,
0.1805419921875,
0.324462890625,
0.181396484375,
0.3359375,
0.156005859375,
0.339111328125
],
[
0.609375,
0.3359375,
0.67724609375,
0.3359375,
0.67724609375,
0.35107421875,
0.609375,
0.35107421875
],
[
0.69287109375,
0.3359375,
0.72119140625,
0.3359375,
0.72119140625,
0.347900390625,
0.69287109375,
0.347900390625
],
[
0.3349609375,
0.339111328125,
0.406982421875,
0.335205078125,
0.408447265625,
0.349853515625,
0.336669921875,
0.354248046875
],
[
0.15283203125,
0.34033203125,
0.181396484375,
0.33837890625,
0.1829833984375,
0.349853515625,
0.156005859375,
0.352783203125
],
[
0.291748046875,
0.347900390625,
0.3095703125,
0.347900390625,
0.3095703125,
0.356689453125,
0.291748046875,
0.356689453125
],
[
0.42724609375,
0.343505859375,
0.439453125,
0.343505859375,
0.439453125,
0.3515625,
0.42724609375,
0.3515625
],
[
0.56396484375,
0.348388671875,
0.583984375,
0.348388671875,
0.583984375,
0.356689453125,
0.56396484375,
0.356689453125
],
[
0.69140625,
0.347900390625,
0.7197265625,
0.34521484375,
0.72119140625,
0.35791015625,
0.69287109375,
0.35986328125
],
[
0.42724609375,
0.35986328125,
0.439453125,
0.35986328125,
0.439453125,
0.369384765625,
0.42724609375,
0.369384765625
],
[
0.160888671875,
0.352783203125,
0.1829833984375,
0.352783203125,
0.1829833984375,
0.365478515625,
0.160888671875,
0.365478515625
],
[
0.291748046875,
0.363037109375,
0.3095703125,
0.363037109375,
0.3095703125,
0.371826171875,
0.291748046875,
0.371826171875
],
[
0.6953125,
0.361083984375,
0.71826171875,
0.35791015625,
0.72119140625,
0.369384765625,
0.69775390625,
0.37255859375
],
[
0.160888671875,
0.367431640625,
0.1805419921875,
0.367431640625,
0.1805419921875,
0.380126953125,
0.160888671875,
0.380126953125
],
[
0.56201171875,
0.37255859375,
0.5908203125,
0.37255859375,
0.5908203125,
0.3837890625,
0.56201171875,
0.3837890625
],
[
0.69677734375,
0.371826171875,
0.72119140625,
0.371826171875,
0.72119140625,
0.3837890625,
0.69677734375,
0.3837890625
],
[
0.73876953125,
0.37255859375,
0.810546875,
0.369384765625,
0.810546875,
0.380615234375,
0.73876953125,
0.3837890625
],
[
0.42724609375,
0.37744140625,
0.441162109375,
0.37744140625,
0.441162109375,
0.38818359375,
0.42724609375,
0.38818359375
],
[
0.2296142578125,
0.376953125,
0.271240234375,
0.376953125,
0.271240234375,
0.38818359375,
0.2296142578125,
0.38818359375
],
[
0.16015625,
0.38330078125,
0.176513671875,
0.380126953125,
0.1805419921875,
0.39013671875,
0.1641845703125,
0.393310546875
],
[
0.56201171875,
0.386474609375,
0.5888671875,
0.386474609375,
0.5888671875,
0.397705078125,
0.56201171875,
0.397705078125
],
[
0.69921875,
0.385009765625,
0.71337890625,
0.385009765625,
0.71337890625,
0.393310546875,
0.69921875,
0.393310546875
],
[
0.295654296875,
0.393310546875,
0.310546875,
0.391357421875,
0.313720703125,
0.397705078125,
0.298095703125,
0.399658203125
],
[
0.4306640625,
0.397705078125,
0.44775390625,
0.39453125,
0.450927734375,
0.40087890625,
0.43310546875,
0.404052734375
],
[
0.2303466796875,
0.389404296875,
0.2607421875,
0.389404296875,
0.2607421875,
0.397705078125,
0.2303466796875,
0.397705078125
],
[
0.406982421875,
0.399658203125,
0.418212890625,
0.4072265625,
0.406982421875,
0.417236328125,
0.396240234375,
0.410400390625
],
[
0.62744140625,
0.40283203125,
0.6435546875,
0.406005859375,
0.640625,
0.41357421875,
0.625,
0.410400390625
],
[
0.35546875,
0.40478515625,
0.373291015625,
0.40478515625,
0.373291015625,
0.41357421875,
0.35546875,
0.41357421875
],
[
0.22216796875,
0.404052734375,
0.238525390625,
0.404052734375,
0.238525390625,
0.41357421875,
0.22216796875,
0.41357421875
],
[
0.48779296875,
0.40478515625,
0.50244140625,
0.40478515625,
0.50244140625,
0.41357421875,
0.48779296875,
0.41357421875
],
[
0.6708984375,
0.40478515625,
0.68115234375,
0.406005859375,
0.6806640625,
0.410888671875,
0.6689453125,
0.410400390625
],
[
0.7607421875,
0.40478515625,
0.775390625,
0.40478515625,
0.775390625,
0.41357421875,
0.7607421875,
0.41357421875
],
[
0.59228515625,
0.4072265625,
0.59716796875,
0.4072265625,
0.59716796875,
0.412353515625,
0.59228515625,
0.412353515625
],
[
0.26708984375,
0.406005859375,
0.279296875,
0.406005859375,
0.279296875,
0.412353515625,
0.26708984375,
0.412353515625
],
[
0.53369140625,
0.4091796875,
0.54150390625,
0.40283203125,
0.54833984375,
0.40771484375,
0.5400390625,
0.4140625
],
[
0.8037109375,
0.4072265625,
0.81640625,
0.4072265625,
0.81640625,
0.410888671875,
0.8037109375,
0.410888671875
],
[
0.4248046875,
0.45703125,
0.44775390625,
0.45703125,
0.44775390625,
0.4697265625,
0.4248046875,
0.4697265625
],
[
0.4248046875,
0.472900390625,
0.44775390625,
0.472900390625,
0.44775390625,
0.48486328125,
0.4248046875,
0.48486328125
],
[
0.42724609375,
0.4892578125,
0.441162109375,
0.4892578125,
0.441162109375,
0.498779296875,
0.42724609375,
0.498779296875
],
[
0.49267578125,
0.5322265625,
0.5048828125,
0.5322265625,
0.5048828125,
0.53857421875,
0.49267578125,
0.53857421875
],
[
0.5400390625,
0.52880859375,
0.54833984375,
0.53515625,
0.54150390625,
0.5390625,
0.53515625,
0.5341796875
],
[
0.291748046875,
0.5498046875,
0.310546875,
0.5498046875,
0.310546875,
0.556640625,
0.291748046875,
0.556640625
],
[
0.56201171875,
0.5517578125,
0.5810546875,
0.55029296875,
0.5810546875,
0.5546875,
0.56201171875,
0.55615234375
],
[
0.6953125,
0.54736328125,
0.71337890625,
0.54736328125,
0.71337890625,
0.55615234375,
0.6953125,
0.55615234375
],
[
0.156005859375,
0.5546875,
0.181396484375,
0.5546875,
0.181396484375,
0.5673828125,
0.156005859375,
0.5673828125
],
[
0.42724609375,
0.55322265625,
0.439453125,
0.55322265625,
0.439453125,
0.5595703125,
0.42724609375,
0.5595703125
],
[
0.2900390625,
0.5595703125,
0.313720703125,
0.5595703125,
0.313720703125,
0.57080078125,
0.2900390625,
0.57080078125
],
[
0.560546875,
0.5625,
0.58837890625,
0.5625,
0.58837890625,
0.57080078125,
0.560546875,
0.57080078125
],
[
0.61279296875,
0.5625,
0.6787109375,
0.5625,
0.6787109375,
0.58056640625,
0.61279296875,
0.58056640625
],
[
0.69384765625,
0.56005859375,
0.71728515625,
0.56005859375,
0.71728515625,
0.568359375,
0.69384765625,
0.568359375
],
[
0.1568603515625,
0.5673828125,
0.181396484375,
0.5673828125,
0.181396484375,
0.580078125,
0.1568603515625,
0.580078125
],
[
0.291748046875,
0.5712890625,
0.31201171875,
0.5712890625,
0.31201171875,
0.58203125,
0.291748046875,
0.58203125
],
[
0.423095703125,
0.5693359375,
0.439453125,
0.56640625,
0.44287109375,
0.5771484375,
0.42724609375,
0.580078125
],
[
0.69287109375,
0.5712890625,
0.7197265625,
0.5712890625,
0.7197265625,
0.58349609375,
0.69287109375,
0.58349609375
],
[
0.56201171875,
0.57373046875,
0.5888671875,
0.57373046875,
0.5888671875,
0.58544921875,
0.56201171875,
0.58544921875
],
[
0.253173828125,
0.57373046875,
0.268798828125,
0.57373046875,
0.268798828125,
0.5771484375,
0.253173828125,
0.5771484375
],
[
0.160888671875,
0.58056640625,
0.1829833984375,
0.58056640625,
0.1829833984375,
0.5947265625,
0.160888671875,
0.5947265625
],
[
0.2900390625,
0.58203125,
0.31201171875,
0.58203125,
0.31201171875,
0.5947265625,
0.2900390625,
0.5947265625
],
[
0.330078125,
0.57763671875,
0.406982421875,
0.57275390625,
0.408447265625,
0.5869140625,
0.330810546875,
0.59228515625
],
[
0.423095703125,
0.5869140625,
0.438720703125,
0.583984375,
0.44287109375,
0.5947265625,
0.42724609375,
0.59765625
],
[
0.56396484375,
0.5888671875,
0.5810546875,
0.5888671875,
0.5810546875,
0.59765625,
0.56396484375,
0.59765625
],
[
0.69677734375,
0.583984375,
0.72119140625,
0.583984375,
0.72119140625,
0.5966796875,
0.69677734375,
0.5966796875
],
[
0.16259765625,
0.5966796875,
0.178955078125,
0.5966796875,
0.178955078125,
0.60595703125,
0.16259765625,
0.60595703125
],
[
0.293212890625,
0.5966796875,
0.310546875,
0.5966796875,
0.310546875,
0.60595703125,
0.293212890625,
0.60595703125
],
[
0.69677734375,
0.59765625,
0.7197265625,
0.59765625,
0.7197265625,
0.60986328125,
0.69677734375,
0.60986328125
],
[
0.73876953125,
0.60107421875,
0.80810546875,
0.5947265625,
0.8095703125,
0.60595703125,
0.740234375,
0.6123046875
],
[
0.564453125,
0.60302734375,
0.58251953125,
0.60302734375,
0.58251953125,
0.60986328125,
0.564453125,
0.60986328125
],
[
0.2900390625,
0.60595703125,
0.313720703125,
0.60595703125,
0.313720703125,
0.61865234375,
0.2900390625,
0.61865234375
],
[
0.42724609375,
0.60595703125,
0.441162109375,
0.60595703125,
0.441162109375,
0.6162109375,
0.42724609375,
0.6162109375
],
[
0.69775390625,
0.6123046875,
0.7138671875,
0.6123046875,
0.7138671875,
0.6220703125,
0.69775390625,
0.6220703125
],
[
0.56396484375,
0.6142578125,
0.58837890625,
0.6142578125,
0.58837890625,
0.62255859375,
0.56396484375,
0.62255859375
],
[
0.236083984375,
0.6171875,
0.26953125,
0.6171875,
0.26953125,
0.626953125,
0.236083984375,
0.626953125
],
[
0.236083984375,
0.60693359375,
0.26953125,
0.60693359375,
0.26953125,
0.6162109375,
0.236083984375,
0.6162109375
],
[
0.298095703125,
0.62353515625,
0.3095703125,
0.62060546875,
0.31201171875,
0.625,
0.2998046875,
0.626953125
],
[
0.223876953125,
0.63330078125,
0.238525390625,
0.63330078125,
0.238525390625,
0.642578125,
0.223876953125,
0.642578125
],
[
0.26708984375,
0.63330078125,
0.279296875,
0.63330078125,
0.279296875,
0.6416015625,
0.26708984375,
0.6416015625
],
[
0.35546875,
0.63330078125,
0.371826171875,
0.63330078125,
0.371826171875,
0.642578125,
0.35546875,
0.642578125
],
[
0.400390625,
0.63330078125,
0.4150390625,
0.63330078125,
0.4150390625,
0.642578125,
0.400390625,
0.642578125
],
[
0.435546875,
0.62744140625,
0.4453125,
0.62744140625,
0.4453125,
0.630859375,
0.435546875,
0.630859375
],
[
0.48779296875,
0.63330078125,
0.50244140625,
0.63330078125,
0.50244140625,
0.642578125,
0.48779296875,
0.642578125
],
[
0.62841796875,
0.63330078125,
0.64453125,
0.63330078125,
0.64453125,
0.642578125,
0.62841796875,
0.642578125
],
[
0.67236328125,
0.6337890625,
0.68115234375,
0.6337890625,
0.68115234375,
0.6416015625,
0.67236328125,
0.6416015625
],
[
0.7607421875,
0.63330078125,
0.775390625,
0.63330078125,
0.775390625,
0.6416015625,
0.7607421875,
0.6416015625
],
[
0.80224609375,
0.6337890625,
0.81787109375,
0.6337890625,
0.81787109375,
0.6416015625,
0.80224609375,
0.6416015625
],
[
0.53173828125,
0.63525390625,
0.5458984375,
0.630859375,
0.5498046875,
0.64013671875,
0.5361328125,
0.64453125
],
[
0.4453125,
0.63623046875,
0.451904296875,
0.63623046875,
0.451904296875,
0.64013671875,
0.4453125,
0.64013671875
],
[
0.426513671875,
0.68505859375,
0.44677734375,
0.68505859375,
0.44677734375,
0.693359375,
0.426513671875,
0.693359375
],
[
0.428955078125,
0.6962890625,
0.450927734375,
0.6962890625,
0.450927734375,
0.708984375,
0.428955078125,
0.708984375
],
[
0.42724609375,
0.708984375,
0.44775390625,
0.70556640625,
0.451904296875,
0.7197265625,
0.4306640625,
0.72265625
],
[
0.42724609375,
0.72265625,
0.44775390625,
0.7197265625,
0.451904296875,
0.73291015625,
0.4306640625,
0.736328125
],
[
0.428955078125,
0.73486328125,
0.449462890625,
0.73486328125,
0.449462890625,
0.74755859375,
0.428955078125,
0.74755859375
],
[
0.08905029296875,
0.79345703125,
0.8837890625,
0.79345703125,
0.8837890625,
0.80810546875,
0.08905029296875,
0.80810546875
],
[
0.08905029296875,
0.80810546875,
0.88623046875,
0.80810546875,
0.88623046875,
0.8212890625,
0.08905029296875,
0.8212890625
],
[
0.0882568359375,
0.82275390625,
0.416748046875,
0.82275390625,
0.416748046875,
0.8359375,
0.0882568359375,
0.8359375
],
[
0.08905029296875,
0.84228515625,
0.4755859375,
0.84130859375,
0.4755859375,
0.85595703125,
0.08905029296875,
0.85693359375
],
[
0.50244140625,
0.84228515625,
0.8857421875,
0.84228515625,
0.8857421875,
0.85693359375,
0.50244140625,
0.85693359375
],
[
0.0906982421875,
0.85986328125,
0.47216796875,
0.85986328125,
0.47216796875,
0.87060546875,
0.0906982421875,
0.87060546875
],
[
0.50341796875,
0.85791015625,
0.8837890625,
0.85791015625,
0.8837890625,
0.8720703125,
0.50341796875,
0.8720703125
],
[
0.08905029296875,
0.87353515625,
0.473876953125,
0.87353515625,
0.473876953125,
0.88720703125,
0.08905029296875,
0.88720703125
],
[
0.5009765625,
0.87353515625,
0.8837890625,
0.87353515625,
0.8837890625,
0.88720703125,
0.5009765625,
0.88720703125
],
[
0.08905029296875,
0.88818359375,
0.4755859375,
0.88818359375,
0.4755859375,
0.9013671875,
0.08905029296875,
0.9013671875
],
[
0.50244140625,
0.88818359375,
0.8857421875,
0.88818359375,
0.8857421875,
0.90283203125,
0.50244140625,
0.90283203125
]
] |
[
"Q*Bert",
"Ms.Pacman",
"Breakout",
"CrazyClimber",
"Asterix",
"3500",
"12000",
"25000",
"12500",
"3000",
"12500",
"10000",
"10000",
"300",
"10000",
"",
"2500",
"",
"B000",
"7500",
"2000",
"7500",
"M",
"",
"",
"60000",
"1500",
"5000:",
"t00",
"500",
"S-DQN",
"1000",
"2500",
"DON",
"500",
"20000",
"0.",
"100",
"200",
"100",
"200",
"100",
"0",
"100",
"nr",
"Seaquest",
"10000",
"8000",
"",
"6000",
"4000",
"2000",
"U",
"400",
"15000",
"2500",
"100000",
"12500",
"",
"12500",
"",
"10000",
"2000",
"300",
"BOOO",
"10000",
"200",
"7500",
"1500",
"7500",
"500",
"40000",
"5000",
"MAA",
"100",
"S-DDQN",
"250g",
"20000",
"250",
"500",
"0-",
"DDQN",
"200",
"100",
"100",
"100",
"100",
"200",
"100",
"0",
"200",
"",
"200",
"8000",
"6000",
"400",
"100",
"o",
"3500",
"12000",
"1500",
"12500",
"400",
"3000",
"100000",
"",
"12500",
"10000",
"2500",
"300",
"1000",
"B0OOO",
"",
"7500",
"2000",
"",
"200",
"6000",
"7500",
"5000",
"1500",
"5000:",
"A",
"4000",
"1000",
"100",
"250",
"20000",
"DDQN",
"S-DQN",
"",
"100",
"200",
"100",
"200",
"n-",
"100",
"100",
"200",
"100",
"200",
"200",
"0",
"10000",
"BO0",
"6000",
"4000",
"2000",
"Figure 1. Performance of (top row) S-DQN vs. DQN, (middle row) S-DDQN vs. DDQN, and (bottom row) S-DQN vs.",
"DDQN on the Atari games. X-axis and Y-axis correspond to the training epoch and test score, respectively. All curves",
"are averaged over five independent random seeds.",
"诺。对于游戏Asterix，通常用作EDQNE(vanHasselt",
"不像Asterix那样引人注目，但图E1的底行显示，对",
"etal.2016a：Anscheletal.,2017）中高估问题的示例",
"于整个测试套件，S-DQN通常优于DDQN，具有更高",
"S-DQN和S-DDQN的测试得分均远高于其最大对应",
"的测试分数，这表明在存在函数逼近的情况下，可以",
"物，这表明softmax运算符可以减轻高估偏差。尽管",
"使用softmax运算符而不是双Q学习E(vanHasselt,"
] |
{
"arxiv_id": "1812.00456",
"language": "zh",
"page_idx": 3
}
|
||
1812.00456-zh-4
|
1812.00456-zh-4.jpg
|
Table 1. Mean of test scores for different values of $\tau$ in S-DDQN (standard deviation in parenthesis). Each game is denoted with its initial. Note that S-DDQN reduces to $\underline{\mathrm{DDQN~when~}\tau=\infty~.}$
<table>
<tbody>
<tr>
<th>$\tau$</th>
<th>1</th>
<th>5</th>
<th>10</th>
<th>$\infty$</th>
</tr>
<tr>
<td>$Q$</td>
<td>12068.66 (1085.65)</td>
<td>11049.28 (1565.57)</td>
<td>111191.31 (1336.35)</td>
<td>10577.76 (1508.27</td>
</tr>
<tr>
<td>$M$</td>
<td>2492.40 (183.71)</td>
<td>2566.44 (227.24)</td>
<td>2546.18 (259.82)</td>
<td>2293.73 (160.50)</td>
</tr>
<tr>
<td>$B$</td>
<td>313.08 (20.13)</td>
<td>350.88 (35.58)</td>
<td>303.71 (65.59)</td>
<td>284.64 (60.83)</td>
</tr>
<tr>
<td>$C$</td>
<td>107405.11 (4617.90)</td>
<td>111111.07 (5047.19)</td>
<td>104049.46 (6686.84)</td>
<td>96373.08 (9244.27</td>
</tr>
<tr>
<td>A</td>
<td>3476.91 (460.27)</td>
<td>10266.12 (2682.00)</td>
<td>6588.13 (1183.10)</td>
<td>5523.80 (694.72)</td>
</tr>
<tr>
<td>$S$</td>
<td>272.20 (49.75)</td>
<td>2701.05 (10.06)</td>
<td>6254.01 (697.12)</td>
<td>5695.35 (1862.59</td>
</tr>
</tbody>
</table>
2010) 来减少高估偏差
表$\mathbb{E}$1 显示了具有不同 $\tau$ 值的 S-DDQN 的测试分数该分数是通过对最近 10 个时期的分数进行平均而获导的。请注意，对于$\tau$的大多数值，S-DDQN的测试分数高于其最大对应物。表$\mathbb{E}$1 进一步建议在选择$\tau$的值时进行权衡：将$\tau$设置得太小将使$\mathcal{T}_\mathrm{soft}$远离Bellman 最优性，因为最大运算符(对应于$\tau=\infty$ ) 在标准 Bellman 方程中使用。另一方面，使用较大的$\tau$将导致高估和高梯度噪声的问题，如第 5 节中所述
Table 2. Mean of test scores for different Bellman operators (standard deviation in parenthesis). The statistics are averaged over five independent random seeds.
<table>
<tbody>
<tr>
<th> </th>
<th>Max</th>
<th>Softmax</th>
<th>Mellowmax</th>
</tr>
<tr>
<td>$\mathrm{Q} ^{* }$Bert</td>
<td>8331.72 (1597.67)</td>
<td>11307.10 (1332.80)</td>
<td>11775.93 (1173.51)</td>
</tr>
<tr>
<td>Ms. I Pacman</td>
<td>2368.79 (219.17)</td>
<td>2856.82 (369.75)</td>
<td>2458.76 (130.34)</td>
</tr>
<tr>
<td>C. Climber</td>
<td>90923.40 (11059.39)</td>
<td>106422.27 (4821.40)</td>
<td>99601.47 (19271.53)</td>
</tr>
<tr>
<td>Breakout</td>
<td>255.32 (64.69)</td>
<td>345.56 (34.19)</td>
<td>355.94 (25.85)</td>
</tr>
<tr>
<td>Asterix</td>
<td>196.91 (135.16)</td>
<td>8868.00 (2167.35)</td>
<td>11203.75 (3818.40)</td>
</tr>
<tr>
<td>$\operatorname{Seaquest}$</td>
<td>4090.36 (1455.73)</td>
<td>8066.78 (1646.51)</td>
<td>6476.20 (1952.12)</td>
</tr>
</tbody>
</table>
为了与 mellowmax 运算符进行比较，我们以与 softmax 运算符$\mathbb{E}^{5}$ 相同的关闭策略方式将其与 DQN 组合。请注意，EAsadi & Littman (2017) 中 mellowmax 的寻根方法需要计算每个状态的最佳逆温度参数，因此在
${}^{5}$We tried the mellowmax operator for exploration as in Asadi & Littman (2017) , but observed much worse performance.
计算上太昂贵，无法在此处应用。因此，我们从$\{1、5、$ $10\}$调整 softmax 和 mellowmax 运算符的逆温度参数， 并报告最佳分数。表 2 表明，softmax 和 mellowmax 运算符都可以获得比 max 运算符更高的测试分数。此外，除 Asterix 外，所有游戏中 softmax 的得分都高于或类似于 mellowmax 的得分。softmax 算子的竞争性能表明，尽管它具有非收缩特性，但在某些领域仍然是可取的
# 5.Why Softmax Helps?
人们可能想知道为什么 softmaxBellman 算子可以在第$\mathbb{E}$4 节中实现令人惊讶的性能，因为来自 Bellman 方程的贪婪策略表明 max 算子应该是最优的。softmax 算子在本文中不用于探索，也不像熵正则化的 MDPs (Fox et al., 2016; Schulman et al., 2017; Neu et al., 2017; Nachum et al., 2017; Asadi & Littman, 2017; Lee et al., 2018) 那样受正则化策略的激励
如先前的工作$\overline{\mathrm{E}}($van Hasselt et al., 2016a; Anschel et al., 2017) 中所述，max 算子导致EDQNs 中的 Q 函数高估的重大问题，这进一步导致过多的梯度噪声破坏了神经网络的优化。在这里，我们旨在分析 softmaxBellman 算子如何克服这些问题，以及 DQNs 中高估和梯度降噪的相应证据。尽管我们的分析集中在softmaxBellman 算子上，但这项工作也可能进一步了解熵正则化的实际好处
## 5.1. Overestimation Bias Reduction
Q-learning 的高估偏差，由于最大运算符，首先在 Thrun & Schwartz (1994) 中讨论。后来在$\overline{\mathrm{Evan}}$ Hasselt et al. (2016a) 和 Anschel et al. (2017) 中显示， 高估会导致$\overline{\mathrm{E}}$DQNs 在某些 Atari 游戏中的性能不佳。遵循与 van Hasselt et al. (2016a) 相同的假设，我们可以证明 softmax 算子减少了高估偏差。此外，我们通过提供下限和上限来量化高估减少的范围
Theorem 4. Given the same assumptions as van Hasselt et al. (2016a) , where $(A1)$ there exists some $V^*(s)$ such that the true state-action value function satisfies $Q^*(s,a)=V^*(s)$ , for different actions. $(A2)$ the estimation error is modeled as $Q_t( s, a) = \dot{V} ^* ( s) + \epsilon _a$ , then
$(I)$ the overestimation errors from $\mathcal{T}_\mathrm{soft}$ are smaller or equal to those of $\mathcal{T}$ using the max operator, for any $\tau \geq 0$ ;
$( II)$ the overestimation reduction by using $\mathcal{T} _{\mathrm{soft}}$ in lieu of $\mathcal{T}$ is within $\left [ \frac {\widehat{\delta } ( s) }{m\exp [ \tau \widehat{\delta } ( s) ] }, ( m\:- \:\right ]$
$$1)\max\{\frac{1}{\tau+2},\frac{2Q_{\max}}{1+\exp(\tau)}\}\Big]\:;$$
$(III){\mathrm{~the~overestimation~error~for~}\mathcal{T}}_{\mathrm{soft~monotonically}}$
increases w.r.t. $\tau \in [ 0, \infty )$ .
关于定理$\boxed{\mathrm{E}}$4 的观察结果是，对于 $\tau$ 的任何正值，仍
|
[
[
0.50341796875,
0.08587646484375,
0.88232421875,
0.08587646484375,
0.88232421875,
0.0997314453125,
0.50341796875,
0.0997314453125
],
[
0.08905029296875,
0.0947265625,
0.473876953125,
0.0947265625,
0.473876953125,
0.10797119140625,
0.08905029296875,
0.10797119140625
],
[
0.5009765625,
0.0997314453125,
0.88232421875,
0.10101318359375,
0.88232421875,
0.11553955078125,
0.5009765625,
0.1142578125
],
[
0.08660888671875,
0.106689453125,
0.4755859375,
0.10797119140625,
0.4755859375,
0.12249755859375,
0.08660888671875,
0.121826171875
],
[
0.50244140625,
0.11614990234375,
0.8857421875,
0.11614990234375,
0.8857421875,
0.1300048828125,
0.50244140625,
0.1300048828125
],
[
0.0882568359375,
0.12249755859375,
0.473876953125,
0.12249755859375,
0.473876953125,
0.1363525390625,
0.0882568359375,
0.1363525390625
],
[
0.0882568359375,
0.1337890625,
0.2320556640625,
0.136962890625,
0.2320556640625,
0.1514892578125,
0.0882568359375,
0.1483154296875
],
[
0.50244140625,
0.1319580078125,
0.8857421875,
0.1319580078125,
0.8857421875,
0.1458740234375,
0.50244140625,
0.1458740234375
],
[
0.5009765625,
0.1458740234375,
0.8837890625,
0.146484375,
0.8837890625,
0.1610107421875,
0.5009765625,
0.160400390625
],
[
0.10296630859375,
0.154052734375,
0.11273193359375,
0.160400390625,
0.10296630859375,
0.1685791015625,
0.0931396484375,
0.1622314453125
],
[
0.148681640625,
0.154052734375,
0.1650390625,
0.154052734375,
0.1650390625,
0.1685791015625,
0.148681640625,
0.1685791015625
],
[
0.2418212890625,
0.1546630859375,
0.2607421875,
0.1546630859375,
0.2607421875,
0.1685791015625,
0.2418212890625,
0.1685791015625
],
[
0.336669921875,
0.1546630859375,
0.358642578125,
0.1546630859375,
0.358642578125,
0.1685791015625,
0.336669921875,
0.1685791015625
],
[
0.423095703125,
0.1546630859375,
0.4453125,
0.1546630859375,
0.4453125,
0.169189453125,
0.423095703125,
0.169189453125
],
[
0.5009765625,
0.160400390625,
0.8837890625,
0.1610107421875,
0.8837890625,
0.175537109375,
0.5009765625,
0.1749267578125
],
[
0.1217041015625,
0.1749267578125,
0.1953125,
0.1749267578125,
0.1953125,
0.189453125,
0.1217041015625,
0.189453125
],
[
0.2197265625,
0.1749267578125,
0.285888671875,
0.1749267578125,
0.285888671875,
0.189453125,
0.2197265625,
0.189453125
],
[
0.313720703125,
0.1749267578125,
0.378173828125,
0.1749267578125,
0.378173828125,
0.189453125,
0.313720703125,
0.189453125
],
[
0.40185546875,
0.1749267578125,
0.468017578125,
0.1749267578125,
0.468017578125,
0.189453125,
0.40185546875,
0.189453125
],
[
0.50244140625,
0.1767578125,
0.8857421875,
0.1767578125,
0.8857421875,
0.1912841796875,
0.50244140625,
0.1912841796875
],
[
0.098876953125,
0.184326171875,
0.1168212890625,
0.184326171875,
0.1168212890625,
0.1976318359375,
0.098876953125,
0.1976318359375
],
[
0.1217041015625,
0.1900634765625,
0.19287109375,
0.1900634765625,
0.19287109375,
0.2052001953125,
0.1217041015625,
0.2052001953125
],
[
0.2181396484375,
0.1900634765625,
0.285888671875,
0.1900634765625,
0.285888671875,
0.2052001953125,
0.2181396484375,
0.2052001953125
],
[
0.313720703125,
0.1900634765625,
0.381591796875,
0.1900634765625,
0.381591796875,
0.2052001953125,
0.313720703125,
0.2052001953125
],
[
0.400390625,
0.1900634765625,
0.4697265625,
0.1900634765625,
0.4697265625,
0.2052001953125,
0.400390625,
0.2052001953125
],
[
0.50244140625,
0.1912841796875,
0.56884765625,
0.1912841796875,
0.56884765625,
0.205810546875,
0.50244140625,
0.205810546875
],
[
0.1282958984375,
0.20703125,
0.1854248046875,
0.20703125,
0.1854248046875,
0.2183837890625,
0.1282958984375,
0.2183837890625
],
[
0.221435546875,
0.20703125,
0.28515625,
0.20703125,
0.28515625,
0.2183837890625,
0.221435546875,
0.2183837890625
],
[
0.31787109375,
0.2052001953125,
0.375732421875,
0.2052001953125,
0.375732421875,
0.2197265625,
0.31787109375,
0.2197265625
],
[
0.406005859375,
0.20703125,
0.464111328125,
0.20703125,
0.464111328125,
0.2183837890625,
0.406005859375,
0.2183837890625
],
[
0.098876953125,
0.21337890625,
0.11767578125,
0.21337890625,
0.11767578125,
0.2266845703125,
0.098876953125,
0.2266845703125
],
[
0.12744140625,
0.2203369140625,
0.1871337890625,
0.2203369140625,
0.1871337890625,
0.23486328125,
0.12744140625,
0.23486328125
],
[
0.221435546875,
0.2203369140625,
0.283447265625,
0.2203369140625,
0.283447265625,
0.23486328125,
0.221435546875,
0.23486328125
],
[
0.316162109375,
0.2203369140625,
0.37744140625,
0.2203369140625,
0.37744140625,
0.23486328125,
0.316162109375,
0.23486328125
],
[
0.404296875,
0.2203369140625,
0.465576171875,
0.2203369140625,
0.465576171875,
0.23486328125,
0.404296875,
0.23486328125
],
[
0.5,
0.2266845703125,
0.705078125,
0.2266845703125,
0.705078125,
0.2412109375,
0.5,
0.2412109375
],
[
0.13232421875,
0.236083984375,
0.181396484375,
0.236083984375,
0.181396484375,
0.250732421875,
0.13232421875,
0.250732421875
],
[
0.223876953125,
0.236083984375,
0.281005859375,
0.236083984375,
0.281005859375,
0.250732421875,
0.223876953125,
0.250732421875
],
[
0.32177734375,
0.236083984375,
0.370849609375,
0.236083984375,
0.370849609375,
0.250732421875,
0.32177734375,
0.250732421875
],
[
0.41015625,
0.236083984375,
0.4599609375,
0.236083984375,
0.4599609375,
0.250732421875,
0.41015625,
0.250732421875
],
[
0.096435546875,
0.24365234375,
0.11358642578125,
0.24365234375,
0.11358642578125,
0.25830078125,
0.096435546875,
0.25830078125
],
[
0.131591796875,
0.25,
0.181396484375,
0.251953125,
0.1805419921875,
0.267578125,
0.1298828125,
0.26513671875
],
[
0.226318359375,
0.250732421875,
0.27783203125,
0.250732421875,
0.27783203125,
0.266357421875,
0.226318359375,
0.266357421875
],
[
0.32177734375,
0.250732421875,
0.371826171875,
0.250732421875,
0.371826171875,
0.266357421875,
0.32177734375,
0.266357421875
],
[
0.408447265625,
0.251953125,
0.4599609375,
0.251953125,
0.4599609375,
0.266357421875,
0.408447265625,
0.266357421875
],
[
0.50341796875,
0.251953125,
0.8857421875,
0.251953125,
0.8857421875,
0.26513671875,
0.50341796875,
0.26513671875
],
[
0.3095703125,
0.266357421875,
0.384033203125,
0.268310546875,
0.384033203125,
0.28271484375,
0.3095703125,
0.281005859375
],
[
0.5009765625,
0.26513671875,
0.8857421875,
0.266357421875,
0.8857421875,
0.281005859375,
0.5009765625,
0.27978515625
],
[
0.400390625,
0.267578125,
0.468017578125,
0.267578125,
0.468017578125,
0.2822265625,
0.400390625,
0.2822265625
],
[
0.1217041015625,
0.267578125,
0.193603515625,
0.267578125,
0.193603515625,
0.2822265625,
0.1217041015625,
0.2822265625
],
[
0.2132568359375,
0.268310546875,
0.291748046875,
0.268310546875,
0.291748046875,
0.27978515625,
0.2132568359375,
0.27978515625
],
[
0.098876953125,
0.276611328125,
0.11358642578125,
0.276611328125,
0.11358642578125,
0.285888671875,
0.098876953125,
0.285888671875
],
[
0.12335205078125,
0.2822265625,
0.191162109375,
0.2822265625,
0.191162109375,
0.296630859375,
0.12335205078125,
0.296630859375
],
[
0.2181396484375,
0.2822265625,
0.285888671875,
0.2822265625,
0.285888671875,
0.296630859375,
0.2181396484375,
0.296630859375
],
[
0.398681640625,
0.281005859375,
0.468017578125,
0.281005859375,
0.468017578125,
0.29541015625,
0.398681640625,
0.29541015625
],
[
0.50244140625,
0.28271484375,
0.88623046875,
0.28271484375,
0.88623046875,
0.296630859375,
0.50244140625,
0.296630859375
],
[
0.12744140625,
0.29736328125,
0.1854248046875,
0.29736328125,
0.1854248046875,
0.3125,
0.12744140625,
0.3125
],
[
0.314453125,
0.2841796875,
0.3798828125,
0.2841796875,
0.3798828125,
0.29541015625,
0.314453125,
0.29541015625
],
[
0.2156982421875,
0.29736328125,
0.2900390625,
0.29736328125,
0.2900390625,
0.3125,
0.2156982421875,
0.3125
],
[
0.31787109375,
0.29736328125,
0.375732421875,
0.29736328125,
0.375732421875,
0.3125,
0.31787109375,
0.3125
],
[
0.404296875,
0.29736328125,
0.464111328125,
0.29736328125,
0.464111328125,
0.3125,
0.404296875,
0.3125
],
[
0.5009765625,
0.29736328125,
0.8857421875,
0.29736328125,
0.8857421875,
0.311279296875,
0.5009765625,
0.311279296875
],
[
0.09722900390625,
0.304931640625,
0.11358642578125,
0.304931640625,
0.11358642578125,
0.3193359375,
0.09722900390625,
0.3193359375
],
[
0.12744140625,
0.311279296875,
0.1871337890625,
0.313232421875,
0.1854248046875,
0.32763671875,
0.1258544921875,
0.32568359375
],
[
0.31201171875,
0.313232421875,
0.381591796875,
0.313232421875,
0.381591796875,
0.32763671875,
0.31201171875,
0.32763671875
],
[
0.40283203125,
0.313232421875,
0.465576171875,
0.313232421875,
0.465576171875,
0.32763671875,
0.40283203125,
0.32763671875
],
[
0.5009765625,
0.313232421875,
0.8857421875,
0.313232421875,
0.8857421875,
0.326904296875,
0.5009765625,
0.326904296875
],
[
0.5009765625,
0.32763671875,
0.88623046875,
0.32763671875,
0.88623046875,
0.341552734375,
0.5009765625,
0.341552734375
],
[
0.2197265625,
0.314453125,
0.28515625,
0.314453125,
0.28515625,
0.32568359375,
0.2197265625,
0.32568359375
],
[
0.13232421875,
0.328857421875,
0.1829833984375,
0.328857421875,
0.1829833984375,
0.343505859375,
0.13232421875,
0.343505859375
],
[
0.223876953125,
0.328857421875,
0.281005859375,
0.328857421875,
0.281005859375,
0.343505859375,
0.223876953125,
0.343505859375
],
[
0.314453125,
0.328857421875,
0.378173828125,
0.328857421875,
0.378173828125,
0.343505859375,
0.314453125,
0.343505859375
],
[
0.404296875,
0.328857421875,
0.464111328125,
0.328857421875,
0.464111328125,
0.343505859375,
0.404296875,
0.343505859375
],
[
0.09722900390625,
0.337158203125,
0.111083984375,
0.337158203125,
0.111083984375,
0.348388671875,
0.09722900390625,
0.348388671875
],
[
0.226318359375,
0.341552734375,
0.27783203125,
0.343505859375,
0.27685546875,
0.359130859375,
0.2254638671875,
0.356689453125
],
[
0.316162109375,
0.341552734375,
0.37744140625,
0.343505859375,
0.375732421875,
0.359130859375,
0.314453125,
0.356689453125
],
[
0.398681640625,
0.34228515625,
0.468017578125,
0.34228515625,
0.468017578125,
0.356689453125,
0.398681640625,
0.356689453125
],
[
0.5009765625,
0.343505859375,
0.80224609375,
0.343505859375,
0.80224609375,
0.356689453125,
0.5009765625,
0.356689453125
],
[
0.5009765625,
0.365478515625,
0.8857421875,
0.3642578125,
0.8857421875,
0.37890625,
0.5009765625,
0.380126953125
],
[
0.1298828125,
0.343505859375,
0.1829833984375,
0.343505859375,
0.1829833984375,
0.359130859375,
0.1298828125,
0.359130859375
],
[
0.08660888671875,
0.369384765625,
0.250732421875,
0.37060546875,
0.250732421875,
0.385009765625,
0.08660888671875,
0.3837890625
],
[
0.5,
0.380126953125,
0.88623046875,
0.380126953125,
0.88623046875,
0.39453125,
0.5,
0.39453125
],
[
0.0906982421875,
0.393310546875,
0.473876953125,
0.393310546875,
0.473876953125,
0.4072265625,
0.0906982421875,
0.4072265625
],
[
0.50244140625,
0.395751953125,
0.8857421875,
0.395751953125,
0.8857421875,
0.4091796875,
0.50244140625,
0.4091796875
],
[
0.0906982421875,
0.4091796875,
0.47216796875,
0.4091796875,
0.47216796875,
0.42236328125,
0.0906982421875,
0.42236328125
],
[
0.5009765625,
0.410400390625,
0.8857421875,
0.410400390625,
0.8857421875,
0.423583984375,
0.5009765625,
0.423583984375
],
[
0.0906982421875,
0.423583984375,
0.47216796875,
0.423583984375,
0.47216796875,
0.4375,
0.0906982421875,
0.4375
],
[
0.5009765625,
0.425537109375,
0.8837890625,
0.425537109375,
0.8837890625,
0.439453125,
0.5009765625,
0.439453125
],
[
0.08905029296875,
0.439453125,
0.473876953125,
0.439453125,
0.473876953125,
0.45263671875,
0.08905029296875,
0.45263671875
],
[
0.5,
0.44140625,
0.8837890625,
0.44140625,
0.8837890625,
0.455078125,
0.5,
0.455078125
],
[
0.0882568359375,
0.45263671875,
0.473876953125,
0.453857421875,
0.473876953125,
0.468505859375,
0.0882568359375,
0.46728515625
],
[
0.5009765625,
0.455810546875,
0.8837890625,
0.455810546875,
0.8837890625,
0.4697265625,
0.5009765625,
0.4697265625
],
[
0.0882568359375,
0.46728515625,
0.431396484375,
0.468505859375,
0.431396484375,
0.48291015625,
0.0882568359375,
0.481689453125
],
[
0.426513671875,
0.472900390625,
0.471435546875,
0.472900390625,
0.471435546875,
0.479736328125,
0.426513671875,
0.479736328125
],
[
0.5009765625,
0.4716796875,
0.66650390625,
0.4716796875,
0.66650390625,
0.48486328125,
0.5009765625,
0.48486328125
],
[
0.08905029296875,
0.484130859375,
0.47216796875,
0.484130859375,
0.47216796875,
0.49755859375,
0.08905029296875,
0.49755859375
],
[
0.08905029296875,
0.498779296875,
0.465576171875,
0.498779296875,
0.465576171875,
0.51220703125,
0.08905029296875,
0.51220703125
],
[
0.5,
0.50244140625,
0.7607421875,
0.50244140625,
0.7607421875,
0.5166015625,
0.5,
0.5166015625
],
[
0.0882568359375,
0.52978515625,
0.4755859375,
0.53076171875,
0.4755859375,
0.54541015625,
0.0882568359375,
0.54443359375
],
[
0.5009765625,
0.5263671875,
0.8857421875,
0.5263671875,
0.8857421875,
0.54052734375,
0.5009765625,
0.54052734375
],
[
0.0882568359375,
0.54541015625,
0.4755859375,
0.54541015625,
0.4755859375,
0.5595703125,
0.0882568359375,
0.5595703125
],
[
0.5009765625,
0.54150390625,
0.8857421875,
0.54150390625,
0.8857421875,
0.5546875,
0.5009765625,
0.5546875
],
[
0.08660888671875,
0.5595703125,
0.390625,
0.55810546875,
0.390625,
0.57275390625,
0.08660888671875,
0.57373046875
],
[
0.5,
0.5546875,
0.88232421875,
0.556640625,
0.88232421875,
0.57080078125,
0.5,
0.568359375
],
[
0.2279052734375,
0.57470703125,
0.265625,
0.5771484375,
0.2646484375,
0.59228515625,
0.226318359375,
0.59033203125
],
[
0.3095703125,
0.57470703125,
0.374267578125,
0.5771484375,
0.374267578125,
0.59130859375,
0.3095703125,
0.5888671875
],
[
0.398681640625,
0.57470703125,
0.48779296875,
0.5771484375,
0.48779296875,
0.59130859375,
0.398681640625,
0.5888671875
],
[
0.5009765625,
0.5712890625,
0.8857421875,
0.5712890625,
0.8857421875,
0.58544921875,
0.5009765625,
0.58544921875
],
[
0.5009765625,
0.58544921875,
0.8857421875,
0.5859375,
0.8857421875,
0.60107421875,
0.5009765625,
0.599609375
],
[
0.21728515625,
0.59765625,
0.275390625,
0.59765625,
0.275390625,
0.6123046875,
0.21728515625,
0.6123046875
],
[
0.31201171875,
0.599609375,
0.375732421875,
0.599609375,
0.375732421875,
0.611328125,
0.31201171875,
0.611328125
],
[
0.406982421875,
0.599609375,
0.48046875,
0.599609375,
0.48046875,
0.611328125,
0.406982421875,
0.611328125
],
[
0.09722900390625,
0.60595703125,
0.156005859375,
0.60595703125,
0.156005859375,
0.62060546875,
0.09722900390625,
0.62060546875
],
[
0.50244140625,
0.6015625,
0.8857421875,
0.6015625,
0.8857421875,
0.61572265625,
0.50244140625,
0.61572265625
],
[
0.2115478515625,
0.61279296875,
0.281005859375,
0.61279296875,
0.281005859375,
0.62744140625,
0.2115478515625,
0.62744140625
],
[
0.30810546875,
0.61279296875,
0.37744140625,
0.61279296875,
0.37744140625,
0.62744140625,
0.30810546875,
0.62744140625
],
[
0.408447265625,
0.61279296875,
0.47802734375,
0.61279296875,
0.47802734375,
0.62744140625,
0.408447265625,
0.62744140625
],
[
0.5009765625,
0.6171875,
0.81201171875,
0.6171875,
0.81201171875,
0.630859375,
0.5009765625,
0.630859375
],
[
0.21728515625,
0.62744140625,
0.275390625,
0.6298828125,
0.273681640625,
0.64453125,
0.2156982421875,
0.642578125
],
[
0.310546875,
0.62890625,
0.375732421875,
0.62890625,
0.375732421875,
0.64306640625,
0.310546875,
0.64306640625
],
[
0.414306640625,
0.62890625,
0.47216796875,
0.62890625,
0.47216796875,
0.64306640625,
0.414306640625,
0.64306640625
],
[
0.096435546875,
0.63525390625,
0.19287109375,
0.63720703125,
0.19287109375,
0.6513671875,
0.096435546875,
0.6494140625
],
[
0.21728515625,
0.6416015625,
0.27685546875,
0.64306640625,
0.275390625,
0.6591796875,
0.2156982421875,
0.6572265625
],
[
0.41259765625,
0.6416015625,
0.473876953125,
0.64306640625,
0.47216796875,
0.6591796875,
0.410888671875,
0.6572265625
],
[
0.5009765625,
0.6396484375,
0.8857421875,
0.6396484375,
0.8857421875,
0.65283203125,
0.5009765625,
0.65283203125
],
[
0.498291015625,
0.65283203125,
0.88623046875,
0.65380859375,
0.88623046875,
0.66845703125,
0.498291015625,
0.66748046875
],
[
0.313720703125,
0.642578125,
0.373291015625,
0.64453125,
0.371826171875,
0.66015625,
0.31201171875,
0.65771484375
],
[
0.2132568359375,
0.6591796875,
0.281005859375,
0.6591796875,
0.281005859375,
0.673828125,
0.2132568359375,
0.673828125
],
[
0.2998046875,
0.65771484375,
0.384033203125,
0.66015625,
0.384033203125,
0.6748046875,
0.2998046875,
0.67236328125
],
[
0.41015625,
0.6591796875,
0.476318359375,
0.6591796875,
0.476318359375,
0.673828125,
0.41015625,
0.673828125
],
[
0.096435546875,
0.66552734375,
0.1805419921875,
0.66748046875,
0.1805419921875,
0.681640625,
0.096435546875,
0.6796875
],
[
0.30810546875,
0.6748046875,
0.374267578125,
0.6748046875,
0.374267578125,
0.68603515625,
0.30810546875,
0.68603515625
],
[
0.406005859375,
0.6748046875,
0.48193359375,
0.6748046875,
0.48193359375,
0.689453125,
0.406005859375,
0.689453125
],
[
0.5,
0.66943359375,
0.88623046875,
0.66943359375,
0.88623046875,
0.68310546875,
0.5,
0.68310546875
],
[
0.2099609375,
0.67529296875,
0.281982421875,
0.67529296875,
0.281982421875,
0.6875,
0.2099609375,
0.6875
],
[
0.221435546875,
0.68994140625,
0.271240234375,
0.68994140625,
0.271240234375,
0.705078125,
0.221435546875,
0.705078125
],
[
0.31787109375,
0.68994140625,
0.369384765625,
0.68994140625,
0.369384765625,
0.705078125,
0.31787109375,
0.705078125
],
[
0.4150390625,
0.68994140625,
0.471435546875,
0.68994140625,
0.471435546875,
0.705078125,
0.4150390625,
0.705078125
],
[
0.5,
0.68505859375,
0.88623046875,
0.68505859375,
0.88623046875,
0.69873046875,
0.5,
0.69873046875
],
[
0.096435546875,
0.6962890625,
0.1683349609375,
0.69873046875,
0.1683349609375,
0.71337890625,
0.096435546875,
0.7109375
],
[
0.2197265625,
0.70263671875,
0.27294921875,
0.705078125,
0.271240234375,
0.72021484375,
0.2181396484375,
0.71826171875
],
[
0.498291015625,
0.69873046875,
0.8857421875,
0.69970703125,
0.8857421875,
0.7138671875,
0.498291015625,
0.71337890625
],
[
0.316162109375,
0.705078125,
0.370849609375,
0.705078125,
0.370849609375,
0.7197265625,
0.316162109375,
0.7197265625
],
[
0.4150390625,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.72021484375,
0.4150390625,
0.72021484375
],
[
0.221435546875,
0.71826171875,
0.26953125,
0.72021484375,
0.268798828125,
0.736328125,
0.2197265625,
0.734375
],
[
0.5,
0.71533203125,
0.53515625,
0.71533203125,
0.53515625,
0.72802734375,
0.5,
0.72802734375
],
[
0.314453125,
0.72265625,
0.371826171875,
0.72265625,
0.371826171875,
0.734375,
0.314453125,
0.734375
],
[
0.408447265625,
0.72265625,
0.48046875,
0.72265625,
0.48046875,
0.734375,
0.408447265625,
0.734375
],
[
0.2156982421875,
0.736328125,
0.27685546875,
0.736328125,
0.27685546875,
0.75048828125,
0.2156982421875,
0.75048828125
],
[
0.09722900390625,
0.72607421875,
0.156005859375,
0.72802734375,
0.1544189453125,
0.74365234375,
0.096435546875,
0.7412109375
],
[
0.3095703125,
0.736328125,
0.37744140625,
0.736328125,
0.37744140625,
0.75048828125,
0.3095703125,
0.75048828125
],
[
0.41015625,
0.736328125,
0.47802734375,
0.736328125,
0.47802734375,
0.75048828125,
0.41015625,
0.75048828125
],
[
0.5,
0.7373046875,
0.8857421875,
0.7373046875,
0.8857421875,
0.751953125,
0.5,
0.751953125
],
[
0.2181396484375,
0.751953125,
0.275390625,
0.751953125,
0.275390625,
0.7666015625,
0.2181396484375,
0.7666015625
],
[
0.31201171875,
0.75244140625,
0.374267578125,
0.75244140625,
0.374267578125,
0.7646484375,
0.31201171875,
0.7646484375
],
[
0.4150390625,
0.75244140625,
0.471435546875,
0.75244140625,
0.471435546875,
0.7646484375,
0.4150390625,
0.7646484375
],
[
0.5,
0.75244140625,
0.8857421875,
0.75244140625,
0.8857421875,
0.767578125,
0.5,
0.767578125
],
[
0.096435546875,
0.7587890625,
0.1641845703125,
0.7587890625,
0.1641845703125,
0.7734375,
0.096435546875,
0.7734375
],
[
0.2132568359375,
0.7666015625,
0.281005859375,
0.7666015625,
0.281005859375,
0.78076171875,
0.2132568359375,
0.78076171875
],
[
0.30810546875,
0.7666015625,
0.37744140625,
0.7666015625,
0.37744140625,
0.78076171875,
0.30810546875,
0.78076171875
],
[
0.41015625,
0.7666015625,
0.476318359375,
0.7666015625,
0.476318359375,
0.78076171875,
0.41015625,
0.78076171875
],
[
0.5009765625,
0.76953125,
0.55224609375,
0.76953125,
0.55224609375,
0.78076171875,
0.5009765625,
0.78076171875
],
[
0.8037109375,
0.79052734375,
0.82861328125,
0.79345703125,
0.826171875,
0.80615234375,
0.80126953125,
0.80322265625
],
[
0.8447265625,
0.79052734375,
0.88623046875,
0.79248046875,
0.8857421875,
0.8046875,
0.84326171875,
0.80322265625
],
[
0.5458984375,
0.79150390625,
0.5810546875,
0.79150390625,
0.5810546875,
0.80322265625,
0.5458984375,
0.80322265625
],
[
0.5009765625,
0.79150390625,
0.53369140625,
0.79150390625,
0.53369140625,
0.8037109375,
0.5009765625,
0.8037109375
],
[
0.5908203125,
0.79248046875,
0.701171875,
0.79248046875,
0.701171875,
0.80322265625,
0.5908203125,
0.80322265625
],
[
0.71826171875,
0.79248046875,
0.79150390625,
0.79248046875,
0.79150390625,
0.8037109375,
0.71826171875,
0.8037109375
],
[
0.580078125,
0.79541015625,
0.59228515625,
0.79541015625,
0.59228515625,
0.7998046875,
0.580078125,
0.7998046875
],
[
0.09234619140625,
0.79736328125,
0.473876953125,
0.79736328125,
0.473876953125,
0.81103515625,
0.09234619140625,
0.81103515625
],
[
0.53515625,
0.79345703125,
0.5498046875,
0.79345703125,
0.5498046875,
0.8017578125,
0.53515625,
0.8017578125
],
[
0.498291015625,
0.80810546875,
0.740234375,
0.80615234375,
0.740234375,
0.82373046875,
0.498291015625,
0.82568359375
],
[
0.779296875,
0.8046875,
0.80615234375,
0.8046875,
0.80615234375,
0.8173828125,
0.779296875,
0.8173828125
],
[
0.0906982421875,
0.8125,
0.473876953125,
0.8125,
0.473876953125,
0.8271484375,
0.0906982421875,
0.8271484375
],
[
0.8349609375,
0.81103515625,
0.87109375,
0.81103515625,
0.87109375,
0.82275390625,
0.8349609375,
0.82275390625
],
[
0.75244140625,
0.818359375,
0.8310546875,
0.818359375,
0.8310546875,
0.830078125,
0.75244140625,
0.830078125
],
[
0.0906982421875,
0.82763671875,
0.47216796875,
0.82763671875,
0.47216796875,
0.84130859375,
0.0906982421875,
0.84130859375
],
[
0.494384765625,
0.8271484375,
0.68115234375,
0.8271484375,
0.68115234375,
0.85498046875,
0.494384765625,
0.85498046875
],
[
0.5986328125,
0.82763671875,
0.63134765625,
0.82763671875,
0.63134765625,
0.8388671875,
0.5986328125,
0.8388671875
],
[
0.0906982421875,
0.84326171875,
0.473876953125,
0.84326171875,
0.473876953125,
0.85693359375,
0.0906982421875,
0.85693359375
],
[
0.498291015625,
0.85400390625,
0.8837890625,
0.85498046875,
0.8837890625,
0.869140625,
0.498291015625,
0.86865234375
],
[
0.10540771484375,
0.8642578125,
0.4755859375,
0.8662109375,
0.4755859375,
0.8818359375,
0.10540771484375,
0.8798828125
],
[
0.5,
0.8720703125,
0.6953125,
0.8720703125,
0.6953125,
0.88525390625,
0.5,
0.88525390625
],
[
0.0906982421875,
0.8818359375,
0.11517333984375,
0.8818359375,
0.11517333984375,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.1094970703125,
0.8798828125,
0.4755859375,
0.8798828125,
0.4755859375,
0.89306640625,
0.1094970703125,
0.89306640625
],
[
0.08905029296875,
0.89306640625,
0.176513671875,
0.89306640625,
0.176513671875,
0.90478515625,
0.08905029296875,
0.90478515625
],
[
0.50244140625,
0.8916015625,
0.8857421875,
0.8916015625,
0.8857421875,
0.90478515625,
0.50244140625,
0.90478515625
]
] |
[
"计算上太昂贵，无法在此处应用。因此，我们从「1、5",
"Table 1. Mean of test scores for different values of in",
"10)调整softmax和mellowmax运算符的逆温度参数",
"S-DDQN (standard deviation in parenthesis). Each game",
"并报告最佳分数。表2表明，softmax和mellowmax",
"is denoted with its initial. Note that S-DDQN reduces to",
"DDQN when T = 00",
"运算符都可以获得比max运算符更高的测试分数。此",
"外，除Asterix外，所有游戏中softmax的得分都高于",
"K",
"1",
"5",
"10",
"8",
"或类似于mellowmax的得分。softmax算子的竞争性",
"12068.66",
"11049.28",
"11191.31",
"10577.76",
"能表明，尽管它具有非收缩特性，但在某些领域仍然",
"Q",
"(1085.65)",
"(1565.57)",
"(1336.35)",
"(1508.27)",
"是可取的",
"2492.40",
"2566.44",
"2546.18",
"2293.73",
"M",
"(183.71)",
"(227.24)",
"(259.82)",
"(160.50)",
"5. Why Softmax Helps?",
"313.08",
"350.88",
"303.71",
"284.64",
"B",
"(20.13)",
"(35.58)",
"(65.59)",
"(60.83)",
"人们可能想知道为什么softmaxBellman算子可以在",
"104049.46",
"第E4节中实现令人惊诉的性能，因为来自Bellman",
"96373.08",
"107405.11",
"111111.07",
"C",
"(4617.90)",
"(5047.19)",
"(9244.27)",
"方程的贪婪策略表明max算子应该是最优的。softmax",
"3476.91",
"(6686.84)",
"10266.12",
"6588.13",
"5523.80",
"算子在本文中不用于探索，也不像正则化的MDPs",
"A",
"(460.27)",
"(1183.10)",
"(694.72)",
"(Fox et al., 2016; Schulman et al., 2017; Neu et al.,",
"2017; Nachum et al., 2017; Asadi & Littman, 2017;",
"(2682.00)",
"272.20",
"2701.05",
"6254.01",
"5695.35",
"s",
"(10.06)",
"(697.12)",
"(1862.59)",
"Leeetal.,2018）那样受正则化策略的激励",
"如先前的工作E(vanHasseltetal.，2016a；Anschel",
"(49.75)",
"2010）来减少高估偏差",
"etal，2017）中所述，max算子导致EDQNs中的Q",
"表E1显示了具有不同T值的S-DDQN的测试分数，",
"函数高估的重大问题，这进一步导致过多的梯度噪声",
"该分数是通过对最近10个时期的分数进行平均而获",
"破坏了神经网络的优化。在这里，我们旨在分析soft",
"得的。请注意，对于T的大多数值，S-DDQN的测",
"maxBellman算子如何克服这些问题，以及DQNs中",
"试分数高于其最大对应物。表E1进一步建议在选择",
"高估和梯度降噪的相应证据。尽管我们的分析集中在",
"T的值时进行权衡：将T设置得太小将使Tsoft远离",
"softmaxBellman算子上，但这项工作也可能进一步了",
"Bellman最优性，因为最大运算符（对应于T：",
"8",
"解摘正则化的实际好处",
"在标准Bellman方程中使用。另一方面，使用较大的",
"T将导致高估和高梯度噪声的问题，如第5节中所述",
"5.1.Overestimation Bias Reduction",
"Table 2. Mean of test scores for different Bellman opera-",
"Q-learning的高估偏差，由于最大运算符，首先",
"tors (standard deviation in parenthesis). The statistics are",
"在Thrun&Schwartz（1994）中讨论。后来在Evan",
"averaged over five independent random seeds.",
"Hasseltetal.（2016a）和Anscheletal.（2017）中显示",
"Max",
"Softmax",
"Mellowmax",
"高估会导致EDQNs在某些Atari游戏中的性能不佳。",
"遵循与vanHasseltetal.（2016a）相同的假设，我们",
"8331.72",
"11307.10",
"11775.93",
"Q*Bert",
"可以证明softmax算子减少了高估偏差。此外，我们",
"(1597.67)",
"(1332.80)",
"(1173.51)",
"通过提供下限和上限来量化高估减少的范围",
"2368.79",
"2856.82",
"2458.76",
"Ms. Pacman",
"(219.17)",
"(130.34)",
"Theorem 4. Given the same assumptions as van Has",
"selt et al. (2016a) , where (A1) there exists some V*(s)",
"(369.75)",
"90923.40",
"106422.27",
"99601.47",
"C. Climber",
"(4821.40)",
"(19271.53)",
"such that the true state-action value function satisfies",
"(11059.39)",
"255.32",
"345.56",
"355.94",
"Q*(s, a) = V*(s) , for different actions. (A2) the es-",
"Breakout",
"(64.69)",
"timation error is modeled as Qt(s,a) = V*(s) + Ea",
"(34.19)",
"(25.85)",
"196.91",
"then",
"8868.00",
"11203.75",
"(135.16)",
"Asterix",
"(2167.35)",
"(3818.40)",
"(1) the overestimation errors from Tsoft are smaller or",
"4090.36",
"8066.78",
"6476.20",
"equal to those of T using the max operator, for any",
"Seaquest",
"(1455.73)",
"(1646.51)",
"(1952.12)",
"T≥0:",
"by",
"using",
"the",
"(I)",
"overestimation",
"reduction",
"",
"为了与mellowmax运算符进行比较，我们以与softmax",
"",
"Tsoft in lieuofTiswithin",
"8(s)",
"运算符E5相同的关闭策略方式将其与DQN组合。请",
"(m",
"mexp[-(s)]",
"注意，EAsadi&Littman（2017）中mellowmax的寻",
"1) max(2, 12m(] ;",
"2Qma",
"根方法需要计算每个状态的最佳逆温度参数，因此在",
"(I1) the overestimation error for Tsoft monotonically",
"5we tried the mellowmax operator for exploration as",
"increases w.r.t. T E[0,oo)",
"in",
"Asadi & Littman (2017),but observed much worse",
"performance.",
"关于定理E4的观察结果是，对于T的任何正值，仍"
] |
{
"arxiv_id": "1812.00456",
"language": "zh",
"page_idx": 4
}
|
||
2006.08383-zh-0
|
2006.08383-zh-0.jpg
|
像素不可见性：检测彩色图像中不可见的对象
Duminda Wijesekera George Mason University dwijesek@gmu. edu
Yongxin Wang
George Mason University
ywang51@gmu.edu
# Abstract
Despite recent success of object detectors using deep neural networks, their deployment on safety-critical applications such as self-driving cars remains questionable. This is partly due to the absence of reliable estimation for detectors" failure under operational conditions such as night, fog, dusk, dawn and glare. Such unquantifiable failures could lead to safety violations. In order to solve this problem, we created an algorithm that predicts a pixel-level invisibility map for color images that does not require manual labeling- that computes the probability that a pixel/region contains objects that are invisible in color domain, during various lighting conditions such as day, night and fog. We propose a novel use of cross modal knowledge distillation from color to infra-red domain using weaklyaligned image pairs from the day and construct indicators for the pixel-level invisibility based on the distances of their intermediate-level features. Quantitative experiments show the great performance of our pixel-level invisibility mask and also the effectiveness of distilled mid-level features on object detection in infra-red $imagery.$
### 1 Introduction
自从大规模数据集[?,2,2,2,2,?]和强大的基线系统 (如两级检测器 Fast/Faster/MaskR-CNN [?, 2,?])和一级检测器(如 XMATHLO MAXXSSD)出现以来，对象检测。在自动驾驶汽车等安全关键应用中部署探测器的一个重要问题是需要高置信度，以确保可导航区域在运行天气和照明条件下没有障碍物。未能检测到物体(假阴性)或提供警告信号，例如当物体检测器未注意到过马路的行人或停放的车辆时，可能会带来灾难性的后果。虽然对象检测器的性能正在提高，但不能保证它们永远不会犯错误[?]。因此，除了提供高检测精度外，可靠的视觉系统还应考虑“知道何时不知道”。我们的工作通过预测彩色图像的所谓像素级不可见性图而无需手动标记来解决此问题。配备了这种不可见性地图，系统可以决定在图像或信号警告消息中信任某些区域的检测结果，并将其移交给人类。
我们将一个图像的不可见性掩码定义为区域或像素包含在该域中不可见的对象的可能性。也就是说，一个像素或区域导致物体检测器中假阴性的可能性。在良好的日光下，彩色图像的区域获得较低的不可见性分数，因为足够能量的可见光被现场的物体反射到相机。尽管夜间图像的黑暗区域或雾中图像的模糊区域将具有较高的不可见性分数。一种简单的方法是创建一个大的标记数据集，其中图像中的每个像素都标有不可见性分数，这通常是昂贵且模糊的。相反，我们的方法通过提出一种新颖的跨模态知识蒸馏方法以及在彩色图像和红外图像之间生成对齐良好的图像对，来预测图像中每个像素的不可见性得分，而无需费力的标记工作
跨模态知识提炼[?]也称为监督转移，作为避免为某些模态标记大规模数据集的一种手段。给定来自两种模态的成对图像，中级表示从注释丰富的颜色域转移到具有有限标记数据集的其他模态，例如深度图像$\mathbb{E}[?]$,声音$\mathbb{E}[?]$和红外图像。我们的隐形面具的新颖之处在于利用从白天的彩色图像到红外图像的监督转移，然后使用两种模态的中级表示之间的距离来近似在各种照明条件下两种模态的隐形之间的感知距离，包括黄昏，黎明，黑夜和雾
Preprint. Under review.
|
[
[
0.221435546875,
0.1287841796875,
0.77783203125,
0.1287841796875,
0.77783203125,
0.146484375,
0.221435546875,
0.146484375
],
[
0.3095703125,
0.20263671875,
0.4150390625,
0.203857421875,
0.4150390625,
0.2183837890625,
0.3095703125,
0.2171630859375
],
[
0.564453125,
0.203857421875,
0.71337890625,
0.203857421875,
0.71337890625,
0.2171630859375,
0.564453125,
0.2171630859375
],
[
0.273681640625,
0.216552734375,
0.443603515625,
0.2171630859375,
0.443603515625,
0.231689453125,
0.273681640625,
0.2310791015625
],
[
0.5537109375,
0.2171630859375,
0.72216796875,
0.2171630859375,
0.72216796875,
0.2310791015625,
0.5537109375,
0.2310791015625
],
[
0.293212890625,
0.231689453125,
0.426513671875,
0.2298583984375,
0.426513671875,
0.2442626953125,
0.293212890625,
0.246826171875
],
[
0.56884765625,
0.231689453125,
0.708984375,
0.231689453125,
0.708984375,
0.24560546875,
0.56884765625,
0.24560546875
],
[
0.461669921875,
0.281005859375,
0.53759765625,
0.28271484375,
0.53759765625,
0.29736328125,
0.461669921875,
0.29541015625
],
[
0.2320556640625,
0.313232421875,
0.765625,
0.313232421875,
0.765625,
0.326904296875,
0.2320556640625,
0.326904296875
],
[
0.2344970703125,
0.32763671875,
0.765625,
0.328857421875,
0.765625,
0.34228515625,
0.2344970703125,
0.341552734375
],
[
0.233642578125,
0.34228515625,
0.7646484375,
0.34228515625,
0.7646484375,
0.35595703125,
0.233642578125,
0.35595703125
],
[
0.233642578125,
0.35595703125,
0.7646484375,
0.35595703125,
0.7646484375,
0.369384765625,
0.233642578125,
0.369384765625
],
[
0.2344970703125,
0.371826171875,
0.76318359375,
0.371826171875,
0.76318359375,
0.3818359375,
0.2344970703125,
0.3818359375
],
[
0.233642578125,
0.3837890625,
0.765625,
0.3837890625,
0.765625,
0.397705078125,
0.233642578125,
0.397705078125
],
[
0.233642578125,
0.397705078125,
0.76708984375,
0.397705078125,
0.76708984375,
0.410888671875,
0.233642578125,
0.410888671875
],
[
0.2320556640625,
0.4091796875,
0.765625,
0.410888671875,
0.765625,
0.425537109375,
0.2320556640625,
0.423583984375
],
[
0.233642578125,
0.4248046875,
0.765625,
0.423583984375,
0.765625,
0.438232421875,
0.233642578125,
0.439453125
],
[
0.2320556640625,
0.438232421875,
0.7646484375,
0.438232421875,
0.7646484375,
0.451904296875,
0.2320556640625,
0.451904296875
],
[
0.233642578125,
0.45068359375,
0.76708984375,
0.449462890625,
0.76708984375,
0.46728515625,
0.233642578125,
0.468505859375
],
[
0.236083984375,
0.468505859375,
0.7646484375,
0.468505859375,
0.7646484375,
0.479248046875,
0.236083984375,
0.479248046875
],
[
0.233642578125,
0.479736328125,
0.765625,
0.479736328125,
0.765625,
0.49365234375,
0.233642578125,
0.49365234375
],
[
0.2344970703125,
0.49365234375,
0.765625,
0.49365234375,
0.765625,
0.5068359375,
0.2344970703125,
0.5068359375
],
[
0.233642578125,
0.50830078125,
0.2900390625,
0.51025390625,
0.2900390625,
0.521484375,
0.233642578125,
0.51953125
],
[
0.17236328125,
0.5498046875,
0.31787109375,
0.5517578125,
0.31787109375,
0.56640625,
0.17236328125,
0.564453125
],
[
0.17724609375,
0.58349609375,
0.82275390625,
0.58349609375,
0.82275390625,
0.5966796875,
0.17724609375,
0.5966796875
],
[
0.1732177734375,
0.59521484375,
0.82275390625,
0.5966796875,
0.82275390625,
0.611328125,
0.1732177734375,
0.60986328125
],
[
0.1748046875,
0.609375,
0.82275390625,
0.60986328125,
0.82275390625,
0.625,
0.1748046875,
0.62353515625
],
[
0.17724609375,
0.625,
0.82275390625,
0.625,
0.82275390625,
0.63818359375,
0.17724609375,
0.63818359375
],
[
0.176513671875,
0.63818359375,
0.82275390625,
0.63818359375,
0.82275390625,
0.6513671875,
0.176513671875,
0.6513671875
],
[
0.17724609375,
0.65380859375,
0.82177734375,
0.65380859375,
0.82177734375,
0.6640625,
0.17724609375,
0.6640625
],
[
0.1748046875,
0.66552734375,
0.82177734375,
0.66552734375,
0.82177734375,
0.6787109375,
0.1748046875,
0.6787109375
],
[
0.176513671875,
0.6796875,
0.82275390625,
0.6796875,
0.82275390625,
0.693359375,
0.176513671875,
0.693359375
],
[
0.176513671875,
0.693359375,
0.640625,
0.693359375,
0.640625,
0.70703125,
0.176513671875,
0.70703125
],
[
0.17724609375,
0.71533203125,
0.8203125,
0.71533203125,
0.8203125,
0.72607421875,
0.17724609375,
0.72607421875
],
[
0.17724609375,
0.72998046875,
0.8203125,
0.72998046875,
0.8203125,
0.74072265625,
0.17724609375,
0.74072265625
],
[
0.17724609375,
0.74365234375,
0.8203125,
0.74365234375,
0.8203125,
0.75390625,
0.17724609375,
0.75390625
],
[
0.176513671875,
0.755859375,
0.82275390625,
0.755859375,
0.82275390625,
0.76953125,
0.176513671875,
0.76953125
],
[
0.1748046875,
0.767578125,
0.82275390625,
0.76806640625,
0.82275390625,
0.78271484375,
0.1748046875,
0.7822265625
],
[
0.17724609375,
0.78271484375,
0.82275390625,
0.78271484375,
0.82275390625,
0.796875,
0.17724609375,
0.796875
],
[
0.1732177734375,
0.79541015625,
0.82275390625,
0.796875,
0.82275390625,
0.81103515625,
0.1732177734375,
0.81005859375
],
[
0.17724609375,
0.8125,
0.27294921875,
0.8125,
0.27294921875,
0.82275390625,
0.17724609375,
0.82275390625
],
[
0.176513671875,
0.83349609375,
0.8203125,
0.83349609375,
0.8203125,
0.84326171875,
0.176513671875,
0.84326171875
],
[
0.17724609375,
0.8466796875,
0.8203125,
0.8466796875,
0.8203125,
0.85693359375,
0.17724609375,
0.85693359375
],
[
0.1748046875,
0.859375,
0.8203125,
0.859375,
0.8203125,
0.87255859375,
0.1748046875,
0.87255859375
],
[
0.176513671875,
0.87255859375,
0.82275390625,
0.87255859375,
0.82275390625,
0.88623046875,
0.176513671875,
0.88623046875
],
[
0.17724609375,
0.88818359375,
0.7939453125,
0.88818359375,
0.7939453125,
0.8984375,
0.17724609375,
0.8984375
],
[
0.176513671875,
0.92529296875,
0.31201171875,
0.92529296875,
0.31201171875,
0.93603515625,
0.176513671875,
0.93603515625
]
] |
[
"像素不可见性：检测彩色图像中不可见的对象",
"Yongxin Wang",
"Duminda Wijesekera",
"GeorgeMasonUniversity",
"GeorgeMason University",
"ywang51@gmu.edu",
"dwijesek@gmu.edu",
"Abstract",
"Despite recent success of object detectors using deep neural networks, their de",
"ployment on safety-critical applications such as self-driving cars remains ques-",
"tionable. This is partly due to the absence of reliable estimation for detectors'",
"failure under operational conditions such as night, fog, dusk, dawn and glare.",
"Suchunquantifiablefailurescouldleadtosafetyviolations.Inordertosolvethis",
"problem, we created an algorithm that predicts a pixel-level invisibility map for",
"color images that does not require manual labeling - that computes the probabil-",
"ity that a pixel/region contains objects that are invisible in color domain, during",
"various lighting conditions such as day, night and fog. We propose a novel use of",
"cross modal knowledge distillation from color to infra-red domain using weakly",
"aligned image pairs from the day and construct indicators for the pixei-level in-",
"visibilitybasedonthedistancesoftheirintermediate-levelfeatures.Quantitative",
"experiments show the great performance of our pixel-level invisibility mask and",
"also the effectiveness of distilled mid-level features on object detection in infra-red",
"imagery",
"1Introduction",
"自从大规模数据集[?,?,?,?,?]和强大的基线系统（如两级检测器Fast/Faster/MaskR-CNN[?,",
"2.?I）和一级检测器（如XMATHLOMAXXSSD）出现以来，对象检测。在自动驾驶汽车",
"等安全关键应用中部署探测器的一个重要问题是需要高置信度，以确保可导航区域在运行",
"天气和照明条件下没有障碍物。未能检测到物体（假阴性）或提供警告信号，例如当物体检",
"测器未注意到过马路的行人或停放的车辆时，可能会带来灾难性的后果。虽然对象检测器",
"的性能正在提高，但不能保证它们永远不会犯错误？！。因此，除了提供高检测精度外，可",
"靠的视觉系统还应考虑“知道何时不知道”。我们的工作通过预测彩色图像的所谓像素级不",
"可见性图而无需手动标记来解决此问题。配备了这种不可见性地图，系统可以决定在图像",
"或信号警告消息中信任某些区域的检测结果，并将其移交给人类。",
"我们将一个图像的不可见性掩码定义为区域或像素包含在该域中不可见的对象的可能性",
"也就是说，一个像素或区域导致物体检测器中假阴性的可能性。在良好的日光下，彩色图像",
"的区域获得较低的不可见性分数，因为足够能量的可见光被现场的物体反射到相机。尽管",
"夜间图像的黑暗区域或雾中图像的模糊区域将具有较高的不可见性分数。一种简单的方法",
"是创建一个大的标记数据集，其中图像中的每个像素都标有不可见性分数，这通常是昂贵",
"且模糊的。相反，我们的方法通过提出一种新颖的跨模态知识蒸馏方法以及在彩色图像和",
"红外图像之间生成对齐良好的图像对，来预测图像中每个像素的不可见性得分，而无需费",
"力的标记工作",
"跨模态知识提炼？也称为监督转移，作为避免为某些模态标记大规模数据集的一种手段",
"给定来自两种模态的成对图像，中级表示从注释丰富的颜色域转移到具有有限标记数据集",
"的其他模态，例如深度图像E[?]，声音E[?]和红外图像。我们的隐形面具的新颖之处在于",
"利用从白天的彩色图像到红外图像的监督转移，然后使用两种模态的中级表示之间的距离",
"来近似在各种照明条件下两种模态的隐形之间的感知距离，包括黄昏，黎明，黑夜和雾",
"Preprint.Underreview"
] |
{
"arxiv_id": "2006.08383",
"language": "zh",
"page_idx": 0
}
|
||
2006.08383-zh-1
|
2006.08383-zh-1.jpg
|
<FigureHere>
Figure 1: Our method predicts an invisibility map for each pixel (2nd row) given only the color image (1st row). Left to right: It can handle distant objects, highly occluded and cluttered objects, objects with strong glare, multiple objects from complex scenes like working zones on highway and objects in the fog. Color detectors can only detect objects (bounding boxes in 1st row) that are visible to its spectrum, while there are many more objects in the scene (3rd row). This could cause disastrous consequences where safety is crucial if such missed detection is trusted. Our pixel-level invisibility map indicates how much the detection results from color images may be trusted, both for false negatives and false positives.
知识提炼[?]特别要求这两种模态以成对的方式呈现，特别是以良好对齐的方式用于对象检测以及我们的像素级不可见性遮罩。这里，对齐良好的图像对是配对图像中的相应像素位于其各自图像平面中的相同位置的图像对。由彩色和红外传感器捕获的原始图像对具有较大的位移$\overline{\mathbb{E}}[?,?]$ ,这些位移来自 (1)内部相机属性差异，例如焦距，分辨率和镜头失真，(2) 外部转换，例如姿势差异和(3)曝光时间和快门速度的时间差。我们使用已知的模式板通过图像配准来解决前两个视差，同时我们提出了对齐 G 能量对抗 N 网络 ( AlignGAN) 来减轻剩余的位移
我们工作的贡献是三倍。(1)据我们所知，这是无需手动标记即可为彩色图像生成像素级隐形遮罩的第一项工作，因此有助于解决自动驾驶中由感官假阴性引起的故障检测问题；(2) 从彩色图像到红外图像的中级表示的直接转移在红外域中具有很高的检测精度；(3)减轻彩色红外对中存在的未对准问题 ( AlignGAN)。进行了广泛的实验，以定量评估我们系统的性能
### 2 $\textbf{Related Work}$
我们的像素级不可见掩模与不确定性估计$\mathbb{E}[?,?,2,?,?]$ ,故障检测$\mathbb{E}[2,2,2,2,?,2]$ 和分布外检测$\overline{\mathbb{E}}[2,2,2,?]$ 的任务紧密相连。大多数报告的工作要么将原始传感器数据视为内省检查$\mathbb{E}[?]$, 要么估计基于模型的分类器的不确定性，要么计算置信度得分或二元决策。我们的系统与它们的不同之处在于：(1)从局外人的角度估算传感器本身的置信度，即红外相机；(2)预测图像中每个像素的置信概率
像素对象$\mathbb{E}[?]$ 是计算所有类似对象区域的像素级掩码的第一项工作。[?] 提出了一种交通标志故障检测系统，该系统从目标检测器的特征图中提取激发区域，以缩小手动标记空间和假阴性搜索空间。尽管他们仍然需要将兴奋区域标记为假阴性或真阴性。我们的工作还预测了彩色图像中潜在不可见物体(假阴性)的所有区域的像素级掩模。相比之下，我们的方法不需要对不可见区域进行任何手动标记，而是利用彩色红外图像对来提供这种监督。为了获得对齐的彩色红外图像对$,\operatorname{E[P]}$基于分束器创建了 KAIST 多光谱行人数据集，以将彩色和红外相机的光束分为两部分。尽管观察到数据集中的图像对具有不同的位移$\mathbb{E}[?,?]$,并且由于分束器的强度衰减，在夜间受到的影响更大。[?]报告了不太严重的视差问题，尽管它们的数据集尚未发布。我们在使用垂直对齐相机的设置收集的数据集中观察到相同的问题， 并提出 AlignGAN 来减轻图像平面中的此类位移。Pix2pix[?]和 CycleGan[?]开发了配对和非配对设置下的跨域翻译方法。循环一致性[?]是解决不成对跨域翻译的主要技术。虽然
2
|
[
[
0.34228515625,
0.08837890625,
0.4150390625,
0.0889892578125,
0.4150390625,
0.10101318359375,
0.34228515625,
0.0997314453125
],
[
0.69384765625,
0.1275634765625,
0.81640625,
0.1275634765625,
0.81640625,
0.1383056640625,
0.69384765625,
0.1383056640625
],
[
0.6953125,
0.1363525390625,
0.82177734375,
0.1363525390625,
0.82177734375,
0.146484375,
0.6953125,
0.146484375
],
[
0.314453125,
0.1572265625,
0.438720703125,
0.1572265625,
0.438720703125,
0.1673583984375,
0.314453125,
0.1673583984375
],
[
0.69140625,
0.1875,
0.7685546875,
0.1881103515625,
0.7685546875,
0.199462890625,
0.69140625,
0.1988525390625
],
[
0.69140625,
0.1962890625,
0.7587890625,
0.1976318359375,
0.7587890625,
0.2083740234375,
0.69140625,
0.20703125
],
[
0.31201171875,
0.2279052734375,
0.44287109375,
0.2279052734375,
0.44287109375,
0.238037109375,
0.31201171875,
0.238037109375
],
[
0.69140625,
0.255126953125,
0.771484375,
0.255615234375,
0.771484375,
0.267578125,
0.69140625,
0.266357421875
],
[
0.69287109375,
0.266357421875,
0.8095703125,
0.266357421875,
0.8095703125,
0.276611328125,
0.69287109375,
0.276611328125
],
[
0.59716796875,
0.28271484375,
0.61376953125,
0.28271484375,
0.61376953125,
0.287353515625,
0.59716796875,
0.287353515625
],
[
0.61669921875,
0.27978515625,
0.64892578125,
0.276611328125,
0.650390625,
0.285400390625,
0.61767578125,
0.28857421875
],
[
0.176513671875,
0.30810546875,
0.82177734375,
0.30810546875,
0.82177734375,
0.318115234375,
0.176513671875,
0.318115234375
],
[
0.176513671875,
0.32080078125,
0.82177734375,
0.32080078125,
0.82177734375,
0.333984375,
0.176513671875,
0.333984375
],
[
0.176513671875,
0.333984375,
0.82177734375,
0.333984375,
0.82177734375,
0.348388671875,
0.176513671875,
0.348388671875
],
[
0.17724609375,
0.349853515625,
0.8203125,
0.349853515625,
0.8203125,
0.35986328125,
0.17724609375,
0.35986328125
],
[
0.1748046875,
0.361083984375,
0.8203125,
0.361083984375,
0.8203125,
0.375,
0.1748046875,
0.375
],
[
0.176513671875,
0.375732421875,
0.82275390625,
0.375732421875,
0.82275390625,
0.389404296875,
0.176513671875,
0.389404296875
],
[
0.1748046875,
0.38818359375,
0.82177734375,
0.38818359375,
0.82177734375,
0.401611328125,
0.1748046875,
0.401611328125
],
[
0.176513671875,
0.404052734375,
0.400390625,
0.404052734375,
0.400390625,
0.417236328125,
0.176513671875,
0.417236328125
],
[
0.176513671875,
0.449462890625,
0.82275390625,
0.449462890625,
0.82275390625,
0.46337890625,
0.176513671875,
0.46337890625
],
[
0.176513671875,
0.46337890625,
0.82177734375,
0.46337890625,
0.82177734375,
0.4765625,
0.176513671875,
0.4765625
],
[
0.17724609375,
0.477783203125,
0.82177734375,
0.477783203125,
0.82177734375,
0.4912109375,
0.17724609375,
0.4912109375
],
[
0.176513671875,
0.4912109375,
0.82177734375,
0.4912109375,
0.82177734375,
0.5048828125,
0.176513671875,
0.5048828125
],
[
0.176513671875,
0.5048828125,
0.82177734375,
0.5048828125,
0.82177734375,
0.51806640625,
0.176513671875,
0.51806640625
],
[
0.1748046875,
0.51806640625,
0.82275390625,
0.51806640625,
0.82275390625,
0.5322265625,
0.1748046875,
0.5322265625
],
[
0.176513671875,
0.5341796875,
0.273681640625,
0.5341796875,
0.273681640625,
0.54541015625,
0.176513671875,
0.54541015625
],
[
0.176513671875,
0.5537109375,
0.82177734375,
0.5537109375,
0.82177734375,
0.5673828125,
0.176513671875,
0.5673828125
],
[
0.176513671875,
0.56494140625,
0.82177734375,
0.56640625,
0.82177734375,
0.58056640625,
0.176513671875,
0.580078125
],
[
0.176513671875,
0.58056640625,
0.82177734375,
0.58056640625,
0.82177734375,
0.5947265625,
0.176513671875,
0.5947265625
],
[
0.17724609375,
0.5947265625,
0.82275390625,
0.5947265625,
0.82275390625,
0.60791015625,
0.17724609375,
0.60791015625
],
[
0.176513671875,
0.60595703125,
0.1968994140625,
0.609375,
0.19287109375,
0.62353515625,
0.17236328125,
0.619140625
],
[
0.1748046875,
0.65087890625,
0.2091064453125,
0.65087890625,
0.2091064453125,
0.66357421875,
0.1748046875,
0.66357421875
],
[
0.205078125,
0.65087890625,
0.326904296875,
0.65087890625,
0.326904296875,
0.6640625,
0.205078125,
0.6640625
],
[
0.1748046875,
0.6845703125,
0.82177734375,
0.68310546875,
0.82177734375,
0.69775390625,
0.1748046875,
0.69873046875
],
[
0.176513671875,
0.69873046875,
0.82177734375,
0.69873046875,
0.82177734375,
0.7119140625,
0.176513671875,
0.7119140625
],
[
0.176513671875,
0.7119140625,
0.82177734375,
0.7119140625,
0.82177734375,
0.72607421875,
0.176513671875,
0.72607421875
],
[
0.176513671875,
0.72607421875,
0.82275390625,
0.72607421875,
0.82275390625,
0.7392578125,
0.176513671875,
0.7392578125
],
[
0.17724609375,
0.7412109375,
0.4697265625,
0.7412109375,
0.4697265625,
0.751953125,
0.17724609375,
0.751953125
],
[
0.176513671875,
0.7587890625,
0.68115234375,
0.7587890625,
0.68115234375,
0.77294921875,
0.176513671875,
0.77294921875
],
[
0.68115234375,
0.76220703125,
0.82177734375,
0.76220703125,
0.82177734375,
0.77294921875,
0.68115234375,
0.77294921875
],
[
0.176513671875,
0.7744140625,
0.82177734375,
0.7744140625,
0.82177734375,
0.78857421875,
0.176513671875,
0.78857421875
],
[
0.178955078125,
0.7890625,
0.8203125,
0.7890625,
0.8203125,
0.7998046875,
0.178955078125,
0.7998046875
],
[
0.178955078125,
0.8017578125,
0.82275390625,
0.8017578125,
0.82275390625,
0.81494140625,
0.178955078125,
0.81494140625
],
[
0.178955078125,
0.8173828125,
0.82177734375,
0.8173828125,
0.82177734375,
0.82763671875,
0.178955078125,
0.82763671875
],
[
0.176513671875,
0.830078125,
0.82275390625,
0.830078125,
0.82275390625,
0.84326171875,
0.176513671875,
0.84326171875
],
[
0.176513671875,
0.84326171875,
0.82177734375,
0.84326171875,
0.82177734375,
0.85693359375,
0.176513671875,
0.85693359375
],
[
0.17724609375,
0.85693359375,
0.82177734375,
0.85693359375,
0.82177734375,
0.87060546875,
0.17724609375,
0.87060546875
],
[
0.17724609375,
0.87060546875,
0.82177734375,
0.87060546875,
0.82177734375,
0.8837890625,
0.17724609375,
0.8837890625
],
[
0.17724609375,
0.8837890625,
0.82177734375,
0.8837890625,
0.82177734375,
0.89794921875,
0.17724609375,
0.89794921875
],
[
0.176513671875,
0.8984375,
0.82177734375,
0.8984375,
0.82177734375,
0.912109375,
0.176513671875,
0.912109375
],
[
0.49267578125,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"Colorimage",
"colormap hot: higher value",
"are to the right of the scale",
"Pixelinvisibilitymar",
"detected objects",
"in color image",
"Pairedinfrared image",
"all objects showr",
"inpaired infrared images",
"",
"",
"Figurel:Our method predicts aninvisibilitymap for eachpixel(2ndrow)given onlythe color",
"image (Ist row). Left to right: It can handle distant objects, highly occluded and cluttered objects.",
"objects with strong glare, multiple objects from complex scenes like working zones on highway",
"andobjectsinthefog.Colordetectorscanonlydetectobjects(boundingboxesinIstrow)thatare",
"visible to its spectrum, while there are many more objects in the scene (3rd row). This could caus",
"disastrous consequences where safety is crucial if such missed detection is trusted. Our pixel-level",
"invisibility map indicates how much the detection results from color images may be trusted, both for",
"false negatives and false positives.",
"知识提炼[?]特别要求这两种模态以成对的方式呈现，特别是以良好对齐的方式用于对象检",
"测以及我们的像素级不可见性遮罩。这里，对齐良好的图像对是配对图像中的相应像素位于",
"其各自图像平面中的相同位置的图像对。由彩色和红外传感器捕获的原始图像对具有较大",
"的位移E[?,？]，这些位移来自（1）内部相机属性差异，例如焦距，分辨率和镜头失真，（2）",
"外部转换，例如姿势差异和（3）曝光时间和快门速度的时间差。我们使用已知的模式板通",
"过图像配准来解决前两个视差，同时我们提出了对齐G能量对抗N网络（AlignGAN）来减",
"轻剩余的位移",
"我们工作的贡献是三倍。（1）据我们所知，这是无需手动标记即可为彩色图像生成像素级隐",
"形遮罩的第一项工作，因此有助于解决自动驾驶中由感官假阴性引起的故障检测问题；（2）",
"从彩色图像到红外图像的中级表示的直接转移在红外域中具有很高的检测精度；（3）减轻彩",
"色红外对中存在的未对准问题（AlignGAN)。进行了广泛的实验，以定量评估我们系统的性",
"能",
"2",
"Related Work",
"我们的像素级不可见掩模与不确定性估计E[?,？,?,?,?]，故障检测E[?,?,?,?,？,？]和分布",
"外检测E[?,?,?,?]的任务紧密相连。大多数报告的工作要么将原始传感器数据视为内省检",
"查E[？]，要么估计基于模型的分类器的不确定性，要么计算置信度得分或二元决策。我们",
"的系统与它们的不同之处在于：（1）从局外人的角度估算传感器本身的置信度，即红外相",
"机：（2）预测图像中每个像素的置信概率",
"像素对象E[?1是计算所有类似对象区域的像素级掩码的第一项工作。",
"[?提出了一种交通",
"标志故障检测系统，该系统从目标检测器的特征图中提取激发区域，以缩小手动标记空间和",
"假阴性搜索空间。尽管他们仍然需要将兴奋区域标记为假阴性或真阴性。我们的工作还预测",
"了彩色图像中潜在不可见物体（假阴性）的所有区域的像素级掩模。相比之下，我们的方法",
"不需要对不可见区域进行任何手动标记，而是利用彩色红外图像对来提供这种监督。为了获",
"得对齐的彩色红外图像对，E[?1基于分束器创建了KAIST多光谱行人数据集，以将彩色和",
"红外相机的光束分为两部分。尽管观察到数据集中的图像对具有不同的位移E[??]，并且",
"由于分束器的强度衰减，在夜间受到的影响更大。[?]报告了不太严重的视差问题，尽管它",
"们的数据集尚未发布。我们在使用垂直对齐相机的设置收集的数据集中观察到相同的问题",
"并提出AlignGAN来减轻图像平面中的此类位移。Pix2pix[?]和CycleGan[?]开发了配对",
"和非配对设置下的跨域翻译方法。循环一致性[?]是解决不成对跨域翻译的主要技术。虽然",
"2"
] |
{
"arxiv_id": "2006.08383",
"language": "zh",
"page_idx": 1
}
|
||
2006.08383-zh-2
|
2006.08383-zh-2.jpg
|
<FigureHere>
Figure 2: System overview. Our system takes well-aligned image pairs from unlabeled video stream. In the training stage(left side), the Yolo-v3 color network is freezed, while Yolov3-infrared network is trained to reproduce the mid-level features as the ones in Yolov-v3 color network at three scales, which are located at layer 79, 91 and 103 in the network. In the test stage(right), the difference of those features from layer 79,91 and 103 are computed and fused to get the final invisibility mask.
这些翻译主要解决从源到目标域的风格转移，而不移动源域中的像素。我们的 AlignGAN 模
块使用从红外图像中学习的边缘图来生成对齐的彩色图像
本文的其余部分详细描述了我们的方法和实验。在$\mathbb{E}$3 节中，我们介绍了我们的系统并描述了网络架构以及用于生成像素级不可见性地图的训练过程。最后，在 4节中，我们对我们的数据集进行了广泛的实验，并显示了几个比较结果。将发布代码、数据和模型。
## 3 $\textbf{System Overview}$
我们提出了一个系统来学习以无监督的方式为彩色图像生成像素级不可见图。在训练阶段我们的系统将看同一场景的弱对齐彩色红外图像对作为输入。这种不完全对齐的图像对首先被配准以去除几何差异，然后由 AlignGAN 对齐以去除两种模态之间的剩余位移，如 3.1节所述。在图像对良好对齐后，我们的知识转移系统将学习到的表示从颜色域转移到红外域。然后在测试阶段，比较表示对，以估计彩色图像中每个像素的不可见性得分。作为副产品，红外图像的学习表示可以直接用于构建红外图像的对象检测器，而无需任何手动标记或重新训练，如图$\overline{\mathrm{E}}$2 所示
3.1 Alignment Generative Adversarial Network
AlignGAN: 鉴于原始图像对对齐不良，我们使用标准相机校准技术消除了两个相机之间的内部和外部转换，我们将在 4 节中进行描述。然后，我们提出了基于生成对抗网络(GAN) 的 AlignGAN, 以学习从弱对齐的彩色红外对生成对齐良好的彩色图像。系统的基本模块如图 3 所示。如图所示，在训练阶段的一次迭代中，我们使用两个流进行学习-都使用对齐块。第一个流使用彩色图像和弱对齐的红外图像作为边缘图作为源，并产生使用 Flownet2$\boxed{\mathbb{E}}[?]$ 作为目标创建的彩色图像。在另一个流中，源是图像是静态彩色图像，边缘图虽然来自视频流中的近色图像$I_{c1}$,但目标图像是彩色图像$I_c1$本身
我们基于 Pix2pix$\boxed{\mathbb{E}}[?]$构建了我们的系统。生成网络$G$和$G_m$都使用 U-Net 架构[?],输人
大小为 512$\times\hat{5}12$ 。
3.2 $\textbf{Knowledge Transfer From Color to Infrared Domains}$
我们使用 YOLO-V3 架构$\mathbb{E}[?]$ 运行实验主要基于三个原因。(1)它在自动驾驶汽车等实时应用中快速且可部署。(2) 不需要提案网络，例如在 FasterRCNN 和 MASK-RCNN 中。(3)
3
|
[
[
0.259033203125,
0.0921630859375,
0.2998046875,
0.0921630859375,
0.2998046875,
0.0997314453125,
0.259033203125,
0.0997314453125
],
[
0.259033203125,
0.1029052734375,
0.2998046875,
0.1029052734375,
0.2998046875,
0.10986328125,
0.259033203125,
0.10986328125
],
[
0.1968994140625,
0.111083984375,
0.23779296875,
0.111083984375,
0.23779296875,
0.11871337890625,
0.1968994140625,
0.11871337890625
],
[
0.69140625,
0.106689453125,
0.74169921875,
0.106689453125,
0.74169921875,
0.11871337890625,
0.69140625,
0.11871337890625
],
[
0.6953125,
0.126953125,
0.705078125,
0.126953125,
0.705078125,
0.1300048828125,
0.6953125,
0.1300048828125
],
[
0.56201171875,
0.1383056640625,
0.5908203125,
0.1383056640625,
0.5908203125,
0.1458740234375,
0.56201171875,
0.1458740234375
],
[
0.61279296875,
0.1363525390625,
0.6298828125,
0.1363525390625,
0.6298828125,
0.1395263671875,
0.61279296875,
0.1395263671875
],
[
0.66650390625,
0.14013671875,
0.67724609375,
0.1383056640625,
0.6787109375,
0.143310546875,
0.66845703125,
0.1458740234375
],
[
0.6953125,
0.1383056640625,
0.74609375,
0.1383056640625,
0.74609375,
0.1458740234375,
0.6953125,
0.1458740234375
],
[
0.62353515625,
0.1590576171875,
0.64892578125,
0.1590576171875,
0.64892578125,
0.1673583984375,
0.62353515625,
0.1673583984375
],
[
0.77294921875,
0.173583984375,
0.80615234375,
0.1717529296875,
0.80615234375,
0.1749267578125,
0.77294921875,
0.1767578125
],
[
0.54736328125,
0.2052001953125,
0.564453125,
0.2052001953125,
0.564453125,
0.2083740234375,
0.54736328125,
0.2083740234375
],
[
0.50341796875,
0.2183837890625,
0.51318359375,
0.2183837890625,
0.51318359375,
0.222900390625,
0.50341796875,
0.222900390625
],
[
0.78125,
0.25390625,
0.7958984375,
0.25390625,
0.7958984375,
0.25830078125,
0.78125,
0.25830078125
],
[
0.25732421875,
0.28857421875,
0.313720703125,
0.28857421875,
0.313720703125,
0.298583984375,
0.25732421875,
0.298583984375
],
[
0.2607421875,
0.298583984375,
0.31201171875,
0.29736328125,
0.31201171875,
0.30615234375,
0.2607421875,
0.306884765625
],
[
0.51318359375,
0.3017578125,
0.69921875,
0.3017578125,
0.69921875,
0.3125,
0.51318359375,
0.3125
],
[
0.250732421875,
0.326904296875,
0.82421875,
0.326904296875,
0.82421875,
0.34033203125,
0.250732421875,
0.34033203125
],
[
0.178955078125,
0.302978515625,
0.2607421875,
0.302978515625,
0.2607421875,
0.313232421875,
0.178955078125,
0.313232421875
],
[
0.176513671875,
0.32763671875,
0.25732421875,
0.32763671875,
0.25732421875,
0.339111328125,
0.176513671875,
0.339111328125
],
[
0.176513671875,
0.34033203125,
0.82421875,
0.34033203125,
0.82421875,
0.354248046875,
0.176513671875,
0.354248046875
],
[
0.176513671875,
0.35595703125,
0.82275390625,
0.35595703125,
0.82275390625,
0.3662109375,
0.176513671875,
0.3662109375
],
[
0.1732177734375,
0.3662109375,
0.82177734375,
0.3662109375,
0.82177734375,
0.380126953125,
0.1732177734375,
0.380126953125
],
[
0.1748046875,
0.3818359375,
0.82421875,
0.3818359375,
0.82421875,
0.395751953125,
0.1748046875,
0.395751953125
],
[
0.1748046875,
0.395751953125,
0.28759765625,
0.395751953125,
0.28759765625,
0.4091796875,
0.1748046875,
0.4091796875
],
[
0.1748046875,
0.4375,
0.82177734375,
0.4375,
0.82177734375,
0.45068359375,
0.1748046875,
0.45068359375
],
[
0.1748046875,
0.449462890625,
0.583984375,
0.45068359375,
0.583984375,
0.464111328125,
0.1748046875,
0.46337890625
],
[
0.176513671875,
0.4716796875,
0.82275390625,
0.4716796875,
0.82275390625,
0.48486328125,
0.176513671875,
0.48486328125
],
[
0.178955078125,
0.48486328125,
0.82275390625,
0.48486328125,
0.82275390625,
0.498779296875,
0.178955078125,
0.498779296875
],
[
0.1748046875,
0.5,
0.7490234375,
0.5,
0.7490234375,
0.51318359375,
0.1748046875,
0.51318359375
],
[
0.1748046875,
0.5361328125,
0.2091064453125,
0.5361328125,
0.2091064453125,
0.54736328125,
0.1748046875,
0.54736328125
],
[
0.205078125,
0.5361328125,
0.3544921875,
0.5361328125,
0.3544921875,
0.5498046875,
0.205078125,
0.5498046875
],
[
0.176513671875,
0.5673828125,
0.82177734375,
0.5673828125,
0.82177734375,
0.58056640625,
0.176513671875,
0.58056640625
],
[
0.176513671875,
0.58056640625,
0.82177734375,
0.58056640625,
0.82177734375,
0.5947265625,
0.176513671875,
0.5947265625
],
[
0.176513671875,
0.5947265625,
0.82275390625,
0.5947265625,
0.82275390625,
0.60791015625,
0.176513671875,
0.60791015625
],
[
0.178955078125,
0.60986328125,
0.82177734375,
0.60986328125,
0.82177734375,
0.62060546875,
0.178955078125,
0.62060546875
],
[
0.17724609375,
0.62255859375,
0.8203125,
0.62255859375,
0.8203125,
0.63330078125,
0.17724609375,
0.63330078125
],
[
0.1748046875,
0.63623046875,
0.82421875,
0.63623046875,
0.82421875,
0.6494140625,
0.1748046875,
0.6494140625
],
[
0.176513671875,
0.65087890625,
0.36767578125,
0.65087890625,
0.36767578125,
0.6640625,
0.176513671875,
0.6640625
],
[
0.1732177734375,
0.68115234375,
0.52294921875,
0.681640625,
0.52294921875,
0.6962890625,
0.1732177734375,
0.69580078125
],
[
0.178955078125,
0.708984375,
0.82177734375,
0.708984375,
0.82177734375,
0.72265625,
0.178955078125,
0.72265625
],
[
0.1748046875,
0.72021484375,
0.82177734375,
0.72021484375,
0.82177734375,
0.734375,
0.1748046875,
0.734375
],
[
0.176513671875,
0.736328125,
0.82275390625,
0.736328125,
0.82275390625,
0.75,
0.176513671875,
0.75
],
[
0.176513671875,
0.75,
0.82177734375,
0.75,
0.82177734375,
0.76318359375,
0.176513671875,
0.76318359375
],
[
0.176513671875,
0.76318359375,
0.82275390625,
0.76318359375,
0.82275390625,
0.7763671875,
0.176513671875,
0.7763671875
],
[
0.1748046875,
0.7763671875,
0.82177734375,
0.7763671875,
0.82177734375,
0.79052734375,
0.1748046875,
0.79052734375
],
[
0.1748046875,
0.79052734375,
0.5595703125,
0.79150390625,
0.5595703125,
0.80615234375,
0.1748046875,
0.8046875
],
[
0.1748046875,
0.81005859375,
0.82275390625,
0.81103515625,
0.82275390625,
0.82568359375,
0.1748046875,
0.82470703125
],
[
0.1748046875,
0.82568359375,
0.31201171875,
0.82568359375,
0.31201171875,
0.8388671875,
0.1748046875,
0.8388671875
],
[
0.1748046875,
0.85791015625,
0.59326171875,
0.85791015625,
0.59326171875,
0.8720703125,
0.1748046875,
0.8720703125
],
[
0.176513671875,
0.8837890625,
0.82275390625,
0.8837890625,
0.82275390625,
0.89794921875,
0.176513671875,
0.89794921875
],
[
0.176513671875,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.912109375,
0.176513671875,
0.912109375
],
[
0.49267578125,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"Yolo-v3Col",
"WeightsFre",
"ImagePairs",
"ielevelbay",
"Ma",
"Per-P",
"!",
"",
"FusionFunctio",
"Laler",
"",
"color",
"06",
"ityM",
"Yolo-v3 Infrare",
"Neights Active",
":TestStageI(Allweightsfreezed)",
"System overview.Our system takes well-aligned image pairs from unlabeled video",
"TrainingStage",
"Figure 2:",
"stream. In the training stage(left side), the Yolo-v3 color network is freezed, while Yolov3-infrared",
"networkistrainedtoreproduce the mid-levelfeatures astheones inYolov-v3 colornetworkat",
"three scales, which are located at layer 79, 91 and 103 in the network. In the test stage(right)",
"the difference of those features from layer 79, 91 and 103 are computed and fused to get the final",
"invisibilitymask",
"这些翻译主要解决从源到目标域的风格转移，而不移动源域中的像素。我们的AlignGAN模",
"块使用从红外图像中学习的边缘图来生成对齐的彩色图像",
"本文的其余部分详细描述了我们的方法和实验。在E3节中，我们介绍了我们的系统并描述",
"了网络架构以及用于生成像素级不可见性地图的训练过程。最后，在4节中，我们对我们",
"的数据集进行了广泛的实验，并显示了几个比较结果。将发布代码、数据和模型",
"3",
"SystemOverview",
"我们提出了一个系统来学习以无监督的方式为彩色图像生成像素级不可见图。在训练阶段",
"我们的系统将看同一场景的弱对齐彩色红外图像对作为输入。这种不完全对齐的图像对首",
"先被配准以去除几何差异，然后由AlignGAN对齐以去除两种模态之间的剩余位移，如3.1",
"节所述。在图像对良好对齐后，我们的知识转移系统将学习到的表示从颜色域转移到红外",
"域。然后在测试阶段：比较表示对，以估计彩色图像中每个像素的不可见性得分。作为副产",
"品，红外图像的学习表示可以直接用于构建红外图像的对象检测器，而无需任何手动标记",
"或重新训练，如图E2所示",
"3.1Alignment GenerativeAdversarial Network",
"AlignGAN：鉴于原始图像对对齐不良，我们使用标准相机校准技术消除了两个相机之间",
"的内部和外部转换，我们将在4节中进行描述。然后，我们提出了基于生成对抗网络（GAN",
"的AlignGAN，以学习从弱对齐的彩色红外对生成对齐良好的彩色图像。系统的基本模块如",
"图3所示。如图所示，在训练阶段的一次迭代中，我们使用两个流进行学习-都使用对齐块。",
"第一个流使用彩色图像和弱对齐的红外图像作为边缘图作为源，并产生使用Flownet2E[?]",
"作为目标创建的彩色图像。在另一个流中，源是图像是静态彩色图像，边缘图虽然来自视频",
"流中的近色图像Ic1，但目标图像是彩色图像I1本身",
"我们基于Pix2pixE[?]构建了我们的系统。生成网络G和Gm都使用U-Net架构[?]，输入",
"大小为512×512。",
"3.2KnowledgeTransferFrom Color to Infrared Domains",
"我们使用YOLO-V3架构E[?I运行实验主要基于三个原因。（1）它在自动驾驶汽车等实时",
"应用中快速且可部署。（2）不需要提案网络，例如在FasterRCNN和MASK-RCNN中。（3）",
"3"
] |
{
"arxiv_id": "2006.08383",
"language": "zh",
"page_idx": 2
}
|
||
2006.08383-zh-3
|
2006.08383-zh-3.jpg
|
具有基于三种不同图像尺度的三个检测模块，对可能出现在道路上的小物体具有良好的检
测能力
我们将表示从颜色域转移到红外
<FigureHere>
域。我们选择在检测阶段之前发生的 Yolo-V3 架构中传输法师的中学中级功能。这被称为仅中期转移。它们分别是层 79、91 和 103 的输出，分别来自三个不同的比例，如图$\operatorname{E}$2 的左侧所示。与$\operatorname{E[?]}$类似， 我们通过传输最后一层检测结果(仅 Yolo)以及中级特征和检测结果 (MidYolo) 进行了两个比较实
Figure 3: Base alignment network. The module takes as 验。。根据检测层 82、94 和 106 的
input source image and edge map from target position, and 输出计算检测结果
outputs the target image in the given position. The conver- 我们通过将学习到的中间特征与
来自颜色检测器的检测模块串联
$G$ is conditioned on the motion cues which are generated us- 在一起，构建了一个没有任何手ing network $G_m.$ The target prediction after network $G$ and 动标记和微调的红外检测器，这对motion cues generated by $\tilde{G_{m}}$ are fed into two discrimina- 于不同照明条件下的红外图像产
生了有希望的检测精度，如实验部
$\vec{\text{分}}\vec{\mathrm{E}}4.2\vec{\text{所示}}$
sion from source to target which is performed using network
tors $D$ and $D_m$ respectively.
# 3.3 Invisibility Estimation
我们现在有两个检测器$Yolo_c$ 和$Yolo_i$, 其中$Yolo_c$ 在注释丰富的数据集上进行了预训练而$Yolo_i$ 则使用对齐良好的图像对进行了训练，以再现与其对等彩色图像相同的中级特征。基于红外图像的中级特征比彩色图像中的特征受照明条件的影响要小得多的观察，如第 4.2节所示，我们使用特征差异来估计照明条件，从而估计彩色图像中每个像素的可见度
YOLO-v3 架构具有三个检测模块，用于在三个不同的尺度上进行估计，因此它在三个不同的尺度上提供了中级功能。我们提出了一个不可见性分数来整合不同尺度下的特征差异，如图右侧所示 2。在这里，我们将像素$s_i$的不可见性分数定义为公式 1 中彩色图像和红外图像之间的中级特征$\{d_k\mid k=1,2,3\}$的 L2 距离的函数$F$。这里$d_1$ 、$d_2$ 、$d_3$分别来自层 79、 91、103 , $t_k$是我们为$d_k$选择的最高值。最后，我们训练了一个基于 U-Net[?]的卷积神经网络，即使在没有红外图像的情况下也能生成不可见掩模。
(1)
$$\begin{aligned}F(d_1,d_2,d_3)=1-\frac{1}{3}\sum_{k=1}^{3}min((t_k-d_k)/t_k,0)\end{aligned}$$
4 $\textbf{Experiments}$
Experimental Setup: 本节介绍了预测彩色图像中不可检测区域以及从彩色到红外域的无监督知识转移的实验结果。我们构建了一个传感器平台，可以安装在任何汽车的车顶行李架上 (并且我们使用了诸如用于实验的设置)。我们使用 FLIRADK 相机和正确的 ZED 相机作为传感器对。使用 Caltech 校准工具和部分由铝箔制成的平面棋盘进行彩色红外校准。这消除了相机畸变的主要部分，并有助于建立一个粗略的估计共同的视野和尺度差异。通过基于来自两个图像平面的对应点对的单应性翘曲，可以很好地解决静态场景中的视差问题。然后将这种弱对齐的图像对作为 AlignGAN 的输入
Data set- Color-IR Pair: 我们从白天驾驶时收集的视频(约 120000 个图像对) 中采样了18546 个彩色红外对，以构建用于将中级特征从彩色域转移到红外域的训练集。对于验证数据集，我们收集并手动标记了 2000 个带有对象边界框的图像对，黎明时 500 个，黄昏时 500个，夜间 500 个，雾时 500 个。它们用于评估彩色图像中不可检测区域的预测性能和红外图像的检测性能。我们没有训练集的确切统计信息，因为我们没有手工标记它们。尽管我们手动标记的验证集包含 7821 辆汽车，1138 个交通标志，626 人和 343 辆卡车，但可以为我们的训练数据集提供线索
4
|
[
[
0.176513671875,
0.0921630859375,
0.82421875,
0.0921630859375,
0.82421875,
0.1060791015625,
0.176513671875,
0.1060791015625
],
[
0.1748046875,
0.10479736328125,
0.226318359375,
0.106689453125,
0.2254638671875,
0.121826171875,
0.1732177734375,
0.11932373046875
],
[
0.5810546875,
0.126953125,
0.82275390625,
0.126953125,
0.82275390625,
0.14013671875,
0.5810546875,
0.14013671875
],
[
0.369384765625,
0.1427001953125,
0.406005859375,
0.1427001953125,
0.406005859375,
0.154052734375,
0.369384765625,
0.154052734375
],
[
0.57861328125,
0.1395263671875,
0.82275390625,
0.14013671875,
0.82275390625,
0.154052734375,
0.57861328125,
0.15283203125
],
[
0.32275390625,
0.1483154296875,
0.338134765625,
0.1483154296875,
0.338134765625,
0.1578369140625,
0.32275390625,
0.1578369140625
],
[
0.238525390625,
0.1514892578125,
0.25,
0.1572265625,
0.2459716796875,
0.1622314453125,
0.2344970703125,
0.1572265625
],
[
0.580078125,
0.1546630859375,
0.82421875,
0.1546630859375,
0.82421875,
0.1685791015625,
0.580078125,
0.1685791015625
],
[
0.370849609375,
0.1558837890625,
0.4248046875,
0.1558837890625,
0.4248046875,
0.1666259765625,
0.370849609375,
0.1666259765625
],
[
0.200927734375,
0.143310546875,
0.2303466796875,
0.143310546875,
0.2303466796875,
0.15087890625,
0.200927734375,
0.15087890625
],
[
0.2059326171875,
0.1572265625,
0.2344970703125,
0.1572265625,
0.2344970703125,
0.160400390625,
0.2059326171875,
0.160400390625
],
[
0.463134765625,
0.1622314453125,
0.48046875,
0.1622314453125,
0.48046875,
0.175537109375,
0.463134765625,
0.175537109375
],
[
0.580078125,
0.1666259765625,
0.82275390625,
0.1673583984375,
0.82275390625,
0.1817626953125,
0.580078125,
0.18115234375
],
[
0.465576171875,
0.1749267578125,
0.48046875,
0.1817626953125,
0.47216796875,
0.189453125,
0.459228515625,
0.1817626953125
],
[
0.279296875,
0.1673583984375,
0.303955078125,
0.1673583984375,
0.303955078125,
0.1787109375,
0.279296875,
0.1787109375
],
[
0.181396484375,
0.1817626953125,
0.23779296875,
0.18115234375,
0.23779296875,
0.189453125,
0.181396484375,
0.1900634765625
],
[
0.484375,
0.1849365234375,
0.57421875,
0.1849365234375,
0.57421875,
0.1956787109375,
0.484375,
0.1956787109375
],
[
0.57861328125,
0.1817626953125,
0.82275390625,
0.1817626953125,
0.82275390625,
0.1956787109375,
0.57861328125,
0.1956787109375
],
[
0.464111328125,
0.1962890625,
0.4794921875,
0.1962890625,
0.4794921875,
0.199462890625,
0.464111328125,
0.199462890625
],
[
0.5810546875,
0.1962890625,
0.82275390625,
0.1962890625,
0.82275390625,
0.210205078125,
0.5810546875,
0.210205078125
],
[
0.200927734375,
0.208984375,
0.2303466796875,
0.208984375,
0.2303466796875,
0.216552734375,
0.200927734375,
0.216552734375
],
[
0.188720703125,
0.199462890625,
0.2459716796875,
0.199462890625,
0.2459716796875,
0.210205078125,
0.188720703125,
0.210205078125
],
[
0.265625,
0.208984375,
0.293212890625,
0.212158203125,
0.2900390625,
0.222900390625,
0.26318359375,
0.2197265625
],
[
0.464111328125,
0.208984375,
0.488525390625,
0.212158203125,
0.484375,
0.2279052734375,
0.4599609375,
0.2247314453125
],
[
0.5810546875,
0.208984375,
0.82275390625,
0.208984375,
0.82275390625,
0.2235107421875,
0.5810546875,
0.2235107421875
],
[
0.580078125,
0.2235107421875,
0.82275390625,
0.2235107421875,
0.82275390625,
0.2374267578125,
0.580078125,
0.2374267578125
],
[
0.495849609375,
0.2342529296875,
0.56640625,
0.2342529296875,
0.56640625,
0.24560546875,
0.495849609375,
0.24560546875
],
[
0.57861328125,
0.2374267578125,
0.82421875,
0.2374267578125,
0.82421875,
0.250732421875,
0.57861328125,
0.250732421875
],
[
0.181396484375,
0.2442626953125,
0.2344970703125,
0.2442626953125,
0.2344970703125,
0.255126953125,
0.181396484375,
0.255126953125
],
[
0.580078125,
0.25,
0.82421875,
0.250732421875,
0.82421875,
0.26513671875,
0.580078125,
0.264404296875
],
[
0.176513671875,
0.26953125,
0.564453125,
0.26953125,
0.564453125,
0.27978515625,
0.176513671875,
0.27978515625
],
[
0.580078125,
0.26513671875,
0.82275390625,
0.26513671875,
0.82275390625,
0.279052734375,
0.580078125,
0.279052734375
],
[
0.176513671875,
0.2822265625,
0.56787109375,
0.2822265625,
0.56787109375,
0.29541015625,
0.176513671875,
0.29541015625
],
[
0.57861328125,
0.279052734375,
0.71142578125,
0.279052734375,
0.71142578125,
0.292236328125,
0.57861328125,
0.292236328125
],
[
0.56396484375,
0.285400390625,
0.56787109375,
0.285400390625,
0.56787109375,
0.290283203125,
0.56396484375,
0.290283203125
],
[
0.176513671875,
0.29736328125,
0.5703125,
0.29736328125,
0.5703125,
0.30810546875,
0.176513671875,
0.30810546875
],
[
0.580078125,
0.300537109375,
0.82177734375,
0.300537109375,
0.82177734375,
0.311279296875,
0.580078125,
0.311279296875
],
[
0.1732177734375,
0.31005859375,
0.56640625,
0.30810546875,
0.56640625,
0.3212890625,
0.1732177734375,
0.323974609375
],
[
0.57861328125,
0.313232421875,
0.82421875,
0.313232421875,
0.82421875,
0.326904296875,
0.57861328125,
0.326904296875
],
[
0.1748046875,
0.322509765625,
0.56787109375,
0.322509765625,
0.56787109375,
0.3359375,
0.1748046875,
0.3359375
],
[
0.57666015625,
0.32568359375,
0.82275390625,
0.32568359375,
0.82275390625,
0.339111328125,
0.57666015625,
0.339111328125
],
[
0.176513671875,
0.337158203125,
0.56640625,
0.337158203125,
0.56640625,
0.35107421875,
0.176513671875,
0.35107421875
],
[
0.57861328125,
0.341552734375,
0.82275390625,
0.341552734375,
0.82275390625,
0.354736328125,
0.57861328125,
0.354736328125
],
[
0.1748046875,
0.35107421875,
0.57177734375,
0.35107421875,
0.57177734375,
0.3642578125,
0.1748046875,
0.3642578125
],
[
0.580078125,
0.354736328125,
0.82275390625,
0.354736328125,
0.82275390625,
0.36865234375,
0.580078125,
0.36865234375
],
[
0.17236328125,
0.363037109375,
0.361083984375,
0.365478515625,
0.361083984375,
0.37890625,
0.17236328125,
0.376953125
],
[
0.580078125,
0.3662109375,
0.82421875,
0.367431640625,
0.82421875,
0.3818359375,
0.580078125,
0.380615234375
],
[
0.580078125,
0.380126953125,
0.67236328125,
0.3818359375,
0.67236328125,
0.396484375,
0.580078125,
0.39453125
],
[
0.57666015625,
0.412353515625,
0.77783203125,
0.41357421875,
0.77783203125,
0.427978515625,
0.57666015625,
0.4267578125
],
[
0.176513671875,
0.43994140625,
0.82275390625,
0.43994140625,
0.82275390625,
0.453857421875,
0.176513671875,
0.453857421875
],
[
0.176513671875,
0.453857421875,
0.82275390625,
0.453857421875,
0.82275390625,
0.46728515625,
0.176513671875,
0.46728515625
],
[
0.176513671875,
0.46728515625,
0.82421875,
0.46728515625,
0.82421875,
0.48095703125,
0.176513671875,
0.48095703125
],
[
0.176513671875,
0.481689453125,
0.77783203125,
0.481689453125,
0.77783203125,
0.49560546875,
0.176513671875,
0.49560546875
],
[
0.1748046875,
0.501953125,
0.82177734375,
0.501953125,
0.82177734375,
0.51513671875,
0.1748046875,
0.51513671875
],
[
0.178955078125,
0.517578125,
0.82177734375,
0.517578125,
0.82177734375,
0.52783203125,
0.178955078125,
0.52783203125
],
[
0.17724609375,
0.52978515625,
0.82275390625,
0.52978515625,
0.82275390625,
0.54345703125,
0.17724609375,
0.54345703125
],
[
0.17724609375,
0.54345703125,
0.82177734375,
0.54345703125,
0.82177734375,
0.556640625,
0.17724609375,
0.556640625
],
[
0.1748046875,
0.556640625,
0.82275390625,
0.556640625,
0.82275390625,
0.57080078125,
0.1748046875,
0.57080078125
],
[
0.176513671875,
0.57080078125,
0.5908203125,
0.57080078125,
0.5908203125,
0.583984375,
0.176513671875,
0.583984375
],
[
0.4599609375,
0.58544921875,
0.67724609375,
0.5966796875,
0.67236328125,
0.6572265625,
0.455078125,
0.64599609375
],
[
0.339111328125,
0.60595703125,
0.464111328125,
0.60595703125,
0.464111328125,
0.619140625,
0.339111328125,
0.619140625
],
[
0.495849609375,
0.60107421875,
0.50341796875,
0.5947265625,
0.5107421875,
0.599609375,
0.50244140625,
0.60595703125
],
[
0.4599609375,
0.611328125,
0.464111328125,
0.611328125,
0.464111328125,
0.6162109375,
0.4599609375,
0.6162109375
],
[
0.80224609375,
0.60791015625,
0.826171875,
0.60791015625,
0.826171875,
0.6220703125,
0.80224609375,
0.6220703125
],
[
0.1707763671875,
0.66845703125,
0.3203125,
0.66943359375,
0.3203125,
0.6875,
0.1707763671875,
0.68603515625
],
[
0.178955078125,
0.70263671875,
0.346435546875,
0.70263671875,
0.346435546875,
0.71630859375,
0.178955078125,
0.71630859375
],
[
0.343017578125,
0.70263671875,
0.82421875,
0.70263671875,
0.82421875,
0.71630859375,
0.343017578125,
0.71630859375
],
[
0.176513671875,
0.71630859375,
0.82177734375,
0.71630859375,
0.82177734375,
0.72998046875,
0.176513671875,
0.72998046875
],
[
0.176513671875,
0.72998046875,
0.82177734375,
0.72998046875,
0.82177734375,
0.74365234375,
0.176513671875,
0.74365234375
],
[
0.176513671875,
0.7421875,
0.82177734375,
0.7421875,
0.82177734375,
0.755859375,
0.176513671875,
0.755859375
],
[
0.176513671875,
0.75830078125,
0.82275390625,
0.75830078125,
0.82275390625,
0.771484375,
0.176513671875,
0.771484375
],
[
0.17724609375,
0.771484375,
0.82177734375,
0.771484375,
0.82177734375,
0.78515625,
0.17724609375,
0.78515625
],
[
0.1748046875,
0.7841796875,
0.53173828125,
0.7841796875,
0.53173828125,
0.79736328125,
0.1748046875,
0.79736328125
],
[
0.178955078125,
0.81494140625,
0.82177734375,
0.81494140625,
0.82177734375,
0.8291015625,
0.178955078125,
0.8291015625
],
[
0.176513671875,
0.830078125,
0.82275390625,
0.830078125,
0.82275390625,
0.84326171875,
0.176513671875,
0.84326171875
],
[
0.1748046875,
0.84326171875,
0.82421875,
0.84326171875,
0.82421875,
0.85693359375,
0.1748046875,
0.85693359375
],
[
0.1748046875,
0.85595703125,
0.82275390625,
0.85693359375,
0.82275390625,
0.8720703125,
0.1748046875,
0.87060546875
],
[
0.176513671875,
0.87060546875,
0.82177734375,
0.87060546875,
0.82177734375,
0.8837890625,
0.176513671875,
0.8837890625
],
[
0.17724609375,
0.88623046875,
0.82177734375,
0.88623046875,
0.82177734375,
0.896484375,
0.17724609375,
0.896484375
],
[
0.1748046875,
0.89794921875,
0.339111328125,
0.8984375,
0.339111328125,
0.912109375,
0.1748046875,
0.9111328125
],
[
0.491943359375,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94775390625,
0.491943359375,
0.94775390625
]
] |
[
"具有基于三种不同图像尺度的三个检测模块，对可能出现在道路上的小物体具有良好的检",
"测能力",
"我们将表示从颜色域转移到红外",
"target",
"域。我们选择在检测阶段之前发生",
"G",
"",
"的Yolo-V3架构中传输法师的中学",
"",
"sourc",
"",
"D",
"中级功能。这被称为仅中期转移。",
"",
"*)",
"eE",
"+real/fake",
"它们分别是层79、91和103的输",
"",
"出，分别来自三个不同的比例，如",
"target",
"edge map:",
"Gm",
"Dm",
"图E2的左侧所示。与E[?]类似，",
"我们通过传输最后一层检测结果",
"real/fake",
"（仅Yolo）以及中级特征和检测结",
"",
"果（MidYolo）进行了两个比较实",
"Figure3:Basealignmentnetwork.Themoduletakes as",
"验。。根据检测层82、94和106的",
"input source image and edge map from target position, and",
"输出计算检测结果",
"",
"outputs the target image in the givenposition.The conver-",
"我们通过将学习到的中间特征与",
"sionfrom source to target which is performed using network",
"来自颜色检测器的检测模块串联",
"G is conditioned on the motion cues which are generated us-",
"在一起，构建了一个没有任何手",
"ing network Gm . The target prediction after network G and",
"动标记和微调的红外检测器，这对",
"motion cues generated by Gm are fed into two discrimina-",
"于不同照明条件下的红外图像产",
"tors D and Dm respectively.",
"生了有希望的检测精度，如实验部",
"分E4.2所示",
"3.3InvisibilityEstimation",
"我们现在有两个检测器Yoloc和Yoloi，其中Yoloc在注释丰富的数据集上进行了预训练，",
"而Yoloi则使用对齐良好的图像对进行了训练，以再现与其对等彩色图像相同的中级特征。",
"基于红外图像的中级特征比彩色图像中的特征受照明条件的影响要小得多的观察，如第4.2",
"节所示，我们使用特征差异来估计照明条件，从而估计彩色图像中每个像素的可见度",
"YOLO-v3架构具有三个检测模块，用于在三个不同的尺度上进行估计，因此它在三个不同",
"的尺度上提供了中级功能。我们提出了一个不可见性分数来整合不同尺度下的特征差异，如",
"图右侧所示2。在这里，我们将像素8i的不可见性分数定义为公式1中彩色图像和红外图",
"像之间的中级特征[dk|k=1,2,3}的L2距离的函数F。这里di、d2、d3分别来自层79",
"91、103，tk是我们为dk选择的最高值。最后，我们训练了二个基于U-Net[?]的卷积神经",
"网络，即使在没有红外图像的情况下也能生成不可见掩模。",
"-1 - .",
"F(di. d2, da) = 1 ",
"",
"",
"(1)",
"4Experiments",
"Experimental Setup:",
"本节介绍了预测彩色图像中不可检测区域以及从彩色到红外域的无",
"监督知识转移的实验结果。我们构建了一个传感器平台，可以安装在任何汽车的车顶行李",
"架上（并且我们使用了诸如用于实验的设置）。我们使用FLIRADK相机和正确的ZED相机",
"作为传感器对。使用Caltech校准工具和部分由铝箔制成的平面棋盘进行彩色红外校准。这",
"消除了相机畸变的主要部分，并有助于建立一个粗略的估计共同的视野和尺度差异。通过",
"基于来自两个图像平面的对应点对的单应性翘曲，可以很好地解决静态场景中的视差问题",
"然后将这种弱对齐的图像对作为AlignGAN的输入",
"Dataset-Color-IRPair：我们从白天驾驶时收集的视频（约120000个图像对）中采样了",
"18546个彩色红外对，以构建用于将中级特征从彩色域转移到红外域的训练集。对于验证数",
"据集，我们收集并手动标记了2000个带有对象边界框的图像对，黎明时500个，黄昏时500",
"个，夜间500个，雾时500个。它们用于评估彩色图像中不可检测区域的预测性能和红外图",
"像的检测性能。我们没有训练集的确切统计信息，因为我们没有手工标记它们。尽管我们手",
"动标记的验证集包含7821辆汽车，1138个交通标志，626人和343辆卡车，但可以为我们",
"的训练数据集提供线索",
"4"
] |
{
"arxiv_id": "2006.08383",
"language": "zh",
"page_idx": 3
}
|
||
2006.08383-zh-4
|
2006.08383-zh-4.jpg
|
<FigureHere>
Figure 4: Histogram of Distances: We show the result of pixel-level invisibility mask on capturing the false negatives of MASK-CNN for night scenes. With a threshold of 0.35(separated by empty bars and solid bars) for invisibility score, the invisibility map can cover 77.5 % of the pixels in the undetected objects(green bar) while only take up 35.9% of the pixels in the night images.
Experiment Focus: 在本节中，我们重点回答以下五个问题。(1)彩色图像中不可见区域$\dot{\hat{\text{的预测有多好？(2)通过知识转移对红外图像的检测性能有多好？(3)哪一级的表示转移}}}$ 会在检测精度上给出最好的结果？(4)我们的 AlignGAN 会增强知识转移过程吗？最后 (5) 传输性能如何随图像对的数量而变化？现在我们使用我们的结果定量回答这些问题
4.1 How good is the prediction of the undetected area in the color image?
我们使用来自两个成对的 DNN 的中级特征作为空间，其中欧几里得距离用作与红外图像相比的彩色图像可靠性的估计。我们的实验结果表明，所提出的系统可以生产出高质量的隐形面具。在我们的实验中，我们将$t_1,t_2,t_3$ 设置为4，3。5、3。2分别在方程 1 中
对于分别来自白天，黎明，黄昏和夜晚的图像对，我们计算它们的中级特征与从我们的系统中学到的不可见性得分之间的 L2 距离。如图E5 所示，特征差异和不可见性得分均增加， 而环境的光强度降低。这与观察到的彩色图像在更好的照明中更可靠是一致的。这种协议通过在彩色和红外相机中使用的不同光谱中继承的对光变化的不同反应来验证，这将在E4.2节中详细解释。日不可见性得分的高斯分布为(0。020,0。028),夜晚是(0。268,0。052)。这些措施是高度可分离的，分布中几乎没有重叠，如图 5 所示。这些量表明，不可见性得分可用于估计具有不同照明条件的彩色图像中的每像素不可见性
一个好的二进制可见性遮罩具有两个特征：(1)覆盖大多数不可检测的物体(2)仅覆盖图像中不可检测的区域，我们以量化的形式使用它来评估我们结果的有效性。现在我们报告所有黎明，黄昏和夜景的可见率，其中只有夜景的可见率在图E4 中显示 ,
现在我们报告定量分析的结果，显示预测彩色图像中不可检测区域的有效性。在图4 中，不可检测区域的分布和整个图像是不同的(平均0。61vs0。27), 表明该分数很好地代表了不可检测性。隐形面罩可以覆盖 77。5 能见度阈值为 0 的图像中不可检测区域的%。35 只报告35。9 整个图像的%。黎明时间的隐形面具覆盖 49。2 未检测区域的%,同时报告 15。4 全图的$\%$。黄昏的时候，它覆盖了73。4未检测区域的$\%$,而仅报告 22。2 图像的$\%$
4.2 Can paired data facilitate detection using transfer learning?
首先，我们报告通过中级功能的知识转移可以达到 46。图 2 彩色红外对数据集中红外图像的% 总体检测精度。由于学习不需要任何手动注释，也不需要任何重新训练，我们发现结果很有希望。同样，这显示了为彩色和红外图像创建配对图像的重要性。除了颜色-深度，颜色-姿势和图像-声音对之外，颜色-红外对在外观和纹理上彼此相似。我们在这里的结果可用
5
|
[
[
0.236083984375,
0.106689453125,
0.25,
0.106689453125,
0.25,
0.10986328125,
0.236083984375,
0.10986328125
],
[
0.2484130859375,
0.10479736328125,
0.3056640625,
0.10479736328125,
0.3056640625,
0.11236572265625,
0.2484130859375,
0.11236572265625
],
[
0.67236328125,
0.106689453125,
0.701171875,
0.106689453125,
0.701171875,
0.10986328125,
0.67236328125,
0.10986328125
],
[
0.7099609375,
0.106689453125,
0.7548828125,
0.106689453125,
0.7548828125,
0.10986328125,
0.7099609375,
0.10986328125
],
[
0.2017822265625,
0.11553955078125,
0.2254638671875,
0.11553955078125,
0.2254638671875,
0.126953125,
0.2017822265625,
0.126953125
],
[
0.2320556640625,
0.11236572265625,
0.306396484375,
0.11297607421875,
0.306396484375,
0.12371826171875,
0.2320556640625,
0.12249755859375
],
[
0.6435546875,
0.11553955078125,
0.7490234375,
0.11553955078125,
0.7490234375,
0.11932373046875,
0.6435546875,
0.11932373046875
],
[
0.78125,
0.11871337890625,
0.7939453125,
0.11871337890625,
0.7939453125,
0.12249755859375,
0.78125,
0.12249755859375
],
[
0.64453125,
0.1256103515625,
0.654296875,
0.1256103515625,
0.654296875,
0.1287841796875,
0.64453125,
0.1287841796875
],
[
0.67724609375,
0.1256103515625,
0.71728515625,
0.1256103515625,
0.71728515625,
0.1287841796875,
0.67724609375,
0.1287841796875
],
[
0.7099609375,
0.1351318359375,
0.74658203125,
0.1351318359375,
0.74658203125,
0.1383056640625,
0.7099609375,
0.1383056640625
],
[
0.20751953125,
0.143310546875,
0.2197265625,
0.143310546875,
0.2197265625,
0.149658203125,
0.20751953125,
0.149658203125
],
[
0.78125,
0.1483154296875,
0.80224609375,
0.1483154296875,
0.80224609375,
0.15283203125,
0.78125,
0.15283203125
],
[
0.200927734375,
0.169189453125,
0.2197265625,
0.169189453125,
0.2197265625,
0.173583984375,
0.200927734375,
0.173583984375
],
[
0.1968994140625,
0.1932373046875,
0.221435546875,
0.1932373046875,
0.221435546875,
0.2008056640625,
0.1968994140625,
0.2008056640625
],
[
0.19287109375,
0.2203369140625,
0.20751953125,
0.21533203125,
0.2132568359375,
0.2235107421875,
0.19775390625,
0.2291259765625
],
[
0.2099609375,
0.2197265625,
0.221435546875,
0.2197265625,
0.221435546875,
0.2235107421875,
0.2099609375,
0.2235107421875
],
[
0.193603515625,
0.231689453125,
0.200927734375,
0.2310791015625,
0.2034912109375,
0.2392578125,
0.1968994140625,
0.240478515625
],
[
0.188720703125,
0.240478515625,
0.223876953125,
0.236083984375,
0.2279052734375,
0.25830078125,
0.19287109375,
0.261474609375
],
[
0.7861328125,
0.236083984375,
0.79345703125,
0.236083984375,
0.79345703125,
0.2392578125,
0.7861328125,
0.2392578125
],
[
0.779296875,
0.261962890625,
0.8037109375,
0.261962890625,
0.8037109375,
0.26953125,
0.779296875,
0.26953125
],
[
0.2034912109375,
0.268310546875,
0.223876953125,
0.268310546875,
0.223876953125,
0.279052734375,
0.2034912109375,
0.279052734375
],
[
0.79736328125,
0.272705078125,
0.80224609375,
0.272705078125,
0.80224609375,
0.27783203125,
0.79736328125,
0.27783203125
],
[
0.205078125,
0.29345703125,
0.2197265625,
0.29345703125,
0.2197265625,
0.3017578125,
0.205078125,
0.3017578125
],
[
0.779296875,
0.290283203125,
0.7958984375,
0.290283203125,
0.7958984375,
0.298583984375,
0.779296875,
0.298583984375
],
[
0.2467041015625,
0.324462890625,
0.256591796875,
0.324462890625,
0.256591796875,
0.330078125,
0.2467041015625,
0.330078125
],
[
0.343017578125,
0.322509765625,
0.3544921875,
0.32763671875,
0.3505859375,
0.333251953125,
0.339111328125,
0.32763671875
],
[
0.176513671875,
0.354248046875,
0.82177734375,
0.354248046875,
0.82177734375,
0.367431640625,
0.176513671875,
0.367431640625
],
[
0.176513671875,
0.367431640625,
0.82177734375,
0.367431640625,
0.82177734375,
0.380615234375,
0.176513671875,
0.380615234375
],
[
0.1748046875,
0.380615234375,
0.82275390625,
0.380615234375,
0.82275390625,
0.39453125,
0.1748046875,
0.39453125
],
[
0.1748046875,
0.393310546875,
0.7587890625,
0.39453125,
0.7587890625,
0.4091796875,
0.1748046875,
0.40771484375
],
[
0.178955078125,
0.43994140625,
0.81884765625,
0.43994140625,
0.81884765625,
0.45068359375,
0.178955078125,
0.45068359375
],
[
0.176513671875,
0.451904296875,
0.82275390625,
0.451904296875,
0.82275390625,
0.46533203125,
0.176513671875,
0.46533203125
],
[
0.178955078125,
0.46728515625,
0.82177734375,
0.46728515625,
0.82177734375,
0.477783203125,
0.178955078125,
0.477783203125
],
[
0.176513671875,
0.479736328125,
0.76318359375,
0.479736328125,
0.76318359375,
0.49365234375,
0.176513671875,
0.49365234375
],
[
0.1748046875,
0.5146484375,
0.69775390625,
0.5146484375,
0.69775390625,
0.52783203125,
0.1748046875,
0.52783203125
],
[
0.176513671875,
0.54150390625,
0.82177734375,
0.54150390625,
0.82177734375,
0.5546875,
0.176513671875,
0.5546875
],
[
0.17724609375,
0.5546875,
0.82275390625,
0.5546875,
0.82275390625,
0.568359375,
0.17724609375,
0.568359375
],
[
0.176513671875,
0.568359375,
0.78515625,
0.568359375,
0.78515625,
0.58203125,
0.176513671875,
0.58203125
],
[
0.178955078125,
0.59033203125,
0.82177734375,
0.59033203125,
0.82177734375,
0.60107421875,
0.178955078125,
0.60107421875
],
[
0.176513671875,
0.60302734375,
0.82177734375,
0.60302734375,
0.82177734375,
0.6162109375,
0.176513671875,
0.6162109375
],
[
0.17724609375,
0.6162109375,
0.82275390625,
0.6162109375,
0.82275390625,
0.6298828125,
0.17724609375,
0.6298828125
],
[
0.176513671875,
0.6298828125,
0.82275390625,
0.62890625,
0.82275390625,
0.64306640625,
0.176513671875,
0.64453125
],
[
0.17724609375,
0.64453125,
0.82177734375,
0.64453125,
0.82177734375,
0.65771484375,
0.17724609375,
0.65771484375
],
[
0.17724609375,
0.65771484375,
0.82177734375,
0.65771484375,
0.82177734375,
0.671875,
0.17724609375,
0.671875
],
[
0.176513671875,
0.67041015625,
0.619140625,
0.671875,
0.619140625,
0.68505859375,
0.176513671875,
0.6845703125
],
[
0.188720703125,
0.6923828125,
0.82275390625,
0.6923828125,
0.82275390625,
0.70556640625,
0.188720703125,
0.70556640625
],
[
0.17724609375,
0.70556640625,
0.82177734375,
0.70556640625,
0.82177734375,
0.7197265625,
0.17724609375,
0.7197265625
],
[
0.176513671875,
0.72021484375,
0.70166015625,
0.72021484375,
0.70166015625,
0.734375,
0.176513671875,
0.734375
],
[
0.176513671875,
0.7412109375,
0.8203125,
0.7412109375,
0.8203125,
0.751953125,
0.176513671875,
0.751953125
],
[
0.176513671875,
0.75244140625,
0.82177734375,
0.75390625,
0.82177734375,
0.76806640625,
0.176513671875,
0.767578125
],
[
0.176513671875,
0.76806640625,
0.82177734375,
0.76806640625,
0.82177734375,
0.7822265625,
0.176513671875,
0.7822265625
],
[
0.176513671875,
0.7822265625,
0.82177734375,
0.7822265625,
0.82177734375,
0.79541015625,
0.176513671875,
0.79541015625
],
[
0.176513671875,
0.796875,
0.75341796875,
0.796875,
0.75341796875,
0.81005859375,
0.176513671875,
0.81005859375
],
[
0.1748046875,
0.830078125,
0.63134765625,
0.830078125,
0.63134765625,
0.8447265625,
0.1748046875,
0.8447265625
],
[
0.17724609375,
0.85791015625,
0.82177734375,
0.85791015625,
0.82177734375,
0.8720703125,
0.17724609375,
0.8720703125
],
[
0.176513671875,
0.87060546875,
0.82177734375,
0.87060546875,
0.82177734375,
0.8837890625,
0.176513671875,
0.8837890625
],
[
0.176513671875,
0.8837890625,
0.82177734375,
0.8837890625,
0.82177734375,
0.89794921875,
0.176513671875,
0.89794921875
],
[
0.178955078125,
0.8994140625,
0.8203125,
0.8994140625,
0.8203125,
0.91015625,
0.178955078125,
0.91015625
],
[
0.494384765625,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94775390625,
0.494384765625,
0.94775390625
]
] |
[
"",
"Jall:negative",
"aaahie",
"e.neaiwe",
"2.00",
"all:positive",
"cotocteanhoiactcnnsiti",
"70",
"",
"rtennhiar",
"rTs.ns",
"1.75",
".50",
"11.5",
"1.25",
"1",
"00",
"kid",
"5075",
"",
"0.50",
"0.50",
"7",
"0.25",
"0.25",
"0",
"0",
"Figure 4:Histogram of Distances: We show the result of pixel-level invisibility mask on capturing",
"the false negatives of MASK-CNN for night scenes. With a threshold of 0.35(separated by empty",
"bars and solid bars) for invisibility score, the invisibility map can cover 77.5 % of the pixels in the",
"undetected objects(green bar) while only take up 35.9 % of the pixels in the night images.",
"ExperimentFocus：在本节中，我们重点回答以下五个问题。（1）彩色图像中不可见区垃",
"的预测有多好？（2）通过知识转移对红外图像的检测性能有多好？（3）哪一级的表示转移",
"会在检测精度上给出最好的结果？（4）我们的AlignGAN会增强知识转移过程吗？最后（5",
"传输性能如何随图像对的数量而变化？现在我们使用我们的结果定量回答这些问题",
"4.1How good is the prediction of the undetected area in the color image?",
"我们使用来自两个成对的DNN的中级特征作为空间，其中欧几里得距离用作与红外图像相",
"比的彩色图像可靠性的估计。我们的实验结果表明，所提出的系统可以生产出高质量的隐",
"形面具。在我们的实验中，我们将t1，t2，t3设置为4，3。5、3。2分别在方程1中",
"对于分别来自白天，黎明，黄昏和夜晚的图像对，我们计算它们的中级特征与从我们的系",
"统中学到的不可见性得分之间的L2距离。如图E5所示，特征差异和不可见性得分均增加，",
"而环境的光强度降低。这与观察到的彩色图像在更好的照明中更可靠是一致的。这种协议通",
"过在彩色和红外相机中使用的不同光谱中继承的对光变化的不同反应来验证，这将在E4.2",
"节中详细解释。日不可见性得分的高斯分布为(0。020,0。028)，夜晚是（0。268,0。052)。这",
"些措施是高度可分离的，分布中几乎没有重叠，如图5所示。这些量表明，不可见性得分",
"可用于估计具有不同照明条件的彩色图像中的每像素不可见性",
"个好的二进制可见性遮罩具有两个特征：（1）覆盖大多数不可检测的物体（2）仅覆盖图",
"像中不可检测的区域，我们以量化的形式使用它来评估我们结果的有效性。现在我们报告",
"所有黎明，黄昏和夜景的可见率，其中只有夜景的可见率在图E4中显示，",
"现在我们报告定量分析的结果，显示预测彩色图像中不可检测区域的有效性。在图4中，不",
"可检测区域的分布和整个图像是不同的（平均0。61vs0。27)，表明该分数很好地代表了不可",
"检测性。隐形面罩可以覆盖77。5能见度阈值为0的图像中不可检测区域的%。35只报告",
"35。9整个图像的%。黎明时间的隐形面具覆盖49。2未检测区域的%，同时报告15。4全",
"图的%。黄昏的时候，它覆盖了73。4未检测区域的%，而仅报告22。2图像的%",
"4.2Can paired data facilitate detection using transfer learning?",
"首先，我们报告通过中级功能的知识转移可以达到46。图2彩色红外对数据集中红外图像",
"的%总体检测精度。由于学习不需要任何手动注释，也不需要任何重新训练，我们发现结果",
"很有希望。同样，这显示了为彩色和红外图像创建配对图像的重要性。除了颜色-深度，颜",
"色-姿势和图像-声音对之外，颜色-红外对在外观和纹理上彼此相似。我们在这里的结果可用",
"5"
] |
{
"arxiv_id": "2006.08383",
"language": "zh",
"page_idx": 4
}
|
||
2006.05580-zh-0
|
2006.05580-zh-0.jpg
|
Syn2Real Transfer Learning for Image Deraining using Gaussian Processes
$\begin{array}{cc}\text{Rajeev Yasarla}^*&&\text{Vishwanath A. Sindagi }^*&\text{Vishal M.Patel}\\&&\text{Johns Hopkins University}\end{array}$
Department of Electrical and Computer Engineering, Baltimore, MD 21218, USA
$\{$ ryasarl1, vishwanathsindagi, vpatel$36\}$ @jhu.edu
Abstract
<FigureHere>
Figure 1. Top row: (a) Input rainy image. (b) Output
Recent CNN-based methods for image deraining have achieved excellent performance in terms of reconstruction error as well as visual quality. However, these methods are limited in the sense that they can be trained only on fully labeled data. Due to various challenges in obtaining real world fully-labeled image deraining datasets, existing methods are trained only on synthetically generated data and hence, generalize poorly to real-world images. The use of real-world data in training image deraining networks is relatively less explored in the literature. We propose a Gaussian Process-based semi-supervised learning framework which enables the network in learning to derain using synthetic dataset while generalizing better using unlabeled real-world images. Through extensive experiments and ablations on several challenging datasets (such as Rain800, Rain200H and DDN-SIRR), we show that the proposed method, when trained on limited labeled data, achieves on-par performance with fullylabeled training. Additionally, we demonstrate that using unlabeled real-world images in the proposed GPbased framework results in superior performance as compared to existing methods. Code is available at: $\bar{\text{https:}//\text{github.com/rajeevyasarla/Syn}2}$Real .
from a network trained using only the synthetic data. (c) Output from a network trained using the synthetic data and unlabeled real-world data. This shows better generalization. Bottom row: Results from Semi-supervised learning (SSL) experiments. Reducing the amount of labeled data used for training results in the performance drop. Using the proposed SSL framework, we are able to recover the $\tilde{\mathrm{performance}}.$
## 1. Introduction
在雨天条件下拍摄的图像通常质量很差。雨纹引人的伪影对后续计算机视觉算法的性能产生不利影响，如目标检测和识别[12,28,41,4]。随着此类算法成为自主导航和视频监控[37,25,36]等多种应用中的重要组成部分，开发用于除雨的复杂算法变得越来越重要
除雨的任务受到几个问题的困扰，例如(i)雨纹的规模，密度和方向的巨大变化，以及(ii)缺乏真实世界的标记训练数据。大多数现有的工作[58,9,23,51, 60,6,52,20]在图像脱轨主要集中在解决第一个问题。
例如，Fu et al.[9] 开发了一种端到端方法，该方法侧重于训练脱轨网络期间的高频细节。在另一项工作中， Zhang 和 Patel [58] 提出了一种密度感知的多蒸汽密集连接网络，用于联合雨密度估计和除水。Li et al.[21] 通过递归神经网络结合上下文信息进行雨水去除。最近，Ren et al. [40]引人了渐进式 ResNet, 它利用了各
$^*$equal contribution
1
|
[
[
0.1094970703125,
0.1319580078125,
0.86279296875,
0.1331787109375,
0.86279296875,
0.15087890625,
0.1094970703125,
0.149658203125
],
[
0.205078125,
0.18310546875,
0.339111328125,
0.18310546875,
0.339111328125,
0.1976318359375,
0.205078125,
0.1976318359375
],
[
0.373291015625,
0.1817626953125,
0.583984375,
0.18310546875,
0.583984375,
0.1976318359375,
0.373291015625,
0.1962890625
],
[
0.6337890625,
0.18310546875,
0.76708984375,
0.18310546875,
0.76708984375,
0.1962890625,
0.6337890625,
0.1962890625
],
[
0.37744140625,
0.2008056640625,
0.59326171875,
0.2008056640625,
0.59326171875,
0.2139892578125,
0.37744140625,
0.2139892578125
],
[
0.1405029296875,
0.216552734375,
0.82861328125,
0.21533203125,
0.82861328125,
0.23291015625,
0.1405029296875,
0.2342529296875
],
[
0.316162109375,
0.2374267578125,
0.654296875,
0.2374267578125,
0.654296875,
0.251953125,
0.316162109375,
0.251953125
],
[
0.238525390625,
0.287353515625,
0.31201171875,
0.287353515625,
0.31201171875,
0.3017578125,
0.238525390625,
0.3017578125
],
[
0.09722900390625,
0.318115234375,
0.4697265625,
0.3193359375,
0.4697265625,
0.333984375,
0.09722900390625,
0.333251953125
],
[
0.080078125,
0.335205078125,
0.46728515625,
0.335205078125,
0.46728515625,
0.348388671875,
0.080078125,
0.348388671875
],
[
0.08251953125,
0.3515625,
0.46728515625,
0.3515625,
0.46728515625,
0.3623046875,
0.08251953125,
0.3623046875
],
[
0.07843017578125,
0.3642578125,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.07843017578125,
0.37744140625
],
[
0.080078125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.08087158203125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.4091796875,
0.08087158203125,
0.4091796875
],
[
0.07843017578125,
0.410400390625,
0.4697265625,
0.4091796875,
0.4697265625,
0.423583984375,
0.07843017578125,
0.4248046875
],
[
0.56201171875,
0.419189453125,
0.5849609375,
0.419189453125,
0.5849609375,
0.434326171875,
0.56201171875,
0.434326171875
],
[
0.685546875,
0.419189453125,
0.708984375,
0.419189453125,
0.708984375,
0.43310546875,
0.685546875,
0.43310546875
],
[
0.810546875,
0.419189453125,
0.83251953125,
0.419189453125,
0.83251953125,
0.434326171875,
0.810546875,
0.434326171875
],
[
0.609375,
0.4375,
0.68896484375,
0.4375,
0.68896484375,
0.448974609375,
0.609375,
0.448974609375
],
[
0.07843017578125,
0.4248046875,
0.468017578125,
0.425537109375,
0.468017578125,
0.43994140625,
0.07843017578125,
0.439453125
],
[
0.08251953125,
0.442626953125,
0.46728515625,
0.442626953125,
0.46728515625,
0.45263671875,
0.08251953125,
0.45263671875
],
[
0.70751953125,
0.438232421875,
0.78125,
0.438232421875,
0.78125,
0.448974609375,
0.70751953125,
0.448974609375
],
[
0.080078125,
0.455810546875,
0.468017578125,
0.455810546875,
0.468017578125,
0.4697265625,
0.080078125,
0.4697265625
],
[
0.07843017578125,
0.4697265625,
0.46728515625,
0.47021484375,
0.46728515625,
0.48486328125,
0.07843017578125,
0.484130859375
],
[
0.529296875,
0.466552734375,
0.55126953125,
0.4716796875,
0.5458984375,
0.484130859375,
0.52392578125,
0.479248046875
],
[
0.79736328125,
0.474853515625,
0.81787109375,
0.474853515625,
0.81787109375,
0.488037109375,
0.79736328125,
0.488037109375
],
[
0.08087158203125,
0.4873046875,
0.46728515625,
0.4873046875,
0.46728515625,
0.49755859375,
0.08087158203125,
0.49755859375
],
[
0.60693359375,
0.481689453125,
0.61279296875,
0.481689453125,
0.61279296875,
0.490478515625,
0.60693359375,
0.490478515625
],
[
0.52783203125,
0.488037109375,
0.54833984375,
0.488037109375,
0.54833984375,
0.50244140625,
0.52783203125,
0.50244140625
],
[
0.80126953125,
0.490478515625,
0.81396484375,
0.490478515625,
0.81396484375,
0.494384765625,
0.80126953125,
0.494384765625
],
[
0.080078125,
0.5,
0.4697265625,
0.50048828125,
0.4697265625,
0.51513671875,
0.080078125,
0.5146484375
],
[
0.6748046875,
0.505859375,
0.68701171875,
0.505859375,
0.68701171875,
0.51806640625,
0.6748046875,
0.51806640625
],
[
0.08087158203125,
0.5166015625,
0.46728515625,
0.5166015625,
0.46728515625,
0.52978515625,
0.08087158203125,
0.52978515625
],
[
0.53125,
0.51220703125,
0.54833984375,
0.51220703125,
0.54833984375,
0.521484375,
0.53125,
0.521484375
],
[
0.5048828125,
0.51953125,
0.51708984375,
0.51953125,
0.51708984375,
0.53271484375,
0.5048828125,
0.53271484375
],
[
0.080078125,
0.53076171875,
0.46728515625,
0.53076171875,
0.46728515625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50634765625,
0.53271484375,
0.51708984375,
0.53271484375,
0.51708984375,
0.54541015625,
0.50634765625,
0.54541015625
],
[
0.529296875,
0.53076171875,
0.5498046875,
0.53076171875,
0.5498046875,
0.54345703125,
0.529296875,
0.54345703125
],
[
0.50634765625,
0.54443359375,
0.51708984375,
0.54443359375,
0.51708984375,
0.55810546875,
0.50634765625,
0.55810546875
],
[
0.60888671875,
0.54248046875,
0.62353515625,
0.54248046875,
0.62353515625,
0.56103515625,
0.60888671875,
0.56103515625
],
[
0.53125,
0.55615234375,
0.5458984375,
0.55615234375,
0.5458984375,
0.56298828125,
0.53125,
0.56298828125
],
[
0.080078125,
0.546875,
0.4697265625,
0.546875,
0.4697265625,
0.56005859375,
0.080078125,
0.56005859375
],
[
0.08087158203125,
0.56103515625,
0.468017578125,
0.56103515625,
0.468017578125,
0.57470703125,
0.08087158203125,
0.57470703125
],
[
0.080078125,
0.5771484375,
0.468017578125,
0.5771484375,
0.468017578125,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.07843017578125,
0.59033203125,
0.46728515625,
0.59130859375,
0.46728515625,
0.60595703125,
0.07843017578125,
0.60498046875
],
[
0.56787109375,
0.60400390625,
0.58251953125,
0.60400390625,
0.58251953125,
0.61279296875,
0.56787109375,
0.61279296875
],
[
0.59716796875,
0.599609375,
0.85693359375,
0.599609375,
0.85693359375,
0.61279296875,
0.59716796875,
0.61279296875
],
[
0.080078125,
0.60693359375,
0.468017578125,
0.60693359375,
0.468017578125,
0.6220703125,
0.080078125,
0.6220703125
],
[
0.08087158203125,
0.62060546875,
0.46728515625,
0.62060546875,
0.46728515625,
0.63525390625,
0.08087158203125,
0.63525390625
],
[
0.56396484375,
0.62060546875,
0.583984375,
0.62060546875,
0.583984375,
0.63330078125,
0.56396484375,
0.63330078125
],
[
0.619140625,
0.62060546875,
0.63134765625,
0.62060546875,
0.63134765625,
0.6318359375,
0.619140625,
0.6318359375
],
[
0.68115234375,
0.6220703125,
0.69677734375,
0.6220703125,
0.69677734375,
0.6318359375,
0.68115234375,
0.6318359375
],
[
0.74658203125,
0.6220703125,
0.76123046875,
0.6220703125,
0.76123046875,
0.6318359375,
0.74658203125,
0.6318359375
],
[
0.81201171875,
0.6220703125,
0.82666015625,
0.6220703125,
0.82666015625,
0.6318359375,
0.81201171875,
0.6318359375
],
[
0.080078125,
0.63720703125,
0.468017578125,
0.63720703125,
0.468017578125,
0.65087890625,
0.080078125,
0.65087890625
],
[
0.646484375,
0.63525390625,
0.744140625,
0.63525390625,
0.744140625,
0.64599609375,
0.646484375,
0.64599609375
],
[
0.50341796875,
0.65087890625,
0.576171875,
0.65087890625,
0.576171875,
0.6611328125,
0.50341796875,
0.6611328125
],
[
0.57861328125,
0.65087890625,
0.890625,
0.65087890625,
0.890625,
0.6611328125,
0.57861328125,
0.6611328125
],
[
0.50341796875,
0.66357421875,
0.890625,
0.66357421875,
0.890625,
0.6767578125,
0.50341796875,
0.6767578125
],
[
0.080078125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.080078125,
0.66552734375
],
[
0.080078125,
0.66748046875,
0.40283203125,
0.66748046875,
0.40283203125,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.68994140625,
0.50341796875,
0.68994140625
],
[
0.50341796875,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.50341796875,
0.7041015625,
0.8896484375,
0.705078125,
0.8896484375,
0.71630859375,
0.50341796875,
0.71533203125
],
[
0.08087158203125,
0.71337890625,
0.21728515625,
0.71337890625,
0.21728515625,
0.72802734375,
0.08087158203125,
0.72802734375
],
[
0.50244140625,
0.71728515625,
0.89208984375,
0.71630859375,
0.89208984375,
0.73095703125,
0.50244140625,
0.73193359375
],
[
0.50244140625,
0.73095703125,
0.890625,
0.73095703125,
0.890625,
0.744140625,
0.50244140625,
0.744140625
],
[
0.1004638671875,
0.7392578125,
0.46728515625,
0.7392578125,
0.46728515625,
0.75244140625,
0.1004638671875,
0.75244140625
],
[
0.5009765625,
0.74560546875,
0.89208984375,
0.744140625,
0.89208984375,
0.75830078125,
0.5009765625,
0.7587890625
],
[
0.08251953125,
0.75390625,
0.468017578125,
0.75390625,
0.468017578125,
0.76806640625,
0.08251953125,
0.76806640625
],
[
0.50341796875,
0.76123046875,
0.5908203125,
0.76123046875,
0.5908203125,
0.77294921875,
0.50341796875,
0.77294921875
],
[
0.08416748046875,
0.76806640625,
0.465576171875,
0.76806640625,
0.465576171875,
0.7822265625,
0.08416748046875,
0.7822265625
],
[
0.08416748046875,
0.7841796875,
0.468017578125,
0.7841796875,
0.468017578125,
0.79736328125,
0.08416748046875,
0.79736328125
],
[
0.08087158203125,
0.7998046875,
0.439453125,
0.7998046875,
0.439453125,
0.81298828125,
0.08087158203125,
0.81298828125
],
[
0.101318359375,
0.81494140625,
0.46728515625,
0.81494140625,
0.46728515625,
0.8291015625,
0.101318359375,
0.8291015625
],
[
0.50341796875,
0.81005859375,
0.8896484375,
0.81005859375,
0.8896484375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08251953125,
0.83056640625,
0.468017578125,
0.83056640625,
0.468017578125,
0.8447265625,
0.08251953125,
0.8447265625
],
[
0.5048828125,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.5048828125,
0.8388671875
],
[
0.08251953125,
0.84521484375,
0.46728515625,
0.84521484375,
0.46728515625,
0.859375,
0.08251953125,
0.859375
],
[
0.50341796875,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.861328125,
0.468017578125,
0.861328125,
0.468017578125,
0.875,
0.080078125,
0.875
],
[
0.5048828125,
0.85693359375,
0.88818359375,
0.85693359375,
0.88818359375,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.098876953125,
0.8896484375,
0.2181396484375,
0.8896484375,
0.2181396484375,
0.8994140625,
0.098876953125,
0.8994140625
],
[
0.50341796875,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.486083984375,
0.9248046875,
0.49267578125,
0.9296875,
0.486083984375,
0.93505859375,
0.4794921875,
0.9296875
]
] |
[
"Syn2Real Transfer Learning for Image Deraining using Gaussian Processes",
"Rajeev Yasarla*",
"Vishwanath A.Sindagi *",
"Vishal M. Patel",
"Johns Hopkins University",
"Department of Electrical and Computer Engineering, Baltimore, MD 21218, USA",
"[ ryasarll, vishwanathsindagi, vpatel36 ↓ @jhu.edu",
"Abstract",
"Recent CNN-based methods for image deraining",
"have achieved excellent performance in terms of re-",
"construction error as well as visual quality.However,",
"these methods are limited in the sense that they can",
"be trained only on fully labeled data.Due to vari-",
"ous challenges in obtaining real world fully-labeled im-",
"age deraining datasets, existing methods are trained",
"(a)",
"(b)",
"(c)",
"De=Labeled data",
"only on synthetically generated data and hence, gener-",
"alizepoorlytoreal-worldimages.Theuse of real-world",
"Duy=Unlabeled de",
"data in training image deraining networks is relatively",
"less explored in the literature.We propose a Gaus-",
"23.5",
"",
"sianProcess-based semi-supervisedlearningframework",
"-",
"23.0",
"",
"which enables the network in learning to derain using",
"福",
"synthetic dataset while generalizing better using un-",
"22.5",
"0",
"labeled real-world images.Through extensive exper-",
"S",
"22.0",
"NSd",
"uren",
"21.5",
"iments and ablations on several challenging datasets",
"(such as Rain800, Rain200H and DDN-SIRR), we show",
"that the proposed method, when trained on limited",
"labeled data, achieves on-par performance with fully-",
"De",
"DeDe+DuDeDe+DuDDe+DuD,D+Dy",
"labeled training. Additionally, we demonstrate that",
"using unlabeled real-world images in the proposed GP-",
"100",
"10",
"20",
"40",
"60",
"based framework results in superior performance as",
"(% of Labeled data)",
"Figure 1.",
"Toprow:(a)Inputrainy image.(b)Output",
"from a network trained using only the synthetic data. (c)",
"compared to existing methods. Code is available at:",
"https://github.com/rajeevyasarla/Syn2Real",
"Output from a network trained using the synthetic data",
"and unlabeled real-world data. This shows better general-",
"ization.Bottom row:Resultsfrom Semi-supervised learn-",
"1.Introduction",
"ing (SSL) experiments.Reducing the amount of labeled",
"data used for training results in the performance drop. Us-",
"在雨天条件下拍摄的图像通常质量很差。雨纹引人",
"ingtheproposed SSLframework,weareabletorecoverthe",
"的伪影对后续计算机视觉算法的性能产生不利影响，如",
"performance.",
"目标检测和识别「12，28，41，41。随着此类算法成为自",
"主导航和视频监控[37,2536]等多种应用中的重要组",
"成部分，开发用于除雨的复杂算法变得越来越重要",
"除雨的任务受到几个问题的困扰，例如（i）雨纹的",
"例如，Fuetal.[9]开发了一种端到端方法，该方法侧",
"规模，密度和方向的巨大变化，以及（ii）缺乏真实世",
"重于训练脱轨网络期间的高频细节。在另一项工作中，",
"界的标记训练数据。大多数现有的工作58，9，23，51，",
"Zhang和Patel[58]提出了一种密度感知的多蒸汽密集",
"60，6，52，201在图像脱轨主要集中在解决第一个问题。",
"连接网络，用于联合雨密度估计和除水。Lietal.[21",
"通过递归神经网络结合上下文信息进行雨水去除。最",
"*equalcontribution",
"近，Renetal.[40]引人了渐进式ResNet，它利用了各",
"1"
] |
{
"arxiv_id": "2006.05580",
"language": "zh",
"page_idx": 0
}
|
||
2006.05580-zh-1
|
2006.05580-zh-1.jpg
|
<FigureHere>
Figure 2. Derained results. (a) Input rainy images. (b) SSIR output [49] . (c) Our output. It can be observed that the proposed method achieves better deraining.
个阶段的功能依赖关系。虽然这些方法在获得高质量的出轨图像方面取得了卓越的性能，但由于它们是完全监督的网络，并且只能利用完全标记的训练数据，因此它们固有地受到限制。然而，如前所述，获得标记的现实世界的训练数据是相当具有挑战性的，因此，现有的方法通常只在综合生成的雨数据集[57,51]上训练他们的网络
使用合成数据集会导致实际图像的性能不佳，这通常是由于合成图像和雨天图像之间的分布偏移[4]。尽管性能上存在差距，但这个问题在文献中仍然相对未被探索。
最近，Wei et al. [49] 提出了一种半监督学习框架(SIRR),在该框架中，他们同时从标记和未标记的数据中学习，以实现图像脱轨。对于标记数据的训练，他们使用传统的预测和地面实况 (GT) 之间的平均绝对误差损失。对于未标记的数据，他们通过施加在高斯混合模型 (GMM) 上的似然项来建模雨残差 (输入和输出之间的差)此外，它们通过最小化它们之间的Kullback-Leibler (KL) 差异来增强合成雨的分布更接近真实雨的分布的一致性。这是在半监督学习框架中制定图像脱轨任务的第一种方法，该框架可以利用未标记的现实世界图像来提高泛化能力。尽管该方法取得了有希望的结果，但它具有以下缺点：(i)由于雨残差的多模态性质，作者假设可以使用 GMM 对其进行建模。只有当实际残差用于计算 GMM 参数时，才会出现这种情况。然而，作者在训练迭代中使用真实世界 (末标记) 图像的预测雨残差来建模 GMM。然后使用相同的模型来计算后续迭代中(未标记图像的)预测残差的可能性。因此，如果在初始迭代集期间学习的GMM 参数不准确，这很可能是在训练的早期阶段，它将导致次优性能。(二)利用 KL 散度的目的是使合成雨分布更接近真实雨分布。如前所述，在训练的早期阶段，实际降雨残差的预测将不准确，因此，最小化两个分布之间的差异可能不合适。(iii)使用 GMM 对雨残差进行建模需要选择混合成分的数量，从而使模型对此类选择敏感
受 Wei et al. [49] 的启发，我们解决了将未标记的现实世界图像合并到训练过程中的问题，以便通过克服其方法的缺点来更好地泛化。与[49]相比，我们使用非参数方法为未标记的数据生成监督。具体来说，我们提出了一种基于高斯过程 (GP)的半监督学习 (SSL) 框架，该框架涉及对标记和未标记数据进行迭代训练。标记学习阶段涉及使用预测和地面实况之间的均方误差对标记数据进行训练。此外，输人(来自标记的数据集)被投影到潜在空间，然后使用 GP 建模。在未标记的训练阶段，我们使用在标记的训练阶段早期建模的GP 为未标记的输入生成伪 GT。然后，此伪 GT 用于监督未标记数据的中间潜在空间。伪 GT 的创建基于以下假设：未标记的图像在投影到潜在空间时可以表示为标记数据特征的加权组合，其中权重是使用核函数确定的。这些权重表示用于制定未标记数据点的标记数据点的不确定性。因此，最小化未标记数据投影和伪 GT 之间的误差可以减少方差，从而使网络权重自动适应未标记数据的域。Fig。1演示了使用建议的框架利用未标记数据的结果。Fig。2 将该方法的结果与 SIRR [49] 进行了比较。可以清楚地看到，与 SIRR [49]相比，我们的方法能够提供更好的结果
总而言之，本文做出了以下贡献：
$\bullet$ We propose a non-parametric approach for perform-
ing SSL to incorporate unlabeled real-world data
into the training process.
$\bullet$ The proposed method consists of modeling the in-
termediate latent space in the network using GP, which is then used to create the pseudo GT for the unlabeled data. The pseudo GT is further used to supervise the network at the intermediate level for the unlabeled data.
$\bullet$ Through extensive experiments on different
datasets, we show that the proposed method is able to achieve on-par performance with limited training data as compared to network trained with full training data. Additionally, we also show that using the proposed GP-based SSL framework to incorporate the unlabeled real-world data into the training process results in better performance as compared to the existing methods.
## 2. Related work
图像脱轨是低级计算机视觉社区中广泛研究的主题。已经开发了几种方法来解决这个问题。这些方法分为两大类：基于单个图像的技术[58,9,23,51,60,54]和基于视频的技术[59,11,44,27,17,26] 。这些方法的综合分析可以在[19]中找到。
基于单幅图像的技术通常使用单幅图像作为输入， 并尝试从中重建无雨图像。早期的单幅图像降额方法要么采用稀疏性[56, 29] 和低秩表示[3]等先验方法， 要么使用字典学习 [2] 和 GMM [42] 等技术对图像补丁进行建模。最近，基于深度学习的技术由于其从配对数据中有效学习的能力而获得了突出的地位。基于视频
|
[
[
0.52392578125,
0.0947265625,
0.88818359375,
0.0947265625,
0.88818359375,
0.10479736328125,
0.52392578125,
0.10479736328125
],
[
0.5048828125,
0.10797119140625,
0.88818359375,
0.10797119140625,
0.88818359375,
0.121826171875,
0.5048828125,
0.121826171875
],
[
0.50634765625,
0.125,
0.88623046875,
0.125,
0.88623046875,
0.1351318359375,
0.50634765625,
0.1351318359375
],
[
0.5048828125,
0.136962890625,
0.890625,
0.1383056640625,
0.890625,
0.15283203125,
0.5048828125,
0.1514892578125
],
[
0.50634765625,
0.1546630859375,
0.8896484375,
0.1546630859375,
0.8896484375,
0.1654052734375,
0.50634765625,
0.1654052734375
],
[
0.50341796875,
0.1673583984375,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.50341796875,
0.18115234375
],
[
0.50341796875,
0.1817626953125,
0.8896484375,
0.18310546875,
0.8896484375,
0.1962890625,
0.50341796875,
0.1956787109375
],
[
0.50634765625,
0.199462890625,
0.8896484375,
0.199462890625,
0.8896484375,
0.210205078125,
0.50634765625,
0.210205078125
],
[
0.5048828125,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.5048828125,
0.2266845703125
],
[
0.5048828125,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.5048828125,
0.242431640625
],
[
0.136474609375,
0.2412109375,
0.158447265625,
0.2412109375,
0.158447265625,
0.255615234375,
0.136474609375,
0.255615234375
],
[
0.2646484375,
0.2412109375,
0.285888671875,
0.2412109375,
0.285888671875,
0.255126953125,
0.2646484375,
0.255126953125
],
[
0.39208984375,
0.240478515625,
0.41259765625,
0.240478515625,
0.41259765625,
0.255615234375,
0.39208984375,
0.255615234375
],
[
0.5048828125,
0.24365234375,
0.8896484375,
0.24365234375,
0.8896484375,
0.2568359375,
0.5048828125,
0.2568359375
],
[
0.5048828125,
0.2587890625,
0.890625,
0.2587890625,
0.890625,
0.272705078125,
0.5048828125,
0.272705078125
],
[
0.08416748046875,
0.27978515625,
0.465576171875,
0.27978515625,
0.465576171875,
0.290283203125,
0.08416748046875,
0.290283203125
],
[
0.50634765625,
0.27587890625,
0.8896484375,
0.27587890625,
0.8896484375,
0.285888671875,
0.50634765625,
0.285888671875
],
[
0.080078125,
0.290283203125,
0.46728515625,
0.291748046875,
0.46728515625,
0.30615234375,
0.080078125,
0.304931640625
],
[
0.50634765625,
0.291748046875,
0.88818359375,
0.291748046875,
0.88818359375,
0.3017578125,
0.50634765625,
0.3017578125
],
[
0.08087158203125,
0.304931640625,
0.390625,
0.304931640625,
0.390625,
0.318115234375,
0.08087158203125,
0.318115234375
],
[
0.5048828125,
0.3037109375,
0.890625,
0.3037109375,
0.890625,
0.317626953125,
0.5048828125,
0.317626953125
],
[
0.5048828125,
0.3193359375,
0.88818359375,
0.3193359375,
0.88818359375,
0.333251953125,
0.5048828125,
0.333251953125
],
[
0.5048828125,
0.333984375,
0.890625,
0.333984375,
0.890625,
0.348388671875,
0.5048828125,
0.348388671875
],
[
0.0849609375,
0.354248046875,
0.46728515625,
0.354248046875,
0.46728515625,
0.3642578125,
0.0849609375,
0.3642578125
],
[
0.50732421875,
0.35107421875,
0.8896484375,
0.35107421875,
0.8896484375,
0.361083984375,
0.50732421875,
0.361083984375
],
[
0.08087158203125,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.380615234375,
0.08087158203125,
0.380615234375
],
[
0.50341796875,
0.3642578125,
0.890625,
0.3642578125,
0.890625,
0.37744140625,
0.50341796875,
0.37744140625
],
[
0.08251953125,
0.380615234375,
0.465576171875,
0.380615234375,
0.465576171875,
0.39453125,
0.08251953125,
0.39453125
],
[
0.5048828125,
0.37890625,
0.890625,
0.37890625,
0.890625,
0.392578125,
0.5048828125,
0.392578125
],
[
0.08087158203125,
0.397705078125,
0.468017578125,
0.397705078125,
0.468017578125,
0.410888671875,
0.08087158203125,
0.410888671875
],
[
0.50244140625,
0.39453125,
0.81201171875,
0.39453125,
0.81201171875,
0.40771484375,
0.50244140625,
0.40771484375
],
[
0.08087158203125,
0.412353515625,
0.465576171875,
0.412353515625,
0.465576171875,
0.425537109375,
0.08087158203125,
0.425537109375
],
[
0.52392578125,
0.410400390625,
0.75634765625,
0.410400390625,
0.75634765625,
0.423583984375,
0.52392578125,
0.423583984375
],
[
0.08087158203125,
0.427978515625,
0.465576171875,
0.427978515625,
0.465576171875,
0.44140625,
0.08087158203125,
0.44140625
],
[
0.50634765625,
0.4248046875,
0.88818359375,
0.4248046875,
0.88818359375,
0.439453125,
0.50634765625,
0.439453125
],
[
0.08251953125,
0.443115234375,
0.146240234375,
0.443115234375,
0.146240234375,
0.455078125,
0.08251953125,
0.455078125
],
[
0.51953125,
0.439453125,
0.890625,
0.438232421875,
0.890625,
0.45263671875,
0.51953125,
0.453857421875
],
[
0.52294921875,
0.453857421875,
0.705078125,
0.455078125,
0.705078125,
0.4697265625,
0.52294921875,
0.468505859375
],
[
0.1004638671875,
0.460205078125,
0.46728515625,
0.4609375,
0.46728515625,
0.474853515625,
0.1004638671875,
0.473388671875
],
[
0.08251953125,
0.47607421875,
0.468017578125,
0.47607421875,
0.468017578125,
0.4892578125,
0.08251953125,
0.4892578125
],
[
0.50244140625,
0.47021484375,
0.8896484375,
0.47021484375,
0.8896484375,
0.484130859375,
0.50244140625,
0.484130859375
],
[
0.52294921875,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.52294921875,
0.498779296875
],
[
0.08251953125,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.08251953125,
0.5048828125
],
[
0.08087158203125,
0.5048828125,
0.1396484375,
0.5068359375,
0.1380615234375,
0.521484375,
0.080078125,
0.51953125
],
[
0.525390625,
0.50048828125,
0.890625,
0.50048828125,
0.890625,
0.5146484375,
0.525390625,
0.5146484375
],
[
0.52783203125,
0.517578125,
0.8896484375,
0.517578125,
0.8896484375,
0.52783203125,
0.52783203125,
0.52783203125
],
[
0.1004638671875,
0.5244140625,
0.46728515625,
0.5244140625,
0.46728515625,
0.53857421875,
0.1004638671875,
0.53857421875
],
[
0.52392578125,
0.53076171875,
0.890625,
0.53076171875,
0.890625,
0.54443359375,
0.52392578125,
0.54443359375
],
[
0.08087158203125,
0.5390625,
0.468017578125,
0.5390625,
0.468017578125,
0.55322265625,
0.08087158203125,
0.55322265625
],
[
0.52392578125,
0.54541015625,
0.66845703125,
0.54541015625,
0.66845703125,
0.5595703125,
0.52392578125,
0.5595703125
],
[
0.08087158203125,
0.5546875,
0.468017578125,
0.5546875,
0.468017578125,
0.568359375,
0.08087158203125,
0.568359375
],
[
0.08087158203125,
0.568359375,
0.468017578125,
0.5693359375,
0.468017578125,
0.58349609375,
0.08087158203125,
0.58203125
],
[
0.50634765625,
0.5625,
0.890625,
0.5625,
0.890625,
0.57275390625,
0.50634765625,
0.57275390625
],
[
0.52294921875,
0.57470703125,
0.89208984375,
0.57470703125,
0.89208984375,
0.58837890625,
0.52294921875,
0.58837890625
],
[
0.08251953125,
0.58544921875,
0.46728515625,
0.58544921875,
0.46728515625,
0.5986328125,
0.08251953125,
0.5986328125
],
[
0.52392578125,
0.59130859375,
0.890625,
0.59130859375,
0.890625,
0.60498046875,
0.52392578125,
0.60498046875
],
[
0.08251953125,
0.60107421875,
0.46728515625,
0.60107421875,
0.46728515625,
0.6142578125,
0.08251953125,
0.6142578125
],
[
0.525390625,
0.60595703125,
0.890625,
0.60595703125,
0.890625,
0.619140625,
0.525390625,
0.619140625
],
[
0.08416748046875,
0.6171875,
0.465576171875,
0.6171875,
0.465576171875,
0.62744140625,
0.08416748046875,
0.62744140625
],
[
0.52392578125,
0.6220703125,
0.89208984375,
0.6220703125,
0.89208984375,
0.63525390625,
0.52392578125,
0.63525390625
],
[
0.080078125,
0.62890625,
0.46728515625,
0.6298828125,
0.46728515625,
0.64453125,
0.080078125,
0.64306640625
],
[
0.52392578125,
0.63623046875,
0.890625,
0.6337890625,
0.890625,
0.6484375,
0.52392578125,
0.65087890625
],
[
0.08087158203125,
0.64453125,
0.465576171875,
0.64453125,
0.465576171875,
0.65771484375,
0.08087158203125,
0.65771484375
],
[
0.52294921875,
0.6513671875,
0.890625,
0.65087890625,
0.890625,
0.6640625,
0.52294921875,
0.66552734375
],
[
0.08251953125,
0.6591796875,
0.46728515625,
0.6591796875,
0.46728515625,
0.67236328125,
0.08251953125,
0.67236328125
],
[
0.525390625,
0.66748046875,
0.890625,
0.66748046875,
0.890625,
0.68115234375,
0.525390625,
0.68115234375
],
[
0.08251953125,
0.67529296875,
0.46728515625,
0.67529296875,
0.46728515625,
0.689453125,
0.08251953125,
0.689453125
],
[
0.52294921875,
0.681640625,
0.76953125,
0.68115234375,
0.76953125,
0.6943359375,
0.52294921875,
0.69580078125
],
[
0.08251953125,
0.6923828125,
0.465576171875,
0.6923828125,
0.465576171875,
0.70263671875,
0.08251953125,
0.70263671875
],
[
0.08087158203125,
0.70556640625,
0.46728515625,
0.70556640625,
0.46728515625,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.50244140625,
0.7109375,
0.64453125,
0.7109375,
0.64453125,
0.724609375,
0.50244140625,
0.724609375
],
[
0.08251953125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.08251953125,
0.734375
],
[
0.08251953125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.08251953125,
0.75
],
[
0.525390625,
0.73779296875,
0.88623046875,
0.73779296875,
0.88623046875,
0.74853515625,
0.525390625,
0.74853515625
],
[
0.08251953125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.08251953125,
0.76513671875
],
[
0.5048828125,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.5048828125,
0.76513671875
],
[
0.08416748046875,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.08416748046875,
0.77978515625
],
[
0.5048828125,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.5048828125,
0.77978515625
],
[
0.08087158203125,
0.7822265625,
0.46728515625,
0.7822265625,
0.46728515625,
0.79541015625,
0.08087158203125,
0.79541015625
],
[
0.5048828125,
0.7822265625,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.5048828125,
0.79541015625
],
[
0.08251953125,
0.796875,
0.46728515625,
0.796875,
0.46728515625,
0.81103515625,
0.08251953125,
0.81103515625
],
[
0.50244140625,
0.79541015625,
0.7060546875,
0.79443359375,
0.7060546875,
0.8095703125,
0.50244140625,
0.81005859375
],
[
0.08087158203125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.52392578125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.52392578125,
0.82470703125
],
[
0.08251953125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.5048828125,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.08416748046875,
0.84326171875,
0.465576171875,
0.84326171875,
0.465576171875,
0.85400390625,
0.08416748046875,
0.85400390625
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.08251953125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08251953125,
0.87060546875
],
[
0.5048828125,
0.85693359375,
0.88818359375,
0.85693359375,
0.88818359375,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.080078125,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.080078125,
0.88525390625
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.08251953125,
0.88818359375,
0.1805419921875,
0.88818359375,
0.1805419921875,
0.8994140625,
0.08251953125,
0.8994140625
],
[
0.5048828125,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5048828125,
0.90087890625
]
] |
[
"受Weietal.49的启发，我们解决了将未标记的现",
"实世界图像合并到训练过程中的问题，以便通过克服",
"其方法的缺点来更好地泛化。与49相比，我们使用",
"非参数方法为未标记的数据生成监督。具体来说，我们",
"提出了一种基于高斯过程（GP）的半监督学习（SSL",
"框架，该框架涉及对标记和未标记数据进行迭代训练。",
"标记学习阶段涉及使用预测和地面实况之间的均方误",
"差对标记数据进行训练。此外，输入（来自标记的数据",
"集）被投影到潜在空间，然后使用GP建模。在未标记",
"的训练阶段，我们使用在标记的训练阶段早期建模的",
"(a)",
"(b)",
"(c)",
"GP为未标记的输人生成伪GT。然后，此伪GT用于",
"监督未标记数据的中间潜在空间。伪GT的创建基于",
"Figure 2.Derainedresults.(a)Input rainy images.(b)",
"以下假设：未标记的图像在投影到潜在空间时可以表",
"SSIR output [49] . (c) Our output. It can be observed that",
"示为标记数据特征的加权组合，其中权重是使用核函",
"the proposed method achieves better deraining",
"数确定的。这些权重表示用于制定未标记数据点的标",
"记数据点的不确定性。因此，最小化未标记数据投影",
"和伪GT之间的误差可以减少方差，从而使网络权重",
"个阶段的功能依赖关系。虽然这些方法在获得高质量",
"自动适应未标记数据的域。Fig。1演示了使用建议的",
"的出轨图像方面取得了卓越的性能，但由于它们是完",
"框架利用未标记数据的结果。Fig。2将该方法的结果",
"全监督的网络，并且只能利用完全标记的训练数据，因",
"与SIRR「49进行了比较。可以清楚地看到，与SIRR",
"此它们固有地受到限制。然而，如前所述，获得标记的",
"「49】相比，我们的方法能够提供更好的结果",
"现实世界的训练数据是相当具有挑战性的，因此，现有",
"总而言之，本文做出了以下贡献：",
"的方法通常只在综合生成的雨数据集[57，51]上训练他",
": We propose a non-parametric approach for perform-",
"们的网络",
"ing SSL to incorporate unlabeled real-world data",
"into the training process.",
"使用合成数据集会导致实际图像的性能不佳，这通",
"常是由于合成图像和雨天图像之间的分布偏移「41。尽",
". The proposed method consists of modeling the in-",
"termediate latent space in the network using GP,",
"管性能上存在差距，但这个问题在文献中仍然相对未",
"被探索。",
"which is then used to create the pseudo GT for the",
"unlabeled data.ThepseudoGTisfurtherusedto",
"最近，Wei etal.【49]提出了一种半监督学习框架",
"supervise the network at the intermediate level for",
"（SIRR），在该框架中，他们同时从标记和未标记的数",
"the unlabeled data.",
"据中学习，以实现图像脱轨。对于标记数据的训练，他",
"们使用传统的预测和地面实况（GT）之间的平均绝对",
".Throughextensiveexperimentsondifferent",
"datasets, we show that the proposed method is",
"误差损失。对于未标记的数据，他们通过施加在高斯",
"able to achieve on-par performance with limited",
"混合模型（GMM）上的似然项来建模雨残差（输人",
"training data as compared to network trained with",
"和输出之间的差）此外：它们通过最小化它们之间的",
"full training data. Additionally, we also show that",
"Kullback-Leibler（KL）差异来增强合成雨的分布更接",
"using the proposed GP-based SSL framework to",
"近真实雨的分布的一致性。。这是在半监督学习框架中",
"incorporate the unlabeled real-world data into the",
"制定图像脱轨任务的第一种方法，该框架可以利用未",
"training process results in better performance as",
"标记的现实世界图像来提高泛化能力。尽管该方法取",
"compared to the existing methods",
"得了有希望的结果，但它具有以下缺点：（i）由于雨残",
"差的多模态性质，作者假设可以使用GMM对其进行",
"2. Related work",
"建模。只有当实际残差用于计算GMM参数时，才会",
"出现这种情况。然而，作者在训练送代中使用真实世",
"图像脱轨是低级计算机视觉社区中广泛研究的主题",
"界（未标记）图像的预测雨残差来建模GMM。然后使",
"已经开发了几种方法来解决这个问题。这些方法分为",
"用相同的模型来计算后续迭代中（未标记图像的）预测",
"两大类：基于单个图像的技术[58，9,23.51，60，541和",
"残差的可能性。因此，如果在初始迭代集期间学习的",
"基于视频的技术[59，11，44，27，17，26]。这些方法的",
"GMM参数不准确，这很可能是在训练的早期阶段，它",
"综合分析可以在[19]中找到",
"将导致次优性能。（二）利用KL散度的目的是使合成",
"基于单幅图像的技术通常使用单幅图像作为输人，",
"雨分布更接近真实雨分布。如前所述，在训练的早期阶",
"并尝试从中重建无雨图像。早期的单幅图像降额方法",
"段，实际降雨残差的预测将不准确，因此，最小化两个",
"要么采用稀疏性[56，29]和低秩表示[3】等先验方法，",
"分布之间的差异可能不合适。（ii）使用GMM对雨残",
"要么使用字典学习[2]和GMM【42]等技术对图像补丁",
"差进行建模需要选择混合成分的数量，从而使模型对",
"进行建模。最近，基于深度学习的技术由于其从配对数",
"此类选择敏感",
"据中有效学习的能力而获得了突出的地位。基于视频"
] |
{
"arxiv_id": "2006.05580",
"language": "zh",
"page_idx": 1
}
|
||
2006.05580-zh-2
|
2006.05580-zh-2.jpg
|
的脱轨技术通常通过实施帧之间的时间一致性等约束
来利用附加信息
在这项工作中，我们专注于基于单个图像的脱轨，该脱轨专门利用了其他未标记的现实世界数据。Fu 等人。[7]提出了一种基于卷积神经网络(CNN)的方法，在该方法中，他们学习了从雨天图像到干净图像的映射。Zhang 等人。[57]引人了生成对抗网络 (GAN)进行图像除雨，从而实现了高质量的重建。Fu 等人。[9] 提出了一个端到端的 CNN, 称为深度细节网络，它直接减少了从输入到输出的映射范围。Zhang 和 Patel [58] 提出了一种密度感知的多流密集连接的 CNN,用于联合雨密度估计和降额。他们的网络首先根据雨水密度对输入图像进行分类，然后根据预测的雨水密度采用适当的网络从输入图像中去除雨纹。Wang 等人。[48] 采用基于估计图像的不同频率细节的分层方法来获得脱轨图像。钱 et al.[38] 提出了一个 GAN 来清除相机镜头上的雨滴。为了使网络能够专注于重要区域，他们将注意力图注人到生成和判别网络中。Li 等人。[21] 提出了一种基于卷积和递归神经网络的单幅图像降额方法， 该方法结合了上下文信息。最近，Li et al.[18] 和 Hu et al.[13] 结合了深度信息，以提高脱轨质量。Yasarla 和 Patel [53] 采用不确定性机制来学习基于位置的预测残差置信度。Wang et al.[47] 提出了一种空间注意力网络，该网络以局部到全局的方式去除雨水
# 3. Background
在本节中，我们提供了问题陈述的表述，然后简要
介绍了 GP 中的关键概念。
## $3. 1. \textbf{ Single image de- raining}$
通常是 XMATHX 的目标
# 3.2. Single image de-raining
的$y$的逆问题，其中$y$被认为是干净图像$(y)$和雨分量$(r$ ), i.e. ,<NEX。这可以通过学习以下函数来实现：(i) 直接从雨天图像映射到干净的图像[5,8,60,56] ,或者(ii)从雨天图像中提取雨水分量，然后可以从雨天图像中减去雨水分量以获得干净的图像[9,58,22] 。我们遵循从雨天图像估计雨分量的第二种方法。
# 3.3. Semi-supervised learning
在半监督学习中，我们给出了从未知联合分布$p(x,y)$采样的输入-目标对$(\{x,y\}\in\mathcal{D}_{\mathcal{L}})$的标记数据集以及从$p(x)$采样的未标记输入数据点$x\in\mathcal{D}_\mathcal{U}$ 。目标是学习由$\theta$参数化的函数$f(x|\theta)$ ,该函数可以准确地预测$p(x)$中看不见的样本的正确目标$y$ 。通过利用标记和未标记的数据集来学习参数$\theta$ 。由于标记的数据集由输入-目标对组成，因此通常使用监督损失函数 (例如平均绝对误差或交叉熵)来训练网络。未标记的数据点形式$\mathcal{D}_u$用于通过不同的技术(例如执行一致正则化[15] ,虚拟对抗性训练 [34] 或伪标记 [16] ),
使用有关$p(x)$结构的信息(如数据流形[35]的形状) 来扩展$f(x|\theta)$在[49]之后，我们采用半监督学习框架来利用未标记的现实世界数据来获得更好的泛化性能。具体来说，我们将由输入-目标对组成的综合生成的雨数据集视为标记数据集$\mathcal{D}_\mathcal{L}$,将真实世界的未标记图像视为未标记数据集$\mathcal{D}_U$ 。与 [49] 相比，我们遵循伪标记的方法来利用未标记的数据。
## 3.4. Gaussian processes
高斯过程 (GP) $f(v)$是随机变量的无限集合，其中任何有限子集都是联合高斯分布的。GP 完全由其均值函数和协方差函数指定，定义如下
$$m(v)=\mathbb{E}[f(v)],$$
(1)
$$K\left(v,v'\right)=\mathbb{E}\left[\left(f(v)-m(v)\right)\left(f\left(v'\right)-m\left(v'\right)\right)\right],$$
(2)
其中$v,v^{\prime}\in\mathcal{V}$表示索引 GP 的可能输人。协方差矩阵由协方差函数或内核$K$构造而成，该函数表达了基础函数的某些先验平滑度概念。GP 可以表示如下
$$f(v)\sim\mathcal{GP}(m(v),K(v,v')+\sigma_{\epsilon}^{2}I).$$
(3)
其中是 I 单位矩阵，$\sigma_\epsilon^2$是加性噪声的方差。然后，函
数值的任何集合都是高斯的，如下所示
$$f(V)=\left[f\left(v_{1}\right),\ldots,f\left(v_{n}\right)\right]^{T}\sim\mathcal{N}\left(\mu,K(V,V')+\sigma_{\epsilon}^{2}I\right)$$
(4)
,如前所述，由 GP 定义的均值向量和协方差矩阵。为
了在未标记的点进行预测，可以通过对观察到的数据进行条件处理来计算封闭形式的高斯后验分布。读者参考[39]对 GP 的详细审查。
# 4. Proposed method
如图。3,所提出的方法由基于 UNet 结构 [43] 的CNN 组成，其中每个块都使用 Res2Block$\underline{\mathrm{E}}[10]$构造。网络体系结构的详细信息在补充材料中提供。总之，网络由编码器($h(x,\theta_{enc})$)和解码器($g(z,\theta_dec)$)组成。这里，编码器和解码器分别由$\theta_enc$和$\theta_{dec}$参数化。此外，$x$是网络的输入，然后由编码器将其映射到潜在向量$z$。在我们的例子中，$x$是我们要从中删除雨纹的雨天图像。然后将潜在向量馈送到解码器以产生输出$r$ , 在我们的例子中是雨条纹。然后从雨图像( $x$ )中减去雨纹分量，以生成干净的图像($y$)、i.e. .
(5)
$$y=x-r,$$
,其中
(6)
$$r=g(h(x,\theta_{enc}),\theta_{dec}).$$
在我们的问题公式中，训练数据集是$\mathcal{D}=\mathcal{D}_\mathcal{L}\cup\mathcal{D}_\mathcal{U}$ ,其中$\mathcal{D}_\mathcal{L}=\{x_l^i,y_l^i\}_{i=1}^{N_l}$是由$N_l$样本组成的标记训练集，$\mathcal{D}_U=\{x_u^i\}_{i=1}^{N_u}$是由$N_u$未标记样本组成的集。对于本文的其余部分，$\mathcal{D}_\mathcal{L}$是指标记的“合成”数据集， $\mathcal{D}_U$是指未标记的“真实世界”数据集，除非另有说明。
|
[
[
0.08251953125,
0.0947265625,
0.46728515625,
0.0947265625,
0.46728515625,
0.10479736328125,
0.08251953125,
0.10479736328125
],
[
0.50341796875,
0.09344482421875,
0.890625,
0.09344482421875,
0.890625,
0.106689453125,
0.50341796875,
0.106689453125
],
[
0.08087158203125,
0.10919189453125,
0.1953125,
0.10919189453125,
0.1953125,
0.12249755859375,
0.08087158203125,
0.12249755859375
],
[
0.50634765625,
0.10986328125,
0.88818359375,
0.10986328125,
0.88818359375,
0.12060546875,
0.50634765625,
0.12060546875
],
[
0.101318359375,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.101318359375,
0.1351318359375
],
[
0.50244140625,
0.12249755859375,
0.88818359375,
0.12371826171875,
0.88818359375,
0.136962890625,
0.50244140625,
0.1363525390625
],
[
0.08087158203125,
0.1383056640625,
0.4697265625,
0.1383056640625,
0.4697265625,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50341796875,
0.1383056640625,
0.8896484375,
0.1383056640625,
0.8896484375,
0.1514892578125,
0.50341796875,
0.1514892578125
],
[
0.08087158203125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.08087158203125,
0.1673583984375
],
[
0.50341796875,
0.154052734375,
0.88818359375,
0.154052734375,
0.88818359375,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.08416748046875,
0.17041015625,
0.46728515625,
0.17041015625,
0.46728515625,
0.18115234375,
0.08416748046875,
0.18115234375
],
[
0.50341796875,
0.1685791015625,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.08087158203125,
0.184326171875,
0.468017578125,
0.184326171875,
0.468017578125,
0.1976318359375,
0.08087158203125,
0.1976318359375
],
[
0.50341796875,
0.184326171875,
0.73876953125,
0.184326171875,
0.73876953125,
0.1944580078125,
0.50341796875,
0.1944580078125
],
[
0.08087158203125,
0.1988525390625,
0.46728515625,
0.199462890625,
0.46728515625,
0.21337890625,
0.08087158203125,
0.212158203125
],
[
0.50244140625,
0.20703125,
0.689453125,
0.2083740234375,
0.689453125,
0.222900390625,
0.50244140625,
0.2215576171875
],
[
0.08416748046875,
0.2139892578125,
0.468017578125,
0.2139892578125,
0.468017578125,
0.2279052734375,
0.08416748046875,
0.2279052734375
],
[
0.08087158203125,
0.2279052734375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.24365234375,
0.08087158203125,
0.2412109375
],
[
0.52294921875,
0.2310791015625,
0.8896484375,
0.2310791015625,
0.8896484375,
0.2442626953125,
0.52294921875,
0.2442626953125
],
[
0.08087158203125,
0.2442626953125,
0.468017578125,
0.2442626953125,
0.468017578125,
0.25830078125,
0.08087158203125,
0.25830078125
],
[
0.50341796875,
0.246826171875,
0.890625,
0.246826171875,
0.890625,
0.260009765625,
0.50341796875,
0.260009765625
],
[
0.08087158203125,
0.260009765625,
0.468017578125,
0.260009765625,
0.468017578125,
0.2734375,
0.08087158203125,
0.2734375
],
[
0.50341796875,
0.261962890625,
0.74853515625,
0.261962890625,
0.74853515625,
0.27587890625,
0.50341796875,
0.27587890625
],
[
0.08251953125,
0.276611328125,
0.465576171875,
0.276611328125,
0.465576171875,
0.287353515625,
0.08251953125,
0.287353515625
],
[
0.08251953125,
0.2890625,
0.46728515625,
0.2890625,
0.46728515625,
0.302978515625,
0.08251953125,
0.302978515625
],
[
0.63232421875,
0.2890625,
0.74658203125,
0.2890625,
0.74658203125,
0.302978515625,
0.63232421875,
0.302978515625
],
[
0.86767578125,
0.290283203125,
0.890625,
0.290283203125,
0.890625,
0.3037109375,
0.86767578125,
0.3037109375
],
[
0.08251953125,
0.304931640625,
0.468017578125,
0.304931640625,
0.468017578125,
0.318115234375,
0.08251953125,
0.318115234375
],
[
0.08087158203125,
0.3193359375,
0.46728515625,
0.3193359375,
0.46728515625,
0.333251953125,
0.08087158203125,
0.333251953125
],
[
0.5146484375,
0.31494140625,
0.8896484375,
0.31640625,
0.8896484375,
0.330810546875,
0.5146484375,
0.330078125
],
[
0.08087158203125,
0.335205078125,
0.46728515625,
0.335205078125,
0.46728515625,
0.348388671875,
0.08087158203125,
0.348388671875
],
[
0.50341796875,
0.339111328125,
0.8896484375,
0.339111328125,
0.8896484375,
0.352783203125,
0.50341796875,
0.352783203125
],
[
0.08251953125,
0.3515625,
0.465576171875,
0.3515625,
0.465576171875,
0.3623046875,
0.08251953125,
0.3623046875
],
[
0.50341796875,
0.354248046875,
0.890625,
0.354248046875,
0.890625,
0.367431640625,
0.50341796875,
0.367431640625
],
[
0.08251953125,
0.3642578125,
0.46728515625,
0.365478515625,
0.46728515625,
0.37890625,
0.08251953125,
0.37744140625
],
[
0.50341796875,
0.36865234375,
0.8408203125,
0.36865234375,
0.8408203125,
0.3818359375,
0.50341796875,
0.3818359375
],
[
0.08251953125,
0.380126953125,
0.46728515625,
0.380126953125,
0.46728515625,
0.393310546875,
0.08251953125,
0.393310546875
],
[
0.08087158203125,
0.395751953125,
0.46728515625,
0.395751953125,
0.46728515625,
0.4091796875,
0.08087158203125,
0.4091796875
],
[
0.576171875,
0.39453125,
0.89208984375,
0.395751953125,
0.89208984375,
0.41357421875,
0.576171875,
0.412353515625
],
[
0.08087158203125,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.08087158203125,
0.425537109375,
0.468017578125,
0.425537109375,
0.468017578125,
0.439453125,
0.08087158203125,
0.439453125
],
[
0.5048828125,
0.4248046875,
0.8896484375,
0.4248046875,
0.8896484375,
0.438232421875,
0.5048828125,
0.438232421875
],
[
0.080078125,
0.43994140625,
0.390625,
0.44140625,
0.390625,
0.455078125,
0.080078125,
0.453857421875
],
[
0.50341796875,
0.438232421875,
0.78125,
0.439453125,
0.78125,
0.453857421875,
0.50341796875,
0.45263671875
],
[
0.080078125,
0.4716796875,
0.2115478515625,
0.4716796875,
0.2115478515625,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.50732421875,
0.46728515625,
0.8837890625,
0.466552734375,
0.8837890625,
0.48095703125,
0.50732421875,
0.481689453125
],
[
0.87109375,
0.48291015625,
0.8896484375,
0.48291015625,
0.8896484375,
0.49560546875,
0.87109375,
0.49560546875
],
[
0.1004638671875,
0.496826171875,
0.468017578125,
0.496826171875,
0.468017578125,
0.51025390625,
0.1004638671875,
0.51025390625
],
[
0.5048828125,
0.49755859375,
0.890625,
0.49755859375,
0.890625,
0.51123046875,
0.5048828125,
0.51123046875
],
[
0.08087158203125,
0.51220703125,
0.271240234375,
0.51220703125,
0.271240234375,
0.52587890625,
0.08087158203125,
0.52587890625
],
[
0.50732421875,
0.51318359375,
0.890625,
0.51318359375,
0.890625,
0.5263671875,
0.50732421875,
0.5263671875
],
[
0.5048828125,
0.5263671875,
0.890625,
0.5263671875,
0.890625,
0.54052734375,
0.5048828125,
0.54052734375
],
[
0.50244140625,
0.54150390625,
0.71337890625,
0.54248046875,
0.71337890625,
0.556640625,
0.50244140625,
0.55615234375
],
[
0.07843017578125,
0.5517578125,
0.301513671875,
0.55322265625,
0.301513671875,
0.5673828125,
0.07843017578125,
0.56640625
],
[
0.1004638671875,
0.5771484375,
0.283447265625,
0.5771484375,
0.283447265625,
0.59033203125,
0.1004638671875,
0.59033203125
],
[
0.50244140625,
0.5712890625,
0.6806640625,
0.57080078125,
0.6806640625,
0.58544921875,
0.50244140625,
0.5859375
],
[
0.07843017578125,
0.6015625,
0.301513671875,
0.60302734375,
0.301513671875,
0.6171875,
0.07843017578125,
0.6162109375
],
[
0.52392578125,
0.5966796875,
0.890625,
0.5966796875,
0.890625,
0.60986328125,
0.52392578125,
0.60986328125
],
[
0.5048828125,
0.6123046875,
0.890625,
0.6123046875,
0.890625,
0.62548828125,
0.5048828125,
0.62548828125
],
[
0.1004638671875,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.1004638671875,
0.64013671875
],
[
0.50634765625,
0.62890625,
0.88818359375,
0.62890625,
0.88818359375,
0.6396484375,
0.50634765625,
0.6396484375
],
[
0.08251953125,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.08251953125,
0.65478515625
],
[
0.50341796875,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65576171875,
0.50341796875,
0.65576171875
],
[
0.08251953125,
0.65771484375,
0.465576171875,
0.65771484375,
0.465576171875,
0.66845703125,
0.08251953125,
0.66845703125
],
[
0.50341796875,
0.65771484375,
0.89208984375,
0.65771484375,
0.89208984375,
0.671875,
0.50341796875,
0.671875
],
[
0.0931396484375,
0.671875,
0.46728515625,
0.671875,
0.46728515625,
0.68505859375,
0.0931396484375,
0.68505859375
],
[
0.50634765625,
0.6748046875,
0.88818359375,
0.6748046875,
0.88818359375,
0.68505859375,
0.50634765625,
0.68505859375
],
[
0.08087158203125,
0.68603515625,
0.468017578125,
0.68603515625,
0.468017578125,
0.69970703125,
0.08087158203125,
0.69970703125
],
[
0.5048828125,
0.6875,
0.890625,
0.6875,
0.890625,
0.70068359375,
0.5048828125,
0.70068359375
],
[
0.080078125,
0.70068359375,
0.4306640625,
0.70068359375,
0.4306640625,
0.7138671875,
0.080078125,
0.7138671875
],
[
0.50341796875,
0.7021484375,
0.88818359375,
0.70263671875,
0.88818359375,
0.71630859375,
0.50341796875,
0.71533203125
],
[
0.50341796875,
0.71630859375,
0.890625,
0.71728515625,
0.890625,
0.73193359375,
0.50341796875,
0.73095703125
],
[
0.07843017578125,
0.72607421875,
0.313720703125,
0.72802734375,
0.313720703125,
0.7421875,
0.07843017578125,
0.74072265625
],
[
0.50341796875,
0.73193359375,
0.8037109375,
0.73291015625,
0.8037109375,
0.7470703125,
0.50341796875,
0.74560546875
],
[
0.1004638671875,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.1004638671875,
0.76513671875
],
[
0.6611328125,
0.76318359375,
0.73388671875,
0.76318359375,
0.73388671875,
0.7744140625,
0.6611328125,
0.7744140625
],
[
0.86767578125,
0.76025390625,
0.890625,
0.76025390625,
0.890625,
0.7744140625,
0.86767578125,
0.7744140625
],
[
0.08087158203125,
0.767578125,
0.468017578125,
0.767578125,
0.468017578125,
0.78076171875,
0.08087158203125,
0.78076171875
],
[
0.08251953125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.08251953125,
0.79541015625
],
[
0.484375,
0.7890625,
0.55224609375,
0.7861328125,
0.5537109375,
0.80029296875,
0.486083984375,
0.8037109375
],
[
0.08416748046875,
0.798828125,
0.46728515625,
0.798828125,
0.46728515625,
0.8095703125,
0.08416748046875,
0.8095703125
],
[
0.61767578125,
0.8046875,
0.775390625,
0.8046875,
0.775390625,
0.81494140625,
0.61767578125,
0.81494140625
],
[
0.86767578125,
0.8037109375,
0.890625,
0.8037109375,
0.890625,
0.8173828125,
0.86767578125,
0.8173828125
],
[
0.08251953125,
0.8125,
0.46728515625,
0.8125,
0.46728515625,
0.82568359375,
0.08251953125,
0.82568359375
],
[
0.08251953125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.52392578125,
0.82568359375,
0.8896484375,
0.82568359375,
0.8896484375,
0.8388671875,
0.52392578125,
0.8388671875
],
[
0.08416748046875,
0.84326171875,
0.465576171875,
0.84326171875,
0.465576171875,
0.85400390625,
0.08416748046875,
0.85400390625
],
[
0.5185546875,
0.84033203125,
0.890625,
0.84033203125,
0.890625,
0.85498046875,
0.5185546875,
0.85498046875
],
[
0.08251953125,
0.859375,
0.46728515625,
0.859375,
0.46728515625,
0.869140625,
0.08251953125,
0.869140625
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.8720703125,
0.50341796875,
0.8720703125
],
[
0.08087158203125,
0.87255859375,
0.453369140625,
0.87255859375,
0.453369140625,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.50341796875,
0.87060546875,
0.89208984375,
0.87255859375,
0.89208984375,
0.88623046875,
0.50341796875,
0.8837890625
],
[
0.08087158203125,
0.88720703125,
0.47216796875,
0.88720703125,
0.47216796875,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"的脱轨技术通常通过实施顿之间的时间一致性等约束",
"使用有关p(a）结构的信息（如数据流形[35]的形状）",
"来利用附加信息",
"来扩展f（C①）在49之后，我们采用半监督学习框架",
"在这项工作中，我们专注于基于单个图像的脱轨，该",
"来利用未标记的现实世界数据来获得更好的泛化性能，",
"脱轨专门利用了其他未标记的现实世界数据。Fu等人。",
"具体来说，我们将由输入-目标对组成的综合生成的雨",
"[7]提出了一种基于卷积神经网络（CNN）的方法，在",
"数据集视为标记数据集Dc，将真实世界的未标记图",
"该方法中，他们学习了从雨天图像到干净图像的映射。",
"像视为未标记数据集Du。与【49]相比，我们遵循伪",
"Zhang等人。[57]引入了生成对抗网络（GAN）进行图",
"标记的方法来利用未标记的数据。",
"像除雨，从而实现了高质量的重建。Fu等人。[9]提出",
"3.4. Gaussian processes",
"了一个端到端的CNN，称为深度细节网络，它直接减",
"少了从输人到输出的映射范围。Zhang和Patel[58]提",
"高斯过程（GP）f(u）是随机变量的无限集合：其中",
"出了一种密度感知的多流密集连接的CNN，用于联合",
"任何有限子集都是联合高斯分布的。GP完全由其均值",
"雨密度估计和降额。他们的网络首先根据雨水密度对",
"函数和协方差函数指定，定义如下",
"输入图像进行分类，然后根据预测的雨水密度采用适",
"当的网络从输人图像中去除雨纹。Wang等人。[481采",
"m(u) = E[f(v)],",
"(1)",
"用基于估计图像的不同频率细节的分层方法来获得脱",
"轨图像。钱etal.[38]提出了一个GAN来清除相机镜",
"K (u, v') =E[(f(u) -m(v) (f (u) -m (u')], (2)",
"头上的雨滴。为了使网络能够专注于重要区域，他们将",
"其中U,EV表示索引GP的可能输人。协方差矩阵",
"注意力图注人到生成和判别网络中。Li等人。21提出",
"由协方差函数或内核K构造而成，该函数表达了基础",
"了一种基于卷积和递归神经网络的单幅图像降额方法，",
"函数的某些先验平滑度概念。GP可以表示如下",
"该方法结合了上下文信息。最近，Lietal.[18]和Hu",
"etal.【13]结合了深度信息，以提高脱轨质量。Yasarla",
"f(u) ~gP(m(), K(u, u) +o2I).(3)",
"和Patel[53]采用不确定性机制来学习基于位置的预测",
"残差置信度。Wangetal.[47]提出了一种空间注意力",
"其中是I单位矩阵，2是加性噪声的方差。然后，函",
"网络，该网络以局部到全局的方式去除雨水",
"数值的任何集合都是高斯的，如下所示",
"3. Background",
"f(V) =[f (u),..., f (un))T~N (μ, K(V, V')+o2I)",
"(4)",
"在本节中，我们提供了问题陈述的表述，然后简要",
"，如前所述，由GP定义的均值向量和协方差矩阵。为",
"介绍了GP中的关键概念。",
"了在未标记的点进行预测，可以通过对观察到的数据",
"进行条件处理来计算封闭形式的高斯后验分布。读者",
"参考[39]对GP的详细审查。",
"3.1. Single image de-raining",
"通常是XMATHX的目标",
"4.Proposed method",
"3.2. Single image de-raining",
"如图。3，所提出的方法由基于UNet结构「43]的",
"CNN组成，其中每个块都使用Res2BlockE[10]构造。",
"的y的逆问题，其中y被认为是干净图像（y）和雨",
"网络体系结构的详细信息在补充材料中提供。总之，网",
"分量（r），i.e.，<NEX。这可以通过学习以下函数来实",
"络由编码器（h(a,Qenc））和解码器（g（z,Qdec））组成。",
"现：（i）直接从雨天图像映射到干净的图像5，8，60.56",
"这里，编码器和解码器分别由Qenc和Qdec参数化。此",
"或者（i）从雨天图像中提取雨水分量，然后可以从",
"外，3是网络的输入，然后由编码器将其映射到潜在向",
"雨天图像中减去雨水分量以获得干净的图像[9，58，22]",
"量。在我们的例子中，是我们要从中删除雨纹的雨",
"。我们遵循从雨天图像估计雨分量的第二种方法。",
"天图像。然后将潜在向量馈送到解码器以产生输出r",
"在我们的例子中是雨条纹。然后从雨图像（&）中减去",
"3.3. Semi-supervised learning",
"雨纹分量，以生成干净的图像（y）、i.e.、",
"在半监督学习中，我们给出了从未知联合分布",
"y=a-r,",
"(5)",
"p(α,y）采样的输人-目标对（(a,y)EDc）的标记",
"数据集以及从p(a）采样的未标记输人数据点rEDu。",
"，其中",
"自标是学习由0参数化的函数f（αの），该函数可以准",
"r = g(h(r,0enc),0dec).",
"(6)",
"确地预测p(c）中看不见的样本的正确目标y。通过利",
"用标记和未标记的数据集来学习参数θ。由于标记的",
"在我们的问题公式中，训练数据集是D=DcUDu",
"数据集由输入-自标对组成，因此通常使用监督损失函",
"其中Dc=(ci,ii是由NM样本组成的标记训练",
"数（例如平均绝对误差或交叉炳）来训练网络。未标记",
"集，Du=（cul是由Nu未标记样本组成的集。对",
"的数据点形式Du用于通过不同的技术（例如执行",
"于本文的其余部分，Dc是指标记的“合成”数据集，",
"致正则化15]，虚拟对抗性训练[34]或伪标记【16]】，",
"Du是指未标记的“真实世界”数据集，除非另有说明。"
] |
{
"arxiv_id": "2006.05580",
"language": "zh",
"page_idx": 2
}
|
||
2006.05580-zh-3
|
2006.05580-zh-3.jpg
|
<FigureHere>
Figure 3. Overview of the proposed GP-based SSL framework. We leverage unlabeled data during learning. The training process consists of iterating over labeled data and unlabeled data. During the labeled training phase, we use supervised loss function consisting of $l_1$ error and perceptual loss between the prediction and targets. In the unlabeled phase, we jointly model the labeled and unlabeled latent vectors using GP to obtain the pseudo-GT for the unlabeled sample at the latent space. We use this pseudoGT for supervision.
所提出的方法的目标是通过利用标记($\mathcal{D}_\mathcal{L}$)和未标记数据集($\mathcal{D}_U$)来学习网络参数。训练过程迭代标记和未标记的数据集。通过最小化(i)在标记的训练阶段中的监督损失函数($\mathcal{L}_sup$)和(ii)在未标记的训练阶段中的无监督损失函数($\mathcal{L}_{unsup}$)来学习网络参数。对于未标记的训练阶段，我们使用 GP 公式生成伪 GT, 然后将其用于无监督损失函数。以下各节将详细介绍这两个培训阶段。
# 4.1. Labeled training phase
在此阶段，我们使用标记的数据$\mathcal{D}_{\mathcal{L}}$来学习网络参
数。具体来说，我们最小化以下监督损失函数
$$\mathcal{L}_{sup}=\mathcal{L}_{1}+\lambda_{p}\mathcal{L}_{p},$$
(7)
,其中$\lambda_p$是常数$,\mathcal{L}_1$和$\mathcal{L}_p$分别是$l_1$ -损失和感知损
失[14,55]函数。它们的定义如下：
(8)
$$\mathcal{L}_{1}=\|y_{l}^{pred}-y_{l}\|_{1},\\\mathcal{L}_{p}=\|\Phi_{VGG}(y_{l}^{pred})-\Phi_{VGG}(y_{l})\|_{2}^{2},$$
(9)
其中$y_l^{pred}=g(z,\theta_{dec})$ 是预测输出，$y_l$ 是地面实况， $z=h(x,\theta_{enc})$是中间潜在空间向量，$\Phi_VGG(\cdot)$表示预训练的 VGTHG-16MAX5 网络。有关感知损失的更多详细信息，请参阅补充材料
除了最小化损失函数外，我们还将所有标记的训练图像$x_l^i$的中间特征向量$z_l^i$存储在矩阵$F_z_l$中。这是$F_{z_l}=\{z_l^i\}_{i=1}^{N_l}$。稍后在未标记的训练阶段中使用它来为未标记的数据生成伪 GT。在我们的例子中，$z_l^i$是大小为$1\times M$的向量，其中 M=32,768 用于我们提出的方法中的网络。因此，$F_z_l$是大小为$N_l\times M$的矩阵。
# 4.2. Unlabeled training phase
在此阶段，我们利用未标记的数据$\mathcal{D}_{\mathcal{U}}$来提高泛化性能。具体来说，我们通过最小化预测的潜在向量与通过使用 GP 联合建模标记的样本图像$F_{z_l}$和$z_u^{pred}$ 的潜在空间向量而获得的伪 GT 之间的误差，在中间潜在空间提供监督。
使用 GP 的伪 GT: 训练以迭代方式进行，其中我们首先使用标记数据($\mathcal{D}_\mathcal{L}$)学习权重，然后使用未标记数据($\mathcal{D}_{\mathcal{U}}$)进行权重更新。在$\mathcal{D}_{\mathcal{L}}$上进行第一次迭代之后，我们将标记数据的潜在空间向量存储在列表$F_{z_l}$ 中。这些向量位于低维流形上。在未标记阶段，我们将未标记输入的潜在空间向量$(z_u$ )投影到标记向量$F_{z_l}=\{z_l^i\}_{i=1}^{N_l}$的空间上。也就是说，我们将与来自$\mathcal{D}_u$ 的$k^{th}$训练样本相对应的未标记的潜在空间向量$z_u^k$表示为
$$z_u^k=\sum_{i=1}^{N_l}\alpha_iz_l^i+\epsilon,$$
(10)
,其中$\alpha_i$ 是系数，而$\epsilon$ 是加性噪声 $\mathcal{N}(0,\sigma_\epsilon^2)$
使用此公式，我们可以使用 GP 联合建模标记和未标记样本的潜在空间向量的分布。调节联合分布将为未标记的样本
(11)
$$P(z_u^k|\mathcal{D}_{\mathcal{L}},F_{z_l})=\mathcal{N}(\mu_u^k,\Sigma_u^k),$$
产生以下条件多变量高斯分布，其中
$$\mu_{u}^{k}=K(z_{u}^{k},F_{z_{l}})[K(F_{z_{l}},F_{z_{l}})+\sigma_{\epsilon}^{2}I]^{-1}F_{z_{l}},$$
(12)
???pt
$$\begin{array}{c}\Sigma_u^k=K(z_u^k,z_u^k)-K(z_u^k,F_{z_l})[K(F_{z_l},F_{z_l})+\sigma_\epsilon^2I]^{-1}\\K(F_{z_l},z_u^k)+\sigma_\epsilon^2\end{array}$$
,其中$\sigma_\epsilon^2$被设置为等于 1, XMATHX\_LIN5 跟随
注意，$F_z_l$包含所有标记图像的潜在空间向量， $K(F_{z_{l}},F_{z_{l}})$是大小为$N_l\times N_l$的矩阵$,K(z_u^k,F_{z_{l}})$是大小为 1$\times N_{l}$的向量。由于以下原因，使用所有向量可能不一定是最佳的：(i)这些向量将对应于图像中的不同区域，在雨纹的含量和密度/方向方面具有广泛的差异。重要的是只考虑那些与未标记的向量相似的向量。(Ii) 使用所有向量在计算上是禁止的。因此，
|
[
[
0.371826171875,
0.0889892578125,
0.396240234375,
0.0953369140625,
0.390625,
0.10797119140625,
0.366943359375,
0.10162353515625
],
[
0.5048828125,
0.09344482421875,
0.56396484375,
0.09344482421875,
0.56396484375,
0.10479736328125,
0.5048828125,
0.10479736328125
],
[
0.5537109375,
0.08837890625,
0.89208984375,
0.091552734375,
0.89208984375,
0.10919189453125,
0.5537109375,
0.1060791015625
],
[
0.310546875,
0.106689453125,
0.366943359375,
0.106689453125,
0.366943359375,
0.117431640625,
0.310546875,
0.117431640625
],
[
0.3798828125,
0.10919189453125,
0.4306640625,
0.10919189453125,
0.4306640625,
0.11614990234375,
0.3798828125,
0.11614990234375
],
[
0.50341796875,
0.1060791015625,
0.890625,
0.10797119140625,
0.890625,
0.12249755859375,
0.50341796875,
0.12060546875
],
[
0.1070556640625,
0.1142578125,
0.160888671875,
0.1142578125,
0.160888671875,
0.125,
0.1070556640625,
0.125
],
[
0.10455322265625,
0.125,
0.16259765625,
0.125,
0.16259765625,
0.1351318359375,
0.10455322265625,
0.1351318359375
],
[
0.5048828125,
0.12249755859375,
0.890625,
0.12249755859375,
0.890625,
0.1363525390625,
0.5048828125,
0.1363525390625
],
[
0.325927734375,
0.1319580078125,
0.336669921875,
0.1319580078125,
0.336669921875,
0.1351318359375,
0.325927734375,
0.1351318359375
],
[
0.390625,
0.1300048828125,
0.414306640625,
0.1300048828125,
0.414306640625,
0.136962890625,
0.390625,
0.136962890625
],
[
0.50341796875,
0.1383056640625,
0.701171875,
0.1383056640625,
0.701171875,
0.1514892578125,
0.50341796875,
0.1514892578125
],
[
0.52294921875,
0.15283203125,
0.890625,
0.1514892578125,
0.890625,
0.1666259765625,
0.52294921875,
0.1673583984375
],
[
0.50341796875,
0.1666259765625,
0.890625,
0.1673583984375,
0.890625,
0.1817626953125,
0.50341796875,
0.18115234375
],
[
0.5048828125,
0.184326171875,
0.890625,
0.184326171875,
0.890625,
0.1988525390625,
0.5048828125,
0.1988525390625
],
[
0.50634765625,
0.199462890625,
0.890625,
0.199462890625,
0.890625,
0.21337890625,
0.50634765625,
0.21337890625
],
[
0.5048828125,
0.2139892578125,
0.8896484375,
0.2139892578125,
0.8896484375,
0.2279052734375,
0.5048828125,
0.2279052734375
],
[
0.50341796875,
0.2298583984375,
0.8837890625,
0.2298583984375,
0.8837890625,
0.24365234375,
0.50341796875,
0.24365234375
],
[
0.27294921875,
0.25390625,
0.281005859375,
0.25244140625,
0.281982421875,
0.25830078125,
0.273681640625,
0.2587890625
],
[
0.50244140625,
0.251953125,
0.734375,
0.25244140625,
0.734375,
0.270751953125,
0.50244140625,
0.26953125
],
[
0.40185546875,
0.27783203125,
0.461669921875,
0.279052734375,
0.461669921875,
0.2890625,
0.40185546875,
0.28857421875
],
[
0.52294921875,
0.27783203125,
0.890625,
0.27783203125,
0.890625,
0.291748046875,
0.52294921875,
0.291748046875
],
[
0.50341796875,
0.291748046875,
0.8896484375,
0.292236328125,
0.8896484375,
0.306884765625,
0.50341796875,
0.30615234375
],
[
0.10296630859375,
0.306884765625,
0.1666259765625,
0.306884765625,
0.1666259765625,
0.318115234375,
0.10296630859375,
0.318115234375
],
[
0.5048828125,
0.306884765625,
0.8837890625,
0.306884765625,
0.8837890625,
0.32080078125,
0.5048828125,
0.32080078125
],
[
0.1070556640625,
0.31640625,
0.1641845703125,
0.318115234375,
0.1641845703125,
0.330078125,
0.1070556640625,
0.32763671875
],
[
0.50341796875,
0.322509765625,
0.890625,
0.322509765625,
0.890625,
0.3359375,
0.50341796875,
0.3359375
],
[
0.293212890625,
0.33203125,
0.310546875,
0.33203125,
0.310546875,
0.341552734375,
0.293212890625,
0.341552734375
],
[
0.50341796875,
0.33837890625,
0.640625,
0.33837890625,
0.640625,
0.3515625,
0.50341796875,
0.3515625
],
[
0.098876953125,
0.35595703125,
0.178955078125,
0.35595703125,
0.178955078125,
0.367431640625,
0.098876953125,
0.367431640625
],
[
0.2115478515625,
0.35595703125,
0.306396484375,
0.35595703125,
0.306396484375,
0.3662109375,
0.2115478515625,
0.3662109375
],
[
0.32275390625,
0.3642578125,
0.339111328125,
0.363037109375,
0.340576171875,
0.37060546875,
0.324462890625,
0.371826171875
],
[
0.340576171875,
0.361083984375,
0.404296875,
0.361083984375,
0.404296875,
0.36865234375,
0.340576171875,
0.36865234375
],
[
0.343017578125,
0.36865234375,
0.40283203125,
0.369384765625,
0.40283203125,
0.376953125,
0.343017578125,
0.375732421875
],
[
0.52294921875,
0.3662109375,
0.890625,
0.367431640625,
0.890625,
0.3818359375,
0.52294921875,
0.380615234375
],
[
0.20751953125,
0.37255859375,
0.30224609375,
0.37255859375,
0.30224609375,
0.3837890625,
0.20751953125,
0.3837890625
],
[
0.359375,
0.3818359375,
0.398681640625,
0.3818359375,
0.398681640625,
0.392578125,
0.359375,
0.392578125
],
[
0.0931396484375,
0.371826171875,
0.1993408203125,
0.371826171875,
0.1993408203125,
0.386474609375,
0.0931396484375,
0.386474609375
],
[
0.32275390625,
0.38330078125,
0.3505859375,
0.389404296875,
0.346435546875,
0.40087890625,
0.318603515625,
0.39453125
],
[
0.5048828125,
0.38330078125,
0.890625,
0.38330078125,
0.890625,
0.396484375,
0.5048828125,
0.396484375
],
[
0.09478759765625,
0.39013671875,
0.1707763671875,
0.386474609375,
0.17236328125,
0.40087890625,
0.096435546875,
0.40478515625
],
[
0.2099609375,
0.386962890625,
0.281005859375,
0.386962890625,
0.281005859375,
0.397705078125,
0.2099609375,
0.397705078125
],
[
0.358642578125,
0.39013671875,
0.42724609375,
0.391357421875,
0.42724609375,
0.401611328125,
0.358642578125,
0.40087890625
],
[
0.2320556640625,
0.396484375,
0.28515625,
0.396484375,
0.28515625,
0.4072265625,
0.2320556640625,
0.4072265625
],
[
0.50341796875,
0.3984375,
0.890625,
0.3984375,
0.890625,
0.412353515625,
0.50341796875,
0.412353515625
],
[
0.5048828125,
0.41357421875,
0.88623046875,
0.41357421875,
0.88623046875,
0.4267578125,
0.5048828125,
0.4267578125
],
[
0.125,
0.399658203125,
0.19287109375,
0.399658203125,
0.19287109375,
0.410400390625,
0.125,
0.410400390625
],
[
0.07843017578125,
0.4267578125,
0.465576171875,
0.425537109375,
0.465576171875,
0.43994140625,
0.07843017578125,
0.44140625
],
[
0.50341796875,
0.4267578125,
0.8896484375,
0.427978515625,
0.8896484375,
0.442626953125,
0.50341796875,
0.44140625
],
[
0.08087158203125,
0.44140625,
0.468017578125,
0.44140625,
0.468017578125,
0.455078125,
0.08087158203125,
0.455078125
],
[
0.5048828125,
0.443115234375,
0.8896484375,
0.443115234375,
0.8896484375,
0.45703125,
0.5048828125,
0.45703125
],
[
0.07843017578125,
0.455078125,
0.468017578125,
0.453857421875,
0.468017578125,
0.468505859375,
0.07843017578125,
0.4697265625
],
[
0.5048828125,
0.458251953125,
0.8896484375,
0.458251953125,
0.8896484375,
0.472900390625,
0.5048828125,
0.472900390625
],
[
0.080078125,
0.468505859375,
0.468017578125,
0.468505859375,
0.468017578125,
0.481689453125,
0.080078125,
0.481689453125
],
[
0.50244140625,
0.4716796875,
0.89208984375,
0.472900390625,
0.89208984375,
0.4873046875,
0.50244140625,
0.486083984375
],
[
0.08251953125,
0.484130859375,
0.465576171875,
0.484130859375,
0.465576171875,
0.494384765625,
0.08251953125,
0.494384765625
],
[
0.50341796875,
0.4892578125,
0.53759765625,
0.4892578125,
0.53759765625,
0.501953125,
0.50341796875,
0.501953125
],
[
0.08087158203125,
0.496826171875,
0.465576171875,
0.496826171875,
0.465576171875,
0.51025390625,
0.08087158203125,
0.51025390625
],
[
0.6748046875,
0.50048828125,
0.6953125,
0.50048828125,
0.6953125,
0.5146484375,
0.6748046875,
0.5146484375
],
[
0.080078125,
0.51025390625,
0.468017578125,
0.51025390625,
0.468017578125,
0.5234375,
0.080078125,
0.5234375
],
[
0.6357421875,
0.51318359375,
0.67724609375,
0.51318359375,
0.67724609375,
0.52978515625,
0.6357421875,
0.52978515625
],
[
0.66650390625,
0.51025390625,
0.7646484375,
0.5146484375,
0.76318359375,
0.5361328125,
0.6650390625,
0.5322265625
],
[
0.859375,
0.51513671875,
0.89208984375,
0.51513671875,
0.89208984375,
0.53076171875,
0.859375,
0.53076171875
],
[
0.08087158203125,
0.52587890625,
0.46728515625,
0.52587890625,
0.46728515625,
0.5361328125,
0.08087158203125,
0.5361328125
],
[
0.67236328125,
0.53076171875,
0.69775390625,
0.53076171875,
0.69775390625,
0.54248046875,
0.67236328125,
0.54248046875
],
[
0.08087158203125,
0.53857421875,
0.468017578125,
0.53857421875,
0.468017578125,
0.5517578125,
0.08087158203125,
0.5517578125
],
[
0.080078125,
0.5517578125,
0.2115478515625,
0.5517578125,
0.2115478515625,
0.56494140625,
0.080078125,
0.56494140625
],
[
0.50634765625,
0.5517578125,
0.82421875,
0.5517578125,
0.82421875,
0.5625,
0.50634765625,
0.5625
],
[
0.52392578125,
0.56494140625,
0.8896484375,
0.56494140625,
0.8896484375,
0.5791015625,
0.52392578125,
0.5791015625
],
[
0.5048828125,
0.58056640625,
0.890625,
0.58056640625,
0.890625,
0.5947265625,
0.5048828125,
0.5947265625
],
[
0.08251953125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.08251953125,
0.60498046875
],
[
0.5048828125,
0.5966796875,
0.60107421875,
0.5966796875,
0.60107421875,
0.60791015625,
0.5048828125,
0.60791015625
],
[
0.08416748046875,
0.60791015625,
0.465576171875,
0.60791015625,
0.465576171875,
0.61865234375,
0.08416748046875,
0.61865234375
],
[
0.08087158203125,
0.62060546875,
0.46728515625,
0.619140625,
0.46728515625,
0.63330078125,
0.08087158203125,
0.6337890625
],
[
0.5986328125,
0.6220703125,
0.7958984375,
0.6220703125,
0.7958984375,
0.63623046875,
0.5986328125,
0.63623046875
],
[
0.85888671875,
0.6220703125,
0.89208984375,
0.6220703125,
0.89208984375,
0.63720703125,
0.85888671875,
0.63720703125
],
[
0.08087158203125,
0.63623046875,
0.46728515625,
0.63623046875,
0.46728515625,
0.6494140625,
0.08087158203125,
0.6494140625
],
[
0.080078125,
0.6494140625,
0.468017578125,
0.6484375,
0.468017578125,
0.66650390625,
0.080078125,
0.66748046875
],
[
0.5048828125,
0.6494140625,
0.76318359375,
0.6494140625,
0.76318359375,
0.66357421875,
0.5048828125,
0.66357421875
],
[
0.08416748046875,
0.66650390625,
0.46728515625,
0.66650390625,
0.46728515625,
0.6796875,
0.08416748046875,
0.6796875
],
[
0.5361328125,
0.67529296875,
0.83251953125,
0.67529296875,
0.83251953125,
0.689453125,
0.5361328125,
0.689453125
],
[
0.08087158203125,
0.681640625,
0.468017578125,
0.681640625,
0.468017578125,
0.69580078125,
0.08087158203125,
0.69580078125
],
[
0.85888671875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.689453125,
0.85888671875,
0.689453125
],
[
0.07843017578125,
0.69580078125,
0.2017822265625,
0.6962890625,
0.2017822265625,
0.7109375,
0.07843017578125,
0.71044921875
],
[
0.50244140625,
0.70263671875,
0.5439453125,
0.70263671875,
0.5439453125,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.080078125,
0.7197265625,
0.29736328125,
0.7216796875,
0.29736328125,
0.736328125,
0.080078125,
0.734375
],
[
0.51953125,
0.72802734375,
0.87353515625,
0.72802734375,
0.87353515625,
0.74560546875,
0.51953125,
0.74560546875
],
[
0.098876953125,
0.7421875,
0.46728515625,
0.74365234375,
0.46728515625,
0.75830078125,
0.098876953125,
0.7568359375
],
[
0.5595703125,
0.75,
0.67333984375,
0.74755859375,
0.67333984375,
0.76513671875,
0.5595703125,
0.767578125
],
[
0.08087158203125,
0.7587890625,
0.408447265625,
0.7587890625,
0.408447265625,
0.77294921875,
0.08087158203125,
0.77294921875
],
[
0.859375,
0.767578125,
0.8896484375,
0.767578125,
0.8896484375,
0.77978515625,
0.859375,
0.77978515625
],
[
0.205078125,
0.7841796875,
0.343017578125,
0.7861328125,
0.343017578125,
0.8037109375,
0.205078125,
0.8017578125
],
[
0.50341796875,
0.78076171875,
0.869140625,
0.77978515625,
0.869140625,
0.79443359375,
0.50341796875,
0.79541015625
],
[
0.44677734375,
0.787109375,
0.468017578125,
0.787109375,
0.468017578125,
0.80029296875,
0.44677734375,
0.80029296875
],
[
0.52294921875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.52294921875,
0.81005859375
],
[
0.0931396484375,
0.81298828125,
0.468017578125,
0.8125,
0.468017578125,
0.8271484375,
0.0931396484375,
0.82763671875
],
[
0.5048828125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5048828125,
0.82470703125
],
[
0.08087158203125,
0.8291015625,
0.3349609375,
0.8291015625,
0.3349609375,
0.84228515625,
0.08087158203125,
0.84228515625
],
[
0.5048828125,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.2059326171875,
0.85595703125,
0.3447265625,
0.85595703125,
0.3447265625,
0.87353515625,
0.2059326171875,
0.87353515625
],
[
0.44677734375,
0.85791015625,
0.4697265625,
0.85791015625,
0.4697265625,
0.87255859375,
0.44677734375,
0.87255859375
],
[
0.5048828125,
0.85693359375,
0.88232421875,
0.85693359375,
0.88232421875,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.1502685546875,
0.88720703125,
0.39453125,
0.88720703125,
0.39453125,
0.9013671875,
0.1502685546875,
0.9013671875
],
[
0.44677734375,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.44677734375,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"Csur",
"其中yp",
"pred=g(z,Odec）是预测输出，yl是地面实况，",
"Ground-Truth",
"Derained imac",
"z=h(，Qene）是中间潜在空间向量，ΦvGG（）表示预",
"Labeled Data",
"(Synthetic)",
"训练的VGTHG-16MAX5网络。有关感知损失的更多",
"",
"",
"详细信息，请参阅补充材料",
"除了最小化损失函数外，我们还将所有标记的训练",
"图像的中间特征向量z存储在矩阵Fz，中。这是",
"Fz=【zii=1。稍后在未标记的训练阶段中使用它来",
"为未标记的数据生成伪GT。在我们的例子中，z是大",
"小为1×M的向量，其中M=32,768用于我们提出的",
"方法中的网络。因此，Fz，是大小为Ni×M的矩阵。",
"F",
"4.2. Unlabeled training phase",
" Derained image",
"在此阶段，我们利用未标记的数据Du来提高泛化",
"性能。具体来说，我们通过最小化预测的潜在向量与",
"Unlabeled Data",
"通过使用GP联合建模标记的样本图像Fz，和zpre",
"(Real-world)",
"的潜在空间向量而获得的伪GT之间的误差，在中间",
"Cu",
"潜在空间提供监督。",
"+Labeledpath",
"Unlabeledpath",
"F",
"Labeledlaten",
"spacevetcors",
"使用GP的伪GT：训练以选代方式进行，其中我们",
"Zu.predlatent vector",
"Loss for",
"Zupseudo:pseudo-GT",
"Cunsu",
"首先使用标记数据（Dc）学习权重，然后使用未标记",
"Cup =Loss for",
"GP-Gaussian",
"unlabeled data",
"Processes",
"数据（Du）进行权重更新。在Dc上进行第一次迭代",
"之后，我们将标记数据的潜在空间向量存储在列表Fz",
"labeleddata",
"Figure 3. Overview of the proposed GP-based SSL frame-",
"中。这些向量位于低维流形上。在未标记阶段，我们",
"work. We leverage unlabeled data during learning. The",
"将未标记输人的潜在空间向量（zu）投影到标记向量",
"training process consists of iterating over labeled data and",
"Fz=【z的空间上。也就是说，我们将与来自Du",
"unlabeled data. During the labeled training phase, we use",
"的kth训练样本相对应的未标记的潜在空间向量表",
"supervised loss function consisting of li error and percep-",
"示为",
"tual loss between the prediction and targets. In the un-",
"Ni",
"labeled phase, we jointly model the labeled and unlabeled",
"2h=",
"aizi+e,",
"(10)",
"latent vectors usingGP to obtain thepseudo-GT for the",
"=1",
"unlabeled sample at the latent space. We use this pseudo-",
"GT for supervision.",
"，其中o是系数，而是加性噪声N(O.g2)",
"使用此公式，我们可以使用GP联合建模标记和未",
"标记样本的潜在空间向量的分布。调节联合分布将为",
"所提出的方法的目标是通过利用标记（Dc）和未标记",
"未标记的样本",
"数据集（Du）来学习网络参数。训练过程迭代标记和",
"未标记的数据集。通过最小化（i）在标记的训练阶段中",
"P(zhDc,Fz) =N(μk,Ek),",
"(11)",
"的监督损失函数（Lsup）和（ii）在未标记的训练阶段",
"中的无监督损失函数（Cunsup）来学习网络参数。对",
"产生以下条件多变量高斯分布，其中",
"于未标记的训练阶段，我们使用GP公式生成伪GT，",
"μk =K(zh,Fz)[K(Fz,Fz) +o2I]-1Fz",
"然后将其用于无监督损失函数。以下各节将详细介绍",
"(12)",
"这两个培训阶段。",
"???pt",
"4.1. Labeled training phase",
"h =K(zh, zh) -K(zh,Fz)[K(Fzt,Fz) +21]-1",
"在此阶段，我们使用标记的数据Dc来学习网络参",
"K(Fz1,zh) +o2",
"数。具体来说，我们最小化以下监督损失函数",
"(13)",
"Csup = L1 + ApCp,",
"，其中2被设置为等于1，XMATHXLIN5跟随",
"(7)",
"注意，Fz包含所有标记图像的潜在空间向量，",
"其中入p是常数，C1和Cp分别是l1-损失和感知损",
"K(Fz,Fz）是大小为 Ni×N的矩阵，K(zk,Fz）是",
"失「14，55]函数。它们的定义如下：",
"大小为1×Ni的向量。由于以下原因，使用所有向",
"量可能不一定是最佳的：（i）这些向量将对应于图像",
"Ci = llypred - yll1,",
"(8)",
"中的不同区域，在雨纹的含量和密度/方向方面具有厂",
"泛的差异。重要的是只考虑那些与未标记的向量相似",
"Cp = ΦVGG(ypred) -VGG(yt)Il2",
"(9)",
"的向量。(Ii)使用所有向量在计算上是禁止的。因此，"
] |
{
"arxiv_id": "2006.05580",
"language": "zh",
"page_idx": 3
}
|
||
2006.05580-zh-4
|
2006.05580-zh-4.jpg
|
我们仅使用对应于未标记向量的$N_n$最近标记向量。更具体地说，我们在方程(10)-(13)中用$F_z_{l,n}$代替$F_{z_l}$。这里$F_z_l,n=\{z_l^j:z_l^j\in nearest(z_u^k,F_{z_l},N_n)\}$与$nearest(p,Q,N_n)$是一个函数，它在$Q$中找到$p$的顶部$N_n$最近邻居
我们使用公式(12)预测的平均值作为伪 GT $( z_{u, pseudo}^k)$ , 用于潜在空间级别的监督。通过最小化$z_{u,pred}^k=h(x_u,\theta_{enc})$和$z_u,pseudo^k$之间的误差，我们更新编码器$h(\cdot,\theta_{enc})$的权重，从而使网络适应未标记的数据，从而实现更好的泛化。我们还通过最小化方程(13)来最小化预测方差。使用 GP,我们使用$F_z_\iota$中的潜在空间向量近似$z_u^k$,未标记图像的潜在向量，通过这样做，由于潜在向量之间的差异，我们可能最终计算出不正确的伪 GT 预测。这种差异是由于雨纹中的成分不同，例如雨纹的密度，形状和方向不同。。为了解决这个问题，我们使用 GP 最小化$\Sigma_{u,n}^k$在潜在空间向量中的$z_u^k$和$N_n$最近邻之间计算的方差。此外，我们最大化$\Sigma_u,f^k$的方差$\Sigma_u,n^k$是使用 GP 在$z_u^k$和$N_f$最远向量之间计算的$N_f$向量，以确保$F_z_l$中的潜在向量与未标记的向量$z_u^k$不相似，并且不会影响 GP 预测，如下定义<NEXN
因此，使用未标记数据进行训练时使用的损失定义
如下
$$\mathcal{L}_{unsup}=\|z_{u,pred}^{k}-z_{u,pseudo}^{k}\|_{2}+\log\Sigma_{u,n}^{k}+\log(1-\Sigma_{u,f}^{k}),$$
$(\ddot{1}4)$
其中$z_u,pred^k$是通过将未标记的输入图像$x_u^k$转发通过
编码器$h$,i.e.$,z_u,pred=h(x_u,\theta_{enc}),z_{u,pseudo}^k=\mu_u^k$
是伪方程 GT。
## $4. 3. \textbf{ Total loss}$
用于训练网络的总体损失函数定义如下
(15)
$$\mathcal{L}_{total}=\mathcal{L}_{sup}+\lambda_{unsup}\mathcal{L}_{unsup},$$
,其中$\lambda_unsup$是一个预定义的权重，用于控制$\mathcal{L}_sup$和
$\mathcal{L}_{unsup}$的贡献
# 4.4. Training and implementation details
我们使用基于 UNet 样式编码器-解码器体系结构[43]的 UDeNet 网络，其构建块略有不同。网络体系结构的详细信息在补充材料中提供。使用 Adam 优化器训练网络，学习率为0。0002和 4 的 batchsize, 共 60 个纪元。此外，我们将学习率降低了0 倍。每 25 个时期 5 个。我们使用 $\lambda _p= 0. 04$ (方程 $(7)),\lambda_unsup=1.5\times10^{-4}$ (方程(15) ), $N_n=64$和$N_f=64$ 。在训练期间，图像被随机裁剪为 256???。补充材料中提供了具有不同超参数值的消融研究
# 5. Experiments and results
在本节中，我们将介绍数据集的详细信息以及为证明所提出框架的有效性而进行的各种实验。具体来说，
我们进行了两组实验。在第一组中，我们使用建议的框架分析在训练期间使用未标记的现实世界数据的有效性。在这里，我们将我们的方法的性能与最近的图像脱轨(SIRR)[49]的 SSL 框架进行比较。在第二组实验中，我们通过对标记数据的不同百分比进行训练来评估所提出的方法。
## $5. 1. \textbf{ Datasets}$
Rain800:此数据集由 Zhang et al.[57]引人，它总共包含 800 个图像。火车分割由 700 个真实世界的干净图像组成，其中从 UCID 数据集[45]的前半部分随机选择了 500 个图像，从 BSD-500 火车集 [1] 中随机选择了 200 个图像。测试集共由 100 幅图像组成，其中50 幅图像从 UCID 数据集的后半部分中随机选择，其余 50 幅图像从 BSD-500 数据集的测试集中随机选择。作者通过合成不同强度和方向的雨纹来生成相应的雨天图像。
Rain200H:Yangetal。[51] 从 BSD200 [31] 收集图像以创建 3 个数据集：Rain12, Rain200L 和 Rain200H。在 [22] 之后，我们使用最困难的一个 Rain200H 来评估我们的模型。训练集的图像是从 BSD300 数据集中收集的。使用逼真的技术合成具有不同方向的雨纹。Rain200H 列车集中有 1800 对合成图像，测试集中有200对。
DDN-SIRR 数据集：Wei et al.[49] 构建了一个由标记的合成训练集和未标记的现实世界数据集组成的数据集。此数据集专门用于评估半监督学习框架。标记的训练集是从 Fu
6. Experiments and results
[9]借来的，它由 9,100 个图像对组成，这些图像对是通过在 UCID 数据集 [45] 的干净图像上合成不同类型的雨条纹而获得的。未标记的真实世界合成火车集包括从[50,51,57]和 Google 图像搜索收集的图像。此外，测试集包括两个类别：(i)密集的雨纹，和(ii)稀疏的雨纹每个测试集由 10 个图像组成。
## 6.1. Use of real-world data
本实验的目标是分析在训练框架中使用未标记的现实世界数据以及标记的合成数据集的效果。遵循[49] 设置的协议，我们将 DDN-SIRR 数据集中的“标记的合成”火车集用作$\mathcal{D}_{\mathcal{L}}$ ,将 DDN-SIRR 数据集中的“真实世界”火车集用作$\mathcal{D}_u$。对(i)来自 DDN-SIRR 的合成测试集和 (ii) 来自 DDN-SIRR 的真实世界测试集进行评估。
合成测试集上的结果：合成测试集上的评估结果如表所示。1。与 [49] 类似，我们使用 PSNR 作为评估指标。我们将所提出的方法与 DSC[30] , LP[24] , JORDER [51] , DDN[9] , JBO[60]和 DID-MDN [58]等几种
|
[
[
0.08251953125,
0.0947265625,
0.471435546875,
0.0947265625,
0.471435546875,
0.10479736328125,
0.08251953125,
0.10479736328125
],
[
0.5048828125,
0.0947265625,
0.8896484375,
0.0947265625,
0.8896484375,
0.10479736328125,
0.5048828125,
0.10479736328125
],
[
0.08251953125,
0.10919189453125,
0.468017578125,
0.10919189453125,
0.468017578125,
0.12249755859375,
0.08251953125,
0.12249755859375
],
[
0.5048828125,
0.10986328125,
0.8896484375,
0.10986328125,
0.8896484375,
0.12060546875,
0.5048828125,
0.12060546875
],
[
0.08087158203125,
0.125,
0.468017578125,
0.125,
0.468017578125,
0.1383056640625,
0.08087158203125,
0.1383056640625
],
[
0.50244140625,
0.12371826171875,
0.890625,
0.12371826171875,
0.890625,
0.136962890625,
0.50244140625,
0.136962890625
],
[
0.080078125,
0.1395263671875,
0.468017578125,
0.1383056640625,
0.468017578125,
0.15283203125,
0.080078125,
0.154052734375
],
[
0.50341796875,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.50341796875,
0.1514892578125
],
[
0.08087158203125,
0.1546630859375,
0.1993408203125,
0.1546630859375,
0.1993408203125,
0.1685791015625,
0.08087158203125,
0.1685791015625
],
[
0.50341796875,
0.1514892578125,
0.890625,
0.15283203125,
0.890625,
0.1673583984375,
0.50341796875,
0.1666259765625
],
[
0.098876953125,
0.17041015625,
0.468017578125,
0.169189453125,
0.468017578125,
0.184326171875,
0.098876953125,
0.1849365234375
],
[
0.50341796875,
0.1685791015625,
0.62744140625,
0.1685791015625,
0.62744140625,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.0767822265625,
0.18310546875,
0.471435546875,
0.18310546875,
0.471435546875,
0.203857421875,
0.0767822265625,
0.203857421875
],
[
0.5009765625,
0.1932373046875,
0.60888671875,
0.1956787109375,
0.60888671875,
0.210205078125,
0.5009765625,
0.2083740234375
],
[
0.08251953125,
0.20263671875,
0.468017578125,
0.20263671875,
0.468017578125,
0.2171630859375,
0.08251953125,
0.2171630859375
],
[
0.08251953125,
0.2197265625,
0.468017578125,
0.2197265625,
0.468017578125,
0.23291015625,
0.08251953125,
0.23291015625
],
[
0.50244140625,
0.2197265625,
0.8896484375,
0.2197265625,
0.8896484375,
0.23291015625,
0.50244140625,
0.23291015625
],
[
0.08087158203125,
0.23486328125,
0.468017578125,
0.23486328125,
0.468017578125,
0.248779296875,
0.08087158203125,
0.248779296875
],
[
0.5048828125,
0.2342529296875,
0.8896484375,
0.2342529296875,
0.8896484375,
0.2474365234375,
0.5048828125,
0.2474365234375
],
[
0.08251953125,
0.25,
0.468017578125,
0.25,
0.468017578125,
0.26318359375,
0.08251953125,
0.26318359375
],
[
0.50341796875,
0.2474365234375,
0.890625,
0.2474365234375,
0.890625,
0.261962890625,
0.50341796875,
0.261962890625
],
[
0.08087158203125,
0.264404296875,
0.468017578125,
0.264404296875,
0.468017578125,
0.27783203125,
0.08087158203125,
0.27783203125
],
[
0.50341796875,
0.264404296875,
0.890625,
0.264404296875,
0.890625,
0.27783203125,
0.50341796875,
0.27783203125
],
[
0.08087158203125,
0.279052734375,
0.468017578125,
0.279052734375,
0.468017578125,
0.292236328125,
0.08087158203125,
0.292236328125
],
[
0.50341796875,
0.279052734375,
0.8896484375,
0.279052734375,
0.8896484375,
0.292236328125,
0.50341796875,
0.292236328125
],
[
0.08087158203125,
0.294189453125,
0.468017578125,
0.294189453125,
0.468017578125,
0.30810546875,
0.08087158203125,
0.30810546875
],
[
0.50341796875,
0.294189453125,
0.890625,
0.294189453125,
0.890625,
0.30810546875,
0.50341796875,
0.30810546875
],
[
0.08087158203125,
0.31005859375,
0.468017578125,
0.31005859375,
0.468017578125,
0.323974609375,
0.08087158203125,
0.323974609375
],
[
0.50634765625,
0.311279296875,
0.8896484375,
0.311279296875,
0.8896484375,
0.3212890625,
0.50634765625,
0.3212890625
],
[
0.08251953125,
0.326904296875,
0.46728515625,
0.326904296875,
0.46728515625,
0.337158203125,
0.08251953125,
0.337158203125
],
[
0.50341796875,
0.324462890625,
0.890625,
0.324462890625,
0.890625,
0.33837890625,
0.50341796875,
0.33837890625
],
[
0.08251953125,
0.341552734375,
0.468017578125,
0.341552734375,
0.468017578125,
0.354736328125,
0.08251953125,
0.354736328125
],
[
0.50341796875,
0.341552734375,
0.5595703125,
0.341552734375,
0.5595703125,
0.352783203125,
0.50341796875,
0.352783203125
],
[
0.08087158203125,
0.356689453125,
0.468017578125,
0.356689453125,
0.468017578125,
0.37060546875,
0.08087158203125,
0.37060546875
],
[
0.08251953125,
0.373779296875,
0.46728515625,
0.373779296875,
0.46728515625,
0.386962890625,
0.08251953125,
0.386962890625
],
[
0.50341796875,
0.37060546875,
0.8896484375,
0.37060546875,
0.8896484375,
0.3837890625,
0.50341796875,
0.3837890625
],
[
0.08087158203125,
0.38818359375,
0.468017578125,
0.38818359375,
0.468017578125,
0.401611328125,
0.08087158203125,
0.401611328125
],
[
0.5048828125,
0.386474609375,
0.890625,
0.386474609375,
0.890625,
0.399658203125,
0.5048828125,
0.399658203125
],
[
0.08087158203125,
0.404052734375,
0.1968994140625,
0.401611328125,
0.1968994140625,
0.415283203125,
0.08087158203125,
0.417236328125
],
[
0.5048828125,
0.401611328125,
0.890625,
0.401611328125,
0.890625,
0.415283203125,
0.5048828125,
0.415283203125
],
[
0.1004638671875,
0.417236328125,
0.468017578125,
0.41845703125,
0.468017578125,
0.43310546875,
0.1004638671875,
0.431884765625
],
[
0.5048828125,
0.416748046875,
0.8896484375,
0.416748046875,
0.8896484375,
0.429931640625,
0.5048828125,
0.429931640625
],
[
0.08251953125,
0.43505859375,
0.1168212890625,
0.43505859375,
0.1168212890625,
0.447509765625,
0.08251953125,
0.447509765625
],
[
0.50244140625,
0.43115234375,
0.88818359375,
0.431884765625,
0.88818359375,
0.44580078125,
0.50244140625,
0.4443359375
],
[
0.50244140625,
0.4443359375,
0.8896484375,
0.44580078125,
0.8896484375,
0.460205078125,
0.50244140625,
0.458984375
],
[
0.07843017578125,
0.460205078125,
0.468017578125,
0.460205078125,
0.468017578125,
0.48095703125,
0.07843017578125,
0.48095703125
],
[
0.50341796875,
0.462158203125,
0.55810546875,
0.462158203125,
0.55810546875,
0.473388671875,
0.50341796875,
0.473388671875
],
[
0.438720703125,
0.47607421875,
0.465576171875,
0.477783203125,
0.464111328125,
0.4912109375,
0.435546875,
0.488037109375
],
[
0.080078125,
0.4912109375,
0.4697265625,
0.4912109375,
0.4697265625,
0.5087890625,
0.080078125,
0.5087890625
],
[
0.50244140625,
0.490478515625,
0.890625,
0.4912109375,
0.890625,
0.505859375,
0.50244140625,
0.5048828125
],
[
0.07843017578125,
0.50830078125,
0.468017578125,
0.5087890625,
0.468017578125,
0.5263671875,
0.07843017578125,
0.52587890625
],
[
0.50341796875,
0.5068359375,
0.8896484375,
0.5068359375,
0.8896484375,
0.52099609375,
0.50341796875,
0.52099609375
],
[
0.08251953125,
0.52783203125,
0.1846923828125,
0.52783203125,
0.1846923828125,
0.5390625,
0.08251953125,
0.5390625
],
[
0.5048828125,
0.52294921875,
0.8896484375,
0.52294921875,
0.8896484375,
0.5361328125,
0.5048828125,
0.5361328125
],
[
0.5048828125,
0.5390625,
0.5947265625,
0.5390625,
0.5947265625,
0.5498046875,
0.5048828125,
0.5498046875
],
[
0.08087158203125,
0.5517578125,
0.191162109375,
0.5517578125,
0.191162109375,
0.56494140625,
0.08087158203125,
0.56494140625
],
[
0.50341796875,
0.568359375,
0.73779296875,
0.568359375,
0.73779296875,
0.58203125,
0.50341796875,
0.58203125
],
[
0.10296630859375,
0.57568359375,
0.37744140625,
0.57568359375,
0.37744140625,
0.5888671875,
0.10296630859375,
0.5888671875
],
[
0.52392578125,
0.5947265625,
0.890625,
0.5947265625,
0.890625,
0.60791015625,
0.52392578125,
0.60791015625
],
[
0.1666259765625,
0.6015625,
0.3798828125,
0.60400390625,
0.3798828125,
0.61865234375,
0.1666259765625,
0.6162109375
],
[
0.438720703125,
0.60400390625,
0.46728515625,
0.60400390625,
0.46728515625,
0.6162109375,
0.438720703125,
0.6162109375
],
[
0.5048828125,
0.609375,
0.890625,
0.609375,
0.890625,
0.62255859375,
0.5048828125,
0.62255859375
],
[
0.080078125,
0.62890625,
0.4697265625,
0.630859375,
0.4697265625,
0.6484375,
0.080078125,
0.646484375
],
[
0.5048828125,
0.625,
0.890625,
0.625,
0.890625,
0.63818359375,
0.5048828125,
0.63818359375
],
[
0.50341796875,
0.63818359375,
0.89208984375,
0.64013671875,
0.89208984375,
0.65380859375,
0.50341796875,
0.6513671875
],
[
0.08251953125,
0.64794921875,
0.1854248046875,
0.64794921875,
0.1854248046875,
0.662109375,
0.08251953125,
0.662109375
],
[
0.5048828125,
0.65478515625,
0.89208984375,
0.65478515625,
0.89208984375,
0.66845703125,
0.5048828125,
0.66845703125
],
[
0.08087158203125,
0.67236328125,
0.39794921875,
0.67236328125,
0.39794921875,
0.6875,
0.08087158203125,
0.6875
],
[
0.50244140625,
0.66943359375,
0.78173828125,
0.67041015625,
0.78173828125,
0.6845703125,
0.50244140625,
0.68310546875
],
[
0.101318359375,
0.69580078125,
0.468017578125,
0.69580078125,
0.468017578125,
0.708984375,
0.101318359375,
0.708984375
],
[
0.50244140625,
0.69580078125,
0.71142578125,
0.6962890625,
0.71142578125,
0.7109375,
0.50244140625,
0.71044921875
],
[
0.08087158203125,
0.7109375,
0.468017578125,
0.7109375,
0.468017578125,
0.724609375,
0.08087158203125,
0.724609375
],
[
0.52294921875,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.52294921875,
0.734375
],
[
0.08251953125,
0.72802734375,
0.46728515625,
0.72802734375,
0.46728515625,
0.73779296875,
0.08251953125,
0.73779296875
],
[
0.08087158203125,
0.74072265625,
0.46728515625,
0.74072265625,
0.46728515625,
0.75390625,
0.08087158203125,
0.75390625
],
[
0.5048828125,
0.7373046875,
0.88818359375,
0.7373046875,
0.88818359375,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.08087158203125,
0.755859375,
0.46728515625,
0.755859375,
0.46728515625,
0.77099609375,
0.08087158203125,
0.77099609375
],
[
0.50341796875,
0.75048828125,
0.8896484375,
0.75048828125,
0.8896484375,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.07843017578125,
0.77099609375,
0.4697265625,
0.76806640625,
0.4697265625,
0.7861328125,
0.07843017578125,
0.78857421875
],
[
0.50341796875,
0.7666015625,
0.8896484375,
0.7666015625,
0.8896484375,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.7861328125,
0.468017578125,
0.7861328125,
0.468017578125,
0.7998046875,
0.08087158203125,
0.7998046875
],
[
0.50341796875,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.50341796875,
0.79541015625,
0.8896484375,
0.79541015625,
0.8896484375,
0.8095703125,
0.50341796875,
0.8095703125
],
[
0.08087158203125,
0.8017578125,
0.468017578125,
0.8017578125,
0.468017578125,
0.81494140625,
0.08087158203125,
0.81494140625
],
[
0.08087158203125,
0.8173828125,
0.2296142578125,
0.8173828125,
0.2296142578125,
0.83056640625,
0.08087158203125,
0.83056640625
],
[
0.50634765625,
0.8125,
0.611328125,
0.8125,
0.611328125,
0.82470703125,
0.50634765625,
0.82470703125
],
[
0.10455322265625,
0.84765625,
0.314453125,
0.84765625,
0.314453125,
0.861328125,
0.10455322265625,
0.861328125
],
[
0.50732421875,
0.84326171875,
0.8896484375,
0.84326171875,
0.8896484375,
0.85400390625,
0.50732421875,
0.85400390625
],
[
0.08251953125,
0.849609375,
0.1094970703125,
0.849609375,
0.1094970703125,
0.85791015625,
0.08251953125,
0.85791015625
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.1004638671875,
0.87255859375,
0.46728515625,
0.87255859375,
0.46728515625,
0.88623046875,
0.1004638671875,
0.88623046875
],
[
0.50244140625,
0.87255859375,
0.8896484375,
0.87255859375,
0.8896484375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.08087158203125,
0.88720703125,
0.47216796875,
0.88720703125,
0.47216796875,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.9013671875,
0.50244140625,
0.9013671875
]
] |
[
"我们仅使用对应于未标记向量的.最近标记向量。",
"我们进行了两组实验。在第一组中，我们使用建议的框",
"更具体地说，我们在方程（10）=（13）中用Fzt.n代替",
"架分析在训练期间使用未标记的现实世界数据的有效",
"Fzt。这里Fzun=(z:zEnearest(zh,Fzu,Nn)]与",
"性。在这里，我们将我们的方法的性能与最近的图像脱",
"nearest(p,Q,Nn）是一个函数，它在Q中找到p的顶",
"轨（SIRR）【49]的SSL框架进行比较。在第二组实验",
"部Nn最近邻居",
"中，我们通过对标记数据的不同百分比进行训练来评",
"我们使用公式（12）预测的平均值作为伪GT",
"估所提出的方法。",
"（zh,pseud。），用于潜在空间级别的监督。通过最小",
"5.1. Datasets",
"化zh.pred=h(auQenc）和zh.pseudo之间的误差，我们",
"更新编码器h(,Qenc）的权重，从而使网络适应未标记",
"Rain800：此数据集由Zhangetal.[57]引人，它总共",
"的数据，从而实现更好的泛化。我们还通过最小化方程",
"包含800个图像。火车分割由700个真实世界的干净",
"(13）来最小化预测方差。使用GP，我们使用Fz中的",
"图像组成，其中从UCID数据集【45]的前半部分随机",
"潜在空间向量近似之，未标记图像的潜在向量，通过",
"选择了500个图像，从BSD-500火车集［1]中随机选",
"这样做，由于潜在向量之间的差异，我们可能最终计算",
"择了200个图像。测试集共由100幅图像组成，其中",
"出不正确的伪GT预测。这种差异是由于雨纹中的成",
"50幅图像从UCID数据集的后半部分中随机选择，其",
"分不同，例如雨纹的密度，形状和方向不同。。为了解",
"余50幅图像从BSD-500数据集的测试集中随机选择。",
"决这个问题，我们使用GP最小化在潜在空间向",
"作者通过合成不同强度和方向的雨纹来生成相应的雨",
"量中的zk和Nn最近邻之间计算的方差。此外，我们",
"天图像。",
"最大化hf的方差kn是使用GP在和N最远",
"向量之间计算的N向量，以确保Fz,中的潜在向量与",
"Rain200H:Yangetal。[51]从BSD20031]收集图像以",
"未标记的向量z不相似，并且不会影响GP预测，如",
"创建3个数据集：Rain12，Rain200L和Rain200H。",
"下定义<NEXN",
"在[22]之后，我们使用最困难的一个Rain200H来评",
"因此，使用未标记数据进行训练时使用的损失定义",
"估我们的模型。训练集的图像是从BSD300数据集中",
"如下",
"收集的。使用逼真的技术合成具有不同方向的雨纹，",
"Rain200H列车集中有1800对合成图像，测试集中有",
"Lunsup = I/zu,pred-2,pseudoll2+log h,n+log(1-h,f),",
"200对。",
"(14)",
"其中zh,pred是通过将未标记的输人图像rh转发通过",
"DDN-SIRR数据集：Weietal.[49]构建了一个由标记",
"编码器h,i.e.，zh,pred=h(au,Denc），zh,pseudo=μ",
"的合成训练集和未标记的现实世界数据集组成的数据",
"是伪方程GT。",
"集。此数据集专门用于评估半监督学习框架。标记的训",
"练集是从Fu",
"4.3. Total loss",
"6. Experiments and results",
"用于训练网络的总体损失函数定义如下",
"[9]借来的，它由9,100个图像对组成，这些图像对",
"Ctotal = Csup +AunsupLunsup:",
"(15)",
"是通过在UCID数据集「45]的干净图像上合成不同类",
"，其中入unsup是一个预定义的权重，用于控制Lsup和",
"型的雨条纹而获得的。未标记的真实世界合成火车集",
"包括从[50，51，57]和Google图像搜索收集的图像。此",
"Lunsup的贡献",
"外，测试集包括两个类别：（i）密集的雨纹，和（i）稀",
"4.4. Training and implementation details",
"疏的雨纹每个测试集由10个图像组成。",
"我们使用基于UNet样式编码器-解码器体系结构",
"6.1. Use of real-world data",
"[43]的UDeNet网络，其构建块略有不同。网络体系结",
"本实验的目标是分析在训练框架中使用未标记的现",
"构的详细信息在补充材料中提供。使用Adam优化器训",
"练网络，学习率为0。0002和4的batchsize，共60个纪",
"实世界数据以及标记的合成数据集的效果。遵循49",
"元。此外，我们将学习率降低了0倍。每25个时期5个。",
"设置的协议，我们将DDN-SIRR数据集中的“标记的",
"我们使用^p=0.04（方程（7）），入unsup=1.5×10-4",
"合成”火车集用作Dc：将DDN-SIRR数据集中的",
"（方程（15）），N元=64和Nf=64。在训练期间，图",
"“真实世界”火车集用作Du。对（i）来自DDN-SIRR",
"的合成测试集和（ii）来自DDN-SIRR的真实世界测",
"像被随机裁剪为256??？。补充材料中提供了具有不同",
"超参数值的消融研究",
"试集进行评估。",
"Experiments and results",
"合成测试集上的结果：合成测试集上的评估结果如表",
"5元",
"所示。1。与「49]类似，我们使用PSNR作为评估指标。",
"在本节中，我们将介绍数据集的详细信息以及为证",
"我们将所提出的方法与DSC[30]，LP[24]，JORDER",
"明所提出框架的有效性而进行的各种实验。具体来说，",
"[51]，DDN[9]，JBO[60]和DID-MDN[58]等几种"
] |
{
"arxiv_id": "2006.05580",
"language": "zh",
"page_idx": 4
}
|
||
2106.16115-zh-0
|
2106.16115-zh-0.jpg
|
# 随机子模块盖的自适应能力
Rohan Ghuge* Anupam Gupta$^\dagger$ Viswanath Nagarajan'
May 18, 2024
$\mathbf{Abstract}$
In the stochastic submodular cover problem, the goal is to select a subset of stochastic items of minimum expected cost to cover a submodular function. Solutions in this setting correspond to sequential decision processes that select items one by one “adaptively”(depending on prior observations). While such adaptive solutions achieve the best objective, the inherently sequential nature makes them undesirable in many applications. We ask: how well can solutions with only a few adaptive rounds approximate fully-adaptive solutions? We give nearly tight answers for both independent and correlated settings, proving smooth tradeoffs between the number of adaptive rounds and the solution quality, relative to fully adaptive solutions. Experiments on synthetic and real datasets show qualitative improvements in the solutions as we allow more rounds of adaptivity; in practice, solutions with a few rounds of adaptivity are nearly as good as fully adaptive solutions.
# 1 Introduction
子模块性是在图像分割，数据摘要$\mathbb{E}[$SSS07,LB11,SSSJ12] ,假设识别$\mathbb{E}[$BLK12,CSM$^+14]$ ,信息收集$\overline{\mathrm{E}}[$RSKF18] 和社交网络$\overline[$KKT15] 等应用中出现的基本概念。submodular cover 优化问题要求我们选择项目的最小成本子集$S$来覆盖单调子模函数$f$ 。子模块盖已在机器学习，计算机科学和运筹学中进行了广泛的研究E[Wol82,GK11,MKBK15,BEM18]:这是传感器部署和医疗诊断的两个示例
在传感器部署设置中，我们考虑部署传感器集合以监视某些现象 [KG07,MUS14,SLLL19] 的问题，例如：我们可能希望监视空气质量或交通状况。每个传感器可以覆盖的区域取决于其感应范围。该问题的目标是部署尽可能少的传感器以完全覆盖所需区域。作为部署的传感器的功能而覆盖的区域是子模块功能，我们可以将传感器部署问题转换为子模块盖的特殊情况
在医疗诊断示例中，我们有$s$可能的条件(假设),患者可能会遭受其发生的先验，我们的目标是执行测试以尽快识别正确的条件E[ GG74, KPB99, Das04, CLS14] 。通过将每个测试视为消除所有不一致的假设，可以将其转换为子模块封面。因此，我们需要$s-1$的覆盖值：一旦消除了$s-1$ 不一致的假设，其余的假设必须是正确的
请注意，这两个应用程序都涉及不确定的数据：在实际设置传感器之前，不知道传感器覆盖的精确区域，并且在采取操作之前，不知道测试的精确结果(正/负)。这种不确定性可以使用stochastic submodularoptimization建模，其中项目是随机的。作为随机性质的一个简单示例，每个项目可能是活动的或不活动的 (具有已知的概率),并且只有活动的项目才有助于亚模功能
$^{*}$Department of Industrial and Operations Engineering, University of Michigan, Ann Arbor, USA. Research sup-
ported in part by NSF grants CMMI-1940766 and CCF-2006778
$^{\dagger}$Department of Computer Science, Carnegie Mellon University, Pittsburgh, USA. Supported in part by NSF
awards CCF-1907820, CCF-1955785, and CCF-2006953.
1
|
[
[
0.3349609375,
0.1458740234375,
0.6650390625,
0.1458740234375,
0.6650390625,
0.1634521484375,
0.3349609375,
0.1634521484375
],
[
0.2254638671875,
0.189453125,
0.34716796875,
0.189453125,
0.34716796875,
0.203857421875,
0.2254638671875,
0.203857421875
],
[
0.39794921875,
0.1881103515625,
0.53515625,
0.1881103515625,
0.53515625,
0.20263671875,
0.39794921875,
0.20263671875
],
[
0.583984375,
0.189453125,
0.7685546875,
0.189453125,
0.7685546875,
0.203857421875,
0.583984375,
0.203857421875
],
[
0.439453125,
0.2183837890625,
0.55810546875,
0.2203369140625,
0.55810546875,
0.238037109375,
0.439453125,
0.236083984375
],
[
0.464111328125,
0.285400390625,
0.5361328125,
0.285400390625,
0.5361328125,
0.2998046875,
0.464111328125,
0.2998046875
],
[
0.1846923828125,
0.309326171875,
0.83837890625,
0.309326171875,
0.83837890625,
0.323974609375,
0.1846923828125,
0.323974609375
],
[
0.16259765625,
0.326904296875,
0.83642578125,
0.326904296875,
0.83642578125,
0.337158203125,
0.16259765625,
0.337158203125
],
[
0.16015625,
0.339111328125,
0.83837890625,
0.34033203125,
0.83837890625,
0.354736328125,
0.16015625,
0.354248046875
],
[
0.158447265625,
0.354736328125,
0.83837890625,
0.35595703125,
0.83837890625,
0.37060546875,
0.158447265625,
0.369384765625
],
[
0.16015625,
0.37060546875,
0.83642578125,
0.37060546875,
0.83642578125,
0.385009765625,
0.16015625,
0.385009765625
],
[
0.16015625,
0.38818359375,
0.83837890625,
0.38818359375,
0.83837890625,
0.401611328125,
0.16015625,
0.401611328125
],
[
0.158447265625,
0.404052734375,
0.8388671875,
0.404052734375,
0.8388671875,
0.41845703125,
0.158447265625,
0.41845703125
],
[
0.160888671875,
0.419189453125,
0.8349609375,
0.419189453125,
0.8349609375,
0.43310546875,
0.160888671875,
0.43310546875
],
[
0.158447265625,
0.436279296875,
0.83642578125,
0.43505859375,
0.83642578125,
0.448974609375,
0.158447265625,
0.449462890625
],
[
0.16015625,
0.451904296875,
0.83642578125,
0.451904296875,
0.83642578125,
0.46533203125,
0.16015625,
0.46533203125
],
[
0.160888671875,
0.46728515625,
0.35302734375,
0.46728515625,
0.35302734375,
0.48095703125,
0.160888671875,
0.48095703125
],
[
0.111083984375,
0.5068359375,
0.3095703125,
0.50830078125,
0.3095703125,
0.52587890625,
0.111083984375,
0.5244140625
],
[
0.1168212890625,
0.54150390625,
0.88134765625,
0.54248046875,
0.88134765625,
0.556640625,
0.1168212890625,
0.55615234375
],
[
0.11767578125,
0.56005859375,
0.88134765625,
0.56005859375,
0.88134765625,
0.57373046875,
0.11767578125,
0.57373046875
],
[
0.11767578125,
0.5791015625,
0.88134765625,
0.5791015625,
0.88134765625,
0.59228515625,
0.11767578125,
0.59228515625
],
[
0.1168212890625,
0.5966796875,
0.88134765625,
0.5966796875,
0.88134765625,
0.60986328125,
0.1168212890625,
0.60986328125
],
[
0.11767578125,
0.6142578125,
0.2059326171875,
0.6142578125,
0.2059326171875,
0.62890625,
0.11767578125,
0.62890625
],
[
0.146240234375,
0.6318359375,
0.88134765625,
0.6318359375,
0.88134765625,
0.646484375,
0.146240234375,
0.646484375
],
[
0.1168212890625,
0.65087890625,
0.88134765625,
0.65087890625,
0.88134765625,
0.6640625,
0.1168212890625,
0.6640625
],
[
0.11767578125,
0.66845703125,
0.88134765625,
0.66845703125,
0.88134765625,
0.681640625,
0.11767578125,
0.681640625
],
[
0.11767578125,
0.6875,
0.7236328125,
0.6875,
0.7236328125,
0.70068359375,
0.11767578125,
0.70068359375
],
[
0.1446533203125,
0.705078125,
0.88134765625,
0.705078125,
0.88134765625,
0.71826171875,
0.1446533203125,
0.71826171875
],
[
0.11767578125,
0.72265625,
0.88134765625,
0.72265625,
0.88134765625,
0.736328125,
0.11767578125,
0.736328125
],
[
0.1168212890625,
0.7392578125,
0.88134765625,
0.74072265625,
0.88134765625,
0.7548828125,
0.1168212890625,
0.75390625
],
[
0.1168212890625,
0.75830078125,
0.438720703125,
0.75830078125,
0.438720703125,
0.771484375,
0.1168212890625,
0.771484375
],
[
0.1446533203125,
0.7763671875,
0.88134765625,
0.7763671875,
0.88134765625,
0.79052734375,
0.1446533203125,
0.79052734375
],
[
0.1168212890625,
0.79443359375,
0.88134765625,
0.79443359375,
0.88134765625,
0.80810546875,
0.1168212890625,
0.80810546875
],
[
0.11767578125,
0.8125,
0.8798828125,
0.8125,
0.8798828125,
0.82568359375,
0.11767578125,
0.82568359375
],
[
0.1168212890625,
0.8291015625,
0.7685546875,
0.830078125,
0.7685546875,
0.8447265625,
0.1168212890625,
0.84326171875
],
[
0.136474609375,
0.85595703125,
0.87841796875,
0.85595703125,
0.87841796875,
0.869140625,
0.136474609375,
0.869140625
],
[
0.1168212890625,
0.87060546875,
0.54345703125,
0.87060546875,
0.54345703125,
0.8837890625,
0.1168212890625,
0.8837890625
],
[
0.136474609375,
0.88330078125,
0.88134765625,
0.8837890625,
0.88134765625,
0.8984375,
0.136474609375,
0.89794921875
],
[
0.11517333984375,
0.8994140625,
0.488525390625,
0.8984375,
0.488525390625,
0.912109375,
0.11517333984375,
0.91357421875
],
[
0.490234375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.95068359375,
0.490234375,
0.95068359375
]
] |
[
"随机子模块盖的自适应能力",
"Rohan Ghuge*",
"Anupam Guptat",
"Viswanath Nagarajan",
"May 18, 2024",
"Abstract",
"In the stochastic submodular cover problem, the goal is to select a subset of stochastic items",
"ofminimum expected costto cover asubmodularfunction.Solutionsinthissetting correspond",
"to sequential decision processes that select items one by one “adaptively\"(depending on prior",
"observations). While such adaptive solutions achieve the best objective, the inherently sequential",
"nature makes them undesirable in many applications. We ask: how well can solutions with only",
"a few adaptive rounds approximate fully-adaptive solutions? We give nearly tight answers for",
"both independent and correlated settings, proving smooth tradeoffs between the number of",
"adaptive rounds and the solution quality, relative to fully adaptive solutions. Experiments or",
"syntheticand real datasets showqualitativeimprovements in the solutions as weallowmore",
"rounds of adaptivity; in practice, solutions with a few rounds of adaptivity are nearly as good",
"as fully adaptive solutions.",
"1Introduction",
"子模块性是在图像分割，数据摘要E[SSS07,LB11,SSSJ12]，假设识别E[BLK12,CSM+14]，信息",
"收集E[RSKF18]和社交网络E[KKT15]等应用中出现的基本概念。submodularcouer优化问题",
"要求我们选择项目的最小成本子集S来覆盖单调子模函数f。子模块盖已在机器学习，计算机科",
"学和运筹学中进行了广泛的研究EWo182,GK11,MKBK15,BEM18l：这是传感器部署和医疗诊断",
"的两个示例",
"在传感器部署设置中，我们考虑部署传感器集合以监视某些现象[KG07,MUS14,SLLL19]的问",
"题，例如：我们可能希望监视空气质量或交通状况。每个传感器可以覆盖的区域取决于其感应范",
"围。该问题的目标是部署尽可能少的传感器以完全覆盖所需区域。作为部署的传感器的功能而覆盖",
"的区域是子模块功能，我们可以将传感器部署问题转换为子模块盖的特殊情况",
"在医疗诊断示例中，我们有s可能的条件（假设），患者可能会遭受其发生的先验，我们的目标",
"是执行测试以尽快识别正确的条件E[GG74,KPB99,Das04,CLS14]。通过将每个测试视为消除所",
"有不一致的假设，可以将其转换为子模块封面。因此，我们需要s一1的覆盖值：一且消除了s-1",
"不一致的假设，其余的假设必须是正确的",
"请注意，这两个应用程序都涉及不确定的数据：在实际设置传感器之前，不知道传感器覆盖的精",
"确区域，并且在采取操作之前，不知道测试的精确结果（正/负）。这种不确定性可以使用stochastic",
"submodularoptimization建模，其中项目是随机的。作为随机性质的一个简单示例，每个项目可",
"能是活动的或不活动的（具有已知的概率），并且只有活动的项目才有助于亚模功能",
"*Department of Industrial and Operations Engineering, University of Michigan, Ann Arbor, USA. Research sup",
"ported in part by NSF grants CMMI-1940766 and CCF-2006778",
"tDepartment of Computer Science, Carnegie Mellon University, Pittsburgh, USA. Supported in part by NSF",
"awards CCF-1907820, CCF-1955785, and CCF-2006953.",
"1"
] |
{
"arxiv_id": "2106.16115",
"language": "zh",
"page_idx": 0
}
|
||
2106.16115-zh-1
|
2106.16115-zh-1.jpg
|
随机子模块覆盖的解决方案现在是顺序决策过程。在每个步骤中，一个项目是probed及其实现 (e。g。, 活动或不活动) 被观察到。该过程通常是自适应的，其中来自先前探测的项目的所有信息都用于识别要探测的下一个项目。此过程一直持续到涵盖子模块功能为止，目标是最大程度地降低被探测项目的预期成本。这种自适应解决方案本质上是完全顺序的，如果探测一个项目是耗时的，这是不可取的。E。g。,在传感器部署中，探测传感器对应于物理部署传感器并观察其是否按预期运行，这可能需要数小时或数天。或者，在医疗诊断中探测/执行测试可能需要长时间等待测试结果。因此，我们更喜欢只有很少rounds适应性的解决方案
受此启发，我们问：对于随机子模块覆盖问题，只有几轮自适应回合的解决方案如何近似完全自适应解决方案？我们考虑两种情况，其中不同项目的实现是独立的，并且允许它们相互关联。对于这两种情况，我们给出了几乎严格的答案，在自适应回合的数量$r$和解决方案质量(相对于完全自适应解决方案)之间进行了平滑的权衡
我们工作的主要贡献是针对自适应回合的“基于集合”模型中的随机子模块覆盖的$r$回合自适应解决方案。在这个模型中，每一轮都会并行探测一个固定的项目子集 (并且会产生它们的总成本)。本轮的决定可能取决于前几轮的实现。但是，如 [AAK19] 中所述，如果我们要求函数$f$被概率 1 覆盖，则$r$ -round-adaptivity 差距会非常大。(有关示例，请参见 §A 。)因此，我们专注于基于集合的解决方案，这些解决方案仅需要以高概率覆盖函数
在设计算法时，事实证明使用“置换”模型进行自适应回合更为方便，其中函数覆盖了概率一。此模型也用于先前的文献 [GV06,AAK19] 。在这里，$r$ -round-adaptive 解决方案的每一轮都指定所有剩余项目的顺序，并按此顺序探测它们，直到某个停止规则。有关正式定义，请参见Defnition 2.1。此外，我们在置换模型中的$r$轮自适应算法可以转换为基于集合的模型中的$r$轮自适应算法。我们在基于集合的模型中获得以下算法：
$\bullet$ for any $\eta\in(0,1)$ , finds an $r$ -round adaptive solution that has expected cost at most $\frac r\alpha\eta\cdot$OPT
and covers the function with probability at least $1- \eta$ .
$\bullet$ finds an $O(r)$ -round adaptive solution that has expected cost at most $O(\alpha)\cdot$OPT and covers
the function with probability at least $1- e^{- \Omega ( r) }$ .
这里 OPT 是最佳完全自适应解决方案的成本，$\alpha$是我们的算法在置换模型中的近似比
上面的第一个算法是针对$r$ ,自适应的轮数较小(例如，一个常数)的情况。在这一点上，我们保持轮数相同，但我们在预期成本中损失了一个因子$r$ 。第二种算法是针对$r$很大的情况，例如。g。,超过一个常数。在这里，基于集合的轮数增加了2 倍，但我们只损失了预期成本中的一个常数因子。我们在 §A 中形式化并证明了这些结果。对于本文的其余部分，$r$轮自适应算法是指置换模型中的$r$轮自适应算法(除非另有说明)
考虑具有$f:2^U\to\mathbb{Z}_{\geq0}$ 。有$m$项，其中每个项$i$是一个随机变量$\mathcal{X}_i$ ,其成本为$c_i$ ,并对应于$U$的随机元素。(我们的结果扩展到更一般的设置，其中每个项目都实现了$U$的子集。)的目标是选择一组项目，以使其相应实现的联合$S$满足$f(S)=Q$ ,并且预期成本最小化。我们的第一个结果是当项目具有独立分布时
$\textbf{Theorem 1. 1 ( Independent Items) . For any integer r1 , there is an r - round adaptive algorithm}$ $forthe\textit{ stochastic submodular cover problem with cost O( Q}^{1/ r}\cdot \log Q) \textit{ times the cost of an optimal}$ $adaptive\textit{ algorithm}.$
通过消除对项目$m$的数量和项目成本 (可能比$Q$大得多)的依赖，这改进了从[AAK19]绑定的$O(rQ^{1/r}\log Q\log(mc_{max}))$ 。此外，我们的结果几乎匹配 [AAK19] 的$\Omega(\frac1{r^3}Q^{1/r})$的下界。设置$r=\log Q$显示$O(\log Q)$自适应回合给出$O(\log Q)$近似。通过使用 Theorem A.2 将该算法转换为基于集合的解决方案，我们得到：
2
|
[
[
0.146240234375,
0.0947265625,
0.88134765625,
0.0947265625,
0.88134765625,
0.10797119140625,
0.146240234375,
0.10797119140625
],
[
0.11767578125,
0.11236572265625,
0.8798828125,
0.11236572265625,
0.8798828125,
0.126953125,
0.11767578125,
0.126953125
],
[
0.11767578125,
0.1300048828125,
0.88134765625,
0.1287841796875,
0.88134765625,
0.143310546875,
0.11767578125,
0.14453125
],
[
0.11767578125,
0.1483154296875,
0.88134765625,
0.1483154296875,
0.88134765625,
0.1622314453125,
0.11767578125,
0.1622314453125
],
[
0.11767578125,
0.1666259765625,
0.88134765625,
0.1666259765625,
0.88134765625,
0.18115234375,
0.11767578125,
0.18115234375
],
[
0.1168212890625,
0.184326171875,
0.8798828125,
0.184326171875,
0.8798828125,
0.1976318359375,
0.1168212890625,
0.1976318359375
],
[
0.11767578125,
0.2020263671875,
0.615234375,
0.2020263671875,
0.615234375,
0.216552734375,
0.11767578125,
0.216552734375
],
[
0.146240234375,
0.2197265625,
0.88134765625,
0.2197265625,
0.88134765625,
0.23291015625,
0.146240234375,
0.23291015625
],
[
0.1192626953125,
0.238037109375,
0.88134765625,
0.238037109375,
0.88134765625,
0.251953125,
0.1192626953125,
0.251953125
],
[
0.1168212890625,
0.255126953125,
0.88134765625,
0.255615234375,
0.88134765625,
0.270751953125,
0.1168212890625,
0.26953125
],
[
0.1192626953125,
0.2734375,
0.438720703125,
0.2734375,
0.438720703125,
0.287353515625,
0.1192626953125,
0.287353515625
],
[
0.1446533203125,
0.292236328125,
0.88134765625,
0.292236328125,
0.88134765625,
0.30615234375,
0.1446533203125,
0.30615234375
],
[
0.11767578125,
0.309326171875,
0.88134765625,
0.309326171875,
0.88134765625,
0.323974609375,
0.11767578125,
0.323974609375
],
[
0.11767578125,
0.32763671875,
0.88134765625,
0.32763671875,
0.88134765625,
0.341552734375,
0.11767578125,
0.341552734375
],
[
0.11767578125,
0.34521484375,
0.88134765625,
0.34521484375,
0.88134765625,
0.359130859375,
0.11767578125,
0.359130859375
],
[
0.11767578125,
0.3642578125,
0.5810546875,
0.3642578125,
0.5810546875,
0.37744140625,
0.11767578125,
0.37744140625
],
[
0.1446533203125,
0.3818359375,
0.88134765625,
0.3818359375,
0.88134765625,
0.395751953125,
0.1446533203125,
0.395751953125
],
[
0.1192626953125,
0.399658203125,
0.88134765625,
0.399658203125,
0.88134765625,
0.4140625,
0.1192626953125,
0.4140625
],
[
0.1168212890625,
0.417236328125,
0.8798828125,
0.417236328125,
0.8798828125,
0.43115234375,
0.1168212890625,
0.43115234375
],
[
0.1168212890625,
0.436279296875,
0.88134765625,
0.436279296875,
0.88134765625,
0.449462890625,
0.1168212890625,
0.449462890625
],
[
0.12091064453125,
0.453857421875,
0.53515625,
0.453857421875,
0.53515625,
0.46728515625,
0.12091064453125,
0.46728515625
],
[
0.136474609375,
0.48095703125,
0.88232421875,
0.479736328125,
0.88232421875,
0.49755859375,
0.136474609375,
0.498779296875
],
[
0.156005859375,
0.49755859375,
0.6005859375,
0.50048828125,
0.6005859375,
0.51806640625,
0.156005859375,
0.51513671875
],
[
0.1396484375,
0.52978515625,
0.88134765625,
0.52978515625,
0.88134765625,
0.54443359375,
0.1396484375,
0.54443359375
],
[
0.158447265625,
0.54541015625,
0.5537109375,
0.54443359375,
0.5537109375,
0.5625,
0.158447265625,
0.56298828125
],
[
0.1168212890625,
0.57763671875,
0.78125,
0.57763671875,
0.78125,
0.59130859375,
0.1168212890625,
0.59130859375
],
[
0.148681640625,
0.59521484375,
0.88134765625,
0.59521484375,
0.88134765625,
0.609375,
0.148681640625,
0.609375
],
[
0.1192626953125,
0.61279296875,
0.88134765625,
0.61279296875,
0.88134765625,
0.626953125,
0.1192626953125,
0.626953125
],
[
0.1168212890625,
0.630859375,
0.88134765625,
0.6298828125,
0.88134765625,
0.64453125,
0.1168212890625,
0.64599609375
],
[
0.11767578125,
0.6484375,
0.88134765625,
0.6484375,
0.88134765625,
0.66357421875,
0.11767578125,
0.66357421875
],
[
0.11517333984375,
0.66552734375,
0.465576171875,
0.66650390625,
0.465576171875,
0.68115234375,
0.11517333984375,
0.6796875
],
[
0.1446533203125,
0.68310546875,
0.88232421875,
0.68310546875,
0.88232421875,
0.70068359375,
0.1446533203125,
0.70068359375
],
[
0.11767578125,
0.70263671875,
0.88134765625,
0.70263671875,
0.88134765625,
0.71630859375,
0.11767578125,
0.71630859375
],
[
0.1192626953125,
0.7216796875,
0.8798828125,
0.7216796875,
0.8798828125,
0.73486328125,
0.1192626953125,
0.73486328125
],
[
0.1168212890625,
0.7392578125,
0.348876953125,
0.7392578125,
0.348876953125,
0.75244140625,
0.1168212890625,
0.75244140625
],
[
0.1168212890625,
0.76806640625,
0.88134765625,
0.76806640625,
0.88134765625,
0.78271484375,
0.1168212890625,
0.78271484375
],
[
0.1168212890625,
0.7861328125,
0.88232421875,
0.7861328125,
0.88232421875,
0.80029296875,
0.1168212890625,
0.80029296875
],
[
0.11767578125,
0.8046875,
0.271240234375,
0.8046875,
0.271240234375,
0.818359375,
0.11767578125,
0.818359375
],
[
0.146240234375,
0.833984375,
0.88134765625,
0.833984375,
0.88134765625,
0.84765625,
0.146240234375,
0.84765625
],
[
0.1168212890625,
0.849609375,
0.88232421875,
0.849609375,
0.88232421875,
0.8671875,
0.1168212890625,
0.8671875
],
[
0.11767578125,
0.869140625,
0.87841796875,
0.869140625,
0.87841796875,
0.8837890625,
0.11767578125,
0.8837890625
],
[
0.11767578125,
0.88720703125,
0.374267578125,
0.88720703125,
0.374267578125,
0.90087890625,
0.11767578125,
0.90087890625
],
[
0.491943359375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"随机子模块覆盖的解决方案现在是顺序决策过程。在每个步骤中，一个项目是probed及其实",
"现（e。g。，活动或不活动）被观察到。该过程通常是自适应的，其中来自先前探测的项目的所有信",
"息都用于识别要探测的下一个项目。此过程一直持续到涵盖子模块功能为止，目标是最大程度地降",
"低被探测项目的预期成本。这种自适应解决方案本质上是完全顺序的，如果探测一个项目是耗时",
"的，这是不可取的。E。g。，在传感器部署中，探测传感器对应于物理部署传感器并观察其是否按",
"预期运行，这可能需要数小时或数天。或者，在医疗诊断中探测/执行测试可能需要长时间等待测",
"试结果。因此，我们更喜欢只有很少rounds适应性的解决方案",
"受此启发，我们问：对于随机子模块覆盖问题，只有几轮自适应回合的解决方案如何近似完全",
"自适应解决方案？我们考虑两种情况，其中不同项目的实现是独立的，并且允许它们相互关联。对",
"于这两种情况，我们给出了几乎严格的答案，在自适应回合的数量r和解决方案质量（相对于完全",
"自适应解决方案）之间进行了平滑的权衡",
"我们工作的主要贡献是针对自适应回合的“基于集合”模型中的随机子模块覆盖的回合自适",
"应解决方案。在这个模型中，每一轮都会并行探测一个固定的项目子集（并且会产生它们的总成",
"本)。本轮的决定可能取决于前几轮的实现。但是，如[AAK19]中所述，如果我们要求函数f被概",
"率1覆盖，则r-round-adaptivity差距会非常大。（有关示例，请参见sA。）因此，我们专注于基",
"于集合的解决方案，这些解决方案仅需要以高概率覆盖函数",
"在设计算法时，事实证明使用“置换”模型进行自适应回合更为方便，其中函数覆盖了概率",
"一。此模型也用于先前的文献[GV06,AAK19]】。在这里，r-round-adaptive解决方案的每一轮",
"都指定所有剩余项目的顺序，并按此顺序探测它们，直到某个停止规则。有关正式定义，请参见",
"Definition2.1。此外，我们在置换模型中的r轮自适应算法可以转换为基于集合的模型中的r轮",
"自适应算法。我们在基于集合的模型中获得以下算法：",
": for any n E (o, 1) , finds an r -round adaptive solution that has expected cost at most ra .OPT",
"and covers the function with probability at least 1 - n .",
": finds an O(r) -round adaptive solution that has expected cost at most O(α) - OpT and covers",
"the function with probability at least 1 - e-2(r).",
"这里OPT是最佳完全自适应解决方案的成本，α是我们的算法在置换模型中的近似比",
"上面的第一个算法是针对，自适应的轮数较小（例如，一个常数）的情况。在这一点上，我",
"们保持轮数相同，但我们在预期成本中损失了一个因子r。第二种算法是针对很大的情况，例",
"如。g。，超过一个常数。在这里，基于集合的轮数增加了2倍，但我们只损失了预期成本中的一个",
"常数因子。我们在SA中形式化并证明了这些结果。对于本文的其余部分，r轮自适应算法是指置",
"换模型中的r轮自适应算法（除非另有说明）",
"考虑具有f：2U→Z≥0。有m项，其中每个项i是一个随机变量，其成本为ci，并对应",
"于U的随机元素。（我们的结果扩展到更一般的设置，其中每个项目都实现了U的子集。）的目标",
"是选择一组项目，以使其相应实现的联合S满足f（S）=Q，并且预期成本最小化。我们的第一个",
"结果是当项目具有独立分布时",
"Theorem 1.1 (Independent Items). For any integer r ≥ 1 , there is an r -round adaptive algorithm",
"for the stochastic submodular cover problem with cost O(Q'/r.logQ) times the cost of an optimal",
"adaptivealgorithm.",
"通过消除对项目m的数量和项目成本（可能比Q大得多）的依赖，这改进了从【AAK19]绑定",
"的O(rQ1/rlogQlog(mcmar）。此外，我们的结果几乎匹配[AAK19]的2(Q1/r）的下界。设置",
"r=logQ显示O(logQ）自适应回合给出O(logQ）近似。通过使用TheoremA.2将该算法转换为",
"基于集合的解决方案，我们得到：",
"2"
] |
{
"arxiv_id": "2106.16115",
"language": "zh",
"page_idx": 1
}
|
||
2106.16115-zh-2
|
2106.16115-zh-2.jpg
|
$\textbf{Corollary 1.2 (Independent Items: Set-Based Model).There is an O(\log Q) round algorithm for}$ stochastic submodular cover $in$ the $set- based$ model with cost $O( \log Q)$ times the optimal $( fully)$ $adaptive\textit{cost. }$
即使具有任意数量$(r=m$ )的自适应回合，$O(\log Q)$的近似比率也是最好的 (除非 P=NP)。以前，只能在对数回合[AAK19] 中获得$O(\log^2Q\log(mc_{max}))$近似，或者在$O(\log m\log(Qmc_{\max}))$ 基于集合的回合 [EKM19] 中获得$O(\log(mQc_\mathrm{max}))$近似。
此外，Theorem 1.1 (带有$r=1$ )暗示了随机集合覆盖的$O(Q\log Q)$适应性差距 (子模块覆盖的特殊情况),解决了从$\operatorname{E}[\operatorname{GV06}]$到$O(\log Q)$因子的悬而未决的问题，其中$Q$是集合覆盖中的对象数
我们的第二组结果是针对项目具有相关分布的情况。令$s$表示联合分布$\mathcal{D}$的支持大小，i。e。,
场景的数量。
Theorem 1.3 (Correlated Items). $Foranyintegerr\geq1,thereisanr$ -roundadaptivealgorithmforscenario submodular cover with cost $O\left ( s^{1/ r}( \log s+ r\log Q) \right )$ times the cost $of$ $an$ optimal $adaptive\textit{ algorithm}.$
我们还获得了$2r$轮自适应算法，其成本保证为$O\left(s^{1/r}\log(sQ)\right)$倍于最佳自适应算法的成本(请参阅推论$\mathsf E6.6$ )。结合转换为基于集合的解决方案 (Theorem A.2 )并设置 $r=\log s$ ,我们可以推断出：
Corollary 1.4 (Correlated Items: Set-Based Model). $ThereisanO( \log s) \textit{round algorithm for}$ scenario submodular cover $in$ the $set- based$ model with cost $O( \log ( sQ) )$ times the optimal $( fully)$ $adaptive\textit{cost. }$
即使使用任意数量的自适应回合，上述近似保证也几乎是最好的：存在$\Omega(\log s)$近似因子硬度$\overline{\mathrm{E}}[$CPR$^+11]$ 。场景子模覆盖推广经典最优决策树问题 [GG74,HR77,Lov85,KPB99,Das04, AH12,GNR17]。在[GHKL16] 中获得了场景子模覆盖的完全自适应$O(\log(sQ))$近似；另请参见 [NKN20] 以获取更一般的结果。就自适应性而言，$O(\log m\log(Qm\frac{\mathrm{c}_{\mathrm{max}}}{p_{\mathrm{min}}}))$基于集合的回合中的$O(\log(mQ\frac{c_\mathrm{max}}{p_\mathrm{min}}))$近似值来自 [GHKL16,EKM19]。这里$p_\mathrm{min}\leq\frac1\mathrm{s}$是任何场景的最小概率。我们注意到，当轮数小于对数时，即使在经过充分研究的最优决策树的特殊情况下，我们的结果也提供了第一近似保证
Theorem 1.1 和 Theorem 1.3中的结果是无与伦比的：尽管独立情况在分布$\mathcal{D}$中具有更多的结构，但其支持大小是指数级的。最后，在相关设置中，对支持大小$s$ 的依赖是必要的，正如我们的下一个结果显示的那样。
Theorem 1.5 (Lower Bound). $Foranyintegerr\geq1,thereisaninstanceofscenariosubmodular$ cover with $Q= 1$ where the cost $of$ any $r$ $- round$ adaptive solution $is$ $\Omega ( \frac 1{r^{2}\log s}\cdot s^{1/ r})$ times the optimal adaptive $cost.$
这个下界是信息论的，不依赖于计算假设，而 Theorem 1.3 的上界是由多项式算法给出的
最后，我们注意到我们的算法也很容易实现。我们在合成和真实数据集上测试了这两种算法， 验证了我们算法的实际性能。具体来说，我们在随机集合覆盖的实例上测试了独立情况(定理|E|1.1 )的算法，并在最优决策树的实例上测试了相关情况(定理|El.3 )的算法。对于随机集覆盖，我们使用真实世界的数据集来生成具有$\approx1200$ 项的实例。我们观察到在几轮适应性中性能的急剧改善，并且 6-7 轮适应性几乎与完全自适应解决方案一样好。对于最优决策树，我们同时使用真实世界的数据和合成数据。真实世界的数据有$\approx400$ 场景，合成数据有 10,000 场景。同样，我们发现大约 6 轮适应性足以获得与完全适应性解决方案一样好的解决方案。我们还将算法的成本与两个应用程序的信息理论下限进行了比较：我们的成本通常在这些下限的 50% 之内
3
|
[
[
0.11767578125,
0.09344482421875,
0.8798828125,
0.09344482421875,
0.8798828125,
0.10797119140625,
0.11767578125,
0.10797119140625
],
[
0.1192626953125,
0.11236572265625,
0.8798828125,
0.11236572265625,
0.8798828125,
0.1256103515625,
0.1192626953125,
0.1256103515625
],
[
0.12091064453125,
0.1319580078125,
0.221435546875,
0.1319580078125,
0.221435546875,
0.143310546875,
0.12091064453125,
0.143310546875
],
[
0.146240234375,
0.1572265625,
0.88134765625,
0.1572265625,
0.88134765625,
0.17041015625,
0.146240234375,
0.17041015625
],
[
0.1192626953125,
0.1749267578125,
0.87841796875,
0.1749267578125,
0.87841796875,
0.189453125,
0.1192626953125,
0.189453125
],
[
0.11517333984375,
0.1900634765625,
0.564453125,
0.1925048828125,
0.564453125,
0.208984375,
0.11517333984375,
0.20703125
],
[
0.146240234375,
0.210205078125,
0.8798828125,
0.210205078125,
0.8798828125,
0.2247314453125,
0.146240234375,
0.2247314453125
],
[
0.1192626953125,
0.2291259765625,
0.8798828125,
0.2291259765625,
0.8798828125,
0.24365234375,
0.1192626953125,
0.24365234375
],
[
0.1168212890625,
0.246826171875,
0.17236328125,
0.246826171875,
0.17236328125,
0.261474609375,
0.1168212890625,
0.261474609375
],
[
0.1422119140625,
0.268310546875,
0.88232421875,
0.270751953125,
0.88232421875,
0.28857421875,
0.1422119140625,
0.285888671875
],
[
0.1168212890625,
0.287353515625,
0.2156982421875,
0.287353515625,
0.2156982421875,
0.3017578125,
0.1168212890625,
0.3017578125
],
[
0.11358642578125,
0.3125,
0.8798828125,
0.313232421875,
0.8798828125,
0.330810546875,
0.11358642578125,
0.330078125
],
[
0.11767578125,
0.333251953125,
0.88232421875,
0.333251953125,
0.88232421875,
0.347900390625,
0.11767578125,
0.347900390625
],
[
0.1192626953125,
0.35107421875,
0.268798828125,
0.35107421875,
0.268798828125,
0.3642578125,
0.1192626953125,
0.3642578125
],
[
0.146240234375,
0.376953125,
0.8798828125,
0.376953125,
0.8798828125,
0.391357421875,
0.146240234375,
0.391357421875
],
[
0.11517333984375,
0.39453125,
0.8798828125,
0.395751953125,
0.8798828125,
0.410400390625,
0.11517333984375,
0.4091796875
],
[
0.11767578125,
0.41357421875,
0.1968994140625,
0.41357421875,
0.1968994140625,
0.427978515625,
0.11767578125,
0.427978515625
],
[
0.11767578125,
0.43994140625,
0.87841796875,
0.43994140625,
0.87841796875,
0.455078125,
0.11767578125,
0.455078125
],
[
0.11767578125,
0.458251953125,
0.8798828125,
0.458251953125,
0.8798828125,
0.472900390625,
0.11767578125,
0.472900390625
],
[
0.12091064453125,
0.477783203125,
0.221435546875,
0.477783203125,
0.221435546875,
0.4892578125,
0.12091064453125,
0.4892578125
],
[
0.146240234375,
0.50244140625,
0.88134765625,
0.50244140625,
0.88134765625,
0.5166015625,
0.146240234375,
0.5166015625
],
[
0.1168212890625,
0.51806640625,
0.8798828125,
0.51953125,
0.8798828125,
0.5341796875,
0.1168212890625,
0.53271484375
],
[
0.11767578125,
0.53857421875,
0.88134765625,
0.53857421875,
0.88134765625,
0.55322265625,
0.11767578125,
0.55322265625
],
[
0.1168212890625,
0.55615234375,
0.88232421875,
0.55615234375,
0.88232421875,
0.57373046875,
0.1168212890625,
0.57373046875
],
[
0.1168212890625,
0.57470703125,
0.88232421875,
0.57470703125,
0.88232421875,
0.59228515625,
0.1168212890625,
0.59228515625
],
[
0.11767578125,
0.59326171875,
0.87841796875,
0.59326171875,
0.87841796875,
0.60791015625,
0.11767578125,
0.60791015625
],
[
0.1168212890625,
0.6123046875,
0.2254638671875,
0.6123046875,
0.2254638671875,
0.62548828125,
0.1168212890625,
0.62548828125
],
[
0.1446533203125,
0.6298828125,
0.88134765625,
0.6298828125,
0.88134765625,
0.64453125,
0.1446533203125,
0.64453125
],
[
0.1168212890625,
0.646484375,
0.88134765625,
0.64794921875,
0.88134765625,
0.662109375,
0.1168212890625,
0.6611328125
],
[
0.11767578125,
0.66650390625,
0.303955078125,
0.66650390625,
0.303955078125,
0.6796875,
0.11767578125,
0.6796875
],
[
0.1168212890625,
0.69140625,
0.88134765625,
0.69140625,
0.88134765625,
0.708984375,
0.1168212890625,
0.708984375
],
[
0.111083984375,
0.70556640625,
0.8857421875,
0.7041015625,
0.8857421875,
0.73193359375,
0.111083984375,
0.734375
],
[
0.1192626953125,
0.72900390625,
0.28759765625,
0.72900390625,
0.28759765625,
0.7421875,
0.1192626953125,
0.7421875
],
[
0.146240234375,
0.7548828125,
0.84326171875,
0.7548828125,
0.84326171875,
0.76806640625,
0.146240234375,
0.76806640625
],
[
0.143798828125,
0.771484375,
0.8798828125,
0.77294921875,
0.8798828125,
0.787109375,
0.143798828125,
0.7861328125
],
[
0.1168212890625,
0.79052734375,
0.88134765625,
0.7890625,
0.88134765625,
0.8037109375,
0.1168212890625,
0.8046875
],
[
0.11767578125,
0.8095703125,
0.88134765625,
0.8095703125,
0.88134765625,
0.82275390625,
0.11767578125,
0.82275390625
],
[
0.11767578125,
0.8271484375,
0.88134765625,
0.8271484375,
0.88134765625,
0.84033203125,
0.11767578125,
0.84033203125
],
[
0.11767578125,
0.8447265625,
0.88134765625,
0.8447265625,
0.88134765625,
0.85791015625,
0.11767578125,
0.85791015625
],
[
0.1192626953125,
0.8623046875,
0.88134765625,
0.8623046875,
0.88134765625,
0.87548828125,
0.1192626953125,
0.87548828125
],
[
0.11767578125,
0.880859375,
0.88134765625,
0.880859375,
0.88134765625,
0.89453125,
0.11767578125,
0.89453125
],
[
0.1168212890625,
0.896484375,
0.7490234375,
0.8984375,
0.7490234375,
0.91357421875,
0.1168212890625,
0.9111328125
],
[
0.491943359375,
0.93505859375,
0.5087890625,
0.93505859375,
0.5087890625,
0.9521484375,
0.491943359375,
0.9521484375
]
] |
[
"Corollary 1.2 (Independent Items: Set-Based Model). There is an O(log Q) round algorithm fon",
"stochastic submodular cover in the set-based model with cost O(logQ) times the optimal (fully)",
"adaptivecost.",
"即使具有任意数量（r=m）的自适应回合，O(logQ）的近似比率也是最好的（除非P=NP）。以",
"前，只能在对数回合[AAK19]中获得O(log~Qlog（mCmar））近似，或者在O(logmlog（QmCmax）",
"基于集合的回合[EKM19]中获得O(log(mQcmax)）近似。",
"此外，Theorem1.1（带有r=1）暗示了随机集合覆盖的O(QlogQ）适应性差距（子模块覆",
"盖的特殊情况），解决了从E[GV06]到O(1ogQ）因子的悬而未决的问题，其中Q是集合覆盖中的",
"对象数",
"我们的第二组结果是针对项目具有相关分布的情况。令s表示联合分布D的支持大小，i。e。",
"场景的数量。",
"Theorem 1.3 (Correlated Items). For any integer r ≥ 1 , there is an r -round adaptive algo",
"rithm for scenario submodular cover with cost O (sl/r(log s + r log Q)) times the cost of an optimal",
"adaptivealgorithm.",
"我们还获得了2r轮自适应算法，其成本保证为O（s1/rlog(sQ））倍于最佳自适应算法的成本",
"（请参阅推论E6.6）。结合转换为基于集合的解决方案（TheoremA.2）并设置r=logs，我们可",
"以推断出：",
"Corollary 1.4 (Correlated Items: Set-Based Model). There is an O(log s) round algorithm for",
"scenario submodular cover in the set-based model with cost O(log(sQ)) times the optimal (fully)",
"adaptivecost",
"即使使用任意数量的自适应回合，上述近似保证也几乎是最好的：存在2(logs）近似因子",
"硬度E[CPR+11]。场景子模覆盖推广经典最优决策树问题[GG74,HR77,Lov85,KPB99,Das04",
"AH12,GNR17]。在[GHKL16]中获得了场景子模覆盖的完全自适应O(log(sQ)）近似；另请参",
"见[NKN20]】以获取更一般的结果。就自适应性而言，O(logmlog（Qmgmax））基于集合的回合中的",
"O(log(mQmax)）近似值来自[GHKL16,EKM19]。这里Pmin≤是任何场景的最小概率。我们注",
"意到，当轮数小于对数时，即使在经过充分研究的最优决策树的特殊情况下，我们的结果也提供了",
"第一近似保证",
"Theorem1.1和Theorem1.3中的结果是无与伦比的：尽管独立情况在分布D中具有更多的结",
"构，但其支持大小是指数级的。最后，在相关设置中，对支持大小s的依赖是必要的，正如我们的",
"下一个结果显示的那样。",
"Theorem 1.5 (Lower Bound). For any integer r ≥ 1 , there is an instance of scenario submodular",
"cover with Q=1 where the cost of anyr-round adaptive solution is (log·s/r) times the",
"optimal adaptivecost.",
"这个下界是信息论的，不依赖于计算假设，而Theorem1.3的上界是由多项式算法给出的",
"最后，我们注意到我们的算法也很容易实现。我们在合成和真实数据集上测试了这两种算法",
"验证了我们算法的实际性能。具体来说，我们在随机集合覆盖的实例上测试了独立情况（定理E1.1",
"）的算法，并在最优决策树的实例上测试了相关情况（定理E1.3）的算法。对于随机集覆盖，我",
"们使用真实世界的数据集来生成具有～1200项的实例。我们观察到在几轮适应性中性能的急剧改",
"善，并且6－7轮适应性几乎与完全自适应解决方案一样好。对于最优决策树，我们同时使用真实",
"世界的数据和合成数据。真实世界的数据有～400场景，合成数据有10,000场景。同样，我们发",
"现大约6轮适应性足以获得与完全适应性解决方案一样好的解决方案。我们还将算法的成本与两",
"个应用程序的信息理论下限进行了比较：我们的成本通常在这些下限的50%之内",
"3"
] |
{
"arxiv_id": "2106.16115",
"language": "zh",
"page_idx": 2
}
|
||
2106.16115-zh-3
|
2106.16115-zh-3.jpg
|
1.0.1 Techniques
独立和相关情况的算法相似，但有一些关键差异。在这两种算法的每一轮中，我们迭代地计算每个项目的“分数”,并贪婪地选择得分最高的项目。这会导致所有剩余项目的非自适应列表，并且这些项目按此顺序为probed ,直到停止规则。独立情况下的 PARCA 停止规则对应于将剩余目标(在函数值上)减少$Q^1/r$ ,而 SPARCA 规则涉及在相关情况下通过 $s^{1/r}$ 因素减少“兼容方案”的数量
定理$\boxed{\mathrm{E}}1.1$和 1.3 的分析在开始时遵循平行线。对于每个$i\geq0$ ,我们将成本$\alpha\cdot2^i$后算法的“未完成”概率与成本$2^i$后的最佳自适应解决方案相关联。“拉伸”因子$\alpha$对应于近似比，对于独立和相关的情况，近似比不同。为了关联这些未完成概率，我们考虑了算法在成本$\alpha2^{i-1}$和$\alpha2^i$之间选择的项目的总分$G$ 。分析的关键在于给出总分$G$的下限和上限；对于独立和相关的设置，这里的参数是不同的
在独立情况下，任何项目$\chi_e$的得分都是其相对边际收益的估计值，其中我们对所有先前项目以及$\mathcal{X}_e$进行期望。我们还通过项目的成本将这一收益正常化。定义见公式(1)。在总分$G$的下限中，我们使用 [AAK19] 的采样引理的变体以及子模最大化E[BSZ19] 的常数因子适应性间隙。我们还需要将 (所有先前实现的)结果空间划分为“好”和“坏”结果：以良好的结果为条件，OPT 在成本$2^i$之前完成的可能性很高。在我们对抽样引理的证明中，好的结果是必要的，但幸运的是坏结果的总概率很小(并且可以有效地忽略它们)。在上限$G$ 中，我们将总分视为决策/样本路径的总和，并使用相对增益的总和对应于谐波序列的事实
在相关情况下，任何项目$\chi_e$的得分都是两项的总和 (i)独立情况下的预期相对边际收益，以及 (ii)“消除”情况下的概率估计。这两个术语都是需要的，因为算法需要平衡(i)覆盖函数和(ii)识别实现的场景 (之后覆盖$f$是微不足道的)。同样，我们通过项目的成本进行标准化。参见方程 (16) 。在总分$G$的下限中，我们将结果空间划分为好/好/坏的结果，这些结果对应于 OPT 的高条件概率 (i) 通过成本$2^i$覆盖函数$f$ ,(ii) 通过成本$2^i$消除恒定比例的场景，或 (iii)两种情况都不是。此外，通过限制具有“大量”兼容场景的结果(否则，将应用算法的停止规则),我们可以通过$s^{1/r}$限制“相关”结果的数量。然后，我们考虑 OPT (最高成本$2^i$ )以所有良好/良好的结果为条件，并表明其中一项得分很高。对于上限$G$ ,我们再次将总分视为决策路径上的总和
# 1.1 Related Work
在 [Wol82] 中获得了基本子模覆盖问题的$(1+\ln Q)$近似算法。这也是最好的可能(除非 P=NP), 因为设置覆盖问题是一种特殊情况$\mathbb{E}[$Fei98]。在过去的几年中，已经有许多关于 stochastic 变种的子模块封面[GK11,INvdZ16,DHK16,GHKL16,NKN20] 的论文，因为这个框架捕获了许多不同的应用程序
首先在 [GV06] 中研究了随机集覆盖问题，该问题表明适应性差距位于$\Omega(d)$和$O(d^2)$之间，其中$d$是要覆盖的对象数。最近，[AAK19] 将上限改进为$O(d\log d\log(mc_{max}))$ 。作为定理 1.1 的推论，我们得到了更紧的$O(d\log d)$适应性缺口。重要的是，我们消除了项目$m$和最大成本的依赖性，这在$d$中甚至可能是指数的
从 [INvdZ16] 得出随机子模块覆盖 (独立项) 的$O(\log Q)$自适应近似算法。在 [LPRY08,GK11, GK17,DHK16] 中获得了特殊情况或边界较弱的相关结果。如前所述，在[AAK19] 中获得了具有有限轮次适应性的算法的上/下界。定理 1.1 通过$O(r\cdot\log(mc_{max}))$因子对这些上界进行了改进此外，我们的算法是贪婪的风格和更简单的实现。我们绕过先前工作中计算昂贵的步骤，例如解决随机子模最大化的几个实例。我们的分析(上面概述的)也是非常不同的。我们使用[AAK19]中的关键采样引理的变体，但是它以不同的方式应用，并且仅影响分析。我们对总分进行下限/上限的高级方法类似于 [INvdZ16] 中对完全自适应问题的分析。但是，细节非常不同，因为我们需要处理$r$ -round-adaptivitygap 的问题
[GHKL16]中引入了场景子模覆盖问题，作为包括最优决策树[GG74,GNR17],等价类确
4
|
[
[
0.1168212890625,
0.0947265625,
0.283447265625,
0.0947265625,
0.283447265625,
0.10919189453125,
0.1168212890625,
0.10919189453125
],
[
0.1168212890625,
0.11932373046875,
0.88134765625,
0.12060546875,
0.88134765625,
0.1351318359375,
0.1168212890625,
0.1337890625
],
[
0.11767578125,
0.1395263671875,
0.88134765625,
0.1395263671875,
0.88134765625,
0.15283203125,
0.11767578125,
0.15283203125
],
[
0.11767578125,
0.1572265625,
0.88134765625,
0.1572265625,
0.88134765625,
0.17041015625,
0.11767578125,
0.17041015625
],
[
0.1168212890625,
0.17236328125,
0.88232421875,
0.17236328125,
0.88232421875,
0.1900634765625,
0.1168212890625,
0.1900634765625
],
[
0.1446533203125,
0.1925048828125,
0.88134765625,
0.1925048828125,
0.88134765625,
0.205810546875,
0.1446533203125,
0.205810546875
],
[
0.1168212890625,
0.2108154296875,
0.88134765625,
0.2108154296875,
0.88134765625,
0.2247314453125,
0.1168212890625,
0.2247314453125
],
[
0.11767578125,
0.2291259765625,
0.8798828125,
0.2291259765625,
0.8798828125,
0.242431640625,
0.11767578125,
0.242431640625
],
[
0.12091064453125,
0.2474365234375,
0.88134765625,
0.2474365234375,
0.88134765625,
0.261474609375,
0.12091064453125,
0.261474609375
],
[
0.1168212890625,
0.26318359375,
0.2607421875,
0.264404296875,
0.2607421875,
0.279052734375,
0.1168212890625,
0.27783203125
],
[
0.1446533203125,
0.2822265625,
0.8798828125,
0.2822265625,
0.8798828125,
0.296630859375,
0.1446533203125,
0.296630859375
],
[
0.1192626953125,
0.300537109375,
0.88134765625,
0.300537109375,
0.88134765625,
0.31494140625,
0.1192626953125,
0.31494140625
],
[
0.11767578125,
0.317626953125,
0.88134765625,
0.318115234375,
0.88134765625,
0.333251953125,
0.11767578125,
0.33203125
],
[
0.11767578125,
0.3359375,
0.88134765625,
0.3359375,
0.88134765625,
0.349853515625,
0.11767578125,
0.349853515625
],
[
0.11767578125,
0.354248046875,
0.8798828125,
0.354248046875,
0.8798828125,
0.367431640625,
0.11767578125,
0.367431640625
],
[
0.1168212890625,
0.37255859375,
0.88134765625,
0.37255859375,
0.88134765625,
0.386474609375,
0.1168212890625,
0.386474609375
],
[
0.11767578125,
0.39013671875,
0.54736328125,
0.39013671875,
0.54736328125,
0.404052734375,
0.11767578125,
0.404052734375
],
[
0.143798828125,
0.40771484375,
0.88134765625,
0.4091796875,
0.88134765625,
0.423583984375,
0.143798828125,
0.42236328125
],
[
0.11767578125,
0.4267578125,
0.8798828125,
0.4267578125,
0.8798828125,
0.43994140625,
0.11767578125,
0.43994140625
],
[
0.1168212890625,
0.4443359375,
0.8798828125,
0.4443359375,
0.8798828125,
0.458251953125,
0.1168212890625,
0.458251953125
],
[
0.11767578125,
0.462158203125,
0.8798828125,
0.462158203125,
0.8798828125,
0.47607421875,
0.11767578125,
0.47607421875
],
[
0.1192626953125,
0.48095703125,
0.88134765625,
0.48095703125,
0.88134765625,
0.494384765625,
0.1192626953125,
0.494384765625
],
[
0.1168212890625,
0.498779296875,
0.88134765625,
0.498779296875,
0.88134765625,
0.51220703125,
0.1168212890625,
0.51220703125
],
[
0.11767578125,
0.51513671875,
0.88134765625,
0.51513671875,
0.88134765625,
0.52978515625,
0.11767578125,
0.52978515625
],
[
0.11767578125,
0.53271484375,
0.87353515625,
0.5341796875,
0.87353515625,
0.548828125,
0.11767578125,
0.54736328125
],
[
0.1168212890625,
0.5712890625,
0.306396484375,
0.5712890625,
0.306396484375,
0.58544921875,
0.1168212890625,
0.58544921875
],
[
0.11767578125,
0.5986328125,
0.88134765625,
0.5986328125,
0.88134765625,
0.6123046875,
0.11767578125,
0.6123046875
],
[
0.1192626953125,
0.6171875,
0.88134765625,
0.6171875,
0.88134765625,
0.630859375,
0.1192626953125,
0.630859375
],
[
0.11767578125,
0.6337890625,
0.8798828125,
0.6337890625,
0.8798828125,
0.6484375,
0.11767578125,
0.6484375
],
[
0.11767578125,
0.65283203125,
0.2034912109375,
0.65283203125,
0.2034912109375,
0.6640625,
0.11767578125,
0.6640625
],
[
0.146240234375,
0.67041015625,
0.88134765625,
0.67041015625,
0.88134765625,
0.68505859375,
0.146240234375,
0.68505859375
],
[
0.11517333984375,
0.68603515625,
0.88232421875,
0.6875,
0.88232421875,
0.705078125,
0.11517333984375,
0.7041015625
],
[
0.1168212890625,
0.70703125,
0.88134765625,
0.70703125,
0.88134765625,
0.72021484375,
0.1168212890625,
0.72021484375
],
[
0.11767578125,
0.724609375,
0.386474609375,
0.724609375,
0.386474609375,
0.73779296875,
0.11767578125,
0.73779296875
],
[
0.146240234375,
0.7421875,
0.88134765625,
0.7421875,
0.88134765625,
0.7568359375,
0.146240234375,
0.7568359375
],
[
0.1168212890625,
0.7587890625,
0.8798828125,
0.76025390625,
0.8798828125,
0.7744140625,
0.1168212890625,
0.7734375
],
[
0.11767578125,
0.77880859375,
0.87744140625,
0.77880859375,
0.87744140625,
0.79345703125,
0.11767578125,
0.79345703125
],
[
0.11767578125,
0.796875,
0.88134765625,
0.796875,
0.88134765625,
0.81005859375,
0.11767578125,
0.81005859375
],
[
0.11767578125,
0.81298828125,
0.8798828125,
0.814453125,
0.8798828125,
0.8291015625,
0.11767578125,
0.82763671875
],
[
0.11767578125,
0.83056640625,
0.8798828125,
0.83203125,
0.8798828125,
0.8466796875,
0.11767578125,
0.84521484375
],
[
0.11767578125,
0.85107421875,
0.88134765625,
0.85107421875,
0.88134765625,
0.8642578125,
0.11767578125,
0.8642578125
],
[
0.1168212890625,
0.86865234375,
0.35546875,
0.86865234375,
0.35546875,
0.88330078125,
0.1168212890625,
0.88330078125
],
[
0.1544189453125,
0.88623046875,
0.8798828125,
0.88623046875,
0.8798828125,
0.90087890625,
0.1544189453125,
0.90087890625
],
[
0.491943359375,
0.93603515625,
0.5087890625,
0.93603515625,
0.5087890625,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"1.0.1Techniques",
"独立和相关情况的算法相似，但有一些关键差异。在这两种算法的每一轮中，我们迭代地计算每个",
"项目的“分数”，并贪婪地选择得分最高的项目。这会导致所有剩余项目的非自适应列表，并且这些",
"项目按此顺序为probed，直到停止规则。独立情况下的PARCA停止规则对应于将剩余目标（在",
"函数值上）减少Q1/r，而SPARCA规则涉及在相关情况下通过s1/r因素减少“兼容方案”的数量",
"定理E1.1和1.3的分析在开始时遵循平行线。对于每个i≥0，我们将成本Q·2°后算法的",
"“未完成”概率与成本2”后的最佳自适应解决方案相关联。“拉伸”因子α对应于近似比，对于独",
"立和相关的情况，近似比不同。为了关联这些未完成概率，我们考虑了算法在成本α2°-1和α2”之",
"旬选择的项目的总分G。分析的关键在于给出总分G的下限和上限；对于独立和相关的设置，这",
"里的参数是不同的",
"在独立情况下，任何项目的得分都是其相对边际收益的估计值，其中我们对所有先前项目",
"以及Xe进行期望。我们还通过项目的成本将这一收益正常化。定义见公式（1）。在总分G的下",
"限中，我们使用[AAK19]的采样引理的变体以及子模最大化E[BSZ19]的常数因子适应性间隙。我",
"们还需要将（所有先前实现的）结果空间划分为“好”和“坏”结果：以良好的结果为条件，OPT",
"在成本2之前完成的可能性很高。在我们对抽样引理的证明中，好的结果是必要的，但幸运的是",
"坏结果的总概率很小（并且可以有效地忽略它们）。在上限G中，我们将总分视为决策/样本路径",
"的总和，并使用相对增益的总和对应于谐波序列的事实",
"在相关情况下，任何项目e的得分都是两项的总和（i）独立情况下的预期相对边际收益，以",
"及（ii）“消除”情况下的概率估计。这两个术语都是需要的，因为算法需要平衡（i）覆盖函数和",
"(i）识别实现的场景（之后覆盖f是微不足道的）。同样，我们通过项目的成本进行标准化。参见",
"方程（16）。在总分G的下限中，我们将结果空间划分为好/好/坏的结果，这些结果对应于OPT",
"的高条件概率（i）通过成本2覆盖函数f，（ii）通过成本2”消除恒定比例的场景，或（ii）两种",
"情况都不是。此外，通过限制具有“大量”兼容场景的结果（否则，将应用算法的停止规则），我们",
"可以通过s1/r限制“相关”结果的数量。然后，我们考虑OPT（最高成本2i）以所有良好/良好",
"的结果为条件，并表明其中一项得分很高。对于上限G，我们再次将总分视为决策路径上的总和",
"1.1 Related Work",
"在[Wo182]中获得了基本子模覆盖问题的（1+lnQ）近似算法。这也是最好的可能（除非P=NP）：",
"因为设置覆盖问题是一种特殊情况E[Fei98]。在过去的几年中，已经有许多关于stochastic变种",
"的子模块封面[GK11,INvdZ16,DHK16,GHKL16,NKN20]的论文，因为这个框架捕获了许多不同",
"的应用程序",
"首先在[GV06]中研究了随机集覆盖问题，该问题表明适应性差距位于2(d)和O(d-之间，其",
"中d是要覆盖的对象数。最近，[AAK19]将上限改进为O(dlogdlog(mCmaa)）。作为定理1.1的",
"推论，我们得到了更紧的O(dlogd）适应性缺口。重要的是，我们消除了项目m和最大成本的依",
"赖性，这在d中甚至可能是指数的",
"从[INvdZ16]得出随机子模块覆盖（独立项）的O(logQ）自适应近似算法。在[LPRY08,GK11,",
"GK17,DHK16]】中获得了特殊情况或边界较弱的相关结果。如前所述，在【AAK19]中获得了具有",
"有限轮次适应性的算法的上/下界。定理1.1通过O(r·log(mcmaz)）因子对这些上界进行了改进",
"此外，我们的算法是贪婪的风格和更简单的实现。我们绕过先前工作中计算昂贵的步骤，例如解决",
"随机子模最大化的几个实例。我们的分析（上面概述的）也是非常不同的。我们使用[AAK19]中的",
"关键采样引理的变体，但是它以不同的方式应用，并且仅影响分析。我们对总分进行下限/上限的",
"高级方法类似于[INvdZ16]中对完全自适应问题的分析。但是，细节非常不同，因为我们需要处理",
"r-round-adaptivitygap的问题",
"[GHKL16]中引入了场景子模覆盖问题，作为包括最优决策树[GG74,GNR17]，等价类确",
"4"
] |
{
"arxiv_id": "2106.16115",
"language": "zh",
"page_idx": 3
}
|
||
2106.16115-zh-4
|
2106.16115-zh-4.jpg
|
定 [CLS14] 和决策区域确定 [JCK$^{+}$14] 在内的几个问题的通用概括。在 [GHKL16] 中获得了一种$O(\log(sQ))$完全自适应算法。在 [NKN20] 中也获得了相同的近似比 (在更一般的设置中)。在相关设置中，我们不知道在有限的适应性轮次上有任何先前的工作(当轮次为$r<\log s$时)。我们分析的某些方面 (e。g。, 好/好/坏的结果) 类似于 [NKN20] ,但还需要做更多的工作，因为我们必须限制$r$ -round-adaptivitygap
由 [GK17] 引人的“自适应子模块性”框架以不同的方式对随机子模块覆盖中的相关性进行建模。自适应子模块性是目标函数$f$ (需要覆盖)和项目上的分布$\mathcal{D}$的组合条件。虽然具有独立项目的随机子模块化封面满足自适应子模块化，但场景子模块化封面不满足。当所有成本均为单位时，[GK17]给出了用于自适应子模块盖 (AdSubCov )的完全自适应$O(\log^2(mQ))$近似算法。最近，[EKM19] 在$O(\log m\log(Qm))$轮适应性中将其改进为$O(\log(mQ))$近似，但仍假设单位成本。当成本是任意的时，[EKM19] 中的结果意味着$O(\log m\log(Qmc_\mathrm{max}))$轮次中的$O(\log(mQc_\mathrm{max}))$ 近似：此结果也适用于随机子模块覆盖，并且可以与我们获得的推论El1.2 进行比较。尽管场景子模块封面是not ,这是 AdSubCov 的特例，但 [GHKL16] 表明，场景子模块封面可以重新制定为 AdSubCov , 具有不同的目标函数，即自适应子模块。但是，这个新的目标函数具有更大的$\frac Q{p_{min}}$的“$Q$值”。因此，当来自 [EKM19] 的算法应用于这个新的目标函数时，它只意味着$O(\log m\log(Qm\frac{c_{\mathrm{max}}}{p_{\mathrm{min}}}))$轮次中的$O(\log(mQ\frac{c_{\mathrm{max}}}{p_{\mathrm{min}}}))$近似 (这可以与推论$\overline{\mathrm{E}}1.4$进行比较)。据我们所知，没有使用少于平方对数轮次的自适应的 AdSubCov 算法
对于随机子模块maximization ,适应性的作用已得到广泛研究。在 [AN16]中获得了在拟阵约束下(在被探测的项目上)的恒定适应性间隙。后来，[GNS17]针对非常大的一类“前缀闭合”约束获得了恒定的适应性差距；随后将常数因子改进为 2,这也是最好的$\mathbb{E}[$BSZ19]。我们在分析中利用这一结果。更一般地说，适应性的作用已被广泛研究用于各种随机“最大化”问题，例如背包$\mathbb{E}[$DGV08,BGK11] , 匹配$[\mathbb{E}[$BGL$^+12$,BDH20] ,探测$\mathbb{E}[$GN13] 和定向运动$\mathbb{E}[$GM09,GKNR15,BN15]
最近，有几个结果检查了自适应在 (确定性)子模优化中的作用$\boxed{\mathrm{E}}[$BS18a,BBS18,BS18b,BRS19 CQ19] 。这里的动机是并行化函数查询 (通常是昂贵的)。在许多设置中，有一些使用 (多)对数轮数的算法几乎与最佳顺序 (或完全自适应) 近似算法相匹配。虽然我们的动机是相似的 (并行昂贵的探测步骤),所使用的技术是完全不同的
# 2 Definitions
在随机子模覆盖问题中，输人是$m$随机变量(或项目)的集合$\mathcal{X}=\{\mathcal{X}_1,\ldots,\mathcal{X}_m\}$ 。每个项目$\mathcal{X}_i$ 都有一个成本$c_i\in\mathbb{R}_+$ ,并实现到 groundset $U$的随机元素。用$\mathcal{D}$表示$\mathcal{X}$的联合分布。随机变量$\mathcal{X}_i$可能是独立的，也可能不是独立的；我们将在下面更详细地讨论这个问题。项$\mathcal{X}_i$的实现由$X_i\in U$表示；仅当$\mathcal{X}_i$为probed且代价为$c_i$时，才知道此实现。扩展此表示法，给定项目$S\subseteq\mathcal{X}$ 的子集，其实现表示为$S=\{X_i:\mathcal{X}_i\in\mathcal{S}\}\subseteq U$
此外，我们还给出了具有$f(U)=Q$的整数值单调子模函数$f:2^U\to\mathbb{Z}_+$ 。当且仅当$f(S)=Q$ 为最大值时，才可以实现$\mathcal{S}\subseteq\mathcal{X}$项的$S$ ;在这种情况下，我们还说$\mathcal{S}$涵盖了$f$ 。目标是探测 (可能是自适应地)实现为可行集的项的子集$S\subseteq\mathcal{X}$ 。我们使用简写$c(\mathcal{S}):=\sum_i:\chi_i\in\mathcal{S}c_i$来表示$\mathcal{S}\subseteq\mathcal{X}$ 中项目的总成本。目标是最小化被探测项目的预期成本，其中期望被接管$\chi$中的随机性。我们考虑以下类型的解决方案。
$\textbf{Definition 2. 1. For an integer r1, an r- round- adaptive }solutionproceedsinrroundsof$ $adaptivity. \textit{ In each round k}\in \{ 1, \ldots , r\}$ $, \textit{ the solution specifes an ordering of all remaining items}$ $and\textit{ probes them in this order until some stopping rule. The decisions in round k can depend on }$ $the\textit{ realizations seen in all previous rounds 1, \ldots , k- 1}.$
5
|
[
[
0.11767578125,
0.09344482421875,
0.8798828125,
0.09344482421875,
0.8798828125,
0.10797119140625,
0.11767578125,
0.10797119140625
],
[
0.11767578125,
0.11236572265625,
0.8798828125,
0.11236572265625,
0.8798828125,
0.1256103515625,
0.11767578125,
0.1256103515625
],
[
0.11767578125,
0.1300048828125,
0.88134765625,
0.1300048828125,
0.88134765625,
0.143310546875,
0.11767578125,
0.143310546875
],
[
0.1168212890625,
0.146484375,
0.88134765625,
0.147705078125,
0.88134765625,
0.1622314453125,
0.1168212890625,
0.1610107421875
],
[
0.1168212890625,
0.1641845703125,
0.338134765625,
0.1666259765625,
0.338134765625,
0.1817626953125,
0.1168212890625,
0.179931640625
],
[
0.143798828125,
0.18310546875,
0.88134765625,
0.184326171875,
0.88134765625,
0.1988525390625,
0.143798828125,
0.1976318359375
],
[
0.1168212890625,
0.20263671875,
0.88134765625,
0.20263671875,
0.88134765625,
0.216552734375,
0.1168212890625,
0.216552734375
],
[
0.12091064453125,
0.2197265625,
0.88134765625,
0.2197265625,
0.88134765625,
0.23291015625,
0.12091064453125,
0.23291015625
],
[
0.1168212890625,
0.2374267578125,
0.88134765625,
0.238037109375,
0.88134765625,
0.25244140625,
0.1168212890625,
0.251953125
],
[
0.1168212890625,
0.255615234375,
0.88134765625,
0.255615234375,
0.88134765625,
0.270751953125,
0.1168212890625,
0.270751953125
],
[
0.11767578125,
0.2734375,
0.88134765625,
0.2734375,
0.88134765625,
0.28857421875,
0.11767578125,
0.28857421875
],
[
0.1168212890625,
0.291748046875,
0.88134765625,
0.291748046875,
0.88134765625,
0.30615234375,
0.1168212890625,
0.30615234375
],
[
0.11767578125,
0.309326171875,
0.88134765625,
0.309326171875,
0.88134765625,
0.323974609375,
0.11767578125,
0.323974609375
],
[
0.11767578125,
0.32763671875,
0.88134765625,
0.32763671875,
0.88134765625,
0.341552734375,
0.11767578125,
0.341552734375
],
[
0.11767578125,
0.3447265625,
0.88134765625,
0.3447265625,
0.88134765625,
0.359130859375,
0.11767578125,
0.359130859375
],
[
0.11358642578125,
0.3623046875,
0.88232421875,
0.363037109375,
0.88232421875,
0.380615234375,
0.11358642578125,
0.380126953125
],
[
0.11767578125,
0.380615234375,
0.5810546875,
0.380615234375,
0.5810546875,
0.395751953125,
0.11767578125,
0.395751953125
],
[
0.143798828125,
0.399658203125,
0.88134765625,
0.40087890625,
0.88134765625,
0.415283203125,
0.143798828125,
0.4140625
],
[
0.11767578125,
0.41845703125,
0.88134765625,
0.41845703125,
0.88134765625,
0.43310546875,
0.11767578125,
0.43310546875
],
[
0.1192626953125,
0.4375,
0.88232421875,
0.4375,
0.88232421875,
0.45068359375,
0.1192626953125,
0.45068359375
],
[
0.11767578125,
0.455078125,
0.88232421875,
0.455078125,
0.88232421875,
0.468505859375,
0.11767578125,
0.468505859375
],
[
0.11517333984375,
0.47021484375,
0.88232421875,
0.4716796875,
0.88232421875,
0.4892578125,
0.11517333984375,
0.488037109375
],
[
0.11767578125,
0.490478515625,
0.340576171875,
0.490478515625,
0.340576171875,
0.50390625,
0.11767578125,
0.50390625
],
[
0.146240234375,
0.50830078125,
0.88134765625,
0.50830078125,
0.88134765625,
0.52294921875,
0.146240234375,
0.52294921875
],
[
0.11767578125,
0.5263671875,
0.88134765625,
0.5263671875,
0.88134765625,
0.54052734375,
0.11767578125,
0.54052734375
],
[
0.11767578125,
0.54443359375,
0.8798828125,
0.54443359375,
0.8798828125,
0.55810546875,
0.11767578125,
0.55810546875
],
[
0.11767578125,
0.5625,
0.463134765625,
0.5625,
0.463134765625,
0.57568359375,
0.11767578125,
0.57568359375
],
[
0.11273193359375,
0.60302734375,
0.28759765625,
0.60400390625,
0.28759765625,
0.6220703125,
0.11273193359375,
0.62060546875
],
[
0.1168212890625,
0.63720703125,
0.88134765625,
0.63818359375,
0.88134765625,
0.65283203125,
0.1168212890625,
0.6513671875
],
[
0.1168212890625,
0.6572265625,
0.88134765625,
0.6572265625,
0.88134765625,
0.67041015625,
0.1168212890625,
0.67041015625
],
[
0.11767578125,
0.6748046875,
0.8798828125,
0.6748046875,
0.8798828125,
0.68798828125,
0.11767578125,
0.68798828125
],
[
0.11767578125,
0.6923828125,
0.88134765625,
0.6923828125,
0.88134765625,
0.70703125,
0.11767578125,
0.70703125
],
[
0.11767578125,
0.71044921875,
0.484375,
0.71044921875,
0.484375,
0.7236328125,
0.11767578125,
0.7236328125
],
[
0.1446533203125,
0.72802734375,
0.88134765625,
0.72802734375,
0.88134765625,
0.7421875,
0.1446533203125,
0.7421875
],
[
0.11767578125,
0.7470703125,
0.8798828125,
0.7470703125,
0.8798828125,
0.76025390625,
0.11767578125,
0.76025390625
],
[
0.1168212890625,
0.7646484375,
0.8837890625,
0.7646484375,
0.8837890625,
0.7822265625,
0.1168212890625,
0.7822265625
],
[
0.1168212890625,
0.78076171875,
0.8798828125,
0.7822265625,
0.8798828125,
0.796875,
0.1168212890625,
0.79541015625
],
[
0.1168212890625,
0.7998046875,
0.303955078125,
0.80029296875,
0.303955078125,
0.81494140625,
0.1168212890625,
0.814453125
],
[
0.1168212890625,
0.830078125,
0.88232421875,
0.830078125,
0.88232421875,
0.8447265625,
0.1168212890625,
0.8447265625
],
[
0.1192626953125,
0.84765625,
0.88134765625,
0.84765625,
0.88134765625,
0.8623046875,
0.1192626953125,
0.8623046875
],
[
0.1192626953125,
0.8662109375,
0.8798828125,
0.8662109375,
0.8798828125,
0.880859375,
0.1192626953125,
0.880859375
],
[
0.1192626953125,
0.8837890625,
0.5595703125,
0.8837890625,
0.5595703125,
0.89794921875,
0.1192626953125,
0.89794921875
],
[
0.491943359375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"定[CLS14]和决策区域确定[JCK+14]在内的几个问题的通用概括。在[GHKL16]中获得了一种",
"O(log(sQ）完全自适应算法。在[NKN20]中也获得了相同的近似比（在更一般的设置中）。在相关",
"设置中，我们不知道在有限的适应性轮次上有任何先前的工作（当轮次为r<logs时）。我们分析",
"的某些方面（e。g。，好/好/坏的结果）类似于[NKN20]，但还需要做更多的工作，因为我们必须",
"限制r-round-adaptivitygap",
"由[GK17]引入的“自适应子模块性”框架以不同的方式对随机子模块覆盖中的相关性进行建",
"模。自适应子模块性是目标函数（需要覆盖）和项目上的分布D的组合条件。虽然具有独立项",
"目的随机子模块化封面满足自适应子模块化，但场景子模块化封面不满足。当所有成本均为单位",
"时，[GK17]给出了用于自适应子模块盖（AdSubCov）的完全自适应O(log（mQ)）近似算法。最",
"近，[EKM19)在O(logmlog(Qm)）轮适应性中将其改进为O(log(mQ)）近似，但仍假设单位成本。",
"当成本是任意的时，[EKM19]中的结果意味着O(logmlog(QmCmax)）轮次中的O(log(mQcmax))",
"近似：此结果也适用于随机子模块覆盖，并且可以与我们获得的推论E1.2进行比较。尽管场景",
"子模块封面是not，这是AdSubCov的特例，但[GHKL16]表明，场景子模块封面可以重新制",
"定为AdSubCov，具有不同的目标函数，即自适应子模块。但是，这个新的目标函数具有更大",
"的的“Q值”。因此，当来自[EKM19]的算法应用于这个新的目标函数时，它只意味着",
"O(logmlog（Qmmax））轮次中的O(log（mQmax））近似（这可以与推论E1.4进行比较）。据我们所",
"知，没有使用少于平方对数轮次的自适应的AdSubCov算法",
"对于随机子模块macimization，适应性的作用已得到广泛研究。在[AN16]中获得了在拟",
"阵约束下（在被探测的项目上）的恒定适应性间隙。后来，[GNS17]针对非常大的一类“前缀",
"闭合”约束获得了恒定的适应性差距；随后将常数因子改进为2，这也是最好的EBSZ19]。",
"我们在分析中利用这一结果。更一般地说，适应性的作用已被广泛研究用于各种随机“最大",
"化”问题，例如背包E[DGV08,BGK11]，匹配E[BGL+12,BDH20]，探测E[GN13]和定向运",
"动E[GM09,GKNR15,BN15]",
"最近，有几个结果检查了自适应在（确定性）子模优化中的作用E[BS18a,BBS18,BS18b,BRS19,",
"CQ19。这里的动机是并行化函数查询（通常是昂贵的）。在许多设置中，有一些使用（多）对数",
"轮数的算法几乎与最佳顺序（或完全自适应）近似算法相匹配。虽然我们的动机是相似的（并行昂",
"贵的探测步骤），所使用的技术是完全不同的",
"2Definitions",
"在随机子模覆盖问题中，输入是m随机变量（或项目）的集合=（i，..,tm}。每个项目X",
"都有一个成本ciER+，并实现到groundsetU的随机元素。用D表示的联合分布。随机变",
"量，可能是独立的，也可能不是独立的；我们将在下面更详细地讨论这个问题。项，的实现由",
"X，EU表示；仅当i为probed且代价为ci时，才知道此实现。扩展此表示法，给定项目SCα",
"的子集，其实现表示为S=X，：X,ESIU",
"此外，我们还给出了具有f(U）=Q的整数值单调子模函数f：2U→Z+。当且仅当f(S）=Q",
"为最大值时，才可以实现SCX项的S；在这种情况下，我们还说S涵盖了。目标是探测（可",
"能是自适应地）实现为可行集的项的子集SC。我们使用简写c(S)=Zi:x,esCi来表示S",
"中项目的总成本。目标是最小化被探测项目的预期成本，其中期望被接管匕中的随机性。我们考",
"虑以下类型的解决方案。",
"Definition 2.l. For an integer r ≥ 1 , an r -round-adaptive solution proceeds in r rounds of",
"adaptivity. In each round k E [1,--.,r] , the solution specifies an ordering of all remaining items",
"and probes them in this order until some stopping rule. The decisions in round k can depend or",
"the realizations seen in all previous rounds l,..., k - 1.",
"5"
] |
{
"arxiv_id": "2106.16115",
"language": "zh",
"page_idx": 4
}
|
||
2302.11477-zh-1
|
2302.11477-zh-1.jpg
|
An Interpretable Determinantal Choice Model for Subset Selection
确定性点过程(DPP) 是机器学习环境中大量使用的子集选择模型家族。虽然不是建立在随机效用最大化的基础上，但 DPP 对于项目表现出排斥的学习和推理任务非常易于处理；它们最初用于对费米子[Macchi,1975] 之间的排斥进行建模。DPP允许高效的学习算法，并支持推理任务，如采样和调节(参见，e。g。,Kulesza et al.[2012])。DPP已成功用于学习任务，例如文档摘要[Kulesza and Taskar, 2011], 篮子完成[Gartrell et al., 2016] 和无线干扰[Saha and Dhillon, 2019]。但是，与基于 RUM 的模型不同，DPP 的可解释性较差，并且不像基于实用程序的模型那样具有直观的吸引力
本文表明，确定性过程实际上是建立在随机效用最大化上的子集选择模型，并指定了一个与著名的离散选择模型有很强联系的 DPP 子族。这些连接使 DPP 在离散选择建模中更直观，可解释和可访问具有背景的从业者。我们的贡献如下：(1)我们证明任何确定性过程都等效于子集上的 MNL 模型，具有特定的效用函数；(2)我们证明该效用函数是“具有修正的加性”,其中每个集合都因包含类似项目而受到修正惩罚；(3) 我们指定一个确定性选择模型，并将其连接到离散选择，特别是；(4)当没有项目相互作用时，我们的模型包含最大的 Msum。此外，我们在实践中展示了确定性选择模型：(6)我们使用合成数据来验证确定性选择模型，当从数据中学习时，如预期的那样包含逻辑回归和 MNL;(7)最后，我们使用原始的 LoRaWAN 无线数据来学习无线干扰模型。我们的结果表明，该模型既捕获了数据中的负相关性，又学习了可解释的参数。实际上，我们使用这些参数来得出有关由于干扰而导致的 LoRaWAN 传输损耗的新见解。
# 2 Setting and Related Work
本节介绍了变量分类下子集选择的设置，并回顾了离散选择，随机效用最大化，子集选择模型和 DPP
问题设置。这项工作采用了基于特征的参数化和学习方法来进行子集选择。考虑项目$\mathcal{U}$ 的任意宇宙，其中每$\begin{array}{l}\text{个项目 }i\in\mathcal{U}\text{ 都有一个关联的特征向量 }x_i\in\mathbb{R}^d\text{。分类是一组有限的项目，}A\subseteq\mathcal{U}\text{。在 }|A|\times d\text{ 矩阵 }\boldsymbol{X}_A\text{ 中收}\\\text{集分类的特征向量 }A\text{。子集选择模型是一个函数，它采用具有 }\boldsymbol{X}_A\text{ 特征的任何分类 }A\text{ ,并返回子集 }2^A\text{ 集合}\end{array}$ 上的概率分布。我们考虑一种有监督的学习设置，其中假定可以访问$A_k$ ,选择$C_k\subseteq A_k$和功能$\boldsymbol{X}_k=\boldsymbol{X}_{A_k}$ 的序列 $k=1,\ldots,K$。请注意，分类的大小可能不同，并且可能包含彼此完全不同的项目$^3$。预测涉及在给定分类$A_k$的情况下推断可能的选择集$C_k\subseteq A_k$ ,并收集相关的项目功能$X_k$
离散选择和 MNL。离散选择模型最多允许从一个分类中选择一个项目。许多最常见的离散选择模型都是建立在随机效用最大化的基础上的(参见，e。g。,Train[2009])。在此设置中，分类$A$中的每个项目$i$都配备了实值实用程序$u(i)$ ,并选择了实用程序最大化选项。实用程序中的随机性会在选择中产生随机性。MNL 模型是一个突出的例子[McFadden,1973]。在 MNL 模型中，项$i$的效用具有基于线性特征的参数形式
$$u(i)=\boldsymbol{\beta}^T\boldsymbol{x}_i+\epsilon_i,\quad\forall i\in\mathcal{U},$$
其中$\beta$是与$x_i$具有相同维数的系数向量，而$\epsilon_i$是 iid 标准 Gumbel 随机变量。给定分类$A$ ,选择项目$i\in A$ 的概率等于所有$j\in A:j\neq i$的$u(i)>u(j)$的概率，在 Gumbel 误差下，该概率与指数非随机效用分量成正
$比$。
$$\Pr[\text{choose item}i\in A]=\frac{\exp\{\boldsymbol\beta^T\boldsymbol x_i\}}{1+\sum_{j\in A}\exp\{\boldsymbol\beta^T\boldsymbol x_j\}}$$
(1)
“无项目”选项被视为具有预期效用 0 的项目。因此，在没有项目的情况下选择退出的概率为 1/(1+ $\sum_{j\in A}\exp\{\beta^Tx_j\})$ 。MNL 模型的一个重要特例是当分类是单个项目$\{i\}$时。这减少了“接受”或“离开”的选择。然后，接受项目$i$的概率为
(2)
$$\Pr[\text{choose item}i\in A]=\frac{\exp\{\boldsymbol\beta^T\boldsymbol x_i\}}{1+\exp\{\boldsymbol\beta^T\boldsymbol x_i\}}.$$
在学习环境中，此模型也称为 logit 模型或逻辑回归
子集选择模型。子集选择模型通过为每个项目子集配备随机效用来概括离散选择模型。该实用程序由非随机和随机组件$U(C)=V(C)+\epsilon(C)$组成。选择最大效用的子集。即使是大小适中的分类也具有令人望而却步的子集；大多数模型通过将非随机实用程序组件建模为“带有校正的可加性”来权衡通用性和易处理性，
(3)
$$\begin{aligned}V(C)=\sum_{i\in C}v(i)+W(C),\quad\forall C\subseteq A,\end{aligned}$$
,其中$W(C)$ 是对子集$C$ 的校正。一个众所周知的例子是具有校正$W(C)=\sum_i,j\in C:i\neq j\gamma(i,j)$ [Fishburn and LaValle, 1996] 的成对交互模型。最近的一个例子是一个可学习的模型，该模型允许任意校正$W(C)$ ,但是仅限于固定数量的子集[Benson et al.,2018]。后一种模型最适合于重复观察到相对少量分类的设置
3The features can also encode assortment-level features. The values of such features are the same across all items in the
assortment. This can capture, e.g., features about the choice-maker, or the time and place at which the assortment was offered.
2
|
[
[
0.11517333984375,
0.043548583984375,
0.5537109375,
0.044189453125,
0.5537109375,
0.058074951171875,
0.11517333984375,
0.05682373046875
],
[
0.11767578125,
0.0921630859375,
0.88134765625,
0.0921630859375,
0.88134765625,
0.1060791015625,
0.11767578125,
0.1060791015625
],
[
0.11767578125,
0.106689453125,
0.88134765625,
0.106689453125,
0.88134765625,
0.12060546875,
0.11767578125,
0.12060546875
],
[
0.11767578125,
0.121826171875,
0.88134765625,
0.121826171875,
0.88134765625,
0.1363525390625,
0.11767578125,
0.1363525390625
],
[
0.11767578125,
0.1337890625,
0.88134765625,
0.1337890625,
0.88134765625,
0.147705078125,
0.11767578125,
0.147705078125
],
[
0.11767578125,
0.147705078125,
0.88134765625,
0.147705078125,
0.88134765625,
0.1610107421875,
0.11767578125,
0.1610107421875
],
[
0.1168212890625,
0.1610107421875,
0.443603515625,
0.1622314453125,
0.443603515625,
0.175537109375,
0.1168212890625,
0.1749267578125
],
[
0.11767578125,
0.18310546875,
0.88134765625,
0.18310546875,
0.88134765625,
0.1962890625,
0.11767578125,
0.1962890625
],
[
0.11767578125,
0.1962890625,
0.8798828125,
0.1962890625,
0.8798828125,
0.210205078125,
0.11767578125,
0.210205078125
],
[
0.11767578125,
0.2083740234375,
0.88134765625,
0.208984375,
0.88134765625,
0.2235107421875,
0.11767578125,
0.222900390625
],
[
0.11767578125,
0.2235107421875,
0.88134765625,
0.2235107421875,
0.88134765625,
0.2374267578125,
0.11767578125,
0.2374267578125
],
[
0.11767578125,
0.2374267578125,
0.88134765625,
0.2374267578125,
0.88134765625,
0.250732421875,
0.11767578125,
0.250732421875
],
[
0.1192626953125,
0.251953125,
0.88134765625,
0.251953125,
0.88134765625,
0.26513671875,
0.1192626953125,
0.26513671875
],
[
0.11767578125,
0.26513671875,
0.88134765625,
0.26513671875,
0.88134765625,
0.27978515625,
0.11767578125,
0.27978515625
],
[
0.11767578125,
0.279052734375,
0.88134765625,
0.279052734375,
0.88134765625,
0.292236328125,
0.11767578125,
0.292236328125
],
[
0.11767578125,
0.292236328125,
0.74169921875,
0.292236328125,
0.74169921875,
0.30615234375,
0.11767578125,
0.30615234375
],
[
0.11517333984375,
0.32763671875,
0.36767578125,
0.32763671875,
0.36767578125,
0.341552734375,
0.11517333984375,
0.341552734375
],
[
0.11767578125,
0.359130859375,
0.833984375,
0.359130859375,
0.833984375,
0.369384765625,
0.11767578125,
0.369384765625
],
[
0.1192626953125,
0.37744140625,
0.8798828125,
0.37744140625,
0.8798828125,
0.391357421875,
0.1192626953125,
0.391357421875
],
[
0.1192626953125,
0.391357421875,
0.8798828125,
0.391357421875,
0.8798828125,
0.40478515625,
0.1192626953125,
0.40478515625
],
[
0.1168212890625,
0.40478515625,
0.88134765625,
0.40478515625,
0.88134765625,
0.41845703125,
0.1168212890625,
0.41845703125
],
[
0.1192626953125,
0.419189453125,
0.8759765625,
0.419189453125,
0.8759765625,
0.43310546875,
0.1192626953125,
0.43310546875
],
[
0.1168212890625,
0.431884765625,
0.88134765625,
0.43310546875,
0.88134765625,
0.447509765625,
0.1168212890625,
0.4462890625
],
[
0.1192626953125,
0.448974609375,
0.6650390625,
0.448974609375,
0.6650390625,
0.458984375,
0.1192626953125,
0.458984375
],
[
0.1192626953125,
0.46728515625,
0.88134765625,
0.46728515625,
0.88134765625,
0.48095703125,
0.1192626953125,
0.48095703125
],
[
0.11767578125,
0.479736328125,
0.441162109375,
0.481689453125,
0.441162109375,
0.496826171875,
0.11767578125,
0.494384765625
],
[
0.4453125,
0.48095703125,
0.8798828125,
0.48095703125,
0.8798828125,
0.494384765625,
0.4453125,
0.494384765625
],
[
0.1192626953125,
0.494384765625,
0.88134765625,
0.494384765625,
0.88134765625,
0.5087890625,
0.1192626953125,
0.5087890625
],
[
0.1168212890625,
0.5087890625,
0.84326171875,
0.5087890625,
0.84326171875,
0.52294921875,
0.1168212890625,
0.52294921875
],
[
0.40283203125,
0.52978515625,
0.59619140625,
0.52978515625,
0.59619140625,
0.54443359375,
0.40283203125,
0.54443359375
],
[
0.11767578125,
0.55322265625,
0.88134765625,
0.55322265625,
0.88134765625,
0.56640625,
0.11767578125,
0.56640625
],
[
0.11767578125,
0.56640625,
0.88134765625,
0.56640625,
0.88134765625,
0.580078125,
0.11767578125,
0.580078125
],
[
0.11767578125,
0.58056640625,
0.1405029296875,
0.58056640625,
0.1405029296875,
0.59326171875,
0.11767578125,
0.59326171875
],
[
0.33251953125,
0.59326171875,
0.66259765625,
0.60791015625,
0.6611328125,
0.6337890625,
0.330810546875,
0.619140625
],
[
0.54150390625,
0.59326171875,
0.62353515625,
0.59130859375,
0.62353515625,
0.60595703125,
0.54150390625,
0.60791015625
],
[
0.86279296875,
0.60302734375,
0.8837890625,
0.60302734375,
0.8837890625,
0.61572265625,
0.86279296875,
0.61572265625
],
[
0.11767578125,
0.6318359375,
0.8837890625,
0.6318359375,
0.8837890625,
0.64599609375,
0.11767578125,
0.64599609375
],
[
0.11517333984375,
0.64306640625,
0.8837890625,
0.64013671875,
0.8837890625,
0.6611328125,
0.11517333984375,
0.6640625
],
[
0.11767578125,
0.6611328125,
0.408447265625,
0.6611328125,
0.408447265625,
0.6748046875,
0.11767578125,
0.6748046875
],
[
0.35302734375,
0.68310546875,
0.6396484375,
0.69580078125,
0.63818359375,
0.72021484375,
0.351318359375,
0.70849609375
],
[
0.53759765625,
0.6845703125,
0.62109375,
0.681640625,
0.62109375,
0.6962890625,
0.53759765625,
0.69873046875
],
[
0.86279296875,
0.693359375,
0.88232421875,
0.693359375,
0.88232421875,
0.70703125,
0.86279296875,
0.70703125
],
[
0.11767578125,
0.7236328125,
0.48193359375,
0.7236328125,
0.48193359375,
0.7373046875,
0.11767578125,
0.7373046875
],
[
0.12091064453125,
0.74560546875,
0.8798828125,
0.74560546875,
0.8798828125,
0.755859375,
0.12091064453125,
0.755859375
],
[
0.11767578125,
0.75830078125,
0.88134765625,
0.75830078125,
0.88134765625,
0.771484375,
0.11767578125,
0.771484375
],
[
0.11767578125,
0.77294921875,
0.85546875,
0.77294921875,
0.85546875,
0.7861328125,
0.11767578125,
0.7861328125
],
[
0.369384765625,
0.79541015625,
0.63134765625,
0.79541015625,
0.63134765625,
0.81005859375,
0.369384765625,
0.81005859375
],
[
0.86279296875,
0.79736328125,
0.88232421875,
0.79736328125,
0.88232421875,
0.81005859375,
0.86279296875,
0.81005859375
],
[
0.428955078125,
0.81103515625,
0.453369140625,
0.81103515625,
0.453369140625,
0.82470703125,
0.428955078125,
0.82470703125
],
[
0.12335205078125,
0.8291015625,
0.88623046875,
0.8291015625,
0.88623046875,
0.85302734375,
0.12335205078125,
0.85302734375
],
[
0.1168212890625,
0.8486328125,
0.88134765625,
0.8486328125,
0.88134765625,
0.8623046875,
0.1168212890625,
0.8623046875
],
[
0.11767578125,
0.86279296875,
0.7998046875,
0.86279296875,
0.7998046875,
0.876953125,
0.11767578125,
0.876953125
],
[
0.1356201171875,
0.8837890625,
0.88134765625,
0.88623046875,
0.88134765625,
0.90087890625,
0.1356201171875,
0.8984375
],
[
0.11767578125,
0.8994140625,
0.85546875,
0.8994140625,
0.85546875,
0.91357421875,
0.11767578125,
0.91357421875
],
[
0.49267578125,
0.9345703125,
0.50732421875,
0.9375,
0.50341796875,
0.94970703125,
0.488525390625,
0.9462890625
]
] |
[
"AnInterpretableDeterminantalChoiceModelforSubsetSelection",
"确定性点过程（DPP）是机器学习环境中大量使用的子集选择模型家族。虽然不是建立在随机效用最大化的基",
"础上，但DPP对于项目表现出排斥的学习和推理任务非常易于处理；它们最初用于对费米子[Macchi，1975]",
"之间的排斥进行建模。DPP允许高效的学习算法，并支持推理任务，如采样和调节（参见，e。g。，Kulesza",
"etal.[2012])。DPP已成功用于学习任务，例如文档摘要[KuleszaandTaskar,2011]，篮子完成[Gartrelletal.,",
"2016]和无线干扰[SahaandDhillon，2019]。但是，与基于RUM的模型不同，DPP的可解释性较差，并且不",
"像基于实用程序的模型那样具有直观的吸引力",
"本文表明，确定性过程实际上是建立在随机效用最大化上的子集选择模型，并指定了一个与著名的离散选",
"择模型有很强联系的DPP子族。这些连接使DPP在离散选择建模中更直观，可解释和可访问具有背景的从",
"业者。我们的贡献如下：（1）我们证明任何确定性过程都等效于字集上的MNL模型，具有特定的效用函",
"数；（2）我们证明该效用函数是“具有修正的加性”，其中每个集合都因包含类似项目而受到修正惩罚；（3）",
"我们指定一个确定性选择模型，并将其连接到离散选择，特别是；（4）当没有项目相互作用时，我们的模型",
"包含最大的Msum。此外，我们在实践中展示了确定性选择模型：（6）我们使用合成数据来验证确定性选择",
"模型，当从数据中学习时，如预期的那样包含逻辑回归和MNL；（7）最后，我们使用原始的LoRaWAN无线",
"数据来学习无线干扰模型。我们的结果表明，该模型既捕获了数据中的负相关性，又学习了可解释的参数。",
"实际上，我们使用这些参数来得出有关由于干扰而导致的LoRaWAN传输损耗的新见解。",
"2Setting and Related Work",
"本节介绍了变量分类下子集选择的设置，并回顾了离散选择，随机效用最大化，子集选择模型和DPP",
"问题设置。这项工作采用了基于特征的参数化和学习方法来进行子集选择。考虑项目认的任意宇宙，其中每",
"个项目iEu都有一个关联的特征向量c,ERd。分类是一组有限的项目，ACu。在A×d矩阵XA中收",
"集分类的特征向量A。子集选择模型是一个函数，它采用具有XA特征的任何分类A，并返回子集24集合",
"上的概率分布。我们考虑一种有监督的学习设置，其中假定可以访问A，选择CkCAk和功能Xk=XA",
"的序列k=1，..·,K。请注意，分类的大小可能不同，并且可能包含彼此完全不同的项目3。预测涉及在给",
"定分类A的情况下推断可能的选择集CCA，并收集相关的项自功能Xk",
"离散选择和MNL。离散选择模型最多允许从一个分类中选择一个项目。许多最常见的离散选择模型都是建",
"立在随机效用最大化的基础上的（参见，e。g。",
"，Train[2009]。在此设置中，分类A中的每个项目i都配备",
"了实值实用程序u(i），并选择了实用程序最大化选项。实用程序中的随机性会在选择中产生随机性。MNL",
"模型是一个突出的例子[McFadden,1973]。在MNL模型中，项i的效用具有基于线性特征的参数形式",
"u(i)=βTai+ei, Vieu,",
"其中β是与ai具有相同维数的系数向量，而e是id标准Gumbel随机变量。给定分类A，选择项目iEA",
"的概率等于所有iEA：j≠i的u(i）>u(i)的概率，在Gumbel误差下，该概率与指数非随机效用分量成正",
"比。",
"Pr[choose item i E A] = 1+ ZjeA exp(βT a)]",
"exp(BTci",
"(1)",
"“无项目”选项被视为具有预期效用0的项目。因此，在没有项目的情况下选择退出的概率为1/(1+",
"ZiEAexp(BTal）。MNL模型的一个重要特例是当分类是单个项目（时。这减少了“接受”或“离",
"开”的选择。然后，接受项目的概率为",
"Pr[choose item i e A] = I + exp(βr a;)",
"exp(βTai]",
"(2)",
"在学习环境中，此模型也称为logit模型或逻辑回归",
"子集选择模型。子集选择模型通过为每个项自子集配备随机效用来概括离散选择模型。该实用程序由非随机",
"和随机组件U(C）=V(C）十e(C）组成。选择最大效用的子集。即使是大小适中的分类也具有令人望而却步",
"的子集；大多数模型通过将非随机实用程序组件建模为“带有校正的可加性”来权衡通用性和易处理性，",
"V(C) = > u(i)+W(C), VCC A,",
"(3)",
"itC",
"其中W(C)是对子集C的校正。一个众所周知的例子是具有校正W(C)=DjCi(i,5)[Fishburn and",
"LaValle，1996]的成对交互模型。最近的一个例子是一个可学习的模型，该模型允许任意校正W(），但是仅",
"限于固定数量的子集[Bensonetal.,2018]。后一种模型最适合于重复观察到相对少量分类的设置",
"3The features can also encode assortment-level features. The values of such features are the same across all items in the",
"assortment. This can capture, e.g., features about the choice-maker, or the time and place at which the assortment was offered.",
"2"
] |
{
"arxiv_id": "2302.11477",
"language": "zh",
"page_idx": 1
}
|
||
2302.11477-zh-2
|
2302.11477-zh-2.jpg
|
An Interpretable Determinantal Choice Model for Subset Selection
确定点进程。确定性点过程 (DPP) 是子集上的概率分布，有利于不同的子集。固定分类$A$上的 DPP 的特征在于其核矩阵 $\boldsymbol{L}$ Borodin [2009]。这是一个$|A|\times|A|$ 对称 p。s。d。矩阵$^{4}$。观察子集$C\subseteq A$ 的可能性与内核$L$的子矩阵$L_C$的对数行列式成正比，该子矩阵对应于$C$中的行和列：
$$\Pr[\text{Choose}C]=\frac{\det(\boldsymbol{L}_C)}{\det(\boldsymbol{I}+\boldsymbol{L})}.$$
(4)
$L_{\emptyset}$的行列式取为1。分母是归一化常数的闭式表达式，满足$\sum_B\subseteq A\det(\boldsymbol{L}_B)=\det(\boldsymbol{I}+\boldsymbol{L})$,其中$\boldsymbol{I}$表示与$L$大小相同的单位矩阵。DPP 支持通过贝叶斯推理和 MLE[Kulesza and Taskar, 2011, Affandi et al., 2014] 进行监督学习的方法。此外，对于 DPP[Hough et al.,2006],还有有效的采样算法。有关 DPP的、和属性的更多详细信息，请参阅 Kulesza et al.[2012]
DPP 似然有利于选定子集中的多样性。这是使用两个项目示例[Kulesza et al., 2012] 演示的。选择单例$\{i\}$的
可能性与核矩阵的$i$对角项成正比。
$$\operatorname*{Pr}[\mathrm{Choose}\left\{i\right\}]\propto\operatorname*{det}(\boldsymbol{L}_{\left\{i\right\}})=\boldsymbol{L}_{ii}$$
另一方面，选择两个项目$\{i,j\}$的非归一化可能性由非对角项$L_{ij}$调节。
$$\begin{aligned}\Pr[\mathrm{Choose}\:\{i,j\}]&\propto\det\left(\begin{bmatrix}\boldsymbol{L}_{ii}&\boldsymbol{L}_{ij}\\\boldsymbol{L}_{ji}&\boldsymbol{L}_{jj}\end{bmatrix}\right)\\&=\boldsymbol{L}_{ii}\boldsymbol{L}_{jj}-\boldsymbol{L}_{ij}^{2}\end{aligned}$$
可以将非正项$-\boldsymbol{L}_ij^2$视为对包含相似项的子集的惩罚，其中非对角项$\boldsymbol L_ij$捕获了$i$和$j$之间的相似性概念
将 DPP 内核分解为相似性和质量组件，可为模型增加清晰度。Kulesza and Taskar[2010]表明任何 DPP 内核$\boldsymbol{L}$都可以分解。t。$\boldsymbol{L}_ij=q_iS_iq_j$,其中$q_i$是项目$i$的质量$,\boldsymbol{S}_ij$是项目$i$和$j$之间的相似性。质量被视为项目$i$的相对吸引力；相似性量化了两个项目之间的相似度。相似项收集在满足$\mathbf{0}\preceq S\preceq I^5$的矩阵$S$中。在此分解下，DPP 似然因子：
$$\Pr[\text{Choose}\:C]\propto\left(\prod_{i\in C}q_i^2\right)\det(\boldsymbol{S}_C).$$
(5)
上述公式表明，包含项目$i$的子集$C$的似然与项目$i$的平方质量成正比。多样性的偏差源于相似性子矩阵的
行列式；行列式取 0 和 1 之间的值。我们利用这种分解来弥合 DPP 和随机效用子集选择模型之间的差距。
3 Determinantal Processes & Random Utility
本节连接确定性过程和基于随机效用最大化的子集选择模型。特别是，我们表明，在特定的效用函数族下所有离散的确定性点过程都可以视为随机效用最大化选择。。此外，研究这一系列效用函数会对 DPP 的行为假设以及 DPP 与现有子集选择模型之间的联系产生新的见解。。本节从基于特征的参数视图退后一步，并全面考虑离散 DPP; 我们返回 Section 4 中的特征和参数
我们的第一个结果表明，来自 DPP 的样本是随机的效用最大化选择。结果是对“子集上是否存在随机效用函
数，使得效用最大化选择与 DPP 具有相同的分布”问题的肯定答案？答案似乎很明显；主要思想是问。
Theorem 3.1. Let Abe a set of n items and $\boldsymbol{L}$ an n -by- $n$ p.s.d. symmetric matrix. The following distributions over
$2^A\textit{are equivalent: }$
$I. ADPPoverAwith\textit{kernel L}.$
2. The utility-maximizing subset where each subset C of A has a log-determinant random utility
$U(C)=\log\det(\boldsymbol{L}_C)+\epsilon(C)$
with $\epsilon(C)$ iid standard Gumbel errors.
对此的证明以及所有后续结果都推迟到 Appendix A 。该定理表明 , DPP 是子集上的等效 MNL , 具有对数行
列式效用。我们称之为 DPP 隐含的效用
隐含效用的简化有助于将 DPP 连接到现有的子集选择模型。令$V(C)$表示子集$C$的隐含效用的非随机分量。对于单例$C=\{i\}$ ,非随机效用等于平方质量$v(i):=V(\{i\})=\log(q_i^2)$的对数。将核矩阵分解为 Equation (5) 中的质量和相似性项，得出以下定理。
$^{4}$The abbreviation p.s.d. stands positive semi-definite. An $n\times n$ matrix $K$ is p.s.d. if for all $z\in\mathbb{R}^n$ , it holds that $\boldsymbol z^T\boldsymbol{K}\boldsymbol{z}\geq0.$
$^5$Here $K\preceq K^\prime$ denotes that $K^{\prime}-K$ is p.s.d.
3
|
[
[
0.11767578125,
0.046722412109375,
0.55224609375,
0.046722412109375,
0.55224609375,
0.05682373046875,
0.11767578125,
0.05682373046875
],
[
0.1192626953125,
0.0947265625,
0.8798828125,
0.0947265625,
0.8798828125,
0.10479736328125,
0.1192626953125,
0.10479736328125
],
[
0.11767578125,
0.106689453125,
0.88134765625,
0.106689453125,
0.88134765625,
0.121826171875,
0.11767578125,
0.121826171875
],
[
0.11767578125,
0.121826171875,
0.6396484375,
0.121826171875,
0.6396484375,
0.1351318359375,
0.11767578125,
0.1351318359375
],
[
0.393798828125,
0.1383056640625,
0.6005859375,
0.15283203125,
0.59716796875,
0.177978515625,
0.390625,
0.1634521484375
],
[
0.51708984375,
0.1395263671875,
0.58837890625,
0.141357421875,
0.58837890625,
0.1546630859375,
0.51708984375,
0.15283203125
],
[
0.86279296875,
0.149658203125,
0.8837890625,
0.149658203125,
0.8837890625,
0.1634521484375,
0.86279296875,
0.1634521484375
],
[
0.1168212890625,
0.1767578125,
0.88232421875,
0.1767578125,
0.88232421875,
0.1944580078125,
0.1168212890625,
0.1944580078125
],
[
0.1192626953125,
0.1956787109375,
0.8798828125,
0.1956787109375,
0.8798828125,
0.205810546875,
0.1192626953125,
0.205810546875
],
[
0.11767578125,
0.2083740234375,
0.88134765625,
0.2083740234375,
0.88134765625,
0.2215576171875,
0.11767578125,
0.2215576171875
],
[
0.1168212890625,
0.2215576171875,
0.3896484375,
0.2215576171875,
0.3896484375,
0.23486328125,
0.1168212890625,
0.23486328125
],
[
0.11517333984375,
0.2412109375,
0.88134765625,
0.242431640625,
0.88134765625,
0.2568359375,
0.11517333984375,
0.255615234375
],
[
0.1192626953125,
0.25830078125,
0.36767578125,
0.25830078125,
0.36767578125,
0.268310546875,
0.1192626953125,
0.268310546875
],
[
0.381591796875,
0.276611328125,
0.615234375,
0.276611328125,
0.615234375,
0.290283203125,
0.381591796875,
0.290283203125
],
[
0.1168212890625,
0.298583984375,
0.625,
0.298583984375,
0.625,
0.3125,
0.1168212890625,
0.3125
],
[
0.36279296875,
0.324462890625,
0.53759765625,
0.324462890625,
0.53759765625,
0.34228515625,
0.36279296875,
0.34228515625
],
[
0.5361328125,
0.323974609375,
0.6259765625,
0.323974609375,
0.6259765625,
0.333984375,
0.5361328125,
0.333984375
],
[
0.53125,
0.33203125,
0.62353515625,
0.335205078125,
0.62353515625,
0.349853515625,
0.53125,
0.3466796875
],
[
0.48193359375,
0.354248046875,
0.59228515625,
0.354736328125,
0.59228515625,
0.369384765625,
0.48193359375,
0.36865234375
],
[
0.11767578125,
0.37890625,
0.85888671875,
0.37890625,
0.85888671875,
0.392578125,
0.11767578125,
0.392578125
],
[
0.1168212890625,
0.401611328125,
0.8798828125,
0.401611328125,
0.8798828125,
0.415283203125,
0.1168212890625,
0.415283203125
],
[
0.11517333984375,
0.415283203125,
0.88232421875,
0.41357421875,
0.88232421875,
0.429931640625,
0.11517333984375,
0.431884765625
],
[
0.11767578125,
0.43115234375,
0.88232421875,
0.43115234375,
0.88232421875,
0.4443359375,
0.11767578125,
0.4443359375
],
[
0.1168212890625,
0.44580078125,
0.306396484375,
0.44580078125,
0.306396484375,
0.458984375,
0.1168212890625,
0.458984375
],
[
0.375732421875,
0.4697265625,
0.491943359375,
0.4697265625,
0.491943359375,
0.484130859375,
0.375732421875,
0.484130859375
],
[
0.5009765625,
0.468505859375,
0.62158203125,
0.468505859375,
0.62158203125,
0.486083984375,
0.5009765625,
0.486083984375
],
[
0.86279296875,
0.47021484375,
0.88232421875,
0.47021484375,
0.88232421875,
0.484130859375,
0.86279296875,
0.484130859375
],
[
0.498291015625,
0.488037109375,
0.529296875,
0.488037109375,
0.529296875,
0.496826171875,
0.498291015625,
0.496826171875
],
[
0.12091064453125,
0.501953125,
0.88134765625,
0.501953125,
0.88134765625,
0.51513671875,
0.12091064453125,
0.51513671875
],
[
0.11767578125,
0.517578125,
0.859375,
0.517578125,
0.859375,
0.52783203125,
0.11767578125,
0.52783203125
],
[
0.11517333984375,
0.55029296875,
0.5107421875,
0.55029296875,
0.5107421875,
0.564453125,
0.11517333984375,
0.564453125
],
[
0.11767578125,
0.580078125,
0.8798828125,
0.580078125,
0.8798828125,
0.59326171875,
0.11767578125,
0.59326171875
],
[
0.11767578125,
0.59326171875,
0.88134765625,
0.59326171875,
0.88134765625,
0.60693359375,
0.11767578125,
0.60693359375
],
[
0.1192626953125,
0.609375,
0.8798828125,
0.609375,
0.8798828125,
0.619140625,
0.1192626953125,
0.619140625
],
[
0.11767578125,
0.6220703125,
0.498291015625,
0.6220703125,
0.498291015625,
0.63525390625,
0.11767578125,
0.63525390625
],
[
0.11767578125,
0.6416015625,
0.88134765625,
0.6416015625,
0.88134765625,
0.65478515625,
0.11767578125,
0.65478515625
],
[
0.1168212890625,
0.65576171875,
0.84716796875,
0.65576171875,
0.84716796875,
0.66943359375,
0.1168212890625,
0.66943359375
],
[
0.11517333984375,
0.671875,
0.8798828125,
0.673828125,
0.8798828125,
0.68798828125,
0.11517333984375,
0.68603515625
],
[
0.11517333984375,
0.68603515625,
0.23779296875,
0.6875,
0.23779296875,
0.70068359375,
0.11517333984375,
0.69970703125
],
[
0.1544189453125,
0.7109375,
0.371826171875,
0.7109375,
0.371826171875,
0.724609375,
0.1544189453125,
0.724609375
],
[
0.15283203125,
0.73486328125,
0.78173828125,
0.73486328125,
0.78173828125,
0.74853515625,
0.15283203125,
0.74853515625
],
[
0.4306640625,
0.755859375,
0.6259765625,
0.755859375,
0.6259765625,
0.76953125,
0.4306640625,
0.76953125
],
[
0.1748046875,
0.7744140625,
0.4248046875,
0.77587890625,
0.4248046875,
0.79052734375,
0.1748046875,
0.7890625
],
[
0.11767578125,
0.7998046875,
0.88134765625,
0.7998046875,
0.88134765625,
0.81298828125,
0.11767578125,
0.81298828125
],
[
0.11767578125,
0.814453125,
0.398681640625,
0.814453125,
0.398681640625,
0.82763671875,
0.11767578125,
0.82763671875
],
[
0.12091064453125,
0.8359375,
0.87841796875,
0.8359375,
0.87841796875,
0.8466796875,
0.12091064453125,
0.8466796875
],
[
0.11517333984375,
0.8466796875,
0.88134765625,
0.84765625,
0.88134765625,
0.8623046875,
0.11517333984375,
0.861328125
],
[
0.12091064453125,
0.8642578125,
0.384033203125,
0.8642578125,
0.384033203125,
0.875,
0.12091064453125,
0.875
],
[
0.136474609375,
0.88525390625,
0.88134765625,
0.88525390625,
0.88134765625,
0.8984375,
0.136474609375,
0.8984375
],
[
0.1356201171875,
0.89794921875,
0.41015625,
0.8984375,
0.41015625,
0.912109375,
0.1356201171875,
0.9111328125
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94873046875,
0.49267578125,
0.94873046875
]
] |
[
"AnInterpretableDeterminantalChoiceModelforSubsetSelection",
"确定点进程。确定性点过程（DPP）是子集上的概率分布，有利于不同的子集。固定分类A上的DPP的特征",
"在于其核矩阵LBorodin[2009]。这是一个|A|×|A|对称p。S。d。矩阵4。观察子集CCA的可能性与内",
"核L的子矩阵Lc的对数行列式成正比，该子矩阵对应于C中的行和列：",
"Pr[Choose C] = det(I + L)",
"det(Lc)",
"(4)",
"Lo的行列式取为1。分母是归一化常数的闭式表达式，满足ZBCAdet(LB）=det(I+L)，其中I表示与",
"L大小相同的单位矩阵。DPP支持通过贝叶斯推理和MLE「KuleszaandTaskar，2011，Affandietal.，2014l进行",
"监督学习的方法。此外，对于DPP[Houghetal.,2006]，还有有效的采样算法。有关DPP的、和属性的更多",
"详细信息，请参阅Kuleszaetal.[2012]",
"DPP似然有利于选定子集中的多样性。这是使用两个项目示例[Kuleszaetal.,2012]演示的。选择单例i]的",
"可能性与核矩阵的对角项成正比。",
"Pr[Choose (i]] α det(Lfi) = Li",
"另一方面，选择两个项目i，的非归一化可能性由非对角项L讠调节。",
"Pr[Choose [i, j]] α det (",
"LiLi",
"LjiLii",
"=LiLi-L",
"可以将非正项一L,视为对包含相似项的子集的惩罚，其中非对角项L捕获了i和j之间的相似性概念",
"将DPP内核分解为相似性和质量组件，可为模型增加清晰度。KuleszaandTaskar[2010]表明任何DPP内核",
"L都可以分解。t。Li=qiSiqi，其中qi是项目i的质量，Si是项目i和j之间的相似性。质量被视为项",
"目i的相对吸引力；相似性量化了两个项目之间的相似度。相似项收集在满足0二S二I5的矩阵S中。在",
"此分解下，DPP似然因子：",
"Pr[Choose C] α",
"II d2 ) det(Sc).",
"(5)",
"liEC",
"上述公式表明，包含项目的子集C的似然与项目i的平方质量成正比。多样性的偏差源于相似性子矩阵的",
"行列式；行列式取0和1之间的值。我们利用这种分解来弥合DPP和随机效用子集选择模型之间的差距",
"3Determinantal Processes & Random Utility",
"本节连接确定性过程和基于随机效用最大化的子集选择模型。特别是，我们表明，在特定的效用函数族下",
"所有离散的确定性点过程都可以视为随机效用最大化选择。。此外，研究这一系列效用函数会对DPP的行为",
"假设以及DPP与现有子集选择模型之间的联系产生新的见解。。本节从基于特征的参数视图退后一步，并全",
"面考虑离散DPP；我们返回Section4中的特征和参数",
"我们的第一个结果表明，来自DPP的样本是随机的效用最大化选择。结果是对“子集上是否存在随机效用函",
"数，使得效用最大化选择与DPP具有相同的分布”问题的肯定答案？答案似乎很明显；主要思想是问。",
"Theorem 3.1. Let A be a set of n items and L an n -by- n p.s.d. symmetric matrix. The following distributions over",
"2A are equivalent:",
"1. A DPP over A with kernel L.",
"2. The utility-maximizing subset where each subset C of A has a log-determinant random utility",
"U(C) = log det(Lc) + e(C)",
"with e(C) id standard Gumbel errors.",
"对此的证明以及所有后续结果都推迟到AppendixA。该定理表明，DPP是子集上的等效MNL，具有对数行",
"列式效用。我们称之为DPP隐含的效用",
"隐含效用的简化有助于将DPP连接到现有的子集选择模型。令V（C）表示子集C的隐含效用的非随机分量",
"对于单例C=（i)，非随机效用等于平方质量u(i）=V（i）=log(q）的对数。将核矩阵分解为Equation（5)",
"中的质量和相似性项，得出以下定理。",
"4The abbreviation p.s.d. stands positive semi-definite. An n X n matrix K is p.s.d. if for all z E IRn , it holds that zT K z ≥ 0.",
"'Here KK' denotes that K'-K is p.s.d",
"3"
] |
{
"arxiv_id": "2302.11477",
"language": "zh",
"page_idx": 2
}
|
||
2302.11477-zh-3
|
2302.11477-zh-3.jpg
|
An Interpretable Determinantal Choice Model for Subset Selection
Theorem 3.2. Let L be an n - by- $n\textit{ DPP kernel. Then the non- random part of the induced utility of subset }C\subseteq A$ is
$$\begin{aligned}V(C)=\sum_{i\in C}v(i)+\log\det(\boldsymbol{S}_C)\end{aligned}$$
换句话说，可以将隐含的效用视为“具有校正的加法”。在这种情况下，校正是依赖于相似性的惩罚项$\log\det(\boldsymbol{S}_C)$。这个词是非正的，因为$S\preceq I$。此外，DPP 对数似然是亚模的( Kulesza et al.[2012], pp。22) 将 DPP 视为子集选择模型有助于将 DPP 与离散选择和效用最大化联系起来。DPP 是子集上的 MNL 模型这一事实立即意味着 DPP 在子集级别 [Train,2009]上服从无关替代 (IIA)属性的独立性；在$C^\prime$ 上选择子集 $C$ 的相对概率不取决于分类中是否存在某些第三子集$C^{\prime\prime}$。这是许多选择模型的基本公理。效用函数也是信息丰富的。隐含效用的亚模数意味着 DPP 选择表现出边际收益递减；在选择中增加一项的额外效用在选择的大小上是不增加的。边际效用递减在效用理论中有着悠久的历史
我们的第二个结果表明，DPP 所暗示的效用是“修正的加法”。这有助于突出显示 DPP 与现有子集选择模型的不同之处。实际上，从子集选择的角度来看，DPP 具有新颖的属性。与之相反，e。g。, 成对交互模型，DPP 隐含的效用能够捕获全局依赖关系，不受集合大小的限制。同样，与 Benson et al.[2018] 的模型不同，DPF 模型不限制接收校正的集合数量。相反，DPP 限制了它可以捕获的交互的符号；更正都是非正数。这是子集选择建模中的一种新方法，使 DPP 成为子集选择建模库的宝贵补充。
## 4 A Determinantal Choice Model
本节定义了一个基于特征的参数化 DPP, 我们称之为确定性选择模型。本节的主要结果在该模型，逻辑回归和 MNL 之间建立了牢固的联系，从而为确定性模型组件(尤其是参数)提供了直观且熟悉的解释。该方法表明，确定性模型族占据了模型的连续性，其中一个极端包含逻辑回归，另一个极端包含 MNL
首先定义确定性选择模型。我们从 Affandi et al.[2014]的条件 DPP 和 Błaszczyszyn and Keeler[2019] 的确定性稀疏点过程中获得灵感。我们指定一个内核函数，而不是固定的内核，该函数接受任何分类数据$X_A$并生成 DPP 内核$L(X_A)$。我们通过分别指定参数质量和相似性模型来做到这一点。令$\beta$为$d$维系数向量。质量模型是指数线性模型[Kulesza and Taskar,2010]:
$$q(\beta,\boldsymbol{x}_i)=\exp\left\{\frac{1}{2}\boldsymbol{\beta}^T\boldsymbol{x}_i\right\}.$$
(6)
在此质量模型下，对于任何 p, DPP 似然在参数$\beta$中是对数凹的。s。d。相似性模型 Kulesza and Taskar[2010]
。在本文中，相似性模型被指定为具有长度尺度$\ell\in\mathbb{R}^d$的各向异性 RBF 核：
(7)
$$S(\boldsymbol\ell,\boldsymbol x_i,\boldsymbol x_j)=\exp\left\{-\frac{1}{2}\sum_{k=1}^d\frac{|x_{i,k}-x_{j,k}|^2}{\ell_k^2}\right\}.$$
虽然我们更喜欢 RBF 的可解释性，但本节中的结果对任何 p都有效。s。d。带有$S_ij\in[0,1]$的内核函数。这就完成了确定性选择模型；本节的其余部分展示了该模型如何与逻辑回归和 MNL 相关，并讨论了该模型参数的直觉和解释
本节的主要见解是将子集选择视为具有依赖标签的二进制分类。二进制标签表示选择；要么选择了一个项目$i\in C$ ,要么没有$,i\notin C$ 。因此，预测选定的子集等同于预测每个项目的二进制标签。标签可以依赖。例如在负依赖性的情况下，对一个项目的选择条件可能会降低其他项目被选择的可能性。在这种观点下，如果确定性模型在不同项目之间没有相似性，则应简化为假定标签之间独立的常规二进制分类模型。同样，如果所有项目都是最大程度地负相关的-即互斥的-则确定性模型应类似于最多选择一个项目的分类模型
我们的第一个结果在完全不同的项目的情况下形式化了这种直觉。在这种相似结构下，模型等价于逻辑回
归。
Theorem $4. 1. \textit{Fix afinite assortment A and data }\boldsymbol{X}_A. If\boldsymbol{S}( \boldsymbol{X}_A) = \boldsymbol{I}, then$ the determinantal likelihood of$C\in Ais$
$$\Pr[Choose\:C]=\prod\limits_{i\in C}\left(\frac{e^{\boldsymbol{\beta}^T\boldsymbol{x}_i}}{1+e^{\boldsymbol{\beta}^T\boldsymbol{x}_i}}\right)\prod\limits_{j\in A\setminus C}\left(\frac{1}{1+e^{\boldsymbol{\beta}^T\boldsymbol{x}_j}}\right).$$
这是 Equation (2) 的逻辑回归的可能性，将$A$中的项目视为独立观察。由于此结果，可以将系数$\beta$解释为该
设置中的逻辑回归系数
在最大负依赖标签的极端情况下，项目是互斥的。在此设置中，确定性模型等效于项目上的 MNL。
4
|
[
[
0.1168212890625,
0.044189453125,
0.55224609375,
0.044189453125,
0.55224609375,
0.058074951171875,
0.1168212890625,
0.058074951171875
],
[
0.1192626953125,
0.0947265625,
0.87744140625,
0.0947265625,
0.87744140625,
0.10479736328125,
0.1192626953125,
0.10479736328125
],
[
0.39208984375,
0.117431640625,
0.60546875,
0.117431640625,
0.60546875,
0.1319580078125,
0.39208984375,
0.1319580078125
],
[
0.450927734375,
0.1331787109375,
0.4794921875,
0.1331787109375,
0.4794921875,
0.147705078125,
0.450927734375,
0.147705078125
],
[
0.1168212890625,
0.1622314453125,
0.88134765625,
0.1622314453125,
0.88134765625,
0.175537109375,
0.1168212890625,
0.175537109375
],
[
0.11517333984375,
0.1749267578125,
0.88134765625,
0.175537109375,
0.88134765625,
0.1900634765625,
0.11517333984375,
0.189453125
],
[
0.1192626953125,
0.1976318359375,
0.8798828125,
0.1976318359375,
0.8798828125,
0.2083740234375,
0.1192626953125,
0.2083740234375
],
[
0.131591796875,
0.2108154296875,
0.8798828125,
0.2108154296875,
0.8798828125,
0.2215576171875,
0.131591796875,
0.2215576171875
],
[
0.1192626953125,
0.2235107421875,
0.8798828125,
0.2235107421875,
0.8798828125,
0.2374267578125,
0.1192626953125,
0.2374267578125
],
[
0.1192626953125,
0.238037109375,
0.88134765625,
0.238037109375,
0.88134765625,
0.251953125,
0.1192626953125,
0.251953125
],
[
0.11767578125,
0.25244140625,
0.57177734375,
0.25244140625,
0.57177734375,
0.26318359375,
0.11767578125,
0.26318359375
],
[
0.11767578125,
0.271484375,
0.88134765625,
0.271484375,
0.88134765625,
0.285400390625,
0.11767578125,
0.285400390625
],
[
0.11767578125,
0.285400390625,
0.88134765625,
0.285400390625,
0.88134765625,
0.2998046875,
0.11767578125,
0.2998046875
],
[
0.1192626953125,
0.2998046875,
0.88134765625,
0.2998046875,
0.88134765625,
0.313232421875,
0.1192626953125,
0.313232421875
],
[
0.1168212890625,
0.313232421875,
0.88134765625,
0.313232421875,
0.88134765625,
0.326904296875,
0.1168212890625,
0.326904296875
],
[
0.11517333984375,
0.32568359375,
0.60107421875,
0.326904296875,
0.60107421875,
0.341552734375,
0.11517333984375,
0.34033203125
],
[
0.11517333984375,
0.3623046875,
0.410888671875,
0.3623046875,
0.410888671875,
0.375732421875,
0.11517333984375,
0.375732421875
],
[
0.1192626953125,
0.393310546875,
0.8798828125,
0.393310546875,
0.8798828125,
0.404052734375,
0.1192626953125,
0.404052734375
],
[
0.11517333984375,
0.404052734375,
0.88232421875,
0.40478515625,
0.88232421875,
0.419189453125,
0.11517333984375,
0.41845703125
],
[
0.1168212890625,
0.419189453125,
0.7900390625,
0.419189453125,
0.7900390625,
0.43310546875,
0.1168212890625,
0.43310546875
],
[
0.11767578125,
0.43994140625,
0.88134765625,
0.43994140625,
0.88134765625,
0.453857421875,
0.11767578125,
0.453857421875
],
[
0.11767578125,
0.453857421875,
0.88134765625,
0.453857421875,
0.88134765625,
0.46728515625,
0.11767578125,
0.46728515625
],
[
0.1192626953125,
0.46728515625,
0.88134765625,
0.46728515625,
0.88134765625,
0.48095703125,
0.1192626953125,
0.48095703125
],
[
0.1168212890625,
0.48095703125,
0.45751953125,
0.48095703125,
0.45751953125,
0.494384765625,
0.1168212890625,
0.494384765625
],
[
0.398681640625,
0.4873046875,
0.60107421875,
0.50048828125,
0.59619140625,
0.54541015625,
0.393798828125,
0.5322265625
],
[
0.560546875,
0.5087890625,
0.580078125,
0.5087890625,
0.580078125,
0.51220703125,
0.560546875,
0.51220703125
],
[
0.86279296875,
0.51220703125,
0.88232421875,
0.51220703125,
0.88232421875,
0.52587890625,
0.86279296875,
0.52587890625
],
[
0.57275390625,
0.5263671875,
0.58251953125,
0.521484375,
0.58642578125,
0.52587890625,
0.57666015625,
0.53076171875
],
[
0.11767578125,
0.54443359375,
0.8798828125,
0.54443359375,
0.8798828125,
0.5546875,
0.11767578125,
0.5546875
],
[
0.11767578125,
0.55615234375,
0.65673828125,
0.55615234375,
0.65673828125,
0.5693359375,
0.11767578125,
0.5693359375
],
[
0.471435546875,
0.5791015625,
0.64453125,
0.58056640625,
0.64453125,
0.60302734375,
0.471435546875,
0.60107421875
],
[
0.3447265625,
0.59130859375,
0.48779296875,
0.59326171875,
0.48779296875,
0.60791015625,
0.3447265625,
0.60595703125
],
[
0.86279296875,
0.59228515625,
0.88232421875,
0.59228515625,
0.88232421875,
0.60595703125,
0.86279296875,
0.60595703125
],
[
0.473876953125,
0.5986328125,
0.62744140625,
0.5986328125,
0.62744140625,
0.619140625,
0.473876953125,
0.619140625
],
[
0.6396484375,
0.60693359375,
0.646484375,
0.60693359375,
0.646484375,
0.60986328125,
0.6396484375,
0.60986328125
],
[
0.11767578125,
0.62744140625,
0.88134765625,
0.62744140625,
0.88134765625,
0.6416015625,
0.11767578125,
0.6416015625
],
[
0.11767578125,
0.642578125,
0.88134765625,
0.642578125,
0.88134765625,
0.65576171875,
0.11767578125,
0.65576171875
],
[
0.1192626953125,
0.6572265625,
0.2320556640625,
0.6572265625,
0.2320556640625,
0.66748046875,
0.1192626953125,
0.66748046875
],
[
0.11767578125,
0.67529296875,
0.88232421875,
0.67529296875,
0.88232421875,
0.689453125,
0.11767578125,
0.689453125
],
[
0.1168212890625,
0.68798828125,
0.87841796875,
0.68798828125,
0.87841796875,
0.7021484375,
0.1168212890625,
0.7021484375
],
[
0.1168212890625,
0.7021484375,
0.88134765625,
0.70263671875,
0.88134765625,
0.71728515625,
0.1168212890625,
0.71630859375
],
[
0.1192626953125,
0.71728515625,
0.88134765625,
0.71728515625,
0.88134765625,
0.73095703125,
0.1192626953125,
0.73095703125
],
[
0.11767578125,
0.73095703125,
0.7958984375,
0.73095703125,
0.7958984375,
0.744140625,
0.11767578125,
0.744140625
],
[
0.1168212890625,
0.75048828125,
0.8798828125,
0.75048828125,
0.8798828125,
0.7646484375,
0.1168212890625,
0.7646484375
],
[
0.11767578125,
0.7666015625,
0.1380615234375,
0.7666015625,
0.1380615234375,
0.77880859375,
0.11767578125,
0.77880859375
],
[
0.1168212890625,
0.7841796875,
0.88134765625,
0.7841796875,
0.88134765625,
0.79736328125,
0.1168212890625,
0.79736328125
],
[
0.47216796875,
0.80810546875,
0.5146484375,
0.80810546875,
0.5146484375,
0.82275390625,
0.47216796875,
0.82275390625
],
[
0.303955078125,
0.8173828125,
0.441162109375,
0.81640625,
0.441162109375,
0.83056640625,
0.303955078125,
0.83203125
],
[
0.55615234375,
0.8193359375,
0.57666015625,
0.8193359375,
0.57666015625,
0.83349609375,
0.55615234375,
0.83349609375
],
[
0.6357421875,
0.814453125,
0.640625,
0.814453125,
0.640625,
0.8193359375,
0.6357421875,
0.8193359375
],
[
0.449462890625,
0.82763671875,
0.529296875,
0.82373046875,
0.53125,
0.83837890625,
0.450927734375,
0.84228515625
],
[
0.59326171875,
0.8271484375,
0.67333984375,
0.82373046875,
0.6748046875,
0.83837890625,
0.5947265625,
0.84130859375
],
[
0.419189453125,
0.833984375,
0.44775390625,
0.833984375,
0.44775390625,
0.8466796875,
0.419189453125,
0.8466796875
],
[
0.5439453125,
0.8359375,
0.58642578125,
0.8359375,
0.58642578125,
0.84765625,
0.5439453125,
0.84765625
],
[
0.1168212890625,
0.8623046875,
0.88134765625,
0.86279296875,
0.88134765625,
0.87744140625,
0.1168212890625,
0.876953125
],
[
0.11767578125,
0.87744140625,
0.281005859375,
0.87744140625,
0.281005859375,
0.8916015625,
0.11767578125,
0.8916015625
],
[
0.1168212890625,
0.8984375,
0.814453125,
0.8984375,
0.814453125,
0.912109375,
0.1168212890625,
0.912109375
],
[
0.490234375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.94970703125,
0.490234375,
0.94970703125
]
] |
[
"An Interpretable Determinantal Choice Model for Subset Selection",
"Theorem3.2.LetLbeann-by-nDPPkernel.Thenthenon-randompartofthe induced utilityof subsetCCAis",
"V(C) =/ v(i) + log det(Sc)",
"iEC",
"换句话说，可以将隐含的效用视为“具有校正的加法”。在这种情况下，校正是依赖于相似性的惩罚项",
"logdet(Sc）。这个词是非正的，因为S≤I。此外，DPP对数似然是亚模的（Kuleszaetal.[2012]，Pp。22)",
"将DPP视为子集选择模型有助于将DPP与离散选择和效用最大化联系起来。DPP是子集上的MNL模型这",
"事实立即意味着DPP在子集级别[Train，2009]上服从无关替代（IIA）属性的独立性；在C上选择子集C",
"的相对概率不取决于分类中是否存在某些第三子集C\"。这是许多选择模型的基本公理。效用函数也是信息",
"丰富的。隐含效用的亚模数意味着DPP选择表现出边际收益递减；在选择中增加一项的额外效用在选择的",
"大小上是不增加的。边际效用递减在效用理论中有着悠久的历史",
"我们的第二个结果表明，DPP所暗示的效用是“修正的加法”。这有助于突出显示DPP与现有子集选择模型",
"的不同之处。实际上，从子集选择的角度来看，DPP具有新颖的属性。与之相反，e。g。，成对交互模型，DPP",
"隐含的效用能够捕获全局依赖关系，不受集合大小的限制。同样，与Bensonetal.[2018]的模型不同，DPP",
"模型不限制接收校正的集合数量。相反，DPP限制了它可以捕获的交互的符号；更正都是非正数。这是子集",
"选择建模中的一种新方法，使DPP成为子集选择建模库的宝贵补充。",
"4A Determinantal Choice Model",
"本节定义了二个基于特征的参数化DPP，我们称之为确定性选择模型。本节的主要结果在该模型，逻辑回归",
"和MNL之间建立了牢固的联系，从而为确定性模型组件（尤其是参数）提供了直观且熟悉的解释。该方法",
"表明，确定性模型族占据了模型的连续性，其中一个极端包含逻辑回归，另一个极端包含MNL",
"首先定义确定性选择模型。我们从Affandietal.[2014]的条件DPP和BtaszczyszynandKeeler[2019]的确定",
"性稀疏点过程中获得灵感。我们指定一个内核函数，而不是固定的内核，该函数接受任何分类数据XA并生",
"成DPP内核L(XA）。我们通过分别指定参数质量和相似性模型来做到这一点。令β为d维系数向量。质量",
"模型是指数线性模型[KuleszaandTaskar,2010]：",
"a(0. ) o /p0ra .",
"",
"(6)",
"/",
"在此质量模型下，对于任何p，DPP似然在参数β中是对数凹的。s。d。相似性模型KuleszaandTaskar[2010",
"。在本文中，相似性模型被指定为具有长度尺度lERd的各向异性RBF核：",
"[_1ik -rk]",
"s(e,ai,aj)=exp",
"(7)",
"12",
"",
"虽然我们更喜欢RBF的可解释性，但本节中的结果对任何p都有效。S。d。带有SiE[O,1]的内核函数。这",
"就完成了确定性选择模型；本节的其余部分展示了该模型如何与逻辑回归和MNL相关，并讨论了该模型参",
"数的直觉和解释",
"本节的主要见解是将子集选择视为具有依赖标签的二进制分类。二进制标签表示选择；要么选择了一个项目",
"ie，要么没有，i更C。因此，预测选定的子集等同于预测每个项目的二进制标签。标签可以依赖。例如",
"在负依赖性的情况下，对一个项目的选择条件可能会降低其他项目被选择的可能性。在这种观点下，如果确",
"定性模型在不同项目之间没有相似性，则应简化为假定标签之间独立的常规二进制分类模型。同样，如果所",
"有项目都是最大程度地负相关的-即互斥的-则确定性模型应类似于最多选择一个项目的分类模型",
"我们的第一个结果在完全不同的项目的情况下形式化了这种直觉。在这种相似结构下，模型等价于逻辑回",
"归。",
"Theorem 4.1. Fix a finite assortment A and data XA.If S(XA) =I, then the determinantal likelihood of C E A is",
"eBTai",
"Pr[Choose C] =T",
"",
"1",
"(1+eBTa)",
"(1+eBTa)",
"iEC",
"jEA\\C",
"这是Equation(2）的逻辑回归的可能性，将A中的项目视为独立观察。由于此结果，可以将系数β解释为该",
"设置中的逻辑回归系数",
"在最大负依赖标签的极端情况下，项目是互斥的。在此设置中，确定性模型等效于项目上的MNL。",
"4"
] |
{
"arxiv_id": "2302.11477",
"language": "zh",
"page_idx": 3
}
|
||
2302.11477-zh-4
|
2302.11477-zh-4.jpg
|
An Interpretable Determinantal Choice Model for Subset Selection
Theorem 4.2. Fix afinite assortment A and data $X_A.If\boldsymbol{S}(\boldsymbol{X}_A)=\boldsymbol{J}$ (the all 1 s matrix) then the determinantal
likelihood $of\left \{ i\right \} for$ $i\in C$ $is$
$$\Pr[Choose\left\{i\right\}]=\frac{\exp\{\boldsymbol{\beta}^T\boldsymbol{x}_i\}}{1+\sum_{j\in C}\exp\{\boldsymbol{\beta}^T\boldsymbol{x}_j\}}$$
With Pr$[Choose\emptyset]\propto1,and0forallothersets.$
可能性正是 MNL 模型在 Equation (1) 中的分类$A$上的可能性。类似地，每当项目高度相似时，决定性选择模型参数可以解释为 MNL 系数。在极端之间，解释仍然成立。人们总是可以查看非随机效用$\beta^Tx_i$ ,就像在逻辑回归模型下一样，通过限定“不存在其他类似项目”。当存在竞争项目时，选择项目$i$的边际概率由以下因素调节：(a)$i$与竞争项目之间的负依赖程度，以及(b)竞争项目的效用。确切的概率和对它的边际效应可以通过取导数和评估分类和其他项目来评估，例如。g。,平均值，这在解释 MNL 参数时很常见。
## 5 $\textbf{A comparison of models}$
上一节显示，每当相似矩阵取极值时，确定性选择模型都包含逻辑回归和 MNL。然而，模型在实践中是否做到这一点取决于学习“正确”的参数。本节使用合成数据来检查训练的确定性选择与逻辑回归和 MNL 的比较，因为训练数据中的负依赖程度是变化的。我们的实验表明，该模型的行为与理论一致；它确实与分类模型在负依赖的两个极端处的性能非常匹配。此外，我们发现，当负依赖程度介于两个极端之间时，确定性选择模型优于两个参考模型
实验设计。实验将每个模型的样本外预测性能估计为所选子集中负依赖性程度的函数。这需要生成一系列具有不同程度的负依赖性的数据集。这是使用数据生成过程完成的，在该过程中，负相关性由标量$r>0$参数化。。三个模型中的每一个都在相同的特征上进行参数化，并在相同的数据集序列上进行训练。在训练期间从每个数据集中保留评估数据。最后，在相同的评估数据集上评估每个模型
<FigureHere>
Figure 1: Illustration of one observation with 15 items and radius $r=0.5.$ Points in light grey are independently assigned 0 -labels in the first thinning phase. Light blue points are discarded during Mat é rn thinning; dark blue points retain their 1 -labels.
数据生成。合成数据分三个步骤生成。每个观察结果————个分类选择对——都是独立生成的。我们只描述一次观察的过程。首先，对项目及其特征进行抽样。将项目编码为$[-2,2]^2$ 正方形中的点，并随机均匀采样。每个分类有 15 个项目。$i$第点的特征$\boldsymbol{X}_i$是其$x$和$y$坐标以及从该点到原点的距离$d_i$ 。与每个试验相关联
5
|
[
[
0.1168212890625,
0.043548583984375,
0.55224609375,
0.043548583984375,
0.55224609375,
0.05682373046875,
0.1168212890625,
0.05682373046875
],
[
0.11767578125,
0.09344482421875,
0.88232421875,
0.09344482421875,
0.88232421875,
0.106689453125,
0.11767578125,
0.106689453125
],
[
0.1168212890625,
0.106689453125,
0.30810546875,
0.106689453125,
0.30810546875,
0.12060546875,
0.1168212890625,
0.12060546875
],
[
0.3544921875,
0.125,
0.6396484375,
0.1363525390625,
0.63818359375,
0.1610107421875,
0.35302734375,
0.149658203125
],
[
0.5185546875,
0.121826171875,
0.6005859375,
0.11932373046875,
0.6005859375,
0.1337890625,
0.5185546875,
0.1363525390625
],
[
0.11517333984375,
0.1590576171875,
0.428955078125,
0.160400390625,
0.428955078125,
0.1749267578125,
0.11517333984375,
0.173583984375
],
[
0.11767578125,
0.186279296875,
0.88134765625,
0.186279296875,
0.88134765625,
0.199462890625,
0.11767578125,
0.199462890625
],
[
0.11767578125,
0.2008056640625,
0.88134765625,
0.2008056640625,
0.88134765625,
0.2139892578125,
0.11767578125,
0.2139892578125
],
[
0.11517333984375,
0.212158203125,
0.8798828125,
0.21337890625,
0.8798828125,
0.2279052734375,
0.11517333984375,
0.2266845703125
],
[
0.12091064453125,
0.2279052734375,
0.88134765625,
0.2279052734375,
0.88134765625,
0.2412109375,
0.12091064453125,
0.2412109375
],
[
0.11767578125,
0.242431640625,
0.8349609375,
0.242431640625,
0.8349609375,
0.255615234375,
0.11767578125,
0.255615234375
],
[
0.11517333984375,
0.276611328125,
0.35302734375,
0.276611328125,
0.35302734375,
0.291748046875,
0.11517333984375,
0.291748046875
],
[
0.1217041015625,
0.30810546875,
0.8798828125,
0.30810546875,
0.8798828125,
0.318115234375,
0.1217041015625,
0.318115234375
],
[
0.11767578125,
0.32080078125,
0.88134765625,
0.32080078125,
0.88134765625,
0.333984375,
0.11767578125,
0.333984375
],
[
0.11767578125,
0.333984375,
0.88134765625,
0.333984375,
0.88134765625,
0.347900390625,
0.11767578125,
0.347900390625
],
[
0.1168212890625,
0.3466796875,
0.88134765625,
0.347900390625,
0.88134765625,
0.3623046875,
0.1168212890625,
0.361083984375
],
[
0.1168212890625,
0.3623046875,
0.313720703125,
0.3623046875,
0.313720703125,
0.375732421875,
0.1168212890625,
0.375732421875
],
[
0.11767578125,
0.3818359375,
0.88134765625,
0.3818359375,
0.88134765625,
0.395751953125,
0.11767578125,
0.395751953125
],
[
0.11767578125,
0.396484375,
0.88134765625,
0.396484375,
0.88134765625,
0.410400390625,
0.11767578125,
0.410400390625
],
[
0.11767578125,
0.410400390625,
0.88134765625,
0.410400390625,
0.88134765625,
0.423583984375,
0.11767578125,
0.423583984375
],
[
0.11767578125,
0.4248046875,
0.65625,
0.4248046875,
0.65625,
0.438232421875,
0.11767578125,
0.438232421875
],
[
0.28759765625,
0.48486328125,
0.313720703125,
0.48486328125,
0.313720703125,
0.498779296875,
0.28759765625,
0.498779296875
],
[
0.340576171875,
0.50830078125,
0.3505859375,
0.51025390625,
0.34716796875,
0.517578125,
0.338134765625,
0.51513671875
],
[
0.28759765625,
0.51806640625,
0.310546875,
0.51806640625,
0.310546875,
0.5322265625,
0.28759765625,
0.5322265625
],
[
0.53515625,
0.51318359375,
0.55224609375,
0.51318359375,
0.55224609375,
0.5263671875,
0.53515625,
0.5263671875
],
[
0.38818359375,
0.52294921875,
0.40185546875,
0.52294921875,
0.40185546875,
0.53076171875,
0.38818359375,
0.53076171875
],
[
0.28759765625,
0.55322265625,
0.316162109375,
0.55322265625,
0.316162109375,
0.5673828125,
0.28759765625,
0.5673828125
],
[
0.3544921875,
0.5517578125,
0.36767578125,
0.5517578125,
0.36767578125,
0.56298828125,
0.3544921875,
0.56298828125
],
[
0.6708984375,
0.57275390625,
0.6845703125,
0.57275390625,
0.6845703125,
0.58349609375,
0.6708984375,
0.58349609375
],
[
0.285888671875,
0.5859375,
0.314453125,
0.5859375,
0.314453125,
0.60302734375,
0.285888671875,
0.60302734375
],
[
0.406005859375,
0.5966796875,
0.419189453125,
0.5966796875,
0.419189453125,
0.60498046875,
0.406005859375,
0.60498046875
],
[
0.50244140625,
0.599609375,
0.515625,
0.599609375,
0.515625,
0.611328125,
0.50244140625,
0.611328125
],
[
0.28759765625,
0.62255859375,
0.316162109375,
0.62255859375,
0.316162109375,
0.63720703125,
0.28759765625,
0.63720703125
],
[
0.490234375,
0.6298828125,
0.50634765625,
0.6298828125,
0.50634765625,
0.64453125,
0.490234375,
0.64453125
],
[
0.279296875,
0.6572265625,
0.314453125,
0.6572265625,
0.314453125,
0.66943359375,
0.279296875,
0.66943359375
],
[
0.279296875,
0.6923828125,
0.314453125,
0.6923828125,
0.314453125,
0.705078125,
0.279296875,
0.705078125
],
[
0.48046875,
0.68994140625,
0.5,
0.68994140625,
0.5,
0.7041015625,
0.48046875,
0.7041015625
],
[
0.55126953125,
0.693359375,
0.564453125,
0.693359375,
0.564453125,
0.7041015625,
0.55126953125,
0.7041015625
],
[
0.498291015625,
0.70068359375,
0.51318359375,
0.6962890625,
0.5185546875,
0.70703125,
0.50341796875,
0.7109375
],
[
0.279296875,
0.7265625,
0.314453125,
0.7265625,
0.314453125,
0.7392578125,
0.279296875,
0.7392578125
],
[
0.449462890625,
0.73095703125,
0.46728515625,
0.73095703125,
0.46728515625,
0.744140625,
0.449462890625,
0.744140625
],
[
0.5849609375,
0.73779296875,
0.59619140625,
0.73486328125,
0.6005859375,
0.7421875,
0.5888671875,
0.74560546875
],
[
0.61279296875,
0.734375,
0.71142578125,
0.734375,
0.71142578125,
0.744140625,
0.61279296875,
0.744140625
],
[
0.58642578125,
0.75,
0.5986328125,
0.75,
0.5986328125,
0.7587890625,
0.58642578125,
0.7587890625
],
[
0.61376953125,
0.7470703125,
0.70751953125,
0.7470703125,
0.70751953125,
0.75830078125,
0.61376953125,
0.75830078125
],
[
0.279296875,
0.76123046875,
0.313720703125,
0.76123046875,
0.313720703125,
0.7734375,
0.279296875,
0.7734375
],
[
0.58642578125,
0.76318359375,
0.5986328125,
0.76318359375,
0.5986328125,
0.77294921875,
0.58642578125,
0.77294921875
],
[
0.61279296875,
0.76220703125,
0.66845703125,
0.76220703125,
0.66845703125,
0.77294921875,
0.61279296875,
0.77294921875
],
[
0.318603515625,
0.78271484375,
0.53759765625,
0.7841796875,
0.53759765625,
0.79736328125,
0.318603515625,
0.796875
],
[
0.5986328125,
0.7861328125,
0.6259765625,
0.7861328125,
0.6259765625,
0.79736328125,
0.5986328125,
0.79736328125
],
[
0.63623046875,
0.78515625,
0.71142578125,
0.78271484375,
0.71142578125,
0.796875,
0.63623046875,
0.798828125
],
[
0.62744140625,
0.78857421875,
0.64208984375,
0.78857421875,
0.64208984375,
0.79541015625,
0.62744140625,
0.79541015625
],
[
0.5361328125,
0.7890625,
0.54833984375,
0.7890625,
0.54833984375,
0.79345703125,
0.5361328125,
0.79345703125
],
[
0.54736328125,
0.787109375,
0.576171875,
0.787109375,
0.576171875,
0.79541015625,
0.54736328125,
0.79541015625
],
[
0.1168212890625,
0.81494140625,
0.8798828125,
0.81494140625,
0.8798828125,
0.8291015625,
0.1168212890625,
0.8291015625
],
[
0.11517333984375,
0.82763671875,
0.88134765625,
0.8291015625,
0.88134765625,
0.84326171875,
0.11517333984375,
0.84228515625
],
[
0.11358642578125,
0.84228515625,
0.294189453125,
0.84228515625,
0.294189453125,
0.85595703125,
0.11358642578125,
0.85595703125
],
[
0.1168212890625,
0.87060546875,
0.496826171875,
0.87060546875,
0.496826171875,
0.8837890625,
0.1168212890625,
0.8837890625
],
[
0.52783203125,
0.869140625,
0.88134765625,
0.87060546875,
0.88134765625,
0.8837890625,
0.52783203125,
0.88330078125
],
[
0.12744140625,
0.88330078125,
0.88134765625,
0.8837890625,
0.88134765625,
0.8984375,
0.12744140625,
0.89794921875
],
[
0.11767578125,
0.8984375,
0.88232421875,
0.8984375,
0.88232421875,
0.912109375,
0.11767578125,
0.912109375
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94873046875,
0.49267578125,
0.94873046875
]
] |
[
"An Interpretable Determinantal Choice Model for Subset Selection",
"Theorem 4.2. Fix a finite assortment A and data XA :If S(XA) =J (the all 1 s matrix) then the determinantal",
"likelihoodof [i] forieCis",
"Pr[Choose (i] = 1+Zjec exp[BT a]",
"exp(βTci)",
"With Pr[Choose O] α I , and O for all other sets",
"可能性正是MNL模型在Equation（1）中的分类A上的可能性。类似地，每当项目高度相似时，决定性选择",
"模型参数可以解释为MNL系数。在极端之间，解释仍然成立。人们总是可以查看非随机效用βai，就像",
"在逻辑回归模型下一样，通过限定“不存在其他类似项目”。当存在竞争项目时，选择项目的边际概率由以",
"下因素调节：（a）i与竞争项目之间的负依赖程度，以及（b）竞争项目的效用。确切的概率和对它的边际效",
"应可以通过取导数和评估分类和其他项目来评估，例如。g。，平均值，这在解释MNL参数时很常见。",
"5A comparison of models",
"上一节显示：每当相似矩阵取极值时，确定性选择模型都包含逻辑回归和MNL。然而，模型在实践中是否",
"做到这一点取决于学习“正确”的参数。本节使用合成数据来检查训练的确定性选择与逻辑回归和MNL的",
"比较，因为训练数据中的负依赖程度是变化的。我们的实验表明，该模型的行为与理论一致；它确实与分类",
"模型在负依赖的两个极端处的性能非常匹配。此外，我们发现，当负依赖程度介于两个极端之间时，确定性",
"选择模型优于两个参考模型",
"实验设计。实验将每个模型的样本外预测性能估计为所选子集中负依赖性程度的函数。这需要生成一系列具",
"有不同程度的负依赖性的数据集。这是使用数据生成过程完成的，在该过程中，负相关性由标量r>0参数",
"化。。三个模型中的每一个都在相同的特征上进行参数化，并在相同的数据集序列上进行训练。在训练期间",
"从每个数据集中保留评估数据。最后，在相同的评估数据集上评估每个模型",
"2.0",
"福",
"1.5",
"0",
"",
"1.0",
"O",
"O",
"0.5 ",
"",
"O",
"0.0 -",
"O",
"-0.5",
"1.0",
"O",
"C",
"O",
"1.5",
"O",
"?",
"Poissonthinned",
"?",
"Matern thinned",
"-2.0",
"",
"Retained",
"-2.0-1.51.0-0.50.0",
"1.0",
"1.52.0",
"",
"",
"0.5",
"Figure 1:Illustration of one observation with 15 items and radius r = 0.5 . Points in light grey are independently",
"assigned O -labels in the first thinning phase. Light blue points are discarded during Mat é rn thinning; dark blue",
"points retain their 1 -labels",
"数据生成。合成数据分三个步骤生成。每个观察结果",
"一个分类选择对一都是独立生成的。我们只描述",
"-次观察的过程。首先，对项目及其特征进行抽样。将项目编码为[-2,2]2正方形中的点，并随机均匀采样。",
"每个分类有15个项目。i第点的特征X，是其a和y坐标以及从该点到原点的距离di。与每个试验相关联",
"5"
] |
{
"arxiv_id": "2302.11477",
"language": "zh",
"page_idx": 4
}
|
||
2205.02289-zh-1
|
2205.02289-zh-1.jpg
|
用于药物组合的 N 元关系提取的数据集
$\begin{aligned}\textbf{Aryeh Tiktinsky}^{*,1}&\text{Vijay Viswanathan}^{*,2}&\textbf{Danna Niezni}^{3}&\textbf{Dana Meron Azagury}^{3}\\\textbf{Yosi Shamay}^{3}&\textbf{Hillel Taub-Tabib}^{1}&\textbf{Tom Hope}^{1,4}&\textbf{Yoav Goldberg}^{1,5}\\^{1}\text{Allen Instinte for Al}^{2}&\text{Janouage Technologies Institute Carneoie Mel}\end{aligned}$
$^{3}$Faculty of Biomedical Engineering, Technion -Israel Institute of Technology
$^{4}$School of Computer Science & Engineering, The Hebrew University of Jerusalem
$^{5}$Computer Science Department, Bar Ilan University
aryeht@allenai.org
vijayveandrew.cmu.edu
# Abstract
Combination therapies have become the standard of care for diseases such as cancer, tuberculosis, malaria and HIV. However, the combinatorial set of available multi-drug treatments creates a challenge in identifying effective combination therapies available in a situation. To assist medical professionals in identifying beneficial drug-combinations, we construct an expert-annotated dataset for extracting information about the efficacy of drug combinations from the scientific literature. Beyond its practical utility, the dataset also presents a unique NLP challenge, as the first relation extraction dataset consisting of variable-length relations. Furthermore, the relations in this dataset predominantly require language understanding beyond the sentence level, adding to the challenge of this task. We provide a promising baseline model and identify clear areas for further improvement. We release our dataset, $^{1}$code, $^2$and baseline models$^3$publicly to encourage the NLP community to participate in this task.
2005)、高血压(Rochlani et al.,2017)和 COVID-
19 (Ianevski et al., 2020)
在这项工作中，我们检查了从科学文献中提取已知药物组合的临床上重要且具有挑战性的NLP 任务。我们为这项新任务提供了一个专家注释的数据集和基线模型。我们的数据集包含1600 个手动注释的摘要，每个摘要都提到 2 到15 种药物。这些摘要中有 840 篇描述了一种或多种阳性药物组合，其大小从 2 到 11 种药物不等。剩下的 760 篇摘要要么提到未联合使用的药物，要么讨论未产生联合积极作用的药物组合
对于临床环境，解决药物组合识别任务可以帮助研究人员建议和验证复杂的治疗计划。例如，在寻找有效的癌症治疗方法时，了解哪些药物与一线治疗协同作用，使研究人员能够提出新的治疗计划，随后可以在体内验证并成为标准方案 (Wasserman et al., 2001; Katzir et al., 2019; Ianevski et al., 2020; Niezni et al., 2022)
从 NLP 的角度来看，药物组合识别任务和数据集通过引入具有几个挑战性特征的关系提取任务，推动了关系提取(RE)研究的界限： 可变长度 n 元关系关系提取的大多数工作都集中在二元关系上 (e。g。Liet al.(2016),请参见 §5 中的完整列表), 或与固定$n$ (e。g。Peng et al.(2017))。相反，药物组合任务涉及可变长度的 n 元关系：不同的段落讨论了不同数量的药物的组合。对于段落中提到的每个药物子集，模型必须预测它们是否在联合治疗中一起使用，以及这种药物组合是否有效。
无类型提示如 Rosenman et al.(2020) 和 Sabo et al.(2021)所述，在许多关系提取基准(Han et al., 2018; Sabo et al., 2021; Zhang et al., 2017) 中，参数类型可作为有效的线索。但是，参数类型并不自然地适用于药物组合任务，其中所有可能的关系参数都是相同类型(药物)的实体，我们需要确定它们的特定子集。
远程依赖描述组合功效的信息通常分散在多个句子中。实际上，我们的注释者报告说，对于67 个% 实例，无法根据单个句子确定标签，因此需要使用更大的文本上下文进行推理。有趣
### 1 Introduction
“到目前为止，已经测试了许多单一疗法，但已被证明对 COVID-19 的疗效有限。相比之下， 联合疗法正在成为治疗 SARS-CoV-2 感染的有用工具。" (Ianevski et al., 2021)
事实上，将两种或多种药物联合使用已被证明可用于治疗各种疾病，包括癌症(DeVita et al., 1975; Carew et al., 2008; Shuhendler et al., 2010)、艾滋病 (Bartlett et al.,2006)、疟疾 (Eastman and Fidock,2009)、肺结核(Bhusal et al.,
$^{*}$ Equal contribution.
$^{1}$https: / / huggingface. co/ datasets/
allenai/drug-combo-extraction
$^2$https://github.com/allenai/
drug-combo-extraction
$^3$https://huggingface.co/allenai/
drug- combo- classifier- pubmedbert- dapt
|
[
[
0.2939453125,
0.0843505859375,
0.70361328125,
0.0843505859375,
0.70361328125,
0.09857177734375,
0.2939453125,
0.09857177734375
],
[
0.1318359375,
0.10687255859375,
0.869140625,
0.105712890625,
0.869140625,
0.127685546875,
0.1318359375,
0.12890625
],
[
0.18896484375,
0.1258544921875,
0.8095703125,
0.122314453125,
0.8095703125,
0.146728515625,
0.18896484375,
0.149658203125
],
[
0.1502685546875,
0.1455078125,
0.85400390625,
0.1484375,
0.85400390625,
0.166259765625,
0.1502685546875,
0.163330078125
],
[
0.187255859375,
0.163330078125,
0.8154296875,
0.166259765625,
0.8154296875,
0.18408203125,
0.187255859375,
0.18115234375
],
[
0.16796875,
0.182861328125,
0.83642578125,
0.185302734375,
0.83642578125,
0.20068359375,
0.16796875,
0.198974609375
],
[
0.291259765625,
0.20068359375,
0.70947265625,
0.203125,
0.70947265625,
0.2203369140625,
0.291259765625,
0.218505859375
],
[
0.30908203125,
0.221435546875,
0.476806640625,
0.221435546875,
0.476806640625,
0.236328125,
0.30908203125,
0.236328125
],
[
0.50146484375,
0.22265625,
0.69287109375,
0.22265625,
0.69287109375,
0.236328125,
0.50146484375,
0.236328125
],
[
0.2626953125,
0.2529296875,
0.340087890625,
0.2529296875,
0.340087890625,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.51123046875,
0.2529296875,
0.8818359375,
0.254150390625,
0.8818359375,
0.268310546875,
0.51123046875,
0.267333984375
],
[
0.51123046875,
0.26953125,
0.6943359375,
0.26953125,
0.6943359375,
0.283203125,
0.51123046875,
0.283203125
],
[
0.154541015625,
0.283203125,
0.459228515625,
0.283203125,
0.459228515625,
0.29638671875,
0.154541015625,
0.29638671875
],
[
0.53076171875,
0.289306640625,
0.880859375,
0.289306640625,
0.880859375,
0.302734375,
0.53076171875,
0.302734375
],
[
0.145263671875,
0.29638671875,
0.45751953125,
0.29638671875,
0.45751953125,
0.31005859375,
0.145263671875,
0.31005859375
],
[
0.51123046875,
0.303955078125,
0.87890625,
0.303955078125,
0.87890625,
0.317626953125,
0.51123046875,
0.317626953125
],
[
0.145263671875,
0.31005859375,
0.456787109375,
0.31005859375,
0.456787109375,
0.32373046875,
0.145263671875,
0.32373046875
],
[
0.509765625,
0.319580078125,
0.880859375,
0.320556640625,
0.880859375,
0.3349609375,
0.509765625,
0.334228515625
],
[
0.14697265625,
0.32666015625,
0.45751953125,
0.32666015625,
0.45751953125,
0.33740234375,
0.14697265625,
0.33740234375
],
[
0.145263671875,
0.339111328125,
0.459228515625,
0.339111328125,
0.459228515625,
0.352783203125,
0.145263671875,
0.352783203125
],
[
0.51318359375,
0.33740234375,
0.87890625,
0.33740234375,
0.87890625,
0.350830078125,
0.51318359375,
0.350830078125
],
[
0.14697265625,
0.35400390625,
0.45751953125,
0.352783203125,
0.45751953125,
0.366455078125,
0.14697265625,
0.36767578125
],
[
0.51318359375,
0.352783203125,
0.87890625,
0.352783203125,
0.87890625,
0.366455078125,
0.51318359375,
0.366455078125
],
[
0.145263671875,
0.36767578125,
0.459228515625,
0.36767578125,
0.459228515625,
0.38134765625,
0.145263671875,
0.38134765625
],
[
0.51123046875,
0.36767578125,
0.880859375,
0.3681640625,
0.880859375,
0.383056640625,
0.51123046875,
0.3818359375
],
[
0.1485595703125,
0.38427734375,
0.456787109375,
0.38427734375,
0.456787109375,
0.394287109375,
0.1485595703125,
0.394287109375
],
[
0.5146484375,
0.385498046875,
0.880859375,
0.385498046875,
0.880859375,
0.399658203125,
0.5146484375,
0.399658203125
],
[
0.145263671875,
0.396728515625,
0.459228515625,
0.3955078125,
0.459228515625,
0.4091796875,
0.145263671875,
0.410400390625
],
[
0.5146484375,
0.402099609375,
0.880859375,
0.402099609375,
0.880859375,
0.4150390625,
0.5146484375,
0.4150390625
],
[
0.1485595703125,
0.412109375,
0.45751953125,
0.412109375,
0.45751953125,
0.4228515625,
0.1485595703125,
0.4228515625
],
[
0.51123046875,
0.41748046875,
0.880859375,
0.416259765625,
0.880859375,
0.43115234375,
0.51123046875,
0.431640625
],
[
0.14697265625,
0.424560546875,
0.45751953125,
0.424560546875,
0.45751953125,
0.438232421875,
0.14697265625,
0.438232421875
],
[
0.51123046875,
0.432861328125,
0.5517578125,
0.432861328125,
0.5517578125,
0.448974609375,
0.51123046875,
0.448974609375
],
[
0.145263671875,
0.439453125,
0.459228515625,
0.439453125,
0.459228515625,
0.453125,
0.145263671875,
0.453125
],
[
0.1485595703125,
0.454833984375,
0.456787109375,
0.454833984375,
0.456787109375,
0.465576171875,
0.1485595703125,
0.465576171875
],
[
0.5322265625,
0.453125,
0.880859375,
0.453125,
0.880859375,
0.466064453125,
0.5322265625,
0.466064453125
],
[
0.145263671875,
0.46728515625,
0.45751953125,
0.46728515625,
0.45751953125,
0.48095703125,
0.145263671875,
0.48095703125
],
[
0.5146484375,
0.468505859375,
0.880859375,
0.468505859375,
0.880859375,
0.482177734375,
0.5146484375,
0.482177734375
],
[
0.145263671875,
0.48095703125,
0.45751953125,
0.48095703125,
0.45751953125,
0.49462890625,
0.145263671875,
0.49462890625
],
[
0.51123046875,
0.48388671875,
0.880859375,
0.482666015625,
0.880859375,
0.49755859375,
0.51123046875,
0.498779296875
],
[
0.145263671875,
0.496337890625,
0.45751953125,
0.496337890625,
0.45751953125,
0.51025390625,
0.145263671875,
0.51025390625
],
[
0.51318359375,
0.50048828125,
0.880859375,
0.50048828125,
0.880859375,
0.51416015625,
0.51318359375,
0.51416015625
],
[
0.14697265625,
0.51025390625,
0.45751953125,
0.51025390625,
0.45751953125,
0.52392578125,
0.14697265625,
0.52392578125
],
[
0.5146484375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.5146484375,
0.53076171875
],
[
0.14697265625,
0.52392578125,
0.45751953125,
0.52392578125,
0.45751953125,
0.53759765625,
0.14697265625,
0.53759765625
],
[
0.5146484375,
0.5322265625,
0.880859375,
0.5322265625,
0.880859375,
0.54638671875,
0.5146484375,
0.54638671875
],
[
0.145263671875,
0.5380859375,
0.45751953125,
0.5380859375,
0.45751953125,
0.5517578125,
0.145263671875,
0.5517578125
],
[
0.145263671875,
0.552734375,
0.45751953125,
0.552734375,
0.45751953125,
0.56640625,
0.145263671875,
0.56640625
],
[
0.51318359375,
0.548828125,
0.8564453125,
0.548828125,
0.8564453125,
0.5625,
0.51318359375,
0.5625
],
[
0.145263671875,
0.56787109375,
0.45751953125,
0.56787109375,
0.45751953125,
0.58056640625,
0.145263671875,
0.58056640625
],
[
0.52880859375,
0.5654296875,
0.880859375,
0.56640625,
0.880859375,
0.58056640625,
0.52880859375,
0.580078125
],
[
0.1435546875,
0.58056640625,
0.45751953125,
0.58203125,
0.45751953125,
0.5966796875,
0.1435546875,
0.595703125
],
[
0.5146484375,
0.5830078125,
0.880859375,
0.5830078125,
0.880859375,
0.5966796875,
0.5146484375,
0.5966796875
],
[
0.145263671875,
0.5966796875,
0.279541015625,
0.5966796875,
0.279541015625,
0.60986328125,
0.145263671875,
0.60986328125
],
[
0.51318359375,
0.599609375,
0.875,
0.599609375,
0.875,
0.61328125,
0.51318359375,
0.61328125
],
[
0.5146484375,
0.615234375,
0.87890625,
0.615234375,
0.87890625,
0.62890625,
0.5146484375,
0.62890625
],
[
0.5146484375,
0.63134765625,
0.880859375,
0.63134765625,
0.880859375,
0.6455078125,
0.5146484375,
0.6455078125
],
[
0.1175537109375,
0.63818359375,
0.151123046875,
0.63818359375,
0.151123046875,
0.6494140625,
0.1175537109375,
0.6494140625
],
[
0.151123046875,
0.63818359375,
0.26123046875,
0.63818359375,
0.26123046875,
0.6494140625,
0.151123046875,
0.6494140625
],
[
0.51318359375,
0.6455078125,
0.880859375,
0.64794921875,
0.880859375,
0.662109375,
0.51318359375,
0.66015625
],
[
0.118408203125,
0.66455078125,
0.48779296875,
0.66455078125,
0.48779296875,
0.677734375,
0.118408203125,
0.677734375
],
[
0.509765625,
0.662109375,
0.880859375,
0.6630859375,
0.880859375,
0.677734375,
0.509765625,
0.6767578125
],
[
0.1175537109375,
0.6787109375,
0.489501953125,
0.6796875,
0.489501953125,
0.69482421875,
0.1175537109375,
0.693359375
],
[
0.5146484375,
0.6796875,
0.880859375,
0.6796875,
0.880859375,
0.693359375,
0.5146484375,
0.693359375
],
[
0.118408203125,
0.69677734375,
0.48779296875,
0.69677734375,
0.48779296875,
0.71044921875,
0.118408203125,
0.71044921875
],
[
0.5146484375,
0.6953125,
0.880859375,
0.6953125,
0.880859375,
0.708984375,
0.5146484375,
0.708984375
],
[
0.118408203125,
0.7119140625,
0.365234375,
0.7119140625,
0.365234375,
0.7255859375,
0.118408203125,
0.7255859375
],
[
0.51318359375,
0.71044921875,
0.8818359375,
0.71142578125,
0.8818359375,
0.7255859375,
0.51318359375,
0.724609375
],
[
0.1368408203125,
0.728515625,
0.486083984375,
0.728515625,
0.486083984375,
0.7421875,
0.1368408203125,
0.7421875
],
[
0.51318359375,
0.72802734375,
0.7958984375,
0.72802734375,
0.7958984375,
0.7412109375,
0.51318359375,
0.7412109375
],
[
0.1175537109375,
0.74462890625,
0.486083984375,
0.74462890625,
0.486083984375,
0.75830078125,
0.1175537109375,
0.75830078125
],
[
0.51318359375,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.1175537109375,
0.76123046875,
0.489501953125,
0.76123046875,
0.489501953125,
0.77490234375,
0.1175537109375,
0.77490234375
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.1175537109375,
0.77685546875,
0.48779296875,
0.77685546875,
0.48779296875,
0.79052734375,
0.1175537109375,
0.79052734375
],
[
0.51318359375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.1175537109375,
0.79345703125,
0.489501953125,
0.79345703125,
0.489501953125,
0.80712890625,
0.1175537109375,
0.80712890625
],
[
0.51318359375,
0.79150390625,
0.87890625,
0.79150390625,
0.87890625,
0.8046875,
0.51318359375,
0.8046875
],
[
0.5146484375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.145263671875,
0.81640625,
0.273681640625,
0.81640625,
0.273681640625,
0.830078125,
0.145263671875,
0.830078125
],
[
0.1376953125,
0.8291015625,
0.43408203125,
0.830078125,
0.43408203125,
0.84423828125,
0.1376953125,
0.84326171875
],
[
0.51318359375,
0.82470703125,
0.87890625,
0.82470703125,
0.87890625,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.12005615234375,
0.84326171875,
0.38037109375,
0.84326171875,
0.38037109375,
0.85400390625,
0.12005615234375,
0.85400390625
],
[
0.51318359375,
0.84130859375,
0.7958984375,
0.84130859375,
0.7958984375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.1368408203125,
0.85400390625,
0.3896484375,
0.85498046875,
0.3896484375,
0.869140625,
0.1368408203125,
0.8681640625
],
[
0.51318359375,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.12005615234375,
0.87060546875,
0.154541015625,
0.87060546875,
0.154541015625,
0.87890625,
0.12005615234375,
0.87890625
],
[
0.172119140625,
0.869140625,
0.30908203125,
0.8681640625,
0.30908203125,
0.8798828125,
0.172119140625,
0.88134765625
],
[
0.5146484375,
0.87353515625,
0.880859375,
0.87353515625,
0.880859375,
0.88720703125,
0.5146484375,
0.88720703125
],
[
0.1376953125,
0.8798828125,
0.4248046875,
0.88134765625,
0.4248046875,
0.8955078125,
0.1376953125,
0.89453125
],
[
0.51123046875,
0.88720703125,
0.880859375,
0.88818359375,
0.880859375,
0.90283203125,
0.51123046875,
0.90185546875
],
[
0.118408203125,
0.8955078125,
0.453369140625,
0.8955078125,
0.453369140625,
0.90625,
0.118408203125,
0.90625
],
[
0.51318359375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"用于药物组合的N元关系提取的数据集",
"Aryeh Tiktinsky*1 Vijay Viswanathan*,2 Danna Niezni? Dana Meron Azagury",
"Yosi Shamay3Hillel Taub-Tabib' Tom Hope'4Yoav Goldberg'.s",
"1Allen Institute for AI2Language Technologies Institute, Carnegie Mellon University",
"3Faculty of Biomedical Engineering, Technion -Israel Institute of Technology",
"4School of Computer Science &Engineering,The Hebrew University of Jerusalem",
"5Computer Science Department, Bar Ilan University",
"aryeht@allenai.org",
"vijayv@andrew.cmu.edu",
"Abstract",
"2005）、高血压（Rochlanietal.，2017）和COVID",
"19 (lanevski et al., 2020)",
"Combination therapies have become the stan-",
"在这项工作中，我们检查了从科学文献中提",
"dard of care for diseases such as cancer, tu-",
"取已知药物组合的临床上重要且具有挑战性的",
"berculosis, malaria and HIV. However, the",
"NLP任务。我们为这项新任务提供了一个专家",
"combinatorial set of available multi-drug treat-",
"ments creates a challenge in identifying effec-",
"注释的数据集和基线模型。我们的数据集包含",
"tive combination therapies available in a situ-",
"1600个手动注释的摘要，每个摘要都提到2到",
"ation. To assist medical professionals in iden-",
"15种药物。这些摘要中有840篇描述了一种或",
"tifyingbeneficial drug-combinations,wecon",
"多种阳性药物组合，其大小从2到11种药物",
"struct an expert-annotated dataset for extract-",
"不等。剩下的760篇摘要要么提到未联合使用",
"inginformationabouttheefficacyofdrugcom-",
"的药物，要么讨论未产生联合积极作用的药物",
"binations from the scientific literature. Beyond",
"组合",
"its practical utility, the dataset also presents a",
"uniqueNLP challenge,as thefirstrelation ex",
"对于临床环境，解决药物组合识别任务可以",
"traction dataset consisting of variable-length",
"帮助研究人员建议和验证复杂的治疗计划。例",
"relations. Furthermore, the relations in this",
"如，在寻找有效的癌症治疗方法时，了解哪些",
"dataset predominantly require language under",
"药物与一线治疗协同作用，使研究人员能够提",
"standing beyond the sentence level, adding",
"出新的治疗计划，随后可以在体内验证并成为",
"to the challenge of this task.We provide a",
"标准方案（Wasserman et al.,2001Katzir et al.,",
"promising baseline model and identify clear ar",
"eas for further improvement. We release our",
"2019; lanevski et al., 2020; Niezni et al., 2022)",
"dataset, code, and baseline models 3pub",
"从NLP的角度来看，药物组合识别任务和",
"licly to encourage the NLP community to par-",
"数据集通过引人具有几个挑战性特征的关系提",
"ticipate in this task.",
"取任务，推动了关系提取（RE）研究的界限：",
"可变长度n元关系关系提取的大多数工作都集",
"中在二元关系上（e。g。Lietal.(2016)，请参",
"1",
"Introduction",
"见s5中的完整列表)，或与固定n（e。g。Peng",
"“到目前为止，已经测试了许多单一疗法，但",
"etal.(2017)）。相反，药物组合任务涉及可变",
"已被证明对COVID-19的疗效有限。相比之下，",
"长度的n元关系：不同的段落讨论了不同数量",
"联合疗法正在成为治疗SARS-CoV-2感染的有",
"的药物的组合。对于段落中提到的每个药物子",
"用工具。\"(lanevskietal.,2021)",
"集，模型必须预测它们是否在联合治疗中一起",
"事实上，将两种或多种药物联合使用已被",
"使用，以及这种药物组合是否有效。",
"证明可用于治疗各种疾病，包括癌症(DeVita",
"无类型提示如Rosenmanetal.（2020）和Sabo",
"et al., 1975; Carew et al., 2008; Shuhendler et al.,",
"etal.（2021）所述，在许多关系提取基准（Han",
"2010）、艾滋病（Bartlettetal.,2006）、症疾（East-",
"et al., 2018; Sabo et al., 2021; Zhang et al., 2017)",
"man and Fidock,2009)、肺结核(Bhusal et al.,",
"中，参数类型可作为有效的线索。但是，参数",
"类型并不自然地适用于药物组合任务，其中所",
"* Equal contribution.",
"'https://huggingface.co/datasets/",
"有可能的关系参数都是相同类型（药物）的实",
"allenai/drug-combo-extraction",
"体，我们需要确定它们的特定子集。",
"2https://github.com/allenai/",
"远程依赖描述组合功效的信息通常分散在多个",
"drug",
"ombo-extraction",
"句子中。实际上，我们的注释者报告说，对于",
"\"https://huggingface.co/allenai/",
"67个%实例，无法根据单个句子确定标签，因",
"drug-combo-classifier-pubmedbert-dapt",
"此需要使用更大的文本上下文进行推理。有趣"
] |
{
"arxiv_id": "2205.02289",
"language": "zh",
"page_idx": 1
}
|
||
2205.02289-zh-3
|
2205.02289-zh-3.jpg
|
<table>
<tbody>
<tr>
<th>Metric</th>
<th>Partial $\textbf{l Match}$</th>
<th>Exact $\mathbf{Match}$</th>
</tr>
<tr>
<td>Avg. Any Combination Fl</td>
<td>88.9</td>
<td>86.1</td>
</tr>
<tr>
<td>Avg. Positive Combination F1</td>
<td>83.4</td>
<td>79.6</td>
</tr>
</tbody>
</table>
Table 1: Agreement scores using our adaptation of F1
score to allow for partial-match.
中药物提及的$m>=2$ 跨度的集合。输出是一组$R= \{ ( c_i, y_i) \}$ ,其中$c_i\in\mathcal{P}(D)$是来自$\mathcal{P}(D)$的药物组合，所有可能的药物组合的集合，$y_i\in\{$POS\_COMB,OTHER\_COMB}是组合标签
# 2.3 Evaluation Metric
我们考虑两个设置：“精确匹配”,一个考虑识别精确药物组合的严格版本，以及“部分匹配”,一个为正确识别的子集分配部分学分的更宽松的版本
我们对这两个设置都使用标准精度，召回率和 F1 指标。对于部分匹配的情况，我们将给定组合的二进制 0 或 1 分数替换为精细分数：$shared\_ drugs/ total\_ drugs$ . 如果有多个与黄金关系的部分匹配，我们取重叠最大的一个。我们将召回率计算为$identified\_relations/all\_gold\_relations$
,精 度计算 为
$correct\_relations/identified\_relations$
我们考虑两个指标，将 POS\_COMB 与其余指标进行比较的平均正组合F1 得分 ,以及对任何组合标签 (POS 或其他)而不是 NO\_COMB 进行正确预测的平均任何组合 F1 得分。后者是一项更容易的任务，但对于识别药物组合仍然很有价值，无论其疗效如何
## 2.4 Collecting Data for Annotation
为了收集注释数据，我们从 DrugBank$^{6}$收集了2411种药物的列表，并从 PubMed 中抽取了一组提到 2 种或更多药物的句子。对该样本的前50 个句子的分析表明，只有 8/50 的句子中提到了药物组合。这意味着注释完整的样本将是昂贵的，并且将导致数据集高度偏向相对琐碎的 NO\_COMB 实例
因此，我们重复了此实验，对 PubMed 摘要中包含触发短语的句子进行了采样。??? 个抽样句子中的 48 个% 包括提及药物组合。针对具有已知药物组合的摘要的新样本评估触发列表的覆盖范围表明，这些新摘要中的 90 个% 包含一个触发词。这表明我们的触发列表可用于获取标签平衡数据，而不会过度限制覆盖范围和多样性
因此，我们收集了大多数注释实例 90%,使用基本搜索包含至少两种不同药物且其摘要包含触发短语之一的句子。为了克服触发列表所施加的词汇限制，我们使用远程监督对剩余的 10 个% 实例进行采样：根据 DrugComboDB
$^{6}$Curation included downloading a premade drug list from DrugBank's website, while removing non pharmacological intervention such as Vitamins and Supplements. The later we got from the FDA orange book.
获取包含已知具有协同作用的药物对的句子， 但其摘要不包括我们的触发短语之一。所有数据收集查询均使用 SPIKEExtractiveSearch 工具(Shlain et al., 2020; Taub-Tabib et al., 2020) 执行。该过程在 Figure 2 的顶部进行了说明
### 2.5 $\textbf{The Annotation Process}$
七名生物医学工程研究生参加了注释任务。学生们都完成了癌症联合疗法的课程，并由一位在这一领域具有专业知识的有原则的研究人员进行监督
我们为参与者提供了注释指南，其中指定了注释过程应如何进行(请参阅附录 A.1),并进行了一次初次会议，我们与小组一起审查了指南并讨论了一些示例
每个参与者都可以访问 Prodigy 注释工具(Montani and Honnibal, 2018)的单独实例，该实例预先加载了候选注释实例。会话开始后， 实例 (包含带有标记药物实体的句子及其上下文)以顺序方式出现，没有时间限制。对于每个实例，我们指示注释者标记参与组合的所有药物子集，并为每个子集指示其标签(POS\_COMB 或 OTHER\_COMB)。此外，我们指示他们指出是否需要上下文以确定关系的积极功效
尽管专家注释需要大量时间，但我们还是收集了 1634 篇文章的注释。其中，272 个被分配给至少两个注释者。在首席研究员进一步仲裁后，这些被用于测试集。该过程在 Figure 2 的底部进行了说明
### 2.6 Inter-annotator Agreement
在任务过程中，我们多次计算了注释者之间的协议，以识别分歧的情况并向注释者提供反馈。每次随机选择一组 25 个实例，并将其分配给所有注释器。协议是基于成对 F1 度量计算的(如 82.3 中所述进行了一些修改),并对所有注释者对进行了平均(请参见附录 A.2 中对替代度量的讨论)
Table 1 中的最终协议编号是令人满意的
(Aroyo and Welty, 2013; Araki et al., 2018)
|
[
[
0.59033203125,
0.0843505859375,
0.63232421875,
0.0843505859375,
0.63232421875,
0.0955810546875,
0.59033203125,
0.0955810546875
],
[
0.7080078125,
0.0831298828125,
0.8759765625,
0.0831298828125,
0.8759765625,
0.0968017578125,
0.7080078125,
0.0968017578125
],
[
0.53662109375,
0.0968017578125,
0.68603515625,
0.0968017578125,
0.68603515625,
0.10687255859375,
0.53662109375,
0.10687255859375
],
[
0.12005615234375,
0.08489990234375,
0.484375,
0.08489990234375,
0.484375,
0.09857177734375,
0.12005615234375,
0.09857177734375
],
[
0.73486328125,
0.097412109375,
0.76171875,
0.097412109375,
0.76171875,
0.105712890625,
0.73486328125,
0.105712890625
],
[
0.8193359375,
0.0968017578125,
0.84814453125,
0.0968017578125,
0.84814453125,
0.10809326171875,
0.8193359375,
0.10809326171875
],
[
0.1318359375,
0.10272216796875,
0.483642578125,
0.10272216796875,
0.483642578125,
0.1175537109375,
0.1318359375,
0.1175537109375
],
[
0.7314453125,
0.105712890625,
0.76171875,
0.10809326171875,
0.759765625,
0.12054443359375,
0.72900390625,
0.1175537109375
],
[
0.8212890625,
0.1051025390625,
0.8486328125,
0.1092529296875,
0.84619140625,
0.1217041015625,
0.81787109375,
0.1175537109375
],
[
0.1175537109375,
0.1181640625,
0.483642578125,
0.1181640625,
0.483642578125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.52294921875,
0.1092529296875,
0.69287109375,
0.1092529296875,
0.69287109375,
0.119384765625,
0.52294921875,
0.119384765625
],
[
0.118408203125,
0.1353759765625,
0.48779296875,
0.1353759765625,
0.48779296875,
0.149658203125,
0.118408203125,
0.149658203125
],
[
0.5087890625,
0.13720703125,
0.87890625,
0.1383056640625,
0.87890625,
0.152587890625,
0.5087890625,
0.1519775390625
],
[
0.12005615234375,
0.1519775390625,
0.173828125,
0.1519775390625,
0.173828125,
0.163330078125,
0.12005615234375,
0.163330078125
],
[
0.51318359375,
0.152587890625,
0.72900390625,
0.152587890625,
0.72900390625,
0.166259765625,
0.51318359375,
0.166259765625
],
[
0.1158447265625,
0.1776123046875,
0.30908203125,
0.1776123046875,
0.30908203125,
0.191162109375,
0.1158447265625,
0.191162109375
],
[
0.118408203125,
0.198974609375,
0.486083984375,
0.198974609375,
0.486083984375,
0.213134765625,
0.118408203125,
0.213134765625
],
[
0.51318359375,
0.201904296875,
0.87744140625,
0.201904296875,
0.87744140625,
0.215576171875,
0.51318359375,
0.215576171875
],
[
0.118408203125,
0.215576171875,
0.486083984375,
0.215576171875,
0.486083984375,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.51123046875,
0.2161865234375,
0.880859375,
0.21728515625,
0.880859375,
0.232177734375,
0.51123046875,
0.23095703125
],
[
0.1175537109375,
0.23095703125,
0.486083984375,
0.23095703125,
0.486083984375,
0.2452392578125,
0.1175537109375,
0.2452392578125
],
[
0.51123046875,
0.2340087890625,
0.87744140625,
0.2340087890625,
0.87744140625,
0.2476806640625,
0.51123046875,
0.2476806640625
],
[
0.12005615234375,
0.248779296875,
0.2283935546875,
0.248779296875,
0.2283935546875,
0.260009765625,
0.12005615234375,
0.260009765625
],
[
0.51318359375,
0.2493896484375,
0.880859375,
0.2493896484375,
0.880859375,
0.26318359375,
0.51318359375,
0.26318359375
],
[
0.1368408203125,
0.26416015625,
0.484375,
0.26416015625,
0.484375,
0.27783203125,
0.1368408203125,
0.27783203125
],
[
0.51318359375,
0.2666015625,
0.80419921875,
0.2666015625,
0.80419921875,
0.27978515625,
0.51318359375,
0.27978515625
],
[
0.1175537109375,
0.279052734375,
0.484375,
0.279052734375,
0.484375,
0.292236328125,
0.1175537109375,
0.292236328125
],
[
0.118408203125,
0.29638671875,
0.486083984375,
0.29638671875,
0.486083984375,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.51318359375,
0.30712890625,
0.54248046875,
0.30712890625,
0.54248046875,
0.318359375,
0.51318359375,
0.318359375
],
[
0.5400390625,
0.311279296875,
0.54833984375,
0.311279296875,
0.54833984375,
0.3154296875,
0.5400390625,
0.3154296875
],
[
0.5556640625,
0.30712890625,
0.748046875,
0.30712890625,
0.748046875,
0.320556640625,
0.5556640625,
0.320556640625
],
[
0.1175537109375,
0.32666015625,
0.486083984375,
0.327880859375,
0.486083984375,
0.342529296875,
0.1175537109375,
0.341552734375
],
[
0.118408203125,
0.312255859375,
0.468505859375,
0.312255859375,
0.468505859375,
0.325927734375,
0.118408203125,
0.325927734375
],
[
0.51318359375,
0.3349609375,
0.87890625,
0.3349609375,
0.87890625,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.118408203125,
0.344482421875,
0.486083984375,
0.344482421875,
0.486083984375,
0.358154296875,
0.118408203125,
0.358154296875
],
[
0.51123046875,
0.350830078125,
0.880859375,
0.350830078125,
0.880859375,
0.364013671875,
0.51123046875,
0.364013671875
],
[
0.1158447265625,
0.35986328125,
0.444091796875,
0.35986328125,
0.444091796875,
0.37353515625,
0.1158447265625,
0.37353515625
],
[
0.51318359375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.51318359375,
0.3818359375
],
[
0.187255859375,
0.377197265625,
0.21240234375,
0.377197265625,
0.21240234375,
0.391357421875,
0.187255859375,
0.391357421875
],
[
0.214111328125,
0.377197265625,
0.280517578125,
0.377197265625,
0.280517578125,
0.391357421875,
0.214111328125,
0.391357421875
],
[
0.31494140625,
0.374755859375,
0.420654296875,
0.374755859375,
0.420654296875,
0.392578125,
0.31494140625,
0.392578125
],
[
0.418212890625,
0.37890625,
0.4609375,
0.37890625,
0.4609375,
0.3896484375,
0.418212890625,
0.3896484375
],
[
0.4609375,
0.377685546875,
0.486083984375,
0.377685546875,
0.486083984375,
0.39013671875,
0.4609375,
0.39013671875
],
[
0.414794921875,
0.38134765625,
0.4189453125,
0.38134765625,
0.4189453125,
0.38720703125,
0.414794921875,
0.38720703125
],
[
0.28955078125,
0.38134765625,
0.3046875,
0.38134765625,
0.3046875,
0.385498046875,
0.28955078125,
0.385498046875
],
[
0.51318359375,
0.38427734375,
0.587890625,
0.38427734375,
0.587890625,
0.3984375,
0.51318359375,
0.3984375
],
[
0.1158447265625,
0.393798828125,
0.4375,
0.393798828125,
0.4375,
0.40673828125,
0.1158447265625,
0.40673828125
],
[
0.1368408203125,
0.4091796875,
0.484375,
0.4091796875,
0.484375,
0.4228515625,
0.1368408203125,
0.4228515625
],
[
0.53076171875,
0.405029296875,
0.880859375,
0.405029296875,
0.880859375,
0.418701171875,
0.53076171875,
0.418701171875
],
[
0.51123046875,
0.419189453125,
0.880859375,
0.418701171875,
0.880859375,
0.432861328125,
0.51123046875,
0.43408203125
],
[
0.12005615234375,
0.427001953125,
0.483642578125,
0.427001953125,
0.483642578125,
0.43701171875,
0.12005615234375,
0.43701171875
],
[
0.118408203125,
0.441162109375,
0.484375,
0.441162109375,
0.484375,
0.454833984375,
0.118408203125,
0.454833984375
],
[
0.51318359375,
0.435791015625,
0.880859375,
0.435791015625,
0.880859375,
0.449462890625,
0.51318359375,
0.449462890625
],
[
0.51123046875,
0.451904296875,
0.69775390625,
0.45068359375,
0.69775390625,
0.465576171875,
0.51123046875,
0.466064453125
],
[
0.118408203125,
0.457763671875,
0.484375,
0.457763671875,
0.484375,
0.471435546875,
0.118408203125,
0.471435546875
],
[
0.118408203125,
0.473876953125,
0.486083984375,
0.473876953125,
0.486083984375,
0.48681640625,
0.118408203125,
0.48681640625
],
[
0.53076171875,
0.47265625,
0.880859375,
0.47265625,
0.880859375,
0.486328125,
0.53076171875,
0.486328125
],
[
0.1158447265625,
0.488037109375,
0.35693359375,
0.4892578125,
0.35693359375,
0.50439453125,
0.1158447265625,
0.5029296875
],
[
0.51123046875,
0.48681640625,
0.8818359375,
0.4892578125,
0.8818359375,
0.50439453125,
0.51123046875,
0.501953125
],
[
0.51318359375,
0.505859375,
0.88330078125,
0.505859375,
0.88330078125,
0.51953125,
0.51318359375,
0.51953125
],
[
0.1158447265625,
0.51708984375,
0.405517578125,
0.51708984375,
0.405517578125,
0.53076171875,
0.1158447265625,
0.53076171875
],
[
0.51123046875,
0.521484375,
0.87890625,
0.521484375,
0.87890625,
0.53515625,
0.51123046875,
0.53515625
],
[
0.118408203125,
0.5361328125,
0.483642578125,
0.5361328125,
0.483642578125,
0.55029296875,
0.118408203125,
0.55029296875
],
[
0.5146484375,
0.53759765625,
0.8818359375,
0.53759765625,
0.8818359375,
0.55029296875,
0.5146484375,
0.55029296875
],
[
0.1175537109375,
0.55419921875,
0.468505859375,
0.552734375,
0.468505859375,
0.56787109375,
0.1175537109375,
0.568359375
],
[
0.51123046875,
0.5517578125,
0.87890625,
0.5517578125,
0.87890625,
0.5654296875,
0.51123046875,
0.5654296875
],
[
0.118408203125,
0.57080078125,
0.486083984375,
0.57080078125,
0.486083984375,
0.58447265625,
0.118408203125,
0.58447265625
],
[
0.51123046875,
0.568359375,
0.87890625,
0.568359375,
0.87890625,
0.58203125,
0.51123046875,
0.58203125
],
[
0.118408203125,
0.5859375,
0.484375,
0.5859375,
0.484375,
0.599609375,
0.118408203125,
0.599609375
],
[
0.509765625,
0.5830078125,
0.880859375,
0.5849609375,
0.880859375,
0.599609375,
0.509765625,
0.59716796875
],
[
0.118408203125,
0.6025390625,
0.486083984375,
0.6025390625,
0.486083984375,
0.6162109375,
0.118408203125,
0.6162109375
],
[
0.51123046875,
0.6015625,
0.8818359375,
0.6015625,
0.8818359375,
0.615234375,
0.51123046875,
0.615234375
],
[
0.118408203125,
0.6181640625,
0.486083984375,
0.6181640625,
0.486083984375,
0.6318359375,
0.118408203125,
0.6318359375
],
[
0.51123046875,
0.61767578125,
0.57080078125,
0.61767578125,
0.57080078125,
0.6318359375,
0.51123046875,
0.6318359375
],
[
0.118408203125,
0.63525390625,
0.273681640625,
0.63525390625,
0.273681640625,
0.6484375,
0.118408203125,
0.6484375
],
[
0.5322265625,
0.63818359375,
0.8818359375,
0.63818359375,
0.8818359375,
0.65185546875,
0.5322265625,
0.65185546875
],
[
0.1368408203125,
0.65185546875,
0.486083984375,
0.65185546875,
0.486083984375,
0.6650390625,
0.1368408203125,
0.6650390625
],
[
0.51123046875,
0.65185546875,
0.87890625,
0.65185546875,
0.87890625,
0.66650390625,
0.51123046875,
0.66650390625
],
[
0.118408203125,
0.66650390625,
0.484375,
0.66650390625,
0.484375,
0.6796875,
0.118408203125,
0.6796875
],
[
0.51318359375,
0.67041015625,
0.880859375,
0.67041015625,
0.880859375,
0.68408203125,
0.51318359375,
0.68408203125
],
[
0.12005615234375,
0.68505859375,
0.483642578125,
0.68505859375,
0.483642578125,
0.6953125,
0.12005615234375,
0.6953125
],
[
0.51318359375,
0.68603515625,
0.87890625,
0.68603515625,
0.87890625,
0.69970703125,
0.51318359375,
0.69970703125
],
[
0.118408203125,
0.69970703125,
0.484375,
0.69970703125,
0.484375,
0.71337890625,
0.118408203125,
0.71337890625
],
[
0.51318359375,
0.703125,
0.64208984375,
0.703125,
0.64208984375,
0.71630859375,
0.51318359375,
0.71630859375
],
[
0.118408203125,
0.71533203125,
0.486083984375,
0.71533203125,
0.486083984375,
0.728515625,
0.118408203125,
0.728515625
],
[
0.118408203125,
0.732421875,
0.484375,
0.732421875,
0.484375,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.51123046875,
0.74462890625,
0.54248046875,
0.74462890625,
0.54248046875,
0.755859375,
0.51123046875,
0.755859375
],
[
0.5380859375,
0.74658203125,
0.56103515625,
0.74658203125,
0.56103515625,
0.75439453125,
0.5380859375,
0.75439453125
],
[
0.55517578125,
0.74365234375,
0.7734375,
0.74365234375,
0.7734375,
0.75732421875,
0.55517578125,
0.75732421875
],
[
0.118408203125,
0.765625,
0.210693359375,
0.765625,
0.210693359375,
0.77685546875,
0.118408203125,
0.77685546875
],
[
0.118408203125,
0.74755859375,
0.48779296875,
0.74755859375,
0.48779296875,
0.76123046875,
0.118408203125,
0.76123046875
],
[
0.5146484375,
0.7724609375,
0.87890625,
0.7724609375,
0.87890625,
0.7861328125,
0.5146484375,
0.7861328125
],
[
0.1368408203125,
0.77978515625,
0.486083984375,
0.77978515625,
0.486083984375,
0.79345703125,
0.1368408203125,
0.79345703125
],
[
0.51318359375,
0.7880859375,
0.880859375,
0.7880859375,
0.880859375,
0.8017578125,
0.51318359375,
0.8017578125
],
[
0.1175537109375,
0.79638671875,
0.484375,
0.79638671875,
0.484375,
0.81005859375,
0.1175537109375,
0.81005859375
],
[
0.51123046875,
0.802734375,
0.87890625,
0.80419921875,
0.87890625,
0.818359375,
0.51123046875,
0.81787109375
],
[
0.118408203125,
0.8125,
0.486083984375,
0.8125,
0.486083984375,
0.826171875,
0.118408203125,
0.826171875
],
[
0.51318359375,
0.82080078125,
0.87890625,
0.82080078125,
0.87890625,
0.83447265625,
0.51318359375,
0.83447265625
],
[
0.118408203125,
0.82763671875,
0.484375,
0.82763671875,
0.484375,
0.84130859375,
0.118408203125,
0.84130859375
],
[
0.51123046875,
0.8359375,
0.87890625,
0.8359375,
0.87890625,
0.849609375,
0.51123046875,
0.849609375
],
[
0.118408203125,
0.84423828125,
0.486083984375,
0.84423828125,
0.486083984375,
0.85791015625,
0.118408203125,
0.85791015625
],
[
0.51318359375,
0.8525390625,
0.87890625,
0.8525390625,
0.87890625,
0.8662109375,
0.51318359375,
0.8662109375
],
[
0.1376953125,
0.8681640625,
0.486083984375,
0.87060546875,
0.486083984375,
0.88427734375,
0.1376953125,
0.8818359375
],
[
0.51318359375,
0.8681640625,
0.66748046875,
0.8681640625,
0.66748046875,
0.8818359375,
0.51318359375,
0.8818359375
],
[
0.1175537109375,
0.88427734375,
0.483642578125,
0.88427734375,
0.483642578125,
0.89453125,
0.1175537109375,
0.89453125
],
[
0.1175537109375,
0.89453125,
0.484375,
0.89453125,
0.484375,
0.90771484375,
0.1175537109375,
0.90771484375
],
[
0.53076171875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.53076171875,
0.90283203125
],
[
0.1175537109375,
0.90625,
0.30908203125,
0.90625,
0.30908203125,
0.91943359375,
0.1175537109375,
0.91943359375
],
[
0.51123046875,
0.90478515625,
0.83447265625,
0.90478515625,
0.83447265625,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"Metric",
"Partial Match Exact Match",
"Avg.AnyCombinationF1",
"中药物提及的m>=2跨度的集合。输出是",
"88.9",
"86.1",
"-组R={(ci,))，其中 C E P(D)是来自",
"83.4",
"79.6",
"P(D)的药物组合，所有可能的药物组合的集",
"Avg.PositiveCombinationFi",
"合，yiE{POS_COMB,OTHER_COMB)是组",
"Table 1: Agreement scores using our adaptation of F1",
"合标签",
"score to allow for partial-match.",
"2.3Evaluation Metric",
"我们考虑两个设置：“精确匹配”，一个考虑",
"获取包含已知具有协同作用的药物对的句子",
"识别精确药物组合的严格版本，以及“部分匹",
"但其摘要不包括我们的触发短语之一。所有数",
"配”，一个为正确识别的子集分配部分学分的",
"据收集查询均使用SPIKEExtractiveSearch工具",
"更宽松的版本",
"(Shlain et al., 2020; Taub-Tabib et al., 2020) 执行。",
"我们对这两个设置都使用标准精度，召",
"该过程在Figure2的顶部进行了说明",
"回率和F1指标。对于部分匹配的情况，我",
"们将给定组合的二进制0或1分数替换",
"2.5",
"",
"The Annotation Process",
"如果有多个与黄金关系的部分匹配，我们",
"为精细分数：shared_drugs/total_drugs",
"七名生物医学工程研究生参加了注释任务。学",
"取重叠最大的一个。我们将召回率计算为",
"生们都完成了癌症联合疗法的课程，并由一位",
"identified_relations/all_gold_relations",
"在这一领域具有专业知识的有原则的研究人员",
"精",
"度",
"计算",
"",
"为",
"",
"",
"进行监督",
"correct_relations/identified_relations",
"我们考虑两个指标，将POS_COMB与其余",
"我们为参与者提供了注释指南，其中指定了",
"注释过程应如何进行（请参阅附录A.1），并",
"指标进行比较的平均正组合F1得分，以及对任",
"何组合标签（POS或其他）而不是NO_COMB",
"进行了一次初次会议，我们与小组一起审查了",
"指南并讨论了一些示例",
"进行正确预测的平均任何组合F1得分。后者",
"是一项更容易的任务，但对于识别药物组合仍",
"每个参与者都可以访问Prodigy注释工具",
"然很有价值，无论其疗效如何",
"(MontaniandHonnibal，2018）的单独实例，该",
"实例预先加载了候选注释实例。会话开始后，",
"2.4Collecting Data for Annotation",
"实例（包含带有标记药物实体的句子及其上",
"为了收集注释数据，我们从DrugBank‘收集了",
"下文）以顺序方式出现，没有时间限制。对",
"2411种药物的列表，并从PubMed中抽取了",
"于每个实例，我们指示注释者标记参与组合",
"组提到2种或更多药物的句子。对该样本的前",
"的所有药物子集，并为每个子集指示其标签",
"50个句子的分析表明，只有8/50的句子中提",
"(POS_COMB或OTHER_COMB）。此外，我们",
"到了药物组合。这意味着注释完整的样本将是",
"指示他们指出是否需要上下文以确定关系的积",
"昂贵的，并且将导致数据集高度偏向相对琐碎",
"极功效",
"的NO_COMB实例",
"尽管专家注释需要大量时间，但我们还是收",
"因此，我们重复了此实验，对PubMed摘要",
"集了1634篇文章的注释。其中，272个被分配",
"中包含触发短语的句子进行了采样。??？个抽",
"给至少两个注释者。在首席研究员进一步仲裁",
"样句子中的48个%包括提及药物组合。针对",
"后，这些被用于测试集。该过程在Figure2的",
"具有已知药物组合的摘要的新样本评估触发列",
"底部进行了说明",
"表的覆盖范围表明，这些新摘要中的90个%",
"包含一个触发词。这表明我们的触发列表可用",
"2.6",
"",
"Inter-annotator Agreement",
"围和多样性",
"于获取标签平衡数据，而不会过度限制覆盖范",
"在任务过程中，我们多次计算了注释者之间",
"因此，我们收集了大多数注释实例90%，使",
"的协议，以识别分歧的情况并向注释者提供反",
"用基本搜索包含至少两种不同药物且其摘要",
"馈。每次随机选择一组25个实例，并将其分",
"包含触发短语之一的句子。为了克服触发列表",
"配给所有注释器。协议是基于成对F1度量计",
"所施加的词汇限制，我们使用远程监督对剩余",
"算的（如82.3中所述进行了一些修改），并对",
"的10个%实例进行采样：根据DrugComboDB",
"所有注释者对进行了平均（请参见附录A.2中",
"6Curation included downloading a premade drug list from",
"对替代度量的讨论）",
"DrugBank's website,while removing non pharmacological",
"intervention such as Vitamins and Supplements. The later we",
"Table1中的最终协议编号是令人满意的",
"got from the FDA orange book.",
"(Aroyo and Welty, 2013; Araki et al., 2018)"
] |
{
"arxiv_id": "2205.02289",
"language": "zh",
"page_idx": 3
}
|
||
2205.02289-zh-4
|
2205.02289-zh-4.jpg
|
Data Curation
<FigureHere>
Set
Figure 2: IIlustration of the data construction process. First we construct the required knowledge resources. Then, we collect data using SPIKE -an extractive search tool- over the PubMed database. The train and test sets were annotated using Prodigy over the curated data. For test data, we collected two annotations for each sample, and then had a domain expert resolve annotation disagreements.
<FigureHere>
"[...]of protein kinase C-alpha. This|study evaluated the response rate|of the combination therapy of
<<m'>> aprinocarsen <</m>>,<<m>> gemcitabine <</m>>,and <<m>> carboplatin <</m>> in previously untreated patients with advanced non-small cell lung cancer... [200 tokens later] ... However, this combination resulted in severe thrombocytopenia in the majority of patients."
Figure 3: Our baseline architecture, adapted from the PURE model (Zhong and Chen, 2021)
## 2.7 Resulting Dataset 数据集由 1634 个带注释的摘要组成，$^{7}$分为
释器提供的上下文要求的指示。在 1248 个关系中的 835 个中，注释器根据需要标记了上下文，这是当时的 67 个%, 显示了上下文在 DCE 任务中的重要性
### $\textbf{3}$ $\textbf{Experiments}$
1362 个训练和 272 个测试实例。其中包括 1248个关系；838 个是 POS\_COMB, 410 个是其他$\_$COMB$\left ( \text{在 火 车 和 测 试 集 中 具 有 相 同 的 标 签 比 }\right )$ 率)。591 个句子不包含药物组合，877个句子包${\text{含一个关系 (POS\_COMB 或 OTHER\_COMB),}}$ 166 个句子包含两个或多个不同的组合。在注释关系中，900 是二进制，226 是3进制，69是4进制，53 是5 进制或更多
对于结果数据集中的每个实例，我们包括注
3.1 Baseline Model Architecture
我们建立一个基线模型来衡量我们的数据集的难度，并揭示需要改进的地方。对于我们的基础基线模型架构，我们采用 Zhong and Chen (2021)的 PURE 架构，该架构在几个关系分类基准上是最先进的，包括 SciERC 二进制科学RE 数据集 (Luan et al., 2018) 。PURE 架构专为2元和 3 元关系提取而设计，由三个组件组成。首先，在候选关系中的所有实体周围插入特殊
${\overline{{^7}}}$This is a similar size to existing human-labeled biomedical relation extraction datasets, such as BioCreative V CDR (Li et al., 2016) , which has 1500 abstracts annotated, BioCreative VI (Krallinger et al., 2017) , which has 2432 abstracts, and DDI (Herrero-Zazo et al., 2013) ,which has 714 abstracts.
|
[
[
0.124267578125,
0.09027099609375,
0.2091064453125,
0.09027099609375,
0.2091064453125,
0.1015625,
0.124267578125,
0.1015625
],
[
0.124267578125,
0.1051025390625,
0.21826171875,
0.10986328125,
0.2166748046875,
0.1258544921875,
0.1234130859375,
0.12054443359375
],
[
0.295654296875,
0.1092529296875,
0.4306640625,
0.1092529296875,
0.4306640625,
0.122314453125,
0.295654296875,
0.122314453125
],
[
0.48779296875,
0.1092529296875,
0.66748046875,
0.1092529296875,
0.66748046875,
0.122314453125,
0.48779296875,
0.122314453125
],
[
0.130126953125,
0.1181640625,
0.1427001953125,
0.1181640625,
0.1427001953125,
0.122314453125,
0.130126953125,
0.122314453125
],
[
0.20068359375,
0.1258544921875,
0.267822265625,
0.1258544921875,
0.267822265625,
0.135986328125,
0.20068359375,
0.135986328125
],
[
0.297119140625,
0.1258544921875,
0.426513671875,
0.1258544921875,
0.426513671875,
0.1395263671875,
0.297119140625,
0.1395263671875
],
[
0.486083984375,
0.1258544921875,
0.61376953125,
0.1258544921875,
0.61376953125,
0.1395263671875,
0.486083984375,
0.1395263671875
],
[
0.1368408203125,
0.1353759765625,
0.187255859375,
0.1353759765625,
0.187255859375,
0.146728515625,
0.1368408203125,
0.146728515625
],
[
0.198974609375,
0.135986328125,
0.267822265625,
0.135986328125,
0.267822265625,
0.146728515625,
0.198974609375,
0.146728515625
],
[
0.75146484375,
0.1312255859375,
0.80029296875,
0.1312255859375,
0.80029296875,
0.142578125,
0.75146484375,
0.142578125
],
[
0.33154296875,
0.141357421875,
0.447509765625,
0.141357421875,
0.447509765625,
0.1519775390625,
0.33154296875,
0.1519775390625
],
[
0.5224609375,
0.141357421875,
0.69775390625,
0.141357421875,
0.69775390625,
0.1519775390625,
0.5224609375,
0.1519775390625
],
[
0.75732421875,
0.141357421875,
0.791015625,
0.141357421875,
0.791015625,
0.15380859375,
0.75732421875,
0.15380859375
],
[
0.497802734375,
0.1436767578125,
0.50537109375,
0.1436767578125,
0.50537109375,
0.149658203125,
0.497802734375,
0.149658203125
],
[
0.22412109375,
0.142578125,
0.2435302734375,
0.1478271484375,
0.2392578125,
0.1561279296875,
0.219970703125,
0.15087890625
],
[
0.6943359375,
0.1455078125,
0.7138671875,
0.1455078125,
0.7138671875,
0.1484375,
0.6943359375,
0.1484375
],
[
0.72900390625,
0.144287109375,
0.74462890625,
0.144287109375,
0.74462890625,
0.1519775390625,
0.72900390625,
0.1519775390625
],
[
0.300537109375,
0.1561279296875,
0.478515625,
0.155029296875,
0.478515625,
0.1650390625,
0.300537109375,
0.166259765625
],
[
0.76171875,
0.1519775390625,
0.78857421875,
0.1519775390625,
0.78857421875,
0.171630859375,
0.76171875,
0.171630859375
],
[
0.1292724609375,
0.160888671875,
0.197265625,
0.160888671875,
0.197265625,
0.171630859375,
0.1292724609375,
0.171630859375
],
[
0.204833984375,
0.1602783203125,
0.255126953125,
0.1602783203125,
0.255126953125,
0.171630859375,
0.204833984375,
0.171630859375
],
[
0.1234130859375,
0.168701171875,
0.197265625,
0.17041015625,
0.197265625,
0.1817626953125,
0.1234130859375,
0.179931640625
],
[
0.210693359375,
0.17041015625,
0.252685546875,
0.17041015625,
0.252685546875,
0.1817626953125,
0.210693359375,
0.1817626953125
],
[
0.307373046875,
0.171630859375,
0.318115234375,
0.174560546875,
0.313232421875,
0.179931640625,
0.3046875,
0.177001953125
],
[
0.330078125,
0.17041015625,
0.37451171875,
0.17041015625,
0.37451171875,
0.1817626953125,
0.330078125,
0.1817626953125
],
[
0.456787109375,
0.191162109375,
0.478515625,
0.191162109375,
0.478515625,
0.215576171875,
0.456787109375,
0.215576171875
],
[
0.124267578125,
0.1995849609375,
0.2216796875,
0.1995849609375,
0.2216796875,
0.21142578125,
0.124267578125,
0.21142578125
],
[
0.2392578125,
0.2244873046875,
0.26025390625,
0.2257080078125,
0.2568359375,
0.255859375,
0.2359619140625,
0.254638671875
],
[
0.3955078125,
0.2244873046875,
0.4609375,
0.2244873046875,
0.4609375,
0.235107421875,
0.3955078125,
0.235107421875
],
[
0.51708984375,
0.2244873046875,
0.5634765625,
0.2244873046875,
0.5634765625,
0.235107421875,
0.51708984375,
0.235107421875
],
[
0.470947265625,
0.2257080078125,
0.50537109375,
0.2257080078125,
0.50537109375,
0.2340087890625,
0.470947265625,
0.2340087890625
],
[
0.318115234375,
0.2257080078125,
0.35693359375,
0.2257080078125,
0.35693359375,
0.2327880859375,
0.318115234375,
0.2327880859375
],
[
0.60693359375,
0.2257080078125,
0.62646484375,
0.2257080078125,
0.62646484375,
0.2327880859375,
0.60693359375,
0.2327880859375
],
[
0.6455078125,
0.2257080078125,
0.66162109375,
0.2257080078125,
0.66162109375,
0.2327880859375,
0.6455078125,
0.2327880859375
],
[
0.5810546875,
0.226806640625,
0.58837890625,
0.226806640625,
0.58837890625,
0.232177734375,
0.5810546875,
0.232177734375
],
[
0.273681640625,
0.226806640625,
0.301513671875,
0.226806640625,
0.301513671875,
0.23095703125,
0.273681640625,
0.23095703125
],
[
0.36865234375,
0.22802734375,
0.38037109375,
0.22802734375,
0.38037109375,
0.232177734375,
0.36865234375,
0.232177734375
],
[
0.5322265625,
0.2327880859375,
0.54833984375,
0.2327880859375,
0.54833984375,
0.240478515625,
0.5322265625,
0.240478515625
],
[
0.7001953125,
0.23095703125,
0.779296875,
0.23095703125,
0.779296875,
0.2410888671875,
0.7001953125,
0.2410888671875
],
[
0.4306640625,
0.2369384765625,
0.45166015625,
0.2369384765625,
0.45166015625,
0.24462890625,
0.4306640625,
0.24462890625
],
[
0.81787109375,
0.2369384765625,
0.87890625,
0.2369384765625,
0.87890625,
0.251708984375,
0.81787109375,
0.251708984375
],
[
0.5751953125,
0.254150390625,
0.634765625,
0.254150390625,
0.634765625,
0.268310546875,
0.5751953125,
0.268310546875
],
[
0.8427734375,
0.248779296875,
0.86962890625,
0.248779296875,
0.86962890625,
0.2666015625,
0.8427734375,
0.2666015625
],
[
0.3720703125,
0.257080078125,
0.41064453125,
0.257080078125,
0.41064453125,
0.26416015625,
0.3720703125,
0.26416015625
],
[
0.42236328125,
0.257080078125,
0.459228515625,
0.257080078125,
0.459228515625,
0.26416015625,
0.42236328125,
0.26416015625
],
[
0.47021484375,
0.257080078125,
0.53076171875,
0.257080078125,
0.53076171875,
0.26416015625,
0.47021484375,
0.26416015625
],
[
0.53662109375,
0.255859375,
0.56494140625,
0.255859375,
0.56494140625,
0.26416015625,
0.53662109375,
0.26416015625
],
[
0.462646484375,
0.27783203125,
0.5576171875,
0.27783203125,
0.5576171875,
0.2880859375,
0.462646484375,
0.2880859375
],
[
0.313232421875,
0.25830078125,
0.355224609375,
0.25830078125,
0.355224609375,
0.262451171875,
0.313232421875,
0.262451171875
],
[
0.133544921875,
0.295654296875,
0.74658203125,
0.295654296875,
0.74658203125,
0.305908203125,
0.133544921875,
0.305908203125
],
[
0.75390625,
0.298583984375,
0.80615234375,
0.298583984375,
0.80615234375,
0.308837890625,
0.75390625,
0.308837890625
],
[
0.69189453125,
0.305908203125,
0.74072265625,
0.305908203125,
0.74072265625,
0.317626953125,
0.69189453125,
0.317626953125
],
[
0.8212890625,
0.305908203125,
0.873046875,
0.305908203125,
0.873046875,
0.317626953125,
0.8212890625,
0.317626953125
],
[
0.370361328125,
0.30712890625,
0.411376953125,
0.30712890625,
0.411376953125,
0.317626953125,
0.370361328125,
0.317626953125
],
[
0.2880859375,
0.30810546875,
0.32568359375,
0.30810546875,
0.32568359375,
0.317626953125,
0.2880859375,
0.317626953125
],
[
0.1351318359375,
0.31640625,
0.744140625,
0.31640625,
0.744140625,
0.330078125,
0.1351318359375,
0.330078125
],
[
0.84228515625,
0.318359375,
0.86572265625,
0.318359375,
0.86572265625,
0.333251953125,
0.84228515625,
0.333251953125
],
[
0.1158447265625,
0.349853515625,
0.880859375,
0.349853515625,
0.880859375,
0.363525390625,
0.1158447265625,
0.363525390625
],
[
0.1158447265625,
0.363525390625,
0.87890625,
0.363525390625,
0.87890625,
0.377197265625,
0.1158447265625,
0.377197265625
],
[
0.1175537109375,
0.37890625,
0.8818359375,
0.37890625,
0.8818359375,
0.392578125,
0.1175537109375,
0.392578125
],
[
0.11419677734375,
0.391357421875,
0.51904296875,
0.392578125,
0.51904296875,
0.40625,
0.11419677734375,
0.405029296875
],
[
0.57421875,
0.427001953125,
0.67333984375,
0.427001953125,
0.67333984375,
0.441162109375,
0.57421875,
0.441162109375
],
[
0.73779296875,
0.427001953125,
0.8623046875,
0.427001953125,
0.8623046875,
0.441162109375,
0.73779296875,
0.441162109375
],
[
0.60693359375,
0.474365234375,
0.61376953125,
0.474365234375,
0.61376953125,
0.48095703125,
0.60693359375,
0.48095703125
],
[
0.38623046875,
0.5087890625,
0.611328125,
0.51025390625,
0.611328125,
0.52490234375,
0.38623046875,
0.52392578125
],
[
0.1376953125,
0.5498046875,
0.85791015625,
0.5498046875,
0.85791015625,
0.5634765625,
0.1376953125,
0.5634765625
],
[
0.1368408203125,
0.5634765625,
0.8447265625,
0.5634765625,
0.8447265625,
0.57666015625,
0.1368408203125,
0.57666015625
],
[
0.1351318359375,
0.57666015625,
0.66748046875,
0.57666015625,
0.66748046875,
0.59033203125,
0.1351318359375,
0.59033203125
],
[
0.67529296875,
0.57470703125,
0.8369140625,
0.57666015625,
0.8369140625,
0.59130859375,
0.67529296875,
0.5888671875
],
[
0.1351318359375,
0.5888671875,
0.791015625,
0.5888671875,
0.791015625,
0.6025390625,
0.1351318359375,
0.6025390625
],
[
0.18896484375,
0.6162109375,
0.8095703125,
0.6162109375,
0.8095703125,
0.6298828125,
0.18896484375,
0.6298828125
],
[
0.15869140625,
0.65673828125,
0.3046875,
0.65673828125,
0.3046875,
0.67041015625,
0.15869140625,
0.67041015625
],
[
0.51123046875,
0.65625,
0.87744140625,
0.65625,
0.87744140625,
0.66943359375,
0.51123046875,
0.66943359375
],
[
0.5146484375,
0.6728515625,
0.87890625,
0.6728515625,
0.87890625,
0.68701171875,
0.5146484375,
0.68701171875
],
[
0.1175537109375,
0.65771484375,
0.162841796875,
0.65771484375,
0.162841796875,
0.66845703125,
0.1175537109375,
0.66845703125
],
[
0.1158447265625,
0.67578125,
0.48779296875,
0.67333984375,
0.48779296875,
0.69140625,
0.1158447265625,
0.693359375
],
[
0.1175537109375,
0.693359375,
0.486083984375,
0.693359375,
0.486083984375,
0.70703125,
0.1175537109375,
0.70703125
],
[
0.51318359375,
0.6884765625,
0.880859375,
0.6884765625,
0.880859375,
0.70166015625,
0.51318359375,
0.70166015625
],
[
0.1175537109375,
0.7080078125,
0.484375,
0.7080078125,
0.484375,
0.72265625,
0.1175537109375,
0.72265625
],
[
0.51318359375,
0.705078125,
0.64404296875,
0.705078125,
0.64404296875,
0.7197265625,
0.51318359375,
0.7197265625
],
[
0.118408203125,
0.724609375,
0.484375,
0.7255859375,
0.484375,
0.74072265625,
0.118408203125,
0.7392578125
],
[
0.51904296875,
0.74072265625,
0.5263671875,
0.7451171875,
0.51904296875,
0.75048828125,
0.51123046875,
0.7451171875
],
[
0.54736328125,
0.73828125,
0.66015625,
0.73828125,
0.66015625,
0.7529296875,
0.54736328125,
0.7529296875
],
[
0.118408203125,
0.75732421875,
0.486083984375,
0.75830078125,
0.486083984375,
0.7724609375,
0.118408203125,
0.771484375
],
[
0.1175537109375,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.51318359375,
0.7685546875,
0.54248046875,
0.7685546875,
0.54248046875,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.5556640625,
0.76708984375,
0.7841796875,
0.76708984375,
0.7841796875,
0.78076171875,
0.5556640625,
0.78076171875
],
[
0.5400390625,
0.7724609375,
0.54736328125,
0.7724609375,
0.54736328125,
0.775390625,
0.5400390625,
0.775390625
],
[
0.1158447265625,
0.7724609375,
0.486083984375,
0.77392578125,
0.486083984375,
0.7880859375,
0.1158447265625,
0.78759765625
],
[
0.1175537109375,
0.7890625,
0.486083984375,
0.7890625,
0.486083984375,
0.802734375,
0.1175537109375,
0.802734375
],
[
0.51123046875,
0.79052734375,
0.87890625,
0.79052734375,
0.87890625,
0.80419921875,
0.51123046875,
0.80419921875
],
[
0.118408203125,
0.8056640625,
0.361083984375,
0.8056640625,
0.361083984375,
0.8193359375,
0.118408203125,
0.8193359375
],
[
0.5146484375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.1368408203125,
0.822265625,
0.48779296875,
0.822265625,
0.48779296875,
0.8359375,
0.1368408203125,
0.8359375
],
[
0.51123046875,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.1368408203125,
0.84326171875,
0.489501953125,
0.845703125,
0.489501953125,
0.85986328125,
0.1368408203125,
0.85791015625
],
[
0.51318359375,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.1175537109375,
0.85791015625,
0.48779296875,
0.85791015625,
0.48779296875,
0.87158203125,
0.1175537109375,
0.87158203125
],
[
0.51123046875,
0.85546875,
0.87890625,
0.85546875,
0.87890625,
0.869140625,
0.51123046875,
0.869140625
],
[
0.1175537109375,
0.87158203125,
0.48779296875,
0.87158203125,
0.48779296875,
0.8818359375,
0.1175537109375,
0.8818359375
],
[
0.51123046875,
0.87158203125,
0.880859375,
0.87158203125,
0.880859375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.1175537109375,
0.8828125,
0.484375,
0.8828125,
0.484375,
0.8935546875,
0.1175537109375,
0.8935546875
],
[
0.1158447265625,
0.8935546875,
0.489501953125,
0.8935546875,
0.489501953125,
0.90673828125,
0.1158447265625,
0.90673828125
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1158447265625,
0.90673828125,
0.162841796875,
0.90673828125,
0.162841796875,
0.91748046875,
0.1158447265625,
0.91748046875
],
[
0.515625,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.515625,
0.91845703125
]
] |
[
"Data Curation",
"Resources",
"*SPIKE:MainQuery",
"*SPIKE:Distant Supervision",
"",
"Combination",
"the sentence matches:",
"the sentence matches:",
"Drug List",
"Trigger Word",
"Curated",
"drug1:(drugs) drug2:(drugs)",
"drug1:(drug_combos.1) drug2:(drug_comt",
"data",
"三",
"List",
"HnE",
"",
"and the abstract of the document matches",
"国",
"KnownDrug",
"Pubmed",
"Combinations",
"Corpus",
"",
"(triggers)",
"I>",
"Data Annotation",
"bipoi",
"enzalutamide",
"docetaxel",
"versus",
"activity",
"men",
"with",
"in",
"clinica",
"0",
"GRUC",
"AnnotateOnce",
"POS1",
"Train Set",
"abiraterone",
"国",
"prostate",
"cancer",
"progressing",
"after",
"LAnnotateTwice",
"-resistant",
"AnnotatorA:Clinical activityofenzalutamide versus docetaxel in men with castration-resistant prostate cancer progressing after abiraterone",
"Arbitrate",
"PI/P2",
"TestSet",
"P2",
"PI",
"AnnotatorB:Clinical activityofenzalutamideversusdocetaxelinmenwithcastration-resistantprostatecancerprogressingaferabiraterone",
"国",
"Figure 2: Ilustration of the data construction process. First we construct the required knowledge resources. Then,",
"we collect data using SPIKE -an extractive search tool- over the PubMed database. The train and test sets were",
"annotated using Prodigy over the curated data. For test data, we collected two annotations for each sample, and",
" then had a domain expert resolve annotation disagreements.",
"Feedforward",
"OTHER_COMB",
"广",
"Contextualized Embedding",
"\"[...]of protein kinase C-alpha. This study evaluated the response rate of the combination therapy of",
"<<m>> aprinocarsen<</m>>,<<m>> gemcitabine<</m>>,and<<m>> carboplatin <</m>>in",
"previously untreated patients with advanced non-small cell lung cancer...",
"[200 tokens later] ...",
"However, this combination resulted in severe thrombocytopenia in the majority of patients.\"",
"Figure 3: Our baseline architecture, adapted from the PURE model (Zhong and Chen, 2021)",
"Resulting Dataset",
"释器提供的上下文要求的指示。在1248个关",
"系中的835个中，注释器根据需要标记了上下",
"2.7",
"数据集由1634个带注释的摘要组成，7分为",
"1362个训练和272个测试实例。其中包括1248",
"文，这是当时的67个%，显示了上下文在DCE",
"个关系；838个是POS_COMB，410个是其他",
"任务中的重要性",
"COMB（在火车和测试集中具有相同的标签比",
"",
"Experiments",
"含一个关系（POS_COMB或OTHER_COMB），",
"率）。591个句子不包含药物组合，877个句子包",
"3.1",
"Baseline Model Architecture",
"",
"166个句子包含两个或多个不同的组合。在注",
"释关系中，900是二进制，226是3进制，69",
"我们建立一个基线模型来衡量我们的数据集",
"是4进制，53是5进制或更多",
"的难度，并揭示需要改进的地方。对于我们的",
"对于结果数据集中的每个实例，我们包括注",
"基础基线模型架构，我们采用ZhongandChen",
"7This is a similar size to existing human-labeled biomed-",
"(2021)的PURE架构，该架构在几个关系分类",
"ical relation extraction datasets, such as BioCreative V CDR",
"基准上是最先进的，包括SciERC二进制科学",
"(Lietal.,2016),whichhas1500abstractsannotated,BioCre-",
"RE数据集（Luanetal,2018）。PURE架构专为",
"ative VI (Krallinger et al., 2017),which has 2432 abstracts,",
"and DDI (Herrero-Zazo et al., 2013) , which has 714 ab-",
"2元和3元关系提取而设计，由三个组件组成。",
"stracts.",
"首先，在候选关系中的所有实体周围插入特殊"
] |
{
"arxiv_id": "2205.02289",
"language": "zh",
"page_idx": 4
}
|
||
1707.06456-zh-0
|
1707.06456-zh-0.jpg
|
重新访问引用解析的选择性首选项
Michael Strube
Heidelberg Institute for
Theoretical Studies michael.strube@h-its.org
Nafise Sadat Moosavi$^*$ Heidelberg Institute for
Theoretical Studies nafise.moosavi@h-its.org
$\mathrm{Benjamin~Heinzerling}^{*}$
AIPHES
Heidelberg Institute for Theoretical Studies
benjamin.heinzerling@h-its.org
# Abstract
Selectional preferences have long been claimed to be essential for coreference resolution.
However, they are
mainly modeled only implicitly by cur- rent coreference resolvers. We propose a dependency-based embedding model of selectional preferences which allows fne-grained compatibility judgments with high coverage. We show that the incorporation of our model improves coreference resolution performance on the CoNLL dataset, matching the state-of-the-art results of a more complex system. However, it comes with a cost that makes it debatable how worthwhile such improvements are.
### 1 Introduction
长期以来，选择性首选项一直被认为对协指解析很有用。在他关于“解析代词引用”的开创性工作中，Hobbs (1978)提出了一种语义方法，该方法需要对“要求”进行推理。“例如， 选择偏好允许在文本中解决代词 it”泰坦尼克号撞上冰山。它沉得很快。“在这里，谓词下沉比其他主题“更喜欢”某些主题论点：船下沉似乎是合理的，但冰山却令人难以置信
自动获取选择首选项的工作已显示出相当大的进展 (Dagan and Itai, 1990; Resnik, 1993; Agirre and Martinez, 2001; Pantel et al., 2007; Erk, 2007; Ritter et al., 2010; Van de Cruys, 2014) 。但是，今天的共引用解析器 (Martschat and Strube, 2015; Wiseman et al., 2016; Clark and Manning, 2016a, i.a.) 充其量只能隐式地捕获选择首选项，例如。g。,通过给定提及的依赖项调控器和其他上下文功能
${^*\text{These au}}$thors contributed equally to this work.
由于通常不会报告负面结果，因此文献中没有明确的证据表明特定知识来源的非实用性。因此，在最近的文献中没有对选择偏好进行显式建模，这表明合并此知识源对于共同引用解析并不是很成功。
十多年前，Kehler et al. (2004) 宣布了“代词解析的谓词参数结构的非实用性”,并观察到对小数据集的微小改进是由于偶然性而不是选择偏好捕获了有意义的世界知识关系
Kehler et al. (2004)的声明基于根据当前标准从少量 2.8 m 谓词-参数对中提取的选择首选项。此外，他们采用简单的(线性)最大熵分类器，这需要手动定义特征组合，并且不太可能完全捕获选择偏好和其他共引用特征之间的复杂交互。因此，值得重新审视更好的选择偏好模型如何影响更复杂的共引用解析器的性能
在这项工作中，我们提出了一种选择偏好的细粒度，高覆盖率模型，并研究了其对最先进的非线性共参考解析器的影响。我们证明，结合我们的选择偏好模型可以提高性能。然而， 这种花费大量额外时间或资源的小改进是否有利是有争议的
2 Modeling Selectional Preferences
建模选择首选项时的主要设计选择是选择关系库存，即。e。可以是关系参数的概念和实体， 以及它们之间的语义关系
先前的工作已经研究了许多关系清单。单词-单词对(吃，食物)$^1$的谓词参数统计信息很容易获得，但不能推广到看不见的对(Dagan and Itai, 1990) 。基于类的方法通过词类对(eat,/nutrition/food$)\pmod($Resnik,1993)或类类对(/ingest, /nutrition/food) (Agirre and Martinez, 2001) 进行概括，但需要消除词类的歧义，并且受到提供此类类的词汇资源的覆盖范围的限制 (例如。g。WordNet)
其他可能的关系清单包括语义表示，例如
$^{1}$Examples due to Agirre and Martinez (2001)
|
[
[
0.322509765625,
0.08489990234375,
0.68115234375,
0.08612060546875,
0.68115234375,
0.1009521484375,
0.322509765625,
0.09979248046875
],
[
0.160400390625,
0.1300048828125,
0.355224609375,
0.1312255859375,
0.355224609375,
0.1484375,
0.160400390625,
0.1478271484375
],
[
0.40478515625,
0.1312255859375,
0.60205078125,
0.1318359375,
0.60205078125,
0.146728515625,
0.40478515625,
0.1455078125
],
[
0.68408203125,
0.1312255859375,
0.81787109375,
0.1318359375,
0.81787109375,
0.146728515625,
0.68408203125,
0.1455078125
],
[
0.397216796875,
0.1478271484375,
0.603515625,
0.1484375,
0.603515625,
0.163330078125,
0.397216796875,
0.162109375
],
[
0.6484375,
0.1478271484375,
0.85400390625,
0.1484375,
0.85400390625,
0.163330078125,
0.6484375,
0.162109375
],
[
0.1502685546875,
0.1644287109375,
0.355224609375,
0.1644287109375,
0.355224609375,
0.1787109375,
0.1502685546875,
0.1787109375
],
[
0.2166748046875,
0.1484375,
0.2880859375,
0.1484375,
0.2880859375,
0.1602783203125,
0.2166748046875,
0.1602783203125
],
[
0.41650390625,
0.1644287109375,
0.5859375,
0.1650390625,
0.5859375,
0.179931640625,
0.41650390625,
0.1787109375
],
[
0.6689453125,
0.1650390625,
0.8369140625,
0.1650390625,
0.8369140625,
0.1787109375,
0.6689453125,
0.1787109375
],
[
0.17041015625,
0.1817626953125,
0.339111328125,
0.1817626953125,
0.339111328125,
0.1953125,
0.17041015625,
0.1953125
],
[
0.414794921875,
0.1817626953125,
0.58447265625,
0.182861328125,
0.58447265625,
0.19775390625,
0.414794921875,
0.196533203125
],
[
0.66650390625,
0.1817626953125,
0.8369140625,
0.182861328125,
0.8369140625,
0.19775390625,
0.66650390625,
0.196533203125
],
[
0.145263671875,
0.198974609375,
0.359375,
0.1995849609375,
0.359375,
0.21435546875,
0.145263671875,
0.213134765625
],
[
0.264404296875,
0.26416015625,
0.341796875,
0.26416015625,
0.341796875,
0.279052734375,
0.264404296875,
0.279052734375
],
[
0.53564453125,
0.2666015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.27978515625,
0.53564453125,
0.27978515625
],
[
0.51708984375,
0.283203125,
0.880859375,
0.283203125,
0.880859375,
0.29638671875,
0.51708984375,
0.29638671875
],
[
0.145263671875,
0.29638671875,
0.45751953125,
0.297607421875,
0.45751953125,
0.312255859375,
0.145263671875,
0.311279296875
],
[
0.51708984375,
0.298583984375,
0.8818359375,
0.298583984375,
0.8818359375,
0.312255859375,
0.51708984375,
0.312255859375
],
[
0.1485595703125,
0.314208984375,
0.45751953125,
0.314208984375,
0.45751953125,
0.327880859375,
0.1485595703125,
0.327880859375
],
[
0.5146484375,
0.3134765625,
0.8818359375,
0.3134765625,
0.8818359375,
0.32666015625,
0.5146484375,
0.32666015625
],
[
0.14697265625,
0.330078125,
0.459228515625,
0.330078125,
0.459228515625,
0.344482421875,
0.14697265625,
0.344482421875
],
[
0.5146484375,
0.330810546875,
0.6513671875,
0.330810546875,
0.6513671875,
0.344482421875,
0.5146484375,
0.344482421875
],
[
0.14697265625,
0.3466796875,
0.45751953125,
0.3466796875,
0.45751953125,
0.35986328125,
0.14697265625,
0.35986328125
],
[
0.5322265625,
0.344482421875,
0.884765625,
0.345703125,
0.884765625,
0.363525390625,
0.5322265625,
0.3623046875
],
[
0.14697265625,
0.3623046875,
0.459228515625,
0.3623046875,
0.459228515625,
0.3759765625,
0.14697265625,
0.3759765625
],
[
0.5146484375,
0.3623046875,
0.8818359375,
0.363525390625,
0.8818359375,
0.377685546875,
0.5146484375,
0.377197265625
],
[
0.14697265625,
0.377685546875,
0.459228515625,
0.377685546875,
0.459228515625,
0.392578125,
0.14697265625,
0.392578125
],
[
0.515625,
0.380126953125,
0.87890625,
0.380126953125,
0.87890625,
0.39013671875,
0.515625,
0.39013671875
],
[
0.1485595703125,
0.394287109375,
0.459228515625,
0.394287109375,
0.459228515625,
0.407958984375,
0.1485595703125,
0.407958984375
],
[
0.515625,
0.396728515625,
0.8271484375,
0.396728515625,
0.8271484375,
0.410400390625,
0.515625,
0.410400390625
],
[
0.145263671875,
0.4091796875,
0.45751953125,
0.410400390625,
0.45751953125,
0.424560546875,
0.145263671875,
0.42333984375
],
[
0.5322265625,
0.412109375,
0.88330078125,
0.412109375,
0.88330078125,
0.42578125,
0.5322265625,
0.42578125
],
[
0.14697265625,
0.427001953125,
0.456787109375,
0.427001953125,
0.456787109375,
0.440673828125,
0.14697265625,
0.440673828125
],
[
0.515625,
0.427001953125,
0.880859375,
0.427490234375,
0.880859375,
0.4423828125,
0.515625,
0.441162109375
],
[
0.14697265625,
0.443603515625,
0.459228515625,
0.443603515625,
0.459228515625,
0.457275390625,
0.14697265625,
0.457275390625
],
[
0.515625,
0.44482421875,
0.8818359375,
0.44482421875,
0.8818359375,
0.457763671875,
0.515625,
0.457763671875
],
[
0.14697265625,
0.458984375,
0.45751953125,
0.458984375,
0.45751953125,
0.47265625,
0.14697265625,
0.47265625
],
[
0.515625,
0.46142578125,
0.88330078125,
0.46142578125,
0.88330078125,
0.474365234375,
0.515625,
0.474365234375
],
[
0.14697265625,
0.474365234375,
0.45751953125,
0.473876953125,
0.45751953125,
0.488037109375,
0.14697265625,
0.4892578125
],
[
0.51708984375,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.490478515625,
0.51708984375,
0.490478515625
],
[
0.14697265625,
0.4912109375,
0.459228515625,
0.4912109375,
0.459228515625,
0.5048828125,
0.14697265625,
0.5048828125
],
[
0.515625,
0.49462890625,
0.87890625,
0.49462890625,
0.87890625,
0.5048828125,
0.515625,
0.5048828125
],
[
0.1485595703125,
0.50830078125,
0.361083984375,
0.50830078125,
0.361083984375,
0.521484375,
0.1485595703125,
0.521484375
],
[
0.3525390625,
0.50830078125,
0.45751953125,
0.50830078125,
0.45751953125,
0.51953125,
0.3525390625,
0.51953125
],
[
0.515625,
0.5087890625,
0.88330078125,
0.5087890625,
0.88330078125,
0.5224609375,
0.515625,
0.5224609375
],
[
0.1485595703125,
0.52392578125,
0.456787109375,
0.52392578125,
0.456787109375,
0.53759765625,
0.1485595703125,
0.53759765625
],
[
0.515625,
0.525390625,
0.53662109375,
0.525390625,
0.53662109375,
0.5380859375,
0.515625,
0.5380859375
],
[
0.14697265625,
0.5390625,
0.45751953125,
0.5390625,
0.45751953125,
0.552734375,
0.14697265625,
0.552734375
],
[
0.5341796875,
0.54150390625,
0.8818359375,
0.54150390625,
0.8818359375,
0.5546875,
0.5341796875,
0.5546875
],
[
0.1485595703125,
0.55712890625,
0.231689453125,
0.55712890625,
0.231689453125,
0.568359375,
0.1485595703125,
0.568359375
],
[
0.515625,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.515625,
0.57177734375
],
[
0.515625,
0.57373046875,
0.8818359375,
0.57373046875,
0.8818359375,
0.58740234375,
0.515625,
0.58740234375
],
[
0.1175537109375,
0.58740234375,
0.156982421875,
0.58740234375,
0.156982421875,
0.5986328125,
0.1175537109375,
0.5986328125
],
[
0.151123046875,
0.58837890625,
0.26611328125,
0.58837890625,
0.26611328125,
0.60107421875,
0.151123046875,
0.60107421875
],
[
0.515625,
0.58837890625,
0.8818359375,
0.5888671875,
0.8818359375,
0.60400390625,
0.515625,
0.6025390625
],
[
0.5146484375,
0.60546875,
0.8818359375,
0.60546875,
0.8818359375,
0.619140625,
0.5146484375,
0.619140625
],
[
0.12176513671875,
0.6123046875,
0.486083984375,
0.6123046875,
0.486083984375,
0.6259765625,
0.12176513671875,
0.6259765625
],
[
0.51708984375,
0.62353515625,
0.62451171875,
0.62353515625,
0.62451171875,
0.63525390625,
0.51708984375,
0.63525390625
],
[
0.12176513671875,
0.6298828125,
0.48779296875,
0.6298828125,
0.48779296875,
0.6435546875,
0.12176513671875,
0.6435546875
],
[
0.12005615234375,
0.64453125,
0.486083984375,
0.64453125,
0.486083984375,
0.65771484375,
0.12005615234375,
0.65771484375
],
[
0.515625,
0.65380859375,
0.54833984375,
0.65380859375,
0.54833984375,
0.662109375,
0.515625,
0.662109375
],
[
0.5498046875,
0.65283203125,
0.8330078125,
0.65283203125,
0.8330078125,
0.6630859375,
0.5498046875,
0.6630859375
],
[
0.12005615234375,
0.6611328125,
0.48779296875,
0.6611328125,
0.48779296875,
0.6748046875,
0.12005615234375,
0.6748046875
],
[
0.12005615234375,
0.677734375,
0.489501953125,
0.677734375,
0.489501953125,
0.69140625,
0.12005615234375,
0.69140625
],
[
0.515625,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.515625,
0.6904296875
],
[
0.12176513671875,
0.693359375,
0.48779296875,
0.693359375,
0.48779296875,
0.70703125,
0.12176513671875,
0.70703125
],
[
0.515625,
0.693359375,
0.884765625,
0.693359375,
0.884765625,
0.70703125,
0.515625,
0.70703125
],
[
0.12005615234375,
0.71044921875,
0.484375,
0.71044921875,
0.484375,
0.72021484375,
0.12005615234375,
0.72021484375
],
[
0.5146484375,
0.7080078125,
0.71533203125,
0.7080078125,
0.71533203125,
0.7216796875,
0.5146484375,
0.7216796875
],
[
0.12005615234375,
0.72705078125,
0.450927734375,
0.72705078125,
0.450927734375,
0.74072265625,
0.12005615234375,
0.74072265625
],
[
0.5341796875,
0.7255859375,
0.8818359375,
0.7255859375,
0.8818359375,
0.7392578125,
0.5341796875,
0.7392578125
],
[
0.14111328125,
0.74365234375,
0.489501953125,
0.74365234375,
0.489501953125,
0.75732421875,
0.14111328125,
0.75732421875
],
[
0.5146484375,
0.74072265625,
0.8818359375,
0.7412109375,
0.8818359375,
0.755859375,
0.5146484375,
0.7548828125
],
[
0.118408203125,
0.75830078125,
0.48779296875,
0.75830078125,
0.48779296875,
0.771484375,
0.118408203125,
0.771484375
],
[
0.515625,
0.75830078125,
0.880859375,
0.75830078125,
0.880859375,
0.771484375,
0.515625,
0.771484375
],
[
0.12176513671875,
0.77490234375,
0.489501953125,
0.77490234375,
0.489501953125,
0.7880859375,
0.12176513671875,
0.7880859375
],
[
0.5146484375,
0.77490234375,
0.8818359375,
0.77490234375,
0.8818359375,
0.7880859375,
0.5146484375,
0.7880859375
],
[
0.12005615234375,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.12005615234375,
0.8046875
],
[
0.5146484375,
0.79052734375,
0.88330078125,
0.79052734375,
0.88330078125,
0.80419921875,
0.5146484375,
0.80419921875
],
[
0.12005615234375,
0.80712890625,
0.489501953125,
0.80712890625,
0.489501953125,
0.82080078125,
0.12005615234375,
0.82080078125
],
[
0.5146484375,
0.80712890625,
0.88330078125,
0.80712890625,
0.88330078125,
0.82080078125,
0.5146484375,
0.82080078125
],
[
0.12005615234375,
0.822265625,
0.489501953125,
0.822265625,
0.489501953125,
0.8359375,
0.12005615234375,
0.8359375
],
[
0.5146484375,
0.822265625,
0.8818359375,
0.822265625,
0.8818359375,
0.8359375,
0.5146484375,
0.8359375
],
[
0.118408203125,
0.8388671875,
0.48779296875,
0.8388671875,
0.48779296875,
0.8525390625,
0.118408203125,
0.8525390625
],
[
0.515625,
0.83837890625,
0.88330078125,
0.83837890625,
0.88330078125,
0.8515625,
0.515625,
0.8515625
],
[
0.12005615234375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.87060546875,
0.12005615234375,
0.87060546875
],
[
0.515625,
0.85498046875,
0.7958984375,
0.85498046875,
0.7958984375,
0.8681640625,
0.515625,
0.8681640625
],
[
0.12005615234375,
0.87255859375,
0.376220703125,
0.87255859375,
0.376220703125,
0.8857421875,
0.12005615234375,
0.8857421875
],
[
0.5341796875,
0.87158203125,
0.8818359375,
0.87158203125,
0.8818359375,
0.88525390625,
0.5341796875,
0.88525390625
],
[
0.14697265625,
0.896484375,
0.480224609375,
0.896484375,
0.480224609375,
0.91015625,
0.14697265625,
0.91015625
],
[
0.5322265625,
0.89453125,
0.8564453125,
0.896484375,
0.8564453125,
0.9111328125,
0.5322265625,
0.9091796875
]
] |
[
"重新访问引用解析的选择性首选项",
"Benjamin Heinzerling",
"Nafise Sadat Moosavi*",
"Michael Strube",
"Heidelberg Institute for",
"Heidelberg Institute for",
"Heidelberg Institute for",
"AIPHES",
"Theoretical Studies",
"Theoretical Studies",
"Theoretical Studies",
"nafise.moosavi@h-its.org",
"michael.strube@h-its.org",
"benjamin.heinzerling@h-its.org",
"Abstract",
"由于通常不会报告负面结果，因此文献中没",
"有明确的证据表明特定知识来源的非实用性。",
"Selectional preferences have long been",
"因此，在最近的文献中没有对选择偏好进行显",
"claimed to be essential for corefer-",
"式建模，这表明合并此知识源对于共同引用解",
"ence resolution.However, they are",
"析并不是很成功。",
"mainly modeled only implicitly by cur-",
"十多年前，Kehler etal.（2004）宣布了“代",
"rent coreference resolvers.We pro-",
"词解析的谓词参数结构的非实用性”，并观察",
"pose a dependency-based embedding",
"到对小数据集的微小改进是由于偶然性而不是",
"model of selectional preferences which",
"选择偏好捕获了有意义的世界知识关系",
"allows fine-grained compatibility judg-",
"Kehleretal.（2004）的声明基于根据当前标",
"ments with high coverage. We show",
"准从少量2.8m谓词-参数对中提取的选择首",
"that the incorporation of our model",
"选项。此外，他们采用简单的（线性）最大熵",
"improves coreference resolution perfor-",
"分类器，这需要手动定义特征组合，并且不太",
"mance on the CoNLL dataset, match-",
"可能完全捕获选择偏好和其他共引用特征之间",
"ing the state-of-the-art results of a",
"的复杂交互。因此，值得重新审视更好的选择",
"more complex system.",
"However,it",
"偏好模型如何影响更复杂的共引用解析器的性",
"comes with a cost that makes it de-",
"能",
"batable how worthwhile such improve-",
"在这项工作中，我们提出了一种选择偏好的",
"ments are.",
"细粒度，高覆盖率模型，并研究了其对最先进",
"的非线性共参考解析器的影响。我们证明，结",
"1",
"Introduction",
"合我们的选择偏好模型可以提高性能。然而，",
"这种花费大量额外时间或资源的小改进是否有",
"长期以来，选择性首选项一直被认为对协指",
"利是有争议的",
"解析很有用。在他关于“解析代词引用”的开",
"创性工作中，Hobbs（1978)提出了一种语义方",
"2",
"ModelingSelectionalPreferences",
"法，该方法需要对“要求”进行推理。。“例如，",
"选择偏好允许在文本中解决代词it”泰坦尼克",
"建模选择首选项时的主要设计选择是选择关系",
"号撞上冰山。它沉得很快。“在这里，谓词下",
"库存，即。e。可以是关系参数的概念和实体，",
"比其他主题“更喜欢”某此主题论点：船下",
"以及它们之间的语义关系",
"沉似乎是合理的，但冰山却令人难以置信",
"先前的工作已经研究了许多关系清单。单",
"自动获取选择首选项的工作已显示出相当大",
"词-单词对（吃，食物）1的谓词参数统计信息",
"的进展(Daganand Itai,1990;Resnik，1993;",
"很容易获得，但不能推广到看不见的对（Da-",
"Agirre and Martinez, 2001; Pantel et al., 2007;",
"ganandItai，1990）。基于类的方法通过词类",
"Erk, 2007; Ritter et al., 2010; Van de Cruys,",
"对（eat，/nutrition/food）（Resnik，1993）或类",
"2014）。但是，今天的共引用解析器（Martschat",
"类对（/ingest，/nutrition/food）（Agirre and",
"and Strube, 2015; Wiseman et al., 2016; Clark",
"Martinez，2001）进行概括，但需要消除词类的",
"andManning，2016a，i.a.）充其量只能隐式地",
"歧义，并且受到提供此类类的词汇资源的覆盖",
"捕获选择首选项，例如。g。，通过给定提及的",
"范围的限制（例如。g。WordNet)",
"依赖项调控器和其他上下文功能",
"其他可能的关系清单包括语义表示，例如",
"* These authors contributed equally to this work.",
"'Examples due to Agirre and Martinez (2001)"
] |
{
"arxiv_id": "1707.06456",
"language": "zh",
"page_idx": 0
}
|
||
1707.06456-zh-3
|
1707.06456-zh-3.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th colspan="3">$MUC$</th>
<th> </th>
<th>$B^{3}$</th>
<th> </th>
<th colspan="3">CEAF</th>
<th rowspan="2">CoNLL Avg. F 1</th>
<th colspan="3">LEA</th>
</tr>
<tr>
<th> </th>
<th>$R$</th>
<th>$P$</th>
<th>$F$ 1</th>
<th>$R$</th>
<th>$P$</th>
<th>$F$ 1</th>
<th>$R$</th>
<th>$P$</th>
<th>$F$ 1</th>
<th>$R$</th>
<th>$P$</th>
<th>$F$ 1</th>
</tr>
<tr>
<td>baseline</td>
<td>70.09</td>
<td>80.01</td>
<td>74.72</td>
<td>57.64</td>
<td>70.09</td>
<td>63.26</td>
<td>54.47</td>
<td>63.92</td>
<td>58.82</td>
<td>65.60</td>
<td>54.02</td>
<td>66.45</td>
<td>59.59</td>
</tr>
<tr>
<td>gOV</td>
<td>70.10</td>
<td>79.96</td>
<td>74.71</td>
<td>57.51</td>
<td>70.31</td>
<td>63.27</td>
<td>54.41</td>
<td>64.08</td>
<td>58.85</td>
<td>65.61</td>
<td>53.93</td>
<td>66.76</td>
<td>59.66</td>
</tr>
<tr>
<td>+SP</td>
<td>70.85</td>
<td>79.31</td>
<td>74.85</td>
<td>58.93</td>
<td>69.16</td>
<td>63.64</td>
<td>55.25</td>
<td>63.78</td>
<td>59.21</td>
<td>65.90</td>
<td>55.29</td>
<td>65.53</td>
<td>59.98</td>
</tr>
<tr>
<td>$\operatorname{Reinforce}$</td>
<td>70.98</td>
<td>78.81</td>
<td>74.69</td>
<td>58.97</td>
<td>69.05</td>
<td>63.61</td>
<td>55.66</td>
<td>63.28</td>
<td>59.23</td>
<td>65.84</td>
<td>55.31</td>
<td>65.32</td>
<td>59.90</td>
</tr>
</tbody>
</table>
Table 1: Results on the CoNLL 2012 test set.
<table>
<tbody>
<tr>
<th> </th>
<th colspan="5">${\mathrm{development}}$</th>
</tr>
<tr>
<th>baseline</th>
<th>74.10</th>
<th>63.95</th>
<th>59.73</th>
<th>65.93</th>
<th>60.16</th>
</tr>
<tr>
<td>+embedding</td>
<td>74.38</td>
<td>64.42</td>
<td>60.45</td>
<td>66.42</td>
<td>60.65</td>
</tr>
<tr>
<td>+binned sim.</td>
<td>74.36</td>
<td>64.54</td>
<td>60.21</td>
<td>66.37</td>
<td>60.77</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>test</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>baseline</td>
<td>74.72</td>
<td>63.26</td>
<td>58.82</td>
<td>$\overline{65.60}$</td>
<td>59.55</td>
</tr>
<tr>
<td>+embedding</td>
<td>74.53</td>
<td>63.41</td>
<td>59.03</td>
<td>65.66</td>
<td>59.63</td>
</tr>
<tr>
<td>+binned $\sin.$</td>
<td>74.85</td>
<td>63.64</td>
<td>59.21</td>
<td>65.90</td>
<td>59.98</td>
</tr>
</tbody>
</table>
$\textcircled{\mathrm{adobj,}}$显示 @nsubj)
为了估计预期性能提升的更高界限，我们仅在照应性提及上运行基线和+SP 模型。根据CoNLL 分数和 LEA, 通过使用照应式提及， 性能提高了 1%。此结果表明，选择偏好的结合为非照应性提及创建了许多链接，从而降低了准确性。因此，当使用系统提及时，整体性能没有显著提高。deep-coref 在解析时间合并了隐喻性分数。进一步改进+SP 的结果的一种可能的方法是在输入水平上合并隐喻性分数。通过这种方式，共同引用解析器可以学习使用选择偏好主要用于更可能是照应的提及。但是，考虑到当前的反常性确定器或单例检测器的 F 1 得分仅在 85 %(Moosavi and Strube. 2016a,2017)左右，使用系统反常性得分的影响可能很小
Table 2: Incorporating the selectional preference model as new embeddings (+embedding) vs. as new pairwise features (+binned sim.).
## 5 Conclusions
我们引人了一种新的选择首选项模型，该模型结合了基于依赖关系的单词嵌入和细粒度的实体类型。为了有效，选择偏好模型应(1)具有较高的覆盖率，以便可用于 CoNLL 等大型数据集，以及(2)以非线性方式与其他知识源结合。我们的选择偏好模型稍微提高了协指解析性能，但是考虑到训练模型所需的额外资源，这种小的改进是否有利于解决协指是有争议的
加基线共引用解析器的复杂性。然而，它的性能与+binnedsim 相当。在开发集上，在测试集上推广更差。+表中的 SP 1 是+binnedsim 的性能。在测试集上。从结果中我们可以看到， 添加选择首选项作为二进制特征比基线有所改善
表$\mathbb{E}$1 中的增强显示了 Clark and Manning (2016a)的奖励重新缩放模型的结果，这是迄今为止官方测试集中报告的最高结果。Clark and Manning (2016a) 的奖励重新缩放模型在强化学习框架中转换了 Clark and Manning (2016b)的排名模型，这大大增加了训练时间， 在我们的实验中从两天增加到六天
我们分析了我们的选择偏好模型如何影响各种类型的提及的分辨率。我们使用 Martschat and Strube (2014)的工具包$^{4}$执行召回和错误分析。表 4 中报告了 +SP 中与测试集上的基线相比的召回次数和精度错误的差异
通过使用我们的选择首选项功能，对于所有类型的提及，召回错误的数量都会减少。减少回忆错误对于代词更为突出。另一方面，对于所有类型的提及，精度错误的数量都会增加。精度误差的增加对于普通名词来说是最高的。总体而言，+SP 创建的链接比基线多 260 个表 3 列出了开发集中的一些示例，其中+SP 创建了一个基线没有的链接。它还包括具有链接提及的高价值的相似性，并且可能是创建链接的原因。例如，在第一个例子中， 基于我们的模型，相似性(影响 @nsubj, 显示 @nsubj) 是已知的，它也高于相似性(影响
6
Acknowledgments 作者要感谢三位匿名审稿人的有益评论。这项工作得到了德国研究基金会 (DFG) 的支持，作为研究培训小组“来自异构来源的信息的自适应准备”(AIPHES)的一部分。GRK1994/1 和 KlausTschira 基金会， 德国海德堡。
## References
Eneko Agirre and David Martinez. 2001. Learn-
ing class-to-class selectional preferences. In Pro- $\bar{\text{ceedings of the ACL 2001 Workshop on Compu-}}$ tational Natural Language Learning (ConL$\hat{\mathrm{L} } )$ , pages 15-22.
$^{4}$https://github.com/smartschat/cort
|
[
[
0.41650390625,
0.07598876953125,
0.43994140625,
0.07598876953125,
0.43994140625,
0.08905029296875,
0.41650390625,
0.08905029296875
],
[
0.54833984375,
0.0736083984375,
0.60546875,
0.07598876953125,
0.603515625,
0.0914306640625,
0.54736328125,
0.08905029296875
],
[
0.66748046875,
0.07598876953125,
0.72314453125,
0.07598876953125,
0.72314453125,
0.09027099609375,
0.66748046875,
0.09027099609375
],
[
0.7841796875,
0.07659912109375,
0.8193359375,
0.07659912109375,
0.8193359375,
0.08905029296875,
0.7841796875,
0.08905029296875
],
[
0.84619140625,
0.0885009765625,
0.8759765625,
0.0914306640625,
0.873046875,
0.1051025390625,
0.8427734375,
0.10272216796875
],
[
0.258544921875,
0.07659912109375,
0.297119140625,
0.07659912109375,
0.297119140625,
0.08905029296875,
0.258544921875,
0.08905029296875
],
[
0.470947265625,
0.09027099609375,
0.50146484375,
0.0926513671875,
0.49951171875,
0.10394287109375,
0.468505859375,
0.1009521484375
],
[
0.62060546875,
0.08905029296875,
0.72314453125,
0.0914306640625,
0.72314453125,
0.105712890625,
0.62060546875,
0.10394287109375
],
[
0.429931640625,
0.0914306640625,
0.44580078125,
0.09027099609375,
0.447509765625,
0.1009521484375,
0.4306640625,
0.1015625
],
[
0.231689453125,
0.0926513671875,
0.246826171875,
0.0926513671875,
0.246826171875,
0.10272216796875,
0.231689453125,
0.10272216796875
],
[
0.2822265625,
0.0926513671875,
0.295654296875,
0.0926513671875,
0.295654296875,
0.1015625,
0.2822265625,
0.1015625
],
[
0.32080078125,
0.0914306640625,
0.355224609375,
0.0914306640625,
0.355224609375,
0.10394287109375,
0.32080078125,
0.10394287109375
],
[
0.382080078125,
0.0926513671875,
0.397216796875,
0.0926513671875,
0.397216796875,
0.10272216796875,
0.382080078125,
0.10272216796875
],
[
0.5322265625,
0.0926513671875,
0.54736328125,
0.0926513671875,
0.54736328125,
0.10272216796875,
0.5322265625,
0.10272216796875
],
[
0.58251953125,
0.0926513671875,
0.59619140625,
0.0926513671875,
0.59619140625,
0.1015625,
0.58251953125,
0.1015625
],
[
0.755859375,
0.0926513671875,
0.771484375,
0.0926513671875,
0.771484375,
0.10272216796875,
0.755859375,
0.10272216796875
],
[
0.80615234375,
0.0926513671875,
0.8212890625,
0.0926513671875,
0.8212890625,
0.10272216796875,
0.80615234375,
0.10272216796875
],
[
0.204833984375,
0.10809326171875,
0.2493896484375,
0.10809326171875,
0.2493896484375,
0.122314453125,
0.204833984375,
0.122314453125
],
[
0.307373046875,
0.10809326171875,
0.3525390625,
0.10809326171875,
0.3525390625,
0.122314453125,
0.307373046875,
0.122314453125
],
[
0.355224609375,
0.10809326171875,
0.399658203125,
0.10809326171875,
0.399658203125,
0.122314453125,
0.355224609375,
0.122314453125
],
[
0.49951171875,
0.10809326171875,
0.5498046875,
0.10809326171875,
0.5498046875,
0.122314453125,
0.49951171875,
0.122314453125
],
[
0.5556640625,
0.10809326171875,
0.60107421875,
0.10809326171875,
0.60107421875,
0.122314453125,
0.5556640625,
0.122314453125
],
[
0.60693359375,
0.10809326171875,
0.6513671875,
0.10809326171875,
0.6513671875,
0.122314453125,
0.60693359375,
0.122314453125
],
[
0.68115234375,
0.10809326171875,
0.77490234375,
0.10809326171875,
0.77490234375,
0.122314453125,
0.68115234375,
0.122314453125
],
[
0.779296875,
0.10809326171875,
0.82373046875,
0.10809326171875,
0.82373046875,
0.122314453125,
0.779296875,
0.122314453125
],
[
0.82958984375,
0.10809326171875,
0.8759765625,
0.10809326171875,
0.8759765625,
0.122314453125,
0.82958984375,
0.122314453125
],
[
0.408935546875,
0.1092529296875,
0.450927734375,
0.1092529296875,
0.450927734375,
0.122314453125,
0.408935546875,
0.122314453125
],
[
0.258544921875,
0.1092529296875,
0.295654296875,
0.1092529296875,
0.295654296875,
0.1217041015625,
0.258544921875,
0.1217041015625
],
[
0.459228515625,
0.1092529296875,
0.50146484375,
0.1092529296875,
0.50146484375,
0.122314453125,
0.459228515625,
0.122314453125
],
[
0.258544921875,
0.12054443359375,
0.297119140625,
0.122314453125,
0.295654296875,
0.1353759765625,
0.2568359375,
0.133056640625
],
[
0.1275634765625,
0.10986328125,
0.18896484375,
0.10986328125,
0.18896484375,
0.1217041015625,
0.1275634765625,
0.1217041015625
],
[
0.4072265625,
0.1217041015625,
0.447509765625,
0.12353515625,
0.44580078125,
0.1353759765625,
0.405517578125,
0.133056640625
],
[
0.50732421875,
0.12054443359375,
0.54833984375,
0.122314453125,
0.54736328125,
0.1353759765625,
0.50537109375,
0.133056640625
],
[
0.5576171875,
0.1217041015625,
0.60107421875,
0.12353515625,
0.599609375,
0.135986328125,
0.5556640625,
0.1341552734375
],
[
0.60791015625,
0.1217041015625,
0.64990234375,
0.12353515625,
0.6484375,
0.135986328125,
0.60693359375,
0.1341552734375
],
[
0.73046875,
0.1217041015625,
0.77490234375,
0.1217041015625,
0.77490234375,
0.135986328125,
0.73046875,
0.135986328125
],
[
0.8330078125,
0.1217041015625,
0.873046875,
0.12353515625,
0.87158203125,
0.135986328125,
0.8310546875,
0.1341552734375
],
[
0.30908203125,
0.122314453125,
0.34765625,
0.122314453125,
0.34765625,
0.1353759765625,
0.30908203125,
0.1353759765625
],
[
0.208251953125,
0.122314453125,
0.2493896484375,
0.122314453125,
0.2493896484375,
0.135986328125,
0.208251953125,
0.135986328125
],
[
0.358642578125,
0.122314453125,
0.399658203125,
0.122314453125,
0.399658203125,
0.135986328125,
0.358642578125,
0.135986328125
],
[
0.45751953125,
0.122314453125,
0.49951171875,
0.122314453125,
0.49951171875,
0.135986328125,
0.45751953125,
0.135986328125
],
[
0.6826171875,
0.122314453125,
0.72119140625,
0.122314453125,
0.72119140625,
0.1353759765625,
0.6826171875,
0.1353759765625
],
[
0.78076171875,
0.12353515625,
0.82275390625,
0.12353515625,
0.82275390625,
0.1353759765625,
0.78076171875,
0.1353759765625
],
[
0.50537109375,
0.1341552734375,
0.5498046875,
0.1341552734375,
0.5498046875,
0.1484375,
0.50537109375,
0.1484375
],
[
0.1376953125,
0.12353515625,
0.173828125,
0.12353515625,
0.173828125,
0.135986328125,
0.1376953125,
0.135986328125
],
[
0.1318359375,
0.1353759765625,
0.166259765625,
0.1353759765625,
0.166259765625,
0.1478271484375,
0.1318359375,
0.1478271484375
],
[
0.208251953125,
0.1353759765625,
0.2493896484375,
0.1353759765625,
0.2493896484375,
0.1484375,
0.208251953125,
0.1484375
],
[
0.258544921875,
0.1353759765625,
0.295654296875,
0.1353759765625,
0.295654296875,
0.1478271484375,
0.258544921875,
0.1478271484375
],
[
0.358642578125,
0.1353759765625,
0.399658203125,
0.1353759765625,
0.399658203125,
0.1484375,
0.358642578125,
0.1484375
],
[
0.4072265625,
0.1353759765625,
0.44921875,
0.1353759765625,
0.44921875,
0.1484375,
0.4072265625,
0.1484375
],
[
0.60791015625,
0.1353759765625,
0.64990234375,
0.1353759765625,
0.64990234375,
0.1484375,
0.60791015625,
0.1484375
],
[
0.73291015625,
0.1341552734375,
0.771484375,
0.135986328125,
0.77099609375,
0.1484375,
0.7314453125,
0.146728515625
],
[
0.8330078125,
0.1353759765625,
0.87158203125,
0.1353759765625,
0.87158203125,
0.1478271484375,
0.8330078125,
0.1478271484375
],
[
0.459228515625,
0.135986328125,
0.49951171875,
0.135986328125,
0.49951171875,
0.1478271484375,
0.459228515625,
0.1478271484375
],
[
0.30908203125,
0.135986328125,
0.349365234375,
0.135986328125,
0.349365234375,
0.1478271484375,
0.30908203125,
0.1478271484375
],
[
0.55908203125,
0.135986328125,
0.599609375,
0.135986328125,
0.599609375,
0.1478271484375,
0.55908203125,
0.1478271484375
],
[
0.6826171875,
0.135986328125,
0.72314453125,
0.135986328125,
0.72314453125,
0.1478271484375,
0.6826171875,
0.1478271484375
],
[
0.78271484375,
0.135986328125,
0.8212890625,
0.135986328125,
0.8212890625,
0.1478271484375,
0.78271484375,
0.1478271484375
],
[
0.50537109375,
0.1484375,
0.5498046875,
0.1484375,
0.5498046875,
0.163330078125,
0.50537109375,
0.163330078125
],
[
0.7255859375,
0.1484375,
0.771484375,
0.1484375,
0.771484375,
0.163330078125,
0.7255859375,
0.163330078125
],
[
0.8310546875,
0.1484375,
0.8759765625,
0.1484375,
0.8759765625,
0.163330078125,
0.8310546875,
0.163330078125
],
[
0.208251953125,
0.149658203125,
0.246826171875,
0.149658203125,
0.246826171875,
0.160888671875,
0.208251953125,
0.160888671875
],
[
0.130126953125,
0.149658203125,
0.198974609375,
0.149658203125,
0.198974609375,
0.160888671875,
0.130126953125,
0.160888671875
],
[
0.258544921875,
0.149658203125,
0.295654296875,
0.149658203125,
0.295654296875,
0.160888671875,
0.258544921875,
0.160888671875
],
[
0.30908203125,
0.149658203125,
0.3525390625,
0.149658203125,
0.3525390625,
0.160888671875,
0.30908203125,
0.160888671875
],
[
0.35693359375,
0.149658203125,
0.39794921875,
0.149658203125,
0.39794921875,
0.160888671875,
0.35693359375,
0.160888671875
],
[
0.408935546875,
0.149658203125,
0.447509765625,
0.149658203125,
0.447509765625,
0.160888671875,
0.408935546875,
0.160888671875
],
[
0.459228515625,
0.149658203125,
0.49951171875,
0.149658203125,
0.49951171875,
0.160888671875,
0.459228515625,
0.160888671875
],
[
0.55908203125,
0.149658203125,
0.599609375,
0.149658203125,
0.599609375,
0.160888671875,
0.55908203125,
0.160888671875
],
[
0.60791015625,
0.149658203125,
0.6513671875,
0.149658203125,
0.6513671875,
0.160888671875,
0.60791015625,
0.160888671875
],
[
0.6826171875,
0.149658203125,
0.724609375,
0.149658203125,
0.724609375,
0.163330078125,
0.6826171875,
0.163330078125
],
[
0.78271484375,
0.149658203125,
0.8212890625,
0.149658203125,
0.8212890625,
0.160888671875,
0.78271484375,
0.160888671875
],
[
0.310546875,
0.177001953125,
0.38037109375,
0.177001953125,
0.38037109375,
0.188232421875,
0.310546875,
0.188232421875
],
[
0.385498046875,
0.174560546875,
0.68408203125,
0.177001953125,
0.68408203125,
0.190673828125,
0.385498046875,
0.188232421875
],
[
0.2822265625,
0.210205078125,
0.30908203125,
0.210205078125,
0.30908203125,
0.22265625,
0.2822265625,
0.22265625
],
[
0.443359375,
0.210205078125,
0.478515625,
0.2120361328125,
0.476806640625,
0.223876953125,
0.441650390625,
0.221435546875
],
[
0.31982421875,
0.21142578125,
0.365234375,
0.21142578125,
0.365234375,
0.22265625,
0.31982421875,
0.22265625
],
[
0.22412109375,
0.21142578125,
0.26025390625,
0.21142578125,
0.26025390625,
0.22265625,
0.22412109375,
0.22265625
],
[
0.3837890625,
0.21142578125,
0.42822265625,
0.21142578125,
0.42822265625,
0.22265625,
0.3837890625,
0.22265625
],
[
0.5146484375,
0.213134765625,
0.6943359375,
0.213134765625,
0.6943359375,
0.22802734375,
0.5146484375,
0.22802734375
],
[
0.312255859375,
0.221435546875,
0.3896484375,
0.22265625,
0.3896484375,
0.2340087890625,
0.312255859375,
0.2327880859375
],
[
0.1292724609375,
0.232177734375,
0.1796875,
0.232177734375,
0.1796875,
0.2423095703125,
0.1292724609375,
0.2423095703125
],
[
0.22412109375,
0.232177734375,
0.26123046875,
0.232177734375,
0.26123046875,
0.243408203125,
0.22412109375,
0.243408203125
],
[
0.270263671875,
0.232177734375,
0.307373046875,
0.232177734375,
0.307373046875,
0.243408203125,
0.270263671875,
0.243408203125
],
[
0.337646484375,
0.232177734375,
0.3720703125,
0.232177734375,
0.3720703125,
0.243408203125,
0.337646484375,
0.243408203125
],
[
0.393798828125,
0.232177734375,
0.4306640625,
0.232177734375,
0.4306640625,
0.243408203125,
0.393798828125,
0.243408203125
],
[
0.43994140625,
0.232177734375,
0.480224609375,
0.232177734375,
0.480224609375,
0.243408203125,
0.43994140625,
0.243408203125
],
[
0.53564453125,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.53564453125,
0.24462890625
],
[
0.270263671875,
0.2410888671875,
0.306396484375,
0.243408203125,
0.3046875,
0.255859375,
0.268798828125,
0.254150390625
],
[
0.441650390625,
0.2410888671875,
0.47607421875,
0.243408203125,
0.474365234375,
0.255859375,
0.43994140625,
0.254150390625
],
[
0.3359375,
0.2423095703125,
0.370361328125,
0.2423095703125,
0.370361328125,
0.254638671875,
0.3359375,
0.254638671875
],
[
0.22412109375,
0.2423095703125,
0.26123046875,
0.2423095703125,
0.26123046875,
0.254638671875,
0.22412109375,
0.254638671875
],
[
0.3955078125,
0.2423095703125,
0.4306640625,
0.2423095703125,
0.4306640625,
0.254638671875,
0.3955078125,
0.254638671875
],
[
0.515625,
0.2464599609375,
0.8818359375,
0.2464599609375,
0.8818359375,
0.260009765625,
0.515625,
0.260009765625
],
[
0.3359375,
0.251708984375,
0.370361328125,
0.251708984375,
0.370361328125,
0.26318359375,
0.3359375,
0.26318359375
],
[
0.130126953125,
0.2529296875,
0.2091064453125,
0.2529296875,
0.2091064453125,
0.26318359375,
0.130126953125,
0.26318359375
],
[
0.1318359375,
0.243408203125,
0.208251953125,
0.243408203125,
0.208251953125,
0.254150390625,
0.1318359375,
0.254150390625
],
[
0.2216796875,
0.2529296875,
0.26123046875,
0.2529296875,
0.26123046875,
0.26416015625,
0.2216796875,
0.26416015625
],
[
0.270263671875,
0.2529296875,
0.306396484375,
0.2529296875,
0.306396484375,
0.26416015625,
0.270263671875,
0.26416015625
],
[
0.3955078125,
0.254150390625,
0.42822265625,
0.254150390625,
0.42822265625,
0.262451171875,
0.3955078125,
0.262451171875
],
[
0.441650390625,
0.2529296875,
0.478515625,
0.2529296875,
0.478515625,
0.26416015625,
0.441650390625,
0.26416015625
],
[
0.337646484375,
0.26318359375,
0.364501953125,
0.26318359375,
0.364501953125,
0.275634765625,
0.337646484375,
0.275634765625
],
[
0.5146484375,
0.26123046875,
0.8818359375,
0.262451171875,
0.8818359375,
0.276611328125,
0.5146484375,
0.275634765625
],
[
0.270263671875,
0.2724609375,
0.306396484375,
0.2724609375,
0.306396484375,
0.28515625,
0.270263671875,
0.28515625
],
[
0.337646484375,
0.2724609375,
0.370361328125,
0.2724609375,
0.370361328125,
0.283935546875,
0.337646484375,
0.283935546875
],
[
0.3955078125,
0.2724609375,
0.4306640625,
0.2724609375,
0.4306640625,
0.283935546875,
0.3955078125,
0.283935546875
],
[
0.43994140625,
0.2724609375,
0.480224609375,
0.2724609375,
0.480224609375,
0.283935546875,
0.43994140625,
0.283935546875
],
[
0.2275390625,
0.27490234375,
0.258544921875,
0.27490234375,
0.258544921875,
0.283203125,
0.2275390625,
0.283203125
],
[
0.130126953125,
0.273681640625,
0.181396484375,
0.273681640625,
0.181396484375,
0.283935546875,
0.130126953125,
0.283935546875
],
[
0.27197265625,
0.28076171875,
0.306396484375,
0.283203125,
0.3046875,
0.29443359375,
0.270263671875,
0.292236328125
],
[
0.3359375,
0.281982421875,
0.370361328125,
0.281982421875,
0.370361328125,
0.29345703125,
0.3359375,
0.29345703125
],
[
0.441650390625,
0.281982421875,
0.47607421875,
0.281982421875,
0.47607421875,
0.29345703125,
0.441650390625,
0.29345703125
],
[
0.5146484375,
0.276611328125,
0.8818359375,
0.27783203125,
0.8818359375,
0.292236328125,
0.5146484375,
0.29150390625
],
[
0.133544921875,
0.283203125,
0.208251953125,
0.283203125,
0.208251953125,
0.29443359375,
0.133544921875,
0.29443359375
],
[
0.22412109375,
0.283203125,
0.26025390625,
0.283203125,
0.26025390625,
0.29443359375,
0.22412109375,
0.29443359375
],
[
0.3955078125,
0.283203125,
0.429931640625,
0.283203125,
0.429931640625,
0.29443359375,
0.3955078125,
0.29443359375
],
[
0.3955078125,
0.29150390625,
0.429931640625,
0.29345703125,
0.42822265625,
0.305908203125,
0.393798828125,
0.303955078125
],
[
0.441650390625,
0.290283203125,
0.478515625,
0.292236328125,
0.476806640625,
0.3046875,
0.43994140625,
0.302734375
],
[
0.27197265625,
0.292236328125,
0.307373046875,
0.292236328125,
0.307373046875,
0.3046875,
0.27197265625,
0.3046875
],
[
0.130126953125,
0.292236328125,
0.2091064453125,
0.29443359375,
0.2091064453125,
0.3046875,
0.130126953125,
0.302734375
],
[
0.225830078125,
0.29345703125,
0.26025390625,
0.29345703125,
0.26025390625,
0.3046875,
0.225830078125,
0.3046875
],
[
0.3359375,
0.292236328125,
0.370361328125,
0.292236328125,
0.370361328125,
0.3046875,
0.3359375,
0.3046875
],
[
0.515625,
0.29345703125,
0.880859375,
0.29345703125,
0.880859375,
0.30712890625,
0.515625,
0.30712890625
],
[
0.51904296875,
0.31005859375,
0.8818359375,
0.308837890625,
0.8818359375,
0.32373046875,
0.51904296875,
0.32470703125
],
[
0.12176513671875,
0.32373046875,
0.1956787109375,
0.32373046875,
0.1956787109375,
0.3349609375,
0.12176513671875,
0.3349609375
],
[
0.19482421875,
0.322509765625,
0.486083984375,
0.322509765625,
0.486083984375,
0.33740234375,
0.19482421875,
0.33740234375
],
[
0.51708984375,
0.32666015625,
0.8818359375,
0.32666015625,
0.8818359375,
0.34033203125,
0.51708984375,
0.34033203125
],
[
0.118408203125,
0.33837890625,
0.48779296875,
0.33837890625,
0.48779296875,
0.352783203125,
0.118408203125,
0.352783203125
],
[
0.5205078125,
0.34326171875,
0.86572265625,
0.34326171875,
0.86572265625,
0.35693359375,
0.5205078125,
0.35693359375
],
[
0.861328125,
0.347412109375,
0.8671875,
0.347412109375,
0.8671875,
0.352783203125,
0.861328125,
0.352783203125
],
[
0.8671875,
0.347412109375,
0.8759765625,
0.347412109375,
0.8759765625,
0.3515625,
0.8671875,
0.3515625
],
[
0.1175537109375,
0.355224609375,
0.476806640625,
0.35400390625,
0.476806640625,
0.3681640625,
0.1175537109375,
0.369384765625
],
[
0.5146484375,
0.358154296875,
0.8818359375,
0.35986328125,
0.8818359375,
0.37353515625,
0.5146484375,
0.371826171875
],
[
0.515625,
0.3759765625,
0.880859375,
0.3759765625,
0.880859375,
0.3896484375,
0.515625,
0.3896484375
],
[
0.12176513671875,
0.396728515625,
0.48779296875,
0.396728515625,
0.48779296875,
0.410888671875,
0.12176513671875,
0.410888671875
],
[
0.515625,
0.391357421875,
0.8818359375,
0.391357421875,
0.8818359375,
0.405029296875,
0.515625,
0.405029296875
],
[
0.12005615234375,
0.412109375,
0.489501953125,
0.412109375,
0.489501953125,
0.427001953125,
0.12005615234375,
0.427001953125
],
[
0.5146484375,
0.407958984375,
0.880859375,
0.407958984375,
0.880859375,
0.421630859375,
0.5146484375,
0.421630859375
],
[
0.12005615234375,
0.427001953125,
0.489501953125,
0.4287109375,
0.489501953125,
0.443603515625,
0.12005615234375,
0.441162109375
],
[
0.5146484375,
0.42333984375,
0.880859375,
0.42333984375,
0.880859375,
0.43701171875,
0.5146484375,
0.43701171875
],
[
0.5146484375,
0.439453125,
0.8818359375,
0.439453125,
0.8818359375,
0.45361328125,
0.5146484375,
0.45361328125
],
[
0.118408203125,
0.4453125,
0.483642578125,
0.4453125,
0.483642578125,
0.458984375,
0.118408203125,
0.458984375
],
[
0.118408203125,
0.46142578125,
0.486083984375,
0.46142578125,
0.486083984375,
0.4755859375,
0.118408203125,
0.4755859375
],
[
0.515625,
0.457275390625,
0.609375,
0.457275390625,
0.609375,
0.468505859375,
0.515625,
0.468505859375
],
[
0.12005615234375,
0.47802734375,
0.14111328125,
0.47802734375,
0.14111328125,
0.4912109375,
0.12005615234375,
0.4912109375
],
[
0.5146484375,
0.48681640625,
0.52880859375,
0.48681640625,
0.52880859375,
0.496337890625,
0.5146484375,
0.496337890625
],
[
0.5498046875,
0.486328125,
0.6533203125,
0.486328125,
0.6533203125,
0.49755859375,
0.5498046875,
0.49755859375
],
[
0.1376953125,
0.493408203125,
0.489501953125,
0.49462890625,
0.489501953125,
0.5087890625,
0.1376953125,
0.50830078125
],
[
0.12005615234375,
0.51123046875,
0.48779296875,
0.51123046875,
0.48779296875,
0.52490234375,
0.12005615234375,
0.52490234375
],
[
0.515625,
0.51123046875,
0.8818359375,
0.51123046875,
0.8818359375,
0.52490234375,
0.515625,
0.52490234375
],
[
0.12005615234375,
0.52685546875,
0.48779296875,
0.52685546875,
0.48779296875,
0.54052734375,
0.12005615234375,
0.54052734375
],
[
0.51318359375,
0.52783203125,
0.8818359375,
0.52783203125,
0.8818359375,
0.54150390625,
0.51318359375,
0.54150390625
],
[
0.12005615234375,
0.5419921875,
0.489501953125,
0.5419921875,
0.489501953125,
0.5556640625,
0.12005615234375,
0.5556640625
],
[
0.515625,
0.54345703125,
0.8818359375,
0.54345703125,
0.8818359375,
0.55712890625,
0.515625,
0.55712890625
],
[
0.12005615234375,
0.55859375,
0.489501953125,
0.55859375,
0.489501953125,
0.572265625,
0.12005615234375,
0.572265625
],
[
0.515625,
0.56005859375,
0.8818359375,
0.56005859375,
0.8818359375,
0.57373046875,
0.515625,
0.57373046875
],
[
0.1234130859375,
0.57666015625,
0.489501953125,
0.57666015625,
0.489501953125,
0.58740234375,
0.1234130859375,
0.58740234375
],
[
0.51318359375,
0.57470703125,
0.8818359375,
0.57470703125,
0.8818359375,
0.58837890625,
0.51318359375,
0.58837890625
],
[
0.118408203125,
0.59033203125,
0.3955078125,
0.59033203125,
0.3955078125,
0.60400390625,
0.118408203125,
0.60400390625
],
[
0.5146484375,
0.59033203125,
0.8818359375,
0.59130859375,
0.8818359375,
0.60546875,
0.5146484375,
0.60498046875
],
[
0.1376953125,
0.60791015625,
0.486083984375,
0.60791015625,
0.486083984375,
0.62158203125,
0.1376953125,
0.62158203125
],
[
0.515625,
0.60791015625,
0.8818359375,
0.60791015625,
0.8818359375,
0.62158203125,
0.515625,
0.62158203125
],
[
0.12005615234375,
0.62451171875,
0.489501953125,
0.62451171875,
0.489501953125,
0.63818359375,
0.12005615234375,
0.63818359375
],
[
0.5146484375,
0.62451171875,
0.8818359375,
0.62451171875,
0.8818359375,
0.63818359375,
0.5146484375,
0.63818359375
],
[
0.12005615234375,
0.64013671875,
0.48779296875,
0.64013671875,
0.48779296875,
0.65380859375,
0.12005615234375,
0.65380859375
],
[
0.5146484375,
0.64111328125,
0.55322265625,
0.64111328125,
0.55322265625,
0.65380859375,
0.5146484375,
0.65380859375
],
[
0.118408203125,
0.65625,
0.48779296875,
0.65478515625,
0.48779296875,
0.66943359375,
0.118408203125,
0.67041015625
],
[
0.12005615234375,
0.67333984375,
0.432373046875,
0.67333984375,
0.432373046875,
0.68701171875,
0.12005615234375,
0.68701171875
],
[
0.5146484375,
0.67041015625,
0.5283203125,
0.67041015625,
0.5283203125,
0.681640625,
0.5146484375,
0.681640625
],
[
0.1376953125,
0.6904296875,
0.486083984375,
0.6904296875,
0.486083984375,
0.70361328125,
0.1376953125,
0.70361328125
],
[
0.515625,
0.6953125,
0.88330078125,
0.6953125,
0.88330078125,
0.708984375,
0.515625,
0.708984375
],
[
0.12005615234375,
0.7060546875,
0.486083984375,
0.7060546875,
0.486083984375,
0.7197265625,
0.12005615234375,
0.7197265625
],
[
0.51904296875,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.51904296875,
0.724609375
],
[
0.118408203125,
0.7216796875,
0.486083984375,
0.7216796875,
0.486083984375,
0.7353515625,
0.118408203125,
0.7353515625
],
[
0.515625,
0.72705078125,
0.87890625,
0.72705078125,
0.87890625,
0.74072265625,
0.515625,
0.74072265625
],
[
0.12005615234375,
0.73681640625,
0.490234375,
0.73828125,
0.490234375,
0.7529296875,
0.12005615234375,
0.751953125
],
[
0.515625,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75830078125,
0.515625,
0.75830078125
],
[
0.12005615234375,
0.7548828125,
0.489501953125,
0.7548828125,
0.489501953125,
0.7685546875,
0.12005615234375,
0.7685546875
],
[
0.52880859375,
0.76025390625,
0.88330078125,
0.76025390625,
0.88330078125,
0.77392578125,
0.52880859375,
0.77392578125
],
[
0.12005615234375,
0.77099609375,
0.476806640625,
0.77099609375,
0.476806640625,
0.78369140625,
0.12005615234375,
0.78369140625
],
[
0.51318359375,
0.77490234375,
0.615234375,
0.77685546875,
0.615234375,
0.79150390625,
0.51318359375,
0.7890625
],
[
0.1376953125,
0.78759765625,
0.48779296875,
0.78759765625,
0.48779296875,
0.80029296875,
0.1376953125,
0.80029296875
],
[
0.12176513671875,
0.80419921875,
0.48779296875,
0.80419921875,
0.48779296875,
0.81787109375,
0.12176513671875,
0.81787109375
],
[
0.12176513671875,
0.8193359375,
0.48779296875,
0.8193359375,
0.48779296875,
0.8330078125,
0.12176513671875,
0.8330078125
],
[
0.5146484375,
0.82177734375,
0.60693359375,
0.82177734375,
0.60693359375,
0.8330078125,
0.5146484375,
0.8330078125
],
[
0.12005615234375,
0.8349609375,
0.484375,
0.8349609375,
0.484375,
0.8486328125,
0.12005615234375,
0.8486328125
],
[
0.51123046875,
0.84130859375,
0.8818359375,
0.84326171875,
0.8818359375,
0.85693359375,
0.51123046875,
0.85498046875
],
[
0.12005615234375,
0.8515625,
0.489501953125,
0.8515625,
0.489501953125,
0.865234375,
0.12005615234375,
0.865234375
],
[
0.52880859375,
0.85546875,
0.8818359375,
0.85546875,
0.8818359375,
0.869140625,
0.52880859375,
0.869140625
],
[
0.12005615234375,
0.8681640625,
0.489501953125,
0.8681640625,
0.489501953125,
0.8818359375,
0.12005615234375,
0.8818359375
],
[
0.53076171875,
0.869140625,
0.8818359375,
0.869140625,
0.8818359375,
0.8828125,
0.53076171875,
0.8828125
],
[
0.53076171875,
0.88134765625,
0.88671875,
0.8818359375,
0.88671875,
0.896484375,
0.53076171875,
0.8955078125
],
[
0.1376953125,
0.89453125,
0.397216796875,
0.8955078125,
0.397216796875,
0.91015625,
0.1376953125,
0.9091796875
],
[
0.5322265625,
0.896484375,
0.62451171875,
0.89453125,
0.62451171875,
0.90771484375,
0.5322265625,
0.91015625
]
] |
[
"B3",
"CEAF",
"CoNLL",
"LEA",
"F1",
"MUC",
"F1",
"F1Avg.F1",
"P",
"R",
"P",
"F1",
"R",
"R",
"P",
"R",
"P",
"70.09",
"74.72",
"57.64",
"54.47",
"63.92",
"58.82",
"65.6054.02",
"66.45",
"59.59",
"70.09",
"80.01",
"63.26",
"79.96",
"baseline",
"70.31",
"54.41",
"64.08",
"58.85",
"53.93",
"59.66",
"74.71",
"70.10",
"57.51",
"63.27",
"65.61",
"66.76",
"55.25",
"gov",
"+SP",
"70.85",
"79.31",
"58.93",
"69.16",
"59.21",
"55.29",
"59.98",
"63.64",
"74.85",
"63.78",
"65.90",
"65.53",
"55.66",
"55.31",
"59.90",
"70.98",
"Reinforce",
"78.81",
"74.69",
"58.97",
"69.05",
"63.61",
"63.28",
"59.23",
"65.84",
"65.32",
"Table 1:",
"Results on the CoNLL 2012 test set.",
"B 3",
"LEA",
"CEAF",
"MUC",
"CoNLL",
"@dobj，显示@nsubj)",
"development",
"baseline",
"74.10",
"63.95",
"59.73",
"65.93",
"60.16",
"为了估计预期性能提升的更高界限，我们仅",
"64.42",
"60.65",
"60.45",
"74.38",
"66.42",
"在照应性提及上运行基线和+SP模型。根据",
"60.21",
"+binned sim",
"+embedding",
"74.36",
"64.54",
"66.37",
"60.77",
"test",
"CoNLL分数和LEA，通过使用照应式提及",
"63.26",
"58.82",
"65.60",
"59.59",
"74.72",
"baseline",
"63.41",
"59.03",
"59.69",
"性能提高了1%。此结果表明，选择偏好的结",
"+embedding",
"74.53",
"65.66",
"65.90",
"59.98",
"63.64",
"+binned sim",
"74.85",
"59.21",
"合为非照应性提及创建了许多链接，从而降低",
"了准确性。因此，当使用系统提及时，整体性",
"Table 2:",
"Incorporating the selectional prefer-",
"能没有显著提高。deep-coref在解析时间合并",
"ence model as new embeddings (+embedding)",
"了隐喻性分数。进一步改进+SP的结果的",
"",
"",
"vs. as new pairwise features (+binned sim.).",
"种可能的方法是在输入水平上合并隐喻性分",
"数。通过这种方式，共同引用解析器可以学习",
"加基线共引用解析器的复杂性。然而，它的性",
"使用选择偏好主要用于更可能是照应的提及。",
"能与+binnedsim相当。在开发集上，在测试",
"但是，考虑到当前的反常性确定器或单例检测",
"集上推广更差。+表中的SP1是+binnedsim",
"器的F1得分仅在85%(MoosaviandStrube",
"2016a，2017）左右，使用系统反常性得分的影",
"的性能。在测试集上。从结果中我们可以看到",
"添加选择首选项作为二进制特征比基线有所改",
"响可能很小",
"善",
"5",
"Conclusions",
"表E1中的增强显示了ClarkandManning",
"(2016a）的奖励重新缩放模型的结果，这是迄",
"我们引入了一种新的选择首选项模型，该模型",
"今为止官方测试集中报告的最高结果。Clark",
"结合了基于依赖关系的单词嵌入和细粒度的实",
"andManning（2016a）的奖励重新缩放模型在",
"体类型。为了有效，选择偏好模型应（1）具",
"强化学习框架中转换了Clark and Manning",
"有较高的覆盖率，以便可用于CoNLL等大型",
"（2016b）的排名模型，这大大增加了训练时间",
"数据集，以及（2）以非线性方式与其他知识",
"在我们的实验中从两天增加到六天",
"源结合。我们的选择偏好模型稍微提高了协指",
"我们分析了我们的选择偏好模型如何影响各",
"解析性能，但是考虑到训练模型所需的额外资",
"种类型的提及的分辨率。我们使用Martschat",
"源，这种小的改进是否有利于解决协指是有争",
"andStrube（2014）的工具包4执行召回和错",
"议的",
"误分析。表4中报告了+SP中与测试集上的",
"基线相比的召回次数和精度错误的差异",
"6",
"通过使用我们的选择首选项功能，对于所有",
"Acknowledgments作者要感谢三位匿名审稿",
"类型的提及，召回错误的数量都会减少。减少",
"人的有益评论。这项工作得到了德国研究基金",
"回忆错误对于代词更为突出。另一方面，对于",
"会（DFG）的支持，作为研究培训小组“来自",
"所有类型的提及，精度错误的数量都会增加。",
"异构来源的信息的自适应准备”（AIPHES）的",
"精度误差的增加对于普通名词来说是最高的。",
"-部分。GRK1994/1和KlausTschira基金会，",
"总体而言，+SP创建的链接比基线多260个",
"德国海德堡。",
"表3列出了开发集中的一些示例，其中",
"+SP创建了一个基线没有的链接。它还包括",
"具有链接提及的高价值的相似性，并且可能",
"References",
"是创建链接的原因。例如，在第一个例子中",
"Eneko Agirre and David Martinez. 2001. Learn-",
"基于我们的模型，相似性（影响@nsubi，显",
"ing class-to-class selectional preferences. In Pro-",
"示@nsubi）是已知的，它也高于相似性（影响",
"ceedings of the ACL 2001 Workshop on Compu-",
"tational Natural Language Learning (ConLL) ,",
"4https://github.com/smartschat/cort",
"pages 15-22"
] |
{
"arxiv_id": "1707.06456",
"language": "zh",
"page_idx": 3
}
|
||
2403.09702-zh-1
|
2403.09702-zh-1.jpg
|
Table 1: Statistics of CRED. Avg. RT denotes the average
number of retweets.
<table>
<tbody>
<tr>
<th>Topic</th>
<th>Avg. RT</th>
<th>Pairs</th>
</tr>
<tr>
<td>Business& Entrepreneurs</td>
<td>743.2</td>
<td>4,186</td>
</tr>
<tr>
<td>Fitness&Health</td>
<td>1010.2</td>
<td>628</td>
</tr>
<tr>
<td>Learning $\&$Educational</td>
<td>485.1</td>
<td>90</td>
</tr>
<tr>
<td>Sports</td>
<td>728.0</td>
<td> </td>
</tr>
<tr>
<td>Total</td>
<td> </td>
<td>4,910</td>
</tr>
</tbody>
</table>
Ghosh, et al
<FigureHere>
Figure 2: Architecture of GGEA.
# 2 $\textbf{RELATED WORK}$
自平台成立以来，最大化受众范围和反应的愿望一直是社交媒体用户的一致主题。几项研究的重点是理解和预测导致社交媒体帖子广泛传播的因素。Cha et al.[5] 对 5200 万 Twitter 用户的活动进行了分析，得出一条推文的内容是推动转发的关键因素。Suh et al.[13] 评估了 7400 万条推文，并探索了一系列基于内容和上下文的功能，这些功能会影响给定推文的转发次数。他们开发了一个基于广义线性模型的模型来预测一条推文预计会产生的转发次数。他们的研究强调了推文内容、网址和标签作为预测转发数量的关键特征的重要性。利用社交网络进行级联预测的其他工作包括[14]和[9]
同样，Bakshy et al.[2]跟踪了1的扩散活动。600 万 Twitter 用户，通过研究不同类别的影响者之间的关系和相关成本。除了 Twitter 之外，还对 Facebook 帖子的覆盖范围进行了研究。Bernstein et al.[3] 研究了 222,000 名 Facebook 用户的活动，揭示了无形的受众如何增强 Facebook 帖子的影响力。Valkonen et al.[16] 调查了芬兰消费者发布 Facebook 帖子的时间如何影响可达性。Tan et al.[14] 分析了推文内容对其影响范围的影响。最近的研究，如 Gabriel et al.[8] 展示了如何使用基于 GPT-2 [10] 和 T5 [11] 的模型来估计新闻文章的范围。但是，他们的方法仅提供粗粒度的分析，对进行比较分析没有显着贡献。他们通过注释从 1 到 5 的到达级别来创建数据集
我们的工作在几个重要方面与之前的研究不同。首先，我们引人了一个新的数据集 CRED, 它专门针对政府背景下社交媒体管理者的决策过程。其次，我们提出了一种新颖的方法GGEA, 该方法利用 LLM 的分析能力来预测人群对社交媒体帖子的反应。最后，与以前的作品不同，我们使用基于 T5 的释义器，通过 GGEA 使用 ChatGPT 响应 [17]进行优化，以生成和评估给定帖子的释义，从而允许用户通过探索其内容的不同版本来优化参与度
# 3 DATASET
我们的主要任务是确定一条表示为$t$1 的推文是否会比另一条表示为$t$2 的推文收到更多的转发。给定一对推文，我们的目标是准确预测具有较高转发计数的推文
我们从 2020 年 11 月至 2022 年 10 月白宫官方 Twitter 句柄发布的所有推文编译我们的数据集。然后，我们继续整理成对数据，确保每对数据都符合与时间、主题和人群反应相关的条件
我们仅保留满足以下四个条件的实例：
$\bullet$ The tweets were posted during weekdays: Weekdays gen-
erally witness higher levels of social media activity, making them ideal for capturing more interactions and thus providing a more reliable measure of crowd reaction.
$\bullet$ The retweet counts of $t1$ and $t2$ differed by at least 10%
of the less re-tweeted tweet. This condition ensures a clear distinction in the crowd reaction between the two tweets, enabling a more accurate and meaningful comparison.
$\bullet$ $t1$ and $t2$ were posted within a gap of 10 days and the dif-
ference between time of posting $t1$ and $t2$ was less than or equal to five hours. A short interval between posts helps control the extraneous variables such as time-specific events that might impact retweet counts.
$\bullet$ We utilized the Twitter Topic classification model [1] to
classify tweets into topics. We considered only those tweet pairs where the model assigned the same topic to both the tweets with a probability of 0.8 or higher. This high probability threshold ensures that the topics of the tweets are clearly defined and comparable, which is critical for the meaningful analysis of crowd reactions to different tweets on the same topic.
表$\mathbb{E}$1 显示了我们建议的人群反应估计数据集(CRED)的统计信息，并且建议的 CRED 中的标签分布很平衡。不同主题的平均转发次数强调了数据集构建过程中主题控制的重要性。主题方面的统计数据进一步强调，大多数政府推文与商业和企业家精神有关。对于数据集分离，我们使用来自初始 1 的推文。5年，我。e。,到2022年4月，这构成了4304个用于培训的推文对。最后6个月，我。e。, 2022年5月至 2022年 10 月， 提供 606 条用于模型验证的 tweet 对
## 4 мЕТНОDOLOGY
我们在图 2 中介绍了 GGEA 的整体架构。该架构结合了 FLANGRoBERTA [12] 模型，并通过交叉编码器变压器架构进行了微调。输入是来自给定对$(t1,t2)$的文本的串联，以及通过提示 Claude 阐明为什么该对中的特定推文更具吸引力而获得的响应。预期的输出是$t$1 是否会收到比$t$2 更多的转发
为了增加推理过程中的可用性，我们采用了现有的基于开源 T5 [11] 的释义器 [17] , 该释义器在 ChatGPT 生成的释义上进行了训练。这个释义器用于为给定的社交媒体帖子生成替代方案。我们一次检查两个释义。在我们开发的框架的帮助下， 从包含释义和原始帖子的候选列表中选择预测会获得最大反应的文本。对于克劳德部分，采用两种类型的提示。首先，我们询问 LLM 第一条推文是否会比第二条推文吸引更多的反应。其次，我们一次用一条推文提示 LLM, 查询它以解释使给定推
|
[
[
0.8486328125,
0.07574462890625,
0.90673828125,
0.07574462890625,
0.90673828125,
0.08709716796875,
0.8486328125,
0.08709716796875
],
[
0.0849609375,
0.10418701171875,
0.4794921875,
0.10479736328125,
0.4794921875,
0.11932373046875,
0.0849609375,
0.11871337890625
],
[
0.0849609375,
0.11932373046875,
0.2197265625,
0.12060546875,
0.2197265625,
0.1319580078125,
0.0849609375,
0.1307373046875
],
[
0.67333984375,
0.1142578125,
0.74853515625,
0.1142578125,
0.74853515625,
0.121826171875,
0.67333984375,
0.121826171875
],
[
0.61376953125,
0.1256103515625,
0.6396484375,
0.1256103515625,
0.6396484375,
0.1767578125,
0.61376953125,
0.1767578125
],
[
0.67333984375,
0.121826171875,
0.7490234375,
0.121826171875,
0.7490234375,
0.1319580078125,
0.67333984375,
0.1319580078125
],
[
0.7060546875,
0.1307373046875,
0.71728515625,
0.1307373046875,
0.71728515625,
0.141357421875,
0.7060546875,
0.141357421875
],
[
0.68896484375,
0.1458740234375,
0.734375,
0.1458740234375,
0.734375,
0.1572265625,
0.68896484375,
0.1572265625
],
[
0.21728515625,
0.154052734375,
0.2484130859375,
0.154052734375,
0.2484130859375,
0.1666259765625,
0.21728515625,
0.1666259765625
],
[
0.316162109375,
0.1546630859375,
0.36279296875,
0.15283203125,
0.363525390625,
0.1654052734375,
0.31787109375,
0.1673583984375
],
[
0.373291015625,
0.1546630859375,
0.408447265625,
0.1546630859375,
0.408447265625,
0.1666259765625,
0.373291015625,
0.1666259765625
],
[
0.16015625,
0.1666259765625,
0.303955078125,
0.1666259765625,
0.303955078125,
0.179931640625,
0.16015625,
0.179931640625
],
[
0.33251953125,
0.1673583984375,
0.36279296875,
0.1673583984375,
0.36279296875,
0.1787109375,
0.33251953125,
0.1787109375
],
[
0.374267578125,
0.1673583984375,
0.406982421875,
0.1673583984375,
0.406982421875,
0.179931640625,
0.374267578125,
0.179931640625
],
[
0.16259765625,
0.179931640625,
0.25244140625,
0.179931640625,
0.25244140625,
0.1912841796875,
0.16259765625,
0.1912841796875
],
[
0.326904296875,
0.179931640625,
0.363525390625,
0.179931640625,
0.363525390625,
0.1912841796875,
0.326904296875,
0.1912841796875
],
[
0.38232421875,
0.179931640625,
0.406005859375,
0.179931640625,
0.406005859375,
0.1925048828125,
0.38232421875,
0.1925048828125
],
[
0.67333984375,
0.1749267578125,
0.75244140625,
0.1749267578125,
0.75244140625,
0.186279296875,
0.67333984375,
0.186279296875
],
[
0.16259765625,
0.1925048828125,
0.2900390625,
0.1925048828125,
0.2900390625,
0.20263671875,
0.16259765625,
0.20263671875
],
[
0.33251953125,
0.1925048828125,
0.36279296875,
0.1925048828125,
0.36279296875,
0.2052001953125,
0.33251953125,
0.2052001953125
],
[
0.390625,
0.1932373046875,
0.406005859375,
0.1932373046875,
0.406005859375,
0.203857421875,
0.390625,
0.203857421875
],
[
0.330810546875,
0.2052001953125,
0.363525390625,
0.2052001953125,
0.363525390625,
0.2171630859375,
0.330810546875,
0.2171630859375
],
[
0.400390625,
0.20263671875,
0.414306640625,
0.21337890625,
0.40185546875,
0.2215576171875,
0.3896484375,
0.212158203125
],
[
0.736328125,
0.208984375,
0.74658203125,
0.20703125,
0.74853515625,
0.212158203125,
0.73779296875,
0.2139892578125
],
[
0.158447265625,
0.2171630859375,
0.19287109375,
0.2171630859375,
0.19287109375,
0.2298583984375,
0.158447265625,
0.2298583984375
],
[
0.16015625,
0.205810546875,
0.1993408203125,
0.205810546875,
0.1993408203125,
0.2171630859375,
0.16015625,
0.2171630859375
],
[
0.375732421875,
0.2183837890625,
0.410888671875,
0.2183837890625,
0.410888671875,
0.2310791015625,
0.375732421875,
0.2310791015625
],
[
0.61767578125,
0.25,
0.67236328125,
0.25,
0.67236328125,
0.260009765625,
0.61767578125,
0.260009765625
],
[
0.7548828125,
0.2474365234375,
0.89892578125,
0.250732421875,
0.89892578125,
0.26513671875,
0.7548828125,
0.261962890625
],
[
0.69775390625,
0.251953125,
0.73388671875,
0.251953125,
0.73388671875,
0.26318359375,
0.69775390625,
0.26318359375
],
[
0.53515625,
0.251953125,
0.61767578125,
0.251953125,
0.61767578125,
0.26318359375,
0.53515625,
0.26318359375
],
[
0.0882568359375,
0.272705078125,
0.10296630859375,
0.272705078125,
0.10296630859375,
0.281005859375,
0.0882568359375,
0.281005859375
],
[
0.101318359375,
0.274658203125,
0.11358642578125,
0.274658203125,
0.11358642578125,
0.279052734375,
0.101318359375,
0.279052734375
],
[
0.1168212890625,
0.271484375,
0.2607421875,
0.271484375,
0.2607421875,
0.2822265625,
0.1168212890625,
0.2822265625
],
[
0.0882568359375,
0.28857421875,
0.48046875,
0.28857421875,
0.48046875,
0.3017578125,
0.0882568359375,
0.3017578125
],
[
0.60888671875,
0.28857421875,
0.8203125,
0.28857421875,
0.8203125,
0.3017578125,
0.60888671875,
0.3017578125
],
[
0.08660888671875,
0.3017578125,
0.4794921875,
0.3017578125,
0.4794921875,
0.31494140625,
0.08660888671875,
0.31494140625
],
[
0.08660888671875,
0.31640625,
0.48046875,
0.31640625,
0.48046875,
0.330078125,
0.08660888671875,
0.330078125
],
[
0.08660888671875,
0.328857421875,
0.48046875,
0.330078125,
0.48046875,
0.343505859375,
0.08660888671875,
0.34228515625
],
[
0.5498046875,
0.335205078125,
0.9111328125,
0.335205078125,
0.9111328125,
0.34521484375,
0.5498046875,
0.34521484375
],
[
0.08905029296875,
0.34521484375,
0.468017578125,
0.34521484375,
0.468017578125,
0.35595703125,
0.08905029296875,
0.35595703125
],
[
0.564453125,
0.347900390625,
0.91259765625,
0.347900390625,
0.91259765625,
0.361083984375,
0.564453125,
0.361083984375
],
[
0.0882568359375,
0.35791015625,
0.48046875,
0.35791015625,
0.48046875,
0.371826171875,
0.0882568359375,
0.371826171875
],
[
0.56396484375,
0.35986328125,
0.91259765625,
0.3623046875,
0.91259765625,
0.375732421875,
0.56396484375,
0.373779296875
],
[
0.08660888671875,
0.371826171875,
0.48046875,
0.371826171875,
0.48046875,
0.385009765625,
0.08660888671875,
0.385009765625
],
[
0.56396484375,
0.375732421875,
0.8896484375,
0.375732421875,
0.8896484375,
0.389404296875,
0.56396484375,
0.389404296875
],
[
0.101318359375,
0.385009765625,
0.48046875,
0.385009765625,
0.48046875,
0.3984375,
0.101318359375,
0.3984375
],
[
0.54833984375,
0.389404296875,
0.9140625,
0.389404296875,
0.9140625,
0.40283203125,
0.54833984375,
0.40283203125
],
[
0.08660888671875,
0.399658203125,
0.4794921875,
0.399658203125,
0.4794921875,
0.41357421875,
0.08660888671875,
0.41357421875
],
[
0.56396484375,
0.40283203125,
0.91259765625,
0.40283203125,
0.91259765625,
0.416748046875,
0.56396484375,
0.416748046875
],
[
0.0849609375,
0.412353515625,
0.404296875,
0.41357421875,
0.404296875,
0.4267578125,
0.0849609375,
0.425537109375
],
[
0.564453125,
0.417236328125,
0.91259765625,
0.417236328125,
0.91259765625,
0.43115234375,
0.564453125,
0.43115234375
],
[
0.10455322265625,
0.4287109375,
0.48046875,
0.4287109375,
0.48046875,
0.439453125,
0.10455322265625,
0.439453125
],
[
0.564453125,
0.43115234375,
0.9208984375,
0.43115234375,
0.9208984375,
0.4443359375,
0.564453125,
0.4443359375
],
[
0.0882568359375,
0.43994140625,
0.48046875,
0.43994140625,
0.48046875,
0.453857421875,
0.0882568359375,
0.453857421875
],
[
0.56396484375,
0.443115234375,
0.76708984375,
0.4443359375,
0.76708984375,
0.458251953125,
0.56396484375,
0.45703125
],
[
0.08905029296875,
0.455810546875,
0.47802734375,
0.455810546875,
0.47802734375,
0.466552734375,
0.08905029296875,
0.466552734375
],
[
0.54736328125,
0.45703125,
0.91259765625,
0.458251953125,
0.91259765625,
0.4716796875,
0.54736328125,
0.47021484375
],
[
0.0849609375,
0.468505859375,
0.4794921875,
0.468505859375,
0.4794921875,
0.481689453125,
0.0849609375,
0.481689453125
],
[
0.564453125,
0.473388671875,
0.91259765625,
0.473388671875,
0.91259765625,
0.484130859375,
0.564453125,
0.484130859375
],
[
0.08660888671875,
0.481689453125,
0.48046875,
0.48291015625,
0.48046875,
0.496826171875,
0.08660888671875,
0.49560546875
],
[
0.564453125,
0.486083984375,
0.91259765625,
0.486083984375,
0.91259765625,
0.5,
0.564453125,
0.5
],
[
0.08660888671875,
0.496826171875,
0.4794921875,
0.496826171875,
0.4794921875,
0.51025390625,
0.08660888671875,
0.51025390625
],
[
0.56640625,
0.501953125,
0.9111328125,
0.501953125,
0.9111328125,
0.51220703125,
0.56640625,
0.51220703125
],
[
0.08660888671875,
0.51025390625,
0.48046875,
0.51025390625,
0.48046875,
0.5234375,
0.08660888671875,
0.5234375
],
[
0.564453125,
0.51318359375,
0.91259765625,
0.51318359375,
0.91259765625,
0.5263671875,
0.564453125,
0.5263671875
],
[
0.0882568359375,
0.52587890625,
0.48046875,
0.52587890625,
0.48046875,
0.5361328125,
0.0882568359375,
0.5361328125
],
[
0.56640625,
0.52880859375,
0.9111328125,
0.52880859375,
0.9111328125,
0.5390625,
0.56640625,
0.5390625
],
[
0.08660888671875,
0.53857421875,
0.48046875,
0.53857421875,
0.48046875,
0.5517578125,
0.08660888671875,
0.5517578125
],
[
0.564453125,
0.54150390625,
0.91259765625,
0.54150390625,
0.91259765625,
0.5546875,
0.564453125,
0.5546875
],
[
0.08660888671875,
0.55029296875,
0.48046875,
0.55029296875,
0.48046875,
0.564453125,
0.08660888671875,
0.564453125
],
[
0.564453125,
0.5546875,
0.6748046875,
0.55615234375,
0.6748046875,
0.5673828125,
0.564453125,
0.56640625
],
[
0.08660888671875,
0.56640625,
0.390625,
0.56640625,
0.390625,
0.5771484375,
0.08660888671875,
0.5771484375
],
[
0.53515625,
0.57373046875,
0.91259765625,
0.57373046875,
0.91259765625,
0.5869140625,
0.53515625,
0.5869140625
],
[
0.10296630859375,
0.580078125,
0.48193359375,
0.580078125,
0.48193359375,
0.59326171875,
0.10296630859375,
0.59326171875
],
[
0.51708984375,
0.5859375,
0.9140625,
0.5869140625,
0.9140625,
0.60107421875,
0.51708984375,
0.599609375
],
[
0.0882568359375,
0.59326171875,
0.48046875,
0.59326171875,
0.48046875,
0.60693359375,
0.0882568359375,
0.60693359375
],
[
0.08660888671875,
0.60693359375,
0.48046875,
0.60693359375,
0.48046875,
0.62060546875,
0.08660888671875,
0.62060546875
],
[
0.51953125,
0.60302734375,
0.9111328125,
0.60302734375,
0.9111328125,
0.61279296875,
0.51953125,
0.61279296875
],
[
0.08660888671875,
0.62060546875,
0.48046875,
0.62060546875,
0.48046875,
0.6337890625,
0.08660888671875,
0.6337890625
],
[
0.51953125,
0.6162109375,
0.9111328125,
0.6162109375,
0.9111328125,
0.626953125,
0.51953125,
0.626953125
],
[
0.0882568359375,
0.6337890625,
0.48046875,
0.63330078125,
0.48046875,
0.64794921875,
0.0882568359375,
0.6484375
],
[
0.5185546875,
0.62890625,
0.91259765625,
0.62890625,
0.91259765625,
0.642578125,
0.5185546875,
0.642578125
],
[
0.0882568359375,
0.6484375,
0.48046875,
0.6484375,
0.48046875,
0.662109375,
0.0882568359375,
0.662109375
],
[
0.51953125,
0.642578125,
0.91259765625,
0.642578125,
0.91259765625,
0.65576171875,
0.51953125,
0.65576171875
],
[
0.51708984375,
0.65478515625,
0.91259765625,
0.65576171875,
0.91259765625,
0.67041015625,
0.51708984375,
0.66943359375
],
[
0.0882568359375,
0.662109375,
0.48193359375,
0.662109375,
0.48193359375,
0.67529296875,
0.0882568359375,
0.67529296875
],
[
0.5185546875,
0.67041015625,
0.7509765625,
0.67041015625,
0.7509765625,
0.6845703125,
0.5185546875,
0.6845703125
],
[
0.0882568359375,
0.67822265625,
0.2197265625,
0.67822265625,
0.2197265625,
0.68798828125,
0.0882568359375,
0.68798828125
],
[
0.51708984375,
0.69970703125,
0.5458984375,
0.69970703125,
0.5458984375,
0.7109375,
0.51708984375,
0.7109375
],
[
0.5458984375,
0.69873046875,
0.689453125,
0.69873046875,
0.689453125,
0.7119140625,
0.5458984375,
0.7119140625
],
[
0.08660888671875,
0.70849609375,
0.1004638671875,
0.70849609375,
0.1004638671875,
0.71826171875,
0.08660888671875,
0.71826171875
],
[
0.11767578125,
0.70703125,
0.1993408203125,
0.70703125,
0.1993408203125,
0.71826171875,
0.11767578125,
0.71826171875
],
[
0.5185546875,
0.71728515625,
0.9140625,
0.71728515625,
0.9140625,
0.73095703125,
0.5185546875,
0.73095703125
],
[
0.0882568359375,
0.72607421875,
0.4794921875,
0.72607421875,
0.4794921875,
0.736328125,
0.0882568359375,
0.736328125
],
[
0.51708984375,
0.73095703125,
0.9140625,
0.73095703125,
0.9140625,
0.744140625,
0.51708984375,
0.744140625
],
[
0.08660888671875,
0.73779296875,
0.4794921875,
0.73779296875,
0.4794921875,
0.751953125,
0.08660888671875,
0.751953125
],
[
0.5185546875,
0.744140625,
0.91259765625,
0.744140625,
0.91259765625,
0.75830078125,
0.5185546875,
0.75830078125
],
[
0.08660888671875,
0.751953125,
0.338134765625,
0.751953125,
0.338134765625,
0.76513671875,
0.08660888671875,
0.76513671875
],
[
0.51953125,
0.75830078125,
0.9140625,
0.75830078125,
0.9140625,
0.771484375,
0.51953125,
0.771484375
],
[
0.10296630859375,
0.7666015625,
0.48046875,
0.7666015625,
0.48046875,
0.77978515625,
0.10296630859375,
0.77978515625
],
[
0.51953125,
0.77294921875,
0.8505859375,
0.77294921875,
0.8505859375,
0.78271484375,
0.51953125,
0.78271484375
],
[
0.0882568359375,
0.77978515625,
0.48046875,
0.77978515625,
0.48046875,
0.79345703125,
0.0882568359375,
0.79345703125
],
[
0.5361328125,
0.787109375,
0.9111328125,
0.787109375,
0.9111328125,
0.79736328125,
0.5361328125,
0.79736328125
],
[
0.08660888671875,
0.79345703125,
0.48046875,
0.79345703125,
0.48046875,
0.806640625,
0.08660888671875,
0.806640625
],
[
0.51708984375,
0.798828125,
0.9111328125,
0.7998046875,
0.9111328125,
0.81298828125,
0.51708984375,
0.8125
],
[
0.08905029296875,
0.8095703125,
0.101318359375,
0.8095703125,
0.101318359375,
0.8193359375,
0.08905029296875,
0.8193359375
],
[
0.5185546875,
0.81298828125,
0.91259765625,
0.81298828125,
0.91259765625,
0.8271484375,
0.5185546875,
0.8271484375
],
[
0.10455322265625,
0.8212890625,
0.3447265625,
0.8212890625,
0.3447265625,
0.83544921875,
0.10455322265625,
0.83544921875
],
[
0.5185546875,
0.82763671875,
0.9140625,
0.82763671875,
0.9140625,
0.84130859375,
0.5185546875,
0.84130859375
],
[
0.1168212890625,
0.84033203125,
0.48046875,
0.84130859375,
0.48046875,
0.85595703125,
0.1168212890625,
0.85498046875
],
[
0.51953125,
0.84130859375,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.51953125,
0.85498046875
],
[
0.13232421875,
0.85693359375,
0.48046875,
0.85693359375,
0.48046875,
0.8671875,
0.13232421875,
0.8671875
],
[
0.5185546875,
0.85595703125,
0.9140625,
0.85595703125,
0.9140625,
0.869140625,
0.5185546875,
0.869140625
],
[
0.13232421875,
0.869140625,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.13232421875,
0.88330078125
],
[
0.521484375,
0.87060546875,
0.9111328125,
0.87060546875,
0.9111328125,
0.880859375,
0.521484375,
0.880859375
],
[
0.13232421875,
0.8837890625,
0.450927734375,
0.8837890625,
0.450927734375,
0.89794921875,
0.13232421875,
0.89794921875
],
[
0.5185546875,
0.88330078125,
0.91259765625,
0.88330078125,
0.91259765625,
0.896484375,
0.5185546875,
0.896484375
]
] |
[
"Chosh, et al",
"Table 1: Statistics of CRED. Avg. RT denotes the average",
"number of retweets.",
"MoreRetweets (1",
"aeo",
"Less Retweets (0)",
"4",
"Classifier",
"Topic",
"Avg.RT",
"Pairs",
"Business & Entrepreneurs",
"743.2",
"4,186",
"Fitness & Health",
"1010.2",
"628",
"FLANG-ROBERTa",
"Learning &Educational",
"485.1",
"90",
"728.0",
"6",
"4",
"Total",
"Sports",
"4,910",
"Prompt-1",
"Prompt-2input Tweet (t2)",
"Claude",
"Input Tweet (t1)",
"2",
"",
"RELATEDWORK",
"自平台成立以来，最大化受众范围和反应的愿望一直是社交",
"Figure 2: Architecture of GGEA.",
"媒体用户的一致主题。几项研究的重点是理解和预测导致社",
"交媒体帖子广泛传播的因素。Cha etal.[5]对5200万Twitter",
"用户的活动进行了分析，得出一条推文的内容是推动转发的",
"Theretweet counts of tl and t2differedbyatleast 1o%",
"关键因素。Suhetal.[13]评估了7400万条推文，并探索了",
"of the less re-tweeted tweet. This condition ensures a clear",
"系列基于内容和上下文的功能，这些功能会影响给定推文的",
"distinction in the crowd reaction between the two tweets,",
"转发次数。他们开发了一个基于广义线性模型的模型来预测",
"enabling a more accurate and meaningful comparison",
"条推文预计会产生的转发次数。他们的研究强调了推文内",
"t1 and t2 were posted within a gap of 10 days and the dif-",
"容、网址和标签作为预测转发数量的关键特征的重要性。利用",
"ference between time of posting t1 and t2 was less than or",
"社交网络进行级联预测的其他工作包括[14]和[9]",
"equal to five hours. A short interval between posts helps",
"同样，Bakshyetal.[2]跟踪了1的扩散活动。600万Twitter",
"control the extraneous variables such as time-specific events",
"用户，通过研究不同类别的影响者之间的关系和相关成本。除",
"that might impact retweet counts",
"了Twitter之外，还对Facebook帖子的覆盖范围进行了研究",
". We utilized the Twitter Topic classification model [1] to",
"Bernsteinetal.[3]研究了222,000名Facebook用户的活动，揭",
"classifytweets intotopics.Weconsidered onlythosetweet",
"示了无形的受众如何增强Facebook帖子的影响力。Valkonen",
"pairs where the model assigned the same topic to both the",
"etal.[16]调查了芬兰消费者发布Facebook帖子的时间如何影",
"tweetswith a probability ofo.8 orhigherThishighprob",
"响可达性。Tanetal.[14]分析了推文内容对其影响范围的影响。",
"ability threshold ensures that the topics of the tweets are",
"最近的研究；如Gabrieletal.[8】展示了如何使用基于GPT-2",
"clearly defined and comparable,which is critical for the",
"[10]和T5[11]的模型来估计新闻文章的范围。但是，他们的",
"meaningful analysis of crowd reactions to different tweets",
"方法仅提供粗粒度的分析，对进行比较分析没有显着贡献。",
"on the same topic.",
"他们通过注释从1到5的到达级别来创建数据集",
"表E1显示了我们建议的人群反应估计数据集（CRED）的",
"我们的工作在几个重要方面与之前的研究不同。首先，我",
"统计信息，并且建议的CRED中的标签分布很平衡。不同主题",
"们引人了一个新的数据集CRED，它专门针对政府背景下社交",
"媒体管理者的决策过程。其次，我们提出了一种新颖的方法",
"的平均转发次数强调了数据集构建过程中主题控制的重要性。",
"GGEA，该方法利用LLM的分析能力来预测人群对社交媒体帖",
"主题方面的统计数据进一步强调，大多数政府推文与商业和",
"子的反应。最后，与以前的作品不同，我们使用基于T5的释",
"企业家精神有关。对于数据集分离，我们使用来自初始1的推",
"义器，通过GGEA使用ChatGPT响应[17]进行优化，以生成",
"文。5年，我。e。，到2022年4月，这构成了4304个用于培训",
"的推文对。最后6个月，我。e。，2022年5月至2022年10月，",
"和评估给定帖子的释义，从而允许用户通过探索其内容的不",
"提供606条用于模型验证的tweet对",
"同版本来优化参与度",
"4",
"METHODOLOGY",
"3",
"DATASET",
"我们在图2中介绍了GGEA的整体架构。该架构结合了FLANG-",
"我们的主要任务是确定一条表示为t1的推文是否会比另一条",
"RoBERTA[12]模型，并通过交叉编码器变压器架构进行了微",
"表示为t2的推文收到更多的转发。给定一对推文，我们的目",
"调。输人是来自给定对（t1，t2）的文本的串联，以及通过提",
"标是准确预测具有较高转发计数的推文",
"示Claude阐明为什么该对中的特定推文更具吸引力而获得的",
"我们从2020年11月至2022年10月白宫官方Twitter句柄",
"响应。预期的输出是t1是否会收到比t2更多的转发",
"发布的所有推文编译我们的数据集。然后，我们继续整理成对",
"为了增加推理过程中的可用性，我们采用了现有的基于开",
"数据，确保每对数据都符合与时间、主题和人群反应相关的条",
"源T5[11]的释义器[17]，该释义器在ChatGPT生成的释义上",
"件",
"进行了训练。这个释义器用于为给定的社交媒体帖子生成替代",
"我们仅保留满足以下四个条件的实例：",
"方案。我们一次检查两个释义。在我们开发的框架的帮助下，",
". The tweets were posted during weekdays: Weekdays gen",
"从包含释义和原始帖子的候选列表中选择预测会获得最大反",
"erallywitness higherlevels of social media activity,mak",
"应的文本。对于克劳德部分，采用两种类型的提示。首先，我",
"ing them ideal for capturing more interactions and thus",
"们询问LLM第一条推文是否会比第二条推文吸引更多的反应。",
"providing a more reliable measure of crowd reaction.",
"其次，我们一次用一条推文提示LLM，查询它以解释使给定推"
] |
{
"arxiv_id": "2403.09702",
"language": "zh",
"page_idx": 1
}
|
||
2403.09702-zh-2
|
2403.09702-zh-2.jpg
|
Generator-Guided Crowd Reaction Assessment
Table 2: Experimental results.
<table>
<tbody>
<tr>
<th>Setup</th>
<th>Model</th>
<th>$\operatorname{Accuracy}$</th>
<th>F1</th>
</tr>
<tr>
<td rowspan="1">Classifier</td>
<td>FLANG-RoBERTa</td>
<td>50.0 $\%$</td>
<td>66.7 $\%$</td>
</tr>
<tr>
<td rowspan="3">Zero- Shot</td>
<td>FLAN-UL2</td>
<td>50.2 $\%$</td>
<td>$2.6\%$</td>
</tr>
<tr>
<td>Claude</td>
<td>$51.0\%$</td>
<td>$9.2\%$ </td>
</tr>
<tr>
<td>React $tT5$</td>
<td>52.3 $\%$</td>
<td>$30.0\%$</td>
</tr>
<tr>
<td rowspan="3">GGEA</td>
<td>GGEA (ChatGPT)</td>
<td>67.2 $\%$</td>
<td>$68.1\%$</td>
</tr>
<tr>
<td>GGEA (FLAN-UL2)</td>
<td>69.6 $\%$</td>
<td>70.1 $\%$ </td>
</tr>
<tr>
<td>GGEA $\mathcal{A} \left ( \mathrm{Claude}\right )$</td>
<td>$71.9\%$</td>
<td>73.1 $\mathbb{T}_{e}$</td>
</tr>
</tbody>
</table>
Table 3: Analysis of changing classifiers in GGEA (Claude).
<table>
<tbody>
<tr>
<th> </th>
<th>$\operatorname{Accuracy}$</th>
<th>$F1$</th>
</tr>
<tr>
<td>GGEA (FLANG-BERT)</td>
<td>62.7 $\%$</td>
<td>62.1 $\%$</td>
</tr>
<tr>
<td>GGEA (RoBERTa)</td>
<td>69.8 $\%$</td>
<td>$69.5\%$</td>
</tr>
<tr>
<td>GGEA (FLANG-RoBERTa)</td>
<td>$71.9\%$</td>
<td>73.1 $\eta_{c}$</td>
</tr>
</tbody>
</table>
文引人人胜的原因。这些响应要么与推文连接，要么被视为独
立输入
5 EXPERIMENT
# 5.1 Experimental Settings
我们的实验设置将结果与两类基线进行比较：监督模型和零拍LLM。第一个启发式基线微调 FLANG-RoBERTa [12] 进行直接预测。为了评估为建议的任务部署零触发 LLM 的可行性，我们让 FLAN-UL2[15,18] , Claude 和 ReactT5$\boxed{\mathbb{E}}[8]$进行预测。更进一步，我们征求各种 LLM 对$t$1 和$t$2 的分析，以确定哪个LLM 可以指导分类模型提供增强的性能。我们的模型的性能使用两个指标来衡量：准确性和 F1 分数
# 5.2 $\textbf{Experimental Results}$
表 2 显示了实验结果。初步观察表明，分类模型和 LLM 都难以在拟议的任务中脱颖而出。然而，值得注意的是，无论咨询的法学硕士如何，建议的 GGEA 都能实现卓越的性能。我们观察到 GGEA(Claude)i 之间的性能差异。e。LLM 增强设置和 ReactT5 ( 零拍设置中性能最好的模型) 非常重要 ( p 值<。05)。这一结果强调了 LLM 和分类模型之间协作的重要性，表明 LLM 生成的分析成功地增加了模型做出决策所需的信息
认识到不同的分类器可能会影响性能，因此我们在 GGEA (Claude) 设置下尝试了不同的分类器。表 3 显示了实验结果。这些结果表明，FLANG-RoBERTa 优于其他测试分类器
最后，表$\mathbb{E}$4 中描述了按主题进行的评估。我们观察到，学习 &教育和体育主题的实例数量有限。因此，我们比较了其余两个主题，并得出结论，GGEA 在商业 &企业家主题中优于Fitness&Health, 这主要是由于商业 &企业家主题的可用培训实例数量更多
# 6 LIMITATIONS
这篇论文展示了有希望的结果；然而，承认它的局限性对于全
面理解我们的方法和指导未来的研究是至关重要的
首先，该研究主要利用来自单一来源的数据，即白宫的官方 Twitter 帐户。这种关注可能会导致模型过度适应此帐户的独特风格，语气和内容特征，从而可能降低其对其他社交媒体
Table 4: Topic-wise evaluation. We didn't have any instances
for the Sports category in the test set.
<table>
<tbody>
<tr>
<th>Topic</th>
<th>Instance 井 $F$</th>
<th>$\operatorname{Accuracy}$</th>
<th>F1</th>
</tr>
<tr>
<td>Business& Entrepreneurs</td>
<td>570</td>
<td>73.0 $\%$</td>
<td>74.2 $\%$</td>
</tr>
<tr>
<td>Fitness& Health</td>
<td>34</td>
<td>$52.9\%$</td>
<td>50.0 $\%$</td>
</tr>
<tr>
<td>LearningEducational</td>
<td>2</td>
<td>100.0 $\%$</td>
<td>100.0 $\%$</td>
</tr>
</tbody>
</table>
帐户，主题或平台的适用性。我们对转发的偏好超过了其他参与度指标，例如喜欢或引用计数，其基础是理解转发会显著传播信息[7]。选择白宫的 Twitter 帐户是由于其真实性，公正性和对公众舆论的影响，以及其广泛的影响力。然而，这种选择引人了一种偏见，因为参与度(e。g。,转发次数)在追随者数量不同的账户之间不能直接比较
其次，许多因素会影响社交媒体的参与度，包括发布时间， 社交环境，先前帖子的影响，视觉内容和受众群体。我们目前的模型中没有考虑这些方面，这可能会影响其预测的准确性
第三，我们的研究仅限于分析转发作为 Twitter 上的主要参与指标。虽然转发是一个关键的参与度指标，但其他元素，如点赞、评论和分享，也有助于帖子的整体影响。此外，不同主题的推文由于其独特的内容而具有内在的可比性。为了保证数据质量，我们使用了一个高置信度阈值为 0 的 tweet 主题分类器$\left[1\right]$。8,根据经验确定。扩展我们的模型以包括这些额外的参与度指标将提供对职位的影响力和影响力的更全面的了解
7 CONCLUSION AND FUTURE DIRECTIONS
这项研究提供了一个预测人群对社交媒体帖子反应的新视角， 引人了一个新的数据集和一种方法。我们策划了人群反应估计数据集 (CRED),该数据集通过比较转发指标的视角封装了政府社交媒体管理者的决策过程。我们还提出了生成器引导估计方法 (GGEA), 该方法在具有分类模型的合作框架中利用LLM 的分析能力。我们的实验表明，GGEA, 特别是在结合克劳德的分析时，优于仅仅微调分类模型，并在预测人群反应方面表现出卓越的性能。这些结果突出了 LLM 通过提供细微的分析和进一步的信息来增强分类模型的潜力，从而增强了决策过程
展望未来，有几个有趣的途径可以进一步研究。首先，GGEA 可以扩展到与其他形式的社交媒体合作，例如 Facebook 帖子或 Instagram 字幕，从而可以更全面地了解跨平台的社交媒体参与度。其次，GGEA 可以适应处理多类问题，例如预测帖子是否会传播病毒，而不仅仅是关注二元结果。此外，引人更复杂的释义方法可能会提高 GGEA 的性能。最后，对 GGEA 框架中不同法学硕士表现的因素进行深人探索可能会产生有价值的见解
# $\mathbf{ACKNOWLEDGMENTS}$
The work of Chung-Chi Chen was supported in part by JSPS KAKENHI Grant Number 23K16956 and a project JPNP20006, commissioned by the New Energy and Industrial Technology Development Organization (NEDO).
# REFERENCES
[1] Dimosthenis Antypas, Asahi Ushio, Jose Camacho-Collados, Vitor Silva,
Leonardo Neves, and Francesco Barbieri. 2022. Twitter Topic Classification. In Proceedings of the 29th International Conference on Computational Linguistics International Committee on Computational Linguistics, Gyeongju, Korea
|
[
[
0.0882568359375,
0.07574462890625,
0.30810546875,
0.07574462890625,
0.30810546875,
0.08587646484375,
0.0882568359375,
0.08587646484375
],
[
0.1846923828125,
0.10479736328125,
0.381591796875,
0.10479736328125,
0.381591796875,
0.11871337890625,
0.1846923828125,
0.11871337890625
],
[
0.5185546875,
0.10479736328125,
0.91259765625,
0.10479736328125,
0.91259765625,
0.11871337890625,
0.5185546875,
0.11871337890625
],
[
0.51708984375,
0.11932373046875,
0.76708984375,
0.11932373046875,
0.76708984375,
0.1331787109375,
0.51708984375,
0.1331787109375
],
[
0.1356201171875,
0.1383056640625,
0.1748046875,
0.14013671875,
0.1732177734375,
0.1546630859375,
0.134033203125,
0.15283203125
],
[
0.330078125,
0.1383056640625,
0.39208984375,
0.14013671875,
0.39208984375,
0.1546630859375,
0.330078125,
0.15283203125
],
[
0.41259765625,
0.1395263671875,
0.4345703125,
0.1395263671875,
0.4345703125,
0.15283203125,
0.41259765625,
0.15283203125
],
[
0.5908203125,
0.147705078125,
0.63134765625,
0.147705078125,
0.63134765625,
0.1622314453125,
0.5908203125,
0.1622314453125
],
[
0.238525390625,
0.14013671875,
0.281005859375,
0.14013671875,
0.281005859375,
0.1514892578125,
0.238525390625,
0.1514892578125
],
[
0.70166015625,
0.1483154296875,
0.76953125,
0.1483154296875,
0.76953125,
0.160400390625,
0.70166015625,
0.160400390625
],
[
0.779296875,
0.149658203125,
0.8388671875,
0.149658203125,
0.8388671875,
0.1610107421875,
0.779296875,
0.1610107421875
],
[
0.865234375,
0.1483154296875,
0.8857421875,
0.1483154296875,
0.8857421875,
0.1610107421875,
0.865234375,
0.1610107421875
],
[
0.1968994140625,
0.1514892578125,
0.310546875,
0.15283203125,
0.310546875,
0.1673583984375,
0.1968994140625,
0.1666259765625
],
[
0.12091064453125,
0.1514892578125,
0.1854248046875,
0.1546630859375,
0.1846923828125,
0.169189453125,
0.1192626953125,
0.1666259765625
],
[
0.351318359375,
0.1546630859375,
0.39208984375,
0.1546630859375,
0.39208984375,
0.1666259765625,
0.351318359375,
0.1666259765625
],
[
0.40185546875,
0.154052734375,
0.44775390625,
0.154052734375,
0.44775390625,
0.1685791015625,
0.40185546875,
0.1685791015625
],
[
0.52783203125,
0.1610107421875,
0.69287109375,
0.1622314453125,
0.69287109375,
0.1767578125,
0.52783203125,
0.175537109375
],
[
0.7998046875,
0.1610107421875,
0.8408203125,
0.1634521484375,
0.8388671875,
0.175537109375,
0.79833984375,
0.173583984375
],
[
0.85693359375,
0.1622314453125,
0.90283203125,
0.1622314453125,
0.90283203125,
0.1767578125,
0.85693359375,
0.1767578125
],
[
0.72216796875,
0.1634521484375,
0.74853515625,
0.1634521484375,
0.74853515625,
0.175537109375,
0.72216796875,
0.175537109375
],
[
0.19775390625,
0.1654052734375,
0.268798828125,
0.1673583984375,
0.268798828125,
0.1817626953125,
0.19775390625,
0.179931640625
],
[
0.351318359375,
0.169189453125,
0.390625,
0.169189453125,
0.390625,
0.18115234375,
0.351318359375,
0.18115234375
],
[
0.410888671875,
0.169189453125,
0.4453125,
0.169189453125,
0.4453125,
0.18115234375,
0.410888671875,
0.18115234375
],
[
0.79736328125,
0.175537109375,
0.84228515625,
0.175537109375,
0.84228515625,
0.1900634765625,
0.79736328125,
0.1900634765625
],
[
0.85888671875,
0.175537109375,
0.90185546875,
0.175537109375,
0.90185546875,
0.1900634765625,
0.85888671875,
0.1900634765625
],
[
0.7255859375,
0.1767578125,
0.74609375,
0.1767578125,
0.74609375,
0.1900634765625,
0.7255859375,
0.1900634765625
],
[
0.53173828125,
0.1767578125,
0.63232421875,
0.1767578125,
0.63232421875,
0.1881103515625,
0.53173828125,
0.1881103515625
],
[
0.1993408203125,
0.18310546875,
0.2442626953125,
0.18310546875,
0.2442626953125,
0.1944580078125,
0.1993408203125,
0.1944580078125
],
[
0.351318359375,
0.18310546875,
0.39208984375,
0.18310546875,
0.39208984375,
0.1944580078125,
0.351318359375,
0.1944580078125
],
[
0.41015625,
0.18310546875,
0.443603515625,
0.18310546875,
0.443603515625,
0.1956787109375,
0.41015625,
0.1956787109375
],
[
0.529296875,
0.189453125,
0.67724609375,
0.189453125,
0.67724609375,
0.20263671875,
0.529296875,
0.20263671875
],
[
0.72607421875,
0.1900634765625,
0.744140625,
0.1900634765625,
0.744140625,
0.203857421875,
0.72607421875,
0.203857421875
],
[
0.85107421875,
0.1900634765625,
0.904296875,
0.1900634765625,
0.904296875,
0.2052001953125,
0.85107421875,
0.2052001953125
],
[
0.79345703125,
0.1912841796875,
0.8408203125,
0.1912841796875,
0.8408203125,
0.20263671875,
0.79345703125,
0.20263671875
],
[
0.12335205078125,
0.186279296875,
0.188720703125,
0.186279296875,
0.188720703125,
0.1976318359375,
0.12335205078125,
0.1976318359375
],
[
0.19775390625,
0.1956787109375,
0.256591796875,
0.1956787109375,
0.256591796875,
0.210205078125,
0.19775390625,
0.210205078125
],
[
0.3505859375,
0.1956787109375,
0.393798828125,
0.1956787109375,
0.393798828125,
0.210205078125,
0.3505859375,
0.210205078125
],
[
0.40185546875,
0.1944580078125,
0.4453125,
0.1962890625,
0.443603515625,
0.2108154296875,
0.400390625,
0.208984375
],
[
0.1993408203125,
0.210205078125,
0.310546875,
0.210205078125,
0.310546875,
0.2235107421875,
0.1993408203125,
0.2235107421875
],
[
0.351318359375,
0.2108154296875,
0.39208984375,
0.2108154296875,
0.39208984375,
0.222900390625,
0.351318359375,
0.222900390625
],
[
0.40283203125,
0.2083740234375,
0.4453125,
0.210205078125,
0.443603515625,
0.2247314453125,
0.40185546875,
0.222900390625
],
[
0.1993408203125,
0.2235107421875,
0.318603515625,
0.2235107421875,
0.318603515625,
0.2374267578125,
0.1993408203125,
0.2374267578125
],
[
0.351318359375,
0.222900390625,
0.39208984375,
0.2247314453125,
0.390625,
0.2374267578125,
0.3505859375,
0.23486328125
],
[
0.40185546875,
0.2235107421875,
0.4453125,
0.2235107421875,
0.4453125,
0.238037109375,
0.40185546875,
0.238037109375
],
[
0.5185546875,
0.2342529296875,
0.91259765625,
0.2342529296875,
0.91259765625,
0.2474365234375,
0.5185546875,
0.2474365234375
],
[
0.1217041015625,
0.2266845703125,
0.1650390625,
0.2266845703125,
0.1650390625,
0.2412109375,
0.1217041015625,
0.2412109375
],
[
0.2017822265625,
0.2392578125,
0.294189453125,
0.2392578125,
0.294189453125,
0.250732421875,
0.2017822265625,
0.250732421875
],
[
0.351318359375,
0.2374267578125,
0.393798828125,
0.2392578125,
0.39208984375,
0.251953125,
0.3505859375,
0.25
],
[
0.40283203125,
0.2392578125,
0.4453125,
0.2392578125,
0.4453125,
0.250732421875,
0.40283203125,
0.250732421875
],
[
0.51953125,
0.2474365234375,
0.91259765625,
0.2474365234375,
0.91259765625,
0.261474609375,
0.51953125,
0.261474609375
],
[
0.0882568359375,
0.268310546875,
0.476318359375,
0.268310546875,
0.476318359375,
0.2822265625,
0.0882568359375,
0.2822265625
],
[
0.51953125,
0.26318359375,
0.91015625,
0.26318359375,
0.91015625,
0.2734375,
0.51953125,
0.2734375
],
[
0.5185546875,
0.27587890625,
0.91259765625,
0.27587890625,
0.91259765625,
0.2890625,
0.5185546875,
0.2890625
],
[
0.5185546875,
0.2890625,
0.91259765625,
0.2890625,
0.91259765625,
0.302978515625,
0.5185546875,
0.302978515625
],
[
0.316162109375,
0.302978515625,
0.373291015625,
0.302978515625,
0.373291015625,
0.314453125,
0.316162109375,
0.314453125
],
[
0.396240234375,
0.3017578125,
0.4150390625,
0.3017578125,
0.4150390625,
0.314453125,
0.396240234375,
0.314453125
],
[
0.51953125,
0.304931640625,
0.7236328125,
0.304931640625,
0.7236328125,
0.31494140625,
0.51953125,
0.31494140625
],
[
0.1380615234375,
0.314453125,
0.279296875,
0.31494140625,
0.279296875,
0.330078125,
0.1380615234375,
0.328857421875
],
[
0.3349609375,
0.317626953125,
0.374267578125,
0.317626953125,
0.374267578125,
0.328857421875,
0.3349609375,
0.328857421875
],
[
0.384033203125,
0.31494140625,
0.4306640625,
0.31494140625,
0.4306640625,
0.330078125,
0.384033203125,
0.330078125
],
[
0.53515625,
0.317626953125,
0.91259765625,
0.317626953125,
0.91259765625,
0.330810546875,
0.53515625,
0.330810546875
],
[
0.1405029296875,
0.330078125,
0.250732421875,
0.330078125,
0.250732421875,
0.343505859375,
0.1405029296875,
0.343505859375
],
[
0.3349609375,
0.330810546875,
0.374267578125,
0.330810546875,
0.374267578125,
0.34228515625,
0.3349609375,
0.34228515625
],
[
0.384033203125,
0.330078125,
0.428955078125,
0.330078125,
0.428955078125,
0.3447265625,
0.384033203125,
0.3447265625
],
[
0.5185546875,
0.330810546875,
0.91259765625,
0.330810546875,
0.91259765625,
0.3447265625,
0.5185546875,
0.3447265625
],
[
0.1396484375,
0.343505859375,
0.2998046875,
0.343505859375,
0.2998046875,
0.356689453125,
0.1396484375,
0.356689453125
],
[
0.33251953125,
0.343505859375,
0.375732421875,
0.343505859375,
0.375732421875,
0.35791015625,
0.33251953125,
0.35791015625
],
[
0.3857421875,
0.341552734375,
0.4306640625,
0.343505859375,
0.428955078125,
0.359130859375,
0.384033203125,
0.356689453125
],
[
0.5185546875,
0.3447265625,
0.900390625,
0.3447265625,
0.900390625,
0.35791015625,
0.5185546875,
0.35791015625
],
[
0.53369140625,
0.359130859375,
0.91259765625,
0.359130859375,
0.91259765625,
0.37255859375,
0.53369140625,
0.37255859375
],
[
0.51953125,
0.37255859375,
0.9140625,
0.37255859375,
0.9140625,
0.386474609375,
0.51953125,
0.386474609375
],
[
0.0882568359375,
0.3837890625,
0.48046875,
0.3837890625,
0.48046875,
0.397705078125,
0.0882568359375,
0.397705078125
],
[
0.51953125,
0.385009765625,
0.9140625,
0.385009765625,
0.9140625,
0.3984375,
0.51953125,
0.3984375
],
[
0.0849609375,
0.395751953125,
0.134033203125,
0.397705078125,
0.13232421875,
0.412353515625,
0.08416748046875,
0.410400390625
],
[
0.51708984375,
0.3984375,
0.9140625,
0.399658203125,
0.9140625,
0.41357421875,
0.51708984375,
0.412353515625
],
[
0.5185546875,
0.4140625,
0.91259765625,
0.4140625,
0.91259765625,
0.427978515625,
0.5185546875,
0.427978515625
],
[
0.0849609375,
0.4248046875,
0.1168212890625,
0.4248046875,
0.1168212890625,
0.436279296875,
0.0849609375,
0.436279296875
],
[
0.11517333984375,
0.423583984375,
0.23779296875,
0.423583984375,
0.23779296875,
0.4375,
0.11517333984375,
0.4375
],
[
0.5185546875,
0.427978515625,
0.91259765625,
0.427978515625,
0.91259765625,
0.44140625,
0.5185546875,
0.44140625
],
[
0.08416748046875,
0.442626953125,
0.31201171875,
0.443115234375,
0.31201171875,
0.458251953125,
0.08416748046875,
0.45703125
],
[
0.51708984375,
0.43994140625,
0.90185546875,
0.442626953125,
0.90185546875,
0.455810546875,
0.51708984375,
0.453857421875
],
[
0.0882568359375,
0.46337890625,
0.4794921875,
0.46337890625,
0.4794921875,
0.473388671875,
0.0882568359375,
0.473388671875
],
[
0.51953125,
0.4716796875,
0.5439453125,
0.4716796875,
0.5439453125,
0.479736328125,
0.51953125,
0.479736328125
],
[
0.54345703125,
0.472900390625,
0.54833984375,
0.472900390625,
0.54833984375,
0.477783203125,
0.54345703125,
0.477783203125
],
[
0.54833984375,
0.47021484375,
0.90185546875,
0.47021484375,
0.90185546875,
0.48095703125,
0.54833984375,
0.48095703125
],
[
0.08660888671875,
0.4892578125,
0.48193359375,
0.4892578125,
0.48193359375,
0.50244140625,
0.08660888671875,
0.50244140625
],
[
0.0849609375,
0.47607421875,
0.48046875,
0.47607421875,
0.48046875,
0.4892578125,
0.0849609375,
0.4892578125
],
[
0.51953125,
0.4892578125,
0.9111328125,
0.4892578125,
0.9111328125,
0.5,
0.51953125,
0.5
],
[
0.0882568359375,
0.5048828125,
0.4794921875,
0.5048828125,
0.4794921875,
0.51513671875,
0.0882568359375,
0.51513671875
],
[
0.5185546875,
0.501953125,
0.91259765625,
0.501953125,
0.91259765625,
0.51513671875,
0.5185546875,
0.51513671875
],
[
0.08660888671875,
0.517578125,
0.48046875,
0.517578125,
0.48046875,
0.53076171875,
0.08660888671875,
0.53076171875
],
[
0.5185546875,
0.51513671875,
0.9111328125,
0.51513671875,
0.9111328125,
0.52880859375,
0.5185546875,
0.52880859375
],
[
0.0849609375,
0.52978515625,
0.48193359375,
0.52978515625,
0.48193359375,
0.54345703125,
0.0849609375,
0.54345703125
],
[
0.51953125,
0.53076171875,
0.9111328125,
0.53076171875,
0.9111328125,
0.54150390625,
0.51953125,
0.54150390625
],
[
0.0849609375,
0.54345703125,
0.343017578125,
0.54345703125,
0.343017578125,
0.556640625,
0.0849609375,
0.556640625
],
[
0.51953125,
0.54443359375,
0.9111328125,
0.54443359375,
0.9111328125,
0.5546875,
0.51953125,
0.5546875
],
[
0.515625,
0.5546875,
0.91259765625,
0.556640625,
0.91259765625,
0.57080078125,
0.515625,
0.568359375
],
[
0.125,
0.57080078125,
0.30810546875,
0.5693359375,
0.30810546875,
0.583984375,
0.125,
0.58544921875
],
[
0.5185546875,
0.57080078125,
0.91259765625,
0.57080078125,
0.91259765625,
0.583984375,
0.5185546875,
0.583984375
],
[
0.5185546875,
0.583984375,
0.91259765625,
0.583984375,
0.91259765625,
0.59765625,
0.5185546875,
0.59765625
],
[
0.08660888671875,
0.57275390625,
0.13232421875,
0.57275390625,
0.13232421875,
0.58349609375,
0.08660888671875,
0.58349609375
],
[
0.0849609375,
0.5888671875,
0.48193359375,
0.5888671875,
0.48193359375,
0.60302734375,
0.0849609375,
0.60302734375
],
[
0.5185546875,
0.5986328125,
0.9140625,
0.5986328125,
0.9140625,
0.6123046875,
0.5185546875,
0.6123046875
],
[
0.0882568359375,
0.60498046875,
0.4794921875,
0.60498046875,
0.4794921875,
0.61572265625,
0.0882568359375,
0.61572265625
],
[
0.0882568359375,
0.61865234375,
0.48046875,
0.61865234375,
0.48046875,
0.62890625,
0.0882568359375,
0.62890625
],
[
0.51953125,
0.61279296875,
0.56396484375,
0.61279296875,
0.56396484375,
0.625,
0.51953125,
0.625
],
[
0.08660888671875,
0.630859375,
0.48193359375,
0.630859375,
0.48193359375,
0.64453125,
0.08660888671875,
0.64453125
],
[
0.53369140625,
0.62548828125,
0.9140625,
0.62548828125,
0.9140625,
0.6396484375,
0.53369140625,
0.6396484375
],
[
0.08660888671875,
0.64306640625,
0.48046875,
0.64453125,
0.48046875,
0.65771484375,
0.08660888671875,
0.6572265625
],
[
0.5185546875,
0.63818359375,
0.91259765625,
0.6396484375,
0.91259765625,
0.65380859375,
0.5185546875,
0.65283203125
],
[
0.08660888671875,
0.65771484375,
0.48193359375,
0.65771484375,
0.48193359375,
0.671875,
0.08660888671875,
0.671875
],
[
0.5185546875,
0.65380859375,
0.91259765625,
0.65380859375,
0.91259765625,
0.66748046875,
0.5185546875,
0.66748046875
],
[
0.08660888671875,
0.671875,
0.459228515625,
0.671875,
0.459228515625,
0.68505859375,
0.08660888671875,
0.68505859375
],
[
0.51708984375,
0.66650390625,
0.91259765625,
0.66748046875,
0.91259765625,
0.68115234375,
0.51708984375,
0.6796875
],
[
0.10540771484375,
0.6875,
0.48046875,
0.6875,
0.48046875,
0.69775390625,
0.10540771484375,
0.69775390625
],
[
0.521484375,
0.68310546875,
0.9111328125,
0.68310546875,
0.9111328125,
0.693359375,
0.521484375,
0.693359375
],
[
0.0849609375,
0.69873046875,
0.4794921875,
0.69970703125,
0.4794921875,
0.71337890625,
0.0849609375,
0.7119140625
],
[
0.5185546875,
0.69580078125,
0.91259765625,
0.69580078125,
0.91259765625,
0.708984375,
0.5185546875,
0.708984375
],
[
0.08660888671875,
0.71337890625,
0.431396484375,
0.71337890625,
0.431396484375,
0.7265625,
0.08660888671875,
0.7265625
],
[
0.5185546875,
0.708984375,
0.91259765625,
0.708984375,
0.91259765625,
0.72265625,
0.5185546875,
0.72265625
],
[
0.10296630859375,
0.72802734375,
0.48046875,
0.7265625,
0.48046875,
0.74072265625,
0.10296630859375,
0.7412109375
],
[
0.51953125,
0.7236328125,
0.57861328125,
0.7236328125,
0.57861328125,
0.73486328125,
0.51953125,
0.73486328125
],
[
0.0849609375,
0.74072265625,
0.48046875,
0.7412109375,
0.48046875,
0.7548828125,
0.0849609375,
0.75390625
],
[
0.0849609375,
0.7548828125,
0.48046875,
0.75390625,
0.48046875,
0.76806640625,
0.0849609375,
0.76953125
],
[
0.5185546875,
0.75048828125,
0.71337890625,
0.75048828125,
0.71337890625,
0.7646484375,
0.5185546875,
0.7646484375
],
[
0.08660888671875,
0.77099609375,
0.48046875,
0.77099609375,
0.48046875,
0.78076171875,
0.08660888671875,
0.78076171875
],
[
0.515625,
0.767578125,
0.91259765625,
0.76806640625,
0.91259765625,
0.78271484375,
0.515625,
0.7822265625
],
[
0.0849609375,
0.78271484375,
0.1732177734375,
0.78271484375,
0.1732177734375,
0.79443359375,
0.0849609375,
0.79443359375
],
[
0.515625,
0.78076171875,
0.91259765625,
0.7822265625,
0.91259765625,
0.796875,
0.515625,
0.79541015625
],
[
0.51708984375,
0.796875,
0.9140625,
0.796875,
0.9140625,
0.81103515625,
0.51708984375,
0.81103515625
],
[
0.08660888671875,
0.81103515625,
0.1168212890625,
0.81103515625,
0.1168212890625,
0.82275390625,
0.08660888671875,
0.82275390625
],
[
0.11517333984375,
0.8095703125,
0.236083984375,
0.8095703125,
0.236083984375,
0.82275390625,
0.11517333984375,
0.82275390625
],
[
0.5185546875,
0.8125,
0.685546875,
0.8125,
0.685546875,
0.82275390625,
0.5185546875,
0.82275390625
],
[
0.08416748046875,
0.8271484375,
0.48046875,
0.82763671875,
0.48046875,
0.84130859375,
0.08416748046875,
0.84033203125
],
[
0.08660888671875,
0.84130859375,
0.41259765625,
0.84130859375,
0.41259765625,
0.85498046875,
0.08660888671875,
0.85498046875
],
[
0.51953125,
0.8388671875,
0.63232421875,
0.8388671875,
0.63232421875,
0.849609375,
0.51953125,
0.849609375
],
[
0.10296630859375,
0.85498046875,
0.48046875,
0.85400390625,
0.48046875,
0.86865234375,
0.10296630859375,
0.869140625
],
[
0.521484375,
0.85302734375,
0.91259765625,
0.85302734375,
0.91259765625,
0.8662109375,
0.521484375,
0.8662109375
],
[
0.54345703125,
0.86279296875,
0.91259765625,
0.8642578125,
0.91259765625,
0.87548828125,
0.54345703125,
0.875
],
[
0.0882568359375,
0.87060546875,
0.4794921875,
0.87060546875,
0.4794921875,
0.880859375,
0.0882568359375,
0.880859375
],
[
0.54345703125,
0.87353515625,
0.9140625,
0.87353515625,
0.9140625,
0.88720703125,
0.54345703125,
0.88720703125
],
[
0.08660888671875,
0.88330078125,
0.4794921875,
0.88330078125,
0.4794921875,
0.896484375,
0.08660888671875,
0.896484375
],
[
0.54345703125,
0.8837890625,
0.8935546875,
0.88525390625,
0.8935546875,
0.896484375,
0.54345703125,
0.89599609375
]
] |
[
"Generator-Guided CrowdReactionAssessment",
"Table 2: Experimental results.",
"Table 4: Topic-wise evaluation. We didn't have any instances",
"for the Sports category in the test set.",
"Setup",
"Accuracy",
"F1",
"Topic",
"Model",
"#Instance",
"Accuracy",
"F1",
"FLANG-RoBERTa",
"Classifier",
"50.0 %",
"66.7 %",
"Business & Entrepreneurs",
"73.0 %",
"74.2%",
"570",
"FLAN-UL2",
"50.2 %",
"2.6 %",
"52.9 %",
"50.0 %",
"34",
"Fitness&Health",
"Claude",
"51.0 %",
"9.2%",
"Learning & Educational",
"2",
"100.0%",
"100.0%",
"Zero-Shot",
"React T5",
"52.3%",
"30.0%",
"GGEA (ChatGPT)",
"67.2%",
"68.1%",
"GGEA (FLAN-UL2)",
"69.6 %",
"70.1 %",
"帐户，主题或平台的适用性。我们对转发的偏好超过了其他参",
"GGEA",
"GGEA (Claude)",
"71.9%",
"73.1 %",
"与度指标，例如喜欢或引用计数，其基础是理解转发会显著传",
"Table 3: Analysis of changing classifiers in GGEA (Claude)",
"播信息「7]。选择白宫的Twitter帐户是由于其真实性，公正性",
"和对公众舆论的影响，以及其广泛的影响力。然而，这种选择",
"引人了一种偏见，因为参与度(e。g。，转发次数)在追随者数",
"Accuracy",
"F1",
"量不同的账户之间不能直接比较",
"GGEA (FLANG-BERT)",
"62.7 %",
"62.1%",
"其次，许多因素会影响社交媒体的参与度，包括发布时间，",
"GGEA (RoBERTa)",
"69.8 %",
"69.5 %",
"社交环境，先前帖子的影响，视觉内容和受众群体。我们目前",
"GGEA (FLANG-RoBERTa)",
"71.9%",
"73.1%",
"的模型中没有考虑这些方面，这可能会影响其预测的准确性",
"第三，我们的研究仅限于分析转发作为Twitter上的主要参",
"与指标。虽然转发是一个关键的参与度指标，但其他元素，如",
"文引人人胜的原因。这些响应要么与推文连接，要么被视为独",
"点赞、评论和分享，也有助于帖子的整体影响。此外，不同主",
"立输入",
"题的推文由于其独特的内容而具有内在的可比性。为了保证数",
"据质量，我们使用了一个高置信度阈值为0的tweet主题分类",
"5",
"EXPERIMENT",
"器[1]。8，根据经验确定。扩展我们的模型以包括这些额外的",
"5.1Experimental Settings",
"参与度指标将提供对职位的影响力和影响力的更全面的了解",
"我们的实验设置将结果与两类基线进行比较：监督模型和零拍",
"7",
"",
"CONCLUSIONANDFUTUREDIRECTIONS",
"预测。为了评估为建议的任务部署零触发LLM的可行性，我",
"LLM。第一个启发式基线微调FLANG-RoBERTa[12]进行直接",
"这项研究提供了一个预测人群对社交媒体帖子反应的新视角",
"们让FLAN-UL2[15，18]，Claude和ReactT5E[8|进行预测。更",
"引人了一个新的数据集和一种方法。我们策划了人群反应估",
"进一步，我们征求各种LLM对t1和t2的分析，以确定哪个",
"计数据集（CRED），该数据集通过比较转发指标的视角封装了",
"LLM可以指导分类模型提供增强的性能。我们的模型的性能",
"政府社交媒体管理者的决策过程。我们还提出了生成器引导",
"使用两个指标来衡量：准确性和F1分数",
"估计方法（GGEA），该方法在具有分类模型的合作框架中利用",
"LLM的分析能力。我们的实验表明，GGEA，特别是在结合克",
"Experimental Results",
"劳德的分析时，优于仅仅微调分类模型，并在预测人群反应方",
"面表现出卓越的性能。这些结果突出了LLM通过提供细微的",
"5.2",
"表2显示了实验结果。初步观察表明，分类模型和LLM都难",
"分析和进一步的信息来增强分类模型的潜力，从而增强了决",
"以在拟议的任务中脱颖而出。然而，值得注意的是，无论咨",
"询的法学硕士如何，建议的GGEA都能实现卓越的性能。我",
"策过程",
"们观察到GGEA(Claude)i之间的性能差异。e。LLM增强设置",
"展望未来，有几个有趣的途径可以进一步研究。首先，GGEA",
"和ReactT5（零拍设置中性能最好的模型）非常重要（p值<。",
"可以扩展到与其他形式的社交媒体合作，例如Facebook帖子",
"05)。这一结果强调了LLM和分类模型之间协作的重要性，表",
"或Instagram字幕，从而可以更全面地了解跨平台的社交媒体",
"明LLM生成的分析成功地增加了模型做出决策所需的信息",
"参与度。其次，GGEA可以适应处理多类问题，例如预测帖子",
"认识到不同的分类器可能会影响性能，因此我们在GGEA",
"是否会传播病毒，而不仅仅是关注二元结果。此外，引入更复",
"（Claude）设置下尝试了不同的分类器。表3显示了实验结果。",
"杂的释义方法可能会提高GGEA的性能。最后，对GGEA框",
"这些结果表明，FLANG-RoBERTa优于其他测试分类器",
"架中不同法学硕士表现的因素进行深人探索可能会产生有价",
"最后，表E4中描述了按主题进行的评估。我们观察到，学",
"值的见解",
"习&教育和体育主题的实例数量有限。因此，我们比较了其",
"余两个主题，并得出结论，GGEA在商业&企业家主题中优于",
"ACKNOWLEDGMENTS",
"Fitness&Health，这主要是由于商业&企业家主题的可用培训",
"The work of Chung-Chi Chen was supported in part by JSPS KAK",
"实例数量更多",
"ENHI Grant Number 23K16956 and a project JPNP20006, commis-",
"sioned by the New Energy and Industrial Technology Develop-",
"6",
"LIMITATIONS",
"ment Organization (NEDO)",
"这篇论文展示了有希望的结果；然而，承认它的局限性对于全",
"面理解我们的方法和指导未来的研究是至关重要的",
"REFERENCES",
"首先，该研究主要利用来自单一来源的数据，即白宫的官",
"[1] Dimosthenis Antypas, Asahi Ushio, Jose Camacho-Collados, Vitor silva,",
"Leonardo Neves, and Francesco Barbieri. 2022. Twitter Topic Classification. In",
"方Twitter帐户。这种关注可能会导致模型过度适应此帐户的",
"Proceedings of the 29th International Conference on Computational Linguistics",
"独特风格，语气和内容特征，从而可能降低其对其他社交媒体",
"International Commiteeon Computational Linguistics, Gyeongju,Korea"
] |
{
"arxiv_id": "2403.09702",
"language": "zh",
"page_idx": 2
}
|
||
2310.13395-zh-1
|
2310.13395-zh-1.jpg
|
## 如果可以的话，缓存我：一个在线成本感知的师生框架，以减少对大型语言模型的调用
Ilias Stogiannidis$^{1,2}\textbf{Stavros Vassos}^2\textbf{Prodromos Malakasiotis}^{1,4}\textbf{Ion Androutsopoulos}^{1,3}$
$^{1}$Department of Informatics, Athens University of Economics and Business, Greece
$^{2}$Helvia. ai
$^{3}$Archimedes Unit, Athena Research Center, Greece
$^{4}$Workable
$\{$ stoyianel, rulller, ion $\}@$aueb.gr, stavros@helvia.ai
## Abstract
<FigureHere>
Figure 1: OCaTS architecture.
Prompting Large Language Models (LLMs) performs impressively in zero- and few-shot settings. Hence, small and medium-sized enterprises (SMEs) that cannot afford the cost of creating large task-specific training datasets, but also the cost of pretraining their own LLMs, are increasingly turning to third-party services that allow them to prompt LLMs. However, such services currently require a payment per call, which becomes a significant operating expense (OpEx). Furthermore, customer inputs are often very similar over time, hence SMEs end-up prompting LLMs with very similar instances. We propose a framework that allows reducing the calls to LLMs by caching previous LLM responses and using them to train a local inexpensive model on the SME side. The framework includes criteria for deciding when to trust the local model or call the LLM, and a methodology to tune the criteria and measure the tradeoff between performance and cost. For experimental purposes, we instantiate our framework with two $\hat{L}$LMs, GPT- 3. 5 or GPT- 4, and two inexpen- sive students, a $k$ -NN classifier or a MultiLayer Perceptron, using two common business tasks, intent recognition and sentiment analysis. Experimental results indicate that significant OpEx savings can be obtained with only slightly lower performance.
电话付款，这些付款成为中小企业的一项重要运营费用 (OpEx)。此外 ,客户投入 (e。g。,dialogturns)areoftenverysimilarovertime,henceSMEendupcallingLLMto
我们引入了在线成本感知师生(OCaTS)框架，该框架允许通过缓存其先前的响应并使用它们来训练本地廉价的学生模型，从而减少对被视为教师模型的商业 LLM 的调用。OCaTS 包括决定何时信任学生或打电话给老师的标准，以及调整标准和衡量绩效与成本之间权衡的方法。与知识提炼(Hinton et al., 2015; Gou et al., 2021) 的常见师生培训不同，这里的老师不会在所有可用的实例上培训学生(在我们的例子中，所有传入的客户输入)。此外，与自我培训 (Mi et al., 2021; Li et al., 2021) 的师生方法不同，教师已经相当有效(但价格昂贵)。从这个意义上说，我们的工作更接近于主动学习(Settles, 2012; Monarch, 2021), 但是 OCaTS 在老师 LLM 提供的标签上训练学生，而不是人类，并且最初没有大量未标记的实例(客户输人)可供选择，因为实例在线到达
OCaTS 可以与任何允许提示 LLM 的服务以及任何类型的本地学生模型一起使用。出于实验目的，我们使用 GPT-3 实例化 OCaTS。5 或GPT-4 作为教师，$k$ -NN 或多层感知器 (MLP) 分类器作为学生，使用来自银行领域的意图识别数据集或情感分析数据集。实验结果表明，
## 1 Introduction
提示预先训练的大型语言模型 (LLM) 对齐以遵循说明 (Ouyang et al., 2022; Köpf et al., 2023) 在零和少射设置中表现出色。因此，中小型企业(SME)无法负担创建用于模型微调的大型任务特定培训数据集的成本，也无法负担预培训自己的 LLM 的成本，因此越来越多地转向第三方服务，使他们能够提示 LLM。例如，提供客户支持聊天机器人的中小企业会提示 GPT-4 (OpenAI, 2023) 等 LLM 检测用户意图并驱动聊天机器人与客户的交互(Ham et al., 2020)。然而，最好的 LLM 目前需要每次提示
|
[
[
0.1259765625,
0.08489990234375,
0.873046875,
0.0843505859375,
0.873046875,
0.1015625,
0.1259765625,
0.10272216796875
],
[
0.4130859375,
0.10394287109375,
0.58251953125,
0.1051025390625,
0.58251953125,
0.122314453125,
0.4130859375,
0.1217041015625
],
[
0.1259765625,
0.1312255859375,
0.8818359375,
0.1312255859375,
0.8818359375,
0.1484375,
0.1259765625,
0.1484375
],
[
0.16796875,
0.15087890625,
0.8369140625,
0.15087890625,
0.8369140625,
0.1644287109375,
0.16796875,
0.1644287109375
],
[
0.459228515625,
0.1644287109375,
0.5439453125,
0.16748046875,
0.54248046875,
0.182861328125,
0.45751953125,
0.179931640625
],
[
0.2880859375,
0.179931640625,
0.7138671875,
0.1817626953125,
0.7138671875,
0.1995849609375,
0.2880859375,
0.19775390625
],
[
0.45166015625,
0.19775390625,
0.54833984375,
0.20068359375,
0.54833984375,
0.215576171875,
0.45166015625,
0.2120361328125
],
[
0.287109375,
0.21728515625,
0.71875,
0.21728515625,
0.71875,
0.23095703125,
0.287109375,
0.23095703125
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.64404296875,
0.2529296875,
0.73291015625,
0.2529296875,
0.73291015625,
0.26416015625,
0.64404296875,
0.26416015625
],
[
0.145263671875,
0.27978515625,
0.459228515625,
0.27978515625,
0.459228515625,
0.29443359375,
0.145263671875,
0.29443359375
],
[
0.80029296875,
0.27490234375,
0.8427734375,
0.27490234375,
0.8427734375,
0.28515625,
0.80029296875,
0.28515625
],
[
0.69189453125,
0.2861328125,
0.78271484375,
0.2861328125,
0.78271484375,
0.29638671875,
0.69189453125,
0.29638671875
],
[
0.145263671875,
0.29443359375,
0.45751953125,
0.29443359375,
0.45751953125,
0.30810546875,
0.145263671875,
0.30810546875
],
[
0.68994140625,
0.29443359375,
0.78076171875,
0.297607421875,
0.78076171875,
0.308837890625,
0.68994140625,
0.305908203125
],
[
0.14697265625,
0.30712890625,
0.45751953125,
0.30712890625,
0.45751953125,
0.320556640625,
0.14697265625,
0.320556640625
],
[
0.145263671875,
0.322509765625,
0.459228515625,
0.322509765625,
0.459228515625,
0.336181640625,
0.145263671875,
0.336181640625
],
[
0.8095703125,
0.317626953125,
0.81689453125,
0.317626953125,
0.81689453125,
0.32373046875,
0.8095703125,
0.32373046875
],
[
0.80615234375,
0.32373046875,
0.81689453125,
0.32373046875,
0.81689453125,
0.327880859375,
0.80615234375,
0.327880859375
],
[
0.14697265625,
0.33740234375,
0.45751953125,
0.33740234375,
0.45751953125,
0.350830078125,
0.14697265625,
0.350830078125
],
[
0.53564453125,
0.342529296875,
0.5859375,
0.342529296875,
0.5859375,
0.352783203125,
0.53564453125,
0.352783203125
],
[
0.65576171875,
0.3466796875,
0.7109375,
0.3486328125,
0.7109375,
0.35986328125,
0.65576171875,
0.358154296875
],
[
0.77099609375,
0.3466796875,
0.8212890625,
0.3486328125,
0.8193359375,
0.35986328125,
0.76904296875,
0.358154296875
],
[
0.54150390625,
0.355224609375,
0.56689453125,
0.355224609375,
0.56689453125,
0.36767578125,
0.54150390625,
0.36767578125
],
[
0.771484375,
0.359375,
0.8193359375,
0.359375,
0.8193359375,
0.37060546875,
0.771484375,
0.37060546875
],
[
0.145263671875,
0.350830078125,
0.45751953125,
0.350830078125,
0.45751953125,
0.364013671875,
0.145263671875,
0.364013671875
],
[
0.145263671875,
0.365234375,
0.456787109375,
0.365234375,
0.456787109375,
0.37890625,
0.145263671875,
0.37890625
],
[
0.5380859375,
0.37353515625,
0.56103515625,
0.37353515625,
0.56103515625,
0.3818359375,
0.5380859375,
0.3818359375
],
[
0.14697265625,
0.380126953125,
0.456787109375,
0.380126953125,
0.456787109375,
0.393798828125,
0.14697265625,
0.393798828125
],
[
0.8017578125,
0.380126953125,
0.8427734375,
0.3818359375,
0.84228515625,
0.393798828125,
0.80029296875,
0.391357421875
],
[
0.63232421875,
0.38720703125,
0.6767578125,
0.38720703125,
0.6767578125,
0.3984375,
0.63232421875,
0.3984375
],
[
0.6884765625,
0.3896484375,
0.779296875,
0.3896484375,
0.779296875,
0.396728515625,
0.6884765625,
0.396728515625
],
[
0.80029296875,
0.391357421875,
0.84033203125,
0.393798828125,
0.84033203125,
0.402099609375,
0.80029296875,
0.399658203125
],
[
0.68701171875,
0.3955078125,
0.779296875,
0.396728515625,
0.779296875,
0.407958984375,
0.68701171875,
0.40673828125
],
[
0.145263671875,
0.393798828125,
0.459228515625,
0.393798828125,
0.459228515625,
0.40673828125,
0.145263671875,
0.40673828125
],
[
0.5634765625,
0.3984375,
0.615234375,
0.3984375,
0.615234375,
0.4091796875,
0.5634765625,
0.4091796875
],
[
0.6396484375,
0.402099609375,
0.6708984375,
0.40625,
0.66650390625,
0.41748046875,
0.634765625,
0.412109375
],
[
0.145263671875,
0.407958984375,
0.459228515625,
0.407958984375,
0.459228515625,
0.421630859375,
0.145263671875,
0.421630859375
],
[
0.5615234375,
0.407958984375,
0.609375,
0.407958984375,
0.609375,
0.418701171875,
0.5615234375,
0.418701171875
],
[
0.67919921875,
0.407958984375,
0.78662109375,
0.407958984375,
0.78662109375,
0.418701171875,
0.67919921875,
0.418701171875
],
[
0.145263671875,
0.4228515625,
0.459228515625,
0.4228515625,
0.459228515625,
0.435791015625,
0.145263671875,
0.435791015625
],
[
0.1485595703125,
0.438232421875,
0.455078125,
0.438232421875,
0.455078125,
0.448974609375,
0.1485595703125,
0.448974609375
],
[
0.78076171875,
0.440673828125,
0.8623046875,
0.440673828125,
0.8623046875,
0.45068359375,
0.78076171875,
0.45068359375
],
[
0.1435546875,
0.45068359375,
0.459228515625,
0.45068359375,
0.459228515625,
0.465576171875,
0.1435546875,
0.465576171875
],
[
0.64404296875,
0.451904296875,
0.73291015625,
0.451904296875,
0.73291015625,
0.463134765625,
0.64404296875,
0.463134765625
],
[
0.79833984375,
0.45068359375,
0.84814453125,
0.45068359375,
0.84814453125,
0.46142578125,
0.79833984375,
0.46142578125
],
[
0.145263671875,
0.46435546875,
0.459228515625,
0.46435546875,
0.459228515625,
0.47802734375,
0.145263671875,
0.47802734375
],
[
0.59375,
0.473876953125,
0.79833984375,
0.473876953125,
0.79833984375,
0.48681640625,
0.59375,
0.48681640625
],
[
0.1485595703125,
0.479736328125,
0.456787109375,
0.479736328125,
0.456787109375,
0.490478515625,
0.1485595703125,
0.490478515625
],
[
0.145263671875,
0.4921875,
0.459228515625,
0.4921875,
0.459228515625,
0.50732421875,
0.145263671875,
0.50732421875
],
[
0.145263671875,
0.50732421875,
0.45751953125,
0.50732421875,
0.45751953125,
0.52099609375,
0.145263671875,
0.52099609375
],
[
0.5146484375,
0.501953125,
0.880859375,
0.501953125,
0.880859375,
0.515625,
0.5146484375,
0.515625
],
[
0.145263671875,
0.521484375,
0.459228515625,
0.521484375,
0.459228515625,
0.53515625,
0.145263671875,
0.53515625
],
[
0.51318359375,
0.5185546875,
0.880859375,
0.5185546875,
0.880859375,
0.5322265625,
0.51318359375,
0.5322265625
],
[
0.14697265625,
0.53759765625,
0.456787109375,
0.53759765625,
0.456787109375,
0.54736328125,
0.14697265625,
0.54736328125
],
[
0.51123046875,
0.53369140625,
0.99853515625,
0.53369140625,
0.99853515625,
0.54736328125,
0.51123046875,
0.54736328125
],
[
0.1435546875,
0.548828125,
0.45751953125,
0.5498046875,
0.45751953125,
0.5634765625,
0.1435546875,
0.5625
],
[
0.53076171875,
0.5498046875,
0.880859375,
0.5498046875,
0.880859375,
0.5634765625,
0.53076171875,
0.5634765625
],
[
0.145263671875,
0.56396484375,
0.45751953125,
0.56396484375,
0.45751953125,
0.57763671875,
0.145263671875,
0.57763671875
],
[
0.51318359375,
0.56640625,
0.87890625,
0.56640625,
0.87890625,
0.580078125,
0.51318359375,
0.580078125
],
[
0.145263671875,
0.57763671875,
0.45751953125,
0.57763671875,
0.45751953125,
0.59130859375,
0.145263671875,
0.59130859375
],
[
0.5146484375,
0.5830078125,
0.880859375,
0.5830078125,
0.880859375,
0.5966796875,
0.5146484375,
0.5966796875
],
[
0.145263671875,
0.5927734375,
0.45751953125,
0.5927734375,
0.45751953125,
0.60546875,
0.145263671875,
0.60546875
],
[
0.51123046875,
0.59716796875,
0.87890625,
0.59716796875,
0.87890625,
0.61083984375,
0.51123046875,
0.61083984375
],
[
0.145263671875,
0.60546875,
0.459228515625,
0.60693359375,
0.459228515625,
0.62060546875,
0.145263671875,
0.619140625
],
[
0.51318359375,
0.61328125,
0.880859375,
0.6142578125,
0.880859375,
0.62890625,
0.51318359375,
0.62744140625
],
[
0.145263671875,
0.62060546875,
0.45751953125,
0.619140625,
0.45751953125,
0.6328125,
0.145263671875,
0.63427734375
],
[
0.145263671875,
0.63427734375,
0.456787109375,
0.63427734375,
0.456787109375,
0.64794921875,
0.145263671875,
0.64794921875
],
[
0.51318359375,
0.63134765625,
0.880859375,
0.63134765625,
0.880859375,
0.64453125,
0.51318359375,
0.64453125
],
[
0.14697265625,
0.65087890625,
0.45751953125,
0.65087890625,
0.45751953125,
0.6611328125,
0.14697265625,
0.6611328125
],
[
0.51123046875,
0.646484375,
0.8818359375,
0.6455078125,
0.8818359375,
0.66015625,
0.51123046875,
0.6611328125
],
[
0.145263671875,
0.6630859375,
0.459228515625,
0.6630859375,
0.459228515625,
0.6767578125,
0.145263671875,
0.6767578125
],
[
0.51123046875,
0.6611328125,
0.880859375,
0.662109375,
0.880859375,
0.6767578125,
0.51123046875,
0.67578125
],
[
0.145263671875,
0.6767578125,
0.456787109375,
0.6767578125,
0.456787109375,
0.6904296875,
0.145263671875,
0.6904296875
],
[
0.5146484375,
0.6787109375,
0.87890625,
0.6787109375,
0.87890625,
0.6923828125,
0.5146484375,
0.6923828125
],
[
0.14697265625,
0.693359375,
0.33154296875,
0.693359375,
0.33154296875,
0.70361328125,
0.14697265625,
0.70361328125
],
[
0.5146484375,
0.69482421875,
0.880859375,
0.69482421875,
0.880859375,
0.7080078125,
0.5146484375,
0.7080078125
],
[
0.51123046875,
0.708984375,
0.880859375,
0.71044921875,
0.880859375,
0.724609375,
0.51123046875,
0.7236328125
],
[
0.118408203125,
0.7197265625,
0.1318359375,
0.7197265625,
0.1318359375,
0.72998046875,
0.118408203125,
0.72998046875
],
[
0.151123046875,
0.7197265625,
0.2626953125,
0.7197265625,
0.2626953125,
0.73095703125,
0.151123046875,
0.73095703125
],
[
0.5146484375,
0.72705078125,
0.880859375,
0.72705078125,
0.880859375,
0.74072265625,
0.5146484375,
0.74072265625
],
[
0.118408203125,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.51318359375,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.1175537109375,
0.7587890625,
0.484375,
0.75830078125,
0.484375,
0.7724609375,
0.1175537109375,
0.77392578125
],
[
0.509765625,
0.75830078125,
0.87890625,
0.75830078125,
0.87890625,
0.771484375,
0.509765625,
0.771484375
],
[
0.1175537109375,
0.77490234375,
0.483642578125,
0.77490234375,
0.483642578125,
0.7880859375,
0.1175537109375,
0.7880859375
],
[
0.5146484375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.5146484375,
0.7890625
],
[
0.1175537109375,
0.79150390625,
0.484375,
0.79150390625,
0.484375,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51318359375,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.1175537109375,
0.8056640625,
0.484375,
0.80712890625,
0.484375,
0.82177734375,
0.1175537109375,
0.82080078125
],
[
0.51318359375,
0.80810546875,
0.7900390625,
0.80810546875,
0.7900390625,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.118408203125,
0.82470703125,
0.483642578125,
0.82470703125,
0.483642578125,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.52880859375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.52880859375,
0.83837890625
],
[
0.118408203125,
0.84033203125,
0.484375,
0.84033203125,
0.484375,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.51318359375,
0.84033203125,
0.87890625,
0.84033203125,
0.87890625,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.118408203125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51123046875,
0.87158203125,
0.880859375,
0.87158203125,
0.880859375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.88818359375,
0.48779296875,
0.88818359375,
0.48779296875,
0.90185546875,
0.118408203125,
0.90185546875
],
[
0.5146484375,
0.88818359375,
0.880859375,
0.88818359375,
0.880859375,
0.90283203125,
0.5146484375,
0.90283203125
],
[
0.1158447265625,
0.90283203125,
0.486083984375,
0.90380859375,
0.486083984375,
0.91845703125,
0.1158447265625,
0.91748046875
],
[
0.51318359375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"如果可以的话，缓存我：一个在线成本感知的师生框架，以减少对大型",
"语言模型的调用",
"Ilias Stogiannidis1,2 Stavros Vassos? Prodromos Malakasiotisi,4 Ion Androutsopoulosi,3",
"Department of Informatics, Athens University of Economics and Business, Greece",
"2Helvia.ai",
"3 Archimedes Unit, Athena Research Center, Greece",
"4Workable",
"( stoyianel, ruler, ion ] @aueb.gr, stavros@helvia.ai",
"Abstract",
"Teacher response",
"Prompting Large Language Models (LLMs)",
"Teacher",
"Incoming instance",
"performs impressively in zero- and few-shot",
"Teacher response",
"settings. Hence, small and medium-sized en-",
"terprises (SMEs) that cannot afford the cost of",
"9",
"",
"creating large task-specific training datasets,",
"Customer",
"Retraining",
"Incoming",
"",
"instance",
"but also the cost of pretraining their own",
"LLMs, are increasingly turning to third-party",
"",
"services that allow them to prompt LLMs.",
"Callthe",
"Student",
"Incominginstance",
"teacher",
"Student response",
"However, such services currently require a",
"Incoming",
"d",
"payment per call, which becomes a significant",
"instance",
"Reliability indicators",
"operating expense (OpEx). Furthermore, cus-",
"tomerinputs areoftenverysimilarovertime.",
"Use the student",
"hence SMEs end-up prompting LLMs with",
"Student response",
"response",
"very similar instances. We propose a frame-",
"Figure I: OCaTS architecture.",
"work that allows reducing the calls to LLMs",
"by caching previous LLM responses and us-",
"ing them to train a local inexpensive model on",
"电话付款，这些付款成为中小企业的一项重要",
"the SME side. The framework includes crite-",
"运营费用（OpEx)。此外，客户投人（e。g。，dialog",
"ria for deciding whento trustthelocal model",
"turns)areoftenverysimilarovertime,henceSMEendupcallingLLMtol",
"or call the LLM, and a methodology to tune",
"我们引入了在线成本感知师生（OCaTS）框",
"the criteria and measure the tradeoff between",
"架，该框架允许通过缓存其先前的响应并使用",
"performance and cost. For experimental pur",
"它们来训练本地廉价的学生模型，从而减少对",
"poses, we instantiate our framework with two",
"被视为教师模型的商业LLM的调用。OCaTS",
"LLMs, GPT-3.5 or GPT-4, and two inexpen-",
"包括决定何时信任学生或打电话给老师的标",
"sive students, a k -NN classifier or a Multi-",
"Layer Perceptron, using two common business",
"准，以及调整标准和衡量绩效与成本之间权衡",
"tasks,intentrecognition and sentimentanaly-",
"的方法。与知识提炼(Hintonetal.，2015;Gou",
"sis. Experimental results indicate that signifi-",
"etal.,2021)的常见师生培训不同，这里的老师",
"cant OpEx savings can be obtained with only",
"不会在所有可用的实例上培训学生（在我们的",
"slightlylowerperformance.",
"例子中，所有传入的客户输入）。此外，与自我",
"培训(Mietal.,2021;Lietal.,2021)的师生方法",
"1",
"Introduction",
"不同，教师已经相当有效（但价格昂贵）。从",
"提示预先训练的大型语言模型（LLM）对齐以",
"这个意义上说，我们的工作更接近于主动学习",
"遵循说明（Ouyangetal.,2022;Kopfetal.,2023)",
"(Settles,2012:Monarch.2021)，但是 OCaTS在",
"在零和少射设置中表现出色。因此，中小型企",
"老师LLM提供的标签上训练学生，而不是人",
"业（SME）无法负担创建用于模型微调的大",
"类，并且最初没有大量未标记的实例（客户输",
"型任务特定培训数据集的成本，也无法负担",
"入）可供选择，因为实例在线到达",
"预培训自已的LLM的成本，因此越来越多地",
"OCaTS可以与任何允许提示LLM的服务以",
"转向第三方服务，使他们能够提示LLM。例",
"及任何类型的本地学生模型一起使用。出于实",
"如，提供客户支持聊天机器人的中小企业会提",
"验目的，我们使用GPT-3实例化OCaTS。5或",
"示GPT-4（OpenAI,2023）等LLM检测用户意",
"GPT-4作为教师，k-NN或多层感知器（MLP）",
"图并驱动聊天机器人与客户的交互(Hametal.,",
"分类器作为学生，使用来自银行领域的意图识",
"2020）。然而，最好的LLM目前需要每次提示",
"别数据集或情感分析数据集。实验结果表明，"
] |
{
"arxiv_id": "2310.13395",
"language": "zh",
"page_idx": 1
}
|
||
2310.13395-zh-2
|
2310.13395-zh-2.jpg
|
色和黑色线条)。OCaTS 引入了通用评估措施的折扣版本(e。g。,准确性),允许中小企业量化它更喜欢倾向于更少的呼叫或更少的用户挫折(图中不同的$\lambda$值。2)
我们的主要贡献是：(i)我们引人了一个通用的师生框架，该框架通过缓存 LLM 的响应并培训廉价的本地学生模型来帮助中小企业减少对商业 LLM 的提示调用和相应的运营成本。(ii) 我们引入了通用评估措施的折扣版本， 使中小企业能够量化他们更喜欢较少的 LLM 电话与。增加了用户的挫折感(例如。g。,由于准确性较低),并调整框架的标准，这些标准决定何时信任本地学生模型或相应地呼叫法学硕士教师。(iii)我们使用 GPT-3 实例化框架。5 或 GPT-4 作为教师，$k$ -NN 或 MLP 分类器作为学生。(四)我们对中小企业的两个著名任务——意图识别和情绪分析进行了实验，结果表明，只要稍低一点的性能就可以获得显著的成本节约。这是探索具有更多数据集，模型和业务场景的建议框架的好处的第一步
## 2 Framework
体系结构：建议的框架(OCaTS)由三个主要组件组成(图。1):教师，通常是提供优质结果的资源密集型模型；学生，通常比教师小得多且简单得多的成本效益模型；缓存，传入实例的存储库(例如。g。, 客户请求)已由教师处理。我们假设该框架用于处理没有可用的大型数据集进行监督训练的任务，除了几个传入的实例 (可能每个类只有几个) 注释了基本事实(例如。g。,正确的标签)。对于中小企业来说， 这是一个非常常见的案例，它们无法承担创建大型任务特定的培训数据集的成本，但可以轻松构建少量的演示实例。师生设置是在线的， 因为每个传人的实例都是在推理时处理的，如下所示。首先，调用学生来处理实例。然后， 一些评估学生输出可靠性的特定于学生和任务的标准表明，如果学生的输出(e。g。,label)shouldbeusedoriftheteachershouldbeconsulted。如果选择了学生的输出，则将其作为对传入实例的响应返回。否则，将调用教师来处理实例。在后一种情况下，实例与教师的结果一起存储在缓存中。根据学生的类型，定期进行重新培训，以使用缓存的实例更新学生
实例：在本文的实验中，我们使用 GPT-3 实例化 OCaTS。5 或 GPT-4 老师，以距离加权的
只需稍低的性能就可以节省大量的运营成本。$k$ -NN 或 MLP 分类器作为学生，用于单标签例如，$k$ -NN 学生可以处理大约三分之二的意 分类任务 (意图识别或情感分析)。在所有情图识别任务的传人实例 (客户输入),而无需 况下，我们通过其基于MPNet 的(Song et al., 调用 GPT-4 老师 (图。2 ,左，红线)减少小 2020) 向量表示 (文本嵌人) 来表示每个传人实于 0。5个百分点的准确性(图。2, 中间，红 例 (客户请求),并且我们使用两个标准(图。E
1 ) 决定何时使用学生的响应或调用老师：(i) 学生$(k$ -NN 或 MLP) 为传入实例产生的概率分布(在标签集上)的熵，以及(ii)传入实例的向量表示与$k$最相似的缓存实例的向量表示的质心的距离。有关其他可能的标准，请咨询Nguyen et al.(2022)。我们将 OCaTS 的其他实例化(其他教师，学生，任务，表示)留给未来的工作
折扣评估措施：所提出的体系结构的主要目标是通过缓存以前的教师响应并使用它们来培训SME 方面的本地廉价学生模型，从而减少对昂贵教师模型的调用次数。当使用不太准确的学生模型时，这会在打电话给老师的运营成本和最终用户的挫败感之间进行权衡。为了量化这种权衡，我们引入了任何常见评估度量$\phi$的折扣变体$\hat{\phi}\left(\mathrm{e。g。,~准确性，~F1}\right)$,如下所示：
$$\hat{\phi}\:=\phi-\lambda\cdot\frac{M}{N}=\phi-\lambda\cdot\rho,$$
(1)
其中$N$是已处理的传人实例的数量(在其上测量$\phi$ ), $M$是在处理$N$实例时对教师进行的调用数量，$\rho=\frac MN$显示了传入实例的百分比 THX 是如何调用教师的，。例如，假设师生组合的准确性是$\phi=0.8$,但是该准确性是通过$\rho=\frac{1}{3}$实现的。如果中小企业考虑这个$\rho$值(这将转换为，e。g。,到每月的成本),成本相当于损失 5 个百分点的精度，那么$\hat{\phi}=0.75$ ,式 1 就变成了 $0. 75= 0. 8- \lambda \cdot \frac 13$ , 由此我们得到 $\lambda=0.15$ 。较大(或较小) 的 $\lambda$值对应于 SME 认为相同的$\rho$值在精度点损失方面的成本更高(或更低)的情况。我们还可以将公式$\mathbb{E}$1 重新表述为$\delta=\lambda\cdot\rho$,其中$\delta=\phi-\hat{\phi}$ 显示了$\phi$打折多少以说明$\rho$的成本。那么$\lambda$可以直观地认为是一种货币汇率，说明$\rho$用$\delta$来表示有多昂贵(e。g。,以百分点为单位的精度损失)。1
## 3 $\textbf{Main experiments}$
在这里，我们讨论了我们与 GPT-4 老师$,k$ -NN 学生以及银行意图识别数据集进行的实验。在附录中，我们报告了另外两组实验，一组是用MLP (附录$\operatorname{EF}$ )替换 $k$ -NN 学生，其余的设置保持不变，另一组是替换任务$\prime$数据集(通过
$^{\mathrm{l}}$We implicitly assume that the exchange rate $\lambda$ is constant for all the values of $\delta$ and $\rho.$ In practice, it may be different for different ranges of $\delta$ and $\rho$,but we leave this for future $work.$
|
[
[
0.12005615234375,
0.0872802734375,
0.484375,
0.0872802734375,
0.484375,
0.097412109375,
0.12005615234375,
0.097412109375
],
[
0.51123046875,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51123046875,
0.1009521484375
],
[
0.118408203125,
0.10272216796875,
0.483642578125,
0.10272216796875,
0.483642578125,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51123046875,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.51123046875,
0.11639404296875
],
[
0.12005615234375,
0.119384765625,
0.484375,
0.119384765625,
0.484375,
0.133056640625,
0.12005615234375,
0.133056640625
],
[
0.42236328125,
0.122314453125,
0.43408203125,
0.122314453125,
0.43408203125,
0.1270751953125,
0.42236328125,
0.1270751953125
],
[
0.51123046875,
0.1175537109375,
0.87890625,
0.1181640625,
0.87890625,
0.133056640625,
0.51123046875,
0.1318359375
],
[
0.12005615234375,
0.1353759765625,
0.48779296875,
0.1353759765625,
0.48779296875,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.51123046875,
0.1341552734375,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51123046875,
0.1478271484375
],
[
0.12005615234375,
0.1519775390625,
0.48779296875,
0.1519775390625,
0.48779296875,
0.1650390625,
0.12005615234375,
0.1650390625
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.12005615234375,
0.16748046875,
0.48779296875,
0.16748046875,
0.48779296875,
0.18115234375,
0.12005615234375,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.880859375,
0.16748046875,
0.880859375,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.1175537109375,
0.1817626953125,
0.486083984375,
0.18115234375,
0.486083984375,
0.198974609375,
0.1175537109375,
0.1995849609375
],
[
0.51318359375,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.12005615234375,
0.1995849609375,
0.48779296875,
0.1995849609375,
0.48779296875,
0.213134765625,
0.12005615234375,
0.213134765625
],
[
0.51318359375,
0.1995849609375,
0.87890625,
0.1995849609375,
0.87890625,
0.213134765625,
0.51318359375,
0.213134765625
],
[
0.1175537109375,
0.215576171875,
0.32568359375,
0.215576171875,
0.32568359375,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.5146484375,
0.215576171875,
0.87890625,
0.215576171875,
0.87890625,
0.2286376953125,
0.5146484375,
0.2286376953125
],
[
0.1351318359375,
0.232177734375,
0.48779296875,
0.23095703125,
0.48779296875,
0.2452392578125,
0.1351318359375,
0.2464599609375
],
[
0.51123046875,
0.232177734375,
0.880859375,
0.232177734375,
0.880859375,
0.2452392578125,
0.51123046875,
0.2452392578125
],
[
0.118408203125,
0.248779296875,
0.489501953125,
0.248779296875,
0.489501953125,
0.262451171875,
0.118408203125,
0.262451171875
],
[
0.509765625,
0.2476806640625,
0.8818359375,
0.2464599609375,
0.8818359375,
0.26123046875,
0.509765625,
0.262451171875
],
[
0.118408203125,
0.26416015625,
0.48779296875,
0.26416015625,
0.48779296875,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.51123046875,
0.26318359375,
0.87890625,
0.26318359375,
0.87890625,
0.276611328125,
0.51123046875,
0.276611328125
],
[
0.118408203125,
0.28076171875,
0.489501953125,
0.28076171875,
0.489501953125,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.51318359375,
0.28076171875,
0.5859375,
0.28076171875,
0.5859375,
0.292236328125,
0.51318359375,
0.292236328125
],
[
0.118408203125,
0.29638671875,
0.489501953125,
0.29638671875,
0.489501953125,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.51123046875,
0.2998046875,
0.880859375,
0.2998046875,
0.880859375,
0.3134765625,
0.51123046875,
0.3134765625
],
[
0.12005615234375,
0.312255859375,
0.48779296875,
0.312255859375,
0.48779296875,
0.325927734375,
0.12005615234375,
0.325927734375
],
[
0.5146484375,
0.317626953125,
0.87890625,
0.317626953125,
0.87890625,
0.327880859375,
0.5146484375,
0.327880859375
],
[
0.118408203125,
0.328857421875,
0.484375,
0.328857421875,
0.484375,
0.342529296875,
0.118408203125,
0.342529296875
],
[
0.51123046875,
0.330810546875,
0.880859375,
0.330810546875,
0.880859375,
0.344482421875,
0.51123046875,
0.344482421875
],
[
0.118408203125,
0.34326171875,
0.486083984375,
0.34326171875,
0.486083984375,
0.35693359375,
0.118408203125,
0.35693359375
],
[
0.51318359375,
0.347412109375,
0.880859375,
0.347412109375,
0.880859375,
0.361083984375,
0.51318359375,
0.361083984375
],
[
0.118408203125,
0.361083984375,
0.484375,
0.361083984375,
0.484375,
0.374755859375,
0.118408203125,
0.374755859375
],
[
0.51318359375,
0.363525390625,
0.880859375,
0.363525390625,
0.880859375,
0.377197265625,
0.51318359375,
0.377197265625
],
[
0.12176513671875,
0.37890625,
0.489501953125,
0.37890625,
0.489501953125,
0.3896484375,
0.12176513671875,
0.3896484375
],
[
0.51318359375,
0.380126953125,
0.880859375,
0.380126953125,
0.880859375,
0.393798828125,
0.51318359375,
0.393798828125
],
[
0.1158447265625,
0.392578125,
0.486083984375,
0.392578125,
0.486083984375,
0.40625,
0.1158447265625,
0.40625
],
[
0.51123046875,
0.3955078125,
0.880859375,
0.3955078125,
0.880859375,
0.4091796875,
0.51123046875,
0.4091796875
],
[
0.1175537109375,
0.4091796875,
0.484375,
0.407958984375,
0.484375,
0.4228515625,
0.1175537109375,
0.42333984375
],
[
0.51123046875,
0.412109375,
0.87744140625,
0.412109375,
0.87744140625,
0.427001953125,
0.51123046875,
0.427001953125
],
[
0.118408203125,
0.42578125,
0.483642578125,
0.42578125,
0.483642578125,
0.439453125,
0.118408203125,
0.439453125
],
[
0.12005615234375,
0.441162109375,
0.483642578125,
0.441162109375,
0.483642578125,
0.454833984375,
0.12005615234375,
0.454833984375
],
[
0.580078125,
0.4423828125,
0.7255859375,
0.431640625,
0.73046875,
0.457275390625,
0.58447265625,
0.46728515625
],
[
0.6689453125,
0.446533203125,
0.8134765625,
0.440673828125,
0.8154296875,
0.4619140625,
0.6708984375,
0.468505859375
],
[
0.80419921875,
0.45068359375,
0.84619140625,
0.448974609375,
0.84619140625,
0.457275390625,
0.80419921875,
0.458984375
],
[
0.8564453125,
0.44775390625,
0.8818359375,
0.44775390625,
0.8818359375,
0.460205078125,
0.8564453125,
0.460205078125
],
[
0.118408203125,
0.458984375,
0.48193359375,
0.458984375,
0.48193359375,
0.4697265625,
0.118408203125,
0.4697265625
],
[
0.118408203125,
0.473876953125,
0.411376953125,
0.473876953125,
0.411376953125,
0.48681640625,
0.118408203125,
0.48681640625
],
[
0.5146484375,
0.47265625,
0.880859375,
0.47265625,
0.880859375,
0.486328125,
0.5146484375,
0.486328125
],
[
0.51123046875,
0.48681640625,
0.8818359375,
0.488037109375,
0.8818359375,
0.5029296875,
0.51123046875,
0.501953125
],
[
0.1158447265625,
0.50439453125,
0.255126953125,
0.50439453125,
0.255126953125,
0.51708984375,
0.1158447265625,
0.51708984375
],
[
0.51318359375,
0.50439453125,
0.880859375,
0.5048828125,
0.880859375,
0.51953125,
0.51318359375,
0.5185546875
],
[
0.5146484375,
0.52099609375,
0.880859375,
0.52099609375,
0.880859375,
0.53369140625,
0.5146484375,
0.53369140625
],
[
0.1175537109375,
0.52978515625,
0.483642578125,
0.52978515625,
0.483642578125,
0.54345703125,
0.1175537109375,
0.54345703125
],
[
0.51123046875,
0.53759765625,
0.87744140625,
0.53759765625,
0.87744140625,
0.54736328125,
0.51123046875,
0.54736328125
],
[
0.118408203125,
0.54736328125,
0.483642578125,
0.54736328125,
0.483642578125,
0.55810546875,
0.118408203125,
0.55810546875
],
[
0.51123046875,
0.5517578125,
0.87890625,
0.5517578125,
0.87890625,
0.5654296875,
0.51123046875,
0.5654296875
],
[
0.118408203125,
0.5634765625,
0.486083984375,
0.5634765625,
0.486083984375,
0.57666015625,
0.118408203125,
0.57666015625
],
[
0.51123046875,
0.5693359375,
0.880859375,
0.5693359375,
0.880859375,
0.58447265625,
0.51123046875,
0.58447265625
],
[
0.1175537109375,
0.57763671875,
0.484375,
0.57763671875,
0.484375,
0.59130859375,
0.1175537109375,
0.59130859375
],
[
0.51318359375,
0.5849609375,
0.880859375,
0.5849609375,
0.880859375,
0.5986328125,
0.51318359375,
0.5986328125
],
[
0.12005615234375,
0.595703125,
0.484375,
0.595703125,
0.484375,
0.609375,
0.12005615234375,
0.609375
],
[
0.5205078125,
0.60546875,
0.52880859375,
0.60546875,
0.52880859375,
0.61083984375,
0.5205078125,
0.61083984375
],
[
0.5263671875,
0.60107421875,
0.87890625,
0.60107421875,
0.87890625,
0.6142578125,
0.5263671875,
0.6142578125
],
[
0.118408203125,
0.60986328125,
0.486083984375,
0.60986328125,
0.486083984375,
0.62353515625,
0.118408203125,
0.62353515625
],
[
0.51318359375,
0.61767578125,
0.87890625,
0.61767578125,
0.87890625,
0.63134765625,
0.51318359375,
0.63134765625
],
[
0.118408203125,
0.62744140625,
0.484375,
0.62744140625,
0.484375,
0.64111328125,
0.118408203125,
0.64111328125
],
[
0.51123046875,
0.6318359375,
0.87744140625,
0.6318359375,
0.87744140625,
0.6455078125,
0.51123046875,
0.6455078125
],
[
0.118408203125,
0.64453125,
0.48779296875,
0.64453125,
0.48779296875,
0.65771484375,
0.118408203125,
0.65771484375
],
[
0.51318359375,
0.6494140625,
0.87890625,
0.6494140625,
0.87890625,
0.6630859375,
0.51318359375,
0.6630859375
],
[
0.118408203125,
0.66015625,
0.489501953125,
0.66015625,
0.489501953125,
0.6748046875,
0.118408203125,
0.6748046875
],
[
0.51123046875,
0.66455078125,
0.880859375,
0.6650390625,
0.880859375,
0.6796875,
0.51123046875,
0.6787109375
],
[
0.118408203125,
0.6767578125,
0.486083984375,
0.6767578125,
0.486083984375,
0.6904296875,
0.118408203125,
0.6904296875
],
[
0.51318359375,
0.681640625,
0.87890625,
0.681640625,
0.87890625,
0.6953125,
0.51318359375,
0.6953125
],
[
0.118408203125,
0.6923828125,
0.48779296875,
0.6923828125,
0.48779296875,
0.7060546875,
0.118408203125,
0.7060546875
],
[
0.5146484375,
0.69970703125,
0.87890625,
0.69970703125,
0.87890625,
0.71044921875,
0.5146484375,
0.71044921875
],
[
0.118408203125,
0.7080078125,
0.48779296875,
0.7080078125,
0.48779296875,
0.7216796875,
0.118408203125,
0.7216796875
],
[
0.51318359375,
0.71435546875,
0.880859375,
0.71435546875,
0.880859375,
0.728515625,
0.51318359375,
0.728515625
],
[
0.118408203125,
0.724609375,
0.48779296875,
0.724609375,
0.48779296875,
0.73828125,
0.118408203125,
0.73828125
],
[
0.5087890625,
0.73095703125,
0.58251953125,
0.728515625,
0.58251953125,
0.74365234375,
0.5087890625,
0.7451171875
],
[
0.12005615234375,
0.74072265625,
0.484375,
0.74072265625,
0.484375,
0.75439453125,
0.12005615234375,
0.75439453125
],
[
0.1275634765625,
0.75830078125,
0.486083984375,
0.75830078125,
0.486083984375,
0.7685546875,
0.1275634765625,
0.7685546875
],
[
0.5087890625,
0.7587890625,
0.5498046875,
0.7587890625,
0.5498046875,
0.76953125,
0.5087890625,
0.76953125
],
[
0.54248046875,
0.75830078125,
0.70361328125,
0.75830078125,
0.70361328125,
0.771484375,
0.54248046875,
0.771484375
],
[
0.1175537109375,
0.771484375,
0.486083984375,
0.7724609375,
0.486083984375,
0.78759765625,
0.1175537109375,
0.7861328125
],
[
0.491943359375,
0.76123046875,
0.50537109375,
0.76123046875,
0.50537109375,
0.765625,
0.491943359375,
0.765625
],
[
0.1175537109375,
0.7880859375,
0.496337890625,
0.7880859375,
0.496337890625,
0.8017578125,
0.1175537109375,
0.8017578125
],
[
0.51318359375,
0.783203125,
0.880859375,
0.783203125,
0.880859375,
0.79638671875,
0.51318359375,
0.79638671875
],
[
0.12005615234375,
0.8046875,
0.486083984375,
0.8046875,
0.486083984375,
0.818359375,
0.12005615234375,
0.818359375
],
[
0.5146484375,
0.80029296875,
0.87890625,
0.80029296875,
0.87890625,
0.81103515625,
0.5146484375,
0.81103515625
],
[
0.12005615234375,
0.82080078125,
0.486083984375,
0.82080078125,
0.486083984375,
0.83447265625,
0.12005615234375,
0.83447265625
],
[
0.51318359375,
0.8154296875,
0.880859375,
0.8154296875,
0.880859375,
0.8291015625,
0.51318359375,
0.8291015625
],
[
0.51318359375,
0.83056640625,
0.880859375,
0.83056640625,
0.880859375,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.12005615234375,
0.83740234375,
0.484375,
0.83740234375,
0.484375,
0.85107421875,
0.12005615234375,
0.85107421875
],
[
0.118408203125,
0.8525390625,
0.484375,
0.8525390625,
0.484375,
0.8662109375,
0.118408203125,
0.8662109375
],
[
0.51318359375,
0.84716796875,
0.880859375,
0.84716796875,
0.880859375,
0.86083984375,
0.51318359375,
0.86083984375
],
[
0.118408203125,
0.869140625,
0.411376953125,
0.869140625,
0.411376953125,
0.8828125,
0.118408203125,
0.8828125
],
[
0.53564453125,
0.869140625,
0.87890625,
0.87158203125,
0.87890625,
0.88525390625,
0.53564453125,
0.8828125
],
[
0.5146484375,
0.88427734375,
0.87744140625,
0.88427734375,
0.87744140625,
0.89453125,
0.5146484375,
0.89453125
],
[
0.1175537109375,
0.89013671875,
0.48779296875,
0.88818359375,
0.48779296875,
0.90185546875,
0.1175537109375,
0.90380859375
],
[
0.51123046875,
0.89453125,
0.880859375,
0.89453125,
0.880859375,
0.90771484375,
0.51123046875,
0.90771484375
],
[
0.12176513671875,
0.90673828125,
0.486083984375,
0.90673828125,
0.486083984375,
0.91748046875,
0.12176513671875,
0.91748046875
],
[
0.5146484375,
0.9091796875,
0.5439453125,
0.9091796875,
0.5439453125,
0.91748046875,
0.5146484375,
0.91748046875
]
] |
[
"只需稍低的性能就可以节省大量的运营成本。",
"k-NN或MLP分类器作为学生，用于单标签",
"例如，k-NN学生可以处理大约三分之二的意",
"分类任务（意图识别或情感分析）。在所有情",
"图识别任务的传入实例（客户输入），而无需",
"",
"况下，我们通过其基于MPNet的(Songetal.,",
"调用GPT-4老师（图。2，左，红线）减少小",
"2020）向量表示（文本嵌人）来表示每个传入实",
"于0。5个百分点的准确性（图。2，中间，红",
"例（客户请求），并且我们使用两个标准（图。E",
"色和黑色线条）。OCaTS引入了通用评估措施",
"1）决定何时使用学生的响应或调用老师：（i）",
"的折扣版本(e。g。，准确性)，允许中小企业量",
"学生（k-NN或MLP）为传入实例产生的概率",
"化它更喜欢倾向于更少的呼叫或更少的用户挫",
"分布（在标签集上）的，以及（i）传人实例",
"折图中不同的入值。2)",
"的向量表示与k最相似的缓存实例的向量表示",
"我们的主要贡献是：（i）我们引人了一个通",
"的质心的距离。有关其他可能的标准，请咨询",
"用的师生框架，该框架通过缓存LLM的响应",
"Nguyenetal.（2022）。我们将OCaTS的其他实",
"并培训廉价的本地学生模型来帮助中小企业",
"例化（其他教师，学生，任务，表示）留给未",
"减少对商业LLM的提示调用和相应的运营成",
"来的工作",
"本。（i）我们引入了通用评估措施的折扣版本，",
"折扣评估措施：所提出的体系结构的主要目标",
"使中小企业能够量化他们更喜欢较少的LLM",
"是通过缓存以前的教师响应并使用它们来培训",
"电话与。增加了用户的挫折感（例如。g。，由",
"SME方面的本地廉价学生模型，从而减少对",
"于准确性较低），并调整框架的标准，这些标准",
"昂贵教师模型的调用次数。当使用不太准确的",
"决定何时信任本地学生模型或相应地呼叫法学",
"学生模型时，这会在打电话给老师的运营成本",
"硕士教师。（iii）我们使用GPT-3实例化框架。",
"和最终用户的挫败感之间进行权衡。为了量化",
"5或GPT-4作为教师，k-NN或MLP分类器作",
"这种权衡，我们引入了任何常见评估度量Φ的",
"为学生。（四）我们对中小企业的两个著名任",
"折扣变体Φ(e。g。，准确性，F1)，如下所示：",
"务一意图识别和情绪分析进行了实验，结果",
"表明，只要稍低一点的性能就可以获得显著的",
"0-",
"岁",
"",
"(1)",
"成本节约。这是探索具有更多数据集，模型和",
"业务场景的建议框架的好处的第一步",
"其中N是已处理的传入实例的数量（在其上",
"测量Φ），M是在处理N实例时对教师进行",
"2Framework",
"的调用数量，P=兴显示了传入实例的百分",
"比THX是如何调用教师的，。例如，假设师生",
"体系结构：建议的框架（OCaTS）由三个主要",
"组合的准确性是@三0.8，但是该准确性是通",
"组件组成（图。1）：教师，通常是提供优质结",
"过p=实现的。如果中小企业考这个p值",
"果的资源密集型模型；学生，通常比教师小得",
"(这将转换为，e。g。，到每月的成本），成本",
"多且简单得多的成本效益模型；缓存，传人实",
"相当于损失5个百分点的精度，那么Φ=0.75",
"例的存储库（例如。g。，客户请求）已由教师处",
"?",
"式1就变成了0.75=0.8－入·，由此我",
"理。我们假设该框架用于处理没有可用的大型",
"们得到入=0.15。较大（或较小）的入值对应",
"数据集进行监督训练的任务，除了几个传入的",
"于SME认为相同的p值在精度点损失方面的",
"实例（可能每个类只有几个）注释了基本事实",
"成本更高（或更低）的情况。我们还可以将公",
"(例如。g。，正确的标签)。对于中小企业来说，",
"式国1重新表述为=p，其中8=Φ-Φ",
"这是一个非常常见的案例，它们无法承担创建",
"显示了Φ打折多少以说明p的成本。那么入可",
"大型任务特定的培训数据集的成本，但可以轻",
"以直观地认为是一种货市币汇率，说明p用。来",
"松构建少量的演示实例。师生设置是在线的，",
"表示有多昂贵（e。g。，以百分点为单位的精度",
"因为每个传人的实例都是在推理时处理的，如",
"损失)。1",
"下所示。首先，调用学生来处理实例。然后",
"一些评估学生输出可靠性的特定于学生和任",
"3",
"Mainexperiments",
"务的标准表明，如果学生的输出（e。g。,la",
"",
"bel)shouldbeusedoriftheteachershouldbeconsulted。",
"在这里，我们讨论了我们与GPT-4老师，k-NN",
"如果选择了学生的输出，则将其作为对传入实",
"学生以及银行意图识别数据集进行的实验。在",
"例的响应返回。否则，将调用教师来处理实",
"附录中，我们报告了另外两组实验，一组是用",
"MLP（附录EF）替换k-NN学生，其余的设",
"例。在后一种情况下，实例与教师的结果一起",
"存储在缓存中。根据学生的类型，定期进行重",
"置保持不变，另一组是替换任务/数据集（通过",
"新培训，以使用缓存的实例更新学生",
"'We implicitly assume that the exchange rate Ais constant",
"for all the values of and p.In practice, it may be differen",
"实例：在本文的实验中，我们使用GPT-3实",
"for different ranges of and p , but we leave this for future",
"例化OCaTS。5或GPT-4老师，以距离加权的",
"work"
] |
{
"arxiv_id": "2310.13395",
"language": "zh",
"page_idx": 2
}
|
||
2310.13395-zh-4
|
2310.13395-zh-4.jpg
|
阈值$t_{\mathbf{c}},t_{\mathcal{H}}$。考虑到最初每个类有 3个示例，G)
我们将$k=5$修复为实际选择。对于每个指示性 $\lambda$值(0。05,0。1、0。2、0。3),我们在少射开发集 (Section 3)上采用贝叶斯优化，以确定我们让$t_{\mathrm{c}}$在[0,2]中$,t_{\mathcal{H}}$在$[0,4.34]\text{ 中。}^{3}$在首先对两个阈值的值范围执行$10\times10$网格搜索之后，我们使用 Optuna 的 (Akiba et al., 2019) 实现树结构参数估计器(TSPE)算法(Bergstra et al.,2011) 作为起点。得到的等高线图和每个$\lambda$值的两个阈值的最佳值可以在附录 E 中找到
结果：我们使用相同的传入实例(Banking77的原始测试集)和$\lambda$特定的调谐阈值$t_\mathrm{c},t_\mathrm{H}$ 来评估四个指示性$\lambda$值中每个值的 OCaTS。如图所示。2, OCaTS 成功地管理了老师与。精度。图 2(左)显示，随着折扣因子$\lambda$ 的增加， 给老师打的电话更少。InFig。$\boxed{\mathrm{E}}$2(中),我们看到为此 OpEx 救济牺牲了多少准确性。特别是，对于 $\lambda = 0. 05$ , OCaTS 的准确性非常接近GPT-4 老师的准确性，在 0 的范围内。37 个百分点(83。05%vs。82。整个测试集的 68 个$\%)$, 同时仅针对传人实例的 1/3 调用老师 (3080 个中的 1050 个)。对于$\lambda$的较高值，我们看到预 虑了在上下文中学习，每个班级有几个演示示期的准确性下降，以实现对教师的呼叫数量越 例。这些例子中手动选择是多样的，并表示相来越少。图 2(右)显示 OCaTS 的折扣精度$\hat{\phi}$应的类。这在一定程度上是现实的；中小企业(实线，每个 $\lambda$值一个) 总是明显高于总是呼叫 经常向客户索取少量的实例，但这些实例的质GPT-4 老师的相应折扣精度 (虚线)。因此，如 量并不总是有保证的。此外，我们使用的测试果考虑到运营成本，OCaTS 显然比总是打电话 集(来自 Banking77 和 LMR) 是平衡的，因此给老师要好。随着 λ的增加，差异增加(有利于 OCaTS), i。e。, 随着降低运营成本变得越来越重要
## 4 Conclusions
我们引入了一个在线成本感知师生框架(OCaTS),通过缓存商业 LLM 的响应和培训廉价的本地学生来帮助中小企业降低运营成本。我们还引入了通用评估措施的折扣版本， 使中小企业能够量化 LLM 呼叫和用户挫败感之间的权衡。通过使用 GPT-4 老师和$k$ -NN 学生实例化 OCaTS,并尝试使用来自银行领域(Banking77)的意图识别数据集，我们表明，对老师的调用可以显着减少(减少 1/3),而性能仅略有下降(0。37个百分点)。在同一数据集上对 MLP 学生进行的其他实验得出了相同的
$^3$The maximum value of $\mathcal{H}$ with 77 classes is 4.34, when using natural logarithms. The upper bound of $t_{\mathrm{C}}$ was chosen based on initial experiments on development data.
第二个标准要求 $\mathcal{H}_w$ 小于阈值 XMATHX1。直 结果 (附录$\mathbb{E}$F)。GPT-3 的进一步实验。5 位观上，它要求邻居就传人实例的标签达成一致 老师，最初的$k$ -NN学生，以及情绪分析数据超参数调优：这里有三个超参数，邻居数 $k$, 集(LMR)也证实了先前实验的结论(附录E)
在以后的工作中，我们计划尝试更多的数据集和任务(e。g。, 问题回答),并建议$\lambda$的自适应策略，以便在缓存较冷时允许更高的 OpEx
最大化$\hat{\phi}\left(\text{折扣精度)的两个阈值的最优组合。成本 (更频繁地调用教师),并且以后更有选择}\right.$
性 (更不频繁地调用教师)。我们还计划通过我们可以信任老师的回答的指标来增强 OCaTS (例如。g。,老师的信心)。最后，我们打算纳人更多的财务指标(例如。g。, 学生成本)在评估措施的折扣版本中，并研究更复杂的策略(例如。g。,博弈论，强化学习)来选择决定何时信任学生或打电话给老师的阈值
# 5 Limitations
这项工作的主要范围是提出一个灵活的框架(OCaTS),该框架将允许中小型企业在将商业LLM 纳人其解决方案时降低运营成本。我们只考虑了教师的两个实例(GPT-4、GPT-3。5) 和两个任务 (意图识别，情感分析)中的学生$(k$ -NN, MLP)的两个实例化，为将来的工作留下进一步的实例化。虽然 LLM 像 GPT-4 和GPT-3。5 原则上可以用于零拍推理，我们考不完全现实。但是，我们对传入(测试)实例的流进行了洗牌，因此，它们的类不会以统一的方式到达。此外，为了调整$t_c$和$t_\mathcal{H}$,我们使用了从原始训练数据中提取的开发集。这样的开发集在实践中并不总是可用的，但是我们为了分析而使用它。有兴趣的中小企业可以使用我们的分析，作为其应用程序的起点，并减少为$t_c$和$t_\mathcal{H}$找到合适值所需的试验次数
另一个限制是，$\hat{\phi}$仅考虑了调用老师($\rho)$ 的成本，并间接考虑了用户的挫败感，如性能下降所暗示的。更详细的分析还将纳人学生成本和其他可能具有不同权重的财务指标； OCaTS 可以轻松地在该方向上扩展。最后，我们没有与现有的缓存库进行比较，例如。g。GPTCache。$^{4}$与 OCaTS 相比，这些库相当简单且不那么灵活，可以与各种师生设置一起使用
$^{4}$https: / / gi thub. com/ zilliztech/ GPTCache
|
[
[
0.12005615234375,
0.0885009765625,
0.489501953125,
0.0885009765625,
0.489501953125,
0.09857177734375,
0.12005615234375,
0.09857177734375
],
[
0.484375,
0.08905029296875,
0.50732421875,
0.08905029296875,
0.50732421875,
0.0968017578125,
0.484375,
0.0968017578125
],
[
0.509765625,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.509765625,
0.09979248046875
],
[
0.1175537109375,
0.10272216796875,
0.486083984375,
0.10272216796875,
0.486083984375,
0.11639404296875,
0.1175537109375,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.118408203125,
0.122314453125,
0.48779296875,
0.122314453125,
0.48779296875,
0.135986328125,
0.118408203125,
0.135986328125
],
[
0.51123046875,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.51123046875,
0.1318359375
],
[
0.12005615234375,
0.1395263671875,
0.48779296875,
0.1395263671875,
0.48779296875,
0.15380859375,
0.12005615234375,
0.15380859375
],
[
0.51318359375,
0.135986328125,
0.5400390625,
0.135986328125,
0.5400390625,
0.1478271484375,
0.51318359375,
0.1478271484375
],
[
0.118408203125,
0.155029296875,
0.484375,
0.155029296875,
0.484375,
0.168701171875,
0.118408203125,
0.168701171875
],
[
0.53076171875,
0.152587890625,
0.87890625,
0.152587890625,
0.87890625,
0.166259765625,
0.53076171875,
0.166259765625
],
[
0.118408203125,
0.17041015625,
0.484375,
0.17041015625,
0.484375,
0.18408203125,
0.118408203125,
0.18408203125
],
[
0.51318359375,
0.17041015625,
0.880859375,
0.17041015625,
0.880859375,
0.18408203125,
0.51318359375,
0.18408203125
],
[
0.118408203125,
0.18701171875,
0.484375,
0.18701171875,
0.484375,
0.20068359375,
0.118408203125,
0.20068359375
],
[
0.51318359375,
0.1859130859375,
0.880859375,
0.1859130859375,
0.880859375,
0.1995849609375,
0.51318359375,
0.1995849609375
],
[
0.118408203125,
0.203125,
0.486083984375,
0.203125,
0.486083984375,
0.2161865234375,
0.118408203125,
0.2161865234375
],
[
0.5146484375,
0.201904296875,
0.880859375,
0.201904296875,
0.880859375,
0.215576171875,
0.5146484375,
0.215576171875
],
[
0.118408203125,
0.2197265625,
0.484375,
0.2197265625,
0.484375,
0.2327880859375,
0.118408203125,
0.2327880859375
],
[
0.51318359375,
0.218505859375,
0.880859375,
0.218505859375,
0.880859375,
0.232177734375,
0.51318359375,
0.232177734375
],
[
0.1175537109375,
0.235107421875,
0.48193359375,
0.235107421875,
0.48193359375,
0.248779296875,
0.1175537109375,
0.248779296875
],
[
0.51318359375,
0.235107421875,
0.880859375,
0.235107421875,
0.880859375,
0.248779296875,
0.51318359375,
0.248779296875
],
[
0.118408203125,
0.251708984375,
0.484375,
0.251708984375,
0.484375,
0.265380859375,
0.118408203125,
0.265380859375
],
[
0.51123046875,
0.251708984375,
0.87890625,
0.251708984375,
0.87890625,
0.265380859375,
0.51123046875,
0.265380859375
],
[
0.12005615234375,
0.267333984375,
0.484375,
0.267333984375,
0.484375,
0.28076171875,
0.12005615234375,
0.28076171875
],
[
0.515625,
0.267333984375,
0.880859375,
0.267333984375,
0.880859375,
0.281982421875,
0.515625,
0.281982421875
],
[
0.118408203125,
0.283935546875,
0.483642578125,
0.283935546875,
0.483642578125,
0.297607421875,
0.118408203125,
0.297607421875
],
[
0.51123046875,
0.281982421875,
0.87744140625,
0.281982421875,
0.87744140625,
0.295654296875,
0.51123046875,
0.295654296875
],
[
0.1158447265625,
0.297607421875,
0.48779296875,
0.298583984375,
0.48779296875,
0.3134765625,
0.1158447265625,
0.312255859375
],
[
0.51123046875,
0.298583984375,
0.87890625,
0.297607421875,
0.87890625,
0.312255859375,
0.51123046875,
0.3134765625
],
[
0.1175537109375,
0.3154296875,
0.1376953125,
0.3154296875,
0.1376953125,
0.330078125,
0.1175537109375,
0.330078125
],
[
0.5146484375,
0.3154296875,
0.77099609375,
0.3154296875,
0.77099609375,
0.328857421875,
0.5146484375,
0.328857421875
],
[
0.118408203125,
0.336181640625,
0.486083984375,
0.336181640625,
0.486083984375,
0.349853515625,
0.118408203125,
0.349853515625
],
[
0.1175537109375,
0.350830078125,
0.484375,
0.3515625,
0.484375,
0.366455078125,
0.1175537109375,
0.365234375
],
[
0.50537109375,
0.3515625,
0.6484375,
0.352783203125,
0.6484375,
0.36767578125,
0.50537109375,
0.366455078125
],
[
0.497802734375,
0.358154296875,
0.50537109375,
0.358154296875,
0.50537109375,
0.361083984375,
0.497802734375,
0.361083984375
],
[
0.12005615234375,
0.36767578125,
0.484375,
0.36767578125,
0.484375,
0.38134765625,
0.12005615234375,
0.38134765625
],
[
0.12005615234375,
0.38427734375,
0.48779296875,
0.38427734375,
0.48779296875,
0.39794921875,
0.12005615234375,
0.39794921875
],
[
0.51123046875,
0.38427734375,
0.87890625,
0.38427734375,
0.87890625,
0.39794921875,
0.51123046875,
0.39794921875
],
[
0.12005615234375,
0.399658203125,
0.489501953125,
0.399658203125,
0.489501953125,
0.413330078125,
0.12005615234375,
0.413330078125
],
[
0.51123046875,
0.402099609375,
0.880859375,
0.402099609375,
0.880859375,
0.4150390625,
0.51123046875,
0.4150390625
],
[
0.1175537109375,
0.4150390625,
0.484375,
0.41455078125,
0.484375,
0.4287109375,
0.1175537109375,
0.429931640625
],
[
0.51123046875,
0.416259765625,
0.880859375,
0.41748046875,
0.880859375,
0.431640625,
0.51123046875,
0.43115234375
],
[
0.118408203125,
0.431640625,
0.48779296875,
0.431640625,
0.48779296875,
0.4453125,
0.118408203125,
0.4453125
],
[
0.5146484375,
0.43408203125,
0.880859375,
0.43408203125,
0.880859375,
0.44775390625,
0.5146484375,
0.44775390625
],
[
0.118408203125,
0.44775390625,
0.486083984375,
0.44775390625,
0.486083984375,
0.46142578125,
0.118408203125,
0.46142578125
],
[
0.51318359375,
0.449462890625,
0.880859375,
0.449462890625,
0.880859375,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.118408203125,
0.46435546875,
0.484375,
0.46435546875,
0.484375,
0.47802734375,
0.118408203125,
0.47802734375
],
[
0.51123046875,
0.465576171875,
0.880859375,
0.465576171875,
0.880859375,
0.478515625,
0.51123046875,
0.478515625
],
[
0.118408203125,
0.48095703125,
0.490234375,
0.48095703125,
0.490234375,
0.49462890625,
0.118408203125,
0.49462890625
],
[
0.509765625,
0.482177734375,
0.87890625,
0.482177734375,
0.87890625,
0.495361328125,
0.509765625,
0.495361328125
],
[
0.118408203125,
0.496337890625,
0.489501953125,
0.496337890625,
0.489501953125,
0.51025390625,
0.118408203125,
0.51025390625
],
[
0.509765625,
0.496337890625,
0.87890625,
0.49755859375,
0.87890625,
0.5126953125,
0.509765625,
0.51123046875
],
[
0.118408203125,
0.5126953125,
0.486083984375,
0.5126953125,
0.486083984375,
0.525390625,
0.118408203125,
0.525390625
],
[
0.51318359375,
0.515625,
0.87890625,
0.515625,
0.87890625,
0.529296875,
0.51318359375,
0.529296875
],
[
0.1175537109375,
0.529296875,
0.484375,
0.529296875,
0.484375,
0.5419921875,
0.1175537109375,
0.5419921875
],
[
0.51123046875,
0.53076171875,
0.880859375,
0.53076171875,
0.880859375,
0.54443359375,
0.51123046875,
0.54443359375
],
[
0.1175537109375,
0.54345703125,
0.486083984375,
0.54345703125,
0.486083984375,
0.55712890625,
0.1175537109375,
0.55712890625
],
[
0.51318359375,
0.54638671875,
0.8818359375,
0.54638671875,
0.8818359375,
0.56005859375,
0.51318359375,
0.56005859375
],
[
0.1175537109375,
0.56005859375,
0.484375,
0.55859375,
0.484375,
0.57373046875,
0.1175537109375,
0.57470703125
],
[
0.51318359375,
0.5625,
0.8818359375,
0.5625,
0.8818359375,
0.576171875,
0.51318359375,
0.576171875
],
[
0.118408203125,
0.57666015625,
0.48779296875,
0.57666015625,
0.48779296875,
0.59033203125,
0.118408203125,
0.59033203125
],
[
0.51318359375,
0.5791015625,
0.87744140625,
0.5791015625,
0.87744140625,
0.5888671875,
0.51318359375,
0.5888671875
],
[
0.118408203125,
0.59326171875,
0.486083984375,
0.59326171875,
0.486083984375,
0.60693359375,
0.118408203125,
0.60693359375
],
[
0.509765625,
0.59326171875,
0.87890625,
0.59326171875,
0.87890625,
0.60693359375,
0.509765625,
0.60693359375
],
[
0.118408203125,
0.609375,
0.486083984375,
0.609375,
0.486083984375,
0.62255859375,
0.118408203125,
0.62255859375
],
[
0.5146484375,
0.6123046875,
0.87890625,
0.6123046875,
0.87890625,
0.62255859375,
0.5146484375,
0.62255859375
],
[
0.118408203125,
0.6259765625,
0.489501953125,
0.6259765625,
0.489501953125,
0.6396484375,
0.118408203125,
0.6396484375
],
[
0.51318359375,
0.626953125,
0.86376953125,
0.626953125,
0.86376953125,
0.64013671875,
0.51318359375,
0.64013671875
],
[
0.1175537109375,
0.64111328125,
0.19482421875,
0.64111328125,
0.19482421875,
0.65625,
0.1175537109375,
0.65625
],
[
0.51123046875,
0.642578125,
0.880859375,
0.6435546875,
0.880859375,
0.65771484375,
0.51123046875,
0.65673828125
],
[
0.515625,
0.6611328125,
0.87890625,
0.6611328125,
0.87890625,
0.67138671875,
0.515625,
0.67138671875
],
[
0.1175537109375,
0.67138671875,
0.156982421875,
0.67138671875,
0.156982421875,
0.68310546875,
0.1175537109375,
0.68310546875
],
[
0.151123046875,
0.67041015625,
0.2568359375,
0.67041015625,
0.2568359375,
0.681640625,
0.151123046875,
0.681640625
],
[
0.5146484375,
0.67578125,
0.880859375,
0.67578125,
0.880859375,
0.689453125,
0.5146484375,
0.689453125
],
[
0.51318359375,
0.6904296875,
0.87890625,
0.6904296875,
0.87890625,
0.70361328125,
0.51318359375,
0.70361328125
],
[
0.1175537109375,
0.69677734375,
0.486083984375,
0.6953125,
0.486083984375,
0.71044921875,
0.1175537109375,
0.71142578125
],
[
0.1175537109375,
0.71142578125,
0.484375,
0.71142578125,
0.484375,
0.724609375,
0.1175537109375,
0.724609375
],
[
0.51318359375,
0.7080078125,
0.880859375,
0.7080078125,
0.880859375,
0.7216796875,
0.51318359375,
0.7216796875
],
[
0.118408203125,
0.728515625,
0.486083984375,
0.728515625,
0.486083984375,
0.7421875,
0.118408203125,
0.7421875
],
[
0.51318359375,
0.7236328125,
0.84033203125,
0.7236328125,
0.84033203125,
0.73681640625,
0.51318359375,
0.73681640625
],
[
0.118408203125,
0.7451171875,
0.48193359375,
0.7451171875,
0.48193359375,
0.755859375,
0.118408203125,
0.755859375
],
[
0.53076171875,
0.7421875,
0.8818359375,
0.7421875,
0.8818359375,
0.755859375,
0.53076171875,
0.755859375
],
[
0.12005615234375,
0.7626953125,
0.486083984375,
0.7626953125,
0.486083984375,
0.7724609375,
0.12005615234375,
0.7724609375
],
[
0.5146484375,
0.76025390625,
0.87890625,
0.76025390625,
0.87890625,
0.77099609375,
0.5146484375,
0.77099609375
],
[
0.118408203125,
0.77783203125,
0.48779296875,
0.77783203125,
0.48779296875,
0.79150390625,
0.118408203125,
0.79150390625
],
[
0.51318359375,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.51318359375,
0.7880859375
],
[
0.118408203125,
0.79345703125,
0.489501953125,
0.79345703125,
0.489501953125,
0.80712890625,
0.118408203125,
0.80712890625
],
[
0.51123046875,
0.7890625,
0.880859375,
0.79150390625,
0.880859375,
0.8046875,
0.51123046875,
0.802734375
],
[
0.118408203125,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.118408203125,
0.82373046875
],
[
0.51123046875,
0.8056640625,
0.880859375,
0.80712890625,
0.880859375,
0.82177734375,
0.51123046875,
0.82080078125
],
[
0.1175537109375,
0.826171875,
0.484375,
0.826171875,
0.484375,
0.8388671875,
0.1175537109375,
0.8388671875
],
[
0.51123046875,
0.82177734375,
0.880859375,
0.82373046875,
0.880859375,
0.83837890625,
0.51123046875,
0.8359375
],
[
0.118408203125,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.8388671875,
0.880859375,
0.8388671875,
0.880859375,
0.8525390625,
0.51318359375,
0.8525390625
],
[
0.118408203125,
0.85791015625,
0.484375,
0.85791015625,
0.484375,
0.87158203125,
0.118408203125,
0.87158203125
],
[
0.51123046875,
0.85400390625,
0.8818359375,
0.85498046875,
0.8818359375,
0.869140625,
0.51123046875,
0.8681640625
],
[
0.515625,
0.87255859375,
0.53076171875,
0.87255859375,
0.53076171875,
0.8828125,
0.515625,
0.8828125
],
[
0.14111328125,
0.8818359375,
0.484375,
0.8818359375,
0.484375,
0.8955078125,
0.14111328125,
0.8955078125
],
[
0.1175537109375,
0.8935546875,
0.484375,
0.89453125,
0.484375,
0.90771484375,
0.1175537109375,
0.90673828125
],
[
0.1175537109375,
0.90625,
0.4189453125,
0.90625,
0.4189453125,
0.91943359375,
0.1175537109375,
0.91943359375
],
[
0.53564453125,
0.90625,
0.8251953125,
0.90625,
0.8251953125,
0.91943359375,
0.53564453125,
0.91943359375
]
] |
[
"第二个标准要求H川小于值XMATHX1。直",
"",
"结果（附录EF）。GPT-3的进一步实验。5位",
"观上，它要求邻居就传入实例的标签达成一致",
"老师，最初的k-NN学生，以及情绪分析数据",
"超参数调优：这里有三个超参数，邻居数k，",
"集（LMR）也证实了先前实验的结论（附录E",
"阈值tc，th。考虑到最初每个类有3个示例，",
"G",
"我们将k=5修复为实际选择。对于每个指示",
"在以后的工作中，我们计划尝试更多的数据",
"性入值(0。05,0。1、0。2、0。3)，我们在少射开",
"集和任务（e。g。，问题回答），并建议入的自适",
"发集（Section3）上采用贝叶斯优化，以确定",
"应策略，以便在缓存较冷时允许更高的OpEx",
"最大化Φ（折扣精度)的两个阈值的最优组合。",
"成本（更频繁地调用教师)，并且以后更有选择",
"我们让tc在[0,2]中，tH在[0,4.34]中。3在首",
"性（更不频繁地调用教师）。我们还计划通过我",
"先对两个阈值的值范围执行10×10网格搜索",
"们可以信任老师的回答的指标来增强OCaTS",
"之后，我们使用Optuna的（Akibaetal，2019）",
"（例如。g。，老师的信心）。最后，我们打算纳入",
"实现树结构参数估计器（TSPE）算法（Bergstra",
"更多的财务指标（例如。g。，学生成本)在评估",
"etal.,2011)作为起点。得到的等高线图和每个",
"措施的折扣版本中，并研究更复杂的策略（例",
"入值的两个值的最佳值可以在附录E中找",
"如。g。，博奔论，强化学习）来选择决定何时",
"到",
"信任学生或打电话给老师的阈值",
"结果：我们使用相同的传人实例（Banking77",
"的原始测试集）和入特定的调谐阈值te，th",
"5Limitations",
"",
"来评估四个指示性入值中每个值的OCaTS。如",
"图所示。2，OCaTS成功地管理了老师与。精",
"这项工作的主要范围是提出一个灵活的框架",
"度。图2（左）显示，随着折扣因子入的增加，",
"（OCaTS），该框架将允许中小型企业在将商业",
"给老师打的电话更少。InFig。E2（中），我们",
"LLM纳人其解决方案时降低运营成本。我们",
"看到为此OpEx救济牺牲了多少准确性。特别",
"只考虑了教师的两个实例（GPT-4、GPT-3。5）",
"是，对于入=0.05，OCaTS的准确性非常接近",
"和两个任务（意图识别，情感分析）中的学生",
"GPT-4老师的准确性，在0的范围内。37个百",
"（k-NN，MLP）的两个实例化，为将来的工作",
"分点（83。05%vS。82。整个测试集的68个%），",
"留下进一步的实例化。虽然LLM像GPT-4和",
"同时仅针对传入实例的1/3调用老师（3080个",
"GPT-3。5原则上可以用于零拍推理，我们考",
"中的1050个）。对于入的较高值，我们看到预",
"虑了在上下文中学习，每个班级有几个演示示",
"期的准确性下降，以实现对教师的呼叫数量越",
"例。这些例子中手动选择是多样的，并表示相",
"来越少。图2（右）显示OCaTS的折扣精度Φ",
"应的类。这在一定程度上是现实的；中小企业",
"（实线，每个入值一个）总是明显高于总是呼叫",
"经常向客户索取少量的实例，但这些实例的质",
"GPT-4老师的相应折扣精度（虚线）。因此，如",
"量并不总是有保证的。此外我们使用的测试",
"果考虑到运营成本，OCaTS显然比总是打电话",
"集（来自Banking77和LMR）是平衡的，因此",
"给老师要好。随着入的增加，差异增加（有利",
"不完全现实。但是，我们对传入（测试）实例",
"于OCaTS)，i。e。，随着降低运营成本变得越",
"的流进行了洗牌，因此，它们的类不会以统",
"来越重要",
"的方式到达。此外，为了调整t。和t，我们",
"使用了从原始训练数据中提取的开发集。这样",
"4",
"Conclusions",
"的开发集在实践中并不总是可用的，但是我们",
"为了分析而使用它。有兴趣的中小企业可以使",
"我们引人了一个在线成本感知师生框架",
"(OCaTS），通过缓存商业LLM的响应和培",
"用我们的分析，作为其应用程序的起点，并减",
"训廉价的本地学生来帮助中小企业降低运营成",
"少为tc和tH找到合适值所需的试验次数",
"本。我们还引入了通用评估措施的折扣版本",
"另一个限制是，Φ仅考虑了调用老师（p）",
"使中小企业能够量化LLM呼叫和用户挫败感",
"的成本，并间接考虑了用户的挫败感，如性",
"之间的权衡。通过使用GPT-4老师和k-NN学",
"能下降所暗示的。更详细的分析还将纳人学",
"生实例化OCaTS，并尝试使用来自银行领域",
"生成本和其他可能具有不同权重的财务指标；",
"（Banking77）的意图识别数据集，我们表明，对",
"OCaTS可以轻松地在该方向上扩展。最后，我",
"老师的调用可以显着减少（减少1/3），而性能",
"们没有与现有的缓存库进行比较，例如。g。",
"仅略有下降（0。37个百分点）。在同一数据集",
"GPTCache。4与OCaTS相比，这些库相当简",
"上对MLP学生进行的其他实验得出了相同的",
"单且不那么灵活，可以与各种师生设置一起使",
"用",
"3 The maximum value of H with 77 classes is 4.34, when",
"using natural logarithms.The upper bound of te was chosen",
"based on initial experiments on development data.",
"thttps://github.com/zilliztech/GPTCache"
] |
{
"arxiv_id": "2310.13395",
"language": "zh",
"page_idx": 4
}
|
||
2206.11484-zh-2
|
2206.11484-zh-2.jpg
|
(Nadeem et al.,2021)等)。(Cao and Daumé III, 2020) 在最近的 150 篇共引用解析论文中提供了顺规范性检查。特别是，他们发现绝大多数论文将语言性别与社会性别混为一谈，并假设社会性别是二元的。他们发现只有一篇论文在共指解析的上下文中明确考虑了他们/他们的人称代词。最近对 NLP (Blodgett et al., 2020) 中的偏见进行了严格的调查，引起了有效的方法学关注，但并未解决在 NLP 偏见研究中包括或排除哪些社会偏见的问题
令人鼓舞的是，有些作品确实研究了大型语言模型中的反酷儿偏见。(Nangia et al., 2020) 将性取向作为其偏见类别之一。但是，他们将同性恋视为“同性恋/非同性恋”二进制文件，并且不考虑影响 LGBTQ 社区特定子组的刻板印象和偏见。(Czarnowska et al.,2021)以相对细微的方式处理酷儿，在“女同性恋”和“代理人”等特定身份的层面上检查模型的公平性。“但是，他们仍然对所有酷儿身份使用相同的模板句子集，因此错过了探查特定身份定型模型的机会。在当今的 NLP 偏见文献中， 缺乏对大型语言模型编码哪些反酷儿刻板印象的探索。此外，现有的少数文献未能解决对LGBTQ+ 社区中边缘化群体的偏见(例如。g。同性恋恐惧症、双性恋恐惧症、跨性别症等。), 使得 NLP 系统中抗 LGBTQ+ 偏差的交叉分析问题没有得到解决
## 3 $\textbf{Methods}$
3.1 Data Collection
我们假设在少数群体成员产生的语言数据上训练的 NLP 模型导致对该社区的偏见较小。作为案例研究，我们收集了1) Twitter上LGBTQ 社区的社交媒体数据，这些数据反映了其成员的语言，以及 2)解决 LGBTQ 问题的美国国家新闻媒体。时间范围固定在 2015 年 1 月 1日至2021年12月31日之间
为了获取 Twitter 数据，我们使用 Twitter 学术 API 对推文进行追溯搜索。对于搜索词，我们使用从“立法追踪器：反跨性别立法”网站$^2$检索的反跨性别法案编号，并集成与反跨性别法案相关的主题标签(i。e。, #transrightsarehumanrights、#transbill、#KillTheBill)。我们还选代分析关于反反式账单的共现标签，以构建更全面的搜索词列表。特别是，我们删除了不
$^2$https://freedomforallamericans.
org/legislative-tracker/
anti-transgender-legislation/
责任的普遍需求上，或者考虑了性别和种族 特定于 LGBTQ+ 社区的主题标签(例如。g。# 偏见，但忽略了同性恋偏见并将性别视为二进 指甲，#化妆，#睫毛)以及那些明确提及色情制 ((Caliskan et al., 2017) , (May et al., 2019) , 销售的人。在 (Bender et al., 2021) 中讨论了过
滤“显式”内容如何无意中使酷儿社区沉默之后，我们选择不删除所有露骨的主题标签。相反，我们只删除了像 #onlyfans, #sellingcontent 和 #transporn 这样的标签，这些标签似乎主要与广告有关，而不是有机的社区讨论。但是， 我们承认我们的方法远非完美，并且可能会导致某些社区成员沉默的意外后果。过滤后，我们的第二次搜索包含了共同出现的主题标签， 总共产生了 2,862,924 条推文。我们称这个语料库为 QueerTwitter
同样，使用开源门户 MediaCloud,我们根据反反式账单号和与反反式账单相关的搜索词进行关键字搜索(i。e。、反反式票据、 反式票据、反反式)或 LGBTQ+ 身份(i。e。,lgbtq,lgbt,gay,lesbian,queer,trans,bisexual)。对于MediaCloud, 我们使用与 LGBTQ+ 身份相关的更一般的搜索字词，因为与 Twitter 相比，使用相同字词时，MediaCloud 产生的结果更少。这导致了 90,495 篇新闻文章的语料库 , 我们称之为 QueerNews
## 3.2 WinoQueer Template Creation
为了以比以前的工作更具体，更细微的方式来衡量反同性恋和反跨性别偏见，我们引人了 WinoQueer, 它是在 (Nangia et al., 2020) 配对句子偏见探测任务之后建模的新基准数据集。据作者所知，我们的基准是第一个探索特定身份的反同性恋和反跨性别偏见的基准，明确解决了同性恋社区中的同性恋恐惧症，双性恋恐惧症，跨性别症和其他有害的刻板印象。我们还认为，这是第一个明确包含他们/他们的人称代词以及 neoprounsey, zie 和 sie 的偏见数据集。这项工作仍在进行中，我们计划在未来开发更完整的 WinoQueer 版本。在这项工作中，我们介绍了我们的基准测试的第一次迭代：WinoQueer-v0
我们首先开发了一组类似于现有的共同参照和性别偏见数据集的模板，但主题集中在同性恋和异性恋关系上。与 Winogender (Rudinger et al., 2018)类似，我们的模式旨在模仿仅由代词或明显性别名称不同的句子对。我们考虑到现有方法的局限性-现有模式不一定能捕捉到性别流动性或性偏好的复杂性。因此，我们的WinoQueer-v0 模式由两种类型组成：
1.Qualitative descriptions: We choose five
adjectives that fall under the following categories— masculine, feminine, positive, and negative. This yields a total of 20 adjec-
|
[
[
0.118408203125,
0.0872802734375,
0.489501953125,
0.0872802734375,
0.489501953125,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.509765625,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.509765625,
0.09979248046875
],
[
0.118408203125,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51123046875,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51123046875,
0.11639404296875
],
[
0.12005615234375,
0.119384765625,
0.47607421875,
0.119384765625,
0.47607421875,
0.133056640625,
0.12005615234375,
0.133056640625
],
[
0.5146484375,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.87744140625,
0.1353759765625,
0.87744140625,
0.1484375,
0.51318359375,
0.1484375
],
[
0.118408203125,
0.15087890625,
0.48779296875,
0.15087890625,
0.48779296875,
0.1644287109375,
0.118408203125,
0.1644287109375
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.12176513671875,
0.16748046875,
0.48779296875,
0.16748046875,
0.48779296875,
0.18115234375,
0.12176513671875,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.880859375,
0.16748046875,
0.880859375,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.118408203125,
0.1817626953125,
0.486083984375,
0.1817626953125,
0.486083984375,
0.1953125,
0.118408203125,
0.1953125
],
[
0.51123046875,
0.1817626953125,
0.87890625,
0.1817626953125,
0.87890625,
0.1953125,
0.51123046875,
0.1953125
],
[
0.12005615234375,
0.20068359375,
0.486083984375,
0.20068359375,
0.486083984375,
0.21142578125,
0.12005615234375,
0.21142578125
],
[
0.51318359375,
0.1995849609375,
0.87744140625,
0.1995849609375,
0.87744140625,
0.213134765625,
0.51318359375,
0.213134765625
],
[
0.118408203125,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.51123046875,
0.215576171875,
0.880859375,
0.215576171875,
0.880859375,
0.2286376953125,
0.51123046875,
0.2286376953125
],
[
0.118408203125,
0.232177734375,
0.486083984375,
0.232177734375,
0.486083984375,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.509765625,
0.2298583984375,
0.8818359375,
0.23095703125,
0.8818359375,
0.2452392578125,
0.509765625,
0.24462890625
],
[
0.118408203125,
0.2476806640625,
0.489501953125,
0.2476806640625,
0.489501953125,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.5146484375,
0.2493896484375,
0.87890625,
0.2493896484375,
0.87890625,
0.260009765625,
0.5146484375,
0.260009765625
],
[
0.118408203125,
0.26318359375,
0.489501953125,
0.26318359375,
0.489501953125,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.51318359375,
0.26416015625,
0.8818359375,
0.26416015625,
0.8818359375,
0.27783203125,
0.51318359375,
0.27783203125
],
[
0.118408203125,
0.27978515625,
0.339111328125,
0.27978515625,
0.339111328125,
0.29345703125,
0.118408203125,
0.29345703125
],
[
0.51318359375,
0.27978515625,
0.6708984375,
0.27978515625,
0.6708984375,
0.29345703125,
0.51318359375,
0.29345703125
],
[
0.1368408203125,
0.29638671875,
0.484375,
0.29638671875,
0.484375,
0.31005859375,
0.1368408203125,
0.31005859375
],
[
0.5322265625,
0.295654296875,
0.880859375,
0.295654296875,
0.880859375,
0.31005859375,
0.5322265625,
0.31005859375
],
[
0.12005615234375,
0.312255859375,
0.48779296875,
0.312255859375,
0.48779296875,
0.325927734375,
0.12005615234375,
0.325927734375
],
[
0.51123046875,
0.312255859375,
0.880859375,
0.312255859375,
0.880859375,
0.325927734375,
0.51123046875,
0.325927734375
],
[
0.12005615234375,
0.330078125,
0.486083984375,
0.330078125,
0.486083984375,
0.34033203125,
0.12005615234375,
0.34033203125
],
[
0.51318359375,
0.328857421875,
0.8818359375,
0.328857421875,
0.8818359375,
0.342529296875,
0.51318359375,
0.342529296875
],
[
0.118408203125,
0.344482421875,
0.48779296875,
0.344482421875,
0.48779296875,
0.358154296875,
0.118408203125,
0.358154296875
],
[
0.51318359375,
0.344482421875,
0.88330078125,
0.344482421875,
0.88330078125,
0.358154296875,
0.51318359375,
0.358154296875
],
[
0.118408203125,
0.361083984375,
0.48193359375,
0.361083984375,
0.48193359375,
0.371826171875,
0.118408203125,
0.371826171875
],
[
0.51123046875,
0.35986328125,
0.87890625,
0.35986328125,
0.87890625,
0.37353515625,
0.51123046875,
0.37353515625
],
[
0.12005615234375,
0.377197265625,
0.48779296875,
0.377197265625,
0.48779296875,
0.39013671875,
0.12005615234375,
0.39013671875
],
[
0.51318359375,
0.377197265625,
0.880859375,
0.377197265625,
0.880859375,
0.39013671875,
0.51318359375,
0.39013671875
],
[
0.1175537109375,
0.392578125,
0.486083984375,
0.391357421875,
0.486083984375,
0.40625,
0.1175537109375,
0.40673828125
],
[
0.5146484375,
0.392578125,
0.87890625,
0.392578125,
0.87890625,
0.40673828125,
0.5146484375,
0.40673828125
],
[
0.118408203125,
0.4091796875,
0.486083984375,
0.4091796875,
0.486083984375,
0.4228515625,
0.118408203125,
0.4228515625
],
[
0.51318359375,
0.4091796875,
0.880859375,
0.4091796875,
0.880859375,
0.4228515625,
0.51318359375,
0.4228515625
],
[
0.1175537109375,
0.42333984375,
0.484375,
0.424560546875,
0.484375,
0.439453125,
0.1175537109375,
0.438232421875
],
[
0.51318359375,
0.424560546875,
0.87890625,
0.424560546875,
0.87890625,
0.438232421875,
0.51318359375,
0.438232421875
],
[
0.118408203125,
0.441162109375,
0.484375,
0.441162109375,
0.484375,
0.454833984375,
0.118408203125,
0.454833984375
],
[
0.51318359375,
0.4423828125,
0.62109375,
0.4423828125,
0.62109375,
0.45361328125,
0.51318359375,
0.45361328125
],
[
0.118408203125,
0.457763671875,
0.484375,
0.457763671875,
0.484375,
0.471435546875,
0.118408203125,
0.471435546875
],
[
0.118408203125,
0.473876953125,
0.484375,
0.473876953125,
0.484375,
0.48681640625,
0.118408203125,
0.48681640625
],
[
0.51123046875,
0.468505859375,
0.8017578125,
0.468505859375,
0.8017578125,
0.482177734375,
0.51123046875,
0.482177734375
],
[
0.118408203125,
0.4892578125,
0.489501953125,
0.4892578125,
0.489501953125,
0.5029296875,
0.118408203125,
0.5029296875
],
[
0.51318359375,
0.4912109375,
0.880859375,
0.4912109375,
0.880859375,
0.5048828125,
0.51318359375,
0.5048828125
],
[
0.1158447265625,
0.50439453125,
0.490234375,
0.505859375,
0.490234375,
0.52099609375,
0.1158447265625,
0.5185546875
],
[
0.51318359375,
0.50732421875,
0.880859375,
0.50732421875,
0.880859375,
0.52099609375,
0.51318359375,
0.52099609375
],
[
0.12005615234375,
0.521484375,
0.489501953125,
0.521484375,
0.489501953125,
0.53515625,
0.12005615234375,
0.53515625
],
[
0.51318359375,
0.5224609375,
0.880859375,
0.5224609375,
0.880859375,
0.5361328125,
0.51318359375,
0.5361328125
],
[
0.12005615234375,
0.53759765625,
0.48779296875,
0.53759765625,
0.48779296875,
0.55029296875,
0.12005615234375,
0.55029296875
],
[
0.509765625,
0.53759765625,
0.880859375,
0.5380859375,
0.880859375,
0.552734375,
0.509765625,
0.5517578125
],
[
0.12005615234375,
0.55419921875,
0.26611328125,
0.55419921875,
0.26611328125,
0.56787109375,
0.12005615234375,
0.56787109375
],
[
0.51318359375,
0.5546875,
0.880859375,
0.5546875,
0.880859375,
0.568359375,
0.51318359375,
0.568359375
],
[
0.5146484375,
0.57080078125,
0.880859375,
0.57080078125,
0.880859375,
0.58447265625,
0.5146484375,
0.58447265625
],
[
0.11419677734375,
0.5830078125,
0.2301025390625,
0.5830078125,
0.2301025390625,
0.59716796875,
0.11419677734375,
0.59716796875
],
[
0.5146484375,
0.58740234375,
0.87744140625,
0.58740234375,
0.87744140625,
0.59716796875,
0.5146484375,
0.59716796875
],
[
0.51123046875,
0.6015625,
0.87890625,
0.6015625,
0.87890625,
0.615234375,
0.51123046875,
0.615234375
],
[
0.1158447265625,
0.609375,
0.287109375,
0.609375,
0.287109375,
0.62255859375,
0.1158447265625,
0.62255859375
],
[
0.5146484375,
0.62060546875,
0.87890625,
0.62060546875,
0.87890625,
0.63134765625,
0.5146484375,
0.63134765625
],
[
0.118408203125,
0.6318359375,
0.48779296875,
0.6318359375,
0.48779296875,
0.6455078125,
0.118408203125,
0.6455078125
],
[
0.51318359375,
0.63525390625,
0.880859375,
0.63525390625,
0.880859375,
0.6494140625,
0.51318359375,
0.6494140625
],
[
0.118408203125,
0.64794921875,
0.48779296875,
0.64794921875,
0.48779296875,
0.6611328125,
0.118408203125,
0.6611328125
],
[
0.51318359375,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.6650390625,
0.51318359375,
0.6650390625
],
[
0.1175537109375,
0.662109375,
0.486083984375,
0.662109375,
0.486083984375,
0.67578125,
0.1175537109375,
0.67578125
],
[
0.5146484375,
0.66943359375,
0.87890625,
0.66943359375,
0.87890625,
0.6796875,
0.5146484375,
0.6796875
],
[
0.118408203125,
0.6787109375,
0.484375,
0.6787109375,
0.484375,
0.6923828125,
0.118408203125,
0.6923828125
],
[
0.51318359375,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69775390625,
0.51318359375,
0.69775390625
],
[
0.118408203125,
0.6953125,
0.484375,
0.6953125,
0.484375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51318359375,
0.69970703125,
0.66015625,
0.69970703125,
0.66015625,
0.71337890625,
0.51318359375,
0.71337890625
],
[
0.118408203125,
0.71142578125,
0.48779296875,
0.71142578125,
0.48779296875,
0.724609375,
0.118408203125,
0.724609375
],
[
0.53076171875,
0.71630859375,
0.880859375,
0.71630859375,
0.880859375,
0.72998046875,
0.53076171875,
0.72998046875
],
[
0.1175537109375,
0.72705078125,
0.345947265625,
0.72705078125,
0.345947265625,
0.74072265625,
0.1175537109375,
0.74072265625
],
[
0.51318359375,
0.73095703125,
0.880859375,
0.732421875,
0.880859375,
0.74658203125,
0.51318359375,
0.7451171875
],
[
0.1351318359375,
0.74365234375,
0.486083984375,
0.74365234375,
0.486083984375,
0.75732421875,
0.1351318359375,
0.75732421875
],
[
0.509765625,
0.74658203125,
0.880859375,
0.7490234375,
0.880859375,
0.76318359375,
0.509765625,
0.76123046875
],
[
0.12005615234375,
0.76123046875,
0.484375,
0.76123046875,
0.484375,
0.771484375,
0.12005615234375,
0.771484375
],
[
0.51318359375,
0.76416015625,
0.880859375,
0.76416015625,
0.880859375,
0.77783203125,
0.51318359375,
0.77783203125
],
[
0.118408203125,
0.77685546875,
0.48779296875,
0.77685546875,
0.48779296875,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.51318359375,
0.779296875,
0.87890625,
0.77978515625,
0.87890625,
0.79443359375,
0.51318359375,
0.79345703125
],
[
0.1158447265625,
0.79150390625,
0.489501953125,
0.7919921875,
0.489501953125,
0.80712890625,
0.1158447265625,
0.8056640625
],
[
0.51318359375,
0.79638671875,
0.87890625,
0.79638671875,
0.87890625,
0.81005859375,
0.51318359375,
0.81005859375
],
[
0.118408203125,
0.80859375,
0.48779296875,
0.80859375,
0.48779296875,
0.822265625,
0.118408203125,
0.822265625
],
[
0.5146484375,
0.8125,
0.880859375,
0.8125,
0.880859375,
0.826171875,
0.5146484375,
0.826171875
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51318359375,
0.8291015625,
0.80029296875,
0.8291015625,
0.80029296875,
0.8427734375,
0.51318359375,
0.8427734375
],
[
0.118408203125,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.52294921875,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.52294921875,
0.87060546875
],
[
0.5498046875,
0.87353515625,
0.880859375,
0.87353515625,
0.880859375,
0.88720703125,
0.5498046875,
0.88720703125
],
[
0.1351318359375,
0.8798828125,
0.424072265625,
0.8828125,
0.424072265625,
0.89794921875,
0.1351318359375,
0.89453125
],
[
0.5498046875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.5498046875,
0.90283203125
],
[
0.12005615234375,
0.8955078125,
0.333251953125,
0.8955078125,
0.333251953125,
0.90625,
0.12005615234375,
0.90625
],
[
0.118408203125,
0.90625,
0.38037109375,
0.90625,
0.38037109375,
0.91943359375,
0.118408203125,
0.91943359375
],
[
0.54833984375,
0.90625,
0.880859375,
0.90625,
0.880859375,
0.91943359375,
0.54833984375,
0.91943359375
]
] |
[
"责任的普遍需求上，或者考虑了性别和种族",
"特定于LGBTQ+社区的主题标签（例如。g。#",
"偏见，但忽略了同性恋偏见并将性别视为二进",
"指甲，#化妆，#睫毛）以及那些明确提及色情",
"制 ((Caliskan et al., 2017)，(May et al., 2019)",
"销售的人。在（Benderetal.2021）中讨论了过",
"(Nadeem et al.,2021)等)。(Cao and Daumé III,",
"滤“显式”内容如何无意中使酷儿社区沉默之",
"2020）在最近的150篇共引用解析论文中提供",
"后，我们选择不删除所有露骨的主题标签。相",
"了顺规范性检查。特别是，他们发现绝大多数",
"反，我们只删除了像#onlyfans，#sellingcontent",
"论文将语言性别与社会性别混为一谈，并假设",
"和#transporn这样的标签，这些标签似乎主要",
"社会性别是二元的。他们发现只有一篇论文在",
"与广告有关，而不是有机的社区讨论。但是",
"共指解析的上下文中明确考虑了他们/他们的",
"我们承认我们的方法远非完美，并且可能会导",
"人称代词。最近对NLP(Blodgettetal.,2020）中",
"致某些社区成员沉默的意外后果。。过滤后，我",
"的偏见进行了严格的调查，引起了有效的方法",
"们的第二次搜索包含了共同出现的主题标签，",
"学关注，但并未解决在NLP偏见研究中包括",
"总共产生了2,862,924条推文。我们称这个语",
"或排除哪些社会偏见的问题",
"料库为QueerTwitter",
"令人鼓舞的是，有些作品确实研究了大型语",
"同样，使用开源门户MediaCloud，我们根",
"言模型中的反酷儿偏见。（Nangiaetal.，2020)",
"据反反式账单号和与反反式账单相关的搜",
"将性取向作为其偏见类别之一。但是，他们",
"索词进行关键字搜索(i。e。、反反式票据、",
"将同性恋视为“同性恋/非同性恋”二进制文",
"反式票据、反反式）或LGBTQ+身份(i。e。",
"件，并月不考虑影响LGBTO社区特定子组的",
",Igbtq,lgbt,gay,lesbian,queer,trans,bisexual)。对于",
"刻板印象和偏见。（Czarnowskaetal.，2021）以",
"MediaCloud，我们使用与LGBTQ+身份相关的",
"相对细微的方式处理酷儿，在“女同性恋”和",
"更一般的搜索字词，因为与Twitter相比，使用",
"“代理人”等特定身份的层面上检查模型的公",
"相同字词时，MediaCloud产生的结果更少。这",
"平性。“但是，他们仍然对所有酷儿身份使用",
"导致了90,495篇新闻文章的语料库，我们称之",
"相同的模板句子集，因此错过了探查特定身份",
"为QueerNews",
"定型模型的机会。在当今的NLP偏见文献中，",
"缺乏对大型语言模型编码哪些反酷儿刻板印",
"3.2WinoQueer Template Creation",
"象的探索。此外，现有的少数文献未能解决对",
"为了以比以前的工作更具体，更细微的方式",
"LGBTQ+社区中边缘化群体的偏见（例如。g。",
"来衡量反同性恋和反跨性别偏见，我们引入",
"同性恋恐惧症、双性恋恐惧症、跨性别症等。），",
"了WinoQueer，它是在（Nangia etal.,2020）配",
"使得NLP系统中抗LGBTQ+偏差的交叉分析",
"对句子偏见探测任务之后建模的新基准数据",
"问题没有得到解决",
"集。据作者所知，我们的基准是第一个探索特",
"定身份的反同性恋和反跨性别偏见的基准，明",
"3Methods",
"确解决了同性恋社区中的同性恋恐惧症，双性",
"恋恐惧症，跨性别症和其他有害的刻板印象。",
"3.1Data Collection",
"我们还认为，这是第一个明确包含他们/他们",
"我们假设在少数群体成员产生的语言数据上训",
"的人称代词以及neoprounsey，zie和sie的偏",
"练的NLP模型导致对该社区的偏见较小。作",
"见数据集。这项工作仍在进行中，我们计划在",
"为案例研究，我们收集了1）Twitter上LGBTQ",
"未来开发更完整的WinoQueer版本。在这项工",
"社区的社交媒体数据，这些数据反映了其成员",
"作中，我们介绍了我们的基准测试的第一次迭",
"的语言，以及2）解决LGBTQ问题的美国国",
"代：WinoQueer-vo",
"家新闻媒体。时间范围固定在2015年1月1",
"我们首先开发了一组类似于现有的共同参照",
"日至2021年12月31日之间",
"和性别偏见数据集的模板，但主题集中在同性",
"为了获取Twitter数据，我们使用Twitter学",
"恋和异性恋关系上。与Winogender（Rudinger",
"术API对推文进行追溯搜索。对于搜索词，我",
"etal.,2018)类似，我们的模式旨在模仿仅由代",
"们使用从“立法追踪器：反跨性别立法”网站",
"词或明显性别名称不同的句子对。我们考虑到",
"2检索的反跨性别法案编号，并集成与反跨性",
"现有方法的局限性-现有模式不一定能捕捉到",
"别法案相关的主题标签（i。e。、#transrightsare-",
"性别流动性或性偏好的复杂性。因此，我们的",
"humanrights、#transbill、#KillTheBill)。我们还",
"WinoQueer-vo模式由两种类型组成：",
"迭代分析关于反反式账单的共现标签，以构建",
"更全面的搜索词列表。特别是，我们删除了不",
"1. Qualitative descriptions:We choose five",
"adjectives that fall under the following",
"2https://freedomforallamericans.",
"categories masculine,feminine, positive,",
"org/legislative-tracker/",
"anti-transgender-legislation/",
"and negative. This yields a total of 20 adjec-"
] |
{
"arxiv_id": "2206.11484",
"language": "zh",
"page_idx": 2
}
|
||
2206.11484-zh-3
|
2206.11484-zh-3.jpg
|
tives.
2. Pairwise interaction: We consider schemas
that take the Cartesian product of two indiinfluence the predictive scoring. We use three templates.
然后，我们使用一组代词和性别名称插入这然后名称 vs。名称。示例如表$\boxed{\mathrm{E}}$1 WinoQueer 该方法引人了一种新颖的伪对数似然度量，用于大型语言模型中的偏差。他们的数据集由成对的刻板印象和刻板印象较少的句子组成。偏误分数是指较刻板的句子的可能性高于较不刻板的句子的可能性的例子的百分比。一个完美的分数是 0。5、i。e。这两个句子同样可能。直觉上，分数大于 0。5 表示 LLM 发现更多的刻板印象句子正常，而较少的刻板印象句子异常或令人惊讶。我们的工作与他们的不同，因为它调查了 LGBTQ+ 社区中特定子组的刻板印象。WinoQueer 基准的未来版本将包括涵盖特定 LGBTQ+ 身份和刻板印象的模板，以及来自更多不同来源的刻板印象，如众包或社区调查
# 3.3 Model Fine-tuning and Analysis
首先，我们使用我们收集的语料库对几个广泛使用的大型预训练语言模型进行微调： $BERT\_base, BERT\_large (Devlin et al., 2019)$ , SpanBERT\_base 和 SpanBERT\_large (Joshi et al., 2020) 。我们为每个模型制作两个版本： 一个在 QueerTwitter 上进行了微调，另一个在 QueerNews 上进行了微调。然后，我们在WinoQueer-v0 数据集上测试了4个现成模型和 8 个微调模型。我们的假设是：
H1: Both the fine-tuned models (BERT and Span-
BERT) will show less anti-queer bias than
their corresponding off-the-shelf models.
H2: The model trained on QueerTwitter data will
show less preference for straight over queer sentences the model trained on QueerNews.
我们还测试了我们的模型的共引用分辨率。特别是，我们评估了 OntoNotes (Weischedel et al., 2017) 共同引用任务和性别歧义代词
LGBTQ+ 语料库的微调不会显著降低 LLM 在下游 NLP 任务上的性能。其次，它探讨了这种微调如何影响二元性别偏见指标的表现。我们最初假设明确包含酷儿观点的模型将显示较少
viduals, such as x and y held hands. The idea 的二元性别偏见。然而，我们对这一假设的结
here is to test whether the pairings of x and y 果充其量是不确定的，如下所述
## 4 $\textbf{Results}$
我们在两个设置中评估我们的模型：掩码语言
些模式。对于成对交互，我们考虑代词与。代词，建模，以量化模型权重和共引用分辨率中编码
的潜在刻板印象，以检查对酷儿数据的微调如
上的评估遵循 (Nangia et al., 2020) 的评估方法，何影响下游性能和 (二进制) 性别偏见
## 4.1 Masked Language Modeling
对于蒙版语言建模实验，我们检查在 QueerTwitter 上微调的模型是否比在 QueerNews 上微调的模型显示出较少的反酷儿刻板印象的证据。我们在 WinoQueer 数据集上遵循 (Nangia et al.,2020)的评估方法。对于屏蔽语言建模任务，我们测试了4个模型： $BERT\_base、BERT\_large (Devlin et al., 2019)$ 、SpanBERT\_base 和 SpanBERT\_large (Joshi et al.,2020),每个模型都遵循 3 个微调约定：
$\bullet$ Out-of-the-box (no finetuning)
$\bullet$ Finetuned on our LGBTQ+ twitter corpus
$\bullet$ Finetuned on our straight media coverage of
queer issues corpus
表 2显示了我们的 WinoQueer 测试集上的结果。所示的偏差分数是模型对更刻板的(直的)比对较不刻板的(酷儿)句子具有更高的伪对数似然的实例的百分比。不出所料， BERT 表现出明显的反酷儿偏见。两种微调语料库都可以有效地减少偏见，其中 QueerTwitter 对于 BERT\_base 略好，而 QueerNews 对于BERT$\_$large 的表现略好。。正如文献中所记录的那样，大模型显示出更多的偏差，并且比基本模型更难消除偏差。我们发现 SpanBERT 在我们的特定测试集上没有显示出同性恋偏见。然而，重要的是要注意，这并不意味着模型中不存在偏差。在 SpanBERT 上，我们观察到新闻语料库上的微调引入了对直句的更多偏好， 而 Twitter 语料库上的微调则没有引入这种偏好。这似乎表明主流媒体报道使用的语言比Twitter 上的酷儿用户更加恐同，并且 LLM 可以在调整新闻时吸收和编码这种偏见。这些结果支持了我们的假设，即经过 Twitter 调整的
(GAP) 任务(Webster et al.,2018), 这是 NLP 模型显示出的刻板印象的证据较少，因为与直系统中 (二进制) 性别偏见的既定度量。共指 接媒体来源相比，同性恋者以更细微和富有同分析有两个目的。首先，它表明我们提出的对 情心的方式谈论自己
|
[
[
0.154541015625,
0.0885009765625,
0.193115234375,
0.0885009765625,
0.193115234375,
0.1009521484375,
0.154541015625,
0.1009521484375
],
[
0.51318359375,
0.0872802734375,
0.87744140625,
0.0872802734375,
0.87744140625,
0.097412109375,
0.51318359375,
0.097412109375
],
[
0.5146484375,
0.10272216796875,
0.87744140625,
0.10272216796875,
0.87744140625,
0.11639404296875,
0.5146484375,
0.11639404296875
],
[
0.1292724609375,
0.114013671875,
0.48779296875,
0.11517333984375,
0.48779296875,
0.12890625,
0.1292724609375,
0.127685546875
],
[
0.5146484375,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.154541015625,
0.1312255859375,
0.489501953125,
0.1312255859375,
0.489501953125,
0.144287109375,
0.154541015625,
0.144287109375
],
[
0.51318359375,
0.1353759765625,
0.87744140625,
0.1353759765625,
0.87744140625,
0.1484375,
0.51318359375,
0.1484375
],
[
0.154541015625,
0.146728515625,
0.489501953125,
0.146728515625,
0.489501953125,
0.1602783203125,
0.154541015625,
0.1602783203125
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.151123046875,
0.162109375,
0.48779296875,
0.163330078125,
0.48779296875,
0.1776123046875,
0.151123046875,
0.177001953125
],
[
0.51318359375,
0.16748046875,
0.771484375,
0.16748046875,
0.771484375,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.15283203125,
0.1787109375,
0.483642578125,
0.1787109375,
0.483642578125,
0.1923828125,
0.15283203125,
0.1923828125
],
[
0.151123046875,
0.193603515625,
0.2301025390625,
0.1953125,
0.2301025390625,
0.210205078125,
0.151123046875,
0.2078857421875
],
[
0.509765625,
0.1953125,
0.61279296875,
0.1953125,
0.61279296875,
0.210205078125,
0.509765625,
0.210205078125
],
[
0.1368408203125,
0.22265625,
0.48779296875,
0.22265625,
0.48779296875,
0.236328125,
0.1368408203125,
0.236328125
],
[
0.51123046875,
0.2203369140625,
0.880859375,
0.2203369140625,
0.880859375,
0.2340087890625,
0.51123046875,
0.2340087890625
],
[
0.1175537109375,
0.2369384765625,
0.490234375,
0.2381591796875,
0.490234375,
0.2529296875,
0.1175537109375,
0.251708984375
],
[
0.51318359375,
0.2381591796875,
0.87744140625,
0.2381591796875,
0.87744140625,
0.248779296875,
0.51318359375,
0.248779296875
],
[
0.1175537109375,
0.254638671875,
0.48779296875,
0.254638671875,
0.48779296875,
0.268310546875,
0.1175537109375,
0.268310546875
],
[
0.5146484375,
0.254150390625,
0.87744140625,
0.254150390625,
0.87744140625,
0.26416015625,
0.5146484375,
0.26416015625
],
[
0.118408203125,
0.26953125,
0.486083984375,
0.26953125,
0.486083984375,
0.283203125,
0.118408203125,
0.283203125
],
[
0.51318359375,
0.26953125,
0.8251953125,
0.26953125,
0.8251953125,
0.283203125,
0.51318359375,
0.283203125
],
[
0.118408203125,
0.287353515625,
0.484375,
0.287353515625,
0.484375,
0.300537109375,
0.118408203125,
0.300537109375
],
[
0.509765625,
0.29345703125,
0.7841796875,
0.29443359375,
0.7841796875,
0.308837890625,
0.509765625,
0.30810546875
],
[
0.12005615234375,
0.303955078125,
0.484375,
0.303955078125,
0.484375,
0.317626953125,
0.12005615234375,
0.317626953125
],
[
0.1175537109375,
0.320556640625,
0.48193359375,
0.320556640625,
0.48193359375,
0.330810546875,
0.1175537109375,
0.330810546875
],
[
0.51318359375,
0.3154296875,
0.880859375,
0.3154296875,
0.880859375,
0.328857421875,
0.51318359375,
0.328857421875
],
[
0.118408203125,
0.336181640625,
0.484375,
0.336181640625,
0.484375,
0.349853515625,
0.118408203125,
0.349853515625
],
[
0.51123046875,
0.334228515625,
0.87744140625,
0.334228515625,
0.87744140625,
0.344482421875,
0.51123046875,
0.344482421875
],
[
0.118408203125,
0.3515625,
0.486083984375,
0.3515625,
0.486083984375,
0.365234375,
0.118408203125,
0.365234375
],
[
0.51318359375,
0.3486328125,
0.87744140625,
0.3486328125,
0.87744140625,
0.359375,
0.51318359375,
0.359375
],
[
0.118408203125,
0.3681640625,
0.48779296875,
0.3681640625,
0.48779296875,
0.3818359375,
0.118408203125,
0.3818359375
],
[
0.51123046875,
0.364013671875,
0.8818359375,
0.364013671875,
0.8818359375,
0.377685546875,
0.51123046875,
0.377685546875
],
[
0.1175537109375,
0.383056640625,
0.484375,
0.383056640625,
0.484375,
0.396728515625,
0.1175537109375,
0.396728515625
],
[
0.509765625,
0.37890625,
0.87890625,
0.37890625,
0.87890625,
0.393798828125,
0.509765625,
0.393798828125
],
[
0.1175537109375,
0.3984375,
0.484375,
0.399658203125,
0.484375,
0.413330078125,
0.1175537109375,
0.412109375
],
[
0.51123046875,
0.396728515625,
0.8818359375,
0.396728515625,
0.8818359375,
0.410400390625,
0.51123046875,
0.410400390625
],
[
0.12005615234375,
0.41748046875,
0.483642578125,
0.41748046875,
0.483642578125,
0.427490234375,
0.12005615234375,
0.427490234375
],
[
0.51318359375,
0.412109375,
0.880859375,
0.412109375,
0.880859375,
0.42578125,
0.51318359375,
0.42578125
],
[
0.118408203125,
0.431640625,
0.486083984375,
0.431640625,
0.486083984375,
0.4453125,
0.118408203125,
0.4453125
],
[
0.50146484375,
0.427001953125,
0.880859375,
0.427490234375,
0.880859375,
0.4423828125,
0.50146484375,
0.441162109375
],
[
0.118408203125,
0.44775390625,
0.48779296875,
0.44775390625,
0.48779296875,
0.46142578125,
0.118408203125,
0.46142578125
],
[
0.51318359375,
0.44482421875,
0.86962890625,
0.44482421875,
0.86962890625,
0.457763671875,
0.51318359375,
0.457763671875
],
[
0.1175537109375,
0.46435546875,
0.483642578125,
0.46435546875,
0.483642578125,
0.47802734375,
0.1175537109375,
0.47802734375
],
[
0.53076171875,
0.4697265625,
0.77490234375,
0.471435546875,
0.77490234375,
0.486328125,
0.53076171875,
0.48388671875
],
[
0.12005615234375,
0.48095703125,
0.486083984375,
0.48095703125,
0.486083984375,
0.49462890625,
0.12005615234375,
0.49462890625
],
[
0.1158447265625,
0.496337890625,
0.14111328125,
0.496337890625,
0.14111328125,
0.51123046875,
0.1158447265625,
0.51123046875
],
[
0.53076171875,
0.496337890625,
0.85791015625,
0.49755859375,
0.85791015625,
0.5126953125,
0.53076171875,
0.51123046875
],
[
0.1158447265625,
0.52392578125,
0.41650390625,
0.52392578125,
0.41650390625,
0.53759765625,
0.1158447265625,
0.53759765625
],
[
0.53076171875,
0.52392578125,
0.8818359375,
0.52490234375,
0.8818359375,
0.5390625,
0.53076171875,
0.5380859375
],
[
0.12005615234375,
0.5458984375,
0.486083984375,
0.5458984375,
0.486083984375,
0.55859375,
0.12005615234375,
0.55859375
],
[
0.54833984375,
0.54150390625,
0.69287109375,
0.54150390625,
0.69287109375,
0.5556640625,
0.54833984375,
0.5556640625
],
[
0.118408203125,
0.56103515625,
0.48779296875,
0.56103515625,
0.48779296875,
0.57470703125,
0.118408203125,
0.57470703125
],
[
0.53076171875,
0.56787109375,
0.880859375,
0.56787109375,
0.880859375,
0.58056640625,
0.53076171875,
0.58056640625
],
[
0.1175537109375,
0.57763671875,
0.486083984375,
0.57763671875,
0.486083984375,
0.59130859375,
0.1175537109375,
0.59130859375
],
[
0.51123046875,
0.58203125,
0.87890625,
0.58203125,
0.87890625,
0.595703125,
0.51123046875,
0.595703125
],
[
0.12176513671875,
0.59716796875,
0.1376953125,
0.59716796875,
0.1376953125,
0.60498046875,
0.12176513671875,
0.60498046875
],
[
0.1318359375,
0.5927734375,
0.486083984375,
0.5927734375,
0.486083984375,
0.60546875,
0.1318359375,
0.60546875
],
[
0.51123046875,
0.599609375,
0.87890625,
0.599609375,
0.87890625,
0.61328125,
0.51123046875,
0.61328125
],
[
0.1158447265625,
0.60791015625,
0.489501953125,
0.609375,
0.489501953125,
0.62353515625,
0.1158447265625,
0.62255859375
],
[
0.51123046875,
0.6142578125,
0.87890625,
0.6142578125,
0.87890625,
0.62744140625,
0.51123046875,
0.62744140625
],
[
0.124267578125,
0.6259765625,
0.486083984375,
0.6259765625,
0.486083984375,
0.6396484375,
0.124267578125,
0.6396484375
],
[
0.51123046875,
0.6298828125,
0.8818359375,
0.63134765625,
0.8818359375,
0.6455078125,
0.51123046875,
0.64453125
],
[
0.1175537109375,
0.64111328125,
0.486083984375,
0.64111328125,
0.486083984375,
0.65478515625,
0.1175537109375,
0.65478515625
],
[
0.51318359375,
0.64794921875,
0.8818359375,
0.64794921875,
0.8818359375,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.1175537109375,
0.65771484375,
0.484375,
0.65771484375,
0.484375,
0.67138671875,
0.1175537109375,
0.67138671875
],
[
0.51123046875,
0.6630859375,
0.87890625,
0.6630859375,
0.87890625,
0.6767578125,
0.51123046875,
0.6767578125
],
[
0.118408203125,
0.6748046875,
0.38037109375,
0.6748046875,
0.38037109375,
0.6884765625,
0.118408203125,
0.6884765625
],
[
0.51123046875,
0.6787109375,
0.8818359375,
0.6796875,
0.8818359375,
0.693359375,
0.51123046875,
0.6923828125
],
[
0.5146484375,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.5146484375,
0.708984375
],
[
0.11083984375,
0.70166015625,
0.486083984375,
0.70166015625,
0.486083984375,
0.71533203125,
0.11083984375,
0.71533203125
],
[
0.5146484375,
0.7119140625,
0.880859375,
0.7119140625,
0.880859375,
0.7255859375,
0.5146484375,
0.7255859375
],
[
0.15283203125,
0.7197265625,
0.486083984375,
0.7197265625,
0.486083984375,
0.73291015625,
0.15283203125,
0.73291015625
],
[
0.51318359375,
0.728515625,
0.8818359375,
0.728515625,
0.8818359375,
0.7421875,
0.51318359375,
0.7421875
],
[
0.154541015625,
0.736328125,
0.455078125,
0.736328125,
0.455078125,
0.75,
0.154541015625,
0.75
],
[
0.51318359375,
0.74462890625,
0.87890625,
0.74462890625,
0.87890625,
0.75830078125,
0.51318359375,
0.75830078125
],
[
0.11083984375,
0.7626953125,
0.486083984375,
0.7626953125,
0.486083984375,
0.775390625,
0.11083984375,
0.775390625
],
[
0.51123046875,
0.7587890625,
0.87890625,
0.7587890625,
0.87890625,
0.7724609375,
0.51123046875,
0.7724609375
],
[
0.154541015625,
0.77978515625,
0.489501953125,
0.77978515625,
0.489501953125,
0.79345703125,
0.154541015625,
0.79345703125
],
[
0.51318359375,
0.77685546875,
0.8818359375,
0.77685546875,
0.8818359375,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.154541015625,
0.7958984375,
0.476806640625,
0.7958984375,
0.476806640625,
0.81005859375,
0.154541015625,
0.81005859375
],
[
0.5146484375,
0.79345703125,
0.87744140625,
0.79345703125,
0.87744140625,
0.80419921875,
0.5146484375,
0.80419921875
],
[
0.51123046875,
0.80810546875,
0.880859375,
0.80859375,
0.880859375,
0.822265625,
0.51123046875,
0.82177734375
],
[
0.1368408203125,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.8388671875,
0.1368408203125,
0.8388671875
],
[
0.51123046875,
0.826171875,
0.87890625,
0.826171875,
0.87890625,
0.8388671875,
0.51123046875,
0.8388671875
],
[
0.1175537109375,
0.84130859375,
0.486083984375,
0.84130859375,
0.486083984375,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.51123046875,
0.84130859375,
0.87890625,
0.84130859375,
0.87890625,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.12005615234375,
0.859375,
0.486083984375,
0.859375,
0.486083984375,
0.869140625,
0.12005615234375,
0.869140625
],
[
0.5146484375,
0.859375,
0.87890625,
0.859375,
0.87890625,
0.869140625,
0.5146484375,
0.869140625
],
[
0.118408203125,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51318359375,
0.87451171875,
0.87744140625,
0.87451171875,
0.87744140625,
0.88525390625,
0.51318359375,
0.88525390625
],
[
0.118408203125,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.12005615234375,
0.90625,
0.48779296875,
0.90625,
0.48779296875,
0.91943359375,
0.12005615234375,
0.91943359375
],
[
0.51318359375,
0.90625,
0.67822265625,
0.90625,
0.67822265625,
0.91943359375,
0.51318359375,
0.91943359375
]
] |
[
"tives.",
"LGBTO+语料库的微调不会显著降低LLM在",
"下游NLP任务上的性能。其次，它探讨了这种",
"2. Pairwise interaction: We consider schemas",
"微调如何影响二元性别偏见指标的表现。我们",
"that take the Cartesian product of two indi-",
"最初假设明确包含酷儿观点的模型将显示较少",
"viduals, such as x and y held hands. The idea",
"的二元性别偏见。然而，我们对这一假设的结",
"here is to test whether the pairings of x and y",
"果充其量是不确定的，如下所述",
"influence the predictive scoring. We use three",
"templates.",
"4Results",
"然后，我们使用一组代词和性别名称插人这",
"我们在两个设置中评估我们的模型：掩码语言",
"些模式。对于成对交互，我们考虑代词与。代词，",
"建模，以量化模型权重和共引用分辨率中编码",
"然后名称vs。名称。示例如表E1WinoQueer",
"的潜在刻板印象，以检香对酷儿数据的微调如",
"上的评估遵循（Nangiaetal.，2020）的评估方法，",
"何影响下游性能和（二进制）性别偏见",
"该方法引入了一种新颖的伪对数似然度量，用",
"4.1Masked Language Modeling",
"于大型语言模型中的偏差。他们的数据集由成",
"对的刻板印象和刻板印象较少的句子组成。偏",
"对于蒙版语言建模实验，我们检查在Queer",
"误分数是指较刻板的句子的可能性高于较不刻",
"Twitter上微调的模型是否比在QueerNews",
"板的句子的可能性的例子的百分比。一个完美",
"上微调的模型显示出较少的反酷儿刻板印",
"的分数是0。5、i。e。这两个句子同样可能。直",
"象的证据。我们在WinoQueer数据集上遵",
"觉上，分数大于0。5表示LLM发现更多的刻",
"循(Nangia etal.，2020）的评估方法。对于",
"板印象句子正常，而较少的刻板印象句子异常",
"屏蔽语言建模任务，我们测试了4个模型：",
"或令人惊边。我们的工作与他们的不同，因为",
"BERT_base、BERT_large (Devlin et al., 2019)",
"它调查了LGBTQ+社区中特定子组的刻板印",
"、SpanBERT_base和SpanBERT_large(Joshi",
"象。WinoQueer基准的未来版本将包括涵盖特",
"etal.，2020），每个模型都遵循3个微调约定：",
"定LGBTQ+身份和刻板印象的模板，以及来",
".Out-of-the-box (no finetuning)",
"自更多不同来源的刻板印象，如众包或社区调",
"查",
"· Finetuned on our LGBTQ+ twitter corpus",
"3.3Model Fine-tuning and Analysis",
": Finetuned on our straight media coverage of",
"首先，我们使用我们收集的语料库对几个",
"queer issues corpus",
"广泛使用的大型预训练语言模型进行微调：",
"表2显示了我们的WinoQueer测试集上",
"BERT_base, BERT_large (Devlin et al., 2019)",
"的结果。所示的偏差分数是模型对更刻板的",
"",
"SpanBERTbase 和 SpanBERT large Joshi",
"（直的）比对较不刻板的（酷儿）句子具有更",
"etal.2020）。我们为每个模型制作两个版本：",
"高的伪对数似然的实例的百分比。不出所料",
"一个在QueerTwitter上进行了微调，另一个",
"BERT表现出明显的反酷儿偏见。两种微调语",
"在QueerNews上进行了微调。然后，我们在",
"料库都可以有效地减少偏见，其中QueerTwit-",
"WinoQueer-vo数据集上测试了4个现成模型",
"ter对于BERT_base略好，而QueerNews对于",
"和8个微调模型。我们的假设是",
"BERTlarge的表现略好。。正如文献中所记录",
"的那样，大模型显示出更多的偏差，并且比基",
"Hl: Both the fine-tuned models (BERT and Span-",
"本模型更难消除偏差。我们发现SpanBERT在",
"BERT) will show less anti-queer bias than",
"我们的特定测试集上没有显示出同性恋偏见。",
"their corresponding off-the-shelf models",
"然而，重要的是要注意，这并不意味着模型中",
"H2:The model trained on QueerTwitter data will",
"不存在偏差。在SpanBERT上，我们观察到新",
"show less preference for straight over queer",
"闻语料库上的微调引入了对直句的更多偏好，",
"sentences the model trained on QueerNews.",
"而Twitter语料库上的微调则没有引人这种偏",
"好。这似乎表明主流媒体报道使用的语言比",
"我们还测试了我们的模型的共引用分辨率。",
"Twitter上的酷儿用户更加恐同，并且LLM可",
"特别是，我们评估了OntoNotes（Weischedel",
"以在调整新闻时吸收和编码这种偏见。这些结",
"etal.，2017)）共同引用任务和性别歧义代词",
"果支持了我们的假设，即经过Twitter调整的",
"(GAP）任务（Websteretal.,2018），这是NLP",
"模型显示出的刻板印象的证据较少，因为与直",
"系统中（二进制）性别偏见的既定度量。共指",
"接媒体来源相比，同性恋者以更细微和富有同",
"分析有两个目的。首先，它表明我们提出的对",
"情心的方式谈论自己"
] |
{
"arxiv_id": "2206.11484",
"language": "zh",
"page_idx": 3
}
|
||
2206.11484-zh-4
|
2206.11484-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>More Stereotypical (Straight)</th>
<th>Less Stereotypical 1(0ueer)</th>
</tr>
</tbody>
</table>
Table 1: Example sentences from WinoQueer-v0 benchmark.
<table>
<tbody>
<tr>
<th>$\mathbf{Model}$</th>
<th>Off- the- shelf</th>
<th>Finetune on OueerNews</th>
<th>Finetune on OueerTwitten</th>
</tr>
<tr>
<td>base $\mathbf{BERTt}$ </td>
<td>73.77</td>
<td>57.32</td>
<td>55.33</td>
</tr>
<tr>
<td>$\mathbf{BERT}$ large</td>
<td>76.33</td>
<td>70.3</td>
<td>$71.7\epsilon$</td>
</tr>
<tr>
<td>SpanBERT base</td>
<td>46.0</td>
<td>50.61</td>
<td>49.64</td>
</tr>
<tr>
<td>SpanBERT large</td>
<td>48.37</td>
<td>56.28</td>
<td>44.04</td>
</tr>
</tbody>
</table>
Table 2: Finetuning on both or our datasets improves anti-queer bias for BERT. SpanBERT does not show significant anti-queer bias. The metric is the percentage of examples for which the LLM's log-likelihood for the straight sentence is higher than for the gay sentence. Scores above 0.5 indicate presence of homophobic bias.
## 4.2 Coreference Resolution
在共引用分辨率设置中，我们对 OntoNotes (Weischedel et al., 2017) 和 GAP(性别歧义代词，(Webster et al.,2018))数据集进行评估。我们评估相同的 4LLM 的性能：BERT\_base, BERT\_large, SpanBERT\_base, SpanBERT\_large。我们使用原始模型创建者发布的预训练模型，这些模型通常在维基百科和其他一般领域数据上进行预训练。我们测试每个模型的三个不同版本：
$\bullet$ Finetune off-the-shelf model on OntoNotes
coreference resolution (task only). We call
this the “vanilla”condition.
$\bullet$ Finetune on LGBTQ+ news corpus then
OntoNotes (content then task)
$\bullet$ Finetune on LGBTQ+ twitter then
OntoNotes (content then task)
在 OntoNotes 测试集上，我们的模型的性能与以前发布的结果相当。正如预期的那样， 较大的模型通常具有更好的性能和更少的性别偏见，而 SpanBERT 的表现优于 BERT。我们观察到相对较少的性能下降从我们的微调。在 Twitter 上微调的 BERT_base 上观察到最严重的降级，损失 1。8%F1 得分与我们的香草BERT\_base 相比。对于较大的模型和在新闻语料库中微调的模型，性能损失甚至更小。OntoNotes 协同引用解析的强大性能表明，我们的微调不会导致灾难性的遗忘，并且我们的模型在下游 NLP 任务中仍然有用
表 4 列出了所有模型的 GAP 数据集上的结果。请注意，表中的基线模型不是基于 BERT 的模型，因此分数要低得多。我们的微调对性
能的影响类似于对 OntoNotes 的影响：观察到非常轻微的性能下降，Twitter 语料库的影响略大。这些实验的有趣结果是，我们的微调通常不会减少模型性能上的性别差异。我们期望对酷儿数据进行微调可以改善性别差异，但实际上在大多数情况下，性别差异会稍微恶化。对于 SpanBERT 模型，新闻对偏见得分的负面影响比 Twitter 略大。。造成这种影响的原因尚不清楚，需要进一步调查
## 5 $\mathbf{Conclusion}$
在这项工作中，我们介绍了 WinoQueer,这是一个新的基准数据集，用于量化大型语言模型中的反 LGBTQ 偏差。我们表明，现成的 LLM 在我们的测试集上显示出实质性的反酷儿偏见。此外，我们证明了这种偏见可以通过在LGBTQ+ 社区的推文集上微调 LLM 来缓解。最后，我们评估了我们的模型的共指分辨率， 以表明这种微调不会导致原始模型权重的灾难性遗忘。该领域的未来工作有许多有希望的方向，包括开发更全面的 WinoQueer 基准测试版本，并使用它来测试更广泛的 LLM
## $\textbf{References}$
Emily M. Bender, Timnit Gebru, Angelina McMillan-
Major, and Shmargaret Shmitchell.2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency , pages 610-623, Virtual Event Canada. $ACM.$
Reuben Binns, Michael Veale, Max Van Kleek, and
Nigel Shadbolt. 2017. Like Trainer, Like Bot? Inheritance of Bias in Algorithmic Content Moderation. In Social Informatics , Lecture Notes in Com-
|
[
[
0.26025390625,
0.0843505859375,
0.69873046875,
0.0843505859375,
0.69873046875,
0.097412109375,
0.26025390625,
0.097412109375
],
[
0.2626953125,
0.1015625,
0.465087890625,
0.1015625,
0.465087890625,
0.11517333984375,
0.2626953125,
0.11517333984375
],
[
0.49951171875,
0.1009521484375,
0.705078125,
0.1015625,
0.705078125,
0.11517333984375,
0.49951171875,
0.114013671875
],
[
0.26611328125,
0.1175537109375,
0.47021484375,
0.1175537109375,
0.47021484375,
0.1312255859375,
0.26611328125,
0.1312255859375
],
[
0.50146484375,
0.1175537109375,
0.72900390625,
0.1175537109375,
0.72900390625,
0.1312255859375,
0.50146484375,
0.1312255859375
],
[
0.2626953125,
0.1341552734375,
0.411376953125,
0.1341552734375,
0.411376953125,
0.1478271484375,
0.2626953125,
0.1478271484375
],
[
0.50146484375,
0.1341552734375,
0.654296875,
0.1341552734375,
0.654296875,
0.1478271484375,
0.50146484375,
0.1478271484375
],
[
0.291259765625,
0.160888671875,
0.705078125,
0.160888671875,
0.705078125,
0.174560546875,
0.291259765625,
0.174560546875
],
[
0.16796875,
0.190673828125,
0.225830078125,
0.190673828125,
0.225830078125,
0.204833984375,
0.16796875,
0.204833984375
],
[
0.31494140625,
0.189453125,
0.82861328125,
0.190673828125,
0.82861328125,
0.204833984375,
0.31494140625,
0.2037353515625
],
[
0.1695556640625,
0.204833984375,
0.264404296875,
0.207275390625,
0.264404296875,
0.221435546875,
0.1695556640625,
0.2197265625
],
[
0.372802734375,
0.207275390625,
0.420654296875,
0.207275390625,
0.420654296875,
0.221435546875,
0.372802734375,
0.221435546875
],
[
0.57080078125,
0.207275390625,
0.61865234375,
0.207275390625,
0.61865234375,
0.221435546875,
0.57080078125,
0.221435546875
],
[
0.779296875,
0.207275390625,
0.82861328125,
0.207275390625,
0.82861328125,
0.221435546875,
0.779296875,
0.221435546875
],
[
0.1695556640625,
0.221435546875,
0.267822265625,
0.223876953125,
0.267822265625,
0.2381591796875,
0.1695556640625,
0.236328125
],
[
0.372802734375,
0.22265625,
0.420654296875,
0.22265625,
0.420654296875,
0.2369384765625,
0.372802734375,
0.2369384765625
],
[
0.580078125,
0.221435546875,
0.61865234375,
0.221435546875,
0.61865234375,
0.2369384765625,
0.580078125,
0.2369384765625
],
[
0.779296875,
0.22265625,
0.82861328125,
0.22265625,
0.82861328125,
0.2369384765625,
0.779296875,
0.2369384765625
],
[
0.38037109375,
0.2381591796875,
0.424072265625,
0.2381591796875,
0.424072265625,
0.254150390625,
0.38037109375,
0.254150390625
],
[
0.57080078125,
0.2381591796875,
0.6171875,
0.2381591796875,
0.6171875,
0.254150390625,
0.57080078125,
0.254150390625
],
[
0.779296875,
0.2381591796875,
0.8271484375,
0.2381591796875,
0.8271484375,
0.254150390625,
0.779296875,
0.254150390625
],
[
0.779296875,
0.2529296875,
0.82861328125,
0.254638671875,
0.8271484375,
0.270751953125,
0.77734375,
0.268310546875
],
[
0.172119140625,
0.2392578125,
0.300537109375,
0.2392578125,
0.300537109375,
0.254150390625,
0.172119140625,
0.254150390625
],
[
0.37451171875,
0.254638671875,
0.424072265625,
0.254638671875,
0.424072265625,
0.26953125,
0.37451171875,
0.26953125
],
[
0.57080078125,
0.254638671875,
0.62060546875,
0.254638671875,
0.62060546875,
0.26953125,
0.57080078125,
0.26953125
],
[
0.1175537109375,
0.283203125,
0.8818359375,
0.283203125,
0.8818359375,
0.29638671875,
0.1175537109375,
0.29638671875
],
[
0.17041015625,
0.255859375,
0.3046875,
0.255859375,
0.3046875,
0.26953125,
0.17041015625,
0.26953125
],
[
0.1175537109375,
0.29638671875,
0.880859375,
0.29638671875,
0.880859375,
0.31005859375,
0.1175537109375,
0.31005859375
],
[
0.1175537109375,
0.312255859375,
0.79443359375,
0.311279296875,
0.79443359375,
0.325927734375,
0.1175537109375,
0.32666015625
],
[
0.1175537109375,
0.3515625,
0.349365234375,
0.3515625,
0.349365234375,
0.365234375,
0.1175537109375,
0.365234375
],
[
0.51123046875,
0.350830078125,
0.87744140625,
0.350830078125,
0.87744140625,
0.365234375,
0.51123046875,
0.365234375
],
[
0.118408203125,
0.37353515625,
0.48779296875,
0.37353515625,
0.48779296875,
0.38720703125,
0.118408203125,
0.38720703125
],
[
0.5146484375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.5146484375,
0.3818359375
],
[
0.118408203125,
0.3896484375,
0.48779296875,
0.3896484375,
0.48779296875,
0.402587890625,
0.118408203125,
0.402587890625
],
[
0.51318359375,
0.385498046875,
0.880859375,
0.385498046875,
0.880859375,
0.3984375,
0.51318359375,
0.3984375
],
[
0.12005615234375,
0.40625,
0.486083984375,
0.40625,
0.486083984375,
0.419189453125,
0.12005615234375,
0.419189453125
],
[
0.51318359375,
0.40087890625,
0.880859375,
0.40087890625,
0.880859375,
0.41455078125,
0.51318359375,
0.41455078125
],
[
0.1175537109375,
0.421630859375,
0.48779296875,
0.42041015625,
0.48779296875,
0.435302734375,
0.1175537109375,
0.435791015625
],
[
0.51318359375,
0.41748046875,
0.880859375,
0.41748046875,
0.880859375,
0.43115234375,
0.51318359375,
0.43115234375
],
[
0.118408203125,
0.43701171875,
0.48779296875,
0.43701171875,
0.48779296875,
0.451904296875,
0.118408203125,
0.451904296875
],
[
0.515625,
0.432861328125,
0.880859375,
0.432861328125,
0.880859375,
0.446533203125,
0.515625,
0.446533203125
],
[
0.1175537109375,
0.45361328125,
0.484375,
0.45361328125,
0.484375,
0.46728515625,
0.1175537109375,
0.46728515625
],
[
0.51318359375,
0.448974609375,
0.87744140625,
0.448974609375,
0.87744140625,
0.4619140625,
0.51318359375,
0.4619140625
],
[
0.118408203125,
0.4697265625,
0.486083984375,
0.4697265625,
0.486083984375,
0.482666015625,
0.118408203125,
0.482666015625
],
[
0.51318359375,
0.46435546875,
0.880859375,
0.46435546875,
0.880859375,
0.478515625,
0.51318359375,
0.478515625
],
[
0.1175537109375,
0.486328125,
0.483642578125,
0.486328125,
0.483642578125,
0.49951171875,
0.1175537109375,
0.49951171875
],
[
0.51123046875,
0.48095703125,
0.69775390625,
0.479736328125,
0.69775390625,
0.49462890625,
0.51123046875,
0.495361328125
],
[
0.118408203125,
0.501953125,
0.310546875,
0.501953125,
0.310546875,
0.515625,
0.118408203125,
0.515625
],
[
0.51123046875,
0.51025390625,
0.5498046875,
0.51025390625,
0.5498046875,
0.5224609375,
0.51123046875,
0.5224609375
],
[
0.54150390625,
0.50830078125,
0.6455078125,
0.51025390625,
0.6455078125,
0.52490234375,
0.54150390625,
0.5224609375
],
[
0.1376953125,
0.52978515625,
0.486083984375,
0.52978515625,
0.486083984375,
0.54443359375,
0.1376953125,
0.54443359375
],
[
0.51318359375,
0.5361328125,
0.880859375,
0.5361328125,
0.880859375,
0.5498046875,
0.51318359375,
0.5498046875
],
[
0.15283203125,
0.54638671875,
0.486083984375,
0.54638671875,
0.486083984375,
0.56005859375,
0.15283203125,
0.56005859375
],
[
0.5322265625,
0.5517578125,
0.880859375,
0.5517578125,
0.880859375,
0.5654296875,
0.5322265625,
0.5654296875
],
[
0.15283203125,
0.5634765625,
0.35693359375,
0.5634765625,
0.35693359375,
0.57666015625,
0.15283203125,
0.57666015625
],
[
0.51318359375,
0.56640625,
0.880859375,
0.56787109375,
0.880859375,
0.58203125,
0.51318359375,
0.58056640625
],
[
0.5146484375,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.5146484375,
0.59716796875
],
[
0.1376953125,
0.59033203125,
0.48779296875,
0.59033203125,
0.48779296875,
0.60498046875,
0.1376953125,
0.60498046875
],
[
0.151123046875,
0.60498046875,
0.376220703125,
0.60546875,
0.376220703125,
0.62060546875,
0.151123046875,
0.619140625
],
[
0.515625,
0.60107421875,
0.880859375,
0.60107421875,
0.880859375,
0.6142578125,
0.515625,
0.6142578125
],
[
0.51123046875,
0.6142578125,
0.880859375,
0.6162109375,
0.880859375,
0.63134765625,
0.51123046875,
0.62890625
],
[
0.225830078125,
0.63525390625,
0.358642578125,
0.63525390625,
0.358642578125,
0.6484375,
0.225830078125,
0.6484375
],
[
0.5146484375,
0.63427734375,
0.87890625,
0.63427734375,
0.87890625,
0.64453125,
0.5146484375,
0.64453125
],
[
0.378662109375,
0.6357421875,
0.43408203125,
0.6357421875,
0.43408203125,
0.64794921875,
0.378662109375,
0.64794921875
],
[
0.139404296875,
0.6357421875,
0.225830078125,
0.6357421875,
0.225830078125,
0.64794921875,
0.139404296875,
0.64794921875
],
[
0.450927734375,
0.6357421875,
0.48779296875,
0.6357421875,
0.48779296875,
0.64794921875,
0.450927734375,
0.64794921875
],
[
0.15283203125,
0.6494140625,
0.376220703125,
0.65087890625,
0.376220703125,
0.6650390625,
0.15283203125,
0.66455078125
],
[
0.5146484375,
0.6484375,
0.880859375,
0.6484375,
0.880859375,
0.662109375,
0.5146484375,
0.662109375
],
[
0.51123046875,
0.66455078125,
0.87744140625,
0.66455078125,
0.87744140625,
0.677734375,
0.51123046875,
0.677734375
],
[
0.1368408203125,
0.6787109375,
0.48779296875,
0.6787109375,
0.48779296875,
0.6923828125,
0.1368408203125,
0.6923828125
],
[
0.51318359375,
0.68115234375,
0.880859375,
0.68115234375,
0.880859375,
0.69482421875,
0.51318359375,
0.69482421875
],
[
0.118408203125,
0.69482421875,
0.489501953125,
0.69482421875,
0.489501953125,
0.7080078125,
0.118408203125,
0.7080078125
],
[
0.51318359375,
0.69775390625,
0.7958984375,
0.69775390625,
0.7958984375,
0.71142578125,
0.51318359375,
0.71142578125
],
[
0.118408203125,
0.71142578125,
0.484375,
0.71142578125,
0.484375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.118408203125,
0.7255859375,
0.484375,
0.7255859375,
0.484375,
0.7392578125,
0.118408203125,
0.7392578125
],
[
0.118408203125,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.51123046875,
0.73828125,
0.60791015625,
0.74072265625,
0.60791015625,
0.7548828125,
0.51123046875,
0.7529296875
],
[
0.1175537109375,
0.76025390625,
0.484375,
0.76025390625,
0.484375,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.76318359375,
0.87744140625,
0.76318359375,
0.87744140625,
0.77392578125,
0.51123046875,
0.77392578125
],
[
0.118408203125,
0.77490234375,
0.484375,
0.77490234375,
0.484375,
0.7880859375,
0.118408203125,
0.7880859375
],
[
0.52880859375,
0.775390625,
0.8818359375,
0.77490234375,
0.8818359375,
0.7890625,
0.52880859375,
0.79052734375
],
[
0.1175537109375,
0.7919921875,
0.484375,
0.7919921875,
0.484375,
0.8056640625,
0.1175537109375,
0.8056640625
],
[
0.52880859375,
0.7890625,
0.8818359375,
0.7890625,
0.8818359375,
0.802734375,
0.52880859375,
0.802734375
],
[
0.1175537109375,
0.80810546875,
0.483642578125,
0.80810546875,
0.483642578125,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.53076171875,
0.802734375,
0.880859375,
0.802734375,
0.880859375,
0.81640625,
0.53076171875,
0.81640625
],
[
0.5283203125,
0.81396484375,
0.87890625,
0.8154296875,
0.87890625,
0.8291015625,
0.5283203125,
0.82763671875
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.52880859375,
0.8291015625,
0.87890625,
0.82763671875,
0.87890625,
0.84130859375,
0.52880859375,
0.8427734375
],
[
0.118408203125,
0.84033203125,
0.484375,
0.84033203125,
0.484375,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.53076171875,
0.8427734375,
0.5751953125,
0.8427734375,
0.5751953125,
0.85400390625,
0.53076171875,
0.85400390625
],
[
0.118408203125,
0.85546875,
0.382080078125,
0.85546875,
0.382080078125,
0.869140625,
0.118408203125,
0.869140625
],
[
0.1368408203125,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1368408203125,
0.8857421875
],
[
0.5146484375,
0.86767578125,
0.880859375,
0.86767578125,
0.880859375,
0.87744140625,
0.5146484375,
0.87744140625
],
[
0.5322265625,
0.88134765625,
0.880859375,
0.88134765625,
0.880859375,
0.89111328125,
0.5322265625,
0.89111328125
],
[
0.118408203125,
0.88818359375,
0.486083984375,
0.88818359375,
0.486083984375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.5322265625,
0.8935546875,
0.880859375,
0.8935546875,
0.880859375,
0.90380859375,
0.5322265625,
0.90380859375
],
[
0.118408203125,
0.90478515625,
0.484375,
0.90478515625,
0.484375,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.52880859375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.52880859375,
0.91845703125
]
] |
[
"More Stereotypical (Straight)Less Stereotypical (Queer)",
"James held hands with her.",
"Jamesheld handswithhim",
"John held hands with Mary.",
"Jennifer held hands with Mary",
"He is divorcing her.",
"She is divorcing her.",
"Table l: Example sentences from WinoQueer-vO benchmark.",
"Model",
"Off-the-shelfFinetuneonQueerNewsFinetuneonQueerTwitter",
"BERT_base",
"73.77",
"57.32",
"55.33",
"BERT_large",
"76.33",
"70.3",
"71.76",
"46.0",
"50.61",
"49.64",
"44.04",
"SpanBERT_base",
"48.37",
"56.28",
"Table 2: Finetuning on both or our datasets improves anti-queer bias for BERT. SpanBERT does not show signifi-",
"SpanBERT_large",
"cant anti-queer bias. The metric is the percentage of examples for which the LLM's log-likelihood for the straight",
"sentence is higher than for the gay sentence. Scores above 0.5 indicate presence of homophobic bias.",
"4.2CoreferenceResolution",
"能的影响类似于对OntoNotes的影响：观察到",
"在共引用分辨率设置中，我们对OntoNotes",
"非常轻微的性能下降，Twitter语料库的影响略",
"(Weischedeletal.，2017）和GAP（性别歧",
"大。这些实验的有趣结果是，我们的微调通常",
"义代词，（Websteretal.，2018））数据集",
"不会减少模型性能上的性别差异。我们期望对",
"进行评估。我们评估相同的4LLM的性",
"酷儿数据进行微调可以改善性别差异，但实际",
"能:BERT_base,BERT_large,SpanBERT_base,",
"上在大多数情况下，性别差异会稍微恶化。对",
"SpanBERT_large。我们使用原始模型创建者发",
"于SpanBERT模型，新闻对偏见得分的负面影",
"布的预训练模型，这些模型通常在维基百科和",
"响比Twitter略大。。造成这种影响的原因尚不",
"其他一般领域数据上进行预训练。我们测试每",
"清楚，需要进一步调查",
"个模型的三个不同版本：",
"5",
"Conclusion",
": Finetune off-the-shelf model on OntoNotes",
"在这项工作中，我们介绍了WinoQueer，这是",
"coreference resolution (task only).We call",
"个新的基准数据集，用于量化大型语言模型",
"this the \"vanilla\" condition.",
"中的反LGBTQ偏差。我们表明，现成的LLM",
"在我们的测试集上显示出实质性的反酷儿偏",
": Finetune on LGBTQ+ news corpus then",
"OntoNotes (content then task)",
"见。此外，我们证明了这种偏见可以通过在",
"LGBTQ+社区的推文集上微调LLM来缓解。",
"onLGBTQ+",
"最后，我们评估了我们的模型的共指分辨率，",
"twitter",
": Finetune",
"then",
"OntoNotes (content then task)",
"以表明这种微调不会导致原始模型权重的灾难",
"性遗忘。该领域的未来工作有许多有希望的方",
"在OntoNotes测试集上，我们的模型的性",
"向，包括开发更全面的WinoQueer基准测试版",
"能与以前发布的结果相当。正如预期的那样，",
"本，并使用它来测试更广泛的LLM",
"较大的模型通常具有更好的性能和更少的性",
"别偏见，而SpanBERT的表现优于BERT。我",
"们观察到相对较少的性能下降从我们的微调。",
"References",
"在Twitter上微调的BERT_base上观察到最严",
"Emily M. Bender. Timnit Gebru.Angelina McMillan",
"重的降级，损失1。8%F1得分与我们的香草",
"Major, and Shmargaret Shmitcheli. 2021. On the",
"BERT_base相比。对于较大的模型和在新闻",
"Dangers of Stochastic Parrots: Can Language Mod-",
"语料库中微调的模型，性能损失甚至更小。",
"els Be Too Big? In Proceedings of the 2021 ACM",
"Conference on Fairness, Accountability, and Trans",
"OntoNotes协同引用解析的强大性能表明，我",
"parency,pages 610-623,Virtual Event Canada.",
"们的微调不会导致灾难性的遗忘，并且我们的",
"ACM.",
"模型在下游NLP任务中仍然有用",
"表4列出了所有模型的GAP数据集上的结",
"ReubenBinns,MichaelVeale,MaxVanKleek,and",
"NigelShadbolt.2o17.LikeTrainer,LikeBot?In-",
"果。请注意，表中的基线模型不是基于BERT",
"heritance ofBias inAlgorithmic ContentModera-",
"的模型，因此分数要低得多。我们的微调对性",
"tion. In Social Informatics , Lecture Notes in Com-"
] |
{
"arxiv_id": "2206.11484",
"language": "zh",
"page_idx": 4
}
|
||
2302.06488-zh-1
|
2302.06488-zh-1.jpg
|
# 为什么话语解析不能泛化？ 对数据多样性影响的彻底调查
Yang Janet Liu and Amir Zeldes
# Department of Linguistics Georgetown University
$\{$ yl879, amir.zeldes $\}$ @georgetown.edu
## $\mathbf{Abstract}$
Recent advances in discourse parsing performance create the impression that, as in other NLP tasks, performance for high-resource languages such as English is finally becoming reliable. In this paper we demonstrate that this is not the case, and thoroughly investigate the impact of data diversity on RST parsing stability. We show that state-of-the-art architectures trained on the standard English newswire benchmark do not generalize well, even within the news domain. Using the two largest RST corpora of English with text from multiple genres, we quantify the impact of genre diversity in training data for achieving generalization to text types unseen during training. Our results show that a heterogeneous training regime is critical for stable and generalizable models, across parser architectures. We also provide error analyses of model outputs and out-of-domain performance. To our knowledge, this study is the first to fully evaluate cross-corpus RST parsing generalizability on complete trees, examine between-genre degradation within an RST corpus, and investigate the impact of genre diversity in training data composition.
## 1 Introduction
语篇解析是对文本不同部分之间的连贯关系进行识别和分类的任务，例如识别一个句子指定了另一个句子中事件的原因，或者从句指示了主句的目的。在层次框架中，例如修辞结构理论(RST, Mann and Thompson 1988 ),解析器构造了一个标记的话语单元组成树，如图 1 所示。这样的树有许多应用：检索特定的关系(例如。g。所有 CONCESSIONS 都是由某个政客在任何演讲中制作的，例如。g。$\overline{\mathrm{E}}$图 1 中的单元26-27), 或者在任意范围的文本中分层查找最中心的单元(图1中的单元 24)。它们还可以有助于下游任务，如文本生成 (Maskharashvili et al.,2021) 和摘要 (例如。g。Louis et al. 2010; Li et al. 2016; Xu et al. 2020; Hewett and Stede
<FigureHere>
Figure 1: RST Fragment from GUM (Zeldes, 2017) . 2022)和对文本(Wan et al., 2019) 的定性分析与比较。1
NLP 的最新进展导致了越来越精确的系统，用于诸如词性标记(Heinzerling and Strube, 2019) 和依赖性解析 (Mrini et al., 2020) 之类的任务，这些任务现在通常可以放心地应用于新颖的数据，至少对于高资源语言，例如英语。同时，根据 1989年《华尔街日报》(WSJ)的数据，主要标准英语 RST 基准 RST 话语树库(RST-DT, Carlson et al. 2003 )的得分不断提高， 给人的印象是语篇解析也变得可靠，并且通过代理，适用于任意文本。这导致使用自动 RST 解析来自动生成大规模数据集，例如 MEGADT (Huber and Carenini, 2020) 或 AMALGUM (Gessler et al., 2020)。但是，有一些迹象表明这种情况过于乐观，一些研究表明，一旦目标域发生变化 (Ferracane et al. 2019; Wang et al. 2019; Linscheid et al. 2021; Atwell et al. 20212022; Nishida and Matsumoto 2022; Yu et al. 2022 ), 基于 RST-DT 的解析中就会出现问题并且对解析器输出的定性检查会对其可靠性产生怀疑
评估英语 RST 解析的稳定性的一个问题是域外 (OOD) 数据的稀缺性，以及不同语料库的注释方案之间的潜在差异。最近，英语GUM 语料库 (Zeldes, 2017) 中越来越多的非新闻专线数据以及非 WSJ 新闻数据的可用性使一些详细的比较成为可能，我们将在下面进行
${\overline{{^{1}}}\text{This applies to other discourse frameworks too, such as}}$
PDTB (Prasad et al., 2019) or SDRT (Asher and Lascarides, 2003) , but we limit the scope of this paper to RST parsing
|
[
[
0.358642578125,
0.0926513671875,
0.64404296875,
0.09320068359375,
0.64404296875,
0.11102294921875,
0.358642578125,
0.10986328125
],
[
0.344970703125,
0.114013671875,
0.654296875,
0.114013671875,
0.654296875,
0.12890625,
0.344970703125,
0.12890625
],
[
0.35693359375,
0.15087890625,
0.646484375,
0.1519775390625,
0.646484375,
0.166259765625,
0.35693359375,
0.1650390625
],
[
0.3955078125,
0.169189453125,
0.609375,
0.169189453125,
0.609375,
0.182861328125,
0.3955078125,
0.182861328125
],
[
0.405517578125,
0.1859130859375,
0.59619140625,
0.1859130859375,
0.59619140625,
0.1995849609375,
0.405517578125,
0.1995849609375
],
[
0.310546875,
0.203125,
0.6943359375,
0.203125,
0.6943359375,
0.2161865234375,
0.310546875,
0.2161865234375
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.5615234375,
0.255859375,
0.5859375,
0.255859375,
0.5859375,
0.26416015625,
0.5615234375,
0.26416015625
],
[
0.1435546875,
0.281982421875,
0.459228515625,
0.283203125,
0.459228515625,
0.297607421875,
0.1435546875,
0.29638671875
],
[
0.5576171875,
0.29150390625,
0.58447265625,
0.29150390625,
0.58447265625,
0.2998046875,
0.5576171875,
0.2998046875
],
[
0.767578125,
0.29150390625,
0.79248046875,
0.29150390625,
0.79248046875,
0.2998046875,
0.767578125,
0.2998046875
],
[
0.145263671875,
0.312255859375,
0.459228515625,
0.312255859375,
0.459228515625,
0.325927734375,
0.145263671875,
0.325927734375
],
[
0.145263671875,
0.29638671875,
0.45751953125,
0.29638671875,
0.45751953125,
0.31005859375,
0.145263671875,
0.31005859375
],
[
0.1435546875,
0.325927734375,
0.4609375,
0.32470703125,
0.4609375,
0.339111328125,
0.1435546875,
0.34033203125
],
[
0.73486328125,
0.32470703125,
0.748046875,
0.328857421875,
0.744140625,
0.33837890625,
0.72900390625,
0.334228515625
],
[
0.81689453125,
0.327880859375,
0.82373046875,
0.327880859375,
0.82373046875,
0.334228515625,
0.81689453125,
0.334228515625
],
[
0.6513671875,
0.328857421875,
0.66015625,
0.328857421875,
0.66015625,
0.334228515625,
0.6513671875,
0.334228515625
],
[
0.53564453125,
0.334228515625,
0.609375,
0.336181640625,
0.609375,
0.3466796875,
0.53564453125,
0.344482421875
],
[
0.5693359375,
0.328857421875,
0.580078125,
0.328857421875,
0.580078125,
0.336181640625,
0.5693359375,
0.336181640625
],
[
0.62109375,
0.33740234375,
0.68603515625,
0.33740234375,
0.68603515625,
0.344482421875,
0.62109375,
0.344482421875
],
[
0.705078125,
0.33740234375,
0.771484375,
0.33740234375,
0.771484375,
0.344482421875,
0.705078125,
0.344482421875
],
[
0.77734375,
0.33740234375,
0.85986328125,
0.33740234375,
0.85986328125,
0.344482421875,
0.77734375,
0.344482421875
],
[
0.5341796875,
0.344482421875,
0.77685546875,
0.344482421875,
0.77685546875,
0.355224609375,
0.5341796875,
0.355224609375
],
[
0.78662109375,
0.344482421875,
0.84814453125,
0.344482421875,
0.84814453125,
0.355224609375,
0.78662109375,
0.355224609375
],
[
0.54736328125,
0.352783203125,
0.60107421875,
0.352783203125,
0.60107421875,
0.363525390625,
0.54736328125,
0.363525390625
],
[
0.145263671875,
0.34033203125,
0.459228515625,
0.34033203125,
0.459228515625,
0.35400390625,
0.145263671875,
0.35400390625
],
[
0.63232421875,
0.352783203125,
0.67822265625,
0.352783203125,
0.67822265625,
0.363525390625,
0.63232421875,
0.363525390625
],
[
0.705078125,
0.352783203125,
0.771484375,
0.352783203125,
0.771484375,
0.363525390625,
0.705078125,
0.363525390625
],
[
0.5498046875,
0.361083984375,
0.59619140625,
0.361083984375,
0.59619140625,
0.371826171875,
0.5498046875,
0.371826171875
],
[
0.145263671875,
0.35400390625,
0.45751953125,
0.35400390625,
0.45751953125,
0.36767578125,
0.145263671875,
0.36767578125
],
[
0.14697265625,
0.3681640625,
0.459228515625,
0.3681640625,
0.459228515625,
0.3818359375,
0.14697265625,
0.3818359375
],
[
0.145263671875,
0.383056640625,
0.459228515625,
0.383056640625,
0.459228515625,
0.396728515625,
0.145263671875,
0.396728515625
],
[
0.51708984375,
0.383056640625,
0.875,
0.383056640625,
0.875,
0.396728515625,
0.51708984375,
0.396728515625
],
[
0.145263671875,
0.39794921875,
0.459228515625,
0.39794921875,
0.459228515625,
0.410888671875,
0.145263671875,
0.410888671875
],
[
0.14697265625,
0.413330078125,
0.45751953125,
0.413330078125,
0.45751953125,
0.42333984375,
0.14697265625,
0.42333984375
],
[
0.51318359375,
0.412109375,
0.880859375,
0.412109375,
0.880859375,
0.427001953125,
0.51318359375,
0.427001953125
],
[
0.145263671875,
0.42578125,
0.45751953125,
0.42578125,
0.45751953125,
0.439453125,
0.145263671875,
0.439453125
],
[
0.5146484375,
0.427490234375,
0.59619140625,
0.427490234375,
0.59619140625,
0.4423828125,
0.5146484375,
0.4423828125
],
[
0.1435546875,
0.439453125,
0.45751953125,
0.439453125,
0.45751953125,
0.453125,
0.1435546875,
0.453125
],
[
0.52880859375,
0.446533203125,
0.880859375,
0.446533203125,
0.880859375,
0.460205078125,
0.52880859375,
0.460205078125
],
[
0.145263671875,
0.45361328125,
0.459228515625,
0.45361328125,
0.459228515625,
0.46728515625,
0.145263671875,
0.46728515625
],
[
0.145263671875,
0.468505859375,
0.459228515625,
0.468505859375,
0.459228515625,
0.482177734375,
0.145263671875,
0.482177734375
],
[
0.51318359375,
0.463134765625,
0.880859375,
0.463134765625,
0.880859375,
0.476806640625,
0.51318359375,
0.476806640625
],
[
0.1435546875,
0.48095703125,
0.456787109375,
0.482177734375,
0.456787109375,
0.496337890625,
0.1435546875,
0.495361328125
],
[
0.51318359375,
0.478515625,
0.880859375,
0.478515625,
0.880859375,
0.4921875,
0.51318359375,
0.4921875
],
[
0.145263671875,
0.496337890625,
0.45751953125,
0.496337890625,
0.45751953125,
0.51025390625,
0.145263671875,
0.51025390625
],
[
0.51318359375,
0.49462890625,
0.87890625,
0.49462890625,
0.87890625,
0.5087890625,
0.51318359375,
0.5087890625
],
[
0.1435546875,
0.51123046875,
0.456787109375,
0.51025390625,
0.456787109375,
0.52392578125,
0.1435546875,
0.52490234375
],
[
0.5146484375,
0.51123046875,
0.8818359375,
0.51123046875,
0.8818359375,
0.52490234375,
0.5146484375,
0.52490234375
],
[
0.145263671875,
0.52490234375,
0.45751953125,
0.52490234375,
0.45751953125,
0.5380859375,
0.145263671875,
0.5380859375
],
[
0.51318359375,
0.52685546875,
0.880859375,
0.525390625,
0.880859375,
0.54052734375,
0.51318359375,
0.54150390625
],
[
0.145263671875,
0.5390625,
0.459228515625,
0.5390625,
0.459228515625,
0.552734375,
0.145263671875,
0.552734375
],
[
0.5146484375,
0.54345703125,
0.880859375,
0.54345703125,
0.880859375,
0.55712890625,
0.5146484375,
0.55712890625
],
[
0.145263671875,
0.55419921875,
0.459228515625,
0.55419921875,
0.459228515625,
0.56787109375,
0.145263671875,
0.56787109375
],
[
0.509765625,
0.55810546875,
0.8818359375,
0.55859375,
0.8818359375,
0.57373046875,
0.509765625,
0.572265625
],
[
0.145263671875,
0.56787109375,
0.45751953125,
0.56787109375,
0.45751953125,
0.58056640625,
0.145263671875,
0.58056640625
],
[
0.51123046875,
0.57470703125,
0.87890625,
0.57470703125,
0.87890625,
0.58837890625,
0.51123046875,
0.58837890625
],
[
0.145263671875,
0.58203125,
0.45751953125,
0.58203125,
0.45751953125,
0.595703125,
0.145263671875,
0.595703125
],
[
0.14697265625,
0.5966796875,
0.456787109375,
0.5966796875,
0.456787109375,
0.60693359375,
0.14697265625,
0.60693359375
],
[
0.5146484375,
0.59130859375,
0.8818359375,
0.59130859375,
0.8818359375,
0.60498046875,
0.5146484375,
0.60498046875
],
[
0.145263671875,
0.609375,
0.456787109375,
0.609375,
0.456787109375,
0.62255859375,
0.145263671875,
0.62255859375
],
[
0.5146484375,
0.60791015625,
0.8818359375,
0.60791015625,
0.8818359375,
0.62158203125,
0.5146484375,
0.62158203125
],
[
0.145263671875,
0.62451171875,
0.459228515625,
0.62451171875,
0.459228515625,
0.63818359375,
0.145263671875,
0.63818359375
],
[
0.51123046875,
0.62255859375,
0.880859375,
0.62255859375,
0.880859375,
0.6357421875,
0.51123046875,
0.6357421875
],
[
0.1427001953125,
0.6396484375,
0.2359619140625,
0.63720703125,
0.2359619140625,
0.65185546875,
0.1427001953125,
0.65380859375
],
[
0.51123046875,
0.64013671875,
0.880859375,
0.64013671875,
0.880859375,
0.65380859375,
0.51123046875,
0.65380859375
],
[
0.51123046875,
0.65478515625,
0.8818359375,
0.65625,
0.8818359375,
0.67041015625,
0.51123046875,
0.66943359375
],
[
0.1351318359375,
0.6728515625,
0.1427001953125,
0.6728515625,
0.1427001953125,
0.6767578125,
0.1351318359375,
0.6767578125
],
[
0.151123046875,
0.66943359375,
0.2626953125,
0.66943359375,
0.2626953125,
0.68115234375,
0.151123046875,
0.68115234375
],
[
0.51318359375,
0.6728515625,
0.880859375,
0.6728515625,
0.880859375,
0.68603515625,
0.51318359375,
0.68603515625
],
[
0.51318359375,
0.6884765625,
0.884765625,
0.6884765625,
0.884765625,
0.70166015625,
0.51318359375,
0.70166015625
],
[
0.11248779296875,
0.6748046875,
0.1259765625,
0.66650390625,
0.1376953125,
0.6748046875,
0.124267578125,
0.68310546875
],
[
0.118408203125,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51318359375,
0.70361328125,
0.880859375,
0.70361328125,
0.880859375,
0.71728515625,
0.51318359375,
0.71728515625
],
[
0.118408203125,
0.71142578125,
0.484375,
0.71142578125,
0.484375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.51123046875,
0.71875,
0.880859375,
0.7197265625,
0.880859375,
0.73388671875,
0.51123046875,
0.73291015625
],
[
0.12176513671875,
0.72705078125,
0.483642578125,
0.72705078125,
0.483642578125,
0.7412109375,
0.12176513671875,
0.7412109375
],
[
0.51123046875,
0.7353515625,
0.8818359375,
0.73388671875,
0.8818359375,
0.7490234375,
0.51123046875,
0.75
],
[
0.118408203125,
0.7421875,
0.484375,
0.7421875,
0.484375,
0.755859375,
0.118408203125,
0.755859375
],
[
0.51318359375,
0.751953125,
0.57080078125,
0.751953125,
0.57080078125,
0.7666015625,
0.51318359375,
0.7666015625
],
[
0.118408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.53076171875,
0.76953125,
0.880859375,
0.77099609375,
0.880859375,
0.78515625,
0.53076171875,
0.78369140625
],
[
0.118408203125,
0.775390625,
0.486083984375,
0.775390625,
0.486083984375,
0.7890625,
0.118408203125,
0.7890625
],
[
0.51123046875,
0.78515625,
0.880859375,
0.7861328125,
0.880859375,
0.80029296875,
0.51123046875,
0.7998046875
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.12005615234375,
0.80810546875,
0.484375,
0.80810546875,
0.484375,
0.822265625,
0.12005615234375,
0.822265625
],
[
0.51318359375,
0.802734375,
0.8818359375,
0.802734375,
0.8818359375,
0.81640625,
0.51318359375,
0.81640625
],
[
0.51318359375,
0.818359375,
0.880859375,
0.818359375,
0.880859375,
0.83203125,
0.51318359375,
0.83203125
],
[
0.118408203125,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.486083984375,
0.84033203125,
0.486083984375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.51318359375,
0.8349609375,
0.880859375,
0.8349609375,
0.880859375,
0.8486328125,
0.51318359375,
0.8486328125
],
[
0.5283203125,
0.849609375,
0.880859375,
0.85107421875,
0.880859375,
0.865234375,
0.5283203125,
0.8642578125
],
[
0.12005615234375,
0.85693359375,
0.486083984375,
0.85693359375,
0.486083984375,
0.87060546875,
0.12005615234375,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.53564453125,
0.87744140625,
0.88330078125,
0.8798828125,
0.88330078125,
0.89794921875,
0.53564453125,
0.8955078125
],
[
0.1158447265625,
0.88720703125,
0.48779296875,
0.88818359375,
0.48779296875,
0.90283203125,
0.1158447265625,
0.90185546875
],
[
0.51123046875,
0.8935546875,
0.880859375,
0.8935546875,
0.880859375,
0.90673828125,
0.51123046875,
0.90673828125
],
[
0.1175537109375,
0.90625,
0.484375,
0.90625,
0.484375,
0.91650390625,
0.1175537109375,
0.91650390625
],
[
0.509765625,
0.90380859375,
0.87158203125,
0.90478515625,
0.87158203125,
0.91943359375,
0.509765625,
0.91845703125
]
] |
[
"为什么话语解析不能泛化？",
"对数据多样性影响的彻底调查",
"Yang Janet Liu and Amir Zeldes",
"Department of Linguistics",
"Georgetown University",
"( yl879, amir.zeldes } @georgetown.edu",
"Abstract",
"24-27",
"Recent advances in discourse parsing perfor-",
"24-25",
"26-27",
"NLP tasks, performance for high-resource lan-",
"mance create the impression that, as in other",
"guages such as English is finally becoming re-",
"26",
"27",
"25",
"Thecommunity",
"24",
"astheUnited",
"Thisperiodof",
"andwillbefull of",
"must support the Nations does now change may last",
"challenges.",
"Arab world",
"liable. In this paper we demonstrate that this",
"in Libya .",
"forsometime",
"transition",
"is not the case, and thoroughly investigate the",
"impact of data diversity on RST parsing sta-",
"bility. We show that state-of-the-art architec-",
"Figure 1: RST Fragment from GUM (Zeldes, 2017) .",
"tures trained on the standard English newswire",
"benchmark donotgeneralizewell,even within",
"2022）和对文本（Wanetal.,2019）的定性分析",
"the news domain. Using the two largest RST",
"与比较。1",
"corpora of English with text from multiple",
"NLP的最新进展导致了越来越精确的系",
"genres, we quantify the impact of genre di-",
"versity in training data for achieving general",
"统，用于诸如词性标记（Heinzerlingand Strube,",
"ization to text types unseen during training.",
"2019)和依赖性解析（Mrinietal.，2020）之类的",
"Our results show that a heterogeneous training",
"任务，这些任务现在通常可以放心地应用于新",
"regime is critical for stable and generalizable",
"颖的数据，至少对于高资源语言，例如英语。",
"models, across parser architectures. We also",
"同时，根据1989年《华尔街日报》（WSJ）的",
"provide error analyses of model outputs and",
"数据，主要标准英语RST基准RST话语树库",
"out-of-domain performance. To our knowl-",
"（RST-DT，Carlsonetal.2003）的得分不断提高，",
"edge, this study is the first to fully evaluate",
"给人的印象是语篇解析也变得可靠，并且通过",
"cross-corpus RST parsing generalizability on",
"completetrees.examinebetween-genredegra",
"代理，适用于任意文本。这导致使用自动RST",
"dation within an RST corpus, and investigate",
"解析来自动生成大规模数据集，例如MEGA-",
"the impact of genre diversity in training data",
"DT (Huber and Carenini.2020)或 AMALGUM",
"composition.",
"(Gessleretal.,2020）。但是，有一些迹象表明，",
"这种情况过于乐观，一些研究表明，一且目标",
"",
"Introduction",
"域发生变化（Ferracaneetal.2019;Wang etal.",
"2019; Linscheid et al. 2021; Atwell et al. 2021 ,",
"1",
"语篇解析是对文本不同部分之间的连贯关系进",
"2022 ; Nishida and Matsumoto 2022 ; Yu et al.",
"行识别和分类的任务，例如识别一个句子指定",
"2022），基于RST-DT的解析中就会出现问题，",
"了另一个句子中事件的原因，或者从句指示了",
"并且对解析器输出的定性检查会对其可靠性产",
"主句的目的。在层次框架中，例如修辞结构理",
"生怀疑",
"论（RST，MannandThompson1988），解析器",
"评估英语RST解析的稳定性的一个问题是",
"构造了一个标记的话语单元组成树，如图1所",
"域外（OOD）数据的稀缺性，以及不同语料",
"示。这样的树有许多应用：检索特定的关系（例",
"如。g。所有CONCESSIONS都是由某个政客在",
"库的注释方案之间的潜在差异。最近，英语",
"GUM语料库（Zeldes，2017）中越来越多的非新",
"任何演讲中制作的，例如。g。E图1中的单元",
"26-27），或者在任意范围的文本中分层查找最",
"闻专线数据以及非WSJ新闻数据的可用性使",
"-些详细的比较成为可能，我们将在下面进行",
"中心的单元（图1中的单元24）。它们还可以",
"有助于下游任务，如文本生成（Maskharashvili",
"IThis applies to other discourseframeworks too, such as",
"etal.,2021)和摘要（例如。g。Louisetal.2010;",
"PDTB (Prasad et al., 2019) or SDRT (Asher and Lascarides,",
"Li et al. 2016: Xu et al. 2020: Hewett and Stede",
"2003), but we limit the scope of this paper to RST parsing."
] |
{
"arxiv_id": "2302.06488",
"language": "zh",
"page_idx": 1
}
|
||
2302.06488-zh-2
|
2302.06488-zh-2.jpg
|
探讨。除了评估 OOD 退化之外，这些用于英语 RST 的大型数据集的存在使我们能够测试不同的联合训练策略，以提高英语 RST 解析的准确性和通用性。$^{2}$
我们在这里的主要目标是：1)展示基于 RSTDT 的英语 RST 解析的可泛化性局限性，并为用户量化；2) 探索可泛化性问题的原因，重点是训练集的体裁组成，为健壮的语篇解析所需的数据类型指明了方向；3)根据我们的实验结果，促进 RST 解析的多体裁基准。总的来说，我们发现，无论模型架构如何，不同的训练数据都能更好地概括看不见的体裁
## 2 $\textbf{Related Work}$
2.1 English RST Corpora
RST 话语树库 (RST-DT, Carlson et al. 2003 )是标准的英语 RST 基准，数据来自 PennTreebank (PTB, Marcus et al.1993)的WSJ 部分。另一个带有人类注释的英语 RST 语料库是 GeorgetownUniversityMmultilayer (GUM) 语料库，该语料库可在线免费获得，涵盖 12 种书面和口头类型 (Zeldes,2017)。GUM 不断增长，每个版本都添加了新数据。对于本文，我们使用了语料库的第 8 版，如表 1 中有关 RST-DT 的信息旁边的描述 (GUM 的当前版本 V9 在提交时尚未提供，现在包含 213 个文档，203K 令牌和 26KEDU, 下一个版本 10 设置为添加 4 个其他新类型)。
<table>
<tbody>
<tr>
<th> </th>
<th>RST- DT</th>
<th>GUMV8</th>
</tr>
<tr>
<td>documents</td>
<td>385</td>
<td>193</td>
</tr>
<tr>
<td>$\operatorname{train/dev/test}$</td>
<td>$347/-^3/38$</td>
<td>145/24/24</td>
</tr>
<tr>
<td>tokens</td>
<td>203,352</td>
<td>180,851</td>
</tr>
<tr>
<td>EDUs</td>
<td>21,789</td>
<td>23,107</td>
</tr>
<tr>
<td>relation instances</td>
<td>20,163</td>
<td>21,903</td>
</tr>
<tr>
<td>relation labels genres</td>
<td>-0,0 78/17 classes 4 1</td>
<td>$\mathbf{E}\cdot\mathbf{A}/\mathbf{O}\cdot\mathbf{A}$ 32/15 classes 12</td>
</tr>
<tr>
<td>relation labels 1</td>
<td>78/17 classes</td>
<td>32/15 classes</td>
</tr>
<tr>
<td>genres</td>
<td>1 4</td>
<td>12</td>
</tr>
</tbody>
</table>
Table 1: Overview of the English RST Corpora.
如表 1 所示，尽管 GUMV8 在令牌和文档数量上小于 RST-DT, 但它包含更多基本话语单元(EDU) 和关系的实例，因为 RST-DT 的新闻专线单元平均更长。对于两个语料库，EDU 分割准则是相同的，并且标签集非常相似，对于不连续的 EDU, 两个语料库都使用 SAMEUNIT 伪关系。GUM 的 12 种流派使进行实验
${}^{2}{\mathrm{Though~we~focus~on~English~in~this~paper,~understand-}}$ ing the size and nature of data needed in English will hopefully shed light on what other languages may require.
$^{3}$There is no established dev partition for RST-DT.
$^{4}$Previous work has identified 4 sub-genres in the WSJ data (Webber, 2009), but these are very unevenly distributed, and 29 RST-DT documents were not included in that analysis.
以研究解析的通用性成为可能 (有关其确切大小细分，请参见附录 B 中的表 12 )。几乎所有关于 RST 成分解析的工作都只使用折叠的粗略关系类(i。e。17个RST-DT 标签);我们将在下面的评估中遵循这一做法，但在我们的分析中也会提到一些与细粒度标签有关的问题。此外，我们对两个语料库中的 s-n 核模式进行了分析，证实了它们非常相似：NS 以 77 种为主。7RST-DT 中的%,类似于 74。8口香糖新闻中的%; 有趣的是，口香糖中的比例总体上略低(70。1%)
还有其他语料库在 RST 依赖关系中进行了注释，例如 SciDTB (Yang and Li, 2018) 和其他语言的完整 RST 组成树库 (有关概述，请参见Zeldes et al. 2021 ); 由于我们专注于分层的英语 RST 成分解析，因此我们将使用表 1 中的语料库进行以下实验：它们在内容，词汇，领域以及我们所有
## 2.2 $\textbf{RST Discourse Parsers}$
已经提出了几种用于 RST 解析的方法，主要通过 BOTTOM-UP 与。TOP-DOWN 方法，以及使用的算法(通常是神经移位-减少架构)。表 2在三个指标上比较了最近的高分解析器：Span (子树是否跨越正确的 EDU), N 核性(边缘是否指向正确的方向)和R关系(标签是否正确)。在 Morey et al.(2017)之后，我们在二叉树上使用更严格的原始 Parseval 过程
如表 2 所示，BOTTOM-UP 和 TOP-DOWN 方法目前非常接近，并且所有最先进的(SOTA) 系统都依赖于变压器字嵌入作为主要输入。为了排除解析器体系结构的影响，我们使用 Guz and Carenini (2020) 的最佳 SpanBERT-NoCoref 设置对 BOTTOM-UP 进行了测试，并使用 Liu et al.(2021)的 TOP-DOWN 进行了测试，因为它们的代码是公共的，并且都具有接近 SOTA 的 S 和 N 指标，这对于整个语料库之间的标签清单之间的差异更为稳健(相关的超参数和XMATHX\_。
### 2.3 Cross-Genre Variation in RST
先前的工作表明，RST 树在文本类型，$^{5}$的关系分布 (Taboada and Lavid, 2003; Zeldes, 2018) 中差异很大，它们的信号方式(例如。g。Eggins and Martin 1997; Liu 2019; Demberg et al. 2019 ), 以及跨语言的变化( Cao et al. 2018 ; Stede and Neumann 2014 ; Iruskieta et al. 2013 ; Redeker et al. 2012 ; da Cunha et al. 2011 ; Pisarevskaya et al. 2017 )
$\overline{{^5\text{We use“tex}}}$t type”,“genre”,and“domain”interchangeably here to mean different kinds of data, though they are not exactly the same (e.g. see Biber 1988 and Lee 2001 ).
|
[
[
0.1175537109375,
0.08612060546875,
0.48779296875,
0.08489990234375,
0.48779296875,
0.09857177734375,
0.1175537109375,
0.09979248046875
],
[
0.51123046875,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.118408203125,
0.1015625,
0.48779296875,
0.1015625,
0.48779296875,
0.11517333984375,
0.118408203125,
0.11517333984375
],
[
0.51318359375,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.5146484375,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.5146484375,
0.1318359375
],
[
0.12005615234375,
0.135986328125,
0.27783203125,
0.135986328125,
0.27783203125,
0.146728515625,
0.12005615234375,
0.146728515625
],
[
0.5146484375,
0.135986328125,
0.87890625,
0.135986328125,
0.87890625,
0.146728515625,
0.5146484375,
0.146728515625
],
[
0.1368408203125,
0.1519775390625,
0.489501953125,
0.1519775390625,
0.489501953125,
0.1650390625,
0.1368408203125,
0.1650390625
],
[
0.51123046875,
0.149658203125,
0.87890625,
0.15087890625,
0.87890625,
0.1650390625,
0.51123046875,
0.1644287109375
],
[
0.1175537109375,
0.16748046875,
0.486083984375,
0.16748046875,
0.486083984375,
0.18115234375,
0.1175537109375,
0.18115234375
],
[
0.51318359375,
0.166259765625,
0.880859375,
0.166259765625,
0.880859375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.1175537109375,
0.1817626953125,
0.484375,
0.1817626953125,
0.484375,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.51318359375,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.12005615234375,
0.1995849609375,
0.486083984375,
0.1995849609375,
0.486083984375,
0.213134765625,
0.12005615234375,
0.213134765625
],
[
0.515625,
0.198974609375,
0.87890625,
0.198974609375,
0.87890625,
0.2120361328125,
0.515625,
0.2120361328125
],
[
0.118408203125,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.51318359375,
0.213134765625,
0.880859375,
0.21435546875,
0.880859375,
0.2286376953125,
0.51318359375,
0.22802734375
],
[
0.118408203125,
0.23095703125,
0.486083984375,
0.23095703125,
0.486083984375,
0.24462890625,
0.118408203125,
0.24462890625
],
[
0.51123046875,
0.23095703125,
0.87890625,
0.23095703125,
0.87890625,
0.24462890625,
0.51123046875,
0.24462890625
],
[
0.118408203125,
0.2476806640625,
0.48779296875,
0.2476806640625,
0.48779296875,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.51123046875,
0.2476806640625,
0.62890625,
0.2476806640625,
0.62890625,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.1175537109375,
0.26416015625,
0.4130859375,
0.26416015625,
0.4130859375,
0.27783203125,
0.1175537109375,
0.27783203125
],
[
0.52880859375,
0.262451171875,
0.87890625,
0.26318359375,
0.87890625,
0.27783203125,
0.52880859375,
0.276611328125
],
[
0.51318359375,
0.27783203125,
0.87890625,
0.27783203125,
0.87890625,
0.29150390625,
0.51318359375,
0.29150390625
],
[
0.1158447265625,
0.29345703125,
0.151123046875,
0.29345703125,
0.151123046875,
0.305908203125,
0.1158447265625,
0.305908203125
],
[
0.1502685546875,
0.292236328125,
0.274658203125,
0.292236328125,
0.274658203125,
0.30712890625,
0.1502685546875,
0.30712890625
],
[
0.5146484375,
0.29443359375,
0.880859375,
0.29443359375,
0.880859375,
0.308837890625,
0.5146484375,
0.308837890625
],
[
0.51318359375,
0.311279296875,
0.880859375,
0.311279296875,
0.880859375,
0.32470703125,
0.51318359375,
0.32470703125
],
[
0.11419677734375,
0.318359375,
0.337646484375,
0.319580078125,
0.337646484375,
0.334228515625,
0.11419677734375,
0.333251953125
],
[
0.51318359375,
0.32666015625,
0.880859375,
0.32666015625,
0.880859375,
0.34033203125,
0.51318359375,
0.34033203125
],
[
0.1158447265625,
0.34033203125,
0.486083984375,
0.34033203125,
0.486083984375,
0.35400390625,
0.1158447265625,
0.35400390625
],
[
0.51123046875,
0.342529296875,
0.880859375,
0.342529296875,
0.880859375,
0.355712890625,
0.51123046875,
0.355712890625
],
[
0.118408203125,
0.358154296875,
0.48779296875,
0.358154296875,
0.48779296875,
0.371826171875,
0.118408203125,
0.371826171875
],
[
0.51318359375,
0.35986328125,
0.64208984375,
0.35986328125,
0.64208984375,
0.374755859375,
0.51318359375,
0.374755859375
],
[
0.118408203125,
0.37353515625,
0.466796875,
0.37353515625,
0.466796875,
0.38720703125,
0.118408203125,
0.38720703125
],
[
0.462646484375,
0.377685546875,
0.466796875,
0.377685546875,
0.466796875,
0.383056640625,
0.462646484375,
0.383056640625
],
[
0.1175537109375,
0.388427734375,
0.48779296875,
0.3896484375,
0.48779296875,
0.40380859375,
0.1175537109375,
0.402587890625
],
[
0.51123046875,
0.385986328125,
0.7421875,
0.385986328125,
0.7421875,
0.399658203125,
0.51123046875,
0.399658203125
],
[
0.1175537109375,
0.40625,
0.48779296875,
0.40625,
0.48779296875,
0.419189453125,
0.1175537109375,
0.419189453125
],
[
0.51318359375,
0.407958984375,
0.880859375,
0.407958984375,
0.880859375,
0.421630859375,
0.51318359375,
0.421630859375
],
[
0.1175537109375,
0.421630859375,
0.486083984375,
0.42041015625,
0.486083984375,
0.435302734375,
0.1175537109375,
0.435791015625
],
[
0.51318359375,
0.42333984375,
0.880859375,
0.42333984375,
0.880859375,
0.43701171875,
0.51318359375,
0.43701171875
],
[
0.118408203125,
0.438232421875,
0.484375,
0.438232421875,
0.484375,
0.451904296875,
0.118408203125,
0.451904296875
],
[
0.5146484375,
0.440673828125,
0.8818359375,
0.440673828125,
0.8818359375,
0.45361328125,
0.5146484375,
0.45361328125
],
[
0.118408203125,
0.454833984375,
0.484375,
0.454833984375,
0.484375,
0.468505859375,
0.118408203125,
0.468505859375
],
[
0.51123046875,
0.454833984375,
0.880859375,
0.4560546875,
0.880859375,
0.47021484375,
0.51123046875,
0.4697265625
],
[
0.12005615234375,
0.47021484375,
0.486083984375,
0.47021484375,
0.486083984375,
0.48388671875,
0.12005615234375,
0.48388671875
],
[
0.51123046875,
0.47265625,
0.880859375,
0.47265625,
0.880859375,
0.486328125,
0.51123046875,
0.486328125
],
[
0.118408203125,
0.486328125,
0.484375,
0.486328125,
0.484375,
0.49951171875,
0.118408203125,
0.49951171875
],
[
0.51318359375,
0.488037109375,
0.880859375,
0.488037109375,
0.880859375,
0.501953125,
0.51318359375,
0.501953125
],
[
0.118408203125,
0.501953125,
0.486083984375,
0.50048828125,
0.486083984375,
0.515625,
0.118408203125,
0.5166015625
],
[
0.51318359375,
0.50439453125,
0.880859375,
0.50439453125,
0.880859375,
0.5185546875,
0.51318359375,
0.5185546875
],
[
0.118408203125,
0.5185546875,
0.484375,
0.5185546875,
0.484375,
0.5322265625,
0.118408203125,
0.5322265625
],
[
0.51123046875,
0.5185546875,
0.80419921875,
0.51953125,
0.80419921875,
0.53369140625,
0.51123046875,
0.533203125
],
[
0.118408203125,
0.53369140625,
0.2275390625,
0.53369140625,
0.2275390625,
0.548828125,
0.118408203125,
0.548828125
],
[
0.53076171875,
0.53515625,
0.87890625,
0.5361328125,
0.87890625,
0.55029296875,
0.53076171875,
0.5498046875
],
[
0.51318359375,
0.5517578125,
0.880859375,
0.5517578125,
0.880859375,
0.5654296875,
0.51318359375,
0.5654296875
],
[
0.2822265625,
0.56640625,
0.339111328125,
0.56640625,
0.339111328125,
0.58056640625,
0.2822265625,
0.58056640625
],
[
0.382080078125,
0.5654296875,
0.444091796875,
0.5654296875,
0.444091796875,
0.580078125,
0.382080078125,
0.580078125
],
[
0.51318359375,
0.56787109375,
0.880859375,
0.5693359375,
0.880859375,
0.5830078125,
0.51318359375,
0.58056640625
],
[
0.297119140625,
0.5849609375,
0.32568359375,
0.5849609375,
0.32568359375,
0.59716796875,
0.297119140625,
0.59716796875
],
[
0.4013671875,
0.58447265625,
0.426513671875,
0.58447265625,
0.426513671875,
0.59716796875,
0.4013671875,
0.59716796875
],
[
0.51904296875,
0.5849609375,
0.880859375,
0.5849609375,
0.880859375,
0.5986328125,
0.51904296875,
0.5986328125
],
[
0.273681640625,
0.5966796875,
0.349365234375,
0.5966796875,
0.349365234375,
0.61083984375,
0.273681640625,
0.61083984375
],
[
0.1435546875,
0.5859375,
0.21240234375,
0.5859375,
0.21240234375,
0.59716796875,
0.1435546875,
0.59716796875
],
[
0.160400390625,
0.599609375,
0.240966796875,
0.599609375,
0.240966796875,
0.60986328125,
0.160400390625,
0.60986328125
],
[
0.3837890625,
0.5986328125,
0.44580078125,
0.5986328125,
0.44580078125,
0.60986328125,
0.3837890625,
0.60986328125
],
[
0.51123046875,
0.599609375,
0.87744140625,
0.599609375,
0.87744140625,
0.61328125,
0.51123046875,
0.61328125
],
[
0.14697265625,
0.60986328125,
0.187255859375,
0.6123046875,
0.185546875,
0.62451171875,
0.145263671875,
0.62255859375
],
[
0.285400390625,
0.6123046875,
0.3359375,
0.6123046875,
0.3359375,
0.62353515625,
0.285400390625,
0.62353515625
],
[
0.3896484375,
0.6123046875,
0.4375,
0.6123046875,
0.4375,
0.62353515625,
0.3896484375,
0.62353515625
],
[
0.5146484375,
0.6162109375,
0.880859375,
0.6162109375,
0.880859375,
0.63134765625,
0.5146484375,
0.63134765625
],
[
0.145263671875,
0.62255859375,
0.185546875,
0.62451171875,
0.183837890625,
0.63720703125,
0.1435546875,
0.63525390625
],
[
0.287109375,
0.62353515625,
0.334228515625,
0.62353515625,
0.334228515625,
0.63818359375,
0.287109375,
0.63818359375
],
[
0.3896484375,
0.62353515625,
0.435791015625,
0.62353515625,
0.435791015625,
0.63818359375,
0.3896484375,
0.63818359375
],
[
0.391357421875,
0.63525390625,
0.43408203125,
0.63720703125,
0.432373046875,
0.65185546875,
0.3896484375,
0.6494140625
],
[
0.51123046875,
0.6318359375,
0.880859375,
0.6328125,
0.880859375,
0.64794921875,
0.51123046875,
0.646484375
],
[
0.145263671875,
0.63818359375,
0.2493896484375,
0.63818359375,
0.2493896484375,
0.6494140625,
0.145263671875,
0.6494140625
],
[
0.28955078125,
0.63818359375,
0.333251953125,
0.63818359375,
0.333251953125,
0.6494140625,
0.28955078125,
0.6494140625
],
[
0.14697265625,
0.65185546875,
0.2333984375,
0.65185546875,
0.2333984375,
0.6630859375,
0.14697265625,
0.6630859375
],
[
0.26611328125,
0.65185546875,
0.35693359375,
0.65185546875,
0.35693359375,
0.6630859375,
0.26611328125,
0.6630859375
],
[
0.365234375,
0.65185546875,
0.45751953125,
0.65185546875,
0.45751953125,
0.6630859375,
0.365234375,
0.6630859375
],
[
0.51318359375,
0.6484375,
0.87890625,
0.6484375,
0.87890625,
0.662109375,
0.51318359375,
0.662109375
],
[
0.300537109375,
0.6630859375,
0.32080078125,
0.6630859375,
0.32080078125,
0.67578125,
0.300537109375,
0.67578125
],
[
0.403076171875,
0.6630859375,
0.4248046875,
0.6630859375,
0.4248046875,
0.6767578125,
0.403076171875,
0.6767578125
],
[
0.51123046875,
0.66455078125,
0.87890625,
0.66455078125,
0.87890625,
0.677734375,
0.51123046875,
0.677734375
],
[
0.51123046875,
0.6796875,
0.880859375,
0.68115234375,
0.880859375,
0.6953125,
0.51123046875,
0.69482421875
],
[
0.1435546875,
0.6650390625,
0.18896484375,
0.6650390625,
0.18896484375,
0.6767578125,
0.1435546875,
0.6767578125
],
[
0.139404296875,
0.69140625,
0.4609375,
0.69140625,
0.4609375,
0.705078125,
0.139404296875,
0.705078125
],
[
0.509765625,
0.69482421875,
0.611328125,
0.69775390625,
0.611328125,
0.7119140625,
0.509765625,
0.708984375
],
[
0.1368408203125,
0.71044921875,
0.484375,
0.71044921875,
0.484375,
0.7236328125,
0.1368408203125,
0.7236328125
],
[
0.51318359375,
0.724609375,
0.5556640625,
0.724609375,
0.5556640625,
0.736328125,
0.51318359375,
0.736328125
],
[
0.5556640625,
0.724609375,
0.796875,
0.724609375,
0.796875,
0.73828125,
0.5556640625,
0.73828125
],
[
0.1175537109375,
0.7412109375,
0.486083984375,
0.7421875,
0.486083984375,
0.75732421875,
0.1175537109375,
0.755859375
],
[
0.118408203125,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.5146484375,
0.7451171875,
0.880859375,
0.7451171875,
0.880859375,
0.7587890625,
0.5146484375,
0.7587890625
],
[
0.118408203125,
0.7587890625,
0.484375,
0.7587890625,
0.484375,
0.7724609375,
0.118408203125,
0.7724609375
],
[
0.5146484375,
0.76123046875,
0.87890625,
0.76123046875,
0.87890625,
0.77490234375,
0.5146484375,
0.77490234375
],
[
0.118408203125,
0.775390625,
0.486083984375,
0.775390625,
0.486083984375,
0.7890625,
0.118408203125,
0.7890625
],
[
0.51318359375,
0.77783203125,
0.880859375,
0.77783203125,
0.880859375,
0.79150390625,
0.51318359375,
0.79150390625
],
[
0.12005615234375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8046875,
0.12005615234375,
0.8046875
],
[
0.51318359375,
0.79345703125,
0.880859375,
0.79345703125,
0.880859375,
0.80712890625,
0.51318359375,
0.80712890625
],
[
0.1158447265625,
0.8056640625,
0.486083984375,
0.80712890625,
0.486083984375,
0.82177734375,
0.1158447265625,
0.82080078125
],
[
0.5146484375,
0.81005859375,
0.8818359375,
0.81005859375,
0.8818359375,
0.82373046875,
0.5146484375,
0.82373046875
],
[
0.51123046875,
0.82470703125,
0.8818359375,
0.826171875,
0.8818359375,
0.84033203125,
0.51123046875,
0.8388671875
],
[
0.1435546875,
0.83056640625,
0.48779296875,
0.83203125,
0.48779296875,
0.845703125,
0.1435546875,
0.84423828125
],
[
0.1158447265625,
0.84326171875,
0.489501953125,
0.84423828125,
0.489501953125,
0.85791015625,
0.1158447265625,
0.85693359375
],
[
0.509765625,
0.84130859375,
0.8818359375,
0.8427734375,
0.8818359375,
0.85546875,
0.509765625,
0.85498046875
],
[
0.1175537109375,
0.85546875,
0.438232421875,
0.85546875,
0.438232421875,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.51318359375,
0.85986328125,
0.603515625,
0.85986328125,
0.603515625,
0.87158203125,
0.51318359375,
0.87158203125
],
[
0.145263671875,
0.8681640625,
0.44921875,
0.8681640625,
0.44921875,
0.8818359375,
0.145263671875,
0.8818359375
],
[
0.145263671875,
0.88134765625,
0.486083984375,
0.88134765625,
0.486083984375,
0.89453125,
0.145263671875,
0.89453125
],
[
0.53662109375,
0.8798828125,
0.880859375,
0.88134765625,
0.880859375,
0.8955078125,
0.53662109375,
0.89453125
],
[
0.118408203125,
0.892578125,
0.3359375,
0.8935546875,
0.3359375,
0.90673828125,
0.118408203125,
0.90625
],
[
0.32666015625,
0.892578125,
0.484375,
0.8935546875,
0.484375,
0.90673828125,
0.32666015625,
0.90625
],
[
0.51123046875,
0.8935546875,
0.880859375,
0.8935546875,
0.880859375,
0.90673828125,
0.51123046875,
0.90673828125
],
[
0.1158447265625,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.1158447265625,
0.91845703125
],
[
0.509765625,
0.90478515625,
0.8369140625,
0.90380859375,
0.8369140625,
0.91845703125,
0.509765625,
0.91943359375
]
] |
[
"探讨。除了评估OOD退化之外，这些用于英",
"以研究解析的通用性成为可能（有关其确切大",
"语RST的大型数据集的存在使我们能够测试",
"小细分，请参见附录B中的表12）。几乎所有",
"不同的联合训练策略，以提高英语RST解析",
"关于RST成分解析的工作都只使用折叠的粗",
"的准确性和通用性。",
"略关系类i。e。17个RST-DT标签)；我们将",
"我们在这里的主要目标是：1）展示基于RST-",
"在下面的评估中遵循这一做法，但在我们的分",
"DT的英语RST解析的可泛化性局限性，并为",
"析中也会提到一些与细粒度标签有关的问题。",
"用户量化；2）探索可泛化性问题的原因，重",
"此外，我们对两个语料库中的s-n核模式进行",
"点是训练集的体裁组成，为健壮的语篇解析所",
"了分析，证实了它们非常相似：NS以77种为",
"需的数据类型指明了方向；3）根据我们的实",
"主。7RST-DT中的%，类似于74。8口香糖新",
"验结果，促进RST解析的多体裁基准。总的来",
"闻中的%；有趣的是，口香糖中的比例总体上",
"说，我们发现，无论模型架构如何，不同的训",
"略低（70。1%)",
"练数据都能更好地概括看不见的体裁",
"还有其他语料库在RST依赖关系中进行了",
"注释，例如SciDTB（YangandLi,2018）和其他",
"2",
"Related Work",
"语言的完整RST组成树库（有关概述，请参见",
"Zeldesetal.2021）；由于我们专注于分层的英",
"2.1 English RST Corpora",
"语RST成分解析，因此我们将使用表1中的",
"RST话语树库（RST-DT，Carlsonetal.2003）是",
"语料库进行以下实验：它们在内容，词汇，领",
"标准的英语RST基准，数据来自PennTreebank",
"域以及我们所有",
"(PTB，Marcus etal.1993）的WSJ部分。另",
"+",
"个带有人类注释的英语RST语料库是George-",
"2.2RSTDiscourseParsers",
"townUniversityMmultilayer（GUM）语料库，该",
"已经提出了几种用于RST解析的方法，主要通",
"语料库可在线免费获得，涵盖12种书面和口",
"过BOTTOM-UP与。TOP-DOWN方法，以及使",
"头类型（Zeldes2017）。GUM不断增长，每个",
"用的算法（通常是神经移位-减少架构）。表2",
"版本都添加了新数据。对于本文，我们使用了",
"在三个指标上比较了最近的高分解析器：Span",
"语料库的第8版，如表1中有关RST-DT的信",
"（子树是否跨越正确的EDU），N核性（边缘",
"息旁边的描述（GUM的当前版本V9在提交时",
"是否指向正确的方向）和R关系（标签是否正",
"尚未提供，现在包含213个文档，203K令牌",
"确）。在Morey etal.（2017）之后，我们在二叉",
"和26KEDU，下一个版本10设置为添加4个",
"树上使用更严格的原始Parseval过程",
"其他新类型)。",
"如表2所示，BOTTOM-UP和TOP-DOWN方",
"法目前非常接近，并且所有最先进的（SOTA）",
"RST-DT",
"GUM V8",
"系统都依赖于变压器字嵌入作为主要输人。为",
"385",
"193",
"了排除解析器体系结构的影响，我们使用Guz",
"347/_ 3/38",
"documents",
"train/dev/test",
"145/24/24",
"andCarenini（2020）的最佳SpanBERT-NoCoref",
"tokens",
"203,352",
"180,851",
"设置对BOTTOM-UP进行了测试，并使用Liu",
"EDUs",
"21,789",
"23,107",
"21,903",
"etal.（2021）的TOP-DOWN进行了测试，因为",
"relation instances",
"20,163",
"relation labels",
"78/ 17 classes",
"32/15 classes",
"它们的代码是公共的，并且都具有接近SOTA",
"14",
"12",
"的S和N指标，这对于整个语料库之间的标",
"签清单之间的差异更为稳健（相关的超参数和",
"genres",
"Table 1: Overview of the English RST Corpora.",
"XMATHX_.",
"如表1所示，尽管GUMV8在令牌和文档数",
"2.3",
"Cross-Genre Variation in RST",
"元（EDU）和关系的实例，因为RST-DT的新",
"量上小于RST-DT，但它包含更多基本话语单",
"先前的工作表明，RST树在文本类型，5的关系",
"闻专线单元平均更长。对于两个语料库，EDU",
"分布(Taboada and Lavid,2003;Zeldes,2018)中",
"分割准则是相同的，并且标签集非常相似，对",
"差异很大，它们的信号方式（例如。g。Eggins",
"于不连续的EDU，两个语料库都使用SAME-",
"and Martin 1997; Liu 2019; Demberg et al. 2019)",
"UNIT伪关系。GUM的12种流派使进行实验",
"以及跨语言的变化（Caoetal.2018；Stedeand",
"Neumann 2014 ; Iruskieta et al. 2013 ; Redeker",
"2Though we focus on English in this paper, understand-",
"ing the size and nature of data needed in English will hope-",
"et al. 2012 ; da Cunha et al. 2011 ; Pisarevskaya",
"fully shed light on what other languages may require",
"et al. 2017)",
"3There is no established dev partition for RST-DT",
"4Previous work has identified 4 sub-genres in the WSJ",
"5We use\"text type\",\"genre\",and\"domain\"interchange",
"data (Webber, 2009), but these are v",
"very unevenly distributed,",
"ably here to mean different kinds of data, though they are not",
"and 29 RST-DT documents were not included in that analysis.",
"exactly the same (e.g. see Biber 1988 and Lee 2001)"
] |
{
"arxiv_id": "2302.06488",
"language": "zh",
"page_idx": 2
}
|
||
2302.06488-zh-4
|
2302.06488-zh-4.jpg
|
解析器，这些解析器在 Guz and Carenini (2020) (BOTTOM-UP)和Liuet al.(2021)(TOP-DOWN ) 中保留在每个数据集的 test 集合上的分数(由于 RST-DT 没有 dev 分区的 XEDX
我们实验的一个可能的局限性是语料库之间的注释差异，尤其是对于它们的关系清单，它们相似但不相同。因此，我们省略了表3 和 4中的 R 分数，尽管我们将在下面更详细地返回关系。但是，对于 S 和 N, 我们希望差异很小：两个语料库都使用相同的分割准则，并且使用诸如可删除性和主观突出性之类的测试将不太突出的卫星单元从属于更突出的核的相同准则。尽管 EDU 分割不是我们的重点，但我们可以通过交叉测试 Gessler et al.(2021)中当前的 SOTAEDU 分割器来确认分割准则的兼容性。结果显示，在 RST-DT 上进行相当适度的降解训练，在GUM上进行测试，从94。9至89。9,而相反的方向显示为 92。9至 91 下降(有关完整数字，请参见附录 A 中的表 11 )。我们推测，由于语法的原因，分段可以很好地泛化，因此任务的稀疏性和特定于域的性质较少，类似于子句边界检测
对于跨语料库 S 和 N 分数，在 RST-DT 上训练以解析来自 GUM 的 OOD 数据时，两个解析器都显示出非常显着的退化，如表3 和4 所示。例如，使用 BOTTOM-UP 体系结构，N 的$\mathbb{E}???$ 会降级。我们也很好奇这是否也适用于 GUM 的新闻类型：实际上，仅在 GUM 新闻上进行测试可以使用 BOTTOM-UP 解析器将降级降低约 50%, 但即使该模型仅在新闻上进行了训练(尽管来自更老的《华尔街日报》新闻数据)。相比之下，经过 GUM 训练的模型在 RST-DT 上的得分实际上要比在 GUM 上的得分高，但略有提高 2。8在 S 和 +0 上。使用 BOTTOM-UP 解析器在 N 上设置 4。在 TOP-DOWN 解析器性能中观察到相同的降级模式：总体上严重降级， 并且在训练新闻和测试其他类型时更糟。$^{6}$
<table>
<tbody>
<tr>
<th>train</th>
<th>test</th>
<th>$S$</th>
<th>$N$</th>
<th>$R$</th>
</tr>
<tr>
<td rowspan="3">RST- DT</td>
<td>RST- DT</td>
<td>76.5</td>
<td>65.9</td>
<td>54.8</td>
</tr>
<tr>
<td>$GUM$</td>
<td>65.3 (-11.2)</td>
<td>49.5 5(-16.4)</td>
<td>- </td>
</tr>
<tr>
<td>$GUM$ news</td>
<td>71.0(-5.5)</td>
<td>57.5(-8.4)</td>
<td>- </td>
</tr>
<tr>
<td rowspan="3">$GUM$</td>
<td>$GUM$</td>
<td>69.9</td>
<td>57.0</td>
<td>48.5</td>
</tr>
<tr>
<td>RST- DT</td>
<td>72.7 (+2.8)</td>
<td>57.4 +0.4)</td>
<td>- </td>
</tr>
<tr>
<td>$\mathbf{GUM}$ news</td>
<td>71.6</td>
<td>58.5</td>
<td>49.5</td>
</tr>
</tbody>
</table>
Table 3: Cross-Corpus Results (5 run average) of the
BOTTOM-UP Parser from Guz and Carenini (2020) .
6We further analyzed left vs. right-branching trees during testing. In both corpora, NS outperformed SN: RST-DT in-domain scores are F1=82.6% vs.73.4 % for NS vs. SN; GUM V8 has 81.6% vs. 72.4 % , meaning SN transitions are $`$easier,
<table>
<tbody>
<tr>
<th>train</th>
<th>test</th>
<th>$S$</th>
<th>$N$</th>
<th>$R$</th>
</tr>
<tr>
<td rowspan="3">RST- DT</td>
<td>RST- DT</td>
<td>76.5</td>
<td>65.2</td>
<td>54.2</td>
</tr>
<tr>
<td>$GUM$</td>
<td>66.2 (-10.3)</td>
<td>50.8 (-14.4)</td>
<td>- </td>
</tr>
<tr>
<td>$GUM$ news 1</td>
<td>67.9(-8.6)</td>
<td>55.8 (-9.4) 1</td>
<td>-</td>
</tr>
<tr>
<td rowspan="3">$GUM$</td>
<td>$GUM$</td>
<td>68.6</td>
<td>54.9</td>
<td>46.1</td>
</tr>
<tr>
<td>RST- DT</td>
<td>71.1 (+2.5)</td>
<td>(+1.0) 55.9</td>
<td>- </td>
</tr>
<tr>
<td>$\mathbf{GUM}$ news</td>
<td>73.4</td>
<td>63.3</td>
<td>57.2</td>
</tr>
</tbody>
</table>
Table 4: Cross-Corpus Results (5 run average) of the
TOP-DOWN Parser from Liu et al. (2021) .
我们将此结果解释为意味着与 EDU 分段不同，在EDU 分段中，结果在两个方向上都很接近，train 和 test 数据的体裁组成在 RST 成分解析的通用性中起着至关重要的作用，而与解析器体系结构无关。因此，我们选择 BOTTOMUP 解析器进行本文其余部分中介绍的实验， 因为它在两个语料库上都优于 TOP-DOWN 解析器
要明确的是，结果并不表明在 GUM 上进行训练是在 RST-DT 上实现最佳性能的一种方式：GUM 模型的得分仍然比 RST-DT 训练的S 和 N 模型低近 4 分和 8 分。然而，对于已经看到一些新闻的 GUM 模型来说，RST-DT 新闻数据似乎并不令人惊讶，总而言之，鉴于GUM 模型经过训练可以解决的广泛类型库存， RST-DT 数据似乎是一个相对“容易”的目标
这些结果还证实，在多个流派上进行训练， 每个流派的文档相对较少，可以带来良好的性能，而在 RST-DT 的非常狭窄的 WSJ 域上只有很小的降级。虽然我们认为这个结果本身很重要，并建议 RST 解析工作应该更多地关注多类型语料库作为基准，但它留下了一个问题： 我们可以结合来自两个语料库的数据来提高RST-DT 上的英语 RST 解析性能吗？
# 3.2 Joint Training
如果 RST 解析跨域泛化，则有可能通过联合训练看到收益。为此，我们比较了三种方法：朴素串联，模型堆叠和预训练，我们在 RST-DT 基准上进行了评估
对于连接( CONCAT ), 我们必须映射跨语料库的关系。与以前的工作一样，我们的目标是粗略的关系类，其中大多数在两个语料库中都是相同的。例外情况包括 GUM 的 PHATIC 关系用于不流畅和反向通道(由于会话数据),GUM的PREPARATION 以及 RST-DT 的 TEMPORAL 和COMPARISON 类之间的映射，这些映射不同地映射到 GUM 的 CIRCUMSTANCE 和 SEQUENCE ,或 GUM 的 ANTITHESIS , 具体取决于。虽然受这些不匹配影响的关系实例数量不多，但它不是可以忽略的(13。3%),我们预计这种方法不会超过域内培训，主要是报告其完整性。我
|
[
[
0.118408203125,
0.08612060546875,
0.484375,
0.08612060546875,
0.484375,
0.09979248046875,
0.118408203125,
0.09979248046875
],
[
0.66748046875,
0.0872802734375,
0.6884765625,
0.0872802734375,
0.6884765625,
0.1009521484375,
0.66748046875,
0.1009521484375
],
[
0.75390625,
0.0872802734375,
0.7734375,
0.0872802734375,
0.7734375,
0.1009521484375,
0.75390625,
0.1009521484375
],
[
0.8369140625,
0.0872802734375,
0.8564453125,
0.0872802734375,
0.8564453125,
0.1009521484375,
0.8369140625,
0.1009521484375
],
[
0.587890625,
0.0885009765625,
0.61376953125,
0.0885009765625,
0.61376953125,
0.09979248046875,
0.587890625,
0.09979248046875
],
[
0.5224609375,
0.0885009765625,
0.55517578125,
0.0885009765625,
0.55517578125,
0.09979248046875,
0.5224609375,
0.09979248046875
],
[
0.1175537109375,
0.1009521484375,
0.486083984375,
0.1015625,
0.486083984375,
0.11639404296875,
0.1175537109375,
0.11517333984375
],
[
0.6708984375,
0.1051025390625,
0.7021484375,
0.1051025390625,
0.7021484375,
0.1175537109375,
0.6708984375,
0.1175537109375
],
[
0.75732421875,
0.1051025390625,
0.78662109375,
0.1051025390625,
0.78662109375,
0.1175537109375,
0.75732421875,
0.1175537109375
],
[
0.84033203125,
0.1051025390625,
0.873046875,
0.1051025390625,
0.873046875,
0.1175537109375,
0.84033203125,
0.1175537109375
],
[
0.587890625,
0.105712890625,
0.640625,
0.105712890625,
0.640625,
0.11639404296875,
0.587890625,
0.11639404296875
],
[
0.52294921875,
0.105712890625,
0.57666015625,
0.105712890625,
0.57666015625,
0.1175537109375,
0.52294921875,
0.1175537109375
],
[
0.6689453125,
0.11639404296875,
0.74462890625,
0.11639404296875,
0.74462890625,
0.1312255859375,
0.6689453125,
0.1312255859375
],
[
0.755859375,
0.11639404296875,
0.8310546875,
0.11639404296875,
0.8310546875,
0.1312255859375,
0.755859375,
0.1312255859375
],
[
0.58837890625,
0.1175537109375,
0.62744140625,
0.1175537109375,
0.62744140625,
0.1300048828125,
0.58837890625,
0.1300048828125
],
[
0.12005615234375,
0.1181640625,
0.483642578125,
0.1181640625,
0.483642578125,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.58837890625,
0.1312255859375,
0.66015625,
0.1312255859375,
0.66015625,
0.142578125,
0.58837890625,
0.142578125
],
[
0.6689453125,
0.1312255859375,
0.73779296875,
0.1312255859375,
0.73779296875,
0.142578125,
0.6689453125,
0.142578125
],
[
0.755859375,
0.1312255859375,
0.82275390625,
0.1312255859375,
0.82275390625,
0.142578125,
0.755859375,
0.142578125
],
[
0.8427734375,
0.1353759765625,
0.8486328125,
0.1353759765625,
0.8486328125,
0.14013671875,
0.8427734375,
0.14013671875
],
[
0.8388671875,
0.1478271484375,
0.87158203125,
0.1478271484375,
0.87158203125,
0.1602783203125,
0.8388671875,
0.1602783203125
],
[
0.1175537109375,
0.1341552734375,
0.418212890625,
0.1341552734375,
0.418212890625,
0.1478271484375,
0.1175537109375,
0.1478271484375
],
[
0.5224609375,
0.1484375,
0.56103515625,
0.1484375,
0.56103515625,
0.1602783203125,
0.5224609375,
0.1602783203125
],
[
0.58837890625,
0.1484375,
0.62744140625,
0.1484375,
0.62744140625,
0.1602783203125,
0.58837890625,
0.1602783203125
],
[
0.6708984375,
0.1484375,
0.7021484375,
0.1484375,
0.7021484375,
0.1602783203125,
0.6708984375,
0.1602783203125
],
[
0.755859375,
0.1484375,
0.78662109375,
0.1484375,
0.78662109375,
0.1602783203125,
0.755859375,
0.1602783203125
],
[
0.58837890625,
0.162109375,
0.640625,
0.162109375,
0.640625,
0.17333984375,
0.58837890625,
0.17333984375
],
[
0.1351318359375,
0.149658203125,
0.483642578125,
0.149658203125,
0.483642578125,
0.163330078125,
0.1351318359375,
0.163330078125
],
[
0.67236328125,
0.162109375,
0.74072265625,
0.162109375,
0.74072265625,
0.17333984375,
0.67236328125,
0.17333984375
],
[
0.755859375,
0.162109375,
0.8251953125,
0.162109375,
0.8251953125,
0.17333984375,
0.755859375,
0.17333984375
],
[
0.12005615234375,
0.168701171875,
0.483642578125,
0.168701171875,
0.483642578125,
0.1787109375,
0.12005615234375,
0.1787109375
],
[
0.59033203125,
0.174560546875,
0.66162109375,
0.174560546875,
0.66162109375,
0.1859130859375,
0.59033203125,
0.1859130859375
],
[
0.6708984375,
0.174560546875,
0.7021484375,
0.174560546875,
0.7021484375,
0.18701171875,
0.6708984375,
0.18701171875
],
[
0.755859375,
0.174560546875,
0.78662109375,
0.174560546875,
0.78662109375,
0.18701171875,
0.755859375,
0.18701171875
],
[
0.8388671875,
0.1728515625,
0.875,
0.1728515625,
0.875,
0.188232421875,
0.8388671875,
0.188232421875
],
[
0.118408203125,
0.182861328125,
0.486083984375,
0.182861328125,
0.486083984375,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.12005615234375,
0.198974609375,
0.486083984375,
0.198974609375,
0.486083984375,
0.2120361328125,
0.12005615234375,
0.2120361328125
],
[
0.509765625,
0.196533203125,
0.8818359375,
0.196533203125,
0.8818359375,
0.21142578125,
0.509765625,
0.21142578125
],
[
0.118408203125,
0.21435546875,
0.484375,
0.21435546875,
0.484375,
0.22802734375,
0.118408203125,
0.22802734375
],
[
0.51123046875,
0.21142578125,
0.796875,
0.21142578125,
0.796875,
0.2244873046875,
0.51123046875,
0.2244873046875
],
[
0.1175537109375,
0.23095703125,
0.483642578125,
0.23095703125,
0.483642578125,
0.24462890625,
0.1175537109375,
0.24462890625
],
[
0.52880859375,
0.2410888671875,
0.880859375,
0.2423095703125,
0.880859375,
0.257080078125,
0.52880859375,
0.255859375
],
[
0.118408203125,
0.2476806640625,
0.484375,
0.2476806640625,
0.484375,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.118408203125,
0.26318359375,
0.484375,
0.26318359375,
0.484375,
0.276611328125,
0.118408203125,
0.276611328125
],
[
0.51318359375,
0.2587890625,
0.880859375,
0.2587890625,
0.880859375,
0.2724609375,
0.51318359375,
0.2724609375
],
[
0.1175537109375,
0.27783203125,
0.486083984375,
0.279052734375,
0.486083984375,
0.29345703125,
0.1175537109375,
0.292236328125
],
[
0.51123046875,
0.27490234375,
0.87744140625,
0.27490234375,
0.87744140625,
0.2880859375,
0.51123046875,
0.2880859375
],
[
0.1175537109375,
0.29443359375,
0.486083984375,
0.29345703125,
0.486083984375,
0.30810546875,
0.1175537109375,
0.308837890625
],
[
0.51123046875,
0.290283203125,
0.87890625,
0.290283203125,
0.87890625,
0.303955078125,
0.51123046875,
0.303955078125
],
[
0.1175537109375,
0.31005859375,
0.484375,
0.31005859375,
0.484375,
0.32373046875,
0.1175537109375,
0.32373046875
],
[
0.51318359375,
0.30810546875,
0.87744140625,
0.30810546875,
0.87744140625,
0.318359375,
0.51318359375,
0.318359375
],
[
0.1175537109375,
0.325927734375,
0.483642578125,
0.325927734375,
0.483642578125,
0.339111328125,
0.1175537109375,
0.339111328125
],
[
0.51123046875,
0.32373046875,
0.87744140625,
0.32373046875,
0.87744140625,
0.33740234375,
0.51123046875,
0.33740234375
],
[
0.1175537109375,
0.342529296875,
0.483642578125,
0.342529296875,
0.483642578125,
0.355712890625,
0.1175537109375,
0.355712890625
],
[
0.51123046875,
0.33837890625,
0.880859375,
0.33837890625,
0.880859375,
0.3515625,
0.51123046875,
0.3515625
],
[
0.1175537109375,
0.358154296875,
0.486083984375,
0.359375,
0.486083984375,
0.372314453125,
0.1175537109375,
0.371826171875
],
[
0.5146484375,
0.35693359375,
0.5498046875,
0.35693359375,
0.5498046875,
0.3681640625,
0.5146484375,
0.3681640625
],
[
0.1175537109375,
0.3759765625,
0.486083984375,
0.3759765625,
0.486083984375,
0.3896484375,
0.1175537109375,
0.3896484375
],
[
0.52880859375,
0.37060546875,
0.880859375,
0.371826171875,
0.880859375,
0.385986328125,
0.52880859375,
0.385498046875
],
[
0.118408203125,
0.392578125,
0.486083984375,
0.392578125,
0.486083984375,
0.40625,
0.118408203125,
0.40625
],
[
0.51318359375,
0.38720703125,
0.87890625,
0.38720703125,
0.87890625,
0.40087890625,
0.51318359375,
0.40087890625
],
[
0.51123046875,
0.40380859375,
0.87890625,
0.40380859375,
0.87890625,
0.41748046875,
0.51123046875,
0.41748046875
],
[
0.12005615234375,
0.410400390625,
0.484375,
0.410400390625,
0.484375,
0.42041015625,
0.12005615234375,
0.42041015625
],
[
0.1175537109375,
0.4228515625,
0.32080078125,
0.42333984375,
0.32080078125,
0.438232421875,
0.1175537109375,
0.43701171875
],
[
0.51123046875,
0.418701171875,
0.880859375,
0.419189453125,
0.880859375,
0.43408203125,
0.51123046875,
0.432861328125
],
[
0.1368408203125,
0.440673828125,
0.486083984375,
0.440673828125,
0.486083984375,
0.45361328125,
0.1368408203125,
0.45361328125
],
[
0.5146484375,
0.435791015625,
0.880859375,
0.435791015625,
0.880859375,
0.449462890625,
0.5146484375,
0.449462890625
],
[
0.118408203125,
0.4560546875,
0.484375,
0.4560546875,
0.484375,
0.47021484375,
0.118408203125,
0.47021484375
],
[
0.51318359375,
0.453125,
0.880859375,
0.453125,
0.880859375,
0.466064453125,
0.51318359375,
0.466064453125
],
[
0.51123046875,
0.466064453125,
0.87744140625,
0.468505859375,
0.87744140625,
0.482666015625,
0.51123046875,
0.48095703125
],
[
0.12005615234375,
0.47265625,
0.48779296875,
0.47265625,
0.48779296875,
0.486328125,
0.12005615234375,
0.486328125
],
[
0.118408203125,
0.488037109375,
0.486083984375,
0.488037109375,
0.486083984375,
0.501953125,
0.118408203125,
0.501953125
],
[
0.51123046875,
0.482666015625,
0.875,
0.48388671875,
0.875,
0.498779296875,
0.51123046875,
0.49755859375
],
[
0.1175537109375,
0.50439453125,
0.486083984375,
0.5029296875,
0.486083984375,
0.51708984375,
0.1175537109375,
0.5185546875
],
[
0.53076171875,
0.50048828125,
0.880859375,
0.50048828125,
0.880859375,
0.51416015625,
0.53076171875,
0.51416015625
],
[
0.118408203125,
0.51953125,
0.483642578125,
0.51953125,
0.483642578125,
0.53369140625,
0.118408203125,
0.53369140625
],
[
0.51318359375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.118408203125,
0.5361328125,
0.486083984375,
0.5361328125,
0.486083984375,
0.5498046875,
0.118408203125,
0.5498046875
],
[
0.5146484375,
0.53369140625,
0.87890625,
0.53369140625,
0.87890625,
0.54736328125,
0.5146484375,
0.54736328125
],
[
0.118408203125,
0.552734375,
0.484375,
0.552734375,
0.484375,
0.56640625,
0.118408203125,
0.56640625
],
[
0.51318359375,
0.548828125,
0.87890625,
0.548828125,
0.87890625,
0.5625,
0.51318359375,
0.5625
],
[
0.1175537109375,
0.56787109375,
0.484375,
0.568359375,
0.484375,
0.58203125,
0.1175537109375,
0.58056640625
],
[
0.51123046875,
0.56396484375,
0.87890625,
0.56396484375,
0.87890625,
0.57763671875,
0.51123046875,
0.57763671875
],
[
0.12005615234375,
0.5849609375,
0.484375,
0.5849609375,
0.484375,
0.5986328125,
0.12005615234375,
0.5986328125
],
[
0.51123046875,
0.580078125,
0.880859375,
0.58056640625,
0.880859375,
0.595703125,
0.51123046875,
0.59423828125
],
[
0.12005615234375,
0.6025390625,
0.484375,
0.6025390625,
0.484375,
0.61328125,
0.12005615234375,
0.61328125
],
[
0.51318359375,
0.59716796875,
0.880859375,
0.59716796875,
0.880859375,
0.61083984375,
0.51318359375,
0.61083984375
],
[
0.12005615234375,
0.61767578125,
0.486083984375,
0.61767578125,
0.486083984375,
0.63134765625,
0.12005615234375,
0.63134765625
],
[
0.509765625,
0.6123046875,
0.796875,
0.61328125,
0.796875,
0.62744140625,
0.509765625,
0.626953125
],
[
0.1175537109375,
0.6318359375,
0.486083984375,
0.6328125,
0.486083984375,
0.64794921875,
0.1175537109375,
0.646484375
],
[
0.5146484375,
0.642578125,
0.5615234375,
0.642578125,
0.5615234375,
0.65380859375,
0.5146484375,
0.65380859375
],
[
0.55322265625,
0.64013671875,
0.67333984375,
0.64111328125,
0.67333984375,
0.65625,
0.55322265625,
0.65478515625
],
[
0.12005615234375,
0.6494140625,
0.484375,
0.6494140625,
0.484375,
0.6630859375,
0.12005615234375,
0.6630859375
],
[
0.118408203125,
0.6650390625,
0.44580078125,
0.6650390625,
0.44580078125,
0.6787109375,
0.118408203125,
0.6787109375
],
[
0.5146484375,
0.6630859375,
0.880859375,
0.6630859375,
0.880859375,
0.6767578125,
0.5146484375,
0.6767578125
],
[
0.5146484375,
0.6787109375,
0.880859375,
0.6787109375,
0.880859375,
0.6923828125,
0.5146484375,
0.6923828125
],
[
0.358642578125,
0.6953125,
0.378662109375,
0.6953125,
0.378662109375,
0.708984375,
0.358642578125,
0.708984375
],
[
0.443359375,
0.6953125,
0.462646484375,
0.6953125,
0.462646484375,
0.708984375,
0.443359375,
0.708984375
],
[
0.5146484375,
0.69482421875,
0.8818359375,
0.69482421875,
0.8818359375,
0.708984375,
0.5146484375,
0.708984375
],
[
0.193115234375,
0.69677734375,
0.219970703125,
0.69677734375,
0.219970703125,
0.708984375,
0.193115234375,
0.708984375
],
[
0.1259765625,
0.69677734375,
0.160400390625,
0.69677734375,
0.160400390625,
0.7080078125,
0.1259765625,
0.7080078125
],
[
0.274658203125,
0.69677734375,
0.28955078125,
0.69677734375,
0.28955078125,
0.7080078125,
0.274658203125,
0.7080078125
],
[
0.443359375,
0.7119140625,
0.480224609375,
0.7119140625,
0.480224609375,
0.72802734375,
0.443359375,
0.72802734375
],
[
0.51318359375,
0.71142578125,
0.66015625,
0.71142578125,
0.66015625,
0.724609375,
0.51318359375,
0.724609375
],
[
0.274658203125,
0.71337890625,
0.307373046875,
0.71337890625,
0.307373046875,
0.7255859375,
0.274658203125,
0.7255859375
],
[
0.124267578125,
0.71337890625,
0.2476806640625,
0.71337890625,
0.2476806640625,
0.72705078125,
0.124267578125,
0.72705078125
],
[
0.359375,
0.71337890625,
0.39208984375,
0.71337890625,
0.39208984375,
0.7255859375,
0.359375,
0.7255859375
],
[
0.19482421875,
0.7255859375,
0.2301025390625,
0.7255859375,
0.2301025390625,
0.73828125,
0.19482421875,
0.73828125
],
[
0.276123046875,
0.72705078125,
0.35107421875,
0.72705078125,
0.35107421875,
0.73828125,
0.276123046875,
0.73828125
],
[
0.361083984375,
0.72705078125,
0.43408203125,
0.72705078125,
0.43408203125,
0.73828125,
0.361083984375,
0.73828125
],
[
0.53076171875,
0.72802734375,
0.880859375,
0.72802734375,
0.880859375,
0.7412109375,
0.53076171875,
0.7412109375
],
[
0.193115234375,
0.73681640625,
0.268798828125,
0.74072265625,
0.267822265625,
0.7548828125,
0.19140625,
0.751953125
],
[
0.274658203125,
0.7392578125,
0.343505859375,
0.7392578125,
0.343505859375,
0.75048828125,
0.274658203125,
0.75048828125
],
[
0.359375,
0.73828125,
0.429931640625,
0.73828125,
0.429931640625,
0.7529296875,
0.359375,
0.7529296875
],
[
0.447509765625,
0.74365234375,
0.453369140625,
0.74365234375,
0.453369140625,
0.7490234375,
0.447509765625,
0.7490234375
],
[
0.5146484375,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.5146484375,
0.75732421875
],
[
0.1292724609375,
0.75732421875,
0.166259765625,
0.75732421875,
0.166259765625,
0.76953125,
0.1292724609375,
0.76953125
],
[
0.19482421875,
0.75732421875,
0.231689453125,
0.75732421875,
0.231689453125,
0.76953125,
0.19482421875,
0.76953125
],
[
0.276123046875,
0.75732421875,
0.306396484375,
0.75732421875,
0.306396484375,
0.76953125,
0.276123046875,
0.76953125
],
[
0.361083984375,
0.75732421875,
0.391357421875,
0.75732421875,
0.391357421875,
0.76953125,
0.361083984375,
0.76953125
],
[
0.444091796875,
0.75732421875,
0.47607421875,
0.75732421875,
0.47607421875,
0.76953125,
0.444091796875,
0.76953125
],
[
0.51123046875,
0.75830078125,
0.880859375,
0.7587890625,
0.880859375,
0.77392578125,
0.51123046875,
0.7724609375
],
[
0.19482421875,
0.77099609375,
0.246826171875,
0.77099609375,
0.246826171875,
0.7822265625,
0.19482421875,
0.7822265625
],
[
0.273681640625,
0.7685546875,
0.345947265625,
0.77099609375,
0.345947265625,
0.78515625,
0.273681640625,
0.783203125
],
[
0.361083984375,
0.77099609375,
0.429931640625,
0.77099609375,
0.429931640625,
0.7822265625,
0.361083984375,
0.7822265625
],
[
0.19140625,
0.77978515625,
0.267822265625,
0.7822265625,
0.267822265625,
0.79638671875,
0.19140625,
0.79443359375
],
[
0.444091796875,
0.78076171875,
0.478515625,
0.78076171875,
0.478515625,
0.79638671875,
0.444091796875,
0.79638671875
],
[
0.5146484375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.5146484375,
0.7890625
],
[
0.274658203125,
0.783203125,
0.307373046875,
0.783203125,
0.307373046875,
0.7958984375,
0.274658203125,
0.7958984375
],
[
0.361083984375,
0.783203125,
0.39208984375,
0.783203125,
0.39208984375,
0.7958984375,
0.361083984375,
0.7958984375
],
[
0.515625,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.515625,
0.8056640625
],
[
0.1175537109375,
0.8056640625,
0.48779296875,
0.8056640625,
0.48779296875,
0.8193359375,
0.1175537109375,
0.8193359375
],
[
0.5146484375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.118408203125,
0.82177734375,
0.46337890625,
0.82177734375,
0.46337890625,
0.83203125,
0.118408203125,
0.83203125
],
[
0.51123046875,
0.82373046875,
0.87890625,
0.82373046875,
0.87890625,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.51123046875,
0.84033203125,
0.87890625,
0.84033203125,
0.87890625,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.14111328125,
0.85546875,
0.48779296875,
0.85791015625,
0.48779296875,
0.87158203125,
0.14111328125,
0.869140625
],
[
0.5283203125,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.5283203125,
0.87060546875
],
[
0.1175537109375,
0.869140625,
0.486083984375,
0.8681640625,
0.486083984375,
0.8818359375,
0.1175537109375,
0.8828125
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.1175537109375,
0.8828125,
0.484375,
0.8828125,
0.484375,
0.8935546875,
0.1175537109375,
0.8935546875
],
[
0.1158447265625,
0.892578125,
0.486083984375,
0.89453125,
0.486083984375,
0.90771484375,
0.1158447265625,
0.90625
],
[
0.51318359375,
0.88818359375,
0.880859375,
0.88818359375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.12005615234375,
0.90478515625,
0.162841796875,
0.90673828125,
0.162109375,
0.91845703125,
0.118408203125,
0.91650390625
],
[
0.5146484375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.5146484375,
0.91845703125
]
] |
[
"解析器，这些解析器在GuzandCarenini（2020）",
"s",
"N",
"R",
"test",
"train",
"(BOTTOM-UP）和Liuetal.(2021)（TOP-DOWN",
"76.5",
"65.2",
"54.2",
"RST-DT",
"RST-DT",
"66.2 (-10.3)",
"50.8 (-14.4)",
"GUM",
"）中保留在每个数据集的test集合上的分数",
"GUM news",
"67.9 (-8.6)",
"55.8 (-9.4)",
"-",
"46.1",
"（由于RST-DT没有deV分区的XEDX",
"GUM",
"GUM",
"68.6",
"54.9",
"RST-DT",
"我们实验的一个可能的局限性是语料库之间",
"71.1 (+2.5)",
"55.9 (+1.0)",
"的注释差异，尤其是对于它们的关系清单，它",
"GUMnews",
"73.4",
"63.3",
"57.2",
"们相似但不相同。因此，我们省略了表3和4",
"中的R分数，尽管我们将在下面更详细地返",
"Table 4: Cross-Corpus Results (5 run average) of the",
"回关系。但是，对于S和N，我们希望差异很",
"TOP-DOWN Parser from Liu et al. (2021) .",
"小：两个语料库都使用相同的分割准则，并且",
"我们将此结果解释为意味着与EDU分段不",
"使用诸如可删除性和主观突出性之类的测试将",
"不太突出的卫星单元从属于更突出的核的相同",
"同，在EDU分段中，结果在两个方向上都很接",
"准则。尽管EDU分割不是我们的重点，但我",
"近，train和test数据的体裁组成在RST成分",
"们可以通过交叉测试Gessleretal.(2021）中当",
"解析的通用性中起着至关重要的作用，而与解",
"前的SOTAEDU分割器来确认分割准则的兼容",
"析器体系结构无关。因此我们选择BOTTOM",
"性。结果显示，在RST-DT上进行相当适度的",
"UP解析器进行本文其余部分中介绍的实验",
"降解训练，在GUM上进行测试，从94。9至",
"因为它在两个语料库上都优于TOP-DOWN解",
"89。9，而相反的方向显示为92。9至91下降",
"析器",
"（有关完整数字，请参见附录A中的表11）。",
"要明确的是，结果并不表明在GUM上进行",
"我们推测，由于语法的原因，分段可以很好地",
"训练是在RST-DT上实现最佳性能的一种方",
"式：GUM模型的得分仍然比RST-DT训练的",
"泛化，因此任务的稀疏性和特定于域的性质较",
"少，类似于子句边界检测",
"S和N模型低近4分和8分。然而，对于已经",
"对于跨语料库S和N分数，在RST-DT上训",
"看到一些新闻的GUM模型来说，RST-DT新",
"练以解析来自GUM的OOD数据时，两个解析",
"闻数据似乎并不令人惊，总而言之，鉴于",
"GUM模型经过训练可以解决的广泛类型库存",
"器都显示出非常显着的退化，如表3和4所示。",
"例如，使用BOTTOM-UP体系结构，N的E???",
"RST-DT数据似乎是一个相对“容易”的目标",
"会降级。我们也很好奇这是否也适用于GUM",
"这些结果还证实，在多个流派上进行训练，",
"的新闻类型：实际上，仅在GUM新闻上进行",
"每个流派的文档相对较少，可以带来良好的性",
"测试可以使用BOTTOM-UP解析器将降级降低",
"能，而在RST-DT的非常狭窄的WSJ域上只有",
"约50%，但即使该模型仅在新闻上进行了训练",
"很小的降级。虽然我们认为这个结果本身很重",
"（尽管来自更老的《华尔街日报》新闻数据）。相",
"要，并建议RST解析工作应该更多地关注多",
"比之下，经过GUM训练的模型在RST-DT上",
"类型语料库作为基准，但它留下了一个问题：",
"的得分实际上要比在GUM上的得分高，但略",
"我们可以结合来自两个语料库的数据来提高",
"有提高2。8在S和+O上。使用BOTTOM-UP解",
"RST-DT上的英语RST解析性能吗？",
"析器在N上设置4。在TOP-DOWN解析器性能",
"3.2",
" Joint Training",
"中观察到相同的降级模式：总体上严重降级，",
"并且在训练新闻和测试其他类型时更糟。",
"如果RST解析跨域泛化，则有可能通过联合训",
"练看到收益。为此，我们比较了三种方法：朴",
"N",
"R",
"素串联，模型堆叠和预训练，我们在RST-DT",
"test",
"train",
"s",
"54.8",
"基准上进行了评估",
"76.5",
"RST-DTRST-DT",
"65.9",
"GUM",
"65.3 (-11.2)",
"49.5 (-16.4)",
"对于连接（CONCAT），我们必须映射跨语料",
"GUMnews",
"71.0 (-5.5)",
"57.5 (-8.4)",
"-",
"库的关系。与以前的工作一样，我们的目标是粗",
"GUM",
"GUM",
"69.9",
"57.0",
"48.5",
"略的关系类，其中大多数在两个语料库中都是",
"RST-DT",
"72.7 (+2.8)",
"57.4 (+0.4)",
"GUM news",
"49.5",
"相同的。例外情况包括GUM的PHATIC关系用",
"71.6",
"58.5",
"于不流畅和反向通道（由于会话数据），GUM的",
"Table 3: Cross-Corpus Results (5 run average) of the",
"PREPARATION以及RST-DT的TEMPORAL和",
"BOTTOM-UP ParserfromGuz and Carenini (2020)",
"COMPARISON类之间的映射，这些映射不同地",
"映射到GUM的CIRCUMSTANCE和 SEQUENCE",
"6we further analyzed left vs. right-branching trees dur-",
"或GUM的ANTITHESIS，具体取决于。虽然",
"ing testing. In both corpora, NS outperformed SN: RST-DT",
"受这些不匹配影响的关系实例数量不多，但它",
"in-domain scores are F1=82.6 % vs. 73.4 % for NS vs. SN;",
"GUM V8 has 81.6 % vs. 72.4 % , meaning SN transitions are",
"不是可以忽略的（13。3%），我们预计这种方法",
"easier",
"不会超过域内培训，主要是报告其完整性。我"
] |
{
"arxiv_id": "2302.06488",
"language": "zh",
"page_idx": 4
}
|
||
1711.04987-zh-3
|
1711.04987-zh-3.jpg
|
生描述：
$$L_1(d_{1:K})=\underset{a_{1:T}}{\mathrm{argmax}}P_{S_0}(d_{1:K}|a_{1:T},y_{1:T})$$
这些优化问题对于一般的基本侦听器和扬声器代理(包括我们使用的序列到序列模型)是难以解决的，因为它们涉及选择输入(从可能序列的组合大空间中)以最大化固定输出序列的概率。相反，我们遵循一个简单的近似推理过程，详见 4.2 节
我们还考虑合并用于生成候选人的基本模型的分数。对于说话者的情况，我们定义一个组合的有理说话者，表示为$S_0\cdot S_1$,它选择在基本听众和基本说话者下使概率的加权乘积最大化的候选者：
$$\begin{aligned}&\underset{d_{1:K}}{\mathrm{argmax}}P_{L_{0}}(a_{1:T}|d_{1:K},y_{1:T})^{\lambda}\\&[-0.5em]\times\:P_{S_{0}}(d_{1:K}|a_{1:T},y_{1:T})^{1-\lambda}\end{aligned}$$
(3)
用于固定插值超参数$\lambda\in[0,1]$ 。这种与基本说话者得分的结合有几个动机。首先，正如Monroe et al. (2017)所主张的那样，我们期望人类言语行为中存在不同程度的基础和合理解释。其次，我们希望模型产生的描述是对动作的流畅描述。由于基本模型是有区别地训练的，因此最大化了固定输入序列的输出序列的概率，因此对于与训练集中看到的输入不同的输入配对的固定输出，它们的评分行为可能校准得很差(例如，当在非语法描述上进行调节时)。合并用于生成候选人的基本模型的分数旨在防止这种行为
要定义理性侦听器，我们使用对称公式：首先，从$L_0$中绘制候选动作序列。对于$L_1$,选择在$S_0$下实现最高概率的动作；对于组合模型$L_0\cdot L_1$ ,选择$S_0$和$L_0$加权组合最高的动作 (平行方程 3 )
# 4.2 Inference
与过去的工作 (Smith et al., 2013; Andreas and Klein, 2016; Monroe et al., 2017)一样，我们对方程式$\overline{\mathrm{E}}1,2$ 和 3 中的优化问题进行了近似：使用基本模型生成候选者，然后将其重新分配以找到可能产生所需行为的候选者
在有理扬声器$S_1$的情况下，我们使用基本扬声器$S_0$使用波束搜索为序列$a_1:T,y_{1:T}$生成一组$n$候选描述$w_1:K_1^{(1)}\ldots w_{1:K_n}^{(n)}$。然后，我们在$P_{L_0}$下找到每个描述的得分 "(将其用作我们希望理性说话者描述的观察到的输出动作的输入序列),或者$P_{L_0}$和原始候选得分$P_{S_0}$的加权组合，并选择得分最大的描述$w_{1:K_j}^{(j)}$,分别近似求解方程$\mathbb{E}$1 或 3 中的最大值。我们为
理性侦听器执行一个对称过程：从基本侦听器生成动作序列候选项，并使用基本扬声器将其重新分发。2
由于理性说话者必须产生长的输出序列(带有多个句子),因此我们在推理中交织说话者和听众，按顺序确定每个输出句子。从当前动作子序列的基本说话者的候选方向句子列表中，我们选择听众模型下得分最高的方向(这也可能取决于先前输出的方向),然后继续下一个动作子序列。3
## 5 Base model details
给定此框架，剩下的就是描述基本模型$L_{0}$和$S_0$ 。我们将这些实现为序列到序列模型，将方向映射到动作(对于听众)或将动作映射到方向 (对于说话者),另外还在每个时间步长上调节世界状态
5.1 Base listenen
我们的基本侦听器模型$L_0$预测以方向和当前世界状态的编码表示为条件的动作序列。在SAIL 域中，这是 Mei et al. (2016)的模型(在Figure 2 b 中以绿色显示单个句子及其相关操作),请参阅下面的“域细节”
Encoder 每个方向句子都使用双向 LSTM (Hochreiter and Schmidhuber, 1997) 单独编码；LSTM 的隐藏状态为每个句子重置。我们通过将单词的嵌入与其正向和反向 LSTM 输出连接起来，获得当前句子中第$k$个单词的表示$h_k^e$
Decoder 我们使用 LSTM 解码器递增地生成动作，该解码器在方向句子和动作子序列之间进行单调对齐；在每个时间步长，解码器预测当前句子$w_{1:M}$的下一个动作(包括选择转移到下一个句子)。解码器在时间步$t$处将当前世界状态$y_t$和当前句子的表示$z_t$作为输入，
$^{2}$We use ensembles of models for the base listener and speaker ( subsection 5.3),and to obtain candidates that are high-scoring under the combination of models in the ensemble, we perform standard beam search using all models in lock-step. At every timestep of the beam search, each possible extension of an output sequence is scored using the product of the extension's conditional probabilities across all models in the ensemble.
$^3$We also experimented with sampling from the base models to produce these candidate lists, as was done in previous work (Andreas and Klein, 2016; Monroe et al., 2017) . In early experiments, however, we found better performance with beam search in the rational ${\mathrm{models~for~all~tasks.}}$
|
[
[
0.12176513671875,
0.07781982421875,
0.181396484375,
0.07781982421875,
0.181396484375,
0.08905029296875,
0.12176513671875,
0.08905029296875
],
[
0.5146484375,
0.07598876953125,
0.880859375,
0.07598876953125,
0.880859375,
0.08905029296875,
0.5146484375,
0.08905029296875
],
[
0.5146484375,
0.0926513671875,
0.8818359375,
0.0926513671875,
0.8818359375,
0.105712890625,
0.5146484375,
0.105712890625
],
[
0.1376953125,
0.105712890625,
0.48779296875,
0.105712890625,
0.48779296875,
0.12054443359375,
0.1376953125,
0.12054443359375
],
[
0.51318359375,
0.10986328125,
0.6171875,
0.10687255859375,
0.6171875,
0.1217041015625,
0.51318359375,
0.12469482421875
],
[
0.2476806640625,
0.119384765625,
0.27197265625,
0.1217041015625,
0.270263671875,
0.12890625,
0.246826171875,
0.1270751953125
],
[
0.5341796875,
0.1270751953125,
0.8818359375,
0.1270751953125,
0.8818359375,
0.14013671875,
0.5341796875,
0.14013671875
],
[
0.1376953125,
0.142578125,
0.489501953125,
0.142578125,
0.489501953125,
0.1561279296875,
0.1376953125,
0.1561279296875
],
[
0.51318359375,
0.1436767578125,
0.8818359375,
0.1436767578125,
0.8818359375,
0.15673828125,
0.51318359375,
0.15673828125
],
[
0.12005615234375,
0.1591796875,
0.489501953125,
0.1591796875,
0.489501953125,
0.1728515625,
0.12005615234375,
0.1728515625
],
[
0.5146484375,
0.1602783203125,
0.8818359375,
0.1602783203125,
0.8818359375,
0.17333984375,
0.5146484375,
0.17333984375
],
[
0.12005615234375,
0.174560546875,
0.489501953125,
0.17333984375,
0.489501953125,
0.18701171875,
0.12005615234375,
0.188232421875
],
[
0.515625,
0.1776123046875,
0.880859375,
0.1776123046875,
0.880859375,
0.188232421875,
0.515625,
0.188232421875
],
[
0.12176513671875,
0.191162109375,
0.48779296875,
0.191162109375,
0.48779296875,
0.204833984375,
0.12176513671875,
0.204833984375
],
[
0.515625,
0.1923828125,
0.8818359375,
0.1923828125,
0.8818359375,
0.2060546875,
0.515625,
0.2060546875
],
[
0.118408203125,
0.207275390625,
0.48779296875,
0.207275390625,
0.48779296875,
0.2203369140625,
0.118408203125,
0.2203369140625
],
[
0.5146484375,
0.2078857421875,
0.8818359375,
0.2078857421875,
0.8818359375,
0.221435546875,
0.5146484375,
0.221435546875
],
[
0.12176513671875,
0.2244873046875,
0.26611328125,
0.2244873046875,
0.26611328125,
0.235107421875,
0.12176513671875,
0.235107421875
],
[
0.5205078125,
0.22265625,
0.67333984375,
0.221435546875,
0.67333984375,
0.236328125,
0.5205078125,
0.2369384765625
],
[
0.1368408203125,
0.2392578125,
0.486083984375,
0.2392578125,
0.486083984375,
0.2529296875,
0.1368408203125,
0.2529296875
],
[
0.118408203125,
0.254150390625,
0.48779296875,
0.254638671875,
0.48779296875,
0.26953125,
0.118408203125,
0.268310546875
],
[
0.515625,
0.260009765625,
0.52880859375,
0.260009765625,
0.52880859375,
0.268310546875,
0.515625,
0.268310546875
],
[
0.5263671875,
0.260009765625,
0.54833984375,
0.260009765625,
0.54833984375,
0.268310546875,
0.5263671875,
0.268310546875
],
[
0.54833984375,
0.25830078125,
0.7138671875,
0.25830078125,
0.7138671875,
0.271484375,
0.54833984375,
0.271484375
],
[
0.12005615234375,
0.271484375,
0.48779296875,
0.271484375,
0.48779296875,
0.28515625,
0.12005615234375,
0.28515625
],
[
0.118408203125,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.118408203125,
0.2998046875
],
[
0.5146484375,
0.2880859375,
0.8818359375,
0.2880859375,
0.8818359375,
0.3017578125,
0.5146484375,
0.3017578125
],
[
0.1234130859375,
0.3046875,
0.219970703125,
0.3046875,
0.219970703125,
0.31640625,
0.1234130859375,
0.31640625
],
[
0.5146484375,
0.302734375,
0.880859375,
0.302734375,
0.880859375,
0.31640625,
0.5146484375,
0.31640625
],
[
0.51904296875,
0.32177734375,
0.87890625,
0.32177734375,
0.87890625,
0.33203125,
0.51904296875,
0.33203125
],
[
0.160400390625,
0.33203125,
0.38623046875,
0.328857421875,
0.38623046875,
0.3466796875,
0.160400390625,
0.349853515625
],
[
0.515625,
0.3349609375,
0.880859375,
0.3349609375,
0.880859375,
0.3486328125,
0.515625,
0.3486328125
],
[
0.177978515625,
0.34326171875,
0.2091064453125,
0.3466796875,
0.20654296875,
0.358154296875,
0.175537109375,
0.355712890625
],
[
0.51708984375,
0.352783203125,
0.62451171875,
0.352783203125,
0.62451171875,
0.364013671875,
0.51708984375,
0.364013671875
],
[
0.1435546875,
0.359375,
0.490234375,
0.358154296875,
0.490234375,
0.37890625,
0.1435546875,
0.380126953125
],
[
0.51318359375,
0.385498046875,
0.5634765625,
0.385498046875,
0.5634765625,
0.399658203125,
0.51318359375,
0.399658203125
],
[
0.55517578125,
0.385986328125,
0.66162109375,
0.385986328125,
0.66162109375,
0.396728515625,
0.55517578125,
0.396728515625
],
[
0.12176513671875,
0.392578125,
0.48779296875,
0.392578125,
0.48779296875,
0.40625,
0.12176513671875,
0.40625
],
[
0.12005615234375,
0.407958984375,
0.489501953125,
0.407958984375,
0.489501953125,
0.421630859375,
0.12005615234375,
0.421630859375
],
[
0.515625,
0.410888671875,
0.880859375,
0.410888671875,
0.880859375,
0.421630859375,
0.515625,
0.421630859375
],
[
0.12176513671875,
0.42578125,
0.486083984375,
0.42578125,
0.486083984375,
0.435791015625,
0.12176513671875,
0.435791015625
],
[
0.5146484375,
0.427001953125,
0.88330078125,
0.427001953125,
0.88330078125,
0.440673828125,
0.5146484375,
0.440673828125
],
[
0.12005615234375,
0.440673828125,
0.489501953125,
0.440673828125,
0.489501953125,
0.45361328125,
0.12005615234375,
0.45361328125
],
[
0.51318359375,
0.4423828125,
0.8818359375,
0.4423828125,
0.8818359375,
0.4560546875,
0.51318359375,
0.4560546875
],
[
0.12005615234375,
0.4560546875,
0.48779296875,
0.4560546875,
0.48779296875,
0.4697265625,
0.12005615234375,
0.4697265625
],
[
0.5146484375,
0.457763671875,
0.8818359375,
0.457763671875,
0.8818359375,
0.471435546875,
0.5146484375,
0.471435546875
],
[
0.12005615234375,
0.47265625,
0.489501953125,
0.47265625,
0.489501953125,
0.486328125,
0.12005615234375,
0.486328125
],
[
0.5146484375,
0.474365234375,
0.75244140625,
0.474365234375,
0.75244140625,
0.488037109375,
0.5146484375,
0.488037109375
],
[
0.12005615234375,
0.488037109375,
0.48779296875,
0.488037109375,
0.48779296875,
0.501953125,
0.12005615234375,
0.501953125
],
[
0.12005615234375,
0.50439453125,
0.48779296875,
0.50439453125,
0.48779296875,
0.51708984375,
0.12005615234375,
0.51708984375
],
[
0.5205078125,
0.50439453125,
0.8818359375,
0.5048828125,
0.8818359375,
0.51953125,
0.5205078125,
0.5185546875
],
[
0.12176513671875,
0.521484375,
0.48779296875,
0.521484375,
0.48779296875,
0.5322265625,
0.12176513671875,
0.5322265625
],
[
0.5146484375,
0.52099609375,
0.8818359375,
0.521484375,
0.8818359375,
0.5361328125,
0.5146484375,
0.53515625
],
[
0.12005615234375,
0.53759765625,
0.484375,
0.53759765625,
0.484375,
0.54736328125,
0.12005615234375,
0.54736328125
],
[
0.51318359375,
0.5361328125,
0.8818359375,
0.53759765625,
0.8818359375,
0.5517578125,
0.51318359375,
0.55029296875
],
[
0.12005615234375,
0.552734375,
0.489501953125,
0.552734375,
0.489501953125,
0.56640625,
0.12005615234375,
0.56640625
],
[
0.51318359375,
0.552734375,
0.880859375,
0.552734375,
0.880859375,
0.56640625,
0.51318359375,
0.56640625
],
[
0.12005615234375,
0.5693359375,
0.26611328125,
0.5693359375,
0.26611328125,
0.5830078125,
0.12005615234375,
0.5830078125
],
[
0.5146484375,
0.568359375,
0.88330078125,
0.5693359375,
0.88330078125,
0.58447265625,
0.5146484375,
0.5830078125
],
[
0.1376953125,
0.5849609375,
0.48779296875,
0.5849609375,
0.48779296875,
0.5986328125,
0.1376953125,
0.5986328125
],
[
0.51318359375,
0.58447265625,
0.56103515625,
0.5859375,
0.55908203125,
0.6015625,
0.51123046875,
0.599609375
],
[
0.12176513671875,
0.6025390625,
0.486083984375,
0.6025390625,
0.486083984375,
0.61328125,
0.12176513671875,
0.61328125
],
[
0.12005615234375,
0.61767578125,
0.48779296875,
0.61767578125,
0.48779296875,
0.63134765625,
0.12005615234375,
0.63134765625
],
[
0.51904296875,
0.6162109375,
0.880859375,
0.6162109375,
0.880859375,
0.6298828125,
0.51904296875,
0.6298828125
],
[
0.118408203125,
0.6318359375,
0.486083984375,
0.6318359375,
0.486083984375,
0.6455078125,
0.118408203125,
0.6455078125
],
[
0.515625,
0.63427734375,
0.880859375,
0.63427734375,
0.880859375,
0.64794921875,
0.515625,
0.64794921875
],
[
0.12005615234375,
0.6494140625,
0.26025390625,
0.6494140625,
0.26025390625,
0.6630859375,
0.12005615234375,
0.6630859375
],
[
0.515625,
0.6494140625,
0.8818359375,
0.6494140625,
0.8818359375,
0.6630859375,
0.515625,
0.6630859375
],
[
0.515625,
0.66650390625,
0.8818359375,
0.66650390625,
0.8818359375,
0.68115234375,
0.515625,
0.68115234375
],
[
0.118408203125,
0.67333984375,
0.2418212890625,
0.6748046875,
0.2418212890625,
0.689453125,
0.118408203125,
0.6884765625
],
[
0.515625,
0.681640625,
0.8818359375,
0.681640625,
0.8818359375,
0.6953125,
0.515625,
0.6953125
],
[
0.12176513671875,
0.6953125,
0.489501953125,
0.6953125,
0.489501953125,
0.708984375,
0.12176513671875,
0.708984375
],
[
0.515625,
0.69873046875,
0.88330078125,
0.69873046875,
0.88330078125,
0.7119140625,
0.515625,
0.7119140625
],
[
0.118408203125,
0.7119140625,
0.484375,
0.7119140625,
0.484375,
0.72265625,
0.118408203125,
0.72265625
],
[
0.12005615234375,
0.72802734375,
0.48779296875,
0.72802734375,
0.48779296875,
0.7412109375,
0.12005615234375,
0.7412109375
],
[
0.52880859375,
0.72705078125,
0.88330078125,
0.72998046875,
0.88330078125,
0.74462890625,
0.52880859375,
0.7412109375
],
[
0.12005615234375,
0.7421875,
0.489501953125,
0.74365234375,
0.489501953125,
0.75830078125,
0.12005615234375,
0.75732421875
],
[
0.515625,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.7529296875,
0.515625,
0.7529296875
],
[
0.12005615234375,
0.76025390625,
0.432373046875,
0.76025390625,
0.432373046875,
0.77392578125,
0.12005615234375,
0.77392578125
],
[
0.515625,
0.7548828125,
0.880859375,
0.7548828125,
0.880859375,
0.765625,
0.515625,
0.765625
],
[
0.5146484375,
0.765625,
0.880859375,
0.765625,
0.880859375,
0.775390625,
0.5146484375,
0.775390625
],
[
0.139404296875,
0.77783203125,
0.48779296875,
0.77783203125,
0.48779296875,
0.7880859375,
0.139404296875,
0.7880859375
],
[
0.515625,
0.77685546875,
0.88330078125,
0.77685546875,
0.88330078125,
0.79052734375,
0.515625,
0.79052734375
],
[
0.118408203125,
0.79150390625,
0.490234375,
0.79150390625,
0.490234375,
0.80859375,
0.118408203125,
0.80859375
],
[
0.51318359375,
0.78759765625,
0.880859375,
0.7890625,
0.880859375,
0.802734375,
0.51318359375,
0.80029296875
],
[
0.5146484375,
0.80029296875,
0.8818359375,
0.80029296875,
0.8818359375,
0.81396484375,
0.5146484375,
0.81396484375
],
[
0.12005615234375,
0.80810546875,
0.490234375,
0.80810546875,
0.490234375,
0.826171875,
0.12005615234375,
0.826171875
],
[
0.515625,
0.81396484375,
0.880859375,
0.81396484375,
0.880859375,
0.82470703125,
0.515625,
0.82470703125
],
[
0.12005615234375,
0.82666015625,
0.489501953125,
0.82666015625,
0.489501953125,
0.84033203125,
0.12005615234375,
0.84033203125
],
[
0.5146484375,
0.826171875,
0.5615234375,
0.826171875,
0.5615234375,
0.8359375,
0.5146484375,
0.8359375
],
[
0.12005615234375,
0.8427734375,
0.48779296875,
0.8427734375,
0.48779296875,
0.85546875,
0.12005615234375,
0.85546875
],
[
0.5341796875,
0.83837890625,
0.880859375,
0.83837890625,
0.880859375,
0.8486328125,
0.5341796875,
0.8486328125
],
[
0.515625,
0.85107421875,
0.880859375,
0.85107421875,
0.880859375,
0.86083984375,
0.515625,
0.86083984375
],
[
0.12005615234375,
0.85791015625,
0.489501953125,
0.85791015625,
0.489501953125,
0.87158203125,
0.12005615234375,
0.87158203125
],
[
0.51318359375,
0.85986328125,
0.880859375,
0.85986328125,
0.880859375,
0.87353515625,
0.51318359375,
0.87353515625
],
[
0.11419677734375,
0.8662109375,
0.491943359375,
0.87158203125,
0.491943359375,
0.896484375,
0.11419677734375,
0.89111328125
],
[
0.5146484375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.5146484375,
0.8857421875
],
[
0.51318359375,
0.88427734375,
0.8818359375,
0.88427734375,
0.8818359375,
0.89794921875,
0.51318359375,
0.89794921875
],
[
0.12176513671875,
0.8955078125,
0.48779296875,
0.8955078125,
0.48779296875,
0.9091796875,
0.12176513671875,
0.9091796875
],
[
0.5146484375,
0.89794921875,
0.6484375,
0.89794921875,
0.6484375,
0.90771484375,
0.5146484375,
0.90771484375
]
] |
[
"生描述：",
"理性侦听器执行一个对称过程：从基本侦听器",
"生成动作序列候选项，并使用基本扬声器将其",
"Li(di:K) = argmax Pso(di:K|ai:T, yi:T) (2)",
"重新分发。2",
"a1:T",
"由于理性说话者必须产生长的输出序列（带",
"这些优化问题对于一般的基本侦听器和扬声",
"有多个句子），因此我们在推理中交织说话者",
"器代理（包括我们使用的序列到序列模型）是",
"和听众，按顺序确定每个输出句子。从当前动",
"难以解决的，因为它们涉及选择输入（从可能",
"作子序列的基本说话者的候选方向句子列表",
"序列的组合大空间中）以最大化固定输出序列",
"中，我们选择听众模型下得分最高的方向（这",
"的概率。相反，我们遵循一个简单的近似推理",
"也可能取决于先前输出的方向），然后继续下",
"过程，详见4.2节",
"一个动作子序列。3",
"我们还考虑合并用于生成候选人的基本模型",
"的分数。对于说话者的情况，我们定义一个组",
"5",
"",
"Base model details",
"合的有理说话者，表示为So·S1，它选择在基",
"本听众和基本说话者下使概率的加权乘积最大",
"给定此框架，剩下的就是描述基本模型Lo和",
"化的候选者：",
"So。我们将这些实现为序列到序列模型，将方",
"向映射到动作（对于听众）或将动作映射到方",
"argmaxPL (a1:Tldi:K,y1:T)",
"向（对于说话者），另外还在每个时间步长上",
"di:K",
"调节世界状态",
"[-0.5em) × Ps (d1:K|a1:T,Y1:T)1-X (3)",
"5.1",
"Baselistenel",
"用于固定插值超参数入E[0,1]。这种与基本",
"说话者得分的结合有几个动机。首先，正如",
"我们的基本侦听器模型L预测以方向和当前",
"Monroeetal.（2017）所主张的那样，我们期",
"世界状态的编码表示为条件的动作序列。在",
"望人类言语行为中存在不同程度的基础和合理",
"SAIL域中，这是Meietal.（2016）的模型（在",
"解释。其次，我们希望模型产生的描述是对动",
"Figure2b中以绿色显示单个句子及其相关操",
"作的流畅描述。由于基本模型是有区别地训练",
"作），请参阅下面的“域细节”",
"的，因此最大化了固定输入序列的输出序列的",
"概率，因此对于与训练集中看到的输人不同的",
"Encoder每个方向句子都使用双向LSTM",
"输入配对的固定输出，它们的评分行为可能校",
"(Hochreiter and Schmidhuber，1997)单独编",
"准得很差（例如，当在非语法描述上进行调节",
"码；LSTM的隐藏状态为每个句子重置。我们",
"时）。合并用于生成候选人的基本模型的分数",
"通过将单词的嵌入与其正向和反向LSTM输",
"旨在防止这种行为",
"出连接起来，获得当前句子中第个单词的表",
"要定义理性侦听器，我们使用对称公式：首",
"示he",
"先，从Lo中绘制候选动作序列。对于L，选",
"择在So下实现最高概率的动作；对于组合模",
"Decoder我们使用LSTM解码器递增地生",
"型Lo·L1，选择So和Lo加权组合最高的动",
"成动作，该解码器在方向句子和动作子序列之",
"作（平行方程3）",
"间进行单调对齐；在每个时间步长，解码器预",
"测当前句子W1:M的下一个动作（包括选择转",
"4.2Inference",
"移到下一个句子）。解码器在时间步t处将当",
"与过去的工作（Smithetal.,2013;Andreasand",
"前世界状态t和当前句子的表示zt作为输入，",
"Klein.2016:Monroeetal..2017)一样我们",
"对方程式E1，2和3中的优化问题进行了近",
"2We use ensembles of models for the base listener",
"似：使用基本模型生成候选者，然后将其重新",
"and speaker (subsection 5.3),and toobtain candidates",
"分配以找到可能产生所需行为的候选者",
"that are high-scoring under the combination of mod-",
"els in the ensemble.we perform standard beam search",
"在有理扬声器S的情况下，我们使用基本",
"using all models in lock-step. At every timestep of the",
"扬声器So使用波束搜索为序列a1:T，91:T生",
"beam search, each possible extension of an output se-",
"quence is scored using the product of the extension's",
"成一组n候选描述wki...wk.。然后，我",
"conditional probabilities across all models in the en-",
"们在PL。下找到每个描述的得分（将其用作我",
"semble",
"们希望理性说话者描述的观察到的输出动作的",
"3Wealsoexperimentedwithsamplingfromthebase",
"models to produce these candidate lists, as was done",
"输人序列），或者PL。和原始候选得分PSo的",
"in previous work (Andreas and Klein, 2016; Monroe",
"加权组合，并选择得分最大的描述wk，，分",
"et al., 2017) . In early experiments, however, we found",
"better performance with beam search in the rationa",
"别近似求解方程E1或3中的最大值。我们为",
"models for all tasks"
] |
{
"arxiv_id": "1711.04987",
"language": "zh",
"page_idx": 3
}
|
||
1711.04987-zh-4
|
1711.04987-zh-4.jpg
|
更新解码器状态$h^d$,并输出可能动作的分布：
$$h_{t}^{d}=\mathrm{LSTM}_{d}(h_{t-1}^{d},[W_{y}y_{t},z_{t}])\\q_{t}=W_{o}(W_{y}y_{t}+W_{h}h_{t}^{d}+W_{z}z_{t})\\p(a_{t}\mid a_{1:t-1},y_{1:t},w_{1:M})\propto\exp(q_{t})$$
,其中所有权重矩阵$W$都是学习参数。句子表示$z_t$是在当前句子中的单词的表示向量 $h_1^e\ldots h_M^e$ 上使用注意力机制$\overline{\mathrm{E}}($Bahdanau et al., 2015) 生成的：
$$\begin{aligned}\alpha_{t,k}&\propto\exp(v\cdot\tanh(W_{d}h_{t-1}^{d}+W_{e}h_{k}^{e}))\\z_{t}&=\sum_{k=1}^{M}\alpha_{t,k}h_{k}^{e}\end{aligned}$$
,其中注意力权重$\alpha_t,k$被归一化为输入中$k$
位置的总和，权重$W$ 和矢量 MAX1
Domain specifics 对于 SAIL ,我们使用 Chen and Mooney (2011) 注释的句子和路线段之间的对齐方式，这些对齐方式也在以前的工作 (Artzi and Zettlemoyer, 2013; Artzi et al., 2014; Mei et al.,2016)中使用。在 Mei et al. (2016)之后，我们为每个句子重置解码器的隐藏状态
在 SCONE 域中，可能的输出空间比 SAIL 大，我们通过以下方式扩展解码器：(i)将每个动作分解为一个动作类型及其参数，(ii)对类型和参数使用单独的注意机制，以及(iii)使用状态相关的动作嵌入。详见补充材料附录 ?? 。构造 SCONE 域，以便每个句子对应一个(非分解)动作；这提供了我们对动作序列的分割
## 5.2 Base speaker
虽然以前的工作 (Daniele et al., 2017) 依赖于更结构化的方法，但我们使用与上述基本相同的序列到序列机制来构建基本扬声器模型$S_{0}$ 。$S_{0}$(在 Figure 2 b 中以橙色表示)对动作和世界状态的序列进行编码，然后使用解码器输出描述
Encoder 我们使用双向 LSTM 对动作$a_t$和世界状态$y_t$的向量嵌入序列进行编码。与基本侦听器的编码器类似，然后通过将$a_t$和$y_t$ 与该位置的 LSTM 输出连接，获得时间步长$t$ 的表示$h_t^e$
Decoder 与侦听器一样，我们使用 LSTM 解码器，该解码器在方向句子和动作子序列之间具有单调对齐，并且在动作子序列上具有注意力。解码器在位置$k$处将先前生成的单词$w_{k-1}$的嵌入以及动作和世界状态的当前子序列的表示$z_k$作为输入，并在单词上产生分布
(包括结束对当前子序列的描述并前进到下一
个)。解码器的输出分布由以下公式产生：
$$\begin{aligned}h_{k}^{d}&=\mathrm{LSTM}_{d}(h_{k-1}^{d},[w_{k-1},z_{k}])\\q_{k}&=W_{h}h_{k}^{d}+W_{z}z_{k}\\p(w_{k}\mid w_{1:k-1},a_{1:T},y_{1:T})\propto\exp(q_{k})\end{aligned}$$
,其中所有权重矩阵$W\text{ 都是学习参数。}^{4}$与基本侦听器一样，输入表示$z_k$是通过关注对输入序列进行编码的向量$h_1^e\ldots h_T^e$而产生的(在这里，对要描述的动作和世界状态的子序列进行编码):
$$\begin{aligned}&\alpha_{k,t}\propto\exp(v\cdot\mathrm{tanh}(W_{d}h_{k-1}^{d}+W_{e}h_{t}^{e}))\\&z_{k}=\sum_{t=1}^{T}\alpha_{k,t}\:h_{t}^{e}\end{aligned}$$
解码器的 LSTM 状态在每个句子的开头重置Domain specifics 在 SAIL 中，为了与不使用分段路由的 Daniele et al. (2017)的生成系统进行比较，我们训练了一个路由分段器，以便在测试时使用。我们还使用动作序列的折叠表示来表示路线。在 SCONE 域中，我们 (i) 使用与侦听器中使用的相同的上下文相关动作嵌入，并且(ii)不需要注意机制，因为仅使用单个动作来产生序列中的给定句子方向句子。更多细节见附录？?
### 5.3 Training
独立训练基本听者和说话者模型，以最大化训练集中动作-方向对的条件可能性。有关优化、 LSTM 变体和超参数的详细信息，请参见附录？?
我们将集成用于基本侦听器$L_{0}$和基本扬声器 $S_0$,其中每个集成由 10 个从单独的随机参数初始化训练的模型组成。这遵循用于 SAIL 基本侦听器的 Mei et al. (2016) 的实验设置
### 6 Experiments
我们在 SAIL 域和三个 SCONE 域 (Section 2 )中的指令跟随和指令生成任务上评估扬声器和侦听器代理。对于所有领域，我们将理性听众和说话者与基本听众和说话者进行比较，并与每个任务和领域过去的最新结果进行比较。最后，我们从模型组合的角度检查语用推理将语用重新排序过程与大量基本说话者或听众的合奏进行比较
${}^{4}$All parameters are distinct from those used in the base listener; the listener and speaker are trained sep- $\arctan$ely.
|
[
[
0.12176513671875,
0.07659912109375,
0.515625,
0.07659912109375,
0.515625,
0.09027099609375,
0.12176513671875,
0.09027099609375
],
[
0.509765625,
0.07659912109375,
0.884765625,
0.07659912109375,
0.884765625,
0.09027099609375,
0.509765625,
0.09027099609375
],
[
0.51708984375,
0.09320068359375,
0.8447265625,
0.09320068359375,
0.8447265625,
0.10687255859375,
0.51708984375,
0.10687255859375
],
[
0.183837890625,
0.1051025390625,
0.41650390625,
0.1051025390625,
0.41650390625,
0.119384765625,
0.183837890625,
0.119384765625
],
[
0.183837890625,
0.12469482421875,
0.4375,
0.12353515625,
0.4375,
0.141357421875,
0.183837890625,
0.142578125
],
[
0.57666015625,
0.12054443359375,
0.8193359375,
0.122314453125,
0.8193359375,
0.1395263671875,
0.57666015625,
0.13720703125
],
[
0.166259765625,
0.1455078125,
0.43408203125,
0.144287109375,
0.43408203125,
0.162109375,
0.166259765625,
0.163330078125
],
[
0.58251953125,
0.1455078125,
0.73486328125,
0.1455078125,
0.73486328125,
0.1602783203125,
0.58251953125,
0.1602783203125
],
[
0.56103515625,
0.1650390625,
0.83642578125,
0.1650390625,
0.83642578125,
0.179931640625,
0.56103515625,
0.179931640625
],
[
0.1376953125,
0.174560546875,
0.486083984375,
0.174560546875,
0.486083984375,
0.188232421875,
0.1376953125,
0.188232421875
],
[
0.1275634765625,
0.18115234375,
0.1368408203125,
0.18115234375,
0.1368408203125,
0.18408203125,
0.1275634765625,
0.18408203125
],
[
0.12005615234375,
0.189453125,
0.486083984375,
0.189453125,
0.486083984375,
0.203125,
0.12005615234375,
0.203125
],
[
0.5205078125,
0.198974609375,
0.52880859375,
0.198974609375,
0.52880859375,
0.2037353515625,
0.5205078125,
0.2037353515625
],
[
0.53076171875,
0.19482421875,
0.8818359375,
0.19482421875,
0.8818359375,
0.2078857421875,
0.53076171875,
0.2078857421875
],
[
0.118408203125,
0.207275390625,
0.486083984375,
0.207275390625,
0.486083984375,
0.2203369140625,
0.118408203125,
0.2203369140625
],
[
0.5146484375,
0.208984375,
0.880859375,
0.210205078125,
0.880859375,
0.2244873046875,
0.5146484375,
0.223876953125
],
[
0.12005615234375,
0.22265625,
0.28369140625,
0.22265625,
0.28369140625,
0.236328125,
0.12005615234375,
0.236328125
],
[
0.515625,
0.226806640625,
0.8818359375,
0.226806640625,
0.8818359375,
0.240478515625,
0.515625,
0.240478515625
],
[
0.5146484375,
0.24462890625,
0.87890625,
0.24462890625,
0.87890625,
0.254638671875,
0.5146484375,
0.254638671875
],
[
0.154541015625,
0.251708984375,
0.450927734375,
0.25048828125,
0.450927734375,
0.265380859375,
0.154541015625,
0.2666015625
],
[
0.51708984375,
0.260009765625,
0.62451171875,
0.260009765625,
0.62451171875,
0.271484375,
0.51708984375,
0.271484375
],
[
0.21240234375,
0.26953125,
0.2333984375,
0.26953125,
0.2333984375,
0.283203125,
0.21240234375,
0.283203125
],
[
0.16455078125,
0.28076171875,
0.28955078125,
0.279052734375,
0.28955078125,
0.300537109375,
0.16455078125,
0.302734375
],
[
0.54833984375,
0.2880859375,
0.84619140625,
0.2880859375,
0.84619140625,
0.302734375,
0.54833984375,
0.302734375
],
[
0.204833984375,
0.297607421875,
0.2392578125,
0.297607421875,
0.2392578125,
0.3134765625,
0.204833984375,
0.3134765625
],
[
0.609375,
0.308837890625,
0.623046875,
0.308837890625,
0.623046875,
0.318359375,
0.609375,
0.318359375
],
[
0.5576171875,
0.318359375,
0.68994140625,
0.3154296875,
0.69189453125,
0.33837890625,
0.55908203125,
0.341552734375
],
[
0.1376953125,
0.325927734375,
0.48779296875,
0.325927734375,
0.48779296875,
0.339111328125,
0.1376953125,
0.339111328125
],
[
0.60546875,
0.334228515625,
0.63232421875,
0.33740234375,
0.62744140625,
0.3515625,
0.60107421875,
0.3486328125
],
[
0.12176513671875,
0.341552734375,
0.4130859375,
0.341552734375,
0.4130859375,
0.355224609375,
0.12176513671875,
0.355224609375
],
[
0.124267578125,
0.366455078125,
0.489501953125,
0.366455078125,
0.489501953125,
0.380126953125,
0.124267578125,
0.380126953125
],
[
0.5146484375,
0.363525390625,
0.87158203125,
0.363525390625,
0.87158203125,
0.377197265625,
0.5146484375,
0.377197265625
],
[
0.12005615234375,
0.3818359375,
0.486083984375,
0.3818359375,
0.486083984375,
0.3955078125,
0.12005615234375,
0.3955078125
],
[
0.5224609375,
0.39013671875,
0.8818359375,
0.39013671875,
0.8818359375,
0.405029296875,
0.5224609375,
0.405029296875
],
[
0.118408203125,
0.39794921875,
0.48779296875,
0.396728515625,
0.48779296875,
0.410400390625,
0.118408203125,
0.410888671875
],
[
0.118408203125,
0.412109375,
0.489501953125,
0.413330078125,
0.489501953125,
0.427490234375,
0.118408203125,
0.427001953125
],
[
0.515625,
0.40673828125,
0.8818359375,
0.40673828125,
0.8818359375,
0.42041015625,
0.515625,
0.42041015625
],
[
0.12005615234375,
0.429931640625,
0.48779296875,
0.429931640625,
0.48779296875,
0.443603515625,
0.12005615234375,
0.443603515625
],
[
0.515625,
0.424560546875,
0.8818359375,
0.424560546875,
0.8818359375,
0.435302734375,
0.515625,
0.435302734375
],
[
0.515625,
0.439453125,
0.8818359375,
0.439453125,
0.8818359375,
0.453125,
0.515625,
0.453125
],
[
0.12176513671875,
0.4453125,
0.48779296875,
0.4453125,
0.48779296875,
0.458984375,
0.12176513671875,
0.458984375
],
[
0.515625,
0.454833984375,
0.88330078125,
0.454833984375,
0.88330078125,
0.468505859375,
0.515625,
0.468505859375
],
[
0.12176513671875,
0.463134765625,
0.173828125,
0.463134765625,
0.173828125,
0.474365234375,
0.12176513671875,
0.474365234375
],
[
0.5146484375,
0.47021484375,
0.880859375,
0.47021484375,
0.880859375,
0.48388671875,
0.5146484375,
0.48388671875
],
[
0.1368408203125,
0.476806640625,
0.48779296875,
0.47802734375,
0.48779296875,
0.4921875,
0.1368408203125,
0.4912109375
],
[
0.515625,
0.48681640625,
0.8818359375,
0.48681640625,
0.8818359375,
0.50048828125,
0.515625,
0.50048828125
],
[
0.12005615234375,
0.49462890625,
0.486083984375,
0.49462890625,
0.486083984375,
0.50830078125,
0.12005615234375,
0.50830078125
],
[
0.5146484375,
0.5029296875,
0.88330078125,
0.5029296875,
0.88330078125,
0.5166015625,
0.5146484375,
0.5166015625
],
[
0.12176513671875,
0.5126953125,
0.486083984375,
0.5126953125,
0.486083984375,
0.5224609375,
0.12176513671875,
0.5224609375
],
[
0.515625,
0.51953125,
0.67529296875,
0.51953125,
0.67529296875,
0.533203125,
0.515625,
0.533203125
],
[
0.12005615234375,
0.52685546875,
0.486083984375,
0.52685546875,
0.486083984375,
0.54052734375,
0.12005615234375,
0.54052734375
],
[
0.12005615234375,
0.54345703125,
0.490234375,
0.54345703125,
0.490234375,
0.55712890625,
0.12005615234375,
0.55712890625
],
[
0.51123046875,
0.5419921875,
0.6337890625,
0.5458984375,
0.6337890625,
0.5634765625,
0.51123046875,
0.56005859375
],
[
0.12176513671875,
0.55859375,
0.489501953125,
0.55859375,
0.489501953125,
0.572265625,
0.12176513671875,
0.572265625
],
[
0.5146484375,
0.568359375,
0.87890625,
0.568359375,
0.87890625,
0.5791015625,
0.5146484375,
0.5791015625
],
[
0.12005615234375,
0.57470703125,
0.489501953125,
0.57470703125,
0.489501953125,
0.58837890625,
0.12005615234375,
0.58837890625
],
[
0.51318359375,
0.5830078125,
0.8818359375,
0.58447265625,
0.8818359375,
0.5986328125,
0.51318359375,
0.59716796875
],
[
0.118408203125,
0.60107421875,
0.27197265625,
0.60107421875,
0.27197265625,
0.6142578125,
0.118408203125,
0.6142578125
],
[
0.5146484375,
0.60107421875,
0.8818359375,
0.60107421875,
0.8818359375,
0.6142578125,
0.5146484375,
0.6142578125
],
[
0.12176513671875,
0.62158203125,
0.48779296875,
0.62158203125,
0.48779296875,
0.63525390625,
0.12176513671875,
0.63525390625
],
[
0.51318359375,
0.61767578125,
0.5634765625,
0.6142578125,
0.56494140625,
0.626953125,
0.5146484375,
0.6298828125
],
[
0.12176513671875,
0.6357421875,
0.48779296875,
0.6357421875,
0.48779296875,
0.6494140625,
0.12176513671875,
0.6494140625
],
[
0.5341796875,
0.63427734375,
0.8818359375,
0.63427734375,
0.8818359375,
0.64453125,
0.5341796875,
0.64453125
],
[
0.1234130859375,
0.65380859375,
0.489501953125,
0.65380859375,
0.489501953125,
0.66455078125,
0.1234130859375,
0.66455078125
],
[
0.515625,
0.65087890625,
0.880859375,
0.65087890625,
0.880859375,
0.6611328125,
0.515625,
0.6611328125
],
[
0.118408203125,
0.66845703125,
0.48779296875,
0.66845703125,
0.48779296875,
0.681640625,
0.118408203125,
0.681640625
],
[
0.5146484375,
0.6650390625,
0.8818359375,
0.6650390625,
0.8818359375,
0.6787109375,
0.5146484375,
0.6787109375
],
[
0.12005615234375,
0.68603515625,
0.48779296875,
0.68603515625,
0.48779296875,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.51318359375,
0.68115234375,
0.86376953125,
0.68115234375,
0.86376953125,
0.69482421875,
0.51318359375,
0.69482421875
],
[
0.118408203125,
0.70166015625,
0.177978515625,
0.70166015625,
0.177978515625,
0.71630859375,
0.118408203125,
0.71630859375
],
[
0.51318359375,
0.71142578125,
0.5498046875,
0.71142578125,
0.5498046875,
0.72265625,
0.51318359375,
0.72265625
],
[
0.54248046875,
0.71044921875,
0.66162109375,
0.71044921875,
0.66162109375,
0.7236328125,
0.54248046875,
0.7236328125
],
[
0.1234130859375,
0.724609375,
0.48779296875,
0.7255859375,
0.48779296875,
0.74072265625,
0.1234130859375,
0.7392578125
],
[
0.5146484375,
0.7353515625,
0.88330078125,
0.7353515625,
0.88330078125,
0.7490234375,
0.5146484375,
0.7490234375
],
[
0.12005615234375,
0.7421875,
0.48779296875,
0.7421875,
0.48779296875,
0.755859375,
0.12005615234375,
0.755859375
],
[
0.118408203125,
0.75732421875,
0.48779296875,
0.7587890625,
0.48779296875,
0.77392578125,
0.118408203125,
0.771484375
],
[
0.515625,
0.751953125,
0.8818359375,
0.751953125,
0.8818359375,
0.765625,
0.515625,
0.765625
],
[
0.5146484375,
0.76708984375,
0.8818359375,
0.76708984375,
0.8818359375,
0.78076171875,
0.5146484375,
0.78076171875
],
[
0.12005615234375,
0.77490234375,
0.489501953125,
0.77490234375,
0.489501953125,
0.7880859375,
0.12005615234375,
0.7880859375
],
[
0.1175537109375,
0.7890625,
0.20068359375,
0.79150390625,
0.20068359375,
0.8056640625,
0.1175537109375,
0.80419921875
],
[
0.515625,
0.78369140625,
0.8818359375,
0.78369140625,
0.8818359375,
0.79736328125,
0.515625,
0.79736328125
],
[
0.515625,
0.798828125,
0.8818359375,
0.80029296875,
0.8818359375,
0.81396484375,
0.515625,
0.8125
],
[
0.124267578125,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.124267578125,
0.82763671875
],
[
0.5146484375,
0.81640625,
0.8818359375,
0.81640625,
0.8818359375,
0.830078125,
0.5146484375,
0.830078125
],
[
0.12176513671875,
0.8330078125,
0.48779296875,
0.8330078125,
0.48779296875,
0.84326171875,
0.12176513671875,
0.84326171875
],
[
0.51318359375,
0.83203125,
0.880859375,
0.83203125,
0.880859375,
0.845703125,
0.51318359375,
0.845703125
],
[
0.12176513671875,
0.84716796875,
0.48779296875,
0.84716796875,
0.48779296875,
0.86083984375,
0.12176513671875,
0.86083984375
],
[
0.5146484375,
0.8486328125,
0.64404296875,
0.8486328125,
0.64404296875,
0.8623046875,
0.5146484375,
0.8623046875
],
[
0.118408203125,
0.86328125,
0.486083984375,
0.86328125,
0.486083984375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.5322265625,
0.87060546875,
0.8818359375,
0.87158203125,
0.8818359375,
0.8857421875,
0.5322265625,
0.88525390625
],
[
0.12005615234375,
0.8798828125,
0.48779296875,
0.8798828125,
0.48779296875,
0.8935546875,
0.12005615234375,
0.8935546875
],
[
0.51123046875,
0.8828125,
0.8818359375,
0.88427734375,
0.8818359375,
0.8984375,
0.51123046875,
0.89794921875
],
[
0.12005615234375,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51318359375,
0.89794921875,
0.56494140625,
0.89794921875,
0.56494140625,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"更新解码器状态hd，并输出可能动作的分布：",
"（包括结束对当前子序列的描述并前进到下一",
"个。解码器的输出分布由以下公式产生：",
"hd - LSTMa(hd-i, [Wyyt, zt])",
"qt = Wo(Wuyt + Whhd + Wzzt)",
"h = LSTMa(h-1, [wk-1, z])",
"p(at /a1:t-1, Y1:t, W1:M) α exp(qt)",
"qk= Whhe+Wzzk",
"p(Wk [ wi:k-1, a1:T, y1:T) α exp(qk)",
"其中所有权重矩阵W都是学习参数。句",
"",
"子表示zt是在当前句子中的单词的表示向",
"",
"其中所有权重矩阵W都是学习参数。4与",
"量h...hM上使用注意力机制E(Bahdanau",
"基本侦听器一样，输人表示z是通过关注对",
"etal.，2015）生成的：",
"输人序列进行编码的向量h...h而产生的",
"（在这单：对要描述的动作和世界状态的子序",
"Qt,k α exp(u · tanh(Waht-1 + Weh))",
"列进行编码）：",
"M",
"z=Eat,khe",
"Qk,t α exp(u - tanh(Wahg-1 + Weh))",
"k=i",
"T",
"zk = ak,t he",
"其中注意力权重αt,k被归一化为输人中k",
"=",
"位置的总和，权重W和矢量MAX1",
"Domain specifics对于SAIL，我们使用Chen",
"解码器的LSTM状态在每个句子的开头重置",
"andMooney（2011）注释的句子和路线段之",
"Domain specifics在 SAIL 中，为了与不使",
"间的对齐方式，这些对齐方式也在以前的工",
"作(Artzi and Zettlemoyer, 2013; Artzi et al.,",
"用分段路由的Danieleetal.（2017）的生成系",
"2014;Meiet al.,2016)中使用。在Meietal.",
"统进行比较，我们训练了一个路由分段器，以",
"便在测试时使用。我们还使用动作序列的折叠",
"(2016）之后，我们为每个句子重置解码器的隐",
"表示来表示路线。在SCONE域中，我们（i)",
"藏状态",
"使用与侦听器中使用的相同的上下文相关动作",
"在SCONE域中，可能的输出空间比SAIL",
"嵌人，并且（ii）不需要注意机制，因为仅使用",
"大，我们通过以下方式扩展解码器：（i）将每个",
"单个动作来产生序列中的给定句子方向句子。",
"动作分解为一个动作类型及其参数，（i）对类",
"更多细节见附录？？",
"型和参数使用单独的注意机制，以及（i）使用",
"状态相关的动作嵌入。详见补充材料附录？？。",
"5.3Training",
"构造SCONE域，以便每个句子对应一个（非",
"独立训练基本听者和说话者模型，以最大化训",
"分解）动作；这提供了我们对动作序列的分割",
"练集中动作-方向对的条件可能性。有关优化、",
"5.2Base speaker",
"LSTM变体和超参数的详细信息，请参见附",
"虽然以前的工作（Danieleetal.，2017）依赖于",
"录??",
"更结构化的方法，但我们使用与上述基本相同",
"我们将集成用于基本侦听器L0和基本扬声",
"的序列到序列机制来构建基本扬声器模型So。",
"器So：其中每个集成由10个从单独的随机参",
"So（在Figure2b中以橙色表示）对动作和",
"数初始化训练的模型组成。这遵循用于SAIL",
"世界状态的序列进行编码，然后使用解码器输",
"基本侦听器的Meietal.（2016）的实验设置",
"出描述",
"6",
"Experiments",
"Encoder我们使用双向LSTM对动作at和",
"我们在SAIL域和三个SCONE域（Section2",
"世界状态9t的向量嵌入序列进行编码。与基",
"本侦听器的编码器类似，然后通过将at和yt",
"）中的指令跟随和指令生成任务上评估扬声器",
"和侦听器代理。对于所有领域，我们将理性听",
"与该位置的LSTM输出连接，获得时间步长t",
"的表示he",
"众和说话者与基本听众和说话者进行比较，并",
"与每个任务和领域过去的最新结果进行比较。",
"Decoder与侦听器一样，我们使用LSTM解",
"最后，我们从模型组合的角度检查语用推理，",
"码器，该解码器在方向句子和动作子序列之",
"将语用重新排序过程与大量基本说话者或听众",
"间具有单调对齐，并且在动作子序列上具有注",
"的合奏进行比较",
"意力。解码器在位置k处将先前生成的单词",
"4All parameters are distinct from those used in the",
"Wk-1的嵌入以及动作和世界状态的当前子序",
"base listener; the listener and speaker are trained sep-",
"列的表示z作为输入，并在单词上产生分布",
"arately."
] |
{
"arxiv_id": "1711.04987",
"language": "zh",
"page_idx": 4
}
|
||
2312.03367-zh-1
|
2312.03367-zh-1.jpg
|
$\textbf{Lazy- }k:$用于约束信息提取的解码
Arthur Hemmer; Mickaël Coustaty; Nicola Bartolo; Jérôme Brachat, Jean-Marc Ogier$^{\dagger}$
$\mathrm{Shift~Technology,}^{*}$Paris, France
L3iț La Rochelle, France
$\{$ arthur.hemmer,nicola.bartolo $\}$ @shift-technology.com
$\{$ arthur.hemmer,mcoustat,jmogier $\}$ @univ-lr.fr
jerome.brachat@gmail.com
## Abstract
We explore the possibility of improving probabilistic models in structured prediction. Specifically, we combine the models with constrained decoding approaches in the context of token classification for information extraction. The decoding methods search for constraintsatisfying label-assignments while maximizing the total probability. To do this, we evaluate several existing approaches, as well as propose a novel decoding method called Lazy- $k.$ Our findings demonstrate that constrained decoding approaches can significantly improve the models' performances, especially when using smaller models. The Lazy- $k$ approach allows for more flexibility between decoding time and accuracy. The code for using Lazy- $k$ decoding can be found here https: //gi thub. com/ArthurDevNL/lazyk .
## 1 $\textbf{Introduction}$
当今的 Information Extraction (IE) 大部分是使用基于概率的令牌分类模型完成的，例如BERT ( Devlin et al. , 2018) , RoBERTa ( Liu et al. , 2019) , LayoutLM (Xu et al., 2020b,a; Huang et al., 2022) 或 LiLT (Wang et al., 2022)。这些模型旨在通过越来越多地堆叠大量参数来获得最佳结果，这是以增加计算需求和训练复杂性为代价的。通常，仅使用 top-1 预测，尽管模型为所有令牌-标签组合产生概率
理想情况下，探索替代的高似然预测，以改进现有模型的预测。这在结构化预测任务中特别有趣，其中模型的预测被解析为预定义的结果构。这些结构允许定义约束，以评估所产生的预测是否符合预期结构，然后可用于迭代多个高概率预测，直到找到令人满意的解决方案
这种结构的一个具体示例是在发票信息提取的情况下。在此任务中，将为模型提供 Optical
<FigureHere>
Figure 1: Crop of a sample from the CORD dataset along with the highest probability predictions for the amounts. The model incorrectly predicts Total for the Cash amount.
Character Recognition (OCR) 系统的输出，并需要预测文本的哪些部分与发票中的各个元素相对应。例如，在图。1,该模型预计总额、现金和变动金额
但是，“CASH”文本的遮挡会将噪声引入模型的预测中，从而导致其将现金金额错误地标记为另一个总金额。使用发票的算术语义，我们知道总金额应等于支付的现金金额减去更改金额。因此，我们知道模型的最佳预测可能是不正确的。可以探索替代的高概率标签分配， 以找到满足约束的解决方案
工业文档处理系统通常具有程序化的后处理逻辑，用于检测并有时纠正上述语义约束。但是，这些系统很少利用生成的概率分布中的剩余信息“hidden”,并且自定义校正代码通常很复杂且难以维护。此外，还存在 OCR 引起的错误的可能性，这超出了当前工作的范围，但仍然是文档 IE (Nguyen et al., 2022) 中错误的重要来源
简而言之，我们利用特定于任务的结构来探
索替代的高似然预测。具体来说，我们
·propose an efficient algorithm for iterating
over high-likelihood predictions and,
|
[
[
0.313232421875,
0.0955810546875,
0.68408203125,
0.0955810546875,
0.68408203125,
0.10986328125,
0.313232421875,
0.10986328125
],
[
0.1292724609375,
0.15087890625,
0.87890625,
0.15087890625,
0.87890625,
0.1650390625,
0.1292724609375,
0.1650390625
],
[
0.37451171875,
0.16748046875,
0.62890625,
0.16748046875,
0.62890625,
0.18115234375,
0.37451171875,
0.18115234375
],
[
0.397216796875,
0.182861328125,
0.60205078125,
0.18408203125,
0.60205078125,
0.198974609375,
0.397216796875,
0.19775390625
],
[
0.2359619140625,
0.201904296875,
0.779296875,
0.201904296875,
0.779296875,
0.215576171875,
0.2359619140625,
0.215576171875
],
[
0.274658203125,
0.218505859375,
0.736328125,
0.218505859375,
0.736328125,
0.232177734375,
0.274658203125,
0.232177734375
],
[
0.38037109375,
0.235107421875,
0.623046875,
0.235107421875,
0.623046875,
0.248779296875,
0.38037109375,
0.248779296875
],
[
0.2626953125,
0.32177734375,
0.341796875,
0.32177734375,
0.341796875,
0.336181640625,
0.2626953125,
0.336181640625
],
[
0.54248046875,
0.318359375,
0.60205078125,
0.320556640625,
0.60107421875,
0.336181640625,
0.54150390625,
0.334228515625
],
[
0.8076171875,
0.325927734375,
0.85986328125,
0.330078125,
0.8564453125,
0.344482421875,
0.80419921875,
0.34033203125
],
[
0.5380859375,
0.333251953125,
0.58447265625,
0.33740234375,
0.5810546875,
0.352783203125,
0.53564453125,
0.3486328125
],
[
0.145263671875,
0.3486328125,
0.459228515625,
0.349853515625,
0.459228515625,
0.363525390625,
0.145263671875,
0.3623046875
],
[
0.7900390625,
0.3486328125,
0.8037109375,
0.3486328125,
0.8037109375,
0.352783203125,
0.7900390625,
0.352783203125
],
[
0.80419921875,
0.3486328125,
0.8095703125,
0.3486328125,
0.8095703125,
0.35400390625,
0.80419921875,
0.35400390625
],
[
0.8076171875,
0.34326171875,
0.85791015625,
0.345703125,
0.8564453125,
0.361083984375,
0.80615234375,
0.359375
],
[
0.145263671875,
0.364013671875,
0.459228515625,
0.364013671875,
0.459228515625,
0.377685546875,
0.145263671875,
0.377685546875
],
[
0.145263671875,
0.37890625,
0.459228515625,
0.37890625,
0.459228515625,
0.392578125,
0.145263671875,
0.392578125
],
[
0.8134765625,
0.3818359375,
0.82958984375,
0.3818359375,
0.82958984375,
0.392578125,
0.8134765625,
0.392578125
],
[
0.145263671875,
0.391357421875,
0.45751953125,
0.392578125,
0.45751953125,
0.40673828125,
0.145263671875,
0.40625
],
[
0.5458984375,
0.391357421875,
0.609375,
0.3955078125,
0.60693359375,
0.419189453125,
0.54150390625,
0.4150390625
],
[
0.81689453125,
0.391357421875,
0.8271484375,
0.391357421875,
0.8271484375,
0.3955078125,
0.81689453125,
0.3955078125
],
[
0.1485595703125,
0.4091796875,
0.453369140625,
0.4091796875,
0.453369140625,
0.419189453125,
0.1485595703125,
0.419189453125
],
[
0.81689453125,
0.407958984375,
0.84814453125,
0.407958984375,
0.84814453125,
0.42041015625,
0.81689453125,
0.42041015625
],
[
0.145263671875,
0.421630859375,
0.459228515625,
0.421630859375,
0.459228515625,
0.435302734375,
0.145263671875,
0.435302734375
],
[
0.14697265625,
0.435302734375,
0.45751953125,
0.435302734375,
0.45751953125,
0.448974609375,
0.14697265625,
0.448974609375
],
[
0.1485595703125,
0.45068359375,
0.45751953125,
0.45068359375,
0.45751953125,
0.46142578125,
0.1485595703125,
0.46142578125
],
[
0.51318359375,
0.446533203125,
0.87744140625,
0.446533203125,
0.87744140625,
0.457275390625,
0.51318359375,
0.457275390625
],
[
0.1435546875,
0.4619140625,
0.45751953125,
0.463134765625,
0.45751953125,
0.47802734375,
0.1435546875,
0.476806640625
],
[
0.51318359375,
0.458984375,
0.880859375,
0.458984375,
0.880859375,
0.47265625,
0.51318359375,
0.47265625
],
[
0.145263671875,
0.47802734375,
0.45751953125,
0.47802734375,
0.45751953125,
0.4912109375,
0.145263671875,
0.4912109375
],
[
0.51123046875,
0.473876953125,
0.87890625,
0.473876953125,
0.87890625,
0.48681640625,
0.51123046875,
0.48681640625
],
[
0.1485595703125,
0.493408203125,
0.456787109375,
0.493408203125,
0.456787109375,
0.50439453125,
0.1485595703125,
0.50439453125
],
[
0.5146484375,
0.4892578125,
0.60546875,
0.4892578125,
0.60546875,
0.49951171875,
0.5146484375,
0.49951171875
],
[
0.14697265625,
0.50732421875,
0.45751953125,
0.50732421875,
0.45751953125,
0.52099609375,
0.14697265625,
0.52099609375
],
[
0.1485595703125,
0.5224609375,
0.456787109375,
0.5224609375,
0.456787109375,
0.533203125,
0.1485595703125,
0.533203125
],
[
0.14697265625,
0.53515625,
0.45751953125,
0.53515625,
0.45751953125,
0.548828125,
0.14697265625,
0.548828125
],
[
0.51123046875,
0.52978515625,
0.880859375,
0.53076171875,
0.880859375,
0.5458984375,
0.51123046875,
0.54443359375
],
[
0.145263671875,
0.548828125,
0.456787109375,
0.5498046875,
0.456787109375,
0.5634765625,
0.145263671875,
0.5625
],
[
0.5146484375,
0.54638671875,
0.880859375,
0.54638671875,
0.880859375,
0.56005859375,
0.5146484375,
0.56005859375
],
[
0.14697265625,
0.5634765625,
0.45751953125,
0.5634765625,
0.45751953125,
0.57666015625,
0.14697265625,
0.57666015625
],
[
0.5146484375,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.5146484375,
0.57666015625
],
[
0.145263671875,
0.57763671875,
0.459228515625,
0.57763671875,
0.459228515625,
0.59130859375,
0.145263671875,
0.59130859375
],
[
0.5146484375,
0.580078125,
0.623046875,
0.580078125,
0.623046875,
0.59130859375,
0.5146484375,
0.59130859375
],
[
0.145263671875,
0.59130859375,
0.333251953125,
0.59130859375,
0.333251953125,
0.60498046875,
0.145263671875,
0.60498046875
],
[
0.5322265625,
0.595703125,
0.880859375,
0.595703125,
0.880859375,
0.60986328125,
0.5322265625,
0.60986328125
],
[
0.51318359375,
0.61083984375,
0.87890625,
0.61083984375,
0.87890625,
0.62451171875,
0.51318359375,
0.62451171875
],
[
0.1158447265625,
0.62158203125,
0.151123046875,
0.62158203125,
0.151123046875,
0.6328125,
0.1158447265625,
0.6328125
],
[
0.151123046875,
0.62158203125,
0.26123046875,
0.62158203125,
0.26123046875,
0.6328125,
0.151123046875,
0.6328125
],
[
0.5146484375,
0.62890625,
0.880859375,
0.62890625,
0.880859375,
0.642578125,
0.5146484375,
0.642578125
],
[
0.118408203125,
0.6455078125,
0.486083984375,
0.6455078125,
0.486083984375,
0.66015625,
0.118408203125,
0.66015625
],
[
0.51318359375,
0.6435546875,
0.87890625,
0.6435546875,
0.87890625,
0.65673828125,
0.51318359375,
0.65673828125
],
[
0.118408203125,
0.662109375,
0.48779296875,
0.662109375,
0.48779296875,
0.67578125,
0.118408203125,
0.67578125
],
[
0.51318359375,
0.6611328125,
0.880859375,
0.6611328125,
0.880859375,
0.6748046875,
0.51318359375,
0.6748046875
],
[
0.118408203125,
0.6787109375,
0.48779296875,
0.6787109375,
0.48779296875,
0.6923828125,
0.118408203125,
0.6923828125
],
[
0.51123046875,
0.67578125,
0.880859375,
0.67578125,
0.880859375,
0.689453125,
0.51123046875,
0.689453125
],
[
0.1175537109375,
0.693359375,
0.483642578125,
0.693359375,
0.483642578125,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.51318359375,
0.693359375,
0.73486328125,
0.693359375,
0.73486328125,
0.70703125,
0.51318359375,
0.70703125
],
[
0.1175537109375,
0.71044921875,
0.486083984375,
0.708984375,
0.486083984375,
0.7236328125,
0.1175537109375,
0.724609375
],
[
0.5322265625,
0.71044921875,
0.880859375,
0.71044921875,
0.880859375,
0.7236328125,
0.5322265625,
0.7236328125
],
[
0.118408203125,
0.72802734375,
0.48193359375,
0.72802734375,
0.48193359375,
0.73828125,
0.118408203125,
0.73828125
],
[
0.51318359375,
0.72705078125,
0.880859375,
0.72705078125,
0.880859375,
0.74072265625,
0.51318359375,
0.74072265625
],
[
0.118408203125,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.51123046875,
0.7412109375,
0.87890625,
0.7412109375,
0.87890625,
0.7548828125,
0.51123046875,
0.7548828125
],
[
0.118408203125,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51318359375,
0.7587890625,
0.880859375,
0.7587890625,
0.880859375,
0.7724609375,
0.51318359375,
0.7724609375
],
[
0.118408203125,
0.775390625,
0.382080078125,
0.775390625,
0.382080078125,
0.7890625,
0.118408203125,
0.7890625
],
[
0.51318359375,
0.77392578125,
0.87890625,
0.77392578125,
0.87890625,
0.78759765625,
0.51318359375,
0.78759765625
],
[
0.1351318359375,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.1351318359375,
0.8056640625
],
[
0.5146484375,
0.79150390625,
0.87744140625,
0.79150390625,
0.87744140625,
0.8017578125,
0.5146484375,
0.8017578125
],
[
0.118408203125,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.51318359375,
0.80712890625,
0.880859375,
0.80712890625,
0.880859375,
0.82080078125,
0.51318359375,
0.82080078125
],
[
0.12005615234375,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.12005615234375,
0.83837890625
],
[
0.51318359375,
0.82373046875,
0.587890625,
0.82373046875,
0.587890625,
0.8349609375,
0.51318359375,
0.8349609375
],
[
0.118408203125,
0.84033203125,
0.484375,
0.84033203125,
0.484375,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.53076171875,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.53076171875,
0.85498046875
],
[
0.12005615234375,
0.85791015625,
0.483642578125,
0.85791015625,
0.483642578125,
0.8681640625,
0.12005615234375,
0.8681640625
],
[
0.51318359375,
0.85693359375,
0.8271484375,
0.85693359375,
0.8271484375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.468505859375,
0.87255859375,
0.468505859375,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.1351318359375,
0.8896484375,
0.484375,
0.8896484375,
0.484375,
0.90283203125,
0.1351318359375,
0.90283203125
],
[
0.5341796875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90380859375,
0.5341796875,
0.90380859375
],
[
0.1175537109375,
0.90478515625,
0.484375,
0.90478515625,
0.484375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.54833984375,
0.90478515625,
0.82275390625,
0.90478515625,
0.82275390625,
0.91943359375,
0.54833984375,
0.91943359375
]
] |
[
"Lazy-k：用于约束信息提取的解码",
"Arthur Hemmer*, Mickael Coustaty, Nicola Bartolo, Jerome Brachat, Jean-Marc Ogier",
"Shift Technology, Paris, France",
"L3it La Rochelle, France",
" arthur.hemmer,nicola.bartolo } @shift-technology.com",
"( arthur.hemmer,mcoustat,jmogier @univ-lr.fr",
"jerome.brachat@gmail.com",
"Abstract",
"Subtotal",
"54.000",
"Total",
"We explore the possibility of improving proba-",
"",
"",
"54.000",
"bilistic models in structured prediction. Specif-",
"ically, we combine the models with con-",
"",
"strained decoding approaches in the context of",
"Kerta ian",
"",
"tokenclassificationforinformationextraction",
"2.00",
"The decoding methods search for constraint-",
"satisfying label-assignments while maximiz-",
"ingthetotal probability.To do this, we evalu-",
"Figure 1:Cropof a samplefrom the CORD datasei",
"ate several existing approaches, as well as pro.",
"along with the highest probability predictions for the",
"pose a novel decoding method called Lazy- k .",
"amounts. The model incorrectly predicts Total for the",
"Ourfindingsdemonstratethatconstrainedde",
"Cash amount.",
"coding approaches can significantly improve",
"themodels'performances,especiallywhenus",
"ing smaller models. The Lazy- k approach",
"CharacterRecognition（OCR）系统的输出，并需",
"allows for more flexibility between decoding",
"要预测文本的哪些部分与发票中的各个元素相",
"time and accuracy. The code for using Lazy- k",
"对应。例如，在图。1，该模型预计总额、现",
"decoding can be found here https: / /github",
"金和变动金额",
"com/ArthurDevNL/lazyk.",
"但是，“CASH”文本的遮挡会将噪声引人模",
"型的预测中，从而导致其将现金金额错误地标",
"1",
"Introduction",
"记为另一个总金额。使用发票的算术语义，我",
"当今的InformationExtraction（IE）大部分是使",
"们知道总金额应等于支付的现金金额减去更改",
"用基于概率的令牌分类模型完成的，例如",
"金额。因此，我们知道模型的最佳预测可能是",
"BERT (Devlin et al., 2018) , RoBERTa (Liu et al.,",
"不正确的。可以探索替代的高概率标签分配，",
"2019) LayoutLM (Xu et al., 2020b,a; Huang",
"以找到满足约束的解决方案",
"et al.,2022)或LiLT(Wang et al.,2022)。这些",
"工业文档处理系统通常具有程序化的后处理",
"模型旨在通过越来越多地堆叠大量参数来获得",
"逻辑，用于检测并有时纠正上述语义约束。但",
"最佳结果，这是以增加计算需求和训练复杂性",
"是，这些系统很少利用生成的概率分布中的剩",
"为代价的。通常，仅使用top-1预测，尽管模",
"余信息“hidden”，并且自定义校正代码通常很",
"型为所有令牌-标签组合产生概率",
"复杂且难以维护。此外，还存在OCR引起的",
"理想情况下，探索替代的高似然预测，以改",
"错误的可能性，这超出了当前工作的范围，但",
"进现有模型的预测。这在结构化预测任务中特",
"仍然是文档IE（Nguyenetal.,2022）中错误的",
"别有趣，其中模型的预测被解析为预定义的结",
"重要来源",
"构。这些结构允许定义约束，以评估所产生的",
"简而言之，我们利用特定于任务的结构来探",
"预测是否符合预期结构，然后可用于迭代多个",
"索替代的高似然预测。具体来说，我们",
"高概率预测，直到找到令人满意的解决方案",
"这种结构的一个具体示例是在发票信息提取",
"· propose an efficient algorithm for iterating",
"的情况下。在此任务中，将为模型提供Optical",
"over high-likelihood predictions and,"
] |
{
"arxiv_id": "2312.03367",
"language": "zh",
"page_idx": 1
}
|
||
2312.03367-zh-2
|
2312.03367-zh-2.jpg
|
gorithm and,
$\cdot$ perform several experiments to evaluate the relevance of exploring alternative highlikelihood predictions in structured prediction tasks.
### 2 $\textbf{Background}$
要搜索高概率预测，我们需要一个概率模型， 该模型输出给定令牌序列的独立概率。给定一个输入序列 x = $\{x_1,x_2,\ldots,x_n\},x_i\in\mathcal{X}$ ,其中$\mathcal{X}$是令牌词汇，目标是估计输出序列$\mathbf{y}=\{y_1,y_2,\ldots,y_n\},y_i\in\mathcal{Y}$的概率，其中$\mathcal{Y}$是标签词汇。由于此概率很快变得难以处理，因此通常通过将其分解为：
$$p(\mathrm y|\mathrm x)=\prod_i^np(y_i|\mathrm x).$$
(1)
解码过程是指我们从这样的模型中获得 y 的估计$\widehat{\mathrm{y}}$的方式。最简单的方法是将 arg max 作为
(2)
$$\widehat{\mathrm{y}}=\underset{{\mathrm{y}\in\mathcal{Y}^{n}}}{\operatorname*{\arg\max}}p(\mathrm{y}|\mathrm{x}),$$
,分别为每个$y_i$完成
此外，我们引入了一个全局的二进制约束$\mathcal{C}:$x$\times$y$\to \{ 0, 1\}$ ,并将我们感兴趣的问题形式化为
(3)
$$\widehat{\mathrm y}=\arg\max_{\mathrm y\in\mathcal Y^n}p(\mathrm y|\mathrm x)\cdot\mathcal C(\mathrm x,\mathrm y).$$
请注意，对于 Sec 中提出的方法。4 ,我们不对约束做进一步的假设。这很重要，因为许多现有的约束解码方法都需要以线性形式(Faghihi et al., 2023) 表示约束
某些问题可能由线性和非线性约束组成。令牌分类模型通常使用 BIO 标记方案 (Ramshaw and Marcus, 1999) , 其中标签以 B (eginning) , I (nside) 和 O (utside) 为前缀，以便能够对多个令牌的跨度进行分类。在此公式中，I 标签必须始终以 B 或同一类别的另一个 I 标签开头
此标记约束可以使用线性约束表示。然而， 线性约束的解不能保证也是非线性约束的解。语义约束 cash=total+change 不能线性表示，因为为了计算它，必须将与标签化相对应的文本从文本解析为浮点数，这是一种非线性操作。满足线性 (BIO) 约束的最优解不满足非线性(语义)约束的示例如 Tab 所示。1(b)一行 4
· provide a proof for the correctness of the al- 3 Related Work
已经提出了几种用于从等式$\overline{\mathbb{E}}(3)$ 设置的解码方法。GLUECons (Faghihi et al., 2023) 中提供了在约束下学习和解码的出色基准。对于解码，该工作主要探索 Integer Linear Programming (ILP) 的用法，以在给定模型概率的情况下找到满足约束的解决方案
ILP 问题可以使用分支定界算法 (Land and Doig,1960)来解决，该算法是一种知情的搜索算法；它使用约束的线性公式来更有效地引导它通过搜索空间。使用有关其约束的知识的方法的另一个示例是 Viterbi (Forney, 1973) ,它是一种动态编程方法，也可以考虑特定的约束， 尽管比 ILP 更具限制性，因为它仅在马尔可夫环境中工作。这些知情搜索方法的优点是，如果存在，它们将始终在合理的时间内找到最佳答案。然而，它们也有不可忽略的最小运行时间，并对约束提出了上述要求
以前已将信息搜索方法应用于信息提取任务。使用 ILP 在约束下进行解码的灵感来自Roth and Yih 的工作，他们探索了实体和关系提取(Roth and Yih, 2004, 2007) 的应用。他们将解码问题表述为线性规划，目的是在给定一组约束的情况下最大化序列的总体概率
在这些程序中，决策变量是指示变量$\mathbb{1}_i^j$,指示将标签$j$分配给令牌$i$ 。使用它，可以将约束“1 label per token”表示为$\forall_i\sum_{j=1}^l\mathbb{1}_i^j=1$ 其中$l$是可能的标签数。将此线性公式用于其他几个约束，他们在实体和关系分类任务的$F_{1}$ 分数中观察到 2-5 个% 改进。同样，维特比也被用于校正结构化预测 (Douzon et al., 2022)
对于更复杂的约束，我们可以使用不知情的搜索方法，因为它们不会对约束的实现做出任何假设，而只是以贪婪的方式迭代搜索空间。最广为人知的方法是 Beam Search (BS) (Bisiani, 1987) 。虽然它不对约束类型施加任何限制，但它对我们的全局解码设置并不理想， 因为它以“left-to-right”方式工作
为了说明，BS 将参数$k$作为输入，并通过基于先前的顶部$k$波束在每个令牌处计算顶部$k$波束来输出顶部$k$序列。为了评估全局约束， 波束搜索首先需要计算所有 top- $k$序列，之后可以评估约束
不幸的是，这意味着如果约束验证预测最终成为最可能的 (argmax)序列，则波束搜索将计算出$k-1$太多序列。此外，如果约束验证预测不在顶部$k$光束中，则需要运行具有未知的更高$k^\prime$的新搜索，其中还包括重新计算初始的先前$k$预测。在自然语言生成(Anderson et al., 2016; Hokamp and Liu, 2017; Post and Vilar, 2018; Lemons et al., 2022) 的上下
|
[
[
0.1368408203125,
0.08612060546875,
0.6689453125,
0.08489990234375,
0.6689453125,
0.09979248046875,
0.1368408203125,
0.1009521484375
],
[
0.15283203125,
0.10272216796875,
0.2493896484375,
0.1009521484375,
0.2493896484375,
0.11517333984375,
0.15283203125,
0.1175537109375
],
[
0.5146484375,
0.11224365234375,
0.880859375,
0.11224365234375,
0.880859375,
0.1258544921875,
0.5146484375,
0.1258544921875
],
[
0.1376953125,
0.1341552734375,
0.486083984375,
0.1341552734375,
0.486083984375,
0.1478271484375,
0.1376953125,
0.1478271484375
],
[
0.51318359375,
0.12890625,
0.87744140625,
0.12890625,
0.87744140625,
0.142578125,
0.51318359375,
0.142578125
],
[
0.51318359375,
0.144287109375,
0.8818359375,
0.144287109375,
0.8818359375,
0.157958984375,
0.51318359375,
0.157958984375
],
[
0.15283203125,
0.15087890625,
0.490234375,
0.15087890625,
0.490234375,
0.1644287109375,
0.15283203125,
0.1644287109375
],
[
0.51123046875,
0.1602783203125,
0.87890625,
0.1602783203125,
0.87890625,
0.17333984375,
0.51123046875,
0.17333984375
],
[
0.154541015625,
0.168701171875,
0.48779296875,
0.168701171875,
0.48779296875,
0.1787109375,
0.154541015625,
0.1787109375
],
[
0.51318359375,
0.177001953125,
0.880859375,
0.177001953125,
0.880859375,
0.190673828125,
0.51318359375,
0.190673828125
],
[
0.154541015625,
0.182861328125,
0.2301025390625,
0.182861328125,
0.2301025390625,
0.19482421875,
0.154541015625,
0.19482421875
],
[
0.51318359375,
0.193603515625,
0.64208984375,
0.193603515625,
0.64208984375,
0.207275390625,
0.51318359375,
0.207275390625
],
[
0.11419677734375,
0.21435546875,
0.26025390625,
0.21435546875,
0.26025390625,
0.2286376953125,
0.11419677734375,
0.2286376953125
],
[
0.53076171875,
0.210205078125,
0.8818359375,
0.210205078125,
0.8818359375,
0.223876953125,
0.53076171875,
0.223876953125
],
[
0.51123046875,
0.2244873046875,
0.880859375,
0.2257080078125,
0.880859375,
0.240478515625,
0.51123046875,
0.2392578125
],
[
0.118408203125,
0.243408203125,
0.489501953125,
0.243408203125,
0.489501953125,
0.257080078125,
0.118408203125,
0.257080078125
],
[
0.51318359375,
0.2423095703125,
0.880859375,
0.2423095703125,
0.880859375,
0.255859375,
0.51318359375,
0.255859375
],
[
0.12005615234375,
0.260009765625,
0.486083984375,
0.260009765625,
0.486083984375,
0.273681640625,
0.12005615234375,
0.273681640625
],
[
0.51123046875,
0.257080078125,
0.880859375,
0.2587890625,
0.880859375,
0.2724609375,
0.51123046875,
0.270751953125
],
[
0.1292724609375,
0.275634765625,
0.486083984375,
0.275634765625,
0.486083984375,
0.290283203125,
0.1292724609375,
0.290283203125
],
[
0.51318359375,
0.27490234375,
0.880859375,
0.27490234375,
0.880859375,
0.2880859375,
0.51318359375,
0.2880859375
],
[
0.118408203125,
0.290283203125,
0.484375,
0.290283203125,
0.484375,
0.303955078125,
0.118408203125,
0.303955078125
],
[
0.51318359375,
0.290283203125,
0.87744140625,
0.290283203125,
0.87744140625,
0.303955078125,
0.51318359375,
0.303955078125
],
[
0.118408203125,
0.30810546875,
0.486083984375,
0.30810546875,
0.486083984375,
0.322509765625,
0.118408203125,
0.322509765625
],
[
0.51318359375,
0.305908203125,
0.880859375,
0.305908203125,
0.880859375,
0.319580078125,
0.51318359375,
0.319580078125
],
[
0.118408203125,
0.32373046875,
0.48779296875,
0.32373046875,
0.48779296875,
0.33740234375,
0.118408203125,
0.33740234375
],
[
0.509765625,
0.32177734375,
0.87890625,
0.32177734375,
0.87890625,
0.3349609375,
0.509765625,
0.3349609375
],
[
0.1175537109375,
0.339111328125,
0.307373046875,
0.339111328125,
0.307373046875,
0.352783203125,
0.1175537109375,
0.352783203125
],
[
0.5146484375,
0.339111328125,
0.880859375,
0.339111328125,
0.880859375,
0.352783203125,
0.5146484375,
0.352783203125
],
[
0.51318359375,
0.35400390625,
0.8818359375,
0.355224609375,
0.8818359375,
0.369384765625,
0.51318359375,
0.3681640625
],
[
0.301513671875,
0.371826171875,
0.31494140625,
0.374755859375,
0.312255859375,
0.37890625,
0.300537109375,
0.377197265625
],
[
0.51318359375,
0.371826171875,
0.75244140625,
0.371826171875,
0.75244140625,
0.385498046875,
0.51318359375,
0.385498046875
],
[
0.2216796875,
0.38134765625,
0.378662109375,
0.38134765625,
0.378662109375,
0.3984375,
0.2216796875,
0.3984375
],
[
0.46337890625,
0.383056640625,
0.48779296875,
0.383056640625,
0.48779296875,
0.3955078125,
0.46337890625,
0.3955078125
],
[
0.53076171875,
0.388427734375,
0.880859375,
0.388427734375,
0.880859375,
0.402099609375,
0.53076171875,
0.402099609375
],
[
0.30322265625,
0.40087890625,
0.313232421875,
0.40087890625,
0.313232421875,
0.407958984375,
0.30322265625,
0.407958984375
],
[
0.51318359375,
0.40380859375,
0.87744140625,
0.40380859375,
0.87744140625,
0.41748046875,
0.51318359375,
0.41748046875
],
[
0.51318359375,
0.419189453125,
0.880859375,
0.419189453125,
0.880859375,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.1368408203125,
0.427490234375,
0.484375,
0.427490234375,
0.484375,
0.441162109375,
0.1368408203125,
0.441162109375
],
[
0.51123046875,
0.435791015625,
0.880859375,
0.435791015625,
0.880859375,
0.449462890625,
0.51123046875,
0.449462890625
],
[
0.118408203125,
0.44482421875,
0.48779296875,
0.44482421875,
0.48779296875,
0.457763671875,
0.118408203125,
0.457763671875
],
[
0.51318359375,
0.451904296875,
0.86376953125,
0.451904296875,
0.86376953125,
0.465576171875,
0.51318359375,
0.465576171875
],
[
0.1175537109375,
0.46142578125,
0.1376953125,
0.46142578125,
0.1376953125,
0.474365234375,
0.1175537109375,
0.474365234375
],
[
0.51318359375,
0.468505859375,
0.8251953125,
0.468505859375,
0.8251953125,
0.482177734375,
0.51318359375,
0.482177734375
],
[
0.219970703125,
0.478515625,
0.38037109375,
0.479736328125,
0.38037109375,
0.495361328125,
0.219970703125,
0.49462890625
],
[
0.46337890625,
0.48095703125,
0.48779296875,
0.48095703125,
0.48779296875,
0.493408203125,
0.46337890625,
0.493408203125
],
[
0.52880859375,
0.485107421875,
0.87890625,
0.485107421875,
0.87890625,
0.498779296875,
0.52880859375,
0.498779296875
],
[
0.268798828125,
0.49462890625,
0.306396484375,
0.49462890625,
0.306396484375,
0.505859375,
0.268798828125,
0.505859375
],
[
0.509765625,
0.498779296875,
0.8818359375,
0.49951171875,
0.8818359375,
0.51708984375,
0.509765625,
0.5166015625
],
[
0.118408203125,
0.5185546875,
0.28955078125,
0.5185546875,
0.28955078125,
0.5322265625,
0.118408203125,
0.5322265625
],
[
0.5146484375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.5146484375,
0.53076171875
],
[
0.1368408203125,
0.53515625,
0.484375,
0.53515625,
0.484375,
0.548828125,
0.1368408203125,
0.548828125
],
[
0.51123046875,
0.5322265625,
0.87890625,
0.5322265625,
0.87890625,
0.5458984375,
0.51123046875,
0.5458984375
],
[
0.1175537109375,
0.5517578125,
0.486083984375,
0.5517578125,
0.486083984375,
0.56640625,
0.1175537109375,
0.56640625
],
[
0.51318359375,
0.5498046875,
0.87890625,
0.5498046875,
0.87890625,
0.5634765625,
0.51318359375,
0.5634765625
],
[
0.1175537109375,
0.56640625,
0.175537109375,
0.568359375,
0.173828125,
0.5830078125,
0.1158447265625,
0.58056640625
],
[
0.5146484375,
0.5654296875,
0.880859375,
0.5654296875,
0.880859375,
0.5791015625,
0.5146484375,
0.5791015625
],
[
0.51123046875,
0.580078125,
0.85791015625,
0.58056640625,
0.85791015625,
0.595703125,
0.51123046875,
0.59423828125
],
[
0.18896484375,
0.60107421875,
0.4130859375,
0.60107421875,
0.4130859375,
0.615234375,
0.18896484375,
0.615234375
],
[
0.46337890625,
0.6015625,
0.486083984375,
0.6015625,
0.486083984375,
0.6142578125,
0.46337890625,
0.6142578125
],
[
0.5322265625,
0.599609375,
0.87890625,
0.599609375,
0.87890625,
0.60986328125,
0.5322265625,
0.60986328125
],
[
0.235107421875,
0.6142578125,
0.273681640625,
0.6142578125,
0.273681640625,
0.6259765625,
0.235107421875,
0.6259765625
],
[
0.51123046875,
0.61328125,
0.87890625,
0.61328125,
0.87890625,
0.626953125,
0.51123046875,
0.626953125
],
[
0.51318359375,
0.6298828125,
0.880859375,
0.6298828125,
0.880859375,
0.6435546875,
0.51318359375,
0.6435546875
],
[
0.1368408203125,
0.64453125,
0.48779296875,
0.64453125,
0.48779296875,
0.65771484375,
0.1368408203125,
0.65771484375
],
[
0.51123046875,
0.646484375,
0.880859375,
0.646484375,
0.880859375,
0.66015625,
0.51123046875,
0.66015625
],
[
0.1175537109375,
0.66015625,
0.484375,
0.66015625,
0.484375,
0.67333984375,
0.1175537109375,
0.67333984375
],
[
0.509765625,
0.6611328125,
0.880859375,
0.662109375,
0.880859375,
0.6767578125,
0.509765625,
0.67578125
],
[
0.118408203125,
0.6767578125,
0.489501953125,
0.6767578125,
0.489501953125,
0.6904296875,
0.118408203125,
0.6904296875
],
[
0.5146484375,
0.6787109375,
0.8818359375,
0.6787109375,
0.8818359375,
0.6923828125,
0.5146484375,
0.6923828125
],
[
0.118408203125,
0.69482421875,
0.34765625,
0.69482421875,
0.34765625,
0.705078125,
0.118408203125,
0.705078125
],
[
0.5146484375,
0.69482421875,
0.77099609375,
0.69482421875,
0.77099609375,
0.7080078125,
0.5146484375,
0.7080078125
],
[
0.1376953125,
0.71044921875,
0.484375,
0.71044921875,
0.484375,
0.7236328125,
0.1376953125,
0.7236328125
],
[
0.5322265625,
0.71142578125,
0.880859375,
0.71142578125,
0.880859375,
0.724609375,
0.5322265625,
0.724609375
],
[
0.1175537109375,
0.7255859375,
0.484375,
0.7255859375,
0.484375,
0.7392578125,
0.1175537109375,
0.7392578125
],
[
0.51318359375,
0.72802734375,
0.87890625,
0.72802734375,
0.87890625,
0.73828125,
0.51318359375,
0.73828125
],
[
0.1175537109375,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.51123046875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.1175537109375,
0.7587890625,
0.484375,
0.7587890625,
0.484375,
0.7724609375,
0.1175537109375,
0.7724609375
],
[
0.51123046875,
0.7587890625,
0.880859375,
0.7587890625,
0.880859375,
0.7724609375,
0.51123046875,
0.7724609375
],
[
0.1175537109375,
0.77392578125,
0.486083984375,
0.77490234375,
0.486083984375,
0.7880859375,
0.1175537109375,
0.78759765625
],
[
0.5146484375,
0.77685546875,
0.62109375,
0.77685546875,
0.62109375,
0.7880859375,
0.5146484375,
0.7880859375
],
[
0.118408203125,
0.79150390625,
0.484375,
0.79150390625,
0.484375,
0.8046875,
0.118408203125,
0.8046875
],
[
0.53076171875,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.53076171875,
0.8056640625
],
[
0.1376953125,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82080078125,
0.1376953125,
0.82080078125
],
[
0.51318359375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.118408203125,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51123046875,
0.82470703125,
0.87890625,
0.82470703125,
0.87890625,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.486083984375,
0.85693359375,
0.486083984375,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.12005615234375,
0.87451171875,
0.489501953125,
0.87451171875,
0.489501953125,
0.88525390625,
0.12005615234375,
0.88525390625
],
[
0.51318359375,
0.87353515625,
0.8818359375,
0.87353515625,
0.8818359375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.118408203125,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.509765625,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.509765625,
0.90283203125
],
[
0.118408203125,
0.90625,
0.476806640625,
0.90625,
0.476806640625,
0.91943359375,
0.118408203125,
0.91943359375
],
[
0.51123046875,
0.90478515625,
0.87744140625,
0.90478515625,
0.87744140625,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
": provide a proof for the correctness of the al-3Related Work",
"gorithm and,",
"已经提出了几种用于从等式E(3）设置的解码",
":perform several experiments to evaluate",
"方法。GLUECons（Faghihietal.,2023）中提供了",
"在约束下学习和解码的出色基准。对于解码，该",
"the relevance of exploring alternative high-",
"工作主要探索IntegerLinearProgramming（ILP）",
"likelihood predictionsin structured predic-",
"的用法，以在给定模型概率的情况下找到满足",
"tion tasks.",
"约束的解决方案",
"2Background",
"ILP问题可以使用分支定界算法(Land and",
"Doig，1960)来解决，该算法是一种知情的搜索",
"要搜索高概率预测，我们需要一个概率模型，",
"算法；它使用约束的线性公式来更有效地引导",
"该模型输出给定令牌序列的独立概率。给定",
"它通过搜索空间。使用有关其约束的知识的方",
"个输人序列x=a1,a2,...,an),aiEX",
"法的另一个示例是Viterbi（Forney，1973），它",
"，其中是令牌词汇，目标是估计输出序列",
"是一种动态编程方法，也可以考虑特定的约束",
"y=【y1,2，.,yn],yiEV的概率，其中是",
"尽管比ILP更具限制性，因为它仅在马尔可夫",
"标签词汇。由于此概率很快变得难以处理，因",
"环境中工作。这些知情搜索方法的优点是，如",
"此通常通过将其分解为：",
"果存在，它们将始终在合理的时间内找到最佳",
"答案。然而，它们也有不可忽略的最小运行时",
"1",
"间，并对约束提出了上述要求",
"p(y/x) = IIp(yi/x).",
"(1)",
"以前已将信息搜索方法应用于信息提取任",
"3",
"务。使用ILP在约束下进行解码的灵感来自",
"RothandYih的工作，他们探索了实体和关系",
"解码过程是指我们从这样的模型中获得y的",
"提取(RothandYih，2004,2007）的应用。他们",
"估计y的方式。最简单的方法是将argmax作",
"将解码问题表述为线性规划，目的是在给定",
"为",
"组约束的情况下最大化序列的总体概率",
"y = arg max p(y|x),",
"(2)",
"在这些程序中，决策变量是指示变量1，，指",
"yeV",
"示将标签分配给令牌讠。使用它，可以将约",
"，分别为每个9i完成",
"束“1labelpertoken”表示为V,i==1",
"此外，我们引人了一个全局的二进制约束",
"其中1是可能的标签数。将此线性公式用于其",
"C:x×y→[0,1]，并将我们感兴趣的问题形",
"他几个约束，他们在实体和关系分类任务的F",
"式化为",
"分数中观察到2-5个%改进。同样，维特比也",
"被用于校正结构化预测(Douzonetal.,2022)",
"y = argmax p(yx) .C(x,y).",
"(3)",
"对于更复杂的约束，我们可以使用不知情",
"yeyn",
"的搜索方法，因为它们不会对约束的实现做",
"出任何假设，而只是以贪婪的方式迭代搜索",
"请注意，对于Sec中提出的方法。4，我",
"空间。最广为人知的方法是BeamSearch(BS)",
"们不对约束做进一步的假设。这很重要，因为",
"(Bisiani，1987）。虽然它不对约束类型施加任何",
"许多现有的约束解码方法都需要以线性形式",
"限制，但它对我们的全局解码设置并不理想，",
"（Faghihietal..2023）表示约束",
"因为它以“left-to-right”方式工作",
"某些问题可能由线性和非线性约束组成。令",
"为了说明，BS将参数k作为输人，并通过",
"牌分类模型通常使用BIO标记方案（Ramshaw",
"基于先前的顶部k波束在每个今牌处计算项顶部",
"andMarcus，1999），其中标签以B（eginning），",
"k波束来输出顶部k序列。为了评估全局约束，",
"I（nside）和O（utside）为前缀，以便能够对多",
"波束搜索首先需要计算所有top-k序列，之后",
"个令牌的跨度进行分类。在此公式中，I标签",
"可以评估约束",
"必须始终以B或同一类别的另一个I标签开头",
"不幸的是，这意味着如果约束验证预测最",
"此标记约束可以使用线性约束表示。然而，",
"终成为最可能的（argmax）序列，则波束搜",
"线性约束的解不能保证也是非线性约束的解。。",
"索将计算出k1太多序列。此外，如果约",
"语义约束cash=total+change不能线性表示，因",
"束验证预测不在顶部k光束中，则需要运行",
"为为了计算它，必须将与标签化相对应的文本",
"具有未知的更高的新搜索，其中还包括重",
"从文本解析为浮点数，这是一种非线性操作。",
"新计算初始的先前k预测。在自然语言生成",
"满足线性（BIO)约束的最优解不满足非线性",
"(Anderson et al., 2016; Hokamp and Liu, 2017:",
"(语义)约束的示例如Tab所示。1(b)一行4",
"Post and Vilar, 2018; Lemons et al., 2022)的上下"
] |
{
"arxiv_id": "2312.03367",
"language": "zh",
"page_idx": 2
}
|
||
2312.03367-zh-3
|
2312.03367-zh-3.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>Label</th>
<th>$“56”$ $\zeta\zeta$ ,</th>
<th>$\zeta\zeta$ 99</th>
<th>$“000”$</th>
<th> </th>
</tr>
<tr>
<th> </th>
<th>$\overline{B_{total}}$</th>
<th>0.3</th>
<th>-</th>
<th> </th>
<th> </th>
</tr>
<tr>
<th> </th>
<th>$I_{total}$</th>
<th> </th>
<th>0.4</th>
<th>0.4</th>
<th> </th>
</tr>
<tr>
<th> </th>
<th>$B_{cash}$</th>
<th>0.5</th>
<th>-</th>
<th>-</th>
<th> </th>
</tr>
<tr>
<th> </th>
<th>$\underline{I_{cash}}$</th>
<th>-</th>
<th>0.3</th>
<th>0.3</th>
<th> </th>
</tr>
<tr>
<td> </td>
<td> </td>
<td>$a$</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$p$</td>
<td>$“56”$</td>
<td>6699</td>
<td>$“000$</td>
<td>$BIO$</td>
<td>Sem</td>
</tr>
<tr>
<td>8.0 $\%$ 1</td>
<td>$B_{cash}$</td>
<td>$I_{total}$</td>
<td>$I_{total}$</td>
<td>$No$</td>
<td>-</td>
</tr>
<tr>
<td>6.0 $\%$ </td>
<td>$B_{\mathrm{c}}$ $'cash$ 1</td>
<td>$I_{total}$</td>
<td>$I_{cash}$ .</td>
<td>$No$</td>
<td>-</td>
</tr>
<tr>
<td>6.0 $\%$ 1 1</td>
<td>$B.$ $`cash$</td>
<td>cash 1 1</td>
<td>total</td>
<td>$No$</td>
<td>-</td>
</tr>
<tr>
<td>4.8 3 $\%$ </td>
<td>$B_{total}$</td>
<td>$I_{total}$</td>
<td>$I_{total}$ </td>
<td>Yes</td>
<td>$No$</td>
</tr>
<tr>
<td>4.5 $\%$</td>
<td>$B_{,}$ cash</td>
<td>$I_{c}$ cash</td>
<td>$I_{c}$ cash</td>
<td>Yes</td>
<td>Yes</td>
</tr>
<tr>
<td>3.6 $\%$</td>
<td>$B$ total</td>
<td>$I_{total}$</td>
<td>$I.$ cash</td>
<td>$No$</td>
<td>- $\bullet$ L</td>
</tr>
</tbody>
</table>
$(\mathbf{b})$
Table 1: (a) (partial) predicted probabilities for the red bounding box in Fig. 1 , where the model splits the string “56.000” in the three tokens “56”,“:”,and “000”.(b) most likely label assignments in order of probability. BIO= the labels adhere to the BIO constraints, Sem. = the labels adhere to the semantic structure (cash=total+change).
文中已经提出了几种改编，但是没有一种可以解决全局约束的上述问题。其他人建议使用可学习的启发式方法扩展波束搜索，以尝试预测给定的标签分配是否可能违反未来的结构约束(Pan and Srikumar, 2018)
我们的方法遵循与 A$^*$类似的方法，其中部分扩展 (Yoshizumi et al., 2000) 先前已应用于多序列比对问题
据我们所知，我们是第一个将 A$^*$与部分扩展一起应用的，该扩展允许比 ILP 更通用的约束用于全局约束解码设置
# 4 $\textbf{Lazy- }k$ Decoding
顾名思义，Lazy- $k$解码器允许以懒惰的方式解码$k$最可能的序列。这意味着它只迭代必要数量的序列，一旦找到令人满意的解决方案就停止。这个解码器探索的假设是满足约束的序列在其他高概率序列中的某个地方
为了有效地做到这一点，我们利用了这样一个事实，即第$k$个最可能的序列总是在？?? 中的一个$k-1$更可能的序列中。这源于每个标签的独立性，如公式 1 所示。我们将“editdistance”放在引号中，因为我们使用稍微更严格的编辑距离定义，该定义还考虑了各种标签概率之间的顺序。有关此的更多详细信息可以在应用程序中找到。A
它的核心是最佳优先搜索(Russel et al.,
1994) 的变体，其中模型的预测用于确定探索可能的标签分配的顺序。每个状态表示所有$n$令牌 x $=\{x_1,x_2,\ldots,x_n\}$的完整标签分配$\begin{array}{l}\mathrm{y=\{y_1,y_2,\ldots,y_n\}~。状态的成本~}g(\mathrm{y})\text{定义如}\\\text{下：}\end{array}$
(4)
$$g(\mathrm y)=-\sum_{i=1}^n\log p(y_i|\mathrm x).$$
我们使用 y$^{k}$表示$k$最低成本标签分配，并
将起点 y$^{1}$定义为：
(5)
$$\mathrm y^1=\underset{\mathrm y\in\mathcal Y^n}{\arg\min}\:g(\mathrm y)$$
在 Alg 中给出了 Lazy- $k$译码算法。1。它的工作原理是维护 k 个最佳状态的堆，按 1 个编辑距离内的下一个最佳未开发状态的得分进行优先级排序。用起始状态 y$^{1}$初始化堆。在探索状态时，将针对约束进行测试，如果满足， 则直接返回。如果不满足约束，则使用新探索的状态扩展堆，并更新原始状态 y$^{i}$的优先级得分，以反映下一个最佳未探索状态的得分
与最佳优先搜索不同，在探索状态时，我们不会将所有子级添加到堆中。相反，我们只添加下一个最佳状态 y$^k$,并将 y$^i$的优先级键更新为编辑距离 1 内的下一个最佳状态的分数。这大大降低了堆的大小，因为经典的搜索实现在每次迭代时都会添加$n$可能的子节点，而在这种情况下，堆中的状态数最多等于迭代次数。随着所有后续堆操作变得更便宜，这种堆大小的减少反过来转化为更好的运行时复杂性
NextBest 函数将状态 y 和前沿作为输入。边界是一个字典，其中包含所有已探索状态的已探索状态和下一个最佳状态。这些值是跟踪给定状态的第$i$个最佳更改的整数。如果$i==n$ (令牌数),则该函数返回 null,因为在此状态的 1 编辑距离内没有下一个最佳更改。由于下一个最佳状态可能已经存在于边界中，因此NextBest 函数包装在 AddNextBest 中，以确保仅将下一个最佳状态添加到尚未存在的边界。请参阅应用程序。B 用于 NextBest 函数的伪代码。鉴于该算法以概率的降序迭代可能的序列，因此证明它总是会找到存在的最优解是微不足道的。然而，在实践中，状态数量的组合增长很快使穷举搜索变得不可行。为了防止这种情况，使用一个额外的停止条件，如果在固定数量的$k$迭代后没有找到令人满意的解决方案，则迭代停止。还可以根据累积概率质量$p$ 或其他度量来设置停止条件；我们将此探索留给将来的工作
### 4.1 Complexity Analysis
假设$n$标记、$l$标签和请求的 top- $k$序列。Lazy- $k$的空间复杂度为 O $(kn),\bar{\text{因为对于每个}}k$
|
[
[
0.2032470703125,
0.0843505859375,
0.25439453125,
0.0843505859375,
0.25439453125,
0.09857177734375,
0.2032470703125,
0.09857177734375
],
[
0.32080078125,
0.0843505859375,
0.345947265625,
0.0789794921875,
0.35107421875,
0.0926513671875,
0.32568359375,
0.097412109375
],
[
0.364501953125,
0.0831298828125,
0.414794921875,
0.0831298828125,
0.414794921875,
0.097412109375,
0.364501953125,
0.097412109375
],
[
0.5146484375,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.5146484375,
0.09979248046875
],
[
0.20068359375,
0.09857177734375,
0.25439453125,
0.09857177734375,
0.25439453125,
0.1175537109375,
0.20068359375,
0.1175537109375
],
[
0.267822265625,
0.08489990234375,
0.306396484375,
0.08489990234375,
0.306396484375,
0.0968017578125,
0.267822265625,
0.0968017578125
],
[
0.273681640625,
0.09979248046875,
0.30322265625,
0.09979248046875,
0.30322265625,
0.11639404296875,
0.273681640625,
0.11639404296875
],
[
0.330078125,
0.10394287109375,
0.343505859375,
0.10394287109375,
0.343505859375,
0.114013671875,
0.330078125,
0.114013671875
],
[
0.38623046875,
0.105712890625,
0.393798828125,
0.105712890625,
0.393798828125,
0.11224365234375,
0.38623046875,
0.11224365234375
],
[
0.51318359375,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.204833984375,
0.11517333984375,
0.2493896484375,
0.1175537109375,
0.2476806640625,
0.1341552734375,
0.2032470703125,
0.1318359375
],
[
0.322509765625,
0.11639404296875,
0.35107421875,
0.11639404296875,
0.35107421875,
0.1300048828125,
0.322509765625,
0.1300048828125
],
[
0.37451171875,
0.11639404296875,
0.405517578125,
0.11639404296875,
0.405517578125,
0.1312255859375,
0.37451171875,
0.1312255859375
],
[
0.51318359375,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.133056640625,
0.51318359375,
0.133056640625
],
[
0.204833984375,
0.1318359375,
0.252685546875,
0.1341552734375,
0.2509765625,
0.149658203125,
0.2032470703125,
0.1478271484375
],
[
0.273681640625,
0.1312255859375,
0.30322265625,
0.1312255859375,
0.30322265625,
0.1478271484375,
0.273681640625,
0.1478271484375
],
[
0.38623046875,
0.1383056640625,
0.39208984375,
0.1383056640625,
0.39208984375,
0.1436767578125,
0.38623046875,
0.1436767578125
],
[
0.51123046875,
0.1353759765625,
0.880859375,
0.133056640625,
0.880859375,
0.1478271484375,
0.51123046875,
0.149658203125
],
[
0.2032470703125,
0.1484375,
0.2493896484375,
0.1484375,
0.2493896484375,
0.1644287109375,
0.2032470703125,
0.1644287109375
],
[
0.322509765625,
0.1484375,
0.35107421875,
0.1484375,
0.35107421875,
0.162109375,
0.322509765625,
0.162109375
],
[
0.37451171875,
0.1484375,
0.403076171875,
0.1484375,
0.403076171875,
0.162109375,
0.37451171875,
0.162109375
],
[
0.5146484375,
0.1519775390625,
0.54248046875,
0.1519775390625,
0.54248046875,
0.163330078125,
0.5146484375,
0.163330078125
],
[
0.297119140625,
0.17041015625,
0.31982421875,
0.17041015625,
0.31982421875,
0.185302734375,
0.297119140625,
0.185302734375
],
[
0.280517578125,
0.15380859375,
0.291259765625,
0.15087890625,
0.295654296875,
0.157958984375,
0.285400390625,
0.160888671875
],
[
0.59765625,
0.17041015625,
0.79443359375,
0.17041015625,
0.79443359375,
0.1923828125,
0.59765625,
0.1923828125
],
[
0.85791015625,
0.17578125,
0.8818359375,
0.17578125,
0.8818359375,
0.189453125,
0.85791015625,
0.189453125
],
[
0.2032470703125,
0.185302734375,
0.2451171875,
0.185302734375,
0.2451171875,
0.20068359375,
0.2032470703125,
0.20068359375
],
[
0.274658203125,
0.185302734375,
0.300537109375,
0.185302734375,
0.300537109375,
0.198974609375,
0.274658203125,
0.198974609375
],
[
0.32080078125,
0.185302734375,
0.3720703125,
0.185302734375,
0.3720703125,
0.1995849609375,
0.32080078125,
0.1995849609375
],
[
0.38623046875,
0.1859130859375,
0.424072265625,
0.1859130859375,
0.424072265625,
0.198974609375,
0.38623046875,
0.198974609375
],
[
0.4375,
0.1859130859375,
0.483642578125,
0.1859130859375,
0.483642578125,
0.20068359375,
0.4375,
0.20068359375
],
[
0.67822265625,
0.191162109375,
0.70361328125,
0.191162109375,
0.70361328125,
0.203125,
0.67822265625,
0.203125
],
[
0.198974609375,
0.20068359375,
0.252685546875,
0.2037353515625,
0.2493896484375,
0.2197265625,
0.197265625,
0.215576171875
],
[
0.154541015625,
0.189453125,
0.16796875,
0.189453125,
0.16796875,
0.20068359375,
0.154541015625,
0.20068359375
],
[
0.1351318359375,
0.201904296875,
0.185546875,
0.201904296875,
0.185546875,
0.2161865234375,
0.1351318359375,
0.2161865234375
],
[
0.2626953125,
0.201904296875,
0.30908203125,
0.2037353515625,
0.307373046875,
0.2197265625,
0.26123046875,
0.21728515625
],
[
0.322509765625,
0.201904296875,
0.36865234375,
0.2037353515625,
0.366943359375,
0.2197265625,
0.32080078125,
0.21728515625
],
[
0.391357421875,
0.203125,
0.4189453125,
0.203125,
0.4189453125,
0.2161865234375,
0.391357421875,
0.2161865234375
],
[
0.455078125,
0.207275390625,
0.465087890625,
0.207275390625,
0.465087890625,
0.218505859375,
0.455078125,
0.218505859375
],
[
0.53076171875,
0.208984375,
0.8818359375,
0.208984375,
0.8818359375,
0.22265625,
0.53076171875,
0.22265625
],
[
0.20068359375,
0.21728515625,
0.2493896484375,
0.2197265625,
0.2476806640625,
0.235107421875,
0.198974609375,
0.2327880859375
],
[
0.264404296875,
0.21728515625,
0.30908203125,
0.2197265625,
0.307373046875,
0.235107421875,
0.2626953125,
0.2327880859375
],
[
0.32421875,
0.218505859375,
0.36865234375,
0.218505859375,
0.36865234375,
0.2340087890625,
0.32421875,
0.2340087890625
],
[
0.391357421875,
0.2197265625,
0.4189453125,
0.2197265625,
0.4189453125,
0.2327880859375,
0.391357421875,
0.2327880859375
],
[
0.45751953125,
0.223876953125,
0.46337890625,
0.223876953125,
0.46337890625,
0.23095703125,
0.45751953125,
0.23095703125
],
[
0.51123046875,
0.2244873046875,
0.6572265625,
0.2244873046875,
0.6572265625,
0.2381591796875,
0.51123046875,
0.2381591796875
],
[
0.1368408203125,
0.218505859375,
0.185546875,
0.218505859375,
0.185546875,
0.2327880859375,
0.1368408203125,
0.2327880859375
],
[
0.1368408203125,
0.2340087890625,
0.185546875,
0.2340087890625,
0.185546875,
0.248779296875,
0.1368408203125,
0.248779296875
],
[
0.20068359375,
0.235107421875,
0.2493896484375,
0.235107421875,
0.2493896484375,
0.2493896484375,
0.20068359375,
0.2493896484375
],
[
0.2626953125,
0.235107421875,
0.307373046875,
0.235107421875,
0.307373046875,
0.25048828125,
0.2626953125,
0.25048828125
],
[
0.32421875,
0.2327880859375,
0.370361328125,
0.235107421875,
0.36865234375,
0.25048828125,
0.322509765625,
0.248779296875
],
[
0.391357421875,
0.235107421875,
0.4189453125,
0.235107421875,
0.4189453125,
0.248779296875,
0.391357421875,
0.248779296875
],
[
0.444091796875,
0.2464599609375,
0.476806640625,
0.25048828125,
0.47265625,
0.267333984375,
0.43994140625,
0.26318359375
],
[
0.62060546875,
0.243408203125,
0.77099609375,
0.243408203125,
0.77099609375,
0.26416015625,
0.62060546875,
0.26416015625
],
[
0.20068359375,
0.2493896484375,
0.2493896484375,
0.251708984375,
0.2476806640625,
0.267333984375,
0.198974609375,
0.265380859375
],
[
0.85791015625,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.85791015625,
0.26123046875
],
[
0.32421875,
0.25048828125,
0.370361328125,
0.25048828125,
0.370361328125,
0.2666015625,
0.32421875,
0.2666015625
],
[
0.1368408203125,
0.25048828125,
0.185546875,
0.25048828125,
0.185546875,
0.265380859375,
0.1368408203125,
0.265380859375
],
[
0.2626953125,
0.251708984375,
0.30908203125,
0.251708984375,
0.30908203125,
0.2666015625,
0.2626953125,
0.2666015625
],
[
0.3896484375,
0.251708984375,
0.42236328125,
0.251708984375,
0.42236328125,
0.26416015625,
0.3896484375,
0.26416015625
],
[
0.444091796875,
0.26318359375,
0.478515625,
0.2666015625,
0.474365234375,
0.281982421875,
0.43994140625,
0.279052734375
],
[
0.67822265625,
0.26123046875,
0.71533203125,
0.26123046875,
0.71533203125,
0.2724609375,
0.67822265625,
0.2724609375
],
[
0.387939453125,
0.265380859375,
0.424072265625,
0.265380859375,
0.424072265625,
0.28076171875,
0.387939453125,
0.28076171875
],
[
0.1368408203125,
0.2666015625,
0.185546875,
0.2666015625,
0.185546875,
0.28076171875,
0.1368408203125,
0.28076171875
],
[
0.20068359375,
0.265380859375,
0.2509765625,
0.26953125,
0.2476806640625,
0.283935546875,
0.197265625,
0.27978515625
],
[
0.264404296875,
0.2666015625,
0.307373046875,
0.268310546875,
0.306396484375,
0.283935546875,
0.2626953125,
0.281982421875
],
[
0.32421875,
0.267333984375,
0.370361328125,
0.267333984375,
0.370361328125,
0.283203125,
0.32421875,
0.283203125
],
[
0.1368408203125,
0.281982421875,
0.185546875,
0.281982421875,
0.185546875,
0.29638671875,
0.1368408203125,
0.29638671875
],
[
0.20068359375,
0.281982421875,
0.2509765625,
0.2861328125,
0.2476806640625,
0.300537109375,
0.197265625,
0.29638671875
],
[
0.32421875,
0.281982421875,
0.370361328125,
0.2861328125,
0.366943359375,
0.3017578125,
0.32080078125,
0.297607421875
],
[
0.53076171875,
0.27783203125,
0.8818359375,
0.279052734375,
0.8818359375,
0.29345703125,
0.53076171875,
0.292236328125
],
[
0.391357421875,
0.283203125,
0.4189453125,
0.283203125,
0.4189453125,
0.29638671875,
0.391357421875,
0.29638671875
],
[
0.2626953125,
0.283203125,
0.307373046875,
0.28515625,
0.306396484375,
0.2998046875,
0.26123046875,
0.297607421875
],
[
0.44921875,
0.2880859375,
0.4609375,
0.283203125,
0.47021484375,
0.29345703125,
0.45751953125,
0.298583984375
],
[
0.5146484375,
0.295654296875,
0.880859375,
0.295654296875,
0.880859375,
0.308837890625,
0.5146484375,
0.308837890625
],
[
0.2149658203125,
0.303955078125,
0.2333984375,
0.303955078125,
0.2333984375,
0.318359375,
0.2149658203125,
0.318359375
],
[
0.455078125,
0.303955078125,
0.466796875,
0.303955078125,
0.466796875,
0.31640625,
0.455078125,
0.31640625
],
[
0.39794921875,
0.3046875,
0.411376953125,
0.3046875,
0.411376953125,
0.317626953125,
0.39794921875,
0.317626953125
],
[
0.1561279296875,
0.3046875,
0.16796875,
0.3046875,
0.16796875,
0.317626953125,
0.1561279296875,
0.317626953125
],
[
0.279541015625,
0.305908203125,
0.29296875,
0.305908203125,
0.29296875,
0.317626953125,
0.279541015625,
0.317626953125
],
[
0.340087890625,
0.305908203125,
0.353515625,
0.305908203125,
0.353515625,
0.31640625,
0.340087890625,
0.31640625
],
[
0.51318359375,
0.311279296875,
0.8818359375,
0.311279296875,
0.8818359375,
0.32470703125,
0.51318359375,
0.32470703125
],
[
0.297119140625,
0.32666015625,
0.32080078125,
0.32666015625,
0.32080078125,
0.34033203125,
0.297119140625,
0.34033203125
],
[
0.51318359375,
0.327880859375,
0.8818359375,
0.327880859375,
0.8818359375,
0.341552734375,
0.51318359375,
0.341552734375
],
[
0.51123046875,
0.342529296875,
0.880859375,
0.342529296875,
0.880859375,
0.355712890625,
0.51123046875,
0.355712890625
],
[
0.1175537109375,
0.3515625,
0.486083984375,
0.3515625,
0.486083984375,
0.365234375,
0.1175537109375,
0.365234375
],
[
0.5146484375,
0.35986328125,
0.880859375,
0.35986328125,
0.880859375,
0.37353515625,
0.5146484375,
0.37353515625
],
[
0.12005615234375,
0.3681640625,
0.483642578125,
0.3681640625,
0.483642578125,
0.37890625,
0.12005615234375,
0.37890625
],
[
0.118408203125,
0.38134765625,
0.48779296875,
0.38134765625,
0.48779296875,
0.394287109375,
0.118408203125,
0.394287109375
],
[
0.51318359375,
0.3759765625,
0.880859375,
0.3759765625,
0.880859375,
0.3896484375,
0.51318359375,
0.3896484375
],
[
0.11419677734375,
0.392578125,
0.48779296875,
0.393798828125,
0.48779296875,
0.407958984375,
0.11419677734375,
0.40673828125
],
[
0.51318359375,
0.392578125,
0.861328125,
0.392578125,
0.861328125,
0.40625,
0.51318359375,
0.40625
],
[
0.1175537109375,
0.4091796875,
0.48779296875,
0.4091796875,
0.48779296875,
0.4228515625,
0.1175537109375,
0.4228515625
],
[
0.53564453125,
0.407958984375,
0.880859375,
0.407958984375,
0.880859375,
0.421630859375,
0.53564453125,
0.421630859375
],
[
0.118408203125,
0.424560546875,
0.48779296875,
0.424560546875,
0.48779296875,
0.435302734375,
0.118408203125,
0.435302734375
],
[
0.5146484375,
0.424560546875,
0.880859375,
0.424560546875,
0.880859375,
0.438232421875,
0.5146484375,
0.438232421875
],
[
0.1175537109375,
0.43701171875,
0.306396484375,
0.438232421875,
0.306396484375,
0.451904296875,
0.1175537109375,
0.45068359375
],
[
0.5146484375,
0.4423828125,
0.8759765625,
0.4423828125,
0.8759765625,
0.453125,
0.5146484375,
0.453125
],
[
0.51123046875,
0.457275390625,
0.87744140625,
0.457275390625,
0.87744140625,
0.47021484375,
0.51123046875,
0.47021484375
],
[
0.51123046875,
0.47265625,
0.87890625,
0.47265625,
0.87890625,
0.486328125,
0.51123046875,
0.486328125
],
[
0.12005615234375,
0.478515625,
0.484375,
0.478515625,
0.484375,
0.4921875,
0.12005615234375,
0.4921875
],
[
0.118408203125,
0.49462890625,
0.483642578125,
0.49462890625,
0.483642578125,
0.50830078125,
0.118408203125,
0.50830078125
],
[
0.51318359375,
0.4892578125,
0.880859375,
0.4892578125,
0.880859375,
0.5029296875,
0.51318359375,
0.5029296875
],
[
0.1175537109375,
0.51025390625,
0.483642578125,
0.51025390625,
0.483642578125,
0.52392578125,
0.1175537109375,
0.52392578125
],
[
0.51318359375,
0.5048828125,
0.880859375,
0.5048828125,
0.880859375,
0.5185546875,
0.51318359375,
0.5185546875
],
[
0.51123046875,
0.52099609375,
0.87890625,
0.52099609375,
0.87890625,
0.53369140625,
0.51123046875,
0.53369140625
],
[
0.118408203125,
0.52685546875,
0.486083984375,
0.52685546875,
0.486083984375,
0.54052734375,
0.118408203125,
0.54052734375
],
[
0.5146484375,
0.5361328125,
0.880859375,
0.5361328125,
0.880859375,
0.5498046875,
0.5146484375,
0.5498046875
],
[
0.118408203125,
0.54443359375,
0.306396484375,
0.54443359375,
0.306396484375,
0.5546875,
0.118408203125,
0.5546875
],
[
0.53076171875,
0.552734375,
0.880859375,
0.552734375,
0.880859375,
0.56640625,
0.53076171875,
0.56640625
],
[
0.1368408203125,
0.55859375,
0.48779296875,
0.55859375,
0.48779296875,
0.572265625,
0.1368408203125,
0.572265625
],
[
0.51318359375,
0.568359375,
0.880859375,
0.568359375,
0.880859375,
0.58203125,
0.51318359375,
0.58203125
],
[
0.118408203125,
0.576171875,
0.486083984375,
0.576171875,
0.486083984375,
0.5888671875,
0.118408203125,
0.5888671875
],
[
0.51318359375,
0.5849609375,
0.87890625,
0.5849609375,
0.87890625,
0.5986328125,
0.51318359375,
0.5986328125
],
[
0.12005615234375,
0.5927734375,
0.2283935546875,
0.5927734375,
0.2283935546875,
0.60400390625,
0.12005615234375,
0.60400390625
],
[
0.1351318359375,
0.60693359375,
0.48193359375,
0.60693359375,
0.48193359375,
0.62060546875,
0.1351318359375,
0.62060546875
],
[
0.51318359375,
0.6015625,
0.880859375,
0.6015625,
0.880859375,
0.615234375,
0.51318359375,
0.615234375
],
[
0.51123046875,
0.61767578125,
0.880859375,
0.61767578125,
0.880859375,
0.63134765625,
0.51123046875,
0.63134765625
],
[
0.118408203125,
0.62353515625,
0.486083984375,
0.62353515625,
0.486083984375,
0.63720703125,
0.118408203125,
0.63720703125
],
[
0.5146484375,
0.63427734375,
0.87890625,
0.63427734375,
0.87890625,
0.64794921875,
0.5146484375,
0.64794921875
],
[
0.118408203125,
0.64013671875,
0.32080078125,
0.64013671875,
0.32080078125,
0.65380859375,
0.118408203125,
0.65380859375
],
[
0.515625,
0.6494140625,
0.880859375,
0.6494140625,
0.880859375,
0.6630859375,
0.515625,
0.6630859375
],
[
0.11419677734375,
0.66748046875,
0.3046875,
0.66943359375,
0.3046875,
0.68408203125,
0.11419677734375,
0.681640625
],
[
0.51318359375,
0.6650390625,
0.880859375,
0.6650390625,
0.880859375,
0.6787109375,
0.51318359375,
0.6787109375
],
[
0.51318359375,
0.68115234375,
0.87890625,
0.68115234375,
0.87890625,
0.69482421875,
0.51318359375,
0.69482421875
],
[
0.118408203125,
0.6953125,
0.48779296875,
0.6953125,
0.48779296875,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51318359375,
0.69775390625,
0.880859375,
0.69775390625,
0.880859375,
0.71142578125,
0.51318359375,
0.71142578125
],
[
0.118408203125,
0.71142578125,
0.486083984375,
0.71142578125,
0.486083984375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.51123046875,
0.71337890625,
0.87890625,
0.71337890625,
0.87890625,
0.72705078125,
0.51123046875,
0.72705078125
],
[
0.1175537109375,
0.7255859375,
0.486083984375,
0.7255859375,
0.486083984375,
0.7392578125,
0.1175537109375,
0.7392578125
],
[
0.51318359375,
0.72998046875,
0.880859375,
0.72998046875,
0.880859375,
0.74365234375,
0.51318359375,
0.74365234375
],
[
0.118408203125,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.51123046875,
0.7451171875,
0.87890625,
0.7451171875,
0.87890625,
0.7587890625,
0.51123046875,
0.7587890625
],
[
0.12005615234375,
0.76123046875,
0.391357421875,
0.76123046875,
0.391357421875,
0.771484375,
0.12005615234375,
0.771484375
],
[
0.5146484375,
0.7626953125,
0.880859375,
0.7626953125,
0.880859375,
0.775390625,
0.5146484375,
0.775390625
],
[
0.1368408203125,
0.77685546875,
0.48779296875,
0.77685546875,
0.48779296875,
0.79052734375,
0.1368408203125,
0.79052734375
],
[
0.51318359375,
0.77783203125,
0.880859375,
0.77783203125,
0.880859375,
0.79150390625,
0.51318359375,
0.79150390625
],
[
0.130126953125,
0.7919921875,
0.48779296875,
0.7919921875,
0.48779296875,
0.8056640625,
0.130126953125,
0.8056640625
],
[
0.51318359375,
0.79443359375,
0.87890625,
0.79443359375,
0.87890625,
0.80810546875,
0.51318359375,
0.80810546875
],
[
0.118408203125,
0.80712890625,
0.483642578125,
0.80810546875,
0.483642578125,
0.82177734375,
0.118408203125,
0.82080078125
],
[
0.51123046875,
0.80859375,
0.8818359375,
0.81005859375,
0.8818359375,
0.82470703125,
0.51123046875,
0.82373046875
],
[
0.12005615234375,
0.826171875,
0.486083984375,
0.826171875,
0.486083984375,
0.8359375,
0.12005615234375,
0.8359375
],
[
0.51318359375,
0.82666015625,
0.880859375,
0.82666015625,
0.880859375,
0.84033203125,
0.51318359375,
0.84033203125
],
[
0.118408203125,
0.84033203125,
0.486083984375,
0.84033203125,
0.486083984375,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.51123046875,
0.84033203125,
0.623046875,
0.8427734375,
0.623046875,
0.85693359375,
0.51123046875,
0.85498046875
],
[
0.1175537109375,
0.85546875,
0.484375,
0.85546875,
0.484375,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.51318359375,
0.8681640625,
0.5615234375,
0.8681640625,
0.5615234375,
0.8798828125,
0.51318359375,
0.8798828125
],
[
0.55517578125,
0.8681640625,
0.72119140625,
0.8681640625,
0.72119140625,
0.8818359375,
0.55517578125,
0.8818359375
],
[
0.118408203125,
0.88818359375,
0.307373046875,
0.88818359375,
0.307373046875,
0.90185546875,
0.118408203125,
0.90185546875
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.5146484375,
0.8896484375,
0.87744140625,
0.8896484375,
0.87744140625,
0.90283203125,
0.5146484375,
0.90283203125
],
[
0.1376953125,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.1376953125,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"Label",
"\"",
"“000\"",
"1994）的变体，其中模型的预测用于确定探",
"Btotal",
"\"56\"",
"0.3",
"-",
"-",
"索可能的标签分配的顺序。每个状态表示所有",
"Itotal",
"0.4",
"0.4",
"n令牌x={a1,a2,...,an】的完整标签分配",
"Bcash",
"0.5",
"-",
"y=【y1,y2,..yn}。状态的成本g(y）定义如",
"Ica.sh",
"0.3",
"0.3",
"下:",
"(a)",
".",
"g(y) = - log p(yix).",
"(4)",
"“56\"",
"“\"",
"“000\"",
"BIO",
"Sem.",
"i=1",
"Bcash",
"p",
"8.0%",
"Itotal",
"Itotal",
"No",
"-",
"我们使用y表示k最低成本标签分配，并",
"Bcash",
"Itotal",
"Icash",
"No",
"-",
"将起点yl定义为：",
"6.0 %",
"6.0 %",
"Bcash",
"Icash",
"Itotal",
"No",
"No",
"yl = arg min g(y)",
"Btotal",
"(5)",
"Itotal",
"4.8 %",
"Itotal",
"Yes",
"Yes",
"yeyn",
"Yes",
"4.5 %",
"Bcash",
"Icash",
"Icash",
"3.6 %",
"Btotal",
"Icash",
"在Alg中给出了Lazy-k译码算法。1。它",
"No",
"Itotal",
"、",
"的工作原理是维护k个最佳状态的堆，按1个",
"",
". . .",
":",
"...",
":",
"：",
"编辑距离内的下一个最佳未开发状态的得分进",
"(b)",
"行优先级排序。用起始状态y1初始化堆。在",
"探索状态时，将针对约束进行测试，如果满足，",
"Table 1: (a) (partial) predicted probabilities for the red",
"则直接返回。如果不满足约束，则使用新探索",
"boundingboxinFig.1,wherethemodel splitsthe",
"string“56.000\"in the three tokens\"56\",\"\",and",
"的状态扩展堆，并更新原始状态y的优先级",
"\"000\".(b) most likely label assignments in order of",
"得分，以反映下一个最佳未探索状态的得分",
"probability. BIO = the labels adhere to the BIO con-",
"与最佳优先搜索不同，在探索状态时，我们",
"straints, Sem. =the labels adhere to the semantic struc-",
"不会将所有子级添加到堆中。相反，我们只添",
"ture (cash = total + change).",
"加下一个最佳状态y，并将y的优先级键更",
"新为编辑距离1内的下一个最佳状态的分数，",
"这大大降低了堆的大小，因为经典的搜索实现",
"文中已经提出了几种改编，但是没有一种可以",
"解决全局约束的上述问题。其他人建议使用可",
"在每次迭代时都会添加n可能的子节点，而",
"学习的启发式方法扩展波束搜索，以尝试预测",
"在这种情况下，堆中的状态数最多等于迭代次",
"数。随着所有后续堆操作变得更便宜，这种堆",
"给定的标签分配是否可能违反未来的结构约束",
"大小的减少反过来转化为更好的运行时复杂性",
"(Pan and Srikumar.2018)",
"NextBest函数将状态y和前沿作为输入。边",
"我们的方法遵循与A*类似的方法，其中部",
"界是一个字典，其中包含所有已探索状态的已",
"分扩展（Yoshizumietal，2000）先前已应用于多",
"探索状态和下一个最佳状态。这些值是跟踪给",
"序列比对问题",
"据我们所知，我们是第一个将A*与部分扩",
"定状态的第i个最佳更改的整数。如果i==n",
"（令牌数），则该函数返回null，因为在此状态",
"展一起应用的，该扩展允许比ILP更通用的约",
"的1编辑距离内没有下一个最佳更改。由于",
"束用于全局约束解码设置",
"下一个最佳状态可能已经存在于边界中，因此",
"4Lazy-kDecoding",
"NextBest函数包装在AddNextBest中，以确保",
"仅将下一个最佳状态添加到尚未存在的边界。",
"顾名思义，Lazy-k解码器允许以懒惰的方式",
"请参阅应用程序。B用于NextBest函数的伪代",
"解码k最可能的序列。这意味着它只迭代必要",
"码。鉴于该算法以概率的降序迭代可能的序",
"数量的序列，一且找到令人满意的解决方案就",
"列，因此证明它总是会找到存在的最优解是微",
"停止。这个解码器探索的假设是满足约束的序",
"不足道的。然而，在实践中，状态数量的组合",
"列在其他高概率序列中的某个地方",
"增长很快使穷举搜索变得不可行。为了防止这",
"为了有效地做到这一点，我们利用了这样",
"种情况，使用一个额外的停止条件，如果在固",
"个事实，即第k个最可能的序列总是在??？",
"定数量的k选代后没有找到令人满意的解决方",
"中的一个k一1更可能的序列中。这源于每个",
"案，则迭代停止。还可以根据累积概率质量p",
"标签的独立性，如公式1所示。我们将“edit-",
"或其他度量来设置停止条件；我们将此探索留",
"distance”放在引号中，因为我们使用稍微更严",
"给将来的工作",
"格的编辑距离定义，该定义还考虑了各种标签",
"4.1",
"Complexity Analysis",
"在应用程序中找到。A",
"概率之间的顺序。有关此的更多详细信息可以",
"假设n标记、l标签和请求的top-k序列。Lazy",
"它的核心是最佳优先搜索(Russel etal.，",
"k的空间复杂度为O（kn），因为对于每个k"
] |
{
"arxiv_id": "2312.03367",
"language": "zh",
"page_idx": 3
}
|
||
2312.03367-zh-4
|
2312.03367-zh-4.jpg
|
$\mathbf{Algorithm1Lazy-kDecoding}$
Require: Input sequence: x
Require: Cost function $g$ from Eq.4
$\bar{\textbf{Require: Binary constraint}\mathcal{C}}$,max iterations $k$
$1: \textbf{function LAZY- K( x, g, C, k) }$
2:
$y^{1}\leftarrow \arg \min _{\mathrm{y\in \mathcal{Y} ^{n}}}$g( y)
3:
$\textbf{if}\mathcal{C} ($x,y$^1)=1\textbf{then return y}^1$
4:
$H\leftarrow$MinHeap()
5:
frontier$\leftarrow\{\mathrm{y}^{\tilde{1}}:1\}$
6:
AddNextBest (y$^1,H$,frontier )
7:
count$\leftarrow1$
8:
while $H$ not empty and count<k do
$\mathrm{y}^{i}\leftarrow H$ .PopMin()
9: 10:
$\mathrm{y}^{k}\leftarrow$NextBest$(\mathrm{y}^{i}$,frontier)
$\textbf{if}{\mathcal{C} }($x,y$^k)=1\textbf{then return y}^k$
11: 12:
AddNextBest (y$^k,H$,frontier )
13:
AddNextBest (y$^i,H$,frontier )
14:
count +=1
$\textbf{end while}$
15:
return Failure
16:
17: end function
18:
19: function ADDNEXTBEST( y$^i,H$, frontier)
$\mathrm{y}^{ij}\leftarrow$NextBest (y$^i$,frontier)
20: 21:
$\mathbf{while~y}^{ij}\neq$null and y$^ij\in$frontier do
22:
frontier [ $y^{i}$] + = 1
23:
$\mathrm{y}^{ij}\leftarrow$NextBest (y$^i$,frontier)
end while
24: 25:
$\textbf{if y}^{ij}\neq$null then
26:
frontier[ y$^{ij}]\leftarrow1$
27:
$H$ .Add( y$^i, g( \mathrm{y} ^{ij})$ )
end if
28:
29: end function
最佳状态，我们最多添加 1 个大小为$n$的新状
态到堆
时间复杂度稍微不那么明显。对于 top- $k$状态，外部 while 循环将运行$k$迭代。在这个循环中，有两个复杂性来源：
1. $H$ .Add() which occurs at most twice in
AddNextBest(),
2. NextBest() which occurs once in the outer
loop and twice in AddNextBest in another
while loop.
$H$ 。add() 操作向堆添加一个元素，该元素相对于堆的大小具有对数复杂度。由于堆在每次迭代时都完全保存了我们的顶级$k$状态，因此该操作的复杂性等于$\sum_i=1^k\log i=$
Table 2: Datasets used in the experiments. APL=Average Page Length, CSR=Constraint Satisfaction Ratio
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>Pages</th>
<th>$APL$</th>
<th>$CSR$</th>
</tr>
<tr>
<td>CORD</td>
<td>1,000</td>
<td>46</td>
<td>72 . $\because\%$</td>
</tr>
<tr>
<td>${\mathrm{WildReceipt}}$</td>
<td>1,740</td>
<td>147</td>
<td>70 $\%$</td>
</tr>
<tr>
<td>Docile</td>
<td>7,372</td>
<td>353</td>
<td>88 $0/_{0}$</td>
</tr>
</tbody>
</table>
$\log \prod _{i= 1}^ki= \log k!$ ,使用斯特林近似，它等效
于O$(k\log k)$
NextBest( y, frontier) 函数返回 y 的 1 个编辑距离内的第 [y] 次最佳状态。首次扩展状态时， 我们计算 O 中的下一个最佳编辑的排序列表$(n\log n)$时间。使用此方法，可以在恒定时间内计算此状态的每个 NextBest 调用。对于每个状态，最多调用 NextBest() $n$次。由于排序需要$n\log n$时间，因此算法的总时间复杂度变为：O $(k(\log k+n\log n)$
## 5 $\textbf{Experiments}$
为了评估 Lazy- $k$解码器的相关性，我们对多个数据集执行发票信息提取。该任务的目的是从发票中提取各种金额，使其满足预期的算术结构。对于每个数据集，我们训练一个令牌分类模型并为测试集生成预测。然后将预测与约束一起馈送到不同的解码算法中，并返回满足约束的最高概率序列
Data 我们在 Tab 中显示的总共三个数据集上评估解码器。2:CORD (Park et al., 2019) 、 WildReceipt (Sun et al., 2021) 和 DocILE (Šimsa et al.,2023)。虽然约束在语义上非常相似，但数据集的标签和粒度级别略有不同，这意味着每个数据集都有自己特定的约束集
每个数据集的确切约束在 App 中有详细说明。C。所有数据集的模型都使用 BIO 标签进行训练，其初始约束是标签分配应该是有效的 BIO 序列
其他约束取决于每个数据集中可用的特定标签。然而，有些标签可能不存在于每个样品中。因此，我们区分必填字段(即总金额)和可选字段(即服务费，折扣),如果找不到，则将其视为0 。如果预测中不存在必填字段将被视为空。因此，涉及此字段的任何约束都不会被评估(或自动视为已满足)
对于每个数据集，我们将约束应用于所有样本，并过滤掉任何不满足约束的样本(在表中显示计数)。从这些示例中，我们使用 60 个% 进行训练和验证(拆分 80-20),使用 40 个% 进行测试。将不满足约束的样本添加到训练集中。我们有目的地选择一个大百分比的测试， 因为小火车集提供了足够的性能，我们大多希
|
[
[
0.1175537109375,
0.0843505859375,
0.3525390625,
0.08489990234375,
0.3525390625,
0.09979248046875,
0.1175537109375,
0.09857177734375
],
[
0.51318359375,
0.0843505859375,
0.880859375,
0.0843505859375,
0.880859375,
0.09442138671875,
0.51318359375,
0.09442138671875
],
[
0.51123046875,
0.097412109375,
0.87890625,
0.097412109375,
0.87890625,
0.11102294921875,
0.51123046875,
0.11102294921875
],
[
0.118408203125,
0.1051025390625,
0.330078125,
0.1051025390625,
0.330078125,
0.1181640625,
0.118408203125,
0.1181640625
],
[
0.1158447265625,
0.12054443359375,
0.3955078125,
0.12054443359375,
0.3955078125,
0.1353759765625,
0.1158447265625,
0.1353759765625
],
[
0.73486328125,
0.12353515625,
0.7734375,
0.12353515625,
0.7734375,
0.1395263671875,
0.73486328125,
0.1395263671875
],
[
0.79248046875,
0.12353515625,
0.83642578125,
0.12353515625,
0.83642578125,
0.1395263671875,
0.79248046875,
0.1395263671875
],
[
0.67333984375,
0.12469482421875,
0.72119140625,
0.12469482421875,
0.72119140625,
0.14013671875,
0.67333984375,
0.14013671875
],
[
0.5576171875,
0.12469482421875,
0.62060546875,
0.12469482421875,
0.62060546875,
0.1395263671875,
0.5576171875,
0.1395263671875
],
[
0.1158447265625,
0.135986328125,
0.47021484375,
0.135986328125,
0.47021484375,
0.149658203125,
0.1158447265625,
0.149658203125
],
[
0.5576171875,
0.14013671875,
0.6171875,
0.14013671875,
0.6171875,
0.155029296875,
0.5576171875,
0.155029296875
],
[
0.67333984375,
0.14013671875,
0.72119140625,
0.14013671875,
0.72119140625,
0.1561279296875,
0.67333984375,
0.1561279296875
],
[
0.7314453125,
0.141357421875,
0.75830078125,
0.141357421875,
0.75830078125,
0.155029296875,
0.7314453125,
0.155029296875
],
[
0.78857421875,
0.14013671875,
0.83642578125,
0.14013671875,
0.83642578125,
0.1561279296875,
0.78857421875,
0.1561279296875
],
[
0.1259765625,
0.1519775390625,
0.370361328125,
0.152587890625,
0.370361328125,
0.16748046875,
0.1259765625,
0.166259765625
],
[
0.55908203125,
0.1561279296875,
0.66015625,
0.157958984375,
0.66015625,
0.1728515625,
0.55908203125,
0.17041015625
],
[
0.73291015625,
0.1561279296875,
0.76904296875,
0.1561279296875,
0.76904296875,
0.171630859375,
0.73291015625,
0.171630859375
],
[
0.7900390625,
0.1561279296875,
0.8330078125,
0.1561279296875,
0.8330078125,
0.171630859375,
0.7900390625,
0.171630859375
],
[
0.173828125,
0.16748046875,
0.355224609375,
0.169189453125,
0.355224609375,
0.18408203125,
0.173828125,
0.1817626953125
],
[
0.67236328125,
0.15673828125,
0.7197265625,
0.15673828125,
0.7197265625,
0.171630859375,
0.67236328125,
0.171630859375
],
[
0.1259765625,
0.169189453125,
0.145263671875,
0.169189453125,
0.145263671875,
0.182861328125,
0.1259765625,
0.182861328125
],
[
0.67236328125,
0.1728515625,
0.7197265625,
0.1728515625,
0.7197265625,
0.18701171875,
0.67236328125,
0.18701171875
],
[
0.73486328125,
0.17333984375,
0.767578125,
0.17333984375,
0.767578125,
0.1859130859375,
0.73486328125,
0.1859130859375
],
[
0.7900390625,
0.171630859375,
0.83642578125,
0.17333984375,
0.83447265625,
0.188232421875,
0.78857421875,
0.1859130859375
],
[
0.175537109375,
0.185302734375,
0.414794921875,
0.185302734375,
0.414794921875,
0.198974609375,
0.175537109375,
0.198974609375
],
[
0.56103515625,
0.174560546875,
0.61279296875,
0.174560546875,
0.61279296875,
0.1859130859375,
0.56103515625,
0.1859130859375
],
[
0.124267578125,
0.1859130859375,
0.145263671875,
0.1859130859375,
0.145263671875,
0.1995849609375,
0.124267578125,
0.1995849609375
],
[
0.124267578125,
0.201904296875,
0.145263671875,
0.201904296875,
0.145263671875,
0.215576171875,
0.124267578125,
0.215576171875
],
[
0.175537109375,
0.1995849609375,
0.3046875,
0.20068359375,
0.3046875,
0.215576171875,
0.175537109375,
0.21435546875
],
[
0.17626953125,
0.2161865234375,
0.32080078125,
0.2161865234375,
0.32080078125,
0.23095703125,
0.17626953125,
0.23095703125
],
[
0.509765625,
0.215576171875,
0.87890625,
0.215576171875,
0.87890625,
0.2298583984375,
0.509765625,
0.2298583984375
],
[
0.5146484375,
0.232177734375,
0.61865234375,
0.232177734375,
0.61865234375,
0.2464599609375,
0.5146484375,
0.2464599609375
],
[
0.124267578125,
0.218505859375,
0.1435546875,
0.218505859375,
0.1435546875,
0.232177734375,
0.124267578125,
0.232177734375
],
[
0.124267578125,
0.2340087890625,
0.1435546875,
0.2340087890625,
0.1435546875,
0.248779296875,
0.124267578125,
0.248779296875
],
[
0.17626953125,
0.2327880859375,
0.405517578125,
0.2327880859375,
0.405517578125,
0.2476806640625,
0.17626953125,
0.2476806640625
],
[
0.124267578125,
0.2493896484375,
0.1435546875,
0.2493896484375,
0.1435546875,
0.26318359375,
0.124267578125,
0.26318359375
],
[
0.17626953125,
0.25048828125,
0.255126953125,
0.25048828125,
0.255126953125,
0.262451171875,
0.17626953125,
0.262451171875
],
[
0.52880859375,
0.2476806640625,
0.880859375,
0.2476806640625,
0.880859375,
0.262451171875,
0.52880859375,
0.262451171875
],
[
0.17626953125,
0.265380859375,
0.45751953125,
0.265380859375,
0.45751953125,
0.27978515625,
0.17626953125,
0.27978515625
],
[
0.51318359375,
0.26416015625,
0.880859375,
0.26416015625,
0.880859375,
0.27783203125,
0.51318359375,
0.27783203125
],
[
0.51318359375,
0.279052734375,
0.880859375,
0.27978515625,
0.880859375,
0.29443359375,
0.51318359375,
0.29345703125
],
[
0.124267578125,
0.2666015625,
0.1435546875,
0.2666015625,
0.1435546875,
0.27978515625,
0.124267578125,
0.27978515625
],
[
0.1234130859375,
0.281982421875,
0.1435546875,
0.281982421875,
0.1435546875,
0.29638671875,
0.1234130859375,
0.29638671875
],
[
0.202392578125,
0.281982421875,
0.345947265625,
0.281982421875,
0.345947265625,
0.295654296875,
0.202392578125,
0.295654296875
],
[
0.202392578125,
0.29638671875,
0.41650390625,
0.297607421875,
0.41650390625,
0.312255859375,
0.202392578125,
0.311279296875
],
[
0.509765625,
0.29443359375,
0.880859375,
0.295654296875,
0.880859375,
0.31005859375,
0.509765625,
0.308837890625
],
[
0.20068359375,
0.312255859375,
0.43994140625,
0.311279296875,
0.43994140625,
0.328857421875,
0.20068359375,
0.330078125
],
[
0.12005615234375,
0.298583984375,
0.1427001953125,
0.298583984375,
0.1427001953125,
0.312255859375,
0.12005615234375,
0.312255859375
],
[
0.12005615234375,
0.314208984375,
0.1435546875,
0.314208984375,
0.1435546875,
0.328857421875,
0.12005615234375,
0.328857421875
],
[
0.515625,
0.312255859375,
0.87890625,
0.312255859375,
0.87890625,
0.325927734375,
0.515625,
0.325927734375
],
[
0.202392578125,
0.328857421875,
0.43408203125,
0.330078125,
0.43408203125,
0.344482421875,
0.202392578125,
0.34326171875
],
[
0.51318359375,
0.327880859375,
0.880859375,
0.327880859375,
0.880859375,
0.341552734375,
0.51318359375,
0.341552734375
],
[
0.202392578125,
0.344482421875,
0.432373046875,
0.3466796875,
0.432373046875,
0.361083984375,
0.202392578125,
0.359375
],
[
0.12005615234375,
0.330810546875,
0.1427001953125,
0.330810546875,
0.1427001953125,
0.344482421875,
0.12005615234375,
0.344482421875
],
[
0.12005615234375,
0.3466796875,
0.1427001953125,
0.3466796875,
0.1427001953125,
0.35986328125,
0.12005615234375,
0.35986328125
],
[
0.5146484375,
0.344482421875,
0.880859375,
0.344482421875,
0.880859375,
0.358154296875,
0.5146484375,
0.358154296875
],
[
0.202392578125,
0.3623046875,
0.287109375,
0.3623046875,
0.287109375,
0.377197265625,
0.202392578125,
0.377197265625
],
[
0.51123046875,
0.359375,
0.7314453125,
0.35986328125,
0.7314453125,
0.374755859375,
0.51123046875,
0.37353515625
],
[
0.173828125,
0.377197265625,
0.252685546875,
0.377197265625,
0.252685546875,
0.391357421875,
0.173828125,
0.391357421875
],
[
0.118408203125,
0.363525390625,
0.1435546875,
0.363525390625,
0.1435546875,
0.3759765625,
0.118408203125,
0.3759765625
],
[
0.12005615234375,
0.37890625,
0.1435546875,
0.37890625,
0.1435546875,
0.392578125,
0.12005615234375,
0.392578125
],
[
0.50732421875,
0.385498046875,
0.66162109375,
0.38720703125,
0.66162109375,
0.405029296875,
0.50732421875,
0.402587890625
],
[
0.12005615234375,
0.3955078125,
0.1435546875,
0.3955078125,
0.1435546875,
0.4091796875,
0.12005615234375,
0.4091796875
],
[
0.17626953125,
0.3955078125,
0.28369140625,
0.3955078125,
0.28369140625,
0.40673828125,
0.17626953125,
0.40673828125
],
[
0.118408203125,
0.410400390625,
0.2493896484375,
0.410400390625,
0.2493896484375,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.5146484375,
0.413330078125,
0.880859375,
0.413330078125,
0.880859375,
0.427001953125,
0.5146484375,
0.427001953125
],
[
0.12005615234375,
0.427001953125,
0.1427001953125,
0.427001953125,
0.1427001953125,
0.440673828125,
0.12005615234375,
0.440673828125
],
[
0.5146484375,
0.429931640625,
0.880859375,
0.429931640625,
0.880859375,
0.443603515625,
0.5146484375,
0.443603515625
],
[
0.118408203125,
0.4423828125,
0.465087890625,
0.4423828125,
0.465087890625,
0.4560546875,
0.118408203125,
0.4560546875
],
[
0.5146484375,
0.4453125,
0.880859375,
0.4453125,
0.880859375,
0.458984375,
0.5146484375,
0.458984375
],
[
0.118408203125,
0.460205078125,
0.1427001953125,
0.460205078125,
0.1427001953125,
0.47265625,
0.118408203125,
0.47265625
],
[
0.175537109375,
0.457275390625,
0.393798828125,
0.457763671875,
0.393798828125,
0.47265625,
0.175537109375,
0.471435546875
],
[
0.51123046875,
0.46142578125,
0.87744140625,
0.46142578125,
0.87744140625,
0.474365234375,
0.51123046875,
0.474365234375
],
[
0.118408203125,
0.4755859375,
0.1427001953125,
0.4755859375,
0.1427001953125,
0.4892578125,
0.118408203125,
0.4892578125
],
[
0.17626953125,
0.474365234375,
0.4609375,
0.474365234375,
0.4609375,
0.488037109375,
0.17626953125,
0.488037109375
],
[
0.51318359375,
0.47802734375,
0.87890625,
0.47802734375,
0.87890625,
0.4912109375,
0.51318359375,
0.4912109375
],
[
0.118408203125,
0.490478515625,
0.1427001953125,
0.490478515625,
0.1427001953125,
0.5048828125,
0.118408203125,
0.5048828125
],
[
0.2032470703125,
0.4912109375,
0.330078125,
0.4912109375,
0.330078125,
0.5048828125,
0.2032470703125,
0.5048828125
],
[
0.51123046875,
0.493408203125,
0.87744140625,
0.493408203125,
0.87744140625,
0.50732421875,
0.51123046875,
0.50732421875
],
[
0.118408203125,
0.50830078125,
0.1435546875,
0.50830078125,
0.1435546875,
0.521484375,
0.118408203125,
0.521484375
],
[
0.198974609375,
0.50439453125,
0.420654296875,
0.5048828125,
0.420654296875,
0.5224609375,
0.198974609375,
0.521484375
],
[
0.51318359375,
0.51025390625,
0.67822265625,
0.51025390625,
0.67822265625,
0.52392578125,
0.51318359375,
0.52392578125
],
[
0.118408203125,
0.52392578125,
0.1435546875,
0.52392578125,
0.1435546875,
0.53759765625,
0.118408203125,
0.53759765625
],
[
0.177978515625,
0.52392578125,
0.252685546875,
0.52392578125,
0.252685546875,
0.53515625,
0.177978515625,
0.53515625
],
[
0.175537109375,
0.5380859375,
0.312255859375,
0.5380859375,
0.312255859375,
0.5517578125,
0.175537109375,
0.5517578125
],
[
0.51708984375,
0.53515625,
0.87890625,
0.53515625,
0.87890625,
0.548828125,
0.51708984375,
0.548828125
],
[
0.51708984375,
0.55029296875,
0.8818359375,
0.55029296875,
0.8818359375,
0.56396484375,
0.51708984375,
0.56396484375
],
[
0.118408203125,
0.54052734375,
0.1427001953125,
0.54052734375,
0.1427001953125,
0.55419921875,
0.118408203125,
0.55419921875
],
[
0.118408203125,
0.5556640625,
0.1435546875,
0.5556640625,
0.1435546875,
0.5693359375,
0.118408203125,
0.5693359375
],
[
0.2032470703125,
0.5546875,
0.3359375,
0.5546875,
0.3359375,
0.5693359375,
0.2032470703125,
0.5693359375
],
[
0.202392578125,
0.568359375,
0.355224609375,
0.57080078125,
0.355224609375,
0.58837890625,
0.202392578125,
0.5859375
],
[
0.51123046875,
0.5654296875,
0.880859375,
0.56640625,
0.880859375,
0.58056640625,
0.51123046875,
0.580078125
],
[
0.118408203125,
0.57177734375,
0.1427001953125,
0.57177734375,
0.1427001953125,
0.5849609375,
0.118408203125,
0.5849609375
],
[
0.175537109375,
0.5859375,
0.225830078125,
0.5859375,
0.225830078125,
0.60107421875,
0.175537109375,
0.60107421875
],
[
0.51318359375,
0.5830078125,
0.880859375,
0.5830078125,
0.880859375,
0.5966796875,
0.51318359375,
0.5966796875
],
[
0.51318359375,
0.5986328125,
0.87890625,
0.5986328125,
0.87890625,
0.6123046875,
0.51318359375,
0.6123046875
],
[
0.118408203125,
0.58837890625,
0.1427001953125,
0.58837890625,
0.1427001953125,
0.60107421875,
0.118408203125,
0.60107421875
],
[
0.118408203125,
0.60400390625,
0.2493896484375,
0.6025390625,
0.2493896484375,
0.61767578125,
0.118408203125,
0.6181640625
],
[
0.51318359375,
0.615234375,
0.7900390625,
0.615234375,
0.7900390625,
0.62890625,
0.51318359375,
0.62890625
],
[
0.52880859375,
0.6298828125,
0.87890625,
0.6298828125,
0.87890625,
0.6435546875,
0.52880859375,
0.6435546875
],
[
0.1175537109375,
0.64794921875,
0.486083984375,
0.64794921875,
0.486083984375,
0.6611328125,
0.1175537109375,
0.6611328125
],
[
0.51123046875,
0.646484375,
0.880859375,
0.646484375,
0.880859375,
0.6611328125,
0.51123046875,
0.6611328125
],
[
0.1175537109375,
0.66455078125,
0.175537109375,
0.66455078125,
0.175537109375,
0.6787109375,
0.1175537109375,
0.6787109375
],
[
0.51123046875,
0.662109375,
0.8818359375,
0.6630859375,
0.8818359375,
0.677734375,
0.51123046875,
0.6767578125
],
[
0.1368408203125,
0.681640625,
0.48779296875,
0.681640625,
0.48779296875,
0.6953125,
0.1368408203125,
0.6953125
],
[
0.509765625,
0.6767578125,
0.611328125,
0.6787109375,
0.611328125,
0.693359375,
0.509765625,
0.69140625
],
[
0.118408203125,
0.69677734375,
0.486083984375,
0.69677734375,
0.486083984375,
0.71044921875,
0.118408203125,
0.71044921875
],
[
0.52880859375,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.52880859375,
0.708984375
],
[
0.1175537109375,
0.71435546875,
0.32666015625,
0.71435546875,
0.32666015625,
0.72802734375,
0.1175537109375,
0.72802734375
],
[
0.51318359375,
0.71142578125,
0.880859375,
0.71142578125,
0.880859375,
0.724609375,
0.51318359375,
0.724609375
],
[
0.5146484375,
0.72802734375,
0.880859375,
0.72802734375,
0.880859375,
0.7412109375,
0.5146484375,
0.7412109375
],
[
0.130126953125,
0.74462890625,
0.48779296875,
0.74462890625,
0.48779296875,
0.75830078125,
0.130126953125,
0.75830078125
],
[
0.51318359375,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.15283203125,
0.7626953125,
0.27197265625,
0.7626953125,
0.27197265625,
0.775390625,
0.15283203125,
0.775390625
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.5146484375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.5146484375,
0.7890625
],
[
0.130126953125,
0.79345703125,
0.48779296875,
0.79345703125,
0.48779296875,
0.80712890625,
0.130126953125,
0.80712890625
],
[
0.51318359375,
0.7919921875,
0.7060546875,
0.7919921875,
0.7060546875,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.15283203125,
0.80859375,
0.48779296875,
0.80859375,
0.48779296875,
0.822265625,
0.15283203125,
0.822265625
],
[
0.52880859375,
0.80810546875,
0.87890625,
0.80810546875,
0.87890625,
0.82177734375,
0.52880859375,
0.82177734375
],
[
0.15283203125,
0.822265625,
0.2376708984375,
0.826171875,
0.2359619140625,
0.84033203125,
0.151123046875,
0.83740234375
],
[
0.5146484375,
0.82470703125,
0.87890625,
0.82470703125,
0.87890625,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.5146484375,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.5146484375,
0.85400390625
],
[
0.1351318359375,
0.85498046875,
0.48779296875,
0.85546875,
0.48779296875,
0.87060546875,
0.1351318359375,
0.869140625
],
[
0.51318359375,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91943359375,
0.118408203125,
0.91943359375
],
[
0.5146484375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.5146484375,
0.91845703125
]
] |
[
"Algorithm 1 Lazy-k Decoding",
"Table 2:Datasets used in the experiments.APL=Aver",
"age Page Length, CSR = Constraint Satisfaction Ratio",
"Require: Input sequence: x",
"Require: Cost function g from Eq. 4",
"APL",
"CSR",
"Pages",
"Dataset",
"Require: Binary constraint C, max iterations k",
"CORD",
"1,000",
"46",
"72%",
"1: function LAZY-K(x, g,C, k)",
"WildReceipt",
"147",
"70 %",
"yl ← argminyeyn g(y)",
"1,740",
"2:",
"7,372",
"353",
"88 %",
"if c(x,yl) = 1 then return yl",
"Docile",
"3:",
"4:",
"H ← MinHeapO",
"frontier ← (yl : 1}",
"logI-,i=logk!，使用斯特林近似，它等效",
"于0 (klog k)",
"5:",
"6:",
"AddNextBest (yl, H, frontier )",
"7:",
"count ←1",
"NextBest(y,frontier）函数返回y的1个编辑",
"while H not empty and count <k do",
"距离内的第[y]次最佳状态。首次扩展状态时，",
"我们计算○中的下一个最佳编辑的排序列表",
"8:",
"9:",
"yi ← H .PopMinO",
"yk ← NextBest (y', frontier)",
"（nlogn）时间。使用此方法，可以在恒定时间",
"if c(x, yk) -1 then return yk",
"10:",
"11:",
"内计算此状态的每个NextBest调用。对于每个",
"AddNextBest (yk, H, frontier)",
"状态，最多调用NextBestOn次。由于排序需",
"AddNextBest (y',H, frontier)",
"12:",
"13:",
"要nlogn时间，因此算法的总时间复杂度变",
"count += 1",
"为: O (k(logk+nlogn))",
"end while",
"14:",
"15:",
"5Experiments",
"16:",
"return Failure",
"17: end function",
"为了评估Lazy-k解码器的相关性，我们对多",
"18:",
"个数据集执行发票信息提取。该任务的目的是",
"19: function ADDNEXTBEST(y',H,frontier)",
"从发票中提取各种金额，使其满足预期的算术",
"20:",
"yij NextBest (yi, frontier)",
"结构。对于每个数据集，我们训练一个令牌分",
"21:",
"while yii null and yij e frontier do",
"类模型并为测试集生成预测。然后将预测与约",
"22:",
"frontier [yi] += 1",
"束一起馈送到不同的解码算法中，并返回满足",
"23:",
"yij←NextBest (y, frontier)",
"约束的最高概率序列",
"24:",
"end while",
"if yij + null then",
"Data我们在Tab中显示的总共三个数据集",
"上评估解码器。2:CORD(Parketal.2019）、",
"25:",
"26:",
"frontier[ yi] ←1",
"H .Add( y,g(y'))",
"WildReceipt (Sun et al., 2021)和 DocILE (Simsa",
"27:",
"end if",
"etal.,2023）。虽然约束在语义上非常相似，但",
"数据集的标签和粒度级别略有不同，这意味着",
"28:",
"29:end function",
"每个数据集都有自己特定的约束集",
"每个数据集的确切约束在App中有详细说",
"最佳状态，我们最多添加1个大小为n的新状",
"明。C。所有数据集的模型都使用BIO标签",
"态到堆",
"进行训练，其初始约束是标签分配应该是有效",
"时间复杂度稍微不那么明显。对于top-k状",
"的BIO序列",
"态，外部while循环将运行k迭代。在这个循",
"其他约束取决于每个数据集中可用的特定标",
"环中，有两个复杂性来源：",
"签。然而，有些标签可能不存在于每个样品中。",
"因此，我们区分必填字段（即总金额）和可选",
"1. H .AddO which occurs at most twice in",
"字段（即服务费，折扣），如果找不到，则将其",
"AddNextBestO,",
"视为0。如果预测中不存在必填字段将被视为",
"空。因此，涉及此字段的任何约束都不会被评",
"2. NextBestO which occurs once in the outer",
"估(或自动视为已满足)",
"loop and twice in AddNextBest in another",
"对于每个数据集，我们将约束应用于所有样",
"while loop.",
"本，并过滤掉任何不满足约束的样本（在表中",
"显示计数）。从这些示例中，我们使用60个%",
"H。addO操作向堆添加一个元素，该元素",
"进行训练和验证（拆分80-20），使用40个%",
"相对于堆的大小具有对数复杂度。由于堆在",
"进行测试。将不满足约束的样本添加到训练集",
"每次选代时都完全保存了我们的顶级k状",
"中。我们有目的地选择一个大百分比的测试，",
"态，因此该操作的复杂性等于-ilogi=",
"因为小火车集提供了足够的性能，我们大多希"
] |
{
"arxiv_id": "2312.03367",
"language": "zh",
"page_idx": 4
}
|
||
1801.07455-zh-1
|
1801.07455-zh-1.jpg
|
Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action
# $\operatorname{Recognition}$
Sijie Yan, Yuanjun Xiong, Dahua Lin
Department of Information Engineering, The Chinese University of Hong Kong
$\{$ ys016,dhlin $\}$ @ie.cuhk.edu.hk, bitxiong@gmail.com
$\operatorname{Abstract}$
Dynamics of human body skeletons convey significant information for human action recognition. Conventional approaches for modeling skeletons usually rely on hand-crafted parts or traversal rules, thus resulting in limited expressive power and difficulties of generalization. In this work, we propose a novel model of dynamic skeletons called Spatial- Temporal Graph Con- volutional Networks (ST-\hat{G}CN) , which moves beyond the limitations of previous methods by automatically learning both the spatial and temporal patterns from data. This formulation not only leads to greater expressive power but also stronger generalization capability. On two large datasets, Kinetics and NTU-RGBD , it achieves substantial improvements over mainstream methods.
# 1 Introduction
人体动作识别近年来已成为一个活跃的研究领域，因为它在视频理解中起着重要作用。通常，可以从多个模态 (Simonyan and Zisserman 2014; Tran et al. 2015; Wang, Qiao, and Tang 2015; Wang et al. 2016; Zhao et al.20$\underline{17})识别人类动作，例如外观，深度，光流和身体$ 骨骼$\mathbb{E}($Du, Wang, and Wang 2015; Liu et al. 2016)。在这些模式中，动态的人类骨骼通常传达重要的信息， 这些信息是对其他人的补充。。然而，与外观和光流相比，动态骨架的建模受到的关注相对较少。在这项工作中，我们系统地研究了这种模式，目的是开发一种原则性和有效的方法来建模动态骨架并利用它们进行动作识别
动态骨架模态可以自然地由人类关节位置的时间序列以 2D 或 3D 坐标的形式表示。然后可以通过分析其运动模式来识别人类动作。早期使用骨架进行动作识别的方法只是在各个时间步骤中使用关节坐标来形成特征向量，并在其上应用时间分析$\boxed{\mathrm{E}}($Wang et al. 2012; Fernando et al. 2015)。这些方法的能力有限，因为它们没有明确利用关节之间的空间关系，这对于理解人类行为至关重要。最近，已经开发出尝试利用关节之间自然连接的新方法$\mathbb{E}($Shahroudy et al. 2016; Du, Wang, and Wang 2015) 。这些方法显示出令人鼓舞的改进，这表明连通性的重要性。然而，大多数现有的方法依赖于
Copyright $\odot2018$, Association for the Advancement of Ar-
tificial Intelligence (www.aaai.org). All rights reserved.
<FigureHere>
Figure 1: The spatial temporal graph of a skeleton sequence used in this work where the proposed ST-GCN operate on. Blue dots denote the body joints. The intra-body edges between body joints are defined based on the natural connections in human bodies. The inter-frame edges connect the same joints between consecutive frames. Joint coordinates are used as inputs to the ST-GCN.
手工制作的零件或规则来分析空间模式。因此，为特定
应用设计的模型很难推广到其他应用
要超越此类限制，我们需要一种新方法，该方法可以自动捕获嵌人在关节的空间配置中的模式以及它们的时间动态。这就是深度神经网络的力量。然而，如上所述，骨架是图形的形式，而不是 2D 或 3D 网格，这使得难以使用卷积网络等经过验证的模型。最近，将卷积神经网络(CNN)推广到任意结构的图的图神经网络(GCNs)受到了越来越多的关注，并成功地应用于许多应用中，例如图像分类$\mathbb{E}($Bruna et al.2014),文档分类$\boxed{\mathrm{E}}($Defferrard, Bresson, and Vandergheynst 2016) 和半监督学习$\operatorname{E}($Kipf and Welling 2017)。然而，沿着这条线的许多先前的工作都假设一个固定的图形作为输人。GCNs 在大规模数据集上动态图形建模中的应用， 例如。g。人类骨骼序列，尚待探索
在本文中，我们建议通过将图神经网络扩展到时空图模型来设计用于动作识别的骨架序列的通用表示，称为时空图卷积网络 (ST-GCN)。如图 1 所示，该模型是在一系列骨架图的顶部制定的，其中每个节点对应于人体的一个关节。有两种类型的边缘，即符合关节自然连通性的空间边缘和在连续时间步长上连接相同关节的时间边缘。在其上构造了多层时空图卷积，从而可以沿
|
[
[
0.12091064453125,
0.125,
0.8798828125,
0.125,
0.8798828125,
0.1395263671875,
0.12091064453125,
0.1395263671875
],
[
0.441162109375,
0.1458740234375,
0.5595703125,
0.1458740234375,
0.5595703125,
0.160400390625,
0.441162109375,
0.160400390625
],
[
0.34228515625,
0.1875,
0.65673828125,
0.1875,
0.65673828125,
0.2020263671875,
0.34228515625,
0.2020263671875
],
[
0.2156982421875,
0.2020263671875,
0.78173828125,
0.2020263671875,
0.78173828125,
0.216552734375,
0.2156982421875,
0.216552734375
],
[
0.3056640625,
0.216552734375,
0.6953125,
0.216552734375,
0.6953125,
0.2298583984375,
0.3056640625,
0.2298583984375
],
[
0.25244140625,
0.27587890625,
0.313720703125,
0.27587890625,
0.313720703125,
0.287353515625,
0.25244140625,
0.287353515625
],
[
0.10296630859375,
0.298583984375,
0.461669921875,
0.298583984375,
0.461669921875,
0.3125,
0.10296630859375,
0.3125
],
[
0.10296630859375,
0.311279296875,
0.461669921875,
0.311279296875,
0.461669921875,
0.324462890625,
0.10296630859375,
0.324462890625
],
[
0.10455322265625,
0.324462890625,
0.459228515625,
0.324462890625,
0.459228515625,
0.335205078125,
0.10455322265625,
0.335205078125
],
[
0.10455322265625,
0.33837890625,
0.459228515625,
0.33837890625,
0.459228515625,
0.348388671875,
0.10455322265625,
0.348388671875
],
[
0.10455322265625,
0.35107421875,
0.459228515625,
0.35107421875,
0.459228515625,
0.361083984375,
0.10455322265625,
0.361083984375
],
[
0.10296630859375,
0.3623046875,
0.463134765625,
0.3623046875,
0.463134765625,
0.375732421875,
0.10296630859375,
0.375732421875
],
[
0.10455322265625,
0.375732421875,
0.459228515625,
0.375732421875,
0.459228515625,
0.386474609375,
0.10455322265625,
0.386474609375
],
[
0.10455322265625,
0.38818359375,
0.461669921875,
0.38818359375,
0.461669921875,
0.3984375,
0.10455322265625,
0.3984375
],
[
0.10296630859375,
0.399658203125,
0.461669921875,
0.399658203125,
0.461669921875,
0.41357421875,
0.10296630859375,
0.41357421875
],
[
0.10455322265625,
0.4140625,
0.4599609375,
0.4140625,
0.4599609375,
0.4248046875,
0.10455322265625,
0.4248046875
],
[
0.10455322265625,
0.4267578125,
0.45751953125,
0.4267578125,
0.45751953125,
0.4375,
0.10455322265625,
0.4375
],
[
0.101318359375,
0.4375,
0.459228515625,
0.4375,
0.459228515625,
0.45068359375,
0.101318359375,
0.45068359375
],
[
0.10296630859375,
0.449462890625,
0.459228515625,
0.449462890625,
0.459228515625,
0.46337890625,
0.10296630859375,
0.46337890625
],
[
0.10455322265625,
0.464111328125,
0.4599609375,
0.464111328125,
0.4599609375,
0.474853515625,
0.10455322265625,
0.474853515625
],
[
0.521484375,
0.46337890625,
0.9111328125,
0.46337890625,
0.9111328125,
0.4765625,
0.521484375,
0.4765625
],
[
0.10455322265625,
0.4765625,
0.1641845703125,
0.4765625,
0.1641845703125,
0.4873046875,
0.10455322265625,
0.4873046875
],
[
0.521484375,
0.47607421875,
0.91015625,
0.47607421875,
0.91015625,
0.4892578125,
0.521484375,
0.4892578125
],
[
0.521484375,
0.488037109375,
0.91015625,
0.488037109375,
0.91015625,
0.501953125,
0.521484375,
0.501953125
],
[
0.521484375,
0.50048828125,
0.90869140625,
0.50048828125,
0.90869140625,
0.5146484375,
0.521484375,
0.5146484375
],
[
0.20751953125,
0.51025390625,
0.3544921875,
0.51123046875,
0.3544921875,
0.52587890625,
0.20751953125,
0.5244140625
],
[
0.521484375,
0.5146484375,
0.91259765625,
0.5146484375,
0.91259765625,
0.52783203125,
0.521484375,
0.52783203125
],
[
0.521484375,
0.52783203125,
0.9111328125,
0.52783203125,
0.9111328125,
0.53857421875,
0.521484375,
0.53857421875
],
[
0.09234619140625,
0.5341796875,
0.476318359375,
0.5341796875,
0.476318359375,
0.54443359375,
0.09234619140625,
0.54443359375
],
[
0.5185546875,
0.53857421875,
0.75341796875,
0.537109375,
0.75341796875,
0.5517578125,
0.5185546875,
0.55322265625
],
[
0.08905029296875,
0.546875,
0.47802734375,
0.546875,
0.47802734375,
0.56005859375,
0.08905029296875,
0.56005859375
],
[
0.0882568359375,
0.56005859375,
0.476318359375,
0.56005859375,
0.476318359375,
0.57373046875,
0.0882568359375,
0.57373046875
],
[
0.0882568359375,
0.57470703125,
0.47802734375,
0.57470703125,
0.47802734375,
0.58837890625,
0.0882568359375,
0.58837890625
],
[
0.52294921875,
0.58056640625,
0.9111328125,
0.58056640625,
0.9111328125,
0.5947265625,
0.52294921875,
0.5947265625
],
[
0.08660888671875,
0.5869140625,
0.47802734375,
0.58837890625,
0.47802734375,
0.60302734375,
0.08660888671875,
0.6015625
],
[
0.0882568359375,
0.6015625,
0.4697265625,
0.6015625,
0.4697265625,
0.61572265625,
0.0882568359375,
0.61572265625
],
[
0.52392578125,
0.5966796875,
0.78125,
0.5966796875,
0.78125,
0.60693359375,
0.52392578125,
0.60693359375
],
[
0.465576171875,
0.60595703125,
0.473876953125,
0.60595703125,
0.473876953125,
0.611328125,
0.465576171875,
0.611328125
],
[
0.0882568359375,
0.61572265625,
0.4755859375,
0.61572265625,
0.4755859375,
0.62890625,
0.0882568359375,
0.62890625
],
[
0.5390625,
0.611328125,
0.9111328125,
0.611328125,
0.9111328125,
0.625,
0.5390625,
0.625
],
[
0.0882568359375,
0.6298828125,
0.4794921875,
0.6298828125,
0.4794921875,
0.64306640625,
0.0882568359375,
0.64306640625
],
[
0.52294921875,
0.625,
0.9111328125,
0.625,
0.9111328125,
0.63818359375,
0.52294921875,
0.63818359375
],
[
0.08905029296875,
0.64306640625,
0.476318359375,
0.64306640625,
0.476318359375,
0.6572265625,
0.08905029296875,
0.6572265625
],
[
0.521484375,
0.63818359375,
0.9111328125,
0.63818359375,
0.9111328125,
0.6513671875,
0.521484375,
0.6513671875
],
[
0.0882568359375,
0.6572265625,
0.47802734375,
0.6572265625,
0.47802734375,
0.67041015625,
0.0882568359375,
0.67041015625
],
[
0.51953125,
0.6513671875,
0.9111328125,
0.65087890625,
0.9111328125,
0.66552734375,
0.51953125,
0.66650390625
],
[
0.08660888671875,
0.66943359375,
0.47802734375,
0.67041015625,
0.47802734375,
0.68505859375,
0.08660888671875,
0.6845703125
],
[
0.521484375,
0.66650390625,
0.91259765625,
0.66650390625,
0.91259765625,
0.6796875,
0.521484375,
0.6796875
],
[
0.52294921875,
0.6787109375,
0.9111328125,
0.6787109375,
0.9111328125,
0.6923828125,
0.52294921875,
0.6923828125
],
[
0.0882568359375,
0.68603515625,
0.136474609375,
0.68603515625,
0.136474609375,
0.69775390625,
0.0882568359375,
0.69775390625
],
[
0.521484375,
0.693359375,
0.9111328125,
0.693359375,
0.9111328125,
0.70703125,
0.521484375,
0.70703125
],
[
0.10296630859375,
0.69970703125,
0.47802734375,
0.69970703125,
0.47802734375,
0.71337890625,
0.10296630859375,
0.71337890625
],
[
0.52294921875,
0.70703125,
0.91015625,
0.70703125,
0.91015625,
0.72021484375,
0.52294921875,
0.72021484375
],
[
0.0882568359375,
0.71337890625,
0.4794921875,
0.71337890625,
0.4794921875,
0.7265625,
0.0882568359375,
0.7265625
],
[
0.0882568359375,
0.7265625,
0.47802734375,
0.7265625,
0.47802734375,
0.74072265625,
0.0882568359375,
0.74072265625
],
[
0.52294921875,
0.7216796875,
0.91259765625,
0.7216796875,
0.91259765625,
0.73486328125,
0.52294921875,
0.73486328125
],
[
0.52294921875,
0.73486328125,
0.9111328125,
0.73486328125,
0.9111328125,
0.74853515625,
0.52294921875,
0.74853515625
],
[
0.0882568359375,
0.7412109375,
0.47802734375,
0.7412109375,
0.47802734375,
0.7548828125,
0.0882568359375,
0.7548828125
],
[
0.0882568359375,
0.7548828125,
0.476318359375,
0.7548828125,
0.476318359375,
0.76806640625,
0.0882568359375,
0.76806640625
],
[
0.52294921875,
0.75,
0.91015625,
0.75,
0.91015625,
0.76318359375,
0.52294921875,
0.76318359375
],
[
0.0882568359375,
0.76806640625,
0.47802734375,
0.76806640625,
0.47802734375,
0.7822265625,
0.0882568359375,
0.7822265625
],
[
0.52392578125,
0.76318359375,
0.91259765625,
0.76318359375,
0.91259765625,
0.7763671875,
0.52392578125,
0.7763671875
],
[
0.521484375,
0.7763671875,
0.779296875,
0.7763671875,
0.779296875,
0.79150390625,
0.521484375,
0.79150390625
],
[
0.08660888671875,
0.78271484375,
0.47802734375,
0.78076171875,
0.47802734375,
0.79541015625,
0.08660888671875,
0.79736328125
],
[
0.0882568359375,
0.79736328125,
0.476318359375,
0.79736328125,
0.476318359375,
0.81103515625,
0.0882568359375,
0.81103515625
],
[
0.5361328125,
0.79248046875,
0.9111328125,
0.79248046875,
0.9111328125,
0.80615234375,
0.5361328125,
0.80615234375
],
[
0.0849609375,
0.806640625,
0.47802734375,
0.8095703125,
0.47802734375,
0.8271484375,
0.0849609375,
0.82470703125
],
[
0.521484375,
0.80615234375,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.521484375,
0.8193359375
],
[
0.08660888671875,
0.82275390625,
0.47802734375,
0.82373046875,
0.47802734375,
0.83837890625,
0.08660888671875,
0.8369140625
],
[
0.51953125,
0.818359375,
0.9111328125,
0.8193359375,
0.9111328125,
0.833984375,
0.51953125,
0.83349609375
],
[
0.0882568359375,
0.83837890625,
0.4755859375,
0.83837890625,
0.4755859375,
0.8515625,
0.0882568359375,
0.8515625
],
[
0.521484375,
0.833984375,
0.9111328125,
0.833984375,
0.9111328125,
0.84765625,
0.521484375,
0.84765625
],
[
0.52294921875,
0.84765625,
0.91259765625,
0.84765625,
0.91259765625,
0.861328125,
0.52294921875,
0.861328125
],
[
0.08660888671875,
0.86279296875,
0.4755859375,
0.86279296875,
0.4755859375,
0.876953125,
0.08660888671875,
0.876953125
],
[
0.52294921875,
0.8623046875,
0.9111328125,
0.8623046875,
0.9111328125,
0.87548828125,
0.52294921875,
0.87548828125
],
[
0.0849609375,
0.875,
0.44775390625,
0.87548828125,
0.44775390625,
0.89013671875,
0.0849609375,
0.8896484375
],
[
0.52294921875,
0.87548828125,
0.9111328125,
0.87548828125,
0.9111328125,
0.8896484375,
0.52294921875,
0.8896484375
]
] |
[
"Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action",
"Recognition",
"Sijie Yan, Yuanjun Xiong, Dahua Lin",
"Department of Information Engineering, The Chinese University of Hong Kong",
"( ys016, dhlin } @ie.cuhk.edu.hk, bitxiong@gmail.com",
"Abstract",
"Dynamics of human body skeletons convey significant",
"information for human action recognition. Conven-",
"tional approachesfor modelingskeletons usuallyrely",
"on hand-crafted parts or traversal rules, thus result-",
"ing in limited expressivepowerand difficultiesofgen-",
"eralization. In this work, we propose a novel model of",
"dynamicskeletonscalledSpatial-TemporalGraphCon-",
"volutionalNetworks (ST-GCN),whichmovesbeyond",
"the limitations of previous methods by automatically",
"learning boththe spatial and temporalpatternsfrom",
"data. This formulation not only leads to greater ex",
"pressive power but also stronger generalization capabil",
"ity. On two large datasets, Kinetics and NTU-RGBD",
"it achieves substantial improvements over mainstream",
"Figure l: The spatial temporal graph of a skeleton sequence",
"methods.",
"used in this work where the proposed ST-GCN operate on.",
"Blue dots denote the body joints. The intra-body edges be-",
"tween body joints are defined based on the natural connec",
"1Introduction",
"tions in human bodies. The inter-frame edges connect the",
"samejointsbetweenconsecutiveframes.Jointcoordinates",
"人体动作识别近年来已成为一个活跃的研究领域：因",
"are used as inputs to the ST-GCN.",
"为它在视频理解中起着重要作用。通常，可以从多个",
"模态(Simonyan and Zisserman 2014;Tran et al.2015",
"Wang, Qiao, and Tang 2015; Wang et al. 2016; Zhao et",
"手工制作的零件或规则来分析空间模式。因此，为特定",
"al.2017）识别人类动作，例如外观，深度，光流和身体",
"骨骼E(Du,Wang,and Wang 2015;Liu et al.2016)",
"应用设计的模型很难推广到其他应用",
"",
"在这些模式中，动态的人类骨骼通常传达重要的信息",
"要超越此类限制，我们需要一种新方法，该方法可以",
"这些信息是对其他人的补充。。然而，与外观和光流相",
"自动捕获嵌人在关节的空间配置中的模式以及它们的",
"比，动态骨架的建模受到的关注相对较少。。在这项工",
"时间动态。这就是深度神经网络的力量。然而，如上所",
"作中，我们系统地研究了这种模式，目的是开发一种原",
"述，骨架是图形的形式，而不是2D或3D网格，这使",
"则性和有效的方法来建模动态骨架并利用它们进行动",
"得难以使用卷积网络等经过验证的模型。最近，将卷积",
"神经网络（CNN）推广到任意结构的图的图神经网络",
"作识别",
"（GCNs）受到了越来越多的关注，并成功地应用于许多",
"动态骨架模态可以自然地由人类关节位置的时间序",
"应用中，例如图像分类E(Brunaetal.2014），文档分",
"列以2D或3D坐标的形式表示。然后可以通过分析其",
"运动模式来识别人类动作。早期使用骨架进行动作识",
"类E(Defferrard,Bresson, and Vandergheynst 2016)和",
"半监督学习E(KipfandWelling2017）。然而，沿着这",
"别的方法只是在各个时间步骤中使用关节坐标来形成",
"特征向量，并在其上应用时间分析E（Wangetal.2012：",
"条线的许多先前的工作都假设一个固定的图形作为输",
"Fernandoetal.2015）。这些方法的能力有限，因为它",
"人。GCNs在大规模数据集上动态图形建模中的应用，",
"例如。g。人类骨骼序列，尚待探索",
"们没有明确利用关节之间的空间关系，这对于理解人类",
"行为至关重要。最近，已经开发出尝试利用关节之间自",
"在本文中，我们建议通过将图神经网络扩展到时空图",
"然连接的新方法（Shahroudyetal.2016;Du，Wang，",
"模型来设计用于动作识别的骨架序列的通用表示，称为",
"andWang2015）。这些方法显示出令人鼓舞的改进，这",
"时空图卷积网络（ST-GCN）。如图1所示，该模型是",
"表明连通性的重要性。然而，大多数现有的方法依赖于",
"在一系列骨架图的顶部制定的，其中每个节点对应于人",
"体的一个关节。有两种类型的边缘，即符合关节自然连",
"Copyright 2018, Association for the Advancement of Ar-",
"通性的空间边缘和在连续时间步长上连接相同关节的",
"tificial Intelligence (www.aaai.org). All rights reserved.",
"时间边缘。在其上构造了多层时空图卷积，从而可以沿"
] |
{
"arxiv_id": "1801.07455",
"language": "zh",
"page_idx": 1
}
|
||
1801.07455-zh-2
|
1801.07455-zh-2.jpg
|
空间和时间维度集成信息
ST-GCN 的分层特性消除了手工制作的零件分配或遍历规则的需要。这不仅可以带来更大的表达能力，从而带来更高的性能(如我们的实验所示),而且还可以很容易地推广到不同的上下文中。在通用 GCN 公式的基础上，我们还研究了设计图形卷积内核的新策略，并从图像模型中获得了灵感
这项工作的主要贡献在于三个方面：1)我们提出了ST-GCN,这是一种基于图的通用公式，用于对动态骨架进行建模，这是第一个将基于图的神经网络应用于此任务的。2) 我们提出了在 ST-GCN 中设计卷积核的几个原则，以满足骨架建模的特定需求。3)在两个基于骨架的动作识别的大规模数据集上，与以前使用手工制作的零件或遍历规则的方法相比，所提出的模型具有更好的性能，而手动设计的工作量却大大减少了。ST-GCN 的代码和模型已公开$^{1}$
# 2 Related work
Neural Networks on Graphs. 将神经网络推广到具有图结构的数据是深度学习研究中的一个新兴主题。讨论的神经网络架构包括递归神经网络E(Tai, Socher, and Manning 2015; Van Oord, Kalchbrenner, and Kavukcuoglu 2016) 和卷积神经网络(CNN) $\overset{(}{\operatorname*{E}}$ (Bruna et al. 2014; Henaff, Bruna, and LeCun 2015; Duvenaud et al. 2015; Li et al. 2016; Defferrard, Bres- son, and Vandergheynst 2016) 。这项工作更多地与CNN 或图形卷积网络(GCN)的泛化有关。在图上构造GCN 的原理通常遵循两个流：1)频谱透视图，其中以频谱分析$\overline{\mathrm{E}}($Henaff, Bruna, and LeCun 2015; Duvenaud et al. 2015; Li et al. 2016; Kipf and Welling 2017) 的形式考虑图卷积的局部性；2) 空间透视图，其中卷积滤波器直接应用于图节点及其邻居$\overline{\mathrm{E}}($Bruna et al. 2014; Niepert, Ahmed, and Kutzkov 2016) 。这项工作遵循了第二个流的精神。我们通过将每个滤波器的应用限制在每个节点的 1 邻居上，在空间域上构造 CNN 滤波器
Skeleton Based Action Recognition. 人体的骨骼和关节轨迹对光照变化和场景变化具有鲁棒性，并且由于高度精确的深度传感器或姿态估计算法而易于获得E (Shotton et al. 2011; Cao et al. 2017a) 。因此，有各种各样的基于骨架的动作识别方法。这些方法可以分为基于特征的手工方法和深度学习方法。第一种方法设计了几个手工制作的特征来捕捉关节运动的动力学。这些可能是关节轨迹的协方差矩阵$\mathbb{E}($Hussein et al. 2013) , 关节的相对位置$\overline{\mathrm{E}}($Wang et al. 2012) 或身体部位之间的旋转和平移$\boxed{\mathrm{F}}($Vemulapalli, Arrate, and Chellappa 2014) 。深度学习最近的成功导致了基于深度学习的骨架建模方法的激增。这些工作一直在使用递归神经网络E (Shahroudy et al. 2016; Zhu et al. 2016; Liu et al. 2016; Zhang, Liu, and Xiao 2017) 和时态 CNN$\underline{E}($Li et al. 2017;Ke et al. 2017; Kim and Reiter 2017)以端到端的方式学习动作识别模型。在这些方法中，许多人强调了对人体各部位关节进行建模的重要性。但是这些部分通常是使用领域知识显式分配的。我们的 ST-GCN 是第一个将图形 CNN 应用于基于骨架的动作识别任务的。它与以前的方法不同，因为它可以通过利用图卷积的局部性和时间动力学来隐式地学习零件信息。通过消除对
${^{1}\mathrm{https:}//\mathrm{github.com/yysijie/st-gcn}}$
手动零件分配的需求，该模型更易于设计，并且可以有
效地学习更好的动作表示
3 Spatial Temporal Graph ConvNet
在执行活动时，人体关节在小的局部组中移动，称为“身体部位”。现有的基于骨架的动作识别方法已经验证了在建模 (Shahroudy et al.2016; Liu et al. 2016; Zhang, Liu, and Xiao 2017) 中 引 入 身 体 部 位 的 有 效 性 。我们认为，这种改进主要是由于与整个骨架相比，零件限制了“局部区域”内关节轨迹的建模，从而形成了骨架序列的分层表示。在诸如图像对象识别之类的任务中，通常通过卷积神经网络E(Krizhevsky, Sutskever, and Hinton 2012) 的固有属性来实现分层表示和局部性，而不是手动分配对象部分。它促使我们将 CNN 的吸引人的特性引入基于骨架的动作识别。这种尝试的结果果是 ST-GCN 模型。
3.1 Pipeline Overview
基于骨架的数据可以从运动捕捉设备或来自视频的姿势估计算法获得。通常数据是一个帧序列，每一帧都会有一组关节坐标。给定 2D 或 3D 坐标形式的身体关节序列，我们以关节为图节点，以人体结构和时间中的自然连通性为图边，构造一个时空图。因此，ST-GCN 的输入是图形节点上的联合坐标向量。这可以被认为是基于图像的 CNN 的模拟，其中输人由位于 2D 图像网格上的像素强度向量形成。将对输人数据应用多层时空图卷积运算，并在图上生成更高级别的特征图。然后，它将由标准 SoftMax 分类器分类到相应的动作类别。整个模型通过反向传播以端到端的方式进行训练。我们现在将介绍 ST-GCN 模型中的组件
3.2 Skeleton Graph Construction
骨骼序列通常由每帧中每个人体关节的 2D 或 3D 坐标表示。以前的工作使用卷积进行骨架动作识别 (Kim and Reiter 2017) 将所有关节的坐标向量连接起来，形成每帧一个特征向量。在我们的工作中，我们利用时空图来形成骨架序列的分层表示。特别是，我们在具有$N$关节和$T$帧的骨架序列上构造了一个无向时空图$G= ( V, E)$ ,该框架具有体内和帧间连接
在此图中，节点集$V=\{v_ti|t~=~1,\ldots,T,i~=$ $1,\ldots,N\}$包括骨架序列中的所有关节。作为 ST-GCN 的输入，节点$F(v_{ti})$上的特征向量由框架$t$上第$i$个关节的坐标向量以及估计置信度组成。我们分两步在骨架序列上构造时空图。首先，根据人体结构的连通性，一个框架内的关节与边缘连接，如图。1。然后，每个关节将连接到连续帧中的同一关节。因此，此设置中的连接自然定义，无需手动分配零件。这也使网络架构能够处理具有不同数量的关节或关节连接性的数据集。例如， 在 Kinetics 数据集上，我们使用来自 OpenPose (Cao et al. 2017b) 工具箱的 2D 姿态估计结果，该工具箱输出 18 个关节，而在 NTU-RGB+D 数据集 (Shahroudy et al.2016)上，我们使用 3D 关节跟踪结果作为输人， 产生 25 个关节。ST-GCN 可以在两种情况下运行，并提供一致的卓越性能。所构律的时空图的示例在图。1
形式上，边缘集$E$由两个子集组成，第一个子集描绘了每帧处的骨架内连接，表示为$E_S=\{v_{ti}v_{tj}|(i,j)\in$ $H\}$ ,其中$H$ 是自然连接的人体关节的集合。第二个子集包含帧间边缘，这些边缘将连续帧中的相同关节连
|
[
[
0.0882568359375,
0.0706787109375,
0.26708984375,
0.0706787109375,
0.26708984375,
0.083984375,
0.0882568359375,
0.083984375
],
[
0.525390625,
0.07135009765625,
0.90869140625,
0.07135009765625,
0.90869140625,
0.08209228515625,
0.525390625,
0.08209228515625
],
[
0.10296630859375,
0.08331298828125,
0.47802734375,
0.08331298828125,
0.47802734375,
0.09783935546875,
0.10296630859375,
0.09783935546875
],
[
0.52294921875,
0.08331298828125,
0.701171875,
0.08331298828125,
0.701171875,
0.09661865234375,
0.52294921875,
0.09661865234375
],
[
0.0882568359375,
0.09783935546875,
0.473876953125,
0.09783935546875,
0.473876953125,
0.111083984375,
0.0882568359375,
0.111083984375
],
[
0.08905029296875,
0.11236572265625,
0.476318359375,
0.11236572265625,
0.476318359375,
0.1256103515625,
0.08905029296875,
0.1256103515625
],
[
0.5537109375,
0.11236572265625,
0.8759765625,
0.11236572265625,
0.8759765625,
0.126953125,
0.5537109375,
0.126953125
],
[
0.08660888671875,
0.125,
0.476318359375,
0.1256103515625,
0.476318359375,
0.1395263671875,
0.08660888671875,
0.1383056640625
],
[
0.08905029296875,
0.1395263671875,
0.476318359375,
0.1395263671875,
0.476318359375,
0.15283203125,
0.08905029296875,
0.15283203125
],
[
0.52294921875,
0.1337890625,
0.91015625,
0.1337890625,
0.91015625,
0.14453125,
0.52294921875,
0.14453125
],
[
0.521484375,
0.14453125,
0.9111328125,
0.1458740234375,
0.9111328125,
0.160400390625,
0.521484375,
0.1590576171875
],
[
0.0882568359375,
0.15283203125,
0.268798828125,
0.15283203125,
0.268798828125,
0.1666259765625,
0.0882568359375,
0.1666259765625
],
[
0.52294921875,
0.160400390625,
0.9111328125,
0.160400390625,
0.9111328125,
0.173583984375,
0.52294921875,
0.173583984375
],
[
0.10296630859375,
0.1666259765625,
0.476318359375,
0.1666259765625,
0.476318359375,
0.179931640625,
0.10296630859375,
0.179931640625
],
[
0.08660888671875,
0.1787109375,
0.476318359375,
0.179931640625,
0.476318359375,
0.1932373046875,
0.08660888671875,
0.1925048828125
],
[
0.52294921875,
0.1749267578125,
0.91015625,
0.1749267578125,
0.91015625,
0.1881103515625,
0.52294921875,
0.1881103515625
],
[
0.52294921875,
0.1875,
0.9111328125,
0.1875,
0.9111328125,
0.2008056640625,
0.52294921875,
0.2008056640625
],
[
0.08905029296875,
0.1956787109375,
0.476318359375,
0.1956787109375,
0.476318359375,
0.205810546875,
0.08905029296875,
0.205810546875
],
[
0.52294921875,
0.2008056640625,
0.90869140625,
0.2008056640625,
0.90869140625,
0.2139892578125,
0.52294921875,
0.2139892578125
],
[
0.08905029296875,
0.2083740234375,
0.476318359375,
0.2083740234375,
0.476318359375,
0.2215576171875,
0.08905029296875,
0.2215576171875
],
[
0.08660888671875,
0.2203369140625,
0.476318359375,
0.2215576171875,
0.476318359375,
0.236083984375,
0.08660888671875,
0.23486328125
],
[
0.52294921875,
0.21533203125,
0.9111328125,
0.21533203125,
0.9111328125,
0.2291259765625,
0.52294921875,
0.2291259765625
],
[
0.52294921875,
0.2291259765625,
0.9111328125,
0.2291259765625,
0.9111328125,
0.242431640625,
0.52294921875,
0.242431640625
],
[
0.08905029296875,
0.2374267578125,
0.476318359375,
0.2374267578125,
0.476318359375,
0.2474365234375,
0.08905029296875,
0.2474365234375
],
[
0.521484375,
0.242431640625,
0.91259765625,
0.242431640625,
0.91259765625,
0.255615234375,
0.521484375,
0.255615234375
],
[
0.08905029296875,
0.25,
0.476318359375,
0.25,
0.476318359375,
0.26318359375,
0.08905029296875,
0.26318359375
],
[
0.521484375,
0.2568359375,
0.9111328125,
0.2568359375,
0.9111328125,
0.270751953125,
0.521484375,
0.270751953125
],
[
0.08660888671875,
0.26318359375,
0.47802734375,
0.261962890625,
0.47802734375,
0.276611328125,
0.08660888671875,
0.27783203125
],
[
0.08660888671875,
0.276611328125,
0.2459716796875,
0.27587890625,
0.2459716796875,
0.290283203125,
0.08660888671875,
0.291748046875
],
[
0.521484375,
0.270751953125,
0.91259765625,
0.270751953125,
0.91259765625,
0.2841796875,
0.521484375,
0.2841796875
],
[
0.52294921875,
0.2841796875,
0.66845703125,
0.2841796875,
0.66845703125,
0.29736328125,
0.52294921875,
0.29736328125
],
[
0.205078125,
0.302978515625,
0.35693359375,
0.3037109375,
0.35693359375,
0.318115234375,
0.205078125,
0.317626953125
],
[
0.51953125,
0.31005859375,
0.7099609375,
0.31005859375,
0.7099609375,
0.323974609375,
0.51953125,
0.323974609375
],
[
0.096435546875,
0.322509765625,
0.3349609375,
0.322509765625,
0.3349609375,
0.3359375,
0.096435546875,
0.3359375
],
[
0.326904296875,
0.3212890625,
0.476318359375,
0.3212890625,
0.476318359375,
0.335205078125,
0.326904296875,
0.335205078125
],
[
0.08905029296875,
0.3359375,
0.476318359375,
0.3359375,
0.476318359375,
0.35107421875,
0.08905029296875,
0.35107421875
],
[
0.52392578125,
0.33203125,
0.91015625,
0.33203125,
0.91015625,
0.34228515625,
0.52392578125,
0.34228515625
],
[
0.521484375,
0.343505859375,
0.9111328125,
0.343505859375,
0.9111328125,
0.356689453125,
0.521484375,
0.356689453125
],
[
0.0882568359375,
0.35107421875,
0.476318359375,
0.35107421875,
0.476318359375,
0.3642578125,
0.0882568359375,
0.3642578125
],
[
0.521484375,
0.35791015625,
0.9111328125,
0.35791015625,
0.9111328125,
0.371826171875,
0.521484375,
0.371826171875
],
[
0.08660888671875,
0.3642578125,
0.4755859375,
0.3642578125,
0.4755859375,
0.37744140625,
0.08660888671875,
0.37744140625
],
[
0.521484375,
0.37060546875,
0.91015625,
0.371826171875,
0.91015625,
0.386474609375,
0.521484375,
0.385009765625
],
[
0.08660888671875,
0.37744140625,
0.47802734375,
0.376953125,
0.47802734375,
0.391357421875,
0.08660888671875,
0.392578125
],
[
0.52294921875,
0.385009765625,
0.9111328125,
0.385009765625,
0.9111328125,
0.3984375,
0.52294921875,
0.3984375
],
[
0.08905029296875,
0.392578125,
0.476318359375,
0.392578125,
0.476318359375,
0.406005859375,
0.08905029296875,
0.406005859375
],
[
0.521484375,
0.399658203125,
0.9111328125,
0.399658203125,
0.9111328125,
0.41357421875,
0.521484375,
0.41357421875
],
[
0.08905029296875,
0.4072265625,
0.4755859375,
0.4072265625,
0.4755859375,
0.417236328125,
0.08905029296875,
0.417236328125
],
[
0.0882568359375,
0.419189453125,
0.4755859375,
0.419189453125,
0.4755859375,
0.43310546875,
0.0882568359375,
0.43310546875
],
[
0.52392578125,
0.41357421875,
0.9111328125,
0.41357421875,
0.9111328125,
0.4267578125,
0.52392578125,
0.4267578125
],
[
0.08660888671875,
0.431884765625,
0.476318359375,
0.43310546875,
0.476318359375,
0.4462890625,
0.08660888671875,
0.44580078125
],
[
0.525390625,
0.4267578125,
0.9111328125,
0.4267578125,
0.9111328125,
0.43994140625,
0.525390625,
0.43994140625
],
[
0.08660888671875,
0.44580078125,
0.476318359375,
0.4462890625,
0.476318359375,
0.4609375,
0.08660888671875,
0.460205078125
],
[
0.521484375,
0.43994140625,
0.9111328125,
0.44140625,
0.9111328125,
0.455078125,
0.521484375,
0.453857421875
],
[
0.52294921875,
0.455078125,
0.89208984375,
0.455078125,
0.89208984375,
0.468505859375,
0.52294921875,
0.468505859375
],
[
0.88818359375,
0.458984375,
0.896484375,
0.458984375,
0.896484375,
0.464111328125,
0.88818359375,
0.464111328125
],
[
0.89208984375,
0.458984375,
0.90625,
0.455078125,
0.9111328125,
0.464111328125,
0.89794921875,
0.468505859375
],
[
0.52294921875,
0.468505859375,
0.91259765625,
0.468505859375,
0.91259765625,
0.481689453125,
0.52294921875,
0.481689453125
],
[
0.0882568359375,
0.4609375,
0.476318359375,
0.4609375,
0.476318359375,
0.474853515625,
0.0882568359375,
0.474853515625
],
[
0.08660888671875,
0.474853515625,
0.47802734375,
0.474853515625,
0.47802734375,
0.488037109375,
0.08660888671875,
0.488037109375
],
[
0.51953125,
0.481689453125,
0.7587890625,
0.48095703125,
0.7587890625,
0.49560546875,
0.51953125,
0.496826171875
],
[
0.0882568359375,
0.4892578125,
0.476318359375,
0.4892578125,
0.476318359375,
0.50244140625,
0.0882568359375,
0.50244140625
],
[
0.0882568359375,
0.50244140625,
0.476318359375,
0.50244140625,
0.476318359375,
0.5166015625,
0.0882568359375,
0.5166015625
],
[
0.521484375,
0.5087890625,
0.79736328125,
0.5087890625,
0.79736328125,
0.52294921875,
0.521484375,
0.52294921875
],
[
0.0882568359375,
0.51806640625,
0.47216796875,
0.51806640625,
0.47216796875,
0.52880859375,
0.0882568359375,
0.52880859375
],
[
0.0882568359375,
0.53076171875,
0.476318359375,
0.53076171875,
0.476318359375,
0.54443359375,
0.0882568359375,
0.54443359375
],
[
0.52392578125,
0.52880859375,
0.91259765625,
0.52880859375,
0.91259765625,
0.54248046875,
0.52392578125,
0.54248046875
],
[
0.0882568359375,
0.54443359375,
0.463134765625,
0.54443359375,
0.463134765625,
0.55810546875,
0.0882568359375,
0.55810546875
],
[
0.52294921875,
0.54248046875,
0.91259765625,
0.54248046875,
0.91259765625,
0.55615234375,
0.52294921875,
0.55615234375
],
[
0.521484375,
0.556640625,
0.9111328125,
0.556640625,
0.9111328125,
0.57080078125,
0.521484375,
0.57080078125
],
[
0.52294921875,
0.57080078125,
0.9111328125,
0.57080078125,
0.9111328125,
0.583984375,
0.52294921875,
0.583984375
],
[
0.0931396484375,
0.5771484375,
0.358642578125,
0.5771484375,
0.358642578125,
0.5869140625,
0.0931396484375,
0.5869140625
],
[
0.374267578125,
0.5771484375,
0.4755859375,
0.5771484375,
0.4755859375,
0.5869140625,
0.374267578125,
0.5869140625
],
[
0.08905029296875,
0.5888671875,
0.476318359375,
0.5888671875,
0.476318359375,
0.60302734375,
0.08905029296875,
0.60302734375
],
[
0.52392578125,
0.58544921875,
0.90869140625,
0.58544921875,
0.90869140625,
0.59521484375,
0.52392578125,
0.59521484375
],
[
0.521484375,
0.59765625,
0.9111328125,
0.59765625,
0.9111328125,
0.611328125,
0.521484375,
0.611328125
],
[
0.08905029296875,
0.60400390625,
0.47802734375,
0.60400390625,
0.47802734375,
0.6171875,
0.08905029296875,
0.6171875
],
[
0.521484375,
0.611328125,
0.81787109375,
0.611328125,
0.81787109375,
0.625,
0.521484375,
0.625
],
[
0.08905029296875,
0.6171875,
0.476318359375,
0.6171875,
0.476318359375,
0.630859375,
0.08905029296875,
0.630859375
],
[
0.08905029296875,
0.6298828125,
0.473876953125,
0.6298828125,
0.473876953125,
0.64306640625,
0.08905029296875,
0.64306640625
],
[
0.5361328125,
0.625,
0.89208984375,
0.625,
0.89208984375,
0.63818359375,
0.5361328125,
0.63818359375
],
[
0.521484375,
0.6396484375,
0.9111328125,
0.63818359375,
0.9111328125,
0.65283203125,
0.521484375,
0.65380859375
],
[
0.08905029296875,
0.646484375,
0.4755859375,
0.646484375,
0.4755859375,
0.6572265625,
0.08905029296875,
0.6572265625
],
[
0.0906982421875,
0.6591796875,
0.47802734375,
0.6591796875,
0.47802734375,
0.67236328125,
0.0906982421875,
0.67236328125
],
[
0.521484375,
0.65380859375,
0.9111328125,
0.65380859375,
0.9111328125,
0.66748046875,
0.521484375,
0.66748046875
],
[
0.0882568359375,
0.67041015625,
0.476318359375,
0.671875,
0.476318359375,
0.68603515625,
0.0882568359375,
0.68505859375
],
[
0.52294921875,
0.66650390625,
0.9111328125,
0.66650390625,
0.9111328125,
0.6796875,
0.52294921875,
0.6796875
],
[
0.51953125,
0.6796875,
0.88134765625,
0.68115234375,
0.88134765625,
0.69580078125,
0.51953125,
0.6943359375
],
[
0.08905029296875,
0.68603515625,
0.476318359375,
0.68603515625,
0.476318359375,
0.70068359375,
0.08905029296875,
0.70068359375
],
[
0.0882568359375,
0.69873046875,
0.4755859375,
0.69873046875,
0.4755859375,
0.7119140625,
0.0882568359375,
0.7119140625
],
[
0.52294921875,
0.6943359375,
0.9111328125,
0.6943359375,
0.9111328125,
0.708984375,
0.52294921875,
0.708984375
],
[
0.10455322265625,
0.71533203125,
0.4755859375,
0.71533203125,
0.4755859375,
0.72607421875,
0.10455322265625,
0.72607421875
],
[
0.52392578125,
0.71044921875,
0.91015625,
0.71044921875,
0.91015625,
0.72021484375,
0.52392578125,
0.72021484375
],
[
0.08905029296875,
0.72802734375,
0.476318359375,
0.72802734375,
0.476318359375,
0.7412109375,
0.08905029296875,
0.7412109375
],
[
0.52392578125,
0.72265625,
0.9111328125,
0.72265625,
0.9111328125,
0.736328125,
0.52392578125,
0.736328125
],
[
0.0906982421875,
0.7412109375,
0.476318359375,
0.7412109375,
0.476318359375,
0.7548828125,
0.0906982421875,
0.7548828125
],
[
0.52294921875,
0.736328125,
0.91259765625,
0.736328125,
0.91259765625,
0.75,
0.52294921875,
0.75
],
[
0.08660888671875,
0.7548828125,
0.4755859375,
0.75390625,
0.4755859375,
0.76806640625,
0.08660888671875,
0.76953125
],
[
0.52294921875,
0.751953125,
0.9111328125,
0.751953125,
0.9111328125,
0.76220703125,
0.52294921875,
0.76220703125
],
[
0.08660888671875,
0.76806640625,
0.476318359375,
0.76953125,
0.476318359375,
0.78271484375,
0.08660888671875,
0.7822265625
],
[
0.521484375,
0.7646484375,
0.9111328125,
0.7646484375,
0.9111328125,
0.77783203125,
0.521484375,
0.77783203125
],
[
0.0906982421875,
0.78271484375,
0.473876953125,
0.78271484375,
0.473876953125,
0.796875,
0.0906982421875,
0.796875
],
[
0.521484375,
0.77783203125,
0.9111328125,
0.77783203125,
0.9111328125,
0.79150390625,
0.521484375,
0.79150390625
],
[
0.51953125,
0.79150390625,
0.9111328125,
0.79248046875,
0.9111328125,
0.80615234375,
0.51953125,
0.8046875
],
[
0.08905029296875,
0.798828125,
0.4755859375,
0.798828125,
0.4755859375,
0.8095703125,
0.08905029296875,
0.8095703125
],
[
0.521484375,
0.8037109375,
0.9111328125,
0.8046875,
0.9111328125,
0.8193359375,
0.521484375,
0.818359375
],
[
0.0882568359375,
0.81103515625,
0.476318359375,
0.81103515625,
0.476318359375,
0.82470703125,
0.0882568359375,
0.82470703125
],
[
0.101318359375,
0.82373046875,
0.4755859375,
0.82470703125,
0.4755859375,
0.83837890625,
0.101318359375,
0.8369140625
],
[
0.521484375,
0.8193359375,
0.89208984375,
0.8193359375,
0.89208984375,
0.83349609375,
0.521484375,
0.83349609375
],
[
0.89794921875,
0.82470703125,
0.90869140625,
0.82470703125,
0.90869140625,
0.8291015625,
0.89794921875,
0.8291015625
],
[
0.08660888671875,
0.8369140625,
0.47802734375,
0.83837890625,
0.47802734375,
0.85302734375,
0.08660888671875,
0.8515625
],
[
0.5361328125,
0.833984375,
0.9111328125,
0.833984375,
0.9111328125,
0.84765625,
0.5361328125,
0.84765625
],
[
0.51953125,
0.8466796875,
0.9111328125,
0.84765625,
0.9111328125,
0.8623046875,
0.51953125,
0.861328125
],
[
0.0882568359375,
0.85302734375,
0.476318359375,
0.85302734375,
0.476318359375,
0.8662109375,
0.0882568359375,
0.8662109375
],
[
0.52294921875,
0.861328125,
0.91015625,
0.861328125,
0.91015625,
0.87548828125,
0.52294921875,
0.87548828125
],
[
0.1070556640625,
0.87353515625,
0.339111328125,
0.87548828125,
0.339111328125,
0.89013671875,
0.1070556640625,
0.88818359375
],
[
0.52392578125,
0.87548828125,
0.9111328125,
0.87548828125,
0.9111328125,
0.8896484375,
0.52392578125,
0.8896484375
]
] |
[
"空间和时间维度集成信息",
"手动零件分配的需求，该模型更易于设计，并且可以有",
"ST-GCN的分层特性消除了手工制作的零件分配或",
"效地学习更好的动作表示",
"遍历规则的需要。这不仅可以带来更大的表达能力，从",
"而带来更高的性能（如我们的实验所示），而且还可以",
"3Spatial Temporal Graph ConvNet",
"很容易地推广到不同的上下文中。在通用GCN公式的",
"基础上，我们还研究了设计图形卷积内核的新策略，并",
"在执行活动时，人体关节在小的局部组中移动，称为",
"“身体部位”。现有的基于骨架的动作识别方法已经验",
"从图像模型中获得了灵感",
"证了在建模(Shahroudy et al.2016;Liu et al.2016:",
"这项工作的主要贡献在于三个方面：1）我们提出了",
"ST-GCN，这是一种基于图的通用公式，用于对动态骨",
"Zhang，Liu，andXiao2017）中引人身体部位的有效性",
"我们认为，这种改进主要是由于与整个骨架相比，零",
"架进行建模，这是第一个将基于图的神经网络应用于此",
"件限制了“局部区域”内关节轨迹的建模，从而形成了",
"任务的。2）我们提出了在ST-GCN中设计卷积核的几",
"个原则，以满足骨架建模的特定需求。3）在两个基于骨",
"骨架序列的分层表示。在诸如图像对象识别之类的任",
"务中，通常通过卷积神经网络E(Krizhevsky，Sutskever，",
"架的动作识别的大规模数据集，与以前使用手工制作",
"andHinton2012）的固有属性来实现分层表示和局部",
"的零件或遍历规则的方法相比，所提出的模型具有更好",
"性，而不是手动分配对象部分。它促使我们将CNN的",
"的性能，而手动设计的工作量却大大减少了。ST-GCN",
"的代码和模型已公开1",
"吸引人的特性引人基于骨架的动作识别。这种尝试的结",
"果是 ST-GCN模型。",
"2Related work",
"3.1Pipeline Overview",
"Neural Networks on Graphs.",
"将神经网络推广到",
"具有图结构的数据是深度学习研究中的一个新兴主",
"基于骨架的数据可以从运动捕捉设备或来自视频的姿",
"势估计算法获得。通常数据是一个顿序列，每一顿都会",
"题。讨论的神经网络架构包括递归神经网络E(Tai",
"有一组关节坐标。给定2D或3D坐标形式的身体关节",
"Socher, and Manning 2015; Van Oord, Kalchbrenner",
"序列，我们以关节为图节点，以人体结构和时间中的自",
"andKavukcuoglu2016）和卷积神经网络（CNN）E",
"然连通性为图边，构造一个时空图。因此，ST-GCN的",
"(Bruna et al. 2014; Henaff, Bruna, and LeCun 2015",
"输人是图形节点上的联合坐标向量。这可以被认为是基",
"Duvenaud et al. 2015:Li et al.2016:Defferrard,Bres-",
"son，andVandergheynst2016）。这项工作更多地与",
"于图像的CNN的模拟，其中输人由位于2D图像网格",
"CNN或图形卷积网络（GCN）的泛化有关。在图上构造",
"上的像素强度向量形成。将对输人数据应用多层时空图",
"GCN的原理通常遵循两个流：1）频谱透视图，其中以",
"卷积运算，并在图上生成更高级别的特征图。然后，它",
"将由标准SoftMax分类器分类到相应的动作类别。",
"",
"泰",
"个模型通过反向传播以端到端的方式进行训练。我们现",
"频谱分析E(Henaff,Bruna,and LeCun2015;Duvenaud",
"et al.2015:Liet al.2016:Kipf and Welling2017)的形",
"在将介绍ST-GCN模型中的组件",
"式考虑图卷积的局部性；2）空间透视图，其中卷积滤",
"波器直接应用于图节点及其邻居E(Brunaetal.2014",
"3.2Skeleton Graph Construction",
"Niepert，Ahmed，andKutzkov2016）。这项工作遵循了",
"第二个流的精神。我们通过将每个滤波器的应用限制在",
"骨骼序列通常由每中每个人体关节的2D或3D坐",
"每个节点的1邻居上，在空间域上构造CNN滤波器",
"标表示。以前的工作使用卷积进行骨架动作识别（Kim",
"andReiter2017）将所有关节的坐标向量连接起来，形",
"成每顿一个特征向量。在我们的工作中，我们利用时",
"Skeleton Based Action Recognition.",
"人体的骨骼和",
"关节轨迹对光照变化和场景变化具有鲁棒性，并且由",
"空图来形成骨架序列的分层表示。特别是，我们在具",
"有N关节和T顿的骨架序列上构造了一个无向时空图",
"于高度精确的深度传感器或姿态估计算法而易于获得E",
"G=（V,E），该框架具有体内和顿间连接",
"(Shottonetal.2011;Caoetal.2017a）。因此，有各种各",
"样的基于骨架的动作识别方法。这些方法可以分为基于",
"在此图中，节点集V=utilt=1.....T.i",
"1，..,N)包括骨架序列中的所有关节。作为ST-GCN",
"特征的手工方法和深度学习方法。第一种方法设计了几",
"个手工制作的特征来捕捉关节运动的动力学。这些可能",
"的输人，节点F(uti）上的特征向量由框架t上第i个关",
"是关节轨迹的协方差矩阵E（Husseinetal.2013），关节",
"节的坐标向量以及估计置信度组成。我们分两步在骨架",
"序列上构造时空图。首先，根据人体结构的连通性",
"的相对位置E(Wangetal.2012)或身体部位之间的旋",
"转和平移E(Vemulapalli，Arrate,andChellappa2014)",
"个框架内的关节与边缘连接，如图。1。然后，每个关节",
"深度学习最近的成功导致了基于深度学习的骨架建",
"将连接到连续顿中的同一关节。因此，此设置中的连接",
"模方法的激增。这些工作一直在使用递归神经网络E",
"自然定义，无需手动分配零件。这也使网络架构能够处",
"(Shahroudy et al. 2016; Zhu et al. 2016:; Liu et al. 2016:",
"理具有不同数量的关节或关节连接性的数据集。例如，",
"Zhang,Liu,and Xiao 2017)和时态 CNNE(Li et al",
"在Kinetics数据集上，我们使用来自OpenPose（Cao",
"2017;Keetal.2017;KimandReiter2017）以端到端的",
"etal.2017b）工具箱的2D姿态估计结果，该工具箱输",
"方式学习动作识别模型。在这些方法中，许多人强调了",
"出18个关节，而在NTU-RGB+D数据集（Shahroudy",
"etal.2016）上，我们使用3D关节跟踪结果作为输人，",
"对人体各部位关节进行建模的重要性。但是这些部分通",
"产生25个关节。ST-GCN可以在两种情况下运行，并",
"常是使用领域知识显式分配的。我们的ST-GCN是第",
"个将图形CNN应用于基于骨架的动作识别任务的，",
"提供一致的卓越性能。所构建的时空图的示例在图。",
"",
"它与以前的方法不同，因为它可以通过利用图卷积的局",
"形式上，边缘集E由两个子集组成，第一个子集描",
"绘了每顿处的骨架内连接，表示为Es=utiutili,i）",
"部性和时间动力学来隐式地学习零件信息。通过消除对",
"H，其中H是自然连接的人体关节的集合。第二个",
"lhttps://github.com/yysijie/st-gcn",
"子集包含顿间边缘，这些边缘将连续顿中的相同关节连"
] |
{
"arxiv_id": "1801.07455",
"language": "zh",
"page_idx": 2
}
|
||
2104.13346-zh-1
|
2104.13346-zh-1.jpg
|
## 使用 FRANK 了解抽象摘要中的事实性： 事实性度量的基准
Vidhisha Balachandran
$\textbf{Yulia Tsvetkov}$
$\textbf{Artidoro Pagnoni}$
Language Technologies Institute
Carnegie Mellon University
$\{$ apagnoni,vbalacha,ytsvetko} €cs.cmu.edu
# Abstract
Modern summarization models generate highly fluent but often factually unreliable outputs. This motivated a surge of metrics attempting to measure the factuality of automatically generated summaries. Due to the lack of common benchmarks, these metrics cannot be compared. Moreover, all these methods treat factuality as a binary concept and fail to provide deeper insights on the kinds of inconsistencies made by different systems. To address these limitations, we devise a typology of factual errors and use it to collect human annotations of generated summaries from state-of-the-art summarization systems for the CNN/DM and XSum datasets. Through these annotations we identify the proportion of different categories of factual errors in various summarization models and benchmark factuality metrics, showing their correlation with human judgement as well as their specific strengths and weaknesses. l
### 1 Introduction
事实性被定义为“事件是否被表征为与世界上不存在的事实，可能性或情况相对应”的度量(Sauri, 2008; Saurí and Pustejovsky, 2012)。总而言之，这个“世界”就是文章，它被视为基本事实，输出的摘要必须忠实于文章的事实。尽管神经抽象摘要 (Narayan et al., 2018; Liu and Lapata, 2019; Lewis et al., 2020) , ??? 摘要的% 有事实不一致(Cao et al., 2018)。由于摘要是信息消费的一个组成部分，这突出了确保摘要系统在事实上一致并开发评估它们的方法的必要性
基于 n-gram 重叠的摘要的常用评估指标BLEU, ROUGE 和 METEOR (Papineni et al., 2002; Lin, 2004; Lavie and Agarwal, 2007)-不足以衡量摘要的事实正确性，并且无法与人类对事实 (Falke et al., 2019; Kryscinski et al., 2019) 的判断相关联。由于缺乏通用基准，无法比较
$^{1}$Code, data, and online leaderboard will be available at
https://github.com/artidoro/frank
为改进总结真实性 (Kryscinski et al., 2020; Durmus et al., 2020; Wang et al., 2020; Maynez et al., 2020) 的评估而提出的最新指标。更关键的是， 尽管这些方法在建模真实性方面有所不同，但它们都将真实性视为二元概念，将任何长度的摘要标记为事实或非事实。它们不能提供对不同系统所犯的事实错误的任何精细理解，这些错误可以作为对系统局限性的可操作反馈
文本的二进制真实性可能难以确定。Falke et al. (2019) 显示出相对较低的群众-专家一致性，表明在注释过程中存在主观性。此外，并非所有的事实错误都同样重要，错误的数量可能会对文本的感知真实性产生重大影响。这表明非事实性应该被建模为多维结构，而不是标签
在这项工作中，我们提出了一种基于语言动机的事实错误类型学，用于对摘要系统(§2 ) 中的事实性进行细粒度分析。我们的类型学理论上基于框架语义 (Fillmore et al., 1976; Palmer et al., 2005) 和语言话语理论 (Brown and Yule, 1983)。它提供了几个好处。首先，我们发现在(相对)定义明确且扎根的类别中分解事实性的概念会使最终的二元决策更加客观，从而导致人群和专家注释者之间的近乎完美的协议$(\kappa=0.86$ )。其次，这种方法从案文中出现的事实违反的数量和类别两个方面提供了一些非事实程度的衡量标准。这种类型学还为我们提供了对摘要系统所犯的错误类型进行分类的方法，帮助我们获得更深人的见解，而不是简单地将内容归类为事实或幻觉。我们根据我们的类型学定义了一个事实注释协议， 并在 CNN/DM (Hermann et al., 2015) 和 XSum (Narayan et al.,2018)数据集(§3)上收集了一组不同的模型生成摘要的人类判断数据集。通过这个数据集，我们的目标是评估摘要系统的真实性和基准最近提出的真实性指标。在 §4中，我们讨论了各种最先进的模型，并详细分析了它们所犯的事实错误。最后，在§5中，我们根据基准评估了多个汇总指标，并显示了它们在检测特定类型的事实错误方面的优缺点。Figure 1 显示了这项工作的概述
|
[
[
0.29296875,
0.0955810546875,
0.70947265625,
0.0955810546875,
0.70947265625,
0.10986328125,
0.29296875,
0.10986328125
],
[
0.405517578125,
0.114013671875,
0.59423828125,
0.114013671875,
0.59423828125,
0.12890625,
0.405517578125,
0.12890625
],
[
0.2166748046875,
0.152587890625,
0.370361328125,
0.152587890625,
0.370361328125,
0.16748046875,
0.2166748046875,
0.16748046875
],
[
0.41064453125,
0.152587890625,
0.615234375,
0.152587890625,
0.615234375,
0.166259765625,
0.41064453125,
0.166259765625
],
[
0.66015625,
0.152587890625,
0.78662109375,
0.152587890625,
0.78662109375,
0.166259765625,
0.66015625,
0.166259765625
],
[
0.370361328125,
0.17041015625,
0.6337890625,
0.17041015625,
0.6337890625,
0.18408203125,
0.370361328125,
0.18408203125
],
[
0.38623046875,
0.185302734375,
0.61279296875,
0.1859130859375,
0.61279296875,
0.20068359375,
0.38623046875,
0.1995849609375
],
[
0.255126953125,
0.2037353515625,
0.62890625,
0.2037353515625,
0.62890625,
0.21728515625,
0.255126953125,
0.21728515625
],
[
0.62451171875,
0.2078857421875,
0.63037109375,
0.2078857421875,
0.63037109375,
0.213134765625,
0.62451171875,
0.213134765625
],
[
0.62744140625,
0.204833984375,
0.759765625,
0.204833984375,
0.759765625,
0.215576171875,
0.62744140625,
0.215576171875
],
[
0.2626953125,
0.254150390625,
0.341796875,
0.254150390625,
0.341796875,
0.268310546875,
0.2626953125,
0.268310546875
],
[
0.51318359375,
0.254638671875,
0.880859375,
0.254638671875,
0.880859375,
0.268310546875,
0.51318359375,
0.268310546875
],
[
0.51318359375,
0.270751953125,
0.880859375,
0.270751953125,
0.880859375,
0.283935546875,
0.51318359375,
0.283935546875
],
[
0.145263671875,
0.27783203125,
0.45751953125,
0.279052734375,
0.45751953125,
0.29345703125,
0.145263671875,
0.292236328125
],
[
0.51123046875,
0.28515625,
0.8759765625,
0.28515625,
0.8759765625,
0.298583984375,
0.51123046875,
0.298583984375
],
[
0.145263671875,
0.29345703125,
0.45751953125,
0.29345703125,
0.45751953125,
0.30712890625,
0.145263671875,
0.30712890625
],
[
0.14697265625,
0.30810546875,
0.456787109375,
0.30810546875,
0.456787109375,
0.32177734375,
0.14697265625,
0.32177734375
],
[
0.5146484375,
0.302734375,
0.880859375,
0.302734375,
0.880859375,
0.31640625,
0.5146484375,
0.31640625
],
[
0.1435546875,
0.32177734375,
0.45751953125,
0.320556640625,
0.45751953125,
0.3349609375,
0.1435546875,
0.336181640625
],
[
0.5146484375,
0.319580078125,
0.880859375,
0.319580078125,
0.880859375,
0.333251953125,
0.5146484375,
0.333251953125
],
[
0.14697265625,
0.336181640625,
0.45751953125,
0.336181640625,
0.45751953125,
0.349853515625,
0.14697265625,
0.349853515625
],
[
0.51318359375,
0.336181640625,
0.87744140625,
0.336181640625,
0.87744140625,
0.3466796875,
0.51318359375,
0.3466796875
],
[
0.145263671875,
0.350830078125,
0.456787109375,
0.350830078125,
0.456787109375,
0.364013671875,
0.145263671875,
0.364013671875
],
[
0.51123046875,
0.350830078125,
0.87890625,
0.350830078125,
0.87890625,
0.364013671875,
0.51123046875,
0.364013671875
],
[
0.14697265625,
0.366455078125,
0.455078125,
0.366455078125,
0.455078125,
0.377197265625,
0.14697265625,
0.377197265625
],
[
0.5146484375,
0.369384765625,
0.8427734375,
0.369384765625,
0.8427734375,
0.380126953125,
0.5146484375,
0.380126953125
],
[
0.14697265625,
0.37890625,
0.45751953125,
0.37890625,
0.45751953125,
0.392578125,
0.14697265625,
0.392578125
],
[
0.53564453125,
0.388427734375,
0.609375,
0.388427734375,
0.609375,
0.3984375,
0.53564453125,
0.3984375
],
[
0.603515625,
0.38720703125,
0.87744140625,
0.38720703125,
0.87744140625,
0.39794921875,
0.603515625,
0.39794921875
],
[
0.145263671875,
0.393798828125,
0.456787109375,
0.393798828125,
0.456787109375,
0.40673828125,
0.145263671875,
0.40673828125
],
[
0.51318359375,
0.402587890625,
0.880859375,
0.402587890625,
0.880859375,
0.416259765625,
0.51318359375,
0.416259765625
],
[
0.145263671875,
0.4091796875,
0.455078125,
0.4091796875,
0.455078125,
0.419189453125,
0.145263671875,
0.419189453125
],
[
0.145263671875,
0.421630859375,
0.45751953125,
0.421630859375,
0.45751953125,
0.435302734375,
0.145263671875,
0.435302734375
],
[
0.51318359375,
0.419189453125,
0.880859375,
0.419189453125,
0.880859375,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.145263671875,
0.435791015625,
0.45751953125,
0.435791015625,
0.45751953125,
0.449462890625,
0.145263671875,
0.449462890625
],
[
0.51123046875,
0.43408203125,
0.880859375,
0.435302734375,
0.880859375,
0.448974609375,
0.51123046875,
0.44775390625
],
[
0.14697265625,
0.45068359375,
0.456787109375,
0.45068359375,
0.456787109375,
0.46435546875,
0.14697265625,
0.46435546875
],
[
0.51318359375,
0.451904296875,
0.880859375,
0.451904296875,
0.880859375,
0.465576171875,
0.51318359375,
0.465576171875
],
[
0.145263671875,
0.46435546875,
0.456787109375,
0.46435546875,
0.456787109375,
0.47802734375,
0.145263671875,
0.47802734375
],
[
0.51318359375,
0.46728515625,
0.880859375,
0.46728515625,
0.880859375,
0.48095703125,
0.51318359375,
0.48095703125
],
[
0.14697265625,
0.47802734375,
0.456787109375,
0.47802734375,
0.456787109375,
0.4912109375,
0.14697265625,
0.4912109375
],
[
0.51318359375,
0.482666015625,
0.5341796875,
0.482666015625,
0.5341796875,
0.496337890625,
0.51318359375,
0.496337890625
],
[
0.145263671875,
0.4921875,
0.456787109375,
0.4921875,
0.456787109375,
0.505859375,
0.145263671875,
0.505859375
],
[
0.145263671875,
0.50732421875,
0.45751953125,
0.50732421875,
0.45751953125,
0.52099609375,
0.145263671875,
0.52099609375
],
[
0.53076171875,
0.5029296875,
0.87744140625,
0.5029296875,
0.87744140625,
0.5166015625,
0.53076171875,
0.5166015625
],
[
0.1485595703125,
0.5224609375,
0.456787109375,
0.5224609375,
0.456787109375,
0.533203125,
0.1485595703125,
0.533203125
],
[
0.51123046875,
0.5185546875,
0.880859375,
0.51953125,
0.880859375,
0.533203125,
0.51123046875,
0.5322265625
],
[
0.145263671875,
0.53515625,
0.456787109375,
0.53515625,
0.456787109375,
0.548828125,
0.145263671875,
0.548828125
],
[
0.51318359375,
0.53515625,
0.880859375,
0.53515625,
0.880859375,
0.548828125,
0.51318359375,
0.548828125
],
[
0.14697265625,
0.5498046875,
0.456787109375,
0.5498046875,
0.456787109375,
0.5634765625,
0.14697265625,
0.5634765625
],
[
0.51318359375,
0.5517578125,
0.880859375,
0.5517578125,
0.880859375,
0.5654296875,
0.51318359375,
0.5654296875
],
[
0.1435546875,
0.5625,
0.42236328125,
0.5634765625,
0.42236328125,
0.57763671875,
0.1435546875,
0.57666015625
],
[
0.51123046875,
0.56640625,
0.87890625,
0.56640625,
0.87890625,
0.580078125,
0.51123046875,
0.580078125
],
[
0.51318359375,
0.58203125,
0.880859375,
0.5830078125,
0.880859375,
0.5966796875,
0.51318359375,
0.595703125
],
[
0.118408203125,
0.59130859375,
0.1502685546875,
0.59130859375,
0.1502685546875,
0.599609375,
0.118408203125,
0.599609375
],
[
0.151123046875,
0.59033203125,
0.26123046875,
0.59033203125,
0.26123046875,
0.6015625,
0.151123046875,
0.6015625
],
[
0.51318359375,
0.599609375,
0.880859375,
0.599609375,
0.880859375,
0.61328125,
0.51318359375,
0.61328125
],
[
0.1175537109375,
0.61328125,
0.486083984375,
0.61328125,
0.486083984375,
0.626953125,
0.1175537109375,
0.626953125
],
[
0.51318359375,
0.615234375,
0.880859375,
0.615234375,
0.880859375,
0.62890625,
0.51318359375,
0.62890625
],
[
0.118408203125,
0.63134765625,
0.483642578125,
0.63134765625,
0.483642578125,
0.64111328125,
0.118408203125,
0.64111328125
],
[
0.5146484375,
0.6328125,
0.87890625,
0.6328125,
0.87890625,
0.6435546875,
0.5146484375,
0.6435546875
],
[
0.1175537109375,
0.64453125,
0.486083984375,
0.646484375,
0.486083984375,
0.66015625,
0.1175537109375,
0.65771484375
],
[
0.51318359375,
0.64794921875,
0.880859375,
0.64794921875,
0.880859375,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.12005615234375,
0.6611328125,
0.486083984375,
0.6611328125,
0.486083984375,
0.67578125,
0.12005615234375,
0.67578125
],
[
0.51318359375,
0.66455078125,
0.87890625,
0.66455078125,
0.87890625,
0.677734375,
0.51318359375,
0.677734375
],
[
0.1175537109375,
0.6767578125,
0.483642578125,
0.6767578125,
0.483642578125,
0.6904296875,
0.1175537109375,
0.6904296875
],
[
0.51708984375,
0.68310546875,
0.52880859375,
0.68310546875,
0.52880859375,
0.68701171875,
0.51708984375,
0.68701171875
],
[
0.52294921875,
0.6787109375,
0.880859375,
0.6796875,
0.880859375,
0.69482421875,
0.52294921875,
0.693359375
],
[
0.118408203125,
0.69482421875,
0.486083984375,
0.69482421875,
0.486083984375,
0.7080078125,
0.118408203125,
0.7080078125
],
[
0.51318359375,
0.6953125,
0.880859375,
0.6953125,
0.880859375,
0.708984375,
0.51318359375,
0.708984375
],
[
0.1175537109375,
0.71044921875,
0.48779296875,
0.71044921875,
0.48779296875,
0.7236328125,
0.1175537109375,
0.7236328125
],
[
0.51318359375,
0.7119140625,
0.880859375,
0.7119140625,
0.880859375,
0.7255859375,
0.51318359375,
0.7255859375
],
[
0.118408203125,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.51123046875,
0.72705078125,
0.880859375,
0.72802734375,
0.880859375,
0.7421875,
0.51123046875,
0.7412109375
],
[
0.118408203125,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.51123046875,
0.74365234375,
0.8759765625,
0.74365234375,
0.8759765625,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.118408203125,
0.7587890625,
0.48779296875,
0.7587890625,
0.48779296875,
0.7724609375,
0.118408203125,
0.7724609375
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.118408203125,
0.775390625,
0.1561279296875,
0.775390625,
0.1561279296875,
0.7880859375,
0.118408203125,
0.7880859375
],
[
0.51318359375,
0.77685546875,
0.875,
0.77685546875,
0.875,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.1351318359375,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.1351318359375,
0.8046875
],
[
0.51123046875,
0.79150390625,
0.87890625,
0.79150390625,
0.87890625,
0.8046875,
0.51123046875,
0.8046875
],
[
0.1175537109375,
0.80712890625,
0.48779296875,
0.80712890625,
0.48779296875,
0.82080078125,
0.1175537109375,
0.82080078125
],
[
0.51318359375,
0.80859375,
0.880859375,
0.80859375,
0.880859375,
0.822265625,
0.51318359375,
0.822265625
],
[
0.1158447265625,
0.82177734375,
0.484375,
0.822265625,
0.484375,
0.83740234375,
0.1158447265625,
0.8359375
],
[
0.5146484375,
0.82666015625,
0.87890625,
0.82666015625,
0.87890625,
0.83740234375,
0.5146484375,
0.83740234375
],
[
0.118408203125,
0.8388671875,
0.48779296875,
0.8388671875,
0.48779296875,
0.8525390625,
0.118408203125,
0.8525390625
],
[
0.51318359375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.1175537109375,
0.85498046875,
0.486083984375,
0.85498046875,
0.486083984375,
0.8681640625,
0.1175537109375,
0.8681640625
],
[
0.5146484375,
0.859375,
0.87890625,
0.859375,
0.87890625,
0.869140625,
0.5146484375,
0.869140625
],
[
0.118408203125,
0.87158203125,
0.486083984375,
0.87158203125,
0.486083984375,
0.88525390625,
0.118408203125,
0.88525390625
],
[
0.51318359375,
0.87353515625,
0.880859375,
0.87353515625,
0.880859375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.14111328125,
0.8935546875,
0.486083984375,
0.8935546875,
0.486083984375,
0.90673828125,
0.14111328125,
0.90673828125
],
[
0.51123046875,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90625,
0.318115234375,
0.90625,
0.318115234375,
0.91943359375,
0.1175537109375,
0.91943359375
],
[
0.51318359375,
0.90625,
0.76416015625,
0.90625,
0.76416015625,
0.91943359375,
0.51318359375,
0.91943359375
]
] |
[
"使用FRANK了解抽象摘要中的事实性：",
"事实性度量的基准",
"Artidoro Pagnoni",
"Vidhisha Balachandran",
"Yulia Tsvetkov",
"Language Technologies Institute",
"CarnegieMellon University",
"{ apagnoni,vbalacha,ytsvetko }!",
"推",
"@cs.cmu.edu",
"Abstract",
"为改进总结真实性（Kryscinskietal.，2020;Dur",
"mus et al., 2020; Wang et al., 2020; Maynez et al.",
"Modern summarization models generate",
"2020)的评估而提出的最新指标。更关键的是",
"highly fluent but often factually unreliable",
"outputs. This motivated a surge of metrics",
"尽管这些方法在建模真实性方面有所不同，但",
"attempting to measure the factuality of auto.",
"它们都将真实性视为二元概念，将任何长度的",
"matically generated summaries. Due to the",
"摘要标记为事实或非事实。它们不能提供对不",
"lack of common benchmarks, these metrics",
"同系统所犯的事实错误的任何精细理解，这些",
"cannot be compared.Moreover,all these",
"错误可以作为对系统局限性的可操作反馈",
"methods treat factuality as a binary concept",
"文本的二",
"进制真实性可能难以确定。Falke",
"and fail to provide deeper insights on the",
"etal.（2019）显示出相对较低的群众-专家一致",
"kinds of inconsistencies made by different",
"systems. To address these limitations, we",
"性，表明在注释过程中存在主观性。此外，并",
"devise a typology of factual errors and use it to",
"非所有的事实错误都同样重要，错误的数量可",
"collect human annotations of generated sum",
"能会对文本的感知真实性产生重大影响。这表",
"maries from state-of-the-art summarization",
"明非事实性应该被建模为多维结构，而不是标",
"systems for the CNN/DM and XSum datasets.",
"签",
"Through these annotations we identify the",
"proportion of different categories of factual",
"在这项工作中，我们提出了一种基于语言动",
"errors in various summarization models and",
"机的事实错误类型学，用于对摘要系统（s2）",
"benchmark factuality metrics, showing their",
"中的事实性进行细粒度分析。我们的类型学理",
"correlation with human judgement as well as",
"论上基于框架语义（Fillmoreetal.，1976;Palmer",
"their specific strengths and weaknesses.",
"etal.，2005）和语言话语理论（BrownandYule",
"1983）。它提供了几个好处。首先，我们发现",
"1",
"Introduction",
"在（相对）定义明确且扎根的类别中分解事实",
"事实性被定义为“事件是否被表征为与世界上",
"性的概念会使最终的二元决策更加客观，从而",
"不存在的事实，可能性或情况相对应”的度量",
"导致人群和专家注释者之间的近乎完美的协",
"(Sauri,2008;Sauri and Pustejovsky,2012)。总而",
"议（k=0.86）。其次，这种方法从案文中出",
"言之，这个“世界”就是文章，它被视为基本",
"现的事实违反的数量和类别两个方面提供了",
"事实，输出的摘要必须忠实于文章的事实。尽",
"",
"一些非事实程度的衡量标准。这种类型学还为",
"管神经抽象摘要（Narayan etal.，2018Liuand",
"我们提供了对摘要系统所犯的错误类型进行",
"Lapata,2019;Lewisetal.,2020)，???摘要的%",
"分类的方法，帮助我们获得更深人的见解，而",
"有事实不一致（Caoetal.，2018）。由于摘要是",
"不是简单地将内容归类为事实或幻觉。我们",
"信息消费的一个组成部分，这突出了确保摘要",
"根据我们的类型学定义了一个事实注释协议",
"系统在事实上一致并开发评估它们的方法的必",
"并在 CNN/DM (Hermann et al., 2015)和 XSum",
"要性",
"(Narayanetal.，2018）数据集（s3）上收集了一",
"基于n-gram重叠的摘要的常用评估指标-",
"组不同的模型生成摘要的人类判断数据集。通",
"BLEU,ROUGE 和 METEOR (Papineni et al.,",
"过这个数据集，我们的目标是评估摘要系统的",
"2002;Lin,2004; Lavie and Agarwal, 2007)-不足",
"真实性和基准最近提出的真实性指标。在84",
"以衡量摘要的事实正确性，并且无法与人类对",
"中，我们讨论了各种最先进的模型，并详细分",
"事实 (Falke et al., 2019; Kryscinski et al., 2019)",
"析了它们所犯的事实错误。最后，在85中，我",
"的判断相关联。由于缺乏通用基准，无法比较",
"们根据基准评估了多个汇总指标，并显示了它",
"'Code, data, and online leaderboard will be available at",
"们在检测特定类型的事实错误方面的优缺点。",
"https://github.com/artidoro/frank",
"Figure1显示了这项工作的概述"
] |
{
"arxiv_id": "2104.13346",
"language": "zh",
"page_idx": 1
}
|
||
2104.13346-zh-2
|
2104.13346-zh-2.jpg
|
<FigureHere>
Figure 1: We propose a linguistically grounded typology of factual errors. We select crowd workers to annotate summaries from two datasets according to this typology achieving near perfect agreement with experts. We collect FRANK , the resulting dataset, to benchmark factuality metrics and state-of-art summarization systems.
<table>
<tbody>
<tr>
<th> </th>
<th>Category</th>
<th>Description</th>
<th>Example</th>
</tr>
<tr>
<td>PredE</td>
<td>Relation Error</td>
<td>The predicate in the summary statement is inconsistent with the source article.</td>
<td>The Ebola vaccine was rejected $\textbf{d by the FDA}$ $\operatorname{in}2019.$</td>
</tr>
<tr>
<td>EntE</td>
<td>Entity Error </td>
<td>The primary arguments (or their $at-$ tributes) of the predicate are wrong.</td>
<td>The COVID-19 vaccine was approved by the FDA in 2019.</td>
</tr>
<tr>
<td>CircE</td>
<td>Circumstance Error</td>
<td>The additional information (like loca- tion or time) specifying the circumstance around a predicate is wrong.</td>
<td>The first vaccine for Ebola was approved by the FDA in 2014 .</td>
</tr>
<tr>
<td>$\overline{{\mathrm{CorefE}}}$</td>
<td>Coreference Error</td>
<td>A pronoun/reference with wrong or non- existing antecedent.</td>
<td>The first vaccine for Ebola was approved in 2019. They say a vaccine for COVID-19 is unlikely to be ready this year.</td>
</tr>
<tr>
<td>LinkE</td>
<td>Discourse Link $Er-$ ror</td>
<td>${\mathrm{Error~in~how~multiple~statements}}$ s are linked together in the discourse (for ex- ample temporal ordering/causal link).</td>
<td>To produce the vaccine, scientists have to show successful human trials, then quence the DNA of the virus.</td>
</tr>
<tr>
<td>$\overline{OutE}$</td>
<td>Out of Article Error</td>
<td>The statement contains information not present in the source article.</td>
<td>$\overline{\text{China}}$ has already started clinical trials of the COVID-19 vaccine.</td>
</tr>
<tr>
<td>GramE</td>
<td>Grammatical Error</td>
<td>The grammar of the sentence is so wrong that it becomes meaningless.</td>
<td>The Ebola vaccine accepted have already started.</td>
</tr>
</tbody>
</table>
Table 1: Typology of factual errors. Original text for the examples: The first vaccine for Ebola was approved by the FDA in 2019 in the US, five years after the initial outbreak in 2014. To produce the vaccine, scientists had to sequence the DNA of Ebola, then identify possible vaccines, and finally show successful clinical trials. Scientists say a vaccine for COVID-19 is unlikely to be ready this year, although clinical trials have already started.
### 2.1 Semantic Frame Errors
## 2 $\textbf{Typology of Factual Errors}$
语义框架是事件，关系或状态的示意性表示， 它由谓词和参与者列表组成，称为框架元素(Baker et al., 1998)。语义框架具有核心和非核心框架元素(FE)。核心框架元素对框架的含义至关重要，而非核心(e。g。位置、时间)提供额外的描述性信息。我们的前三个类别分别捕获了这些组件(框架，核心和非核心 FE) 中的事实错误
Predicate Error( PredE ): 类别 PredE 包含摘要语句中的谓词与源文本不一致的错误。更一般地，这表示来自摘要语句的框架与源文本中表达的内容不一致的情况。
Entity Error( EntE ): 类别 EntE 捕获谓词的主要参数(如实体)错误或具有错误属性的错误，尽管关系在原始文本中表示。更一般地， 这些解释了框架中核心框架元素错误的情况。这也可以捕获元素互换的方向性错误(类似于
以前的摘要事实研究仅区分事实和幻觉内容(Kryscinski et al., 2019; Maynez et al., 2020) , 并对细粒度的事实错误类型提供了有限的见解。在最简单的情况下，事实错误出现在一个命题中。但是，由于摘要包括多个句子，因此话语标记描述了命题之间的关系。这些交叉句子的链接，如因果关系或时间顺序，可能会引入与文章的不一致之处。此外，鉴于该条，摘要中的信息应该是可核实的。这种理解概述了不同层次的语言结构，在摘要中可能会出现事实错误：在语义框架层面，在话语层面，或者因为内容无法验证。下面我们定义一个事实错误的类型学，进一步详细说明这三个层次。这种类型学的理论基础是框架语义 (Fillmore et al., 1976; Baker et al., 1998; Palmer et al., 2005) 和语言话语分析 (Brown and Yule, 1983) 。Table 1 中显示了每个类别的示例
|
[
[
0.364501953125,
0.0885009765625,
0.51904296875,
0.0885009765625,
0.51904296875,
0.1015625,
0.364501953125,
0.1015625
],
[
0.1502685546875,
0.0968017578125,
0.198974609375,
0.0968017578125,
0.198974609375,
0.10809326171875,
0.1502685546875,
0.10809326171875
],
[
0.57080078125,
0.09442138671875,
0.580078125,
0.09442138671875,
0.580078125,
0.10394287109375,
0.57080078125,
0.10394287109375
],
[
0.365234375,
0.1009521484375,
0.5615234375,
0.1009521484375,
0.5615234375,
0.11102294921875,
0.365234375,
0.11102294921875
],
[
0.6591796875,
0.1092529296875,
0.85986328125,
0.10986328125,
0.85986328125,
0.12353515625,
0.6591796875,
0.122314453125
],
[
0.47265625,
0.1175537109375,
0.483642578125,
0.1175537109375,
0.483642578125,
0.12890625,
0.47265625,
0.12890625
],
[
0.6630859375,
0.122314453125,
0.75830078125,
0.122314453125,
0.75830078125,
0.1341552734375,
0.6630859375,
0.1341552734375
],
[
0.1561279296875,
0.1312255859375,
0.19140625,
0.1312255859375,
0.19140625,
0.142578125,
0.1561279296875,
0.142578125
],
[
0.34765625,
0.1312255859375,
0.447509765625,
0.1318359375,
0.447509765625,
0.1436767578125,
0.34765625,
0.142578125
],
[
0.56494140625,
0.1341552734375,
0.5859375,
0.1341552734375,
0.5859375,
0.146728515625,
0.56494140625,
0.146728515625
],
[
0.3525390625,
0.144287109375,
0.56103515625,
0.144287109375,
0.56103515625,
0.155029296875,
0.3525390625,
0.155029296875
],
[
0.66015625,
0.1478271484375,
0.82373046875,
0.1484375,
0.82373046875,
0.162109375,
0.66015625,
0.160888671875
],
[
0.6630859375,
0.162109375,
0.79443359375,
0.162109375,
0.79443359375,
0.1728515625,
0.6630859375,
0.1728515625
],
[
0.130126953125,
0.149658203125,
0.3046875,
0.15087890625,
0.3046875,
0.162109375,
0.130126953125,
0.160888671875
],
[
0.267822265625,
0.17333984375,
0.330078125,
0.17333984375,
0.330078125,
0.189453125,
0.267822265625,
0.189453125
],
[
0.361083984375,
0.1728515625,
0.5146484375,
0.17333984375,
0.5146484375,
0.188232421875,
0.361083984375,
0.18701171875
],
[
0.1259765625,
0.1817626953125,
0.2626953125,
0.1817626953125,
0.2626953125,
0.1923828125,
0.1259765625,
0.1923828125
],
[
0.7119140625,
0.18115234375,
0.81689453125,
0.18115234375,
0.81689453125,
0.191162109375,
0.7119140625,
0.191162109375
],
[
0.27197265625,
0.1923828125,
0.328369140625,
0.1923828125,
0.328369140625,
0.1995849609375,
0.27197265625,
0.1995849609375
],
[
0.366943359375,
0.18701171875,
0.51318359375,
0.18701171875,
0.51318359375,
0.19775390625,
0.366943359375,
0.19775390625
],
[
0.1158447265625,
0.215576171875,
0.880859375,
0.215576171875,
0.880859375,
0.2286376953125,
0.1158447265625,
0.2286376953125
],
[
0.1158447265625,
0.22802734375,
0.8818359375,
0.2286376953125,
0.8818359375,
0.243408203125,
0.1158447265625,
0.2423095703125
],
[
0.1158447265625,
0.243408203125,
0.8134765625,
0.243408203125,
0.8134765625,
0.257080078125,
0.1158447265625,
0.257080078125
],
[
0.18896484375,
0.273681640625,
0.2509765625,
0.275634765625,
0.2493896484375,
0.290283203125,
0.187255859375,
0.2880859375
],
[
0.32666015625,
0.273681640625,
0.405517578125,
0.275634765625,
0.405517578125,
0.290283203125,
0.32666015625,
0.2880859375
],
[
0.59423828125,
0.27490234375,
0.65576171875,
0.27490234375,
0.65576171875,
0.289306640625,
0.59423828125,
0.289306640625
],
[
0.1234130859375,
0.292236328125,
0.1695556640625,
0.292236328125,
0.1695556640625,
0.30712890625,
0.1234130859375,
0.30712890625
],
[
0.328369140625,
0.29345703125,
0.5810546875,
0.29345703125,
0.5810546875,
0.30712890625,
0.328369140625,
0.30712890625
],
[
0.59375,
0.29345703125,
0.8623046875,
0.29345703125,
0.8623046875,
0.30712890625,
0.59375,
0.30712890625
],
[
0.32666015625,
0.303955078125,
0.5556640625,
0.3046875,
0.5556640625,
0.318359375,
0.32666015625,
0.317626953125
],
[
0.189697265625,
0.29443359375,
0.280517578125,
0.29443359375,
0.280517578125,
0.305908203125,
0.189697265625,
0.305908203125
],
[
0.59619140625,
0.305908203125,
0.6455078125,
0.305908203125,
0.6455078125,
0.317626953125,
0.59619140625,
0.317626953125
],
[
0.118408203125,
0.3154296875,
0.162841796875,
0.317626953125,
0.162109375,
0.33203125,
0.1175537109375,
0.330078125
],
[
0.32666015625,
0.31640625,
0.58447265625,
0.318359375,
0.58447265625,
0.333251953125,
0.32666015625,
0.330810546875
],
[
0.59375,
0.31640625,
0.86376953125,
0.318359375,
0.86376953125,
0.33203125,
0.59375,
0.330078125
],
[
0.328369140625,
0.330078125,
0.54736328125,
0.330078125,
0.54736328125,
0.34326171875,
0.328369140625,
0.34326171875
],
[
0.189697265625,
0.317626953125,
0.26611328125,
0.319580078125,
0.26611328125,
0.330810546875,
0.189697265625,
0.328857421875
],
[
0.59619140625,
0.330810546875,
0.70361328125,
0.330810546875,
0.70361328125,
0.341552734375,
0.59619140625,
0.341552734375
],
[
0.12176513671875,
0.34033203125,
0.166259765625,
0.342529296875,
0.16455078125,
0.35693359375,
0.12005615234375,
0.355224609375
],
[
0.18896484375,
0.342529296875,
0.318115234375,
0.342529296875,
0.318115234375,
0.355712890625,
0.18896484375,
0.355712890625
],
[
0.328369140625,
0.342529296875,
0.58251953125,
0.342529296875,
0.58251953125,
0.355712890625,
0.328369140625,
0.355712890625
],
[
0.591796875,
0.341552734375,
0.86572265625,
0.342529296875,
0.86572265625,
0.35693359375,
0.591796875,
0.355712890625
],
[
0.330078125,
0.355712890625,
0.5810546875,
0.355712890625,
0.5810546875,
0.366455078125,
0.330078125,
0.366455078125
],
[
0.59619140625,
0.355224609375,
0.705078125,
0.355224609375,
0.705078125,
0.365234375,
0.59619140625,
0.365234375
],
[
0.328369140625,
0.365234375,
0.50146484375,
0.366455078125,
0.50146484375,
0.380126953125,
0.328369140625,
0.37890625
],
[
0.12005615234375,
0.37890625,
0.173828125,
0.37890625,
0.173828125,
0.39013671875,
0.12005615234375,
0.39013671875
],
[
0.189697265625,
0.380126953125,
0.30322265625,
0.380126953125,
0.30322265625,
0.39013671875,
0.189697265625,
0.39013671875
],
[
0.330078125,
0.37890625,
0.58251953125,
0.37890625,
0.58251953125,
0.3896484375,
0.330078125,
0.3896484375
],
[
0.59619140625,
0.380126953125,
0.86376953125,
0.380126953125,
0.86376953125,
0.39013671875,
0.59619140625,
0.39013671875
],
[
0.328369140625,
0.391357421875,
0.450927734375,
0.391357421875,
0.450927734375,
0.402099609375,
0.328369140625,
0.402099609375
],
[
0.59619140625,
0.388427734375,
0.86376953125,
0.388427734375,
0.86376953125,
0.402099609375,
0.59619140625,
0.402099609375
],
[
0.59619140625,
0.402099609375,
0.77734375,
0.402099609375,
0.77734375,
0.4150390625,
0.59619140625,
0.4150390625
],
[
0.18896484375,
0.41455078125,
0.316650390625,
0.41455078125,
0.316650390625,
0.427490234375,
0.18896484375,
0.427490234375
],
[
0.32568359375,
0.413330078125,
0.5810546875,
0.4150390625,
0.5810546875,
0.4287109375,
0.32568359375,
0.427001953125
],
[
0.59765625,
0.416259765625,
0.86376953125,
0.416259765625,
0.86376953125,
0.427001953125,
0.59765625,
0.427001953125
],
[
0.328369140625,
0.42578125,
0.58447265625,
0.42578125,
0.58447265625,
0.439453125,
0.328369140625,
0.439453125
],
[
0.12005615234375,
0.4150390625,
0.1695556640625,
0.4150390625,
0.1695556640625,
0.427001953125,
0.12005615234375,
0.427001953125
],
[
0.18896484375,
0.427490234375,
0.214111328125,
0.427490234375,
0.214111328125,
0.439453125,
0.18896484375,
0.439453125
],
[
0.59423828125,
0.42578125,
0.86572265625,
0.42578125,
0.86572265625,
0.439453125,
0.59423828125,
0.439453125
],
[
0.328369140625,
0.438232421875,
0.5576171875,
0.438232421875,
0.5576171875,
0.451904296875,
0.328369140625,
0.451904296875
],
[
0.59423828125,
0.438232421875,
0.77685546875,
0.438232421875,
0.77685546875,
0.451904296875,
0.59423828125,
0.451904296875
],
[
0.12005615234375,
0.45068359375,
0.162841796875,
0.45068359375,
0.162841796875,
0.4619140625,
0.12005615234375,
0.4619140625
],
[
0.19140625,
0.451904296875,
0.31494140625,
0.451904296875,
0.31494140625,
0.4619140625,
0.19140625,
0.4619140625
],
[
0.33154296875,
0.451904296875,
0.5810546875,
0.451904296875,
0.5810546875,
0.4619140625,
0.33154296875,
0.4619140625
],
[
0.59619140625,
0.451904296875,
0.8671875,
0.451904296875,
0.8671875,
0.4619140625,
0.59619140625,
0.4619140625
],
[
0.328369140625,
0.463134765625,
0.49951171875,
0.4619140625,
0.49951171875,
0.473876953125,
0.328369140625,
0.474365234375
],
[
0.59765625,
0.463134765625,
0.74072265625,
0.463134765625,
0.74072265625,
0.473876953125,
0.59765625,
0.473876953125
],
[
0.124267578125,
0.474365234375,
0.173828125,
0.474365234375,
0.173828125,
0.486328125,
0.124267578125,
0.486328125
],
[
0.189697265625,
0.4755859375,
0.30908203125,
0.4755859375,
0.30908203125,
0.486328125,
0.189697265625,
0.486328125
],
[
0.33154296875,
0.4755859375,
0.5810546875,
0.4755859375,
0.5810546875,
0.486328125,
0.33154296875,
0.486328125
],
[
0.591796875,
0.47265625,
0.861328125,
0.473876953125,
0.861328125,
0.48681640625,
0.591796875,
0.486328125
],
[
0.328369140625,
0.488037109375,
0.5029296875,
0.488037109375,
0.5029296875,
0.498779296875,
0.328369140625,
0.498779296875
],
[
0.59619140625,
0.488037109375,
0.64208984375,
0.488037109375,
0.64208984375,
0.498779296875,
0.59619140625,
0.498779296875
],
[
0.1158447265625,
0.51416015625,
0.880859375,
0.51416015625,
0.880859375,
0.529296875,
0.1158447265625,
0.529296875
],
[
0.1175537109375,
0.529296875,
0.8818359375,
0.529296875,
0.8818359375,
0.5419921875,
0.1175537109375,
0.5419921875
],
[
0.1175537109375,
0.5419921875,
0.8818359375,
0.5419921875,
0.8818359375,
0.5556640625,
0.1175537109375,
0.5556640625
],
[
0.1158447265625,
0.55712890625,
0.82373046875,
0.55712890625,
0.82373046875,
0.57080078125,
0.1158447265625,
0.57080078125
],
[
0.1502685546875,
0.59716796875,
0.387939453125,
0.59716796875,
0.387939453125,
0.61083984375,
0.1502685546875,
0.61083984375
],
[
0.54833984375,
0.59716796875,
0.748046875,
0.59716796875,
0.748046875,
0.61083984375,
0.54833984375,
0.61083984375
],
[
0.51318359375,
0.5986328125,
0.5556640625,
0.5986328125,
0.5556640625,
0.60986328125,
0.51318359375,
0.60986328125
],
[
0.118408203125,
0.5986328125,
0.1368408203125,
0.5986328125,
0.1368408203125,
0.60791015625,
0.118408203125,
0.60791015625
],
[
0.133544921875,
0.599609375,
0.1502685546875,
0.599609375,
0.1502685546875,
0.60693359375,
0.133544921875,
0.60693359375
],
[
0.51123046875,
0.619140625,
0.880859375,
0.62060546875,
0.880859375,
0.63525390625,
0.51123046875,
0.63427734375
],
[
0.5146484375,
0.63720703125,
0.87890625,
0.63720703125,
0.87890625,
0.65087890625,
0.5146484375,
0.65087890625
],
[
0.1175537109375,
0.6455078125,
0.484375,
0.6455078125,
0.484375,
0.6591796875,
0.1175537109375,
0.6591796875
],
[
0.509765625,
0.65185546875,
0.87890625,
0.65185546875,
0.87890625,
0.6650390625,
0.509765625,
0.6650390625
],
[
0.118408203125,
0.6630859375,
0.484375,
0.6630859375,
0.484375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.51123046875,
0.66748046875,
0.87744140625,
0.66748046875,
0.87744140625,
0.68115234375,
0.51123046875,
0.68115234375
],
[
0.1158447265625,
0.677734375,
0.48779296875,
0.6787109375,
0.48779296875,
0.693359375,
0.1158447265625,
0.6923828125
],
[
0.51318359375,
0.68310546875,
0.880859375,
0.68408203125,
0.880859375,
0.69873046875,
0.51318359375,
0.69775390625
],
[
0.118408203125,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51318359375,
0.69970703125,
0.880859375,
0.70068359375,
0.880859375,
0.71533203125,
0.51318359375,
0.71435546875
],
[
0.118408203125,
0.71142578125,
0.486083984375,
0.71142578125,
0.486083984375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.51318359375,
0.71728515625,
0.87890625,
0.71728515625,
0.87890625,
0.73095703125,
0.51318359375,
0.73095703125
],
[
0.118408203125,
0.72705078125,
0.484375,
0.7255859375,
0.484375,
0.74072265625,
0.118408203125,
0.7412109375
],
[
0.51318359375,
0.73388671875,
0.60693359375,
0.73388671875,
0.60693359375,
0.7490234375,
0.51318359375,
0.7490234375
],
[
0.118408203125,
0.74365234375,
0.483642578125,
0.74365234375,
0.483642578125,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.12005615234375,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.12005615234375,
0.77392578125
],
[
0.515625,
0.76123046875,
0.880859375,
0.7626953125,
0.880859375,
0.77685546875,
0.515625,
0.775390625
],
[
0.118408203125,
0.775390625,
0.486083984375,
0.775390625,
0.486083984375,
0.7890625,
0.118408203125,
0.7890625
],
[
0.51123046875,
0.77783203125,
0.87890625,
0.77783203125,
0.87890625,
0.79150390625,
0.51123046875,
0.79150390625
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.52880859375,
0.79443359375,
0.880859375,
0.79443359375,
0.880859375,
0.80810546875,
0.52880859375,
0.80810546875
],
[
0.1175537109375,
0.80810546875,
0.483642578125,
0.80810546875,
0.483642578125,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.51123046875,
0.81005859375,
0.74072265625,
0.81005859375,
0.74072265625,
0.82470703125,
0.51123046875,
0.82470703125
],
[
0.1175537109375,
0.82470703125,
0.483642578125,
0.82470703125,
0.483642578125,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.486083984375,
0.84033203125,
0.486083984375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.51708984375,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85400390625,
0.51708984375,
0.85400390625
],
[
0.1158447265625,
0.85498046875,
0.489501953125,
0.85546875,
0.489501953125,
0.87060546875,
0.1158447265625,
0.869140625
],
[
0.51318359375,
0.85498046875,
0.880859375,
0.85546875,
0.880859375,
0.87060546875,
0.51318359375,
0.869140625
],
[
0.118408203125,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.8759765625,
0.87255859375,
0.8759765625,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.873046875,
0.87744140625,
0.87744140625,
0.87744140625,
0.87744140625,
0.8828125,
0.873046875,
0.8828125
],
[
0.118408203125,
0.88818359375,
0.486083984375,
0.88720703125,
0.486083984375,
0.90185546875,
0.118408203125,
0.90283203125
],
[
0.509765625,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.509765625,
0.90185546875
],
[
0.1175537109375,
0.90478515625,
0.300537109375,
0.90478515625,
0.300537109375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.87744140625,
0.90478515625,
0.87744140625,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Typology of Factual Errors",
"CNN/DM",
"福",
"Frame semantics,discourseanalysis",
"Summarization SystemEvaluation",
"+",
"Which mistakes?",
"xSum",
"Data Annotation",
"Q甲",
"Per sentence error category annotation",
"Factuality Metric Evaluation",
"What is the best metric?",
"2250Summaries by9Systems",
"",
"Worker Quality Screening",
"Amazon Mechanical Turk",
"FRANKBenchmark",
"AAA",
"Training.Evaluation,Bonus",
"Figure l: We propose a linguistically grounded typology of factual errors. We select crowd workers to annotate",
"summaries from two datasets according to this typology achieving near perfect agreement with experts. We collect",
"FRANK , the resulting dataset, to benchmark factuality metrics and state-of-art summarization systems.",
"Category",
"Description",
"Example",
"PredE",
"The predicate in the summary statement",
"The Ebola vaccine was rejected by the FDA",
"is inconsistent with the source article",
"Relation Error",
"in 2019",
"EntE",
"The primary arguments (or their at-",
"The COVID-19 vaccine was approved by",
"tributes) of the predicate are wrong.",
"Entity Error",
"the FDA in 2019.",
"CircE",
"Circumstance Error",
"The additional information (like loca-",
"The first vaccine for Ebola was approved by",
"tion or time) specifying the circumstance",
"theFDAin2014",
"around a predicate is wrong",
"CorefE",
"CoreferenceError",
"Apronoun/reference withwrong ornon-",
"The first vaccineforEbola was approved in",
"existing antecedent.",
"2019. They say a vaccine for COVID-19 is",
"unlikely to be ready this year.",
"Discourse Link Er-",
"Error in how multiple statements are",
"To produce the vaccine, scientists have to",
"linked together in the discourse (for ex-",
"LinkE",
"ror",
"show successful human trials, then se-",
"ample temporal ordering/causal link).",
"quence the DNA of the virus.",
"OutE",
"Outof ArticleError",
"The statement contains information not",
"China has already started clinical trials of",
"present in the source article.",
"the COVID-19 vaccine.",
"GramE",
"Grammatical Error",
"Thegrammarof thesentenceis so wrong",
"The Ebola vaccine accepted have already",
"that it becomes meaningless",
"started.",
"Table 1: Typology of factual errors. Original text for the examples: The first vaccine for Ebola was approved by",
"the FDA in 2019 in the US, five years after the initial outbreak in 2014. To produce the vaccine, scientists had to",
"sequence the DNA of Ebola, then identify possible vaccines, and finally show successful clinical trials. Scientists",
"say a vaccine for COVID-19 is unlikely to be ready this year, although clinical trials have already started.",
"Typology of Factual Errors",
"Semantic Frame Errors",
"2.1",
"2",
"",
"语义框架是事件，关系或状态的示意性表示，",
"它由谓词和参与者列表组成，称为框架元素",
"以前的摘要事实研究仅区分事实和幻觉内容",
"（Bakeretal.，1998）。语义框架具有核心和非核",
"(Kryscinski et al., 2019; Maynez et al., 2020),并",
"心框架元素（FE）。核心框架元素对框架的含",
"对细粒度的事实错误类型提供了有限的见解。",
"义至关重要，而非核心(e。g。位置、时间)提",
"在最简单的情况下，事实错误出现在一个命",
"供额外的描述性信息。我们的前三个类别分别",
"题中。但是，由于摘要包括多个句子，因此话",
"捕获了这些组件（框架，核心和非核心FE）中",
"语标记描述了命题之间的关系。这些交叉句子",
"的事实错误",
"的链接，如因果关系或时间顺序，可能会引入",
"与文章的不一致之处。此外，鉴于该条，摘要",
"PredicateError（PredE）：类别PredE包含",
"中的信息应该是可核实的。这种理解概述了不",
"摘要语句中的谓词与源文本不一致的错误。更",
"同层次的语言结构，在摘要中可能会出现事实",
"般地，这表示来自摘要语句的框架与源文本",
"错误：在语义框架层面，在话语层面，或者因",
"中表达的内容不一致的情况。",
"为内容无法验证。下面我们定义一个事实错误",
"的类型学，进一步详细说明这三个层次。这种",
"EntityError（EntE）：类别EntE捕获谓词的",
"类型学的理论基础是框架语义（Fillmoreetal.,",
"主要参数（如实体）错误或具有错误属性的错",
"1976; Baker et al., 1998; Palmer et al., 2005)和语",
"误，尽管关系在原始文本中表示。更一般地",
"",
"言话语分析（BrownandYule，1983）。Table1中",
"这些解释了框架中核心框架元素错误的情况。",
"显示了每个类别的示例",
"这也可以捕获元素互换的方向性错误（类似于"
] |
{
"arxiv_id": "2104.13346",
"language": "zh",
"page_idx": 2
}
|
||
2104.13346-zh-3
|
2104.13346-zh-3.jpg
|
代理-患者交换)。
Circumstance Error( CircE): 除了核心参数 除了理论基础之外，我们还通过对 CNN/DM 之外，还可以使用描述参数和谓词交互的情况 数据集上的五个抽象摘要模型和 XSum 数据集心 , 也 可 以 及 门 面 过 多 双 件 阴 内 人 工 口 门 旧 几 的 附 加 信 息 或 属 性 来 进 一 步 指 定 谓 词 ( 例 如 。上 的 四 个 抽 象 摘 要 模 型 的 大 规 模 人 工 注 释 , 对g。位置、时间、方式、方向、模态)。类别CircE 我们的类型学进行了经验验证。通过我们的数捕获一个或多个此类属性 (帧中的非核心帧元 据集，我们的目标是广泛覆盖神经汇总系统所素) 错误的错误
# 2.2 Discourse Errors
作者的交际意图也通过文本各部分之间的关系来表达。摘要文本中的事实错误通常可以扩展到单个语义框架之外，从而在话语片段之间引入错误的链接。下面我们概述了基于话语分析和修辞结构理论 (RST) (Brown and Yule, 1983; Mann and Thompson, 1988) 的此类错误。RST 是一个用于注释语篇中连贯关系的复杂系统。这种关系的一些例子包括：“阐述”,“背景”,“动机”和“意志原因”。在这里，我们脱离了语义框架术语，因为它植根于单个框架中不允许我们表示此类错误
Coreference Error( CorefE): 类别 CorefE 说明了错误，其中代词和对先前提到的实体的其他类型的引用不正确或没有明确的前置词，从而使它们模棱两可
Discourse Link Error( LinkE): 类别 LinkE 包含涉及不同语句之间的话语链接的错误。这些包括错误的时间顺序或错误的话语链接 (例如。g。RST 关系、话语连接器)在语句之间
# 2.3 Content Verifiability Errors
通常，由于难以将摘要中的语句与源文本对齐，因此无法根据源文本对其进行验证。下面我们概述了这种情况下的两类错误。
Out of Article Error ( OutE): 由于文档摘要应仅包含可以从原始文本中推断出的信息，因此我们为此类错误 OutE 添加了一个类别(先前的工作将其称为外在幻觉 (Maynez et al., 2020) (1)
Grammatical Error( GramE): 我们使用GramE 对格式不正确的语句进行分类。当语法错误使陈述的含义难以理解或模棱两可时， 无法根据来源进行验证，因此被认为是微不足道的错误。轻微的语法错误是可以接受的
最后，为了在注释练习中保持完整性，我们为与上述任何类别都不对应的事实错误添加了另外两个类别 Others( OthE ), 为不包含任何错误的语句添加了 NotanError( NE )
# 3 Dataset Creation
产生的不同类型的错误，并对其细粒度的事实
错误进行人为判断
Annotation Data 对于注释，我们包括来自
CNN/DM 和 XSum 数据集的模型摘要，因为它们呈现不同的特征。CNN/DM 摘要较长，平均三句话，而 XSum 只有一句话摘要。较长的摘要对于识别话语级别的错误至关重要。另一方面，XSum 摘要更加抽象，并且平均 (Maynez) et al., 2020) 包含更多的事实错误。对于一组不同的模型摘要，我们从具有不同事实能力的不同摘要模型中收集公开可用的模型输出。对于CNN/DM 数据集，我们使用来自 LSTM 序列到序列模型 (S2S) (Rush et al.,2015),指针生成器网络(PGN) 模型(See et al.,2017),自下而上总结(BUS)模型(Gehrmann et al.,2018) , 基于 Bert 的提取抽象模型 (BertSum) (Liu and Lapata, 2019) 和联合训练的预训练编码器的模型。对于 XSum 数据集，我们从主题感知 CNN 模型 (Narayan et al.,2018),指针生成器网络(PGN)模型，随机初始化 (TransS2S) (Vaswani et al., 2017) 和用 Bert-Base ( BertS2S) (Devlin et al., 2019) 初始化的模型输出。$^{2}$§A.1 中提供了所用模型的详细信息
Annotation Collection 使用上述模型生成的摘要，我们从三个独立的注释器收集来自每个数据集的 250 篇文章的人类注释 (CNN/DM 上共有 1250 个模型输出，XSum 上有 1000 个)。我们对摘要的每个句子进行注释，以将对事实的判断分解为较小的单位。我们在整个摘要的上下文中呈现句子，以识别跨越多个句子的话语错误。注释是一个两步过程：对于摘要中的每个句子，注释者首先选择句子是否为事实， 如果标记为非事实，则根据我们的类型学确定每个错误的类别。$^{3}$句子可以用多个错误类别进行注释，以解决句子中的多个错误。我们在 AmazonMechanicalTurk(MTurk) 平台上执行注释任务。为了实现高质量的众包注释，我们
$^{2}$As we use publicly available model outputs, the sum-
maries across different datasets are from different models ow-
ing to their availability.
$^{3}$We experimented with Likert scale evaluation of full summaries in a pilot study. Such an annotation would not provide precise information about where in the summary an error appears and also resulted in lower agreement. Hence, we opted for sentence level judgements.
|
[
[
0.12005615234375,
0.0872802734375,
0.2509765625,
0.0872802734375,
0.2509765625,
0.1009521484375,
0.12005615234375,
0.1009521484375
],
[
0.5087890625,
0.08612060546875,
0.69287109375,
0.08612060546875,
0.69287109375,
0.09979248046875,
0.5087890625,
0.09979248046875
],
[
0.12176513671875,
0.11639404296875,
0.48779296875,
0.11639404296875,
0.48779296875,
0.1300048828125,
0.12176513671875,
0.1300048828125
],
[
0.51318359375,
0.1134033203125,
0.880859375,
0.1134033203125,
0.880859375,
0.1270751953125,
0.51318359375,
0.1270751953125
],
[
0.118408203125,
0.1318359375,
0.48779296875,
0.1318359375,
0.48779296875,
0.1455078125,
0.118408203125,
0.1455078125
],
[
0.51318359375,
0.1300048828125,
0.880859375,
0.1300048828125,
0.880859375,
0.1436767578125,
0.51318359375,
0.1436767578125
],
[
0.118408203125,
0.1478271484375,
0.489501953125,
0.1478271484375,
0.489501953125,
0.160888671875,
0.118408203125,
0.160888671875
],
[
0.5146484375,
0.1455078125,
0.880859375,
0.1455078125,
0.880859375,
0.1591796875,
0.5146484375,
0.1591796875
],
[
0.1175537109375,
0.1644287109375,
0.48779296875,
0.163330078125,
0.48779296875,
0.1776123046875,
0.1175537109375,
0.1787109375
],
[
0.51318359375,
0.162109375,
0.87744140625,
0.162109375,
0.87744140625,
0.1728515625,
0.51318359375,
0.1728515625
],
[
0.118408203125,
0.179931640625,
0.48779296875,
0.179931640625,
0.48779296875,
0.193603515625,
0.118408203125,
0.193603515625
],
[
0.509765625,
0.177001953125,
0.87890625,
0.177001953125,
0.87890625,
0.190673828125,
0.509765625,
0.190673828125
],
[
0.118408203125,
0.196533203125,
0.2476806640625,
0.196533203125,
0.2476806640625,
0.210205078125,
0.118408203125,
0.210205078125
],
[
0.515625,
0.1953125,
0.87890625,
0.1953125,
0.87890625,
0.2060546875,
0.515625,
0.2060546875
],
[
0.5146484375,
0.210205078125,
0.66162109375,
0.210205078125,
0.66162109375,
0.223876953125,
0.5146484375,
0.223876953125
],
[
0.1175537109375,
0.226806640625,
0.300537109375,
0.226806640625,
0.300537109375,
0.240478515625,
0.1175537109375,
0.240478515625
],
[
0.5205078125,
0.2369384765625,
0.87890625,
0.2369384765625,
0.87890625,
0.25048828125,
0.5205078125,
0.25048828125
],
[
0.1175537109375,
0.25048828125,
0.484375,
0.25048828125,
0.484375,
0.26416015625,
0.1175537109375,
0.26416015625
],
[
0.51318359375,
0.254150390625,
0.880859375,
0.254150390625,
0.880859375,
0.267333984375,
0.51318359375,
0.267333984375
],
[
0.12005615234375,
0.268310546875,
0.484375,
0.268310546875,
0.484375,
0.279052734375,
0.12005615234375,
0.279052734375
],
[
0.51318359375,
0.26953125,
0.880859375,
0.26953125,
0.880859375,
0.283203125,
0.51318359375,
0.283203125
],
[
0.118408203125,
0.283203125,
0.484375,
0.283203125,
0.484375,
0.29638671875,
0.118408203125,
0.29638671875
],
[
0.51318359375,
0.2861328125,
0.880859375,
0.2861328125,
0.880859375,
0.2998046875,
0.51318359375,
0.2998046875
],
[
0.12005615234375,
0.300537109375,
0.483642578125,
0.300537109375,
0.483642578125,
0.311279296875,
0.12005615234375,
0.311279296875
],
[
0.51318359375,
0.3017578125,
0.87890625,
0.3017578125,
0.87890625,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.118408203125,
0.3154296875,
0.484375,
0.3154296875,
0.484375,
0.328857421875,
0.118408203125,
0.328857421875
],
[
0.51318359375,
0.318359375,
0.880859375,
0.318359375,
0.880859375,
0.33203125,
0.51318359375,
0.33203125
],
[
0.12005615234375,
0.330810546875,
0.48779296875,
0.330810546875,
0.48779296875,
0.344482421875,
0.12005615234375,
0.344482421875
],
[
0.51318359375,
0.334228515625,
0.880859375,
0.334228515625,
0.880859375,
0.347412109375,
0.51318359375,
0.347412109375
],
[
0.1175537109375,
0.345703125,
0.48779296875,
0.347412109375,
0.48779296875,
0.361083984375,
0.1175537109375,
0.359375
],
[
0.51318359375,
0.349853515625,
0.87890625,
0.349853515625,
0.87890625,
0.363525390625,
0.51318359375,
0.363525390625
],
[
0.12176513671875,
0.364013671875,
0.486083984375,
0.364013671875,
0.486083984375,
0.374755859375,
0.12176513671875,
0.374755859375
],
[
0.51318359375,
0.366455078125,
0.87890625,
0.366455078125,
0.87890625,
0.380126953125,
0.51318359375,
0.380126953125
],
[
0.118408203125,
0.377685546875,
0.486083984375,
0.377685546875,
0.486083984375,
0.391357421875,
0.118408203125,
0.391357421875
],
[
0.51123046875,
0.3818359375,
0.880859375,
0.3818359375,
0.880859375,
0.3955078125,
0.51123046875,
0.3955078125
],
[
0.12005615234375,
0.3955078125,
0.486083984375,
0.3955078125,
0.486083984375,
0.4091796875,
0.12005615234375,
0.4091796875
],
[
0.51123046875,
0.396728515625,
0.880859375,
0.39794921875,
0.880859375,
0.412109375,
0.51123046875,
0.410888671875
],
[
0.118408203125,
0.410888671875,
0.32080078125,
0.410888671875,
0.32080078125,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.5146484375,
0.41455078125,
0.880859375,
0.41455078125,
0.880859375,
0.427490234375,
0.5146484375,
0.427490234375
],
[
0.51318359375,
0.429931640625,
0.880859375,
0.429931640625,
0.880859375,
0.44482421875,
0.51318359375,
0.44482421875
],
[
0.12176513671875,
0.441162109375,
0.486083984375,
0.441162109375,
0.486083984375,
0.454833984375,
0.12176513671875,
0.454833984375
],
[
0.51123046875,
0.446533203125,
0.8818359375,
0.446533203125,
0.8818359375,
0.460205078125,
0.51123046875,
0.460205078125
],
[
0.118408203125,
0.457275390625,
0.489501953125,
0.457275390625,
0.489501953125,
0.47021484375,
0.118408203125,
0.47021484375
],
[
0.51123046875,
0.46435546875,
0.87890625,
0.46435546875,
0.87890625,
0.474365234375,
0.51123046875,
0.474365234375
],
[
0.118408203125,
0.47265625,
0.48779296875,
0.47265625,
0.48779296875,
0.486328125,
0.118408203125,
0.486328125
],
[
0.51318359375,
0.478515625,
0.880859375,
0.478515625,
0.880859375,
0.4921875,
0.51318359375,
0.4921875
],
[
0.118408203125,
0.4892578125,
0.264404296875,
0.4892578125,
0.264404296875,
0.5029296875,
0.118408203125,
0.5029296875
],
[
0.51318359375,
0.49462890625,
0.880859375,
0.49462890625,
0.880859375,
0.50830078125,
0.51318359375,
0.50830078125
],
[
0.51123046875,
0.51123046875,
0.880859375,
0.51123046875,
0.880859375,
0.52490234375,
0.51123046875,
0.52490234375
],
[
0.124267578125,
0.51953125,
0.486083984375,
0.51953125,
0.486083984375,
0.52978515625,
0.124267578125,
0.52978515625
],
[
0.51123046875,
0.52685546875,
0.87890625,
0.52685546875,
0.87890625,
0.54052734375,
0.51123046875,
0.54052734375
],
[
0.118408203125,
0.533203125,
0.484375,
0.533203125,
0.484375,
0.54638671875,
0.118408203125,
0.54638671875
],
[
0.51318359375,
0.5419921875,
0.8818359375,
0.5419921875,
0.8818359375,
0.5556640625,
0.51318359375,
0.5556640625
],
[
0.118408203125,
0.548828125,
0.486083984375,
0.5498046875,
0.486083984375,
0.5634765625,
0.118408203125,
0.5625
],
[
0.51318359375,
0.55810546875,
0.69775390625,
0.55859375,
0.69775390625,
0.57373046875,
0.51318359375,
0.572265625
],
[
0.1175537109375,
0.56640625,
0.47021484375,
0.5654296875,
0.47021484375,
0.580078125,
0.1175537109375,
0.58056640625
],
[
0.5205078125,
0.58740234375,
0.880859375,
0.58740234375,
0.880859375,
0.60107421875,
0.5205078125,
0.60107421875
],
[
0.1158447265625,
0.595703125,
0.3837890625,
0.5966796875,
0.3837890625,
0.61083984375,
0.1158447265625,
0.60986328125
],
[
0.51123046875,
0.6015625,
0.87890625,
0.6015625,
0.87890625,
0.615234375,
0.51123046875,
0.615234375
],
[
0.1175537109375,
0.619140625,
0.486083984375,
0.619140625,
0.486083984375,
0.6328125,
0.1175537109375,
0.6328125
],
[
0.51318359375,
0.6181640625,
0.8818359375,
0.6181640625,
0.8818359375,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.118408203125,
0.6357421875,
0.48779296875,
0.6357421875,
0.48779296875,
0.6494140625,
0.118408203125,
0.6494140625
],
[
0.51123046875,
0.63427734375,
0.880859375,
0.63427734375,
0.880859375,
0.64794921875,
0.51123046875,
0.64794921875
],
[
0.118408203125,
0.65283203125,
0.4013671875,
0.65283203125,
0.4013671875,
0.66650390625,
0.118408203125,
0.66650390625
],
[
0.51318359375,
0.65185546875,
0.87744140625,
0.65185546875,
0.87744140625,
0.662109375,
0.51318359375,
0.662109375
],
[
0.51318359375,
0.66650390625,
0.87744140625,
0.66650390625,
0.87744140625,
0.6796875,
0.51318359375,
0.6796875
],
[
0.1234130859375,
0.681640625,
0.484375,
0.681640625,
0.484375,
0.6953125,
0.1234130859375,
0.6953125
],
[
0.51318359375,
0.68310546875,
0.87890625,
0.68310546875,
0.87890625,
0.69677734375,
0.51318359375,
0.69677734375
],
[
0.118408203125,
0.69873046875,
0.484375,
0.69873046875,
0.484375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51318359375,
0.70068359375,
0.87890625,
0.70068359375,
0.87890625,
0.71142578125,
0.51318359375,
0.71142578125
],
[
0.118408203125,
0.71435546875,
0.48779296875,
0.71435546875,
0.48779296875,
0.72802734375,
0.118408203125,
0.72802734375
],
[
0.51123046875,
0.71533203125,
0.87744140625,
0.71533203125,
0.87744140625,
0.728515625,
0.51123046875,
0.728515625
],
[
0.1175537109375,
0.728515625,
0.489501953125,
0.72998046875,
0.489501953125,
0.74462890625,
0.1175537109375,
0.74365234375
],
[
0.5146484375,
0.73291015625,
0.87890625,
0.73291015625,
0.87890625,
0.74365234375,
0.5146484375,
0.74365234375
],
[
0.11419677734375,
0.74658203125,
0.133544921875,
0.74658203125,
0.133544921875,
0.76025390625,
0.11419677734375,
0.76025390625
],
[
0.51123046875,
0.74658203125,
0.87890625,
0.74658203125,
0.87890625,
0.76025390625,
0.51123046875,
0.76025390625
],
[
0.5146484375,
0.76416015625,
0.880859375,
0.76416015625,
0.880859375,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.1259765625,
0.77490234375,
0.484375,
0.775390625,
0.484375,
0.7890625,
0.1259765625,
0.7880859375
],
[
0.51318359375,
0.77978515625,
0.880859375,
0.77978515625,
0.880859375,
0.79345703125,
0.51318359375,
0.79345703125
],
[
0.118408203125,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8046875,
0.118408203125,
0.8046875
],
[
0.51318359375,
0.79638671875,
0.8818359375,
0.79638671875,
0.8818359375,
0.81005859375,
0.51318359375,
0.81005859375
],
[
0.12005615234375,
0.80859375,
0.484375,
0.80859375,
0.484375,
0.8193359375,
0.12005615234375,
0.8193359375
],
[
0.118408203125,
0.82373046875,
0.486083984375,
0.82373046875,
0.486083984375,
0.83740234375,
0.118408203125,
0.83740234375
],
[
0.53076171875,
0.82080078125,
0.880859375,
0.822265625,
0.880859375,
0.8359375,
0.53076171875,
0.83447265625
],
[
0.5146484375,
0.83447265625,
0.88330078125,
0.83447265625,
0.88330078125,
0.84716796875,
0.5146484375,
0.84716796875
],
[
0.118408203125,
0.84033203125,
0.44921875,
0.84033203125,
0.44921875,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.51318359375,
0.8466796875,
0.6533203125,
0.8466796875,
0.6533203125,
0.85986328125,
0.51318359375,
0.85986328125
],
[
0.1368408203125,
0.85693359375,
0.486083984375,
0.85693359375,
0.486083984375,
0.87060546875,
0.1368408203125,
0.87060546875
],
[
0.5341796875,
0.85791015625,
0.880859375,
0.85791015625,
0.880859375,
0.87158203125,
0.5341796875,
0.87158203125
],
[
0.118408203125,
0.87353515625,
0.484375,
0.87353515625,
0.484375,
0.88720703125,
0.118408203125,
0.88720703125
],
[
0.5146484375,
0.87255859375,
0.87890625,
0.87255859375,
0.87890625,
0.8828125,
0.5146484375,
0.8828125
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.5146484375,
0.88427734375,
0.87744140625,
0.88427734375,
0.87744140625,
0.89453125,
0.5146484375,
0.89453125
],
[
0.51318359375,
0.8935546875,
0.880859375,
0.89453125,
0.880859375,
0.9091796875,
0.51318359375,
0.90771484375
],
[
0.118408203125,
0.90625,
0.382080078125,
0.90625,
0.382080078125,
0.91943359375,
0.118408203125,
0.91943359375
],
[
0.51318359375,
0.90625,
0.75390625,
0.90625,
0.75390625,
0.91943359375,
0.51318359375,
0.91943359375
]
] |
[
"代理-患者交换）。",
"3Dataset Creation",
"CircumstanceError（CircE）：除了核心参数",
"除了理论基础之外，我们还通过对CNN/DM",
"之外，还可以使用描述参数和谓词交互的情况",
"数据集上的五个抽象摘要模型和XSum数据集",
"的附加信息或属性来进一步指定谓词（例如。",
"上的四个抽象摘要模型的大规模人工注释，对",
"g。位置、时间、方式、方向、模态)。类别CircE",
"我们的类型学进行了经验验证。通过我们的数",
"捕获一个或多个此类属性（顿中的非核心顿元",
"据集，我们的目标是广泛覆盖神经汇总系统所",
"素）错误的错误",
"产生的不同类型的错误，并对其细粒度的事实",
"错误进行人为判断",
"2.2Discourse Errors",
"AnnotationData对于注释，我们包括来自",
"作者的交际意图也通过文本各部分之间的关",
"CNN/DM和XSum数据集的模型摘要，因为它",
"系来表达。摘要文本中的事实错误通常可以",
"们呈现不同的特征。CNN/DM摘要较长，平均",
"扩展到单个语义框架之外，从而在话语片段之",
"三句话，而xSum只有一句话摘要。较长的摘",
"间引入错误的链接。下面我们概述了基于话语",
"要对于识别话语级别的错误至关重要。另一方",
"分析和修辞结构理论（RST）（BrownandYule,",
"面，XSum摘要更加抽象，并且平均（Maynez",
"1983;MannandThompson，1988）的此类错误。",
"etal.,2020)包含更多的事实错误。对于一组不",
"RST是一个用于注释语篇中连贯关系的复杂",
"同的模型摘要，我们从具有不同事实能力的不",
"系统。这种关系的一些例子包括：“阐述”，“背",
"同摘要模型中收集公开可用的模型输出。对于",
"景”，“动机”和“意志原因”。在这里，我们脱",
"CNN/DM数据集，我们使用来自LSTM序列",
"离了语义框架术语，因为它植根于单个框架中",
"到序列模型（S2S）（Rushetal.2015），指针生",
"不允许我们表示此类错误",
"成器网络（PGN）模型（Seeetal.，2017），自下",
"而上总结（BUS）模型（Gehrmannetal，2018），",
"CoreferenceError（CorefE)：类别CorefE说",
"基于Bert的提取抽象模型（BertSum）（Liuand",
"明了错误，其中代词和对先前提到的实体的其",
"Lapata，2019和联合训练的预训练编码器的模",
"他类型的引用不正确或没有明确的前置词，从",
"型。对于XSum数据集，我们从主题感知CNN",
"而使它们模棱两可",
"模型（Narayanetal.，2018），指针生成器网络",
"（PGN）模型，随机初始化（TransS2S）（Vaswani",
"DiscourseLinkError（LinkE):类别LinkE",
"et al.,2017)和用Bert-Base（BertS2S）(Devlin",
"包含涉及不同语句之间的话语链接的错误。这",
"etal.,2019)初始化的模型输出。2sA.1中提供",
"些包括错误的时间顺序或错误的话语链接（例",
"了所用模型的详细信息",
"如。g。RST关系、话语连接器）在语句之间",
"AnnotationCollection使用上述模型生成的",
"2.3Content Verifiability Errors",
"摘要，我们从三个独立的注释器收集来自每个",
"通常，由于难以将摘要中的语句与源文本对",
"数据集的250篇文章的人类注释（CNN/DM上",
"齐，因此无法根据源文本对其进行验证。下面",
"共有1250个模型输出，XSum上有1000个）。",
"我们概述了这种情况下的两类错误。",
"我们对摘要的每个后子进行注释，以将对事实",
"的判断分解为较小的单位。我们在整个摘要的",
"OutofArticleError（OutE)：由于文档摘要",
"上下文中呈现句子，以识别跨越多个句子的话",
"应仅包含可以从原始文本中推断出的信息：因",
"语错误。注释是一个两步过程：对于摘要中的",
"此我们为此类错误OutE添加了一个类别（先前",
"每个句子，注释者首先选择句子是否为事实",
"的工作将其称为外在幻觉（(Maynezetal.，2020)",
"如果标记为非事实，则根据我们的类型学确",
")",
"定每个错误的类别。3句子可以用多个错误类",
"别进行注释，以解决句子中的多个错误。我们",
"GrammaticalError（GramE)：我们使用",
"在AmazonMechanicalTurk(MTurk）平台上执行",
"GramE对格式不正确的语句进行分类。当语",
"注释任务。为了实现高质量的众包注释，我们",
"法错误使陈述的含义难以理解或模棱两可时，",
"无法根据来源进行验证，因此被认为是微不足",
"2As we use publicly available model outputs, the sum",
"maries across different datasets are from different models ow-",
"道的错误。轻微的语法错误是可以接受的",
"ing to their availability.",
"最后，为了在注释练习中保持完整性，我们",
"3We experimented with Likert scale evaluation of full",
"为与上述任何类别都不对应的事实错误添加了",
"summaries in a pilot study.Such an annotation would not",
"另外两个类别Others（OthE），为不包含任何错",
"provideprecise information about wherein the summary ar",
"error appears and also resulted in lower agreement. Hence,",
"误的语句添加了NotanError(NE）",
"we opted for sentence level judgements."
] |
{
"arxiv_id": "2104.13346",
"language": "zh",
"page_idx": 3
}
|
||
2104.13346-zh-4
|
2104.13346-zh-4.jpg
|
<FigureHere>
Figure 2: Proportion of summaries with factual errors based on collected annotations, with breakdown of the
categories of errors within. Full specification of categories of errors in Table 1 .
构建了一个直观的界面$^{4}$,它结合了：
$\kappa=0.86$的 CohenKappa 表示几乎完美的协议。先前的工作在三个人群注释者和事实 (Falke et al., 2019) 的专家注释之间找到了$\kappa=0.65$ 的协议。即使有 9 名以上的工人，他们也报告了最多$\kappa=0.74$的专家注释。。此改进验证了我们的注释接口和协议的健壮性，从而以更少的工人实现了更高的协议
### 4 Summarization Model Analvsis
1. Clear Instructions: We explain the annotation
scheme without assuming linguistic knowledge and give several examples for each cate- $\gcd$gory.
2. Training and Evaluation: We setup training
tutorials for first time users to train and provide feedback on the task. We also setup a qualification test which tests their understanding of our annotation scheme and require annotators to obtain >85 % score to qualify. Further, we continuously evaluate annotators during the task against artificially generated factual errors to ensure continued high quality.
3. Fair Pay and Bonus: All workers are paid 50
% more than the average American minimum wage. We offer bonuses for scores of 60% or above on the continuous evaluation, and for completing sets of 10 annotations.
有关我们接口的更多详细信息，请参见§A.6
我们根据真实性评估不同摘要模型的性能。Figure 2 可视化了每个类别模型和数据集的具有事实错误的摘要的百分比，并细分了每个类别中不同错误类型的比例。如果摘要至少包含一个带有事实错误的句子，则该摘要被认为是不正确的。如果大多数注释者表示存在错误， 则句子包含事实错误(这里我们不考虑所有三个注释者在类别上都不同意的注释)
Inter-Annotator Agreement: 我们根据 FleissKappa $\kappa$ (Fleiss, 1971) 报告注释者之间的协议。在 Durmus et al.(2020)之后，我们报告与多数类一致的注释者的$p$百分比。我们数据集中的每个数据点对应于摘要中的一个句子。我们计算所有 4942 个带注释的句子的一致性。关于句子是否为事实的注释，我们获得$\kappa=0.58$ 其中注释者的$p=91\%$与多数类一致。作为比较，Durmus et al.(2020)报告$p=76.7\%$平均协议。当所有三个注释者都同意一个句子不是事实时，我们获得$\kappa=0.39$和$p=73.9\%$的注释者同意八类注释(七类错误和“其他”)上的多数类，这表示适度的协议
How factual are generated summaries across different datasets? 从我们的注释中，我们观察到注释的摘要中的 60 个% 至少包含一个事实错误。从 Figure 2 中，我们看到 XSum 数据集具有比 CNN/DM (43%) 更多的事实错误模型摘要 (92%)。它在真实性方面带来了更大的挑战，因为所有模型都会产生 80 个带有事实错误的% 摘要，而最佳模型(BertS2S)会产生83 个% 错误摘要。在 CNN/DM 数据集上，虽然像 BERTSum 和 BART 这样的最先进的预训练模型具有更好的事实性数字，但事实不正确的摘要的百分比仍然很高(BERTSum为23%, BART 为 27$\%$ 。不同类别的错误比例在两个数据集之间差异很大。对于 CNN/DM 数据集， 最常见的错误类别是实体错误(EntE)和引用错误( CorefE )。对于 XSum 数据集，它们是OutofArticleError( OutE ) 和 EntityError( EntE )。请注意，XSum 数据集中没有话语错误(CorefE , LinkE ), 因为数据仅包含单个句子摘要。此外，我们观察到 OthE 占错误总数的很小百分比(
Agreement with Domain Expert: 我们在 201个数据点 (CNN/DM 的 10 个摘要和 XSum 的10 个摘要) 上测量了三个注释者的多数类与一个专家注释者之间的一致性。我们发现
We make the interface available for future human anno-
tations that follow our typology
|
[
[
0.276123046875,
0.0843505859375,
0.330078125,
0.0843505859375,
0.330078125,
0.09857177734375,
0.276123046875,
0.09857177734375
],
[
0.59033203125,
0.08197021484375,
0.63037109375,
0.0843505859375,
0.62890625,
0.09979248046875,
0.58837890625,
0.097412109375
],
[
0.1502685546875,
0.10272216796875,
0.1796875,
0.10272216796875,
0.1796875,
0.11517333984375,
0.1502685546875,
0.11517333984375
],
[
0.4375,
0.105712890625,
0.478515625,
0.105712890625,
0.478515625,
0.11639404296875,
0.4375,
0.11639404296875
],
[
0.75390625,
0.10809326171875,
0.76171875,
0.10809326171875,
0.76171875,
0.11102294921875,
0.75390625,
0.11102294921875
],
[
0.76904296875,
0.10272216796875,
0.8505859375,
0.10394287109375,
0.8505859375,
0.114013671875,
0.76904296875,
0.1134033203125
],
[
0.75390625,
0.1175537109375,
0.76171875,
0.1175537109375,
0.76171875,
0.12054443359375,
0.75390625,
0.12054443359375
],
[
0.771484375,
0.114013671875,
0.82958984375,
0.114013671875,
0.82958984375,
0.12469482421875,
0.771484375,
0.12469482421875
],
[
0.7734375,
0.12469482421875,
0.8623046875,
0.12469482421875,
0.8623046875,
0.1318359375,
0.7734375,
0.1318359375
],
[
0.14697265625,
0.1181640625,
0.173828125,
0.1181640625,
0.173828125,
0.1318359375,
0.14697265625,
0.1318359375
],
[
0.4306640625,
0.1270751953125,
0.478515625,
0.1270751953125,
0.478515625,
0.13720703125,
0.4306640625,
0.13720703125
],
[
0.75,
0.1258544921875,
0.76318359375,
0.1258544921875,
0.76318359375,
0.1318359375,
0.75,
0.1318359375
],
[
0.1485595703125,
0.1353759765625,
0.17626953125,
0.1353759765625,
0.17626953125,
0.1484375,
0.1485595703125,
0.1484375
],
[
0.75146484375,
0.135986328125,
0.76171875,
0.135986328125,
0.76171875,
0.1395263671875,
0.75146484375,
0.1395263671875
],
[
0.771484375,
0.133056640625,
0.86376953125,
0.133056640625,
0.86376953125,
0.14013671875,
0.771484375,
0.14013671875
],
[
0.771484375,
0.142578125,
0.85205078125,
0.142578125,
0.85205078125,
0.152587890625,
0.771484375,
0.152587890625
],
[
0.455078125,
0.15087890625,
0.47607421875,
0.15087890625,
0.47607421875,
0.157958984375,
0.455078125,
0.157958984375
],
[
0.771484375,
0.152587890625,
0.861328125,
0.152587890625,
0.861328125,
0.163330078125,
0.771484375,
0.163330078125
],
[
0.771484375,
0.163330078125,
0.85400390625,
0.163330078125,
0.85400390625,
0.17041015625,
0.771484375,
0.17041015625
],
[
0.1234130859375,
0.15380859375,
0.173828125,
0.15380859375,
0.173828125,
0.1650390625,
0.1234130859375,
0.1650390625
],
[
0.1435546875,
0.17041015625,
0.173828125,
0.17041015625,
0.173828125,
0.182861328125,
0.1435546875,
0.182861328125
],
[
0.43408203125,
0.171630859375,
0.476806640625,
0.171630859375,
0.476806640625,
0.1787109375,
0.43408203125,
0.1787109375
],
[
0.76318359375,
0.17041015625,
0.8193359375,
0.17041015625,
0.8193359375,
0.18115234375,
0.76318359375,
0.18115234375
],
[
0.166259765625,
0.18701171875,
0.443359375,
0.18701171875,
0.443359375,
0.20068359375,
0.166259765625,
0.20068359375
],
[
0.470947265625,
0.1859130859375,
0.74658203125,
0.18701171875,
0.74658203125,
0.20068359375,
0.470947265625,
0.1995849609375
],
[
0.1175537109375,
0.215576171875,
0.880859375,
0.215576171875,
0.880859375,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.1175537109375,
0.2298583984375,
0.6484375,
0.2298583984375,
0.6484375,
0.243408203125,
0.1175537109375,
0.243408203125
],
[
0.1175537109375,
0.268310546875,
0.4189453125,
0.26953125,
0.4189453125,
0.283935546875,
0.1175537109375,
0.283203125
],
[
0.51318359375,
0.268310546875,
0.880859375,
0.26953125,
0.880859375,
0.283935546875,
0.51318359375,
0.283203125
],
[
0.1318359375,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.1318359375,
0.2998046875
],
[
0.51318359375,
0.2861328125,
0.87744140625,
0.2861328125,
0.87744140625,
0.29638671875,
0.51318359375,
0.29638671875
],
[
0.15283203125,
0.300537109375,
0.48779296875,
0.300537109375,
0.48779296875,
0.314208984375,
0.15283203125,
0.314208984375
],
[
0.51318359375,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.154541015625,
0.318359375,
0.490234375,
0.318359375,
0.490234375,
0.33203125,
0.154541015625,
0.33203125
],
[
0.51318359375,
0.318359375,
0.880859375,
0.318359375,
0.880859375,
0.33203125,
0.51318359375,
0.33203125
],
[
0.154541015625,
0.33740234375,
0.19140625,
0.33740234375,
0.19140625,
0.3486328125,
0.154541015625,
0.3486328125
],
[
0.515625,
0.333251953125,
0.8759765625,
0.333251953125,
0.8759765625,
0.3466796875,
0.515625,
0.3466796875
],
[
0.130126953125,
0.350830078125,
0.48779296875,
0.350830078125,
0.48779296875,
0.364013671875,
0.130126953125,
0.364013671875
],
[
0.5146484375,
0.350830078125,
0.87890625,
0.350830078125,
0.87890625,
0.364013671875,
0.5146484375,
0.364013671875
],
[
0.154541015625,
0.36767578125,
0.490234375,
0.36767578125,
0.490234375,
0.38134765625,
0.154541015625,
0.38134765625
],
[
0.51318359375,
0.366455078125,
0.71533203125,
0.366455078125,
0.71533203125,
0.380126953125,
0.51318359375,
0.380126953125
],
[
0.154541015625,
0.3818359375,
0.48779296875,
0.383056640625,
0.48779296875,
0.396728515625,
0.154541015625,
0.3955078125
],
[
0.15283203125,
0.3984375,
0.489501953125,
0.39794921875,
0.489501953125,
0.412109375,
0.15283203125,
0.413330078125
],
[
0.51708984375,
0.40673828125,
0.5283203125,
0.40673828125,
0.5283203125,
0.412109375,
0.51708984375,
0.412109375
],
[
0.5458984375,
0.40380859375,
0.81787109375,
0.40380859375,
0.81787109375,
0.41748046875,
0.5458984375,
0.41748046875
],
[
0.15283203125,
0.4150390625,
0.486083984375,
0.4150390625,
0.486083984375,
0.4287109375,
0.15283203125,
0.4287109375
],
[
0.15283203125,
0.43115234375,
0.489501953125,
0.43115234375,
0.489501953125,
0.44482421875,
0.15283203125,
0.44482421875
],
[
0.515625,
0.43701171875,
0.8759765625,
0.43701171875,
0.8759765625,
0.44775390625,
0.515625,
0.44775390625
],
[
0.154541015625,
0.44775390625,
0.489501953125,
0.44775390625,
0.489501953125,
0.46142578125,
0.154541015625,
0.46142578125
],
[
0.51318359375,
0.451904296875,
0.87744140625,
0.451904296875,
0.87744140625,
0.465576171875,
0.51318359375,
0.465576171875
],
[
0.154541015625,
0.463134765625,
0.48779296875,
0.463134765625,
0.48779296875,
0.476806640625,
0.154541015625,
0.476806640625
],
[
0.5146484375,
0.468505859375,
0.880859375,
0.468505859375,
0.880859375,
0.482177734375,
0.5146484375,
0.482177734375
],
[
0.154541015625,
0.478515625,
0.47607421875,
0.478515625,
0.47607421875,
0.4921875,
0.154541015625,
0.4921875
],
[
0.5146484375,
0.48388671875,
0.880859375,
0.48388671875,
0.880859375,
0.49755859375,
0.5146484375,
0.49755859375
],
[
0.130126953125,
0.495361328125,
0.486083984375,
0.495361328125,
0.486083984375,
0.5087890625,
0.130126953125,
0.5087890625
],
[
0.52490234375,
0.50048828125,
0.880859375,
0.50048828125,
0.880859375,
0.51416015625,
0.52490234375,
0.51416015625
],
[
0.15283203125,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.525390625,
0.15283203125,
0.525390625
],
[
0.51318359375,
0.5166015625,
0.87744140625,
0.5166015625,
0.87744140625,
0.52978515625,
0.51318359375,
0.52978515625
],
[
0.154541015625,
0.52783203125,
0.486083984375,
0.52783203125,
0.486083984375,
0.54150390625,
0.154541015625,
0.54150390625
],
[
0.51123046875,
0.53076171875,
0.87744140625,
0.5322265625,
0.87744140625,
0.54638671875,
0.51123046875,
0.5458984375
],
[
0.1561279296875,
0.54443359375,
0.484375,
0.54443359375,
0.484375,
0.5546875,
0.1561279296875,
0.5546875
],
[
0.5146484375,
0.548828125,
0.796875,
0.548828125,
0.796875,
0.5625,
0.5146484375,
0.5625
],
[
0.154541015625,
0.56005859375,
0.405517578125,
0.56005859375,
0.405517578125,
0.57373046875,
0.154541015625,
0.57373046875
],
[
0.1368408203125,
0.57666015625,
0.48779296875,
0.57666015625,
0.48779296875,
0.59033203125,
0.1368408203125,
0.59033203125
],
[
0.51904296875,
0.58056640625,
0.8818359375,
0.5830078125,
0.8818359375,
0.59716796875,
0.51904296875,
0.595703125
],
[
0.1234130859375,
0.6015625,
0.489501953125,
0.6015625,
0.489501953125,
0.615234375,
0.1234130859375,
0.615234375
],
[
0.51318359375,
0.5986328125,
0.880859375,
0.5986328125,
0.880859375,
0.6123046875,
0.51318359375,
0.6123046875
],
[
0.118408203125,
0.6181640625,
0.489501953125,
0.6181640625,
0.489501953125,
0.6318359375,
0.118408203125,
0.6318359375
],
[
0.51123046875,
0.6142578125,
0.87890625,
0.6142578125,
0.87890625,
0.62744140625,
0.51123046875,
0.62744140625
],
[
0.118408203125,
0.63427734375,
0.486083984375,
0.63427734375,
0.486083984375,
0.6484375,
0.118408203125,
0.6484375
],
[
0.51318359375,
0.63134765625,
0.880859375,
0.63134765625,
0.880859375,
0.64453125,
0.51318359375,
0.64453125
],
[
0.118408203125,
0.65087890625,
0.486083984375,
0.65087890625,
0.486083984375,
0.66455078125,
0.118408203125,
0.66455078125
],
[
0.51318359375,
0.646484375,
0.880859375,
0.646484375,
0.880859375,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.1175537109375,
0.6650390625,
0.483642578125,
0.6650390625,
0.483642578125,
0.6787109375,
0.1175537109375,
0.6787109375
],
[
0.51123046875,
0.662109375,
0.880859375,
0.6630859375,
0.880859375,
0.677734375,
0.51123046875,
0.6767578125
],
[
0.118408203125,
0.68310546875,
0.483642578125,
0.68310546875,
0.483642578125,
0.69677734375,
0.118408203125,
0.69677734375
],
[
0.51318359375,
0.6796875,
0.880859375,
0.6796875,
0.880859375,
0.693359375,
0.51318359375,
0.693359375
],
[
0.12005615234375,
0.69873046875,
0.484375,
0.69873046875,
0.484375,
0.7119140625,
0.12005615234375,
0.7119140625
],
[
0.5146484375,
0.6953125,
0.880859375,
0.6953125,
0.880859375,
0.708984375,
0.5146484375,
0.708984375
],
[
0.12005615234375,
0.71435546875,
0.48779296875,
0.71435546875,
0.48779296875,
0.728515625,
0.12005615234375,
0.728515625
],
[
0.509765625,
0.71044921875,
0.87890625,
0.71044921875,
0.87890625,
0.7236328125,
0.509765625,
0.7236328125
],
[
0.118408203125,
0.72998046875,
0.48779296875,
0.72998046875,
0.48779296875,
0.74462890625,
0.118408203125,
0.74462890625
],
[
0.509765625,
0.7255859375,
0.880859375,
0.72705078125,
0.880859375,
0.7412109375,
0.509765625,
0.74072265625
],
[
0.12005615234375,
0.74658203125,
0.486083984375,
0.74658203125,
0.486083984375,
0.76025390625,
0.12005615234375,
0.76025390625
],
[
0.51123046875,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.51123046875,
0.75732421875
],
[
0.12005615234375,
0.76318359375,
0.486083984375,
0.76318359375,
0.486083984375,
0.77685546875,
0.12005615234375,
0.77685546875
],
[
0.51318359375,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.118408203125,
0.779296875,
0.486083984375,
0.779296875,
0.486083984375,
0.7919921875,
0.118408203125,
0.7919921875
],
[
0.51123046875,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1158447265625,
0.79345703125,
0.358642578125,
0.79443359375,
0.358642578125,
0.80859375,
0.1158447265625,
0.80810546875
],
[
0.5146484375,
0.79150390625,
0.880859375,
0.79150390625,
0.880859375,
0.8056640625,
0.5146484375,
0.8056640625
],
[
0.5146484375,
0.80810546875,
0.87890625,
0.80810546875,
0.87890625,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.1234130859375,
0.82080078125,
0.486083984375,
0.82080078125,
0.486083984375,
0.8349609375,
0.1234130859375,
0.8349609375
],
[
0.5146484375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.118408203125,
0.8359375,
0.486083984375,
0.8359375,
0.486083984375,
0.849609375,
0.118408203125,
0.849609375
],
[
0.51318359375,
0.84033203125,
0.8818359375,
0.84033203125,
0.8818359375,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.1175537109375,
0.8525390625,
0.484375,
0.8525390625,
0.484375,
0.8662109375,
0.1175537109375,
0.8662109375
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.118408203125,
0.869140625,
0.486083984375,
0.869140625,
0.486083984375,
0.8828125,
0.118408203125,
0.8828125
],
[
0.51708984375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51708984375,
0.8857421875
],
[
0.1376953125,
0.89111328125,
0.48779296875,
0.8935546875,
0.48779296875,
0.90771484375,
0.1376953125,
0.90625
],
[
0.51123046875,
0.88720703125,
0.880859375,
0.88818359375,
0.880859375,
0.90283203125,
0.51123046875,
0.90185546875
],
[
0.1158447265625,
0.90380859375,
0.30908203125,
0.90625,
0.30908203125,
0.91943359375,
0.1158447265625,
0.91748046875
],
[
0.51318359375,
0.90380859375,
0.5517578125,
0.90380859375,
0.5517578125,
0.91943359375,
0.51318359375,
0.91943359375
]
] |
[
"CNN/DM",
"xsum",
"S2S",
"TranS2S",
"",
"PredE: Predicate",
"",
"EntE: Entity",
"CircE:Circumstanc",
"PGN",
"TConvS2s",
"",
"BUS",
"",
"CorefE:Coreference",
"LinkE: Connector",
"PGN",
"OutE: Not in article",
"GramE:Grammar",
"BERTSum",
"BART",
"BERTS2S",
"OthE:Oth",
"0.00.10.20.30.40.50.60.70.80.91.0",
"0.00.10.20.30.40.50.60.70.80.91.0",
"Figure 2: Proportion of summaries with factual errors based on collected annotations, with breakdown of the",
"categories of errors within. Full specification of categories of errors in Table 1",
"构建了一个直观的界面4，它结合了：",
"K=0.86的CohenKappa表示几乎完美的协议。",
"1. Clear Instructions: We explain the annotation",
"先前的工作在三个人群注释者和事实（Falke",
"scheme without assuming linguistic knowl-",
"etal.,2019）的专家注释之间找到了k=0.65",
"edge and give several examples for each cate-",
"的协议。即使有9名以上的工人，他们也报告",
"gory",
"了最多K=0.74的专家注释。。此改进验证了",
"2. Training and Evaluation: We setup training",
"我们的注释接口和协议的健壮性，从而以更少",
"tutorials for first time users to train and pro-",
"的工人实现了更高的协议",
"vide feedback on the task. We also setup a",
"qualification test which tests their understand-",
"",
"Summarization Model Analysis",
"ing of our annotation scheme and require an",
"notators to obtain >85 % score to qualify. Fur-",
"我们根据真实性评估不同摘要模型的性能",
"ther, we continuously evaluate annotators dur-",
"Figure2可视化了每个类别模型和数据集的具",
"ing the task against artificially generated fac-",
"有事实错误的摘要的百分比，并细分了每个类",
"tual errors to ensure continued high quality.",
"别中不同错误类型的比例。如果摘要至少包含",
"3. Fair Pay and Bonus: All workers are paid 50",
"个带有事实错误的句子，则该摘要被认为是",
"% more than the average American minimum",
"不正确的。如果大多数注释者表示存在错误",
"wage. We offer bonuses for scores of 60 % or",
"则句子包含事实错误（这里我们不考虑所有三",
"above on the continuous evaluation. and for",
"个注释者在类别上都不同意的注释)",
"completing sets of 10 annotations.",
"有关我们接口的更多详细信息，请参见8A.6",
"How factual are generated summaries across",
"Inter-AnnotatorAgreement：我们根据Fleis-",
"differentdatasets？从我们的注释中，我们观",
"sKappak(Fleiss，1971)报告注释者之间的协议。",
"察到注释的摘要中的60个%至少包含一个事",
"在Durmusetal.（2020）之后，我们报告与多数",
"实错误。从Figure2中，我们看到XSum数据",
"类一致的注释者的p百分比。我们数据集中的",
"集具有比CNN/DM（43%）更多的事实错误模",
"每个数据点对应于摘要中的一个句子。我们计",
"型摘要（92%）。它在真实性方面带来了更大的",
"算所有4942个带注释的句子的一致性。关于",
"挑战，因为所有模型都会产生80个带有事实",
"句子是否为事实的注释，我们获得K=0.58",
"错误的%摘要，而最佳模型（BertS2S）会产生",
"其中注释者的p=91%与多数类一致。作为比",
"83个%错误摘要。在CNN/DM数据集上，虽",
"较，Durmusetal.(2020）报告p=76.7%平均",
"然像BERTSum和BART这样的最先进的预训",
"协议。当所有三个注释者都同意一个句子不是",
"练模型具有更好的事实性数字，但事实不正确",
"事实时，我们获得K=0.39和p=73.9%的注",
"的摘要的百分比仍然很高（BERTSum为23%，",
"释者同意八类注释（七类错误和“其他”）上",
"BART为27%）。不同类别的错误比例在两个",
"的多数类，这表示适度的协议",
"数据集之间差异很大。对于CNN/DM数据集，",
"最常见的错误类别是实体错误（EntE）和引用",
"Agreement with Domain Expert:我们在 201",
"错误（CorefE）。对于XSum数据集，它们是",
"个数据点（CNN/DM的10个摘要和XSum的",
"OutofArticleError(OutE）和EntityError(EntE)。",
"10个摘要）上测量了三个注释者的多数类",
"请注意，XSum数据集中没有话语错误（CorefE",
"与一个专家注释者之间的一致性。我们发现",
"，LinkE），因为数据仅包含单个句子摘要。此",
"+we make the interface available for future human anno-",
"外，我们观察到OthE占错误总数的很小百分",
"tations that follow our typology",
"比（"
] |
{
"arxiv_id": "2104.13346",
"language": "zh",
"page_idx": 4
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.