id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2209.07430-zh-4
|
2209.07430-zh-4.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>Example</th>
<th>Step</th>
<th>Relevant Spans</th>
</tr>
<tr>
<td rowspan="3">Comparison</td>
<td>Context: Blind Shaft is a 2003 film about a pair of brutal con artists operating in the illegal coal mines of present day northern</td>
<td>Interpreting the question</td>
<td>came out relation: <film, release date> film entities: Blind Shaft, The Mask Of Fu Manchu earlier: date comparison $\operatorname{target}:$ $min($ ${(\:\mathrm{release}}$ date Blind Shaft , release date The Mask of Fu Manchu )</td>
</tr>
<tr>
<td>China. The Mask Of Fu Manchu is a 1932 pre- Code adventure film directed by Charles Brabin. Question: Which film came out earlier, Blind Shaft or The Mask OfFuManchu? </td>
<td>Identifying relevant information through referential equality</td>
<td>$\mathrm{Blind~Shaft~}_{\mathrm{q}}:=\mathrm{Blind~Shaft~}_{\mathrm{c}}$ The Mask Of Fu Manchu $_\mathrm{q}:=$ The Mask Of Fu Manchu c . came out $_{\mathrm{q}}:=<$date, film> construction c release dates: < Blind Shaft , 2003>, < The Mask O $FuManchu,1932>$</td>
</tr>
<tr>
<td>Answer: I ne Nlask Or Fu Manchu</td>
<td>Value comparison</td>
<td>solution: earlier $_\mathrm{q}:=\min_{\mathrm{c}}$ $min(1932,2003)=1932$</td>
</tr>
<tr>
<td rowspan="3">Coreference</td>
<td>Context: Barack Obama was the 44th president of the US. He was</td>
<td>Interpreting the question</td>
<td>born relation: <person, location> Hawaii:location target: born:<Hawaii ,UNK></td>
</tr>
<tr>
<td>born in Hawaii. Question: Who was born in Hawaii? 1.7 $n$</td>
<td>$\operatorname{Identifying}$ relevant information through referential equality</td>
<td>$\mathrm{Hawaii}_{\mathrm{q}}:=\mathrm{Hawaii}_{\mathrm{c}}$ born relation:<he, Hawaii></td>
</tr>
<tr>
<td>$\boldsymbol{UL}$ ${\mathrm{ramu.}}$</td>
<td>Coref. resolution</td>
<td><Barack Obama, he> solution: born<Barack Obama. Hawaii></td>
</tr>
</tbody>
</table>
Table 1: The basic reasoning steps for answering comparison and coreference questions.
instances in that dataset for which different kinds of explana-
tions align (according to our Def. 2 ).
我们将高对齐分数解释为两种解释都是忠实的证据，并且模型确实执行了预期的推理步骤
# 3 Methodology
3.1 Datasets and Models
对于 Coreference, 我们使用 Quoref E(Dasigi et al., 2019)数据集 (20K 训练和 2。4K 验证实例), 其中要求注释者为给定文本设计问题， 以便回答这些问题需要解决回指。为了进行比较，我们对 HotpotQA (Yang et al.,2018) 和2WikiMultiHopQA (Ho et al., 2020) 中的问题进行了采样：两个数据集的问题用它们的推理类型(桥接或比较)进行了手动注释。我们选择包含比较形容词或副词的“比较”问题(23K 训练，3K 验证实例)。这些资源基于维基百科， 并有多个段落作为上下文，但回答问题所需的句子(通常为 2-3)被标记为“支持事实”。由于我们不关注多跳信息检索技能，因此我们将上下文限制为这些句子
我们使用 BERT 系列的五个预先训练的基于 Transformer 的编码器进行实验：RoBERTa$_\mathrm{large}$E(Liu et al.,2019) , $\mathrm{BERT}_{\text{large- cased}}$, $\mathrm{BERT}_{\text{base- cased}}\underline {{\mathrm{E} }}( \underline {{\mathrm{Devlin~et~al. , }}}$ 2019), BERT$_\text{medium}$和 BERT$_\text{small}\boxed{\mathrm{E}}($Turc et al., 2019; Bhargava et al., 2021)。这些 BERT 模型的主要区别在于架构块的结构和参数的数量， 而 RoBERTa 也有不同的训练语料库和优化。
由于较大的模型被证明可以更好地概括某些用例 (Hendrycks et al., 2020; Bhargava et al., 2021) ,我们调查它们是否也更有可能是正确的原因我们 fine-tune 每个编码器使用 Devlin et al.
(2019)中的体系结构 (有关详细信息，请参见附录),并在验证集上对其进行评估(因为测试集不是公开的)。我们在提取 QA 中使用标准评估指标：F1-Score (预测和“黄金”答案之间的令牌重叠百分比，在所有数据点上平均) 和 Exact-match (预测答案与“黄金”答案匹配的数据点数量)
## 3.2 Counterfactual Explanations
我们的推理公式(Table 1)包括三个用于共同引用和比较的基本步骤：解释问题，通过引用相等来识别相关信息以及对所识别信息的目标操作 (共同引用解析或值比较)。我们专注于最后一步，因为：(a)它隐式地依赖于对问题和上下文的正确语义解析；(b)我们数据中的引用相等在很大程度上是微不足道的：大多数实体在问题和文本中具有相同的表面形式
一个明显的语义上有效的扰动应该改变预测(从而测试模型对比较操作的理解)是将比较形容词替换为其反义词(Figure 3d)。由于我们的示例仅包含 6 个用作比较运算符的令牌， 因此我们手动定义适当的替换。$^{8}$
对于共同引用问题，有能力的 RC 模型至少
$^{8}$earlier$\leftrightarrow$later, first$\to$later, more recently$\to$earlier,
older$\leftrightarrow$younger.
|
[
[
0.16796875,
0.0872802734375,
0.2275390625,
0.0872802734375,
0.2275390625,
0.1015625,
0.16796875,
0.1015625
],
[
0.39208984375,
0.0885009765625,
0.4248046875,
0.0885009765625,
0.4248046875,
0.1009521484375,
0.39208984375,
0.1009521484375
],
[
0.53564453125,
0.0872802734375,
0.63037109375,
0.0885009765625,
0.63037109375,
0.09979248046875,
0.53564453125,
0.09857177734375
],
[
0.39208984375,
0.10272216796875,
0.470947265625,
0.1051025390625,
0.470947265625,
0.119384765625,
0.39208984375,
0.1175537109375
],
[
0.496337890625,
0.105712890625,
0.5224609375,
0.105712890625,
0.5224609375,
0.1175537109375,
0.496337890625,
0.1175537109375
],
[
0.5341796875,
0.1051025390625,
0.771484375,
0.1051025390625,
0.771484375,
0.1181640625,
0.5341796875,
0.1181640625
],
[
0.393798828125,
0.119384765625,
0.44921875,
0.119384765625,
0.44921875,
0.1300048828125,
0.393798828125,
0.1300048828125
],
[
0.53662109375,
0.1181640625,
0.85400390625,
0.1181640625,
0.85400390625,
0.12890625,
0.53662109375,
0.12890625
],
[
0.53662109375,
0.1312255859375,
0.68603515625,
0.1312255859375,
0.68603515625,
0.141357421875,
0.53662109375,
0.141357421875
],
[
0.172119140625,
0.1217041015625,
0.364501953125,
0.1217041015625,
0.364501953125,
0.1353759765625,
0.172119140625,
0.1353759765625
],
[
0.16796875,
0.1341552734375,
0.3525390625,
0.1341552734375,
0.3525390625,
0.1478271484375,
0.16796875,
0.1478271484375
],
[
0.60107421875,
0.142578125,
0.7021484375,
0.142578125,
0.7021484375,
0.15380859375,
0.60107421875,
0.15380859375
],
[
0.7109375,
0.141357421875,
0.8037109375,
0.144287109375,
0.8037109375,
0.1561279296875,
0.7109375,
0.152587890625
],
[
0.8095703125,
0.142578125,
0.85791015625,
0.142578125,
0.85791015625,
0.152587890625,
0.8095703125,
0.152587890625
],
[
0.53662109375,
0.1436767578125,
0.5810546875,
0.1436767578125,
0.5810546875,
0.15380859375,
0.53662109375,
0.15380859375
],
[
0.16796875,
0.144287109375,
0.376220703125,
0.144287109375,
0.376220703125,
0.157958984375,
0.16796875,
0.157958984375
],
[
0.166259765625,
0.157958984375,
0.3525390625,
0.15673828125,
0.3525390625,
0.17041015625,
0.166259765625,
0.171630859375
],
[
0.53564453125,
0.152587890625,
0.67919921875,
0.15380859375,
0.67919921875,
0.16748046875,
0.53564453125,
0.166259765625
],
[
0.145263671875,
0.166259765625,
0.151123046875,
0.166259765625,
0.151123046875,
0.17333984375,
0.145263671875,
0.17333984375
],
[
0.166259765625,
0.168701171875,
0.36865234375,
0.169189453125,
0.36865234375,
0.182861328125,
0.166259765625,
0.1817626953125
],
[
0.39208984375,
0.171630859375,
0.52294921875,
0.171630859375,
0.52294921875,
0.185302734375,
0.39208984375,
0.185302734375
],
[
0.5322265625,
0.169189453125,
0.7197265625,
0.171630859375,
0.7197265625,
0.185302734375,
0.5322265625,
0.182861328125
],
[
0.166259765625,
0.18115234375,
0.378662109375,
0.1817626953125,
0.378662109375,
0.193603515625,
0.166259765625,
0.1923828125
],
[
0.145263671875,
0.17333984375,
0.151123046875,
0.17333984375,
0.151123046875,
0.1787109375,
0.145263671875,
0.1787109375
],
[
0.14111328125,
0.1817626953125,
0.151123046875,
0.18115234375,
0.154541015625,
0.189453125,
0.1435546875,
0.191162109375
],
[
0.391357421875,
0.182861328125,
0.51708984375,
0.18408203125,
0.51708984375,
0.1953125,
0.391357421875,
0.19482421875
],
[
0.5341796875,
0.182861328125,
0.85986328125,
0.182861328125,
0.85986328125,
0.19775390625,
0.5341796875,
0.19775390625
],
[
0.16796875,
0.19482421875,
0.3359375,
0.19482421875,
0.3359375,
0.204833984375,
0.16796875,
0.204833984375
],
[
0.5341796875,
0.19482421875,
0.603515625,
0.196533203125,
0.603515625,
0.2078857421875,
0.5341796875,
0.2060546875
],
[
0.39208984375,
0.1953125,
0.51123046875,
0.196533203125,
0.51123046875,
0.2078857421875,
0.39208984375,
0.207275390625
],
[
0.145263671875,
0.1953125,
0.151123046875,
0.1953125,
0.151123046875,
0.2037353515625,
0.145263671875,
0.2037353515625
],
[
0.14111328125,
0.2060546875,
0.1561279296875,
0.208984375,
0.151123046875,
0.21728515625,
0.1368408203125,
0.213134765625
],
[
0.166259765625,
0.204833984375,
0.366943359375,
0.204833984375,
0.366943359375,
0.218505859375,
0.166259765625,
0.218505859375
],
[
0.5341796875,
0.2060546875,
0.7841796875,
0.204833984375,
0.7841796875,
0.2197265625,
0.5341796875,
0.2203369140625
],
[
0.1695556640625,
0.218505859375,
0.365234375,
0.218505859375,
0.365234375,
0.2286376953125,
0.1695556640625,
0.2286376953125
],
[
0.53564453125,
0.218505859375,
0.861328125,
0.218505859375,
0.861328125,
0.232177734375,
0.53564453125,
0.232177734375
],
[
0.1695556640625,
0.2298583984375,
0.267822265625,
0.2298583984375,
0.267822265625,
0.2410888671875,
0.1695556640625,
0.2410888671875
],
[
0.53564453125,
0.2286376953125,
0.66162109375,
0.23095703125,
0.66162109375,
0.24462890625,
0.53564453125,
0.2423095703125
],
[
0.16796875,
0.240478515625,
0.33154296875,
0.240478515625,
0.33154296875,
0.254150390625,
0.16796875,
0.254150390625
],
[
0.391357421875,
0.2464599609375,
0.50732421875,
0.2493896484375,
0.50732421875,
0.26416015625,
0.391357421875,
0.26123046875
],
[
0.5341796875,
0.2464599609375,
0.705078125,
0.248779296875,
0.705078125,
0.26318359375,
0.5341796875,
0.26123046875
],
[
0.16796875,
0.254150390625,
0.2216796875,
0.254150390625,
0.2216796875,
0.265380859375,
0.16796875,
0.265380859375
],
[
0.53662109375,
0.262451171875,
0.7001953125,
0.262451171875,
0.7001953125,
0.2724609375,
0.53662109375,
0.2724609375
],
[
0.39208984375,
0.279052734375,
0.468505859375,
0.28076171875,
0.468505859375,
0.292236328125,
0.39208984375,
0.290283203125
],
[
0.49951171875,
0.28076171875,
0.5224609375,
0.28076171875,
0.5224609375,
0.29150390625,
0.49951171875,
0.29150390625
],
[
0.5341796875,
0.279052734375,
0.73779296875,
0.27978515625,
0.73779296875,
0.29150390625,
0.5341796875,
0.290283203125
],
[
0.39208984375,
0.292236328125,
0.44921875,
0.292236328125,
0.44921875,
0.303955078125,
0.39208984375,
0.303955078125
],
[
0.53662109375,
0.292236328125,
0.6455078125,
0.292236328125,
0.6455078125,
0.302734375,
0.53662109375,
0.302734375
],
[
0.53564453125,
0.302734375,
0.72900390625,
0.3017578125,
0.72900390625,
0.3154296875,
0.53564453125,
0.31640625
],
[
0.17041015625,
0.29345703125,
0.372802734375,
0.29345703125,
0.372802734375,
0.30712890625,
0.17041015625,
0.30712890625
],
[
0.166259765625,
0.303955078125,
0.370361328125,
0.303955078125,
0.370361328125,
0.317626953125,
0.166259765625,
0.317626953125
],
[
0.1485595703125,
0.31005859375,
0.15283203125,
0.314208984375,
0.1485595703125,
0.317626953125,
0.1427001953125,
0.3134765625
],
[
0.145263671875,
0.320556640625,
0.154541015625,
0.320556640625,
0.154541015625,
0.330078125,
0.145263671875,
0.330078125
],
[
0.16796875,
0.318359375,
0.264404296875,
0.318359375,
0.264404296875,
0.330078125,
0.16796875,
0.330078125
],
[
0.39208984375,
0.320556640625,
0.52490234375,
0.320556640625,
0.52490234375,
0.334228515625,
0.39208984375,
0.334228515625
],
[
0.5341796875,
0.320556640625,
0.67236328125,
0.320556640625,
0.67236328125,
0.334228515625,
0.5341796875,
0.334228515625
],
[
0.1427001953125,
0.327880859375,
0.15283203125,
0.327880859375,
0.15283203125,
0.341552734375,
0.1427001953125,
0.341552734375
],
[
0.166259765625,
0.328857421875,
0.339111328125,
0.328857421875,
0.339111328125,
0.342529296875,
0.166259765625,
0.342529296875
],
[
0.39208984375,
0.333251953125,
0.5205078125,
0.333251953125,
0.5205078125,
0.3466796875,
0.39208984375,
0.3466796875
],
[
0.5341796875,
0.333251953125,
0.71533203125,
0.333251953125,
0.71533203125,
0.3466796875,
0.5341796875,
0.3466796875
],
[
0.145263671875,
0.34033203125,
0.151123046875,
0.34033203125,
0.151123046875,
0.3486328125,
0.145263671875,
0.3486328125
],
[
0.16796875,
0.341552734375,
0.2216796875,
0.341552734375,
0.2216796875,
0.352783203125,
0.16796875,
0.352783203125
],
[
0.391357421875,
0.345703125,
0.509765625,
0.3466796875,
0.509765625,
0.358154296875,
0.391357421875,
0.35693359375
],
[
0.1427001953125,
0.35400390625,
0.15283203125,
0.35400390625,
0.15283203125,
0.361083984375,
0.1427001953125,
0.361083984375
],
[
0.16796875,
0.35400390625,
0.318115234375,
0.35400390625,
0.318115234375,
0.364013671875,
0.16796875,
0.364013671875
],
[
0.39208984375,
0.364013671875,
0.49951171875,
0.364013671875,
0.49951171875,
0.3759765625,
0.39208984375,
0.3759765625
],
[
0.5341796875,
0.363525390625,
0.6767578125,
0.363525390625,
0.6767578125,
0.377197265625,
0.5341796875,
0.377197265625
],
[
0.53564453125,
0.374755859375,
0.791015625,
0.37353515625,
0.791015625,
0.38720703125,
0.53564453125,
0.388427734375
],
[
0.2283935546875,
0.402099609375,
0.765625,
0.402099609375,
0.765625,
0.4150390625,
0.2283935546875,
0.4150390625
],
[
0.1175537109375,
0.4423828125,
0.490234375,
0.4423828125,
0.490234375,
0.4560546875,
0.1175537109375,
0.4560546875
],
[
0.5146484375,
0.441162109375,
0.87890625,
0.441162109375,
0.87890625,
0.454833984375,
0.5146484375,
0.454833984375
],
[
0.1175537109375,
0.4560546875,
0.345947265625,
0.4560546875,
0.345947265625,
0.466064453125,
0.1175537109375,
0.466064453125
],
[
0.51318359375,
0.457275390625,
0.880859375,
0.457275390625,
0.880859375,
0.47021484375,
0.51318359375,
0.47021484375
],
[
0.1351318359375,
0.4755859375,
0.484375,
0.4755859375,
0.484375,
0.4892578125,
0.1351318359375,
0.4892578125
],
[
0.52880859375,
0.473876953125,
0.87890625,
0.473876953125,
0.87890625,
0.48681640625,
0.52880859375,
0.48681640625
],
[
0.118408203125,
0.4921875,
0.486083984375,
0.4921875,
0.486083984375,
0.505859375,
0.118408203125,
0.505859375
],
[
0.53076171875,
0.490478515625,
0.880859375,
0.490478515625,
0.880859375,
0.50439453125,
0.53076171875,
0.50439453125
],
[
0.51123046875,
0.5048828125,
0.880859375,
0.505859375,
0.880859375,
0.52099609375,
0.51123046875,
0.51953125
],
[
0.1158447265625,
0.52099609375,
0.151123046875,
0.52099609375,
0.151123046875,
0.5322265625,
0.1158447265625,
0.5322265625
],
[
0.1485595703125,
0.51708984375,
0.26611328125,
0.52099609375,
0.26611328125,
0.53515625,
0.1485595703125,
0.5322265625
],
[
0.51318359375,
0.5224609375,
0.880859375,
0.5224609375,
0.880859375,
0.5361328125,
0.51318359375,
0.5361328125
],
[
0.51318359375,
0.5380859375,
0.880859375,
0.5380859375,
0.880859375,
0.5517578125,
0.51318359375,
0.5517578125
],
[
0.1158447265625,
0.54443359375,
0.328369140625,
0.54443359375,
0.328369140625,
0.55859375,
0.1158447265625,
0.55859375
],
[
0.51318359375,
0.5546875,
0.87890625,
0.5546875,
0.87890625,
0.568359375,
0.51318359375,
0.568359375
],
[
0.1175537109375,
0.56640625,
0.486083984375,
0.5654296875,
0.486083984375,
0.580078125,
0.1175537109375,
0.58056640625
],
[
0.51318359375,
0.57080078125,
0.8818359375,
0.57080078125,
0.8818359375,
0.5849609375,
0.51318359375,
0.5849609375
],
[
0.1158447265625,
0.58056640625,
0.486083984375,
0.58203125,
0.486083984375,
0.5966796875,
0.1158447265625,
0.595703125
],
[
0.51318359375,
0.58740234375,
0.880859375,
0.58740234375,
0.880859375,
0.60107421875,
0.51318359375,
0.60107421875
],
[
0.1175537109375,
0.59716796875,
0.483642578125,
0.59716796875,
0.483642578125,
0.61083984375,
0.1175537109375,
0.61083984375
],
[
0.509765625,
0.6015625,
0.63232421875,
0.6015625,
0.63232421875,
0.6162109375,
0.509765625,
0.6162109375
],
[
0.1175537109375,
0.6142578125,
0.484375,
0.61328125,
0.484375,
0.62744140625,
0.1175537109375,
0.62890625
],
[
0.118408203125,
0.63134765625,
0.486083984375,
0.63134765625,
0.486083984375,
0.64453125,
0.118408203125,
0.64453125
],
[
0.51123046875,
0.6328125,
0.5400390625,
0.6328125,
0.5400390625,
0.6455078125,
0.51123046875,
0.6455078125
],
[
0.5556640625,
0.6318359375,
0.78662109375,
0.6328125,
0.78662109375,
0.646484375,
0.5556640625,
0.6455078125
],
[
0.1158447265625,
0.6455078125,
0.48779296875,
0.646484375,
0.48779296875,
0.6611328125,
0.1158447265625,
0.66015625
],
[
0.51123046875,
0.65380859375,
0.87890625,
0.65380859375,
0.87890625,
0.66845703125,
0.51123046875,
0.66845703125
],
[
0.118408203125,
0.6630859375,
0.486083984375,
0.6630859375,
0.486083984375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.509765625,
0.67041015625,
0.880859375,
0.67138671875,
0.880859375,
0.68603515625,
0.509765625,
0.68505859375
],
[
0.1175537109375,
0.6767578125,
0.486083984375,
0.677734375,
0.486083984375,
0.6923828125,
0.1175537109375,
0.69140625
],
[
0.51318359375,
0.6884765625,
0.8818359375,
0.6884765625,
0.8818359375,
0.70166015625,
0.51318359375,
0.70166015625
],
[
0.118408203125,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51123046875,
0.70166015625,
0.880859375,
0.70361328125,
0.880859375,
0.71875,
0.51123046875,
0.71630859375
],
[
0.1175537109375,
0.71044921875,
0.483642578125,
0.71044921875,
0.483642578125,
0.7236328125,
0.1175537109375,
0.7236328125
],
[
0.51123046875,
0.71875,
0.87890625,
0.71875,
0.87890625,
0.732421875,
0.51123046875,
0.732421875
],
[
0.118408203125,
0.72802734375,
0.484375,
0.72802734375,
0.484375,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.51318359375,
0.7353515625,
0.880859375,
0.7353515625,
0.880859375,
0.75,
0.51318359375,
0.75
],
[
0.1175537109375,
0.7421875,
0.483642578125,
0.7421875,
0.483642578125,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.51318359375,
0.751953125,
0.880859375,
0.751953125,
0.880859375,
0.765625,
0.51318359375,
0.765625
],
[
0.118408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51123046875,
0.7666015625,
0.8447265625,
0.76708984375,
0.8447265625,
0.7822265625,
0.51123046875,
0.78076171875
],
[
0.118408203125,
0.77490234375,
0.301513671875,
0.77490234375,
0.301513671875,
0.7890625,
0.118408203125,
0.7890625
],
[
0.133544921875,
0.79052734375,
0.48779296875,
0.79150390625,
0.48779296875,
0.8056640625,
0.133544921875,
0.8046875
],
[
0.54833984375,
0.78515625,
0.880859375,
0.78515625,
0.880859375,
0.798828125,
0.54833984375,
0.798828125
],
[
0.51123046875,
0.80029296875,
0.880859375,
0.80029296875,
0.880859375,
0.81396484375,
0.51123046875,
0.81396484375
],
[
0.118408203125,
0.80712890625,
0.486083984375,
0.80712890625,
0.486083984375,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.1175537109375,
0.822265625,
0.47265625,
0.822265625,
0.47265625,
0.84033203125,
0.1175537109375,
0.84033203125
],
[
0.51318359375,
0.81787109375,
0.880859375,
0.81787109375,
0.880859375,
0.83056640625,
0.51318359375,
0.83056640625
],
[
0.51123046875,
0.8330078125,
0.880859375,
0.8330078125,
0.880859375,
0.8466796875,
0.51123046875,
0.8466796875
],
[
0.1158447265625,
0.8388671875,
0.48779296875,
0.8388671875,
0.48779296875,
0.85693359375,
0.1158447265625,
0.85693359375
],
[
0.51318359375,
0.849609375,
0.76904296875,
0.849609375,
0.76904296875,
0.86328125,
0.51318359375,
0.86328125
],
[
0.76171875,
0.84716796875,
0.77685546875,
0.849609375,
0.77490234375,
0.85693359375,
0.759765625,
0.85498046875
],
[
0.1175537109375,
0.85546875,
0.48779296875,
0.85546875,
0.48779296875,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.1158447265625,
0.87158203125,
0.486083984375,
0.87060546875,
0.486083984375,
0.88525390625,
0.1158447265625,
0.8857421875
],
[
0.5322265625,
0.8662109375,
0.880859375,
0.8662109375,
0.880859375,
0.8798828125,
0.5322265625,
0.8798828125
],
[
0.1175537109375,
0.88720703125,
0.48779296875,
0.88818359375,
0.48779296875,
0.90283203125,
0.1175537109375,
0.90185546875
],
[
0.5322265625,
0.89111328125,
0.8818359375,
0.8935546875,
0.8818359375,
0.90771484375,
0.5322265625,
0.90625
],
[
0.118408203125,
0.90478515625,
0.489501953125,
0.90478515625,
0.489501953125,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.62109375,
0.90673828125,
0.62109375,
0.91845703125,
0.51123046875,
0.91650390625
]
] |
[
"Example",
"Step",
"Relevant Spans",
"Interpreting",
"the",
"came out relation: <film, release date>",
"question",
"film entities: Blind Shaft, The Mask Of Fu Manchu",
"earlier: date comparison",
"Context : Blind Shaft is a 2003",
"film about a pair of brutal con",
"min( release",
"date Blind Shaft,",
"release",
"target:",
"artists operating in the illegal coal",
"mines of present day northern",
"date The Mask ofFu Manchu)",
"",
"China. The Mask Of Fu Manchu",
"Identifying relevant",
"Blind Shaft g =Blind Shaft e",
"isa1932pre-Codeadventurefilm",
"T",
"a",
"information through",
"The Mask Of Fu Manchu q := The Mask Of Fu",
"directed by Charles Brabin.",
"Manchu e.",
"referential equality",
"",
"O",
"Question : Which film came out",
"came out q := <date, film> construction e",
"earlier,Blind Shaft orTheMask",
"release dates: < Blind Shaft , 2003 >,< The Mask Of",
"Of Fu Manchu?",
"Fu Manchu , 1932 >",
"Answer : The Mask Of Fu",
"Value comparison",
"solution: earlier q := min c",
"Manchu",
"min(1932,2003)=1932",
"Interpreting",
"the",
"bornrelation:<person,location>",
"question",
"Hawaii : location",
"target: born : <Hawaii , UNK>",
"Context: Barack Obama was the",
"44th president of the US.He was",
"J",
"S",
"born in Hawaii.",
"Identifying relevant",
"Hawaii q - Hawaii c",
"",
"Question: Who was born in",
"information through",
"born relation: < he, Hawaii >",
"e.",
"Hawaii?",
"referential equality",
"O",
"Answer:BarackObama.",
"Coref. resolution",
"<Barack Obama, he >",
"solution: born < Barack Obama, Hawaii >",
"Table 1: The basic reasoning steps for answering comparison and coreference questions",
"instances in that dataset for which different kinds of explana-",
"由于较大的模型被证明可以更好地概括某些用",
"tions align (according to our Def. 2).",
"例 (Hendrycks et al.,2020; Bhargava et al.,2021)",
"我们将高对齐分数解释为两种解释都是忠实",
"我们调查它们是否也更有可能是正确的原因",
"的证据，并且模型确实执行了预期的推理步骤",
"我们fine-tune每个编码器使用Devlinetal.",
"(2019）中的体系结构（有关详细信息，请参见",
"3",
"Methodology",
"附录），并在验证集上对其进行评估（因为测",
"试集不是公开的）。我们在提取QA中使用标",
"3.1Datasets and Models",
"准评估指标：F1-Score（预测和“黄金”答案之",
"对于Coreference，我们使用QuorefE(Dasigi",
"间的令牌重叠百分比，在所有数据点上平均）",
"etal.,2019）数据集（20K训练和2。4K验证",
"和Exact-match（预测答案与“黄金”答案匹配",
"实例），其中要求注释者为给定文本设计问题",
"的数据点数量)",
"以便回答这些问题需要解决回指。为了进行",
"比较，我们对HotpotQA（Yangetal.，2018）和",
"3.2",
"Counterfactual Explanations",
"2WikiMultiHopQA（Hoetal.，2020）中的问题进",
"我们的推理公式（Table1）包括三个用于共同",
"行了采样：两个数据集的问题用它们的推理类",
"引用和比较的基本步骤：解释问题，通过引用",
"型（桥接或比较）进行了手动注释。我们选择",
"相等来识别相关信息以及对所识别信息的目标",
"包含比较形容词或副词的“比较”问题（23K",
"操作（共同引用解析或值比较）。我们专注于",
"训练，3K验证实例）。这些资源基于维基百科",
"最后一步，因为：（a）它隐式地依赖于对问题",
"并有多个段落作为上下文，但回答问题所需的",
"和上下文的正确语义解析；（b）我们数据中的",
"句子（通常为2-3）被标记为“支持事实”。由",
"引用相等在很大程度上是微不足道的：大多数",
"于我们不关注多跳信息检索技能，因此我们将",
"实体在问题和文本中具有相同的表面形式",
"上下文限制为这些句子",
"我们使用BERT系列的五个预先",
"个明显的语义上有效的扰动应该改变预测",
"（从而测试模型对比较操作的理解）是将比较",
"训练的基于Transformer的编码器进",
"行实验：RoBERTalargeE(Liu et al.，2019)",
"形容词替换为其反义词（Figure3d）。由于我",
"们的示例仅包含6个用作比较运算符的令牌，",
"BERTlarge-cased, BERTbase-casedE(Devlin et al.,",
"因此我们手动定义适当的替换。",
"8",
"2019)，BERTmedium和 BERTsmallE(Turc et al.,",
"2019;Bhargavaetal.,2021)。这些BERT模型",
"对于共同引用问题，有能力的RC模型至少",
"的主要区别在于架构块的结构和参数的数量，",
"\"earlier later, first → later, more recently → earlier,",
"而RoBERTa也有不同的训练语料库和优化。",
"olderyounger."
] |
{
"arxiv_id": "2209.07430",
"language": "zh",
"page_idx": 4
}
|
||
2004.14355-zh-0
|
2004.14355-zh-0.jpg
|
# 学习如何消除歧义： 元学习，用于少射词义消除歧义
Nithin Holla
ILLC, University of Amsterdam
nithin. holla7@ gmail. com
Pushkar Mishra Facebook AI
pushkarmishra@fb.com
Helen Yannakoudakis
Ekaterina Shutova
Dept. of Informatics, King's College London ILLC, University of Amsterdam
helen. yannakoudakis@ kcl. ac. uk
e.shutova@uva.nl
$\operatorname{Abstract}$
The success of deep learning methods hinges on the availability of large training datasets annotated for the task of interest. In contrast to human intelligence, these methods lack versatility and struggle to learn and adapt quickly to new tasks, where labeled data is scarce. Metalearning aims to solve this problem by training a model on a large number of few-shot tasks, with an objective to learn new tasks quickly from a small number of examples. In this paper, we propose a meta-learning framework for few-shot word sense disambiguation (WSD),where the goal is to learn to disambiguate unseen words from only a few labeled instances. Meta-learning approaches have so far been typically tested in an $N$ -way, $K$ -shot classification setting where each task has $N$ classes with $K$ examples per class. Owing to its nature, WSD deviates from this controlled setup and requires the models to handle a large number of highly unbalanced classes. We extend several popular meta-learning approaches to this scenario, and analyze their strengths and weaknesses in this new challenging setting.
2016)或基于知识的 (Lesk, 1986; Agirre et al., 2014; Moro et al.,2014)。虽然监督方法通常优于基于知识的方法 (Raganato et al., 2017a)
但它们需要用词义手动注释数据，而大规模生产这些数据很昂贵。这些方法还倾向于独立地学习每个单词的分类模型，因此可能对注释数据量有限的单词表现不佳。然而，涉及所有单词 (Raganato et al., 2017b) 的单个监督模型的替代方案仍然不能充分解决稀有单词(Kumar et al., 2019) 的问题
另一方面，人类具有从少数示例(Lake et al., 2015)中学习的非凡能力。这激发了研究人员研究使机器学习模型能够做到这一点的技术。一种这样的方法是迁移学习 (Caruana, 1993)
它旨在通过在任务之间传输特征来提高模型的数据效率。但是，在缺少大量用于微调(Yogatama et al., 2019) 的特定于任务的数据的情况下，它仍然无法推广到新任务。元学习，被称为学习(Schmidhuber,1987; Bengio et al., 1991; Thrun and Pratt, 1998) ,是一种替代范式，它借鉴了过去的经验，以便快速学习和适应新任务：该模型在许多相关任务上进行了训练，以便它可以解决看不见的任务仅使用少量的训练示例。典型的元学习设置由两个部分组成：一个学习者，从其小型训练数据中适应每个任务；以及一个元学习者，通过获取所有任务中通用的知识来指导学习者
元学习最近已成为一种有前途的少发学习方法。它在计算机视觉(Triantafillou et al., 2020; Fontanini et al., 2020; Hendryx et al., 2019; Wang et al., 2020) 和强化学习 (Wang et al., 2016; Duan et al., 2016; Alet et al., 2020) 方面取得了成功。它最近也进入了 NLP,并已应用于机器翻译 (Gu et al.,2018) ,关系(Obamuyide and Vlachos, 2019b) 和文本 (Yu et al., 2018) 分类以及句子级语义任务 (Dou et al., 2019; Bansal et al., 2019)
在本文中，我们介绍了 WSD 的第一个元学习方法。我们提出了一些模型，这些模型可以学习仅从几个标记的示例中快速消除新
## 1 Introduction
自然语言本质上是模棱两可的，许多单词具有一系列可能的含义。词义消歧(WSD)是自然语言理解的核心任务，其目标是从预定义的词义清单中将单词与其正确的上下文含义相关联。。WSD 已被证明可以改善下游任务， 例如机器翻译 (Chan et al.,2007) 和信息检索(Zhong and Ng,2012)。但是，它被认为是一个 AI 完整的问题 (Navigli, 2009) -它需要对语言以及现实世界知识的复杂理解
WSD 的方法通常依赖于(半)监督学习(Zhong and Ng, 2010; Melamud et al., 2016; $\text{K\r {a}geb\" {a}ckandSalomonsson, 2016; Yuanetal. }$,
|
[
[
0.399658203125,
0.07659912109375,
0.603515625,
0.07781982421875,
0.603515625,
0.0955810546875,
0.399658203125,
0.09442138671875
],
[
0.334228515625,
0.09857177734375,
0.66650390625,
0.09857177734375,
0.66650390625,
0.1134033203125,
0.334228515625,
0.1134033203125
],
[
0.280517578125,
0.1258544921875,
0.39208984375,
0.127685546875,
0.39208984375,
0.142578125,
0.280517578125,
0.14013671875
],
[
0.599609375,
0.127685546875,
0.73779296875,
0.127685546875,
0.73779296875,
0.141357421875,
0.599609375,
0.141357421875
],
[
0.197265625,
0.1436767578125,
0.47607421875,
0.144287109375,
0.47607421875,
0.1591796875,
0.197265625,
0.157958984375
],
[
0.61376953125,
0.1455078125,
0.72314453125,
0.1455078125,
0.72314453125,
0.15673828125,
0.61376953125,
0.15673828125
],
[
0.231689453125,
0.160888671875,
0.443359375,
0.160888671875,
0.443359375,
0.174560546875,
0.231689453125,
0.174560546875
],
[
0.568359375,
0.160888671875,
0.767578125,
0.160888671875,
0.767578125,
0.17578125,
0.568359375,
0.17578125
],
[
0.2392578125,
0.1923828125,
0.42822265625,
0.193603515625,
0.42822265625,
0.2078857421875,
0.2392578125,
0.207275390625
],
[
0.58837890625,
0.193603515625,
0.748046875,
0.193603515625,
0.748046875,
0.2078857421875,
0.58837890625,
0.2078857421875
],
[
0.14111328125,
0.210205078125,
0.80615234375,
0.210205078125,
0.80615234375,
0.2244873046875,
0.14111328125,
0.2244873046875
],
[
0.197265625,
0.2257080078125,
0.466796875,
0.226806640625,
0.466796875,
0.2410888671875,
0.197265625,
0.240478515625
],
[
0.59375,
0.22802734375,
0.744140625,
0.22802734375,
0.744140625,
0.2410888671875,
0.59375,
0.2410888671875
],
[
0.26611328125,
0.265380859375,
0.343505859375,
0.265380859375,
0.343505859375,
0.27978515625,
0.26611328125,
0.27978515625
],
[
0.515625,
0.2666015625,
0.880859375,
0.2666015625,
0.880859375,
0.27978515625,
0.515625,
0.27978515625
],
[
0.5146484375,
0.28076171875,
0.880859375,
0.28076171875,
0.880859375,
0.295654296875,
0.5146484375,
0.295654296875
],
[
0.1485595703125,
0.29345703125,
0.459228515625,
0.29345703125,
0.459228515625,
0.30810546875,
0.1485595703125,
0.30810546875
],
[
0.515625,
0.298583984375,
0.880859375,
0.298583984375,
0.880859375,
0.312255859375,
0.515625,
0.312255859375
],
[
0.1485595703125,
0.30810546875,
0.459228515625,
0.30810546875,
0.459228515625,
0.32177734375,
0.1485595703125,
0.32177734375
],
[
0.5341796875,
0.31640625,
0.880859375,
0.31640625,
0.880859375,
0.32666015625,
0.5341796875,
0.32666015625
],
[
0.14697265625,
0.322509765625,
0.459228515625,
0.322509765625,
0.459228515625,
0.336181640625,
0.14697265625,
0.336181640625
],
[
0.5146484375,
0.328857421875,
0.88330078125,
0.330078125,
0.88330078125,
0.344482421875,
0.5146484375,
0.34326171875
],
[
0.1485595703125,
0.33740234375,
0.45751953125,
0.33740234375,
0.45751953125,
0.350830078125,
0.1485595703125,
0.350830078125
],
[
0.14697265625,
0.349853515625,
0.459228515625,
0.350830078125,
0.459228515625,
0.365234375,
0.14697265625,
0.364013671875
],
[
0.51708984375,
0.3466796875,
0.8818359375,
0.3466796875,
0.8818359375,
0.35986328125,
0.51708984375,
0.35986328125
],
[
0.1485595703125,
0.365234375,
0.45751953125,
0.365234375,
0.45751953125,
0.37890625,
0.1485595703125,
0.37890625
],
[
0.515625,
0.363525390625,
0.8818359375,
0.363525390625,
0.8818359375,
0.377197265625,
0.515625,
0.377197265625
],
[
0.1502685546875,
0.38134765625,
0.45751953125,
0.38134765625,
0.45751953125,
0.391357421875,
0.1502685546875,
0.391357421875
],
[
0.5146484375,
0.377685546875,
0.880859375,
0.377685546875,
0.880859375,
0.391357421875,
0.5146484375,
0.391357421875
],
[
0.14697265625,
0.393798828125,
0.45751953125,
0.393798828125,
0.45751953125,
0.40673828125,
0.14697265625,
0.40673828125
],
[
0.515625,
0.3955078125,
0.88330078125,
0.3955078125,
0.88330078125,
0.4091796875,
0.515625,
0.4091796875
],
[
0.14697265625,
0.407958984375,
0.4609375,
0.407958984375,
0.4609375,
0.421630859375,
0.14697265625,
0.421630859375
],
[
0.5146484375,
0.410888671875,
0.74072265625,
0.410888671875,
0.74072265625,
0.424560546875,
0.5146484375,
0.424560546875
],
[
0.1485595703125,
0.421630859375,
0.459228515625,
0.421630859375,
0.459228515625,
0.435302734375,
0.1485595703125,
0.435302734375
],
[
0.5341796875,
0.427490234375,
0.880859375,
0.427490234375,
0.880859375,
0.441162109375,
0.5341796875,
0.441162109375
],
[
0.1485595703125,
0.435791015625,
0.459228515625,
0.435791015625,
0.459228515625,
0.449462890625,
0.1485595703125,
0.449462890625
],
[
0.5146484375,
0.443603515625,
0.8818359375,
0.443603515625,
0.8818359375,
0.457275390625,
0.5146484375,
0.457275390625
],
[
0.1485595703125,
0.45068359375,
0.45751953125,
0.45068359375,
0.45751953125,
0.46435546875,
0.1485595703125,
0.46435546875
],
[
0.1485595703125,
0.46435546875,
0.459228515625,
0.46435546875,
0.459228515625,
0.47802734375,
0.1485595703125,
0.47802734375
],
[
0.515625,
0.460205078125,
0.880859375,
0.460205078125,
0.880859375,
0.473876953125,
0.515625,
0.473876953125
],
[
0.1485595703125,
0.478515625,
0.459228515625,
0.478515625,
0.459228515625,
0.4921875,
0.1485595703125,
0.4921875
],
[
0.5341796875,
0.47802734375,
0.880859375,
0.47802734375,
0.880859375,
0.488037109375,
0.5341796875,
0.488037109375
],
[
0.1485595703125,
0.493408203125,
0.459228515625,
0.493408203125,
0.459228515625,
0.50732421875,
0.1485595703125,
0.50732421875
],
[
0.5341796875,
0.4921875,
0.8818359375,
0.4921875,
0.8818359375,
0.505859375,
0.5341796875,
0.505859375
],
[
0.151123046875,
0.5087890625,
0.453369140625,
0.5087890625,
0.453369140625,
0.51953125,
0.151123046875,
0.51953125
],
[
0.51708984375,
0.50830078125,
0.8818359375,
0.50830078125,
0.8818359375,
0.521484375,
0.51708984375,
0.521484375
],
[
0.1485595703125,
0.521484375,
0.459228515625,
0.521484375,
0.459228515625,
0.53515625,
0.1485595703125,
0.53515625
],
[
0.5146484375,
0.52490234375,
0.880859375,
0.52490234375,
0.880859375,
0.5380859375,
0.5146484375,
0.5380859375
],
[
0.1485595703125,
0.5361328125,
0.459228515625,
0.5361328125,
0.459228515625,
0.5498046875,
0.1485595703125,
0.5498046875
],
[
0.515625,
0.54052734375,
0.8818359375,
0.54052734375,
0.8818359375,
0.55419921875,
0.515625,
0.55419921875
],
[
0.1485595703125,
0.548828125,
0.45751953125,
0.548828125,
0.45751953125,
0.5625,
0.1485595703125,
0.5625
],
[
0.5146484375,
0.5546875,
0.8818359375,
0.5556640625,
0.8818359375,
0.57080078125,
0.5146484375,
0.5693359375
],
[
0.1485595703125,
0.5634765625,
0.459228515625,
0.5634765625,
0.459228515625,
0.57763671875,
0.1485595703125,
0.57763671875
],
[
0.14697265625,
0.57763671875,
0.459228515625,
0.57763671875,
0.459228515625,
0.59130859375,
0.14697265625,
0.59130859375
],
[
0.5146484375,
0.572265625,
0.8818359375,
0.572265625,
0.8818359375,
0.5859375,
0.5146484375,
0.5859375
],
[
0.1485595703125,
0.5927734375,
0.459228515625,
0.5927734375,
0.459228515625,
0.60546875,
0.1485595703125,
0.60546875
],
[
0.515625,
0.5888671875,
0.8818359375,
0.5888671875,
0.8818359375,
0.6025390625,
0.515625,
0.6025390625
],
[
0.1485595703125,
0.60791015625,
0.45751953125,
0.60791015625,
0.45751953125,
0.6181640625,
0.1485595703125,
0.6181640625
],
[
0.51708984375,
0.60498046875,
0.8818359375,
0.60498046875,
0.8818359375,
0.6181640625,
0.51708984375,
0.6181640625
],
[
0.1502685546875,
0.62255859375,
0.45751953125,
0.62255859375,
0.45751953125,
0.6328125,
0.1502685546875,
0.6328125
],
[
0.5146484375,
0.62060546875,
0.8818359375,
0.619140625,
0.8818359375,
0.63427734375,
0.5146484375,
0.63525390625
],
[
0.1485595703125,
0.63525390625,
0.45751953125,
0.63525390625,
0.45751953125,
0.6484375,
0.1485595703125,
0.6484375
],
[
0.515625,
0.63720703125,
0.880859375,
0.63720703125,
0.880859375,
0.65087890625,
0.515625,
0.65087890625
],
[
0.1485595703125,
0.6494140625,
0.45751953125,
0.6494140625,
0.45751953125,
0.6630859375,
0.1485595703125,
0.6630859375
],
[
0.515625,
0.65283203125,
0.8818359375,
0.65283203125,
0.8818359375,
0.66650390625,
0.515625,
0.66650390625
],
[
0.145263671875,
0.662109375,
0.376220703125,
0.66455078125,
0.376220703125,
0.6787109375,
0.145263671875,
0.6767578125
],
[
0.515625,
0.66943359375,
0.8818359375,
0.66943359375,
0.8818359375,
0.68310546875,
0.515625,
0.68310546875
],
[
0.515625,
0.68505859375,
0.8076171875,
0.68505859375,
0.8076171875,
0.69873046875,
0.515625,
0.69873046875
],
[
0.118408203125,
0.693359375,
0.1435546875,
0.693359375,
0.1435546875,
0.705078125,
0.118408203125,
0.705078125
],
[
0.15283203125,
0.69140625,
0.26611328125,
0.6923828125,
0.26611328125,
0.70703125,
0.15283203125,
0.7060546875
],
[
0.53564453125,
0.70166015625,
0.880859375,
0.70166015625,
0.880859375,
0.71533203125,
0.53564453125,
0.71533203125
],
[
0.12176513671875,
0.71875,
0.489501953125,
0.71875,
0.489501953125,
0.732421875,
0.12176513671875,
0.732421875
],
[
0.515625,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.732421875,
0.515625,
0.732421875
],
[
0.12005615234375,
0.73388671875,
0.489501953125,
0.73388671875,
0.489501953125,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.5146484375,
0.73388671875,
0.880859375,
0.73388671875,
0.880859375,
0.74755859375,
0.5146484375,
0.74755859375
],
[
0.12176513671875,
0.75048828125,
0.48779296875,
0.75048828125,
0.48779296875,
0.76416015625,
0.12176513671875,
0.76416015625
],
[
0.5146484375,
0.7490234375,
0.8818359375,
0.75,
0.8818359375,
0.76416015625,
0.5146484375,
0.76318359375
],
[
0.12005615234375,
0.76708984375,
0.484375,
0.76708984375,
0.484375,
0.77783203125,
0.12005615234375,
0.77783203125
],
[
0.5146484375,
0.7666015625,
0.880859375,
0.7666015625,
0.880859375,
0.77978515625,
0.5146484375,
0.77978515625
],
[
0.12176513671875,
0.7822265625,
0.48779296875,
0.7822265625,
0.48779296875,
0.7958984375,
0.12176513671875,
0.7958984375
],
[
0.515625,
0.7822265625,
0.880859375,
0.7822265625,
0.880859375,
0.7958984375,
0.515625,
0.7958984375
],
[
0.12176513671875,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.12176513671875,
0.8125
],
[
0.515625,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.515625,
0.8125
],
[
0.118408203125,
0.81396484375,
0.47265625,
0.81396484375,
0.47265625,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.468505859375,
0.818359375,
0.474365234375,
0.818359375,
0.474365234375,
0.82373046875,
0.468505859375,
0.82373046875
],
[
0.474365234375,
0.818359375,
0.483642578125,
0.818359375,
0.483642578125,
0.82373046875,
0.474365234375,
0.82373046875
],
[
0.51318359375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.12176513671875,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.12176513671875,
0.84423828125
],
[
0.51318359375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.12176513671875,
0.8466796875,
0.3779296875,
0.8466796875,
0.3779296875,
0.85986328125,
0.12176513671875,
0.85986328125
],
[
0.5146484375,
0.8466796875,
0.771484375,
0.8466796875,
0.771484375,
0.85986328125,
0.5146484375,
0.85986328125
],
[
0.139404296875,
0.86328125,
0.484375,
0.86328125,
0.484375,
0.876953125,
0.139404296875,
0.876953125
],
[
0.5322265625,
0.8623046875,
0.8818359375,
0.8623046875,
0.8818359375,
0.876953125,
0.5322265625,
0.876953125
],
[
0.118408203125,
0.87890625,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.118408203125,
0.8935546875
],
[
0.51318359375,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.491943359375,
0.8955078125,
0.491943359375,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.515625,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"学习如何消除歧义：",
"元学习，用于少射词义消除歧义",
"Nithin Holla",
"Pushkar Mishra",
"ILLC, University of Amsterdam",
"Facebook Al",
"nithin.holla7@gmail.com",
"pushkarmishra@fb.com",
"Helen Yannakoudakis",
"Ekaterina Shutova",
"Dept. of Informatics, King's College London ILLC, University of Amsterdam",
"helen.yannakoudakis@kcl.ac.uk",
"e.shutova@uva.nl",
"Abstract",
"2016）或基于知识的（Lesk，1986；Agirreetal.",
"2014；Moroetal.，2014）。虽然监督方法通常",
"The success of deep learning methods",
"优于基于知识的方法（Raganatoetal.，2017a)",
"hinges on the availability of large train",
"但它们需要用词义手动注释数据，而大规",
"ing datasets annotated for the task of in-",
"模生产这些数据很昂贵。这些方法还倾向于独",
"terest. In contrast to human intelligence,",
"these methods lack versatility and strug-",
"立地学习每个单词的分类模型，因此可能对注",
"gle to learn and adapt quickly to new",
"释数据量有限的单词表现不佳。然而，涉及所",
"tasks,wherelabeled datais scarce.Meta",
"有单词（Raganato et al.2017b）的单个监督",
"learning aims to solve this problem by",
"模型的替代方案仍然不能充分解决稀有单词",
"training a model on a large number of",
"(Kumaretal.,2019）的问题",
"few-shot tasks, with an objective to learn",
"另一方面，人类具有从少数示例（Lakeetal.",
"new tasks quickly from a small number",
"2015）中学习的非凡能力。这激发了研究人员",
"of examples. In this paper, we propose a",
"meta-learning framework for few-shot word",
"研究使机器学习模型能够做到这一点的技术。",
"sense disambiguation (WSD), where the",
"种这样的方法是迁移学习（Caruana，1993）",
"goal is to learn to disambiguate unseen",
"它旨在通过在任务之间传输特征来提高模",
"words from only afewlabeled instances",
"型的数据效率。但是，在缺少大量用于微调",
"Meta-learning approaches have so far been",
"（Yogatamaetal.，2019）的特定于任务的数据",
"typically tested in an N -way, K -shot clas",
"的情况下，它仍然无法推广到新任务。元学",
"sification setting where each task has N",
"习，被称为学习（Schmidhuber，1987；Bengio",
"classes with K examples per class. Ow-",
"ing to its nature, WSD deviates from this",
"etal.,1991;Thrun andPratt,1998)，是一种",
"controlled setup and requires the models",
"替代范式，它借鉴了过去的经验，以便快速学",
"to handle a large number of highly unbal",
"习和适应新任务：该模型在许多相关任务上进",
"anced classes.We extend severalpopular",
"行了训练，以便它可以解决看不见的任务仅使",
"meta-learning approaches to this scenario,",
"用少量的训练示例。典型的元学习设置由两个",
"and analyze their strengths and weaknesses",
"部分组成：二个学习者，从其小型训练数据中",
"in this new challenging setting.",
"适应每个任务；以及一个元学习者，通过获取",
"所有任务中通用的知识来指导学习者",
"1 ",
"Introduction",
"元学习最近已成为一种有前途的少发学习",
"自然语言本质上是模棱两可的，许多单词具",
"方法。它在计算机视觉（Triantafillou etal.",
"有一系列可能的含义。词义消歧（WSD）是",
"2020; Fontanini et al., 2020; Hendryx et al.",
"自然语言理解的核心任务，其目标是从预定义",
"2019；Wangetal.，2020）和强化学习（Wang",
"的词义清单中将单词与其正确的上下文含义",
"et al., 2016; Duan et al., 2016; Alet et al., 2020)",
"相关联。。WSD已被证明可以改善下游任务，",
"方面取得了成功。它最近也进入了NLP，并",
"例如机器翻译（Chanetal.，2007）和信息检索",
"已应用于机器翻译（Guetal.，2018），关系",
"(ZhongandNg，2012）。但是，它被认为是",
"-",
"",
"(Obamuyide and Vlachos,2019b)和文本(Yu",
"个AI完整的问题（Navigli，2009）-它需要对语",
"etal.，2018）分类以及句子级语义任务（Dou",
"言以及现实世界知识的复杂理解",
"et al., 2019; Bansal et al., 2019)",
"WSD的方法通常依赖于（半）监督学习",
"在本文中，我们介绍了WSD的第一个元",
"(Zhong and Ng, 2010; Melamud et al., 2016;",
"学习方法。我们提出了一些模型，这些模型",
"Kageback and Salomonsson, 2016; Yuan et al.",
"可以学习仅从几个标记的示例中快速消除新"
] |
{
"arxiv_id": "2004.14355",
"language": "zh",
"page_idx": 0
}
|
||
2004.14355-zh-1
|
2004.14355-zh-1.jpg
|
词的歧义。由于其性质，WSD 在句子中表现出单词间的依赖性，具有大量的类以及不可避免的类不平衡；与大多数当前元学习方法中的受控设置相比，所有这些都提出了新的挑战。为了解决这些挑战，我们将三种流行的元学习算法扩展到此任务：原型网络(Snell et al.,2017) ,与模型无关的元学习 (MAML) (Finn et al., 2017) 及其混合-ProtoMAML (Triantafillou et al., 2020) 。我们使用三种底层模型架构来研究元学习，即递归网络，多层感知器 (MLP) 和转换器 (Vaswani et al.,2017) , 并尝试使用不同数量的句子进行特定于任务的微调。我们通过在一组看不见的新词上进行测试来评估模型的快速适应能力，从而证明了它从少量示例中学习新词的能力
由于没有可用的少量 WSD 基准，因此我们创建了公开可用的 WSD 数据集的少量版本。我们发布了我们的代码以及用于生成少量数据设置的脚本，以促进进一步的研究。$^{1}$
## 2 Related Work
2.1 Meta-learning
与“传统”机器学习方法相比，元学习涉及一种不同的范式，称为情节学习。元学习中的训练集和测试集分别称为元训练集($\mathcal{D}_meta-train$ )和元测试集($\mathcal{D}_meta-test$)。两组均由情节组成，而不是单个数据点。每个情节构成一个任务 $\mathcal{T} _i$ , 其 中 包 括 少 量 用 于 适 应 的 训 练 示 例 - 支 持 集 $\mathcal{D} _{support}^{( i) }$,以 及 一 组 单 独 的 用 于 评 估 的 示 特 集 $\mathcal{D} _{support}^{( i) }$ , 以 及 一 组 单 独 的 用 于 评 估 的 示例-查询集$\mathcal{D}_query$ 。元学习的典型设置是平衡的$N$ -way , $K$ -shot 设置，其中每个剧集都有$N$类，其支持集中每个类都有$K$示例
元学习算法大致分为三种类型：基于度量的 (Koch et al., 2015; Vinyals et al., 2016; Sung et al., 2018; Snell et al., 2017),基于模型的 (Santoro et al., 2016; Munkhdalai and Yu, 2017) 和基于优化的 (Ravi and Larochelle, 2017; Finn et al., 2017; Nichol et al., 2018) 。基于度量的方法首先将每个情节中的示例嵌入到高维空间中，通常使用神经网络。接下来， 他们基于测量与支持示例的相似性的核函数， 获得所有查询示例的标签上的概率分布。基于模型的方法旨在直接通过其体系结构实现快速学习。他们通常使用外部存储器，以便记住过去遇到的关键示例。基于优化的方法在其目标函数中明确包括泛化性，并针对同一目标函数进行优化。在本文中，我们尝试了基于度量和基于优化的方法，以及它们的混合方法。
$^{1}$https://github.com/Nithin-Holla/MetaWSD
### 2.2 Meta-learning in NLP
NLP 中的元学习仍处于起步阶段。Gu et al. (2018)将元学习应用于神经机器翻译问题，他们对将高资源语言翻译成英语进行元训练，并对将低资源语言翻译成英语进行元测试。Obamuyide and Vlachos (2019b) 使用元学习进行关系分类，Obamuyide and Vlachos (2019a) 用于终身学习环境中的关系提取。Chen et al. (2019)考虑关系学习并将元学习应用于知识图中的少拍链接预测。Dou et al.(2019) 对来自 GLUE 基准 (Wang et al., 2018) 的某些高资源任务执行元训练，并对来自同一基准的某些低资源任务执行元测试。Bansal et al. (2019)提出了一个 softmax 参数生成器组件， 可以在元训练任务中启用不同数量的类。他们选择 GLUE 中的任务以及 SNLI (Bowman et al., 2015) 进行元训练，并使用实体类型，关系分类，情感分类，文本分类和科学 NLI 作为测试任务。元学习也被探索用于少数文本分类(Yu et al., 2018; Geng et al., 2019; Jiang et al., 2018; Sun et al., 2019)。Wu et al.(2019) 采用元强化学习技术进行多标签分类，并进行了实体类型和文本分类的实验。Hu et al. (2019) 使用元学习来学习词汇外单词的表示形式，并将其作为回归任务
## 2.3 Supervised WSD
早期的 WSD 监督学习方法依赖于从上下文单词 (Lee and Ng, 2002; Navigli, 2009; Zhong and Ng, 2010) 中提取的手工特征。后来的工作使用单词嵌入作为分类 (Taghipour and Ng, 2015; Rothe and Schütze, 2015; Iacobacci et al., 2016) 的功能。随着深度学习的兴起，LSTM (Hochreiter and Schmidhuber, 1997)模型成为流行的 (Melamud et al., 2016; Kågebäck and Salomonsson, 2016; Yuan et al., 2016)。虽然大多数工作每个单词训练单个模型，但 Raganato et al. (2017b) 设计了一个具有大量输出单元的 LSTM 模型来消除所有单词的歧义。Peters et al. (2018) 通过与上下文化ELMo (Peters et al.,2018)嵌入匹配的最近邻执行 WSD。Hadiwinoto et al. (2019)使用来自BERT (Devlin et al., 2019) 的预先训练的上下文化表示作为特征。GlossBERT (Huang et al. 2019)结合了 WordNet (Miller et al., 1990) 的意义定义，以形成上下文-光泽对，同时微调WSD 的 BERT。通过考虑 WordNet 的图形结构，EWISER (Bevilacqua and Navigli, 2020) 达到了当前最先进的 F1 分数 80。1% 上的基准 Raganato et al. (2017a)
|
[
[
0.12005615234375,
0.07598876953125,
0.48779296875,
0.07598876953125,
0.48779296875,
0.08905029296875,
0.12005615234375,
0.08905029296875
],
[
0.509765625,
0.07659912109375,
0.73876953125,
0.07659912109375,
0.73876953125,
0.0914306640625,
0.509765625,
0.0914306640625
],
[
0.12005615234375,
0.09320068359375,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10394287109375,
0.880859375,
0.1181640625,
0.51318359375,
0.1175537109375
],
[
0.12005615234375,
0.10986328125,
0.489501953125,
0.10986328125,
0.489501953125,
0.12353515625,
0.12005615234375,
0.12353515625
],
[
0.51318359375,
0.119384765625,
0.88330078125,
0.12054443359375,
0.88330078125,
0.1353759765625,
0.51318359375,
0.1341552734375
],
[
0.12176513671875,
0.1258544921875,
0.48779296875,
0.1258544921875,
0.48779296875,
0.1395263671875,
0.12176513671875,
0.1395263671875
],
[
0.12005615234375,
0.141357421875,
0.48779296875,
0.141357421875,
0.48779296875,
0.155029296875,
0.12005615234375,
0.155029296875
],
[
0.515625,
0.13720703125,
0.8818359375,
0.13720703125,
0.8818359375,
0.15087890625,
0.515625,
0.15087890625
],
[
0.118408203125,
0.1561279296875,
0.486083984375,
0.1561279296875,
0.486083984375,
0.169189453125,
0.118408203125,
0.169189453125
],
[
0.5146484375,
0.152587890625,
0.8818359375,
0.152587890625,
0.8818359375,
0.166259765625,
0.5146484375,
0.166259765625
],
[
0.12005615234375,
0.17333984375,
0.490234375,
0.17333984375,
0.490234375,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.51318359375,
0.168701171875,
0.8818359375,
0.169189453125,
0.8818359375,
0.182861328125,
0.51318359375,
0.1817626953125
],
[
0.118408203125,
0.188232421875,
0.489501953125,
0.188232421875,
0.489501953125,
0.201904296875,
0.118408203125,
0.201904296875
],
[
0.515625,
0.185302734375,
0.880859375,
0.185302734375,
0.880859375,
0.198974609375,
0.515625,
0.198974609375
],
[
0.12176513671875,
0.2060546875,
0.48779296875,
0.2060546875,
0.48779296875,
0.2197265625,
0.12176513671875,
0.2197265625
],
[
0.515625,
0.201904296875,
0.880859375,
0.201904296875,
0.880859375,
0.215576171875,
0.515625,
0.215576171875
],
[
0.12176513671875,
0.221435546875,
0.489501953125,
0.221435546875,
0.489501953125,
0.235107421875,
0.12176513671875,
0.235107421875
],
[
0.515625,
0.21728515625,
0.8818359375,
0.21728515625,
0.8818359375,
0.23095703125,
0.515625,
0.23095703125
],
[
0.12005615234375,
0.2381591796875,
0.490234375,
0.2381591796875,
0.490234375,
0.251708984375,
0.12005615234375,
0.251708984375
],
[
0.51708984375,
0.2327880859375,
0.8818359375,
0.2327880859375,
0.8818359375,
0.2464599609375,
0.51708984375,
0.2464599609375
],
[
0.118408203125,
0.254150390625,
0.48779296875,
0.254150390625,
0.48779296875,
0.267333984375,
0.118408203125,
0.267333984375
],
[
0.5146484375,
0.2493896484375,
0.8818359375,
0.2493896484375,
0.8818359375,
0.26318359375,
0.5146484375,
0.26318359375
],
[
0.12176513671875,
0.271484375,
0.486083984375,
0.271484375,
0.486083984375,
0.281982421875,
0.12176513671875,
0.281982421875
],
[
0.515625,
0.2666015625,
0.87890625,
0.2666015625,
0.87890625,
0.276611328125,
0.515625,
0.276611328125
],
[
0.12176513671875,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.12176513671875,
0.2998046875
],
[
0.515625,
0.281982421875,
0.880859375,
0.281982421875,
0.880859375,
0.295654296875,
0.515625,
0.295654296875
],
[
0.12005615234375,
0.3017578125,
0.358642578125,
0.3017578125,
0.358642578125,
0.3154296875,
0.12005615234375,
0.3154296875
],
[
0.5146484375,
0.29638671875,
0.8818359375,
0.297607421875,
0.8818359375,
0.312255859375,
0.5146484375,
0.311279296875
],
[
0.51708984375,
0.314208984375,
0.88330078125,
0.314208984375,
0.88330078125,
0.327880859375,
0.51708984375,
0.327880859375
],
[
0.139404296875,
0.320556640625,
0.48779296875,
0.320556640625,
0.48779296875,
0.334228515625,
0.139404296875,
0.334228515625
],
[
0.515625,
0.330078125,
0.8818359375,
0.330078125,
0.8818359375,
0.34326171875,
0.515625,
0.34326171875
],
[
0.12005615234375,
0.336181640625,
0.489501953125,
0.336181640625,
0.489501953125,
0.349853515625,
0.12005615234375,
0.349853515625
],
[
0.51318359375,
0.344482421875,
0.8818359375,
0.345703125,
0.8818359375,
0.35986328125,
0.51318359375,
0.359375
],
[
0.12005615234375,
0.3515625,
0.48779296875,
0.3515625,
0.48779296875,
0.365234375,
0.12005615234375,
0.365234375
],
[
0.51708984375,
0.3623046875,
0.8818359375,
0.3623046875,
0.8818359375,
0.3759765625,
0.51708984375,
0.3759765625
],
[
0.12005615234375,
0.3681640625,
0.4189453125,
0.3681640625,
0.4189453125,
0.3818359375,
0.12005615234375,
0.3818359375
],
[
0.51708984375,
0.377685546875,
0.8818359375,
0.377685546875,
0.8818359375,
0.392578125,
0.51708984375,
0.392578125
],
[
0.5146484375,
0.392578125,
0.87890625,
0.392578125,
0.87890625,
0.40673828125,
0.5146484375,
0.40673828125
],
[
0.1158447265625,
0.40087890625,
0.15869140625,
0.40087890625,
0.15869140625,
0.416259765625,
0.1158447265625,
0.416259765625
],
[
0.15283203125,
0.402099609375,
0.276123046875,
0.402099609375,
0.276123046875,
0.4150390625,
0.15283203125,
0.4150390625
],
[
0.515625,
0.410400390625,
0.8818359375,
0.410400390625,
0.8818359375,
0.42333984375,
0.515625,
0.42333984375
],
[
0.1175537109375,
0.429931640625,
0.280517578125,
0.43115234375,
0.280517578125,
0.446533203125,
0.1175537109375,
0.4453125
],
[
0.51708984375,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.440673828125,
0.51708984375,
0.440673828125
],
[
0.51708984375,
0.4423828125,
0.880859375,
0.4423828125,
0.880859375,
0.4560546875,
0.51708984375,
0.4560546875
],
[
0.118408203125,
0.45361328125,
0.47265625,
0.454833984375,
0.47265625,
0.4697265625,
0.118408203125,
0.468505859375
],
[
0.466796875,
0.457763671875,
0.486083984375,
0.457763671875,
0.486083984375,
0.465576171875,
0.466796875,
0.465576171875
],
[
0.515625,
0.458984375,
0.8818359375,
0.458984375,
0.8818359375,
0.47265625,
0.515625,
0.47265625
],
[
0.12005615234375,
0.471435546875,
0.489501953125,
0.471435546875,
0.489501953125,
0.485107421875,
0.12005615234375,
0.485107421875
],
[
0.5146484375,
0.474365234375,
0.66162109375,
0.474365234375,
0.66162109375,
0.488037109375,
0.5146484375,
0.488037109375
],
[
0.118408203125,
0.482666015625,
0.486083984375,
0.486328125,
0.486083984375,
0.5048828125,
0.118408203125,
0.501953125
],
[
0.12005615234375,
0.50439453125,
0.486083984375,
0.50439453125,
0.486083984375,
0.51708984375,
0.12005615234375,
0.51708984375
],
[
0.51318359375,
0.51318359375,
0.69873046875,
0.51318359375,
0.69873046875,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.12176513671875,
0.51953125,
0.489501953125,
0.51953125,
0.489501953125,
0.533203125,
0.12176513671875,
0.533203125
],
[
0.12005615234375,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.12005615234375,
0.5498046875
],
[
0.515625,
0.54150390625,
0.880859375,
0.54150390625,
0.880859375,
0.5546875,
0.515625,
0.5546875
],
[
0.11419677734375,
0.5498046875,
0.490234375,
0.55029296875,
0.490234375,
0.572265625,
0.11419677734375,
0.57177734375
],
[
0.51708984375,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.51708984375,
0.57177734375
],
[
0.1175537109375,
0.57177734375,
0.489501953125,
0.572265625,
0.489501953125,
0.59033203125,
0.1175537109375,
0.5888671875
],
[
0.51318359375,
0.572265625,
0.880859375,
0.572265625,
0.880859375,
0.5859375,
0.51318359375,
0.5859375
],
[
0.118408203125,
0.5888671875,
0.486083984375,
0.5888671875,
0.486083984375,
0.6025390625,
0.118408203125,
0.6025390625
],
[
0.51318359375,
0.58837890625,
0.880859375,
0.5888671875,
0.880859375,
0.60400390625,
0.51318359375,
0.6025390625
],
[
0.118408203125,
0.60546875,
0.429931640625,
0.60546875,
0.429931640625,
0.619140625,
0.118408203125,
0.619140625
],
[
0.5146484375,
0.60546875,
0.88330078125,
0.60546875,
0.88330078125,
0.619140625,
0.5146484375,
0.619140625
],
[
0.139404296875,
0.62255859375,
0.486083984375,
0.62255859375,
0.486083984375,
0.6357421875,
0.139404296875,
0.6357421875
],
[
0.51318359375,
0.619140625,
0.8818359375,
0.62060546875,
0.8818359375,
0.63525390625,
0.51318359375,
0.63427734375
],
[
0.12005615234375,
0.64013671875,
0.489501953125,
0.64013671875,
0.489501953125,
0.65380859375,
0.12005615234375,
0.65380859375
],
[
0.5146484375,
0.6357421875,
0.880859375,
0.63720703125,
0.880859375,
0.65185546875,
0.5146484375,
0.65087890625
],
[
0.12005615234375,
0.65625,
0.48779296875,
0.65625,
0.48779296875,
0.66943359375,
0.12005615234375,
0.66943359375
],
[
0.515625,
0.65380859375,
0.8818359375,
0.65380859375,
0.8818359375,
0.66748046875,
0.515625,
0.66748046875
],
[
0.12005615234375,
0.6728515625,
0.489501953125,
0.6728515625,
0.489501953125,
0.68603515625,
0.12005615234375,
0.68603515625
],
[
0.5146484375,
0.67041015625,
0.87744140625,
0.67041015625,
0.87744140625,
0.68115234375,
0.5146484375,
0.68115234375
],
[
0.12005615234375,
0.689453125,
0.489501953125,
0.689453125,
0.489501953125,
0.703125,
0.12005615234375,
0.703125
],
[
0.51318359375,
0.68505859375,
0.880859375,
0.68505859375,
0.880859375,
0.69873046875,
0.51318359375,
0.69873046875
],
[
0.51708984375,
0.70361328125,
0.62060546875,
0.70361328125,
0.62060546875,
0.71435546875,
0.51708984375,
0.71435546875
],
[
0.64990234375,
0.70166015625,
0.8818359375,
0.70166015625,
0.8818359375,
0.71533203125,
0.64990234375,
0.71533203125
],
[
0.51708984375,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.51708984375,
0.732421875
],
[
0.12005615234375,
0.705078125,
0.478515625,
0.705078125,
0.478515625,
0.71875,
0.12005615234375,
0.71875
],
[
0.118408203125,
0.7197265625,
0.486083984375,
0.7197265625,
0.486083984375,
0.73291015625,
0.118408203125,
0.73291015625
],
[
0.12005615234375,
0.73681640625,
0.48193359375,
0.73681640625,
0.48193359375,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.515625,
0.7353515625,
0.8818359375,
0.7353515625,
0.8818359375,
0.7490234375,
0.515625,
0.7490234375
],
[
0.12005615234375,
0.7529296875,
0.483642578125,
0.7529296875,
0.483642578125,
0.7666015625,
0.12005615234375,
0.7666015625
],
[
0.5146484375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.12176513671875,
0.77099609375,
0.483642578125,
0.77099609375,
0.483642578125,
0.78076171875,
0.12176513671875,
0.78076171875
],
[
0.515625,
0.7666015625,
0.880859375,
0.7666015625,
0.880859375,
0.77978515625,
0.515625,
0.77978515625
],
[
0.12005615234375,
0.78515625,
0.48779296875,
0.78515625,
0.48779296875,
0.798828125,
0.12005615234375,
0.798828125
],
[
0.5146484375,
0.783203125,
0.880859375,
0.783203125,
0.880859375,
0.79638671875,
0.5146484375,
0.79638671875
],
[
0.118408203125,
0.8017578125,
0.486083984375,
0.8017578125,
0.486083984375,
0.8154296875,
0.118408203125,
0.8154296875
],
[
0.515625,
0.798828125,
0.880859375,
0.798828125,
0.880859375,
0.8125,
0.515625,
0.8125
],
[
0.12176513671875,
0.81787109375,
0.48779296875,
0.81787109375,
0.48779296875,
0.83056640625,
0.12176513671875,
0.83056640625
],
[
0.5146484375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.8291015625,
0.5146484375,
0.8291015625
],
[
0.12176513671875,
0.8330078125,
0.48779296875,
0.8330078125,
0.48779296875,
0.8466796875,
0.12176513671875,
0.8466796875
],
[
0.515625,
0.8330078125,
0.880859375,
0.8330078125,
0.880859375,
0.84326171875,
0.515625,
0.84326171875
],
[
0.12176513671875,
0.849609375,
0.486083984375,
0.849609375,
0.486083984375,
0.86328125,
0.12176513671875,
0.86328125
],
[
0.51318359375,
0.845703125,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51318359375,
0.85986328125
],
[
0.12005615234375,
0.8662109375,
0.43994140625,
0.8662109375,
0.43994140625,
0.8798828125,
0.12005615234375,
0.8798828125
],
[
0.515625,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.515625,
0.876953125
],
[
0.5146484375,
0.87890625,
0.8818359375,
0.87744140625,
0.8818359375,
0.892578125,
0.5146484375,
0.8935546875
],
[
0.139404296875,
0.89453125,
0.450927734375,
0.8955078125,
0.450927734375,
0.91015625,
0.139404296875,
0.9091796875
],
[
0.515625,
0.8955078125,
0.73046875,
0.8955078125,
0.73046875,
0.91015625,
0.515625,
0.91015625
]
] |
[
"词的歧义。由于其性质，WSD在句子中表现",
"2.2Meta-learning in NLP",
"出单词间的依赖性，具有大量的类以及不可",
"NLP中的元学习仍处于起步阶段。Guetal.",
"避免的类不平衡；与大多数当前元学习方法",
"（2018）将元学习应用于神经机器翻译问题，他",
"中的受控设置相比，所有这些都提出了新的",
"挑战。为了解决这些挑战，我们将三种流行",
"们对将高资源语言翻译成英语进行元训练，并",
"的元学习算法扩展到此任务：原型网络（Snell",
"对将低资源语言翻译成英语进行元测试。Oba",
"etal.，2017），与模型无关的元学习（MAML）",
"muyideandVlachos（2019b）使用元学习进行",
"(Finn et al.,2017)及其混合-ProtoMAML（Tri-",
"关系分类，ObamuyideandVlachos（2019a）",
"antafillouetal.，2020）。我们使用三种底层模",
"用于终身学习环境中的关系提取。Chenetal.",
"型架构来研究元学习，即递归网络，多层感知",
"(2019）考虑关系学习并将元学习应用于知识",
"器（MLP）和转换器（Vaswanietal.，2017），",
"图中的少拍链接预测。Douetal.（2019）对",
"并尝试使用不同数量的句子进行特定于任务的",
"来自GLUE基准（Wangetal.，2018）的某些",
"微调。我们通过在一组看不见的新词上进行测",
"高资源任务执行元训练，并对来自同一基准",
"试来评估模型的快速适应能力，从而证明了它",
"的某些低资源任务执行元测试。Bansaletal",
"从少量示例中学习新词的能力",
"(2019）提出了一个softmax参数生成器组件，",
"可以在元训练任务中启用不同数量的类。他",
"由于没有可用的少量WSD基准，因此我们",
"们选择GLUE中的任务以及SNLI（Bowman",
"创建了公开可用的WSD数据集的少量版本。",
"etal.，2015）进行元训练，并使用实体类型，关",
"我们发布了我们的代码以及用于生成少量数据",
"系分类，情感分类，文本分类和科学NLI作为",
"设置的脚本，以促进进一步的研究。1",
"测试任务。元学习也被探索用于少数文本分类",
"(Yu et al., 2018: Geng et al., 2019: Jiang et al.",
"2",
"Related Work",
"2018;Sun etal.,2019)。Wuet al.(2019)采",
"2.1 Meta-learning",
"用元强化学习技术进行多标签分类，并进行了",
"实体类型和文本分类的实验。Huetal.（2019）",
"与“传统”机器学习方法相比，元学习涉及",
"一",
"使用元学习来学习词汇外单词的表示形式，并",
"种不同的范式，称为情节学习。元学习中的训",
"将其作为回归任务",
"练集和测试集分别称为元训练集（Dmeta-train",
"）和元测试集（Dmeta-test）。两组均由情节组",
"2.3Supervised WSD",
"成，而不是单个数据点。每个情节构成一个任",
"务T，其中包括少量用于适应的训练示例-支",
"早期的WSD监督学习方法依赖于从上下",
"持集 Delupport，以及一组单独的用于评估的示",
"文单词（LeeandNg，2002；Navigli，2009;",
"例-查询集Dquery。元学习的典型设置是平衡",
"Zhong andNg，2010）中提取的手工特征。后",
"的N-way，K-shot设置，其中每个剧集都有",
"来的工作使用单词嵌人作为分类（Taghipour",
"N类，其支持集中每个类都有K示例",
"and Ng, 2015; Rothe and Schitze, 2015; Ia",
"元学习算法大致分为三种类型：基于度量",
"cobacci etal.，2016）的功能。随着深度学习",
"的 (Koch et al., 2015; Vinyals et al., 2016;",
"的兴起，LSTM(Hochreiter and Schmidhuber,",
"Sung et al.,2018;Snell et al.,2017)，基于",
"1997）模型成为流行的（Melamudetal.，2016;",
"模型的(Santoro etal.,2016;Munkhdalaiand",
"Kageback and Salomonsson. 2016: Yuan et al.",
"Yu，2017）和基于优化的（RaviandLarochelle,",
"2016）。虽然大多数工作每个单词训练单个模",
"型，但Ragal",
"etal.（2017b）设计了一个具",
"有大量输出单元的LSTM模型来消除所有单",
"2017; Finn et al., 2017; Nichol et al., 2018)",
"基于度量的方法首先将每个情节中的示例嵌入",
"到高维空间中，通常使用神经网络。接下来",
"词的歧义。Petersetal.（2018）通过与上下文化",
"他们基于测量与支持示例的相似性的核函数",
"ELMo（Petersetal.，2018）嵌人匹配的最近邻",
"获得所有香询示例的标签上的概率分布。基于",
"执行WSD。Hadiwinotoetal.（2019）使用来自",
"模型的方法旨在直接通过其体系结构实现快速",
"BERT（Devlinetal.，2019）的预先训练的上下",
"学习。他们通常使用外部存储器，以便记住过",
"文化表示作为特征。GlossBERT（Huangetal.",
"去遇到的关键示例。基于优化的方法在其目标",
"2019）结合了WordNet（Miller et al.,1990）的",
"函数中明确包括泛化性，并针对同一目标函数",
"意义定义，以形成上下文-光泽对，同时微调",
"进行优化。在本文中，我们尝试了基于度量和",
"WSD的BERT。通过考虑WordNet的图形结",
"基于优化的方法，以及它们的混合方法。",
"构，EWISER (Bevilacqua and Navigli, 2020)",
"达到了当前最先进的F1分数80。1%上的基",
"Ihttps://github.com/Nithin-Holla/MetaWSD",
"准 Raganato et al. (2017a)"
] |
{
"arxiv_id": "2004.14355",
"language": "zh",
"page_idx": 1
}
|
||
2004.14355-zh-2
|
2004.14355-zh-2.jpg
|
## 3 Task and Dataset
我们将 WSD 视为少数单词级分类问题，其中在给定单词的句子上下文的情况下将其分配给单词。由于不同的单词可能具有不同的感官数量，并且句子可能具有多个歧义的单词，因此$N$ -way , $K$ -shot 分类的标准设置在我们的情况下不成立。具体来说，不同的剧集可以有不同数量的类和每个类不同数量的示例-一个更现实的设置 (Triantafillou et al., 2020)
Dataset 我们使用 SemCor 语料库 (Miller et al., 1994) , 该语料库由 Yuan et al. (2016) $^{2}$的新牛津美国词典中的感官手动注释。使用37,176注释的句子，这是最大的带有意义注释的英语语料库之一。我们根据要消除歧义的单词对语料库中的句子进行分组，然后使用 60 :15:25拆分将单词随机分为不相交的元训练，元验证和元测试集。一个句子可能会多次出现同一个单词，在这种情况下，我们会对所有单词进行预测。我们考虑四种不同的设置， 支持集大小$| S| = 4$ , 8 , 16 和 32 句子。元训练/元测试集中的不同单词数分别为 985/270 、 985/259 、799/197 和 580/129 。结果数据集的详细统计信息见附录 A.1
Training episodes 在元训练集中，支持集和查询集具有相同数量的句子。由于发作次数不足，我们在元训练期间每集使用一个单词的初始实验产生了较差的结果。为了克服这个问题并设计一个合适的元训练设置，我们改为创建带有多个注释单词的剧集。具体来说，每一集都由$r$采样的单词$\{z_j\}_j=1^r$和$min(\lfloor|S|/r\rfloor,\nu(z_j))$ 的每个单词的感觉组成， 其中$\nu(z_j)$是单词$z_j$的感觉数。因此，元训练集中的每个任务都是$r$单词在多达$|S|$感官之间的歧义消除。我们为$|S|=4$设置$r= 2$ , 为其余设置$r=4$ 。然后对包含这些感觉的句子进行采样，以获取支持和查询集，以便尽可能平衡类。例如，对于$|S|=8$,我们首先为每个单词选择 4 个单词和 2 个感官，然后为每个单词-感官对采样一个句子。感官的标签在不同的情节中被洗牌，我。e。,当在另一集中采样时，一种感觉可以有不同的标签。这是元学习的关键，因为它可以防止记忆(Yin et al., 2020) 。我们构建元训练集的方法的优势在于， 它允许生成组合的大量任务。在这里，我们使用10，000元训练集的总数
Evaluation episodes 对于元验证和元测试集，每个情节都对应于消除单个单词歧义的任
$^2$https://tinyurl.com/wsdcrp. Note that the corpus does not have a standard train/validation/test split.
务。在将句子拆分为支持集和查询集时，我们确保查询集中的意义存在于支持集中。此外， 我们只考虑在其查询集中具有两种或多种意义的单词。发作在不同数量的感官中的分布见附录 A.1 。请注意，与元训练任务不同，我们的元测试任务代表了自然的数据分布，因此允许我们在现实环境中测试我们的模型
## 4 Methods
我们的模型由三个组件组成：一个编码器，它将句子中的单词作为输入，并为每个单词生成上下文表示，一个隐藏的线性层，将这些表示投影到另一个空间，以及一个输出线性层，产生感官上的概率分布。编码器和隐藏层在所有任务中共享-我们将此块表示为具有共享参数$\boldsymbol{\theta}$的$f_{\boldsymbol{\theta}}$ 。输出层为每个任务随机初始化$\mathcal{T}_i$ (i。e。情节) -我们将其表示为带有参数$\phi_i$的$g_{\phi_i}$ 。$\theta$是元学习的，而$\phi_i$是为每个任务独立学习的。
### 4.1 Model Architectures
我们使用三种不同的体系结构进行实验：(1) 单层双向 GRU (Cho et al.,2014) ,其 GloVe 嵌入 (Pennington et al., 2014) 作为未微调的输入；(2) 未微调的 ELMo (Peters et al., 2018)嵌入，然后是MLP；(3)BERT_{BASE (Devlin et al.,2019)。我们三种不同模型的架构-GloVe+GRU,ELMo+MLP 和 BERT-如图1 所示
4.2 Meta-learning Methods
Prototypical Networks 由 Snell et al. (2017) 提出，原型网络是一种基于度量的方法。由$\theta$ 参数化的嵌入网络$f_{\boldsymbol{\theta}}$用于为每个类生成一个原型向量，作为该类的所有支持数据点的嵌入的平均向量。假设$S_c$表示包含类$c\in C$的示例的支持集的子集，则原型$\mu_c$为：
(1)
$$\mu_c=\frac{1}{|S_c|}\sum\limits_{\boldsymbol{x}_i\in S_c}f_{\boldsymbol{\theta}}(\boldsymbol{x}_i)$$
给定在嵌入空间上定义的距离函数，查询点在类上的分布计算为到类原型的负距离上的softmax
我们从支持示例的共享块$f_{\boldsymbol{\theta}}$的输出中生成原型(每个意义一个)。我们不使用$g_{\phi_i}$,而是根据距离函数获得查询示例的概率分布。参数$\theta$在使用 Adam 优化器 (Kingma and Ba, 2015) 的每个情节之后更新：
$$\theta\leftarrow\mathrm{Adam}(\mathcal{L}_{\mathcal{T}_i}^q,\boldsymbol{\theta},\beta)$$
(2)
其中$\mathcal{L}_{\mathcal{T}_i}^q$是查询集上的交叉熵损失，$\beta$是元学
习率
|
[
[
0.118408203125,
0.07659912109375,
0.30908203125,
0.07659912109375,
0.30908203125,
0.09027099609375,
0.118408203125,
0.09027099609375
],
[
0.515625,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.0914306640625,
0.515625,
0.0914306640625
],
[
0.515625,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.515625,
0.10687255859375
],
[
0.12005615234375,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.12005615234375,
0.11639404296875
],
[
0.515625,
0.11102294921875,
0.880859375,
0.11102294921875,
0.880859375,
0.1217041015625,
0.515625,
0.1217041015625
],
[
0.12005615234375,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.5146484375,
0.1258544921875,
0.8818359375,
0.1258544921875,
0.8818359375,
0.1395263671875,
0.5146484375,
0.1395263671875
],
[
0.12176513671875,
0.13720703125,
0.486083984375,
0.13720703125,
0.486083984375,
0.1478271484375,
0.12176513671875,
0.1478271484375
],
[
0.5146484375,
0.141357421875,
0.8818359375,
0.141357421875,
0.8818359375,
0.155029296875,
0.5146484375,
0.155029296875
],
[
0.12176513671875,
0.1519775390625,
0.489501953125,
0.1519775390625,
0.489501953125,
0.1650390625,
0.12176513671875,
0.1650390625
],
[
0.515625,
0.157958984375,
0.8818359375,
0.157958984375,
0.8818359375,
0.171630859375,
0.515625,
0.171630859375
],
[
0.118408203125,
0.166259765625,
0.48779296875,
0.16748046875,
0.48779296875,
0.1817626953125,
0.118408203125,
0.18115234375
],
[
0.5146484375,
0.17333984375,
0.791015625,
0.17333984375,
0.791015625,
0.18701171875,
0.5146484375,
0.18701171875
],
[
0.12005615234375,
0.18408203125,
0.48779296875,
0.18408203125,
0.48779296875,
0.19775390625,
0.12005615234375,
0.19775390625
],
[
0.12005615234375,
0.1995849609375,
0.486083984375,
0.1995849609375,
0.486083984375,
0.213134765625,
0.12005615234375,
0.213134765625
],
[
0.509765625,
0.1995849609375,
0.62744140625,
0.1995849609375,
0.62744140625,
0.213134765625,
0.509765625,
0.213134765625
],
[
0.12005615234375,
0.215576171875,
0.4248046875,
0.215576171875,
0.4248046875,
0.2286376953125,
0.12005615234375,
0.2286376953125
],
[
0.5146484375,
0.22265625,
0.8818359375,
0.223876953125,
0.8818359375,
0.2381591796875,
0.5146484375,
0.2369384765625
],
[
0.1259765625,
0.2423095703125,
0.48779296875,
0.2423095703125,
0.48779296875,
0.255859375,
0.1259765625,
0.255859375
],
[
0.5146484375,
0.240478515625,
0.88330078125,
0.240478515625,
0.88330078125,
0.254150390625,
0.5146484375,
0.254150390625
],
[
0.12005615234375,
0.260009765625,
0.484375,
0.260009765625,
0.484375,
0.270751953125,
0.12005615234375,
0.270751953125
],
[
0.515625,
0.255859375,
0.8818359375,
0.255859375,
0.8818359375,
0.26953125,
0.515625,
0.26953125
],
[
0.1158447265625,
0.270751953125,
0.489501953125,
0.2724609375,
0.489501953125,
0.290283203125,
0.1158447265625,
0.2880859375
],
[
0.5146484375,
0.2724609375,
0.88330078125,
0.2724609375,
0.88330078125,
0.2861328125,
0.5146484375,
0.2861328125
],
[
0.1175537109375,
0.289306640625,
0.489501953125,
0.290283203125,
0.489501953125,
0.3046875,
0.1175537109375,
0.303955078125
],
[
0.5146484375,
0.289306640625,
0.8818359375,
0.289306640625,
0.8818359375,
0.302734375,
0.5146484375,
0.302734375
],
[
0.12005615234375,
0.30712890625,
0.484375,
0.30712890625,
0.484375,
0.317626953125,
0.12005615234375,
0.317626953125
],
[
0.5146484375,
0.3046875,
0.88330078125,
0.3046875,
0.88330078125,
0.318359375,
0.5146484375,
0.318359375
],
[
0.12005615234375,
0.322509765625,
0.489501953125,
0.322509765625,
0.489501953125,
0.336181640625,
0.12005615234375,
0.336181640625
],
[
0.5146484375,
0.320556640625,
0.88330078125,
0.320556640625,
0.88330078125,
0.334228515625,
0.5146484375,
0.334228515625
],
[
0.118408203125,
0.33740234375,
0.489501953125,
0.33837890625,
0.489501953125,
0.3515625,
0.118408203125,
0.350830078125
],
[
0.51123046875,
0.3349609375,
0.87890625,
0.33740234375,
0.87890625,
0.3515625,
0.51123046875,
0.349853515625
],
[
0.51708984375,
0.355224609375,
0.53076171875,
0.355224609375,
0.53076171875,
0.363525390625,
0.51708984375,
0.363525390625
],
[
0.52490234375,
0.352783203125,
0.880859375,
0.352783203125,
0.880859375,
0.366455078125,
0.52490234375,
0.366455078125
],
[
0.51318359375,
0.3681640625,
0.5634765625,
0.3681640625,
0.5634765625,
0.383056640625,
0.51318359375,
0.383056640625
],
[
0.12005615234375,
0.355712890625,
0.489501953125,
0.355712890625,
0.489501953125,
0.369384765625,
0.12005615234375,
0.369384765625
],
[
0.12005615234375,
0.371826171875,
0.489501953125,
0.371826171875,
0.489501953125,
0.385498046875,
0.12005615234375,
0.385498046875
],
[
0.12005615234375,
0.38720703125,
0.484375,
0.38720703125,
0.484375,
0.40087890625,
0.12005615234375,
0.40087890625
],
[
0.51318359375,
0.394287109375,
0.72705078125,
0.394287109375,
0.72705078125,
0.407958984375,
0.51318359375,
0.407958984375
],
[
0.1234130859375,
0.405029296875,
0.486083984375,
0.405029296875,
0.486083984375,
0.4150390625,
0.1234130859375,
0.4150390625
],
[
0.12005615234375,
0.419189453125,
0.478515625,
0.419189453125,
0.478515625,
0.432861328125,
0.12005615234375,
0.432861328125
],
[
0.5146484375,
0.41455078125,
0.88330078125,
0.41455078125,
0.88330078125,
0.427490234375,
0.5146484375,
0.427490234375
],
[
0.474365234375,
0.4228515625,
0.48779296875,
0.4228515625,
0.48779296875,
0.4287109375,
0.474365234375,
0.4287109375
],
[
0.118408203125,
0.435302734375,
0.486083984375,
0.435302734375,
0.486083984375,
0.448974609375,
0.118408203125,
0.448974609375
],
[
0.515625,
0.43115234375,
0.88330078125,
0.43115234375,
0.88330078125,
0.44482421875,
0.515625,
0.44482421875
],
[
0.5146484375,
0.446533203125,
0.880859375,
0.446533203125,
0.880859375,
0.460205078125,
0.5146484375,
0.460205078125
],
[
0.12176513671875,
0.453125,
0.337646484375,
0.453125,
0.337646484375,
0.463134765625,
0.12176513671875,
0.463134765625
],
[
0.51318359375,
0.463134765625,
0.8818359375,
0.463134765625,
0.8818359375,
0.476806640625,
0.51318359375,
0.476806640625
],
[
0.1275634765625,
0.47802734375,
0.489501953125,
0.47802734375,
0.489501953125,
0.4921875,
0.1275634765625,
0.4921875
],
[
0.51123046875,
0.473876953125,
0.88330078125,
0.476806640625,
0.88330078125,
0.49755859375,
0.51123046875,
0.49462890625
],
[
0.12005615234375,
0.49462890625,
0.489501953125,
0.49462890625,
0.489501953125,
0.50830078125,
0.12005615234375,
0.50830078125
],
[
0.5146484375,
0.496337890625,
0.87890625,
0.496337890625,
0.87890625,
0.50732421875,
0.5146484375,
0.50732421875
],
[
0.118408203125,
0.51025390625,
0.486083984375,
0.51025390625,
0.486083984375,
0.52392578125,
0.118408203125,
0.52392578125
],
[
0.5146484375,
0.51123046875,
0.8818359375,
0.51123046875,
0.8818359375,
0.52490234375,
0.5146484375,
0.52490234375
],
[
0.12176513671875,
0.52685546875,
0.48779296875,
0.52685546875,
0.48779296875,
0.54052734375,
0.12176513671875,
0.54052734375
],
[
0.51123046875,
0.52783203125,
0.568359375,
0.52783203125,
0.568359375,
0.5419921875,
0.51123046875,
0.5419921875
],
[
0.12005615234375,
0.5419921875,
0.48779296875,
0.5419921875,
0.48779296875,
0.5556640625,
0.12005615234375,
0.5556640625
],
[
0.51318359375,
0.552734375,
0.75146484375,
0.552734375,
0.75146484375,
0.56640625,
0.51318359375,
0.56640625
],
[
0.12176513671875,
0.56005859375,
0.486083984375,
0.56005859375,
0.486083984375,
0.57080078125,
0.12176513671875,
0.57080078125
],
[
0.1175537109375,
0.57177734375,
0.489501953125,
0.57373046875,
0.489501953125,
0.59130859375,
0.1175537109375,
0.5888671875
],
[
0.51708984375,
0.57177734375,
0.8818359375,
0.572265625,
0.8818359375,
0.58740234375,
0.51708984375,
0.5859375
],
[
0.12176513671875,
0.59130859375,
0.489501953125,
0.59130859375,
0.489501953125,
0.60498046875,
0.12176513671875,
0.60498046875
],
[
0.5146484375,
0.5888671875,
0.88330078125,
0.5888671875,
0.88330078125,
0.6025390625,
0.5146484375,
0.6025390625
],
[
0.12005615234375,
0.60693359375,
0.48779296875,
0.60693359375,
0.48779296875,
0.62060546875,
0.12005615234375,
0.62060546875
],
[
0.5146484375,
0.60546875,
0.8818359375,
0.60546875,
0.8818359375,
0.619140625,
0.5146484375,
0.619140625
],
[
0.12176513671875,
0.62451171875,
0.486083984375,
0.62451171875,
0.486083984375,
0.63525390625,
0.12176513671875,
0.63525390625
],
[
0.5146484375,
0.62060546875,
0.880859375,
0.62060546875,
0.880859375,
0.63427734375,
0.5146484375,
0.63427734375
],
[
0.118408203125,
0.63818359375,
0.490234375,
0.63818359375,
0.490234375,
0.65185546875,
0.118408203125,
0.65185546875
],
[
0.5146484375,
0.63720703125,
0.880859375,
0.63720703125,
0.880859375,
0.65087890625,
0.5146484375,
0.65087890625
],
[
0.12005615234375,
0.65380859375,
0.48779296875,
0.65380859375,
0.48779296875,
0.66748046875,
0.12005615234375,
0.66748046875
],
[
0.51318359375,
0.65283203125,
0.7958984375,
0.65380859375,
0.7958984375,
0.66845703125,
0.51318359375,
0.66748046875
],
[
0.12005615234375,
0.67041015625,
0.48779296875,
0.67041015625,
0.48779296875,
0.68408203125,
0.12005615234375,
0.68408203125
],
[
0.6455078125,
0.66650390625,
0.8017578125,
0.6728515625,
0.796875,
0.72705078125,
0.640625,
0.72021484375
],
[
0.66650390625,
0.6748046875,
0.7001953125,
0.6787109375,
0.69775390625,
0.689453125,
0.6630859375,
0.68505859375
],
[
0.61279296875,
0.68115234375,
0.6689453125,
0.6884765625,
0.6650390625,
0.703125,
0.60791015625,
0.6953125
],
[
0.85791015625,
0.681640625,
0.8818359375,
0.681640625,
0.8818359375,
0.69677734375,
0.85791015625,
0.69677734375
],
[
0.12005615234375,
0.70166015625,
0.486083984375,
0.70166015625,
0.486083984375,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.118408203125,
0.68603515625,
0.486083984375,
0.68603515625,
0.486083984375,
0.69970703125,
0.118408203125,
0.69970703125
],
[
0.12005615234375,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.12005615234375,
0.732421875
],
[
0.5146484375,
0.7197265625,
0.8818359375,
0.7197265625,
0.8818359375,
0.73291015625,
0.5146484375,
0.73291015625
],
[
0.12005615234375,
0.736328125,
0.48779296875,
0.736328125,
0.48779296875,
0.75,
0.12005615234375,
0.75
],
[
0.5146484375,
0.736328125,
0.8818359375,
0.736328125,
0.8818359375,
0.75,
0.5146484375,
0.75
],
[
0.12176513671875,
0.75439453125,
0.48779296875,
0.75439453125,
0.48779296875,
0.76416015625,
0.12176513671875,
0.76416015625
],
[
0.5146484375,
0.7529296875,
0.580078125,
0.7529296875,
0.580078125,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.12176513671875,
0.7685546875,
0.48779296875,
0.7685546875,
0.48779296875,
0.7822265625,
0.12176513671875,
0.7822265625
],
[
0.5322265625,
0.7685546875,
0.88330078125,
0.7685546875,
0.88330078125,
0.7822265625,
0.5322265625,
0.7822265625
],
[
0.12005615234375,
0.78369140625,
0.489501953125,
0.78369140625,
0.489501953125,
0.79736328125,
0.12005615234375,
0.79736328125
],
[
0.51123046875,
0.78076171875,
0.88330078125,
0.783203125,
0.88330078125,
0.80029296875,
0.51123046875,
0.798828125
],
[
0.1234130859375,
0.8017578125,
0.48779296875,
0.8017578125,
0.48779296875,
0.8125,
0.1234130859375,
0.8125
],
[
0.51708984375,
0.8017578125,
0.880859375,
0.8017578125,
0.880859375,
0.8125,
0.51708984375,
0.8125
],
[
0.12005615234375,
0.81640625,
0.334228515625,
0.81640625,
0.334228515625,
0.830078125,
0.12005615234375,
0.830078125
],
[
0.5146484375,
0.81640625,
0.8818359375,
0.81640625,
0.8818359375,
0.830078125,
0.5146484375,
0.830078125
],
[
0.51318359375,
0.83056640625,
0.73779296875,
0.83203125,
0.73779296875,
0.8466796875,
0.51318359375,
0.845703125
],
[
0.1259765625,
0.84130859375,
0.48779296875,
0.8427734375,
0.48779296875,
0.85693359375,
0.1259765625,
0.85546875
],
[
0.61376953125,
0.85546875,
0.7841796875,
0.85546875,
0.7841796875,
0.87060546875,
0.61376953125,
0.87060546875
],
[
0.8564453125,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.869140625,
0.8564453125,
0.869140625
],
[
0.515625,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.89453125,
0.515625,
0.89453125
],
[
0.12005615234375,
0.859375,
0.48779296875,
0.859375,
0.48779296875,
0.87255859375,
0.12005615234375,
0.87255859375
],
[
0.1368408203125,
0.8818359375,
0.48779296875,
0.88427734375,
0.48779296875,
0.8984375,
0.1368408203125,
0.896484375
],
[
0.12005615234375,
0.8955078125,
0.47265625,
0.896484375,
0.47265625,
0.91015625,
0.12005615234375,
0.9091796875
],
[
0.5146484375,
0.8935546875,
0.55517578125,
0.8955078125,
0.55322265625,
0.9111328125,
0.51318359375,
0.9091796875
]
] |
[
"3Task and Dataset",
"务。在将句子拆分为支持集和查询集时，我们",
"确保查询集中的意义存在于支持集中。此外，",
"我们将WSD视为少数单词级分类问题，其中",
"我们只考虑在其查询集中具有两种或多种意义",
"在给定单词的句子上下文的情况下将其分配给",
"的单词。发作在不同数量的感官中的分布见附",
"单词。由手不同的单词可能具有不同的感官数",
"录A.1。请注意，与元训练任务不同，我们的",
"量，并且句子可能具有多个歧义的单词，因此",
"元测试任务代表了自然的数据分布，因此允许",
"N-way，K-shot分类的标准设置在我们的情",
"我们在现实环境中测试我们的模型",
"况下不成立。具体来说，不同的剧集可以有不",
"同数量的类和每个类不同数量的示例-一个更",
"4Methods",
"现实的设置（Triantafillouetal，2020）",
"我们的模型由三个组件组成：一个编码器，它",
"Dataset我们使用SemCor语料库（Miller",
"将句子中的单词作为输入，并为每个单词生成",
"etal.1994）该语料库由Yuanetal.（2016）",
"上下文表示，一个隐藏的线性层，将这些表示",
"2的新牛津美国词典中的感官手动注释。使用",
"投影到另一个空间，以及一个输出线性层，产",
"37,176注释的句子，这是最大的带有意义注释",
"生感官上的概率分布。编码器和隐藏层在所有",
"的英语语料库之一。我们根据要消除歧义的单",
"任务中共享-我们将此块表示为具有共享参数",
"词对语料库中的句子进行分组，然后使用60",
"e的fe。输出层为每个任务随机初始化T（i。",
"：15：25拆分将单词随机分为不相交的元训",
"e。情节）-我们将其表示为带有参数Φi的9p",
"0",
"θ是元学习的，而Φi是为每个任务独立学",
"习的。",
"练，元验证和元测试集。一个句子可能会多次",
"出现同一个单词，在这种情况下，我们会对所",
"有单词进行预测。我们考虑四种不同的设置",
"4.1Model Architectures",
"支持集天小S=4，8，16和32句子。元训",
"练/元测试集中的不同单词数分别为985/270",
"我们使用三种不同的体系结构进行实验：（1）",
"",
"985/259、799/197和580/129。结果数据集",
"单层双向GRU（Choetal.，2014），其GloVe",
"嵌人（Penningtonetal.，2014）作为未微调的",
"的详细统计信息见附录A.1",
"输人；（2）未微调的ELMo(Peterset al.",
"Trainingepisodes在元训练集中，支持集",
"2018）嵌人，然后是MLP；（3）BERTBASE",
"和查询集具有相同数量的句子。由于发作次",
"（Devlinetal.，2019）。我们三种不同模型的架",
"数不足，我们在元训练期间每集使用一个单",
"构-GloVe+GRU，ELMo+MLP和BERT-如图",
"词的初始实验产生了较差的结果。为了克服",
"1所示",
"这个问题并设计一个合适的元训练设置，我",
"4.2Meta-learning Methods",
"们改为创建带有多个注释单词的剧集。具体",
"来说，每一集都由采样的单词{=1和",
"Prototypical Networks由Snell et al.(2017)",
"min(LIS//r],v(zi)）的每个单词的感觉组成，",
"提出，原型网络是一种基于度量的方法。由θ",
"其中v(zi）是单词z的感觉数。因此，元训",
"参数化的嵌人网络fe用于为每个类生成一个",
"练集中的每个任务都是r单词在多达ST感官",
"原型向量，作为该类的所有支持数据点的嵌人",
"之间的歧义消除。我们为ISI=4设置r=2，",
"的平均向量。假设S表示包含类CEC的示",
"为其余设置r=4。然后对包含这些感觉的句",
"例的支持集的子集，则原型μc为：",
"子进行采样，以获取支持和查询集，以便尽可",
"BZ",
"1",
"e=1",
"(1)",
"每个单词选择4个单词和2个感官，然后为每",
"能平衡类。例如，对于IS=8，我们首先为",
"个单词-感官对采样一个句子。感官的标签在",
"给定在嵌入空间上定义的距离函数，查询点",
"不同的情节中被洗牌，我。e。，当在另一集中",
"在类上的分布计算为到类原型的负距离上的",
"采样时，一种感觉可以有不同的标签。这是元",
"softmax",
"学习的关键，因为它可以防止记忆（Yinetal.，",
"我们从支持示例的共享块fe的输出中生成",
"2020）。我们构建元训练集的方法的优势在于，",
"原型（每个意义一个）。我们不使用9中；，而",
"它充许生成组合的大量任务。在这里，我们使",
"是根据距离函数获得查询示例的概率分布。参",
"用10,000元训练集的总数",
"数在使用Adam优化器（KingmaandBa,",
"2015）的每个情节之后更新：",
"Evaluationepisodes对于元验证和元测试",
"0←Adam(CT,0,β)",
"(2)",
"其中C%是查询集上的交叉损失，β是元学",
"集，每个情节都对应于消除单个单词歧义的任",
"2https://tinyurl.com/wsdcrp.Notethat the corpus",
"does not have a standard train/validation/test split.",
"习率"
] |
{
"arxiv_id": "2004.14355",
"language": "zh",
"page_idx": 2
}
|
||
2004.14355-zh-4
|
2004.14355-zh-4.jpg
|
BERT 输出执行最近邻匹配，但不使用 GloVe 平均值和标准差。我们注意到，结果在$|S|$设嵌入，因为它们对于所有感官都是相同的。我 置中不能直接比较，因为根据它们的表述，它们将此基线称为 NearestNeighbor
Non-episodic training 它是一个单一的模型， 在所有任务上进行训练，没有任何区别-它合并了支持和查询集，并使用迷你批处理进行训练。因此，输出层与任务无关，输出单元的数量等于数据集中的感官总数。输出层的 softmax 仅在小批量中的相关类上获取。而不是每个任务的$\phi_i$ ,我们现在有一个单一的$\phi$ 。在训练期间，每个小批的参数更新为：
$$\theta,\phi\leftarrow\mathrm{Adam}(\mathcal{L}_{\mathcal{T}_i},\boldsymbol{\theta},\phi,\alpha)$$
(9)
其中$\alpha$是学习率。在元测试阶段，我们在每个任务的支持集上独立地微调训练模型(以情节方式),如下所示：
(10)
$$\theta_i',\phi_i'\leftarrow\mathrm{SGD}(\mathcal{L}_{\mathcal{T}_i},\boldsymbol{\theta},\boldsymbol{\phi},\alpha,\gamma,m)$$
, 其中损失是在支持示例上计算的$,\alpha$是以前的学习者学习率，$\gamma$是输出学习率。我们将此模型称为 NE-Baseline
机初始化的模型开始执行元测试。这相当于从零开始对每集的支持示例进行培训。我们在前缀 EF-前面加上前缀来表示
5 Experiments and Results
## 5.1 Experimental setup
我们使用元验证集为模型选择最佳超参数。所的度量。如果指标在两个时期内没有改善，我们通过终止训练来提前停止。GloVe+GRU 中隐藏状态的大小为 256 ,对于 GloVe+GRU , ELMo+MLP 和 BERT, 共享线性层的大小分别为 64 、256 和 192 。共享线性层的激活函数固定数量的 5612 单元。我们使用 higher 包(Grefenstette et al.,2019)来实现 MAML 变体
## 5.2 Results
们涉及不同的元测试事件
GloVe+GRU 所有元学习方法的性能均优于其 EF 对应方法，表明从元训练集中成功学习。当$|S|=8,16,32$时，FOMAML 无法超越NE-Baseline 以及其他元学习方法的 EF 版本。有趣的是，单独的元测试通常比 NE-Baseline 模型更好，这表明后者不能有效地从元训练集中转移知识。ProtoNet 是性能最好的模型$(|S|=8$除外), ProtoMAML 紧随其后
ELMo+MLP 与基于 GloVe 的模型相比， NearestNeighbor, NE-Baseline 和 EF 方法的得分更高，这可以归因于输入嵌入是上下文的。由于元训练，ProtoNet 和 ProtoFOMAML 仍然比 EF 同行有所改善。ProtoFOMAML 优于所有$\left|S\right|$的其他方法，FOMAML 相对较弱
BERT 除 NE-Baseline 和 FOMAML 外，所有方法的分数都高于以前体系结构的分数。基于 BERT 的 ProtoNet 以及 ProtoFOMAML 优于所有$\left|S\right|$的所有其他方法。此外，对于$| S| = 4, 8$ ,ProtoFOMAML 优于 ProtoNet,对
Episodic fne-tuning baseline 对于每种元学 于$|S|=16,32$ ,反之亦然。总体而言，在整个架习方法，我们都包含一个变体，该变体仅从随 构中，我们看到 NE-Baseline 和 FOMAML 始
终表现不佳，而 ProtoNet 和 ProtoFOMAML 是最有效的方法。此外，它们只需 4 个训练示例即可实现较高的消歧性能，在许多情况下， 它们接近针对个人感官的一次性分类设置(请参见附录 A.1 )。模型也相对稳定，如运行期间的低标准偏差所示
Effect of second-order gradients 我们进一
选的评估指标是元验证集中所有单词的宏 F1 步实验了 ProtoMAML,包括二阶梯度。在分数的平均值。我们在元测试集上报告相同 表 2 中，我们将其 F1 分数与 ProtoNet 和
ProtoFOMAML 一起报告。对于 BERT, 我们训练 ProtoMAML,同时仅对顶层进行微调， 并且由于其较高的计算成本，仅对一个内循环更新步骤进行微调。我们还训练了一个等效的ProtoFOMAML 变体进行公平的比较。我们可
对于 GloVe+GRU 为 tanh,对于 ELMo+MLP 以观察到 ProtoMAML 在大多数情况下获得和 BERT 为 ReLU。对于 FOMAML、Proto- 与 ProtoFOMAML 相似的分数，表明一阶近FOMAML 和 ProtoMAML, 批处理大小设置 似的有效性。在某些情况下，ProtoFOMAML 为 16 任务。这些输出层在每个情节中都会 的得分高于 ProtoMAML, 这可能是由于后者重新初始化，而在 NE-Baseline 中，它具有 引起的过度拟合效应。根据这些结果，我们认
为一阶 ProtoFOMAML 足以完成这项任务
5.3 Analysis
Effect of number of episodes 我们首先调查使用更多的元训练集是否总是转化为更高的
在表 1 中，我们报告了元测试集中所有单词的 性能。随着元训练次数的增加，我们绘制了我平均宏 F1 分数。我们报告了五次独立运行的 们的一个高得分模型-带有 BERT 的 ProtoNet
|
[
[
0.12005615234375,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.51123046875,
0.07598876953125,
0.880859375,
0.07598876953125,
0.880859375,
0.08905029296875,
0.51123046875,
0.08905029296875
],
[
0.12005615234375,
0.0926513671875,
0.48779296875,
0.0926513671875,
0.48779296875,
0.105712890625,
0.12005615234375,
0.105712890625
],
[
0.497802734375,
0.097412109375,
0.5146484375,
0.097412109375,
0.5146484375,
0.1009521484375,
0.497802734375,
0.1009521484375
],
[
0.5146484375,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.12005615234375,
0.1092529296875,
0.387939453125,
0.1092529296875,
0.387939453125,
0.12353515625,
0.12005615234375,
0.12353515625
],
[
0.515625,
0.10986328125,
0.71533203125,
0.10986328125,
0.71533203125,
0.12353515625,
0.515625,
0.12353515625
],
[
0.124267578125,
0.1341552734375,
0.490234375,
0.1341552734375,
0.490234375,
0.1478271484375,
0.124267578125,
0.1478271484375
],
[
0.51904296875,
0.1318359375,
0.880859375,
0.1318359375,
0.880859375,
0.1455078125,
0.51904296875,
0.1455078125
],
[
0.12176513671875,
0.15087890625,
0.489501953125,
0.15087890625,
0.489501953125,
0.1644287109375,
0.12176513671875,
0.1644287109375
],
[
0.51318359375,
0.149658203125,
0.8818359375,
0.149658203125,
0.8818359375,
0.163330078125,
0.51318359375,
0.163330078125
],
[
0.12005615234375,
0.166259765625,
0.48779296875,
0.166259765625,
0.48779296875,
0.179931640625,
0.12005615234375,
0.179931640625
],
[
0.51318359375,
0.1644287109375,
0.8818359375,
0.1644287109375,
0.8818359375,
0.1776123046875,
0.51318359375,
0.1776123046875
],
[
0.12005615234375,
0.182861328125,
0.489501953125,
0.182861328125,
0.489501953125,
0.196533203125,
0.12005615234375,
0.196533203125
],
[
0.51123046875,
0.179931640625,
0.88330078125,
0.18115234375,
0.88330078125,
0.1953125,
0.51123046875,
0.19482421875
],
[
0.12005615234375,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.12005615234375,
0.2120361328125
],
[
0.5146484375,
0.19775390625,
0.8818359375,
0.19775390625,
0.8818359375,
0.21142578125,
0.5146484375,
0.21142578125
],
[
0.12005615234375,
0.21435546875,
0.48779296875,
0.21435546875,
0.48779296875,
0.22802734375,
0.12005615234375,
0.22802734375
],
[
0.51318359375,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51318359375,
0.22802734375
],
[
0.12005615234375,
0.23095703125,
0.490234375,
0.23095703125,
0.490234375,
0.24462890625,
0.12005615234375,
0.24462890625
],
[
0.51318359375,
0.2298583984375,
0.880859375,
0.2298583984375,
0.880859375,
0.243408203125,
0.51318359375,
0.243408203125
],
[
0.12005615234375,
0.2464599609375,
0.366943359375,
0.2464599609375,
0.366943359375,
0.260009765625,
0.12005615234375,
0.260009765625
],
[
0.51123046875,
0.24462890625,
0.8310546875,
0.2452392578125,
0.8310546875,
0.260009765625,
0.51123046875,
0.2587890625
],
[
0.52294921875,
0.268310546875,
0.880859375,
0.26953125,
0.880859375,
0.283935546875,
0.52294921875,
0.283203125
],
[
0.197265625,
0.276611328125,
0.411376953125,
0.276611328125,
0.411376953125,
0.290283203125,
0.197265625,
0.290283203125
],
[
0.4189453125,
0.27978515625,
0.453369140625,
0.27978515625,
0.453369140625,
0.2880859375,
0.4189453125,
0.2880859375
],
[
0.462646484375,
0.276611328125,
0.48779296875,
0.276611328125,
0.48779296875,
0.289306640625,
0.462646484375,
0.289306640625
],
[
0.5146484375,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.5146484375,
0.2998046875
],
[
0.12005615234375,
0.305908203125,
0.489501953125,
0.305908203125,
0.489501953125,
0.319580078125,
0.12005615234375,
0.319580078125
],
[
0.515625,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.515625,
0.31640625
],
[
0.12005615234375,
0.32177734375,
0.489501953125,
0.32177734375,
0.489501953125,
0.3349609375,
0.12005615234375,
0.3349609375
],
[
0.515625,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.33203125,
0.515625,
0.33203125
],
[
0.12005615234375,
0.33837890625,
0.26611328125,
0.33837890625,
0.26611328125,
0.3515625,
0.12005615234375,
0.3515625
],
[
0.515625,
0.334228515625,
0.8818359375,
0.334228515625,
0.8818359375,
0.347412109375,
0.515625,
0.347412109375
],
[
0.515625,
0.350830078125,
0.85986328125,
0.350830078125,
0.85986328125,
0.364013671875,
0.515625,
0.364013671875
],
[
0.162109375,
0.36767578125,
0.418212890625,
0.36767578125,
0.418212890625,
0.3818359375,
0.162109375,
0.3818359375
],
[
0.4189453125,
0.37060546875,
0.453369140625,
0.37060546875,
0.453369140625,
0.377685546875,
0.4189453125,
0.377685546875
],
[
0.44921875,
0.366455078125,
0.486083984375,
0.3681640625,
0.484375,
0.38134765625,
0.447509765625,
0.37890625
],
[
0.5205078125,
0.374755859375,
0.88330078125,
0.374755859375,
0.88330078125,
0.388427734375,
0.5205078125,
0.388427734375
],
[
0.1351318359375,
0.396728515625,
0.489501953125,
0.396728515625,
0.489501953125,
0.410400390625,
0.1351318359375,
0.410400390625
],
[
0.515625,
0.391357421875,
0.88330078125,
0.391357421875,
0.88330078125,
0.405029296875,
0.515625,
0.405029296875
],
[
0.1275634765625,
0.402099609375,
0.1368408203125,
0.402099609375,
0.1368408203125,
0.40625,
0.1275634765625,
0.40625
],
[
0.12005615234375,
0.412109375,
0.489501953125,
0.412109375,
0.489501953125,
0.42578125,
0.12005615234375,
0.42578125
],
[
0.5146484375,
0.40673828125,
0.8818359375,
0.40673828125,
0.8818359375,
0.42041015625,
0.5146484375,
0.42041015625
],
[
0.118408203125,
0.427001953125,
0.298828125,
0.427490234375,
0.298828125,
0.4423828125,
0.118408203125,
0.441162109375
],
[
0.5146484375,
0.421630859375,
0.88330078125,
0.4228515625,
0.88330078125,
0.43701171875,
0.5146484375,
0.435791015625
],
[
0.515625,
0.438232421875,
0.8818359375,
0.43701171875,
0.8818359375,
0.451904296875,
0.515625,
0.453125
],
[
0.1259765625,
0.453125,
0.489501953125,
0.453125,
0.489501953125,
0.466064453125,
0.1259765625,
0.466064453125
],
[
0.515625,
0.454833984375,
0.88330078125,
0.454833984375,
0.88330078125,
0.468505859375,
0.515625,
0.468505859375
],
[
0.12005615234375,
0.4697265625,
0.489501953125,
0.4697265625,
0.489501953125,
0.482666015625,
0.12005615234375,
0.482666015625
],
[
0.5146484375,
0.47021484375,
0.88330078125,
0.47021484375,
0.88330078125,
0.48388671875,
0.5146484375,
0.48388671875
],
[
0.12005615234375,
0.485107421875,
0.489501953125,
0.485107421875,
0.489501953125,
0.498779296875,
0.12005615234375,
0.498779296875
],
[
0.515625,
0.488037109375,
0.8818359375,
0.488037109375,
0.8818359375,
0.498779296875,
0.515625,
0.498779296875
],
[
0.12005615234375,
0.501953125,
0.48779296875,
0.501953125,
0.48779296875,
0.515625,
0.12005615234375,
0.515625
],
[
0.5146484375,
0.5029296875,
0.880859375,
0.5029296875,
0.880859375,
0.5166015625,
0.5146484375,
0.5166015625
],
[
0.12005615234375,
0.51708984375,
0.340087890625,
0.51708984375,
0.340087890625,
0.53076171875,
0.12005615234375,
0.53076171875
],
[
0.5146484375,
0.5185546875,
0.8818359375,
0.51953125,
0.8818359375,
0.533203125,
0.5146484375,
0.5322265625
],
[
0.51708984375,
0.5361328125,
0.8818359375,
0.5361328125,
0.8818359375,
0.5498046875,
0.51708984375,
0.5498046875
],
[
0.118408203125,
0.5458984375,
0.372802734375,
0.5458984375,
0.372802734375,
0.56005859375,
0.118408203125,
0.56005859375
],
[
0.515625,
0.5517578125,
0.8818359375,
0.5517578125,
0.8818359375,
0.5654296875,
0.515625,
0.5654296875
],
[
0.1175537109375,
0.568359375,
0.32666015625,
0.57080078125,
0.32666015625,
0.5849609375,
0.1175537109375,
0.5830078125
],
[
0.5146484375,
0.56640625,
0.68115234375,
0.568359375,
0.68115234375,
0.58203125,
0.5146484375,
0.580078125
],
[
0.12005615234375,
0.59130859375,
0.48779296875,
0.59130859375,
0.48779296875,
0.60498046875,
0.12005615234375,
0.60498046875
],
[
0.5224609375,
0.59033203125,
0.8818359375,
0.59130859375,
0.8818359375,
0.60546875,
0.5224609375,
0.60498046875
],
[
0.12005615234375,
0.60693359375,
0.489501953125,
0.60693359375,
0.489501953125,
0.62060546875,
0.12005615234375,
0.62060546875
],
[
0.51318359375,
0.60693359375,
0.8818359375,
0.60693359375,
0.8818359375,
0.62060546875,
0.51318359375,
0.62060546875
],
[
0.12176513671875,
0.62451171875,
0.48779296875,
0.62451171875,
0.48779296875,
0.63525390625,
0.12176513671875,
0.63525390625
],
[
0.515625,
0.6259765625,
0.8818359375,
0.6259765625,
0.8818359375,
0.6357421875,
0.515625,
0.6357421875
],
[
0.12005615234375,
0.6396484375,
0.489501953125,
0.6396484375,
0.489501953125,
0.65283203125,
0.12005615234375,
0.65283203125
],
[
0.51318359375,
0.64013671875,
0.8818359375,
0.64013671875,
0.8818359375,
0.65380859375,
0.51318359375,
0.65380859375
],
[
0.118408203125,
0.65478515625,
0.489501953125,
0.65380859375,
0.489501953125,
0.66845703125,
0.118408203125,
0.66943359375
],
[
0.51318359375,
0.65478515625,
0.8818359375,
0.65625,
0.8818359375,
0.67041015625,
0.51318359375,
0.66943359375
],
[
0.12005615234375,
0.67041015625,
0.489501953125,
0.66943359375,
0.489501953125,
0.68408203125,
0.12005615234375,
0.68505859375
],
[
0.515625,
0.6728515625,
0.88330078125,
0.6728515625,
0.88330078125,
0.68603515625,
0.515625,
0.68603515625
],
[
0.118408203125,
0.68603515625,
0.486083984375,
0.68603515625,
0.486083984375,
0.69970703125,
0.118408203125,
0.69970703125
],
[
0.51318359375,
0.68701171875,
0.880859375,
0.68701171875,
0.880859375,
0.70068359375,
0.51318359375,
0.70068359375
],
[
0.12005615234375,
0.70361328125,
0.48779296875,
0.70361328125,
0.48779296875,
0.71728515625,
0.12005615234375,
0.71728515625
],
[
0.51318359375,
0.703125,
0.8818359375,
0.70361328125,
0.8818359375,
0.71875,
0.51318359375,
0.71728515625
],
[
0.12005615234375,
0.7197265625,
0.48779296875,
0.7197265625,
0.48779296875,
0.73291015625,
0.12005615234375,
0.73291015625
],
[
0.515625,
0.72021484375,
0.8818359375,
0.72021484375,
0.8818359375,
0.73388671875,
0.515625,
0.73388671875
],
[
0.12005615234375,
0.736328125,
0.489501953125,
0.736328125,
0.489501953125,
0.75,
0.12005615234375,
0.75
],
[
0.515625,
0.736328125,
0.8818359375,
0.736328125,
0.8818359375,
0.75,
0.515625,
0.75
],
[
0.12005615234375,
0.751953125,
0.489501953125,
0.751953125,
0.489501953125,
0.765625,
0.12005615234375,
0.765625
],
[
0.5146484375,
0.7529296875,
0.8818359375,
0.7529296875,
0.8818359375,
0.7666015625,
0.5146484375,
0.7666015625
],
[
0.12005615234375,
0.7685546875,
0.48779296875,
0.7685546875,
0.48779296875,
0.7822265625,
0.12005615234375,
0.7822265625
],
[
0.51318359375,
0.76708984375,
0.880859375,
0.76708984375,
0.880859375,
0.78076171875,
0.51318359375,
0.78076171875
],
[
0.12176513671875,
0.78369140625,
0.48779296875,
0.78369140625,
0.48779296875,
0.79736328125,
0.12176513671875,
0.79736328125
],
[
0.5146484375,
0.78515625,
0.8818359375,
0.78515625,
0.8818359375,
0.798828125,
0.5146484375,
0.798828125
],
[
0.12176513671875,
0.8017578125,
0.484375,
0.8017578125,
0.484375,
0.8125,
0.12176513671875,
0.8125
],
[
0.5146484375,
0.8017578125,
0.85791015625,
0.8017578125,
0.85791015625,
0.8154296875,
0.5146484375,
0.8154296875
],
[
0.12176513671875,
0.81640625,
0.48779296875,
0.81640625,
0.48779296875,
0.830078125,
0.12176513671875,
0.830078125
],
[
0.118408203125,
0.83203125,
0.1427001953125,
0.83203125,
0.1427001953125,
0.8466796875,
0.118408203125,
0.8466796875
],
[
0.51318359375,
0.82666015625,
0.6337890625,
0.82666015625,
0.6337890625,
0.84130859375,
0.51318359375,
0.84130859375
],
[
0.5205078125,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.5205078125,
0.86083984375
],
[
0.118408203125,
0.85791015625,
0.2283935546875,
0.85791015625,
0.2283935546875,
0.87255859375,
0.118408203125,
0.87255859375
],
[
0.5146484375,
0.8623046875,
0.8818359375,
0.86328125,
0.8818359375,
0.87744140625,
0.5146484375,
0.876953125
],
[
0.12176513671875,
0.8798828125,
0.48779296875,
0.8798828125,
0.48779296875,
0.8935546875,
0.12176513671875,
0.8935546875
],
[
0.51318359375,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51123046875,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.51123046875,
0.9091796875
]
] |
[
"BERT输出执行最近邻匹配，但不使用GloVe",
"平均值和标准差。我们注意到，结果在ISI设",
"嵌入，因为它们对于所有感官都是相同的。我",
"",
"置中不能直接比较，因为根据它们的表述，它",
"们将此基线称为NearestNeighbor",
"们涉及不同的元测试事件",
"Non-episodictraining，它是一个单一的模型，",
"GloVe+GRU所有元学习方法的性能均优于",
"在所有任务上进行训练，没有任何区别-它合",
"其EF对应方法，表明从元训练集中成功学",
"并了支持和查询集，并使用迷你批处理进行训",
"习。当|S|=8.16.32时，FOMAML无法超越",
"练。因此，输出层与任务无关，输出单元的数量",
"NE-Baseline以及其他元学习方法的EF版本。",
"等于数据集中的感官总数。输出层的softmax",
"有趣的是，单独的元测试通常比NE-Baseline",
"仅在小批量中的相关类上获取。而不是每个任",
"模型更好，这表明后者不能有效地从元训练",
"务的Φi，我们现在有一个单一的Φ。在训练",
"集中转移知识。ProtoNet是性能最好的模型",
"期间，每个小批的参数更新为：",
"（IS=8除外），ProtoMAML紧随其后",
"ELMo+MLP与基于GloVe的模型相比",
"0, Φ ← Adam(LT,0,Φ,α)",
"",
"(9)",
"NearestNeighbor，NE-Baseline和EF方法的",
"其中α是学习率。在元测试阶段，我们在每个",
"得分更高，这可以归因于输人嵌入是上下文的。",
"任务的支持集上独立地微调训练模型（以情节",
"由于元训练，ProtoNet和ProtoFOMAML仍",
"方式），如下所示：",
"然比EF同行有所改善。ProtoFOMAML优于",
"所有SI的其他方法，FOMAML相对较弱",
"0,,Φ, 一 SGD(LT, 0, Φ,α, , m)",
"",
"(10)",
"BERT除NE-Baseline和FOMAML外，所",
"其中损失是在支持示例上计算的，Q是以前",
"有方法的分数都高于以前体系结构的分数。基",
"",
"的学习者学习率，是输出学习率。我们将此",
"于BERT的ProtoNet以及ProtoFOMAML",
"模型称为NE-Baseline",
"优于所有ISI的所有其他方法。此外，对于",
"[S|=4,8ProtoFOMAML优于ProtoNet，对",
"Episodicfine-tuningbaseline对于每种元学",
"于S=16,32，反之亦然。总体而言，在整个架",
"习方法，我们都包含一个变体，该变体仅从随",
"构中，我们看到NE-Baseline和FOMAML始",
"机初始化的模型开始执行元测试。这相当于从",
"终表现不佳，而ProtoNet和ProtoFOMAML",
"零开始对每集的支持示例进行培训。我们在前",
"是最有效的方法。此外，它们只需4个训练示",
"缀EF-前面加上前缀来表示",
"例即可实现较高的消歧性能，在许多情况下，",
"它们接近针对个人感官的一次性分类设置（请",
"5 Experiments and Results",
"参见附录A.1）。模型也相对稳定，如运行期",
"5.1Experimental setup",
"间的低标准偏差所示",
"我们使用元验证集为模型选择最佳超参数。所",
"Effect of second-ordergradients 我们进一",
"选的评估指标是元验证集中所有单词的宏F1",
"步实验了ProtoMAML，包括二阶梯度。在",
"分数的平均值。我们在元测试集上报告相同",
"表2中，我们将其F1分数与ProtoNet和",
"的度量。如果指标在两个时期内没有改善，我",
"ProtoFOMAML一起报告。对于BERT，我们",
"们通过终止训练来提前停止。GloVe+GRU中",
"训练ProtoMAML，同时仅对顶层进行微调，",
"隐藏状态的大小为256，对于GloVe+GRU、",
"并且由于其较高的计算成本，仅对一个内循环",
"ELMo+MLP和BERT，共享线性层的大小分",
"更新步骤进行微调。我们还训练了一个等效的",
"别为64、256和192。共享线性层的激活函数",
"ProtoFOMAML变体进行公平的比较。我们可",
"对于GloVe+GRU为tanh，对于ELMo+MLP",
"以观察到ProtoMAML在大多数情况下获得",
"和BERT为ReLU。对于FOMAML、Proto-",
"与ProtoFOMAML相似的分数，表明一阶近",
"FOMAML和ProtoMAML，批处理大小设置",
"似的有效性。在某些情况下，ProtoFOMAML",
"为16任务。这些输出层在每个情节中都会",
"的得分高于ProtoMAML，这可能是由于后者",
"重新初始化，而在NE-Baseline中，它具有",
"引起的过度拟合效应。根据这些结果，我们认",
"固定数量的5612单元。我们使用higher包",
"为一阶ProtoFOMAML足以完成这项任务",
"(Grefenstetteetal.2019）来实现MAML变",
"体",
"5.3Analysis",
"Effectofnumberof episodes我们首先调查",
"5.2 Results",
"使用更多的元训练集是否总是转化为更高的",
"在表1中，我们报告了元测试集中所有单词的",
"性能。随着元训练次数的增加，我们绘制了我",
"平均宏F1分数。我们报告了五次独立运行的",
"们的一个高得分模型-带有BERT的ProtoNet"
] |
{
"arxiv_id": "2004.14355",
"language": "zh",
"page_idx": 4
}
|
||
2303.14933-zh-1
|
2303.14933-zh-1.jpg
|
Table 1. Review of common VQA databases, where "UGC+Compression’ refers to manually encoding the UGC videos with
different compression settings.
<table>
<tbody>
<tr>
<th>Database</th>
<th>Year</th>
<th>$\operatorname{Duration}$ $\mathbf{n/s}$</th>
<th>Ref. $N$</th>
<th>$\operatorname{Scope}$</th>
<th>Subjective Evaluating Format</th>
</tr>
<tr>
<td>CVD2014 $\lceil31\rceil$</td>
<td>2014</td>
<td>10-25</td>
<td> </td>
<td>In- capture</td>
<td>In- lab</td>
</tr>
<tr>
<td>LIVE- Qualcomm [12]</td>
<td>2016</td>
<td>15</td>
<td> </td>
<td>In- capture</td>
<td>In- lab</td>
</tr>
<tr>
<td>$\boxed{14}$ KoNViD- 1k</td>
<td>$\overline{2017}$</td>
<td>$\overline{8}$</td>
<td> </td>
<td>In-the-wild</td>
<td>Crowdsourced</td>
</tr>
<tr>
<td>LIVE- VQC $\lceil36\rceil$</td>
<td>2018</td>
<td>10</td>
<td> </td>
<td>In- the- wild</td>
<td>Crowdsourced</td>
</tr>
<tr>
<td>YouTube- UGC [45]</td>
<td>2019</td>
<td>20</td>
<td> </td>
<td>In- the- wild</td>
<td>Crowdsourced</td>
</tr>
<tr>
<td>LSVQ [51]</td>
<td>2021</td>
<td>5-12</td>
<td> </td>
<td>In- the- wild</td>
<td>Crowdsourced</td>
</tr>
<tr>
<td>UGC- VIDEO $\boxed{21}$</td>
<td>$\overline{2019}$</td>
<td>$\overline{>10}$</td>
<td>$\overline{50}$</td>
<td>UGC+ - Compressior</td>
<td>In- lab</td>
</tr>
<tr>
<td>LIVE- WC $53^{\prime}$</td>
<td>2020</td>
<td>10</td>
<td>55</td>
<td>UGC+ .1 Compressior</td>
<td>In- lab</td>
</tr>
<tr>
<td>(Subset) [46] YT- UGC 1</td>
<td>2021</td>
<td>20</td>
<td>18!</td>
<td>$\mathop{\mathrm{UGC}}+$ - Compressior</td>
<td>In- lab</td>
</tr>
<tr>
<td>$[1^{\prime}$ ICME2021</td>
<td>2021</td>
<td> </td>
<td>1,00</td>
<td>$\mathop{\mathrm{UGC}}+$ - Compressior</td>
<td>In- lab</td>
</tr>
<tr>
<td>TaoLive( proposed) </td>
<td>2022</td>
<td>8</td>
<td>418</td>
<td>UGC+ 1 Compressior</td>
<td>In- lab</td>
</tr>
</tbody>
</table>
型提取语义特征；通过特定的手工图像失真描述符(i。e。模糊、噪声、块效应、曝光和色彩);并通过预训练的 3D-CNN 模型从视频剪辑中提取运动特征。与现有的 UGCVQA 算法相比，MD-VQA 从多个维度衡量视觉质量，这些维度对应着影响直播视频质量的关键因素，从而具有更好的可解释性和性能。本文的贡献总结如下：
<FigureHere>
Figure 2. Comparison of the quality distribution of reference videos between the TaoLive and ICME2021 [1] databases. The source UGC videos in the ICME2021 database are centered on high-quality levels while the quality levels of the source UGC videos in the TaoLive database are more diverse.
$\cdot$ We build a large-scale UGC Live VQA database
targeted at the compression artifacts on the UGC live videos. We collect 418 raw UGC live videos that are diverse in content, distortion, and quality. Then 8 encoding settings are used, which provides 3,762 compressed UGC live videos in total.
$\bullet$ We carry out a well-controlled in-lab subjective
experiment. 44 participants are invited to participate in the subjective experiment and a total of $\dot{1}65,528$ subjective annotations are gathered.
$\bullet$ A multi-dimensional NR-VQA model is proposed,
using pretrained 2D-CNN, handcrafted distortion descriptors, and pretrained 3D-CNN for the semantic, distortion, and motion features extraction respectively. The extracted features are then spatio-temporally fused to obtain the video-level quality score. The extensive experimental results validate the effectiveness of the proposed method.
机捕获的视频组成。然而，所提到的捕获内数据库的规模相对较小，所包含的失真要简单得多。后来，诸如KoNViD-1k$\left[14\right]$, YouTube-UGC$\left[45\right]$和 LSVQ$\left[51\right]$之类的 UGCVQA 数据库从在线平台上收集了野外 UGC 视频，这些视频具有更多样化的内容和失真，并大大促进了 UGCVQA 任务的开发
对于 UGC 实时视频，我们可以将其视为内捕获的UGC 视频 ,然后是压缩失真，其中野外失真和压缩失真都会对视觉质量产生重大影响。尽管 UGC-VIDEO[21] , LIVE-WC [53] 和 YT-UGC + [46] 等一些作品试图评估由常见压缩算法引起的 UGC 视频的质量，但这些数据库的相对较小的大小使得难以支持主流的数据驱动模型，例如。g。第 2.2 节中基于深度学习的模型。最近的 ICME2021 [1] 数据库规模很大。然而，如图所示。2, ICME2021 中的源 UGC 视频表现出高视觉质量，因此不能反映实时流媒体系统中的真实源视频质量分布。更重要的是，没有提到的数据库包括从实际的直播平台收集的源视频
# 2. Related Works
# $2. 1. \textbf{ VQA Databases}$
## 2.2. UGC VQA models
在过去的十年中，已经执行了许多 VQA 数据库[12,14,31,36,41,42,46,51,53]来解决 VQA 问题的挑战，并且在表 1 中展示了对通用 VQA 数据库的回顾。早期的 VQA 数据库[41,42]通常收集有限数量的源视频，并手动引人压缩和传输错误等失真来生成失真视频。这样的数据库在内容和扭曲方面不太多样化，不适合 UGC 视频的范围。然后，CVD2014[31] , LIVEQualcomm [12]和 LIVE-VQC [36]数据库由真实摄像
基于手工制作：基于手工制作的 VQA 模型提取质量感知特征以对空间和时间失真进行建模，例如自然场景统计 (NSS) 特征[23,29,33] ,工件[18,27,39] , 运动[10,18,39]等。例如，VIIDEO [29]给出了从自然
|
[
[
0.08087158203125,
0.0889892578125,
0.89208984375,
0.0889892578125,
0.89208984375,
0.1029052734375,
0.08087158203125,
0.1029052734375
],
[
0.080078125,
0.10418701171875,
0.279296875,
0.10418701171875,
0.279296875,
0.117431640625,
0.080078125,
0.117431640625
],
[
0.58837890625,
0.125,
0.6337890625,
0.126953125,
0.63232421875,
0.141357421875,
0.58642578125,
0.1395263671875
],
[
0.69287109375,
0.125,
0.89453125,
0.1256103515625,
0.89453125,
0.14013671875,
0.69287109375,
0.1395263671875
],
[
0.396240234375,
0.1256103515625,
0.471435546875,
0.1256103515625,
0.471435546875,
0.14013671875,
0.396240234375,
0.14013671875
],
[
0.26708984375,
0.126953125,
0.3056640625,
0.126953125,
0.3056640625,
0.14013671875,
0.26708984375,
0.14013671875
],
[
0.48779296875,
0.126953125,
0.525390625,
0.126953125,
0.525390625,
0.14013671875,
0.48779296875,
0.14013671875
],
[
0.3095703125,
0.1275634765625,
0.38818359375,
0.1275634765625,
0.38818359375,
0.1395263671875,
0.3095703125,
0.1395263671875
],
[
0.1396484375,
0.1275634765625,
0.2034912109375,
0.1275634765625,
0.2034912109375,
0.1395263671875,
0.1396484375,
0.1395263671875
],
[
0.12091064453125,
0.1427001953125,
0.21728515625,
0.14453125,
0.21728515625,
0.1590576171875,
0.12091064453125,
0.1572265625
],
[
0.330078125,
0.14453125,
0.370849609375,
0.14453125,
0.370849609375,
0.1590576171875,
0.330078125,
0.1590576171875
],
[
0.491943359375,
0.143310546875,
0.52294921875,
0.143310546875,
0.52294921875,
0.1590576171875,
0.491943359375,
0.1590576171875
],
[
0.771484375,
0.14453125,
0.81640625,
0.14453125,
0.81640625,
0.1590576171875,
0.771484375,
0.1590576171875
],
[
0.576171875,
0.146484375,
0.64794921875,
0.146484375,
0.64794921875,
0.1578369140625,
0.576171875,
0.1578369140625
],
[
0.265625,
0.1458740234375,
0.2998046875,
0.1458740234375,
0.2998046875,
0.1578369140625,
0.265625,
0.1578369140625
],
[
0.428955078125,
0.147705078125,
0.439453125,
0.147705078125,
0.439453125,
0.1717529296875,
0.428955078125,
0.1717529296875
],
[
0.09722900390625,
0.1558837890625,
0.2418212890625,
0.1572265625,
0.2418212890625,
0.1717529296875,
0.09722900390625,
0.17041015625
],
[
0.26708984375,
0.1578369140625,
0.2998046875,
0.1578369140625,
0.2998046875,
0.17041015625,
0.26708984375,
0.17041015625
],
[
0.339111328125,
0.1578369140625,
0.361083984375,
0.1578369140625,
0.361083984375,
0.1717529296875,
0.339111328125,
0.1717529296875
],
[
0.49267578125,
0.1578369140625,
0.51953125,
0.1578369140625,
0.51953125,
0.17041015625,
0.49267578125,
0.17041015625
],
[
0.576171875,
0.1590576171875,
0.646484375,
0.1590576171875,
0.646484375,
0.17041015625,
0.576171875,
0.17041015625
],
[
0.77294921875,
0.1590576171875,
0.814453125,
0.1590576171875,
0.814453125,
0.17041015625,
0.77294921875,
0.17041015625
],
[
0.11767578125,
0.1717529296875,
0.21728515625,
0.1717529296875,
0.21728515625,
0.18310546875,
0.11767578125,
0.18310546875
],
[
0.26708984375,
0.1717529296875,
0.2998046875,
0.1717529296875,
0.2998046875,
0.184326171875,
0.26708984375,
0.184326171875
],
[
0.340576171875,
0.1717529296875,
0.358642578125,
0.1717529296875,
0.358642578125,
0.1849365234375,
0.340576171875,
0.1849365234375
],
[
0.488525390625,
0.17041015625,
0.52392578125,
0.17041015625,
0.52392578125,
0.18310546875,
0.488525390625,
0.18310546875
],
[
0.57421875,
0.1717529296875,
0.64892578125,
0.1717529296875,
0.64892578125,
0.18310546875,
0.57421875,
0.18310546875
],
[
0.74658203125,
0.1717529296875,
0.84228515625,
0.1717529296875,
0.84228515625,
0.18310546875,
0.74658203125,
0.18310546875
],
[
0.490234375,
0.1787109375,
0.52392578125,
0.18310546875,
0.521484375,
0.1988525390625,
0.486083984375,
0.1944580078125
],
[
0.26708984375,
0.1817626953125,
0.2998046875,
0.184326171875,
0.298095703125,
0.1962890625,
0.265625,
0.1944580078125
],
[
0.340576171875,
0.18310546875,
0.361083984375,
0.18310546875,
0.361083984375,
0.1962890625,
0.340576171875,
0.1962890625
],
[
0.57421875,
0.18310546875,
0.64892578125,
0.18310546875,
0.64892578125,
0.1944580078125,
0.57421875,
0.1944580078125
],
[
0.1192626953125,
0.184326171875,
0.223876953125,
0.184326171875,
0.223876953125,
0.1962890625,
0.1192626953125,
0.1962890625
],
[
0.431396484375,
0.173583984375,
0.439453125,
0.173583984375,
0.439453125,
0.18310546875,
0.431396484375,
0.18310546875
],
[
0.4306640625,
0.186279296875,
0.438720703125,
0.186279296875,
0.438720703125,
0.1956787109375,
0.4306640625,
0.1956787109375
],
[
0.74658203125,
0.184326171875,
0.84228515625,
0.184326171875,
0.84228515625,
0.1956787109375,
0.74658203125,
0.1956787109375
],
[
0.26708984375,
0.1944580078125,
0.2998046875,
0.1962890625,
0.298095703125,
0.208984375,
0.265625,
0.20703125
],
[
0.486083984375,
0.1932373046875,
0.525390625,
0.1956787109375,
0.52392578125,
0.2108154296875,
0.484375,
0.208984375
],
[
0.339111328125,
0.1956787109375,
0.361083984375,
0.1956787109375,
0.361083984375,
0.208984375,
0.339111328125,
0.208984375
],
[
0.1070556640625,
0.1956787109375,
0.236083984375,
0.1956787109375,
0.236083984375,
0.208984375,
0.1070556640625,
0.208984375
],
[
0.57421875,
0.1962890625,
0.64892578125,
0.1962890625,
0.64892578125,
0.2083740234375,
0.57421875,
0.2083740234375
],
[
0.74658203125,
0.1962890625,
0.84228515625,
0.1962890625,
0.84228515625,
0.2083740234375,
0.74658203125,
0.2083740234375
],
[
0.334228515625,
0.2052001953125,
0.36767578125,
0.2083740234375,
0.365234375,
0.222900390625,
0.330810546875,
0.2197265625
],
[
0.428955078125,
0.1976318359375,
0.439453125,
0.1976318359375,
0.439453125,
0.2203369140625,
0.428955078125,
0.2203369140625
],
[
0.1356201171875,
0.2083740234375,
0.205078125,
0.208984375,
0.205078125,
0.2203369140625,
0.1356201171875,
0.2197265625
],
[
0.26708984375,
0.2083740234375,
0.298095703125,
0.2083740234375,
0.298095703125,
0.2203369140625,
0.26708984375,
0.2203369140625
],
[
0.48193359375,
0.20703125,
0.529296875,
0.20703125,
0.529296875,
0.2215576171875,
0.48193359375,
0.2215576171875
],
[
0.57421875,
0.208984375,
0.64892578125,
0.208984375,
0.64892578125,
0.2203369140625,
0.57421875,
0.2203369140625
],
[
0.74609375,
0.208984375,
0.8408203125,
0.208984375,
0.8408203125,
0.2203369140625,
0.74609375,
0.2203369140625
],
[
0.11273193359375,
0.2203369140625,
0.2303466796875,
0.2203369140625,
0.2303466796875,
0.2342529296875,
0.11273193359375,
0.2342529296875
],
[
0.426513671875,
0.2203369140625,
0.4453125,
0.2203369140625,
0.4453125,
0.2342529296875,
0.426513671875,
0.2342529296875
],
[
0.5390625,
0.2197265625,
0.68310546875,
0.2203369140625,
0.68310546875,
0.23486328125,
0.5390625,
0.2342529296875
],
[
0.3349609375,
0.2215576171875,
0.365234375,
0.2215576171875,
0.365234375,
0.2342529296875,
0.3349609375,
0.2342529296875
],
[
0.26708984375,
0.2215576171875,
0.2998046875,
0.2215576171875,
0.2998046875,
0.2342529296875,
0.26708984375,
0.2342529296875
],
[
0.491943359375,
0.2215576171875,
0.51953125,
0.2215576171875,
0.51953125,
0.2342529296875,
0.491943359375,
0.2342529296875
],
[
0.77294921875,
0.2215576171875,
0.814453125,
0.2215576171875,
0.814453125,
0.23291015625,
0.77294921875,
0.23291015625
],
[
0.265625,
0.2310791015625,
0.301513671875,
0.2342529296875,
0.2998046875,
0.248779296875,
0.26318359375,
0.246826171875
],
[
0.53759765625,
0.231689453125,
0.68310546875,
0.23291015625,
0.68310546875,
0.2474365234375,
0.53759765625,
0.246826171875
],
[
0.771484375,
0.23291015625,
0.81640625,
0.23291015625,
0.81640625,
0.2474365234375,
0.771484375,
0.2474365234375
],
[
0.339111328125,
0.2342529296875,
0.361083984375,
0.2342529296875,
0.361083984375,
0.2474365234375,
0.339111328125,
0.2474365234375
],
[
0.12335205078125,
0.2342529296875,
0.2197265625,
0.2342529296875,
0.2197265625,
0.24560546875,
0.12335205078125,
0.24560546875
],
[
0.4248046875,
0.2342529296875,
0.44677734375,
0.2342529296875,
0.44677734375,
0.246826171875,
0.4248046875,
0.246826171875
],
[
0.491943359375,
0.2342529296875,
0.51953125,
0.2342529296875,
0.51953125,
0.246826171875,
0.491943359375,
0.246826171875
],
[
0.490234375,
0.24560546875,
0.521484375,
0.24560546875,
0.521484375,
0.261474609375,
0.490234375,
0.261474609375
],
[
0.5390625,
0.24560546875,
0.68310546875,
0.2474365234375,
0.68310546875,
0.261474609375,
0.5390625,
0.2587890625
],
[
0.771484375,
0.24560546875,
0.81640625,
0.24560546875,
0.81640625,
0.260009765625,
0.771484375,
0.260009765625
],
[
0.339111328125,
0.246826171875,
0.361083984375,
0.246826171875,
0.361083984375,
0.260009765625,
0.339111328125,
0.260009765625
],
[
0.0882568359375,
0.246826171875,
0.256591796875,
0.246826171875,
0.256591796875,
0.260009765625,
0.0882568359375,
0.260009765625
],
[
0.265625,
0.2474365234375,
0.298095703125,
0.2474365234375,
0.298095703125,
0.260009765625,
0.265625,
0.260009765625
],
[
0.42236328125,
0.246826171875,
0.450927734375,
0.246826171875,
0.450927734375,
0.261474609375,
0.42236328125,
0.261474609375
],
[
0.1217041015625,
0.2568359375,
0.21728515625,
0.2587890625,
0.21728515625,
0.2734375,
0.1217041015625,
0.271484375
],
[
0.4150390625,
0.2587890625,
0.455810546875,
0.2587890625,
0.455810546875,
0.2734375,
0.4150390625,
0.2734375
],
[
0.5390625,
0.25830078125,
0.68310546875,
0.260009765625,
0.68310546875,
0.2734375,
0.5390625,
0.271484375
],
[
0.48779296875,
0.260009765625,
0.52392578125,
0.260009765625,
0.52392578125,
0.271484375,
0.48779296875,
0.271484375
],
[
0.265625,
0.260009765625,
0.2998046875,
0.260009765625,
0.2998046875,
0.272705078125,
0.265625,
0.272705078125
],
[
0.346435546875,
0.264404296875,
0.3505859375,
0.264404296875,
0.351318359375,
0.26953125,
0.34716796875,
0.26953125
],
[
0.77294921875,
0.260009765625,
0.814453125,
0.260009765625,
0.814453125,
0.271484375,
0.77294921875,
0.271484375
],
[
0.1070556640625,
0.26953125,
0.2344970703125,
0.270751953125,
0.2344970703125,
0.285400390625,
0.1070556640625,
0.2841796875
],
[
0.484375,
0.270751953125,
0.525390625,
0.270751953125,
0.525390625,
0.285400390625,
0.484375,
0.285400390625
],
[
0.5400390625,
0.270751953125,
0.68310546875,
0.270751953125,
0.68310546875,
0.2841796875,
0.5400390625,
0.2841796875
],
[
0.76953125,
0.270751953125,
0.81640625,
0.270751953125,
0.81640625,
0.285400390625,
0.76953125,
0.285400390625
],
[
0.343017578125,
0.271484375,
0.35693359375,
0.271484375,
0.35693359375,
0.28271484375,
0.343017578125,
0.28271484375
],
[
0.26708984375,
0.271484375,
0.301513671875,
0.271484375,
0.301513671875,
0.2841796875,
0.26708984375,
0.2841796875
],
[
0.420654296875,
0.271484375,
0.450927734375,
0.271484375,
0.450927734375,
0.2841796875,
0.420654296875,
0.2841796875
],
[
0.080078125,
0.2998046875,
0.468017578125,
0.300537109375,
0.468017578125,
0.31494140625,
0.080078125,
0.314453125
],
[
0.51953125,
0.298583984375,
0.52685546875,
0.3037109375,
0.51953125,
0.309326171875,
0.51318359375,
0.3037109375
],
[
0.7138671875,
0.311279296875,
0.72119140625,
0.31005859375,
0.72119140625,
0.313232421875,
0.7158203125,
0.314453125
],
[
0.74267578125,
0.30615234375,
0.78515625,
0.30615234375,
0.78515625,
0.313232421875,
0.74267578125,
0.313232421875
],
[
0.080078125,
0.31640625,
0.468017578125,
0.31640625,
0.468017578125,
0.330078125,
0.080078125,
0.330078125
],
[
0.5185546875,
0.317626953125,
0.525390625,
0.317626953125,
0.525390625,
0.323974609375,
0.5185546875,
0.323974609375
],
[
0.73046875,
0.314453125,
0.79150390625,
0.314453125,
0.79150390625,
0.3212890625,
0.73046875,
0.3212890625
],
[
0.69287109375,
0.32080078125,
0.72607421875,
0.317626953125,
0.73046875,
0.343505859375,
0.69677734375,
0.3466796875
],
[
0.708984375,
0.3212890625,
0.7255859375,
0.3212890625,
0.7255859375,
0.328857421875,
0.708984375,
0.328857421875
],
[
0.08087158203125,
0.33203125,
0.46728515625,
0.33203125,
0.46728515625,
0.34521484375,
0.08087158203125,
0.34521484375
],
[
0.5048828125,
0.333251953125,
0.529296875,
0.333251953125,
0.529296875,
0.354248046875,
0.5048828125,
0.354248046875
],
[
0.69775390625,
0.33837890625,
0.72802734375,
0.33837890625,
0.72802734375,
0.35595703125,
0.69775390625,
0.35595703125
],
[
0.08087158203125,
0.3466796875,
0.468017578125,
0.3466796875,
0.468017578125,
0.35986328125,
0.08087158203125,
0.35986328125
],
[
0.70751953125,
0.354736328125,
0.72607421875,
0.354736328125,
0.72607421875,
0.367431640625,
0.70751953125,
0.367431640625
],
[
0.08087158203125,
0.3623046875,
0.468017578125,
0.3623046875,
0.468017578125,
0.375732421875,
0.08087158203125,
0.375732421875
],
[
0.5810546875,
0.3662109375,
0.615234375,
0.3662109375,
0.615234375,
0.373779296875,
0.5810546875,
0.373779296875
],
[
0.650390625,
0.365478515625,
0.68896484375,
0.365478515625,
0.68896484375,
0.375732421875,
0.650390625,
0.375732421875
],
[
0.71142578125,
0.369384765625,
0.7255859375,
0.369384765625,
0.7255859375,
0.380126953125,
0.71142578125,
0.380126953125
],
[
0.65185546875,
0.37060546875,
0.689453125,
0.37255859375,
0.68896484375,
0.3837890625,
0.650390625,
0.3818359375
],
[
0.5185546875,
0.367431640625,
0.525390625,
0.367431640625,
0.525390625,
0.373779296875,
0.5185546875,
0.373779296875
],
[
0.08251953125,
0.37744140625,
0.468017578125,
0.37744140625,
0.468017578125,
0.391357421875,
0.08251953125,
0.391357421875
],
[
0.580078125,
0.371826171875,
0.61669921875,
0.371826171875,
0.61669921875,
0.38330078125,
0.580078125,
0.38330078125
],
[
0.55224609375,
0.37890625,
0.6357421875,
0.389404296875,
0.63232421875,
0.404052734375,
0.5498046875,
0.393310546875
],
[
0.63134765625,
0.3837890625,
0.685546875,
0.3837890625,
0.685546875,
0.395751953125,
0.63134765625,
0.395751953125
],
[
0.08251953125,
0.393310546875,
0.12091064453125,
0.393310546875,
0.12091064453125,
0.4072265625,
0.08251953125,
0.4072265625
],
[
0.7861328125,
0.389404296875,
0.82275390625,
0.389404296875,
0.82275390625,
0.40087890625,
0.7861328125,
0.40087890625
],
[
0.5888671875,
0.40087890625,
0.611328125,
0.40087890625,
0.611328125,
0.4140625,
0.5888671875,
0.4140625
],
[
0.78515625,
0.3984375,
0.80810546875,
0.401611328125,
0.8037109375,
0.416748046875,
0.78125,
0.41357421875
],
[
0.0931396484375,
0.41845703125,
0.468017578125,
0.41845703125,
0.468017578125,
0.431884765625,
0.0931396484375,
0.431884765625
],
[
0.50244140625,
0.4267578125,
0.890625,
0.4267578125,
0.890625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.111083984375,
0.434326171875,
0.468017578125,
0.431884765625,
0.468017578125,
0.44580078125,
0.111083984375,
0.447509765625
],
[
0.50244140625,
0.43994140625,
0.88818359375,
0.43994140625,
0.88818359375,
0.453857421875,
0.50244140625,
0.453857421875
],
[
0.111083984375,
0.447509765625,
0.4697265625,
0.448974609375,
0.4697265625,
0.462158203125,
0.111083984375,
0.4609375
],
[
0.50244140625,
0.455078125,
0.8896484375,
0.455078125,
0.8896484375,
0.468505859375,
0.50244140625,
0.468505859375
],
[
0.11273193359375,
0.464111328125,
0.468017578125,
0.464111328125,
0.468017578125,
0.477783203125,
0.11273193359375,
0.477783203125
],
[
0.5009765625,
0.468505859375,
0.8896484375,
0.468505859375,
0.8896484375,
0.481689453125,
0.5009765625,
0.481689453125
],
[
0.11273193359375,
0.477783203125,
0.471435546875,
0.479248046875,
0.471435546875,
0.49365234375,
0.11273193359375,
0.492431640625
],
[
0.50341796875,
0.484130859375,
0.890625,
0.484130859375,
0.890625,
0.494384765625,
0.50341796875,
0.494384765625
],
[
0.11273193359375,
0.49365234375,
0.420654296875,
0.49365234375,
0.420654296875,
0.5068359375,
0.11273193359375,
0.5068359375
],
[
0.50244140625,
0.496826171875,
0.61767578125,
0.496826171875,
0.61767578125,
0.51025390625,
0.50244140625,
0.51025390625
],
[
0.09478759765625,
0.51806640625,
0.4697265625,
0.51806640625,
0.4697265625,
0.5322265625,
0.09478759765625,
0.5322265625
],
[
0.111083984375,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.111083984375,
0.54736328125
],
[
0.5048828125,
0.52978515625,
0.890625,
0.52978515625,
0.890625,
0.54345703125,
0.5048828125,
0.54345703125
],
[
0.11358642578125,
0.55029296875,
0.4697265625,
0.55029296875,
0.4697265625,
0.56103515625,
0.11358642578125,
0.56103515625
],
[
0.5048828125,
0.54541015625,
0.890625,
0.54541015625,
0.890625,
0.5595703125,
0.5048828125,
0.5595703125
],
[
0.111083984375,
0.5625,
0.431396484375,
0.56298828125,
0.431396484375,
0.57763671875,
0.111083984375,
0.5771484375
],
[
0.50341796875,
0.5595703125,
0.8896484375,
0.5595703125,
0.8896484375,
0.57373046875,
0.50341796875,
0.57373046875
],
[
0.50341796875,
0.57470703125,
0.89208984375,
0.57470703125,
0.89208984375,
0.58837890625,
0.50341796875,
0.58837890625
],
[
0.09478759765625,
0.5869140625,
0.468017578125,
0.5869140625,
0.468017578125,
0.6015625,
0.09478759765625,
0.6015625
],
[
0.50341796875,
0.59033203125,
0.89208984375,
0.59033203125,
0.89208984375,
0.60400390625,
0.50341796875,
0.60400390625
],
[
0.1094970703125,
0.60302734375,
0.4697265625,
0.6015625,
0.4697265625,
0.6162109375,
0.1094970703125,
0.6171875
],
[
0.5009765625,
0.60400390625,
0.705078125,
0.60498046875,
0.705078125,
0.619140625,
0.5009765625,
0.61865234375
],
[
0.111083984375,
0.6171875,
0.465576171875,
0.6171875,
0.465576171875,
0.630859375,
0.111083984375,
0.630859375
],
[
0.52294921875,
0.62060546875,
0.89208984375,
0.62060546875,
0.89208984375,
0.63525390625,
0.52294921875,
0.63525390625
],
[
0.11273193359375,
0.6337890625,
0.46728515625,
0.6337890625,
0.46728515625,
0.64794921875,
0.11273193359375,
0.64794921875
],
[
0.50244140625,
0.63525390625,
0.890625,
0.63623046875,
0.890625,
0.65087890625,
0.50244140625,
0.6494140625
],
[
0.11273193359375,
0.6484375,
0.468017578125,
0.6484375,
0.468017578125,
0.662109375,
0.11273193359375,
0.662109375
],
[
0.50341796875,
0.6513671875,
0.8896484375,
0.6513671875,
0.8896484375,
0.66552734375,
0.50341796875,
0.66552734375
],
[
0.11273193359375,
0.6640625,
0.4697265625,
0.6640625,
0.4697265625,
0.67822265625,
0.11273193359375,
0.67822265625
],
[
0.51318359375,
0.66552734375,
0.88818359375,
0.66650390625,
0.88818359375,
0.68115234375,
0.51318359375,
0.6796875
],
[
0.111083984375,
0.6787109375,
0.471435546875,
0.6787109375,
0.471435546875,
0.693359375,
0.111083984375,
0.693359375
],
[
0.50341796875,
0.681640625,
0.8896484375,
0.681640625,
0.8896484375,
0.69580078125,
0.50341796875,
0.69580078125
],
[
0.11273193359375,
0.6943359375,
0.46728515625,
0.6943359375,
0.46728515625,
0.70849609375,
0.11273193359375,
0.70849609375
],
[
0.50244140625,
0.69580078125,
0.890625,
0.6962890625,
0.890625,
0.7109375,
0.50244140625,
0.71044921875
],
[
0.50341796875,
0.7119140625,
0.890625,
0.7119140625,
0.890625,
0.7265625,
0.50341796875,
0.7265625
],
[
0.50341796875,
0.7265625,
0.890625,
0.7265625,
0.890625,
0.7412109375,
0.50341796875,
0.7412109375
],
[
0.08087158203125,
0.734375,
0.111083984375,
0.734375,
0.111083984375,
0.7470703125,
0.08087158203125,
0.7470703125
],
[
0.10455322265625,
0.73291015625,
0.236083984375,
0.73291015625,
0.236083984375,
0.74755859375,
0.10455322265625,
0.74755859375
],
[
0.50244140625,
0.7421875,
0.890625,
0.7412109375,
0.890625,
0.755859375,
0.50244140625,
0.7568359375
],
[
0.07843017578125,
0.755859375,
0.24267578125,
0.7568359375,
0.24267578125,
0.771484375,
0.07843017578125,
0.77099609375
],
[
0.50341796875,
0.75830078125,
0.890625,
0.75830078125,
0.890625,
0.771484375,
0.50341796875,
0.771484375
],
[
0.5048828125,
0.7734375,
0.890625,
0.7734375,
0.890625,
0.787109375,
0.5048828125,
0.787109375
],
[
0.1004638671875,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.1004638671875,
0.79443359375
],
[
0.50341796875,
0.78857421875,
0.6689453125,
0.78857421875,
0.6689453125,
0.8017578125,
0.50341796875,
0.8017578125
],
[
0.08087158203125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.080078125,
0.8095703125,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.080078125,
0.82373046875
],
[
0.50341796875,
0.81640625,
0.685546875,
0.81640625,
0.685546875,
0.830078125,
0.50341796875,
0.830078125
],
[
0.08251953125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.08251953125,
0.8388671875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.52392578125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.52392578125,
0.85498046875
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.50341796875,
0.8720703125,
0.88623046875,
0.8720703125,
0.88623046875,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.08087158203125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88623046875,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"Table 1. Review of common VQA databases, where 'UGC+Compression' refers to manually encoding the UGC videos with",
"different compression settings.",
"Scope",
"Subjective Evaluating Format",
"Ref. Num.",
"Year",
"Scale",
"Duration/s",
"Database",
"CVD2014[31]",
"10-25",
"234",
" In-lab",
"In-capture",
"2014",
"1",
"LIVE-Qualcomm[12]",
"2016",
"15",
"208",
"In-capture",
"In-lab",
"KoNViD-1k 14",
"2017",
"8",
"1,200",
"In-the-wild",
"Crowdsourced",
"585",
"2018",
"10",
"In-the-wild",
"LIVE-VQC[36]",
"1",
"-",
"Crowdsourced",
"2019",
"1,500",
"20",
"YouTube-UGC[45]",
"In-the-wild",
"Crowdsourced",
"5-12",
"1",
"LSVQ[51]",
"2021",
"39,075",
"In-the-wild",
"Crowdsourced",
"UGC-VIDEO 21",
"50",
"UGC+Compression",
">10",
"2019",
"550",
"In-lab",
"2020",
"UGC + Compression",
"In-lab",
"10",
"LIVE-WC[53]",
"55",
"275",
"567",
"UGC+Compression",
" In-lab",
"20",
"YT-UGC + (Subset) [46]",
"2021",
"189",
"ICME2021[]",
"1,000",
"UGC + Compression",
"8.000",
"2021",
"1",
"In-lab",
"TaoLive(proposed)",
"3,762",
"UGC + Compression",
" In-lab",
"8",
"2022",
"418",
"型提取语义特征；通过特定的手工图像失真描述符（i。",
"ss",
"",
"TaoLive Ref",
"e。模糊、噪声、块效应、曝光和色彩)；并通过预训练",
".5",
"ICME2021Ref",
"",
"0.25",
"的3D-CNN模型从视频剪辑中提取运动特征。与现有",
"",
"1015",
"的UGCVQA算法相比，MD-VQA从多个维度衡量视",
"0.10",
"觉质量，这些维度对应着影响直播视频质量的关键因",
"Av:3.62",
"Avg4.47",
"0.05",
"Std:0.186",
".0",
"素，从而具有更好的可解释性和性能。本文的贡献总结",
"Std:0.369",
"TaoLive_ Rel Database",
"ICME2021_Ref",
"如下：",
"5MOs",
"(a)",
"(b)",
": We build a large-scale UGC Live VQA database",
"Figure 2. Comparison of the quality distribution of ref-",
"targeted at the compression artifacts on the UGC",
"erence videos between the TaoLive and ICME2021 [",
"livevideos.We collect 418 raw UGC live videos",
"databases. The source UGC videos in the ICME2021",
"that are diverse in content, distortion, and quality.",
"database are centered on high-quality levels while the qual-",
"Then 8 encoding settings are used, which provides",
"itylevelsofthesourceUGCvideosintheTaoLivedatabase",
"3,762 compressed UGC live videos in total.",
"are more diverse.",
". We carry out a well-controlled in-lab subjective",
"experiment. 44 participants are invited to partic-",
"机捕获的视频组成。然而，所提到的捕获内数据库的",
"ipate in the subjective experiment and a total of",
"规模相对较小，所包含的失真要简单得多。后来，诸如",
"165,528 subjective annotations are gathered.",
"KoNViD-1k[14], YouTube-UGC[45] 和 LSVQ[5]] 之",
"类的UGCVQA数据库从在线平台上收集了野外UGC",
": A multi-dimensional NR-VQA model is proposed,",
"视频，这些视频具有更多样化的内容和失真，并大大促",
"using pretrained 2D-CNN, handcrafted distortion",
"进了UGCVQA任务的开发",
"descriptors, and pretrained 3D-CNN for the se-",
"对于UGC实时视频，我们可以将其视为内捕获的",
"mantic, distortion, and motion features extrac-",
"UGC视频，然后是压缩失真，其中野外失真和压缩失真",
"tion respectively. The extracted features are then",
"都会对视觉质量产生重大影响。尽管UGC-VIDEO[21]",
"spatio-temporally fused to obtain the video-level",
"LIVE-WC[53】和YT-UGC+【46]等一些作品试图",
"quality score. The extensive experimental results",
"评估由常见压缩算法引起的UGC视频的质量，但这",
"validate the effectiveness of the proposed method.",
"些数据库的相对较小的大小使得难以支持主流的数据",
"驱动模型，例如。g。第2.2节中基于深度学习的模型。",
"最近的ICME2021「11数据库规模很大。然而，如图所",
"2.",
"Related Works",
"示。2，ICME2021中的源UGC视频表现出高视觉质",
"2.1. VQA Databases",
"量，因此不能反映实时流媒体系统中的真实源视频质",
"量分布。更重要的是，没有提到的数据库包括从实际的",
"在过去的十年中，已经执行了许多VQA数据库",
"直播平台收集的源视频",
"[12,14,31,36,41,42,46,51,53]来解决VQA问题的挑",
"战，并且在表1中展示了对通用VQA数据库的回顾。",
"2.2. UGC VQA models",
"早期的VQA数据库【41.421通常收集有限数量的源视",
"频，并手动引人压缩和传输错误等失真来生成失真视",
"基于手工制作：基于手工制作的VQA模型提取质",
"频。这样的数据库在内容和扭曲方面不太多样化，不",
"量感知特征以对空间和时间失真进行建模，例如自然",
"适合UGC视频的范围。然后，CVD2014[31]，LIVE-",
"场景统计（NSS）特征[23,29,33]，工件[18,27,39]",
"Qualcomm[12]和LIVE-VQC[36]数据库由真实摄像",
"运动[10,18,39]等。例如，VIIDEO[29]给出了从自然"
] |
{
"arxiv_id": "2303.14933",
"language": "zh",
"page_idx": 1
}
|
||
2303.14933-zh-2
|
2303.14933-zh-2.jpg
|
<FigureHere>
Figure 3. Sample frames of the videos in the proposed TaoLive Database, where the resolutions are marked on the top right. Additionally, the frame samples are cropped to delete some sensitive content such as human faces and watermarks for exhibition.
视频中收集的内在统计规律，并根据这些规律评估视频质量。V-BLIINDS [33] 通过使用时空 NSS 模型和运动表示模型来评估视频质量。TLVQM [18] 分两步计算低复杂度和高复杂度质量感知功能，以获得最终的视频质量。VIDEVAL [39] 仔细选择主流 NR-VQA 模型中具有代表性的质量感知功能，并将代表性功能回归为质量值
基于深度学习：考虑到深度神经网络(DNN)的巨大参数和 VQA 数据库相对较小的规模，一些 VQA 方法使用预训练的 DNN 模型进行特征提取。VSFA[20] 使用预训练的 DNN 模型提取深层语义特征，并使用门控循环单元 (GRU) 学习质量感知表示。为了增强对视频运动特征的理解，一些研究[19,37,51]进一步尝试使用在视频动作识别数据库上预训练的 3D-CNN 模型来提取运动特征，以帮助检测视频运动失真并产生良好的性能。之后，进行了一些基于 Transformer 的VQA 方法。LSCT [52] 首先提取帧感知质量特征，然后将特征输入到长短期卷积转换器中，以预测视频质量。FAST/FASTER-VQA [48,49] 提出了网格迷你补丁采样，并将视频采样结果形成为片段，然后将其放入片段修改后的视频 swin 转换器 [24] 中，以表示视频质量
3. UGC Live Dataset and Human Study
$3. 1. \textbf{ Videos Collection}$
如图所示。1 ,用户将其直播视频流上传到平台，平台将视频流以不同的比特率压缩，并根据互联网带宽或观看者的选择将其中之一分发给观看者。为了反映捕获的实时视频的真实质量分布，我们首先从中国非常受欢迎的实时平台 Taolive [2]平台收集大规模未压缩的原始 UGC 视频。然后，我们手动选择包含技术，
时尚，食品，日常生活，金融销售等场景的原始 UGC 视频，以确保内容多样性。最后，我们收集了418 个原始 UGC 视频 (110 个视频的分辨率为 720P,而 318个视频的分辨率为 1080P),每个原始 UGC 视频被裁剪为大约 8s 的持续时间作为源 UGC 实时视频
我们使用开源压缩工具 FFmpeg [3] 通过 H 的 8 个恒定速率因子 (CRF) 压缩源 UGC 实时视频。265 包括 16、20、24、28、32、36、40 和 44,以接近实时平台的分发过程。因此，生成了3，344=418个？?? 个压缩的 UGC 视频，总共收集了3，762=3,344+418个 UGC 视频进行评估，其样本如图。3
## 3.2. Human Study
人类研究是在控制良好的环境中进行的。44 人，包括 20 名男性和 24 名女性被邀请参加主观实验。观众坐在大约 1。5 倍的屏幕高度 (45 厘米)与正常的室内照明和视频播放的 iMac 显示器，支持分辨率高达4096???。在观众开始评估教资会直播视频之前，先做一个简短的介绍，让观众熟悉设备和质量评估任务。我们将实验分成 76 个小部分，每个部分包含 50 个 UGC 直播视频，没有内容重叠。观众单独参加所有会议，每次会议持续约 30 分钟。会议之间至少有 1 小时的休息时间，每个科目一天内不得参加超过 2 次会议。在会话期间，每个视频仅播放一次，观看者可以将视频质量从 1 到 5 进行评分，最小间隔为0。1。我们确保每个UGC 直播视频由 44 位受邀观众进行评估，并最终收集 165,528=3,762??? 个主观评分
## 3.3. Subjective Data Analysis
根据 ITU-RBT 的建议。500-13[6],我们计算 z 分
数作为 UGC 直播视频的质量标签：
(1)
$$z_{ij}=\frac{r_{ij}-\mu_i}{\sigma_i},$$
其中$r_{ij}$表示第$i$个主体在第$j$ 个 UGC 直 播 视 频上给出的质量评级$,\mu_i=\frac1{N_i}\sum_{j=1}^{N_i}r_{ij}~,~\sigma_i~=$ $\sqrt {\frac 1{N_i- 1}\sum _{j= 1}^{N_i}\left ( r_{ij}- \mu _i\right ) }$和 $N_i$ 是 由 主 体 $i$ 评 估 的 UGC 古 埒 词 医 的 数 目 处 厂 , 我 们 用 中 [ C] 中 将 节 的 产 品 折直播视频的数量。然后，我们根据[6]中推荐的主题拒绝程序，从不可靠的主题中删除质量标签。最后，将z 分数线性重新缩放为[1,5] ,并通过对重新缩放的z 分数进行平均来获得 UGC 视频$j$的平均意见得分(MOS):
(2)
$$MOS_j=\frac{1}{M}\sum_{i=1}^{M}z_{ij}^{'},$$
其中$MOS_j$表示第$j$个 UGC 视频的 MOS, $M$是
有效主题的数量，$z_ij^{\prime}$是重新缩放的 z 分数
我们进一步从 C$\overset{\mathrm{O}}{\operatorname*{RF}}$和分辨率角度绘制 MOS 分布。如图所示。4a ,保守的 CRF 参数选择 (16???) 对 UGC 实时视频引入了轻微的感知质量损失。当 CRF 从 28增加到 44 时，感知质量的下降趋势更加明显。此外，
|
[
[
0.50341796875,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.5048828125,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.5048828125,
0.121826171875
],
[
0.50341796875,
0.12249755859375,
0.89208984375,
0.12249755859375,
0.89208984375,
0.1363525390625,
0.50341796875,
0.1363525390625
],
[
0.5048828125,
0.136962890625,
0.89208984375,
0.136962890625,
0.89208984375,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.416748046875,
0.15283203125,
0.43310546875,
0.15283203125,
0.43310546875,
0.1610107421875,
0.416748046875,
0.1610107421875
],
[
0.50341796875,
0.15283203125,
0.8388671875,
0.15283203125,
0.8388671875,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.42724609375,
0.1717529296875,
0.455810546875,
0.1717529296875,
0.455810546875,
0.1787109375,
0.42724609375,
0.1787109375
],
[
0.52392578125,
0.1685791015625,
0.8896484375,
0.1685791015625,
0.8896484375,
0.1817626953125,
0.52392578125,
0.1817626953125
],
[
0.50341796875,
0.184326171875,
0.890625,
0.184326171875,
0.890625,
0.1976318359375,
0.50341796875,
0.1976318359375
],
[
0.23779296875,
0.173583984375,
0.25,
0.173583984375,
0.25,
0.177978515625,
0.23779296875,
0.177978515625
],
[
0.50244140625,
0.1988525390625,
0.88818359375,
0.1988525390625,
0.88818359375,
0.212158203125,
0.50244140625,
0.212158203125
],
[
0.414306640625,
0.2108154296875,
0.42724609375,
0.2108154296875,
0.42724609375,
0.2139892578125,
0.414306640625,
0.2139892578125
],
[
0.5048828125,
0.2139892578125,
0.8896484375,
0.2139892578125,
0.8896484375,
0.2279052734375,
0.5048828125,
0.2279052734375
],
[
0.50341796875,
0.2266845703125,
0.890625,
0.2279052734375,
0.890625,
0.242431640625,
0.50341796875,
0.2412109375
],
[
0.50341796875,
0.24365234375,
0.72607421875,
0.24365234375,
0.72607421875,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.5009765625,
0.267578125,
0.64794921875,
0.268310546875,
0.64794921875,
0.28271484375,
0.5009765625,
0.2822265625
],
[
0.52392578125,
0.292236328125,
0.89208984375,
0.292236328125,
0.89208984375,
0.30615234375,
0.52392578125,
0.30615234375
],
[
0.50634765625,
0.309326171875,
0.8896484375,
0.309326171875,
0.8896484375,
0.3193359375,
0.50634765625,
0.3193359375
],
[
0.5048828125,
0.322509765625,
0.890625,
0.322509765625,
0.890625,
0.3359375,
0.5048828125,
0.3359375
],
[
0.08087158203125,
0.33203125,
0.46728515625,
0.33203125,
0.46728515625,
0.34521484375,
0.08087158203125,
0.34521484375
],
[
0.5048828125,
0.337158203125,
0.890625,
0.337158203125,
0.890625,
0.35107421875,
0.5048828125,
0.35107421875
],
[
0.08251953125,
0.347900390625,
0.46728515625,
0.347900390625,
0.46728515625,
0.35791015625,
0.08251953125,
0.35791015625
],
[
0.50244140625,
0.35107421875,
0.890625,
0.3515625,
0.890625,
0.3662109375,
0.50244140625,
0.365478515625
],
[
0.08087158203125,
0.35986328125,
0.468017578125,
0.35986328125,
0.468017578125,
0.373779296875,
0.08087158203125,
0.373779296875
],
[
0.515625,
0.36865234375,
0.890625,
0.36865234375,
0.890625,
0.3818359375,
0.515625,
0.3818359375
],
[
0.08251953125,
0.375732421875,
0.464111328125,
0.375732421875,
0.464111328125,
0.386474609375,
0.08251953125,
0.386474609375
],
[
0.5048828125,
0.38330078125,
0.890625,
0.38330078125,
0.890625,
0.396484375,
0.5048828125,
0.396484375
],
[
0.08087158203125,
0.389404296875,
0.23779296875,
0.389404296875,
0.23779296875,
0.399658203125,
0.08087158203125,
0.399658203125
],
[
0.5048828125,
0.3984375,
0.890625,
0.3984375,
0.890625,
0.412353515625,
0.5048828125,
0.412353515625
],
[
0.08416748046875,
0.4140625,
0.46728515625,
0.4140625,
0.46728515625,
0.4248046875,
0.08416748046875,
0.4248046875
],
[
0.50341796875,
0.41357421875,
0.890625,
0.41357421875,
0.890625,
0.4267578125,
0.50341796875,
0.4267578125
],
[
0.08087158203125,
0.427978515625,
0.468017578125,
0.427978515625,
0.468017578125,
0.44140625,
0.08087158203125,
0.44140625
],
[
0.50341796875,
0.4287109375,
0.890625,
0.4287109375,
0.890625,
0.442626953125,
0.50341796875,
0.442626953125
],
[
0.08087158203125,
0.443115234375,
0.46728515625,
0.443115234375,
0.46728515625,
0.45703125,
0.08087158203125,
0.45703125
],
[
0.5048828125,
0.443115234375,
0.8896484375,
0.443115234375,
0.8896484375,
0.45703125,
0.5048828125,
0.45703125
],
[
0.08251953125,
0.458251953125,
0.468017578125,
0.458251953125,
0.468017578125,
0.4716796875,
0.08251953125,
0.4716796875
],
[
0.5048828125,
0.458251953125,
0.8896484375,
0.458251953125,
0.8896484375,
0.4716796875,
0.5048828125,
0.4716796875
],
[
0.08087158203125,
0.473388671875,
0.468017578125,
0.473388671875,
0.468017578125,
0.4873046875,
0.08087158203125,
0.4873046875
],
[
0.50244140625,
0.4716796875,
0.890625,
0.472900390625,
0.890625,
0.4873046875,
0.50244140625,
0.486083984375
],
[
0.08087158203125,
0.4873046875,
0.468017578125,
0.488037109375,
0.468017578125,
0.50244140625,
0.08087158203125,
0.501953125
],
[
0.50341796875,
0.4892578125,
0.73876953125,
0.4892578125,
0.73876953125,
0.50244140625,
0.50341796875,
0.50244140625
],
[
0.08251953125,
0.5048828125,
0.146240234375,
0.5048828125,
0.146240234375,
0.5166015625,
0.08251953125,
0.5166015625
],
[
0.1004638671875,
0.51953125,
0.468017578125,
0.51953125,
0.468017578125,
0.53271484375,
0.1004638671875,
0.53271484375
],
[
0.50341796875,
0.5146484375,
0.73193359375,
0.5146484375,
0.73193359375,
0.52783203125,
0.50341796875,
0.52783203125
],
[
0.08087158203125,
0.5341796875,
0.46728515625,
0.5341796875,
0.46728515625,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.52294921875,
0.5361328125,
0.8896484375,
0.537109375,
0.8896484375,
0.55029296875,
0.52294921875,
0.5498046875
],
[
0.08087158203125,
0.5498046875,
0.465576171875,
0.5498046875,
0.465576171875,
0.56298828125,
0.08087158203125,
0.56298828125
],
[
0.50341796875,
0.5517578125,
0.7548828125,
0.5517578125,
0.7548828125,
0.56494140625,
0.50341796875,
0.56494140625
],
[
0.08251953125,
0.564453125,
0.46728515625,
0.564453125,
0.46728515625,
0.57763671875,
0.08251953125,
0.57763671875
],
[
0.08087158203125,
0.57763671875,
0.468017578125,
0.5791015625,
0.468017578125,
0.59326171875,
0.08087158203125,
0.59228515625
],
[
0.6396484375,
0.58349609375,
0.74609375,
0.57275390625,
0.7490234375,
0.59326171875,
0.6435546875,
0.60400390625
],
[
0.869140625,
0.5869140625,
0.890625,
0.5869140625,
0.890625,
0.599609375,
0.869140625,
0.599609375
],
[
0.08087158203125,
0.5947265625,
0.46728515625,
0.5947265625,
0.46728515625,
0.60791015625,
0.08087158203125,
0.60791015625
],
[
0.7060546875,
0.59765625,
0.73046875,
0.59765625,
0.73046875,
0.60595703125,
0.7060546875,
0.60595703125
],
[
0.08251953125,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.08251953125,
0.62353515625
],
[
0.50341796875,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.6298828125,
0.50341796875,
0.6298828125
],
[
0.08251953125,
0.62353515625,
0.465576171875,
0.62353515625,
0.465576171875,
0.63720703125,
0.08251953125,
0.63720703125
],
[
0.5009765625,
0.630859375,
0.8896484375,
0.63330078125,
0.8896484375,
0.6494140625,
0.5009765625,
0.64794921875
],
[
0.76318359375,
0.630859375,
0.79150390625,
0.630859375,
0.79150390625,
0.63818359375,
0.76318359375,
0.63818359375
],
[
0.08087158203125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.08087158203125,
0.65380859375
],
[
0.08087158203125,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66845703125,
0.08087158203125,
0.66845703125
],
[
0.5009765625,
0.65283203125,
0.8935546875,
0.6484375,
0.8935546875,
0.66943359375,
0.5009765625,
0.673828125
],
[
0.08251953125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.08251953125,
0.6845703125
],
[
0.5048828125,
0.673828125,
0.890625,
0.673828125,
0.890625,
0.6875,
0.5048828125,
0.6875
],
[
0.08251953125,
0.68505859375,
0.468017578125,
0.68505859375,
0.468017578125,
0.69873046875,
0.08251953125,
0.69873046875
],
[
0.50244140625,
0.6875,
0.890625,
0.68798828125,
0.890625,
0.70263671875,
0.50244140625,
0.7021484375
],
[
0.08416748046875,
0.70068359375,
0.46728515625,
0.70068359375,
0.46728515625,
0.7138671875,
0.08416748046875,
0.7138671875
],
[
0.50244140625,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.08251953125,
0.71630859375,
0.468017578125,
0.71630859375,
0.468017578125,
0.72998046875,
0.08251953125,
0.72998046875
],
[
0.50244140625,
0.71826171875,
0.8896484375,
0.71826171875,
0.8896484375,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.08087158203125,
0.73095703125,
0.1004638671875,
0.73095703125,
0.1004638671875,
0.744140625,
0.08087158203125,
0.744140625
],
[
0.50244140625,
0.73486328125,
0.56201171875,
0.73486328125,
0.56201171875,
0.7470703125,
0.50244140625,
0.7470703125
],
[
0.07843017578125,
0.76123046875,
0.426513671875,
0.76220703125,
0.426513671875,
0.7763671875,
0.07843017578125,
0.77587890625
],
[
0.66845703125,
0.7568359375,
0.77783203125,
0.7568359375,
0.77783203125,
0.8037109375,
0.66845703125,
0.8037109375
],
[
0.71826171875,
0.76318359375,
0.73876953125,
0.76318359375,
0.73876953125,
0.771484375,
0.71826171875,
0.771484375
],
[
0.6259765625,
0.77294921875,
0.69140625,
0.77587890625,
0.689453125,
0.7890625,
0.625,
0.7861328125
],
[
0.869140625,
0.7744140625,
0.890625,
0.7744140625,
0.890625,
0.7890625,
0.869140625,
0.7890625
],
[
0.080078125,
0.787109375,
0.253173828125,
0.787109375,
0.253173828125,
0.80029296875,
0.080078125,
0.80029296875
],
[
0.7197265625,
0.79150390625,
0.74169921875,
0.79150390625,
0.74169921875,
0.798828125,
0.7197265625,
0.798828125
],
[
0.101318359375,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.101318359375,
0.82470703125
],
[
0.50634765625,
0.81494140625,
0.521484375,
0.81494140625,
0.521484375,
0.8193359375,
0.50634765625,
0.8193359375
],
[
0.51708984375,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82373046875,
0.51708984375,
0.82373046875
],
[
0.08087158203125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50341796875,
0.8271484375,
0.80810546875,
0.8271484375,
0.80810546875,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.08251953125,
0.84326171875,
0.46728515625,
0.84326171875,
0.46728515625,
0.85400390625,
0.08251953125,
0.85400390625
],
[
0.52392578125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.52392578125,
0.85498046875
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.5048828125,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"时尚，食品，日常生活，金融销售等场景的原始UGC",
"视频，以确保内容多样性。最后，我们收集了418个",
"原始UGC视频（110个视频的分辨率为720P，而318",
"个视频的分辨率为1080P），每个原始UGC视频被裁",
"",
"剪为大约8s的持续时间作为源UGC实时视频",
"1080P",
"我们使用开源压缩工具FFmpeg[3]通过H的8个",
"恒定速率因子（CRF）压缩源UGC实时视频。265包",
"621",
"括16、20、24、28、32、36、40和44，以接近实时平",
"",
"台的分发过程。因此，生成了3,344=418个??？个压缩",
"的UGC视频，总共收集了3,762=3,344+418个UGC",
"视频进行评估，其样本如图。3",
"3.2. Human Study",
"人类研究是在控制良好的环境中进行的。44人，包",
"括20名男性和24名女性被邀请参加主观实验。观众",
"坐在大约1。5倍的屏幕高度（45厘米）与正常的室",
"Figure 3. Sample frames of the videos in the proposed Tao-",
"内照明和视频播放的iMac显示器，支持分辨率高达",
"Live Database, where the resolutions are marked on the",
"4096??？。在观众开始评估教资会直播视频之前，先做",
"top right. Additionally, the frame samples are cropped to",
"个简短的介绍，让观众熟悉设备和质量评估任务。我",
"delete some sensitivecontent such as humanfacesand wa",
"们将实验分成76个小部分，每个部分包含50个UGC",
"termarks for exhibition.",
"直播视频，没有内容重叠。观众单独参加所有会议，每",
"视频中收集的内在统计规律，并根据这些规律评估视",
"次会议持续约30分钟。会议之间至少有1小时的休息",
"频质量。V-BLIINDS331通过使用时空NSS模型和运",
"时间，每个科目一天内不得参加超过2次会议。在会",
"动表示模型来评估视频质量。TLVQM[18]分两步计算",
"话期间，每个视频仅播放一次，观看者可以将视频质量",
"低复杂度和高复杂度质量感知功能，以获得最终的视",
"从1到5进行评分，最小间隔为0。1。我们确保每个",
"频质量。VIDEVAL39]仔细选择主流NR-VQA模型",
"UGC直播视频由44位受邀观众进行评估，并最终收",
"中具有代表性的质量感知功能，并将代表性功能回归",
"集165,528=3,762??？个主观评分",
"为质量值",
"基于深度学习：考虑到深度神经网络（DNN）的巨",
"3.3. Subjective Data Analysis",
"大参数和VQA数据库相对较小的规模，一些VQA方",
"根据ITU-RBT的建议。500-13[6]，我们计算z分",
"法使用预训练的DNN模型进行特征提取。VSFA[20",
"数作为UGC直播视频的质量标签：",
"使用预训练的DNN模型提取深层语义特征，并使用",
"门控循环单元（GRU）学习质量感知表示。为了增强",
"zi=-",
"(1)",
"对视频运动特征的理解，一些研究[19,37,51]进一步",
"Qi",
"尝试使用在视频动作识别数据库上预训练的3D-CNN",
"其中ri表示第i个主体在第j个UGC直播视",
"模型来提取运动特征，以帮助检测视频运动失真并产",
"频上给出的质量评级，==ri、=",
"JN",
"生良好的性能。之后，进行了一些基于Transformer的",
"VQA方法。LSCT52首先提取顿感知质量特征，然",
"（ri-a）和N；是由主体i评估的UGC",
"后将特征输入到长短期卷积转换器中，以预测视频质",
"直播视频的数量。然后，我们根据[6]中推荐的主题拒",
"量。FAST/FASTER-VQA「48，49】提出了网格迷你补",
"绝程序，从不可靠的主题中删除质量标签。最后，将",
"丁采样，并将视频采样结果形成为片段，然后将其放人",
"Z分数线性重新缩放为[1,5]，并通过对重新缩放的",
"片段修改后的视频swin转换器[24]中，以表示视频质",
"Z分数进行平均来获得UGC视频j的平均意见得分",
"量",
"(MOS):",
"3. UGC Live Dataset and Human Study",
"",
"M",
"MOS,=",
"(2)",
"3.1. Videos Collection",
"1-1",
"如图所示。1，用户将其直播视频流上传到平台，平",
"",
"其中MOS,表示第i个UGC视频的MOS，M是",
"台将视频流以不同的比特率压缩，并根据互联网带宽",
"有效主题的数量，z是重新缩放的z分数",
"或观看者的选择将其中之一分发给观看者。为了反映",
"我们进一步从CRF和分辨率角度绘制MOS分布。",
"捕获的实时视频的真实质量分布，我们首先从中国非",
"如图所示。4a，保守的CRF参数选择（16???）对UGC",
"常受欢迎的实时平台Taolive[2]平台收集大规模未压",
"实时视频引人了轻微的感知质量损失。当CRF从28",
"缩的原始UGC视频。然后，我们手动选择包含技术，",
"增加到44时，感知质量的下降趋势更加明显。此外，"
] |
{
"arxiv_id": "2303.14933",
"language": "zh",
"page_idx": 2
}
|
||
2303.14933-zh-4
|
2303.14933-zh-4.jpg
|
<FigureHere>
G Global Average Pooling $\textcircled{\mathrm{A}}$ Absolute Error Operation $\textcircled{\mathrm{M}}$ Multilayer Perceptron C Concatenation
Figure 5. The framework of the proposed method, where the semantic, distortion, and motion features are extracted by the pretrained EffcientNetV2 [38] , handcrafted distortion descriptors, and pretrained ResNet3D-18 [13] respectively. The absolute error between the adjacent frames’ semantic and distortion features is used to reflect the temporal quality fluctuations. Finally, the multi-dimensional features are spatio-temporally fused and regressed into quality values
语义和失真特征之间的绝对误差用于时间质量波动反
量：
映：
$$Q=1/\frac{n}{r}\sum_{1}^{\frac{n}{r}}Q^{i},$$
(9)
(5)
$$SF_{2k}^{i'}=|SF_{2k}^{i}-SF_{2k-1}^{i}|,k\in\{1,\cdots,L\},\\DF_{2k}^{i'}=|DF_{2k}^{i}-DF_{2k-1}^{i}|,k\in\{1,\cdots,L\},$$
其中$Q$是视频质量值，$\frac nr$表示剪辑数。我们只使用均
方误差(MSE)作为损失函数：
其中$SF_{2k}^{i^{\prime}}$和$DF_{2k}^{i^{\prime}}$表示相邻语义和失真特征之间的
绝对误差。然后可以得出时空融合为：
(10)
$$\begin{aligned}Loss=\frac{1}{n}\sum_{m=1}^{n}\left(Q_m'-Q_m\right)^2\end{aligned}$$
$$SD_{2k}^{i}=\omega(SF_{2k}^{i})\oplus\omega(DF_{2k}^{i})\oplus\omega(SF_{2k}^{i^{\prime}})\oplus\omega(DF_{2k}^{i^{\prime}}),\\STF^{i}=W_{L}^{1}(SD^{i^{T}}),$$
其中$n$表示小批量中的视频数量，$Q_m^\prime$和$Q_m$分别是
主观质量标签和预测质量级别
,其中$\oplus(\cdot)$表示时间上的学习特征$,\omega(\cdot)$表示可学习
### 5. Experiment
的多层感知器 ( MLP ), $SD_{2k}^{i}\in\mathbb{R}^{1\times N_{SD}}$表示从语义和失真特征获得的帧级空间融合特征，$SD^{i^T}\in\mathbb{R}^{N_{SD}\times L}$ 是 XMATHX 语义级的可转换结果。为了进一步引入质量感知运动特征，我们将时空特征与运动特征连接起来：
$$F^i=STF^{i^T}\oplus\omega(MF^i),$$
在本节中，我们首先给出实验设置的详细信息。然后在建议的 TaoLive 数据库和其他两个 UGC 压缩 VQA 模型上，将建议的 MD-VQA 模型与其他主流 VQA 模型进行验证。进行消融研究和跨数据库验证，以研究不同特征组的贡献和 VQA 模型的泛化能力。最后，我们在两个野外 UGCVQA 数据库上测试了提出的 MDVQA 模型。
(7)
其中$STF^{i^T}\in\mathbb{R}^{1\times N_{SD}}$,最终的剪辑级质量感知表示
$\begin{array}{ll}{{\overbrace{F^{i}\in\mathbb{R}^{1\times(N_{SD}+N_{M}^{\prime})}}}}&{{,}}&{{\mathrm{~ax}\text{公时分件放贝里心知权}}}\\{{\text{ 和 }N_{M}^{\prime}\text{ 在 MLP 操作后调整了运}}}\end{array}$
动特征的通道数
## 5.1. Benchmark Databases
# 4.3. Feature Regression
在上述特征提取过程之后，我们使用三阶段全连接
层将剪辑级质量感知表示$F^i$回归为质量值：
选择建议的 TaoLive 数据库和两个压缩的UGCVQA 数据库，包括 LIVE-WC [53]和 YT-UGC +[46]数据库作为基准数据库。对于所有数据库，我们遵循惯例，并以 80%-20% 训练测试比率拆分数据库。此外，所有数据库都单独验证。为了全面评估 VQA 模型的稳定性和性能，随机进行 30 次拆分，并将平均结果记录为最终性能
$$Q^i=FC(F^i),$$
(8)
其中$FC(\cdot)$表示全连接层，$Q_i$表示剪辑$C_i$的质量值。因此，可以通过平均池化获得整体 UGC 实时视频质
|
[
[
0.455810546875,
0.0889892578125,
0.5361328125,
0.0889892578125,
0.5361328125,
0.10101318359375,
0.455810546875,
0.10101318359375
],
[
0.45751953125,
0.10101318359375,
0.529296875,
0.10162353515625,
0.529296875,
0.11297607421875,
0.45751953125,
0.11236572265625
],
[
0.740234375,
0.09783935546875,
0.76708984375,
0.09783935546875,
0.76708984375,
0.1142578125,
0.740234375,
0.1142578125
],
[
0.1282958984375,
0.11553955078125,
0.148681640625,
0.11553955078125,
0.148681640625,
0.1307373046875,
0.1282958984375,
0.1307373046875
],
[
0.59716796875,
0.11614990234375,
0.63134765625,
0.11614990234375,
0.63134765625,
0.1319580078125,
0.59716796875,
0.1319580078125
],
[
0.74169921875,
0.126953125,
0.76318359375,
0.126953125,
0.76318359375,
0.141357421875,
0.74169921875,
0.141357421875
],
[
0.765625,
0.126953125,
0.775390625,
0.12249755859375,
0.779296875,
0.1275634765625,
0.76953125,
0.1331787109375
],
[
0.486083984375,
0.1337890625,
0.53125,
0.1337890625,
0.53125,
0.1483154296875,
0.486083984375,
0.1483154296875
],
[
0.5986328125,
0.1458740234375,
0.6435546875,
0.149658203125,
0.640625,
0.1641845703125,
0.59619140625,
0.160400390625
],
[
0.66650390625,
0.141357421875,
0.67333984375,
0.141357421875,
0.67333984375,
0.146484375,
0.66650390625,
0.146484375
],
[
0.449462890625,
0.15283203125,
0.55126953125,
0.15283203125,
0.55126953125,
0.1641845703125,
0.449462890625,
0.1641845703125
],
[
0.7060546875,
0.147705078125,
0.734375,
0.147705078125,
0.734375,
0.1622314453125,
0.7060546875,
0.1622314453125
],
[
0.12744140625,
0.1610107421875,
0.146240234375,
0.1610107421875,
0.146240234375,
0.1767578125,
0.12744140625,
0.1767578125
],
[
0.4345703125,
0.169189453125,
0.450927734375,
0.169189453125,
0.450927734375,
0.1767578125,
0.4345703125,
0.1767578125
],
[
0.494384765625,
0.169189453125,
0.5009765625,
0.17041015625,
0.498291015625,
0.1787109375,
0.491943359375,
0.177978515625
],
[
0.60693359375,
0.18310546875,
0.61767578125,
0.18310546875,
0.61767578125,
0.1976318359375,
0.60693359375,
0.1976318359375
],
[
0.6787109375,
0.177978515625,
0.701171875,
0.177978515625,
0.701171875,
0.1912841796875,
0.6787109375,
0.1912841796875
],
[
0.3896484375,
0.1849365234375,
0.396240234375,
0.1849365234375,
0.396240234375,
0.1900634765625,
0.3896484375,
0.1900634765625
],
[
0.4306640625,
0.184326171875,
0.451904296875,
0.184326171875,
0.451904296875,
0.1925048828125,
0.4306640625,
0.1925048828125
],
[
0.214111328125,
0.1944580078125,
0.223876953125,
0.1944580078125,
0.223876953125,
0.2083740234375,
0.214111328125,
0.2083740234375
],
[
0.49267578125,
0.189453125,
0.5,
0.189453125,
0.5,
0.1956787109375,
0.49267578125,
0.1956787109375
],
[
0.74609375,
0.1944580078125,
0.7861328125,
0.1962890625,
0.7861328125,
0.203857421875,
0.74609375,
0.2020263671875
],
[
0.78369140625,
0.1962890625,
0.82861328125,
0.1962890625,
0.82861328125,
0.2083740234375,
0.78369140625,
0.2083740234375
],
[
0.59326171875,
0.1976318359375,
0.62841796875,
0.2020263671875,
0.6259765625,
0.2139892578125,
0.5908203125,
0.210205078125
],
[
0.419189453125,
0.2008056640625,
0.464111328125,
0.2008056640625,
0.464111328125,
0.2052001953125,
0.419189453125,
0.2052001953125
],
[
0.68115234375,
0.20703125,
0.69677734375,
0.20703125,
0.69677734375,
0.2171630859375,
0.68115234375,
0.2171630859375
],
[
0.70361328125,
0.20703125,
0.71337890625,
0.20703125,
0.71337890625,
0.210205078125,
0.70361328125,
0.210205078125
],
[
0.765625,
0.20703125,
0.8056640625,
0.20703125,
0.8056640625,
0.2183837890625,
0.765625,
0.2183837890625
],
[
0.423095703125,
0.21533203125,
0.45751953125,
0.21533203125,
0.45751953125,
0.222900390625,
0.423095703125,
0.222900390625
],
[
0.61376953125,
0.216552734375,
0.619140625,
0.216552734375,
0.619140625,
0.2215576171875,
0.61376953125,
0.2215576171875
],
[
0.125,
0.2266845703125,
0.1502685546875,
0.2266845703125,
0.1502685546875,
0.24560546875,
0.125,
0.24560546875
],
[
0.428955078125,
0.2279052734375,
0.453369140625,
0.2279052734375,
0.453369140625,
0.2392578125,
0.428955078125,
0.2392578125
],
[
0.49267578125,
0.2259521484375,
0.50244140625,
0.2259521484375,
0.50244140625,
0.23291015625,
0.49267578125,
0.23291015625
],
[
0.6005859375,
0.2279052734375,
0.63818359375,
0.2310791015625,
0.63623046875,
0.24365234375,
0.5986328125,
0.240478515625
],
[
0.6640625,
0.2291259765625,
0.69140625,
0.2310791015625,
0.689453125,
0.242431640625,
0.6611328125,
0.2392578125
],
[
0.7900390625,
0.2298583984375,
0.82275390625,
0.2298583984375,
0.82275390625,
0.24560546875,
0.7900390625,
0.24560546875
],
[
0.26318359375,
0.2374267578125,
0.26708984375,
0.2374267578125,
0.26708984375,
0.25,
0.26318359375,
0.25
],
[
0.4453125,
0.242431640625,
0.5537109375,
0.242431640625,
0.5537109375,
0.255615234375,
0.4453125,
0.255615234375
],
[
0.7509765625,
0.251953125,
0.76953125,
0.251953125,
0.76953125,
0.26513671875,
0.7509765625,
0.26513671875
],
[
0.35693359375,
0.2587890625,
0.4345703125,
0.2587890625,
0.4345703125,
0.26953125,
0.35693359375,
0.26953125
],
[
0.68701171875,
0.281005859375,
0.705078125,
0.2841796875,
0.701171875,
0.2998046875,
0.68115234375,
0.296630859375
],
[
0.80224609375,
0.281005859375,
0.81201171875,
0.281005859375,
0.81201171875,
0.28857421875,
0.80224609375,
0.28857421875
],
[
0.73779296875,
0.300537109375,
0.8349609375,
0.300537109375,
0.8349609375,
0.31494140625,
0.73779296875,
0.31494140625
],
[
0.191162109375,
0.311279296875,
0.2418212890625,
0.313232421875,
0.240234375,
0.328857421875,
0.1895751953125,
0.326904296875
],
[
0.61669921875,
0.306884765625,
0.64453125,
0.306884765625,
0.64453125,
0.3212890625,
0.61669921875,
0.3212890625
],
[
0.459228515625,
0.3212890625,
0.5498046875,
0.3212890625,
0.5498046875,
0.33203125,
0.459228515625,
0.33203125
],
[
0.143798828125,
0.333984375,
0.65283203125,
0.333984375,
0.65283203125,
0.348388671875,
0.143798828125,
0.348388671875
],
[
0.6640625,
0.337158203125,
0.6845703125,
0.337158203125,
0.6845703125,
0.347900390625,
0.6640625,
0.347900390625
],
[
0.6787109375,
0.341552734375,
0.685546875,
0.3359375,
0.69140625,
0.34033203125,
0.6845703125,
0.34521484375
],
[
0.68701171875,
0.3359375,
0.77392578125,
0.3359375,
0.77392578125,
0.347900390625,
0.68701171875,
0.347900390625
],
[
0.080078125,
0.359130859375,
0.89208984375,
0.35986328125,
0.89208984375,
0.375,
0.080078125,
0.373779296875
],
[
0.080078125,
0.375,
0.8896484375,
0.375,
0.8896484375,
0.38818359375,
0.080078125,
0.38818359375
],
[
0.08087158203125,
0.38818359375,
0.890625,
0.38818359375,
0.890625,
0.401611328125,
0.08087158203125,
0.401611328125
],
[
0.080078125,
0.401611328125,
0.833984375,
0.401611328125,
0.833984375,
0.415283203125,
0.080078125,
0.415283203125
],
[
0.08087158203125,
0.43310546875,
0.468017578125,
0.43310546875,
0.468017578125,
0.4462890625,
0.08087158203125,
0.4462890625
],
[
0.50244140625,
0.43310546875,
0.52783203125,
0.43310546875,
0.52783203125,
0.4462890625,
0.50244140625,
0.4462890625
],
[
0.080078125,
0.447509765625,
0.10540771484375,
0.447509765625,
0.10540771484375,
0.4609375,
0.080078125,
0.4609375
],
[
0.71826171875,
0.443115234375,
0.73046875,
0.45068359375,
0.7197265625,
0.4609375,
0.70751953125,
0.455078125
],
[
0.689453125,
0.45263671875,
0.75634765625,
0.462158203125,
0.75244140625,
0.477783203125,
0.685546875,
0.468505859375
],
[
0.63818359375,
0.462158203125,
0.6748046875,
0.462158203125,
0.6748046875,
0.473388671875,
0.63818359375,
0.473388671875
],
[
0.86669921875,
0.460205078125,
0.8935546875,
0.460205078125,
0.8935546875,
0.4765625,
0.86669921875,
0.4765625
],
[
0.1282958984375,
0.472900390625,
0.416748046875,
0.472900390625,
0.416748046875,
0.4873046875,
0.1282958984375,
0.4873046875
],
[
0.44677734375,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.44677734375,
0.49755859375
],
[
0.1258544921875,
0.494384765625,
0.42236328125,
0.494384765625,
0.42236328125,
0.5087890625,
0.1258544921875,
0.5087890625
],
[
0.5048828125,
0.49560546875,
0.890625,
0.49560546875,
0.890625,
0.5087890625,
0.5048828125,
0.5087890625
],
[
0.5048828125,
0.51123046875,
0.7255859375,
0.51123046875,
0.7255859375,
0.5244140625,
0.5048828125,
0.5244140625
],
[
0.08251953125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.53515625,
0.08251953125,
0.53515625
],
[
0.08087158203125,
0.5361328125,
0.3505859375,
0.5361328125,
0.3505859375,
0.5498046875,
0.08087158203125,
0.5498046875
],
[
0.6611328125,
0.54150390625,
0.6748046875,
0.54150390625,
0.6748046875,
0.55029296875,
0.6611328125,
0.55029296875
],
[
0.6650390625,
0.54150390625,
0.7958984375,
0.54443359375,
0.7939453125,
0.5673828125,
0.6640625,
0.564453125
],
[
0.60107421875,
0.548828125,
0.650390625,
0.548828125,
0.650390625,
0.56005859375,
0.60107421875,
0.56005859375
],
[
0.85888671875,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56298828125,
0.85888671875,
0.56298828125
],
[
0.0882568359375,
0.5595703125,
0.45751953125,
0.55810546875,
0.45751953125,
0.57568359375,
0.0882568359375,
0.5771484375
],
[
0.65625,
0.55615234375,
0.70361328125,
0.564453125,
0.69921875,
0.580078125,
0.65185546875,
0.5712890625
],
[
0.2099609375,
0.58349609375,
0.3544921875,
0.58349609375,
0.3544921875,
0.59765625,
0.2099609375,
0.59765625
],
[
0.5048828125,
0.58837890625,
0.890625,
0.58837890625,
0.890625,
0.60302734375,
0.5048828125,
0.60302734375
],
[
0.44677734375,
0.599609375,
0.46728515625,
0.599609375,
0.46728515625,
0.61279296875,
0.44677734375,
0.61279296875
],
[
0.50341796875,
0.60400390625,
0.71728515625,
0.60400390625,
0.71728515625,
0.6171875,
0.50341796875,
0.6171875
],
[
0.096435546875,
0.6142578125,
0.46728515625,
0.6142578125,
0.46728515625,
0.62744140625,
0.096435546875,
0.62744140625
],
[
0.0767822265625,
0.62548828125,
0.471435546875,
0.62353515625,
0.471435546875,
0.64599609375,
0.0767822265625,
0.64794921875
],
[
0.50244140625,
0.63330078125,
0.63134765625,
0.63330078125,
0.63134765625,
0.646484375,
0.50244140625,
0.646484375
],
[
0.080078125,
0.64599609375,
0.468017578125,
0.642578125,
0.468017578125,
0.66015625,
0.080078125,
0.66357421875
],
[
0.521484375,
0.65576171875,
0.890625,
0.6572265625,
0.890625,
0.671875,
0.521484375,
0.67041015625
],
[
0.08251953125,
0.662109375,
0.465576171875,
0.662109375,
0.465576171875,
0.67529296875,
0.08251953125,
0.67529296875
],
[
0.08251953125,
0.67822265625,
0.468017578125,
0.67822265625,
0.468017578125,
0.69140625,
0.08251953125,
0.69140625
],
[
0.5048828125,
0.67236328125,
0.8896484375,
0.67236328125,
0.8896484375,
0.68603515625,
0.5048828125,
0.68603515625
],
[
0.50341796875,
0.68603515625,
0.89208984375,
0.6875,
0.89208984375,
0.7021484375,
0.50341796875,
0.70068359375
],
[
0.080078125,
0.6923828125,
0.12335205078125,
0.6923828125,
0.12335205078125,
0.70703125,
0.080078125,
0.70703125
],
[
0.1846923828125,
0.705078125,
0.36279296875,
0.70703125,
0.36279296875,
0.724609375,
0.1846923828125,
0.72265625
],
[
0.5048828125,
0.70263671875,
0.890625,
0.70263671875,
0.890625,
0.71630859375,
0.5048828125,
0.71630859375
],
[
0.44677734375,
0.71044921875,
0.468017578125,
0.71044921875,
0.468017578125,
0.7236328125,
0.44677734375,
0.7236328125
],
[
0.5048828125,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.5048828125,
0.73193359375
],
[
0.0767822265625,
0.72607421875,
0.471435546875,
0.72802734375,
0.471435546875,
0.74755859375,
0.0767822265625,
0.74560546875
],
[
0.5048828125,
0.73291015625,
0.8896484375,
0.73291015625,
0.8896484375,
0.74755859375,
0.5048828125,
0.74755859375
],
[
0.07843017578125,
0.74072265625,
0.471435546875,
0.744140625,
0.471435546875,
0.76513671875,
0.07843017578125,
0.76123046875
],
[
0.5048828125,
0.74853515625,
0.5849609375,
0.74853515625,
0.5849609375,
0.76025390625,
0.5048828125,
0.76025390625
],
[
0.08087158203125,
0.76220703125,
0.1968994140625,
0.76220703125,
0.1968994140625,
0.77587890625,
0.08087158203125,
0.77587890625
],
[
0.50244140625,
0.771484375,
0.7158203125,
0.77294921875,
0.7158203125,
0.787109375,
0.50244140625,
0.7861328125
],
[
0.08087158203125,
0.7841796875,
0.265625,
0.7841796875,
0.265625,
0.79736328125,
0.08087158203125,
0.79736328125
],
[
0.52294921875,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.52294921875,
0.81005859375
],
[
0.1004638671875,
0.80810546875,
0.468017578125,
0.80810546875,
0.468017578125,
0.8212890625,
0.1004638671875,
0.8212890625
],
[
0.50244140625,
0.8095703125,
0.890625,
0.81005859375,
0.890625,
0.82470703125,
0.50244140625,
0.82373046875
],
[
0.08087158203125,
0.82275390625,
0.39453125,
0.82275390625,
0.39453125,
0.8369140625,
0.08087158203125,
0.8369140625
],
[
0.50341796875,
0.82470703125,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.50341796875,
0.8388671875
],
[
0.22216796875,
0.8466796875,
0.325927734375,
0.8466796875,
0.325927734375,
0.861328125,
0.22216796875,
0.861328125
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.44677734375,
0.84765625,
0.4697265625,
0.84765625,
0.4697265625,
0.861328125,
0.44677734375,
0.861328125
],
[
0.50341796875,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.08087158203125,
0.869140625,
0.468017578125,
0.87060546875,
0.468017578125,
0.88525390625,
0.08087158203125,
0.8837890625
],
[
0.5048828125,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.08087158203125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88525390625,
0.63623046875,
0.88623046875,
0.63623046875,
0.90087890625,
0.50341796875,
0.8994140625
]
] |
[
"EfficientNetV2",
"Last4Stages",
"?",
"Fi",
"SF2k",
"?",
"O",
"@&C",
"SF2k-1",
"A",
"Semantic Features",
"SF2k",
"F2",
"Blur",
"",
"",
"?",
"国",
"Noise",
"..",
"日",
"Snatio-t",
"temporal",
"DF2K",
"BlockEfect",
"M",
"-",
"Fusion",
"Exposure",
"",
"F2L",
"Color",
"H",
"DF2k-",
"DF2K",
"STFi",
"",
"Distortion Features",
"?",
"ResNet3D-18",
"Q'",
"F",
"QualityRegression",
"Clip Ci",
"MFi",
"Motion Features",
"@ Global Average Pooling @ Absolute Error Operation ? Multilayer Perceptron",
"O",
"",
"Concatenation",
"Figure 5.The framework of the proposed method, where the semantic, distortion, and motion features are extracted",
"by the pretrained EfficientNetV2 [38] , handcrafted distortion descriptors, and pretrained ResNet3D-18 [13] respectively.",
"The absolute error between the adjacent frames' semantic and distortion features is used to reflect the temporal quality",
"fluctuations. Finally, the multi-dimensional features are spatio-temporally fused and regressed into quality values.",
"语义和失真特征之间的绝对误差用于时间质量波动反",
"量：",
"映：",
"A",
"nQ",
"Q=",
"(9)",
"SF2h=|SF2k-SF2h-1l,k E[1,,L]",
"(5)",
"DF2k = [DF2k -DF2k-1l,k E [1,..,L]",
"其中Q是视频质量值，表示剪辑数。我们只使用均",
"方误差（MSE）作为损失函数：",
"其中SF2k和DF2k表示相邻语义和失真特征之间的",
"绝对误差。然后可以得出时空融合为：",
"1",
"Z(Qm-Qm)",
"Loss-",
"(10)",
"SD2h = w(SF2h) 田w(DF2k) ①w(SF2k)Dw(DF2k),",
"nm=1",
"STF = W(SDi),",
"其中n表示小批量中的视频数量，Qm和Qm分别是",
"(6)",
"主观质量标签和预测质量级别",
"其中④(）表示时间上的学习特征，w(）表示可学习",
"的多层感知器（MLP)，SDERIXNsD表示从语义和",
"5.Experiment",
"失真特征获得的顿级空间融合特征，SDiTEIRNsDxL",
"在本节中，我们首先给出实验设置的详细信息。然后",
"是XMATHX语义级的可转换结果。。为了进一步引人",
"质量感知运动特征，我们将时空特征与运动特征连接",
"在建议的TaoLive数据库和其他两个UGC压缩VQA",
"模型上，将建议的MD-VQA模型与其他主流VQA模",
"起来：",
"F\" = STF @w(MF\"),",
"型进行验证。进行消融研究和跨数据库验证，以研究",
"(7)",
"不同特征组的贡献和VQA模型的泛化能力。最后，我",
"其中STFitERIxNsD，最终的剪辑级质量感知表示",
"们在两个野外UGCVQA数据库上测试了提出的MD-",
"FiERIx(NsD+NM）和NM在MLP操作后调整了运",
"VQA模型。",
"动特征的通道数",
"5.1.Benchmark Databases",
"4.3.FeatureRegression",
"选择建议的TaoLive数据库和两个压缩的",
"在上述特征提取过程之后，我们使用三阶段全连接",
"UGCVQA数据库，包括LIVE-WC[53】和YT-UGC",
"层将剪辑级质量感知表示Fi回归为质量值：",
"+「46]数据库作为基准数据库。对于所有数据库，我们",
"Q= FC(F),",
"遵循惯例，并以80%-20%训练测试比率拆分数据库。",
"(8)",
"此外，所有数据库都单独验证。为了全面评估VQA模",
"其中FC（）表示全连接层，Qi表示剪辑C的质量值。",
"型的稳定性和性能，随机进行30次拆分，并将平均结",
"因此，可以通过平均池化获得整体UGC实时视频质",
"果记录为最终性能"
] |
{
"arxiv_id": "2303.14933",
"language": "zh",
"page_idx": 4
}
|
||
2004.14837-zh-0
|
2004.14837-zh-0.jpg
|
# 来自神经机器翻译的精确单词对齐诱导
Yun Chen$^*1$, Yang Liu$^2$,Guanhua Chen$^3$, Xin Jiang$^4$,Qun Liu$^4$ $^{1}$Shanghai University of Finance and Economics, Shanghai, China
$^{2}$Tsinghua University, Beijing, China
$^{3}$The University of Hong Kong, Hong Kong, China
$^{4}$Huawei Noah's Ark Lab, Hong Kong, China
yunchen@sufe.edu.cn, liuyang2011@tsinghua.edu.cn,
$ghchen@ee.hku.hk,\{~jiang.xin,~qun.liu~\}~@huawei.com$
## Abstract
Despite its original goal to jointly learn to align and translate, prior researches suggest that Transformer captures poor word alignments through its attention mechanism.
In this paper, we show that
attention weights DO capture accurate word alignments and propose two novel word alignment induction methods ShiftAtt and Shift-AET .The main idea is to induce alignments at the step when the tobe-aligned target token is the decoder input rather than the decoder output as in previous work. Shift-Att is an interpretation method that induces alignments from the attention weights of Transformer and does not require parameter update or architecture change. Shift-AET extracts alignments from an additional alignment module which is tightly integrated into Transformer and trained in isolation with supervision from symmetrized Shift-Att alignments. Experiments on three publicly available datasets demonstrate that both methods perform better than their corresponding neural baselines and Shift-AET significantly outperforms GIZA++ by 1.4- $4.\bar{8}$ AER points.$^1$
## 1 Introduction
单词对齐的任务是从并行语料库中找到词典翻译等效项E(Brown et al.,1993)。它是自然语言处理(NLP)的基本任务之一，由社区(Dyer et al., 2013; Brown et al., 1993; Liu and Sun, 2015) 广泛研究。单词对齐在许多情况下都很有用，例如错误分析$\overline{\mathrm{E}}($Ding et al.,2017; Li et al., 2019) ,引入覆盖率和生育力模型E (Tu et al., 2016) , 在交互式机器翻译$\underline{\mathrm{E}}($Hasler et al., 2018; Chen et al., 2020) 中插人外部约束
$^{*}$Corresponding author. Part of the work was done
when Yun was in Huawei Noah, s Ark Lab.
$^{1}$Code canbe found at https://github.com/
sufe- nlp/transformer-alignment .
<FigureHere>
Figure 1: An example to compare our method Shift-Att and the baseline Naive-Att. The left is an attention map from the third decoder layer of the vanilla Transformer and the right are the induced alignments. Shift-Att induces alignments for target word $y_i$ at decoding step $i+1$ when $y_i$ is the decoder input, while Naive-Att at step $i$ when $y_i$ is the decoder output.
以及在计算机辅助翻译E(Dagan et al., 1993)
中为人类翻译人员提供指导
单词对齐是统计机器翻译$\operatorname{E}($Koehn et al. 2003, SMT)中管道的一部分，但对于神经机器翻译$\operatorname{E}($Bahdanau et al., 2015, NMT) 不一定需要。NMT 中的注意机制在功能上没有发挥源与目标之间的单词对齐的作用，至少与 SMT 中的模拟方式不同。很难解释注意力激活和提取有意义的单词对齐，尤其是从Transformer (Garg et al.,2019) 。结果 ,使用最广泛的单词对齐工具仍然是外部统计模型，例如 Fast-Align (Dyer et al., 2013) 和 GIZA++ (Brown et al., 1993; Och and Ney, 2003)
最近，社区中重新出现了研究 Transformer$\allowbreak ( {\mathrm{mod}} $Ding et al. , 2019; Li et al. , 2019) 的 ) 单词对齐方式的兴趣。一个简单的解决方案
|
[
[
0.297119140625,
0.0843505859375,
0.705078125,
0.0843505859375,
0.705078125,
0.1015625,
0.297119140625,
0.1015625
],
[
0.235107421875,
0.119384765625,
0.78271484375,
0.119384765625,
0.78271484375,
0.1341552734375,
0.235107421875,
0.1341552734375
],
[
0.2166748046875,
0.1341552734375,
0.791015625,
0.1353759765625,
0.791015625,
0.152587890625,
0.2166748046875,
0.1519775390625
],
[
0.341796875,
0.149658203125,
0.66748046875,
0.152587890625,
0.66748046875,
0.171630859375,
0.341796875,
0.168701171875
],
[
0.2822265625,
0.166259765625,
0.724609375,
0.168701171875,
0.724609375,
0.18701171875,
0.2822265625,
0.185302734375
],
[
0.306396484375,
0.18408203125,
0.70361328125,
0.185302734375,
0.70361328125,
0.203125,
0.306396484375,
0.201904296875
],
[
0.276123046875,
0.2037353515625,
0.72705078125,
0.2037353515625,
0.72705078125,
0.21728515625,
0.276123046875,
0.21728515625
],
[
0.264404296875,
0.2203369140625,
0.74462890625,
0.2203369140625,
0.74462890625,
0.235107421875,
0.264404296875,
0.235107421875
],
[
0.2626953125,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.28076171875,
0.2626953125,
0.279052734375
],
[
0.14697265625,
0.290283203125,
0.4609375,
0.290283203125,
0.4609375,
0.303955078125,
0.14697265625,
0.303955078125
],
[
0.8212890625,
0.298583984375,
0.8447265625,
0.298583984375,
0.8447265625,
0.3017578125,
0.8212890625,
0.3017578125
],
[
0.14697265625,
0.3046875,
0.4609375,
0.3046875,
0.4609375,
0.318359375,
0.14697265625,
0.318359375
],
[
0.5224609375,
0.31005859375,
0.5517578125,
0.312255859375,
0.54833984375,
0.32666015625,
0.51904296875,
0.32373046875
],
[
0.145263671875,
0.318359375,
0.4609375,
0.317626953125,
0.4609375,
0.33203125,
0.145263671875,
0.333251953125
],
[
0.7421875,
0.3134765625,
0.76904296875,
0.3134765625,
0.76904296875,
0.325927734375,
0.7421875,
0.325927734375
],
[
0.515625,
0.327880859375,
0.55322265625,
0.327880859375,
0.55322265625,
0.34033203125,
0.515625,
0.34033203125
],
[
0.66748046875,
0.325927734375,
0.72900390625,
0.325927734375,
0.72900390625,
0.336181640625,
0.66748046875,
0.336181640625
],
[
0.14697265625,
0.333251953125,
0.4609375,
0.333251953125,
0.4609375,
0.3466796875,
0.14697265625,
0.3466796875
],
[
0.72900390625,
0.330078125,
0.771484375,
0.330078125,
0.771484375,
0.34033203125,
0.72900390625,
0.34033203125
],
[
0.52490234375,
0.342529296875,
0.5498046875,
0.342529296875,
0.5498046875,
0.355712890625,
0.52490234375,
0.355712890625
],
[
0.74658203125,
0.345703125,
0.76904296875,
0.345703125,
0.76904296875,
0.355712890625,
0.74658203125,
0.355712890625
],
[
0.2166748046875,
0.3466796875,
0.4609375,
0.3466796875,
0.4609375,
0.361083984375,
0.2166748046875,
0.361083984375
],
[
0.14697265625,
0.3486328125,
0.2166748046875,
0.3486328125,
0.2166748046875,
0.359375,
0.14697265625,
0.359375
],
[
0.14697265625,
0.35986328125,
0.459228515625,
0.35986328125,
0.459228515625,
0.37353515625,
0.14697265625,
0.37353515625
],
[
0.14697265625,
0.3759765625,
0.4609375,
0.3759765625,
0.4609375,
0.3896484375,
0.14697265625,
0.3896484375
],
[
0.14697265625,
0.3896484375,
0.4609375,
0.3896484375,
0.4609375,
0.402587890625,
0.14697265625,
0.402587890625
],
[
0.14697265625,
0.40380859375,
0.4609375,
0.40380859375,
0.4609375,
0.41748046875,
0.14697265625,
0.41748046875
],
[
0.736328125,
0.4091796875,
0.767578125,
0.4091796875,
0.767578125,
0.42041015625,
0.736328125,
0.42041015625
],
[
0.14697265625,
0.418701171875,
0.4609375,
0.418701171875,
0.4609375,
0.431640625,
0.14697265625,
0.431640625
],
[
0.53076171875,
0.42578125,
0.6533203125,
0.42578125,
0.6533203125,
0.439453125,
0.53076171875,
0.439453125
],
[
0.74462890625,
0.42578125,
0.76416015625,
0.42578125,
0.76416015625,
0.435302734375,
0.74462890625,
0.435302734375
],
[
0.5322265625,
0.43701171875,
0.72705078125,
0.43701171875,
0.72705078125,
0.45068359375,
0.5322265625,
0.45068359375
],
[
0.145263671875,
0.43115234375,
0.4609375,
0.432861328125,
0.4609375,
0.446533203125,
0.145263671875,
0.44482421875
],
[
0.14697265625,
0.446533203125,
0.4609375,
0.446533203125,
0.4609375,
0.460205078125,
0.14697265625,
0.460205078125
],
[
0.5341796875,
0.449462890625,
0.73291015625,
0.449462890625,
0.73291015625,
0.460205078125,
0.5341796875,
0.460205078125
],
[
0.779296875,
0.45068359375,
0.8154296875,
0.454833984375,
0.81201171875,
0.47021484375,
0.77490234375,
0.466064453125
],
[
0.8212890625,
0.457275390625,
0.83447265625,
0.457275390625,
0.83447265625,
0.466064453125,
0.8212890625,
0.466064453125
],
[
0.84619140625,
0.454833984375,
0.8623046875,
0.454833984375,
0.8623046875,
0.46435546875,
0.84619140625,
0.46435546875
],
[
0.14697265625,
0.46142578125,
0.4609375,
0.46142578125,
0.4609375,
0.474365234375,
0.14697265625,
0.474365234375
],
[
0.7958984375,
0.465576171875,
0.80615234375,
0.465576171875,
0.80615234375,
0.4697265625,
0.7958984375,
0.4697265625
],
[
0.1485595703125,
0.476806640625,
0.45751953125,
0.476806640625,
0.45751953125,
0.48681640625,
0.1485595703125,
0.48681640625
],
[
0.14697265625,
0.4892578125,
0.4609375,
0.4892578125,
0.4609375,
0.5029296875,
0.14697265625,
0.5029296875
],
[
0.14697265625,
0.5029296875,
0.4609375,
0.5029296875,
0.4609375,
0.5166015625,
0.14697265625,
0.5166015625
],
[
0.51318359375,
0.5029296875,
0.88330078125,
0.50439453125,
0.88330078125,
0.5185546875,
0.51318359375,
0.51708984375
],
[
0.14697265625,
0.51708984375,
0.459228515625,
0.51708984375,
0.459228515625,
0.53076171875,
0.14697265625,
0.53076171875
],
[
0.51123046875,
0.5185546875,
0.88330078125,
0.5185546875,
0.88330078125,
0.5322265625,
0.51123046875,
0.5322265625
],
[
0.1485595703125,
0.533203125,
0.459228515625,
0.533203125,
0.459228515625,
0.54345703125,
0.1485595703125,
0.54345703125
],
[
0.51318359375,
0.53369140625,
0.880859375,
0.53369140625,
0.880859375,
0.54443359375,
0.51318359375,
0.54443359375
],
[
0.14697265625,
0.5458984375,
0.4609375,
0.5458984375,
0.4609375,
0.55859375,
0.14697265625,
0.55859375
],
[
0.51123046875,
0.54638671875,
0.88330078125,
0.54638671875,
0.88330078125,
0.56005859375,
0.51123046875,
0.56005859375
],
[
0.14697265625,
0.56005859375,
0.4609375,
0.56005859375,
0.4609375,
0.57373046875,
0.14697265625,
0.57373046875
],
[
0.51123046875,
0.56103515625,
0.88330078125,
0.56103515625,
0.88330078125,
0.57470703125,
0.51123046875,
0.57470703125
],
[
0.14697265625,
0.57373046875,
0.4609375,
0.57373046875,
0.4609375,
0.58837890625,
0.14697265625,
0.58837890625
],
[
0.51123046875,
0.576171875,
0.884765625,
0.57470703125,
0.884765625,
0.5888671875,
0.51123046875,
0.59033203125
],
[
0.14697265625,
0.59033203125,
0.2166748046875,
0.59033203125,
0.2166748046875,
0.60107421875,
0.14697265625,
0.60107421875
],
[
0.2149658203125,
0.58837890625,
0.459228515625,
0.58837890625,
0.459228515625,
0.6015625,
0.2149658203125,
0.6015625
],
[
0.51318359375,
0.59033203125,
0.884765625,
0.59033203125,
0.884765625,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.1485595703125,
0.60400390625,
0.459228515625,
0.60400390625,
0.459228515625,
0.6142578125,
0.1485595703125,
0.6142578125
],
[
0.51123046875,
0.60498046875,
0.69873046875,
0.60400390625,
0.69873046875,
0.6181640625,
0.51123046875,
0.619140625
],
[
0.145263671875,
0.61767578125,
0.459228515625,
0.61767578125,
0.459228515625,
0.63134765625,
0.145263671875,
0.63134765625
],
[
0.1435546875,
0.63134765625,
0.459228515625,
0.6298828125,
0.459228515625,
0.64453125,
0.1435546875,
0.6455078125
],
[
0.14697265625,
0.6455078125,
0.4609375,
0.6455078125,
0.4609375,
0.6591796875,
0.14697265625,
0.6591796875
],
[
0.515625,
0.6484375,
0.8818359375,
0.6484375,
0.8818359375,
0.6630859375,
0.515625,
0.6630859375
],
[
0.14697265625,
0.6591796875,
0.279541015625,
0.6591796875,
0.279541015625,
0.6728515625,
0.14697265625,
0.6728515625
],
[
0.515625,
0.66455078125,
0.736328125,
0.66455078125,
0.736328125,
0.6787109375,
0.515625,
0.6787109375
],
[
0.154541015625,
0.68505859375,
0.264404296875,
0.68505859375,
0.264404296875,
0.69677734375,
0.154541015625,
0.69677734375
],
[
0.5341796875,
0.68310546875,
0.8818359375,
0.68310546875,
0.8818359375,
0.69775390625,
0.5341796875,
0.69775390625
],
[
0.515625,
0.69970703125,
0.8818359375,
0.69970703125,
0.8818359375,
0.71337890625,
0.515625,
0.71337890625
],
[
0.12005615234375,
0.68603515625,
0.1318359375,
0.68603515625,
0.1318359375,
0.6953125,
0.12005615234375,
0.6953125
],
[
0.12005615234375,
0.708984375,
0.48779296875,
0.708984375,
0.48779296875,
0.72265625,
0.12005615234375,
0.72265625
],
[
0.51318359375,
0.71533203125,
0.87158203125,
0.71435546875,
0.87158203125,
0.728515625,
0.51318359375,
0.72998046875
],
[
0.12005615234375,
0.7255859375,
0.486083984375,
0.7255859375,
0.486083984375,
0.7392578125,
0.12005615234375,
0.7392578125
],
[
0.5146484375,
0.73095703125,
0.8818359375,
0.73095703125,
0.8818359375,
0.7451171875,
0.5146484375,
0.7451171875
],
[
0.12176513671875,
0.74072265625,
0.48779296875,
0.74072265625,
0.48779296875,
0.7548828125,
0.12176513671875,
0.7548828125
],
[
0.5146484375,
0.7451171875,
0.8818359375,
0.74658203125,
0.8818359375,
0.76123046875,
0.5146484375,
0.76025390625
],
[
0.1175537109375,
0.755859375,
0.48779296875,
0.755859375,
0.48779296875,
0.76953125,
0.1175537109375,
0.76953125
],
[
0.515625,
0.7626953125,
0.8818359375,
0.76318359375,
0.8818359375,
0.77783203125,
0.515625,
0.77685546875
],
[
0.118408203125,
0.7724609375,
0.484375,
0.7724609375,
0.484375,
0.7861328125,
0.118408203125,
0.7861328125
],
[
0.515625,
0.78076171875,
0.8818359375,
0.78076171875,
0.8818359375,
0.79443359375,
0.515625,
0.79443359375
],
[
0.118408203125,
0.7880859375,
0.486083984375,
0.7880859375,
0.486083984375,
0.8017578125,
0.118408203125,
0.8017578125
],
[
0.5146484375,
0.79638671875,
0.8818359375,
0.79638671875,
0.8818359375,
0.81005859375,
0.5146484375,
0.81005859375
],
[
0.1175537109375,
0.8046875,
0.48779296875,
0.8046875,
0.48779296875,
0.8193359375,
0.1175537109375,
0.8193359375
],
[
0.515625,
0.8125,
0.88330078125,
0.8125,
0.88330078125,
0.82666015625,
0.515625,
0.82666015625
],
[
0.118408203125,
0.82177734375,
0.489501953125,
0.82177734375,
0.489501953125,
0.8349609375,
0.118408203125,
0.8349609375
],
[
0.515625,
0.8291015625,
0.8818359375,
0.8291015625,
0.8818359375,
0.8427734375,
0.515625,
0.8427734375
],
[
0.118408203125,
0.83740234375,
0.486083984375,
0.83740234375,
0.486083984375,
0.85107421875,
0.118408203125,
0.85107421875
],
[
0.5146484375,
0.84423828125,
0.84033203125,
0.84423828125,
0.84033203125,
0.859375,
0.5146484375,
0.859375
],
[
0.145263671875,
0.859375,
0.489501953125,
0.859375,
0.489501953125,
0.87255859375,
0.145263671875,
0.87255859375
],
[
0.5322265625,
0.86328125,
0.884765625,
0.86328125,
0.884765625,
0.876953125,
0.5322265625,
0.876953125
],
[
0.1158447265625,
0.869140625,
0.4189453125,
0.87158203125,
0.4189453125,
0.88525390625,
0.1158447265625,
0.8828125
],
[
0.139404296875,
0.8818359375,
0.489501953125,
0.88427734375,
0.489501953125,
0.8984375,
0.139404296875,
0.896484375
],
[
0.5146484375,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.118408203125,
0.89794921875,
0.341796875,
0.89794921875,
0.341796875,
0.90771484375,
0.118408203125,
0.90771484375
],
[
0.51708984375,
0.89453125,
0.8818359375,
0.89453125,
0.8818359375,
0.9091796875,
0.51708984375,
0.9091796875
]
] |
[
"来自神经机器翻译的精确单词对齐诱导",
"Yun Chen*1, Yang Liu?, Guanhua Chen3, Xin Jiang4, Qun Liu4",
"1Shanghai University of Finance and Economics, Shanghai, China",
"2Tsinghua University, Beijing, China",
"3The University of Hong Kong, Hong Kong, China",
"4Huawei Noah's Ark Lab, Hong Kong, China",
"yunchen@sufe.edu.cn, liuyang2011@tsinghua.edu.cn.",
"ghchen@eee.hku.hk, { jiang.xin, qun.liu } @huawei.com",
"Abstract",
"Despite its original goal to jointly learn to",
"",
"align and translate, prior researches sug-",
"das",
"gest that Transformer captures poor word",
"das",
"weiB",
"NAIVE-ATT.",
"alignments through its attention mech-",
"weiB",
"ich",
"ich",
"In thispaper,we show that",
"anism.",
"attention weights DO capture accurate",
"word alignments and propose two novel",
"word alignment induction methods Shift-",
"Att and Shift-AET . The main idea is to",
"weiB",
"induce alignments at the step when the to-",
"Source: das weiB ich.",
"ich",
"Dec. input: <bos>i understand this.",
"be-aligned target token is the decoder in-",
"put rather than the decoder output as in",
"Dec. output: i understand this.<eos>",
"Tbosi",
"Ln",
"th",
"previous work..Shift-Att is an interpreta-",
"",
"tion method that induces alignments from",
"the attention weights of Transformer and",
"does not require parameter update or archi-",
"Figure l:An example to compare our method",
"tecture change. Shift-AET extracts align-",
"Shift-Att and the baseline Naive-Att : The left",
"ments from an additional alignment mod-",
"is an attention map from the third decoder layer",
"ule which is tightly integrated into Trans-",
"of the vanilla Transformer and the right are the",
"former and trained in isolation with super-",
"induced alignments. Shift-Att induces alignments",
"vision from symmetrized Shift-Att align-",
"for target word yi at decoding step i+1 when yi is",
"ments.",
"Experiments on three publicly",
"the decoder input, while Naive-Att at step i when",
"available datasets demonstrate that both",
"Yi is the decoder output.",
"methods perform better than their corre",
"sponding neural baselines and Shift-AET",
"significantly outperforms GIZA++ by 1.4-",
"以及在计算机辅助翻译(Dagan etal.，1993)",
"4.8 AER points. 1",
"中为人类翻译人员提供指导",
"Introduction",
"单词对齐是统计机器翻译E(Koehn et al.,",
"2003，SMT）中管道的一部分，但对于神经机",
"1",
"单词对齐的任务是从并行语料库中找到词典",
"器翻译E(Bahdanauetal.，2015,NMT）不",
"翻译等效项E(Brownetal.，1993）。它是自",
"定需要。NMT中的注意机制在功能上没有",
"然语言处理（NLP）的基本任务之一，由社区",
"发挥源与目标之间的单词对齐的作用，至少",
"(Dyer et al., 2013: Brown et al., 1993: Liu and",
"与SMT中的模拟方式不同。很难解释注意",
"Sun，2015）广泛研究。单词对齐在许多情况下",
"力激活和提取有意义的单词对齐，尤其是从",
"都很有用，例如错误分析E(Dingetal.，2017：",
"Transformer（Gargetal.,2019）。结果，使用最",
"Lietal.，2019），引人覆盖率和生育力模型E",
"广泛的单词对齐工具仍然是外部统计模型，例",
"(Tuetal.,2016），在交互式机器翻译E(Hasler",
"如 Fast-Align (Dyer et al., 2013)和 GIZA++",
"etal.，2018;Chenetal.,2020）中插人外部约束",
"(Brown et al., 1993; Och and Ney, 2003)",
"*Corresponding author. Part of the work was done",
"最近，社区中重新出现了研究Trans-",
"when Yun was in Huawei Noah' s Ark Lab.",
"ICodecan be found athttps://github.com/",
"former (Ding et al., 2019; Li et al., 2019) 的",
"sufe-nlp/transformer-alignment",
"单词对齐方式的兴趣。一个简单的解决方案"
] |
{
"arxiv_id": "2004.14837",
"language": "zh",
"page_idx": 0
}
|
||
2004.14837-zh-2
|
2004.14837-zh-2.jpg
|
好地表示$y_{i-1}$而不是$y_i$ ,尤其是对于底层。其次，由于$W_i,j^l$是在观察$y_i$之前计算的，因此很难为目标令牌$y_i$引人对齐的源令牌，如第 1 节中所述
因此，有必要开发新的对齐感应方法。此方法应该能够(???)考虑$z_i^l,y_i$和$y_i-1$的关系，并且(???) 使要对齐的目标令牌适应对齐感应
## 3 Method
在本节中，我们提出了两种新颖的对齐诱导方法 Shift-Att 和 Shift-AET 。两种方法都通过在目标令牌是解码器输入的步骤中计算对齐分数来调整与要对齐的目标令牌的对齐感应
3.1 Shift-Att:Alignment from Vanilla
## Transformer
(Garg
et al., 2019) 在步骤$i$诱导目标令牌$y_i$的对
Alignment Induction Naive-Att
齐，当$y_i$是解码器输出时，并定义对齐分数矩阵与方程。他们通过评估测试集上所有层的AER 来找到最佳层$l$来提取比对
相反，当$y_i$是解码器输人时，我们建议在步骤$i+1$中诱导目标令牌$y_i$的对齐。我们将比对得分矩阵$S$定义为：
$$S_{i,j}=W_{i+1,j}^{l}\quad(1\leq i\leq|\boldsymbol{y}|,1\leq j\leq|\boldsymbol{x}|).$$
这是因为$W_{i+1,j}^l$度量了$z_{i+1}^l$ 和 $h_j$ 之间的相关性，并且我们使用$z_i+1^l$和$h_j$分别表示$y_i$ 和$x_j$ 。使用对齐分数矩阵$S$ ,我们可以使用公式$\overline{\mathrm{E}}$3 提取单词对齐 $A$ 。我们称这个方法为 Shift-Att 。Fig。1 显示了一个对齐归纳示例，用于比较 Naive-Att 和 Shift-Att
Shift-Att 使用$z_i+1^l$表示要对齐的目标标记$y_i$ ,而 Naive-Att 使用 $z_i^l$。我们认为使用 $z_i+1^l$ 更好。首先，在底层，我们假设$z_i+1^l$可以比输出$y_{i+1}$更好地表示解码器输入$y_i$ 。因此，我们可以使用$z_i+1^l$和小$l$来表示$y_i$ 。其次，在观察$y_i$之后计算$z_i+1^l$,这表明 Shift-Att 能够使要对齐的目标令牌适应对齐感应
我们提出的方法涉及从源到目标和目标到源的香草变压器模型诱导对齐。在 Zenkel et al.(2019) 之后，我们使用增长对角启发式 (Koehn et al., 2005) 合并双向对齐
Layer Selection Criterion 为了选择最佳层$l_b$来诱导对齐，我们提出了一种替代层选择标准，而无需手动标记单词对齐。实验表明，该准则与 AER 度量值具有很好的相关性。
给定并行句子对$\langle x,y\rangle$ ,我们训练源到目标模型$\theta_{x\to y}$和目标到源模型$\theta_y\to x$ 。我们假
<FigureHere>
Figure 2: Illustration of the alignment module at decoding step $i.$ The decoder input token is $y_i-1$ ,while the output token is $y_i.$The alignment module predicts $S_{i-1}$ , the alignment scores corresponding to the input target token $y_{i-1}.$ During ${\mathrm{the~alignment~module~training~process,~parameters}}$ of the blue blocks are frozen, and only parameters of the orange blocks are updated.
设从这两个模型中提取的单词对齐应该彼此一致 (Cheng et al., 2016) 。因此，我们通过计算验证集上的 AER 分数来评估比对的质量，其中源到目标的比对作为假设，目标到源的比对作为参考。对于每个模型，我们可以从$L$不同的层获得$L$单词对齐。总的来说，我们获得了$L\times L$ AER 分数。我们选择 AER 得分最低的一个，其对应的源到目标和目标到源模型层是我们将在测试时用于提取比对的层：
$$l_{\mathrm{b},\boldsymbol{x}\to\boldsymbol{y}},l_{\mathrm{b},\boldsymbol{y}\to\boldsymbol{x}}=\underset{i,j}{\mathrm{argmin}}\operatorname{AER}(\boldsymbol{A}_{\boldsymbol{x}\to\boldsymbol{y}}^{i},\boldsymbol{A}_{\boldsymbol{y}\to\boldsymbol{x}}^{j}).$$
3.2 Shift-AET:Alignment from
Alignment-Enhanced Transformer
为了进一步提高对准精度，我们提出了 ShiftAET , 这是一种字对齐感应方法，用于从对齐增强转换器 ( AET) 中提取对齐。AET 通过单独的对齐模块扩展了 Transformer 架构，该模块在每个步骤中观察底层 Transformer 的隐藏状态，并预测当前解码器输入的对齐分数。请注意，此模块是一个即插即用组件，它既不对基础 NMT 模型进行任何更改，也不影响翻译质量
图。2 说明了 AET 在解码步骤$i$的对齐模块。我们仅在$l_b$层添加对齐模块，这是使用Shift-Att 提取对齐的最佳层。对齐模块执行类似于编码器-解码器注意子层的多头注意。它以编码器输出$\boldsymbol{h}=\{h_1,...,h_{|\boldsymbol{x}|}\}$和当前解码器隐藏状态$\tilde{z}_i^{l_b}$在层$l_b$内作为输入，并输出$S_{i-1}$ ,与目标字$y_i-1$相对应的对齐分数：
$$S_{i-1}=\frac{1}{N}\sum_{n}\mathrm{softmax}(\frac{(\boldsymbol{hG}_{n}^{K})(\tilde{z}_{i}^{l_{b}}\boldsymbol{G}_{n}^{Q})^{\top}}{\sqrt{d_{k}}}),$$
(5)
|
[
[
0.12005615234375,
0.07659912109375,
0.484375,
0.07659912109375,
0.484375,
0.0914306640625,
0.12005615234375,
0.0914306640625
],
[
0.76171875,
0.0736083984375,
0.86572265625,
0.0748291015625,
0.86572265625,
0.08612060546875,
0.76171875,
0.08489990234375
],
[
0.611328125,
0.0831298828125,
0.68603515625,
0.0831298828125,
0.68603515625,
0.09320068359375,
0.611328125,
0.09320068359375
],
[
0.12005615234375,
0.0926513671875,
0.48779296875,
0.0926513671875,
0.48779296875,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.51904296875,
0.09027099609375,
0.58837890625,
0.0926513671875,
0.58837890625,
0.10394287109375,
0.51904296875,
0.1015625
],
[
0.79443359375,
0.09442138671875,
0.8369140625,
0.0968017578125,
0.83642578125,
0.10986328125,
0.79248046875,
0.10809326171875
],
[
0.12176513671875,
0.1092529296875,
0.489501953125,
0.1092529296875,
0.489501953125,
0.12353515625,
0.12176513671875,
0.12353515625
],
[
0.611328125,
0.11224365234375,
0.68408203125,
0.11224365234375,
0.68408203125,
0.12353515625,
0.611328125,
0.12353515625
],
[
0.12176513671875,
0.1258544921875,
0.2333984375,
0.1258544921875,
0.2333984375,
0.13720703125,
0.12176513671875,
0.13720703125
],
[
0.77490234375,
0.122314453125,
0.84619140625,
0.122314453125,
0.84619140625,
0.1341552734375,
0.77490234375,
0.1341552734375
],
[
0.599609375,
0.13720703125,
0.68994140625,
0.13720703125,
0.68994140625,
0.1484375,
0.599609375,
0.1484375
],
[
0.78857421875,
0.1341552734375,
0.8330078125,
0.1341552734375,
0.8330078125,
0.141357421875,
0.78857421875,
0.141357421875
],
[
0.139404296875,
0.14013671875,
0.484375,
0.14013671875,
0.484375,
0.15380859375,
0.139404296875,
0.15380859375
],
[
0.59375,
0.1395263671875,
0.60107421875,
0.1395263671875,
0.60107421875,
0.144287109375,
0.59375,
0.144287109375
],
[
0.51708984375,
0.15087890625,
0.5615234375,
0.15087890625,
0.5615234375,
0.160888671875,
0.51708984375,
0.160888671875
],
[
0.77734375,
0.1455078125,
0.78662109375,
0.1455078125,
0.78662109375,
0.149658203125,
0.77734375,
0.149658203125
],
[
0.118408203125,
0.1561279296875,
0.48779296875,
0.15673828125,
0.48779296875,
0.171630859375,
0.118408203125,
0.17041015625
],
[
0.75146484375,
0.157958984375,
0.8017578125,
0.157958984375,
0.8017578125,
0.1728515625,
0.75146484375,
0.1728515625
],
[
0.8212890625,
0.157958984375,
0.8623046875,
0.1602783203125,
0.861328125,
0.1728515625,
0.8193359375,
0.17041015625
],
[
0.61279296875,
0.16748046875,
0.67822265625,
0.16748046875,
0.67822265625,
0.1787109375,
0.61279296875,
0.1787109375
],
[
0.12005615234375,
0.17333984375,
0.48779296875,
0.17333984375,
0.48779296875,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.76416015625,
0.188232421875,
0.779296875,
0.188232421875,
0.779296875,
0.198974609375,
0.76416015625,
0.198974609375
],
[
0.8212890625,
0.18701171875,
0.86376953125,
0.18701171875,
0.86376953125,
0.198974609375,
0.8212890625,
0.198974609375
],
[
0.60107421875,
0.19482421875,
0.69189453125,
0.1923828125,
0.69189453125,
0.207275390625,
0.60107421875,
0.208984375
],
[
0.1175537109375,
0.21728515625,
0.225830078125,
0.21728515625,
0.225830078125,
0.232177734375,
0.1175537109375,
0.232177734375
],
[
0.12005615234375,
0.191162109375,
0.156982421875,
0.191162109375,
0.156982421875,
0.2037353515625,
0.12005615234375,
0.2037353515625
],
[
0.5146484375,
0.2203369140625,
0.88330078125,
0.2203369140625,
0.88330078125,
0.2340087890625,
0.5146484375,
0.2340087890625
],
[
0.51318359375,
0.235107421875,
0.8818359375,
0.235107421875,
0.8818359375,
0.248779296875,
0.51318359375,
0.248779296875
],
[
0.12005615234375,
0.243408203125,
0.48779296875,
0.243408203125,
0.48779296875,
0.257080078125,
0.12005615234375,
0.257080078125
],
[
0.51904296875,
0.2476806640625,
0.76318359375,
0.2493896484375,
0.76318359375,
0.26318359375,
0.51904296875,
0.26123046875
],
[
0.76318359375,
0.248779296875,
0.880859375,
0.2493896484375,
0.880859375,
0.26123046875,
0.76318359375,
0.260009765625
],
[
0.12005615234375,
0.2587890625,
0.48779296875,
0.2587890625,
0.48779296875,
0.2724609375,
0.12005615234375,
0.2724609375
],
[
0.51318359375,
0.26318359375,
0.8818359375,
0.26318359375,
0.8818359375,
0.27783203125,
0.51318359375,
0.27783203125
],
[
0.12005615234375,
0.275634765625,
0.48779296875,
0.275634765625,
0.48779296875,
0.289306640625,
0.12005615234375,
0.289306640625
],
[
0.5146484375,
0.27783203125,
0.88330078125,
0.27783203125,
0.88330078125,
0.29150390625,
0.5146484375,
0.29150390625
],
[
0.12005615234375,
0.29150390625,
0.450927734375,
0.29150390625,
0.450927734375,
0.3046875,
0.12005615234375,
0.3046875
],
[
0.51318359375,
0.290283203125,
0.8818359375,
0.292236328125,
0.8818359375,
0.305908203125,
0.51318359375,
0.303955078125
],
[
0.51318359375,
0.305908203125,
0.88330078125,
0.305908203125,
0.88330078125,
0.319580078125,
0.51318359375,
0.319580078125
],
[
0.12005615234375,
0.319580078125,
0.1695556640625,
0.319580078125,
0.1695556640625,
0.330810546875,
0.12005615234375,
0.330810546875
],
[
0.162109375,
0.318359375,
0.447509765625,
0.318359375,
0.447509765625,
0.33203125,
0.162109375,
0.33203125
],
[
0.51123046875,
0.319580078125,
0.759765625,
0.319580078125,
0.759765625,
0.333251953125,
0.51123046875,
0.333251953125
],
[
0.15869140625,
0.333251953125,
0.26123046875,
0.3349609375,
0.26123046875,
0.349853515625,
0.15869140625,
0.347412109375
],
[
0.130126953125,
0.355224609375,
0.4072265625,
0.35400390625,
0.4072265625,
0.3681640625,
0.130126953125,
0.369384765625
],
[
0.438232421875,
0.35693359375,
0.48779296875,
0.35693359375,
0.48779296875,
0.3681640625,
0.438232421875,
0.3681640625
],
[
0.515625,
0.359375,
0.87744140625,
0.359375,
0.87744140625,
0.372314453125,
0.515625,
0.372314453125
],
[
0.12005615234375,
0.371826171875,
0.48779296875,
0.371826171875,
0.48779296875,
0.385498046875,
0.12005615234375,
0.385498046875
],
[
0.5146484375,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.3896484375,
0.5146484375,
0.3896484375
],
[
0.118408203125,
0.385986328125,
0.486083984375,
0.385986328125,
0.486083984375,
0.399658203125,
0.118408203125,
0.399658203125
],
[
0.5146484375,
0.391357421875,
0.88330078125,
0.391357421875,
0.88330078125,
0.405029296875,
0.5146484375,
0.405029296875
],
[
0.12176513671875,
0.40380859375,
0.489501953125,
0.40380859375,
0.489501953125,
0.41748046875,
0.12176513671875,
0.41748046875
],
[
0.5146484375,
0.40625,
0.8818359375,
0.407958984375,
0.8818359375,
0.4228515625,
0.5146484375,
0.42041015625
],
[
0.12005615234375,
0.419189453125,
0.38623046875,
0.419189453125,
0.38623046875,
0.432861328125,
0.12005615234375,
0.432861328125
],
[
0.515625,
0.42333984375,
0.8818359375,
0.42333984375,
0.8818359375,
0.43701171875,
0.515625,
0.43701171875
],
[
0.1376953125,
0.435791015625,
0.48779296875,
0.435791015625,
0.48779296875,
0.449462890625,
0.1376953125,
0.449462890625
],
[
0.515625,
0.440673828125,
0.880859375,
0.440673828125,
0.880859375,
0.45361328125,
0.515625,
0.45361328125
],
[
0.118408203125,
0.45068359375,
0.48779296875,
0.451904296875,
0.48779296875,
0.466064453125,
0.118408203125,
0.465576171875
],
[
0.51318359375,
0.45361328125,
0.8818359375,
0.4560546875,
0.8818359375,
0.47021484375,
0.51318359375,
0.468505859375
],
[
0.12176513671875,
0.468505859375,
0.31982421875,
0.468505859375,
0.31982421875,
0.482177734375,
0.12176513671875,
0.482177734375
],
[
0.52880859375,
0.471435546875,
0.8818359375,
0.471435546875,
0.8818359375,
0.486328125,
0.52880859375,
0.486328125
],
[
0.5146484375,
0.48681640625,
0.79833984375,
0.488037109375,
0.79833984375,
0.5029296875,
0.5146484375,
0.501953125
],
[
0.1259765625,
0.49462890625,
0.489501953125,
0.495361328125,
0.489501953125,
0.51708984375,
0.1259765625,
0.5166015625
],
[
0.515625,
0.5126953125,
0.87744140625,
0.5126953125,
0.87744140625,
0.533203125,
0.515625,
0.533203125
],
[
0.118408203125,
0.529296875,
0.489501953125,
0.529296875,
0.489501953125,
0.54638671875,
0.118408203125,
0.54638671875
],
[
0.6708984375,
0.525390625,
0.6943359375,
0.52978515625,
0.68994140625,
0.5419921875,
0.66650390625,
0.5380859375
],
[
0.12005615234375,
0.54638671875,
0.48779296875,
0.54736328125,
0.48779296875,
0.5625,
0.12005615234375,
0.56103515625
],
[
0.5146484375,
0.5517578125,
0.5615234375,
0.5517578125,
0.5615234375,
0.5634765625,
0.5146484375,
0.5634765625
],
[
0.5576171875,
0.55029296875,
0.79248046875,
0.55029296875,
0.79248046875,
0.56396484375,
0.5576171875,
0.56396484375
],
[
0.12176513671875,
0.5634765625,
0.486083984375,
0.5634765625,
0.486083984375,
0.57763671875,
0.12176513671875,
0.57763671875
],
[
0.5576171875,
0.568359375,
0.82861328125,
0.568359375,
0.82861328125,
0.5791015625,
0.5576171875,
0.5791015625
],
[
0.12005615234375,
0.57763671875,
0.48779296875,
0.5791015625,
0.48779296875,
0.59326171875,
0.12005615234375,
0.5927734375
],
[
0.515625,
0.58740234375,
0.88330078125,
0.58740234375,
0.88330078125,
0.60107421875,
0.515625,
0.60107421875
],
[
0.12005615234375,
0.59326171875,
0.48779296875,
0.59423828125,
0.48779296875,
0.609375,
0.12005615234375,
0.60791015625
],
[
0.515625,
0.6025390625,
0.8818359375,
0.6025390625,
0.8818359375,
0.6162109375,
0.515625,
0.6162109375
],
[
0.12005615234375,
0.6123046875,
0.4189453125,
0.60986328125,
0.4189453125,
0.62353515625,
0.12005615234375,
0.6259765625
],
[
0.515625,
0.6181640625,
0.8818359375,
0.6181640625,
0.8818359375,
0.6318359375,
0.515625,
0.6318359375
],
[
0.1368408203125,
0.62744140625,
0.48779296875,
0.62744140625,
0.48779296875,
0.64111328125,
0.1368408203125,
0.64111328125
],
[
0.5146484375,
0.63427734375,
0.8818359375,
0.63525390625,
0.8818359375,
0.6494140625,
0.5146484375,
0.6484375
],
[
0.118408203125,
0.642578125,
0.486083984375,
0.642578125,
0.486083984375,
0.66015625,
0.118408203125,
0.66015625
],
[
0.5146484375,
0.65185546875,
0.8818359375,
0.65185546875,
0.8818359375,
0.6650390625,
0.5146484375,
0.6650390625
],
[
0.12176513671875,
0.66015625,
0.486083984375,
0.66015625,
0.486083984375,
0.6748046875,
0.12176513671875,
0.6748046875
],
[
0.51318359375,
0.66748046875,
0.8818359375,
0.66845703125,
0.8818359375,
0.681640625,
0.51318359375,
0.68115234375
],
[
0.12005615234375,
0.6767578125,
0.489501953125,
0.6767578125,
0.489501953125,
0.6904296875,
0.12005615234375,
0.6904296875
],
[
0.5146484375,
0.68408203125,
0.8818359375,
0.68408203125,
0.8818359375,
0.69775390625,
0.5146484375,
0.69775390625
],
[
0.12005615234375,
0.6923828125,
0.48779296875,
0.6923828125,
0.48779296875,
0.70703125,
0.12005615234375,
0.70703125
],
[
0.515625,
0.69970703125,
0.8818359375,
0.69970703125,
0.8818359375,
0.71337890625,
0.515625,
0.71337890625
],
[
0.12176513671875,
0.708984375,
0.48779296875,
0.708984375,
0.48779296875,
0.7236328125,
0.12176513671875,
0.7236328125
],
[
0.515625,
0.71630859375,
0.55322265625,
0.71630859375,
0.55322265625,
0.728515625,
0.515625,
0.728515625
],
[
0.12005615234375,
0.724609375,
0.414794921875,
0.724609375,
0.414794921875,
0.73828125,
0.12005615234375,
0.73828125
],
[
0.5341796875,
0.732421875,
0.8818359375,
0.732421875,
0.8818359375,
0.7451171875,
0.5341796875,
0.7451171875
],
[
0.1376953125,
0.7412109375,
0.486083984375,
0.7412109375,
0.486083984375,
0.7548828125,
0.1376953125,
0.7548828125
],
[
0.51318359375,
0.74658203125,
0.880859375,
0.74658203125,
0.880859375,
0.76025390625,
0.51318359375,
0.76025390625
],
[
0.12005615234375,
0.75732421875,
0.48779296875,
0.75732421875,
0.48779296875,
0.77099609375,
0.12005615234375,
0.77099609375
],
[
0.5146484375,
0.7626953125,
0.88330078125,
0.76318359375,
0.88330078125,
0.77783203125,
0.5146484375,
0.77685546875
],
[
0.118408203125,
0.77392578125,
0.484375,
0.77392578125,
0.484375,
0.78369140625,
0.118408203125,
0.78369140625
],
[
0.5146484375,
0.779296875,
0.8818359375,
0.77978515625,
0.8818359375,
0.79443359375,
0.5146484375,
0.79345703125
],
[
0.12005615234375,
0.7890625,
0.426513671875,
0.7890625,
0.426513671875,
0.802734375,
0.12005615234375,
0.802734375
],
[
0.5146484375,
0.7958984375,
0.8818359375,
0.79638671875,
0.8818359375,
0.81103515625,
0.5146484375,
0.81005859375
],
[
0.1259765625,
0.8134765625,
0.489501953125,
0.8134765625,
0.489501953125,
0.82763671875,
0.1259765625,
0.82763671875
],
[
0.51318359375,
0.81396484375,
0.884765625,
0.8154296875,
0.884765625,
0.830078125,
0.51318359375,
0.8291015625
],
[
0.118408203125,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51123046875,
0.83056640625,
0.80419921875,
0.830078125,
0.80419921875,
0.84423828125,
0.51123046875,
0.845703125
],
[
0.118408203125,
0.8466796875,
0.48779296875,
0.8466796875,
0.48779296875,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.59765625,
0.86083984375,
0.6337890625,
0.8642578125,
0.63037109375,
0.87451171875,
0.59423828125,
0.87158203125
],
[
0.71728515625,
0.85693359375,
0.8505859375,
0.85498046875,
0.8505859375,
0.87255859375,
0.71728515625,
0.87451171875
],
[
0.5341796875,
0.865234375,
0.62890625,
0.876953125,
0.62451171875,
0.892578125,
0.52880859375,
0.88134765625
],
[
0.623046875,
0.865234375,
0.71728515625,
0.86767578125,
0.71728515625,
0.8818359375,
0.623046875,
0.8798828125
],
[
0.84619140625,
0.87158203125,
0.8623046875,
0.87158203125,
0.8623046875,
0.8798828125,
0.84619140625,
0.8798828125
],
[
0.12005615234375,
0.8623046875,
0.4375,
0.8623046875,
0.4375,
0.8759765625,
0.12005615234375,
0.8759765625
],
[
0.1368408203125,
0.87744140625,
0.484375,
0.87890625,
0.484375,
0.8935546875,
0.1368408203125,
0.892578125
],
[
0.62744140625,
0.8818359375,
0.640625,
0.8896484375,
0.63037109375,
0.89794921875,
0.61865234375,
0.89111328125
],
[
0.765625,
0.876953125,
0.8037109375,
0.876953125,
0.8037109375,
0.892578125,
0.765625,
0.892578125
],
[
0.12005615234375,
0.8955078125,
0.48779296875,
0.8955078125,
0.48779296875,
0.91015625,
0.12005615234375,
0.91015625
],
[
0.8564453125,
0.8935546875,
0.884765625,
0.8935546875,
0.884765625,
0.9111328125,
0.8564453125,
0.9111328125
]
] |
[
"好地表示yi-1而不是i，尤其是对于底层，",
"Alignment Scores Si-1",
"Hidden state zib",
"其次，由于W是在观察yi之前计算的，因",
"Ih Dec. Layer",
"Average",
"此很难为目标令牌引人对齐的源令牌，如",
"Feed Forward",
"第1节中所述",
"Scaled MatMul",
"Enc-DecAttention",
"+SoftMax",
"因此，有必要开发新的对齐感应方法。此方",
"",
"Enc.Out",
"小",
"法应该能够（???）考虑，yi和9i-1的关",
"Linear",
"Linear",
"Self-attention",
"系，并且（？?？）使要对齐的目标令牌适应对齐",
"2b",
"Enc. Out",
"Hidden state z/b-1",
"3 Method",
"感应",
"Figure 2: Illustration of the alignment module at",
"decoding step i. The decoder input token is yi-1",
"在本节中，我们提出了两种新颖的对齐诱导方",
"while the outputtoken is yi.",
"The alignment",
"法Shift-Att和Shift-AET。两种方法都通过",
"module predicts Si-1 , the alignment scores corre",
"在目标令牌是解码器输入的步骤中计算对齐分",
"sponding to the input target token i-1 .During",
"数来调整与要对齐的目标令牌的对齐感应",
"the alignment module training process, parameters",
"of the blue blocks are frozen, and only parameters",
"3.1",
"Shift-Att : Alignment from Vanilla",
"of the orange blocks are updated.",
"Transformer",
"Alignment Induction Naive-Att",
"(Garg",
"设从这两个模型中提取的单词对齐应该彼此一",
"etal.，2019）在步骤i诱导目标令牌yi的对",
"致（Chengetal.，2016）。因此，我们通过计算",
"齐，当出是解码器输出时，并定义对齐分数",
"验证集上的AER分数来评估比对的质量，其",
"矩阵与方程。他们通过评估测试集上所有层的",
"中源到目标的比对作为假设，目标到源的比对",
"AER来找到最佳层1来提取比对",
"作为参考。对于每个模型，我们可以从L不同",
"相反，当yi是解码器输入时，我们建议在",
"的层获得L单词对齐。总的来说，我们获得了",
"步骤i+1中诱导目标令牌yi的对齐。我们将",
"L×LAER分数。我们选择AER得分最低的",
"比对得分矩阵S定义为：",
"个，其对应的源到目标和目标到源模型层是",
"我们将在测试时用于提取比对的层：",
"Sij=Wi+1,3 (1≤i≤lyl,1≤j≤[al). (4)",
"lb,a-y,lb,y- = argmin AER(A-y, Ag→a)",
"这是因为Wi+1,j度量了zi+1和hj之间的相",
"i.j",
"关性，并且我们使用z+1和hi分别表示yi",
"3.2",
"Shift-AET: Alignment from",
"和&i。使用对齐分数矩阵S，我们可以使用",
"Alignment-EnhancedTransformer",
"公式E3提取单词对齐A。我们称这个方法",
"为了进一步提高对准精度，我们提出了Shift-",
"为Shift-Att。Fig。1显示了一个对齐归纳示",
"AET，这是一种字对齐感应方法，用于从对齐",
"例，用于比较Naive-Att和Shift-Att",
"增强转换器（AET）中提取对齐。AET通过单",
"Shift-Att使用zi+1表示要对齐的目标标记",
"独的对齐模块扩展了Transformer架构，该模",
"yi，而Naive-Att使用z。我们认为使用z+1",
"块在每个步骤中观察底层Transformer的隐藏",
"更好。首先，在底层，我们假设2+1可以比输",
"状态，并预测当前解码器输入的对齐分数。请",
"出9i+1更好地表示解码器输入9i。因此，我",
"注意，此模块是一个即插即用组件，它既不对",
"们可以使用2+1和小l来表示i。其次，在",
"基础NMT模型进行任何更改，也不影响翻译",
"观察yi之后计算zi+1，这表明Shift-Att能",
"质量",
"够使要对齐的目标令牌适应对齐感应",
"图。2说明了AET在解码步骤i的对齐模",
"我们提出的方法涉及从源到目标和目标到",
"块。我们仅在lh层添加对齐模块，这是使用",
"源的香草变压器模型诱导对齐。在Zenkel",
"Shift-Att提取对齐的最佳层。对齐模块执行类",
"etal（2019）之后，我们使用增长对角启发",
"似于编码器-解码器注意子层的多头注意。它",
"式（Koehnetal.，2005）合并双向对齐",
"以编码器输出h={h1,,hal}和当前解码器",
"Layer SelectionCriterion为了选择最佳层",
"隐藏状态在层lb内作为输入，并输出Si-1",
"lb来诱导对齐，我们提出了一种替代层选择标",
"，与目标字i-1相对应的对齐分数：",
"准，而无需手动标记单词对齐。实验表明，该",
"17",
"(hGK)(zG)T",
"Si-1 = N Z",
"softmax(",
"）",
"准则与AER度量值具有很好的相关性。",
"给定并行句子对《α,y），我们训练源到目",
"T",
"Vdk",
"标模型a→y和目标到源模型ay→a。我们假",
"(5)"
] |
{
"arxiv_id": "2004.14837",
"language": "zh",
"page_idx": 2
}
|
||
2004.14837-zh-3
|
2004.14837-zh-3.jpg
|
<table>
<tbody>
<tr>
<th>${\mathrm{Dataset}}$</th>
<th>$\overline{\text{Train}}$</th>
<th>${\overline{{Validation}}}$</th>
<th>$\overline{\text{Test}}$</th>
</tr>
<tr>
<td>$\overline{\mathrm{de-en}}$</td>
<td>1.9M</td>
<td>$\overline{994}$</td>
<td>$\overline{508}$</td>
</tr>
<tr>
<td>fr-en</td>
<td>1.1M</td>
<td>1,000</td>
<td>447</td>
</tr>
<tr>
<td>ro-en</td>
<td>0.5M</td>
<td>999</td>
<td>248</td>
</tr>
</tbody>
</table>
Table 1: Number of sentences in each dataset
其中$G_n^K,G_n^Q\in\mathbb{R}^{d_{\mathrm{model}}\times d_k}$是$n$的关键和查询投影矩阵注意第 XMATHX\_。由于我们只关心注意力权重，因此在此模块中省略了与价值相关的参数和计算
为了训练对齐模块，我们使用从香草变压器模型中提取的对称化 Shift-Att 对齐作为标签。具体来说，当底层变压器是预先训练和固定的(图。$\overline {\mathrm{E} }2$ ), 我们使用$\overline\mathrm{E}$Garg et al.(2019) 以下的损失函数训练对齐模块：
$$\mathcal{L}_a=-\frac{1}{|\boldsymbol{y}|}\sum_{i=1}^{|\boldsymbol{y}|}\sum_{j=1}^{|\boldsymbol{x}|}\big(\hat{A}_{i,j}^p\odot\log S_{i,j}\big),$$
(6)
其中$S=\{S_{1};...;S_{|\boldsymbol{y}|}\}$是对齐模块预测的对齐分数矩阵，$\hat{A}^p$表示归一化的参考对称化Shift-Att 对齐。$^{2}$通过这种方式，我们将在两个 vanillaTransformer 模型$\theta_{x\to y}$和$\theta_{y\to x}$中隐式学习的对齐知识转移到单个 AET 模型的对齐模块中
一日对对齐模块讲行了训练、我们就可以在给定平行句子对的情况下从中提取对齐分数$S$ ,并使用公式$\mathbb{E}$3 诱导对齐 $A$
## 4 Experiments
# 4.1 Settings
Dataset 我们遵循先前的工作$\overline{\mathrm{E}}($Zenkel et al., 2019, 2020) 进行数据设置，并对德语-英语 (de-en)$^3$,罗马尼亚语-英语(ro-en)和法语-英语 (fr-en)$^{4}$的公开可用数据集。由于未提供验证集，因此我们按照 Ding et al. (2019) 将预处理前训练数据的最后 1,000 个句子设置为验证集。我们学习了具有 10k 合并操作的联合源和目标字节对编码$\overline{\mathrm{E}}($Sennrich et al., 2016)。表 1显示了详细的数据统计
NMT Systems 我们使用 fairseq-py $^{5}$实现Transformer, 并使用$\overline\text{EDing et al.(2019)之}$
后的 transformer\_iwslt\_de\_en 模型配置。我
$^{2}$We simply normalize rows corresponding to target tokens that are aligned to at least one source token of $\hat{A}$ .
$^3$https://www-i6.informatik.rwth-aachen.de/
$\underset{41}{\operatorname*{\operatorname*{goldAlignment}}}/$
$^{4}$http://web.eecs.umich.edu/~mihalcea/wpt/index.
html
$^{5}$https://github.com/pytorch/fairseq
们使用 36K 令牌的批量大小训练模型，并将Transformer 和 AET 的最大更新分别设置为50K 和 10K。AET 的最后一个检查点用于评估。所有模型都在两个翻译方向上进行训练，并使用 Zenkel et al. (2019) 中的脚本使用$\underset{a}{\operatorname*{grow-diag}}\underline{{\mathrm{E}}}($Koehn et al.,2005)进行对称化。Evaluation 我们使用对齐错误率 (Och and Ney, 2000, AER) 评估我们方法的对齐质量。由于单词对齐对于 1 节中讨论的许多下游任务都很有用，因此我们还使用度量 BLEUE (Papineni et al., 2002) 评估了字典指导解码的方法，这是对齐归纳的下游任务。更多细节见4.3节
Baselines 我们将我们的方法与两个统计基线 Fast-Align 和 GIZA++ 以及其他九个基线进行比较：
$\bullet$ Naive-Att $\boxed{\mathrm{E}}($Garg et al.,2019):我们在 2.2节中讨论的方法，该方法从变压器倒数第二层的注意力权重中引起对齐
$\bullet$ Naive-Att-LA (Garg et al.,2019):没有层选择的 Naive-Att 方法。它从所有层的平均注意力权重中诱导出对齐
$\bullet$ Shift-Att-LA:没有层选择的 Shift-Att 方法。
它从所有层的平均注意力权重中诱导出对齐
$\bullet$ SmoothGrad $\boxed{\mathrm{E}}($Li et al.,2016):从单词显着性诱导对齐的方法，该方法是通过以多个嘈杂的句子对作为输入对基于梯度的显着性得分进行平均来计算的
$\bullet$SD-SmoothGrad$\boxtimes$E(Ding $et al. , 2019) :$ SmoothGrad 的改进版本，它在单热输入向量上定义了显着性，而不是单词嵌人。
$\bullet$PD (Li et al.,2019):通过迭代地屏蔽每个源令牌并测量预测差异来计算 Transformer 的对齐分数的方法
$\bullet$ AddSGD $\mathsf{E}($Zenkel et al., 2019):在 Transformer 顶部显式添加额外关注层并直接优化其激活以预测要对齐的目标令牌的方法
$\bullet$Mtl-Fullc $\underline{\mathrm{E}}($Garg et al., 2019):在多任务学习框架中训练单个模型以同时预测目标句子和对齐方式的方法。在预测对齐方式时，模型观察完整的目标句子，并使用对称的 Naive-Att 对齐方式作为标签
$\bullet$ Mtl-Fullc-GZ (Garg et al.,2019):与 MtlFullc 相同的方法，除了使用对称的 GIZA++ 比对作为标签。这是一种统计和神经方法，因为它依赖于 GIZA++ 比对
在这九个基线和我们提出的方法中SmoothGrad,SD-SmoothGrad 和 PD 使用
${^{6}\mathrm{https://git}}$hub.com/lilt/alignment-scripts
|
[
[
0.1695556640625,
0.07183837890625,
0.291259765625,
0.07183837890625,
0.291259765625,
0.08489990234375,
0.1695556640625,
0.08489990234375
],
[
0.307373046875,
0.07183837890625,
0.378662109375,
0.0736083984375,
0.378662109375,
0.08489990234375,
0.307373046875,
0.0831298828125
],
[
0.397216796875,
0.0736083984375,
0.43408203125,
0.0736083984375,
0.43408203125,
0.08489990234375,
0.397216796875,
0.08489990234375
],
[
0.515625,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.515625,
0.09027099609375
],
[
0.175537109375,
0.08489990234375,
0.2166748046875,
0.0872802734375,
0.2149658203125,
0.09857177734375,
0.173828125,
0.0968017578125
],
[
0.2493896484375,
0.08612060546875,
0.291259765625,
0.08612060546875,
0.291259765625,
0.097412109375,
0.2493896484375,
0.097412109375
],
[
0.353515625,
0.08612060546875,
0.382080078125,
0.08612060546875,
0.382080078125,
0.097412109375,
0.353515625,
0.097412109375
],
[
0.399658203125,
0.08612060546875,
0.432373046875,
0.08612060546875,
0.432373046875,
0.097412109375,
0.399658203125,
0.097412109375
],
[
0.177978515625,
0.0968017578125,
0.214111328125,
0.0968017578125,
0.214111328125,
0.1092529296875,
0.177978515625,
0.1092529296875
],
[
0.252685546875,
0.0968017578125,
0.28955078125,
0.0968017578125,
0.28955078125,
0.1092529296875,
0.252685546875,
0.1092529296875
],
[
0.4013671875,
0.0968017578125,
0.429931640625,
0.0968017578125,
0.429931640625,
0.1092529296875,
0.4013671875,
0.1092529296875
],
[
0.5146484375,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.2451171875,
0.10809326171875,
0.291259765625,
0.10809326171875,
0.291259765625,
0.122314453125,
0.2451171875,
0.122314453125
],
[
0.341796875,
0.097412109375,
0.382080078125,
0.097412109375,
0.382080078125,
0.1092529296875,
0.341796875,
0.1092529296875
],
[
0.353515625,
0.1092529296875,
0.3955078125,
0.1092529296875,
0.3955078125,
0.122314453125,
0.353515625,
0.122314453125
],
[
0.5146484375,
0.1092529296875,
0.8818359375,
0.1092529296875,
0.8818359375,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.39794921875,
0.10986328125,
0.435791015625,
0.10986328125,
0.435791015625,
0.1217041015625,
0.39794921875,
0.1217041015625
],
[
0.17041015625,
0.10986328125,
0.214111328125,
0.10809326171875,
0.2149658203125,
0.12054443359375,
0.172119140625,
0.122314453125
],
[
0.5146484375,
0.12469482421875,
0.880859375,
0.12469482421875,
0.880859375,
0.1383056640625,
0.5146484375,
0.1383056640625
],
[
0.133544921875,
0.1341552734375,
0.470947265625,
0.1341552734375,
0.470947265625,
0.1478271484375,
0.133544921875,
0.1478271484375
],
[
0.515625,
0.141357421875,
0.8818359375,
0.141357421875,
0.8818359375,
0.155029296875,
0.515625,
0.155029296875
],
[
0.51318359375,
0.15673828125,
0.88330078125,
0.1561279296875,
0.88330078125,
0.17041015625,
0.51318359375,
0.171630859375
],
[
0.1158447265625,
0.169189453125,
0.490234375,
0.17041015625,
0.490234375,
0.191162109375,
0.1158447265625,
0.190673828125
],
[
0.118408203125,
0.190673828125,
0.486083984375,
0.190673828125,
0.486083984375,
0.2037353515625,
0.118408203125,
0.2037353515625
],
[
0.5205078125,
0.198974609375,
0.88330078125,
0.198974609375,
0.88330078125,
0.2120361328125,
0.5205078125,
0.2120361328125
],
[
0.118408203125,
0.2078857421875,
0.48779296875,
0.2078857421875,
0.48779296875,
0.221435546875,
0.118408203125,
0.221435546875
],
[
0.51318359375,
0.213134765625,
0.88330078125,
0.21435546875,
0.88330078125,
0.2286376953125,
0.51318359375,
0.22802734375
],
[
0.12005615234375,
0.2244873046875,
0.267822265625,
0.2244873046875,
0.267822265625,
0.2381591796875,
0.12005615234375,
0.2381591796875
],
[
0.515625,
0.23095703125,
0.88330078125,
0.23095703125,
0.88330078125,
0.24462890625,
0.515625,
0.24462890625
],
[
0.1376953125,
0.240478515625,
0.48779296875,
0.240478515625,
0.48779296875,
0.254150390625,
0.1376953125,
0.254150390625
],
[
0.515625,
0.2476806640625,
0.884765625,
0.2476806640625,
0.884765625,
0.26123046875,
0.515625,
0.26123046875
],
[
0.12005615234375,
0.255859375,
0.490234375,
0.255859375,
0.490234375,
0.26953125,
0.12005615234375,
0.26953125
],
[
0.51318359375,
0.262451171875,
0.8818359375,
0.26318359375,
0.8818359375,
0.27783203125,
0.51318359375,
0.276611328125
],
[
0.12005615234375,
0.2724609375,
0.48779296875,
0.2724609375,
0.48779296875,
0.2861328125,
0.12005615234375,
0.2861328125
],
[
0.5146484375,
0.279052734375,
0.880859375,
0.279052734375,
0.880859375,
0.292236328125,
0.5146484375,
0.292236328125
],
[
0.118408203125,
0.287353515625,
0.486083984375,
0.2880859375,
0.486083984375,
0.302734375,
0.118408203125,
0.3017578125
],
[
0.51318359375,
0.29345703125,
0.56494140625,
0.295654296875,
0.5634765625,
0.31005859375,
0.51123046875,
0.30810546875
],
[
0.12005615234375,
0.3046875,
0.33154296875,
0.3046875,
0.33154296875,
0.318359375,
0.12005615234375,
0.318359375
],
[
0.5205078125,
0.319580078125,
0.88330078125,
0.320556640625,
0.88330078125,
0.3349609375,
0.5205078125,
0.334228515625
],
[
0.2376708984375,
0.33203125,
0.291259765625,
0.33203125,
0.291259765625,
0.3466796875,
0.2376708984375,
0.3466796875
],
[
0.21826171875,
0.34033203125,
0.235107421875,
0.34033203125,
0.235107421875,
0.349853515625,
0.21826171875,
0.349853515625
],
[
0.51318359375,
0.3349609375,
0.88330078125,
0.336181640625,
0.88330078125,
0.350830078125,
0.51318359375,
0.349853515625
],
[
0.2301025390625,
0.344482421875,
0.2392578125,
0.349853515625,
0.2359619140625,
0.352783203125,
0.2275390625,
0.347412109375
],
[
0.2359619140625,
0.34326171875,
0.432373046875,
0.347412109375,
0.432373046875,
0.366455078125,
0.2359619140625,
0.3623046875
],
[
0.462646484375,
0.3466796875,
0.489501953125,
0.349853515625,
0.484375,
0.363525390625,
0.459228515625,
0.359375
],
[
0.21240234375,
0.352783203125,
0.297119140625,
0.3623046875,
0.29296875,
0.38134765625,
0.208251953125,
0.371826171875
],
[
0.1502685546875,
0.3466796875,
0.197265625,
0.3486328125,
0.1956787109375,
0.363525390625,
0.1485595703125,
0.361083984375
],
[
0.515625,
0.35400390625,
0.59765625,
0.35400390625,
0.59765625,
0.365234375,
0.515625,
0.365234375
],
[
0.5146484375,
0.369384765625,
0.88330078125,
0.369384765625,
0.88330078125,
0.383056640625,
0.5146484375,
0.383056640625
],
[
0.118408203125,
0.39013671875,
0.48779296875,
0.391357421875,
0.48779296875,
0.40625,
0.118408203125,
0.405029296875
],
[
0.515625,
0.385986328125,
0.8818359375,
0.385986328125,
0.8818359375,
0.399658203125,
0.515625,
0.399658203125
],
[
0.51318359375,
0.402099609375,
0.71533203125,
0.402099609375,
0.71533203125,
0.4150390625,
0.51318359375,
0.4150390625
],
[
0.12005615234375,
0.410400390625,
0.48779296875,
0.410400390625,
0.48779296875,
0.42333984375,
0.12005615234375,
0.42333984375
],
[
0.5146484375,
0.418701171875,
0.88330078125,
0.418701171875,
0.88330078125,
0.431640625,
0.5146484375,
0.431640625
],
[
0.12005615234375,
0.42578125,
0.48779296875,
0.42578125,
0.48779296875,
0.439453125,
0.12005615234375,
0.439453125
],
[
0.5146484375,
0.435791015625,
0.8818359375,
0.435791015625,
0.8818359375,
0.446533203125,
0.5146484375,
0.446533203125
],
[
0.12005615234375,
0.4423828125,
0.48779296875,
0.4423828125,
0.48779296875,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.5146484375,
0.45068359375,
0.7001953125,
0.45068359375,
0.7001953125,
0.46435546875,
0.5146484375,
0.46435546875
],
[
0.118408203125,
0.457763671875,
0.486083984375,
0.457763671875,
0.486083984375,
0.471435546875,
0.118408203125,
0.471435546875
],
[
0.515625,
0.46728515625,
0.88330078125,
0.46728515625,
0.88330078125,
0.48095703125,
0.515625,
0.48095703125
],
[
0.12176513671875,
0.4755859375,
0.210693359375,
0.4755859375,
0.210693359375,
0.48681640625,
0.12176513671875,
0.48681640625
],
[
0.515625,
0.482666015625,
0.86376953125,
0.482666015625,
0.86376953125,
0.496337890625,
0.515625,
0.496337890625
],
[
0.1502685546875,
0.4892578125,
0.48779296875,
0.490478515625,
0.48779296875,
0.5048828125,
0.1502685546875,
0.50439453125
],
[
0.51318359375,
0.498779296875,
0.8818359375,
0.49951171875,
0.8818359375,
0.51318359375,
0.51318359375,
0.5126953125
],
[
0.118408203125,
0.50732421875,
0.48779296875,
0.50732421875,
0.48779296875,
0.52099609375,
0.118408203125,
0.52099609375
],
[
0.515625,
0.515625,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.515625,
0.529296875
],
[
0.1175537109375,
0.5224609375,
0.378662109375,
0.5224609375,
0.378662109375,
0.5361328125,
0.1175537109375,
0.5361328125
],
[
0.5146484375,
0.53076171875,
0.8818359375,
0.52978515625,
0.8818359375,
0.54443359375,
0.5146484375,
0.5458984375
],
[
0.118408203125,
0.5517578125,
0.151123046875,
0.5517578125,
0.151123046875,
0.56396484375,
0.118408203125,
0.56396484375
],
[
0.1561279296875,
0.552734375,
0.264404296875,
0.552734375,
0.264404296875,
0.56396484375,
0.1561279296875,
0.56396484375
],
[
0.5146484375,
0.54736328125,
0.66162109375,
0.54736328125,
0.66162109375,
0.56103515625,
0.5146484375,
0.56103515625
],
[
0.5146484375,
0.5634765625,
0.88818359375,
0.5634765625,
0.88818359375,
0.57763671875,
0.5146484375,
0.57763671875
],
[
0.118408203125,
0.576171875,
0.17041015625,
0.576171875,
0.17041015625,
0.59033203125,
0.118408203125,
0.59033203125
],
[
0.162109375,
0.57470703125,
0.235107421875,
0.57763671875,
0.2333984375,
0.5927734375,
0.160400390625,
0.5888671875
],
[
0.51318359375,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59326171875,
0.51318359375,
0.59326171875
],
[
0.130126953125,
0.59716796875,
0.489501953125,
0.59716796875,
0.489501953125,
0.6123046875,
0.130126953125,
0.6123046875
],
[
0.51318359375,
0.59423828125,
0.81787109375,
0.595703125,
0.81787109375,
0.60986328125,
0.51318359375,
0.609375
],
[
0.12005615234375,
0.6142578125,
0.489501953125,
0.6142578125,
0.489501953125,
0.62744140625,
0.12005615234375,
0.62744140625
],
[
0.51318359375,
0.6123046875,
0.8818359375,
0.6123046875,
0.8818359375,
0.6259765625,
0.51318359375,
0.6259765625
],
[
0.12005615234375,
0.6298828125,
0.489501953125,
0.6298828125,
0.489501953125,
0.64453125,
0.12005615234375,
0.64453125
],
[
0.51318359375,
0.626953125,
0.88330078125,
0.62744140625,
0.88330078125,
0.642578125,
0.51318359375,
0.64111328125
],
[
0.118408203125,
0.64453125,
0.486083984375,
0.64453125,
0.486083984375,
0.6591796875,
0.118408203125,
0.6591796875
],
[
0.5146484375,
0.64453125,
0.6630859375,
0.64453125,
0.6630859375,
0.65771484375,
0.5146484375,
0.65771484375
],
[
0.12005615234375,
0.662109375,
0.48779296875,
0.662109375,
0.48779296875,
0.6767578125,
0.12005615234375,
0.6767578125
],
[
0.51318359375,
0.6611328125,
0.8818359375,
0.6611328125,
0.8818359375,
0.6748046875,
0.51318359375,
0.6748046875
],
[
0.118408203125,
0.6767578125,
0.486083984375,
0.6767578125,
0.486083984375,
0.6904296875,
0.118408203125,
0.6904296875
],
[
0.51123046875,
0.67578125,
0.8818359375,
0.677734375,
0.8818359375,
0.69140625,
0.51123046875,
0.689453125
],
[
0.12005615234375,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.12005615234375,
0.7080078125
],
[
0.5146484375,
0.693359375,
0.8271484375,
0.693359375,
0.8271484375,
0.70703125,
0.5146484375,
0.70703125
],
[
0.12176513671875,
0.7119140625,
0.489501953125,
0.7119140625,
0.489501953125,
0.72265625,
0.12176513671875,
0.72265625
],
[
0.5146484375,
0.71044921875,
0.8818359375,
0.71044921875,
0.8818359375,
0.7236328125,
0.5146484375,
0.7236328125
],
[
0.12005615234375,
0.72705078125,
0.418212890625,
0.72705078125,
0.418212890625,
0.74072265625,
0.12005615234375,
0.74072265625
],
[
0.51904296875,
0.72802734375,
0.880859375,
0.72802734375,
0.880859375,
0.73828125,
0.51904296875,
0.73828125
],
[
0.515625,
0.7421875,
0.8818359375,
0.7421875,
0.8818359375,
0.755859375,
0.515625,
0.755859375
],
[
0.1259765625,
0.751953125,
0.489501953125,
0.75048828125,
0.489501953125,
0.765625,
0.1259765625,
0.7666015625
],
[
0.51318359375,
0.75830078125,
0.8818359375,
0.75732421875,
0.8818359375,
0.77099609375,
0.51318359375,
0.771484375
],
[
0.1175537109375,
0.76708984375,
0.484375,
0.76708984375,
0.484375,
0.78076171875,
0.1175537109375,
0.78076171875
],
[
0.5146484375,
0.77490234375,
0.6650390625,
0.77490234375,
0.6650390625,
0.7880859375,
0.5146484375,
0.7880859375
],
[
0.12176513671875,
0.783203125,
0.48779296875,
0.783203125,
0.48779296875,
0.79638671875,
0.12176513671875,
0.79638671875
],
[
0.51318359375,
0.7890625,
0.88330078125,
0.79052734375,
0.88330078125,
0.8046875,
0.51318359375,
0.80419921875
],
[
0.1368408203125,
0.80712890625,
0.48779296875,
0.80859375,
0.48779296875,
0.822265625,
0.1368408203125,
0.82080078125
],
[
0.5146484375,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.5146484375,
0.82080078125
],
[
0.118408203125,
0.8193359375,
0.48779296875,
0.82080078125,
0.48779296875,
0.83447265625,
0.118408203125,
0.8330078125
],
[
0.5146484375,
0.822265625,
0.880859375,
0.822265625,
0.880859375,
0.8359375,
0.5146484375,
0.8359375
],
[
0.12176513671875,
0.8349609375,
0.1376953125,
0.8349609375,
0.1376953125,
0.84326171875,
0.12176513671875,
0.84326171875
],
[
0.515625,
0.8388671875,
0.73291015625,
0.8388671875,
0.73291015625,
0.8525390625,
0.515625,
0.8525390625
],
[
0.139404296875,
0.8466796875,
0.444091796875,
0.8466796875,
0.444091796875,
0.85986328125,
0.139404296875,
0.85986328125
],
[
0.12005615234375,
0.85986328125,
0.225830078125,
0.85986328125,
0.225830078125,
0.87060546875,
0.12005615234375,
0.87060546875
],
[
0.5322265625,
0.85546875,
0.88330078125,
0.85546875,
0.88330078125,
0.869140625,
0.5322265625,
0.869140625
],
[
0.139404296875,
0.87060546875,
0.489501953125,
0.87060546875,
0.489501953125,
0.88525390625,
0.139404296875,
0.88525390625
],
[
0.51318359375,
0.87158203125,
0.8818359375,
0.87158203125,
0.8818359375,
0.88525390625,
0.51318359375,
0.88525390625
],
[
0.12005615234375,
0.88427734375,
0.154541015625,
0.88427734375,
0.154541015625,
0.8955078125,
0.12005615234375,
0.8955078125
],
[
0.139404296875,
0.896484375,
0.39208984375,
0.896484375,
0.39208984375,
0.91015625,
0.139404296875,
0.91015625
],
[
0.5322265625,
0.89453125,
0.8251953125,
0.896484375,
0.8251953125,
0.9111328125,
0.5322265625,
0.9091796875
]
] |
[
"DatasetTrain",
"Validation",
"Test",
"们使用36K令牌的批量大小训练模型，并将",
"de-en",
"1.9M",
"994",
"508",
"fr-en",
"1.1M",
"447",
"Transformer和AET的最大更新分别设置为",
"0.5M",
"1,000",
"999",
"50K和10K。AET的最后一个检查点用于",
"248",
"ro-en",
"评估。所有模型都在两个翻译方向上进行训",
"Table 1: Number of sentences in each dataset.",
"练，并使用Zenkeletal.（2019）中的脚本使用",
"grow-diag(Koehnetal.，2005）进行对称化。",
"其中 GK,G Rdmodelxdk是 n 的关键和查询",
"投影矩阵注意第XMATHX。由于我们只关",
"Evaluation我们使用对齐错误率（Ochand",
"心注意力权重，因此在此模块中省略了与价值",
"Ney，2000，AER）评估我们方法的对齐质量。",
"相关的参数和计算",
"由于单词对齐对于1节中讨论的许多下游任",
"为了训练对齐模块，我们使用从香草变压器",
"务都很有用，因此我们还使用度量BLEUE",
"模型中提取的对称化Shift-Att对齐作为标签。",
"(Papinenietal.，2002）评估了字典指导解码的",
"具体来说，当底层变压器是预先训练和固定的",
"方法，这是对齐归纳的下游任务。更多细节见",
"（图。E2），我们使用EGargetal.（2019）以下",
"4.3节",
"的损失函数训练对齐模块：",
"Baselines我们将我们的方法与两个统计基",
"lyl lal",
"1",
"线Fast-Align和GIZA++以及其他九个基线",
"",
"ZZ(AP, o log Si.j),",
"(6)",
"13台会",
"Ca =",
"进行比较：",
".Naive-AttE(Garg et al.,2019)：我们在2.2",
"其中S={Si;.;Sul是对齐模块预测的对",
"节中讨论的方法，该方法从变压器倒数第二层",
"的注意力权重中引起对齐",
"齐分数矩阵，AP表示归一化的参考对称化",
"·Naive-Att-LA（Gargetal.,2019)：没有层",
"Shift-Att对齐。2通过这种方式，我们将在两",
"选择的Naive-Att方法。它从所有层的平均注",
"个vanillaTransformer模型Qa→y和y→a中",
"意力权重中诱导出对齐",
"隐式学习的对齐知识转移到单个AET模型的",
"·Shift-Att-LA：没有层选择的Shift-Att方法。",
"对齐模块中",
"它从所有层的平均注意力权重中诱导出对齐",
"旦对对齐模块进行了训练，我们就可以在",
"·SmoothGradE(Lietal.,2016）：从单词显",
"给定平行句子对的情况下从中提取对齐分数",
"着性诱导对齐的方法，该方法是通过以多个噜",
"S，并使用公式E3诱导对齐A",
"杂的句子对作为输人对基于梯度的显着性得分",
"4",
"Experiments",
"进行平均来计算的",
". SD-SmoothGrad E(Ding et al., 2019) :",
"4.1",
"Settings",
"SmoothGrad的改进版本，它在单热输人向",
"Dataset 我们遵循先前的工作E(Zenkel",
"量上定义了显着性，而不是单词嵌入。",
"etal，2019,2020）进行数据设置，并对德语-英",
"·PD（Lietal.,2019）：通过迭代地屏蔽每",
"语（de-en）3，罗马尼亚语-英语（ro-en）和法",
"个源令牌并测量预测差异来计算Transformer",
"语-英语（fr-en）4的公开可用数据集。由于未",
"的对齐分数的方法",
"提供验证集，因此我们按照Dingetal.（2019)",
"·AddSGD E(Zenkel et al., 2019):在 Trans",
"将预处理前训练数据的最后1.000个句子设",
"former顶部显式添加额外关注层并直接优化",
"置为验证集。我们学习了具有10k合并操作",
"其激活以预测要对齐的目标令牌的方法",
"的联合源和目标字节对编码E（Sennrichetal.",
"·Mtl-FullcE(Gargetal.,2019）：在多任务学",
"2016）。表1显示了详细的数据统计",
"习框架中训练单个模型以同时预测自标句子和",
"对齐方式的方法。在预测对齐方式时，模型观",
"NMTSystems我们使用fairseq-py5实现",
"察完整的目标句子，并使用对称的Naive-Att",
"Transformer，并使用EDing et al.（2019）之",
"对齐方式作为标签",
"后的transformeriwsltdeen模型配置。我",
".Mtl-Fullc-GZ(Garg et al., 2019):与Mtl-",
"2we simply normalize rows corresponding to target",
"Fullc相同的方法，除了使用对称的GIZA++",
"tokens that are aligned to at least one source token of",
"比对作为标签。这是一种统计和神经方法，因",
"A",
"为它依赖于GIZA++比对",
"3https://www-i6.informatik.rwth-aachen.de/",
"goldAlignment/",
"在这九个基线和我们提出的方法中",
"4http://web.eecs.umich.edu/~mihalcea/wpt/index.",
"SmoothGrad，SD-SmoothGrad 和 PD使用",
"html",
"\"https://github.com/pytorch/fairseq",
"6https://github.com/lilt/alignment-scripts"
] |
{
"arxiv_id": "2004.14837",
"language": "zh",
"page_idx": 3
}
|
||
2004.14837-zh-4
|
2004.14837-zh-4.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th rowspan="2">Inter. </th>
<th rowspan="2">Fullc</th>
<th colspan="3">$de-en$</th>
<th> </th>
<th colspan="2">fr- en</th>
<th colspan="3">ro-en</th>
</tr>
<tr>
<th>de$\to$en</th>
<th>$\to$de en</th>
<th>bidir</th>
<th>fr$\to \epsilon$ en</th>
<th>$\to$fr en </th>
<th>bidir</th>
<th>ro $\rightarrow$ en</th>
<th>en $\rightarrow$ ro</th>
<th>bidi</th>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>Stati</td>
<td>${\text{tical Met}\mathrm{P}}$</td>
<td>hods</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Fast- Align TE (Dver et al.. 2013</td>
<td> </td>
<td>$\overline{\mathrm{Y}}$</td>
<td>28.5</td>
<td>30.4</td>
<td>25.7</td>
<td>16.3</td>
<td>17.1</td>
<td>12.1</td>
<td>33.6</td>
<td>36.8</td>
<td>31.</td>
</tr>
<tr>
<td>GIZA+ + Brown et al.. 1993</td>
<td> </td>
<td>Y</td>
<td>18.8</td>
<td>19.6</td>
<td>17.8</td>
<td>7.1</td>
<td>7.2</td>
<td>6.1</td>
<td>27.4</td>
<td>28.7</td>
<td>26.(</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>Net</td>
<td>$\iota ral$ Metho</td>
<td>$)ds$</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Naive- Att Garg et al. , 2019</td>
<td>$Y$</td>
<td>N</td>
<td>33.3</td>
<td>36.5</td>
<td>28.1</td>
<td>27.5</td>
<td>23.6</td>
<td>16.0</td>
<td>33.6</td>
<td>35.1</td>
<td>$\overline{30.C}$</td>
</tr>
<tr>
<td>Naive- Att- LA $^{\prime}$Garg et al..2019</td>
<td>$Y$</td>
<td>$N$</td>
<td>40.9</td>
<td>50.8</td>
<td>39.8</td>
<td>32.4</td>
<td>29.8</td>
<td>21.2</td>
<td>37.5</td>
<td>35.5</td>
<td>32.7</td>
</tr>
<tr>
<td>Shift-Att-LA</td>
<td>$Y$</td>
<td>$N$</td>
<td>54.7</td>
<td>46.2</td>
<td>45.5</td>
<td>60.5</td>
<td>46.9</td>
<td>55.1</td>
<td>66.1</td>
<td>60.4</td>
<td>65.</td>
</tr>
<tr>
<td>SmoothGrad (Li et al., 2016</td>
<td>$Y$</td>
<td>$N$</td>
<td>36.4</td>
<td>45.8</td>
<td>30.3</td>
<td>25.5</td>
<td>27.0</td>
<td>15.6</td>
<td>41.3</td>
<td>39.9</td>
<td>33.7</td>
</tr>
<tr>
<td>SD-SmoothGrad 1 (Ding et al., 2019</td>
<td></td>
<td>$N$</td>
<td>36.4</td>
<td>43.0</td>
<td>29.0</td>
<td>25.9</td>
<td>29.7</td>
<td>15.3</td>
<td>41.2</td>
<td>41.4</td>
<td>32.7</td>
</tr>
<tr>
<td>$PD$ (Li et al., 2019</td>
<td>Y</td>
<td>$N$</td>
<td>38.1</td>
<td>44.8</td>
<td>34.4</td>
<td>32.4</td>
<td>31.1</td>
<td>23.1</td>
<td>40.2</td>
<td>40.8</td>
<td>35.1 1 </td>
</tr>
<tr>
<td>AddSGD Zenkel et al. , 2019</td>
<td>N</td>
<td>$N$</td>
<td>26.6</td>
<td>30.4</td>
<td>21.2</td>
<td>20.5</td>
<td>23.8</td>
<td>10.0</td>
<td>32.3</td>
<td>34.8</td>
<td>27.1 1</td>
</tr>
<tr>
<td>Mtl-Fullc Garg et al. , 2019</td>
<td>N</td>
<td>$Y$</td>
<td> </td>
<td> </td>
<td>20.2</td>
<td> </td>
<td> </td>
<td>7.7</td>
<td> </td>
<td> </td>
<td>26.(</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>Statistical</td>
<td>Neural 十</td>
<td>Metho</td>
<td>$\overline{\mathrm{ds}}$</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Mtl-Fullc-GZ (Garg et al., 2019</td>
<td>N</td>
<td>$Y$</td>
<td> </td>
<td> </td>
<td>16.0</td>
<td> </td>
<td> </td>
<td>4.6</td>
<td> </td>
<td> </td>
<td>$\overline{23.1}$</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>Our $\Gamma$</td>
<td>eural Met</td>
<td>hods</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Shift-Att</td>
<td>$Y$</td>
<td>$\overline{N}$</td>
<td>20.9</td>
<td>25.7</td>
<td>17.9</td>
<td>17.1</td>
<td>16.1</td>
<td>6.6</td>
<td>27.4</td>
<td>26.0</td>
<td>$\overline{23.1}$</td>
</tr>
<tr>
<td>Shift-AET</td>
<td>N</td>
<td>N</td>
<td>15.8</td>
<td>19.2</td>
<td>15.4</td>
<td>9.9</td>
<td>10.5</td>
<td>4.7</td>
<td>22.7</td>
<td>23.6</td>
<td>21.</td>
</tr>
</tbody>
</table>
Table 2: AER on the test set with different alignment methods. bidir are symmetrized alignment results. The column Inter. represents whether the method is an interpretation method that can extract alignments from a pretrained vanilla Transformer model. The column Fullc denotes whether full target sentence is used to extract alignments at test time. The lower AER, the better. We mark best symmetrized interpretation results of vanilla Transformer with underlines, and best symmetrized results among all with boldface.
<table>
<tbody>
<tr>
<th>Method</th>
<th>$\overline{\mathrm{de-en}}$</th>
<th>$\overline{\mathrm{fr-en}}$</th>
<th>ro-en</th>
</tr>
<tr>
<td>BAO- Guided</td>
<td>16.3</td>
<td>$\overline{5.0}$</td>
<td>$\overline{23.4}$</td>
</tr>
<tr>
<td>Shift-AET</td>
<td>15.4</td>
<td>4.7</td>
<td>21.2</td>
</tr>
</tbody>
</table>
特征重要性度量来诱导对齐，而其他来自某种形式的注意力权重。请注意，具有特征重要性度量的方法的计算成本远高于具有注意力权重的方法。7
Table 3: Comparison of our method Shift-AET
with BAO-Guided
(Zenkel et al., 2020).We
report the symmetrized AER on the test set.
## 4.2 Alignment Results
Comparison with Baselines 表 2 将我们的方法与所有基线进行比较。首先，Shift-Att 是一种用于香草变压器的纯解释方法，其性能明显优于 Fast-Align 和所有神经基线，并且与GIZA++的性能相当。例如，它在不同语言对上的 8.7- 11.1 AER 点优于 SD-SmoothGrad , 这是一种具有特征重要性度量的最新方法可以从香草转换器中提取对齐。Shift-Att 的成功表明，vanillaTransformer 以隐含的方式捕获了对齐信息，如果选择了正确的解码步骤和层来诱导对齐，则可以从注意力权重中揭示
其次，方法 Shift-AET 实现了新的最新技术，明显优于所有基线。它通过 1.4-4.8 AER 在不同的语言对上对 GIZA++ 进行了改进。表明可以构建比 GIZA++ 更好的神经对准器而无需使用从统计对准器生成的任何对准来引导训练。与 MTL-Fullc-GZ 相比，我们还发现Shift-AET 的性能稍好 (de-en 和 ro-en)或同等 (fr-en),这是一种使用 GIZA++ 对齐来引导训练的方法。我们评估模型尺寸：香草变
${\overline{{^7}\text{For each s}}}$entence pair, PD forwards once with
压器和 AET 中的参数数量为 36。8M 和 37。3M,并发现 AET 仅引入 1。4% 附加参数到香草变压器。总之，通过使用对称化的 Shift-Att 对齐来监督对齐模块，Shift-AET 比 Shift-Att 和 GIZA++ 有所改进，参数增加可忽略不计， 并且不会影响翻译质量
Comparison with Zenkel et al. (2020) 与我们的工作同时，Zenkel et al. (2020) 提出了一种可以超越 GIZA++ 的神经对准器。表 3 比较了 Shift-AET 和最佳方法 BAO-Guided 的性能 (Birdir。Att。Opt。+引导)在 Zenkel et al.(2020) 中。我们观察到 Shift-AET 在对准精度方面比 BAO-Guided 表现更好
Shift-AET 也比 BAO-Guided 简单得多。BAO-Guided 的训练包括三个阶段：(???)在源到目标和目标到源的方向上训练香草变压器；(???)训练对齐层，并通过双向注意力优化在训练集上提取对齐。此对齐提取过程的计算成本很高，因为双向注意力优化会针对训练集中的每个句子对分别微调模型参数；(???) 以提取的对齐方式为指导，重新训练对齐层。相比之下，Shift-AET 可以在两个阶段进行更快的训练，并且不涉及双向注意力优化
与 Mtl-Fullc $\boxed{\mathrm{E}}($Garg et al.,2019)类似，
$|\boldsymbol{x}|+1$ masked sentence pairs as the input, while SmoothGrad and SD-SmoothGrad forward and backward once with $m(\begin{array}{cc}m&=&30&\mathrm{in}&\mathrm{Ding~et~al.~(2019)}\end{array}$ ) noisy sentence pairs as the input. In contrast, attention weights based methods forward once with one sentence pair as the input.
|
[
[
0.465087890625,
0.0736083984375,
0.497802734375,
0.0736083984375,
0.497802734375,
0.08489990234375,
0.465087890625,
0.08489990234375
],
[
0.615234375,
0.07244873046875,
0.6455078125,
0.07244873046875,
0.6455078125,
0.08489990234375,
0.615234375,
0.08489990234375
],
[
0.1275634765625,
0.0789794921875,
0.175537109375,
0.0789794921875,
0.175537109375,
0.09027099609375,
0.1275634765625,
0.09027099609375
],
[
0.76318359375,
0.0736083984375,
0.79443359375,
0.0736083984375,
0.79443359375,
0.08489990234375,
0.76318359375,
0.08489990234375
],
[
0.330078125,
0.08013916015625,
0.364501953125,
0.08013916015625,
0.364501953125,
0.0885009765625,
0.330078125,
0.0885009765625
],
[
0.366943359375,
0.08013916015625,
0.399658203125,
0.08013916015625,
0.399658203125,
0.0885009765625,
0.366943359375,
0.0885009765625
],
[
0.51708984375,
0.0831298828125,
0.5556640625,
0.0831298828125,
0.5556640625,
0.09442138671875,
0.51708984375,
0.09442138671875
],
[
0.61376953125,
0.0831298828125,
0.6630859375,
0.0831298828125,
0.6630859375,
0.09442138671875,
0.61376953125,
0.09442138671875
],
[
0.66162109375,
0.0831298828125,
0.7021484375,
0.0831298828125,
0.7021484375,
0.09442138671875,
0.66162109375,
0.09442138671875
],
[
0.8134765625,
0.0831298828125,
0.8505859375,
0.0831298828125,
0.8505859375,
0.09442138671875,
0.8134765625,
0.09442138671875
],
[
0.465087890625,
0.0843505859375,
0.52294921875,
0.0843505859375,
0.52294921875,
0.09442138671875,
0.465087890625,
0.09442138671875
],
[
0.4375,
0.08489990234375,
0.459228515625,
0.08489990234375,
0.459228515625,
0.09320068359375,
0.4375,
0.09320068359375
],
[
0.580078125,
0.08489990234375,
0.60791015625,
0.08489990234375,
0.60791015625,
0.09320068359375,
0.580078125,
0.09320068359375
],
[
0.7080078125,
0.08489990234375,
0.724609375,
0.08489990234375,
0.724609375,
0.09320068359375,
0.7080078125,
0.09320068359375
],
[
0.72314453125,
0.08489990234375,
0.75732421875,
0.08489990234375,
0.75732421875,
0.09320068359375,
0.72314453125,
0.09320068359375
],
[
0.76416015625,
0.08489990234375,
0.7841796875,
0.08489990234375,
0.7841796875,
0.09320068359375,
0.76416015625,
0.09320068359375
],
[
0.5615234375,
0.08612060546875,
0.57080078125,
0.08612060546875,
0.57080078125,
0.0914306640625,
0.5615234375,
0.0914306640625
],
[
0.411376953125,
0.08612060546875,
0.4248046875,
0.08612060546875,
0.4248046875,
0.0926513671875,
0.411376953125,
0.0926513671875
],
[
0.78271484375,
0.08612060546875,
0.81201171875,
0.08612060546875,
0.81201171875,
0.09320068359375,
0.78271484375,
0.09320068359375
],
[
0.43408203125,
0.097412109375,
0.54150390625,
0.097412109375,
0.54150390625,
0.1092529296875,
0.43408203125,
0.1092529296875
],
[
0.1234130859375,
0.1051025390625,
0.287109375,
0.105712890625,
0.287109375,
0.119384765625,
0.1234130859375,
0.1181640625
],
[
0.63818359375,
0.105712890625,
0.6650390625,
0.1092529296875,
0.66162109375,
0.12054443359375,
0.634765625,
0.1175537109375
],
[
0.72900390625,
0.105712890625,
0.75830078125,
0.1092529296875,
0.755859375,
0.119384765625,
0.7255859375,
0.11639404296875
],
[
0.78515625,
0.10687255859375,
0.8134765625,
0.10687255859375,
0.8134765625,
0.1181640625,
0.78515625,
0.1181640625
],
[
0.82275390625,
0.10687255859375,
0.85205078125,
0.10687255859375,
0.85205078125,
0.1181640625,
0.82275390625,
0.1181640625
],
[
0.5810546875,
0.10687255859375,
0.609375,
0.10687255859375,
0.609375,
0.119384765625,
0.5810546875,
0.119384765625
],
[
0.43408203125,
0.1092529296875,
0.45751953125,
0.1092529296875,
0.45751953125,
0.1175537109375,
0.43408203125,
0.1175537109375
],
[
0.491943359375,
0.1092529296875,
0.51708984375,
0.1092529296875,
0.51708984375,
0.1175537109375,
0.491943359375,
0.1175537109375
],
[
0.5283203125,
0.1092529296875,
0.55322265625,
0.1092529296875,
0.55322265625,
0.1175537109375,
0.5283203125,
0.1175537109375
],
[
0.67529296875,
0.1092529296875,
0.7001953125,
0.1092529296875,
0.7001953125,
0.1175537109375,
0.67529296875,
0.1175537109375
],
[
0.432373046875,
0.1175537109375,
0.459228515625,
0.1175537109375,
0.459228515625,
0.1300048828125,
0.432373046875,
0.1300048828125
],
[
0.39208984375,
0.10986328125,
0.403076171875,
0.10986328125,
0.403076171875,
0.1175537109375,
0.39208984375,
0.1175537109375
],
[
0.490234375,
0.1175537109375,
0.51904296875,
0.1175537109375,
0.51904296875,
0.12890625,
0.490234375,
0.12890625
],
[
0.5263671875,
0.1175537109375,
0.5556640625,
0.1175537109375,
0.5556640625,
0.12890625,
0.5263671875,
0.12890625
],
[
0.5859375,
0.1175537109375,
0.609375,
0.1175537109375,
0.609375,
0.1300048828125,
0.5859375,
0.1300048828125
],
[
0.6767578125,
0.1175537109375,
0.7021484375,
0.1175537109375,
0.7021484375,
0.12890625,
0.6767578125,
0.12890625
],
[
0.72900390625,
0.1175537109375,
0.75830078125,
0.1175537109375,
0.75830078125,
0.1300048828125,
0.72900390625,
0.1300048828125
],
[
0.7841796875,
0.1175537109375,
0.8134765625,
0.1175537109375,
0.8134765625,
0.1300048828125,
0.7841796875,
0.1300048828125
],
[
0.3896484375,
0.1181640625,
0.403076171875,
0.1181640625,
0.403076171875,
0.127685546875,
0.3896484375,
0.127685546875
],
[
0.1292724609375,
0.1181640625,
0.28955078125,
0.1181640625,
0.28955078125,
0.12890625,
0.1292724609375,
0.12890625
],
[
0.640625,
0.1181640625,
0.6630859375,
0.1181640625,
0.6630859375,
0.12890625,
0.640625,
0.12890625
],
[
0.82275390625,
0.1181640625,
0.85205078125,
0.1181640625,
0.85205078125,
0.1300048828125,
0.82275390625,
0.1300048828125
],
[
0.443359375,
0.12890625,
0.53076171875,
0.12890625,
0.53076171875,
0.1395263671875,
0.443359375,
0.1395263671875
],
[
0.5810546875,
0.1395263671875,
0.609375,
0.1395263671875,
0.609375,
0.15087890625,
0.5810546875,
0.15087890625
],
[
0.78515625,
0.1395263671875,
0.81005859375,
0.1395263671875,
0.81005859375,
0.15087890625,
0.78515625,
0.15087890625
],
[
0.82275390625,
0.1395263671875,
0.85205078125,
0.1395263671875,
0.85205078125,
0.15087890625,
0.82275390625,
0.15087890625
],
[
0.432373046875,
0.14013671875,
0.459228515625,
0.14013671875,
0.459228515625,
0.1519775390625,
0.432373046875,
0.1519775390625
],
[
0.1259765625,
0.14013671875,
0.285400390625,
0.14013671875,
0.285400390625,
0.15087890625,
0.1259765625,
0.15087890625
],
[
0.355224609375,
0.141357421875,
0.36279296875,
0.146728515625,
0.35693359375,
0.1519775390625,
0.34765625,
0.1455078125
],
[
0.3896484375,
0.141357421875,
0.4013671875,
0.141357421875,
0.4013671875,
0.15087890625,
0.3896484375,
0.15087890625
],
[
0.491943359375,
0.141357421875,
0.51708984375,
0.141357421875,
0.51708984375,
0.149658203125,
0.491943359375,
0.149658203125
],
[
0.5263671875,
0.141357421875,
0.55322265625,
0.141357421875,
0.55322265625,
0.149658203125,
0.5263671875,
0.149658203125
],
[
0.63818359375,
0.141357421875,
0.6630859375,
0.141357421875,
0.6630859375,
0.149658203125,
0.63818359375,
0.149658203125
],
[
0.67529296875,
0.141357421875,
0.7001953125,
0.141357421875,
0.7001953125,
0.149658203125,
0.67529296875,
0.149658203125
],
[
0.72900390625,
0.14013671875,
0.75732421875,
0.14013671875,
0.75732421875,
0.1519775390625,
0.72900390625,
0.1519775390625
],
[
0.432373046875,
0.1478271484375,
0.459228515625,
0.15087890625,
0.456787109375,
0.162109375,
0.429931640625,
0.1591796875
],
[
0.5810546875,
0.1478271484375,
0.609375,
0.15087890625,
0.60791015625,
0.162109375,
0.57861328125,
0.1591796875
],
[
0.7841796875,
0.1484375,
0.8134765625,
0.1519775390625,
0.81005859375,
0.162109375,
0.78076171875,
0.1591796875
],
[
0.490234375,
0.149658203125,
0.51904296875,
0.149658203125,
0.51904296875,
0.160888671875,
0.490234375,
0.160888671875
],
[
0.387939453125,
0.149658203125,
0.4013671875,
0.149658203125,
0.4013671875,
0.160888671875,
0.387939453125,
0.160888671875
],
[
0.5263671875,
0.149658203125,
0.55322265625,
0.149658203125,
0.55322265625,
0.160888671875,
0.5263671875,
0.160888671875
],
[
0.634765625,
0.149658203125,
0.6630859375,
0.149658203125,
0.6630859375,
0.160888671875,
0.634765625,
0.160888671875
],
[
0.67236328125,
0.149658203125,
0.7001953125,
0.149658203125,
0.7001953125,
0.160888671875,
0.67236328125,
0.160888671875
],
[
0.72900390625,
0.149658203125,
0.75732421875,
0.149658203125,
0.75732421875,
0.160888671875,
0.72900390625,
0.160888671875
],
[
0.349365234375,
0.15087890625,
0.36279296875,
0.15087890625,
0.36279296875,
0.160888671875,
0.349365234375,
0.160888671875
],
[
0.1292724609375,
0.15087890625,
0.306396484375,
0.15087890625,
0.306396484375,
0.160888671875,
0.1292724609375,
0.160888671875
],
[
0.82275390625,
0.15087890625,
0.8505859375,
0.15087890625,
0.8505859375,
0.162109375,
0.82275390625,
0.162109375
],
[
0.1275634765625,
0.1591796875,
0.2032470703125,
0.1602783203125,
0.2032470703125,
0.171630859375,
0.1275634765625,
0.17041015625
],
[
0.349365234375,
0.1602783203125,
0.36279296875,
0.1602783203125,
0.36279296875,
0.17041015625,
0.349365234375,
0.17041015625
],
[
0.5263671875,
0.1602783203125,
0.55517578125,
0.1602783203125,
0.55517578125,
0.1728515625,
0.5263671875,
0.1728515625
],
[
0.72900390625,
0.1602783203125,
0.75732421875,
0.1602783203125,
0.75732421875,
0.1728515625,
0.72900390625,
0.1728515625
],
[
0.4306640625,
0.160888671875,
0.459228515625,
0.160888671875,
0.459228515625,
0.1728515625,
0.4306640625,
0.1728515625
],
[
0.387939453125,
0.160888671875,
0.403076171875,
0.160888671875,
0.403076171875,
0.171630859375,
0.387939453125,
0.171630859375
],
[
0.490234375,
0.160888671875,
0.51904296875,
0.160888671875,
0.51904296875,
0.1728515625,
0.490234375,
0.1728515625
],
[
0.5810546875,
0.160888671875,
0.609375,
0.160888671875,
0.609375,
0.1728515625,
0.5810546875,
0.1728515625
],
[
0.634765625,
0.160888671875,
0.66650390625,
0.160888671875,
0.66650390625,
0.1728515625,
0.634765625,
0.1728515625
],
[
0.67333984375,
0.162109375,
0.69873046875,
0.162109375,
0.69873046875,
0.17041015625,
0.67333984375,
0.17041015625
],
[
0.78662109375,
0.162109375,
0.8134765625,
0.162109375,
0.8134765625,
0.17041015625,
0.78662109375,
0.17041015625
],
[
0.8212890625,
0.160888671875,
0.8505859375,
0.160888671875,
0.8505859375,
0.1728515625,
0.8212890625,
0.1728515625
],
[
0.349365234375,
0.17041015625,
0.36279296875,
0.17041015625,
0.36279296875,
0.1817626953125,
0.349365234375,
0.1817626953125
],
[
0.490234375,
0.169189453125,
0.5205078125,
0.1728515625,
0.51708984375,
0.182861328125,
0.48779296875,
0.179931640625
],
[
0.5263671875,
0.17041015625,
0.55517578125,
0.17041015625,
0.55517578125,
0.1817626953125,
0.5263671875,
0.1817626953125
],
[
0.5810546875,
0.17041015625,
0.609375,
0.17041015625,
0.609375,
0.182861328125,
0.5810546875,
0.182861328125
],
[
0.634765625,
0.17041015625,
0.6650390625,
0.17041015625,
0.6650390625,
0.1817626953125,
0.634765625,
0.1817626953125
],
[
0.67333984375,
0.17041015625,
0.7001953125,
0.17041015625,
0.7001953125,
0.1817626953125,
0.67333984375,
0.1817626953125
],
[
0.72900390625,
0.17041015625,
0.75732421875,
0.17041015625,
0.75732421875,
0.182861328125,
0.72900390625,
0.182861328125
],
[
0.78515625,
0.17041015625,
0.8154296875,
0.17041015625,
0.8154296875,
0.1817626953125,
0.78515625,
0.1817626953125
],
[
0.82275390625,
0.17041015625,
0.8505859375,
0.17041015625,
0.8505859375,
0.182861328125,
0.82275390625,
0.182861328125
],
[
0.432373046875,
0.171630859375,
0.459228515625,
0.171630859375,
0.459228515625,
0.182861328125,
0.432373046875,
0.182861328125
],
[
0.1292724609375,
0.171630859375,
0.28369140625,
0.171630859375,
0.28369140625,
0.1817626953125,
0.1292724609375,
0.1817626953125
],
[
0.3837890625,
0.1728515625,
0.39794921875,
0.168701171875,
0.405517578125,
0.179931640625,
0.391357421875,
0.18408203125
],
[
0.349365234375,
0.18115234375,
0.36279296875,
0.18115234375,
0.36279296875,
0.191162109375,
0.349365234375,
0.191162109375
],
[
0.432373046875,
0.18115234375,
0.459228515625,
0.18115234375,
0.459228515625,
0.193603515625,
0.432373046875,
0.193603515625
],
[
0.490234375,
0.18115234375,
0.51904296875,
0.18115234375,
0.51904296875,
0.193603515625,
0.490234375,
0.193603515625
],
[
0.52490234375,
0.18115234375,
0.55517578125,
0.18115234375,
0.55517578125,
0.193603515625,
0.52490234375,
0.193603515625
],
[
0.5810546875,
0.18115234375,
0.609375,
0.18115234375,
0.609375,
0.193603515625,
0.5810546875,
0.193603515625
],
[
0.634765625,
0.179931640625,
0.6650390625,
0.182861328125,
0.66162109375,
0.19482421875,
0.6337890625,
0.191162109375
],
[
0.67333984375,
0.18115234375,
0.7001953125,
0.18115234375,
0.7001953125,
0.193603515625,
0.67333984375,
0.193603515625
],
[
0.78515625,
0.18115234375,
0.8154296875,
0.18115234375,
0.8154296875,
0.193603515625,
0.78515625,
0.193603515625
],
[
0.387939453125,
0.1817626953125,
0.4013671875,
0.1817626953125,
0.4013671875,
0.1923828125,
0.387939453125,
0.1923828125
],
[
0.1292724609375,
0.1817626953125,
0.31982421875,
0.1817626953125,
0.31982421875,
0.1923828125,
0.1292724609375,
0.1923828125
],
[
0.72900390625,
0.1817626953125,
0.75732421875,
0.1817626953125,
0.75732421875,
0.193603515625,
0.72900390625,
0.193603515625
],
[
0.82275390625,
0.1817626953125,
0.8505859375,
0.1817626953125,
0.8505859375,
0.193603515625,
0.82275390625,
0.193603515625
],
[
0.355224609375,
0.189453125,
0.366943359375,
0.19775390625,
0.355224609375,
0.2060546875,
0.343505859375,
0.19775390625
],
[
0.5263671875,
0.191162109375,
0.55517578125,
0.191162109375,
0.55517578125,
0.2037353515625,
0.5263671875,
0.2037353515625
],
[
0.67333984375,
0.191162109375,
0.69873046875,
0.191162109375,
0.69873046875,
0.2037353515625,
0.67333984375,
0.2037353515625
],
[
0.73046875,
0.191162109375,
0.75732421875,
0.191162109375,
0.75732421875,
0.2037353515625,
0.73046875,
0.2037353515625
],
[
0.432373046875,
0.1923828125,
0.459228515625,
0.1923828125,
0.459228515625,
0.2037353515625,
0.432373046875,
0.2037353515625
],
[
0.1292724609375,
0.1923828125,
0.2359619140625,
0.1923828125,
0.2359619140625,
0.2037353515625,
0.1292724609375,
0.2037353515625
],
[
0.490234375,
0.1923828125,
0.51904296875,
0.1923828125,
0.51904296875,
0.2037353515625,
0.490234375,
0.2037353515625
],
[
0.5810546875,
0.1923828125,
0.60791015625,
0.1923828125,
0.60791015625,
0.2037353515625,
0.5810546875,
0.2037353515625
],
[
0.634765625,
0.1923828125,
0.6630859375,
0.1923828125,
0.6630859375,
0.2037353515625,
0.634765625,
0.2037353515625
],
[
0.78515625,
0.1923828125,
0.8154296875,
0.1923828125,
0.8154296875,
0.2037353515625,
0.78515625,
0.2037353515625
],
[
0.82275390625,
0.1923828125,
0.8505859375,
0.1923828125,
0.8505859375,
0.2037353515625,
0.82275390625,
0.2037353515625
],
[
0.634765625,
0.201904296875,
0.6650390625,
0.204833984375,
0.66162109375,
0.2161865234375,
0.6337890625,
0.213134765625
],
[
0.385498046875,
0.193603515625,
0.399658203125,
0.190673828125,
0.40478515625,
0.201904296875,
0.3896484375,
0.204833984375
],
[
0.1292724609375,
0.203125,
0.287109375,
0.201904296875,
0.287109375,
0.213134765625,
0.1292724609375,
0.21435546875
],
[
0.34765625,
0.203125,
0.364501953125,
0.203125,
0.364501953125,
0.223876953125,
0.34765625,
0.223876953125
],
[
0.4306640625,
0.203125,
0.459228515625,
0.203125,
0.459228515625,
0.215576171875,
0.4306640625,
0.215576171875
],
[
0.490234375,
0.203125,
0.5205078125,
0.203125,
0.5205078125,
0.21435546875,
0.490234375,
0.21435546875
],
[
0.5263671875,
0.203125,
0.55517578125,
0.203125,
0.55517578125,
0.21435546875,
0.5263671875,
0.21435546875
],
[
0.5810546875,
0.203125,
0.609375,
0.203125,
0.609375,
0.215576171875,
0.5810546875,
0.215576171875
],
[
0.67333984375,
0.203125,
0.7001953125,
0.203125,
0.7001953125,
0.21435546875,
0.67333984375,
0.21435546875
],
[
0.72900390625,
0.203125,
0.75732421875,
0.203125,
0.75732421875,
0.215576171875,
0.72900390625,
0.215576171875
],
[
0.7841796875,
0.201904296875,
0.8154296875,
0.2037353515625,
0.81201171875,
0.2161865234375,
0.78271484375,
0.213134765625
],
[
0.82275390625,
0.203125,
0.8505859375,
0.203125,
0.8505859375,
0.21435546875,
0.82275390625,
0.21435546875
],
[
0.39208984375,
0.210205078125,
0.405517578125,
0.215576171875,
0.39794921875,
0.2244873046875,
0.385498046875,
0.2197265625
],
[
0.387939453125,
0.2037353515625,
0.4013671875,
0.2037353515625,
0.4013671875,
0.21435546875,
0.387939453125,
0.21435546875
],
[
0.1275634765625,
0.2120361328125,
0.2822265625,
0.213134765625,
0.2822265625,
0.2244873046875,
0.1275634765625,
0.223876953125
],
[
0.5263671875,
0.21142578125,
0.5556640625,
0.21435546875,
0.55322265625,
0.2244873046875,
0.52294921875,
0.221435546875
],
[
0.6767578125,
0.213134765625,
0.70361328125,
0.213134765625,
0.70361328125,
0.2244873046875,
0.6767578125,
0.2244873046875
],
[
0.82275390625,
0.213134765625,
0.85205078125,
0.213134765625,
0.85205078125,
0.2257080078125,
0.82275390625,
0.2257080078125
],
[
0.408935546875,
0.22265625,
0.56689453125,
0.22265625,
0.56689453125,
0.236328125,
0.408935546875,
0.236328125
],
[
0.67822265625,
0.2340087890625,
0.70947265625,
0.2340087890625,
0.70947265625,
0.248779296875,
0.67822265625,
0.248779296875
],
[
0.82275390625,
0.2327880859375,
0.85546875,
0.235107421875,
0.85400390625,
0.2476806640625,
0.8212890625,
0.2452392578125
],
[
0.5263671875,
0.235107421875,
0.56103515625,
0.235107421875,
0.56103515625,
0.2464599609375,
0.5263671875,
0.2464599609375
],
[
0.1234130859375,
0.235107421875,
0.300537109375,
0.235107421875,
0.300537109375,
0.2452392578125,
0.1234130859375,
0.2452392578125
],
[
0.34765625,
0.236328125,
0.36279296875,
0.236328125,
0.36279296875,
0.2452392578125,
0.34765625,
0.2452392578125
],
[
0.38623046875,
0.2381591796875,
0.40478515625,
0.2340087890625,
0.4072265625,
0.2423095703125,
0.3896484375,
0.2464599609375
],
[
0.432373046875,
0.2464599609375,
0.5439453125,
0.2464599609375,
0.5439453125,
0.257080078125,
0.432373046875,
0.257080078125
],
[
0.124267578125,
0.255859375,
0.181396484375,
0.255859375,
0.181396484375,
0.267333984375,
0.124267578125,
0.267333984375
],
[
0.58251953125,
0.255859375,
0.609375,
0.255859375,
0.609375,
0.268310546875,
0.58251953125,
0.268310546875
],
[
0.634765625,
0.255859375,
0.6630859375,
0.255859375,
0.6630859375,
0.268310546875,
0.634765625,
0.268310546875
],
[
0.73046875,
0.255859375,
0.75732421875,
0.255859375,
0.75732421875,
0.268310546875,
0.73046875,
0.268310546875
],
[
0.78515625,
0.254150390625,
0.8154296875,
0.255859375,
0.8134765625,
0.268310546875,
0.78271484375,
0.2666015625
],
[
0.490234375,
0.257080078125,
0.51904296875,
0.257080078125,
0.51904296875,
0.268310546875,
0.490234375,
0.268310546875
],
[
0.432373046875,
0.257080078125,
0.4609375,
0.257080078125,
0.4609375,
0.268310546875,
0.432373046875,
0.268310546875
],
[
0.52490234375,
0.257080078125,
0.5576171875,
0.257080078125,
0.5576171875,
0.268310546875,
0.52490234375,
0.268310546875
],
[
0.8212890625,
0.257080078125,
0.85205078125,
0.257080078125,
0.85205078125,
0.268310546875,
0.8212890625,
0.268310546875
],
[
0.3896484375,
0.25830078125,
0.403076171875,
0.25830078125,
0.403076171875,
0.267333984375,
0.3896484375,
0.267333984375
],
[
0.35107421875,
0.25830078125,
0.36279296875,
0.25830078125,
0.36279296875,
0.267333984375,
0.35107421875,
0.267333984375
],
[
0.68115234375,
0.25830078125,
0.7001953125,
0.25830078125,
0.7001953125,
0.267333984375,
0.68115234375,
0.267333984375
],
[
0.1275634765625,
0.2666015625,
0.187255859375,
0.2666015625,
0.187255859375,
0.27783203125,
0.1275634765625,
0.27783203125
],
[
0.63623046875,
0.2666015625,
0.6630859375,
0.2666015625,
0.6630859375,
0.27783203125,
0.63623046875,
0.27783203125
],
[
0.72900390625,
0.2666015625,
0.755859375,
0.2666015625,
0.755859375,
0.27783203125,
0.72900390625,
0.27783203125
],
[
0.78515625,
0.2666015625,
0.8154296875,
0.2666015625,
0.8154296875,
0.27783203125,
0.78515625,
0.27783203125
],
[
0.3896484375,
0.267333984375,
0.403076171875,
0.267333984375,
0.403076171875,
0.275634765625,
0.3896484375,
0.275634765625
],
[
0.349365234375,
0.267333984375,
0.36279296875,
0.267333984375,
0.36279296875,
0.275634765625,
0.349365234375,
0.275634765625
],
[
0.435791015625,
0.267333984375,
0.45751953125,
0.267333984375,
0.45751953125,
0.27783203125,
0.435791015625,
0.27783203125
],
[
0.58837890625,
0.267333984375,
0.60791015625,
0.267333984375,
0.60791015625,
0.276611328125,
0.58837890625,
0.276611328125
],
[
0.68115234375,
0.267333984375,
0.7001953125,
0.267333984375,
0.7001953125,
0.276611328125,
0.68115234375,
0.276611328125
],
[
0.8212890625,
0.267333984375,
0.8505859375,
0.267333984375,
0.8505859375,
0.279052734375,
0.8212890625,
0.279052734375
],
[
0.52490234375,
0.268310546875,
0.55517578125,
0.268310546875,
0.55517578125,
0.276611328125,
0.52490234375,
0.276611328125
],
[
0.491943359375,
0.268310546875,
0.51708984375,
0.268310546875,
0.51708984375,
0.276611328125,
0.491943359375,
0.276611328125
],
[
0.118408203125,
0.289306640625,
0.880859375,
0.289306640625,
0.880859375,
0.302734375,
0.118408203125,
0.302734375
],
[
0.118408203125,
0.3046875,
0.884765625,
0.3046875,
0.884765625,
0.318359375,
0.118408203125,
0.318359375
],
[
0.118408203125,
0.319580078125,
0.88330078125,
0.319580078125,
0.88330078125,
0.333251953125,
0.118408203125,
0.333251953125
],
[
0.118408203125,
0.333251953125,
0.67529296875,
0.33203125,
0.67529296875,
0.345703125,
0.118408203125,
0.3466796875
],
[
0.6689453125,
0.334228515625,
0.8818359375,
0.333251953125,
0.8818359375,
0.344482421875,
0.6689453125,
0.345703125
],
[
0.118408203125,
0.347412109375,
0.884765625,
0.347412109375,
0.884765625,
0.361083984375,
0.118408203125,
0.361083984375
],
[
0.1175537109375,
0.3623046875,
0.22412109375,
0.363525390625,
0.22412109375,
0.374755859375,
0.1175537109375,
0.37353515625
],
[
0.5634765625,
0.399658203125,
0.62744140625,
0.399658203125,
0.62744140625,
0.410888671875,
0.5634765625,
0.410888671875
],
[
0.6708984375,
0.399658203125,
0.72119140625,
0.399658203125,
0.72119140625,
0.410888671875,
0.6708984375,
0.410888671875
],
[
0.78515625,
0.40087890625,
0.82958984375,
0.40087890625,
0.82958984375,
0.412109375,
0.78515625,
0.412109375
],
[
0.73779296875,
0.402099609375,
0.771484375,
0.402099609375,
0.771484375,
0.410400390625,
0.73779296875,
0.410400390625
],
[
0.12005615234375,
0.402099609375,
0.48779296875,
0.402099609375,
0.48779296875,
0.4150390625,
0.12005615234375,
0.4150390625
],
[
0.5634765625,
0.410400390625,
0.67529296875,
0.412109375,
0.67529296875,
0.42333984375,
0.5634765625,
0.421630859375
],
[
0.6708984375,
0.412109375,
0.71533203125,
0.412109375,
0.71533203125,
0.42333984375,
0.6708984375,
0.42333984375
],
[
0.7314453125,
0.412109375,
0.76318359375,
0.412109375,
0.76318359375,
0.42333984375,
0.7314453125,
0.42333984375
],
[
0.78515625,
0.412109375,
0.82275390625,
0.412109375,
0.82275390625,
0.42333984375,
0.78515625,
0.42333984375
],
[
0.78662109375,
0.421630859375,
0.8212890625,
0.42333984375,
0.8193359375,
0.435791015625,
0.78515625,
0.43408203125
],
[
0.57080078125,
0.42333984375,
0.6484375,
0.42333984375,
0.6484375,
0.435302734375,
0.57080078125,
0.435302734375
],
[
0.67822265625,
0.42333984375,
0.7138671875,
0.42333984375,
0.7138671875,
0.435302734375,
0.67822265625,
0.435302734375
],
[
0.73486328125,
0.42333984375,
0.76171875,
0.42333984375,
0.76171875,
0.435302734375,
0.73486328125,
0.435302734375
],
[
0.12005615234375,
0.43408203125,
0.486083984375,
0.43408203125,
0.486083984375,
0.44775390625,
0.12005615234375,
0.44775390625
],
[
0.51123046875,
0.446533203125,
0.88330078125,
0.44775390625,
0.88330078125,
0.4619140625,
0.51123046875,
0.46142578125
],
[
0.1175537109375,
0.45068359375,
0.2032470703125,
0.44775390625,
0.204833984375,
0.463134765625,
0.118408203125,
0.466064453125
],
[
0.118408203125,
0.41748046875,
0.48779296875,
0.416259765625,
0.48779296875,
0.43115234375,
0.118408203125,
0.431640625
],
[
0.5146484375,
0.46435546875,
0.65576171875,
0.46435546875,
0.65576171875,
0.474365234375,
0.5146484375,
0.474365234375
],
[
0.6591796875,
0.46728515625,
0.67236328125,
0.46728515625,
0.67236328125,
0.471435546875,
0.6591796875,
0.471435546875
],
[
0.6708984375,
0.4619140625,
0.88330078125,
0.4619140625,
0.88330078125,
0.4755859375,
0.6708984375,
0.4755859375
],
[
0.12005615234375,
0.476806640625,
0.154541015625,
0.476806640625,
0.154541015625,
0.488037109375,
0.12005615234375,
0.488037109375
],
[
0.162109375,
0.476806640625,
0.31982421875,
0.476806640625,
0.31982421875,
0.490478515625,
0.162109375,
0.490478515625
],
[
0.51318359375,
0.476806640625,
0.8447265625,
0.476806640625,
0.8447265625,
0.490478515625,
0.51318359375,
0.490478515625
],
[
0.1275634765625,
0.496337890625,
0.48779296875,
0.496337890625,
0.48779296875,
0.51025390625,
0.1275634765625,
0.51025390625
],
[
0.151123046875,
0.478515625,
0.16796875,
0.478515625,
0.16796875,
0.48681640625,
0.151123046875,
0.48681640625
],
[
0.12005615234375,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.52685546875,
0.12005615234375,
0.52685546875
],
[
0.51318359375,
0.5166015625,
0.88330078125,
0.515625,
0.88330078125,
0.52978515625,
0.51318359375,
0.53076171875
],
[
0.1275634765625,
0.529296875,
0.48779296875,
0.529296875,
0.48779296875,
0.54345703125,
0.1275634765625,
0.54345703125
],
[
0.51123046875,
0.53076171875,
0.8818359375,
0.5322265625,
0.8818359375,
0.54638671875,
0.51123046875,
0.5458984375
],
[
0.118408203125,
0.54443359375,
0.48779296875,
0.54345703125,
0.48779296875,
0.55810546875,
0.118408203125,
0.55859375
],
[
0.5146484375,
0.548828125,
0.88330078125,
0.548828125,
0.88330078125,
0.5625,
0.5146484375,
0.5625
],
[
0.12005615234375,
0.56103515625,
0.48779296875,
0.56103515625,
0.48779296875,
0.57470703125,
0.12005615234375,
0.57470703125
],
[
0.51318359375,
0.56396484375,
0.88330078125,
0.5634765625,
0.88330078125,
0.57763671875,
0.51318359375,
0.5791015625
],
[
0.12005615234375,
0.576171875,
0.48779296875,
0.576171875,
0.48779296875,
0.5888671875,
0.12005615234375,
0.5888671875
],
[
0.515625,
0.58056640625,
0.88330078125,
0.58056640625,
0.88330078125,
0.59423828125,
0.515625,
0.59423828125
],
[
0.1376953125,
0.59326171875,
0.484375,
0.59326171875,
0.484375,
0.60693359375,
0.1376953125,
0.60693359375
],
[
0.480224609375,
0.5986328125,
0.484375,
0.5986328125,
0.484375,
0.60400390625,
0.480224609375,
0.60400390625
],
[
0.5146484375,
0.595703125,
0.7001953125,
0.5966796875,
0.7001953125,
0.61083984375,
0.5146484375,
0.60986328125
],
[
0.12005615234375,
0.609375,
0.48779296875,
0.609375,
0.48779296875,
0.62255859375,
0.12005615234375,
0.62255859375
],
[
0.118408203125,
0.62451171875,
0.486083984375,
0.62451171875,
0.486083984375,
0.63818359375,
0.118408203125,
0.63818359375
],
[
0.5205078125,
0.619140625,
0.88330078125,
0.62060546875,
0.88330078125,
0.63525390625,
0.5205078125,
0.63427734375
],
[
0.12005615234375,
0.64111328125,
0.48779296875,
0.64111328125,
0.48779296875,
0.65478515625,
0.12005615234375,
0.65478515625
],
[
0.515625,
0.63720703125,
0.86572265625,
0.63720703125,
0.86572265625,
0.65087890625,
0.515625,
0.65087890625
],
[
0.12005615234375,
0.65673828125,
0.47021484375,
0.65673828125,
0.47021484375,
0.67138671875,
0.12005615234375,
0.67138671875
],
[
0.5146484375,
0.65185546875,
0.88330078125,
0.65283203125,
0.88330078125,
0.66748046875,
0.5146484375,
0.66650390625
],
[
0.1368408203125,
0.67333984375,
0.48779296875,
0.67333984375,
0.48779296875,
0.68701171875,
0.1368408203125,
0.68701171875
],
[
0.5146484375,
0.66943359375,
0.8818359375,
0.66943359375,
0.8818359375,
0.68408203125,
0.5146484375,
0.68408203125
],
[
0.12005615234375,
0.6904296875,
0.489501953125,
0.6904296875,
0.489501953125,
0.70361328125,
0.12005615234375,
0.70361328125
],
[
0.5146484375,
0.68603515625,
0.88330078125,
0.68603515625,
0.88330078125,
0.69970703125,
0.5146484375,
0.69970703125
],
[
0.12005615234375,
0.7060546875,
0.484375,
0.7060546875,
0.484375,
0.72021484375,
0.12005615234375,
0.72021484375
],
[
0.51123046875,
0.70068359375,
0.880859375,
0.70068359375,
0.880859375,
0.71435546875,
0.51123046875,
0.71435546875
],
[
0.12005615234375,
0.72265625,
0.484375,
0.72265625,
0.484375,
0.736328125,
0.12005615234375,
0.736328125
],
[
0.51708984375,
0.7197265625,
0.8154296875,
0.7197265625,
0.8154296875,
0.72998046875,
0.51708984375,
0.72998046875
],
[
0.118408203125,
0.73681640625,
0.484375,
0.73681640625,
0.484375,
0.75048828125,
0.118408203125,
0.75048828125
],
[
0.53076171875,
0.73291015625,
0.8818359375,
0.73291015625,
0.8818359375,
0.74658203125,
0.53076171875,
0.74658203125
],
[
0.118408203125,
0.7529296875,
0.48779296875,
0.75439453125,
0.48779296875,
0.7685546875,
0.118408203125,
0.76708984375
],
[
0.5146484375,
0.75,
0.88330078125,
0.75,
0.88330078125,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.1175537109375,
0.7685546875,
0.48779296875,
0.76953125,
0.48779296875,
0.78369140625,
0.1175537109375,
0.783203125
],
[
0.515625,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.515625,
0.77978515625
],
[
0.12005615234375,
0.7861328125,
0.48779296875,
0.7861328125,
0.48779296875,
0.7998046875,
0.12005615234375,
0.7998046875
],
[
0.51318359375,
0.78076171875,
0.88330078125,
0.7822265625,
0.88330078125,
0.79638671875,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.8017578125,
0.489501953125,
0.80029296875,
0.489501953125,
0.8154296875,
0.118408203125,
0.81640625
],
[
0.5146484375,
0.798828125,
0.880859375,
0.798828125,
0.880859375,
0.8125,
0.5146484375,
0.8125
],
[
0.515625,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.8291015625,
0.515625,
0.8291015625
],
[
0.1376953125,
0.82373046875,
0.489501953125,
0.82373046875,
0.489501953125,
0.83740234375,
0.1376953125,
0.83740234375
],
[
0.5146484375,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.12005615234375,
0.83837890625,
0.486083984375,
0.83837890625,
0.486083984375,
0.8486328125,
0.12005615234375,
0.8486328125
],
[
0.118408203125,
0.849609375,
0.48779296875,
0.849609375,
0.48779296875,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.515625,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.515625,
0.86083984375
],
[
0.118408203125,
0.85986328125,
0.489501953125,
0.85986328125,
0.489501953125,
0.87353515625,
0.118408203125,
0.87353515625
],
[
0.5146484375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.5146484375,
0.876953125
],
[
0.118408203125,
0.87158203125,
0.490234375,
0.87158203125,
0.490234375,
0.88525390625,
0.118408203125,
0.88525390625
],
[
0.51318359375,
0.87744140625,
0.82861328125,
0.87890625,
0.82861328125,
0.8935546875,
0.51318359375,
0.892578125
],
[
0.118408203125,
0.88427734375,
0.489501953125,
0.88427734375,
0.489501953125,
0.89794921875,
0.118408203125,
0.89794921875
],
[
0.118408203125,
0.89794921875,
0.300537109375,
0.89794921875,
0.300537109375,
0.90771484375,
0.118408203125,
0.90771484375
],
[
0.53564453125,
0.89453125,
0.884765625,
0.89453125,
0.884765625,
0.9091796875,
0.53564453125,
0.9091796875
]
] |
[
"de-en",
"fr-en",
"Method",
"ro-en",
"Inter.",
"Fulle",
"bidir",
"en → fr",
"bidir",
"bidir",
"en → de",
"en",
"+en",
"ro",
"→en",
"en",
"f",
"de",
"→ro",
"Statistical Methods",
"Fast-Align (Dyer et al., 2013)",
"17.1",
"33.6",
"36.8",
"31.8",
"16.3",
"28.5",
"30.4",
"25.7",
"12.1",
"18.8",
"Y",
"19.6",
"17.8",
"7.1",
"6.1",
"27.4",
"28.7",
"Y",
"GIZA++ (Brown et al.,1993)",
"7.2",
"26.0",
"Neural Methods",
"27.5",
"35.1",
"30.9",
"33.3",
"Naive-Att (Garg et al., 2019)",
"A",
"N",
"36.5",
"28.1",
"23.6",
"16.0",
"33.6",
"40.9",
"32.4",
"35.5",
"50.8",
"N",
"39.8",
"29.8",
"21.2",
"37.5",
"Y",
"Naive-Att-LA (Garg et al., 2019)",
"32.7",
"Shift-Att-LA",
"Y",
"45.5",
"66.1",
"54.7",
"N",
"46.2",
"60.5",
"46.9",
"55.1",
"60.4",
"65.3",
"Y",
"45.8",
"30.3",
"25.5",
"27.0",
"15.6",
"41.3",
"39.9",
"33.7",
"36.4",
"SmoothGrad (Li et al., 2016)",
"N",
"Y",
"36.4",
"43.0",
"29.0",
"25.9",
"29.7",
"15.3",
"41.4",
"N",
"SD-SmoothGrad (Ding et al., 2019)",
"41.2",
"32.7",
"4",
"34.4",
"23.1",
"40.2",
"38.1",
"PD (Li et al., 2019)",
"44.8",
"32.4",
"31.1",
"40.8",
"35.6",
"23.8",
"N",
"AddSGD (Zenkel et al.,2019)",
"Z.2",
"26.6",
"30.4",
"21.2",
"20.5",
"10.0",
"32.3",
"34.8",
"27.6",
"A",
"N",
"Mtl-Fullc (Garg et al., 2019)",
"20.2",
"7.7",
"26.0",
"Statistical + Neural Methods",
"4.6",
"23.1",
"16.0",
"Mtl-Fullc-GZ (Garg et al.,2019)",
"N",
"Y",
"Our Neural Methods",
"Shift-Att",
"17.1",
"16.1",
"27.4",
"26.0",
"25.7",
"20.9",
"17.9",
"23.9",
"N",
"Y",
"6.6",
"Shift-AET",
"10.5",
"22.7",
"23.6",
"N",
"N",
"15.8",
"9.9",
"4.7",
"21.2",
"15.4",
"19.2",
"Table 2: AER on the test set with different alignment methods. bidir are symmetrized alignment results.",
"The column Inter. represents whether the method is an interpretation method that can extract alignments",
"from a pretrained vanilla Transformer model. The column Fullc denotes whether full target sentence",
"is used to extract alignments at test time.The lowerAER, thebetter.",
"Wemarkbestsymmetrized",
"interpretation results of vanilla Transformer with underlines, and best symmetrized results among all",
"with boldface.",
"Method",
"de-en",
"ro-en",
"fr-en",
"特征重要性度量来诱导对齐，而其他来自某种",
"BAO-Guided",
"16.3",
"5.0",
"23.4",
"21.2",
"Shift-AET",
"15.4",
"4.7",
"度量的方法的计算成本远高于具有注意力权重",
"Table 3:Comparison of our method Shift-AET",
"的方法。7",
"形式的注意力权重。请注意，具有特征重要性",
"with BAO-Guided",
"",
"(Zenkel et al., 2020).We",
"4.2",
" Alignment Results",
"report the symmetrized AER on the test set.",
"Comparison withBaselines表2将我们的",
"",
"方法与所有基线进行比较。首先，Shift-Att是",
"压器和AET中的参数数量为36。8M和37。",
"一种用于香草变压器的纯解释方法，其性能明",
"3M，并发现AET仅引入1。4%附加参数到香",
"显优于Fast-Align和所有神经基线，并且与",
"草变压器。总之，通过使用对称化的Shift-Att",
"GIZA十＋的性能相当。例如，它在不同语言对",
"对齐来监督对齐模块，Shift-AET比Shift-Att",
"上的8.7-11.1AER点优于SD-SmoothGrad",
"和GIZA++有所改进，参数增加可忽略不计，",
"这是一种具有特征重要性度量的最新方法",
"",
"并且不会影响翻译质量",
"可以从香草转换器中提取对齐。Shift-Att的成",
"功表明，vanillaTransformer以隐含的方式捕",
"Comparison with Zenkel et al.(2020)与我",
"获了对齐信息，如果选择了正确的解码步骤和",
"们的工作同时，Zenkeletal.（2020）提出了",
"层来诱导对齐，则可以从注意力权重中揭示",
"种可以超越GIZA++的神经对准器。表3比",
"其次，方法Shift-AET实现了新的最新技",
"较了Shift-AET和最佳方法BAO-Guided 的",
"术，明显优于所有基线。它通过1.4-4.8AER",
"性能(Birdir。Att。Opt。+引导）在Zenkel",
"在不同的语言对上对GIZA++进行了改进",
"etal.（2020）中。我们观察到Shift-AET在对",
"表明可以构建比GIZA++更好的神经对准器",
"准精度方面比BAO-Guided表现更好",
"而无需使用从统计对准器生成的任何对准来引",
"Shift-AET也比BAO-Guided简单得多。",
"导训练。与MTL-Fullc-GZ相比，我们还发现",
"BAO-Guided的训练包括三个阶段：（???）在",
"Shift-AET的性能稍好（de-en和ro-en）或同",
"源到目标和目标到源的方向上训练香草变压",
"等（fr-en），这是一种使用GIZA++对齐来",
"器；（???）训练对齐层，并通过双向注意力优",
"引导训练的方法。我们评估模型尺寸：香草变",
"化在训练集上提取对齐。此对齐提取过程的计",
"算成本很高，因为双向注意力优化会针对训练",
"7For each sentence pair, PD forwards once with",
"集中的每个句子对分别微调模型参数；（???）",
"+1masked sentencepairsas the input,while",
"SmoothGradand SD-SmoothGradforwardand back-",
"以提取的对齐方式为指导，重新训练对齐层。",
"ward once with m (m = 30 in Ding et al. (2019)",
"相比之下，Shift-AET可以在两个阶段进行更",
" noisy sentence pairs as the input. In contrast, at-",
"快的训练，并且不涉及双向注意力优化",
"tention weights based methods forward once with one",
"sentence pair as the input.",
"与Mtl-Fullc E(Garg et al.，2019)类似,"
] |
{
"arxiv_id": "2004.14837",
"language": "zh",
"page_idx": 4
}
|
||
2312.03577-zh-1
|
2312.03577-zh-1.jpg
|
$\textbf{Eojin Jeon}^{1*},\textbf{Mingyu Lee}^{1*},\textbf{Juhyeong Park}^{1}$,
Yeachan Kim$^1,\textbf{Wing-Lam Mok}^1$,SangKeun Lee$^1,2$
$^{1}$Department of Artificial Intelligence $^2$Department of Computer Science and Engineering
Korea University, Seoul, Republic of Korea
{ skdlcm456, decon9201, johnida, yeachan, wlmokac, yalphy } @korea.ac.kr
## $\mathbf{Abstract}$
<FigureHere>
Figure 1: Comparison between (a) existing debiasing methods and (b) our proposed framework. (a): identify biased examples by using an auxiliary model trained with the multi-class learning objective, the same as the main model. (b): introduce intermediate binary classifiers, called bias experts, between the auxiliary model and the main model.
Biases in the dataset often enable the model to achieve high performance on in-distribution data, while poorly performing on out-ofdistribution data. To mitigate the detrimental effect of the bias on the networks, previous works have proposed debiasing methods that down-weight the biased examples identified by an auxiliary model, which is trained with explicit bias labels. However, finding a type of bias in datasets is a costly process. Therefore, recent studies have attempted to make the auxiliary model biased without the guidance (or annotation) of bias labels, by constraining the model's training environment or the capability of the model itself. Despite the promising debiasing results of recent works, the multi-class learning objective, which has been naively used to train the auxiliary model, may harm the bias mitigation effect due to its regularization effect and competitive nature across classes. As an alternative, we propose a new debiasing framework that introduces binary classifiers between the auxiliary model and the main model, coined bias experts. Specifically, each bias expert is trained on a binary classification task derived from the multi-class classification task via the One-vs-Rest approach. Experimental results demonstrate that our proposed strategy improves the bias identification ability of the auxiliary model. Consequently, our debiased model consistently outperforms the state-of-the-art on various challenge datasets.$^{\mathrm{l}}$
(Kim et al., 2019; Schuster et al., 2019; Clark et al., 2019; Mahabadi et al., 2020) (Bahng et al., 2020)( )(Sanh et al., 2021) (Liu et al., 2021) (Utama et al., 2020b; Kim et al., 2022) GCE (Nam et al.,2020)
(gsoftmaxcross-entropy): Feldman et al. (2019)
softmax1 (Wen et al., 2022)
1 One-vs-Rest (Rifkin and Klautau, 2004)
Nam et al. (2020)
· We show that the multi-class learning objective may harm the bias identification ability of the auxiliary model.
·We propose a novel debiasing approach based on bias experts to improve the out-ofdistribution performance of existing debiasing methods.
$\bullet$ Through various empirical evaluations, we show that introducing bias experts improves debiasing performance.
## 1 Introduction
(He et al., 2016) (Devlin et al., 2019) (Ribeiro et al., 2016; Zhu et al., 2017; Gururangan et al., 2018; McCoy et al., 2019) NLI (Gururangan et al., 2018; Poliak et al., 2018) $^{2}$
$^*$These authors contributed equally to this work.
$^{1}$Our code is available at https://github.com/jej127/
Bias-Experts .
$^2$Bias-conflicting examples indicate examples that cannot
be correctly predicted by solely relying on biases.
|
[
[
0.300537109375,
0.1341552734375,
0.7021484375,
0.1341552734375,
0.7021484375,
0.1519775390625,
0.300537109375,
0.1519775390625
],
[
0.280517578125,
0.152587890625,
0.7197265625,
0.15087890625,
0.7197265625,
0.168701171875,
0.280517578125,
0.17041015625
],
[
0.14111328125,
0.169189453125,
0.8623046875,
0.171630859375,
0.8623046875,
0.1859130859375,
0.14111328125,
0.18408203125
],
[
0.32568359375,
0.18701171875,
0.68115234375,
0.18701171875,
0.68115234375,
0.20068359375,
0.32568359375,
0.20068359375
],
[
0.145263671875,
0.204833984375,
0.869140625,
0.204833984375,
0.869140625,
0.2197265625,
0.145263671875,
0.2197265625
],
[
0.2626953125,
0.2529296875,
0.343505859375,
0.2529296875,
0.343505859375,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.5458984375,
0.254150390625,
0.5810546875,
0.254150390625,
0.5810546875,
0.265380859375,
0.5458984375,
0.265380859375
],
[
0.73876953125,
0.254150390625,
0.7734375,
0.254150390625,
0.7734375,
0.265380859375,
0.73876953125,
0.265380859375
],
[
0.54248046875,
0.26416015625,
0.58447265625,
0.26416015625,
0.58447265625,
0.275634765625,
0.54248046875,
0.275634765625
],
[
0.736328125,
0.262451171875,
0.77734375,
0.26416015625,
0.77685546875,
0.276611328125,
0.73486328125,
0.27490234375
],
[
0.145263671875,
0.27978515625,
0.459228515625,
0.27978515625,
0.459228515625,
0.29345703125,
0.145263671875,
0.29345703125
],
[
0.145263671875,
0.29443359375,
0.459228515625,
0.29443359375,
0.459228515625,
0.30810546875,
0.145263671875,
0.30810546875
],
[
0.145263671875,
0.30810546875,
0.4609375,
0.30810546875,
0.4609375,
0.32177734375,
0.145263671875,
0.32177734375
],
[
0.646484375,
0.32177734375,
0.6708984375,
0.317626953125,
0.67529296875,
0.328857421875,
0.6513671875,
0.333251953125
],
[
0.8193359375,
0.319580078125,
0.83447265625,
0.322509765625,
0.82958984375,
0.330810546875,
0.8154296875,
0.32666015625
],
[
0.6484375,
0.33203125,
0.69775390625,
0.33203125,
0.69775390625,
0.342529296875,
0.6484375,
0.342529296875
],
[
0.1435546875,
0.322509765625,
0.459228515625,
0.322509765625,
0.459228515625,
0.336181640625,
0.1435546875,
0.336181640625
],
[
0.73291015625,
0.33203125,
0.78076171875,
0.33203125,
0.78076171875,
0.342529296875,
0.73291015625,
0.342529296875
],
[
0.81005859375,
0.33203125,
0.8623046875,
0.33203125,
0.8623046875,
0.342529296875,
0.81005859375,
0.342529296875
],
[
0.5283203125,
0.336181640625,
0.59423828125,
0.33837890625,
0.59423828125,
0.349853515625,
0.5283203125,
0.347412109375
],
[
0.145263671875,
0.336181640625,
0.45751953125,
0.33740234375,
0.45751953125,
0.350830078125,
0.145263671875,
0.349853515625
],
[
0.65576171875,
0.34326171875,
0.69189453125,
0.34326171875,
0.69189453125,
0.355224609375,
0.65576171875,
0.355224609375
],
[
0.74072265625,
0.341552734375,
0.77099609375,
0.342529296875,
0.76904296875,
0.35400390625,
0.73876953125,
0.352783203125
],
[
0.82275390625,
0.34326171875,
0.85400390625,
0.34326171875,
0.85400390625,
0.355224609375,
0.82275390625,
0.355224609375
],
[
0.1435546875,
0.350830078125,
0.459228515625,
0.350830078125,
0.459228515625,
0.364013671875,
0.1435546875,
0.364013671875
],
[
0.53076171875,
0.349853515625,
0.58837890625,
0.349853515625,
0.58837890625,
0.361083984375,
0.53076171875,
0.361083984375
],
[
0.64990234375,
0.352783203125,
0.69775390625,
0.352783203125,
0.69775390625,
0.363525390625,
0.64990234375,
0.363525390625
],
[
0.72705078125,
0.352783203125,
0.78271484375,
0.352783203125,
0.78271484375,
0.363525390625,
0.72705078125,
0.363525390625
],
[
0.53662109375,
0.35986328125,
0.58251953125,
0.35986328125,
0.58251953125,
0.37060546875,
0.53662109375,
0.37060546875
],
[
0.8134765625,
0.355224609375,
0.85791015625,
0.355224609375,
0.85791015625,
0.3623046875,
0.8134765625,
0.3623046875
],
[
0.145263671875,
0.364013671875,
0.45751953125,
0.364013671875,
0.45751953125,
0.377685546875,
0.145263671875,
0.377685546875
],
[
0.1435546875,
0.380126953125,
0.459228515625,
0.380126953125,
0.459228515625,
0.393798828125,
0.1435546875,
0.393798828125
],
[
0.72705078125,
0.377685546875,
0.78271484375,
0.377685546875,
0.78271484375,
0.388427734375,
0.72705078125,
0.388427734375
],
[
0.73486328125,
0.385986328125,
0.77685546875,
0.385986328125,
0.77685546875,
0.39794921875,
0.73486328125,
0.39794921875
],
[
0.145263671875,
0.393798828125,
0.459228515625,
0.393798828125,
0.459228515625,
0.40673828125,
0.145263671875,
0.40673828125
],
[
0.5400390625,
0.40380859375,
0.5859375,
0.40625,
0.58447265625,
0.421630859375,
0.5380859375,
0.419189453125
],
[
0.75390625,
0.399658203125,
0.759765625,
0.399658203125,
0.759765625,
0.40673828125,
0.75390625,
0.40673828125
],
[
0.145263671875,
0.407958984375,
0.45751953125,
0.407958984375,
0.45751953125,
0.421630859375,
0.145263671875,
0.421630859375
],
[
0.73291015625,
0.40625,
0.779296875,
0.40625,
0.779296875,
0.42041015625,
0.73291015625,
0.42041015625
],
[
0.1435546875,
0.4228515625,
0.459228515625,
0.4228515625,
0.459228515625,
0.435791015625,
0.1435546875,
0.435791015625
],
[
0.5498046875,
0.42333984375,
0.5751953125,
0.42333984375,
0.5751953125,
0.43701171875,
0.5498046875,
0.43701171875
],
[
0.7421875,
0.421630859375,
0.771484375,
0.421630859375,
0.771484375,
0.438232421875,
0.7421875,
0.438232421875
],
[
0.14697265625,
0.438232421875,
0.45751953125,
0.438232421875,
0.45751953125,
0.448974609375,
0.14697265625,
0.448974609375
],
[
0.145263671875,
0.45068359375,
0.45751953125,
0.45068359375,
0.45751953125,
0.46435546875,
0.145263671875,
0.46435546875
],
[
0.51123046875,
0.448974609375,
0.8818359375,
0.448974609375,
0.8818359375,
0.4619140625,
0.51123046875,
0.4619140625
],
[
0.1427001953125,
0.463134765625,
0.45751953125,
0.46435546875,
0.45751953125,
0.47802734375,
0.1427001953125,
0.476806640625
],
[
0.51318359375,
0.46435546875,
0.87890625,
0.46435546875,
0.87890625,
0.474365234375,
0.51318359375,
0.474365234375
],
[
0.1435546875,
0.478515625,
0.459228515625,
0.478515625,
0.459228515625,
0.4921875,
0.1435546875,
0.4921875
],
[
0.51123046875,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.490478515625,
0.51123046875,
0.490478515625
],
[
0.14697265625,
0.49462890625,
0.456787109375,
0.49462890625,
0.456787109375,
0.5048828125,
0.14697265625,
0.5048828125
],
[
0.51123046875,
0.4912109375,
0.8818359375,
0.4912109375,
0.8818359375,
0.5048828125,
0.51123046875,
0.5048828125
],
[
0.145263671875,
0.50732421875,
0.45751953125,
0.50732421875,
0.45751953125,
0.52099609375,
0.145263671875,
0.52099609375
],
[
0.51318359375,
0.50732421875,
0.8818359375,
0.50732421875,
0.8818359375,
0.51708984375,
0.51318359375,
0.51708984375
],
[
0.145263671875,
0.521484375,
0.459228515625,
0.521484375,
0.459228515625,
0.53515625,
0.145263671875,
0.53515625
],
[
0.51123046875,
0.51953125,
0.8818359375,
0.51953125,
0.8818359375,
0.533203125,
0.51123046875,
0.533203125
],
[
0.14697265625,
0.53759765625,
0.45751953125,
0.53759765625,
0.45751953125,
0.54736328125,
0.14697265625,
0.54736328125
],
[
0.51123046875,
0.53369140625,
0.6513671875,
0.53369140625,
0.6513671875,
0.54736328125,
0.51123046875,
0.54736328125
],
[
0.1435546875,
0.548828125,
0.45751953125,
0.5498046875,
0.45751953125,
0.5634765625,
0.1435546875,
0.5625
],
[
0.1435546875,
0.56396484375,
0.45751953125,
0.56396484375,
0.45751953125,
0.57763671875,
0.1435546875,
0.57763671875
],
[
0.53564453125,
0.57080078125,
0.8818359375,
0.57080078125,
0.8818359375,
0.58447265625,
0.53564453125,
0.58447265625
],
[
0.145263671875,
0.5791015625,
0.45751953125,
0.5791015625,
0.45751953125,
0.5927734375,
0.145263671875,
0.5927734375
],
[
0.509765625,
0.5859375,
0.8818359375,
0.58740234375,
0.8818359375,
0.6015625,
0.509765625,
0.60107421875
],
[
0.1435546875,
0.5927734375,
0.45751953125,
0.5927734375,
0.45751953125,
0.60546875,
0.1435546875,
0.60546875
],
[
0.145263671875,
0.60693359375,
0.45751953125,
0.60693359375,
0.45751953125,
0.62060546875,
0.145263671875,
0.62060546875
],
[
0.51123046875,
0.6025390625,
0.875,
0.6025390625,
0.875,
0.6162109375,
0.51123046875,
0.6162109375
],
[
0.1435546875,
0.62060546875,
0.459228515625,
0.62060546875,
0.459228515625,
0.63427734375,
0.1435546875,
0.63427734375
],
[
0.51123046875,
0.619140625,
0.8818359375,
0.619140625,
0.8818359375,
0.6328125,
0.51123046875,
0.6328125
],
[
0.14697265625,
0.6357421875,
0.45751953125,
0.6357421875,
0.45751953125,
0.646484375,
0.14697265625,
0.646484375
],
[
0.51318359375,
0.63720703125,
0.60205078125,
0.63720703125,
0.60205078125,
0.6484375,
0.51318359375,
0.6484375
],
[
0.145263671875,
0.6494140625,
0.45751953125,
0.6494140625,
0.45751953125,
0.6630859375,
0.145263671875,
0.6630859375
],
[
0.53076171875,
0.65185546875,
0.8818359375,
0.65185546875,
0.8818359375,
0.6650390625,
0.53076171875,
0.6650390625
],
[
0.145263671875,
0.6630859375,
0.4609375,
0.6630859375,
0.4609375,
0.6767578125,
0.145263671875,
0.6767578125
],
[
0.51123046875,
0.66845703125,
0.71875,
0.66845703125,
0.71875,
0.681640625,
0.51123046875,
0.681640625
],
[
0.145263671875,
0.677734375,
0.459228515625,
0.677734375,
0.459228515625,
0.69140625,
0.145263671875,
0.69140625
],
[
0.5400390625,
0.68408203125,
0.873046875,
0.68408203125,
0.873046875,
0.69775390625,
0.5400390625,
0.69775390625
],
[
0.14697265625,
0.693359375,
0.45751953125,
0.693359375,
0.45751953125,
0.70361328125,
0.14697265625,
0.70361328125
],
[
0.51123046875,
0.69970703125,
0.6455078125,
0.69970703125,
0.6455078125,
0.71337890625,
0.51123046875,
0.71337890625
],
[
0.1435546875,
0.7060546875,
0.4609375,
0.7060546875,
0.4609375,
0.7197265625,
0.1435546875,
0.7197265625
],
[
0.145263671875,
0.72021484375,
0.4609375,
0.72021484375,
0.4609375,
0.73388671875,
0.145263671875,
0.73388671875
],
[
0.52880859375,
0.71875,
0.5400390625,
0.71875,
0.5400390625,
0.72998046875,
0.52880859375,
0.72998046875
],
[
0.1435546875,
0.73388671875,
0.255126953125,
0.732421875,
0.255126953125,
0.74658203125,
0.1435546875,
0.7490234375
],
[
0.5341796875,
0.7412109375,
0.8818359375,
0.7412109375,
0.8818359375,
0.7548828125,
0.5341796875,
0.7548828125
],
[
0.54736328125,
0.755859375,
0.87890625,
0.755859375,
0.87890625,
0.77099609375,
0.54736328125,
0.77099609375
],
[
0.118408203125,
0.76318359375,
0.1318359375,
0.76318359375,
0.1318359375,
0.771484375,
0.118408203125,
0.771484375
],
[
0.15283203125,
0.7626953125,
0.2626953125,
0.7626953125,
0.2626953125,
0.77392578125,
0.15283203125,
0.77392578125
],
[
0.5458984375,
0.771484375,
0.71875,
0.7724609375,
0.71875,
0.78759765625,
0.5458984375,
0.7861328125
],
[
0.1234130859375,
0.7861328125,
0.48779296875,
0.7861328125,
0.48779296875,
0.7998046875,
0.1234130859375,
0.7998046875
],
[
0.1158447265625,
0.80029296875,
0.486083984375,
0.8017578125,
0.486083984375,
0.81640625,
0.1158447265625,
0.8154296875
],
[
0.5322265625,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8134765625,
0.5322265625,
0.8134765625
],
[
0.1175537109375,
0.818359375,
0.48779296875,
0.818359375,
0.48779296875,
0.83203125,
0.1175537109375,
0.83203125
],
[
0.54736328125,
0.8154296875,
0.88330078125,
0.8154296875,
0.88330078125,
0.8291015625,
0.54736328125,
0.8291015625
],
[
0.11419677734375,
0.83203125,
0.32421875,
0.830078125,
0.32421875,
0.84716796875,
0.11419677734375,
0.849609375
],
[
0.54833984375,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.845703125,
0.54833984375,
0.845703125
],
[
0.54833984375,
0.8466796875,
0.646484375,
0.8466796875,
0.646484375,
0.86083984375,
0.54833984375,
0.86083984375
],
[
0.1435546875,
0.85546875,
0.435791015625,
0.85546875,
0.435791015625,
0.869140625,
0.1435546875,
0.869140625
],
[
0.1376953125,
0.8662109375,
0.48779296875,
0.86767578125,
0.48779296875,
0.8818359375,
0.1376953125,
0.88134765625
],
[
0.53076171875,
0.87060546875,
0.8818359375,
0.87255859375,
0.8818359375,
0.88720703125,
0.53076171875,
0.88525390625
],
[
0.1175537109375,
0.88134765625,
0.2166748046875,
0.88134765625,
0.2166748046875,
0.892578125,
0.1175537109375,
0.892578125
],
[
0.1351318359375,
0.89111328125,
0.489501953125,
0.8935546875,
0.489501953125,
0.90771484375,
0.1351318359375,
0.90625
],
[
0.5458984375,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.5458984375,
0.90185546875
],
[
0.11419677734375,
0.90380859375,
0.420654296875,
0.90478515625,
0.420654296875,
0.91943359375,
0.11419677734375,
0.91845703125
],
[
0.54833984375,
0.90478515625,
0.72314453125,
0.90478515625,
0.72314453125,
0.91943359375,
0.54833984375,
0.91943359375
]
] |
[
"Eojin Jeonl* , Mingyu Leel*, Juhyeong Park',",
"Yeachan Kim', Wing-Lam Mok', SangKeun Leel,2",
"IDepartment of Artificial Intelligence2Department of Computer Science and Engineering",
"Korea University, Seoul, Republic of Korea",
" skdlcm456, decon9201,johnida,yeachan,wlmokac,yalphy}@korea.ac.kr",
"Abstract",
"Main",
"Main",
"Model",
"Model",
"Biases in the dataset often enable the model",
"to achieve high performance on in-distribution",
"data, while poorly performing on out-of-",
"LA",
"IC",
"Class A?",
"distribution data. To mitigate the detrimental",
"Class B2",
"Class C?",
"Which elass?",
"effect of the bias on the networks, previous",
"Bias",
"Bias",
"Bias",
"works have proposed debiasing methods that",
"Auxiliary",
"Expert 1",
"Expert 2",
"Model",
"Expert",
"down-weight the biased examples identified",
"by an auxiliary model, which is trained with",
"Auxiliary",
"Model",
"explicit bias labels. However, finding a type of",
"Inputs",
"",
"bias in datasets is a costly process. Therefore,",
"Inputs",
"recent studies have attempted to make the aux-",
"(a)",
"(b)",
"iliary model biased withouttheguidance (or",
"annotation) of bias labels, by constraining the",
"Figure l: Comparison between (a) existing debiasing",
"model's training environment or the capability",
"methodsand(b)ourproposedframework.(a):identify",
"of the model itself. Despite the promising de-",
"biased examples by using an auxiliary model trained",
"biasing results ofrecent works,the multi-class",
"with the multi-class learning objective, the same as the",
"learning objective, which has been naively",
"mainmodel.(b):introduceintermediatebinaryclassi-",
"used to train the auxiliary model, may harm",
"fiers, called bias experts, between the auxiliary model",
"the biasmitigationeffect due to itsregular-",
"and the main model.",
"ization effect and competitive nature across",
"classes. As an alternative, we propose a new",
"(Kim et al., 2019; Schuster et al., 2019; Clark",
"debiasing framework that introduces binary",
"et al., 2019; Mahabadi et al., 2020) (Bahng et al.,",
"classifiers between the auxiliary model and the",
"main model, coined bias experts. Specifically,",
"2020) O (Sanh et al., 2021) (Liu et al., 2021)",
"each bias expert is trained on a binary classi-",
"(Utama et al., 2020b; Kim et al., 2022) GCE (Nam",
"fication task derived from the multi-class clas-",
"et al.,2020)",
"sification task via the One-vs-Rest approach.",
"(gsoftmaxcross-entropy): Feldman et al. (2019)",
"Experimental results demonstrate that our pro-",
"softmax1 (Wen et al., 2022)",
"posed strategy improves the bias identifica-",
"1 One-vs-Rest (Rifkin and Klautau, 2004)",
"tionability of the auxiliary model.Conse-",
"Nam et al. (2020)",
"quently, our debiased model consistently out-",
"performs the state-of-the-art on various chal-",
":",
"lenge datasets.",
". We show that the multi-class learning objec-",
"tive may harm the bias identification ability",
"1",
"Introduction",
"of the auxiliary model.",
"(He et al., 2016) (Devlin et al., 2019) (Ribeiro",
"et al., 2016; Zhu et al., 2017; Gururangan et al.,",
". We propose a novel debiasing approach",
"2018; McCoy et al., 2019) NLI (Gururangan et al.,",
"based on bias experts to improve the out-of-",
"2018; Poliak et al., 2018) 2",
"distribution performance of existing debias-",
"ing methods.",
"*These authors contributed equally to this work.",
"'Our code is available at https: //github.com/jej127/",
". Through various empirical evaluations, we",
"Bias-Experts",
"2Bias-conflicting examples indicate examples that cannot",
"show that introducing bias experts improves",
"be correctly predicted by solely relying on biases.",
"debiasing performance."
] |
{
"arxiv_id": "2312.03577",
"language": "zh",
"page_idx": 1
}
|
||
2312.03577-zh-4
|
2312.03577-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>$\mathbf{Method}$</th>
<th>Known biases</th>
<th>dev</th>
<th>$\mathbf{MNLI}$ HANS</th>
<th>$\dot{\text{jap}}$ </th>
<th>dev</th>
<th>$\mathbf{FEVER}$ symm</th>
<th>$R$ Gap</th>
<th>dev</th>
<th>$QQP$ PAWS</th>
<th>Gap 1</th>
</tr>
<tr>
<th>BERT-base (Devlin et al..2019)</th>
<th>$X$</th>
<th>$\overline{84.5}$</th>
<th>62.4</th>
<th>$\overline{22.1}$</th>
<th>$\overline{85.6}$</th>
<th>$\overline{63.1}$</th>
<th>22.5</th>
<th>91.0</th>
<th>33.5</th>
<th>$\overline{57.5}$</th>
</tr>
<tr>
<td>Reweighting ( Clark et al. , 2019) </td>
<td>$√$</td>
<td></td>
<td>$\overline{83.5}$</td>
<td>69.2</td>
<td>$\overline{14.3}$</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>-</td>
<td>Reweighting (Schuster et al., 2019)</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>84.6</td>
<td>66.5</td>
<td>18.1</td>
<td>-</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>PoE (Clark et al.,2019)</td>
<td>1</td>
<td>82.9</td>
<td>67.9</td>
<td>15.0</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>Conf- reg $I$</td>
<td>(Utama et al., 2020a)</td>
<td>84.3</td>
<td>69.1</td>
<td>15.7</td>
<td>86.4</td>
<td>66.2</td>
<td>20.2</td>
<td>-</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>Reweighting (Utama et al..2020b)</td>
<td>$X$</td>
<td>$\overline{82.3}$</td>
<td>69.7</td>
<td>12.6</td>
<td>$\underline{87.1}$</td>
<td>65.5</td>
<td>21.6</td>
<td>85.2</td>
<td>$\underline{57.4}$</td>
<td>27.8</td>
</tr>
<tr>
<td>PoE (Utama et al..2020b)</td>
<td>$X$</td>
<td>81.9</td>
<td>66.8</td>
<td>15.1</td>
<td>85.9</td>
<td>65.8</td>
<td>20.1</td>
<td>86.1</td>
<td>56.3</td>
<td>29.8</td>
</tr>
<tr>
<td>PoE (Sanh et al., 2021)</td>
<td>$X$</td>
<td>$\underline{83.3}$</td>
<td>67.9</td>
<td>15.4</td>
<td>84.8</td>
<td>65.7</td>
<td>$\underline{19.1}$</td>
<td>88.0</td>
<td>46.4</td>
<td>41.6</td>
</tr>
<tr>
<td>Conf- reg (Utama et al..2020b) </td>
<td>$X$</td>
<td>84.3</td>
<td>67.1</td>
<td>17.2</td>
<td>87.6</td>
<td>$\underline{66.0}$</td>
<td>21.6</td>
<td>89.0</td>
<td>43.0</td>
<td>46.0</td>
</tr>
<tr>
<td>Self-Debiasing (Ghaddar et al., 2021)</td>
<td>$X$</td>
<td>83.2</td>
<td>$\underline{71.2}$</td>
<td>$\underline{12.0}$</td>
<td>-</td>
<td> </td>
<td>-</td>
<td>90.2</td>
<td>46.5</td>
<td>43.7</td>
</tr>
<tr>
<td>Bias Experts ( ours) </td>
<td>$X$</td>
<td>82.7</td>
<td>72.6</td>
<td>$\overline{10.1}$</td>
<td>85.6</td>
<td>68.1</td>
<td>17.5</td>
<td>86.8</td>
<td>58.1</td>
<td>$\underline{28.7}$</td>
</tr>
</tbody>
</table>
Table 1: Performances of models evaluated on MNLI, FEVER, QQP, and their corresponding challenge test sets. We also report the difference between in- and out-of-distribution performances as Gap. The results of baselines using prior knowledge of biases are from the original paper. We mark the best and the second-best performance in bold and $\underline{\text{underline}}$,respectively.
### 5 $\textbf{Related Work}$
5.1
Debiasing with prior knowledge of bias
NLI (He et al., 2019; Mahabadi et al., 2020) VQA (Cadène et al.,2019) (Bahng et al., 2020) (Schuster et al., 2019) (Utama et al., 2020a; Mahabadi et al., 2020) PoE (Clark et al., 2019)
<table>
<tbody>
<tr>
<th>Ablative Setting</th>
<th>$\mathbf{MNLI}$</th>
<th>$\mathbf{HANS}$ Ent NEnt Avg</th>
</tr>
<tr>
<td>Bias Experts (ours)</td>
<td>82.7</td>
<td>93.5 51.8 72.6</td>
</tr>
<tr>
<td>(1) w/o Amp. </td>
<td>82.8</td>
<td>94.9 45.0 70.0</td>
</tr>
<tr>
<td>(2) w/o OvR</td>
<td>83.0</td>
<td>94.4 42.7 68.5</td>
</tr>
<tr>
<td>(3) w/o Amp. and OvR</td>
<td>83.1</td>
<td>94.5 41.4 67.9</td>
</tr>
</tbody>
</table>
# 5.2 Debiasing without targeting a specific bias
Table 2: Ablation studies using accuracy on MNLI and
HANS. We mark the best performance in bold .
(Utama et al., 2020b) (Bras et al., 2020; Clark et al., 2020; Sanh et al., 2021; Yaghoobzadeh et al., 2021; Ghaddar et al., 2021) GCE (Nam et al., 2020) (Liu et al., 2021) Kim et al.(2022) Ahn et al.(2023) Lee et al.(2021) Lee et al.(2023) Lyu et al. (2023)
## 6 $\mathbf{Conclusion}$
One-vs-Rest
<table>
<tbody>
<tr>
<th>$\mathbf{Method}$</th>
<th>$\mathbf{MNLI}$</th>
<th>HANS</th>
</tr>
<tr>
<td>BERT- base</td>
<td>84.5</td>
<td>62.4</td>
</tr>
<tr>
<td>PoE (Utama et al..2020b</td>
<td>81.9</td>
<td>66.8</td>
</tr>
<tr>
<td>PoE Sanh et al..2021)</td>
<td>83.3</td>
<td>67.9</td>
</tr>
<tr>
<td>PoE+Ours$(T=2$k)</td>
<td>83.8</td>
<td>68.0</td>
</tr>
<tr>
<td>PoE+Ours$(T=10$k)</td>
<td>83.3</td>
<td>71.0</td>
</tr>
<tr>
<td>Bias Experts (ours</td>
<td>82.7</td>
<td>72.6</td>
</tr>
</tbody>
</table>
7
Limitations
1NLUNLI2-322-54-20NLU
2 Utama et al.(2020a)$\lambda_1\lambda_2\alpha$Wen et al. (2022) $\lambda _{1}= ( k- 1) / k$ $\lambda _{2}= 1/ k$ $k$ $\lambda _{1}$ $\lambda _{2}$ $\alpha$ HANS099 $\alpha$ 110-880 (Utama et al., 2020b; Sanh et al., 2021; Kim et al., 2022; Lyu et al., 2023)
Table 3: Trade-off results between in- and out-ofdistribution performances evaluated on MNLI and HANS. $T$ denotes the number of training steps for bias experts. We mark the best and the second-best performance in bold and $\underline{\text{underline}}$,respectively.
8
Ethics Statement ie)
|
[
[
0.405517578125,
0.08197021484375,
0.466796875,
0.0843505859375,
0.465087890625,
0.09979248046875,
0.40478515625,
0.097412109375
],
[
0.5087890625,
0.0831298828125,
0.5615234375,
0.0831298828125,
0.5615234375,
0.09857177734375,
0.5087890625,
0.09857177734375
],
[
0.640625,
0.0831298828125,
0.70361328125,
0.0831298828125,
0.70361328125,
0.09857177734375,
0.640625,
0.09857177734375
],
[
0.78857421875,
0.0831298828125,
0.82958984375,
0.0831298828125,
0.82958984375,
0.09979248046875,
0.78857421875,
0.09979248046875
],
[
0.12176513671875,
0.0926513671875,
0.187255859375,
0.0926513671875,
0.187255859375,
0.10687255859375,
0.12176513671875,
0.10687255859375
],
[
0.50390625,
0.097412109375,
0.60205078125,
0.09979248046875,
0.60205078125,
0.11517333984375,
0.50390625,
0.1134033203125
],
[
0.78076171875,
0.097412109375,
0.8759765625,
0.09979248046875,
0.8759765625,
0.11517333984375,
0.78076171875,
0.1134033203125
],
[
0.7021484375,
0.09979248046875,
0.74072265625,
0.09979248046875,
0.74072265625,
0.11517333984375,
0.7021484375,
0.11517333984375
],
[
0.408935546875,
0.09979248046875,
0.4609375,
0.09979248046875,
0.4609375,
0.11517333984375,
0.408935546875,
0.11517333984375
],
[
0.64404296875,
0.1009521484375,
0.7001953125,
0.1009521484375,
0.7001953125,
0.11517333984375,
0.64404296875,
0.11517333984375
],
[
0.74658203125,
0.1009521484375,
0.78271484375,
0.1009521484375,
0.78271484375,
0.1134033203125,
0.74658203125,
0.1134033203125
],
[
0.609375,
0.1015625,
0.64208984375,
0.1015625,
0.64208984375,
0.1134033203125,
0.609375,
0.1134033203125
],
[
0.470947265625,
0.1015625,
0.50537109375,
0.1015625,
0.50537109375,
0.1134033203125,
0.470947265625,
0.1134033203125
],
[
0.56494140625,
0.11517333984375,
0.60205078125,
0.11517333984375,
0.60205078125,
0.1312255859375,
0.56494140625,
0.1312255859375
],
[
0.60107421875,
0.11639404296875,
0.6455078125,
0.11639404296875,
0.6455078125,
0.1312255859375,
0.60107421875,
0.1312255859375
],
[
0.7001953125,
0.114013671875,
0.7841796875,
0.114013671875,
0.7841796875,
0.1318359375,
0.7001953125,
0.1318359375
],
[
0.51708984375,
0.1175537109375,
0.5556640625,
0.1175537109375,
0.5556640625,
0.1300048828125,
0.51708984375,
0.1300048828125
],
[
0.466796875,
0.1175537109375,
0.50537109375,
0.1175537109375,
0.50537109375,
0.1300048828125,
0.466796875,
0.1300048828125
],
[
0.64990234375,
0.1175537109375,
0.69189453125,
0.1175537109375,
0.69189453125,
0.1312255859375,
0.64990234375,
0.1312255859375
],
[
0.791015625,
0.1175537109375,
0.82861328125,
0.1175537109375,
0.82861328125,
0.1300048828125,
0.791015625,
0.1300048828125
],
[
0.8388671875,
0.1175537109375,
0.8759765625,
0.1175537109375,
0.8759765625,
0.1300048828125,
0.8388671875,
0.1300048828125
],
[
0.42822265625,
0.1181640625,
0.43994140625,
0.1181640625,
0.43994140625,
0.1300048828125,
0.42822265625,
0.1300048828125
],
[
0.124267578125,
0.1181640625,
0.359375,
0.1181640625,
0.359375,
0.12890625,
0.124267578125,
0.12890625
],
[
0.466796875,
0.133056640625,
0.50732421875,
0.133056640625,
0.50732421875,
0.1484375,
0.466796875,
0.1484375
],
[
0.56494140625,
0.1318359375,
0.60205078125,
0.1318359375,
0.60205078125,
0.1478271484375,
0.56494140625,
0.1478271484375
],
[
0.51708984375,
0.1341552734375,
0.55517578125,
0.1341552734375,
0.55517578125,
0.146728515625,
0.51708984375,
0.146728515625
],
[
0.1234130859375,
0.1341552734375,
0.36279296875,
0.1341552734375,
0.36279296875,
0.1478271484375,
0.1234130859375,
0.1478271484375
],
[
0.4248046875,
0.1353759765625,
0.43994140625,
0.1353759765625,
0.43994140625,
0.146728515625,
0.4248046875,
0.146728515625
],
[
0.71728515625,
0.1353759765625,
0.72705078125,
0.1353759765625,
0.72705078125,
0.146728515625,
0.71728515625,
0.146728515625
],
[
0.8095703125,
0.135986328125,
0.8154296875,
0.1383056640625,
0.81005859375,
0.1455078125,
0.80419921875,
0.144287109375
],
[
0.6708984375,
0.1383056640625,
0.67822265625,
0.1383056640625,
0.67822265625,
0.144287109375,
0.6708984375,
0.144287109375
],
[
0.623046875,
0.1383056640625,
0.62890625,
0.1383056640625,
0.62890625,
0.144287109375,
0.623046875,
0.144287109375
],
[
0.759765625,
0.14013671875,
0.76416015625,
0.14013671875,
0.76416015625,
0.1455078125,
0.759765625,
0.1455078125
],
[
0.85400390625,
0.14013671875,
0.85986328125,
0.14013671875,
0.85986328125,
0.1455078125,
0.85400390625,
0.1455078125
],
[
0.424072265625,
0.1484375,
0.441650390625,
0.1484375,
0.441650390625,
0.163330078125,
0.424072265625,
0.163330078125
],
[
0.6484375,
0.1484375,
0.69287109375,
0.1484375,
0.69287109375,
0.163330078125,
0.6484375,
0.163330078125
],
[
0.7021484375,
0.1484375,
0.73876953125,
0.1484375,
0.73876953125,
0.1644287109375,
0.7021484375,
0.1644287109375
],
[
0.60693359375,
0.149658203125,
0.64404296875,
0.149658203125,
0.64404296875,
0.162109375,
0.60693359375,
0.162109375
],
[
0.124267578125,
0.149658203125,
0.385498046875,
0.149658203125,
0.385498046875,
0.1644287109375,
0.124267578125,
0.1644287109375
],
[
0.52880859375,
0.1519775390625,
0.54150390625,
0.1519775390625,
0.54150390625,
0.162109375,
0.52880859375,
0.162109375
],
[
0.80615234375,
0.1519775390625,
0.81201171875,
0.1519775390625,
0.81201171875,
0.1591796875,
0.80615234375,
0.1591796875
],
[
0.759765625,
0.152587890625,
0.76416015625,
0.152587890625,
0.76416015625,
0.160888671875,
0.759765625,
0.160888671875
],
[
0.483642578125,
0.152587890625,
0.489501953125,
0.152587890625,
0.489501953125,
0.1602783203125,
0.483642578125,
0.1602783203125
],
[
0.85400390625,
0.152587890625,
0.85986328125,
0.152587890625,
0.85986328125,
0.160888671875,
0.85400390625,
0.160888671875
],
[
0.466796875,
0.1650390625,
0.50732421875,
0.1650390625,
0.50732421875,
0.18115234375,
0.466796875,
0.18115234375
],
[
0.515625,
0.1650390625,
0.5556640625,
0.1650390625,
0.5556640625,
0.18115234375,
0.515625,
0.18115234375
],
[
0.56494140625,
0.1650390625,
0.603515625,
0.1650390625,
0.603515625,
0.18115234375,
0.56494140625,
0.18115234375
],
[
0.426513671875,
0.166259765625,
0.43994140625,
0.166259765625,
0.43994140625,
0.1787109375,
0.426513671875,
0.1787109375
],
[
0.124267578125,
0.166259765625,
0.301513671875,
0.166259765625,
0.301513671875,
0.179931640625,
0.124267578125,
0.179931640625
],
[
0.67236328125,
0.16748046875,
0.68115234375,
0.17333984375,
0.67333984375,
0.1787109375,
0.6650390625,
0.1728515625
],
[
0.8037109375,
0.16748046875,
0.8154296875,
0.16748046875,
0.8154296875,
0.1787109375,
0.8037109375,
0.1787109375
],
[
0.8505859375,
0.16748046875,
0.8623046875,
0.16748046875,
0.8623046875,
0.1787109375,
0.8505859375,
0.1787109375
],
[
0.75146484375,
0.168701171875,
0.76318359375,
0.1644287109375,
0.771484375,
0.17578125,
0.759765625,
0.179931640625
],
[
0.623046875,
0.169189453125,
0.62890625,
0.169189453125,
0.62890625,
0.177001953125,
0.623046875,
0.177001953125
],
[
0.71533203125,
0.168701171875,
0.72705078125,
0.168701171875,
0.72705078125,
0.1787109375,
0.71533203125,
0.1787109375
],
[
0.466796875,
0.18115234375,
0.50537109375,
0.18115234375,
0.50537109375,
0.196533203125,
0.466796875,
0.196533203125
],
[
0.5146484375,
0.18115234375,
0.55517578125,
0.18115234375,
0.55517578125,
0.196533203125,
0.5146484375,
0.196533203125
],
[
0.56494140625,
0.179931640625,
0.646484375,
0.179931640625,
0.646484375,
0.19775390625,
0.56494140625,
0.19775390625
],
[
0.6533203125,
0.18115234375,
0.69287109375,
0.18115234375,
0.69287109375,
0.196533203125,
0.6533203125,
0.196533203125
],
[
0.7021484375,
0.18115234375,
0.7421875,
0.18115234375,
0.7421875,
0.196533203125,
0.7021484375,
0.196533203125
],
[
0.76318359375,
0.18115234375,
0.771484375,
0.18701171875,
0.76318359375,
0.193603515625,
0.75390625,
0.18701171875
],
[
0.426513671875,
0.1817626953125,
0.43994140625,
0.1817626953125,
0.43994140625,
0.19482421875,
0.426513671875,
0.19482421875
],
[
0.124267578125,
0.1817626953125,
0.3525390625,
0.1817626953125,
0.3525390625,
0.196533203125,
0.124267578125,
0.196533203125
],
[
0.8037109375,
0.1817626953125,
0.8154296875,
0.1817626953125,
0.8154296875,
0.19482421875,
0.8037109375,
0.19482421875
],
[
0.8505859375,
0.18408203125,
0.8623046875,
0.18408203125,
0.8623046875,
0.19482421875,
0.8505859375,
0.19482421875
],
[
0.56494140625,
0.196533203125,
0.6455078125,
0.196533203125,
0.6455078125,
0.21435546875,
0.56494140625,
0.21435546875
],
[
0.7001953125,
0.196533203125,
0.7841796875,
0.196533203125,
0.7841796875,
0.21435546875,
0.7001953125,
0.21435546875
],
[
0.6533203125,
0.19775390625,
0.69287109375,
0.19775390625,
0.69287109375,
0.213134765625,
0.6533203125,
0.213134765625
],
[
0.1234130859375,
0.198974609375,
0.38037109375,
0.198974609375,
0.38037109375,
0.213134765625,
0.1234130859375,
0.213134765625
],
[
0.466796875,
0.19775390625,
0.50732421875,
0.19775390625,
0.50732421875,
0.213134765625,
0.466796875,
0.213134765625
],
[
0.515625,
0.1995849609375,
0.55322265625,
0.1995849609375,
0.55322265625,
0.2120361328125,
0.515625,
0.2120361328125
],
[
0.791015625,
0.1995849609375,
0.82861328125,
0.1995849609375,
0.82861328125,
0.2120361328125,
0.791015625,
0.2120361328125
],
[
0.83447265625,
0.198974609375,
0.8759765625,
0.196533203125,
0.87744140625,
0.2120361328125,
0.83642578125,
0.21435546875
],
[
0.12176513671875,
0.213134765625,
0.31982421875,
0.21435546875,
0.31982421875,
0.2286376953125,
0.12176513671875,
0.22802734375
],
[
0.426513671875,
0.20068359375,
0.43994140625,
0.20068359375,
0.43994140625,
0.2120361328125,
0.426513671875,
0.2120361328125
],
[
0.56494140625,
0.213134765625,
0.603515625,
0.213134765625,
0.603515625,
0.2286376953125,
0.56494140625,
0.2286376953125
],
[
0.7021484375,
0.213134765625,
0.74462890625,
0.213134765625,
0.74462890625,
0.2286376953125,
0.7021484375,
0.2286376953125
],
[
0.7421875,
0.21435546875,
0.78271484375,
0.21435546875,
0.78271484375,
0.2298583984375,
0.7421875,
0.2298583984375
],
[
0.7900390625,
0.21435546875,
0.82958984375,
0.21435546875,
0.82958984375,
0.2298583984375,
0.7900390625,
0.2298583984375
],
[
0.51708984375,
0.215576171875,
0.55517578125,
0.215576171875,
0.55517578125,
0.22802734375,
0.51708984375,
0.22802734375
],
[
0.468505859375,
0.215576171875,
0.50537109375,
0.215576171875,
0.50537109375,
0.22802734375,
0.468505859375,
0.22802734375
],
[
0.603515625,
0.215576171875,
0.6455078125,
0.215576171875,
0.6455078125,
0.2286376953125,
0.603515625,
0.2286376953125
],
[
0.654296875,
0.215576171875,
0.69189453125,
0.215576171875,
0.69189453125,
0.22802734375,
0.654296875,
0.22802734375
],
[
0.8369140625,
0.215576171875,
0.875,
0.215576171875,
0.875,
0.22802734375,
0.8369140625,
0.22802734375
],
[
0.5634765625,
0.2286376953125,
0.646484375,
0.2286376953125,
0.646484375,
0.2464599609375,
0.5634765625,
0.2464599609375
],
[
0.426513671875,
0.2161865234375,
0.43994140625,
0.2161865234375,
0.43994140625,
0.22802734375,
0.426513671875,
0.22802734375
],
[
0.466796875,
0.2298583984375,
0.50732421875,
0.2298583984375,
0.50732421875,
0.2452392578125,
0.466796875,
0.2452392578125
],
[
0.515625,
0.2298583984375,
0.5556640625,
0.2298583984375,
0.5556640625,
0.2452392578125,
0.515625,
0.2452392578125
],
[
0.6533203125,
0.2298583984375,
0.6943359375,
0.2298583984375,
0.6943359375,
0.2452392578125,
0.6533203125,
0.2452392578125
],
[
0.7021484375,
0.2298583984375,
0.74462890625,
0.2298583984375,
0.74462890625,
0.2452392578125,
0.7021484375,
0.2452392578125
],
[
0.7421875,
0.2298583984375,
0.78271484375,
0.2298583984375,
0.78271484375,
0.2452392578125,
0.7421875,
0.2452392578125
],
[
0.7900390625,
0.2298583984375,
0.8310546875,
0.2298583984375,
0.8310546875,
0.2452392578125,
0.7900390625,
0.2452392578125
],
[
0.83447265625,
0.2298583984375,
0.8759765625,
0.2298583984375,
0.8759765625,
0.2452392578125,
0.83447265625,
0.2452392578125
],
[
0.426513671875,
0.232177734375,
0.441650390625,
0.232177734375,
0.441650390625,
0.243408203125,
0.426513671875,
0.243408203125
],
[
0.1234130859375,
0.23095703125,
0.297119140625,
0.23095703125,
0.297119140625,
0.24462890625,
0.1234130859375,
0.24462890625
],
[
0.5146484375,
0.2452392578125,
0.55517578125,
0.2452392578125,
0.55517578125,
0.26123046875,
0.5146484375,
0.26123046875
],
[
0.56494140625,
0.2452392578125,
0.60791015625,
0.2452392578125,
0.60791015625,
0.26123046875,
0.56494140625,
0.26123046875
],
[
0.60107421875,
0.2452392578125,
0.6455078125,
0.2452392578125,
0.6455078125,
0.26123046875,
0.60107421875,
0.26123046875
],
[
0.7001953125,
0.24462890625,
0.7841796875,
0.24462890625,
0.7841796875,
0.262451171875,
0.7001953125,
0.262451171875
],
[
0.468505859375,
0.2476806640625,
0.50537109375,
0.2476806640625,
0.50537109375,
0.260009765625,
0.468505859375,
0.260009765625
],
[
0.124267578125,
0.2464599609375,
0.355224609375,
0.2464599609375,
0.355224609375,
0.26123046875,
0.124267578125,
0.26123046875
],
[
0.654296875,
0.2476806640625,
0.69287109375,
0.2476806640625,
0.69287109375,
0.260009765625,
0.654296875,
0.260009765625
],
[
0.791015625,
0.2476806640625,
0.82958984375,
0.2476806640625,
0.82958984375,
0.260009765625,
0.791015625,
0.260009765625
],
[
0.83642578125,
0.2476806640625,
0.87744140625,
0.2476806640625,
0.87744140625,
0.26123046875,
0.83642578125,
0.26123046875
],
[
0.515625,
0.260009765625,
0.5556640625,
0.262451171875,
0.55517578125,
0.27783203125,
0.5146484375,
0.275634765625
],
[
0.426513671875,
0.248779296875,
0.43994140625,
0.248779296875,
0.43994140625,
0.260009765625,
0.426513671875,
0.260009765625
],
[
0.124267578125,
0.262451171875,
0.403076171875,
0.262451171875,
0.403076171875,
0.275634765625,
0.124267578125,
0.275634765625
],
[
0.465087890625,
0.262451171875,
0.50732421875,
0.262451171875,
0.50732421875,
0.27783203125,
0.465087890625,
0.27783203125
],
[
0.56494140625,
0.26123046875,
0.603515625,
0.26123046875,
0.603515625,
0.276611328125,
0.56494140625,
0.276611328125
],
[
0.744140625,
0.26318359375,
0.78076171875,
0.26318359375,
0.78076171875,
0.275634765625,
0.744140625,
0.275634765625
],
[
0.791015625,
0.26318359375,
0.82958984375,
0.26318359375,
0.82958984375,
0.275634765625,
0.791015625,
0.275634765625
],
[
0.83642578125,
0.26318359375,
0.875,
0.26318359375,
0.875,
0.275634765625,
0.83642578125,
0.275634765625
],
[
0.66748046875,
0.2666015625,
0.67919921875,
0.2666015625,
0.67919921875,
0.275634765625,
0.66748046875,
0.275634765625
],
[
0.432373046875,
0.26318359375,
0.443359375,
0.270751953125,
0.43408203125,
0.275634765625,
0.4248046875,
0.26953125
],
[
0.43408203125,
0.27783203125,
0.44580078125,
0.28515625,
0.43408203125,
0.295654296875,
0.42236328125,
0.2880859375
],
[
0.465087890625,
0.279052734375,
0.50732421875,
0.279052734375,
0.50732421875,
0.29443359375,
0.465087890625,
0.29443359375
],
[
0.56689453125,
0.279052734375,
0.6455078125,
0.279052734375,
0.6455078125,
0.29345703125,
0.56689453125,
0.29345703125
],
[
0.640625,
0.27978515625,
0.69189453125,
0.27978515625,
0.69189453125,
0.29443359375,
0.640625,
0.29443359375
],
[
0.51708984375,
0.28076171875,
0.55517578125,
0.28076171875,
0.55517578125,
0.29345703125,
0.51708984375,
0.29345703125
],
[
0.1234130859375,
0.27978515625,
0.270263671875,
0.27978515625,
0.270263671875,
0.29345703125,
0.1234130859375,
0.29345703125
],
[
0.705078125,
0.28076171875,
0.78271484375,
0.28076171875,
0.78271484375,
0.292236328125,
0.705078125,
0.292236328125
],
[
0.791015625,
0.28076171875,
0.8271484375,
0.28076171875,
0.8271484375,
0.29345703125,
0.791015625,
0.29345703125
],
[
0.8388671875,
0.28076171875,
0.875,
0.28076171875,
0.875,
0.29345703125,
0.8388671875,
0.29345703125
],
[
0.1158447265625,
0.311279296875,
0.880859375,
0.312255859375,
0.880859375,
0.32666015625,
0.1158447265625,
0.325927734375
],
[
0.118408203125,
0.328857421875,
0.880859375,
0.328857421875,
0.880859375,
0.339111328125,
0.118408203125,
0.339111328125
],
[
0.1158447265625,
0.341552734375,
0.8818359375,
0.341552734375,
0.8818359375,
0.355712890625,
0.1158447265625,
0.355712890625
],
[
0.1158447265625,
0.355712890625,
0.341796875,
0.355712890625,
0.341796875,
0.369384765625,
0.1158447265625,
0.369384765625
],
[
0.5087890625,
0.396728515625,
0.52880859375,
0.391357421875,
0.53564453125,
0.40380859375,
0.5146484375,
0.4091796875
],
[
0.54248046875,
0.3984375,
0.54736328125,
0.3984375,
0.54736328125,
0.40380859375,
0.54248046875,
0.40380859375
],
[
0.5439453125,
0.3955078125,
0.6708984375,
0.3955078125,
0.6708984375,
0.4091796875,
0.5439453125,
0.4091796875
],
[
0.51318359375,
0.42333984375,
0.54248046875,
0.42333984375,
0.54248046875,
0.435791015625,
0.51318359375,
0.435791015625
],
[
0.53076171875,
0.399658203125,
0.54150390625,
0.399658203125,
0.54150390625,
0.402587890625,
0.53076171875,
0.402587890625
],
[
0.5556640625,
0.42333984375,
0.8671875,
0.42333984375,
0.8671875,
0.43701171875,
0.5556640625,
0.43701171875
],
[
0.300537109375,
0.435302734375,
0.355224609375,
0.435302734375,
0.355224609375,
0.449462890625,
0.300537109375,
0.449462890625
],
[
0.391357421875,
0.435302734375,
0.447509765625,
0.435302734375,
0.447509765625,
0.449462890625,
0.391357421875,
0.449462890625
],
[
0.12176513671875,
0.4423828125,
0.2451171875,
0.44482421875,
0.2451171875,
0.458984375,
0.12176513671875,
0.457275390625
],
[
0.35693359375,
0.448974609375,
0.483642578125,
0.45068359375,
0.483642578125,
0.468505859375,
0.35693359375,
0.466064453125
],
[
0.51123046875,
0.446533203125,
0.880859375,
0.446533203125,
0.880859375,
0.460205078125,
0.51123046875,
0.460205078125
],
[
0.118408203125,
0.46728515625,
0.267822265625,
0.468505859375,
0.267822265625,
0.482666015625,
0.118408203125,
0.482177734375
],
[
0.355224609375,
0.468505859375,
0.48193359375,
0.468505859375,
0.48193359375,
0.482177734375,
0.355224609375,
0.482177734375
],
[
0.51318359375,
0.463134765625,
0.8818359375,
0.463134765625,
0.8818359375,
0.476806640625,
0.51318359375,
0.476806640625
],
[
0.307373046875,
0.4697265625,
0.344970703125,
0.4697265625,
0.344970703125,
0.482177734375,
0.307373046875,
0.482177734375
],
[
0.12176513671875,
0.482666015625,
0.2225341796875,
0.485107421875,
0.2225341796875,
0.49951171875,
0.12176513671875,
0.49755859375
],
[
0.307373046875,
0.482666015625,
0.34765625,
0.482666015625,
0.34765625,
0.498779296875,
0.307373046875,
0.498779296875
],
[
0.35693359375,
0.482666015625,
0.441650390625,
0.482666015625,
0.441650390625,
0.49755859375,
0.35693359375,
0.49755859375
],
[
0.51318359375,
0.478515625,
0.88330078125,
0.478515625,
0.88330078125,
0.4921875,
0.51318359375,
0.4921875
],
[
0.435791015625,
0.485107421875,
0.480224609375,
0.485107421875,
0.480224609375,
0.496337890625,
0.435791015625,
0.496337890625
],
[
0.1234130859375,
0.49951171875,
0.219970703125,
0.49951171875,
0.219970703125,
0.51416015625,
0.1234130859375,
0.51416015625
],
[
0.35693359375,
0.49951171875,
0.48193359375,
0.49951171875,
0.48193359375,
0.51416015625,
0.35693359375,
0.51416015625
],
[
0.51318359375,
0.49462890625,
0.8330078125,
0.49462890625,
0.8330078125,
0.50830078125,
0.51318359375,
0.50830078125
],
[
0.30908203125,
0.50048828125,
0.345947265625,
0.50048828125,
0.345947265625,
0.51318359375,
0.30908203125,
0.51318359375
],
[
0.12176513671875,
0.5166015625,
0.298828125,
0.5166015625,
0.298828125,
0.52978515625,
0.12176513671875,
0.52978515625
],
[
0.3046875,
0.51708984375,
0.344970703125,
0.51708984375,
0.344970703125,
0.529296875,
0.3046875,
0.529296875
],
[
0.355224609375,
0.515625,
0.48193359375,
0.515625,
0.48193359375,
0.52978515625,
0.355224609375,
0.52978515625
],
[
0.51318359375,
0.52490234375,
0.54150390625,
0.52490234375,
0.54150390625,
0.53759765625,
0.51318359375,
0.53759765625
],
[
0.55517578125,
0.52392578125,
0.85400390625,
0.52490234375,
0.85400390625,
0.5390625,
0.55517578125,
0.5380859375
],
[
0.5498046875,
0.5390625,
0.591796875,
0.54150390625,
0.59033203125,
0.5556640625,
0.54833984375,
0.55419921875
],
[
0.11419677734375,
0.548828125,
0.48779296875,
0.5498046875,
0.48779296875,
0.56396484375,
0.11419677734375,
0.5634765625
],
[
0.11419677734375,
0.5634765625,
0.4375,
0.56396484375,
0.4375,
0.5791015625,
0.11419677734375,
0.57763671875
],
[
0.51904296875,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.51904296875,
0.57666015625
],
[
0.51318359375,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59326171875,
0.51318359375,
0.59326171875
],
[
0.51123046875,
0.595703125,
0.8818359375,
0.5966796875,
0.8818359375,
0.60986328125,
0.51123046875,
0.609375
],
[
0.51123046875,
0.61083984375,
0.880859375,
0.61083984375,
0.880859375,
0.62451171875,
0.51123046875,
0.62451171875
],
[
0.51318359375,
0.62744140625,
0.8759765625,
0.62744140625,
0.8759765625,
0.642578125,
0.51318359375,
0.642578125
],
[
0.51123046875,
0.64453125,
0.63818359375,
0.64453125,
0.63818359375,
0.65771484375,
0.51123046875,
0.65771484375
],
[
0.130126953125,
0.6728515625,
0.193115234375,
0.6728515625,
0.193115234375,
0.68408203125,
0.130126953125,
0.68408203125
],
[
0.344970703125,
0.6728515625,
0.39794921875,
0.6728515625,
0.39794921875,
0.68408203125,
0.344970703125,
0.68408203125
],
[
0.418212890625,
0.6728515625,
0.47265625,
0.6728515625,
0.47265625,
0.68408203125,
0.418212890625,
0.68408203125
],
[
0.51318359375,
0.6767578125,
0.52490234375,
0.6767578125,
0.52490234375,
0.68701171875,
0.51318359375,
0.68701171875
],
[
0.5458984375,
0.6748046875,
0.646484375,
0.6748046875,
0.646484375,
0.689453125,
0.5458984375,
0.689453125
],
[
0.1292724609375,
0.68603515625,
0.219970703125,
0.6884765625,
0.219970703125,
0.703125,
0.1292724609375,
0.70068359375
],
[
0.3525390625,
0.68701171875,
0.39208984375,
0.68701171875,
0.39208984375,
0.703125,
0.3525390625,
0.703125
],
[
0.426513671875,
0.68701171875,
0.465087890625,
0.68701171875,
0.465087890625,
0.703125,
0.426513671875,
0.703125
],
[
0.130126953125,
0.703125,
0.32568359375,
0.703125,
0.32568359375,
0.71630859375,
0.130126953125,
0.71630859375
],
[
0.353515625,
0.703125,
0.391357421875,
0.703125,
0.391357421875,
0.71875,
0.353515625,
0.71875
],
[
0.426513671875,
0.705078125,
0.46337890625,
0.705078125,
0.46337890625,
0.71728515625,
0.426513671875,
0.71728515625
],
[
0.51123046875,
0.70361328125,
0.60546875,
0.70361328125,
0.60546875,
0.71533203125,
0.51123046875,
0.71533203125
],
[
0.130126953125,
0.72021484375,
0.3046875,
0.72021484375,
0.3046875,
0.73388671875,
0.130126953125,
0.73388671875
],
[
0.353515625,
0.72021484375,
0.391357421875,
0.72021484375,
0.391357421875,
0.73291015625,
0.353515625,
0.73291015625
],
[
0.426513671875,
0.72021484375,
0.46337890625,
0.72021484375,
0.46337890625,
0.73291015625,
0.426513671875,
0.73291015625
],
[
0.1292724609375,
0.7353515625,
0.291259765625,
0.7353515625,
0.291259765625,
0.7490234375,
0.1292724609375,
0.7490234375
],
[
0.353515625,
0.7353515625,
0.39208984375,
0.7353515625,
0.39208984375,
0.75048828125,
0.353515625,
0.75048828125
],
[
0.426513671875,
0.7353515625,
0.465087890625,
0.7353515625,
0.465087890625,
0.75048828125,
0.426513671875,
0.75048828125
],
[
0.51123046875,
0.73388671875,
0.5263671875,
0.73388671875,
0.5263671875,
0.75,
0.51123046875,
0.75
],
[
0.353515625,
0.75048828125,
0.39208984375,
0.75048828125,
0.39208984375,
0.76708984375,
0.353515625,
0.76708984375
],
[
0.426513671875,
0.75048828125,
0.466796875,
0.75048828125,
0.466796875,
0.76708984375,
0.426513671875,
0.76708984375
],
[
0.51123046875,
0.76123046875,
0.599609375,
0.76318359375,
0.599609375,
0.77685546875,
0.51123046875,
0.77490234375
],
[
0.130126953125,
0.7529296875,
0.301513671875,
0.7529296875,
0.301513671875,
0.7666015625,
0.130126953125,
0.7666015625
],
[
0.1275634765625,
0.76708984375,
0.276123046875,
0.76953125,
0.276123046875,
0.78369140625,
0.1275634765625,
0.7822265625
],
[
0.3525390625,
0.7685546875,
0.39208984375,
0.7685546875,
0.39208984375,
0.78369140625,
0.3525390625,
0.78369140625
],
[
0.426513671875,
0.7685546875,
0.466796875,
0.7685546875,
0.466796875,
0.78369140625,
0.426513671875,
0.78369140625
],
[
0.5322265625,
0.779296875,
0.74072265625,
0.779296875,
0.74072265625,
0.7919921875,
0.5322265625,
0.7919921875
],
[
0.52880859375,
0.79638671875,
0.880859375,
0.79638671875,
0.880859375,
0.81103515625,
0.52880859375,
0.81103515625
],
[
0.1158447265625,
0.802734375,
0.48779296875,
0.802734375,
0.48779296875,
0.81640625,
0.1158447265625,
0.81640625
],
[
0.1158447265625,
0.81787109375,
0.486083984375,
0.81787109375,
0.486083984375,
0.83056640625,
0.1158447265625,
0.83056640625
],
[
0.51123046875,
0.8134765625,
0.8818359375,
0.8134765625,
0.8818359375,
0.82666015625,
0.51123046875,
0.82666015625
],
[
0.11419677734375,
0.830078125,
0.489501953125,
0.83056640625,
0.489501953125,
0.845703125,
0.11419677734375,
0.84423828125
],
[
0.51123046875,
0.8291015625,
0.880859375,
0.8291015625,
0.880859375,
0.8427734375,
0.51123046875,
0.8427734375
],
[
0.1158447265625,
0.845703125,
0.490234375,
0.845703125,
0.490234375,
0.859375,
0.1158447265625,
0.859375
],
[
0.51318359375,
0.84423828125,
0.85400390625,
0.84423828125,
0.85400390625,
0.859375,
0.51318359375,
0.859375
],
[
0.1158447265625,
0.85986328125,
0.405517578125,
0.85986328125,
0.405517578125,
0.87353515625,
0.1158447265625,
0.87353515625
],
[
0.509765625,
0.8759765625,
0.5283203125,
0.8759765625,
0.5283203125,
0.89013671875,
0.509765625,
0.89013671875
],
[
0.509765625,
0.90283203125,
0.66748046875,
0.90380859375,
0.66748046875,
0.91845703125,
0.509765625,
0.91748046875
]
] |
[
"Known",
"MNLI",
"FEVER",
"QQP",
"Method",
"HANS Gap",
"PAWS Gap",
"Gap",
"biases",
"symm.",
"dev",
"dev",
"dev",
"22.1",
"85.6",
"22.591.0",
"62.4",
"84.5",
"63.1",
"33.5",
"57.5",
"X",
"BERT-base (Devlin et al., 2019)",
"83.5",
"14.3",
"69.2",
"Reweighting (Clark et al., 2019)",
"V",
"",
"1",
"-",
"",
"",
"-",
"V",
"66.5",
"18.1",
"84.6",
"Reweighting (Schuster et al., 2019)",
"-",
"I",
"",
"1",
"I",
"82.9",
"67.9",
"15.0",
"Y",
"PoE (Clark et al., 2019)",
".",
"-",
"-",
"、",
"",
"-",
"84.3",
"69.1",
"15.786.4",
"66.2",
"20.2",
"",
"1",
"Conf-reg (Utama et al., 2020a)",
"-",
"-",
"12.687.1",
"21.685.2",
"65.5",
"Reweighting (Utama et al., 2020b)",
"82.3",
"69.7",
"57.4",
"27.8",
"PoE (Utama et al., 2020b)",
"X",
"15.1",
"20.1",
"86.1",
"56.3",
"66.8",
"81.9",
"85.9",
"65.8",
"29.8",
"15.484.8",
"X",
"83.3",
"67.9",
"65.7",
"19.1",
"88.0",
"46.4",
"41.6",
"X",
"PoE (Sanh et al., 2021)",
"67.1",
"17.2",
"87.6",
"21.689.0",
"84.3",
"Conf-reg (Utama et al., 2020b)",
"66.0",
"43.0",
"46.0",
"71.2",
"",
"Self-Debiasing (Ghaddar et al., 2021)",
"83.2",
"12.0",
"90.2",
"46.5",
"43.7",
"-",
"+",
"+",
"82.7",
"10.185.6",
"68.1",
"72.6",
"Bias Experts (ours)",
"17.586.8",
"58.1",
"28.7",
"Table l: Performances of models evaluated on MNLI, FEVER, QQP, and their corresponding challenge test sets.",
"Wealsoreportthedifferencebetweenin-andout-of-distributionperformances asGap.Theresultsof baselines",
"using prior knowledge of biases are from the original paper. We mark the best and the second-best performance in",
"bold and underline , respectively.",
"5",
"",
"Related Work",
"5.1",
"",
"Debiasing with prior knowledge of bias",
"MNLI",
"HANS",
"Ablative Setting",
"Ent NEnt Avg",
"NLI (He et al., 2019; Mahabadi et al., 2020)",
"Bias Experts (ours)",
"93.5 51.8 72.6",
"VQA (Cadene et al., 2019) (Bahng et al., 2020)",
"82.7",
"(1) w/o Amp.",
"82.8",
"94.9 45.0",
"(Schuster et al., 2019) (Utama et al., 2020a; Ma-",
"70.0",
"(2) w/o OvR",
"94.4 42.7 68.5",
"habadi et al., 2020) PoE (Clark et al., 2019)",
"83.0",
"(3) w/o Amp. and OvR",
"83.1",
"94.5 41.4 67.9",
"5.2",
"Debiasing without targeting a specific",
"bias",
"Table 2: Ablation studies using accuracy on MNLI and",
"HANS. We mark the best performance in bold .",
"(Utama et al., 2020b) (Bras et al., 2020; Clark",
"et al., 2020; Sanh et al., 2021; Yaghoobzadeh et al.,",
"2021; Ghaddar et al., 2021)GCE (Nam et al.,",
"2020) (Liu et al., 2021) Kim et al. (2022) Ahn",
"et al. (2023) Lee et al. (2021) Lee et al. (2023)",
"Lyu et al. (2023)",
"Method",
"MNLI",
"HANS",
"6",
"Conclusion",
"BERT-base",
"84.5",
"62.4",
"PoE (Utama et al.. 2020b)",
"81.9",
"66.8",
"One-vs-Rest",
"PoE (Sanh et al., 2021)",
"83.3",
"67.9",
"PoE + Ours (T = 2k)",
"83.8",
"68.0",
"7",
"83.3",
"71.0",
"Limitations",
"PoE + Ours (T = 10k)",
"Bias Experts (ours)",
"82.7",
"72.6",
"1NLUNLI2-322-54-20NLU",
"2 Utama et al. (2020a) A1 2 Q Wen et al.",
"Table 3: Trade-off results between in- and out-of-",
"distribution performances evaluated on MNLI and",
"(2022) /i =(k-1)/k2=1/k k入1 入2 α",
"HANS. T denotes the number of training steps for bias",
"HANS099 α 110-880 (Utama et al., 2020b; Sanh",
"experts. We mark the best and the second-best perfor-",
"et al., 2021; Kim et al., 2022; Lyu et al., 2023)",
"mance in bold and underline, respectively.",
"8",
"Ethics Statement ie)"
] |
{
"arxiv_id": "2312.03577",
"language": "zh",
"page_idx": 4
}
|
||
1905.10726-zh-0
|
1905.10726-zh-0.jpg
|
SemBleu : AMR 解析评估的稳健度量
Linfeng Song and Daniel Gildea Department of Computer Science
University of Rochester Rochester, NY 14627
Abstract
Evaluating AMR parsing accuracy involves comparing pairs of AMR graphs. The major evaluation metric, Smatch (Cai and Knight, 2013) , searches for one-toone mappings between the nodes of two AMRs with a greedy hill-climbing algorithm, which leads to search errors. We propose SemBleu , a robust metric that extends Bleu (Papineni et al.,2002) to AMRs. It does not suffer from search errors and considers non-local correspondences in addition to local ones. SemBleu is fully content-driven and punishes situations where a system's output does not preserve most information from the input. Preliminary experiments on both sentence and corpus levels show that SemBleu has slightly higher consistency with human judgments than Smatch . Our code is available at http://github.com/ freesunshine0316/sembleu .
## 1 Introduction
抽象含义表示(AMR)(Banarescu et al., 2013) 是一种语义形式主义，其中句子的含义被编码为有根的有向图。图 1 显示了两个 AMR 图， 其中节点(例如“girl”和“leave-11”)表示AMR 概念，边缘(例如“ARG0”和“ARG1”) 表示概念之间的关系。由于对更好的自然语言理解的需求，将句子解析为 AMR 的任务越来越受到关注
尽管在 AMR 解析 (Flanigan et al., 2014; Artzi et al., 2015; Pust et al., 2015; Peng et al., 2015; Buys and Blunsom, 2017; Konstas et al., 2017; Wang and Xue, 2017; Ballesteros and Al-Onaizan, 2017; Lyu and Titov, 2018; Peng et al., 2018; Groschwitz et al., 2018; Guo and Lu, 2018) 上进行了大量工作，但很少关注评估解析结果，而将 Smatch (Cai and Knight, 2013) 作为唯一的整体性能指标。Damonte
<FigureHere>
Figure 1: Two AMR examples meaning“The girl asked the boy to leave." and“The woman made two pies.", respectively. Their Smatch score is 25 ( $\%$ ).
<FigureHere>
Figure 2: Average, minimal and maximal Smatch scores over 100 runs on 100 sentences. The running time increases from 6.6 seconds $(\begin{array}{c}r=1)\text{ to 21.0}\end{array}\bar{(}$ $r= 4$ ).
et al.(2017)基于 Smatch 的节点映射开发
了一套细粒度的性能度量(见下文)
Smatch 有两个主要缺点：首先，它基于贪梦的爬山来找到两个 AMR 之间的一对一节点映射 (找到确切的最佳映射是 NP 完全的)。搜索错误削弱了其作为度量的鲁棒性。为了增强鲁棒性，爬山搜索通过随机重启执行多次。这降低了效率，更重要的是，不能消除搜索错误。图 2 显示了 Smatch 分数的均值和错误界限， 它是 100 个句子上 100 次运行的$r$重新启动次数的函数。我们可以看到，当$r$很大时，方差仍然很大。此外，通过与其他研究人员的通信，我们了解到，以前关于 AMR 解析的论文使用$r$的不同值来报告 Smatch 分数
另一个问题是 Smatch 将一个节点映射到另
|
[
[
0.300537109375,
0.08489990234375,
0.7001953125,
0.08489990234375,
0.7001953125,
0.09979248046875,
0.300537109375,
0.09979248046875
],
[
0.36279296875,
0.13720703125,
0.646484375,
0.13720703125,
0.646484375,
0.1519775390625,
0.36279296875,
0.1519775390625
],
[
0.359375,
0.15380859375,
0.64990234375,
0.15380859375,
0.64990234375,
0.16748046875,
0.359375,
0.16748046875
],
[
0.403076171875,
0.169189453125,
0.60546875,
0.169189453125,
0.60546875,
0.182861328125,
0.403076171875,
0.182861328125
],
[
0.41064453125,
0.185302734375,
0.59765625,
0.185302734375,
0.59765625,
0.203125,
0.41064453125,
0.203125
],
[
0.2626953125,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.27978515625,
0.2626953125,
0.27783203125
],
[
0.59033203125,
0.270751953125,
0.62060546875,
0.270751953125,
0.62060546875,
0.279052734375,
0.59033203125,
0.279052734375
],
[
0.76318359375,
0.26953125,
0.81201171875,
0.26953125,
0.81201171875,
0.27978515625,
0.76318359375,
0.27978515625
],
[
0.5380859375,
0.283935546875,
0.57666015625,
0.2861328125,
0.5751953125,
0.297607421875,
0.53662109375,
0.295654296875
],
[
0.64404296875,
0.2861328125,
0.67919921875,
0.2861328125,
0.67919921875,
0.297607421875,
0.64404296875,
0.297607421875
],
[
0.71728515625,
0.287353515625,
0.75146484375,
0.287353515625,
0.75146484375,
0.295654296875,
0.71728515625,
0.295654296875
],
[
0.8193359375,
0.2861328125,
0.85546875,
0.2861328125,
0.85546875,
0.297607421875,
0.8193359375,
0.297607421875
],
[
0.14697265625,
0.295654296875,
0.462646484375,
0.295654296875,
0.462646484375,
0.308837890625,
0.14697265625,
0.308837890625
],
[
0.53662109375,
0.305908203125,
0.57861328125,
0.305908203125,
0.57861328125,
0.317626953125,
0.53662109375,
0.317626953125
],
[
0.62890625,
0.305908203125,
0.68701171875,
0.305908203125,
0.68701171875,
0.31640625,
0.62890625,
0.31640625
],
[
0.7109375,
0.305908203125,
0.76318359375,
0.305908203125,
0.76318359375,
0.317626953125,
0.7109375,
0.317626953125
],
[
0.81689453125,
0.305908203125,
0.8505859375,
0.305908203125,
0.8505859375,
0.317626953125,
0.81689453125,
0.317626953125
],
[
0.5751953125,
0.308837890625,
0.58447265625,
0.308837890625,
0.58447265625,
0.314208984375,
0.5751953125,
0.314208984375
],
[
0.42236328125,
0.30810546875,
0.459228515625,
0.31005859375,
0.45751953125,
0.322509765625,
0.420654296875,
0.320556640625
],
[
0.80419921875,
0.308837890625,
0.8154296875,
0.305908203125,
0.8193359375,
0.311279296875,
0.8095703125,
0.314208984375
],
[
0.1435546875,
0.32470703125,
0.4609375,
0.322509765625,
0.4609375,
0.336181640625,
0.1435546875,
0.33837890625
],
[
0.14697265625,
0.31005859375,
0.41650390625,
0.31005859375,
0.41650390625,
0.32373046875,
0.14697265625,
0.32373046875
],
[
0.63037109375,
0.330810546875,
0.640625,
0.330810546875,
0.640625,
0.334228515625,
0.63037109375,
0.334228515625
],
[
0.8017578125,
0.330078125,
0.82275390625,
0.330078125,
0.82275390625,
0.33740234375,
0.8017578125,
0.33740234375
],
[
0.14697265625,
0.33837890625,
0.4609375,
0.33837890625,
0.4609375,
0.352783203125,
0.14697265625,
0.352783203125
],
[
0.14697265625,
0.352783203125,
0.4609375,
0.352783203125,
0.4609375,
0.366455078125,
0.14697265625,
0.366455078125
],
[
0.6337890625,
0.349853515625,
0.6826171875,
0.349853515625,
0.6826171875,
0.35986328125,
0.6337890625,
0.35986328125
],
[
0.60107421875,
0.355712890625,
0.615234375,
0.355712890625,
0.615234375,
0.365234375,
0.60107421875,
0.365234375
],
[
0.779296875,
0.35693359375,
0.79248046875,
0.35693359375,
0.79248046875,
0.363525390625,
0.779296875,
0.363525390625
],
[
0.14697265625,
0.366455078125,
0.456787109375,
0.366455078125,
0.456787109375,
0.380126953125,
0.14697265625,
0.380126953125
],
[
0.145263671875,
0.380126953125,
0.4609375,
0.380126953125,
0.4609375,
0.393798828125,
0.145263671875,
0.393798828125
],
[
0.51318359375,
0.374755859375,
0.88330078125,
0.3759765625,
0.88330078125,
0.39013671875,
0.51318359375,
0.3896484375
],
[
0.1435546875,
0.394287109375,
0.459228515625,
0.393798828125,
0.459228515625,
0.407958984375,
0.1435546875,
0.4091796875
],
[
0.51318359375,
0.39013671875,
0.88330078125,
0.39013671875,
0.88330078125,
0.40380859375,
0.51318359375,
0.40380859375
],
[
0.14697265625,
0.4091796875,
0.4609375,
0.4091796875,
0.4609375,
0.4228515625,
0.14697265625,
0.4228515625
],
[
0.51123046875,
0.405029296875,
0.88330078125,
0.405029296875,
0.88330078125,
0.418701171875,
0.51123046875,
0.418701171875
],
[
0.51123046875,
0.416259765625,
0.5517578125,
0.418701171875,
0.54833984375,
0.43408203125,
0.509765625,
0.43115234375
],
[
0.14697265625,
0.4228515625,
0.4609375,
0.4228515625,
0.4609375,
0.435791015625,
0.14697265625,
0.435791015625
],
[
0.14697265625,
0.43701171875,
0.459228515625,
0.43701171875,
0.459228515625,
0.45068359375,
0.14697265625,
0.45068359375
],
[
0.14697265625,
0.451904296875,
0.41064453125,
0.451904296875,
0.41064453125,
0.465576171875,
0.14697265625,
0.465576171875
],
[
0.414794921875,
0.453125,
0.459228515625,
0.453125,
0.459228515625,
0.46435546875,
0.414794921875,
0.46435546875
],
[
0.5458984375,
0.448974609375,
0.57275390625,
0.448974609375,
0.57275390625,
0.46142578125,
0.5458984375,
0.46142578125
],
[
0.14697265625,
0.466064453125,
0.459228515625,
0.466064453125,
0.459228515625,
0.479736328125,
0.14697265625,
0.479736328125
],
[
0.54248046875,
0.4697265625,
0.57275390625,
0.4697265625,
0.57275390625,
0.482177734375,
0.54248046875,
0.482177734375
],
[
0.145263671875,
0.48095703125,
0.4609375,
0.48095703125,
0.4609375,
0.49462890625,
0.145263671875,
0.49462890625
],
[
0.53564453125,
0.48095703125,
0.54833984375,
0.48095703125,
0.54833984375,
0.4912109375,
0.53564453125,
0.4912109375
],
[
0.14697265625,
0.49462890625,
0.459228515625,
0.49462890625,
0.459228515625,
0.50830078125,
0.14697265625,
0.50830078125
],
[
0.53564453125,
0.4921875,
0.56689453125,
0.490478515625,
0.568359375,
0.501953125,
0.53662109375,
0.50439453125
],
[
0.14697265625,
0.5087890625,
0.459228515625,
0.5087890625,
0.459228515625,
0.5224609375,
0.14697265625,
0.5224609375
],
[
0.1485595703125,
0.52490234375,
0.456787109375,
0.52490234375,
0.456787109375,
0.53515625,
0.1485595703125,
0.53515625
],
[
0.14697265625,
0.53759765625,
0.459228515625,
0.53759765625,
0.459228515625,
0.55029296875,
0.14697265625,
0.55029296875
],
[
0.14697265625,
0.55029296875,
0.411376953125,
0.55029296875,
0.411376953125,
0.56396484375,
0.14697265625,
0.56396484375
],
[
0.42822265625,
0.5517578125,
0.459228515625,
0.5517578125,
0.459228515625,
0.5634765625,
0.42822265625,
0.5634765625
],
[
0.54833984375,
0.55859375,
0.56689453125,
0.55859375,
0.56689453125,
0.56640625,
0.54833984375,
0.56640625
],
[
0.14697265625,
0.5654296875,
0.459228515625,
0.5654296875,
0.459228515625,
0.5791015625,
0.14697265625,
0.5791015625
],
[
0.6455078125,
0.56787109375,
0.7841796875,
0.56787109375,
0.7841796875,
0.58056640625,
0.6455078125,
0.58056640625
],
[
0.14697265625,
0.5791015625,
0.349365234375,
0.5791015625,
0.349365234375,
0.5927734375,
0.14697265625,
0.5927734375
],
[
0.51318359375,
0.5888671875,
0.8818359375,
0.5888671875,
0.8818359375,
0.6025390625,
0.51318359375,
0.6025390625
],
[
0.151123046875,
0.60791015625,
0.267822265625,
0.609375,
0.267822265625,
0.62353515625,
0.151123046875,
0.62255859375
],
[
0.51123046875,
0.60400390625,
0.88330078125,
0.60498046875,
0.88330078125,
0.619140625,
0.51123046875,
0.6181640625
],
[
0.1175537109375,
0.61083984375,
0.1318359375,
0.60986328125,
0.1351318359375,
0.62158203125,
0.118408203125,
0.62255859375
],
[
0.51318359375,
0.619140625,
0.88330078125,
0.619140625,
0.88330078125,
0.6328125,
0.51318359375,
0.6328125
],
[
0.12005615234375,
0.6357421875,
0.48779296875,
0.6357421875,
0.48779296875,
0.6494140625,
0.12005615234375,
0.6494140625
],
[
0.5146484375,
0.63525390625,
0.57080078125,
0.63525390625,
0.57080078125,
0.646484375,
0.5146484375,
0.646484375
],
[
0.12005615234375,
0.65185546875,
0.48779296875,
0.65185546875,
0.48779296875,
0.66650390625,
0.12005615234375,
0.66650390625
],
[
0.12176513671875,
0.66845703125,
0.484375,
0.66845703125,
0.484375,
0.681640625,
0.12176513671875,
0.681640625
],
[
0.51318359375,
0.6650390625,
0.8818359375,
0.6650390625,
0.8818359375,
0.6787109375,
0.51318359375,
0.6787109375
],
[
0.12005615234375,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69873046875,
0.12005615234375,
0.69873046875
],
[
0.515625,
0.68115234375,
0.79833984375,
0.68115234375,
0.79833984375,
0.69482421875,
0.515625,
0.69482421875
],
[
0.12176513671875,
0.69970703125,
0.48779296875,
0.69970703125,
0.48779296875,
0.71435546875,
0.12176513671875,
0.71435546875
],
[
0.53076171875,
0.69970703125,
0.8818359375,
0.69970703125,
0.8818359375,
0.71337890625,
0.53076171875,
0.71337890625
],
[
0.12176513671875,
0.71728515625,
0.486083984375,
0.71728515625,
0.486083984375,
0.73095703125,
0.12176513671875,
0.73095703125
],
[
0.5146484375,
0.71435546875,
0.880859375,
0.71435546875,
0.880859375,
0.72802734375,
0.5146484375,
0.72802734375
],
[
0.118408203125,
0.73291015625,
0.48779296875,
0.732421875,
0.48779296875,
0.74658203125,
0.118408203125,
0.74755859375
],
[
0.5146484375,
0.732421875,
0.8818359375,
0.732421875,
0.8818359375,
0.7451171875,
0.5146484375,
0.7451171875
],
[
0.118408203125,
0.74755859375,
0.21240234375,
0.75,
0.21240234375,
0.76416015625,
0.118408203125,
0.7626953125
],
[
0.515625,
0.74755859375,
0.88330078125,
0.74755859375,
0.88330078125,
0.76123046875,
0.515625,
0.76123046875
],
[
0.139404296875,
0.765625,
0.489501953125,
0.765625,
0.489501953125,
0.77978515625,
0.139404296875,
0.77978515625
],
[
0.5146484375,
0.76318359375,
0.880859375,
0.76318359375,
0.880859375,
0.77685546875,
0.5146484375,
0.77685546875
],
[
0.12005615234375,
0.7822265625,
0.490234375,
0.7822265625,
0.490234375,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.515625,
0.78076171875,
0.8818359375,
0.78076171875,
0.8818359375,
0.79443359375,
0.515625,
0.79443359375
],
[
0.12005615234375,
0.79736328125,
0.490234375,
0.79736328125,
0.490234375,
0.8125,
0.12005615234375,
0.8125
],
[
0.51708984375,
0.79638671875,
0.8818359375,
0.79638671875,
0.8818359375,
0.81005859375,
0.51708984375,
0.81005859375
],
[
0.12005615234375,
0.81396484375,
0.490234375,
0.81396484375,
0.490234375,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.515625,
0.8125,
0.8818359375,
0.8125,
0.8818359375,
0.826171875,
0.515625,
0.826171875
],
[
0.12176513671875,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.12176513671875,
0.84423828125
],
[
0.5146484375,
0.82666015625,
0.8818359375,
0.82763671875,
0.8818359375,
0.8427734375,
0.5146484375,
0.84130859375
],
[
0.1175537109375,
0.8466796875,
0.48779296875,
0.8466796875,
0.48779296875,
0.85986328125,
0.1175537109375,
0.85986328125
],
[
0.515625,
0.84423828125,
0.8818359375,
0.84423828125,
0.8818359375,
0.85791015625,
0.515625,
0.85791015625
],
[
0.118408203125,
0.86083984375,
0.489501953125,
0.8623046875,
0.489501953125,
0.876953125,
0.118408203125,
0.8759765625
],
[
0.5146484375,
0.85986328125,
0.8818359375,
0.85986328125,
0.8818359375,
0.87451171875,
0.5146484375,
0.87451171875
],
[
0.12176513671875,
0.87890625,
0.489501953125,
0.87890625,
0.489501953125,
0.892578125,
0.12176513671875,
0.892578125
],
[
0.515625,
0.876953125,
0.81005859375,
0.876953125,
0.81005859375,
0.89013671875,
0.515625,
0.89013671875
],
[
0.12005615234375,
0.89453125,
0.489501953125,
0.89453125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.5341796875,
0.89453125,
0.880859375,
0.89453125,
0.880859375,
0.9091796875,
0.5341796875,
0.9091796875
]
] |
[
"SemBleu：AMR解析评估的稳健度量",
"Linfeng Song and Daniel Gildea",
"Department of Computer Science",
"University of Rochester",
"Rochester, NY 14627",
"Abstract",
"ask-01",
"make-01",
":ARGO",
"ARG1",
"ARGO",
"ARG1",
"Evaluating AMR parsing accuracy involves",
"Cgirl",
"leave-11",
"woman",
"pie",
"",
"The",
"",
"major evaluation metric, Smatch (Cai",
"comparing pairs of AMR graphs.",
"4B",
"quat",
"and Knight, 2013), searches for one-to-",
"one mappings between the nodes of two",
"boy",
"(a)",
"(b)",
"AMRs with a greedy hill-climbing algo-",
"rithm, which leads to search errors. We",
"Figure l: Two AMR examples meaning“The girl",
"propose SemBleu , a robust metric that",
"asked the boy to leave.\"and \"The woman made",
"extends Bleu (Papineni et al., 2002) to",
"two pies.\", respectively. Their Smatch score is 25 (",
"% ).",
"AMRs.It does not suffer from search",
"errors and considers non-local correspon-",
"dences in addition to local ones.",
"Sem-",
"59.4",
"Bleu is fully content-driven and punishes",
"59.2",
"situations where a system's output does",
"",
"not preserve most information from the in-",
"59.0",
"put. Preliminary experiments on both sen-",
"tence and corpus levels show that Sem-",
"Bleu has slightly higher consistency with",
"human judgments than Smatch :",
"Our",
"58.4",
"code is available at http://github.com/",
"Number of random restarts ",
"freesunshine0316/sembleu.",
"Figure 2: Average, minimal and maximal Smatch",
"Introduction",
"scores over 100 runs on 100 sentences. The running",
"1",
"time increases from 6.6 seconds (r =1) to 21.0 (",
"抽象含义表示（AMR）（Banarescuetal.，2013）",
"r =4).",
"是一种语义形式主义，其中句子的含义被编码",
"为有根的有向图。图1显示了两个AMR图",
"etal.（2017）基于Smatch的节点映射开发",
"其中节点（例如“girl”和“leave-11”）表示",
"了一套细粒度的性能度量（见下文）",
"AMR概念，边缘（例如“ARGO”和“ARG1\"）",
"Smatch有两个主要缺点：首先，它基于贪",
"表示概念之间的关系。由于对更好的自然语言",
"婪的爬山来找到两个AMR之间的一对一节点",
"理解的需求，将句子解析为AMR的任务越来",
"映射（找到确切的最佳映射是NP完全的）。搜",
"越受到关注",
"索错误削弱了其作为度量的鲁棒性。为了增强",
"尽管在AMR解析（Flaniganetal.，2014;",
"鲁棒性，爬山搜索通过随机重启执行多次。这",
"Artzi et al., 2015; Pust et al., 2015; Peng et al.,",
"降低了效率，更重要的是，不能消除搜索错误。",
"2015; Buys and Blunsom, 2017; Konstas et al.,",
"图2显示了Smatch分数的均值和错误界限，",
"2017; Wang and Xue, 2017; Ballesteros and",
"它是100个句子上100次运行的r重新启动",
"Al-Onaizan, 2017; Lyu and Titov, 2018; Peng",
"次数的函数。我们可以看到，当『很大时，方",
"et al., 2018: Groschwitz et al., 2018: Guo and",
"差仍然很大。此外，通过与其他研究人员的通",
"Lu，2018）上进行了大量工作，但很少关注评",
"信，我们了解到，以前关于AMR解析的论文",
"估解析结果，而将Smatch（CaiandKnight，",
"使用r的不同值来报告Smatch分数",
"2013）作为唯一的整体性能指标。Damonte",
"另一个问题是Smatch将一个节点映射到另"
] |
{
"arxiv_id": "1905.10726",
"language": "zh",
"page_idx": 0
}
|
||
1905.10726-zh-1
|
1905.10726-zh-1.jpg
|
可以认为两个不同的边是相同的，例如图 1中的“ask-01:ARG1leave-11”和“make-01: ARG1pie" 。这可能会导致两个完全不同的AMR 的分数过大。如图 1 所示，Smatch 给两个 AMR 的分数为 25%, 意思是“女孩让男孩离开”和“女人做了两个馅饼”,它们传达了明显不同的含义。$^{1}$对于边缘标签类型很少的两个不同的 AMR,情况可能会更糟，如果它们之间的所有边缘对都意外匹配，则得分可能达到 50%
为了解决上述问题，我们引人了 SemBleu ,这是一种用于比较 AMR 图的准确指标。SemBleu 扩展了 Bleu (Papineni et al., 2002) , 已被证明可有效评估各种文本生成任务，例如机器翻译和数据到文本的生成。通常，Bleu 分数是通过将预测句子的 $n$ -gram $(n$ 最多 4 个) 与参考句子的$n$ -gram 进行比较而计算得出的精度分数。为了惩罚非常短的预测，它乘以一个简短的惩罚，小于 1。0 表示短于其参考的预测。为了使 Bleu 适用于比较 AMR, 我们将每个 AMR 节点(例如“ask-01”)视为一个单元，并将每对直接连接的 AMR 节点及其关系(例如“ask-01:ARG0girl”)作为一个 bigram。高阶$n$ -gram (例如“ask-01: ARG1leave-11: ARG0boy" ) 以类似的方式定义
SemBleu 比 Smatch 有几个优点。首先，它给出了每对 AMR 的精确分数，没有搜索错误。其次，它是非常有效的计算。在 1368 对AMR 的数据集上，SemBleu 取0。5 秒，而Smatch 使用同一台机器需要近 2 分钟。第三， 除了节点到节点和边到边的对应关系之外，它还捕获高阶关系。这为评估 AMR 解析质量的标准 Smatch 指标提供了补充判断。最后，它不会对代表完全不同含义的 AMR 给予过大的信任
我们的初步评估表明，在语料库级别和句子级别的评估上，SemBleu 比 Smatch 具有更高的人类判断一致性。我们还表明，由 SemBleu 提取的$n$ -克的数量在 AMR 标度中大致是线性的。对几个最近模型的输出的评估表明， SemBleu 在结果排名上与 Smatch 基本一致， 但偶尔会有分歧2 Our metric
我们的方法基于 Bleu , 我们对此进行了简要介绍，然后展示了如何将其扩展为匹配 AMR 图
$^{1}$https://amr.isi.edu/eval/smatch/compare.html
gives more details.
一个节点，而不管其实际内容如何，并且在 2.1 Preliminary knowledge on Bleu
比较两个边时仅考虑边标签。因此，Smatch 如公式 1 所示，候选$c$和参考$z$的 Bleu 分数
是通过将修改后的精度与简洁惩罚($BP$)相
乘来计算的。。
(1)
$$\text{Bleu}=BP\cdot\exp\left(\sum_{k=1}^nw_k\log p_k\right)$$
$BP$定义为$e^\min\{1-\frac{|\boldsymbol{z}|}{|\boldsymbol{c}|},0\}$,该值小于 1。当候选长度( $|c|$ )小于参考长度( $|z|$ )时为$0。p_k$ 和$w_k$是匹配$k$ -gram 的精度和权重，$p_k$定义为
$$p_k=\frac{|kgram(\boldsymbol z)\cap kgram(\boldsymbol c)|}{|kgram(\boldsymbol c)|},$$
(2)
,其中kgram是用于从其输入中提取所有$k$
-gram 的函数
### 2.2 SemBleu
要引入 SemBleu ,我们进行以下更改以使Bleu 适应 AMR 图。首先，我们将每个 AMR $(G$ )的大小定义为节点数加上边数：$|G|=$ $|G.nodes|+|G.edges|$。此大小用于计算简洁惩罚(BP)。直观上，边携带重要的关系信息。此外，我们观察到许多情况下，系统生成的 AMR 保留了引用中的大多数概念，但错过了许多边缘
另一个更改是对$n$ -gram 提取函数(方程式2中的kgram )。AMR 是有向无环图，因此我们开始从根提取$n$ -gram。这类似于开始从句子左端点提取普通的$n$ -gram。请注意，每个$n$ -gram 的顺序仅由其中的节点数决定。例如，“ask-01: ARG0girl”被认为是 bigram,而不是 trigram。我们的$n$ -gram 提取方法采用广度优先遍历来枚举提取$n$ -gram 的每个可能的起始节点。从每个起始节点$p$中，它从$p$开始提取所有可能的$k$ -gram $(1\leq k\leq n$)。在每个节点上，它首先存储当前的$k$ -gram,然后枚举节点的每个后代并继续。以图 1 中的AMR 图为例，用我们的方法提取的$n$ -gram 如表 1 所示
处理逆关系 AMR 的一个重要特征是逆关系，例如“ask-01:ARGOgirl”$\Rightarrow“girl:ARG0-$ ofask-01",以保持扎根和无环的特性。原始关系和反向关系都具有相同的语义含义。在Smatch 之后，在计算 SemBleu 分数之前，我们通过将所有反向关系恢复为原始关系来统一两种类型的关系
效率$\overline{\mathrm{EE}}$F作为一个重要因素，SemBleu 的效率在很大程度上取决于提取的$n$ -克的数量。一个潜在的问题是，对于非常密集的图，可能
|
[
[
0.1351318359375,
0.07659912109375,
0.490234375,
0.07659912109375,
0.490234375,
0.09027099609375,
0.1351318359375,
0.09027099609375
],
[
0.486083984375,
0.07598876953125,
0.8154296875,
0.07659912109375,
0.8154296875,
0.0914306640625,
0.486083984375,
0.09027099609375
],
[
0.12176513671875,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.12176513671875,
0.10687255859375
],
[
0.5146484375,
0.097412109375,
0.8818359375,
0.097412109375,
0.8818359375,
0.11102294921875,
0.5146484375,
0.11102294921875
],
[
0.12176513671875,
0.10986328125,
0.490234375,
0.10986328125,
0.490234375,
0.12353515625,
0.12176513671875,
0.12353515625
],
[
0.5146484375,
0.114013671875,
0.8818359375,
0.114013671875,
0.8818359375,
0.127685546875,
0.5146484375,
0.127685546875
],
[
0.1234130859375,
0.1270751953125,
0.490234375,
0.1270751953125,
0.490234375,
0.13720703125,
0.1234130859375,
0.13720703125
],
[
0.515625,
0.1318359375,
0.62451171875,
0.1318359375,
0.62451171875,
0.1436767578125,
0.515625,
0.1436767578125
],
[
0.12005615234375,
0.14013671875,
0.486083984375,
0.14013671875,
0.486083984375,
0.15380859375,
0.12005615234375,
0.15380859375
],
[
0.118408203125,
0.1561279296875,
0.48779296875,
0.15673828125,
0.48779296875,
0.171630859375,
0.118408203125,
0.17041015625
],
[
0.56494140625,
0.16748046875,
0.7021484375,
0.169189453125,
0.7021484375,
0.18408203125,
0.56494140625,
0.1817626953125
],
[
0.7138671875,
0.168701171875,
0.81787109375,
0.168701171875,
0.81787109375,
0.18701171875,
0.7138671875,
0.18701171875
],
[
0.85791015625,
0.17041015625,
0.880859375,
0.17041015625,
0.880859375,
0.18408203125,
0.85791015625,
0.18408203125
],
[
0.70947265625,
0.1859130859375,
0.744140625,
0.1859130859375,
0.744140625,
0.198974609375,
0.70947265625,
0.198974609375
],
[
0.12005615234375,
0.17333984375,
0.486083984375,
0.17333984375,
0.486083984375,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.12005615234375,
0.190673828125,
0.484375,
0.190673828125,
0.484375,
0.20068359375,
0.12005615234375,
0.20068359375
],
[
0.12176513671875,
0.204833984375,
0.486083984375,
0.204833984375,
0.486083984375,
0.218505859375,
0.12176513671875,
0.218505859375
],
[
0.5087890625,
0.208984375,
0.88671875,
0.210205078125,
0.88671875,
0.235107421875,
0.5087890625,
0.2340087890625
],
[
0.12005615234375,
0.221435546875,
0.489501953125,
0.221435546875,
0.489501953125,
0.235107421875,
0.12005615234375,
0.235107421875
],
[
0.51318359375,
0.232177734375,
0.880859375,
0.2340087890625,
0.880859375,
0.2493896484375,
0.51318359375,
0.2476806640625
],
[
0.12176513671875,
0.2381591796875,
0.48779296875,
0.2381591796875,
0.48779296875,
0.2529296875,
0.12176513671875,
0.2529296875
],
[
0.5146484375,
0.248779296875,
0.8818359375,
0.2493896484375,
0.8818359375,
0.26416015625,
0.5146484375,
0.26318359375
],
[
0.12176513671875,
0.254638671875,
0.214111328125,
0.254638671875,
0.214111328125,
0.2666015625,
0.12176513671875,
0.2666015625
],
[
0.1368408203125,
0.268310546875,
0.490234375,
0.267333984375,
0.490234375,
0.28515625,
0.1368408203125,
0.2861328125
],
[
0.5146484375,
0.267333984375,
0.53662109375,
0.267333984375,
0.53662109375,
0.28076171875,
0.5146484375,
0.28076171875
],
[
0.623046875,
0.27978515625,
0.81201171875,
0.27978515625,
0.81201171875,
0.29443359375,
0.623046875,
0.29443359375
],
[
0.1376953125,
0.287353515625,
0.489501953125,
0.287353515625,
0.489501953125,
0.300537109375,
0.1376953125,
0.300537109375
],
[
0.57666015625,
0.290283203125,
0.62451171875,
0.290283203125,
0.62451171875,
0.3046875,
0.57666015625,
0.3046875
],
[
0.85791015625,
0.289306640625,
0.8818359375,
0.289306640625,
0.8818359375,
0.302734375,
0.85791015625,
0.302734375
],
[
0.12005615234375,
0.302734375,
0.478515625,
0.302734375,
0.478515625,
0.31640625,
0.12005615234375,
0.31640625
],
[
0.67236328125,
0.297607421875,
0.76171875,
0.297607421875,
0.76171875,
0.312255859375,
0.67236328125,
0.312255859375
],
[
0.474365234375,
0.30810546875,
0.483642578125,
0.30810546875,
0.483642578125,
0.3134765625,
0.474365234375,
0.3134765625
],
[
0.12176513671875,
0.320556640625,
0.486083984375,
0.320556640625,
0.486083984375,
0.330810546875,
0.12176513671875,
0.330810546875
],
[
0.5087890625,
0.318359375,
0.88330078125,
0.319580078125,
0.88330078125,
0.334228515625,
0.5087890625,
0.333251953125
],
[
0.118408203125,
0.334228515625,
0.486083984375,
0.334228515625,
0.486083984375,
0.347412109375,
0.118408203125,
0.347412109375
],
[
0.51123046875,
0.336181640625,
0.62646484375,
0.3349609375,
0.62646484375,
0.349853515625,
0.51123046875,
0.350830078125
],
[
0.12005615234375,
0.350830078125,
0.48779296875,
0.349853515625,
0.48779296875,
0.364013671875,
0.12005615234375,
0.365234375
],
[
0.12176513671875,
0.36767578125,
0.48779296875,
0.36767578125,
0.48779296875,
0.38134765625,
0.12176513671875,
0.38134765625
],
[
0.51318359375,
0.3623046875,
0.634765625,
0.3623046875,
0.634765625,
0.377197265625,
0.51318359375,
0.377197265625
],
[
0.12005615234375,
0.383056640625,
0.48779296875,
0.383056640625,
0.48779296875,
0.396728515625,
0.12005615234375,
0.396728515625
],
[
0.515625,
0.383056640625,
0.8818359375,
0.383056640625,
0.8818359375,
0.39794921875,
0.515625,
0.39794921875
],
[
0.12176513671875,
0.399658203125,
0.489501953125,
0.399658203125,
0.489501953125,
0.413330078125,
0.12176513671875,
0.413330078125
],
[
0.5146484375,
0.399658203125,
0.8818359375,
0.399658203125,
0.8818359375,
0.413330078125,
0.5146484375,
0.413330078125
],
[
0.12005615234375,
0.4150390625,
0.48779296875,
0.4150390625,
0.48779296875,
0.4287109375,
0.12005615234375,
0.4287109375
],
[
0.51123046875,
0.4150390625,
0.87890625,
0.4150390625,
0.87890625,
0.4287109375,
0.51123046875,
0.4287109375
],
[
0.12005615234375,
0.431640625,
0.489501953125,
0.431640625,
0.489501953125,
0.4453125,
0.12005615234375,
0.4453125
],
[
0.515625,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.515625,
0.4453125
],
[
0.12005615234375,
0.44775390625,
0.489501953125,
0.44775390625,
0.489501953125,
0.46142578125,
0.12005615234375,
0.46142578125
],
[
0.5146484375,
0.44775390625,
0.880859375,
0.44775390625,
0.880859375,
0.46142578125,
0.5146484375,
0.46142578125
],
[
0.1175537109375,
0.4619140625,
0.491943359375,
0.463134765625,
0.491943359375,
0.47802734375,
0.1175537109375,
0.476806640625
],
[
0.515625,
0.46435546875,
0.88330078125,
0.46435546875,
0.88330078125,
0.47802734375,
0.515625,
0.47802734375
],
[
0.1175537109375,
0.478515625,
0.495361328125,
0.476806640625,
0.495361328125,
0.493408203125,
0.1175537109375,
0.495361328125
],
[
0.515625,
0.479736328125,
0.8818359375,
0.479736328125,
0.8818359375,
0.493408203125,
0.515625,
0.493408203125
],
[
0.12005615234375,
0.495361328125,
0.372802734375,
0.495361328125,
0.372802734375,
0.5087890625,
0.12005615234375,
0.5087890625
],
[
0.5205078125,
0.496337890625,
0.60693359375,
0.496337890625,
0.60693359375,
0.50830078125,
0.5205078125,
0.50830078125
],
[
0.1376953125,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.525390625,
0.1376953125,
0.525390625
],
[
0.5322265625,
0.5126953125,
0.88330078125,
0.5126953125,
0.88330078125,
0.525390625,
0.5322265625,
0.525390625
],
[
0.12005615234375,
0.529296875,
0.489501953125,
0.529296875,
0.489501953125,
0.5419921875,
0.12005615234375,
0.5419921875
],
[
0.51318359375,
0.52783203125,
0.8818359375,
0.52783203125,
0.8818359375,
0.54150390625,
0.51318359375,
0.54150390625
],
[
0.12176513671875,
0.54443359375,
0.489501953125,
0.54443359375,
0.489501953125,
0.55810546875,
0.12176513671875,
0.55810546875
],
[
0.5146484375,
0.54443359375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55810546875,
0.5146484375,
0.55810546875
],
[
0.12005615234375,
0.56103515625,
0.489501953125,
0.56103515625,
0.489501953125,
0.57470703125,
0.12005615234375,
0.57470703125
],
[
0.5146484375,
0.56005859375,
0.880859375,
0.56005859375,
0.880859375,
0.57373046875,
0.5146484375,
0.57373046875
],
[
0.12176513671875,
0.5791015625,
0.489501953125,
0.5791015625,
0.489501953125,
0.5888671875,
0.12176513671875,
0.5888671875
],
[
0.515625,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59130859375,
0.515625,
0.59130859375
],
[
0.12005615234375,
0.59326171875,
0.48779296875,
0.59326171875,
0.48779296875,
0.60693359375,
0.12005615234375,
0.60693359375
],
[
0.5146484375,
0.5927734375,
0.8818359375,
0.59130859375,
0.8818359375,
0.60546875,
0.5146484375,
0.60693359375
],
[
0.118408203125,
0.60791015625,
0.486083984375,
0.60791015625,
0.486083984375,
0.62158203125,
0.118408203125,
0.62158203125
],
[
0.5146484375,
0.609375,
0.8818359375,
0.609375,
0.8818359375,
0.62255859375,
0.5146484375,
0.62255859375
],
[
0.12005615234375,
0.6259765625,
0.48779296875,
0.6259765625,
0.48779296875,
0.6396484375,
0.12005615234375,
0.6396484375
],
[
0.51318359375,
0.62353515625,
0.880859375,
0.62353515625,
0.880859375,
0.63720703125,
0.51318359375,
0.63720703125
],
[
0.12005615234375,
0.64111328125,
0.48779296875,
0.64111328125,
0.48779296875,
0.65478515625,
0.12005615234375,
0.65478515625
],
[
0.5146484375,
0.64111328125,
0.8818359375,
0.64111328125,
0.8818359375,
0.65478515625,
0.5146484375,
0.65478515625
],
[
0.12176513671875,
0.65771484375,
0.156982421875,
0.65771484375,
0.156982421875,
0.67041015625,
0.12176513671875,
0.67041015625
],
[
0.515625,
0.65673828125,
0.88330078125,
0.65673828125,
0.88330078125,
0.67041015625,
0.515625,
0.67041015625
],
[
0.1368408203125,
0.6748046875,
0.48779296875,
0.6748046875,
0.48779296875,
0.6884765625,
0.1368408203125,
0.6884765625
],
[
0.5146484375,
0.67333984375,
0.88330078125,
0.67333984375,
0.88330078125,
0.68701171875,
0.5146484375,
0.68701171875
],
[
0.12176513671875,
0.69140625,
0.486083984375,
0.69140625,
0.486083984375,
0.70166015625,
0.12176513671875,
0.70166015625
],
[
0.51318359375,
0.6884765625,
0.880859375,
0.6884765625,
0.880859375,
0.70166015625,
0.51318359375,
0.70166015625
],
[
0.12176513671875,
0.7080078125,
0.486083984375,
0.7080078125,
0.486083984375,
0.71875,
0.12176513671875,
0.71875
],
[
0.51318359375,
0.70361328125,
0.8818359375,
0.7060546875,
0.8818359375,
0.72021484375,
0.51318359375,
0.71875
],
[
0.118408203125,
0.7216796875,
0.486083984375,
0.7216796875,
0.486083984375,
0.7353515625,
0.118408203125,
0.7353515625
],
[
0.51708984375,
0.72265625,
0.609375,
0.72265625,
0.609375,
0.73388671875,
0.51708984375,
0.73388671875
],
[
0.12005615234375,
0.7392578125,
0.490234375,
0.7392578125,
0.490234375,
0.7529296875,
0.12005615234375,
0.7529296875
],
[
0.5341796875,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.5341796875,
0.75732421875
],
[
0.12005615234375,
0.7548828125,
0.495361328125,
0.7548828125,
0.495361328125,
0.7685546875,
0.12005615234375,
0.7685546875
],
[
0.5146484375,
0.76025390625,
0.8818359375,
0.76025390625,
0.8818359375,
0.77392578125,
0.5146484375,
0.77392578125
],
[
0.118408203125,
0.76953125,
0.2509765625,
0.77099609375,
0.2509765625,
0.78515625,
0.118408203125,
0.78369140625
],
[
0.5146484375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.5146484375,
0.7890625
],
[
0.51708984375,
0.7919921875,
0.88330078125,
0.7919921875,
0.88330078125,
0.8056640625,
0.51708984375,
0.8056640625
],
[
0.1175537109375,
0.7998046875,
0.252685546875,
0.7998046875,
0.252685546875,
0.81396484375,
0.1175537109375,
0.81396484375
],
[
0.5146484375,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.12005615234375,
0.82763671875,
0.48779296875,
0.82763671875,
0.48779296875,
0.84130859375,
0.12005615234375,
0.84130859375
],
[
0.5146484375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.12176513671875,
0.84326171875,
0.489501953125,
0.84326171875,
0.489501953125,
0.85693359375,
0.12176513671875,
0.85693359375
],
[
0.5146484375,
0.84130859375,
0.6455078125,
0.84130859375,
0.6455078125,
0.85498046875,
0.5146484375,
0.85498046875
],
[
0.12176513671875,
0.86083984375,
0.1376953125,
0.86083984375,
0.1376953125,
0.87158203125,
0.12176513671875,
0.87158203125
],
[
0.5341796875,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.5341796875,
0.876953125
],
[
0.139404296875,
0.8818359375,
0.47021484375,
0.88427734375,
0.47021484375,
0.8984375,
0.139404296875,
0.896484375
],
[
0.515625,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.515625,
0.8935546875
],
[
0.12005615234375,
0.89794921875,
0.2435302734375,
0.896484375,
0.2435302734375,
0.90771484375,
0.12005615234375,
0.9091796875
],
[
0.51904296875,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.51904296875,
0.9091796875
]
] |
[
"个节点，而不管其实际内容如何，并且在",
"2.1PreliminaryknowledgeonBleu",
"比较两个边时仅考虑边标签。因此，Smatch",
"如公式1所示，候选c和参考的Bleu分数",
"可以认为两个不同的边是相同的，例如图1",
"是通过将修改后的精度与简洁惩罚（BP）相",
"中的“ask-0l:ARG1leave-11”和“make-0l:",
"乘来计算的。。",
"ARG1pie”。这可能会导致两个完全不同的",
"AMR的分数过大。如图1所示，Smatch给",
"Bleu = BP.exp",
"Zwklog pk",
"(1)",
"(k=1",
"两个AMR的分数为25%，意思是“女孩让男",
"孩离开”和“女人做了两个馅饼”，它们传达",
"了明显不同的含义。1对于边缘标签类型很少",
"BP定义为emin(1-哥,0)，该值小于1。当候",
"的两个不同的AMR，情况可能会更糟，如果",
"选长度（lcl）小于参考长度（lzl）时为0。Pk",
"它们之间的所有边缘对都意外匹配，则得分可",
"和wk是匹配k-gram的精度和权重，Pk定义",
"能达到50%",
"为了解决上述问题，我们引入了SemBleu",
"为",
"[kgram(z) n kgram(c)l",
"这是一种用于比较AMR图的准确指标。",
"Pk =",
"(2)",
"SemBleu 扩展了 Bleu (Papineni et al., 2002)",
"[kgram(c)",
"",
"已被证明可有效评估各种文本生成任务，例如",
"，其中kgram是用于从其输入中提取所有k",
"机器翻译和数据到文本的生成。通常，Bleu分",
"-gram的函数",
"数是通过将预测句子的n-gram（n最多4个）",
"与参考句子的n-gram进行比较而计算得出的",
"2.2 SemBleu",
"精度分数。为了惩罚非常短的预测，它乘以一",
"要引人SemBleu，我们进行以下更改以使",
"个简短的惩罚，小于1。0表示短于其参考的",
"Bleu适应AMR图。首先，我们将每个AMR",
"预测。为了使Bleu适用于比较AMR，我们将",
"（G）的大小定义为节点数加上边数：IG",
"每个AMR节点（例如“ask-01\"）视为一个单",
"|G.nodes|+G.edgesl。此大小用于计算简洁",
"元，并将每对直接连接的AMR节点及其关系",
"惩罚（BP）。直观上，边携带重要的关系信",
"（例如“ask-01：ARG0girl\"）作为一个bigram。",
"息。此外，我们观察到许多情况下，系统生成",
"高阶n-gram（例如“ask-01：ARG1leave-11:",
"的AMR保留了引用中的大多数概念，但错过",
"ARGOboy\"）以类似的方式定义",
"了许多边缘",
"SemBleu比Smatch有几个优点。首先，它",
"另一个更改是对n-gram提取函数（方程式",
"给出了每对AMR的精确分数，没有搜索错",
"2中的kgram）。AMR是有向无环图，因此",
"误。其次，它是非常有效的计算。在1368对",
"我们开始从根提取n-gram。这类似于开始从",
"AMR的数据集上，SemBleu取0。5秒，而",
"句子左端点提取普通的n-gram。请注意，每",
"Smatch使用同一台机器需要近2分钟。第三，",
"个n-gram的顺序仅由其中的节点数决定。例",
"除了节点到节点和边到边的对应关系之外，它",
"如，“ask-01：ARGOgirl”被认为是bigram，而",
"还捕获高阶关系。这为评估AMR解析质量的",
"不是trigram。我们的n-gram提取方法采用",
"标准Smatch指标提供了补充判断。最后，它",
"广度优先遍历来枚举提取n-gram的每个可能",
"不会对代表完全不同含义的AMR给予过大的",
"的起始节点。从每个起始节点p中，它从p开",
"信任",
"始提取所有可能的k-gram（1≤k≤n）。在",
"我们的初步评估表明，在语料库级别和句子",
"每个节点上，它首先存储当前的k-gram，然",
"级别的评估上，SemBleu比Smatch具有更高",
"后枚举节点的每个后代并继续。以图1中的",
"的人类判断一致性。我们还表明，由SemBleu",
"AMR图为例，用我们的方法提取的n-gram",
"提取的n-克的数量在AMR标度中大致是",
"如表1所示",
"线性的。对几个最近模型的输出的评估表明，",
"处理逆关系AMR的一个重要特征是逆关",
"SemBleu在结果排名上与Smatch基本一致，",
"系，例如“ask-01：ARGOgirl\"→“girl：ARG0-",
"但偶尔会有分歧",
"ofask-01”，以保持扎根和无环的特性。原始",
"关系和反向关系都具有相同的语义含义。在",
"2 Our metric",
"Smatch之后，在计算SemBleu分数之前，我",
"我们的方法基于Bleu，我们对此进行了简要",
"们通过将所有反向关系恢复为原始关系来统一",
"介绍，然后展示了如何将其扩展为匹配AMR",
"两种类型的关系",
"图",
"效率EE作为一个重要因素，SemBleu的效",
"\"https://amr.isi.edu/eval/smatch/compare.html",
"率在很大程度上取决于提取的n-克的数量。",
"gives more details.",
"一个潜在的问题是，对于非常密集的图，可能"
] |
{
"arxiv_id": "1905.10726",
"language": "zh",
"page_idx": 1
}
|
||
1905.10726-zh-2
|
1905.10726-zh-2.jpg
|
<table>
<tbody>
<tr>
<th>$Fg$</th>
<th>$\eta_{i}$</th>
<th>Extracted $n$ -grams</th>
</tr>
<tr>
<td> </td>
<td>1 $\begin{array}{c}{1}\\{-}\end{array}$ -</td>
<td>ask-01; girl; leave-11; boy ask- 01$\cdot$ARG0 oirl$\cdot$</td>
</tr>
<tr>
<td>$^{\prime }$a) </td>
<td>2</td>
<td>ask-01:ARG1 leave-11; 10717</td>
</tr>
<tr>
<td> </td>
<td>1 -</td>
<td>$\mathbf{\upsilon}_{v}$ woman; make-01; pie; 2</td>
</tr>
<tr>
<td> </td>
<td>$\bar{3}$ - -</td>
<td>pie:quant 2; $\bar{\mathrm{make-}}\bar{0}\bar{1}^{-}\bar{\mathrm{AR}}\bar{\mathrm{G}}\bar{\mathrm{~bie}}$ ${\mathrm{:quant~}}\bar{2};$</td>
</tr>
</tbody>
</table>
Table 1: $n$ -grams (separated by “,”) extracted from the AMRs in Figure 1 with our extraction algorithm. Fg represents the corresponding subfigure.
有大量提取的$n$ -gram。对于具有$N$节点的全连接图，有$O(N^n)$可能的$n$ -gram。幸运的是，AMR 是非常稀疏的树状图 (Chiang et al., 2018)。对于具有$N$节点的树，$n$ -gram 的数量以$O(n\cdot N)$ 为界，$O(n\cdot N)$在树尺度上是线性的。作为树状图，我们期望从 AMR 中提取的$n$ -gram 的数量在图尺度上大致呈线性。我们的实验从经验上证实了这一期望
2.3 Comparison with Smatch
通常，Smatch 将比较两个 AMR 的问题分解为比较最小单位：节点和边。它将每个 AMR 视为节点和边的袋子，然后计算关于正确映射的节点和边的 F1 分数。给定两个 AMR, Smatch 通过最大化总体 F1 分数来搜索图节点之间的一对一映射，并且边到边映射由节点到节点映射自动确定。由于获得最佳映射是NP 完全的 (通过从子图同构中进行约简),因此它使用贪婪的爬山算法来查找映射，该映射很可能是次优的。一个关键的区别是 SemBleu 通常比 Smatch 考虑更多的全局功能。这两个指标唯一的共同点是节点到节点的对应关系 (也称为 SemBleu 的 unigram)。SemBleu 的每个 bigram 由两个 AMR 节点和一个连接它们的边组成，因此 bigram 已经捕获了比Smatch 更大的上下文。此外 , SemBleu 的高阶$n$ -gram 捕获了更大的对应关系。这可能是一种权衡。通常，更多的高阶匹配表示更好的解析性能，但有时我们希望将部分正确的结果与完全错误的结果区分开来。因此，将 Smatch 与 SemBleu 结合使用可能会给出更全面的判断
另一个区别是确定边(关系)等价的方式。Smatch 仅检查边缘标签，因此具有相同标签
但传达不同含义的两个边缘可以被 Smatch 视为等效。$^{2}$另一方面，SemBleu 不仅考虑边缘标签，还考虑其头部和尾部的内容，如表 1 中提取的$n$ -gram 所示
以图 1 中的 AMR 为例，Smatch 将 (a) 中的“girl”,“ask-01”和“leave-11”映射为(b) 中的“woman”,“make-01”和“pie”。结果，它认为(a)中的“ask-01:ARG0girl”和“ask-01: ARG1leave-11”被正确映射到(b)中的“make- 01: ARG0woman"和“make-01: ARG1pie” 这没有意义。相反，SemBleu 不认为这些边是正确匹配的
## 3 Experiments
我们将 SemBleu 与 Smatch 在来自LDC2015E86 测试集的 100 个句子中的 4 个系统的输出上进行了比较。这些系统是：CAMR, $^{3}{\mathrm{JAMR},}^{4}{\mathrm{Gros}}{\mathrm{(Groschwitz~et~al.,~2018)}}$和Lyu (Lyu and Titov, 2018) 。对于每个句子， 在 Callison-Burch et al.(2010) 之后，注释者决定相对顺序，而不是所有系统的完整顺序。特别是，4 个系统输出被随机分组为 2 对，以进行 2 次比较。对于每对，我们请 3 位注释者决定哪一个更好，并选择多数票作为最终判断。所有注释者在 AMR 相关研究方面都有几年的经验，并且判断是基于他们对系统生成的AMR 保留参考 AMR 含义的印象。在 200 个比较中，注释者完全同意 142 个，占 71 个%。通过判断，我们研究了句子和语料库级别上两个度量的一致性
我们考虑 SemBleu 的所有 unigram, bigram 和 trigram, 并且权重(等式 1 中的$w_k$ s)是等效的(每个为 1/3)。对于句子级别的评估，我们按照以前的工作使用 NIST 几何平滑 (Chen and Cherry, 2014)。在 Smatch 之后，在提取$n$ -gram 进行比较之前，反转诸如“ARG0-of” 之类的反向关系
3.1 Corpus-level experiment
对于语料库级别的比较，我们为每个系统分配一个人类分数，该分数等于系统输出的首选次数
我们的四个系统的人类得分分别为 30、33、 63 和 74。他们的 SemBleu 得分分别为 28、30、 38和 41, Smatch 得分分别为 56、56、63 和67。SemBleu 通常更符合人类的判断。特别是， 对于 Smatch 分数，CAMR 和 JAMR 之间存在联系，而 SemBleu 分数更具区别性。在计
${{}^{2}\text{One examp}}$le is shown in the Smatch tutorial https:
//amr.isi.edu/eval/smatch/tutorial.html.
$^{3}$https://github.com/c-amr/camr $^{4}$https://github.com/jflanigan/jamr
|
[
[
0.130126953125,
0.0748291015625,
0.19140625,
0.0748291015625,
0.19140625,
0.09027099609375,
0.130126953125,
0.09027099609375
],
[
0.198974609375,
0.0736083984375,
0.364501953125,
0.0748291015625,
0.364501953125,
0.08905029296875,
0.198974609375,
0.0885009765625
],
[
0.19482421875,
0.08013916015625,
0.202392578125,
0.08013916015625,
0.202392578125,
0.0843505859375,
0.19482421875,
0.0843505859375
],
[
0.5146484375,
0.0748291015625,
0.8818359375,
0.07598876953125,
0.8818359375,
0.09027099609375,
0.5146484375,
0.08905029296875
],
[
0.166259765625,
0.0955810546875,
0.414794921875,
0.0968017578125,
0.414794921875,
0.114013671875,
0.166259765625,
0.1134033203125
],
[
0.5146484375,
0.0914306640625,
0.880859375,
0.0914306640625,
0.880859375,
0.1051025390625,
0.5146484375,
0.1051025390625
],
[
0.2032470703125,
0.1134033203125,
0.358642578125,
0.1134033203125,
0.358642578125,
0.1270751953125,
0.2032470703125,
0.1270751953125
],
[
0.5146484375,
0.1092529296875,
0.880859375,
0.1092529296875,
0.880859375,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.202392578125,
0.127685546875,
0.393798828125,
0.12890625,
0.393798828125,
0.1436767578125,
0.202392578125,
0.142578125
],
[
0.51318359375,
0.12353515625,
0.68408203125,
0.12469482421875,
0.68408203125,
0.1395263671875,
0.51318359375,
0.1383056640625
],
[
0.130126953125,
0.1312255859375,
0.160400390625,
0.1312255859375,
0.160400390625,
0.1436767578125,
0.130126953125,
0.1436767578125
],
[
0.16455078125,
0.133056640625,
0.172119140625,
0.133056640625,
0.172119140625,
0.1383056640625,
0.16455078125,
0.1383056640625
],
[
0.166259765625,
0.1341552734375,
0.1822509765625,
0.127685546875,
0.189697265625,
0.13720703125,
0.173828125,
0.1436767578125
],
[
0.187255859375,
0.1318359375,
0.2032470703125,
0.1318359375,
0.2032470703125,
0.1395263671875,
0.187255859375,
0.1395263671875
],
[
0.202392578125,
0.141357421875,
0.372802734375,
0.144287109375,
0.372802734375,
0.162109375,
0.202392578125,
0.1591796875
],
[
0.5322265625,
0.1395263671875,
0.88330078125,
0.14013671875,
0.88330078125,
0.155029296875,
0.5322265625,
0.15380859375
],
[
0.162109375,
0.157958984375,
0.51904296875,
0.1561279296875,
0.51904296875,
0.17578125,
0.162109375,
0.1776123046875
],
[
0.51318359375,
0.15673828125,
0.884765625,
0.15673828125,
0.884765625,
0.17041015625,
0.51318359375,
0.17041015625
],
[
0.515625,
0.1728515625,
0.8818359375,
0.1728515625,
0.8818359375,
0.1859130859375,
0.515625,
0.1859130859375
],
[
0.173828125,
0.1859130859375,
0.18896484375,
0.1859130859375,
0.18896484375,
0.1953125,
0.173828125,
0.1953125
],
[
0.2032470703125,
0.18408203125,
0.397216796875,
0.18408203125,
0.397216796875,
0.198974609375,
0.2032470703125,
0.198974609375
],
[
0.5146484375,
0.188232421875,
0.88671875,
0.18701171875,
0.88671875,
0.201904296875,
0.5146484375,
0.203125
],
[
0.202392578125,
0.198974609375,
0.403076171875,
0.1995849609375,
0.403076171875,
0.21435546875,
0.202392578125,
0.213134765625
],
[
0.5146484375,
0.204833984375,
0.880859375,
0.204833984375,
0.880859375,
0.218505859375,
0.5146484375,
0.218505859375
],
[
0.1292724609375,
0.2161865234375,
0.17626953125,
0.2161865234375,
0.17626953125,
0.23095703125,
0.1292724609375,
0.23095703125
],
[
0.17041015625,
0.218505859375,
0.18896484375,
0.218505859375,
0.18896484375,
0.22802734375,
0.17041015625,
0.22802734375
],
[
0.1822509765625,
0.215576171875,
0.370361328125,
0.2161865234375,
0.370361328125,
0.23095703125,
0.1822509765625,
0.2298583984375
],
[
0.5146484375,
0.2203369140625,
0.8818359375,
0.2203369140625,
0.8818359375,
0.2340087890625,
0.5146484375,
0.2340087890625
],
[
0.202392578125,
0.2340087890625,
0.307373046875,
0.232177734375,
0.307373046875,
0.2464599609375,
0.202392578125,
0.248779296875
],
[
0.5146484375,
0.2381591796875,
0.8818359375,
0.2381591796875,
0.8818359375,
0.251708984375,
0.5146484375,
0.251708984375
],
[
0.16455078125,
0.24462890625,
0.459228515625,
0.2452392578125,
0.459228515625,
0.26318359375,
0.16455078125,
0.262451171875
],
[
0.5146484375,
0.254150390625,
0.609375,
0.254150390625,
0.609375,
0.268310546875,
0.5146484375,
0.268310546875
],
[
0.12005615234375,
0.27783203125,
0.489501953125,
0.27783203125,
0.489501953125,
0.292236328125,
0.12005615234375,
0.292236328125
],
[
0.51318359375,
0.283203125,
0.52880859375,
0.283203125,
0.52880859375,
0.292236328125,
0.51318359375,
0.292236328125
],
[
0.5498046875,
0.281982421875,
0.66015625,
0.281982421875,
0.66015625,
0.29345703125,
0.5498046875,
0.29345703125
],
[
0.12005615234375,
0.29345703125,
0.48779296875,
0.29345703125,
0.48779296875,
0.30712890625,
0.12005615234375,
0.30712890625
],
[
0.12005615234375,
0.30810546875,
0.48779296875,
0.30810546875,
0.48779296875,
0.32177734375,
0.12005615234375,
0.32177734375
],
[
0.51318359375,
0.3046875,
0.8818359375,
0.305908203125,
0.8818359375,
0.320556640625,
0.51318359375,
0.319580078125
],
[
0.118408203125,
0.32373046875,
0.1502685546875,
0.32373046875,
0.1502685546875,
0.3349609375,
0.118408203125,
0.3349609375
],
[
0.5146484375,
0.320556640625,
0.88330078125,
0.32177734375,
0.88330078125,
0.336181640625,
0.5146484375,
0.3349609375
],
[
0.5146484375,
0.33837890625,
0.88330078125,
0.33837890625,
0.88330078125,
0.352783203125,
0.5146484375,
0.352783203125
],
[
0.51318359375,
0.352783203125,
0.88330078125,
0.352783203125,
0.88330078125,
0.36767578125,
0.51318359375,
0.36767578125
],
[
0.12176513671875,
0.363525390625,
0.48779296875,
0.363525390625,
0.48779296875,
0.377197265625,
0.12176513671875,
0.377197265625
],
[
0.51123046875,
0.3681640625,
0.8818359375,
0.369384765625,
0.8818359375,
0.38427734375,
0.51123046875,
0.383056640625
],
[
0.12005615234375,
0.377685546875,
0.48779296875,
0.37890625,
0.48779296875,
0.393798828125,
0.12005615234375,
0.392578125
],
[
0.5146484375,
0.385986328125,
0.8818359375,
0.385986328125,
0.8818359375,
0.399658203125,
0.5146484375,
0.399658203125
],
[
0.12005615234375,
0.393798828125,
0.48779296875,
0.394287109375,
0.48779296875,
0.4091796875,
0.12005615234375,
0.407958984375
],
[
0.515625,
0.402587890625,
0.8818359375,
0.402587890625,
0.8818359375,
0.416259765625,
0.515625,
0.416259765625
],
[
0.118408203125,
0.410888671875,
0.486083984375,
0.410888671875,
0.486083984375,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.5146484375,
0.418701171875,
0.8818359375,
0.418701171875,
0.8818359375,
0.431640625,
0.5146484375,
0.431640625
],
[
0.12176513671875,
0.427001953125,
0.48779296875,
0.427001953125,
0.48779296875,
0.440673828125,
0.12176513671875,
0.440673828125
],
[
0.51318359375,
0.432861328125,
0.8818359375,
0.43408203125,
0.8818359375,
0.448974609375,
0.51318359375,
0.44775390625
],
[
0.12005615234375,
0.443603515625,
0.48779296875,
0.443603515625,
0.48779296875,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.515625,
0.45068359375,
0.8818359375,
0.45068359375,
0.8818359375,
0.46435546875,
0.515625,
0.46435546875
],
[
0.12005615234375,
0.458984375,
0.490234375,
0.458984375,
0.490234375,
0.473876953125,
0.12005615234375,
0.473876953125
],
[
0.5146484375,
0.465576171875,
0.8818359375,
0.466064453125,
0.8818359375,
0.48095703125,
0.5146484375,
0.479736328125
],
[
0.12005615234375,
0.4755859375,
0.41650390625,
0.4755859375,
0.41650390625,
0.4892578125,
0.12005615234375,
0.4892578125
],
[
0.515625,
0.482666015625,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.515625,
0.496337890625
],
[
0.118408203125,
0.5029296875,
0.3720703125,
0.5029296875,
0.3720703125,
0.51708984375,
0.118408203125,
0.51708984375
],
[
0.515625,
0.498779296875,
0.8818359375,
0.498779296875,
0.8818359375,
0.51318359375,
0.515625,
0.51318359375
],
[
0.51708984375,
0.515625,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.51708984375,
0.529296875
],
[
0.12005615234375,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5380859375,
0.12005615234375,
0.5380859375
],
[
0.515625,
0.53076171875,
0.8818359375,
0.53076171875,
0.8818359375,
0.54443359375,
0.515625,
0.54443359375
],
[
0.12176513671875,
0.54052734375,
0.48779296875,
0.54052734375,
0.48779296875,
0.5546875,
0.12176513671875,
0.5546875
],
[
0.5146484375,
0.5458984375,
0.6455078125,
0.54638671875,
0.6455078125,
0.56103515625,
0.5146484375,
0.56005859375
],
[
0.12005615234375,
0.55712890625,
0.489501953125,
0.55712890625,
0.489501953125,
0.57080078125,
0.12005615234375,
0.57080078125
],
[
0.5322265625,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.5322265625,
0.57666015625
],
[
0.12005615234375,
0.572265625,
0.490234375,
0.572265625,
0.490234375,
0.5859375,
0.12005615234375,
0.5859375
],
[
0.5146484375,
0.57763671875,
0.8818359375,
0.5791015625,
0.8818359375,
0.59326171875,
0.5146484375,
0.5927734375
],
[
0.118408203125,
0.58740234375,
0.48779296875,
0.58837890625,
0.48779296875,
0.6025390625,
0.118408203125,
0.6015625
],
[
0.5146484375,
0.59423828125,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.5146484375,
0.60791015625
],
[
0.12176513671875,
0.60498046875,
0.48779296875,
0.60498046875,
0.48779296875,
0.6181640625,
0.12176513671875,
0.6181640625
],
[
0.5146484375,
0.60986328125,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.5146484375,
0.62353515625
],
[
0.12176513671875,
0.62060546875,
0.486083984375,
0.62060546875,
0.486083984375,
0.63427734375,
0.12176513671875,
0.63427734375
],
[
0.51123046875,
0.6259765625,
0.8818359375,
0.626953125,
0.8818359375,
0.64111328125,
0.51123046875,
0.64013671875
],
[
0.118408203125,
0.6357421875,
0.486083984375,
0.6357421875,
0.486083984375,
0.6494140625,
0.118408203125,
0.6494140625
],
[
0.5146484375,
0.6435546875,
0.8818359375,
0.6435546875,
0.8818359375,
0.65771484375,
0.5146484375,
0.65771484375
],
[
0.12176513671875,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.66650390625,
0.12176513671875,
0.66650390625
],
[
0.5146484375,
0.66015625,
0.6455078125,
0.66015625,
0.6455078125,
0.6748046875,
0.5146484375,
0.6748046875
],
[
0.12005615234375,
0.66943359375,
0.48779296875,
0.66943359375,
0.48779296875,
0.68310546875,
0.12005615234375,
0.68310546875
],
[
0.51318359375,
0.68603515625,
0.54833984375,
0.68603515625,
0.54833984375,
0.69775390625,
0.51318359375,
0.69775390625
],
[
0.55908203125,
0.68505859375,
0.75732421875,
0.68505859375,
0.75732421875,
0.69873046875,
0.55908203125,
0.69873046875
],
[
0.12176513671875,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.12176513671875,
0.71533203125
],
[
0.12176513671875,
0.68701171875,
0.486083984375,
0.68701171875,
0.486083984375,
0.69775390625,
0.12176513671875,
0.69775390625
],
[
0.51318359375,
0.705078125,
0.88330078125,
0.7060546875,
0.88330078125,
0.72021484375,
0.51318359375,
0.7197265625
],
[
0.12176513671875,
0.71728515625,
0.48779296875,
0.71728515625,
0.48779296875,
0.732421875,
0.12176513671875,
0.732421875
],
[
0.52490234375,
0.72021484375,
0.88330078125,
0.7216796875,
0.88330078125,
0.736328125,
0.52490234375,
0.7353515625
],
[
0.12005615234375,
0.73291015625,
0.48779296875,
0.73291015625,
0.48779296875,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.515625,
0.7412109375,
0.53564453125,
0.7412109375,
0.53564453125,
0.75048828125,
0.515625,
0.75048828125
],
[
0.12176513671875,
0.75,
0.48779296875,
0.75,
0.48779296875,
0.76318359375,
0.12176513671875,
0.76318359375
],
[
0.5322265625,
0.7548828125,
0.88330078125,
0.7548828125,
0.88330078125,
0.7685546875,
0.5322265625,
0.7685546875
],
[
0.12005615234375,
0.765625,
0.48779296875,
0.765625,
0.48779296875,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.5146484375,
0.77099609375,
0.88330078125,
0.77099609375,
0.88330078125,
0.78515625,
0.5146484375,
0.78515625
],
[
0.12005615234375,
0.783203125,
0.470947265625,
0.783203125,
0.470947265625,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.5146484375,
0.78759765625,
0.88330078125,
0.78759765625,
0.88330078125,
0.80029296875,
0.5146484375,
0.80029296875
],
[
0.12005615234375,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.12005615234375,
0.8125
],
[
0.51318359375,
0.802734375,
0.880859375,
0.802734375,
0.880859375,
0.81640625,
0.51318359375,
0.81640625
],
[
0.118408203125,
0.81396484375,
0.486083984375,
0.81396484375,
0.486083984375,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.5146484375,
0.8193359375,
0.8818359375,
0.8193359375,
0.8818359375,
0.8330078125,
0.5146484375,
0.8330078125
],
[
0.118408203125,
0.830078125,
0.48779296875,
0.8291015625,
0.48779296875,
0.84326171875,
0.118408203125,
0.84423828125
],
[
0.515625,
0.8349609375,
0.880859375,
0.8349609375,
0.880859375,
0.8486328125,
0.515625,
0.8486328125
],
[
0.12005615234375,
0.84423828125,
0.48779296875,
0.845703125,
0.48779296875,
0.85986328125,
0.12005615234375,
0.859375
],
[
0.5322265625,
0.85546875,
0.8818359375,
0.85791015625,
0.8818359375,
0.87255859375,
0.5322265625,
0.87060546875
],
[
0.12005615234375,
0.8623046875,
0.139404296875,
0.8623046875,
0.139404296875,
0.8759765625,
0.12005615234375,
0.8759765625
],
[
0.5146484375,
0.87060546875,
0.8017578125,
0.87060546875,
0.8017578125,
0.88427734375,
0.5146484375,
0.88427734375
],
[
0.1376953125,
0.87890625,
0.48779296875,
0.87890625,
0.48779296875,
0.892578125,
0.1376953125,
0.892578125
],
[
0.53564453125,
0.8828125,
0.76416015625,
0.8828125,
0.76416015625,
0.896484375,
0.53564453125,
0.896484375
],
[
0.12005615234375,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.5341796875,
0.89453125,
0.78076171875,
0.8955078125,
0.78076171875,
0.91015625,
0.5341796875,
0.9091796875
]
] |
[
"Fgn",
"Extracted n -grams",
"",
"但传达不同含义的两个边缘可以被Smatch视",
"1. ask-01; girl; leave-1l; boy",
"为等效。2另一方面：SemBleu不仅考虑边缘",
"ask-01 :ARG0 girl;",
"标签，还考虑其头部和尾部的内容，如表1中",
"ask-01 :ARG1 leave-11;",
"提取的n-gram所示",
"(a)",
"",
"2",
"",
"leave-11 :ARGO boy;",
"以图1中的AMR为例，Smatch将（a）中",
"\"3ask-01:ARGT leave-if :ARGO boy; ",
"的“girl\"，“ask-01”和“leave-11”映射为（b)",
"中的“woman”，“make-01”和“pie\"。结果，它",
"1",
"woman; make-0l; pie; 2",
"认为（a）中的“ask-01：ARG0girl\"和“ask-01：",
"make-01 :ARG0 woman;",
"ARG1leave-11\"被正确映射到（b）中的“make-",
"(b)",
"2",
"make-01 :ARG1pie;",
"01:ARGOwoman\" 和“make-01:ARG1pie\"",
"pie :quant 2;",
"这没有意义。相反，SemBleu不认为这些边是",
"3-make-01 :ARGI pie :quant 2;",
"正确匹配的",
"Table 1: n -grams (separated by \";\") extracted",
"3",
"Experiments",
"from the AMRs in Figure 1 with our extraction",
"algorithm. Fg represents the corresponding subfig-",
"我们将SemBleu与Smatch在来自",
"ure.",
"LDC2015E86测试集的100个句子中的4个系",
"统的输出上进行了比较。这些系统是：CAMR，",
"3 JAMR, 4 Gros (Groschwitz et al., 2018)和",
"有大量提取的n-gram。对于具有N节点的",
"Lyu（LyuandTitov，2018）。对于每个句子",
"全连接图，有O(Nn）可能的n-gram。幸运的",
"在Callison-Burchetal.（2010）之后，注释者",
"是，AMR是非常稀疏的树状图（Chiang etal.",
"决定相对顺序，而不是所有系统的完整顺序。",
"2018）。对于具有N节点的树，n-gram的数",
"特别是，4个系统输出被随机分组为2对，以",
"量以O(n·N）为界，O(n·N）在树尺度上是",
"进行2次比较。对于每对，我们请3位注释",
"线性的。作为树状图，我们期望从AMR中提",
"者决定哪一个更好，并选择多数票作为最终判",
"取的n-gram的数量在图尺度上大致呈线性。",
"断。所有注释者在AMR相关研究方面都有几",
"我们的实验从经验上证实了这一期望",
"年的经验，并且判断是基于他们对系统生成的",
"2.3Comparison with Smatch",
"AMR保留参考AMR含义的印象。在200个",
"比较中，注释者完全同意142个，占71个%。",
"通常，Smatch将比较两个AMR的问题分解",
"通过判断，我们研究了句子和语料库级别上两",
"为比较最小单位：节点和边。它将每个AMR",
"个度量的一致性",
"视为节点和边的袋子，然后计算关于正确映",
"我们考虑SemBleu的所有unigram，bigram",
"射的节点和边的F1分数。给定两个AMR，",
"和trigram，并且权重（等式1中的wks）是等",
"Smatch通过最大化总体F1分数来搜索图节",
"效的（每个为1/3）。对于句子级别的评估，我",
"点之间的一对一映射，并且边到边映射由节",
"们按照以前的工作使用NIST几何平滑（Chen",
"点到节点映射自动确定。由于获得最佳映射是",
"andCherry，2014）。在Smatch之后，在提取",
"NP完全的（通过从子图同构中进行约简），因",
"n-gram进行比较之前，反转诸如“ARGO-of\"",
"此它使用贪婪的爬山算法来查找映射，该映射",
"之类的反向关系",
"很可能是次优的。一个关键的区别是SemBleu",
"3.1",
"Corpus-level experiment",
"个指标唯一的共同点是节点到节点的对应关",
"通常比Smatch考虑更多的全局功能。这两",
"对于语料库级别的比较，我们为每个系统分配",
"系（也称为SemBleu的unigram）。SemBleu",
"一个人类分数，该分数等于系统输出的首选次",
"的每个bigram由两个AMR节点和一个连",
"数",
"接它们的边组成，因此bigram已经捕获了比",
"我们的四个系统的人类得分分别为30、33、",
"Smatch更大的上下文。此外，SemBleu的高阶",
"63和74。他们的SemBleu得分分别为28、30、",
"n-gram捕获了更大的对应关系。这可能是",
"38和41，Smatch得分分别为56、56、63和",
"种权衡。通常，更多的高阶匹配表示更好的解",
"67。SemBleu通常更符合人类的判断。特别是",
"析性能，但有时我们希望将部分正确的结果与",
"对于Smatch分数，CAMR和JAMR之间存",
"完全错误的结果区分开来。。因此，将Smatch",
"在联系，而SemBleu分数更具区别性。在计",
"与SemBleu结合使用可能会给出更全面的判",
"20ne example is shown in the Smatch tutorial https:",
"断",
"//amr.isi.edu/eval/smatch/tutorial.html.",
"另一个区别是确定边（关系）等价的方式。",
"3https://github.com/c-amr/camr",
"Smatch仅检查边缘标签，因此具有相同标签",
"4https://github.com/jflanigan/jamr"
] |
{
"arxiv_id": "1905.10726",
"language": "zh",
"page_idx": 2
}
|
||
1905.10726-zh-4
|
1905.10726-zh-4.jpg
|
Titov, 2018) , Gros (Groschwitz et al., 2018) ,vanNood(van Noord and Bos,2017)和Guo (Guo and Lu, 2018) 的输出在我们的 SemBleu 上进行评估。对于 CAMR 和 JAMR, 我们通过运行发布的系统获得它们的输出。SemBleu 大部分与 Smatch 一致，除了 Guo 和 Gros 之间的顺序。这可能是因为郭与参考文献有更多的高阶对应关系
## 4 Conclusion
虽然人们可能期望在速度和与人类判断的相关性之间进行权衡，但 SemBleu 似乎在两个维度上都优于 Smatch 。与人类判断的相关性的改进来自于 SemBleu 考虑了输入图的较大片段。速度的提高来自于避免搜索两个图之间的映射。在实践中，可以通过简单地考虑顶点标签及其邻居的标签，通过它们出现的$n$ -gram 来识别顶点映射。SemBleu 可用于比较其他类型的图，包括循环图
Acknowledgments 我们非常感谢 LisaJin 和ParkerRiley 的注释。我们感谢王志国 (亚马逊AWS)、苏劲松(厦门大学) 和匿名审稿人的深刻评论。NSF 奖 IIS-1813823 支持的研究
References
Yoav Artzi, Kenton Lee, and Luke Zettlemoyer.
2015. Broad-coverage CCG semantic parsing with AMR. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , pages 1699-1710.
Miguel Ballesteros and Yaser Al-Onaizan. 2017.
AMR parsing using stack-LSTMs. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing , pages 1269-1275.
Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Grifftt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract $\operatorname*{meaning}_{\mathrm{u}}$representation for sembanking. In Proceedings of the 7th Linguistic Annotation Work- $\bar{\text{shop and Interoperability with Discourse }}$,pages 178-186.
Jan Buys and Phil Blunsom. 2017. Robust incre-
mental neural semantic graph parsing. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1{:}Long Papers) , pages $1215-\bar{1226}.$
Shu Cai and Kevin Knight. 2013. Smatch: an eval-
uation metric for semantic feature structures. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages 748-752.
Chris Callison-Burch, Philipp Koehn, Christof
Monz, Kay Peterson, Mark Przybocki, and Omar F Zaidan. 2010. Findings of the 2010 joint workshop on statistical machine translation and metrics for machine translation. In Proceedings of the Joint Fifth Workshop on Statistical Machine Translation and MetricsMATR , pages 17- 53.
Boxing Chen and Colin Cherry. 2014. A sys-
tematic comparison of smoothing techniques for sentence-level BLEU. In Proceedings of the Ninth Workshop on Statistical Machine Translation ,pages 362-367.
David Chiang, Frank Drewes, Daniel Gildea, Adam
Lopez, and Giorgio Satta. 2018. Weighted DAG automata for semantic graphs. Computational Linguistics , 44( 1) : 119- 186.
Marco Damonte, Shay B. Cohen, and Giorgio
Satta. 2017. An incremental parser for abstract meaning representation. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers , pages 536-546.
Jeffrey Flanigan, Sam Thomson, Jaime Carbonell,
Chris Dyer, and Noah A. Smith. 2014. A discriminative graph-based parser for the abstract meaning representation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) ,pages $1426-\bar{1}436.$
Jonas Groschwitz, Matthias Lindemann, Meaghan
Fowlie, Mark Johnson, and Alexander Koller. 2018. AMR dependency parsing with a typed semantic algebra. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 1831-1841.
Zhijiang Guo and Wei Lu. 2018. Better transition-
$\mathop{\text{based AMR parsing with a refned search space.}}$ In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing ,pages 1712-1722.
Philipp Koehn. 2004. Statistical significance tests
for machine translation evaluation. In Proceedings of the 2004 conference on empirical methods in natural language processing , pages 388- 395.
Ioannis Konstas, Srinivasan Iyer, Mark Yatskar,
Yejin Choi, and Luke Zettlemoyer. 2017. Neural AMR: Sequence-to-sequence models for parsing and generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 146-157.
Chunchuan Lyu and Ivan Titov. 2018. AMR pars- $\operatorname*{ing}_{\mathrm{~as~graph~prediction~with~latent~alignment.}}$ In Proceedings of the 56th Annual Meeting of
|
[
[
0.12005615234375,
0.07659912109375,
0.48779296875,
0.07659912109375,
0.48779296875,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.51318359375,
0.07659912109375,
0.884765625,
0.07659912109375,
0.884765625,
0.09027099609375,
0.51318359375,
0.09027099609375
],
[
0.124267578125,
0.0926513671875,
0.489501953125,
0.0914306640625,
0.489501953125,
0.105712890625,
0.124267578125,
0.10687255859375
],
[
0.52880859375,
0.0885009765625,
0.88330078125,
0.0885009765625,
0.88330078125,
0.1015625,
0.52880859375,
0.1015625
],
[
0.12005615234375,
0.1092529296875,
0.489501953125,
0.1092529296875,
0.489501953125,
0.122314453125,
0.12005615234375,
0.122314453125
],
[
0.5341796875,
0.10394287109375,
0.8818359375,
0.10394287109375,
0.8818359375,
0.114013671875,
0.5341796875,
0.114013671875
],
[
0.53076171875,
0.1175537109375,
0.8818359375,
0.1175537109375,
0.8818359375,
0.127685546875,
0.53076171875,
0.127685546875
],
[
0.124267578125,
0.12469482421875,
0.489501953125,
0.12469482421875,
0.489501953125,
0.1383056640625,
0.124267578125,
0.1383056640625
],
[
0.52880859375,
0.1270751953125,
0.884765625,
0.12890625,
0.884765625,
0.1436767578125,
0.52880859375,
0.141357421875
],
[
0.12005615234375,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.12005615234375,
0.155029296875
],
[
0.52880859375,
0.14013671875,
0.8818359375,
0.141357421875,
0.8818359375,
0.1561279296875,
0.52880859375,
0.155029296875
],
[
0.118408203125,
0.1561279296875,
0.484375,
0.1561279296875,
0.484375,
0.169189453125,
0.118408203125,
0.169189453125
],
[
0.52880859375,
0.152587890625,
0.880859375,
0.155029296875,
0.880859375,
0.169189453125,
0.52880859375,
0.16748046875
],
[
0.12005615234375,
0.1728515625,
0.48779296875,
0.1728515625,
0.48779296875,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.53076171875,
0.168701171875,
0.5576171875,
0.168701171875,
0.5576171875,
0.18115234375,
0.53076171875,
0.18115234375
],
[
0.118408203125,
0.189453125,
0.2493896484375,
0.189453125,
0.2493896484375,
0.203125,
0.118408203125,
0.203125
],
[
0.51318359375,
0.1923828125,
0.88330078125,
0.1923828125,
0.88330078125,
0.2060546875,
0.51318359375,
0.2060546875
],
[
0.53076171875,
0.2060546875,
0.884765625,
0.2060546875,
0.884765625,
0.2197265625,
0.53076171875,
0.2197265625
],
[
0.118408203125,
0.2161865234375,
0.14111328125,
0.2161865234375,
0.14111328125,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.151123046875,
0.21435546875,
0.252685546875,
0.2161865234375,
0.252685546875,
0.23095703125,
0.151123046875,
0.2286376953125
],
[
0.5322265625,
0.218505859375,
0.88330078125,
0.218505859375,
0.88330078125,
0.232177734375,
0.5322265625,
0.232177734375
],
[
0.5322265625,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.5322265625,
0.24462890625
],
[
0.12005615234375,
0.2392578125,
0.486083984375,
0.240478515625,
0.486083984375,
0.254638671875,
0.12005615234375,
0.254150390625
],
[
0.53076171875,
0.24462890625,
0.7021484375,
0.24462890625,
0.7021484375,
0.2587890625,
0.53076171875,
0.2587890625
],
[
0.12005615234375,
0.257080078125,
0.48779296875,
0.257080078125,
0.48779296875,
0.270751953125,
0.12005615234375,
0.270751953125
],
[
0.118408203125,
0.271484375,
0.486083984375,
0.271484375,
0.486083984375,
0.28515625,
0.118408203125,
0.28515625
],
[
0.51318359375,
0.268310546875,
0.88330078125,
0.268310546875,
0.88330078125,
0.281982421875,
0.51318359375,
0.281982421875
],
[
0.52880859375,
0.28076171875,
0.8818359375,
0.28076171875,
0.8818359375,
0.29443359375,
0.52880859375,
0.29443359375
],
[
0.12005615234375,
0.2880859375,
0.48779296875,
0.2880859375,
0.48779296875,
0.302734375,
0.12005615234375,
0.302734375
],
[
0.53076171875,
0.295654296875,
0.884765625,
0.295654296875,
0.884765625,
0.308837890625,
0.53076171875,
0.308837890625
],
[
0.12005615234375,
0.3046875,
0.489501953125,
0.3046875,
0.489501953125,
0.318359375,
0.12005615234375,
0.318359375
],
[
0.52880859375,
0.30712890625,
0.73486328125,
0.30712890625,
0.73486328125,
0.320556640625,
0.52880859375,
0.320556640625
],
[
0.118408203125,
0.319580078125,
0.489501953125,
0.320556640625,
0.489501953125,
0.3349609375,
0.118408203125,
0.334228515625
],
[
0.118408203125,
0.3349609375,
0.489501953125,
0.33740234375,
0.489501953125,
0.350830078125,
0.118408203125,
0.3486328125
],
[
0.51123046875,
0.330810546875,
0.884765625,
0.33203125,
0.884765625,
0.3466796875,
0.51123046875,
0.345703125
],
[
0.53076171875,
0.345703125,
0.880859375,
0.345703125,
0.880859375,
0.355712890625,
0.53076171875,
0.355712890625
],
[
0.12005615234375,
0.352783203125,
0.48779296875,
0.352783203125,
0.48779296875,
0.366455078125,
0.12005615234375,
0.366455078125
],
[
0.53076171875,
0.359375,
0.88330078125,
0.359375,
0.88330078125,
0.372314453125,
0.53076171875,
0.372314453125
],
[
0.118408203125,
0.36767578125,
0.287109375,
0.3681640625,
0.287109375,
0.383056640625,
0.118408203125,
0.3818359375
],
[
0.5322265625,
0.371826171875,
0.884765625,
0.371826171875,
0.884765625,
0.385498046875,
0.5322265625,
0.385498046875
],
[
0.5322265625,
0.38427734375,
0.88330078125,
0.38427734375,
0.88330078125,
0.39794921875,
0.5322265625,
0.39794921875
],
[
0.1259765625,
0.393798828125,
0.48779296875,
0.393798828125,
0.48779296875,
0.40673828125,
0.1259765625,
0.40673828125
],
[
0.53076171875,
0.39794921875,
0.80615234375,
0.39794921875,
0.80615234375,
0.410888671875,
0.53076171875,
0.410888671875
],
[
0.118408203125,
0.4091796875,
0.48779296875,
0.4091796875,
0.48779296875,
0.4228515625,
0.118408203125,
0.4228515625
],
[
0.12005615234375,
0.42578125,
0.489501953125,
0.42578125,
0.489501953125,
0.439453125,
0.12005615234375,
0.439453125
],
[
0.5146484375,
0.421630859375,
0.8818359375,
0.421630859375,
0.8818359375,
0.435302734375,
0.5146484375,
0.435302734375
],
[
0.53076171875,
0.43408203125,
0.8818359375,
0.43408203125,
0.8818359375,
0.448974609375,
0.53076171875,
0.448974609375
],
[
0.12005615234375,
0.441162109375,
0.453369140625,
0.441162109375,
0.453369140625,
0.454833984375,
0.12005615234375,
0.454833984375
],
[
0.52880859375,
0.44775390625,
0.884765625,
0.446533203125,
0.884765625,
0.460205078125,
0.52880859375,
0.46142578125
],
[
0.5322265625,
0.460205078125,
0.884765625,
0.460205078125,
0.884765625,
0.473876953125,
0.5322265625,
0.473876953125
],
[
0.53076171875,
0.473876953125,
0.884765625,
0.473876953125,
0.884765625,
0.48681640625,
0.53076171875,
0.48681640625
],
[
0.12176513671875,
0.485107421875,
0.21240234375,
0.485107421875,
0.21240234375,
0.496337890625,
0.12176513671875,
0.496337890625
],
[
0.5322265625,
0.48681640625,
0.88330078125,
0.48681640625,
0.88330078125,
0.50048828125,
0.5322265625,
0.50048828125
],
[
0.12005615234375,
0.50439453125,
0.48779296875,
0.50439453125,
0.48779296875,
0.5185546875,
0.12005615234375,
0.5185546875
],
[
0.52880859375,
0.50048828125,
0.7138671875,
0.498779296875,
0.7138671875,
0.5126953125,
0.52880859375,
0.51416015625
],
[
0.1351318359375,
0.5166015625,
0.489501953125,
0.51708984375,
0.489501953125,
0.5322265625,
0.1351318359375,
0.53076171875
],
[
0.1368408203125,
0.53076171875,
0.489501953125,
0.53076171875,
0.489501953125,
0.54443359375,
0.1368408203125,
0.54443359375
],
[
0.515625,
0.525390625,
0.88330078125,
0.525390625,
0.88330078125,
0.5361328125,
0.515625,
0.5361328125
],
[
0.52880859375,
0.5361328125,
0.8818359375,
0.5361328125,
0.8818359375,
0.5498046875,
0.52880859375,
0.5498046875
],
[
0.1351318359375,
0.5419921875,
0.48779296875,
0.54345703125,
0.48779296875,
0.55810546875,
0.1351318359375,
0.55712890625
],
[
0.1351318359375,
0.5556640625,
0.35693359375,
0.5556640625,
0.35693359375,
0.5693359375,
0.1351318359375,
0.5693359375
],
[
0.53076171875,
0.55029296875,
0.884765625,
0.55029296875,
0.884765625,
0.56396484375,
0.53076171875,
0.56396484375
],
[
0.52880859375,
0.5625,
0.884765625,
0.5625,
0.884765625,
0.576171875,
0.52880859375,
0.576171875
],
[
0.118408203125,
0.5791015625,
0.486083984375,
0.5791015625,
0.486083984375,
0.5927734375,
0.118408203125,
0.5927734375
],
[
0.5322265625,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59033203125,
0.5322265625,
0.59033203125
],
[
0.1368408203125,
0.5927734375,
0.48779296875,
0.5927734375,
0.48779296875,
0.60546875,
0.1368408203125,
0.60546875
],
[
0.53076171875,
0.58837890625,
0.88671875,
0.5888671875,
0.88671875,
0.60400390625,
0.53076171875,
0.6025390625
],
[
0.1351318359375,
0.60498046875,
0.48779296875,
0.60498046875,
0.48779296875,
0.6181640625,
0.1351318359375,
0.6181640625
],
[
0.52880859375,
0.60400390625,
0.66015625,
0.60107421875,
0.66015625,
0.6142578125,
0.52880859375,
0.61767578125
],
[
0.133544921875,
0.61767578125,
0.489501953125,
0.6181640625,
0.489501953125,
0.6328125,
0.133544921875,
0.6318359375
],
[
0.51318359375,
0.6259765625,
0.880859375,
0.6259765625,
0.880859375,
0.6396484375,
0.51318359375,
0.6396484375
],
[
0.1376953125,
0.6328125,
0.219970703125,
0.6328125,
0.219970703125,
0.64453125,
0.1376953125,
0.64453125
],
[
0.53076171875,
0.6396484375,
0.88330078125,
0.6396484375,
0.88330078125,
0.65283203125,
0.53076171875,
0.65283203125
],
[
0.118408203125,
0.65380859375,
0.489501953125,
0.65380859375,
0.489501953125,
0.66748046875,
0.118408203125,
0.66748046875
],
[
0.5322265625,
0.65478515625,
0.8818359375,
0.65478515625,
0.8818359375,
0.6650390625,
0.5322265625,
0.6650390625
],
[
0.1351318359375,
0.66650390625,
0.486083984375,
0.66650390625,
0.486083984375,
0.6796875,
0.1351318359375,
0.6796875
],
[
0.53076171875,
0.6650390625,
0.88671875,
0.66650390625,
0.88671875,
0.68115234375,
0.53076171875,
0.6796875
],
[
0.1368408203125,
0.68115234375,
0.489501953125,
0.68115234375,
0.489501953125,
0.69482421875,
0.1368408203125,
0.69482421875
],
[
0.53076171875,
0.6796875,
0.6689453125,
0.6767578125,
0.6689453125,
0.6904296875,
0.53076171875,
0.693359375
],
[
0.1351318359375,
0.6923828125,
0.486083984375,
0.6923828125,
0.486083984375,
0.7060546875,
0.1351318359375,
0.7060546875
],
[
0.1368408203125,
0.70703125,
0.489501953125,
0.70703125,
0.489501953125,
0.72021484375,
0.1368408203125,
0.72021484375
],
[
0.515625,
0.705078125,
0.88330078125,
0.705078125,
0.88330078125,
0.71533203125,
0.515625,
0.71533203125
],
[
0.133544921875,
0.7197265625,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.133544921875,
0.73291015625
],
[
0.53076171875,
0.71630859375,
0.88330078125,
0.71630859375,
0.88330078125,
0.72998046875,
0.53076171875,
0.72998046875
],
[
0.1351318359375,
0.73291015625,
0.489501953125,
0.73291015625,
0.489501953125,
0.74658203125,
0.1351318359375,
0.74658203125
],
[
0.53076171875,
0.72998046875,
0.8818359375,
0.72998046875,
0.8818359375,
0.74072265625,
0.53076171875,
0.74072265625
],
[
0.1368408203125,
0.7451171875,
0.202392578125,
0.74755859375,
0.202392578125,
0.7587890625,
0.1368408203125,
0.75732421875
],
[
0.53076171875,
0.7421875,
0.88330078125,
0.7421875,
0.88330078125,
0.75732421875,
0.53076171875,
0.75732421875
],
[
0.12005615234375,
0.76953125,
0.48779296875,
0.76953125,
0.48779296875,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.51318359375,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.1351318359375,
0.7822265625,
0.48779296875,
0.7822265625,
0.48779296875,
0.7958984375,
0.1351318359375,
0.7958984375
],
[
0.5322265625,
0.77978515625,
0.88330078125,
0.77978515625,
0.88330078125,
0.79345703125,
0.5322265625,
0.79345703125
],
[
0.1351318359375,
0.79443359375,
0.48779296875,
0.79443359375,
0.48779296875,
0.80810546875,
0.1351318359375,
0.80810546875
],
[
0.53076171875,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.53076171875,
0.8056640625
],
[
0.1351318359375,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1351318359375,
0.82177734375
],
[
0.53076171875,
0.8056640625,
0.884765625,
0.8056640625,
0.884765625,
0.8193359375,
0.53076171875,
0.8193359375
],
[
0.1351318359375,
0.8193359375,
0.4013671875,
0.82080078125,
0.4013671875,
0.8349609375,
0.1351318359375,
0.83447265625
],
[
0.5322265625,
0.818359375,
0.8818359375,
0.818359375,
0.8818359375,
0.83203125,
0.5322265625,
0.83203125
],
[
0.53076171875,
0.83056640625,
0.88671875,
0.83203125,
0.88671875,
0.8466796875,
0.53076171875,
0.845703125
],
[
0.118408203125,
0.84423828125,
0.484375,
0.84423828125,
0.484375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.52880859375,
0.845703125,
0.6455078125,
0.84423828125,
0.6455078125,
0.859375,
0.52880859375,
0.85986328125
],
[
0.1368408203125,
0.85546875,
0.490234375,
0.85546875,
0.490234375,
0.869140625,
0.1368408203125,
0.869140625
],
[
0.1368408203125,
0.869140625,
0.489501953125,
0.869140625,
0.489501953125,
0.8828125,
0.1368408203125,
0.8828125
],
[
0.51318359375,
0.8681640625,
0.8818359375,
0.869140625,
0.8818359375,
0.8828125,
0.51318359375,
0.8818359375
],
[
0.139404296875,
0.88427734375,
0.486083984375,
0.88427734375,
0.486083984375,
0.89453125,
0.139404296875,
0.89453125
],
[
0.53076171875,
0.8828125,
0.88330078125,
0.8828125,
0.88330078125,
0.896484375,
0.53076171875,
0.896484375
],
[
0.1368408203125,
0.896484375,
0.42236328125,
0.896484375,
0.42236328125,
0.91015625,
0.1368408203125,
0.91015625
],
[
0.53076171875,
0.896484375,
0.88671875,
0.896484375,
0.88671875,
0.91015625,
0.53076171875,
0.91015625
]
] |
[
"Titov, 2018),Gros (Groschwitz et al., 2018)",
"Chris Callison-Burch, Philipp Koehn, Christof",
", vanNood (van Noord and Bos, 2017) 和 Guo",
"Monz, Kay Peterson, Mark Przybocki, and",
"(GuoandLu，2018）的输出在我们的SemBleu",
"Omar FZaidan.2010.Findings of the2010 joint",
"workshop on statistical machinetranslation and",
"上进行评估。对于CAMR和JAMR，我们通",
"metrics for machine translation. In Proceedings",
"过运行发布的系统获得它们的输出。SemBleu",
"of the Joint Fifth Workshop on Statistical Ma-",
"大部分与Smatch一致，除了Guo和Gros之",
"chine Translation and MetricsMATR , pages 17-",
"间的顺序。这可能是因为郭与参考文献有更多",
"53.",
"的高阶对应关系",
"Boxing Chen and Colin Cherry. 2014. A sys-",
"tematic comparison of smoothing techniques for",
"4.",
"Conclusion",
"sentence-level BLEU..In Proceedings of the",
"Ninth Workshop on Statistical Machine Trans-",
"虽然人们可能期望在速度和与人类判断的相关",
"lation , pages 362-367.",
"性之间进行权衡，但SemBleu似乎在两个维",
"度上都优于Smatch。与人类判断的相关性的",
"David Chiang, Frank Drewes, Daniel Gildea, Adam",
"Lopez, and Giorgio Satta. 2018. Weighted DAG",
"改进来自于SemBleu考虑了输入图的较大片",
"automata for semantic graphs.Computational",
"段。速度的提高来自于避免搜索两个图之间的",
"Linguistics , 44(1):119-186.",
"映射。在实践中，可以通过简单地考虑顶点标",
"签及其邻居的标签，通过它们出现的n-gram",
"Marco Damonte, Shay B. Cohen, and Giorgio",
"Satta.2017.An incremental parserfor abstract",
"来识别顶点映射。SemBleu可用于比较其他类",
"meaning representation. In Proceedings of the",
"型的图，包括循环图",
"15th Conference of the European Chapter of the",
"Association for Computational Linguistics: Vol-",
"Acknowledgments我们非常感谢LisaJin和",
"ume 1, Long Papers , pages 536-546.",
"ParkerRiley的注释。我们感谢王志国（亚马逊",
"AWS)、苏劲松（厦门大学）和匿名审稿人的深",
"Jeffrey Flanigan, Sam Thomson, Jaime Carbonell.",
"Chris Dyer, and Noah A. Smith. 2014. A dis-",
"刻评论。NSF奖IIS-1813823支持的研究",
"criminative graph-based parser for the abstract",
"meaning representation. In Proceedings of the",
"52nd Annual Meeting of the Association for",
"References",
"Computational Linguistics (Volume 1: Long Pa-",
"Yoav Artzi, Kenton Lee, and Luke Zettlemoyer.",
"pers), pages 1426-1436.",
"2015.Broad-coverage CCG semantic parsing",
"with AMR. In Proceedings of the 2015 Confer-",
"Jonas Groschwitz,Matthias Lindemann,Meaghan",
"Fowlie, Mark Johnson, and Alexander Koller",
"ence on Empirical Methods in Natural Language",
"Processing pages 1699-1710.",
"2018.AMR dependency parsing with a typed",
" semantic algebra. In Proceedings of the 56th",
"Miguel Ballesteros and Yaser Al-Onaizan. 2017",
"Annual Meeting of the Association for Compu-",
"AMR parsing using stack-LSTMs. In Proceed-",
"tational Linguistics (Volume l: Long Papers),",
"ings of the 2017 Conference on Empirical Meth-",
"pages 1831-1841.",
"ods in Natural Language Processing ,pages",
"Zhijiang Guo and Wei Lu. 2018. Better transition",
"1269-1275.",
"based AMR parsing with a refined search space.",
"Laura Banarescu, Claire Bonial, Shu Cai,",
"In Proceedings of the 2018 Conference onEm-",
"Madalina Georgescu, Kira Griffitt, Ulf Herm-",
"pirical Methods in Natural Language Processing",
"jakob, Kevin Knight, Philipp Koehn, Martha",
", pages 1712-1722.",
"Palmer, and Nathan Schneider. 2013. Abstract",
"meaning representation for sembanking. In Pro-",
"Philipp Koehn.2004.Statistical significance tests",
"ceedings of the 7th Linguistic Annotation Work-",
"for machine translation evaluation. In Proceed-",
"shop and Interoperability with Discourse, pages",
"ings of the 2004 conference on empirical methods",
"178-186.",
"in natural language processing , pages 388-395.",
"Jan Buvs and Phil Blunsom. 2017. Robust incre-",
"Ioannis Konstas, Srinivasan Iyer, Mark Yatskar,",
"mental neural semantic graph parsing. In Pro-",
"Yejin Choi, and Luke Zettlemoyer. 2017. Neu-",
"ceedings of the 55th Annual Meeting of the Asso-",
"ral AMR: Sequence-to-sequence models for pars-",
"ciation for Computational Linguistics (Volume",
"ing and generation. In Proceedings of the 55th",
"1: Long Papers) , pages 1215-1226.",
"Annual Meeting of the Association for Compu-",
"tational Linguistics (Volume l: Long Papers),",
"Shu Cai and Kevin Knight. 2013. Smatch: an eval-",
"pages 146-157.",
"uation metric for semantic feature structures. In",
"Proceedings of the 5lst Annual Meeting of the",
"Chunchuan Lyu and Ivan Titov. 2018. AMR pars-",
"Association forComputationalLinguistics (Vol-",
"ing as graph prediction with latent alignment.",
"ume 2: Short Papers), pages 748-752.",
"In Proceedings of the 56th Annual Meeting of"
] |
{
"arxiv_id": "1905.10726",
"language": "zh",
"page_idx": 4
}
|
||
2106.01654-zh-0
|
2106.01654-zh-0.jpg
|
通过外部因果陈述的自监督表征学习改进事件因果关系识别
Xinyu Zuo$^1,2$, Pengfei Cao$^1,2$, Yubo Chen$^1,2$, Kang Liu$^1,2$, Jun Zhao$^1,2$,
Weihua Peng$^3$ and Yuguang Chen$^3$
$^{1}$National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China
$^{2}$School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China
$^{3}$Beijing Baidu Netcom Science Technology Co., Ltd
$\{$ xinyu.zuo,pengfei.cao,yubo.chen,kliu,jzhao $\}$ @nlpr.ia.ac.cn
$\{$ pengweihua,chenyuguang $\}$ @baidu.com
## $\operatorname{Abstract}$
Current models for event causality identification (ECI)mainly adopt a supervised framework, which heavily rely on labeled data for training. Unfortunately, the scale of current annotated datasets is relatively limited, which cannot provide sufficient support for models to capture useful indicators from causal statements, especially for handing those new, unseen cases. To alleviate this problem, we propose a novel approach, shortly named CauSeRL, which leverages external causal statements for event causality identification. First of all, we design a self-supervised framework to learn context-specific causal patterns from external causal statements. Then, we adopt a contrastive transfer strategy to incorporate the learned contextspecific causal patterns into the target ECI model. Experimental results show that our method significantly outperforms previous methods on EventStoryLine and CausalTimeBank (+2.0 and +3.4 points on F1 value respectively).
## 1 Introduction
事件因果关系识别 (ECI) 旨在识别文本中事件之间的因果关系，这可以为深入的文本理解(Girju, 2003; Oh et al., 2013, 2017) 提供至关重要的线索。例如，在图 1 中，ECI 系统应在$S_1$中识别上述事件的两个因果关系：注意到$E1\stackrel{cause}{\longrightarrow}$警报$\stackrel{\:}{\text{ 和警报 }}_{E3}\stackrel{cause}{\longrightarrow}$运行$_E2$
迄今为止，大多数现有方法都将此任务视为分类问题，并且通常在注释数据(Hashimoto et al., 2014; Riaz and Girju, 2014b; Mirza and Tonelli, 2016; Hu and Walker, 2017b; Gao et al., 2019) 上训练 ECI 模型。但是，当前注释数据集的规模相对有限，其中迄今为止最大的数据集 EventStoryLine (Caselli and Vossen, 2017)仅包含 258 个文档，4316 个句子和 1770
<FigureHere>
Figure 1: $S_1$ is a labeled data that contains unseen causal events and their statement when training; $S_2$ is an external causal statement; The bottom $\bar{\text{illustrates the context-specific causal pattern in }S_{2}}$ could help identify the causality of unseen events in$S_1.$
个因果事件对。结果，在有限的注释示例中， 现有的 ECI 模型无法轻松地从因果陈述中捕获有用的指标，尤其是在处理那些新的，看不见的案例时
为了解决这个问题，Liu et al.(2020)采用了外部事件相关知识库(KB)来增强因果关系推断，其中这些 KB 存储了某些给定事件之间的固有因果关系。对于 KB 中那些看不见的事件和未标记的因果关系，Liu et al. (2020) 提出了一种基于提及掩码的推理器，以增强因果语句的表示。然而，这种基于提及掩码的推理器仍然只在人类注释的例子上训练。它仍然会受到数据限制，没有能力处理看不见的上下文。此外，Zuo et al. (2020)通过远程监督的标记训练数据提高了 ECI 的性能。然而，他们的模型仍然局限于自动生成数据的不满意质量
为了解决注释不充分的示例问题，我们采用了大量外部因果语句 (Sap et al., 2018; Mostafazadeh et al.,2020) ,这些语句可以支持上下文特定因果模式 (Liu et al., 2020) 的充分证据，以理解事件因果关系。例如，在图 1中，外部因果语句$S_{2}$支持的上下文特定因果模式有助于识别$S_{1}$中事件注意到的$_{E1}$和事件警报的$_E_{3}$的因果关系，这在仅使用标记数
|
[
[
0.19140625,
0.08612060546875,
0.81201171875,
0.08612060546875,
0.81201171875,
0.1009521484375,
0.19140625,
0.1009521484375
],
[
0.2032470703125,
0.12353515625,
0.8017578125,
0.12353515625,
0.8017578125,
0.1383056640625,
0.2032470703125,
0.1383056640625
],
[
0.349365234375,
0.141357421875,
0.66015625,
0.141357421875,
0.66015625,
0.1561279296875,
0.349365234375,
0.1561279296875
],
[
0.10662841796875,
0.155029296875,
0.90185546875,
0.1561279296875,
0.90185546875,
0.17333984375,
0.10662841796875,
0.1728515625
],
[
0.1099853515625,
0.171630859375,
0.89990234375,
0.1728515625,
0.89990234375,
0.190673828125,
0.1099853515625,
0.189453125
],
[
0.276123046875,
0.189453125,
0.73291015625,
0.190673828125,
0.73291015625,
0.204833984375,
0.276123046875,
0.2037353515625
],
[
0.2435302734375,
0.2078857421875,
0.767578125,
0.2078857421875,
0.767578125,
0.22265625,
0.2435302734375,
0.22265625
],
[
0.32421875,
0.2244873046875,
0.68603515625,
0.223876953125,
0.68603515625,
0.2381591796875,
0.32421875,
0.2392578125
],
[
0.2626953125,
0.275634765625,
0.343505859375,
0.27783203125,
0.343505859375,
0.29345703125,
0.2626953125,
0.29150390625
],
[
0.59619140625,
0.27978515625,
0.6396484375,
0.27978515625,
0.6396484375,
0.290283203125,
0.59619140625,
0.290283203125
],
[
0.6396484375,
0.281982421875,
0.67919921875,
0.281982421875,
0.67919921875,
0.289306640625,
0.6396484375,
0.289306640625
],
[
0.67822265625,
0.28076171875,
0.73291015625,
0.28076171875,
0.73291015625,
0.2880859375,
0.67822265625,
0.2880859375
],
[
0.63623046875,
0.29345703125,
0.6962890625,
0.29345703125,
0.6962890625,
0.300537109375,
0.63623046875,
0.300537109375
],
[
0.69775390625,
0.29443359375,
0.7314453125,
0.295654296875,
0.7314453125,
0.302734375,
0.69775390625,
0.3017578125
],
[
0.79833984375,
0.295654296875,
0.8447265625,
0.29638671875,
0.8447265625,
0.3046875,
0.79833984375,
0.303955078125
],
[
0.75390625,
0.298583984375,
0.77099609375,
0.298583984375,
0.77099609375,
0.302734375,
0.75390625,
0.302734375
],
[
0.60546875,
0.298583984375,
0.623046875,
0.298583984375,
0.623046875,
0.3017578125,
0.60546875,
0.3017578125
],
[
0.1485595703125,
0.305908203125,
0.4609375,
0.305908203125,
0.4609375,
0.319580078125,
0.1485595703125,
0.319580078125
],
[
0.59423828125,
0.305908203125,
0.7421875,
0.305908203125,
0.7421875,
0.31640625,
0.59423828125,
0.31640625
],
[
0.56689453125,
0.319580078125,
0.5751953125,
0.319580078125,
0.5751953125,
0.32373046875,
0.56689453125,
0.32373046875
],
[
0.60107421875,
0.318359375,
0.6591796875,
0.318359375,
0.6591796875,
0.322509765625,
0.60107421875,
0.322509765625
],
[
0.59619140625,
0.327880859375,
0.80419921875,
0.327880859375,
0.80419921875,
0.33837890625,
0.59619140625,
0.33837890625
],
[
0.14697265625,
0.320556640625,
0.4609375,
0.320556640625,
0.4609375,
0.334228515625,
0.14697265625,
0.334228515625
],
[
0.14697265625,
0.334228515625,
0.459228515625,
0.334228515625,
0.459228515625,
0.347412109375,
0.14697265625,
0.347412109375
],
[
0.6455078125,
0.345703125,
0.66650390625,
0.345703125,
0.66650390625,
0.3486328125,
0.6455078125,
0.3486328125
],
[
0.71875,
0.345703125,
0.748046875,
0.345703125,
0.748046875,
0.3486328125,
0.71875,
0.3486328125
],
[
0.30322265625,
0.347412109375,
0.459228515625,
0.347412109375,
0.459228515625,
0.361083984375,
0.30322265625,
0.361083984375
],
[
0.14697265625,
0.3486328125,
0.298828125,
0.3486328125,
0.298828125,
0.3623046875,
0.14697265625,
0.3623046875
],
[
0.64404296875,
0.358154296875,
0.75146484375,
0.358154296875,
0.75146484375,
0.365234375,
0.64404296875,
0.365234375
],
[
0.7958984375,
0.352783203125,
0.81201171875,
0.352783203125,
0.81201171875,
0.355712890625,
0.7958984375,
0.355712890625
],
[
0.1485595703125,
0.3623046875,
0.4609375,
0.3623046875,
0.4609375,
0.3759765625,
0.1485595703125,
0.3759765625
],
[
0.69287109375,
0.37060546875,
0.7119140625,
0.37060546875,
0.7119140625,
0.374755859375,
0.69287109375,
0.374755859375
],
[
0.14697265625,
0.377197265625,
0.4609375,
0.377197265625,
0.4609375,
0.39013671875,
0.14697265625,
0.39013671875
],
[
0.14697265625,
0.39013671875,
0.459228515625,
0.39013671875,
0.459228515625,
0.40380859375,
0.14697265625,
0.40380859375
],
[
0.145263671875,
0.40380859375,
0.4609375,
0.405029296875,
0.4609375,
0.419189453125,
0.145263671875,
0.418701171875
],
[
0.515625,
0.405029296875,
0.880859375,
0.405029296875,
0.880859375,
0.4150390625,
0.515625,
0.4150390625
],
[
0.14697265625,
0.418701171875,
0.45751953125,
0.419189453125,
0.45751953125,
0.432861328125,
0.14697265625,
0.431640625
],
[
0.51318359375,
0.418701171875,
0.88330078125,
0.418701171875,
0.88330078125,
0.431640625,
0.51318359375,
0.431640625
],
[
0.1485595703125,
0.43408203125,
0.459228515625,
0.43408203125,
0.459228515625,
0.44775390625,
0.1485595703125,
0.44775390625
],
[
0.515625,
0.43408203125,
0.87890625,
0.43408203125,
0.87890625,
0.44482421875,
0.515625,
0.44482421875
],
[
0.14697265625,
0.44775390625,
0.459228515625,
0.44775390625,
0.459228515625,
0.46142578125,
0.14697265625,
0.46142578125
],
[
0.5146484375,
0.446533203125,
0.88330078125,
0.446533203125,
0.88330078125,
0.460205078125,
0.5146484375,
0.460205078125
],
[
0.1485595703125,
0.4619140625,
0.45751953125,
0.46142578125,
0.45751953125,
0.474365234375,
0.1485595703125,
0.4755859375
],
[
0.51318359375,
0.46142578125,
0.88330078125,
0.46142578125,
0.88330078125,
0.474365234375,
0.51318359375,
0.474365234375
],
[
0.14697265625,
0.4755859375,
0.4609375,
0.4755859375,
0.4609375,
0.4892578125,
0.14697265625,
0.4892578125
],
[
0.51318359375,
0.476806640625,
0.5634765625,
0.476806640625,
0.5634765625,
0.488037109375,
0.51318359375,
0.488037109375
],
[
0.14697265625,
0.490478515625,
0.459228515625,
0.490478515625,
0.459228515625,
0.50439453125,
0.14697265625,
0.50439453125
],
[
0.1502685546875,
0.505859375,
0.45751953125,
0.505859375,
0.45751953125,
0.5166015625,
0.1502685546875,
0.5166015625
],
[
0.14697265625,
0.5185546875,
0.459228515625,
0.5185546875,
0.459228515625,
0.5322265625,
0.14697265625,
0.5322265625
],
[
0.515625,
0.52099609375,
0.880859375,
0.52099609375,
0.880859375,
0.53369140625,
0.515625,
0.53369140625
],
[
0.1485595703125,
0.533203125,
0.459228515625,
0.533203125,
0.459228515625,
0.54638671875,
0.1485595703125,
0.54638671875
],
[
0.5146484375,
0.53515625,
0.880859375,
0.53515625,
0.880859375,
0.548828125,
0.5146484375,
0.548828125
],
[
0.145263671875,
0.54638671875,
0.4609375,
0.5458984375,
0.4609375,
0.56005859375,
0.145263671875,
0.56103515625
],
[
0.515625,
0.5517578125,
0.88330078125,
0.5517578125,
0.88330078125,
0.56640625,
0.515625,
0.56640625
],
[
0.14697265625,
0.56103515625,
0.4609375,
0.56103515625,
0.4609375,
0.57470703125,
0.14697265625,
0.57470703125
],
[
0.51708984375,
0.5693359375,
0.60693359375,
0.5693359375,
0.60693359375,
0.58056640625,
0.51708984375,
0.58056640625
],
[
0.14697265625,
0.576171875,
0.459228515625,
0.576171875,
0.459228515625,
0.5888671875,
0.14697265625,
0.5888671875
],
[
0.1485595703125,
0.59130859375,
0.45751953125,
0.59130859375,
0.45751953125,
0.6015625,
0.1485595703125,
0.6015625
],
[
0.5341796875,
0.58740234375,
0.8818359375,
0.58740234375,
0.8818359375,
0.60107421875,
0.5341796875,
0.60107421875
],
[
0.1485595703125,
0.60400390625,
0.4609375,
0.60400390625,
0.4609375,
0.61767578125,
0.1485595703125,
0.61767578125
],
[
0.515625,
0.6015625,
0.880859375,
0.6015625,
0.880859375,
0.615234375,
0.515625,
0.615234375
],
[
0.14697265625,
0.6181640625,
0.4609375,
0.6181640625,
0.4609375,
0.6318359375,
0.14697265625,
0.6318359375
],
[
0.51708984375,
0.6181640625,
0.880859375,
0.6181640625,
0.880859375,
0.6318359375,
0.51708984375,
0.6318359375
],
[
0.14697265625,
0.6328125,
0.2880859375,
0.6328125,
0.2880859375,
0.646484375,
0.14697265625,
0.646484375
],
[
0.51708984375,
0.63525390625,
0.8818359375,
0.63525390625,
0.8818359375,
0.6484375,
0.51708984375,
0.6484375
],
[
0.515625,
0.65087890625,
0.880859375,
0.65087890625,
0.880859375,
0.66455078125,
0.515625,
0.66455078125
],
[
0.118408203125,
0.6630859375,
0.1351318359375,
0.6630859375,
0.1351318359375,
0.67138671875,
0.118408203125,
0.67138671875
],
[
0.15283203125,
0.6591796875,
0.26611328125,
0.6611328125,
0.26611328125,
0.67578125,
0.15283203125,
0.67333984375
],
[
0.515625,
0.66748046875,
0.880859375,
0.66748046875,
0.880859375,
0.681640625,
0.515625,
0.681640625
],
[
0.118408203125,
0.68408203125,
0.486083984375,
0.68408203125,
0.486083984375,
0.69775390625,
0.118408203125,
0.69775390625
],
[
0.515625,
0.68408203125,
0.8818359375,
0.68408203125,
0.8818359375,
0.69775390625,
0.515625,
0.69775390625
],
[
0.12005615234375,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.5146484375,
0.69873046875,
0.880859375,
0.69873046875,
0.880859375,
0.7119140625,
0.5146484375,
0.7119140625
],
[
0.118408203125,
0.71728515625,
0.48779296875,
0.71728515625,
0.48779296875,
0.73095703125,
0.118408203125,
0.73095703125
],
[
0.51708984375,
0.71630859375,
0.8818359375,
0.71630859375,
0.8818359375,
0.72998046875,
0.51708984375,
0.72998046875
],
[
0.12005615234375,
0.73291015625,
0.486083984375,
0.73291015625,
0.486083984375,
0.74658203125,
0.12005615234375,
0.74658203125
],
[
0.51708984375,
0.732421875,
0.8818359375,
0.732421875,
0.8818359375,
0.7451171875,
0.51708984375,
0.7451171875
],
[
0.12005615234375,
0.75048828125,
0.486083984375,
0.75048828125,
0.486083984375,
0.76416015625,
0.12005615234375,
0.76416015625
],
[
0.5146484375,
0.74658203125,
0.880859375,
0.74658203125,
0.880859375,
0.76025390625,
0.5146484375,
0.76025390625
],
[
0.15869140625,
0.76416015625,
0.343505859375,
0.7666015625,
0.343505859375,
0.7822265625,
0.15869140625,
0.77978515625
],
[
0.35107421875,
0.7626953125,
0.444091796875,
0.7685546875,
0.443359375,
0.78369140625,
0.349365234375,
0.77783203125
],
[
0.515625,
0.76416015625,
0.8818359375,
0.76416015625,
0.8818359375,
0.77783203125,
0.515625,
0.77783203125
],
[
0.1234130859375,
0.77099609375,
0.1435546875,
0.77099609375,
0.1435546875,
0.779296875,
0.1234130859375,
0.779296875
],
[
0.1351318359375,
0.78076171875,
0.486083984375,
0.7822265625,
0.486083984375,
0.79638671875,
0.1351318359375,
0.7958984375
],
[
0.5341796875,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.5341796875,
0.79638671875
],
[
0.118408203125,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.118408203125,
0.8125
],
[
0.515625,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.515625,
0.8125
],
[
0.118408203125,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.5146484375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.5146484375,
0.82763671875
],
[
0.118408203125,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51318359375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.12005615234375,
0.8466796875,
0.489501953125,
0.8466796875,
0.489501953125,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.515625,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.515625,
0.86083984375
],
[
0.12176513671875,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.87744140625,
0.12176513671875,
0.87744140625
],
[
0.515625,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.515625,
0.876953125
],
[
0.12176513671875,
0.87890625,
0.490234375,
0.87890625,
0.490234375,
0.892578125,
0.12176513671875,
0.892578125
],
[
0.515625,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.515625,
0.8935546875
],
[
0.12005615234375,
0.89453125,
0.489501953125,
0.89453125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.91015625,
0.5146484375,
0.91015625
]
] |
[
"通过外部因果陈述的自监督表征学习改进事件因果关系识别",
"Xinyu Zuo1,2, Pengfei Cao1,2, Yubo Chen1,2, Kang Liul,2, Jun Zhao1,2",
"Weihua Peng and Yuguang Chen3",
"INational Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China",
"2School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China",
"3Beijing BaiduNetcom ScienceTechnology Co.,Ltd",
"[ xinyu.zuo,pengfei.cao,yubo.chen,kliu,jzhao } @nlpr.ia.ac.cn",
"[pengweihua,chenyuguang ↓ @baidu.com",
"Abstract",
"The captain",
"noticed",
"euthepirates",
"othedeckand",
"alertede",
"emergency",
"CiC",
"T",
"Current models for event causality identi-",
"Billy finds his childhood teddy bear.",
".",
"Causps/Enahle",
"Billy gives his childhood teddy bear to his daughter",
"fication (ECI) mainly adopt a supervised",
"framework, which heavily rely on labeled",
"Pah",
"m",
"Unfortunately,the",
"data for training.",
"Entinvcall/give/alert/.. [Entity",
"",
"scale of current annotated datasets is rel-",
"Can",
"atively limited, which cannot provide suf-",
"ficient support for models to capture use",
"ful indicators from causal statements, espe-",
"Figure l:S,is alabeled datathat contains unseen",
"cially for handing those new, unseen cases.",
"causal events and their statement when training;",
"To alleviate this problem, we propose a",
"S2 is an external causal statement;Thebottom",
"novel approach, shortly named CauSeRL,",
"illustrates the context-specific causal pattern in S2",
"which leverages external causal statements",
"could help identify the causality of unseen events",
"for event causality identification. First",
"in Si:",
"of all, we design a self-supervised frame",
"work to learn context-specific causal pat",
"terns from external causal statements.",
"个因果事件对。结果，在有限的注释示例中",
"Then, we adopt a contrastive transfer strat-",
"现有的ECI模型无法轻松地从因果陈述中捕",
"egy to incorporate the learned context-",
"获有用的指标，尤其是在处理那些新的，看不",
"specific causal patterns into the target ECI",
"见的案例时",
"model. Experimental results show that our",
"methodsignificantlyoutperformsprevious",
"为了解决这个问题，Liuetal.（2020）采用",
"methods on EventStoryLine and Causal-",
"了外部事件相关知识库（KB）来增强因果关",
"TimeBank (+2.0 and +3.4 points on F1",
"系推断，其中这些KB存储了某些给定事件之",
"value respectively).",
"间的固有因果关系。对于KB中那些看不见",
"的事件和未标记的因果关系，Liuetal.（2020)",
"1",
"Introduction",
"提出了一种基于提及掩码的推理器，以增强因",
"事件因果关系识别（ECI）旨在识别文本中事",
"果语句的表示。然而，这种基于提及掩码的推",
"件之间的因果关系，这可以为深人的文本理解",
"理器仍然只在人类注释的例子上训练。它仍然",
"(Girju，2003:Ohetal.,2013,2017）提供至关",
"会受到数据限制，没有能力处理看不见的上下",
"重要的线索。例如，在图1中，ECI系统应在",
"文。此外，Zuoetal.（2020）通过远程监督的",
"S1中识别上述事件的两个因果关系：注意到",
"标记训练数据提高了ECI的性能。然而，他们",
"use警报E3和警报E3",
"ause运行E2",
"的模型仍然局限于自动生成数据的不满意质量",
"E1",
"迄今为止，大多数现有方法都将此任务视为",
"为了解决注释不充分的示例问题，我们",
"分类问题，并且通常在注释数据（Hashimoto",
"采用了大量外部因果语句（Sapetal.，2018;",
"et al., 2014; Riaz and Girju, 2014b; Mirza",
"Mostafazadehetal.，2020），这些语句可以支",
"and Tonelli, 2016: Hu and Walker, 2017b; Ga0",
"持上下文特定因果模式（Liuetal.,2020）的充",
"etal.，2019）上训练ECI模型。但是，当前注",
"分证据，以理解事件因果关系。例如，在图1",
"释数据集的规模相对有限，其中迄今为止最大",
"中，外部因果语句S2支持的上下文特定因果",
"的数据集EventStoryLine（CaselliandVossen，",
"模式有助于识别S1中事件注意到的E1和事",
"2017）仅包含258个文档，4316个句子和1770",
"件警报的E3的因果关系，这在仅使用标记数"
] |
{
"arxiv_id": "2106.01654",
"language": "zh",
"page_idx": 0
}
|
||
2106.01654-zh-1
|
2106.01654-zh-1.jpg
|
据进行训练时是看不见的。但是，与 ECI 任务的注释示例不同，外部因果语句中没有事件注释。因此，模型很难从中学习特定于上下文的因果模式来识别事件因果关系。为了解决这个问题，受 Grill et al. (2020)的启发，我们设计了一个自监督表示学习框架，以从外部因果语句中学习增强的因果表示。具体来说，我们迭代地采样两个外部因果语句，然后将它们中的每一个作为目标，以学习它们之间的共性。从直觉上讲，我们认为通过自我监督在不同因果陈述之间学习到的共性反映了这种特定于上下文的因果模式，这有助于在看不见的情况下识别事件因果关系
此外，为了将从外部因果语句中学到的特定于上下文的因果模式合并到目标 ECI 模型中， 我们采用了对比转移策略。具体而言，我们将自监督表示学习模块视为掌握丰富的外部因果陈述的教师模型，将目标 ECI 模型视为学生模型。从方法论上讲，我们使学生模型编码的因果事件的表示应接近教师模型掌握的因果表示，而使非因果事件的表示远离。这样，教师和学生模型之间的相互信息可以最大化 (Tian et al.,2020)。然后，可以将学习到的特定于上下文的因果模式自然地转移到 ECI 模型中， 并可以改进泛化
在实验中，我们在两个基准上评估我们的模型。实验结果表明，我们的模型达到了 SOTA 性能。然后，具体的证明表明，我们的自我监督的基于对比的框架对于特定于上下文的因果模式学习和转移的有效性
总之，贡献如下：
$\cdot$We propose a novel approach, shortly named CauSeRL , which could leverage external causal statements to identify the causalities between events.
$\bullet$ First of all, we design a self-supervised
framework to learn context-specific causal patterns from external causal statements. Then, we adopt a contrastive transfer strategy to incorporate the learned context-specific causal patterns into target ECI model for identification.
$\bullet$Experimental results on two benchmarks show that our model achieves the best performance.
## 2 Related Work
Event Causality Identification 到目前为止， 识别文本中隐含的因果关系已引起越来越多的
关注 (Hu and Walker, 2017a; Riaz and Girju, 2014b; Hashimoto et al., 2014; Riaz and Girju, 2014a, 2010; Do et al., 2011; Hidey and McKeown, 2016; Beamer and Girju, 2009; Hu et al., 2017; Hu and Walker, 2017b)。最近，一些关于事件因果关系的基准已经发布。Mirza et al (2014) , Mirza and Tonelli (2016)用基于规则的多筛法结合事件时间关系提取事件的因果关系。Mirza and Tonelli (2014) 注释了事件因果关系的因果-时间库。Caselli and Vossen (2017) 注释了 EventStoryLine 语料库，用于基于时间和因果关系注释的数据集 (Mostafazadeh et al., 2016) 在 320 个短篇小说中识别事件因果关系。Dunietz et al. (2017)呈现 BECauSE20,新版 BECauSE (Dunietz et al., 2015)的因果关系等七种关系
基于上述基准，Gao et al. (2019)对文档级结构进行建模，以识别事件的因果关系。Liu et al. (2020)通过提及掩码泛化和外部 KB 确定了事件因果关系。Zuo et al.(2020)通过远距离自动标记的训练数据提高了 ECI 的性能。然而，这些方法只依赖于小规模的标记数据。在本文中，我们引入外部因果陈述来帮助识别事件因果关系
Self-Supervised Representation Learning 自监督表示学习关心生成通常对许多任务(Weng, 2019) 有帮助的良好功能。Wu et al. (2018)提出了 MemoryBank, 它存储了所有数据的表示形式，并对一组随机的密钥进行了采样作为负面示例。He et al. (2020) 提供了一个无监督学习视觉表示的框架 MoCo,作为动态字典查找。Chen et al. (2020) 提出了 SimCLR, 它通过对比损失最大化同一样本的不同增强视图之间的一致性来学习视觉输入的表示。Grill et al. (2020)声称一种新颖的表示学习框架依赖于两个神经网络 BYOL,而不使用负样本。CURL (Srinivas et al., 2020) 将上述思想应用于强化学习。受其启发，我们设计了一个自监督框架，从外部因果陈述中学习特定于上下文的因果模式，并采用对比转移策略将其纳人目标 ECI 模型
3 Methodology
如图 2 所示，CauSeRL 的整个流水线过程分
为两个主要阶段
$\bullet$ Self-supervised causal representation
3.1 In
learning(SelfRL,Sec.
this stage, we design a self-supervised representation learning module to learn enhanced causal representations by
|
[
[
0.118408203125,
0.07598876953125,
0.48779296875,
0.07598876953125,
0.48779296875,
0.08905029296875,
0.118408203125,
0.08905029296875
],
[
0.51318359375,
0.07598876953125,
0.880859375,
0.07659912109375,
0.880859375,
0.0914306640625,
0.51318359375,
0.09027099609375
],
[
0.12005615234375,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.51318359375,
0.0914306640625,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10809326171875,
0.51318359375,
0.105712890625
],
[
0.12005615234375,
0.10986328125,
0.48779296875,
0.10986328125,
0.48779296875,
0.12353515625,
0.12005615234375,
0.12353515625
],
[
0.5146484375,
0.1092529296875,
0.884765625,
0.1092529296875,
0.884765625,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.12005615234375,
0.1258544921875,
0.486083984375,
0.1258544921875,
0.486083984375,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.51318359375,
0.1258544921875,
0.880859375,
0.1258544921875,
0.880859375,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.12176513671875,
0.141357421875,
0.48779296875,
0.141357421875,
0.48779296875,
0.155029296875,
0.12176513671875,
0.155029296875
],
[
0.5146484375,
0.141357421875,
0.8818359375,
0.141357421875,
0.8818359375,
0.155029296875,
0.5146484375,
0.155029296875
],
[
0.1234130859375,
0.157958984375,
0.48779296875,
0.157958984375,
0.48779296875,
0.171630859375,
0.1234130859375,
0.171630859375
],
[
0.5146484375,
0.1561279296875,
0.8818359375,
0.157958984375,
0.8818359375,
0.171630859375,
0.5146484375,
0.169189453125
],
[
0.12176513671875,
0.17333984375,
0.48779296875,
0.17333984375,
0.48779296875,
0.18701171875,
0.12176513671875,
0.18701171875
],
[
0.5146484375,
0.17333984375,
0.8818359375,
0.17333984375,
0.8818359375,
0.18701171875,
0.5146484375,
0.18701171875
],
[
0.12176513671875,
0.189453125,
0.48779296875,
0.189453125,
0.48779296875,
0.203125,
0.12176513671875,
0.203125
],
[
0.5146484375,
0.188232421875,
0.880859375,
0.188232421875,
0.880859375,
0.201904296875,
0.5146484375,
0.201904296875
],
[
0.12005615234375,
0.2060546875,
0.489501953125,
0.2060546875,
0.489501953125,
0.2197265625,
0.12005615234375,
0.2197265625
],
[
0.515625,
0.2060546875,
0.8818359375,
0.2060546875,
0.8818359375,
0.2197265625,
0.515625,
0.2197265625
],
[
0.12005615234375,
0.221435546875,
0.489501953125,
0.221435546875,
0.489501953125,
0.235107421875,
0.12005615234375,
0.235107421875
],
[
0.5146484375,
0.2203369140625,
0.880859375,
0.2203369140625,
0.880859375,
0.235107421875,
0.5146484375,
0.235107421875
],
[
0.12005615234375,
0.2381591796875,
0.489501953125,
0.2381591796875,
0.489501953125,
0.251708984375,
0.12005615234375,
0.251708984375
],
[
0.5146484375,
0.2381591796875,
0.8818359375,
0.2381591796875,
0.8818359375,
0.251708984375,
0.5146484375,
0.251708984375
],
[
0.12176513671875,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.267333984375,
0.12176513671875,
0.267333984375
],
[
0.5146484375,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.5146484375,
0.267333984375
],
[
0.12005615234375,
0.270751953125,
0.2476806640625,
0.270751953125,
0.2476806640625,
0.283935546875,
0.12005615234375,
0.283935546875
],
[
0.515625,
0.270751953125,
0.87890625,
0.270751953125,
0.87890625,
0.28076171875,
0.515625,
0.28076171875
],
[
0.1376953125,
0.287353515625,
0.489501953125,
0.287353515625,
0.489501953125,
0.300537109375,
0.1376953125,
0.300537109375
],
[
0.5146484375,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.5146484375,
0.2998046875
],
[
0.12005615234375,
0.302734375,
0.489501953125,
0.302734375,
0.489501953125,
0.31640625,
0.12005615234375,
0.31640625
],
[
0.51123046875,
0.300537109375,
0.8818359375,
0.3017578125,
0.8818359375,
0.31640625,
0.51123046875,
0.3154296875
],
[
0.12005615234375,
0.319580078125,
0.484375,
0.319580078125,
0.484375,
0.330078125,
0.12005615234375,
0.330078125
],
[
0.5146484375,
0.318359375,
0.66162109375,
0.318359375,
0.66162109375,
0.33203125,
0.5146484375,
0.33203125
],
[
0.12176513671875,
0.3349609375,
0.48779296875,
0.3349609375,
0.48779296875,
0.3486328125,
0.12176513671875,
0.3486328125
],
[
0.5322265625,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.5322265625,
0.3486328125
],
[
0.12176513671875,
0.3515625,
0.48779296875,
0.3515625,
0.48779296875,
0.365234375,
0.12176513671875,
0.365234375
],
[
0.5146484375,
0.3515625,
0.8818359375,
0.3515625,
0.8818359375,
0.365234375,
0.5146484375,
0.365234375
],
[
0.12176513671875,
0.36767578125,
0.48779296875,
0.36767578125,
0.48779296875,
0.38134765625,
0.12176513671875,
0.38134765625
],
[
0.5146484375,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.5146484375,
0.38134765625
],
[
0.12005615234375,
0.3818359375,
0.486083984375,
0.3818359375,
0.486083984375,
0.3955078125,
0.12005615234375,
0.3955078125
],
[
0.515625,
0.383056640625,
0.8818359375,
0.383056640625,
0.8818359375,
0.396728515625,
0.515625,
0.396728515625
],
[
0.12176513671875,
0.399658203125,
0.489501953125,
0.399658203125,
0.489501953125,
0.413330078125,
0.12176513671875,
0.413330078125
],
[
0.5146484375,
0.399658203125,
0.8818359375,
0.399658203125,
0.8818359375,
0.413330078125,
0.5146484375,
0.413330078125
],
[
0.12005615234375,
0.4150390625,
0.489501953125,
0.4150390625,
0.489501953125,
0.4287109375,
0.12005615234375,
0.4287109375
],
[
0.5146484375,
0.4150390625,
0.8818359375,
0.4150390625,
0.8818359375,
0.4287109375,
0.5146484375,
0.4287109375
],
[
0.12005615234375,
0.431640625,
0.489501953125,
0.431640625,
0.489501953125,
0.4453125,
0.12005615234375,
0.4453125
],
[
0.51318359375,
0.43115234375,
0.88330078125,
0.43115234375,
0.88330078125,
0.44482421875,
0.51318359375,
0.44482421875
],
[
0.12005615234375,
0.44775390625,
0.491943359375,
0.44775390625,
0.491943359375,
0.46142578125,
0.12005615234375,
0.46142578125
],
[
0.515625,
0.448974609375,
0.62646484375,
0.448974609375,
0.62646484375,
0.460205078125,
0.515625,
0.460205078125
],
[
0.118408203125,
0.46435546875,
0.2493896484375,
0.46435546875,
0.2493896484375,
0.47802734375,
0.118408203125,
0.47802734375
],
[
0.51904296875,
0.473876953125,
0.8818359375,
0.473876953125,
0.8818359375,
0.48681640625,
0.51904296875,
0.48681640625
],
[
0.1368408203125,
0.479736328125,
0.48779296875,
0.478515625,
0.48779296875,
0.493408203125,
0.1368408203125,
0.49462890625
],
[
0.5146484375,
0.4892578125,
0.8818359375,
0.4892578125,
0.8818359375,
0.5029296875,
0.5146484375,
0.5029296875
],
[
0.12005615234375,
0.496337890625,
0.48779296875,
0.496337890625,
0.48779296875,
0.51025390625,
0.12005615234375,
0.51025390625
],
[
0.5146484375,
0.505859375,
0.8818359375,
0.505859375,
0.8818359375,
0.51953125,
0.5146484375,
0.51953125
],
[
0.12005615234375,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.525390625,
0.12005615234375,
0.525390625
],
[
0.51318359375,
0.51953125,
0.8818359375,
0.52099609375,
0.8818359375,
0.53515625,
0.51318359375,
0.53369140625
],
[
0.12176513671875,
0.529296875,
0.48779296875,
0.529296875,
0.48779296875,
0.5419921875,
0.12176513671875,
0.5419921875
],
[
0.51318359375,
0.5361328125,
0.880859375,
0.5361328125,
0.880859375,
0.5498046875,
0.51318359375,
0.5498046875
],
[
0.12005615234375,
0.54443359375,
0.322509765625,
0.54443359375,
0.322509765625,
0.55810546875,
0.12005615234375,
0.55810546875
],
[
0.5146484375,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.56787109375,
0.5146484375,
0.56787109375
],
[
0.1376953125,
0.56005859375,
0.276123046875,
0.56005859375,
0.276123046875,
0.57373046875,
0.1376953125,
0.57373046875
],
[
0.51708984375,
0.57080078125,
0.87890625,
0.57080078125,
0.87890625,
0.58056640625,
0.51708984375,
0.58056640625
],
[
0.1351318359375,
0.59033203125,
0.48779296875,
0.59033203125,
0.48779296875,
0.60400390625,
0.1351318359375,
0.60400390625
],
[
0.515625,
0.5859375,
0.880859375,
0.5859375,
0.880859375,
0.599609375,
0.515625,
0.599609375
],
[
0.154541015625,
0.60498046875,
0.489501953125,
0.60546875,
0.489501953125,
0.62060546875,
0.154541015625,
0.619140625
],
[
0.515625,
0.6015625,
0.8818359375,
0.6015625,
0.8818359375,
0.615234375,
0.515625,
0.615234375
],
[
0.1561279296875,
0.62255859375,
0.489501953125,
0.62255859375,
0.489501953125,
0.6357421875,
0.1561279296875,
0.6357421875
],
[
0.5146484375,
0.61767578125,
0.880859375,
0.61767578125,
0.880859375,
0.63134765625,
0.5146484375,
0.63134765625
],
[
0.154541015625,
0.63720703125,
0.370361328125,
0.63818359375,
0.370361328125,
0.65283203125,
0.154541015625,
0.65185546875
],
[
0.515625,
0.63525390625,
0.8818359375,
0.63525390625,
0.8818359375,
0.6484375,
0.515625,
0.6484375
],
[
0.515625,
0.65087890625,
0.8818359375,
0.65087890625,
0.8818359375,
0.66455078125,
0.515625,
0.66455078125
],
[
0.1368408203125,
0.66748046875,
0.48779296875,
0.66748046875,
0.48779296875,
0.68115234375,
0.1368408203125,
0.68115234375
],
[
0.515625,
0.66650390625,
0.880859375,
0.66650390625,
0.880859375,
0.6796875,
0.515625,
0.6796875
],
[
0.1561279296875,
0.68310546875,
0.489501953125,
0.68310546875,
0.489501953125,
0.69677734375,
0.1561279296875,
0.69677734375
],
[
0.5146484375,
0.68115234375,
0.8818359375,
0.681640625,
0.8818359375,
0.69677734375,
0.5146484375,
0.6953125
],
[
0.154541015625,
0.69970703125,
0.489501953125,
0.69873046875,
0.489501953125,
0.7119140625,
0.154541015625,
0.71337890625
],
[
0.515625,
0.69873046875,
0.8818359375,
0.69873046875,
0.8818359375,
0.7119140625,
0.515625,
0.7119140625
],
[
0.1561279296875,
0.71533203125,
0.489501953125,
0.71533203125,
0.489501953125,
0.728515625,
0.1561279296875,
0.728515625
],
[
0.5146484375,
0.71337890625,
0.87890625,
0.71337890625,
0.87890625,
0.72705078125,
0.5146484375,
0.72705078125
],
[
0.1561279296875,
0.73095703125,
0.489501953125,
0.73095703125,
0.489501953125,
0.7451171875,
0.1561279296875,
0.7451171875
],
[
0.51318359375,
0.73095703125,
0.615234375,
0.73095703125,
0.615234375,
0.7451171875,
0.51318359375,
0.7451171875
],
[
0.1561279296875,
0.74755859375,
0.48779296875,
0.74755859375,
0.48779296875,
0.76123046875,
0.1561279296875,
0.76123046875
],
[
0.5146484375,
0.76123046875,
0.5498046875,
0.76123046875,
0.5498046875,
0.7724609375,
0.5146484375,
0.7724609375
],
[
0.54736328125,
0.75732421875,
0.6630859375,
0.7587890625,
0.6630859375,
0.77490234375,
0.54736328125,
0.7724609375
],
[
0.15283203125,
0.76416015625,
0.418212890625,
0.76318359375,
0.418212890625,
0.77783203125,
0.15283203125,
0.779296875
],
[
0.515625,
0.78515625,
0.880859375,
0.78515625,
0.880859375,
0.798828125,
0.515625,
0.798828125
],
[
0.1351318359375,
0.79150390625,
0.489501953125,
0.79052734375,
0.489501953125,
0.8046875,
0.1351318359375,
0.8056640625
],
[
0.515625,
0.8017578125,
0.6455078125,
0.8017578125,
0.6455078125,
0.8154296875,
0.515625,
0.8154296875
],
[
0.1561279296875,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.1561279296875,
0.82177734375
],
[
0.1561279296875,
0.82470703125,
0.235107421875,
0.82470703125,
0.235107421875,
0.8388671875,
0.1561279296875,
0.8388671875
],
[
0.53076171875,
0.83056640625,
0.67529296875,
0.83056640625,
0.67529296875,
0.845703125,
0.53076171875,
0.845703125
],
[
0.6708984375,
0.830078125,
0.8818359375,
0.83203125,
0.8818359375,
0.8466796875,
0.6708984375,
0.84423828125
],
[
0.5498046875,
0.84716796875,
0.8671875,
0.84716796875,
0.8671875,
0.8623046875,
0.5498046875,
0.8623046875
],
[
0.1175537109375,
0.85400390625,
0.274658203125,
0.85400390625,
0.274658203125,
0.86767578125,
0.1175537109375,
0.86767578125
],
[
0.861328125,
0.85107421875,
0.880859375,
0.85107421875,
0.880859375,
0.859375,
0.861328125,
0.859375
],
[
0.5498046875,
0.8623046875,
0.884765625,
0.8623046875,
0.884765625,
0.8798828125,
0.5498046875,
0.8798828125
],
[
0.124267578125,
0.87744140625,
0.486083984375,
0.87890625,
0.486083984375,
0.8935546875,
0.124267578125,
0.892578125
],
[
0.5498046875,
0.8798828125,
0.880859375,
0.8798828125,
0.880859375,
0.89453125,
0.5498046875,
0.89453125
],
[
0.1234130859375,
0.89794921875,
0.486083984375,
0.89794921875,
0.486083984375,
0.90771484375,
0.1234130859375,
0.90771484375
],
[
0.5498046875,
0.89453125,
0.8818359375,
0.8955078125,
0.8818359375,
0.91015625,
0.5498046875,
0.9091796875
]
] |
[
"据进行训练时是看不见的。但是，与ECI任务",
"关注(Hu and Walker,2017a;Riaz and Girju,",
"的注释示例不同，外部因果语句中没有事件注",
"2014b; Hashimoto et al., 2014; Riaz and Girju,",
"释。因此，模型很难从中学习特定于上下文的",
"2014a, 2010; Do et al., 2011; Hidey and McKe-",
"因果模式来识别事件因果关系。为了解决这个",
"own, 2016; Beamer and Girju, 2009; Hu et al.",
"问题，受Grilletal.（2020）的启发，我们设计",
"2017HuandWalker,2017b）。最近，一些关",
"了一个自监督表示学习框架，以从外部因果语",
"于事件因果关系的基准已经发布。Mirzaetal.",
"句中学习增强的因果表示。具体来说，我们送",
"(2014），MirzaandTonelli（2016）用基于规则",
"代地采样两个外部因果语句，然后将它们中的",
"的多筛法结合事件时间关系提取事件的因果关",
"每一个作为目标，以学习它们之间的共性。从",
"系。MirzaandTonelli（2014）注释了事件因果",
"直觉上讲，我们认为通过自我监督在不同因果",
"关系的因果-时间库。CaselliandVossen（2017）",
"陈述之间学习到的共性反映了这种特定于上下",
"注释了EventStoryLine语料库，用于基于时",
"文的因果模式，这有助于在看不见的情况下识",
"间和因果关系注释的数据集（Mostafazadeh",
"别事件因果关系",
"etal2016）在320个短篇小说中识别事件因",
"此外，为了将从外部因果语句中学到的特定",
"果关系。Dunietzetal.（2017）呈现BECauSE2。",
"于上下文的因果模式合并到目标ECI模型中，",
"0，新版BECauSE（Dunietzetal.,2015）的因",
"我们采用了对比转移策略。具体而言，我们将",
"果关系等七种关系",
"自监督表示学习模块视为掌握丰富的外部因果",
"基于上述基准，Gaoetal.（2019）对文档级",
"陈述的教师模型，将目标ECI模型视为学生",
"结构进行建模，以识别事件的因果关系。Liu",
"模型。从方法论上讲，我们使学生模型编码的",
"etal.（2020）通过提及掩码泛化和外部KB确",
"因果事件的表示应接近教师模型掌握的因果表",
"定了事件因果关系。Zuoetal.（2020）通过远",
"示，而使非因果事件的表示远离。这样，教师",
"距离自动标记的训练数据提高了ECI的性能。",
"和学生模型之间的相互信息可以最大化（Tian",
"然而，这些方法只依赖于小规模的标记数据。",
"etal.，2020）。然后，可以将学习到的特定于",
"在本文中，我们引人外部因果陈述来帮助识别",
"上下文的因果模式自然地转移到ECI模型中，",
"事件因果关系",
"并可以改进泛化",
"Self-Supervised Representation Learning自",
"在实验中，我们在两个基准上评估我们的模",
"监督表示学习关心生成通常对许多任务",
"型。实验结果表明，我们的模型达到了SOTA",
"(Weng，2019）有帮助的良好功能。Wuetal.",
"性能。然后，具体的证明表明，我们的自我监",
"(2018）提出了MemoryBank，它存储了所有数",
"督的基于对比的框架对于特定于上下文的因果",
"据的表示形式，并对一组随机的密钥进行了采",
"模式学习和转移的有效性",
"样作为负面示例。Heetal.（2020）提供了一个",
"总之，贡献如下：",
"无监督学习视觉表示的框架MoCo：作为动态",
": We propose a novel approach, shortly",
"字典查找。Chenetal.（2020）提出了SimCLR",
"named CauSeRL ,which could leverage",
"它通过对比损失最大化同一样本的不同增强视",
"external causal statements to identify the",
"图之间的一致性来学习视觉输入的表示。Grill",
"causalities between events.",
"etal.（2020）声称一种新颖的表示学习框架依",
"赖于两个神经网络BYOL，而不使用负样本。",
" First of all, we design a self-supervised",
"CURL（Srinivasetal，2020）将上述思想应用",
"framework to learn context-specific causal",
"于强化学习。受其启发，我们设计了一个自监",
"patterns from external causal statements.",
"督框架，从外部因果陈述中学习特定于上下文",
"Then, we adopt a contrastive trans-",
"的因果模式，并采用对比转移策略将其纳入目",
"fer strategy to incorporate the learned",
"标ECI模型",
"context-specific causal patterns into tar-",
"3",
"Methodology",
"get ECI model for identification.",
"如图2所示，CauSeRL的整个流水线过程分",
". Experimental results on two benchmarks",
"为两个主要阶段",
"show that our model achieves the best per-",
"formance.",
": Self-supervised",
"causalrepresentation",
"learning（SelfRLSec.3.1).",
"2Related Work",
"In",
"this stage, we design a self-supervised",
"EventCausalityIdentification到目前为止，",
"representation learning module to learn",
"识别文本中隐含的因果关系已引起越来越多的",
"enhanced causal representations by"
] |
{
"arxiv_id": "2106.01654",
"language": "zh",
"page_idx": 1
}
|
||
2106.01654-zh-4
|
2106.01654-zh-4.jpg
|
<table>
<tbody>
<tr>
<td>${\overline{{\mathrm{Resource}}}}$</td>
<td>Original Statement ${\textbf{E Form}}$</td>
<td>$\overline{\mathrm{Form}}$ onverter lallsa</td>
</tr>
<tr>
<td>GLU- SPE</td>
<td>$\overline{\mathrm{Billy}}$ ${\mathrm{finds}}$ his $\mathop{\mathrm{childhood}}$ $\overline{\mathrm{teddy}}$ bear > , $Cause/Enable$ $e~>~{\mathrm{Billy}}~{\mathrm{gives}}~{\mathrm{his}}~{\mathrm{childhood}}$ teddy bear to his daughter</td>
<td>Billy fnds his childhood teddy bear billy gives his childhood teddy bear ta his daughter.</td>
</tr>
<tr>
<td>GLU- GEN</td>
<td>Someone\_A ${\overline{{\mathrm{finds}}}}$ Something\_A > $Cause/Enable>$Someone\_A gives Something\_A to Someone B</td>
<td>Someone\_A finds Something\_A, Some one\_A gives Something\_A to Some- one B.</td>
</tr>
<tr>
<td>ATOMIC</td>
<td>PersonX follows PersonY into room>$oWant>$to know why PersonX is following them</td>
<td>PersonX follows PersonY into room, to know why PersonX is following them.</td>
</tr>
<tr>
<td>DISTANT</td>
<td>Fisk was shot to death by his mistress's new lover and Fisk's ex-business partner.</td>
<td>Fisk was shot to death by his mistress's new lover and Fisk's ex-business partner.</td>
</tr>
</tbody>
</table>
Table 1: The original and converted form (the input form of SelfRL) of different causal statements from three resources. GLU-SPE and GLU-GEN denote the specific and general statements from GLUCOSE respectively.
获取四种类型的外部因果语句。表 1 说明了SelfRL 的原始形式和转换后的输入形式 (Sec。3.1 )来自三个不同资源的因果陈述。
其中，$P^+$和$P$分别是一个批次中的因果事件对和所有事件对，$T$是调节浓度水平的温度$\mathcal{D}$是$\ell_2$ -距离函数，用于测量两个表示的距离Learning of Event Causality Identifier 对于事件因果关系标识符的训练，我们将对比损失添加到基本分类损失中，这可以指导标识符从SelfRL 学习增强因果表示中隐含的特定于上下文的因果模式。如算法 1 所示，我们最小化$L_\lambda$ , 随 机 梯 度 更 新 $\lambda$ 如下：
$\bullet$ GLUCOSE (Mostafazadeh et al., 2020):
a large-scale dataset of implicit commonsense knowledge, encoded as causal explanatory mini-theories inspired by cognitive psychology. Each GLUCOSE explanation is stated both as a specific statement (grounded in a given context, GLUSPE in Table 1 ) and a corresponding general rule ( $applicable to other contexts$, GLU-GEN in Table 1 ).
$$\begin{array}{c}\mathcal{L}_\lambda=\mathcal{L}_\lambda^{stu}+\mathcal{L}_\lambda^{con},\\\lambda\leftarrow\eta_{stu}\nabla_\lambda\mathcal{L}_\lambda,\end{array}$$
(8) (9)
$\bullet$ ATOMIC (Sap et al., 2018): an atlas of
machine commonsense, as a step toward
addressing the rich spectrum of inferen-
其中，$\eta_{stu}$是标识符的学习率。为了进行评估， 我们在没有对比传递模块的情况下预测输入事件对的因果关系。此外，$\mathcal{L}_\lambda^{con}$中的$T$间接起到调整$\mathcal{L}_\lambda^{stu}$和$\mathcal{L}_\lambda^{con}$的影响权重的作用。具体来说，对于教学，我们以在线网络的学习过的$Enc_{\theta}$作为编码器，冻结其参数，以提供外部因果语句的增强因果表示，以进行对比表示传递
tial knowledge that is crucial for common-
sense reasoning.
$\bullet$ DISTANT (Zuo et al., 2020): the auto-
matically labeled training data for ECI via distant supervision that expresses the causal semantics between events.
4 Experiments
### 4.1 Experimental Setup
Dataset and Evaluation Metrics for ECI 我们的实验在两个主要基准上进行，包${\text{括：EventStoryLinev}0。9}{\mathrm{~(ESC)~(Caselli~and}}$ Vossen,2017) 上述；和(2)因果时间银行(CTB) (Mirza and Tonelli, 2014),其中包含 184个文档，6813 个事件和 318 个因果事件对。与以前的方法相同，我们使用 ESC 的最后两个主题作为两个数据集的开发集。对于评估，我们${\text{采用 Precision(P)、Recall(R)和 F1-score(F1)}}$ 作为评估指标。我们分别对 ESC 和 CTB 进行 5 倍和 10 倍交叉验证，与以前的方法相同。所有结果均为三次独立实验的平均值
Parameters Settings 在实现中，所有 BERT 模块都在 BERT-Base 体系结构$^{2}$上实现，该体系结构具有 12 层，768 层和 12头。我们采用单层 BiLSTM (Hochreiter and Schmidhuber, 1997)作为$Enc_\theta$和$Enc_\mathrm{\delta}$ 。对于参数，我们将 SelfRL($\eta_tea$)和标识符($\eta_stu$)的学习率分别设置为 1e-5 和 2e-5。对比传输模块和BiLSTM 的隐藏层中的空间大小均设置为 50。并且我们分别设置 SelfRL 中移动平均线的衰减率$\tau$和对比损失$\mathcal{L}_\lambda^{con}$的温度为 0。996 和0。1 在开发集上调谐。此外，我们还在开发集上将 SelfRL 和标识符的批处理大小分别调整为 48 和 16。我们应用早期停止和 AdamW 梯
Data Preparation for Self-Supervised Causal Representation Learning 我们从三种资源中
$^2$https://github.com/google-research/bert
|
[
[
0.2880859375,
0.07244873046875,
0.515625,
0.07244873046875,
0.515625,
0.08612060546875,
0.2880859375,
0.08612060546875
],
[
0.609375,
0.07244873046875,
0.8505859375,
0.07244873046875,
0.8505859375,
0.08612060546875,
0.609375,
0.08612060546875
],
[
0.587890625,
0.0843505859375,
0.86572265625,
0.08489990234375,
0.86572265625,
0.09979248046875,
0.587890625,
0.09857177734375
],
[
0.1318359375,
0.0736083984375,
0.204833984375,
0.0736083984375,
0.204833984375,
0.08489990234375,
0.1318359375,
0.08489990234375
],
[
0.28369140625,
0.08489990234375,
0.32421875,
0.0872802734375,
0.322509765625,
0.09857177734375,
0.2822265625,
0.0968017578125
],
[
0.337646484375,
0.08612060546875,
0.365234375,
0.08612060546875,
0.365234375,
0.097412109375,
0.337646484375,
0.097412109375
],
[
0.38623046875,
0.08612060546875,
0.45751953125,
0.08612060546875,
0.45751953125,
0.097412109375,
0.38623046875,
0.097412109375
],
[
0.474365234375,
0.0872802734375,
0.51904296875,
0.0872802734375,
0.51904296875,
0.097412109375,
0.474365234375,
0.097412109375
],
[
0.54150390625,
0.0885009765625,
0.5751953125,
0.0885009765625,
0.5751953125,
0.0968017578125,
0.54150390625,
0.0968017578125
],
[
0.225830078125,
0.0968017578125,
0.5751953125,
0.0968017578125,
0.5751953125,
0.10986328125,
0.225830078125,
0.10986328125
],
[
0.2301025390625,
0.0872802734375,
0.267822265625,
0.0872802734375,
0.267822265625,
0.0968017578125,
0.2301025390625,
0.0968017578125
],
[
0.1368408203125,
0.097412109375,
0.2091064453125,
0.097412109375,
0.2091064453125,
0.1092529296875,
0.1368408203125,
0.1092529296875
],
[
0.59033203125,
0.0968017578125,
0.869140625,
0.0968017578125,
0.869140625,
0.10986328125,
0.59033203125,
0.10986328125
],
[
0.2275390625,
0.10809326171875,
0.414794921875,
0.1092529296875,
0.414794921875,
0.122314453125,
0.2275390625,
0.1217041015625
],
[
0.59375,
0.10986328125,
0.6826171875,
0.10986328125,
0.6826171875,
0.1217041015625,
0.59375,
0.1217041015625
],
[
0.2283935546875,
0.122314453125,
0.316650390625,
0.122314453125,
0.316650390625,
0.1341552734375,
0.2283935546875,
0.1341552734375
],
[
0.349365234375,
0.122314453125,
0.387939453125,
0.122314453125,
0.387939453125,
0.1341552734375,
0.349365234375,
0.1341552734375
],
[
0.42236328125,
0.122314453125,
0.52294921875,
0.122314453125,
0.52294921875,
0.1341552734375,
0.42236328125,
0.1341552734375
],
[
0.5634765625,
0.1258544921875,
0.57421875,
0.1258544921875,
0.57421875,
0.1312255859375,
0.5634765625,
0.1312255859375
],
[
0.59033203125,
0.1217041015625,
0.86962890625,
0.1217041015625,
0.86962890625,
0.1353759765625,
0.59033203125,
0.1353759765625
],
[
0.1351318359375,
0.1341552734375,
0.210693359375,
0.1341552734375,
0.210693359375,
0.1455078125,
0.1351318359375,
0.1455078125
],
[
0.2275390625,
0.1318359375,
0.57666015625,
0.1318359375,
0.57666015625,
0.1455078125,
0.2275390625,
0.1455078125
],
[
0.591796875,
0.1341552734375,
0.869140625,
0.1341552734375,
0.869140625,
0.1478271484375,
0.591796875,
0.1478271484375
],
[
0.2283935546875,
0.146728515625,
0.334228515625,
0.146728515625,
0.334228515625,
0.157958984375,
0.2283935546875,
0.157958984375
],
[
0.59033203125,
0.146728515625,
0.64404296875,
0.146728515625,
0.64404296875,
0.157958984375,
0.59033203125,
0.157958984375
],
[
0.2275390625,
0.15673828125,
0.580078125,
0.15673828125,
0.580078125,
0.17041015625,
0.2275390625,
0.17041015625
],
[
0.59033203125,
0.15673828125,
0.869140625,
0.15673828125,
0.869140625,
0.17041015625,
0.59033203125,
0.17041015625
],
[
0.139404296875,
0.1644287109375,
0.2091064453125,
0.1644287109375,
0.2091064453125,
0.17578125,
0.139404296875,
0.17578125
],
[
0.225830078125,
0.16748046875,
0.484375,
0.169189453125,
0.484375,
0.182861328125,
0.225830078125,
0.18115234375
],
[
0.59033203125,
0.169189453125,
0.8486328125,
0.169189453125,
0.8486328125,
0.182861328125,
0.59033203125,
0.182861328125
],
[
0.225830078125,
0.179931640625,
0.57666015625,
0.179931640625,
0.57666015625,
0.193603515625,
0.225830078125,
0.193603515625
],
[
0.59033203125,
0.18115234375,
0.869140625,
0.18115234375,
0.869140625,
0.19482421875,
0.59033203125,
0.19482421875
],
[
0.1351318359375,
0.188232421875,
0.210693359375,
0.188232421875,
0.210693359375,
0.1995849609375,
0.1351318359375,
0.1995849609375
],
[
0.2275390625,
0.193603515625,
0.43994140625,
0.193603515625,
0.43994140625,
0.207275390625,
0.2275390625,
0.207275390625
],
[
0.587890625,
0.1923828125,
0.86962890625,
0.193603515625,
0.86962890625,
0.2078857421875,
0.587890625,
0.207275390625
],
[
0.118408203125,
0.218505859375,
0.8818359375,
0.218505859375,
0.8818359375,
0.232177734375,
0.118408203125,
0.232177734375
],
[
0.12005615234375,
0.2327880859375,
0.88330078125,
0.2327880859375,
0.88330078125,
0.2464599609375,
0.12005615234375,
0.2464599609375
],
[
0.12005615234375,
0.248779296875,
0.2091064453125,
0.248779296875,
0.2091064453125,
0.260009765625,
0.12005615234375,
0.260009765625
],
[
0.12005615234375,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.12005615234375,
0.2998046875
],
[
0.51318359375,
0.2861328125,
0.87890625,
0.2861328125,
0.87890625,
0.2998046875,
0.51318359375,
0.2998046875
],
[
0.12005615234375,
0.302734375,
0.490234375,
0.302734375,
0.490234375,
0.317626953125,
0.12005615234375,
0.317626953125
],
[
0.51123046875,
0.3017578125,
0.88330078125,
0.302734375,
0.88330078125,
0.317626953125,
0.51123046875,
0.31640625
],
[
0.118408203125,
0.319580078125,
0.489501953125,
0.319580078125,
0.489501953125,
0.333251953125,
0.118408203125,
0.333251953125
],
[
0.51318359375,
0.319580078125,
0.80419921875,
0.319580078125,
0.80419921875,
0.333251953125,
0.51318359375,
0.333251953125
],
[
0.1234130859375,
0.342529296875,
0.486083984375,
0.342529296875,
0.486083984375,
0.355712890625,
0.1234130859375,
0.355712890625
],
[
0.52880859375,
0.34326171875,
0.884765625,
0.344482421875,
0.884765625,
0.359375,
0.52880859375,
0.358154296875
],
[
0.12005615234375,
0.359375,
0.48779296875,
0.359375,
0.48779296875,
0.372314453125,
0.12005615234375,
0.372314453125
],
[
0.54833984375,
0.361083984375,
0.884765625,
0.35986328125,
0.884765625,
0.374755859375,
0.54833984375,
0.3759765625
],
[
0.12176513671875,
0.377685546875,
0.486083984375,
0.377685546875,
0.486083984375,
0.388427734375,
0.12176513671875,
0.388427734375
],
[
0.5517578125,
0.377685546875,
0.88330078125,
0.377685546875,
0.88330078125,
0.391357421875,
0.5517578125,
0.391357421875
],
[
0.12005615234375,
0.392578125,
0.48779296875,
0.392578125,
0.48779296875,
0.40625,
0.12005615234375,
0.40625
],
[
0.5517578125,
0.393798828125,
0.884765625,
0.393798828125,
0.884765625,
0.40673828125,
0.5517578125,
0.40673828125
],
[
0.1234130859375,
0.407958984375,
0.48779296875,
0.407958984375,
0.48779296875,
0.4228515625,
0.1234130859375,
0.4228515625
],
[
0.5498046875,
0.4091796875,
0.88671875,
0.407958984375,
0.88671875,
0.4228515625,
0.5498046875,
0.42333984375
],
[
0.12005615234375,
0.424560546875,
0.344970703125,
0.424560546875,
0.344970703125,
0.438232421875,
0.12005615234375,
0.438232421875
],
[
0.5498046875,
0.42333984375,
0.88671875,
0.424560546875,
0.88671875,
0.439453125,
0.5498046875,
0.438232421875
],
[
0.2301025390625,
0.4453125,
0.3779296875,
0.44775390625,
0.3779296875,
0.466064453125,
0.2301025390625,
0.46435546875
],
[
0.5517578125,
0.441162109375,
0.884765625,
0.441162109375,
0.884765625,
0.454833984375,
0.5517578125,
0.454833984375
],
[
0.4609375,
0.44775390625,
0.490234375,
0.44775390625,
0.490234375,
0.465576171875,
0.4609375,
0.465576171875
],
[
0.5498046875,
0.4560546875,
0.884765625,
0.457763671875,
0.884765625,
0.47265625,
0.5498046875,
0.47021484375
],
[
0.2392578125,
0.46728515625,
0.365234375,
0.4697265625,
0.365234375,
0.48388671875,
0.2392578125,
0.482177734375
],
[
0.462646484375,
0.468505859375,
0.48779296875,
0.468505859375,
0.48779296875,
0.482666015625,
0.462646484375,
0.482666015625
],
[
0.55322265625,
0.474365234375,
0.88330078125,
0.474365234375,
0.88330078125,
0.488037109375,
0.55322265625,
0.488037109375
],
[
0.12005615234375,
0.493408203125,
0.484375,
0.493408203125,
0.484375,
0.50830078125,
0.12005615234375,
0.50830078125
],
[
0.5498046875,
0.488037109375,
0.74462890625,
0.4892578125,
0.74462890625,
0.50439453125,
0.5498046875,
0.5029296875
],
[
0.12005615234375,
0.51025390625,
0.484375,
0.51025390625,
0.484375,
0.52392578125,
0.12005615234375,
0.52392578125
],
[
0.53076171875,
0.51416015625,
0.884765625,
0.515625,
0.884765625,
0.52978515625,
0.53076171875,
0.529296875
],
[
0.12176513671875,
0.525390625,
0.48779296875,
0.525390625,
0.48779296875,
0.5390625,
0.12176513671875,
0.5390625
],
[
0.5498046875,
0.533203125,
0.88330078125,
0.533203125,
0.88330078125,
0.54638671875,
0.5498046875,
0.54638671875
],
[
0.12005615234375,
0.5419921875,
0.48779296875,
0.5419921875,
0.48779296875,
0.5556640625,
0.12005615234375,
0.5556640625
],
[
0.5498046875,
0.54638671875,
0.88330078125,
0.54736328125,
0.88330078125,
0.5625,
0.5498046875,
0.56103515625
],
[
0.1234130859375,
0.56005859375,
0.486083984375,
0.56005859375,
0.486083984375,
0.57080078125,
0.1234130859375,
0.57080078125
],
[
0.5498046875,
0.56396484375,
0.88330078125,
0.56396484375,
0.88330078125,
0.57763671875,
0.5498046875,
0.57763671875
],
[
0.12176513671875,
0.57470703125,
0.489501953125,
0.57470703125,
0.489501953125,
0.58837890625,
0.12176513671875,
0.58837890625
],
[
0.54833984375,
0.580078125,
0.67919921875,
0.58056640625,
0.67919921875,
0.595703125,
0.54833984375,
0.59423828125
],
[
0.118408203125,
0.5888671875,
0.484375,
0.5888671875,
0.484375,
0.6025390625,
0.118408203125,
0.6025390625
],
[
0.12005615234375,
0.60693359375,
0.14111328125,
0.60693359375,
0.14111328125,
0.62060546875,
0.12005615234375,
0.62060546875
],
[
0.5322265625,
0.60498046875,
0.88330078125,
0.60546875,
0.88330078125,
0.62060546875,
0.5322265625,
0.619140625
],
[
0.5498046875,
0.62255859375,
0.88330078125,
0.62255859375,
0.88330078125,
0.6357421875,
0.5498046875,
0.6357421875
],
[
0.118408203125,
0.6328125,
0.151123046875,
0.6328125,
0.151123046875,
0.6455078125,
0.118408203125,
0.6455078125
],
[
0.15283203125,
0.6328125,
0.267822265625,
0.6328125,
0.267822265625,
0.64794921875,
0.15283203125,
0.64794921875
],
[
0.54833984375,
0.63720703125,
0.884765625,
0.63818359375,
0.884765625,
0.65283203125,
0.54833984375,
0.65185546875
],
[
0.162841796875,
0.65673828125,
0.328369140625,
0.65673828125,
0.328369140625,
0.67138671875,
0.162841796875,
0.67138671875
],
[
0.54833984375,
0.65380859375,
0.81689453125,
0.65478515625,
0.81689453125,
0.66943359375,
0.54833984375,
0.66845703125
],
[
0.130126953125,
0.677734375,
0.489501953125,
0.677734375,
0.489501953125,
0.69140625,
0.130126953125,
0.69140625
],
[
0.12005615234375,
0.65771484375,
0.1695556640625,
0.65771484375,
0.1695556640625,
0.66943359375,
0.12005615234375,
0.66943359375
],
[
0.5205078125,
0.6796875,
0.88330078125,
0.6796875,
0.88330078125,
0.69482421875,
0.5205078125,
0.69482421875
],
[
0.12005615234375,
0.69482421875,
0.489501953125,
0.69482421875,
0.489501953125,
0.7080078125,
0.12005615234375,
0.7080078125
],
[
0.5146484375,
0.6953125,
0.88330078125,
0.6953125,
0.88330078125,
0.71044921875,
0.5146484375,
0.71044921875
],
[
0.12005615234375,
0.71044921875,
0.489501953125,
0.71044921875,
0.489501953125,
0.7236328125,
0.12005615234375,
0.7236328125
],
[
0.515625,
0.7119140625,
0.8818359375,
0.7119140625,
0.8818359375,
0.7255859375,
0.515625,
0.7255859375
],
[
0.118408203125,
0.72705078125,
0.490234375,
0.72705078125,
0.490234375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.515625,
0.728515625,
0.8818359375,
0.728515625,
0.8818359375,
0.7421875,
0.515625,
0.7421875
],
[
0.12005615234375,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.12005615234375,
0.755859375
],
[
0.515625,
0.74462890625,
0.8818359375,
0.74462890625,
0.8818359375,
0.7587890625,
0.515625,
0.7587890625
],
[
0.12005615234375,
0.75732421875,
0.48779296875,
0.75830078125,
0.48779296875,
0.7724609375,
0.12005615234375,
0.771484375
],
[
0.51123046875,
0.75830078125,
0.8818359375,
0.7587890625,
0.8818359375,
0.77685546875,
0.51123046875,
0.775390625
],
[
0.12176513671875,
0.77490234375,
0.48779296875,
0.77490234375,
0.48779296875,
0.7880859375,
0.12176513671875,
0.7880859375
],
[
0.515625,
0.77685546875,
0.8818359375,
0.77685546875,
0.8818359375,
0.79052734375,
0.515625,
0.79052734375
],
[
0.118408203125,
0.7890625,
0.486083984375,
0.7890625,
0.486083984375,
0.802734375,
0.118408203125,
0.802734375
],
[
0.5146484375,
0.7919921875,
0.88330078125,
0.7919921875,
0.88330078125,
0.8056640625,
0.5146484375,
0.8056640625
],
[
0.12005615234375,
0.80712890625,
0.489501953125,
0.80712890625,
0.489501953125,
0.82080078125,
0.12005615234375,
0.82080078125
],
[
0.515625,
0.80859375,
0.8818359375,
0.80859375,
0.8818359375,
0.822265625,
0.515625,
0.822265625
],
[
0.12005615234375,
0.822265625,
0.48779296875,
0.822265625,
0.48779296875,
0.8359375,
0.12005615234375,
0.8359375
],
[
0.51318359375,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.12005615234375,
0.8388671875,
0.490234375,
0.8388671875,
0.490234375,
0.8525390625,
0.12005615234375,
0.8525390625
],
[
0.5146484375,
0.84130859375,
0.8818359375,
0.84130859375,
0.8818359375,
0.85498046875,
0.5146484375,
0.85498046875
],
[
0.12005615234375,
0.85498046875,
0.414794921875,
0.85498046875,
0.414794921875,
0.869140625,
0.12005615234375,
0.869140625
],
[
0.51904296875,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51904296875,
0.87060546875
],
[
0.124267578125,
0.87890625,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.124267578125,
0.8935546875
],
[
0.515625,
0.87353515625,
0.88330078125,
0.87353515625,
0.88330078125,
0.88720703125,
0.515625,
0.88720703125
],
[
0.1175537109375,
0.89453125,
0.337646484375,
0.8955078125,
0.337646484375,
0.91015625,
0.1175537109375,
0.9091796875
],
[
0.33154296875,
0.8955078125,
0.486083984375,
0.8955078125,
0.486083984375,
0.9091796875,
0.33154296875,
0.9091796875
],
[
0.5341796875,
0.8955078125,
0.8212890625,
0.8955078125,
0.8212890625,
0.91015625,
0.5341796875,
0.91015625
]
] |
[
"Original Causal Statement Form",
"Converted Causal Statement Form",
"Billy finds his childhood teddy bear,",
"Resource",
"finds",
"his",
"childhood",
"teddy",
"bear",
"> Cause/Enable > Billy gives his childhood",
"Billy",
"GLU-SPE",
"billy gives his childhood teddy bear to",
"teddy bear to his daughter",
"his daughter.",
"Someone_A",
"finds",
"Something_A",
">",
"Someone_A finds Something_A, Some-",
"GLU-GEN",
"Cause/Enable > Someone_A gives Something_A",
"one_A gives Something_A to Some-",
"to Someone_B",
"one_ B.",
"PersonXfollows PersonYintoroom>oWant>to",
"PersonXfollows PersonY into room, to",
"ATOMIC",
"know why PersonX is following them",
"know why PersonX is following them.",
"Fisk was shot to death by his mistress's new lover",
"Fisk was shot to death by his mistress's",
"DISTANT",
"and Fisk's ex-business partner.",
"new lover and Fisk's ex-business partner.",
"Table 1: The original and converted form (the input form of SelfRL) of different causal statements from",
"three resources. GLU-SPE and GLU-GEN denote the specific and general statements from GLUCOSE",
"respectively.",
"其中，P+和P分别是一个批次中的因果事件",
"获取四种类型的外部因果语句。表1说明了",
"对和所有事件对，T是调节浓度水平的温度，",
"SelfRL的原始形式和转换后的输人形式（Sec。",
"D是l2-距离函数，用于测量两个表示的距离",
"3.1）来自三个不同资源的因果陈述。",
"LearningofEventCausalityIdentifier对于",
": GLUCOSE (Mostafazadeh et al., 2020) :",
"事件因果关系标识符的训练，我们将对比损失",
"a large-scale dataset of implicit common-",
"添加到基本分类损失中，这可以指导标识符从",
"sense knowledge, encoded as causal ex-",
"SelfRL学习增强因果表示中隐含的特定于上",
"planatory mini-theories inspired by cogni-",
"下文的因果模式。如算法1所示，我们最小化",
"tive psychology. Each GLUCOSE expla-",
"L>，随机梯度更新入如下：",
"nation is stated both as a specific state-",
"Cx = Cstu + Ccon,",
"ment (grounded in a given context, GLU-",
"(8)",
"SPE in Table 1）and a corresponding",
"nstuV",
"(9)",
"general rule (applicable to other contexts,",
"其中，nstu是标识符的学习率。为了进行评估",
"GLU-GEN in Table 1 ).",
"我们在没有对比传递模块的情况下预测输入事",
". ATOMIC (Sap et al., 2018) : an atlas of",
"件对的因果关系。此外，Lcon中的T间接起",
"machine commonsense, as a step toward",
"到调整Cstu和Ccon的影响权重的作用。具体",
"addressing the rich spectrum of inferen-",
"来说，对于教学，我们以在线网络的学习过的",
"tial knowledge that is crucial for common-",
"Ence作为编码器，冻结其参数，以提供外部",
"sense reasoning.",
"因果语句的增强因果表示，以进行对比表示传",
"递",
". DISTANT (Zuo et al., 2020) : the auto-",
"matically labeled training data for ECI",
"4",
"Experiments",
"via distant supervision that expresses the",
"Experimental Setup",
"causal semantics between events.",
"Dataset and Evaluation Metrics for ECI",
"4.1",
"ParametersSettings在实现中，所有BERT",
"我们的实验在两个主要基准上进行，包",
"模块都在BERT-Base体系结构2上实现，该",
"括: EventStoryLinev0。9 (ESC) (Caselli and",
"体系结构具有12层，768层和12头。我们采用",
"Vossen，2017）上述；和(2)因果时间银行（CTB）",
"单层BiLSTM(Hochreiter and Schmidhuber,",
"(MirzaandTonelli，2014），其中包含184个",
"1997）作为Ence和Encs。对于参数，我们",
"文档，6813个事件和318个因果事件对。与以",
"将SelfRL（ntea）和标识符（nstu）的学习",
"前的方法相同，我们使用ESC的最后两个主",
"率分别设置为1e-5和2e-5。对比传输模块和",
"题作为两个数据集的开发集。对于评估，我们",
"BiLSTM的隐藏层中的空间大小均设置为50。",
"采用Precision(P)、Recall(R)和F1-score(F1)",
"并且我们分别设置SelfRL中移动平均线的衰",
"作为评估指标。我们分别对ESC和CTB进",
"减率T和对比损失Lcon的温度为0。996和",
"行5倍和10倍交叉验证，与以前的方法相同。",
"0。1在开发集上调谐。此外，我们还在开发集",
"所有结果均为三次独立实验的平均值",
"上将SelfRL和标识符的批处理大小分别调整",
"Data Preparation for Self-Supervised Causal",
"为48和16。我们应用早期停止和AdamW梯",
"Representation Learning",
"我们从三种资源中",
"2https://github.com/google-research/bert"
] |
{
"arxiv_id": "2106.01654",
"language": "zh",
"page_idx": 4
}
|
||
2303.01112-zh-2
|
2303.01112-zh-2.jpg
|
modifications made to the original architecture. 他们研究的主要发现之一是，当 ViT 在较大的数据集上进行预训练时，例如 ImageNet-21k 和 JFT-300M, 下游任务的准确性随数据集的大小 and models 。还努力将 SSL 用于预训练视觉转换器，例如。g。DINO [9] ,BEiT [5] ,MAE [14] 。但是，ImageNet-1k 微调的最高精度仍然是通过使用 JFT-3B$\underline{\mathrm{E}}[43]$进行有监督的预训练的视觉转换器获得的
最大的数据集，例如 JFT-300M/3B/4BE
[29]
和
[34, 43] , Instagram-3。5B
[35]
$\operatorname{\dot{Y}FCC}(\bar{\text{Yahoo!FlickrCreativeCommons}})$
都是预训练视觉转换器的优秀候选者。然而，对这些数据集的访问仅限于选定的一组研究人员。为了解决此问题，LAION-400M/5B 数据集[32,33]已根据允许商业用途的许可许可公开。LAION -5B 包含 5。850亿个 CLIP 过滤的图像-文本对，可用于训练大型视觉+语言模型。然而，由于 LAION 数据集也不是精选的，因此它的使用也受到与其他大型数据集相关的相同关注 (i。e。,社会偏见、隐私和版权问题的存在)[7]。与大型数据集相关的这些道德问题是一个严重的问题。For example, one of the most widely used datasets in computer vision - ImageNet [11], contains images that violate privacy protection laws and biased samples, can lead to unfairness in models trained on the dataset [40]. 此外，由于数据集中的歧视性标签， 诸如 80MTinyImages$\underline{\mathbb{F}}[1]$之类的数据集的公开发布实际上已被暂停$\mathbb{E}[6]$
SSL 可以避免与歧视性标签相关的问题，还可以使我们摆脱对这些庞大数据集进行标记的繁琐任务。但是，SSL 尚未达到与 SL 相同的精度水平。Moreover, SSL still relies on natural images and shares the limitations of large datasets e.g. societal bias, privacy and copyright.
2.2. Formula-Driven Supervised Learning
已经提出了一种用于预训练视觉转换器的替代方法，其中从数学公式生成的合成图像用于执行公式驱动的监督学习 (FDSL) $\operatorname{E}[21,30]$ 。已经发表了一些关于 FDSL 的后续研究，这些研究扩展了这种方法的应用，并提高了合成数据集的质量$\mathbb{E}[3,16,19,20,39]$ FDSL 一直在迅速发展，并且图像质量已经提高到现在超过了使用 ImageNet-21k$\underline{\mathrm{F}}[19]$可以实现的预训练水平。此外，FDSL 具有相当大的潜力，因为它产生了可以不断改进的动态数据集，这与 JFT 和 LAION 等静态数据集不同，后者只能增加数据集的数量。FDSL 的另一个优点是它完全没有与自然图像相关的常见道德问题，例如社会偏见，隐私和版权侵权。FDSL 也不需要从主机网站下载大量图像，因为具有公式的任何人都可以在本地生成相同的合成图像。这些优势使采取当前可以匹配 ImageNet-21k 的 FDSL 并将其质量提高到可以匹配 JFT-300M 的程度成为值得的努
$^{1}$The full version has not yet been released, but a copy of
YFCC-15M is available.
力。However, previous FDSL methods have several outstanding issues: 1) difficult to control the variety of the contour outline of the synthetic images, 2) lack of smoothness of the contour, 3) unclear why and how to generate better synthetic images for pre-training. $\bar{\mathrm{Therefore,~an~exhaustive~investigation~regarding~the}}$ shape of contours in synthetic images is required.
## 3. Method
受 deBroglie 在量子物理学领域的原子模型的概念启发，我们提出了 visual atomic renderer ,可用于生成合成图像，以学习计算机视觉任务中的视觉表示。基于可以从平滑波函数的组合中获得预训练所需的各种对象形状的想法，我们在下面分两步介绍我们的 method 。首先，我们定义了可视化原子 renderer ,即 produces 二维可视化原子。其次，我们提出了一种使用 VisualAtom, which is synthetic image dataset generated by our proposed the visual atomic renderen 的预训练方法。The visual atomic renderer can assign many variations to the contour outlines and control these variations by changing the parameter ranges.
## 3.1. Visual Atomic Renderer
可视化原子 renderer $\mathcal{V}$将原子$\mathcal{A}$可视化为图像。更具体地说，视觉原子 renderer produces and renders a visual atom $\mathcal{V}(\mathcal{A})\subset\mathbb{R}^2$ from an atom $\mathcal{A}$.在这里，我们通过参数化原子的每个分量来定义$\nu$
Parameterized orbits. 我们在 2D 欧几里得 space 中通
过具有两个参数$(a_k,b_k)\in\mathbb{R}^2$的椭圆定义轨道$O_k:$
$$O_k=\{(x,y)^\top:F_k(x,y)=0\}\subset\mathbb{R}^2,$$
(1)
其中$F_k$是由
$$F_k(x,y)=\frac{x^2}{a_k^2}+\frac{y^2}{b_k^2}-1.$$
(2)
给出的椭圆的隐式函数。。为了减少制作轨道序列时的参数数量，我们引人了间隔参数$c$ ,并在给定随机采样的$(a_1,b_1)$的情况下将参数值递归确定为$(a_{k+1},b_{k+1})=(a_{k}+c,b_{k}+c)$
Parameterized waves.波函数$\Phi_k$的最简单定义可以由
单个正弦波给出：
(3)
$$\Phi_k(\theta)=\lambda\sin(n\theta),$$
其中，$n\in\mathbb{N}$是频率参数，$\lambda\in\mathbb{R}$是振幅参数。请注意，
$n$需要是自然数才能满足条件$\Phi_k(\theta)=\Phi_k(\theta+2\pi).$
为了具有足够多的波，视觉原子 renderer 假设每个轨道具有两个正弦波的混合和一个小噪声，如下所示：
$$\Phi_k(\theta)=\lambda_1\sin(n_1\theta)+\lambda_2\sin(n_2\theta)+\eta\epsilon(\theta),$$
其中$,n_1,n_2\in\mathbb{N}$ 是频率参数$,\lambda_1,\lambda_2\in\mathbb{R}$ 是振幅参数$\eta\in\mathbb{R}$是噪声级参数 XMATHX4。波函数(w/o 噪声)
|
[
[
0.08416748046875,
0.0947265625,
0.468017578125,
0.0947265625,
0.468017578125,
0.10479736328125,
0.08416748046875,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.106689453125,
0.50244140625,
0.106689453125
],
[
0.08251953125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08251953125,
0.121826171875
],
[
0.50341796875,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.08251953125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08251953125,
0.1351318359375
],
[
0.5009765625,
0.12371826171875,
0.8896484375,
0.12371826171875,
0.8896484375,
0.136962890625,
0.5009765625,
0.136962890625
],
[
0.08087158203125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50244140625,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.50244140625,
0.1514892578125
],
[
0.080078125,
0.15283203125,
0.46728515625,
0.1514892578125,
0.46728515625,
0.1666259765625,
0.080078125,
0.1673583984375
],
[
0.50244140625,
0.154052734375,
0.88818359375,
0.154052734375,
0.88818359375,
0.1673583984375,
0.50244140625,
0.1673583984375
],
[
0.080078125,
0.1666259765625,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.080078125,
0.18115234375
],
[
0.51318359375,
0.1685791015625,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.51318359375,
0.1817626953125
],
[
0.08251953125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.08251953125,
0.1962890625
],
[
0.50244140625,
0.18310546875,
0.85546875,
0.18310546875,
0.85546875,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.08087158203125,
0.1988525390625,
0.2607421875,
0.1988525390625,
0.2607421875,
0.212158203125,
0.08087158203125,
0.212158203125
],
[
0.09722900390625,
0.21337890625,
0.468017578125,
0.212158203125,
0.468017578125,
0.2266845703125,
0.09722900390625,
0.2279052734375
],
[
0.5009765625,
0.2108154296875,
0.6005859375,
0.2108154296875,
0.6005859375,
0.2259521484375,
0.5009765625,
0.2259521484375
],
[
0.176513671875,
0.2291259765625,
0.3203125,
0.2291259765625,
0.3203125,
0.242431640625,
0.176513671875,
0.242431640625
],
[
0.3857421875,
0.2279052734375,
0.4697265625,
0.2279052734375,
0.4697265625,
0.24560546875,
0.3857421875,
0.24560546875
],
[
0.521484375,
0.23486328125,
0.890625,
0.236083984375,
0.890625,
0.250732421875,
0.521484375,
0.25
],
[
0.08087158203125,
0.24365234375,
0.361083984375,
0.24365234375,
0.361083984375,
0.2568359375,
0.08087158203125,
0.2568359375
],
[
0.08251953125,
0.2298583984375,
0.1502685546875,
0.2298583984375,
0.1502685546875,
0.2412109375,
0.08251953125,
0.2412109375
],
[
0.439453125,
0.2442626953125,
0.46728515625,
0.2442626953125,
0.46728515625,
0.2568359375,
0.439453125,
0.2568359375
],
[
0.50341796875,
0.251953125,
0.890625,
0.251953125,
0.890625,
0.26513671875,
0.50341796875,
0.26513671875
],
[
0.08087158203125,
0.25830078125,
0.468017578125,
0.25830078125,
0.468017578125,
0.272705078125,
0.08087158203125,
0.272705078125
],
[
0.50341796875,
0.267578125,
0.89208984375,
0.267578125,
0.89208984375,
0.281005859375,
0.50341796875,
0.281005859375
],
[
0.08251953125,
0.27587890625,
0.46728515625,
0.27587890625,
0.46728515625,
0.285888671875,
0.08251953125,
0.285888671875
],
[
0.5048828125,
0.28271484375,
0.890625,
0.28271484375,
0.890625,
0.296630859375,
0.5048828125,
0.296630859375
],
[
0.08087158203125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.5048828125,
0.29736328125,
0.890625,
0.29736328125,
0.890625,
0.311279296875,
0.5048828125,
0.311279296875
],
[
0.08087158203125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.08087158203125,
0.317626953125
],
[
0.50244140625,
0.311279296875,
0.890625,
0.3125,
0.890625,
0.326904296875,
0.50244140625,
0.32568359375
],
[
0.08087158203125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.08087158203125,
0.333251953125
],
[
0.5107421875,
0.33203125,
0.52294921875,
0.33203125,
0.52294921875,
0.33837890625,
0.5107421875,
0.33837890625
],
[
0.51708984375,
0.32763671875,
0.88232421875,
0.32763671875,
0.88232421875,
0.341552734375,
0.51708984375,
0.341552734375
],
[
0.08087158203125,
0.333984375,
0.468017578125,
0.333251953125,
0.468017578125,
0.347900390625,
0.08087158203125,
0.348388671875
],
[
0.50341796875,
0.341552734375,
0.890625,
0.34228515625,
0.890625,
0.356689453125,
0.50341796875,
0.35595703125
],
[
0.08251953125,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.08251953125,
0.363037109375
],
[
0.5048828125,
0.35791015625,
0.88818359375,
0.35791015625,
0.88818359375,
0.371826171875,
0.5048828125,
0.371826171875
],
[
0.08087158203125,
0.365478515625,
0.46728515625,
0.365478515625,
0.46728515625,
0.37890625,
0.08087158203125,
0.37890625
],
[
0.50244140625,
0.371826171875,
0.8896484375,
0.37255859375,
0.8896484375,
0.386962890625,
0.50244140625,
0.386474609375
],
[
0.08087158203125,
0.380126953125,
0.46728515625,
0.380126953125,
0.46728515625,
0.393310546875,
0.08087158203125,
0.393310546875
],
[
0.5048828125,
0.38818359375,
0.890625,
0.38818359375,
0.890625,
0.401611328125,
0.5048828125,
0.401611328125
],
[
0.08251953125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.4091796875,
0.08251953125,
0.4091796875
],
[
0.5009765625,
0.401611328125,
0.865234375,
0.40283203125,
0.865234375,
0.417236328125,
0.5009765625,
0.416748046875
],
[
0.08087158203125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.08087158203125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.439453125,
0.08087158203125,
0.439453125
],
[
0.50244140625,
0.4248046875,
0.7294921875,
0.425537109375,
0.7294921875,
0.43994140625,
0.50244140625,
0.439453125
],
[
0.08087158203125,
0.43994140625,
0.46728515625,
0.43994140625,
0.46728515625,
0.453857421875,
0.08087158203125,
0.453857421875
],
[
0.52294921875,
0.448974609375,
0.890625,
0.449462890625,
0.890625,
0.464111328125,
0.52294921875,
0.46337890625
],
[
0.08087158203125,
0.455810546875,
0.468017578125,
0.455810546875,
0.468017578125,
0.4697265625,
0.08087158203125,
0.4697265625
],
[
0.50244140625,
0.464111328125,
0.890625,
0.46533203125,
0.890625,
0.479248046875,
0.50244140625,
0.477783203125
],
[
0.08087158203125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.484130859375,
0.08087158203125,
0.484130859375
],
[
0.08087158203125,
0.48486328125,
0.2132568359375,
0.48486328125,
0.2132568359375,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.50341796875,
0.479248046875,
0.89208984375,
0.479736328125,
0.89208984375,
0.494384765625,
0.50341796875,
0.49365234375
],
[
0.098876953125,
0.50048828125,
0.46728515625,
0.50048828125,
0.46728515625,
0.5146484375,
0.098876953125,
0.5146484375
],
[
0.5048828125,
0.49560546875,
0.78369140625,
0.49560546875,
0.78369140625,
0.5087890625,
0.5048828125,
0.5087890625
],
[
0.50244140625,
0.51025390625,
0.890625,
0.5087890625,
0.890625,
0.5234375,
0.50244140625,
0.5244140625
],
[
0.08087158203125,
0.5166015625,
0.46728515625,
0.5166015625,
0.46728515625,
0.52978515625,
0.08087158203125,
0.52978515625
],
[
0.08087158203125,
0.53076171875,
0.46728515625,
0.53076171875,
0.46728515625,
0.54443359375,
0.08087158203125,
0.54443359375
],
[
0.50341796875,
0.52587890625,
0.87353515625,
0.52587890625,
0.87353515625,
0.5390625,
0.50341796875,
0.5390625
],
[
0.07843017578125,
0.54443359375,
0.46728515625,
0.54541015625,
0.46728515625,
0.56005859375,
0.07843017578125,
0.5595703125
],
[
0.5703125,
0.5517578125,
0.82666015625,
0.5517578125,
0.82666015625,
0.56494140625,
0.5703125,
0.56494140625
],
[
0.86669921875,
0.55322265625,
0.890625,
0.55322265625,
0.890625,
0.56640625,
0.86669921875,
0.56640625
],
[
0.07843017578125,
0.56005859375,
0.468017578125,
0.56103515625,
0.468017578125,
0.57568359375,
0.07843017578125,
0.57470703125
],
[
0.08087158203125,
0.5791015625,
0.1519775390625,
0.5791015625,
0.1519775390625,
0.5888671875,
0.08087158203125,
0.5888671875
],
[
0.50244140625,
0.5771484375,
0.5986328125,
0.5791015625,
0.5986328125,
0.59326171875,
0.50244140625,
0.59130859375
],
[
0.080078125,
0.60107421875,
0.408447265625,
0.6015625,
0.408447265625,
0.6162109375,
0.080078125,
0.61572265625
],
[
0.689453125,
0.60107421875,
0.765625,
0.60107421875,
0.765625,
0.61865234375,
0.689453125,
0.61865234375
],
[
0.61279296875,
0.6123046875,
0.68701171875,
0.6123046875,
0.68701171875,
0.626953125,
0.61279296875,
0.626953125
],
[
0.68115234375,
0.6162109375,
0.78515625,
0.60498046875,
0.7900390625,
0.630859375,
0.68701171875,
0.642578125
],
[
0.86669921875,
0.611328125,
0.8935546875,
0.611328125,
0.8935546875,
0.62744140625,
0.86669921875,
0.62744140625
],
[
0.10455322265625,
0.626953125,
0.465576171875,
0.626953125,
0.465576171875,
0.63720703125,
0.10455322265625,
0.63720703125
],
[
0.67724609375,
0.62353515625,
0.689453125,
0.62353515625,
0.689453125,
0.62744140625,
0.67724609375,
0.62744140625
],
[
0.08087158203125,
0.64013671875,
0.468017578125,
0.64013671875,
0.468017578125,
0.65380859375,
0.08087158203125,
0.65380859375
],
[
0.50341796875,
0.64599609375,
0.8896484375,
0.64599609375,
0.8896484375,
0.6591796875,
0.50341796875,
0.6591796875
],
[
0.08087158203125,
0.65380859375,
0.46728515625,
0.65380859375,
0.46728515625,
0.66748046875,
0.08087158203125,
0.66748046875
],
[
0.50341796875,
0.6611328125,
0.890625,
0.6611328125,
0.890625,
0.6748046875,
0.50341796875,
0.6748046875
],
[
0.08251953125,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.6845703125,
0.08251953125,
0.6845703125
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.68994140625,
0.50341796875,
0.68994140625
],
[
0.08416748046875,
0.6875,
0.465576171875,
0.6875,
0.465576171875,
0.69775390625,
0.08416748046875,
0.69775390625
],
[
0.5048828125,
0.6923828125,
0.7158203125,
0.6923828125,
0.7158203125,
0.70556640625,
0.5048828125,
0.70556640625
],
[
0.08087158203125,
0.70068359375,
0.468017578125,
0.70068359375,
0.468017578125,
0.71533203125,
0.08087158203125,
0.71533203125
],
[
0.50341796875,
0.70703125,
0.8896484375,
0.70703125,
0.8896484375,
0.72021484375,
0.50341796875,
0.72021484375
],
[
0.08087158203125,
0.71533203125,
0.46728515625,
0.71533203125,
0.46728515625,
0.72900390625,
0.08087158203125,
0.72900390625
],
[
0.50341796875,
0.7216796875,
0.62841796875,
0.7216796875,
0.62841796875,
0.73486328125,
0.50341796875,
0.73486328125
],
[
0.08087158203125,
0.73095703125,
0.46728515625,
0.73095703125,
0.46728515625,
0.744140625,
0.08087158203125,
0.744140625
],
[
0.08087158203125,
0.74560546875,
0.468017578125,
0.744140625,
0.468017578125,
0.7587890625,
0.08087158203125,
0.76025390625
],
[
0.63232421875,
0.74755859375,
0.76318359375,
0.74755859375,
0.76318359375,
0.76123046875,
0.63232421875,
0.76123046875
],
[
0.86767578125,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76318359375,
0.86767578125,
0.76318359375
],
[
0.08087158203125,
0.76123046875,
0.46728515625,
0.76025390625,
0.46728515625,
0.7734375,
0.08087158203125,
0.7744140625
],
[
0.08087158203125,
0.77587890625,
0.46728515625,
0.77587890625,
0.46728515625,
0.7890625,
0.08087158203125,
0.7890625
],
[
0.50341796875,
0.7744140625,
0.89208984375,
0.7744140625,
0.89208984375,
0.78857421875,
0.50341796875,
0.78857421875
],
[
0.08251953125,
0.79150390625,
0.468017578125,
0.79150390625,
0.468017578125,
0.8046875,
0.08251953125,
0.8046875
],
[
0.50244140625,
0.7890625,
0.865234375,
0.7890625,
0.865234375,
0.80322265625,
0.50244140625,
0.80322265625
],
[
0.08251953125,
0.80810546875,
0.46728515625,
0.80810546875,
0.46728515625,
0.818359375,
0.08251953125,
0.818359375
],
[
0.52392578125,
0.8037109375,
0.8896484375,
0.8037109375,
0.8896484375,
0.8173828125,
0.52392578125,
0.8173828125
],
[
0.08087158203125,
0.8212890625,
0.468017578125,
0.8212890625,
0.468017578125,
0.83544921875,
0.08087158203125,
0.83544921875
],
[
0.5048828125,
0.8212890625,
0.8857421875,
0.8212890625,
0.8857421875,
0.83203125,
0.5048828125,
0.83203125
],
[
0.08251953125,
0.8359375,
0.46728515625,
0.8359375,
0.46728515625,
0.849609375,
0.08251953125,
0.849609375
],
[
0.54736328125,
0.8447265625,
0.890625,
0.8466796875,
0.890625,
0.861328125,
0.54736328125,
0.859375
],
[
0.08251953125,
0.8515625,
0.46728515625,
0.8515625,
0.46728515625,
0.86572265625,
0.08251953125,
0.86572265625
],
[
0.101318359375,
0.87548828125,
0.4697265625,
0.875,
0.4697265625,
0.88818359375,
0.101318359375,
0.8896484375
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.5048828125,
0.88623046875
],
[
0.08087158203125,
0.88623046875,
0.2279052734375,
0.88818359375,
0.2279052734375,
0.9013671875,
0.08087158203125,
0.8994140625
],
[
0.5048828125,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.9013671875,
0.5048828125,
0.9013671875
]
] |
[
"modificationsmadetotheoriginal architecture.他",
"力。However, previous FDSL methods have several",
"们研究的主要发现之一是，当ViT在较大的数据集上",
"outstanding issues: 1) difficult to control the variety",
"进行预训练时，例如ImageNet-21k和JFT-300M，下",
"of the contour outline of the synthetic images, 2)lack",
"游任务的准确性随数据集的大小andmodels。还努力",
"of smoothness of the contour, 3) unclear why and how",
"将SSL用于预训练视觉转换器，例如。g。DINO[9]",
"to generate better synthetic images for pre-training",
",BEiT[5],MAE【14]。但是，ImageNet-1k微调的最",
"Therefore, an exhaustive investigation regarding the",
"高精度仍然是通过使用JFT-3BE[43]进行有监督的预",
"shape of contours in synthetic images is required.",
"训练的视觉转换器获得的",
"最大的数据集，例如JFT-300M/3B/4BE",
"3. Method",
"，Instagram-3。5B",
"[29]”和",
"受deBroglie在量子物理学领域的原子模型的概念",
"YFCC(Yahoo!FlickrCreativeCommons)",
"[34,43]",
"[35]",
"启发，我们提出了visualatomicrenderer，可用于",
"都是预训练视觉转换器的优秀候选者。然而，对这些",
"生成合成图像，以学习计算机视觉任务中的视觉表",
"数据集的访问仅限于选定的一组研究人员。为了解决",
"示。基于可以从平滑波函数的组合中获得预训练所需",
"此问题，LAION-400M/5B数据集【32,33]】已根据允",
"的各种对象形状的想法，我们在下面分两步介绍我们",
"许商业用途的许可许可公开。LAION-5B包含5。850",
"的method。首先，我们定义了可视化原子renderer",
"亿个CLIP过滤的图像-文本对，可用于训练大型视",
"",
"即produces二维可视化原子。其次，我们提出了",
"觉+语言模型。然而，由于LAION数据集也不是",
"种使用 VisualAtom, which is synthetic image dataset",
"精选的，因此它的使用也受到与其他大型数据集相关",
"generated by our proposed the visual atomic renderel",
"的相同关注(i。e。，社会偏见、隐私和版权问题的存",
"的预训练方法。Thevisual atomic renderer can assign",
"在）[7。与大型数据集相关的这些道德问题是一个严",
"many variations to the contour outlines and control",
"重的问题。For example,oneofthemostwidelyused",
"these variations by changing the parameter ranges.",
"datasets in computer vision - ImageNet [ll], contains",
"images that violate privacy protection laws and biased",
"3.1. Visual Atomic Renderer",
"samples, can lead to unfairness in models trained on",
"可视化原子rendererV将原子A可视化为图像。更",
"thedataset【40]：此外，由于数据集中的歧视性标签，",
"具体地说，视觉原子rendererproducesandrendersa",
"诸如80MTinyImagesE[1]之类的数据集的公开发布实",
"际上已被暂停E[6]",
"visual atomV(A)cR2fromanatomA.在这里，我",
"SSL可以避免与歧视性标签相关的问题，还可以使",
"们通过参数化原子的每个分量来定义V",
"Parameterizedorbits.我们在2D欧几里得space中通",
"我们摆脱对这些庞大数据集进行标记的繁琐任务。但",
"是，SSL尚未达到与SL相同的精度水平。Moreover，",
"过具有两个参数（ak，bk）ER²的椭圆定义轨道Ok：",
"SSL still relies on natural images and shares the limi-",
"Ok = ((c, y) /: Fr(c,y) = O) C R?,",
"(1)",
"tations of large datasets e.g. societal bias, privacy and",
"copyright.",
"其中F是由",
"2.2. Formula-Driven Supervised Learning",
"r?y?",
"Fr(c,y) =",
"+-1.",
"(2)",
"已经提出了一种用于预训练视觉转换器的替代方",
"",
"法，其中从数学公式生成的合成图像用于执行公式驱",
"给出的椭圆的隐式函数。。为了减少制作轨道序列",
"动的监督学习（FDSL）E[21,301。已经发表了一些",
"时的参数数量，我们引人了间隔参数，并在给定",
"关于FDSL的后续研究，这些研究扩展了这种方法的",
"随机采样的（a1,b1）的情况下将参数值递归确定为",
"应用，并提高了合成数据集的质量F3.16.19.20,39",
"(ak+1,bk+1)=(ak+c,bk+)",
"FDSL一直在迅速发展，并且图像质量已经提高到现",
"Parameterizedwaves.波函数Φk的最简单定义可以由",
"在超过了使用ImageNet-21kE[19]可以实现的预训练",
"单个正弦波给出：",
"水平。此外，FDSL具有相当大的潜力，因为它产生了",
"可以不断改进的动态数据集，这与JFT和LAION等",
"Φ(0) = Asin(n0),",
"(3)",
"静态数据集不同，后者只能增加数据集的数量。FDSL",
"的另一个优点是它完全没有与自然图像相关的常见道",
"其中，nEN是频率参数，ER是振幅参数。请注意，",
"德问题，例如社会偏见，隐私和版权侵权。FDSL也",
"n需要是自然数才能满足条件Φk(①）=Φ（0+2元）",
"不需要从主机网站下载大量图像，因为具有公式的任",
"为了具有足够多的波，视觉原子renderer假设每个",
"何人都可以在本地生成相同的合成图像。这些优势使",
"轨道具有两个正弦波的混合和一个小噪声，如下所示：",
"采取当前可以匹配ImageNet-21k的FDSL并将其质",
"Φr(0)=入1sin(n10)+入2 sin(n20)+ne(0),(4)",
"量提高到可以匹配JFT-300M的程度成为值得的努",
"The full version has not yet been released, but a copy of",
"其中，n1,n2EN是频率参数，1,入2ER是振幅参数",
"YFCC-15M is available.",
"mER是噪声级参数XMATHX4。波函数(w/o噪声)"
] |
{
"arxiv_id": "2303.01112",
"language": "zh",
"page_idx": 2
}
|
||
2104.09425-zh-0
|
2104.09425-zh-0.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2022}}$
鲁棒学习满足生成模型：代理分布可以提高对抗性鲁
棒性吗？
Vikash Sehwag$^{\bullet\dagger}$, Saeed Mahloujifar$^{\bullet}$, Tinashe Handina$^{\bullet}$, Sihui Dai$^{\bullet}$, Chong Xiang$^{\bullet}$
Mung Chiang$^{\bullet}$, Prateek Mittal$^{\bullet}$
$^{\bullet}$Princeton University, $^{\bullet}$Caltech, $^{\bullet}$Purdue University
1
Abstract
While additional training data improves the robustness of deep neural networks against adversarial examples, it presents the challenge of curating a large number of specific real-world samples. We circumvent this challenge by using additional data from proxy distributions learned by advanced generative models. We frst seek to formally understand the transfer of robustness from classifiers trained on proxy distributions to the real data distribution. We prove that the difference between the robustness of a classifier on the two distributions is upper bounded by the conditional Wasserstein distance between them. Next we use proxy distributions to significantly improve the performance of adversarial training on five different datasets. For example, we improve robust accuracy by up to 7.5 % and 6.7 % in $\ell_{\infty}$ and $\ell_2$ threat model over baselines that are not using proxy distributions on the CIFAR-10 dataset. We also improve certified robust accuracy by 7.6% on the CIFAR-10 dataset. We further demonstrate that different generative models bring a disparate improvement in the performance in robust training. We propose a robust discrimination approach to characterize the impact of individual generative models and further provide a deeper understanding of why current state-of-the-art in diffusion-based generative models are a better choice for proxy distribution than generative adversarial networks.
$\operatorname{Introduction}$
深度神经网络是强大的工具，但它们的成功在很大程度上取决于训练数据量 (Sun et al. 2017; Mahajan et al., 2018)。最近的工作表明，对于提高针对对抗性示例的鲁棒性训练E (Biggio et al., 2013; Szegedy et al., 2014; Biggio & Roli, 2018) , 额外的训练数据是有帮助的$\overline{\mathrm{E}}($Carmon et al., 2019; Schmidt et al., 2018a; Uesato et al., 2019; Deng et al., 2021)。然而，从实际数据分布中整理更多的现实世界数据通常是具有挑战性和成本的 (Recht et al., 2019) 。为了规避这一挑战，我们问：是否可以使用代理分布来增强健壮的训练，即。e。, real data 分布的近似值？特别是，来自代理分布的额外样本可能更便宜，可以提高稳健性。Ifso,cangenerativemodelsthataretrainedonlimitedtrainingimagesinsmallscaledatasets (Le-cangenerativemodelsthataretrainedonlimitedtrainingimagesinsmallscaledatasets (Le-cangenerativemedonedeltaredonedtaretrainedonlimitedtrainingimagesinsmallscaledatasets ( Cun'$\&$ Cortes, 2010; Krizhevsky et al., 2014) ,actassuchaproxydistribution? $^1$
当从生成模型中训练合成样本时，一个自然的问题是合成数据的鲁棒性是否也会转移到现
实世界的数据。即使是这样，我们是否可以确定能够实现这种合成到真实的稳健性转移的
$\overset{\dagger}{\operatorname*{\operatorname*{\operatorname*{Corresponding~author:~vvikash@princeton.edu}}}}$
Our code is available at https: / / github. com/ inspire- group/ proxy- distributions .
$^{1}$Proxy distributions may not necessarily be modeled by generative models. When a proxy distribution is the output of a generative model, we call it synthetic distribution and refer to data sampled from it as synthetic data .
1
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.176513671875,
0.10162353515625,
0.82177734375,
0.10418701171875,
0.82177734375,
0.125,
0.176513671875,
0.12249755859375
],
[
0.176513671875,
0.126953125,
0.275390625,
0.126953125,
0.275390625,
0.1483154296875,
0.176513671875,
0.1483154296875
],
[
0.188720703125,
0.1749267578125,
0.7998046875,
0.173583984375,
0.7998046875,
0.1912841796875,
0.188720703125,
0.1925048828125
],
[
0.1829833984375,
0.1925048828125,
0.416748046875,
0.1925048828125,
0.416748046875,
0.20703125,
0.1829833984375,
0.20703125
],
[
0.1829833984375,
0.2203369140625,
0.56640625,
0.2215576171875,
0.56640625,
0.2392578125,
0.1829833984375,
0.238037109375
],
[
0.461669921875,
0.272705078125,
0.53759765625,
0.274658203125,
0.53759765625,
0.292236328125,
0.461669921875,
0.290283203125
],
[
0.2344970703125,
0.306884765625,
0.76123046875,
0.306884765625,
0.76123046875,
0.32080078125,
0.2344970703125,
0.32080078125
],
[
0.233642578125,
0.3193359375,
0.7646484375,
0.3193359375,
0.7646484375,
0.333984375,
0.233642578125,
0.333984375
],
[
0.2320556640625,
0.333984375,
0.765625,
0.335205078125,
0.765625,
0.349853515625,
0.2320556640625,
0.348388671875
],
[
0.233642578125,
0.349853515625,
0.76123046875,
0.349853515625,
0.76123046875,
0.363037109375,
0.233642578125,
0.363037109375
],
[
0.2344970703125,
0.3642578125,
0.76123046875,
0.3642578125,
0.76123046875,
0.375,
0.2344970703125,
0.375
],
[
0.233642578125,
0.376953125,
0.7646484375,
0.376953125,
0.7646484375,
0.39013671875,
0.233642578125,
0.39013671875
],
[
0.233642578125,
0.39013671875,
0.7646484375,
0.39013671875,
0.7646484375,
0.404052734375,
0.233642578125,
0.404052734375
],
[
0.233642578125,
0.404052734375,
0.765625,
0.404052734375,
0.765625,
0.417236328125,
0.233642578125,
0.417236328125
],
[
0.2320556640625,
0.416748046875,
0.7646484375,
0.417236328125,
0.7646484375,
0.431884765625,
0.2320556640625,
0.43115234375
],
[
0.2344970703125,
0.43115234375,
0.76318359375,
0.43115234375,
0.76318359375,
0.44580078125,
0.2344970703125,
0.44580078125
],
[
0.2320556640625,
0.44580078125,
0.765625,
0.4443359375,
0.765625,
0.458984375,
0.2320556640625,
0.460205078125
],
[
0.233642578125,
0.458984375,
0.7646484375,
0.458984375,
0.7646484375,
0.473388671875,
0.233642578125,
0.473388671875
],
[
0.2320556640625,
0.473388671875,
0.7646484375,
0.473388671875,
0.7646484375,
0.4873046875,
0.2320556640625,
0.4873046875
],
[
0.233642578125,
0.4873046875,
0.76318359375,
0.4873046875,
0.76318359375,
0.50048828125,
0.233642578125,
0.50048828125
],
[
0.233642578125,
0.50048828125,
0.765625,
0.50048828125,
0.765625,
0.51513671875,
0.233642578125,
0.51513671875
],
[
0.2320556640625,
0.51513671875,
0.76123046875,
0.51513671875,
0.76123046875,
0.52880859375,
0.2320556640625,
0.52880859375
],
[
0.2344970703125,
0.53076171875,
0.76318359375,
0.53076171875,
0.76318359375,
0.54150390625,
0.2344970703125,
0.54150390625
],
[
0.236083984375,
0.54443359375,
0.7646484375,
0.54443359375,
0.7646484375,
0.5546875,
0.236083984375,
0.5546875
],
[
0.2344970703125,
0.556640625,
0.765625,
0.556640625,
0.765625,
0.57080078125,
0.2344970703125,
0.57080078125
],
[
0.2344970703125,
0.57080078125,
0.301513671875,
0.57080078125,
0.301513671875,
0.58056640625,
0.2344970703125,
0.58056640625
],
[
0.17236328125,
0.611328125,
0.2034912109375,
0.611328125,
0.2034912109375,
0.62353515625,
0.17236328125,
0.62353515625
],
[
0.2034912109375,
0.6123046875,
0.31787109375,
0.6123046875,
0.31787109375,
0.62548828125,
0.2034912109375,
0.62548828125
],
[
0.176513671875,
0.6845703125,
0.8203125,
0.6845703125,
0.8203125,
0.69775390625,
0.176513671875,
0.69775390625
],
[
0.1732177734375,
0.69775390625,
0.82275390625,
0.6962890625,
0.82275390625,
0.7109375,
0.1732177734375,
0.7119140625
],
[
0.178955078125,
0.7138671875,
0.82177734375,
0.7138671875,
0.82177734375,
0.724609375,
0.178955078125,
0.724609375
],
[
0.176513671875,
0.72607421875,
0.82421875,
0.72607421875,
0.82421875,
0.7392578125,
0.176513671875,
0.7392578125
],
[
0.176513671875,
0.74072265625,
0.82177734375,
0.74072265625,
0.82177734375,
0.75390625,
0.176513671875,
0.75390625
],
[
0.176513671875,
0.75390625,
0.82421875,
0.75390625,
0.82421875,
0.767578125,
0.176513671875,
0.767578125
],
[
0.1748046875,
0.767578125,
0.82177734375,
0.76806640625,
0.82177734375,
0.7822265625,
0.1748046875,
0.78076171875
],
[
0.1748046875,
0.7822265625,
0.77294921875,
0.7822265625,
0.77294921875,
0.79541015625,
0.1748046875,
0.79541015625
],
[
0.775390625,
0.787109375,
0.78515625,
0.787109375,
0.78515625,
0.79150390625,
0.775390625,
0.79150390625
],
[
0.79150390625,
0.78515625,
0.81884765625,
0.78515625,
0.81884765625,
0.79345703125,
0.79150390625,
0.79345703125
],
[
0.1732177734375,
0.79736328125,
0.7294921875,
0.796875,
0.7294921875,
0.81103515625,
0.1732177734375,
0.8125
],
[
0.176513671875,
0.8193359375,
0.419189453125,
0.8212890625,
0.419189453125,
0.83203125,
0.176513671875,
0.830078125
],
[
0.410888671875,
0.818359375,
0.82275390625,
0.8193359375,
0.82275390625,
0.83349609375,
0.410888671875,
0.83203125
],
[
0.176513671875,
0.833984375,
0.82275390625,
0.833984375,
0.82275390625,
0.84765625,
0.176513671875,
0.84765625
],
[
0.1993408203125,
0.85693359375,
0.52294921875,
0.85791015625,
0.52294921875,
0.87255859375,
0.1993408203125,
0.8720703125
],
[
0.200927734375,
0.87255859375,
0.7294921875,
0.87255859375,
0.7294921875,
0.88623046875,
0.200927734375,
0.88623046875
],
[
0.1968994140625,
0.88623046875,
0.82275390625,
0.88623046875,
0.82275390625,
0.90087890625,
0.1968994140625,
0.90087890625
],
[
0.1748046875,
0.89794921875,
0.82275390625,
0.8994140625,
0.82275390625,
0.9140625,
0.1748046875,
0.912109375
],
[
0.176513671875,
0.9140625,
0.406982421875,
0.9140625,
0.406982421875,
0.9248046875,
0.176513671875,
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 2022",
"鲁棒学习满足生成模型：代理分布可以提高对抗性鲁",
"棒性吗？",
"Vikash Sehwag*t, Saeed Mahloujifar*, Tinashe Handina , Sihui Dai*, Chong Xiang",
"Mung Chiang*, Prateek Mittal*",
"*Princeton University, 'Caltech, *Purdue University",
"Abstract",
"While additional training data improves the robustness of deep neural net.",
"works against adversarial examples, it presents the challenge of curating a",
"large number of specific real-world samples. We circumvent this challenge",
"by using additional data from proxy distributions learned by advanced gen",
"erativemodels. Wefirst seek to formally understand the transfer of robust.",
"ness from classifiers trained on proxy distributions to the real data distri-",
"bution. We prove that the difference between the robustness of a classifier",
"on the two distributions is upper bounded by the conditional Wasserstein",
"distance between them. Next we use proxy distributions to significantly",
"improve the performance of adversarial training on five different datasets.",
"For example, we improve robust accuracy by up to 7.5 % and 6.7% in",
"loo and l2 threat model over baselines that are not using proxy distribu-",
"tions on the CIFAR-10 dataset. We also improve certified robust accuracy",
"by 7.6 % on the CIFAR-10 dataset.We further demonstrate that differ-",
"ent generative models bring a disparate improvement in the performance in",
"robust training. We propose a robust discrimination approach to character",
"izetheimpact ofindividualgenerativemodelsandfurtherprovideadeeper",
"understanding of why current state-of-the-art in diffusion-basedgenerative",
"models are a better choice for proxy distribution than generative adversarial",
"networks.",
"1",
"Introduction",
"深度神经网络是强大的工具，但它们的成功在很大程度上取决于训练数据量(Sunetal.",
"2017；Mahajanetal.，2018）。最近的工作表明，对于提高针对对抗性示例的鲁棒性训练F",
"（Biggioetal.，2013Szegedyetal，2014；Biggio&Roli，2018），额外的训练数据是有帮助",
"的E(Carmon et al., 2019; Schmidt et al., 2018a; Uesato et al., 2019; Deng et al., 2021) 。然",
"而，从实际数据分布中整理更多的现实世界数据通常是具有挑战性和成本的（Rechtetal.",
"2019）。为了规避这一挑战，我们问：是否可以使用代理分布来增强健壮的训练，即。e。，real",
"data分布的近似值？特别是，来自代理分布的额外样本可能更便宜，可以提高稳健性。",
"Ifso,cangenerativemodelsthataretrainedonlimitedtrainingimagesinsmallscaledatasets",
"",
"(Le",
"Cun & Cortes, 2010; Krizhevsky et al., 2014) ,actassuchaproxydistribution?1",
"当从生成模型中训练合成样本时，",
"，一个自然的问题是合成数据的鲁棒性是否也会转移到现",
"实世界的数据。即使是这样，我们是否可以确定能够实现这种合成到真实的稳健性转移的",
"t Corresponding author: vvikash@princeton.edu",
"Our code is available at https://github.com/inspire-group/proxy-distributions .",
"'Proxy distributions may not necessarily be modeled by generative models. When a proxy",
"distribution is the output of a generative model, we call it synthetic distribution and refer to data",
"sampled from it as synthetic data:",
"1"
] |
{
"arxiv_id": "2104.09425",
"language": "zh",
"page_idx": 0
}
|
||
2104.09425-zh-1
|
2104.09425-zh-1.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2022}}$
合成数据的特征，并根据这些特征优化我们对生成模型的选择？最后，我们还可以优化单个
合成样本的选择以最大化稳健性转移吗？
$\mathbb{Q}$。1 什么时候健壮性从代理分发转移到真正的数据分发？这个问题对于更好地理解代理分发是否会有所帮助是至关重要的。For a classifier trained on only synthetic samples, we argue that in addition to empirical and generalization error, distribution shift penalty also $\bar{\text{determines its robustness on the real data distribution.我们证明了这个惩罚是由 proxy 分}}$ 布和 real data 分布之间的 conditional Wasserstein distance 上限。因此，相对于 conditional Wasserstein distance , 健壮性将从与 real data 分布非常接近的代理分布转移
$Q。2$ How effective are proxy distributions in boosting adversarial robustness on real-world dataset? Our experimental results on fve datasets demonstrate that the use of samples from PrOxy distributions in Robust Training (PORT) can significantly improve robustness. In particular, PORT achieves up to to $7.5\%$ improvement in adversarial robustness over existing state-of-the-art (Croce et al., 2020). Its improvement is consistent across different threat models $(\ell_{\infty}$ or $\ell_{2})$, network architectures, datasets, and robustness criteria (empirical or certified robustness). We also uncover that synthetic images from diffusion-based generative models are most helpful in improving robustness on real datasets. We further $\begin{array}{c}\mathrm{investigate~the~use~of~proxy~distributions~in~robust~training.~In~particular,~we~investigate}\\\mathrm{investigate~the~use~of~proxy~distributions~in~robust~training.~In~particular,~we~investigate}\end{array}$ why current state-of-the-art in diffusion-based models are significantly more helpful than their counterparts, such as generative adversarial networks (GANs).
Q.3 Can we develop a metric to characterize which proxy distribution will be most helpful in robust training? Our theory motivates the design of a measure of proximity between two $\begin{array}{c}\text{distributions that incorporates the geometry of the distributions and can empirically predict}\\\text{distributions that incorporates the geometry of the distributions and can empirically predict}\end{array}$ the transfer of robustness. 我们提出了一种基于鲁棒学习的方法，在该方法中，我们使用鉴别器的成功来区分合成数据和真实数据的不利扰动样本，以此作为接近度的度量。区分合成数据和真实数据是一种常见的做法 (Goodfellow et al., 2014; Gui et al., 2020) , however, we find that considering adversarial perturbations on synthetic and real data samples is the key to making the discriminator an effective measure for this task. 我们证明，随着扰动大小的增加，鉴别器成功率的下降率可以有效地测量接近度，并且可以准确地预测来自不同生成模型的鲁棒性的相对转移
我们还利用强大的鉴别器来识别最有用的合成样本。我们使用每个合成样本与 real data 分布的接近度，称为合成分数，作为判断其重要性的指标。可以使用来自鉴别器的输出概率来计算此分数，该鉴别器经过了严格的训练，可以区分来自 proxy 和 real data 分布的不利扰动样本。We demonstrate that selecting synthetic images based on their synthetic scores can further improve performance. 贡献。我们做出以下主要贡献。
$\bullet$ We provide an analytical bound on the transfer of robustness from proxy distributions
to real data distribution using the notion of conditional Wasserstein distance and further
validate it experimentally.
$\bullet$ Overall, using additional synthetic images, we improve both clean and robust accuracy on five different datasets. In particular, we improve robust accuracy by up to $7.5\%$ and $6.7\%$ in $\ell_\infty$ and $\ell_2$ threat models, respectively, and certified robust accuracy( $\ell_2$ )by 7.6 $\%$ on the CIFAR-$\bar{1}0$ dataset.
$\bullet$ When selecting a proxy distribution, we show that existing metrics, such as FID, fail to
determine the synthetic-to-real transfer of robustness. We propose a new metric ( ARC) based on the distinguishability of adversarially perturbed synthetic and real data, that accurately determines the performance transfer.
$\bullet$ We also develop a metric, named synthetic score, to determine the importance of each
synthetic sample in synthetic-to-real robustness transfer. We demonstrate that choosing samples with lower synthetic scores provide better results than randomly selected samples.
2 Integrating proxy distributions in robust training
在本节中，我们建议使用来自代理分布的样本来提高鲁棒性。我们首先提供从 proxy 到 real data 分布的对抗性鲁棒性转移的分析界限。接下来，使用鲁棒鉴别器，我们提供了一个基于我们的分析界的度量，它可以准确地确定不同代理分布在鲁棒性转移方面的相对排名。我们的度量可以根据经验计算 (使用来自两个分布的样本),并且不需要了解 proxy 或 real data
2
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.17724609375,
0.1060791015625,
0.82177734375,
0.1060791015625,
0.82177734375,
0.11932373046875,
0.17724609375,
0.11932373046875
],
[
0.176513671875,
0.11932373046875,
0.461669921875,
0.11932373046875,
0.461669921875,
0.1331787109375,
0.176513671875,
0.1331787109375
],
[
0.17724609375,
0.1427001953125,
0.82177734375,
0.1427001953125,
0.82177734375,
0.15283203125,
0.17724609375,
0.15283203125
],
[
0.176513671875,
0.154052734375,
0.82177734375,
0.154052734375,
0.82177734375,
0.1673583984375,
0.176513671875,
0.1673583984375
],
[
0.176513671875,
0.1685791015625,
0.82275390625,
0.1685791015625,
0.82275390625,
0.1817626953125,
0.176513671875,
0.1817626953125
],
[
0.1748046875,
0.18115234375,
0.82177734375,
0.1817626953125,
0.82177734375,
0.1962890625,
0.1748046875,
0.1956787109375
],
[
0.176513671875,
0.1976318359375,
0.8203125,
0.1976318359375,
0.8203125,
0.2083740234375,
0.176513671875,
0.2083740234375
],
[
0.176513671875,
0.210205078125,
0.7294921875,
0.210205078125,
0.7294921875,
0.2235107421875,
0.176513671875,
0.2235107421875
],
[
0.1748046875,
0.2298583984375,
0.82421875,
0.2298583984375,
0.82421875,
0.2442626953125,
0.1748046875,
0.2442626953125
],
[
0.176513671875,
0.24560546875,
0.82275390625,
0.24560546875,
0.82275390625,
0.2587890625,
0.176513671875,
0.2587890625
],
[
0.1748046875,
0.2587890625,
0.82177734375,
0.2587890625,
0.82177734375,
0.272705078125,
0.1748046875,
0.272705078125
],
[
0.1748046875,
0.272705078125,
0.82177734375,
0.272705078125,
0.82177734375,
0.287353515625,
0.1748046875,
0.287353515625
],
[
0.176513671875,
0.287353515625,
0.82275390625,
0.287353515625,
0.82275390625,
0.300537109375,
0.176513671875,
0.300537109375
],
[
0.176513671875,
0.300537109375,
0.82177734375,
0.300537109375,
0.82177734375,
0.314453125,
0.176513671875,
0.314453125
],
[
0.1748046875,
0.314453125,
0.82421875,
0.314453125,
0.82421875,
0.32763671875,
0.1748046875,
0.32763671875
],
[
0.176513671875,
0.32763671875,
0.82275390625,
0.32763671875,
0.82275390625,
0.341552734375,
0.176513671875,
0.341552734375
],
[
0.1748046875,
0.341552734375,
0.82177734375,
0.341552734375,
0.82177734375,
0.354736328125,
0.1748046875,
0.354736328125
],
[
0.176513671875,
0.35595703125,
0.82275390625,
0.35595703125,
0.82275390625,
0.369384765625,
0.176513671875,
0.369384765625
],
[
0.176513671875,
0.371826171875,
0.66015625,
0.371826171875,
0.66015625,
0.3818359375,
0.176513671875,
0.3818359375
],
[
0.176513671875,
0.391357421875,
0.82421875,
0.391357421875,
0.82421875,
0.40478515625,
0.176513671875,
0.40478515625
],
[
0.1748046875,
0.40478515625,
0.82177734375,
0.40478515625,
0.82177734375,
0.41845703125,
0.1748046875,
0.41845703125
],
[
0.1748046875,
0.41845703125,
0.82177734375,
0.41845703125,
0.82177734375,
0.431884765625,
0.1748046875,
0.431884765625
],
[
0.176513671875,
0.43310546875,
0.82177734375,
0.43310546875,
0.82177734375,
0.4462890625,
0.176513671875,
0.4462890625
],
[
0.1748046875,
0.4462890625,
0.8203125,
0.4462890625,
0.8203125,
0.460205078125,
0.1748046875,
0.460205078125
],
[
0.176513671875,
0.460205078125,
0.82177734375,
0.460205078125,
0.82177734375,
0.473388671875,
0.176513671875,
0.473388671875
],
[
0.17724609375,
0.473388671875,
0.82275390625,
0.473388671875,
0.82275390625,
0.4873046875,
0.17724609375,
0.4873046875
],
[
0.17724609375,
0.4892578125,
0.8203125,
0.4892578125,
0.8203125,
0.5,
0.17724609375,
0.5
],
[
0.176513671875,
0.501953125,
0.82275390625,
0.501953125,
0.82275390625,
0.51513671875,
0.176513671875,
0.51513671875
],
[
0.17724609375,
0.517578125,
0.386474609375,
0.517578125,
0.386474609375,
0.52783203125,
0.17724609375,
0.52783203125
],
[
0.17724609375,
0.53857421875,
0.82177734375,
0.53857421875,
0.82177734375,
0.548828125,
0.17724609375,
0.548828125
],
[
0.176513671875,
0.55029296875,
0.82275390625,
0.55029296875,
0.82275390625,
0.564453125,
0.176513671875,
0.564453125
],
[
0.1748046875,
0.56298828125,
0.82177734375,
0.56298828125,
0.82177734375,
0.5771484375,
0.1748046875,
0.5771484375
],
[
0.176513671875,
0.5791015625,
0.82275390625,
0.5791015625,
0.82275390625,
0.59228515625,
0.176513671875,
0.59228515625
],
[
0.1748046875,
0.59326171875,
0.6435546875,
0.59326171875,
0.6435546875,
0.60693359375,
0.1748046875,
0.60693359375
],
[
0.1732177734375,
0.61279296875,
0.82275390625,
0.61279296875,
0.82275390625,
0.62744140625,
0.1732177734375,
0.62744140625
],
[
0.1895751953125,
0.62744140625,
0.82275390625,
0.62744140625,
0.82275390625,
0.6416015625,
0.1895751953125,
0.6416015625
],
[
0.191162109375,
0.6416015625,
0.378173828125,
0.6416015625,
0.378173828125,
0.65478515625,
0.191162109375,
0.65478515625
],
[
0.1732177734375,
0.65478515625,
0.82177734375,
0.65478515625,
0.82177734375,
0.66845703125,
0.1732177734375,
0.66845703125
],
[
0.1895751953125,
0.66748046875,
0.82177734375,
0.66748046875,
0.82177734375,
0.68115234375,
0.1895751953125,
0.68115234375
],
[
0.191162109375,
0.68310546875,
0.82421875,
0.68310546875,
0.82421875,
0.6962890625,
0.191162109375,
0.6962890625
],
[
0.191162109375,
0.6962890625,
0.39453125,
0.6962890625,
0.39453125,
0.71044921875,
0.191162109375,
0.71044921875
],
[
0.1732177734375,
0.71044921875,
0.82275390625,
0.71044921875,
0.82275390625,
0.724609375,
0.1732177734375,
0.724609375
],
[
0.191162109375,
0.724609375,
0.82275390625,
0.724609375,
0.82275390625,
0.73779296875,
0.191162109375,
0.73779296875
],
[
0.191162109375,
0.73779296875,
0.82275390625,
0.73779296875,
0.82275390625,
0.751953125,
0.191162109375,
0.751953125
],
[
0.1953125,
0.75244140625,
0.53515625,
0.75244140625,
0.53515625,
0.76318359375,
0.1953125,
0.76318359375
],
[
0.1732177734375,
0.76513671875,
0.82177734375,
0.76513671875,
0.82177734375,
0.77880859375,
0.1732177734375,
0.77880859375
],
[
0.1895751953125,
0.77880859375,
0.82275390625,
0.77978515625,
0.82275390625,
0.79443359375,
0.1895751953125,
0.79345703125
],
[
0.19287109375,
0.79443359375,
0.82177734375,
0.79443359375,
0.82177734375,
0.80810546875,
0.19287109375,
0.80810546875
],
[
0.17236328125,
0.833984375,
0.625,
0.83544921875,
0.625,
0.849609375,
0.17236328125,
0.8486328125
],
[
0.176513671875,
0.869140625,
0.82275390625,
0.869140625,
0.82275390625,
0.88330078125,
0.176513671875,
0.88330078125
],
[
0.1748046875,
0.88330078125,
0.82177734375,
0.8818359375,
0.82177734375,
0.896484375,
0.1748046875,
0.89794921875
],
[
0.176513671875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.82177734375,
0.9111328125,
0.82177734375,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.5,
0.94580078125,
0.51318359375,
0.95556640625,
0.5,
0.96533203125,
0.486083984375,
0.955078125
]
] |
[
"Published as a conference paper at ICLR 2022",
"合成数据的特征，并根据这些特征优化我们对生成模型的选择？最后，我们还可以优化单个",
"合成样本的选择以最大化稳健性转移吗？",
"Q。1什么时候健狂性从代理分发转移到真正的数据分发？这个问题对于更好地理解代理",
"分发是否会有所帮助是至关重要的。For a classifier trained on only synthetic samples，we",
"argue that in addition to empirical and generalization error, distribution shift penalty also",
"determinesitsrobustness ontherealdatadistribution.我们证明了这个惩罚是由proxy分",
"布和realdata分布之间的conditionalWassersteindistance上限。因此，相对于conditiona",
"Wassersteindistance，健壮性将从与realdata分布非常接近的代理分布转移",
"Q。 2 How effective are proxy distributions in boosting adversarial robustness on real-world",
"dataset? Our experimental results on five datasets demonstrate that the use of samples",
"from PrOxy distributions in Robust Training (PORT) can significantly improve robustness.",
"In particular, PORT achieves up to to 7.5 % improvement in adversarial robustness over",
"existing state-of-the-art (Croce et al., 2020). Its improvement is consistent across different",
"threat models ( lo or l2 ), network architectures, datasets, and robustness criteria (em-",
"pirical or certified robustness).We also uncover that synthetic images from diffusion-based",
"generative models are most helpful in improving robustness on real datasets. We further",
"investigate the use of proxy distributions in robust training. In particular, we investigate",
"why current state-of-the-art in diffusion-based models are significantly more helpful than",
"their counterparts,such as generative adversarialnetworks (GANs).",
"Q.3 Can we developametric to characterize which proxy distributionwill be most helpful",
"in robust training? Our theory motivates the design of a measure of proximity between two",
"distributions that incorporates the geometry of the distributions and can empirically predict",
"thetransferof robustness.我们提出了一种基于鲁棒学习的方法，在该方法中，我们使用鉴",
"别器的成功来区分合成数据和真实数据的不利扰动样本，以此作为接近度的度量。区分合",
"成数据和真实数据是一种常见的做法（Goodfellowetal.，2014；Guietal.，2020），however",
"we find that considering adversarial perturbations on synthetic and real data samples is the",
"keytomakingthediscriminatoraneffectivemeasureforthistask.我们证明，随着扰动大",
"小的增加，鉴别器成功率的下降率可以有效地测量接近度，并且可以准确地预测来自不同",
"生成模型的鲁棒性的相对转移",
"我们还利用强大的鉴别器来识别最有用的合成样本。我们使用每个合成样本与realdata分",
"布的接近度，称为合成分数，作为判断其重要性的指标。可以使用来自鉴别器的输出概率",
"来计算此分数，该鉴别器经过了严格的训练，可以区分来自proxy和realdata分布的不利",
"扰动样本。We demonstrate that selecting synthetic images based on their synthetic scores",
"canfurtherimproveperformance.贡献。我们做出以下主要贡献。",
": We provide an analytical bound on the transfer of robustness from proxy distributions",
"to real data distribution using the notion of conditional Wasserstein distance and further",
"validate it experimentally.",
". Overall, using additional synthetic images, we improve both clean and robust accuracy",
"on five different datasets. In particular, we improve robust accuracy by up to 7.5 % and",
"6.7% in loo and l2 threat models, respectively, and certified robust accuracy (l2) by 7.6",
"% on the CIFAR-10 dataset.",
": When selecting a proxy distribution, we show that existing metrics, such as FID, fail to",
"determine the symthetic-to-real transfer of robustness. We propose a new metric (ARC)",
"based on the distinguishability of adversarially perturbed synthetic and real data, that",
"accuratelydeterminestheperformancetransfer.",
": We also develop a metric, named symthetic score, to determine the importance of each",
"synthetic sample in synthetic-to-real robustness transfer. We demonstrate that choosing",
"samples with lower synthetic scores provide better results than randomly selected samples.",
"2Integrating proxydistributionsinrobusttraining",
"在本节中，我们建议使用来自代理分布的样本来提高鲁棒性。我们首先提供从proxy到real",
"data分布的对抗性鲁棒性转移的分析界限。接下来，使用鲁棒鉴别器，我们提供了一个基于",
"我们的分析界的度量，它可以准确地确定不同代理分布在鲁棒性转移方面的相对排名。我们",
"的度量可以根据经验计算（使用来自两个分布的样本），并且不需要了解proxy或realdata",
"2"
] |
{
"arxiv_id": "2104.09425",
"language": "zh",
"page_idx": 1
}
|
||
2104.09425-zh-2
|
2104.09425-zh-2.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2022}}$
分布。最后，我们提出了我们的鲁棒训练公式 (PORT),它使用生成模型生成的合成样本
以及真实样本
表示法。我们用$\chi$表示输人空间，并将相应的标签空间表示为$\gamma$ 。从$\chi\times\gamma$支持的联合分布$D$中采样数据。对于标签$y$ ,我们使用$D\mid y$表示类$y$的条件分布。我们将代理分发表示为$\tilde{D}$。我们用$f:\mathcal{X}\to\mathcal{Z}$表示用于分类的神经网络，由$\theta$参数化，它将输入图像映射到输出概率向量$(z$ )。我们使用$h$来引用输出标签的分类函数。对于从分布$D$采样的集合$\mathcal{S}$ ,我们使用$\hat{S}$表示相对于集合$\mathcal{S}$的经验分布。我们使用$\mathcal{S}\sim D$表示来自分布$D$的数据集的采样。我们使用$(x,y)\leftarrow D$来表示$D$中单个点的采样
2.1 Understanding transfer of adversarial robustness between data distributions
Since our goal is to use samples from proxy distribution to improve robustness on real data distribution, we first study the transfer of adversarial robustness between two data distributions.
Definition 1 ( Average Robustness) . We define average robustness for a classiffer $h$ on a distribution $\bar{D}$ according to a distance metric $d$ as follows: $\operatorname{Rob}_d(h,D)=$ $\mathbb{E}_{(x,y)\leftarrow D}[\operatorname*{inf}_{h(x^{\prime})\neq y}d(x^{\prime},x)].$ where classifier $h:\mathcal{X}\to\mathcal{Y}$ predicts class label of an input $\mathrm{sample~}^{2}.$
此定义是指每个$\mathbb{E}$样本到最接近的对抗性示例的预期距离
将健壮性从 proxy 正式转移到 real data 分布。在从代理分布进行鲁棒学习时，当训练集是从代理分布$\tilde{D}$采样的$n$标记示例的集合$\mathcal{S}$时，我们有兴趣限制由学习算法($L$)在分布$D$上获得的分类器的平均鲁棒性。特别是，我们希望在 transferred average robustness i 上$\mathop{\text{提供一个下限。e、}\mathbb{E}}_{h\leftarrow L(S)}^{\mathcal{S}\sim\tilde{D}}[\operatorname{Rob}_{d}(h,D)].$
为了更好地理解这个量，假设$h$是在从$\tilde{D}$采样的集合$\mathcal{S}$上训练的分类器，使用算法$L$ 。
我们将 Rob$_d(h,D)$分解为三个量，如下所示：
$\operatorname{Rob}_{d}(h,D)=+\operatorname{Rob}_{d}(h,D)-\operatorname{Rob}_{d}(h,\tilde{D})\boldsymbol{\longleftarrow}++\operatorname{Rob}_{d}(h,\tilde{D})-\operatorname{Rob}_{d}(h,\hat{\mathcal{S}})\boldsymbol{\longleftarrow}+\operatorname{Rob}_{d}(h,\hat{\mathcal{S}}).$
使用此分解，通过期望的线性和三角不等式，我们可以通过
$$\underbrace{\mathbb{E}_{\mathcal{S}\leftarrow\bar{D}^{n}}\left[\mathrm{Rob}_{d}(h,\hat{\mathcal{S}})\right]}_{Empiricalrobustness}-\underbrace{\left|\mathbb{E}_{\mathcal{S}\leftarrow\bar{D}^{n}}\left[\mathrm{Rob}_{d}(h,\tilde{D})-\mathrm{Rob}_{d}(h,\hat{\mathcal{S}})\right]\right|}_{Generalizationpenalty}-\underbrace{\left|\mathbb{E}_{\mathcal{S}\leftarrow\bar{D}^{n}}\left[\mathrm{Rob}_{d}(h,D)-\mathrm{Rob}_{d}(h,\tilde{D})\right]\right|}_{Distribution-shiftpenalty}$$
从下面约束传递的平均鲁棒性
正如上面的分解所表明的那样，为了约束平均鲁棒性，我们需要同时约束泛化惩罚和分布移位惩罚。在多个作品 (Cullina et al., 2018; Montasser et al., 2019; Schmidt et al., 2018b) 之前，已经对泛化惩罚进行了严格的研究。因此，我们专注于限制分布移位惩罚。我们的目标是在分布偏移惩罚上提供一个界限，该界限与手中的分类器无关，并且仅与分布的属性有关。有了这个目标，我们定义了两个分布之间的距离的概念
Definition 2 (Conditional Wasserstein distance). For two labeled distributions $D$ and $\tilde{D}$
supported on $\mathcal{X}\times\mathcal{Y}$, we define cwd according to a distance metric $d$ as follows:
$$\mathrm{cwd}_{d}(D,\tilde{D})=\underset{(\cdot,y)\leftarrow D}{\operatorname*{\mathbb{E}}}\underset{J\in\mathcal{J}(D|y,\tilde{D}|y)}{\operatorname*{\operatorname*{inf}}}\underset{(x,x^{\prime})\leftarrow J}{\operatorname*{\mathbb{E}}}[d(x,x^{\prime})]\underset{\nearrow}{\operatorname*{\times}}$$
where $\mathcal{J}(D,\tilde{D})$ is the set of joint distributions whose marginals are identical to $D$ and $\tilde{D}.$ 它只是每个类的条件分布之间的 Wasserstein 距离的期望。现在，我们准备好陈述我们的主要定理，该定理限制了仅基于两个分布的 Wasserstein 距离的任何学习算法的分布移位惩罚Theorem 1 (Bounding distribution-shift penalty). Let $D$ and $\tilde{D}$ be two labeled distributions supported on $\mathcal{X}\times\mathcal{Y}$ with identical label distributions, i.e., $\forall y^* \in \mathcal{Y} , \mathrm{Pr}_{( x, y) \leftarrow D}[ y= y^* ] =$ $\Pr _{( x, y) \leftarrow \tilde{D} }[ y= y^{* }]$ .Then for any classifier $h:\mathcal{X}\to\mathcal{Y}$
$$|\mathrm{Rob}_{d}(h,\tilde{D})-\mathrm{Rob}_{d}(h,D)|\leq\mathrm{cwd}_{d}(D,\tilde{D}).$$
$^{2}$This notion of robustness is also used in Gilmer et al. (2018); Diochnos et al. (2018) and
Mahloujifar et al. (2019a) .
3
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.176513671875,
0.1060791015625,
0.82275390625,
0.1060791015625,
0.82275390625,
0.11932373046875,
0.176513671875,
0.11932373046875
],
[
0.17724609375,
0.121826171875,
0.273681640625,
0.121826171875,
0.273681640625,
0.1319580078125,
0.17724609375,
0.1319580078125
],
[
0.176513671875,
0.141357421875,
0.82275390625,
0.141357421875,
0.82275390625,
0.1546630859375,
0.176513671875,
0.1546630859375
],
[
0.176513671875,
0.154052734375,
0.82177734375,
0.154052734375,
0.82177734375,
0.1673583984375,
0.176513671875,
0.1673583984375
],
[
0.176513671875,
0.169189453125,
0.82177734375,
0.169189453125,
0.82177734375,
0.184326171875,
0.176513671875,
0.184326171875
],
[
0.176513671875,
0.184326171875,
0.82275390625,
0.184326171875,
0.82275390625,
0.1976318359375,
0.176513671875,
0.1976318359375
],
[
0.1748046875,
0.199462890625,
0.82275390625,
0.199462890625,
0.82275390625,
0.21337890625,
0.1748046875,
0.21337890625
],
[
0.1748046875,
0.2139892578125,
0.5986328125,
0.2139892578125,
0.5986328125,
0.2279052734375,
0.1748046875,
0.2279052734375
],
[
0.1732177734375,
0.2442626953125,
0.7607421875,
0.2442626953125,
0.7607421875,
0.25830078125,
0.1732177734375,
0.25830078125
],
[
0.1748046875,
0.26953125,
0.82275390625,
0.26953125,
0.82275390625,
0.28271484375,
0.1748046875,
0.28271484375
],
[
0.17724609375,
0.285400390625,
0.82177734375,
0.285400390625,
0.82177734375,
0.29541015625,
0.17724609375,
0.29541015625
],
[
0.1748046875,
0.296630859375,
0.268798828125,
0.298583984375,
0.268798828125,
0.31005859375,
0.1748046875,
0.30810546875
],
[
0.1748046875,
0.313232421875,
0.826171875,
0.313232421875,
0.826171875,
0.32763671875,
0.1748046875,
0.32763671875
],
[
0.176513671875,
0.32763671875,
0.82275390625,
0.32763671875,
0.82275390625,
0.341552734375,
0.176513671875,
0.341552734375
],
[
0.17236328125,
0.343505859375,
0.82421875,
0.341552734375,
0.82421875,
0.359130859375,
0.17236328125,
0.361083984375
],
[
0.1732177734375,
0.363037109375,
0.2467041015625,
0.359130859375,
0.2484130859375,
0.373779296875,
0.1748046875,
0.37744140625
],
[
0.17724609375,
0.38818359375,
0.5810546875,
0.38818359375,
0.5810546875,
0.3984375,
0.17724609375,
0.3984375
],
[
0.176513671875,
0.40771484375,
0.82177734375,
0.40771484375,
0.82177734375,
0.421630859375,
0.176513671875,
0.421630859375
],
[
0.1748046875,
0.421630859375,
0.82275390625,
0.42236328125,
0.82275390625,
0.4375,
0.1748046875,
0.436279296875
],
[
0.176513671875,
0.4375,
0.82275390625,
0.4375,
0.82275390625,
0.45068359375,
0.176513671875,
0.45068359375
],
[
0.1748046875,
0.451904296875,
0.465576171875,
0.451904296875,
0.465576171875,
0.466552734375,
0.1748046875,
0.466552734375
],
[
0.32177734375,
0.458984375,
0.371826171875,
0.462158203125,
0.370849609375,
0.47607421875,
0.3203125,
0.472900390625
],
[
0.1748046875,
0.48291015625,
0.82421875,
0.484130859375,
0.82421875,
0.498779296875,
0.1748046875,
0.49755859375
],
[
0.176513671875,
0.498779296875,
0.5009765625,
0.498779296875,
0.5009765625,
0.51220703125,
0.176513671875,
0.51220703125
],
[
0.1748046875,
0.51806640625,
0.859375,
0.51806640625,
0.859375,
0.53271484375,
0.1748046875,
0.53271484375
],
[
0.1748046875,
0.5390625,
0.59716796875,
0.5390625,
0.59716796875,
0.55322265625,
0.1748046875,
0.55322265625
],
[
0.19287109375,
0.55810546875,
0.8203125,
0.55810546875,
0.8203125,
0.57275390625,
0.19287109375,
0.57275390625
],
[
0.1805419921875,
0.5712890625,
0.22216796875,
0.5712890625,
0.22216796875,
0.5869140625,
0.1805419921875,
0.5869140625
],
[
0.3349609375,
0.57080078125,
0.37744140625,
0.57275390625,
0.375732421875,
0.58837890625,
0.334228515625,
0.5859375
],
[
0.59228515625,
0.57275390625,
0.6337890625,
0.57275390625,
0.6337890625,
0.5869140625,
0.59228515625,
0.5869140625
],
[
0.178955078125,
0.59130859375,
0.2998046875,
0.59130859375,
0.2998046875,
0.60498046875,
0.178955078125,
0.60498046875
],
[
0.375732421875,
0.59228515625,
0.51123046875,
0.59228515625,
0.51123046875,
0.60595703125,
0.375732421875,
0.60595703125
],
[
0.62109375,
0.59228515625,
0.77783203125,
0.59228515625,
0.77783203125,
0.60595703125,
0.62109375,
0.60595703125
],
[
0.176513671875,
0.609375,
0.3896484375,
0.609375,
0.3896484375,
0.62255859375,
0.176513671875,
0.62255859375
],
[
0.176513671875,
0.62890625,
0.82177734375,
0.6298828125,
0.82177734375,
0.64306640625,
0.176513671875,
0.642578125
],
[
0.17724609375,
0.64599609375,
0.5458984375,
0.64599609375,
0.5458984375,
0.65576171875,
0.17724609375,
0.65576171875
],
[
0.5556640625,
0.64599609375,
0.8203125,
0.64599609375,
0.8203125,
0.65576171875,
0.5556640625,
0.65576171875
],
[
0.1748046875,
0.6572265625,
0.8203125,
0.6572265625,
0.8203125,
0.67041015625,
0.1748046875,
0.67041015625
],
[
0.1748046875,
0.67041015625,
0.82275390625,
0.671875,
0.82275390625,
0.68603515625,
0.1748046875,
0.68505859375
],
[
0.176513671875,
0.6875,
0.615234375,
0.6875,
0.615234375,
0.69775390625,
0.176513671875,
0.69775390625
],
[
0.1748046875,
0.70263671875,
0.82421875,
0.7021484375,
0.82421875,
0.71630859375,
0.1748046875,
0.71728515625
],
[
0.176513671875,
0.71826171875,
0.74658203125,
0.71826171875,
0.74658203125,
0.73193359375,
0.176513671875,
0.73193359375
],
[
0.293212890625,
0.73779296875,
0.48046875,
0.736328125,
0.48046875,
0.75048828125,
0.293212890625,
0.75244140625
],
[
0.47802734375,
0.736328125,
0.5439453125,
0.73779296875,
0.5439453125,
0.751953125,
0.47802734375,
0.75
],
[
0.57177734375,
0.73779296875,
0.69921875,
0.73779296875,
0.69921875,
0.75244140625,
0.57177734375,
0.75244140625
],
[
0.396240234375,
0.74853515625,
0.611328125,
0.74853515625,
0.611328125,
0.76220703125,
0.396240234375,
0.76220703125
],
[
0.1748046875,
0.767578125,
0.8203125,
0.7666015625,
0.8203125,
0.78076171875,
0.1748046875,
0.7822265625
],
[
0.17724609375,
0.79248046875,
0.82177734375,
0.79248046875,
0.82177734375,
0.80615234375,
0.17724609375,
0.80615234375
],
[
0.176513671875,
0.80615234375,
0.82177734375,
0.80615234375,
0.82177734375,
0.8193359375,
0.176513671875,
0.8193359375
],
[
0.1748046875,
0.82373046875,
0.82275390625,
0.82470703125,
0.82275390625,
0.8388671875,
0.1748046875,
0.83837890625
],
[
0.176513671875,
0.84033203125,
0.82275390625,
0.84033203125,
0.82275390625,
0.85498046875,
0.176513671875,
0.85498046875
],
[
0.1732177734375,
0.85693359375,
0.5556640625,
0.85595703125,
0.5556640625,
0.87060546875,
0.1732177734375,
0.8720703125
],
[
0.3505859375,
0.87744140625,
0.64794921875,
0.876953125,
0.64794921875,
0.8916015625,
0.3505859375,
0.892578125
],
[
0.19287109375,
0.89453125,
0.82421875,
0.89794921875,
0.82421875,
0.91552734375,
0.19287109375,
0.912109375
],
[
0.1732177734375,
0.91015625,
0.35302734375,
0.912109375,
0.35302734375,
0.92529296875,
0.1732177734375,
0.923828125
],
[
0.491943359375,
0.94873046875,
0.5087890625,
0.94873046875,
0.5087890625,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2022",
"分布。最后，我们提出了我们的鲁棒训练公式（PORT），它使用生成模型生成的合成样本",
"以及真实样本",
"表示法。我们用表示输人空间，并将相应的标签空间表示为。从×支持的联合",
"分布D中采样数据。对于标签y，我们使用DIy表示类y的条件分布。我们将代理分发",
"表示为D。我们用f：一2表示用于分类的神经网络，由θ参数化，它将输人图像映射",
"到输出概率向量（z）。我们使用h来引用输出标签的分类函数。对于从分布D采样的集合",
"S，我们使用S表示相对于集合S的经验分布。我们使用S~D表示来自分布D的数据",
"集的采样。我们使用（α,y）←D来表示D中单个点的采样",
"2.1 Understanding transfer of adversarial robustness between data distributions",
"Since our goal is to use samples from proxy distribution to improve robustness on real",
"datadistribution,wefirst studythe transferofadversarialrobustness between two data",
"distributions.",
"Definition 1 (Average Robustness). We define average robustness for a classifier h",
"on a distribution D according to a distance metric d as follows:Robd(h,D) =",
"E(r,y)-D[infh(n)+y d(r,a)]. where classifier h : X → predicts class label of an input",
"sample 2.",
"此定义是指每个E样本到最接近的对抗性示例的预期距离",
"将健壮性从proxy正式转移到realdata分布。在从代理分布进行鲁棒学习时，当训练集是",
"从代理分布D采样的n标记示例的集合S时，我们有兴趣限制由学习算法（L）在分布",
"D上获得的分类器的平均鲁棒性。特别是，我们希望在transferred averagerobustnessi上",
"提供一个下限。e、Es-D[Roba(h,D)].",
"h=L(S)",
"为了更好地理解这个量，假设h是在从D采样的集合S上训练的分类器，使用算法L。",
"我们将Robd(h,D）分解为三个量，如下所示：",
"Roba(h, D) =Roba(h, D) - Roba(h, D) +Roba(h, D) - Roba(h, S) + Roba(h, S)",
"使用此分解，通过期望的线性和三角不等式，我们可以通过",
"E[Roba(h,S)] -E, [Roba(h,D) -Roba(h,S)] -/E[Robd(h,D) -Roba(h, D)]] ",
"h-L(S)",
"hs-L(s)",
"h-L(s)",
"Empiricalrobustnes",
"Generalizationpenalty",
"Distribution-shiftpenalty",
"从下面约束传递的平均鲁棒性",
"正如上面的分解所表明的那样，为了约束平均鲁棒性，我们需要同时约束泛化惩罚和分布",
"移位惩罚。在多个作品（Cullinaetal.，2018：Monta",
"ieretal.,2019:Schmidtetal.,2018b",
"之前，已经对泛化惩罚进行了严格的研究。因此，我们专注于限制分布移位惩罚。我们的目",
"标是在分布偏移惩罚上提供一个界限，该界限与手中的分类器无关，并且仅与分布的属性",
"有关。有了这个自标，我们定义了两个分布之间的距离的概念",
"Definition 2 (Conditional Wasserstein distance). For two labeled distributions D and D",
"supported on X × , we define cwd according to a distance metric d as follows:",
"cwda(D,D)=EX",
"inf",
"E[d(a,r)]X",
"(,y)-D /JEJ(Dly,Dly) (a,r')-",
"where J(D, D) is the set of joint distributions whose marginals are identical to D and D ",
"它只是每个类的条件分布之间的Wasserstein距离的期望。现在，我们准备好陈述我们的主",
"要定理，该定理限制了仅基于两个分布的Wasserstein距离的任何学习算法的分布移位惩罚",
"Theorem 1 (Bounding distribution-shift penalty). Let D and D be two labeled distributions",
"supported on X × y with identical label distributions, i.e., Vy* E ,Pr(r,y)-D[y = y*] =",
"Pr(a,y)-D[y = y*1 . Then for any classifier h : X -→ y)",
"[Roba(h, D) -Roba(h, D)/ ≤ cwda(D, D).",
"2This notion of robustness is also used in Gilmer et al. (2018); Diochnos et al. (2018) and",
"Mahloujifaretal.(2019a)",
"3"
] |
{
"arxiv_id": "2104.09425",
"language": "zh",
"page_idx": 2
}
|
||
2104.09425-zh-3
|
2104.09425-zh-3.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2022}}$
定理 1 展示了如何通过最小化两个分布之间的条件 Wasserstein 距离来限制分布偏移惩罚。We provide its proof and empirical validation in Appendix A 。Even if we successfully reduce the generalization penalty, the distribution-shift penalty may remain the dominant factor in transferred robustness. 该定理使我们能够将注意力从稳健的泛化转移到创建基本分布接近原始分布的生成模型。In Appendix A, we also provide two other theorems about the tightness of Theorem 1 and the effect of combining clean and proxy data together.
2.2 Which proxy distribution to choose? - Approximating conditional Wasserstein
distance using robust discriminators
我们的分析结果表明，conditional Wasserstein distance 到 real data 较小的代理分布可以提高鲁棒性。这意味着在代理分发中优化 conditional Wasserstein distance (e。g。, 生成模型) 可能会导致对抗鲁棒性的提高。不幸的是，仅使用分布 (Mahloujifar et al.,2019b; Bhagoji et al., 2019) 中的样本来经验地计算 conditional Wasserstein distance 通常是一项艰巨的任务。在本节中，我们介绍一个名为 ARC 的指标作为 conditional Wasserstein distance 的代理，当我们只能对数据分布进行样本访问时，可以计算该指标。在定义我们的度量之前，我们需要定义两个分布之间的鲁棒判别的概念
<FigureHere>
( b) Perturbed
( a) Original samples.
samples.
鲁棒判别。我们的指标基于鉴别器可以在实际和代理分布的样本之间进行可靠区分的程度。与典型的鉴别器相比，即使有对手可以通过大小为$\epsilon$的扰动来扰动采样实例，鲁棒的鉴别器也应该是准确的。直观上，如果存在两个发行版$D$和$\tilde{D}$的成功且强大的鉴别器，则意味着即使进行大小为$\epsilon$的扰动，也没有对手可以使$D$的实例看起来像$\tilde{D}$的实例。。这意味着$D$和$\tilde{D}$彼此远离(图$\mathbb{E}$1b )。On the other hand, if no robust discriminator exists, then it means that Figure 1: Why robust discriminathe adversary can"align“most of the mass of proxy tion is effective. (a) Without addistribution with that of real distribution by making versarial perturbations, a non-robust perturbations of size at most $\epsilon.$ This notion is closely discriminator can perfectly distinrelated to the optimal transport between the proxy guish both proxy distributions from and real distribution. In particular, the optimal ad- the real distribution. (b) Under adversary here corresponds to a transport $J\in \mathcal{J} ( D, \tilde{D} )$ versarial perturbations, classification that minimizes Pr$_{( x, x^{\prime }) \leftarrow J( D, \tilde{D} ) }[ d( x, x^{\prime })$ $\leq \epsilon ]$ (Com- between real and proxy-1 is much pare with Definition 2).更具体地说，鲁棒鉴别器 harder compared to real and proxy- $\sigma \text{ 的 目 标 是 最 大 化 其 鲁 棒 精 度 }^{3}: 2$ distribution. The success of a ro- $\mathrm{Racc}_{\epsilon}(\sigma,D,\tilde{D})=\frac{1}{2}\cdotp\mathrm{Pr}_{x\leftarrow D}\huge\diagup\forall x^{\prime}\in Ball_{\epsilon}(x);\sigma(x^{\prime})=\underset{\text{trIONingipity of vea}}{\operatorname*{\text{tributions.}}}\overset{\text{fiNe}}{\operatorname*{\operatorname*{\operatorname*{\operatorname*{\operatorname*{mosipity~of~of~at}}}}}}\text{h'e Br}\partial\Re_{\xi}(x\text{fis}^{\prime}\sigma(x^{\prime})$
我们使用 Racc$_{\epsilon}^*(D,\tilde{D})$表示最佳现有鲁棒鉴别器的精度，即 Racc$_\epsilon ^* ( D, \tilde{D} )$ = $\operatorname*{min}_\sigma\operatorname{Racc}_\epsilon(\sigma,D,\tilde{D})$ 。请注意，随着$\epsilon$从 0 增长到$\infty$ ,任何鉴别器的鲁棒精度将最多降低到 50%。如果两个分布$D$和$\tilde{D}$彼此接近，则精度下降的速度要快得多，因为对手可以通过添加较小的扰动来使两个分布的图像看起来相似。我们使用这种直觉并定义我们的
2 distribution. The success of a ro-
度量 ARC 如下：
$$\mathrm{ARC}(D,\tilde{D})=\int_0^\infty(\mathrm{Racc}_\epsilon^*(D,\tilde{D})-\frac{1}{2})d\epsilon.$$
换句话说，ARC 度量等于鲁棒判别精度 v 下的面积。s。扰动曲线 (该曲线的样本见图 3 。)在 3.2 节中，我们讨论了如何仅使用来自 real data 和 proxy 分布的样本来计算此度量。When selecting a proxy distribution, we choose the one with the smallest ARC value, since it is expected to be closest to real data distribution. We also demonstrate that ARC metric $\hat{\text{is directly related to conditional Wasserstein distance. 特别是，我们有以下定理：}}$
Theorem 2. For any two distributions $\tilde{D}$ and $D$ with equal class probabilities we have cwd$(D,\tilde{D})\geq4\cdot$ARC$( D, \tilde{D} )$ .Moreover, if for all labels $y$ , $( D\mid y)$ and $(\tilde{D}\mid y)$ are two concentric uniform spheres, then we have cwd$(D,\tilde{D})=4\cdot$ARC$(D,\tilde{D})$
$\overline{^3\text{All the definitions of accuracy a}}$nd robust accuracy in this section are defined for true distribu-
tions. However, in the experiments we work with their empirical variant.
4
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.17724609375,
0.10479736328125,
0.82177734375,
0.10479736328125,
0.82177734375,
0.11932373046875,
0.17724609375,
0.11932373046875
],
[
0.176513671875,
0.11932373046875,
0.82177734375,
0.11932373046875,
0.82177734375,
0.1331787109375,
0.176513671875,
0.1331787109375
],
[
0.17724609375,
0.1331787109375,
0.82177734375,
0.1331787109375,
0.82177734375,
0.146484375,
0.17724609375,
0.146484375
],
[
0.1732177734375,
0.1458740234375,
0.8203125,
0.146484375,
0.8203125,
0.160400390625,
0.1732177734375,
0.1590576171875
],
[
0.17724609375,
0.1622314453125,
0.82177734375,
0.1622314453125,
0.82177734375,
0.17236328125,
0.17724609375,
0.17236328125
],
[
0.176513671875,
0.1749267578125,
0.79345703125,
0.1749267578125,
0.79345703125,
0.1881103515625,
0.176513671875,
0.1881103515625
],
[
0.1748046875,
0.20703125,
0.2091064453125,
0.20703125,
0.2091064453125,
0.2183837890625,
0.1748046875,
0.2183837890625
],
[
0.20751953125,
0.208984375,
0.21728515625,
0.208984375,
0.21728515625,
0.212158203125,
0.20751953125,
0.212158203125
],
[
0.214111328125,
0.205810546875,
0.77685546875,
0.205810546875,
0.77685546875,
0.2197265625,
0.214111328125,
0.2197265625
],
[
0.2115478515625,
0.2197265625,
0.473876953125,
0.2197265625,
0.473876953125,
0.23291015625,
0.2115478515625,
0.23291015625
],
[
0.17724609375,
0.246826171875,
0.8203125,
0.246826171875,
0.8203125,
0.2568359375,
0.17724609375,
0.2568359375
],
[
0.176513671875,
0.2587890625,
0.734375,
0.2587890625,
0.734375,
0.272705078125,
0.176513671875,
0.272705078125
],
[
0.734375,
0.260009765625,
0.8203125,
0.260009765625,
0.8203125,
0.270751953125,
0.734375,
0.270751953125
],
[
0.176513671875,
0.271484375,
0.82275390625,
0.270751953125,
0.82275390625,
0.285400390625,
0.176513671875,
0.285888671875
],
[
0.1732177734375,
0.285400390625,
0.82275390625,
0.285888671875,
0.82275390625,
0.300537109375,
0.1732177734375,
0.2998046875
],
[
0.176513671875,
0.300537109375,
0.82275390625,
0.300537109375,
0.82275390625,
0.314453125,
0.176513671875,
0.314453125
],
[
0.176513671875,
0.314453125,
0.82275390625,
0.314453125,
0.82275390625,
0.32763671875,
0.176513671875,
0.32763671875
],
[
0.176513671875,
0.328857421875,
0.48779296875,
0.328857421875,
0.48779296875,
0.34228515625,
0.176513671875,
0.34228515625
],
[
0.176513671875,
0.349853515625,
0.54833984375,
0.349853515625,
0.54833984375,
0.363037109375,
0.176513671875,
0.363037109375
],
[
0.1748046875,
0.3623046875,
0.54833984375,
0.363037109375,
0.54833984375,
0.376953125,
0.1748046875,
0.375732421875
],
[
0.77685546875,
0.356689453125,
0.7861328125,
0.356689453125,
0.7861328125,
0.361083984375,
0.77685546875,
0.361083984375
],
[
0.176513671875,
0.376953125,
0.54833984375,
0.376953125,
0.54833984375,
0.39013671875,
0.176513671875,
0.39013671875
],
[
0.176513671875,
0.391357421875,
0.54833984375,
0.391357421875,
0.54833984375,
0.40478515625,
0.176513671875,
0.40478515625
],
[
0.178955078125,
0.40771484375,
0.5458984375,
0.40771484375,
0.5458984375,
0.41845703125,
0.178955078125,
0.41845703125
],
[
0.775390625,
0.40771484375,
0.78759765625,
0.40771484375,
0.78759765625,
0.4140625,
0.775390625,
0.4140625
],
[
0.176513671875,
0.419189453125,
0.54833984375,
0.419189453125,
0.54833984375,
0.43310546875,
0.176513671875,
0.43310546875
],
[
0.701171875,
0.42236328125,
0.7294921875,
0.42236328125,
0.7294921875,
0.429931640625,
0.701171875,
0.429931640625
],
[
0.73779296875,
0.421630859375,
0.814453125,
0.421630859375,
0.814453125,
0.431884765625,
0.73779296875,
0.431884765625
],
[
0.176513671875,
0.43505859375,
0.54736328125,
0.43505859375,
0.54736328125,
0.448974609375,
0.176513671875,
0.448974609375
],
[
0.56640625,
0.442626953125,
0.69140625,
0.442626953125,
0.69140625,
0.455810546875,
0.56640625,
0.455810546875
],
[
0.71728515625,
0.4375,
0.7998046875,
0.438232421875,
0.7998046875,
0.449462890625,
0.71728515625,
0.448974609375
],
[
0.176513671875,
0.45068359375,
0.5458984375,
0.45068359375,
0.5458984375,
0.464111328125,
0.176513671875,
0.464111328125
],
[
0.73388671875,
0.449462890625,
0.779296875,
0.449462890625,
0.779296875,
0.460205078125,
0.73388671875,
0.460205078125
],
[
0.176513671875,
0.466552734375,
0.54736328125,
0.466552734375,
0.54736328125,
0.4765625,
0.176513671875,
0.4765625
],
[
0.56396484375,
0.466552734375,
0.8203125,
0.466552734375,
0.8203125,
0.479736328125,
0.56396484375,
0.479736328125
],
[
0.176513671875,
0.479248046875,
0.5498046875,
0.479248046875,
0.5498046875,
0.492431640625,
0.176513671875,
0.492431640625
],
[
0.56201171875,
0.479736328125,
0.82177734375,
0.479736328125,
0.82177734375,
0.49365234375,
0.56201171875,
0.49365234375
],
[
0.1748046875,
0.492431640625,
0.54833984375,
0.492431640625,
0.54833984375,
0.505859375,
0.1748046875,
0.505859375
],
[
0.564453125,
0.49365234375,
0.82275390625,
0.49365234375,
0.82275390625,
0.5068359375,
0.564453125,
0.5068359375
],
[
0.1748046875,
0.505859375,
0.54833984375,
0.505859375,
0.54833984375,
0.51953125,
0.1748046875,
0.51953125
],
[
0.56396484375,
0.50830078125,
0.82177734375,
0.50830078125,
0.82177734375,
0.521484375,
0.56396484375,
0.521484375
],
[
0.176513671875,
0.51806640625,
0.5458984375,
0.52099609375,
0.5458984375,
0.5341796875,
0.176513671875,
0.5322265625
],
[
0.56396484375,
0.521484375,
0.82177734375,
0.521484375,
0.82177734375,
0.53515625,
0.56396484375,
0.53515625
],
[
0.1748046875,
0.5341796875,
0.54833984375,
0.5341796875,
0.54833984375,
0.54736328125,
0.1748046875,
0.54736328125
],
[
0.56201171875,
0.53515625,
0.82177734375,
0.53515625,
0.82177734375,
0.548828125,
0.56201171875,
0.548828125
],
[
0.17724609375,
0.548828125,
0.5498046875,
0.54736328125,
0.5498046875,
0.56103515625,
0.17724609375,
0.5625
],
[
0.560546875,
0.5498046875,
0.82177734375,
0.5498046875,
0.82177734375,
0.56298828125,
0.560546875,
0.56298828125
],
[
0.1732177734375,
0.564453125,
0.5537109375,
0.564453125,
0.5537109375,
0.58203125,
0.1732177734375,
0.58203125
],
[
0.560546875,
0.56298828125,
0.82275390625,
0.56298828125,
0.82275390625,
0.5771484375,
0.560546875,
0.5771484375
],
[
0.1732177734375,
0.58056640625,
0.5498046875,
0.580078125,
0.5498046875,
0.5947265625,
0.1732177734375,
0.59521484375
],
[
0.560546875,
0.5771484375,
0.8203125,
0.57763671875,
0.8203125,
0.59130859375,
0.560546875,
0.59033203125
],
[
0.56396484375,
0.59130859375,
0.8203125,
0.59130859375,
0.8203125,
0.60498046875,
0.56396484375,
0.60498046875
],
[
0.1748046875,
0.59765625,
0.4150390625,
0.59765625,
0.4150390625,
0.611328125,
0.1748046875,
0.611328125
],
[
0.56396484375,
0.60498046875,
0.82275390625,
0.60498046875,
0.82275390625,
0.61865234375,
0.56396484375,
0.61865234375
],
[
0.1732177734375,
0.61865234375,
0.90869140625,
0.61865234375,
0.90869140625,
0.6396484375,
0.1732177734375,
0.6396484375
],
[
0.564453125,
0.6337890625,
0.63623046875,
0.6337890625,
0.63623046875,
0.64599609375,
0.564453125,
0.64599609375
],
[
0.176513671875,
0.6572265625,
0.7939453125,
0.6572265625,
0.7939453125,
0.671875,
0.176513671875,
0.671875
],
[
0.176513671875,
0.6748046875,
0.82177734375,
0.6748046875,
0.82177734375,
0.68798828125,
0.176513671875,
0.68798828125
],
[
0.176513671875,
0.689453125,
0.8203125,
0.689453125,
0.8203125,
0.70263671875,
0.176513671875,
0.70263671875
],
[
0.1748046875,
0.70263671875,
0.82177734375,
0.7041015625,
0.82177734375,
0.71826171875,
0.1748046875,
0.71728515625
],
[
0.176513671875,
0.7197265625,
0.294189453125,
0.7197265625,
0.294189453125,
0.72998046875,
0.176513671875,
0.72998046875
],
[
0.358642578125,
0.73779296875,
0.4599609375,
0.73779296875,
0.4599609375,
0.75244140625,
0.358642578125,
0.75244140625
],
[
0.476318359375,
0.73291015625,
0.64208984375,
0.7373046875,
0.640625,
0.7587890625,
0.4755859375,
0.7548828125
],
[
0.1748046875,
0.7646484375,
0.826171875,
0.76513671875,
0.826171875,
0.77978515625,
0.1748046875,
0.77880859375
],
[
0.1846923828125,
0.77978515625,
0.82275390625,
0.77978515625,
0.82275390625,
0.79345703125,
0.1846923828125,
0.79345703125
],
[
0.176513671875,
0.79345703125,
0.82177734375,
0.79345703125,
0.82177734375,
0.806640625,
0.176513671875,
0.806640625
],
[
0.1748046875,
0.80810546875,
0.82275390625,
0.80810546875,
0.82275390625,
0.8212890625,
0.1748046875,
0.8212890625
],
[
0.1732177734375,
0.82080078125,
0.75732421875,
0.8212890625,
0.75732421875,
0.8359375,
0.1732177734375,
0.83544921875
],
[
0.176513671875,
0.84033203125,
0.82275390625,
0.84033203125,
0.82275390625,
0.85498046875,
0.176513671875,
0.85498046875
],
[
0.1748046875,
0.85693359375,
0.82275390625,
0.85693359375,
0.82275390625,
0.87060546875,
0.1748046875,
0.87060546875
],
[
0.176513671875,
0.87353515625,
0.68701171875,
0.87353515625,
0.68701171875,
0.88720703125,
0.176513671875,
0.88720703125
],
[
0.19287109375,
0.89453125,
0.82275390625,
0.89794921875,
0.82275390625,
0.91552734375,
0.19287109375,
0.912109375
],
[
0.17724609375,
0.91357421875,
0.65283203125,
0.91357421875,
0.65283203125,
0.923828125,
0.17724609375,
0.923828125
],
[
0.490234375,
0.95263671875,
0.50341796875,
0.94775390625,
0.5107421875,
0.95849609375,
0.496826171875,
0.96337890625
]
] |
[
"Published as a conference paper at ICLR 2022",
"定理1展示了如何通过最小化两个分布之间的条件Wasserstein距离来限制分布偏移惩",
"罚。 We provide its proof and empirical validation in Appendix A 。 Even if we successfully",
"reduce the generalization penalty, the distribution-shift penalty may remain the dominant",
"factorintransferredrobustness.该定理使我们能够将注意力从稳健的泛化转移到创建基本",
"分布接近原始分布的生成模型。InAppendixA：wealsoprovidetwoothertheoremsabout",
"the tightness of Theorem 1 and the effect of combining clean and proxy data together.",
"2.2",
"",
"Which proxy distribution to choose? - Approximating conditional Wasserstein",
"distance using robust discriminators",
"我们的分析结果表明：conditionalWassersteindistance到realdata较小的代理分布可以提",
"高鲁棒性。这意味着在代理分发中优化conditionalWassersteindistance（e。g。",
"，生成模型",
"可能会导致对抗鲁棒性的提高。不幸的是，仅使用分布（Mahloujifaretal.，2019b；Bhagoji",
"etal.，2019）中的样本来经验地计算conditionalWassersteindistance通常是一项艰巨的任",
"务。在本节中，我们介绍一个名为ARC的指标作为conditionalWassersteindistance的代",
"理，当我们只能对数据分布进行样本访问时，可以计算该指标。在定义我们的度量之前，我",
"们需要定义两个分布之间的鲁棒判别的概念",
"鲁棒判别。我们的指标基于鉴别器可以在实际和代",
"理分布的样本之间进行可靠区分的程度。与典型的鉴",
"O",
"别器相比，即使有对手可以通过大小为e的扰动来",
"扰动采样实例，鲁棒的鉴别器也应该是准确的。直观",
"上，如果存在两个发行版D和D的成功且强大的鉴",
"",
"别器，则意味着即使进行大小为ε的扰动，也没有对",
"Real",
"Proxy-1 Proxy-2",
"手可以使D的实例看起来像D的实例。。这意味着",
"(a) Original samples.",
"(b)Perturbed",
"D和D彼此远离（图Eib）。Ontheotherhand",
"samples",
"ifnorobust discriminatorexists,thenitmeansthat",
"Figure l: Why robust discrimina.",
"the adversary can \"align\" most of the mass of proxy",
"tion is effective.(a) Without ad-",
"distribution with that of real distribution by making",
"versarial perturbations, a non-robust",
"perturbations of size at most e . This notion is closely",
"discriminator can perfectly distin",
"related to the optimal transport between the proxy",
"guish both proxy distributions from",
"and real distribution. In particular, the optimal ad-",
"the real distribution: (b) Under ad-",
"versary here corresponds to a transport J e J(D, D)",
"versarial perturbations, classification",
"that minimizes Pr(r,r)-J(D,D)[d(r,a) ≤ el (Com-",
"between real and proxy-l is much",
"parewithDefinition2)：更具体地说，鲁棒鉴别器",
"hardercomparedtoreal andproxy",
"2 distribution. The success of a ro-",
"α的目标是最大化其鲁棒精度3：",
"bust discriminator is dependent on",
"Racce(o, D, D) = ·Prr-D X Va e Ball(a);o(a) -thsprokipity Bfxavahd Bay(djso(a) = 0 X",
"trbutions",
"我们使用Racc*(D,D）表示最佳现有鲁棒鉴别器的精度，即Racc*（D,D)",
"min。Racceo,D,D）。请注意，随着e从0增长到o，任何鉴别器的鲁棒精度将最多",
"降低到50%。如果两个分布D和D彼此接近，则精度下降的速度要快得多，因为对手可",
"以通过添加较小的扰动来使两个分布的图像看起来相似。我们使用这种直觉并定义我们的",
"度量ARC如下：",
"ARC(D,D) =",
" (Racc (D, D) - )de.",
"换句话说，ARC度量等于鲁棒判别精度v下的面积。S。扰动曲线（该曲线的样本见图3",
"）在3.2节中，我们讨论了如何仅使用来自realdata和proxy分布的样本来计算此度量。",
"When selecting a proxy distribution, we choose the one with the smallest ARC value, since",
"it is expected to be closest to real data distribution. We also demonstrate that ARC metric",
"isdirectlyrelated toconditionalWassersteindistance.特别是，我们有以下定理：",
"Theorem 2. For any two distributions D and D with equal class probabilities we have",
"cwd(D, D) ≥ 4 : ARC(D, D) . Moreover, if for all labels y , (D I y) and (D I y) are two",
"concentric uniform spheres, then we have cwd(D,D) = 4 · ARC(D, D) .",
"3All the definitions of accuracy and robust accuracy in this section are defined for true distribu-",
"tions.However, in the experiments we work with their empirical variant",
"4"
] |
{
"arxiv_id": "2104.09425",
"language": "zh",
"page_idx": 3
}
|
||
2104.09425-zh-4
|
2104.09425-zh-4.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2022}}$
为什么非鲁棒判别不是有效的度量。深度神经网络通常用作合成图像和真实图像之间的鉴别器 (Goodfellow et al., 2014; Gui et al., 2020) 。在我们的实验中，我们发现即使是四层卷积网络也可以在所有选定的代理分布上实现接近 100 的% 分类精度 (附录EC.1 )。此结果表明，非鲁棒性判别不是接近度的有效度量(如图Ela 所示)
2.3 Bringing it all together: Improving adversarial robustness using proxy distributions
鲁棒训练概述。鲁棒训练的关键目标是求解$min_{\theta }\mathbb{E} _{( x, y) \sim D}L_{adv}( \theta , x, y, \Omega )$ ,其中$L_adv$由训练目标决定，$\Omega$ 是威胁模型。在对抗训练$\boxed{\mathrm{E}}($Madry et al., 2018; Zhang et al., 2019; Carmon et al., 2019; Pang et al., 2021) 中，我们通过选择$L_adv(.)=\ell(\theta,PG\tilde{D}(x,\Omega),y)$来针对基于投影梯度下降 (PGD) 的对抗示例进行训练。在随机平滑 (Cohen et al., 2019; Salman et al. 2019; Sehwag et al., 2020; Carmon et al.,2019)中，目标是实现经认证的鲁棒性 (Wong $\&$ Kolter, 2018) , 我们的目标是通过选择$L_{adv}^{\prime}(.)=\ell(\theta,x,y)+\beta D_{kl}(f_{\theta}(x),f_{\theta}(x+\delta))$来实现对随机高斯噪声的不变性，其中$\delta\sim\mathcal{N}(0,\sigma^2I)$和$D_kl(...)$是 KL-散度。这个损失函数(Zheng et al., 2016) 被称为鲁棒稳定性训练 (RST), 其性能优于使用$\ell ( \theta , x+ \delta , y)$ ( Carmon et al. , 2019)。我们将对抗训练和随机平滑中的$L_adv$分别称为$L_{pgd}$和$L_{smooth}$ 。我们将用于训练分类器的交叉熵损失函数表示为$\ell(.)$,其中$\ell(\theta,x,y)=\langle-log(f_\theta(x)),y\rangle$和$\boldsymbol{y}$是$y$的独热向量编码
我们的目标是在稳健训练 (PORT) 中集成来自代理分布的合成样本。特别是，我们使用以
下强大的训练公式，将来自真实训练数据集的样本与合成样本相结合。
$$\begin{matrix}min&L_{agg};&L_{agg}=\gamma&\mathbb{E}_{(x,y)\sim D}L_{adv}(\theta,x,y,\Omega)+(1-\gamma)&\mathbb{E}_{(x,y)\sim\bar{D}}L_{adv}(\theta,x,y,\Omega)\end{matrix}$$
其中$\gamma\in[0,1]$和$L_{agg}$是$D$和$\tilde{D}$上的合计对抗损失。根据训练目标，我们选择$L_{pgd}$或$L_{smooth}$作为$L_adv$。我们使用来自 real data 分布$(D)$的可用训练样本和从代理分布$(\tilde{D}$ ) 采样的一组合成图像来近似汇总损失。我们通过超参数搜索选择$\gamma$
Which synthetic samples to choose? 特别是，来自代理分布 ($\tilde{D}$ )的$N$合成样本的集合 ($S$ )导致对真实数据的最大传输鲁棒性，即。e。,$\mathbb{E}_{f\leftarrow\hat{L}(S)}[\mathrm{Rob}_d(f,D)]$ ,其中$\hat{L}$是一种鲁棒的学习算法。虽然它可以通过许多不同的方式解决，但我们遵循基于拒绝抽样的方法。特别是，我们测量每个合成样本与 real data 分布的接近度，并将其用作接受或拒绝样本的度量。我们使用来自训练有素的 robust 鉴别器的输出概率分数作为接近度的度量。特别是，我们将 synthetic-score $(x)=\min_{x_i\in Ball_\epsilon(x)}\sigma(x_i)$定义为每个合成样本的最终得分，其中得分最低的样本在选择中优先考虑
# 3 Experimental results
首先，我们证明了我们提出的使用代理分布的方法在$\ell_\infty$和$\ell_2$威胁模型中的对抗性训练和随机平滑的性能方面都有显着改善 (E3.1 节)。We also uncover that diffusion-based generative models are more effective proxy distributions than generative adversarial networks (GANs) on multiple datasets. Next we delve deeper into why diffusion-based generative models are highly effective proxy distributions (Section 3.2). We find that the key reason behind the success of diffusion based models is the that there is no robust discriminator that can robustly distinguish between the samples from diffusion based models and real data distribution. We demonstrate that our proposed metric $(\begin{array}{c}\mathrm{ARC})\end{array}$provides a better characterization of the effectiveness of diffusion-based models than other widely used metrics. Finally, we investigate the effectiveness of adaptively selecting synthetic images in our framework.
3.1 Improving robustness using synthetic images from proxy distributions
安装程序。We consider five datasets, namely CIFAR-10 (Krizhevsky et al., 2014), CIFAR- 100 (Krizhevsky et al., 2014), CelebA (Liu et al., 2015), AFHQ (Choi et al., 2020), and ImageNet ( Deng et al. , 2009) . We also work with commonly used $l_{\infty }$ and $l_2$ threat models for adversarial attacks and use AutoAttack (Croce &Hein, 2020) to measure robustness. We provide extensive details on the setup in Appendix B.
5
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.17724609375,
0.1060791015625,
0.82177734375,
0.1060791015625,
0.82177734375,
0.11932373046875,
0.17724609375,
0.11932373046875
],
[
0.17724609375,
0.11932373046875,
0.82177734375,
0.11932373046875,
0.82177734375,
0.1331787109375,
0.17724609375,
0.1331787109375
],
[
0.176513671875,
0.1331787109375,
0.82177734375,
0.1331787109375,
0.82177734375,
0.146484375,
0.176513671875,
0.146484375
],
[
0.176513671875,
0.147705078125,
0.62841796875,
0.147705078125,
0.62841796875,
0.1610107421875,
0.176513671875,
0.1610107421875
],
[
0.1732177734375,
0.179931640625,
0.81201171875,
0.179931640625,
0.81201171875,
0.1944580078125,
0.1732177734375,
0.1944580078125
],
[
0.17724609375,
0.2083740234375,
0.82275390625,
0.2083740234375,
0.82275390625,
0.222900390625,
0.17724609375,
0.222900390625
],
[
0.52685546875,
0.2215576171875,
0.560546875,
0.2215576171875,
0.560546875,
0.2291259765625,
0.52685546875,
0.2291259765625
],
[
0.17724609375,
0.2310791015625,
0.82177734375,
0.2310791015625,
0.82177734375,
0.2442626953125,
0.17724609375,
0.2442626953125
],
[
0.1748046875,
0.24560546875,
0.82177734375,
0.24560546875,
0.82177734375,
0.2587890625,
0.1748046875,
0.2587890625
],
[
0.176513671875,
0.2587890625,
0.82177734375,
0.2587890625,
0.82177734375,
0.272705078125,
0.176513671875,
0.272705078125
],
[
0.1748046875,
0.271484375,
0.82275390625,
0.272705078125,
0.82275390625,
0.287353515625,
0.1748046875,
0.285888671875
],
[
0.1748046875,
0.287353515625,
0.82275390625,
0.287353515625,
0.82275390625,
0.300537109375,
0.1748046875,
0.300537109375
],
[
0.1748046875,
0.3017578125,
0.82275390625,
0.302978515625,
0.82275390625,
0.317626953125,
0.1748046875,
0.31640625
],
[
0.176513671875,
0.317626953125,
0.82177734375,
0.317626953125,
0.82177734375,
0.330810546875,
0.176513671875,
0.330810546875
],
[
0.1748046875,
0.330078125,
0.82421875,
0.330078125,
0.82421875,
0.347900390625,
0.1748046875,
0.347900390625
],
[
0.1748046875,
0.343505859375,
0.82275390625,
0.3447265625,
0.82275390625,
0.359130859375,
0.1748046875,
0.35791015625
],
[
0.17724609375,
0.361083984375,
0.2418212890625,
0.361083984375,
0.2418212890625,
0.371826171875,
0.17724609375,
0.371826171875
],
[
0.176513671875,
0.380126953125,
0.82275390625,
0.380126953125,
0.82275390625,
0.393310546875,
0.176513671875,
0.393310546875
],
[
0.17724609375,
0.39453125,
0.6748046875,
0.39453125,
0.6748046875,
0.40771484375,
0.17724609375,
0.40771484375
],
[
0.23779296875,
0.41845703125,
0.7587890625,
0.417236328125,
0.7587890625,
0.43505859375,
0.23779296875,
0.436279296875
],
[
0.3896484375,
0.43310546875,
0.4306640625,
0.43310546875,
0.4306640625,
0.43994140625,
0.3896484375,
0.43994140625
],
[
0.60888671875,
0.431884765625,
0.65673828125,
0.431884765625,
0.65673828125,
0.442626953125,
0.60888671875,
0.442626953125
],
[
0.176513671875,
0.45703125,
0.82275390625,
0.45703125,
0.82275390625,
0.4716796875,
0.176513671875,
0.4716796875
],
[
0.1748046875,
0.473388671875,
0.82275390625,
0.472900390625,
0.82275390625,
0.4873046875,
0.1748046875,
0.488037109375
],
[
0.176513671875,
0.490478515625,
0.6650390625,
0.490478515625,
0.6650390625,
0.50048828125,
0.176513671875,
0.50048828125
],
[
0.1748046875,
0.51123046875,
0.82421875,
0.51123046875,
0.82421875,
0.5244140625,
0.1748046875,
0.5244140625
],
[
0.1748046875,
0.52587890625,
0.82421875,
0.52587890625,
0.82421875,
0.54345703125,
0.1748046875,
0.54345703125
],
[
0.176513671875,
0.54345703125,
0.82421875,
0.54345703125,
0.82421875,
0.556640625,
0.176513671875,
0.556640625
],
[
0.17724609375,
0.5595703125,
0.82177734375,
0.5595703125,
0.82177734375,
0.5693359375,
0.17724609375,
0.5693359375
],
[
0.1748046875,
0.5693359375,
0.82275390625,
0.5712890625,
0.82275390625,
0.58544921875,
0.1748046875,
0.58349609375
],
[
0.1732177734375,
0.583984375,
0.82421875,
0.583984375,
0.82421875,
0.6015625,
0.1732177734375,
0.6015625
],
[
0.17724609375,
0.6015625,
0.370849609375,
0.6015625,
0.370849609375,
0.6123046875,
0.17724609375,
0.6123046875
],
[
0.1707763671875,
0.63525390625,
0.3857421875,
0.63623046875,
0.3857421875,
0.65380859375,
0.1707763671875,
0.65283203125
],
[
0.17724609375,
0.671875,
0.8203125,
0.671875,
0.8203125,
0.681640625,
0.17724609375,
0.681640625
],
[
0.176513671875,
0.6845703125,
0.82177734375,
0.6845703125,
0.82177734375,
0.69775390625,
0.176513671875,
0.69775390625
],
[
0.1732177734375,
0.6962890625,
0.8203125,
0.6962890625,
0.8203125,
0.71044921875,
0.1732177734375,
0.71044921875
],
[
0.176513671875,
0.7119140625,
0.82177734375,
0.7119140625,
0.82177734375,
0.72607421875,
0.176513671875,
0.72607421875
],
[
0.1748046875,
0.724609375,
0.82177734375,
0.724609375,
0.82177734375,
0.73779296875,
0.1748046875,
0.73779296875
],
[
0.1748046875,
0.73779296875,
0.82177734375,
0.73779296875,
0.82177734375,
0.751953125,
0.1748046875,
0.751953125
],
[
0.176513671875,
0.75390625,
0.8203125,
0.75390625,
0.8203125,
0.767578125,
0.176513671875,
0.767578125
],
[
0.1748046875,
0.7666015625,
0.82421875,
0.767578125,
0.82421875,
0.7822265625,
0.1748046875,
0.78076171875
],
[
0.176513671875,
0.78076171875,
0.82275390625,
0.78076171875,
0.82275390625,
0.79443359375,
0.176513671875,
0.79443359375
],
[
0.1732177734375,
0.79443359375,
0.79150390625,
0.79541015625,
0.79150390625,
0.81005859375,
0.1732177734375,
0.8095703125
],
[
0.1748046875,
0.8291015625,
0.71728515625,
0.8291015625,
0.71728515625,
0.84326171875,
0.1748046875,
0.84326171875
],
[
0.176513671875,
0.85498046875,
0.8203125,
0.85498046875,
0.8203125,
0.869140625,
0.176513671875,
0.869140625
],
[
0.176513671875,
0.869140625,
0.82177734375,
0.869140625,
0.82177734375,
0.88330078125,
0.176513671875,
0.88330078125
],
[
0.176513671875,
0.8837890625,
0.82421875,
0.8837890625,
0.82421875,
0.89794921875,
0.176513671875,
0.89794921875
],
[
0.176513671875,
0.89794921875,
0.82177734375,
0.89794921875,
0.82177734375,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.176513671875,
0.91357421875,
0.58642578125,
0.91357421875,
0.58642578125,
0.923828125,
0.176513671875,
0.923828125
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.96337890625,
0.491943359375,
0.96337890625
]
] |
[
"Published as a conference paper at ICLR 2022",
"为什么非鲁棒判别不是有效的度量。深度神经网络通常用作合成图像和真实图像之间的鉴",
"别器（Goodfellowetal.，2014Guietal.，2020）。在我们的实验中，我们发现即使是四层",
"卷积网络也可以在所有选定的代理分布上实现接近100的%分类精度（附录EC.1）。此结",
"果表明，非鲁棒性判别不是接近度的有效度量（如图E1a所示）",
"2.3Bringing it all together: Improving adversarial robustness using proxy distributions",
"鲁棒训练概述。鲁棒训练的关键目标是求解minELadu(e,a,y,2），其中Ladu由训练",
"r,y)~",
"目标决定，2是威胁模型。在对抗训练E(Madryetal.，2018；Zhangetal.，2019;Carmon",
"etal.,2019;Pangetal.，2021）中，我们通过选择Ladu(）=l(0,PGD(c2),y）来针对基于",
"投影梯度下降（PGD）的对抗示例进行训练。在随机平滑（Cohenetal.2019;Salmanetal.",
"2019;Sehwagetal.，2020；Carmonetal.，2019）中，目标是实现经认证的鲁棒性（Wong&",
"Kolter，2018），我们的目标是通过选择Ladu)=l(0,a，y)+βDkz(fe(r)，f(α+)）来实现对",
"随机高斯噪声的不变性，其中~N(O,I）和Dkz(）是KL-散度。这个损失函数（Zheng",
"etal.,2016)被称为鲁棒稳定性训练(RST)，其性能优于使用e(0,a+,y）（Carmonetal.",
"2019）。我们将对抗训练和随机平滑中的Ladu分别称为Lpgd和Lsmooth。我们将用于训练",
"分类器的交叉嫡损失函数表示为l(），其中l(o,a,y）=<-log(fe(a),y）和y是y的独热",
"向量编码",
"我们的目标是在稳健训练（PORT）中集成来自代理分布的合成样本。特别是，我们使用以",
"下强大的训练公式，将来自真实训练数据集的样本与合成样本相结合。",
"min Laggi Lagg =ELad(,r, y, 2) + (1 -),ELad(0,r, y, 2)",
"T.)~",
"(a,y)~D",
"其中E[0,1]和Lagg是D和D上的合计对抗损失。根据训练目标，我们选择Lpgd或",
"Lsmooth作为Ladu。我们使用来自realdata分布（D）的可用训练样本和从代理分布（D",
"）采样的一组合成图像来近似汇总损失。我们通过超参数搜索选择",
"Whichsyntheticsamplestochoose？特别是，来自代理分布（D）的N合成样本的集合（S",
"）导致对真实数据的最大传输鲁棒性，即。e。，Ef-L(s)[Roba(f,D)】，其中L是一种鲁棒",
"的学习算法。虽然它可以通过许多不同的方式解决，但我们遵循基于拒绝抽样的方法。特别",
"是，我们测量每个合成样本与realdata分布的接近度，并将其用作接受或拒绝样本的度量。",
"我们使用来自训练有素的robust鉴别器的输出概率分数作为接近度的度量。特别是，我们",
"将 synthetic-score（a）=mina;eBalle(a)a(ci）定义为每个合成样本的最终得分，其中得分最",
"低的样本在选择中优先考虑",
"3 Experimental results",
"首先，我们证明了我们提出的使用代理分布的方法在1和2威模型中的对抗性训练和随",
"机平滑的性能方面都有显着改善（E3.1节）。Wealsouncoverthatdiffusion-basedgenerative",
"models are more effective proxy distributions than generative adversarial networks (GANs)",
"on multiple datasets. Next we delve deeper into why diffusion-based generative models are",
"highly effective proxy distributions (Section 3.2). We find that the key reason behind the",
"success of diffusion based models is the that there is no robust discriminator that can robustly",
"distinguish between the samples from diffusion based models and real data distribution",
"We demonstrate that our proposed metric ( ARC) provides a better characterization of",
"the effectiveness of diffusion-based models than other widely used metrics. Finally, we",
"investigate the effectiveness of adaptively selecting synthetic images in our framework.",
"3.1Improving robustness using synthetic images from proxy distributions",
"安装程序。We consider fivedatasets,namely CIFAR-10(Krizhevsky et al., 2014),CIFAR",
"100 (Krizhevsky et al., 2014), CelebA (Liu et al., 2015), AFHQ (Choi et al., 2020), and",
"ImageNet (Deng et al., 2009). We also work with commonly used loo and l2 threat models",
"for adversarial attacks and use AutoAttack (Croce & Hein, 2020) to measure robustness.",
"Weprovideextensivedetailson thesetup inAppendix B.",
"5"
] |
{
"arxiv_id": "2104.09425",
"language": "zh",
"page_idx": 4
}
|
||
2210.04307-zh-0
|
2210.04307-zh-0.jpg
|
THMATHX\_ MAXXINE THMALINE\_1\_
XMATHX\_ <NEXT\_LNE XMATHISH\_。基于自我注意的语言模型 (SAMs) 从领域不可知的大量训练语料库[?]中捕获更广泛的上下文。在特定于领域的语料库上微调 SAM 可实现特定于领域的上下文捕获 [??]。然而，SAM 架构本质上是黑盒 [?]。因此微调的 SAM 体系结构不适合对开放研究目标进行可靠的评估：(R1) 相关的特定领域上下文覆盖范围，以及 (R2)知识上下文的影响与下游任务中的数据上下文进行权衡[??]。我们提出了一种改进的自我注意架构，注人了知识的自我注意转换器 (KSAT) 来解决这些目标。KSAT 在选定的特定领域任务上表现良好 (请参阅 1.2 ), 同时有助于对 R1 和 R2 进行可靠的人类可理解的评估。因此，KSAT 为促进 AI 用户信任和满意度[??]迈出了重要的一步。
# 1 васкаюшими $^{[\mathrm{m}]}$
## 1.1 Related Work
与 R1 和 R2 相关并包含多个知识上下文的现有方法可以根据知识注人技术大致分为(1)知识调制的 SAM 和(2)知识注人的基于输人嵌人的 SAM [??]。前者使用知识来指导 SAM 中的自我注意机制，后者在将输入传递到 SAM 之前将知识嵌入到向量空间中。在这里，我们简要总结一下它们对 R1 和 R2 的贡献。Category(1) 和 Category(2) 方法的域覆盖率都是通过特定于域的任务描述 (R1) 的性能来评估的。这些方法的消融突出了知识上下文 (R2)的贡献。但是，检查模型组件 (投影矩阵和向量)的数值输出并不容易使其易于理解。。需要可解释的 AI 技术 (将其转换为人类可理解的信息的数值输出的后处理)来确认作者的观点[?]。基于后处理的解释是对特定输人的 SAM 推理的局部近似，因此不呈现全局图片，从而对 SAM 评估的有效性产生怀疑。KSAT 提出了一种SAM 架构，其数值输出有助于对 R1 和 R2 进行可靠的人类可理解的评估
1.2 Task Description, Data, and External Knowledge Sources
尽管 KSAT 体系结构广泛适用于任何特定领域的任务，但我们选择了Gaur 等人的自杀倾向心理健康诊断辅助的特定任务。[?]。我们将此数据集表示为 MHDA。数据包含有关自杀相关子 reddits 的 Reddit 帖子的高质量专家注释。注释方法确保了来自测量伪影的最小噪声和专家注释者之间的高度一致性。我们使用哥伦比亚自杀严重程度评定量表 (CSSRS) 中包含的临床确定的诊断过程信息。图1 (a,b) 说明了自杀模式可能出现的各种上下文 (每个树路径代表一个上下文)。任务是预测自杀模式，即-适应症， 观念 1,观念 2 和行为或企图。图 1 (c)显示了来自 MHDA 数据集的示例，并增加了知识上下文注释。我们将增强的数据集表示为 k-MHDA。k-MHDA 包含帖子和句子级别的知识上下文注释 (请参见图 1 (c))。我们将 k-MHDA 的构造细节从 CSSRS 知识和MHDA 数据推迟到附录部分 A ,因为它不是论文$^1$的主要焦点。
2 КSАТ- РROPOSED АRCHITECTURE
对于香草 SAM 层$l$中的输人$x$,引人了$CLS$令牌，该令牌对下游任务的数据中的信息进行编码。例如，对于分类$Z_\mathrm{CLS}$,来自最终层$L$ 的 CLS 令牌的表示通过输出类概率的 softmax 层传递。在 KSAT 层中，我们引人了一个$Z_{KCLS}$ 令牌，该令牌对该层的知识进行编码。$Z_{KCLS}$值由输入$x$句子与输人$x$的知识上下文注释中的概念之间的图形上下文确定。例如，考虑输人$x$类似于图 1 (c) 中所示的帖子 PostX1 和 PostX2, 然后图 1 (d) 中的图形说明了显示句子概念连接的图形上下文。人们会期望两个帖子中的句子S1 和 S2 的$Z_{KCLS}$编码将是相似的，因为它们具有相似的图上下文。$KG_{bias}$项确保$Z_{KCLS}$中的值捕获此行为 (请参见图 2 (b))
# 2.1 The KSAT layer
与香草 SAM 层相比，每个 KSAT 层都有以下主要区别：
2.1.1 Layer Parameters and Outputs. 让$Y$表示与每个上下文相关的自杀结果集：行为指示或无，想法1，。让$x$表示一个输入帖子
(例如帖子，请参见图 1 (c))。对于每个输人$x$,每个KSAT 层$l$输出由公式 1 给出的每个结果$y\in Y$的概率向量：
(1)
$W,Z_{CLS}(x)$和$Z_KCLS(x)$的尺寸为 384×1。在 MHDA 语料库上进行微调，用于嵌人输入[?]。回想一下，术语$KG_{bias}$用于确保$Z_{CLS}(x)$对表示为图形的知识上下文进行编码(2.1.2中的详细信息)。$KG_\boldsymbol{bias}$的维度大小为$1\times1$ 。标量项$\alpha_l$用作权衡因子，使用凸组合来调制数据和知识上下文之间的贡献 (请参见公式 1)。我们将 KSAT 层输出与香草 SAM 层进行对比，其中类似的概率向量仅在最后一层输出，而不是在每一层输出(见图2)。用于计算自注意矩阵的香草 SAM 层中的参数(查询，键和值投影矩阵) 保留在 KSAT 层[?] 中。与 vanillaSAM 层不同，在 vanillaSAM 层中 , 参数可以跨层共享，也可以不跨层共享，因为不同的层对不同的知识上下文进行编码。
$$P_{l}^{ksat}(y|x)=\sigma(W^{T}(\alpha_{l}Z_{KCLS}(x)+(1-\alpha_{l})Z_{CLS}(x))+KG_{bias}(x))$$
2.1.2 Encoding Craph Contexts in Layers. 即使每个 KSAT 层$\iota$输出每个结果的概率向量$y\in\{$指示或无，Ideation1, Ideation2, 行为或尝试},表示每个上下文，它在$Z_{KCLS}$中只编码一个上下文。我们展示了如何通过引用图1 (c,d)中的帖子PostX1和 PostX2以及图上下文$\underline\text{Ideation1}$和 Behavior or Attempt 来计算$KG_{bias}$项
$ComputingSentence-Concept$ Connection Vectors: 我们首先为帖子中的每个句子计算一个句子-概念连接向量。对于帖子X1，句子S1 和 S2 都连接到概念集$\{$C1, C2, C3$\}$中的概念 C1(请参见图 1(b))。因此，它们的句子-概念连接向量都计算为：[1,0,0]。类似地，对于 PostX2, S1 和 S2 的句子-概念连接向量计算为：[1,1,1]
$^{1}$We will release the k-MHDA dataset, along with code to construct it along with the KSAT code for reproducibility of results
|
[
[
0.10455322265625,
0.10797119140625,
0.339111328125,
0.10797119140625,
0.339111328125,
0.121826171875,
0.10455322265625,
0.121826171875
],
[
0.10296630859375,
0.121826171875,
0.9111328125,
0.121826171875,
0.9111328125,
0.1351318359375,
0.10296630859375,
0.1351318359375
],
[
0.0882568359375,
0.136962890625,
0.91015625,
0.136962890625,
0.91015625,
0.147705078125,
0.0882568359375,
0.147705078125
],
[
0.08660888671875,
0.1483154296875,
0.9111328125,
0.149658203125,
0.9111328125,
0.1634521484375,
0.08660888671875,
0.1622314453125
],
[
0.08905029296875,
0.1654052734375,
0.9111328125,
0.1654052734375,
0.9111328125,
0.175537109375,
0.08905029296875,
0.175537109375
],
[
0.0882568359375,
0.177978515625,
0.91259765625,
0.177978515625,
0.91259765625,
0.1912841796875,
0.0882568359375,
0.1912841796875
],
[
0.08905029296875,
0.1932373046875,
0.51953125,
0.1932373046875,
0.51953125,
0.203857421875,
0.08905029296875,
0.203857421875
],
[
0.0849609375,
0.2183837890625,
0.2459716796875,
0.2183837890625,
0.2459716796875,
0.231689453125,
0.0849609375,
0.231689453125
],
[
0.0882568359375,
0.238037109375,
0.2459716796875,
0.238037109375,
0.2459716796875,
0.251953125,
0.0882568359375,
0.251953125
],
[
0.0882568359375,
0.255615234375,
0.9111328125,
0.255615234375,
0.9111328125,
0.26953125,
0.0882568359375,
0.26953125
],
[
0.0882568359375,
0.271484375,
0.91015625,
0.271484375,
0.91015625,
0.2822265625,
0.0882568359375,
0.2822265625
],
[
0.08660888671875,
0.2841796875,
0.91259765625,
0.2841796875,
0.91259765625,
0.29736328125,
0.08660888671875,
0.29736328125
],
[
0.0882568359375,
0.29736328125,
0.9111328125,
0.29736328125,
0.9111328125,
0.311279296875,
0.0882568359375,
0.311279296875
],
[
0.08905029296875,
0.313232421875,
0.91015625,
0.313232421875,
0.91015625,
0.323974609375,
0.08905029296875,
0.323974609375
],
[
0.0882568359375,
0.32568359375,
0.91259765625,
0.32568359375,
0.91259765625,
0.339111328125,
0.0882568359375,
0.339111328125
],
[
0.0849609375,
0.33837890625,
0.54833984375,
0.339111328125,
0.54833984375,
0.352783203125,
0.0849609375,
0.3515625
],
[
0.08660888671875,
0.3662109375,
0.58837890625,
0.3662109375,
0.58837890625,
0.380126953125,
0.08660888671875,
0.380126953125
],
[
0.08660888671875,
0.38330078125,
0.91259765625,
0.3837890625,
0.91259765625,
0.397705078125,
0.08660888671875,
0.396484375
],
[
0.0882568359375,
0.397705078125,
0.91259765625,
0.397705078125,
0.91259765625,
0.410888671875,
0.0882568359375,
0.410888671875
],
[
0.0882568359375,
0.412353515625,
0.91259765625,
0.412353515625,
0.91259765625,
0.425537109375,
0.0882568359375,
0.425537109375
],
[
0.0882568359375,
0.425537109375,
0.90869140625,
0.425537109375,
0.90869140625,
0.439453125,
0.0882568359375,
0.439453125
],
[
0.0882568359375,
0.439453125,
0.91015625,
0.439453125,
0.91015625,
0.45263671875,
0.0882568359375,
0.45263671875
],
[
0.0882568359375,
0.45263671875,
0.9111328125,
0.453857421875,
0.9111328125,
0.46533203125,
0.0882568359375,
0.464111328125
],
[
0.0849609375,
0.46533203125,
0.495849609375,
0.466552734375,
0.495849609375,
0.48095703125,
0.0849609375,
0.479736328125
],
[
0.08416748046875,
0.49365234375,
0.423095703125,
0.49365234375,
0.423095703125,
0.5068359375,
0.08416748046875,
0.5068359375
],
[
0.0882568359375,
0.51220703125,
0.91259765625,
0.51220703125,
0.91259765625,
0.52587890625,
0.0882568359375,
0.52587890625
],
[
0.0882568359375,
0.52587890625,
0.9111328125,
0.52587890625,
0.9111328125,
0.5390625,
0.0882568359375,
0.5390625
],
[
0.08660888671875,
0.5390625,
0.9111328125,
0.5390625,
0.9111328125,
0.55322265625,
0.08660888671875,
0.55322265625
],
[
0.0882568359375,
0.55322265625,
0.91015625,
0.55322265625,
0.91015625,
0.56640625,
0.0882568359375,
0.56640625
],
[
0.0849609375,
0.5673828125,
0.90869140625,
0.5673828125,
0.90869140625,
0.58203125,
0.0849609375,
0.58203125
],
[
0.0849609375,
0.5947265625,
0.25732421875,
0.5947265625,
0.25732421875,
0.60791015625,
0.0849609375,
0.60791015625
],
[
0.08660888671875,
0.6123046875,
0.435546875,
0.6123046875,
0.435546875,
0.62548828125,
0.08660888671875,
0.62548828125
],
[
0.0882568359375,
0.6337890625,
0.9111328125,
0.6337890625,
0.9111328125,
0.64794921875,
0.0882568359375,
0.64794921875
],
[
0.08660888671875,
0.64794921875,
0.81640625,
0.64794921875,
0.81640625,
0.6611328125,
0.08660888671875,
0.6611328125
],
[
0.295654296875,
0.66748046875,
0.701171875,
0.66845703125,
0.701171875,
0.68310546875,
0.295654296875,
0.681640625
],
[
0.896484375,
0.67041015625,
0.9111328125,
0.67041015625,
0.9111328125,
0.68115234375,
0.896484375,
0.68115234375
],
[
0.08660888671875,
0.689453125,
0.91259765625,
0.68994140625,
0.91259765625,
0.705078125,
0.08660888671875,
0.7041015625
],
[
0.08905029296875,
0.70556640625,
0.91015625,
0.70556640625,
0.91015625,
0.71630859375,
0.08905029296875,
0.71630859375
],
[
0.08660888671875,
0.71728515625,
0.9111328125,
0.71728515625,
0.9111328125,
0.73095703125,
0.08660888671875,
0.73095703125
],
[
0.0882568359375,
0.73193359375,
0.9111328125,
0.73193359375,
0.9111328125,
0.74560546875,
0.0882568359375,
0.74560546875
],
[
0.0882568359375,
0.74560546875,
0.91259765625,
0.74560546875,
0.91259765625,
0.7587890625,
0.0882568359375,
0.7587890625
],
[
0.08905029296875,
0.76025390625,
0.253173828125,
0.76025390625,
0.253173828125,
0.77099609375,
0.08905029296875,
0.77099609375
],
[
0.08660888671875,
0.78076171875,
0.9111328125,
0.78076171875,
0.9111328125,
0.79443359375,
0.08660888671875,
0.79443359375
],
[
0.08660888671875,
0.79443359375,
0.91259765625,
0.79443359375,
0.91259765625,
0.8095703125,
0.08660888671875,
0.8095703125
],
[
0.0882568359375,
0.8095703125,
0.51318359375,
0.8095703125,
0.51318359375,
0.82275390625,
0.0882568359375,
0.82275390625
],
[
0.108642578125,
0.83349609375,
0.9111328125,
0.83349609375,
0.9111328125,
0.8466796875,
0.108642578125,
0.8466796875
],
[
0.08660888671875,
0.8466796875,
0.91259765625,
0.8466796875,
0.91259765625,
0.85986328125,
0.08660888671875,
0.85986328125
],
[
0.08660888671875,
0.85986328125,
0.495849609375,
0.85986328125,
0.495849609375,
0.87353515625,
0.08660888671875,
0.87353515625
],
[
0.08660888671875,
0.8818359375,
0.6708984375,
0.8837890625,
0.6708984375,
0.8984375,
0.08660888671875,
0.896484375
]
] |
[
"THMATHX_MAXXINE THMALINE_1",
"XMATHX<NEXT_LINEXMATHISH_。基于自我注意的语言模型（SAMs）从领域不可知的大量训练语料库[？】中捕获更广泛",
"的上下文。在特定于领域的语料库上微调SAM可实现特定于领域的上下文捕获[？？】。然而，SAM架构本质上是黑盒[？]。因此",
"微调的SAM体系结构不适合对开放研究目标进行可靠的评估：（R1）相关的特定领域上下文覆盖范围，以及（R2）知识上下文的",
"影响与下游任务中的数据上下文进行权衡[？？1。我们提出了一种改进的自我注意架构，注入了知识的自我注意转换器（KSAT",
"来解决这些目标。KSAT在选定的特定领域任务上表现良好（请参阅1.2），同时有助于对R1和R2进行可靠的人类可理解的评估。",
"因此，KSAT为促进AI用户信任和满意度[？？】迈出了重要的一步。。",
"1BACKGROUND",
"1.1 Related Work",
"与R1和R2相关并包含多个知识上下文的现有方法可以根据知识注人技术大致分为（1）知识调制的SAM和（2）知识注人的基于",
"输入嵌入的SAM[？？]。前者使用知识来指导SAM中的自我注意机制，后者在将输人传递到SAM之前将知识嵌入到向量空间中。",
"在这里，我们简要总结一下它们对R1和R2的贡献。Category(1)和Category(2)方法的域覆盖率都是通过特定于域的任务描述(R1)",
"的性能来评估的。这些方法的消融突出了知识上下文(R2)的贡献。但是，检查模型组件（投影矩阵和向量）的数值输出并不容易",
"使其易于理解。需要可解释的AI技术（将其转换为人类可理解的信息的数值输出的后处理）来确认作者的观点[？】。基于后处",
"理的解释是对特定输入的SAM推理的局部近似，因此不呈现全局图片，从而对SAM评估的有效性产生怀疑。KSAT提出了一种",
"SAM架构，其数值输出有助于对R1和R2进行可靠的人类可理解的评估",
"1.2TaskDescription,Data,andExternalKnowledgeSources",
"尽管KSAT体系结构广泛适用于任何特定领域的任务，但我们选择了Gaur等人的自杀倾向心理健康诊断辅助的特定任务。[？]。",
"我们将此数据集表示为MHDA。数据包含有关自杀相关子reddits的Reddit帖子的高质量专家注释。注释方法确保了来自测量伪",
"影的最小噪声和专家注释者之间的高度一致性。我们使用哥伦比亚自杀严重程度评定量表（CSSRS）中包含的临床确定的诊断过",
"程信息。图1（a，b）说明了自杀模式可能出现的各种上下文（每个树路径代表一个上下文）。任务是预测自杀模式，即-适应症",
"观念1，观念2和行为或企图。图1(c）显示了来自MHDA数据集的示例，并增加了知识上下文注释。我们将增强的数据集表示",
"为k-MHDA。k-MHDA包含帖子和句子级别的知识上下文注释（请参见图1（c））。我们将k-MHDA的构造细节从CSSRS知识和",
"MHDA数据推迟到附录部分A，因为它不是论文1的主要焦点。",
"2KSAT-PROPOSEDARCHITECTURE",
"对于香草SAM层I中的输入x，引人了CLS令牌，该令牌对下游任务的数据中的信息进行编码。例如，对于分类ZcLS，来自最",
"终层L的CLS令牌的表示通过输出类概率的softmax层传递。在KSAT层中，我们引人了一个ZkCLS令牌，该令牌对该层的知识",
"进行编码。ZkCLs值由输人x句子与输人x的知识上下文注释中的概念之间的图形上下文确定。例如，考虑输人x类似于图1（c）",
"中所示的帖子PostX1和PostX2，然后图1（d）中的图形说明了显示句子概念连接的图形上下文。人们会期望两个帖子中的句子",
"S1和S2的ZkCLs编码将是相似的，因为它们具有相似的图上下文。KGbias项确保ZkCLs中的值捕获此行为（请参见图2（b)）",
"2.1The KSATlayer",
"与香草SAM层相比，每个KSAT层都有以下主要区别：",
"2.1.1LayerParametersandOutputs.让Y表示与每个上下文相关的自杀结果集：（行为指示或无，想法1，。让x表示一个输人帖子",
"（例如帖子，请参见图1（c））。对于每个输人x，每个KSAT层I输出由公式1给出的每个结果yEY的概率向量：",
"Pksat (ylx) = o(WT (αiZkKCLs(x) +(1 -α)ZcLs(x) +KGbias(x))",
"(1)",
"W，ZcLs(x）和ZkCLs(x)的尺寸为384×1。在MHDA语料库上进行微调，用于嵌人输人[?]。回想一下，术语KGbias用于确保",
"ZcLs(x）对表示为图形的知识上下文进行编码（2.1.2中的详细信息）。KGbiαs的维度大小为1×1。标量项αl用作权衡因子，使用",
"凸组合来调制数据和知识上下文之间的贡献（请参见公式1）。我们将KSAT层输出与香草SAM层进行对比，其中类似的概率向",
"量仅在最后一层输出，而不是在每一层输出（见图2）。用于计算自注意矩阵的香草SAM层中的参数（查询，键和值投影矩阵）",
"保留在KSAT层[？］中。与vanillaSAM层不同，在vanillaSAM层中，参数可以跨层共享，也可以不跨层共享，因为不同的层对不",
"同的知识上下文进行编码。",
"2.1.2EncodingGraphContextsinLayers，即使每个KSAT层I输出每个结果的概率向量yE(指示或无，Ideation1，Ideation2，行为",
"或尝试)，表示每个上下文，它在ZkCLs中只编码一个上下文。我们展示了如何通过引用图1（c，d）中的帖子PostX1和PostX2",
"以及图上下文Ideation1和BehaviororAttempt来计算KGbias项",
"ComputingSentence-ConceptConnectionVectors：我们首先为帖子中的每个句子计算一个句子-概念连接向量。对于帖子X1，句子",
"S1和S2都连接到概念集(C1，C2，C3)中的概念C1（请参见图1（b))。因此，它们的句子-概念连接向量都计算为：[1,0,0]。类似",
"地，对于PostX2，S1和S2的句子-概念连接向量计算为：[1,1,1]",
"I'We will release the k-MHDA dataset, along with code to construct it along with the KSAT code for reproducibility of results."
] |
{
"arxiv_id": "2210.04307",
"language": "zh",
"page_idx": 0
}
|
||
2210.04307-zh-1
|
2210.04307-zh-1.jpg
|
$S1$
$S2$
$C1$
$\underline{ldeation1}$
<FigureHere>
Tree Paths Representing
Columbia Suicide Severity Rating
$(\mathfrak{c})\underline{PostX1}$
S1: I don't feel like waking up anymore $\to$ Concept 1 (T)
S2: It wouldn't be the worst thing if I didn't wake up honestly $\to$Concept 1 (T)
Context: Concept 1 $(\mathbb{T})\to\underline{ldeation1}$
$C3$
$\underline{PostX2}$
S1: I could jump off the ledge l am sitting on $\to$ Concept 1 (T), Concept 2 (T), Concept 3 (T) S2: Orl could do it with the gun on my lap $\to$ Concept 1 (T), Concept 2 (T), Concept 3 (T) Context: Concept 1 (T), Concept 2 (T), Concept 3 (T) $\to\underline{Behavior~or~Attempt}$
Sentence-Concept Graph
Contexts
$S2$
$S1$
$C1$
$C2$
Behavior or Attempt
Figure 1: The knowledge source we utilize for our experiments is the Columbia Suicide Severity Rating Scale (CSSRS) and the data are posts from Reddit subreddits related to suicide.(a) Shows the knowledge contexts (tree-path represents a context) present in the CSSRS leading to varying suicidal patterns.( b) Shows the different concepts listed in the CSSRS.( c) Shows the posts and sentences from the data annotated with relevant knowledge concepts and contexts from the CSSRS.( d) Shows how sentences from the input posts ( Post X1 and Post X2 ) and concepts from the CSSRS are graphically connected in different contexts (In this case, the $\underline\text{Ideation1}$ and the Behavior or Attempt contexts).
ComputingGraph Context Distances: 回想一下，KSAT 层将单个帖子作为输入 (等式 1 中的$x)$。将句子 S1 的句子概念向量表示为$C_{S1}$ ,要计算 $KG_{bias}(x)$ ,我们首先需要计算帖子 PostX1 和 PostX2 的图形上下文距离：$d(C_{S1},C_{S2})$。如果帖子有两个以上的句子，则图形上下文距离将包括所有对(例如：三个句子的$d(C_{S1},C_{S2}),d(C_{S2},C_{S3}),d(C_{S1},C_{S3}))$。$x$后的一对句子($S_i,S_j)$的术语$d(C_{Si},C_{Sj})$捕获了句子之间基于图形上下文的距离。直观上，具有等效图上下文的句子应该具有$d(S_i,S_j)=0$ 。我们在实验中使用汉明距离
因此，输入$x$的$KG_bias$项由等式 2 给出为：
(2)
$$KG_{bias}(x)=-\sum_{(S_i,S_j)\in x}\frac{(Z_{KCLS}(x)[S_i]-Z_{KCLS}(x)[S_j])^2}{d(S_i,S_j)+\epsilon}$$
等式 2 中$KG_{bias}(x)$项的公式鼓励在相同图形上下文中的句子的$Z_\mathrm{KCLS}$表示相似。$\varepsilon$术语是为了防止除以零错误。
# 2.2 Aggregating KSAT Layer Outputs
结合由方程 1 给出的 KSAT 层概率是应用程序域相关的。对于集合中的自杀结果：{指示或无，想法 }1,想法 2, 行为或尝试{,可以}
合理地预期自杀意念 (想法 1 和想法 2) 先于企图自杀的行为 (行为或尝试)。在其他应用程序域中，上下文可以彼此独立
在我们的实验中，我们按照结果指示或无，想法1，想法 2, 行为或尝试的顺序堆叠四个 KSAT 层 (该顺序源自图 1 (a) 中的树结构)。通常，对于相关概率结果$X$ 和$Y$,其中$X$在$Y$之前，概率 $P(X,Y)$ 将被建模为 $P(Y\mid X)P(X)$ 。但是，由于我们以特定顺序堆叠 KSAT 层，因此我们使用乘积近似$(P(X,Y)=P(X)P(Y))$,因为信息通过 KSAT 层向上传播。因此，来自 KSAT 层的最终层概率使用公式 3 计算为：
(3)
$$P_{final}^{ksat}(y|x)=\prod_{l}P_{l}^{ksat}(y|x),$$
其中$P_l^{ksat}(y|x)$由公式 1 给出
|
[
[
0.12744140625,
0.11614990234375,
0.279296875,
0.11871337890625,
0.279296875,
0.1319580078125,
0.12744140625,
0.1300048828125
],
[
0.47216796875,
0.117431640625,
0.66845703125,
0.117431640625,
0.66845703125,
0.1307373046875,
0.47216796875,
0.1307373046875
],
[
0.69775390625,
0.117431640625,
0.7197265625,
0.121826171875,
0.7158203125,
0.1337890625,
0.69384765625,
0.1300048828125
],
[
0.73193359375,
0.12060546875,
0.87841796875,
0.12060546875,
0.87841796875,
0.1307373046875,
0.73193359375,
0.1307373046875
],
[
0.439453125,
0.12060546875,
0.461669921875,
0.12060546875,
0.461669921875,
0.1331787109375,
0.439453125,
0.1331787109375
],
[
0.10540771484375,
0.121826171875,
0.1258544921875,
0.121826171875,
0.1258544921875,
0.1319580078125,
0.10540771484375,
0.1319580078125
],
[
0.131591796875,
0.1287841796875,
0.253173828125,
0.1300048828125,
0.253173828125,
0.14453125,
0.131591796875,
0.143310546875
],
[
0.47216796875,
0.1300048828125,
0.625,
0.1300048828125,
0.625,
0.143310546875,
0.47216796875,
0.143310546875
],
[
0.73193359375,
0.1319580078125,
0.7861328125,
0.1319580078125,
0.7861328125,
0.143310546875,
0.73193359375,
0.143310546875
],
[
0.134033203125,
0.143310546875,
0.188720703125,
0.143310546875,
0.188720703125,
0.1546630859375,
0.134033203125,
0.1546630859375
],
[
0.32177734375,
0.149658203125,
0.3447265625,
0.149658203125,
0.3447265625,
0.1622314453125,
0.32177734375,
0.1622314453125
],
[
0.273681640625,
0.1622314453125,
0.310546875,
0.1622314453125,
0.310546875,
0.173583984375,
0.273681640625,
0.173583984375
],
[
0.36767578125,
0.1641845703125,
0.398681640625,
0.1641845703125,
0.398681640625,
0.17236328125,
0.36767578125,
0.17236328125
],
[
0.4794921875,
0.1590576171875,
0.611328125,
0.1590576171875,
0.611328125,
0.169189453125,
0.4794921875,
0.169189453125
],
[
0.74267578125,
0.17041015625,
0.75634765625,
0.17041015625,
0.75634765625,
0.179931640625,
0.74267578125,
0.179931640625
],
[
0.82275390625,
0.1685791015625,
0.8447265625,
0.1685791015625,
0.8447265625,
0.18115234375,
0.82275390625,
0.18115234375
],
[
0.28515625,
0.175537109375,
0.2998046875,
0.175537109375,
0.2998046875,
0.184326171875,
0.28515625,
0.184326171875
],
[
0.2548828125,
0.1900634765625,
0.27783203125,
0.1900634765625,
0.27783203125,
0.20263671875,
0.2548828125,
0.20263671875
],
[
0.326904296875,
0.1925048828125,
0.44677734375,
0.1925048828125,
0.44677734375,
0.205810546875,
0.326904296875,
0.205810546875
],
[
0.47802734375,
0.1956787109375,
0.62841796875,
0.1956787109375,
0.62841796875,
0.208984375,
0.47802734375,
0.208984375
],
[
0.298095703125,
0.20703125,
0.306396484375,
0.20703125,
0.306396484375,
0.212158203125,
0.298095703125,
0.212158203125
],
[
0.4794921875,
0.2083740234375,
0.59619140625,
0.2083740234375,
0.59619140625,
0.2215576171875,
0.4794921875,
0.2215576171875
],
[
0.78173828125,
0.2171630859375,
0.8056640625,
0.2171630859375,
0.8056640625,
0.2291259765625,
0.78173828125,
0.2291259765625
],
[
0.188720703125,
0.231689453125,
0.2115478515625,
0.231689453125,
0.2115478515625,
0.2442626953125,
0.188720703125,
0.2442626953125
],
[
0.281005859375,
0.2298583984375,
0.348876953125,
0.231689453125,
0.348876953125,
0.246826171875,
0.281005859375,
0.2442626953125
],
[
0.8408203125,
0.2374267578125,
0.896484375,
0.23486328125,
0.896484375,
0.246826171875,
0.8408203125,
0.248779296875
],
[
0.2320556640625,
0.248779296875,
0.240234375,
0.248779296875,
0.240234375,
0.25390625,
0.2320556640625,
0.25390625
],
[
0.48193359375,
0.2474365234375,
0.65283203125,
0.2474365234375,
0.65283203125,
0.261474609375,
0.48193359375,
0.261474609375
],
[
0.24267578125,
0.261474609375,
0.2484130859375,
0.261474609375,
0.2484130859375,
0.266357421875,
0.24267578125,
0.266357421875
],
[
0.48046875,
0.261962890625,
0.62353515625,
0.261962890625,
0.62353515625,
0.272705078125,
0.48046875,
0.272705078125
],
[
0.11358642578125,
0.27587890625,
0.1854248046875,
0.27587890625,
0.1854248046875,
0.287353515625,
0.11358642578125,
0.287353515625
],
[
0.2197265625,
0.27587890625,
0.285888671875,
0.27587890625,
0.285888671875,
0.290283203125,
0.2197265625,
0.290283203125
],
[
0.48046875,
0.2734375,
0.640625,
0.2734375,
0.640625,
0.2841796875,
0.48046875,
0.2841796875
],
[
0.125,
0.2890625,
0.176513671875,
0.2890625,
0.176513671875,
0.300537109375,
0.125,
0.300537109375
],
[
0.744140625,
0.31005859375,
0.75634765625,
0.31005859375,
0.75634765625,
0.31640625,
0.744140625,
0.31640625
],
[
0.82421875,
0.309326171875,
0.84228515625,
0.309326171875,
0.84228515625,
0.317626953125,
0.82421875,
0.317626953125
],
[
0.10296630859375,
0.323974609375,
0.176513671875,
0.323974609375,
0.176513671875,
0.335205078125,
0.10296630859375,
0.335205078125
],
[
0.12744140625,
0.333984375,
0.370849609375,
0.335205078125,
0.370849609375,
0.348388671875,
0.12744140625,
0.347900390625
],
[
0.363525390625,
0.333984375,
0.451904296875,
0.3359375,
0.451904296875,
0.347900390625,
0.363525390625,
0.34521484375
],
[
0.1258544921875,
0.34521484375,
0.57275390625,
0.3466796875,
0.57275390625,
0.361083984375,
0.1258544921875,
0.35986328125
],
[
0.78369140625,
0.354736328125,
0.8056640625,
0.354736328125,
0.8056640625,
0.367431640625,
0.78369140625,
0.367431640625
],
[
0.86279296875,
0.354248046875,
0.8837890625,
0.354248046875,
0.8837890625,
0.3662109375,
0.86279296875,
0.3662109375
],
[
0.7060546875,
0.35595703125,
0.7236328125,
0.35595703125,
0.7236328125,
0.3642578125,
0.7060546875,
0.3642578125
],
[
0.12744140625,
0.359130859375,
0.3447265625,
0.35986328125,
0.3447265625,
0.371826171875,
0.12744140625,
0.37060546875
],
[
0.12744140625,
0.3837890625,
0.17724609375,
0.3837890625,
0.17724609375,
0.395751953125,
0.12744140625,
0.395751953125
],
[
0.1282958984375,
0.396484375,
0.66015625,
0.396484375,
0.66015625,
0.4072265625,
0.1282958984375,
0.4072265625
],
[
0.12744140625,
0.406005859375,
0.37744140625,
0.406005859375,
0.37744140625,
0.419189453125,
0.12744140625,
0.419189453125
],
[
0.370849609375,
0.40771484375,
0.64892578125,
0.40771484375,
0.64892578125,
0.41845703125,
0.370849609375,
0.41845703125
],
[
0.7900390625,
0.406005859375,
0.89794921875,
0.406005859375,
0.89794921875,
0.419189453125,
0.7900390625,
0.419189453125
],
[
0.12744140625,
0.41845703125,
0.5908203125,
0.41845703125,
0.5908203125,
0.431884765625,
0.12744140625,
0.431884765625
],
[
0.44287109375,
0.42236328125,
0.451904296875,
0.42236328125,
0.451904296875,
0.4267578125,
0.44287109375,
0.4267578125
],
[
0.0849609375,
0.4609375,
0.9140625,
0.462158203125,
0.9140625,
0.47607421875,
0.0849609375,
0.474853515625
],
[
0.08660888671875,
0.472900390625,
0.9140625,
0.472900390625,
0.9140625,
0.486083984375,
0.08660888671875,
0.486083984375
],
[
0.08660888671875,
0.484130859375,
0.91259765625,
0.484130859375,
0.91259765625,
0.494384765625,
0.08660888671875,
0.494384765625
],
[
0.08660888671875,
0.494384765625,
0.9140625,
0.494384765625,
0.9140625,
0.5048828125,
0.08660888671875,
0.5048828125
],
[
0.08660888671875,
0.50390625,
0.43701171875,
0.50390625,
0.43701171875,
0.5146484375,
0.08660888671875,
0.5146484375
],
[
0.1070556640625,
0.55615234375,
0.91259765625,
0.55615234375,
0.91259765625,
0.5693359375,
0.1070556640625,
0.5693359375
],
[
0.0882568359375,
0.57080078125,
0.91259765625,
0.57080078125,
0.91259765625,
0.583984375,
0.0882568359375,
0.583984375
],
[
0.08660888671875,
0.58203125,
0.91259765625,
0.58349609375,
0.91259765625,
0.59765625,
0.08660888671875,
0.5966796875
],
[
0.0849609375,
0.59765625,
0.91259765625,
0.59765625,
0.91259765625,
0.611328125,
0.0849609375,
0.611328125
],
[
0.08660888671875,
0.6123046875,
0.146240234375,
0.6123046875,
0.146240234375,
0.62353515625,
0.08660888671875,
0.62353515625
],
[
0.10296630859375,
0.62548828125,
0.3857421875,
0.62548828125,
0.3857421875,
0.6396484375,
0.10296630859375,
0.6396484375
],
[
0.468017578125,
0.65283203125,
0.6748046875,
0.65087890625,
0.6748046875,
0.66552734375,
0.468017578125,
0.66748046875
],
[
0.32177734375,
0.6611328125,
0.420654296875,
0.6611328125,
0.420654296875,
0.67529296875,
0.32177734375,
0.67529296875
],
[
0.43701171875,
0.66015625,
0.451904296875,
0.66015625,
0.451904296875,
0.67041015625,
0.43701171875,
0.67041015625
],
[
0.8935546875,
0.6611328125,
0.9140625,
0.6611328125,
0.9140625,
0.67529296875,
0.8935546875,
0.67529296875
],
[
0.5361328125,
0.66845703125,
0.61279296875,
0.66845703125,
0.61279296875,
0.68310546875,
0.5361328125,
0.68310546875
],
[
0.416748046875,
0.67822265625,
0.4697265625,
0.67822265625,
0.4697265625,
0.689453125,
0.416748046875,
0.689453125
],
[
0.0882568359375,
0.70263671875,
0.8095703125,
0.70263671875,
0.8095703125,
0.71728515625,
0.0882568359375,
0.71728515625
],
[
0.0882568359375,
0.736328125,
0.11358642578125,
0.736328125,
0.11358642578125,
0.744140625,
0.0882568359375,
0.744140625
],
[
0.11767578125,
0.7392578125,
0.1298828125,
0.7392578125,
0.1298828125,
0.7421875,
0.11767578125,
0.7421875
],
[
0.1258544921875,
0.734375,
0.40185546875,
0.734375,
0.40185546875,
0.74755859375,
0.1258544921875,
0.74755859375
],
[
0.08660888671875,
0.751953125,
0.9140625,
0.751953125,
0.9140625,
0.76513671875,
0.08660888671875,
0.76513671875
],
[
0.0882568359375,
0.76513671875,
0.86669921875,
0.76513671875,
0.86669921875,
0.77880859375,
0.0882568359375,
0.77880859375
],
[
0.10296630859375,
0.77880859375,
0.91259765625,
0.77880859375,
0.91259765625,
0.79248046875,
0.10296630859375,
0.79248046875
],
[
0.08660888671875,
0.79345703125,
0.91259765625,
0.79345703125,
0.91259765625,
0.806640625,
0.08660888671875,
0.806640625
],
[
0.08660888671875,
0.806640625,
0.91259765625,
0.806640625,
0.91259765625,
0.82080078125,
0.08660888671875,
0.82080078125
],
[
0.0882568359375,
0.82275390625,
0.2418212890625,
0.82275390625,
0.2418212890625,
0.83349609375,
0.0882568359375,
0.83349609375
],
[
0.410888671875,
0.8447265625,
0.49267578125,
0.83837890625,
0.495849609375,
0.859375,
0.414306640625,
0.86572265625
],
[
0.5087890625,
0.84033203125,
0.58642578125,
0.84228515625,
0.58642578125,
0.85986328125,
0.5087890625,
0.85791015625
],
[
0.419189453125,
0.8466796875,
0.44287109375,
0.84521484375,
0.44287109375,
0.849609375,
0.419189453125,
0.85107421875
],
[
0.89453125,
0.8447265625,
0.9150390625,
0.8447265625,
0.9150390625,
0.85791015625,
0.89453125,
0.85791015625
],
[
0.0849609375,
0.8798828125,
0.27685546875,
0.8798828125,
0.27685546875,
0.89794921875,
0.0849609375,
0.89794921875
]
] |
[
"TreePaths Representing",
"Columbia Suicide Severity Rating",
"(d)",
"Sentence-ConceptGraph",
"(b)",
"(a)",
"Different Knowledge",
"Scale (CSSRS) Concepts",
"Contexts",
"Contexts",
"C1",
"True",
"Faise",
"C1:WishtobeDead",
"S",
"S2",
"K",
"C2",
"Indication orNone",
"C2:Non-SpecificActive",
"口",
"Suicidal Thoughts",
"C1",
"C3",
"Ideation1",
"Ideation1",
"F",
"C3:Active Suicidal ldeation",
"A",
"withAnyMethods(Not",
"Behavioror",
"Ideation2",
"Plan)without IntenttoAct",
"Attempt",
"S",
"S2",
"(c) Post X1",
"s1:Idon'tfeel like waking up anymore-",
"Concept1(T)",
"S2: It wouldn't be the worst thing if I didn't wake up honestly - Concept 1 (T)",
"C1",
"C2",
"C3",
"Context:Concept1(T)-→ldeation1",
"PostX2",
"S1:Icould jumpofftheledgeIamsittingonConcept1(T),Concept2(T),Concept3(T)",
"S2: OrI could do it with the gun on my lap",
"Concept1(T),Concept2(T),Concept3(T)",
"Behavior or Attempt",
"Context:Concept1 (T),Concept2 (T),Concept3(T)-BehaviororAttempt",
"",
"Figure l: The knowledge source we utilize for our experiments is the Columbia Suicide Severity Rating Scale (CSSRS) and the data are posts from Reddit",
"subreddits related to suicide. (a ) Shows the knowledge contexts (tree-path represents a context) present in the CSSRs leading to varying suicidal patterns. (b)",
"Shows thedifferent concepts listed in the CsSRS. (c)Shows theposts and sentences from the data annotated with relevant knowledge concepts and contexts from",
"the CSSRS.(d)Shows how sentences from the input posts (Post X1 and Post X2)and concepts from the CSSRS are graphically connected in different contexts",
"(In this case,theIdeationl and the BehaviororAttempt contexts),",
"ComputingGraphContextDistances：回想一下，KSAT层将单个帖子作为输人（等式1中的x）。将句子S1的句子概念向量表示",
"为Cs1，要计算KGbias（x），我们首先需要计算帖子PostX1和PostX2的图形上下文距离：d(Cs1,Cs2）。如果帖子有两个以上的句",
"子，则图形上下文距离将包括所有对（例如：三个句子的d(Cs1,Cs2），d(Cs2,Cs3），d(Cs1,Cs3））。x后的一对句子（S,S,）的术语",
"d(Csi.Csi）捕获了句子之间基于图形上下文的距离。直观上，具有等效图上下文的句子应该具有d(Si,S,）=0。我们在实验中使用",
"汉明距离",
"因此，输人x的KGbias项由等式2给出为：",
"(ZKCLs(x)[Si] -ZKCLs(x)[S,l)2",
"KGbias(x) = --",
"",
"(2)",
"d(Si,S) +e",
"(Si,S,)Ex",
"等式2中KGbias(x）项的公式鼓励在相同图形上下文中的句子的ZkCLs表示相似。ε术语是为了防止除以零错误。",
"2.2",
"",
"AggregatingKSATLayerOutputs",
"结合由方程1给出的KSAT层概率是应用程序域相关的。对于集合中的自杀结果：（指示或无，想法1，想法2，行为或尝试)，可以",
"合理地预期自杀意念（想法1和想法2）先于企图自杀的行为（行为或尝试）。在其他应用程序域中，上下文可以彼被此独立",
"在我们的实验中，我们按照结果指示或无，想法1，想法2，行为或尝试的顺序堆叠四个KSAT层（该顺序源自图1（a）中的树",
"结构)。通常，对于相关概率结果X和Y，其中X在Y之前，概率P(X,Y)将被建模为P(YIX)P(X)。但是，由于我们以特定顺",
"序堆叠KSAT层，因此我们使用乘积近似（P(X,Y）=P(X)P(Y）），因为信息通过KSAT层向上传播。因此，来自KSAT层的最终层",
"概率使用公式3计算为：",
"Pkinal (ylx) =",
"Ipksat (ylx),",
"oksa",
"(3)",
"其中Pksat(ylx)由公式1给出"
] |
{
"arxiv_id": "2210.04307",
"language": "zh",
"page_idx": 1
}
|
||
2210.04307-zh-3
|
2210.04307-zh-3.jpg
|
<FigureHere>
Figure 3: Shows that the first two Behavior or Attempt posts in the list of posts are far off based on the $Z_{CLS}$ representations( false ), whereas KSAT represen
tations depict them as closer i.e d(.)<ε( true ). The $\alpha$ shows the data and knowledge contributions for the samples.
# $A$ $CONSTRUCTING$ K-MHDA
MHDA 数据集中有 500 Reddit 帖子。CSSRS 中的知识上下文可以用树表示 (见图 1(d))。我们可以构造一个概率决策树，它接受$x$
后的输入，并从叶子中输出结果$y$。我们可以用代数形式写出树，如方程 4 所示。
$$P(y\mid x,\{\theta_i\})=\sum_{y\in Outcomes}p_y\prod_{i=1}^3\sum_{x_{sub}\in x}\left(cos_-sim\biggl(x_{sub}^R,q_i^R\biggr)\geq\theta_i\right)\geq0.5$$
(4)
$p_y$是每个结果的真实概率。索引$i$遍历图 1(d)中的 3 概念。$x_{sub}$表示输人帖子的一个子片段(1 句子、2 句子等。)。$q_i$表示$i$索引的 3 概念中的概念文本。$x_{sub}^\mathrm{R}$和 $q_i^\mathrm{R}$ 是使用 Reimers 等人发布的句子转换器的帖子子片段和概念文本的表示。[?]。方程 5 确定在后子片段$x_{sub}$中是否存在概念$q_i$ 。首先，我们计算它们的句子转换表示$x_{sub}^R$和$q_i^R$之间的余弦相似性。如果结果值为$\geq\theta_i$ ,则确定$x_sub$中存在概念$q_i$ ,否则确定$x_sub$中不存在概念$q_i$ 。
方程 4 中的$\sum_x_{sub}\in x(\cdot)\geq0.5$是 V 运算的代数形式，因为我们确定概念$q_i$存在于后$x$中，如果任何后片段$x_sub\in x$显示存在概
念$y$。
$$\left(cos_sim\biggl(x_{sub}^R,q_i^R\biggr)\geq\theta_i\right)$$
(5)
(6)
然后，我们可以在给定输入，结果对$(x,y)$和参数$\{\theta_i\}$的情况下评估伯努利损失$\mathcal{L}:$
$$\mathcal{L}(x,y,\{\theta_{i}\})=P(y\mid x,\{\theta_{i}\})log(P(y\mid x,\{\theta_{i}\}))+\\(1-P(y\mid x,\{\theta_{i}\}))log(1-P(y\mid x,\{\theta_{i}\}))$$
我们使用网格搜索来查找参数 $\left\{\theta_i\right\}$的配置，并发布具有最大值的子片段 $x_{sub}$
$\prod_{(x,y)\in MHDA}\mathcal{L}(x,y,\{\theta_i\})$ 。我们在-1到 1 (余弦函数的范围) 的范围内改变每个单独的$\theta _i$ , 并且$x_{sub}$从集合$\{1,2,3\}$中取值
推理是在决策树分类器中进行的，每个分支的概念存在或不存在，使用公式 5 进行评估
|
[
[
0.58837890625,
0.2052001953125,
0.81884765625,
0.205810546875,
0.81884765625,
0.2197265625,
0.58837890625,
0.2183837890625
],
[
0.56884765625,
0.2139892578125,
0.580078125,
0.21337890625,
0.580078125,
0.2171630859375,
0.56884765625,
0.2183837890625
],
[
0.419189453125,
0.2215576171875,
0.50244140625,
0.2215576171875,
0.50244140625,
0.236083984375,
0.419189453125,
0.236083984375
],
[
0.5888671875,
0.2197265625,
0.7685546875,
0.2197265625,
0.7685546875,
0.23291015625,
0.5888671875,
0.23291015625
],
[
0.418212890625,
0.238037109375,
0.50244140625,
0.238037109375,
0.50244140625,
0.255615234375,
0.418212890625,
0.255615234375
],
[
0.56884765625,
0.251953125,
0.5849609375,
0.251953125,
0.5849609375,
0.2587890625,
0.56884765625,
0.2587890625
],
[
0.5908203125,
0.246826171875,
0.830078125,
0.246826171875,
0.830078125,
0.2568359375,
0.5908203125,
0.2568359375
],
[
0.396240234375,
0.25830078125,
0.5009765625,
0.25830078125,
0.5009765625,
0.272705078125,
0.396240234375,
0.272705078125
],
[
0.583984375,
0.25390625,
0.5908203125,
0.248779296875,
0.5947265625,
0.251953125,
0.58837890625,
0.2568359375
],
[
0.59228515625,
0.2587890625,
0.6650390625,
0.2587890625,
0.6650390625,
0.26953125,
0.59228515625,
0.26953125
],
[
0.5888671875,
0.270751953125,
0.73193359375,
0.270751953125,
0.73193359375,
0.2841796875,
0.5888671875,
0.2841796875
],
[
0.56884765625,
0.3017578125,
0.58837890625,
0.304931640625,
0.58642578125,
0.313232421875,
0.56787109375,
0.311279296875
],
[
0.5810546875,
0.296630859375,
0.8349609375,
0.296630859375,
0.8349609375,
0.31005859375,
0.5810546875,
0.31005859375
],
[
0.59326171875,
0.309326171875,
0.83837890625,
0.309326171875,
0.83837890625,
0.322509765625,
0.59326171875,
0.322509765625
],
[
0.59228515625,
0.32080078125,
0.8408203125,
0.32080078125,
0.8408203125,
0.333984375,
0.59228515625,
0.333984375
],
[
0.5908203125,
0.3359375,
0.64208984375,
0.3359375,
0.64208984375,
0.347900390625,
0.5908203125,
0.347900390625
],
[
0.58837890625,
0.35986328125,
0.810546875,
0.361083984375,
0.810546875,
0.375,
0.58837890625,
0.373779296875
],
[
0.5703125,
0.36865234375,
0.58251953125,
0.36865234375,
0.58251953125,
0.375,
0.5703125,
0.375
],
[
0.5888671875,
0.37255859375,
0.79736328125,
0.373779296875,
0.79736328125,
0.38818359375,
0.5888671875,
0.386962890625
],
[
0.5908203125,
0.385009765625,
0.6435546875,
0.386962890625,
0.64208984375,
0.401611328125,
0.5888671875,
0.399658203125
],
[
0.08660888671875,
0.5341796875,
0.91259765625,
0.5341796875,
0.91259765625,
0.54736328125,
0.08660888671875,
0.54736328125
],
[
0.0882568359375,
0.54736328125,
0.67626953125,
0.54736328125,
0.67626953125,
0.5546875,
0.0882568359375,
0.5546875
],
[
0.0882568359375,
0.58203125,
0.1217041015625,
0.58203125,
0.1217041015625,
0.59326171875,
0.0882568359375,
0.59326171875
],
[
0.11767578125,
0.58056640625,
0.3447265625,
0.58056640625,
0.3447265625,
0.5947265625,
0.11767578125,
0.5947265625
],
[
0.0849609375,
0.5986328125,
0.91259765625,
0.5986328125,
0.91259765625,
0.6123046875,
0.0849609375,
0.6123046875
],
[
0.08660888671875,
0.61279296875,
0.60546875,
0.61279296875,
0.60546875,
0.626953125,
0.08660888671875,
0.626953125
],
[
0.324462890625,
0.64013671875,
0.400390625,
0.64013671875,
0.400390625,
0.6513671875,
0.324462890625,
0.6513671875
],
[
0.57666015625,
0.6416015625,
0.646484375,
0.6416015625,
0.646484375,
0.6484375,
0.57666015625,
0.6484375
],
[
0.6435546875,
0.6416015625,
0.6689453125,
0.6416015625,
0.6689453125,
0.6494140625,
0.6435546875,
0.6494140625
],
[
0.900390625,
0.6416015625,
0.91259765625,
0.6416015625,
0.91259765625,
0.6513671875,
0.900390625,
0.6513671875
],
[
0.53173828125,
0.642578125,
0.57177734375,
0.642578125,
0.57177734375,
0.6494140625,
0.53173828125,
0.6494140625
],
[
0.45751953125,
0.64306640625,
0.4794921875,
0.64306640625,
0.4794921875,
0.6513671875,
0.45751953125,
0.6513671875
],
[
0.496826171875,
0.65478515625,
0.51123046875,
0.65283203125,
0.51318359375,
0.6572265625,
0.498291015625,
0.6591796875
],
[
0.08660888671875,
0.66748046875,
0.9111328125,
0.66748046875,
0.9111328125,
0.681640625,
0.08660888671875,
0.681640625
],
[
0.0849609375,
0.68115234375,
0.91259765625,
0.68115234375,
0.91259765625,
0.69873046875,
0.0849609375,
0.69873046875
],
[
0.0849609375,
0.69580078125,
0.9140625,
0.69775390625,
0.9140625,
0.71533203125,
0.0849609375,
0.71337890625
],
[
0.08660888671875,
0.71337890625,
0.44287109375,
0.71337890625,
0.44287109375,
0.72802734375,
0.08660888671875,
0.72802734375
],
[
0.10455322265625,
0.72900390625,
0.91259765625,
0.72900390625,
0.91259765625,
0.74365234375,
0.10455322265625,
0.74365234375
],
[
0.0849609375,
0.7421875,
0.1258544921875,
0.7421875,
0.1258544921875,
0.7568359375,
0.0849609375,
0.7568359375
],
[
0.431396484375,
0.767578125,
0.48779296875,
0.767578125,
0.48779296875,
0.7822265625,
0.431396484375,
0.7822265625
],
[
0.8935546875,
0.76806640625,
0.9140625,
0.76806640625,
0.9140625,
0.7822265625,
0.8935546875,
0.7822265625
],
[
0.08660888671875,
0.796875,
0.6259765625,
0.796875,
0.6259765625,
0.81005859375,
0.08660888671875,
0.81005859375
],
[
0.35693359375,
0.81494140625,
0.6435546875,
0.81494140625,
0.6435546875,
0.8291015625,
0.35693359375,
0.8291015625
],
[
0.8935546875,
0.82568359375,
0.9140625,
0.82568359375,
0.9140625,
0.8388671875,
0.8935546875,
0.8388671875
],
[
0.39794921875,
0.83349609375,
0.64208984375,
0.83349609375,
0.64208984375,
0.8466796875,
0.39794921875,
0.8466796875
],
[
0.08660888671875,
0.85400390625,
0.57275390625,
0.85400390625,
0.57275390625,
0.8671875,
0.08660888671875,
0.8671875
],
[
0.101318359375,
0.8671875,
0.91259765625,
0.8642578125,
0.91259765625,
0.8818359375,
0.101318359375,
0.88525390625
],
[
0.10296630859375,
0.88330078125,
0.6611328125,
0.88330078125,
0.6611328125,
0.896484375,
0.10296630859375,
0.896484375
]
] |
[
"Ihave lost thewill to live honestly..1",
"",
"d(.)=0",
"cant...I justcant..Ideation2",
"d(.)>0",
"O",
"Contemplatingdowningthisentirejack",
"a1-a",
"",
".lifeissht",
".Behavior orAttempt",
"®",
"Icould jumpoftheledgeIamsittingon",
".OrIcoulddo itwiththegunonmylap",
". haha.. maybeI justdid..Behavioror",
"Attempt",
"I have taken like5benzos..fingers",
"",
"crossed Idont wake! Behavior or",
"Attempt",
"Figure 3: Shows that the first two Behavior or Attempt posts in the list of posts are far off based on the Zcs representations (false ), whereas KSAT represen",
"tationsdepictthemascloseri.ed(.)<e(true).Theα showsthedataandknowledgecontributionsforthesamples",
"A",
" CONSTRUCTING K-MHDA",
"MHDA数据集中有500Reddit帖子。CSSRS中的知识上下文可以用树表示（见图1(d)）。我们可以构造一个概率决策树，它接受x",
"后的输入，并从叶子中输出结果y。我们可以用代数形式写出树，如方程4所示。",
"P(y / x, (0)) =",
"RaR>A",
"≥0.",
"(4)",
"cos sim",
"Py",
"ee",
"Py是每个结果的真实概率。索引i遍历图1(d)中的3概念。xsub表示输入帖子的一个子片段（1句子、2句子等。）。qi表示i索引",
"的3概念中的概念文本。xRub和qR是使用Reimers等人发布的句子转换器的帖子子片段和概念文本的表示。[?]。方程5确定在",
"后子片段xsub中是否存在概念qi。首先，我们计算它们的句子转换表示xR,和之间的余弦相似性。如果结果值为≥Qi，则确",
"定xsub中存在概念qi，否则确定xsub中不存在概念qi",
"方程4中的ZxsubEx（.）≥0.5是V运算的代数形式，因为我们确定概念qi存在于后x中，如果任何后片段xsubEx显示存在概",
"念y。",
"cos_sim",
"(5)",
"然后，我们可以在给定输人，结果对（x,y)和参数(0i)的情况下评估伯努利损失L：",
"L(x, y, (0;) = P(y I x, (0))log(P(y / x, (0;))+",
"(6)",
"(1 - P(y /x, (0;)log(1 - P(y / x, {0)))",
"我们使用网格搜索来查找参数（0）的配置，并发布具有最大值的子片段xsub",
"I(x.y)eMHDAL(x,y,(0;))。我们在-1到1（余弦函数的范围）的范围内改变每个单独的i，并且xsub从集合(1,2,3)中取值",
"推理是在决策树分类器中进行的，每个分支的概念存在或不存在，使用公式5进行评估"
] |
{
"arxiv_id": "2210.04307",
"language": "zh",
"page_idx": 3
}
|
||
1711.05415-zh-0
|
1711.05415-zh-0.jpg
|
DNA-GAN: 从多属性图像中学习解纠缠表示
Taihong Xiao, Jiapeng Hong & Jinwen Ma
Department of Information Science, School of Mathematical Sciences and LMAM
Peking University, Beijing, 100871, China
$\{$ xiaotaihong, jphong $\}$ @pku.edu.cn jwma@math.pku.edu.cn
## Abstract
Disentangling factors of variation has become a very challenging problem on representation learning. Existing algorithms suffer from many limitations, such as unpredictable disentangling factors, poor quality of generated images from encodings, lack of identity information, etc. In this paper, we propose a supervised learning model called DNA-GAN which tries to disentangle different factors or attributes of images. The latent representations of images are DNA-like, in which each individual piece (of the encoding) represents an independent factor of the variation. By annihilating the re- $\dot{\text{cessive piece and swapping a certain piece of one latent representation with}}$ that of the other one, we obtain two different representations which could be decoded into two kinds of images with the existence of the corresponding attribute being changed. In order to obtain realistic images and also disentangled representations, we further introduce the discriminator for adversarial training. Experiments on Multi-PIE and CelebA datasets finally demonstrate that our proposed method is effective for factors disentangling and even overcome certain limitations of the existing methods.
1 Introduction
机器学习 (ML) 算法的性能取决于数据表示形式，因为不同的表示形式可能会在数据背后纠缠不同的解释因素。尽管先验知识可以帮助我们设计表示，但由于特征工程是劳动密集型的并且需要领域专家知识，因此无法满足各种 AI 领域中 ML 算法的巨大需求。因此，可以自动学习数据的良好表示的 ML 算法肯定会使人们在构建分类器或预测器时更容易提取有用的信息
在 Bengio et al. (2013) 中讨论的学习良好表示的所有标准中，解开变化因素是一个重要的因素，有助于分离各种解释因素。例如，给定人脸图像，我们可以获得有关该人的各种面部特征，包括性别，发型，面部表情，带/不带眼镜等。然而，训练一个能够处理纠缠在单个图像中的不同面部特征或属性的分类器是相当困难的。如果我们可以获得面部图像的解开表示，我们可以为多个属性训练单个分类器
在本文中，我们提出了一种称为 DNA-GAN 的监督方法来获取图像的解纠缠表示。DNAGAN 的想法是由 DNA 双螺旋结构驱动的，其中不同种类的性状分别编码在不同的 DNA 片段中。我们做出类似的假设，即图像中的不同视觉属性由其潜在表示中的不同编码片段控制。在 DNA-GAN 中，使用编码器将图像编码为属性相关部分和属性无关部分，其中属性相关部分中的不同片段编码不同属性的信息，属性无关部分编码其他信息。例如，给定一个面部图像，我们试图获得一个潜在的表示，每个单独的部分控制不同的属性，如发型，性别，表情等。这些属性应编码为潜在表示中与属性相关的部分，而其他信息(例如背景)应编码为与属性无关的部分。通过消灭隐性片段并交换某些片段，我们可以获得一些可以解码成新图像的新颖杂交。借助对抗性鉴别器丢失和重建丢失，DNA-GAN 可以重建输人图像并生成具有新属性的新图像。通过迭代训练过程，每个属性都逐渐与其他属性分离。最后，我们能够从潜在表示中获得解开的表示
我们的工作贡献总结如下：
1
|
[
[
0.1748046875,
0.10418701171875,
0.72216796875,
0.1060791015625,
0.72216796875,
0.12371826171875,
0.1748046875,
0.121826171875
],
[
0.1846923828125,
0.15283203125,
0.490234375,
0.15283203125,
0.490234375,
0.1666259765625,
0.1846923828125,
0.1666259765625
],
[
0.1854248046875,
0.1673583984375,
0.76708984375,
0.1673583984375,
0.76708984375,
0.18115234375,
0.1854248046875,
0.18115234375
],
[
0.1846923828125,
0.18115234375,
0.48193359375,
0.18115234375,
0.48193359375,
0.1944580078125,
0.1846923828125,
0.1944580078125
],
[
0.188720703125,
0.1944580078125,
0.646484375,
0.1932373046875,
0.646484375,
0.2083740234375,
0.188720703125,
0.208984375
],
[
0.461669921875,
0.24365234375,
0.53759765625,
0.24560546875,
0.53759765625,
0.261474609375,
0.461669921875,
0.2587890625
],
[
0.233642578125,
0.285888671875,
0.76123046875,
0.285888671875,
0.76123046875,
0.2998046875,
0.233642578125,
0.2998046875
],
[
0.2344970703125,
0.300537109375,
0.76123046875,
0.300537109375,
0.76123046875,
0.314453125,
0.2344970703125,
0.314453125
],
[
0.2320556640625,
0.313232421875,
0.765625,
0.314453125,
0.765625,
0.328857421875,
0.2320556640625,
0.32763671875
],
[
0.2320556640625,
0.32763671875,
0.765625,
0.328857421875,
0.765625,
0.34228515625,
0.2320556640625,
0.341552734375
],
[
0.2344970703125,
0.34228515625,
0.76123046875,
0.34228515625,
0.76123046875,
0.35595703125,
0.2344970703125,
0.35595703125
],
[
0.233642578125,
0.35595703125,
0.7646484375,
0.35595703125,
0.7646484375,
0.369384765625,
0.233642578125,
0.369384765625
],
[
0.233642578125,
0.369384765625,
0.7646484375,
0.36865234375,
0.7646484375,
0.38330078125,
0.233642578125,
0.3837890625
],
[
0.2344970703125,
0.3837890625,
0.76123046875,
0.3837890625,
0.76123046875,
0.397705078125,
0.2344970703125,
0.397705078125
],
[
0.233642578125,
0.397705078125,
0.7646484375,
0.397705078125,
0.7646484375,
0.412353515625,
0.233642578125,
0.412353515625
],
[
0.2344970703125,
0.41357421875,
0.7646484375,
0.41357421875,
0.7646484375,
0.423583984375,
0.2344970703125,
0.423583984375
],
[
0.233642578125,
0.425537109375,
0.7607421875,
0.425537109375,
0.7607421875,
0.439453125,
0.233642578125,
0.439453125
],
[
0.233642578125,
0.439453125,
0.7646484375,
0.439453125,
0.7646484375,
0.45263671875,
0.233642578125,
0.45263671875
],
[
0.2344970703125,
0.453857421875,
0.76123046875,
0.453857421875,
0.76123046875,
0.46728515625,
0.2344970703125,
0.46728515625
],
[
0.2344970703125,
0.46728515625,
0.76318359375,
0.46728515625,
0.76318359375,
0.48095703125,
0.2344970703125,
0.48095703125
],
[
0.233642578125,
0.48095703125,
0.765625,
0.48095703125,
0.765625,
0.494384765625,
0.233642578125,
0.494384765625
],
[
0.233642578125,
0.494384765625,
0.68115234375,
0.494384765625,
0.68115234375,
0.50830078125,
0.233642578125,
0.50830078125
],
[
0.17236328125,
0.556640625,
0.31787109375,
0.5595703125,
0.31787109375,
0.57373046875,
0.17236328125,
0.5712890625
],
[
0.176513671875,
0.59765625,
0.82275390625,
0.59765625,
0.82275390625,
0.611328125,
0.176513671875,
0.611328125
],
[
0.176513671875,
0.6123046875,
0.82177734375,
0.6123046875,
0.82177734375,
0.62548828125,
0.176513671875,
0.62548828125
],
[
0.176513671875,
0.62548828125,
0.82177734375,
0.62548828125,
0.82177734375,
0.6396484375,
0.176513671875,
0.6396484375
],
[
0.1748046875,
0.63818359375,
0.82177734375,
0.63818359375,
0.82177734375,
0.6513671875,
0.1748046875,
0.6513671875
],
[
0.17724609375,
0.65478515625,
0.256591796875,
0.65478515625,
0.256591796875,
0.66552734375,
0.17724609375,
0.66552734375
],
[
0.1748046875,
0.673828125,
0.8203125,
0.673828125,
0.8203125,
0.6875,
0.1748046875,
0.6875
],
[
0.176513671875,
0.68798828125,
0.82275390625,
0.68798828125,
0.82275390625,
0.7021484375,
0.176513671875,
0.7021484375
],
[
0.176513671875,
0.7021484375,
0.82275390625,
0.7021484375,
0.82275390625,
0.71533203125,
0.176513671875,
0.71533203125
],
[
0.176513671875,
0.71630859375,
0.82275390625,
0.71630859375,
0.82275390625,
0.72998046875,
0.176513671875,
0.72998046875
],
[
0.1748046875,
0.73095703125,
0.486083984375,
0.73095703125,
0.486083984375,
0.744140625,
0.1748046875,
0.744140625
],
[
0.176513671875,
0.751953125,
0.8203125,
0.751953125,
0.8203125,
0.76513671875,
0.176513671875,
0.76513671875
],
[
0.176513671875,
0.76513671875,
0.82275390625,
0.76513671875,
0.82275390625,
0.77880859375,
0.176513671875,
0.77880859375
],
[
0.17724609375,
0.77978515625,
0.82275390625,
0.77978515625,
0.82275390625,
0.79345703125,
0.17724609375,
0.79345703125
],
[
0.1732177734375,
0.79150390625,
0.82275390625,
0.79248046875,
0.82275390625,
0.806640625,
0.1732177734375,
0.80615234375
],
[
0.17724609375,
0.80810546875,
0.8095703125,
0.80810546875,
0.8095703125,
0.818359375,
0.17724609375,
0.818359375
],
[
0.176513671875,
0.82080078125,
0.82275390625,
0.82080078125,
0.82275390625,
0.833984375,
0.176513671875,
0.833984375
],
[
0.176513671875,
0.833984375,
0.82275390625,
0.833984375,
0.82275390625,
0.84765625,
0.176513671875,
0.84765625
],
[
0.176513671875,
0.84765625,
0.82275390625,
0.84765625,
0.82275390625,
0.861328125,
0.176513671875,
0.861328125
],
[
0.176513671875,
0.8623046875,
0.82177734375,
0.8623046875,
0.82177734375,
0.87548828125,
0.176513671875,
0.87548828125
],
[
0.17724609375,
0.87744140625,
0.82275390625,
0.87744140625,
0.82275390625,
0.88818359375,
0.17724609375,
0.88818359375
],
[
0.176513671875,
0.89013671875,
0.486083984375,
0.89013671875,
0.486083984375,
0.90380859375,
0.176513671875,
0.90380859375
],
[
0.17724609375,
0.91357421875,
0.359375,
0.91357421875,
0.359375,
0.923828125,
0.17724609375,
0.923828125
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"DNA-GAN：从多属性图像中学习解纠缠表示",
"Taihong Xiao, Jiapeng Hong & Jinwen Ma",
"Department of Information Science, School of Mathematical Sciences and LMAM",
"Peking University, Beijing, 100871, China",
"[xiaotaihong,jphong@pku.edu.cnjwma@math.pku.edu.cn",
"Abstract",
"Disentangling factors of variation has become a very challenging problem",
"on representation learning. Existing algorithms suffer from many limita.",
"tions, such as unpredictable disentangling factors, poor quality of generated",
"images from encodings, lack of identity information,etc.In this paper, we",
"propose a supervised learning model called DNA-GAN which tries to disen",
"tangle different factors or attributes of images. The latent representations",
"of images are DNA-like, in which each individual piece (of the encoding)",
"represents an independent factor of the variation. By annihilating the re-",
"cessive piece and swapping a certain piece of one latent representation with",
"that oftheother one,we obtain two differentrepresentations which could",
"be decoded into two kinds of images with the existence of the correspond",
"ing attribute being changed. In order to obtain realistic images and also",
"disentangled representations, we further introduce the discriminator for ad.",
"versarial training. Experiments on Multi-PIE and CelebA datasets finally",
"demonstrate that ourproposed method is effective forfactors disentangling",
"and even overcome certain limitations of the existing methods.",
"1Introduction",
"机器学习（ML）算法的性能取决于数据表示形式，因为不同的表示形式可能会在数据背后",
"纠缠不同的解释因素。尽管先验知识可以帮助我们设计表示，但由于特征工程是劳动密集",
"型的并且需要领域专家知识，因此无法满足各种AI领域中ML算法的巨大需求。因此，可",
"以自动学习数据的良好表示的ML算法肯定会使人们在构建分类器或预测器时更容易提取",
"有用的信息",
"在Bengioetal.（2013）中讨论的学习良好表示的所有标准中，解开变化因素是一个重要的",
"因素，有助于分离各种解释因素。例如，给定人脸图像，我们可以获得有关该人的各种面部",
"特征，包括性别，发型，面部表情，带不带眼镜等。然而，训练一个能够处理纠缠在单个",
"图像中的不同面部特征或属性的分类器是相当困难的。如果我们可以获得面部图像的解开",
"表示，我们可以为多个属性训练单个分类器",
"在本文中，我们提出了一种称为DNA-GAN的监督方法来获取图像的解纠缠表示。DNA",
"GAN的想法是由DNA双螺旋结构驱动的，其中不同种类的性状分别编码在不同的DNA",
"片段中。我们做出类似的假设，即图像中的不同视觉属性由其潜在表示中的不同编码片段",
"控制。在DNA-GAN中，使用编码器将图像编码为属性相关部分和属性无关部分，其中属",
"性相关部分中的不同片段编码不同属性的信息，属性无关部分编码其他信息。例如，给定",
"个面部图像，我们试图获得一个潜在的表示，每个单独的部分控制不同的属性，如发型，性",
"别，表情等。这些属性应编码为潜在表示中与属性相关的部分，而其他信息（例如背景）应",
"编码为与属性无关的部分。通过消灭隐性片段并交换某些片段，我们可以获得一些可以解",
"码成新图像的新颖杂交。借助对抗性鉴别器丢失和重建丢失，DNA-GAN可以重建输人图",
"像并生成具有新属性的新图像。通过送代训练过程，每个属性都遂渐与其他属性分离。最",
"后，我们能够从潜在表示中获得解开的表示",
"我们的工作贡献总结如下",
"1"
] |
{
"arxiv_id": "1711.05415",
"language": "zh",
"page_idx": 0
}
|
||
1711.05415-zh-1
|
1711.05415-zh-1.jpg
|
1. We propose a supervised algorithm called DNA-GAN that is able to disentangle
multiple attributes as demonstrated by the experiments of interpolating multiple $\bar{\text{attributes on Multi-PIE }}($Gross et al., $\dot{2} 010)$and CelebA ( Liu et al. , 2015) datasets.
2. We introduce the annihilating operation that prevents from trivial solutions: the
attribute-relevant part encodes information of the whole image instead of a certain
attribute.
3. We employ an iterative training method to address the problem of unbalanced
multi-attribute image data, which was theoretically proved to be more efficient
than random image pairs.
2 Related Work
传统的表示学习算法主要集中在 (1)概率图模型，其特征在于受限玻尔兹曼机(RBM)$\boxed{\mathrm{E}}$ (Smolensky,1986) ,自动编码器( AE) 及其变体；(2)流形学习和几何方法，例如主成分分析(PCA)E(Pearson,1901) ,局部线性嵌人 (LLE)$\mathbb{E}($Roweis $\&$ Saul,2000) ,THMAXXXXCC 等。然而，最近的研究积极地集中在开发学习表示数据分布的深度概率模型上。。实际上， Kingma & Welling (2013)采用了显式模型分布，并利用变分推理来学习其参数。随着生成对抗网络 (GAN)E(Goodfellow et al.,2014)的建立，已经开发了许多隐式模型
在半监督设置中，Siddharth et al. (2016) 尝试通过使用辅助变量来学习解开的表示。Bouchacourt et al.(2017)提出了 ML-VAE, 可以从一组分组观察中学习解开的表示。在无监督设置中，InfoGAN$\boxed{\mathrm{E}}($Chen et al., 2016) 试图通过引入辅助网络来近似后验来最大化潜在变量和观测值的小子集之间的互信息。然而，它在很大程度上依赖于分布的先验选择，并遭受不稳定的训练。改编自 VAE 的另一种流行的无监督方法$\beta$ -VAEE(Higgins et al., 2016) 在近似后验和先验之间的 KL 距离上施加了很大的压力。然而，无监督的方法并没有将特定的含义锚定到解开中。
与我们的方法更紧密地联系在一起，监督方法利用标记数据的优势，并尝试按预期解开因素。DC-IGN (Kulkarni et al., 2015) 要求活动属性通过在小批量中按平均值馈送其他属性来解释某些变化因素。TD-GAN (Wang et al., 2017) 使用标签映射网络来提高解纠缠表示的质量，这与通过解纠缠网络从图像中提取的表示一致。此外，通过实施对抗性训练策略， 提高了生成图像的质量。但是，应该对身份信息进行标记，以便在交换属性时保留 id 信息， 这使得将其应用到许多其他没有 id 标签的数据集中受到限制。IcGAN (Perarnau et al., 2016)是一种多阶段训练算法，它首先利用 cGAN (Mirza &Osindero, 2014)学习从潜在表示和条件信息到真实图像的映射，然后以监督的方式学习其从图像到潜在表示和条件的逆映射。整体效果取决于每个训练阶段，因此很难获得令人满意的图像。与这些模型不同，我们的模型既不需要标签中的显式 ID 信息，也不需要多阶段训练
许多作品已经研究了使用基于 GAN 的体系结构的不成对图像数据之间的图像到图像转换， 例如 Isola et al. (2016) , Taigman et al. (2016) , Zhu et al. (2017) , Liu et al. (2017)和Zhou et al. (2017) 。有趣的是，这些模型需要一种类似于我们设置的0/1弱监督形式。但是，它们被限制在两个图像域中，这两个域相对于单个属性彼此相反。。我们的模型与它们不同，因为我们推广到多属性图像数据的情况。具体而言，我们采用迭代训练策略来克服不平衡多属性图像数据集上的训练困难
3 DNA-GAN Approach
在本节中，我们正式介绍我们的方法。在我们的设置中考虑了一组多标签图像的$\mathscr{X}$和一组标签$\mathscr{Y}$。令$\left\{(\mathbb{X}^1,\mathbb{Y}^1),\ldots,(\mathbb{X}^m,\mathbb{Y}^m)\right\}$表示整个训练数据集，其中 X$^i\in\mathscr{X}$ 是第 $i$ 个图像，其标签为$\operatorname{\bar{Y}}^i\in\mathscr{U}$。小写字母$m$表示集合$\mathscr{X}$中的样本数，$n$表示属性数。标签$\mathrm{Y} ^{i}= ( \mathrm{y} _{1}^{i}, \ldots , \mathrm{y} _{n}^{i})$ 是 $n\text{ 维 向 量 , 其 中 每 个 元 素 表 示 X}^{i}$ 是 否 具 有 特 定 属 性 。例 如 , 在 标 签 具 $\text{女 二 个 问 时 间 为 一 四 位 }^{\:i}$ 他 们 的 时 间 下 上 一 个 体 计 $\mathrm{~V}^{i}( 10, 1)$ 他 不 部 网 络 $\mathrm{V} ^{i}$ 它 址 公 共有三个候选 [刘海，眼镜，微笑] 的情况下，标签为 Y$^i=(1,0,1)$的面部图像 X$^i$应描绘带有刘海且没有眼镜的笑脸
2
|
[
[
0.2132568359375,
0.10479736328125,
0.82275390625,
0.10479736328125,
0.82275390625,
0.11932373046875,
0.2132568359375,
0.11932373046875
],
[
0.2344970703125,
0.11932373046875,
0.82177734375,
0.11932373046875,
0.82177734375,
0.1331787109375,
0.2344970703125,
0.1331787109375
],
[
0.236083984375,
0.1331787109375,
0.82177734375,
0.1331787109375,
0.82177734375,
0.146484375,
0.236083984375,
0.146484375
],
[
0.2115478515625,
0.1654052734375,
0.82275390625,
0.1654052734375,
0.82275390625,
0.179931640625,
0.2115478515625,
0.179931640625
],
[
0.233642578125,
0.1787109375,
0.82275390625,
0.179931640625,
0.82275390625,
0.1944580078125,
0.233642578125,
0.1932373046875
],
[
0.2320556640625,
0.1912841796875,
0.303955078125,
0.1932373046875,
0.303955078125,
0.2083740234375,
0.2320556640625,
0.205810546875
],
[
0.2132568359375,
0.2259521484375,
0.82275390625,
0.2259521484375,
0.82275390625,
0.2392578125,
0.2132568359375,
0.2392578125
],
[
0.2344970703125,
0.2392578125,
0.82275390625,
0.2392578125,
0.82275390625,
0.25390625,
0.2344970703125,
0.25390625
],
[
0.233642578125,
0.25244140625,
0.42236328125,
0.25390625,
0.42236328125,
0.268310546875,
0.233642578125,
0.267578125
],
[
0.1732177734375,
0.3037109375,
0.326904296875,
0.3037109375,
0.326904296875,
0.317626953125,
0.1732177734375,
0.317626953125
],
[
0.176513671875,
0.34228515625,
0.82421875,
0.34228515625,
0.82421875,
0.35595703125,
0.176513671875,
0.35595703125
],
[
0.17724609375,
0.35595703125,
0.82275390625,
0.35595703125,
0.82275390625,
0.369384765625,
0.17724609375,
0.369384765625
],
[
0.178955078125,
0.371826171875,
0.82275390625,
0.371826171875,
0.82275390625,
0.3818359375,
0.178955078125,
0.3818359375
],
[
0.176513671875,
0.38330078125,
0.82275390625,
0.385009765625,
0.82275390625,
0.3984375,
0.176513671875,
0.396484375
],
[
0.176513671875,
0.397705078125,
0.82421875,
0.397705078125,
0.82421875,
0.410888671875,
0.176513671875,
0.410888671875
],
[
0.176513671875,
0.412353515625,
0.74609375,
0.412353515625,
0.74609375,
0.425537109375,
0.176513671875,
0.425537109375
],
[
0.176513671875,
0.43310546875,
0.8203125,
0.43310546875,
0.8203125,
0.4462890625,
0.176513671875,
0.4462890625
],
[
0.176513671875,
0.4462890625,
0.82275390625,
0.44580078125,
0.82275390625,
0.460205078125,
0.176513671875,
0.4609375
],
[
0.17724609375,
0.4609375,
0.82177734375,
0.4609375,
0.82177734375,
0.474853515625,
0.17724609375,
0.474853515625
],
[
0.176513671875,
0.47607421875,
0.82177734375,
0.47607421875,
0.82177734375,
0.4892578125,
0.176513671875,
0.4892578125
],
[
0.1748046875,
0.4873046875,
0.82275390625,
0.488037109375,
0.82275390625,
0.50244140625,
0.1748046875,
0.501953125
],
[
0.176513671875,
0.501953125,
0.82275390625,
0.501953125,
0.82275390625,
0.51513671875,
0.176513671875,
0.51513671875
],
[
0.17724609375,
0.517578125,
0.313720703125,
0.517578125,
0.313720703125,
0.52783203125,
0.17724609375,
0.52783203125
],
[
0.178955078125,
0.5390625,
0.82177734375,
0.5390625,
0.82177734375,
0.5498046875,
0.178955078125,
0.5498046875
],
[
0.176513671875,
0.55029296875,
0.82177734375,
0.55029296875,
0.82177734375,
0.564453125,
0.176513671875,
0.564453125
],
[
0.1748046875,
0.564453125,
0.82177734375,
0.56494140625,
0.82177734375,
0.5791015625,
0.1748046875,
0.57763671875
],
[
0.17724609375,
0.58056640625,
0.8203125,
0.58056640625,
0.8203125,
0.59130859375,
0.17724609375,
0.59130859375
],
[
0.17724609375,
0.5947265625,
0.82177734375,
0.5947265625,
0.82177734375,
0.60498046875,
0.17724609375,
0.60498046875
],
[
0.1748046875,
0.60595703125,
0.8203125,
0.60595703125,
0.8203125,
0.619140625,
0.1748046875,
0.619140625
],
[
0.176513671875,
0.6220703125,
0.82275390625,
0.6220703125,
0.82275390625,
0.63525390625,
0.176513671875,
0.63525390625
],
[
0.178955078125,
0.63623046875,
0.82177734375,
0.63623046875,
0.82177734375,
0.646484375,
0.178955078125,
0.646484375
],
[
0.176513671875,
0.6484375,
0.82275390625,
0.6484375,
0.82275390625,
0.662109375,
0.176513671875,
0.662109375
],
[
0.176513671875,
0.66357421875,
0.6298828125,
0.66357421875,
0.6298828125,
0.6767578125,
0.176513671875,
0.6767578125
],
[
0.176513671875,
0.681640625,
0.82177734375,
0.68310546875,
0.82177734375,
0.6962890625,
0.176513671875,
0.69580078125
],
[
0.178955078125,
0.69873046875,
0.82177734375,
0.69873046875,
0.82177734375,
0.708984375,
0.178955078125,
0.708984375
],
[
0.176513671875,
0.7109375,
0.82275390625,
0.7109375,
0.82275390625,
0.724609375,
0.176513671875,
0.724609375
],
[
0.176513671875,
0.72607421875,
0.82275390625,
0.72607421875,
0.82275390625,
0.7392578125,
0.176513671875,
0.7392578125
],
[
0.17724609375,
0.7392578125,
0.82275390625,
0.7392578125,
0.82275390625,
0.75244140625,
0.17724609375,
0.75244140625
],
[
0.176513671875,
0.75244140625,
0.43701171875,
0.75244140625,
0.43701171875,
0.7666015625,
0.176513671875,
0.7666015625
],
[
0.1707763671875,
0.80029296875,
0.390625,
0.8017578125,
0.390625,
0.81640625,
0.1707763671875,
0.81494140625
],
[
0.176513671875,
0.84130859375,
0.82177734375,
0.84130859375,
0.82177734375,
0.85498046875,
0.176513671875,
0.85498046875
],
[
0.1748046875,
0.85400390625,
0.82275390625,
0.85498046875,
0.82275390625,
0.869140625,
0.1748046875,
0.86865234375
],
[
0.176513671875,
0.86865234375,
0.82177734375,
0.86865234375,
0.82177734375,
0.88330078125,
0.176513671875,
0.88330078125
],
[
0.1748046875,
0.8818359375,
0.82177734375,
0.88330078125,
0.82177734375,
0.89794921875,
0.1748046875,
0.896484375
],
[
0.176513671875,
0.89794921875,
0.82177734375,
0.89794921875,
0.82177734375,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.176513671875,
0.9111328125,
0.35546875,
0.9111328125,
0.35546875,
0.9248046875,
0.176513671875,
0.9248046875
],
[
0.490234375,
0.94873046875,
0.50732421875,
0.94873046875,
0.50732421875,
0.9619140625,
0.490234375,
0.9619140625
]
] |
[
"1: We propose a supervised algorithm called DNA-GAN that is able to disentangle",
"multiple attributes as demonstrated by the experiments of interpolating multiple",
"attributes on Multi-PIE (Gross et al., 2010) and CelebA (Liu et al., 2015) datasets.",
"2. We introduce the annihilating operation that prevents from trivial solutions: the",
"attribute-relevant part encodes information of the whole image instead of a certain",
"attribute.",
"3. We employ an iterative training method to address the problem of unbalanced",
"multi-attribute image data, which was theoretically proved to be more efficient",
"than random image pairs.",
"2Related Work",
"传统的表示学习算法主要集中在（1）概率图模型，其特征在于受限玻尔兹曼机（RBM)E",
"(Smolensky，1986），自动编码器（AE）及其变体；（2）流形学习和几何方法，例如主成分分析",
"PCA)(Pearson，1901)，局部线性人（LLE)E(Roweis&Saul，2000），THMAXXXXCC",
"等。然而，最近的研究积极地集中在开发学习表示数据分布的深度概率模型上。。实际上",
"Kingma&Welling（2013）采用了显式模型分布，并利用变分推理来学习其参数。随着生成",
"对抗网络（GAN)E(Goodfellowetal.，2014）的建立，已经开发了许多隐式模型",
"在半监督设置中，Siddharthetal.（2016）尝试通过使用辅助变量来学习解开的表示。Boucha",
"courtetal.（2017）提出了ML-VAE，可以从一组分组观察中学习解开的表示。在无监督设",
"置中，InfoGANE(Chenetal.，2016）试图通过引人辅助网络来近似后验来最大化潜在变量",
"和观测值的小子集之间的互信息。然而，它在很大程度上依赖于分布的先验选择，并遭受不",
"稳定的训练。改编自VAE的另一种流行的无监督方法β-VAEE(Higginsetal.,2016)在近",
"似后验和先验之间的KL距离上施加了很大的压力。然而，无监督的方法并没有将特定的",
"含义锚定到解开中。",
"与我们的方法更紧密地联系在一起，监督方法利用标记数据的优势，并尝试按预期解开因",
"素。DC-IGN（Kulkarnietal.，2015）要求活动属性通过在小批量中按平均值馈送其他属性",
"来解释某些变化因素。TD-GAN（Wangetal.，2017）使用标签映射网络来提高解纠缠表示",
"的质量，这与通过解纠缠网络从图像中提取的表示一致。此外，通过实施对抗性训练策略",
"提高了生成图像的质量。但是，应该对身份信息进行标记，以便在交换属性时保留id信息，",
"这使得将其应用到许多其他没有id标签的数据集中受到限制。IcGAN(Perarnauetal.",
"2016）是一种多阶段训练算法，它首先利用cGAN（Mirza&Osindero，2014）学习从潜在表",
"示和条件信息到真实图像的映射：然后以监督的方式学习其从图像到潜在表示和条件的递",
"映射。整体效果取决于每个训练阶段，因此很难获得令人满意的图像。与这些模型不同，我",
"们的模型既不需要标签中的显式ID信息，也不需要多阶段训练",
"许多作品已经研究了使用基于GAN的体系结构的不成对图像数据之间的图像到图像转换",
"如Isolaetal.(2016)，Taigmanetal.(2016)，Zhuetal.（2017)Liuetal.(2017)和",
"Zhouetal.（2017）。有趣的是，这些模型需要一种类似于我们设置的0/1弱监督形式。但",
"是，它们被限制在两个图像域中，这两个域相对于单个属性彼此相反。。我们的模型与它们",
"不同，因为我们推广到多属性图像数据的情况。具体而言，我们采用送代训练策略来克服不",
"平衡多属性图像数据集上的训练困难",
"3DNA-GANApproach",
"在本节中，我们正式介绍我们的方法。在我们的设置中考虑了一组多标签图像的%和一",
"组标签。令(Xl,Yi)，...，(Xm,Ym）)表示整个训练数据集，其中XiE%是第i个",
"图像，其标签为Yie?。小写字母m表示集合%中的样本数，n表示属性数。标签",
"Yi=（yi....,y）是n维向量，其中每个元素表示Xi是否具有特定属性。例如，在标签具",
"有三个候选[刘海，眼镜，微笑]的情况下，标签为Yi=（1,0,1）的面部图像Xi应描绘带",
"有刘海且没有眼镜的笑脸",
"2"
] |
{
"arxiv_id": "1711.05415",
"language": "zh",
"page_idx": 1
}
|
||
1711.05415-zh-2
|
1711.05415-zh-2.jpg
|
## 3.1 Model
如图$\boxed{\mathrm{E}}1$ 所示，DNA-GAN 主要由三个部分组成：编码器 (Enc ),解码器 (Dec )和鉴别
器 (D )。编码器将真实世界的图像$A$和$B$映射为两个潜在的解纠缠表示
$$\mathrm{Enc}(A)=[a_{1},\ldots,a_{i},\ldots,a_{n},z_{a}],\quad\mathrm{Enc}(B)=[b_{1},\ldots,b_{i},\ldots,b_{n},z_{b}]$$
(1)
,其中$[a_1,\ldots,a_i,\ldots,a_n]$称为属性相关部分 ,而$z_a$称为属性无关部分。$a_i$应该是一个DNA 片段，它控制 y$i$ ,即标签中的第$i$个属性，而$z_a$用于保持未出现在属性列表中的其他静默因素以及图像标识信息。同样的道理也适用于 Enc$(B)$
<FigureHere>
Figure 1: DNA-GAN architecture.
我们在框架中每次都关注一个属性。假设我们在$i$ -th 属性。$A$和$B$需要具有不同的标签， i。e。, 分 别 为 ( $y_{1}^{A}$, $\ldots$, $1_{i}^{A}$, $\ldots$, $y_{n}^{A}$) 和 ( $y_{1}^{B}$, $\ldots$, $0_{i}^{B}$, $\ldots$, $y_{n}^{B}$) 。在 此 设 置 下 , $A$始终用于主导模式，而$B$用于隐性模式。我们直接复制 Enc$(\dot{A})$作为$A_1$的潜在表示，并消灭 Enc$(B)$副本中的$b_i$作为$B_1$的潜在表示。an 灭操作意味着将所有元素替换为零，并且在解开属性中起着关键作用，这将在E3.3节中详细讨论。通过交换$a_i$和$0_i$,我们获得了两个新的潜在表示$[a_1,\ldots,0_i,\ldots,a_n,z_a]$和$[b_1,\ldots,a_i,\ldots,b_n,z_b]$ ,它们应该分别解码为$A_2$和$B_2$ 。通过解码器 Dec , 我们可以得到四个新生成的图像$\dot{A} _1$ , $B_1$ , $A_2$和$B_2$ 。
$$\begin{array}{ll}\operatorname{Dec}([a_1,\dots,a_i,\dots,a_n,z_a])=A_1,&\operatorname{Dec}([b_1,\dots,0_i,\dots,b_n,z_b])=B_1\\\operatorname{Dec}([a_1,\dots,0_i,\dots,a_n,z_a])=A_2,&\operatorname{Dec}([b_1,\dots,a_i,\dots,b_n,z_b])=B_2\end{array}$$
(2)
在这四个子代中，$A_1$和$B_1$分别是$A$和$B$的直接重建，而$A_2$和$B_2$是新颖的杂交品种。$A$和$A_1$ 、$B$和$B_1$之间的重建损失分别保证了直接重建样本的质量。此外，使用对抗鉴别器 D 有助于使生成的样本$A_2$与$B$无法区分 ,而$B_{2}$与$A$无法区分 ,我们可以考虑将属性相关信息编码在$a_i$中
3.2 Loss Functions
$\begin{array}{l}\text{给定两个图像 }A\text{ 和 }B\text{ 及其标签 Y}^{A}=(\mathrm{y}_{1}^{A},\ldots,1_{i}^{A},\ldots,\mathrm{y}_{n}^{A})\text{ 和 Y}^{B}=(\mathrm{y}_{1}^{B},\ldots,0_{i}^{B},\ldots,\mathrm{y}_{n}^{B})\\\text{它们在 }i\text{ 位置不同，数据流可以用 }(1)\text{ 和 }(2)\text{ 概括。我们强制 }B\text{ 的 }i\text{ 潜在编码为零},\text{ 以防}\end{array}$
止琐碎的解决方案，我们将在 3.3 节中讨论
编码器和解码器接收两种类型的损失：(1)重建损失，
$$L_{reconstruct}=\|A-A_1\|_1+\|B-B_1\|_1$$
(3)
,用于测量一系列编码和解码后的重建质量；(2)标准 GAN 损失，
$$L_{GAN}=-\mathbb{E}[\log(\mathrm D(A_2|\mathrm y_i^A=1))]-\mathbb{E}[\log(\mathrm D(B_2|\mathrm y_i^B=0))]$$
(4)
3
|
[
[
0.1748046875,
0.1060791015625,
0.2646484375,
0.1060791015625,
0.2646484375,
0.117431640625,
0.1748046875,
0.117431640625
],
[
0.176513671875,
0.1300048828125,
0.82421875,
0.1300048828125,
0.82421875,
0.143310546875,
0.176513671875,
0.143310546875
],
[
0.1748046875,
0.14453125,
0.70751953125,
0.14453125,
0.70751953125,
0.1578369140625,
0.1748046875,
0.1578369140625
],
[
0.2646484375,
0.1666259765625,
0.734375,
0.1666259765625,
0.734375,
0.179931640625,
0.2646484375,
0.179931640625
],
[
0.80126953125,
0.1666259765625,
0.82421875,
0.1666259765625,
0.82421875,
0.179931640625,
0.80126953125,
0.179931640625
],
[
0.1895751953125,
0.1881103515625,
0.2646484375,
0.189453125,
0.2646484375,
0.2008056640625,
0.1895751953125,
0.199462890625
],
[
0.281005859375,
0.1881103515625,
0.82177734375,
0.186279296875,
0.82177734375,
0.2008056640625,
0.281005859375,
0.20263671875
],
[
0.1748046875,
0.2020263671875,
0.351318359375,
0.2020263671875,
0.351318359375,
0.216552734375,
0.1748046875,
0.216552734375
],
[
0.34716796875,
0.20263671875,
0.82275390625,
0.20263671875,
0.82275390625,
0.216552734375,
0.34716796875,
0.216552734375
],
[
0.1748046875,
0.2139892578125,
0.60693359375,
0.21533203125,
0.60693359375,
0.2298583984375,
0.1748046875,
0.2291259765625
],
[
0.5986328125,
0.274658203125,
0.7138671875,
0.274658203125,
0.7138671875,
0.285400390625,
0.5986328125,
0.285400390625
],
[
0.366943359375,
0.296630859375,
0.381591796875,
0.296630859375,
0.381591796875,
0.339111328125,
0.366943359375,
0.339111328125
],
[
0.398681640625,
0.29736328125,
0.44677734375,
0.29736328125,
0.44677734375,
0.3125,
0.398681640625,
0.3125
],
[
0.46728515625,
0.302978515625,
0.48193359375,
0.2998046875,
0.486083984375,
0.30810546875,
0.471435546875,
0.311279296875
],
[
0.50732421875,
0.2998046875,
0.56640625,
0.2998046875,
0.56640625,
0.311279296875,
0.50732421875,
0.311279296875
],
[
0.400390625,
0.32568359375,
0.4453125,
0.32568359375,
0.4453125,
0.337158203125,
0.400390625,
0.337158203125
],
[
0.5107421875,
0.32763671875,
0.53369140625,
0.32763671875,
0.53369140625,
0.3359375,
0.5107421875,
0.3359375
],
[
0.5400390625,
0.326904296875,
0.560546875,
0.326904296875,
0.560546875,
0.335205078125,
0.5400390625,
0.335205078125
],
[
0.650390625,
0.347900390625,
0.65625,
0.347900390625,
0.65625,
0.352783203125,
0.650390625,
0.352783203125
],
[
0.3349609375,
0.33203125,
0.34716796875,
0.33203125,
0.34716796875,
0.33837890625,
0.3349609375,
0.33837890625
],
[
0.619140625,
0.348388671875,
0.6396484375,
0.348388671875,
0.6396484375,
0.361083984375,
0.619140625,
0.361083984375
],
[
0.398681640625,
0.356689453125,
0.4453125,
0.356689453125,
0.4453125,
0.36865234375,
0.398681640625,
0.36865234375
],
[
0.5087890625,
0.35791015625,
0.55810546875,
0.35791015625,
0.55810546875,
0.36865234375,
0.5087890625,
0.36865234375
],
[
0.58642578125,
0.37890625,
0.6005859375,
0.37890625,
0.6005859375,
0.42236328125,
0.58642578125,
0.42236328125
],
[
0.62158203125,
0.380126953125,
0.6337890625,
0.380126953125,
0.6337890625,
0.39013671875,
0.62158203125,
0.39013671875
],
[
0.400390625,
0.4072265625,
0.44287109375,
0.4072265625,
0.44287109375,
0.417236328125,
0.400390625,
0.417236328125
],
[
0.400390625,
0.38330078125,
0.441162109375,
0.38330078125,
0.441162109375,
0.39453125,
0.400390625,
0.39453125
],
[
0.5146484375,
0.410400390625,
0.52685546875,
0.410400390625,
0.52685546875,
0.4140625,
0.5146484375,
0.4140625
],
[
0.62353515625,
0.410888671875,
0.6357421875,
0.410888671875,
0.6357421875,
0.421630859375,
0.62353515625,
0.421630859375
],
[
0.398681640625,
0.431884765625,
0.453369140625,
0.431884765625,
0.453369140625,
0.443115234375,
0.398681640625,
0.443115234375
],
[
0.5107421875,
0.434326171875,
0.54150390625,
0.434326171875,
0.54150390625,
0.442626953125,
0.5107421875,
0.442626953125
],
[
0.5400390625,
0.434326171875,
0.55615234375,
0.434326171875,
0.55615234375,
0.442626953125,
0.5400390625,
0.442626953125
],
[
0.62109375,
0.439453125,
0.6396484375,
0.439453125,
0.6396484375,
0.455078125,
0.62109375,
0.455078125
],
[
0.283447265625,
0.458984375,
0.396240234375,
0.458984375,
0.396240234375,
0.473388671875,
0.283447265625,
0.473388671875
],
[
0.338134765625,
0.48291015625,
0.34228515625,
0.48291015625,
0.34228515625,
0.488037109375,
0.338134765625,
0.488037109375
],
[
0.375732421875,
0.5087890625,
0.62353515625,
0.5087890625,
0.62353515625,
0.52294921875,
0.375732421875,
0.52294921875
],
[
0.176513671875,
0.53857421875,
0.82177734375,
0.53857421875,
0.82177734375,
0.55322265625,
0.176513671875,
0.55322265625
],
[
0.1748046875,
0.5517578125,
0.5390625,
0.5517578125,
0.5390625,
0.56640625,
0.1748046875,
0.56640625
],
[
0.53173828125,
0.55322265625,
0.82177734375,
0.55322265625,
0.82177734375,
0.56640625,
0.53173828125,
0.56640625
],
[
0.176513671875,
0.56640625,
0.82275390625,
0.56640625,
0.82275390625,
0.58056640625,
0.176513671875,
0.58056640625
],
[
0.17724609375,
0.580078125,
0.82177734375,
0.580078125,
0.82177734375,
0.5947265625,
0.17724609375,
0.5947265625
],
[
0.176513671875,
0.5947265625,
0.82177734375,
0.5947265625,
0.82177734375,
0.60791015625,
0.176513671875,
0.60791015625
],
[
0.17724609375,
0.609375,
0.25244140625,
0.60986328125,
0.25244140625,
0.6220703125,
0.17724609375,
0.62060546875
],
[
0.30810546875,
0.60791015625,
0.4306640625,
0.60693359375,
0.4306640625,
0.6220703125,
0.30810546875,
0.62255859375
],
[
0.483642578125,
0.609375,
0.82275390625,
0.609375,
0.82275390625,
0.62255859375,
0.483642578125,
0.62255859375
],
[
0.4306640625,
0.61279296875,
0.476318359375,
0.61279296875,
0.476318359375,
0.62060546875,
0.4306640625,
0.62060546875
],
[
0.2548828125,
0.60986328125,
0.2998046875,
0.60986328125,
0.2998046875,
0.62060546875,
0.2548828125,
0.62060546875
],
[
0.176513671875,
0.62255859375,
0.69140625,
0.62255859375,
0.69140625,
0.63623046875,
0.176513671875,
0.63623046875
],
[
0.256591796875,
0.6416015625,
0.740234375,
0.6416015625,
0.740234375,
0.6591796875,
0.256591796875,
0.6591796875
],
[
0.80126953125,
0.65283203125,
0.82421875,
0.65283203125,
0.82421875,
0.66748046875,
0.80126953125,
0.66748046875
],
[
0.256591796875,
0.6591796875,
0.74267578125,
0.6591796875,
0.74267578125,
0.6767578125,
0.256591796875,
0.6767578125
],
[
0.17724609375,
0.69140625,
0.82177734375,
0.69140625,
0.82177734375,
0.705078125,
0.17724609375,
0.705078125
],
[
0.17724609375,
0.705078125,
0.82275390625,
0.705078125,
0.82275390625,
0.7197265625,
0.17724609375,
0.7197265625
],
[
0.176513671875,
0.71826171875,
0.82275390625,
0.71826171875,
0.82275390625,
0.73291015625,
0.176513671875,
0.73291015625
],
[
0.1748046875,
0.73291015625,
0.330810546875,
0.73291015625,
0.330810546875,
0.74755859375,
0.1748046875,
0.74755859375
],
[
0.1732177734375,
0.7646484375,
0.325927734375,
0.7646484375,
0.325927734375,
0.77783203125,
0.1732177734375,
0.77783203125
],
[
0.176513671875,
0.7890625,
0.81640625,
0.7890625,
0.81640625,
0.8037109375,
0.176513671875,
0.8037109375
],
[
0.176513671875,
0.8037109375,
0.82177734375,
0.8037109375,
0.82177734375,
0.8173828125,
0.176513671875,
0.8173828125
],
[
0.17724609375,
0.8193359375,
0.490234375,
0.8193359375,
0.490234375,
0.830078125,
0.17724609375,
0.830078125
],
[
0.176513671875,
0.8388671875,
0.5556640625,
0.8388671875,
0.5556640625,
0.85302734375,
0.176513671875,
0.85302734375
],
[
0.361083984375,
0.861328125,
0.6337890625,
0.861328125,
0.6337890625,
0.87548828125,
0.361083984375,
0.87548828125
],
[
0.80126953125,
0.859375,
0.826171875,
0.859375,
0.826171875,
0.876953125,
0.80126953125,
0.876953125
],
[
0.176513671875,
0.88525390625,
0.65283203125,
0.88525390625,
0.65283203125,
0.8984375,
0.176513671875,
0.8984375
],
[
0.29736328125,
0.90576171875,
0.69921875,
0.90478515625,
0.69921875,
0.91943359375,
0.29736328125,
0.92041015625
],
[
0.80126953125,
0.90478515625,
0.826171875,
0.90478515625,
0.826171875,
0.92236328125,
0.80126953125,
0.92236328125
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"3.1 Model",
"如图E1所示，DNA-GAN主要由三个部分组成：编码器（Enc），解码器（Dec）和鉴别",
"器（D）。编码器将真实世界的图像A和B映射为两个潜在的解纠缠表示",
"Enc(A) = [a1,...,ai,...,an, zal, Enc(B) = [b1,..., bi,..., bn, zb]",
"(1)",
"其中[a.",
"，ai，...，anl称为属性相关部分，而za称为属性无关部分。ai应该是一个",
"DNA片段，它控制yi，",
"即标签中的第i个属性，而za用于保持未出现在属性列表中的其",
"他静默因素以及图像标识信息。同样的道理也适用于Enc（B）",
"Reconstruction loss",
"Encoder",
"aia2",
"a",
"anza",
"b1b2",
"6m",
"Zh",
"",
"B",
"A1",
"a1a2",
"anZa",
"",
"B",
"a1a2",
"b1b2",
"",
"A,",
"61b2.",
"bn",
"26",
"B2",
"Discriminator loss",
"",
"Figure 1: DNA-GAN architecture.",
"我们在框架中每次都关注一个属性。假设我们在i-th属性。A和B需要具有不同的标签",
"i。e。, 分别为 (yA,...,1A,...,yA)和 (yB,...,OB,",
"...，yB）。在此设置下，A始终用于主导",
"模式，而B用于隐性模式。我们直接复制Enc(A）作为A1的潜在表示，并消灭Enc(B)副",
"本中的b；作为B1的潜在表示。an灭操作意味着将所有元素替换为零，并且在解开属性中",
"起着关键作用，这将在E3.3节中详细讨论。通过交换ai和0，，我们获得了两个新的潜在",
"表示a1..",
",an,za] 和[b1,..",
"：,bn，zbl，它们应该分别解码为A2和B2。通",
".ai..",
".,O...",
"过解码器Dec，我们可以得到四个新生成的图像A1，B1，A2和B2。",
"Dec([a,..., ai,...,an, zal) = Ai, Dec([bi,..., Oi,...,bn, zbl) = Bi",
"(2)",
"Dec([a1,.., O, ...,an, zal) = A2, Dec([b1,..., ai,...,bn, zb]) = B2",
"在这四个子代中，A1和B1分别是A和B的直接重建，而A2和B2是新颖的杂交品种。",
"A和A1、B和B1之间的重建损失分别保证了直接重建样本的质量。此外，使用对抗鉴别",
"器D有助于使生成的样本A2与B无法区分，而B2与A无法区分，我们可以考虑将属性",
"相关信息编码在i中",
"3.2Loss Functions",
"给定两个图像A和B及其标签YA=(yA,...,1A,...,yA）和YB=（yB,..,OB,...,yB）",
"它们在i位置不同，数据流可以用（1）和（2）概括。我们强制B的i潜在编码为零，以防",
"止琐碎的解决方案，我们将在3.3节中讨论",
"编码器和解码器接收两种类型的损失：（1）重建损失，",
"Lreconstruct =A-Ailli +B-Billi",
"(3)",
"，用于测量一系列编码和解码后的重建质量；（2）标准GAN损失，",
"LGAN = -E[log(D(A2lyA = 1)I -E[log(D(B2lyB = 0)",
"(4)",
"3"
] |
{
"arxiv_id": "1711.05415",
"language": "zh",
"page_idx": 2
}
|
||
1711.05415-zh-3
|
1711.05415-zh-3.jpg
|
,用于测量生成的图像的真实度。鉴别器将生成的图像及其标签的第$i$个元素作为输人，并输出一个数字，该数字指示输人图像的真实度。数字越大，图像越逼真。省略系数，编码器和解码器的损失函数为
$$L_{G}=L_{reconstruct}+L_{GAN}.$$
(5)
鉴别器 D 接收标准 GAN 鉴别器丢失
(6)
(7)
$$\begin{aligned}&L_{D_{1}}=-\mathbb{E}[\log(\mathrm{D}(A|\mathrm{y}_{i}^{A}=1))]-\mathbb{E}[\log(1-\mathrm{D}(B_{2}|\mathrm{y}_{i}^{A}=1))]\\&L_{D_{0}}=-\mathbb{E}[\log(\mathrm{D}(B|\mathrm{y}_{i}^{B}=0))]-\mathbb{E}[\log(1-\mathrm{D}(A_{2}|\mathrm{y}_{i}^{B}=0))]\\&L_{D}=L_{D_{1}}+L_{D_{0}}\end{aligned}$$
(8)
,其中$L_{D_1}$驱动 D 从$B_2$告诉$A,L_{D_0}$驱动 D 从$A_2$告诉$B$ 3.3 Annihilating Operation Prevents from Trivial Solutions
通过实验，我们可以观察到我们的模型存在琐碎的解决方案，而无需 an 灭操作。我们只是以 单 属 性 的 情 况 为 例 。假 设 Enc$( A) = [ a, z_a]$ 和 Enc$( B) = [ b, z_b]$ , 我 们 可 以 得 到 四 个 孩 子 , 而无需 an 灭操作
$$A_1=\mathrm{Dec}([a,z_a]),\quad B_1=\mathrm{Dec}([b,z_b]),\quad A_2=\mathrm{Dec}([b,z_a]),\quad B_2=\mathrm{Dec}([a,z_b]).$$
重建损失使其从潜在编码到图像可逆。对抗性鉴别器 D 应该通过告诉$A_{2}$是否看起来像$B$ 一样真实，$B_{2}$是否看起来像$A$一样真实来将属性与其他信息分离。众所周知，当达到纳什均衡时，生成对抗网络会获得最佳解决方案。但是如果没有 an 灭操作，则整个图像的信息可以集中编码到属性相关部分中，这意味着
$$\operatorname{Enc}(A)=[a,0],\quad\operatorname{Enc}(B)=[b,0].$$
(10)
这样，我们有以下四个孩子：
$$A_1=\mathrm{Dec}([a,0]),\quad B_1=\mathrm{Dec}([b,0]),\quad A_2=\mathrm{Dec}([b,0]),\quad B_2=\mathrm{Dec}([a,0]).$$
在这种情况下，鉴别器 D 无法将$A_2$与$B$共享相同的编码，因为它们。通过重建过程，$A_2$ 和$B$会导致相同的图像，这与我们的期望相反，即$A_2$应使用从$B$借用的属性来描绘$A$ 中的人。同样的情况也出现在$B_2$和$A$的情况下
为了摆脱这些琐碎的解决方案，我们通过将隐性模式$b\text{ 替换为相同大小的零张量 }^{1}$,i 来采用 an 灭操作。e。, 把它敲掉，让另一部分活着。实际上，如果将整个图像的信息仅编码到属性相关部分中，则在这种情况下的四个子将是
$$A_1=\mathrm{Dec}([a,0]),\quad B_1=\mathrm{Dec}([0,0]),\quad A_2=\mathrm{Dec}([0,0]),\quad B_2=\mathrm{Dec}([a,0]).$$
因此，$B_1$和$A_2$的编码根本不包含任何信息，因此，$B_1$和$A_2$的相应图像中的人应该与我们的所有目标相同。5。因此，在实现湮灭操作时，属性无关部分对于编码图像的某些信息是必要的
## 3.4 Iterative Training
为了降低解开多个属性的难度，我们采取迭代训练的策略：我们每次在某个位置使用一对标签相反的图像来更新我们的模型。假设我们在第$i$个位置，图像$A$ 的标签是$( \mathrm{y} _{1}^{A}, \ldots , 1_{i}^{A}, \ldots , \mathrm{y} _{n}^{A})$ ,而图像 $B$ 的标签是 (y$_1^{B},\ldots,0_{i}^{B},\ldots,\mathrm{y}_{n}^{B})$。在每次迭代过程中，随着$i$从 1 到$n$反复进行，我们的模型通过这样一对图像可以逐个解开多个属性
与随机图像对的训练相比，迭代训练被证明是更有效的。图像的随机对意味着每次随机选
择图像对而没有标签约束。一对具有不同标签的图像称为 useful pair
我们从理论上表明，我们的迭代训练机制比随机图像对更有效，特别是当数据集不平衡时。
所有的证明都可以在 中找到。
$^{1}$Use tf.zeros\_like() in TensorFlow implementation.
4
|
[
[
0.1895751953125,
0.1060791015625,
0.82177734375,
0.1060791015625,
0.82177734375,
0.11932373046875,
0.1895751953125,
0.11932373046875
],
[
0.176513671875,
0.11932373046875,
0.82275390625,
0.11932373046875,
0.82275390625,
0.1331787109375,
0.176513671875,
0.1331787109375
],
[
0.176513671875,
0.1331787109375,
0.339111328125,
0.1331787109375,
0.339111328125,
0.146484375,
0.176513671875,
0.146484375
],
[
0.40283203125,
0.1572265625,
0.59326171875,
0.1578369140625,
0.59326171875,
0.17236328125,
0.40283203125,
0.1717529296875
],
[
0.80224609375,
0.1578369140625,
0.82421875,
0.1578369140625,
0.82421875,
0.17236328125,
0.80224609375,
0.17236328125
],
[
0.17724609375,
0.1925048828125,
0.44287109375,
0.1925048828125,
0.44287109375,
0.20263671875,
0.17724609375,
0.20263671875
],
[
0.295654296875,
0.21533203125,
0.701171875,
0.21533203125,
0.701171875,
0.2298583984375,
0.295654296875,
0.2298583984375
],
[
0.80126953125,
0.2139892578125,
0.82421875,
0.2139892578125,
0.82421875,
0.23291015625,
0.80126953125,
0.23291015625
],
[
0.295654296875,
0.23486328125,
0.70166015625,
0.23486328125,
0.70166015625,
0.25,
0.295654296875,
0.25
],
[
0.80126953125,
0.2342529296875,
0.826171875,
0.2342529296875,
0.826171875,
0.251953125,
0.80126953125,
0.251953125
],
[
0.2998046875,
0.25244140625,
0.423095703125,
0.25390625,
0.423095703125,
0.268310546875,
0.2998046875,
0.267578125
],
[
0.80224609375,
0.25244140625,
0.82275390625,
0.25244140625,
0.82275390625,
0.268310546875,
0.80224609375,
0.268310546875
],
[
0.1748046875,
0.27978515625,
0.62744140625,
0.27978515625,
0.62744140625,
0.29345703125,
0.1748046875,
0.29345703125
],
[
0.1748046875,
0.3125,
0.61376953125,
0.3125,
0.61376953125,
0.32568359375,
0.1748046875,
0.32568359375
],
[
0.176513671875,
0.339111328125,
0.82177734375,
0.339111328125,
0.82177734375,
0.352783203125,
0.176513671875,
0.352783203125
],
[
0.176513671875,
0.352783203125,
0.82275390625,
0.352783203125,
0.82275390625,
0.3662109375,
0.176513671875,
0.3662109375
],
[
0.176513671875,
0.367431640625,
0.30224609375,
0.367431640625,
0.30224609375,
0.380615234375,
0.176513671875,
0.380615234375
],
[
0.223876953125,
0.391357421875,
0.82275390625,
0.391357421875,
0.82275390625,
0.40478515625,
0.223876953125,
0.40478515625
],
[
0.176513671875,
0.415283203125,
0.82421875,
0.415283203125,
0.82421875,
0.4287109375,
0.176513671875,
0.4287109375
],
[
0.178955078125,
0.429931640625,
0.82177734375,
0.429931640625,
0.82177734375,
0.443115234375,
0.178955078125,
0.443115234375
],
[
0.176513671875,
0.443115234375,
0.82275390625,
0.443115234375,
0.82275390625,
0.45703125,
0.176513671875,
0.45703125
],
[
0.176513671875,
0.45703125,
0.50244140625,
0.45703125,
0.50244140625,
0.47021484375,
0.176513671875,
0.47021484375
],
[
0.378173828125,
0.48095703125,
0.619140625,
0.48095703125,
0.619140625,
0.494384765625,
0.378173828125,
0.494384765625
],
[
0.79345703125,
0.479736328125,
0.82421875,
0.479736328125,
0.82421875,
0.49560546875,
0.79345703125,
0.49560546875
],
[
0.176513671875,
0.5087890625,
0.37744140625,
0.5087890625,
0.37744140625,
0.51953125,
0.176513671875,
0.51953125
],
[
0.236083984375,
0.5322265625,
0.82275390625,
0.5322265625,
0.82275390625,
0.54541015625,
0.236083984375,
0.54541015625
],
[
0.176513671875,
0.55810546875,
0.82275390625,
0.55810546875,
0.82275390625,
0.5712890625,
0.176513671875,
0.5712890625
],
[
0.1748046875,
0.57080078125,
0.82275390625,
0.57080078125,
0.82275390625,
0.583984375,
0.1748046875,
0.583984375
],
[
0.17724609375,
0.5869140625,
0.51953125,
0.5869140625,
0.51953125,
0.59765625,
0.17724609375,
0.59765625
],
[
0.17724609375,
0.60693359375,
0.82275390625,
0.60693359375,
0.82275390625,
0.62060546875,
0.17724609375,
0.62060546875
],
[
0.17724609375,
0.62060546875,
0.82177734375,
0.62060546875,
0.82177734375,
0.6337890625,
0.17724609375,
0.6337890625
],
[
0.176513671875,
0.6337890625,
0.5185546875,
0.6337890625,
0.5185546875,
0.64794921875,
0.176513671875,
0.64794921875
],
[
0.236083984375,
0.6591796875,
0.82275390625,
0.6591796875,
0.82275390625,
0.67236328125,
0.236083984375,
0.67236328125
],
[
0.17724609375,
0.68603515625,
0.82177734375,
0.68603515625,
0.82177734375,
0.6962890625,
0.17724609375,
0.6962890625
],
[
0.17724609375,
0.69873046875,
0.82275390625,
0.69873046875,
0.82275390625,
0.7119140625,
0.17724609375,
0.7119140625
],
[
0.176513671875,
0.7138671875,
0.2418212890625,
0.7138671875,
0.2418212890625,
0.72607421875,
0.176513671875,
0.72607421875
],
[
0.1732177734375,
0.74365234375,
0.346435546875,
0.74560546875,
0.346435546875,
0.76025390625,
0.1732177734375,
0.75830078125
],
[
0.17724609375,
0.7734375,
0.8203125,
0.7734375,
0.8203125,
0.7841796875,
0.17724609375,
0.7841796875
],
[
0.1871337890625,
0.78515625,
0.82177734375,
0.78515625,
0.82177734375,
0.798828125,
0.1871337890625,
0.798828125
],
[
0.176513671875,
0.7998046875,
0.82177734375,
0.7998046875,
0.82177734375,
0.814453125,
0.176513671875,
0.814453125
],
[
0.1748046875,
0.8125,
0.7158203125,
0.81298828125,
0.7158203125,
0.82763671875,
0.1748046875,
0.8271484375
],
[
0.176513671875,
0.83544921875,
0.82275390625,
0.83544921875,
0.82275390625,
0.8486328125,
0.176513671875,
0.8486328125
],
[
0.176513671875,
0.85107421875,
0.66015625,
0.85107421875,
0.66015625,
0.861328125,
0.176513671875,
0.861328125
],
[
0.176513671875,
0.87060546875,
0.82275390625,
0.87060546875,
0.82275390625,
0.8837890625,
0.176513671875,
0.8837890625
],
[
0.1748046875,
0.88330078125,
0.390625,
0.8837890625,
0.390625,
0.8984375,
0.1748046875,
0.89794921875
],
[
0.193603515625,
0.9091796875,
0.53173828125,
0.912109375,
0.53173828125,
0.9267578125,
0.193603515625,
0.923828125
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.9609375,
0.49267578125,
0.9609375
]
] |
[
"用于测量生成的图像的真实度。鉴别器将生成的图像及其标签的第i个元素作为输入，并",
"输出一个数字，该数字指示输人图像的真实度。数字越大，图像越逼真。省略系数，编码器",
"和解码器的损失函数为",
"LG=Lreconstruct +LGAN",
"(5)",
"鉴别器D接收标准GAN鉴别器丢失",
"LD, =-E[log(D(A|y =1))] -E[log(1-D(B2lyA =1)))",
"(6)",
"LD。=-E[log(D(BlyB = 0))) -E[log(1 -D(A2lyB = 0)))",
"(7)",
"LD= LDi+ LDo",
"(8)",
"，其中LDI驱动D从B2告诉A，LD驱动D从A2告诉B",
"3.3Annihilating Operation Prevents from Trivial Solutions",
"通过实验，我们可以观察到我们的模型存在碎的解决方案，而无需a灭操作。我们只是",
"以单属性的情况为例。假设Enc(A)=[a,zal和Enc(B=[b,zb]，我们可以得到四个孩子",
"而无需an灭操作",
"Ai =Dec([a, zal), Bi =Dec([b, zb]), A2 =Dec([b, zal), B2 = Dec([a, zb]l). (9)",
"重建损失使其从潜在编码到图像可逆。对抗性鉴别器D应该通过告诉A2是否看起来像B",
"一样真实，B2是否看起来像A一样真实来将属性与其他信息分离。众所周知，当达到纳什",
"均衡时，生成对抗网络会获得最佳解决方案。。但是如果没有an灭操作，则整个图像的信",
"息可以集中编码到属性相关部分中，这意味着",
"Enc(A) =[a,O], Enc(B) =[b,O],",
"(10)",
"这样，我们有以下四个孩子：",
"Ai =Dec([a,0]),Bi =Dec([b,0]),A2 =Dec([b,0]), B2 = Dec([a, 0]). (11)",
"在这种情况下，鉴别器D无法将A2与B共享相同的编码，因为它们。通过重建过程，A2",
"和B会导致相同的图像，这与我们的期望相反，即A2应使用从B借用的属性来描绘A",
"中的人。同样的情况也出现在B2和A的情况下",
"为了摆脱这些碎的解决方案，我们通过将隐性模式6替换为相同大小的零张量1，i来采",
"用an灭操作。e。，把它敲掉，让另一部分活着。实际上，如果将整个图像的信息仅编码到",
"属性相关部分中，则在这种情况下的四个子将是",
"Ai =Dec([a, 0),Bi =Dec([0, 0]),A2 =Dec([0,0]),B2 =Dec([a, 0]). (12)",
"因此，B1和A2的编码根本不包含任何信息：因此，B1和A2的相应图像中的人应该与我",
"们的所有目标相同。5。因此，在实现灭操作时，属性无关部分对于编码图像的某些信息",
"是必要的",
"3.4Iterative Training",
"为了降低解开多个属性的难度，我们采取迭代训练的策略：我们每次在某个位置使用",
"对标签相反的图像来更新我们的模型。假设我们在第i个位置，图像A的标签是",
"(yi，...,A，..,yn），而图像B的标签是（yP，..o，...,yn）。在每次迭代过程中，随着",
"i从1到n反复进行，我们的模型通过这样一对图像可以逐个解开多个属性",
"与随机图像对的训练相比，送代训练被证明是更有效的。图像的随机对意味着每次随机选",
"择图像对而没有标签约束。一对具有不同标签的图像称为usefulpail",
"我们从理论上表明，我们的选代训练机制比随机图像对更有效，特别是当数据集不平衡时。",
"所有的证明都可以在中找到。",
"1Use tf.zeros_like()in TensorFlow implementation",
"4"
] |
{
"arxiv_id": "1711.05415",
"language": "zh",
"page_idx": 3
}
|
||
1711.05415-zh-4
|
1711.05415-zh-4.jpg
|
Theorem 1. Let $\mathscr{X} = \{ ( \mathrm{X} ^{1}, \mathrm{Y} ^{1}) , \ldots , ( \mathrm{X} ^{m}, \mathrm{Y} ^{m}) \}$ denote the whole multi-attribute image dataset, where X$^i$ is a multi-attribute image and its label Y$^i=(\mathrm{y}_1^i,\ldots,\mathrm{y}_n^i)$ is an $n-$ $\operatorname*{dimensional}$ vector. There are totally $2^n$ kinds of labels, denoted by $\mathscr{L} = \{ l_1, \ldots , _{2^n}\} .$ The number of images with label $l_i$ is $m_i$ ,and $\sum_{i=1}^{2^n}m_i=m.$ To select all useful pairs at least once, the expected numbers of iterations needed for randomly selecting pairs and for iterative training are denoted by E$_1$ and E$_2$ respectively. Then,
(13)
$$\begin{aligned}&\mathrm{E}_{1}=m^{2}\left(1+\frac{1}{2}+\cdots+\frac{1}{m^{2}-\sum_{i=1}^{2^{n}}m_{i}^{2}}\right)\\&\mathrm{E}_{2}\leq2n\cdot\max_{s=1,\ldots,n}\sum_{i\in I_{s},j\in J_{s}}m_{i}m_{j}\left(1+\frac{1}{2}+\cdots+\frac{1}{m^{2}-\sum_{k_{1}=1}^{2^{n-1}}(m_{i_{k_{1}}}+m_{j_{k_{1}}})^{2}}\right)\end{aligned}$$
(14)
where $I_s$ represents the indices of labels where the $s$ -th element is 1 , and $J_s$ represents the
indices of labels where the $s$-th element is 0.
Definition 1. (Balancedness) Define the balancedness of a dataset $\mathscr{X}$ described above with
respect to the $s$-th attribute as follows:
$$\rho_s=\frac{\sum_{i\in I_s}m_i}{\sum_{j\in J_s}m_j}$$
(15)
where $I_s$ represents the indices of labels where the s-th element is 1 , and $J_s$ represents the
indices of labels where the $s$-th element is 0.
Theorem 2. We have $\mathrm{E}_2\leq\mathrm{E}_1$,when
$$n\leq\min_{s}\frac{(\rho_s+1)^2}{2\rho_s}.$$
(16)
$\mathrm{Specifically, ~E}_{2}\leq \mathrm{E} _{1}{\mathrm{~holds~true~for~all~}}n\leq 2$ .
函数$(\rho+1)^2/(2\rho)$的属性与平衡性的定义非常吻合，因为它对于$\rho$和$1/\rho$具有相同的值， 这对于不同的标记方法是不变的。随着数据集变得更加不平衡，它的值变得更大。最小值是在$\rho=1$处获得的 , 这是平衡数据集的情况
定理 2 证明了当属性的数量满足标准 (16) 时，迭代训练机制总是比随机图像对更有效。随着数据集变得更加不平衡，$(\rho_s+1)^2/(2\rho_s)$变得更大，这意味着(16)可以更容易地得到满足。更重要的是，迭代训练有助于稳定不平衡数据集上的训练过程。例如，给定一个双属性数据集，每种数据的数量如下：
Table 1: The example of an unbalanced two-attribute dataset.
$$\begin{array}{c|c|c|c|c|c}\hline\text{Label}&(0,0)&(0,1)&(1,0)&(1,1)\\\hline\text{Number of data}&1&1&m&m\\\hline\end{array}$$
如果$m\gg1$是一个非常大的数字，那么很有可能我们将通过随机选择对每次选择一对标签为(1,0)和(1,1)的图像。我们忽略标签为(1,0)和(1,0)或(1,1)和(1,1)的一对图像，尽管这两种情况被选择的概率相等。因为他们不是有用的对，所以不参加培训。在这种情况下，大多数时候模型是针对第二个属性进行训练的，这将导致最终学习的模型对第一个属性的效果较差。但是，迭代训练可以防止这种情况发生，因为我们针对两个属性均匀地更新模型
### 4 Experimental Results
在本节中，我们对两个现实世界的数据集进行了不同类型的实验，以验证我们方法的有效性。我们使用由 5e-5 的学习率和动量0初始化的 RMSProp (Sutskever et al., 2013)优化方法。所有神经网络在卷积或解卷积后都配备了 BatchNormalization$\boxed{\mathrm{E}}($Ioffe &Szegedy, 2015) 。我们使用 LeakyRelu (Maas et al., 2013) 作为编码器中的激活函数。此外，我们采用 WassersteinGAN (Arjovsky et al.,2017) 中提到的策略进行稳定训练。我们将所有图像按照 9:1 的比例分为训练图像和测试图像。以下所有结果均来自没有樱桃采摘的测试图像。更多细节可以在 https://github.com/Prinsphield/DNA-GAN 上找到
5
|
[
[
0.176513671875,
0.10418701171875,
0.82275390625,
0.1060791015625,
0.82275390625,
0.12060546875,
0.176513671875,
0.11871337890625
],
[
0.1732177734375,
0.11614990234375,
0.81396484375,
0.117431640625,
0.81396484375,
0.1351318359375,
0.1732177734375,
0.1337890625
],
[
0.1732177734375,
0.1307373046875,
0.82275390625,
0.1331787109375,
0.82275390625,
0.147705078125,
0.1732177734375,
0.1458740234375
],
[
0.1748046875,
0.149658203125,
0.53125,
0.149658203125,
0.53125,
0.1634521484375,
0.1748046875,
0.1634521484375
],
[
0.521484375,
0.15087890625,
0.82275390625,
0.15087890625,
0.82275390625,
0.1641845703125,
0.521484375,
0.1641845703125
],
[
0.1748046875,
0.1641845703125,
0.82421875,
0.1641845703125,
0.82421875,
0.1787109375,
0.1748046875,
0.1787109375
],
[
0.1748046875,
0.1787109375,
0.625,
0.1787109375,
0.625,
0.189453125,
0.1748046875,
0.189453125
],
[
0.306396484375,
0.203857421875,
0.325927734375,
0.203857421875,
0.325927734375,
0.212158203125,
0.306396484375,
0.212158203125
],
[
0.4345703125,
0.2020263671875,
0.44677734375,
0.2020263671875,
0.44677734375,
0.21337890625,
0.4345703125,
0.21337890625
],
[
0.273681640625,
0.210205078125,
0.316162109375,
0.21533203125,
0.314453125,
0.2235107421875,
0.27294921875,
0.2183837890625
],
[
0.79150390625,
0.20703125,
0.826171875,
0.210205078125,
0.82421875,
0.2259521484375,
0.7890625,
0.222900390625
],
[
0.314453125,
0.21337890625,
0.375732421875,
0.212158203125,
0.375732421875,
0.222900390625,
0.314453125,
0.2235107421875
],
[
0.316162109375,
0.2183837890625,
0.330810546875,
0.2235107421875,
0.326904296875,
0.231689453125,
0.310546875,
0.2266845703125
],
[
0.2099609375,
0.2108154296875,
0.27783203125,
0.2108154296875,
0.27783203125,
0.222900390625,
0.2099609375,
0.222900390625
],
[
0.38232421875,
0.2183837890625,
0.5,
0.2183837890625,
0.5,
0.236083984375,
0.38232421875,
0.236083984375
],
[
0.48193359375,
0.246826171875,
0.488525390625,
0.251953125,
0.483642578125,
0.255615234375,
0.476318359375,
0.250732421875
],
[
0.640625,
0.2442626953125,
0.65283203125,
0.2442626953125,
0.65283203125,
0.2568359375,
0.640625,
0.2568359375
],
[
0.2091064453125,
0.25244140625,
0.326904296875,
0.25244140625,
0.326904296875,
0.266357421875,
0.2091064453125,
0.266357421875
],
[
0.44287109375,
0.251953125,
0.48193359375,
0.255615234375,
0.48046875,
0.266357421875,
0.441162109375,
0.26318359375
],
[
0.79345703125,
0.25244140625,
0.82421875,
0.25244140625,
0.82421875,
0.268310546875,
0.79345703125,
0.268310546875
],
[
0.3896484375,
0.255126953125,
0.43701171875,
0.255126953125,
0.43701171875,
0.266357421875,
0.3896484375,
0.266357421875
],
[
0.35302734375,
0.255126953125,
0.370849609375,
0.255126953125,
0.370849609375,
0.264404296875,
0.35302734375,
0.264404296875
],
[
0.488525390625,
0.25830078125,
0.5107421875,
0.25830078125,
0.5107421875,
0.261962890625,
0.488525390625,
0.261962890625
],
[
0.5458984375,
0.260009765625,
0.74609375,
0.26318359375,
0.74609375,
0.285400390625,
0.5458984375,
0.2822265625
],
[
0.476318359375,
0.261962890625,
0.488525390625,
0.261962890625,
0.488525390625,
0.272705078125,
0.476318359375,
0.272705078125
],
[
0.285888671875,
0.26513671875,
0.314453125,
0.26513671875,
0.314453125,
0.26953125,
0.285888671875,
0.26953125
],
[
0.313720703125,
0.268310546875,
0.32275390625,
0.268310546875,
0.32275390625,
0.271484375,
0.313720703125,
0.271484375
],
[
0.324462890625,
0.270751953125,
0.3896484375,
0.270751953125,
0.3896484375,
0.2822265625,
0.324462890625,
0.2822265625
],
[
0.176513671875,
0.287353515625,
0.82421875,
0.287353515625,
0.82421875,
0.300537109375,
0.176513671875,
0.300537109375
],
[
0.176513671875,
0.302978515625,
0.498291015625,
0.302978515625,
0.498291015625,
0.313232421875,
0.176513671875,
0.313232421875
],
[
0.1748046875,
0.318115234375,
0.82421875,
0.318115234375,
0.82421875,
0.33203125,
0.1748046875,
0.33203125
],
[
0.1748046875,
0.333251953125,
0.4599609375,
0.33203125,
0.4599609375,
0.34521484375,
0.1748046875,
0.3466796875
],
[
0.483642578125,
0.35107421875,
0.55224609375,
0.354248046875,
0.55126953125,
0.37255859375,
0.48193359375,
0.369384765625
],
[
0.443603515625,
0.3623046875,
0.486083984375,
0.367431640625,
0.483642578125,
0.380126953125,
0.441162109375,
0.375
],
[
0.7939453125,
0.359130859375,
0.826171875,
0.3623046875,
0.82177734375,
0.37890625,
0.7900390625,
0.375732421875
],
[
0.48046875,
0.371826171875,
0.5537109375,
0.373779296875,
0.5537109375,
0.389404296875,
0.48046875,
0.386962890625
],
[
0.1748046875,
0.391357421875,
0.82421875,
0.392578125,
0.82421875,
0.4072265625,
0.1748046875,
0.406005859375
],
[
0.1732177734375,
0.406005859375,
0.5048828125,
0.4072265625,
0.5048828125,
0.42041015625,
0.1732177734375,
0.419189453125
],
[
0.176513671875,
0.4248046875,
0.44287109375,
0.4248046875,
0.44287109375,
0.438232421875,
0.176513671875,
0.438232421875
],
[
0.426513671875,
0.451904296875,
0.56201171875,
0.4375,
0.56640625,
0.458251953125,
0.4306640625,
0.472900390625
],
[
0.7939453125,
0.449462890625,
0.826171875,
0.45263671875,
0.82177734375,
0.4697265625,
0.7900390625,
0.466552734375
],
[
0.51318359375,
0.46337890625,
0.54150390625,
0.46337890625,
0.54150390625,
0.47607421875,
0.51318359375,
0.47607421875
],
[
0.1748046875,
0.48095703125,
0.50341796875,
0.48095703125,
0.50341796875,
0.494384765625,
0.1748046875,
0.494384765625
],
[
0.176513671875,
0.505859375,
0.8203125,
0.505859375,
0.8203125,
0.51953125,
0.176513671875,
0.51953125
],
[
0.1748046875,
0.51806640625,
0.82275390625,
0.51953125,
0.82275390625,
0.5341796875,
0.1748046875,
0.53271484375
],
[
0.1732177734375,
0.53271484375,
0.486083984375,
0.5341796875,
0.486083984375,
0.54736328125,
0.1732177734375,
0.546875
],
[
0.17724609375,
0.556640625,
0.82177734375,
0.556640625,
0.82177734375,
0.5673828125,
0.17724609375,
0.5673828125
],
[
0.176513671875,
0.568359375,
0.82275390625,
0.568359375,
0.82275390625,
0.58349609375,
0.176513671875,
0.58349609375
],
[
0.17724609375,
0.58349609375,
0.82275390625,
0.58349609375,
0.82275390625,
0.5966796875,
0.17724609375,
0.5966796875
],
[
0.17724609375,
0.5986328125,
0.39453125,
0.5986328125,
0.39453125,
0.609375,
0.17724609375,
0.609375
],
[
0.27685546875,
0.62255859375,
0.72119140625,
0.62255859375,
0.72119140625,
0.63623046875,
0.27685546875,
0.63623046875
],
[
0.365234375,
0.6484375,
0.408447265625,
0.6484375,
0.408447265625,
0.66357421875,
0.365234375,
0.66357421875
],
[
0.453369140625,
0.6513671875,
0.5107421875,
0.6513671875,
0.5107421875,
0.66357421875,
0.453369140625,
0.66357421875
],
[
0.51123046875,
0.6513671875,
0.5595703125,
0.6513671875,
0.5595703125,
0.66357421875,
0.51123046875,
0.66357421875
],
[
0.56884765625,
0.65087890625,
0.67724609375,
0.65087890625,
0.67724609375,
0.6640625,
0.56884765625,
0.6640625
],
[
0.3203125,
0.66357421875,
0.490234375,
0.66357421875,
0.490234375,
0.6767578125,
0.3203125,
0.6767578125
],
[
0.525390625,
0.66650390625,
0.54736328125,
0.6640625,
0.54833984375,
0.67236328125,
0.52685546875,
0.6748046875
],
[
0.5849609375,
0.66650390625,
0.60546875,
0.66650390625,
0.60546875,
0.6767578125,
0.5849609375,
0.6767578125
],
[
0.640625,
0.66357421875,
0.66259765625,
0.66748046875,
0.65869140625,
0.6787109375,
0.63623046875,
0.6748046875
],
[
0.176513671875,
0.693359375,
0.82275390625,
0.693359375,
0.82275390625,
0.70703125,
0.176513671875,
0.70703125
],
[
0.176513671875,
0.70849609375,
0.82177734375,
0.70849609375,
0.82177734375,
0.7216796875,
0.176513671875,
0.7216796875
],
[
0.17724609375,
0.7216796875,
0.82177734375,
0.7216796875,
0.82177734375,
0.73486328125,
0.17724609375,
0.73486328125
],
[
0.17724609375,
0.73486328125,
0.82275390625,
0.73486328125,
0.82275390625,
0.74853515625,
0.17724609375,
0.74853515625
],
[
0.1748046875,
0.74853515625,
0.82275390625,
0.75,
0.82275390625,
0.76318359375,
0.1748046875,
0.76220703125
],
[
0.176513671875,
0.7646484375,
0.2099609375,
0.7646484375,
0.2099609375,
0.7763671875,
0.176513671875,
0.7763671875
],
[
0.1732177734375,
0.79736328125,
0.3896484375,
0.79736328125,
0.3896484375,
0.8125,
0.1732177734375,
0.8125
],
[
0.176513671875,
0.82763671875,
0.82275390625,
0.82763671875,
0.82275390625,
0.84130859375,
0.176513671875,
0.84130859375
],
[
0.176513671875,
0.84130859375,
0.82275390625,
0.84130859375,
0.82275390625,
0.85498046875,
0.176513671875,
0.85498046875
],
[
0.176513671875,
0.85498046875,
0.82177734375,
0.85498046875,
0.82177734375,
0.869140625,
0.176513671875,
0.869140625
],
[
0.1748046875,
0.869140625,
0.82177734375,
0.869140625,
0.82177734375,
0.88330078125,
0.1748046875,
0.88330078125
],
[
0.17724609375,
0.88525390625,
0.82177734375,
0.88525390625,
0.82177734375,
0.89599609375,
0.17724609375,
0.89599609375
],
[
0.176513671875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.6640625,
0.91015625,
0.6640625,
0.9248046875,
0.1748046875,
0.92529296875
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Theorem 1. Let = {(Xl,Yl),...,(Xm,Ym)) denote the whole multi-attribute image",
"dataset, where x is a multi-attribute image and its label Yi = (yi,..:,yn) is an n",
"dimensional vector. There are totally 2n kinds of labels, denoted by = [li,..., 2n] . The",
"number of images with label l, is mi , and ?",
"-, mi = m : To select all useful pairs at",
"least once, the expected numbers of iterations needed for randomly selecting pairs and for",
"iterative training are denoted byE, and E2respectively.Then",
"1",
"1",
"11+",
"(13)",
"+·-",
"2",
"Ei = m?",
"m2-L2=1m?",
"",
"1",
"E2≤2n· max",
"1+",
"(14)",
"mimj",
">",
"",
"m2-Dr=1(mik1+mjn1)2",
"2",
"1",
"",
"iEls,jEJ",
"where I, represents the indices of labels where the s -th element is 1 , and J, represents the",
"indices of labels where the s -th element is 0",
"Definition 1. (Balancedness) Define the balancedness of a dataset described above with",
"respect to the s -th attribute as follows:",
"Ziel,mi",
"ps=",
"(15)",
"ZjeJ.mj",
"where I, represents the indices of labels where the s -th element is I , and J, represents the",
"indices of labels where the s -th element is 0.",
"Theorem2.We have E2≤Ei,when",
"n ≤min (es + 1)?",
"(16)",
"2ps",
"Specifically, E2≤Ei holds true for all n ≤ 2.",
"函数（β+1)²/(2p）的属性与平衡性的定义非常吻合，因为它对于p和1/p具有相同的值",
"这对于不同的标记方法是不变的。随着数据集变得更加不平衡，它的值变得更大。最小值是",
"在p=1处获得的，这是平衡数据集的情况",
"定理2证明了当属性的数量满足标准（16）时，送代训练机制总是比随机图像对更有效。随",
"着数据集变得更加不平衡，（ps+1)²/(2ps）变得更大，这意味着（16）可以更容易地得到满",
"足。更重要的是，迭代训练有助于稳定不平衡数据集上的训练过程。例如，给定一个双属性",
"数据集，每种数据的数量如下：",
"Table l: The example of an unbalanced two-attribute dataset.",
"Label",
"(0,0)",
"(0,1)",
"(1,0)/(1,1)",
"Numberof data1",
"1",
"m",
"m",
"如果㎡》1是一个非常大的数字，那么很有可能我们将通过随机选择对每次选择一对标签",
"为（1,0)和（1,1）的图像。我们忽略标签为（1,0)和(1,0)或（1,1)和（1,1）的一对图像，尽",
"管这两种情况被选择的概率相等。因为他们不是有用的对，所以不参加培训。在这种情况",
"下，大多数时候模型是针对第二个属性进行训练的，这将导致最终学习的模型对第一个属",
"性的效果较差。但是，迭代训练可以防止这种情况发生，因为我们针对两个属性均匀地更新",
"模型",
"4Experimental Results",
"在本节中，我们对两个现实世界的数据集进行了不同类型的实验，以验证我们方法的有效",
"性。我们使用由5e-5的学习率和动量0初始化的RMSProp（Sutskever etal.2013）优",
"化方法。所有神经网络在卷积或解卷积后都配备了BatchNormalizationE(Ioffe&Szegedy",
"2015）。我们使用LeakyRelu（Maasetal.，2013）作为编码器中的激活函数。此外，我们采",
"用WassersteinGAN（Arjovskyetal.，2017）中提到的策略进行稳定训练。我们将所有图像",
"按照9:1的比例分为训练图像和测试图像。以下所有结果均来自没有樱桃采摘的测试图像。",
"更多细节可以在https://github.com/Prinsphield/DNA-GAN上找到",
"5"
] |
{
"arxiv_id": "1711.05415",
"language": "zh",
"page_idx": 4
}
|
||
2210.03310-zh-0
|
2210.03310-zh-0.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
# 具有局部损失的前向梯度缩放
# $\textbf{Mengye Ren}^{1*},\textbf{Simon Kornblith}^2,\textbf{Renjie Liao}^3,\textbf{Geoffrey Hinton}^{2,4}$ $^{1}\mathrm{NYU,~}^{2}\mathrm{Google,~}^{3}\mathrm{UBC,~}^{4}$Vector Institute
ABSTRACT
Forward gradient learning computes a noisy directional gradient and is a biologically plausible alternative to backprop for learning deep neural networks However, the standard forward gradient algorithm, when applied naively, suffers from high variance when the number of parameters to be learned is large. In this paper, we propose a series of architectural and algorithmic modifications that together make forward gradient learning practical for standard deep learning benchmark tasks. We show that it is possible to substantially reduce the variance of the forward gradient estimator by applying perturbations to activations rather than weights. We further improve the scalability of forward gradient by introducing a large number of local greedy loss functions, each of which involves only a small number of learnable parameters, and a new MLPMixer-inspired architecture, LocalMixer, that is more suitable for local learning. Our approach matches backprop on MNIST and CIFAR-10 and significantly outperforms previously proposed backprop-free algorithms on ImageNet. Code is released at https://github.com/google-research/google-research, tree/master/local\_forward\_gradient.
### 1 INTRODUCTION
当今大多数深度神经网络都是使用反向传播算法进行训练的 (a。k。a。backprop) (Werbos, 1974; LeCun, 1985; Rumelhart et al., 1986) , 它通过将误差信号从损失函数向后传播到每一层来有效地计算权重参数的梯度。尽管人工神经网络最初受到生物神经元的启发，但由于大脑不会形成对称的反向连接或执行同步计算，因此 backprop一直被认为是“生物学上不可信的”。。从工程的角度来看，backprop 与大规模的模型并行性不兼容，并限制了潜在的硬件设计。。这些问题要求深度网络采用截然不同的学习算法
过去，已经尝试通过引人随机向后权重$\overline{\mathbb{E}}(\underline{\mathrm{Lillicrap~et~al.,2016;~Nøkland,~2016)~来解决上述权}}$ 重传输问题，但是发现它们在诸如 ImageNet$\boxed{\mathrm{E}}($Bartunov et al.,2018)之类的较大数据集上的扩展性很差。解决全局同步问题，几篇论文表明，贪婪的局部损失函数几乎可以和端到端学习一样好 (Belilovsky et al., 2019; Löwe et al., 2019; Xiong et al., 2020)。但是，他们仍然依赖于 backprop 来学习每个本地模块中的许多内部层
另一方面，基于权重扰动的方法直接将损失信号发送回权重连接，因此不需要任何后向权重。在前向传递中，网络对突触连接添加了轻微的扰动，然后将权重更新乘以损失的负变化。以前曾提出将重量扰动作为 backprop$\boxed{\mathrm{E}}($Xie & Seung, 1999; Seung, 2003; Fiete & Seung, 2006) 的生物学上合理的替代方法。。代替直接扰动权重，还可以使用正向模式自动微分(AD)来计算沿扰动方向的最终损失的方向梯度$\boxed{\mathrm{E}}($Pearlmutter,1994)。基于前向模式 AD 的算法最近在深度学习 (Baydin et al., 2022; Silver et al., 2022) 的背景下重新受到关注。然而，现有的方法受到维数的诅咒，估计梯度的方差太高，无法有效地训练大型网络
在本文中，我们重新审视了活动扰动$\operatorname{E}($Le Cun et al., 1988; Widrow & Lehr, 1990; Fiete & Seung, 2006) 作为权重扰动的替代方法。As previous works focused on specific settings, 我们探索了在具有挑战性的视觉任务上训练的大型网络的普遍适用性。我们证明了活动扰动比权重扰动产生更低的方差梯度估计，并为我们的算法提供了基于连续时间速率的解释。我们通过设计具有许多局部贪婪损失函数的体系结构，将网络隔离为局部模块，从而减少每个损失的可学习参数数量，从而直接解决前向梯度学习的可伸缩性问题。与仅沿深度维度增加局部损耗的先前工作不同，我们发现具有补丁式和通道组式损耗也很关键。最后，受
$^{*}$Work done as a visiting faculty researcher at Google. Correspondence to: mengye@cs.nyu.edu.
1
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.1805419921875,
0.10479736328125,
0.5390625,
0.10479736328125,
0.5390625,
0.12249755859375,
0.1805419921875,
0.12249755859375
],
[
0.1829833984375,
0.15283203125,
0.65673828125,
0.15087890625,
0.65673828125,
0.1654052734375,
0.1829833984375,
0.1673583984375
],
[
0.1846923828125,
0.1654052734375,
0.455810546875,
0.1666259765625,
0.455810546875,
0.18115234375,
0.1846923828125,
0.179931640625
],
[
0.451904296875,
0.2171630859375,
0.5458984375,
0.2197265625,
0.5458984375,
0.2342529296875,
0.451904296875,
0.231689453125
],
[
0.233642578125,
0.25390625,
0.765625,
0.25390625,
0.765625,
0.267578125,
0.233642578125,
0.267578125
],
[
0.2320556640625,
0.266357421875,
0.76123046875,
0.266357421875,
0.76123046875,
0.27978515625,
0.2320556640625,
0.27978515625
],
[
0.233642578125,
0.281005859375,
0.7646484375,
0.281005859375,
0.7646484375,
0.294189453125,
0.233642578125,
0.294189453125
],
[
0.2344970703125,
0.29541015625,
0.76318359375,
0.29541015625,
0.76318359375,
0.30615234375,
0.2344970703125,
0.30615234375
],
[
0.233642578125,
0.309326171875,
0.7646484375,
0.309326171875,
0.7646484375,
0.322509765625,
0.233642578125,
0.322509765625
],
[
0.2320556640625,
0.3212890625,
0.765625,
0.322509765625,
0.765625,
0.337158203125,
0.2320556640625,
0.3359375
],
[
0.2344970703125,
0.33837890625,
0.7646484375,
0.33837890625,
0.7646484375,
0.348388671875,
0.2344970703125,
0.348388671875
],
[
0.2344970703125,
0.35107421875,
0.7646484375,
0.35107421875,
0.7646484375,
0.3642578125,
0.2344970703125,
0.3642578125
],
[
0.233642578125,
0.3642578125,
0.7646484375,
0.3642578125,
0.7646484375,
0.37744140625,
0.233642578125,
0.37744140625
],
[
0.2320556640625,
0.376953125,
0.7646484375,
0.376953125,
0.7646484375,
0.39013671875,
0.2320556640625,
0.39013671875
],
[
0.2320556640625,
0.391357421875,
0.7646484375,
0.391357421875,
0.7646484375,
0.40478515625,
0.2320556640625,
0.40478515625
],
[
0.233642578125,
0.406005859375,
0.765625,
0.406005859375,
0.765625,
0.419189453125,
0.233642578125,
0.419189453125
],
[
0.2344970703125,
0.419189453125,
0.7646484375,
0.419189453125,
0.7646484375,
0.434326171875,
0.2344970703125,
0.434326171875
],
[
0.233642578125,
0.434326171875,
0.765625,
0.434326171875,
0.765625,
0.447509765625,
0.233642578125,
0.447509765625
],
[
0.2344970703125,
0.447509765625,
0.765625,
0.447509765625,
0.765625,
0.4609375,
0.2344970703125,
0.4609375
],
[
0.2344970703125,
0.4609375,
0.576171875,
0.4609375,
0.576171875,
0.474853515625,
0.2344970703125,
0.474853515625
],
[
0.1748046875,
0.498779296875,
0.340576171875,
0.5,
0.340576171875,
0.51318359375,
0.1748046875,
0.51220703125
],
[
0.17724609375,
0.52587890625,
0.82177734375,
0.52587890625,
0.82177734375,
0.5390625,
0.17724609375,
0.5390625
],
[
0.176513671875,
0.537109375,
0.82275390625,
0.53857421875,
0.82275390625,
0.55322265625,
0.176513671875,
0.5517578125
],
[
0.176513671875,
0.55322265625,
0.82275390625,
0.55322265625,
0.82275390625,
0.56640625,
0.176513671875,
0.56640625
],
[
0.1732177734375,
0.56640625,
0.82177734375,
0.56494140625,
0.82177734375,
0.580078125,
0.1732177734375,
0.58056640625
],
[
0.176513671875,
0.58056640625,
0.82275390625,
0.58056640625,
0.82275390625,
0.5947265625,
0.176513671875,
0.5947265625
],
[
0.1748046875,
0.59326171875,
0.576171875,
0.5947265625,
0.576171875,
0.60791015625,
0.1748046875,
0.60693359375
],
[
0.176513671875,
0.61572265625,
0.82275390625,
0.61572265625,
0.82275390625,
0.62890625,
0.176513671875,
0.62890625
],
[
0.17724609375,
0.6298828125,
0.82177734375,
0.6298828125,
0.82177734375,
0.64306640625,
0.17724609375,
0.64306640625
],
[
0.1748046875,
0.64306640625,
0.82177734375,
0.64306640625,
0.82177734375,
0.6572265625,
0.1748046875,
0.6572265625
],
[
0.176513671875,
0.6572265625,
0.82275390625,
0.6572265625,
0.82275390625,
0.671875,
0.176513671875,
0.671875
],
[
0.1748046875,
0.67041015625,
0.52392578125,
0.671875,
0.52392578125,
0.68603515625,
0.1748046875,
0.68505859375
],
[
0.176513671875,
0.6923828125,
0.82275390625,
0.6923828125,
0.82275390625,
0.70556640625,
0.176513671875,
0.70556640625
],
[
0.1748046875,
0.705078125,
0.82177734375,
0.70703125,
0.82177734375,
0.72021484375,
0.1748046875,
0.71826171875
],
[
0.176513671875,
0.7197265625,
0.82275390625,
0.7197265625,
0.82275390625,
0.734375,
0.176513671875,
0.734375
],
[
0.176513671875,
0.734375,
0.82275390625,
0.734375,
0.82275390625,
0.74755859375,
0.176513671875,
0.74755859375
],
[
0.176513671875,
0.74755859375,
0.82275390625,
0.74755859375,
0.82275390625,
0.76123046875,
0.176513671875,
0.76123046875
],
[
0.1748046875,
0.76025390625,
0.82177734375,
0.76025390625,
0.82177734375,
0.7744140625,
0.1748046875,
0.7744140625
],
[
0.176513671875,
0.77587890625,
0.68310546875,
0.77587890625,
0.68310546875,
0.7890625,
0.176513671875,
0.7890625
],
[
0.176513671875,
0.796875,
0.82275390625,
0.796875,
0.82275390625,
0.81005859375,
0.176513671875,
0.81005859375
],
[
0.1748046875,
0.81103515625,
0.82275390625,
0.81103515625,
0.82275390625,
0.82470703125,
0.1748046875,
0.82470703125
],
[
0.1748046875,
0.82470703125,
0.82275390625,
0.82470703125,
0.82275390625,
0.83837890625,
0.1748046875,
0.83837890625
],
[
0.176513671875,
0.83837890625,
0.82421875,
0.83837890625,
0.82421875,
0.8515625,
0.176513671875,
0.8515625
],
[
0.176513671875,
0.8515625,
0.82275390625,
0.8515625,
0.82275390625,
0.86572265625,
0.176513671875,
0.86572265625
],
[
0.176513671875,
0.8662109375,
0.82177734375,
0.8662109375,
0.82177734375,
0.8798828125,
0.176513671875,
0.8798828125
],
[
0.176513671875,
0.8818359375,
0.82177734375,
0.8818359375,
0.82177734375,
0.892578125,
0.176513671875,
0.892578125
],
[
0.193603515625,
0.91015625,
0.7890625,
0.912109375,
0.7890625,
0.92529296875,
0.193603515625,
0.923828125
],
[
0.491943359375,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.96044921875,
0.491943359375,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2023",
"具有局部损失的前向梯度缩放",
"Mengye Ren *, Simon Kornblith?, Renjie Liao3, Geoffrey Hinton2.4",
"1NYU,2Google, 3UBC,4Vector Institute",
"ABSTRACT",
"Forward gradient learning computes a noisy directional gradient and is a bio-",
"logically plausible alternative to backprop for learning deep neural networks",
"However, the standard forward gradient algorithm, when applied naively, suffers",
"from high variance whenthe number of parameters tobe learned islarge.In",
"this paper, we propose a series of architectural and algorithmic modifications",
"that together make forward gradient learning practical for standard deep learning",
"benchmarktasks.Weshowthatitispossibleto substantiallyreducethevariance",
"of the forward gradient estimator by applying perturbations to activations",
"rather than weights.We further improve the scalability of forward gradient by",
"introducing a large number of local greedy loss functions, each of which involves",
"only a small number of learnable parameters, and a new MLPMixer-inspired",
"architecture, LocalMixer, that is more suitable for local learning. Our approach",
"matches backprop on MNIST and CIFAR-10 and significantly outperforms",
"previously proposed backprop-free algorithms on ImageNet.Code is released",
"athttps://github.com/google-research/google-research/",
"tree/master/local_forward_gradient.",
"1INTRODUCTION",
"当今大多数深度神经网络都是使用反向传播算法进行训练的（a。k。a。backprop）（Werbos，",
"1974;LeCun，1985;Rumelhartetal.，1986），它通过将误差信号从损失函数向后传播到每一层",
"来有效地计算权重参数的梯度。尽管人工神经网络最初受到生物神经元的启发，但由于大",
"脑不会形成对称的反向连接或执行同步计算，因此backprop一直被认为是“生物学上不可",
"信的”。。从工程的角度来看，backprop与大规模的模型并行性不兼容，并限制了潜在的硬件",
"设计。。这些问题要求深度网络采用截然不同的学习算法",
"过去，已经尝试通过引人随机向后权重E(Lillicrapetal.，2016;Nokland,2016)来解决上述权",
"重传输问题，但是发现它们在诸如ImageNetE(Bartunovetal.，2018）之类的较大数据集上的",
"扩展性很差。解决全局同步问题，几篇论文表明，贪婪的局部损失函数几乎可以和端到端学",
"习一样好（Belilovskyetal.,2019;Loweetal.,2019；Xiongetal.,2020）。但是，他们仍然依赖",
"于backprop来学习每个本地模块中的许多内部层",
"另一方面，基于权重扰动的方法直接将损失信号发送回权重连接，因此不需要任何后向权",
"重。在前向传递中，网络对突触连接添加了轻微的扰动，然后将权重更新乘以损失的负变化。",
"以前曾提出将重量扰动作为backpropE(Xie&Seung，1999;Seung,2003;Fiete&Seung，2006)",
"的生物学上合理的替代方法。。代替直接扰动权重，还可以使用正向模式自动微分（AD）来",
"计算沿扰动方向的最终损失的方向梯度E(Pearlmutter，1994）。基于前向模式AD的算法最",
"近在深度学习(Baydinetal.,2022;Silveretal.,2022)的背景下重新受到关注。然而，现有的",
"方法受到维数的诅咒，估计梯度的方差太高，无法有效地训练大型网络",
"在本文中，我们重新审视了活动扰动E(LeCunetal.，1988;Widrow&Lehr，1990;Fiete&",
"Seung，2006）作为权重扰动的替代方法。Aspreviousworksfocusedonspecificsettings，我们",
"探索了在具有挑战性的视觉任务上训练的大型网络的普遍适用性。我们证明了活动扰动比",
"权重扰动产生更低的方差梯度估计，并为我们的算法提供了基于连续时间速率的解释。我",
"们通过设计具有许多局部贪婪损失函数的体系结构，将网络隔离为局部模块，从而减少每",
"个损失的可学习参数数量，从而直接解决前向梯度学习的可伸缩性问题。与仅沿深度维度",
"增加局部损耗的先前工作不同，我们发现具有补丁式和通道组式损耗也很关键。最后，受",
"*Work done as a visiting faculty researcher at Google.Correspondence to: mengye@ cs.nyu.edu.",
"1"
] |
{
"arxiv_id": "2210.03310",
"language": "zh",
"page_idx": 0
}
|
||
2210.03310-zh-1
|
2210.03310-zh-1.jpg
|
Published as a conference paper at ICLR 2023
MLPMixer$\boxed{\mathrm{F}}$(Tolstikhin et al., 2021) 设计的启发，我们设计了一个名为 LocalMixer 的网络，该
网络具有线性令牌混合层和分组通道，以更好地与本地学习兼容
我们在监督和自监督图像分类问题上评估了我们的局部贪婪前向梯度算法。在 MNIST 和CIFAR-10 上，我们的学习算法的性能与 backprop 相当，在 ImageNet 上，它的性能明显优于其他使用非对称正向和反向权重的生物学上合理的替代方案。尽管我们还没有在更大规模的问题上完全匹配 backprop, 但我们认为局部损失设计可能是生物学上合理的学习算法和下一代模型并行计算的关键因素。
# 2 RELATED WORK
自感知器时代以来，神经网络学习算法的设计，特别是可以由生物大脑实现的算法，一直是人们关注的焦点。Whittington& Bogacz (2019) 和 Lillicrap et al.(2020) 的综述论文系统地总结了生物学上合理的深度学习的进展。在这里，我们在以下子主题中讨论相关工作。
Forward gradient and reinforcement learning. 我们的工作利用了正向模式自动微分(AD), 这是由$\boxed{\mathrm{E}}$Wengert (1964)首次提出的。后来它被用来学习递归神经网络 (Williams &Zipser, 1989) 和计算 Hessian 向量乘积 (Pearlmutter, 1994)。使用正向模式 AD 计算真实梯度需要完整的雅可比行列式，这通常很大且计算昂贵。最近，Baydin et al.(2022) 和 Silver et al.(2022) 提出了基于方向梯度沿随机 or learned 扰动方向更新权重。他们发现这种方法足以解决小规模问题。这个通用的算法家族也与强化学习(RL)和进化策略(ES)有关，因为在每种情况下，网络都会获得全局奖励。RL 和 ES 在神经网络(Whitley,1993; Stanley&Miikkulainen, 2002; Salimans et al., 2017) 中有着悠久的应用历史，它们对某些连续的控制和决策任务是有效的。Clark et al.(2021)发现全局信用分配也可以在向量神经网络中很好地工作，其中权重仅存在于向量化的神经元组之间
Greedy local learning. 已经有许多尝试使用局部贪婪学习目标来训练深度神经网络。贪婪分层预训练 (Bengio et al.,2006; Hinton et al.,2006; Vincent et al., 2010)一次训练一个单独的层或模块，以贪婪地优化目标。局部损耗通常应用于不同的层或剩余阶段，使用常见的监督和自监督损耗公式 (Belilovsky et al., 2019; Nøkland & Eidnes, 2019; Löwe et al., 2019; Belilovsky et al., 2020) 。Xiong et al. (2020); Gomez et al. (2020) 建议使用重叠损失来减少贪婪学习的影响。Patel et al.(2022) 建议将网络拆分为神经元组。Laskin et al.(2020) 将贪婪局部学习应用于模型并行性训练，Wang et al.(2021)提出增加局部重构损失以保留信息。然而，过去十年中提出的大多数本地学习方法都依赖于 backprop 来计算本地模块内的权重更新。一个例外是 Nøkland & Eidnes (2019) 的工作，它通过使用分层目标以及相似性损失或反馈对齐机制来避免 backprop。门控线性网络及其变体 (Veness et al., 2017; 2021; Sezener et al., 2021) 要求每个神经元进行预测，并在避免灾难性遗忘方面显示出有趣的结果。从理论的角度来看，Baldi & Sadowski (2016) 提供了关于为什么局部学习可能比全局学习更糟糕的见解和证明
Asymmetric feedback weights. Backprop 依赖于权重对称：后向权重与前向权重相同。过去的研究已经研究了这种约束是否必要。Lillicrap et al. (2016) 提出了使用随机和固定后向权重的反馈对齐(FA),并发现它可以支持神经网络中的错误驱动学习。直接 FA (Nøkland,2016) 使用单个后向层将损耗函数布线回每一层。也有一些方法旨在显式更新向后权重。再循环 (Hinton & " $\&$ McClelland, 1987) 和目标传播 (TP) (Bengio, 2014; Lee et al., 2015; Bartunov et al., 2018) 使用局部重建目标来学习单独的前向和后向权重作为彼此的近似逆。Laddernetworks (Rasmus et al., 2015) 发现局部重建目标和不对称权重可以帮助实现强大的半监督学习性能。但是Bartunov et al. (2018)报告了 FA 和 TP 算法都无法扩展到较大的问题，例如 ImageNet,其错误率超过 90%。Liao et al. (2016); Xiao et al. (2019) 建议的符号对称 (SS),其中每个后向连接权重与前向连接权重共享相同的符号。Akrout et al.(2019) 提出的权重镜像和改进的Kolen-Pollack 算法$\boxed{\mathrm{F}}$(Kolen & Pollack, 1994) 来对齐正向和反向权重。Woo et al.(2021) 建议使用下面几层的活动进行更新，以避免双向连接。与这些作品相比，我们通过仅使用奖励(及其变化率)而不是向后权重来规避权重对称性以及更一般的网络对称性问题。
Biologically plausible perturbation learning. 正向梯度与生物学环境中的扰动学习有关。传统上，神经可塑性学习规则的重点是将权重更新作为神经元$\mathbb{E}($Hebb,1949;Widrow&Hoff, 1960; Oja, 1982; Bienenstock et al., 1982; Abbott & Nelson, 2000) 的输入和输出活动的函数。另一方面，权重扰动学习$\boxed{\mathrm{E}}($Jabri &Flower,1992) 更为通用，因为它允许任何形式的全局奖励$\boxed{\mathrm{E}}($Schultz et al.,1997)。它是在基于额定和基于峰值的公式中开发的(Xie&Seung, 1999; Seung,2003)。活动 (或节点)扰动是在浅层网络$\mathbb{E}($Le Cun et al.,1988; Widrow & Lehr, 1990) 中提出的，后来在基于尖峰的连续时间网络$\overline{\mathrm{E}}($Fiete & Seung,2006) 中提出的，它被解释为神经元电导的扰动。Werfel et al.(2003) 表明，backprop 比扰动学习具有更快的收敛速度，并
2
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.176513671875,
0.10418701171875,
0.82421875,
0.10479736328125,
0.82421875,
0.11932373046875,
0.176513671875,
0.11871337890625
],
[
0.176513671875,
0.11932373046875,
0.63232421875,
0.11932373046875,
0.63232421875,
0.1331787109375,
0.176513671875,
0.1331787109375
],
[
0.17724609375,
0.1427001953125,
0.82177734375,
0.1427001953125,
0.82177734375,
0.15283203125,
0.17724609375,
0.15283203125
],
[
0.1748046875,
0.15283203125,
0.82177734375,
0.154052734375,
0.82177734375,
0.1685791015625,
0.1748046875,
0.1673583984375
],
[
0.17724609375,
0.169189453125,
0.8203125,
0.169189453125,
0.8203125,
0.179931640625,
0.17724609375,
0.179931640625
],
[
0.176513671875,
0.18310546875,
0.82177734375,
0.18310546875,
0.82177734375,
0.1962890625,
0.176513671875,
0.1962890625
],
[
0.176513671875,
0.1962890625,
0.410888671875,
0.1962890625,
0.410888671875,
0.210205078125,
0.176513671875,
0.210205078125
],
[
0.1748046875,
0.2203369140625,
0.35302734375,
0.2203369140625,
0.35302734375,
0.2342529296875,
0.1748046875,
0.2342529296875
],
[
0.17724609375,
0.24560546875,
0.82177734375,
0.24560546875,
0.82177734375,
0.2587890625,
0.17724609375,
0.2587890625
],
[
0.176513671875,
0.2587890625,
0.82275390625,
0.2587890625,
0.82275390625,
0.272705078125,
0.176513671875,
0.272705078125
],
[
0.176513671875,
0.272705078125,
0.80126953125,
0.272705078125,
0.80126953125,
0.285888671875,
0.176513671875,
0.285888671875
],
[
0.178955078125,
0.294189453125,
0.82275390625,
0.294189453125,
0.82275390625,
0.30810546875,
0.178955078125,
0.30810546875
],
[
0.17724609375,
0.31005859375,
0.8203125,
0.31005859375,
0.8203125,
0.32080078125,
0.17724609375,
0.32080078125
],
[
0.17724609375,
0.32080078125,
0.82177734375,
0.3212890625,
0.82177734375,
0.3359375,
0.17724609375,
0.335205078125
],
[
0.176513671875,
0.335205078125,
0.82177734375,
0.335205078125,
0.82177734375,
0.348388671875,
0.176513671875,
0.348388671875
],
[
0.176513671875,
0.349853515625,
0.82177734375,
0.349853515625,
0.82177734375,
0.363037109375,
0.176513671875,
0.363037109375
],
[
0.176513671875,
0.363037109375,
0.82275390625,
0.363037109375,
0.82275390625,
0.376953125,
0.176513671875,
0.376953125
],
[
0.178955078125,
0.376953125,
0.82177734375,
0.376953125,
0.82177734375,
0.39013671875,
0.178955078125,
0.39013671875
],
[
0.176513671875,
0.391357421875,
0.82177734375,
0.391357421875,
0.82177734375,
0.40478515625,
0.176513671875,
0.40478515625
],
[
0.1748046875,
0.40478515625,
0.82177734375,
0.40478515625,
0.82177734375,
0.41845703125,
0.1748046875,
0.41845703125
],
[
0.178955078125,
0.42041015625,
0.40283203125,
0.42041015625,
0.40283203125,
0.43115234375,
0.178955078125,
0.43115234375
],
[
0.178955078125,
0.43994140625,
0.82177734375,
0.43994140625,
0.82177734375,
0.453857421875,
0.178955078125,
0.453857421875
],
[
0.17724609375,
0.453857421875,
0.82177734375,
0.453857421875,
0.82177734375,
0.46728515625,
0.17724609375,
0.46728515625
],
[
0.176513671875,
0.4697265625,
0.81884765625,
0.4697265625,
0.81884765625,
0.479736328125,
0.176513671875,
0.479736328125
],
[
0.176513671875,
0.48095703125,
0.82177734375,
0.481689453125,
0.82177734375,
0.49560546875,
0.176513671875,
0.494384765625
],
[
0.17724609375,
0.49755859375,
0.8203125,
0.49755859375,
0.8203125,
0.50830078125,
0.17724609375,
0.50830078125
],
[
0.176513671875,
0.5087890625,
0.82177734375,
0.5087890625,
0.82177734375,
0.52294921875,
0.176513671875,
0.52294921875
],
[
0.176513671875,
0.52294921875,
0.82177734375,
0.52294921875,
0.82177734375,
0.537109375,
0.176513671875,
0.537109375
],
[
0.176513671875,
0.537109375,
0.82177734375,
0.537109375,
0.82177734375,
0.55029296875,
0.176513671875,
0.55029296875
],
[
0.17724609375,
0.55029296875,
0.82177734375,
0.55029296875,
0.82177734375,
0.564453125,
0.17724609375,
0.564453125
],
[
0.176513671875,
0.564453125,
0.82177734375,
0.564453125,
0.82177734375,
0.5791015625,
0.176513671875,
0.5791015625
],
[
0.176513671875,
0.5791015625,
0.82275390625,
0.5791015625,
0.82275390625,
0.59228515625,
0.176513671875,
0.59228515625
],
[
0.1732177734375,
0.59130859375,
0.76123046875,
0.59130859375,
0.76123046875,
0.60498046875,
0.1732177734375,
0.60498046875
],
[
0.17724609375,
0.6123046875,
0.82177734375,
0.61279296875,
0.82177734375,
0.62744140625,
0.17724609375,
0.626953125
],
[
0.176513671875,
0.62744140625,
0.82177734375,
0.62744140625,
0.82177734375,
0.6416015625,
0.176513671875,
0.6416015625
],
[
0.178955078125,
0.64306640625,
0.8203125,
0.64306640625,
0.8203125,
0.65380859375,
0.178955078125,
0.65380859375
],
[
0.17724609375,
0.65576171875,
0.8203125,
0.65576171875,
0.8203125,
0.66943359375,
0.17724609375,
0.66943359375
],
[
0.1748046875,
0.66845703125,
0.82275390625,
0.66845703125,
0.82275390625,
0.681640625,
0.1748046875,
0.681640625
],
[
0.17724609375,
0.681640625,
0.8203125,
0.68310546875,
0.8203125,
0.6962890625,
0.17724609375,
0.69580078125
],
[
0.17724609375,
0.69873046875,
0.8203125,
0.69873046875,
0.8203125,
0.708984375,
0.17724609375,
0.708984375
],
[
0.176513671875,
0.71044921875,
0.82177734375,
0.71044921875,
0.82177734375,
0.724609375,
0.176513671875,
0.724609375
],
[
0.1748046875,
0.7236328125,
0.82177734375,
0.724609375,
0.82177734375,
0.73779296875,
0.1748046875,
0.7373046875
],
[
0.17724609375,
0.73779296875,
0.82177734375,
0.73779296875,
0.82177734375,
0.751953125,
0.17724609375,
0.751953125
],
[
0.176513671875,
0.751953125,
0.82177734375,
0.751953125,
0.82177734375,
0.76513671875,
0.176513671875,
0.76513671875
],
[
0.17724609375,
0.7666015625,
0.82177734375,
0.7666015625,
0.82177734375,
0.77978515625,
0.17724609375,
0.77978515625
],
[
0.17724609375,
0.7822265625,
0.73876953125,
0.7822265625,
0.73876953125,
0.79248046875,
0.17724609375,
0.79248046875
],
[
0.178955078125,
0.8017578125,
0.82177734375,
0.8017578125,
0.82177734375,
0.81494140625,
0.178955078125,
0.81494140625
],
[
0.17724609375,
0.81640625,
0.82177734375,
0.81640625,
0.82177734375,
0.8271484375,
0.17724609375,
0.8271484375
],
[
0.1805419921875,
0.83056640625,
0.81640625,
0.83056640625,
0.81640625,
0.84130859375,
0.1805419921875,
0.84130859375
],
[
0.17724609375,
0.84228515625,
0.82177734375,
0.84228515625,
0.82177734375,
0.85595703125,
0.17724609375,
0.85595703125
],
[
0.17724609375,
0.85595703125,
0.82177734375,
0.85595703125,
0.82177734375,
0.869140625,
0.17724609375,
0.869140625
],
[
0.1748046875,
0.87060546875,
0.82177734375,
0.869140625,
0.82177734375,
0.88330078125,
0.1748046875,
0.8837890625
],
[
0.17724609375,
0.8837890625,
0.82177734375,
0.8837890625,
0.82177734375,
0.89794921875,
0.17724609375,
0.89794921875
],
[
0.17724609375,
0.8994140625,
0.8203125,
0.8994140625,
0.8203125,
0.91015625,
0.17724609375,
0.91015625
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2023",
"MLPMixerE(Tolstikhinetal.，2021）设计的启发，我们设计了一个名为LocalMixer的网络，该",
"网络具有线性令牌混合层和分组通道，以更好地与本地学习兼容",
"我们在监督和自监督图像分类问题上评估了我们的局部贪楼前向梯度算法。在MNIST和",
"CIFAR-1o上，我们的学习算法的性能与backprop相当，在ImageNet上，它的性能明显优于",
"其他使用非对称正向和反向权重的生物学上合理的替代方案。尽管我们还没有在更大规模",
"的问题上完全匹配backprop，但我们认为局部损失设计可能是生物学上合理的学习算法和",
"下一代模型并行计算的关键因素。",
"2RELATED WORK",
"自感知器时代以来，神经网络学习算法的设计，特别是可以由生物大脑实现的算法，一直",
"是人们关注的焦点。Whittington&Bogacz(2019）和Lillicrapetal.（2020）的综述论文系统",
"地总结了生物学上合理的深度学习的进展。在这里，我们在以下子主题中讨论相关工作。",
"Forwardgradientandreinforcementlearning.我们的工作利用了正向模式自动微分（AD）",
"这是由E/Wengert（1964）首次提出的。后来它被用来学习递归神经网络（Williams&Zipser",
"1989）和计算Hessian向量乘积（Pearlmutter，1994）。使用正向模式AD计算真实梯度需要完",
"整的雅可比行列式，这通常很大且计算昂贵。最近，Baydinetal.（2022)和Silveretal.(2022）",
"提出了基于方向梯度沿随机orlearned扰动方向更新权重。他们发现这种方法足以解决小规",
"模问题。这个通用的算法家族也与强化学习（RL）和进化策略（ES）有关，因为在每种情况",
"下，网络都会获得全局奖励。RL和ES在神经网络（Whitley，1993;Stanley&Mikkulainen",
"2002；Salimansetal.，2017）中有着悠久的应用历史，它们对某些连续的控制和决策任务是有",
"效的。Clarketal.(2021）发现全局信用分配也可以在向量神经网络中很好地工作，其中权重",
"仅存在于向量化的神经元组之间",
"Greedylocallearning.已经有许多尝试使用局部贪婪学习目标来训练深度神经网络。贪婪分",
"层预训练（Bengioetal.，2006;Hintonetal.,2006；Vincentetal.2010）一次训练一个单独的层",
"或模块，以贪婪地优化目标。局部损耗通常应用于不同的层或剩余阶段：使用常见的监督和",
"自监督损耗公式（Belilovskyetal.,2019;Nokland&Eidnes,2019;Loweetal.，2019;Belilovsky",
"etal.，2020）。Xiongetal.（2020)；Gomezetal.（2020）建议使用重叠损失来减少贪婪学习的影",
"响。Pateletal.(2022）建议将网络拆分为神经元组。Laskinetal.(2020)将贪婪局部学习应用",
"于模型并行性训练，Wangetal.(2021)提出增加局部重构损失以保留信息。然而，过去十年",
"中提出的大多数本地学习方法都依赖于backprop来计算本地模块内的权重更新。一个例外",
"是Nokland&Eidnes(2019)的工作，它通过使用分层目标以及相似性损失或反馈对齐机制来",
"避免backprop。门控线性网络及其变体（Venessetal.，2017:2021;Sezeneretal.，2021）要求每",
"个神经元进行预测，并在避免灾难性遗忘方面显示出有趣的结果。从理论的角度来看，Baldi",
"&Sadowski(2016)提供了关于为什么局部学习可能比全局学习更糟糕的见解和证明",
"Asymmetricfeedbackweights.Backprop依赖于权重对称：后向权重与前向权重相同。过去的",
"研究已经研究了这种约束是否必要。Lillicrapetal.(2016)提出了使用随机和固定后向权重的",
"反馈对齐（FA)并发现它可以支持神经网络中的错误驱动学习。直接FA（Nokland,2016）使用",
"单个后向层将损耗函数布线回每一层。也有一些方法旨在显式更新向后权重。再循环（Hintor",
"&McClelland,1987)和目标传播(TP）(Bengio,2014;Leeetal.,2015;Bartunov et al.,2018)使",
"用局部重建目标来学习单独的前向和后向权重作为彼此的近似逆。Laddernetworks(Rasmus",
"etal.2015）发现局部重建自标和不对称权重可以帮助实现强大的半监督学习性能。但是",
"Bartunovetal.（2018）报告了FA和TP算法都无法扩展到较大的问题，例如ImageNet，其",
"错误率超过90%。Liaoetal.（2016)；Xiaoetal.（2019）建议的符号对称（SS），其中每个后",
"向连接权重与前向连接权重共享相同的符号。Akroutetal.(2019)提出的权重镜像和改进的",
"Kolen-Pollack算法E(Kolen&Pollack，1994）来对齐正向和反向权重。Wooetal.(2021）建议",
"使用下面几层的活动进行更新，以避免双向连接。与这些作品相比，我们通过仅使用奖励",
"（及其变化率）而不是向后权重来规避权重对称性以及更一般的网络对称性问题",
"Biologicallyplausibleperturbationlearning.正向梯度与生物学环境中的扰动学习有关。传",
"统上，神经可塑性学习规则的重点是将权重更新作为神经元E(Hebb，1949：Widrow&Hoff",
"1960；Oja，1982；Bienenstocketal.，1982；Abbott&Nelson，2000）的输人和输出活动的函数。",
"另一方面，权重扰动学习E(Jabri&Flower，1992）更为通用，因为它允许任何形式的全局奖",
"励E(Schultzetal.1997）。它是在基于额定和基于峰值的公式中开发的（Xie&Seung，1999",
"Seung，2003）。活动（或节点）扰动是在浅层网络E(LeCunetal.，1988;Widrow&Lehr，1990)",
"中提出的，后来在基于尖峰的连续时间网络E(Fiete&Seung，2006）中提出的，它被解释为",
"神经元电导的扰动。Werfeletal.（2003）表明，backprop比扰动学习具有更快的收敛速度，并",
"2"
] |
{
"arxiv_id": "2210.03310",
"language": "zh",
"page_idx": 1
}
|
||
2210.03310-zh-2
|
2210.03310-zh-2.jpg
|
Published as a conference paper at ICLR 2023
<table>
<tbody>
<tr>
<th> </th>
<th>${\mathrm{Unbiased?}}$</th>
<th>Avg. Variance (shared)</th>
<th>Avg. Variance $\because1$ (independent</th>
</tr>
<tr>
<td>$g_{w}$</td>
<td>Yes</td>
<td>$pq+2$ $V+(pq+1)S$ $Nr$</td>
<td>$pq+2$ $V+$ $pq+1$ $S$ $Nr$ $N$</td>
</tr>
<tr>
<td>$g_{a}$</td>
<td>Yes</td>
<td>$(q+$ $\cdot1/.$ </td>
<td> </td>
</tr>
</tbody>
</table>
Table 1: Comparing weight $(g_w)$ and activity $(g_a)$ perturbation. $V=$dimension-wise avg. gradient
variance, $S=$dimension-wise avg. squared gradient norm; $p=$fan-in; $q=$fan-out; $N=$batch size.
且活动扰动战胜了另一个因素的权重扰动。在我们的工作中，我们表明与权重扰动相比，活
动扰动具有较低的梯度估计方差。
3 FORWARD GRADIENT LEARNING
在本节中，我们将回顾并建立学习算法的技术背景。我们首先回顾了正向模式自动微分
(AD)技术。其次，我们在权重空间或活动空间中提出了两种不同类型的扰动
3.1 FORWARD-MODE AUTOMATIC DIFFERENTIATION (AD)
令$f:\mathbb{R}^m\mapsto\mathbb{R}^n$。$f$的雅可比行列式$J_f$是大小为$n\times m$的矩阵。正向模式 AD 计算矩阵向
量乘积$J_f\boldsymbol{v}$ ,其中$v\in\mathbb{R}^m$ 。它被定义为在$x$上沿$v$评估的方向梯度：
$$J_{f}\boldsymbol{v}:=\lim_{\delta\mapsto0}\frac{f(\boldsymbol{x}+\delta\boldsymbol{v})-f(\boldsymbol{x})}{\delta}.$$
(1)
为了进行比较，backprop (也称为反向模式 AD) 计算向量雅可比乘积$vJ_f$ ,其中$v\in\mathbb{R}^n$ 对应于链规则中的最后一项。与反向模式 AD 相比，正向模式 AD 只需要一次正向通过，这是用衍生信息增强的。要计算计算图中节点的雅可比向量积，首先将使用$v$扩展输人节点， 该$v$是要相乘的向量。然后对于其他节点，我们将$(x,x^{\prime})$的元组作为输入发送，并计算一个元组$(y,y^{\prime})$作为输出，其中$x^{\prime}$和$y^{\prime}$是节点$x$和节点$y$的中间导数，$i.e.y^{\prime}=\frac{d\boldsymbol{y}}{d\boldsymbol{x}}x^{\prime}$之间是 MAXXX。在 JAX 库 (Bradbury et al.,2018) 中，正向模式 AD 实现为 jax. jvp
3.2 WEIGHT-PERTURBED FORWARD GRADIENT
生成权重更新的权重扰动最初是在 (Barto et al.,1983; Xie& Seung, 1999; Seung, 2003) 中探索的。Baydin et al.(2022) 使用正向模式 AD 的技术来实现权重扰动，在数值稳定性方面优于有限差分。令$w_ij$为单元$i$和$j$之间的权重连接，$f$为损失函数。我们可以通过从零均值单位方差高斯分布中提取具有 iid 元素$v_{ij}$的随机矩阵来估计梯度。估计器为
$$g_{w}(w_{ij})=\left(\sum_{i'j'}\nabla w_{i'j'}v_{i'j'}\right)v_{ij}.$$
(2)
直观上，该估计器对随机扰动方向$v_{ij}$进行采样，并通过使用正向模式执行点积来测试其如何与真实梯度$\nabla w_{i^{\prime}j^{\prime}}$对齐，然后将标量对齐与扰动方向再次相乘。Baydin et al.(2022) 将使用正向模式 AD 的这种形式的梯度估计称为“正向梯度”。为了与我们稍后详细描述的另一种形式的扰动区分开，我们将其称为“权重扰动前向梯度”,或简称为“权重扰动”
3.3 ACTIVITY-PERTURBED FORWARD GRADIENT
扰动权重的另一种方法是扰动活动，这可以减少每个示例的扰动维数。在限制性假设下，最初在 Le Cun et al. (1988); Widrow & Lehr (1990) 中探索了活动扰动。在这里，我们介绍一种使用活动扰动估计梯度的一般方法。它在生物学上可能是合理的，因为它可能对应于每个神经元$\mathbb{E}($Fiete &Seung,2006)中电导的扰动。。在这里，为了简单起见，我们专注于基于离散时间速率的公式。令$x_i$表示第$i$个突触前神经元的活动，$z_j$表示非线性激活函数之前的第$j$个突触后神经元的活动，$u_j$是$z_j$的扰动。活动扰动的前向梯度估计器是
$$g_a(w_{ij})=x_i\left(\sum_{j'}\nabla z_{j'}u_{j'}\right)u_j,$$
(3)
,其中$\nabla\boldsymbol{z}$和$u$之间的内积再次通过使用前向模式 AD 来计算
3.4 THEORETICAL PROPERTIES
在本节中，我们旨在分析正向梯度估计器的期望和方差属性。我们将分析重点放在一个权
重矩阵$\{w_{ij}\}$的梯度上，但结论也适用于具有多个权重矩阵的网络
表 1总结了$^{1}$的理论结果。在批量大小为 $N$ 的情况下，独立扰动可以实现 $1/N$ 的方差减小，而共享扰动具有由平方梯度范数主导的恒定方差项。但是，在执行独立权重扰动时，不
$^{1}$All proofs can be found in Appendix ?? and 9. Numerical simulation results can be found in Appendix 10
3
|
[
[
0.176513671875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.310546875,
0.083984375,
0.73388671875,
0.083984375,
0.73388671875,
0.09783935546875,
0.310546875,
0.09783935546875
],
[
0.34228515625,
0.10479736328125,
0.366943359375,
0.10479736328125,
0.366943359375,
0.117431640625,
0.34228515625,
0.117431640625
],
[
0.410888671875,
0.1029052734375,
0.5361328125,
0.10479736328125,
0.5361328125,
0.11871337890625,
0.410888671875,
0.11614990234375
],
[
0.59716796875,
0.10479736328125,
0.69384765625,
0.10479736328125,
0.69384765625,
0.11614990234375,
0.59716796875,
0.11614990234375
],
[
0.4150390625,
0.11297607421875,
0.529296875,
0.11614990234375,
0.529296875,
0.1337890625,
0.4150390625,
0.1307373046875
],
[
0.26953125,
0.1060791015625,
0.306396484375,
0.1060791015625,
0.306396484375,
0.117431640625,
0.26953125,
0.117431640625
],
[
0.334228515625,
0.11871337890625,
0.36767578125,
0.11871337890625,
0.36767578125,
0.1300048828125,
0.334228515625,
0.1300048828125
],
[
0.5986328125,
0.1142578125,
0.69140625,
0.1142578125,
0.69140625,
0.1319580078125,
0.5986328125,
0.1319580078125
],
[
0.1748046875,
0.14453125,
0.82421875,
0.14453125,
0.82421875,
0.1578369140625,
0.1748046875,
0.1578369140625
],
[
0.26953125,
0.11932373046875,
0.30810546875,
0.11932373046875,
0.30810546875,
0.1331787109375,
0.26953125,
0.1331787109375
],
[
0.1732177734375,
0.1578369140625,
0.79736328125,
0.1590576171875,
0.79736328125,
0.17236328125,
0.1732177734375,
0.1717529296875
],
[
0.17724609375,
0.189453125,
0.82275390625,
0.189453125,
0.82275390625,
0.20263671875,
0.17724609375,
0.20263671875
],
[
0.17724609375,
0.2052001953125,
0.410888671875,
0.2052001953125,
0.410888671875,
0.21533203125,
0.17724609375,
0.21533203125
],
[
0.17236328125,
0.2259521484375,
0.484375,
0.2266845703125,
0.484375,
0.240478515625,
0.17236328125,
0.2392578125
],
[
0.176513671875,
0.251953125,
0.82177734375,
0.251953125,
0.82177734375,
0.26513671875,
0.176513671875,
0.26513671875
],
[
0.1732177734375,
0.26513671875,
0.7255859375,
0.26513671875,
0.7255859375,
0.279052734375,
0.1732177734375,
0.279052734375
],
[
0.1748046875,
0.285888671875,
0.5947265625,
0.287353515625,
0.5947265625,
0.300537109375,
0.1748046875,
0.2998046875
],
[
0.176513671875,
0.30810546875,
0.82275390625,
0.309326171875,
0.82275390625,
0.323974609375,
0.176513671875,
0.322509765625
],
[
0.17724609375,
0.32568359375,
0.6748046875,
0.32568359375,
0.6748046875,
0.3359375,
0.17724609375,
0.3359375
],
[
0.3896484375,
0.34521484375,
0.48193359375,
0.35107421875,
0.48046875,
0.36865234375,
0.38818359375,
0.363037109375
],
[
0.47216796875,
0.34033203125,
0.6044921875,
0.341552734375,
0.6044921875,
0.35595703125,
0.47216796875,
0.354736328125
],
[
0.80224609375,
0.349853515625,
0.826171875,
0.349853515625,
0.826171875,
0.363037109375,
0.80224609375,
0.363037109375
],
[
0.53759765625,
0.35791015625,
0.5439453125,
0.363037109375,
0.53759765625,
0.36865234375,
0.53125,
0.363037109375
],
[
0.176513671875,
0.371826171875,
0.8203125,
0.371826171875,
0.8203125,
0.385009765625,
0.176513671875,
0.385009765625
],
[
0.176513671875,
0.386474609375,
0.82275390625,
0.386474609375,
0.82275390625,
0.399658203125,
0.176513671875,
0.399658203125
],
[
0.17724609375,
0.40087890625,
0.82275390625,
0.40087890625,
0.82275390625,
0.4140625,
0.17724609375,
0.4140625
],
[
0.17724609375,
0.4140625,
0.81396484375,
0.4140625,
0.81396484375,
0.427978515625,
0.17724609375,
0.427978515625
],
[
0.176513671875,
0.429931640625,
0.82275390625,
0.4287109375,
0.82275390625,
0.443115234375,
0.176513671875,
0.4443359375
],
[
0.1748046875,
0.4443359375,
0.75341796875,
0.4462890625,
0.75341796875,
0.4609375,
0.1748046875,
0.458984375
],
[
0.176513671875,
0.468505859375,
0.5146484375,
0.468505859375,
0.5146484375,
0.479248046875,
0.176513671875,
0.479248046875
],
[
0.176513671875,
0.490478515625,
0.82275390625,
0.490478515625,
0.82275390625,
0.50390625,
0.176513671875,
0.50390625
],
[
0.176513671875,
0.50390625,
0.82177734375,
0.50390625,
0.82177734375,
0.517578125,
0.176513671875,
0.517578125
],
[
0.17724609375,
0.51953125,
0.82177734375,
0.51953125,
0.82177734375,
0.52978515625,
0.17724609375,
0.52978515625
],
[
0.176513671875,
0.5322265625,
0.71728515625,
0.5322265625,
0.71728515625,
0.54541015625,
0.176513671875,
0.54541015625
],
[
0.378173828125,
0.5517578125,
0.61669921875,
0.5517578125,
0.61669921875,
0.5693359375,
0.378173828125,
0.5693359375
],
[
0.8037109375,
0.5537109375,
0.82421875,
0.5537109375,
0.82421875,
0.56640625,
0.8037109375,
0.56640625
],
[
0.17724609375,
0.57568359375,
0.82275390625,
0.57568359375,
0.82275390625,
0.5888671875,
0.17724609375,
0.5888671875
],
[
0.176513671875,
0.5888671875,
0.82275390625,
0.5888671875,
0.82275390625,
0.60302734375,
0.176513671875,
0.60302734375
],
[
0.176513671875,
0.60302734375,
0.81787109375,
0.60302734375,
0.81787109375,
0.6162109375,
0.176513671875,
0.6162109375
],
[
0.176513671875,
0.6171875,
0.7587890625,
0.6171875,
0.7587890625,
0.630859375,
0.176513671875,
0.630859375
],
[
0.176513671875,
0.63818359375,
0.525390625,
0.6396484375,
0.525390625,
0.65087890625,
0.176513671875,
0.6494140625
],
[
0.176513671875,
0.6611328125,
0.82421875,
0.6611328125,
0.82421875,
0.6748046875,
0.176513671875,
0.6748046875
],
[
0.17724609375,
0.6767578125,
0.82177734375,
0.6767578125,
0.82177734375,
0.6875,
0.17724609375,
0.6875
],
[
0.17724609375,
0.68798828125,
0.82177734375,
0.68798828125,
0.82177734375,
0.7021484375,
0.17724609375,
0.7021484375
],
[
0.176513671875,
0.7021484375,
0.82275390625,
0.7021484375,
0.82275390625,
0.71630859375,
0.176513671875,
0.71630859375
],
[
0.17724609375,
0.71630859375,
0.82275390625,
0.71630859375,
0.82275390625,
0.72998046875,
0.17724609375,
0.72998046875
],
[
0.176513671875,
0.73095703125,
0.740234375,
0.73095703125,
0.740234375,
0.744140625,
0.176513671875,
0.744140625
],
[
0.38818359375,
0.75,
0.60888671875,
0.75048828125,
0.60888671875,
0.76806640625,
0.38818359375,
0.767578125
],
[
0.8037109375,
0.75244140625,
0.82421875,
0.75244140625,
0.82421875,
0.7666015625,
0.8037109375,
0.7666015625
],
[
0.176513671875,
0.77294921875,
0.615234375,
0.771484375,
0.615234375,
0.7861328125,
0.176513671875,
0.787109375
],
[
0.1732177734375,
0.79345703125,
0.406982421875,
0.79443359375,
0.406982421875,
0.8095703125,
0.1732177734375,
0.80810546875
],
[
0.176513671875,
0.818359375,
0.82275390625,
0.818359375,
0.82275390625,
0.83203125,
0.176513671875,
0.83203125
],
[
0.176513671875,
0.83203125,
0.646484375,
0.83203125,
0.646484375,
0.84521484375,
0.176513671875,
0.84521484375
],
[
0.176513671875,
0.85498046875,
0.82275390625,
0.85498046875,
0.82275390625,
0.86865234375,
0.176513671875,
0.86865234375
],
[
0.17724609375,
0.87060546875,
0.82177734375,
0.87060546875,
0.82177734375,
0.880859375,
0.17724609375,
0.880859375
],
[
0.193603515625,
0.89794921875,
0.82421875,
0.8984375,
0.82421875,
0.91357421875,
0.193603515625,
0.912109375
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.9609375,
0.49267578125,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2023",
"Unbiased? Avg. Variance (shared)Avg. Variance (independent)",
"Yes",
"pqt2V+ (pq+1)S",
"pqV+tis",
"2+(q+1)s",
"gw()",
"Yes",
"+",
"Table l: Comparing weight(gu)and activity(ga)perturbation.V=dimension-wise avg.gradient",
"ga()",
"variance, S=dimension-wise avg.squared gradient norm; p=fan-in; q =fan-out; N =batch size",
"且活动扰动战胜了另一个因素的权重扰动。在我们的工作中，我们表明与权重扰动相比，活",
"动扰动具有较低的梯度估计方差。",
"3FORWARDGRADIENTLEARNING",
"在本节中，我们将回顾并建立学习算法的技术背景。我们首先回顾了正向模式自动微分",
"（AD）技术。其次，我们在权重空间或活动空间中提出了两种不同类型的扰动",
"3.1FORWARD-MODEAUTOMATICDIFFERENTIATION(AD)",
"令f：Rm→Rn。f的雅可比行列式J是大小为n×m的矩阵。正向模式AD计算矩阵向",
"量乘积JU，其中ERm。它被定义为在&上沿U评估的方向梯度：",
"Jyu := lim",
"f(α+du)-f(α)",
"(1)",
"S",
"为了进行比较，backprop（也称为反向模式AD）计算向量雅可比乘积uJt，其中uER\"",
"对应于链规则中的最后一项。与反向模式AD相比，正向模式AD只需要一次正向通过，这",
"是用衍生信息增强的。要计算计算图中节点的雅可比向量积，首先将使用扩展输入节点，",
"该是要相乘的向量。然后对于其他节点，我们将（α，'）的元组作为输人发送，并计算-",
"个元组（y,y'）作为输出，其中a'和y是节点和节点y的中间导数，i.e.y=翌a之间",
"是MAXXX。在JAX库（Bradburyetal.，2018）中，正向模式AD实现为jax.jvp",
"3.2WEIGHT-PERTURBEDFORWARDGRADIENT",
"生成权重更新的权重扰动最初是在（Bartoetal.，1983;Xie&Seung，1999；Seung，2003）中探",
"索的。Baydinetal.(2022）使用正向模式AD的技术来实现权重扰动，在数值稳定性方面优",
"于有限差分。令W为单元和之间的权重连接，f为损失函数。我们可以通过从零均值",
"单位方差高斯分布中提取具有id元素Uii的随机矩阵来估计梯度。估计器为",
"gu(wis)=(EijVwigUnj)vi",
"(2)",
"直观上，该估计器对随机扰动方向U进行采样，并通过使用正向模式执行点积来测试其如",
"何与真实梯度Vwii对齐，然后将标量对齐与扰动方向再次相乘。Baydinetal.(2022）将使",
"用正向模式AD的这种形式的梯度估计称为“正向梯度”。为了与我们稍后详细描述的另一",
"种形式的扰动区分开，我们将其称为“权重扰动前向梯度”，或简称为“权重扰动”",
"3.3ACTIVITY-PERTURBEDFORWARDGRADIENT",
"扰动权重的另一种方法是扰动活动，这可以减少每个示例的扰动维数。在限制性假设下，最",
"初在LeCunetal.（1988）：Widrow&Lehr（1990）中探索了活动扰动。在这里，我们介绍一种",
"使用活动扰动估计梯度的一般方法。它在生物学上可能是合理的，因为它可能对应于每个",
"神经元E(Fiete&Seung，2006)中电导的扰动。。在这里，为了简单起见，我们专注于基于",
"离散时间速率的公式。令表示第i个突触前神经元的活动，之；表示非线性激活函数之前",
"的第i个突触后神经元的活动，u；是z的扰动。活动扰动的前向梯度估计器是",
"ga(wi)= ri (g Vzgruj)uj,",
"(3)",
"，其中√z和u之间的内积再次通过使用前向模式AD来计算",
"3.4THEORETICALPROPERTIES",
"在本节中，我们旨在分析正向梯度估计器的期望和方差属性。我们将分析重点放在一个权",
"重矩阵【Wi}的梯度上，但结论也适用于具有多个权重矩阵的网络",
"表1总结了1的理论结果。在批量大小为N的情况下，独立扰动可以实现1/N的方差减",
"小，而共享扰动具有由平方梯度范数主导的恒定方差项。但是，在执行独立权重扰动时，不",
"'All proofs can be found in Appendix ?? and 9.Numerical simulation results can be found in Appendix 10",
"3"
] |
{
"arxiv_id": "2210.03310",
"language": "zh",
"page_idx": 2
}
|
||
2210.03310-zh-3
|
2210.03310-zh-3.jpg
|
Published as a conference paper at ICLR 2023
<FigureHere>
<FigureHere>
Figure 2: A LocalMixer residual block with local losses. Token mixing consists of a linear layer
and channels are grouped in the channel mixing layers. Layer norm is applied before and after every linear layer. LN=Layer Norm; FC=Fully Connected layer; A=Activation function (ReLU); T= Transpose.
能对矩阵乘法进行批处理，因为每个示例的激活向量都与不同的权重矩阵相乘。相比之下， 独立活动扰动允许批处理矩阵乘法。此外，与权重扰动相比，活动扰动具有扇人(p)倍的方差，因为扰动元素的数量是输出单元的数量，而不是整个权重矩阵的大小。活动扰动的唯一缺点是存储中间激活所需的内存，以换取$Np$方差减少的因素。然而，对于活动和权重扰动，方差仍然随着更大的网络而增长。在 4 节中，我们将通过引人局部损失函数来进一步减小方差
3.5 CONTINUOUS-TIME RATE-BASED MODELS
正向模式 AD 可以看作是在连续时间物理系统中计算一阶时间导数。假设在计算图的节点之间传递的元组是$(x,\dot{x})$ ,其中$\dot{x}$是$x$随时间的变化。然后，计算与正向模式 AD 相同。对于每个节点，$\dot{y}=\frac{d\boldsymbol{y}}{dx}\dot{x}$,其中$\frac d\boldsymbol{y}{dx}$是输出和输人之间的雅可比行列式。请注意，在物理系统中，我们不必通过运行两个正向传递来显式执行微分运算。相反，一阶导数信息在模拟信号中很容易获得，我们只需要将输出信号插入微分器电路
因此，连续时间系统的活动扰动学习规则是$\dot{w}_{ij}\propto x_i\dot{y}_j\dot{r}$,其中$x_i$是突触前活动$,\dot{y}_j$是突触后活动的变化率，这是一小段时间的扰动方向，$\ddot r$是奖励的负变化率(或。奖励控制学习是Hebbian 还是反 Hebbian。Hinton et al.(2007) 和 Bengio et al. (2017)都建议使用突触前活动与突触后活动变化率的乘积。但是，他们没有考虑使用奖励变化率作为调制器，而是依靠另一组反馈权重来通过输入传达误差信号。相比之下，我们表明，通过广播奖励的变化率，我们实际上可以绕过重量运输问题
3.6 ACTIVATION SPARSITY AND NORMALIZATION FUNCTIONS
在具有 ReLU 激活的网络中，我们可以利用 ReLU 稀疏性来实现进一步的方差减小，因为未
激活的单元将具有零梯度，因此我们不应该扰动这些单元，并将扰动设置为零
标准化层通常在深度神经网络中添加在线性层之后。为了计算活动扰动中的正确梯度，我们还需要在权重更新规则中考虑归一化。由于没有向后权重连接，一种选择是简单地在归一化层上应用 backprop。但是，我们还发现，在使用层归一化时忽略归一化层的梯度也是可以的。
4 SCALING WITH LOCAL LossES
正如我们在上一节中所解释的那样，扰动学习可能会遭受维数的诅咒：方差随着扰动维数的增加而增长，在深度网络中，通常有数百万个参数同时变化。限制可学习维度数量的一种方法是将网络划分为子模块，每个子模块都有一个单独的损失函数。在本节中，我们将探讨增加局部损失数量以驯服方差的几种方法
$1) \textbf{ Blockwise loss. 首 先 , 我 们 将 网 络 深 人 划 分 为 模 块 。每 个 模 块 由 几 层 组 成 。在 每 个 模 块 的 }$ 末尾，我们计算一个损失函数，该损失用于更新该模块中的参数。这种方法等效于在模块之间添加“停止梯度”运算符。以前在 Belilovsky et al.(2019)和 Löwe et al.(2019) 中探讨了这种局部贪婪损失。
4
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.31787109375,
0.08837890625,
0.3447265625,
0.08837890625,
0.3447265625,
0.0997314453125,
0.31787109375,
0.0997314453125
],
[
0.51318359375,
0.08837890625,
0.54345703125,
0.08837890625,
0.54345703125,
0.0997314453125,
0.51318359375,
0.0997314453125
],
[
0.74169921875,
0.08837890625,
0.77685546875,
0.08837890625,
0.77685546875,
0.0997314453125,
0.74169921875,
0.0997314453125
],
[
0.31787109375,
0.0997314453125,
0.343017578125,
0.0997314453125,
0.343017578125,
0.10797119140625,
0.31787109375,
0.10797119140625
],
[
0.51318359375,
0.0997314453125,
0.5439453125,
0.0997314453125,
0.5439453125,
0.10797119140625,
0.51318359375,
0.10797119140625
],
[
0.740234375,
0.0985107421875,
0.78515625,
0.0985107421875,
0.78515625,
0.10986328125,
0.740234375,
0.10986328125
],
[
0.223876953125,
0.12060546875,
0.27294921875,
0.12249755859375,
0.271240234375,
0.1351318359375,
0.22216796875,
0.1331787109375
],
[
0.50341796875,
0.11553955078125,
0.5087890625,
0.117431640625,
0.50341796875,
0.125,
0.5,
0.12249755859375
],
[
0.32177734375,
0.121826171875,
0.370849609375,
0.12371826171875,
0.370849609375,
0.1351318359375,
0.32177734375,
0.1331787109375
],
[
0.381591796875,
0.12371826171875,
0.43310546875,
0.12371826171875,
0.43310546875,
0.1337890625,
0.381591796875,
0.1337890625
],
[
0.560546875,
0.121826171875,
0.5986328125,
0.12371826171875,
0.59716796875,
0.1337890625,
0.5595703125,
0.1319580078125
],
[
0.4345703125,
0.1300048828125,
0.5009765625,
0.1300048828125,
0.5009765625,
0.141357421875,
0.4345703125,
0.141357421875
],
[
0.61767578125,
0.1275634765625,
0.66259765625,
0.1275634765625,
0.66259765625,
0.1307373046875,
0.61767578125,
0.1307373046875
],
[
0.68701171875,
0.1300048828125,
0.7294921875,
0.1300048828125,
0.7294921875,
0.141357421875,
0.68701171875,
0.141357421875
],
[
0.27783203125,
0.1307373046875,
0.3095703125,
0.1307373046875,
0.3095703125,
0.1427001953125,
0.27783203125,
0.1427001953125
],
[
0.2181396484375,
0.1319580078125,
0.281005859375,
0.1351318359375,
0.279296875,
0.149658203125,
0.21728515625,
0.146484375
],
[
0.318603515625,
0.1331787109375,
0.4345703125,
0.1331787109375,
0.4345703125,
0.15087890625,
0.318603515625,
0.15087890625
],
[
0.496826171875,
0.1319580078125,
0.51318359375,
0.1319580078125,
0.51318359375,
0.14013671875,
0.496826171875,
0.14013671875
],
[
0.56201171875,
0.1351318359375,
0.5908203125,
0.1351318359375,
0.5908203125,
0.1427001953125,
0.56201171875,
0.1427001953125
],
[
0.58837890625,
0.136962890625,
0.625,
0.136962890625,
0.625,
0.147705078125,
0.58837890625,
0.147705078125
],
[
0.61669921875,
0.1337890625,
0.65869140625,
0.1337890625,
0.65869140625,
0.14453125,
0.61669921875,
0.14453125
],
[
0.178955078125,
0.1673583984375,
0.81640625,
0.1673583984375,
0.81640625,
0.177978515625,
0.178955078125,
0.177978515625
],
[
0.2548828125,
0.189453125,
0.2900390625,
0.189453125,
0.2900390625,
0.1962890625,
0.2548828125,
0.1962890625
],
[
0.4345703125,
0.189453125,
0.465576171875,
0.189453125,
0.465576171875,
0.1962890625,
0.4345703125,
0.1962890625
],
[
0.6708984375,
0.208984375,
0.7900390625,
0.208984375,
0.7900390625,
0.250732421875,
0.6708984375,
0.250732421875
],
[
0.24267578125,
0.276611328125,
0.279296875,
0.276611328125,
0.279296875,
0.2841796875,
0.24267578125,
0.2841796875
],
[
0.410888671875,
0.276611328125,
0.44677734375,
0.276611328125,
0.44677734375,
0.2841796875,
0.410888671875,
0.2841796875
],
[
0.55224609375,
0.276611328125,
0.583984375,
0.276611328125,
0.583984375,
0.2841796875,
0.55224609375,
0.2841796875
],
[
0.6337890625,
0.2890625,
0.6650390625,
0.2890625,
0.6650390625,
0.296630859375,
0.6337890625,
0.296630859375
],
[
0.214111328125,
0.27978515625,
0.2344970703125,
0.27978515625,
0.2344970703125,
0.28271484375,
0.214111328125,
0.28271484375
],
[
0.1748046875,
0.3017578125,
0.82275390625,
0.302978515625,
0.82275390625,
0.317626953125,
0.1748046875,
0.31640625
],
[
0.1748046875,
0.317626953125,
0.82275390625,
0.317626953125,
0.82275390625,
0.330810546875,
0.1748046875,
0.330810546875
],
[
0.1748046875,
0.330810546875,
0.82421875,
0.330810546875,
0.82421875,
0.3447265625,
0.1748046875,
0.3447265625
],
[
0.1732177734375,
0.34228515625,
0.265625,
0.34521484375,
0.265625,
0.359130859375,
0.1732177734375,
0.35595703125
],
[
0.1748046875,
0.37255859375,
0.82275390625,
0.373779296875,
0.82275390625,
0.386962890625,
0.1748046875,
0.386474609375
],
[
0.176513671875,
0.386962890625,
0.82275390625,
0.386962890625,
0.82275390625,
0.40087890625,
0.176513671875,
0.40087890625
],
[
0.176513671875,
0.40087890625,
0.82275390625,
0.40087890625,
0.82275390625,
0.4140625,
0.176513671875,
0.4140625
],
[
0.1829833984375,
0.41845703125,
0.1895751953125,
0.41845703125,
0.1895751953125,
0.42236328125,
0.1829833984375,
0.42236328125
],
[
0.188720703125,
0.415283203125,
0.82275390625,
0.415283203125,
0.82275390625,
0.4287109375,
0.188720703125,
0.4287109375
],
[
0.176513671875,
0.4287109375,
0.82275390625,
0.4287109375,
0.82275390625,
0.442626953125,
0.176513671875,
0.442626953125
],
[
0.1748046875,
0.4443359375,
0.240234375,
0.4443359375,
0.240234375,
0.455078125,
0.1748046875,
0.455078125
],
[
0.1748046875,
0.46337890625,
0.51123046875,
0.46337890625,
0.51123046875,
0.4765625,
0.1748046875,
0.4765625
],
[
0.176513671875,
0.4873046875,
0.82177734375,
0.4873046875,
0.82177734375,
0.50048828125,
0.176513671875,
0.50048828125
],
[
0.176513671875,
0.50048828125,
0.82275390625,
0.50048828125,
0.82275390625,
0.5146484375,
0.176513671875,
0.5146484375
],
[
0.176513671875,
0.5166015625,
0.82275390625,
0.5166015625,
0.82275390625,
0.52978515625,
0.176513671875,
0.52978515625
],
[
0.176513671875,
0.5322265625,
0.82275390625,
0.5322265625,
0.82275390625,
0.54541015625,
0.176513671875,
0.54541015625
],
[
0.176513671875,
0.546875,
0.56884765625,
0.546875,
0.56884765625,
0.56005859375,
0.176513671875,
0.56005859375
],
[
0.176513671875,
0.5673828125,
0.82275390625,
0.5673828125,
0.82275390625,
0.58203125,
0.176513671875,
0.58203125
],
[
0.176513671875,
0.58056640625,
0.82275390625,
0.58056640625,
0.82275390625,
0.5947265625,
0.176513671875,
0.5947265625
],
[
0.176513671875,
0.5966796875,
0.82177734375,
0.5966796875,
0.82177734375,
0.60693359375,
0.176513671875,
0.60693359375
],
[
0.176513671875,
0.609375,
0.810546875,
0.609375,
0.810546875,
0.62255859375,
0.176513671875,
0.62255859375
],
[
0.176513671875,
0.62255859375,
0.82275390625,
0.62255859375,
0.82275390625,
0.63623046875,
0.176513671875,
0.63623046875
],
[
0.17724609375,
0.63818359375,
0.386474609375,
0.63818359375,
0.386474609375,
0.6484375,
0.17724609375,
0.6484375
],
[
0.1732177734375,
0.6572265625,
0.619140625,
0.65771484375,
0.619140625,
0.671875,
0.1732177734375,
0.67041015625
],
[
0.176513671875,
0.68115234375,
0.82275390625,
0.68115234375,
0.82275390625,
0.6943359375,
0.176513671875,
0.6943359375
],
[
0.1748046875,
0.693359375,
0.7294921875,
0.693359375,
0.7294921875,
0.70703125,
0.1748046875,
0.70703125
],
[
0.176513671875,
0.71630859375,
0.82275390625,
0.71630859375,
0.82275390625,
0.72998046875,
0.176513671875,
0.72998046875
],
[
0.176513671875,
0.72998046875,
0.82275390625,
0.72998046875,
0.82275390625,
0.74365234375,
0.176513671875,
0.74365234375
],
[
0.1895751953125,
0.74365234375,
0.82177734375,
0.74365234375,
0.82177734375,
0.75830078125,
0.1895751953125,
0.75830078125
],
[
0.176513671875,
0.7587890625,
0.2156982421875,
0.7587890625,
0.2156982421875,
0.77099609375,
0.176513671875,
0.77099609375
],
[
0.17236328125,
0.77978515625,
0.47216796875,
0.7822265625,
0.47216796875,
0.79541015625,
0.17236328125,
0.79345703125
],
[
0.17724609375,
0.80615234375,
0.82275390625,
0.80615234375,
0.82275390625,
0.8193359375,
0.17724609375,
0.8193359375
],
[
0.17724609375,
0.8212890625,
0.82177734375,
0.8212890625,
0.82177734375,
0.83203125,
0.17724609375,
0.83203125
],
[
0.176513671875,
0.833984375,
0.82275390625,
0.833984375,
0.82275390625,
0.84765625,
0.176513671875,
0.84765625
],
[
0.176513671875,
0.84765625,
0.4697265625,
0.84765625,
0.4697265625,
0.861328125,
0.176513671875,
0.861328125
],
[
0.1805419921875,
0.869140625,
0.82177734375,
0.869140625,
0.82177734375,
0.88330078125,
0.1805419921875,
0.88330078125
],
[
0.17724609375,
0.88330078125,
0.8203125,
0.88330078125,
0.8203125,
0.896484375,
0.17724609375,
0.896484375
],
[
0.176513671875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.176513671875,
0.9111328125,
0.298095703125,
0.9111328125,
0.298095703125,
0.9248046875,
0.176513671875,
0.9248046875
],
[
0.491943359375,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.96044921875,
0.491943359375,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2023",
"Local",
"Local",
"Local",
"Losses",
"Losses",
"Losses",
"Fchannel",
"",
"fToken",
"channel",
"Token",
"++-A",
"channa",
"+)→(A",
"A",
"Mixing",
"MixingMixing",
"..",
"Mixin",
"19#",
"Mixing",
"Figurel:ALocalMixernetworkconsistsof severalmixerblocks.A=Activationfunction(ReLU)",
"TokenM",
"nelMix",
"",
"N.XCXHW",
"NXHWX.C",
"NXHWX",
"+HA",
"广",
"Figure 2: A LocalMixer residual block with local losses. Token mixing consists of a linear layer",
"and channels are grouped in the channel mixing layers. Layer norm is applied before and after",
"every linear layer. LN=Layer Norm; FC-Fully Connected layer; A=Activation function (ReLU);",
"T=Transpose.",
"能对矩阵乘法进行批处理，因为每个示例的激活向量都与不同的权重矩阵相乘。相比之下",
"独立活动扰动允许批处理矩阵乘法。此外，与权重扰动相比，活动扰动具有扇入（p）倍的",
"方差，因为扰动元素的数量是输出单元的数量，而不是整个权重矩阵的大小。活动扰动的唯",
"",
"-缺点是存储中间激活所需的内存，以换取Np方差减少的因素。然而，对于活动和权重扰",
"动，方差仍然随着更大的网络而增长。在4节中，我们将通过引人局部损失函数来进一步",
"减小方差",
"3.5CONTINUOUS-TIME RATE-BASED MODELS",
"正向模式AD可以看作是在连续时间物理系统中计算一阶时间导数。假设在计算图的节点",
"之间传递的元组是（α,），其中ε是a随时间的变化。然后，计算与正向模式AD相同。对",
"于每个节点，9三，其中是输出和输入之间的雅可比行列式。请注意，在物理系统",
"中，我们不必通过运行两个正向传递来显式执行微分运算。相反，一阶导数信息在模拟信号",
"中很容易获得，我们只需要将输出信号插人微分器电路",
"因此，连续时间系统的活动扰动学习规则是i号αci,r，其中a,是突触前活动，9i是突触",
"后活动的变化率，这是一小段时间的扰动方向，是奖励的负变化率（或。奖励控制学习是",
"Hebbian还是反Hebbian。Hintonetal.（2007）和Bengioetal.（2017）都建议使用突触前活动与",
"突触后活动变化率的乘积。但是，他们没有考虑使用奖励变化率作为调制器，而是依靠另",
"组反馈权重来通过输入传达误差信号。相比之下，我们表明，通过广播奖励的变化率，我们",
"实际上可以绕过重量运输间题",
"3.6ACTIVATIONSPARSITYANDNORMALIZATIONFUNCTIONS",
"在具有ReLU激活的网络中，我们可以利用ReLU稀疏性来实现进一步的方差减小，因为未",
"激活的单元将具有零梯度，因此我们不应该扰动这些单元，并将扰动设置为零",
"标准化层通常在深度神经网络中添加在线性层之后。为了计算活动扰动中的正确梯度，我",
"们还需要在权重更新规则中考虑归一化。由于没有向后权重连接，一种选择是简单地在归",
"化层上应用backprop。但是，我们还发现，在使用层归一化时忽略归一化层的梯度也是可",
"以的。",
"4SCALINGWITHLOCALLOSSES",
"正如我们在上一节中所解释的那样，扰动学习可能会遭受维数的诅咒：方差随着扰动维数",
"的增加而增长，在深度网络中，通常有数百万个参数同时变化。限制可学习维度数量的一种",
"方法是将网络划分为子模块，每个子模块都有一个单独的损失函数。在本节中，我们将探讨",
"增加局部损失数量以驯服方差的几种方法",
"1)Blockwiseloss。首先，我们将网络深人划分为模块。每个模块由儿层组成。在每个模块的",
"末尾，我们计算一个损失函数，该损失用于更新该模块中的参数。这种方法等效于在模块之",
"间添加“停止梯度”运算符。以前在Belilovskyetal.(2019）和Loweetal.（2019）中探讨了这",
"种局部贪婪损失。",
"4"
] |
{
"arxiv_id": "2210.03310",
"language": "zh",
"page_idx": 3
}
|
||
2210.03310-zh-4
|
2210.03310-zh-4.jpg
|
Published as a conference paper at ICLR 2023
<FigureHere>
Figure 3: Feature aggregator designs. A) In the conventional design, average pooling is performed to aggregate features from different spatial locations. B) We propose the replicated design, features are first concatenated across groups and then averaged across spatial locations. We create copies of the same feature with different stop gradient masks so that we obtain more local losses instead of a global one. The stop gradient mask makes sure that perturbation in one spatial group corresponds to its loss function. The numerical value of the loss function is the same as the conventional design.
2)Patchwise loss. 图像等感官输人信号具有空间维度。我们将沿着这些空间维度逐段应用单独的损失。在 VisionTransformer 架构 (Vaswani et al.,2017; Dosovitskiy et al.,2021)中，每个空间标记表示图像中的一个补丁。在现代深度网络中，通常会共享每个空间位置的参数， 以提高数据效率并降低内存带宽利用率。虽然天真的重量共享在生物学上是不合理的，但我们仍然认为在这项工作中共享重量。通过在补丁之间添加知识蒸馏$\overline{\mathrm{E}}($Hinton et al., 2015) 损失，可以模拟权重共享的效果。
3)Groupwise loss. 最后，我们转向通道维度。为了创建多个损失，我们将通道拆分为多个组，每个组都附加到一个损失函数 (Patel et al.,2022)。为防止组之间相互通信，通道仅连接使用通道组来提高计算效率(Krizhevskv et al. 2012: Ioannou et al. 2017: Xie et al. 2017). 而在我们的工作中，添加组会增加损失总数，从而减少方差
Feature aggregators. 朴素地将损失分别应用于空间和通道维度会导致次优的性能，因为每个维度仅包含本地信息。对于标准任务(如分类)的损失，模型需要输入的全局视图来做出决策。标准体系结构通过在最终分类层之前执行全局平均池化层来获得此全局视图。因此， 我们探索在局部损失函数之前聚合来自其他组和空间斑块的信息的策略。
我们希望在不减少维度总数的情况下执行聚合。因此，我们提出了一种用于特征聚合的复制设计，如图 3 所示。首先，通道组被复制并相互通信，但是除了活动组本身之外的每个组都被停止梯度屏蔽，以便其他组不影响前向梯度计算：
(4)
$${\boldsymbol{x}}_{p,g}=[\mathrm{StopGrad}(x_{p,1}...x_{p,g-1}),x_{p,g},\mathrm{StopGrad}(x_{p,g+1},...,x_{p,G})],$$
其中$p$和$g$分别索引补丁和组。同样，每个空间位置也被复制，通信和屏蔽，然后在本地平
均：
(5)
$$\overline{\boldsymbol{x}}_{p,g}=\frac{1}{P}\left(\boldsymbol{x}_{p,g}+\sum_{p^{\prime}\neq p}\mathrm{StopGrad}(\boldsymbol{x}_{p^{\prime},g})\right).$$
特征聚合的输出与常规全局平均池化层的输出相同。不同的是，这里的损失是重复的，并且
在每个损失中激活不同的补丁组
Learning objectives. 我们考虑监督分类损失和对比信息损失$\mathbb{E}($van den Oord et al., 2018; Chen et al., 2020), 这是图像表示学习中最常用的两种损失。对于监督分类，我们在Cinel et al. , 2020) , 这 定 图 像 农 亦 手 刁 干 取 吊 用 的 网 件 顶 大 。对 丁 m首 方 失 , 我 们 在 聚 合 特 征 的 顶 部 附 加 一 个 共 享 线 性 层 (跨 $p, g$ 共 享 ) , 以 实 现 交 叉 熵 损 失 : $L_{p, g}^{s}=$ 绿 合 特 征 的 顶 部 附 加 一 个 共 享 线 性 层 (跨 $p, g$ 共 享 ) , 以 实 现 交 叉 熵 损 失 : $L_{p, g}^{s}=$ $-\sum_kt_k\log$softmax$(W_l\overline{\boldsymbol{x}}_{p,g})_k$。每个组和修补程序位置的损失值相同
对于对比学习，线性层成为线性特征投影仪。假设$x_n^{(1)}$和$x_n^{(2)}$ 是$n$示例的两个不同视图，则对比学习的 InfoNCE 损失为：
$$L_{p,g}^{c}=-\sum_{n}\log\frac{(W\overline{\boldsymbol x}_{n,p,g}^{(1)})^{\top}\mathrm{StopGrad}(W\overline{\boldsymbol x}_{n}^{(2)})}{\sum_{m}(W\overline{\boldsymbol x}_{n,p,g}^{(1)})^{\top}\mathrm{StopGrad}(W\overline{\boldsymbol x}_{m}^{(2)})}.$$
(6)
<FigureHere>
请注意，我们在第二个视图上添加了停止梯度运算符。通常没有必要在 InfoNCE 损失中添加此停止梯度；但是，我们发现基于扰动的方法需要停止梯度，否则损失不会下降。这可能是因为我们在两个视图上共享扰动，
5
Figure 4: Importance of StopGradient in the InfoNCE loss, using M/8 on CIFAR-10 with 256 channels 1 group.
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.2467041015625,
0.0770263671875,
0.343017578125,
0.0770263671875,
0.343017578125,
0.08837890625,
0.2467041015625,
0.08837890625
],
[
0.42724609375,
0.080810546875,
0.48046875,
0.080810546875,
0.48046875,
0.091552734375,
0.42724609375,
0.091552734375
],
[
0.54736328125,
0.0802001953125,
0.57421875,
0.0802001953125,
0.57421875,
0.091552734375,
0.54736328125,
0.091552734375
],
[
0.42724609375,
0.09783935546875,
0.56787109375,
0.09783935546875,
0.56787109375,
0.111083984375,
0.42724609375,
0.111083984375
],
[
0.439453125,
0.11236572265625,
0.48046875,
0.11236572265625,
0.48046875,
0.12371826171875,
0.439453125,
0.12371826171875
],
[
0.2484130859375,
0.126953125,
0.328369140625,
0.126953125,
0.328369140625,
0.1383056640625,
0.2484130859375,
0.1383056640625
],
[
0.2459716796875,
0.146484375,
0.408447265625,
0.1458740234375,
0.408447265625,
0.1572265625,
0.2459716796875,
0.1578369140625
],
[
0.55810546875,
0.1483154296875,
0.65185546875,
0.1483154296875,
0.65185546875,
0.1558837890625,
0.55810546875,
0.1558837890625
],
[
0.67333984375,
0.147705078125,
0.74609375,
0.147705078125,
0.74609375,
0.1578369140625,
0.67333984375,
0.1578369140625
],
[
0.64453125,
0.1654052734375,
0.6640625,
0.1654052734375,
0.6640625,
0.17236328125,
0.64453125,
0.17236328125
],
[
0.250732421875,
0.173583984375,
0.261474609375,
0.173583984375,
0.261474609375,
0.18310546875,
0.250732421875,
0.18310546875
],
[
0.35693359375,
0.177978515625,
0.371826171875,
0.177978515625,
0.371826171875,
0.1817626953125,
0.35693359375,
0.1817626953125
],
[
0.490234375,
0.1767578125,
0.498291015625,
0.1749267578125,
0.5,
0.18115234375,
0.49267578125,
0.1817626953125
],
[
0.2548828125,
0.1881103515625,
0.2607421875,
0.1881103515625,
0.2607421875,
0.1944580078125,
0.2548828125,
0.1944580078125
],
[
0.494384765625,
0.1881103515625,
0.5,
0.1881103515625,
0.5,
0.1944580078125,
0.494384765625,
0.1944580078125
],
[
0.6396484375,
0.189453125,
0.646484375,
0.189453125,
0.646484375,
0.1932373046875,
0.6396484375,
0.1932373046875
],
[
0.261474609375,
0.199462890625,
0.275390625,
0.2008056640625,
0.273681640625,
0.2083740234375,
0.2607421875,
0.20703125
],
[
0.62158203125,
0.1988525390625,
0.6650390625,
0.1988525390625,
0.6650390625,
0.208984375,
0.62158203125,
0.208984375
],
[
0.339111328125,
0.2108154296875,
0.43701171875,
0.2108154296875,
0.43701171875,
0.2215576171875,
0.339111328125,
0.2215576171875
],
[
0.5556640625,
0.2108154296875,
0.6689453125,
0.2108154296875,
0.6689453125,
0.2247314453125,
0.5556640625,
0.2247314453125
],
[
0.176513671875,
0.2291259765625,
0.82421875,
0.2291259765625,
0.82421875,
0.24365234375,
0.176513671875,
0.24365234375
],
[
0.1748046875,
0.24365234375,
0.82421875,
0.24365234375,
0.82421875,
0.2568359375,
0.1748046875,
0.2568359375
],
[
0.1748046875,
0.2568359375,
0.82421875,
0.2568359375,
0.82421875,
0.270751953125,
0.1748046875,
0.270751953125
],
[
0.1748046875,
0.271484375,
0.82421875,
0.271484375,
0.82421875,
0.285400390625,
0.1748046875,
0.285400390625
],
[
0.176513671875,
0.285400390625,
0.82421875,
0.285400390625,
0.82421875,
0.298583984375,
0.176513671875,
0.298583984375
],
[
0.1748046875,
0.29736328125,
0.8037109375,
0.29736328125,
0.8037109375,
0.311279296875,
0.1748046875,
0.311279296875
],
[
0.178955078125,
0.3125,
0.82177734375,
0.3125,
0.82177734375,
0.32568359375,
0.178955078125,
0.32568359375
],
[
0.17724609375,
0.32568359375,
0.82275390625,
0.32568359375,
0.82275390625,
0.339111328125,
0.17724609375,
0.339111328125
],
[
0.176513671875,
0.33837890625,
0.82177734375,
0.33837890625,
0.82177734375,
0.3515625,
0.176513671875,
0.3515625
],
[
0.176513671875,
0.354248046875,
0.82275390625,
0.354248046875,
0.82275390625,
0.367431640625,
0.176513671875,
0.367431640625
],
[
0.1748046875,
0.3662109375,
0.82275390625,
0.3662109375,
0.82275390625,
0.380126953125,
0.1748046875,
0.380126953125
],
[
0.176513671875,
0.380615234375,
0.410888671875,
0.380615234375,
0.410888671875,
0.39453125,
0.176513671875,
0.39453125
],
[
0.178955078125,
0.401611328125,
0.82275390625,
0.401611328125,
0.82275390625,
0.415283203125,
0.178955078125,
0.415283203125
],
[
0.17724609375,
0.417236328125,
0.82177734375,
0.417236328125,
0.82177734375,
0.427978515625,
0.17724609375,
0.427978515625
],
[
0.1748046875,
0.4287109375,
0.82421875,
0.4287109375,
0.82421875,
0.4462890625,
0.1748046875,
0.4462890625
],
[
0.176513671875,
0.443115234375,
0.82421875,
0.443115234375,
0.82421875,
0.45703125,
0.176513671875,
0.45703125
],
[
0.17724609375,
0.458984375,
0.583984375,
0.458984375,
0.583984375,
0.4697265625,
0.17724609375,
0.4697265625
],
[
0.178955078125,
0.479248046875,
0.82177734375,
0.477783203125,
0.82177734375,
0.492431640625,
0.178955078125,
0.49365234375
],
[
0.17724609375,
0.492431640625,
0.82177734375,
0.492431640625,
0.82177734375,
0.505859375,
0.17724609375,
0.505859375
],
[
0.176513671875,
0.5068359375,
0.82275390625,
0.5068359375,
0.82275390625,
0.52099609375,
0.176513671875,
0.52099609375
],
[
0.17724609375,
0.52294921875,
0.68701171875,
0.52294921875,
0.68701171875,
0.53271484375,
0.17724609375,
0.53271484375
],
[
0.176513671875,
0.54248046875,
0.82275390625,
0.54248046875,
0.82275390625,
0.55615234375,
0.176513671875,
0.55615234375
],
[
0.17724609375,
0.55615234375,
0.82275390625,
0.55615234375,
0.82275390625,
0.5693359375,
0.17724609375,
0.5693359375
],
[
0.1748046875,
0.5693359375,
0.576171875,
0.5693359375,
0.576171875,
0.58349609375,
0.1748046875,
0.58349609375
],
[
0.26953125,
0.5859375,
0.72802734375,
0.5859375,
0.72802734375,
0.60107421875,
0.26953125,
0.60107421875
],
[
0.80224609375,
0.5859375,
0.82421875,
0.5859375,
0.82421875,
0.5986328125,
0.80224609375,
0.5986328125
],
[
0.17724609375,
0.60400390625,
0.82177734375,
0.60400390625,
0.82177734375,
0.6171875,
0.17724609375,
0.6171875
],
[
0.178955078125,
0.6142578125,
0.200927734375,
0.619140625,
0.1968994140625,
0.63330078125,
0.1732177734375,
0.62744140625
],
[
0.3447265625,
0.63720703125,
0.63623046875,
0.63720703125,
0.63623046875,
0.65478515625,
0.3447265625,
0.65478515625
],
[
0.8037109375,
0.63818359375,
0.826171875,
0.63818359375,
0.826171875,
0.65283203125,
0.8037109375,
0.65283203125
],
[
0.176513671875,
0.6591796875,
0.82275390625,
0.6591796875,
0.82275390625,
0.67236328125,
0.176513671875,
0.67236328125
],
[
0.176513671875,
0.673828125,
0.404296875,
0.673828125,
0.404296875,
0.6875,
0.176513671875,
0.6875
],
[
0.1805419921875,
0.6943359375,
0.82275390625,
0.6943359375,
0.82275390625,
0.708984375,
0.1805419921875,
0.708984375
],
[
0.176513671875,
0.708984375,
0.82275390625,
0.708984375,
0.82275390625,
0.72265625,
0.176513671875,
0.72265625
],
[
0.17236328125,
0.7197265625,
0.82275390625,
0.7216796875,
0.82275390625,
0.7392578125,
0.17236328125,
0.7373046875
],
[
0.181396484375,
0.74365234375,
0.1895751953125,
0.74365234375,
0.1895751953125,
0.74853515625,
0.181396484375,
0.74853515625
],
[
0.1854248046875,
0.7392578125,
0.66650390625,
0.7392578125,
0.66650390625,
0.75244140625,
0.1854248046875,
0.75244140625
],
[
0.64453125,
0.76318359375,
0.74658203125,
0.7587890625,
0.74853515625,
0.7763671875,
0.646484375,
0.78076171875
],
[
0.736328125,
0.76025390625,
0.82421875,
0.75830078125,
0.82421875,
0.7763671875,
0.736328125,
0.77880859375
],
[
0.1748046875,
0.77783203125,
0.736328125,
0.7763671875,
0.736328125,
0.79150390625,
0.1748046875,
0.79248046875
],
[
0.176513671875,
0.7646484375,
0.65625,
0.7646484375,
0.65625,
0.77783203125,
0.176513671875,
0.77783203125
],
[
0.73046875,
0.78076171875,
0.74609375,
0.78076171875,
0.74609375,
0.78857421875,
0.73046875,
0.78857421875
],
[
0.74658203125,
0.77978515625,
0.82421875,
0.77978515625,
0.82421875,
0.79150390625,
0.74658203125,
0.79150390625
],
[
0.31787109375,
0.80615234375,
0.52783203125,
0.80615234375,
0.52783203125,
0.82373046875,
0.31787109375,
0.82373046875
],
[
0.60888671875,
0.81103515625,
0.63818359375,
0.81103515625,
0.63818359375,
0.818359375,
0.60888671875,
0.818359375
],
[
0.7685546875,
0.81005859375,
0.79150390625,
0.81005859375,
0.79150390625,
0.8173828125,
0.7685546875,
0.8173828125
],
[
0.2998046875,
0.82470703125,
0.5439453125,
0.82470703125,
0.5439453125,
0.84521484375,
0.2998046875,
0.84521484375
],
[
0.1829833984375,
0.8125,
0.32177734375,
0.814453125,
0.32177734375,
0.8359375,
0.1829833984375,
0.833984375
],
[
0.256591796875,
0.83544921875,
0.281005859375,
0.83544921875,
0.281005859375,
0.84326171875,
0.256591796875,
0.84326171875
],
[
0.279296875,
0.8359375,
0.298095703125,
0.8359375,
0.298095703125,
0.84033203125,
0.279296875,
0.84033203125
],
[
0.6259765625,
0.830078125,
0.81396484375,
0.83837890625,
0.81201171875,
0.85400390625,
0.625,
0.84521484375
],
[
0.54150390625,
0.8447265625,
0.56201171875,
0.8447265625,
0.56201171875,
0.85791015625,
0.54150390625,
0.85791015625
],
[
0.6611328125,
0.85498046875,
0.81787109375,
0.87060546875,
0.814453125,
0.88818359375,
0.65869140625,
0.87255859375
],
[
0.5908203125,
0.861328125,
0.5947265625,
0.861328125,
0.5947265625,
0.86865234375,
0.5908203125,
0.86865234375
],
[
0.17724609375,
0.869140625,
0.560546875,
0.869140625,
0.560546875,
0.88330078125,
0.17724609375,
0.88330078125
],
[
0.5908203125,
0.8720703125,
0.59619140625,
0.8720703125,
0.59619140625,
0.88330078125,
0.5908203125,
0.88330078125
],
[
0.176513671875,
0.88330078125,
0.560546875,
0.88330078125,
0.560546875,
0.896484375,
0.176513671875,
0.896484375
],
[
0.66259765625,
0.87744140625,
0.814453125,
0.88720703125,
0.81396484375,
0.9013671875,
0.6611328125,
0.892578125
],
[
0.17724609375,
0.89794921875,
0.56201171875,
0.89794921875,
0.56201171875,
0.9111328125,
0.17724609375,
0.9111328125
],
[
0.17724609375,
0.9111328125,
0.560546875,
0.9111328125,
0.560546875,
0.9248046875,
0.17724609375,
0.9248046875
],
[
0.646484375,
0.92041015625,
0.79345703125,
0.92041015625,
0.79345703125,
0.93115234375,
0.646484375,
0.93115234375
],
[
0.49267578125,
0.94873046875,
0.50634765625,
0.94873046875,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
],
[
0.576171875,
0.9443359375,
0.82421875,
0.9443359375,
0.82421875,
0.95849609375,
0.576171875,
0.95849609375
],
[
0.57666015625,
0.95849609375,
0.82275390625,
0.95849609375,
0.82275390625,
0.9716796875,
0.57666015625,
0.9716796875
],
[
0.57666015625,
0.97265625,
0.82177734375,
0.97265625,
0.82177734375,
0.986328125,
0.57666015625,
0.986328125
]
] |
[
"Published as a conference paper at ICLR 2023",
"A.Conventional",
"NXHWXC",
"NXC",
"Avg",
"Patches",
"B.Replicated",
"NXHWXGXC/GNXHWXGXC",
"NXHWXHWXGXC",
"NXHWXGXC",
"vg",
"D",
"",
"Q",
"",
"",
"",
"",
"Avg",
"Stack +StopGradient",
"AvgPool +stopGradient",
"Figure 3: Feature aggregator designs. A) In the conventional design, average pooling is performed",
"to aggregate features from different spatial locations. B) We propose the replicated design, features",
"are first concatenated across groups and then averaged across spatial locations. We create copies of",
"the same feature with different stop gradient masks so that we obtain more local losses instead of a",
"global one. The stop gradient mask makes sure that perturbation in one spatial group corresponds to",
"its loss function. The numerical value of the loss function is the same as the conventional design",
"2)Patchwiseloss.图像等感官输入信号具有空间维度。我们将沿着这些空间维度逐段应用",
"单独的损失。在VisionTransformer架构（Vaswanietal.2017;Dosovitskiyetal.,2021）中，每",
"个空间标记表示图像中的一个补丁。在现代深度网络中，通常会共享每个空间位置的参数",
"以提高数据效率并降低内存带宽利用率。虽然天真的重量共享在生物学上是不合理的，但",
"我们仍然认为在这项工作中共享重量。通过在补丁之间添加知识蒸馏E(Hintonetal.2015)",
"损失，可以模拟权重共享的效果。",
"3)Groupwiseloss.最后，我们转向通道维度。为了创建多个损失，我们将通道拆分为多个",
"组，每个组都附加到一个损失函数（Pateletal..2022）。为防止组之间相互通信，通道仅连接",
"到同一组中的其他通道。单个组g的分组线性层计算为g,3=Z,wg,,jg,，。以前的工作",
"使用通道组来提高计算效率（Krizhevskyetal.,2012;Ioannouetal.2017;Xieetal.,2017），而",
"在我们的工作中，添加组会增加损失总数，从而减少方差",
"Featureaggregators.朴素地将损失分别应用于空间和通道维度会导致次优的性能，因为每",
"个维度仅包含本地信息。对于标准任务（如分类）的损失，模型需要输人的全局视图来做出",
"决策。标准体系结构通过在最终分类层之前执行全局平均池化层来获得此全局视图。因此，",
"我们探索在局部损失函数之前聚合来自其他组和空间斑块的信息的策略。",
"我们希望在不减少维度总数的情况下执行聚合。因此，我们提出了一种用于特征聚合的复",
"制设计，如图3所示。首先，通道组被复制并相互通信，但是除了活动组本身之外的每个",
"组都被停止梯度屏蔽，以便其他组不影响前向梯度计算：",
"Cp.g =[StopGrad(αp,1..Cp.g-1), p.g, StopGrad(ap,g+1,.., ap,G)],",
"(4)",
"其中P和9分别索引补丁和组。同样，每个空间位置也被复制，通信和屏蔽，然后在本地平",
"均：",
"p.g=(ap,g+ppStopGrad(ap,g)",
"(5)",
"特征聚合的输出与常规全局平均池化层的输出相同。不同的是，这里的损失是重复的，并且",
"在每个损失中激活不同的补丁组",
"Learningobjectives.我们考虑监督分类损失和对比信息损失E(vandenOordetal，2018;",
"Chenetal.，2020），这是图像表示学习中最常用的两种损失。对于监督分类，我们在",
"聚合特征的顶部附加一个共享线性层（跨p,9共享），以实现交叉嫡损失：Lp.g=",
"-",
"Zktklogsoftmax(Wipg）k。每个组和修补程序位置的损失值相同",
"假设)",
"）和α(2)",
"是n示例的两个不同视图，则对比学习的InfoNCE",
"对于对比学习，线性层成为线性特征投影仪。",
"",
"损失为：",
"(W(,p,g) T StopGrad(W(2)",
"Origina",
"StopGr",
"m(Wp.g) StopGrad(W()",
"Lp,g = -log ",
"n",
"",
"",
"(6)",
"意事意意事",
"S0",
"请注意，我们在第二个视图上添加了停止梯度运算符。",
"Bul",
"通常没有必要在InfoNCE损失中添加此停止梯度；但",
"",
"是，我们发现基于扰动的方法需要停止梯度，否则损失",
"不会下降。这可能是因为我们在两个视图上共享扰动，",
"20000400006000080000",
"5",
"Figure 4:Importance of StopGradi-",
"ent in the InfoNCE loss, using M/8 on",
"CIFAR-10 with 256 channels 1 group"
] |
{
"arxiv_id": "2210.03310",
"language": "zh",
"page_idx": 4
}
|
||
1902.01785-zh-0
|
1902.01785-zh-0.jpg
|
Homogeneous Linear Inequality Constraints for Neural Network Activations
Thomas Frerix
Matthias Nießner
thomas.frerix@tum.de
${\mathrm{niessner@tum.de}}$
Technical University of Munich
Daniel Cremers
cremers@tum.de
Abstract
We propose a method to impose homogeneous linear inequality constraints of the form $Ax\leq0$ on neural network activations. The proposed method allows a data-driven training approach to be combined with modeling prior knowledge about the task. One way to achieve this task is by means of a projection step at test time after unconstrained training. However, this is an expensive operation. By directly incorporating the constraints into the architecture, we can significantly speed-up inference at test time; for instance, our experiments show a speed-up of up to two orders of magnitude over a projection method. Our algorithm computes a suitable parameterization of the feasible set at initialization and uses standard variants of stochastic gradient descent to fnd solutions to the constrained network. Thus, the modeling constraints are always satisfied during training. Crucially, our approach avoids to solve an optimization problem at each training step or to manually trade-off data and constraint fdelity with additional hyperparameters. We consider constrained generative modeling as an important application domain and experimentally demonstrate the proposed method by constraining a variational autoencoder.
## 1. Introduction
深度学习模型[14]在需要利用微妙相关性的任务中取得了显著成功，例如计算机视觉[11]和序列学习[20] 。通常，人类对一项任务有很强的先验知识，例如。g。, 基于对称、几何或物理。以纯数据驱动的方式学习这种先验假设是低效的，在某些情况下，可能根本不可行。尽管某些先验知识已成功应用-例如通过卷积体系结构[13]进行平移对称-在深度网络的训练中结合更通用的建模假设仍然是一个开放的挑战。最近，生成神经网络已经显著提高了[8,10]。使用这样的模型，控制生成过程超越数据驱动的黑盒方法尤为重要
在本文中，我们提出了一种通过$Ax\leq0$形式的齐
<FigureHere>
Figure 1. Samples drawn from a variational autoencoder
trained on MNIST without constraints (left) and with a checkerboard constraint on the output domain (right). For a pixel intensity domain [-1,1] , the checkerboard constraint forces the image tiles to have average positive or negative brightness.
次线性不等式约束对深度学习模型的激活施加先验知识的方法。我们通过对可行集进行适当的参数化直接施加这些约束。这种方法的两个主要优点是：
$\bullet$ The constraints are hard-constraints in the sense
that they are satisfied at any point during training
and inference.
$\bullet$ Inference on the constrained network incurs no
overhead compared to unconstrained inference
总而言之，我们方法的主要贡献是重新参数化，它将齐次线性不等式硬约束纳入神经网络激活，并允许有效的测试时间预测，即。e。,我们的方法快到两个数量级。该模型可以通过随机梯度下降的标准变体进行优化。作为生成建模中的应用，我们证明了我们的方法能够从变分自编码器生成真实样本，同时满足施加的约束
## 2. Related work
各种作品都介绍了对神经网络激活施加某种类型的
硬约束的方法
$\text{M\' {a}rquez- Neilaetal. [ 15] 将 通 用 可 微 分 等 式 约 束 公 }$ 式化为软约束，并采用拉格朗日方法来训练其模型。虽然这是一种约束优化的原则方法，但它不能很好地扩展到具有大量参数的实际深度神经网络模型。为了使他们的方法在计算上易于处理，在每个训练步骤中选
|
[
[
0.1070556640625,
0.1351318359375,
0.86669921875,
0.1351318359375,
0.86669921875,
0.149658203125,
0.1070556640625,
0.149658203125
],
[
0.223876953125,
0.18115234375,
0.351318359375,
0.18310546875,
0.351318359375,
0.2008056640625,
0.223876953125,
0.1988525390625
],
[
0.419189453125,
0.184326171875,
0.56787109375,
0.184326171875,
0.56787109375,
0.1988525390625,
0.419189453125,
0.1988525390625
],
[
0.6259765625,
0.184326171875,
0.75634765625,
0.184326171875,
0.75634765625,
0.1988525390625,
0.6259765625,
0.1988525390625
],
[
0.214111328125,
0.2020263671875,
0.361083984375,
0.2020263671875,
0.361083984375,
0.216552734375,
0.214111328125,
0.216552734375
],
[
0.43701171875,
0.20263671875,
0.5498046875,
0.20263671875,
0.5498046875,
0.216552734375,
0.43701171875,
0.216552734375
],
[
0.6337890625,
0.2020263671875,
0.74658203125,
0.2008056640625,
0.74658203125,
0.21533203125,
0.6337890625,
0.216552734375
],
[
0.3544921875,
0.2266845703125,
0.61669921875,
0.2266845703125,
0.61669921875,
0.240478515625,
0.3544921875,
0.240478515625
],
[
0.238525390625,
0.27783203125,
0.313720703125,
0.27783203125,
0.313720703125,
0.292236328125,
0.238525390625,
0.292236328125
],
[
0.560546875,
0.285400390625,
0.82666015625,
0.27587890625,
0.83251953125,
0.36865234375,
0.56640625,
0.37744140625
],
[
0.101318359375,
0.311279296875,
0.46728515625,
0.311279296875,
0.46728515625,
0.324462890625,
0.101318359375,
0.324462890625
],
[
0.08087158203125,
0.326904296875,
0.465576171875,
0.326904296875,
0.465576171875,
0.34033203125,
0.08087158203125,
0.34033203125
],
[
0.08087158203125,
0.341552734375,
0.468017578125,
0.341552734375,
0.468017578125,
0.354736328125,
0.08087158203125,
0.354736328125
],
[
0.080078125,
0.356689453125,
0.468017578125,
0.35595703125,
0.468017578125,
0.37060546875,
0.080078125,
0.371826171875
],
[
0.08087158203125,
0.37060546875,
0.46728515625,
0.37060546875,
0.46728515625,
0.385009765625,
0.08087158203125,
0.385009765625
],
[
0.07843017578125,
0.385009765625,
0.468017578125,
0.386962890625,
0.468017578125,
0.401611328125,
0.07843017578125,
0.399658203125
],
[
0.5048828125,
0.386962890625,
0.88818359375,
0.386962890625,
0.88818359375,
0.397705078125,
0.5048828125,
0.397705078125
],
[
0.080078125,
0.401611328125,
0.468017578125,
0.401611328125,
0.468017578125,
0.415283203125,
0.080078125,
0.415283203125
],
[
0.5048828125,
0.40087890625,
0.8896484375,
0.40087890625,
0.8896484375,
0.410888671875,
0.5048828125,
0.410888671875
],
[
0.08087158203125,
0.417236328125,
0.46728515625,
0.417236328125,
0.46728515625,
0.43115234375,
0.08087158203125,
0.43115234375
],
[
0.5048828125,
0.4140625,
0.8896484375,
0.4140625,
0.8896484375,
0.4248046875,
0.5048828125,
0.4248046875
],
[
0.50244140625,
0.4267578125,
0.88818359375,
0.4267578125,
0.88818359375,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.08251953125,
0.434326171875,
0.46728515625,
0.434326171875,
0.46728515625,
0.4443359375,
0.08251953125,
0.4443359375
],
[
0.50341796875,
0.43994140625,
0.8896484375,
0.44140625,
0.8896484375,
0.455078125,
0.50341796875,
0.453857421875
],
[
0.08087158203125,
0.447509765625,
0.46728515625,
0.447509765625,
0.46728515625,
0.4609375,
0.08087158203125,
0.4609375
],
[
0.5048828125,
0.45703125,
0.63232421875,
0.45703125,
0.63232421875,
0.46728515625,
0.5048828125,
0.46728515625
],
[
0.08251953125,
0.464111328125,
0.46728515625,
0.464111328125,
0.46728515625,
0.474853515625,
0.08251953125,
0.474853515625
],
[
0.08087158203125,
0.477783203125,
0.468017578125,
0.477783203125,
0.468017578125,
0.4912109375,
0.08087158203125,
0.4912109375
],
[
0.08087158203125,
0.492431640625,
0.468017578125,
0.492431640625,
0.468017578125,
0.505859375,
0.08087158203125,
0.505859375
],
[
0.5048828125,
0.496826171875,
0.8896484375,
0.496826171875,
0.8896484375,
0.5068359375,
0.5048828125,
0.5068359375
],
[
0.08087158203125,
0.50830078125,
0.468017578125,
0.50830078125,
0.468017578125,
0.521484375,
0.08087158203125,
0.521484375
],
[
0.5048828125,
0.51025390625,
0.890625,
0.51025390625,
0.890625,
0.5234375,
0.5048828125,
0.5234375
],
[
0.08251953125,
0.5244140625,
0.46728515625,
0.5244140625,
0.46728515625,
0.53515625,
0.08251953125,
0.53515625
],
[
0.50244140625,
0.52587890625,
0.81884765625,
0.52587890625,
0.81884765625,
0.5390625,
0.50244140625,
0.5390625
],
[
0.08087158203125,
0.537109375,
0.46728515625,
0.537109375,
0.46728515625,
0.55029296875,
0.08087158203125,
0.55029296875
],
[
0.080078125,
0.55029296875,
0.465576171875,
0.55322265625,
0.465576171875,
0.5673828125,
0.080078125,
0.56494140625
],
[
0.51708984375,
0.5517578125,
0.890625,
0.5517578125,
0.890625,
0.56494140625,
0.51708984375,
0.56494140625
],
[
0.08087158203125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.08087158203125,
0.58203125
],
[
0.53515625,
0.56494140625,
0.89208984375,
0.56640625,
0.89208984375,
0.58056640625,
0.53515625,
0.580078125
],
[
0.08251953125,
0.58544921875,
0.465576171875,
0.58544921875,
0.465576171875,
0.59521484375,
0.08251953125,
0.59521484375
],
[
0.5361328125,
0.58349609375,
0.6337890625,
0.58349609375,
0.6337890625,
0.5947265625,
0.5361328125,
0.5947265625
],
[
0.08087158203125,
0.5986328125,
0.468017578125,
0.5986328125,
0.468017578125,
0.6123046875,
0.08087158203125,
0.6123046875
],
[
0.080078125,
0.6123046875,
0.468017578125,
0.61279296875,
0.468017578125,
0.62744140625,
0.080078125,
0.626953125
],
[
0.51708984375,
0.609375,
0.890625,
0.609375,
0.890625,
0.62255859375,
0.51708984375,
0.62255859375
],
[
0.08087158203125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.642578125,
0.08087158203125,
0.642578125
],
[
0.5361328125,
0.62548828125,
0.86767578125,
0.62548828125,
0.86767578125,
0.63623046875,
0.5361328125,
0.63623046875
],
[
0.080078125,
0.642578125,
0.46728515625,
0.64306640625,
0.46728515625,
0.65771484375,
0.080078125,
0.6572265625
],
[
0.5048828125,
0.6484375,
0.89208984375,
0.6494140625,
0.89208984375,
0.6640625,
0.5048828125,
0.66357421875
],
[
0.080078125,
0.6591796875,
0.221435546875,
0.6591796875,
0.221435546875,
0.67236328125,
0.080078125,
0.67236328125
],
[
0.5048828125,
0.66552734375,
0.8896484375,
0.66552734375,
0.8896484375,
0.6787109375,
0.5048828125,
0.6787109375
],
[
0.5048828125,
0.6796875,
0.890625,
0.6796875,
0.890625,
0.693359375,
0.5048828125,
0.693359375
],
[
0.5048828125,
0.6962890625,
0.890625,
0.6962890625,
0.890625,
0.70703125,
0.5048828125,
0.70703125
],
[
0.080078125,
0.70849609375,
0.21728515625,
0.70849609375,
0.21728515625,
0.72265625,
0.080078125,
0.72265625
],
[
0.5048828125,
0.71044921875,
0.890625,
0.71044921875,
0.890625,
0.7236328125,
0.5048828125,
0.7236328125
],
[
0.50341796875,
0.72607421875,
0.89208984375,
0.72607421875,
0.89208984375,
0.7392578125,
0.50341796875,
0.7392578125
],
[
0.101318359375,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.101318359375,
0.74755859375
],
[
0.50341796875,
0.7412109375,
0.52294921875,
0.7412109375,
0.52294921875,
0.7548828125,
0.50341796875,
0.7548828125
],
[
0.08087158203125,
0.75,
0.46728515625,
0.75,
0.46728515625,
0.76318359375,
0.08087158203125,
0.76318359375
],
[
0.08416748046875,
0.76318359375,
0.46728515625,
0.76513671875,
0.46728515625,
0.77978515625,
0.08416748046875,
0.77783203125
],
[
0.50244140625,
0.76953125,
0.6435546875,
0.76953125,
0.6435546875,
0.78271484375,
0.50244140625,
0.78271484375
],
[
0.08251953125,
0.77978515625,
0.468017578125,
0.77978515625,
0.468017578125,
0.79345703125,
0.08251953125,
0.79345703125
],
[
0.080078125,
0.79345703125,
0.468017578125,
0.79443359375,
0.468017578125,
0.8095703125,
0.080078125,
0.80810546875
],
[
0.52294921875,
0.79443359375,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.52294921875,
0.8095703125
],
[
0.08251953125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.08251953125,
0.82373046875
],
[
0.50244140625,
0.81005859375,
0.60302734375,
0.8125,
0.60302734375,
0.8271484375,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.08087158203125,
0.8388671875
],
[
0.52294921875,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84033203125,
0.52294921875,
0.84033203125
],
[
0.08251953125,
0.84228515625,
0.46728515625,
0.84228515625,
0.46728515625,
0.85302734375,
0.08251953125,
0.85302734375
],
[
0.50341796875,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.08251953125,
0.85498046875,
0.468017578125,
0.85498046875,
0.468017578125,
0.86865234375,
0.08251953125,
0.86865234375
],
[
0.50634765625,
0.859375,
0.88623046875,
0.859375,
0.88623046875,
0.869140625,
0.50634765625,
0.869140625
],
[
0.08087158203125,
0.87060546875,
0.359375,
0.87060546875,
0.359375,
0.8837890625,
0.08087158203125,
0.8837890625
],
[
0.5048828125,
0.87255859375,
0.890625,
0.87255859375,
0.890625,
0.88623046875,
0.5048828125,
0.88623046875
],
[
0.101318359375,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.101318359375,
0.90087890625
],
[
0.5048828125,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.5048828125,
0.90087890625
]
] |
[
"Homogeneous Linear Inequality Constraints for Neural Network Activations",
"Thomas Frerix",
"Matthias NieBner",
"Daniel Cremers",
"thomas.frerix@tum.de",
"niessner@tum.de",
"cremers@tum.de",
"Technical University of Munich",
"Abstract",
"oFD",
"We propose a method to impose homogeneous lin-",
"ear inequality constraints of the form A ≤ 0 on neu",
"ral network activations. The proposed method allows",
"a data-driven training approach to be combined with",
"modeling prior knowledge about the task. One way",
"to achieve this task is by means of a projection step",
"Figure l.Samples drawn from a variational autoencoder",
"at test time after unconstrained training. However,",
"trained on MNIST without constraints (left) and with a",
"this is an expensive operation. By directly incorpo-",
"checkerboard constraint on the output domain (right).For",
"a pixel intensity domain [-1,1] , the checkerboard con-",
"rating the constraints into the architecture,we can",
"straint forces the image tiles to have average positive or",
"significantly speed-up inference at test time; for in-",
"negativebrightness",
"stance,our experiments show a speed-up of up to two",
"orders of magnitude over a projection method. Our",
"algorithm computes a suitable parameterization of the",
"次线性不等式约束对深度学习模型的激活施加先验知",
"feasible set at initialization and uses standard variants",
"识的方法。我们通过对可行集进行适当的参数化直接",
"of stochastic gradient descent to find solutions to the",
"施加这些约束。这种方法的两个主要优点是",
"constrained network. Thus, the modeling constraints",
"are always satisfied during training. Crucially, our ap",
". The constraints are hard-constraints in the sense",
"proach avoids to solve an optimization problem at each",
"that they are satisfied at any point during training",
"training step or to manuallytrade-off data and con-",
"and inference",
"straint fidelity with additional hyperparameters. We",
"consider constrained generative modeling as an impor-",
". Inference on the constrained network incurs no",
"tant application domain and experimentally demon-",
"overhead comparedto unconstrainedinference",
"strate the proposed method by constraining a varia-",
"总而言之，我们方法的主要贡献是重新参数化，它将",
"tional autoencoder.",
"齐次线性不等式硬约束纳入神经网络激活，并允许有",
"效的测试时间预测，即。e。，我们的方法快到两个数量",
"级。该模型可以通过随机梯度下降的标准变体进行优",
"1. Introduction",
"化。作为生成建模中的应用，我们证明了我们的方法能",
"够从变分自编码器生成真实样本，同时满足施加的约",
"深度学习模型[14]在需要利用微妙相关性的任务中",
"束",
"取得了显著成功，例如计算机视觉[11]和序列学习[20]",
"。通常，人类对一项任务有很强的先验知识，例如。g。",
"2. Related work",
"基于对称、几何或物理。以纯数据驱动的方式学习这种",
"先验假设是低效的，在某些情况下，可能根本不可行。",
"各种作品都介绍了对神经网络激活施加某种类型的",
"尽管某些先验知识已成功应用-例如通过卷积体系结构",
"硬约束的方法",
"[13]进行平移对称-在深度网络的训练中结合更通用的",
"Marquez-Neilaetal.[15]将通用可微分等式约束公",
"建模假设仍然是一个开放的挑战。最近，生成神经网络",
"式化为软约束，并采用拉格朗日方法来训练其模型。虽",
"已经显著提高了[8，10]。使用这样的模型，控制生成",
"然这是一种纳束优化的原则方法，但它不能很好地扩",
"过程超越数据驱动的黑盒方法尤为重要",
"展到具有大量参数的实际深度神经网络模型。为了使",
"在本文中，我们提出了一种通过A≤0形式的齐",
"他们的方法在计算上易于处理，在每个训练步骤中选"
] |
{
"arxiv_id": "1902.01785",
"language": "zh",
"page_idx": 0
}
|
||
1902.01785-zh-1
|
1902.01785-zh-1.jpg
|
择约束的子集。此外，这些约束是局部线性化的；因此，
不能保证在参数更新后满足该子集
是，此方法涉及在每个训练步骤中解决(凸)优化问题。此外，整体收敛路径取决于交替优化步骤的组合方式， 这引人了必须调整的额外超参数。Briq et al.[4] 通过在单纯形上实现正交投影的层来解决弱监督分割问题， 从而将激活直接限制为概率分布。此优化问题可以有效地解决，但不能推广到其他类型的不等式约束
OptNet 是 Amos and Kolter [1]提出的一种解决通用二次程序作为可微网络层的方法。OptNet 通过二次程序的一阶最优性条件进行反向传播，并且可以将线性不等式约束作为特例强制执行。该公式是灵活的；但是，它随变量和约束的数量而变化。因此，训练大规模深度学习模型变得非常昂贵
最后，有几篇作品提出了针对特定应用的手工解决方案，例如骨架预测[21]和刚体运动预测[5]。相反， 为了避免费力的架构设计，我们主张对约束类进行一般建模的价值。在实践中，这使得约束方法更容易为更
广泛的问题类别
Contribution 在这项工作中，我们解决了对神经网络激活施加齐次线性不等式约束的问题。我们不是在训练期间解决优化问题，而是将此任务分为初始化时的可行性步骤和训练期间的最优性步骤。在初始化时， 我们计算约束集(多面体锥)的适当参数化，并使用神经网络训练算法在此可行集中找到一个好的解决方案。从概念上讲，我们在初始化期间权衡计算成本，以获得在测试时没有开销的模型。所提出的方法被实现为一个神经网络层，该神经网络层由一组齐次线性不等式指定，其输出对可行集进行参数化
3. Linear constraints for deep learning models
我们考虑一个通用的$L$层神经网络$F_\theta$,其模型参数$\theta$用于输入$x$ ,如下所示：。g。, 全连接或卷积层，$\sigma\text{ 是元素非线性}^{1}$, e。g。, S 形或整流线性单元(ReLU)。在监督学习中，训练目标$y$是已知的，并且损失$\mathcal{L}_y(F_\theta(x))$作为网络参数$\theta$的函数被最小化。分类任务的典型损失是网络输出和经验目标分布之间的交叉熵，而均方误差通常用于回归任务。所提出的方法可用于约束任何线性激活$z^{(l)}=f_{\theta_l}^{(l)}(a^{(l-1)})$或非线性激活$a^{(l)}=\sigma(z^{(l)})$ 。在大多数情况下，希望约束输出$F_{\theta}(x)$
$d$维中$m$线性不等式约束的可行集是凸多面体
$$\mathcal C:=\left\{z\bigg|Az\le b,A\in\mathbb R^{m\times d},b\in\mathbb R^m\right\}\subseteq\mathbb R^d\:.$$
(1,
$^{1}$Formally, $\sigma$ maps between different spaces for different layers and may also be a different element-wise non-linearity for each layer. We omit such details in favor of notational simplicity.
通过多面体的分解定理获得了凸多面体$\mathcal{C}$的合适描述。
对 于 弱 监 督 分 割 的 特 定 问 题 , Pathak et al. [ 18] 提 Theorem 1 (Decomposition of polyhedra, 出 了 一 种 优 化 方 案 , 该 方 案 在 优 化 深 度 学 习 模 型 和 将 Minkowski- Weyl) . A set $\mathcal{C} \subset \mathbb{R} ^d$ is a convex 受 约 束 的 分 布 拟 合 到 这 些 中 间 模 型 之 间 进 行 交 替 。但 polyhedron of the form ( 1) if and only if
$$\begin{aligned}\mathcal{C}&=\mathrm{conv}(v_{1},\ldots,v_{n})+\mathrm{cone}(r_{1},\ldots,r_{s})\\&=\left\{\sum_{i=1}^{n}\lambda_{i}v_{i}+\sum_{j=1}^{s}\mu_{j}r_{j}\bigg|\lambda_{i},\mu_{j}\geq0,\sum_{i=1}^{n}\lambda_{i}=1\right\}\end{aligned}$$
(2
$\begin{array}{ll}{\mathrm{for}}&{{\mathrm{fnitely}}}&{{\mathrm{many}}}&{{\mathrm{vertices}}}&{{\{v_{1},\ldots,v_{n}\}}}&{{\mathrm{and}}}\\{{\mathrm{rays}}}&{{\{r_{1},\ldots,r_{s}\}}}&{{\mathrm{Furthermore},}}&{{\mathcal{C}}}&{{=}}\\{{\{z|Az\leq0,A\in\mathbb{R}^{m\times d}\}}}&{{\mathrm{if~and~only~if}}}\end{array}$
$$\mathcal{C}=\mathrm{cone}(r_1,\ldots,r_s)$$
for fnitely many rays $\{ r_1, \ldots , r_s\}$ .
(3)
该定理指出，半空间(半空间或 H 表示)的交集可以写成多面体顶点的凸组合和某些射线的圆锥形组合(顶点或 V 表示)的 Minkowski 和。可以通过双重描述方法[7,16]在这两个观点之间进行算法切换，我们将在下面讨论。因此，在对不等式约束进行建模时很自然的 H 表示可以转换为 V 表示 , 可以将其合并到基于梯度的神经网络训练中
在本文中，我们专注于(3)形式的齐次约束，其可行集是多面体锥。由于此集的特殊结构，我们可以避免使用 (2) 中的凸组合参数，这在数值上是有利的 (第 3.5节),并且我们可以有效地组合建模约束和域约束，例如图像的[-1,1]像素域 (第 3.3 节)。这种多面体锥体如图 2 所示
## 3.1. Double description method
双重描述方法在线性不等式系统的半空间和顶点表示之间进行转换。它最初由 Motzkin et al.[16] 提出， 并由 Fukuda and Prodon $[7]\text{ 进一步完善。}^{2}$在这里， 我们只对齐次约束(3)从 H 表示到 V 表示的转换感兴趣，
$$\to\:\mathrm{cone}(r_1,\ldots,r_s)\:.$$
(4)
核心算法如下进行。让$A$的行定义一组齐次不等式， 并让$R=[r_1,\ldots,r_s]$为矩阵，其列是相应圆锥的射线。在这里，$(A,R)$形成双描述对。该算法以以下方式从$(A^{k},R^{k})$迭代构建双描述对$(A^k+1,R^{k+1})$ 。$A^{k}$中的行表示$A$行的$k$子集，因此定义了与$R^k$关联的凸多面体。将单行添加到$A^k$会引入一个额外的半空间约束， 该约束对应于超平面。如果$R^k$的两列$r_i,r_j$的向量$r_{i}-r_{j}$ 与此超平面相交，则将此交点添加到$R^k$。从$R^k$中删除被附加超平面截断的现有光线。结果是双描述对$(A^k+1,R^{k+1})$ 。此程序如图 2 所示
$^2$In our experiments we use pycddlib , which is a Python wrap-
per of Fukuda's cddlib .
|
[
[
0.08087158203125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.50341796875,
0.0921630859375,
0.88623046875,
0.0921630859375,
0.88623046875,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.08087158203125,
0.106689453125,
0.325927734375,
0.10797119140625,
0.325927734375,
0.12249755859375,
0.08087158203125,
0.121826171875
],
[
0.5048828125,
0.11932373046875,
0.57177734375,
0.11932373046875,
0.57177734375,
0.1307373046875,
0.5048828125,
0.1307373046875
],
[
0.57177734375,
0.11932373046875,
0.63232421875,
0.11932373046875,
0.63232421875,
0.1300048828125,
0.57177734375,
0.1300048828125
],
[
0.6259765625,
0.11871337890625,
0.81396484375,
0.11871337890625,
0.81396484375,
0.1319580078125,
0.6259765625,
0.1319580078125
],
[
0.101318359375,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.101318359375,
0.1351318359375
],
[
0.80615234375,
0.12060546875,
0.88818359375,
0.12060546875,
0.88818359375,
0.1319580078125,
0.80615234375,
0.1319580078125
],
[
0.08087158203125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50341796875,
0.1331787109375,
0.89208984375,
0.1331787109375,
0.89208984375,
0.147705078125,
0.50341796875,
0.147705078125
],
[
0.08087158203125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.08087158203125,
0.1666259765625
],
[
0.5009765625,
0.149658203125,
0.7958984375,
0.1483154296875,
0.7958984375,
0.1622314453125,
0.5009765625,
0.1634521484375
],
[
0.08251953125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.08251953125,
0.1817626953125
],
[
0.51123046875,
0.175537109375,
0.7890625,
0.1767578125,
0.7890625,
0.1912841796875,
0.51123046875,
0.1900634765625
],
[
0.08087158203125,
0.18310546875,
0.465576171875,
0.18310546875,
0.465576171875,
0.1962890625,
0.08087158203125,
0.1962890625
],
[
0.08087158203125,
0.1988525390625,
0.46728515625,
0.1988525390625,
0.46728515625,
0.212158203125,
0.08087158203125,
0.212158203125
],
[
0.56201171875,
0.20703125,
0.861328125,
0.2083740234375,
0.861328125,
0.2298583984375,
0.56201171875,
0.2291259765625
],
[
0.08087158203125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.85693359375,
0.2139892578125,
0.87109375,
0.2139892578125,
0.87109375,
0.222900390625,
0.85693359375,
0.222900390625
],
[
0.86767578125,
0.21337890625,
0.88623046875,
0.21337890625,
0.88623046875,
0.2259521484375,
0.86767578125,
0.2259521484375
],
[
0.63232421875,
0.2279052734375,
0.6650390625,
0.2279052734375,
0.6650390625,
0.2392578125,
0.63232421875,
0.2392578125
],
[
0.771484375,
0.2266845703125,
0.82666015625,
0.2266845703125,
0.82666015625,
0.2412109375,
0.771484375,
0.2412109375
],
[
0.5537109375,
0.2298583984375,
0.583984375,
0.2298583984375,
0.583984375,
0.238037109375,
0.5537109375,
0.238037109375
],
[
0.08251953125,
0.2291259765625,
0.46728515625,
0.2291259765625,
0.46728515625,
0.242431640625,
0.08251953125,
0.242431640625
],
[
0.08087158203125,
0.24365234375,
0.42236328125,
0.24365234375,
0.42236328125,
0.2568359375,
0.08087158203125,
0.2568359375
],
[
0.50341796875,
0.2568359375,
0.5458984375,
0.2568359375,
0.5458984375,
0.268310546875,
0.50341796875,
0.268310546875
],
[
0.5361328125,
0.25390625,
0.8935546875,
0.25244140625,
0.8935546875,
0.270751953125,
0.5361328125,
0.271484375
],
[
0.7138671875,
0.26953125,
0.8486328125,
0.270751953125,
0.8486328125,
0.285400390625,
0.7138671875,
0.2841796875
],
[
0.1004638671875,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.1004638671875,
0.272705078125
],
[
0.08416748046875,
0.27587890625,
0.465576171875,
0.27587890625,
0.465576171875,
0.285888671875,
0.08416748046875,
0.285888671875
],
[
0.50244140625,
0.271484375,
0.64208984375,
0.271484375,
0.64208984375,
0.285400390625,
0.50244140625,
0.285400390625
],
[
0.87841796875,
0.27587890625,
0.88623046875,
0.27587890625,
0.88623046875,
0.27978515625,
0.87841796875,
0.27978515625
],
[
0.08087158203125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.50244140625,
0.2841796875,
0.76953125,
0.28271484375,
0.76953125,
0.300537109375,
0.50244140625,
0.3017578125
],
[
0.08087158203125,
0.3037109375,
0.46728515625,
0.3037109375,
0.46728515625,
0.317626953125,
0.08087158203125,
0.317626953125
],
[
0.62353515625,
0.311279296875,
0.7685546875,
0.313232421875,
0.7685546875,
0.330078125,
0.62353515625,
0.32763671875
],
[
0.08251953125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.08251953125,
0.333251953125
],
[
0.86767578125,
0.314453125,
0.890625,
0.314453125,
0.890625,
0.328857421875,
0.86767578125,
0.328857421875
],
[
0.080078125,
0.333984375,
0.293212890625,
0.335205078125,
0.293212890625,
0.348388671875,
0.080078125,
0.347900390625
],
[
0.50244140625,
0.34033203125,
0.75341796875,
0.34228515625,
0.75341796875,
0.356689453125,
0.50244140625,
0.354736328125
],
[
0.1004638671875,
0.349853515625,
0.468017578125,
0.349853515625,
0.468017578125,
0.363037109375,
0.1004638671875,
0.363037109375
],
[
0.08251953125,
0.365478515625,
0.468017578125,
0.365478515625,
0.468017578125,
0.37890625,
0.08251953125,
0.37890625
],
[
0.52392578125,
0.367431640625,
0.890625,
0.367431640625,
0.890625,
0.380615234375,
0.52392578125,
0.380615234375
],
[
0.08251953125,
0.380126953125,
0.465576171875,
0.380126953125,
0.465576171875,
0.393310546875,
0.08251953125,
0.393310546875
],
[
0.50341796875,
0.38330078125,
0.8896484375,
0.38330078125,
0.8896484375,
0.396484375,
0.50341796875,
0.396484375
],
[
0.08251953125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.4091796875,
0.08251953125,
0.4091796875
],
[
0.50341796875,
0.399658203125,
0.8896484375,
0.399658203125,
0.8896484375,
0.410400390625,
0.50341796875,
0.410400390625
],
[
0.080078125,
0.410400390625,
0.1968994140625,
0.410400390625,
0.1968994140625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.5048828125,
0.41357421875,
0.890625,
0.41357421875,
0.890625,
0.4267578125,
0.5048828125,
0.4267578125
],
[
0.5048828125,
0.4287109375,
0.890625,
0.4287109375,
0.890625,
0.442626953125,
0.5048828125,
0.442626953125
],
[
0.08660888671875,
0.443115234375,
0.468017578125,
0.443115234375,
0.468017578125,
0.45703125,
0.08660888671875,
0.45703125
],
[
0.50341796875,
0.4443359375,
0.890625,
0.4443359375,
0.890625,
0.458251953125,
0.50341796875,
0.458251953125
],
[
0.080078125,
0.45703125,
0.468017578125,
0.458251953125,
0.468017578125,
0.4716796875,
0.080078125,
0.47021484375
],
[
0.50341796875,
0.458984375,
0.65185546875,
0.458984375,
0.65185546875,
0.472900390625,
0.50341796875,
0.472900390625
],
[
0.08087158203125,
0.472900390625,
0.46728515625,
0.473388671875,
0.46728515625,
0.4873046875,
0.08087158203125,
0.486083984375
],
[
0.52294921875,
0.474853515625,
0.890625,
0.474853515625,
0.890625,
0.488037109375,
0.52294921875,
0.488037109375
],
[
0.08087158203125,
0.4892578125,
0.4697265625,
0.4892578125,
0.4697265625,
0.50244140625,
0.08087158203125,
0.50244140625
],
[
0.50341796875,
0.4892578125,
0.89208984375,
0.4892578125,
0.89208984375,
0.50244140625,
0.50341796875,
0.50244140625
],
[
0.08251953125,
0.50390625,
0.468017578125,
0.50390625,
0.468017578125,
0.517578125,
0.08251953125,
0.517578125
],
[
0.50341796875,
0.50390625,
0.89208984375,
0.50390625,
0.89208984375,
0.517578125,
0.50341796875,
0.517578125
],
[
0.08251953125,
0.52099609375,
0.46728515625,
0.52099609375,
0.46728515625,
0.53076171875,
0.08251953125,
0.53076171875
],
[
0.50341796875,
0.51953125,
0.8896484375,
0.51953125,
0.8896484375,
0.53271484375,
0.50341796875,
0.53271484375
],
[
0.08251953125,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.08251953125,
0.54736328125
],
[
0.50341796875,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.50341796875,
0.548828125
],
[
0.08087158203125,
0.548828125,
0.451904296875,
0.548828125,
0.451904296875,
0.5625,
0.08087158203125,
0.5625
],
[
0.50341796875,
0.5498046875,
0.5947265625,
0.5517578125,
0.5947265625,
0.56298828125,
0.50341796875,
0.56103515625
],
[
0.080078125,
0.5625,
0.468017578125,
0.56298828125,
0.468017578125,
0.57763671875,
0.080078125,
0.5771484375
],
[
0.08087158203125,
0.580078125,
0.326904296875,
0.580078125,
0.326904296875,
0.59326171875,
0.08087158203125,
0.59326171875
],
[
0.50244140625,
0.57470703125,
0.7490234375,
0.57373046875,
0.7490234375,
0.58837890625,
0.50244140625,
0.5888671875
],
[
0.52392578125,
0.5986328125,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.52392578125,
0.6123046875
],
[
0.080078125,
0.609375,
0.468017578125,
0.609375,
0.468017578125,
0.62353515625,
0.080078125,
0.62353515625
],
[
0.5048828125,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.5048828125,
0.626953125
],
[
0.5048828125,
0.62744140625,
0.890625,
0.62744140625,
0.890625,
0.642578125,
0.5048828125,
0.642578125
],
[
0.101318359375,
0.6337890625,
0.468017578125,
0.6337890625,
0.468017578125,
0.64794921875,
0.101318359375,
0.64794921875
],
[
0.5048828125,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.6572265625,
0.5048828125,
0.6572265625
],
[
0.08087158203125,
0.6494140625,
0.468017578125,
0.6494140625,
0.468017578125,
0.6640625,
0.08087158203125,
0.6640625
],
[
0.07843017578125,
0.662109375,
0.4697265625,
0.6611328125,
0.4697265625,
0.6787109375,
0.07843017578125,
0.6796875
],
[
0.50341796875,
0.6591796875,
0.52685546875,
0.6591796875,
0.52685546875,
0.671875,
0.50341796875,
0.671875
],
[
0.08251953125,
0.6767578125,
0.465576171875,
0.6767578125,
0.465576171875,
0.69140625,
0.08251953125,
0.69140625
],
[
0.6259765625,
0.68603515625,
0.76953125,
0.6875,
0.76953125,
0.7021484375,
0.6259765625,
0.70068359375
],
[
0.08087158203125,
0.693359375,
0.46728515625,
0.693359375,
0.46728515625,
0.70703125,
0.08087158203125,
0.70703125
],
[
0.869140625,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.7021484375,
0.869140625,
0.7021484375
],
[
0.08087158203125,
0.708984375,
0.468017578125,
0.708984375,
0.468017578125,
0.72265625,
0.08087158203125,
0.72265625
],
[
0.5048828125,
0.71533203125,
0.890625,
0.71533203125,
0.890625,
0.72900390625,
0.5048828125,
0.72900390625
],
[
0.08251953125,
0.724609375,
0.468017578125,
0.724609375,
0.468017578125,
0.73779296875,
0.08251953125,
0.73779296875
],
[
0.50341796875,
0.72900390625,
0.890625,
0.72998046875,
0.890625,
0.744140625,
0.50341796875,
0.74365234375
],
[
0.08087158203125,
0.7392578125,
0.4697265625,
0.7392578125,
0.4697265625,
0.7568359375,
0.08087158203125,
0.7568359375
],
[
0.50341796875,
0.74560546875,
0.8896484375,
0.74560546875,
0.8896484375,
0.7587890625,
0.50341796875,
0.7587890625
],
[
0.08087158203125,
0.75830078125,
0.468017578125,
0.75830078125,
0.468017578125,
0.77294921875,
0.08087158203125,
0.77294921875
],
[
0.5048828125,
0.7587890625,
0.890625,
0.7587890625,
0.890625,
0.7734375,
0.5048828125,
0.7734375
],
[
0.08251953125,
0.7744140625,
0.1217041015625,
0.7744140625,
0.1217041015625,
0.7861328125,
0.08251953125,
0.7861328125
],
[
0.5048828125,
0.7744140625,
0.890625,
0.7744140625,
0.890625,
0.7890625,
0.5048828125,
0.7890625
],
[
0.09722900390625,
0.787109375,
0.431396484375,
0.787109375,
0.431396484375,
0.80029296875,
0.09722900390625,
0.80029296875
],
[
0.5048828125,
0.79052734375,
0.88818359375,
0.79052734375,
0.88818359375,
0.8037109375,
0.5048828125,
0.8037109375
],
[
0.50244140625,
0.8046875,
0.8896484375,
0.8046875,
0.8896484375,
0.8193359375,
0.50244140625,
0.8193359375
],
[
0.36767578125,
0.82373046875,
0.449462890625,
0.82568359375,
0.449462890625,
0.84033203125,
0.36767578125,
0.83837890625
],
[
0.5048828125,
0.82080078125,
0.890625,
0.82080078125,
0.890625,
0.833984375,
0.5048828125,
0.833984375
],
[
0.108642578125,
0.82763671875,
0.1380615234375,
0.82763671875,
0.1380615234375,
0.8388671875,
0.108642578125,
0.8388671875
],
[
0.156005859375,
0.82470703125,
0.35693359375,
0.82275390625,
0.35693359375,
0.8388671875,
0.156005859375,
0.84130859375
],
[
0.44287109375,
0.8291015625,
0.46728515625,
0.8291015625,
0.46728515625,
0.84033203125,
0.44287109375,
0.84033203125
],
[
0.50341796875,
0.833984375,
0.890625,
0.8359375,
0.890625,
0.849609375,
0.50341796875,
0.84765625
],
[
0.50244140625,
0.849609375,
0.79736328125,
0.85107421875,
0.79736328125,
0.86572265625,
0.50244140625,
0.8642578125
],
[
0.09234619140625,
0.861328125,
0.464111328125,
0.86279296875,
0.464111328125,
0.87744140625,
0.09234619140625,
0.87548828125
],
[
0.08087158203125,
0.87548828125,
0.46728515625,
0.87548828125,
0.46728515625,
0.8896484375,
0.08087158203125,
0.8896484375
],
[
0.5185546875,
0.87353515625,
0.8896484375,
0.87548828125,
0.8896484375,
0.89013671875,
0.5185546875,
0.88818359375
],
[
0.080078125,
0.88720703125,
0.4453125,
0.88818359375,
0.4453125,
0.9013671875,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88818359375,
0.64892578125,
0.88818359375,
0.64892578125,
0.9013671875,
0.50341796875,
0.9013671875
]
] |
[
"择约束的子集。此外，这些约束是局部线性化的；因此，",
"通过多面体的分解定理获得了凸多面体C的合适描述，",
"不能保证在参数更新后满足该子集",
"Theorem",
"1",
"(Decompositionof",
"对于弱监督分割的特定问题，Pathaketal.18l提",
"polyhedra",
"出了一种优化方案，该方案在优化深度学习模型和将",
"Minkowski-Weyl). A set C C Rd is a convex",
"受约束的分布拟合到这些中间模型之间进行交替。但",
"polyhedron of the form (1) if and only if",
"是，此方法涉及在每个训练步骤中解决（凸）优化问题。",
"C = conv(ui,..., Un) + cone(ri,...,rs)",
"此外，整体收敛路径取决于交替优化步骤的组合方式",
"这引人了必须调整的额外超参数。Briqet al.[4]通过",
"0=",
"在单纯形上实现正交投影的层来解决弱监督分割问题，",
"",
"(2",
"=l",
"=1",
".3=1",
"从而将激活直接限制为概率分布。此优化问题可以有",
"效地解决，但不能推广到其他类型的不等式约束",
"for",
"finitely many vertices [ui,...,Un) and",
"Furthermore,C",
"OptNet是AmosandKolter[1]提出的一种解决通",
"用二次程序作为可微网络层的方法。OptNet通过二次",
"rays(ri,...,rs]",
"-",
"程序的一阶最优性条件进行反向传播，并且可以将线",
"[z]Az ≤ o, AE Rmxd] if and only if",
"性不等式约束作为特例强制执行。该公式是灵活的；但",
"C = cone(r1,...,rs)",
"是，它随变量和约束的数量而变化。。因此，训练大规",
"(3)",
"模深度学习模型变得非常昂贵",
"for finitely many rays (ri,...,rs]",
"最后，有几篇作品提出了针对特定应用的手工解决",
"方案，例如骨架预测[21]和刚体运动预测[5]。相反,",
"该定理指出，半空间（半空间或H表示）的交集可",
"为了避免费力的架构设计，我们主张对约束类进行一",
"以写成多面体顶点的凸组合和某些射线的圆锥形组合",
"般建模的价值。在实践中，这使得约束方法更容易为更",
"（顶点或V表示）的Minkowski和。可以通过双重描述",
"广泛的问题类别",
"方法[7，16]在这两个观点之间进行算法切换，我们将",
"在下面讨论。因此，在对不等式约束进行建模时很自然",
"Contribution在这项工作中，我们解决了对神经网",
"的H表示可以转换为V表示，可以将其合并到基于梯",
"络激活施加齐次线性不等式约束的问题。我们不是在",
"度的神经网络训练中",
"训练期间解决优化问题，而是将此任务分为初始化时",
"在本文中，我们专注于（3）形式的齐次约束，其可行",
"的可行性步骤和训练期间的最优性步骤。在初始化时，",
"集是多面体锥。由于此集的特殊结构，我们可以避免使",
"我们计算约束集（多面体锥）的适当参数化，并使用神",
"用（2）中的凸组合参数，这在数值上是有利的（第3.5",
"经网络训练算法在此可行集中找到一个好的解决方案。",
"节），并且我们可以有效地组合建模约束和域约束，例",
"从概念上讲，我们在初始化期间权衡计算成本，以获得",
"如图像的【-1,1]像素域（第3.3节）。这种多面体锥体",
"在测试时没有开销的模型。所提出的方法被实现为",
"如图2所示",
"个神经网络层，该神经网络层由一组齐次线性不等式",
"指定，其输出对可行集进行参数化",
"3.1. Double description method",
"双重描述方法在线性不等式系统的半空间和顶点表",
"3. Linear constraints for deep learning models",
"示之间进行转换。它最初由Motzkinetal.[16]提出，",
"并由 Fukuda and Prodon [7] 进一步完善。2在这里,",
"我们考虑一个通用的L层神经网络Fe，其模型参",
"我们只对齐次约束（3）从H表示到V表示的转换感兴",
"数用于输入，如下所示：。g。，全连接或卷积",
"层，α是元素非线性1，e。g。，S形或整流线性单元",
"趣，",
"(ReLU)。在监督学习中，训练目标y是已知的，并且",
"→ cone(ri,...,rs) .",
"损失Cu(Fe(a)）作为网络参数的函数被最小化。分",
"(4)",
"类任务的典型损失是网络输出和经验目标分布之间的",
"核心算法如下进行。让A的行定义一组齐次不等式，",
"交叉摘，而均方误差通常用于回归任务。所提出的方法",
"并让R=[r1...,rs]为矩阵，其列是相应圆锥的射线。",
"可用于约束任何线性激活()=fe(a(l-1)）或非线性",
"在这里，（A,R）形成双描述对。该算法以以下方式从",
"激活α①）=α(z()）。在大多数情况下，希望约束输出",
"（A,Rk）迭代构建双描述对（Ak+1,Rk+1）。Ak中的行",
"Fe(a)",
"表示A行的k子集，因此定义了与Rh关联的凸多面",
"d维中m线性不等式约束的可行集是凸多面体",
"体。将单行添加到A会引人一个额外的半空间约束",
"该约束对应于超平面。如果Rk的两列ri，ri的向量",
"CRd.",
"ri-ri与此超平面相交，则将此交点添加到Rk。从",
"C:=",
"2Az <b, Ae Rmxd,bE Rm",
"(1)",
"Rk中删除被附加超平面截断的现有光线。结果是双描",
"述对（Ak+1,Rk+1）。此程序如图2所示",
"IFormally, g maps between different spaces for different layers",
"and may also be a different element-wise non-linearity for each",
"2In our experiments we use pycddlib,which is a Python wrap",
"layer. We omit such details in favor of notational simplicity.",
"per of Fukuda's cddlib."
] |
{
"arxiv_id": "1902.01785",
"language": "zh",
"page_idx": 1
}
|
||
1902.01785-zh-2
|
1902.01785-zh-2.jpg
|
<FigureHere>
Figure 2. Diagram illustrating an iteration of the dou-
ble description method. Adding a constraint to the $k$- constraint set $A^k$ at iteration $k+1$ introduces a hyperplane $H.$ The intersection points of $H$ with the boundary of the current polyhedron $R^k$ (marked by o ) are added as rays $r_6$ and $r_7$ to the polyhedral cone. The ray $r_2$ is cut-off by the hyperplane $H$ and is removed from $R^k.$ The result is the next iterate $R^{k+ 1}$ .
添加超平面可能会大大增加中间表示中的光线数量， 这反过来又在随后的迭代中以组合方式做出贡献。实际上，存在最坏情况的多面体，该算法的指数运行时间是不等式数和输入维数以及射线数[3,6]的函数。总的来说，可以期望该算法仅对具有相当少量的不等式$m$ 和维数$d$的问题有效。
3.2. Integration in neural network architectures
我们通过神经网络层对齐次形式(3)进行参数化。该层将数据的某些(潜在)表示形式作为输人，该表示形式映射到满足所需硬约束的激活。该算法具有线性不等式约束的 H 表示，即。e。, $d$维中$m$约束的矩阵$A\in \mathbb{R} ^{m\times d}$ ,用于指定可行集(3)。在初始化时，我们通过双重描述方法将其转换为 V 表示 (第 3.1 节)。这对应于计算一组射线$\{r_1,\ldots,r_s\}$来表示多面体锥。在训练过程中，使用神经网络训练算法在可行集内进行优化。此过程中有两个关键方面。首先，如$\overline{\mathrm{E}}3.1$ 节所述，双重描述方法的运行时复杂性可能过高。从概念上讲，所提出的方法允许在初始化时花费大量的计算时
间，以获得在训练和测试时非常有效的算法。其次，我们必须确保从潜在表示到参数的映射与训练算法很好地集成。我们假设模型是用基于梯度的反向传播训练的，这在当前的深度学习应用中是很常见的。。约束层包括一个批处理归一化层和一个仿射映射(带有偏差的全连接层),然后是元素方式的绝对值函数，该函数可确保圆锥形组合参数所需的非负性。理论上，任何函数$f:\mathbb{R}\to\mathbb{R}_{\geq0}$都将满足此要求；但是，必须注意不要干扰反向传播的梯度
3.3. Combining modeling and domain constraints
域约束通常表示为单位框约束$\mathcal{B}:=\{x\in\mathbb{R}^d|-1\leq$ $x_i\leq 1\}$ ,例如图像的像素域。框约束特别不适合使用双重描述方法进行转换，因为顶点的数量在维度上是指数的。因此，我们区分建模约束和域约束，仅将前者转换为 V 表示。基于此表示，我们在建模约束集中获得一个点$x\in\mathcal{C}$ 。但是，这一点可能不在单元框$\mathcal{B}$中。为了到达交点$\mathcal{C}\cap\mathcal{B}$中的一个点，如果$x\notin \mathcal{B}$ , $\hat{x} = x/ \max \{ \| x\| _\infty , 1\}$ ,我们通过其无穷大范数对$x$进行归一化。实际上，$\hat{x}\in\mathcal{C}\cap\mathcal{B}$由于按正常数缩放仍保留在圆锥体中，即。e。,如果 $x\in\mathcal{C}$,则 $\alpha x\in\mathcal{C}\forall\alpha\geq0$
3.4. Applications of homogeneous constraints
$Ax\leq0$形式的约束的自然应用是一组二进制分类器的参数化。如果$A$的每一行$a_i$都是这样的二进制分类器，那么本文提出的方法将集合$\{x|a_i^Tx\leq0\forall i\}$参数化。因此，可以保证神经网络激活满足一组二进制标准。另一个领域是表达神经网络激活之间的某些直接关系。值得注意的是，可以通过$x_i+1\geq x_i$保证单调性和通过$x_{i+1}-2x_i+x_{i-1}\geq0$保证凸性等数学属性
3.5. Extension to general linear constraints
所提出的方法利用多面体锥体的特殊结构来有效地结合建模和域约束(第$\overline{\mathrm{E}}3.3$ 节)。$Ax\leq b$ 形式的一般线性不等式约束，而对$A$和$b$没有限制，可能需要(2)的圆锥和凸分量用于其 V 表示。在这种情况下，可以使用本文的主要方法，即。e。,我们层还需要预测凸组合参数。但是，我们观察到收敛缓慢，这归因于凸组合参数的单纯形参数化。我们使用 softmax 函数$\text{口纽口参数时中纪D多数化。仅们仅们}^{\text{ 301010014x 四双}}$ $f( x) _i= \exp ( x_i) / \sum _{j= 1}^m\exp ( x_j)$ 来 强 制 (2) 中 凸 组 合 参 数 的 幼 击 $> 0\sum _{m}^m- 1\text{ 业 众 }_{\text{ 时 日 十 五 世 }}$ 数的约束$\lambda_i\geq0,\sum_{i=1}^{m}\lambda_{i}=1$ 。当一个$x_i$明显大于其他向量条目时，此函数具有消失的梯度。此外，这种最一般的设置不允许有效地合并域约束，因为这将需要对一般凸多面体和单位框的交集进行有效的参数化
## 4. Numerical results
我们将所提出的约束参数化算法与在没有约束的情况下进行训练但在测试时需要投影步骤的算法进行了比较。我们称之为后一种算法测试时间投影。所提出的算法在可行集上进行优化，而投影仅限于在该集合的边界上产生解。我们在两种不同的设置中分析这些算法。在初始实验中，我们学习了约束集上的正交投影，
|
[
[
0.5048828125,
0.09344482421875,
0.89208984375,
0.09344482421875,
0.89208984375,
0.106689453125,
0.5048828125,
0.106689453125
],
[
0.50634765625,
0.10986328125,
0.88818359375,
0.10986328125,
0.88818359375,
0.12060546875,
0.50634765625,
0.12060546875
],
[
0.324462890625,
0.11871337890625,
0.330810546875,
0.12371826171875,
0.325927734375,
0.1287841796875,
0.318603515625,
0.12371826171875
],
[
0.5048828125,
0.125,
0.8896484375,
0.125,
0.8896484375,
0.1351318359375,
0.5048828125,
0.1351318359375
],
[
0.5048828125,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.5048828125,
0.154052734375,
0.890625,
0.154052734375,
0.890625,
0.1673583984375,
0.5048828125,
0.1673583984375
],
[
0.236083984375,
0.1641845703125,
0.256591796875,
0.1641845703125,
0.256591796875,
0.1749267578125,
0.236083984375,
0.1749267578125
],
[
0.5048828125,
0.1685791015625,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.5048828125,
0.1817626953125
],
[
0.50341796875,
0.18310546875,
0.8896484375,
0.184326171875,
0.8896484375,
0.1976318359375,
0.50341796875,
0.1962890625
],
[
0.50341796875,
0.1988525390625,
0.890625,
0.1988525390625,
0.890625,
0.212158203125,
0.50341796875,
0.212158203125
],
[
0.5048828125,
0.2139892578125,
0.66845703125,
0.2139892578125,
0.66845703125,
0.2279052734375,
0.5048828125,
0.2279052734375
],
[
0.50244140625,
0.238037109375,
0.88818359375,
0.2392578125,
0.88818359375,
0.25390625,
0.50244140625,
0.25244140625
],
[
0.52392578125,
0.261962890625,
0.890625,
0.261962890625,
0.890625,
0.27587890625,
0.52392578125,
0.27587890625
],
[
0.50244140625,
0.279052734375,
0.890625,
0.27783203125,
0.890625,
0.291748046875,
0.50244140625,
0.292236328125
],
[
0.2467041015625,
0.29736328125,
0.2607421875,
0.29736328125,
0.2607421875,
0.30810546875,
0.2467041015625,
0.30810546875
],
[
0.5048828125,
0.294189453125,
0.8896484375,
0.294189453125,
0.8896484375,
0.304931640625,
0.5048828125,
0.304931640625
],
[
0.32275390625,
0.306884765625,
0.33251953125,
0.306884765625,
0.33251953125,
0.311279296875,
0.32275390625,
0.311279296875
],
[
0.50634765625,
0.31005859375,
0.8896484375,
0.31005859375,
0.8896484375,
0.32080078125,
0.50634765625,
0.32080078125
],
[
0.5048828125,
0.323974609375,
0.890625,
0.323974609375,
0.890625,
0.337158203125,
0.5048828125,
0.337158203125
],
[
0.5048828125,
0.339111328125,
0.890625,
0.339111328125,
0.890625,
0.352783203125,
0.5048828125,
0.352783203125
],
[
0.50244140625,
0.352783203125,
0.88623046875,
0.3515625,
0.88623046875,
0.3662109375,
0.50244140625,
0.367431640625
],
[
0.5048828125,
0.37060546875,
0.8896484375,
0.37060546875,
0.8896484375,
0.380615234375,
0.5048828125,
0.380615234375
],
[
0.5048828125,
0.3837890625,
0.890625,
0.3837890625,
0.890625,
0.397705078125,
0.5048828125,
0.397705078125
],
[
0.5048828125,
0.399658203125,
0.890625,
0.399658203125,
0.890625,
0.41357421875,
0.5048828125,
0.41357421875
],
[
0.50244140625,
0.423583984375,
0.85546875,
0.425537109375,
0.85546875,
0.439453125,
0.50244140625,
0.4375
],
[
0.52392578125,
0.448974609375,
0.890625,
0.448974609375,
0.890625,
0.462158203125,
0.52392578125,
0.462158203125
],
[
0.50341796875,
0.462158203125,
0.8896484375,
0.46337890625,
0.8896484375,
0.477783203125,
0.50341796875,
0.4765625
],
[
0.08087158203125,
0.479736328125,
0.468017578125,
0.479736328125,
0.468017578125,
0.49365234375,
0.08087158203125,
0.49365234375
],
[
0.50341796875,
0.479248046875,
0.890625,
0.479248046875,
0.890625,
0.492431640625,
0.50341796875,
0.492431640625
],
[
0.080078125,
0.492431640625,
0.468017578125,
0.49365234375,
0.468017578125,
0.5068359375,
0.080078125,
0.505859375
],
[
0.5048828125,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5068359375,
0.5048828125,
0.5068359375
],
[
0.09234619140625,
0.5087890625,
0.465576171875,
0.5087890625,
0.465576171875,
0.51953125,
0.09234619140625,
0.51953125
],
[
0.50341796875,
0.5087890625,
0.890625,
0.5087890625,
0.890625,
0.52294921875,
0.50341796875,
0.52294921875
],
[
0.08251953125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.53515625,
0.08251953125,
0.53515625
],
[
0.5048828125,
0.5234375,
0.890625,
0.5234375,
0.890625,
0.53857421875,
0.5048828125,
0.53857421875
],
[
0.08087158203125,
0.5341796875,
0.46728515625,
0.53515625,
0.46728515625,
0.5498046875,
0.08087158203125,
0.548828125
],
[
0.5048828125,
0.5390625,
0.86669921875,
0.5390625,
0.86669921875,
0.55322265625,
0.5048828125,
0.55322265625
],
[
0.08251953125,
0.55029296875,
0.465576171875,
0.55029296875,
0.465576171875,
0.56103515625,
0.08251953125,
0.56103515625
],
[
0.08087158203125,
0.56298828125,
0.468017578125,
0.56298828125,
0.468017578125,
0.5771484375,
0.08087158203125,
0.5771484375
],
[
0.50244140625,
0.564453125,
0.8349609375,
0.564453125,
0.8349609375,
0.57763671875,
0.50244140625,
0.57763671875
],
[
0.08087158203125,
0.57568359375,
0.2091064453125,
0.57568359375,
0.2091064453125,
0.5888671875,
0.08087158203125,
0.5888671875
],
[
0.52392578125,
0.58837890625,
0.890625,
0.58837890625,
0.890625,
0.6015625,
0.52392578125,
0.6015625
],
[
0.50341796875,
0.60302734375,
0.8759765625,
0.60302734375,
0.8759765625,
0.6162109375,
0.50341796875,
0.6162109375
],
[
0.1004638671875,
0.61279296875,
0.468017578125,
0.61279296875,
0.468017578125,
0.626953125,
0.1004638671875,
0.626953125
],
[
0.50341796875,
0.61865234375,
0.8896484375,
0.61865234375,
0.8896484375,
0.6318359375,
0.50341796875,
0.6318359375
],
[
0.08251953125,
0.6298828125,
0.46728515625,
0.6298828125,
0.46728515625,
0.64013671875,
0.08251953125,
0.64013671875
],
[
0.50634765625,
0.63525390625,
0.88623046875,
0.63525390625,
0.88623046875,
0.64599609375,
0.50634765625,
0.64599609375
],
[
0.080078125,
0.642578125,
0.46728515625,
0.64306640625,
0.46728515625,
0.6572265625,
0.080078125,
0.65576171875
],
[
0.50244140625,
0.64794921875,
0.88818359375,
0.6484375,
0.88818359375,
0.662109375,
0.50244140625,
0.6611328125
],
[
0.08251953125,
0.6591796875,
0.46728515625,
0.6591796875,
0.46728515625,
0.67236328125,
0.08251953125,
0.67236328125
],
[
0.50341796875,
0.662109375,
0.88818359375,
0.66357421875,
0.88818359375,
0.67822265625,
0.50341796875,
0.6767578125
],
[
0.08251953125,
0.67529296875,
0.465576171875,
0.67529296875,
0.465576171875,
0.68603515625,
0.08251953125,
0.68603515625
],
[
0.5048828125,
0.6787109375,
0.890625,
0.6787109375,
0.890625,
0.6923828125,
0.5048828125,
0.6923828125
],
[
0.08087158203125,
0.689453125,
0.240234375,
0.689453125,
0.240234375,
0.70263671875,
0.08087158203125,
0.70263671875
],
[
0.5048828125,
0.6943359375,
0.890625,
0.6943359375,
0.890625,
0.708984375,
0.5048828125,
0.708984375
],
[
0.08087158203125,
0.71337890625,
0.451904296875,
0.71337890625,
0.451904296875,
0.72802734375,
0.08087158203125,
0.72802734375
],
[
0.50341796875,
0.70849609375,
0.89208984375,
0.70849609375,
0.89208984375,
0.72607421875,
0.50341796875,
0.72607421875
],
[
0.5048828125,
0.72607421875,
0.890625,
0.72607421875,
0.890625,
0.7392578125,
0.5048828125,
0.7392578125
],
[
0.101318359375,
0.73779296875,
0.465576171875,
0.73779296875,
0.465576171875,
0.74853515625,
0.101318359375,
0.74853515625
],
[
0.50634765625,
0.7412109375,
0.890625,
0.7412109375,
0.890625,
0.7548828125,
0.50634765625,
0.7548828125
],
[
0.08251953125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.08251953125,
0.76513671875
],
[
0.50341796875,
0.755859375,
0.86376953125,
0.755859375,
0.86376953125,
0.76953125,
0.50341796875,
0.76953125
],
[
0.08416748046875,
0.76806640625,
0.465576171875,
0.76806640625,
0.465576171875,
0.77880859375,
0.08416748046875,
0.77880859375
],
[
0.08087158203125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.08087158203125,
0.79541015625
],
[
0.080078125,
0.79248046875,
0.468017578125,
0.79541015625,
0.468017578125,
0.81298828125,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.78857421875,
0.6787109375,
0.78857421875,
0.6787109375,
0.798828125,
0.50341796875,
0.798828125
],
[
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
],
[
0.08087158203125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84130859375,
0.08087158203125,
0.84130859375
],
[
0.50341796875,
0.8271484375,
0.88818359375,
0.8271484375,
0.88818359375,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.08416748046875,
0.84326171875,
0.46728515625,
0.84326171875,
0.46728515625,
0.85400390625,
0.08416748046875,
0.85400390625
],
[
0.5048828125,
0.84326171875,
0.8896484375,
0.84326171875,
0.8896484375,
0.85400390625,
0.5048828125,
0.85400390625
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.08087158203125,
0.87060546875
],
[
0.5048828125,
0.859375,
0.88818359375,
0.859375,
0.88818359375,
0.869140625,
0.5048828125,
0.869140625
],
[
0.08251953125,
0.87353515625,
0.46728515625,
0.87353515625,
0.46728515625,
0.8837890625,
0.08251953125,
0.8837890625
],
[
0.5048828125,
0.87353515625,
0.8896484375,
0.87353515625,
0.8896484375,
0.8837890625,
0.5048828125,
0.8837890625
],
[
0.08087158203125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.5048828125,
0.8896484375,
0.890625,
0.8896484375,
0.890625,
0.8994140625,
0.5048828125,
0.8994140625
]
] |
[
"间，以获得在训练和测试时非常有效的算法。其次，我",
"们必须确保从潜在表示到参数的映射与训练算法很好",
"?",
"地集成。我们假设模型是用基于梯度的反向传播训练",
"的，这在当前的深度学习应用中是很常见的。。约束层",
"包括一个批处理归一化层和一个仿射映射（带有偏差",
"Rk",
"的全连接层），然后是元素方式的绝对值函数，该函数",
"可确保圆锥形组合参数所需的非负性。理论上，任何函",
"数f：R→R>0都将满足此要求；但是，必须注意不",
"要干扰反向传播的梯度",
"3.3. Combining modeling and domain constraints",
"域约束通常表示为单位框约束B=αERd-1≤",
"Ti≤1，例如图像的像素域。框约束特别不适合使",
"r6",
"用双重描述方法进行转换，因为顶点的数量在维度上",
"",
"是指数的。因此，我们区分建模纳束和域纳束，仅将",
"前者转换为V表示。基于此表示，我们在建模约束集",
"中获得一个点EC。但是，这一点可能不在单元框",
"B中。为了到达交点CnB中的一个点，如果生B",
"=c/maxa：1，我们通过其无穷大范数对进",
"行归一化。实际上，ECnB由于按正常数缩放仍保",
"留在圆锥体中，即。e。，如果aEC，则αaEVα≥O",
"3.4.Applications of homogeneous constraints",
"Ar≤0形式的约束的自然应用是一组二进制分类器",
"的参数化。如果A的每一行ai都是这样的二进制分",
"Figure 2. Diagram illustrating an iteration of the dou-",
"类器，那么本文提出的方法将集合（αa≤0Vi参",
"ble description method. Adding a constraint to the k-",
"数化。因此，可以保证神经网络激活满足一组二进制标",
"onstraintsetAat iterationk+1introducesahyperplane",
"准。另一个领域是表达神经网络激活之间的某些直接",
"H . The intersection points of H with the boundary of the",
"关系。值得注意的是，可以通过1≥r保证单调性",
"current polyhedron Rk (marked by o ) are added as rays T6",
"和通过a+1-2ai+ai-1≥0保证凸性等数学属性",
"and r7to the polyhedral cone.The ray r2 is cut-off bythe",
"hyperplane H and is removed from Rk. The result is the",
"3.5. Extension to general linear constraints",
"next iterate Rk+1.",
"所提出的方法利用多面体锥体的特殊结构来有效地",
"结合建模和域约束（第E3.3节）。Ac≤b形式的",
"添加超平面可能会大大增加中间表示中的光线数量，",
"般线性不等式约束，而对A和6没有限制，可能需要",
"这反过来又在随后的送代中以组合方式做出责献。实",
"（2）的圆锥和凸分量用于其V表示。在这种情况下，可",
"际上，存在最坏情况的多面体，该算法的指数运行时间",
"以使用本文的主要方法，即。e。，我们层还需要预测",
"是不等式数和输人维数以及射线数[3，6]的函数。总的",
"凸组合参数。但是，我们观察到收敛缓慢，这归因于",
"来说，可以期望该算法仅对具有相当少量的不等式M",
"凸组合参数的单纯形参数化。我们使用softmax函数",
"和维数d的问题有效。",
"f(α);=exp(c)/Z=1exp(cs）来强制(2)中凸组合参",
"3.2. Integration in neural network architectures",
"数的约束入≥0,㎡1入=1。当一个明显大于其",
"他向量条目时，此函数具有消失的梯度。此外，这种最",
"我们通过神经网络层对齐次形式（3）进行参数化。",
"一般的设置不允许有效地合并域约束，因为这将需要",
"该层将数据的某些（潜在）表示形式作为输人，该表示",
"对一般凸多面体和单位框的交集进行有效的参数化",
"形式映射到满足所需硬约束的激活。该算法具有线性",
"不等式约束的H表示，即。e。，d维中m约束的矩阵",
"AeRmxd，用于指定可行集（3）。在初始化时，我们",
"4.Numericalresults",
"通过双重描述方法将其转换为V表示（第3.1节）。这",
"我们将所提出的约束参数化算法与在没有约束的情",
"对应于计算一组射线{r1，...，rs}来表示多面体锥。在",
"况下进行训练但在测试时需要投影步骤的算法进行了",
"训练过程中，使用神经网络训练算法在可行集内进行",
"比较。我们称之为后一种算法测试时间投影。所提出的",
"优化。此过程中有两个关键方面。首先，如E3.1节所",
"算法在可行集上进行优化，而投影仅限手在该集合的",
"述：双重描述方法的运行时复杂性可能过高。从概念上",
"边界上产生解。我们在两种不同的设置中分析这些算",
"讲，所提出的方法允许在初始化时花费大量的计算时",
"法。在初始实验中，我们学习了纳束集上的正交投影，"
] |
{
"arxiv_id": "1902.01785",
"language": "zh",
"page_idx": 2
}
|
||
1403.1349-zh-0
|
1403.1349-zh-0.jpg
|
学习软线性约束及其在引文字段提取中的应用
Sam Anzaroot Alexandre Passos David Belanger Andrew McCallum
Department of Computer Science University of Massachusetts, Amherst
$\{$anzaroot, apassos, belanger, mccallum $\}@$cs.umass.edu
$\operatorname{Abstract}$
Accurately segmenting a citation string into felds for authors, titles, etc. is a challenging task because the output typically obeys various global constraints. Previous work has shown that modeling soft constraints , where the model is encouraged, but not require to obey the constraints, can substantially improve segmentation performance. On the other hand, for imposing hard constraints , dual decomposition is a popular technique for efficient prediction given existing algorithms for unconstrained inference. We extend dual decomposition to perform prediction subject to soft constraints. Moreover, with a technique for performing inference given soft constraints, it is easy to automatically generate large families of constraints and learn their costs with a simple convex optimization problem during training. This allows us to obtain substantial gains in accuracy on a new, challenging citation extraction dataset.
## 1 Introduction
引文字段提取，是信息提取的一个实例，是将研究论文引文字符串分割并标记为其组成部分的任务，包括作者，编辑，年份，期刊，卷，会议地点等。这项任务很重要，因为引文数据通常仅以纯文本形式提供；但是，对于许多科学计量任务，例如绘制科学子社区，发现研究趋势以及分析研究人员网络，必须拥有准确的书目信息结构化数据库。自动引文字段提取需要进一步研究，因为它尚未达到可以在实际系统中实际部署的精度水平
隐马尔可夫模型和线性链条件随机场(CRFs)以前已应用于引文提取E(Hetzner,
2008; Peng and McCallum, 2004)。这些模型支持高效的动态编程推理，但仅对输出标签序列中的 local 依赖项进行建模。然而，引用具有这些模型未捕获的强 global 正则性。例如，许多书籍引用同时包含 author 部分和editor 部分，但没有一个具有两个不相交的author 部分。由于线性链模型无法捕获超过马尔可夫依赖关系，因此这些模型有时会将editor 错误地标记为第二作者。如果我们可以强制执行只有一个 author 部分的全局约束， 则可以提高准确性
将此类全局约束添加到易处理模型中的一个框架是 constrained inference ,在该框架中， 在推理时，原始模型会对输出进行限制，以使其遵守某些全局规则。当硬约束可以编码为输出变量上的线性方程，并且基础模型的推理任务可以作为线性优化时，可以将此约束推理问题表述为 integer linear program (ILP) (Roth and Yih, 2004)。或者，可以采用 dual ${\mathrm{decomposition~(Rush~et~al.,2010)~。双重分解}}$ 相对于 ILP 的优势在于，它可以将原始模型的现有推理算法作为黑匣子。这样的模块化算法易于实现，并且在实践中效果很好，为大多数示例提供了最优性证书
上述两种方法以前已应用于对模型的输出施加 hard 约束。另一方面，最近的工作证明了通过施加 soft 约束 (Chang et al.,2012)在引文字段提取方面的改进。在这里，该模型不需要遵守全局约束，而只是为违反它们而支付罚款
本文介绍了一种通过对偶分解施加软约束的新方法。我们还提出了一种方法，用于学习预测问题因违反这些软约束而受到的惩罚。由于我们的学习方法使许多惩罚变为零，因此它允许从业者执行“约束选择”,其中可以考虑大量自动生成的候选全局约束，并自动将其剔除为较小的有用约束集，可以在测试时快速运行使用我们的新方法，我们不仅能够合并Chang 等人的所有软全局约束。(2012),但也有更复杂的数据驱动约束，同时也提供了比他们的光束搜索技术更强的最优性证书。在
|
[
[
0.26123046875,
0.08489990234375,
0.73876953125,
0.08612060546875,
0.73876953125,
0.1009521484375,
0.26123046875,
0.09979248046875
],
[
0.189697265625,
0.1300048828125,
0.8134765625,
0.1300048828125,
0.8134765625,
0.144287109375,
0.189697265625,
0.144287109375
],
[
0.35693359375,
0.1478271484375,
0.6455078125,
0.1478271484375,
0.6455078125,
0.160888671875,
0.35693359375,
0.160888671875
],
[
0.3359375,
0.162109375,
0.6650390625,
0.1644287109375,
0.6650390625,
0.1787109375,
0.3359375,
0.177001953125
],
[
0.255126953125,
0.18115234375,
0.75,
0.18115234375,
0.75,
0.1953125,
0.255126953125,
0.1953125
],
[
0.2626953125,
0.2423095703125,
0.344970703125,
0.2423095703125,
0.344970703125,
0.260009765625,
0.2626953125,
0.260009765625
],
[
0.5146484375,
0.24462890625,
0.8818359375,
0.24462890625,
0.8818359375,
0.2587890625,
0.5146484375,
0.2587890625
],
[
0.515625,
0.26123046875,
0.8818359375,
0.26123046875,
0.8818359375,
0.27490234375,
0.515625,
0.27490234375
],
[
0.14697265625,
0.27490234375,
0.459228515625,
0.275634765625,
0.459228515625,
0.290283203125,
0.14697265625,
0.289306640625
],
[
0.515625,
0.27783203125,
0.87744140625,
0.27783203125,
0.87744140625,
0.2880859375,
0.515625,
0.2880859375
],
[
0.14697265625,
0.29150390625,
0.459228515625,
0.29150390625,
0.459228515625,
0.3046875,
0.14697265625,
0.3046875
],
[
0.515625,
0.29345703125,
0.8818359375,
0.29345703125,
0.8818359375,
0.30712890625,
0.515625,
0.30712890625
],
[
0.14697265625,
0.30810546875,
0.45751953125,
0.30810546875,
0.45751953125,
0.32177734375,
0.14697265625,
0.32177734375
],
[
0.5146484375,
0.308837890625,
0.8818359375,
0.30810546875,
0.8818359375,
0.322509765625,
0.5146484375,
0.32373046875
],
[
0.145263671875,
0.32373046875,
0.456787109375,
0.322509765625,
0.456787109375,
0.33740234375,
0.145263671875,
0.33837890625
],
[
0.5146484375,
0.325927734375,
0.8818359375,
0.325927734375,
0.8818359375,
0.339111328125,
0.5146484375,
0.339111328125
],
[
0.1485595703125,
0.341552734375,
0.240966796875,
0.341552734375,
0.240966796875,
0.352783203125,
0.1485595703125,
0.352783203125
],
[
0.231689453125,
0.33837890625,
0.45751953125,
0.339111328125,
0.45751953125,
0.35400390625,
0.231689453125,
0.352783203125
],
[
0.5146484375,
0.342529296875,
0.8818359375,
0.342529296875,
0.8818359375,
0.355712890625,
0.5146484375,
0.355712890625
],
[
0.14697265625,
0.355712890625,
0.459228515625,
0.355712890625,
0.459228515625,
0.369384765625,
0.14697265625,
0.369384765625
],
[
0.51708984375,
0.358154296875,
0.8818359375,
0.358154296875,
0.8818359375,
0.371826171875,
0.51708984375,
0.371826171875
],
[
0.14697265625,
0.371826171875,
0.45751953125,
0.371826171875,
0.45751953125,
0.385498046875,
0.14697265625,
0.385498046875
],
[
0.515625,
0.374755859375,
0.87890625,
0.374755859375,
0.87890625,
0.385498046875,
0.515625,
0.385498046875
],
[
0.1485595703125,
0.388427734375,
0.45751953125,
0.388427734375,
0.45751953125,
0.402099609375,
0.1485595703125,
0.402099609375
],
[
0.51318359375,
0.3896484375,
0.8818359375,
0.3896484375,
0.8818359375,
0.402587890625,
0.51318359375,
0.402587890625
],
[
0.1485595703125,
0.405029296875,
0.45751953125,
0.405029296875,
0.45751953125,
0.418701171875,
0.1485595703125,
0.418701171875
],
[
0.515625,
0.40625,
0.6650390625,
0.40625,
0.6650390625,
0.419189453125,
0.515625,
0.419189453125
],
[
0.1485595703125,
0.421630859375,
0.252685546875,
0.421630859375,
0.252685546875,
0.432861328125,
0.1485595703125,
0.432861328125
],
[
0.246826171875,
0.42041015625,
0.459228515625,
0.42041015625,
0.459228515625,
0.43408203125,
0.246826171875,
0.43408203125
],
[
0.5341796875,
0.42333984375,
0.8818359375,
0.42333984375,
0.8818359375,
0.43701171875,
0.5341796875,
0.43701171875
],
[
0.14697265625,
0.435791015625,
0.45751953125,
0.435791015625,
0.45751953125,
0.449462890625,
0.14697265625,
0.449462890625
],
[
0.5146484375,
0.438232421875,
0.8818359375,
0.439453125,
0.8818359375,
0.45361328125,
0.5146484375,
0.453125
],
[
0.1485595703125,
0.453125,
0.456787109375,
0.453125,
0.456787109375,
0.466064453125,
0.1485595703125,
0.466064453125
],
[
0.515625,
0.4560546875,
0.88330078125,
0.4560546875,
0.88330078125,
0.4697265625,
0.515625,
0.4697265625
],
[
0.1485595703125,
0.468505859375,
0.45751953125,
0.468505859375,
0.45751953125,
0.482177734375,
0.1485595703125,
0.482177734375
],
[
0.515625,
0.47265625,
0.8818359375,
0.47265625,
0.8818359375,
0.486328125,
0.515625,
0.486328125
],
[
0.1485595703125,
0.485107421875,
0.45751953125,
0.485107421875,
0.45751953125,
0.498779296875,
0.1485595703125,
0.498779296875
],
[
0.515625,
0.488037109375,
0.8818359375,
0.488037109375,
0.8818359375,
0.501953125,
0.515625,
0.501953125
],
[
0.1485595703125,
0.49951171875,
0.453369140625,
0.50048828125,
0.453369140625,
0.51416015625,
0.1485595703125,
0.51318359375
],
[
0.515625,
0.50439453125,
0.8818359375,
0.50439453125,
0.8818359375,
0.51708984375,
0.515625,
0.51708984375
],
[
0.1485595703125,
0.5185546875,
0.455078125,
0.5185546875,
0.455078125,
0.529296875,
0.1485595703125,
0.529296875
],
[
0.5146484375,
0.52099609375,
0.884765625,
0.52099609375,
0.884765625,
0.53369140625,
0.5146484375,
0.53369140625
],
[
0.1485595703125,
0.533203125,
0.45751953125,
0.533203125,
0.45751953125,
0.54638671875,
0.1485595703125,
0.54638671875
],
[
0.51708984375,
0.53759765625,
0.8818359375,
0.53759765625,
0.8818359375,
0.55029296875,
0.51708984375,
0.55029296875
],
[
0.1485595703125,
0.55029296875,
0.453369140625,
0.55029296875,
0.453369140625,
0.56103515625,
0.1485595703125,
0.56103515625
],
[
0.5146484375,
0.552734375,
0.8818359375,
0.552734375,
0.8818359375,
0.56640625,
0.5146484375,
0.56640625
],
[
0.145263671875,
0.5654296875,
0.456787109375,
0.56640625,
0.456787109375,
0.580078125,
0.145263671875,
0.5791015625
],
[
0.515625,
0.5693359375,
0.87890625,
0.5693359375,
0.87890625,
0.580078125,
0.515625,
0.580078125
],
[
0.1485595703125,
0.58203125,
0.459228515625,
0.58203125,
0.459228515625,
0.595703125,
0.1485595703125,
0.595703125
],
[
0.515625,
0.5849609375,
0.8818359375,
0.5849609375,
0.8818359375,
0.5986328125,
0.515625,
0.5986328125
],
[
0.14697265625,
0.59716796875,
0.459228515625,
0.59716796875,
0.459228515625,
0.61083984375,
0.14697265625,
0.61083984375
],
[
0.5146484375,
0.60107421875,
0.8818359375,
0.599609375,
0.8818359375,
0.6142578125,
0.5146484375,
0.615234375
],
[
0.1485595703125,
0.61328125,
0.45751953125,
0.61328125,
0.45751953125,
0.62744140625,
0.1485595703125,
0.62744140625
],
[
0.5146484375,
0.6162109375,
0.71728515625,
0.6162109375,
0.71728515625,
0.6298828125,
0.5146484375,
0.6298828125
],
[
0.1485595703125,
0.6298828125,
0.459228515625,
0.6298828125,
0.459228515625,
0.6435546875,
0.1485595703125,
0.6435546875
],
[
0.53564453125,
0.63427734375,
0.8818359375,
0.63427734375,
0.8818359375,
0.64794921875,
0.53564453125,
0.64794921875
],
[
0.1485595703125,
0.6455078125,
0.45751953125,
0.6455078125,
0.45751953125,
0.66015625,
0.1485595703125,
0.66015625
],
[
0.515625,
0.65087890625,
0.880859375,
0.65087890625,
0.880859375,
0.6650390625,
0.515625,
0.6650390625
],
[
0.14697265625,
0.662109375,
0.397216796875,
0.662109375,
0.397216796875,
0.67578125,
0.14697265625,
0.67578125
],
[
0.515625,
0.66748046875,
0.8818359375,
0.66748046875,
0.8818359375,
0.68115234375,
0.515625,
0.68115234375
],
[
0.51708984375,
0.68310546875,
0.8818359375,
0.68310546875,
0.8818359375,
0.69677734375,
0.51708984375,
0.69677734375
],
[
0.118408203125,
0.6923828125,
0.151123046875,
0.6923828125,
0.151123046875,
0.705078125,
0.118408203125,
0.705078125
],
[
0.1485595703125,
0.6953125,
0.154541015625,
0.6953125,
0.154541015625,
0.70068359375,
0.1485595703125,
0.70068359375
],
[
0.1502685546875,
0.6904296875,
0.264404296875,
0.6904296875,
0.264404296875,
0.705078125,
0.1502685546875,
0.705078125
],
[
0.51708984375,
0.69873046875,
0.880859375,
0.69873046875,
0.880859375,
0.7119140625,
0.51708984375,
0.7119140625
],
[
0.118408203125,
0.71728515625,
0.48779296875,
0.71728515625,
0.48779296875,
0.73095703125,
0.118408203125,
0.73095703125
],
[
0.51318359375,
0.71533203125,
0.5380859375,
0.71533203125,
0.5380859375,
0.72802734375,
0.51318359375,
0.72802734375
],
[
0.118408203125,
0.73291015625,
0.484375,
0.73291015625,
0.484375,
0.74658203125,
0.118408203125,
0.74658203125
],
[
0.53076171875,
0.73291015625,
0.8818359375,
0.73291015625,
0.8818359375,
0.74658203125,
0.53076171875,
0.74658203125
],
[
0.12005615234375,
0.75048828125,
0.48779296875,
0.75048828125,
0.48779296875,
0.76416015625,
0.12005615234375,
0.76416015625
],
[
0.5146484375,
0.75,
0.87890625,
0.75,
0.87890625,
0.76025390625,
0.5146484375,
0.76025390625
],
[
0.12005615234375,
0.7666015625,
0.48779296875,
0.7666015625,
0.48779296875,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.5146484375,
0.76318359375,
0.8818359375,
0.76416015625,
0.8818359375,
0.779296875,
0.5146484375,
0.77783203125
],
[
0.12005615234375,
0.7822265625,
0.489501953125,
0.7822265625,
0.489501953125,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.51318359375,
0.77978515625,
0.8818359375,
0.77978515625,
0.8818359375,
0.79345703125,
0.51318359375,
0.79345703125
],
[
0.118408203125,
0.798828125,
0.486083984375,
0.798828125,
0.486083984375,
0.8125,
0.118408203125,
0.8125
],
[
0.5146484375,
0.79736328125,
0.88330078125,
0.79736328125,
0.88330078125,
0.81103515625,
0.5146484375,
0.81103515625
],
[
0.12005615234375,
0.81640625,
0.486083984375,
0.81640625,
0.486083984375,
0.82666015625,
0.12005615234375,
0.82666015625
],
[
0.515625,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.515625,
0.82763671875
],
[
0.12176513671875,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.12176513671875,
0.84423828125
],
[
0.5146484375,
0.830078125,
0.8818359375,
0.830078125,
0.8818359375,
0.84326171875,
0.5146484375,
0.84326171875
],
[
0.12005615234375,
0.845703125,
0.489501953125,
0.8466796875,
0.489501953125,
0.86083984375,
0.12005615234375,
0.85986328125
],
[
0.5341796875,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.5341796875,
0.86083984375
],
[
0.12005615234375,
0.86328125,
0.322509765625,
0.86328125,
0.322509765625,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.5146484375,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.87744140625,
0.5146484375,
0.87744140625
],
[
0.139404296875,
0.8798828125,
0.489501953125,
0.8798828125,
0.489501953125,
0.8935546875,
0.139404296875,
0.8935546875
],
[
0.515625,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.515625,
0.8935546875
],
[
0.1175537109375,
0.89453125,
0.490234375,
0.8955078125,
0.490234375,
0.91015625,
0.1175537109375,
0.9091796875
],
[
0.515625,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"学习软线性约束及其在引文字段提取中的应用",
"Sam Anzaroot Alexandre Passos David Belanger Andrew McCallum",
"Department of Computer Science",
"University of Massachusetts, Amherst",
"[ anzaroot, apassos, belanger, mccallum } @cs.umass.edu",
"Abstract",
"2008;Peng and McCallum，2004)。这些模",
"型支持高效的动态编程推理，但仅对输出标签",
"Accurately segmenting a citation string",
"序列中的local，依赖项进行建模。然而，引用",
"into fields for authors, titles, etc. is",
"具有这些模型未捕获的强global正则性。例",
"a challenging task because the out-",
"如，许多书籍引用同时包含author部分和",
"put typically obeys various global con-",
"editor部分，但没有一个具有两个不相交的",
"straints.",
"Previous work has shown",
"author部分。由于线性链模型无法捕获超过",
"that modeling soft constraints , where",
"马尔可夫依赖关系，因此这些模型有时会将",
"the model is encouraged, but not re-",
"editor错误地标记为第二作者。如果我们可以",
"quire to obey the constraints, can sub-",
"强制执行只有一个author部分的全局约束，",
"stantially improve segmentation per-",
"则可以提高准确性",
"formance.",
"On the other hand, for",
"将此类全局约束添加到易处理模型中的一个",
"imposing hard constraints , dual de-",
"框架是constrainedinference，在该框架中",
"composition is a popular technique for",
"在推理时，原始模型会对输出进行限制，以使",
"efficient prediction given existing al-",
"其遵守某些全局规则。当硬约束可以编码为",
"gorithms for unconstrained inference.",
"输出变量上的线性方程，并且基础模型的推",
"We extend dual decomposition to per",
"理任务可以作为线性优化时，可以将此约束推",
"form prediction subject to soft con",
"理问题表述为integerlinearprogram（ILP）",
"straints. Moreover, with a technique",
"RothandYih，2004）。或者，可以采用dual",
"forperforminginferencegivensoftcon",
"decomposition（Rushetal.,2010）。双重分解",
"straints, it is easy to automatically gen-",
"相对于LP的优势在于，它可以将原始模型",
"erate large families of constraints and",
"的现有推理算法作为黑匣子。这样的模块化算",
"learn their costs with a simple convex",
"法易于实现，并且在实践中效果很好，为大多",
"optimization problem during training.",
"数示例提供了最优性证书",
"This allows us to obtain substantial",
"上述两种方法以前已应用于对模型的输出施",
"gains in accuracy on a new, challeng-",
"加hard约束。另一方面，最近的工作证明了",
"ing citation extraction dataset.",
"通过施加soft约束（Changetal.，2012）在引",
"文字段提取方面的改进。在这里，该模型不需",
"1",
"",
"Introduction",
"要遵守全局约束，而只是为违反它们而支付罚",
"引文字段提取，是信息提取的一个实例，是将",
"款",
"研究论文引文字符串分割并标记为其组成部分",
"本文介绍了一种通过对偶分解施加软约束的",
"的任务，包括作者，编辑，年份，期刊，卷，会",
"新方法。我们还提出了一种方法，用于学习预",
"议地点等。这项任务很重要，因为引文数据通",
"测问题因违反这些软约束而受到的惩罚。由于",
"常仅以纯文本形式提供；但是，对于许多科学",
"我们的学习方法使许多惩罚变为零，因此它允",
"计量任务，例如绘制科学子社区，发现研究趋",
"许从业者执行“约束选择”，其中可以考虑大",
"势以及分析研究人员网络，必须拥有准确的书",
"量自动生成的候选全局约束，并自动将其剔除",
"目信息结构化数据库。。自动引文字段提取需",
"为较小的有用约束集，可以在测试时快速运行",
"要进一步研究，因为它尚未达到可以在实际系",
"使用我们的新方法，我们不仅能够合并",
"统中实际部署的精度水平",
"Chang等人的所有软全局约束。（2012），但",
"隐马尔可夫模型和线性链条件随机场",
"也有更复杂的数据驱动约束，同时也提供了",
"（CRFs）以前已应用于引文提取E(Hetzner，",
"比他们的光束搜索技术更强的最优性证书。在"
] |
{
"arxiv_id": "1403.1349",
"language": "zh",
"page_idx": 0
}
|
||
1403.1349-zh-1
|
1403.1349-zh-1.jpg
|
4 .ref-marker[ J.first D. middle Monk ,last person $]_\text{authors}[$ Cardinal Functions on Boolean Algebra , ]title[ Lectures in Mathematics ,ETH Zurich , series Birkh ause Verlag , publisher Basel , Boston , Berlin , address 1990. year date $\int_{\mathrm{venue}}$
Figure 1: Example labeled citation
<table>
<tbody>
<tr>
<td>$a_{t}^{(t)}$ =argmax $_{u\in\mathcal{U}}\left\langle w+A^{T}\lambda,y\right\rangle$</td>
</tr>
<tr>
<td>$\lambda^{(t)}=\Pi_{0\leq\cdot}\left[\int_{0}^{(t)}dt\right]$ $\lambda^{(t-1)}-\eta^{(t)}(Ay-b)$</td>
</tr>
</tbody>
</table>
一个新的、更具代表性和挑战性的引文字段提取数据集上，我们表明我们的方法实现了 17。9 与线性链条件随机场相比，% 减少了错误。此外，我们证明了我们的推理技术可以使用Chang 等人的约束并从中受益。(2012),但在这些之上包括我们的数据驱动约束是有益的。虽然本文专注于引文字段提取的应用，但这里介绍的新颖方法很容易推广到许多具有全局输出规则的问题
2.2 Dual Decomposition for Global
Constraints
# 2 Background
2.1 Structured Linear Models
为了执行受各种全局约束的预测，我们可能需要使用其他约束来增强问题$\overline{\mathrm{E}}(1)$ 。双重分解是在这种情况下执行 MAP 推理的一种流行方法，因为它利用了基本问题中 MAP 的已知算法，其中没有添加这些额外的约束(Ko- ${\mathrm{modakis~et~al.,~2007;~Sontag~et~al.,~2011;~Rush}}$ and Collins, 2012)。在这种情况下，MAP 问题可以表述为类似于方程 (1)的结构化线性模型，为此我们有 MAP 算法，但是我们施加了一些额外的约束$Ay\leq b$ ,这些约束不再允许我们使用该算法。换句话说，我们考虑任意矩阵$A$和向量$b$的问题
$$\begin{array}{rl}max.&\langle w,y\rangle\\s.t.&y\in\mathcal{U}\\&Ay\leq b,\end{array}$$
(2)
我们采用的整体建模技术是将软约束添加到一个简单的模型中，我们已经有了一个有效的预测算法。对于此基础模型，我们采用链结构条件随机场 (CRF),因为已证明 CRF 的性能优于其他简单的无约束模型，例如用于引文提取 (Peng and McCallum, 2004) 的隐马尔可夫模型。我们通过执行 MAP 推断 (Koller and Friedman, 2009) 来产生预测
CRF 中的 MAP 推理任务可以表示为具有线性目标$\mathsf E($Sontag, 2010; Sontag et al., 2011) 的优化问题。在这里，我们为图形模型中每个因子的每个候选设置定义一个二进制指标变量。这些指标变量中的每一个都与因子在具有指标变量的相应值时所获得的分数相关联。由于 CRF 中某些$y$的对数概率与所有因素的得分之和成正比，因此我们可以将指标变量连接为向量$y$,将得分连接为向量$w$,并将 MAP 问题写为
。我们可以将此问题的拉格朗日写为
$$L(y,\lambda)=\langle w,y\rangle+\lambda^T(Ay-b).$$
(3)
重新分组项并在原始变量上最大化，我们有对
偶问题
(4
$$\boldsymbol{min.}_\lambda D(\lambda)=\max_{y\in\mathcal{U}}\left\langle w+A^T\lambda,y\right\rangle-\lambda^Tb.$$
$$\begin{matrix}max.&\langle w,y\rangle\\s.t.&y\in\mathcal{U},\end{matrix}$$
(1)
对于任何$\lambda$ ,我们可以评估对偶目标$D( \lambda )$ , 因为在$\mathbb{E}(4)$ 中的最大化具有与原始问题相同。此外，对于最大化该内部优化问题的$y^*,D(\lambda)$ 的次梯度为$Ay^*-b$ 。因此，我们可以用投影次梯度法 (Boyd and Vandenberghe, 2004) 最小化$D(\lambda)$ ,并且在评估$D(\lambda^*)$时可以获得最佳的$y$ 。请注意，$D(\lambda)$的次梯度是在$y$上最大化时$\lambda$违反每个约束的量
算法 1 描述了对偶分解的基本投影次梯度下降算法。投影运算符$\Pi$包括将$\lambda$的所有负坐标截断为 0。这是必要的，因为$\lambda$是不等式约束的对偶变量向量。当$y^{(t)}$满足相等的每个约束或 $\lambda$ 的相应分量为 0 时，由于互补松弛$\boxed{\mathrm{E}}$ (Boyd and Vandenberghe, 2004) , 算法已收敛
, 其中集合$U$ 表示指标变量的有效配置集。在这里，约束是所有相邻因素在$y$的重叠部分上都一致
结构化线性模型是一般的模型族，其中预测需要解决$\overline{\mathrm{E}}(1)$ 形式的问题，并且它们并不总是对应于概率模型。我们在后面的章节中介绍的用于处理软全局约束和学习这些约束的惩罚的算法可以应用于一般的结构化线性模型，而不仅仅是 CRF,只要我们有一个可用的算法来执行 MAP 推理
|
[
[
0.1502685546875,
0.0736083984375,
0.39208984375,
0.0736083984375,
0.39208984375,
0.0843505859375,
0.1502685546875,
0.0843505859375
],
[
0.426513671875,
0.0706787109375,
0.8818359375,
0.07244873046875,
0.8818359375,
0.08612060546875,
0.426513671875,
0.0843505859375
],
[
0.1259765625,
0.07781982421875,
0.15283203125,
0.07781982421875,
0.15283203125,
0.08197021484375,
0.1259765625,
0.08197021484375
],
[
0.118408203125,
0.0843505859375,
0.87744140625,
0.0843505859375,
0.87744140625,
0.09857177734375,
0.118408203125,
0.09857177734375
],
[
0.12005615234375,
0.09857177734375,
0.1822509765625,
0.09857177734375,
0.1822509765625,
0.1092529296875,
0.12005615234375,
0.1092529296875
],
[
0.35693359375,
0.12353515625,
0.63818359375,
0.12353515625,
0.63818359375,
0.13720703125,
0.35693359375,
0.13720703125
],
[
0.133544921875,
0.160888671875,
0.48779296875,
0.160888671875,
0.48779296875,
0.174560546875,
0.133544921875,
0.174560546875
],
[
0.5146484375,
0.162109375,
0.8818359375,
0.162109375,
0.8818359375,
0.177001953125,
0.5146484375,
0.177001953125
],
[
0.12176513671875,
0.179931640625,
0.489501953125,
0.179931640625,
0.489501953125,
0.190673828125,
0.12176513671875,
0.190673828125
],
[
0.5146484375,
0.1787109375,
0.85205078125,
0.1787109375,
0.85205078125,
0.1923828125,
0.5146484375,
0.1923828125
],
[
0.12005615234375,
0.19482421875,
0.484375,
0.19482421875,
0.484375,
0.2078857421875,
0.12005615234375,
0.2078857421875
],
[
0.5205078125,
0.1923828125,
0.7314453125,
0.193603515625,
0.7314453125,
0.2078857421875,
0.5205078125,
0.207275390625
],
[
0.52490234375,
0.208984375,
0.54150390625,
0.208984375,
0.54150390625,
0.218505859375,
0.52490234375,
0.218505859375
],
[
0.54150390625,
0.210205078125,
0.5693359375,
0.210205078125,
0.5693359375,
0.218505859375,
0.54150390625,
0.218505859375
],
[
0.5634765625,
0.207275390625,
0.78076171875,
0.207275390625,
0.78076171875,
0.221435546875,
0.5634765625,
0.221435546875
],
[
0.56494140625,
0.2203369140625,
0.8017578125,
0.2203369140625,
0.8017578125,
0.2410888671875,
0.56494140625,
0.2410888671875
],
[
0.12005615234375,
0.2120361328125,
0.483642578125,
0.2120361328125,
0.483642578125,
0.22265625,
0.12005615234375,
0.22265625
],
[
0.12005615234375,
0.22802734375,
0.48779296875,
0.22802734375,
0.48779296875,
0.2410888671875,
0.12005615234375,
0.2410888671875
],
[
0.51904296875,
0.232177734375,
0.5341796875,
0.223876953125,
0.5439453125,
0.232177734375,
0.52880859375,
0.240478515625
],
[
0.12005615234375,
0.243408203125,
0.48779296875,
0.243408203125,
0.48779296875,
0.257080078125,
0.12005615234375,
0.257080078125
],
[
0.12005615234375,
0.26123046875,
0.484375,
0.26123046875,
0.484375,
0.271484375,
0.12005615234375,
0.271484375
],
[
0.515625,
0.271484375,
0.54150390625,
0.271484375,
0.54150390625,
0.283203125,
0.515625,
0.283203125
],
[
0.56103515625,
0.270751953125,
0.8154296875,
0.270751953125,
0.8154296875,
0.283935546875,
0.56103515625,
0.283935546875
],
[
0.5458984375,
0.27490234375,
0.5556640625,
0.27490234375,
0.5556640625,
0.279052734375,
0.5458984375,
0.279052734375
],
[
0.118408203125,
0.27490234375,
0.486083984375,
0.27490234375,
0.486083984375,
0.2880859375,
0.118408203125,
0.2880859375
],
[
0.12176513671875,
0.29345703125,
0.267822265625,
0.29345703125,
0.267822265625,
0.303955078125,
0.12176513671875,
0.303955078125
],
[
0.5576171875,
0.2880859375,
0.6513671875,
0.2880859375,
0.6513671875,
0.2998046875,
0.5576171875,
0.2998046875
],
[
0.51708984375,
0.30810546875,
0.8818359375,
0.30810546875,
0.8818359375,
0.32177734375,
0.51708984375,
0.32177734375
],
[
0.5146484375,
0.322509765625,
0.880859375,
0.322509765625,
0.880859375,
0.336181640625,
0.5146484375,
0.336181640625
],
[
0.1175537109375,
0.330810546875,
0.15869140625,
0.330810546875,
0.15869140625,
0.3466796875,
0.1175537109375,
0.3466796875
],
[
0.15283203125,
0.33203125,
0.26025390625,
0.33203125,
0.26025390625,
0.3466796875,
0.15283203125,
0.3466796875
],
[
0.51318359375,
0.33837890625,
0.88330078125,
0.339111328125,
0.88330078125,
0.35400390625,
0.51318359375,
0.352783203125
],
[
0.51318359375,
0.355224609375,
0.880859375,
0.355224609375,
0.880859375,
0.3681640625,
0.51318359375,
0.3681640625
],
[
0.118408203125,
0.366455078125,
0.17041015625,
0.366455078125,
0.17041015625,
0.38134765625,
0.118408203125,
0.38134765625
],
[
0.166259765625,
0.3681640625,
0.372802734375,
0.3681640625,
0.372802734375,
0.37890625,
0.166259765625,
0.37890625
],
[
0.515625,
0.372314453125,
0.880859375,
0.372314453125,
0.880859375,
0.385986328125,
0.515625,
0.385986328125
],
[
0.51318359375,
0.388427734375,
0.8818359375,
0.388427734375,
0.8818359375,
0.402099609375,
0.51318359375,
0.402099609375
],
[
0.12005615234375,
0.394287109375,
0.470947265625,
0.394287109375,
0.470947265625,
0.407958984375,
0.12005615234375,
0.407958984375
],
[
0.51318359375,
0.40380859375,
0.8818359375,
0.40380859375,
0.8818359375,
0.41748046875,
0.51318359375,
0.41748046875
],
[
0.118408203125,
0.410888671875,
0.486083984375,
0.410888671875,
0.486083984375,
0.424560546875,
0.118408203125,
0.424560546875
],
[
0.515625,
0.419189453125,
0.8818359375,
0.419189453125,
0.8818359375,
0.43408203125,
0.515625,
0.43408203125
],
[
0.118408203125,
0.427001953125,
0.486083984375,
0.427001953125,
0.486083984375,
0.440673828125,
0.118408203125,
0.440673828125
],
[
0.5146484375,
0.435791015625,
0.880859375,
0.435791015625,
0.880859375,
0.449462890625,
0.5146484375,
0.449462890625
],
[
0.12005615234375,
0.44482421875,
0.484375,
0.44482421875,
0.484375,
0.454833984375,
0.12005615234375,
0.454833984375
],
[
0.5283203125,
0.45068359375,
0.8818359375,
0.451904296875,
0.8818359375,
0.466064453125,
0.5283203125,
0.465576171875
],
[
0.12005615234375,
0.460205078125,
0.48779296875,
0.460205078125,
0.48779296875,
0.473876953125,
0.12005615234375,
0.473876953125
],
[
0.51318359375,
0.468505859375,
0.8818359375,
0.468505859375,
0.8818359375,
0.482177734375,
0.51318359375,
0.482177734375
],
[
0.12005615234375,
0.4755859375,
0.48779296875,
0.4755859375,
0.48779296875,
0.4892578125,
0.12005615234375,
0.4892578125
],
[
0.5146484375,
0.48388671875,
0.68994140625,
0.48388671875,
0.68994140625,
0.49755859375,
0.5146484375,
0.49755859375
],
[
0.12005615234375,
0.4912109375,
0.48779296875,
0.4912109375,
0.48779296875,
0.5048828125,
0.12005615234375,
0.5048828125
],
[
0.12005615234375,
0.50830078125,
0.349365234375,
0.50830078125,
0.349365234375,
0.521484375,
0.12005615234375,
0.521484375
],
[
0.634765625,
0.50830078125,
0.74462890625,
0.50830078125,
0.74462890625,
0.525390625,
0.634765625,
0.525390625
],
[
0.6533203125,
0.52490234375,
0.74462890625,
0.52490234375,
0.74462890625,
0.5390625,
0.6533203125,
0.5390625
],
[
0.8564453125,
0.52392578125,
0.884765625,
0.52392578125,
0.884765625,
0.54150390625,
0.8564453125,
0.54150390625
],
[
0.6962890625,
0.54052734375,
0.759765625,
0.54052734375,
0.759765625,
0.5546875,
0.6962890625,
0.5546875
],
[
0.1368408203125,
0.525390625,
0.486083984375,
0.52685546875,
0.486083984375,
0.54150390625,
0.1368408203125,
0.54052734375
],
[
0.118408203125,
0.5419921875,
0.48779296875,
0.54345703125,
0.48779296875,
0.55810546875,
0.118408203125,
0.55712890625
],
[
0.12005615234375,
0.56005859375,
0.48779296875,
0.56005859375,
0.48779296875,
0.57373046875,
0.12005615234375,
0.57373046875
],
[
0.5205078125,
0.56005859375,
0.8154296875,
0.56005859375,
0.8154296875,
0.57373046875,
0.5205078125,
0.57373046875
],
[
0.12005615234375,
0.576171875,
0.48779296875,
0.576171875,
0.48779296875,
0.5888671875,
0.12005615234375,
0.5888671875
],
[
0.57861328125,
0.5859375,
0.82373046875,
0.5859375,
0.82373046875,
0.599609375,
0.57861328125,
0.599609375
],
[
0.81787109375,
0.5888671875,
0.85791015625,
0.5888671875,
0.85791015625,
0.5966796875,
0.81787109375,
0.5966796875
],
[
0.85546875,
0.58740234375,
0.880859375,
0.58740234375,
0.880859375,
0.599609375,
0.85546875,
0.599609375
],
[
0.12005615234375,
0.60791015625,
0.48779296875,
0.60791015625,
0.48779296875,
0.62158203125,
0.12005615234375,
0.62158203125
],
[
0.12176513671875,
0.5927734375,
0.483642578125,
0.5927734375,
0.483642578125,
0.6025390625,
0.12176513671875,
0.6025390625
],
[
0.5146484375,
0.60986328125,
0.8818359375,
0.61083984375,
0.8818359375,
0.6259765625,
0.5146484375,
0.62451171875
],
[
0.12176513671875,
0.6259765625,
0.48779296875,
0.6259765625,
0.48779296875,
0.6357421875,
0.12176513671875,
0.6357421875
],
[
0.515625,
0.62890625,
0.5693359375,
0.62890625,
0.5693359375,
0.64013671875,
0.515625,
0.64013671875
],
[
0.12005615234375,
0.64013671875,
0.48779296875,
0.64013671875,
0.48779296875,
0.65380859375,
0.12005615234375,
0.65380859375
],
[
0.12005615234375,
0.65625,
0.486083984375,
0.65625,
0.486083984375,
0.67041015625,
0.12005615234375,
0.67041015625
],
[
0.5263671875,
0.65283203125,
0.8818359375,
0.65283203125,
0.8818359375,
0.66748046875,
0.5263671875,
0.66748046875
],
[
0.6484375,
0.66650390625,
0.67529296875,
0.66455078125,
0.6767578125,
0.6728515625,
0.64990234375,
0.6748046875
],
[
0.12005615234375,
0.6728515625,
0.1956787109375,
0.6728515625,
0.1956787109375,
0.68701171875,
0.12005615234375,
0.68701171875
],
[
0.5146484375,
0.68505859375,
0.884765625,
0.68603515625,
0.884765625,
0.70068359375,
0.5146484375,
0.69970703125
],
[
0.246826171875,
0.69970703125,
0.353515625,
0.69970703125,
0.353515625,
0.71435546875,
0.246826171875,
0.71435546875
],
[
0.462646484375,
0.7080078125,
0.489501953125,
0.7080078125,
0.489501953125,
0.72021484375,
0.462646484375,
0.72021484375
],
[
0.515625,
0.703125,
0.8818359375,
0.703125,
0.8818359375,
0.71630859375,
0.515625,
0.71630859375
],
[
0.2626953125,
0.71533203125,
0.358642578125,
0.71533203125,
0.358642578125,
0.72998046875,
0.2626953125,
0.72998046875
],
[
0.515625,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.515625,
0.732421875
],
[
0.5146484375,
0.73388671875,
0.87890625,
0.73388671875,
0.87890625,
0.74755859375,
0.5146484375,
0.74755859375
],
[
0.145263671875,
0.7451171875,
0.48779296875,
0.7451171875,
0.48779296875,
0.7587890625,
0.145263671875,
0.7587890625
],
[
0.51318359375,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.51318359375,
0.76318359375
],
[
0.12005615234375,
0.76123046875,
0.486083984375,
0.76123046875,
0.486083984375,
0.77490234375,
0.12005615234375,
0.77490234375
],
[
0.5146484375,
0.765625,
0.880859375,
0.765625,
0.880859375,
0.779296875,
0.5146484375,
0.779296875
],
[
0.12005615234375,
0.779296875,
0.19482421875,
0.779296875,
0.19482421875,
0.79345703125,
0.12005615234375,
0.79345703125
],
[
0.515625,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.515625,
0.79638671875
],
[
0.1376953125,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.1376953125,
0.8125
],
[
0.51318359375,
0.798828125,
0.73876953125,
0.798828125,
0.73876953125,
0.8125,
0.51318359375,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.48779296875,
0.81396484375,
0.48779296875,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.53076171875,
0.8154296875,
0.8818359375,
0.8154296875,
0.8818359375,
0.8291015625,
0.53076171875,
0.8291015625
],
[
0.12005615234375,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.51708984375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.51708984375,
0.84423828125
],
[
0.12005615234375,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.515625,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.515625,
0.86083984375
],
[
0.12005615234375,
0.86328125,
0.48779296875,
0.86328125,
0.48779296875,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.51318359375,
0.86328125,
0.880859375,
0.86328125,
0.880859375,
0.876953125,
0.51318359375,
0.876953125
],
[
0.118408203125,
0.87744140625,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.118408203125,
0.892578125
],
[
0.51318359375,
0.87890625,
0.88330078125,
0.87744140625,
0.88330078125,
0.892578125,
0.51318359375,
0.8935546875
],
[
0.118408203125,
0.8955078125,
0.267822265625,
0.8955078125,
0.267822265625,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.515625,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"-marker[J.first D.middle Monk,las",
"on lauthors[ Cardinal Functions on Boolean Algebra, Jtitle [ Lectures",
"F-rof",
"in Mathematics , ETH Zurich , series Birkh\"ause Verlag , publisher Basel , Boston , Berlin , address i990 . yea",
"dateJvenue",
"Figure l: Example labeled citation",
"个新的、更具代表性和挑战性的引文字段提",
"Algorithm 1 DD: projected subgradient for",
"取数据集上，我们表明我们的方法实现了17。",
"dual decomposition with hard constraints",
"9与线性链条件随机场相比，%减少了错误，",
"1: while has not converged do",
"2:",
"",
"y(t) = argmaxyeu (w + AT入,y)",
"^() = Io.[<(t-1) - n((Ay - 6)]",
"此外，我们证明了我们的推理技术可以使用",
"Chang等人的约束并从中受益。。（2012），但",
"3",
"在这些之上包括我们的数据驱动约束是有益",
"的。虽然本文专注于引文字段提取的应用，但",
"2.2",
"Dual Decomposition for Global",
"",
"这里介绍的新颖方法很容易推广到许多具有全",
"局输出规则的问题",
"Constraints",
"为了执行受各种全局约束的预测，我们可能",
"需要使用其他约束来增强问题E(1）。双重分",
"2",
"Background",
"解是在这种情况下执行MAP推理的一种流",
"行方法，因为它利用了基本问题中MAP的已",
"2.1",
"Structured Linear Models",
"知算法，其中没有添加这些额外的约束（Ko-",
"modakis et al., 2007; Sontag et al., 2011; Rush",
"我们采用的整体建模技术是将软约束添加到",
"andCollins，2012）。在这种情况下，MAP问",
"个简单的模型中，我们已经有了一个有效的预",
"题可以表述为类似于方程（1）的结构化线性模",
"测算法。对于此基础模型，我们采用链结构条",
"型，为此我们有MAP算法，但是我们施加了",
"件随机场（CRF），因为已证明CRF的性能",
"-些额外的约束Ay≤b，这些约束不再允许",
"优于其他简单的无约束模型，例如用于引文提",
"我们使用该算法。换句话说，我们考虑任意矩",
"取（PengandMcCallum，2004）的隐马尔可夫",
"阵 A 和向量 b 的问题",
"模型。我们通过执行MAP推断（Kollerand",
"Friedman，2009）来产生预测",
"ma.(w,y)",
"s.t. yeu",
"(2)",
"Ay ≤ b,",
"CRF中的MAP推理任务可以表示为具有",
"线性目标E(Sontag，2010;Sontagetal.,2011)",
"的优化问题。在这里，我们为图形模型中每个",
"。我们可以将此问题的拉格朗日写为",
"因子的每个候选设置定义一个二进制指标变",
"L(y, ^) = (w, y) + ^T(Ay -b).",
"",
"(3)",
"指标变量的相应值时所获得的分数相关联。由",
"量。这些指标变量中的每一个都与因子在具有",
"重新分组项并在原始变量上最大化，我们有对",
"于CRF中某些y的对数概率与所有因素的得",
"偶问题",
"分之和成正比，因此我们可以将指标变量连接",
"为向量y，将得分连接为向量W，并将MAP",
"min.^D() = max(w + ATX, y) - ^Tb. (4)",
"Ye",
"问题写为",
"对于任何入，我们可以评估对偶目标D(>），",
"mar.(w,y)",
"(1)",
"因为在E(4)中的最大化具有与原始问题相同。",
"s.t. y eu,",
"此外，对于最大化该内部优化问题的y*，D()",
"的次梯度为Ay*一b。因此，我们可以用投影",
"其中集合U表示指标变量的有效配置集。",
"次梯度法（BoydandVandenberghe，2004）最",
"在这里，约束是所有相邻因素在9的重叠部分",
"小化D(^），并且在评估D(*）时可以获得最",
"上都一致",
"佳的y。请注意，D()的次梯度是在y上最",
"结构化线性模型是一般的模型族，其中预测",
"大化时入违反每个约束的量",
"需要解决E(1）形式的问题，并且它们并不总",
"算法1描述了对偶分解的基本投影次梯度",
"是对应于概率模型。我们在后面的章节中介绍",
"下降算法。投影运算符ⅡI包括将入的所有负",
"的用于处理软全局约束和学习这些约束的惩罚",
"坐标截断为0。这是必要的，因为入是不等式",
"的算法可以应用于一般的结构化线性模型，而",
"约束的对偶变量向量。当y(t)满足相等的每个",
"不仅仅是CRF，只要我们有一个可用的算法",
"约束或入的相应分量为0时，由于互补松弛E",
"来执行MAP推理",
"(BoydandVandenberghe，2004），算法已收敛"
] |
{
"arxiv_id": "1403.1349",
"language": "zh",
"page_idx": 1
}
|
||
1403.1349-zh-2
|
1403.1349-zh-2.jpg
|
3 Soft Constraints in Dual
## Decomposition
我们现在介绍算法$\overline{\mathrm{E}}$1 的扩展以处理软约束。在我们的公式中，软约束模型对每个未满足的约束施加惩罚，与违反的数量成比例。因此， 我们的推导类似于如何通过引人辅助松弛变量(Cortes and Vapnik, 1995) 从硬裕度 SVM 中导出软裕度 SVM。请注意，在执行受软约束的 MAP 时，最优解可能不满足某些约束，因为这样做会使模型的分数降低太多
考虑以下形式的优化问题：
$$\begin{array}{rl}max.&\langle w,y\rangle-\langle c,z\rangle\\s.t.&y\in\mathcal U\\&Ay-b\leq z\\&-z\leq0,\end{array}$$
(5)
对于正$c_i$ ,很明显，最佳$z_i$将等于违反$a_i^Ty\leq b_i$的程度。因此，我们支付的成本是违反第$i$个约束的程度的$c_i$倍，这反映了松弛变量如何用于表示 SVM 的铰链损失。请注意， $c_i$必须为正，否则此线性程序是无界的，并且可以通过将$z_i$设置为无穷大来获得最佳值
使用与 2.2 节中类似的构造，我们将拉格朗
日写为：
$$\begin{array}{c}L(y,z,\lambda,\mu)=\langle w,y\rangle-\langle c,z\rangle\\+\lambda^T(Ay-b-z)+\mu^T(-z).\end{array}$$
(6)
关于 $z$ 的最优性约束告诉我们$-c-\lambda-\mu=0$
,因此$\mu=-c-\lambda$。代人，我们有
$$L(y,\lambda)=\langle w,y\rangle+\lambda^T(Ay-b),$$
(7)
,除了$\mu=-c-\lambda$暗示$\mu$为正的$\lambda\leq c$的约束。由于此拉格朗日具有与方程(3)相同的形式，因此我们还可以推导出一个对偶问题该对偶问题与方程 (4)中的对偶问题相同，并带有附加约束，即每个$\lambda_i$不能大于其成本$c_i$ 。换句话说，对偶问题不能惩罚违反约束的行为，而原始中的软约束模型在违反约束时会惩罚您。
这个优化问题仍然可以用投影次梯度下降来解决，并在算法 2 中进行了描述。对算法$\overline{\mathrm{E}}$1的唯一修改是用$\Pi_{0\leq\cdot\leq c}$替换坐标投影$\Pi_{0\leq\cdot}$, 以及我们如何检查收敛。现在，我们检查 (5) 的 KKT 条件，其中对于每个约束$i$ , 约束满足相等，$\lambda_i=0$或$\lambda_i=c_i$
因此，实现软约束对偶分解与实现硬约束对偶分解一样容易，并且每次迭代的复杂度相同。我们鼓励将软约束对偶分解进一步应用于现有和新的 NLP 问题
<table>
<tbody>
<tr>
<td>$\overline{\mathrm{Al}_{\xi}}$ 1 1 1 L 1 I 1</td>
<td>rithm 2 Soft-DD: projected subgradient</td>
</tr>
<tr>
<td>1: </td>
<td>$\mathrm{hil}_{t}$ $\overline{\mathrm{l~do}}$ converged</td>
</tr>
<tr>
<td>2: 1</td>
<td>$.(t)$ $=\operatorname{argmax}_{y\in\mathcal{U}}\left\langle w+A^{T}\lambda,y\right\rangle$ $y$</td>
</tr>
<tr>
<td>3:</td>
<td>$\lambda^{(t)}$ -1</td>
</tr>
</tbody>
</table>
## 3.1 Learning Penalties
使用 softv 时的一个注意事项。s。硬约束是软约束提出了一个新的训练问题，因为我们需要选择向量$c$ ,违反约束的惩罚。问题$\overline{\mathrm{F}}(5)$在上一节中的一个重要属性是，它对应于$y$和$z$ 上的结构化线性模型。因此，我们可以应用已知的训练算法来估计结构化线性模型的参数， 以选择$c$
我们所需要采用的结构化感知器算法(Collins, 2002) 或结构化 SVM 算法 (Tsochantaridis et al., 2004) 是一个黑盒过程，用于在给定任意成本向量的情况下在结构化线性模型中执行 MAP 推理。幸运的是$,\operatorname{\overline{E}}(5)$ 的 MAP 问题可以在算法$\overline{\mathrm{E}}$2 中使用 Soft-DD 解决。每个惩罚$c_i$必须是非负的；否则，方程 (5)中的优化问题是不明确的。这可以通过对感知器的简单修改和结构化 SVM 目标的次梯度下降优化来确保，只需在每次学习迭代中将$c$按坐标截断为非负即可
从直觉上讲，感知器更新会在满足基本事实而不是推断的预测时增加约束的惩罚，而在满足预测而不是基本事实时减少约束的惩罚。由于我们将惩罚截断为0，这表明我们将学习三类约束的惩罚为 0: 不符合基本真理的约束， 符合基本真理但在实践中通过在基本 CRF 模型中执行推理而得到满足的约束，以及在实践中得到满足的约束，这是对其他一些约束施加非零惩罚的副作用。类似的分析适用于结构化SVM 方法
因此，我们不仅可以将学习惩罚的值视为参数调整，还可以将其视为执行“约束选择”的一种手段，因为可以忽略惩罚为 0 的约束。此属性允许我们考虑大型约束族，从中自动识别有用的约束族
我们发现，尽管在理论上没有必要，但与其他模型参数分开学习对保持的开发集的约束是有益的，因为在训练过程中，由于过度拟合， 大多数约束都得到了满足，这导致低估了相关的惩罚
4 Citation Extraction Data
我们考虑在 Anzaroot 和 McCallum (2013) 中首次引入的 UMass 引文数据集。它有来自许多学术领域的 1800 多个引文，摘自 arXiv。此
|
[
[
0.118408203125,
0.07659912109375,
0.366943359375,
0.07659912109375,
0.366943359375,
0.09027099609375,
0.118408203125,
0.09027099609375
],
[
0.5146484375,
0.0748291015625,
0.880859375,
0.07598876953125,
0.880859375,
0.09027099609375,
0.5146484375,
0.08905029296875
],
[
0.1502685546875,
0.09320068359375,
0.27783203125,
0.09320068359375,
0.27783203125,
0.10687255859375,
0.1502685546875,
0.10687255859375
],
[
0.51318359375,
0.0914306640625,
0.86962890625,
0.0914306640625,
0.86962890625,
0.1051025390625,
0.51318359375,
0.1051025390625
],
[
0.5205078125,
0.10809326171875,
0.7314453125,
0.10809326171875,
0.7314453125,
0.1217041015625,
0.5205078125,
0.1217041015625
],
[
0.12176513671875,
0.12054443359375,
0.48779296875,
0.12054443359375,
0.48779296875,
0.1312255859375,
0.12176513671875,
0.1312255859375
],
[
0.52490234375,
0.12469482421875,
0.5439453125,
0.12469482421875,
0.5439453125,
0.133056640625,
0.52490234375,
0.133056640625
],
[
0.55908203125,
0.12054443359375,
0.78271484375,
0.12054443359375,
0.78271484375,
0.1383056640625,
0.55908203125,
0.1383056640625
],
[
0.5458984375,
0.1270751953125,
0.56494140625,
0.1270751953125,
0.56494140625,
0.1312255859375,
0.5458984375,
0.1312255859375
],
[
0.12005615234375,
0.1353759765625,
0.48779296875,
0.1353759765625,
0.48779296875,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.5283203125,
0.1383056640625,
0.54248046875,
0.1436767578125,
0.53564453125,
0.152587890625,
0.5205078125,
0.1478271484375
],
[
0.56689453125,
0.13720703125,
0.81689453125,
0.13720703125,
0.81689453125,
0.155029296875,
0.56689453125,
0.155029296875
],
[
0.12005615234375,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.1650390625,
0.12005615234375,
0.1650390625
],
[
0.12176513671875,
0.16748046875,
0.48779296875,
0.16748046875,
0.48779296875,
0.18115234375,
0.12176513671875,
0.18115234375
],
[
0.1234130859375,
0.182861328125,
0.486083984375,
0.182861328125,
0.486083984375,
0.196533203125,
0.1234130859375,
0.196533203125
],
[
0.5146484375,
0.189453125,
0.5556640625,
0.189453125,
0.5556640625,
0.20068359375,
0.5146484375,
0.20068359375
],
[
0.55908203125,
0.18701171875,
0.7138671875,
0.188232421875,
0.7138671875,
0.203125,
0.55908203125,
0.201904296875
],
[
0.12176513671875,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.12176513671875,
0.2120361328125
],
[
0.12005615234375,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.12005615234375,
0.2286376953125
],
[
0.5146484375,
0.210205078125,
0.8818359375,
0.210205078125,
0.8818359375,
0.223876953125,
0.5146484375,
0.223876953125
],
[
0.12005615234375,
0.232177734375,
0.3955078125,
0.232177734375,
0.3955078125,
0.2452392578125,
0.12005615234375,
0.2452392578125
],
[
0.515625,
0.226806640625,
0.8818359375,
0.226806640625,
0.8818359375,
0.240478515625,
0.515625,
0.240478515625
],
[
0.1376953125,
0.2476806640625,
0.345947265625,
0.2476806640625,
0.345947265625,
0.26123046875,
0.1376953125,
0.26123046875
],
[
0.515625,
0.24462890625,
0.87890625,
0.24462890625,
0.87890625,
0.254638671875,
0.515625,
0.254638671875
],
[
0.51708984375,
0.25830078125,
0.88330078125,
0.25830078125,
0.88330078125,
0.271484375,
0.51708984375,
0.271484375
],
[
0.21826171875,
0.276611328125,
0.38623046875,
0.276611328125,
0.38623046875,
0.29150390625,
0.21826171875,
0.29150390625
],
[
0.51708984375,
0.27490234375,
0.880859375,
0.27490234375,
0.880859375,
0.2880859375,
0.51708984375,
0.2880859375
],
[
0.2333984375,
0.292236328125,
0.32666015625,
0.290283203125,
0.32666015625,
0.30810546875,
0.2333984375,
0.31005859375
],
[
0.515625,
0.290283203125,
0.88330078125,
0.290283203125,
0.88330078125,
0.303955078125,
0.515625,
0.303955078125
],
[
0.276123046875,
0.30712890625,
0.36865234375,
0.308837890625,
0.36865234375,
0.32373046875,
0.276123046875,
0.32177734375
],
[
0.46337890625,
0.3017578125,
0.48779296875,
0.3017578125,
0.48779296875,
0.314208984375,
0.46337890625,
0.314208984375
],
[
0.515625,
0.30810546875,
0.5859375,
0.30810546875,
0.5859375,
0.319580078125,
0.515625,
0.319580078125
],
[
0.27783203125,
0.32470703125,
0.343505859375,
0.32470703125,
0.343505859375,
0.339111328125,
0.27783203125,
0.339111328125
],
[
0.5322265625,
0.32373046875,
0.88330078125,
0.32373046875,
0.88330078125,
0.33740234375,
0.5322265625,
0.33740234375
],
[
0.5146484375,
0.33837890625,
0.880859375,
0.339111328125,
0.880859375,
0.35400390625,
0.5146484375,
0.352783203125
],
[
0.1292724609375,
0.3486328125,
0.48779296875,
0.3486328125,
0.48779296875,
0.3623046875,
0.1292724609375,
0.3623046875
],
[
0.5146484375,
0.355712890625,
0.88330078125,
0.355712890625,
0.88330078125,
0.369384765625,
0.5146484375,
0.369384765625
],
[
0.118408203125,
0.364013671875,
0.489501953125,
0.363525390625,
0.489501953125,
0.377685546875,
0.118408203125,
0.37890625
],
[
0.118408203125,
0.37890625,
0.48779296875,
0.38134765625,
0.48779296875,
0.3955078125,
0.118408203125,
0.393798828125
],
[
0.515625,
0.37353515625,
0.880859375,
0.37353515625,
0.880859375,
0.38427734375,
0.515625,
0.38427734375
],
[
0.515625,
0.38720703125,
0.8818359375,
0.38720703125,
0.8818359375,
0.40087890625,
0.515625,
0.40087890625
],
[
0.12005615234375,
0.396728515625,
0.489501953125,
0.396728515625,
0.489501953125,
0.410400390625,
0.12005615234375,
0.410400390625
],
[
0.515625,
0.40380859375,
0.8818359375,
0.40380859375,
0.8818359375,
0.41748046875,
0.515625,
0.41748046875
],
[
0.118408203125,
0.412109375,
0.489501953125,
0.410888671875,
0.489501953125,
0.42578125,
0.118408203125,
0.427001953125
],
[
0.5146484375,
0.419189453125,
0.880859375,
0.419189453125,
0.880859375,
0.432861328125,
0.5146484375,
0.432861328125
],
[
0.12176513671875,
0.4287109375,
0.45751953125,
0.4287109375,
0.45751953125,
0.4423828125,
0.12176513671875,
0.4423828125
],
[
0.515625,
0.435791015625,
0.8818359375,
0.435791015625,
0.8818359375,
0.449462890625,
0.515625,
0.449462890625
],
[
0.139404296875,
0.4453125,
0.48779296875,
0.4453125,
0.48779296875,
0.458984375,
0.139404296875,
0.458984375
],
[
0.5146484375,
0.45068359375,
0.8818359375,
0.451904296875,
0.8818359375,
0.466064453125,
0.5146484375,
0.465576171875
],
[
0.12005615234375,
0.46142578125,
0.185546875,
0.46142578125,
0.185546875,
0.4755859375,
0.12005615234375,
0.4755859375
],
[
0.515625,
0.468505859375,
0.88330078125,
0.468505859375,
0.88330078125,
0.482177734375,
0.515625,
0.482177734375
],
[
0.5146484375,
0.485107421875,
0.64404296875,
0.485107421875,
0.64404296875,
0.498779296875,
0.5146484375,
0.498779296875
],
[
0.12176513671875,
0.495361328125,
0.343505859375,
0.495361328125,
0.343505859375,
0.5087890625,
0.12176513671875,
0.5087890625
],
[
0.465087890625,
0.50439453125,
0.48779296875,
0.50439453125,
0.48779296875,
0.51416015625,
0.465087890625,
0.51416015625
],
[
0.5341796875,
0.501953125,
0.88330078125,
0.501953125,
0.88330078125,
0.515625,
0.5341796875,
0.515625
],
[
0.2376708984375,
0.51318359375,
0.447509765625,
0.51318359375,
0.447509765625,
0.52783203125,
0.2376708984375,
0.52783203125
],
[
0.5146484375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.5146484375,
0.53076171875
],
[
0.515625,
0.533203125,
0.8818359375,
0.53369140625,
0.8818359375,
0.54736328125,
0.515625,
0.54638671875
],
[
0.12005615234375,
0.54443359375,
0.48779296875,
0.54345703125,
0.48779296875,
0.55810546875,
0.12005615234375,
0.55859375
],
[
0.515625,
0.5498046875,
0.88330078125,
0.5498046875,
0.88330078125,
0.5634765625,
0.515625,
0.5634765625
],
[
0.12005615234375,
0.56103515625,
0.405517578125,
0.56103515625,
0.405517578125,
0.57470703125,
0.12005615234375,
0.57470703125
],
[
0.51708984375,
0.5654296875,
0.880859375,
0.5654296875,
0.880859375,
0.5791015625,
0.51708984375,
0.5791015625
],
[
0.5146484375,
0.58056640625,
0.8818359375,
0.580078125,
0.8818359375,
0.59423828125,
0.5146484375,
0.595703125
],
[
0.183837890625,
0.59033203125,
0.489501953125,
0.59033203125,
0.489501953125,
0.60498046875,
0.183837890625,
0.60498046875
],
[
0.5146484375,
0.595703125,
0.88330078125,
0.5966796875,
0.88330078125,
0.61083984375,
0.5146484375,
0.60986328125
],
[
0.5146484375,
0.6123046875,
0.880859375,
0.6123046875,
0.880859375,
0.6259765625,
0.5146484375,
0.6259765625
],
[
0.133544921875,
0.62060546875,
0.48779296875,
0.619140625,
0.48779296875,
0.63427734375,
0.133544921875,
0.63525390625
],
[
0.1275634765625,
0.62744140625,
0.1368408203125,
0.62744140625,
0.1368408203125,
0.63134765625,
0.1275634765625,
0.63134765625
],
[
0.515625,
0.6298828125,
0.88330078125,
0.6298828125,
0.88330078125,
0.6435546875,
0.515625,
0.6435546875
],
[
0.12005615234375,
0.63720703125,
0.48779296875,
0.63720703125,
0.48779296875,
0.65087890625,
0.12005615234375,
0.65087890625
],
[
0.12005615234375,
0.65283203125,
0.489501953125,
0.65283203125,
0.489501953125,
0.66650390625,
0.12005615234375,
0.66650390625
],
[
0.5146484375,
0.64794921875,
0.599609375,
0.64794921875,
0.599609375,
0.6591796875,
0.5146484375,
0.6591796875
],
[
0.5341796875,
0.662109375,
0.8818359375,
0.662109375,
0.8818359375,
0.6767578125,
0.5341796875,
0.6767578125
],
[
0.12176513671875,
0.66943359375,
0.48779296875,
0.66943359375,
0.48779296875,
0.68310546875,
0.12176513671875,
0.68310546875
],
[
0.51318359375,
0.677734375,
0.880859375,
0.677734375,
0.880859375,
0.69140625,
0.51318359375,
0.69140625
],
[
0.118408203125,
0.68408203125,
0.489501953125,
0.68603515625,
0.489501953125,
0.69970703125,
0.118408203125,
0.69775390625
],
[
0.515625,
0.69482421875,
0.88330078125,
0.69482421875,
0.88330078125,
0.7080078125,
0.515625,
0.7080078125
],
[
0.1275634765625,
0.70703125,
0.1368408203125,
0.70703125,
0.1368408203125,
0.71142578125,
0.1275634765625,
0.71142578125
],
[
0.1351318359375,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.1351318359375,
0.71533203125
],
[
0.515625,
0.71142578125,
0.88330078125,
0.71142578125,
0.88330078125,
0.724609375,
0.515625,
0.724609375
],
[
0.12176513671875,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.12176513671875,
0.732421875
],
[
0.51708984375,
0.72802734375,
0.62646484375,
0.72802734375,
0.62646484375,
0.7392578125,
0.51708984375,
0.7392578125
],
[
0.12005615234375,
0.7353515625,
0.162841796875,
0.7353515625,
0.162841796875,
0.74658203125,
0.12005615234375,
0.74658203125
],
[
0.5322265625,
0.74365234375,
0.88330078125,
0.74365234375,
0.88330078125,
0.75732421875,
0.5322265625,
0.75732421875
],
[
0.1376953125,
0.751953125,
0.486083984375,
0.751953125,
0.486083984375,
0.7626953125,
0.1376953125,
0.7626953125
],
[
0.51318359375,
0.75830078125,
0.880859375,
0.75830078125,
0.880859375,
0.771484375,
0.51318359375,
0.771484375
],
[
0.12005615234375,
0.7666015625,
0.48779296875,
0.765625,
0.48779296875,
0.77978515625,
0.12005615234375,
0.78076171875
],
[
0.5146484375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.5146484375,
0.7890625
],
[
0.12005615234375,
0.783203125,
0.484375,
0.783203125,
0.484375,
0.79736328125,
0.12005615234375,
0.79736328125
],
[
0.12005615234375,
0.798828125,
0.484375,
0.798828125,
0.484375,
0.8125,
0.12005615234375,
0.8125
],
[
0.515625,
0.79345703125,
0.8818359375,
0.79345703125,
0.8818359375,
0.80419921875,
0.515625,
0.80419921875
],
[
0.12005615234375,
0.81396484375,
0.486083984375,
0.81396484375,
0.486083984375,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.5146484375,
0.80859375,
0.57080078125,
0.80859375,
0.57080078125,
0.82080078125,
0.5146484375,
0.82080078125
],
[
0.118408203125,
0.830078125,
0.328369140625,
0.83056640625,
0.328369140625,
0.845703125,
0.118408203125,
0.84423828125
],
[
0.515625,
0.84033203125,
0.54833984375,
0.84033203125,
0.54833984375,
0.84716796875,
0.515625,
0.84716796875
],
[
0.5498046875,
0.83740234375,
0.76904296875,
0.83740234375,
0.76904296875,
0.85107421875,
0.5498046875,
0.85107421875
],
[
0.139404296875,
0.84716796875,
0.48779296875,
0.84716796875,
0.48779296875,
0.86083984375,
0.139404296875,
0.86083984375
],
[
0.12005615234375,
0.8623046875,
0.489501953125,
0.86328125,
0.489501953125,
0.87744140625,
0.12005615234375,
0.876953125
],
[
0.5146484375,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.5146484375,
0.876953125
],
[
0.12176513671875,
0.8798828125,
0.489501953125,
0.8798828125,
0.489501953125,
0.8935546875,
0.12176513671875,
0.8935546875
],
[
0.515625,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.515625,
0.8935546875
],
[
0.118408203125,
0.8955078125,
0.297119140625,
0.8955078125,
0.297119140625,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"3Soft Constraints in Dual",
"Algorithm 2 Soft-DD: projected subgradient",
"Decomposition",
"for dual decomposition with soft constraints",
"l: while has not converged do",
"我们现在介绍算法E1的扩展以处理软约束。",
"2:",
"y(t) = argmaxyeu (w + AT>,y)",
"",
"在我们的公式中，软约束模型对每个未满足的",
"3.",
"(0) = Io≤.≤e[(t-1) - n((Ay - b)",
"约束施加惩罚，与违反的数量成比例。因此",
"我们的推导类似于如何通过引人辅助松弛变量",
"(CortesandVapnik，1995）从硬裕度SVM中",
"3.1",
"Learning Penalties",
"导出软裕度SVM。请注意，在执行受软约束",
"的MAP时，最优解可能不满足某些约束，因",
"使用softv时的一个注意事项。S。硬约束是软",
"为这样做会使模型的分数降低太多",
"约束提出了一个新的训练问题，因为我们需要",
"考虑以下形式的优化问题：",
"选择向量℃，违反约束的惩罚。问题E（5）在",
"上一节中的一个重要属性是，它对应于9和z",
"mac. (w,y) - (c,z)",
"上的结构化线性模型。因此，我们可以应用已",
"s.t. yeu",
"知的训练算法来估计结构化线性模型的参数，",
"Ay-b≤z",
"(5)",
"以选择c",
"-2≤0,",
"我们所需要采用的结构化感知器算法",
"（Collins，2002）或结构化SVM算法（Tsochan-",
"对于正ci，很明显，最佳zi将等于违反",
"taridisetal.，2004）是一个黑盒过程，用于在",
"ay≤bi的程度。因此，我们支付的成本是违",
"反第i个约束的程度的C倍，这反映了松弛",
"给定任意成本向量的情况下在结构化线性模型",
"中执行MAP推理。幸运的是,E(5的MAP",
"变量如何用于表示SVM的铰链损失。请注意，",
"问题可以在算法E2中使用Soft-DD解决。每",
"Ci必须为正，否则此线性程序是无界的，并且",
"个惩罚Ci必须是非负的；否则，方程（5）中的",
"可以通过将设置为无穷大来获得最佳值",
"优化问题是不明确的。这可以通过对感知器的",
"使用与2.2节中类似的构造，我们将拉格朗",
"简单修改和结构化SVM目标的次梯度下降优",
"日写为：",
"化来确保，只需在每次学习送代中将℃按坐标",
"截断为非负即可",
"L(y, z, 入, μ) = (w,y) - (c,z)",
"(6)",
"从直觉上讲，感知器更新会在满足基本事实",
"+T(Ay-b-2)+μT(-2).",
"而不是推断的预测时增加约束的惩罚，而在满",
"足预测而不是基本事实时减少约束的惩罚。由",
"关于z的最优性约束告诉我们-c--μ=0",
"于我们将惩罚截断为0，这表明我们将学习三",
"，因此μ=一c一入。代入，我们有",
"类约束的惩罚为0：不符合基本真理的约束",
"符合基本真理但在实践中通过在基本CRF模",
"L(y, ) = (w,y) +^T(Ay -b),(7)",
"型中执行推理而得到满足的约束，以及在实践",
"中得到满足的约束，这是对其他一些约束施加",
"除了μ=c-入暗示μ为正的入≤c的",
"",
"非零惩罚的副作用。类似的分析适用于结构化",
"约束。由于此拉格朗日具有与方程（3）相同的",
"形式，因此我们还可以推导出一个对偶问题，",
"SVM方法",
"因此，我们不仅可以将学习惩罚的值视为参",
"该对偶问题与方程（4）中的对偶问题相同，并",
"数调整，还可以将其视为执行“约束选择”的",
"带有附加约束，即每个入，不能大于其成本ci",
"一种手段，因为可以忽略惩罚为0的约束。此",
"",
"换句话说，对偶问题不能惩罚违反约束的行",
"属性允许我们考虑大型约束族，从中自动识别",
"为，而原始中的软约束模型在违反约束时会惩",
"有用的约束族",
"罚您。",
"我们发现，尽管在理论上没有必要，但与其",
"这个优化问题仍然可以用投影次梯度下降来",
"他模型参数分开学习对保持的开发集的约束是",
"解决，并在算法2中进行了描述。对算法E1",
"有益的，因为在训练过程中，由于过度拟合",
"的唯一修改是用IIo<:≤c替换坐标投影IIo<：",
"以及我们如何检查收敛。现在，我们检查（5）",
"大多数约束都得到了满足，这导致低估了相关",
"的KKT条件，其中对于每个约束i，约束满",
"的惩罚",
"足相等，入i=0或入i=Ci",
"4",
"Citation Extraction Data",
"因此，实现软约束对偶分解与实现硬约束对",
"偶分解一样容易，并且每次迭代的复杂度相",
"我们考虑在Anzaroot和McCallum（2013）中",
"同。我们鼓励将软约束对偶分解进一步应用于",
"首次引人的UMass引文数据集。它有来自许",
"现有和新的NLP问题",
"多学术领域的1800多个引文，摘自arXiv。此"
] |
{
"arxiv_id": "1403.1349",
"language": "zh",
"page_idx": 2
}
|
||
1403.1349-zh-3
|
1403.1349-zh-3.jpg
|
数据集包含粗粒度和细粒度标签；例如，它包含所有作者的部分、每个作者的部分以及每个作者的名字和姓氏的标签。开发集中有 660 个引文，测试集中有 367 个引文。UMass 数据集中的标签是来自分层定义架构的标签的串联。例如，作者的名字被标记为：作者/人/第一个。此外，使用层次结构中每个级别的 BIO 标签模式对单个令牌进行标记。BIO 是信息提取任务的常用标记模式。BIO 标签允许标记上的单个标签标记分段信息以及分段的标签。在此架构中，开始段的标签前面加上 B, 延续段的标签前面加上 I, 在此架构中没有标签的令牌前面加上 O 标签。例如，在分层 BIO 标签模式中，第二个作者的名字中的第一个令牌可以标记为：I-authors/B-person/B-first。可以在图 1中查看此数据集中标记为引用的示例。
5 Global Constraints for Citation
## $\operatorname{Extraction}$
5.1 Constraint Templates
我们现在描述我们考虑用于引文提取的全局约束族。请注意，这些约束都是线性的，因为它们仅取决于每个可能的条件随机场标签的计数。此外，由于我们的标签是生物编码的，因此可以通过计算 B 标签来计算每个引文标签本身在句子中出现的频率。我们描述的前两个约束系列对于任何序列标记任务都是通用的， 而最后一个特定于分层标记，例如 UMass 数据集中可用的。我们的序列输出表示为$y$ ,此序列的元素为$y_k$ 。如果$y_k$在索引$i$处具有1 , 否 则 0将$[[y_k=i]]$表示为输出 1 的函数。在这里，$y_k$表示 CRF 的输出标记，因此如果$[[y_k=i]]=1$,则我们认为$y_k$被赋予了索引为$i$的标签。
## 5.2 Singleton Constraints
单例约束确保每个标签在引用中最多出现一次。这些是与 Chang 等人在引文场提取中使用的相同的全局约束。(2012)。我们将$s(i)$ 定义为在引用中预测具有索引$i$的标签的次数，形式为：
$$s(i)=\sum_{y_k\in y}\left[[y_k=i]\right]$$
每个标签最多只能出现一次的约束采用以下形
式：。我们将$z_1(i,j)$定义为
$$z_1(i,j)=\sum_{y_k\in y}\left[[y_k=i]\right]+\sum_{y_k\in y}\left[[y_k=j]\right]$$
,将$z_2(i,j)$定义为
$$\begin{aligned}z_2(i,j)=\sum_{y_k\in y}\left[[y_k=i]\right]-\sum_{y_k\in y}\left[[y_k=j]\right]\end{aligned}$$
我们考虑以下形式的所有约束：请注意，这些约束中的一些对是冗余的或逻辑上不兼容的。但是，我们将它们用作软约束，因此模型的输出不一定会满足这些约束，从而消除了对执行逻辑上不可能的输出的担忧。此外，在 3.1节中，我们描述了学习惩罚的过程如何将一些惩罚驱动到0，这有效地将它们从我们考虑的约束集中删除。例如，可以证明，我们永远不会使用$\overline{\mathrm{E}}3.1$ 节中描述的感知器式算法来学习某些逻辑上不兼容的约束对的非零惩罚。
5.3 Hierarchical Equality Constraints
引文数据集中的标签是分层标签。这意味着标签是层次结构中所有级别的串联。我们可以创建仅依赖于层次结构中的一个或几个元素的约束。我们将$C(x,i)$定义为返回 1 的函数，如果输出$x$在层次结构中包含标签$i$ ,否则为0。我们将$e(i,j)$定义为
$$e(i,j)=\sum_{y_k\in y}\left[[C(y_k,i)]\right]-\sum_{y_k\in y}\left[[C(y_k,j)]\right]$$
分层相等约束采用以下形式：
$$\begin{matrix}e(i,j)&\ge&0\\e(i,j)&\le&0\end{matrix}$$
(8) (9)
5.4 Local constraints
我们将链结构 CRF 的输出标记限制为有效的BIO 编码。在没有全局约束的情况下使用时， 这既提高了基础模型的性能，又确保了我们施加的全局约束的有效性，因为它们仅在 B 标签上运行。标签是有效的 BIO 的约束可以表示为序列中相邻标签的成对约束的集合。当在CRF 中执行 MAP 推理时，可以通过修改维特比算法的动态程序以仅允许有效的相邻标签对，而不是使用双重分解来强制执行这些约束。。
# 5.5 Constraint Pruning
虽然 3.1 节中的技术可以很容易地在训练时处理大量约束，但这在计算上可能很昂贵，特别是在考虑非常大的约束族的情况下。这是有问题的，因为我们考虑的一些约束族的大小随候选标签的数量呈二次方增长，在 UMass 数据集中大约有 100 个。这样的家庭由两个不同的标签类型的计数的总和必须是有界的约束(一
|
[
[
0.12005615234375,
0.07598876953125,
0.48779296875,
0.07598876953125,
0.48779296875,
0.08905029296875,
0.12005615234375,
0.08905029296875
],
[
0.12176513671875,
0.09320068359375,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.12176513671875,
0.10687255859375
],
[
0.12005615234375,
0.1092529296875,
0.486083984375,
0.10809326171875,
0.486083984375,
0.122314453125,
0.12005615234375,
0.12353515625
],
[
0.54248046875,
0.10394287109375,
0.85546875,
0.10394287109375,
0.85546875,
0.1217041015625,
0.54248046875,
0.1217041015625
],
[
0.62109375,
0.12353515625,
0.6572265625,
0.12353515625,
0.6572265625,
0.135986328125,
0.62109375,
0.135986328125
],
[
0.74462890625,
0.12469482421875,
0.78076171875,
0.12469482421875,
0.78076171875,
0.135986328125,
0.74462890625,
0.135986328125
],
[
0.12176513671875,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.12176513671875,
0.155029296875
],
[
0.12005615234375,
0.1258544921875,
0.48779296875,
0.1258544921875,
0.48779296875,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.515625,
0.1484375,
0.8818359375,
0.1484375,
0.8818359375,
0.162109375,
0.515625,
0.162109375
],
[
0.12005615234375,
0.157958984375,
0.489501953125,
0.157958984375,
0.489501953125,
0.171630859375,
0.12005615234375,
0.171630859375
],
[
0.515625,
0.1644287109375,
0.88330078125,
0.1644287109375,
0.88330078125,
0.1776123046875,
0.515625,
0.1776123046875
],
[
0.118408203125,
0.1728515625,
0.48779296875,
0.1728515625,
0.48779296875,
0.1859130859375,
0.118408203125,
0.1859130859375
],
[
0.515625,
0.18115234375,
0.8818359375,
0.18115234375,
0.8818359375,
0.19482421875,
0.515625,
0.19482421875
],
[
0.12005615234375,
0.189453125,
0.489501953125,
0.189453125,
0.489501953125,
0.203125,
0.12005615234375,
0.203125
],
[
0.515625,
0.196533203125,
0.8818359375,
0.196533203125,
0.8818359375,
0.210205078125,
0.515625,
0.210205078125
],
[
0.12005615234375,
0.2060546875,
0.48779296875,
0.2060546875,
0.48779296875,
0.2197265625,
0.12005615234375,
0.2197265625
],
[
0.515625,
0.2120361328125,
0.8818359375,
0.2120361328125,
0.8818359375,
0.2257080078125,
0.515625,
0.2257080078125
],
[
0.118408203125,
0.221435546875,
0.48779296875,
0.221435546875,
0.48779296875,
0.235107421875,
0.118408203125,
0.235107421875
],
[
0.515625,
0.2286376953125,
0.8818359375,
0.2286376953125,
0.8818359375,
0.2423095703125,
0.515625,
0.2423095703125
],
[
0.12005615234375,
0.2381591796875,
0.48779296875,
0.2381591796875,
0.48779296875,
0.251708984375,
0.12005615234375,
0.251708984375
],
[
0.515625,
0.2452392578125,
0.8818359375,
0.2452392578125,
0.8818359375,
0.2587890625,
0.515625,
0.2587890625
],
[
0.12005615234375,
0.254150390625,
0.48779296875,
0.254150390625,
0.48779296875,
0.267333984375,
0.12005615234375,
0.267333984375
],
[
0.5146484375,
0.26123046875,
0.87890625,
0.26123046875,
0.87890625,
0.27490234375,
0.5146484375,
0.27490234375
],
[
0.12005615234375,
0.26953125,
0.48779296875,
0.268310546875,
0.48779296875,
0.283203125,
0.12005615234375,
0.283935546875
],
[
0.515625,
0.276611328125,
0.8818359375,
0.276611328125,
0.8818359375,
0.290283203125,
0.515625,
0.290283203125
],
[
0.12005615234375,
0.28515625,
0.486083984375,
0.28515625,
0.486083984375,
0.298583984375,
0.12005615234375,
0.298583984375
],
[
0.5146484375,
0.29345703125,
0.81787109375,
0.29345703125,
0.81787109375,
0.30712890625,
0.5146484375,
0.30712890625
],
[
0.118408203125,
0.3017578125,
0.489501953125,
0.300537109375,
0.489501953125,
0.3154296875,
0.118408203125,
0.31640625
],
[
0.12005615234375,
0.318359375,
0.42236328125,
0.318359375,
0.42236328125,
0.33203125,
0.12005615234375,
0.33203125
],
[
0.5146484375,
0.319580078125,
0.8330078125,
0.319580078125,
0.8330078125,
0.333251953125,
0.5146484375,
0.333251953125
],
[
0.1175537109375,
0.34326171875,
0.42822265625,
0.344482421875,
0.42822265625,
0.359375,
0.1175537109375,
0.358154296875
],
[
0.515625,
0.34033203125,
0.8818359375,
0.34033203125,
0.8818359375,
0.35400390625,
0.515625,
0.35400390625
],
[
0.145263671875,
0.35986328125,
0.2435302734375,
0.3623046875,
0.2435302734375,
0.377197265625,
0.145263671875,
0.374755859375
],
[
0.5146484375,
0.355712890625,
0.8818359375,
0.355712890625,
0.8818359375,
0.369384765625,
0.5146484375,
0.369384765625
],
[
0.5146484375,
0.371826171875,
0.880859375,
0.371826171875,
0.880859375,
0.385498046875,
0.5146484375,
0.385498046875
],
[
0.12005615234375,
0.38720703125,
0.1695556640625,
0.38720703125,
0.1695556640625,
0.3984375,
0.12005615234375,
0.3984375
],
[
0.160400390625,
0.385498046875,
0.341796875,
0.385986328125,
0.341796875,
0.40087890625,
0.160400390625,
0.399658203125
],
[
0.515625,
0.388427734375,
0.8818359375,
0.388427734375,
0.8818359375,
0.402099609375,
0.515625,
0.402099609375
],
[
0.12005615234375,
0.40673828125,
0.489501953125,
0.40673828125,
0.489501953125,
0.42041015625,
0.12005615234375,
0.42041015625
],
[
0.51318359375,
0.402587890625,
0.8818359375,
0.40380859375,
0.8818359375,
0.418701171875,
0.51318359375,
0.41748046875
],
[
0.118408203125,
0.42333984375,
0.486083984375,
0.42333984375,
0.486083984375,
0.43701171875,
0.118408203125,
0.43701171875
],
[
0.51123046875,
0.419189453125,
0.7119140625,
0.42041015625,
0.7119140625,
0.435302734375,
0.51123046875,
0.43408203125
],
[
0.12176513671875,
0.441162109375,
0.483642578125,
0.441162109375,
0.483642578125,
0.451904296875,
0.12176513671875,
0.451904296875
],
[
0.118408203125,
0.4560546875,
0.486083984375,
0.4560546875,
0.486083984375,
0.4697265625,
0.118408203125,
0.4697265625
],
[
0.12005615234375,
0.471435546875,
0.48779296875,
0.471435546875,
0.48779296875,
0.485107421875,
0.12005615234375,
0.485107421875
],
[
0.53564453125,
0.466064453125,
0.861328125,
0.466064453125,
0.861328125,
0.479736328125,
0.53564453125,
0.479736328125
],
[
0.603515625,
0.482177734375,
0.6396484375,
0.482177734375,
0.6396484375,
0.49755859375,
0.603515625,
0.49755859375
],
[
0.7421875,
0.48388671875,
0.77685546875,
0.48388671875,
0.77685546875,
0.495361328125,
0.7421875,
0.495361328125
],
[
0.12005615234375,
0.50439453125,
0.484375,
0.50439453125,
0.484375,
0.51708984375,
0.12005615234375,
0.51708984375
],
[
0.12005615234375,
0.48681640625,
0.48779296875,
0.48681640625,
0.48779296875,
0.50048828125,
0.12005615234375,
0.50048828125
],
[
0.515625,
0.51025390625,
0.73876953125,
0.51025390625,
0.73876953125,
0.52099609375,
0.515625,
0.52099609375
],
[
0.12176513671875,
0.51953125,
0.48779296875,
0.51953125,
0.48779296875,
0.533203125,
0.12176513671875,
0.533203125
],
[
0.118408203125,
0.53369140625,
0.489501953125,
0.53515625,
0.489501953125,
0.5498046875,
0.118408203125,
0.548828125
],
[
0.646484375,
0.53759765625,
0.75244140625,
0.53759765625,
0.75244140625,
0.5517578125,
0.646484375,
0.5517578125
],
[
0.8564453125,
0.5361328125,
0.884765625,
0.5361328125,
0.884765625,
0.55419921875,
0.8564453125,
0.55419921875
],
[
0.12176513671875,
0.5517578125,
0.486083984375,
0.5517578125,
0.486083984375,
0.56640625,
0.12176513671875,
0.56640625
],
[
0.6455078125,
0.55712890625,
0.75390625,
0.5556640625,
0.75390625,
0.57177734375,
0.6455078125,
0.572265625
],
[
0.8564453125,
0.5556640625,
0.88330078125,
0.5556640625,
0.88330078125,
0.57373046875,
0.8564453125,
0.57373046875
],
[
0.12005615234375,
0.568359375,
0.48779296875,
0.568359375,
0.48779296875,
0.58203125,
0.12005615234375,
0.58203125
],
[
0.12005615234375,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.5986328125,
0.12005615234375,
0.5986328125
],
[
0.5146484375,
0.58837890625,
0.7001953125,
0.58837890625,
0.7001953125,
0.5986328125,
0.5146484375,
0.5986328125
],
[
0.12005615234375,
0.60107421875,
0.48779296875,
0.60107421875,
0.48779296875,
0.615234375,
0.12005615234375,
0.615234375
],
[
0.5146484375,
0.60791015625,
0.880859375,
0.60791015625,
0.880859375,
0.62158203125,
0.5146484375,
0.62158203125
],
[
0.1175537109375,
0.615234375,
0.1956787109375,
0.61767578125,
0.1956787109375,
0.6318359375,
0.1175537109375,
0.6298828125
],
[
0.5146484375,
0.62353515625,
0.8818359375,
0.62353515625,
0.8818359375,
0.63720703125,
0.5146484375,
0.63720703125
],
[
0.12005615234375,
0.642578125,
0.340087890625,
0.642578125,
0.340087890625,
0.65625,
0.12005615234375,
0.65625
],
[
0.51318359375,
0.63818359375,
0.88330078125,
0.6396484375,
0.88330078125,
0.65380859375,
0.51318359375,
0.65283203125
],
[
0.5146484375,
0.65478515625,
0.880859375,
0.65478515625,
0.880859375,
0.66845703125,
0.5146484375,
0.66845703125
],
[
0.12005615234375,
0.6611328125,
0.470947265625,
0.6611328125,
0.470947265625,
0.6748046875,
0.12005615234375,
0.6748046875
],
[
0.468505859375,
0.66650390625,
0.474365234375,
0.66650390625,
0.474365234375,
0.67138671875,
0.468505859375,
0.67138671875
],
[
0.474365234375,
0.66650390625,
0.480224609375,
0.66650390625,
0.480224609375,
0.67138671875,
0.474365234375,
0.67138671875
],
[
0.51318359375,
0.67041015625,
0.8818359375,
0.67041015625,
0.8818359375,
0.68408203125,
0.51318359375,
0.68408203125
],
[
0.12005615234375,
0.6787109375,
0.48779296875,
0.6787109375,
0.48779296875,
0.6923828125,
0.12005615234375,
0.6923828125
],
[
0.5146484375,
0.68603515625,
0.8818359375,
0.68701171875,
0.8818359375,
0.70166015625,
0.5146484375,
0.70068359375
],
[
0.12176513671875,
0.69482421875,
0.32568359375,
0.69482421875,
0.32568359375,
0.7080078125,
0.12176513671875,
0.7080078125
],
[
0.316650390625,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.316650390625,
0.7080078125
],
[
0.51318359375,
0.70166015625,
0.8818359375,
0.703125,
0.8818359375,
0.71728515625,
0.51318359375,
0.71630859375
],
[
0.12005615234375,
0.71142578125,
0.48779296875,
0.71142578125,
0.48779296875,
0.724609375,
0.12005615234375,
0.724609375
],
[
0.515625,
0.7197265625,
0.8818359375,
0.7197265625,
0.8818359375,
0.73291015625,
0.515625,
0.73291015625
],
[
0.12005615234375,
0.72705078125,
0.2166748046875,
0.72705078125,
0.2166748046875,
0.73828125,
0.12005615234375,
0.73828125
],
[
0.5146484375,
0.73388671875,
0.8818359375,
0.7353515625,
0.8818359375,
0.75,
0.5146484375,
0.7490234375
],
[
0.2216796875,
0.7548828125,
0.3837890625,
0.75732421875,
0.3837890625,
0.775390625,
0.2216796875,
0.77392578125
],
[
0.5146484375,
0.751953125,
0.55908203125,
0.751953125,
0.55908203125,
0.7666015625,
0.5146484375,
0.7666015625
],
[
0.27783203125,
0.77685546875,
0.313232421875,
0.77685546875,
0.313232421875,
0.7890625,
0.27783203125,
0.7890625
],
[
0.51123046875,
0.77685546875,
0.72119140625,
0.779296875,
0.72119140625,
0.79345703125,
0.51123046875,
0.79150390625
],
[
0.12176513671875,
0.7958984375,
0.489501953125,
0.7958984375,
0.489501953125,
0.80859375,
0.12176513671875,
0.80859375
],
[
0.497802734375,
0.8017578125,
0.5087890625,
0.8017578125,
0.5087890625,
0.8056640625,
0.497802734375,
0.8056640625
],
[
0.5146484375,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.5146484375,
0.8125
],
[
0.12176513671875,
0.8125,
0.341796875,
0.8125,
0.341796875,
0.826171875,
0.12176513671875,
0.826171875
],
[
0.515625,
0.81640625,
0.87890625,
0.81640625,
0.87890625,
0.82666015625,
0.515625,
0.82666015625
],
[
0.51708984375,
0.83203125,
0.8818359375,
0.83203125,
0.8818359375,
0.845703125,
0.51708984375,
0.845703125
],
[
0.14697265625,
0.8525390625,
0.459228515625,
0.8525390625,
0.459228515625,
0.87060546875,
0.14697265625,
0.87060546875
],
[
0.515625,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.515625,
0.86083984375
],
[
0.515625,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.515625,
0.876953125
],
[
0.2275390625,
0.87158203125,
0.26123046875,
0.87158203125,
0.26123046875,
0.88427734375,
0.2275390625,
0.88427734375
],
[
0.3525390625,
0.87158203125,
0.38623046875,
0.87158203125,
0.38623046875,
0.88427734375,
0.3525390625,
0.88427734375
],
[
0.5146484375,
0.87890625,
0.880859375,
0.87890625,
0.880859375,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.27783203125,
0.8955078125,
0.27783203125,
0.91015625,
0.12005615234375,
0.91015625
],
[
0.5146484375,
0.8955078125,
0.880859375,
0.8955078125,
0.880859375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"数据集包含粗粒度和细粒度标签；例如，它包",
"含所有作者的部分、每个作者的部分以及每个",
"作者的名字和姓氏的标签。开发集中有660个",
"22(i, 5] = Z [yk = ] - Z [yk =]",
"YhEy",
"ykEy",
"中的标签是来自分层定义架构的标签的串联。",
"引文，测试集中有367个引文。UMass数据集",
"我们考虑以下形式的所有约束：请注意，这些",
"例如，作者的名字被标记为：作者/人/第一个。",
"约束中的一些对是穴余的或逻辑上不兼容的。",
"此外，使用层次结构中每个级别的BIO标签",
"但是，我们将它们用作软约束，因此模型的输",
"模式对单个令牌进行标记。BIO是信息提取任",
"出不一定会满足这些约束，从而消除了对执行",
"务的常用标记模式。BIO标签允许标记上的单",
"逻辑上不可能的输出的担忧。此外，在3.1节",
"个标签标记分段信息以及分段的标签。在此架",
"中，我们描述了学习惩罚的过程如何将一些惩",
"构中，开始段的标签前面加上B，延续段的标",
"罚驱动到0，这有效地将它们从我们考虑的约",
"签前面加上I，在此架构中没有标签的令牌前",
"束集中删除。例如，可以证明，我们永远不会",
"面加上O标签。例如，在分层BIO标签模式",
"使用E3.1节中描述的感知器式算法来学习某",
"中，第二个作者的名字中的第一个令牌可以标",
"些逻辑上不兼容的约束对的非零惩罚。",
"记为：I-authors/B-person/B-first。可以在图1",
"中查看此数据集中标记为引用的示例。",
"5.3Hierarchical Equality Constraints",
"5Global Constraints for Citation",
"引文数据集中的标签是分层标签。这意味着标",
"Extraction",
"签是层次结构中所有级别的串联。我们可以创",
"建仅依赖于层次结构中的一个或几个元素的约",
"5.1",
"Constraint Templates",
"束。我们将C(c,i）定义为返回1的函数，如",
"我们现在描述我们考虑用于引文提取的全局",
"果输出在层次结构中包含标签i，否则为",
"约束族。请注意，这些约束都是线性的，因为",
"0。我们将e(i,j)定义为",
"它们仅取决于每个可能的条件随机场标签的计",
"数。此外，由于我们的标签是生物编码的，因",
"此可以通过计算B标签来计算每个引文标签",
"e(i,) = >[C(yk, i)]l - [C(yk,j)]",
"ykEy",
"YEy",
"约束系列对于任何序列标记任务都是通用的",
"本身在句子中出现的频率。我们描述的前两个",
"分层相等约束采用以下形式",
"而最后一个特定于分层标记，例如UMass数",
"据集中可用的。我们的序列输出表示为9，此",
"e(i,j) ≥ 0",
"(8)",
"序列的元素为yk。如果yh在索引i处具有",
"e(i,j) ≤ 0",
"(9)",
"1，否则0将[yk=i]]表示为输出1的函数。",
"在这里，Yk表示CRF的输出标记，因此如果",
"5.4Localconstraints",
"[yk=]=1，则我们认为yk被赋予了索引为",
"我们将链结构CRF的输出标记限制为有效的",
"i的标签。",
"BIO编码。在没有全局约束的情况下使用时",
"5.2Singleton Constraints",
"这既提高了基础模型的性能，又确保了我们施",
"加的全局约束的有效性，因为它们仅在B标",
"单例约束确保每个标签在引用中最多出现",
"-",
"",
"签上运行。标签是有效的BIO的约束可以表",
"次。这些是与Chang等人在引文场提取中使",
"示为序列中相邻标签的成对约束的集合。当在",
"用的相同的全局约束。（",
"(2012）。我们将s(i)",
"CRF中执行MAP推理时，可以通过修改维",
"定义为在引用中预测具有索引的标签的次",
"特比算法的动态程序以仅允许有效的相邻标",
"数，形式为",
"签对，而不是使用双重分解来强制执行这些约",
"s(i) = [yk: = i]",
"束。。",
"ykEy",
"5.5Constraint Pruning",
"每个标签最多只能出现一次的约束采用以下形",
"",
"虽然3.1节中的技术可以很容易地在训练时处",
"式：。我们将z1(i,j）定义为",
"理大量约束，但这在计算上可能很昂贵，特别",
"是在考虑非常大的约束族的情况下。这是有问",
"zi(i, 3] = Z [yk = i] + [yk = il]",
"题的，因为我们考虑的一些约束族的大小随候",
"选标签的数量呈二次方增长，在UMass数据",
"ykEy",
"ykEy",
"集中大约有100个。这样的家庭由两个不同的",
"，将22(i,j）定义为",
"标签类型的计数的总和必须是有界的约束（一"
] |
{
"arxiv_id": "1403.1349",
"language": "zh",
"page_idx": 3
}
|
||
1810.12443-zh-0
|
1810.12443-zh-0.jpg
|
学习更好的单词内部结构以进行序列标记
Yingwei Xin, Ethan Hart, Vibhuti Mahajan, Jean-David Ruvini
eBay Research
2025 Hamilton Ave, San Jose, CA 95125, USA
$\{$ yixin,ejhart,vibmahajan,jean-david.ruvini $\}$@ebay.com
# Abstract
Character-based neural models have recently proven very useful for many NLP tasks. However, there is a gap of sophistication between methods for learning representations of sentences and words. While most character models for learning representations of sentences are deep and complex, models for learning representations of words are shallow and simple. Also, in spite of considerable research on learning character embeddings, it is still not clear which kind of architecture is the best for capturing character-to-word representations. To address these questions, we first investigate the gaps between methods for learning word and sentence representations. We conduct detailed experiments and comparisons of different state-of-theart convolutional models, and also investigate the advantages and disadvantages of their constituents. Furthermore, we propose IntNet, a funnel-shaped wide convolutional neural architecture with no downsampling for learning representations of the internal structure of words by composing their characters from limited, supervised training corpora. We evaluate our proposed model on six sequence labeling datasets, including named entity recognition, part-of-speech tagging, and syntactic chunking. Our in-depth analysis shows that IntNet significantly outperforms other character embedding models and obtains new state-of-the-art performance without relying on any external knowledge or resources.
# 1 Introduction
序列标记是为数据序列的每个元素分配标签或类的任务，并且是许多自然语言处理(NLP) 任务的第一阶段之一。例如，命名实体识别(NER)旨在将句子中的单词分类为几个预定
义的感兴趣类别，例如人员，组织，位置等。词性 (POS) 标记为输入句子中的每个单词分配词性。语法分块将文本分为语法相关的，不重叠的单词组。序列标记是一个具有挑战性的问题，因为人工注释非常昂贵，通常只有少量的标记数据可用
大多数传统的序列标记系统都被严重依赖特征工程的线性统计模型所主导。因此，精心构建的手工特征和特定领域的知识被广泛用于解决这些任务。不幸的是，开发特定领域的知识和手工制作的功能是昂贵的。最近，使用字符级信息的神经网络已成功用于最小化特征工程的需求。基于字符的建模基本上有两个线程，一个侧重于学习语义和语法(Zhang et al., 2015; Conneau et al., 2017) 的句子表示；另一个侧重于学习单词的表示，以消除单词形状信息 (Lample et al., 2016; Ma and Hovy, 2016) 的手工特征
从后面的线程中出现了两种学习序列标记的字符表示的主要最新方法。一种基于 RNN 并使用双向 LSTM 或 GRU 来学习正向和反向字符信息 (Ling et al., 2015; Lample et al., 2016; Yang et al.,2017) 。另一种方法是基于CNN, 每个单词周围都有一个固定大小的窗口来创建字符级表示 (Santos and Zadrozny, 2014; Chiu and Nichols, 2016; Ma and Hovy, 2016)。但是，与基于单词的方法相比，基于字符的学习句子表示方法之间的复杂程度存在差距。。我们发现，大多数最先进的基于字符的CNN 单词模型都使用卷积，然后使用最大池化作为浅层特征提取器，这与具有深人而复杂的句子体系结构的 CNN 模型有很大不同。尽管对学习字符嵌入进行了大量研究，但仍不清楚哪种体系结构最适合捕获字符到单词的表示形式
因此，许多问题仍然悬而未决：
$\bullet$ Why is there a gap between methods for
learning representations of sentences and
words? How can this gap be bridged?
|
[
[
0.2880859375,
0.08612060546875,
0.71533203125,
0.08612060546875,
0.71533203125,
0.1009521484375,
0.2880859375,
0.1009521484375
],
[
0.2301025390625,
0.13720703125,
0.78076171875,
0.13720703125,
0.78076171875,
0.1519775390625,
0.2301025390625,
0.1519775390625
],
[
0.43994140625,
0.15380859375,
0.57080078125,
0.15380859375,
0.57080078125,
0.168701171875,
0.43994140625,
0.168701171875
],
[
0.3046875,
0.17041015625,
0.70361328125,
0.17041015625,
0.70361328125,
0.18408203125,
0.3046875,
0.18408203125
],
[
0.26025390625,
0.18701171875,
0.748046875,
0.18701171875,
0.748046875,
0.20068359375,
0.26025390625,
0.20068359375
],
[
0.2626953125,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.28076171875,
0.2626953125,
0.279052734375
],
[
0.51904296875,
0.2666015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.27978515625,
0.51904296875,
0.27978515625
],
[
0.51708984375,
0.281982421875,
0.880859375,
0.281982421875,
0.880859375,
0.295654296875,
0.51708984375,
0.295654296875
],
[
0.1485595703125,
0.29443359375,
0.455078125,
0.29443359375,
0.455078125,
0.30810546875,
0.1485595703125,
0.30810546875
],
[
0.515625,
0.298583984375,
0.8818359375,
0.298583984375,
0.8818359375,
0.312255859375,
0.515625,
0.312255859375
],
[
0.14697265625,
0.30810546875,
0.45751953125,
0.30712890625,
0.45751953125,
0.32177734375,
0.14697265625,
0.322509765625
],
[
0.515625,
0.314208984375,
0.8818359375,
0.314208984375,
0.8818359375,
0.327880859375,
0.515625,
0.327880859375
],
[
0.145263671875,
0.32177734375,
0.45751953125,
0.322509765625,
0.45751953125,
0.33740234375,
0.145263671875,
0.336181640625
],
[
0.51708984375,
0.330810546875,
0.8818359375,
0.330810546875,
0.8818359375,
0.344482421875,
0.51708984375,
0.344482421875
],
[
0.1502685546875,
0.339111328125,
0.456787109375,
0.339111328125,
0.456787109375,
0.349853515625,
0.1502685546875,
0.349853515625
],
[
0.14697265625,
0.3515625,
0.459228515625,
0.3515625,
0.459228515625,
0.365234375,
0.14697265625,
0.365234375
],
[
0.515625,
0.3466796875,
0.64404296875,
0.3466796875,
0.64404296875,
0.361083984375,
0.515625,
0.361083984375
],
[
0.14697265625,
0.365234375,
0.45751953125,
0.366455078125,
0.45751953125,
0.380126953125,
0.14697265625,
0.37890625
],
[
0.5322265625,
0.363525390625,
0.880859375,
0.363525390625,
0.880859375,
0.377197265625,
0.5322265625,
0.377197265625
],
[
0.1485595703125,
0.380126953125,
0.45751953125,
0.380126953125,
0.45751953125,
0.393798828125,
0.1485595703125,
0.393798828125
],
[
0.515625,
0.380126953125,
0.8818359375,
0.380126953125,
0.8818359375,
0.393798828125,
0.515625,
0.393798828125
],
[
0.14697265625,
0.392578125,
0.459228515625,
0.393798828125,
0.459228515625,
0.407958984375,
0.14697265625,
0.40673828125
],
[
0.515625,
0.3955078125,
0.8818359375,
0.3955078125,
0.8818359375,
0.4091796875,
0.515625,
0.4091796875
],
[
0.1485595703125,
0.407958984375,
0.459228515625,
0.407958984375,
0.459228515625,
0.421630859375,
0.1485595703125,
0.421630859375
],
[
0.5146484375,
0.410400390625,
0.8818359375,
0.410888671875,
0.8818359375,
0.42578125,
0.5146484375,
0.424560546875
],
[
0.14697265625,
0.4228515625,
0.45751953125,
0.4228515625,
0.45751953125,
0.435791015625,
0.14697265625,
0.435791015625
],
[
0.51708984375,
0.427490234375,
0.8818359375,
0.427490234375,
0.8818359375,
0.441162109375,
0.51708984375,
0.441162109375
],
[
0.14697265625,
0.435791015625,
0.459228515625,
0.435791015625,
0.459228515625,
0.449462890625,
0.14697265625,
0.449462890625
],
[
0.5146484375,
0.443603515625,
0.8818359375,
0.443603515625,
0.8818359375,
0.457275390625,
0.5146484375,
0.457275390625
],
[
0.1485595703125,
0.45068359375,
0.45751953125,
0.45068359375,
0.45751953125,
0.46435546875,
0.1485595703125,
0.46435546875
],
[
0.14697265625,
0.465576171875,
0.455078125,
0.465576171875,
0.455078125,
0.478515625,
0.14697265625,
0.478515625
],
[
0.51708984375,
0.460205078125,
0.88330078125,
0.460205078125,
0.88330078125,
0.473876953125,
0.51708984375,
0.473876953125
],
[
0.1502685546875,
0.48095703125,
0.456787109375,
0.48095703125,
0.456787109375,
0.4912109375,
0.1502685546875,
0.4912109375
],
[
0.515625,
0.4755859375,
0.88330078125,
0.4755859375,
0.88330078125,
0.490478515625,
0.515625,
0.490478515625
],
[
0.14697265625,
0.493408203125,
0.459228515625,
0.493408203125,
0.459228515625,
0.50732421875,
0.14697265625,
0.50732421875
],
[
0.515625,
0.4921875,
0.880859375,
0.4921875,
0.880859375,
0.505859375,
0.515625,
0.505859375
],
[
0.14697265625,
0.50830078125,
0.45751953125,
0.50830078125,
0.45751953125,
0.521484375,
0.14697265625,
0.521484375
],
[
0.515625,
0.50732421875,
0.8818359375,
0.50830078125,
0.8818359375,
0.5224609375,
0.515625,
0.521484375
],
[
0.145263671875,
0.52099609375,
0.459228515625,
0.521484375,
0.459228515625,
0.5361328125,
0.145263671875,
0.53515625
],
[
0.515625,
0.52490234375,
0.8818359375,
0.52490234375,
0.8818359375,
0.5380859375,
0.515625,
0.5380859375
],
[
0.1485595703125,
0.53759765625,
0.456787109375,
0.53759765625,
0.456787109375,
0.54736328125,
0.1485595703125,
0.54736328125
],
[
0.515625,
0.54150390625,
0.60693359375,
0.54150390625,
0.60693359375,
0.552734375,
0.515625,
0.552734375
],
[
0.1485595703125,
0.5498046875,
0.45751953125,
0.5498046875,
0.45751953125,
0.5634765625,
0.1485595703125,
0.5634765625
],
[
0.53076171875,
0.5556640625,
0.88330078125,
0.55712890625,
0.88330078125,
0.57177734375,
0.53076171875,
0.57080078125
],
[
0.14697265625,
0.56396484375,
0.4609375,
0.56396484375,
0.4609375,
0.57763671875,
0.14697265625,
0.57763671875
],
[
0.1485595703125,
0.5791015625,
0.45751953125,
0.580078125,
0.45751953125,
0.59130859375,
0.1485595703125,
0.59033203125
],
[
0.515625,
0.57373046875,
0.8818359375,
0.57373046875,
0.8818359375,
0.58740234375,
0.515625,
0.58740234375
],
[
0.14697265625,
0.5927734375,
0.456787109375,
0.5927734375,
0.456787109375,
0.60546875,
0.14697265625,
0.60546875
],
[
0.5146484375,
0.58837890625,
0.880859375,
0.58837890625,
0.880859375,
0.6015625,
0.5146484375,
0.6015625
],
[
0.14697265625,
0.60693359375,
0.456787109375,
0.60693359375,
0.456787109375,
0.62060546875,
0.14697265625,
0.62060546875
],
[
0.515625,
0.60546875,
0.8818359375,
0.60498046875,
0.8818359375,
0.619140625,
0.515625,
0.62060546875
],
[
0.1485595703125,
0.62060546875,
0.459228515625,
0.62060546875,
0.459228515625,
0.63427734375,
0.1485595703125,
0.63427734375
],
[
0.515625,
0.62158203125,
0.8818359375,
0.62158203125,
0.8818359375,
0.6357421875,
0.515625,
0.6357421875
],
[
0.14697265625,
0.63525390625,
0.45751953125,
0.63525390625,
0.45751953125,
0.6484375,
0.14697265625,
0.6484375
],
[
0.5146484375,
0.6357421875,
0.8818359375,
0.63720703125,
0.8818359375,
0.65185546875,
0.5146484375,
0.65087890625
],
[
0.1485595703125,
0.65185546875,
0.455078125,
0.65185546875,
0.455078125,
0.662109375,
0.1485595703125,
0.662109375
],
[
0.515625,
0.65283203125,
0.88330078125,
0.65478515625,
0.88330078125,
0.66943359375,
0.515625,
0.66748046875
],
[
0.1485595703125,
0.66455078125,
0.459228515625,
0.66455078125,
0.459228515625,
0.677734375,
0.1485595703125,
0.677734375
],
[
0.5146484375,
0.66845703125,
0.8818359375,
0.66943359375,
0.8818359375,
0.68408203125,
0.5146484375,
0.68310546875
],
[
0.1485595703125,
0.677734375,
0.459228515625,
0.677734375,
0.459228515625,
0.6923828125,
0.1485595703125,
0.6923828125
],
[
0.5146484375,
0.68603515625,
0.8818359375,
0.68603515625,
0.8818359375,
0.69970703125,
0.5146484375,
0.69970703125
],
[
0.1485595703125,
0.6923828125,
0.45751953125,
0.6923828125,
0.45751953125,
0.7060546875,
0.1485595703125,
0.7060546875
],
[
0.14697265625,
0.70703125,
0.45751953125,
0.70703125,
0.45751953125,
0.72021484375,
0.14697265625,
0.72021484375
],
[
0.51708984375,
0.70166015625,
0.8818359375,
0.70166015625,
0.8818359375,
0.71533203125,
0.51708984375,
0.71533203125
],
[
0.1485595703125,
0.72021484375,
0.459228515625,
0.72021484375,
0.459228515625,
0.73388671875,
0.1485595703125,
0.73388671875
],
[
0.51708984375,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.51708984375,
0.732421875
],
[
0.14697265625,
0.7353515625,
0.459228515625,
0.7353515625,
0.459228515625,
0.7490234375,
0.14697265625,
0.7490234375
],
[
0.5146484375,
0.73291015625,
0.8818359375,
0.73388671875,
0.8818359375,
0.7490234375,
0.5146484375,
0.74755859375
],
[
0.14697265625,
0.7490234375,
0.45751953125,
0.7490234375,
0.45751953125,
0.7626953125,
0.14697265625,
0.7626953125
],
[
0.515625,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.515625,
0.76318359375
],
[
0.14697265625,
0.76318359375,
0.459228515625,
0.76318359375,
0.459228515625,
0.77685546875,
0.14697265625,
0.77685546875
],
[
0.515625,
0.76708984375,
0.8818359375,
0.76708984375,
0.8818359375,
0.78076171875,
0.515625,
0.78076171875
],
[
0.14697265625,
0.77685546875,
0.456787109375,
0.77685546875,
0.456787109375,
0.79052734375,
0.14697265625,
0.79052734375
],
[
0.515625,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.515625,
0.79638671875
],
[
0.14697265625,
0.79345703125,
0.20654296875,
0.79345703125,
0.20654296875,
0.8046875,
0.14697265625,
0.8046875
],
[
0.5146484375,
0.798828125,
0.880859375,
0.798828125,
0.880859375,
0.8125,
0.5146484375,
0.8125
],
[
0.151123046875,
0.8193359375,
0.26611328125,
0.82080078125,
0.26611328125,
0.8349609375,
0.151123046875,
0.83447265625
],
[
0.5146484375,
0.81396484375,
0.55517578125,
0.81396484375,
0.55517578125,
0.830078125,
0.5146484375,
0.830078125
],
[
0.118408203125,
0.82177734375,
0.1435546875,
0.82177734375,
0.1435546875,
0.8330078125,
0.118408203125,
0.8330078125
],
[
0.14111328125,
0.82470703125,
0.1485595703125,
0.82470703125,
0.1485595703125,
0.830078125,
0.14111328125,
0.830078125
],
[
0.5341796875,
0.83203125,
0.779296875,
0.83203125,
0.779296875,
0.845703125,
0.5341796875,
0.845703125
],
[
0.12005615234375,
0.845703125,
0.48779296875,
0.8466796875,
0.48779296875,
0.86083984375,
0.12005615234375,
0.85986328125
],
[
0.12176513671875,
0.86328125,
0.48779296875,
0.86328125,
0.48779296875,
0.876953125,
0.12176513671875,
0.876953125
],
[
0.53076171875,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.87744140625,
0.53076171875,
0.87744140625
],
[
0.12005615234375,
0.87890625,
0.486083984375,
0.87890625,
0.486083984375,
0.892578125,
0.12005615234375,
0.892578125
],
[
0.54736328125,
0.87890625,
0.88330078125,
0.87744140625,
0.88330078125,
0.892578125,
0.54736328125,
0.8935546875
],
[
0.118408203125,
0.8955078125,
0.486083984375,
0.8955078125,
0.486083984375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.5498046875,
0.8955078125,
0.85546875,
0.8955078125,
0.85546875,
0.9091796875,
0.5498046875,
0.9091796875
]
] |
[
"学习更好的单词内部结构以进行序列标记",
"Yingwei Xin, Ethan Hart, Vibhuti Mahajan, Jean-David Ruvini",
"eBay Research",
"2025 Hamilton Ave, San Jose, CA 95125,USA",
" yixin,ejhart,vibmahajan.jean-david.ruvini 1 @ebay.com",
"Abstract",
"义的感兴趣类别，例如人员，组织，位置等。",
"词性（POS）标记为输入句子中的每个单词分",
"Character-based neural models have re",
"配词性。语法分块将文本分为语法相关的，不",
"cently proven very useful for many NLP",
"重叠的单词组。序列标记是一个具有挑战性的",
"tasks. However, there is a gap of sophisti-",
"问题，因为人工注释非常昂贵，通常只有少量",
"cationbetweenmethodsforlearningrepre-",
"sentations of sentences and words. While",
"的标记数据可用",
"most character models for learning repre-",
"大多数传统的序列标记系统都被严重依赖特",
"sentations of sentences are deep and com-",
"征工程的线性统计模型所主导。因此，精心构",
"plex, models for learning representations",
"建的手工特征和特定领域的知识被广泛用于",
"of words are shallow and simple.Also,",
"解决这些任务。不幸的是，开发特定领域的知",
"in spite of considerable research on learn-",
"识和手工制作的功能是昂贵的。最近，使用字",
"ing character embeddings, it is still not",
"符级信息的神经网络已成功用于最小化特征",
"clear which kind of architecture is the best",
"for capturing character-to-word represen",
"工程的需求。基于字符的建模基本上有两个线",
"tations.To addressthesequestions,we",
"程，一个侧重于学习语义和语法（Zhang et al.,",
"first investigate the gaps between methods",
"2015；Conneauetal.，2017）的句子表示；另一",
"for learning word and sentence represen-",
"个侧重于学习单词的表示，以消除单词形状信",
"tations. We conduct detailed experiments",
"息 (Lample et al.,2016;Ma and Hovy,2016)",
"and comparisons of different state-of-the-",
"的手工特征",
"art convolutional models, and also investi-",
"从后面的线程中出现了两种学习序列标记",
"gate the advantages and disadvantages of",
"their constituents.Furthermore, we pro-",
"的字符表示的主要最新方法。一种基于RNN，",
"pose IntNet, a funnel-shaped wide convo-",
"并使用双向LSTM或GRU来学习正向和反",
"lutional neural architecture with no down-",
"向字符信息（Lingetal.，2015;Lampleetal.,",
"sampling for learning representations of",
"2016；Yangetal.，2017）。另一种方法是基于",
"the internal structure of words by compos-",
"CNN，每个单词周围都有一个固定大小的窗",
"ing their characters from limited, super",
"口来创建字符级表示（SantosandZadrozny，",
"vised training corpora. We evaluate our",
"2014; Chiu and Nichols, 2016; Ma and Hovy,",
"proposed model on six sequence labeling",
"2016）。但是，与基于单词的方法相比，基于",
"datasets, including named entity recogni-",
"tion, part-of-speech tagging, and syntac-",
"字符的学习句子表示方法之间的复杂程度存在",
"tic chunking. Our in-depth analysis shows",
"差距。。我们发现，大多数最先进的基于字符的",
"that IntNet significantly outperforms other",
"CNN单词模型都使用卷积，然后使用最大池",
"character embedding models and obtains",
"化作为浅层特征提取器，这与具有深入而复杂",
"new state-of-the-art performance without",
"的句子体系结构的CNN模型有很大不同。尽",
"relying on any external knowledge or re-",
"管对学习字符嵌人进行了大量研究，但仍不清",
"sources.",
"楚哪种体系结构最适合捕获字符到单词的表示",
"Introduction",
"形式",
"1. ",
"",
"因此，许多问题仍然悬而未决：",
"序列标记是为数据序列的每个元素分配标签或",
"类的任务，并且是许多自然语言处理（NLP）",
". Why is there a gap between methods for",
"任务的第一阶段之一。例如，命名实体识别",
"learning representations of sentences and",
"（NER）旨在将句子中的单词分类为几个预定",
"words? How can this gap be bridged?"
] |
{
"arxiv_id": "1810.12443",
"language": "zh",
"page_idx": 0
}
|
||
1810.12443-zh-1
|
1810.12443-zh-1.jpg
|
$\bullet$ How do state-of-the-art character embed-
ding models differ in term of perfor-
mance?
$\bullet$ What kind of neural network architecture is better for learning the internal structure of a word? Deep or shallow? Narrow or wide?
为了回答这些问题，我们首先研究卷积体系结构的学习单词表示和句子表示之间的差距。最直接的想法是添加更多的卷积层，这些卷积层遵循学习句子表示的方法。有趣的是，我们观察到准确性并没有增加太多，并且发现当我们增加网络的深度时，准确性会下降。此观察结果表明，学习单词内部结构的字符表示与句子有很大不同，这也可能解释了基于字符的CNN 模型在表示单词和句子方面存在差距的原因之一
在本文中，我们介绍了来自自然语言处理和计算机视觉的不同最新卷积模型的详细实验和比较。我们还研究了它们的某些成分在不同卷积体系结构上的优缺点。此外，我们提出了一个漏斗形的宽卷积神经网络 IntNet,用于通过组成字符来学习单词的内部结构。与以前基于 CNN 的方法不同，我们的漏斗形 IntNet 探索了更深入，更广泛的体系结构，而没有下采样，用于从有限的监督训练语料库中学习字符到单词的表示。最后，我们将 IntNet 模型与LSTM-CRF 相结合，该模型捕获单词形状和上下文信息，并联合解码标签以进行序列标记
本文的主要贡献如下：
$\bullet$We conduct detailed studies on inves-
tigating the gap between learning word representations and sentence representations.
$\bullet$ We provide in-depth experiments and em-
pirical comparisons of different convolu-
tional models and explore the advantages
and disadvantages of their components
for learning character-to-word representa-
tions.
$\bullet$ We propose a funnel-shaped wide con-
volutional neural architecture with no down-sampling that focuses on learning a better internal structure of words.
$\bullet$ Our proposed compositional character-to-
word model combined with LSTM-CRF achieves state-of-the-art performance for various sequence labeling tasks.
本文的组织如下$:\mathbb{E}$2 节描述了相关工作的多个线程。第 3 节介绍了神经网络的整体架构。第 4 节提供了有关实验设置和比较方法的详细信息。第 5 节报告了不同基准的模型结果，并进行了详细分析和讨论
## 2 Related Work
关于本文的主题，存在三个相关工作线程：(i) 来自不同领域的不同卷积体系结构；(ii)单词的字符嵌入模型；(iii) 使用深度神经网络进行序列标记
跨域的 CNN 模型。卷积神经网络 (CNN) 在从原始信号中提取信息方面非常有用。在NLP 领域，Kim (2014)率先提出了带有单词嵌人的浅层 CNN, 用于句子分类。Zhang et al. (2015)通过直接提取字符级信息来学习句子的语义结构表示，提出了具有 6 个卷积层的 CNN。最近，Conneau et al.(2017) 提出了一种具有 29 个卷积层的 VDCNN 架构， 该架构使用残差连接进行文本分类。然而，一项关于随机丢弃用于训练深度残差网络的层的研究 (Huang et al.,2016)表明，可能不需要所有层，并强调 ResNet (He et al., 2016) 中存在一定程度的冗余。此外，一些研究显示了广泛架构的有希望的结果，例如，广泛的 ResNet (Zagoruyko and Komodakis, 2016) Inception-ResNet (Szegedy et al., 2017) 和
DenseNet (Huang et al.,2017)。这些模型使用字符级信息来学习表示是针对句子而不是单词
字符嵌入模型。Santos and Zadrozny (2014) 提出了一种 CNN 模型来学习单词的字符表示形式，以取代手工制作的词性标记功能。Ling et al. (2015) 提出了一种针对字符的双向 LSTM, 以用作学习字符到单词表示的输人。Chiu and Nichols (2016)通过应用 Santos and Zadrozny (2014)中基于 CNN 的字符嵌入模型，提出了一种带有词典的双向 LSTMCNN, 用于命名实体识别。Plank et al. (2016) 通过遵循 Ling et al. (2015)中基于 LSTM 的字符嵌入模型，提出了一种具有辅助损失的双LSTM 模型，用于多语言词性标记。Cotterell and Heigold (2017) 提出了一种用于神经形态标记的字符级迁移学习模型
序列标记。Collobert et al.(2011) 首先提出了一种基于 CNN-CRF 的方法，该方法可以从单词中学习重要特征，并且几乎不需要手工制作的特征。Huang et al. (2015) 通过使用单词嵌入和手工制作的特征进行序列标记，提出了双向 LSTM-CRF 模型。Lample et al. (2016) 将 Ling et al.(2015)的基于 LSTM 的字符
|
[
[
0.1368408203125,
0.07659912109375,
0.48779296875,
0.07659912109375,
0.48779296875,
0.09027099609375,
0.1368408203125,
0.09027099609375
],
[
0.5322265625,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.5322265625,
0.09027099609375
],
[
0.1561279296875,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.1561279296875,
0.10687255859375
],
[
0.515625,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.515625,
0.10687255859375
],
[
0.15283203125,
0.1092529296875,
0.2166748046875,
0.10687255859375,
0.21826171875,
0.1217041015625,
0.154541015625,
0.12353515625
],
[
0.5146484375,
0.10809326171875,
0.8818359375,
0.1092529296875,
0.8818359375,
0.12353515625,
0.5146484375,
0.122314453125
],
[
0.5146484375,
0.1258544921875,
0.880859375,
0.1258544921875,
0.880859375,
0.1395263671875,
0.5146484375,
0.1395263671875
],
[
0.1368408203125,
0.1353759765625,
0.489501953125,
0.1353759765625,
0.489501953125,
0.1484375,
0.1368408203125,
0.1484375
],
[
0.51318359375,
0.14013671875,
0.771484375,
0.141357421875,
0.771484375,
0.1561279296875,
0.51318359375,
0.155029296875
],
[
0.154541015625,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.1650390625,
0.154541015625,
0.1650390625
],
[
0.1561279296875,
0.16748046875,
0.489501953125,
0.16748046875,
0.489501953125,
0.18115234375,
0.1561279296875,
0.18115234375
],
[
0.51318359375,
0.171630859375,
0.67236328125,
0.171630859375,
0.67236328125,
0.1859130859375,
0.51318359375,
0.1859130859375
],
[
0.1561279296875,
0.185302734375,
0.225830078125,
0.185302734375,
0.225830078125,
0.196533203125,
0.1561279296875,
0.196533203125
],
[
0.515625,
0.1995849609375,
0.8818359375,
0.1995849609375,
0.8818359375,
0.213134765625,
0.515625,
0.213134765625
],
[
0.139404296875,
0.21142578125,
0.48779296875,
0.21142578125,
0.48779296875,
0.2244873046875,
0.139404296875,
0.2244873046875
],
[
0.515625,
0.2161865234375,
0.8818359375,
0.2161865234375,
0.8818359375,
0.2298583984375,
0.515625,
0.2298583984375
],
[
0.12005615234375,
0.226806640625,
0.489501953125,
0.226806640625,
0.489501953125,
0.240478515625,
0.12005615234375,
0.240478515625
],
[
0.515625,
0.2340087890625,
0.880859375,
0.2340087890625,
0.880859375,
0.24462890625,
0.515625,
0.24462890625
],
[
0.12005615234375,
0.243408203125,
0.48779296875,
0.243408203125,
0.48779296875,
0.257080078125,
0.12005615234375,
0.257080078125
],
[
0.5146484375,
0.2493896484375,
0.58837890625,
0.2493896484375,
0.58837890625,
0.26123046875,
0.5146484375,
0.26123046875
],
[
0.12005615234375,
0.2587890625,
0.489501953125,
0.2587890625,
0.489501953125,
0.2724609375,
0.12005615234375,
0.2724609375
],
[
0.5322265625,
0.265380859375,
0.880859375,
0.265380859375,
0.880859375,
0.279052734375,
0.5322265625,
0.279052734375
],
[
0.12005615234375,
0.275634765625,
0.486083984375,
0.275634765625,
0.486083984375,
0.289306640625,
0.12005615234375,
0.289306640625
],
[
0.5146484375,
0.281982421875,
0.88330078125,
0.281982421875,
0.88330078125,
0.295654296875,
0.5146484375,
0.295654296875
],
[
0.12176513671875,
0.29345703125,
0.486083984375,
0.29345703125,
0.486083984375,
0.303955078125,
0.12176513671875,
0.303955078125
],
[
0.51318359375,
0.295654296875,
0.8818359375,
0.297607421875,
0.8818359375,
0.312255859375,
0.51318359375,
0.31005859375
],
[
0.12176513671875,
0.30810546875,
0.48779296875,
0.30810546875,
0.48779296875,
0.32177734375,
0.12176513671875,
0.32177734375
],
[
0.5146484375,
0.3134765625,
0.8818359375,
0.314208984375,
0.8818359375,
0.328857421875,
0.5146484375,
0.327880859375
],
[
0.12176513671875,
0.32373046875,
0.48779296875,
0.32373046875,
0.48779296875,
0.33740234375,
0.12176513671875,
0.33740234375
],
[
0.5146484375,
0.330078125,
0.880859375,
0.330078125,
0.880859375,
0.34326171875,
0.5146484375,
0.34326171875
],
[
0.118408203125,
0.33837890625,
0.48779296875,
0.339111328125,
0.48779296875,
0.35400390625,
0.118408203125,
0.352783203125
],
[
0.515625,
0.344482421875,
0.8818359375,
0.345703125,
0.8818359375,
0.35986328125,
0.515625,
0.359375
],
[
0.12176513671875,
0.35693359375,
0.1822509765625,
0.35693359375,
0.1822509765625,
0.3681640625,
0.12176513671875,
0.3681640625
],
[
0.5146484375,
0.361083984375,
0.880859375,
0.361083984375,
0.880859375,
0.374755859375,
0.5146484375,
0.374755859375
],
[
0.1376953125,
0.371826171875,
0.48779296875,
0.371826171875,
0.48779296875,
0.385498046875,
0.1376953125,
0.385498046875
],
[
0.515625,
0.377685546875,
0.88330078125,
0.377685546875,
0.88330078125,
0.391357421875,
0.515625,
0.391357421875
],
[
0.12005615234375,
0.388427734375,
0.48779296875,
0.388427734375,
0.48779296875,
0.402099609375,
0.12005615234375,
0.402099609375
],
[
0.51708984375,
0.3955078125,
0.85546875,
0.3955078125,
0.85546875,
0.40625,
0.51708984375,
0.40625
],
[
0.12176513671875,
0.40380859375,
0.48779296875,
0.40380859375,
0.48779296875,
0.41748046875,
0.12176513671875,
0.41748046875
],
[
0.515625,
0.410400390625,
0.8818359375,
0.410400390625,
0.8818359375,
0.42333984375,
0.515625,
0.42333984375
],
[
0.118408203125,
0.419189453125,
0.476806640625,
0.419189453125,
0.476806640625,
0.432861328125,
0.118408203125,
0.432861328125
],
[
0.5146484375,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.440673828125,
0.5146484375,
0.440673828125
],
[
0.118408203125,
0.435791015625,
0.486083984375,
0.435791015625,
0.486083984375,
0.449462890625,
0.118408203125,
0.449462890625
],
[
0.51318359375,
0.4423828125,
0.880859375,
0.441162109375,
0.880859375,
0.4560546875,
0.51318359375,
0.457275390625
],
[
0.12005615234375,
0.453125,
0.48779296875,
0.453125,
0.48779296875,
0.466064453125,
0.12005615234375,
0.466064453125
],
[
0.5146484375,
0.457763671875,
0.880859375,
0.457763671875,
0.880859375,
0.47265625,
0.5146484375,
0.47265625
],
[
0.12005615234375,
0.468505859375,
0.489501953125,
0.468505859375,
0.489501953125,
0.482177734375,
0.12005615234375,
0.482177734375
],
[
0.51318359375,
0.473876953125,
0.8818359375,
0.474365234375,
0.8818359375,
0.4892578125,
0.51318359375,
0.488037109375
],
[
0.12005615234375,
0.485107421875,
0.489501953125,
0.485107421875,
0.489501953125,
0.498779296875,
0.12005615234375,
0.498779296875
],
[
0.5146484375,
0.4912109375,
0.880859375,
0.4912109375,
0.880859375,
0.5048828125,
0.5146484375,
0.5048828125
],
[
0.118408203125,
0.50048828125,
0.486083984375,
0.50048828125,
0.486083984375,
0.51416015625,
0.118408203125,
0.51416015625
],
[
0.52490234375,
0.50732421875,
0.8818359375,
0.50732421875,
0.8818359375,
0.52099609375,
0.52490234375,
0.52099609375
],
[
0.12005615234375,
0.5166015625,
0.48779296875,
0.515625,
0.48779296875,
0.52978515625,
0.12005615234375,
0.53076171875
],
[
0.5146484375,
0.5224609375,
0.88330078125,
0.5224609375,
0.88330078125,
0.53759765625,
0.5146484375,
0.53759765625
],
[
0.12005615234375,
0.533203125,
0.48779296875,
0.533203125,
0.48779296875,
0.54638671875,
0.12005615234375,
0.54638671875
],
[
0.515625,
0.5390625,
0.8818359375,
0.5390625,
0.8818359375,
0.552734375,
0.515625,
0.552734375
],
[
0.12176513671875,
0.548828125,
0.48779296875,
0.548828125,
0.48779296875,
0.5625,
0.12176513671875,
0.5625
],
[
0.515625,
0.5556640625,
0.53662109375,
0.5556640625,
0.53662109375,
0.568359375,
0.515625,
0.568359375
],
[
0.1351318359375,
0.5634765625,
0.312255859375,
0.56396484375,
0.312255859375,
0.5791015625,
0.1351318359375,
0.57763671875
],
[
0.5322265625,
0.57177734375,
0.8818359375,
0.57177734375,
0.8818359375,
0.5859375,
0.5322265625,
0.5859375
],
[
0.1376953125,
0.59130859375,
0.484375,
0.59130859375,
0.484375,
0.60498046875,
0.1376953125,
0.60498046875
],
[
0.515625,
0.58837890625,
0.88330078125,
0.58837890625,
0.88330078125,
0.6025390625,
0.515625,
0.6025390625
],
[
0.154541015625,
0.60693359375,
0.489501953125,
0.60546875,
0.489501953125,
0.62060546875,
0.154541015625,
0.62158203125
],
[
0.51318359375,
0.60400390625,
0.880859375,
0.60400390625,
0.880859375,
0.61767578125,
0.51318359375,
0.61767578125
],
[
0.1561279296875,
0.62353515625,
0.48779296875,
0.62353515625,
0.48779296875,
0.63720703125,
0.1561279296875,
0.63720703125
],
[
0.5146484375,
0.62060546875,
0.8818359375,
0.62060546875,
0.8818359375,
0.63427734375,
0.5146484375,
0.63427734375
],
[
0.154541015625,
0.6396484375,
0.20068359375,
0.6396484375,
0.20068359375,
0.65380859375,
0.154541015625,
0.65380859375
],
[
0.515625,
0.63525390625,
0.8818359375,
0.6357421875,
0.8818359375,
0.65087890625,
0.515625,
0.6494140625
],
[
0.5146484375,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.6650390625,
0.5146484375,
0.6650390625
],
[
0.1368408203125,
0.66650390625,
0.483642578125,
0.66650390625,
0.483642578125,
0.6796875,
0.1368408203125,
0.6796875
],
[
0.5146484375,
0.66943359375,
0.8818359375,
0.66943359375,
0.8818359375,
0.68310546875,
0.5146484375,
0.68310546875
],
[
0.1561279296875,
0.681640625,
0.484375,
0.681640625,
0.484375,
0.6953125,
0.1561279296875,
0.6953125
],
[
0.5146484375,
0.68408203125,
0.87890625,
0.68408203125,
0.87890625,
0.69775390625,
0.5146484375,
0.69775390625
],
[
0.154541015625,
0.69677734375,
0.489501953125,
0.69775390625,
0.489501953125,
0.7119140625,
0.154541015625,
0.71142578125
],
[
0.515625,
0.70166015625,
0.880859375,
0.70166015625,
0.880859375,
0.71533203125,
0.515625,
0.71533203125
],
[
0.1561279296875,
0.71435546875,
0.489501953125,
0.71435546875,
0.489501953125,
0.72802734375,
0.1561279296875,
0.72802734375
],
[
0.515625,
0.71728515625,
0.8818359375,
0.71728515625,
0.8818359375,
0.73095703125,
0.515625,
0.73095703125
],
[
0.154541015625,
0.72998046875,
0.48779296875,
0.72998046875,
0.48779296875,
0.74462890625,
0.154541015625,
0.74462890625
],
[
0.515625,
0.73388671875,
0.8818359375,
0.73388671875,
0.8818359375,
0.74755859375,
0.515625,
0.74755859375
],
[
0.154541015625,
0.74462890625,
0.204833984375,
0.74658203125,
0.2032470703125,
0.76123046875,
0.15283203125,
0.7587890625
],
[
0.5146484375,
0.75,
0.8818359375,
0.75,
0.8818359375,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.515625,
0.76416015625,
0.88330078125,
0.765625,
0.88330078125,
0.77978515625,
0.515625,
0.779296875
],
[
0.1351318359375,
0.7724609375,
0.484375,
0.7724609375,
0.484375,
0.7861328125,
0.1351318359375,
0.7861328125
],
[
0.5146484375,
0.7822265625,
0.736328125,
0.7822265625,
0.736328125,
0.7958984375,
0.5146484375,
0.7958984375
],
[
0.1561279296875,
0.7880859375,
0.48779296875,
0.7880859375,
0.48779296875,
0.802734375,
0.1561279296875,
0.802734375
],
[
0.52880859375,
0.7958984375,
0.88330078125,
0.79638671875,
0.88330078125,
0.81396484375,
0.52880859375,
0.8134765625
],
[
0.154541015625,
0.802734375,
0.48779296875,
0.8046875,
0.48779296875,
0.8193359375,
0.154541015625,
0.81787109375
],
[
0.51904296875,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51904296875,
0.82763671875
],
[
0.154541015625,
0.82080078125,
0.429931640625,
0.82080078125,
0.429931640625,
0.83447265625,
0.154541015625,
0.83447265625
],
[
0.515625,
0.83056640625,
0.880859375,
0.83056640625,
0.880859375,
0.84423828125,
0.515625,
0.84423828125
],
[
0.133544921875,
0.845703125,
0.490234375,
0.84423828125,
0.490234375,
0.8623046875,
0.133544921875,
0.86328125
],
[
0.515625,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.515625,
0.86083984375
],
[
0.1561279296875,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.1561279296875,
0.876953125
],
[
0.51318359375,
0.8623046875,
0.880859375,
0.86328125,
0.880859375,
0.87744140625,
0.51318359375,
0.876953125
],
[
0.154541015625,
0.8798828125,
0.490234375,
0.8798828125,
0.490234375,
0.8935546875,
0.154541015625,
0.8935546875
],
[
0.515625,
0.87890625,
0.880859375,
0.87890625,
0.880859375,
0.8935546875,
0.515625,
0.8935546875
],
[
0.1561279296875,
0.8955078125,
0.41064453125,
0.8955078125,
0.41064453125,
0.91015625,
0.1561279296875,
0.91015625
],
[
0.515625,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.91015625,
0.515625,
0.91015625
]
] |
[
": How do state-of-the-art character embed-",
"本文的组织如下：E2节描述了相关工作的",
"ding models differ in term of perfor-",
"多个线程。第3节介绍了神经网络的整体架",
"mance?",
"构。第4节提供了有关实验设置和比较方法",
"的详细信息。第5节报告了不同基准的模型",
". What kind of neural network architecture",
"结果，并进行了详细分析和讨论",
"is better for learning the internal struc",
"ture of a word? Deep or shallow? Narrow",
"2Related Work",
"or wide?",
"关于本文的主题，存在三个相关工作线程：（i）",
"为了回答这些问题，我们首先研究卷积体系",
"来自不同领域的不同卷积体系结构；（ii）单词",
"结构的学习单词表示和句子表示之间的差距。",
"的字符嵌人模型；（i）使用深度神经网络进行",
"最直接的想法是添加更多的卷积层，这些卷",
"序列标记",
"积层遵循学习句子表示的方法。有趣的是，我",
"跨域的CNN模型。卷积神经网络（CNN)",
"们观察到准确性并没有增加太多，并且发现当",
"在从原始信号中提取信息方面非常有用。在",
"我们增加网络的深度时，准确性会下降。此观",
"NLP领域，Kim（2014)率先提出了带有单",
"察结果表明，学习单词内部结构的字符表示与",
"词嵌人的浅层CNN，用于句子分类。Zhang",
"句子有很大不同，这也可能解释了基于字符的",
"etal.（2015）通过直接提取字符级信息来学习",
"CNN模型在表示单词和句子方面存在差距的",
"句子的语义结构表示，提出了具有6个卷积",
"原因之一",
"层的CNN。最近，Conneauetal.（2017）提",
"在本文中，我们介绍了来自自然语言处理和",
"出了一种具有29个卷积层的VDCNN架构，",
"计算机视觉的不同最新卷积模型的详细实验和",
"该架构使用残差连接进行文本分类。然而，",
"比较。我们还研究了它们的某些成分在不同卷",
"项关于随机丢弃用于训练深度残差网络的层",
"积体系结构上的优缺点。此外，我们提出了",
"的研究（Huangetal.，2016）表明，可能不需",
"个漏斗形的宽卷积神经网络IntNet，用于通",
"要所有层，并强调ResNet（Heet al.，2016)",
"过组成字符来学习单词的内部结构。与以前基",
"中存在一定程度的穴余。此外，一些研究显",
"于CNN的方法不同，我们的漏斗形IntNet探",
"示了广泛架构的有希望的结果，例如，广泛",
"索了更深人，更广泛的体系结构，而没有下采",
"的 ResNet (Zagoruyko and Komodakis, 2016)",
"样，用于从有限的监督训练语料库中学习字符",
"Inception-ResNet (Szegedy et al., 2017)和",
"到单词的表示。最后，我们将IntNet模型与",
"DenseNet（Huang et al.，2017）。这些模型使",
"LSTM-CRF相结合，该模型捕获单词形状和",
"用字符级信息来学习表示是针对句子而不是单",
"上下文信息，并联合解码标签以进行序列标记",
"词",
"本文的主要贡献如下：",
"字符嵌人模型。SantosandZadrozny（2014)",
": We conduct detailed studies on inves-",
"提出了一种CNN模型来学习单词的字符表",
"tigating the gap between learning word",
"示形式，以取代手工制作的词性标记功能",
"representations and sentence representa-",
"Lingetal.（2015）提出了一种针对字符的双",
"tions.",
"向LSTM，以用作学习字符到单词表示的输",
"人。ChiuandNichols（2016）通过应用Santos",
": We provide in-depth experiments and em",
"andZadrozny（2014）中基于CNN的字符嵌",
"pirical comparisons of different convolu-",
"入模型，提出了一种带有词典的双向LSTM",
"tional models and explore the advantages",
"CNN，用于命名实体识别。Planketal.（2016)）",
"and disadvantages of their components",
"通过遵循Lingetal.（2015）中基于LSTM的",
"for learning character-to-word representa-",
"字符嵌入模型，提出了一种具有辅助损失的双",
"tions.",
"LSTM模型，用于多语言词性标记。Cotterell",
"andHeigold（2017）提出了一种用于神经形态",
". We propose a funnel-shaped wide con",
"标记的字符级迁移学习模型",
"volutional neural architecture with no",
"序列标记。Collobert et al.（2011）首先提出",
"down-sampling that focuses on learning a",
"了一种基于CNN-CRF的方法，该方法可以从",
"better internal structure of words.",
"单词中学习重要特征，并且几乎不需要手工制",
": Our proposed compositional character-to-",
"作的特征。Huangetal.（2015）通过使用单词",
"word model combined with LSTM-CRF",
"嵌入和手工制作的特征进行序列标记，提出了",
"achieves state-of-the-art performance for",
"双向LSTM-CRF模型。Lampleetal.(2016)",
"various sequence labeling tasks.",
"将Lingetal.（2015）的基于LSTM的字符"
] |
{
"arxiv_id": "1810.12443",
"language": "zh",
"page_idx": 1
}
|
||
1810.12443-zh-3
|
1810.12443-zh-3.jpg
|
$$\begin{aligned}&i_{t}=\sigma(W_{zi}z_{t}+W_{hi}h_{t-1}+W_{ci}c_{t-1}+b_{i})\\&f_{t}=\sigma(W_{zf}z_{t}+W_{hf}h_{t-1}+W_{cf}c_{t-1}+b_{f})\\&\widetilde{c}_{t}=tanh(W_{zc}z_{t}+W_{hc}h_{t-1}+b_{c})\\&c_{t}=f_{t}\odot c_{t-1}+i_{t}\odot\widetilde{c}_{t}\\&o_{t}=\sigma(W_{zo}z_{t}+W_{ho}h_{t-1}+W_{co}c_{t}+b_{o})\\&h_{t}=o_{t}\odot tanh(c_{t}),\end{aligned}$$
<FigureHere>
Figure 1: The main architecture of IntNet.
,它采用与特定过滤器相对应的最大值。这个想法是为每个特征图捕获具有最高值的最重要特征。最后，我们将所有显著特征连接在一起作为这个词的表示：
$$z=[\hat{c_0},\hat{c_1},\ldots\hat{c_u}],$$
(6)
,其中$u$是显着特征的数量，等于最后一层中输出特征图的总数。如果每个函数$\mathcal{F}_l$产生$p$ 特征图，我们得到$(p_0+p\times\frac{L-1}{2})$ 表示，其中$p_0$是初始卷积层中输出特征图的数量
### 3.2 Bi-directional RNN
给定由 IntNet 在公式 6 中计算的字符到单词表示，我们表示输入向量$( z_1$ $, z_2$ $, . . , z_n$ )表示句子。LSTM (Hochreiter and Schmidhuber, 1997)返回序列$(h_1,h_2,。,h_n)$,表示每一步的顺序信息。我们使用以下实现：
,其中 $\sigma$ 是元素的 sigmoid 函数，$\odot$是元素的乘积。$z_t$是时间$t$和$i_t$的输入向量，$f_t,o_t$, $c_t$是输入门，忘记门，输出门和单元向量，所有这些都与隐藏向量$h_t$大小相同。$W_zi$ $, W_zf$ , $W_{zo}$ , $W_{zc}$表示输入$z_t$的不同门的权重矩阵$;W_{hi},W_{hf},W_{ho},W_{hc}$表示隐藏状态$\odot$ 的权重矩阵。正向 LSTM 和反向 LSTM 分别计算句子左右上下文的$\overrightarrow{h_t}$ 和$\overleftarrow{h_t}$ 的表示。我们连接两个隐藏状态以形成双向 LSTM[ $\overrightarrow{h_t},\overleftarrow{h_t}]$ 的输出，用于从两侧捕获上下文信息
## 3.3 Scoring Function
我们不是独立预测每个标签，而是考虑邻域中标签之间的相关性，并通过利用条件随机场(Lafferty et al.,2001)联合解码给定输入句子的最佳标签链。形式上，标签的顺序定义为：
$$y=(y_1,y_2,...,y_T).$$
(7)
要为每个位置$t$定义评分函数$f(h,y)$, 我们将隐藏状态$h_t^w$与由标签$y_t$索引的参数向量$w_{y_t}$相乘，以获得双向 LSTM 网络输出的分数矩阵。因此，函数$f$可以写成：
$$f(h,y)=\sum_{t=1}^Tw_{y_t}h_t^w+\sum_{t=1}^T\mathrm{A}_{y_{t-1},y_t}.$$
(8)
在等式 8 中$,A$是过渡分数的矩阵 , A$_i,j$表示从标记$i$到标记$j$的过渡分数，$y_1$是句子的开始标记。令$\mathcal{Y}(h)$表示 $h$ 的可能标签序列集。序列的概率模型在给定$h$的所有可能的标签序列$y$上定义了一系列条件概率$p(y|h)~$, 其形式如下：
(9)
$$p(y|h)=\frac{e^{f(\mathrm{h,y})}}{\sum_{y'\in\mathcal{Y}(\mathrm{h})}e^{f(h,\mathrm{y'})}}.$$
3.4 Objective Function and Inference
对于端到端网络训练，我们使用最大条件似然
估计来最大化正确标签序列的对数概率：
|
[
[
0.210693359375,
0.07659912109375,
0.40478515625,
0.07659912109375,
0.40478515625,
0.0872802734375,
0.210693359375,
0.0872802734375
],
[
0.187255859375,
0.105712890625,
0.2301025390625,
0.105712890625,
0.2301025390625,
0.11639404296875,
0.187255859375,
0.11639404296875
],
[
0.382080078125,
0.105712890625,
0.426513671875,
0.105712890625,
0.426513671875,
0.11639404296875,
0.382080078125,
0.11639404296875
],
[
0.5322265625,
0.10687255859375,
0.84814453125,
0.10687255859375,
0.84814453125,
0.12054443359375,
0.5322265625,
0.12054443359375
],
[
0.301513671875,
0.1175537109375,
0.30908203125,
0.1175537109375,
0.30908203125,
0.12353515625,
0.301513671875,
0.12353515625
],
[
0.5341796875,
0.1258544921875,
0.8623046875,
0.1258544921875,
0.8623046875,
0.14013671875,
0.5341796875,
0.14013671875
],
[
0.274658203125,
0.13720703125,
0.33154296875,
0.13720703125,
0.33154296875,
0.1478271484375,
0.274658203125,
0.1478271484375
],
[
0.2880859375,
0.149658203125,
0.322509765625,
0.149658203125,
0.322509765625,
0.157958984375,
0.2880859375,
0.157958984375
],
[
0.53076171875,
0.1455078125,
0.796875,
0.1455078125,
0.796875,
0.1602783203125,
0.53076171875,
0.1602783203125
],
[
0.53076171875,
0.1650390625,
0.7138671875,
0.1650390625,
0.7138671875,
0.179931640625,
0.53076171875,
0.179931640625
],
[
0.53076171875,
0.185302734375,
0.84033203125,
0.185302734375,
0.84033203125,
0.198974609375,
0.53076171875,
0.198974609375
],
[
0.2509765625,
0.193603515625,
0.26123046875,
0.193603515625,
0.26123046875,
0.196533203125,
0.2509765625,
0.196533203125
],
[
0.5283203125,
0.20068359375,
0.68603515625,
0.203125,
0.68603515625,
0.2203369140625,
0.5283203125,
0.218505859375
],
[
0.361083984375,
0.22802734375,
0.372802734375,
0.22802734375,
0.372802734375,
0.232177734375,
0.361083984375,
0.232177734375
],
[
0.5263671875,
0.2529296875,
0.8818359375,
0.254150390625,
0.8818359375,
0.268310546875,
0.5263671875,
0.267333984375
],
[
0.252685546875,
0.268310546875,
0.365234375,
0.268310546875,
0.365234375,
0.279052734375,
0.252685546875,
0.279052734375
],
[
0.5146484375,
0.268310546875,
0.8818359375,
0.270751953125,
0.8818359375,
0.28515625,
0.5146484375,
0.283203125
],
[
0.5146484375,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.5146484375,
0.2998046875
],
[
0.5146484375,
0.300537109375,
0.8818359375,
0.3017578125,
0.8818359375,
0.31640625,
0.5146484375,
0.3154296875
],
[
0.53076171875,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.333251953125,
0.53076171875,
0.333251953125
],
[
0.515625,
0.333251953125,
0.8818359375,
0.334228515625,
0.8818359375,
0.3486328125,
0.515625,
0.347412109375
],
[
0.5146484375,
0.350830078125,
0.8818359375,
0.349853515625,
0.8818359375,
0.364013671875,
0.5146484375,
0.365234375
],
[
0.2301025390625,
0.35986328125,
0.2568359375,
0.35986328125,
0.2568359375,
0.3681640625,
0.2301025390625,
0.3681640625
],
[
0.2880859375,
0.35986328125,
0.316650390625,
0.35986328125,
0.316650390625,
0.36767578125,
0.2880859375,
0.36767578125
],
[
0.361083984375,
0.361083984375,
0.36865234375,
0.361083984375,
0.36865234375,
0.366455078125,
0.361083984375,
0.366455078125
],
[
0.2301025390625,
0.37060546875,
0.26025390625,
0.37060546875,
0.26025390625,
0.37890625,
0.2301025390625,
0.37890625
],
[
0.353515625,
0.37060546875,
0.385498046875,
0.37060546875,
0.385498046875,
0.37890625,
0.353515625,
0.37890625
],
[
0.515625,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.515625,
0.38134765625
],
[
0.5146484375,
0.38427734375,
0.8818359375,
0.383056640625,
0.8818359375,
0.39794921875,
0.5146484375,
0.3984375
],
[
0.29296875,
0.371826171875,
0.322509765625,
0.371826171875,
0.322509765625,
0.37890625,
0.29296875,
0.37890625
],
[
0.2626953125,
0.40380859375,
0.344970703125,
0.40380859375,
0.344970703125,
0.41455078125,
0.2626953125,
0.41455078125
],
[
0.5146484375,
0.40087890625,
0.8076171875,
0.40087890625,
0.8076171875,
0.41455078125,
0.5146484375,
0.41455078125
],
[
0.301513671875,
0.41748046875,
0.30908203125,
0.41748046875,
0.30908203125,
0.424560546875,
0.301513671875,
0.424560546875
],
[
0.5146484375,
0.4287109375,
0.5615234375,
0.4287109375,
0.5615234375,
0.439453125,
0.5146484375,
0.439453125
],
[
0.55908203125,
0.427490234375,
0.7001953125,
0.427490234375,
0.7001953125,
0.441162109375,
0.55908203125,
0.441162109375
],
[
0.515625,
0.448974609375,
0.88330078125,
0.448974609375,
0.88330078125,
0.4619140625,
0.515625,
0.4619140625
],
[
0.51708984375,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.51708984375,
0.47802734375
],
[
0.515625,
0.48095703125,
0.880859375,
0.48095703125,
0.880859375,
0.49462890625,
0.515625,
0.49462890625
],
[
0.5146484375,
0.495361328125,
0.87158203125,
0.496337890625,
0.87158203125,
0.51123046875,
0.5146484375,
0.51025390625
],
[
0.623046875,
0.53076171875,
0.7734375,
0.53076171875,
0.7734375,
0.5458984375,
0.623046875,
0.5458984375
],
[
0.8564453125,
0.53076171875,
0.88330078125,
0.53076171875,
0.88330078125,
0.54443359375,
0.8564453125,
0.54443359375
],
[
0.2451171875,
0.55712890625,
0.274658203125,
0.55810546875,
0.274658203125,
0.56640625,
0.2451171875,
0.5654296875
],
[
0.5322265625,
0.5546875,
0.8818359375,
0.5546875,
0.8818359375,
0.5693359375,
0.5322265625,
0.5693359375
],
[
0.515625,
0.57177734375,
0.88330078125,
0.57177734375,
0.88330078125,
0.5859375,
0.515625,
0.5859375
],
[
0.1376953125,
0.58837890625,
0.459228515625,
0.58837890625,
0.459228515625,
0.6015625,
0.1376953125,
0.6015625
],
[
0.51708984375,
0.58740234375,
0.8818359375,
0.58740234375,
0.8818359375,
0.6015625,
0.51708984375,
0.6015625
],
[
0.515625,
0.6025390625,
0.81787109375,
0.6025390625,
0.81787109375,
0.6162109375,
0.515625,
0.6162109375
],
[
0.1368408203125,
0.62744140625,
0.486083984375,
0.62744140625,
0.486083984375,
0.64111328125,
0.1368408203125,
0.64111328125
],
[
0.12005615234375,
0.642578125,
0.484375,
0.642578125,
0.484375,
0.65625,
0.12005615234375,
0.65625
],
[
0.64404296875,
0.642578125,
0.6650390625,
0.642578125,
0.6650390625,
0.65625,
0.64404296875,
0.65625
],
[
0.74462890625,
0.642578125,
0.76416015625,
0.642578125,
0.76416015625,
0.65625,
0.74462890625,
0.65625
],
[
0.63232421875,
0.65087890625,
0.84814453125,
0.65380859375,
0.84814453125,
0.6787109375,
0.63232421875,
0.67578125
],
[
0.118408203125,
0.6591796875,
0.486083984375,
0.6591796875,
0.486083984375,
0.6728515625,
0.118408203125,
0.6728515625
],
[
0.5615234375,
0.6591796875,
0.646484375,
0.6591796875,
0.646484375,
0.67333984375,
0.5615234375,
0.67333984375
],
[
0.85791015625,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.6728515625,
0.85791015625,
0.6728515625
],
[
0.12176513671875,
0.67578125,
0.274658203125,
0.67578125,
0.274658203125,
0.689453125,
0.12176513671875,
0.689453125
],
[
0.640625,
0.67578125,
0.66748046875,
0.67578125,
0.66748046875,
0.6884765625,
0.640625,
0.6884765625
],
[
0.74072265625,
0.67578125,
0.767578125,
0.67578125,
0.767578125,
0.6884765625,
0.74072265625,
0.6884765625
],
[
0.5341796875,
0.70166015625,
0.88330078125,
0.70166015625,
0.88330078125,
0.71533203125,
0.5341796875,
0.71533203125
],
[
0.2333984375,
0.708984375,
0.372802734375,
0.708984375,
0.372802734375,
0.7236328125,
0.2333984375,
0.7236328125
],
[
0.462646484375,
0.708984375,
0.48779296875,
0.708984375,
0.48779296875,
0.72265625,
0.462646484375,
0.72265625
],
[
0.51708984375,
0.71728515625,
0.8818359375,
0.71728515625,
0.8818359375,
0.732421875,
0.51708984375,
0.732421875
],
[
0.1318359375,
0.73388671875,
0.486083984375,
0.73388671875,
0.486083984375,
0.74755859375,
0.1318359375,
0.74755859375
],
[
0.5146484375,
0.73291015625,
0.880859375,
0.73291015625,
0.880859375,
0.74658203125,
0.5146484375,
0.74658203125
],
[
0.1275634765625,
0.74072265625,
0.1368408203125,
0.74072265625,
0.1368408203125,
0.74462890625,
0.1275634765625,
0.74462890625
],
[
0.12176513671875,
0.75,
0.489501953125,
0.75,
0.489501953125,
0.76416015625,
0.12176513671875,
0.76416015625
],
[
0.5146484375,
0.7490234375,
0.880859375,
0.7490234375,
0.880859375,
0.7626953125,
0.5146484375,
0.7626953125
],
[
0.1175537109375,
0.76416015625,
0.48779296875,
0.7626953125,
0.48779296875,
0.78076171875,
0.1175537109375,
0.783203125
],
[
0.5146484375,
0.76416015625,
0.884765625,
0.765625,
0.884765625,
0.77978515625,
0.5146484375,
0.779296875
],
[
0.1175537109375,
0.7822265625,
0.4189453125,
0.78076171875,
0.4189453125,
0.7958984375,
0.1175537109375,
0.79638671875
],
[
0.51708984375,
0.783203125,
0.615234375,
0.783203125,
0.615234375,
0.79443359375,
0.51708984375,
0.79443359375
],
[
0.118408203125,
0.80859375,
0.32568359375,
0.80859375,
0.32568359375,
0.822265625,
0.118408203125,
0.822265625
],
[
0.705078125,
0.80810546875,
0.75732421875,
0.8056640625,
0.75830078125,
0.82177734375,
0.7060546875,
0.82373046875
],
[
0.59033203125,
0.82080078125,
0.6513671875,
0.82080078125,
0.6513671875,
0.8359375,
0.59033203125,
0.8359375
],
[
0.85546875,
0.8193359375,
0.884765625,
0.8193359375,
0.884765625,
0.8359375,
0.85546875,
0.8359375
],
[
0.12005615234375,
0.830078125,
0.48779296875,
0.830078125,
0.48779296875,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.6650390625,
0.82763671875,
0.8017578125,
0.82470703125,
0.8037109375,
0.8466796875,
0.66650390625,
0.849609375
],
[
0.12176513671875,
0.84716796875,
0.48779296875,
0.84716796875,
0.48779296875,
0.8623046875,
0.12176513671875,
0.8623046875
],
[
0.5146484375,
0.85791015625,
0.5615234375,
0.85791015625,
0.5615234375,
0.869140625,
0.5146484375,
0.869140625
],
[
0.55908203125,
0.85693359375,
0.8310546875,
0.85546875,
0.8310546875,
0.87060546875,
0.55908203125,
0.87158203125
],
[
0.12005615234375,
0.87890625,
0.47265625,
0.87890625,
0.47265625,
0.892578125,
0.12005615234375,
0.892578125
],
[
0.118408203125,
0.8623046875,
0.48779296875,
0.86083984375,
0.48779296875,
0.8759765625,
0.118408203125,
0.876953125
],
[
0.466796875,
0.8818359375,
0.48779296875,
0.8818359375,
0.48779296875,
0.8896484375,
0.466796875,
0.8896484375
],
[
0.515625,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.892578125,
0.515625,
0.892578125
],
[
0.12005615234375,
0.89453125,
0.403076171875,
0.89453125,
0.403076171875,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.83447265625,
0.8955078125,
0.83447265625,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"characterto\"word\"representations",
"",
"",
"it = o(Wzizt + Whiht-1+ WciCt-1+bi)",
"A",
"ft = o(Wzfzt + Whfht-1 +WefCt-1 +bf)",
"Max-over-time",
"pooling",
"Ct = tanh(Wzczt + Whcht-1 +bc)",
"Ct= ftoct-1+ itoct",
"Ot = o(Wzozt +Whoht-1 + WcoCt +bo)",
"",
"ht = Ot o tanh(ct),",
"to",
"其中α是元素的sigmoid函数，是元素的",
"Conv,n x1,bottleneck",
"乘积。zt是时间t和it的输人向量，ft，Ot",
"Ct是输入门，忘记门，输出门和单元向量，所",
"有这些都与隐藏向量ht大小相同。Wzi，Wzf",
"Wzo，Wzc表示输人zt的不同门的权重矩",
"阵；Whi，Wht，Who，Whc表示隐藏状态",
"的权重矩阵。正向LSTM和反向LSTM分别",
"Conv",
"Con",
"o",
"nx3",
"nx5",
"计算句子左右上下文的ht和ht的表示。我们",
"连接两个隐藏状态以形成双向LSTM[ht,ht",
"nx4",
"Conv,nx1,botieneck",
"的输出，用于从两侧捕获上下文信息",
"4",
"3.3",
"Scoring Function",
"我们不是独立预测每个标签，而是考虑邻域",
"中标签之间的相关性，并通过利用条件随机场",
"(Laffertyetal.，2001）联合解码给定输入句子",
"的最佳标签链。形式上，标签的顺序定义为：",
"y = (y1, y2, .., yT).",
"(7)",
"Co",
"要为每个位置t定义评分函数f（h，y）",
"我们将隐藏状态hw与由标签yt索引的参数",
"Figure I:The main architecture of IntNet.",
"向量wu相乘，以获得双向LSTM网络输出",
"的分数矩阵。因此，函数f可以写成",
"它采用与特定过滤器相对应的最大值。这个",
"想法是为每个特征图捕获具有最高值的最重要",
"T",
"T",
"Zwurhe +ZAu-1ut",
"特征。最后，我们将所有显著特征连接在一起",
"f(h,y) =",
"(8)",
"作为这个词的表示：",
"t=1",
"t=1",
"在等式8中，A是过渡分数的矩阵，Ai表",
"2 = [cCo, Gi, ... Cu],",
"(6)",
"示从标记i到标记j的过渡分数，1是句子",
"其中u是显着特征的数量，等于最后一层中",
"的开始标记。令（h）表示h的可能标签序",
"",
"输出特征图的总数。如果每个函数Fi产生p",
"列集。序列的概率模型在给定h的所有可能的",
"特征图，我们得到（po+p×-）表示，其中",
"标签序列上定义了一系列条件概率p(ylh），",
"po是初始卷积层中输出特征图的数量",
"其形式如下：",
"3.2Bi-directional RNN",
"ef(h,y)",
"p(ylh) :",
"(9)",
"给定由IntNet在公式6中计算的字符到单词",
"Ey'ey(h) ef(h.y)",
"表示，我们表示输人向量（21，22，。，2m）表",
"3.4",
"Objective Function and Inference",
"1997）返回序列（h1，h2，。，hn），表示每",
"示句子。LSTM (Hochreiter and Schmidhuber,",
"-",
"对于端到端网络训练，我们使用最大条件似然",
"步的顺序信息。我们使用以下实现：",
"估计来最大化正确标签序列的对数概率："
] |
{
"arxiv_id": "1810.12443",
"language": "zh",
"page_idx": 3
}
|
||
1810.12443-zh-4
|
1810.12443-zh-4.jpg
|
$$log(p(y|h))=f(h,y)-log\left(\sum_{y'\in\mathcal{Y}(\mathrm{h})}e^{f(h,\mathrm{y}')}\right).$$
在解码时，我们预测获得以下给出的最高分
数的标签序列：
(10)
$$y^*=\arg\max_{y'\in\mathcal{Y}(\mathrm{h})}f(h,\mathrm{y'}).$$
可以通过动态编程有效地计算目标函数及其梯度；为了进行推理，我们使用维特比算法来找到使分数最大化的最佳标签路径
# 4 Experiments
4.1 Datasets
我们对序列标记任务的六个标准数据集进行了实验，i。e。命名实体识别、词性标注和句法分块。为了测试我们提出的模型的有效性，我们不使用语言特定的资源 (如地名词典)、外部知识 (如迁移学习、联合训练)、手工制作的特征或任何字符预处理，我们不将任何罕见的单词替换为 UNKNOWN
命名实体识别。CoNLL-2002 和 CoNLL2003数据集 (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003) 包含西班牙语，荷兰语，英语和德语的命名实体标签作为单独的数据集。这四个数据集包含不同类型的命名实体：位置、人员、组织和杂项实体。与某些方法不同，我们不将验证集与训练集组合在一起。尽管 POS 标签可用于这些数据集， 但我们不会将其用作附加信息，这使我们的方法与迁移学习的方法不同
词性标记。PennTreebank (PTB) (Marcus et al.,1993)的《华尔街日报》(WSJ)部分包含 25 个部分，并将每个单词分类为 45 个POS 标签中的一个。我们采用标准拆分，使用第 0-18 节作为训练数据，第 19-21 节作为开发数据，第 22-24 节作为测试数据
语法分块。CoNLL2000 分块任务 (Tjong Kim Sang and Buchholz, 2000)使用《华尔街日报》语料库中的第 15-18 节进行培训，第20 节进行测试。它定义了 11 种语法块类型(例如。g。, NP, VP, ADJP), 我们采用标准拆分并从训练集中抽取 1000 个句子作为开发集
4.2 Training Settings
初始化。字符嵌人的维度大小为 32,使用均匀分布随机初始化。对于在训练期间更新的随机初始化的单词嵌入，我们采用相同的初
始化方法。对于 IntNet, 初始卷积的过滤器大小为 32,其他卷积的过滤器大小为 16。我们为所有内核使用了大小为[3,4,5]的过滤器。IntNet-5 和 IntNet-9 的卷积层数分别为 5 和9,我们采用了与 ResNet 相同的权重初始化。我们使用预训练的单词嵌入进行初始化，英语使用 GloVe??? 维单词嵌入，西班牙语，荷兰语和德语使用 fastText??? 维单词嵌入。双向LSTM 的状态大小设置为 256。我们对 NER 和 Chunking 采用标准的 BIOES 标记方案
优化。我们采用具有动量的小批量随机梯度下降。批量大小、动量和学习率设置为 10、 0。9 和$\eta _t= \frac {\eta _0}{1+ \rho t}$ ,其中$\eta_0$是初始学习率0。01和$\rho=0.05$是衰减比，梯度限幅的值为 5。Dropout 应用于 IntNet、LSTMs 和 CRF 的输入，其比率为0。5 是固定的，但在 IntNet 内部没有退出
### 4.3 Compared Methods
为了解决 1 节中的那些悬而未决的问题，我们对跨不同领域的不同最新字符嵌人模型进行了详细的实验和经验比较。首先，我们使用随机初始化词嵌入的 LSTM-CRF 作为我们的初始基线。我们在序列标记中采用了两种最先进的方法，分别表示为 char-LSTM (Lample et al., 2016)和 char-CNN (Ma and Hovy, 2016)。我们将更多的层添加到 char-CNN 模型中，并将其分别称为 char-CNN-5 和 char-CNN-9, 分别用于 5 和 9 个卷积层。此外，我们将剩余连接添加到 char-CNN-9 并将其称为 char-ResNet. 此外，我们应用基于 char-ResNet 的 3 个密集块，我们称之为 char-DenseNet,以比较剩余连接和密集连接之间的差异。最后，我们将我们提出的使用不同卷积层的模型称为 charIntNet-5 和 char-IntNet-9
## 5 Results and Analysis
5.1 Character-to-word Models
表 1 显示了六个基准数据集上不同字符到单词模型的性能。对于序列标记，char-LSTM 和char-CNN 是当前最先进的字符嵌入模型，用于学习字符到单词的表示。我们观察到，在大多数情况下，char-LSTM 的性能优于 charCNN,但是，char-CNN 使用卷积层，然后使用最大池化作为浅层特征提取器，这没有探索CNN 的全部潜力
因此，我们基于 char-CNN 实现了两个变体，分别称为 char-CNN-5 和 char-CNN-9。结果表明，对于大多数数据集，当我们直接添加更多层时，F1 得分并没有太大提高。当我们不断增加深度时，我们也观察到一些精度
|
[
[
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
],
[
0.515625,
0.10986328125,
0.88330078125,
0.10986328125,
0.88330078125,
0.12353515625,
0.515625,
0.12353515625
],
[
0.12176513671875,
0.1181640625,
0.3359375,
0.1181640625,
0.3359375,
0.1318359375,
0.12176513671875,
0.1318359375
],
[
0.364501953125,
0.11639404296875,
0.4609375,
0.114013671875,
0.4609375,
0.12890625,
0.364501953125,
0.1312255859375
],
[
0.51318359375,
0.1258544921875,
0.8818359375,
0.1258544921875,
0.8818359375,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.344970703125,
0.1341552734375,
0.405517578125,
0.1341552734375,
0.405517578125,
0.149658203125,
0.344970703125,
0.149658203125
],
[
0.51318359375,
0.14013671875,
0.880859375,
0.14013671875,
0.880859375,
0.15380859375,
0.51318359375,
0.15380859375
],
[
0.1376953125,
0.160888671875,
0.48779296875,
0.162109375,
0.48779296875,
0.177001953125,
0.1376953125,
0.17578125
],
[
0.515625,
0.157958984375,
0.8818359375,
0.157958984375,
0.8818359375,
0.171630859375,
0.515625,
0.171630859375
],
[
0.5146484375,
0.1728515625,
0.880859375,
0.1728515625,
0.880859375,
0.1859130859375,
0.5146484375,
0.1859130859375
],
[
0.12005615234375,
0.1787109375,
0.2392578125,
0.1787109375,
0.2392578125,
0.1923828125,
0.12005615234375,
0.1923828125
],
[
0.5146484375,
0.188232421875,
0.880859375,
0.188232421875,
0.880859375,
0.201904296875,
0.5146484375,
0.201904296875
],
[
0.204833984375,
0.210205078125,
0.399658203125,
0.208984375,
0.399658203125,
0.226806640625,
0.204833984375,
0.22802734375
],
[
0.5146484375,
0.2060546875,
0.8818359375,
0.2060546875,
0.8818359375,
0.2197265625,
0.5146484375,
0.2197265625
],
[
0.453369140625,
0.2120361328125,
0.48779296875,
0.2120361328125,
0.48779296875,
0.2244873046875,
0.453369140625,
0.2244873046875
],
[
0.27783203125,
0.223876953125,
0.33154296875,
0.223876953125,
0.33154296875,
0.235107421875,
0.27783203125,
0.235107421875
],
[
0.515625,
0.221435546875,
0.85791015625,
0.221435546875,
0.85791015625,
0.235107421875,
0.515625,
0.235107421875
],
[
0.139404296875,
0.243408203125,
0.48779296875,
0.243408203125,
0.48779296875,
0.257080078125,
0.139404296875,
0.257080078125
],
[
0.5341796875,
0.2381591796875,
0.88330078125,
0.2381591796875,
0.88330078125,
0.251708984375,
0.5341796875,
0.251708984375
],
[
0.5146484375,
0.254150390625,
0.88330078125,
0.254150390625,
0.88330078125,
0.267333984375,
0.5146484375,
0.267333984375
],
[
0.118408203125,
0.260009765625,
0.486083984375,
0.260009765625,
0.486083984375,
0.273681640625,
0.118408203125,
0.273681640625
],
[
0.118408203125,
0.275634765625,
0.3955078125,
0.276611328125,
0.3955078125,
0.290283203125,
0.118408203125,
0.289306640625
],
[
0.51318359375,
0.270751953125,
0.88330078125,
0.270751953125,
0.88330078125,
0.28515625,
0.51318359375,
0.28515625
],
[
0.5146484375,
0.287353515625,
0.88330078125,
0.287353515625,
0.88330078125,
0.3017578125,
0.5146484375,
0.3017578125
],
[
0.118408203125,
0.303955078125,
0.156982421875,
0.303955078125,
0.156982421875,
0.31640625,
0.118408203125,
0.31640625
],
[
0.151123046875,
0.302734375,
0.267822265625,
0.303955078125,
0.267822265625,
0.318359375,
0.151123046875,
0.317626953125
],
[
0.5146484375,
0.303955078125,
0.8818359375,
0.303955078125,
0.8818359375,
0.317626953125,
0.5146484375,
0.317626953125
],
[
0.5146484375,
0.318359375,
0.8818359375,
0.319580078125,
0.8818359375,
0.334228515625,
0.5146484375,
0.333251953125
],
[
0.1175537109375,
0.32666015625,
0.2392578125,
0.327880859375,
0.2392578125,
0.342529296875,
0.1175537109375,
0.341552734375
],
[
0.515625,
0.3349609375,
0.62451171875,
0.3349609375,
0.62451171875,
0.349853515625,
0.515625,
0.349853515625
],
[
0.12005615234375,
0.349853515625,
0.484375,
0.349853515625,
0.484375,
0.363525390625,
0.12005615234375,
0.363525390625
],
[
0.118408203125,
0.364013671875,
0.48779296875,
0.364013671875,
0.48779296875,
0.377685546875,
0.118408203125,
0.377685546875
],
[
0.51318359375,
0.363525390625,
0.72314453125,
0.363525390625,
0.72314453125,
0.377685546875,
0.51318359375,
0.377685546875
],
[
0.118408203125,
0.380126953125,
0.489501953125,
0.38134765625,
0.489501953125,
0.3955078125,
0.118408203125,
0.394287109375
],
[
0.51318359375,
0.38427734375,
0.880859375,
0.38427734375,
0.880859375,
0.39794921875,
0.51318359375,
0.39794921875
],
[
0.12005615234375,
0.39794921875,
0.489501953125,
0.39794921875,
0.489501953125,
0.410888671875,
0.12005615234375,
0.410888671875
],
[
0.5146484375,
0.40087890625,
0.880859375,
0.40087890625,
0.880859375,
0.41455078125,
0.5146484375,
0.41455078125
],
[
0.12005615234375,
0.41455078125,
0.48779296875,
0.41455078125,
0.48779296875,
0.427490234375,
0.12005615234375,
0.427490234375
],
[
0.51123046875,
0.416259765625,
0.8818359375,
0.41748046875,
0.8818359375,
0.431640625,
0.51123046875,
0.43115234375
],
[
0.12005615234375,
0.429931640625,
0.48779296875,
0.429931640625,
0.48779296875,
0.443603515625,
0.12005615234375,
0.443603515625
],
[
0.5146484375,
0.43408203125,
0.8818359375,
0.43408203125,
0.8818359375,
0.44775390625,
0.5146484375,
0.44775390625
],
[
0.12176513671875,
0.4453125,
0.301513671875,
0.4453125,
0.301513671875,
0.458984375,
0.12176513671875,
0.458984375
],
[
0.515625,
0.449462890625,
0.880859375,
0.449462890625,
0.880859375,
0.463134765625,
0.515625,
0.463134765625
],
[
0.1376953125,
0.46142578125,
0.489501953125,
0.460205078125,
0.489501953125,
0.474365234375,
0.1376953125,
0.4755859375
],
[
0.5146484375,
0.465576171875,
0.880859375,
0.465576171875,
0.880859375,
0.478515625,
0.5146484375,
0.478515625
],
[
0.118408203125,
0.47802734375,
0.48779296875,
0.47802734375,
0.48779296875,
0.4921875,
0.118408203125,
0.4921875
],
[
0.5146484375,
0.482177734375,
0.8818359375,
0.482177734375,
0.8818359375,
0.495361328125,
0.5146484375,
0.495361328125
],
[
0.118408203125,
0.493408203125,
0.489501953125,
0.4921875,
0.489501953125,
0.50732421875,
0.118408203125,
0.50830078125
],
[
0.5146484375,
0.49755859375,
0.8818359375,
0.49755859375,
0.8818359375,
0.51123046875,
0.5146484375,
0.51123046875
],
[
0.12005615234375,
0.51025390625,
0.489501953125,
0.51025390625,
0.489501953125,
0.52392578125,
0.12005615234375,
0.52392578125
],
[
0.51318359375,
0.5126953125,
0.8818359375,
0.51318359375,
0.8818359375,
0.52783203125,
0.51318359375,
0.52685546875
],
[
0.12005615234375,
0.52685546875,
0.489501953125,
0.52685546875,
0.489501953125,
0.54052734375,
0.12005615234375,
0.54052734375
],
[
0.515625,
0.52978515625,
0.8818359375,
0.52978515625,
0.8818359375,
0.54345703125,
0.515625,
0.54345703125
],
[
0.12005615234375,
0.5419921875,
0.491943359375,
0.5419921875,
0.491943359375,
0.5556640625,
0.12005615234375,
0.5556640625
],
[
0.515625,
0.5458984375,
0.8818359375,
0.5458984375,
0.8818359375,
0.55859375,
0.515625,
0.55859375
],
[
0.12176513671875,
0.55859375,
0.489501953125,
0.55859375,
0.489501953125,
0.572265625,
0.12176513671875,
0.572265625
],
[
0.5146484375,
0.56005859375,
0.8818359375,
0.56103515625,
0.8818359375,
0.576171875,
0.5146484375,
0.57470703125
],
[
0.12176513671875,
0.57373046875,
0.490234375,
0.57373046875,
0.490234375,
0.58740234375,
0.12176513671875,
0.58740234375
],
[
0.51318359375,
0.57666015625,
0.880859375,
0.57666015625,
0.880859375,
0.59033203125,
0.51318359375,
0.59033203125
],
[
0.12005615234375,
0.59130859375,
0.48779296875,
0.59130859375,
0.48779296875,
0.60498046875,
0.12005615234375,
0.60498046875
],
[
0.515625,
0.59423828125,
0.8818359375,
0.59423828125,
0.8818359375,
0.60791015625,
0.515625,
0.60791015625
],
[
0.12005615234375,
0.60693359375,
0.32421875,
0.60693359375,
0.32421875,
0.62060546875,
0.12005615234375,
0.62060546875
],
[
0.515625,
0.60986328125,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.515625,
0.62353515625
],
[
0.1376953125,
0.62255859375,
0.489501953125,
0.62255859375,
0.489501953125,
0.6357421875,
0.1376953125,
0.6357421875
],
[
0.51318359375,
0.626953125,
0.72314453125,
0.626953125,
0.72314453125,
0.64013671875,
0.51318359375,
0.64013671875
],
[
0.12005615234375,
0.6396484375,
0.48779296875,
0.6396484375,
0.48779296875,
0.65283203125,
0.12005615234375,
0.65283203125
],
[
0.12176513671875,
0.65478515625,
0.48779296875,
0.65478515625,
0.48779296875,
0.66943359375,
0.12176513671875,
0.66943359375
],
[
0.5146484375,
0.65771484375,
0.5283203125,
0.65771484375,
0.5283203125,
0.66650390625,
0.5146484375,
0.66650390625
],
[
0.54736328125,
0.65380859375,
0.7314453125,
0.65478515625,
0.7314453125,
0.66943359375,
0.54736328125,
0.66845703125
],
[
0.53076171875,
0.66015625,
0.54736328125,
0.6591796875,
0.54736328125,
0.6630859375,
0.53076171875,
0.66455078125
],
[
0.12005615234375,
0.67041015625,
0.486083984375,
0.67041015625,
0.486083984375,
0.68408203125,
0.12005615234375,
0.68408203125
],
[
0.51123046875,
0.6787109375,
0.77490234375,
0.6796875,
0.77490234375,
0.69482421875,
0.51123046875,
0.693359375
],
[
0.12005615234375,
0.68701171875,
0.48779296875,
0.68701171875,
0.48779296875,
0.70068359375,
0.12005615234375,
0.70068359375
],
[
0.12005615234375,
0.703125,
0.3955078125,
0.703125,
0.3955078125,
0.71728515625,
0.12005615234375,
0.71728515625
],
[
0.51708984375,
0.70361328125,
0.880859375,
0.70361328125,
0.880859375,
0.71435546875,
0.51708984375,
0.71435546875
],
[
0.1376953125,
0.7197265625,
0.489501953125,
0.7197265625,
0.489501953125,
0.73291015625,
0.1376953125,
0.73291015625
],
[
0.515625,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.515625,
0.732421875
],
[
0.118408203125,
0.736328125,
0.489501953125,
0.736328125,
0.489501953125,
0.75,
0.118408203125,
0.75
],
[
0.51318359375,
0.73291015625,
0.8818359375,
0.73388671875,
0.8818359375,
0.74755859375,
0.51318359375,
0.74658203125
],
[
0.12005615234375,
0.751953125,
0.48779296875,
0.751953125,
0.48779296875,
0.765625,
0.12005615234375,
0.765625
],
[
0.5146484375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.12005615234375,
0.7685546875,
0.48779296875,
0.7685546875,
0.48779296875,
0.7822265625,
0.12005615234375,
0.7822265625
],
[
0.5146484375,
0.7666015625,
0.880859375,
0.765625,
0.880859375,
0.77978515625,
0.5146484375,
0.78076171875
],
[
0.12176513671875,
0.78369140625,
0.489501953125,
0.78369140625,
0.489501953125,
0.798828125,
0.12176513671875,
0.798828125
],
[
0.515625,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.515625,
0.79638671875
],
[
0.12176513671875,
0.8017578125,
0.486083984375,
0.8017578125,
0.486083984375,
0.8125,
0.12176513671875,
0.8125
],
[
0.515625,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.515625,
0.8125
],
[
0.12005615234375,
0.81640625,
0.14111328125,
0.81640625,
0.14111328125,
0.830078125,
0.12005615234375,
0.830078125
],
[
0.5146484375,
0.8134765625,
0.654296875,
0.81396484375,
0.654296875,
0.8291015625,
0.5146484375,
0.82763671875
],
[
0.5341796875,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.5341796875,
0.84423828125
],
[
0.118408203125,
0.84033203125,
0.306396484375,
0.8427734375,
0.306396484375,
0.85791015625,
0.118408203125,
0.85546875
],
[
0.515625,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.515625,
0.86083984375
],
[
0.12176513671875,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.12176513671875,
0.876953125
],
[
0.5146484375,
0.86328125,
0.88330078125,
0.86328125,
0.88330078125,
0.876953125,
0.5146484375,
0.876953125
],
[
0.12176513671875,
0.8798828125,
0.489501953125,
0.8798828125,
0.489501953125,
0.8935546875,
0.12176513671875,
0.8935546875
],
[
0.5146484375,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.12176513671875,
0.8955078125,
0.48779296875,
0.8955078125,
0.48779296875,
0.9091796875,
0.12176513671875,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"始化方法。对于IntNet，初始卷积的过滤器",
"大小为32，其他卷积的过滤器大小为16。我",
"们为所有内核使用了大小为[3,4,5]的过滤器。",
"log(p(y|h)) = f(h, y) - log",
"Tef(h,y)",
"IntNet-5和IntNet-9的卷积层数分别为5和",
"(y'ey(h)",
"9，我们采用了与ResNet相同的权重初始化。",
"在解码时，我们预测获得以下给出的最高分",
"我们使用预训练的单词嵌人进行初始化，英语",
"使用GloVe??？维单词嵌入，西班牙语，荷兰",
"数的标签序列：",
"语和德语使用fastText??？维单词嵌。双向",
"y* = arg max,f(h,y').",
"LSTM的状态大小设置为256。我们对NER",
"(10)",
"y'ev(h)",
"和Chunking采用标准的BIOES标记方案",
"可以通过动态编程有效地计算目标函数及其",
"优化。我们采用具有动量的小批量随机梯",
"度下降。批量大小、动量和学习率设置为10、",
"梯度；为了进行推理，我们使用维特比算法来",
"找到使分数最大化的最佳标签路径",
"0。9和nt=+，其中o是初始学习率0。",
"01和p=0.05是衰减比，梯度限幅的值为5。",
"4",
"Experiments",
"Dropout应用于IntNet、LSTMs和CRF的",
"输入，其比率为0。5是固定的，但在IntNet",
"4.1 Datasets",
"内部没有退出",
"我们对序列标记任务的六个标准数据集进行了",
"实验，i。e。命名实体识别、词性标注和句法",
"4.3Compared Methods",
"分块。为了测试我们提出的模型的有效性，我",
"为了解决1节中的那些悬而未决的问题，我们",
"们不使用语言特定的资源（如地名词典）、外部",
"对跨不同领域的不同最新字符嵌入模型进行了",
"知识（如迁移学习、联合训练）、手工制作的特",
"详细的实验和经验比较。首先，我们使用随机",
"征或任何字符预处理，我们不将任何罕见的单",
"初始化词嵌人的LSTM-CRF作为我们的初始",
"词替换为UNKNOWN",
"基线。我们在序列标记中采用了两种最先进的",
"命名实体识别。CoNLL-2002和CoNLL2003",
"方法，分别表示为char-LSTM（Lampleetal.，",
"数据集（TjongKimSang，2002；Tjong",
"2016)和 char-CNN (Ma and Hovy,2016)。我",
"KimSangandDeMeulder，2003）包含西班",
"们将更多的层添加到char-CNN模型中，并将",
"牙语，荷兰语，英语和德语的命名实体标签作",
"其分别称为char-CNN-5和char-CNN-9，分别",
"为单独的数据集。这四个数据集包含不同类型",
"用于5和9个卷积层。此外，我们将剩余连接",
"的命名实体：位置、人员、组织和杂项实体。",
"添加到char-CNN-9并将其称为char-ResNet。",
"与某些方法不同，我们不将验证集与训练集组",
"此外，我们应用基于char-ResNet的3个密",
"合在一起。尽管POS标签可用于这些数据集，",
"集块，我们称之为char-DenseNet，以比较剩",
"但我们不会将其用作附加信息，这使我们的方",
"余连接和密集连接之间的差异。最后，我们将",
"法与迁移学习的方法不同",
"我们提出的使用不同卷积层的模型称为char-",
"词性标记。PennTreebank（PTB）（Marcus",
"IntNet-5和 char-IntNet-9",
"etal.，1993）的《华尔街日报》（WSJ）部分",
"包含25个部分，并将每个单词分类为45个",
"5",
"Results and Analysis",
"",
"POS标签中的一个。我们采用标准拆分，使用",
"5.1Character-to-word Models",
"第0-18节作为训练数据，第19-21节作为开",
"发数据，第22-24节作为测试数据",
"表1显示了六个基准数据集上不同字符到单",
"语法分块。CoNLL2000分块任务（Tjong",
"词模型的性能。对于序列标记，char-LSTM和",
"KimSang andBuchholz，2000）使用《华尔",
"char-CNN是当前最先进的字符嵌人模型，用",
"街日报》语料库中的第15-18节进行培训，第",
"于学习字符到单词的表示。我们观察到，在",
"20节进行测试。它定义了11种语法块类型",
"大多数情况下，char-LSTM的性能优于char-",
"例如。g。，NP，VP，ADJP），我们采用标准",
"CNN，但是，char-CNN使用卷积层，然后使",
"拆分并从训练集中抽取1000个句子作为开发",
"用最大池化作为浅层特征提取器，这没有探索",
"集",
"CNN的全部潜力",
"因此，我们基于char-CNN实现了两个变",
"4.2Training Settings",
"体，分别称为char-CNN-5和char-CNN-9。结",
"初始化。字符嵌人的维度大小为32，使用均",
"果表明，对于大多数数据集，当我们直接添",
"匀分布随机初始化。对于在训练期间更新的",
"加更多层时，F1得分并没有太大提高。当我",
"随机初始化的单词嵌入，我们采用相同的初",
"们不断增加深度时，我们也观察到一些精度"
] |
{
"arxiv_id": "1810.12443",
"language": "zh",
"page_idx": 4
}
|
||
2305.00132-zh-0
|
2305.00132-zh-0.jpg
|
LD-GAN: Low-Dimensional Generative Adversarial Network for Spectral Image
Generation with Variance Regularization
Emmanuel Martinez, Roman Jacome, Alejandra Hernandez-Rojas and Henry Arguelld
Universidad Industrial de Santander
$\{$ emmanuel$2162134@$correo,roman$2162474@$correo,maria.hernandez$26@$correo,henarfu® $\}.\text{uis.edu.co }^*$
$^*$This paper was supported by the Vicerrectoría de Investi-
gación y Extensión UIS, project code 3735
## Abstract
Deep learning methods are state-of-the-art for spectral image (SI) computational tasks. However, these methods are constrained in their performance since available datasets are limited due to the highly expensive and long acquisition time. Usually, data augmentation techniques are employed to mitigate the lack of data. Surpassing classical augmentation methods, such as geometric transformations, GANs enable diverse augmentation by learning and sampling from the data distribution. Nevertheless, GAN-based SI generation is challenging since the high-dimensionality nature of this kind of data hinders the convergence of the GAN training yielding to suboptimal generation. To surmount this limitation, we propose low-dimensional GAN (LD-GAN), where we train the GAN employing a low-dimensional representation of the dataset with the latent space of a pretrained autoencoder network. Thus, we generate new low-dimensional samples which are then mapped to the SI dimension with the pretrained decoder network. Besides, we propose a statistical regularization to control the low-dimensional representation variance for the autoencoder training and to achieve high diversity of samples generated with the GAN. We validate our method LDGAN as data augmentation strategy for compressive spectral imaging, SI super-resolution, and RBG to spectral tasks with improvements varying from 0.5 to 1 [dB] in each task respectively. We perform comparisons against the non-data augmentation training, traditional DA, and with the same GAN adjusted and trained to generate the full -sized SIs
The code of this paper can be found in https:
//github.com/romanjacome99/LD\_GAN.git
## 1. Introduction
光谱成像涉及获取？??D 图像 which contain specific electromagnetic radiation from light ,称为光谱图像(SI)。光谱信息允许估计不同材料的独特特征和分布。SI 已用于与对象分类[24] ,图像分割[9] ,显着对象检测 [47] 和对象跟踪 [17] 相关的计算机视觉任务，这些任务已广泛应用于医疗应用[27],地球观测[30],食品质量[35]和监视[10]等领域
数据驱动的深度学习(DL)方法的最新进展为 SI 处理，获取及其应用 [32] 开辟了新的领域。一些示例包括高光谱和多光谱图像融合[22,23,43] ,分类[6,24] 快照压缩光谱成像(CSI)恢复[7,20]或将 RGB 图像映射到 SI [44] 的恢复方法。虽然 DL 在广泛的应用中取得巨大成功的主要原因之一是模型可以提取大型数据集 [26] 的内在结构，从而提高其泛化性能，但在 SI 应用中，数据集的可用样本数有限由于昂贵且采集时间长[14]。因此，DL 方法的性能仍然局限于受限的可用数据
为了解决这个问题，采用了数据增强(DA)策略[39] 。传统的 DA 执行几何变换，例如翻转，旋转和 resizing 原始数据集以生成新的训练样本。然而，这些方法在原始数据集上产生很少的不变性，这很容易被网络学习。最近的方法采用生成对抗网络(GAN)[13],基于学习原始数据集的概率分布并从学习的分布 [2] 生成合成样本来合成新样本。在 SI 生成的上下文中，many works have been developed as a data augmentation strategy for SI classification [15] ,where conditional GANs are employed to generate new labeled data. A more related work is SHS-GAN[16] ,which 是一个条件 GAN,它通过生成器从 RGB 图像映射 SI 来创建新的 SI 样本，并采用评论家网络在映射的 SI 中注人光谱信息。此外，已经提出了基于变分自动编码器(VAE)的方法，以在[33] 中仅生成频谱签名。Unlike current SI generation methods that are conditioned on labels or RGB images, we look to generate new SI samples from unconditional GANs leveraging the available spectral image datasets without additional information
However ,由于 SI 样本的高维性和有限的训练，用于 SI 生成的训练 GAN 具有挑战性。理论研究通过
1
|
[
[
0.08416748046875,
0.1307373046875,
0.8857421875,
0.1337890625,
0.8857421875,
0.1514892578125,
0.08416748046875,
0.1483154296875
],
[
0.279296875,
0.1572265625,
0.69140625,
0.1578369140625,
0.69140625,
0.17236328125,
0.279296875,
0.1717529296875
],
[
0.125,
0.2052001953125,
0.84619140625,
0.2052001953125,
0.84619140625,
0.2197265625,
0.125,
0.2197265625
],
[
0.33251953125,
0.2235107421875,
0.640625,
0.2235107421875,
0.640625,
0.2374267578125,
0.33251953125,
0.2374267578125
],
[
0.148681640625,
0.242431640625,
0.826171875,
0.242431640625,
0.826171875,
0.255615234375,
0.148681640625,
0.255615234375
],
[
0.23779296875,
0.292236328125,
0.313720703125,
0.292236328125,
0.313720703125,
0.306884765625,
0.23779296875,
0.306884765625
],
[
0.50341796875,
0.292236328125,
0.6396484375,
0.292236328125,
0.6396484375,
0.306884765625,
0.50341796875,
0.306884765625
],
[
0.52294921875,
0.3193359375,
0.890625,
0.318115234375,
0.890625,
0.333251953125,
0.52294921875,
0.333984375
],
[
0.098876953125,
0.32763671875,
0.46728515625,
0.32763671875,
0.46728515625,
0.34228515625,
0.098876953125,
0.34228515625
],
[
0.50341796875,
0.335205078125,
0.890625,
0.335205078125,
0.890625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.08087158203125,
0.34521484375,
0.468017578125,
0.34521484375,
0.468017578125,
0.359130859375,
0.08087158203125,
0.359130859375
],
[
0.50244140625,
0.35107421875,
0.890625,
0.35107421875,
0.890625,
0.365478515625,
0.50244140625,
0.365478515625
],
[
0.080078125,
0.35986328125,
0.468017578125,
0.35986328125,
0.468017578125,
0.373779296875,
0.080078125,
0.373779296875
],
[
0.50244140625,
0.365478515625,
0.8896484375,
0.365478515625,
0.8896484375,
0.37890625,
0.50244140625,
0.37890625
],
[
0.080078125,
0.375732421875,
0.468017578125,
0.375732421875,
0.468017578125,
0.389404296875,
0.080078125,
0.389404296875
],
[
0.5048828125,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.39453125,
0.5048828125,
0.39453125
],
[
0.080078125,
0.391357421875,
0.46728515625,
0.391357421875,
0.46728515625,
0.40478515625,
0.080078125,
0.40478515625
],
[
0.5048828125,
0.396484375,
0.89208984375,
0.396484375,
0.89208984375,
0.410400390625,
0.5048828125,
0.410400390625
],
[
0.08087158203125,
0.406005859375,
0.4697265625,
0.406005859375,
0.4697265625,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.50341796875,
0.410888671875,
0.72607421875,
0.410888671875,
0.72607421875,
0.4248046875,
0.50341796875,
0.4248046875
],
[
0.080078125,
0.421630859375,
0.468017578125,
0.421630859375,
0.468017578125,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.525390625,
0.4287109375,
0.890625,
0.4287109375,
0.890625,
0.442626953125,
0.525390625,
0.442626953125
],
[
0.07843017578125,
0.43505859375,
0.46728515625,
0.434326171875,
0.46728515625,
0.448974609375,
0.07843017578125,
0.449462890625
],
[
0.50341796875,
0.443115234375,
0.89208984375,
0.443115234375,
0.89208984375,
0.458251953125,
0.50341796875,
0.458251953125
],
[
0.080078125,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.5048828125,
0.458984375,
0.8837890625,
0.458984375,
0.8837890625,
0.472900390625,
0.5048828125,
0.472900390625
],
[
0.080078125,
0.466552734375,
0.46728515625,
0.466552734375,
0.46728515625,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.50341796875,
0.473388671875,
0.890625,
0.473388671875,
0.890625,
0.4873046875,
0.50341796875,
0.4873046875
],
[
0.08087158203125,
0.48291015625,
0.465576171875,
0.48291015625,
0.465576171875,
0.49365234375,
0.08087158203125,
0.49365234375
],
[
0.50341796875,
0.488037109375,
0.8896484375,
0.488037109375,
0.8896484375,
0.50244140625,
0.50341796875,
0.50244140625
],
[
0.08087158203125,
0.496826171875,
0.468017578125,
0.496826171875,
0.468017578125,
0.51123046875,
0.08087158203125,
0.51123046875
],
[
0.50341796875,
0.50390625,
0.89208984375,
0.50390625,
0.89208984375,
0.517578125,
0.50341796875,
0.517578125
],
[
0.080078125,
0.51123046875,
0.468017578125,
0.51123046875,
0.468017578125,
0.5244140625,
0.080078125,
0.5244140625
],
[
0.08087158203125,
0.5263671875,
0.468017578125,
0.5263671875,
0.468017578125,
0.54052734375,
0.08087158203125,
0.54052734375
],
[
0.5048828125,
0.52099609375,
0.8896484375,
0.52099609375,
0.8896484375,
0.53076171875,
0.5048828125,
0.53076171875
],
[
0.080078125,
0.5390625,
0.468017578125,
0.54150390625,
0.468017578125,
0.55615234375,
0.080078125,
0.5537109375
],
[
0.5048828125,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54736328125,
0.5048828125,
0.54736328125
],
[
0.5048828125,
0.5498046875,
0.8896484375,
0.5498046875,
0.8896484375,
0.56298828125,
0.5048828125,
0.56298828125
],
[
0.080078125,
0.55615234375,
0.4697265625,
0.5546875,
0.4697265625,
0.5693359375,
0.080078125,
0.57080078125
],
[
0.08087158203125,
0.5712890625,
0.46728515625,
0.5712890625,
0.46728515625,
0.5859375,
0.08087158203125,
0.5859375
],
[
0.5048828125,
0.56640625,
0.5556640625,
0.56640625,
0.5556640625,
0.57763671875,
0.5048828125,
0.57763671875
],
[
0.52392578125,
0.58056640625,
0.8896484375,
0.58203125,
0.8896484375,
0.5966796875,
0.52392578125,
0.59521484375
],
[
0.08087158203125,
0.5869140625,
0.468017578125,
0.5869140625,
0.468017578125,
0.60107421875,
0.08087158203125,
0.60107421875
],
[
0.080078125,
0.6015625,
0.46728515625,
0.6015625,
0.46728515625,
0.61572265625,
0.080078125,
0.61572265625
],
[
0.5087890625,
0.5966796875,
0.89208984375,
0.59765625,
0.89208984375,
0.6123046875,
0.5087890625,
0.611328125
],
[
0.080078125,
0.6171875,
0.46728515625,
0.6171875,
0.46728515625,
0.630859375,
0.080078125,
0.630859375
],
[
0.5048828125,
0.6123046875,
0.890625,
0.6123046875,
0.890625,
0.62548828125,
0.5048828125,
0.62548828125
],
[
0.5048828125,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.64013671875,
0.5048828125,
0.64013671875
],
[
0.08087158203125,
0.6337890625,
0.468017578125,
0.6337890625,
0.468017578125,
0.64453125,
0.08087158203125,
0.64453125
],
[
0.080078125,
0.64794921875,
0.4697265625,
0.64794921875,
0.4697265625,
0.6611328125,
0.080078125,
0.6611328125
],
[
0.5048828125,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65576171875,
0.5048828125,
0.65576171875
],
[
0.080078125,
0.662109375,
0.46728515625,
0.662109375,
0.46728515625,
0.67529296875,
0.080078125,
0.67529296875
],
[
0.5048828125,
0.65771484375,
0.890625,
0.65771484375,
0.890625,
0.671875,
0.5048828125,
0.671875
],
[
0.080078125,
0.6767578125,
0.46728515625,
0.6767578125,
0.46728515625,
0.68994140625,
0.080078125,
0.68994140625
],
[
0.5048828125,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.68603515625,
0.5048828125,
0.68603515625
],
[
0.50244140625,
0.68603515625,
0.890625,
0.68798828125,
0.890625,
0.70263671875,
0.50244140625,
0.70068359375
],
[
0.080078125,
0.6923828125,
0.46728515625,
0.6923828125,
0.46728515625,
0.70556640625,
0.080078125,
0.70556640625
],
[
0.07843017578125,
0.70703125,
0.468017578125,
0.70556640625,
0.468017578125,
0.72021484375,
0.07843017578125,
0.7216796875
],
[
0.50341796875,
0.70263671875,
0.890625,
0.70263671875,
0.890625,
0.71630859375,
0.50341796875,
0.71630859375
],
[
0.080078125,
0.72265625,
0.4697265625,
0.72265625,
0.4697265625,
0.736328125,
0.080078125,
0.736328125
],
[
0.50341796875,
0.71826171875,
0.89208984375,
0.71826171875,
0.89208984375,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.370849609375,
0.736328125,
0.4697265625,
0.73779296875,
0.4697265625,
0.75244140625,
0.370849609375,
0.75048828125
],
[
0.5048828125,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.5048828125,
0.7470703125
],
[
0.5048828125,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76220703125,
0.5048828125,
0.76220703125
],
[
0.080078125,
0.7373046875,
0.37744140625,
0.7373046875,
0.37744140625,
0.75048828125,
0.080078125,
0.75048828125
],
[
0.080078125,
0.75390625,
0.46728515625,
0.75390625,
0.46728515625,
0.767578125,
0.080078125,
0.767578125
],
[
0.07843017578125,
0.767578125,
0.468017578125,
0.7666015625,
0.468017578125,
0.78076171875,
0.07843017578125,
0.7822265625
],
[
0.50341796875,
0.76220703125,
0.890625,
0.76318359375,
0.890625,
0.77783203125,
0.50341796875,
0.7763671875
],
[
0.07843017578125,
0.78271484375,
0.4697265625,
0.7822265625,
0.4697265625,
0.796875,
0.07843017578125,
0.79736328125
],
[
0.5048828125,
0.77880859375,
0.890625,
0.77880859375,
0.890625,
0.79248046875,
0.5048828125,
0.79248046875
],
[
0.11273193359375,
0.796875,
0.468017578125,
0.79736328125,
0.468017578125,
0.8125,
0.11273193359375,
0.81103515625
],
[
0.50244140625,
0.79345703125,
0.890625,
0.79248046875,
0.890625,
0.806640625,
0.50244140625,
0.80810546875
],
[
0.08087158203125,
0.81298828125,
0.400390625,
0.81298828125,
0.400390625,
0.82763671875,
0.08087158203125,
0.82763671875
],
[
0.50341796875,
0.8095703125,
0.8896484375,
0.8095703125,
0.8896484375,
0.82275390625,
0.50341796875,
0.82275390625
],
[
0.50341796875,
0.82373046875,
0.890625,
0.82373046875,
0.890625,
0.8369140625,
0.50341796875,
0.8369140625
],
[
0.50244140625,
0.8388671875,
0.890625,
0.8388671875,
0.890625,
0.85302734375,
0.50244140625,
0.85302734375
],
[
0.50341796875,
0.85595703125,
0.83837890625,
0.85595703125,
0.83837890625,
0.8662109375,
0.50341796875,
0.8662109375
],
[
0.096435546875,
0.87353515625,
0.468017578125,
0.875,
0.468017578125,
0.8896484375,
0.096435546875,
0.88818359375
],
[
0.52294921875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.52294921875,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.34228515625,
0.88720703125,
0.34228515625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88623046875,
0.890625,
0.88623046875,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.48193359375,
0.92822265625,
0.490234375,
0.92822265625,
0.490234375,
0.93310546875,
0.48193359375,
0.93310546875
]
] |
[
"LD-GAN: Low-Dimensional Generative Adversarial Network for Spectral Image",
"Generation with Variance Regularization",
"Emmanuel Martinez, Roman Jacome, Alejandra Hernandez-Rojas and Henry Arguello",
"Universidad Industrial de Santander",
"emmanuel2162134@correo,roman2162474@correo,maria.hernandez26@correo,henarfu@1.uis.edu.co",
"Abstract",
"1.Introduction",
"光谱成像涉及获取???D图像whichcontainspecific",
"Deep learning methods are state-of-the-art for spec-",
"electromagneticradiationfromlight，称为光谱图像",
"tral image (SI) computational tasks. However, these",
"（SI）。光谱信息允许估计不同材料的独特特征和分布。",
"methods are constrained in their performance since",
"SI已用于与对象分类[24］，图像分割[9]，显着对象检",
"available datasets are limited due to the highly ex-",
"测「47和对象跟踪【17]相关的计算机视觉任务，这些",
"pensive and long acquisition time. Usually, data aug-",
"任务已广泛应用于医疗应用[27]，地球观测【30]，食",
"mentation techniques are employed to mitigate the lack",
"品质量[35]和监视[10]等领域",
"of data. Surpassing classical augmentation methods,",
"数据驱动的深度学习（DL）方法的最新进展为SI处",
"such as geometric transformations, GANs enable di-",
"理，获取及其应用[32开辟了新的领域。一些示例包",
"verse augmentation by learning and sampling from the",
"括高光谱和多光谱图像融合[22,23,43]】，分类[6,24]",
"data distribution. Nevertheless, GAN-based SI gener-",
"快照压缩光谱成像（CSI）恢复[7,20]或将RGB图像",
"ation is challenging since the high-dimensionality na-",
"映射到 SI[441的恢复方法。虽然DL在广泛的应用中",
"ture of this kind of data hinders the convergence of the",
"取得巨大成功的主要原因之一是模型可以提取大型数",
"GAN training yielding to suboptimal generation. To",
"surmount this limitation, we propose low-dimensional",
"据集26的内在结构，从而提高其泛化性能，但在SI",
"GAN (LD-GAN), where we train the GAN employing",
"应用中，数据集的可用样本数有限由于昂贵且采集时",
"间长14]。因此，DL方法的性能仍然局限于受限的可",
"a low-dimensional representation of the dataset with",
"the latent space of a pretrained autoencoder network.",
"用数据",
"为了解决这个问题，采用了数据增强（DA)策略[39]",
"Thus, we generate new low-dimensional samples which",
"are then mapped to the SI dimension with the pre-",
"。传统的DA执行几何变换，例如翻转，旋转和resizing",
"trained decoder network. Besides, we propose a sta-",
"原始数据集以生成新的训练样本。然而，这些方法在原",
"始数据集上产生很少的不变性，这很容易被网络学习。。",
"tistical regularization to control the low-dimensional",
"representation variance for theautoencoder train-",
"最近的方法采用生成对抗网络（GAN）[13I，基于学习",
"ing and to achieve high diversity of samples gener-",
"原始数据集的概率分布并从学习的分布【2】生成合成样",
"ated with the GAN. We validate our method LD-",
"本来合成新样本。在SI生成的上下文中，manyworks",
"have been developed as a data augmentation strategy",
"GAN as data augmentation strategy for compres-",
"sive spectral imaging, SI super-resolution, and RBG",
"for SI classification [15], where conditional GANs are",
"to spectral tasks with improvements varying from 0.5",
"employed to generate new labeled data. A more related",
"Weperform",
"workisSHS-GAN[16]，which是一个条件GAN，它",
"通过生成器从RGB图像映射SI来创建新的SI样",
"to 1 [dBl in each task respectively.",
"comparisons against the non-data augmentation train-",
"ing, traditional DA, and with the same GAN ad-",
"本，并采用评论家网络在映射的SI中注入光谱信息。",
"justed and trained to generate the full -sized SIs",
"此外，已经提出了基于变分自动编码器（VAE）的方",
"The code of this paper can be found in https:",
"法，以在[33]中仅生成频谱签名。Unlikecurrent SI",
"//github.com/romanjacome99/LD_GAN.git",
"generation methods that are conditioned on labels or",
"RGB images, we look to generate new SI samples from",
"unconditional GANs leveraging the available spectral",
"imagedatasets without additional information.",
"*This paper was supported by the Vicerrectoria de Investi-",
"However，由于SI样本的高维性和有限的训练，用",
"gaci6n y Extension UIS, project code 3735",
"于SI生成的训练GAN具有挑战性。理论研究通过",
""
] |
{
"arxiv_id": "2305.00132",
"language": "zh",
"page_idx": 0
}
|
||
2305.00132-zh-1
|
2305.00132-zh-1.jpg
|
推导 GAN 估计器的收敛速度来描述这些问题，得出对于高维数据和少量样本，收敛速度会降低，从而产生样本生成[5,19,28]的次优性能。实验上，投影的GAN [38] 通过使用从多个特征或编码网络获得的 LD 图像数据集训练 GAN 来解决此问题。与预计的 GAN 不同，我们希望开发一种简单的策略来生成用于训练GAN 的 SI 的唯一 LD 表示。This approach differs from existing generative models based on dimensionality reduction such as VAE [25]. These methods aim to learn the dataset distribution via latent variables that represent the mean and variance of the distribution and with a Kullback-Leibler divergence regularization approximate this distribution to Gaussian. Here, we propose to learn the latent distribution with a GAN and decode the image with a pre-trained decoder network. Additionally LD representations of SIs are very suitable since the high dimensionality of SIs produces high information redundancy in this type of data. Other SI applications have harnessed this in tasks such as the recovery of SIs from CSI measurements [29] , classification [1] or unmixing [45].
所提出的用于生成 SI 的 LD-GAN 使用 SI 数据集来训练自动编码器(AE)网络，以获得 LD 图像数据集。然后，使用 LD 图像数据集对 GAN 进行对抗性训练，以生成新的 LD 图像样本。最后，通过 AE 网络对生成的 LD 图像样本进行解码，得到生成的 SIs。我们在 AE 训练中提出了 LD 表示图像上的统计正则化函数。该正则化函数促进了沿着训练数据批次的 LD 空间上的方差最小化，从而具有更紧凑的 SI 表示，从而产生最佳解码。此标准遵循收缩 AE,该收缩 AE 通过减少与解码器输人[37] 有关的表示的可变性来改善LD 表示。此外，由于我们正在减少生成的数据的维数，为了产生更多样化的 LD 样本，导致 SI 的更多可变性，通过最大化 LD 图像的方差，生成的 LD 图像被正则化用于 GAN 训练
Since we are proposing an LD adversarial training method, we also develop a high dimensional adversarial training with the same GAN model in order to perform comparisons but adjusting the subnetworks to generate the full-sized SIs, which we will call from now on SGAN .实验结果表明，与 S-GAN 相比，所提出的LD-GAN 改善了 GAN 的收敛性。生成的 SI 的视觉质量进一步支持了这种改进，这表明与 S-GAN 相比，所提出的方法产生了更逼真的空间和光谱信息。我们验证了所提出的 LD-GAN 作为用于 CSI 恢复，SI 空间光谱超分辨率和 RGB 到光谱任务的 DA 技术。我们提出了从 0 变化的改进。5 到 1[dB] 与所提出的方法相比，与非 DA 训练，传统 DA,并与 GAN 训练以生成全光谱图像
本文的其余部分组织如下。2 节被公式化为传统的GAN。此后，在 3节中，我们制定了建议的 LD-GAN。然后，4 部分描述了 AE 潜在空间和 GAN 生成的 LD 样本上的统计正则化。第 5 节介绍了用于验证所提出
方法的 SI 应用程序的数学公式。实验结果在 6 节中传
播。最后，在 7 中给出了结论
## 2. Generative Adversarial Network
给定一个数据集$\left\{x^{(i)}\right\}_{i=1}^K$,带有$K$图像和$x^(i)\in$ $\mathbb{R}^{MNL}$,其中$M,N$是空间维度$,L$是光谱带的数量， 并假设数据集遵循分布$p_\mathrm{data}(x)$ ,这是未知的。然后， 生成网络$\mathcal{G}$将从数据$x$中学习分布$p_g$ 。对于生成网络，在输入噪声变量$p_\boldsymbol{z}(\boldsymbol{z})$上假设先验，其中$z\in\mathbb{R}^m$ 和$m$是随机变量维度，通常是高斯分布，它们映射到所需的生成图像。然后，还定义了一个判别网络$\mathcal{D}$ 。此网络将接收生成的样本或数据集样本。GAN[13] 的损失函数为
$$\begin{aligned}\mathcal{L}_{gan}(\mathcal{D},\mathcal{G},\mathcal{I})&=\mathbb{E}_{\boldsymbol{x}\sim p_{\mathrm{data}}(\boldsymbol{x})}\left[\log\left(\mathcal{D}(\mathcal{I}(\boldsymbol{x}))\right)\right]+\\&\mathbb{E}_{\boldsymbol{z}\sim p_{\boldsymbol{z}}(\boldsymbol{z})}\left[\log\left(1-\mathcal{D}\left(\mathcal{G}(\boldsymbol{z})\right)\right)\right],\end{aligned}$$
其中$\mathcal{G}$和$\mathcal{D}$分别是生成器和分配器网络。此外，$\mathcal{I}$ 是一个标识运算符。在对抗性训练中，每个网络$\{\mathcal{G},\mathcal{D}\}$ 竞争 s 以实现 its 目标：$\mathcal{G}$将生成假样本，而$\mathcal{D}$将预测接收到的样本是真还是假。这称为对抗性训练，可以表示为
(2)
$$\{\hat{\mathcal{G}},\hat{\mathcal{D}}\}=\underset{\mathcal{D}}{\arg\min}\underset{\mathcal{G}}{\arg\max}\:\mathcal{L}_{gan}(\mathcal{D},\mathcal{G},\mathcal{I}).$$
尽管 GAN 在生成图像方面取得了最先进的结果，但是当使用的数据集样本具有高维数且样本数量有限时，通过对抗性训练实现 GAN 的正确收敛是具有挑战性的， 例如 SI 的生成。在下面的部分中，我们将分析 GAN 的收敛性
2.1. Analysis on the GAN Convergence Rate
对 GAN 的收敛性质进行了理论研究，涉及数据的维数和训练样本的数量。例如，在[28]中，得出了估计分布与真实源数据分布之间的总变化度量的渐近边界，其中样本复杂度取决于数据的平方维数(在我们的示例中为$MNL$ ), 以及训练数据集 [28, Theorem 19]的样本数的对数因子。此外，在[19]中，通过积分概率度量的收敛速度提供了 GAN 的错误分析，从而产生了在训练样本数量和数据[19,Theorem 5] 维度方面增加的样本复杂性。此外，当源数据可以用 LD 流形[19, Theorem 19] 表示时，收敛速度得到了提高。因此，基于对 GAN 收敛的这些理论见解，我们建议生成SI 的 LD 表示，以改善 GAN 的收敛，并通过预训练的AE 网络对其进行解码来生成高光谱维 SI, 如下所述。
## 3. Generative Adversarial Networks with LowDimensional Representation
本节将描述建议的 LD-GAN, 其中 LD 样本由 GAN
生成，合成 SI 通过解码器网络解码 LD 获得。
2
|
[
[
0.08251953125,
0.09344482421875,
0.468017578125,
0.09344482421875,
0.468017578125,
0.106689453125,
0.08251953125,
0.106689453125
],
[
0.5048828125,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.5048828125,
0.1060791015625
],
[
0.08416748046875,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.08416748046875,
0.12060546875
],
[
0.50341796875,
0.10919189453125,
0.7197265625,
0.10919189453125,
0.7197265625,
0.12249755859375,
0.50341796875,
0.12249755859375
],
[
0.08251953125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08251953125,
0.1351318359375
],
[
0.08087158203125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50244140625,
0.141357421875,
0.80126953125,
0.141357421875,
0.80126953125,
0.1546630859375,
0.50244140625,
0.1546630859375
],
[
0.08416748046875,
0.1546630859375,
0.46728515625,
0.1546630859375,
0.46728515625,
0.1654052734375,
0.08416748046875,
0.1654052734375
],
[
0.08251953125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.08251953125,
0.1817626953125
],
[
0.521484375,
0.1634521484375,
0.8935546875,
0.1634521484375,
0.8935546875,
0.184326171875,
0.521484375,
0.184326171875
],
[
0.08087158203125,
0.1817626953125,
0.468017578125,
0.18310546875,
0.468017578125,
0.1976318359375,
0.08087158203125,
0.1962890625
],
[
0.5009765625,
0.179931640625,
0.8896484375,
0.18310546875,
0.8896484375,
0.2008056640625,
0.5009765625,
0.1976318359375
],
[
0.080078125,
0.1988525390625,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.080078125,
0.212158203125
],
[
0.50244140625,
0.1988525390625,
0.88818359375,
0.199462890625,
0.88818359375,
0.2139892578125,
0.50244140625,
0.21337890625
],
[
0.08087158203125,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2279052734375,
0.08087158203125,
0.2279052734375
],
[
0.50244140625,
0.21337890625,
0.8896484375,
0.2139892578125,
0.8896484375,
0.2291259765625,
0.50244140625,
0.2279052734375
],
[
0.080078125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.080078125,
0.242431640625
],
[
0.50341796875,
0.2298583984375,
0.88623046875,
0.2298583984375,
0.88623046875,
0.24365234375,
0.50341796875,
0.24365234375
],
[
0.07843017578125,
0.2442626953125,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.07843017578125,
0.25830078125
],
[
0.50634765625,
0.246826171875,
0.88818359375,
0.246826171875,
0.88818359375,
0.2568359375,
0.50634765625,
0.2568359375
],
[
0.07843017578125,
0.2568359375,
0.46728515625,
0.2587890625,
0.46728515625,
0.272705078125,
0.07843017578125,
0.270751953125
],
[
0.5048828125,
0.260009765625,
0.890625,
0.260009765625,
0.890625,
0.2734375,
0.5048828125,
0.2734375
],
[
0.08087158203125,
0.274658203125,
0.46728515625,
0.274658203125,
0.46728515625,
0.28857421875,
0.08087158203125,
0.28857421875
],
[
0.5048828125,
0.274658203125,
0.890625,
0.274658203125,
0.890625,
0.28857421875,
0.5048828125,
0.28857421875
],
[
0.080078125,
0.2890625,
0.468017578125,
0.28857421875,
0.468017578125,
0.302978515625,
0.080078125,
0.3037109375
],
[
0.5048828125,
0.291748046875,
0.56884765625,
0.291748046875,
0.56884765625,
0.302978515625,
0.5048828125,
0.302978515625
],
[
0.08087158203125,
0.3037109375,
0.465576171875,
0.3037109375,
0.465576171875,
0.317626953125,
0.08087158203125,
0.317626953125
],
[
0.08251953125,
0.3212890625,
0.1258544921875,
0.3212890625,
0.1258544921875,
0.33203125,
0.08251953125,
0.33203125
],
[
0.15283203125,
0.3193359375,
0.46728515625,
0.3193359375,
0.46728515625,
0.333251953125,
0.15283203125,
0.333251953125
],
[
0.15283203125,
0.322509765625,
0.1568603515625,
0.322509765625,
0.1568603515625,
0.32763671875,
0.15283203125,
0.32763671875
],
[
0.52392578125,
0.3212890625,
0.8408203125,
0.3212890625,
0.8408203125,
0.3359375,
0.52392578125,
0.3359375
],
[
0.08416748046875,
0.3359375,
0.463134765625,
0.3359375,
0.463134765625,
0.3466796875,
0.08416748046875,
0.3466796875
],
[
0.6357421875,
0.33837890625,
0.89208984375,
0.337158203125,
0.89208984375,
0.354736328125,
0.6357421875,
0.35595703125
],
[
0.08251953125,
0.348388671875,
0.463134765625,
0.348388671875,
0.463134765625,
0.3623046875,
0.08251953125,
0.3623046875
],
[
0.08087158203125,
0.363037109375,
0.46728515625,
0.363037109375,
0.46728515625,
0.376953125,
0.08087158203125,
0.376953125
],
[
0.50732421875,
0.371826171875,
0.890625,
0.371826171875,
0.890625,
0.3818359375,
0.50732421875,
0.3818359375
],
[
0.08087158203125,
0.380126953125,
0.47216796875,
0.380126953125,
0.47216796875,
0.393310546875,
0.08087158203125,
0.393310546875
],
[
0.50244140625,
0.3837890625,
0.88818359375,
0.385009765625,
0.88818359375,
0.3984375,
0.50244140625,
0.397705078125
],
[
0.07843017578125,
0.393310546875,
0.325927734375,
0.395751953125,
0.325927734375,
0.4091796875,
0.07843017578125,
0.4072265625
],
[
0.5048828125,
0.40087890625,
0.8896484375,
0.40087890625,
0.8896484375,
0.4140625,
0.5048828125,
0.4140625
],
[
0.10296630859375,
0.412353515625,
0.468017578125,
0.412353515625,
0.468017578125,
0.4267578125,
0.10296630859375,
0.4267578125
],
[
0.5048828125,
0.415283203125,
0.890625,
0.415283203125,
0.890625,
0.4287109375,
0.5048828125,
0.4287109375
],
[
0.08087158203125,
0.427978515625,
0.468017578125,
0.427978515625,
0.468017578125,
0.44140625,
0.08087158203125,
0.44140625
],
[
0.50341796875,
0.431884765625,
0.5537109375,
0.431884765625,
0.5537109375,
0.443115234375,
0.50341796875,
0.443115234375
],
[
0.08087158203125,
0.443115234375,
0.46728515625,
0.443115234375,
0.46728515625,
0.45703125,
0.08087158203125,
0.45703125
],
[
0.08251953125,
0.458984375,
0.46728515625,
0.458984375,
0.46728515625,
0.472900390625,
0.08251953125,
0.472900390625
],
[
0.55126953125,
0.458251953125,
0.89208984375,
0.460205078125,
0.89208984375,
0.477783203125,
0.55126953125,
0.47607421875
],
[
0.08087158203125,
0.473388671875,
0.46728515625,
0.473388671875,
0.46728515625,
0.4873046875,
0.08087158203125,
0.4873046875
],
[
0.64892578125,
0.477783203125,
0.65625,
0.477783203125,
0.65625,
0.48291015625,
0.64892578125,
0.48291015625
],
[
0.708984375,
0.474853515625,
0.71728515625,
0.479736328125,
0.71142578125,
0.48486328125,
0.70361328125,
0.479736328125
],
[
0.080078125,
0.488037109375,
0.468017578125,
0.4873046875,
0.468017578125,
0.501953125,
0.080078125,
0.50244140625
],
[
0.08251953125,
0.50244140625,
0.468017578125,
0.50244140625,
0.468017578125,
0.5166015625,
0.08251953125,
0.5166015625
],
[
0.50341796875,
0.498779296875,
0.8896484375,
0.5,
0.8896484375,
0.5146484375,
0.50341796875,
0.51318359375
],
[
0.50341796875,
0.5146484375,
0.890625,
0.51513671875,
0.890625,
0.52880859375,
0.50341796875,
0.52783203125
],
[
0.08251953125,
0.52099609375,
0.465576171875,
0.52099609375,
0.465576171875,
0.53076171875,
0.08251953125,
0.53076171875
],
[
0.08251953125,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.08251953125,
0.54736328125
],
[
0.50341796875,
0.52880859375,
0.890625,
0.52978515625,
0.890625,
0.54443359375,
0.50341796875,
0.54345703125
],
[
0.08087158203125,
0.548828125,
0.468017578125,
0.548828125,
0.468017578125,
0.5625,
0.08087158203125,
0.5625
],
[
0.50341796875,
0.54541015625,
0.890625,
0.54541015625,
0.890625,
0.5595703125,
0.50341796875,
0.5595703125
],
[
0.080078125,
0.5625,
0.468017578125,
0.56298828125,
0.468017578125,
0.57763671875,
0.080078125,
0.5771484375
],
[
0.5009765625,
0.5595703125,
0.5703125,
0.56103515625,
0.5703125,
0.57568359375,
0.5009765625,
0.57373046875
],
[
0.08087158203125,
0.5791015625,
0.46728515625,
0.5791015625,
0.46728515625,
0.59228515625,
0.08087158203125,
0.59228515625
],
[
0.08087158203125,
0.5947265625,
0.468017578125,
0.5947265625,
0.468017578125,
0.60791015625,
0.08087158203125,
0.60791015625
],
[
0.50341796875,
0.5888671875,
0.84619140625,
0.5888671875,
0.84619140625,
0.60302734375,
0.50341796875,
0.60302734375
],
[
0.080078125,
0.60791015625,
0.2459716796875,
0.609375,
0.2459716796875,
0.62353515625,
0.080078125,
0.62255859375
],
[
0.52294921875,
0.6123046875,
0.890625,
0.61279296875,
0.890625,
0.62744140625,
0.52294921875,
0.626953125
],
[
0.098876953125,
0.62744140625,
0.468017578125,
0.62744140625,
0.468017578125,
0.642578125,
0.098876953125,
0.642578125
],
[
0.50341796875,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.50341796875,
0.642578125
],
[
0.08251953125,
0.64453125,
0.465576171875,
0.64453125,
0.465576171875,
0.65478515625,
0.08251953125,
0.65478515625
],
[
0.50341796875,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50341796875,
0.65771484375
],
[
0.08087158203125,
0.65771484375,
0.46728515625,
0.65771484375,
0.46728515625,
0.671875,
0.08087158203125,
0.671875
],
[
0.5048828125,
0.6611328125,
0.8896484375,
0.6611328125,
0.8896484375,
0.671875,
0.5048828125,
0.671875
],
[
0.08087158203125,
0.67236328125,
0.468017578125,
0.67236328125,
0.468017578125,
0.68603515625,
0.08087158203125,
0.68603515625
],
[
0.50341796875,
0.673828125,
0.88818359375,
0.673828125,
0.88818359375,
0.6875,
0.50341796875,
0.6875
],
[
0.08251953125,
0.689453125,
0.46728515625,
0.689453125,
0.46728515625,
0.69970703125,
0.08251953125,
0.69970703125
],
[
0.5048828125,
0.689453125,
0.890625,
0.689453125,
0.890625,
0.70263671875,
0.5048828125,
0.70263671875
],
[
0.08087158203125,
0.70263671875,
0.468017578125,
0.70263671875,
0.468017578125,
0.71630859375,
0.08087158203125,
0.71630859375
],
[
0.50341796875,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.71728515625,
0.50341796875,
0.71728515625
],
[
0.08087158203125,
0.71826171875,
0.46728515625,
0.71826171875,
0.46728515625,
0.73193359375,
0.08087158203125,
0.73193359375
],
[
0.5048828125,
0.72021484375,
0.8896484375,
0.72021484375,
0.8896484375,
0.73095703125,
0.5048828125,
0.73095703125
],
[
0.08416748046875,
0.73486328125,
0.46728515625,
0.73486328125,
0.46728515625,
0.74560546875,
0.08416748046875,
0.74560546875
],
[
0.5048828125,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.5048828125,
0.74853515625
],
[
0.08087158203125,
0.74853515625,
0.46728515625,
0.75,
0.46728515625,
0.76123046875,
0.08087158203125,
0.76025390625
],
[
0.50341796875,
0.75,
0.8896484375,
0.75,
0.8896484375,
0.76318359375,
0.50341796875,
0.76318359375
],
[
0.08251953125,
0.76318359375,
0.468017578125,
0.76318359375,
0.468017578125,
0.7763671875,
0.08251953125,
0.7763671875
],
[
0.50341796875,
0.7646484375,
0.89208984375,
0.7646484375,
0.89208984375,
0.77783203125,
0.50341796875,
0.77783203125
],
[
0.08251953125,
0.77880859375,
0.468017578125,
0.77880859375,
0.468017578125,
0.79248046875,
0.08251953125,
0.79248046875
],
[
0.5009765625,
0.77880859375,
0.8896484375,
0.77880859375,
0.8896484375,
0.79248046875,
0.5009765625,
0.79248046875
],
[
0.08251953125,
0.79541015625,
0.46728515625,
0.79541015625,
0.46728515625,
0.80615234375,
0.08251953125,
0.80615234375
],
[
0.50244140625,
0.79541015625,
0.88623046875,
0.79541015625,
0.88623046875,
0.8095703125,
0.50244140625,
0.8095703125
],
[
0.08251953125,
0.8095703125,
0.468017578125,
0.8095703125,
0.468017578125,
0.82275390625,
0.08251953125,
0.82275390625
],
[
0.08251953125,
0.82470703125,
0.1650390625,
0.82470703125,
0.1650390625,
0.8359375,
0.08251953125,
0.8359375
],
[
0.50341796875,
0.8291015625,
0.8896484375,
0.8291015625,
0.8896484375,
0.84228515625,
0.50341796875,
0.84228515625
],
[
0.1004638671875,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.1004638671875,
0.85498046875
],
[
0.53125,
0.8447265625,
0.7685546875,
0.84521484375,
0.7685546875,
0.85986328125,
0.53125,
0.859375
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.87060546875,
0.468017578125,
0.88525390625,
0.080078125,
0.88623046875
],
[
0.52294921875,
0.8720703125,
0.8896484375,
0.8720703125,
0.8896484375,
0.88525390625,
0.52294921875,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8310546875,
0.88720703125,
0.8310546875,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.939453125,
0.47802734375,
0.939453125
]
] |
[
"推导GAN估计器的收敛速度来描述这些问题，得出",
"方法的SI应用程序的数学公式。实验结果在6节中传",
"对于高维数据和少量样本，收敛速度会降低，从而产",
"播。最后，在7中给出了结论",
"生样本生成5，19.28的次优性能。实验上，投影的",
"GAN[38]通过使用从多个特征或编码网络获得的LD",
"2.Generative Adversarial Network",
"图像数据集训练GAN来解决此问题。与预计的GAN",
"不同，我们希望开发一种简单的策略来生成用于训练",
"给定一个数据集【a())1，带有K图像和a()e",
"GAN的SI的唯一LD表示。This approachdiffers",
"RMNL，其中M,N是空间维度，L是光谱带的数量，",
"from existing generative models based on dimensional-",
"并假设数据集遵循分布Pdata（α），这是未知的。然后",
"ity reduction such as VAE [25] . These methods aim to",
"生成网络g将从数据a中学习分布pg。对于生成网",
"learn the dataset distribution via latent variables that",
"络，在输人噪声变量Pz（z）上假设先验，其中zER\"",
"represent the mean and variance of the distribution",
"和㎡是随机变量维度，通常是高斯分布，它们映射到",
"and with a Kullback-Leibler divergence regularization",
"所需的生成图像。然后，还定义了一个判别网络D。此",
"approximate this distribution to Gaussian. Here, we",
"网络将接收生成的样本或数据集样本。GAN[13]的损",
"propose to learn the latent distribution with a GAN",
"失函数为",
"and decode the image with a pre-trained decoder net",
"work..",
"Additionally LD representations of SIs are",
"",
"Lgan(D, G,I) = Ea~pdata(a) [log (D(I(α)))] +",
"verysuitablesincethehighdimensionalityof SIspro",
"Ez~pz(z) [1og(1 -D(G(z)] , (1)",
"duces high information redundancy in this type of data",
"Other SI applications have harnessed this in tasks such",
"，其中G和D分别是生成器和分配器网络。此外，工",
"as the recovery of SIs from CSI measurements [29] ,",
"是一个标识运算符。在对抗性训练中，每个网络(9,D",
"classification[1]or unmixing[45]",
"竞争s以实现its目标：G将生成假样本，而D将预",
"所提出的用于生成SI的LD-GAN使用SI数据集",
"测接收到的样本是真还是假。这称为对抗性训练，可以",
"来训练自动编码器（AE）网络，以获得LD图像数据",
"表示为",
"集。然后，使用LD图像数据集对GAN进行对抗性",
"训练，以生成新的LD图像样本。最后，通过AE网",
"[9, D) = argmin arg max Cgan(D,g,I).(2)",
"络对生成的LD图像样本进行解码，得到生成的SIs。",
"L",
"C",
"我们在AE训练中提出了LD表示图像上的统计正则",
"化函数。该正则化函数促进了沿着训练数据批次的LD",
"尽管GAN在生成图像方面取得了最先进的结果，但是",
"当使用的数据集样本具有高维数且样本数量有限时，通",
"空间上的方差最小化，从而具有更紧凑的SI表示，从",
"而产生最佳解码。此标准遵循收缩AE，该收缩AE通",
"过对抗性训练实现GAN的正确收敛是具有挑战性的，",
"过减少与解码器输人[37]有关的表示的可变性来改善",
"例如SI的生成。在下面的部分中，我们将分析GAN",
"LD表示。此外，由于我们正在减少生成的数据的维",
"的收敛性",
"数，为了产生更多样化的LD样本，导致SI的更多可",
"变性，通过最大化LD图像的方差，生成的LD图像被",
"2.1. Analysis on the GAN Convergence Rate",
"正则化用于GAN训练",
"对GAN的收敛性质进行了理论研究，涉及数据的",
"Since we are proposing an LD adversarial training",
"维数和训练样本的数量。例如，在[28]中，得出了估",
"method,wealsodevelopahighdimensionaladversarial",
"计分布与真实源数据分布之间的总变化度量的渐近边",
"training with the same GAN model in order to perform",
"界，其中样本复杂度取决于数据的平方维数（在我们",
"comparisons but adjusting the subnetworks to generate",
"的示例中为MNL），以及训练数据集[28，Theorem",
"thefull-sized SIs,whichwewill call fromnowonS-",
"19]的样本数的对数因子。此外，在【19]中，通过积分",
"GAN：实验结果表明，与S-GAN相比，所提出的",
"概率度量的收敛速度提供了GAN的错误分析，从而",
"LD-GAN改善了GAN的收敛性。生成的SI的视觉质",
"产生了在训练样本数量和数据19，Theorem5]维度方",
"量进一步支持了这种改进，这表明与S-GAN相比，所",
"面增加的样本复杂性。此外，当源数据可以用LD流",
"提出的方法产生了更逼真的空间和光谱信息。我们验",
"形[19，Theorem19]表示时，收敛速度得到了提高。因",
"证了所提出的LD-GAN作为用于CSI恢复，SI空间",
"此，基于对GAN收敛的这些理论见解，我们建议生成",
"光谱超分辨率和RGB到光谱任务的DA技术。我们",
"SI的LD表示，以改善GAN的收敛，并通过预训练的",
"提出了从0变化的改进。5到1dB与所提出的方法相",
"AE网络对其进行解码来生成高光谱维SI，如下所述。",
"比，与非DA训练，传统DA，并与GAN训练以生成",
"全光谱图像",
"3. Generative Adversarial Networks with Low-",
"本文的其余部分组织如下。2节被公式化为传统的",
"DimensionalRepresentation",
"GAN。此后，在3节中，我们制定了建议的LD-GAN。",
"然后，4部分描述了AE潜在空间和GAN生成的LD",
"本节将描述建议的LD-GAN，其中LD样本由GAN",
"样本上的统计正则化。第5节介绍了用于验证所提出",
"生成，合成SI通过解码器网络解码LD获得。",
"2"
] |
{
"arxiv_id": "2305.00132",
"language": "zh",
"page_idx": 1
}
|
||
2305.00132-zh-2
|
2305.00132-zh-2.jpg
|
<FigureHere>
Figure 1. Autoencoder and LD-GAN networks. a) The AE network allows obtaining an LD representation of the SI through an encoder subnetwork and then reconstructs the original SI with a decoder subnetwork. b) The LD representation obtained by the AE is employed to train a GAN to generate new LD representation images. c) Once the GAN is trained, the new LD images can be decoded to recover SIs.
## 3.1. Spectral Image Autoencoder
使用 AE 网络获得具有$K$样本的 SI 数据集$\{x^{(i)}\}_{i=1}^K$ 的 LD 表示。两个子网给出了 AE 的结构， 如图所示。1 (a)。编码器网络$E$将压缩 SI 的 spectral 信息，解码器网络$D$将解码基础 SI。这些子网的优化由下式给出
正则化方法，该方法允许在 LD 空间中紧凑地表示 SI 数据集，从而提高了 AE 恢复性能。然后，我们在生成的 LD 空间上使用方差最大化进行 GAN 训练，以在generated SI 数据集上产生多样化的数据和更高的质量。此基于方差的正则化准则也已用于自监督学习[8] 和稀疏编码[12] 。首先，定义包含 AE 的 LD 表示的集$\left.\mathcal{A}=\{e^{(i)}\}_{i=1}^{K},\text{并定义生成的 LD 图像集 by GAN}\right.$ as $\mathcal{W} = \{ \boldsymbol{d}^{( i) }\} _{i= 1}^{K}$。$\mathcal{A}$ 的方差由 $\sigma_{\mathcal{A}}^{2}\in\mathbb{R}_{+}^{d}$ 表示，而GAN 生成的表示$\sigma_{\mathcal{W}}^2\in\mathbb{R}_+^d$在训练批次维度上以像素为单位。建议的统计正则化函数由
$$\{\hat{E},\hat{D}\}=\sum\limits_{i=1}^K||\boldsymbol{x}^{(i)}-D(E(\boldsymbol{x}^{(i)}))||_2^2,$$
(3)
(5)
$$\mathcal{R}(\cdot)=\|\sigma_{(\cdot)}^2\|_2.$$
其中$\hat{E}$和$\hat{D}$分别是编码器和解码器。因此，SI 的 LD 表示可以作为$b^(i)=E(\boldsymbol{x}^{(i)})$获得，其中$b^(i)\in\mathbb{R}^MNc$ 和$c<L$ 。最后，可以通过将该 LD 图像解码为$\hat{x}^{(i)}=$ $D(\boldsymbol{b}^{(i)})$来恢复 SI
给出，然后，正则化的 AE 训练如下执行
$$\{\hat{E},\hat{D}\}=\arg\min_{E,D}||\boldsymbol{x}^{(i)}-D(E(\boldsymbol{x}^{(i)}))||_{2}^{2}+\mu_{ae}\mathcal{R}(\mathcal{A}).$$
## 3.2. Low-Dimensional Generative Adversarial Network
(6)
,最终的 GAN 优化问题由
我们 employ 将针对 a SI 数据集的 LD 表示进行优化的 GAN。然后，来自方程式 2 的对抗性训练可以重写为
$$\{\hat{\mathcal{G}},\hat{\mathcal{D}}\}=\underset{\mathcal{D}}{\arg\min}\underset{\mathcal{G}}{\arg\max}\mathcal{L}_{gan}(\mathcal{D},\mathcal{G},\hat{E})-\mu_{gan}\mathcal{R}(\mathcal{W}).$$
(7)
(4)
$$\{\hat{\mathcal{G}},\hat{\mathcal{D}}\}=\underset{\mathcal{D}}{\arg\min}\underset{\mathcal{G}}{\arg\max}\:\mathcal{L}_{gan}(\mathcal{D},\mathcal{G},\hat{E}),$$
给出。注意，方差正则化函数中的负号是因为我们想要最大化所生成的 SI 的表示。在 AE 和 GAN 这两个优化问题中，参数$\mu_{ae}$和$\mu_{gan}$是正则化超参数，它们控制我们将 AE LD 空间集中多少或我们将 GAN 生成的图像的可变性增加多少。
其中生成网络可以生成 LD 图像作为$e=\hat{\mathcal{G}}(z)$ ,并且可以使用解码器作为$\hat{f}=\hat{D}(\boldsymbol{e})$获得新的 SI 样本。
$\begin{array}{ll}{4.}&{{\mathrm{Statistical~Regularization~for~the~AE~and}}}\\{{\mathrm{GAN}}}\end{array}$
为了改善 AE 的 LD 表示和 GAN 生成的 LD 图像的多样性，我们在 AE 训练中提出了一种方差最小化
## $5.{\mathrm{~Spectral~Image~Applications}}$
为了验证我们提出的方法的性能，我们解决了以下
基于 DL 的 SI 任务：
3
|
[
[
0.101318359375,
0.0947265625,
0.1544189453125,
0.0947265625,
0.1544189453125,
0.1060791015625,
0.101318359375,
0.1060791015625
],
[
0.27294921875,
0.09344482421875,
0.29736328125,
0.09344482421875,
0.29736328125,
0.1060791015625,
0.27294921875,
0.1060791015625
],
[
0.3896484375,
0.0947265625,
0.483642578125,
0.0947265625,
0.483642578125,
0.1060791015625,
0.3896484375,
0.1060791015625
],
[
0.54833984375,
0.0953369140625,
0.68896484375,
0.10919189453125,
0.685546875,
0.126953125,
0.5458984375,
0.11297607421875
],
[
0.685546875,
0.0985107421875,
0.75634765625,
0.10101318359375,
0.75634765625,
0.11553955078125,
0.685546875,
0.11297607421875
],
[
0.313720703125,
0.0997314453125,
0.369384765625,
0.10162353515625,
0.369384765625,
0.11297607421875,
0.313720703125,
0.111083984375
],
[
0.1070556640625,
0.10479736328125,
0.1478271484375,
0.106689453125,
0.146240234375,
0.11871337890625,
0.10540771484375,
0.11614990234375
],
[
0.193603515625,
0.10101318359375,
0.2484130859375,
0.10101318359375,
0.2484130859375,
0.11236572265625,
0.193603515625,
0.11236572265625
],
[
0.26318359375,
0.10479736328125,
0.3056640625,
0.106689453125,
0.303955078125,
0.11871337890625,
0.261474609375,
0.11614990234375
],
[
0.38818359375,
0.1060791015625,
0.48193359375,
0.1060791015625,
0.48193359375,
0.117431640625,
0.38818359375,
0.117431640625
],
[
0.56787109375,
0.11297607421875,
0.60693359375,
0.11297607421875,
0.60693359375,
0.12371826171875,
0.56787109375,
0.12371826171875
],
[
0.6787109375,
0.111083984375,
0.75244140625,
0.11297607421875,
0.75244140625,
0.125,
0.6787109375,
0.12249755859375
],
[
0.10455322265625,
0.12249755859375,
0.404296875,
0.12249755859375,
0.404296875,
0.184326171875,
0.10455322265625,
0.184326171875
],
[
0.370849609375,
0.1331787109375,
0.455810546875,
0.1331787109375,
0.455810546875,
0.1787109375,
0.370849609375,
0.1787109375
],
[
0.609375,
0.1558837890625,
0.6259765625,
0.1558837890625,
0.6259765625,
0.1641845703125,
0.609375,
0.1641845703125
],
[
0.611328125,
0.1654052734375,
0.625,
0.1654052734375,
0.625,
0.169189453125,
0.611328125,
0.169189453125
],
[
0.7490234375,
0.1641845703125,
0.833984375,
0.1641845703125,
0.833984375,
0.175537109375,
0.7490234375,
0.175537109375
],
[
0.435546875,
0.189453125,
0.4599609375,
0.186279296875,
0.463134765625,
0.199462890625,
0.438720703125,
0.20263671875
],
[
0.6298828125,
0.189453125,
0.66650390625,
0.189453125,
0.66650390625,
0.2052001953125,
0.6298828125,
0.2052001953125
],
[
0.27783203125,
0.1900634765625,
0.301513671875,
0.1900634765625,
0.301513671875,
0.20263671875,
0.27783203125,
0.20263671875
],
[
0.08905029296875,
0.1900634765625,
0.1478271484375,
0.1881103515625,
0.1478271484375,
0.199462890625,
0.08905029296875,
0.2020263671875
],
[
0.19775390625,
0.1912841796875,
0.2442626953125,
0.1912841796875,
0.2442626953125,
0.20263671875,
0.19775390625,
0.20263671875
],
[
0.330810546875,
0.1912841796875,
0.37744140625,
0.1912841796875,
0.37744140625,
0.20263671875,
0.330810546875,
0.20263671875
],
[
0.5908203125,
0.1925048828125,
0.60302734375,
0.1925048828125,
0.60302734375,
0.2008056640625,
0.5908203125,
0.2008056640625
],
[
0.705078125,
0.1944580078125,
0.71142578125,
0.1944580078125,
0.71142578125,
0.199462890625,
0.705078125,
0.199462890625
],
[
0.08416748046875,
0.2108154296875,
0.28515625,
0.2291259765625,
0.281982421875,
0.246826171875,
0.08087158203125,
0.2291259765625
],
[
0.83251953125,
0.205810546875,
0.86767578125,
0.205810546875,
0.86767578125,
0.2183837890625,
0.83251953125,
0.2183837890625
],
[
0.52294921875,
0.21337890625,
0.576171875,
0.21337890625,
0.576171875,
0.2247314453125,
0.52294921875,
0.2247314453125
],
[
0.611328125,
0.216552734375,
0.66845703125,
0.2183837890625,
0.66650390625,
0.23291015625,
0.609375,
0.2310791015625
],
[
0.69677734375,
0.212158203125,
0.72119140625,
0.212158203125,
0.72119140625,
0.2235107421875,
0.69677734375,
0.2235107421875
],
[
0.396240234375,
0.2171630859375,
0.464111328125,
0.2171630859375,
0.464111328125,
0.2291259765625,
0.396240234375,
0.2291259765625
],
[
0.2156982421875,
0.2171630859375,
0.281982421875,
0.2171630859375,
0.281982421875,
0.2291259765625,
0.2156982421875,
0.2291259765625
],
[
0.30224609375,
0.2235107421875,
0.35693359375,
0.2235107421875,
0.35693359375,
0.2342529296875,
0.30224609375,
0.2342529296875
],
[
0.53173828125,
0.222900390625,
0.5703125,
0.2247314453125,
0.56884765625,
0.236083984375,
0.53125,
0.2342529296875
],
[
0.68896484375,
0.2247314453125,
0.7294921875,
0.2247314453125,
0.7294921875,
0.23486328125,
0.68896484375,
0.23486328125
],
[
0.84228515625,
0.2247314453125,
0.875,
0.2247314453125,
0.875,
0.2374267578125,
0.84228515625,
0.2374267578125
],
[
0.38232421875,
0.2279052734375,
0.476318359375,
0.2291259765625,
0.476318359375,
0.240478515625,
0.38232421875,
0.2392578125
],
[
0.09722900390625,
0.2291259765625,
0.136474609375,
0.2291259765625,
0.136474609375,
0.2392578125,
0.09722900390625,
0.2392578125
],
[
0.84228515625,
0.236083984375,
0.8759765625,
0.236083984375,
0.8759765625,
0.2474365234375,
0.84228515625,
0.2474365234375
],
[
0.5537109375,
0.24365234375,
0.68115234375,
0.24365234375,
0.68115234375,
0.304931640625,
0.5537109375,
0.304931640625
],
[
0.77783203125,
0.25,
0.8095703125,
0.25244140625,
0.8056640625,
0.268310546875,
0.77392578125,
0.26513671875
],
[
0.12744140625,
0.26513671875,
0.188720703125,
0.26513671875,
0.188720703125,
0.28857421875,
0.12744140625,
0.28857421875
],
[
0.28515625,
0.274658203125,
0.295654296875,
0.274658203125,
0.295654296875,
0.279052734375,
0.28515625,
0.279052734375
],
[
0.62353515625,
0.309326171875,
0.65673828125,
0.309326171875,
0.65673828125,
0.3212890625,
0.62353515625,
0.3212890625
],
[
0.69775390625,
0.31005859375,
0.71142578125,
0.306884765625,
0.7158203125,
0.318115234375,
0.70166015625,
0.3212890625
],
[
0.54736328125,
0.313232421875,
0.5537109375,
0.313232421875,
0.5537109375,
0.3193359375,
0.54736328125,
0.3193359375
],
[
0.313720703125,
0.314453125,
0.34716796875,
0.317626953125,
0.3447265625,
0.33203125,
0.310546875,
0.328857421875
],
[
0.51123046875,
0.313232421875,
0.52392578125,
0.3193359375,
0.51708984375,
0.328857421875,
0.50341796875,
0.322509765625
],
[
0.1641845703125,
0.31494140625,
0.193603515625,
0.317626953125,
0.19287109375,
0.328857421875,
0.16259765625,
0.326904296875
],
[
0.0882568359375,
0.31494140625,
0.134033203125,
0.31494140625,
0.134033203125,
0.330810546875,
0.0882568359375,
0.330810546875
],
[
0.4248046875,
0.317626953125,
0.43701171875,
0.317626953125,
0.43701171875,
0.32763671875,
0.4248046875,
0.32763671875
],
[
0.080078125,
0.35595703125,
0.8935546875,
0.35595703125,
0.8935546875,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.080078125,
0.369384765625,
0.89208984375,
0.369384765625,
0.89208984375,
0.38330078125,
0.080078125,
0.38330078125
],
[
0.080078125,
0.38330078125,
0.89208984375,
0.38330078125,
0.89208984375,
0.397705078125,
0.080078125,
0.397705078125
],
[
0.080078125,
0.397705078125,
0.4150390625,
0.397705078125,
0.4150390625,
0.410888671875,
0.080078125,
0.410888671875
],
[
0.080078125,
0.42041015625,
0.338134765625,
0.42041015625,
0.338134765625,
0.43505859375,
0.080078125,
0.43505859375
],
[
0.50244140625,
0.419189453125,
0.890625,
0.42041015625,
0.890625,
0.43505859375,
0.50244140625,
0.434326171875
],
[
0.5048828125,
0.436279296875,
0.890625,
0.436279296875,
0.890625,
0.449462890625,
0.5048828125,
0.449462890625
],
[
0.098876953125,
0.442626953125,
0.468017578125,
0.443115234375,
0.468017578125,
0.458251953125,
0.098876953125,
0.45703125
],
[
0.08087158203125,
0.455810546875,
0.471435546875,
0.458251953125,
0.471435546875,
0.47607421875,
0.08087158203125,
0.473388671875
],
[
0.50341796875,
0.451904296875,
0.890625,
0.451904296875,
0.890625,
0.46533203125,
0.50341796875,
0.46533203125
],
[
0.50244140625,
0.466552734375,
0.890625,
0.46533203125,
0.890625,
0.479736328125,
0.50244140625,
0.48095703125
],
[
0.08251953125,
0.474853515625,
0.468017578125,
0.474853515625,
0.468017578125,
0.488037109375,
0.08251953125,
0.488037109375
],
[
0.5048828125,
0.481689453125,
0.88818359375,
0.481689453125,
0.88818359375,
0.49560546875,
0.5048828125,
0.49560546875
],
[
0.080078125,
0.4892578125,
0.468017578125,
0.488037109375,
0.468017578125,
0.50244140625,
0.080078125,
0.50390625
],
[
0.5048828125,
0.496826171875,
0.890625,
0.496826171875,
0.890625,
0.51025390625,
0.5048828125,
0.51025390625
],
[
0.08087158203125,
0.5048828125,
0.1650390625,
0.50244140625,
0.1650390625,
0.517578125,
0.08087158203125,
0.51953125
],
[
0.50244140625,
0.5087890625,
0.89208984375,
0.5087890625,
0.89208984375,
0.5263671875,
0.50244140625,
0.5263671875
],
[
0.2254638671875,
0.52880859375,
0.240234375,
0.52880859375,
0.240234375,
0.53857421875,
0.2254638671875,
0.53857421875
],
[
0.5,
0.5234375,
0.8935546875,
0.52587890625,
0.8935546875,
0.546875,
0.5,
0.54443359375
],
[
0.146240234375,
0.5361328125,
0.40185546875,
0.5361328125,
0.40185546875,
0.55810546875,
0.146240234375,
0.55810546875
],
[
0.4453125,
0.54248046875,
0.468017578125,
0.54248046875,
0.468017578125,
0.55615234375,
0.4453125,
0.55615234375
],
[
0.50341796875,
0.54345703125,
0.890625,
0.54345703125,
0.890625,
0.55810546875,
0.50341796875,
0.55810546875
],
[
0.22216796875,
0.5595703125,
0.24267578125,
0.5595703125,
0.24267578125,
0.5693359375,
0.22216796875,
0.5693359375
],
[
0.50244140625,
0.55810546875,
0.74658203125,
0.55810546875,
0.74658203125,
0.5712890625,
0.50244140625,
0.5712890625
],
[
0.0906982421875,
0.58056640625,
0.468017578125,
0.58056640625,
0.468017578125,
0.59521484375,
0.0906982421875,
0.59521484375
],
[
0.640625,
0.583984375,
0.7490234375,
0.583984375,
0.7490234375,
0.5986328125,
0.640625,
0.5986328125
],
[
0.86669921875,
0.583984375,
0.8935546875,
0.583984375,
0.8935546875,
0.60107421875,
0.86669921875,
0.60107421875
],
[
0.080078125,
0.59326171875,
0.4697265625,
0.59228515625,
0.4697265625,
0.611328125,
0.080078125,
0.6123046875
],
[
0.07843017578125,
0.6123046875,
0.468017578125,
0.609375,
0.468017578125,
0.626953125,
0.07843017578125,
0.6298828125
],
[
0.50341796875,
0.61279296875,
0.7998046875,
0.61279296875,
0.7998046875,
0.626953125,
0.50341796875,
0.626953125
],
[
0.080078125,
0.626953125,
0.2091064453125,
0.62744140625,
0.2091064453125,
0.642578125,
0.080078125,
0.6416015625
],
[
0.5107421875,
0.6337890625,
0.88134765625,
0.63525390625,
0.88134765625,
0.65576171875,
0.5107421875,
0.65478515625
],
[
0.080078125,
0.6513671875,
0.46728515625,
0.6513671875,
0.46728515625,
0.66650390625,
0.080078125,
0.66650390625
],
[
0.5986328125,
0.6513671875,
0.63232421875,
0.6513671875,
0.63232421875,
0.66357421875,
0.5986328125,
0.66357421875
],
[
0.1192626953125,
0.66748046875,
0.1641845703125,
0.66748046875,
0.1641845703125,
0.681640625,
0.1192626953125,
0.681640625
],
[
0.86669921875,
0.66357421875,
0.8935546875,
0.66357421875,
0.8935546875,
0.6796875,
0.86669921875,
0.6796875
],
[
0.50634765625,
0.67822265625,
0.69775390625,
0.67822265625,
0.69775390625,
0.6923828125,
0.50634765625,
0.6923828125
],
[
0.1004638671875,
0.68994140625,
0.46728515625,
0.68994140625,
0.46728515625,
0.705078125,
0.1004638671875,
0.705078125
],
[
0.08087158203125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.50341796875,
0.7021484375,
0.8896484375,
0.7041015625,
0.8896484375,
0.7216796875,
0.50341796875,
0.7197265625
],
[
0.6005859375,
0.71728515625,
0.60888671875,
0.7236328125,
0.6005859375,
0.72998046875,
0.59228515625,
0.7236328125
],
[
0.64892578125,
0.72021484375,
0.6650390625,
0.72021484375,
0.6650390625,
0.72900390625,
0.64892578125,
0.72900390625
],
[
0.86669921875,
0.72998046875,
0.8935546875,
0.72998046875,
0.8935546875,
0.7470703125,
0.86669921875,
0.7470703125
],
[
0.08087158203125,
0.7216796875,
0.13232421875,
0.7216796875,
0.13232421875,
0.736328125,
0.08087158203125,
0.736328125
],
[
0.1282958984375,
0.74365234375,
0.42236328125,
0.744140625,
0.42236328125,
0.76318359375,
0.1282958984375,
0.76220703125
],
[
0.44677734375,
0.74853515625,
0.46728515625,
0.74853515625,
0.46728515625,
0.76123046875,
0.44677734375,
0.76123046875
],
[
0.2197265625,
0.76123046875,
0.233642578125,
0.76123046875,
0.233642578125,
0.76953125,
0.2197265625,
0.76953125
],
[
0.273681640625,
0.76123046875,
0.289306640625,
0.76123046875,
0.289306640625,
0.77099609375,
0.273681640625,
0.77099609375
],
[
0.5048828125,
0.75830078125,
0.890625,
0.75830078125,
0.890625,
0.771484375,
0.5048828125,
0.771484375
],
[
0.50341796875,
0.771484375,
0.89208984375,
0.771484375,
0.89208984375,
0.7861328125,
0.50341796875,
0.7861328125
],
[
0.0849609375,
0.7841796875,
0.468017578125,
0.7841796875,
0.468017578125,
0.79736328125,
0.0849609375,
0.79736328125
],
[
0.50341796875,
0.787109375,
0.89208984375,
0.787109375,
0.89208984375,
0.8017578125,
0.50341796875,
0.8017578125
],
[
0.08251953125,
0.80029296875,
0.451904296875,
0.80029296875,
0.451904296875,
0.814453125,
0.08251953125,
0.814453125
],
[
0.50341796875,
0.8017578125,
0.89208984375,
0.8017578125,
0.89208984375,
0.81640625,
0.50341796875,
0.81640625
],
[
0.50341796875,
0.8173828125,
0.69287109375,
0.8173828125,
0.69287109375,
0.83056640625,
0.50341796875,
0.83056640625
],
[
0.080078125,
0.830078125,
0.111083984375,
0.830078125,
0.111083984375,
0.84130859375,
0.080078125,
0.84130859375
],
[
0.11273193359375,
0.8291015625,
0.46728515625,
0.8291015625,
0.46728515625,
0.84326171875,
0.11273193359375,
0.84326171875
],
[
0.108642578125,
0.84521484375,
0.156005859375,
0.84521484375,
0.156005859375,
0.85986328125,
0.108642578125,
0.85986328125
],
[
0.5048828125,
0.84765625,
0.53515625,
0.84765625,
0.53515625,
0.859375,
0.5048828125,
0.859375
],
[
0.52783203125,
0.8466796875,
0.771484375,
0.8466796875,
0.771484375,
0.861328125,
0.52783203125,
0.861328125
],
[
0.101318359375,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.101318359375,
0.88525390625
],
[
0.52392578125,
0.8720703125,
0.8896484375,
0.8720703125,
0.8896484375,
0.88525390625,
0.52392578125,
0.88525390625
],
[
0.08087158203125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.65185546875,
0.88623046875,
0.65185546875,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"Spectral",
"LD",
"Reconstructed",
"Random Generator",
"Generated",
"Decoder",
"Image",
"Encoder",
"Image",
"Spectral Image",
"Input",
"LDImage",
"",
"",
"?",
"",
"Discriminator",
"0",
"G(z)",
"6()",
"a)())",
"E()",
"D(b(0))",
"z",
"O",
"Random Generator LD Image",
"",
"Spectral",
"Encoder",
"LD",
"Generated",
"Generated",
"Decoder",
"Image",
"Image",
"True",
"Spectral Image",
"Input",
"False",
"上",
"D()",
"",
"",
"E(a)",
"6",
"T",
"D(e)",
"b)",
"G(z)",
"z",
"a",
"Figure I. Autoencoder and LD-GAN networks.a) The AE network allows obtaining an LD representation of the SI",
"through an encoder subnetwork and then reconstructs the original SI with a decoder subnetwork. b) The LD representation",
"obtained by the AE is employed to train a GAN to generate new LD representation images. c) Once the GAN is trained,",
"the new LD images can be decoded to recover SIs.",
"3.1.Spectral Image Autoencoder",
"正则化方法，该方法允许在LD空间中紧凑地表示SI",
"数据集，从而提高了AE恢复性能。然后，我们在生成",
"使用AE网络获得具有K样本的SI数据集",
"[α()]，的LD表示。两个子网给出了AE的结构，",
"的LD空间上使用方差最大化进行GAN训练，以在",
"generated SI数据集上产生多样化的数据和更高的质",
"如图所示。1（a)。编码器网络E将压缩SI的spectral",
"量。此基于方差的正则化准则也已用于自监督学习8",
"信息，解码器网络D将解码基础SI。这些子网的优化",
"和稀疏编码[12]。首先，定义包含AE的LD表示的",
"由下式给出",
"集A={e()K1，并定义生成的LD图像集byGAN",
"K",
"asW=(d。A的方差由ER表示，而",
"[E, D) =I() - D(E((),",
"(3)",
"GAN生成的表示αwER4在训练批次维度上以像素",
"i=1",
"为单位。建议的统计正则化函数由",
"其中E和D分别是编码器和解码器。因此，SI的LD",
"R() = α ll2",
"(5)",
"表示可以作为6(）=E(a(）获得，其中b)RMNc",
"和c≤L。最后，可以通过将该LD图像解码为(i）=",
"给出，然后，正则化的AE训练如下执行",
"D(b(i))来恢复 SI",
"[E, D) = arg min Ila() - D(E(a()2 + μaeR(A).",
"3.2. Low-Dimensional Generative Adversarial Net-",
"E,D",
"work",
"(6)",
"，最终的GAN优化问题由",
"我们employ将针对a SI数据集的LD表示进行",
"优化的GAN。然后，来自方程式2的对抗性训练可以",
"[G, D) = arg min arg max Cgan(D, G, E) - Hgan R(W).",
"0",
"G",
"(7)",
"重写为",
"(G, D) = arg min arg max Cgan(D, g, E),",
"(4)",
"D",
"g",
"给出。注意，方差正则化函数中的负号是因为我们想要",
"最大化所生成的 SI的表示。在AE和 GAN这两个优",
"其中生成网络可以生成LD图像作为e=(z），并",
"化问题中，参数μae和μgan是正则化超参数，它们控",
"且可以使用解码器作为f=D(e)获得新的 SI样本。",
"制我们将AELD空间集中多少或我们将GAN生成",
"的图像的可变性增加多少。",
"4.",
"Statistical Regularization for the AE and",
"GAN",
"5.",
" Spectral Image Applications",
"为了改善AE的LD表示和GAN生成的LD图像",
"为了验证我们提出的方法的性能，我们解决了以下",
"的多样性，我们在AE训练中提出了一种方差最小化",
"基于 DL 的 SI 任务：",
"3"
] |
{
"arxiv_id": "2305.00132",
"language": "zh",
"page_idx": 2
}
|
||
2305.00132-zh-3
|
2305.00132-zh-3.jpg
|
CSIRecovery [7]: CSI 是一种技术，它允许通过高维数据 and 的编码投影，在单次拍摄[46]中使用 2D 传感器感测 SI 的整个 3D 信息，这是一种恢复底层 SI 的计算算法。已经提出了几种用于 CSI 的光学架构。我们采用了编码孔径快照光谱成像仪(CASSI)[40],它是 CSI [4]中的旗舰体系结构。这种感测现象在矩阵向量 乘 法 中 表 示 为 $\boldsymbol{y}= \boldsymbol{H}\boldsymbol{x}+ \boldsymbol{n}$ , 其 中 $\boldsymbol{y}\in \mathbb{R} ^{M( N+ L- 1) }$ 是 压 缩 的 测 量 值 , $\boldsymbol{H}\in \mathbb{R} ^{M( N+ L- 1) \times MNL}$ 是 CASSI 感 测 矩 阵 , $x\in \mathbb{R} ^{MNL}$ 是 SI 的 向 量 化 $, \boldsymbol{n}\in \mathbb{R} ^{M( N+ L- 1) }$ 是来自采集过程的加性噪声
单幅图像空间光谱图像超分辨率：SI 的一项众所周知的任务是从低空间光谱分辨率 SI [44] 恢复高空间光谱分辨率 SI。在此任务中，可以通过抽取矩阵$D\in\mathbb{R}^{\frac{MN}{k_s}\left(\frac L{k_l}\right)\times MNL}$在空间和光谱上对 SI 进行下采样，其中$k_s$和$k_l$分别表示 SI 的空间和光谱分辨率的抽取因子。低空间光谱分辨率图像表示为$y=Dx+n$
从 RGB 图像恢复 SI: 另一个引起研究界极大关注的任务是从 RGB 图像到 SI [3] 的映射。此任务包括从具有较少光谱信息的计数的 RGB 图像中恢复 SI,并考虑已知的光谱响应函数 $\boldsymbol{R}\in\mathbb{R}^{MN3\times MNL}$。然后， RGB 图像可以表示为$y=Rx+n$
Since the mentioned tasks are ill-posed problems, the objective is to recover the SI $x$ from the observed data $y=Ax$, ,其中$A$可以根据所选问题表示传感矩阵$H$ ,抽取矩阵$D$或光谱响应函数$R$ 。然后，我们可以通过优化问题
$$\hat{\boldsymbol{\theta}}=\arg\min_{\boldsymbol{\theta}}\frac{1}{K}\sum_{k=1}^{K}\|\boldsymbol{x}^{(k)}-\mathcal{M}_{\boldsymbol{\theta}}(\boldsymbol{y}^{(k)})\|_{2}^{2},$$
(8)
These are ill-posed inverse problems that are challenging. Therefore, we aim to increase the performance of the network $\mathcal{M}_\theta$ for each case by adding synthetic samples generated by the LD-GAN. 来解决基于 DL 的 SI 计算任务
## 6. Numerical Experiments
在本节中，我们对 LD 表示具有不同通道的 AE 网络进行了几次实验。对于 GAN 架构，深度卷积生成对抗网络 [36] 适应于所采用数据集的空间大小，并且针对从 AE 网络获得的 LD 图像数据集对整个尺寸的 SI 数据集进行实验。所有实验均在带有 NVIDIARTX3090图形卡的 GPU 上进行。用于所有实验的数据集是经过预处理的 ARAD1K [3] ,将空间分辨率降低到$256\times256$并保持 31 光谱带。This dataset contains 900 samples for training and 50 samples for testing. 我们提取具有$128\times128$空间分辨率的独特补丁，总共获得 3600 个用于训练的补丁和 200 个用于测试的补丁。Recovery performance is measured with the peak-noise-to-signal ratio (PSNR) and the structural similarity index measure (SSIM) [18] .
# 6.1. AE Experiments
SI 的压缩仅在频谱维度上执行，其中评估了编码器子网的输出通道数$c=\{1,3,8,16\}$的几个值。The encoder and subnetwork consist of 7 convolutional layers with ReLU activations, which decrease the number of feature channels from 16c to $c$ and the decoder network has also 7 layers increasing the number of features from $c$ to 16$c$ and a final convolutional layer with $L=31$ features to match the SI spectral dimension 表 1 显示，对于评估的通道数，$c=3$具有最佳性能。因此，当应用所提出的方法时，对于 SI 应用程序的性能，将高度考虑此压缩级别
<table>
<tbody>
<tr>
<th>Channels</th>
<th>$PSNR$ [dB$]\uparrow$</th>
<th>SSIM</th>
</tr>
<tr>
<td>1</td>
<td>33.72</td>
<td>0.9513</td>
</tr>
<tr>
<td>3</td>
<td>39.32</td>
<td>0.9739</td>
</tr>
<tr>
<td>8</td>
<td>38.37</td>
<td>0.9722</td>
</tr>
<tr>
<td>16</td>
<td>38.31</td>
<td>0.9679</td>
</tr>
</tbody>
</table>
Table 1. AE recovery performance for different latent chan-
nel sizes. The highlighted green values are the best perfor-
mance and the blue ones are the second best.
## 6.2. GAN Experiments
我们使用 LD images from SI 数据集，使用上面提到的 AE 来优化 GAN。所有 GAN 实验均经过 50个时期的训练，鉴别器和生成器的学习速率均为$2e^{-4}$ 生成器和鉴别器网络采用 2D 转置卷积和 2D 卷积
层 , respectively. Each subnetwork counts with batch normalization, and Leaky ReLU activations, after the convolutional layer, following the implementation given by [36] 。GAN 的所有实验都采用随机输入$z\in\mathbb{R}^100\sim$ $\mathcal{N}(\mathbf{0},\mathbf{1})$。InFig。$\mathsf{E}$2,显示了具有$c=\{3,8\}$的 LDGAN,具有统计正则化的 LD-GAN 和 S-GAN 的一些生成的 LD 样本及其各自的 SI。为了比较所提出方法的有效性，我们采用所采用的 GAN 直接生成没有LD 表示的 SI, 表示为 S-GAN。InFig。3,显示了不同 channels $c$的鉴别器损失函数。当在 LD image dataset 上训练 GAN 时，性能正确显示此行为。This performance is far from the theoretical optimum when training is done with S-GAN.
我们使用建议的 LD-GAN 分析了生成的光谱信息， 并与 S-GAN 进行了比较。为此，我们采用了广泛使用的 SI 分析技术，即线性混合模型 [11] 。该模型指出， SI 的给定像素的光谱特征由$q$最具代表性的 SI 特征的线性组合表示。这些特征表示为端成员，线性模型的系数称为丰度。为了提取 SI ,we employ 的端元和丰度，高光谱顶点分量分析(HyperVCA)[31] algorithm Specifically ,我们使用 HyperVCA 从原始 ARAD 数据集中，从 LD-GAN 和 S-GAN 生成的数据集中提取$q=4$末端成员 ,如图. 4 所示 ,平均为 100 张图像。所获得的结果表明，所提出的 LD-GAN 的生成的平均端元与原始数据集非常相似。此结果可以得出结论，生成的光谱数据遵循与原始数据集类似的行为。
4
|
[
[
0.1004638671875,
0.091552734375,
0.46728515625,
0.091552734375,
0.46728515625,
0.10479736328125,
0.1004638671875,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.6689453125,
0.0921630859375,
0.6689453125,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.521484375,
0.11614990234375,
0.8896484375,
0.11614990234375,
0.8896484375,
0.1300048828125,
0.521484375,
0.1300048828125
],
[
0.08416748046875,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08416748046875,
0.1351318359375
],
[
0.08087158203125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.5048828125,
0.1331787109375,
0.8896484375,
0.1331787109375,
0.8896484375,
0.143310546875,
0.5048828125,
0.143310546875
],
[
0.50341796875,
0.1458740234375,
0.890625,
0.1458740234375,
0.890625,
0.160400390625,
0.50341796875,
0.160400390625
],
[
0.08251953125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.08251953125,
0.1666259765625
],
[
0.50341796875,
0.1610107421875,
0.890625,
0.1610107421875,
0.890625,
0.1749267578125,
0.50341796875,
0.1749267578125
],
[
0.08251953125,
0.1673583984375,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.08251953125,
0.1817626953125
],
[
0.080078125,
0.179931640625,
0.471435546875,
0.1767578125,
0.471435546875,
0.1976318359375,
0.080078125,
0.2008056640625
],
[
0.5009765625,
0.175537109375,
0.890625,
0.175537109375,
0.890625,
0.189453125,
0.5009765625,
0.189453125
],
[
0.080078125,
0.1925048828125,
0.471435546875,
0.1925048828125,
0.471435546875,
0.21337890625,
0.080078125,
0.21337890625
],
[
0.50341796875,
0.1925048828125,
0.890625,
0.1925048828125,
0.890625,
0.205810546875,
0.50341796875,
0.205810546875
],
[
0.08087158203125,
0.210205078125,
0.468017578125,
0.210205078125,
0.468017578125,
0.2279052734375,
0.08087158203125,
0.2279052734375
],
[
0.50341796875,
0.205810546875,
0.890625,
0.20703125,
0.890625,
0.2203369140625,
0.50341796875,
0.2197265625
],
[
0.08087158203125,
0.2279052734375,
0.27783203125,
0.2279052734375,
0.27783203125,
0.2412109375,
0.08087158203125,
0.2412109375
],
[
0.5048828125,
0.222900390625,
0.87353515625,
0.222900390625,
0.87353515625,
0.236083984375,
0.5048828125,
0.236083984375
],
[
0.50341796875,
0.2374267578125,
0.890625,
0.2374267578125,
0.890625,
0.250732421875,
0.50341796875,
0.250732421875
],
[
0.101318359375,
0.2442626953125,
0.468017578125,
0.24365234375,
0.468017578125,
0.25830078125,
0.101318359375,
0.2587890625
],
[
0.5048828125,
0.25244140625,
0.890625,
0.25244140625,
0.890625,
0.266357421875,
0.5048828125,
0.266357421875
],
[
0.08251953125,
0.261962890625,
0.46728515625,
0.261962890625,
0.46728515625,
0.272705078125,
0.08251953125,
0.272705078125
],
[
0.50244140625,
0.266357421875,
0.701171875,
0.267578125,
0.701171875,
0.2822265625,
0.50244140625,
0.281005859375
],
[
0.08087158203125,
0.27587890625,
0.468017578125,
0.27587890625,
0.468017578125,
0.2890625,
0.08087158203125,
0.2890625
],
[
0.0767822265625,
0.2841796875,
0.47216796875,
0.2890625,
0.47216796875,
0.31494140625,
0.0767822265625,
0.31005859375
],
[
0.5986328125,
0.298583984375,
0.78759765625,
0.298583984375,
0.78759765625,
0.3125,
0.5986328125,
0.3125
],
[
0.080078125,
0.31005859375,
0.468017578125,
0.311279296875,
0.468017578125,
0.32568359375,
0.080078125,
0.324462890625
],
[
0.6845703125,
0.31494140625,
0.7158203125,
0.31494140625,
0.7158203125,
0.32763671875,
0.6845703125,
0.32763671875
],
[
0.744140625,
0.31494140625,
0.7861328125,
0.31494140625,
0.7861328125,
0.326904296875,
0.744140625,
0.326904296875
],
[
0.08087158203125,
0.32568359375,
0.46728515625,
0.32568359375,
0.46728515625,
0.339111328125,
0.08087158203125,
0.339111328125
],
[
0.62158203125,
0.31640625,
0.6337890625,
0.31640625,
0.6337890625,
0.324462890625,
0.62158203125,
0.324462890625
],
[
0.6845703125,
0.32568359375,
0.7138671875,
0.32568359375,
0.7138671875,
0.33837890625,
0.6845703125,
0.33837890625
],
[
0.74658203125,
0.326904296875,
0.7861328125,
0.326904296875,
0.7861328125,
0.33837890625,
0.74658203125,
0.33837890625
],
[
0.6845703125,
0.335205078125,
0.71337890625,
0.337158203125,
0.71142578125,
0.349853515625,
0.68310546875,
0.347900390625
],
[
0.62158203125,
0.328857421875,
0.6337890625,
0.328857421875,
0.6337890625,
0.337158203125,
0.62158203125,
0.337158203125
],
[
0.6259765625,
0.33837890625,
0.63134765625,
0.341552734375,
0.62841796875,
0.34521484375,
0.62158203125,
0.343505859375
],
[
0.74609375,
0.337158203125,
0.78759765625,
0.337158203125,
0.78759765625,
0.348388671875,
0.74609375,
0.348388671875
],
[
0.101318359375,
0.343505859375,
0.468017578125,
0.343505859375,
0.468017578125,
0.356689453125,
0.101318359375,
0.356689453125
],
[
0.6845703125,
0.347900390625,
0.71337890625,
0.347900390625,
0.71337890625,
0.359130859375,
0.6845703125,
0.359130859375
],
[
0.74609375,
0.347900390625,
0.78515625,
0.347900390625,
0.78515625,
0.359130859375,
0.74609375,
0.359130859375
],
[
0.08087158203125,
0.359130859375,
0.46728515625,
0.359130859375,
0.46728515625,
0.37255859375,
0.08087158203125,
0.37255859375
],
[
0.619140625,
0.35107421875,
0.63232421875,
0.3466796875,
0.63623046875,
0.35595703125,
0.62353515625,
0.359130859375
],
[
0.50244140625,
0.363037109375,
0.88818359375,
0.363037109375,
0.88818359375,
0.37744140625,
0.50244140625,
0.37744140625
],
[
0.08251953125,
0.375,
0.46728515625,
0.375,
0.46728515625,
0.38818359375,
0.08251953125,
0.38818359375
],
[
0.50341796875,
0.376953125,
0.88623046875,
0.37744140625,
0.88623046875,
0.391357421875,
0.50341796875,
0.39013671875
],
[
0.0767822265625,
0.38330078125,
0.47216796875,
0.3837890625,
0.47216796875,
0.406005859375,
0.0767822265625,
0.40478515625
],
[
0.50244140625,
0.391357421875,
0.8056640625,
0.39013671875,
0.8056640625,
0.404052734375,
0.50244140625,
0.40478515625
],
[
0.08087158203125,
0.40478515625,
0.33251953125,
0.40478515625,
0.33251953125,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.098876953125,
0.421630859375,
0.46728515625,
0.421630859375,
0.46728515625,
0.43505859375,
0.098876953125,
0.43505859375
],
[
0.50244140625,
0.423583984375,
0.68310546875,
0.423583984375,
0.68310546875,
0.4375,
0.50244140625,
0.4375
],
[
0.08251953125,
0.438232421875,
0.46728515625,
0.438232421875,
0.46728515625,
0.448974609375,
0.08251953125,
0.448974609375
],
[
0.08087158203125,
0.451904296875,
0.468017578125,
0.451904296875,
0.468017578125,
0.46533203125,
0.08087158203125,
0.46533203125
],
[
0.52294921875,
0.4462890625,
0.890625,
0.4462890625,
0.890625,
0.4609375,
0.52294921875,
0.4609375
],
[
0.08087158203125,
0.46728515625,
0.468017578125,
0.46728515625,
0.468017578125,
0.48095703125,
0.08087158203125,
0.48095703125
],
[
0.50341796875,
0.46337890625,
0.89208984375,
0.46337890625,
0.89208984375,
0.4765625,
0.50341796875,
0.4765625
],
[
0.08087158203125,
0.481689453125,
0.2279052734375,
0.481689453125,
0.2279052734375,
0.49560546875,
0.08087158203125,
0.49560546875
],
[
0.5048828125,
0.477783203125,
0.8896484375,
0.477783203125,
0.8896484375,
0.4912109375,
0.5048828125,
0.4912109375
],
[
0.5048828125,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.5048828125,
0.5068359375
],
[
0.2156982421875,
0.505859375,
0.4248046875,
0.51953125,
0.420654296875,
0.55615234375,
0.2115478515625,
0.54248046875
],
[
0.2459716796875,
0.51220703125,
0.2607421875,
0.51220703125,
0.2607421875,
0.52099609375,
0.2459716796875,
0.52099609375
],
[
0.5048828125,
0.5087890625,
0.890625,
0.5087890625,
0.890625,
0.52294921875,
0.5048828125,
0.52294921875
],
[
0.125,
0.5166015625,
0.2607421875,
0.521484375,
0.25732421875,
0.55810546875,
0.1217041015625,
0.55322265625
],
[
0.44287109375,
0.5263671875,
0.465576171875,
0.5263671875,
0.465576171875,
0.53857421875,
0.44287109375,
0.53857421875
],
[
0.50634765625,
0.52587890625,
0.8896484375,
0.52587890625,
0.8896484375,
0.5361328125,
0.50634765625,
0.5361328125
],
[
0.4150390625,
0.53076171875,
0.439453125,
0.52978515625,
0.439453125,
0.5341796875,
0.4150390625,
0.53515625
],
[
0.410888671875,
0.52978515625,
0.4150390625,
0.52978515625,
0.4150390625,
0.53515625,
0.410888671875,
0.53515625
],
[
0.2303466796875,
0.5361328125,
0.26708984375,
0.54345703125,
0.26318359375,
0.55615234375,
0.226318359375,
0.548828125
],
[
0.5048828125,
0.53857421875,
0.890625,
0.53857421875,
0.890625,
0.5517578125,
0.5048828125,
0.5517578125
],
[
0.5009765625,
0.5517578125,
0.89208984375,
0.5498046875,
0.89208984375,
0.5673828125,
0.5009765625,
0.5693359375
],
[
0.0849609375,
0.56640625,
0.46728515625,
0.5673828125,
0.46728515625,
0.58203125,
0.0849609375,
0.58056640625
],
[
0.50341796875,
0.568359375,
0.88818359375,
0.568359375,
0.88818359375,
0.58203125,
0.50341796875,
0.58203125
],
[
0.08251953125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.08251953125,
0.5966796875
],
[
0.50244140625,
0.58203125,
0.88232421875,
0.58349609375,
0.88232421875,
0.59765625,
0.50244140625,
0.5966796875
],
[
0.08087158203125,
0.5986328125,
0.464111328125,
0.5986328125,
0.464111328125,
0.6123046875,
0.08087158203125,
0.6123046875
],
[
0.50341796875,
0.59765625,
0.88818359375,
0.5986328125,
0.88818359375,
0.6123046875,
0.50341796875,
0.611328125
],
[
0.08087158203125,
0.61279296875,
0.468017578125,
0.61279296875,
0.468017578125,
0.62744140625,
0.08087158203125,
0.62744140625
],
[
0.5048828125,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.5048828125,
0.62744140625
],
[
0.08087158203125,
0.62890625,
0.1478271484375,
0.62890625,
0.1478271484375,
0.64306640625,
0.08087158203125,
0.64306640625
],
[
0.50341796875,
0.62744140625,
0.8896484375,
0.62890625,
0.8896484375,
0.64306640625,
0.50341796875,
0.642578125
],
[
0.5048828125,
0.64306640625,
0.88818359375,
0.64306640625,
0.88818359375,
0.65771484375,
0.5048828125,
0.65771484375
],
[
0.08087158203125,
0.66357421875,
0.3095703125,
0.66357421875,
0.3095703125,
0.67822265625,
0.08087158203125,
0.67822265625
],
[
0.50341796875,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50341796875,
0.673828125
],
[
0.50341796875,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.101318359375,
0.6923828125,
0.465576171875,
0.6923828125,
0.465576171875,
0.70263671875,
0.101318359375,
0.70263671875
],
[
0.5048828125,
0.69140625,
0.7158203125,
0.69140625,
0.7158203125,
0.7021484375,
0.5048828125,
0.7021484375
],
[
0.08087158203125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.521484375,
0.705078125,
0.88818359375,
0.70556640625,
0.88818359375,
0.72021484375,
0.521484375,
0.7197265625
],
[
0.08087158203125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.08087158203125,
0.734375
],
[
0.5048828125,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.5048828125,
0.734375
],
[
0.08251953125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08251953125,
0.74755859375
],
[
0.50341796875,
0.736328125,
0.88818359375,
0.736328125,
0.88818359375,
0.75,
0.50341796875,
0.75
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75,
0.468017578125,
0.7646484375,
0.080078125,
0.76513671875
],
[
0.50244140625,
0.75,
0.890625,
0.75048828125,
0.890625,
0.76513671875,
0.50244140625,
0.7646484375
],
[
0.08251953125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.08251953125,
0.77978515625
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.78076171875,
0.46728515625,
0.78076171875,
0.46728515625,
0.79443359375,
0.08087158203125,
0.79443359375
],
[
0.50634765625,
0.78271484375,
0.88818359375,
0.78271484375,
0.88818359375,
0.79345703125,
0.50634765625,
0.79345703125
],
[
0.08087158203125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.8896484375,
0.79443359375,
0.8896484375,
0.8095703125,
0.50244140625,
0.81103515625
],
[
0.08087158203125,
0.81103515625,
0.463134765625,
0.81103515625,
0.463134765625,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.5107421875,
0.81640625,
0.521484375,
0.81640625,
0.521484375,
0.82080078125,
0.5107421875,
0.82080078125
],
[
0.52294921875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.52294921875,
0.82470703125
],
[
0.08087158203125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50244140625,
0.82568359375,
0.8896484375,
0.82568359375,
0.8896484375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84033203125,
0.468017578125,
0.84130859375,
0.468017578125,
0.85595703125,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.08087158203125,
0.85693359375,
0.46728515625,
0.85693359375,
0.46728515625,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.50341796875,
0.85595703125,
0.89208984375,
0.85595703125,
0.89208984375,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50341796875,
0.87060546875,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88525390625,
0.35302734375,
0.88720703125,
0.35302734375,
0.9013671875,
0.080078125,
0.8994140625
],
[
0.50341796875,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"CSIRecovery「7T：CSI是一种技术，它允许通过高",
"6.1. AE Experiments",
"维数据and的编码投影，在单次拍摄【46]中使用2D",
"SI的压缩仅在频谱维度上执行，其中评估了编码",
"传感器感测SI的整个3D信息，这是一种恢复底层SI",
"的计算算法。已经提出了几种用于CSI的光学架构。我",
"器子网的输出通道数c=1.3.8.161的个值。The",
"encoder and subnetwork consist of 7 convolutional lay-",
"们采用了编码孔径快照光谱成像仪（CASSI）「401，它",
"ers with ReLU activations, which decrease the number",
"是CSI【4]】中的旗舰体系结构。这种感测现象在矩阵向",
"量乘法中表示为y=Ha+n，其中ueRM(N+z-1)",
"of feature channels from 16c to c and the decoder",
"是压缩的测量值，HERM(N+L-1)xMNL是CASSI感",
"network has also 7 layers increasing the number of fea-",
"测矩阵，ERMNL是SI的向量化，nERM(N+L-1)",
"tures from c to16c and a final convolutional layer with",
"是来自采集过程的加性噪声",
"L = 31 features to match the SI spectral dimension",
"表1显示，对于评估的通道数，C=3具有最佳性能。",
"单幅图像空间光谱图像超分辨率：SI的一项众所",
"因此，当应用所提出的方法时，对于SI应用程序的性",
"周知的任务是从低空间光谱分辨率SI441恢复高空",
"能，将高度考虑此压缩级别",
"间光谱分辨率SI。在此任务中，可以通过抽取矩阵",
"DER(右)XMNL在空间和光谱上对 SI进行下采",
"ChannelsPSNR [dB] + SSIM +",
"样，其中k。和ki分别表示SI的空间和光谱分辨率的",
"33.72",
"0.9513",
"抽取因子。低空间光谱分辨率图像表示为y=Da+n",
"1",
"39.32",
"0.9739",
"38.37",
"3",
"S",
"0.9722",
"从RGB图像恢复SI：另一个引起研究界极大关注",
"38.31",
"0.9679",
"的任务是从RGB图像到SI31的映射。此任务包括从",
"16",
"Table 1.AE recovery performance for different latent chan",
"具有较少光谱信息的计数的RGB图像中恢复SI，并",
"nel sizes. The highlighted green values are the best perfor",
"考虑已知的光谱响应函数RERMN3xMNL。然后，",
"mance and the blue ones are the second best.",
"RGB图像可以表示为y=Rac+n",
"Since the mentioned tasks are ill-posed problems,",
"6.2. GAN Experiments",
"the objective isto recover the S from the observed",
"datay=Aa，，其中A可以根据所选问题表示传感",
"我们使用LDimagesfromSI数据集，使用上面",
"矩阵H，抽取矩阵D或光谱响应函数R。然后，我",
"提到的AE来优化GAN。所有GAN实验均经过50",
"们可以通过优化问题",
"个时期的训练，鉴别器和生成器的学习速率均为2e-4",
"。生成器和鉴别器网络采用2D转置卷积和2D卷积",
"2 1() - o(g(0)1,",
"K",
"层, respectively. Each subnetwork counts with batch",
"=angin",
"(8)",
"normalization,and LeakyReLU activations,afterthe",
"",
":+",
"Kk=1",
"convolutional layer, following the implementation given",
"by[36]】。GAN的所有实验都采用随机输人zER100～",
"These are ill-posed inverse problems that are challeng-",
"N(0,1)。InFig。E2，显示了具有c={3,8)的LD",
"ing. Therefore, we aim to increase the performance of",
"GAN，具有统计正则化的LD-GAN和S-GAN的",
"the network Me for each case by adding synthetic sam",
"些生成的LD样本及其各自的SI。为了比较所提出方",
"ples generated bytheLD-GAN.来解决基于DL的 SI",
"法的有效性，我们采用所采用的GAN直接生成没有",
"计算任务",
"LD表示的SI，表示为S-GAN。InFig。3，显示了",
"不同 channels c的鉴别器损失函数。当在LD image",
"6. Numerical Experiments",
"dataset上训练GAN时，性能正确显示此行为。This",
"performance is far from the theoretical optimum when",
"在本节中，我们对LD表示具有不同通道的AE网络",
"training is done with S-GAN.",
"进行了几次实验。对于GAN架构，深度卷积生成对抗",
"我们使用建议的LD-GAN分析了生成的光谱信息",
"网络[36]适应于所采用数据集的空间大小，并且针对",
"并与S-GAN进行了比较。为此，我们采用了广泛使用",
"从AE网络获得的LD图像数据集对整个尺寸的SI数",
"的SI分析技术，即线性混合模型[11]。该模型指出",
"据集进行实验。所有实验均在带有NVIDIARTX3090",
"SI 的给定像素的光谱特征由 q最具代表性的SI特征",
"图形卡的GPU上进行。用于所有实验的数据集是",
"的线性组合表示。这些特征表示为端成员，线性模型的",
"经过预处理的ARAD1K31，将空间分辨率降低到",
"系数称为丰度。为了提取SI，weemploy的端元和丰",
"256×256并保持31光谱带。This dataset contains",
"度，高光谱顶点分量分析（HyperVCA）[31】algorithm",
"900 samples for training and 50 samples for testing",
"",
"Specifically，我们使用HyperVCA从原始ARAD",
"我们提取具有128×128空间分辨率的独特补丁，总",
"数据集中，从LD-GAN和S-GAN生成的数据集中提",
"共获得3600个用于训练的补丁和200个用于测试",
"取9=4末端成员，如图4所示，平均为100张",
"的补丁。 Recovery performance is measured with the",
"图像。所获得的结果表明，所提出的LD-GAN的生成",
"peak-noise-to-signal ratio (PSNR) and the structural",
"的平均端元与原始数据集非常相似。此结果可以得出",
"similarity index measure (SSIM) [18] ",
"结论，生成的光谱数据遵循与原始数据集类似的行为。",
"4"
] |
{
"arxiv_id": "2305.00132",
"language": "zh",
"page_idx": 3
}
|
||
2305.00132-zh-4
|
2305.00132-zh-4.jpg
|
<FigureHere>
Figure 2. Generated images for an LD representation with a) 3 bands and b) 8 bands. c) Generated SI samples from LD representation employing the statistical regularizer with $\mu_{gan}=\mu_{ae}=0.001.$ For a-c) the top images refer to the generated LD images and the bottom images are the obtained SI by decoding the LD representation. d) Generated SIs without an LD representation. The bottom row represents a false RGB mapping of generated SIs from their LD representations, except for S- GAN, which SIs is directly generated by the GAN.
虽然 S-GAN 在频谱上有许多伪影，但这可能会对基于
DL 的 SI 应用的 DA 策略的网络训练产生负面影响
training and validation, the spectral response provided by the ARAD dataset [3] was used . For the training of this network, the learning rate decays exponentially to achieve a more stable convergence.
## 6.3. DA Applications
我们通过生成合成 SI 数据集并将其用作基于 DL, CSI, SSSR 和 RGB 的光谱任务的 DA 来验证所提出的方法。以下所有网络均经过 100 个时期的训练，学习率为 1$e^{-3}$。根据获得的结果，我们报告了每个实验的测试图像的最佳性能。
$\bullet$ SI super-resolution setting: We employed a UNET ${\mathrm{network~with~4~downsampling~and~4~upsam-}}$ pling convolutional blocks to reconstruct the highresolution SI from low-spatial low-spectral measurements. The selection of the type of architecture is based on recent SSSR works [42] . The spatial and spectral decimator factors are set as $k_s= 2, k_l= 4$ .
$\bullet$ CSI recovery setting: The unrolled network deep
spatial-spectral prior [41] was used for CSI recovery. The measurements were simulated through the CASSI system using a random CA with 50 % of transmittance.
# 6.4. Quantitative Results
$\bullet$ SI recovery from RGB images Setting: A variation
of the UNET architecture described for the SSSR case is employed. UNET-based models have been previously proposed for RGB to spectral task [44] . To obtain the RGB input image for the network
对于每个 SI 应用程序，我们使用 20% , 50% 和$100\%$新样本考虑 LD-GAN 和 S-GAN 生成的每个 SI 数据集执行恢复任务，(百分比基于原始数据集的样本数)。如图所示。$\color{red}{\operatorname*{E}}\color{red}{5}$, LD-GAN 与$c=3$实现了更高的 PSNR 性能。此外，当进行 S-GAN 和 LD-GAN 之间的比较时，我们可以确定不仅 LD-GAN 更好，而且S-GAN 的性能低于基线 in several cases 。我们还观察到，使用 DA 生成的数据数量越多，所提出的方法对 SI 应用程序的性能的提高就越高
<FigureHere>
## 6.5. Qualitative Results
从测试数据集重建的 SI 的视觉结果如图。6,其中大量的新样本允许在 PSNR 方面有更高的性能。从地面实况 SI 及其重建中提取的光谱特征表明，使用LDSI 图像进行的训练实现了与地面实况数据更相似的密度。结果表明，使用建议数据集中的合成样本训练的模型在每个任务中都具有更高的重建质量
Figure 3. Discriminator loss function convergence for different channels. $L=31$ represents the performance by S- GAN.
5
|
[
[
0.35546875,
0.091552734375,
0.418212890625,
0.091552734375,
0.418212890625,
0.10162353515625,
0.35546875,
0.10162353515625
],
[
0.76708984375,
0.09027099609375,
0.81884765625,
0.09027099609375,
0.81884765625,
0.10162353515625,
0.76708984375,
0.10162353515625
],
[
0.08416748046875,
0.111083984375,
0.10296630859375,
0.111083984375,
0.1004638671875,
0.1654052734375,
0.08087158203125,
0.1654052734375
],
[
0.08660888671875,
0.1749267578125,
0.101318359375,
0.1749267578125,
0.1004638671875,
0.2310791015625,
0.0849609375,
0.2310791015625
],
[
0.69140625,
0.1944580078125,
0.6953125,
0.1944580078125,
0.6953125,
0.199462890625,
0.69140625,
0.199462890625
],
[
0.5087890625,
0.246826171875,
0.650390625,
0.246826171875,
0.650390625,
0.260009765625,
0.5087890625,
0.260009765625
],
[
0.7607421875,
0.246826171875,
0.826171875,
0.246826171875,
0.826171875,
0.261474609375,
0.7607421875,
0.261474609375
],
[
0.361083984375,
0.2474365234375,
0.4150390625,
0.2474365234375,
0.4150390625,
0.2587890625,
0.361083984375,
0.2587890625
],
[
0.1666259765625,
0.248779296875,
0.22216796875,
0.248779296875,
0.22216796875,
0.2587890625,
0.1666259765625,
0.2587890625
],
[
0.08087158203125,
0.2734375,
0.890625,
0.2734375,
0.890625,
0.287353515625,
0.08087158203125,
0.287353515625
],
[
0.080078125,
0.28857421875,
0.89208984375,
0.28857421875,
0.89208984375,
0.302978515625,
0.080078125,
0.302978515625
],
[
0.080078125,
0.3017578125,
0.890625,
0.3017578125,
0.890625,
0.31494140625,
0.080078125,
0.31494140625
],
[
0.080078125,
0.317626953125,
0.188720703125,
0.31640625,
0.188720703125,
0.32763671875,
0.080078125,
0.328857421875
],
[
0.1846923828125,
0.31494140625,
0.890625,
0.31494140625,
0.890625,
0.328857421875,
0.1846923828125,
0.328857421875
],
[
0.080078125,
0.330078125,
0.453369140625,
0.330078125,
0.453369140625,
0.343505859375,
0.080078125,
0.343505859375
],
[
0.08251953125,
0.352783203125,
0.46728515625,
0.352783203125,
0.46728515625,
0.3662109375,
0.08251953125,
0.3662109375
],
[
0.53369140625,
0.3515625,
0.890625,
0.352783203125,
0.890625,
0.367431640625,
0.53369140625,
0.3662109375
],
[
0.080078125,
0.3662109375,
0.44775390625,
0.367431640625,
0.44775390625,
0.3818359375,
0.080078125,
0.380615234375
],
[
0.5361328125,
0.36865234375,
0.89208984375,
0.36865234375,
0.89208984375,
0.3818359375,
0.5361328125,
0.3818359375
],
[
0.53759765625,
0.3837890625,
0.8896484375,
0.3837890625,
0.8896484375,
0.39453125,
0.53759765625,
0.39453125
],
[
0.080078125,
0.393310546875,
0.2484130859375,
0.39453125,
0.2484130859375,
0.4091796875,
0.080078125,
0.40771484375
],
[
0.53515625,
0.3984375,
0.89208984375,
0.3984375,
0.89208984375,
0.412353515625,
0.53515625,
0.412353515625
],
[
0.101318359375,
0.41845703125,
0.468017578125,
0.41845703125,
0.468017578125,
0.431884765625,
0.101318359375,
0.431884765625
],
[
0.5361328125,
0.415283203125,
0.60302734375,
0.415283203125,
0.60302734375,
0.4267578125,
0.5361328125,
0.4267578125
],
[
0.080078125,
0.431884765625,
0.46728515625,
0.43310546875,
0.46728515625,
0.447509765625,
0.080078125,
0.4462890625
],
[
0.08087158203125,
0.448974609375,
0.46728515625,
0.448974609375,
0.46728515625,
0.462158203125,
0.08087158203125,
0.462158203125
],
[
0.521484375,
0.460205078125,
0.5361328125,
0.458251953125,
0.53759765625,
0.462158203125,
0.52294921875,
0.464111328125
],
[
0.53369140625,
0.455078125,
0.890625,
0.455078125,
0.890625,
0.4697265625,
0.53369140625,
0.4697265625
],
[
0.08251953125,
0.46337890625,
0.468017578125,
0.46337890625,
0.468017578125,
0.477783203125,
0.08251953125,
0.477783203125
],
[
0.53515625,
0.4697265625,
0.890625,
0.47021484375,
0.890625,
0.48486328125,
0.53515625,
0.484130859375
],
[
0.08087158203125,
0.479248046875,
0.253173828125,
0.479248046875,
0.253173828125,
0.492431640625,
0.08087158203125,
0.492431640625
],
[
0.5361328125,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.5361328125,
0.498779296875
],
[
0.09478759765625,
0.50390625,
0.468017578125,
0.505859375,
0.468017578125,
0.52099609375,
0.09478759765625,
0.51806640625
],
[
0.53515625,
0.50048828125,
0.890625,
0.50048828125,
0.890625,
0.5146484375,
0.53515625,
0.5146484375
],
[
0.1094970703125,
0.52099609375,
0.46728515625,
0.51953125,
0.46728515625,
0.5341796875,
0.1094970703125,
0.53515625
],
[
0.53515625,
0.5166015625,
0.890625,
0.5166015625,
0.890625,
0.52978515625,
0.53515625,
0.52978515625
],
[
0.111083984375,
0.53515625,
0.465576171875,
0.53515625,
0.465576171875,
0.548828125,
0.111083984375,
0.548828125
],
[
0.53515625,
0.53076171875,
0.890625,
0.53076171875,
0.890625,
0.54443359375,
0.53515625,
0.54443359375
],
[
0.11273193359375,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.56494140625,
0.11273193359375,
0.56494140625
],
[
0.5361328125,
0.546875,
0.89208984375,
0.546875,
0.89208984375,
0.56005859375,
0.5361328125,
0.56005859375
],
[
0.11273193359375,
0.56494140625,
0.236083984375,
0.56640625,
0.236083984375,
0.580078125,
0.11273193359375,
0.5791015625
],
[
0.53515625,
0.56103515625,
0.64453125,
0.56103515625,
0.64453125,
0.57470703125,
0.53515625,
0.57470703125
],
[
0.09478759765625,
0.59228515625,
0.46728515625,
0.59228515625,
0.46728515625,
0.60595703125,
0.09478759765625,
0.60595703125
],
[
0.50244140625,
0.5986328125,
0.69921875,
0.5986328125,
0.69921875,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.111083984375,
0.609375,
0.468017578125,
0.609375,
0.468017578125,
0.62255859375,
0.111083984375,
0.62255859375
],
[
0.11273193359375,
0.62353515625,
0.468017578125,
0.62353515625,
0.468017578125,
0.63720703125,
0.11273193359375,
0.63720703125
],
[
0.52294921875,
0.62548828125,
0.89208984375,
0.62548828125,
0.89208984375,
0.6396484375,
0.52294921875,
0.6396484375
],
[
0.111083984375,
0.63818359375,
0.465576171875,
0.63818359375,
0.465576171875,
0.6513671875,
0.111083984375,
0.6513671875
],
[
0.5048828125,
0.64013671875,
0.890625,
0.64013671875,
0.890625,
0.65478515625,
0.5048828125,
0.65478515625
],
[
0.11273193359375,
0.65380859375,
0.4697265625,
0.65283203125,
0.4697265625,
0.66748046875,
0.11273193359375,
0.66845703125
],
[
0.5048828125,
0.65576171875,
0.890625,
0.65576171875,
0.890625,
0.66943359375,
0.5048828125,
0.66943359375
],
[
0.5048828125,
0.67041015625,
0.890625,
0.67041015625,
0.890625,
0.68505859375,
0.5048828125,
0.68505859375
],
[
0.2091064453125,
0.68994140625,
0.365234375,
0.69140625,
0.365234375,
0.70556640625,
0.2091064453125,
0.705078125
],
[
0.50341796875,
0.68603515625,
0.88818359375,
0.68603515625,
0.88818359375,
0.69970703125,
0.50341796875,
0.69970703125
],
[
0.5048828125,
0.70068359375,
0.8896484375,
0.70068359375,
0.8896484375,
0.71533203125,
0.5048828125,
0.71533203125
],
[
0.398681640625,
0.708984375,
0.443603515625,
0.708984375,
0.443603515625,
0.72021484375,
0.398681640625,
0.72021484375
],
[
0.098876953125,
0.71630859375,
0.12335205078125,
0.71630859375,
0.12335205078125,
0.72900390625,
0.098876953125,
0.72900390625
],
[
0.50244140625,
0.71533203125,
0.8896484375,
0.71533203125,
0.8896484375,
0.72900390625,
0.50244140625,
0.72900390625
],
[
0.098876953125,
0.73193359375,
0.1217041015625,
0.73193359375,
0.1217041015625,
0.74560546875,
0.098876953125,
0.74560546875
],
[
0.50341796875,
0.72998046875,
0.890625,
0.73095703125,
0.890625,
0.74560546875,
0.50341796875,
0.744140625
],
[
0.09722900390625,
0.75,
0.12335205078125,
0.75,
0.12335205078125,
0.76318359375,
0.09722900390625,
0.76318359375
],
[
0.50341796875,
0.7470703125,
0.7587890625,
0.7470703125,
0.7587890625,
0.76025390625,
0.50341796875,
0.76025390625
],
[
0.09722900390625,
0.767578125,
0.1258544921875,
0.767578125,
0.1258544921875,
0.7822265625,
0.09722900390625,
0.7822265625
],
[
0.09722900390625,
0.78515625,
0.12335205078125,
0.78515625,
0.12335205078125,
0.798828125,
0.09722900390625,
0.798828125
],
[
0.50341796875,
0.7841796875,
0.689453125,
0.7841796875,
0.689453125,
0.79736328125,
0.50341796875,
0.79736328125
],
[
0.098876953125,
0.8017578125,
0.125,
0.8017578125,
0.125,
0.81494140625,
0.098876953125,
0.81494140625
],
[
0.2303466796875,
0.81103515625,
0.253173828125,
0.81103515625,
0.253173828125,
0.8193359375,
0.2303466796875,
0.8193359375
],
[
0.52392578125,
0.81005859375,
0.89208984375,
0.81005859375,
0.89208984375,
0.82373046875,
0.52392578125,
0.82373046875
],
[
0.5048828125,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.5048828125,
0.8388671875
],
[
0.17724609375,
0.81298828125,
0.1846923828125,
0.81298828125,
0.1846923828125,
0.81640625,
0.17724609375,
0.81640625
],
[
0.1895751953125,
0.8271484375,
0.2115478515625,
0.8271484375,
0.2115478515625,
0.8388671875,
0.1895751953125,
0.8388671875
],
[
0.30810546875,
0.8271484375,
0.330078125,
0.8271484375,
0.330078125,
0.8388671875,
0.30810546875,
0.8388671875
],
[
0.369384765625,
0.8271484375,
0.390625,
0.8271484375,
0.390625,
0.8388671875,
0.369384765625,
0.8388671875
],
[
0.428955078125,
0.8271484375,
0.449462890625,
0.8271484375,
0.449462890625,
0.8388671875,
0.428955078125,
0.8388671875
],
[
0.25,
0.82763671875,
0.26953125,
0.82763671875,
0.26953125,
0.83837890625,
0.25,
0.83837890625
],
[
0.13232421875,
0.82763671875,
0.146240234375,
0.82763671875,
0.146240234375,
0.83837890625,
0.13232421875,
0.83837890625
],
[
0.2607421875,
0.84033203125,
0.318603515625,
0.84033203125,
0.318603515625,
0.85498046875,
0.2607421875,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.07843017578125,
0.85498046875,
0.468017578125,
0.85400390625,
0.468017578125,
0.86865234375,
0.07843017578125,
0.869140625
],
[
0.50341796875,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.50341796875,
0.869140625
],
[
0.080078125,
0.869140625,
0.4697265625,
0.869140625,
0.4697265625,
0.88330078125,
0.080078125,
0.88330078125
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88330078125,
0.13232421875,
0.88330078125,
0.13232421875,
0.89453125,
0.080078125,
0.89453125
],
[
0.50341796875,
0.88720703125,
0.79833984375,
0.88720703125,
0.79833984375,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"LD-GAN",
"S-GAN",
"",
"",
"",
"c)Regularized Samples",
"d) L =31",
"b)c=8",
"a)c=3",
"Figure 2. Generated images for an LD representation with a) 3 bands and b) 8 bands. c) Generated SI samples from LD",
"representation employing the statistical regularizer with μgan = μae = 0.001 . For a-c) the top images refer to the generated",
"LD images and the bottom images are the obtained SI by decoding the LD representation. d) Generated SIs without an LD",
"representation.",
"The bottom row represents a false RGB mapping of generated SIs from their LD representations, except",
"for S-GAN, which SIs is directly generated by the GAN.",
"虽然S-GAN在频谱上有许多伪影，但这可能会对基于",
"training and validation, the spectral response pro-",
"DL的SI应用的DA策略的网络训练产生负面影响",
"vided by the ARAD dataset [3l was used For",
"thetraining of thisnetwork.thelearningratede-",
"6.3.DA Applications",
"cays exponentially to achieve a more stable con-",
"我们通过生成合成SI数据集并将其用作基于DL，",
"vergence.",
"CSI，SSSR和RGB的光谱任务的DA来验证所提出",
"的方法。以下所有网络均经过100个时期的训练，学",
"",
"SI super-resolution setting: We employed a UNET",
"习率为1e-3。根据获得的结果，我们报告了每个实验",
"network with 4 downsampling and 4 upsam-",
"的测试图像的最佳性能。",
"pling convolutional blocks to reconstruct the high-",
".CSI recovery setting: The unrolled network deep",
"resolution SI from low-spatial low-spectral mea-",
"spatial-spectral prior [41] was used for CSI recov-",
"surements. The selection of the type of architec-",
"ery.,The measurements were simulated through",
"ture is based on recent SSSR works [42] . The",
"the CASSI system using a random CA with 50 %",
"spatial and spectral decimator factors are set as",
"of transmittance.",
"ks=2,kl =4.",
": SI recovery from RGB images Setting: A variation",
"6.4.QuantitativeResults",
"of the UNET architecture described for the SSSR",
"case is employed. UNET-based models have been",
"对于每个SI应用程序，我们使用20%，50%和",
"previously proposed for RGB to spectral task 44",
"100%新样本考虑LD-GAN和S-GAN生成的每个SI",
". To obtain the RGB input image for the network",
"数据集执行恢复任务，（百分比基于原始数据集的样本",
"数）。如图所示。E5，LD-GAN与c=3实现了更高",
"Discriminator Loss",
"的PSNR性能。此外，当进行S-GAN和LD-GAN之",
"间的比较时，我们可以确定不仅LD-GAN更好，而且",
"L=31",
"1.2",
"S-GAN 的性能低于基线 in several cases 。我们还观",
"1.0",
"察到，使用DA生成的数据数量越多，所提出的方法",
"0.8",
"对SI应用程序的性能的提高就越高",
"0.6",
"0.4",
"6.5. Qualitative Results",
"0.2",
"16",
"从测试数据集重建的SI的视觉结果如图。6；其",
"中大量的新样本允许在PSNR方面有更高的性能。从",
"U",
"10",
"30",
"40",
"50",
"20",
"0",
"Epochs",
"地面实况SI及其重建中提取的光谱特征表明，使用",
"Figure 3. Discriminator loss function convergence for dif-",
"LDSI图像进行的训练实现了与地面实况数据更相似的",
"ferent channels. L = 31 represents the performance by",
"密度。结果表明，使用建议数据集中的合成样本训练的",
"S-GAN.",
"模型在每个任务中都具有更高的重建质量",
"5"
] |
{
"arxiv_id": "2305.00132",
"language": "zh",
"page_idx": 4
}
|
||
2404.01532-zh-1
|
2404.01532-zh-1.jpg
|
用于
自动回归事件时态图生成的集合对齐框架
$\mathbf{Xingwei~Tan}^{1,2},\textbf{Yuxiang Zhou}^2,\textbf{Gabriele Pergola}^1,\textbf{Yulan He}^{1,2,3}$
$^{1}$Department of Computer Science, University of Warwick, UK
$^{2}$Department of Informatics, King's College London, UK
$^{3}$The Alan Turing Institute, UK
$\{$ Xingwei.Tan, Gabriele.Pergola.1 $\}$ @warwick.ac.uk
$\{$ Yuxiang.Zhou, Yulan.He $\}$ @kcl.ac.uk
## $\mathbf{Abstract}$
Event temporal graphs have been shown as convenient and effective representations of complex temporal relations between events in text. Recent studies, which employ pre-trained language models to auto-regressively generate linearised graphs for constructing event temporal graphs, have shown promising results. However, these methods have often led to suboptimal graph generation as the linearised graphs exhibit set characteristics which are instead treated sequentially by language models. This discrepancy stems from the conventional text generation objectives, leading to erroneous penalisation of correct predictions caused by the misalignment of elements in target sequences. To address these challenges, we reframe the task as a conditional set generation problem, proposing a Set-aligning Framework tailored for the effective utilisation of Large Language Models (LLMs). The framework incorporates data augmentations and set-property regularisations designed to alleviate text generation loss penalties associated with the linearised graph edge sequences, thus encouraging the generation of more relation edges. Experimental results show that our framework surpasses existing baselines for event temporal graph generation. Furthermore, under zero-shot settings, the structural knowledge introduced through our framework notably improves model generalisation, particularly when the training examples available are limited. 1
### 1 $\textbf{Introduction}$
理解长文档中提到的事件之间的时间关系对于使用清晰的叙述对复杂文本进行建模至关重要。事件时间关系理解的广泛采用的基准之一是 SemEval2013TempEval-3$\underline{\mathrm{E}}($UzZaman et al.,2013) , 需要直接从原始文本端到端生成事件时间图。事件时间图是时间信息的自然表
$^{1}$For access to experimental code and data, please refer to: https://github.com/Xingwei-Warwick/Set-AligningEvent-Temporal-Graph-Generation
示，节点表示事件，边缘表示事件之间的时间
关系，例如“之前”,“之后”或“同时”
现有研究通常通过两步管道来解决构建事件时态图的问题，第一步侧重于检测文本中的事件，第二步侧重于对它们之间的时间关系进行分类 (McDowell et al., 2017; Ning et al., 2018b) 。然而，这种基于管道的方法受到众所周知的限制，包括(i)在每个步骤中都需要细粒度注释；(ii) 整个管道中错误传播的可能性。在第一步中，事件提取器旨在在给定文档中找到尽可能多的事件触发器，从而导致包含许多琐碎的事件，这些事件通常与叙述缺乏相关性，并且与其他事件没有关系。因此，时间关系提取的下一步将成为许多嘈杂事件的负担，从而显着影响模型的整体准确性和效率
为了解决这些限制，Madaan and Yang (2021) 通过条件文本生成直接生成事件时态图来引入任务的重新表述。这种方法允许使用预先训练的语言模型，更重要的是，克服了与管道架构相关的典型限制。虽然这种方法涉及微调文本生成模型，如 GPT-2, 用于生成线性事件时间图作为序列，但它没有考虑一个重要方面。具体来说，它没有考虑到目标序列(i。e。事件时间关系列表)是顺序不变的，因此应将其视为集合而不是有序序列。例如，以下两个序列表示相同的时间图：
S1:[(Cuomo leaving his office, before, speak to reporters),
$\cdots($Cuomo leaving, before, met with representatives)]
S2:[(Cuomo leaving, before, met with representatives),
$\cdots($Cuomo leaving his office, before, speak to reporters)]
在这种情况下，常规的文本生成损失将(错误地)产生很高的值，因为相应位置的大多数令牌都不匹配，即使事件关系相同。此问题对模型性能有不利影响，原因有几个。首先，它不鼓励语言模型生成额外的边。生成更多的边意味着边集中有更多的潜在排列，使其不太可能与目标匹配。其次，如果序列中最初生成的边缘与目标中的边缘在令牌计数上不同，则会导致所有后续边缘与目标不对齐，即使它们相同， 也会导致高损失值
在这项工作中，我们提出了一个集合对齐框
|
[
[
0.476806640625,
0.0872802734375,
0.5263671875,
0.0872802734375,
0.5263671875,
0.1051025390625,
0.476806640625,
0.1051025390625
],
[
0.287109375,
0.10687255859375,
0.71533203125,
0.10687255859375,
0.71533203125,
0.1217041015625,
0.287109375,
0.1217041015625
],
[
0.22412109375,
0.133056640625,
0.779296875,
0.1312255859375,
0.779296875,
0.1484375,
0.22412109375,
0.15087890625
],
[
0.246826171875,
0.149658203125,
0.75732421875,
0.15087890625,
0.75732421875,
0.1650390625,
0.246826171875,
0.1644287109375
],
[
0.27197265625,
0.16748046875,
0.7314453125,
0.16748046875,
0.7314453125,
0.1817626953125,
0.27197265625,
0.1817626953125
],
[
0.372802734375,
0.1817626953125,
0.62744140625,
0.182861328125,
0.62744140625,
0.19775390625,
0.372802734375,
0.196533203125
],
[
0.255126953125,
0.20068359375,
0.75830078125,
0.20068359375,
0.75830078125,
0.21435546875,
0.255126953125,
0.21435546875
],
[
0.32080078125,
0.21728515625,
0.69287109375,
0.21728515625,
0.69287109375,
0.23095703125,
0.32080078125,
0.23095703125
],
[
0.2626953125,
0.254150390625,
0.341796875,
0.254150390625,
0.341796875,
0.268310546875,
0.2626953125,
0.268310546875
],
[
0.51318359375,
0.254638671875,
0.87890625,
0.254638671875,
0.87890625,
0.268310546875,
0.51318359375,
0.268310546875
],
[
0.51318359375,
0.270751953125,
0.8330078125,
0.270751953125,
0.8330078125,
0.283935546875,
0.51318359375,
0.283935546875
],
[
0.145263671875,
0.281982421875,
0.45751953125,
0.281982421875,
0.45751953125,
0.295654296875,
0.145263671875,
0.295654296875
],
[
0.53076171875,
0.2861328125,
0.880859375,
0.2861328125,
0.880859375,
0.2998046875,
0.53076171875,
0.2998046875
],
[
0.14697265625,
0.297607421875,
0.45751953125,
0.297607421875,
0.45751953125,
0.30810546875,
0.14697265625,
0.30810546875
],
[
0.51318359375,
0.3017578125,
0.880859375,
0.3017578125,
0.880859375,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.145263671875,
0.308837890625,
0.45751953125,
0.308837890625,
0.45751953125,
0.322509765625,
0.145263671875,
0.322509765625
],
[
0.51123046875,
0.31640625,
0.87890625,
0.31640625,
0.87890625,
0.330078125,
0.51123046875,
0.330078125
],
[
0.145263671875,
0.32373046875,
0.45751953125,
0.32373046875,
0.45751953125,
0.33740234375,
0.145263671875,
0.33740234375
],
[
0.145263671875,
0.33740234375,
0.45751953125,
0.33740234375,
0.45751953125,
0.350830078125,
0.145263671875,
0.350830078125
],
[
0.51318359375,
0.334228515625,
0.880859375,
0.334228515625,
0.880859375,
0.347412109375,
0.51318359375,
0.347412109375
],
[
0.14697265625,
0.3515625,
0.45751953125,
0.3515625,
0.45751953125,
0.365234375,
0.14697265625,
0.365234375
],
[
0.51708984375,
0.352783203125,
0.5341796875,
0.352783203125,
0.5341796875,
0.35986328125,
0.51708984375,
0.35986328125
],
[
0.53076171875,
0.349853515625,
0.880859375,
0.349853515625,
0.880859375,
0.363525390625,
0.53076171875,
0.363525390625
],
[
0.1435546875,
0.365234375,
0.456787109375,
0.364013671875,
0.456787109375,
0.377685546875,
0.1435546875,
0.37890625
],
[
0.5146484375,
0.365234375,
0.880859375,
0.365234375,
0.880859375,
0.37890625,
0.5146484375,
0.37890625
],
[
0.14697265625,
0.38134765625,
0.456787109375,
0.38134765625,
0.456787109375,
0.391357421875,
0.14697265625,
0.391357421875
],
[
0.51123046875,
0.380126953125,
0.880859375,
0.38134765625,
0.880859375,
0.3955078125,
0.51123046875,
0.394287109375
],
[
0.14697265625,
0.393798828125,
0.45751953125,
0.393798828125,
0.45751953125,
0.40673828125,
0.14697265625,
0.40673828125
],
[
0.51708984375,
0.396728515625,
0.880859375,
0.396728515625,
0.880859375,
0.410400390625,
0.51708984375,
0.410400390625
],
[
0.1435546875,
0.40673828125,
0.459228515625,
0.40625,
0.459228515625,
0.419189453125,
0.1435546875,
0.42041015625
],
[
0.515625,
0.412109375,
0.880859375,
0.412109375,
0.880859375,
0.42578125,
0.515625,
0.42578125
],
[
0.145263671875,
0.421630859375,
0.45751953125,
0.421630859375,
0.45751953125,
0.435302734375,
0.145263671875,
0.435302734375
],
[
0.1435546875,
0.432861328125,
0.45751953125,
0.43408203125,
0.45751953125,
0.448974609375,
0.1435546875,
0.44775390625
],
[
0.51123046875,
0.427490234375,
0.87890625,
0.427490234375,
0.87890625,
0.441162109375,
0.51123046875,
0.441162109375
],
[
0.145263671875,
0.449462890625,
0.45751953125,
0.449462890625,
0.45751953125,
0.463134765625,
0.145263671875,
0.463134765625
],
[
0.5146484375,
0.44482421875,
0.880859375,
0.44482421875,
0.880859375,
0.457763671875,
0.5146484375,
0.457763671875
],
[
0.14697265625,
0.463134765625,
0.45751953125,
0.463134765625,
0.45751953125,
0.476806640625,
0.14697265625,
0.476806640625
],
[
0.51318359375,
0.460205078125,
0.8818359375,
0.460205078125,
0.8818359375,
0.473876953125,
0.51318359375,
0.473876953125
],
[
0.145263671875,
0.476806640625,
0.45751953125,
0.476806640625,
0.45751953125,
0.490478515625,
0.145263671875,
0.490478515625
],
[
0.51318359375,
0.4755859375,
0.771484375,
0.4755859375,
0.771484375,
0.4892578125,
0.51318359375,
0.4892578125
],
[
0.145263671875,
0.4912109375,
0.45751953125,
0.4912109375,
0.45751953125,
0.5048828125,
0.145263671875,
0.5048828125
],
[
0.5322265625,
0.4912109375,
0.880859375,
0.4912109375,
0.880859375,
0.5048828125,
0.5322265625,
0.5048828125
],
[
0.145263671875,
0.5048828125,
0.45751953125,
0.5048828125,
0.45751953125,
0.5185546875,
0.145263671875,
0.5185546875
],
[
0.51318359375,
0.5087890625,
0.87744140625,
0.5087890625,
0.87744140625,
0.51953125,
0.51318359375,
0.51953125
],
[
0.14697265625,
0.5185546875,
0.45751953125,
0.5185546875,
0.45751953125,
0.5322265625,
0.14697265625,
0.5322265625
],
[
0.51318359375,
0.52392578125,
0.880859375,
0.52392578125,
0.880859375,
0.53759765625,
0.51318359375,
0.53759765625
],
[
0.14697265625,
0.53369140625,
0.456787109375,
0.53369140625,
0.456787109375,
0.54443359375,
0.14697265625,
0.54443359375
],
[
0.51318359375,
0.5390625,
0.880859375,
0.5390625,
0.880859375,
0.552734375,
0.51318359375,
0.552734375
],
[
0.145263671875,
0.54638671875,
0.45751953125,
0.54638671875,
0.45751953125,
0.56103515625,
0.145263671875,
0.56103515625
],
[
0.145263671875,
0.56005859375,
0.456787109375,
0.56005859375,
0.456787109375,
0.57373046875,
0.145263671875,
0.57373046875
],
[
0.51318359375,
0.5546875,
0.880859375,
0.5546875,
0.880859375,
0.568359375,
0.51318359375,
0.568359375
],
[
0.14697265625,
0.57373046875,
0.456787109375,
0.57373046875,
0.456787109375,
0.58740234375,
0.14697265625,
0.58740234375
],
[
0.51318359375,
0.57080078125,
0.87744140625,
0.57080078125,
0.87744140625,
0.58056640625,
0.51318359375,
0.58056640625
],
[
0.145263671875,
0.58837890625,
0.459228515625,
0.58837890625,
0.459228515625,
0.6015625,
0.145263671875,
0.6015625
],
[
0.5146484375,
0.5859375,
0.880859375,
0.5859375,
0.880859375,
0.599609375,
0.5146484375,
0.599609375
],
[
0.145263671875,
0.6015625,
0.455078125,
0.6025390625,
0.455078125,
0.6162109375,
0.145263671875,
0.615234375
],
[
0.51318359375,
0.6015625,
0.880859375,
0.6015625,
0.880859375,
0.615234375,
0.51318359375,
0.615234375
],
[
0.14697265625,
0.6162109375,
0.45751953125,
0.6162109375,
0.45751953125,
0.6298828125,
0.14697265625,
0.6298828125
],
[
0.51318359375,
0.6181640625,
0.880859375,
0.6181640625,
0.880859375,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.145263671875,
0.6298828125,
0.459228515625,
0.6298828125,
0.459228515625,
0.6435546875,
0.145263671875,
0.6435546875
],
[
0.51318359375,
0.6328125,
0.880859375,
0.6328125,
0.880859375,
0.646484375,
0.51318359375,
0.646484375
],
[
0.14697265625,
0.6455078125,
0.456787109375,
0.6455078125,
0.456787109375,
0.65625,
0.14697265625,
0.65625
],
[
0.51318359375,
0.6494140625,
0.64990234375,
0.6494140625,
0.64990234375,
0.6630859375,
0.51318359375,
0.6630859375
],
[
0.145263671875,
0.65771484375,
0.45751953125,
0.65771484375,
0.45751953125,
0.67138671875,
0.145263671875,
0.67138671875
],
[
0.14697265625,
0.67138671875,
0.45751953125,
0.67138671875,
0.45751953125,
0.68505859375,
0.14697265625,
0.68505859375
],
[
0.145263671875,
0.68505859375,
0.45751953125,
0.68505859375,
0.45751953125,
0.69873046875,
0.145263671875,
0.69873046875
],
[
0.509765625,
0.68115234375,
0.8623046875,
0.68115234375,
0.8623046875,
0.69482421875,
0.509765625,
0.69482421875
],
[
0.14697265625,
0.69970703125,
0.459228515625,
0.69970703125,
0.459228515625,
0.71337890625,
0.14697265625,
0.71337890625
],
[
0.54833984375,
0.69677734375,
0.85205078125,
0.69677734375,
0.85205078125,
0.70703125,
0.54833984375,
0.70703125
],
[
0.1435546875,
0.7119140625,
0.459228515625,
0.71337890625,
0.459228515625,
0.72802734375,
0.1435546875,
0.72705078125
],
[
0.51318359375,
0.7080078125,
0.8330078125,
0.708984375,
0.8330078125,
0.72021484375,
0.51318359375,
0.7197265625
],
[
0.5458984375,
0.72021484375,
0.87890625,
0.72021484375,
0.87890625,
0.73095703125,
0.5458984375,
0.73095703125
],
[
0.14697265625,
0.72802734375,
0.235107421875,
0.72802734375,
0.235107421875,
0.7392578125,
0.14697265625,
0.7392578125
],
[
0.51318359375,
0.7451171875,
0.87890625,
0.7451171875,
0.87890625,
0.7587890625,
0.51318359375,
0.7587890625
],
[
0.1175537109375,
0.75732421875,
0.1351318359375,
0.75732421875,
0.1351318359375,
0.7666015625,
0.1175537109375,
0.7666015625
],
[
0.1351318359375,
0.75830078125,
0.1427001953125,
0.75830078125,
0.1427001953125,
0.76318359375,
0.1351318359375,
0.76318359375
],
[
0.1485595703125,
0.75439453125,
0.264404296875,
0.7548828125,
0.264404296875,
0.76953125,
0.1485595703125,
0.7685546875
],
[
0.51123046875,
0.7626953125,
0.87890625,
0.7626953125,
0.87890625,
0.775390625,
0.51123046875,
0.775390625
],
[
0.1175537109375,
0.77978515625,
0.486083984375,
0.77978515625,
0.486083984375,
0.79345703125,
0.1175537109375,
0.79345703125
],
[
0.51318359375,
0.779296875,
0.880859375,
0.779296875,
0.880859375,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.1175537109375,
0.79443359375,
0.486083984375,
0.7958984375,
0.486083984375,
0.80859375,
0.1175537109375,
0.80810546875
],
[
0.51318359375,
0.79443359375,
0.880859375,
0.79443359375,
0.880859375,
0.80810546875,
0.51318359375,
0.80810546875
],
[
0.1175537109375,
0.81103515625,
0.484375,
0.81103515625,
0.484375,
0.82470703125,
0.1175537109375,
0.82470703125
],
[
0.51318359375,
0.81005859375,
0.880859375,
0.81005859375,
0.880859375,
0.82373046875,
0.51318359375,
0.82373046875
],
[
0.1158447265625,
0.826171875,
0.486083984375,
0.82666015625,
0.486083984375,
0.84130859375,
0.1158447265625,
0.84033203125
],
[
0.5146484375,
0.82763671875,
0.87890625,
0.82763671875,
0.87890625,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.1175537109375,
0.84326171875,
0.486083984375,
0.84326171875,
0.486083984375,
0.85693359375,
0.1175537109375,
0.85693359375
],
[
0.51318359375,
0.8427734375,
0.880859375,
0.8427734375,
0.880859375,
0.85546875,
0.51318359375,
0.85546875
],
[
0.118408203125,
0.859375,
0.486083984375,
0.859375,
0.486083984375,
0.87255859375,
0.118408203125,
0.87255859375
],
[
0.51318359375,
0.85791015625,
0.880859375,
0.85791015625,
0.880859375,
0.87158203125,
0.51318359375,
0.87158203125
],
[
0.51318359375,
0.87353515625,
0.8818359375,
0.87353515625,
0.8818359375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.1376953125,
0.88134765625,
0.48779296875,
0.8828125,
0.48779296875,
0.89794921875,
0.1376953125,
0.8955078125
],
[
0.118408203125,
0.89453125,
0.175537109375,
0.89794921875,
0.173828125,
0.9091796875,
0.1175537109375,
0.90625
],
[
0.5146484375,
0.8896484375,
0.66162109375,
0.8896484375,
0.66162109375,
0.90283203125,
0.5146484375,
0.90283203125
],
[
0.1695556640625,
0.896484375,
0.486083984375,
0.896484375,
0.486083984375,
0.90673828125,
0.1695556640625,
0.90673828125
],
[
0.1175537109375,
0.90625,
0.333251953125,
0.90625,
0.333251953125,
0.91943359375,
0.1175537109375,
0.91943359375
],
[
0.53076171875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.53076171875,
0.91845703125
]
] |
[
"用于",
"自动回归事件时态图生成的集合对齐框架",
"Xingwei Tan'2, Yuxiang Zhou?, Gabriele Pergola', Yulan Hel,2,3",
"1Department of Computer Science, University of Warwick, UK",
"2Department of Informatics, King's College London, UK",
"3The Alan Turing Institute, UK",
"( Xingwei.Tan, Gabriele.Pergola.1 } @warwick.ac.uk",
" Yuxiang.Zhou, Yulan.He @kcl.ac.uk",
"Abstract",
"示，节点表示事件，边缘表示事件之间的时间",
"关系，例如“之前”，“之后”或“同时”",
"Event temporal graphs have been shown as",
"现有研究通常通过两步管道来解决构建事件",
"convenient and effective representations of",
"时态图的问题，第一步侧重于检测文本中的事",
"complex temporal relations between events in",
"件，第二步侧重于对它们之间的时间关系进行",
"text. Recent studies, which employ pre-trained",
"language models to auto-regressively generate",
"分类(McDowell et al.,2017;Ning et al.,2018b)",
"linearised graphs for constructing event tem-",
"",
"然而，这种基于管道的方法受到众所周知的",
"poral graphs, have shown promising results.",
"限制，包括（i）在每个步骤中都需要细粒度注",
"However,these methods have often ledto",
"释；（i）整个管道中错误传播的可能性。在第",
"suboptimal graph generation as the linearised",
"一步中，事件提取器旨在在给定文档中找到尽",
"graphs exhibit set characteristics which are in-",
"可能多的事件触发器，从而导致包含许多碎",
"stead treated sequentially by language mod",
"els.This discrepancy stems from the con-",
"的事件，这些事件通常与叙述缺乏相关性，并",
"ventional text generation objectives, leading",
"且与其他事件没有关系。因此，时间关系提取",
"to erroneous penalisation of correct predic-",
"的下一步将成为许多噜杂事件的负担，从而显",
"tions caused by the misalignment of elements",
"着影响模型的整体准确性和效率",
"in target sequences. To address these chal-",
"为了解决这些限制，MadaanandYang（2021)",
"lenges, we reframe the task as a conditional set",
"通过条件文本生成直接生成事件时态图来引人",
"generation problem, proposing a Set-aligning",
"任务的重新表述。这种方法允许使用预先训练",
"Framework tailored for the effective utilisa",
"的语言模型，更重要的是，克服了与管道架构",
"tion of Large Language Models (LLMs). The",
"framework incorporates data augmentations",
"相关的典型限制。虽然这种方法涉及微调文本",
"and set-property regularisations designed to al",
"生成模型，如GPT-2，用于生成线性事件时间",
"leviate text generation loss penalties associ-",
"图作为序列，但它没有考虑一个重要方面。具",
"ated with the linearised graph edge sequences",
"体来说，它没有考虑到目标序列(i。e。事件时",
"thus encouraging the generation of more re-",
"间关系列表）是顺序不变的，因此应将其视为",
"lation edges. Experimental results show that",
"集合而不是有序序列。例如，以下两个序列表",
"our framework surpasses existing baselines",
"示相同的时间图：",
"for event temporal graph generation. Further-",
"more, under zero-shot settings, the structural",
"knowledge introduced through our framework",
"S1: [(Cuomo leaving his office, before, speak to reporters)",
"notably improves model generalisation, partic-",
"+(Cuomo leaving,before,metwith representatives))",
"ularly when the training examples available",
"s2:[(Cuomo leaving, before, met with representatives)",
".-(Cuomo leaving his office, before, speak to reporters)",
"are limited.",
"在这种情况下，常规的文本生成损失将（错误",
"1 ",
"",
"Introduction",
"地）产生很高的值，因为相应位置的大多数令",
"理解长文档中提到的事件之间的时间关系对",
"牌都不匹配，即使事件关系相同。此问题对模",
"于使用清晰的叙述对复杂文本进行建模至",
"型性能有不利影响，原因有几个。首先，它不",
"关重要。事件时间关系理解的广泛采用的基",
"鼓励语言模型生成额外的边。生成更多的边意",
"准之一是SemEval2013TempEval-3E(UzZaman",
"味着边集中有更多的潜在排列，使其不太可能",
"etal.,2013），需要直接从原始文本端到端生成",
"与目标匹配。其次，如果序列中最初生成的边",
"事件时间图。事件时间图是时间信息的自然表",
"缘与目标中的边缘在令牌计数上不同，则会导",
"致所有后续边缘与目标不对齐，即使它们相同，",
"'For access to experimental code and data, please re-",
"fer to:",
"也会导致高损失值",
"https://github.com/Xingwei-Warwick/Set-Aligning",
"Event-Temporal-Graph-Generation",
"在这项工作中，我们提出了一个集合对齐框"
] |
{
"arxiv_id": "2404.01532",
"language": "zh",
"page_idx": 1
}
|
||
2404.01532-zh-2
|
2404.01532-zh-2.jpg
|
架 (SAF),该框架可以有效地利用 LLM 进行自回归事件时态图生成。SAF 结合了一组新颖的正则化，称为集合属性正则化(SPR),以及增强的数据，旨在通过减轻对目标序列的惩罚来解决与在上下文化图生成中使用 LM 损失相关的问题。例如，上面的 S1 和 S2 是同一边缘集的不同序列。即使 S1 具有与目标边缘序列相同的阶数，因此具有比 S2 更低的 LM 损耗它们都将被添加相同的 SPR。因此，它们的损失值的相对差异变小，这避免了模型向 S1 的特定边阶过拟合。此外，如果模型探索在 S2 之后再生成一个边，并且该边是正确的，则 SPR 值将减小，而 LM 损失可能会增加
使用建议的 SAF, 我们从弱监督的 T5 (Raffel et al., 2020) 系列中微调语言模型。此外，我们引入了第一个基于纽约时报的上下文化事件时态图生成的人工注释数据集，我们将其与现有的事件关系提取数据集结合起来，以评估 SAF 框架的有效性。在新注释的《纽约时报》语料库$^{2}$上进行的实验表明，SAF显着增加了生成的边缘数量，从而提高了召回率。此外，我们评估了我们的方法在现有的句子级事件时间关系提取数据集上的性能，即 MATRES (Ning et al., 2018a) 和 TB-Dense (Cassidy et al., 2014) ,在零拍设置下，我们发现通过建议的 SAF引人的结构知识当可用的训练示例有限时，对模型泛化的影响更大
我们的贡献有三个方面：
$\bullet$ We introduce a model-agnostic framework, called SAF, for event temporal graph generation. SAF incorporates novel Set-Aligning regularisations, data augmentation, and weak supervision techniques.
·We offer a human-annotated test set and a weakly-supervised dataset specifically designed for document-level event temporal generation.
$\bullet$ Our extensive experimental results in various settings demonstrate the effectiveness of our proposed model. Our thorough analysis shows that our SAF framework encourages language models to generate at least 24 % more edges than previous graph generation approaches across various datasets.
## 2 $\textbf{Related Work}$
2.1 Event Temporal Graph
事件时态图提取的任务是评估将原始文本作为输入和输出的端到端系统的重要任务 TimeML
$^{2}$https://doi.org/10.35111/77ba-9x74
注释 (i。e。, 时间关系) (UzZaman et al., 2013) 。该任务的早期尝试包括 CAEVO (McDowell et al., 2017) 和 Cogcomptime (Ning et al., 2018b) ,它们依赖于统计和基于规则的方法的结合。近年来，人们越来越努力地使用基于神经网络的方法 (Ning et al., 2019; Han et al., 2019a; Tan et al., 2021a) 开发专门的子系统。大型语言模型的出现为端到端学习铺平了道路，将时间图生成视为条件文本生成 (Madaan and Yang, 2021) 。为了解决 Madaan and Yang (2021) 中尚未探索的集合错位问题，我们提出了一个基于一组新颖的正则化的框架，旨在增强自回归事件时间图的生成
值得注意的是，还有另一个相关且更广泛认可的任务，称为时间关系提取，旨在对预先提取的事件 (Wang et al., 2020; Wen and Ji, 2021; Tan et al., 2023) 之间的时间链接类型进行分类。虽然 Han et al. (2019b) 提出了事件和事件时间关系的联合提取模型，但它们依赖于事件提取监督信号，而我们的工作不需要
# 2.2 Graph Generation with Language Models
使用语言模型生成图形已在许多领域进行了探索。例如，Bosselut et al. (2019)在 ATOMIC 常识知识图 (Sap et al., 2019) 上微调 GPT。Melnyk et al. (2022) 提出了一种基于 T5 的多阶段知识生成系统。然而，这些研究并没有在一代中生成整个图形。相反，Madaan et al.(2021) 使用图生成器和图校正器的组合生成了推理图， 用于在不可行推理中进行查询。Zaratiana et al. (2023) 使用自回归 LM 生成实体和实体关系但它们没有考虑目标的集合属性。与它们不同，我们专注于生成序列的集合属性，这在输人文档和输出序列都相当长的设置中尤为重要
## 2.3 Conditional Set Generation
文本生成模型主要设计用于生成具有严格线性顺序的文本，使其对于生成集而言不是最佳的。最近的 NLP 研究已经承认了这一局限性， 其中已经努力使 seq2seq 框架适应诸如多标签分类和关键字生成 (Qin et al., 2019; Ye et al., 2021)之类的任务。Vinyals et al.(2016) 研究了将集合用作文本生成模型的输入或目标输出的一般挑战。他们发现，在这两种情况下，集合中元素的顺序对收敛和最终困惑都有重大影响。这意味着可能存在输入或输出集序列的最佳顺序，他们建议允许模型在训练期间搜索此顺序。Madaan et al.(2022) 没有采用穷举搜索而是提出了一种数据增强方法来增强顺序不变性，并将集合的基数添加到目标序列中，以确保正确的基数。虽然以前的研究已经解决了多标签预测和关键短语生成，但我们的工作深人
|
[
[
0.12005615234375,
0.0872802734375,
0.486083984375,
0.0872802734375,
0.486083984375,
0.097412109375,
0.12005615234375,
0.097412109375
],
[
0.51318359375,
0.0872802734375,
0.87890625,
0.0872802734375,
0.87890625,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.118408203125,
0.1015625,
0.486083984375,
0.1015625,
0.486083984375,
0.11517333984375,
0.118408203125,
0.11517333984375
],
[
0.5146484375,
0.105712890625,
0.53662109375,
0.105712890625,
0.53662109375,
0.1134033203125,
0.5146484375,
0.1134033203125
],
[
0.5322265625,
0.10394287109375,
0.87890625,
0.10394287109375,
0.87890625,
0.114013671875,
0.5322265625,
0.114013671875
],
[
0.118408203125,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.51318359375,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.118408203125,
0.1341552734375,
0.484375,
0.1341552734375,
0.484375,
0.1478271484375,
0.118408203125,
0.1478271484375
],
[
0.51708984375,
0.1341552734375,
0.880859375,
0.1341552734375,
0.880859375,
0.1478271484375,
0.51708984375,
0.1478271484375
],
[
0.12005615234375,
0.1519775390625,
0.483642578125,
0.1519775390625,
0.483642578125,
0.162109375,
0.12005615234375,
0.162109375
],
[
0.51123046875,
0.1484375,
0.87744140625,
0.1484375,
0.87744140625,
0.162109375,
0.51123046875,
0.162109375
],
[
0.118408203125,
0.1650390625,
0.48779296875,
0.1650390625,
0.48779296875,
0.1787109375,
0.118408203125,
0.1787109375
],
[
0.51123046875,
0.1650390625,
0.87890625,
0.1650390625,
0.87890625,
0.1787109375,
0.51123046875,
0.1787109375
],
[
0.118408203125,
0.18115234375,
0.484375,
0.18115234375,
0.484375,
0.19482421875,
0.118408203125,
0.19482421875
],
[
0.51318359375,
0.18115234375,
0.880859375,
0.18115234375,
0.880859375,
0.19482421875,
0.51318359375,
0.19482421875
],
[
0.118408203125,
0.19775390625,
0.489501953125,
0.19775390625,
0.489501953125,
0.21142578125,
0.118408203125,
0.21142578125
],
[
0.51318359375,
0.19775390625,
0.880859375,
0.19775390625,
0.880859375,
0.21142578125,
0.51318359375,
0.21142578125
],
[
0.1175537109375,
0.2120361328125,
0.484375,
0.2120361328125,
0.484375,
0.2257080078125,
0.1175537109375,
0.2257080078125
],
[
0.51318359375,
0.21435546875,
0.87890625,
0.21435546875,
0.87890625,
0.2244873046875,
0.51318359375,
0.2244873046875
],
[
0.118408203125,
0.2286376953125,
0.484375,
0.2286376953125,
0.484375,
0.2423095703125,
0.118408203125,
0.2423095703125
],
[
0.51123046875,
0.2286376953125,
0.880859375,
0.2286376953125,
0.880859375,
0.2423095703125,
0.51123046875,
0.2423095703125
],
[
0.118408203125,
0.24462890625,
0.484375,
0.24462890625,
0.484375,
0.25830078125,
0.118408203125,
0.25830078125
],
[
0.51123046875,
0.243408203125,
0.87890625,
0.243408203125,
0.87890625,
0.257080078125,
0.51123046875,
0.257080078125
],
[
0.118408203125,
0.260009765625,
0.484375,
0.260009765625,
0.484375,
0.273681640625,
0.118408203125,
0.273681640625
],
[
0.5146484375,
0.260009765625,
0.8818359375,
0.260009765625,
0.8818359375,
0.273681640625,
0.5146484375,
0.273681640625
],
[
0.118408203125,
0.275634765625,
0.39208984375,
0.275634765625,
0.39208984375,
0.289306640625,
0.118408203125,
0.289306640625
],
[
0.515625,
0.276611328125,
0.60546875,
0.276611328125,
0.60546875,
0.2880859375,
0.515625,
0.2880859375
],
[
0.1376953125,
0.29150390625,
0.484375,
0.29150390625,
0.484375,
0.3046875,
0.1376953125,
0.3046875
],
[
0.53076171875,
0.290283203125,
0.87890625,
0.29150390625,
0.87890625,
0.305908203125,
0.53076171875,
0.3046875
],
[
0.12005615234375,
0.308837890625,
0.486083984375,
0.308837890625,
0.486083984375,
0.319580078125,
0.12005615234375,
0.319580078125
],
[
0.51318359375,
0.305908203125,
0.87890625,
0.305908203125,
0.87890625,
0.319580078125,
0.51318359375,
0.319580078125
],
[
0.12005615234375,
0.32470703125,
0.483642578125,
0.32470703125,
0.483642578125,
0.3349609375,
0.12005615234375,
0.3349609375
],
[
0.51318359375,
0.32373046875,
0.880859375,
0.32373046875,
0.880859375,
0.33740234375,
0.51318359375,
0.33740234375
],
[
0.118408203125,
0.339111328125,
0.484375,
0.339111328125,
0.484375,
0.352783203125,
0.118408203125,
0.352783203125
],
[
0.51318359375,
0.339111328125,
0.880859375,
0.339111328125,
0.880859375,
0.352783203125,
0.51318359375,
0.352783203125
],
[
0.118408203125,
0.355224609375,
0.48779296875,
0.355224609375,
0.48779296875,
0.3681640625,
0.118408203125,
0.3681640625
],
[
0.51123046875,
0.352783203125,
0.87890625,
0.35400390625,
0.87890625,
0.3681640625,
0.51123046875,
0.36767578125
],
[
0.118408203125,
0.37060546875,
0.484375,
0.37060546875,
0.484375,
0.38427734375,
0.118408203125,
0.38427734375
],
[
0.51123046875,
0.369384765625,
0.87890625,
0.369384765625,
0.87890625,
0.383056640625,
0.51123046875,
0.383056640625
],
[
0.118408203125,
0.385986328125,
0.486083984375,
0.385986328125,
0.486083984375,
0.399658203125,
0.118408203125,
0.399658203125
],
[
0.51318359375,
0.385986328125,
0.77099609375,
0.385986328125,
0.77099609375,
0.399658203125,
0.51318359375,
0.399658203125
],
[
0.12005615234375,
0.40380859375,
0.483642578125,
0.40380859375,
0.483642578125,
0.41455078125,
0.12005615234375,
0.41455078125
],
[
0.51123046875,
0.410888671875,
0.8251953125,
0.412109375,
0.8251953125,
0.427001953125,
0.51123046875,
0.42578125
],
[
0.12005615234375,
0.419189453125,
0.483642578125,
0.419189453125,
0.483642578125,
0.429931640625,
0.12005615234375,
0.429931640625
],
[
0.5517578125,
0.427490234375,
0.61376953125,
0.427490234375,
0.61376953125,
0.4423828125,
0.5517578125,
0.4423828125
],
[
0.12005615234375,
0.43408203125,
0.484375,
0.43408203125,
0.484375,
0.44775390625,
0.12005615234375,
0.44775390625
],
[
0.118408203125,
0.448974609375,
0.486083984375,
0.448974609375,
0.486083984375,
0.4619140625,
0.118408203125,
0.4619140625
],
[
0.5146484375,
0.448974609375,
0.8818359375,
0.448974609375,
0.8818359375,
0.4619140625,
0.5146484375,
0.4619140625
],
[
0.1275634765625,
0.466064453125,
0.483642578125,
0.466064453125,
0.483642578125,
0.476806640625,
0.1275634765625,
0.476806640625
],
[
0.51318359375,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.118408203125,
0.48095703125,
0.486083984375,
0.48095703125,
0.486083984375,
0.49462890625,
0.118408203125,
0.49462890625
],
[
0.51123046875,
0.479736328125,
0.880859375,
0.479736328125,
0.880859375,
0.493408203125,
0.51123046875,
0.493408203125
],
[
0.118408203125,
0.496337890625,
0.26611328125,
0.496337890625,
0.26611328125,
0.51025390625,
0.118408203125,
0.51025390625
],
[
0.5146484375,
0.495361328125,
0.880859375,
0.495361328125,
0.880859375,
0.5087890625,
0.5146484375,
0.5087890625
],
[
0.1368408203125,
0.51318359375,
0.32666015625,
0.51318359375,
0.32666015625,
0.52685546875,
0.1368408203125,
0.52685546875
],
[
0.5146484375,
0.51123046875,
0.880859375,
0.51123046875,
0.880859375,
0.52490234375,
0.5146484375,
0.52490234375
],
[
0.51123046875,
0.52783203125,
0.880859375,
0.52783203125,
0.880859375,
0.54150390625,
0.51123046875,
0.54150390625
],
[
0.1368408203125,
0.53759765625,
0.484375,
0.5361328125,
0.484375,
0.55029296875,
0.1368408203125,
0.5517578125
],
[
0.5146484375,
0.54345703125,
0.8818359375,
0.54345703125,
0.8818359375,
0.55712890625,
0.5146484375,
0.55712890625
],
[
0.154541015625,
0.55419921875,
0.489501953125,
0.55419921875,
0.489501953125,
0.56787109375,
0.154541015625,
0.56787109375
],
[
0.5146484375,
0.55859375,
0.87890625,
0.55859375,
0.87890625,
0.572265625,
0.5146484375,
0.572265625
],
[
0.154541015625,
0.5693359375,
0.484375,
0.5693359375,
0.484375,
0.5830078125,
0.154541015625,
0.5830078125
],
[
0.51318359375,
0.57470703125,
0.880859375,
0.57470703125,
0.880859375,
0.58837890625,
0.51318359375,
0.58837890625
],
[
0.154541015625,
0.5849609375,
0.48779296875,
0.5849609375,
0.48779296875,
0.5986328125,
0.154541015625,
0.5986328125
],
[
0.5146484375,
0.59033203125,
0.880859375,
0.59033203125,
0.880859375,
0.60400390625,
0.5146484375,
0.60400390625
],
[
0.154541015625,
0.6015625,
0.32666015625,
0.6015625,
0.32666015625,
0.615234375,
0.154541015625,
0.615234375
],
[
0.5146484375,
0.60693359375,
0.880859375,
0.60693359375,
0.880859375,
0.62060546875,
0.5146484375,
0.62060546875
],
[
0.1376953125,
0.626953125,
0.48779296875,
0.626953125,
0.48779296875,
0.64013671875,
0.1376953125,
0.64013671875
],
[
0.51318359375,
0.62255859375,
0.880859375,
0.62255859375,
0.880859375,
0.6357421875,
0.51318359375,
0.6357421875
],
[
0.15283203125,
0.64111328125,
0.489501953125,
0.64111328125,
0.489501953125,
0.65478515625,
0.15283203125,
0.65478515625
],
[
0.509765625,
0.646484375,
0.7734375,
0.646484375,
0.7734375,
0.66015625,
0.509765625,
0.66015625
],
[
0.154541015625,
0.6591796875,
0.48779296875,
0.6591796875,
0.48779296875,
0.6728515625,
0.154541015625,
0.6728515625
],
[
0.5146484375,
0.66845703125,
0.880859375,
0.66845703125,
0.880859375,
0.681640625,
0.5146484375,
0.681640625
],
[
0.151123046875,
0.6748046875,
0.2376708984375,
0.6728515625,
0.2376708984375,
0.68701171875,
0.151123046875,
0.689453125
],
[
0.51318359375,
0.68310546875,
0.880859375,
0.68408203125,
0.880859375,
0.69873046875,
0.51318359375,
0.69775390625
],
[
0.1376953125,
0.69970703125,
0.495361328125,
0.69970703125,
0.495361328125,
0.71337890625,
0.1376953125,
0.71337890625
],
[
0.51318359375,
0.70068359375,
0.880859375,
0.70068359375,
0.880859375,
0.71435546875,
0.51318359375,
0.71435546875
],
[
0.15283203125,
0.71728515625,
0.489501953125,
0.71533203125,
0.489501953125,
0.728515625,
0.15283203125,
0.73095703125
],
[
0.509765625,
0.71435546875,
0.880859375,
0.71533203125,
0.880859375,
0.72998046875,
0.509765625,
0.728515625
],
[
0.154541015625,
0.732421875,
0.489501953125,
0.73095703125,
0.489501953125,
0.7451171875,
0.154541015625,
0.74658203125
],
[
0.51318359375,
0.732421875,
0.87890625,
0.732421875,
0.87890625,
0.7451171875,
0.51318359375,
0.7451171875
],
[
0.15283203125,
0.74658203125,
0.48779296875,
0.74755859375,
0.48779296875,
0.7626953125,
0.15283203125,
0.76123046875
],
[
0.51123046875,
0.7451171875,
0.880859375,
0.74658203125,
0.880859375,
0.76123046875,
0.51123046875,
0.76025390625
],
[
0.154541015625,
0.76318359375,
0.489501953125,
0.76318359375,
0.489501953125,
0.77783203125,
0.154541015625,
0.77783203125
],
[
0.51318359375,
0.76318359375,
0.880859375,
0.76318359375,
0.880859375,
0.77685546875,
0.51318359375,
0.77685546875
],
[
0.154541015625,
0.77978515625,
0.490234375,
0.77978515625,
0.490234375,
0.79345703125,
0.154541015625,
0.79345703125
],
[
0.497802734375,
0.78369140625,
0.50537109375,
0.78369140625,
0.50537109375,
0.78759765625,
0.497802734375,
0.78759765625
],
[
0.51318359375,
0.779296875,
0.880859375,
0.779296875,
0.880859375,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.15283203125,
0.7958984375,
0.4130859375,
0.79443359375,
0.4130859375,
0.80859375,
0.15283203125,
0.81005859375
],
[
0.5146484375,
0.79443359375,
0.880859375,
0.79443359375,
0.880859375,
0.80810546875,
0.5146484375,
0.80810546875
],
[
0.515625,
0.8125,
0.87890625,
0.8125,
0.87890625,
0.822265625,
0.515625,
0.822265625
],
[
0.11419677734375,
0.82177734375,
0.273681640625,
0.82177734375,
0.273681640625,
0.8359375,
0.11419677734375,
0.8359375
],
[
0.5146484375,
0.82666015625,
0.8818359375,
0.82666015625,
0.8818359375,
0.84033203125,
0.5146484375,
0.84033203125
],
[
0.1175537109375,
0.8466796875,
0.345947265625,
0.8466796875,
0.345947265625,
0.86083984375,
0.1175537109375,
0.86083984375
],
[
0.5146484375,
0.8427734375,
0.880859375,
0.8427734375,
0.880859375,
0.85546875,
0.5146484375,
0.85546875
],
[
0.5146484375,
0.85791015625,
0.87890625,
0.85791015625,
0.87890625,
0.87158203125,
0.5146484375,
0.87158203125
],
[
0.118408203125,
0.869140625,
0.484375,
0.869140625,
0.484375,
0.8798828125,
0.118408203125,
0.8798828125
],
[
0.51318359375,
0.87353515625,
0.87890625,
0.87353515625,
0.87890625,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.118408203125,
0.88525390625,
0.484375,
0.88525390625,
0.484375,
0.8955078125,
0.118408203125,
0.8955078125
],
[
0.51318359375,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.139404296875,
0.90625,
0.403076171875,
0.90625,
0.403076171875,
0.91943359375,
0.139404296875,
0.91943359375
],
[
0.515625,
0.90673828125,
0.87890625,
0.90673828125,
0.87890625,
0.91748046875,
0.515625,
0.91748046875
]
] |
[
"架（SAE），该框架可以有效地利用LLM进行",
"注释（i。e。，时间关系）（UzZamanetal.,2013）",
"自回归事件时态图生成。SAF结合了一组新颖",
"8",
"该任务的早期尝试包括CAEVO（McDowell",
"的正则化，称为集合属性正则化（SPR），以及",
"et al., 2017) 和 Cogcomptime (Ning et al., 2018b)",
"增强的数据，旨在通过减轻对目标序列的惩罚",
"，它们依赖于统计和基于规则的方法的结合。",
"来解决与在上下文化图生成中使用LM损失相",
"近年来，人们越来越努力地使用基于神经网络",
"关的问题。例如，上面的S1和S2是同一边缘",
"的方法 (Ning et al., 2019; Han et al., 2019a; Tan",
"集的不同序列。即使S1具有与目标边缘序列",
"etal.,2021a)开发专门的子系统。大型语言模型",
"相同的阶数，因此具有比S2更低的LM损耗，",
"的出现为端到端学习铺平了道路，将时间图生",
"它们都将被添加相同的SPR。因此，它们的损",
"成视为条件文本生成（MadaanandYang，2021）",
"失值的相对差异变小，这避免了模型向S1的",
"。为了解决MadaanandYang（2021）中尚未探",
"特定边阶过拟合。此外，如果模型探索在S2之",
"索的集合错位问题，我们提出了一个基于一组",
"后再生成一个边，并且该边是正确的，则SPR",
"新颖的正则化的框架，旨在增强自回归事件时",
"值将减小，而LM损失可能会增加",
"间图的生成",
"使用建议的SAF，我们从弱监督的T5(Raffel",
"值得注意的是，还有另一个相关且更广泛认",
"etal.2020）系列中微调语言模型。此外，我们",
"可的任务，称为时间关系提取，旨在对预先提",
"引入了第一个基于纽约时报的上下文化事件时",
"取的事件（Wangetal.,2020;Wenand Ji,2021;",
"态图生成的人工注释数据集，我们将其与现有",
"Tanetal.,2023）之间的时间链接类型进行分类。",
"的事件关系提取数据集结合起来，以评估SAF",
"虽然Hanetal.(2019b)提出了事件和事件时间",
"框架的有效性。在新注释的《纽约时报》语料",
"关系的联合提取模型，但它们依赖于事件提取",
"库2上进行的实验表明，SAF显着增加了生成",
"监督信号，而我们的工作不需要",
"的边缘数量，从而提高了召回率。此外，我们",
"2.2Graph Generation with Language",
"评估了我们的方法在现有的句子级事件时间",
"Models",
"关系提取数据集上的性能，即MATRES（Ning",
"et al., 2018a) 和 TB-Dense (Cassidy et al., 2014)",
"使用语言模型生成图形已在许多领域进行了探",
"在零拍设置下，我们发现通过建议的SAF引",
"索。例如，Bosselut et al.(2019)在ATOMIC常",
"入的结构知识当可用的训练示例有限时，对模",
"识知识图（Sapetal.，2019）上微调GPT。Melnyk",
"型泛化的影响更大",
"etal.(2022)提出了一种基于T5的多阶段知识",
"我们的贡献有三个方面：",
"生成系统。然而，这些研究并没有在一代中",
"生成整个图形。相反，Madaanetal.(2021）使",
": We introduce a model-agnostic framework,",
"用图生成器和图校正器的组合生成了推理图，",
"called SAF, for event temporal graph gener-",
"用于在不可行推理中进行查询。Zaratianaetal.",
"ation. SAF incorporates novel Set-Aligning",
"(2023）使用自回归LM生成实体和实体关系，",
"regularisations, data augmentation, and weak",
"但它们没有考虑目标的集合属性。与它们不",
"supervision techniques.",
"同，我们专注于生成序列的集合属性，这在输",
". We offer a human-annotated test set and",
"入文档和输出序列都相当长的设置中尤为重要",
"a weakly-supervised dataset specifically de-",
"2.3Conditional Set Generation",
"signed for document-level event temporal",
"文本生成模型主要设计用于生成具有严格线",
"generation.",
"性顺序的文本，使其对于生成集而言不是最佳",
": Our extensive experimental results in vari-",
"的。最近的NLP研究已经承认了这一局限性，",
"ous settings demonstrate the effectiveness of",
"其中已经努力使seq2seq框架适应诸如多标签",
"our proposed model. Our thorough analysis",
"分类和关键字生成（Qinetal.，2019;Yeetal..",
"shows that our SAF framework encourages",
"2021）之类的任务。Vinyalsetal.（2016)研究了",
"language models to generate at least 24 %",
"将集合用作文本生成模型的输人或目标输出",
"more edges than previous graph generation",
"",
"的一般挑战。他们发现，在这两种情况下，集",
"approaches across various datasets.",
"合中元素的顺序对收敛和最终困惑都有重大影",
"响。这意味着可能存在输入或输出集序列的最",
"2 Related Work",
"佳顺序，他们建议允许模型在训练期间搜索此",
"2.1 Event Temporal Graph",
"顺序。Madaanetal.（2022）没有采用穷举搜索，",
"而是提出了一种数据增强方法来增强顺序不变",
"事件时态图提取的任务是评估将原始文本作为",
"性，并将集合的基数添加到目标序列中，以确",
"输人和输出的端到端系统的重要任务TimeML",
"保正确的基数。虽然以前的研究已经解决了多",
"2https://doi.org/10.35111/77ba-9x74",
"标签预测和关键短语生成，但我们的工作深入"
] |
{
"arxiv_id": "2404.01532",
"language": "zh",
"page_idx": 2
}
|
||
2404.01532-zh-3
|
2404.01532-zh-3.jpg
|
10 Document: Governor Cuomo (el: leaving) his office in Albany yesterday to (e2: speak) to
<FigureHere>
Figure 1: Set-Aligning framework (SAF).
3.1
Event Temporal Graph Modelling as
Edge Set Generation
研究了事件时态图生成所带来的独特挑战，其
中涉及长序列和部分有序属性
从更一般的意义上讲，来自计算机视觉的对象检测任务还涉及集合预测 (Chen et al., 2022) 。Carion et al. (2020) 使用并行解码基于对象查询生成集合中的元素。Tan et al.(2021b) 在具有非自回归解码器的名称实体识别中采用了类似的方法。与图像中的实体不同，我们任务中的集合元素(事件关系)不是具体的空间对象或文本跨度，而是长度变化并分散在每个文档中。这使得对象查询和非自回归解码在我们的设置中不适用。
事件时态图是没有孤立顶点的有向图。图中的每个边描述了两个事件之间的时间关系，不允许自循环。在 Madaan and Yang (2021)之后，我们通过使用 DOT 图描述语言$\operatorname{E}($Gansner, 2006) 将它们线性化为字符串来表示这些图(图1中所示的示例)。鉴于事件时态图没有孤立的顶点，序列本质上代表了图的边集
我们对生成字符串$y$的概率进行建模，该字符串是事件时间图$G$的线性化表示，以文档$X=(x_1,x_2,...,x_n)$为条件，使用语言模型：
### 3 $\textbf{Set- Aligning Framework}$
$$p_{\text{LM}}(y|X)=\prod_{t=1}^Tp(y_t|X,y_{<t})$$
(1)
,其中$y$是以 DOT 表示法格式化的字符串
### 3.2 Data Augmentation
Madaan and Yang (2021) 首先探索了使用神经语言建模生成端到端事件时间图的可能性。然而，自那时以来，这项任务仍未得到充分探索有许多未解决的问题。要详细说明，第一个问题是 Madaan and Yang (2021)框架图生成是一个常规的序列生成问题，而从根本上说是一个集合生成问题。其次，他们构建的数据集主要由小型图组成，未能在文档级理解方面挑战模型。最后，他们的研究主要集中在 GPT-2 上， 而 LLM 的景观随着具有不同结构的模型的出现而发展(例如。g。、编码器-解码器)和新范式(e。g。,上下文学习)近年来。在这项研究中，我们解决了这三个方面，以增强对序列到序列时间图生成的理解
尽管我们提出的框架被设计为与模型无关， 但有几个因素导致我们选择 Flan-T5 作为我们实验的基础模型：(i)根据我们的初步实验Flan-T5-base 在性能方面与。资源消耗，使我们能够测试更多的变体；(ii)由于其在处理冗长文档中的综合信息方面的效率，其编码器-解码器结构非常适合文档级图形生成
事件时态图生成的目标序列本质上是集合，而不是严格排序的文本序列。因此，传统的文本生成丢失可能会无意中惩罚令牌顺序，并迫使元素的排列与目标序列中的顺序相匹配，这不一定是最佳顺序。此强制执行的顺序可能会导致性能欠佳的 (Vinyals et al.,2016)。一个潜在的解决方案是引人集合元素的随机排列作为增强的训练示例，这已经在多标签分类和关键短语生成 (Madaan et al., 2022) 等任务中被证明是有效的。具体来说，在事件时间图生成的上下文中，元素对应于目标字符串中的边。表示边的子字符串被随机洗牌，而字符串的其余部分保持不变
将地面实况边集的集基数前置到生成目标也可能有助于约束生成模型，以避免过度生成(Madaan et al.,2022)。然而，尽管生成的边缘数量显着减少，但我们初步实验中的此类尝试
|
[
[
0.14697265625,
0.0955810546875,
0.154541015625,
0.0955810546875,
0.154541015625,
0.1009521484375,
0.14697265625,
0.1009521484375
],
[
0.1561279296875,
0.09027099609375,
0.611328125,
0.09027099609375,
0.611328125,
0.1009521484375,
0.1561279296875,
0.1009521484375
],
[
0.2225341796875,
0.1015625,
0.599609375,
0.1015625,
0.599609375,
0.11224365234375,
0.2225341796875,
0.11224365234375
],
[
0.22412109375,
0.11224365234375,
0.61865234375,
0.11224365234375,
0.61865234375,
0.122314453125,
0.22412109375,
0.122314453125
],
[
0.2216796875,
0.119384765625,
0.587890625,
0.12054443359375,
0.587890625,
0.1341552734375,
0.2216796875,
0.133056640625
],
[
0.372802734375,
0.1383056640625,
0.393798828125,
0.1383056640625,
0.393798828125,
0.15380859375,
0.372802734375,
0.15380859375
],
[
0.6826171875,
0.1455078125,
0.82861328125,
0.1455078125,
0.82861328125,
0.1561279296875,
0.6826171875,
0.1561279296875
],
[
0.64208984375,
0.15673828125,
0.84814453125,
0.1591796875,
0.84814453125,
0.17333984375,
0.64208984375,
0.171630859375
],
[
0.366943359375,
0.163330078125,
0.399658203125,
0.1650390625,
0.39794921875,
0.1776123046875,
0.365234375,
0.17578125
],
[
0.478515625,
0.163330078125,
0.611328125,
0.163330078125,
0.611328125,
0.17041015625,
0.478515625,
0.17041015625
],
[
0.646484375,
0.16748046875,
0.8212890625,
0.17041015625,
0.8212890625,
0.18408203125,
0.646484375,
0.18115234375
],
[
0.21826171875,
0.17578125,
0.2509765625,
0.17578125,
0.2509765625,
0.18701171875,
0.21826171875,
0.18701171875
],
[
0.2032470703125,
0.1644287109375,
0.267822265625,
0.1644287109375,
0.267822265625,
0.17578125,
0.2032470703125,
0.17578125
],
[
0.339111328125,
0.17578125,
0.42822265625,
0.17578125,
0.42822265625,
0.18701171875,
0.339111328125,
0.18701171875
],
[
0.6455078125,
0.1776123046875,
0.8193359375,
0.179931640625,
0.8193359375,
0.19482421875,
0.6455078125,
0.1923828125
],
[
0.372802734375,
0.20068359375,
0.393798828125,
0.20068359375,
0.393798828125,
0.21728515625,
0.372802734375,
0.21728515625
],
[
0.7109375,
0.203125,
0.7197265625,
0.203125,
0.7197265625,
0.207275390625,
0.7109375,
0.207275390625
],
[
0.78271484375,
0.198974609375,
0.82861328125,
0.203125,
0.8251953125,
0.2120361328125,
0.78076171875,
0.2078857421875
],
[
0.1259765625,
0.22802734375,
0.175537109375,
0.22802734375,
0.175537109375,
0.2381591796875,
0.1259765625,
0.2381591796875
],
[
0.5439453125,
0.2298583984375,
0.623046875,
0.2286376953125,
0.623046875,
0.240478515625,
0.5439453125,
0.2410888671875
],
[
0.72314453125,
0.226806640625,
0.80615234375,
0.221435546875,
0.8076171875,
0.232177734375,
0.724609375,
0.2369384765625
],
[
0.133544921875,
0.2369384765625,
0.4306640625,
0.2369384765625,
0.4306640625,
0.24462890625,
0.133544921875,
0.24462890625
],
[
0.474365234375,
0.240478515625,
0.51904296875,
0.243408203125,
0.51708984375,
0.254150390625,
0.47265625,
0.25048828125
],
[
0.54833984375,
0.2423095703125,
0.62060546875,
0.2423095703125,
0.62060546875,
0.2529296875,
0.54833984375,
0.2529296875
],
[
0.74072265625,
0.2369384765625,
0.77490234375,
0.2369384765625,
0.77490234375,
0.24462890625,
0.74072265625,
0.24462890625
],
[
0.1318359375,
0.24462890625,
0.4609375,
0.24462890625,
0.4609375,
0.254638671875,
0.1318359375,
0.254638671875
],
[
0.1318359375,
0.2529296875,
0.465087890625,
0.2529296875,
0.465087890625,
0.26318359375,
0.1318359375,
0.26318359375
],
[
0.6533203125,
0.248779296875,
0.6650390625,
0.248779296875,
0.6650390625,
0.2529296875,
0.6533203125,
0.2529296875
],
[
0.77685546875,
0.254150390625,
0.81005859375,
0.254150390625,
0.81005859375,
0.26123046875,
0.77685546875,
0.26123046875
],
[
0.1292724609375,
0.262451171875,
0.1485595703125,
0.262451171875,
0.1485595703125,
0.271484375,
0.1292724609375,
0.271484375
],
[
0.52880859375,
0.26123046875,
0.6396484375,
0.26123046875,
0.6396484375,
0.271484375,
0.52880859375,
0.271484375
],
[
0.54833984375,
0.2724609375,
0.62109375,
0.2724609375,
0.62109375,
0.283935546875,
0.54833984375,
0.283935546875
],
[
0.358642578125,
0.31005859375,
0.63818359375,
0.31005859375,
0.63818359375,
0.320556640625,
0.358642578125,
0.320556640625
],
[
0.118408203125,
0.336181640625,
0.489501953125,
0.336181640625,
0.489501953125,
0.349853515625,
0.118408203125,
0.349853515625
],
[
0.51123046875,
0.33837890625,
0.5556640625,
0.33837890625,
0.5556640625,
0.349853515625,
0.51123046875,
0.349853515625
],
[
0.55322265625,
0.336181640625,
0.8486328125,
0.33740234375,
0.8486328125,
0.3515625,
0.55322265625,
0.350830078125
],
[
0.12005615234375,
0.352783203125,
0.35693359375,
0.352783203125,
0.35693359375,
0.366455078125,
0.12005615234375,
0.366455078125
],
[
0.5517578125,
0.3515625,
0.71728515625,
0.3515625,
0.71728515625,
0.365234375,
0.5517578125,
0.365234375
],
[
0.1368408203125,
0.372314453125,
0.486083984375,
0.372314453125,
0.486083984375,
0.385986328125,
0.1368408203125,
0.385986328125
],
[
0.5146484375,
0.374755859375,
0.87890625,
0.374755859375,
0.87890625,
0.388427734375,
0.5146484375,
0.388427734375
],
[
0.118408203125,
0.388427734375,
0.48779296875,
0.388427734375,
0.48779296875,
0.402099609375,
0.118408203125,
0.402099609375
],
[
0.51318359375,
0.39013671875,
0.880859375,
0.39013671875,
0.880859375,
0.40380859375,
0.51318359375,
0.40380859375
],
[
0.133544921875,
0.40380859375,
0.486083984375,
0.40380859375,
0.486083984375,
0.41748046875,
0.133544921875,
0.41748046875
],
[
0.51318359375,
0.407958984375,
0.87890625,
0.407958984375,
0.87890625,
0.418701171875,
0.51318359375,
0.418701171875
],
[
0.118408203125,
0.419189453125,
0.486083984375,
0.419189453125,
0.486083984375,
0.432861328125,
0.118408203125,
0.432861328125
],
[
0.51318359375,
0.421630859375,
0.8818359375,
0.421630859375,
0.8818359375,
0.435791015625,
0.51318359375,
0.435791015625
],
[
0.12005615234375,
0.435302734375,
0.486083984375,
0.435302734375,
0.486083984375,
0.448974609375,
0.12005615234375,
0.448974609375
],
[
0.51123046875,
0.43701171875,
0.87890625,
0.435791015625,
0.87890625,
0.45068359375,
0.51123046875,
0.451904296875
],
[
0.118408203125,
0.45068359375,
0.486083984375,
0.45068359375,
0.486083984375,
0.46435546875,
0.118408203125,
0.46435546875
],
[
0.51318359375,
0.453125,
0.880859375,
0.453125,
0.880859375,
0.466064453125,
0.51318359375,
0.466064453125
],
[
0.1175537109375,
0.46728515625,
0.483642578125,
0.46728515625,
0.483642578125,
0.48095703125,
0.1175537109375,
0.48095703125
],
[
0.5146484375,
0.4697265625,
0.77099609375,
0.4697265625,
0.77099609375,
0.482666015625,
0.5146484375,
0.482666015625
],
[
0.118408203125,
0.482177734375,
0.486083984375,
0.482177734375,
0.486083984375,
0.495361328125,
0.118408203125,
0.495361328125
],
[
0.53076171875,
0.485107421875,
0.880859375,
0.485107421875,
0.880859375,
0.498779296875,
0.53076171875,
0.498779296875
],
[
0.118408203125,
0.498779296875,
0.484375,
0.498779296875,
0.484375,
0.5126953125,
0.118408203125,
0.5126953125
],
[
0.51318359375,
0.501953125,
0.880859375,
0.501953125,
0.880859375,
0.515625,
0.51318359375,
0.515625
],
[
0.118408203125,
0.51416015625,
0.2376708984375,
0.51416015625,
0.2376708984375,
0.52783203125,
0.118408203125,
0.52783203125
],
[
0.509765625,
0.51416015625,
0.861328125,
0.515625,
0.861328125,
0.533203125,
0.509765625,
0.5322265625
],
[
0.68994140625,
0.5458984375,
0.7060546875,
0.5458984375,
0.7060546875,
0.5556640625,
0.68994140625,
0.5556640625
],
[
0.1158447265625,
0.55419921875,
0.151123046875,
0.55419921875,
0.151123046875,
0.5654296875,
0.1158447265625,
0.5654296875
],
[
0.1502685546875,
0.552734375,
0.366943359375,
0.552734375,
0.366943359375,
0.56787109375,
0.1502685546875,
0.56787109375
],
[
0.5810546875,
0.55712890625,
0.8095703125,
0.55712890625,
0.8095703125,
0.57763671875,
0.5810546875,
0.57763671875
],
[
0.85791015625,
0.56103515625,
0.8818359375,
0.56103515625,
0.8818359375,
0.57373046875,
0.85791015625,
0.57373046875
],
[
0.68603515625,
0.5791015625,
0.7109375,
0.5791015625,
0.7109375,
0.58740234375,
0.68603515625,
0.58740234375
],
[
0.1175537109375,
0.5859375,
0.483642578125,
0.5859375,
0.483642578125,
0.5966796875,
0.1175537109375,
0.5966796875
],
[
0.118408203125,
0.6015625,
0.486083984375,
0.6015625,
0.486083984375,
0.615234375,
0.118408203125,
0.615234375
],
[
0.5263671875,
0.60107421875,
0.85205078125,
0.6015625,
0.85205078125,
0.6162109375,
0.5263671875,
0.615234375
],
[
0.51708984375,
0.60693359375,
0.53076171875,
0.60693359375,
0.53076171875,
0.61328125,
0.51708984375,
0.61328125
],
[
0.118408203125,
0.6181640625,
0.484375,
0.6181640625,
0.484375,
0.6318359375,
0.118408203125,
0.6318359375
],
[
0.51318359375,
0.63134765625,
0.54150390625,
0.63134765625,
0.54150390625,
0.642578125,
0.51318359375,
0.642578125
],
[
0.54736328125,
0.62890625,
0.71533203125,
0.6298828125,
0.71533203125,
0.64453125,
0.54736328125,
0.6435546875
],
[
0.118408203125,
0.6494140625,
0.47265625,
0.6494140625,
0.47265625,
0.6630859375,
0.118408203125,
0.6630859375
],
[
0.1175537109375,
0.6318359375,
0.483642578125,
0.6318359375,
0.483642578125,
0.6455078125,
0.1175537109375,
0.6455078125
],
[
0.51123046875,
0.65087890625,
0.880859375,
0.65087890625,
0.880859375,
0.66455078125,
0.51123046875,
0.66455078125
],
[
0.118408203125,
0.66455078125,
0.483642578125,
0.66455078125,
0.483642578125,
0.677734375,
0.118408203125,
0.677734375
],
[
0.509765625,
0.66650390625,
0.87890625,
0.66650390625,
0.87890625,
0.6796875,
0.509765625,
0.6796875
],
[
0.1175537109375,
0.6796875,
0.484375,
0.6796875,
0.484375,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.51318359375,
0.68310546875,
0.880859375,
0.68310546875,
0.880859375,
0.69677734375,
0.51318359375,
0.69677734375
],
[
0.118408203125,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51318359375,
0.69970703125,
0.880859375,
0.69970703125,
0.880859375,
0.71337890625,
0.51318359375,
0.71337890625
],
[
0.1175537109375,
0.71044921875,
0.486083984375,
0.71142578125,
0.486083984375,
0.7255859375,
0.1175537109375,
0.724609375
],
[
0.5263671875,
0.71533203125,
0.880859375,
0.71533203125,
0.880859375,
0.728515625,
0.5263671875,
0.728515625
],
[
0.1175537109375,
0.72705078125,
0.483642578125,
0.72705078125,
0.483642578125,
0.74072265625,
0.1175537109375,
0.74072265625
],
[
0.51123046875,
0.72998046875,
0.87890625,
0.72998046875,
0.87890625,
0.74365234375,
0.51123046875,
0.74365234375
],
[
0.118408203125,
0.74365234375,
0.48779296875,
0.74365234375,
0.48779296875,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.51318359375,
0.74658203125,
0.880859375,
0.74658203125,
0.880859375,
0.76025390625,
0.51318359375,
0.76025390625
],
[
0.1175537109375,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77490234375,
0.1175537109375,
0.77490234375
],
[
0.51318359375,
0.7626953125,
0.880859375,
0.7626953125,
0.880859375,
0.775390625,
0.51318359375,
0.775390625
],
[
0.12005615234375,
0.77490234375,
0.486083984375,
0.77490234375,
0.486083984375,
0.7880859375,
0.12005615234375,
0.7880859375
],
[
0.51318359375,
0.77783203125,
0.880859375,
0.77783203125,
0.880859375,
0.79150390625,
0.51318359375,
0.79150390625
],
[
0.118408203125,
0.79150390625,
0.30322265625,
0.79150390625,
0.30322265625,
0.8046875,
0.118408203125,
0.8046875
],
[
0.51318359375,
0.7919921875,
0.87744140625,
0.7919921875,
0.87744140625,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.1368408203125,
0.81005859375,
0.484375,
0.81005859375,
0.484375,
0.82373046875,
0.1368408203125,
0.82373046875
],
[
0.5146484375,
0.81005859375,
0.880859375,
0.81005859375,
0.880859375,
0.82373046875,
0.5146484375,
0.82373046875
],
[
0.118408203125,
0.826171875,
0.486083984375,
0.826171875,
0.486083984375,
0.8388671875,
0.118408203125,
0.8388671875
],
[
0.51123046875,
0.82373046875,
0.880859375,
0.82470703125,
0.880859375,
0.8388671875,
0.51123046875,
0.83837890625
],
[
0.12005615234375,
0.84326171875,
0.48193359375,
0.84326171875,
0.48193359375,
0.85400390625,
0.12005615234375,
0.85400390625
],
[
0.51123046875,
0.84033203125,
0.587890625,
0.83837890625,
0.587890625,
0.8525390625,
0.51123046875,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.5283203125,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.5283203125,
0.87060546875
],
[
0.118408203125,
0.87353515625,
0.484375,
0.87353515625,
0.484375,
0.88720703125,
0.118408203125,
0.88720703125
],
[
0.51318359375,
0.87353515625,
0.8818359375,
0.87353515625,
0.8818359375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.12005615234375,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.12005615234375,
0.90283203125
],
[
0.509765625,
0.88818359375,
0.87890625,
0.88818359375,
0.87890625,
0.90185546875,
0.509765625,
0.90185546875
],
[
0.118408203125,
0.90478515625,
0.393798828125,
0.90478515625,
0.393798828125,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.5146484375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.5146484375,
0.91845703125
]
] |
[
"",
"Document:Governor Cuomo (el:leaving)his office in Albanyyesterday to (e2: speak)to",
"reporters after he (e4: met)with representatives of groups pushing for higher",
"ethical standards for public officials in the state.\"We (e3: wanted)the Governor",
"andtheLegislaturetoknow thatthefeeling isoutthere.\"saidPaulElisha.",
"",
"Set-Property Regularisations",
"-Matching Regularisation:RHausdorff",
"Large",
"-CombinedwithLMLoss",
"-Cardinality Regularisation:Reard",
"Data",
"Augmented",
"LanguageModel",
"-DuplicationRegularisation:Raupt",
"+",
"2",
"Raausdorf",
"digraph G (",
"EdgeSet-Gold",
"EdgeEmbedding",
"\"Gov.Cuomoleavinghisoffice\"->\"Tospeaktoreporters\"[l=before];",
"Parsing",
"[ei,e2..,en]",
"e2:speal",
"\"Gov. Cuomo leaving his office\"->\"He met with representatives\" [I-before];",
"\"He met with representatives\"->\"Gov. Cuomo leaving his office\" [I - before];",
"",
"distonca",
"[]",
"Edge Set-Generated",
"[ei,e2,.,en]",
"Figure l:Set-Aligningframework (SAF)",
"研究了事件时态图生成所带来的独特挑战，其",
"3.1",
"Event Temporal Graph Modelling as",
"中涉及长序列和部分有序属性",
"Edge Set Generation",
"从更一般的意义上讲，来自计算机视觉的对",
"事件时态图是没有孤立顶点的有向图。图中的",
"象检测任务还涉及集合预测（Chenetal.,2022)",
"每个边描述了两个事件之间的时间关系，不允",
"Carionetal.(2020)使用并行解码基于对象查",
"许自循环。在MadaanandYang（2021）之后，我",
"询生成集合中的元素。Tanetal.(2021b）在具",
"们通过使用DOT图描述语言E(Gansner,2006)",
"有非自回归解码器的名称实体识别中采用了类",
"将它们线性化为字符串来表示这些图（图1中",
"似的方法。与图像中的实体不同，我们任务中",
"所示的示例）。鉴于事件时态图没有孤立的顶",
"的集合元素（事件关系）不是具体的空间对象",
"点，序列本质上代表了图的边集",
"或文本跨度，而是长度变化并分散在每个文档",
"我们对生成字符串9的概率进行建模，该字",
"中。这使得对象查询和非自回归解码在我们的",
"符串是事件时间图G的线性化表示，以文档",
"设置中不适用。",
"X=（a1,c2..an）为条件，使用语言模型：",
"T",
"3",
"Set-Aligning Framework",
"PLM(g|X) = IIp(yt/X, y<t)",
"(1)",
"t=1",
"MadaanandYang（2021）首先探索了使用神经",
"语言建模生成端到端事件时间图的可能性。然",
"其中y是以DOT表示法格式化的字符串",
"",
"而，自那时以来，这项任务仍未得到充分探索，",
"3.2",
"Data Augmentation",
"题是MadaanandYang（2021）框架图生成是",
"有许多未解决的问题。要详细说明，第一个问",
"事件时态图生成的目标序列本质上是集合，而",
"个常规的序列生成问题，而从根本上说是一个",
"不是严格排序的文本序列。因此，传统的文本",
"集合生成问题。其次，他们构建的数据集主要",
"生成丢失可能会无意中惩罚令牌顺序，并迫使",
"由小型图组成，未能在文档级理解方面挑战模",
"元素的排列与目标序列中的顺序相匹配，这不",
"型。最后，他们的研究主要集中在GPT-2上，",
"定是最佳顺序。此强制执行的顺序可能会导",
"而LLM的景观随着具有不同结构的模型的出",
"致性能欠佳的（Vinyalsetal.,2016）。一个潜在",
"现而发展（例如。g。、编码器-解码器）和新范",
"的解决方案是引人集合元素的随机排列作为增",
"式(e。g。，上下文学习)近年来。在这项研究",
"强的训练示例，这已经在多标签分类和关键短",
"中，我们解决了这三个方面，以增强对序列到",
"语生成（Madaanetal.,2022)等任务中被证明是",
"序列时间图生成的理解",
"有效的。具体来说，在事件时间图生成的上下",
"尽管我们提出的框架被设计为与模型无关，",
"文中，元素对应于目标字符串中的边。表示边",
"但有几个因素导致我们选择Flan-T5作为我们",
"的子字符串被随机洗牌，而字符串的其余部分",
"实验的基础模型：（i）根据我们的初步实验",
"保持不变",
"Flan-T5-base在性能方面与。资源消耗，使我们",
"将地面实况边集的集基数前置到生成目标",
"能够测试更多的变体；（i）由于其在处理兀长",
"也可能有助于约束生成模型，以避免过度生成",
"文档中的综合信息方面的效率，其编码器-解",
"(Madaanetal.，2022）。然而，尽管生成的边缘",
"码器结构非常适合文档级图形生成",
"数量显着减少，但我们初步实验中的此类尝试"
] |
{
"arxiv_id": "2404.01532",
"language": "zh",
"page_idx": 3
}
|
||
2404.01532-zh-4
|
2404.01532-zh-4.jpg
|
导致边缘$F_1$得分下降了 4%。因此，我们决定
不将基数纳入最终框架
# 33 Set Property Regularisations (SPR)
简单地添加增强数据来训练模型并不能解决集合对齐的基本问题。在这种方法中出现了几个挑战。首先，添加所有排列是不现实的，尤其是在处理包含众多事件关系的长文档时，因为训练数据将以与目标集基数的阶乘成比例的速率增长。更重要的是，对于每个增强的示例， 损失函数仍将惩罚集合中未观察到的排列。这会使训练不稳定
核心挑战在于找到一种有效的方法来将线性化的目标图与线性化的生成图进行比较，而不像传统的文本生成那样依赖于严格的逐个令牌比较。为了解决这个问题，我们建议对生成目标进行修改。由于线性化图本质上表示图的边集，我们可以将图的比较问题简化为集合比较问题。我们的方法涉及几个组件。首先，我们添加一组基数正则化，以鼓励模型生成足够数量的时间关系边。然后，我们引入重复正则化来惩罚边缘集中元素的任何重复。最后，我们设计了一个集合匹配正则化，用于评估目标边集中的元素与生成的边集中的元素之间的语义相似性。总的来说，上述正则化被称为集合属性正则化 (SPR)。它们通过加权平均与令牌级交叉熵损失相结合
要计算集合属性正则化，需要首先从给定训练输入的语言模型中采样图形字符串。然后， 将此序列解析为边$E$的列表，其中每个边$e$是由头事件，关系类型和尾事件$(h,r,t)$组成的三元组。使用基于规则的解析器进行解析，该解析器将图形文本字符串转换为结构化数据表示。当边缘加载到结构化列表中时，可以计算边缘和重复边缘的数量。令$\mathcal{E}$表示$E$中所有唯一边的集合。集合基数正则化和重复正则化的值可以计算如下：
(2)
(3)
$$\begin{aligned}\mathcal{E}&=\{e|e\in E\}\\R_{\mathrm{dupl}}&=\frac{|E|-|\mathcal{E}|}{|\mathcal{E}|}\\R_{\mathrm{card}}&=\frac{\mathrm{abs}(|\mathcal{E}^{\prime}|-|\mathcal{E}|)}{|\mathcal{E}|}\end{aligned}$$
(4)
其中函数 abs(·)表示取绝对值，$\mathcal{E}^\prime$表示地面实
况边缘集
为了计算集合匹配正则化，我们通过比较两个集合中边的语义相似性来评估生成的集合和目标集合之间的相似性。我们将解码器对相应令牌的表示的最后一层作为事件和关系类型的语义表示。然后，我们将这些表示连接为每个边的语义表示：
$$z_{h}=H_{[h_{1},h_{2},...,h_{m}]}\text{(5)}\\z_{r}=H_{[r_{1},r_{2},...,r_{s}]}\text{(6)}\\z_{t}=H_{[t_{1},t_{2},...,t_{n}]}\text{(7)}\\\bar{e}=\begin{bmatrix}\mathrm{pool}(z_h);\mathrm{pool}(z_r);\mathrm{pool}(z_t))\end{bmatrix}\text{(8)}$$
其中$H$是解码器的最后一层隐藏状态。$[ h_{1}, . . . , h_{m}]$ $、 [ r_{1}, . . . , r_{s}]$ 和$[t_1,...,t_n]$ 分别是头事件、关系类型和尾事件的索引。$z_h,z_r,z_t$分别表示头事件，关系类型和尾事件的语义表示。pool($\cdot) 表示平均池函数。\bar{e} 表示边的语义$ 表示
我们现在拥有两组嵌入：一组包含从目标图中提取的边嵌入，另一组包含从生成的图中派生的边嵌入。从本质上讲，它们可以被视为表示空间中的两组点。因此，我们可以通过测量表示空间中两个点集(流形)之间的距离来测量两个图的相似性。Hausdorff 距离最初定义为测量度量空间内两个子集之间的间距，最近在机器学习中发现了用于测量两组嵌人(Schutze et al., 2012; Wang et al., 2023) 之间的距离的应用。我们计算平均 Hausdorff 距离作为度量：
$$\begin{aligned}d_{H}(\mathcal{E}^{\prime},\mathcal{E})=&\frac{1}{|\mathcal{E}^{\prime}|}\sum_{\bar{e^{\prime}}\in\mathcal{E^{\prime}}}\operatorname*{min}_{\bar{e}\in\mathcal{E}}d_{cos}(\bar{e^{\prime}},\bar{e})\\&+\frac{1}{|\mathcal{E}|}\sum_{\bar{e}\in\mathcal{E}}\operatorname*{min}_{\bar{e^{\prime}}\in\mathcal{E^{\prime}}}d_{cos}(\bar{e^{\prime}},\bar{e})\end{aligned}$$
(9)
,其中边缘对的距离由余弦距离$d_cos(\cdot)$计算
当模型以与目标序列不同的顺序生成集合元素时，令牌级交叉熵损失将很高。如果模型生成更多正确的元素作为顺序错误的序列的后缀，则损失值可能会进一步增加。但是，SPR 将具有较低的值，从而减轻了由令牌级交叉熵损失引起的生成更多元素的沮丧
# 3.4 Fine-tuning with Set Property Regularisations
与基于并行解码 (Carion et al., 2020; Tan et al., 2021b)的集合预测方法不同，SPR 不能直接用作自回归生成中的主要目标。这有两个主要原因。第一个原因是获得 SPR 需要从解码器采样，这会大大降低训练速度。此外，第二个原因是语言模型将难以准确地生成 DOT 格式的序列，因为学习这种格式的令牌依赖需要语言建模目标。因此，序列解析器将无法识别序列中的任何有效边，从而导致高 SPR 值并阻碍训练
为了避免上述问题，我们在一定数量的微调选代之后引人了 SPR。一旦模型在生成正确的
|
[
[
0.12005615234375,
0.0872802734375,
0.484375,
0.0872802734375,
0.484375,
0.1009521484375,
0.12005615234375,
0.1009521484375
],
[
0.118408203125,
0.1015625,
0.30322265625,
0.1015625,
0.30322265625,
0.11517333984375,
0.118408203125,
0.11517333984375
],
[
0.56103515625,
0.11517333984375,
0.70361328125,
0.1181640625,
0.70361328125,
0.133056640625,
0.56103515625,
0.1300048828125
],
[
0.8564453125,
0.11517333984375,
0.8818359375,
0.11517333984375,
0.8818359375,
0.1300048828125,
0.8564453125,
0.1300048828125
],
[
0.11419677734375,
0.1270751953125,
0.43994140625,
0.127685546875,
0.43994140625,
0.142578125,
0.11419677734375,
0.141357421875
],
[
0.56103515625,
0.133056640625,
0.6943359375,
0.13720703125,
0.69287109375,
0.155029296875,
0.55908203125,
0.15087890625
],
[
0.8564453125,
0.133056640625,
0.884765625,
0.13720703125,
0.880859375,
0.15380859375,
0.85205078125,
0.149658203125
],
[
0.118408203125,
0.1484375,
0.484375,
0.1484375,
0.484375,
0.162109375,
0.118408203125,
0.162109375
],
[
0.5615234375,
0.155029296875,
0.69189453125,
0.1591796875,
0.68994140625,
0.177001953125,
0.56103515625,
0.1728515625
],
[
0.8564453125,
0.155029296875,
0.8818359375,
0.157958984375,
0.87890625,
0.1728515625,
0.85400390625,
0.169189453125
],
[
0.12176513671875,
0.166259765625,
0.483642578125,
0.166259765625,
0.483642578125,
0.177001953125,
0.12176513671875,
0.177001953125
],
[
0.118408203125,
0.18115234375,
0.486083984375,
0.18115234375,
0.486083984375,
0.19482421875,
0.118408203125,
0.19482421875
],
[
0.5693359375,
0.1776123046875,
0.8818359375,
0.1776123046875,
0.8818359375,
0.1923828125,
0.5693359375,
0.1923828125
],
[
0.118408203125,
0.1953125,
0.486083984375,
0.1953125,
0.486083984375,
0.208984375,
0.118408203125,
0.208984375
],
[
0.52880859375,
0.2037353515625,
0.8818359375,
0.204833984375,
0.8818359375,
0.2197265625,
0.52880859375,
0.218505859375
],
[
0.12176513671875,
0.213134765625,
0.484375,
0.213134765625,
0.484375,
0.223876953125,
0.12176513671875,
0.223876953125
],
[
0.5224609375,
0.21142578125,
0.52880859375,
0.21142578125,
0.52880859375,
0.215576171875,
0.5224609375,
0.215576171875
],
[
0.51123046875,
0.2197265625,
0.880859375,
0.2203369140625,
0.880859375,
0.235107421875,
0.51123046875,
0.2340087890625
],
[
0.12005615234375,
0.2286376953125,
0.486083984375,
0.2286376953125,
0.486083984375,
0.2392578125,
0.12005615234375,
0.2392578125
],
[
0.51318359375,
0.235107421875,
0.880859375,
0.236328125,
0.880859375,
0.25048828125,
0.51318359375,
0.2493896484375
],
[
0.118408203125,
0.243408203125,
0.486083984375,
0.243408203125,
0.486083984375,
0.257080078125,
0.118408203125,
0.257080078125
],
[
0.5146484375,
0.2529296875,
0.880859375,
0.2529296875,
0.880859375,
0.2666015625,
0.5146484375,
0.2666015625
],
[
0.118408203125,
0.2587890625,
0.2476806640625,
0.2587890625,
0.2476806640625,
0.2724609375,
0.118408203125,
0.2724609375
],
[
0.51123046875,
0.268310546875,
0.880859375,
0.267333984375,
0.880859375,
0.281982421875,
0.51123046875,
0.283203125
],
[
0.1368408203125,
0.27490234375,
0.486083984375,
0.27490234375,
0.486083984375,
0.2880859375,
0.1368408203125,
0.2880859375
],
[
0.12005615234375,
0.290283203125,
0.48779296875,
0.290283203125,
0.48779296875,
0.303955078125,
0.12005615234375,
0.303955078125
],
[
0.51318359375,
0.2861328125,
0.5498046875,
0.2861328125,
0.5498046875,
0.297607421875,
0.51318359375,
0.297607421875
],
[
0.52880859375,
0.2998046875,
0.87890625,
0.2998046875,
0.87890625,
0.3134765625,
0.52880859375,
0.3134765625
],
[
0.12005615234375,
0.30712890625,
0.483642578125,
0.30712890625,
0.483642578125,
0.317626953125,
0.12005615234375,
0.317626953125
],
[
0.118408203125,
0.322509765625,
0.483642578125,
0.322509765625,
0.483642578125,
0.336181640625,
0.118408203125,
0.336181640625
],
[
0.5146484375,
0.317626953125,
0.87890625,
0.317626953125,
0.87890625,
0.327880859375,
0.5146484375,
0.327880859375
],
[
0.118408203125,
0.33740234375,
0.484375,
0.33740234375,
0.484375,
0.350830078125,
0.118408203125,
0.350830078125
],
[
0.51318359375,
0.33203125,
0.880859375,
0.33203125,
0.880859375,
0.345703125,
0.51318359375,
0.345703125
],
[
0.1175537109375,
0.352783203125,
0.483642578125,
0.352783203125,
0.483642578125,
0.366455078125,
0.1175537109375,
0.366455078125
],
[
0.51318359375,
0.347412109375,
0.87890625,
0.347412109375,
0.87890625,
0.361083984375,
0.51318359375,
0.361083984375
],
[
0.12005615234375,
0.369384765625,
0.486083984375,
0.369384765625,
0.486083984375,
0.383056640625,
0.12005615234375,
0.383056640625
],
[
0.51318359375,
0.364013671875,
0.87890625,
0.364013671875,
0.87890625,
0.377685546875,
0.51318359375,
0.377685546875
],
[
0.51318359375,
0.37890625,
0.880859375,
0.37890625,
0.880859375,
0.392578125,
0.51318359375,
0.392578125
],
[
0.118408203125,
0.385498046875,
0.484375,
0.385498046875,
0.484375,
0.3984375,
0.118408203125,
0.3984375
],
[
0.51123046875,
0.393798828125,
0.87890625,
0.393798828125,
0.87890625,
0.40673828125,
0.51123046875,
0.40673828125
],
[
0.12005615234375,
0.402587890625,
0.48193359375,
0.402587890625,
0.48193359375,
0.413330078125,
0.12005615234375,
0.413330078125
],
[
0.51123046875,
0.410400390625,
0.87890625,
0.410400390625,
0.87890625,
0.42333984375,
0.51123046875,
0.42333984375
],
[
0.12005615234375,
0.418701171875,
0.484375,
0.418701171875,
0.484375,
0.4287109375,
0.12005615234375,
0.4287109375
],
[
0.51318359375,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.440673828125,
0.51318359375,
0.440673828125
],
[
0.12005615234375,
0.43408203125,
0.483642578125,
0.43408203125,
0.483642578125,
0.44482421875,
0.12005615234375,
0.44482421875
],
[
0.118408203125,
0.448974609375,
0.486083984375,
0.448974609375,
0.486083984375,
0.4619140625,
0.118408203125,
0.4619140625
],
[
0.5146484375,
0.443603515625,
0.85791015625,
0.443603515625,
0.85791015625,
0.45361328125,
0.5146484375,
0.45361328125
],
[
0.1175537109375,
0.46435546875,
0.483642578125,
0.46435546875,
0.483642578125,
0.47802734375,
0.1175537109375,
0.47802734375
],
[
0.118408203125,
0.479736328125,
0.486083984375,
0.479736328125,
0.486083984375,
0.493408203125,
0.118408203125,
0.493408203125
],
[
0.6484375,
0.486328125,
0.65576171875,
0.486328125,
0.65576171875,
0.4912109375,
0.6484375,
0.4912109375
],
[
0.6708984375,
0.485107421875,
0.8154296875,
0.4892578125,
0.8134765625,
0.50732421875,
0.6689453125,
0.5029296875
],
[
0.118408203125,
0.495361328125,
0.26611328125,
0.495361328125,
0.26611328125,
0.5087890625,
0.118408203125,
0.5087890625
],
[
0.5439453125,
0.4912109375,
0.62646484375,
0.4912109375,
0.62646484375,
0.5029296875,
0.5439453125,
0.5029296875
],
[
0.623046875,
0.495361328125,
0.63037109375,
0.495361328125,
0.63037109375,
0.50048828125,
0.623046875,
0.50048828125
],
[
0.6396484375,
0.496337890625,
0.67333984375,
0.496337890625,
0.67333984375,
0.5126953125,
0.6396484375,
0.5126953125
],
[
0.7001953125,
0.501953125,
0.73046875,
0.501953125,
0.73046875,
0.5087890625,
0.7001953125,
0.5087890625
],
[
0.1368408203125,
0.51123046875,
0.484375,
0.51123046875,
0.484375,
0.52490234375,
0.1368408203125,
0.52490234375
],
[
0.6708984375,
0.51025390625,
0.7001953125,
0.51025390625,
0.7001953125,
0.5185546875,
0.6708984375,
0.5185546875
],
[
0.118408203125,
0.52783203125,
0.484375,
0.52783203125,
0.484375,
0.54150390625,
0.118408203125,
0.54150390625
],
[
0.6708984375,
0.525390625,
0.67822265625,
0.525390625,
0.67822265625,
0.533203125,
0.6708984375,
0.533203125
],
[
0.68603515625,
0.52783203125,
0.861328125,
0.53076171875,
0.861328125,
0.5458984375,
0.68603515625,
0.5419921875
],
[
0.6484375,
0.53369140625,
0.66650390625,
0.5380859375,
0.66162109375,
0.54736328125,
0.64404296875,
0.54345703125
],
[
0.65576171875,
0.53369140625,
0.755859375,
0.54150390625,
0.75244140625,
0.56396484375,
0.6533203125,
0.55712890625
],
[
0.85546875,
0.5322265625,
0.880859375,
0.5322265625,
0.880859375,
0.54345703125,
0.85546875,
0.54345703125
],
[
0.12005615234375,
0.54443359375,
0.483642578125,
0.54443359375,
0.483642578125,
0.5546875,
0.12005615234375,
0.5546875
],
[
0.118408203125,
0.55810546875,
0.483642578125,
0.55810546875,
0.483642578125,
0.57177734375,
0.118408203125,
0.57177734375
],
[
0.12176513671875,
0.576171875,
0.484375,
0.576171875,
0.484375,
0.5859375,
0.12176513671875,
0.5859375
],
[
0.51123046875,
0.57666015625,
0.53076171875,
0.57666015625,
0.53076171875,
0.5849609375,
0.51123046875,
0.5849609375
],
[
0.52294921875,
0.572265625,
0.8623046875,
0.572265625,
0.8623046875,
0.5859375,
0.52294921875,
0.5859375
],
[
0.12005615234375,
0.5927734375,
0.484375,
0.5927734375,
0.484375,
0.6025390625,
0.12005615234375,
0.6025390625
],
[
0.53076171875,
0.5888671875,
0.880859375,
0.5888671875,
0.880859375,
0.6025390625,
0.53076171875,
0.6025390625
],
[
0.118408203125,
0.60693359375,
0.483642578125,
0.60693359375,
0.483642578125,
0.61767578125,
0.118408203125,
0.61767578125
],
[
0.5146484375,
0.60693359375,
0.87744140625,
0.60693359375,
0.87744140625,
0.61767578125,
0.5146484375,
0.61767578125
],
[
0.118408203125,
0.62158203125,
0.486083984375,
0.62158203125,
0.486083984375,
0.63525390625,
0.118408203125,
0.63525390625
],
[
0.5146484375,
0.62158203125,
0.87744140625,
0.62158203125,
0.87744140625,
0.6318359375,
0.5146484375,
0.6318359375
],
[
0.1175537109375,
0.6357421875,
0.48779296875,
0.63720703125,
0.48779296875,
0.65185546875,
0.1175537109375,
0.65087890625
],
[
0.51318359375,
0.6357421875,
0.880859375,
0.6357421875,
0.880859375,
0.6494140625,
0.51318359375,
0.6494140625
],
[
0.12005615234375,
0.65478515625,
0.273681640625,
0.65478515625,
0.273681640625,
0.6650390625,
0.12005615234375,
0.6650390625
],
[
0.51318359375,
0.65283203125,
0.880859375,
0.65283203125,
0.880859375,
0.66650390625,
0.51318359375,
0.66650390625
],
[
0.51318359375,
0.66845703125,
0.771484375,
0.66845703125,
0.771484375,
0.681640625,
0.51318359375,
0.681640625
],
[
0.2376708984375,
0.693359375,
0.353515625,
0.6953125,
0.353515625,
0.71044921875,
0.2376708984375,
0.7080078125
],
[
0.462646484375,
0.69677734375,
0.490234375,
0.69677734375,
0.490234375,
0.71044921875,
0.462646484375,
0.71044921875
],
[
0.50537109375,
0.69140625,
0.79443359375,
0.693359375,
0.79443359375,
0.71142578125,
0.50537109375,
0.708984375
],
[
0.48779296875,
0.69873046875,
0.50732421875,
0.69873046875,
0.50732421875,
0.7060546875,
0.48779296875,
0.7060546875
],
[
0.274658203125,
0.71533203125,
0.34765625,
0.71533203125,
0.34765625,
0.72998046875,
0.274658203125,
0.72998046875
],
[
0.55322265625,
0.71142578125,
0.67822265625,
0.71142578125,
0.67822265625,
0.724609375,
0.55322265625,
0.724609375
],
[
0.214111328125,
0.7236328125,
0.276123046875,
0.7255859375,
0.274658203125,
0.74072265625,
0.21240234375,
0.73828125
],
[
0.27197265625,
0.7236328125,
0.279541015625,
0.7236328125,
0.279541015625,
0.72802734375,
0.27197265625,
0.72802734375
],
[
0.462646484375,
0.724609375,
0.489501953125,
0.724609375,
0.489501953125,
0.73828125,
0.462646484375,
0.73828125
],
[
0.301513671875,
0.73291015625,
0.32568359375,
0.73291015625,
0.32568359375,
0.74658203125,
0.301513671875,
0.74658203125
],
[
0.51318359375,
0.72998046875,
0.87890625,
0.73095703125,
0.87890625,
0.7451171875,
0.51318359375,
0.74462890625
],
[
0.51123046875,
0.7451171875,
0.87890625,
0.74658203125,
0.87890625,
0.76123046875,
0.51123046875,
0.76025390625
],
[
0.27783203125,
0.7529296875,
0.38623046875,
0.7529296875,
0.38623046875,
0.76416015625,
0.27783203125,
0.76416015625
],
[
0.2166748046875,
0.76123046875,
0.274658203125,
0.76123046875,
0.274658203125,
0.7724609375,
0.2166748046875,
0.7724609375
],
[
0.46337890625,
0.76025390625,
0.489501953125,
0.76025390625,
0.489501953125,
0.7724609375,
0.46337890625,
0.7724609375
],
[
0.484375,
0.76318359375,
0.50390625,
0.76318359375,
0.50390625,
0.77099609375,
0.484375,
0.77099609375
],
[
0.51318359375,
0.76318359375,
0.880859375,
0.76318359375,
0.880859375,
0.77685546875,
0.51318359375,
0.77685546875
],
[
0.5146484375,
0.779296875,
0.87744140625,
0.779296875,
0.87744140625,
0.7890625,
0.5146484375,
0.7890625
],
[
0.322509765625,
0.7685546875,
0.344970703125,
0.7685546875,
0.344970703125,
0.7822265625,
0.322509765625,
0.7822265625
],
[
0.118408203125,
0.79443359375,
0.48779296875,
0.79443359375,
0.48779296875,
0.80810546875,
0.118408203125,
0.80810546875
],
[
0.491943359375,
0.798828125,
0.5029296875,
0.798828125,
0.5029296875,
0.802734375,
0.491943359375,
0.802734375
],
[
0.51123046875,
0.79345703125,
0.87890625,
0.79345703125,
0.87890625,
0.80712890625,
0.51123046875,
0.80712890625
],
[
0.12005615234375,
0.81103515625,
0.19140625,
0.81103515625,
0.19140625,
0.822265625,
0.12005615234375,
0.822265625
],
[
0.51318359375,
0.80859375,
0.880859375,
0.81005859375,
0.880859375,
0.82373046875,
0.51318359375,
0.822265625
],
[
0.1368408203125,
0.82666015625,
0.486083984375,
0.82666015625,
0.486083984375,
0.84033203125,
0.1368408203125,
0.84033203125
],
[
0.51123046875,
0.82470703125,
0.87890625,
0.82666015625,
0.87890625,
0.84033203125,
0.51123046875,
0.83837890625
],
[
0.118408203125,
0.8427734375,
0.484375,
0.8427734375,
0.484375,
0.85546875,
0.118408203125,
0.85546875
],
[
0.51123046875,
0.8427734375,
0.87890625,
0.8427734375,
0.87890625,
0.85546875,
0.51123046875,
0.85546875
],
[
0.12176513671875,
0.859375,
0.484375,
0.859375,
0.484375,
0.869140625,
0.12176513671875,
0.869140625
],
[
0.51318359375,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.12005615234375,
0.87353515625,
0.484375,
0.87353515625,
0.484375,
0.88720703125,
0.12005615234375,
0.88720703125
],
[
0.51318359375,
0.8759765625,
0.5322265625,
0.8759765625,
0.5322265625,
0.88525390625,
0.51318359375,
0.88525390625
],
[
0.118408203125,
0.8896484375,
0.484375,
0.8896484375,
0.484375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.53076171875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.53076171875,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.2359619140625,
0.90478515625,
0.2359619140625,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.5146484375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.5146484375,
0.91845703125
]
] |
[
"导致边缘F1得分下降了4%。因此，我们决定",
"不将基数纳入最终框架",
"Zh = H[h1,h2...m]",
"(5)",
"3.3Set Property Regularisations (SPR)",
"Zr = H[r,r2... s]",
"(6)",
"简单地添加增强数据来训练模型并不能解决集",
"Zt = H[, ...]",
"(7)",
"合对齐的基本问题。在这种方法中出现了几个",
"挑战。首先，添加所有排列是不现实的，尤其",
"é =[pool(zh); pool(zr); pool(zt)) (8)",
"是在处理包含众多事件关系的长文档时，因为",
"其中H是解码器的最后一层隐藏状态。",
"训练数据将以与目标集基数的阶乘成比例的速",
"",
"[hi.hm]、[ri,rs]和[ti,.tn]分别是头",
"率增长。更重要的是，对于每个增强的示例，",
"事件、关系类型和尾事件的索引。zh，zr，zt分",
"损失函数仍将惩罚集合中未观察到的排列。这",
"别表示头事件，关系类型和尾事件的语义表",
"会使训练不稳定",
"示。pool(）表示平均池函数。é表示边的语义",
"核心挑战在于找到一种有效的方法来将线性",
"化的目标图与线性化的生成图进行比较，而不",
"表示",
"我们现在拥有两组嵌人：一组包含从目标图",
"像传统的文本生成那样依赖于严格的逐个今牌",
"比较。为了解决这个问题，我们建议对生成目",
"中提取的边嵌入，另一组包含从生成的图中派",
"标进行修改。由于线性化图本质上表示图的边",
"生的边嵌入。从本质上讲，它们可以被视为表",
"集，我们可以将图的比较问题简化为集合比较",
"示空间中的两组点。因此，我们可以通过测量",
"问题。我们的方法涉及几个组件。首先，我们",
"表示空间中两个点集（流形）之间的距离来测",
"量两个图的相似性。Hausdorff距离最初定义为",
"添加一组基数正则化，以鼓励模型生成足够数",
"测量度量空间内两个子集之间的间距，最近在",
"量的时间关系边。然后，我们引人重复正则化",
"机器学习中发现了用于测量两组嵌入（Schutze",
"来惩罚边缘集中元素的任何重复。最后：我们",
"etal.,2012；Wangetal.,2023）之间的距离的应",
"设计了一个集合配正则化，用于评估自标边",
"集中的元素与生成的边集中的元素之间的语义",
"用。我们计算平均Hausdorff距离作为度量",
"相似性。总的来说，上述正则化被称为集合属",
"性正则化(SPR)。它们通过加权平均与令牌级",
"",
"I min dcos(e',e)",
"交叉熵损失相结合",
"dH(8,)",
"=",
"",
"ee",
"要计算集合属性正则化，需要首先从给定训",
"e'Es",
"练输人的语言模型中采样图形字符串。然后",
"I",
"min dcos(el,é)",
"+",
"eeeees",
"(9)",
"将此序列解析为边E的列表，其中每个边e是",
"由头事件，关系类型和尾事件（h,r,t）组成的",
"三元组。使用基于规则的解析器进行解析，该",
"",
"其中边缘对的距离由余弦距离dcos（）计算",
"解析器将图形文本字符串转换为结构化数据表",
"当模型以与目标序列不同的顺序生成集合元",
"示。当边缘加载到结构化列表中时，可以计算",
"素时，令牌级交叉损失将很高。如果模型生",
"边缘和重复边缘的数量。令&表示E中所有",
"成更多正确的元素作为顺序错误的序列的后",
"唯一边的集合。集合基数正则化和重复正则化",
"缀，则损失值可能会进一步增加。但是，SPR",
"的值可以计算如下：",
"将具有较低的值，从而减轻了由令牌级交叉熵",
"损失引起的生成更多元素的沮丧",
"&={eleEE)",
"(2) ",
"3.4Fine-tuning with Set Property",
"",
"E-8",
"Regularisations",
"Rdupl =",
"",
"(3)",
"[]",
"与基于并行解码(Carion etal.,2020；Tanetal.",
"2021b）的集合预测方法不同，SPR不能直接用",
"abs(8-8D)",
"Rcard =",
"(4) ",
"",
"作自回归生成中的主要目标。这有两个主要原",
"因。第一个原因是获得SPR需要从解码器采",
"[8]",
"其中函数abs(）表示取绝对值，'表示地面实",
"",
"样，这会大大降低训练速度。此外，第二个原",
"况边缘集",
"因是语言模型将难以准确地生成DOT格式的",
"为了计算集合匹配正则化，我们通过比较两",
"序列，因为学习这种格式的令牌依赖需要语言",
"个集合中边的语义相似性来评估生成的集合和",
"建模目标。因此，序列解析器将无法识别序列",
"自标集合之间的相似性。我们将解码器对相应",
"中的任何有效边，从而导致高SPR值并阻碍训",
"令牌的表示的最后一层作为事件和关系类型的",
"练",
"语义表示。然后，我们将这些表示连接为每个",
"为了避免上述问题，我们在一定数量的微调",
"边的语义表示：",
"迭代之后引入了SPR。一且模型在生成正确的"
] |
{
"arxiv_id": "2404.01532",
"language": "zh",
"page_idx": 4
}
|
||
2204.07135-zh-1
|
2204.07135-zh-1.jpg
|
用于大规模会话 AI 系统中技能路由的可扩展和强大的自学习
$\textbf{Mohammad Kachuee, Jinseok Nam, Sarthak Ahuja, Jin- Myung Won, Sungjin Lee}$
Amazon Alexa AI
Seattle, WA
## Abstract
Skill routing is an important component in large-scale conversational systems. In contrast to traditional rule-based skill routing, state-ofthe-art systems use a model-based approach to enable natural conversations. To provide supervision signal required to train such models, ideas such as human annotation, replication of a rule-based system, relabeling based on user paraphrases, and bandit-based learning were suggested. However, these approaches: (a) do not scale in terms of the number of skills and skill on-boarding, (b) require a very costly expert annotation/rule-design, (c) introduce risks in the user experience with each model update. In this paper, we present a scalable selflearning approach to explore routing alternatives without causing abrupt policy changes that break the user experience, learn from the user interaction, and incrementally improve the routing via frequent model refreshes. To enable such robust frequent model updates, we suggest a simple and effective approach that ensures controlled policy updates for individual domains, followed by an off-policy evaluation for making deployment decisions without any need for lengthy A/B experimentation. We conduct various offline and online A/B experiments on a commercial large-scale conversational system to demonstrate the effectiveness of the proposed method in real-world production settings.
### 1 $\textbf{Introduction}$
大型智能会话系统，例如 AppleSiri, AmazonAlexa, GoogleAssistant 和 MicrosoftCortana , 是从传统人机交互过渡到无缝自然交互的组成部分。会话系统是多个组件的复杂相互作用， 从硬件和信号处理块到机器学习模型。Figure_ 1 显示了处理用户请求的主要处理步骤的概述： $(i)$自动语音识别 (ASR)块转录了话语，同时生成了转录置信度信号和其他语音特征，例如用户的情感$(ii)$ ,自然语言处理 (NLU)根据用户意图以及指定的实体分辨率和插槽生成
<FigureHere>
Figure 1: An overview of the major processing steps to
handle a user request in a conversational system
了一组排名的解释。。为了提供训练技能路由模型所需的监督，文献中提出了不同的方法例如复制基于规则的系统，使用人工注释以及基于用户释义的重新标记$\overline{\mathrm{E}}($Park et al., 2020by Sarikaya, 2017; Sammut, 2001)。使用人工注释非常昂贵，并且周转时间长，这对于经常引人新技能的现实世界的大型系统来说是不切实际的。另一方面，诸如 Park et al.(2020b)引人的重新标记方法仅限于我们观察到足够高精度的重新短语的情况
从可伸缩性和周转时间的角度来看，训练模型然后在每次模型部署之前进行长时间的 A/B 实验的传统方法导致模型更新频率有限，通常不足以跟上新技能的引入和其他流量变化。另一种方法是将问题表述为上下文强盗，并直接旨在最大化用户满意度 (Karampatziakis et al. 2019)。这种方法可以在监督方面更具可扩展性，因为用户满意度已经是会话系统中的一个既定指标 (Kachuee et al.,2021)。此外，可以使用非政策评估来减少进行 A/B 实验的需要。但是，在大规模生产系统中，仅依靠用户满意度最大化可能会导致由于强盗探索甚至在非策略学习中估计错误而导致的不稳定性E (Sachdeva et al., 2020; Joachims et al., 2018)。本文提出了一种新颖的基于上下文强盗学习的自学习方法，以不断探索替代决策，获取用户反
|
[
[
0.181396484375,
0.0955810546875,
0.81787109375,
0.0955810546875,
0.81787109375,
0.10986328125,
0.181396484375,
0.10986328125
],
[
0.1485595703125,
0.1436767578125,
0.85791015625,
0.1455078125,
0.85791015625,
0.1602783203125,
0.1485595703125,
0.157958984375
],
[
0.42822265625,
0.162109375,
0.57666015625,
0.162109375,
0.57666015625,
0.177001953125,
0.42822265625,
0.177001953125
],
[
0.453369140625,
0.1776123046875,
0.5498046875,
0.1776123046875,
0.5498046875,
0.1923828125,
0.453369140625,
0.1923828125
],
[
0.264404296875,
0.254638671875,
0.340087890625,
0.254638671875,
0.340087890625,
0.2666015625,
0.264404296875,
0.2666015625
],
[
0.7109375,
0.254150390625,
0.744140625,
0.254150390625,
0.744140625,
0.2666015625,
0.7109375,
0.2666015625
],
[
0.56689453125,
0.2724609375,
0.6591796875,
0.2724609375,
0.6591796875,
0.283203125,
0.56689453125,
0.283203125
],
[
0.1435546875,
0.27978515625,
0.45751953125,
0.28076171875,
0.45751953125,
0.295654296875,
0.1435546875,
0.29443359375
],
[
0.7109375,
0.27783203125,
0.85400390625,
0.27783203125,
0.85400390625,
0.2880859375,
0.7109375,
0.2880859375
],
[
0.55908203125,
0.287353515625,
0.5693359375,
0.287353515625,
0.5693359375,
0.292236328125,
0.55908203125,
0.292236328125
],
[
0.14697265625,
0.297607421875,
0.45751953125,
0.297607421875,
0.45751953125,
0.30810546875,
0.14697265625,
0.30810546875
],
[
0.62890625,
0.303955078125,
0.66748046875,
0.303955078125,
0.66748046875,
0.31640625,
0.62890625,
0.31640625
],
[
0.7841796875,
0.30712890625,
0.79248046875,
0.30712890625,
0.79248046875,
0.312255859375,
0.7841796875,
0.312255859375
],
[
0.7109375,
0.3154296875,
0.74462890625,
0.3154296875,
0.74462890625,
0.327880859375,
0.7109375,
0.327880859375
],
[
0.1435546875,
0.30810546875,
0.459228515625,
0.308837890625,
0.459228515625,
0.32373046875,
0.1435546875,
0.322509765625
],
[
0.52294921875,
0.31640625,
0.55517578125,
0.318359375,
0.55322265625,
0.330810546875,
0.5224609375,
0.328857421875
],
[
0.14697265625,
0.325927734375,
0.456787109375,
0.325927734375,
0.456787109375,
0.336181640625,
0.14697265625,
0.336181640625
],
[
0.634765625,
0.327880859375,
0.64404296875,
0.327880859375,
0.64404296875,
0.330810546875,
0.634765625,
0.330810546875
],
[
0.7080078125,
0.328857421875,
0.80029296875,
0.328857421875,
0.80029296875,
0.339111328125,
0.7080078125,
0.339111328125
],
[
0.145263671875,
0.33837890625,
0.45751953125,
0.33837890625,
0.45751953125,
0.3515625,
0.145263671875,
0.3515625
],
[
0.57080078125,
0.333251953125,
0.59423828125,
0.333251953125,
0.59423828125,
0.33740234375,
0.57080078125,
0.33740234375
],
[
0.634765625,
0.341552734375,
0.6484375,
0.341552734375,
0.6484375,
0.344482421875,
0.634765625,
0.344482421875
],
[
0.7109375,
0.34033203125,
0.81689453125,
0.34033203125,
0.81689453125,
0.347412109375,
0.7109375,
0.347412109375
],
[
0.81689453125,
0.342529296875,
0.84619140625,
0.342529296875,
0.84619140625,
0.3466796875,
0.81689453125,
0.3466796875
],
[
0.145263671875,
0.3515625,
0.456787109375,
0.3515625,
0.456787109375,
0.365234375,
0.145263671875,
0.365234375
],
[
0.7060546875,
0.3486328125,
0.8193359375,
0.3486328125,
0.8193359375,
0.359375,
0.7060546875,
0.359375
],
[
0.1485595703125,
0.3681640625,
0.456787109375,
0.3681640625,
0.456787109375,
0.37890625,
0.1485595703125,
0.37890625
],
[
0.78662109375,
0.365234375,
0.79443359375,
0.37060546875,
0.78662109375,
0.3759765625,
0.779296875,
0.37060546875
],
[
0.5224609375,
0.377685546875,
0.5576171875,
0.377685546875,
0.5576171875,
0.39013671875,
0.5224609375,
0.39013671875
],
[
0.70947265625,
0.3759765625,
0.7958984375,
0.377685546875,
0.7958984375,
0.392578125,
0.70947265625,
0.39013671875
],
[
0.51904296875,
0.393798828125,
0.6455078125,
0.393798828125,
0.6455078125,
0.40673828125,
0.51904296875,
0.40673828125
],
[
0.145263671875,
0.38134765625,
0.45751953125,
0.38134765625,
0.45751953125,
0.394287109375,
0.145263671875,
0.394287109375
],
[
0.14697265625,
0.394287109375,
0.45751953125,
0.394287109375,
0.45751953125,
0.407958984375,
0.14697265625,
0.407958984375
],
[
0.70947265625,
0.394287109375,
0.8154296875,
0.394287109375,
0.8154296875,
0.405029296875,
0.70947265625,
0.405029296875
],
[
0.14697265625,
0.4091796875,
0.45751953125,
0.4091796875,
0.45751953125,
0.4228515625,
0.14697265625,
0.4228515625
],
[
0.5205078125,
0.4091796875,
0.61376953125,
0.4091796875,
0.61376953125,
0.42041015625,
0.5205078125,
0.42041015625
],
[
0.70947265625,
0.4091796875,
0.85400390625,
0.4091796875,
0.85400390625,
0.419189453125,
0.70947265625,
0.419189453125
],
[
0.145263671875,
0.4228515625,
0.459228515625,
0.4228515625,
0.459228515625,
0.435791015625,
0.145263671875,
0.435791015625
],
[
0.1485595703125,
0.439453125,
0.455078125,
0.439453125,
0.455078125,
0.449462890625,
0.1485595703125,
0.449462890625
],
[
0.51318359375,
0.440673828125,
0.880859375,
0.441162109375,
0.880859375,
0.454833984375,
0.51318359375,
0.45361328125
],
[
0.14697265625,
0.451904296875,
0.45751953125,
0.451904296875,
0.45751953125,
0.465576171875,
0.14697265625,
0.465576171875
],
[
0.51123046875,
0.454833984375,
0.84033203125,
0.454833984375,
0.84033203125,
0.468505859375,
0.51123046875,
0.468505859375
],
[
0.14697265625,
0.466064453125,
0.459228515625,
0.466064453125,
0.459228515625,
0.479736328125,
0.14697265625,
0.479736328125
],
[
0.1435546875,
0.479736328125,
0.459228515625,
0.478515625,
0.459228515625,
0.493408203125,
0.1435546875,
0.49462890625
],
[
0.1435546875,
0.49462890625,
0.45751953125,
0.493408203125,
0.45751953125,
0.50830078125,
0.1435546875,
0.5087890625
],
[
0.515625,
0.50048828125,
0.880859375,
0.50048828125,
0.880859375,
0.51416015625,
0.515625,
0.51416015625
],
[
0.145263671875,
0.5087890625,
0.45751953125,
0.5087890625,
0.45751953125,
0.5224609375,
0.145263671875,
0.5224609375
],
[
0.1435546875,
0.521484375,
0.45751953125,
0.5224609375,
0.45751953125,
0.5361328125,
0.1435546875,
0.53515625
],
[
0.5146484375,
0.5166015625,
0.87744140625,
0.5166015625,
0.87744140625,
0.52978515625,
0.5146484375,
0.52978515625
],
[
0.14697265625,
0.5361328125,
0.45751953125,
0.53759765625,
0.45751953125,
0.55029296875,
0.14697265625,
0.5498046875
],
[
0.5146484375,
0.533203125,
0.880859375,
0.533203125,
0.880859375,
0.54638671875,
0.5146484375,
0.54638671875
],
[
0.145263671875,
0.55029296875,
0.45751953125,
0.55029296875,
0.45751953125,
0.56396484375,
0.145263671875,
0.56396484375
],
[
0.51123046875,
0.54736328125,
0.87890625,
0.54736328125,
0.87890625,
0.56103515625,
0.51123046875,
0.56103515625
],
[
0.145263671875,
0.56396484375,
0.456787109375,
0.5654296875,
0.456787109375,
0.5791015625,
0.145263671875,
0.57763671875
],
[
0.51318359375,
0.56396484375,
0.880859375,
0.56396484375,
0.880859375,
0.57763671875,
0.51318359375,
0.57763671875
],
[
0.145263671875,
0.580078125,
0.45751953125,
0.580078125,
0.45751953125,
0.59326171875,
0.145263671875,
0.59326171875
],
[
0.5146484375,
0.58056640625,
0.880859375,
0.58056640625,
0.880859375,
0.59423828125,
0.5146484375,
0.59423828125
],
[
0.145263671875,
0.59326171875,
0.459228515625,
0.59326171875,
0.459228515625,
0.60693359375,
0.145263671875,
0.60693359375
],
[
0.51318359375,
0.5966796875,
0.880859375,
0.5966796875,
0.880859375,
0.60986328125,
0.51318359375,
0.60986328125
],
[
0.14697265625,
0.60791015625,
0.45751953125,
0.60791015625,
0.45751953125,
0.62158203125,
0.14697265625,
0.62158203125
],
[
0.51123046875,
0.61083984375,
0.880859375,
0.6123046875,
0.880859375,
0.626953125,
0.51123046875,
0.6259765625
],
[
0.1435546875,
0.62158203125,
0.45751953125,
0.62060546875,
0.45751953125,
0.63525390625,
0.1435546875,
0.6357421875
],
[
0.51318359375,
0.62744140625,
0.87890625,
0.62744140625,
0.87890625,
0.64111328125,
0.51318359375,
0.64111328125
],
[
0.14697265625,
0.6357421875,
0.45751953125,
0.6357421875,
0.45751953125,
0.6494140625,
0.14697265625,
0.6494140625
],
[
0.5146484375,
0.64453125,
0.64208984375,
0.64453125,
0.64208984375,
0.6591796875,
0.5146484375,
0.6591796875
],
[
0.145263671875,
0.65087890625,
0.45751953125,
0.65087890625,
0.45751953125,
0.66455078125,
0.145263671875,
0.66455078125
],
[
0.14697265625,
0.66650390625,
0.456787109375,
0.66650390625,
0.456787109375,
0.6767578125,
0.14697265625,
0.6767578125
],
[
0.5322265625,
0.6630859375,
0.880859375,
0.6630859375,
0.880859375,
0.677734375,
0.5322265625,
0.677734375
],
[
0.1485595703125,
0.68115234375,
0.456787109375,
0.68115234375,
0.456787109375,
0.69140625,
0.1485595703125,
0.69140625
],
[
0.5146484375,
0.6796875,
0.880859375,
0.6796875,
0.880859375,
0.693359375,
0.5146484375,
0.693359375
],
[
0.14697265625,
0.693359375,
0.456787109375,
0.693359375,
0.456787109375,
0.70361328125,
0.14697265625,
0.70361328125
],
[
0.51318359375,
0.69482421875,
0.87890625,
0.69482421875,
0.87890625,
0.7080078125,
0.51318359375,
0.7080078125
],
[
0.1435546875,
0.7060546875,
0.2333984375,
0.7080078125,
0.2333984375,
0.7216796875,
0.1435546875,
0.7197265625
],
[
0.5146484375,
0.7119140625,
0.880859375,
0.7119140625,
0.880859375,
0.7255859375,
0.5146484375,
0.7255859375
],
[
0.5205078125,
0.732421875,
0.52880859375,
0.732421875,
0.52880859375,
0.736328125,
0.5205078125,
0.736328125
],
[
0.5283203125,
0.72802734375,
0.880859375,
0.72802734375,
0.880859375,
0.7412109375,
0.5283203125,
0.7412109375
],
[
0.1158447265625,
0.7353515625,
0.145263671875,
0.7353515625,
0.145263671875,
0.74658203125,
0.1158447265625,
0.74658203125
],
[
0.1485595703125,
0.73291015625,
0.264404296875,
0.7353515625,
0.264404296875,
0.7490234375,
0.1485595703125,
0.74658203125
],
[
0.5146484375,
0.7421875,
0.880859375,
0.74365234375,
0.880859375,
0.75830078125,
0.5146484375,
0.75732421875
],
[
0.118408203125,
0.76025390625,
0.489501953125,
0.76025390625,
0.489501953125,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.118408203125,
0.77685546875,
0.484375,
0.77685546875,
0.484375,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.51318359375,
0.77685546875,
0.87158203125,
0.77685546875,
0.87158203125,
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,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.51318359375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.12005615234375,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.12005615234375,
0.83837890625
],
[
0.5146484375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.489501953125,
0.84130859375,
0.489501953125,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.87890625,
0.84130859375,
0.87890625,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.484375,
0.8896484375,
0.484375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.5146484375,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.5146484375,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"用于大规模会话AI系统中技能路由的可扩展和强大的自学习",
"Mohammad Kachuee, Jinseok Nam, Sarthak Ahuja, Jin-Myung Won, Sungjin Lee",
"Amazon Alexa AI",
"Seattle, WA",
"Abstract",
"ASR",
"Turnon the light",
"Skill routing is an important component in",
"Transcript:\"turn on the light'",
"",
"large-scale conversational systems.In contrast",
"Agent",
"Y",
"NLU",
"to traditional rule-based skill routing, state-of-",
"User",
"the-art systems use amodel-based approachto",
"",
"Interpretation#1:",
"enable natural conversations. To provide su-",
"K",
"-",
"intent:smart-home.t",
"Un-o",
"pervision signal required to train such models,",
"slots: (device:\"light\")",
"ideassuchashumanannotation,replicationof",
"Y",
"Skill",
"Skill Routing",
"generate response:\"ok\"",
"a rule-based system, relabeling based on user",
"paraphrases, and bandit-based learning were",
"skill:1p-smart-home",
"suggested. However, these approaches: (a) do",
"switchthe lighton",
"Interp.:smart-home,turn-on",
"not scale in terms of the number of skills and",
"skillon-boarding,(b)requireaverycostlyex",
"Figure I: An overview of the major processing steps to",
"pert annotation/rule-design, (c) introduce risks",
"handle a user request in a conversational system.",
"in the user experience with each model up-",
"date. In this paper, we present a scalable self-",
"learning approach to explore routing alterna-",
"了一组排名的解释。。为了提供训练技能路由",
"tives without causing abrupt policy changes",
"that break the user experience, learn from the",
"模型所需的监督，文献中提出了不同的方法",
"user interaction, and incrementally improve",
"例如复制基于规则的系统，使用人工注释以及",
"the routing via frequent model refreshes. To",
"基于用户释义的重新标记EParketal.，2020b",
"enable such robust frequent model updates, we",
"Sarikaya,2017；Sammut,2001）。使用人工注释",
"suggest a simple and effective approach that",
"非常昂贵，并且周转时间长，这对于经常引入",
"ensures controlled policy updates for individ-",
"新技能的现实世界的大型系统来说是不切实际",
"ual domains, followed by an off-policy evalua-",
"的。另一方面，诸如Parketal.（2020b）引人的",
"tion for making deployment decisions without",
"重新标记方法仅限于我们观察到足够高精度的",
"any need for lengthy A/B experimentation. We",
"重新短语的情况",
"conduct various offline and online A/B exper",
"iments ona commercial large-scaleconversa",
"从可伸缩性和周转时间的角度来看，训练模",
"tionalsystemtodemonstratetheeffectiveness",
"型然后在每次模型部署之前进行长时间的A/B",
"oftheproposedmethodinreal-worldproduc",
"实验的传统方法导致模型更新频率有限，通常",
"tion settings.",
"不足以跟上新技能的引人和其他流量变化。另",
"",
"-种方法是将问题表述为上下文强盗，并直接",
"1",
"Introduction",
"旨在最大化用户满意度（Karampatziakisetal.，",
"大型智能会话系统，例如AppleSiri，Ama-",
"2019）。这种方法可以在监督方面更具可扩展",
"zonAlexa，GoogleAssistant 和 MicrosoftCortana",
"性，因为用户满意度已经是会话系统中的一",
"是从传统人机交互过渡到无缝自然交互的组成",
"个既定指标（Kachueeetal.,2021）。此外，可",
"部分。会话系统是多个组件的复杂相互作用，",
"以使用非政策评估来减少进行A/B实验的需",
"从硬件和信号处理块到机器学习模型。Figure",
"要。但是，在大规模生产系统中，仅依靠用户",
"1显示了处理用户请求的主要处理步骤的概述：",
"满意度最大化可能会导致由于强盗探索甚至",
"(i）自动语音识别（ASR）块转录了话语，同",
"在非策略学习中估计错误而导致的不稳定性E",
"时生成了转录置信度信号和其他语音特征，例",
"(Sachdeva et al., 2020; Joachims et al., 2018) 。本",
"如用户的情感（i），自然语言处理（NLU）根",
"文提出了一种新颖的基于上下文强盗学习的自",
"据用户意图以及指定的实体分辨率和插槽生成",
"学习方法，以不断探索替代决策，获取用户反"
] |
{
"arxiv_id": "2204.07135",
"language": "zh",
"page_idx": 1
}
|
||
2204.07135-zh-2
|
2204.07135-zh-2.jpg
|
用户体验中断。建议的方法简单而有效，因为后进行广泛的测试，而不是传统的 A/B 分析。此方法可在实际服务设置中实现低周转时间模型刷新，同时保持业务关键型用例的最佳用户体验。为了验证所提出方法的有效性，我们在真实的大型商业对话系统中对真实的客户流量进行了广泛的离线和在线 A/B 实验。
## 2 $\textbf{Related Works}$
会话系统中的第一代技能路由使用基于规则的系统来服务用户的请求。这些规则可以在多个级别和不同的信号上定义，例如预先录制的语音，话语记录或 NLU 解释 (Sarikaya, 2017; Sammut,2001)。但是，基于规则的实现无法概括和理解自然语言变体。基于规则的路由系统的另一个重要缺点是在处理大量相互竞争的技能和规则时出现的可伸缩性问题 (Jadhav and Thorat, 2020; Agostaro et al., 2005)。基于模型的会话系统使用机器学习模型来理解用户的话语并选择最佳技能来满足请求。基于模型的系统可以超越基于规则的系统的能力，因为机器学习模型可以潜在地理解请求的语义含义 (Park et al., 2020b)。请注意，尽管承诺更好的泛化和可扩展性，但在现实世界的大规模系统中，从基于规则的方法到基于模型的方法的过渡具有挑战性，因为复杂的模型因缺乏鲁棒性和可解释性而闻名。
为模型训练提供监督是训练技能路由模型的重要考虑因素。基于规则的系统可用于向模型提供监督信号，希望模型能够在提供的训练示例之外泛化。考虑到鲁棒性方面，这种复制目标相对简单且可取；但是，实际上，它可能不会超出基于规则的方法E(Li et al., 2021)的范围。。另一项工作是基于通过检测重新措辞的话语来重新标记样本$\overline{\mathrm{F}}($Park et al.,2020b),因为当代理无法正确响应时，用户倾向于重新措辞并重复。但是，这种重新标记仅涵盖仅提供有限路由改进机会的流量子集的校正模式。例如，用户可能决定放弃对话，而不是转述相同的请求
考虑到用户满意度是对话系统的主要目标，
可以使用满意度作为监督信号来指导路由决 词向量和双向LSTM 对话语文本进行编码。通策。对话系统中的用户满意度测量和预测在文 过求和操作减少嵌人向量的序列，并连接上下献中得到了广泛的研究(Kachuee et al.,2021;
馈并学习改进技能路由决策。由于频繁的模型 Park et al., 2020a; Bodigutla et al., 2019; Jiang 刷新是自学习循环的一部分，因此我们建议采 et al.,2015)。一种可能的方法是将技能路由用混合策略体系结构来控制策略偏差，以确保 问题表述为上下文强盗问题，旨在最大程度地对用户体验的一致和强大的改进。e。, 不会 提高用户满意度$\overline{\mathrm{E}}($Karampatziakis et al., 2019) 导致突然的策略更改，从而导致某些用例上的 。它可以在用户-代理交互中以用户体验为指
导，积极探索替代候选人。然而，在现实世界
它支持每个 NLU 意图的不同级别的鲁棒性-敏 的生产系统中，控制代理行为的变化至关重感性此外，拟议的方法依赖于非政策评估，然 要，因为强盗学习中的过度探索或偏离政策估
计错误可能会导致意外行为。
## 3 $\textbf{Proposed Method}$
## 3.1 Problem Definition
我们考虑会话系统中技能路由的一般问题。具体来说，我们定义了不同的 NLU 解释对(例如。g。、域、意图、插槽等。)和技能(e。g。天气技能或购物技能)作为路由候选，i。e。我们政策的行动空间。给定一组路由候选及其相应的上下文信号，在向量空间中编码为$X= \{ x_1\ldots x_T| x_i\in \mathcal{R} ^d\}$ ,技能路由代理的任务是选择一个路由候选$a\in\{1\ldots T\}$来为用户服务
此外，我们假设存在一个当前的，不一定是最优的，由$\Pi_0(a|X)$ 表示的策略。任务是从非策略设置中从当前策略交互中收集的经验中学习，以训练由$\theta,\Pi_\theta(a|X)$参数化的新策略， 旨在提高用户满意度。在这里，在采取行动之后，代理观察到奖励信号$r$ ,这是用户满意度的度量。奖励信号本身由多个组件组成，例如隐式/显式用户反馈和机器学习模型
## 3.2 Self-Learning Process
图 2 显示了建议的自学习过程的概述。首先， 从当前策略中收集一批已记录的交互(在图中用$HP_i$表示)。然后，我们使用 off-policy 学习来更新策略，使用记录流量的拆分(第 3.4节和第 3.5 节)。在实际部署之前和之后对新策略进行评估，从而可以使用护栏指标来做出部署决策以及跟踪新模型的实际在线性能(第3.6节)
## 3.3 Model Architecture
图 3 显示了模型体系结构的概述。模型的输人包括每个候选人的 NLU 解释和技能，以及ASR 转录和一组不同的上下文信号 (例如。g。ASR 置信度、设备类型、设备状态等。)在候选人之间共享
我们使用特征大小与唯一值数量的平方根成比例的嵌入矩阵对分类特征进行编码。使用单文信号以获得最终表示 i。e。$x\in\mathcal{R}^d$ 。最后，
|
[
[
0.12005615234375,
0.0872802734375,
0.48779296875,
0.0872802734375,
0.48779296875,
0.1009521484375,
0.12005615234375,
0.1009521484375
],
[
0.51123046875,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51123046875,
0.1009521484375
],
[
0.12176513671875,
0.1051025390625,
0.483642578125,
0.1051025390625,
0.483642578125,
0.11517333984375,
0.12176513671875,
0.11517333984375
],
[
0.51123046875,
0.10272216796875,
0.87744140625,
0.10272216796875,
0.87744140625,
0.11639404296875,
0.51123046875,
0.11639404296875
],
[
0.12176513671875,
0.12054443359375,
0.484375,
0.12054443359375,
0.484375,
0.1312255859375,
0.12176513671875,
0.1312255859375
],
[
0.5146484375,
0.12054443359375,
0.87744140625,
0.12054443359375,
0.87744140625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.1175537109375,
0.1341552734375,
0.484375,
0.1353759765625,
0.484375,
0.149658203125,
0.1175537109375,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.12005615234375,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.1650390625,
0.12005615234375,
0.1650390625
],
[
0.515625,
0.155029296875,
0.53662109375,
0.155029296875,
0.53662109375,
0.162109375,
0.515625,
0.162109375
],
[
0.5322265625,
0.152587890625,
0.87744140625,
0.152587890625,
0.87744140625,
0.163330078125,
0.5322265625,
0.163330078125
],
[
0.118408203125,
0.16748046875,
0.486083984375,
0.16748046875,
0.486083984375,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.87890625,
0.16748046875,
0.87890625,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.12005615234375,
0.182861328125,
0.486083984375,
0.182861328125,
0.486083984375,
0.196533203125,
0.12005615234375,
0.196533203125
],
[
0.51318359375,
0.182861328125,
0.87890625,
0.182861328125,
0.87890625,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.118408203125,
0.1995849609375,
0.48779296875,
0.1995849609375,
0.48779296875,
0.213134765625,
0.118408203125,
0.213134765625
],
[
0.509765625,
0.198974609375,
0.87890625,
0.198974609375,
0.87890625,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.118408203125,
0.215576171875,
0.490234375,
0.215576171875,
0.490234375,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.51123046875,
0.215576171875,
0.744140625,
0.215576171875,
0.744140625,
0.2286376953125,
0.51123046875,
0.2286376953125
],
[
0.118408203125,
0.232177734375,
0.486083984375,
0.232177734375,
0.486083984375,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.118408203125,
0.2476806640625,
0.489501953125,
0.2476806640625,
0.489501953125,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.50732421875,
0.243408203125,
0.7001953125,
0.243408203125,
0.7001953125,
0.25830078125,
0.50732421875,
0.25830078125
],
[
0.490234375,
0.2493896484375,
0.50390625,
0.2476806640625,
0.50537109375,
0.2529296875,
0.491943359375,
0.254638671875
],
[
0.118408203125,
0.26416015625,
0.48779296875,
0.26416015625,
0.48779296875,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.5087890625,
0.268310546875,
0.70947265625,
0.26953125,
0.70947265625,
0.283935546875,
0.5087890625,
0.283203125
],
[
0.1175537109375,
0.279052734375,
0.484375,
0.279052734375,
0.484375,
0.292236328125,
0.1175537109375,
0.292236328125
],
[
0.118408203125,
0.29638671875,
0.403076171875,
0.29638671875,
0.403076171875,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.51318359375,
0.292236328125,
0.8759765625,
0.292236328125,
0.8759765625,
0.302734375,
0.51318359375,
0.302734375
],
[
0.51123046875,
0.30712890625,
0.87890625,
0.30712890625,
0.87890625,
0.320556640625,
0.51123046875,
0.320556640625
],
[
0.11419677734375,
0.32666015625,
0.2822265625,
0.32666015625,
0.2822265625,
0.341552734375,
0.11419677734375,
0.341552734375
],
[
0.51123046875,
0.32373046875,
0.8818359375,
0.32470703125,
0.8818359375,
0.339111328125,
0.51123046875,
0.33837890625
],
[
0.51123046875,
0.339111328125,
0.8818359375,
0.34033203125,
0.8818359375,
0.35400390625,
0.51123046875,
0.352783203125
],
[
0.118408203125,
0.35693359375,
0.484375,
0.35693359375,
0.484375,
0.36767578125,
0.118408203125,
0.36767578125
],
[
0.51318359375,
0.355712890625,
0.880859375,
0.355712890625,
0.880859375,
0.369384765625,
0.51318359375,
0.369384765625
],
[
0.118408203125,
0.371826171875,
0.48779296875,
0.371826171875,
0.48779296875,
0.385498046875,
0.118408203125,
0.385498046875
],
[
0.51123046875,
0.372314453125,
0.880859375,
0.372314453125,
0.880859375,
0.385986328125,
0.51123046875,
0.385986328125
],
[
0.1175537109375,
0.385986328125,
0.486083984375,
0.38720703125,
0.486083984375,
0.402099609375,
0.1175537109375,
0.40087890625
],
[
0.51123046875,
0.388427734375,
0.880859375,
0.388427734375,
0.880859375,
0.402099609375,
0.51123046875,
0.402099609375
],
[
0.118408203125,
0.40380859375,
0.486083984375,
0.40380859375,
0.486083984375,
0.41748046875,
0.118408203125,
0.41748046875
],
[
0.51318359375,
0.405029296875,
0.880859375,
0.405029296875,
0.880859375,
0.418701171875,
0.51318359375,
0.418701171875
],
[
0.118408203125,
0.421630859375,
0.483642578125,
0.421630859375,
0.483642578125,
0.431640625,
0.118408203125,
0.431640625
],
[
0.51318359375,
0.42041015625,
0.54833984375,
0.42041015625,
0.54833984375,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.12005615234375,
0.435791015625,
0.484375,
0.435791015625,
0.484375,
0.449462890625,
0.12005615234375,
0.449462890625
],
[
0.5322265625,
0.435791015625,
0.880859375,
0.435791015625,
0.880859375,
0.45068359375,
0.5322265625,
0.45068359375
],
[
0.12005615234375,
0.45361328125,
0.483642578125,
0.45361328125,
0.483642578125,
0.46435546875,
0.12005615234375,
0.46435546875
],
[
0.51318359375,
0.453125,
0.880859375,
0.453125,
0.880859375,
0.466064453125,
0.51318359375,
0.466064453125
],
[
0.12005615234375,
0.47021484375,
0.483642578125,
0.47021484375,
0.483642578125,
0.48095703125,
0.12005615234375,
0.48095703125
],
[
0.51318359375,
0.468505859375,
0.87890625,
0.468505859375,
0.87890625,
0.482177734375,
0.51318359375,
0.482177734375
],
[
0.1175537109375,
0.485107421875,
0.484375,
0.485107421875,
0.484375,
0.498779296875,
0.1175537109375,
0.498779296875
],
[
0.51318359375,
0.48681640625,
0.875,
0.48681640625,
0.875,
0.49755859375,
0.51318359375,
0.49755859375
],
[
0.118408203125,
0.50048828125,
0.484375,
0.50048828125,
0.484375,
0.51416015625,
0.118408203125,
0.51416015625
],
[
0.5146484375,
0.50048828125,
0.87890625,
0.501953125,
0.87890625,
0.515625,
0.5146484375,
0.51416015625
],
[
0.12005615234375,
0.5185546875,
0.484375,
0.5185546875,
0.484375,
0.529296875,
0.12005615234375,
0.529296875
],
[
0.51318359375,
0.51953125,
0.8759765625,
0.51953125,
0.8759765625,
0.52978515625,
0.51318359375,
0.52978515625
],
[
0.118408203125,
0.533203125,
0.486083984375,
0.533203125,
0.486083984375,
0.54638671875,
0.118408203125,
0.54638671875
],
[
0.5146484375,
0.53515625,
0.87744140625,
0.53515625,
0.87744140625,
0.5458984375,
0.5146484375,
0.5458984375
],
[
0.118408203125,
0.548828125,
0.484375,
0.548828125,
0.484375,
0.5625,
0.118408203125,
0.5625
],
[
0.51318359375,
0.5498046875,
0.79248046875,
0.5498046875,
0.79248046875,
0.5634765625,
0.51318359375,
0.5634765625
],
[
0.12005615234375,
0.5654296875,
0.484375,
0.5654296875,
0.484375,
0.5791015625,
0.12005615234375,
0.5791015625
],
[
0.51123046875,
0.57763671875,
0.5615234375,
0.57763671875,
0.5615234375,
0.58837890625,
0.51123046875,
0.58837890625
],
[
0.55517578125,
0.57666015625,
0.73046875,
0.57666015625,
0.73046875,
0.59033203125,
0.55517578125,
0.59033203125
],
[
0.118408203125,
0.59716796875,
0.48779296875,
0.59716796875,
0.48779296875,
0.61083984375,
0.118408203125,
0.61083984375
],
[
0.1175537109375,
0.580078125,
0.486083984375,
0.580078125,
0.486083984375,
0.59326171875,
0.1175537109375,
0.59326171875
],
[
0.51123046875,
0.5986328125,
0.8818359375,
0.5986328125,
0.8818359375,
0.6123046875,
0.51123046875,
0.6123046875
],
[
0.118408203125,
0.61328125,
0.48779296875,
0.61328125,
0.48779296875,
0.626953125,
0.118408203125,
0.626953125
],
[
0.51318359375,
0.615234375,
0.8759765625,
0.615234375,
0.8759765625,
0.6259765625,
0.51318359375,
0.6259765625
],
[
0.118408203125,
0.62890625,
0.28955078125,
0.62890625,
0.28955078125,
0.642578125,
0.118408203125,
0.642578125
],
[
0.51318359375,
0.63134765625,
0.880859375,
0.63134765625,
0.880859375,
0.64453125,
0.51318359375,
0.64453125
],
[
0.1368408203125,
0.646484375,
0.484375,
0.646484375,
0.484375,
0.66015625,
0.1368408203125,
0.66015625
],
[
0.51123046875,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.66015625,
0.51123046875,
0.66015625
],
[
0.118408203125,
0.6630859375,
0.486083984375,
0.6630859375,
0.486083984375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.51123046875,
0.662109375,
0.880859375,
0.662109375,
0.880859375,
0.67578125,
0.51123046875,
0.67578125
],
[
0.118408203125,
0.6787109375,
0.486083984375,
0.6787109375,
0.486083984375,
0.6923828125,
0.118408203125,
0.6923828125
],
[
0.51318359375,
0.6787109375,
0.87890625,
0.6787109375,
0.87890625,
0.6923828125,
0.51318359375,
0.6923828125
],
[
0.12005615234375,
0.69482421875,
0.483642578125,
0.69482421875,
0.483642578125,
0.7080078125,
0.12005615234375,
0.7080078125
],
[
0.51123046875,
0.693359375,
0.880859375,
0.69482421875,
0.880859375,
0.708984375,
0.51123046875,
0.7080078125
],
[
0.12005615234375,
0.71142578125,
0.48779296875,
0.71142578125,
0.48779296875,
0.724609375,
0.12005615234375,
0.724609375
],
[
0.50732421875,
0.71142578125,
0.5693359375,
0.71142578125,
0.5693359375,
0.7255859375,
0.50732421875,
0.7255859375
],
[
0.1175537109375,
0.72705078125,
0.486083984375,
0.7255859375,
0.486083984375,
0.74072265625,
0.1175537109375,
0.7412109375
],
[
0.118408203125,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.509765625,
0.73828125,
0.7119140625,
0.73828125,
0.7119140625,
0.751953125,
0.509765625,
0.751953125
],
[
0.118408203125,
0.75830078125,
0.484375,
0.75830078125,
0.484375,
0.771484375,
0.118408203125,
0.771484375
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.118408203125,
0.775390625,
0.483642578125,
0.775390625,
0.483642578125,
0.7861328125,
0.118408203125,
0.7861328125
],
[
0.51318359375,
0.77685546875,
0.880859375,
0.77685546875,
0.880859375,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.118408203125,
0.79150390625,
0.486083984375,
0.79150390625,
0.486083984375,
0.8046875,
0.118408203125,
0.8046875
],
[
0.51123046875,
0.79052734375,
0.8818359375,
0.79345703125,
0.8818359375,
0.80810546875,
0.51123046875,
0.8046875
],
[
0.118408203125,
0.80810546875,
0.484375,
0.80810546875,
0.484375,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.51123046875,
0.80712890625,
0.8818359375,
0.80859375,
0.8818359375,
0.822265625,
0.51123046875,
0.82080078125
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.5146484375,
0.826171875,
0.62451171875,
0.826171875,
0.62451171875,
0.83740234375,
0.5146484375,
0.83740234375
],
[
0.1175537109375,
0.8388671875,
0.175537109375,
0.8388671875,
0.175537109375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.53076171875,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.53076171875,
0.85498046875
],
[
0.1368408203125,
0.85693359375,
0.489501953125,
0.85693359375,
0.489501953125,
0.87060546875,
0.1368408203125,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.12176513671875,
0.87451171875,
0.48779296875,
0.87451171875,
0.48779296875,
0.88525390625,
0.12176513671875,
0.88525390625
],
[
0.51123046875,
0.87255859375,
0.87890625,
0.87255859375,
0.87890625,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.118408203125,
0.90625,
0.489501953125,
0.90625,
0.489501953125,
0.91943359375,
0.118408203125,
0.91943359375
],
[
0.51123046875,
0.90380859375,
0.8818359375,
0.90478515625,
0.8818359375,
0.91943359375,
0.51123046875,
0.91845703125
]
] |
[
"馈并学习改进技能路由决策。由于频繁的模型",
"Park et al., 2020a; Bodigutla et al., 2019; Jiang",
"刷新是自学习循环的一部分，因此我们建议采",
"etal.,2015）。一种可能的方法是将技能路由",
"用混合策略体系结构来控制策略偏差，以确保",
"问题表述为上下文强盗问题，旨在最大程度地",
"对用户体验的一致和强大的改进。。e。，不会",
"提高用户满意度E(Karampatziakisetal.，2019)",
"导致突然的策略更改，从而导致某些用例上的",
"9",
"它可以在用户-代理交互中以用户体验为指",
"用户体验中断。建议的方法简单而有效，因为",
"导，积极探索替代候选人。然而，在现实世界",
"它支持每个NLU意图的不同级别的鲁棒性-敏",
"的生产系统中，控制代理行为的变化至关重",
"感性此外，拟议的方法依赖于非政策评估，然",
"要，因为强盗学习中的过度探索或偏离政策估",
"后进行广泛的测试，而不是传统的A/B分析。",
"计错误可能会导致意外行为。",
"此方法可在实际服务设置中实现低周转时间模",
"型刷新，同时保持业务关键型用例的最佳用户",
"3Proposed Method",
"",
"体验。为了验证所提出方法的有效性，我们在",
"3.1ProblemDefinition",
"真实的大型商业对话系统中对真实的客户流量",
"进行了广泛的离线和在线A/B实验。",
"我们考虑会话系统中技能路由的一般问题。具",
"体来说，我们定义了不同的NLU解释对（例",
"2Related Works",
"如。g。、域、意图、插槽等。）和技能(e。g。、",
"天气技能或购物技能）作为路由候选，i。e。",
"会话系统中的第一代技能路由使用基于规则",
"我们政策的行动空间。给定一组路由候选及",
"的系统来服务用户的请求。这些规则可以在多",
"其相应的上下文信号，在向量空间中编码为",
"个级别和不同的信号上定义，例如预先录制的",
"X=[ai..Ta;ERd】，技能路由代理的任",
"语音，话语记录或NLU解释（Sarikaya，2017;",
"务是选择一个路由候选αE{1...T}来为用户",
"Sammut，2001）。但是，基于规则的实现无法",
"服务",
"概括和理解自然语言变体。基于规则的路由系",
"此外，我们假设存在一个当前的，不一定是",
"统的另一个重要缺点是在处理大量相互竞争",
"最优的，由IIo(αX)表示的策略。任务是从非",
"的技能和规则时出现的可伸缩性问题Jadhav",
"策略设置中从当前策略交互中收集的经验中学",
"and Thorat, 2020; Agostaro et al., 2005)。基于",
"习，以训练由e，Ie(aX）参数化的新策略",
"模型的会话系统使用机器学习模型来理解用",
"旨在提高用户满意度。在这里，在采取行动之",
"户的话语并选择最佳技能来满足请求。基于模",
"后，代理观察到奖励信号，这是用户满意度",
"型的系统可以超越基于规则的系统的能力，因",
"的度量。奖励信号本身由多个组件组成，例如",
"为机器学习模型可以潜在地理解请求的语义含",
"隐式/显式用户反馈和机器学习模型",
"义(Parketal.,2020b）。请注意，尽管承诺更好",
"3.2",
" Self-Learning Process",
"统中，从基于规则的方法到基于模型的方法的",
"的泛化和可扩展性，但在现实世界的大规模系",
"图2显示了建议的自学习过程的概述。首先，",
"过渡具有挑战性，因为复杂的模型因缺乏鲁棒",
"从当前策略中收集一批已记录的交互（在图中",
"性和可解释性而闻名。",
"用HP，表示）。然后，我们使用off-policy学",
"为模型训练提供监督是训练技能路由模型的",
"习来更新策略，使用记录流量的拆分（第3.4",
"重要考虑因素。基于规则的系统可用于向模型",
"节和第3.5节）。在实际部署之前和之后对新",
"提供监督信号，希望模型能够在提供的训练示",
"策略进行评估，从而可以使用护栏指标来做出",
"例之外泛化。考虑到鲁棒性方面，这种复制目",
"部署决策以及跟踪新模型的实际在线性能（第",
"标相对简单且可取；但是，实际上，它可能不",
"3.6节)",
"会超出基于规则的方法E(Lietal.,2021)的范",
"围。。另一项工作是基于通过检测重新措辞的",
"3.3Model Architecture",
"话语来重新标记样本E(Parketal.，2020b），因",
"图3显示了模型体系结构的概述。模型的输",
"为当代理无法正确响应时，用户倾向于重新措",
"入包括每个候选人的NLU解释和技能，以及",
"辞并重复。但是，这种重新标记仅涵盖仅提供",
"ASR转录和一组不同的上下文信号（例如。g。",
"有限路由改进机会的流量子集的校正模式。例",
"ASR置信度、设备类型、设备状态等。）在候",
"如，用户可能决定放弃对话，而不是转述相同",
"选人之间共享",
"的请求",
"我们使用特征大小与唯一值数量的平方根成",
"考虑到用户满意度是对话系统的主要目标，",
"比例的嵌入矩阵对分类特征进行编码。使用单",
"可以使用满意度作为监督信号来指导路由决",
"词向量和双向LSTM对话语文本进行编码。通",
"策。对话系统中的用户满意度测量和预测在文",
"过求和操作减少嵌入向量的序列，并连接上下",
"献中得到了广泛的研究（Kachueeetal.，202l；",
"文信号以获得最终表示i。e。αERd。最后，"
] |
{
"arxiv_id": "2204.07135",
"language": "zh",
"page_idx": 2
}
|
||
2204.07135-zh-3
|
2204.07135-zh-3.jpg
|
根据 NLU 解释置信度对编码假设集$X$进行排序，并通过双向 LSTM, 两个完全连接的层以及 softmax 激活来输出$\Pi_\theta(X)$策略的操作概率
## 3.4 Model Training
我们定义了两个训练目标：复制策略(RP)和学习策略 (LP)。RP 目标尝试训练复制记录的操作的模型。具体来说，我们将 RP 损失函数定义为最小化：
$$L_{RP}=\mathbb{E}_{X,a\sim\mathbb{D}}\sum_{i=1}^{T}-\mathbf{1}[a=i]log(\Pi_{\theta}(a|X)).$$
(1)
简而言之，(1)是一种交叉熵损失，鼓励新策略
为复制记录的操作的操作分配最高分数。我们还探索了其他替代方案，例如 KL 发散或软蒸馏目标，但发现简单的交叉熵目标非常稳定， 并显示出出色的复制性能
我们将 LP 损失函数定义为非策略上下文强
<FigureHere>
Figure 2: An overview of the self-learning process: model training, RPDR computation, pre-deployment evaluation, and post-deployment evaluation.
<FigureHere>
Figure 3: An overview of the proposed model architecture: a set of hypothesis are encoded as vectors ( $x_1\ldots x_T$ )and fed to a bi-directional LSTM which is followed by a shared MLP and a softmax layer to normalize the predicted candidate selection probabilities.
盗目标，例如反向倾向评分(IPS)目标：
(2)
$$L_{LP}=\mathbb{E}_{X,a,r\sim\mathbb{D}}=r\frac{\Pi_\theta(a|\boldsymbol{x})}{\Pi_0(a|\boldsymbol{x})},$$
$r$是观察到的采取操作的奖励$a$记录在数据集中。(2)的目标是训练旨在最大化预期回报的政策。在这里，为了简单起见，我们使用香草IPS 估计器；然而，任何其他非政策强盗目标(e。g。, 双鲁棒估计器)可以用来代替
## 3.5 Hybrid Policy
在生产系统中，任何策略更新都会直接影响用户体验。训练具有一般报酬最大化目标的新政策，而不对行为的变化进行任何控制，会带来各种实际风险。例如，新模型可能会降低尾域的技能路由质量，同时显示出更好的平均性能。作为另一个示例，新策略可能会积极探索替代方案，考虑到关闭策略设置中的周转时间，可能会导致广泛的负面体验，直到下一个模型刷新。为了缓解此问题并限制单个模型更新中策略行为的更改，我们引人了使用混合策略(HP) 的思想。HP 由使用 RP 和 LP 目标训练的两个内部模型组成。由于 RP 会复制当前行为，而LP 会尝试通过随机选择 RP 或 LP 来最大化奖励，因此我们可以通过做出其他决策来在当前行为的复制与奖励功能的潜在改进之间建立平衡
具体来说，要创建 HP 模型，我们首先使用RP 和 LP 目标训练两个单独的模型。然后，我们使用验证集来计算 LP 为每个数据段复制当前策略的速率，计算如下：
$$\kappa_j=\mathbb{E}_{X\sim\mathbb{D}_j}[1-\frac{|\Pi_\theta(X)-\Pi_0(X)|_1}{2}],$$
(3)
其中$j$是每个数据段的索引$,\kappa_j$是新 LP 策略复制当前策略的预期速率。在这项工作中，我们将数据段定义为基于得分最高的 NLU 意图。此外，我们为所有数据段设置了所需的最小复制速率$(\tilde{\kappa})($例如。g。$\tilde{\kappa}=99\%$)。为了实现所需的复制级别，我们将参考策略决策率(RPDR) 定义为：
(4)
$$RPDR_j=\begin{cases}0&\text{if}\kappa_j\ge\tilde{\kappa}\\\frac{\tilde{\kappa}-\kappa_j}{1-\kappa_j}&\text{otherwise}\end{cases}.$$
直观地，假设 RP 是一个很好的复制模型，通过使用 RPDR 随机决定是 LP 还是 RP 应该处理每个样本，我们可以为每个段实现所需的最小复制级别。最终的 HP 模型由 LP, RP 和每个意图的预先计算的 RPDR 值的字典组成。有关 HP 的说明，请参见图 4 。
|
[
[
0.133544921875,
0.0831298828125,
0.231689453125,
0.0831298828125,
0.231689453125,
0.09320068359375,
0.133544921875,
0.09320068359375
],
[
0.51318359375,
0.08612060546875,
0.84033203125,
0.08612060546875,
0.84033203125,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.1502685546875,
0.09320068359375,
0.21240234375,
0.09320068359375,
0.21240234375,
0.1009521484375,
0.1502685546875,
0.1009521484375
],
[
0.1796875,
0.1051025390625,
0.185546875,
0.1051025390625,
0.185546875,
0.11102294921875,
0.1796875,
0.11102294921875
],
[
0.57861328125,
0.11517333984375,
0.8134765625,
0.1270751953125,
0.81201171875,
0.1484375,
0.57666015625,
0.13720703125
],
[
0.73876953125,
0.11102294921875,
0.8076171875,
0.1134033203125,
0.8076171875,
0.127685546875,
0.73876953125,
0.1258544921875
],
[
0.276123046875,
0.12469482421875,
0.306396484375,
0.12469482421875,
0.306396484375,
0.133056640625,
0.276123046875,
0.133056640625
],
[
0.3525390625,
0.127685546875,
0.44580078125,
0.127685546875,
0.44580078125,
0.1383056640625,
0.3525390625,
0.1383056640625
],
[
0.85791015625,
0.122314453125,
0.8818359375,
0.122314453125,
0.8818359375,
0.1353759765625,
0.85791015625,
0.1353759765625
],
[
0.15869140625,
0.1312255859375,
0.20654296875,
0.133056640625,
0.204833984375,
0.144287109375,
0.156982421875,
0.142578125
],
[
0.1561279296875,
0.14013671875,
0.20654296875,
0.142578125,
0.20654296875,
0.152587890625,
0.1561279296875,
0.15087890625
],
[
0.366943359375,
0.13720703125,
0.4306640625,
0.13720703125,
0.4306640625,
0.1478271484375,
0.366943359375,
0.1478271484375
],
[
0.1796875,
0.15673828125,
0.18896484375,
0.15673828125,
0.18896484375,
0.166259765625,
0.1796875,
0.166259765625
],
[
0.2493896484375,
0.157958984375,
0.255126953125,
0.157958984375,
0.255126953125,
0.1650390625,
0.2493896484375,
0.1650390625
],
[
0.51123046875,
0.1591796875,
0.87890625,
0.1591796875,
0.87890625,
0.1728515625,
0.51123046875,
0.1728515625
],
[
0.1485595703125,
0.174560546875,
0.177978515625,
0.174560546875,
0.177978515625,
0.18701171875,
0.1485595703125,
0.18701171875
],
[
0.36865234375,
0.174560546875,
0.40478515625,
0.174560546875,
0.40478515625,
0.18701171875,
0.36865234375,
0.18701171875
],
[
0.5146484375,
0.174560546875,
0.87890625,
0.174560546875,
0.87890625,
0.188232421875,
0.5146484375,
0.188232421875
],
[
0.307373046875,
0.190673828125,
0.32080078125,
0.1923828125,
0.31982421875,
0.196533203125,
0.306396484375,
0.19482421875
],
[
0.51318359375,
0.191162109375,
0.880859375,
0.191162109375,
0.880859375,
0.204833984375,
0.51318359375,
0.204833984375
],
[
0.25439453125,
0.1923828125,
0.26123046875,
0.1923828125,
0.26123046875,
0.196533203125,
0.25439453125,
0.196533203125
],
[
0.154541015625,
0.1923828125,
0.181396484375,
0.1923828125,
0.181396484375,
0.2037353515625,
0.154541015625,
0.2037353515625
],
[
0.370361328125,
0.1923828125,
0.39794921875,
0.1923828125,
0.39794921875,
0.2037353515625,
0.370361328125,
0.2037353515625
],
[
0.393798828125,
0.1923828125,
0.426513671875,
0.1923828125,
0.426513671875,
0.2037353515625,
0.393798828125,
0.2037353515625
],
[
0.246826171875,
0.1995849609375,
0.274658203125,
0.1995849609375,
0.274658203125,
0.207275390625,
0.246826171875,
0.207275390625
],
[
0.185546875,
0.193603515625,
0.208251953125,
0.193603515625,
0.208251953125,
0.201904296875,
0.185546875,
0.201904296875
],
[
0.300537109375,
0.1995849609375,
0.337646484375,
0.1995849609375,
0.337646484375,
0.207275390625,
0.300537109375,
0.207275390625
],
[
0.162109375,
0.208984375,
0.2032470703125,
0.208984375,
0.2032470703125,
0.2203369140625,
0.162109375,
0.2203369140625
],
[
0.38037109375,
0.208984375,
0.420654296875,
0.208984375,
0.420654296875,
0.2203369140625,
0.38037109375,
0.2203369140625
],
[
0.51123046875,
0.207275390625,
0.8818359375,
0.207275390625,
0.8818359375,
0.2203369140625,
0.51123046875,
0.2203369140625
],
[
0.509765625,
0.22265625,
0.8134765625,
0.221435546875,
0.8134765625,
0.236328125,
0.509765625,
0.2369384765625
],
[
0.1158447265625,
0.236328125,
0.484375,
0.2381591796875,
0.484375,
0.251708984375,
0.1158447265625,
0.2493896484375
],
[
0.51318359375,
0.2493896484375,
0.5556640625,
0.2493896484375,
0.5556640625,
0.26123046875,
0.51318359375,
0.26123046875
],
[
0.54736328125,
0.248779296875,
0.6689453125,
0.2493896484375,
0.6689453125,
0.26416015625,
0.54736328125,
0.26318359375
],
[
0.1175537109375,
0.265380859375,
0.414794921875,
0.265380859375,
0.414794921875,
0.279052734375,
0.1175537109375,
0.279052734375
],
[
0.1175537109375,
0.25048828125,
0.48779296875,
0.25048828125,
0.48779296875,
0.265380859375,
0.1175537109375,
0.265380859375
],
[
0.51318359375,
0.26953125,
0.880859375,
0.270751953125,
0.880859375,
0.28515625,
0.51318359375,
0.283935546875
],
[
0.51318359375,
0.28515625,
0.880859375,
0.287353515625,
0.880859375,
0.3017578125,
0.51318359375,
0.2998046875
],
[
0.432373046875,
0.3017578125,
0.44921875,
0.305908203125,
0.444091796875,
0.314208984375,
0.42822265625,
0.31005859375
],
[
0.51318359375,
0.3017578125,
0.87890625,
0.3017578125,
0.87890625,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.5146484375,
0.319580078125,
0.87890625,
0.319580078125,
0.87890625,
0.330078125,
0.5146484375,
0.330078125
],
[
0.37451171875,
0.305908203125,
0.385498046875,
0.305908203125,
0.385498046875,
0.311279296875,
0.37451171875,
0.311279296875
],
[
0.15869140625,
0.322509765625,
0.19140625,
0.322509765625,
0.19140625,
0.334228515625,
0.15869140625,
0.334228515625
],
[
0.208251953125,
0.32177734375,
0.2822265625,
0.322509765625,
0.2822265625,
0.334228515625,
0.208251953125,
0.333251953125
],
[
0.51318359375,
0.334228515625,
0.8759765625,
0.334228515625,
0.8759765625,
0.347412109375,
0.51318359375,
0.347412109375
],
[
0.21240234375,
0.3466796875,
0.270263671875,
0.3466796875,
0.270263671875,
0.358154296875,
0.21240234375,
0.358154296875
],
[
0.214111328125,
0.35693359375,
0.258544921875,
0.359375,
0.2568359375,
0.371826171875,
0.21240234375,
0.369384765625
],
[
0.5146484375,
0.3515625,
0.880859375,
0.3515625,
0.880859375,
0.365234375,
0.5146484375,
0.365234375
],
[
0.156982421875,
0.363525390625,
0.19140625,
0.365234375,
0.189697265625,
0.377685546875,
0.1561279296875,
0.3759765625
],
[
0.214111328125,
0.369384765625,
0.25439453125,
0.371826171875,
0.252685546875,
0.383056640625,
0.21240234375,
0.38134765625
],
[
0.5146484375,
0.36767578125,
0.87890625,
0.36767578125,
0.87890625,
0.38134765625,
0.5146484375,
0.38134765625
],
[
0.210693359375,
0.383056640625,
0.26611328125,
0.383056640625,
0.26611328125,
0.394287109375,
0.210693359375,
0.394287109375
],
[
0.4375,
0.385986328125,
0.441650390625,
0.385986328125,
0.441650390625,
0.393798828125,
0.4375,
0.393798828125
],
[
0.5146484375,
0.38427734375,
0.87890625,
0.38427734375,
0.87890625,
0.39794921875,
0.5146484375,
0.39794921875
],
[
0.382080078125,
0.394287109375,
0.441650390625,
0.396728515625,
0.43994140625,
0.410888671875,
0.38037109375,
0.4091796875
],
[
0.51318359375,
0.399658203125,
0.87744140625,
0.399658203125,
0.87744140625,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.15869140625,
0.40673828125,
0.189697265625,
0.40673828125,
0.189697265625,
0.418701171875,
0.15869140625,
0.418701171875
],
[
0.2091064453125,
0.40673828125,
0.2451171875,
0.40673828125,
0.2451171875,
0.419189453125,
0.2091064453125,
0.419189453125
],
[
0.411376953125,
0.421630859375,
0.456787109375,
0.421630859375,
0.456787109375,
0.431640625,
0.411376953125,
0.431640625
],
[
0.51318359375,
0.416259765625,
0.880859375,
0.416259765625,
0.880859375,
0.429931640625,
0.51318359375,
0.429931640625
],
[
0.364501953125,
0.4228515625,
0.393798828125,
0.4228515625,
0.393798828125,
0.43115234375,
0.364501953125,
0.43115234375
],
[
0.2091064453125,
0.435302734375,
0.2626953125,
0.435302734375,
0.2626953125,
0.446533203125,
0.2091064453125,
0.446533203125
],
[
0.38623046875,
0.435791015625,
0.4375,
0.435791015625,
0.4375,
0.44775390625,
0.38623046875,
0.44775390625
],
[
0.51318359375,
0.431640625,
0.87890625,
0.431640625,
0.87890625,
0.4453125,
0.51318359375,
0.4453125
],
[
0.1485595703125,
0.4423828125,
0.197265625,
0.4423828125,
0.197265625,
0.45361328125,
0.1485595703125,
0.45361328125
],
[
0.210693359375,
0.4453125,
0.26123046875,
0.44775390625,
0.26025390625,
0.458984375,
0.2091064453125,
0.457275390625
],
[
0.424072265625,
0.448974609375,
0.44921875,
0.448974609375,
0.44921875,
0.460205078125,
0.424072265625,
0.460205078125
],
[
0.51318359375,
0.446533203125,
0.880859375,
0.44775390625,
0.880859375,
0.4619140625,
0.51318359375,
0.46142578125
],
[
0.4013671875,
0.451904296875,
0.414794921875,
0.451904296875,
0.414794921875,
0.4560546875,
0.4013671875,
0.4560546875
],
[
0.3720703125,
0.449462890625,
0.3896484375,
0.449462890625,
0.3896484375,
0.458984375,
0.3720703125,
0.458984375
],
[
0.361083984375,
0.46142578125,
0.3955078125,
0.46142578125,
0.3955078125,
0.47265625,
0.361083984375,
0.47265625
],
[
0.424072265625,
0.46142578125,
0.459228515625,
0.46142578125,
0.459228515625,
0.47265625,
0.424072265625,
0.47265625
],
[
0.51123046875,
0.4619140625,
0.8818359375,
0.463134765625,
0.8818359375,
0.47802734375,
0.51123046875,
0.476806640625
],
[
0.51123046875,
0.479736328125,
0.880859375,
0.478515625,
0.880859375,
0.493408203125,
0.51123046875,
0.49462890625
],
[
0.1158447265625,
0.48681640625,
0.48779296875,
0.486328125,
0.48779296875,
0.50048828125,
0.1158447265625,
0.501953125
],
[
0.51123046875,
0.496337890625,
0.87744140625,
0.496337890625,
0.87744140625,
0.51025390625,
0.51123046875,
0.51025390625
],
[
0.1158447265625,
0.5029296875,
0.486083984375,
0.5029296875,
0.486083984375,
0.5166015625,
0.1158447265625,
0.5166015625
],
[
0.118408203125,
0.5185546875,
0.486083984375,
0.5166015625,
0.486083984375,
0.52783203125,
0.118408203125,
0.52978515625
],
[
0.51318359375,
0.51318359375,
0.5341796875,
0.51318359375,
0.5341796875,
0.525390625,
0.51318359375,
0.525390625
],
[
0.1175537109375,
0.53076171875,
0.489501953125,
0.53076171875,
0.489501953125,
0.54443359375,
0.1175537109375,
0.54443359375
],
[
0.53076171875,
0.529296875,
0.87890625,
0.529296875,
0.87890625,
0.5419921875,
0.53076171875,
0.5419921875
],
[
0.1175537109375,
0.5458984375,
0.48193359375,
0.5458984375,
0.48193359375,
0.55859375,
0.1175537109375,
0.55859375
],
[
0.509765625,
0.54345703125,
0.87890625,
0.54345703125,
0.87890625,
0.55712890625,
0.509765625,
0.55712890625
],
[
0.51318359375,
0.55859375,
0.880859375,
0.56005859375,
0.880859375,
0.57470703125,
0.51318359375,
0.57373046875
],
[
0.5146484375,
0.57666015625,
0.724609375,
0.57666015625,
0.724609375,
0.59033203125,
0.5146484375,
0.59033203125
],
[
0.118408203125,
0.59033203125,
0.486083984375,
0.59033203125,
0.486083984375,
0.60400390625,
0.118408203125,
0.60400390625
],
[
0.1175537109375,
0.60498046875,
0.484375,
0.60546875,
0.484375,
0.62060546875,
0.1175537109375,
0.619140625
],
[
0.6708984375,
0.60400390625,
0.8251953125,
0.60400390625,
0.8251953125,
0.6181640625,
0.6708984375,
0.6181640625
],
[
0.5380859375,
0.61328125,
0.64990234375,
0.61083984375,
0.64990234375,
0.6259765625,
0.5380859375,
0.62744140625
],
[
0.85986328125,
0.6142578125,
0.87890625,
0.6142578125,
0.87890625,
0.62353515625,
0.85986328125,
0.62353515625
],
[
0.118408203125,
0.62255859375,
0.486083984375,
0.62255859375,
0.486083984375,
0.6357421875,
0.118408203125,
0.6357421875
],
[
0.74072265625,
0.62255859375,
0.75390625,
0.62255859375,
0.75390625,
0.63427734375,
0.74072265625,
0.63427734375
],
[
0.5146484375,
0.6455078125,
0.880859375,
0.6455078125,
0.880859375,
0.66015625,
0.5146484375,
0.66015625
],
[
0.11419677734375,
0.65625,
0.28955078125,
0.65771484375,
0.28955078125,
0.6728515625,
0.11419677734375,
0.67041015625
],
[
0.5146484375,
0.6611328125,
0.880859375,
0.6611328125,
0.880859375,
0.6748046875,
0.5146484375,
0.6748046875
],
[
0.51318359375,
0.6767578125,
0.87890625,
0.6767578125,
0.87890625,
0.6904296875,
0.51318359375,
0.6904296875
],
[
0.118408203125,
0.68310546875,
0.486083984375,
0.68310546875,
0.486083984375,
0.69677734375,
0.118408203125,
0.69677734375
],
[
0.51318359375,
0.69482421875,
0.87890625,
0.69482421875,
0.87890625,
0.7080078125,
0.51318359375,
0.7080078125
],
[
0.12005615234375,
0.70068359375,
0.484375,
0.70068359375,
0.484375,
0.71142578125,
0.12005615234375,
0.71142578125
],
[
0.118408203125,
0.71435546875,
0.486083984375,
0.71435546875,
0.486083984375,
0.72802734375,
0.118408203125,
0.72802734375
],
[
0.5146484375,
0.71044921875,
0.87890625,
0.71044921875,
0.87890625,
0.724609375,
0.5146484375,
0.724609375
],
[
0.1175537109375,
0.73095703125,
0.2359619140625,
0.73095703125,
0.2359619140625,
0.7451171875,
0.1175537109375,
0.7451171875
],
[
0.51318359375,
0.72705078125,
0.8818359375,
0.72705078125,
0.8818359375,
0.74072265625,
0.51318359375,
0.74072265625
],
[
0.51318359375,
0.74365234375,
0.5751953125,
0.74365234375,
0.5751953125,
0.7548828125,
0.51318359375,
0.7548828125
],
[
0.2568359375,
0.76318359375,
0.27783203125,
0.76318359375,
0.27783203125,
0.77685546875,
0.2568359375,
0.77685546875
],
[
0.67236328125,
0.771484375,
0.69189453125,
0.771484375,
0.69189453125,
0.78515625,
0.67236328125,
0.78515625
],
[
0.73291015625,
0.7724609375,
0.8095703125,
0.7724609375,
0.8095703125,
0.78759765625,
0.73291015625,
0.78759765625
],
[
0.29296875,
0.779296875,
0.468505859375,
0.779296875,
0.468505859375,
0.7919921875,
0.29296875,
0.7919921875
],
[
0.276123046875,
0.78076171875,
0.298828125,
0.78076171875,
0.298828125,
0.7880859375,
0.276123046875,
0.7880859375
],
[
0.130126953125,
0.775390625,
0.285400390625,
0.775390625,
0.285400390625,
0.79736328125,
0.130126953125,
0.79736328125
],
[
0.57421875,
0.78076171875,
0.64404296875,
0.783203125,
0.64404296875,
0.79638671875,
0.57421875,
0.79443359375
],
[
0.68115234375,
0.7861328125,
0.71533203125,
0.79150390625,
0.7119140625,
0.80029296875,
0.67822265625,
0.7958984375
],
[
0.85791015625,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.85791015625,
0.79638671875
],
[
0.25439453125,
0.79443359375,
0.280517578125,
0.79443359375,
0.280517578125,
0.80712890625,
0.25439453125,
0.80712890625
],
[
0.67236328125,
0.7958984375,
0.71533203125,
0.798828125,
0.7138671875,
0.81005859375,
0.6708984375,
0.80712890625
],
[
0.72900390625,
0.79150390625,
0.81201171875,
0.79150390625,
0.81201171875,
0.8056640625,
0.72900390625,
0.8056640625
],
[
0.462646484375,
0.8056640625,
0.486083984375,
0.8056640625,
0.486083984375,
0.818359375,
0.462646484375,
0.818359375
],
[
0.118408203125,
0.822265625,
0.486083984375,
0.822265625,
0.486083984375,
0.8359375,
0.118408203125,
0.8359375
],
[
0.51318359375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.118408203125,
0.83837890625,
0.48779296875,
0.83837890625,
0.48779296875,
0.8515625,
0.118408203125,
0.8515625
],
[
0.51123046875,
0.84033203125,
0.87890625,
0.84033203125,
0.87890625,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.1175537109375,
0.85400390625,
0.484375,
0.85400390625,
0.484375,
0.86767578125,
0.1175537109375,
0.86767578125
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.118408203125,
0.87060546875,
0.489501953125,
0.87060546875,
0.489501953125,
0.88427734375,
0.118408203125,
0.88427734375
],
[
0.51123046875,
0.87255859375,
0.87890625,
0.87255859375,
0.87890625,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.8857421875,
0.32080078125,
0.8857421875,
0.32080078125,
0.8994140625,
0.118408203125,
0.8994140625
],
[
0.515625,
0.88818359375,
0.880859375,
0.88818359375,
0.880859375,
0.90283203125,
0.515625,
0.90283203125
],
[
0.1368408203125,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.1368408203125,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.7421875,
0.90478515625,
0.7421875,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"Post-deployment",
"盗目标，例如反向倾向评分（IPS）目标：",
"Evaluation",
"不",
"LLP = Ex,a,r~D =rIlo(aa),",
"Ie(aa)",
"fold-Out",
"Pre-deployment",
"(2)",
"Logged",
"Traffic",
"Evaluation",
"个",
"OW",
"r是观察到的采取操作的奖励α记录在数据集",
"HPi",
"HPi+1",
"中。(2)的目标是训练旨在最大化预期回报的",
"Pn",
"政策。在这里，为了简单起见，我们使用香草",
"F",
"RP",
"RP",
"LP",
"P/RP",
"LP",
"RPDR",
"RPDR",
"RPDR",
"IPS估计器；然而，任何其他非政策强盗目标",
"(e。g。，双鲁棒估计器)可以用来代替",
"Figure 2:An overview of the self-learning process:",
"3.5",
"Hybrid Policy",
"evaluation, and post-deployment evaluation.",
"model training, RPDR computation, pre-deployment",
"在生产系统中，任何策略更新都会直接影响用",
"户体验。训练具有一般报酬最大化目标的新政",
"XT",
"策，而不对行为的变化进行任何控制，会带来",
"各种实际风险。例如，新模型可能会降低尾域",
"X1",
"ASR",
"Utterance",
"的技能路由质量，同时显示出更好的平均性能",
"Domain",
"Intent",
"作为另一个示例，新策略可能会积极探索替代",
"NLU",
"Slots",
"方案，考虑到关闭策略设置中的周转时间，可",
"LScores",
"",
"能会导致广泛的负面体验，直到下一个模型刷",
"Bi-LSTM",
"新。为了缓解此问题并限制单个模型更新中策",
"Skill",
"Skill",
".MLP",
"略行为的更改，我们引入了使用混合策略（HP)",
"MLP",
"Device",
"Softmax",
"的思想。HP由使用RP和LP目标训练的两个",
"Context",
"Status",
"0",
"内部模型组成。由于RP会复制当前行为，而",
".",
"O",
"p(a)",
"p(at)",
"LP会尝试通过随机选择RP或LP来最大化奖",
"励，因此我们可以通过做出其他决策来在当前",
"Figure 3: An overview of the proposed model archi-",
"行为的复制与奖励功能的潜在改进之间建立平",
"tecture: a set of hypothesis are encoded as vectors (",
"&i..&T)and fed toabi-directional LSTMwhichis",
"衡",
"followed by a shared MLP and a softmax layer to nor-",
"具体来说，要创建HP模型，我们首先使用",
"malize the predicted candidate selection probabilities.",
"RP和LP目标训练两个单独的模型。然后，我",
"们使用验证集来计算LP为每个数据段复制当",
"前策略的速率，计算如下：",
"根据NLU解释置信度对编码假设集X进行排",
"序，并通过双向LSTM，两个完全连接的层以",
"[IIe(X) -Io(X)l1,",
"Kj=Ex~D,[1",
"(3)",
"及softmax激活来输出IIe(X）策略的操作概率",
"2",
"其中j是每个数据段的索引，ri是新LP策略",
"3.4Model Training",
"复制当前策略的预期速率。在这项工作中，我",
"们将数据段定义为基于得分最高的NLU意图。",
"我们定义了两个训练目标：复制策略(RP)和",
"此外，我们为所有数据段设置了所需的最小复",
"学习策略（LP)。RP目标尝试训练复制记录的",
"操作的模型。具体来说，我们将RP损失函数",
"制速率（K）（例如。g。K=99%）。为了实现所需",
"定义为最小化：",
"的复制级别，我们将参考策略决策率（RPDR）",
"定义为：",
"T",
"0",
"ifkj≥k",
"1[a = illog(Ie(a|X))",
"二",
"LRP = Ex,a~D ",
"RPDR;",
"R-K",
"(4)",
"i=1",
"(1-K",
"otherwise",
"(1)",
"简而言之，（1）是一种交叉炳损失，鼓励新策略",
"直观地，假设RP是一个很好的复制模型，通",
"为复制记录的操作的操作分配最高分数。我们",
"过使用RPDR随机决定是LP还是RP应该处",
"还探索了其他替代方案，例如KL发散或软蒸",
"理每个样本，我们可以为每个段实现所需的最",
"馏目标，但发现简单的交叉摘目标非常稳定，",
"小复制级别。最终的HP模型由LP，RP和每",
"并显示出出色的复制性能",
"个意图的预先计算的RPDR值的字典组成。有",
"我们将LP损失函数定义为非策略上下文强",
"关HP的说明，请参见图4。"
] |
{
"arxiv_id": "2204.07135",
"language": "zh",
"page_idx": 3
}
|
||
2204.07135-zh-4
|
2204.07135-zh-4.jpg
|
<FigureHere>
Figure 4: The hybrid policy consists of the LP and RP models as well as the pre-computed RPDR values. At the inference time, the RPDR value corresponding to the NLU top intent used to stochastically decide which model handles that sample.
要更新 HP, 根据 LP/RP 更新频率，每次在建模数据分割上训练其中一个模型，然后计算RPDR 值 (请参见图$\overline{\mathrm{E}}$2 )。我们更新 LP 型号的频率比 RP(e。g。,LP 每日更新，RP 每周更新)。此决定背后的原因是为了限制路由行为的更改更长时间。更新频率较低的 RP 模型充当移动平均过滤器，逐渐吸收 LP 行为
# 3.6 Pre/Post- Deployment Evaluation
创建新的 HP 后，在部署之前使用脱离策略评估 (OPE)来评估新策略。表 1显示了 OPE 分析为每个数据段报告的主要指标的摘要(这里是顶级 NLU 解释的域意图)。在部署前评估中， 将一组专家定义的护栏应用于评估结果，以确保可靠的模型更新，尤其是对于关键业务案例。如果新模型未达到护栏条件，部署将被中止，并由人类专家负责调查问题。否则，随着新模型的自动训练和部署，自学习循环将继续根据用户反馈逐步优化策略行为。这个自动化过程有效地解除了自我学习系统不必要的人为干预或 A/B 实验所需的高周转时间
OPE 在部署之前提供了有关模型性能的宝贵见解；但是，由于通常用于绑定 IPS 权重的权重限幅以及由于日志数据集覆盖问题而导致的高方差，OPE 估计可能会受到估计偏差的影响 (Swaminathan et al., 2016; Joachims et al., 2018; Sachdeva et al.,2020)。因此，跟踪已部署策略的部署后性能并测量经验复制和用户体验指标至关重要。
## 4 Experiments
# 4.1 Setup
为了评估所提出的自学习技能路由方法，我们在现实世界的生产环境中进行了广泛的在线和离线实验。在本节中，我们使用术语基线来指代类似于 Park et al.(2020b) 建议的重新标记方法的策略实现
我们进行了在线 A/B 实验，涉及约 6M 个
<table>
<tbody>
<tr>
<th>Metric</th>
<th>Description</th>
</tr>
<tr>
<td>Replication (defect/non-defect)</td>
<td>rate of $\Pi_\theta$ making actions similar to $\Pi_{0}$</td>
</tr>
<tr>
<td rowspan="2">L1-distance</td>
<td>average of Ll-distance</td>
</tr>
<tr>
<td>between $\Pi_{0}$ and $\Pi_{\theta}$</td>
</tr>
<tr>
<td rowspan="2">STD of L1-distance</td>
<td>std of Ll-distance</td>
</tr>
<tr>
<td>between $\Pi_{0}$ and $\Pi_\theta$</td>
</tr>
<tr>
<td>Expected reward</td>
<td>IPS weighted reward for $\Pi_{\theta}$ (counterfactual estimation)</td>
</tr>
<tr>
<td>Expected IPS weight</td>
<td>average IPS weight (ideally equal to 1.0)</td>
</tr>
<tr>
<td>Stochastic exploration</td>
<td>the rate of not selecting</td>
</tr>
<tr>
<td>(defect/non-defect)</td>
<td>the highest scoring candidate</td>
</tr>
</tbody>
</table>
Table 1: The summary of main metrics used in the pre-
deployment evaluation.
独特的客户，其中基线策略服务于控制，而自我学习模型服务于治疗客户。我们以每周一次HP 的节奏训练了四个连续的自学习 HP(由HP1 至 HP4 表示)。每个模型都在两周的治疗数据的流量窗口上训练，除了第一个治疗模型(HP1)在实验前收集的基线的记录数据上训练。由于生产中的 A/B 插槽可用性限制，我们决定不在此 A/B 实验中更新 RP,并使用 OPE 分析来评估经过训练的 RP 模型的性能。因此， 我们使用了一个固定的 RP 模型，该模型复制了基线策略，并专注于在整个实验过程中更新LP。我们将个人意图的最小复制所需级别( $\tilde{\kappa}$ ) 设置为 90%
此外，我们进行了初步的 A/B 实验，包括在49 天内进行了7 次 LP 和 2 次 RP 模型更新 , 证明了在很长一段时间内稳定，稳定的改进。但是，由于某些部署问题，模型更新的时间表受到了影响，我们决定在附录中介绍这些结果。
## 4.2 Results
图 5 显示了 A/B 测试模型的治疗和对照之间的差异百分比。从图中可以看出，所提出的自学习模型提高了平均奖励，显示出比迭代改进的总体趋势。请注意，在高度优化的生产系统中，1% 的改进被认为是用户体验的重大改进。在这里，我们使用具有八个重新采样的引导方法来查找 95% 置信区间，并在图中用填充区域显示它们。请注意，每个报告的值是一周内收集的大约 40M 话语的平均值。比较 HP3 和HP4 模型的性能，我们可以看到 OPE 预测的第四次模型刷新的回归。但是，由于奖励回归没有超过我们部署前的护栏公差值，因此继续进行部署
|
[
[
0.252685546875,
0.08612060546875,
0.27783203125,
0.08612060546875,
0.27783203125,
0.097412109375,
0.252685546875,
0.097412109375
],
[
0.2939453125,
0.0914306640625,
0.355224609375,
0.0914306640625,
0.355224609375,
0.10272216796875,
0.2939453125,
0.10272216796875
],
[
0.52490234375,
0.09027099609375,
0.57421875,
0.09027099609375,
0.57421875,
0.1051025390625,
0.52490234375,
0.1051025390625
],
[
0.68603515625,
0.0926513671875,
0.75830078125,
0.0926513671875,
0.75830078125,
0.10394287109375,
0.68603515625,
0.10394287109375
],
[
0.52294921875,
0.10687255859375,
0.60546875,
0.1092529296875,
0.60546875,
0.12353515625,
0.52294921875,
0.1217041015625
],
[
0.68408203125,
0.1092529296875,
0.8486328125,
0.1092529296875,
0.8486328125,
0.122314453125,
0.68408203125,
0.122314453125
],
[
0.5263671875,
0.1270751953125,
0.64990234375,
0.1270751953125,
0.64990234375,
0.14013671875,
0.5263671875,
0.14013671875
],
[
0.68603515625,
0.127685546875,
0.767578125,
0.127685546875,
0.767578125,
0.1395263671875,
0.68603515625,
0.1395263671875
],
[
0.52490234375,
0.142578125,
0.60546875,
0.1455078125,
0.603515625,
0.1602783203125,
0.52294921875,
0.15673828125
],
[
0.6826171875,
0.1436767578125,
0.82958984375,
0.1436767578125,
0.82958984375,
0.157958984375,
0.6826171875,
0.157958984375
],
[
0.6826171875,
0.160888671875,
0.81005859375,
0.162109375,
0.81005859375,
0.177001953125,
0.6826171875,
0.17578125
],
[
0.52490234375,
0.1787109375,
0.65576171875,
0.18115234375,
0.65576171875,
0.1953125,
0.52490234375,
0.193603515625
],
[
0.6826171875,
0.179931640625,
0.8017578125,
0.18115234375,
0.8017578125,
0.1953125,
0.6826171875,
0.19482421875
],
[
0.1175537109375,
0.1923828125,
0.486083984375,
0.1923828125,
0.486083984375,
0.2060546875,
0.1175537109375,
0.2060546875
],
[
0.68408203125,
0.198974609375,
0.81005859375,
0.198974609375,
0.81005859375,
0.2120361328125,
0.68408203125,
0.2120361328125
],
[
0.118408203125,
0.2078857421875,
0.484375,
0.2078857421875,
0.484375,
0.218505859375,
0.118408203125,
0.218505859375
],
[
0.5283203125,
0.218505859375,
0.634765625,
0.218505859375,
0.634765625,
0.2298583984375,
0.5283203125,
0.2298583984375
],
[
0.68408203125,
0.21728515625,
0.8623046875,
0.21728515625,
0.8623046875,
0.23095703125,
0.68408203125,
0.23095703125
],
[
0.1175537109375,
0.235107421875,
0.486083984375,
0.235107421875,
0.486083984375,
0.248779296875,
0.1175537109375,
0.248779296875
],
[
0.118408203125,
0.22265625,
0.483642578125,
0.22265625,
0.483642578125,
0.2327880859375,
0.118408203125,
0.2327880859375
],
[
0.68603515625,
0.235107421875,
0.8564453125,
0.235107421875,
0.8564453125,
0.248779296875,
0.68603515625,
0.248779296875
],
[
0.1175537109375,
0.2493896484375,
0.298828125,
0.25048828125,
0.298828125,
0.262451171875,
0.1175537109375,
0.26123046875
],
[
0.5283203125,
0.2529296875,
0.66162109375,
0.2529296875,
0.66162109375,
0.2666015625,
0.5283203125,
0.2666015625
],
[
0.6826171875,
0.2529296875,
0.8095703125,
0.2529296875,
0.8095703125,
0.2666015625,
0.6826171875,
0.2666015625
],
[
0.68408203125,
0.271484375,
0.81689453125,
0.271484375,
0.81689453125,
0.28515625,
0.68408203125,
0.28515625
],
[
0.5263671875,
0.287353515625,
0.67333984375,
0.2880859375,
0.67333984375,
0.302734375,
0.5263671875,
0.3017578125
],
[
0.67919921875,
0.289306640625,
0.83447265625,
0.289306640625,
0.83447265625,
0.302734375,
0.67919921875,
0.302734375
],
[
0.118408203125,
0.30712890625,
0.486083984375,
0.30712890625,
0.486083984375,
0.320556640625,
0.118408203125,
0.320556640625
],
[
0.1368408203125,
0.29150390625,
0.486083984375,
0.29150390625,
0.486083984375,
0.3046875,
0.1368408203125,
0.3046875
],
[
0.5283203125,
0.30712890625,
0.64990234375,
0.30712890625,
0.64990234375,
0.320556640625,
0.5283203125,
0.320556640625
],
[
0.68408203125,
0.30712890625,
0.86962890625,
0.30712890625,
0.86962890625,
0.320556640625,
0.68408203125,
0.320556640625
],
[
0.1175537109375,
0.322509765625,
0.484375,
0.322509765625,
0.484375,
0.336181640625,
0.1175537109375,
0.336181640625
],
[
0.509765625,
0.333251953125,
0.87890625,
0.333251953125,
0.87890625,
0.3466796875,
0.509765625,
0.3466796875
],
[
0.118408203125,
0.339111328125,
0.48779296875,
0.339111328125,
0.48779296875,
0.35400390625,
0.118408203125,
0.35400390625
],
[
0.5146484375,
0.350830078125,
0.66650390625,
0.350830078125,
0.66650390625,
0.361083984375,
0.5146484375,
0.361083984375
],
[
0.12005615234375,
0.35693359375,
0.484375,
0.35693359375,
0.484375,
0.36767578125,
0.12005615234375,
0.36767578125
],
[
0.1175537109375,
0.37060546875,
0.484375,
0.37060546875,
0.484375,
0.38427734375,
0.1175537109375,
0.38427734375
],
[
0.118408203125,
0.388427734375,
0.36865234375,
0.388427734375,
0.36865234375,
0.402099609375,
0.118408203125,
0.402099609375
],
[
0.51318359375,
0.39013671875,
0.87890625,
0.39013671875,
0.87890625,
0.40380859375,
0.51318359375,
0.40380859375
],
[
0.51318359375,
0.40625,
0.880859375,
0.40625,
0.880859375,
0.419189453125,
0.51318359375,
0.419189453125
],
[
0.1175537109375,
0.416259765625,
0.426513671875,
0.416259765625,
0.426513671875,
0.429931640625,
0.1175537109375,
0.429931640625
],
[
0.51123046875,
0.421630859375,
0.87890625,
0.421630859375,
0.87890625,
0.435302734375,
0.51123046875,
0.435302734375
],
[
0.118408203125,
0.438232421875,
0.486083984375,
0.438232421875,
0.486083984375,
0.451904296875,
0.118408203125,
0.451904296875
],
[
0.51123046875,
0.435791015625,
0.880859375,
0.43701171875,
0.880859375,
0.451904296875,
0.51123046875,
0.45068359375
],
[
0.118408203125,
0.454833984375,
0.486083984375,
0.454833984375,
0.486083984375,
0.468505859375,
0.118408203125,
0.468505859375
],
[
0.5146484375,
0.454833984375,
0.87744140625,
0.454833984375,
0.87744140625,
0.465576171875,
0.5146484375,
0.465576171875
],
[
0.118408203125,
0.47021484375,
0.484375,
0.47021484375,
0.484375,
0.48388671875,
0.118408203125,
0.48388671875
],
[
0.509765625,
0.4697265625,
0.87890625,
0.4697265625,
0.87890625,
0.482666015625,
0.509765625,
0.482666015625
],
[
0.12005615234375,
0.488037109375,
0.48193359375,
0.488037109375,
0.48193359375,
0.498779296875,
0.12005615234375,
0.498779296875
],
[
0.51123046875,
0.486328125,
0.87890625,
0.486328125,
0.87890625,
0.49951171875,
0.51123046875,
0.49951171875
],
[
0.1175537109375,
0.5029296875,
0.483642578125,
0.5029296875,
0.483642578125,
0.5166015625,
0.1175537109375,
0.5166015625
],
[
0.5146484375,
0.50439453125,
0.880859375,
0.50439453125,
0.880859375,
0.51416015625,
0.5146484375,
0.51416015625
],
[
0.1175537109375,
0.5185546875,
0.484375,
0.51708984375,
0.484375,
0.5322265625,
0.1175537109375,
0.533203125
],
[
0.51318359375,
0.5185546875,
0.880859375,
0.5185546875,
0.880859375,
0.5322265625,
0.51318359375,
0.5322265625
],
[
0.118408203125,
0.53369140625,
0.483642578125,
0.53369140625,
0.483642578125,
0.54736328125,
0.118408203125,
0.54736328125
],
[
0.51318359375,
0.53515625,
0.880859375,
0.53515625,
0.880859375,
0.548828125,
0.51318359375,
0.548828125
],
[
0.1175537109375,
0.55029296875,
0.484375,
0.55029296875,
0.484375,
0.56396484375,
0.1175537109375,
0.56396484375
],
[
0.5146484375,
0.55029296875,
0.880859375,
0.55029296875,
0.880859375,
0.56396484375,
0.5146484375,
0.56396484375
],
[
0.118408203125,
0.56787109375,
0.486083984375,
0.56787109375,
0.486083984375,
0.58056640625,
0.118408203125,
0.58056640625
],
[
0.51123046875,
0.5654296875,
0.880859375,
0.5654296875,
0.880859375,
0.5791015625,
0.51123046875,
0.5791015625
],
[
0.12005615234375,
0.58447265625,
0.483642578125,
0.58447265625,
0.483642578125,
0.59423828125,
0.12005615234375,
0.59423828125
],
[
0.51123046875,
0.5830078125,
0.60693359375,
0.58056640625,
0.60693359375,
0.595703125,
0.51123046875,
0.59716796875
],
[
0.118408203125,
0.599609375,
0.486083984375,
0.599609375,
0.486083984375,
0.61328125,
0.118408203125,
0.61328125
],
[
0.52880859375,
0.5986328125,
0.87890625,
0.5986328125,
0.87890625,
0.6123046875,
0.52880859375,
0.6123046875
],
[
0.12005615234375,
0.6162109375,
0.3955078125,
0.6162109375,
0.3955078125,
0.6298828125,
0.12005615234375,
0.6298828125
],
[
0.51123046875,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.51123046875,
0.62890625
],
[
0.1351318359375,
0.63134765625,
0.48779296875,
0.6318359375,
0.48779296875,
0.646484375,
0.1351318359375,
0.6455078125
],
[
0.51318359375,
0.6318359375,
0.87890625,
0.6318359375,
0.87890625,
0.642578125,
0.51318359375,
0.642578125
],
[
0.118408203125,
0.6484375,
0.486083984375,
0.6484375,
0.486083984375,
0.662109375,
0.118408203125,
0.662109375
],
[
0.51318359375,
0.64794921875,
0.880859375,
0.64794921875,
0.880859375,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.12005615234375,
0.66650390625,
0.484375,
0.66650390625,
0.484375,
0.6767578125,
0.12005615234375,
0.6767578125
],
[
0.51123046875,
0.66455078125,
0.86962890625,
0.66455078125,
0.86962890625,
0.677734375,
0.51123046875,
0.677734375
],
[
0.118408203125,
0.68115234375,
0.484375,
0.68115234375,
0.484375,
0.69482421875,
0.118408203125,
0.69482421875
],
[
0.51318359375,
0.69140625,
0.5615234375,
0.69140625,
0.5615234375,
0.703125,
0.51318359375,
0.703125
],
[
0.55517578125,
0.6904296875,
0.61865234375,
0.6904296875,
0.61865234375,
0.705078125,
0.55517578125,
0.705078125
],
[
0.118408203125,
0.69677734375,
0.484375,
0.69677734375,
0.484375,
0.71044921875,
0.118408203125,
0.71044921875
],
[
0.1175537109375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.7255859375,
0.1175537109375,
0.7255859375
],
[
0.51123046875,
0.7119140625,
0.880859375,
0.7119140625,
0.880859375,
0.7255859375,
0.51123046875,
0.7255859375
],
[
0.118408203125,
0.728515625,
0.48779296875,
0.728515625,
0.48779296875,
0.7421875,
0.118408203125,
0.7421875
],
[
0.51123046875,
0.72705078125,
0.87744140625,
0.72705078125,
0.87744140625,
0.74072265625,
0.51123046875,
0.74072265625
],
[
0.1175537109375,
0.74365234375,
0.25439453125,
0.7451171875,
0.25439453125,
0.7587890625,
0.1175537109375,
0.75732421875
],
[
0.51318359375,
0.74462890625,
0.8818359375,
0.74462890625,
0.8818359375,
0.75830078125,
0.51318359375,
0.75830078125
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.1175537109375,
0.77490234375,
0.14697265625,
0.77490234375,
0.14697265625,
0.78759765625,
0.1175537109375,
0.78759765625
],
[
0.1502685546875,
0.77490234375,
0.264404296875,
0.77490234375,
0.264404296875,
0.7890625,
0.1502685546875,
0.7890625
],
[
0.51318359375,
0.77685546875,
0.87890625,
0.77685546875,
0.87890625,
0.78759765625,
0.51318359375,
0.78759765625
],
[
0.51318359375,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.11419677734375,
0.7998046875,
0.210693359375,
0.8017578125,
0.210693359375,
0.81640625,
0.11419677734375,
0.81396484375
],
[
0.5146484375,
0.81005859375,
0.87744140625,
0.81005859375,
0.87744140625,
0.82080078125,
0.5146484375,
0.82080078125
],
[
0.1175537109375,
0.82373046875,
0.48779296875,
0.822265625,
0.48779296875,
0.83740234375,
0.1175537109375,
0.83837890625
],
[
0.51123046875,
0.82373046875,
0.880859375,
0.826171875,
0.880859375,
0.8388671875,
0.51123046875,
0.83740234375
],
[
0.118408203125,
0.84033203125,
0.48779296875,
0.84033203125,
0.48779296875,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.509765625,
0.8388671875,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.509765625,
0.8525390625
],
[
0.118408203125,
0.85791015625,
0.484375,
0.85791015625,
0.484375,
0.8681640625,
0.118408203125,
0.8681640625
],
[
0.51123046875,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.51123046875,
0.87158203125,
0.8818359375,
0.87255859375,
0.8818359375,
0.88720703125,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.89013671875,
0.2283935546875,
0.89013671875,
0.2283935546875,
0.90087890625,
0.118408203125,
0.90087890625
],
[
0.51318359375,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.1351318359375,
0.90478515625,
0.486083984375,
0.90380859375,
0.486083984375,
0.91845703125,
0.1351318359375,
0.91943359375
],
[
0.51123046875,
0.90478515625,
0.58837890625,
0.90478515625,
0.58837890625,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"HP",
">RPDR",
"Metric",
"Description",
"Replication",
"rate of IIe making actions",
"(defect/non-defect)",
"similar to Io",
"LI-distance",
"average of L1-distance",
"between Io and Ie",
"STD of LI-distance",
"std of L1-distance",
"Figure 4: The hybrid policy consists of the LP and RP",
"between Io and IIe",
"models as well as the pre-computed RPDR values. At",
"Expected reward",
"IPS weighted reward for IIg",
"the NLU top intent used to stochastically decide which",
"theinferencetime,theRPDR valuecorresponding to",
"(counterfactual estimation)",
"model handlesthat sample",
"Expected IPS weight",
"average IPS weight",
"(ideally equal to 1.0)",
"Stochastic exploration",
"the rate of not selecting",
"建模数据分割上训练其中一个模型，然后计算",
"要更新HP，根据LP/RP更新频率，每次在",
"(defect/non-defect)",
"the highest scoring candidate",
"RPDR值（请参见图E2）。我们更新LP型号的",
"Table l: The summary of main metrics used in the pre",
"频率比RP(e。g。,LP每日更新,RP每周更新)。",
"deploymentevaluation",
"此决定背后的原因是为了限制路由行为的更改",
"更长时间。更新频率较低的RP模型充当移动",
"平均过滤器，逐渐吸收LP行为",
"独特的客户，其中基线策略服务于控制，而自",
"我学习模型服务于治疗客户。我们以每周一次",
"3.6Pre/Post-Deployment Evaluation",
"HP的节奏训练了四个连续的自学习HP（由",
"创建新的HP后，在部署之前使用脱离策略评",
"HP1至HP4表示）。每个模型都在两周的治疗",
"估（OPE）来评估新策略。表1显示了OPE分",
"数据的流量窗口上训练，除了第一个治疗模型",
"析为每个数据段报告的主要指标的摘要（这里",
"（HP1）在实验前收集的基线的记录数据上训",
"是顶级NLU解释的域意图）。在部署前评估中",
"练。由于生产中的A/B插槽可用性限制，我们",
"将一组专家定义的护栏应用于评估结果，以",
"决定不在此A/B实验中更新RP，并使用OPE",
"确保可靠的模型更新，尤其是对于关键业务案",
"分析来评估经过训练的RP模型的性能。因此，",
"例。如果新模型未达到护栏条件，部署将被中",
"我们使用了一个固定的RP模型，该模型复制",
"止，并由人类专家负责调查问题。否则，随着",
"了基线策略，并专注于在整个实验过程中更新",
"新模型的自动训练和部署，自学习循环将继续",
"LP。我们将个人意图的最小复制所需级别（）",
"根据用户反馈逐步优化策略行为。这个自动化",
"设置为90%",
"过程有效地解除了自我学习系统不必要的人为",
"此外，我们进行了初步的A/B实验，包括在",
"干预或A/B实验所需的高周转时间",
"49天内进行了7次LP和2次RP模型更新，证",
"OPE在部署之前提供了有关模型性能的宝",
"明了在很长一段时间内稳定，稳定的改进。但",
"贵见解；但是，由于通常用于绑定IPS权重的",
"是，由于某些部署问题，模型更新的时间表受",
"权重限幅以及由于日志数据集覆盖问题而导致",
"到了影响，我们决定在附录中介绍这些结果。",
"的高方差，OPE估计可能会受到估计偏差的",
"4.2",
"Results",
"影响 (Swaminathan et al., 2016; Joachims et al.,",
"2018Sachdevaetal.，2020）。因此，跟踪已部",
"图5显示了A/B测试模型的治疗和对照之间",
"署策略的部署后性能并测量经验复制和用户体",
"的差异百分比。从图中可以看出，所提出的自",
"验指标至关重要。",
"学习模型提高了平均奖励，显示出比迭代改进",
"的总体趋势。请注意，在高度优化的生产系统",
"4",
"Experiments",
"中，1%的改进被认为是用户体验的重大改进。",
"在这里，我们使用具有八个重新采样的引导方",
"4.1Setup",
"法来查找95%置信区间，并在图中用填充区",
"为了评估所提出的自学习技能路由方法，我们",
"域显示它们。请注意，每个报告的值是一周内",
"在现实世界的生产环境中进行了广泛的在线和",
"收集的大约40M话语的平均值。比较HP3和",
"离线实验。在本节中，我们使用术语基线来指",
"HP4模型的性能，我们可以看到OPE预测的",
"代类似于Parketal.(2020b）建议的重新标记方",
"第四次模型刷新的回归。但是，由于奖励回归",
"法的策略实现",
"没有超过我们部署前的护栏公差值，因此继续",
"我们进行了在线A/B实验，涉及约6M个",
"进行部署"
] |
{
"arxiv_id": "2204.07135",
"language": "zh",
"page_idx": 4
}
|
||
1701.05946-zh-0
|
1701.05946-zh-0.jpg
|
GSP—机构设计的灰姑娘$^*$
Christopher A. Wilkens$^\dagger$ Ruggiero Cavallo$^{\ddagger }$ Rad Niazadeh$^{\$}$
May 18, 2024
$\mathbf{Abstract}$
Nearly fifteen years ago, Google unveiled the generalized second price (GSP) auction. By all theoretical accounts including their own [28] ,this was the wrong auction-the VickreyClarke-Groves (VCG) auction would have been the proper choice - yet GSP has succeeded spectacularly. We give a deep justification for GSP's success: advertisers' preferences map to a model we call $value\textit{ maximization ; they do not maximize profit as the standard theory would}$ believe. For value maximizers, GSP is the truthful auction [4]. Moreover, this implies an axiomatization of GSP— it is an auction whose prices are truthful for value maximizers— that can be applied much more broadly than the simple model for which GSP was originally designed. In particular, applying it to arbitrary single-parameter domains recovers the folklore definition of GSP. Through the lens of value maximization, GSP metamorphosizes into a powerful auction, sound in its principles and elegant in its simplicity.
# 1 Introduction
2002 年 2月，Google 推出了按点击付费的 AdWords 拍卖，其创新将改变行业：竞争对手使用第一种价格模型在搜索结果页面上出售广告，而$^{1}$Google 选择了第二种价格范式$\mathbb{E}[28]$。好处很明显[16] ,在 5 年内，搜索广告行业已经标准化了谷歌的通用第二价格 (GSP) 拍卖。GSP 存在一个问题：这不是 Vickrey 第二次价格拍卖的适当概括，而是属于 Vickrey-Clarke-Groves (VCG) 机制的标题。Google 在发布后仅几个月就意识到了这一点，但 GSP 已经嗡嗡作响，根深蒂固，无法取代$\mathbb{E}[28]$
从所有合理的角度来看，GSP 蓬勃发展，令研究界大为震惊。我们的集体第一反应只是“不能太糟糕”: 合理均衡中的收入和福利并不比 VCG [27,17]差。随后进行了许多关于 GSP 的研究-最近的理由超出了 GSP 的简单性：GSP 定价与用户点击广告的可能性 $^2$无关，因此 GSP 的均衡保证不会受到点击建模中的错误的影响$\mathbb{E}[22,15]$。然而，在这些和许多其他结果中，社区的隐含观点一直保持不变：普惠制是一只丑小鸭，需要不断的借口
$^{*}$A preliminary work“On the Truthfulness of GSP”[12] was presented at the 11th Ad Auctions Workshop and
covered a number of results in this paper.
$^{\dagger}$Yahoo Research, Sunnyvale, CA $^{\ddagger}$Yahoo Research, New York, NY $^{8}$Cornell University, Ithaca, NY. Part of this work was done while the author was an intern at Yahoo.
$^{1}$Overture.com, formerly GoTo.com and later bought by Yahoo, introduced the first sponsored search auctions in 1996 and was still the dominant player in 2002. Their auction had pay-your-bid pricing that has since been dubbed the generalized frst price auction (GFP).
$^{2}$Technically, the“ad effect”term in the click probability appears in both the optimization and pricing, but these can be thought of as bidder weights; analysis of GSP's equilibria does not require that they are accurate. This simplicity also makes implementation easier and is a practical reason GSP is common
1
|
[
[
0.343017578125,
0.1427001953125,
0.66650390625,
0.143310546875,
0.66650390625,
0.1622314453125,
0.343017578125,
0.1610107421875
],
[
0.2156982421875,
0.186279296875,
0.423095703125,
0.1849365234375,
0.423095703125,
0.199462890625,
0.2156982421875,
0.2008056640625
],
[
0.464111328125,
0.1875,
0.62109375,
0.1849365234375,
0.62109375,
0.199462890625,
0.464111328125,
0.2020263671875
],
[
0.6611328125,
0.186279296875,
0.7900390625,
0.18310546875,
0.7900390625,
0.2008056640625,
0.6611328125,
0.203857421875
],
[
0.441162109375,
0.2171630859375,
0.5556640625,
0.2171630859375,
0.5556640625,
0.231689453125,
0.441162109375,
0.231689453125
],
[
0.439453125,
0.189453125,
0.45751953125,
0.189453125,
0.45751953125,
0.1962890625,
0.439453125,
0.1962890625
],
[
0.464111328125,
0.279052734375,
0.5361328125,
0.279052734375,
0.5361328125,
0.29345703125,
0.464111328125,
0.29345703125
],
[
0.1846923828125,
0.300537109375,
0.83837890625,
0.300537109375,
0.83837890625,
0.314453125,
0.1846923828125,
0.314453125
],
[
0.16259765625,
0.31640625,
0.73046875,
0.31640625,
0.73046875,
0.330078125,
0.16259765625,
0.330078125
],
[
0.740234375,
0.31494140625,
0.8349609375,
0.317626953125,
0.8349609375,
0.328857421875,
0.740234375,
0.326904296875
],
[
0.160888671875,
0.330810546875,
0.654296875,
0.330810546875,
0.654296875,
0.3447265625,
0.160888671875,
0.3447265625
],
[
0.65185546875,
0.330078125,
0.83837890625,
0.330078125,
0.83837890625,
0.343505859375,
0.65185546875,
0.343505859375
],
[
0.16015625,
0.3466796875,
0.83837890625,
0.3466796875,
0.83837890625,
0.35986328125,
0.16015625,
0.35986328125
],
[
0.160888671875,
0.361083984375,
0.83837890625,
0.361083984375,
0.83837890625,
0.375,
0.160888671875,
0.375
],
[
0.16015625,
0.376953125,
0.83837890625,
0.376953125,
0.83837890625,
0.39013671875,
0.16015625,
0.39013671875
],
[
0.16259765625,
0.392578125,
0.8349609375,
0.392578125,
0.8349609375,
0.40283203125,
0.16259765625,
0.40283203125
],
[
0.16259765625,
0.4091796875,
0.8349609375,
0.4091796875,
0.8349609375,
0.419189453125,
0.16259765625,
0.419189453125
],
[
0.16015625,
0.42236328125,
0.83837890625,
0.42236328125,
0.83837890625,
0.436279296875,
0.16015625,
0.436279296875
],
[
0.158447265625,
0.43505859375,
0.83642578125,
0.436279296875,
0.83642578125,
0.45068359375,
0.158447265625,
0.449462890625
],
[
0.16015625,
0.451904296875,
0.52392578125,
0.451904296875,
0.52392578125,
0.46533203125,
0.16015625,
0.46533203125
],
[
0.11358642578125,
0.492431640625,
0.160888671875,
0.492431640625,
0.160888671875,
0.5068359375,
0.11358642578125,
0.5068359375
],
[
0.1568603515625,
0.492431640625,
0.30810546875,
0.49365234375,
0.30810546875,
0.50830078125,
0.1568603515625,
0.5068359375
],
[
0.1168212890625,
0.5244140625,
0.8798828125,
0.5244140625,
0.8798828125,
0.5390625,
0.1168212890625,
0.5390625
],
[
0.1258544921875,
0.54248046875,
0.88134765625,
0.54248046875,
0.88134765625,
0.556640625,
0.1258544921875,
0.556640625
],
[
0.11767578125,
0.5595703125,
0.87744140625,
0.5595703125,
0.87744140625,
0.57373046875,
0.11767578125,
0.57373046875
],
[
0.1168212890625,
0.5771484375,
0.88134765625,
0.5771484375,
0.88134765625,
0.59033203125,
0.1168212890625,
0.59033203125
],
[
0.1168212890625,
0.59228515625,
0.88134765625,
0.59326171875,
0.88134765625,
0.60791015625,
0.1168212890625,
0.60693359375
],
[
0.11517333984375,
0.60986328125,
0.19775390625,
0.60986328125,
0.19775390625,
0.625,
0.11517333984375,
0.625
],
[
0.146240234375,
0.62744140625,
0.88134765625,
0.62744140625,
0.88134765625,
0.6416015625,
0.146240234375,
0.6416015625
],
[
0.1168212890625,
0.64306640625,
0.88134765625,
0.64453125,
0.88134765625,
0.6591796875,
0.1168212890625,
0.65771484375
],
[
0.11767578125,
0.662109375,
0.88134765625,
0.662109375,
0.88134765625,
0.67529296875,
0.11767578125,
0.67529296875
],
[
0.11517333984375,
0.67822265625,
0.88134765625,
0.6787109375,
0.88134765625,
0.693359375,
0.11517333984375,
0.6923828125
],
[
0.1168212890625,
0.69580078125,
0.56396484375,
0.69580078125,
0.56396484375,
0.708984375,
0.1168212890625,
0.708984375
],
[
0.136474609375,
0.7216796875,
0.88232421875,
0.7216796875,
0.88232421875,
0.73486328125,
0.136474609375,
0.73486328125
],
[
0.11358642578125,
0.734375,
0.390625,
0.73486328125,
0.390625,
0.75,
0.11358642578125,
0.74853515625
],
[
0.1380615234375,
0.74755859375,
0.35546875,
0.74853515625,
0.35546875,
0.76220703125,
0.1380615234375,
0.76123046875
],
[
0.1356201171875,
0.76220703125,
0.3544921875,
0.7646484375,
0.3544921875,
0.77783203125,
0.1356201171875,
0.77587890625
],
[
0.1356201171875,
0.77783203125,
0.80615234375,
0.77783203125,
0.80615234375,
0.79150390625,
0.1356201171875,
0.79150390625
],
[
0.1356201171875,
0.79052734375,
0.88134765625,
0.79150390625,
0.88134765625,
0.8046875,
0.1356201171875,
0.8037109375
],
[
0.1168212890625,
0.8046875,
0.88134765625,
0.8037109375,
0.88134765625,
0.818359375,
0.1168212890625,
0.8193359375
],
[
0.11767578125,
0.8212890625,
0.38818359375,
0.8212890625,
0.38818359375,
0.83203125,
0.11767578125,
0.83203125
],
[
0.1356201171875,
0.83203125,
0.88134765625,
0.83349609375,
0.88134765625,
0.84765625,
0.1356201171875,
0.8466796875
],
[
0.1168212890625,
0.84765625,
0.88232421875,
0.84765625,
0.88232421875,
0.861328125,
0.1168212890625,
0.861328125
],
[
0.11517333984375,
0.861328125,
0.68701171875,
0.8623046875,
0.68701171875,
0.87548828125,
0.11517333984375,
0.875
],
[
0.491943359375,
0.89794921875,
0.50732421875,
0.89794921875,
0.50732421875,
0.912109375,
0.491943359375,
0.912109375
]
] |
[
"GSP一机构设计的灰姑娘*",
"Christopher A. Wilkenst",
"Ruggiero Cavallot",
"Rad Niazadehs",
"May 18, 2024",
"",
"Abstract",
"Nearly fifteen years ago, Google unveiled the generalized second price (GSP) auction. By",
"all theoretical accounts including their own [28] , this was the wrong auction",
"theVickrey",
"Clarke-Groves (VCG) auction would have been the proper choice",
"yet GSP has succeeded",
"spectacularly. We give a deep justification for GsP's success: advertisers preferences map to a",
"model we call value marimization ; they do not maximize profit as the standard theory would",
"believe. For value maximizers, GSP is the truthful auction[4]: Moreover, this implies an",
"axiomatizationof GSPitisanauctionwhosepricesaretruthfulforvaluemaximizersthat",
"canbeappliedmuchmorebroadlythanthesimplemodelforwhichGsPwasoriginallydesigned",
"In particular, applying it to arbitrary single-parameter domains recovers the folklore definition",
"of GSP. Through the lens of value maximization, GSP metamorphosizes into a powerful auction,",
"sound in its principles and elegant in its simplicity",
"1",
"Introduction",
"2002年2月，Google推出了按点击付费的AdWords拍卖，其创新将改变行业：竞争对手使用第",
"一种价格模型在搜索结果页面上出售广告，而1Google选择了第二种价格范式E[28]。好处很明",
"显[16]，在5年内，搜索广告行业已经标准化了谷歌的通用第二价格（GSP）拍卖。GSP存在",
"个问题：这不是Vickrey第二次价格拍卖的适当概括，而是属于Vickrey-Clarke-Groves（VCG）机",
"制的标题。Google在发布后仅几个月就意识到了这一点，但GSP已经喻喻作响，根深蒂固，无法",
"取代E[28]",
"从所有合理的角度来看，GSP蓬勃发展，令研究界大为震惊。我们的集体第一反应只是“不能",
"太糟糕”：合理均衡中的收入和福利并不比VCG[27，17]差。随后进行了许多关于GSP的研究-最",
"近的理由超出了GSP的简单性：GSP定价与用户点击广告的可能性2无关，因此GSP的均衡保",
"证不会受到点击建模中的错误的影响E[22，15]。然而，在这些和许多其他结果中，社区的隐含观",
"点一直保持不变：普惠制是一只丑小鸭，需要不断的借口",
"*A preliminary work “On the Truthfulness of GSP\"[12] was presented at the 1lth Ad Auctions Workshop and",
"covered a number of results in this paper",
"YahooResearch,Sunnyvale,CA",
"YahooResearch, NewYork, NY",
"sCornell University, Ithaca, NY. Part of this work was done while the author was an intern at Yahoo",
"1Overture.com, formerly GoTo.com and later bought by Yahoo, introduced the first sponsored search auctions in",
"1996 and was still the dominant player in 2002. Their auction had pay-your-bid pricing that has since been dubbed",
"thegeneralizedfirstpriceauction(GFP)",
"2Technically, the “\"ad effect\" term in the click probability appears in both the optimization and pricing, but these",
"can be thought of as bidder weights; analysis of GSP's equilibria does not require that they are accurate. This",
"simplicity also makes implementation easier and is a practical reason GSP is common.",
"1"
] |
{
"arxiv_id": "1701.05946",
"language": "zh",
"page_idx": 0
}
|
||
1701.05946-zh-1
|
1701.05946-zh-1.jpg
|
我们将过去的分析转到头上，并提出 GSP 的失控成功来自一个更简单的解释：GSPistruthful ,当然，众所周知，VCG 是真实的$forprofit\textit{ maximizers 。实 际 上 , 广 告 商 很 少 为 利 润 而 优 化 }$ 而且，正如我们将讨论的那样，他们的目标和行为更接近于我们称为$value\textit{ maximization 的 模 型 。}$ 事实上，很容易看出普惠制对这类投标人是真实的——Aggarwal 等人。[4] 在一般偏好模型的极限情况下正好注意到这一点一面临的挑战是认为广告商的偏好generally符合我们的价值最大化模型，我们通过两种方式做到这一点。首先，我们基于广告商的传统目标，认为价值最大化模型是正确的最小理论模型。其次，我们证明并通过经验验证，具有适度 ROI 约束的任何投标人的行为都将等同于价值最大化者，从而使价值最大化成为适合广告客户偏好的稳健选择
此外，我们的工作提出了 GSP 的基本公理定义：拍卖使广告商的价值总和最大化，并为价值最大化者收取真实的价格。这是第一项为(e。g。)当搜索广告超出其当前模型时，推广GSP。在此过程中，我们证明了民间传说的定义，即普惠制价格是维持相同分配所需的最低出价
通过这些结果，我们对普惠制有了新的认识。它作为一种真实的拍卖获得了坚实的理论基础， 被价值最大化偏好可以以多种方式证明的事实所放大。这种美德因其优雅的朴素而增强，使其成为名副其实的灰姑娘-错误的拍卖结果恰到好处
广告客户可以分为两类：brand和performance 。品牌广告商的目标是长期增长和知名度。他们的目标是接触到尽可能多的人，产生尽可能多的点击，捕获尽可能多的市场份额，产生尽可能多的收入，等等。相比之下，广告客户关心的是立竿见影的效果。他们针对直接从广告中生成的销售， 注册或其他所谓的conversions进行优化
Preferences of Brand Advertisers 品牌广告商带着明确的目标来到市场。他们通常有一个任务，以满足一个特定的业务目标——展示给观众的印象，产生点击，或最大化收入——驱动的长期考虑，而不是眼前的利润。因此，成本虽然很重要，但只是作为一种约束进人他们的偏好-品牌广告商将有一个预算和限制，他们愿意平均支付印象/点击/转化，但否则直接优化他们的任务
Preferences of Performance Advetisers 性能广告客户优化价值和成本之间的直接权衡。几十年来，投资回报率 (ROI) 一直是衡量所有类型广告之间权衡的标准指标。ROI 衡量获得的利润(“ 回 报 ” ) 与 支 付 的 成 本 或 价 格 ( “ 投 资 ” ) 的 比 率 , i。e。, 成 本 中 的 利 润 密 度 :
1.1 Preferences of Online Advertisers
$$\text{ROI}=\frac{\text{Value-Price}}{\text{Price}}=\frac{\text{Profit}}{\text{Price}}$$
作为密度度量，ROI 的无约束最大化是不明智的；$^3$取而代之的是，广告商带来了 ROI 约束并进
行了优化 (e。g。最大化收入和/或点击)。例如，以下故事在 Yahoo 上定期播放：
1. Advertiser X designates a small budget for testing a Yahoo advertising product.
2. Advertiser X measures ROI-if X is happy with its ROI, it increases its budget hoping to
$maintain\textit{ the same ROI ; if it is unhappy, it withdraws. }$
此行为也反映在标准行业工具中：Google 的 AdWords 广告系列管理工具在保持目标平均 CPC 和预算的同时购买尽可能多的广告(平均 CPC 约束对应于 ROI 约束，而边际 CPC 约束则适合于利润最大化)。雅虎的展示广告产品还提供了具有“CPC 目标”和“CPA 目标”的选项，这些选项可以购买展示次数，同时旨在分别实现目标平均每次点击成本或每次转换成本
$^{3}$Unconstrained maximization of ROI would likely cause an advertiser to buy only the single cheapest impression
or click available.
2
|
[
[
0.1446533203125,
0.09344482421875,
0.8837890625,
0.09344482421875,
0.8837890625,
0.10797119140625,
0.1446533203125,
0.10797119140625
],
[
0.11767578125,
0.111083984375,
0.8798828125,
0.111083984375,
0.8798828125,
0.125,
0.11767578125,
0.125
],
[
0.1192626953125,
0.1275634765625,
0.88134765625,
0.1275634765625,
0.88134765625,
0.141357421875,
0.1192626953125,
0.141357421875
],
[
0.11767578125,
0.1458740234375,
0.8798828125,
0.1458740234375,
0.8798828125,
0.1590576171875,
0.11767578125,
0.1590576171875
],
[
0.1192626953125,
0.1622314453125,
0.88134765625,
0.1622314453125,
0.88134765625,
0.1767578125,
0.1192626953125,
0.1767578125
],
[
0.11767578125,
0.1787109375,
0.88134765625,
0.1787109375,
0.88134765625,
0.1925048828125,
0.11767578125,
0.1925048828125
],
[
0.11767578125,
0.1976318359375,
0.88134765625,
0.1976318359375,
0.88134765625,
0.2108154296875,
0.11767578125,
0.2108154296875
],
[
0.1168212890625,
0.2139892578125,
0.7236328125,
0.2139892578125,
0.7236328125,
0.2279052734375,
0.1168212890625,
0.2279052734375
],
[
0.146240234375,
0.23291015625,
0.8798828125,
0.23291015625,
0.8798828125,
0.24365234375,
0.146240234375,
0.24365234375
],
[
0.1192626953125,
0.248779296875,
0.88134765625,
0.248779296875,
0.88134765625,
0.26318359375,
0.1192626953125,
0.26318359375
],
[
0.1168212890625,
0.26513671875,
0.77685546875,
0.26513671875,
0.77685546875,
0.279052734375,
0.1168212890625,
0.279052734375
],
[
0.1446533203125,
0.281005859375,
0.8798828125,
0.2822265625,
0.8798828125,
0.296630859375,
0.1446533203125,
0.29541015625
],
[
0.1192626953125,
0.2998046875,
0.88134765625,
0.2998046875,
0.88134765625,
0.313232421875,
0.1192626953125,
0.313232421875
],
[
0.1192626953125,
0.31640625,
0.461669921875,
0.31640625,
0.461669921875,
0.330078125,
0.1192626953125,
0.330078125
],
[
0.1168212890625,
0.352783203125,
0.494384765625,
0.352783203125,
0.494384765625,
0.3662109375,
0.1168212890625,
0.3662109375
],
[
0.11767578125,
0.37890625,
0.8798828125,
0.37890625,
0.8798828125,
0.393310546875,
0.11767578125,
0.393310546875
],
[
0.1192626953125,
0.395751953125,
0.8798828125,
0.395751953125,
0.8798828125,
0.4091796875,
0.1192626953125,
0.4091796875
],
[
0.11767578125,
0.41357421875,
0.88134765625,
0.41357421875,
0.88134765625,
0.4267578125,
0.11767578125,
0.4267578125
],
[
0.1168212890625,
0.429931640625,
0.438720703125,
0.43115234375,
0.438720703125,
0.44580078125,
0.1168212890625,
0.4443359375
],
[
0.12335205078125,
0.46533203125,
0.88134765625,
0.466552734375,
0.88134765625,
0.48095703125,
0.12335205078125,
0.479736328125
],
[
0.1192626953125,
0.484130859375,
0.88134765625,
0.484130859375,
0.88134765625,
0.49755859375,
0.1192626953125,
0.49755859375
],
[
0.11767578125,
0.50048828125,
0.88134765625,
0.50048828125,
0.88134765625,
0.51513671875,
0.11767578125,
0.51513671875
],
[
0.11767578125,
0.51806640625,
0.84619140625,
0.51806640625,
0.84619140625,
0.5322265625,
0.11767578125,
0.5322265625
],
[
0.12335205078125,
0.5537109375,
0.88134765625,
0.5537109375,
0.88134765625,
0.5673828125,
0.12335205078125,
0.5673828125
],
[
0.12091064453125,
0.57080078125,
0.88134765625,
0.57080078125,
0.88134765625,
0.583984375,
0.12091064453125,
0.583984375
],
[
0.11767578125,
0.58837890625,
0.744140625,
0.58837890625,
0.744140625,
0.6015625,
0.11767578125,
0.6015625
],
[
0.366943359375,
0.611328125,
0.61767578125,
0.60400390625,
0.619140625,
0.6318359375,
0.36767578125,
0.6396484375
],
[
0.453369140625,
0.6318359375,
0.5048828125,
0.6318359375,
0.5048828125,
0.646484375,
0.453369140625,
0.646484375
],
[
0.5595703125,
0.6298828125,
0.609375,
0.6318359375,
0.60888671875,
0.64794921875,
0.55810546875,
0.64599609375
],
[
0.1192626953125,
0.65576171875,
0.88134765625,
0.65576171875,
0.88134765625,
0.67041015625,
0.1192626953125,
0.67041015625
],
[
0.1168212890625,
0.673828125,
0.74658203125,
0.673828125,
0.74658203125,
0.68798828125,
0.1168212890625,
0.68798828125
],
[
0.136474609375,
0.69873046875,
0.78515625,
0.69970703125,
0.78515625,
0.7138671875,
0.136474609375,
0.71337890625
],
[
0.134033203125,
0.724609375,
0.88134765625,
0.72607421875,
0.88134765625,
0.74365234375,
0.134033203125,
0.7421875
],
[
0.16259765625,
0.744140625,
0.6005859375,
0.744140625,
0.6005859375,
0.7587890625,
0.16259765625,
0.7587890625
],
[
0.1192626953125,
0.76953125,
0.87841796875,
0.76953125,
0.87841796875,
0.7841796875,
0.1192626953125,
0.7841796875
],
[
0.1168212890625,
0.787109375,
0.88134765625,
0.78857421875,
0.88134765625,
0.80322265625,
0.1168212890625,
0.8017578125
],
[
0.11767578125,
0.80615234375,
0.8798828125,
0.80615234375,
0.8798828125,
0.8193359375,
0.11767578125,
0.8193359375
],
[
0.11767578125,
0.82373046875,
0.7060546875,
0.82373046875,
0.7060546875,
0.8369140625,
0.11767578125,
0.8369140625
],
[
0.134033203125,
0.84521484375,
0.88134765625,
0.84765625,
0.88134765625,
0.8623046875,
0.134033203125,
0.85986328125
],
[
0.11767578125,
0.86279296875,
0.2296142578125,
0.86279296875,
0.2296142578125,
0.87353515625,
0.11767578125,
0.87353515625
],
[
0.491943359375,
0.89794921875,
0.50732421875,
0.89794921875,
0.50732421875,
0.91357421875,
0.491943359375,
0.91357421875
]
] |
[
"我们将过去的分析转到头上，并提出GSP的失控成功来自一个更简单的解释：GSPistruthful",
"。当然，众所周知，VCG是真实的forprofitmacimizers。实际上，广告商很少为利润而优化",
"而且，正如我们将讨论的那样，他们的目标和行为更接近于我们称为valuemarimization的模型。",
"事实上，很容易看出普惠制对这类投标人是真实的一一Aggarwal等人。[4]在一般偏好模型的极",
"限情况下正好注意到这一点一面临的挑战是认为广告商的偏好generally符合我们的价值最大化模",
"型，我们通过两种方式做到这一点。首先，我们基于广告商的传统目标，认为价值最大化模型是正",
"确的最小理论模型。其次，我们证明并通过经验验证，具有适度ROI约束的任何投标人的行为都",
"将等同于价值最大化者，从而使价值最大化成为适合广告客户偏好的稳健选择",
"此外，我们的工作提出GSP的基本公理定义：拍卖使厂告商的价值总和最天化，并为价值最",
"大化者收取真实的价格。这是第一项为（e。g。）当搜索广告超出其当前模型时，推广GSP。在此",
"过程中，我们证明了民间传说的定义，即普惠制价格是维持相同分配所需的最低出价",
"通过这些结果，我们对普惠制有了新的认识。它作为一种真实的拍卖获得了坚实的理论基础",
"被价值最大化偏好可以以多种方式证明的事实所放大。这种美德因其优雅的朴素而增强，使其成为",
"名副其实的灰姑娘-错误的拍卖结果恰到好处",
"1.1Preferences of Online Advertisers",
"广告客户可以分为两类：brand和performance。品牌广告商的目标是长期增长和知名度。他们",
"的目标是接触到尽可能多的人，产生尽可能多的点击，捕获尽可能多的市场份额，产生尽可能多的",
"收入，等等。相比之下，广告客户关心的是立竿见影的效果。他们针对直接从广告中生成的销售：",
"注册或其他所谓的conversions进行优化",
"PreferencesofBrandAdvertisers品牌广告商带着明确的目标来到市场。他们通常有一个",
"任务，以满足一个特定的业务目标一一展示给观众的印象，产生点击，或最大化收入一一驱动的长",
"期考虑，而不是眼前的利润。因此，成本虽然很重要，但只是作为一种约束进人他们的偏好-品牌",
"广告商将有一个预算和限制，他们愿意平均支付印象/点击/转化，但否则直接优化他们的任务",
"PreferencesofPerformanceAdvetisers性能广告客户优化价值和成本之间的直接权衡。几",
"十年来，投资回报率（ROI）一直是衡量所有类型广告之间权衡的标准指标。ROI衡量获得的利润",
"（“回报”）与支付的成本或价格（“投资”）的比率，i。e。，成本中的利润密度：",
"ROI=Value-Price_ Profit",
"Price",
"Price",
"作为密度度量，ROI的无约束最大化是不明智的；3取而代之的是，广告商带来了ROI约束并进",
"行了优化（e。g。最大化收人和/或点击)。例如，以下故事在Yahoo上定期播放：",
"1. Advertiser X designates a small budget for testing a Yahoo advertising product.",
"2. Advertiser X measures ROI - if X is happy with its ROI, it increases its budget hoping to",
"maintain the same ROI; if it isunhappy,it withdraws.",
"此行为也反映在标准行业工具中：Google的AdWords广告系列管理工具在保持目标平均CPC和",
"预算的同时购买尽可能多的广告（平均CPC约束对应于ROI约束，而边际CPC约束则适合于利",
"润最大化）。雅虎的展示广告产品还提供了具有“CPC目标”和“CPA目标”的选项，这些选项可",
"以购买展示次数，同时旨在分别实现目标平均每次点击成本或每次转换成本",
"3Unconstrained maximization of ROI would likely cause an advertiser to buy only the single cheapest impression",
"orclickavailable",
"2"
] |
{
"arxiv_id": "1701.05946",
"language": "zh",
"page_idx": 1
}
|
||
1701.05946-zh-2
|
1701.05946-zh-2.jpg
|
性能广告客户以 ROI 为中心的行为需要评论，因为它doesnot最大化利润。标准直觉表明积极在价值和成本之间进行权衡的广告商应该最大化利润，但很容易看出，ROI 约束无法购买一些有利可图的广告。解决这种冲突的一种方法是得出结论，直觉是错误的，也许是因为公司经常像关心利润一样关心收人和利润率。另一个解决方案是得出结论，ROI 是一种根深蒂固的启发式方法，可以最大程度地提高固定资源的有效性，可以在拍卖之间或平台之间管理广告预算E [11,20,26,29,8] (决定如何在 Yahoo 和 Google 之间出价),甚至可以管理广告客户组织内的资源(在营销和工程之间分配预算)。然而，无论如何，广告客户的行为很少是利润最大化的
Advertisers as Value Maximizers 我们刚刚看到广告商的目标符合一个简单的范例，我们称之为$value\textit{ maximizing : 他 们 在 支 出 限 制 的 情 况 下 最 大 化 他 们 购 买 的 东 西 ( 价 值 ) 。最 简 单 的 支 }$ 出约束是总价格不超过所购买商品的总价值。这激发了我们的最小理论模型：$^{4}$
$\textbf{Definition 1 }($Informal$).A$ simple value maximizer $maximizes\textit{value v}_iwhile\textit{maintaining budget}$
$balance$, $i. e.$ while keeping payment $p_{i}\leq v_{i}$ ; when value $is$ $equal$, $a$ lower price $is$ $preferred.$
令人惊讶的是，这个简单的模型甚至可以捕获具有预算和 ROI 约束的广告客户的拍卖级偏好。预算限制可以在拍卖层面被忽略，因为它们通常比出价大几个数量级。相反，ROI 约束在拍卖级别上具有实质性影响-至少需要 ROI $\gamma$的值最大化器将需要$\frac vi(o)-p_i{p_i}\geq\gamma$ ,因此$p_i\leq\frac{v_i(o)}{1+\gamma}$ -但这看起来像是具有不同值的简单值最大化器：
$\textbf{Observation 1. 1 ( Informal) . If an auction is truthful for simple value maximizers, value maxi- }$
$mizerswithROI\textit{constraints will“ truthfully” report v}_{i}^{\prime }= \frac {v_{i}}{1+ \gamma }$ .
因此，如果一个 naTHXMAXXXXX。e。他们报告他们愿意支付的费用(拍卖人不需要知道$\gamma$
,也不需要对所有竞标者都一样)。我们稍后将更正式地讨论这些论点
1.2 Related Work
Aggarwal 等人。[4]设计赞助搜索 (槽) 拍卖在一般匹配模型。在他们的模型中，每个广告客户都有一个特定于插槽的价值和一个特定于插槽的最高价格；广告客户在其最高价格约束下最大化利润。在价值较大的限制下，他们将广告商称为maximumprice投标人，直接与我们的价值最大化者相对应，并观察到 GSP 对于最高价格投标人是真实的。他们将最高价格竞标者视为市场的特殊角落，而不是所有广告客户的一般定义，并且他们不会以结果为基础来证明/解释 GSP
有限的实证文献研究了投标人的行为。瓦里安[27]测试出价是否可以解释为一些拟线性投标人的均衡，并发现误差不大；但这并不能解决我们的情况，因为假设投标人是价值最大化者将导致零误差 (每个人都只是在说实话)。Athey 和 Nekipelov$\mathbb{E}[7]$假定广告商具有广泛的异质目标，并对其行为进行建模
代理最大化价值的想法很常见，尽管不是在机制设计中。例如，传统的消费者选择和市场均衡研究始于代理商，他们根据自己的禀赋E[21] 选择他们可以负担得起的最喜欢的捆绑包。。在机制设计的背景下，只有有限的文献研究了非准线性投标人。Fadaei 和 Bichler 的并发工作介绍了“价值出价者”的相关模型$\mathbb{E}[18]$ ,然后研究了收益最大化的一种形式。我们的模型明显不同，因为我们说当价值保持不变时，竞标者更喜欢较低的价格，而 Fadaei 和 Bichler 暗示竞标者严格地喜欢花钱。Alaei 等人。[5] 展示了如何利用非拟线性单位需求设置中的瓦尔拉斯均衡来构建广告拍卖， 其中估计误差打破了拟线性。几篇论文研究了非拟线性偏好的一般真实机制[2,23] 。这些论文为多单元和单元需求设置开发了 VCG 模拟的公理特征
$^{4}$A similar model was discovered independently in concurrent work by Fadaei and Bichler [18]. See related work
3
|
[
[
0.1446533203125,
0.0921630859375,
0.87841796875,
0.0921630859375,
0.87841796875,
0.106689453125,
0.1446533203125,
0.106689453125
],
[
0.11767578125,
0.111083984375,
0.88134765625,
0.111083984375,
0.88134765625,
0.125,
0.11767578125,
0.125
],
[
0.1282958984375,
0.1319580078125,
0.1356201171875,
0.1319580078125,
0.1356201171875,
0.1363525390625,
0.1282958984375,
0.1363525390625
],
[
0.131591796875,
0.1275634765625,
0.88134765625,
0.1275634765625,
0.88134765625,
0.141357421875,
0.131591796875,
0.141357421875
],
[
0.11767578125,
0.1458740234375,
0.88134765625,
0.1458740234375,
0.88134765625,
0.1590576171875,
0.11767578125,
0.1590576171875
],
[
0.1192626953125,
0.1622314453125,
0.88232421875,
0.1622314453125,
0.88232421875,
0.175537109375,
0.1192626953125,
0.175537109375
],
[
0.1192626953125,
0.1787109375,
0.88134765625,
0.1787109375,
0.88134765625,
0.1932373046875,
0.1192626953125,
0.1932373046875
],
[
0.11767578125,
0.1976318359375,
0.8203125,
0.1976318359375,
0.8203125,
0.2108154296875,
0.11767578125,
0.2108154296875
],
[
0.12335205078125,
0.231689453125,
0.8798828125,
0.23291015625,
0.8798828125,
0.2474365234375,
0.12335205078125,
0.246826171875
],
[
0.11767578125,
0.250732421875,
0.8798828125,
0.250732421875,
0.8798828125,
0.264404296875,
0.11767578125,
0.264404296875
],
[
0.11767578125,
0.266357421875,
0.73388671875,
0.26513671875,
0.73388671875,
0.27978515625,
0.11767578125,
0.281005859375
],
[
0.1168212890625,
0.294189453125,
0.8798828125,
0.294189453125,
0.8798828125,
0.309326171875,
0.1168212890625,
0.309326171875
],
[
0.11517333984375,
0.311279296875,
0.82861328125,
0.31005859375,
0.82861328125,
0.32763671875,
0.11517333984375,
0.328857421875
],
[
0.146240234375,
0.341552734375,
0.88134765625,
0.341552734375,
0.88134765625,
0.35595703125,
0.146240234375,
0.35595703125
],
[
0.11767578125,
0.359130859375,
0.88134765625,
0.359130859375,
0.88134765625,
0.37255859375,
0.11767578125,
0.37255859375
],
[
0.111083984375,
0.365478515625,
0.88623046875,
0.369384765625,
0.88623046875,
0.397705078125,
0.111083984375,
0.393310546875
],
[
0.1168212890625,
0.39453125,
0.41015625,
0.39453125,
0.41015625,
0.40771484375,
0.1168212890625,
0.40771484375
],
[
0.1168212890625,
0.421630859375,
0.8798828125,
0.42236328125,
0.8798828125,
0.4375,
0.1168212890625,
0.436279296875
],
[
0.11517333984375,
0.43505859375,
0.60888671875,
0.438232421875,
0.60888671875,
0.458984375,
0.11517333984375,
0.455810546875
],
[
0.1422119140625,
0.46728515625,
0.8837890625,
0.468505859375,
0.8837890625,
0.486083984375,
0.1422119140625,
0.48486328125
],
[
0.1168212890625,
0.488037109375,
0.66015625,
0.488037109375,
0.66015625,
0.501953125,
0.1168212890625,
0.501953125
],
[
0.1168212890625,
0.52294921875,
0.3056640625,
0.52294921875,
0.3056640625,
0.537109375,
0.1168212890625,
0.537109375
],
[
0.11767578125,
0.5498046875,
0.88134765625,
0.5498046875,
0.88134765625,
0.564453125,
0.11767578125,
0.564453125
],
[
0.11767578125,
0.5673828125,
0.88134765625,
0.5673828125,
0.88134765625,
0.58056640625,
0.11767578125,
0.58056640625
],
[
0.1168212890625,
0.58544921875,
0.53125,
0.58544921875,
0.53125,
0.5986328125,
0.1168212890625,
0.5986328125
],
[
0.56640625,
0.58544921875,
0.88134765625,
0.58544921875,
0.88134765625,
0.5986328125,
0.56640625,
0.5986328125
],
[
0.11767578125,
0.6015625,
0.88134765625,
0.6015625,
0.88134765625,
0.61572265625,
0.11767578125,
0.61572265625
],
[
0.1168212890625,
0.61865234375,
0.7900390625,
0.6171875,
0.7900390625,
0.6318359375,
0.1168212890625,
0.63330078125
],
[
0.1446533203125,
0.63623046875,
0.88134765625,
0.63623046875,
0.88134765625,
0.6494140625,
0.1446533203125,
0.6494140625
],
[
0.1192626953125,
0.65283203125,
0.88134765625,
0.65283203125,
0.88134765625,
0.66650390625,
0.1192626953125,
0.66650390625
],
[
0.11767578125,
0.66845703125,
0.8798828125,
0.66845703125,
0.8798828125,
0.68310546875,
0.11767578125,
0.68310546875
],
[
0.11767578125,
0.6875,
0.24267578125,
0.6875,
0.24267578125,
0.70068359375,
0.11767578125,
0.70068359375
],
[
0.146240234375,
0.7041015625,
0.88134765625,
0.7041015625,
0.88134765625,
0.71728515625,
0.146240234375,
0.71728515625
],
[
0.11767578125,
0.7216796875,
0.8798828125,
0.7216796875,
0.8798828125,
0.73486328125,
0.11767578125,
0.73486328125
],
[
0.11767578125,
0.73779296875,
0.88134765625,
0.73779296875,
0.88134765625,
0.751953125,
0.11767578125,
0.751953125
],
[
0.11767578125,
0.755859375,
0.88134765625,
0.755859375,
0.88134765625,
0.76953125,
0.11767578125,
0.76953125
],
[
0.11767578125,
0.77294921875,
0.88232421875,
0.77294921875,
0.88232421875,
0.787109375,
0.11767578125,
0.787109375
],
[
0.11767578125,
0.79052734375,
0.88134765625,
0.79052734375,
0.88134765625,
0.8037109375,
0.11767578125,
0.8037109375
],
[
0.11767578125,
0.806640625,
0.88134765625,
0.806640625,
0.88134765625,
0.82080078125,
0.11767578125,
0.82080078125
],
[
0.11767578125,
0.82275390625,
0.52685546875,
0.82373046875,
0.52685546875,
0.83837890625,
0.11767578125,
0.8369140625
],
[
0.1356201171875,
0.8486328125,
0.8798828125,
0.849609375,
0.8798828125,
0.8642578125,
0.1356201171875,
0.86279296875
],
[
0.491943359375,
0.89794921875,
0.5087890625,
0.89794921875,
0.5087890625,
0.91357421875,
0.491943359375,
0.91357421875
]
] |
[
"性能广告客户以ROI为中心的行为需要评论，因为它doesnot最大化利润。标准直觉表明",
"积极在价值和成本之间进行权衡的广告商应该最大化利润，但很容易看出，ROI约束无法购买",
"",
"些有利可图的广告。解决这种冲突的一种方法是得出结论，直觉是错误的，也许是因为公司经",
"常像关心利润一样关心收人和利润率。另一个解决方案是得出结论，ROI是一种根深蒂固的启",
"发式方法，可以最大程度地提高固定资源的有效性，可以在拍卖之间或平台之间管理广告预算E",
"[11，20,26，29，8]（决定如何在Yahoo和Google之间出价），甚至可以管理广告客户组织内的资",
"源（在营销和工程之间分配预算）。然而，无论如何，广告客户的行为很少是利润最大化的",
"AdvertisersasValueMaximizers我们刚刚看到广告商的目标符合一个简单的范例，我们",
"称之为valuemarimizing：他们在支出限制的情况下最大化他们购买的东西（价值）。最简单的支",
"出约束是总价格不超过所购买商品的总价值。这激发了我们的最小理论模型：4",
"Definition 1 (Informal). A simple value maximizer marimizes value vi while maintaining budgei",
"balance, i.e. while keeping payment pi ≤ vi ; when value is equal, a lower price is preferred.",
"令人惊牙的是，这个简单的模型甚至可以捕获具有预算和ROI约束的广告客户的拍卖级偏好。",
"预算限制可以在拍卖层面被忽略，因为它们通常比出价大几个数量级。相反，ROI约束在拍卖级别",
"上具有实质性影响-至少需要ROI的值最大化器将需要“≥，因此Pi≤一但这看起",
"来像是具有不同值的简单值最大化器：",
"Observation 1.1 (Informal). If an auction is truthful for simple value marimizers, value mari-",
"mizers with ROI constraints will “truthfully\" report u', = .",
"因此，如果一个naTHXMAXXXXX。e。他们报告他们愿意支付的费用（拍卖人不需要知道",
"，也不需要对所有竞标者都一样）。我们稍后将更正式地讨论这些论点",
"1.2Related Work",
"Aggarwal等人。[4]设计赞助搜索（槽）拍卖在一般匹配模型。在他们的模型中，每个广告客户都",
"有一个特定于插槽的价值和一个特定于插槽的最高价格；广告客户在其最高价格约束下最大化利",
"润。在价值较大的限制下，他们将广告商称为macim",
"price投标人，直接与我们的价值最大化",
"者相对应，并观察到GSP对于最高价格投标人是真实的。他们将最高价格竞标者视为市场的特殊",
"角落，而不是所有广告客户的一般定义，并且他们不会以结果为基础来证明/解释GSP",
"有限的实证文献研究了投标人的行为。瓦里安[27]测试出价是否可以解释为一些拟线性投标人",
"的均衡，并发现误差不大；但这并不能解决我们的情况，因为假设投标人是价值最大化者将导致零",
"误差（每个人都只是在说实话）。Athey和NekipelovE[7]假定广告商具有广泛的异质目标，并对",
"其行为进行建模",
"代理最大化价值的想法很常见，尽管不是在机制设计中。例如，传统的消费者选择和市场均衡",
"研究始于代理商，他们根据自已的票赋E21]选择他们可以负担得起的最喜欢的捆绑包。。在机制",
"设计的背景下，只有有限的文献研究了非准线性投标人。Fadaei和Bichler的并发工作介绍了“价",
"值出价者”的相关模型E[18]，然后研究了收益最大化的一种形式。我们的模型明显不同，因为我",
"们说当价值保持不变时，竞标者更喜欢较低的价格，而Fadaei和Bichler暗示竞标者严格地喜欢",
"花钱。Alaei等人。[5]展示了如何利用非拟线性单位需求设置中的瓦尔拉斯均衡来构建广告拍卖，",
"其中估计误差打破了拟线性。几篇论文研究了非拟线性偏好的一般真实机制[2，23]。这些论文为",
"多单元和单元需求设置开发了VCG模拟的公理特征",
"4A similar model was discovered independently in concurrent work by Fadaei and Bichler [18] . See related work",
"3"
] |
{
"arxiv_id": "1701.05946",
"language": "zh",
"page_idx": 2
}
|
||
1701.05946-zh-3
|
1701.05946-zh-3.jpg
|
qasilinear 广告文献中的一些线程也是相关的。一项工作研究了使用 ROI 作为启发式方法在不同拍卖中设置出价的效果[11,20,26,29,8]。另一种工作是使用行为经济学的思想来捕捉观察到的行为[25] 。最后，一个小线程研究了 GSP 以及将其推广到新环境的问题。拟线性世界中 GSP 的规范分析是[17,27] 。试图推广 GSP 并考虑其在标准模型之外的性能，包括[13,19,1,3]
## 2 Model and Preliminaries
一组$n$广告将根据配置(结果)$o\in\mathcal{O}$放置在页面上。对于每个配置，用户点击广告$i$ ($i$的分配)
的可能性由$x_i(o)$给出；当用户点击时，广告商$i$支付价格$p_i$
每个广告客户每次点击都有一个私有值$v_i$和一个公共$^{5}$完整且可传递的首选项$\preceq_i$,而不是结
果-价格对$(o,p)$
我们将在整篇论文中引用一些不同的偏好模型。首先，我们提到标准的拟线性偏好：
$\textbf{Definition 2. Preferences are quasi- linear }if\textit{a bidder prefers to maximize the utility function}$
$u_i= ( v_i- p_i) x_i$ , $i. e.$
$$(o,p_i)\succeq_i(o',p'_i)\Leftrightarrow(v_i-p_i)x_i(o)\geq(v_i-p'_i)x_i(o')\:.$$
本文的大部分内容将侧重于各种形式的价值最大化偏好。为任意约束$c$定义了一般的值最大化
首选项
$\textbf{Definition 3. Preferences are value- maximzing }subject\textit{ to constraints C if they always prefer anditiont }$
$outcomewithhighervalueas\textit{long as none of its constraints are violated}.$
$$v_{i}x_{i}(o)>v_{i}x_{i}(o^{\prime})\:,\:(o,p_{i})\:satisfies\:\mathcal{C}\quad\Rightarrow\quad(o,p_{i})\succ_{i}(o^{\prime},p_{i}^{\prime})\:,$$
$and~when~the~value~(allocation)~is~the~same,~a~lower~price~is~preferred:$
$$x_{i}(o)=x_{i}(o')\:,\:p_{i}<p_{i}'\:,\:(o,p_{i})\:satisfies\:\mathcal{C}\Rightarrow(o,p_{i})\succ_{i}(o',p_{i}')\:.$$
$Moreover, any\textit{ outcome is preferred to the one violating C}:$
$$(o',p'_i)\:violates\:\mathcal{C}\Rightarrow(o,p_i)\succeq_i(o,p'_i)\:.$$
我们将主要使用两种特定类型的单调约束：
$\textbf{Definition 4. }A\textit{bidder has simple value- maximizing preferences }if\textit{the bidder is value- }maximizing$
subject $to$ the constraint that $( o, p_i)$ must satisfy $p_i\leq v_i$ .
$\textbf{Definition 5. }A\textit{ bidder has value- maximizing preferences with an ROI constraint }\gamma \textit{ if the bidder}$
$is\textit{ value- maximizing with the constraint }\frac {v_i- p_i}{p_i}\geq \gamma$ .
Auctions and Truthfulness 拍卖要求投标人报告其价值参数，并为投标人提交的投标$b$指
定结果$f(\boldsymbol{b})$和付款$p_i(\boldsymbol{b})$ 。函数$f$被称为$social\textit{choice function , 它 意 味 着 分 配 }x_i( \boldsymbol{b})$
如果报告投标人的真实私人信息 (例如。g。出价$b_i=v_i$ )是一种主导策略：
$^{5}$We assume this is public for simplicity, but it is not necessary for the results in our paper
4
|
[
[
0.143798828125,
0.09344482421875,
0.8798828125,
0.0921630859375,
0.8798828125,
0.106689453125,
0.143798828125,
0.10797119140625
],
[
0.1192626953125,
0.111083984375,
0.8798828125,
0.111083984375,
0.8798828125,
0.125,
0.1192626953125,
0.125
],
[
0.1192626953125,
0.1275634765625,
0.8798828125,
0.1275634765625,
0.8798828125,
0.1427001953125,
0.1192626953125,
0.1427001953125
],
[
0.1192626953125,
0.1458740234375,
0.84619140625,
0.1458740234375,
0.84619140625,
0.1590576171875,
0.1192626953125,
0.1590576171875
],
[
0.11517333984375,
0.1875,
0.44677734375,
0.1875,
0.44677734375,
0.2020263671875,
0.11517333984375,
0.2020263671875
],
[
0.11767578125,
0.2197265625,
0.88134765625,
0.2197265625,
0.88134765625,
0.2342529296875,
0.11767578125,
0.2342529296875
],
[
0.1192626953125,
0.2374267578125,
0.5986328125,
0.2374267578125,
0.5986328125,
0.250732421875,
0.1192626953125,
0.250732421875
],
[
0.143798828125,
0.250732421875,
0.88134765625,
0.251953125,
0.88134765625,
0.26953125,
0.143798828125,
0.268310546875
],
[
0.11767578125,
0.270751953125,
0.240234375,
0.270751953125,
0.240234375,
0.285400390625,
0.11767578125,
0.285400390625
],
[
0.146240234375,
0.28857421875,
0.7939453125,
0.28857421875,
0.7939453125,
0.3017578125,
0.146240234375,
0.3017578125
],
[
0.11358642578125,
0.313232421875,
0.8837890625,
0.314453125,
0.8837890625,
0.33203125,
0.11358642578125,
0.330810546875
],
[
0.1168212890625,
0.333984375,
0.281982421875,
0.333984375,
0.281982421875,
0.348388671875,
0.1168212890625,
0.348388671875
],
[
0.298095703125,
0.363037109375,
0.705078125,
0.363037109375,
0.705078125,
0.380615234375,
0.298095703125,
0.380615234375
],
[
0.1446533203125,
0.393310546875,
0.8798828125,
0.39453125,
0.8798828125,
0.4091796875,
0.1446533203125,
0.40771484375
],
[
0.11767578125,
0.412353515625,
0.17236328125,
0.412353515625,
0.17236328125,
0.4267578125,
0.11767578125,
0.4267578125
],
[
0.1168212890625,
0.43994140625,
0.8798828125,
0.43994140625,
0.8798828125,
0.455078125,
0.1168212890625,
0.455078125
],
[
0.11767578125,
0.458251953125,
0.68701171875,
0.458251953125,
0.68701171875,
0.472900390625,
0.11767578125,
0.472900390625
],
[
0.25732421875,
0.4892578125,
0.744140625,
0.4892578125,
0.744140625,
0.50390625,
0.25732421875,
0.50390625
],
[
0.12091064453125,
0.51953125,
0.66845703125,
0.51953125,
0.66845703125,
0.5341796875,
0.12091064453125,
0.5341796875
],
[
0.25244140625,
0.5517578125,
0.74609375,
0.5517578125,
0.74609375,
0.56640625,
0.25244140625,
0.56640625
],
[
0.1168212890625,
0.58056640625,
0.58251953125,
0.58056640625,
0.58251953125,
0.5986328125,
0.1168212890625,
0.5986328125
],
[
0.3544921875,
0.61279296875,
0.646484375,
0.61279296875,
0.646484375,
0.62744140625,
0.3544921875,
0.62744140625
],
[
0.146240234375,
0.64306640625,
0.47216796875,
0.64306640625,
0.47216796875,
0.65771484375,
0.146240234375,
0.65771484375
],
[
0.11358642578125,
0.66943359375,
0.88134765625,
0.67041015625,
0.88134765625,
0.68798828125,
0.11358642578125,
0.6875
],
[
0.11767578125,
0.68603515625,
0.564453125,
0.68798828125,
0.564453125,
0.70556640625,
0.11767578125,
0.7041015625
],
[
0.11767578125,
0.71728515625,
0.88134765625,
0.71728515625,
0.88134765625,
0.73193359375,
0.11767578125,
0.73193359375
],
[
0.11358642578125,
0.72998046875,
0.515625,
0.73193359375,
0.515625,
0.75244140625,
0.11358642578125,
0.75048828125
],
[
0.441162109375,
0.74853515625,
0.451904296875,
0.74853515625,
0.451904296875,
0.751953125,
0.441162109375,
0.751953125
],
[
0.1258544921875,
0.77294921875,
0.88134765625,
0.77294921875,
0.88134765625,
0.7861328125,
0.1258544921875,
0.7861328125
],
[
0.1192626953125,
0.79052734375,
0.80126953125,
0.79052734375,
0.80126953125,
0.8037109375,
0.1192626953125,
0.8037109375
],
[
0.146240234375,
0.806640625,
0.73193359375,
0.806640625,
0.73193359375,
0.82080078125,
0.146240234375,
0.82080078125
],
[
0.13232421875,
0.83056640625,
0.740234375,
0.83349609375,
0.740234375,
0.84765625,
0.13232421875,
0.84521484375
],
[
0.49267578125,
0.89794921875,
0.5087890625,
0.89794921875,
0.5087890625,
0.91357421875,
0.49267578125,
0.91357421875
]
] |
[
"qasilinear广告文献中的一些线程也是相关的。一项工作研究了使用ROI作为启发式方法在不",
"同拍卖中设置出价的效果[11，20,26，29，8]。另一种工作是使用行为经济学的思想来捕提观察到",
"的行为[25]。最后，一个小线程研究了GSP以及将其推广到新环境的问题。拟线性世界中GSP",
"的规范分析是[17，27]。试图推广GSP并考虑其在标准模型之外的性能，包括[13，19，1，3]",
"2Model and Preliminaries",
"一组n广告将根据配置（结果）oEO放置在页面上。对于每个配置，用户点击广告i（i的分配）",
"的可能性由&（o）给出；当用户点击时，广告商i支付价格pi",
"每个广告客户每次点击都有一个私有值U和一个公共5完整且可传递的首选项≤i，而不是结",
"果-价格对(o,p)",
"我们将在整篇论文中引用一些不同的偏好模型。首先，我们提到标准的拟线性偏好：",
"Definition 2. Preferences are quasi-linear if a bidder prefers to marimize the utility function",
"ui = (vi -pi)ri, i.e.",
"(o, pi) ≥i (o, p)) 台 (vi - pi)c;(o) ≥ (vi - p')c;(o) :",
"本文的大部分内容将侧重于各种形式的价值最大化偏好。为任意约束C定义了一般的值最大化",
"首选项",
"Definition 3. Preferences are value-maximzing subject to constraints C if they always prefer ar",
"outcome with higher value as long as none of its constraints are violated:",
"ici(o) >vici(o) ,(o,pi) satisfies C→(o,pi) >i (o,pi),",
"and when the value (allocation) is the same, a lower price is preferred:",
"ci(o) = ci(o) , Pi<pi, (o, pi) satisfies C → (o,pi) i (o',p) :",
"Moreover, any outcome is preferred to the one violating C :",
"(o,p,) violates C = (o, pi) ≥i (o, p,) :",
"我们将主要使用两种特定类型的单调约束：",
"Definition 4. A bidder has simple value-maximizing preferences if the bidder is value-marimizing",
"subject to the constraint that (o, pi) must satisfy pi ≤ vi .",
"Definition 5. A bidder has value-maximizing preferences with an ROI constraint if the bidder",
"is value-marimizing with the constraint \"ppi ≥ .",
"",
"AuctionsandTruthfulness拍卖要求投标人报告其价值参数，并为投标人提交的投标b指",
"定结果f(b）和付款p;(b）。函数f被称为socialchoicefunction，它意味着分配ci(b）",
"如果报告投标人的真实私人信息（例如。g。出价b，=）是一种主导策略：",
"5We assume this is public for simplicity, but it is not necessary for the results in our paper.",
"4"
] |
{
"arxiv_id": "1701.05946",
"language": "zh",
"page_idx": 3
}
|
||
1701.05946-zh-4
|
1701.05946-zh-4.jpg
|
$\textbf{Definition 6. }A\textit{mechanism M}= ( f, \boldsymbol{p}) \textit{is dominant strategy incentive compatible }( DSIC) \textit{if}$
$and\textit{ only if}$
$$\forall v_i,b,:\quad(f(b_{-i},v_i),p_i(b_{-i},v_i))\succeq_i(f(b_{-i},b_i),p_i(b_{-i},b_i))$$
在本文中，当真实的出价会导致平局时，我们将在连续出价空间中遇到麻烦；为了解决这个问
题，我们引入了 DSIC 的轻微弱化，它允许对一组“从未发生”的出价进行任意行为：
$\textbf{Definition 7. }A\textit{ mechanism M}= ( x, \boldsymbol{p}) \textit{ is dominant strategy incentive compatible almost every}$
where$(DSIC$-$AE)if,foranyiandbidsb_{-i},wehave$
$$(f(b_{-i},v_i),p_i(b_{-i},v_i))\succeq_i(f(b_{-i},b_i),p_i(b_{-i},b_i))$$
$for~all~value~functions~v_{i}~and~deviations~b_{i}~,~except~a~set~of~v_{i}~with~measure~zero.$
我们松散地使用真实来描述 DSIC 或 DSIC-AE
$\textbf{Remark 1. }Adding\textit{ a minimum bid increment would eliminate the need for DSIC- AE ( the next- 1) }$
highest bid $is$ always $well- defined)$ $at$ $a$ high cost $in$ $readability.$
Advertising Slot Auctions and GSP 我们将与用于研究赞助搜索拍卖的标准可分离点击率(CTR)框架一起使用。在可分离模型中$,n$广告竞争$m$插槽。存在插槽$(\alpha_1,\ldots,\alpha_m)$和$(\beta_1,\ldots,\beta_n)$ ,因此，当在插槽 $j$ 中显示 ad $i$ 时，用户有可能单击它：
Pr[click on ad $i$ when shown in slot $j]=\alpha_j\beta_i$,
标准 GSP 拍卖最初是在此设置中定义的：
$\textbf{Definition 8. The generalized second price }auction ( GSP) proceeds as follows:$
$1. \textit{ Bidders are sorted by }\beta _{i}b_{i}$ $; \textit{the bidder ranked j gets slot j }.$
$2. A\textit{ bidder is charged a price- per- click equal to the minimum bid required to keep the same slot}$
3 GSP is Truthful
3.1 Value Maximizers and GSP
我们之前认为，广告商的目标是通过受 ROI 约束的价值最大化来很好地建模的。我们还暗示这等
同于我们现在正式化的简单价值最大化模型：
$\textbf{Lemma 1. If a deterministic auction is truthful for simple value maximizers, it will be a dominant}$ $strategyforavaluemaximizerwithvaluev_{i}andROIconstraint\gamma toreportwhatititiswillingto$ $pay$, $v_i^{\prime }=$ $\frac {v_i}{1+ \gamma }$ .
备注我们不再期望投标人报告$v_i$ , 而是期望他们报告他们愿意支付的金额。如果我们想要一个真正的直接启示机制，我们可以要求投标人报告$\gamma$,但在实践中，报告一个人愿意支付的东西比真正值得的东西更自然
有了这个引理，它仍然表明 GSP 对于简单的价值最大化是真实的。令人惊讶的是，这几乎是微不足道的，Aggarwal 等人观察到了这一点。一般偏好模型中的 $[4]\text{。}^6$我们从直觉的第一原理中绘制一个证明：
Theorem 1 (Application of Aggarwal et al. [4]). GSP is truthful $^{7}\textit{for value maximizers with ROI}$
$\begin{array}{l}{{^{6}\text{The proof via }[4]\text{ observes that value max}}}\\{{\mathrm{d}\text{OI constraints can be mapped to their maximum price}}}\\{{\mathrm{d}\text{d}\text{an modol}}}\end{array}$
bidder model.
$^{7}$Effectively DSIC-AE (see Definition 7).
5
|
[
[
0.1168212890625,
0.09344482421875,
0.8837890625,
0.09344482421875,
0.8837890625,
0.10797119140625,
0.1168212890625,
0.10797119140625
],
[
0.1192626953125,
0.111083984375,
0.2059326171875,
0.111083984375,
0.2059326171875,
0.1256103515625,
0.1192626953125,
0.1256103515625
],
[
0.279296875,
0.126953125,
0.71728515625,
0.1275634765625,
0.71728515625,
0.1427001953125,
0.279296875,
0.141357421875
],
[
0.143798828125,
0.149658203125,
0.8798828125,
0.15087890625,
0.8798828125,
0.1654052734375,
0.143798828125,
0.1641845703125
],
[
0.1168212890625,
0.1673583984375,
0.78369140625,
0.1673583984375,
0.78369140625,
0.1817626953125,
0.1168212890625,
0.1817626953125
],
[
0.11358642578125,
0.1875,
0.88134765625,
0.1881103515625,
0.88134765625,
0.205810546875,
0.11358642578125,
0.2052001953125
],
[
0.1168212890625,
0.20703125,
0.5458984375,
0.20703125,
0.5458984375,
0.2215576171875,
0.1168212890625,
0.2215576171875
],
[
0.3203125,
0.2310791015625,
0.6787109375,
0.2310791015625,
0.6787109375,
0.24560546875,
0.3203125,
0.24560546875
],
[
0.11517333984375,
0.25390625,
0.74609375,
0.255126953125,
0.74609375,
0.26953125,
0.11517333984375,
0.268310546875
],
[
0.1446533203125,
0.276611328125,
0.51123046875,
0.27587890625,
0.51123046875,
0.290283203125,
0.1446533203125,
0.291748046875
],
[
0.1168212890625,
0.298583984375,
0.8798828125,
0.298583984375,
0.8798828125,
0.313232421875,
0.1168212890625,
0.313232421875
],
[
0.11517333984375,
0.31494140625,
0.60546875,
0.31640625,
0.60546875,
0.330810546875,
0.11517333984375,
0.330078125
],
[
0.125,
0.35107421875,
0.88134765625,
0.35107421875,
0.88134765625,
0.365478515625,
0.125,
0.365478515625
],
[
0.11273193359375,
0.365478515625,
0.8798828125,
0.367431640625,
0.8798828125,
0.385009765625,
0.11273193359375,
0.38330078125
],
[
0.11767578125,
0.386474609375,
0.56201171875,
0.386474609375,
0.56201171875,
0.399658203125,
0.11767578125,
0.399658203125
],
[
0.31787109375,
0.410400390625,
0.6806640625,
0.410400390625,
0.6806640625,
0.4248046875,
0.31787109375,
0.4248046875
],
[
0.146240234375,
0.43310546875,
0.468017578125,
0.43310546875,
0.468017578125,
0.447509765625,
0.146240234375,
0.447509765625
],
[
0.11517333984375,
0.453857421875,
0.7490234375,
0.455810546875,
0.7490234375,
0.47021484375,
0.11517333984375,
0.468505859375
],
[
0.134033203125,
0.474853515625,
0.6259765625,
0.4765625,
0.6259765625,
0.494384765625,
0.134033203125,
0.492431640625
],
[
0.1380615234375,
0.50390625,
0.8798828125,
0.50390625,
0.8798828125,
0.51806640625,
0.1380615234375,
0.51806640625
],
[
0.11358642578125,
0.54345703125,
0.1641845703125,
0.54345703125,
0.1641845703125,
0.5595703125,
0.11358642578125,
0.5595703125
],
[
0.160888671875,
0.54345703125,
0.34716796875,
0.54345703125,
0.34716796875,
0.55810546875,
0.160888671875,
0.55810546875
],
[
0.11517333984375,
0.5771484375,
0.435546875,
0.5771484375,
0.435546875,
0.59130859375,
0.11517333984375,
0.59130859375
],
[
0.11767578125,
0.60302734375,
0.88134765625,
0.60302734375,
0.88134765625,
0.6171875,
0.11767578125,
0.6171875
],
[
0.1192626953125,
0.62060546875,
0.465576171875,
0.62060546875,
0.465576171875,
0.6337890625,
0.1192626953125,
0.6337890625
],
[
0.1168212890625,
0.64306640625,
0.88232421875,
0.64306640625,
0.88232421875,
0.6572265625,
0.1168212890625,
0.6572265625
],
[
0.11767578125,
0.66015625,
0.88232421875,
0.66015625,
0.88232421875,
0.6748046875,
0.11767578125,
0.6748046875
],
[
0.11517333984375,
0.6748046875,
0.2344970703125,
0.67822265625,
0.2344970703125,
0.69580078125,
0.11517333984375,
0.6923828125
],
[
0.1478271484375,
0.70068359375,
0.88134765625,
0.70068359375,
0.88134765625,
0.71533203125,
0.1478271484375,
0.71533203125
],
[
0.11767578125,
0.71826171875,
0.88134765625,
0.71826171875,
0.88134765625,
0.73291015625,
0.11767578125,
0.73291015625
],
[
0.1168212890625,
0.73486328125,
0.294189453125,
0.734375,
0.294189453125,
0.74853515625,
0.1168212890625,
0.75
],
[
0.146240234375,
0.75048828125,
0.8798828125,
0.75048828125,
0.8798828125,
0.76513671875,
0.146240234375,
0.76513671875
],
[
0.11767578125,
0.76806640625,
0.87841796875,
0.76806640625,
0.87841796875,
0.78271484375,
0.11767578125,
0.78271484375
],
[
0.11767578125,
0.7861328125,
0.236083984375,
0.7861328125,
0.236083984375,
0.80029296875,
0.11767578125,
0.80029296875
],
[
0.1168212890625,
0.8095703125,
0.88232421875,
0.8095703125,
0.88232421875,
0.82373046875,
0.1168212890625,
0.82373046875
],
[
0.1356201171875,
0.83349609375,
0.88134765625,
0.83349609375,
0.88134765625,
0.84765625,
0.1356201171875,
0.84765625
],
[
0.11767578125,
0.84765625,
0.205078125,
0.84765625,
0.205078125,
0.859375,
0.11767578125,
0.859375
],
[
0.136474609375,
0.861328125,
0.414306640625,
0.861328125,
0.414306640625,
0.875,
0.136474609375,
0.875
],
[
0.491943359375,
0.89794921875,
0.50732421875,
0.89794921875,
0.50732421875,
0.91357421875,
0.491943359375,
0.91357421875
]
] |
[
"Definition 6. A mechanism M = (f,p) is dominant strategy incentive compatible (DsIC) if",
"and onlyif",
"Voi,b,:(f(b-i, v),pi(b-i, vi)) ≥i (f(b-i,bi),pi(b_z,b))",
"在本文中，当真实的出价会导致平局时，我们将在连续出价空间中遇到麻烦；为了解决这个问",
"题，我们引入了DSIC的轻微弱化，它允许对一组“从未发生”的出价进行任意行为",
"Definition 7. A mechanism M = (c, p) is dominant strategy incentive compatible almost every-",
"where (DSIC-AE) if, for any i and bids b_i,we have",
"(f(b_i, vi), Pi(b_i, Vi)) ≥i (f(b_i, bi), pi(b_z, bi))",
"for all alue functions vi and deviations bi , ercept a set of vi with measure zero",
"我们松散地使用真实来描述DSIC或DSIC-AE",
"Remark 1. Adding a minimum bid increment would eliminate the need for DsIC-AE (the nert-",
"highest bid is always well-defined) at a high cost in readability.",
"AdvertisingSlotAuctionsandGSP我们将与用于研究赞助搜索拍卖的标准可分离点击率",
"（CTR）框架一起使用。在可分离模型中，n广告竞争m插槽。存在插槽（α1，...,αm）和（β1，...,βm）",
"，因此，当在插槽j中显示adi时，用户有可能单击它：",
"Pr[click on ad i when shown in slot j J = ajβi,",
"标准GSP拍卖最初是在此设置中定义的：",
"Definition 8. The generalized second price auction (GsP) proceeds as follows",
"1. Bidders are sorted by Bibi ; the bidder ranked j gets slot j .",
"2. A bidder is charged a price-per-click equal to the minimum bid required to keep the same slot",
"3",
"GSP is Truthful",
"3.1 Value Maximizers and GSP",
"我们之前认为，广告商的目标是通过受ROI约束的价值最大化来很好地建模的。我们还暗示这等",
"同于我们现在正式化的简单价值最大化模型：",
"Lemma 1. If a deterministic auction is truthful for simple value marimizers, it will be a dominant",
"strategy for a value marimizer with value vi and ROI constraint to report what it is willing to",
"pay, U, = 1 :",
"备注我们不再期望投标人报告Ui，而是期望他们报告他们愿意支付的金额。如果我们想要一个",
"真正的直接启示机制，我们可以要求投标人报告，但在实践中，报告一个人愿意支付的东西比",
"真正值得的东西更自然",
"有了这个引理，它仍然表明GSP对于简单的价值最大化是真实的。令人惊的是，这几乎是微",
"不足道的，Aggarwal等人观察到了这一点。一般偏好模型中的[4]。6我们从直觉的第一原理中",
"绘制一个证明：",
"Theorem 1 (Application of Aggarwal et al. [4]). GsP is truthful for alue marimizers with ROl",
"The proof via [4] observes that value maximizers with ROI constraints can be mapped to their maximum price",
"bidder model",
"7Effectively DSIC-AE (see Definition 7).",
"5"
] |
{
"arxiv_id": "1701.05946",
"language": "zh",
"page_idx": 4
}
|
||
2305.08703-zh-1
|
2305.08703-zh-1.jpg
|
## 模式适应性知识图谱构建
$\mathbf{Hongbin~Ye}^{1,2,4},\mathbf{Honghao~Gui}^{1,4},\mathbf{Xin~Xu}^{1,4},\mathbf{Xi~Chen}^3,\mathbf{Huajun~Chen}^{1,4},\mathbf{Ningyu~Zhang}^{1,4^*}_{11}\text{Zhejiang University }^2\text{Zhejiang Lab }^3$Platform and Content Group, Tencent
$^{4}$ Zhejiang University- Ant Group Joint Laboratory of Knowledge Graph
yehongbin@zhejianglab.com, jasonxchen@tencent.com,
$\{$ guihonghao, xxucs, huajunsir, zhangningyu $\}$ @zju.edu.cn
## $\mathbf{Abstract}$
Conventional Knowledge Graph Construction (KGC) approaches typically follow the static information extraction paradigm with a closed set of pre-defined schema. As a result, such approaches fall short when applied to dynamic scenarios or domains, whereas a new type of knowledge emerges. This necessitates a system that can handle evolving schema automatically to extract information for KGC. To address this need, we propose a new task called schema-adaptable KGC, which aims to continually extract entity, relation, and event based on a dynamically changing schema graph without re-training. We first split and convert existing datasets based on three principles to build a benchmark, i.e., horizontal schema expansion, vertical schema expansion, and hybrid schema expansion; then investigate the schema-adaptable performance of several wellknown approaches such as Text2Event, TANL, UIE and GPT-3.5. We further propose a simple yet effective baseline dubbed ADAKGC ,which contains schema-enriched prefix instructor and schema-conditioned dynamic decoding to better handle evolving schema. Comprehensive experimental results illustrate that ADAKGC can outperform baselines but still have room for improvement. We hope the proposed work can deliver benefits to the com- ${\hat{\text{munity}}^{1}}.$
### 1 $\textbf{Introduction}$
知识图谱构建(KGC)通常通过信息提取获得了广泛的经验成功，并且可以为各种 NLP 任务提供后端支持，例如问答(Saxena et al.,2020; Shang et al., 2022; Zhang et al., 2022a) , 常识推理(Yasunaga et al., 2021; Zhang et al., 2022d) 等。传统的 KGC 任务，包括命名实体识别 (NER) (Liu et al.,2021; Wang et al.,2021), 关系提取(RE) (Chen et al., 2022a; Zheng et al., 2021; Yao
${^{*}\text{Corresp}}$onding author.
$^{1}$Code and datasets available at https://github.com/
zjunlp/AdaKGC .
<FigureHere>
Figure 1: Knowledge Graph Construction (KGC) with
dynamic updates of schema .
et al., 2022) 和事件提取( EE)(Huang et al., 2018; Liu et al., 2020; Lu et al., 2021; Lou et al., 2023) 是“反应性的”,依赖于来自最终用户的静态预定义模式。但是，如图 1 所示，模式可能会随着场景适应而发展，这使得以前的模型在不重新训练的情况下难以使用
请注意，现有的信息提取系统只能通过预定义的模式并在固定框架上执行一次性培训来处理固定数量的类。最好对变化做出反应(例如。g。, 不断发展的模式) 到现有的 KG, 使系统的行为更加“主动”,就像人类可以处理灵活的知识更新一样。早期，几种方法引入了增量学习(Cao et al., 2020; Wang et al., 2019; Shen et al., 2020; Cui et al., 2021b) 来不断学习新类。在这种情况下，提取系统从类增量数据流中学习，但在适应新类时通常会在旧类上遭受明显的性能下降。换句话说，以前的研究强调与灾难性遗忘 (Thrun, 1998) 作斗争。但是，对于模式演化的场景，提取模型的动态泛化性起着至关重要的作用，需要从本体演化的角度进行检查
因此，我们提出了一种新颖的 KGC 任务，称为模式适应性 KGC, 其中要求模型具有表示和适应补充知识提取的能力。我们首先根据演化模式方向的三个原则(水平模式扩展，垂直模式扩展和混合模式扩展)在 NER, RE $^2$和
${^{2}}$We regard RE as relational triple extraction in this paper.
|
[
[
0.366943359375,
0.0926513671875,
0.63232421875,
0.0926513671875,
0.63232421875,
0.10986328125,
0.366943359375,
0.10986328125
],
[
0.1275634765625,
0.1318359375,
0.87890625,
0.1300048828125,
0.87890625,
0.1478271484375,
0.1275634765625,
0.149658203125
],
[
0.19482421875,
0.15087890625,
0.81201171875,
0.15087890625,
0.81201171875,
0.1650390625,
0.19482421875,
0.1650390625
],
[
0.204833984375,
0.163330078125,
0.79833984375,
0.1650390625,
0.79833984375,
0.182861328125,
0.204833984375,
0.18115234375
],
[
0.264404296875,
0.18408203125,
0.73779296875,
0.18408203125,
0.73779296875,
0.19775390625,
0.264404296875,
0.19775390625
],
[
0.235107421875,
0.20068359375,
0.779296875,
0.1995849609375,
0.779296875,
0.21435546875,
0.235107421875,
0.215576171875
],
[
0.2626953125,
0.2529296875,
0.343505859375,
0.2529296875,
0.343505859375,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.68115234375,
0.265380859375,
0.6943359375,
0.265380859375,
0.6943359375,
0.27490234375,
0.68115234375,
0.27490234375
],
[
0.69189453125,
0.26123046875,
0.71875,
0.26123046875,
0.71875,
0.2724609375,
0.69189453125,
0.2724609375
],
[
0.8037109375,
0.2666015625,
0.8134765625,
0.2666015625,
0.8134765625,
0.26953125,
0.8037109375,
0.26953125
],
[
0.75146484375,
0.26953125,
0.771484375,
0.26953125,
0.771484375,
0.281982421875,
0.75146484375,
0.281982421875
],
[
0.5517578125,
0.2666015625,
0.603515625,
0.2666015625,
0.603515625,
0.287353515625,
0.5517578125,
0.287353515625
],
[
0.79833984375,
0.270751953125,
0.8251953125,
0.27978515625,
0.81689453125,
0.29150390625,
0.7900390625,
0.281982421875
],
[
0.8037109375,
0.26953125,
0.8134765625,
0.26953125,
0.8134765625,
0.273681640625,
0.8037109375,
0.273681640625
],
[
0.67529296875,
0.271484375,
0.72119140625,
0.279052734375,
0.71728515625,
0.29150390625,
0.6708984375,
0.283935546875
],
[
0.63623046875,
0.271484375,
0.6484375,
0.271484375,
0.6484375,
0.283203125,
0.63623046875,
0.283203125
],
[
0.7421875,
0.2724609375,
0.755859375,
0.2724609375,
0.755859375,
0.281982421875,
0.7421875,
0.281982421875
],
[
0.82958984375,
0.2724609375,
0.83642578125,
0.276611328125,
0.8310546875,
0.27978515625,
0.8251953125,
0.275634765625
],
[
0.145263671875,
0.28076171875,
0.459228515625,
0.28076171875,
0.459228515625,
0.29443359375,
0.145263671875,
0.29443359375
],
[
0.61865234375,
0.273681640625,
0.6337890625,
0.273681640625,
0.6337890625,
0.283203125,
0.61865234375,
0.283203125
],
[
0.8154296875,
0.28076171875,
0.83447265625,
0.27783203125,
0.8369140625,
0.287353515625,
0.81787109375,
0.290283203125
],
[
0.57275390625,
0.290283203125,
0.59765625,
0.2880859375,
0.599609375,
0.29638671875,
0.57421875,
0.298583984375
],
[
0.69287109375,
0.290283203125,
0.7197265625,
0.290283203125,
0.7197265625,
0.297607421875,
0.69287109375,
0.297607421875
],
[
0.79833984375,
0.2880859375,
0.84228515625,
0.2880859375,
0.84228515625,
0.298583984375,
0.79833984375,
0.298583984375
],
[
0.145263671875,
0.295654296875,
0.45751953125,
0.295654296875,
0.45751953125,
0.308837890625,
0.145263671875,
0.308837890625
],
[
0.67919921875,
0.29150390625,
0.69189453125,
0.29150390625,
0.69189453125,
0.295654296875,
0.67919921875,
0.295654296875
],
[
0.145263671875,
0.308837890625,
0.459228515625,
0.308837890625,
0.459228515625,
0.322509765625,
0.145263671875,
0.322509765625
],
[
0.72705078125,
0.30712890625,
0.86962890625,
0.30712890625,
0.86962890625,
0.317626953125,
0.72705078125,
0.317626953125
],
[
0.145263671875,
0.32373046875,
0.459228515625,
0.32373046875,
0.459228515625,
0.33740234375,
0.145263671875,
0.33740234375
],
[
0.68408203125,
0.319580078125,
0.69775390625,
0.319580078125,
0.69775390625,
0.328857421875,
0.68408203125,
0.328857421875
],
[
0.54150390625,
0.32666015625,
0.62744140625,
0.32666015625,
0.62744140625,
0.33837890625,
0.54150390625,
0.33837890625
],
[
0.71728515625,
0.32470703125,
0.80419921875,
0.32470703125,
0.80419921875,
0.3349609375,
0.71728515625,
0.3349609375
],
[
0.145263671875,
0.33740234375,
0.456787109375,
0.33740234375,
0.456787109375,
0.350830078125,
0.145263671875,
0.350830078125
],
[
0.5439453125,
0.339111328125,
0.623046875,
0.339111328125,
0.623046875,
0.3466796875,
0.5439453125,
0.3466796875
],
[
0.145263671875,
0.352783203125,
0.459228515625,
0.352783203125,
0.459228515625,
0.366455078125,
0.145263671875,
0.366455078125
],
[
0.57421875,
0.35693359375,
0.5859375,
0.35693359375,
0.5859375,
0.366455078125,
0.57421875,
0.366455078125
],
[
0.145263671875,
0.366455078125,
0.459228515625,
0.366455078125,
0.459228515625,
0.380126953125,
0.145263671875,
0.380126953125
],
[
0.145263671875,
0.38134765625,
0.459228515625,
0.38134765625,
0.459228515625,
0.394287109375,
0.145263671875,
0.394287109375
],
[
0.59033203125,
0.38134765625,
0.60205078125,
0.38720703125,
0.59423828125,
0.393798828125,
0.58447265625,
0.38720703125
],
[
0.796875,
0.3818359375,
0.8037109375,
0.3818359375,
0.8037109375,
0.38720703125,
0.796875,
0.38720703125
],
[
0.14697265625,
0.3955078125,
0.45751953125,
0.3955078125,
0.45751953125,
0.40625,
0.14697265625,
0.40625
],
[
0.145263671875,
0.4091796875,
0.459228515625,
0.4091796875,
0.459228515625,
0.4228515625,
0.145263671875,
0.4228515625
],
[
0.145263671875,
0.4228515625,
0.4609375,
0.4228515625,
0.4609375,
0.435791015625,
0.145263671875,
0.435791015625
],
[
0.51318359375,
0.418701171875,
0.8818359375,
0.418701171875,
0.8818359375,
0.431640625,
0.51318359375,
0.431640625
],
[
0.145263671875,
0.43701171875,
0.459228515625,
0.43701171875,
0.459228515625,
0.45068359375,
0.145263671875,
0.45068359375
],
[
0.51318359375,
0.432861328125,
0.7060546875,
0.432861328125,
0.7060546875,
0.446533203125,
0.51318359375,
0.446533203125
],
[
0.1435546875,
0.451904296875,
0.459228515625,
0.451904296875,
0.459228515625,
0.465576171875,
0.1435546875,
0.465576171875
],
[
0.1435546875,
0.465576171875,
0.459228515625,
0.465576171875,
0.459228515625,
0.478515625,
0.1435546875,
0.478515625
],
[
0.145263671875,
0.479736328125,
0.459228515625,
0.479736328125,
0.459228515625,
0.493408203125,
0.145263671875,
0.493408203125
],
[
0.51318359375,
0.474365234375,
0.8818359375,
0.474365234375,
0.8818359375,
0.488037109375,
0.51318359375,
0.488037109375
],
[
0.14697265625,
0.495361328125,
0.45751953125,
0.495361328125,
0.45751953125,
0.505859375,
0.14697265625,
0.505859375
],
[
0.51318359375,
0.4921875,
0.87890625,
0.4921875,
0.87890625,
0.5029296875,
0.51318359375,
0.5029296875
],
[
0.14697265625,
0.51025390625,
0.45751953125,
0.51025390625,
0.45751953125,
0.52099609375,
0.14697265625,
0.52099609375
],
[
0.5146484375,
0.50732421875,
0.880859375,
0.50732421875,
0.880859375,
0.52099609375,
0.5146484375,
0.52099609375
],
[
0.145263671875,
0.5224609375,
0.459228515625,
0.5224609375,
0.459228515625,
0.5361328125,
0.145263671875,
0.5361328125
],
[
0.51318359375,
0.5224609375,
0.880859375,
0.5224609375,
0.880859375,
0.53759765625,
0.51318359375,
0.53759765625
],
[
0.145263671875,
0.53759765625,
0.459228515625,
0.53759765625,
0.459228515625,
0.55029296875,
0.145263671875,
0.55029296875
],
[
0.51318359375,
0.5390625,
0.880859375,
0.5390625,
0.880859375,
0.552734375,
0.51318359375,
0.552734375
],
[
0.145263671875,
0.55029296875,
0.45751953125,
0.55029296875,
0.45751953125,
0.56396484375,
0.145263671875,
0.56396484375
],
[
0.5146484375,
0.5556640625,
0.73486328125,
0.5556640625,
0.73486328125,
0.5693359375,
0.5146484375,
0.5693359375
],
[
0.1435546875,
0.5634765625,
0.459228515625,
0.56396484375,
0.459228515625,
0.5791015625,
0.1435546875,
0.57763671875
],
[
0.53076171875,
0.572265625,
0.880859375,
0.572265625,
0.880859375,
0.5859375,
0.53076171875,
0.5859375
],
[
0.1435546875,
0.5791015625,
0.456787109375,
0.57763671875,
0.456787109375,
0.59130859375,
0.1435546875,
0.5927734375
],
[
0.5146484375,
0.58740234375,
0.8818359375,
0.58837890625,
0.8818359375,
0.6025390625,
0.5146484375,
0.6015625
],
[
0.14697265625,
0.59326171875,
0.459228515625,
0.59326171875,
0.459228515625,
0.60693359375,
0.14697265625,
0.60693359375
],
[
0.14697265625,
0.609375,
0.45751953125,
0.609375,
0.45751953125,
0.619140625,
0.14697265625,
0.619140625
],
[
0.5146484375,
0.60498046875,
0.880859375,
0.60498046875,
0.880859375,
0.6181640625,
0.5146484375,
0.6181640625
],
[
0.1435546875,
0.62158203125,
0.45751953125,
0.62158203125,
0.45751953125,
0.63525390625,
0.1435546875,
0.63525390625
],
[
0.51318359375,
0.619140625,
0.87744140625,
0.619140625,
0.87744140625,
0.6328125,
0.51318359375,
0.6328125
],
[
0.145263671875,
0.6357421875,
0.45751953125,
0.6357421875,
0.45751953125,
0.6494140625,
0.145263671875,
0.6494140625
],
[
0.51123046875,
0.63525390625,
0.87744140625,
0.63525390625,
0.87744140625,
0.6494140625,
0.51123046875,
0.6494140625
],
[
0.145263671875,
0.65087890625,
0.459228515625,
0.65087890625,
0.459228515625,
0.66455078125,
0.145263671875,
0.66455078125
],
[
0.51318359375,
0.65283203125,
0.880859375,
0.65283203125,
0.880859375,
0.66650390625,
0.51318359375,
0.66650390625
],
[
0.145263671875,
0.66455078125,
0.45751953125,
0.66455078125,
0.45751953125,
0.677734375,
0.145263671875,
0.677734375
],
[
0.51318359375,
0.66748046875,
0.880859375,
0.66748046875,
0.880859375,
0.68115234375,
0.51318359375,
0.68115234375
],
[
0.1435546875,
0.6787109375,
0.459228515625,
0.677734375,
0.459228515625,
0.6923828125,
0.1435546875,
0.693359375
],
[
0.51123046875,
0.68310546875,
0.87744140625,
0.68408203125,
0.87744140625,
0.69873046875,
0.51123046875,
0.69775390625
],
[
0.145263671875,
0.69482421875,
0.210693359375,
0.69482421875,
0.210693359375,
0.7060546875,
0.145263671875,
0.7060546875
],
[
0.51123046875,
0.69970703125,
0.87890625,
0.69970703125,
0.87890625,
0.71337890625,
0.51123046875,
0.71337890625
],
[
0.151123046875,
0.72021484375,
0.2626953125,
0.72021484375,
0.2626953125,
0.732421875,
0.151123046875,
0.732421875
],
[
0.51318359375,
0.71630859375,
0.87890625,
0.71630859375,
0.87890625,
0.72998046875,
0.51318359375,
0.72998046875
],
[
0.118408203125,
0.72265625,
0.1318359375,
0.72265625,
0.1318359375,
0.73095703125,
0.118408203125,
0.73095703125
],
[
0.51123046875,
0.732421875,
0.880859375,
0.73291015625,
0.880859375,
0.74755859375,
0.51123046875,
0.74658203125
],
[
0.1175537109375,
0.74365234375,
0.486083984375,
0.7451171875,
0.486083984375,
0.76025390625,
0.1175537109375,
0.75830078125
],
[
0.51318359375,
0.7490234375,
0.87890625,
0.7490234375,
0.87890625,
0.7626953125,
0.51318359375,
0.7626953125
],
[
0.12176513671875,
0.76025390625,
0.484375,
0.76025390625,
0.484375,
0.77490234375,
0.12176513671875,
0.77490234375
],
[
0.51123046875,
0.76416015625,
0.87744140625,
0.76416015625,
0.87744140625,
0.77783203125,
0.51123046875,
0.77783203125
],
[
0.118408203125,
0.77783203125,
0.486083984375,
0.77783203125,
0.486083984375,
0.79150390625,
0.118408203125,
0.79150390625
],
[
0.51318359375,
0.78076171875,
0.87890625,
0.78076171875,
0.87890625,
0.79443359375,
0.51318359375,
0.79443359375
],
[
0.1158447265625,
0.79345703125,
0.48779296875,
0.7919921875,
0.48779296875,
0.80712890625,
0.1158447265625,
0.80810546875
],
[
0.51318359375,
0.79736328125,
0.5341796875,
0.79736328125,
0.5341796875,
0.81005859375,
0.51318359375,
0.81005859375
],
[
0.1158447265625,
0.81005859375,
0.486083984375,
0.81005859375,
0.486083984375,
0.82373046875,
0.1158447265625,
0.82373046875
],
[
0.52880859375,
0.8134765625,
0.8818359375,
0.81396484375,
0.8818359375,
0.8291015625,
0.52880859375,
0.82763671875
],
[
0.118408203125,
0.826171875,
0.48779296875,
0.826171875,
0.48779296875,
0.84033203125,
0.118408203125,
0.84033203125
],
[
0.5146484375,
0.83056640625,
0.880859375,
0.83056640625,
0.880859375,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.1175537109375,
0.8427734375,
0.486083984375,
0.8427734375,
0.486083984375,
0.85546875,
0.1175537109375,
0.85546875
],
[
0.51123046875,
0.845703125,
0.880859375,
0.8466796875,
0.880859375,
0.86083984375,
0.51123046875,
0.85986328125
],
[
0.1158447265625,
0.85791015625,
0.486083984375,
0.85791015625,
0.486083984375,
0.87158203125,
0.1158447265625,
0.87158203125
],
[
0.5146484375,
0.86328125,
0.880859375,
0.86328125,
0.880859375,
0.876953125,
0.5146484375,
0.876953125
],
[
0.15283203125,
0.8798828125,
0.30322265625,
0.8798828125,
0.30322265625,
0.89453125,
0.15283203125,
0.89453125
],
[
0.5146484375,
0.87744140625,
0.8818359375,
0.87744140625,
0.8818359375,
0.892578125,
0.5146484375,
0.892578125
],
[
0.1427001953125,
0.89111328125,
0.48779296875,
0.892578125,
0.48779296875,
0.90673828125,
0.1427001953125,
0.90625
],
[
0.1175537109375,
0.90673828125,
0.2225341796875,
0.90673828125,
0.2225341796875,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.5380859375,
0.90380859375,
0.8818359375,
0.90478515625,
0.8818359375,
0.91943359375,
0.5380859375,
0.91845703125
]
] |
[
"模式适应性知识图谱构建",
"Hongbin Yel,2.4, Honghao Guil-4, Xin Xul-4, Xi Chen, Huajun Chen'-4, Ningyu Zhangl.4",
"1 Zhejiang University 2 Zhejiang Lab 3 Platform and Content Group, Tencent",
"4 Zhejiang University - Ant Group Joint Laboratory of Knowledge Graph",
"yehongbin@zhejianglab.com,jasonxchen@tencent.com,",
" guihonghao,xxucs,huajunsir,zhangningyu } @zju.edu.cn",
"Abstract",
"a",
"1o",
"",
"",
"",
"",
"人",
"",
"A",
"一",
"",
"Conventional Knowledge Graph Construction",
"M",
"8",
"ance",
"tance",
"Instance3",
"(KGC) approaches typically follow the static",
"Ins",
"information extraction paradigm with a closed",
"InstanceKnowledgeGraph",
"set of pre-defined schema . As a result, such",
"",
"Ontology Graph",
"Task Documents",
"approaches fall short when applied to dynamic",
"Schema Chang",
"scenarios or domains, whereas a new type of",
"O",
"knowledge emerges. This necessitates a sys-",
"tem that can handle evolving schema automat-",
"C",
"福",
"icallytoextractinformationforKGC.Toad-",
"dress this need, we propose a new task called",
"schema-adaptable KGC, which aims to contin-",
"Figure 1:Knowledge Graph Construction (KGC) with",
"ually extract entity, relation, and event based",
"dynamic updates of schema",
"on a dynamically changing schema graph with-",
"out re-training. We first split and convert ex-",
"isting datasets based on three principles to",
"etal.,2022）和事件提取（EE)（Huangetal.2018;",
"build a benchmark, i.e.,horizontal schema ex-",
"Liu et al.. 2020: Lu et al., 2021: Lou et al.. 2023)",
"pansion,vertical schema expansion, and hy",
"是“反应性的”，依赖于来自最终用户的静态",
"brid schema expansion; then investigate the",
"预定义模式。但是，如图1所示，模式可能会",
"schema-adaptable performance of several well-",
"随着场景适应而发展，这使得以前的模型在不",
"known approaches such as Text2Event, TANL,",
"重新训练的情况下难以使用",
"UIE and GPT-3.5. We further propose a sim-",
"请注意，现有的信息提取系统只能通过预定",
"ple yet effective baseline dubbed ADAKGC",
"义的模式并在固定框架上执行一次性培训来",
", which contains schema-enriched prefix in-",
"structorand schema-conditioneddynamicde-",
"处理固定数量的类。最好对变化做出反应（例",
"coding to better handle evolving schema.",
"如。g。，不断发展的模式)到现有的KG，使系",
"Comprehensive experimental results illustrate",
"统的行为更加“主动”，就像人类可以处理灵",
"that ADAKGC can outperform baselines but",
"活的知识更新一样。早期，几种方法引入了增",
"still have room for improvement. We hope the",
"量学习 (Cao et al., 2020: Wang et al.,2019; Shen",
"proposed work can deliver benefits to the com-",
"etal.,2020;Cuietal.，2021b）来不断学习新类，",
"munity",
"在这种情况下，提取系统从类增量数据流中学",
"Introduction",
"习，但在适应新类时通常会在旧类上遭受明显",
"1",
"的性能下降。换句话说，以前的研究强调与灾",
"知识图谱构建（KGC）通常通过信息提取获得",
"难性遗忘（Thrun，1998）作斗争。但是，对于模",
"了广泛的经验成功，并且可以为各种NLP任",
"式演化的场景，提取模型的动态泛化性起着至",
"务提供后端支持，例如问答（Saxenaetal.，2020;",
"关重要的作用，需要从本体演化的角度进行检",
"Shang etal.,2022;Zhang et al.,2022a)，常识推",
"查",
"理 (Yasunaga et al., 2021; Zhang et al., 2022d)等。",
"因此，我们提出了一种新颖的KGC任务，称",
"传统的KGC任务，包括命名实体识别（NER）",
"为模式适应性KGC，其中要求模型具有表示",
"(Liuetal.,2021;Wangetal.,2021)，关系提取",
"和适应补充知识提取的能力。我们首先根据演",
"(RE) (Chen et al., 2022a; Zheng et al., 2021; Yao",
"化模式方向的三个原则（水平模式扩展，垂直",
"*Corresponding author.",
"模式扩展和混合模式扩展）在NER，RE²和",
"'Code and datasets available at https://github.com/",
"zjunlp/AdaKGC",
"2We regard RE as relational triple extraction in this paper."
] |
{
"arxiv_id": "2305.08703",
"language": "zh",
"page_idx": 1
}
|
||
2305.08703-zh-2
|
2305.08703-zh-2.jpg
|
(Paolini et al., 2021)、UIE (Lu et al., 2022) 和GPT-3 的方法。5 (Ouyang et al.,2022) 无法有效提取给定复杂演化模式的信息。我们认为主要问题如下：1) 如何学习动态和可概括的模式表示作为提取条件；2)如何精确提取受新更新模式约束的新实例
为此，我们提出了一个简单的基线，称为ADAKGC ,它引人了富含模式的前缀指导者(SPI)来表示和传递可学习的特定于模式的知识。在每个迭代阶段，我们从当前模式图线性转换为可学习的提示，使用本体名称初始化并连接到特定于任务的前缀。为了鼓励解码器理解动态模式，我们使用了模式条件动态解码(SDD)策略，该策略构造了特定于模式的词汇表到输出空间的解码路径。当模式改变时， 我们动态地构造一个新的特里树来调整输出空间。请注意，ADAKGC 与模型无关，可以处理各种具有挑战性的模式演化方案。我们将这项工作的贡献总结为：
$\cdot$ We introduce a new task of the schemaadaptable KGC to meet the schema evolution requirements, which is a new branch that has not been well-explored to the best of our knowledge.
$\cdot$ We propose a new baseline ADAKGC , which includes schema-enriched prefix instructors and schema-conditioned dynamic decoding strategy, and experimentally demonstrate the adaptability.
$\cdot$ We release the schema-adaptable KGC benchmark, which imposes new challenges and presents new research opportunities for the NLP community.
2 $\textbf{Problem Statement and Overview}$
2.1 Background of KGC
KGC一直是(Lu et al., 2022; Zhang et al., 2022b) 的一项有前途的研究挑战，现有的基准测试利用定义良好的模式来指导知识图构建，重点是生成特定领域的知识图或聚合异构结构化数据库。例如，FEW-NERD (Ding et al.,2021)由粗粒度和细粒度实体类型定义组成，用于定位和分类来自非结构化自然语言的命名实体。NYT (Riedel et al., 2010a) 根据特定的分类结构专门从文本数据源中提取关系三重实例。ACE2005 (Ntroduction) 根据上下文识别触发器和事件类型，并且每个都有自己的事件参数，以填充槽
EE 的三个任务上构建数据集。通过实证分析，的方式描述。此外，TAC-KBP(Ellis et al., 2014) 我们注意到 Text2Event (Lu et al., 2021) 、TANL 旨在利用现有的通用域结构化数据源并扩展采
用描述性文本作为附加信息的实体链接。OAEI (Euzenat et al., 2011) 基于两个或多个现有本体或知识图之间的对齐创建集成本体。在本文中， 我们专注于从非结构化文本中提取知识实例的工作，这被视为模式约束预测(结构预测)任务
2.2 $\textbf{Definition of Schema- adaptable KGC}$
在现实世界中，KGC 系统从非结构化文本中提取结构化知识，并根据频繁调整的模式将其规范化为实例图。给定一组模式图$\mathcal{S}=\left\{s_1,s_2,...,s_n\right\}$,可适应模式的 KGC 的任务是为每次迭代生成一组模式约束实例$\mathcal{G}=\{g_1,g_2,...,g_n\}$。假设有一个在初始训练集上训练的模型$\mathcal{M}_\theta=LM(\mathcal{D}_{train}^1|S_1)$ ,之后更新模式的标记数据不可用。提供了一个模式自适应数据流$\left\{\mathcal{D}^{(1)},\mathcal{D}^{(2)},\ldots,\mathcal{D}^{(N)}\right\}$来评估模型对模式动态更新的适应性。每个$\mathcal{D}^(k)$都包含开发/测试数据 $\left(\mathcal{D}_\mathrm{dev}^k,\mathcal{D}_\mathrm{test}^k\right)$ 和模式图 $s_k$ 。请注意，该模型不会被重新训练，但希望能够利用不断发展的模式来提取信息。面临的挑战是，该模型有望在测试集$\mathcal{D}_\mathrm{test}^k$的每次迭代中表现良好，该测试集包含为更新的模式更改的黄金实例
## 2.3 Dataset Construction Process
如图$\operatorname{E}$2 所示，我们针对不同类型的模式演化设计了三个原则，并应用 Algorithm$\boxed{\mathrm{F}}1$构建了用于评估的数据集：(1)水平模式扩展要求模式添加相同级别的新类节点，可以将其视为一种形式的类增量学习实例，。基于对相邻新类的泛化效应，我们可以评估模式特征的传输能力。(2) 垂直模式扩展要求模式添加父类的子类。基于对子类的泛化效应，我们可以评估模式特征的继承和派生能力。(3)混合模式扩展要求模式在每次迭代时随机地水平或垂直扩展节点，这总结了模式图并代表了它们潜在的共同进化模式。更多细节在附录 A.1 中，除了上述结构扩展之外，我们还从语义的角度进一步探讨了类似的节点替换
2.4 Schema-adaptable KGC Benchmark
模式适应性 KGC 有两个挑战。首先，由于模式在每次迭代中都会更新，因此需要将模式演变信息动态注入模型中。其次，由于 KGC 的输出目标通常是特定于需求的，因此应根据模式自适应地调整提取结果。我们详细介绍了以下几个香草基线，并在 §3 中介绍了建议的ADAKGC
|
[
[
0.118408203125,
0.0872802734375,
0.5029296875,
0.0872802734375,
0.5029296875,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.51318359375,
0.0872802734375,
0.8759765625,
0.0872802734375,
0.8759765625,
0.097412109375,
0.51318359375,
0.097412109375
],
[
0.118408203125,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.5146484375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.5146484375,
0.11639404296875
],
[
0.12005615234375,
0.12054443359375,
0.484375,
0.12054443359375,
0.484375,
0.1312255859375,
0.12005615234375,
0.1312255859375
],
[
0.5146484375,
0.12054443359375,
0.87744140625,
0.12054443359375,
0.87744140625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.118408203125,
0.1353759765625,
0.486083984375,
0.1353759765625,
0.486083984375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.118408203125,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.5146484375,
0.152587890625,
0.8759765625,
0.152587890625,
0.8759765625,
0.163330078125,
0.5146484375,
0.163330078125
],
[
0.118408203125,
0.16748046875,
0.486083984375,
0.16748046875,
0.486083984375,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.51123046875,
0.166259765625,
0.87744140625,
0.166259765625,
0.87744140625,
0.179931640625,
0.51123046875,
0.179931640625
],
[
0.118408203125,
0.182861328125,
0.484375,
0.182861328125,
0.484375,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.509765625,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.509765625,
0.196533203125
],
[
0.118408203125,
0.1995849609375,
0.28369140625,
0.1995849609375,
0.28369140625,
0.213134765625,
0.118408203125,
0.213134765625
],
[
0.51318359375,
0.198974609375,
0.5341796875,
0.198974609375,
0.5341796875,
0.213134765625,
0.51318359375,
0.213134765625
],
[
0.1351318359375,
0.215576171875,
0.484375,
0.21435546875,
0.484375,
0.2286376953125,
0.1351318359375,
0.2298583984375
],
[
0.118408203125,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.51123046875,
0.22802734375,
0.85400390625,
0.2286376953125,
0.85400390625,
0.243408203125,
0.51123046875,
0.2423095703125
],
[
0.1175537109375,
0.2464599609375,
0.486083984375,
0.248779296875,
0.486083984375,
0.262451171875,
0.1175537109375,
0.260009765625
],
[
0.51123046875,
0.251708984375,
0.880859375,
0.251708984375,
0.880859375,
0.265380859375,
0.51123046875,
0.265380859375
],
[
0.12005615234375,
0.26416015625,
0.489501953125,
0.26416015625,
0.489501953125,
0.27783203125,
0.12005615234375,
0.27783203125
],
[
0.51123046875,
0.268310546875,
0.880859375,
0.268310546875,
0.880859375,
0.281982421875,
0.51123046875,
0.281982421875
],
[
0.118408203125,
0.27978515625,
0.48779296875,
0.27978515625,
0.48779296875,
0.29345703125,
0.118408203125,
0.29345703125
],
[
0.51123046875,
0.283935546875,
0.87890625,
0.283935546875,
0.87890625,
0.297607421875,
0.51123046875,
0.297607421875
],
[
0.1175537109375,
0.29443359375,
0.484375,
0.29443359375,
0.484375,
0.30810546875,
0.1175537109375,
0.30810546875
],
[
0.5087890625,
0.297607421875,
0.8818359375,
0.298583984375,
0.8818359375,
0.31640625,
0.5087890625,
0.3154296875
],
[
0.118408203125,
0.312255859375,
0.483642578125,
0.312255859375,
0.483642578125,
0.322509765625,
0.118408203125,
0.322509765625
],
[
0.51318359375,
0.31640625,
0.87890625,
0.31640625,
0.87890625,
0.330078125,
0.51318359375,
0.330078125
],
[
0.118408203125,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.118408203125,
0.341552734375
],
[
0.5087890625,
0.330078125,
0.8818359375,
0.330810546875,
0.8818359375,
0.3486328125,
0.5087890625,
0.347412109375
],
[
0.12005615234375,
0.345703125,
0.48779296875,
0.345703125,
0.48779296875,
0.355712890625,
0.12005615234375,
0.355712890625
],
[
0.51123046875,
0.3486328125,
0.880859375,
0.3486328125,
0.880859375,
0.3623046875,
0.51123046875,
0.3623046875
],
[
0.12005615234375,
0.3623046875,
0.486083984375,
0.3623046875,
0.486083984375,
0.372314453125,
0.12005615234375,
0.372314453125
],
[
0.51318359375,
0.364013671875,
0.880859375,
0.364013671875,
0.880859375,
0.377685546875,
0.51318359375,
0.377685546875
],
[
0.118408203125,
0.377197265625,
0.48779296875,
0.377197265625,
0.48779296875,
0.39013671875,
0.118408203125,
0.39013671875
],
[
0.51318359375,
0.380126953125,
0.880859375,
0.37890625,
0.880859375,
0.393798828125,
0.51318359375,
0.394287109375
],
[
0.118408203125,
0.393798828125,
0.48779296875,
0.393798828125,
0.48779296875,
0.40673828125,
0.118408203125,
0.40673828125
],
[
0.51318359375,
0.396728515625,
0.880859375,
0.396728515625,
0.880859375,
0.410888671875,
0.51318359375,
0.410888671875
],
[
0.118408203125,
0.4091796875,
0.273681640625,
0.4091796875,
0.273681640625,
0.4228515625,
0.118408203125,
0.4228515625
],
[
0.51318359375,
0.413330078125,
0.880859375,
0.41455078125,
0.880859375,
0.4287109375,
0.51318359375,
0.427490234375
],
[
0.515625,
0.429931640625,
0.8818359375,
0.429931640625,
0.8818359375,
0.443603515625,
0.515625,
0.443603515625
],
[
0.1368408203125,
0.435791015625,
0.490234375,
0.43701171875,
0.490234375,
0.451904296875,
0.1368408203125,
0.45068359375
],
[
0.515625,
0.44775390625,
0.87890625,
0.44775390625,
0.87890625,
0.457763671875,
0.515625,
0.457763671875
],
[
0.151123046875,
0.45361328125,
0.489501953125,
0.453125,
0.489501953125,
0.466064453125,
0.151123046875,
0.46728515625
],
[
0.5146484375,
0.4619140625,
0.87890625,
0.4619140625,
0.87890625,
0.4755859375,
0.5146484375,
0.4755859375
],
[
0.154541015625,
0.47021484375,
0.484375,
0.4697265625,
0.484375,
0.48095703125,
0.154541015625,
0.482177734375
],
[
0.51123046875,
0.476806640625,
0.880859375,
0.47802734375,
0.880859375,
0.4921875,
0.51123046875,
0.4912109375
],
[
0.154541015625,
0.485107421875,
0.484375,
0.485107421875,
0.484375,
0.498779296875,
0.154541015625,
0.498779296875
],
[
0.51123046875,
0.493408203125,
0.58837890625,
0.495361328125,
0.58837890625,
0.51025390625,
0.51123046875,
0.50830078125
],
[
0.154541015625,
0.5029296875,
0.240966796875,
0.5029296875,
0.240966796875,
0.51416015625,
0.154541015625,
0.51416015625
],
[
0.5146484375,
0.525390625,
0.5615234375,
0.525390625,
0.5615234375,
0.53759765625,
0.5146484375,
0.53759765625
],
[
0.55517578125,
0.52490234375,
0.791015625,
0.52490234375,
0.791015625,
0.5380859375,
0.55517578125,
0.5380859375
],
[
0.15283203125,
0.5458984375,
0.489501953125,
0.5458984375,
0.489501953125,
0.55859375,
0.15283203125,
0.55859375
],
[
0.1351318359375,
0.52685546875,
0.50146484375,
0.525390625,
0.50146484375,
0.54345703125,
0.1351318359375,
0.54443359375
],
[
0.51123046875,
0.5458984375,
0.880859375,
0.54638671875,
0.880859375,
0.56103515625,
0.51123046875,
0.56005859375
],
[
0.154541015625,
0.56103515625,
0.48779296875,
0.56103515625,
0.48779296875,
0.57470703125,
0.154541015625,
0.57470703125
],
[
0.51318359375,
0.5625,
0.87744140625,
0.5625,
0.87744140625,
0.576171875,
0.51318359375,
0.576171875
],
[
0.154541015625,
0.57763671875,
0.486083984375,
0.57763671875,
0.486083984375,
0.5927734375,
0.154541015625,
0.5927734375
],
[
0.5146484375,
0.5791015625,
0.8818359375,
0.5791015625,
0.8818359375,
0.5927734375,
0.5146484375,
0.5927734375
],
[
0.15283203125,
0.5927734375,
0.365234375,
0.59326171875,
0.365234375,
0.60791015625,
0.15283203125,
0.60693359375
],
[
0.515625,
0.5966796875,
0.86376953125,
0.5966796875,
0.86376953125,
0.60693359375,
0.515625,
0.60693359375
],
[
0.5146484375,
0.61083984375,
0.880859375,
0.61083984375,
0.880859375,
0.62451171875,
0.5146484375,
0.62451171875
],
[
0.1376953125,
0.62060546875,
0.48779296875,
0.62060546875,
0.48779296875,
0.63427734375,
0.1376953125,
0.63427734375
],
[
0.5146484375,
0.62890625,
0.87890625,
0.62890625,
0.87890625,
0.6396484375,
0.5146484375,
0.6396484375
],
[
0.15283203125,
0.63720703125,
0.48779296875,
0.63720703125,
0.48779296875,
0.65087890625,
0.15283203125,
0.65087890625
],
[
0.51318359375,
0.6435546875,
0.880859375,
0.6435546875,
0.880859375,
0.65673828125,
0.51318359375,
0.65673828125
],
[
0.154541015625,
0.65380859375,
0.48779296875,
0.65380859375,
0.48779296875,
0.66748046875,
0.154541015625,
0.66748046875
],
[
0.51318359375,
0.66015625,
0.880859375,
0.66015625,
0.880859375,
0.67333984375,
0.51318359375,
0.67333984375
],
[
0.151123046875,
0.66748046875,
0.310546875,
0.66943359375,
0.310546875,
0.68310546875,
0.151123046875,
0.68115234375
],
[
0.5146484375,
0.6767578125,
0.880859375,
0.6767578125,
0.880859375,
0.6904296875,
0.5146484375,
0.6904296875
],
[
0.1158447265625,
0.69775390625,
0.441650390625,
0.69775390625,
0.441650390625,
0.71142578125,
0.1158447265625,
0.71142578125
],
[
0.5146484375,
0.6923828125,
0.880859375,
0.6923828125,
0.880859375,
0.7060546875,
0.5146484375,
0.7060546875
],
[
0.51318359375,
0.70703125,
0.87890625,
0.70703125,
0.87890625,
0.72021484375,
0.51318359375,
0.72021484375
],
[
0.1175537109375,
0.7236328125,
0.328369140625,
0.7236328125,
0.328369140625,
0.73681640625,
0.1175537109375,
0.73681640625
],
[
0.51318359375,
0.72265625,
0.880859375,
0.7236328125,
0.880859375,
0.73828125,
0.51318359375,
0.73681640625
],
[
0.1175537109375,
0.74462890625,
0.484375,
0.74462890625,
0.484375,
0.75830078125,
0.1175537109375,
0.75830078125
],
[
0.51318359375,
0.74072265625,
0.87890625,
0.74072265625,
0.87890625,
0.75439453125,
0.51318359375,
0.75439453125
],
[
0.118408203125,
0.76025390625,
0.484375,
0.76025390625,
0.484375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.509765625,
0.7548828125,
0.6962890625,
0.7548828125,
0.6962890625,
0.7685546875,
0.509765625,
0.7685546875
],
[
0.118408203125,
0.77685546875,
0.484375,
0.77685546875,
0.484375,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.51123046875,
0.7861328125,
0.84619140625,
0.7861328125,
0.84619140625,
0.7998046875,
0.51123046875,
0.7998046875
],
[
0.118408203125,
0.79345703125,
0.483642578125,
0.79345703125,
0.483642578125,
0.80419921875,
0.118408203125,
0.80419921875
],
[
0.118408203125,
0.80859375,
0.48779296875,
0.80859375,
0.48779296875,
0.822265625,
0.118408203125,
0.822265625
],
[
0.51123046875,
0.80712890625,
0.880859375,
0.80810546875,
0.880859375,
0.822265625,
0.51123046875,
0.82177734375
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
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,
0.486083984375,
0.84130859375,
0.486083984375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51318359375,
0.859375,
0.87744140625,
0.859375,
0.87744140625,
0.869140625,
0.51318359375,
0.869140625
],
[
0.1175537109375,
0.87255859375,
0.48779296875,
0.87158203125,
0.48779296875,
0.88525390625,
0.1175537109375,
0.8857421875
],
[
0.5146484375,
0.87451171875,
0.8759765625,
0.87451171875,
0.8759765625,
0.88525390625,
0.5146484375,
0.88525390625
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90380859375,
0.48779296875,
0.90478515625,
0.48779296875,
0.91943359375,
0.1175537109375,
0.91845703125
],
[
0.51318359375,
0.90625,
0.591796875,
0.90625,
0.591796875,
0.91748046875,
0.51318359375,
0.91748046875
]
] |
[
"EE的三个任务上构建数据集。通过实证分析，",
"的方式描述。此外，TAC-KBP(Ellisetal.2014",
"我们注意到Text2Event(Luet al.,2021）、TANL",
"旨在利用现有的通用域结构化数据源并扩展采",
"(Paolinietal.,2021)、UIE（Luetal.,2022)和",
"用描述性文本作为附加信息的实体链接。OAE",
"GPT-3的方法。5（Ouyangetal.,2022）无法有效",
"(Euzenatetal.，2011)基于两个或多个现有本体",
"提取给定复杂演化模式的信息。我们认为主要",
"或知识图之间的对齐创建集成本体。在本文中",
"问题如下：1）如何学习动态和可概括的模式",
"我们专注于从非结构化文本中提取知识实例的",
"表示作为提取条件；2）如何精确提取受新更",
"工作，这被视为模式约束预测（结构预测）任",
"新模式约束的新实例",
"务",
"为此，我们提出了一个简单的基线，称为",
"ADAKGC，它引入了富含模式的前缀指导者",
"2.2Definition of Schema-adaptableKGC",
"（SPI）来表示和传递可学习的特定于模式的知",
"在现实世界中，KGC系统从非结构化文本",
"识。在每个选代阶段，我们从当前模式图线性",
"中提取结构化知识，并根据频繁调整的模",
"转换为可学习的提示，使用本体名称初始化",
"式将其规范化为实例图。给定一组模式图",
"并连接到特定于任务的前缀。为了鼓励解码器",
"S=s1,82，，sn，可适应模式的KGC的",
"理解动态模式，我们使用了模式条件动态解码",
"任务是为每次选代生成一组模式约束实例",
"(SDD）策略，该策略构造了特定于模式的词",
"g=91,92，…,9n）。假设有一个在初始训练",
"汇表到输出空间的解码路径。当模式改变时，",
"集上训练的模型Mg=LM(Dtrain/Si），之后",
"我们动态地构造一个新的特单树来调整输出空",
"更新模式的标记数据不可用。提供了一个模式",
"间。请注意，ADAKGC与模型无关，可以处理",
"自适应数据流D(1),D(2),..,D(M)来评估",
"各种具有挑战性的模式演化方案。我们将这项",
"模型对模式动态更新的适应性。每个D(K)都",
"工作的贡献总结为：",
"包含开发/测试数据（Dlev，Dtest）和模式图Sk。",
"请注意，该模型不会被重新训练，但希望能够",
". We introduce a new task of the schema-",
"利用不断发展的模式来提取信息。面临的挑战",
"adaptable KGC to meet the schema evolu-",
"是，该模型有望在测试集Dest的每次迭代中",
"tion requirements, which is a new branch that",
"表现良好，该测试集包含为更新的模式更改的",
"has not been well-explored to the best of our",
"黄金实例",
"knowledge.",
"2.3",
"Dataset Construction Process",
"which includes schema-enriched prefix",
"·We propose a new baseline ADAKGC,",
"如图E2所示，我们针对不同类型的模式演化",
"instructors and schema-conditioned dynamic",
"设计了三个原则，并应用AlgorithmE1构建了",
"decoding strategy,and experimentally",
"用于评估的数据集：（1）水平模式扩展要求模",
"demonstrate the adaptability",
"式添加相同级别的新类节点，可以将其视为",
"种形式的类增量学习实例，。基于对相邻新类",
": We release the schema-adaptable KGC",
"的泛化效应，我们可以评估模式特征的传输能",
"benchmark, which imposes new challenges",
"力。（2）垂直模式扩展要求模式添加父类的子",
"and presents new research opportunities for",
"类。基于对子类的泛化效应，我们可以评估模",
"the NLP community.",
"式特征的继承和派生能力。（3）混合模式扩展",
"2Problem Statement and Overview",
"要求模式在每次选代时随机地水平或垂直扩展",
"节点，这总结了模式图并代表了它们潜在的共",
"2.1 Background of KGC",
"同进化模式。更多细节在附录A.1中，除了上",
"KGC一直是 (Lu et al., 2022; Zhang et al., 2022b)",
"述结构扩展之外，我们还从语义的角度进一步",
"的一项有前途的研究挑战，现有的基准测试利",
"探讨了类似的节点替换",
"用定义良好的模式来指导知识图构建，重点是",
"2.4Schema-adaptable KGC Benchmark",
"生成特定领域的知识图或聚合异构结构化数据",
"库。例如，FEW-NERD(Dingetal.,2021)由粗",
"模式适应性KGC有两个挑战。首先，由于模",
"粒度和细粒度实体类型定义组成，用于定位和",
"式在每次迭代中都会更新，因此需要将模式演",
"分类来自非结构化自然语言的命名实体。NYT",
"变信息动态注入模型中。其次，由于KGC的",
"(Riedeletal.，2010a）根据特定的分类结构专门",
"输出自标通常是特定于需求的，因此应根据",
"从文本数据源中提取关系三重实例。ACE2005",
"模式自适应地调整提取结果。我们详细介绍了",
"(Ntroduction）根据上下文识别触发器和事件类",
"以下几个香草基线，并在$3中介绍了建议的",
"型，并且每个都有自己的事件参数，以填充槽",
"ADAKGC"
] |
{
"arxiv_id": "2305.08703",
"language": "zh",
"page_idx": 2
}
|
||
2305.08703-zh-3
|
2305.08703-zh-3.jpg
|
<FigureHere>
Figure 2: Illustration of schema-adaptable KGC with iterative schema evolution. Schema graph nodes (colored $\bullet$ ) are evolved by adding a new class node (colored $\color{red}{\bullet}$ )in horizontal schema expansion while is inherited by a new subclass node (colored $\bullet)$ in vertical schema expansion.
<FigureHere>
Figure 3: The overview of ADA ptive K nowledge G raph C onstruction ADAKGC .
3 $\textbf{The Proposed Baseline: ADAKGC}$
3.1 0verview
如图E3 所示，ADAKGC 利用预先训练的编码器-解码器语言模型(LM) T5(Raffel et al., 2020a) 作为模式适应性 KGC 任务的基本体系结构。具体来说，让编码器输入$X_en=[S;X]$ 是模式$S$和输入$X$的串联。在解码过程中， LM 计算在给定先前令牌$y_{<t}$的情况下生成新令牌$y_t$的条件概率：
Vanilla Baselines: 模式适应性 KGC 可以被认为是一种结构化的预测语言任务，它通过结构的通用性在类节点之间传递信息。TANL (Paolini et al., 2021)引入了一个增强的自然语言翻译任务，从中可以隐式提取与模式相关的信息。TEXT2EVENT (Lu et al., 2021) 是用于事件提取的统一序列到结构架构，具有用于推理期间事件模式知识注人的受限解码算法。UIE (Lu et al., 2022) 是一个统一的文本到结构生成框架，可以对不同的 IE 任务进行统一建模，并通过基于模式的提示机制自适应地生成目标序列。GPT-3 (Brown et al., 2020) 是一种大型语言模型 (LLM),可以作为模式适应性 KGC 的基准。尽管当前专注于结构化提取的工作可以通过静态类型的知识实现出色的性能，但它们通常不了解模式的演变。为了澄清这个问题，我们引人了一个简单而有效的基线，称为模 式- ADApositive KnowledgeGraphConstructure ADAKGC
$$p(Y_{de}\mid X_{en})=\prod\limits_{t=1}^{|Y|}p\left(y_t\mid y_{<t},S,X\right)$$
(1)
我们使用预训练参数$\theta$初始化模型。在这里， $p_\theta$是一个可训练的语言模型发行版。在第$k$次迭代中，我们对以下对数似然目标执行梯度更新：
|
[
[
0.133544921875,
0.08905029296875,
0.189697265625,
0.08905029296875,
0.189697265625,
0.0968017578125,
0.133544921875,
0.0968017578125
],
[
0.429931640625,
0.0926513671875,
0.57080078125,
0.0914306640625,
0.57080078125,
0.10272216796875,
0.429931640625,
0.10394287109375
],
[
0.7109375,
0.08905029296875,
0.85986328125,
0.08905029296875,
0.85986328125,
0.09979248046875,
0.7109375,
0.09979248046875
],
[
0.1275634765625,
0.0968017578125,
0.28369140625,
0.0968017578125,
0.28369140625,
0.10986328125,
0.1275634765625,
0.10986328125
],
[
0.432373046875,
0.1051025390625,
0.57421875,
0.1051025390625,
0.57421875,
0.11517333984375,
0.432373046875,
0.11517333984375
],
[
0.71533203125,
0.1015625,
0.85791015625,
0.10272216796875,
0.85791015625,
0.114013671875,
0.71533203125,
0.1134033203125
],
[
0.130126953125,
0.1092529296875,
0.29296875,
0.1092529296875,
0.29296875,
0.119384765625,
0.130126953125,
0.119384765625
],
[
0.1292724609375,
0.1181640625,
0.298828125,
0.119384765625,
0.298828125,
0.1312255859375,
0.1292724609375,
0.1300048828125
],
[
0.441650390625,
0.1181640625,
0.56494140625,
0.1181640625,
0.56494140625,
0.12890625,
0.441650390625,
0.12890625
],
[
0.724609375,
0.11639404296875,
0.8505859375,
0.11639404296875,
0.8505859375,
0.1270751953125,
0.724609375,
0.1270751953125
],
[
0.71875,
0.1300048828125,
0.8564453125,
0.1300048828125,
0.8564453125,
0.14013671875,
0.71875,
0.14013671875
],
[
0.139404296875,
0.13720703125,
0.1822509765625,
0.1395263671875,
0.181396484375,
0.149658203125,
0.1376953125,
0.1478271484375
],
[
0.36865234375,
0.135986328125,
0.52880859375,
0.135986328125,
0.52880859375,
0.146728515625,
0.36865234375,
0.146728515625
],
[
0.391357421875,
0.146728515625,
0.5146484375,
0.146728515625,
0.5146484375,
0.15673828125,
0.391357421875,
0.15673828125
],
[
0.634765625,
0.1455078125,
0.77685546875,
0.1455078125,
0.77685546875,
0.1561279296875,
0.634765625,
0.1561279296875
],
[
0.640625,
0.155029296875,
0.77099609375,
0.155029296875,
0.77099609375,
0.168701171875,
0.640625,
0.168701171875
],
[
0.3720703125,
0.157958984375,
0.52880859375,
0.157958984375,
0.52880859375,
0.1650390625,
0.3720703125,
0.1650390625
],
[
0.18896484375,
0.1644287109375,
0.202392578125,
0.160888671875,
0.204833984375,
0.168701171875,
0.193115234375,
0.171630859375
],
[
0.62646484375,
0.166259765625,
0.78271484375,
0.166259765625,
0.78271484375,
0.177001953125,
0.62646484375,
0.177001953125
],
[
0.62060546875,
0.1787109375,
0.7841796875,
0.1787109375,
0.7841796875,
0.1859130859375,
0.62060546875,
0.1859130859375
],
[
0.41064453125,
0.179931640625,
0.495361328125,
0.179931640625,
0.495361328125,
0.18701171875,
0.41064453125,
0.18701171875
],
[
0.1435546875,
0.149658203125,
0.181396484375,
0.149658203125,
0.181396484375,
0.1591796875,
0.1435546875,
0.1591796875
],
[
0.14697265625,
0.182861328125,
0.2822265625,
0.182861328125,
0.2822265625,
0.193603515625,
0.14697265625,
0.193603515625
],
[
0.491943359375,
0.179931640625,
0.5498046875,
0.179931640625,
0.5498046875,
0.18701171875,
0.491943359375,
0.18701171875
],
[
0.139404296875,
0.20068359375,
0.285400390625,
0.20068359375,
0.285400390625,
0.21435546875,
0.139404296875,
0.21435546875
],
[
0.414794921875,
0.19775390625,
0.5458984375,
0.19775390625,
0.5458984375,
0.2078857421875,
0.414794921875,
0.2078857421875
],
[
0.6708984375,
0.201904296875,
0.68701171875,
0.201904296875,
0.68701171875,
0.21142578125,
0.6708984375,
0.21142578125
],
[
0.426513671875,
0.21435546875,
0.5810546875,
0.21435546875,
0.5810546875,
0.2244873046875,
0.426513671875,
0.2244873046875
],
[
0.68994140625,
0.21435546875,
0.84228515625,
0.21435546875,
0.84228515625,
0.2244873046875,
0.68994140625,
0.2244873046875
],
[
0.4306640625,
0.2257080078125,
0.5751953125,
0.2257080078125,
0.5751953125,
0.2392578125,
0.4306640625,
0.2392578125
],
[
0.7001953125,
0.22802734375,
0.82958984375,
0.22802734375,
0.82958984375,
0.2381591796875,
0.7001953125,
0.2381591796875
],
[
0.130126953125,
0.2392578125,
0.231689453125,
0.2392578125,
0.231689453125,
0.2493896484375,
0.130126953125,
0.2493896484375
],
[
0.634765625,
0.236328125,
0.6455078125,
0.236328125,
0.6455078125,
0.240478515625,
0.634765625,
0.240478515625
],
[
0.3896484375,
0.2476806640625,
0.54736328125,
0.2476806640625,
0.54736328125,
0.25830078125,
0.3896484375,
0.25830078125
],
[
0.64990234375,
0.2464599609375,
0.8134765625,
0.2464599609375,
0.8134765625,
0.260009765625,
0.64990234375,
0.260009765625
],
[
0.6767578125,
0.2587890625,
0.79248046875,
0.2587890625,
0.79248046875,
0.26953125,
0.6767578125,
0.26953125
],
[
0.139404296875,
0.25048828125,
0.285400390625,
0.25048828125,
0.285400390625,
0.26416015625,
0.139404296875,
0.26416015625
],
[
0.151123046875,
0.262451171875,
0.276123046875,
0.262451171875,
0.276123046875,
0.275634765625,
0.151123046875,
0.275634765625
],
[
0.4130859375,
0.260009765625,
0.53564453125,
0.260009765625,
0.53564453125,
0.267333984375,
0.4130859375,
0.267333984375
],
[
0.1175537109375,
0.290283203125,
0.880859375,
0.290283203125,
0.880859375,
0.303955078125,
0.1175537109375,
0.303955078125
],
[
0.1175537109375,
0.3046875,
0.8818359375,
0.3046875,
0.8818359375,
0.318359375,
0.1175537109375,
0.318359375
],
[
0.1158447265625,
0.317626953125,
0.497802734375,
0.317626953125,
0.497802734375,
0.330810546875,
0.1158447265625,
0.330810546875
],
[
0.47265625,
0.355224609375,
0.51904296875,
0.355224609375,
0.51904296875,
0.3623046875,
0.47265625,
0.3623046875
],
[
0.68994140625,
0.3515625,
0.80615234375,
0.352783203125,
0.80615234375,
0.363525390625,
0.68994140625,
0.3623046875
],
[
0.47021484375,
0.361083984375,
0.67333984375,
0.35986328125,
0.67333984375,
0.37060546875,
0.47021484375,
0.371826171875
],
[
0.6884765625,
0.35986328125,
0.861328125,
0.35986328125,
0.861328125,
0.37060546875,
0.6884765625,
0.37060546875
],
[
0.291259765625,
0.363525390625,
0.39794921875,
0.363525390625,
0.39794921875,
0.37353515625,
0.291259765625,
0.37353515625
],
[
0.145263671875,
0.363525390625,
0.2032470703125,
0.363525390625,
0.2032470703125,
0.37060546875,
0.145263671875,
0.37060546875
],
[
0.470947265625,
0.37060546875,
0.67529296875,
0.37060546875,
0.67529296875,
0.38134765625,
0.470947265625,
0.38134765625
],
[
0.69287109375,
0.37060546875,
0.8623046875,
0.37060546875,
0.8623046875,
0.377685546875,
0.69287109375,
0.377685546875
],
[
0.47021484375,
0.377197265625,
0.56103515625,
0.37890625,
0.56103515625,
0.3896484375,
0.47021484375,
0.38720703125
],
[
0.68701171875,
0.377685546875,
0.81005859375,
0.37890625,
0.81005859375,
0.38720703125,
0.68701171875,
0.385986328125
],
[
0.26025390625,
0.40087890625,
0.424072265625,
0.40087890625,
0.424072265625,
0.410888671875,
0.26025390625,
0.410888671875
],
[
0.76318359375,
0.3955078125,
0.85546875,
0.3955078125,
0.85546875,
0.40625,
0.76318359375,
0.40625
],
[
0.69287109375,
0.41455078125,
0.81005859375,
0.41455078125,
0.81005859375,
0.421630859375,
0.69287109375,
0.421630859375
],
[
0.69287109375,
0.4228515625,
0.72119140625,
0.4228515625,
0.72119140625,
0.429931640625,
0.69287109375,
0.429931640625
],
[
0.75390625,
0.4228515625,
0.85986328125,
0.4228515625,
0.85986328125,
0.429931640625,
0.75390625,
0.429931640625
],
[
0.14111328125,
0.43115234375,
0.160400390625,
0.431640625,
0.160400390625,
0.435302734375,
0.14111328125,
0.43408203125
],
[
0.57080078125,
0.43115234375,
0.6513671875,
0.43115234375,
0.6513671875,
0.438232421875,
0.57080078125,
0.438232421875
],
[
0.69287109375,
0.429931640625,
0.8623046875,
0.429931640625,
0.8623046875,
0.440673828125,
0.69287109375,
0.440673828125
],
[
0.5146484375,
0.431640625,
0.53076171875,
0.431640625,
0.53076171875,
0.439453125,
0.5146484375,
0.439453125
],
[
0.16455078125,
0.431640625,
0.202392578125,
0.431640625,
0.202392578125,
0.435791015625,
0.16455078125,
0.435791015625
],
[
0.490234375,
0.441162109375,
0.65576171875,
0.441162109375,
0.65576171875,
0.448974609375,
0.490234375,
0.448974609375
],
[
0.6943359375,
0.439453125,
0.81005859375,
0.439453125,
0.81005859375,
0.446533203125,
0.6943359375,
0.446533203125
],
[
0.69189453125,
0.448974609375,
0.71875,
0.448974609375,
0.71875,
0.4560546875,
0.69189453125,
0.4560546875
],
[
0.7900390625,
0.45068359375,
0.8095703125,
0.45068359375,
0.8095703125,
0.45361328125,
0.7900390625,
0.45361328125
],
[
0.8388671875,
0.45068359375,
0.8505859375,
0.45068359375,
0.8505859375,
0.45361328125,
0.8388671875,
0.45361328125
],
[
0.5029296875,
0.463134765625,
0.5439453125,
0.463134765625,
0.5439453125,
0.47021484375,
0.5029296875,
0.47021484375
],
[
0.599609375,
0.463134765625,
0.64404296875,
0.463134765625,
0.64404296875,
0.47021484375,
0.599609375,
0.47021484375
],
[
0.77685546875,
0.458984375,
0.78271484375,
0.458984375,
0.78271484375,
0.465576171875,
0.77685546875,
0.465576171875
],
[
0.172119140625,
0.47265625,
0.1796875,
0.47265625,
0.1796875,
0.476806640625,
0.172119140625,
0.476806640625
],
[
0.2626953125,
0.46435546875,
0.4248046875,
0.46435546875,
0.4248046875,
0.474365234375,
0.2626953125,
0.474365234375
],
[
0.55517578125,
0.478515625,
0.75146484375,
0.478515625,
0.75146484375,
0.4892578125,
0.55517578125,
0.4892578125
],
[
0.17626953125,
0.4892578125,
0.189697265625,
0.4892578125,
0.189697265625,
0.493408203125,
0.17626953125,
0.493408203125
],
[
0.47607421875,
0.49755859375,
0.49365234375,
0.49755859375,
0.49365234375,
0.501953125,
0.47607421875,
0.501953125
],
[
0.5576171875,
0.495361328125,
0.63232421875,
0.4912109375,
0.6337890625,
0.505859375,
0.55908203125,
0.51025390625
],
[
0.26025390625,
0.498779296875,
0.34765625,
0.498779296875,
0.34765625,
0.5087890625,
0.26025390625,
0.5087890625
],
[
0.1376953125,
0.498779296875,
0.219970703125,
0.498779296875,
0.219970703125,
0.5087890625,
0.1376953125,
0.5087890625
],
[
0.450927734375,
0.498779296875,
0.47607421875,
0.498779296875,
0.47607421875,
0.5029296875,
0.450927734375,
0.5029296875
],
[
0.252685546875,
0.50830078125,
0.35693359375,
0.50732421875,
0.35693359375,
0.51708984375,
0.252685546875,
0.5185546875
],
[
0.455078125,
0.5087890625,
0.491943359375,
0.5087890625,
0.491943359375,
0.5166015625,
0.455078125,
0.5166015625
],
[
0.8369140625,
0.5126953125,
0.85791015625,
0.5166015625,
0.85546875,
0.52392578125,
0.83447265625,
0.52099609375
],
[
0.432373046875,
0.51318359375,
0.44580078125,
0.51318359375,
0.44580078125,
0.51708984375,
0.432373046875,
0.51708984375
],
[
0.130126953125,
0.51025390625,
0.231689453125,
0.51025390625,
0.231689453125,
0.51708984375,
0.130126953125,
0.51708984375
],
[
0.246826171875,
0.51708984375,
0.353515625,
0.51708984375,
0.353515625,
0.52490234375,
0.246826171875,
0.52490234375
],
[
0.611328125,
0.5166015625,
0.62109375,
0.5166015625,
0.62109375,
0.521484375,
0.611328125,
0.521484375
],
[
0.43994140625,
0.521484375,
0.453369140625,
0.521484375,
0.453369140625,
0.52783203125,
0.43994140625,
0.52783203125
],
[
0.130126953125,
0.5185546875,
0.2333984375,
0.5185546875,
0.2333984375,
0.525390625,
0.130126953125,
0.525390625
],
[
0.25439453125,
0.525390625,
0.343505859375,
0.525390625,
0.343505859375,
0.533203125,
0.25439453125,
0.533203125
],
[
0.6708984375,
0.521484375,
0.68603515625,
0.521484375,
0.68603515625,
0.52978515625,
0.6708984375,
0.52978515625
],
[
0.478515625,
0.52685546875,
0.5029296875,
0.52490234375,
0.5029296875,
0.52783203125,
0.478515625,
0.52978515625
],
[
0.130126953125,
0.52685546875,
0.21826171875,
0.52685546875,
0.21826171875,
0.53369140625,
0.130126953125,
0.53369140625
],
[
0.724609375,
0.52685546875,
0.7314453125,
0.52685546875,
0.7314453125,
0.52978515625,
0.724609375,
0.52978515625
],
[
0.255126953125,
0.533203125,
0.387939453125,
0.533203125,
0.387939453125,
0.54052734375,
0.255126953125,
0.54052734375
],
[
0.611328125,
0.533203125,
0.623046875,
0.533203125,
0.623046875,
0.5361328125,
0.611328125,
0.5361328125
],
[
0.8310546875,
0.53076171875,
0.85986328125,
0.53515625,
0.8564453125,
0.54150390625,
0.82861328125,
0.53759765625
],
[
0.2149658203125,
0.56005859375,
0.779296875,
0.56005859375,
0.779296875,
0.57080078125,
0.2149658203125,
0.57080078125
],
[
0.1259765625,
0.60107421875,
0.486083984375,
0.60107421875,
0.486083984375,
0.61083984375,
0.1259765625,
0.61083984375
],
[
0.509765625,
0.60107421875,
0.52880859375,
0.60107421875,
0.52880859375,
0.609375,
0.509765625,
0.609375
],
[
0.5263671875,
0.6015625,
0.54833984375,
0.6015625,
0.54833984375,
0.609375,
0.5263671875,
0.609375
],
[
0.54248046875,
0.5986328125,
0.84619140625,
0.5986328125,
0.84619140625,
0.6123046875,
0.54248046875,
0.6123046875
],
[
0.118408203125,
0.615234375,
0.486083984375,
0.615234375,
0.486083984375,
0.62890625,
0.118408203125,
0.62890625
],
[
0.497802734375,
0.60400390625,
0.50537109375,
0.60400390625,
0.50537109375,
0.60693359375,
0.497802734375,
0.60693359375
],
[
0.118408203125,
0.63134765625,
0.484375,
0.63134765625,
0.484375,
0.64453125,
0.118408203125,
0.64453125
],
[
0.51318359375,
0.63134765625,
0.54150390625,
0.63134765625,
0.54150390625,
0.642578125,
0.51318359375,
0.642578125
],
[
0.55517578125,
0.62890625,
0.634765625,
0.63134765625,
0.634765625,
0.64453125,
0.55517578125,
0.642578125
],
[
0.1175537109375,
0.64794921875,
0.486083984375,
0.64794921875,
0.486083984375,
0.6611328125,
0.1175537109375,
0.6611328125
],
[
0.51318359375,
0.65673828125,
0.880859375,
0.65673828125,
0.880859375,
0.67041015625,
0.51318359375,
0.67041015625
],
[
0.12005615234375,
0.66455078125,
0.484375,
0.66455078125,
0.484375,
0.677734375,
0.12005615234375,
0.677734375
],
[
0.51123046875,
0.67138671875,
0.87744140625,
0.67138671875,
0.87744140625,
0.68505859375,
0.51123046875,
0.68505859375
],
[
0.118408203125,
0.6787109375,
0.484375,
0.6787109375,
0.484375,
0.6923828125,
0.118408203125,
0.6923828125
],
[
0.1175537109375,
0.69482421875,
0.486083984375,
0.6953125,
0.486083984375,
0.71044921875,
0.1175537109375,
0.708984375
],
[
0.51318359375,
0.689453125,
0.880859375,
0.689453125,
0.880859375,
0.703125,
0.51318359375,
0.703125
],
[
0.1175537109375,
0.71142578125,
0.48779296875,
0.71044921875,
0.48779296875,
0.724609375,
0.1175537109375,
0.7255859375
],
[
0.51318359375,
0.7060546875,
0.880859375,
0.7060546875,
0.880859375,
0.7197265625,
0.51318359375,
0.7197265625
],
[
0.51318359375,
0.7216796875,
0.87744140625,
0.7216796875,
0.87744140625,
0.7353515625,
0.51318359375,
0.7353515625
],
[
0.118408203125,
0.72802734375,
0.486083984375,
0.72802734375,
0.486083984375,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.509765625,
0.736328125,
0.8818359375,
0.73681640625,
0.8818359375,
0.751953125,
0.509765625,
0.75048828125
],
[
0.1175537109375,
0.7421875,
0.483642578125,
0.7421875,
0.483642578125,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.5146484375,
0.7529296875,
0.67529296875,
0.7529296875,
0.67529296875,
0.76708984375,
0.5146484375,
0.76708984375
],
[
0.118408203125,
0.76025390625,
0.484375,
0.76025390625,
0.484375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.1175537109375,
0.77490234375,
0.48193359375,
0.77490234375,
0.48193359375,
0.7880859375,
0.1175537109375,
0.7880859375
],
[
0.118408203125,
0.7919921875,
0.484375,
0.7919921875,
0.484375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.68115234375,
0.7880859375,
0.70361328125,
0.7880859375,
0.70361328125,
0.80029296875,
0.68115234375,
0.80029296875
],
[
0.5576171875,
0.802734375,
0.83642578125,
0.802734375,
0.83642578125,
0.82080078125,
0.5576171875,
0.82080078125
],
[
0.8564453125,
0.8046875,
0.8818359375,
0.8046875,
0.8818359375,
0.81787109375,
0.8564453125,
0.81787109375
],
[
0.67822265625,
0.82177734375,
0.705078125,
0.82177734375,
0.705078125,
0.8330078125,
0.67822265625,
0.8330078125
],
[
0.118408203125,
0.80810546875,
0.483642578125,
0.80810546875,
0.483642578125,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.12005615234375,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.12005615234375,
0.85498046875
],
[
0.1175537109375,
0.85693359375,
0.483642578125,
0.85693359375,
0.483642578125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.52880859375,
0.85498046875,
0.880859375,
0.85546875,
0.880859375,
0.87060546875,
0.52880859375,
0.869140625
],
[
0.118408203125,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.880859375,
0.87158203125,
0.880859375,
0.8857421875,
0.51123046875,
0.88720703125
],
[
0.118408203125,
0.8896484375,
0.197265625,
0.8896484375,
0.197265625,
0.90087890625,
0.118408203125,
0.90087890625
],
[
0.51318359375,
0.88818359375,
0.880859375,
0.88818359375,
0.880859375,
0.90185546875,
0.51318359375,
0.90185546875
],
[
0.51318359375,
0.90478515625,
0.5400390625,
0.90478515625,
0.5400390625,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"Input Text:",
"(people,location,location)",
"(people,location,location)",
"Professor Dr.Weldon with his",
"(location,containslocation)",
"(locationcontains,location)",
"son Charles traveled throughout",
"the world, living in Paris, France.",
"(people,children,person)",
"(people,children,person)",
"(people,profession,person)",
"Schema",
"(Dr.Weldon.place lived,Paris)",
"(France,contains,Paris)",
"(Dr.Weldon,location,Paris)",
"(France,contains,Paris)",
"(DrWeldon.children.Charles)",
"",
"(Dr.Weldon.children.children)",
"(Dr.Weldon.profession.Professo!",
"HorizontalSchem",
"Graph",
"(people.location,location)",
"aExpansion",
"(location,containslocation)",
"Vertical Schema Expansion",
"",
"(peopleplacelivedlocation)",
"(people,placelived,location)",
"(location,contains,location)",
"(location,capital,location)",
"Instance Extraction:",
"BR",
"(Dr.Weldon,place lived,Paris)",
"(Dr. Weldon,place lived, Paris)",
"(France.capital,Paris)",
"(Dr. Weldon,location,Paris)",
"(France,contains,Paris)",
"(France.contains.Paris)",
"Figure 2: Illustration of schema-adaptable KGC with iterative schema evolution. Schema graph nodes (colored ",
") are evolved by adding a new class node (colored ) in horizontal schema expansion while is inherited by a new",
"subclassnode(colored)invertical schemaexpansion",
"InputText:",
"Iteration n Output Instance:",
"Subsequent to his retirement, Professor Dr.Weldon",
"(France,contains,Paris)(Dr.Weldon,place",
"Schema-pecific Prefix",
"Iteration n schen",
"with his son Charles traveled throughout the world.",
"lived.Paris)(Dr.Weldon.place lived.France",
"living inParisFran",
"(Dr.Weldon,children,Charles)",
"tpC, tpC2 tpC; tpCg tpCs-PAD-<PAD<PAD",
"ice Database Update",
"Iterationn+IOutputInstanc",
"(France",
"s.Paris)(Dr.Weldon.place",
"tn",
"inputtextembedding",
"live,Paris) (Dr. Weldon,place lived,Sydney",
"spe",
"ntlsehe",
"Schema-enriched PrefixInstructor",
"(Dr.Weldon.children.Charles",
"(Paris",
"O",
"T",
"Encoder",
"Dncoder",
"1",
"",
"tpc, tpC2tpC, tpc, tpCs tpC<PADs-PAD",
"Schema-conditioned Dynamic Decoding",
"n+",
"Pat",
"区lterationn",
"Iteration n+1 trie tree",
"Iteration n trie tree",
"locatios",
"(people,place lived,location)",
"ation>Fr",
"OR",
"ho",
"(people,place lived.location)",
"(location,contains,location)",
"",
"大",
"(location,contains,location)",
"(people,children,person)",
"贝",
".",
"(people.children.person)",
"",
"(capitaladministrativedivisions,cou",
"",
"3",
"Figure3:Theoverviewof ADAptiveKnowledgeGraphC onstructionADAKGC",
"VanillaBaselines：模式适应性KGC可以被",
"3",
"",
"TheProposed Baseline:ADAKGC",
"认为是一种结构化的预测语言任务，它通过",
"",
"结构的通用性在类节点之间传递信息。TANL",
"3.1 ",
"Overview",
"(Paolinietal.，2021)引入了一个增强的自然语",
"如图E3所示，ADAKGC利用预先训练的编",
"言翻译任务，从中可以隐式提取与模式相关的",
"码器-解码器语言模型（LM）T5（Raffeletal.",
"信息。TEXT2EVENT(Luetal.,2021）是用于事",
"件提取的统一序列到结构架构，具有用于推理",
"2020a）作为模式适应性KGC任务的基本体系",
"期间事件模式知识注入的受限解码算法。UIE",
"结构。具体来说，让编码器输人Xen=[S;X]",
"是模式 S和输人X的串联。在解码过程中",
"(Luetal.，2022)是一个统一的文本到结构生成",
"LM计算在给定先前令牌y<t的情况下生成新",
"框架，可以对不同的E任务进行统一建模，并",
"令牌yt的条件概率：",
"通过基于模式的提示机制自适应地生成目标序",
"列。GPT-3（Brownetal.2020）是一种大型语言",
"模型（LLM），可以作为模式适应性KGC的",
"[Y]",
"p(Yde I Xen) = IIp(yt I y<t, S, X)",
"(1)",
"t=1",
"基准。尽管当前专注于结构化提取的工作可",
"以通过静态类型的知识实现出色的性能，但",
"它们通常不了解模式的演变。为了澄清这个问",
"题，我们引入了一个简单而有效的基线，称为",
"我们使用预训练参数?初始化模型。在这里，",
"模式-ADApositiveKnowledgeGraphConstructure",
"pe是一个可训练的语言模型发行版。在第k次",
"ADAKGC",
"迭代中，我们对以下对数似然目标执行梯度更",
"新："
] |
{
"arxiv_id": "2305.08703",
"language": "zh",
"page_idx": 3
}
|
||
2307.05440-zh-1
|
2307.05440-zh-1.jpg
|
ISLTranslate: 用于翻译印度手语的数据集
$\mathbf{Abhinav~Joshi}^{1}$ $\textbf{Susmit Agrawal}^{2}$ $\textbf{Ashutosh Modi}^{1}$
$^{1}$Indian Institute of Technology Kanpur (IIT Kanpur)
$^2$Indian Institute of Technology Hyderabad (IIT Hyderabad)
$\{$ ajoshi, ashutoshm $\}$ @cse.iitk.ac.in,ai22mtech12002@iith.ac.in
## $\mathbf{Abstract}$
Sign languages are the primary means of communication for many hard-of-hearing people worldwide. Recently, to bridge the communication gap between the hard-of-hearing community and the rest of the population, several sign language translation datasets have been proposed to enable the development of statistical sign language translation systems. However, there is a dearth of sign language resources for the Indian sign language. This resource paper introduces ISLTranslate, a translation dataset for continuous Indian Sign Language (ISL) consisting of 31k ISL-English sentence/phrase pairs. To the best of our knowledge, it is the largest translation dataset for continuous Indian Sign Language. We provide a detailed analysis of the dataset. To validate the performance of existing end-to-end Sign language to spoken language translation systems, we benchmark the created dataset with a transformer-based model for ISL translation.
## 1 $\textbf{Introduction}$
全球约有 4.3 亿听力障碍者$^1$,其中 6300 万在印度$^{2}$。手语是听力障碍社区的主要交流方式。尽管自然语言处理技术在过去五年中已显示出巨大的进步，但主要是由于注释资源和大型语言模型 (Tunstall et al., 2022) 的可用性，像手语这样具有身体形态的语言仍然缺乏有效的语言处理系统。最近，手语研究已开始在 NLP 社区 (Yin et al., 2021; Koller et al., 2015; Sincan and Keles, 2020; Xu et al., 2022; Albanie et al., 2020; Jiang et al., 2021; Moryossef et al., 2020; Joshi et al., 2022) 中引起关注。翻译数据集的可用性改善了手语的 NLP 系统的研究和开发，例如 ASL(美国手语)(Li et al., 2020), BSL(英国手语)(Albanie et al., 2021)和
$^{\mathrm{l}}$https://www.who.int/news-room/fact-sheets/
detail/deafness- and-hearing- loss
$$^{2}$https://nhm.gov.in/index1.php?lang=1\&$level=28
sublinkid=1051&lid=606
Let's discuss
<FigureHere>
Figure 1: An example showing the translation of the
phrase“Let's discuss" in Indian Sign Language.
DGS ( DeutscheGebärdensprache) (Camgoz et al. 2018)。另一方面，专注于印度手语的工作量较少。主要原因是印度手语(ISL)的大型注释数据集不可用。ISL 作为印度庞大而多样化的人口的交流媒介，仍然面临着认证翻译人员的不足(印度只有 300 名认证手语口译员$^{3}$),这使得印度的口语和手语之间的差距更加突出。本文旨在通过管理印度手语的新翻译数据集来弥合这一差距：ISLTranslate,具有31，222 ISL-英语对
由于 ISL 的认证手语翻译人员较少，因此听力障碍社区缺乏教育材料。印度的许多政府和非政府组织最近开始通过在 ISL 中创建标准化的教育内容来弥合这一差距。创建的内容有助于为听力障碍的儿童建立基本词汇，并帮助人们使用口语学习和教授儿童 ISL。考虑到词汇的标准化表示和简单性，我们选择这些内容来管理 ISL-英语翻译数据集。我们选择特定于教育材料的内容，这些内容是标准化的，并在印度各地用于初等教育。因此，创建的内容中使用的词汇涵盖了不同的主题(例如。g。数学、科学、英语)使用常见的日常单词
ISL 是一种低资源语言，从训练机器学习模型的角度来看，用于通信的身体模态的存在使其更加需要资源。在手势级别注释手语(将相似的手势分组到不同的符号句子中)具有挑战性且不可扩展。此外，在过去，研究人员已经尝试将标志翻译成光泽表示和光泽到书面语言翻译(Sign2Gloss2Text (Camgoz et al.,2018))。光泽
The statistic is as per the Indian government organiza-
tion Indian Sign Language Research and Training Centre
(ISLRTC): http://islrtc.nic.in/
|
[
[
0.276123046875,
0.0955810546875,
0.72119140625,
0.0955810546875,
0.72119140625,
0.10986328125,
0.276123046875,
0.10986328125
],
[
0.2493896484375,
0.1395263671875,
0.38037109375,
0.1395263671875,
0.38037109375,
0.15380859375,
0.2493896484375,
0.15380859375
],
[
0.42236328125,
0.1395263671875,
0.57275390625,
0.1395263671875,
0.57275390625,
0.15380859375,
0.42236328125,
0.15380859375
],
[
0.611328125,
0.1395263671875,
0.75146484375,
0.1383056640625,
0.75146484375,
0.152587890625,
0.611328125,
0.15380859375
],
[
0.287109375,
0.155029296875,
0.71728515625,
0.15673828125,
0.71728515625,
0.171630859375,
0.287109375,
0.169189453125
],
[
0.26025390625,
0.171630859375,
0.744140625,
0.1728515625,
0.744140625,
0.18701171875,
0.26025390625,
0.1859130859375
],
[
0.189697265625,
0.189453125,
0.8251953125,
0.189453125,
0.8251953125,
0.2037353515625,
0.189697265625,
0.2037353515625
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.67919921875,
0.251708984375,
0.736328125,
0.251708984375,
0.736328125,
0.262451171875,
0.67919921875,
0.262451171875
],
[
0.145263671875,
0.281982421875,
0.459228515625,
0.281982421875,
0.459228515625,
0.295654296875,
0.145263671875,
0.295654296875
],
[
0.6455078125,
0.276611328125,
0.6884765625,
0.276611328125,
0.6884765625,
0.287353515625,
0.6455078125,
0.287353515625
],
[
0.145263671875,
0.29638671875,
0.459228515625,
0.29638671875,
0.459228515625,
0.31005859375,
0.145263671875,
0.31005859375
],
[
0.59619140625,
0.305908203125,
0.646484375,
0.305908203125,
0.646484375,
0.31640625,
0.59619140625,
0.31640625
],
[
0.145263671875,
0.312255859375,
0.45751953125,
0.312255859375,
0.45751953125,
0.322509765625,
0.145263671875,
0.322509765625
],
[
0.78271484375,
0.30712890625,
0.8193359375,
0.30712890625,
0.8193359375,
0.31640625,
0.78271484375,
0.31640625
],
[
0.1435546875,
0.32470703125,
0.459228515625,
0.32373046875,
0.459228515625,
0.33837890625,
0.1435546875,
0.339111328125
],
[
0.52880859375,
0.327880859375,
0.7119140625,
0.32373046875,
0.7138671875,
0.352783203125,
0.53076171875,
0.35693359375
],
[
0.755859375,
0.333251953125,
0.7900390625,
0.333251953125,
0.7900390625,
0.349853515625,
0.755859375,
0.349853515625
],
[
0.779296875,
0.33203125,
0.84228515625,
0.33203125,
0.84228515625,
0.350830078125,
0.779296875,
0.350830078125
],
[
0.145263671875,
0.339111328125,
0.459228515625,
0.339111328125,
0.459228515625,
0.352783203125,
0.145263671875,
0.352783203125
],
[
0.145263671875,
0.352783203125,
0.459228515625,
0.352783203125,
0.459228515625,
0.366455078125,
0.145263671875,
0.366455078125
],
[
0.145263671875,
0.36767578125,
0.459228515625,
0.36767578125,
0.459228515625,
0.38134765625,
0.145263671875,
0.38134765625
],
[
0.51318359375,
0.366455078125,
0.8818359375,
0.366455078125,
0.8818359375,
0.380126953125,
0.51318359375,
0.380126953125
],
[
0.1435546875,
0.380126953125,
0.4609375,
0.38134765625,
0.4609375,
0.3955078125,
0.1435546875,
0.394287109375
],
[
0.51123046875,
0.38134765625,
0.83642578125,
0.38134765625,
0.83642578125,
0.394287109375,
0.51123046875,
0.394287109375
],
[
0.1435546875,
0.393798828125,
0.459228515625,
0.3955078125,
0.459228515625,
0.410400390625,
0.1435546875,
0.407958984375
],
[
0.145263671875,
0.410400390625,
0.456787109375,
0.410400390625,
0.456787109375,
0.42333984375,
0.145263671875,
0.42333984375
],
[
0.51123046875,
0.40625,
0.87890625,
0.40673828125,
0.87890625,
0.421630859375,
0.51123046875,
0.42041015625
],
[
0.1435546875,
0.42333984375,
0.45751953125,
0.4228515625,
0.45751953125,
0.43701171875,
0.1435546875,
0.438232421875
],
[
0.51123046875,
0.42333984375,
0.880859375,
0.42333984375,
0.880859375,
0.43701171875,
0.51123046875,
0.43701171875
],
[
0.145263671875,
0.438232421875,
0.45751953125,
0.438232421875,
0.45751953125,
0.451904296875,
0.145263671875,
0.451904296875
],
[
0.51318359375,
0.439453125,
0.880859375,
0.439453125,
0.880859375,
0.453125,
0.51318359375,
0.453125
],
[
0.1427001953125,
0.451904296875,
0.45751953125,
0.45068359375,
0.45751953125,
0.465576171875,
0.1427001953125,
0.466064453125
],
[
0.51318359375,
0.4560546875,
0.880859375,
0.4560546875,
0.880859375,
0.4697265625,
0.51318359375,
0.4697265625
],
[
0.14697265625,
0.46728515625,
0.45751953125,
0.46728515625,
0.45751953125,
0.48095703125,
0.14697265625,
0.48095703125
],
[
0.5146484375,
0.471435546875,
0.880859375,
0.471435546875,
0.880859375,
0.485107421875,
0.5146484375,
0.485107421875
],
[
0.145263671875,
0.48095703125,
0.459228515625,
0.48095703125,
0.459228515625,
0.49462890625,
0.145263671875,
0.49462890625
],
[
0.51123046875,
0.486328125,
0.880859375,
0.48681640625,
0.880859375,
0.501953125,
0.51123046875,
0.50048828125
],
[
0.145263671875,
0.495361328125,
0.459228515625,
0.495361328125,
0.459228515625,
0.5087890625,
0.145263671875,
0.5087890625
],
[
0.51318359375,
0.501953125,
0.880859375,
0.50439453125,
0.880859375,
0.5185546875,
0.51318359375,
0.5166015625
],
[
0.14697265625,
0.5087890625,
0.459228515625,
0.5087890625,
0.459228515625,
0.5224609375,
0.14697265625,
0.5224609375
],
[
0.145263671875,
0.52392578125,
0.459228515625,
0.52392578125,
0.459228515625,
0.53759765625,
0.145263671875,
0.53759765625
],
[
0.51318359375,
0.51953125,
0.880859375,
0.51953125,
0.880859375,
0.533203125,
0.51318359375,
0.533203125
],
[
0.145263671875,
0.5380859375,
0.459228515625,
0.5380859375,
0.459228515625,
0.5517578125,
0.145263671875,
0.5517578125
],
[
0.51123046875,
0.53515625,
0.8818359375,
0.53515625,
0.8818359375,
0.548828125,
0.51123046875,
0.548828125
],
[
0.145263671875,
0.5517578125,
0.459228515625,
0.5517578125,
0.459228515625,
0.5654296875,
0.145263671875,
0.5654296875
],
[
0.509765625,
0.5498046875,
0.60205078125,
0.5517578125,
0.60205078125,
0.56640625,
0.509765625,
0.56396484375
],
[
0.145263671875,
0.56787109375,
0.45751953125,
0.56787109375,
0.45751953125,
0.57763671875,
0.145263671875,
0.57763671875
],
[
0.51318359375,
0.568359375,
0.880859375,
0.568359375,
0.880859375,
0.58203125,
0.51318359375,
0.58203125
],
[
0.14697265625,
0.5791015625,
0.19140625,
0.58056640625,
0.189697265625,
0.59326171875,
0.145263671875,
0.59130859375
],
[
0.51318359375,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.5986328125,
0.51318359375,
0.5986328125
],
[
0.51318359375,
0.599609375,
0.87890625,
0.599609375,
0.87890625,
0.61328125,
0.51318359375,
0.61328125
],
[
0.1175537109375,
0.60986328125,
0.1318359375,
0.60986328125,
0.1318359375,
0.62060546875,
0.1175537109375,
0.62060546875
],
[
0.1485595703125,
0.60791015625,
0.264404296875,
0.609375,
0.264404296875,
0.62353515625,
0.1485595703125,
0.62255859375
],
[
0.5146484375,
0.61767578125,
0.8759765625,
0.61767578125,
0.8759765625,
0.62744140625,
0.5146484375,
0.62744140625
],
[
0.1175537109375,
0.63427734375,
0.486083984375,
0.6328125,
0.486083984375,
0.64794921875,
0.1175537109375,
0.6484375
],
[
0.5146484375,
0.6328125,
0.880859375,
0.6328125,
0.880859375,
0.646484375,
0.5146484375,
0.646484375
],
[
0.118408203125,
0.6494140625,
0.486083984375,
0.6494140625,
0.486083984375,
0.66455078125,
0.118408203125,
0.66455078125
],
[
0.51123046875,
0.64794921875,
0.87890625,
0.64794921875,
0.87890625,
0.6611328125,
0.51123046875,
0.6611328125
],
[
0.118408203125,
0.66748046875,
0.486083984375,
0.66748046875,
0.486083984375,
0.68115234375,
0.118408203125,
0.68115234375
],
[
0.51123046875,
0.6630859375,
0.880859375,
0.66455078125,
0.880859375,
0.6787109375,
0.51123046875,
0.677734375
],
[
0.118408203125,
0.68310546875,
0.486083984375,
0.68310546875,
0.486083984375,
0.69677734375,
0.118408203125,
0.69677734375
],
[
0.51318359375,
0.68115234375,
0.87890625,
0.68115234375,
0.87890625,
0.69482421875,
0.51318359375,
0.69482421875
],
[
0.118408203125,
0.69873046875,
0.484375,
0.69873046875,
0.484375,
0.7119140625,
0.118408203125,
0.7119140625
],
[
0.51318359375,
0.69775390625,
0.880859375,
0.69775390625,
0.880859375,
0.71142578125,
0.51318359375,
0.71142578125
],
[
0.118408203125,
0.71533203125,
0.486083984375,
0.71533203125,
0.486083984375,
0.728515625,
0.118408203125,
0.728515625
],
[
0.51318359375,
0.71337890625,
0.87890625,
0.71337890625,
0.87890625,
0.72705078125,
0.51318359375,
0.72705078125
],
[
0.1175537109375,
0.72998046875,
0.486083984375,
0.72998046875,
0.486083984375,
0.74365234375,
0.1175537109375,
0.74365234375
],
[
0.51318359375,
0.72998046875,
0.8818359375,
0.72998046875,
0.8818359375,
0.74462890625,
0.51318359375,
0.74462890625
],
[
0.118408203125,
0.74755859375,
0.489501953125,
0.74755859375,
0.489501953125,
0.76123046875,
0.118408203125,
0.76123046875
],
[
0.51123046875,
0.7451171875,
0.76416015625,
0.7451171875,
0.76416015625,
0.7587890625,
0.51123046875,
0.7587890625
],
[
0.1158447265625,
0.7626953125,
0.48779296875,
0.7626953125,
0.48779296875,
0.775390625,
0.1158447265625,
0.775390625
],
[
0.51123046875,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.77490234375,
0.51123046875,
0.77490234375
],
[
0.1175537109375,
0.77978515625,
0.48779296875,
0.77978515625,
0.48779296875,
0.79345703125,
0.1175537109375,
0.79345703125
],
[
0.51318359375,
0.77783203125,
0.880859375,
0.77783203125,
0.880859375,
0.79150390625,
0.51318359375,
0.79150390625
],
[
0.1175537109375,
0.7958984375,
0.486083984375,
0.7958984375,
0.486083984375,
0.80859375,
0.1175537109375,
0.80859375
],
[
0.5146484375,
0.79345703125,
0.880859375,
0.79345703125,
0.880859375,
0.80712890625,
0.5146484375,
0.80712890625
],
[
0.118408203125,
0.81103515625,
0.486083984375,
0.81103515625,
0.486083984375,
0.826171875,
0.118408203125,
0.826171875
],
[
0.51318359375,
0.80859375,
0.8818359375,
0.80859375,
0.8818359375,
0.822265625,
0.51318359375,
0.822265625
],
[
0.118408203125,
0.82763671875,
0.48779296875,
0.82763671875,
0.48779296875,
0.84130859375,
0.118408203125,
0.84130859375
],
[
0.51318359375,
0.82373046875,
0.8818359375,
0.826171875,
0.8818359375,
0.84033203125,
0.51318359375,
0.83837890625
],
[
0.1175537109375,
0.84326171875,
0.486083984375,
0.84326171875,
0.486083984375,
0.85693359375,
0.1175537109375,
0.85693359375
],
[
0.5146484375,
0.84326171875,
0.87890625,
0.84326171875,
0.87890625,
0.85400390625,
0.5146484375,
0.85400390625
],
[
0.51123046875,
0.85791015625,
0.880859375,
0.85791015625,
0.880859375,
0.87255859375,
0.51123046875,
0.87255859375
],
[
0.139404296875,
0.8662109375,
0.46337890625,
0.86767578125,
0.46337890625,
0.8818359375,
0.139404296875,
0.88134765625
],
[
0.1158447265625,
0.87890625,
0.364501953125,
0.8798828125,
0.364501953125,
0.8935546875,
0.1158447265625,
0.892578125
],
[
0.5322265625,
0.88134765625,
0.8818359375,
0.88134765625,
0.8818359375,
0.8955078125,
0.5322265625,
0.8955078125
],
[
0.1368408203125,
0.89111328125,
0.48779296875,
0.8935546875,
0.48779296875,
0.90771484375,
0.1368408203125,
0.90625
],
[
0.51123046875,
0.8935546875,
0.8818359375,
0.8935546875,
0.8818359375,
0.90673828125,
0.51123046875,
0.90673828125
],
[
0.1158447265625,
0.90478515625,
0.287109375,
0.90625,
0.287109375,
0.91748046875,
0.1158447265625,
0.91650390625
],
[
0.51123046875,
0.90478515625,
0.744140625,
0.90478515625,
0.744140625,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"ISLTranslate：用于翻译印度手语的数据集",
"AbhinavJoshil",
"Susmit Agrawal?",
"AshutoshModi",
"1Indian Institute of Technology Kanpur (IIT Kanpur)",
"2Indian Institute of Technology Hyderabad (IIT Hyderabad)",
"(ajoshi,ashutoshm @cse.iitk.ac.in,ai22mtech12002@iith.ac.in",
"Abstract",
"Let'sdiscuss",
"Sign languages are the primary means of com-",
"let's",
"munication for many hard-of-hearing people",
"discuss",
"worldwide.Recently,to bridge the communi-",
"let's",
"cation gap between the hard-of-hearing com-",
"雪重重骨大小",
"大",
"",
"munity and the rest of the population, several",
"sign language translation datasets have been",
"proposed to enable the development of statis-",
"Figure 1: An example showing the translation of the",
"tical sign language translation systems. How-",
"phrase \"Let's discuss\" in Indian Sign Language.",
"ever, there is a dearth of sign language re-",
"sources for the Indian sign language. This",
"DGS (DeutscheGebardensprache) (Camgoz et al.,",
"resource paper introduces ISLTranslate,a",
"2018）。另一方面，专注于印度手语的工作量",
"translation dataset for continuous Indian Sign",
"较少。主要原因是印度手语（ISL）的大型注释",
"Language (ISL) consisting of 31k ISL-English",
"数据集不可用。ISL作为印度庞大而多样化的",
"sentence/phrase pairs.To the best of our",
"人口的交流媒介，仍然面临着认证翻译人员的",
"knowledge, it is the largest translation dataset",
"不足（印度只有300名认证手语口译员3），这",
"for continuous Indian Sign Language. We pro-",
"使得印度的口语和手语之间的差距更加突出。",
"vide a detailed analysis of the dataset. To val-",
"idate the performance of existing end-to-end",
"本文旨在通过管理印度手语的新翻译数据集",
"Sign language to spoken language translation",
"来弥合这一差距：ISLTranslate，具有31,222",
"systems, we benchmark the created dataset",
"ISL-英语对",
"withatransformer-basedmodelforISLtrans-",
"由于ISL的认证手语翻译人员较少，因此听力",
"lation.",
"障碍社区缺乏教育材料。印度的许多政府和非",
"政府组织最近开始通过在ISL中创建标准化的",
"1",
"Introduction",
"教育内容来弥合这一差距。创建的内容有助于",
"全球约有4.3亿听力障碍者1，其中6300万",
"为听力障碍的儿童建立基本词汇，并帮助人们",
"在印度？。手语是听力障碍社区的主要交流",
"使用口语学习和教授儿童ISL。考虑到词汇的",
"方式。尽管自然语言处理技术在过去五年中",
"标准化表示和简单性，我们选择这些内容来管",
"已显示出巨大的进步，但主要是由于注释资源",
"理ISL-英语翻译数据集。我们选择特定于教育",
"和大型语言模型（Tunstalletal.，2022）的可用",
"材料的内容，这些内容是标准化的，并在印度",
"性，像手语这样具有身体形态的语言仍然缺乏",
"各地用于初等教育。因此，创建的内容中使用",
"有效的语言处理系统。最近，手语研究已开始",
"的词汇涵盖了不同的主题（例如。g。、数学、科",
"在 NLP 社区 (Yin et al., 2021; Koller et al., 2015;",
"学、英语）使用常见的日常单词",
"Sincan and Keles. 2020: Xu et al.. 2022: Albanie",
"ISL是一种低资源语言，从训练机器学习模型",
"et al., 2020; Jiang et al., 2021; Moryossef et al.,",
"的角度来看，用于通信的身体模态的存在使其",
"2020；Joshietal.，2022）中引起关注。翻译数",
"更加需要资源。在手势级别注释手语（将相似",
"据集的可用性改善了手语的NLP系统的研",
"的手势分组到不同的符号句子中）具有挑战性",
"究和开发，例如ASL（美国手语）（Lietal.,",
"且不可扩展。此外，在过去，研究人员已经尝试",
"2020），BSL（英国手语）（Albanieetal.,2021)和",
"将标志翻译成光泽表示和光泽到书面语言翻译",
"(Sign2Gloss2Text (Camgoz et al.,2018))。光泽",
"'https://www.who.int/news-room/fact-sheets/",
"detail/deafness-and-hearing-loss",
"3The statistic is as per the Indian government organiza-",
"2https://nhm.gov.in/index1.php?lang=1&level=2&",
"tion Indian Sign Language Research and Training Centre",
"sublinkid=1051&lid=606",
"(ISLRTC):http://islrtc.nic.in/"
] |
{
"arxiv_id": "2307.05440",
"language": "zh",
"page_idx": 1
}
|
||
2307.05440-zh-2
|
2307.05440-zh-2.jpg
|
是赋予已签名手势的文本标签。数据集中符号句子的光泽标签的存在有助于翻译系统在符号翻译的粒度级别上工作。但是，为有符号的句子生成光泽表示是数据注释的另一个挑战。对于 ISLTranslate, 我们提出了端到端 ISL 到英语翻译的任务。图1 显示了来自 ISLTranslate 的 ISL 标志视频的示例。该示例显示了句子“Let'sdiscussion”的翻译，其中签名者通过将皱着眉头的手同时向上抬起手掌来传达“让我们。“与文本到文本的翻译任务相比，标志视频中存在的连续性使其更具挑战性，因为为运动构建标记化表示是一个具有挑战性的问题。总的来说，在这篇资源论文中，我们做出了以下贡献：
·We create a large ISL-English translation dataset with more than 31,2222 ISL-English pair sentences/phrases. The dataset covers a wide range of daily communication words with a vocabulary size of 11,655
We believe making this dataset available for the NLP community will facilitate future research in sign languages with a significant societal impact. Moreover, though not attempted in this paper, we hope that ISLTranslate could also be useful for sign language generation research. The dataset is made available at: https://github.com/ Exploration-Lab/ISLTranslate.
· We propose a baseline model for end-to-end ISL-English translation inspired by sign language transformer (Camgoz et al., 2020) .
2
## $\textbf{Related Work}$
与自然语言相比，手语使用身体形态，包括手的形状和位置，头部运动(如点头/摇晃),眼睛注视，手指拼写和面部表情。随着手，眼，头和面部表情的特征并行进行，与口语相比，它变得更加丰富，在口语中，连续的口语句子可以看作是发音单元的串联版本。此外，从语言和研究的角度来看，从 3 个维度的连续运动进行翻译使手语翻译更具挑战性和令人兴奋
手语翻译数据集：近年来提出了各种手语翻译数据集(Yin et al.,2021)。特别是对于美国手语(ASL),已经有一些关于创建数据集 (Martinez et al., 2002; Dreuw et al., 2007) 的早期工作，其中通过要求本地签名者对内容进行签名来在工作室中收集数据集。已经提出了其他数据集，用于中国手语(Zhou et al., 2021),韩国手语(Ko et al.,2018), 瑞士德语手语-DeutschweizerGebardensprache(DSGS)和佛兰芒手语-VlaamseGebarentaal (VGT) (Camgöz
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>Lang. </th>
<th>Sentences</th>
<th>Vocab. </th>
</tr>
<tr>
<td>PurdueRVL- SLLL (Martinez et al., 2002)</td>
<td>$ASL$</td>
<td>2.5k</td>
<td>104</td>
</tr>
<tr>
<td>Boston 104 (Dreuw et al., 2007)</td>
<td>$ASL$</td>
<td>201</td>
<td>103</td>
</tr>
<tr>
<td>How2Sign (Duarte et al., 2021)</td>
<td>$ASL$</td>
<td>$35k$</td>
<td>$16k$</td>
</tr>
<tr>
<td>OpenASL (Shi et al., 2022)</td>
<td>$ASL$</td>
<td>$98k$</td>
<td>$33k$</td>
</tr>
<tr>
<td>BOBSL (Albanie et al., 2021)</td>
<td>$BSL$</td>
<td>1.9k</td>
<td>78k</td>
</tr>
<tr>
<td>$\operatorname{CSL}\operatorname{Daily}$ (Zhou et al., 2021)</td>
<td>$CSL$</td>
<td>20.6k</td>
<td>$2\mathbf{k}$</td>
</tr>
<tr>
<td>${\mathrm{Phoenix-2014T}}$ ${\mathrm{amgoz~et~al.2018}}$</td>
<td>$DGS$</td>
<td>8.2</td>
<td>$3K$</td>
</tr>
<tr>
<td>150L vt M $\mathbf{1}\cup$ SWISSTXT-Weather ${\mathrm{amg\"{o}zetal.,2021)}}$ $\textbf{( 1) }$</td>
<td>$DSGS$</td>
<td>811</td>
<td>$1k$</td>
</tr>
<tr>
<td>TEOL VI -0س1/ SWISSTXT-News ${\mathrm{amg\"{o}zetal.2021}})$ ${\mathrm{T}}_{:}$</td>
<td>$DSGS$</td>
<td>$6k$</td>
<td>$10k$</td>
</tr>
<tr>
<td>$KETI$ (Ko et al...2018)</td>
<td>KSL</td>
<td>14.6k</td>
<td>419</td>
</tr>
<tr>
<td>$ur,wvio$ VRT- News (Camgöz et al., 2021)</td>
<td>$VGT$</td>
<td>$7.1k$</td>
<td>$7k$</td>
</tr>
<tr>
<td>ISL-CSLRT and Nataraian (Flakkiva 2021)</td>
<td>$ISL$</td>
<td>100</td>
<td>-</td>
</tr>
<tr>
<td>ISLTranslate(ours)</td>
<td>$ISL$</td>
<td>$31k$</td>
<td>$11k$</td>
</tr>
</tbody>
</table>
Table 1: Comparison of continuous sign language trans-
lation datasets.
et al.,2021)。在这项工作中，我们专门针对印度手语，并提出了一个包含 ISL 视频-英语翻译对的数据集
端到端手语翻译系统：大多数现有的手语翻译方法(Camgoz et al., 2018; De Coster and Dambre, 2022; De Coster et al., 2021) 依赖于翻译的中间光泽标签。由于光泽与视频片段对齐， 它们提供了精细的一对一映射，有助于在学习有效的视频表示时进行监督学习。以前的工作(Camgoz et al., 2018) 已报告在没有光泽标签的BLEU-4 分数中下降了约 10.0。然而，考虑到光泽级注释的注释成本，考虑无光泽的歌唱语言翻译方法变得势在必行。此外，连续手语中的光泽映射可能会从手语中删除语法方面。关于手语翻译的其他最新作品包括 Voskou et al. (2021); Yin and Read (2020), 它试图消除对训练的光泽序列的要求，并提出了一种基于转换器的端到端翻译体系结构。我们还遵循无光泽的 ISL 翻译方法
## 3 ISLTranslate
ISLTranslate 是从 ISLRTC 组织制作的公开教育视频创建的数据集，可通过 YouTube 获得。创建这些视频是为了向听力障碍的儿童提供学校水平的教育。视频涵盖了 ISL 中的NCERT$^{4}$标准化英语教育内容。由于这些视
${^{4}\mathrm{https://en.wikipedia.org/wiki/National}}_{-}$ Council\_of\_Educational\_Research\_and\_Training
|
[
[
0.12005615234375,
0.08612060546875,
0.484375,
0.08612060546875,
0.484375,
0.09979248046875,
0.12005615234375,
0.09979248046875
],
[
0.5283203125,
0.0885009765625,
0.57421875,
0.0885009765625,
0.57421875,
0.09979248046875,
0.5283203125,
0.09979248046875
],
[
0.71875,
0.0872802734375,
0.8671875,
0.0872802734375,
0.8671875,
0.1015625,
0.71875,
0.1015625
],
[
0.12005615234375,
0.10272216796875,
0.484375,
0.10272216796875,
0.484375,
0.11639404296875,
0.12005615234375,
0.11639404296875
],
[
0.5263671875,
0.1051025390625,
0.64404296875,
0.1051025390625,
0.64404296875,
0.1181640625,
0.5263671875,
0.1181640625
],
[
0.724609375,
0.11224365234375,
0.75732421875,
0.11224365234375,
0.75732421875,
0.12469482421875,
0.724609375,
0.12469482421875
],
[
0.77490234375,
0.11224365234375,
0.8076171875,
0.11224365234375,
0.8076171875,
0.12469482421875,
0.77490234375,
0.12469482421875
],
[
0.8330078125,
0.11224365234375,
0.861328125,
0.11224365234375,
0.861328125,
0.12469482421875,
0.8330078125,
0.12469482421875
],
[
0.118408203125,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.5263671875,
0.1175537109375,
0.6650390625,
0.1175537109375,
0.6650390625,
0.1312255859375,
0.5263671875,
0.1312255859375
],
[
0.12005615234375,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.5283203125,
0.1300048828125,
0.59765625,
0.1300048828125,
0.59765625,
0.141357421875,
0.5283203125,
0.141357421875
],
[
0.77685546875,
0.1353759765625,
0.8076171875,
0.1353759765625,
0.8076171875,
0.149658203125,
0.77685546875,
0.149658203125
],
[
0.8310546875,
0.1353759765625,
0.8623046875,
0.1353759765625,
0.8623046875,
0.149658203125,
0.8310546875,
0.149658203125
],
[
0.5283203125,
0.142578125,
0.6484375,
0.142578125,
0.6484375,
0.152587890625,
0.5283203125,
0.152587890625
],
[
0.72314453125,
0.135986328125,
0.755859375,
0.135986328125,
0.755859375,
0.1484375,
0.72314453125,
0.1484375
],
[
0.12005615234375,
0.15087890625,
0.486083984375,
0.15087890625,
0.486083984375,
0.1644287109375,
0.12005615234375,
0.1644287109375
],
[
0.5283203125,
0.155029296875,
0.591796875,
0.155029296875,
0.591796875,
0.1650390625,
0.5283203125,
0.1650390625
],
[
0.77734375,
0.1591796875,
0.80615234375,
0.1591796875,
0.80615234375,
0.1728515625,
0.77734375,
0.1728515625
],
[
0.82958984375,
0.157958984375,
0.8623046875,
0.157958984375,
0.8623046875,
0.174560546875,
0.82958984375,
0.174560546875
],
[
0.5263671875,
0.1644287109375,
0.6513671875,
0.1644287109375,
0.6513671875,
0.1776123046875,
0.5263671875,
0.1776123046875
],
[
0.724609375,
0.1602783203125,
0.75732421875,
0.1602783203125,
0.75732421875,
0.1728515625,
0.724609375,
0.1728515625
],
[
0.118408203125,
0.16748046875,
0.484375,
0.16748046875,
0.484375,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.5263671875,
0.1776123046875,
0.587890625,
0.1776123046875,
0.587890625,
0.189453125,
0.5263671875,
0.189453125
],
[
0.72314453125,
0.1817626953125,
0.75830078125,
0.1817626953125,
0.75830078125,
0.19775390625,
0.72314453125,
0.19775390625
],
[
0.77685546875,
0.182861328125,
0.8076171875,
0.182861328125,
0.8076171875,
0.19775390625,
0.77685546875,
0.19775390625
],
[
0.8330078125,
0.182861328125,
0.85986328125,
0.182861328125,
0.85986328125,
0.196533203125,
0.8330078125,
0.196533203125
],
[
0.5263671875,
0.188232421875,
0.62744140625,
0.188232421875,
0.62744140625,
0.1995849609375,
0.5263671875,
0.1995849609375
],
[
0.118408203125,
0.182861328125,
0.484375,
0.182861328125,
0.484375,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.1175537109375,
0.198974609375,
0.486083984375,
0.198974609375,
0.486083984375,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.5283203125,
0.20068359375,
0.5751953125,
0.20068359375,
0.5751953125,
0.2120361328125,
0.5283203125,
0.2120361328125
],
[
0.72119140625,
0.204833984375,
0.75732421875,
0.204833984375,
0.75732421875,
0.2203369140625,
0.72119140625,
0.2203369140625
],
[
0.77490234375,
0.204833984375,
0.8095703125,
0.204833984375,
0.8095703125,
0.2203369140625,
0.77490234375,
0.2203369140625
],
[
0.5263671875,
0.21142578125,
0.6572265625,
0.21142578125,
0.6572265625,
0.2244873046875,
0.5263671875,
0.2244873046875
],
[
0.8310546875,
0.207275390625,
0.861328125,
0.207275390625,
0.861328125,
0.2197265625,
0.8310546875,
0.2197265625
],
[
0.118408203125,
0.21435546875,
0.486083984375,
0.21435546875,
0.486083984375,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.5263671875,
0.2244873046875,
0.59375,
0.2244873046875,
0.59375,
0.236328125,
0.5263671875,
0.236328125
],
[
0.76904296875,
0.2298583984375,
0.8134765625,
0.2298583984375,
0.8134765625,
0.2452392578125,
0.76904296875,
0.2452392578125
],
[
0.83447265625,
0.2298583984375,
0.8564453125,
0.2298583984375,
0.8564453125,
0.24462890625,
0.83447265625,
0.24462890625
],
[
0.724609375,
0.23095703125,
0.755859375,
0.23095703125,
0.755859375,
0.243408203125,
0.724609375,
0.243408203125
],
[
0.1175537109375,
0.23095703125,
0.484375,
0.23095703125,
0.484375,
0.24462890625,
0.1175537109375,
0.24462890625
],
[
0.5263671875,
0.2340087890625,
0.64208984375,
0.2340087890625,
0.64208984375,
0.248779296875,
0.5263671875,
0.248779296875
],
[
0.118408203125,
0.2476806640625,
0.484375,
0.2476806640625,
0.484375,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.5283203125,
0.2476806640625,
0.62109375,
0.2476806640625,
0.62109375,
0.2587890625,
0.5283203125,
0.2587890625
],
[
0.7197265625,
0.251708984375,
0.759765625,
0.254150390625,
0.75732421875,
0.268310546875,
0.71875,
0.2666015625
],
[
0.77734375,
0.254638671875,
0.80615234375,
0.254638671875,
0.80615234375,
0.267333984375,
0.77734375,
0.267333984375
],
[
0.8330078125,
0.254150390625,
0.85791015625,
0.254150390625,
0.85791015625,
0.268310546875,
0.8330078125,
0.268310546875
],
[
0.1175537109375,
0.262451171875,
0.483642578125,
0.262451171875,
0.483642578125,
0.275634765625,
0.1175537109375,
0.275634765625
],
[
0.5263671875,
0.2587890625,
0.66015625,
0.2587890625,
0.66015625,
0.2724609375,
0.5263671875,
0.2724609375
],
[
0.5263671875,
0.270751953125,
0.654296875,
0.271484375,
0.654296875,
0.283203125,
0.5263671875,
0.281982421875
],
[
0.71728515625,
0.276611328125,
0.76171875,
0.276611328125,
0.76171875,
0.29150390625,
0.71728515625,
0.29150390625
],
[
0.77685546875,
0.27783203125,
0.8076171875,
0.27783203125,
0.8076171875,
0.292236328125,
0.77685546875,
0.292236328125
],
[
0.83642578125,
0.27783203125,
0.85791015625,
0.27783203125,
0.85791015625,
0.29150390625,
0.83642578125,
0.29150390625
],
[
0.5263671875,
0.283935546875,
0.66015625,
0.283203125,
0.66015625,
0.29443359375,
0.5263671875,
0.295654296875
],
[
0.5263671875,
0.29443359375,
0.640625,
0.29638671875,
0.640625,
0.30810546875,
0.5263671875,
0.305908203125
],
[
0.118408203125,
0.27978515625,
0.486083984375,
0.27978515625,
0.486083984375,
0.29345703125,
0.118408203125,
0.29345703125
],
[
0.12005615234375,
0.295654296875,
0.1822509765625,
0.295654296875,
0.1822509765625,
0.31005859375,
0.12005615234375,
0.31005859375
],
[
0.71728515625,
0.300537109375,
0.76171875,
0.300537109375,
0.76171875,
0.3154296875,
0.71728515625,
0.3154296875
],
[
0.78076171875,
0.3017578125,
0.8037109375,
0.3017578125,
0.8037109375,
0.3154296875,
0.78076171875,
0.3154296875
],
[
0.8310546875,
0.2998046875,
0.8623046875,
0.2998046875,
0.8623046875,
0.31640625,
0.8310546875,
0.31640625
],
[
0.1376953125,
0.312255859375,
0.486083984375,
0.312255859375,
0.486083984375,
0.32666015625,
0.1376953125,
0.32666015625
],
[
0.5263671875,
0.30712890625,
0.66162109375,
0.30712890625,
0.66162109375,
0.320556640625,
0.5263671875,
0.320556640625
],
[
0.5263671875,
0.319580078125,
0.56494140625,
0.319580078125,
0.56494140625,
0.330810546875,
0.5263671875,
0.330810546875
],
[
0.72314453125,
0.32373046875,
0.759765625,
0.32373046875,
0.759765625,
0.339111328125,
0.72314453125,
0.339111328125
],
[
0.771484375,
0.322509765625,
0.8134765625,
0.32470703125,
0.81201171875,
0.339111328125,
0.77099609375,
0.33740234375
],
[
0.8310546875,
0.32470703125,
0.85986328125,
0.32470703125,
0.85986328125,
0.33837890625,
0.8310546875,
0.33837890625
],
[
0.5283203125,
0.330810546875,
0.62744140625,
0.330810546875,
0.62744140625,
0.342529296875,
0.5283203125,
0.342529296875
],
[
0.5283203125,
0.341552734375,
0.59423828125,
0.34326171875,
0.59423828125,
0.355224609375,
0.5283203125,
0.352783203125
],
[
0.154541015625,
0.327880859375,
0.486083984375,
0.327880859375,
0.486083984375,
0.342529296875,
0.154541015625,
0.342529296875
],
[
0.15283203125,
0.345703125,
0.339111328125,
0.345703125,
0.339111328125,
0.359375,
0.15283203125,
0.359375
],
[
0.3359375,
0.347412109375,
0.3525390625,
0.347412109375,
0.3525390625,
0.355224609375,
0.3359375,
0.355224609375
],
[
0.345947265625,
0.344482421875,
0.489501953125,
0.344482421875,
0.489501953125,
0.358154296875,
0.345947265625,
0.358154296875
],
[
0.77490234375,
0.347412109375,
0.8095703125,
0.347412109375,
0.8095703125,
0.363525390625,
0.77490234375,
0.363525390625
],
[
0.83642578125,
0.347412109375,
0.8564453125,
0.347412109375,
0.8564453125,
0.3623046875,
0.83642578125,
0.3623046875
],
[
0.5263671875,
0.35400390625,
0.66162109375,
0.35400390625,
0.66162109375,
0.36767578125,
0.5263671875,
0.36767578125
],
[
0.72314453125,
0.3486328125,
0.75732421875,
0.3486328125,
0.75732421875,
0.361083984375,
0.72314453125,
0.361083984375
],
[
0.15283203125,
0.361083984375,
0.486083984375,
0.361083984375,
0.486083984375,
0.374755859375,
0.15283203125,
0.374755859375
],
[
0.15283203125,
0.377197265625,
0.48779296875,
0.377197265625,
0.48779296875,
0.39013671875,
0.15283203125,
0.39013671875
],
[
0.5283203125,
0.372314453125,
0.603515625,
0.372314453125,
0.603515625,
0.38427734375,
0.5283203125,
0.38427734375
],
[
0.7255859375,
0.37890625,
0.75146484375,
0.37890625,
0.75146484375,
0.391357421875,
0.7255859375,
0.391357421875
],
[
0.77734375,
0.37890625,
0.80615234375,
0.37890625,
0.80615234375,
0.392578125,
0.77734375,
0.392578125
],
[
0.8369140625,
0.385498046875,
0.84619140625,
0.38134765625,
0.8505859375,
0.385986328125,
0.8427734375,
0.39013671875
],
[
0.1695556640625,
0.391357421875,
0.48779296875,
0.392578125,
0.48779296875,
0.40673828125,
0.1695556640625,
0.40625
],
[
0.5263671875,
0.38427734375,
0.71875,
0.38427734375,
0.71875,
0.39794921875,
0.5263671875,
0.39794921875
],
[
0.5263671875,
0.396728515625,
0.66162109375,
0.396728515625,
0.66162109375,
0.410400390625,
0.5263671875,
0.410400390625
],
[
0.7255859375,
0.396728515625,
0.75390625,
0.396728515625,
0.75390625,
0.4091796875,
0.7255859375,
0.4091796875
],
[
0.77734375,
0.396728515625,
0.80615234375,
0.396728515625,
0.80615234375,
0.410400390625,
0.77734375,
0.410400390625
],
[
0.8330078125,
0.396728515625,
0.86376953125,
0.396728515625,
0.86376953125,
0.410888671875,
0.8330078125,
0.410888671875
],
[
0.15283203125,
0.4091796875,
0.489501953125,
0.4091796875,
0.489501953125,
0.42333984375,
0.15283203125,
0.42333984375
],
[
0.160400390625,
0.3984375,
0.172119140625,
0.3984375,
0.172119140625,
0.402099609375,
0.160400390625,
0.402099609375
],
[
0.15283203125,
0.42578125,
0.489501953125,
0.42578125,
0.489501953125,
0.439453125,
0.15283203125,
0.439453125
],
[
0.509765625,
0.42333984375,
0.8818359375,
0.42578125,
0.8818359375,
0.439453125,
0.509765625,
0.43701171875
],
[
0.154541015625,
0.443603515625,
0.483642578125,
0.443603515625,
0.483642578125,
0.45361328125,
0.154541015625,
0.45361328125
],
[
0.51318359375,
0.440673828125,
0.61376953125,
0.440673828125,
0.61376953125,
0.451904296875,
0.51318359375,
0.451904296875
],
[
0.15283203125,
0.457763671875,
0.48779296875,
0.457763671875,
0.48779296875,
0.471435546875,
0.15283203125,
0.471435546875
],
[
0.51123046875,
0.46435546875,
0.87744140625,
0.46435546875,
0.87744140625,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.15283203125,
0.473876953125,
0.486083984375,
0.473876953125,
0.486083984375,
0.48681640625,
0.15283203125,
0.48681640625
],
[
0.5146484375,
0.48095703125,
0.880859375,
0.48095703125,
0.880859375,
0.49462890625,
0.5146484375,
0.49462890625
],
[
0.15283203125,
0.490478515625,
0.486083984375,
0.490478515625,
0.486083984375,
0.50439453125,
0.15283203125,
0.50439453125
],
[
0.51123046875,
0.49462890625,
0.62451171875,
0.496337890625,
0.62451171875,
0.51123046875,
0.51123046875,
0.5087890625
],
[
0.15283203125,
0.505859375,
0.486083984375,
0.505859375,
0.486083984375,
0.51953125,
0.15283203125,
0.51953125
],
[
0.51318359375,
0.51318359375,
0.880859375,
0.51318359375,
0.880859375,
0.52685546875,
0.51318359375,
0.52685546875
],
[
0.15283203125,
0.521484375,
0.4248046875,
0.521484375,
0.4248046875,
0.53515625,
0.15283203125,
0.53515625
],
[
0.51318359375,
0.529296875,
0.880859375,
0.529296875,
0.880859375,
0.5419921875,
0.51318359375,
0.5419921875
],
[
0.1376953125,
0.53759765625,
0.48779296875,
0.53759765625,
0.48779296875,
0.5517578125,
0.1376953125,
0.5517578125
],
[
0.51318359375,
0.54443359375,
0.880859375,
0.54443359375,
0.880859375,
0.55810546875,
0.51318359375,
0.55810546875
],
[
0.15283203125,
0.55419921875,
0.48779296875,
0.55419921875,
0.48779296875,
0.56787109375,
0.15283203125,
0.56787109375
],
[
0.51318359375,
0.56005859375,
0.8759765625,
0.56005859375,
0.8759765625,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.151123046875,
0.57080078125,
0.46337890625,
0.5693359375,
0.46337890625,
0.58447265625,
0.151123046875,
0.5849609375
],
[
0.5146484375,
0.57763671875,
0.880859375,
0.57763671875,
0.880859375,
0.59130859375,
0.5146484375,
0.59130859375
],
[
0.51318359375,
0.5927734375,
0.87890625,
0.5927734375,
0.87890625,
0.60546875,
0.51318359375,
0.60546875
],
[
0.1158447265625,
0.6015625,
0.145263671875,
0.6015625,
0.145263671875,
0.6142578125,
0.1158447265625,
0.6142578125
],
[
0.151123046875,
0.6015625,
0.274658203125,
0.6015625,
0.274658203125,
0.615234375,
0.151123046875,
0.615234375
],
[
0.51318359375,
0.60986328125,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.118408203125,
0.6298828125,
0.486083984375,
0.6298828125,
0.486083984375,
0.6435546875,
0.118408203125,
0.6435546875
],
[
0.51318359375,
0.6259765625,
0.880859375,
0.6259765625,
0.880859375,
0.6396484375,
0.51318359375,
0.6396484375
],
[
0.1175537109375,
0.6455078125,
0.484375,
0.6455078125,
0.484375,
0.6591796875,
0.1175537109375,
0.6591796875
],
[
0.51318359375,
0.642578125,
0.87744140625,
0.642578125,
0.87744140625,
0.65283203125,
0.51318359375,
0.65283203125
],
[
0.1175537109375,
0.6611328125,
0.484375,
0.6611328125,
0.484375,
0.6748046875,
0.1175537109375,
0.6748046875
],
[
0.5146484375,
0.65771484375,
0.87890625,
0.65771484375,
0.87890625,
0.67138671875,
0.5146484375,
0.67138671875
],
[
0.1175537109375,
0.677734375,
0.486083984375,
0.6767578125,
0.486083984375,
0.69140625,
0.1175537109375,
0.6923828125
],
[
0.509765625,
0.6728515625,
0.87890625,
0.6728515625,
0.87890625,
0.68603515625,
0.509765625,
0.68603515625
],
[
0.118408203125,
0.69482421875,
0.483642578125,
0.69482421875,
0.483642578125,
0.7080078125,
0.118408203125,
0.7080078125
],
[
0.5146484375,
0.6904296875,
0.87890625,
0.6904296875,
0.87890625,
0.70361328125,
0.5146484375,
0.70361328125
],
[
0.1175537109375,
0.708984375,
0.48193359375,
0.708984375,
0.48193359375,
0.72265625,
0.1175537109375,
0.72265625
],
[
0.51123046875,
0.7060546875,
0.880859375,
0.7060546875,
0.880859375,
0.7197265625,
0.51123046875,
0.7197265625
],
[
0.51123046875,
0.72021484375,
0.880859375,
0.7216796875,
0.880859375,
0.736328125,
0.51123046875,
0.7353515625
],
[
0.118408203125,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.118408203125,
0.7421875,
0.44921875,
0.7421875,
0.44921875,
0.755859375,
0.118408203125,
0.755859375
],
[
0.51123046875,
0.73681640625,
0.880859375,
0.73828125,
0.880859375,
0.7529296875,
0.51123046875,
0.751953125
],
[
0.509765625,
0.75439453125,
0.64404296875,
0.7548828125,
0.64404296875,
0.76953125,
0.509765625,
0.7685546875
],
[
0.12005615234375,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.12005615234375,
0.77392578125
],
[
0.118408203125,
0.775390625,
0.486083984375,
0.775390625,
0.486083984375,
0.7890625,
0.118408203125,
0.7890625
],
[
0.509765625,
0.783203125,
0.5498046875,
0.783203125,
0.5498046875,
0.79345703125,
0.509765625,
0.79345703125
],
[
0.54248046875,
0.7822265625,
0.6708984375,
0.7822265625,
0.6708984375,
0.7958984375,
0.54248046875,
0.7958984375
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.1175537109375,
0.80810546875,
0.484375,
0.80810546875,
0.484375,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.51123046875,
0.80712890625,
0.880859375,
0.80712890625,
0.880859375,
0.82080078125,
0.51123046875,
0.82080078125
],
[
0.1175537109375,
0.822265625,
0.486083984375,
0.82373046875,
0.486083984375,
0.83837890625,
0.1175537109375,
0.83740234375
],
[
0.51318359375,
0.822265625,
0.8818359375,
0.822265625,
0.8818359375,
0.8359375,
0.51318359375,
0.8359375
],
[
0.1175537109375,
0.84033203125,
0.483642578125,
0.84033203125,
0.483642578125,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.51318359375,
0.8388671875,
0.880859375,
0.8388671875,
0.880859375,
0.8525390625,
0.51318359375,
0.8525390625
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51318359375,
0.85498046875,
0.880859375,
0.85498046875,
0.880859375,
0.8681640625,
0.51318359375,
0.8681640625
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.509765625,
0.87060546875,
0.880859375,
0.87060546875,
0.880859375,
0.88525390625,
0.509765625,
0.88525390625
],
[
0.118408203125,
0.88818359375,
0.48779296875,
0.88818359375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.5322265625,
0.89111328125,
0.82958984375,
0.892578125,
0.82958984375,
0.90673828125,
0.5322265625,
0.90625
],
[
0.1175537109375,
0.90283203125,
0.48779296875,
0.90478515625,
0.48779296875,
0.91943359375,
0.1175537109375,
0.91748046875
],
[
0.509765625,
0.90380859375,
0.84619140625,
0.90625,
0.84619140625,
0.91943359375,
0.509765625,
0.91748046875
]
] |
[
"是赋予已签名手势的文本标签。数据集中符号",
"Dataset",
"Lang. Sentences Vocab.",
"句子的光泽标签的存在有助于翻译系统在符号",
"Purdue RVL-SLLL",
"ASL",
"2.5k",
"104",
"翻译的粒度级别上工作。但是，为有符号的句",
"(Martinez et al., 2002)",
"子生成光泽表示是数据注释的另一个挑战。对",
"Boston104",
"201",
"103",
"(Dreuw et al., 2007)",
"ASL",
"于ISLTranslate，我们提出了端到端ISL到英",
"How2Sign",
"35k",
"16k",
"(Duarte et al., 2021)",
"ASL",
"语翻译的任务。图1显示了来自ISLTranslate",
"OpenASL",
"ASL",
"98k",
"33k",
"(Shi et al., 2022)",
"的ISL标志视频的示例。该示例显示了句子",
"“Let'sdiscussion”的翻译，其中签名者通过将",
"BOBSL",
"BSL",
"1.9k",
"(Albanie et al., 2021)",
"78k",
"皱着眉头的手同时向上抬起手掌来传达“让我",
"CSL Daily",
"20.6k",
"2k",
"CSL",
"们。“与文本到文本的翻译任务相比，标志视",
"(Zhou et al., 2021)",
"频中存在的连续性使其更具挑战性，因为为运",
"Phoenix-2014T",
"DGS",
"8.2",
"3K",
"动构建标记化表示是一个具有挑战性的问题。",
"(Camgoz et al., 2018)",
"SWISSTXT-Weather",
"DSGS",
"811",
"1k",
"(Camgoz et al.,2021)",
"SWISSTXT-News",
"总的来说，在这篇资源论文中，我们做出了以",
"下贡献：",
"DSGS",
"6k",
"10k",
". We create a large ISL-English translation",
"(Camgoz et al., 2021)",
"KETI",
"KSL",
"14.6k",
"419",
"(Ko et al., 2018)",
"VRT-NewS",
"dataset with more than 31, 222 ISL-English",
"pair sentences/phrases.",
"",
"The dataset cov-",
"7.1k",
"7k",
"(Camgoz et al., 2021)",
"VGT",
"ers a wide range of daily communication",
"words with a vocabulary size of 11,655",
"ISL-CSLRT",
"ISL",
"100",
".",
"We believe making this dataset available",
"(Elakkiya and Natarajan, 2021)",
"ISLTranslate (ours)",
"ISL",
"31k",
"11k",
"for the NLP community will facilitate fu-",
"",
"ture research in sign languages with a sig-",
"Table I: Comparison of continuous sign language trans-",
"nificant societalimpact.Moreover,though",
"lation datasets.",
"not attempted in this paper, we hope that",
"etal.,2021）。在这项工作中，我们专门针对印",
"ISLTranslate could also be useful for sign",
"度手语，并提出了一个包含ISL视频-英语翻",
"language generation research. The dataset",
"译对的数据集",
"is made available at: https://github.com/",
"端到端手语翻译系统：大多数现有的手语",
"Exploration-Lab/ISLTranslate.",
"翻译方法（Camgozetal.,2018;DeCosterand",
": We propose a baseline model for end-to-end",
"Dambre,2022;De Coster et al.,2021）依赖于翻",
"ISL-English translation inspired by sign lan-",
"译的中间光泽标签。由于光泽与视频片段对齐",
"guage transformer (Camgoz et al., 2020) .",
"它们提供了精细的一对一映射，有助于在学习",
"有效的视频表示时进行监督学习。以前的工作",
"2",
"Related Work",
"(Camgozetal.，2018）已报告在没有光泽标签的",
"与自然语言相比，手语使用身体形态，包括手",
"BLEU-4分数中下降了约10.0。然而，考虑到",
"的形状和位置，头部运动（如点头/摇晃），眼睛",
"光泽级注释的注释成本，考虑无光泽的歌唱语",
"注视，手指拼写和面部表情。随着手，眼，头",
"言翻译方法变得势在必行。此外，连续手语中",
"和面部表情的特征并行进行，与口语相比，它",
"的光泽映射可能会从手语中删除语法方面。关",
"变得更加丰富，在口语中，连续的口语句子可",
"于手语翻译的其他最新作品包括Voskouetal.",
"以看作是发音单元的串联版本。此外，从语言",
"(2021)；YinandRead（2020），它试图消除对训",
"练的光泽序列的要求，并提出了一种基于转换",
"和研究的角度来看，从3个维度的连续运动进",
"行翻译使手语翻译更具挑战性和令人兴奋",
"器的端到端翻译体系结构。我们还遵循无光泽",
"的ISL翻译方法",
"手语翻译数据集：近年来提出了各种手语翻",
"译数据集（Yinetal.，2021）。特别是对于美",
"3",
" ISLTranslate",
"国手语（ASL），已经有一些关于创建数据",
"集(Martinez et al.,2002;Dreuw et al.,2007)的",
"ISLTranslate是从ISLRTC组织制作的公开",
"早期工作，其中通过要求本地签名者对内容",
"教育视频创建的数据集，可通过YouTube获",
"进行签名来在工作室中收集数据集。已经提",
"得。创建这些视频是为了向听力障碍的儿童",
"出了其他数据集，用于中国手语(Zhouetal.,",
"提供学校水平的教育。视频涵盖了ISL中的",
"2021），韩国手语（Koetal.,2018），瑞士德语手",
"NCERT4标准化英语教育内容。由于这些视",
"语-DeutschweizerGebardensprache（DSGS）和佛",
"thttps://en.wikipedia.org/wiki/National.",
"兰芒手语-VlaamseGebarentaal（VGT）（Camgoz",
"Councii_of_Educational_Research_and_Training"
] |
{
"arxiv_id": "2307.05440",
"language": "zh",
"page_idx": 2
}
|
||
2307.05440-zh-4
|
2307.05440-zh-4.jpg
|
于生成的文本与书中章节的文本相同，因此验证生成的样本很容易，并且可以通过手动将它们与教科书进行匹配来完成。总的来说，我们发现自动转录的文本质量很高；然而，错误生成的文本是在书中内容的帮助下手动纠正的
图 2显示了一个长句子的示例(来自ISLTranslate)及其在 ISL 中的翻译。图中的框架被分组为英语单词，并描绘了ISL 中的连续通信。注意句子中类似的单词，“符号/标志”和“语言”。“(另请参阅手语‘‘的视觉表示)。与其他自然语言一样，需要不同长度的表示(字符/手势)来传达不同的单词。在ISLTranslate 中，我们仅限于句子$\prime$短语级别的翻译。数据集分为训练、验证和测试拆分(应用程序中的详细信息。B)。App。图3显示了各种拆分中样本数的分布
与其他连续手语数据集的比较：我们主要与基于视频的数据集进行比较，该数据集包含成对的连续签名视频以及表 1 中相应的书面语言翻译。据我们所知，我们为 ISL 提出了最大的数据集
数据清理和预处理：视频 (e。g。,应用程序。$\operatorname{Fig。4}$)包含图片对应的书页。我们通过将人脸位置视为参考点并删除视频中剩余的背景， 将签名者从视频中裁剪出来
ISLTranslate 中的噪声消除：由于ISLTranslate 由使用可用音频信号中的暂停从较长的视频中剪辑的视频组成，因此数据集中的噪声可能会以多种方式蔓延。在翻译音频中的文本时，签名者可能会使用不同的符号， 这些符号可能不是相应英语句子的逐字翻译。此外，尽管背景中的音频与视频中的相应符号对齐，但在某些情况下，与相应的符号表示相比，音频可能会很快，并且可能会在句子的开头或结尾遗漏几个单词。我们还发现了一些情况，在叙述故事时，音频中的人通过修改语音来扮演角色，使其听起来像指定的角色在说句子。例如，在一个老鼠正在说话的故事中，说话者可能会改变声音并增加音调以模拟老鼠所说的对话，而不是说出“老鼠说”语句后面的句子。相比之下，在手语视频中，一个人在将句子翻译成 ISL 时可能会也可能不会扮演鼠标的角色
ISLTranslate 验证：为了验证数据集中存在的句子/短语 ISL-英语对的可靠性，我们在经过认证的 ISL 签名者的帮助下。由于认证的 ISL 签名者的可用性有限，我们只能使用一个小的随机选择的符号文本对样本(291 对)进行人工翻
$^{6}$https://www.youtube.com/watch?v=SInKhy-06qA
段。此外，这些片段通过 SOTA 语音到文本模 译和验证。我们请 ISL 讲师翻译视频 (应用程型 (Radford et al., 2022) 来生成相应的文本。由 序中的详细信息。D)。签名者为每个视频提供
一个参考翻译。表2 显示了 ISL 讲师创建的句子示例。为了定量估计数据集中翻译的可靠性， 我们将数据集中的英语翻译文本与 ISL 讲师提供的文本进行比较。表3 显示了 ISLTranslate 中 291 句子的翻译分数。总体而言，BLEU-4分数为 48.94 , ROUGE-L (Lin, 2004) 为 60.44 $, \textbf{WER( 字 错 误 率 ) 为 61. 88。为 了 提 供 比 较 }$ 的参考，对于文本到文本的翻译，人类翻译的 BLEU 得分在 30-50 之间 (如 Papineni et al. (2002)所报道，在来自 40 个一般新闻故事的约 500 个句子的测试语料库中，人类翻译者得分为 34。68 反对四个参考文献)。我们推测， 与 BLEU 得分为 48 的 ISLTranslate 中的翻译相比，具有很高的可靠性。93 与经认证的 ISL 签名者提供的参考翻译进行了比较。理想情况下，最好为视频中的同一签名句子提供多个参考翻译；然而，高注释工作以及认证的 ISL 签名者的较低可用性使其成为一项具有挑战性的任务
# 4 Baseline Results
给定一个句子的符号视频，手语翻译的任务是将其翻译成口语句子(在我们的例子中为英语)。为了对 ISLTranslate 进行基准测试， 我们为 ISL 到英语翻译创建了一个基准体系结构。我们提出了一种受手语转换器(SLT) (Camgoz et al., 2020) 启发的 ISL 姿势到英语翻译基线 (称为 Pose-SLT)。手语转换器将图像功能与转换器一起用于从签名视频生成文本翻译。然而，考虑到姿态估计模型 (Selvaraj et al., 2022)的更快的实时推断，我们使用姿态而不是图像作为输入。我们使用 Mediapipe 姿态估计管道$^{7}$。Saunders et al.(2020)使用了类似的基于 SLT 的姿势到文本方法，该方法提出了用于端到端手语制作的渐进式转换器，并使用基于 SLT 的姿势到文本来通过反向翻译验证生成的关键点(生成的姿势关键点到文本翻译)。虽然基于姿势的方法处理速度更快，但它们的性能通常不如基于图像的方法。对于整体关键点的选择，我们遵循 Selvaraj et al.(2022), 它返回 75 个关键点的 3D 坐标 (不包括面网格)。此外，我们通过将肩部关键点的中点置于中心并使用鼻子关键点和肩部中点之间的距离缩放关键点来归一化每个帧的关键点。我们使用标准的 BLEU 和 ROUGE 分数来评估获得的英语翻译(App 中的模型超参数详细信息。E)。表4 显示了针对建议的体系结构获得的结果。糟
$^7$https://ai.googleblog.com/2020/12/
mediapipe-holistic-simultaneous-face.html
|
[
[
0.12005615234375,
0.0872802734375,
0.48779296875,
0.0872802734375,
0.48779296875,
0.1009521484375,
0.12005615234375,
0.1009521484375
],
[
0.51123046875,
0.08612060546875,
0.87890625,
0.08612060546875,
0.87890625,
0.09979248046875,
0.51123046875,
0.09979248046875
],
[
0.118408203125,
0.10272216796875,
0.484375,
0.10272216796875,
0.484375,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.12176513671875,
0.12054443359375,
0.483642578125,
0.12054443359375,
0.483642578125,
0.1312255859375,
0.12176513671875,
0.1312255859375
],
[
0.5322265625,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.5322265625,
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
],
[
0.118408203125,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.12005615234375,
0.166259765625,
0.484375,
0.166259765625,
0.484375,
0.179931640625,
0.12005615234375,
0.179931640625
],
[
0.51318359375,
0.16748046875,
0.87890625,
0.16748046875,
0.87890625,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.12005615234375,
0.182861328125,
0.466796875,
0.182861328125,
0.466796875,
0.196533203125,
0.12005615234375,
0.196533203125
],
[
0.5146484375,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.5146484375,
0.196533203125
],
[
0.118408203125,
0.20068359375,
0.484375,
0.20068359375,
0.484375,
0.21435546875,
0.118408203125,
0.21435546875
],
[
0.509765625,
0.198974609375,
0.87890625,
0.198974609375,
0.87890625,
0.2120361328125,
0.509765625,
0.2120361328125
],
[
0.1175537109375,
0.2161865234375,
0.483642578125,
0.2161865234375,
0.483642578125,
0.2298583984375,
0.1175537109375,
0.2298583984375
],
[
0.52880859375,
0.215576171875,
0.880859375,
0.215576171875,
0.880859375,
0.2286376953125,
0.52880859375,
0.2286376953125
],
[
0.118408203125,
0.2327880859375,
0.48779296875,
0.2327880859375,
0.48779296875,
0.2464599609375,
0.118408203125,
0.2464599609375
],
[
0.509765625,
0.232177734375,
0.87890625,
0.232177734375,
0.87890625,
0.2452392578125,
0.509765625,
0.2452392578125
],
[
0.118408203125,
0.248779296875,
0.48779296875,
0.248779296875,
0.48779296875,
0.262451171875,
0.118408203125,
0.262451171875
],
[
0.509765625,
0.2476806640625,
0.87890625,
0.2476806640625,
0.87890625,
0.26123046875,
0.509765625,
0.26123046875
],
[
0.1175537109375,
0.26318359375,
0.486083984375,
0.26318359375,
0.486083984375,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.51123046875,
0.26416015625,
0.880859375,
0.26416015625,
0.880859375,
0.27783203125,
0.51123046875,
0.27783203125
],
[
0.118408203125,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.51123046875,
0.27978515625,
0.880859375,
0.27978515625,
0.880859375,
0.29345703125,
0.51123046875,
0.29345703125
],
[
0.118408203125,
0.29638671875,
0.48779296875,
0.29638671875,
0.48779296875,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.51123046875,
0.295654296875,
0.8818359375,
0.295654296875,
0.8818359375,
0.308837890625,
0.51123046875,
0.308837890625
],
[
0.118408203125,
0.312255859375,
0.48779296875,
0.312255859375,
0.48779296875,
0.325927734375,
0.118408203125,
0.325927734375
],
[
0.51318359375,
0.312255859375,
0.880859375,
0.312255859375,
0.880859375,
0.325927734375,
0.51318359375,
0.325927734375
],
[
0.12005615234375,
0.330078125,
0.483642578125,
0.330078125,
0.483642578125,
0.34033203125,
0.12005615234375,
0.34033203125
],
[
0.51318359375,
0.327880859375,
0.880859375,
0.327880859375,
0.880859375,
0.341552734375,
0.51318359375,
0.341552734375
],
[
0.118408203125,
0.345703125,
0.48779296875,
0.345703125,
0.48779296875,
0.35986328125,
0.118408203125,
0.35986328125
],
[
0.5146484375,
0.345703125,
0.87890625,
0.345703125,
0.87890625,
0.355712890625,
0.5146484375,
0.355712890625
],
[
0.12005615234375,
0.361083984375,
0.339111328125,
0.361083984375,
0.339111328125,
0.374755859375,
0.12005615234375,
0.374755859375
],
[
0.515625,
0.35986328125,
0.880859375,
0.35986328125,
0.880859375,
0.37353515625,
0.515625,
0.37353515625
],
[
0.118408203125,
0.377685546875,
0.486083984375,
0.377685546875,
0.486083984375,
0.391357421875,
0.118408203125,
0.391357421875
],
[
0.51123046875,
0.377197265625,
0.880859375,
0.377197265625,
0.880859375,
0.39013671875,
0.51123046875,
0.39013671875
],
[
0.12176513671875,
0.3955078125,
0.483642578125,
0.3955078125,
0.483642578125,
0.40625,
0.12176513671875,
0.40625
],
[
0.51318359375,
0.393798828125,
0.880859375,
0.393798828125,
0.880859375,
0.40673828125,
0.51318359375,
0.40673828125
],
[
0.1175537109375,
0.4091796875,
0.484375,
0.410400390625,
0.484375,
0.42333984375,
0.1175537109375,
0.4228515625
],
[
0.51123046875,
0.410400390625,
0.54833984375,
0.410400390625,
0.54833984375,
0.4228515625,
0.51123046875,
0.4228515625
],
[
0.118408203125,
0.427001953125,
0.484375,
0.427001953125,
0.484375,
0.440673828125,
0.118408203125,
0.440673828125
],
[
0.118408203125,
0.443603515625,
0.154541015625,
0.443603515625,
0.154541015625,
0.4560546875,
0.118408203125,
0.4560546875
],
[
0.51318359375,
0.438232421875,
0.6884765625,
0.438232421875,
0.6884765625,
0.451904296875,
0.51318359375,
0.451904296875
],
[
0.1175537109375,
0.458984375,
0.489501953125,
0.457763671875,
0.489501953125,
0.47265625,
0.1175537109375,
0.473876953125
],
[
0.51318359375,
0.466064453125,
0.87890625,
0.466064453125,
0.87890625,
0.479736328125,
0.51318359375,
0.479736328125
],
[
0.118408203125,
0.4755859375,
0.48779296875,
0.4755859375,
0.48779296875,
0.4892578125,
0.118408203125,
0.4892578125
],
[
0.51318359375,
0.482177734375,
0.880859375,
0.482177734375,
0.880859375,
0.495361328125,
0.51318359375,
0.495361328125
],
[
0.118408203125,
0.4912109375,
0.490234375,
0.4912109375,
0.490234375,
0.5048828125,
0.118408203125,
0.5048828125
],
[
0.51318359375,
0.49755859375,
0.880859375,
0.49755859375,
0.880859375,
0.51123046875,
0.51318359375,
0.51123046875
],
[
0.12005615234375,
0.5087890625,
0.337646484375,
0.5087890625,
0.337646484375,
0.51953125,
0.12005615234375,
0.51953125
],
[
0.5146484375,
0.51416015625,
0.880859375,
0.51416015625,
0.880859375,
0.52783203125,
0.5146484375,
0.52783203125
],
[
0.1368408203125,
0.52490234375,
0.486083984375,
0.52490234375,
0.486083984375,
0.5380859375,
0.1368408203125,
0.5380859375
],
[
0.51123046875,
0.53076171875,
0.8818359375,
0.53076171875,
0.8818359375,
0.54443359375,
0.51123046875,
0.54443359375
],
[
0.1175537109375,
0.5390625,
0.486083984375,
0.5390625,
0.486083984375,
0.552734375,
0.1175537109375,
0.552734375
],
[
0.509765625,
0.5458984375,
0.87890625,
0.5458984375,
0.87890625,
0.55859375,
0.509765625,
0.55859375
],
[
0.118408203125,
0.55712890625,
0.486083984375,
0.55712890625,
0.486083984375,
0.57080078125,
0.118408203125,
0.57080078125
],
[
0.51123046875,
0.56103515625,
0.87890625,
0.56103515625,
0.87890625,
0.57470703125,
0.51123046875,
0.57470703125
],
[
0.12005615234375,
0.57177734375,
0.486083984375,
0.57177734375,
0.486083984375,
0.5849609375,
0.12005615234375,
0.5849609375
],
[
0.51318359375,
0.5791015625,
0.87744140625,
0.5791015625,
0.87744140625,
0.5888671875,
0.51318359375,
0.5888671875
],
[
0.118408203125,
0.58740234375,
0.489501953125,
0.58837890625,
0.489501953125,
0.6025390625,
0.118408203125,
0.6015625
],
[
0.51318359375,
0.59423828125,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.51318359375,
0.60791015625
],
[
0.118408203125,
0.60498046875,
0.48779296875,
0.60498046875,
0.48779296875,
0.6181640625,
0.118408203125,
0.6181640625
],
[
0.5146484375,
0.6123046875,
0.87890625,
0.6123046875,
0.87890625,
0.62255859375,
0.5146484375,
0.62255859375
],
[
0.12005615234375,
0.62060546875,
0.484375,
0.62060546875,
0.484375,
0.63427734375,
0.12005615234375,
0.63427734375
],
[
0.51123046875,
0.6259765625,
0.87890625,
0.6259765625,
0.87890625,
0.6396484375,
0.51123046875,
0.6396484375
],
[
0.12005615234375,
0.63720703125,
0.48193359375,
0.63720703125,
0.48193359375,
0.64794921875,
0.12005615234375,
0.64794921875
],
[
0.51318359375,
0.6435546875,
0.880859375,
0.6435546875,
0.880859375,
0.65673828125,
0.51318359375,
0.65673828125
],
[
0.12005615234375,
0.65380859375,
0.484375,
0.65380859375,
0.484375,
0.66748046875,
0.12005615234375,
0.66748046875
],
[
0.51123046875,
0.65771484375,
0.87744140625,
0.65771484375,
0.87744140625,
0.67138671875,
0.51123046875,
0.67138671875
],
[
0.1175537109375,
0.66845703125,
0.483642578125,
0.66845703125,
0.483642578125,
0.681640625,
0.1175537109375,
0.681640625
],
[
0.51318359375,
0.6748046875,
0.880859375,
0.6748046875,
0.880859375,
0.6884765625,
0.51318359375,
0.6884765625
],
[
0.12005615234375,
0.68701171875,
0.483642578125,
0.68701171875,
0.483642578125,
0.69775390625,
0.12005615234375,
0.69775390625
],
[
0.51318359375,
0.69140625,
0.87890625,
0.69140625,
0.87890625,
0.705078125,
0.51318359375,
0.705078125
],
[
0.118408203125,
0.70166015625,
0.484375,
0.70166015625,
0.484375,
0.71533203125,
0.118408203125,
0.71533203125
],
[
0.51123046875,
0.70703125,
0.87890625,
0.70703125,
0.87890625,
0.72021484375,
0.51123046875,
0.72021484375
],
[
0.12005615234375,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.12005615234375,
0.732421875
],
[
0.51123046875,
0.72265625,
0.87890625,
0.72265625,
0.87890625,
0.736328125,
0.51123046875,
0.736328125
],
[
0.118408203125,
0.73388671875,
0.486083984375,
0.73388671875,
0.486083984375,
0.74755859375,
0.118408203125,
0.74755859375
],
[
0.5146484375,
0.7412109375,
0.87890625,
0.7412109375,
0.87890625,
0.751953125,
0.5146484375,
0.751953125
],
[
0.12005615234375,
0.75,
0.484375,
0.75,
0.484375,
0.76318359375,
0.12005615234375,
0.76318359375
],
[
0.51318359375,
0.755859375,
0.880859375,
0.755859375,
0.880859375,
0.76953125,
0.51318359375,
0.76953125
],
[
0.12005615234375,
0.765625,
0.48779296875,
0.765625,
0.48779296875,
0.779296875,
0.12005615234375,
0.779296875
],
[
0.51123046875,
0.77099609375,
0.880859375,
0.77099609375,
0.880859375,
0.78369140625,
0.51123046875,
0.78369140625
],
[
0.1175537109375,
0.78076171875,
0.486083984375,
0.78076171875,
0.486083984375,
0.79443359375,
0.1175537109375,
0.79443359375
],
[
0.51318359375,
0.7861328125,
0.87890625,
0.7861328125,
0.87890625,
0.7998046875,
0.51318359375,
0.7998046875
],
[
0.1175537109375,
0.79736328125,
0.173828125,
0.79736328125,
0.173828125,
0.8125,
0.1175537109375,
0.8125
],
[
0.5146484375,
0.80419921875,
0.880859375,
0.80419921875,
0.880859375,
0.81787109375,
0.5146484375,
0.81787109375
],
[
0.118408203125,
0.8154296875,
0.484375,
0.8154296875,
0.484375,
0.8291015625,
0.118408203125,
0.8291015625
],
[
0.51123046875,
0.8193359375,
0.87890625,
0.8193359375,
0.87890625,
0.8330078125,
0.51123046875,
0.8330078125
],
[
0.12005615234375,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.51123046875,
0.83447265625,
0.880859375,
0.8349609375,
0.880859375,
0.849609375,
0.51123046875,
0.8486328125
],
[
0.118408203125,
0.8466796875,
0.486083984375,
0.8466796875,
0.486083984375,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.51318359375,
0.8515625,
0.8818359375,
0.8515625,
0.8818359375,
0.865234375,
0.51318359375,
0.865234375
],
[
0.12005615234375,
0.86328125,
0.484375,
0.86328125,
0.484375,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.51123046875,
0.8681640625,
0.87890625,
0.8681640625,
0.87890625,
0.8818359375,
0.51123046875,
0.8818359375
],
[
0.1175537109375,
0.87744140625,
0.484375,
0.87890625,
0.484375,
0.892578125,
0.1175537109375,
0.89111328125
],
[
0.5322265625,
0.892578125,
0.7958984375,
0.8935546875,
0.7958984375,
0.90771484375,
0.5322265625,
0.90673828125
],
[
0.139404296875,
0.90625,
0.470947265625,
0.90625,
0.470947265625,
0.91943359375,
0.139404296875,
0.91943359375
],
[
0.51318359375,
0.90625,
0.8251953125,
0.90625,
0.8251953125,
0.92041015625,
0.51318359375,
0.92041015625
]
] |
[
"段。此外，这些片段通过SOTA语音到文本模",
"译和验证。我们请ISL讲师翻译视频（应用程",
"型(Radfordetal.,2022)来生成相应的文本。由",
"序中的详细信息。D）。签名者为每个视频提供",
"于生成的文本与书中章节的文本相同，因此验",
"个参考翻译。表2显示了ISL讲师创建的句",
"证生成的样本很容易，并且可以通过手动将它",
"子示例。为了定量估计数据集中翻译的可靠性",
"们与教科书进行匹配来完成。总的来说，我们",
"我们将数据集中的英语翻译文本与ISL讲师提",
"发现自动转录的文本质量很高；然而，错误生",
"供的文本进行比较。表3显示了ISLTranslate",
"成的文本是在书中内容的帮助下手动纠正的",
"中291句子的翻译分数。总体而言，BLEU-4",
"图2显示了一个长句子的示例（来自",
"分数为48.94，ROUGE-L（Lin，2004）为60.44",
"ISLTranslate）及其在ISL中的翻译。图中",
"WER（字错误率）为61.88。为了提供比较",
"的框架被分组为英语单词，并描绘了ISL中的",
"的参考，对于文本到文本的翻译，人类翻译",
"连续通信。注意句子中类似的单词，“符号/标",
"的BLEU得分在30-50之间（如Papinenietal.",
"志”和“语言”。“（另请参阅手语6的视觉",
"(2002）所报道，在来自40个一般新闻故事的",
"表示）。与其他自然语言一样，需要不同长度",
"约500个句子的测试语料库中，人类翻译者得",
"的表示（字符/手势）来传达不同的单词。在",
"分为34。68反对四个参考文献）。我们推测，",
"ISLTranslate中，我们仅限于句子/短语级别",
"与BLEU得分为48的ISLTranslate中的翻译",
"的翻译。数据集分为训练、验证和测试拆分",
"相比，具有很高的可靠性。93与经认证的ISL",
"（应用程序中的详细信息。B）。App。图3显示",
"签名者提供的参考翻译进行了比较。理想情况",
"了各种拆分中样本数的分布",
"下，最好为视频中的同一签名句子提供多个参",
"与其他连续手语数据集的比较：我们主要与基",
"考翻译；然而，高注释工作以及认证的ISL签",
"于视频的数据集进行比较，该数据集包含成对",
"名者的较低可用性使其成为一项具有挑战性的",
"的连续签名视频以及表1中相应的书面语言翻",
"任务",
"译。据我们所知，我们为ISL提出了最大的数",
"据集",
"4BaselineResults",
"数据清理和预处理：视频（e。g。，应用程序。",
"给定一个句子的符号视频，手语翻译的任务",
"Fig。4）包含图片对应的书页。我们通过将人",
"是将其翻译成口语句子（在我们的例子中为",
"脸位置视为参考点并删除视频中剩余的背景，",
"英语）。为了对ISLTranslate进行基准测试，",
"将签名者从视频中裁剪出来",
"我们为ISL到英语翻译创建了一个基准体系",
"ISLTranslate中的噪声消除：由于",
"结构。我们提出了一种受手语转换器（SLT）",
"ISLTranslate由使用可用音频信号中的暂停",
"（Camgozetal.，2020）启发的ISL姿势到英语翻",
"从较长的视频中剪辑的视频组成，因此数据集",
"译基线（称为Pose-SLT）。手语转换器将图像",
"中的噪声可能会以多种方式蔓延。在翻译音频",
"功能与转换器一起用干从签名视频生成文本翻",
"中的文本时，签名者可能会使用不同的符号，",
"译。然而，考虑到姿态估计模型(Selvarajetal.,",
"这些符号可能不是相应英语句子的逐字翻译。",
"2022）的更快的实时推断，我们使用姿态而不",
"此外，尽管背景中的音频与视频中的相应符号",
"是图像作为输人。我们使用Mediapipe姿态估",
"对齐，但在某些情况下，与相应的符号表示相",
"计管道/。Saundersetal.（2020）使用了类似的",
"比，音频可能会很快，并且可能会在句子的开",
"基于SLT的姿势到文本方法，该方法提出了用",
"头或结尾遗漏几个单词。我们还发现了一些情",
"于端到端手语制作的渐进式转换器，并使用基",
"况，在叙述故事时，音频中的人通过修改语音",
"于SLT的姿势到文本来通过反向翻译验证生",
"来扮演角色，使其听起来像指定的角色在说句",
"成的关键点（生成的姿势关键点到文本翻译）",
"子。例如，在一个老鼠正在说话的故事中，说",
"虽然基于姿势的方法处理速度更快，但它们的",
"话者可能会改变声音并增加音调以模拟老鼠所",
"性能通常不如基于图像的方法。对于整体关键",
"说的对话，而不是说出“老鼠说”语句后面的",
"点的选择，我们遵循Selvarajetal.（2022），它",
"句子。相比之下，在手语视频中，一个人在将",
"返回75个关键点的3D坐标（不包括面网格）。",
"句子翻译成ISL时可能会也可能不会扮演鼠标",
"此外，我们通过将肩部关键点的中点置于中心",
"的角色",
"并使用鼻子关键点和肩部中点之间的距离缩放",
"ISLTranslate验证：为了验证数据集中存在的",
"关键点来归一化每个的关键点。我们使用标",
"句子/短语ISL-英语对的可靠性，我们在经过认",
"准的BLEU和ROUGE分数来评估获得的英语",
"证的ISL签名者的帮助下。由于认证的ISL签",
"翻译（App中的模型超参数详细信息。E）。表",
"名者的可用性有限，我们只能使用一个小的随",
"4显示了针对建议的体系结构获得的结果。糟",
"机选择的符号文本对样本（291对）进行人工翻",
"7https://ai.googleblog.com/2020/12/",
"‘https://www.youtube.com/watch?v=SInKhy-06qA",
"mediapipe-holistic-simultaneous-face.html"
] |
{
"arxiv_id": "2307.05440",
"language": "zh",
"page_idx": 4
}
|
||
1612.00394-zh-1
|
1612.00394-zh-1.jpg
|
Definition Modeling: Learning to define word embeddings in natural language
Thanapon Noraset, Chen Liang, Larry Birnbaum, and Doug Downey
Department of Electrical Engineering & Computer Science
Northwestern University, Evanston IIL 60208, USA
$\{$ nor, chenliang2013 $\}$ @u.northwestern.edu, $\{$ l-birnbaum,d-downey $\}$ @northwestern.edu
$\operatorname{Abstract}$
Distributed representations of words have been shown to capture lexical semantics, as demonstrated by their effectiveness in word similarity and analogical relation tasks. But, these tasks only evaluate lexical semantics indirectly. In this paper, we study whether it is possible to utilize distributed representations to generate dictionary defnitions of words, as a more direct and transparent representation of the embeddings' semantics. We introduce definition modeling , the task of generating a definition for a given word and its embedding. We present several definition model architectures based on recurrent neural networks, and experiment with the models over multiple data sets. Our results show that a model that controls dependencies between the word being defined and the definition words performs significantly better, and that a character-level convolution layer designed to leverage morphology can complement word-level embeddings. Finally, an error analysis suggests that the errors made by a definition model may provide insight into the shortcomings of word embeddings.
# 1 Introduction
单词或单词 embeddings 的分布式表示形式是许多自然语言处理(NLP)模型(Turian,Ratinov, and Bengio 2010;Huang et al. 2014) 中的关键组件。最近，已经引人了几种神经网络技术来从未标记的文本数据(Mikolov et al. 2013a; Pennington, Socher, and Manning 2014; Yogatama et al. 2015) 中学习高质量的单词嵌人。嵌人已被证明可以捕获词汇语法和语义。例如，众所周知， 附近的嵌入更可能表示同义词 (Landauer and Dumais 1997)或同一类 (Downey, Schoenmackers, and Etzioni 2007)中的词。最近，嵌入之间的向量偏移已显示为反映类比关系 (Mikolov, Yih, and Zweig 2013) 。但是，诸如单词相似度和类比之类的任务仅间接评估嵌入的词汇信息
在这项工作中，我们研究是否可以使用单词嵌入来生成其相应单词的自然语言定义。字典定义作为词义的直接和明确的陈述。因此，与单词相似性和类比关系任务相比，定义生成可以被认为是嵌人所捕获的语法和语义的更透明的视图。我们介绍定义建模：估计文本定义的
Copyright $\odot2017$, Association for the Advancement of Ar-
tificial Intelligence (www.aaai.org). All rights reserved.
<table>
<tbody>
<tr>
<th>Word</th>
<th>Generated definition</th>
</tr>
<tr>
<td>$\overline{\mathrm{brawler}}$</td>
<td>a person who fights</td>
</tr>
<tr>
<td>prance</td>
<td>to walk in a lofty manner</td>
</tr>
<tr>
<td>valueless</td>
<td>not useful</td>
</tr>
</tbody>
</table>
Table 1: Selected examples of generated defnitions. The model has been trained on occurrences of each example word in running text, but not on the definitions.
概率的任务，给定一个被定义的单词及其嵌人。具体来说，对于给定的一组单词嵌人，在单词和定义对的语料库上训练定义模型。然后，根据每个单词的嵌人情况， 对模型进行测试，以了解它们在训练过程中未看到的单词的模型定义
本文研究的定义模型基于递归神经网络(RNN) 模型 (Elman 1990; Hochreiter and Schmidhuber 1997) 。RNN 模型在许多序列预测和自然语言生成任务 (Cho et al. 2014; Sutskever, Vinyals, and Le 2014; Karpathy and Fei-Fei 2014; Wen et al. 2015a) 上建立了新的最先进的性能。字典定义的一个重要特征是，定义中只有一部分单词强烈依赖于所定义的单词。例如，表1中“女性”定义中的“女性”一词取决于所定义的单词。为了捕获不同程度的依赖关系，我们引人了一个门控更新函数，该函数经过训练以控制用于生成每个定义词的被定义词的信息。此外，由于所定义单词的语素在定义中起着至关重要的作用，因此我们使用字符级卷积神经网络 (CNN) 进行实验，以测试它是否可以为单词嵌入提供补充信息。我们的最佳模型可以生成流畅而准确的定义，如表 1 所示。我们注意到，表中没有一个定义与训练期间看到的任何定义完全匹配
我们的贡献如下：(1)我们介绍了定义建模任务，并基于 RNN 语言模型提出了该任务的概率模型。(2)在不同模型架构和单词特征的实验中，我们表明门函数通
|
[
[
0.1070556640625,
0.12249755859375,
0.890625,
0.12371826171875,
0.890625,
0.141357421875,
0.1070556640625,
0.14013671875
],
[
0.2059326171875,
0.169189453125,
0.79150390625,
0.169189453125,
0.79150390625,
0.184326171875,
0.2059326171875,
0.184326171875
],
[
0.289306640625,
0.184326171875,
0.7099609375,
0.184326171875,
0.7099609375,
0.1988525390625,
0.289306640625,
0.1988525390625
],
[
0.316162109375,
0.1956787109375,
0.6787109375,
0.1962890625,
0.6787109375,
0.2108154296875,
0.316162109375,
0.210205078125
],
[
0.178955078125,
0.2108154296875,
0.82177734375,
0.2108154296875,
0.82177734375,
0.2259521484375,
0.178955078125,
0.2259521484375
],
[
0.529296875,
0.26953125,
0.57666015625,
0.271484375,
0.576171875,
0.285888671875,
0.52783203125,
0.2841796875
],
[
0.64892578125,
0.270751953125,
0.7998046875,
0.271484375,
0.7998046875,
0.285888671875,
0.64892578125,
0.285400390625
],
[
0.53173828125,
0.285888671875,
0.5908203125,
0.285888671875,
0.5908203125,
0.300537109375,
0.53173828125,
0.300537109375
],
[
0.25,
0.2734375,
0.316162109375,
0.2734375,
0.316162109375,
0.28857421875,
0.25,
0.28857421875
],
[
0.64892578125,
0.287353515625,
0.79150390625,
0.287353515625,
0.79150390625,
0.300537109375,
0.64892578125,
0.300537109375
],
[
0.10296630859375,
0.29736328125,
0.463134765625,
0.29736328125,
0.463134765625,
0.311279296875,
0.10296630859375,
0.311279296875
],
[
0.53369140625,
0.3017578125,
0.60546875,
0.3017578125,
0.60546875,
0.313232421875,
0.53369140625,
0.313232421875
],
[
0.650390625,
0.300537109375,
0.89208984375,
0.300537109375,
0.89208984375,
0.314453125,
0.650390625,
0.314453125
],
[
0.10296630859375,
0.31005859375,
0.463134765625,
0.31005859375,
0.463134765625,
0.323974609375,
0.10296630859375,
0.323974609375
],
[
0.53369140625,
0.31494140625,
0.61279296875,
0.31494140625,
0.61279296875,
0.326904296875,
0.53369140625,
0.326904296875
],
[
0.650390625,
0.31640625,
0.8037109375,
0.31640625,
0.8037109375,
0.326904296875,
0.650390625,
0.326904296875
],
[
0.10455322265625,
0.324462890625,
0.4599609375,
0.324462890625,
0.4599609375,
0.335205078125,
0.10455322265625,
0.335205078125
],
[
0.53369140625,
0.330078125,
0.57421875,
0.330078125,
0.57421875,
0.341552734375,
0.53369140625,
0.341552734375
],
[
0.650390625,
0.328857421875,
0.83251953125,
0.328857421875,
0.83251953125,
0.34228515625,
0.650390625,
0.34228515625
],
[
0.10296630859375,
0.335205078125,
0.463134765625,
0.335205078125,
0.463134765625,
0.348388671875,
0.10296630859375,
0.348388671875
],
[
0.53173828125,
0.34228515625,
0.5986328125,
0.34228515625,
0.5986328125,
0.356689453125,
0.53173828125,
0.356689453125
],
[
0.650390625,
0.34228515625,
0.88623046875,
0.34228515625,
0.88623046875,
0.35595703125,
0.650390625,
0.35595703125
],
[
0.650390625,
0.35595703125,
0.85302734375,
0.35595703125,
0.85302734375,
0.369384765625,
0.650390625,
0.369384765625
],
[
0.101318359375,
0.3466796875,
0.461669921875,
0.347900390625,
0.461669921875,
0.3623046875,
0.101318359375,
0.361083984375
],
[
0.10296630859375,
0.359130859375,
0.461669921875,
0.359130859375,
0.461669921875,
0.37255859375,
0.10296630859375,
0.37255859375
],
[
0.53369140625,
0.356689453125,
0.60107421875,
0.356689453125,
0.60107421875,
0.36865234375,
0.53369140625,
0.36865234375
],
[
0.53369140625,
0.369384765625,
0.6337890625,
0.369384765625,
0.6337890625,
0.380615234375,
0.53369140625,
0.380615234375
],
[
0.64892578125,
0.369384765625,
0.88623046875,
0.369384765625,
0.88623046875,
0.38330078125,
0.64892578125,
0.38330078125
],
[
0.65185546875,
0.3837890625,
0.74267578125,
0.3837890625,
0.74267578125,
0.395751953125,
0.65185546875,
0.395751953125
],
[
0.10455322265625,
0.375,
0.461669921875,
0.375,
0.461669921875,
0.385009765625,
0.10455322265625,
0.385009765625
],
[
0.10455322265625,
0.386962890625,
0.4599609375,
0.386962890625,
0.4599609375,
0.397705078125,
0.10455322265625,
0.397705078125
],
[
0.101318359375,
0.397705078125,
0.464111328125,
0.397705078125,
0.464111328125,
0.412353515625,
0.101318359375,
0.412353515625
],
[
0.53125,
0.397705078125,
0.6044921875,
0.395751953125,
0.6044921875,
0.410400390625,
0.53125,
0.412353515625
],
[
0.650390625,
0.397705078125,
0.88623046875,
0.397705078125,
0.88623046875,
0.410888671875,
0.650390625,
0.410888671875
],
[
0.101318359375,
0.410888671875,
0.461669921875,
0.410400390625,
0.461669921875,
0.423583984375,
0.101318359375,
0.4248046875
],
[
0.53369140625,
0.41357421875,
0.583984375,
0.41357421875,
0.583984375,
0.4248046875,
0.53369140625,
0.4248046875
],
[
0.650390625,
0.410888671875,
0.833984375,
0.410888671875,
0.833984375,
0.4248046875,
0.650390625,
0.4248046875
],
[
0.101318359375,
0.423583984375,
0.4599609375,
0.423583984375,
0.4599609375,
0.4375,
0.101318359375,
0.4375
],
[
0.53369140625,
0.4267578125,
0.580078125,
0.4267578125,
0.580078125,
0.438232421875,
0.53369140625,
0.438232421875
],
[
0.64892578125,
0.4248046875,
0.89794921875,
0.423583984375,
0.89794921875,
0.438232421875,
0.64892578125,
0.439453125
],
[
0.101318359375,
0.436279296875,
0.461669921875,
0.4375,
0.461669921875,
0.448974609375,
0.101318359375,
0.447509765625
],
[
0.53369140625,
0.43994140625,
0.583984375,
0.43994140625,
0.583984375,
0.451904296875,
0.53369140625,
0.451904296875
],
[
0.64892578125,
0.439453125,
0.8349609375,
0.439453125,
0.8349609375,
0.45263671875,
0.64892578125,
0.45263671875
],
[
0.101318359375,
0.447509765625,
0.461669921875,
0.448974609375,
0.461669921875,
0.462158203125,
0.101318359375,
0.4609375
],
[
0.53369140625,
0.453857421875,
0.5986328125,
0.453857421875,
0.5986328125,
0.46533203125,
0.53369140625,
0.46533203125
],
[
0.650390625,
0.453857421875,
0.7236328125,
0.453857421875,
0.7236328125,
0.46533203125,
0.650390625,
0.46533203125
],
[
0.10455322265625,
0.46337890625,
0.461669921875,
0.46337890625,
0.461669921875,
0.473388671875,
0.10455322265625,
0.473388671875
],
[
0.10455322265625,
0.47607421875,
0.459228515625,
0.47607421875,
0.459228515625,
0.486083984375,
0.10455322265625,
0.486083984375
],
[
0.51953125,
0.477783203125,
0.9111328125,
0.479248046875,
0.9111328125,
0.49365234375,
0.51953125,
0.492431640625
],
[
0.10455322265625,
0.488037109375,
0.4599609375,
0.488037109375,
0.4599609375,
0.498779296875,
0.10455322265625,
0.498779296875
],
[
0.10296630859375,
0.498779296875,
0.461669921875,
0.498779296875,
0.461669921875,
0.51220703125,
0.10296630859375,
0.51220703125
],
[
0.521484375,
0.49365234375,
0.91259765625,
0.49365234375,
0.91259765625,
0.5068359375,
0.521484375,
0.5068359375
],
[
0.10296630859375,
0.51123046875,
0.463134765625,
0.51123046875,
0.463134765625,
0.5244140625,
0.10296630859375,
0.5244140625
],
[
0.521484375,
0.50830078125,
0.86767578125,
0.50830078125,
0.86767578125,
0.521484375,
0.521484375,
0.521484375
],
[
0.10455322265625,
0.52587890625,
0.4599609375,
0.52587890625,
0.4599609375,
0.5361328125,
0.10455322265625,
0.5361328125
],
[
0.10296630859375,
0.537109375,
0.464111328125,
0.537109375,
0.464111328125,
0.55029296875,
0.10296630859375,
0.55029296875
],
[
0.10296630859375,
0.5498046875,
0.22216796875,
0.5498046875,
0.22216796875,
0.56298828125,
0.10296630859375,
0.56298828125
],
[
0.521484375,
0.5517578125,
0.9111328125,
0.5517578125,
0.9111328125,
0.56494140625,
0.521484375,
0.56494140625
],
[
0.521484375,
0.56640625,
0.9111328125,
0.56640625,
0.9111328125,
0.580078125,
0.521484375,
0.580078125
],
[
0.2115478515625,
0.583984375,
0.2467041015625,
0.583984375,
0.2467041015625,
0.5966796875,
0.2115478515625,
0.5966796875
],
[
0.24267578125,
0.583984375,
0.3544921875,
0.583984375,
0.3544921875,
0.59765625,
0.24267578125,
0.59765625
],
[
0.521484375,
0.580078125,
0.90869140625,
0.580078125,
0.90869140625,
0.59326171875,
0.521484375,
0.59326171875
],
[
0.521484375,
0.59326171875,
0.9111328125,
0.59326171875,
0.9111328125,
0.60693359375,
0.521484375,
0.60693359375
],
[
0.0882568359375,
0.60400390625,
0.476318359375,
0.60400390625,
0.476318359375,
0.6171875,
0.0882568359375,
0.6171875
],
[
0.52294921875,
0.609375,
0.61767578125,
0.609375,
0.61767578125,
0.62060546875,
0.52294921875,
0.62060546875
],
[
0.08905029296875,
0.6171875,
0.47802734375,
0.6171875,
0.47802734375,
0.6318359375,
0.08905029296875,
0.6318359375
],
[
0.5390625,
0.62548828125,
0.9111328125,
0.62548828125,
0.9111328125,
0.63623046875,
0.5390625,
0.63623046875
],
[
0.08660888671875,
0.6318359375,
0.476318359375,
0.6318359375,
0.476318359375,
0.64599609375,
0.08660888671875,
0.64599609375
],
[
0.51953125,
0.63623046875,
0.91259765625,
0.63818359375,
0.91259765625,
0.6513671875,
0.51953125,
0.6494140625
],
[
0.0906982421875,
0.646484375,
0.473876953125,
0.646484375,
0.473876953125,
0.6572265625,
0.0906982421875,
0.6572265625
],
[
0.521484375,
0.6513671875,
0.9111328125,
0.6513671875,
0.9111328125,
0.66552734375,
0.521484375,
0.66552734375
],
[
0.08660888671875,
0.6591796875,
0.47802734375,
0.6591796875,
0.47802734375,
0.673828125,
0.08660888671875,
0.673828125
],
[
0.51953125,
0.66552734375,
0.9111328125,
0.66552734375,
0.9111328125,
0.6787109375,
0.51953125,
0.6787109375
],
[
0.0882568359375,
0.67236328125,
0.47802734375,
0.67236328125,
0.47802734375,
0.68603515625,
0.0882568359375,
0.68603515625
],
[
0.51953125,
0.6787109375,
0.91259765625,
0.6796875,
0.91259765625,
0.693359375,
0.51953125,
0.6923828125
],
[
0.0882568359375,
0.6875,
0.47802734375,
0.6875,
0.47802734375,
0.70068359375,
0.0882568359375,
0.70068359375
],
[
0.52294921875,
0.693359375,
0.9111328125,
0.693359375,
0.9111328125,
0.70703125,
0.52294921875,
0.70703125
],
[
0.0882568359375,
0.70068359375,
0.47802734375,
0.70068359375,
0.47802734375,
0.7138671875,
0.0882568359375,
0.7138671875
],
[
0.52294921875,
0.70703125,
0.9111328125,
0.70703125,
0.9111328125,
0.72021484375,
0.52294921875,
0.72021484375
],
[
0.0882568359375,
0.7138671875,
0.47802734375,
0.7138671875,
0.47802734375,
0.72802734375,
0.0882568359375,
0.72802734375
],
[
0.521484375,
0.72021484375,
0.91015625,
0.72021484375,
0.91015625,
0.734375,
0.521484375,
0.734375
],
[
0.08660888671875,
0.72900390625,
0.47802734375,
0.72900390625,
0.47802734375,
0.7421875,
0.08660888671875,
0.7421875
],
[
0.51953125,
0.734375,
0.91259765625,
0.73291015625,
0.91259765625,
0.74755859375,
0.51953125,
0.74853515625
],
[
0.08660888671875,
0.7421875,
0.47802734375,
0.7421875,
0.47802734375,
0.755859375,
0.08660888671875,
0.755859375
],
[
0.521484375,
0.74853515625,
0.9111328125,
0.74853515625,
0.9111328125,
0.76220703125,
0.521484375,
0.76220703125
],
[
0.0882568359375,
0.755859375,
0.47802734375,
0.755859375,
0.47802734375,
0.76953125,
0.0882568359375,
0.76953125
],
[
0.521484375,
0.76318359375,
0.9111328125,
0.76318359375,
0.9111328125,
0.7763671875,
0.521484375,
0.7763671875
],
[
0.0882568359375,
0.771484375,
0.136474609375,
0.771484375,
0.136474609375,
0.78271484375,
0.0882568359375,
0.78271484375
],
[
0.521484375,
0.7763671875,
0.9111328125,
0.7763671875,
0.9111328125,
0.79052734375,
0.521484375,
0.79052734375
],
[
0.10296630859375,
0.78271484375,
0.476318359375,
0.78271484375,
0.476318359375,
0.79736328125,
0.10296630859375,
0.79736328125
],
[
0.521484375,
0.79052734375,
0.9111328125,
0.79052734375,
0.9111328125,
0.8037109375,
0.521484375,
0.8037109375
],
[
0.0882568359375,
0.79736328125,
0.476318359375,
0.79736328125,
0.476318359375,
0.81103515625,
0.0882568359375,
0.81103515625
],
[
0.51953125,
0.80322265625,
0.9111328125,
0.8037109375,
0.9111328125,
0.818359375,
0.51953125,
0.8173828125
],
[
0.08660888671875,
0.8125,
0.4755859375,
0.8125,
0.4755859375,
0.82568359375,
0.08660888671875,
0.82568359375
],
[
0.52294921875,
0.818359375,
0.91259765625,
0.818359375,
0.91259765625,
0.83203125,
0.52294921875,
0.83203125
],
[
0.0882568359375,
0.82568359375,
0.476318359375,
0.82568359375,
0.476318359375,
0.8388671875,
0.0882568359375,
0.8388671875
],
[
0.521484375,
0.83203125,
0.7509765625,
0.83203125,
0.7509765625,
0.84521484375,
0.521484375,
0.84521484375
],
[
0.0882568359375,
0.8388671875,
0.476318359375,
0.8388671875,
0.476318359375,
0.85302734375,
0.0882568359375,
0.85302734375
],
[
0.53759765625,
0.84765625,
0.900390625,
0.84765625,
0.900390625,
0.861328125,
0.53759765625,
0.861328125
],
[
0.89453125,
0.849609375,
0.91015625,
0.849609375,
0.91015625,
0.859375,
0.89453125,
0.859375
],
[
0.08660888671875,
0.86279296875,
0.4755859375,
0.86279296875,
0.4755859375,
0.876953125,
0.08660888671875,
0.876953125
],
[
0.52294921875,
0.861328125,
0.9140625,
0.861328125,
0.9140625,
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,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.52294921875,
0.8896484375
]
] |
[
"Definition Modeling: Learning to define word embeddings in natural language",
"Thanapon Noraset, Chen Liang, Larry Birnbaum, and Doug Downey",
"Department of Electrical Engineering & Computer Science",
"Northwestern University, Evanston IL 60208, USA",
"[ nor, chenliang2013 } @u.northwestern.edu, { 1-birnbaum,d-downey } @northwestern.edu",
"Word",
"Generated definition",
"brawler",
"Abstract",
"a person who fights",
"Distributed representations of words have been shown",
"butterfish",
"a marine fish of the atlantic coast",
"to capture lexical semantics, as demonstrated by their",
"continually",
"in a constant manner",
"effectiveness in word similarity and analogical relation",
"creek",
"a narrow stream of water",
"tasks. But, these tasks only evaluate lexical semantics",
"feminine",
"having the character of a woman",
"the quality of being childish",
"indirectly. In this paper, we study whether it is pos-",
"sible to utilize distributed representations to generate",
"juvenility",
"mathematical",
"of or pertaining to the science of",
"mathematics",
"dictionary definitions of words, as a more direct and",
"transparent representation of the embeddings'seman-",
"tics. We introduce definition modeling , the task of",
"negotiate",
"to make a contract or agreement",
"generating a definition for a given word and its embed-",
"prance",
"to walk in a lofty manner",
"ding. We present several definition model architectures",
"resent",
"to have a feeling of anger or dislike",
"based on recurrent neural networks, and experiment",
"similar",
"having the same qualities",
"with the models over multiple data sets. Our results",
"valueless",
"not useful",
"showthat amodelthat controlsdependenciesbetween",
"the word being defined and the definition words per-",
"Table 1: Selected examples of generated definitions. The",
"forms significantly better, and that a character-level",
"convolution layer designed to leverage morphology can",
"model has been trained on occurrences of each example",
"complement word-level embeddings. Finally, an error",
"word in running text, but not on the definitions.",
"analysissuggests that the errors made bya definition",
"model may provide insight into the shortcomings of",
"word embeddings.",
"概率的任务，给定一个被定义的单词及其嵌入。具体来",
"说，对于给定的一组单词嵌入，在单词和定义对的语料",
"1",
"Introduction",
"库上训练定义模型。然后，根据每个单词的嵌人情况",
"对模型进行测试，以了解它们在训练过程中未看到的单",
"单词或单词embeddings的分布式表示形式是许多自",
"词的模型定义",
"然语言处理（NLP）模型（Turian，Ratinov，andBengio",
"本文研究的定义模型基于递归神经网络（RNN）模",
"2010;Huangetal.2014）中的关键组件。最近，已经引人",
"型(Elman 1990:Hochreiterand Schmidhuber 1997)。",
"了儿种神经网络技术来从未标记的文本数据（Mikolov",
"RNN模型在许多序列预测和自然语言生成任务（Cho",
"et al. 2013a; Pennington, Socher, and Manning 2014;",
"et al. 2014; Sutskever, Vinyals, and Le 2014; Karpathy",
"Yogatamaetal.2015）中学习高质量的单词嵌人。嵌人",
"andFei-Fei2014;Wenetal.2015a）上建立了新的最先",
"已被证明可以捕获词汇语法和语义。例如，众所周知，",
"进的性能。字典定义的一个重要特征是，定义中只有一",
"附近的嵌人更可能表示同义词（LandauerandDumais",
"部分单词强烈依赖于所定义的单词。例如，表1中“女",
"1997）或同一类（Downey,Schoenmackers，andEtzioni",
"性”定义中的“女性”一词取决于所定义的单词。为了",
"2007）中的词。最近，嵌入之间的向量偏移已显示为反",
"捕获不同程度的依赖关系，我们引人了二个门控更新",
"映类比关系（Mikolov，Yih，andZweig2013）。但是，诸",
"函数，该函数经过训练以控制用于生成每个定义词的被",
"如单词相似度和类比之类的任务仅间接评估嵌人的词",
"定义词的信息。此外，由于所定义单词的语素在定义中",
"汇信息",
"起着至关重要的作用，因此我们使用字符级卷积神经网",
"在这项工作中，我们研究是否可以使用单词嵌入来生",
"络（CNN）进行实验，以测试它是否可以为单词嵌人提",
"成其相应单词的自然语言定义。字典定义作为词义的直",
"供补充信息。我们的最佳模型可以生成流畅而准确的定",
"接和明确的陈述。因此，与单词相似性和类比关系任务",
"义，如表1所示。我们注意到，表中没有一个定义与训",
"相比，定义生成可以被认为是嵌人所捕获的语法和语义",
"练期间看到的任何定义完全匹配",
"的更透明的视图。我们介绍定义建模：估计文本定义的",
"我们的贡献如下：（1）我们介绍了定义建模任务，",
"并",
"Copyright 2017, Association for the Advancement of Ar-",
"基于RNN语言模型提出了该任务的概率模型。（2）在",
"tificial Intelligence (www.aaai.org). All rights reserved.",
"不同模型架构和单词特征的实验中，我们表明门函数通"
] |
{
"arxiv_id": "1612.00394",
"language": "zh",
"page_idx": 1
}
|
||
1612.00394-zh-2
|
1612.00394-zh-2.jpg
|
过$\sim10\%$ 改善了 RNN 语言模型在定义建模任务上的困惑，而字符级 CNN 通过$\sim5\%$进一步改善了困惑。(3) 我们还表明，定义模型可用于执行先前工作中研究的反向词典任务，其目标是将给定的定义与其相应的单词相匹配。我们的模型实现了 11。8 与 Hill 等人以前的最新技术相比，%在精度上的绝对增益。(2016)。(4)最后，我们的错误分析表明，训练有素的单词嵌入集对生成的定义的质量起着重要作用，并且某些错误类型表明单词嵌人中编码的信息存在缺陷
2 Previous Work
我们的目标是研究 RNN 模型，该模型通过训练字典定义的示例来学习定义单词嵌人。尽管字典语料库已在NLP 中得到广泛使用，但据我们所知，以前的工作都没有尝试创建定义的生成模型。早期的工作集中在来自定义的 extracting 语义信息上。例如，Chodorow (1985) 以及 Klavans 和 Whitman (2001) 构建了词典中单词的分类法。Dolan 等人。(1993)和 Vanderwende 等人。(2005)从定义中提取语义表示，以填充词汇知识库
在分布式语义中，单词由实数的密集向量表示，而不是语义谓词。最近，字典定义已被用来学习这样的嵌人。例如，Wang 等人。(2015) 将定义文本中的单词用作 Word2Vec 算法 (Mikolov et al. 2013b)的“上下文” 单词形式。Hill 等人。(2016) 使用字典定义对组合性进行建模，并使用反向字典任务评估模型。虽然这些作品从定义中学习单词或短语嵌人，但我们只关注从现有(固定)嵌人生成定义。我们的实验表明，我们的模型优于希尔等人的模型。反向字典任务上的(2016)
我们的工作采用嵌入模型来生成自然语言。最近在与我们不同的任务上的各种工作中也采取了类似的方法。Dinu 和 Baroni (2014)提出了一种方法，该方法使用嵌人将单个单词映射到表示相同含义的更长短语。同样，Li 等人。(2015) 研究如何将段落或文档编码为嵌人，并从编码中重建原始文本。其他最近的工作，如图像字幕生成 (Karpathy and Fei-Fei 2014) 和口语对话生成 (Wen et al. 2015a) 也与我们的工作有关，因为单词序列是从单个输入向量生成的。我们的模型架构受到序列到序列模型 (Cho et al. 2014; Sutskever, Vinyals, and Le 2014)的启发，但是定义建模是截然不同的，因为它是 word 到序列的任务
3 Dictionary Definitions
在本节中，我们首先通过对现有词典的研究来研究定义内容和结构。然后，我们描述我们的新数据集，并定义我们的任务和指标
3.1 Definition Content and Structure
在现有的词典中，单个定义通常由属和差异(Chodorow, Byrd, and Heidorn 1985; Montemagni and Vanderwende 1992) 组成。genus 是所定义单词的广义类， 而 differentiae 使该单词与同一类中的其他单词不同。例如，
Phosphorescent: emitting light without apprecia-
ble heat as by slow oxidation of phosphorous
“发光”是一个属，并且“没有适用的热量。“是一个不同的。此外，定义往往包括常见的模式，如“。“或”一个有。”(Markowitz, Ahlswede, and Evens 1986)。但是，每个字典的模式和样式通常是唯一的
<table>
<tbody>
<tr>
<th>$\overline{\mathrm{Split}}$</th>
<th>$\tan$</th>
<th>valid</th>
<th>test</th>
</tr>
<tr>
<td>$\overline{\mathrm{Words}}$ $\#$</td>
<td>20,298</td>
<td>$\overline{1,127}$</td>
<td>$\overline{1,129}$</td>
</tr>
<tr>
<td>Entries</td>
<td>146,486</td>
<td>8,087</td>
<td>8,352</td>
</tr>
<tr>
<td>Tokens</td>
<td>1,406,440</td>
<td>77,948</td>
<td>79,699</td>
</tr>
<tr>
<td>Avg $\operatorname{length}$</td>
<td>6.60</td>
<td>6.64</td>
<td>6.54</td>
</tr>
</tbody>
</table>
Table 2: Basic statistics of the common word definitions corpus. Splits are mutually exclusive in the words being defined.
属+差异 ( G+ D) 结构不是定义的唯一模式。例如，
下面的条目显示不同的结构。
Eradication : the act of plucking up by the roots;
a rooting out; extirpation; utter destruction
这组定义包括同义词 (“extirppation”), G+D 结构的反面 (“utterdestruction”)和未分类结构(“rootingout")
3.2 Corpus: Preprocessing and Analysis
字典语料库以数字格式提供，但专为人类消费而设计， 需要预处理才能用于机器学习。字典包含非定义文本， 以帮助人类读者，例如。g。Wordnik$^{1}$中“渐变”的条目包含字段(“数学”)和示例用法(“as,铁路的渐变线。)。此外，许多条目包含多个定义，通常(但不总是) $\stackrel{\sim}{\operatorname*{\text{用}}}\stackrel{\sim}{\operatorname*{\operatorname*{,}}}\stackrel{\sim}{\operatorname*{,}}\stackrel{\sim}{\operatorname*{,}}\stackrel{\sim}{\operatorname*{,}}$分隔。
我们需要一个语料库，其中每个条目仅包含一个定义的单词和一个定义。我们从 GCIDE $^{2}$解析字典条目，并预处理 WordNet 的光泽，并删除字段和用法。解析器和预处理脚本可以在 https://github 上找到。com/northanapon/dict-definition
为了为机器学习实验创建一个合理大小的语料库，我们从 GoogleWeb1T 语料库 (Brants and Franz 2006) 中的 50k 个最常见的单词中采样了大约 20k 个单词，删除了功能单词。此外，在用“”拆分之前，我们将字典中每个单词的条目数限制为三个(以便定义的每个单词在我们的语料库中可以重复多次)。清理和修剪后，语料库的词汇量为 29k。其他语料库统计见表 2 。
我们通过使用以下四种结构之一手动标记每个定义来分析语料库中定义的基础结构：G+D,D+G,Syn (同义词)和 Misc。总的来说，我们从 100 个随机选择的单词中检查了 680 个定义。结果见表 3 。我们重申早期的研究表明，G+D 结构在这两个词典中占主导地位。但是，也存在其他结构，突出了字典建模任务中固有的挑战。此外，我们观察到属术语有时是通用的(e. $\begin{array}{ll}{\mathrm{g。,“—个”或“那个”),以及其他特定时间(e。g。“倡}}\\{\mathrm{g。,“—个}}\end{array}$ $导者”)$
## 3.3 Dictionary Defnition Tasks
在定义建模(DM)任务中，给我们一个输人单词$w^*$
并输出任何给定文本$D$作为输入单词定义的可能性。换句话说，我们估计$P(D|w^*)$。我们假设我们的定义模型可以访问一组词嵌人，这些词嵌入是从用于训练定义模型的定义语料库以外的一些语料库中估计的
DM 是语言建模的特例，与语言建模一样，定义模型的性能可以通过使用测试语料库的困惑来衡量。较低的
$^{1}$https://www.wordnik.com/words/gradient
$^{2}$http://gcide.gnu.org.ua/
|
[
[
0.60888671875,
0.0662841796875,
0.64892578125,
0.0662841796875,
0.64892578125,
0.080810546875,
0.60888671875,
0.080810546875
],
[
0.69775390625,
0.06817626953125,
0.73779296875,
0.06817626953125,
0.73779296875,
0.0802001953125,
0.69775390625,
0.0802001953125
],
[
0.76318359375,
0.06817626953125,
0.80126953125,
0.06817626953125,
0.80126953125,
0.0802001953125,
0.76318359375,
0.0802001953125
],
[
0.8349609375,
0.06817626953125,
0.86767578125,
0.06817626953125,
0.86767578125,
0.0802001953125,
0.8349609375,
0.0802001953125
],
[
0.580078125,
0.080810546875,
0.64794921875,
0.080810546875,
0.64794921875,
0.0953369140625,
0.580078125,
0.0953369140625
],
[
0.08905029296875,
0.07135009765625,
0.4755859375,
0.07135009765625,
0.4755859375,
0.08209228515625,
0.08905029296875,
0.08209228515625
],
[
0.685546875,
0.080810546875,
0.73779296875,
0.080810546875,
0.73779296875,
0.0953369140625,
0.685546875,
0.0953369140625
],
[
0.7607421875,
0.080810546875,
0.80126953125,
0.080810546875,
0.80126953125,
0.0953369140625,
0.7607421875,
0.0953369140625
],
[
0.82421875,
0.080810546875,
0.869140625,
0.080810546875,
0.869140625,
0.0953369140625,
0.82421875,
0.0953369140625
],
[
0.57421875,
0.09344482421875,
0.64794921875,
0.0953369140625,
0.64794921875,
0.10986328125,
0.57421875,
0.10797119140625
],
[
0.08660888671875,
0.08209228515625,
0.47802734375,
0.08331298828125,
0.47802734375,
0.09783935546875,
0.08660888671875,
0.09661865234375
],
[
0.7587890625,
0.0953369140625,
0.80126953125,
0.0953369140625,
0.80126953125,
0.10986328125,
0.7587890625,
0.10986328125
],
[
0.82275390625,
0.0953369140625,
0.86669921875,
0.0953369140625,
0.86669921875,
0.10986328125,
0.82275390625,
0.10986328125
],
[
0.68115234375,
0.09661865234375,
0.736328125,
0.09661865234375,
0.736328125,
0.10797119140625,
0.68115234375,
0.10797119140625
],
[
0.0882568359375,
0.09783935546875,
0.476318359375,
0.09783935546875,
0.476318359375,
0.111083984375,
0.0882568359375,
0.111083984375
],
[
0.576171875,
0.10919189453125,
0.64794921875,
0.10919189453125,
0.64794921875,
0.12371826171875,
0.576171875,
0.12371826171875
],
[
0.75341796875,
0.10986328125,
0.7998046875,
0.10986328125,
0.7998046875,
0.121826171875,
0.75341796875,
0.121826171875
],
[
0.81884765625,
0.10986328125,
0.865234375,
0.10986328125,
0.865234375,
0.121826171875,
0.81884765625,
0.121826171875
],
[
0.66845703125,
0.111083984375,
0.734375,
0.111083984375,
0.734375,
0.12249755859375,
0.66845703125,
0.12249755859375
],
[
0.08905029296875,
0.11236572265625,
0.47802734375,
0.11236572265625,
0.47802734375,
0.1256103515625,
0.08905029296875,
0.1256103515625
],
[
0.705078125,
0.12371826171875,
0.736328125,
0.12371826171875,
0.736328125,
0.1363525390625,
0.705078125,
0.1363525390625
],
[
0.7685546875,
0.12371826171875,
0.80126953125,
0.12371826171875,
0.80126953125,
0.1363525390625,
0.7685546875,
0.1363525390625
],
[
0.83251953125,
0.12371826171875,
0.865234375,
0.12371826171875,
0.865234375,
0.1363525390625,
0.83251953125,
0.1363525390625
],
[
0.56884765625,
0.125,
0.64794921875,
0.125,
0.64794921875,
0.1363525390625,
0.56884765625,
0.1363525390625
],
[
0.08905029296875,
0.125,
0.4755859375,
0.1256103515625,
0.4755859375,
0.1395263671875,
0.08905029296875,
0.1383056640625
],
[
0.08905029296875,
0.1395263671875,
0.47802734375,
0.1395263671875,
0.47802734375,
0.15283203125,
0.08905029296875,
0.15283203125
],
[
0.08660888671875,
0.1514892578125,
0.476318359375,
0.15087890625,
0.476318359375,
0.1654052734375,
0.08660888671875,
0.1666259765625
],
[
0.521484375,
0.15087890625,
0.91259765625,
0.15087890625,
0.91259765625,
0.1641845703125,
0.521484375,
0.1641845703125
],
[
0.08905029296875,
0.1654052734375,
0.476318359375,
0.1654052734375,
0.476318359375,
0.1787109375,
0.08905029296875,
0.1787109375
],
[
0.521484375,
0.1654052734375,
0.9111328125,
0.1654052734375,
0.9111328125,
0.1787109375,
0.521484375,
0.1787109375
],
[
0.08660888671875,
0.1787109375,
0.316162109375,
0.179931640625,
0.316162109375,
0.1944580078125,
0.08660888671875,
0.1932373046875
],
[
0.52294921875,
0.179931640625,
0.576171875,
0.179931640625,
0.576171875,
0.1900634765625,
0.52294921875,
0.1900634765625
],
[
0.200927734375,
0.20703125,
0.36279296875,
0.20703125,
0.36279296875,
0.2203369140625,
0.200927734375,
0.2203369140625
],
[
0.5361328125,
0.21533203125,
0.91259765625,
0.216552734375,
0.91259765625,
0.2310791015625,
0.5361328125,
0.2298583984375
],
[
0.0882568359375,
0.2259521484375,
0.476318359375,
0.2259521484375,
0.476318359375,
0.2392578125,
0.0882568359375,
0.2392578125
],
[
0.52392578125,
0.2310791015625,
0.7236328125,
0.2310791015625,
0.7236328125,
0.2442626953125,
0.52392578125,
0.2442626953125
],
[
0.0906982421875,
0.2392578125,
0.47802734375,
0.2392578125,
0.47802734375,
0.25244140625,
0.0906982421875,
0.25244140625
],
[
0.08660888671875,
0.251953125,
0.476318359375,
0.25244140625,
0.476318359375,
0.267578125,
0.08660888671875,
0.266357421875
],
[
0.53515625,
0.25,
0.896484375,
0.251953125,
0.896484375,
0.26513671875,
0.53515625,
0.26318359375
],
[
0.0882568359375,
0.267578125,
0.476318359375,
0.267578125,
0.476318359375,
0.281005859375,
0.0882568359375,
0.281005859375
],
[
0.5361328125,
0.266357421875,
0.8505859375,
0.26513671875,
0.8505859375,
0.276611328125,
0.5361328125,
0.27783203125
],
[
0.0906982421875,
0.281005859375,
0.476318359375,
0.281005859375,
0.476318359375,
0.294189453125,
0.0906982421875,
0.294189453125
],
[
0.521484375,
0.2841796875,
0.91015625,
0.2841796875,
0.91015625,
0.29736328125,
0.521484375,
0.29736328125
],
[
0.0882568359375,
0.294189453125,
0.47802734375,
0.294189453125,
0.47802734375,
0.30810546875,
0.0882568359375,
0.30810546875
],
[
0.521484375,
0.298583984375,
0.9111328125,
0.298583984375,
0.9111328125,
0.3125,
0.521484375,
0.3125
],
[
0.08905029296875,
0.31005859375,
0.476318359375,
0.31005859375,
0.476318359375,
0.32080078125,
0.08905029296875,
0.32080078125
],
[
0.0882568359375,
0.322509765625,
0.449462890625,
0.322509765625,
0.449462890625,
0.3359375,
0.0882568359375,
0.3359375
],
[
0.521484375,
0.32568359375,
0.8486328125,
0.32568359375,
0.8486328125,
0.339111328125,
0.521484375,
0.339111328125
],
[
0.10296630859375,
0.335205078125,
0.476318359375,
0.3359375,
0.476318359375,
0.349853515625,
0.10296630859375,
0.348388671875
],
[
0.0882568359375,
0.349853515625,
0.47802734375,
0.349853515625,
0.47802734375,
0.363037109375,
0.0882568359375,
0.363037109375
],
[
0.52294921875,
0.3447265625,
0.91259765625,
0.3447265625,
0.91259765625,
0.35791015625,
0.52294921875,
0.35791015625
],
[
0.0882568359375,
0.3623046875,
0.476318359375,
0.363037109375,
0.476318359375,
0.37744140625,
0.0882568359375,
0.376953125
],
[
0.52294921875,
0.359130859375,
0.91015625,
0.359130859375,
0.91015625,
0.37255859375,
0.52294921875,
0.37255859375
],
[
0.5185546875,
0.37060546875,
0.91259765625,
0.369384765625,
0.91259765625,
0.386962890625,
0.5185546875,
0.38818359375
],
[
0.0882568359375,
0.37744140625,
0.47802734375,
0.37744140625,
0.47802734375,
0.391357421875,
0.0882568359375,
0.391357421875
],
[
0.08905029296875,
0.391357421875,
0.47802734375,
0.391357421875,
0.47802734375,
0.40478515625,
0.08905029296875,
0.40478515625
],
[
0.52392578125,
0.386962890625,
0.9111328125,
0.386962890625,
0.9111328125,
0.40087890625,
0.52392578125,
0.40087890625
],
[
0.51953125,
0.3984375,
0.91259765625,
0.399658203125,
0.91259765625,
0.4140625,
0.51953125,
0.41357421875
],
[
0.08905029296875,
0.4072265625,
0.476318359375,
0.4072265625,
0.476318359375,
0.417236328125,
0.08905029296875,
0.417236328125
],
[
0.51953125,
0.410400390625,
0.61767578125,
0.412353515625,
0.61767578125,
0.427978515625,
0.51953125,
0.425537109375
],
[
0.08905029296875,
0.419189453125,
0.476318359375,
0.419189453125,
0.476318359375,
0.43310546875,
0.08905029296875,
0.43310546875
],
[
0.5361328125,
0.4267578125,
0.9111328125,
0.427978515625,
0.9111328125,
0.442626953125,
0.5361328125,
0.44140625
],
[
0.0882568359375,
0.43505859375,
0.476318359375,
0.43505859375,
0.476318359375,
0.44580078125,
0.0882568359375,
0.44580078125
],
[
0.52392578125,
0.44140625,
0.9111328125,
0.44140625,
0.9111328125,
0.455810546875,
0.52392578125,
0.455810546875
],
[
0.08905029296875,
0.448974609375,
0.410888671875,
0.448974609375,
0.410888671875,
0.458984375,
0.08905029296875,
0.458984375
],
[
0.521484375,
0.45263671875,
0.91259765625,
0.453857421875,
0.91259765625,
0.4716796875,
0.521484375,
0.47021484375
],
[
0.10455322265625,
0.4609375,
0.47802734375,
0.4609375,
0.47802734375,
0.474853515625,
0.10455322265625,
0.474853515625
],
[
0.0882568359375,
0.474853515625,
0.4755859375,
0.474853515625,
0.4755859375,
0.488037109375,
0.0882568359375,
0.488037109375
],
[
0.521484375,
0.47021484375,
0.91259765625,
0.47021484375,
0.91259765625,
0.484130859375,
0.521484375,
0.484130859375
],
[
0.0882568359375,
0.488037109375,
0.476318359375,
0.488037109375,
0.476318359375,
0.501953125,
0.0882568359375,
0.501953125
],
[
0.521484375,
0.484130859375,
0.7587890625,
0.484130859375,
0.7587890625,
0.49755859375,
0.521484375,
0.49755859375
],
[
0.08905029296875,
0.501953125,
0.465576171875,
0.501953125,
0.465576171875,
0.51513671875,
0.08905029296875,
0.51513671875
],
[
0.53759765625,
0.49755859375,
0.91259765625,
0.49755859375,
0.91259765625,
0.51123046875,
0.53759765625,
0.51123046875
],
[
0.461669921875,
0.5048828125,
0.4755859375,
0.5048828125,
0.4755859375,
0.51318359375,
0.461669921875,
0.51318359375
],
[
0.0882568359375,
0.5166015625,
0.476318359375,
0.5166015625,
0.476318359375,
0.52978515625,
0.0882568359375,
0.52978515625
],
[
0.52294921875,
0.51318359375,
0.91015625,
0.51318359375,
0.91015625,
0.5234375,
0.52294921875,
0.5234375
],
[
0.0906982421875,
0.52978515625,
0.47802734375,
0.52978515625,
0.47802734375,
0.54345703125,
0.0906982421875,
0.54345703125
],
[
0.521484375,
0.52587890625,
0.91259765625,
0.52587890625,
0.91259765625,
0.5390625,
0.521484375,
0.5390625
],
[
0.0882568359375,
0.54345703125,
0.476318359375,
0.54248046875,
0.476318359375,
0.55615234375,
0.0882568359375,
0.556640625
],
[
0.52294921875,
0.54150390625,
0.91015625,
0.54150390625,
0.91015625,
0.5517578125,
0.52294921875,
0.5517578125
],
[
0.08905029296875,
0.55810546875,
0.47802734375,
0.55810546875,
0.47802734375,
0.5712890625,
0.08905029296875,
0.5712890625
],
[
0.52294921875,
0.55322265625,
0.9111328125,
0.55322265625,
0.9111328125,
0.56640625,
0.52294921875,
0.56640625
],
[
0.0882568359375,
0.5712890625,
0.476318359375,
0.5712890625,
0.476318359375,
0.58544921875,
0.0882568359375,
0.58544921875
],
[
0.51953125,
0.56640625,
0.91259765625,
0.5673828125,
0.91259765625,
0.58056640625,
0.51953125,
0.580078125
],
[
0.0882568359375,
0.58544921875,
0.47802734375,
0.58544921875,
0.47802734375,
0.5986328125,
0.0882568359375,
0.5986328125
],
[
0.51953125,
0.580078125,
0.8486328125,
0.580078125,
0.8486328125,
0.59326171875,
0.51953125,
0.59326171875
],
[
0.08660888671875,
0.5986328125,
0.476318359375,
0.599609375,
0.476318359375,
0.61279296875,
0.08660888671875,
0.6123046875
],
[
0.53759765625,
0.59521484375,
0.91015625,
0.59521484375,
0.91015625,
0.609375,
0.53759765625,
0.609375
],
[
0.0882568359375,
0.61279296875,
0.253173828125,
0.61279296875,
0.253173828125,
0.626953125,
0.0882568359375,
0.626953125
],
[
0.52294921875,
0.609375,
0.91259765625,
0.609375,
0.91259765625,
0.62255859375,
0.52294921875,
0.62255859375
],
[
0.52294921875,
0.62353515625,
0.9111328125,
0.62353515625,
0.9111328125,
0.63720703125,
0.52294921875,
0.63720703125
],
[
0.1690673828125,
0.64013671875,
0.393798828125,
0.64013671875,
0.393798828125,
0.65380859375,
0.1690673828125,
0.65380859375
],
[
0.521484375,
0.63525390625,
0.9111328125,
0.63623046875,
0.9111328125,
0.65087890625,
0.521484375,
0.6494140625
],
[
0.52294921875,
0.65087890625,
0.9111328125,
0.65087890625,
0.9111328125,
0.6640625,
0.52294921875,
0.6640625
],
[
0.0882568359375,
0.65771484375,
0.476318359375,
0.65771484375,
0.476318359375,
0.671875,
0.0882568359375,
0.671875
],
[
0.52294921875,
0.6640625,
0.91259765625,
0.6640625,
0.91259765625,
0.67822265625,
0.52294921875,
0.67822265625
],
[
0.08660888671875,
0.67041015625,
0.476318359375,
0.671875,
0.476318359375,
0.68603515625,
0.08660888671875,
0.68505859375
],
[
0.52294921875,
0.6787109375,
0.9111328125,
0.6787109375,
0.9111328125,
0.6923828125,
0.52294921875,
0.6923828125
],
[
0.08660888671875,
0.68505859375,
0.2197265625,
0.68505859375,
0.2197265625,
0.69873046875,
0.08660888671875,
0.69873046875
],
[
0.52294921875,
0.6943359375,
0.55126953125,
0.6943359375,
0.55126953125,
0.70556640625,
0.52294921875,
0.70556640625
],
[
0.54736328125,
0.68994140625,
0.91259765625,
0.6923828125,
0.91259765625,
0.70703125,
0.54736328125,
0.705078125
],
[
0.0849609375,
0.71044921875,
0.393798828125,
0.71044921875,
0.393798828125,
0.7236328125,
0.0849609375,
0.7236328125
],
[
0.521484375,
0.705078125,
0.580078125,
0.705078125,
0.580078125,
0.7197265625,
0.521484375,
0.7197265625
],
[
0.08905029296875,
0.72900390625,
0.47802734375,
0.72900390625,
0.47802734375,
0.7421875,
0.08905029296875,
0.7421875
],
[
0.521484375,
0.73486328125,
0.78369140625,
0.73486328125,
0.78369140625,
0.74560546875,
0.521484375,
0.74560546875
],
[
0.08660888671875,
0.7421875,
0.476318359375,
0.7421875,
0.476318359375,
0.755859375,
0.08660888671875,
0.755859375
],
[
0.0882568359375,
0.7568359375,
0.4755859375,
0.7568359375,
0.4755859375,
0.77099609375,
0.0882568359375,
0.77099609375
],
[
0.521484375,
0.751953125,
0.91015625,
0.75244140625,
0.91015625,
0.7666015625,
0.521484375,
0.76513671875
],
[
0.08660888671875,
0.76806640625,
0.4794921875,
0.76953125,
0.4794921875,
0.7841796875,
0.08660888671875,
0.78271484375
],
[
0.5361328125,
0.7666015625,
0.91015625,
0.7666015625,
0.91015625,
0.77978515625,
0.5361328125,
0.77978515625
],
[
0.0882568359375,
0.78271484375,
0.1258544921875,
0.78515625,
0.125,
0.79736328125,
0.08660888671875,
0.79541015625
],
[
0.51953125,
0.77880859375,
0.91015625,
0.77978515625,
0.91015625,
0.79443359375,
0.51953125,
0.79345703125
],
[
0.52294921875,
0.79345703125,
0.9111328125,
0.79345703125,
0.9111328125,
0.806640625,
0.52294921875,
0.806640625
],
[
0.10296630859375,
0.8017578125,
0.45751953125,
0.8017578125,
0.45751953125,
0.81494140625,
0.10296630859375,
0.81494140625
],
[
0.525390625,
0.81005859375,
0.86279296875,
0.81005859375,
0.86279296875,
0.82080078125,
0.525390625,
0.82080078125
],
[
0.10296630859375,
0.81640625,
0.4248046875,
0.8173828125,
0.4248046875,
0.8291015625,
0.10296630859375,
0.82763671875
],
[
0.53515625,
0.82080078125,
0.9111328125,
0.8212890625,
0.9111328125,
0.8359375,
0.53515625,
0.83544921875
],
[
0.08660888671875,
0.83203125,
0.431396484375,
0.83349609375,
0.431396484375,
0.8466796875,
0.08660888671875,
0.84521484375
],
[
0.4248046875,
0.83544921875,
0.4755859375,
0.83544921875,
0.4755859375,
0.84521484375,
0.4248046875,
0.84521484375
],
[
0.521484375,
0.8359375,
0.9111328125,
0.8359375,
0.9111328125,
0.849609375,
0.521484375,
0.849609375
],
[
0.0882568359375,
0.8466796875,
0.455810546875,
0.84521484375,
0.455810546875,
0.85986328125,
0.0882568359375,
0.861328125
],
[
0.08905029296875,
0.861328125,
0.47802734375,
0.861328125,
0.47802734375,
0.87548828125,
0.08905029296875,
0.87548828125
],
[
0.5400390625,
0.85986328125,
0.830078125,
0.8623046875,
0.830078125,
0.876953125,
0.5400390625,
0.875
],
[
0.0882568359375,
0.87548828125,
0.381591796875,
0.87548828125,
0.381591796875,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.5400390625,
0.87353515625,
0.7099609375,
0.87548828125,
0.7099609375,
0.89013671875,
0.5400390625,
0.88818359375
]
] |
[
"Split",
"train",
"valid",
"test",
"#Words",
"过~10%改善了RNN语言模型在定义建模任务上的困",
"20,298",
"1,127",
"1,129",
"#Entries",
"惑，而字符级CNN通过～5%进一步改善了困感。（3）",
"8,087",
"8,352",
"146.486",
"我们还表明，定义模型可用于执行先前工作中研究的反",
"# Tokens",
"77.948",
"79,699",
"1,406,440",
"向词典任务，其目标是将给定的定义与其相应的单词相",
"6.60",
"6.64",
"6.54",
"Avg length",
"匹配。我们的模型实现了11。8与Hill等人以前的最",
"新技术相比，%在精度上的绝对增益。（2016）。（4）最",
"后，我们的错误分析表明，训练有素的单词嵌入集对生",
"Table 2: Basic statistics of the common word definitions",
"成的定义的质量起着重要作用，并且某些错误类型表明",
"corpus. Splits are mutually exclusive in the words being",
"单词嵌入中编码的信息存在缺陷",
"defined.",
"2Previous Work",
"属+差异（G+D）结构不是定义的唯一模式。例如，",
"我们的目标是研究RNN模型，该模型通过训练字典定",
"下面的条目显示不同的结构。",
"义的示例来学习定义单词嵌入。尽管字典语料库已在",
"NLP中得到广泛使用，但据我们所知，以前的工作都没",
"Eradication : the act of plucking up by the roots;",
"有尝试创建定义的生成模型。早期的工作集中在来自定",
"a rooting out; extirpation:utter destruction",
"义的extracting语义信息上。例如，Chodorow（1985）",
"这组定义包括同义词（“extirppation\"），G+D结构的反",
"以及Klavans和Whitman（2001）构建了词典中单词",
"面（“utterdestruction\"）和未分类结构（“rootingout”）",
"的分类法。Dolan等人。（1993）和Vanderwende等人。",
"(2005）从定义中提取语义表示，以填充词汇知识库",
"3.2Corpus: Preprocessing and Analysis",
"在分布式语义中，单词由实数的密集向量表示，而不",
"是语义谓词。最近，字典定义已被用来学习这样的嵌",
"字典语料库以数字格式提供，但专为人类消费而设计，",
"人。例如，Wang等人。（2015）将定义文本中的单词用",
"需要预处理才能用于机器学习。字典包含非定义文本",
"以帮助人类读者，例如。g。Wordnik1中“渐变”的条",
"作Word2Vec算法（Mikolovetal.2013b）的“上下文”",
"单词形式。Hill等人。（2016）使用字典定义对组合性进",
"目包含字段（“数学”）和示例用法（“as，铁路的渐变",
"线。）。此外，许多条目包含多个定义，通常（但不总是）",
"行建模，并使用反向字典任务评估模型。虽然这些作品",
"用“”分隔。",
"从定义中学习单词或短语嵌人，但我们只关注从现有",
"我们需要一个语料库，其中每个条目仅包含一个定",
"（固定）嵌人生成定义。我们的实验表明，我们的模型优",
"义的单词和一个定义。我们从GCIDE2解析字典条",
"于希尔等人的模型。反向字典任务上的（2016",
"自，并预处理WordNet的光泽，并删除字段和用法。",
"我们的工作采用嵌人模型来生成自然语言。最近在与",
"我们不同的任务上的各种工作中也采取了类似的方法，",
"解析器和预处理脚本可以在https://github上找到。",
"Dinu 和Baroni(2014）提出了一种方法，该方法使用",
"com/northanapon/dict-definition",
"嵌入将单个单词映射到表示相同含义的更长短语。",
"为了为机器学习实验创建一个合理大小的语料库，我",
"同",
"样，Li等人。（2015）研究如何将段落或文档编码为嵌",
"们从GoogleWeb1T语料库（BrantsandFranz2006）中",
"人，并从编码中重建原始文本。其他最近的工作，如图",
"的50k个最常见的单词中采样了大约20k个单词，删",
"像字幕生成（KarpathyandFei-Fei2014）和口语对话生",
"除了功能单词。此外，在用“”拆分之前，我们将字典",
"成（Wenetal.2015a）也与我们的工作有关，因为单词",
"中每个单词的条目数限制为三个（以便定义的每个单词",
"序列是从单个输人向量生成的。我们的模型架构受到序",
"在我们的语料库中可以重复多次）。清理和修剪后，语",
"列到序列模型（Choetal.2014；Sutskever，Vinyals，and",
"料库的词汇量为29k。其他语料库统计见表2",
"Le2014）的启发，但是定义建模是截然不同的，因为它",
"我们通过使用以下四种结构之一手动标记每个定义",
"是word到序列的任务",
"来分析语料库中定义的基础结构：G+D，D+G，Sym",
"（同义词）和Misc。总的来说，我们从100个随机选择",
"3Dictionary Definitions",
"的单词中检查了680个定义。结果见表3。我们重申",
"早期的研究表明，G+D结构在这两个词典中占主导地",
"在本节中，我们首先通过对现有词典的研究来研究定义",
"位。但是，也存在其他结构，突出了字典建模任务中固",
"内容和结构。然后，我们描述我们的新数据集，并定义",
"有的挑战。此外，我们观察到属术语有时是通用的（e。",
"我们的任务和指标",
"g.,",
"“二个”或“那个”），以及其他特定时间（e。g。“倡",
"3.1 Definition Content and Structure",
"导者”）",
"在现有的词典中，单个定义通常由属和差异（Chodorow,",
"3.3Dictionary Definition Tasks",
"Byrd, and Heidorn 1985: Montemagni and Vander-",
"wende1992）组成。genus是所定义单词的广义类",
"在定义建模（DM）任务中，给我们一个输人单词w*",
"而differentiae使该单词与同一类中的其他单词不同。",
"并输出任何给定文本D作为输入单词定义的可能性。",
"例如，",
"换句话说，我们估计P(Dw*）。我们假设我们的定义",
"模型可以访问一组词嵌入，这些词嵌入是从用于训练定",
"Phosphorescent : emitting light without apprecia",
"义模型的定义语料库以外的一些语料库中估计的",
"ble heat as by slow oxidation of phosphorous",
"DM是语言建模的特例，与语言建模一样，定义模型",
"“发光”是一个属，并且“没有适用的热量。“是",
"一个不",
"的性能可以通过使用测试语料库的困惑来衡量。较低的",
"同的。此外，定义往往包括常见的模式，如“。“或”",
"个有。”(Markowitz,Ahlswede,and Evens 1986)。但",
"1https://www.wordnik.com/words/gradient",
"是，每个字典的模式和样式通常是唯一的",
"2http://gcide.gnu.org.ua/"
] |
{
"arxiv_id": "1612.00394",
"language": "zh",
"page_idx": 2
}
|
||
1612.00394-zh-3
|
1612.00394-zh-3.jpg
|
<table>
<tbody>
<tr>
<th>$\overline{\mathrm{Label}}$</th>
<th>$\overline{WN}$</th>
<th>$\overline{\mathrm{GC}}$</th>
<th>$\overline{{\mathrm{Example}}}$</th>
</tr>
<tr>
<td>$\overline{\mathrm{G+D}}$</td>
<td>$\overline{85\:\%}$</td>
<td>$\overline{50\:\%}$</td>
<td>to divide into thin plates</td>
</tr>
<tr>
<td>$D+G$</td>
<td>$\%$ $7^{(}$</td>
<td>$\%_0$ 9</td>
<td>deer a voung</td>
</tr>
<tr>
<td>Syn</td>
<td>$\%_0$ 1</td>
<td>$\%_{0}$ 32</td>
<td>energy</td>
</tr>
<tr>
<td>Misc. </td>
<td>$4\%$</td>
<td>$\%_{0}$ $8^{(}$ </td>
<td>in a diagonal $\operatorname{direction}$</td>
</tr>
<tr>
<td>Error</td>
<td>$3\%$</td>
<td>$\%_{0}$ 1</td>
<td>intensifiers used as</td>
</tr>
<tr>
<td>$\overline{\mathrm{Total}}$</td>
<td>$\overline{256}$</td>
<td>424</td>
<td> </td>
</tr>
</tbody>
</table>
Table 3: The number of manually labeled structures of dictionary defnitions in WordNet (WN) and GCIDE (GC). G+D is genus followed by differentiae , and D+G is the reverse. Syn is a synonym. The words defined in the example column are“laminate”,“fawn",“activity”, $“$diagonally", and“absolutely”.
困惑表明该模型在捕获定义结构和所定义单词的语义
方面更准确
除了困惑度测量之外，我们还可以使用其他任务来进一步评估字典定义模型，包括定义生成以及反向和正向字典任务。在定义生成中，模型生成测试词的定义。在我们的实验中，我们使用手动检查和 BLEU 分数来评估生成的定义，BLEU 分数是生成文本的自动指标。反向和正向字典任务是排名任务，其中定义模型根据一组测试单词对应于给定定义的可能性对其进行排名$({\mathrm{~reverse~dictionary~}}$任务)或对给定单词的一组测试定义进行排名( forward dictionary 任务)(Hill et al. 2016)。字典定义模型通过使用预测似然$P(D|w^*)$ 作为排名分数来实现这一点
定义模型的目标是在给定定义$w^*$的单词的情况下预测定义($D=[w_1,...,w_T]$)的概率。我们的模型假设生成定义文本的第$t$个单词$w_t$的概率取决于先前的单词和正在定义的单词 (Eq 1)。概率分布通常由 softmax 函数(Eq2)
(1)
$$\begin{aligned}&p(D|w^{*})=\prod_{t=1}^{T}p(w_{t}|w_{1},..,w_{t-1},w^{*})\\&p(w_{t}=j|w_{1},..,w_{t-1},w^{*})\propto e^{W_{j}h_{t}/\tau}\end{aligned}$$
(2)
近似，其中$W_j$是与单词$j$相关的参数矩阵，$h_t$是汇总到目前为止在令牌$t$上输入的向量，并且$\tau$是指定为温度的超参数，除非设置为。请注意，在我们的表达式中，定义为$w^*$的单词在所有时间步骤中都作为附加的条件变量存在
本文探讨的定义模型基于递归神经网络语言模型单元组成，其中每个单元在每个时间步$t$处读取一个单词$w_t$,并输出公式 2 的隐藏表示$h_t$ 。
$$h_t=g(v_{t-1},h_{t-1},v^*)$$
(3)
其中$g$是一个递归非线性函数，$v_t$表示单词$w_t$的嵌人
$(\bar{\text{向量表示}})$,而 $v^*$ 同样是所定义单词的嵌人
## 4.1 Model Architectures
调节 RNN 语言模型的一种自然方法是向网络提供在第一步定义的单词，作为“种子”信息的形式。种子方
法已被证明是有效的 RNN 的其他任务 (Kalchbrenner and Blunsom 2013; Karpathy and Fei-Fei 2014)。在这里，我们遵循 Sutskever 等人的简单方法。(2011) ,其中种子添加在文本的开头。在我们的例子中，被定义的单词被添加到定义的开头。请注意，我们忽略了种子本身在测试时的预测概率分布
3 节显示了定义具有常见的模式。我们假设所定义的单词对于携带语义信息的定义部分应该相对更重要，而对于由功能和停止词组成的模式或结构则不那么重要。此外，Wen 等人。(2015b) 表明，在每个时间步长提供恒定的种子输入会使口语对话生成器的整体性能恶化因此，受 GRU 更新门 (Cho et al. 2014)的启发，我们使用类似 GRU 的更新函数将循环单元的输出更新为：
(4) (5)
$$\begin{aligned}&z_{t}=\sigma(W_{z}[v^{*};h_{t}]+b_{z})\\&r_{t}=\sigma(W_{r}[v^{*};h_{t}]+b_{r})\\&\tilde{h_{t}}=tanh(W_{h}[(r_{t}\odot v^{*});h_{t}]+b_{h})\\&h_{t}=(1-z_{t})\odot h_{t}+z_{t}\odot\tilde{h_{t}}\end{aligned}$$
(6)
(7)
其中$\sigma$是 sigmoid 函数，$[a;b]$表示向量级联，$\odot$表示逐元素乘法。来自 Eq 3的$h_t$如 Eq 7 中所示进行更新。在每个时间步长，$z_t$是一个更新门，用于控制RNN 单元的输出变化量，而$r_t$是一个重置门，用于控制允许定义的单词中有多少信息。我们将此模型命名为$\operatorname{Gated}($G)
在本小节的其余部分中，我们将介绍三个基线模型体系结构，它们删除了 Gated 的部分。在我们的实验中我们将比较 Gated 的性能与基线，以衡量我们架构的每个部分的贡献。首先，我们将模型简化为标准 RNNLM, 其中
(8)
$$h_t=g(v_{t-1},h_{t-1})$$
标准模型 only 在第一步 (从种子)接收有关$w^*$的信
息。我们将此基线称为种子(S)
将整个定义中定义的单词合并的一种简单方法是简单地将其嵌人$v^*$作为每个时间步的常量输入。我们将$\bar{\text{此模型称为 Input(I)}}:$
(9)
$$h_t=g([v^*;v_{t-1}],h_{t-1})$$
(Mikolov and Zweig 2012)。或者，该模型可以利用$v^*$ 来更新来自 RNN 单元的隐藏表示，名为 Hidden (H)。Hidden 的更新函数是：
$$h_t=tanh(W_h[v^*;h_t]+b_h)$$
(10)
其中$W_h$ 是权重矩阵，$b_h$ 是偏置。在 Hidden 中，我们使用 Eq 10 从 Eq 3 更新$h_t$ 。这类似于 Eq 7 中类似GRU 的架构，没有门 (i。e。$r_t$和$z_t$始终是 1s 的向量)
4.2 Other Features
除了模型体系结构之外，我们还探讨了从正在定义的单词派生的其他单词特征是否可以为单词嵌人提供补充信息。我们专注于两个不同的功能：词缀和 hypernym 嵌入。要在 DM 中添加这些功能，我们只需将嵌人$v^*$ 与其他特征向量连接
|
[
[
0.101318359375,
0.0662841796875,
0.1502685546875,
0.06817626953125,
0.1502685546875,
0.0802001953125,
0.101318359375,
0.07763671875
],
[
0.1748046875,
0.06756591796875,
0.2091064453125,
0.06756591796875,
0.2091064453125,
0.0802001953125,
0.1748046875,
0.0802001953125
],
[
0.233642578125,
0.06817626953125,
0.261474609375,
0.06817626953125,
0.261474609375,
0.0802001953125,
0.233642578125,
0.0802001953125
],
[
0.27685546875,
0.06817626953125,
0.34228515625,
0.06817626953125,
0.34228515625,
0.0802001953125,
0.27685546875,
0.0802001953125
],
[
0.52294921875,
0.07135009765625,
0.91015625,
0.07135009765625,
0.91015625,
0.08209228515625,
0.52294921875,
0.08209228515625
],
[
0.1070556640625,
0.080810546875,
0.1519775390625,
0.080810546875,
0.1519775390625,
0.0953369140625,
0.1070556640625,
0.0953369140625
],
[
0.1666259765625,
0.080810546875,
0.2091064453125,
0.080810546875,
0.2091064453125,
0.0953369140625,
0.1666259765625,
0.0953369140625
],
[
0.221435546875,
0.07891845703125,
0.261474609375,
0.080810546875,
0.2607421875,
0.09661865234375,
0.2197265625,
0.0947265625
],
[
0.27685546875,
0.08209228515625,
0.4599609375,
0.08209228515625,
0.4599609375,
0.0953369140625,
0.27685546875,
0.0953369140625
],
[
0.51953125,
0.08209228515625,
0.9111328125,
0.08331298828125,
0.9111328125,
0.09783935546875,
0.51953125,
0.09661865234375
],
[
0.108642578125,
0.0953369140625,
0.1502685546875,
0.0953369140625,
0.1502685546875,
0.10986328125,
0.108642578125,
0.10986328125
],
[
0.176513671875,
0.0953369140625,
0.2059326171875,
0.0953369140625,
0.2059326171875,
0.10797119140625,
0.176513671875,
0.10797119140625
],
[
0.2279052734375,
0.0947265625,
0.261474609375,
0.0947265625,
0.261474609375,
0.10986328125,
0.2279052734375,
0.10986328125
],
[
0.27783203125,
0.09783935546875,
0.371826171875,
0.09783935546875,
0.371826171875,
0.10919189453125,
0.27783203125,
0.10919189453125
],
[
0.52294921875,
0.09783935546875,
0.9111328125,
0.09783935546875,
0.9111328125,
0.111083984375,
0.52294921875,
0.111083984375
],
[
0.176513671875,
0.10919189453125,
0.20751953125,
0.10919189453125,
0.20751953125,
0.121826171875,
0.176513671875,
0.121826171875
],
[
0.221435546875,
0.10919189453125,
0.261474609375,
0.10919189453125,
0.261474609375,
0.12371826171875,
0.221435546875,
0.12371826171875
],
[
0.279296875,
0.10986328125,
0.325927734375,
0.11236572265625,
0.324462890625,
0.125,
0.27783203125,
0.12249755859375
],
[
0.521484375,
0.10986328125,
0.9111328125,
0.111083984375,
0.9111328125,
0.1256103515625,
0.521484375,
0.125
],
[
0.17724609375,
0.121826171875,
0.20751953125,
0.12371826171875,
0.2059326171875,
0.136962890625,
0.176513671875,
0.1351318359375
],
[
0.11767578125,
0.10986328125,
0.148681640625,
0.10986328125,
0.148681640625,
0.12249755859375,
0.11767578125,
0.12249755859375
],
[
0.1070556640625,
0.12249755859375,
0.148681640625,
0.12249755859375,
0.148681640625,
0.136962890625,
0.1070556640625,
0.136962890625
],
[
0.2296142578125,
0.12371826171875,
0.2607421875,
0.12371826171875,
0.2607421875,
0.1363525390625,
0.2296142578125,
0.1363525390625
],
[
0.27685546875,
0.12371826171875,
0.439453125,
0.12371826171875,
0.439453125,
0.136962890625,
0.27685546875,
0.136962890625
],
[
0.52294921875,
0.1256103515625,
0.91259765625,
0.1256103515625,
0.91259765625,
0.1395263671875,
0.52294921875,
0.1395263671875
],
[
0.1070556640625,
0.136962890625,
0.1502685546875,
0.136962890625,
0.1502685546875,
0.1514892578125,
0.1070556640625,
0.1514892578125
],
[
0.17724609375,
0.136962890625,
0.2059326171875,
0.136962890625,
0.2059326171875,
0.149658203125,
0.17724609375,
0.149658203125
],
[
0.2296142578125,
0.136962890625,
0.2607421875,
0.136962890625,
0.2607421875,
0.149658203125,
0.2296142578125,
0.149658203125
],
[
0.27685546875,
0.1383056640625,
0.416748046875,
0.1383056640625,
0.416748046875,
0.1514892578125,
0.27685546875,
0.1514892578125
],
[
0.52294921875,
0.1383056640625,
0.71826171875,
0.1383056640625,
0.71826171875,
0.1514892578125,
0.52294921875,
0.1514892578125
],
[
0.10540771484375,
0.15087890625,
0.15283203125,
0.15087890625,
0.15283203125,
0.1654052734375,
0.10540771484375,
0.1654052734375
],
[
0.176513671875,
0.15087890625,
0.221435546875,
0.15087890625,
0.221435546875,
0.1654052734375,
0.176513671875,
0.1654052734375
],
[
0.233642578125,
0.1514892578125,
0.2646484375,
0.1514892578125,
0.2646484375,
0.1641845703125,
0.233642578125,
0.1641845703125
],
[
0.5361328125,
0.1546630859375,
0.9111328125,
0.1546630859375,
0.9111328125,
0.1685791015625,
0.5361328125,
0.1685791015625
],
[
0.52294921875,
0.1685791015625,
0.9111328125,
0.1685791015625,
0.9111328125,
0.1817626953125,
0.52294921875,
0.1817626953125
],
[
0.08660888671875,
0.1787109375,
0.4794921875,
0.1787109375,
0.4794921875,
0.1925048828125,
0.08660888671875,
0.1925048828125
],
[
0.52294921875,
0.184326171875,
0.90869140625,
0.184326171875,
0.90869140625,
0.1944580078125,
0.52294921875,
0.1944580078125
],
[
0.0849609375,
0.1912841796875,
0.47802734375,
0.1900634765625,
0.47802734375,
0.2052001953125,
0.0849609375,
0.205810546875
],
[
0.521484375,
0.1962890625,
0.91259765625,
0.1962890625,
0.91259765625,
0.210205078125,
0.521484375,
0.210205078125
],
[
0.0882568359375,
0.205810546875,
0.47802734375,
0.205810546875,
0.47802734375,
0.2197265625,
0.0882568359375,
0.2197265625
],
[
0.521484375,
0.210205078125,
0.896484375,
0.210205078125,
0.896484375,
0.2235107421875,
0.521484375,
0.2235107421875
],
[
0.08660888671875,
0.2203369140625,
0.4794921875,
0.2203369140625,
0.4794921875,
0.2342529296875,
0.08660888671875,
0.2342529296875
],
[
0.53515625,
0.2235107421875,
0.9111328125,
0.2247314453125,
0.9111328125,
0.2392578125,
0.53515625,
0.238037109375
],
[
0.08660888671875,
0.23291015625,
0.47802734375,
0.23291015625,
0.47802734375,
0.2474365234375,
0.08660888671875,
0.2474365234375
],
[
0.52294921875,
0.2392578125,
0.9111328125,
0.2392578125,
0.9111328125,
0.25244140625,
0.52294921875,
0.25244140625
],
[
0.08905029296875,
0.25,
0.30810546875,
0.25,
0.30810546875,
0.260009765625,
0.08905029296875,
0.260009765625
],
[
0.52294921875,
0.25390625,
0.5458984375,
0.25390625,
0.5458984375,
0.266357421875,
0.52294921875,
0.266357421875
],
[
0.6005859375,
0.279052734375,
0.76318359375,
0.279052734375,
0.76318359375,
0.29345703125,
0.6005859375,
0.29345703125
],
[
0.8896484375,
0.279052734375,
0.91259765625,
0.279052734375,
0.91259765625,
0.29345703125,
0.8896484375,
0.29345703125
],
[
0.0882568359375,
0.28857421875,
0.47802734375,
0.28857421875,
0.47802734375,
0.3017578125,
0.0882568359375,
0.3017578125
],
[
0.5986328125,
0.296630859375,
0.76318359375,
0.296630859375,
0.76318359375,
0.31005859375,
0.5986328125,
0.31005859375
],
[
0.08905029296875,
0.302978515625,
0.1683349609375,
0.302978515625,
0.1683349609375,
0.314453125,
0.08905029296875,
0.314453125
],
[
0.8896484375,
0.29736328125,
0.91259765625,
0.29736328125,
0.91259765625,
0.311279296875,
0.8896484375,
0.311279296875
],
[
0.10455322265625,
0.317626953125,
0.476318359375,
0.317626953125,
0.476318359375,
0.32763671875,
0.10455322265625,
0.32763671875
],
[
0.59619140625,
0.31640625,
0.833984375,
0.317626953125,
0.833984375,
0.33203125,
0.59619140625,
0.330810546875
],
[
0.8896484375,
0.318115234375,
0.91259765625,
0.318115234375,
0.91259765625,
0.33203125,
0.8896484375,
0.33203125
],
[
0.1004638671875,
0.330078125,
0.47802734375,
0.330078125,
0.47802734375,
0.343505859375,
0.1004638671875,
0.343505859375
],
[
0.59716796875,
0.33837890625,
0.79150390625,
0.33837890625,
0.79150390625,
0.352783203125,
0.59716796875,
0.352783203125
],
[
0.8896484375,
0.339111328125,
0.91259765625,
0.339111328125,
0.91259765625,
0.352783203125,
0.8896484375,
0.352783203125
],
[
0.0882568359375,
0.35791015625,
0.476318359375,
0.35791015625,
0.476318359375,
0.371826171875,
0.0882568359375,
0.371826171875
],
[
0.08905029296875,
0.343505859375,
0.47802734375,
0.343505859375,
0.47802734375,
0.356689453125,
0.08905029296875,
0.356689453125
],
[
0.52392578125,
0.365478515625,
0.91259765625,
0.3662109375,
0.91259765625,
0.380126953125,
0.52392578125,
0.37890625
],
[
0.0882568359375,
0.371826171875,
0.47802734375,
0.371826171875,
0.47802734375,
0.385009765625,
0.0882568359375,
0.385009765625
],
[
0.08660888671875,
0.3837890625,
0.476318359375,
0.38330078125,
0.476318359375,
0.397705078125,
0.08660888671875,
0.3984375
],
[
0.521484375,
0.380126953125,
0.9111328125,
0.380126953125,
0.9111328125,
0.393310546875,
0.521484375,
0.393310546875
],
[
0.098876953125,
0.3984375,
0.4755859375,
0.3984375,
0.4755859375,
0.412353515625,
0.098876953125,
0.412353515625
],
[
0.52294921875,
0.392578125,
0.9111328125,
0.392578125,
0.9111328125,
0.4072265625,
0.52294921875,
0.4072265625
],
[
0.51953125,
0.406005859375,
0.9111328125,
0.4072265625,
0.9111328125,
0.421630859375,
0.51953125,
0.42041015625
],
[
0.08660888671875,
0.41357421875,
0.47802734375,
0.41357421875,
0.47802734375,
0.4267578125,
0.08660888671875,
0.4267578125
],
[
0.08905029296875,
0.4267578125,
0.476318359375,
0.4267578125,
0.476318359375,
0.43994140625,
0.08905029296875,
0.43994140625
],
[
0.52294921875,
0.421630859375,
0.91259765625,
0.421630859375,
0.91259765625,
0.43505859375,
0.52294921875,
0.43505859375
],
[
0.08660888671875,
0.43994140625,
0.47802734375,
0.43994140625,
0.47802734375,
0.453857421875,
0.08660888671875,
0.453857421875
],
[
0.52294921875,
0.436279296875,
0.5888671875,
0.436279296875,
0.5888671875,
0.447509765625,
0.52294921875,
0.447509765625
],
[
0.08660888671875,
0.455078125,
0.26708984375,
0.455078125,
0.26708984375,
0.468505859375,
0.08660888671875,
0.468505859375
],
[
0.5390625,
0.451904296875,
0.91015625,
0.451904296875,
0.91015625,
0.462158203125,
0.5390625,
0.462158203125
],
[
0.52392578125,
0.464111328125,
0.91015625,
0.464111328125,
0.91015625,
0.477783203125,
0.52392578125,
0.477783203125
],
[
0.2320556640625,
0.479736328125,
0.330810546875,
0.481689453125,
0.330810546875,
0.496826171875,
0.2320556640625,
0.494384765625
],
[
0.521484375,
0.4765625,
0.9111328125,
0.477783203125,
0.9111328125,
0.492431640625,
0.521484375,
0.4912109375
],
[
0.52294921875,
0.492431640625,
0.91259765625,
0.492431640625,
0.91259765625,
0.505859375,
0.52294921875,
0.505859375
],
[
0.0882568359375,
0.50048828125,
0.476318359375,
0.50048828125,
0.476318359375,
0.5146484375,
0.0882568359375,
0.5146484375
],
[
0.52294921875,
0.505859375,
0.5537109375,
0.505859375,
0.5537109375,
0.51806640625,
0.52294921875,
0.51806640625
],
[
0.0882568359375,
0.5146484375,
0.47802734375,
0.5146484375,
0.47802734375,
0.52880859375,
0.0882568359375,
0.52880859375
],
[
0.0849609375,
0.5244140625,
0.47802734375,
0.5263671875,
0.47802734375,
0.54443359375,
0.0849609375,
0.54248046875
],
[
0.65185546875,
0.52587890625,
0.78125,
0.52587890625,
0.78125,
0.5390625,
0.65185546875,
0.5390625
],
[
0.890625,
0.52587890625,
0.91259765625,
0.52587890625,
0.91259765625,
0.5390625,
0.890625,
0.5390625
],
[
0.08905029296875,
0.54248046875,
0.47802734375,
0.54248046875,
0.47802734375,
0.55615234375,
0.08905029296875,
0.55615234375
],
[
0.51953125,
0.54736328125,
0.9111328125,
0.548828125,
0.9111328125,
0.56298828125,
0.51953125,
0.5625
],
[
0.0882568359375,
0.556640625,
0.1666259765625,
0.556640625,
0.1666259765625,
0.568359375,
0.0882568359375,
0.568359375
],
[
0.52294921875,
0.56298828125,
0.7490234375,
0.56298828125,
0.7490234375,
0.5771484375,
0.52294921875,
0.5771484375
],
[
0.2418212890625,
0.5771484375,
0.256591796875,
0.5771484375,
0.256591796875,
0.5869140625,
0.2418212890625,
0.5869140625
],
[
0.53759765625,
0.5791015625,
0.9111328125,
0.5791015625,
0.9111328125,
0.59228515625,
0.53759765625,
0.59228515625
],
[
0.1544189453125,
0.58837890625,
0.410888671875,
0.58837890625,
0.410888671875,
0.609375,
0.1544189453125,
0.609375
],
[
0.455810546875,
0.59228515625,
0.47802734375,
0.59228515625,
0.47802734375,
0.60498046875,
0.455810546875,
0.60498046875
],
[
0.52294921875,
0.59228515625,
0.9111328125,
0.59228515625,
0.9111328125,
0.60595703125,
0.52294921875,
0.60595703125
],
[
0.23779296875,
0.609375,
0.25732421875,
0.609375,
0.25732421875,
0.6171875,
0.23779296875,
0.6171875
],
[
0.521484375,
0.60498046875,
0.67333984375,
0.60693359375,
0.67333984375,
0.62060546875,
0.521484375,
0.61865234375
],
[
0.15283203125,
0.625,
0.41015625,
0.61865234375,
0.41015625,
0.63623046875,
0.15283203125,
0.642578125
],
[
0.455810546875,
0.62353515625,
0.47802734375,
0.62353515625,
0.47802734375,
0.63818359375,
0.455810546875,
0.63818359375
],
[
0.6337890625,
0.630859375,
0.79736328125,
0.6318359375,
0.79736328125,
0.646484375,
0.6337890625,
0.64599609375
],
[
0.8896484375,
0.6318359375,
0.91259765625,
0.6318359375,
0.91259765625,
0.646484375,
0.8896484375,
0.646484375
],
[
0.0882568359375,
0.64794921875,
0.47802734375,
0.64794921875,
0.47802734375,
0.6611328125,
0.0882568359375,
0.6611328125
],
[
0.08905029296875,
0.6611328125,
0.47802734375,
0.6611328125,
0.47802734375,
0.67529296875,
0.08905029296875,
0.67529296875
],
[
0.52294921875,
0.65771484375,
0.9111328125,
0.65771484375,
0.9111328125,
0.671875,
0.52294921875,
0.671875
],
[
0.0882568359375,
0.67529296875,
0.476318359375,
0.67529296875,
0.476318359375,
0.689453125,
0.0882568359375,
0.689453125
],
[
0.52294921875,
0.671875,
0.91259765625,
0.671875,
0.91259765625,
0.68603515625,
0.52294921875,
0.68603515625
],
[
0.08660888671875,
0.689453125,
0.476318359375,
0.68798828125,
0.476318359375,
0.70263671875,
0.08660888671875,
0.7041015625
],
[
0.5185546875,
0.6845703125,
0.6845703125,
0.68505859375,
0.6845703125,
0.69970703125,
0.5185546875,
0.69873046875
],
[
0.08905029296875,
0.705078125,
0.1854248046875,
0.705078125,
0.1854248046875,
0.71630859375,
0.08905029296875,
0.71630859375
],
[
0.62109375,
0.7119140625,
0.81396484375,
0.7119140625,
0.81396484375,
0.72607421875,
0.62109375,
0.72607421875
],
[
0.8798828125,
0.7109375,
0.91259765625,
0.7109375,
0.91259765625,
0.7265625,
0.8798828125,
0.7265625
],
[
0.10296630859375,
0.71728515625,
0.476318359375,
0.71728515625,
0.476318359375,
0.73095703125,
0.10296630859375,
0.73095703125
],
[
0.08660888671875,
0.73095703125,
0.476318359375,
0.73095703125,
0.476318359375,
0.744140625,
0.08660888671875,
0.744140625
],
[
0.52392578125,
0.7373046875,
0.9111328125,
0.7373046875,
0.9111328125,
0.75048828125,
0.52392578125,
0.75048828125
],
[
0.08905029296875,
0.74560546875,
0.476318359375,
0.74560546875,
0.476318359375,
0.7587890625,
0.08905029296875,
0.7587890625
],
[
0.52392578125,
0.751953125,
0.9111328125,
0.751953125,
0.9111328125,
0.76513671875,
0.52392578125,
0.76513671875
],
[
0.08660888671875,
0.75830078125,
0.36767578125,
0.7587890625,
0.36767578125,
0.7734375,
0.08660888671875,
0.77294921875
],
[
0.521484375,
0.76513671875,
0.9111328125,
0.76513671875,
0.9111328125,
0.77880859375,
0.521484375,
0.77880859375
],
[
0.2059326171875,
0.7822265625,
0.358642578125,
0.7822265625,
0.358642578125,
0.79541015625,
0.2059326171875,
0.79541015625
],
[
0.455810546875,
0.7822265625,
0.47802734375,
0.7822265625,
0.47802734375,
0.79541015625,
0.455810546875,
0.79541015625
],
[
0.521484375,
0.796875,
0.68896484375,
0.796875,
0.68896484375,
0.81005859375,
0.521484375,
0.81005859375
],
[
0.08905029296875,
0.8037109375,
0.476318359375,
0.8037109375,
0.476318359375,
0.8173828125,
0.08905029296875,
0.8173828125
],
[
0.0849609375,
0.818359375,
0.404296875,
0.818359375,
0.404296875,
0.83203125,
0.0849609375,
0.83203125
],
[
0.52392578125,
0.8212890625,
0.91015625,
0.8212890625,
0.91015625,
0.83203125,
0.52392578125,
0.83203125
],
[
0.52294921875,
0.833984375,
0.9111328125,
0.833984375,
0.9111328125,
0.84765625,
0.52294921875,
0.84765625
],
[
0.0882568359375,
0.84326171875,
0.294189453125,
0.84326171875,
0.294189453125,
0.85693359375,
0.0882568359375,
0.85693359375
],
[
0.52294921875,
0.84765625,
0.9111328125,
0.84765625,
0.9111328125,
0.861328125,
0.52294921875,
0.861328125
],
[
0.0882568359375,
0.861328125,
0.476318359375,
0.861328125,
0.476318359375,
0.875,
0.0882568359375,
0.875
],
[
0.521484375,
0.861328125,
0.90869140625,
0.861328125,
0.90869140625,
0.87548828125,
0.521484375,
0.87548828125
],
[
0.0882568359375,
0.87548828125,
0.476318359375,
0.87548828125,
0.476318359375,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.52294921875,
0.87548828125,
0.6689453125,
0.87548828125,
0.6689453125,
0.8896484375,
0.52294921875,
0.8896484375
]
] |
[
"Label",
"WN",
"GC",
"Example",
"法已被证明是有效的RNN的其他任务（Kalchbrenner",
"G+D",
"85%",
"50%",
"to divide into thin plates",
"and Blunsom 2013;Karpathy and Fei-Fei 2014)。在这",
"D+G",
"7 %",
"9 %",
"a young deer",
"里，我们遵循Sutskever等人的简单方法。，（2011），其",
"1 %",
"32 %",
"energy",
"中种子添加在文本的开头。在我们的例子中，被定义的",
"4%",
"Syn",
"Misc.",
"8%",
"in a diagonal direction",
"单词被添加到定义的开头。请注意，我们忽略了种子本",
"Error",
"3 %",
"1 %",
"used as intensifiers",
"身在测试时的预测概率分布",
"Total",
"256",
"424",
"3节显示了定义具有常见的模式。我们假设所定义的",
"单词对于携带语义信息的定义部分应该相对更重要，而",
"Table 3: The number of manually labeled structures of",
"对于由功能和停正词组成的模式或结构则不那么重要",
"dictionary definitions in WordNet (WN) and GCIDE",
"此外，Wen等人。（2015b）表明，在每个时间步长提供",
"(GC). G+D is genus followed by differentiae , and D+G",
"恒定的种子输人会使口语对话生成器的整体性能恶化",
"is the reverse. Syn is a synonym. The words defined in",
"因此，受GRU更新门（Choetal.2014）的启发，我",
"the example column are\"laminate\",\"fawn\",“activity\",",
"们使用类似GRU的更新函数将循环单元的输出更新",
"\"diagonally\",and\"absolutely\".",
"为：",
"zt = o(Wz[u*; ht] + bz)",
"(4)",
"困惑表明该模型在捕获定义结构和所定义单词的语义",
"rt = a(Wr[u*; htl + br)",
"方面更准确",
"(5)",
"除了困惑度测量之外，我们还可以使用其他任务来进",
"ht = tanh(Wh[(rt *); htl + bh)",
"(6)",
"步评估字典定义模型，包括定义生成以及反向和正",
"ht=(1-z)oht+zoht",
"(7)",
"在我们的实验中，我们使用手动检查和BLEU分数来",
"向字典任务。在定义生成中，模型生成测试词的定义。",
"，其中是sigmoid函数，[a;b]表示向量级联，表",
"评估生成的定义，BLEU分数是生成文本的自动指标。",
"反向和正向字典任务是排名任务，其中定义模型根据",
"示逐元素乘法。来自Eq3的ht如Eq7中所示进行",
"组测试单词对应于给定定义的可能性对其进行排名",
"更新。在每个时间步长，z是一个更新门，用于控制",
"RNN单元的输出变化量，而rt是一个重置门，用于控",
"（reversedictionary任务）或对给定单词的一组测试",
"定义进行排名（forwarddictionary任务）（Hilletal",
"制允许定义的单词中有多少信息。我们将此模型命名为",
"2016）。字典定义模型通过使用预测似然P(Dw*）作",
"Gated(G)",
"为排名分数来实现这一点",
"在本小节的其余部分中，我们将介绍三个基线模型体",
"系结构，它们删除了Gated的部分。在我们的实验中",
"4Models",
"我们将比较Gated的性能与基线，以衡量我们架构的每",
"个部分的贡献。首先，我们将模型简化为标准RNNLM，",
"定义模型的目标是在给定定义w*的单词的情况下预测",
"其中",
"定义（D=[w1,…,WT]）的概率。我们的模型假设生成",
"定义文本的第t个单词wt的概率取决于先前的单词和",
"ht = g(Ut-1,ht-1)",
"(8)",
"正在定义的单词（Eq1）。概率分布通常由softmax函",
"标准模型only在第一步（从种子）接收有关w*的信",
"数（Eq2）",
"息。我们将此基线称为种子（S)",
"T",
"将整个定义中定义的单词合并的一种简单方法是简",
"p(Dw*) = IIp(wt/w1,.. w-1, w*)",
"(1)",
"单地将其嵌人*作为每个时间步的常量输人。我们将",
"t=l",
"此模型称为Input(I)：",
"p(wt = jlwi, ., wt-1, w*) αew,ht/t",
"(2)",
"ht = g([u*; Ut-il, ht-1)",
"(9)",
"近似，其中W是与单词i相关的参数矩阵，ht是汇",
"总到目前为止在令牌t上输入的向量，并且T是指定为",
"(MikolovandZweig2012）。或者，该模型可以利用*",
"温度的超参数，除非设置为。请注意，在我们的表达式",
"来更新来自RNN单元的隐藏表示，名为Hidden（H)。",
"中，定义为w*的单词在所有时间步骤中都作为附加的",
"Hidden的更新函数是：",
"条件变量存在",
"ht = tanh(Wh[u*; htl + bh)",
"(10)",
"本文探讨的定义模型基于递归神经网络语言模型",
"(RNNLM) (Mikolov et al.2010)。RNNLM 由 RNN",
"其中Wh是权重矩阵，bh是偏置。在Hidden中，我们",
"单元组成，其中每个单元在每个时间步t处读取一个单",
"使用Eq10从Eg3更新ht。这类似于Eq7中类似",
"词wt，并输出公式2的隐藏表示ht。",
"GRU的架构，没有门 (i。e。rt 和 zt 始终是 1s的向量)",
"ht = g(vt-1,ht-1,u*)",
"(3)",
"4.2Other Features",
"其中g是一个递归非线性函数，Ut表示单词Wt的嵌人",
"（向量表示），而*同样是所定义单词的嵌入",
"除了模型体系结构之外，我们还探讨了从正在定义的单",
"词派生的其他单词特征是否可以为单词嵌入提供补充",
"4.1Model Architectures",
"信息。我们专注于两个不同的功能：词缀和hypernym",
"调节RNN语言模型的一种自然方法是向网络提供在",
"嵌人。要在DM中添加这些功能，我们只需将嵌入",
"第一步定义的单词，作为“种子”信息的形式。种子方",
"与其他特征向量连接"
] |
{
"arxiv_id": "1612.00394",
"language": "zh",
"page_idx": 3
}
|
||
1612.00394-zh-4
|
1612.00394-zh-4.jpg
|
Affixes 英语和其他语言中的许多单词都由组成的语素组成。例如，单词“capitalist”包含一个词根“capital” 和一个后缀“-ist”。知道给定词根的语义以及词缀如何修改含义的模型可以准确地定义词根的任何形态变体。但是，自动将单词分解为语素并推导词缀的语义并不是一件容易的事
我们尝试通过使用字符级信息来捕获单词中的前缀和后缀。我们采用字符级卷积网络来检测词缀(LeCun et al. 1990) 。具体来说，$w^*$表示为具有独热编码的字符序列。在左侧和右侧添加一个填充字符，以指示单词的开头和结尾。然后，我们在字符序列上应用多个不同长度的内核，并使用最大池化创建最终功能(Kim et al. 2016)。我们假设表示为 CH 的卷积输人将允许模型识别词缀如何改变单词含义的规律性
Hypernym Embeddings 正如我们在 3 节中讨论的那样，字典定义通常遵循属+差异的结构。我们试图通过向模型提供从 hypernym 关系数据库中提取的适当属的知识来利用这种结构。特别是，我们从 WebIsA 数据库 (Seitner et al. 2016) 中获得了 hypernym,该数据库采用类似赫斯特的模式 (Hearst 1992) 从 Web 中提取hypernym 关系。然后，我们为模型提供一个附加的输人向量，称为 HE, 该向量等于数据库中定义的单词的顶级$k$ hypernym 的加权和。在我们的实验中，$k=5$ 和权重与资源中的频率成线性比例。例如，“小鹿”的前 5个 hypernym 和频率是“color”:149,“deer”: $135,“$animal”: 132。0,“野生动物”: 82。0,“年轻”: 68。0。
5 Experiments and Results
我们现在介绍评估定义模型的实验。我们使用训练集训练多个模型体系结构，并使用测试集对 3.3 节中描述的所有三个任务进行模型评估。我们使用有效集来搜索学习超参数。请注意，所定义的单词在三个集合中是互斥的，因此我们的实验评估了模型对新单词的概括程度， 而不是对相同单词的其他定义或意义的概括程度
所有模型都使用来自 Word2Vec 项目 $^{3}$和 2层 LSTM 网络的同一组固定的预训练单词嵌人作为 RNN 组件(Hochreiter and Schmidhuber 1997)。嵌人层和 LSTM 隐藏层各有 300 个单元。对于词缀检测器，字符级 CNN 的内核长度为 2-6, 大小为 $\{10、30、40、40、40\}$,步幅为1。在训练过程中，我们使用 Adam 最大化对数似然目标，Adam 是随机梯度的变体 (Kingma and Ba 2014)。学习率为 0。001 ,并且在连续 4 个历元后停止训练 ,验证性能没有明显改善。我们实验的源代码和数据集可以在https://github 上找到。com/websail-nu/torch-defseq
## 5.1 Definition Modeling
首先，我们比较了使用模型中定义的单词的不同方法。结果如表 4 的第一部分所示。我们看到门控更新 (S+G) 提高了种子的性能，而其他架构(S+I 和 S+H)则没有。结果与我们的假设一致，即被定义的单词与定义中的某些单词比与其他单词更相关，并且门更新可以识别这一点。我们在 6 节中进一步探讨了门的行为
接下来，我们评估语言特征的贡献。我们看到词缀(S+G+CH)进一步改进了模型，表明字符级信息可以补充从上下文中学到的单词嵌人。也许令人惊讶的是，
$^{3}$https://code.google.com/archive/p/word2vec/
<table>
<tbody>
<tr>
<th>Model</th>
<th>$\#$ ${\overline{{\mathrm{Params}}}}$ $\frac{1}{n}=\frac{1}{n}$</th>
<th>${\overline{{\mathrm{Perplexity}}}}$</th>
</tr>
<tr>
<td>Seed</td>
<td>10.2m</td>
<td>56.35C</td>
</tr>
<tr>
<td>$S+I$</td>
<td>10.6m</td>
<td>57.372</td>
</tr>
<tr>
<td>$S+G$</td>
<td>10.8m</td>
<td>50</td>
</tr>
<tr>
<td>S+ G+ CH+ HE</td>
<td>11.7m</td>
<td>48.168</td>
</tr>
</tbody>
</table>
Table 4: Perplexity evaluated on dictionary entries in
the test set (lower is better).
<table>
<tbody>
<tr>
<th>Model</th>
<th>$\overline{\mathrm{GC}}$</th>
<th>$\overline{WN}$</th>
<th>$\overline{\mathrm{Avg}}$</th>
<th>$\overline{\mathrm{Merged}}$</th>
</tr>
<tr>
<td>$\overline{\mathrm{Inter}}$</td>
<td>$\overline{27.90}$</td>
<td>$\overline{21.15}$</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$NE$</td>
<td>29.56</td>
<td>21.42</td>
<td>25.49</td>
<td>34.51</td>
</tr>
<tr>
<td>NE- WN</td>
<td>22.70</td>
<td>27.42</td>
<td>25.06</td>
<td>32.16</td>
</tr>
<tr>
<td>NE- GC</td>
<td>33.22</td>
<td>17.77</td>
<td>25.49</td>
<td>35.45</td>
</tr>
<tr>
<td>$\overline{\mathrm{Seed}}$</td>
<td>26.69</td>
<td>$\overline{22.46}$</td>
<td>$\overline{24.58}$</td>
<td>30.46</td>
</tr>
<tr>
<td>$S+I$</td>
<td>28.44</td>
<td>21.77</td>
<td>25.10</td>
<td>31.58</td>
</tr>
<tr>
<td>$S+H$</td>
<td>27.43</td>
<td>18.82</td>
<td>23.13</td>
<td>29.66</td>
</tr>
<tr>
<td>$S+G$</td>
<td>30.86</td>
<td>23.15</td>
<td>27.01</td>
<td>34.72</td>
</tr>
<tr>
<td>S+ G+ CH</td>
<td>31.12</td>
<td>$\overline{24.11}$</td>
<td>$\overline{27.62}$</td>
<td>35.78</td>
</tr>
<tr>
<td>S+ G+ CH+ HE</td>
<td>31.10</td>
<td>23.81</td>
<td>27.46</td>
<td>35.28</td>
</tr>
<tr>
<td>Additional $\operatorname{experi}$</td>
<td>ments</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\overline{\mathrm{Seed}^*}$</td>
<td>27.24</td>
<td>22.78</td>
<td>$\overline{25.01}$</td>
<td>31.15</td>
</tr>
<tr>
<td>$\mathrm{S+G+CH+HE}^{*}$</td>
<td>33.39</td>
<td>25.91</td>
<td>29.65</td>
<td>38.35</td>
</tr>
<tr>
<td>Random Emb</td>
<td>22.09</td>
<td>20.05</td>
<td>21.07</td>
<td>24.77</td>
</tr>
</tbody>
</table>
Table 5: Equally-weighted BLEU scores for up to 4- grams, on definitions evaluated using different reference dictionaries (results are not comparable between $\operatorname{columns}).$
hypernym 嵌人(S+G+CH+HE) 对性能的贡献不清楚。我们怀疑 hypernym 词的多个嵌人的平均值可能是定义中属的较差表示。利用 hypernyms 的更复杂的方法是末来的工作
# 5.2 Definition Generation
在本实验中，我们评估了模型生成的定义的质量。我们计算输出定义和字典定义之间的 BLEU 分数，以衡量生成的质量。解码算法只是从模型预测的单词概率分布中一次采样一个令牌。我们使用接近贪婪算法的温度(公式 2 中的 $\tau$ )为每个定义的单词采样 40 个定义(0。05或0。1,从有效集中选择)并报告平均 BLEU 得分。为了帮助解释 BLEU 分数，我们还报告了输出在训练或测试集中找到的定义的三种基线方法的分数。第一个基线 Inter 返回来自另一个字典的测试集单词的定义。因此，它的分数反映了语义正确的定义，但在风格上与目标字典不同。其他基线(NE-WN 和 NE-GC)从训练集中返回最接近所定义单词的嵌人定义。在一个单词有多个定义的情况下，我们在平均总分之前对 BLEU 分数进行微平均
表 5 显示了给定不同参考词典的生成定义的 BLEU 分数。表中的 AVG 和 Merge 是聚合 BLEU 分数的两种方法。AVG 通过使用每个字典作为地面实况来平均BLEU 分数。合并通过使用两个字典的联合来计算分数。首先，我们可以看到，当根据另一个字典(Inter
|
[
[
0.56201171875,
0.0662841796875,
0.61767578125,
0.0662841796875,
0.61767578125,
0.080810546875,
0.56201171875,
0.080810546875
],
[
0.6845703125,
0.0662841796875,
0.865234375,
0.0662841796875,
0.865234375,
0.0802001953125,
0.6845703125,
0.0802001953125
],
[
0.72607421875,
0.080810546875,
0.77294921875,
0.08331298828125,
0.771484375,
0.0985107421875,
0.7255859375,
0.09661865234375
],
[
0.10296630859375,
0.06817626953125,
0.47802734375,
0.0694580078125,
0.47802734375,
0.08331298828125,
0.10296630859375,
0.08209228515625
],
[
0.08905029296875,
0.08331298828125,
0.47802734375,
0.08331298828125,
0.47802734375,
0.09661865234375,
0.08905029296875,
0.09661865234375
],
[
0.56787109375,
0.083984375,
0.60546875,
0.083984375,
0.60546875,
0.09783935546875,
0.56787109375,
0.09783935546875
],
[
0.81640625,
0.085205078125,
0.865234375,
0.085205078125,
0.865234375,
0.09661865234375,
0.81640625,
0.09661865234375
],
[
0.08660888671875,
0.0953369140625,
0.47802734375,
0.09661865234375,
0.47802734375,
0.111083984375,
0.08660888671875,
0.10986328125
],
[
0.56787109375,
0.09661865234375,
0.60302734375,
0.09661865234375,
0.60302734375,
0.11236572265625,
0.56787109375,
0.11236572265625
],
[
0.72607421875,
0.09783935546875,
0.771484375,
0.09783935546875,
0.771484375,
0.11236572265625,
0.72607421875,
0.11236572265625
],
[
0.814453125,
0.09783935546875,
0.865234375,
0.09783935546875,
0.865234375,
0.11236572265625,
0.814453125,
0.11236572265625
],
[
0.0882568359375,
0.111083984375,
0.4794921875,
0.111083984375,
0.4794921875,
0.125,
0.0882568359375,
0.125
],
[
0.56884765625,
0.11236572265625,
0.60693359375,
0.11236572265625,
0.60693359375,
0.1256103515625,
0.56884765625,
0.1256103515625
],
[
0.72607421875,
0.111083984375,
0.77294921875,
0.111083984375,
0.77294921875,
0.1256103515625,
0.72607421875,
0.1256103515625
],
[
0.81640625,
0.11297607421875,
0.86376953125,
0.11297607421875,
0.86376953125,
0.125,
0.81640625,
0.125
],
[
0.0882568359375,
0.1256103515625,
0.47802734375,
0.1256103515625,
0.47802734375,
0.1395263671875,
0.0882568359375,
0.1395263671875
],
[
0.56787109375,
0.1256103515625,
0.60888671875,
0.1256103515625,
0.60888671875,
0.14013671875,
0.56787109375,
0.14013671875
],
[
0.72607421875,
0.1256103515625,
0.77294921875,
0.1256103515625,
0.77294921875,
0.14013671875,
0.72607421875,
0.14013671875
],
[
0.81640625,
0.126953125,
0.86376953125,
0.126953125,
0.86376953125,
0.1383056640625,
0.81640625,
0.1383056640625
],
[
0.1004638671875,
0.136962890625,
0.188720703125,
0.1395263671875,
0.188720703125,
0.154052734375,
0.1004638671875,
0.1514892578125
],
[
0.7255859375,
0.1395263671875,
0.77294921875,
0.1395263671875,
0.77294921875,
0.154052734375,
0.7255859375,
0.154052734375
],
[
0.814453125,
0.1395263671875,
0.86669921875,
0.1395263671875,
0.86669921875,
0.154052734375,
0.814453125,
0.154052734375
],
[
0.10455322265625,
0.15283203125,
0.47802734375,
0.15283203125,
0.47802734375,
0.1666259765625,
0.10455322265625,
0.1666259765625
],
[
0.56640625,
0.14013671875,
0.64453125,
0.1383056640625,
0.64453125,
0.15283203125,
0.56640625,
0.1546630859375
],
[
0.56640625,
0.15283203125,
0.6806640625,
0.1514892578125,
0.6806640625,
0.1666259765625,
0.56640625,
0.1673583984375
],
[
0.7255859375,
0.15283203125,
0.77392578125,
0.15283203125,
0.77392578125,
0.1673583984375,
0.7255859375,
0.1673583984375
],
[
0.81787109375,
0.1546630859375,
0.86376953125,
0.1546630859375,
0.86376953125,
0.1666259765625,
0.81787109375,
0.1666259765625
],
[
0.0882568359375,
0.1673583984375,
0.47802734375,
0.1673583984375,
0.47802734375,
0.18115234375,
0.0882568359375,
0.18115234375
],
[
0.08660888671875,
0.18115234375,
0.476318359375,
0.18115234375,
0.476318359375,
0.1944580078125,
0.08660888671875,
0.1944580078125
],
[
0.51953125,
0.1817626953125,
0.91259765625,
0.18310546875,
0.91259765625,
0.1962890625,
0.51953125,
0.1956787109375
],
[
0.08660888671875,
0.1932373046875,
0.47802734375,
0.1944580078125,
0.47802734375,
0.208984375,
0.08660888671875,
0.2083740234375
],
[
0.51953125,
0.1976318359375,
0.7236328125,
0.1976318359375,
0.7236328125,
0.2108154296875,
0.51953125,
0.2108154296875
],
[
0.0882568359375,
0.208984375,
0.476318359375,
0.208984375,
0.476318359375,
0.222900390625,
0.0882568359375,
0.222900390625
],
[
0.08905029296875,
0.222900390625,
0.4755859375,
0.222900390625,
0.4755859375,
0.236083984375,
0.08905029296875,
0.236083984375
],
[
0.52685546875,
0.2266845703125,
0.58251953125,
0.2266845703125,
0.58251953125,
0.2412109375,
0.52685546875,
0.2412109375
],
[
0.6845703125,
0.2266845703125,
0.72216796875,
0.2266845703125,
0.72216796875,
0.240478515625,
0.6845703125,
0.240478515625
],
[
0.736328125,
0.2279052734375,
0.771484375,
0.2279052734375,
0.771484375,
0.2392578125,
0.736328125,
0.2392578125
],
[
0.79345703125,
0.2279052734375,
0.90625,
0.2259521484375,
0.90625,
0.2412109375,
0.79345703125,
0.24365234375
],
[
0.08660888671875,
0.236083984375,
0.47802734375,
0.236083984375,
0.47802734375,
0.25,
0.08660888671875,
0.25
],
[
0.53125,
0.242431640625,
0.57275390625,
0.242431640625,
0.57275390625,
0.2568359375,
0.53125,
0.2568359375
],
[
0.6689453125,
0.242431640625,
0.71337890625,
0.242431640625,
0.71337890625,
0.2568359375,
0.6689453125,
0.2568359375
],
[
0.72802734375,
0.242431640625,
0.76953125,
0.242431640625,
0.76953125,
0.2568359375,
0.72802734375,
0.2568359375
],
[
0.0882568359375,
0.25,
0.334228515625,
0.25,
0.334228515625,
0.26318359375,
0.0882568359375,
0.26318359375
],
[
0.81884765625,
0.248779296875,
0.82275390625,
0.248779296875,
0.82275390625,
0.25390625,
0.81884765625,
0.25390625
],
[
0.53515625,
0.2568359375,
0.560546875,
0.2568359375,
0.560546875,
0.270751953125,
0.53515625,
0.270751953125
],
[
0.7294921875,
0.2568359375,
0.76953125,
0.2568359375,
0.76953125,
0.271484375,
0.7294921875,
0.271484375
],
[
0.78515625,
0.2568359375,
0.82666015625,
0.2568359375,
0.82666015625,
0.271484375,
0.78515625,
0.271484375
],
[
0.85888671875,
0.2568359375,
0.89892578125,
0.2568359375,
0.89892578125,
0.271484375,
0.85888671875,
0.271484375
],
[
0.7294921875,
0.270751953125,
0.76953125,
0.270751953125,
0.76953125,
0.285400390625,
0.7294921875,
0.285400390625
],
[
0.67236328125,
0.25830078125,
0.71142578125,
0.25830078125,
0.71142578125,
0.26953125,
0.67236328125,
0.26953125
],
[
0.78515625,
0.270751953125,
0.82666015625,
0.270751953125,
0.82666015625,
0.285400390625,
0.78515625,
0.285400390625
],
[
0.85888671875,
0.270751953125,
0.900390625,
0.270751953125,
0.900390625,
0.285400390625,
0.85888671875,
0.285400390625
],
[
0.5361328125,
0.271484375,
0.59326171875,
0.271484375,
0.59326171875,
0.28271484375,
0.5361328125,
0.28271484375
],
[
0.10296630859375,
0.271484375,
0.47802734375,
0.271484375,
0.47802734375,
0.285888671875,
0.10296630859375,
0.285888671875
],
[
0.67236328125,
0.271484375,
0.71142578125,
0.271484375,
0.71142578125,
0.28271484375,
0.67236328125,
0.28271484375
],
[
0.53369140625,
0.2822265625,
0.59228515625,
0.2841796875,
0.5908203125,
0.298583984375,
0.53173828125,
0.296630859375
],
[
0.6689453125,
0.2841796875,
0.7138671875,
0.2841796875,
0.7138671875,
0.298583984375,
0.6689453125,
0.298583984375
],
[
0.7294921875,
0.2841796875,
0.76953125,
0.2841796875,
0.76953125,
0.298583984375,
0.7294921875,
0.298583984375
],
[
0.78369140625,
0.2841796875,
0.82666015625,
0.2841796875,
0.82666015625,
0.298583984375,
0.78369140625,
0.298583984375
],
[
0.85888671875,
0.2841796875,
0.900390625,
0.2841796875,
0.900390625,
0.298583984375,
0.85888671875,
0.298583984375
],
[
0.67236328125,
0.29736328125,
0.71142578125,
0.2998046875,
0.7099609375,
0.314453125,
0.6708984375,
0.3125
],
[
0.0882568359375,
0.285888671875,
0.476318359375,
0.285888671875,
0.476318359375,
0.2998046875,
0.0882568359375,
0.2998046875
],
[
0.7294921875,
0.298583984375,
0.76953125,
0.298583984375,
0.76953125,
0.313232421875,
0.7294921875,
0.313232421875
],
[
0.78369140625,
0.298583984375,
0.82666015625,
0.298583984375,
0.82666015625,
0.313232421875,
0.78369140625,
0.313232421875
],
[
0.85693359375,
0.298583984375,
0.900390625,
0.298583984375,
0.900390625,
0.313232421875,
0.85693359375,
0.313232421875
],
[
0.53369140625,
0.2998046875,
0.56884765625,
0.2998046875,
0.56884765625,
0.3125,
0.53369140625,
0.3125
],
[
0.0882568359375,
0.2998046875,
0.47802734375,
0.298583984375,
0.47802734375,
0.313232421875,
0.0882568359375,
0.314453125
],
[
0.0882568359375,
0.313232421875,
0.476318359375,
0.313232421875,
0.476318359375,
0.326904296875,
0.0882568359375,
0.326904296875
],
[
0.53369140625,
0.31005859375,
0.56884765625,
0.313232421875,
0.56640625,
0.32763671875,
0.53125,
0.324462890625
],
[
0.6708984375,
0.313232421875,
0.7138671875,
0.313232421875,
0.7138671875,
0.32763671875,
0.6708984375,
0.32763671875
],
[
0.78515625,
0.313232421875,
0.82666015625,
0.313232421875,
0.82666015625,
0.32763671875,
0.78515625,
0.32763671875
],
[
0.85693359375,
0.313232421875,
0.900390625,
0.313232421875,
0.900390625,
0.32763671875,
0.85693359375,
0.32763671875
],
[
0.53369140625,
0.32568359375,
0.57177734375,
0.32568359375,
0.57177734375,
0.341552734375,
0.53369140625,
0.341552734375
],
[
0.73046875,
0.314453125,
0.7685546875,
0.314453125,
0.7685546875,
0.32568359375,
0.73046875,
0.32568359375
],
[
0.6708984375,
0.326904296875,
0.7138671875,
0.326904296875,
0.7138671875,
0.341552734375,
0.6708984375,
0.341552734375
],
[
0.7294921875,
0.326904296875,
0.76953125,
0.326904296875,
0.76953125,
0.341552734375,
0.7294921875,
0.341552734375
],
[
0.78515625,
0.326904296875,
0.82666015625,
0.326904296875,
0.82666015625,
0.341552734375,
0.78515625,
0.341552734375
],
[
0.85693359375,
0.326904296875,
0.89892578125,
0.326904296875,
0.89892578125,
0.341552734375,
0.85693359375,
0.341552734375
],
[
0.53173828125,
0.339111328125,
0.57275390625,
0.339111328125,
0.57275390625,
0.354248046875,
0.53173828125,
0.354248046875
],
[
0.0882568359375,
0.32763671875,
0.476318359375,
0.32763671875,
0.476318359375,
0.341552734375,
0.0882568359375,
0.341552734375
],
[
0.6708984375,
0.34033203125,
0.7138671875,
0.34033203125,
0.7138671875,
0.354736328125,
0.6708984375,
0.354736328125
],
[
0.72802734375,
0.34033203125,
0.76953125,
0.34033203125,
0.76953125,
0.354736328125,
0.72802734375,
0.354736328125
],
[
0.78369140625,
0.34033203125,
0.826171875,
0.34033203125,
0.826171875,
0.354736328125,
0.78369140625,
0.354736328125
],
[
0.85888671875,
0.34033203125,
0.900390625,
0.34033203125,
0.900390625,
0.354736328125,
0.85888671875,
0.354736328125
],
[
0.08660888671875,
0.354736328125,
0.476318359375,
0.354736328125,
0.476318359375,
0.36865234375,
0.08660888671875,
0.36865234375
],
[
0.0882568359375,
0.341552734375,
0.47802734375,
0.341552734375,
0.47802734375,
0.354736328125,
0.0882568359375,
0.354736328125
],
[
0.53369140625,
0.354736328125,
0.60888671875,
0.354736328125,
0.60888671875,
0.365478515625,
0.53369140625,
0.365478515625
],
[
0.6708984375,
0.354736328125,
0.71337890625,
0.354736328125,
0.71337890625,
0.369384765625,
0.6708984375,
0.369384765625
],
[
0.78369140625,
0.354736328125,
0.82666015625,
0.354736328125,
0.82666015625,
0.369384765625,
0.78369140625,
0.369384765625
],
[
0.85693359375,
0.354736328125,
0.900390625,
0.354736328125,
0.900390625,
0.369384765625,
0.85693359375,
0.369384765625
],
[
0.0882568359375,
0.36865234375,
0.476318359375,
0.367431640625,
0.476318359375,
0.3818359375,
0.0882568359375,
0.38330078125
],
[
0.73046875,
0.35595703125,
0.76708984375,
0.35595703125,
0.76708984375,
0.367431640625,
0.73046875,
0.367431640625
],
[
0.53369140625,
0.36865234375,
0.6435546875,
0.36865234375,
0.6435546875,
0.3818359375,
0.53369140625,
0.3818359375
],
[
0.72802734375,
0.36865234375,
0.76953125,
0.36865234375,
0.76953125,
0.38330078125,
0.72802734375,
0.38330078125
],
[
0.78369140625,
0.36865234375,
0.82666015625,
0.36865234375,
0.82666015625,
0.38330078125,
0.78369140625,
0.38330078125
],
[
0.85888671875,
0.36865234375,
0.900390625,
0.36865234375,
0.900390625,
0.38330078125,
0.85888671875,
0.38330078125
],
[
0.0882568359375,
0.3818359375,
0.47802734375,
0.3818359375,
0.47802734375,
0.395751953125,
0.0882568359375,
0.395751953125
],
[
0.67236328125,
0.369384765625,
0.71337890625,
0.369384765625,
0.71337890625,
0.380615234375,
0.67236328125,
0.380615234375
],
[
0.52685546875,
0.3818359375,
0.70361328125,
0.38330078125,
0.70361328125,
0.397705078125,
0.52685546875,
0.396484375
],
[
0.0882568359375,
0.396484375,
0.47802734375,
0.396484375,
0.47802734375,
0.410400390625,
0.0882568359375,
0.410400390625
],
[
0.6708984375,
0.396484375,
0.7138671875,
0.396484375,
0.7138671875,
0.412353515625,
0.6708984375,
0.412353515625
],
[
0.72607421875,
0.396484375,
0.77294921875,
0.396484375,
0.77294921875,
0.410888671875,
0.72607421875,
0.410888671875
],
[
0.85693359375,
0.396484375,
0.904296875,
0.396484375,
0.904296875,
0.410888671875,
0.85693359375,
0.410888671875
],
[
0.77685546875,
0.3984375,
0.826171875,
0.3984375,
0.826171875,
0.410400390625,
0.77685546875,
0.410400390625
],
[
0.53173828125,
0.397705078125,
0.576171875,
0.397705078125,
0.576171875,
0.4091796875,
0.53173828125,
0.4091796875
],
[
0.78515625,
0.4091796875,
0.82666015625,
0.410888671875,
0.826171875,
0.4267578125,
0.78369140625,
0.4248046875
],
[
0.85693359375,
0.410400390625,
0.89892578125,
0.410400390625,
0.89892578125,
0.425537109375,
0.85693359375,
0.425537109375
],
[
0.53369140625,
0.410888671875,
0.64794921875,
0.410888671875,
0.64794921875,
0.4248046875,
0.53369140625,
0.4248046875
],
[
0.0882568359375,
0.410888671875,
0.4794921875,
0.4091796875,
0.4794921875,
0.42236328125,
0.0882568359375,
0.4248046875
],
[
0.67236328125,
0.412353515625,
0.71142578125,
0.412353515625,
0.71142578125,
0.423583984375,
0.67236328125,
0.423583984375
],
[
0.73046875,
0.410888671875,
0.7685546875,
0.410888671875,
0.7685546875,
0.4248046875,
0.73046875,
0.4248046875
],
[
0.08660888671875,
0.423583984375,
0.48046875,
0.42236328125,
0.48046875,
0.4375,
0.08660888671875,
0.438232421875
],
[
0.6708984375,
0.42236328125,
0.7138671875,
0.4248046875,
0.71337890625,
0.43994140625,
0.6689453125,
0.438232421875
],
[
0.72802734375,
0.4248046875,
0.76953125,
0.4248046875,
0.76953125,
0.439453125,
0.72802734375,
0.439453125
],
[
0.78173828125,
0.4248046875,
0.82666015625,
0.4248046875,
0.82666015625,
0.439453125,
0.78173828125,
0.439453125
],
[
0.85888671875,
0.4248046875,
0.900390625,
0.4248046875,
0.900390625,
0.439453125,
0.85888671875,
0.439453125
],
[
0.0882568359375,
0.439453125,
0.1380615234375,
0.439453125,
0.1380615234375,
0.45068359375,
0.0882568359375,
0.45068359375
],
[
0.53515625,
0.4267578125,
0.6357421875,
0.4267578125,
0.6357421875,
0.4375,
0.53515625,
0.4375
],
[
0.51953125,
0.453857421875,
0.9111328125,
0.455078125,
0.9111328125,
0.4697265625,
0.51953125,
0.468505859375
],
[
0.156005859375,
0.46533203125,
0.406005859375,
0.466552734375,
0.406005859375,
0.479736328125,
0.156005859375,
0.479248046875
],
[
0.521484375,
0.4697265625,
0.91015625,
0.4697265625,
0.91015625,
0.48291015625,
0.521484375,
0.48291015625
],
[
0.0882568359375,
0.486083984375,
0.476318359375,
0.486083984375,
0.476318359375,
0.5,
0.0882568359375,
0.5
],
[
0.521484375,
0.48291015625,
0.91259765625,
0.48291015625,
0.91259765625,
0.496826171875,
0.521484375,
0.496826171875
],
[
0.0882568359375,
0.5,
0.476318359375,
0.5,
0.476318359375,
0.51318359375,
0.0882568359375,
0.51318359375
],
[
0.521484375,
0.498779296875,
0.5888671875,
0.498779296875,
0.5888671875,
0.51025390625,
0.521484375,
0.51025390625
],
[
0.0882568359375,
0.51318359375,
0.476318359375,
0.51318359375,
0.476318359375,
0.5263671875,
0.0882568359375,
0.5263671875
],
[
0.08905029296875,
0.52783203125,
0.476318359375,
0.52783203125,
0.476318359375,
0.54150390625,
0.08905029296875,
0.54150390625
],
[
0.0882568359375,
0.54150390625,
0.4794921875,
0.54150390625,
0.4794921875,
0.5546875,
0.0882568359375,
0.5546875
],
[
0.521484375,
0.53857421875,
0.90869140625,
0.53857421875,
0.90869140625,
0.5517578125,
0.521484375,
0.5517578125
],
[
0.0882568359375,
0.55615234375,
0.4306640625,
0.55615234375,
0.4306640625,
0.5693359375,
0.0882568359375,
0.5693359375
],
[
0.52294921875,
0.5517578125,
0.91015625,
0.5517578125,
0.91015625,
0.56494140625,
0.52294921875,
0.56494140625
],
[
0.10455322265625,
0.5693359375,
0.47802734375,
0.5693359375,
0.47802734375,
0.58349609375,
0.10455322265625,
0.58349609375
],
[
0.521484375,
0.564453125,
0.9111328125,
0.56494140625,
0.9111328125,
0.580078125,
0.521484375,
0.5791015625
],
[
0.0882568359375,
0.58349609375,
0.47802734375,
0.58349609375,
0.47802734375,
0.5966796875,
0.0882568359375,
0.5966796875
],
[
0.52294921875,
0.58056640625,
0.6044921875,
0.58056640625,
0.6044921875,
0.59228515625,
0.52294921875,
0.59228515625
],
[
0.0882568359375,
0.5966796875,
0.47802734375,
0.5966796875,
0.47802734375,
0.60986328125,
0.0882568359375,
0.60986328125
],
[
0.521484375,
0.60791015625,
0.56787109375,
0.60791015625,
0.56787109375,
0.61865234375,
0.521484375,
0.61865234375
],
[
0.560546875,
0.60693359375,
0.73779296875,
0.60693359375,
0.73779296875,
0.62060546875,
0.560546875,
0.62060546875
],
[
0.0882568359375,
0.625,
0.47802734375,
0.625,
0.47802734375,
0.63818359375,
0.0882568359375,
0.63818359375
],
[
0.08905029296875,
0.611328125,
0.47802734375,
0.611328125,
0.47802734375,
0.625,
0.08905029296875,
0.625
],
[
0.521484375,
0.62548828125,
0.9111328125,
0.62548828125,
0.9111328125,
0.6396484375,
0.521484375,
0.6396484375
],
[
0.08660888671875,
0.63720703125,
0.476318359375,
0.63818359375,
0.476318359375,
0.65283203125,
0.08660888671875,
0.6513671875
],
[
0.521484375,
0.64013671875,
0.9111328125,
0.64013671875,
0.9111328125,
0.65380859375,
0.521484375,
0.65380859375
],
[
0.08660888671875,
0.6513671875,
0.4794921875,
0.65283203125,
0.4794921875,
0.66748046875,
0.08660888671875,
0.66650390625
],
[
0.521484375,
0.65380859375,
0.9111328125,
0.65380859375,
0.9111328125,
0.66748046875,
0.521484375,
0.66748046875
],
[
0.0882568359375,
0.66650390625,
0.47802734375,
0.66650390625,
0.47802734375,
0.6796875,
0.0882568359375,
0.6796875
],
[
0.521484375,
0.66650390625,
0.9111328125,
0.66650390625,
0.9111328125,
0.6796875,
0.521484375,
0.6796875
],
[
0.08905029296875,
0.681640625,
0.476318359375,
0.681640625,
0.476318359375,
0.6923828125,
0.08905029296875,
0.6923828125
],
[
0.51953125,
0.6796875,
0.91259765625,
0.68115234375,
0.91259765625,
0.69580078125,
0.51953125,
0.6943359375
],
[
0.0849609375,
0.6943359375,
0.468017578125,
0.6943359375,
0.468017578125,
0.70849609375,
0.0849609375,
0.70849609375
],
[
0.51953125,
0.6923828125,
0.9111328125,
0.6943359375,
0.9111328125,
0.708984375,
0.51953125,
0.70703125
],
[
0.52294921875,
0.70849609375,
0.91259765625,
0.70849609375,
0.91259765625,
0.7216796875,
0.52294921875,
0.7216796875
],
[
0.0849609375,
0.7197265625,
0.293212890625,
0.7216796875,
0.293212890625,
0.73486328125,
0.0849609375,
0.73291015625
],
[
0.52294921875,
0.7216796875,
0.9111328125,
0.7216796875,
0.9111328125,
0.73486328125,
0.52294921875,
0.73486328125
],
[
0.0906982421875,
0.7412109375,
0.47216796875,
0.7412109375,
0.47216796875,
0.751953125,
0.0906982421875,
0.751953125
],
[
0.52294921875,
0.736328125,
0.91259765625,
0.736328125,
0.91259765625,
0.75,
0.52294921875,
0.75
],
[
0.0882568359375,
0.75390625,
0.47802734375,
0.75390625,
0.47802734375,
0.767578125,
0.0882568359375,
0.767578125
],
[
0.52294921875,
0.751953125,
0.91015625,
0.751953125,
0.91015625,
0.76220703125,
0.52294921875,
0.76220703125
],
[
0.0882568359375,
0.767578125,
0.476318359375,
0.767578125,
0.476318359375,
0.78076171875,
0.0882568359375,
0.78076171875
],
[
0.52392578125,
0.7646484375,
0.9111328125,
0.7646484375,
0.9111328125,
0.77783203125,
0.52392578125,
0.77783203125
],
[
0.08660888671875,
0.78076171875,
0.476318359375,
0.77978515625,
0.476318359375,
0.79345703125,
0.08660888671875,
0.79443359375
],
[
0.521484375,
0.77587890625,
0.91015625,
0.77783203125,
0.91015625,
0.79248046875,
0.521484375,
0.79052734375
],
[
0.0882568359375,
0.79541015625,
0.476318359375,
0.79541015625,
0.476318359375,
0.8095703125,
0.0882568359375,
0.8095703125
],
[
0.521484375,
0.79150390625,
0.9111328125,
0.79052734375,
0.9111328125,
0.8046875,
0.521484375,
0.80615234375
],
[
0.0849609375,
0.806640625,
0.416748046875,
0.80810546875,
0.416748046875,
0.82275390625,
0.0849609375,
0.8212890625
],
[
0.52392578125,
0.806640625,
0.619140625,
0.806640625,
0.619140625,
0.818359375,
0.52392578125,
0.818359375
],
[
0.10296630859375,
0.82373046875,
0.47802734375,
0.82373046875,
0.47802734375,
0.8369140625,
0.10296630859375,
0.8369140625
],
[
0.53759765625,
0.8193359375,
0.9111328125,
0.8193359375,
0.9111328125,
0.83349609375,
0.53759765625,
0.83349609375
],
[
0.0882568359375,
0.8369140625,
0.4755859375,
0.8369140625,
0.4755859375,
0.85107421875,
0.0882568359375,
0.85107421875
],
[
0.52294921875,
0.83349609375,
0.9111328125,
0.83349609375,
0.9111328125,
0.84765625,
0.52294921875,
0.84765625
],
[
0.0882568359375,
0.85107421875,
0.4755859375,
0.85107421875,
0.4755859375,
0.8642578125,
0.0882568359375,
0.8642578125
],
[
0.521484375,
0.8466796875,
0.9111328125,
0.84521484375,
0.9111328125,
0.85986328125,
0.521484375,
0.861328125
],
[
0.51953125,
0.85986328125,
0.91015625,
0.861328125,
0.91015625,
0.87548828125,
0.51953125,
0.875
],
[
0.1070556640625,
0.87353515625,
0.418212890625,
0.87548828125,
0.418212890625,
0.89013671875,
0.1070556640625,
0.88818359375
],
[
0.52294921875,
0.87548828125,
0.9111328125,
0.87548828125,
0.9111328125,
0.8896484375,
0.52294921875,
0.8896484375
]
] |
[
"Model",
"#ParamsPerplexity",
"10.2m",
"Afixes英语和其他语言中的许多单词都由组成的语",
"素组成。例如，单词“capitalist”包含一个词根“capital\"",
"Seed",
"56.350",
"和一个后缀“-ist”。知道给定词根的语义以及词缀如何",
"S+1",
"10.6m",
"57.372",
"修改含义的模型可以准确地定义词根的任何形态变体。",
"S+H",
"10.4m",
"58.147",
"但是，自动将单词分解为语素并推导词缀的语义并不是",
"S+G",
"10.8m",
"50.949",
"件容易的事",
"11.1m",
"48.566",
"我们尝试通过使用字符级信息来捕获单词中的前缀",
"S+G+CH",
"S+G+CH+HE",
"11.7m",
"48.168",
"和后缀。我们采用字符级卷积网络来检测词缀（LeCun",
"etal.1990）。具体来说，w*表示为具有独热编码的字",
"Table 4:Perplexity evaluated on dictionary entries in",
"符序列。在左侧和右侧添加一个填充字符，以指示单词",
"the test set (lower is better)",
"的开头和结尾。然后，我们在字符序列上应用多个不同",
"长度的内核，并使用最大池化创建最终功能（Kimetal",
"Model",
"GC",
"WN",
"AvgMerged",
"2016）。我们假设表示为CH的卷积输人将允许模型识",
"Inter",
"27.90",
"21.15",
"别词缀如何改变单词含义的规律性",
"I",
"NE",
"21.42",
"25.49",
"34.51",
"27.42",
"29.56",
"25.06",
"32.16",
"NE-WN",
"HypernymEmbeddings正如我们在3节中讨论的那",
"22.70",
"NE-GC",
"33.22",
"17.77",
"25.49",
"35.45",
"26.69",
"样，字典定义通常遵循属十差异的结构。我们试图通",
"22.46",
"24.58",
"30.46",
"Seed",
"过向模型提供从hypernym关系数据库中提取的适当属",
"的知识来利用这种结构。特别是，我们从WebIsA数据",
"S+I",
"28.44",
"25.10",
"31.58",
"S+H",
"21.77",
"27.43",
"18.82",
"23.13",
"29.66",
"S+G",
"库（Seitneretal.2016）中获得了hypernym，该数据库",
"30.86",
"23.15",
"27.01",
"34.72",
"hypernym关系。然后，我们为模型提供一个附加的输",
"采用类似赫斯特的模式（Hearst1992）从Web中提取",
"S+G+CH",
"31.12",
"27.62",
"35.78",
"入向量，称为HE，该向量等于数据库中定义的单词的",
"24.11",
"S+G+CH+HE",
"23.81",
"27.46",
"35.28",
"顶级khypernym 的加权和。在我们的实验中，k= 5",
"31.10",
"Additional experiments",
"和权重与资源中的频率成线性比例。例如，“小鹿”的",
"27.24",
"22.78",
"31.15",
"25.01",
"Seed*",
"29.65",
"38.35",
"S+G+CH+HE*",
"前5个hypernym和频率是“color”：149，deer”：",
"33.39",
"25.91",
"135，“animal\"：132。0，“野生动物”\"：82。0，“年轻\"：",
"22.09",
"20.05",
"21.07",
"24.77",
"68.0.",
"Random Emb",
"Table 5:Equally-weighted BLEU scores for up to 4-",
"5ExperimentsandResults",
"grams, on definitions evaluated using different refer-",
"我们现在介绍评估定义模型的实验。我们使用训练集训",
"ence dictionaries (results are not comparable between",
"练多个模型体系结构，并使用测试集对3.3节中描述的",
"columns)",
"所有三个任务进行模型评估。我们使用有效集来搜索学",
"习超参数。请注意，所定义的单词在三个集合中是互",
"的，因此我们的实验评估了模型对新单词的概括程度，",
"hypernym嵌人（S+G+CH+HE）对性能的贡献不清楚",
"而不是对相同单词的其他定义或意义的概括程度",
"我们怀疑hypernym词的多个嵌人的平均值可能是定义",
"所有模型都使用来自Word2Vec项目3和2层LSTM",
"中属的较差表示。利用hypernyms的更复杂的方法是",
"网络的同一组固定的预训练单词嵌入作为RNN组件",
"未来的工作",
"(Hochreiterand Schmidhuber1997)。嵌人层和LSTM",
"5.2",
"Definition Generation",
"的内核长度为2-6，大小为[10、30、40、40、40]，步幅为",
"隐藏层各有300个单元。对于词缀检测器，字符级CNN",
"在本实验中，我们评估了模型生成的定义的质量。我们",
"1。在训练过程中，我们使用Adam最大化对数似然目",
"计算输出定义和字典定义之间的BLEU分数，以衡量",
"标，Adam是随机梯度的变体（KingmaandBa2014）。",
"生成的质量。解码算法只是从模型预测的单词概率分",
"学习率为0。001，并且在连续4个历元后停止训练，验证",
"布中一次采样一个令牌。我们使用接近贪婪算法的温度",
"性能没有明显改善。我们实验的源代码和数据集可以在",
"（公式2中的－）为每个定义的单词采样40个定义（0。",
"https://github 上找到。com/websail-nu/torch-defseq",
"05或0。1，从有效集中选择）并报告平均BLEU得分。",
"为了帮助解释BLEU分数，我们还报告了输出在训练",
"5.1Definition Modeling",
"或测试集中找到的定义的三种基线方法的分数。第一个",
"首先，我们比较了使用模型中定义的单词的不同方法",
"基线Inter返回来自另一个字典的测试集单词的定义。",
"结果如表4的第一部分所示。我们看到门控更新（S+G）",
"因此，它的分数反映了语义正确的定义，但在风格上与",
"提高了种子的性能，而其他架构（S+I和S+H）则没",
"目标字典不同。其他基线（NE-WN和NE-GC）从训练",
"有。结果与我们的假设一致，即被定义的单词与定义中",
"集中返回最接近所定义单词的嵌入定义。在一个单词有",
"的某些单词比与其他单词更相关，并且门更新可以识别",
"多个定义的情况下，我们在平均总分之前对BLEU分",
"这一点。我们在6节中进一步探讨了门的行为",
"数进行微平均",
"接下来，我们评估语言特征的贡献。我们看到词缀",
"表5显示了给定不同参考词典的生成定义的BLEU",
"(S+G+CH）进一步改进了模型，表明字符级信息可以",
"分数。表中的AVG和Merge是聚合BLEU分数的两",
"补充从上下文中学到的单词嵌人。也许令人惊牙的是",
"种方法。AVG通过使用每个字典作为地面实况来平均",
"BLEU分数。合并通过使用两个字典的联合来计算分",
"3https://code.google.com/archive/p/word2vec/",
"数。首先，我们可以看到，当根据另一个字典（Inter"
] |
{
"arxiv_id": "1612.00394",
"language": "zh",
"page_idx": 4
}
|
||
2001.11003-zh-1
|
2001.11003-zh-1.jpg
|
势，但通过有效地将每个节点视为连接到图中的所有其他节点而忽略了图拓扑。相反，基于局部聚合的模型根据输入图中定义的相邻节点来学习每个节点的表示。这种方法有效地利用了图的拓扑结构，图的结构对节点表示有很强的影响 (Xu et al., 2018) 但是，由于需要更多的图编码层，因此远程节点之间的编码关系可能具有挑战性，这也会传播噪声(Li et al., 2018) 。例如，图 1 a 表示一个 KG, 其相应的文本如图 1 b 所示。请注意，实体在图中的连接方式与它们的自然语言描述在文本中的关联方式之间存在不匹配。文本中语法相关的某些实体在图中未连接。例如，在句子“对于链接预测任务，首先我们使用 DistMult 方法学习节点嵌人。”,而实体提及依赖于相同的动词，在图中，节点嵌入节点与链接预测和 DistMult 节点没有显式连接，它们位于不同的连接组件中。此示例说明了在输入图中编码远程信息的重要性。如图 1 c 所示，全局编码器能够学习节点嵌入的节点表示，该节点嵌入从非连接实体(如 DistMult)捕获信息。通过对所有节点之间的远程连接进行建模，我们可以捕获这些丢失的链接，因为已知 KG 是高度不完整的 (Dong et al., 2014; Schlichtkrull et al., 2018)
相比之下，本地策略通过更丰富的邻域信息来细化节点表示，因为共享同一邻域的节点表现出很强的同质性：两个相似的实体比随机连接的可能性要大得多。因此，本地上下文用KG三元组的本地信息丰富了节点表示。例如， 在图 1 a 中，GAT 通过 GNN 到达节点嵌入。这种传递关系可以由本地编码器捕获，如图$\overline{\mathbb{E}}$ 1 d 所示。捕获这种形式的关系也可以支持句子级别的文本生成
在本文中，我们研究了结合了全局和局部节点聚合的新颖的图形到文本体系结构，从而从两种策略中获得了好处。特别是，我们提出了一个基于图形注意力网络(GAT) (Veličković et al., 2018) 的统一图形到文本框架。作为该框架的一部分，我们根据经验比较了两种主要体系结构：在执行本地节点聚合之前执行全局节点聚合的级联体系结构，以及同时执行全局和本地聚合的并行体系结构。虽然级联架构允许本地编码器利用全局编码功能，但并行架构允许更多独立功能相互补充。为了进一步考虑细粒度集成，我们还考虑了全局和局部编码器的逐层集成
广泛的实验表明，我们的方法在从 KG 生成文本的两个基准上始终优于最近的模型。据我们所知，我们是第一个考虑在图形到文本生成中集成全局和局部上下文聚合的人，也是第一个提出用于组合全局和局部节点上下文的统一
GAT 结构的人
## 2 $\textbf{Related Work}$
图形到文本生成的早期努力采用统计方法(Flanigan et al., 2016; Pourdamghani et al., 2016; Song et al., 2017)。最近，通过利用基于 LSTM, GNN 和 Transformers 的编码器机制，几种神经graph-to-text 模型取得了成功。
AMR-to-Text Generation. 已经提出了各种神经模型来从抽象含义表示(AMR)图生成句子。Konstas et al.(2017) 通过将输入图线性化为节点和边的序列，为该任务提供了第一种神经方法。Song et al.(2018)提出了图递归网络(GRN)来直接对 AMR 节点进行编码， 而 Beck et al.(2018) 则基于门控 GNN 开发了一个模型。然而，这两种方法都只采用本地节点聚合策略。Damonte and Cohen (2019) 结合图卷积网络(GCN)和 LSTM 以学习互补节点上下文。但是，与 Transformers 和 GNN 不同，LSTM 生成受节点顺序影响的节点表示。Ribeiro et al. (2019) 开发了一个基于不同 GNN 的模型，该模型学习了同时对 AMR 图的自顶向下和自底向上视图进行编码的节点表示， 而 Guo et al. (2019) 则利用了 GNN 中的密集连接。最近，Wang et al. (2020) 提出了一种基于Transformers 的本地图编码器，该编码器对传人和传出邻居使用了单独的注意力。最近的方法 (Zhu et al., 2019; Cai and Lam, 2020) 也采用了 Transformers, 但是学习了全球化的节点表示，对图形路径进行建模以捕获结构关系
$\mathbf{KG-to-Text~Generation.~}$在这项工作中，我们专注于从 KG 生成文本。与有根和连接图的AMR 相比，KG 没有定义的拓扑，不同数据集之间的拓扑可能差异很大，这使得生成过程要求更高。KG 是稀疏结构，可能包含大量关系。此外，我们通常对从 KG 生成多句子文本感兴趣，这涉及解决文档规划问题(Konstas and Lapata, 2013)
最近用于 KG 到文本生成的神经方法简单地线性化 KG 三元组，从而丢失图结构信息。例如，Colin and Gardent (2018) , Moryossef et al.(2019) 和 Adapt (Gardent et al.,2017) 使用 LSTM/GRU 对 WebNLG 图进行编码。Castro Ferreira et al.(2019) 系统地比较了用于从WebNLG 图生成文本的管道和端到端模型。Trisedya et al.(2018)开发了一个基于 LSTM 的图形编码器，用于捕获三元组内部和之间的关系。以前的工作还研究了如何使用 GNN 或 Transformers 显式编码图结构。Marcheggiani and Perez Beltrachini (2018)提出了一种基于
|
[
[
0.12005615234375,
0.07781982421875,
0.486083984375,
0.07781982421875,
0.486083984375,
0.0885009765625,
0.12005615234375,
0.0885009765625
],
[
0.5146484375,
0.07781982421875,
0.62890625,
0.07781982421875,
0.62890625,
0.09027099609375,
0.5146484375,
0.09027099609375
],
[
0.118408203125,
0.0926513671875,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.118408203125,
0.105712890625
],
[
0.5146484375,
0.1092529296875,
0.5283203125,
0.1092529296875,
0.5283203125,
0.1175537109375,
0.5146484375,
0.1175537109375
],
[
0.5263671875,
0.1092529296875,
0.54833984375,
0.1092529296875,
0.54833984375,
0.11639404296875,
0.5263671875,
0.11639404296875
],
[
0.54736328125,
0.105712890625,
0.6708984375,
0.105712890625,
0.6708984375,
0.119384765625,
0.54736328125,
0.119384765625
],
[
0.118408203125,
0.1258544921875,
0.484375,
0.1258544921875,
0.484375,
0.1395263671875,
0.118408203125,
0.1395263671875
],
[
0.12005615234375,
0.10986328125,
0.48779296875,
0.10986328125,
0.48779296875,
0.12353515625,
0.12005615234375,
0.12353515625
],
[
0.515625,
0.1341552734375,
0.87890625,
0.1341552734375,
0.87890625,
0.144287109375,
0.515625,
0.144287109375
],
[
0.1234130859375,
0.1436767578125,
0.486083984375,
0.1436767578125,
0.486083984375,
0.15380859375,
0.1234130859375,
0.15380859375
],
[
0.5146484375,
0.15087890625,
0.8818359375,
0.15087890625,
0.8818359375,
0.1644287109375,
0.5146484375,
0.1644287109375
],
[
0.12005615234375,
0.157958984375,
0.48779296875,
0.157958984375,
0.48779296875,
0.171630859375,
0.12005615234375,
0.171630859375
],
[
0.51123046875,
0.1644287109375,
0.880859375,
0.166259765625,
0.880859375,
0.179931640625,
0.51123046875,
0.1776123046875
],
[
0.12005615234375,
0.17333984375,
0.489501953125,
0.17333984375,
0.489501953125,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.5146484375,
0.1817626953125,
0.8818359375,
0.1817626953125,
0.8818359375,
0.1953125,
0.5146484375,
0.1953125
],
[
0.118408203125,
0.188232421875,
0.486083984375,
0.188232421875,
0.486083984375,
0.201904296875,
0.118408203125,
0.201904296875
],
[
0.51318359375,
0.198974609375,
0.75146484375,
0.19775390625,
0.75146484375,
0.2120361328125,
0.51318359375,
0.213134765625
],
[
0.1376953125,
0.2060546875,
0.489501953125,
0.2060546875,
0.489501953125,
0.2197265625,
0.1376953125,
0.2197265625
],
[
0.12176513671875,
0.221435546875,
0.48779296875,
0.221435546875,
0.48779296875,
0.235107421875,
0.12176513671875,
0.235107421875
],
[
0.5224609375,
0.2244873046875,
0.8818359375,
0.2244873046875,
0.8818359375,
0.2381591796875,
0.5224609375,
0.2381591796875
],
[
0.12005615234375,
0.2381591796875,
0.489501953125,
0.2381591796875,
0.489501953125,
0.251708984375,
0.12005615234375,
0.251708984375
],
[
0.515625,
0.240478515625,
0.88330078125,
0.240478515625,
0.88330078125,
0.254150390625,
0.515625,
0.254150390625
],
[
0.12005615234375,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.267333984375,
0.12005615234375,
0.267333984375
],
[
0.515625,
0.257080078125,
0.88330078125,
0.257080078125,
0.88330078125,
0.270751953125,
0.515625,
0.270751953125
],
[
0.12176513671875,
0.270751953125,
0.48779296875,
0.270751953125,
0.48779296875,
0.283935546875,
0.12176513671875,
0.283935546875
],
[
0.5146484375,
0.2724609375,
0.86572265625,
0.2724609375,
0.86572265625,
0.2861328125,
0.5146484375,
0.2861328125
],
[
0.85986328125,
0.275634765625,
0.8759765625,
0.27490234375,
0.8759765625,
0.279052734375,
0.85986328125,
0.27978515625
],
[
0.118408203125,
0.28515625,
0.486083984375,
0.28515625,
0.486083984375,
0.298583984375,
0.118408203125,
0.298583984375
],
[
0.5146484375,
0.289306640625,
0.8818359375,
0.289306640625,
0.8818359375,
0.302734375,
0.5146484375,
0.302734375
],
[
0.1234130859375,
0.302734375,
0.48779296875,
0.302734375,
0.48779296875,
0.31640625,
0.1234130859375,
0.31640625
],
[
0.515625,
0.3046875,
0.8818359375,
0.3046875,
0.8818359375,
0.318359375,
0.515625,
0.318359375
],
[
0.12176513671875,
0.318359375,
0.489501953125,
0.318359375,
0.489501953125,
0.33203125,
0.12176513671875,
0.33203125
],
[
0.5146484375,
0.32177734375,
0.8818359375,
0.32177734375,
0.8818359375,
0.3349609375,
0.5146484375,
0.3349609375
],
[
0.118408203125,
0.334228515625,
0.486083984375,
0.334228515625,
0.486083984375,
0.347412109375,
0.118408203125,
0.347412109375
],
[
0.52490234375,
0.33740234375,
0.8818359375,
0.33740234375,
0.8818359375,
0.350830078125,
0.52490234375,
0.350830078125
],
[
0.12176513671875,
0.350830078125,
0.48779296875,
0.350830078125,
0.48779296875,
0.364013671875,
0.12176513671875,
0.364013671875
],
[
0.51708984375,
0.355224609375,
0.880859375,
0.355224609375,
0.880859375,
0.365234375,
0.51708984375,
0.365234375
],
[
0.1234130859375,
0.3681640625,
0.486083984375,
0.3681640625,
0.486083984375,
0.37890625,
0.1234130859375,
0.37890625
],
[
0.515625,
0.369384765625,
0.8818359375,
0.369384765625,
0.8818359375,
0.383056640625,
0.515625,
0.383056640625
],
[
0.12176513671875,
0.383056640625,
0.489501953125,
0.383056640625,
0.489501953125,
0.396728515625,
0.12176513671875,
0.396728515625
],
[
0.515625,
0.385498046875,
0.88330078125,
0.385498046875,
0.88330078125,
0.399658203125,
0.515625,
0.399658203125
],
[
0.12005615234375,
0.399658203125,
0.48779296875,
0.399658203125,
0.48779296875,
0.413330078125,
0.12005615234375,
0.413330078125
],
[
0.515625,
0.402099609375,
0.8818359375,
0.402099609375,
0.8818359375,
0.4150390625,
0.515625,
0.4150390625
],
[
0.12005615234375,
0.4150390625,
0.489501953125,
0.4150390625,
0.489501953125,
0.4287109375,
0.12005615234375,
0.4287109375
],
[
0.51318359375,
0.416259765625,
0.8818359375,
0.41748046875,
0.8818359375,
0.431640625,
0.51318359375,
0.43115234375
],
[
0.118408203125,
0.429931640625,
0.48779296875,
0.43115234375,
0.48779296875,
0.4453125,
0.118408203125,
0.44482421875
],
[
0.515625,
0.43408203125,
0.880859375,
0.43408203125,
0.880859375,
0.44775390625,
0.515625,
0.44775390625
],
[
0.12005615234375,
0.44775390625,
0.39794921875,
0.44775390625,
0.39794921875,
0.46142578125,
0.12005615234375,
0.46142578125
],
[
0.515625,
0.449462890625,
0.880859375,
0.449462890625,
0.880859375,
0.463134765625,
0.515625,
0.463134765625
],
[
0.1376953125,
0.468505859375,
0.486083984375,
0.468505859375,
0.486083984375,
0.482177734375,
0.1376953125,
0.482177734375
],
[
0.51708984375,
0.46728515625,
0.880859375,
0.46728515625,
0.880859375,
0.47802734375,
0.51708984375,
0.47802734375
],
[
0.12005615234375,
0.486328125,
0.486083984375,
0.486328125,
0.486083984375,
0.496337890625,
0.12005615234375,
0.496337890625
],
[
0.5146484375,
0.482666015625,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.5146484375,
0.496337890625
],
[
0.12176513671875,
0.501953125,
0.486083984375,
0.501953125,
0.486083984375,
0.5126953125,
0.12176513671875,
0.5126953125
],
[
0.51318359375,
0.498779296875,
0.8818359375,
0.498779296875,
0.8818359375,
0.5126953125,
0.51318359375,
0.5126953125
],
[
0.12176513671875,
0.5166015625,
0.486083984375,
0.5166015625,
0.486083984375,
0.52978515625,
0.12176513671875,
0.52978515625
],
[
0.515625,
0.5166015625,
0.880859375,
0.5166015625,
0.880859375,
0.52685546875,
0.515625,
0.52685546875
],
[
0.118408203125,
0.53076171875,
0.49365234375,
0.5322265625,
0.49365234375,
0.5458984375,
0.118408203125,
0.54443359375
],
[
0.51318359375,
0.529296875,
0.8818359375,
0.52978515625,
0.8818359375,
0.54443359375,
0.51318359375,
0.54345703125
],
[
0.118408203125,
0.54736328125,
0.489501953125,
0.548828125,
0.489501953125,
0.5625,
0.118408203125,
0.56103515625
],
[
0.5146484375,
0.54443359375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55859375,
0.5146484375,
0.55859375
],
[
0.118408203125,
0.5634765625,
0.48779296875,
0.5634765625,
0.48779296875,
0.57666015625,
0.118408203125,
0.57666015625
],
[
0.51318359375,
0.56103515625,
0.84619140625,
0.5625,
0.84619140625,
0.57666015625,
0.51318359375,
0.576171875
],
[
0.118408203125,
0.5791015625,
0.48779296875,
0.580078125,
0.48779296875,
0.59423828125,
0.118408203125,
0.59326171875
],
[
0.51904296875,
0.58740234375,
0.88330078125,
0.5888671875,
0.88330078125,
0.60400390625,
0.51904296875,
0.6015625
],
[
0.12176513671875,
0.59716796875,
0.268798828125,
0.59716796875,
0.268798828125,
0.61083984375,
0.12176513671875,
0.61083984375
],
[
0.5146484375,
0.60400390625,
0.8818359375,
0.60498046875,
0.8818359375,
0.619140625,
0.5146484375,
0.6181640625
],
[
0.1376953125,
0.61767578125,
0.48779296875,
0.61767578125,
0.48779296875,
0.63134765625,
0.1376953125,
0.63134765625
],
[
0.5146484375,
0.62158203125,
0.88330078125,
0.62158203125,
0.88330078125,
0.63525390625,
0.5146484375,
0.63525390625
],
[
0.12005615234375,
0.6328125,
0.486083984375,
0.6328125,
0.486083984375,
0.646484375,
0.12005615234375,
0.646484375
],
[
0.5146484375,
0.63720703125,
0.8818359375,
0.63720703125,
0.8818359375,
0.65087890625,
0.5146484375,
0.65087890625
],
[
0.12176513671875,
0.65087890625,
0.483642578125,
0.65087890625,
0.483642578125,
0.6611328125,
0.12176513671875,
0.6611328125
],
[
0.51318359375,
0.65283203125,
0.880859375,
0.65283203125,
0.880859375,
0.66650390625,
0.51318359375,
0.66650390625
],
[
0.1318359375,
0.66650390625,
0.489501953125,
0.66650390625,
0.489501953125,
0.6796875,
0.1318359375,
0.6796875
],
[
0.5146484375,
0.66845703125,
0.8818359375,
0.66845703125,
0.8818359375,
0.681640625,
0.5146484375,
0.681640625
],
[
0.118408203125,
0.68115234375,
0.48779296875,
0.68115234375,
0.48779296875,
0.69482421875,
0.118408203125,
0.69482421875
],
[
0.5146484375,
0.68603515625,
0.88330078125,
0.68603515625,
0.88330078125,
0.69970703125,
0.5146484375,
0.69970703125
],
[
0.12176513671875,
0.69970703125,
0.48779296875,
0.69970703125,
0.48779296875,
0.71044921875,
0.12176513671875,
0.71044921875
],
[
0.51123046875,
0.70166015625,
0.6171875,
0.69970703125,
0.6171875,
0.71435546875,
0.51123046875,
0.71630859375
],
[
0.12176513671875,
0.71435546875,
0.48779296875,
0.71435546875,
0.48779296875,
0.72802734375,
0.12176513671875,
0.72802734375
],
[
0.5322265625,
0.71728515625,
0.8818359375,
0.71875,
0.8818359375,
0.73291015625,
0.5322265625,
0.732421875
],
[
0.12176513671875,
0.72998046875,
0.48779296875,
0.72998046875,
0.48779296875,
0.74365234375,
0.12176513671875,
0.74365234375
],
[
0.5146484375,
0.7353515625,
0.8818359375,
0.7353515625,
0.8818359375,
0.7490234375,
0.5146484375,
0.7490234375
],
[
0.12005615234375,
0.74462890625,
0.486083984375,
0.74462890625,
0.486083984375,
0.75830078125,
0.12005615234375,
0.75830078125
],
[
0.5146484375,
0.75048828125,
0.88330078125,
0.75048828125,
0.88330078125,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.12176513671875,
0.7626953125,
0.489501953125,
0.7626953125,
0.489501953125,
0.775390625,
0.12176513671875,
0.775390625
],
[
0.5146484375,
0.76708984375,
0.87890625,
0.76708984375,
0.87890625,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.12005615234375,
0.77783203125,
0.48779296875,
0.77783203125,
0.48779296875,
0.79150390625,
0.12005615234375,
0.79150390625
],
[
0.5146484375,
0.783203125,
0.880859375,
0.783203125,
0.880859375,
0.79638671875,
0.5146484375,
0.79638671875
],
[
0.12005615234375,
0.79443359375,
0.48779296875,
0.79443359375,
0.48779296875,
0.80810546875,
0.12005615234375,
0.80810546875
],
[
0.515625,
0.80029296875,
0.880859375,
0.80029296875,
0.880859375,
0.81103515625,
0.515625,
0.81103515625
],
[
0.12005615234375,
0.8125,
0.193115234375,
0.8125,
0.193115234375,
0.82373046875,
0.12005615234375,
0.82373046875
],
[
0.51318359375,
0.8134765625,
0.88330078125,
0.81396484375,
0.88330078125,
0.8291015625,
0.51318359375,
0.82763671875
],
[
0.1368408203125,
0.83056640625,
0.489501953125,
0.830078125,
0.489501953125,
0.84423828125,
0.1368408203125,
0.845703125
],
[
0.51123046875,
0.83203125,
0.880859375,
0.83203125,
0.880859375,
0.845703125,
0.51123046875,
0.845703125
],
[
0.12005615234375,
0.84716796875,
0.48779296875,
0.84716796875,
0.48779296875,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.5146484375,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.12176513671875,
0.865234375,
0.48779296875,
0.865234375,
0.48779296875,
0.8759765625,
0.12176513671875,
0.8759765625
],
[
0.515625,
0.865234375,
0.880859375,
0.865234375,
0.880859375,
0.8759765625,
0.515625,
0.8759765625
],
[
0.12005615234375,
0.87744140625,
0.515625,
0.87890625,
0.515625,
0.8935546875,
0.12005615234375,
0.892578125
],
[
0.51318359375,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.118408203125,
0.89453125,
0.515625,
0.896484375,
0.515625,
0.91015625,
0.118408203125,
0.90771484375
],
[
0.51318359375,
0.89794921875,
0.87890625,
0.89794921875,
0.87890625,
0.90771484375,
0.51318359375,
0.90771484375
]
] |
[
"势，但通过有效地将每个节点视为连接到图中",
"GAT结构的人",
"的所有其他节点而忽略了图拓扑。相反，基于",
"2",
"",
"Related Work",
"来学习每个节点的表示。这种方法有效地利用",
"局部聚合的模型根据输入图中定义的相邻节点",
"图形到文本生成的早期努力采用统计方法",
"了图的拓扑结构：图的结构对节点表示有很强",
"(Flanigan et al., 2016; Pourdamghani et al., 2016;",
"的影响（(Xuetal.,2018）但是，由于需要更多的",
"Songetal.,2017）。最近，通过利用基于LSTM",
"图编码层，因此远程节点之间的编码关系可能",
"GNN和Transformers的编码器机制，几种神经",
"具有挑战性，这也会传播噪声（Lietal.,2018)",
"graph-to-text模型取得了成功。",
"例如，图1a表示一个KG，其相应的文本",
"如图1b所示。请注意，实体在图中的连接方",
"AMR-to-TextGeneration.已经提出了各种",
"式与它们的自然语言描述在文本中的关联方式",
"神经模型来从抽象含义表示（AMR）图生成",
"之间存在不匹配。文本中语法相关的某些实体",
"句子。Konstasetal.(2017）通过将输入图线性",
"在图中未连接。例如，在句子“对于链接预测",
"化为节点和边的序列，为该任务提供了第",
"",
"任务，首先我们使用DistMult方法学习节点嵌",
"种神经方法。Songetal.(2018)提出了图递归",
"入。”，而实体提及依赖于相同的动词，在图",
"网络（GRN）来直接对AMR节点进行编码，",
"中，节点嵌入节点与链接预测和DistMult节点",
"而Becketal.(2018）则基于门控GNN开发了",
"没有显式连接，它们位于不同的连接组件中。",
"一个模型。然而，这两种方法都只采用本地节",
"此示例说明了在输人图中编码远程信息的重",
"点聚合策略。DamonteandCohen（2019）结合",
"要性。如图1c所示，全局编码器能够学习节",
"图卷积网络（GCN）和LSTM以学习互补节",
"点嵌入的节点表示，该节点嵌入从非连接实体",
"点上下文。但是，与Transformers和GNN不",
"（如DistMult）捕获信息。通过对所有节点之间",
"同，LSTM生成受节点顺序影响的节点表示。",
"的远程连接进行建模，我们可以捕获这些丢失",
"Ribeiro etal.(2019）开发了一个基于不同GNN",
"的链接，因为已知KG是高度不完整的（Dong",
"的模型，该模型学习了同时对AMR图的自",
"et al., 2014; Schlichtkrull et al., 2018)",
"顶向下和自底向上视图进行编码的节点表示",
"相比之下，本地策略通过更丰富的邻域信息",
"而Guoetal.（2019）则利用了GNN中的密集连",
"来细化节点表示，因为共享同一邻域的节点",
"接。最近，Wangetal.(2020）提出了一种基于",
"表现出很强的同质性：两个相似的实体比随机",
"Transformers的本地图编码器，该编码器对传",
"连接的可能性要大得多。因此，本地上下文用",
"入和传出邻居使用了单独的注意力。最近的方",
"KG三元组的本地信息丰富了节点表示。例如，",
"法(Zhuetal.,2019;Caiand Lam,2020)也采用",
"在图1a中，GAT通过GNN到达节点嵌人。",
"了Transformers，但是学习了全球化的节点表",
"这种传递关系可以由本地编码器捕获，如图E",
"示，对图形路径进行建模以捕获结构关系",
"1d所示。捕获这种形式的关系也可以支持句",
"KG-to-TextGeneration。在这项工作中，我",
"子级别的文本生成",
"们专注于从KG生成文本。与有根和连接图的",
"在本文中，我们研究了结合了全局和局部节",
"AMR相比，KG没有定义的拓扑，不同数据集",
"点聚合的新颖的图形到文本体系结构，从而从",
"之间的拓扑可能差异很大，这使得生成过程要",
"两种策略中获得了好处。特别是，我们提出了",
"求更高。KG是稀疏结构，可能包含大量关系。",
"个基于图形注意力网络（GAT）（Velickovic",
"此外，我们通常对从KG生成多句子文本感",
"etal.2018)的统一图形到文本框架。作为该框",
"兴趣，这涉及解决文档规划问题（Konstasand",
"架的一部分，我们根据经验比较了两种主要体",
"Lapata, 2013)",
"系结构：在执行本地节点聚合之前执行全局节",
"最近用于KG到文本生成的神经方法简",
"点聚合的级联体系结构，以及同时执行全局和",
"单地线性化KG三元组，从而丢失图结构信",
"本地聚合的并行体系结构。虽然级联架构允许",
"息。例如，Colin andGardent(2018），Moryossef",
"本地编码器利用全局编码功能，但并行架构允",
"et al.(2019)和Adapt (Gardent et al..2017)使",
"许更多独立功能相互补充。为了进一步考虑细",
"用LSTM/GRU对WebNLG图进行编码。Cas-",
"粒度集成，我们还考虑了全局和局部编码器的",
"troFerreiraetal.（2019）系统地比较了用于从",
"逐层集成",
"WebNLG图生成文本的管道和端到端模型。",
"广泛的实验表明，我们的方法在从KG生成",
"Trisedya etal.（2018）开发了一个基于LSTM",
"文本的两个基准上始终优于最近的模型。据我",
"的图形编码器，用于捕获三元组内部和之间",
"们所知，我们是第一个考虑在图形到文本生成",
"的关系。以前的工作还研究了如何使用GNN",
"中集成全局和局部上下文聚合的人，也是第一",
"或Transformers显式编码图结构。Marcheggiani",
"个提出用于组合全局和局部节点上下文的统一",
"andPerezBeltrachini（2018）提出厂种基于"
] |
{
"arxiv_id": "2001.11003",
"language": "zh",
"page_idx": 1
}
|
||
2001.11003-zh-2
|
2001.11003-zh-2.jpg
|
和局部消息传递机制，捕获互补的图上下文
Integrating Global Information. 最近的工作试图集成全局上下文，以便在图形到文本生成中学习更好的节点表示。为此，现有方法采用人工全局节点与其他节点进行消息交换。这种策略可以被视为扩展了图形结构，但使用了类似的消息传递机制。特别是，Koncel-Kedziorski et al.(2019) 将全局节点添加到图中，并使用其表示来初始化解码器。最近，Guo et al.(2019) 和 Cai and Lam (2020) 还采用了一个人工全局节点，该节点具有与所有其他节点的直接边， 以允许全局消息交换以进行 AMR 到文本的生成。同样，Zhang et al.(2018)使用 GRN 模型的全局节点进行句子表示。与上述方法不同， 我们通过调查 model 替代方案而不是 graph structure更改，考虑在node 级别而不是 graph 级别集成全局和局部上下文。此外，我们将GAT 和 Transformer 架构集成到一个统一的全局-局部模型中。
### 3 Graph-to-Text Model
本节首先描述(i)用于从输入(第回3.1节)创建关系图的图转换，以及(ii)我们基于图注意切相关，它为建模全局节点上下文提供了一个方便的原型。然后，(iii)我们提出了组合全局和局部图编码器的策略(第3.5节)。最后，(iv) 我们描述了解码和训练过程(第3.6节)
## 3.1 Graph Preparation
我们将 KG 表示为具有实体节点$e\in\mathcal{V}_e$和标记边$\left(e_h,r,e_t\right)\in\mathcal{E}_e\text{的多关系图 }^2\mathcal{G}_e=\left(\mathcal{V}_e,\mathcal{E}_e,\mathcal{R}\right)$ ,其中$r\in\mathcal{R}$表示从实体$e_h$到$e_t$之间存在的关系。3
与其他当前方法 (Koncel-Kedziorski et al., 2019; Moryossef et al., 2019) 不同，我们将实
$^{2}$In this paper, multi-relational graphs refer to directed
graphs with labelled edges.
$^3\mathcal{R}$ contains relations both in canonical direction (e.g. used-for) and in inverse direction (e.g. used-for-inv), so that the models consider the differences in the incoming and outgoing relations.
GCN 的编码器，该编码器明确考虑了本地节 体表示为一组节点。例如，图 1中的 KG 节点点上下文，并且与 LSTM 相比显示出优越的性 “节点嵌人”将由两个节点表示，一个用于令牌能。最近，Koncel-Kedziorski et al.(2019)提出“节点”,另一个用于令牌“嵌人”。形式上，我了一种基于 Transformer 的方法，该方法通过 们将每个$G_e$ 转换为一个新的图$\mathcal{G}=\left(\mathcal{V},\mathcal{E},\mathcal{R}\right)$, 遵循自我关注策略参与节点邻域来计算节点表 其中实体$e\in\mathcal{V}_e$的每个令牌都成为节点$v\in\mathcal{V}$ 示。相比之下，我们的模型专注于不同的全局 。我们将每个边$(e_h,r,e_t)\in\mathcal{E}_e$转换为一组边
(具有相同的关系$r)$,并将$e_h$的每个令牌连接到$e_t$的每个令牌。也就是说，当且仅当存在边缘$(e_h,r,e_t)\in\mathcal{E}_e$时，边缘$(u,r,v)$将属于$\mathcal{E}$ ,使得$u\in e_h$ 和$v\in e_t$,其中$e_h$ 和$e_t$被视为令牌集。我们用一个嵌人的$h_v^0\in\mathbb{R}^{d_v}$表示每个节点$v\in\mathcal{V}$,从其相应的令牌生成。新的图形$\mathcal{G}$增加了模型的表示能力，因为它允许在令牌级别而不是实体级别学习节点嵌入。这对于文本生成特别重要，因为它允许模型更加灵活，捕获实体令牌之间更丰富的关系。这也允许模型学习源和目标令牌之间的关系和注意力函数。但是，它具有消除多词实体的自然顺序顺序的副作用。为了保留这些信息，我们使用位置嵌人(Vaswani et al.,2017) , i。e。, $h_v^{0}$成为$v$的相应令牌嵌入和位置嵌人的总和
## 3.2 Graph Neural Networks(GNN)
多层 GNN 通过信息传播方案，基于其上下文节点邻居和边缘特征，通过迭代学习节点$v$的表示向量$h_v$来工作。更正式地说，第$l$层聚合了$v$ 的上下文节点的表示：
$$h_{\mathcal{N}(v)}^{(l)}=\text{AGGR}^{(l)}\big(\big\{\big(h_u^{(l-1)},r_{uv}\big):u\in\mathcal{N}(v)\big\}\big)\:,$$
网络 (GAT) (Veličković et al., 2018) 的框架的 其中 AGGR$^{(l)}$(.)是一个聚合函数，由第$l$层上图编码器，用于处理全局(第 3.3 节) 和局部 的所有节点共享。$r_uv$表示$u$和$v$之间的关系。(第 3.4节)节点上下文。我们采用 GAT 是因为 $\mathcal{N} ( v)$ 是 $v$ 的一组上下文节点。在大多数 GNN 它与 Transformer 架构 (Vaswani et al.2017)密 中，上下文节点是与$v$相邻的节点。$h_\mathcal{N}(v)^{(l)}$是
$\mathcal{N}(v)$在$l$层的聚合上下文表示。$h_\mathcal{N}(v)^{(l)}$用于更
新$v$的表示形式：
$$h_v^{(l)}=\text{COMBINE}^{(l)}\Big(h_v^{(l-1)}\:,h_{\mathcal{N}(v)}^{(l)}\Big)\:.$$
在$L$迭代之后，节点的表示对其$L$跳邻域内的结构信息进行编码。AGGR$^{(l)}(.)$和$\mathrm{COMBINE}^{(l)}(.)$的选择因特定的 GNN 模型而异。AGGR$^(l)(.)$的一个示例是$\mathcal{N}(v)$的表示形式的总和。COMBINE$^{(l)}(.)$的一个示例是特征转换后的串联
# 3.3 Global Graph Encoder
全局图编码器基于图的所有节点聚合用于更新每个节点的全局上下文(请参见图 1 c)。我们使用注意力机制作为消息传递方案，将Transformer 的自注意力网络结构扩展为 GAT
|
[
[
0.12005615234375,
0.07598876953125,
0.489501953125,
0.07598876953125,
0.489501953125,
0.08905029296875,
0.12005615234375,
0.08905029296875
],
[
0.490234375,
0.08013916015625,
0.51318359375,
0.08013916015625,
0.51318359375,
0.0872802734375,
0.490234375,
0.0872802734375
],
[
0.509765625,
0.07598876953125,
0.8818359375,
0.07659912109375,
0.8818359375,
0.0914306640625,
0.509765625,
0.09027099609375
],
[
0.12176513671875,
0.09320068359375,
0.490234375,
0.09320068359375,
0.490234375,
0.10687255859375,
0.12176513671875,
0.10687255859375
],
[
0.5029296875,
0.097412109375,
0.51123046875,
0.097412109375,
0.51123046875,
0.1009521484375,
0.5029296875,
0.1009521484375
],
[
0.50732421875,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.50732421875,
0.10687255859375
],
[
0.12005615234375,
0.1092529296875,
0.48779296875,
0.1092529296875,
0.48779296875,
0.122314453125,
0.12005615234375,
0.122314453125
],
[
0.5087890625,
0.10809326171875,
0.88330078125,
0.1092529296875,
0.88330078125,
0.12353515625,
0.5087890625,
0.122314453125
],
[
0.12176513671875,
0.1258544921875,
0.489501953125,
0.1258544921875,
0.489501953125,
0.1395263671875,
0.12176513671875,
0.1395263671875
],
[
0.51318359375,
0.12353515625,
0.884765625,
0.1258544921875,
0.884765625,
0.14013671875,
0.51318359375,
0.1383056640625
],
[
0.12176513671875,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.12176513671875,
0.155029296875
],
[
0.5146484375,
0.141357421875,
0.8818359375,
0.141357421875,
0.8818359375,
0.155029296875,
0.5146484375,
0.155029296875
],
[
0.12005615234375,
0.15673828125,
0.486083984375,
0.15673828125,
0.486083984375,
0.17041015625,
0.12005615234375,
0.17041015625
],
[
0.5146484375,
0.15673828125,
0.87890625,
0.15673828125,
0.87890625,
0.17041015625,
0.5146484375,
0.17041015625
],
[
0.12005615234375,
0.17333984375,
0.47021484375,
0.17333984375,
0.47021484375,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.5146484375,
0.17333984375,
0.8818359375,
0.17333984375,
0.8818359375,
0.18701171875,
0.5146484375,
0.18701171875
],
[
0.5146484375,
0.188232421875,
0.880859375,
0.188232421875,
0.880859375,
0.201904296875,
0.5146484375,
0.201904296875
],
[
0.12176513671875,
0.1995849609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.213134765625,
0.12176513671875,
0.21435546875
],
[
0.51318359375,
0.2060546875,
0.8818359375,
0.2060546875,
0.8818359375,
0.2203369140625,
0.51318359375,
0.2203369140625
],
[
0.12176513671875,
0.2161865234375,
0.489501953125,
0.2161865234375,
0.489501953125,
0.2298583984375,
0.12176513671875,
0.2298583984375
],
[
0.5322265625,
0.221435546875,
0.8818359375,
0.221435546875,
0.8818359375,
0.236328125,
0.5322265625,
0.236328125
],
[
0.1234130859375,
0.2340087890625,
0.486083984375,
0.2340087890625,
0.486083984375,
0.24462890625,
0.1234130859375,
0.24462890625
],
[
0.5146484375,
0.2381591796875,
0.8818359375,
0.2381591796875,
0.8818359375,
0.251708984375,
0.5146484375,
0.251708984375
],
[
0.1234130859375,
0.248779296875,
0.48779296875,
0.248779296875,
0.48779296875,
0.262451171875,
0.1234130859375,
0.262451171875
],
[
0.51318359375,
0.254150390625,
0.880859375,
0.254150390625,
0.880859375,
0.267333984375,
0.51318359375,
0.267333984375
],
[
0.12176513671875,
0.26416015625,
0.489501953125,
0.26416015625,
0.489501953125,
0.27783203125,
0.12176513671875,
0.27783203125
],
[
0.51708984375,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.51708984375,
0.283935546875
],
[
0.12005615234375,
0.27978515625,
0.48779296875,
0.279052734375,
0.48779296875,
0.29345703125,
0.12005615234375,
0.29443359375
],
[
0.51708984375,
0.2880859375,
0.87890625,
0.2880859375,
0.87890625,
0.298583984375,
0.51708984375,
0.298583984375
],
[
0.12005615234375,
0.29638671875,
0.489501953125,
0.29638671875,
0.489501953125,
0.31005859375,
0.12005615234375,
0.31005859375
],
[
0.515625,
0.302734375,
0.87744140625,
0.302734375,
0.87744140625,
0.3134765625,
0.515625,
0.3134765625
],
[
0.12176513671875,
0.3134765625,
0.48779296875,
0.3134765625,
0.48779296875,
0.32666015625,
0.12176513671875,
0.32666015625
],
[
0.5146484375,
0.317626953125,
0.880859375,
0.317626953125,
0.880859375,
0.330810546875,
0.5146484375,
0.330810546875
],
[
0.12176513671875,
0.328857421875,
0.489501953125,
0.328857421875,
0.489501953125,
0.342529296875,
0.12176513671875,
0.342529296875
],
[
0.5146484375,
0.334228515625,
0.880859375,
0.334228515625,
0.880859375,
0.347412109375,
0.5146484375,
0.347412109375
],
[
0.1234130859375,
0.3466796875,
0.486083984375,
0.3466796875,
0.486083984375,
0.35693359375,
0.1234130859375,
0.35693359375
],
[
0.515625,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.364013671875,
0.515625,
0.364013671875
],
[
0.12176513671875,
0.361083984375,
0.48779296875,
0.361083984375,
0.48779296875,
0.374755859375,
0.12176513671875,
0.374755859375
],
[
0.5146484375,
0.365234375,
0.880859375,
0.366455078125,
0.880859375,
0.38134765625,
0.5146484375,
0.380126953125
],
[
0.12005615234375,
0.377197265625,
0.48779296875,
0.3759765625,
0.48779296875,
0.39013671875,
0.12005615234375,
0.391357421875
],
[
0.51318359375,
0.38134765625,
0.8818359375,
0.3818359375,
0.8818359375,
0.396728515625,
0.51318359375,
0.3955078125
],
[
0.12005615234375,
0.392578125,
0.489501953125,
0.392578125,
0.489501953125,
0.40625,
0.12005615234375,
0.40625
],
[
0.5146484375,
0.399658203125,
0.8271484375,
0.399658203125,
0.8271484375,
0.413330078125,
0.5146484375,
0.413330078125
],
[
0.12005615234375,
0.4091796875,
0.48779296875,
0.4091796875,
0.48779296875,
0.42333984375,
0.12005615234375,
0.42333984375
],
[
0.12005615234375,
0.42578125,
0.489501953125,
0.42578125,
0.489501953125,
0.439453125,
0.12005615234375,
0.439453125
],
[
0.51318359375,
0.42578125,
0.8095703125,
0.42578125,
0.8095703125,
0.440673828125,
0.51318359375,
0.440673828125
],
[
0.118408203125,
0.441162109375,
0.486083984375,
0.441162109375,
0.486083984375,
0.454833984375,
0.118408203125,
0.454833984375
],
[
0.515625,
0.44775390625,
0.8818359375,
0.44775390625,
0.8818359375,
0.46142578125,
0.515625,
0.46142578125
],
[
0.118408203125,
0.4560546875,
0.489501953125,
0.457763671875,
0.489501953125,
0.471435546875,
0.118408203125,
0.4697265625
],
[
0.5146484375,
0.46435546875,
0.880859375,
0.46435546875,
0.880859375,
0.47802734375,
0.5146484375,
0.47802734375
],
[
0.12005615234375,
0.474365234375,
0.246826171875,
0.474365234375,
0.246826171875,
0.488037109375,
0.12005615234375,
0.488037109375
],
[
0.515625,
0.479736328125,
0.8818359375,
0.478515625,
0.8818359375,
0.493408203125,
0.515625,
0.49462890625
],
[
0.1158447265625,
0.501953125,
0.340087890625,
0.5029296875,
0.340087890625,
0.51708984375,
0.1158447265625,
0.5166015625
],
[
0.51904296875,
0.496337890625,
0.72705078125,
0.496337890625,
0.72705078125,
0.51025390625,
0.51904296875,
0.51025390625
],
[
0.51123046875,
0.521484375,
0.884765625,
0.521484375,
0.884765625,
0.5498046875,
0.51123046875,
0.5498046875
],
[
0.12176513671875,
0.529296875,
0.486083984375,
0.529296875,
0.486083984375,
0.5419921875,
0.12176513671875,
0.5419921875
],
[
0.12005615234375,
0.54443359375,
0.48779296875,
0.5458984375,
0.48779296875,
0.56005859375,
0.12005615234375,
0.55859375
],
[
0.1234130859375,
0.5634765625,
0.486083984375,
0.5634765625,
0.486083984375,
0.57373046875,
0.1234130859375,
0.57373046875
],
[
0.51123046875,
0.55810546875,
0.88330078125,
0.55859375,
0.88330078125,
0.57666015625,
0.51123046875,
0.576171875
],
[
0.12005615234375,
0.57666015625,
0.48779296875,
0.57666015625,
0.48779296875,
0.59033203125,
0.12005615234375,
0.59033203125
],
[
0.515625,
0.57763671875,
0.88330078125,
0.57763671875,
0.88330078125,
0.59130859375,
0.515625,
0.59130859375
],
[
0.12005615234375,
0.595703125,
0.486083984375,
0.595703125,
0.486083984375,
0.60546875,
0.12005615234375,
0.60546875
],
[
0.51123046875,
0.59326171875,
0.8818359375,
0.59326171875,
0.8818359375,
0.60693359375,
0.51123046875,
0.60693359375
],
[
0.5087890625,
0.6025390625,
0.88671875,
0.60791015625,
0.88671875,
0.6328125,
0.5087890625,
0.62744140625
],
[
0.12005615234375,
0.609375,
0.48779296875,
0.609375,
0.48779296875,
0.62255859375,
0.12005615234375,
0.62255859375
],
[
0.12005615234375,
0.626953125,
0.484375,
0.626953125,
0.484375,
0.64013671875,
0.12005615234375,
0.64013671875
],
[
0.5087890625,
0.63134765625,
0.88671875,
0.63134765625,
0.88671875,
0.65478515625,
0.5087890625,
0.65478515625
],
[
0.12176513671875,
0.64111328125,
0.486083984375,
0.64111328125,
0.486083984375,
0.65478515625,
0.12176513671875,
0.65478515625
],
[
0.5146484375,
0.65185546875,
0.6513671875,
0.65185546875,
0.6513671875,
0.6650390625,
0.5146484375,
0.6650390625
],
[
0.12005615234375,
0.65771484375,
0.486083984375,
0.65771484375,
0.486083984375,
0.67138671875,
0.12005615234375,
0.67138671875
],
[
0.12005615234375,
0.6748046875,
0.435791015625,
0.6748046875,
0.435791015625,
0.6884765625,
0.12005615234375,
0.6884765625
],
[
0.5576171875,
0.6748046875,
0.82861328125,
0.6796875,
0.82861328125,
0.705078125,
0.5576171875,
0.69970703125
],
[
0.118408203125,
0.703125,
0.316650390625,
0.703125,
0.316650390625,
0.71630859375,
0.118408203125,
0.71630859375
],
[
0.53076171875,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.53076171875,
0.732421875
],
[
0.12176513671875,
0.724609375,
0.48779296875,
0.724609375,
0.48779296875,
0.73828125,
0.12176513671875,
0.73828125
],
[
0.51123046875,
0.732421875,
0.880859375,
0.732421875,
0.880859375,
0.74658203125,
0.51123046875,
0.74658203125
],
[
0.118408203125,
0.74072265625,
0.48779296875,
0.7392578125,
0.48779296875,
0.75439453125,
0.118408203125,
0.7548828125
],
[
0.5146484375,
0.75,
0.880859375,
0.75,
0.880859375,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.1259765625,
0.75732421875,
0.48779296875,
0.75732421875,
0.48779296875,
0.77099609375,
0.1259765625,
0.77099609375
],
[
0.5146484375,
0.76416015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.78076171875,
0.5146484375,
0.779296875
],
[
0.1175537109375,
0.77392578125,
0.183837890625,
0.76953125,
0.187255859375,
0.78369140625,
0.12005615234375,
0.7880859375
],
[
0.5146484375,
0.78076171875,
0.880859375,
0.78076171875,
0.880859375,
0.7958984375,
0.5146484375,
0.7958984375
],
[
0.139404296875,
0.7890625,
0.489501953125,
0.7890625,
0.489501953125,
0.802734375,
0.139404296875,
0.802734375
],
[
0.118408203125,
0.80419921875,
0.489501953125,
0.8046875,
0.489501953125,
0.8193359375,
0.118408203125,
0.818359375
],
[
0.51123046875,
0.798828125,
0.62744140625,
0.79736328125,
0.62744140625,
0.8125,
0.51123046875,
0.8134765625
],
[
0.1376953125,
0.8291015625,
0.489501953125,
0.830078125,
0.489501953125,
0.84423828125,
0.1376953125,
0.84326171875
],
[
0.51318359375,
0.826171875,
0.744140625,
0.826171875,
0.744140625,
0.8388671875,
0.51318359375,
0.8388671875
],
[
0.118408203125,
0.84423828125,
0.2822265625,
0.84326171875,
0.2822265625,
0.85791015625,
0.118408203125,
0.859375
],
[
0.51708984375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.51708984375,
0.86083984375
],
[
0.14111328125,
0.85546875,
0.489501953125,
0.85791015625,
0.489501953125,
0.87158203125,
0.14111328125,
0.869140625
],
[
0.51318359375,
0.8642578125,
0.8818359375,
0.8642578125,
0.8818359375,
0.87744140625,
0.51318359375,
0.87744140625
],
[
0.12176513671875,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8828125,
0.12176513671875,
0.8828125
],
[
0.12005615234375,
0.8828125,
0.489501953125,
0.88427734375,
0.489501953125,
0.89794921875,
0.12005615234375,
0.896484375
],
[
0.51318359375,
0.8798828125,
0.880859375,
0.8798828125,
0.880859375,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.12176513671875,
0.8984375,
0.21240234375,
0.8984375,
0.21240234375,
0.9091796875,
0.12176513671875,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.880859375,
0.8955078125,
0.880859375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"GCN的编码器，该编码器明确考虑了本地节",
"",
"体表示为一组节点。例如，图1中的KG节点",
"点上下文，并且与LSTM相比显示出优越的性",
"",
"“节点嵌”将由两个节点表示，一个用于令牌",
"能。最近，Koncel-Kedziorskietal.(2019）提出",
"“节点”，另一个用于令牌“嵌入”。形式上，我",
"了一种基于Transformer的方法，该方法通过",
"们将每个9e转换为一个新的图G=（V,&,R），",
"遵循自我关注策略参与节点邻域来计算节点表",
"其中实体eEVe的每个令牌都成为节点EV",
"示。相比之下，我们的模型专注于不同的全局",
"。我们将每个边（eh，r,et）Ee转换为一组边",
"和局部消息传递机制，捕获互补的图上下文",
"（具有相同的关系r），并将eh的每个令牌连",
"接到et的每个令牌。也就是说，当且仅当存在",
"IntegratingGlobalInformation。最近的工作",
"边缘(eh,r,et)E&e时，边缘（u,r,u)将属于8",
"试图集成全局上下文，以便在图形到文本生成",
"使得uEeh和Eét，其中éh和et被视",
"中学习更好的节点表示。为此，现有方法采用",
"为令牌集。我们用一个嵌人的hoERdu表示",
"人工全局节点与其他节点进行消息交换。这种",
"每个节点EV，从其相应的令牌生成。新的",
"策略可以被视为扩展了图形结构，但使用了类",
"图形9增加了模型的表示能力，因为它允许在",
"似的消息传递机制。特别是，Koncel-Kedziorski",
"令牌级别而不是实体级别学习节点嵌入。这对",
"etal.(2019)将全局节点添加到图中，并使用其",
"干文本生成特别重要，因为它允许模型更加灵",
"表示来初始化解码器。最近，Guoetal.(2019)",
"活，捕获实体令牌之间更丰富的关系。这也允",
"和Caiand Lam(2020)还采用了一个人工全局",
"许模型学习源和目标令牌之间的关系和注意力",
"节点，该节点具有与所有其他节点的直接边",
"函数。但是，它具有消除多词实体的自然顺序",
"以允许全局消息交换以进行AMR到文本的生",
"顺序的副作用。为了保留这些信息，我们使用",
"成。同样，Zhangetal.（2018）使用GRN模型",
"位置嵌人（Vaswanietal.,2017），i。e。，h成",
"的全局节点进行句子表示。与上述方法不同，",
"为的相应令牌嵌人和位置嵌人的总和",
"我们通过调查model替代方案而不是graph",
"structure更改，考虑在node级别而不是graph",
"3.2Graph Neural Networks (GNN)",
"级别集成全局和局部上下文。此外，我们将",
"多层GNN通过信息传播方案，基于其上下文",
"GAT和Transformer架构集成到一个统一的全",
"节点邻居和边缘特征，通过迭代学习节点的",
"局-局部模型中。",
"表示向量h来工作。更正式地说，第1层聚合",
"3Graph-to-Text Model",
"了U的上下文节点的表示：",
"h%() = AGGR((h-1), ru) : uE N(0),",
"本节首先描述（i）用于从输人（第E3.1节）创",
"建关系图的图转换，以及（ii）我们基于图注意",
"网络（GAT）（Velickovicetal2018）的框架的",
"其中AGGR(\")(）是一个聚合函数，由第1层上",
"图编码器，用于处理全局（第3.3节）和局部",
"的所有节点共享。ru表示u和u之间的关系。",
"（第3.4节）节点上下文。我们采用GAT是因为",
"N(u）是的一组上下文节点。在大多数GNN",
"中，上下文节点是与相邻的节点。h()是",
"它与Transformer架构（Vaswanietal，2017）密",
"切相关，它为建模全局节点上下文提供了一个",
"N()在1层的聚合上下文表示。h()用于更",
"方便的原型。然后，（i）我们提出了组合全局",
"新的表示形式：",
"和局部图编码器的策略（第3.5节）。最后，（iv）",
"我们描述了解码和训练过程（第3.6节）",
"hg = COMBINE( (h-1), (0)",
"3.1Graph Preparation",
"在L迭代之后，节点的表示对其L跳",
"我们将KG表示为具有实体节点eEVe和标记",
"邻域内的结构信息进行编码。AGGR()(）和",
"边(eh,r,et)E&e的多关系图?Ge=(Ve,&e,R)",
"COMBINE()（）的选择因特定的GNN模型而",
"其中rER表示从实体eh到et之间存在的",
"异。AGGR(）的一个示例是N(u)的表示形",
"关系。3",
"式的总和。COMBINE()）的一个示例是特征",
"与其他当前方法（Koncel-Kedziorski etal.，",
"2019；Moryossefetal.，2019）不同，我们将实",
"转换后的串联",
"2In this paper, multi-relational graphs refer to directed",
"3.3Global Graph Encoder",
"graphs with labelled edges",
"全局图编码器基于图的所有节点聚合用于更",
"3R contains relations both in canonical direction (e.g.",
"新每个节点的全局上下文（请参见图1℃）",
"used-for) and in inverse direction (e.g. used-for-inv), so that",
"the models consider the differences in the incoming and out-",
"我们使用注意力机制作为消息传递方案，将",
"goingrelations",
"Transformer的自注意力网络结构扩展为GAT"
] |
{
"arxiv_id": "2001.11003",
"language": "zh",
"page_idx": 2
}
|
||
2001.11003-zh-3
|
2001.11003-zh-3.jpg
|
积层，该节点将输入特征表示$h_v$作为输人，采
用 AGGR$^{(l)}(.)$作为：
$$h_{\mathcal{N}(v)}=\sum_{u\in\mathcal{V}}\alpha_{vu}\:W_{g}\:h_{u}\:,$$
(1)
, 其中$W_g\in\mathbb{R}^{d_v\times d_z}$ 是模型参数。注意力权重
$\alpha_{vu}$的计算公式为：
(2)
$$\alpha_{vu}=\frac{\exp(e_{vu})}{\sum_{k\in\mathcal{V}}\exp(e_{vk})}\:,$$
其中，
$$e_{vu}=\begin{pmatrix}\begin{pmatrix}W_qh_v\end{pmatrix}^\top\begin{pmatrix}W_kh_u\end{pmatrix}\end{pmatrix}/d_z$$
(3)
是注意力函数，它测量节点$u$的特征对节点$v$ 的全局重要性。$W_q,W_k\in\mathbb{R}^{d_v\times d_z}$是模型参数$d_z$是比例因子。为了捕获节点之间的不同关系，计算并连接了$K$ 独立的全局卷积：
$$\hat{h}_{\mathcal{N}(v)}=\left|\right|_{k=1}^Kh_{\mathcal{N}(v)}^{(k)}\:.$$
(4)
最后，我们使用层归一化(LayerNorm)和完全连接的前馈网络(FFN)定义 COMBINE$^{(l)}(.)$ ,其方式与转换器架构类似：
(5)
$$\begin{aligned}\hat h_v&=\mathrm{LayerNorm}(\hat h_{\mathcal{N}(v)}+h_v)\:,\\h_v^{global}&=\mathrm{FFN}(\hat h_v)\:+\:\hat h_{\mathcal{N}(v)}+h_v\:.\end{aligned}$$
(6)
请注意，全局编码器创建具有$\mathcal{O}(n^2)$边的人工完整图，并且不考虑边关系。特别是，如果考虑标记的边缘，则自注意空间复杂度将增加到$\Theta(|\mathcal{R}|n^2)$
## 3.4 Local Graph Encoder
表示$h_v^{global}$捕获从$v$到图中所有其他节点的宏关系。然而，这种表示缺乏关于$v$的局部邻域和图拓扑的结构信息。此外，它不会捕获节点之间的标记边(关系)(参见等式$\mathbb{E}$1 和 3)。为了捕获这些关键的图属性并施加强烈的关系归纳偏差，我们构建了一个图编码器，通过使用 GAT 的修改版本来聚合本地上下文，并增加了关系权重。特别是，我们计算节点$v\in\mathcal{V}$ 的局部卷积层，采用 AGGR$^(l)(.)$为：
(7)
$$h_{\mathcal{N}(v)}=\sum_{u\in\mathcal{N}(v)}\alpha_{vu}\:W_{r}\:h_{u}\:,$$
,其中$W_r\in\mathbb{R}^d_v\times d_z$编码$u$和$v$之间的关系$r\in\mathcal{R}$。$\mathcal{N}(v)$是与$v$和$v$本身相邻的一组节点。注意系数$\alpha_vu$的计算公式为：
(8)
$$\alpha_{vu}=\frac{\exp(e_{vu})}{\sum_{k\in\mathcal{N}(v)}\exp(e_{vk})}\:,$$
结构。特别是，我们计算节点$v\in\mathcal{V}$的全局卷 其中
$$e_{vu}=\sigma\big(a^\top[W_vh_v\parallel W_rh_u]\big)$$
(9)
是注意函数，考虑到边缘标签，该函数计算相邻节点的局部重要性。$\sigma$ 是激活函数，$\parallel$表示串联$,W_v\in\mathbb{R}^{d_v\times d_z}$和$a\in\mathbb{R}^2d_z$是模型参数
我们采用多头注意力来学习不同角度的局部关系，如公式$\mathbb{E}$4,生成 $\hat{h}_{\mathcal{N}(v)}$。最后，我们将 COMBINE$^{(l)}(.)$定义为：
$$h_v^{local}=\mathrm{RNN}(h_v,\hat{h}_{\mathcal{N}(v)})\:,$$
(10)
,其中我们采用门控循环单元(GRU)(Cho et al., 2014) 作为 RNN。GRU 促进本地层之间的信息传播。这种选择的动机是最近的作品(Xu et al., 2018; Dehmamy et al., 2019) , 从理论上证明了层之间共享信息有助于结构信号的传播。在类似的方向上，AMR 到文本生成模型在 GNN 层之间采用 LSTMs (Song et al., 2017) 和密集连接 (Guo et al., 2019)
3.5 Combining Global and Local Encodings
我们的目标是实现能够对输入图的全局和局部方面进行编码的图编码器。我们假设这两个信息源是互补的，两者的结合丰富了文本生成的节点表示。为了检验这一假设，我们研究了不同的组合架构
直观上，有两种用于集成两种类型的表示的一般方法。第一个是连接全局和局部上下文的向量，我们称之为parallel表示。第二种是形成一个管道，其中首先获得一个全局表示，然后将其用作输入，用于根据本地节点上下文计算精细表示。我们称这种方法为 cascaded 表示
可以在模型级别执行并行和级联集成，将全局和局部图编码器视为两个表示学习单元，而忽略内部结构。但是，由于我们的模型采用了多层体系结构，其中每一层都在表示中进行了抽象，因此我们可以替代地考虑在层级别上进行集成，以便可以捕获全局和局部上下文之间的更多交互。因此，我们提出了四种集成架构， 如图 2 所示。所有模型都有相同的目的，它们的相对优势应该通过经验来评估
Parallel Graph Encoding (PGE). 在此设置
中，我们以完全并行的结构组成全局和局部图编码器 (图E[2 a)。请注意，每个图形编码器可以有不同数量的层和注意头。最终的节点表示是两个图编码器的最后一层的局部和全局节
|
[
[
0.12005615234375,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.48779296875,
0.07659912109375,
0.55908203125,
0.07659912109375,
0.55908203125,
0.0914306640625,
0.48779296875,
0.0914306640625
],
[
0.12176513671875,
0.09320068359375,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.12176513671875,
0.10687255859375
],
[
0.118408203125,
0.10809326171875,
0.280517578125,
0.1092529296875,
0.280517578125,
0.12353515625,
0.118408203125,
0.122314453125
],
[
0.591796875,
0.11224365234375,
0.80419921875,
0.11224365234375,
0.80419921875,
0.1270751953125,
0.591796875,
0.1270751953125
],
[
0.8564453125,
0.1134033203125,
0.88330078125,
0.1134033203125,
0.88330078125,
0.1258544921875,
0.8564453125,
0.1258544921875
],
[
0.19140625,
0.1353759765625,
0.279541015625,
0.13720703125,
0.279541015625,
0.155029296875,
0.19140625,
0.152587890625
],
[
0.270263671875,
0.1383056640625,
0.45751953125,
0.1341552734375,
0.45751953125,
0.1519775390625,
0.270263671875,
0.1561279296875
],
[
0.462646484375,
0.1395263671875,
0.48779296875,
0.1395263671875,
0.48779296875,
0.15087890625,
0.462646484375,
0.15087890625
],
[
0.51708984375,
0.1478271484375,
0.8818359375,
0.1478271484375,
0.8818359375,
0.160888671875,
0.51708984375,
0.160888671875
],
[
0.118408203125,
0.163330078125,
0.490234375,
0.163330078125,
0.490234375,
0.18408203125,
0.118408203125,
0.18408203125
],
[
0.515625,
0.1644287109375,
0.8818359375,
0.1644287109375,
0.8818359375,
0.1776123046875,
0.515625,
0.1776123046875
],
[
0.51318359375,
0.17578125,
0.8671875,
0.177001953125,
0.8671875,
0.19482421875,
0.51318359375,
0.193603515625
],
[
0.12176513671875,
0.18408203125,
0.27197265625,
0.18408203125,
0.27197265625,
0.19775390625,
0.12176513671875,
0.19775390625
],
[
0.5341796875,
0.19775390625,
0.88330078125,
0.19775390625,
0.88330078125,
0.21142578125,
0.5341796875,
0.21142578125
],
[
0.2880859375,
0.208984375,
0.358642578125,
0.207275390625,
0.359375,
0.22265625,
0.28955078125,
0.2244873046875
],
[
0.2091064453125,
0.215576171875,
0.270263671875,
0.2203369140625,
0.267822265625,
0.2369384765625,
0.20654296875,
0.232177734375
],
[
0.51123046875,
0.21142578125,
0.88330078125,
0.2120361328125,
0.88330078125,
0.2298583984375,
0.51123046875,
0.2286376953125
],
[
0.466796875,
0.2197265625,
0.489501953125,
0.2197265625,
0.489501953125,
0.2298583984375,
0.466796875,
0.2298583984375
],
[
0.48779296875,
0.2203369140625,
0.50537109375,
0.21728515625,
0.5087890625,
0.2244873046875,
0.490234375,
0.22802734375
],
[
0.51318359375,
0.2286376953125,
0.73046875,
0.2298583984375,
0.73046875,
0.2476806640625,
0.51318359375,
0.2464599609375
],
[
0.118408203125,
0.251708984375,
0.166259765625,
0.251708984375,
0.166259765625,
0.2666015625,
0.118408203125,
0.2666015625
],
[
0.264404296875,
0.2244873046875,
0.39794921875,
0.2244873046875,
0.39794921875,
0.243408203125,
0.264404296875,
0.243408203125
],
[
0.599609375,
0.26416015625,
0.8017578125,
0.265380859375,
0.8017578125,
0.283203125,
0.599609375,
0.281982421875
],
[
0.8486328125,
0.268310546875,
0.8818359375,
0.268310546875,
0.8818359375,
0.28076171875,
0.8486328125,
0.28076171875
],
[
0.185546875,
0.276611328125,
0.491943359375,
0.276611328125,
0.491943359375,
0.297607421875,
0.185546875,
0.297607421875
],
[
0.515625,
0.3017578125,
0.880859375,
0.3017578125,
0.880859375,
0.3154296875,
0.515625,
0.3154296875
],
[
0.12005615234375,
0.308837890625,
0.48779296875,
0.308837890625,
0.48779296875,
0.322509765625,
0.12005615234375,
0.322509765625
],
[
0.1175537109375,
0.32177734375,
0.49365234375,
0.32177734375,
0.49365234375,
0.342529296875,
0.1175537109375,
0.342529296875
],
[
0.5146484375,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.33203125,
0.5146484375,
0.33203125
],
[
0.51708984375,
0.336181640625,
0.880859375,
0.336181640625,
0.880859375,
0.3466796875,
0.51708984375,
0.3466796875
],
[
0.12005615234375,
0.342529296875,
0.48779296875,
0.342529296875,
0.48779296875,
0.355712890625,
0.12005615234375,
0.355712890625
],
[
0.5146484375,
0.349853515625,
0.880859375,
0.349853515625,
0.880859375,
0.363525390625,
0.5146484375,
0.363525390625
],
[
0.12005615234375,
0.358154296875,
0.426513671875,
0.358154296875,
0.426513671875,
0.371826171875,
0.12005615234375,
0.371826171875
],
[
0.51904296875,
0.36767578125,
0.880859375,
0.36767578125,
0.880859375,
0.38134765625,
0.51904296875,
0.38134765625
],
[
0.2216796875,
0.385498046875,
0.307373046875,
0.38720703125,
0.306396484375,
0.40625,
0.219970703125,
0.40380859375
],
[
0.51318359375,
0.3818359375,
0.87890625,
0.3818359375,
0.87890625,
0.3955078125,
0.51318359375,
0.3955078125
],
[
0.291259765625,
0.3896484375,
0.361083984375,
0.377685546875,
0.36865234375,
0.39794921875,
0.297119140625,
0.4091796875
],
[
0.32080078125,
0.39013671875,
0.38623046875,
0.3955078125,
0.3837890625,
0.4091796875,
0.318115234375,
0.40380859375
],
[
0.465087890625,
0.3896484375,
0.489501953125,
0.3896484375,
0.489501953125,
0.402099609375,
0.465087890625,
0.402099609375
],
[
0.5146484375,
0.39794921875,
0.8818359375,
0.3984375,
0.8818359375,
0.413330078125,
0.5146484375,
0.412109375
],
[
0.515625,
0.4150390625,
0.7421875,
0.4150390625,
0.7421875,
0.4287109375,
0.515625,
0.4287109375
],
[
0.1376953125,
0.429931640625,
0.48779296875,
0.429931640625,
0.48779296875,
0.44482421875,
0.1376953125,
0.44482421875
],
[
0.12176513671875,
0.446533203125,
0.489501953125,
0.446533203125,
0.489501953125,
0.46142578125,
0.12176513671875,
0.46142578125
],
[
0.51123046875,
0.446533203125,
0.8759765625,
0.448974609375,
0.8759765625,
0.463134765625,
0.51123046875,
0.46142578125
],
[
0.491943359375,
0.453125,
0.50537109375,
0.453125,
0.50537109375,
0.457275390625,
0.491943359375,
0.457275390625
],
[
0.12005615234375,
0.463134765625,
0.349365234375,
0.463134765625,
0.349365234375,
0.47802734375,
0.12005615234375,
0.47802734375
],
[
0.515625,
0.473876953125,
0.88330078125,
0.473876953125,
0.88330078125,
0.48681640625,
0.515625,
0.48681640625
],
[
0.1956787109375,
0.486328125,
0.490234375,
0.488037109375,
0.490234375,
0.5048828125,
0.1956787109375,
0.5029296875
],
[
0.515625,
0.4892578125,
0.8818359375,
0.4892578125,
0.8818359375,
0.5029296875,
0.515625,
0.5029296875
],
[
0.16796875,
0.505859375,
0.490234375,
0.50830078125,
0.490234375,
0.529296875,
0.16796875,
0.52685546875
],
[
0.5146484375,
0.505859375,
0.880859375,
0.505859375,
0.880859375,
0.51953125,
0.5146484375,
0.51953125
],
[
0.51708984375,
0.521484375,
0.88330078125,
0.521484375,
0.88330078125,
0.53515625,
0.51708984375,
0.53515625
],
[
0.12005615234375,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.12005615234375,
0.5498046875
],
[
0.51708984375,
0.5390625,
0.62451171875,
0.5390625,
0.62451171875,
0.55029296875,
0.51708984375,
0.55029296875
],
[
0.12176513671875,
0.5546875,
0.486083984375,
0.5546875,
0.486083984375,
0.5654296875,
0.12176513671875,
0.5654296875
],
[
0.5341796875,
0.5556640625,
0.8818359375,
0.5556640625,
0.8818359375,
0.5693359375,
0.5341796875,
0.5693359375
],
[
0.12176513671875,
0.57080078125,
0.48779296875,
0.57080078125,
0.48779296875,
0.58447265625,
0.12176513671875,
0.58447265625
],
[
0.51123046875,
0.57080078125,
0.8818359375,
0.57177734375,
0.8818359375,
0.5859375,
0.51123046875,
0.5849609375
],
[
0.12176513671875,
0.58740234375,
0.1956787109375,
0.58740234375,
0.1956787109375,
0.5986328125,
0.12176513671875,
0.5986328125
],
[
0.515625,
0.58740234375,
0.880859375,
0.58740234375,
0.880859375,
0.60107421875,
0.515625,
0.60107421875
],
[
0.51708984375,
0.60400390625,
0.8818359375,
0.60400390625,
0.8818359375,
0.61767578125,
0.51708984375,
0.61767578125
],
[
0.12005615234375,
0.6123046875,
0.337646484375,
0.6123046875,
0.337646484375,
0.626953125,
0.12005615234375,
0.626953125
],
[
0.515625,
0.6181640625,
0.880859375,
0.6181640625,
0.880859375,
0.6318359375,
0.515625,
0.6318359375
],
[
0.1175537109375,
0.626953125,
0.490234375,
0.626953125,
0.490234375,
0.64794921875,
0.1175537109375,
0.64794921875
],
[
0.51123046875,
0.63525390625,
0.88330078125,
0.6357421875,
0.88330078125,
0.65087890625,
0.51123046875,
0.6494140625
],
[
0.12176513671875,
0.6494140625,
0.489501953125,
0.6494140625,
0.489501953125,
0.6630859375,
0.12176513671875,
0.6630859375
],
[
0.515625,
0.65478515625,
0.5341796875,
0.65478515625,
0.5341796875,
0.66455078125,
0.515625,
0.66455078125
],
[
0.118408203125,
0.66455078125,
0.486083984375,
0.66455078125,
0.486083984375,
0.677734375,
0.118408203125,
0.677734375
],
[
0.5341796875,
0.67041015625,
0.8818359375,
0.67041015625,
0.8818359375,
0.68408203125,
0.5341796875,
0.68408203125
],
[
0.12005615234375,
0.68115234375,
0.490234375,
0.68115234375,
0.490234375,
0.69482421875,
0.12005615234375,
0.69482421875
],
[
0.515625,
0.68701171875,
0.8818359375,
0.68701171875,
0.8818359375,
0.70068359375,
0.515625,
0.70068359375
],
[
0.12176513671875,
0.69775390625,
0.489501953125,
0.69775390625,
0.489501953125,
0.71142578125,
0.12176513671875,
0.71142578125
],
[
0.51318359375,
0.70166015625,
0.87890625,
0.70166015625,
0.87890625,
0.71533203125,
0.51318359375,
0.71533203125
],
[
0.12005615234375,
0.71337890625,
0.489501953125,
0.71337890625,
0.489501953125,
0.72705078125,
0.12005615234375,
0.72705078125
],
[
0.515625,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.732421875,
0.515625,
0.732421875
],
[
0.12176513671875,
0.728515625,
0.48779296875,
0.728515625,
0.48779296875,
0.7421875,
0.12176513671875,
0.7421875
],
[
0.5146484375,
0.7353515625,
0.8818359375,
0.7353515625,
0.8818359375,
0.7490234375,
0.5146484375,
0.7490234375
],
[
0.12176513671875,
0.7451171875,
0.489501953125,
0.7451171875,
0.489501953125,
0.7587890625,
0.12176513671875,
0.7587890625
],
[
0.51318359375,
0.75048828125,
0.880859375,
0.75048828125,
0.880859375,
0.76416015625,
0.51318359375,
0.76416015625
],
[
0.12005615234375,
0.76025390625,
0.4072265625,
0.76025390625,
0.4072265625,
0.77490234375,
0.12005615234375,
0.77490234375
],
[
0.5146484375,
0.7666015625,
0.880859375,
0.7666015625,
0.880859375,
0.77978515625,
0.5146484375,
0.77978515625
],
[
0.515625,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.515625,
0.79638671875
],
[
0.177978515625,
0.79052734375,
0.273681640625,
0.78369140625,
0.276123046875,
0.80419921875,
0.181396484375,
0.81005859375
],
[
0.26611328125,
0.7919921875,
0.46337890625,
0.783203125,
0.465087890625,
0.8056640625,
0.267822265625,
0.8154296875
],
[
0.4609375,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.80419921875,
0.4609375,
0.80419921875
],
[
0.515625,
0.7998046875,
0.771484375,
0.7998046875,
0.771484375,
0.8134765625,
0.515625,
0.8134765625
],
[
0.12005615234375,
0.8193359375,
0.489501953125,
0.8193359375,
0.489501953125,
0.83740234375,
0.12005615234375,
0.83740234375
],
[
0.5205078125,
0.830078125,
0.8818359375,
0.83056640625,
0.8818359375,
0.845703125,
0.5205078125,
0.84423828125
],
[
0.118408203125,
0.83740234375,
0.48779296875,
0.83837890625,
0.48779296875,
0.8525390625,
0.118408203125,
0.8515625
],
[
0.51708984375,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.51708984375,
0.86083984375
],
[
0.12005615234375,
0.85498046875,
0.38623046875,
0.85498046875,
0.38623046875,
0.8681640625,
0.12005615234375,
0.8681640625
],
[
0.5146484375,
0.86328125,
0.880859375,
0.86328125,
0.880859375,
0.876953125,
0.5146484375,
0.876953125
],
[
0.28955078125,
0.87744140625,
0.361083984375,
0.87744140625,
0.361083984375,
0.8935546875,
0.28955078125,
0.8935546875
],
[
0.51708984375,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.51708984375,
0.8935546875
],
[
0.198974609375,
0.8857421875,
0.255126953125,
0.89111328125,
0.252685546875,
0.90673828125,
0.1956787109375,
0.90185546875
],
[
0.465087890625,
0.88818359375,
0.489501953125,
0.88818359375,
0.489501953125,
0.90087890625,
0.465087890625,
0.90087890625
],
[
0.2493896484375,
0.89453125,
0.41064453125,
0.892578125,
0.41064453125,
0.91455078125,
0.2493896484375,
0.91650390625
],
[
0.515625,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"结构。特别是，我们计算节点UEV的全局卷",
"其中,",
"积层，该节点将输入特征表示h，作为输入，采",
"用AGGR()（）作为：",
"evu = o(a' [W,h, Il W.hul)",
"(9)",
"hr(u) = 2",
"uey Quu Wghu,",
"(1)",
"是注意函数，考虑到边缘标签，该函数计算相",
"，其中W。ERduxd是模型参数。注意力权重",
"邻节点的局部重要性。α是激活函数，Ⅱ表示",
"串联，WuERduxd和aER2d是模型参数",
"Quu的计算公式为：",
"我们采用多头注意力来学习不同角度的局部",
"exp(evu)",
"Qu=",
"关系，如公式E4，生成hn(o）。最后，我们",
"(2)",
"",
"将COMBINE(\")（）定义为：",
"其中,",
"Zkev exp(euk) \"",
"hlocal = RNN(hy, hN(u) ,",
"(10)",
"eu = (Wghu) (Wthu)) /dz (3)",
"2其中我们采用门控循环单元（GRU）（Cho",
"是注意力函数，它测量节点u的特征对节点V",
"的全局重要性。Wa,WsERdxd是模型参数，",
"etal.,2014）作为RNN。GRU促进本地层之间",
"的信息传播。这种选择的动机是最近的作品",
"dz是比例因子。为了捕获节点之间的不同关",
"(Xu et al.,2018;Dehmamy et al.,2019)，从理论",
"系，计算并连接了K独立的全局卷积",
"上证明了层之间共享信息有助于结构信号的传",
"hN() =Il",
"播。在类似的方向上，AMR到文本生成模型",
"k-,h()",
"hN(u)",
"(4)",
"在GNN层之间采用LSTMs(Songetal.,2017)",
"和密集连接（Guoetal.,2019)",
"最后，我们使用层归一化（LayerNorm）和完",
"全连接的前馈网络（FFN）定义COMBINE(）（）",
"3.5Combining Global and Local Encodings",
"",
"，其方式与转换器架构类似：",
"我们的目标是实现能够对输入图的全局和局部",
"hu = LayerNorm(hn() + hu), (5)",
"方面进行编码的图编码器。我们假设这两个信",
"hglobal =FFN(hy) + hN(o) +hy.(6)",
"息源是互补的，两者的结合丰富了文本生成的",
"节点表示。为了检验这一假设，我们研究了不",
"请注意，全局编码器创建具有O(n2）边的人工",
"同的组合架构",
"完整图，并且不考虑边关系。特别是，如果考",
"直观上，有两种用于集成两种类型的表示的",
"虑标记的边缘，则自注意空间复杂度将增加到",
"一般方法。第一个是连接全局和局部上下文的",
"O(R/n-)",
"向量，我们称之为parallel表示。第二种是形",
"成一个管道，其中首先获得一个全局表示，然",
"3.4Local Graph Encoder",
"后将其用作输入，用于根据本地节点上下文计",
"表示hglobal捕获从到图中所有其他节点的",
"算精细表示。我们称这种方法为cascaded表",
"宏关系。然而，这种表示缺乏关于的局部邻",
"示",
"域和图拓扑的结构信息。此外，它不会捕获节",
"可以在模型级别执行并行和级联集成，将全",
"点之间的标记边（关系）（参见等式E1和3）。",
"局和局部图编码器视为两个表示学习单元，而",
"为了捕获这些关键的图属性并施加强烈的关系",
"忽略内部结构。但是，由于我们的模型采用了",
"归纳偏差，我们构建了一个图编码器，通过使",
"多层体系结构，其中每一层都在表示中进行了",
"用GAT的修改版本来聚合本地上下文，并增",
"抽象，因此我们可以替代地考虑在层级别上进",
"加了关系权重。特别是，我们计算节点EV",
"行集成，以便可以捕获全局和局部上下文之间",
"的局部卷积层，采用AGGR①）为：",
"的更多交互。因此，我们提出了四种集成架构",
"如图2所示。所有模型都有相同的目的，它们",
"hN(u) =",
"ueN(e) Qu W, hu,",
"(7)",
"的相对优势应该通过经验来评估",
"，其中W,eRduxdz编码u和u之间的关系",
"Parallel GraphEncoding (PGE).在此设置",
"rER。(u)是与u和本身相邻的一组节",
"中，我们以完全并行的结构组成全局和局部图",
"点。注意系数αuu的计算公式为：",
"编码器（图E2a）。请注意，每个图形编码器",
"exp(evu)",
"可以有不同数量的层和注意头。最终的节点表",
"Qu=",
"(8)",
"Ehen(u) exp(euk) '",
"示是两个图编码器的最后一层的局部和全局节"
] |
{
"arxiv_id": "2001.11003",
"language": "zh",
"page_idx": 3
}
|
||
2001.11003-zh-4
|
2001.11003-zh-4.jpg
|
<FigureHere>
Figure 2: Overview of the proposed encoder architectures. (a) Parallel Graph Encoder( PGE )with separated parallel global and local node encoders. (b) Cascaded Graph Encoder( CGE )with separated cascaded encoders. c) PGE-LW : global and local node representations are concatenated layer-wise. d) CGE-LW : Both node representations are cascaded layer-wise.
<table>
<tbody>
<tr>
<th> </th>
<th>$\#$train</th>
<th>$\#$dev</th>
<th>$\#$test</th>
<th>$\#$ relations </th>
<th>$\#$entities avg</th>
<th>$g\#$nodes avg 1 4</th>
<th>$\#edges$ avg </th>
<th>$\arg\#\operatorname{CC}$</th>
<th>length $\arg$</th>
</tr>
<tr>
<td>AGENDA</td>
<td>38,720</td>
<td>1,000</td>
<td>1,000</td>
<td>7</td>
<td>12.4</td>
<td>44.3</td>
<td>68.6</td>
<td>19.1</td>
<td>140.3</td>
</tr>
<tr>
<td>WebNLG</td>
<td>18,102</td>
<td>872</td>
<td>971</td>
<td>373</td>
<td>4.0</td>
<td>34.9</td>
<td>101.0</td>
<td>1.5</td>
<td>24.2</td>
</tr>
</tbody>
</table>
Table l: Data statistics. Nodes, edges and CC values are calculated after the graph transformation. The average
values are calculated for all splits (training, dev and test sets). CC refers to the number of connected components.
点表示的串联：
2 d)。特别是，我们计算每个编码器层如下：
$$\begin{aligned}h_{v}^{global}&=\mathrm{GE}(h_{v}^{0},\{h_{u}^{0}:u\in\mathcal{V}\})&h_{v}^{global}=\mathrm{GE}_{l}(h_{v}^{l-1},\{h_{u}^{l-1}:u\in\mathcal{V}\})\\h_{v}^{local}&=\mathrm{LE}(h_{v}^{0},\{h_{u}^{0}:u\in\mathcal{N}(v)\})&h_{v}^{l}=\mathrm{LE}_{l}(h_{v}^{global},\{h_{u}^{global}:u\in\mathcal{N}(v)\}).\\h_{v}&=[\:h_{v}^{global}\parallel h_{v}^{local}\:]\:,\quad(11)\end{aligned}$$
### 3.6 Decoder and Training
其中 GE 和 LE 分别表示全局和局部图编码器。$h_v^0$是在两个编码器的第一层中使用的初始节点嵌人。
Cascaded Graph Encoding (CGE). 我们级联本地和全局图编码器，如图E2 b 所示。我们首先计算全局上下文化的节点嵌入，然后使用本地节点上下文对其进行改进。$h_v^0$是全局编码器的初始输入$,h_v^{global}$是本地编码器的初始输入。特别是，最终的节点表示计算如下：
我们的解码器遵循变压器解码器(Vaswani et al., 2017) 的核心架构。通过在编码器的输出(节点嵌人$h_v$)和先前生成的令牌(令牌嵌人)上执行多头关注来更新每个时间步$t$。我们设置中的另一个挑战是生成多句子输出。为了鼓励模型生成更长的文本，我们采用长度惩罚 (Wu et al., 2016) 来改进纯最大概率波束搜索
训练模型以优化每个黄金标准输出文本的负对数似然。我们采用标签平滑正则化来防止模型在训练过程中过于自信地预测令牌，并且泛化不佳
$$\begin{aligned}h_{v}^{global}&=\mathrm{GE}(h_{v}^{0},\{h_{u}^{0}:u\in\mathcal{V}\})\\h_{v}&=\mathrm{LE}(h_{v}^{global},\{h_{u}^{global}:u\in\mathcal{N}(v)\}).\end{aligned}$$
### 4 Data and Preprocessing
Layer-wise Parallel Graph Encoding. 为了允许两种类型的图形上下文信息之间的细粒度交互，我们还以分层(LW)方式组合了编码器。如图$\overline{\mathrm{E}}$2 c 所示，对于每个图层，我们在并行结构( PGE-LW )中同时采用全局和局部编码器。更准确地说，每个编码器层的计算如下：
$$\begin{aligned}h_{v}^{global}&=\mathrm{GE}_{l}(h_{v}^{l-1},\{h_{u}^{l-1}:u\in\mathcal{V}\})\\h_{v}^{local}&=\mathrm{LE}_{l}(h_{v}^{l-1},\{h_{u}^{l-1}:u\in\mathcal{N}(v)\})\\h_{v}^{l}&=[\:h_{v}^{global}\parallel h_{v}^{local}\:]\:,\end{aligned}$$
我们证明了我们的模型在两个数据集上的有效性：AGENDA (Koncel-Kedziorski et al., 2019) 和 WebNLG (Gardent et al.,2017)。表 1 显示了两个数据集的统计信息
$\mathbf{AGENDA.}$在此数据集中，KG与从12个顶级 AI 会议记录中提取的科学摘要配对。每个实例由论文标题、KG 和论文摘要组成。实体对应于通常是多词表达的科学术语(合并了共同引用实体)。我们将标题中的每个令牌视为一个节点，创建一个以标题和 KG 令牌为节点的唯一图。如表 1 所示，平均输出长度相当大，因为目标输出是多句子摘要。
(13)
其中 GE$_l$和 LE$_l$分别指全局和局部图编码器的
$l$层。
Layer-wise Cascaded Graph Encoding. 我们还建议分层级联图形编码器( CGE-LW ,图$\boxed{\mathrm{E}}$
|
[
[
0.139404296875,
0.08013916015625,
0.306396484375,
0.08013916015625,
0.306396484375,
0.09027099609375,
0.139404296875,
0.09027099609375
],
[
0.328369140625,
0.08013916015625,
0.496337890625,
0.08013916015625,
0.496337890625,
0.09027099609375,
0.328369140625,
0.09027099609375
],
[
0.51904296875,
0.08013916015625,
0.68701171875,
0.08013916015625,
0.68701171875,
0.09027099609375,
0.51904296875,
0.09027099609375
],
[
0.70947265625,
0.08013916015625,
0.87890625,
0.08013916015625,
0.87890625,
0.09027099609375,
0.70947265625,
0.09027099609375
],
[
0.313232421875,
0.0968017578125,
0.50146484375,
0.097412109375,
0.50146484375,
0.11102294921875,
0.313232421875,
0.10986328125
],
[
0.74462890625,
0.1009521484375,
0.8447265625,
0.1009521484375,
0.8447265625,
0.10809326171875,
0.74462890625,
0.10809326171875
],
[
0.219970703125,
0.1015625,
0.300537109375,
0.10272216796875,
0.300537109375,
0.114013671875,
0.219970703125,
0.1134033203125
],
[
0.1435546875,
0.1015625,
0.22412109375,
0.1015625,
0.22412109375,
0.1134033203125,
0.1435546875,
0.1134033203125
],
[
0.52490234375,
0.10272216796875,
0.60791015625,
0.10272216796875,
0.60791015625,
0.1134033203125,
0.52490234375,
0.1134033203125
],
[
0.60205078125,
0.10272216796875,
0.67822265625,
0.10272216796875,
0.67822265625,
0.1134033203125,
0.60205078125,
0.1134033203125
],
[
0.50390625,
0.10986328125,
0.52294921875,
0.10986328125,
0.52294921875,
0.114013671875,
0.50390625,
0.114013671875
],
[
0.6943359375,
0.10986328125,
0.7080078125,
0.10986328125,
0.7080078125,
0.114013671875,
0.6943359375,
0.114013671875
],
[
0.6171875,
0.11224365234375,
0.6708984375,
0.11224365234375,
0.6708984375,
0.122314453125,
0.6171875,
0.122314453125
],
[
0.139404296875,
0.11102294921875,
0.306396484375,
0.1134033203125,
0.306396484375,
0.12469482421875,
0.139404296875,
0.122314453125
],
[
0.32080078125,
0.11639404296875,
0.49951171875,
0.1181640625,
0.49951171875,
0.1318359375,
0.32080078125,
0.1300048828125
],
[
0.54150390625,
0.114013671875,
0.58447265625,
0.114013671875,
0.58447265625,
0.1217041015625,
0.54150390625,
0.1217041015625
],
[
0.7421875,
0.11639404296875,
0.84619140625,
0.11639404296875,
0.84619140625,
0.1270751953125,
0.7421875,
0.1270751953125
],
[
0.1351318359375,
0.12469482421875,
0.1502685546875,
0.12469482421875,
0.1502685546875,
0.133056640625,
0.1351318359375,
0.133056640625
],
[
0.2225341796875,
0.1258544921875,
0.2392578125,
0.1258544921875,
0.2392578125,
0.133056640625,
0.2225341796875,
0.133056640625
],
[
0.172119140625,
0.1341552734375,
0.274658203125,
0.135986328125,
0.274658203125,
0.15087890625,
0.172119140625,
0.1484375
],
[
0.366943359375,
0.13720703125,
0.46337890625,
0.13720703125,
0.46337890625,
0.1484375,
0.366943359375,
0.1484375
],
[
0.5556640625,
0.13720703125,
0.654296875,
0.1383056640625,
0.654296875,
0.149658203125,
0.5556640625,
0.1484375
],
[
0.744140625,
0.1341552734375,
0.8447265625,
0.135986328125,
0.8447265625,
0.15087890625,
0.744140625,
0.1484375
],
[
0.118408203125,
0.162109375,
0.884765625,
0.162109375,
0.884765625,
0.17578125,
0.118408203125,
0.17578125
],
[
0.118408203125,
0.17578125,
0.8818359375,
0.17578125,
0.8818359375,
0.190673828125,
0.118408203125,
0.190673828125
],
[
0.118408203125,
0.190673828125,
0.88330078125,
0.190673828125,
0.88330078125,
0.2037353515625,
0.118408203125,
0.2037353515625
],
[
0.118408203125,
0.204833984375,
0.3896484375,
0.204833984375,
0.3896484375,
0.218505859375,
0.118408203125,
0.218505859375
],
[
0.2091064453125,
0.235107421875,
0.79443359375,
0.235107421875,
0.79443359375,
0.248779296875,
0.2091064453125,
0.248779296875
],
[
0.78515625,
0.236328125,
0.873046875,
0.236328125,
0.873046875,
0.2476806640625,
0.78515625,
0.2476806640625
],
[
0.1275634765625,
0.251708984375,
0.353515625,
0.251708984375,
0.353515625,
0.265380859375,
0.1275634765625,
0.265380859375
],
[
0.405517578125,
0.251708984375,
0.4189453125,
0.254638671875,
0.414794921875,
0.26416015625,
0.4013671875,
0.26123046875
],
[
0.476806640625,
0.2529296875,
0.5087890625,
0.2529296875,
0.5087890625,
0.26416015625,
0.476806640625,
0.26416015625
],
[
0.57275390625,
0.2529296875,
0.603515625,
0.2529296875,
0.603515625,
0.26416015625,
0.57275390625,
0.26416015625
],
[
0.66015625,
0.2529296875,
0.6943359375,
0.2529296875,
0.6943359375,
0.26416015625,
0.66015625,
0.26416015625
],
[
0.7421875,
0.2529296875,
0.7734375,
0.2529296875,
0.7734375,
0.26416015625,
0.7421875,
0.26416015625
],
[
0.81689453125,
0.2529296875,
0.85546875,
0.2529296875,
0.85546875,
0.26416015625,
0.81689453125,
0.26416015625
],
[
0.130126953125,
0.26318359375,
0.258544921875,
0.26318359375,
0.258544921875,
0.276611328125,
0.130126953125,
0.276611328125
],
[
0.2626953125,
0.26416015625,
0.30908203125,
0.26416015625,
0.30908203125,
0.275634765625,
0.2626953125,
0.275634765625
],
[
0.3046875,
0.26416015625,
0.3525390625,
0.26416015625,
0.3525390625,
0.275634765625,
0.3046875,
0.275634765625
],
[
0.38623046875,
0.26318359375,
0.418212890625,
0.26318359375,
0.418212890625,
0.275634765625,
0.38623046875,
0.275634765625
],
[
0.478515625,
0.26416015625,
0.51123046875,
0.26416015625,
0.51123046875,
0.275634765625,
0.478515625,
0.275634765625
],
[
0.57080078125,
0.26416015625,
0.60546875,
0.26416015625,
0.60546875,
0.275634765625,
0.57080078125,
0.275634765625
],
[
0.654296875,
0.26416015625,
0.6962890625,
0.26416015625,
0.6962890625,
0.275634765625,
0.654296875,
0.275634765625
],
[
0.74462890625,
0.26416015625,
0.77490234375,
0.26416015625,
0.77490234375,
0.275634765625,
0.74462890625,
0.275634765625
],
[
0.8212890625,
0.26416015625,
0.85546875,
0.26416015625,
0.85546875,
0.276611328125,
0.8212890625,
0.276611328125
],
[
0.1175537109375,
0.283935546875,
0.88330078125,
0.2861328125,
0.88330078125,
0.2998046875,
0.1175537109375,
0.297607421875
],
[
0.118408203125,
0.300537109375,
0.87890625,
0.300537109375,
0.87890625,
0.314208984375,
0.118408203125,
0.314208984375
],
[
0.118408203125,
0.328857421875,
0.2392578125,
0.327880859375,
0.2392578125,
0.342529296875,
0.118408203125,
0.34326171875
],
[
0.51318359375,
0.327880859375,
0.86376953125,
0.327880859375,
0.86376953125,
0.341552734375,
0.51318359375,
0.341552734375
],
[
0.16796875,
0.355224609375,
0.408935546875,
0.35693359375,
0.408935546875,
0.374755859375,
0.16796875,
0.372314453125
],
[
0.51318359375,
0.35693359375,
0.78857421875,
0.359375,
0.78857421875,
0.377197265625,
0.51318359375,
0.374755859375
],
[
0.173828125,
0.3759765625,
0.435791015625,
0.3759765625,
0.435791015625,
0.39794921875,
0.173828125,
0.39794921875
],
[
0.5400390625,
0.377197265625,
0.884765625,
0.377685546875,
0.884765625,
0.399658203125,
0.5400390625,
0.3984375
],
[
0.19482421875,
0.39794921875,
0.37451171875,
0.39794921875,
0.37451171875,
0.418701171875,
0.19482421875,
0.418701171875
],
[
0.453369140625,
0.402587890625,
0.48779296875,
0.402587890625,
0.48779296875,
0.4150390625,
0.453369140625,
0.4150390625
],
[
0.118408203125,
0.43115234375,
0.489501953125,
0.431640625,
0.489501953125,
0.446533203125,
0.118408203125,
0.4453125
],
[
0.118408203125,
0.446533203125,
0.489501953125,
0.44775390625,
0.489501953125,
0.4619140625,
0.118408203125,
0.46142578125
],
[
0.509765625,
0.4423828125,
0.736328125,
0.44482421875,
0.736328125,
0.4619140625,
0.509765625,
0.460205078125
],
[
0.12005615234375,
0.46435546875,
0.185546875,
0.46435546875,
0.185546875,
0.478515625,
0.12005615234375,
0.478515625
],
[
0.5146484375,
0.4697265625,
0.88330078125,
0.4697265625,
0.88330078125,
0.482666015625,
0.5146484375,
0.482666015625
],
[
0.124267578125,
0.490478515625,
0.489501953125,
0.490478515625,
0.489501953125,
0.5048828125,
0.124267578125,
0.5048828125
],
[
0.51318359375,
0.485107421875,
0.880859375,
0.485107421875,
0.880859375,
0.498779296875,
0.51318359375,
0.498779296875
],
[
0.12005615234375,
0.50732421875,
0.48779296875,
0.50732421875,
0.48779296875,
0.52099609375,
0.12005615234375,
0.52099609375
],
[
0.5146484375,
0.501953125,
0.8818359375,
0.501953125,
0.8818359375,
0.515625,
0.5146484375,
0.515625
],
[
0.12176513671875,
0.5224609375,
0.48779296875,
0.5224609375,
0.48779296875,
0.5361328125,
0.12176513671875,
0.5361328125
],
[
0.5146484375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.5146484375,
0.53076171875
],
[
0.12005615234375,
0.5380859375,
0.48779296875,
0.5380859375,
0.48779296875,
0.552734375,
0.12005615234375,
0.552734375
],
[
0.5146484375,
0.533203125,
0.880859375,
0.533203125,
0.880859375,
0.54638671875,
0.5146484375,
0.54638671875
],
[
0.118408203125,
0.5498046875,
0.491943359375,
0.5498046875,
0.491943359375,
0.57080078125,
0.118408203125,
0.57080078125
],
[
0.51904296875,
0.548828125,
0.8818359375,
0.548828125,
0.8818359375,
0.5634765625,
0.51904296875,
0.5634765625
],
[
0.12005615234375,
0.57080078125,
0.45751953125,
0.57080078125,
0.45751953125,
0.5849609375,
0.12005615234375,
0.5849609375
],
[
0.515625,
0.5654296875,
0.8818359375,
0.5654296875,
0.8818359375,
0.5791015625,
0.515625,
0.5791015625
],
[
0.5146484375,
0.580078125,
0.53662109375,
0.58056640625,
0.53564453125,
0.595703125,
0.51318359375,
0.59423828125
],
[
0.1292724609375,
0.59326171875,
0.3720703125,
0.595703125,
0.3720703125,
0.61767578125,
0.1292724609375,
0.615234375
],
[
0.5341796875,
0.5986328125,
0.880859375,
0.5986328125,
0.880859375,
0.6123046875,
0.5341796875,
0.6123046875
],
[
0.156982421875,
0.6181640625,
0.490234375,
0.6181640625,
0.490234375,
0.6396484375,
0.156982421875,
0.6396484375
],
[
0.515625,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.515625,
0.62890625
],
[
0.515625,
0.63134765625,
0.8818359375,
0.63134765625,
0.8818359375,
0.64453125,
0.515625,
0.64453125
],
[
0.5146484375,
0.646484375,
0.57421875,
0.646484375,
0.57421875,
0.6611328125,
0.5146484375,
0.6611328125
],
[
0.1234130859375,
0.65283203125,
0.4248046875,
0.65283203125,
0.4248046875,
0.66650390625,
0.1234130859375,
0.66650390625
],
[
0.418212890625,
0.65478515625,
0.486083984375,
0.65478515625,
0.486083984375,
0.66650390625,
0.418212890625,
0.66650390625
],
[
0.12176513671875,
0.66943359375,
0.48779296875,
0.66943359375,
0.48779296875,
0.68310546875,
0.12176513671875,
0.68310546875
],
[
0.515625,
0.68115234375,
0.52880859375,
0.68115234375,
0.52880859375,
0.689453125,
0.515625,
0.689453125
],
[
0.54150390625,
0.681640625,
0.54736328125,
0.681640625,
0.54736328125,
0.68701171875,
0.54150390625,
0.68701171875
],
[
0.5458984375,
0.6767578125,
0.75732421875,
0.6787109375,
0.75732421875,
0.693359375,
0.5458984375,
0.69140625
],
[
0.12176513671875,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69873046875,
0.12176513671875,
0.69873046875
],
[
0.118408203125,
0.69970703125,
0.48779296875,
0.70068359375,
0.48779296875,
0.71533203125,
0.118408203125,
0.71435546875
],
[
0.515625,
0.7060546875,
0.8818359375,
0.7060546875,
0.8818359375,
0.7197265625,
0.515625,
0.7197265625
],
[
0.12005615234375,
0.71728515625,
0.48779296875,
0.71630859375,
0.48779296875,
0.73095703125,
0.12005615234375,
0.732421875
],
[
0.5146484375,
0.7216796875,
0.8818359375,
0.72021484375,
0.8818359375,
0.7353515625,
0.5146484375,
0.736328125
],
[
0.12005615234375,
0.73388671875,
0.478515625,
0.73388671875,
0.478515625,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.515625,
0.73828125,
0.88330078125,
0.73828125,
0.88330078125,
0.7529296875,
0.515625,
0.7529296875
],
[
0.1502685546875,
0.7587890625,
0.4248046875,
0.76025390625,
0.4248046875,
0.779296875,
0.1502685546875,
0.77783203125
],
[
0.515625,
0.75439453125,
0.71728515625,
0.75439453125,
0.71728515625,
0.76708984375,
0.515625,
0.76708984375
],
[
0.156982421875,
0.78076171875,
0.453369140625,
0.78076171875,
0.453369140625,
0.8017578125,
0.156982421875,
0.8017578125
],
[
0.51904296875,
0.78076171875,
0.8818359375,
0.78076171875,
0.8818359375,
0.79443359375,
0.51904296875,
0.79443359375
],
[
0.177978515625,
0.8017578125,
0.344970703125,
0.8017578125,
0.344970703125,
0.822265625,
0.177978515625,
0.822265625
],
[
0.5146484375,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.5146484375,
0.8125
],
[
0.456787109375,
0.80712890625,
0.48779296875,
0.80712890625,
0.48779296875,
0.8193359375,
0.456787109375,
0.8193359375
],
[
0.515625,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.515625,
0.82763671875
],
[
0.118408203125,
0.8349609375,
0.48779296875,
0.8359375,
0.48779296875,
0.85107421875,
0.118408203125,
0.849609375
],
[
0.5146484375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.12005615234375,
0.8525390625,
0.1561279296875,
0.8525390625,
0.1561279296875,
0.865234375,
0.12005615234375,
0.865234375
],
[
0.51708984375,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.51708984375,
0.86083984375
],
[
0.52490234375,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.87744140625,
0.52490234375,
0.87744140625
],
[
0.12176513671875,
0.87744140625,
0.453369140625,
0.87890625,
0.453369140625,
0.8935546875,
0.12176513671875,
0.892578125
],
[
0.44921875,
0.8798828125,
0.486083984375,
0.8798828125,
0.486083984375,
0.89111328125,
0.44921875,
0.89111328125
],
[
0.515625,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.515625,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.4306640625,
0.8955078125,
0.4306640625,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.424072265625,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.90673828125,
0.424072265625,
0.90673828125
],
[
0.515625,
0.8955078125,
0.779296875,
0.8955078125,
0.779296875,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"(ContextualisedEmbeddings",
"ContextualisedEmbeddings",
"(ContextualisedEmbeddings",
"(Contextualised Embeddings",
"MxLocal Node Encoder Layer",
"LocalNodeEncoder",
"LocalNode",
"Global Node",
"Global Node",
"Local Node",
"NYI",
"N",
"Encoder",
"(EncoderLayerEncoderLayer",
"NxGlobal Node Encoder Layer",
"Encoder",
"Global Node Encoder",
"Nx",
"Mx",
"Node Embeddings",
"Node Embeddings",
"Node Embeddings",
" Node Embeddings",
"Figure 2:Overview of the proposed encoder architectures. (a) Parallel Graph Encoder (PGE with separated",
"parallel global and local node encoders. (b) Cascaded Graph Encoder ( CGE ) with separated cascaded encoders.",
"c) PGE-Lw :global and local node representations are concatenated layer-wise.d)CGE-Lw :Both node",
"representations are cascaded layer-wise.",
"# train # dev # test # relations avg #entities avg # nodes avg # edges avg #CC",
"Cavglength",
"AGENDA 38,720 1,000 1,000",
"7",
"12.4",
"44.3",
"68.6",
"19.1",
"140.3",
"WebNLG 18,102",
"872",
"971",
"373",
"4.0",
"34.9",
"101.0",
"1.5",
"24.2",
"Table 1: Data statistics. Nodes, edges and CC values are calculated after the graph transformation. The average",
"values are calculated for all splits (training, dev and test sets). CC refers to the number of connected components.",
"点表示的串联：",
"2d)。特别是，我们计算每个编码器层如下",
"hglobal =- GE(ho, (ha : u E V))",
"hglobal = GE;(h'-1, [hl-l : u V))",
"hlocal = LE(h, (h : uE N(u))",
"h', = LE;(hglobal,[hglobal: uE N(u),. (14)",
"hy = [hglobal |Ihlocal ],",
"(11)",
"其中GE和LE分别表示全局和局部图编码器。",
"ho是在两个编码器的第一层中使用的初始节",
"3.6Decoder and Training",
"点嵌入。",
"我们的解码器遵循变压器解码器（Vaswani",
"CascadedGraphEncoding(CGE).我们级联",
"etal.，2017）的核心架构。通过在编码器的输",
"本地和全局图编码器，如图E2b所示。我们",
"出（节点嵌入h）和先前生成的令牌（令牌嵌",
"首先计算全局上下文化的节点嵌入，然后使用",
"人）上执行多头关注来更新每个时间步t。我",
"本地节点上下文对其进行改进。ho是全局编",
"们设置中的另一个挑战是生成多句子输出。为",
"码器的初始输人，hgiobal 是本地编码器的初始",
"了鼓励模型生成更长的文本，我们采用长度惩",
"输入。特别是，最终的节点表示计算如下：",
"罚（Wuetal.2016）来改进纯最大概率波束搜",
"索",
"hglobal = GE(hg, (hl : u E V))",
"训练模型以优化每个黄金标准输出文本的负",
"hu = LE(hglobal,[hglobal : u E N(u)). (12)",
"对数似然。我们采用标签平滑正则化来防止模",
"型在训练过程中过于自信地预测令牌，并且泛",
"化不佳",
"Layer-wise Parallel Graph Encoding.",
"为了允",
"许两种类型的图形上下文信息之间的细粒度交",
"4",
"",
"Data and Preprocessing",
"互，我们还以分层（LW）方式组合了编码器。",
"如图E2c所示，对于每个图层，我们在并行结",
"我们证明了我们的模型在两个数据集上的有",
"构（PGE-LW）中同时采用全局和局部编码",
"效性：AGENDA (Koncel-Kedziorski et al., 2019)",
"器。更准确地说，每个编码器层的计算如下：",
"和 WebNLG(Gardent et al.,2017)。表 1显示",
"hglobal = GE(h-1, [hi-1 : u E V))",
"了两个数据集的统计信息",
"hlocal = LE(h-1, [ht-I : E N(u)",
"AGENDA.在此数据集中，KG与从12个顶",
"h', - [hglobal | hlocal ]",
"级AI会议记录中提取的科学摘要配对。每个",
"(13)",
"实例由论文标题、KG和论文摘要组成。实体",
"其中GE,和LEi分别指全局和局部图编码器的",
"对应于通常是多词表达的科学术语（合并了共",
"1层。",
"同引用实体）。我们将标题中的每个令牌视为",
"个节点，创建一个以标题和KG令牌为节点",
"Layer-wise Cascaded Graph Encoding.",
"我们",
"的唯一图。如表1所示，平均输出长度相当",
"还建议分层级联图形编码器（CGE-LW",
"图F",
"大，因为目标输出是多句子摘要。"
] |
{
"arxiv_id": "2001.11003",
"language": "zh",
"page_idx": 4
}
|
||
1708.02982-zh-3
|
1708.02982-zh-3.jpg
|
<FigureHere>
$(\mathbf{b})$ A Channel
(c)B Channel
(a) Original Image
Figure 3: The original image is converted to the LAB color space and the A and B channels are shown. In the A channel, the different shades of red look the same and make a bright square region; and the different shades of green look the same and make a dark ring. This makes it easy to detect the red to green border in the A channel. In the B channel, the different shades of red and green become differentiable, making it possible to read the binary code for the tag.
randomized forests to learn and detect planar objects. In practice, these algorithms do not achieve detection accuracies on par with detection algorithms specifically designed for marker detection. With the increased popularity of deep learning based detection approaches, it is possible that better detection can be achieved; however, ChromaTag, which does not require a GPU, is faster than current deep learning approaches which do require a GPU [27,29].
<FigureHere>
# 3. ChromaTag Design
Figure 4 provides an example of a ChromaTag. The inner red square and green ring are used for rejecting false positives (Section 3.1) and the outer black and white rings are used for precise localization of the tag (Section 3.2).
## 3.1. Efficient False Positive Rejection
Figure 4: An example ChromaTag. The red and green regions are easily detectable and limit false positives. The two shades of red and two shades of green are used to embed the tag code. The outer black-white border provides full spatial resolution for accurate localization of the tag corners.
Opponent color spaces offer large gradients between opposing colors in each channel. Figure 5 depicts the LAB color space, where red and green are opposing colors in the A channel and blue and yellow are opposing colors in the B channel. Figure 5 also depicts how two (or more) shades of green or red can have the same value in the A channel, but different values in the B channel. ChromaTag's color configuration is designed based on these properties.
The red center surrounded by the green ring has a large gradient in the A channel (Figure 3b), which rarely occurs in natural scenes (empirically validated in Figure 10). Thus, we can quickly detect tags with high precision and recall by scanning the image in steps of $N$ pixels and thresholding the A channel difference of neighboring steps.
ChromaTag encodes the binary code in the B channel (Figure 3c), which has little effect on the A channel intensity (Figure 3b). Since every tag includes low and high B values (encoding 0 and 1) in both the green and red area, the
encode marker signature. These tags (AprilTag, RuneTag, and CCTag) represent the current state-of-the-art for fiducial markers. They have demonstrated accurate detection and are commonly used. ChromaTag uses a new detection approach (taking advantage of color on the markers) to achieve processing times significantly faster than other tags while still maintaining similar detection accuracies.
Several works perform planar object detection using machine learning approaches. Early work by Claus et. al [11,12] employs a cascaded bayes and nearest neighbor classification scheme for marker detection. More recent work of Ozuysal et al.[25] and Lepetit et al.[22] uses
|
[
[
0.441162109375,
0.25244140625,
0.53125,
0.250732421875,
0.53125,
0.26513671875,
0.441162109375,
0.267578125
],
[
0.705078125,
0.25244140625,
0.79345703125,
0.250732421875,
0.79345703125,
0.26513671875,
0.705078125,
0.267578125
],
[
0.08087158203125,
0.281005859375,
0.89208984375,
0.281005859375,
0.89208984375,
0.294189453125,
0.08087158203125,
0.294189453125
],
[
0.1666259765625,
0.25390625,
0.279296875,
0.25390625,
0.279296875,
0.268310546875,
0.1666259765625,
0.268310546875
],
[
0.080078125,
0.29541015625,
0.890625,
0.29541015625,
0.890625,
0.309326171875,
0.080078125,
0.309326171875
],
[
0.08087158203125,
0.313232421875,
0.890625,
0.313232421875,
0.890625,
0.323974609375,
0.08087158203125,
0.323974609375
],
[
0.07843017578125,
0.324462890625,
0.689453125,
0.326904296875,
0.689453125,
0.341552734375,
0.07843017578125,
0.339111328125
],
[
0.50341796875,
0.36865234375,
0.89208984375,
0.36865234375,
0.89208984375,
0.3818359375,
0.50341796875,
0.3818359375
],
[
0.50244140625,
0.3837890625,
0.8896484375,
0.38330078125,
0.8896484375,
0.397705078125,
0.50244140625,
0.3984375
],
[
0.50341796875,
0.3984375,
0.89208984375,
0.3984375,
0.89208984375,
0.412353515625,
0.50341796875,
0.412353515625
],
[
0.50341796875,
0.4140625,
0.890625,
0.4140625,
0.890625,
0.427978515625,
0.50341796875,
0.427978515625
],
[
0.50341796875,
0.429931640625,
0.8896484375,
0.429931640625,
0.8896484375,
0.443115234375,
0.50341796875,
0.443115234375
],
[
0.50341796875,
0.4443359375,
0.89208984375,
0.4443359375,
0.89208984375,
0.458251953125,
0.50341796875,
0.458251953125
],
[
0.50244140625,
0.458251953125,
0.89208984375,
0.458984375,
0.89208984375,
0.473388671875,
0.50244140625,
0.472900390625
],
[
0.50341796875,
0.474853515625,
0.79833984375,
0.474853515625,
0.79833984375,
0.488037109375,
0.50341796875,
0.488037109375
],
[
0.50244140625,
0.5048828125,
0.6845703125,
0.5068359375,
0.6845703125,
0.521484375,
0.50244140625,
0.51953125
],
[
0.521484375,
0.5322265625,
0.890625,
0.5322265625,
0.890625,
0.54541015625,
0.521484375,
0.54541015625
],
[
0.50341796875,
0.54736328125,
0.89208984375,
0.54736328125,
0.89208984375,
0.56103515625,
0.50341796875,
0.56103515625
],
[
0.50341796875,
0.5625,
0.89208984375,
0.5625,
0.89208984375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.50341796875,
0.57763671875,
0.869140625,
0.57763671875,
0.869140625,
0.59130859375,
0.50341796875,
0.59130859375
],
[
0.50341796875,
0.60498046875,
0.78515625,
0.60498046875,
0.78515625,
0.61865234375,
0.50341796875,
0.61865234375
],
[
0.08087158203125,
0.6162109375,
0.465576171875,
0.6162109375,
0.465576171875,
0.6298828125,
0.08087158203125,
0.6298828125
],
[
0.08087158203125,
0.6318359375,
0.4697265625,
0.6318359375,
0.4697265625,
0.64599609375,
0.08087158203125,
0.64599609375
],
[
0.521484375,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.521484375,
0.642578125
],
[
0.080078125,
0.64794921875,
0.468017578125,
0.64794921875,
0.468017578125,
0.6611328125,
0.080078125,
0.6611328125
],
[
0.5009765625,
0.64306640625,
0.890625,
0.642578125,
0.890625,
0.6572265625,
0.5009765625,
0.65771484375
],
[
0.080078125,
0.662109375,
0.4697265625,
0.662109375,
0.4697265625,
0.6767578125,
0.080078125,
0.6767578125
],
[
0.50341796875,
0.6591796875,
0.89208984375,
0.6591796875,
0.89208984375,
0.67236328125,
0.50341796875,
0.67236328125
],
[
0.08087158203125,
0.6787109375,
0.428955078125,
0.6787109375,
0.428955078125,
0.689453125,
0.08087158203125,
0.689453125
],
[
0.50341796875,
0.673828125,
0.89208984375,
0.673828125,
0.89208984375,
0.6875,
0.50341796875,
0.6875
],
[
0.5009765625,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.7021484375,
0.5009765625,
0.7021484375
],
[
0.50341796875,
0.7041015625,
0.890625,
0.7041015625,
0.890625,
0.71728515625,
0.50341796875,
0.71728515625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.080078125,
0.734375
],
[
0.50341796875,
0.71826171875,
0.8935546875,
0.71826171875,
0.8935546875,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.5009765625,
0.73291015625,
0.8388671875,
0.73291015625,
0.8388671875,
0.7470703125,
0.5009765625,
0.7470703125
],
[
0.08087158203125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76220703125,
0.08087158203125,
0.76220703125
],
[
0.521484375,
0.74853515625,
0.890625,
0.75,
0.890625,
0.7646484375,
0.521484375,
0.76318359375
],
[
0.08087158203125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.080078125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.77978515625,
0.890625,
0.77978515625,
0.890625,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.79541015625,
0.4697265625,
0.79541015625,
0.4697265625,
0.8095703125,
0.080078125,
0.8095703125
],
[
0.50341796875,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.8095703125,
0.50341796875,
0.8095703125
],
[
0.080078125,
0.81103515625,
0.44677734375,
0.81103515625,
0.44677734375,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50244140625,
0.8095703125,
0.89208984375,
0.81005859375,
0.89208984375,
0.82470703125,
0.50244140625,
0.82373046875
],
[
0.09722900390625,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.09722900390625,
0.84033203125
],
[
0.50341796875,
0.82568359375,
0.8037109375,
0.82568359375,
0.8037109375,
0.8388671875,
0.50341796875,
0.8388671875
],
[
0.07843017578125,
0.84228515625,
0.46728515625,
0.84033203125,
0.46728515625,
0.85400390625,
0.07843017578125,
0.85595703125
],
[
0.521484375,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.521484375,
0.85498046875
],
[
0.08087158203125,
0.85791015625,
0.465576171875,
0.85791015625,
0.465576171875,
0.86865234375,
0.08087158203125,
0.86865234375
],
[
0.5048828125,
0.85791015625,
0.8896484375,
0.85791015625,
0.8896484375,
0.86865234375,
0.5048828125,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"(b) A Channel",
"(c) B Channel",
"Figure 3: The original image is converted to the LAB color space and the A and B channels are shown. In the A channel, the",
"(a) Original Image",
"different shades of red look the same and make a bright square region; and the different shades of green look the same and",
"make a dark ring. This makes it easy to detect the red to green border in the A channel. In the B channel, the different shades",
"of red and green become differentiable, making it possible to read the binary code for the tag.",
"randomized forests to learn and detect planar objects.In",
"practice, these algorithms do not achieve detection accura-",
"cies on par with detection algorithms specifically designed",
"for marker detection. With the increased popularity of deep",
"learning based detection approaches, it is possible that bet-",
"ter detection can be achieved; however, ChromaTag, which",
"does not require a GPU, is faster than current deep learning",
"approaches which do require a GPU [27,29].",
"3.ChromaTagDesign",
"Figure 4 provides an example of a ChromaTag. The in-",
"ner red square and green ring are used for rejecting false",
"positives (Section 3.1) and the outer black and white rings",
"are used for precise localization of the tag (Section 3.2).",
"3.1. Efficient False Positive Rejection",
"Figure 4: An example ChromaTag. The red and green re-",
"gions are easily detectable and limit false positives.The two",
"Opponent color spaces offer large gradients between op-",
"shades of red and two shades of green are used to embed the",
"posing colors in each channel. Figure 5 depicts the LAB",
"tag code. The outer black-white border provides full spatial",
"color space, where red and green are opposing colors in the",
"resolutionforaccuratelocalizationof thetag corners.",
"A channel and blue and yellow are opposing colors in the",
"B channel. Figure 5 also depicts how two (or more) shades",
"of green or red can have the same value in the A channel,",
"encode marker signature. These tags (AprilTag, RuneTag,",
"but different values in the B channel. ChromaTag's color",
"and CCTag) represent the current state-of-the-art for fidu-",
"configuration is designed based on these properties.",
"cialmarkers.They havedemonstratedaccuratedetection",
"The red center surrounded by the green ring has a large",
"andarecommonlyused.ChromaTag usesa new detec-",
"gradient in the A channel (Figure 3b), which rarely occurs",
"tion approach (taking advantage of color on the markers)",
"in natural scenes (empirically validated in Figure 10). Thus,",
"to achieve processing times significantly faster than other",
"we can quickly detect tags with high precision and recall by",
"tags while still maintaining similar detection accuracies.",
"scanning the image in steps of N pixels and thresholding",
"Several works perform planar object detection using",
"the A channel difference of neighboring steps.",
"machine learning approaches.Early work by Claus et.",
"ChromaTag encodes the binary code in the B channel",
"al[1l,12lemploys a cascadedbayes andnearestneigh-",
"(Figure3c),whichhaslittleeffectontheAchannel inten-",
"bor classification scheme for marker detection. More re-",
"sity (Figure 3b). Since every tag includes low and high B",
"cent work of Ozuysal et al. [25] and Lepetit et al. [22] uses",
"values (encoding O and 1) in both the green and red area, the"
] |
{
"arxiv_id": "1708.02982",
"language": "zh",
"page_idx": 3
}
|
||
1708.02982-zh-4
|
1708.02982-zh-4.jpg
|
<FigureHere>
Figure 5: The LAB colorspace. ChromaTag uses red to green borders because they have a large image gradient in the A space. ChromaTag uses two shades of red and two shades of green as representations of 0 and 1 to embed a code in the tag. These different shades of red and green are only differentiable in the B channel, so detection is done in the A channel and decoding is done in the B channel. RGB values for the different red and green colors are shown.
Algorithm 1:ChromaTag Detection
${\mathrm{Im=Input~RGB~image}}$
$\mathbf{N}=$Stepsize in pixels (4 in our experiments)
ADiffThresh=Threshold for initial detection (25 in
our experiments)
Dets=Struct to hold detections
TmpDet=Holds detections as they are built
$\mathbf{for}$ $i= 0;$ $i< Im. Rows( ) ;$ $i+ = N$ $\mathbf{do}$
OldA= ConvertToLAB( Im( i, j) ) $\mathbf{for}$ $j= 0;$ $j< Im. Cols( ) ;$ $j+ = N$ $\mathbf{do}$
if InPreviousDetectionArea(i,j,Dets) then
j= MoveJToEndOfDetection( i, j) OldA= ConvertToLAB( Im( i, j) ) continue
[L,A,B]=ConvertToLAB( Im(i,j) )
$\textbf{if }A- OldA> ADiffThresh\textbf{ then}$
if InitialScan(Im,i,j,A,TmpDet) then
if BuildPolygon(Im,TmpDet) then
if PolyToQuad(Im,TmpDet) then
$\textbf{if Decode}( Im, TmpDet) \textbf{ then}$
Dets.Add(TmpDet)
$\overline{\mathrm{OldA}}=$A
thresholds are adapted per tag to account for variations due to lighting or printing. The tag detection is verified based on the code value, enabling high precision.
We found LAB to be more robust to color printer and lighting variations than other color spaces such as YUV. We use hashing to speed computation of LAB values.
## $3. 2. \textbf{Precise Localization}$
Precise localization is required to decode the tag and recover camera pose. Chrominance has lower effective resolution than luminance due to the Bayer pattern filters used in common cameras. ChromaTag is designed with outer black and white concentric rectangles to provide high contrast and high resolution borders for precise corner localization.
## 4. ChromaTag Detection
Algorithm 1 outlines ChromaTag detection. Substantial effort was required to make each step of the detection computationally efficient (700+ FPS!) and robust to variable lighting and perspective.
FindADiff: The first step is to find potential tag locations. Assuming that tag cells are at least $N/2$ pixels wide, we search over pixels on a grid of every $N^{t\hat{h}}$ row and column. If a sampled pixel at location $(i,j)$ is in an already-detected area (InPreviousDetectionArea(i,j,Dets)), $j$ is moved to the next grid location outside the tag detection $(MoveJToEndOfDetection(i,j)).$ Otherwise, the pixel is converted to the LAB space $( ConvertToLAB( \textit{Im( i, j) }) ) $, and the A channel intensity is compared to the previous grid location's A channel intensity $( A$ $- OldA$ $> ADiffThresh) .$ If
the difference is greater than ADiffThresh, detection commences and the A value is set as ReferenceRed; otherwise, the loop continues to the next sampled pixel. In our experiments, N=4 and $ADiffThresh=25.$
InitialScan( Im,i,j,A,TmpDet ): The next step is to reject red locations that are not tags so that no further processing is done at that location. From the grid location $(i,j)$,we scan left, right, up and down as shown in Figure 6a. Each scan continues until $M$ successive pixels have A channel differences greater than BorderThresh when compared to ReferenceRed. If successive pixels are found, scanning has entered the green region and the border $(u,\nu)$ location and pixel value $(ReferenceGreen)$ are remembered. If successive pixels are not found, we return false and the grid location is abandoned. In our experiments, $M=3$ and $BorderThresh=5.$
If all four scans (up, down, left, right) find the green region, we average the $(u,\nu)$ locations on the red-green border to estimate the center of the tag and repeat the scan. If any scan fails, we return false and the grid location is abandoned. When the center location converges, we continue the scans through the green region to find the green-black and black-white borders. These scans compare against a set ReferenceGreen or ReferenceBlack respectively and use the
|
[
[
0.08660888671875,
0.08709716796875,
0.1871337890625,
0.0889892578125,
0.1871337890625,
0.106689453125,
0.08660888671875,
0.10479736328125
],
[
0.2900390625,
0.10101318359375,
0.3056640625,
0.10101318359375,
0.3056640625,
0.10797119140625,
0.2900390625,
0.10797119140625
],
[
0.5107421875,
0.09661865234375,
0.75341796875,
0.09661865234375,
0.75341796875,
0.10986328125,
0.5107421875,
0.10986328125
],
[
0.08905029296875,
0.10919189453125,
0.188720703125,
0.10919189453125,
0.188720703125,
0.12371826171875,
0.08905029296875,
0.12371826171875
],
[
0.521484375,
0.11553955078125,
0.67724609375,
0.11553955078125,
0.67724609375,
0.1287841796875,
0.521484375,
0.1287841796875
],
[
0.101318359375,
0.1307373046875,
0.134033203125,
0.1307373046875,
0.134033203125,
0.1483154296875,
0.101318359375,
0.1483154296875
],
[
0.521484375,
0.1287841796875,
0.81787109375,
0.1300048828125,
0.81787109375,
0.14453125,
0.521484375,
0.143310546875
],
[
0.419189453125,
0.1383056640625,
0.44287109375,
0.1383056640625,
0.44287109375,
0.1514892578125,
0.419189453125,
0.1514892578125
],
[
0.52294921875,
0.143310546875,
0.865234375,
0.143310546875,
0.865234375,
0.1572265625,
0.52294921875,
0.1572265625
],
[
0.529296875,
0.160400390625,
0.64453125,
0.160400390625,
0.64453125,
0.173583984375,
0.529296875,
0.173583984375
],
[
0.52392578125,
0.1767578125,
0.73046875,
0.1767578125,
0.73046875,
0.1875,
0.52392578125,
0.1875
],
[
0.52294921875,
0.1900634765625,
0.81396484375,
0.1900634765625,
0.81396484375,
0.203857421875,
0.52294921875,
0.203857421875
],
[
0.09722900390625,
0.2108154296875,
0.1356201171875,
0.2108154296875,
0.1356201171875,
0.231689453125,
0.09722900390625,
0.231689453125
],
[
0.426513671875,
0.2139892578125,
0.441162109375,
0.216552734375,
0.439453125,
0.2266845703125,
0.423095703125,
0.2235107421875
],
[
0.52294921875,
0.2203369140625,
0.744140625,
0.2203369140625,
0.744140625,
0.2342529296875,
0.52294921875,
0.2342529296875
],
[
0.36279296875,
0.23486328125,
0.459228515625,
0.23486328125,
0.459228515625,
0.25,
0.36279296875,
0.25
],
[
0.54736328125,
0.236083984375,
0.7646484375,
0.236083984375,
0.7646484375,
0.25,
0.54736328125,
0.25
],
[
0.36279296875,
0.25244140625,
0.4599609375,
0.25244140625,
0.4599609375,
0.267578125,
0.36279296875,
0.267578125
],
[
0.54833984375,
0.250732421875,
0.765625,
0.250732421875,
0.765625,
0.264404296875,
0.54833984375,
0.264404296875
],
[
0.57177734375,
0.266357421875,
0.84619140625,
0.266357421875,
0.84619140625,
0.27978515625,
0.57177734375,
0.27978515625
],
[
0.080078125,
0.281005859375,
0.4697265625,
0.2822265625,
0.4697265625,
0.296630859375,
0.080078125,
0.29541015625
],
[
0.59619140625,
0.2822265625,
0.81201171875,
0.2822265625,
0.81201171875,
0.29541015625,
0.59619140625,
0.29541015625
],
[
0.08087158203125,
0.2998046875,
0.468017578125,
0.2998046875,
0.468017578125,
0.31005859375,
0.08087158203125,
0.31005859375
],
[
0.59619140625,
0.294189453125,
0.81640625,
0.296630859375,
0.81640625,
0.311279296875,
0.59619140625,
0.309326171875
],
[
0.080078125,
0.313232421875,
0.4697265625,
0.313232421875,
0.4697265625,
0.326904296875,
0.080078125,
0.326904296875
],
[
0.5947265625,
0.313232421875,
0.65673828125,
0.313232421875,
0.65673828125,
0.324462890625,
0.5947265625,
0.324462890625
],
[
0.07843017578125,
0.32763671875,
0.4697265625,
0.326904296875,
0.4697265625,
0.341552734375,
0.07843017578125,
0.34228515625
],
[
0.080078125,
0.34228515625,
0.46728515625,
0.34228515625,
0.46728515625,
0.35595703125,
0.080078125,
0.35595703125
],
[
0.5703125,
0.34228515625,
0.80810546875,
0.3447265625,
0.80810546875,
0.3623046875,
0.5703125,
0.35986328125
],
[
0.080078125,
0.35791015625,
0.468017578125,
0.35791015625,
0.468017578125,
0.371826171875,
0.080078125,
0.371826171875
],
[
0.57177734375,
0.361083984375,
0.78173828125,
0.361083984375,
0.78173828125,
0.375,
0.57177734375,
0.375
],
[
0.08087158203125,
0.375,
0.46728515625,
0.375,
0.46728515625,
0.385009765625,
0.08087158203125,
0.385009765625
],
[
0.59716796875,
0.376953125,
0.83642578125,
0.376953125,
0.83642578125,
0.39013671875,
0.59716796875,
0.39013671875
],
[
0.080078125,
0.389404296875,
0.439453125,
0.389404296875,
0.439453125,
0.40283203125,
0.080078125,
0.40283203125
],
[
0.62109375,
0.39013671875,
0.84716796875,
0.391357421875,
0.84716796875,
0.406005859375,
0.62109375,
0.40478515625
],
[
0.646484375,
0.4072265625,
0.86376953125,
0.4072265625,
0.86376953125,
0.42041015625,
0.646484375,
0.42041015625
],
[
0.6708984375,
0.42236328125,
0.861328125,
0.42236328125,
0.861328125,
0.436279296875,
0.6708984375,
0.436279296875
],
[
0.080078125,
0.431884765625,
0.468017578125,
0.431884765625,
0.468017578125,
0.44580078125,
0.080078125,
0.44580078125
],
[
0.68896484375,
0.436279296875,
0.82861328125,
0.4375,
0.82861328125,
0.451904296875,
0.68896484375,
0.45068359375
],
[
0.080078125,
0.4462890625,
0.4697265625,
0.4462890625,
0.4697265625,
0.4609375,
0.080078125,
0.4609375
],
[
0.080078125,
0.462158203125,
0.361083984375,
0.462158203125,
0.361083984375,
0.47607421875,
0.080078125,
0.47607421875
],
[
0.098876953125,
0.477783203125,
0.4697265625,
0.477783203125,
0.4697265625,
0.4912109375,
0.098876953125,
0.4912109375
],
[
0.56884765625,
0.474853515625,
0.6396484375,
0.474853515625,
0.6396484375,
0.486083984375,
0.56884765625,
0.486083984375
],
[
0.080078125,
0.49365234375,
0.4697265625,
0.49365234375,
0.4697265625,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.07843017578125,
0.50830078125,
0.406005859375,
0.50830078125,
0.406005859375,
0.521484375,
0.07843017578125,
0.521484375
],
[
0.080078125,
0.5341796875,
0.26708984375,
0.5341796875,
0.26708984375,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.50244140625,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54736328125,
0.50244140625,
0.54736328125
],
[
0.5009765625,
0.5498046875,
0.8896484375,
0.548828125,
0.8896484375,
0.5625,
0.5009765625,
0.56298828125
],
[
0.098876953125,
0.556640625,
0.468017578125,
0.556640625,
0.468017578125,
0.57080078125,
0.098876953125,
0.57080078125
],
[
0.50341796875,
0.56640625,
0.8896484375,
0.56640625,
0.8896484375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.08087158203125,
0.57470703125,
0.46728515625,
0.57470703125,
0.46728515625,
0.58544921875,
0.08087158203125,
0.58544921875
],
[
0.5009765625,
0.580078125,
0.7255859375,
0.580078125,
0.7255859375,
0.59326171875,
0.5009765625,
0.59326171875
],
[
0.080078125,
0.58837890625,
0.468017578125,
0.58837890625,
0.468017578125,
0.6015625,
0.080078125,
0.6015625
],
[
0.521484375,
0.59765625,
0.89208984375,
0.59765625,
0.89208984375,
0.611328125,
0.521484375,
0.611328125
],
[
0.080078125,
0.60400390625,
0.4697265625,
0.60400390625,
0.4697265625,
0.6171875,
0.080078125,
0.6171875
],
[
0.080078125,
0.61865234375,
0.468017578125,
0.61865234375,
0.468017578125,
0.6318359375,
0.080078125,
0.6318359375
],
[
0.50244140625,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.50244140625,
0.626953125
],
[
0.50341796875,
0.62744140625,
0.89208984375,
0.62744140625,
0.89208984375,
0.6416015625,
0.50341796875,
0.6416015625
],
[
0.07843017578125,
0.6337890625,
0.435546875,
0.6337890625,
0.435546875,
0.64794921875,
0.07843017578125,
0.64794921875
],
[
0.50244140625,
0.642578125,
0.89208984375,
0.642578125,
0.89208984375,
0.65576171875,
0.50244140625,
0.65576171875
],
[
0.5009765625,
0.65771484375,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.5009765625,
0.671875
],
[
0.08087158203125,
0.6640625,
0.285888671875,
0.6640625,
0.285888671875,
0.6787109375,
0.08087158203125,
0.6787109375
],
[
0.50244140625,
0.67236328125,
0.8935546875,
0.67236328125,
0.8935546875,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.098876953125,
0.689453125,
0.468017578125,
0.689453125,
0.468017578125,
0.70263671875,
0.098876953125,
0.70263671875
],
[
0.50244140625,
0.68798828125,
0.89208984375,
0.68798828125,
0.89208984375,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.50244140625,
0.7041015625,
0.89208984375,
0.7041015625,
0.89208984375,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.07843017578125,
0.71826171875,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.07843017578125,
0.73291015625
],
[
0.50244140625,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50244140625,
0.73193359375
],
[
0.080078125,
0.73486328125,
0.273681640625,
0.73486328125,
0.273681640625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50244140625,
0.73291015625,
0.89208984375,
0.73291015625,
0.89208984375,
0.7470703125,
0.50244140625,
0.7470703125
],
[
0.1004638671875,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.1004638671875,
0.76513671875
],
[
0.50244140625,
0.74853515625,
0.89208984375,
0.74853515625,
0.89208984375,
0.76220703125,
0.50244140625,
0.76220703125
],
[
0.080078125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.50244140625,
0.76220703125,
0.64794921875,
0.76318359375,
0.64794921875,
0.77783203125,
0.50244140625,
0.7763671875
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.521484375,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.521484375,
0.79443359375
],
[
0.07843017578125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81103515625,
0.07843017578125,
0.81103515625
],
[
0.50341796875,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50341796875,
0.81005859375
],
[
0.080078125,
0.8125,
0.468017578125,
0.8125,
0.468017578125,
0.82568359375,
0.080078125,
0.82568359375
],
[
0.50244140625,
0.8125,
0.89208984375,
0.8125,
0.89208984375,
0.82568359375,
0.50244140625,
0.82568359375
],
[
0.08087158203125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.080078125,
0.84228515625,
0.468017578125,
0.84228515625,
0.468017578125,
0.85595703125,
0.080078125,
0.85595703125
],
[
0.50244140625,
0.84228515625,
0.89208984375,
0.84228515625,
0.89208984375,
0.85595703125,
0.50244140625,
0.85595703125
],
[
0.080078125,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.8935546875,
0.85693359375,
0.8935546875,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.07843017578125,
0.88525390625
],
[
0.5009765625,
0.87060546875,
0.8935546875,
0.8720703125,
0.8935546875,
0.88623046875,
0.5009765625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"[191.255.0]",
"-",
"Algorithm 1: ChromaTag Detection",
"[166,255,77]",
"Im = Input RGB image",
"-A",
"N = Stepsize in pixels (4 in our experiments)",
"B",
"ADiffThresh = Threshold for initial detection (25 in",
"our experiments)",
"Dets=Structto holddetections",
"TmpDet = Holds detections as they are built",
"-B",
"A",
"for i=0; i < Im.Rows(); i+=N do",
"[255,41:149]",
"OldA = ConvertToLAB(Im(i.j))",
"[254.40.255",
"for j=0; j< Im.Cols(); j+=N do",
"if InPreviousDetectionArea(i.j,Dets) then",
"Figure 5:The LAB colorspace. ChromaTag uses red to",
"j =MoveJToEndOfDetection(i,j)",
"greenbordersbecausetheyhavealargeimagegradientin",
"OldA = ConvertToLAB( Im(i,j) )",
"the A space. ChromaTag uses two shades of red and two",
"continue",
"shades of green as representations of 0 and 1 to embed a",
"code in the tag. These different shades of red and green are",
"[L,A,B] = ConvertToLAB( Im(i.j) )",
"only differentiable in the B channel, so detection is done in",
"if A - OldA > ADiffThresh then",
"theAchannel anddecodingisdoneintheBchannel.RGB",
"if InitialScan(Im,i,j,A, TmpDet) then",
"values for the different red and green colors are shown.",
"if BuildPolygon(Im, TmpDet) then",
"if PolyToQuad(Im,TmpDet) then",
"if Decode(Im,TmpDet) then",
"thresholds are adapted per tag to account for variations due",
"Dets.Add(TmpDet)",
"to lighting or printing. The tag detection is verified based",
"on the code value, enabling high precision.",
"We found LAB to be more robust to color printer and",
"OldA=A",
"lighting variations than other color spaces such as YUV. We",
"use hashing to speed computation of LAB values.",
"3.2. Precise Localization",
"the difference is greater than ADiffThresh, detection com-",
"mences and the A value is set as ReferenceRed; otherwise,",
"Precise localization is required to decode the tag and re-",
"theloop continuesto thenextsampledpixel.In ourexperi-",
"covercamera pose.Chrominancehaslowereffectivereso-",
"ments, N=4 and ADiffThresh=25.",
"lution than luminance due to the Bayer pattern filters used in",
"InitialScan( Im,i.j,A,TmpDet ): The next step is to re-",
"common cameras. ChromaTag is designed with outer black",
"and white concentric rectangles to provide high contrast and",
"ject red locations that are not tags so that no further process-",
"ing is done at that location. From the grid location (i,j), we",
"high resolution borders for precise corner localization.",
"scan left, right, up and down as shown in Figure 6a. Each",
"scan continues until M successive pixels have A channel",
"4. ChromaTag Detection",
"differences greater than BorderThresh when compared",
"Algorithm 1 outlines ChromaTag detection. Substan-",
"to ReferenceRed. If successive pixels are found, scanning",
"tial effort was required to make each step of the detection",
"has entered the green region and the border (u,v) location",
"computationally efficient (700+ FPS!) and robust to vari-",
"and pixel value (ReferenceGreen) are remembered. If suc-",
"able lighting and perspective.",
"cessive pixels are not found, we return false and the grid",
"FindADiff: The first step is to find potential tag lo-",
"location is abandoned. In our experiments, M = 3 and",
"cations. Assuming that tag cells are at least N/2 pixels",
"BorderThresh = 5.",
"wide, we search over pixels on a grid of every Nth row",
"If all four scans (up, down, left, right) find the green re-",
"and column. If a sampled pixel at location (i,j) is in an",
"gion, we average the (u,v) locations on the red-green bor-",
"already-detected area (InPreviousDetectionArea(i.j,Dets)),",
"der to estimate the center of the tag and repeat the scan. If",
"j is moved to the next grid location outside the tag detec-",
"any scan fails, we return false and the grid location is aban-",
"tion (MoveJToEndOfDetection(i,j). Otherwise, the pixel is",
"doned. When the center location converges, we continue",
"converted to the LAB space (ConvertToLAB( Im(i,j) ), and",
"the scans through the green region to find the green-black",
"the A channel intensity is compared to the previous grid lo-",
"and black-white borders. These scans compare against a set",
"cation's A channel intensity (A - OldA > ADiffThresh). If",
"ReferenceGreen or ReferenceBlack respectively and use the"
] |
{
"arxiv_id": "1708.02982",
"language": "zh",
"page_idx": 4
}
|
||
2310.10971-zh-0
|
2310.10971-zh-0.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2024}}$
# 上下文感知元学习
# $\textbf{Christopher Fifty}^1,\textbf{Dennis Duan}^{1,2},\textbf{Ronald G. Junkins}^1$, $\mathbf{Ehsan~Amid}^{3},\mathbf{Jure~Leskovec}^{1},\mathbf{Christopher~Ré}^{1},\mathbf{Sebastian~Thrun}^{1}$ $^{1}$Stanford University, $^2$Google, $^3$Google DeepMind fifty@cs.stanford.com
ABSTRACT
Large Language Models like ChatGPT demonstrate a remarkable capacity to learn new concepts during inference without any fine-tuning. However, visual models trained to detect new objects during inference have been unable to replicate this ability, and instead either perform poorly or require meta-training and/or finetuning on similar objects. In this work, we propose a meta-learning algorithm that emulates Large Language Models by learning new visual concepts during inference without fine-tuning. Our approach leverages a frozen pre-trained feature extractor, and analogous to in-context learning, recasts visual meta-learning as sequence modeling over datapoints with known labels and a test datapoint with an unknown label. On 8 out of 11 few-shot image classification benchmarks, our approach—without meta-training or fine-tuning—exceeds or matches the state-of- $\hat{\mathrm{the-art~algorithm,~P>M>F,~which~is~meta-trained~on~these~benchmarks.~Our~code}}$ is available at https://github.com/cfifty/CAML .
1 INTRODUCTION
元学习是指从少量演示中学习新概念的能力 (Lake et al.,2015)。在机器智能取得显著进步的十年中，它仍然是人类性能继续超越机器的领域 (Brown et al.,2020)。为了匹配人类的能力，为了开发能够像人类一样学习和思考的机器，我们必须开发能够从几个例子中学习新概念的机器智能。
深度学习的许多应用都将学习算法应用于大量的训练数据集；但是，从极少数的训练示例中学习会带来挑战$\overline{\mathbb{E}}($Lake et al., 2017; Garnelo et al., 2018)。这一挑战导致了两种主要的评估设置：域内和跨域。域内设置评估元学习者在特定域内接受类似任务培训后快速适应新任务的能力。为此设置设计的模型通常非常快，但对目标域$\boxed{\mathrm{E}}($Chen et al.,2019)以外的任务的泛化能力很差。同时，跨域设置评估元学习者适应以前看不见的域中的任务的能力。为此设置设计的方法具有很高的适应性，但在推理过程中速度很慢，因为它们需要在支持集 (Guo et al., 2020; Oh et al., 2022; Hu et al., 2022) 上进行微调。关键的是，两种情况下的元学习者都不同于人类快速推广到新任务的能力
最近，大型语言模型 (LLM) 在自然语言中解决了同时快速和通用元学习的问题。像 ChatGPT 这样的 LLM 可以通过一种称为上下文学习的能力快速推广到新任务 (Brown et al.,2020) 。然而，它仍然是一个开放的问题，在计算机视觉。即使是最好的视觉元学习算法也无法部署到类似 ChatGPT 的系统中，因为此类系统所需的模型可以(1)推广到训练时未知的广泛任务集，并且(2)实时执行，而无需花费时间对模型进行微调。LLM 表现出了两种能力的非凡能力；但是，当前的视觉元学习者可能只能满足一个要求或另一个要求$\overline{\mathrm{E}}($Hu et al.,2022) 为了衡量实现快速和通用视觉元学习者这一目标的进展，我们开发了一种评估范式，我们称之为通用元学习。通用元学习测量模型快速学习新图像类的能力。它评估跨跨许多不同图像分类任务的各种元学习基准的模型，而无需对任何基准的训练集进行元训练或在推理期间对支持集进行微调。我们专注于少量图像分类的应用-而不是像绘画或分割这样的密集预测任务-因为已经为这些应用探索了通用设置 (Bar et al.,2022;Zhang et al.,2023; Wang et al., 2023; Kim et al., 2023; Butoi et al., 2023)
除了通用设置中的基准方法之外，我们还介绍了一种实现强大通用性能的元学习器。从LLM 的上下文学习中汲取灵感，我们将$n$ -way- $k$ -shot 图像分类重新表述为支持集和未知查询图像上的非因果序列建模。具体来说，给定每个类的$k$示例的$n$方式分类，我们在
1
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.178955078125,
0.10418701171875,
0.39794921875,
0.10418701171875,
0.39794921875,
0.125,
0.178955078125,
0.125
],
[
0.188720703125,
0.1514892578125,
0.5849609375,
0.15087890625,
0.5849609375,
0.1654052734375,
0.188720703125,
0.1666259765625
],
[
0.1854248046875,
0.1666259765625,
0.646484375,
0.1666259765625,
0.646484375,
0.18115234375,
0.1854248046875,
0.18115234375
],
[
0.1871337890625,
0.18115234375,
0.52294921875,
0.18115234375,
0.52294921875,
0.1944580078125,
0.1871337890625,
0.1944580078125
],
[
0.1871337890625,
0.1976318359375,
0.390625,
0.1976318359375,
0.390625,
0.2083740234375,
0.1871337890625,
0.2083740234375
],
[
0.453369140625,
0.2474365234375,
0.54345703125,
0.2474365234375,
0.54345703125,
0.2587890625,
0.453369140625,
0.2587890625
],
[
0.233642578125,
0.279052734375,
0.765625,
0.27783203125,
0.765625,
0.292236328125,
0.233642578125,
0.29345703125
],
[
0.233642578125,
0.292236328125,
0.62353515625,
0.292236328125,
0.62353515625,
0.30615234375,
0.233642578125,
0.30615234375
],
[
0.615234375,
0.292236328125,
0.76708984375,
0.292236328125,
0.76708984375,
0.30615234375,
0.615234375,
0.30615234375
],
[
0.233642578125,
0.306884765625,
0.765625,
0.306884765625,
0.765625,
0.32080078125,
0.233642578125,
0.32080078125
],
[
0.2344970703125,
0.3212890625,
0.765625,
0.3212890625,
0.765625,
0.335205078125,
0.2344970703125,
0.335205078125
],
[
0.233642578125,
0.335205078125,
0.765625,
0.335205078125,
0.765625,
0.348388671875,
0.233642578125,
0.348388671875
],
[
0.233642578125,
0.348388671875,
0.765625,
0.348388671875,
0.765625,
0.3623046875,
0.233642578125,
0.3623046875
],
[
0.233642578125,
0.363037109375,
0.765625,
0.363037109375,
0.765625,
0.376953125,
0.233642578125,
0.376953125
],
[
0.236083984375,
0.37744140625,
0.7646484375,
0.37744140625,
0.7646484375,
0.38818359375,
0.236083984375,
0.38818359375
],
[
0.2344970703125,
0.39013671875,
0.7646484375,
0.39013671875,
0.7646484375,
0.404052734375,
0.2344970703125,
0.404052734375
],
[
0.233642578125,
0.40283203125,
0.765625,
0.404052734375,
0.765625,
0.41845703125,
0.233642578125,
0.417236328125
],
[
0.2344970703125,
0.41845703125,
0.76708984375,
0.41845703125,
0.76708984375,
0.431884765625,
0.2344970703125,
0.431884765625
],
[
0.233642578125,
0.43115234375,
0.765625,
0.43115234375,
0.765625,
0.4443359375,
0.233642578125,
0.4443359375
],
[
0.233642578125,
0.44580078125,
0.5498046875,
0.44580078125,
0.5498046875,
0.458984375,
0.233642578125,
0.458984375
],
[
0.176513671875,
0.4892578125,
0.2091064453125,
0.4892578125,
0.2091064453125,
0.50048828125,
0.176513671875,
0.50048828125
],
[
0.2059326171875,
0.4892578125,
0.340576171875,
0.4892578125,
0.340576171875,
0.50244140625,
0.2059326171875,
0.50244140625
],
[
0.178955078125,
0.52294921875,
0.8203125,
0.52294921875,
0.8203125,
0.53271484375,
0.178955078125,
0.53271484375
],
[
0.176513671875,
0.53515625,
0.82275390625,
0.53515625,
0.82275390625,
0.548828125,
0.176513671875,
0.548828125
],
[
0.17724609375,
0.548828125,
0.82275390625,
0.548828125,
0.82275390625,
0.5625,
0.17724609375,
0.5625
],
[
0.176513671875,
0.56298828125,
0.29736328125,
0.56298828125,
0.29736328125,
0.5771484375,
0.176513671875,
0.5771484375
],
[
0.176513671875,
0.583984375,
0.82275390625,
0.583984375,
0.82275390625,
0.59765625,
0.176513671875,
0.59765625
],
[
0.176513671875,
0.59765625,
0.82275390625,
0.59765625,
0.82275390625,
0.611328125,
0.176513671875,
0.611328125
],
[
0.176513671875,
0.6123046875,
0.82275390625,
0.6123046875,
0.82275390625,
0.62548828125,
0.176513671875,
0.62548828125
],
[
0.176513671875,
0.625,
0.82177734375,
0.625,
0.82177734375,
0.63818359375,
0.176513671875,
0.63818359375
],
[
0.176513671875,
0.6396484375,
0.82421875,
0.6396484375,
0.82421875,
0.65283203125,
0.176513671875,
0.65283203125
],
[
0.1748046875,
0.6513671875,
0.82275390625,
0.65283203125,
0.82275390625,
0.66748046875,
0.1748046875,
0.66650390625
],
[
0.176513671875,
0.66748046875,
0.82421875,
0.66748046875,
0.82421875,
0.68115234375,
0.176513671875,
0.68115234375
],
[
0.176513671875,
0.68115234375,
0.435546875,
0.68115234375,
0.435546875,
0.6943359375,
0.176513671875,
0.6943359375
],
[
0.1748046875,
0.70068359375,
0.82421875,
0.7021484375,
0.82421875,
0.71533203125,
0.1748046875,
0.7138671875
],
[
0.176513671875,
0.71630859375,
0.82177734375,
0.71630859375,
0.82177734375,
0.72998046875,
0.176513671875,
0.72998046875
],
[
0.176513671875,
0.72998046875,
0.82275390625,
0.72998046875,
0.82275390625,
0.74365234375,
0.176513671875,
0.74365234375
],
[
0.176513671875,
0.74365234375,
0.82275390625,
0.74365234375,
0.82275390625,
0.7568359375,
0.176513671875,
0.7568359375
],
[
0.17724609375,
0.75830078125,
0.82275390625,
0.75830078125,
0.82275390625,
0.771484375,
0.17724609375,
0.771484375
],
[
0.1748046875,
0.77099609375,
0.82177734375,
0.771484375,
0.82177734375,
0.78515625,
0.1748046875,
0.7841796875
],
[
0.176513671875,
0.79248046875,
0.82275390625,
0.79248046875,
0.82275390625,
0.80615234375,
0.176513671875,
0.80615234375
],
[
0.176513671875,
0.80615234375,
0.82275390625,
0.80615234375,
0.82275390625,
0.8193359375,
0.176513671875,
0.8193359375
],
[
0.176513671875,
0.82080078125,
0.82275390625,
0.82080078125,
0.82275390625,
0.833984375,
0.176513671875,
0.833984375
],
[
0.176513671875,
0.833984375,
0.82275390625,
0.833984375,
0.82275390625,
0.84765625,
0.176513671875,
0.84765625
],
[
0.176513671875,
0.8486328125,
0.82421875,
0.8486328125,
0.82421875,
0.8623046875,
0.176513671875,
0.8623046875
],
[
0.1748046875,
0.861328125,
0.449462890625,
0.8623046875,
0.449462890625,
0.87548828125,
0.1748046875,
0.875
],
[
0.176513671875,
0.88330078125,
0.82275390625,
0.88330078125,
0.82275390625,
0.896484375,
0.176513671875,
0.896484375
],
[
0.176513671875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.176513671875,
0.9111328125,
0.82275390625,
0.9111328125,
0.82275390625,
0.9248046875,
0.176513671875,
0.9248046875
],
[
0.491943359375,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.96044921875,
0.491943359375,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2024",
"上下文感知元学习",
"Christopher Fifty', Dennis Duanl,2, Ronald G. Junkins",
"Ehsan Amid'3, Jure Leskovec', Christopher Re', Sebastian Thrun",
"1 Stanford University,2Google, 3Google DeepMind",
"fifty@cs.stanford.com",
"ABSTRACT",
"Large Language Models like ChatGPT demonstrate a remarkable capacity to learn",
"new concepts during inference without any fine-tuning. I",
"However, visual mod-",
"els trained to detect new objects during inference have been unable to replicate",
"this ability, and instead either perform poorly or require meta-training and/or fine-",
"tuning on similar objects. In this work, we propose a meta-learning algorithm",
"that emulates Large Language Models by learning new visual concepts during in-",
"ference without fine-tuning. Our approach leverages a frozen pre-trained feature",
"extractor,andanalogoustoin-contextlearning,recastsvisualmeta-learningasse",
"quence modeling over datapoints with known labels and a test datapoint with an",
"unknown label. On 8 out of 11 few-shot image classification benchmarks, our",
"approach-without meta-training or fine-tuning--exceeds or matches the state-of-",
"the-art algorithm, P>M>F, which is meta-trained on these benchmarks. Our code",
"isavailableat https://github.com/cfifty/CAML",
"1.",
"INTRODUCTION",
"元学习是指从少量演示中学习新概念的能力（Lakeetal..2015）。在机器智能取得显著进步",
"的十年中，它仍然是人类性能继续超越机器的领域（Brownetal,2020）。为了匹配人类的能",
"力，为了开发能够像人类一样学习和思考的机器，我们必须开发能够从几个例子中学习新",
"概念的机器智能。",
"深度学习的许多应用都将学习算法应用于大量的训练数据集；但是，从极少数的训练示例中",
"学习会带来挑战E（Lakeetal.，2017；Garneloetal.，2018）。这一挑战导致了两种主要的评估",
"设置：域内和跨域。域内设置评估元学习者在特定域内接受类似任务培训后快速适应新任",
"务的能力。为此设置设计的模型通常非常快，但对目标域E(Chenetal,2019)以外的任务的",
"泛化能力很差。同时，跨域设置评估元学习者适应以前看不见的域中的任务的能力。为此设",
"置设计的方法具有很高的适应性，但在推理过程中速度很慢，因为它们需要在支持集（Guo",
"etal，2020；Ohetal.，2022；Huetal.,2022）上进行微调。关键的是，两种情况下的元学习者都",
"不同于人类快速推广到新任务的能力",
"最近，大型语言模型（LLM）在自然语言中解决了同时快速和通用元学习的问题。像ChatGPT",
"这样的LLM可以通过一种称为上下文学习的能力快速推广到新任务（Brownetal.，2020）。",
"然而，它仍然是一个开放的问题，在计算机视觉。即使是最好的视觉元学习算法也无法部署",
"到类似ChatGPT的系统中，因为此类系统所需的模型可以（1）推广到训练时未知的广泛任",
"务集，并且（2）实时执行，而无需花费时间对模型进行微调。LLM表现出了两种能力的非",
"凡能力；但是，当前的视觉元学习者可能只能满足一个要求或另一个要求E（Huetal.,2022)",
"为了衡量实现快速和通用视觉元学习者这一目标的进展，我们开发了一种评估范式，我们称",
"之为通用元学习。通用元学习测量模型快速学习新图像类的能力。它评估跨跨许多不同图",
"像分类任务的各种元学习基准的模型，而无需对任何基准的训练集进行元训练或在推理期",
"间对支持集进行微调。我们专注于少量图像分类的应用-而不是像绘画或分割这样的密集预",
"测任务-因为已经为这些应用探索了通用设置（Baretal.,2022;Zhangetal.,2023；Wangetal.,",
"2023;Kimetal.,2023;Butoietal.,2023)",
"除了通用设置中的基准方法之外，我们还介绍了一种实现强大通用性能的元学习器。从",
"LLM的上下文学习中取灵感，我们将n-way-k-shot图像分类重新表述为支持集和未知",
"查询图像上的非因果序列建模。具体来说，给定每个类的k示例的n方式分类，我们在",
"1"
] |
{
"arxiv_id": "2310.10971",
"language": "zh",
"page_idx": 0
}
|
||
2310.10971-zh-1
|
2310.10971-zh-1.jpg
|
Published as a conference paper at ICLR 2024
$\{ ( x_i, y_i) \} _{i= 1}^{nk}$ (图像，标签)支持集对和未标记的查询图像 $x_{nk+1}$ 上训练非因果模型，以预测查询图像的标签。此公式使元学习者在其参数空间中推断出新类，从而使其能够在推理过程中学习新的视觉概念，而无需进行微调。由于其“在上下文中”学习视觉信息的能力， 我们将我们的方法称为上下文感知元学习 (CAML)
总之，我们的贡献是双重的。首先，我们开发了一种元学习评估范式，该范式近似于类似ChatGPT 的应用程序中视觉元学习者的表现。其次，我们设计了一种在这种情况下运行良好的元学习算法。我们的实证研究结果表明，CAML 在通用环境中优于其他元学习者。值得注意的是，CAML在通用设置中的性能通常与最先进的元学习算法 P>M>FE(Hu et al., 2022) 的域内性能相匹配-甚至超过了-在每个下游基准上直接训练
### 2 RELATEDWORK
Meta-Learning as Causal Sequence Modeling. Several of the earliest meta-learning algorithms were formulated as causal sequence modeling problems. Hochreiter et al.(2001) leverage a LSTM (Hochreiter & Schmidhuber, 1997) to model extensions to semi-linear and quadratic functions, and two decades later, Graves et al. (2014); Santoro et al. (2016); Kaiser et al. (2017) build upon this approach by integrating a form of external memory that the LSTM can read to and write from memory to develop Neural Turing Machines. With the advent of self-attention (Vaswani et al., 2017), Mishra et al. (2017) predict the labels of query images by first composing a sequence of (image, label) pairs and then feeding it through a stack of interleaved causal self-attention and temporal convolution layers. Kirsch et al.(2022) replaces the stack of interleaved causal self-attention and temporal convolution layers with a Transformer encoder; however, their approach is also causal in the input sequence by composing a sequence of (image, label of previous image) pairs. Both Mishra et al.(2017) and Kirsch et al.(2022) are conceptually similar to our work; however, the causal property of both approaches breaks an important symmetry in meta-learning, namely invariance to permutations of the support set (Garnelo et al., 2018; Müller et al., 2021). In Section 5.2 , we observe a performance gap between both approaches and CAML and hypothesize the causal approach actually forces a subtly more difficult modeling problem by imposing a causality inductive bias on a fundamentally non-causal prediction task.
跨域元学习。跨域元学习是指一种具有挑战性的评估范式，其中元训练和推理时间数据分布显着不同 (Chen et al.,2019)。最近的工作发现，利用自监督预训练-或基础模型特征提取器-可以显着提高跨域性能$\mathbb{E}($Hu et al, 2022; Zhang et al, 2021)。此外，关于支持集的微调几乎总是优于元学习，而无需在此设置中进行微调 (Guo et al.,2020; Oh et al.,2022; Phoo & Hariharan, 2020; Islam et al.,2021)。虽然有效，但微调对于以类似于像 ChatGPT 这样的 LLM 的方式部署可视化元学习模型是禁止的，因为在每个用户查询上微调模型参数的延迟和内存成本是站不住脚的。因此，我们提出了通用设置来衡量元学习者学习对推理过程中看到的任何任务进行分类而无需微调的能力
In-Context Learning for Dense Prediction Tasks. Many recent works have explored in-context learning for other applications of computer vision. Bar et al.(2022) casts in-context learning as image in-painting by first concatenating demonstration images with a query image and then using a vision model to fill-in-the-blank within this concatenated image. Building on this work, Zhang et al. (2023) explores what demonstrations lead to strong in-painting performance and Wang et al. (2023) generalizes the approach by formulating other visual applications like segmentation, depth estimation, etc. as in-painting. Other approaches explore in-context learning for applications like scene understanding (Balazevic et al., 2024),medical image segmentation (Butoi et al., 2023), and more generally dense prediction tasks (Kim et al., 2023) . Like these approaches, we study visual in-context learning; however, this work focuses on few-shot image classification rather than dense prediction tasks.
## 3 APPROACH
我们调整了 LLM 中支持上下文学习的思想-即学习在单个正向传递中从支持集演示的上下文中对查询进行分类-到图像分类。但是，与上下文学习不同，视觉元学习者应该是非因果的：在支持集中将一个示例放在另一个示例之前不会产生因果关系$\overline{\mathrm{E}}($Garnelo et al.,2018 Müller et al., 2021)
2
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.1748046875,
0.1029052734375,
0.82421875,
0.10418701171875,
0.82421875,
0.121826171875,
0.1748046875,
0.12060546875
],
[
0.1748046875,
0.117431640625,
0.82275390625,
0.11871337890625,
0.82275390625,
0.1331787109375,
0.1748046875,
0.1319580078125
],
[
0.1748046875,
0.1319580078125,
0.82275390625,
0.1319580078125,
0.82275390625,
0.146484375,
0.1748046875,
0.146484375
],
[
0.1732177734375,
0.146484375,
0.54345703125,
0.1458740234375,
0.54345703125,
0.160400390625,
0.1732177734375,
0.1610107421875
],
[
0.17724609375,
0.1685791015625,
0.82275390625,
0.1685791015625,
0.82275390625,
0.1817626953125,
0.17724609375,
0.1817626953125
],
[
0.1748046875,
0.1817626953125,
0.82275390625,
0.1817626953125,
0.82275390625,
0.1956787109375,
0.1748046875,
0.1956787109375
],
[
0.1748046875,
0.1944580078125,
0.82177734375,
0.1956787109375,
0.82177734375,
0.210205078125,
0.1748046875,
0.208984375
],
[
0.176513671875,
0.210205078125,
0.82275390625,
0.210205078125,
0.82275390625,
0.2235107421875,
0.176513671875,
0.2235107421875
],
[
0.176513671875,
0.2235107421875,
0.59619140625,
0.2235107421875,
0.59619140625,
0.2374267578125,
0.176513671875,
0.2374267578125
],
[
0.1748046875,
0.261962890625,
0.2091064453125,
0.264404296875,
0.20751953125,
0.276611328125,
0.1732177734375,
0.274658203125
],
[
0.205078125,
0.264404296875,
0.3505859375,
0.264404296875,
0.3505859375,
0.27783203125,
0.205078125,
0.27783203125
],
[
0.1748046875,
0.296630859375,
0.82421875,
0.296630859375,
0.82421875,
0.311279296875,
0.1748046875,
0.311279296875
],
[
0.1748046875,
0.311279296875,
0.82421875,
0.311279296875,
0.82421875,
0.324462890625,
0.1748046875,
0.324462890625
],
[
0.176513671875,
0.326904296875,
0.82177734375,
0.326904296875,
0.82177734375,
0.337158203125,
0.176513671875,
0.337158203125
],
[
0.1748046875,
0.33837890625,
0.82421875,
0.33837890625,
0.82421875,
0.3515625,
0.1748046875,
0.3515625
],
[
0.176513671875,
0.352783203125,
0.82421875,
0.352783203125,
0.82421875,
0.3662109375,
0.176513671875,
0.3662109375
],
[
0.1748046875,
0.367431640625,
0.82275390625,
0.367431640625,
0.82275390625,
0.380615234375,
0.1748046875,
0.380615234375
],
[
0.1732177734375,
0.37890625,
0.826171875,
0.380126953125,
0.826171875,
0.39453125,
0.1732177734375,
0.393310546875
],
[
0.17724609375,
0.395751953125,
0.82275390625,
0.395751953125,
0.82275390625,
0.406005859375,
0.17724609375,
0.406005859375
],
[
0.1748046875,
0.40771484375,
0.82421875,
0.40771484375,
0.82421875,
0.421630859375,
0.1748046875,
0.421630859375
],
[
0.1732177734375,
0.42041015625,
0.82421875,
0.421630859375,
0.82421875,
0.436279296875,
0.1732177734375,
0.43505859375
],
[
0.1748046875,
0.436279296875,
0.82421875,
0.436279296875,
0.82421875,
0.449462890625,
0.1748046875,
0.449462890625
],
[
0.1732177734375,
0.449462890625,
0.82275390625,
0.449462890625,
0.82275390625,
0.46337890625,
0.1732177734375,
0.46337890625
],
[
0.1748046875,
0.46337890625,
0.82421875,
0.46337890625,
0.82421875,
0.477783203125,
0.1748046875,
0.477783203125
],
[
0.176513671875,
0.477783203125,
0.826171875,
0.477783203125,
0.826171875,
0.4912109375,
0.176513671875,
0.4912109375
],
[
0.176513671875,
0.4912109375,
0.82421875,
0.4912109375,
0.82421875,
0.5048828125,
0.176513671875,
0.5048828125
],
[
0.1732177734375,
0.5048828125,
0.82275390625,
0.5048828125,
0.82275390625,
0.51953125,
0.1732177734375,
0.51953125
],
[
0.1748046875,
0.52099609375,
0.459228515625,
0.52099609375,
0.459228515625,
0.53076171875,
0.1748046875,
0.53076171875
],
[
0.176513671875,
0.54150390625,
0.82177734375,
0.54150390625,
0.82177734375,
0.5517578125,
0.176513671875,
0.5517578125
],
[
0.176513671875,
0.55322265625,
0.82275390625,
0.55322265625,
0.82275390625,
0.56640625,
0.176513671875,
0.56640625
],
[
0.176513671875,
0.5673828125,
0.82275390625,
0.5673828125,
0.82275390625,
0.58056640625,
0.176513671875,
0.58056640625
],
[
0.176513671875,
0.58056640625,
0.82177734375,
0.58056640625,
0.82177734375,
0.5947265625,
0.176513671875,
0.5947265625
],
[
0.176513671875,
0.59521484375,
0.82177734375,
0.59521484375,
0.82177734375,
0.60595703125,
0.176513671875,
0.60595703125
],
[
0.176513671875,
0.60791015625,
0.82177734375,
0.60791015625,
0.82177734375,
0.6220703125,
0.176513671875,
0.6220703125
],
[
0.176513671875,
0.62255859375,
0.82275390625,
0.62255859375,
0.82275390625,
0.63623046875,
0.176513671875,
0.63623046875
],
[
0.1748046875,
0.63720703125,
0.450927734375,
0.63720703125,
0.450927734375,
0.64794921875,
0.1748046875,
0.64794921875
],
[
0.1732177734375,
0.6572265625,
0.826171875,
0.65771484375,
0.826171875,
0.67236328125,
0.1732177734375,
0.671875
],
[
0.1748046875,
0.671875,
0.82421875,
0.671875,
0.82421875,
0.68603515625,
0.1748046875,
0.68603515625
],
[
0.1732177734375,
0.68603515625,
0.82666015625,
0.6875,
0.82666015625,
0.70068359375,
0.1732177734375,
0.69970703125
],
[
0.1732177734375,
0.69970703125,
0.8203125,
0.69970703125,
0.8203125,
0.71337890625,
0.1732177734375,
0.71337890625
],
[
0.1748046875,
0.7138671875,
0.82421875,
0.7138671875,
0.82421875,
0.72802734375,
0.1748046875,
0.72802734375
],
[
0.1748046875,
0.72802734375,
0.82275390625,
0.7265625,
0.82275390625,
0.74072265625,
0.1748046875,
0.7412109375
],
[
0.1748046875,
0.7412109375,
0.82275390625,
0.7412109375,
0.82275390625,
0.7548828125,
0.1748046875,
0.7548828125
],
[
0.1748046875,
0.7548828125,
0.82421875,
0.7548828125,
0.82421875,
0.76806640625,
0.1748046875,
0.76806640625
],
[
0.1748046875,
0.76806640625,
0.82421875,
0.76806640625,
0.82421875,
0.78271484375,
0.1748046875,
0.78271484375
],
[
0.1748046875,
0.78271484375,
0.82275390625,
0.78271484375,
0.82275390625,
0.796875,
0.1748046875,
0.796875
],
[
0.17236328125,
0.796875,
0.283447265625,
0.79541015625,
0.283447265625,
0.81005859375,
0.17236328125,
0.81103515625
],
[
0.1748046875,
0.8359375,
0.2091064453125,
0.8359375,
0.2091064453125,
0.8486328125,
0.1748046875,
0.8486328125
],
[
0.2059326171875,
0.83544921875,
0.3056640625,
0.8369140625,
0.3056640625,
0.8515625,
0.2059326171875,
0.849609375
],
[
0.1748046875,
0.869140625,
0.82177734375,
0.869140625,
0.82177734375,
0.88330078125,
0.1748046875,
0.88330078125
],
[
0.17724609375,
0.88330078125,
0.82177734375,
0.88330078125,
0.82177734375,
0.896484375,
0.17724609375,
0.896484375
],
[
0.1748046875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.1748046875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.3056640625,
0.9111328125,
0.3056640625,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.49267578125,
0.94873046875,
0.5048828125,
0.94873046875,
0.5048828125,
0.9609375,
0.49267578125,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2024",
"((ci,yi))n1（图像，标签）支持集对和未标记的查询图像ank+1上训练非因果模型，以预",
"测查询图像的标签。此公式使元学习者在其参数空间中推断出新类，从而使其能够在推理",
"过程中学习新的视觉概念，而无需进行微调。由于其“在上下文中”学习视觉信息的能力，",
"我们将我们的方法称为上下文感知元学习（CAML）",
"总之，我们的贡献是双重的。首先，我们开发了一种元学习评估范式，该范式近似于类似",
"ChatGPT的应用程序中视觉元学习者的表现。其次，我们设计了一种在这种情况下运行良好",
"的元学习算法。我们的实证研究结果表明，CAML在通用环境中优于其他元学习者。值得",
"注意的是，CAML在通用设置中的性能通常与最先进的元学习算法P>M>FE(Huetal.，2022)",
"的域内性能相匹配-甚至超过了-在每个下游基准上直接训练",
"2",
"RELATEDWORK",
"Meta-Learning as Causal Sequence Modeling. Several of the earliest meta-learning algorithms",
"wereformulated as causal sequence modeling problems.Hochreiter et al.(2001)leverage a",
"LSTM(Hochreiter&Schmidhuber,1997)tomodelextensionstosemi-linearandquadraticfunc",
"tions, and two decades later, Graves et al. (2014); Santoro et al. (2016); Kaiser et al. (2017) build",
"upon this approach by integrating a form of external memory that the LSTM can read to and write",
"from memory to develop Neural Turing Machines.With the advent of self-attention (Vaswani et al.,",
"2017), Mishra et al. (2017) predict the labels of query images by first composing a sequence of (im-",
"age,label)pairsandthenfeedingitthroughastackofinterleavedcausalself-attentionandtemporal",
"convolution layers. Kirsch et al. (2022) replaces the stack of interleaved causal self-attention and",
"temporal convolution layers with a Transformer encoder; however, their approach is also causal in",
"the input sequence by composing a sequence of (image, label of previous image)pairs.Both Mishra",
"et al. (2017) and Kirsch et al. (2022) are conceptually similar to our work; however, the causal",
"property of both approaches breaks an important symmetry in meta-learning, namely invariance to",
"permutations of the support set (Garnelo et al., 2018; Miller et al., 2021).In Section 5.2, we ob-",
"serve a performance gap between both approaches and CAML and hypothesize the causal approach",
"actually forces a subtly more difficult modeling problem by imposing a causality inductive bias on",
"afundamentallynon-causalpredictiontask",
"跨域元学习。跨域元学习是指一种具有挑战性的评估范式，其中元训练和推理时间数据分",
"布显着不同（Chenetal.，2019）。最近的工作发现，利用自监督预训练-或基础模型特征提取",
"器-可以显着提高跨域性能E(Huetal.，2022Zhangetal.，2021）。此外，关于支持集的微调",
"几乎总是优于元学习，而无需在此设置中进行微调（Guoetal.，2020;Ohetal，2022;Phoo&",
"Hariharan.2020：Islametal.2021）。虽然有效，但微调对于以类似于像ChatGPT这样的LLM",
"的方式部署可视化元学习模型是禁止的，因为在每个用户查询上微调模型参数的延迟和内",
"存成本是站不住脚的。因此，我们提出了通用设置来衡量元学习者学习对推理过程中看到",
"的任何任务进行分类而无需微调的能力",
"In-Context Learning for Dense Prediction Tasks. Many recent works have explored in-context learn-",
"ing for other applications of computer vision. Bar et al. (2022) casts in-context learning as image",
"in-painting by first concatenating demonstration images with a query image and then using a vi-",
"sion model to fill-in-the-blank within this concatenated image. Building on this work, Zhang et al",
"(2023) explores what demonstrations lead to strong in-painting performance and Wang et al. (2023)",
"generalizes the approach byformulating other visual applications like segmentation, depth estima",
"tion, etc. as in-painting. Other approaches explore in-context learning for applications like scene",
"understanding (Balazevic et al., 2024),medical image segmentation (Butoi et al.,2023), and",
"more generally dense prediction tasks (Kim et al., 2023). Like these approaches, we study visual",
"in-context learning; however, this work focuses on few-shot image classification rather than dense",
"prediction tasks.",
"3",
"APPROACH",
"我们调整了LLM中支持上下文学习的思想-即学习在单个正向传递中从支持集演示的上下",
"文中对查询进行分类-到图像分类。但是，与上下文学习不同，视觉元学习者应该是非因果",
"的：在支持集中将一个示例放在另一个示例之前不会产生因果关系E（Garmeloetal，2018",
"Muiller et al., 2021)",
"2"
] |
{
"arxiv_id": "2310.10971",
"language": "zh",
"page_idx": 1
}
|
||
2310.10971-zh-2
|
2310.10971-zh-2.jpg
|
Published as a conference paper at ICLR 2024
<FigureHere>
Figure 1: Overview of CAML. Query and support set images are encoded with a pre-trained feature extractor and then concatenated with their corresponding ELMES label embeddings. We feed the resulting sequence of concatenated vectors into a non-casual sequence model and extract the query vector from the output sequence to predict its class.
体系结构。Figure 1 中显示了 CAML 的概述。它由三个不同的组件组成：(1) 冻结的预训练图像编码器，(2)固定的等长和最大等角集(ELMES)类编码器，以及(3)非因果序列模型。虽然预先训练的图像编码器和非因果序列模型是众所周知的，但为了编码标签信息，我们引入了 ELMES 编码器。ELMES 编码器是标签与一组长度相等且最大等角的向量之间的双射映射。从历史上看，标签是用独热向量编码的；但是在 Section 4 中，我们证明了互斥类的 ELMES 编码允许序列模型最大程度地识别支持集中的类
正如在 Figure 1 中可视化的那样，CAML 首先使用冻结的预训练特征提取器对查询和支持集图像进行编码。至关重要的是，预先训练的图像编码器的嵌入空间将图像提取为低维表示，以便具有相似内容和视觉特征的图像具有相似的嵌人。支持集的类使用 ELMES 类编码器进行编码；但是，由于查询的类是未知的，因此我们使用在大规模预训练期间学习的特殊可学习的“未知令牌” 嵌人。CAML 然后将每个图像嵌人与其相应的查询嵌入连接起来，以形成一个输人序列
通过 Figure 1 进行，该序列被馈送到非因果序列模型中，即。e。转换器编码器，用于在查询和支持设置点的完整上下文上调整输出表示。这实现了动态和实时分类；来自查询和支持集图像的视觉特征可以相互比较，以确定特定的视觉特征—如内容、纹理等。--用于对查询进行分类。从非因果序列模型的输出序列中，我们选择与输人序列中的查询位置相同的元素，并将此向量通过浅 MLP 来预测查询的标签
大规模预培训。由于我们的重点是通用元学习--而 CAML 在推理过程中可能会遇到任何新的视觉概念---我们在 ImageNet-1k$\boxed{\mathrm{F}}($Deng et al.,2009) , Fungi$\boxed{\mathrm{E}}($Schroeder &Cui, 2018) . MSCOCO$\underline{\mathbb{F}}($Lin et al., 2014) 和 WikiArt$\underline{\mathbb{E}}($Saleh& Elgammal, 2015) 的少量图像分类任务上预先训练 CAML的非因果序列。我们之所以选择这些数据集，是因为它们涵盖了通用对象识别 (ImageNet-1k, MSCOCO), 细粒度图像分类 (Fungi) 和非自然图像分类 (WikiArt)。为了避免扭曲预训练的图像编码器的嵌人空间，我们冻结这个模块，只在预训练期间更新序列模型的参数。同样，由于 ELMES 最小化了支持集内检测类的熵，因此标签编码器也被冻结。在预训练，元训练和微调的上下文中，CAML 仅需要预训练，并避免在元学习基准的训练/验证拆分上进行元训练或在推理过程中对支持集进行微调
4 THEORETICAL ANALYSIS
在本节中，我们通过考虑元学习算法中所需的对称性来激发我们对 ELMES 类编码器的选择。两个重要的对称性是(1)将支持集类分配给数字标签的不变性和(2)输人序列排序中排列的不变性。第一个不变性意味着类嵌人必须是等角且范数相等的，并且 ELMES 配置最小化了检测任何给定类的可学习模型参数的熵。后来，我们证明了 ELMES 也满足第二对称性。由于空间限制，所有证明和许多定义，属性，引理和定理都分配给 ??。我们从 ELMES 的正式定义开始
3
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.23779296875,
0.0662841796875,
0.363525390625,
0.06756591796875,
0.363525390625,
0.08209228515625,
0.23779296875,
0.080810546875
],
[
0.60693359375,
0.06756591796875,
0.744140625,
0.06756591796875,
0.744140625,
0.080810546875,
0.60693359375,
0.080810546875
],
[
0.17724609375,
0.091552734375,
0.253173828125,
0.091552734375,
0.253173828125,
0.1029052734375,
0.17724609375,
0.1029052734375
],
[
0.281005859375,
0.09027099609375,
0.396240234375,
0.091552734375,
0.396240234375,
0.1029052734375,
0.281005859375,
0.10162353515625
],
[
0.6787109375,
0.10162353515625,
0.72216796875,
0.10162353515625,
0.72216796875,
0.11614990234375,
0.6787109375,
0.11614990234375
],
[
0.72607421875,
0.0997314453125,
0.7900390625,
0.10162353515625,
0.7900390625,
0.11614990234375,
0.72607421875,
0.1142578125
],
[
0.238525390625,
0.10797119140625,
0.419189453125,
0.10797119140625,
0.419189453125,
0.15283203125,
0.238525390625,
0.15283203125
],
[
0.68310546875,
0.117431640625,
0.72119140625,
0.117431640625,
0.72119140625,
0.1307373046875,
0.68310546875,
0.1307373046875
],
[
0.74853515625,
0.117431640625,
0.771484375,
0.117431640625,
0.771484375,
0.1300048828125,
0.74853515625,
0.1300048828125
],
[
0.68115234375,
0.1300048828125,
0.72216796875,
0.1300048828125,
0.72216796875,
0.14453125,
0.68115234375,
0.14453125
],
[
0.57177734375,
0.121826171875,
0.6044921875,
0.121826171875,
0.6044921875,
0.1337890625,
0.57177734375,
0.1337890625
],
[
0.2181396484375,
0.1307373046875,
0.25,
0.1307373046875,
0.25,
0.143310546875,
0.2181396484375,
0.143310546875
],
[
0.74658203125,
0.1307373046875,
0.771484375,
0.1307373046875,
0.771484375,
0.1427001953125,
0.74658203125,
0.1427001953125
],
[
0.68115234375,
0.143310546875,
0.72216796875,
0.143310546875,
0.72216796875,
0.1578369140625,
0.68115234375,
0.1578369140625
],
[
0.74658203125,
0.14453125,
0.77294921875,
0.14453125,
0.77294921875,
0.1558837890625,
0.74658203125,
0.1558837890625
],
[
0.226318359375,
0.1641845703125,
0.39208984375,
0.1641845703125,
0.39208984375,
0.177978515625,
0.226318359375,
0.177978515625
],
[
0.63818359375,
0.1787109375,
0.64453125,
0.1787109375,
0.64453125,
0.1817626953125,
0.63818359375,
0.1817626953125
],
[
0.69775390625,
0.1767578125,
0.71826171875,
0.1767578125,
0.71826171875,
0.184326171875,
0.69775390625,
0.184326171875
],
[
0.22216796875,
0.1925048828125,
0.24267578125,
0.189453125,
0.2442626953125,
0.1976318359375,
0.223876953125,
0.199462890625
],
[
0.2484130859375,
0.189453125,
0.283447265625,
0.1925048828125,
0.281982421875,
0.2008056640625,
0.2467041015625,
0.1976318359375
],
[
0.28759765625,
0.1932373046875,
0.298095703125,
0.1932373046875,
0.298095703125,
0.1962890625,
0.28759765625,
0.1962890625
],
[
0.32275390625,
0.1925048828125,
0.343017578125,
0.1925048828125,
0.343017578125,
0.199462890625,
0.32275390625,
0.199462890625
],
[
0.455810546875,
0.1962890625,
0.53369140625,
0.1962890625,
0.53369140625,
0.2083740234375,
0.455810546875,
0.2083740234375
],
[
0.45751953125,
0.205810546875,
0.5361328125,
0.208984375,
0.53515625,
0.2235107421875,
0.455810546875,
0.2203369140625
],
[
0.5888671875,
0.20703125,
0.765625,
0.20703125,
0.765625,
0.2203369140625,
0.5888671875,
0.2203369140625
],
[
0.2017822265625,
0.21533203125,
0.236083984375,
0.21533203125,
0.236083984375,
0.2266845703125,
0.2017822265625,
0.2266845703125
],
[
0.261474609375,
0.216552734375,
0.301513671875,
0.216552734375,
0.301513671875,
0.2266845703125,
0.261474609375,
0.2266845703125
],
[
0.32275390625,
0.2171630859375,
0.35693359375,
0.2171630859375,
0.35693359375,
0.2247314453125,
0.32275390625,
0.2247314453125
],
[
0.238525390625,
0.2392578125,
0.3798828125,
0.2392578125,
0.3798828125,
0.25244140625,
0.238525390625,
0.25244140625
],
[
0.1846923828125,
0.25244140625,
0.2467041015625,
0.255615234375,
0.2467041015625,
0.266357421875,
0.1846923828125,
0.26318359375
],
[
0.2484130859375,
0.2587890625,
0.42236328125,
0.260009765625,
0.42236328125,
0.274658203125,
0.2484130859375,
0.2734375
],
[
0.1968994140625,
0.266357421875,
0.2303466796875,
0.266357421875,
0.2303466796875,
0.27587890625,
0.1968994140625,
0.27587890625
],
[
0.193603515625,
0.2734375,
0.259033203125,
0.27587890625,
0.259033203125,
0.287353515625,
0.193603515625,
0.285400390625
],
[
0.281982421875,
0.27978515625,
0.404296875,
0.27978515625,
0.404296875,
0.29345703125,
0.281982421875,
0.29345703125
],
[
0.176513671875,
0.3017578125,
0.82275390625,
0.3017578125,
0.82275390625,
0.31494140625,
0.176513671875,
0.31494140625
],
[
0.1732177734375,
0.313232421875,
0.82275390625,
0.313232421875,
0.82275390625,
0.326904296875,
0.1732177734375,
0.326904296875
],
[
0.176513671875,
0.328857421875,
0.82275390625,
0.328857421875,
0.82275390625,
0.339111328125,
0.176513671875,
0.339111328125
],
[
0.1748046875,
0.341552734375,
0.28515625,
0.341552734375,
0.28515625,
0.3515625,
0.1748046875,
0.3515625
],
[
0.1748046875,
0.37890625,
0.82275390625,
0.380126953125,
0.82275390625,
0.39453125,
0.1748046875,
0.393310546875
],
[
0.176513671875,
0.39453125,
0.82275390625,
0.39453125,
0.82275390625,
0.40771484375,
0.176513671875,
0.40771484375
],
[
0.1748046875,
0.40771484375,
0.82177734375,
0.40771484375,
0.82177734375,
0.421630859375,
0.1748046875,
0.421630859375
],
[
0.176513671875,
0.42041015625,
0.82275390625,
0.421630859375,
0.82275390625,
0.436279296875,
0.176513671875,
0.43505859375
],
[
0.176513671875,
0.436279296875,
0.82275390625,
0.436279296875,
0.82275390625,
0.449462890625,
0.176513671875,
0.449462890625
],
[
0.1748046875,
0.448974609375,
0.61376953125,
0.449462890625,
0.61376953125,
0.46337890625,
0.1748046875,
0.462158203125
],
[
0.178955078125,
0.47021484375,
0.82275390625,
0.47021484375,
0.82275390625,
0.484130859375,
0.178955078125,
0.484130859375
],
[
0.17724609375,
0.486083984375,
0.8203125,
0.486083984375,
0.8203125,
0.496826171875,
0.17724609375,
0.496826171875
],
[
0.1748046875,
0.498779296875,
0.82275390625,
0.49755859375,
0.82275390625,
0.51220703125,
0.1748046875,
0.51318359375
],
[
0.17724609375,
0.5146484375,
0.82177734375,
0.5146484375,
0.82177734375,
0.5244140625,
0.17724609375,
0.5244140625
],
[
0.176513671875,
0.5263671875,
0.82275390625,
0.5263671875,
0.82275390625,
0.54052734375,
0.176513671875,
0.54052734375
],
[
0.1748046875,
0.54052734375,
0.30810546875,
0.54052734375,
0.30810546875,
0.5537109375,
0.1748046875,
0.5537109375
],
[
0.17724609375,
0.56298828125,
0.82177734375,
0.56298828125,
0.82177734375,
0.57373046875,
0.17724609375,
0.57373046875
],
[
0.17724609375,
0.57568359375,
0.82275390625,
0.57568359375,
0.82275390625,
0.5888671875,
0.17724609375,
0.5888671875
],
[
0.1748046875,
0.58837890625,
0.82177734375,
0.58837890625,
0.82177734375,
0.6015625,
0.1748046875,
0.6015625
],
[
0.1748046875,
0.60302734375,
0.82177734375,
0.60302734375,
0.82177734375,
0.6162109375,
0.1748046875,
0.6162109375
],
[
0.1748046875,
0.6171875,
0.5439453125,
0.6171875,
0.5439453125,
0.630859375,
0.1748046875,
0.630859375
],
[
0.17724609375,
0.63818359375,
0.82275390625,
0.63818359375,
0.82275390625,
0.6513671875,
0.17724609375,
0.6513671875
],
[
0.176513671875,
0.6513671875,
0.82177734375,
0.6513671875,
0.82177734375,
0.66552734375,
0.176513671875,
0.66552734375
],
[
0.176513671875,
0.66748046875,
0.82177734375,
0.66748046875,
0.82177734375,
0.67822265625,
0.176513671875,
0.67822265625
],
[
0.17724609375,
0.6796875,
0.82275390625,
0.6796875,
0.82275390625,
0.693359375,
0.17724609375,
0.693359375
],
[
0.176513671875,
0.693359375,
0.82275390625,
0.693359375,
0.82275390625,
0.70703125,
0.176513671875,
0.70703125
],
[
0.178955078125,
0.70849609375,
0.82177734375,
0.70849609375,
0.82177734375,
0.7216796875,
0.178955078125,
0.7216796875
],
[
0.17724609375,
0.72021484375,
0.82177734375,
0.72021484375,
0.82177734375,
0.734375,
0.17724609375,
0.734375
],
[
0.176513671875,
0.73486328125,
0.82275390625,
0.73486328125,
0.82275390625,
0.74853515625,
0.176513671875,
0.74853515625
],
[
0.176513671875,
0.75,
0.6044921875,
0.75,
0.6044921875,
0.76318359375,
0.176513671875,
0.76318359375
],
[
0.176513671875,
0.80322265625,
0.20751953125,
0.80322265625,
0.20751953125,
0.81103515625,
0.176513671875,
0.81103515625
],
[
0.20751953125,
0.80029296875,
0.423095703125,
0.80029296875,
0.423095703125,
0.814453125,
0.20751953125,
0.814453125
],
[
0.176513671875,
0.84130859375,
0.82421875,
0.84130859375,
0.82421875,
0.85498046875,
0.176513671875,
0.85498046875
],
[
0.176513671875,
0.85498046875,
0.82275390625,
0.85498046875,
0.82275390625,
0.86865234375,
0.176513671875,
0.86865234375
],
[
0.1748046875,
0.869140625,
0.82275390625,
0.869140625,
0.82275390625,
0.88330078125,
0.1748046875,
0.88330078125
],
[
0.17724609375,
0.88330078125,
0.82275390625,
0.88330078125,
0.82275390625,
0.896484375,
0.17724609375,
0.896484375
],
[
0.1748046875,
0.89599609375,
0.82421875,
0.896484375,
0.82421875,
0.9111328125,
0.1748046875,
0.91015625
],
[
0.1748046875,
0.9111328125,
0.2900390625,
0.9111328125,
0.2900390625,
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 2024",
"Parameters Frozen",
"Parameters Trained",
"QueryImage",
"Support Set Images",
"Class",
"[P(Class)",
"卫人",
"[1, 0, 0]",
"0.03",
"[0, 1, 0] ",
"MLP",
"",
"0.06",
"[0, 0, 1] ",
"0.91",
"Pre-Trained ImageEncoder",
"",
"",
"",
"",
"",
"",
"Concatenate",
"Embeddings",
"Non-Causal SequenceModel",
"*",
"",
"",
"ELMES Class Encoder",
"*Unknown",
"[1, 0, 0] [0, 1,0] [0, 0, ]",
"[class]",
"Embedding",
"Support Set Classes",
"Figure l: Overview of CAML. Query and support set images are encoded with a pre-trained feature extractor",
"and then concatenated with their corresponding ELMEs label embeddings. We feed the resulting sequence of",
"concatenated vectors into a non-casual sequencemodel and extractthequeryvector from the output sequence",
"to predict its class.",
"体系结构。Figure1中显示了CAML的概述。它由三个不同的组件组成：（1）冻结的预训练",
"图像编码器，（2）固定的等长和最大等角集（ELMES）类编码器，以及（3）非因果序列模",
"型。虽然预先训练的图像编码器和非因果序列模型是众所周知的，但为了编码标签信息，我",
"们引人了ELMES编码器。ELMES编码器是标签与一组长度相等且最大等角的向量之间的",
"双射映射。从历史上看，标签是用独热向量编码的；但是在Section4中，我们证明了互斥",
"类的ELMES编码允许序列模型最大程度地识别支持集中的类",
"正如在Figure1中可视化的那样，CAML首先使用冻结的预训练特征提取器对查询和支持",
"集图像进行编码。至关重要的是预先训练的图像编码器的嵌入空间将图像提取为低维表",
"示，以便具有相似内容和视觉特征的图像具有相似的嵌人。支持集的类使用ELMES类编码",
"器进行编码；但是，由于查询的类是未知的，因此我们使用在大规模预训练期间学习的特殊",
"可学习的“未知令牌”嵌入。CAML然后将每个图像嵌人与其相应的查询嵌人连接起来，以",
"形成一个输人序列",
"通过Figure1进行，该序列被馈送到非因果序列模型中，即。e。转换器编码器，用于在查",
"询和支持设置点的完整上下文上调整输出表示。这实现了动态和实时分类；来自查询和支",
"持集图像的视觉特征可以相互比较，以确定特定的视觉特征一如内容、纹理等。--用于对",
"查询进行分类。从非因果序列模型的输出序列中，我们选择与输入序列中的查询位置相同",
"的元素，并将此向量通过浅MLP来预测查询的标签",
"大规模预培训。由于我们的重点是通用元学习---而CAML在推理过程中可能会遇到任何",
"新的视觉概念---我们在ImageNet-1kE（Dengetal.，2009），FungiE（Schroeder&Cui,2018）",
"MSCOCOEl（Linetal.，2014）和WikiArtE（Saleh&Elgammal，2015）的少量图像分类任务上预",
"先训练CAML的非因果序列。我们之所以选择这些数据集，是因为它们涵盖了通用对象识",
"别（ImageNet-1k，MSCOcO），细粒度图像分类（Fungi）和非自然图像分类（WikiArt）。为",
"了避免扭曲预训练的图像编码器的嵌入空间，我们冻结这个模块，只在预训练期间更新序",
"列模型的参数。同样，由于ELMES最小化了支持集内检测类的熵，因此标签编码器也被冻",
"结。在预训练，元训练和微调的上下文中，CAML仅需要预训练，并避免在元学习基准的训",
"练/验证拆分上进行元训练或在推理过程中对支持集进行微调",
"4",
"THEORETICALANALYSIS",
"在本节中，我们通过考虑元学习算法中所需的对称性来激发我们对ELMES类编码器的选",
"择。两个重要的对称性是（1）将支持集类分配给数字标签的不变性和（2）输入序列排序中",
"排列的不变性。第一个不变性意味着类嵌人必须是等角且范数相等的，并且ELMES配置最",
"小化了检测任何给定类的可学习模型参数的熵。后来，我们证明了ELMES也满足第二对称",
"性。由于空间限制，所有证明和许多定义，属性，引理和定理都分配给？？。我们从ELMES",
"的正式定义开始",
"3"
] |
{
"arxiv_id": "2310.10971",
"language": "zh",
"page_idx": 2
}
|
||
2310.10971-zh-3
|
2310.10971-zh-3.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2024}}$
4.1 EQUAL LENGTH AND MAXIMALLY EQUIANGULAR SET OF VECTORS
$\textbf{Definition 1. An Equal Length and Maximally Equiangular Set ( ELMES) is a set of non- zeros }$ $vectors\left \{ \phi _{j}\right \} _{j= 1}^{d}, \phi _{j}\in \mathbb{R} ^{d+ k}forsome$ $k\geq 0$ and $d> 1$, such that $\forall j\neq j^{\prime }, \| \phi _{j}\| = \| \phi _{j^{\prime }}\|$ and $\langle \phi _{j}$ , $\phi _{j^{\prime }}\rangle = \frac {- 1}{d- 1}$ . $Simply$, all vectors $in$ this set are equal length and maximally $equiangular.$
向量的等角集和最大等角集 (ELMES) 与表示理论中的等角紧帧$\overline{\mathrm{F}}($Welch,1974;Fickus et al. 2018) 以及最近的神经崩溃作品中突出显示的单纯形等角紧帧$\overline{\mathrm{E}}$在 XX 末端探索软最大层几何 2。我们提供了在？?中比较这些结构的其他讨论，并提供了ELMES 作为浸人$\mathbb{R}^d+k$中的常规$d$ -单纯形的直观视图
## 4.2 LABEL SYMMETRY
将支持类分配给数字标签的对称性是元学习算法的重要属性。例如，如果我们有支持集类$\{$ 塔，熊，树$\},\{$的映射熊->1,塔->2,树->3 }应该为查询点产生与不同映射{ 熊->2, 塔->3树->1 }相同的预测。为了探索这种对称性，我们研究了模型如何使用类嵌入
根据我们在 Section 3 中的公式，我们将演示向量表示为嵌人$\rho$的图像和嵌入$\phi$的标签的串联$:[\rho\mid\phi]$。这个向量直接输人到自我注意机制中，在那里我们矩阵乘以键、查询和价值自我注意头。为简单起见，只采用这些矩阵中的一个，头维$k:$
$$[\rho\mid\phi]\begin{bmatrix}\dfrac{\Gamma_1}{\psi_1}&...&\Gamma_k\\\end{bmatrix}=[\langle\rho\:,\:\Gamma_1\rangle\quad...\quad\langle\rho\:,\:\Gamma_k\rangle]+[\langle\phi\:,\:\psi_1\rangle\quad...\quad\langle\phi\:,\:\psi_k\rangle]$$
此转换的输出将是两个向量的总和：一个由图像嵌人$\rho$和可学习$\{\Gamma_i\}_i=1^k$之间的内积组成，另一个由类嵌人$\phi$和可学习$\{\psi_i\}_i=1^k$组成。请注意，由于$\langle\rho,\Gamma_i\rangle$和$\langle\phi,\psi_i\rangle$之间的添加，Equation (1)意味着 CAML 对于将标签分配给支持集类并不是不变的；但是，我们可以将类嵌入$\left\{\phi\right\}_j=1^d$的几何形状限制为原则上相对于标签对称性。专门针对 $i\neq j\neq k$ 、 $\langle \phi _{i}$ , $\phi _{j}\rangle = \langle \phi _{i}$ , $\phi _{k}\rangle$ 和$\|\phi_i\|=\|\phi_{j}\|$
类似于卷积滤波器学习匹配图像中的模式，我们的分析假设可学习的$[ \psi _1$ ... $\psi _k]$将收敛到向量，该向量通过受某些约束的单个类嵌人来最大化内积。在此假设下，我们询问$d$类嵌人$\{\phi\}_{j=1}^d$的哪种几何形状允许可学习的$\psi_i$向量最明确地检测单个类嵌人。为了回答这个问题，我们为$d-$类集上的每个$\psi_i$定义了一个概率质量函数，以便最大化$j^th$类的概率与最大化$\left\langle\phi_j\:,\:\psi_i\right\rangle$并同样最小化 $k\neq j$ 的 $\left\langle\phi_k\:,\:\psi_i\right\rangle$一致。
$\begin{array}{rcl}\textbf{Definition 2. Let X be a discrete Random Variable taking on values in \{1,2,\ldots,d\}.}&For\\learnable&vector&\psi_{i}\:,~define~probability~mass~function~p_{\psi_{i}}(X&=&j)\: as~the~softmax~over\end{array}$
$[ \langle \phi _1$ , $\psi _i\rangle$ $\ldots$ $\langle \phi _d$ , $\psi _i\rangle ]$ $so$ $that:$
$$p_{\psi_i}(X=j)=\frac{e^{\|\psi_i\|\|\phi_j\|\cos(\theta_{i,j})}}{\sum_{k=1}^de^{\|\psi_i\|\|\phi_j\|\cos(\theta_{i,k})}}$$
where $\theta_{i,j}$ is the angle between $\phi_j$ and $\psi_i.$
我们说$\psi_i$学习检测类$j$,当$1\leq k\leq d$的$p_\psi_i(X=j)>p_{\psi_i}(X=k)$与$k\neq j$ 。通过将类嵌人分配给支持类的对称性，我们可以假设学习检测类$i$的$\psi_i$的数量与学习检测所有对$(i,j)$ 的类$j$的$\psi_j$的数量相似。我们还利用标签分配中的对称性来支持集合类，以做出以下假设。每个假设的理由位于？?中。
Assumption 1. Suppose $\{\psi_i\}_{i=1}^k$ are learnable class detectors of unit norm with at least one $\psi_i$
detecting each class $1\leq i\leq d$ . The probability $p_{\psi _j}( X= j) = p_{\psi _i}( X= i)$ for $1\leq i, j\leq d$ .
Assumption 2. Define $p_\psi_{i}(X\:=\:i)\backslash\{\phi_{l}\}_{l=(m+1)}^{d}$ as the probability of $\psi_i$ detecting $\phi_i$ from the setofvectors $\{ \phi _{j}\} _{j= 1}^{m}, m$ < $d$ . Then the probability $p_{\psi _{j}}( X$ = $j) \backslash \{ \phi _{l}\} _{l= ( m+ 1) }^{d}$ = $p_{\psi _{i}}( X$ = $i) \backslash \{ \phi _{l}\} _{l= ( m+ 1) }^{d}for$ $1\leq i, j\leq m$ and $m\geq 2$ .
Assumption 3. When $\psi _i= \frac {\phi _i}{\| \phi _i\| }$, $p_{\psi _i}( X= i)$ $is$ $maximized.$
当 XMATHX。
Theorem 1. The set of class embeddings $\{ \phi _j\} _{j= 1}^d$ $\forall j$ $, 1\leq j\leq d$ that maximizes $p_{\psi _j}( X= j)$ $is$
$necessarilyanELMES.$
4
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.176513671875,
0.1060791015625,
0.6953125,
0.1060791015625,
0.6953125,
0.11932373046875,
0.176513671875,
0.11932373046875
],
[
0.1748046875,
0.1319580078125,
0.82177734375,
0.1319580078125,
0.82177734375,
0.1458740234375,
0.1748046875,
0.1458740234375
],
[
0.1732177734375,
0.143310546875,
0.826171875,
0.141357421875,
0.826171875,
0.1622314453125,
0.1732177734375,
0.1641845703125
],
[
0.1732177734375,
0.1610107421875,
0.7900390625,
0.160400390625,
0.7900390625,
0.177978515625,
0.1732177734375,
0.1787109375
],
[
0.17724609375,
0.1912841796875,
0.8203125,
0.1912841796875,
0.8203125,
0.2052001953125,
0.17724609375,
0.2052001953125
],
[
0.1748046875,
0.2052001953125,
0.8203125,
0.2052001953125,
0.8203125,
0.2183837890625,
0.1748046875,
0.2183837890625
],
[
0.176513671875,
0.2171630859375,
0.82275390625,
0.2171630859375,
0.82275390625,
0.231689453125,
0.176513671875,
0.231689453125
],
[
0.176513671875,
0.231689453125,
0.36279296875,
0.231689453125,
0.36279296875,
0.24560546875,
0.176513671875,
0.24560546875
],
[
0.1748046875,
0.26318359375,
0.351318359375,
0.26318359375,
0.351318359375,
0.276611328125,
0.1748046875,
0.276611328125
],
[
0.176513671875,
0.2890625,
0.82177734375,
0.2890625,
0.82177734375,
0.302978515625,
0.176513671875,
0.302978515625
],
[
0.176513671875,
0.3037109375,
0.82275390625,
0.3037109375,
0.82275390625,
0.317626953125,
0.176513671875,
0.317626953125
],
[
0.176513671875,
0.317626953125,
0.72119140625,
0.317626953125,
0.72119140625,
0.330810546875,
0.176513671875,
0.330810546875
],
[
0.176513671875,
0.33837890625,
0.8203125,
0.33837890625,
0.8203125,
0.3515625,
0.176513671875,
0.3515625
],
[
0.176513671875,
0.3515625,
0.8203125,
0.3515625,
0.8203125,
0.365478515625,
0.176513671875,
0.365478515625
],
[
0.176513671875,
0.365478515625,
0.611328125,
0.365478515625,
0.611328125,
0.37890625,
0.176513671875,
0.37890625
],
[
0.2254638671875,
0.37890625,
0.83642578125,
0.37890625,
0.83642578125,
0.4248046875,
0.2254638671875,
0.4248046875
],
[
0.17724609375,
0.4248046875,
0.82177734375,
0.4248046875,
0.82177734375,
0.438232421875,
0.17724609375,
0.438232421875
],
[
0.17724609375,
0.43994140625,
0.82177734375,
0.43994140625,
0.82177734375,
0.453857421875,
0.17724609375,
0.453857421875
],
[
0.176513671875,
0.453857421875,
0.82275390625,
0.453857421875,
0.82275390625,
0.46728515625,
0.176513671875,
0.46728515625
],
[
0.176513671875,
0.466552734375,
0.826171875,
0.46728515625,
0.826171875,
0.481689453125,
0.176513671875,
0.48095703125
],
[
0.176513671875,
0.484130859375,
0.431396484375,
0.484130859375,
0.431396484375,
0.49755859375,
0.176513671875,
0.49755859375
],
[
0.17724609375,
0.5048828125,
0.71826171875,
0.5048828125,
0.71826171875,
0.51806640625,
0.17724609375,
0.51806640625
],
[
0.7294921875,
0.505859375,
0.8203125,
0.505859375,
0.8203125,
0.517578125,
0.7294921875,
0.517578125
],
[
0.176513671875,
0.51806640625,
0.82177734375,
0.51806640625,
0.82177734375,
0.5322265625,
0.176513671875,
0.5322265625
],
[
0.176513671875,
0.53271484375,
0.82177734375,
0.53271484375,
0.82177734375,
0.546875,
0.176513671875,
0.546875
],
[
0.17724609375,
0.5498046875,
0.82275390625,
0.5498046875,
0.82275390625,
0.56298828125,
0.17724609375,
0.56298828125
],
[
0.17724609375,
0.564453125,
0.55224609375,
0.564453125,
0.55224609375,
0.57763671875,
0.17724609375,
0.57763671875
],
[
0.176513671875,
0.58056640625,
0.82275390625,
0.58056640625,
0.82275390625,
0.5947265625,
0.176513671875,
0.5947265625
],
[
0.1748046875,
0.59228515625,
0.82177734375,
0.59326171875,
0.82177734375,
0.611328125,
0.1748046875,
0.60986328125
],
[
0.1748046875,
0.609375,
0.41015625,
0.60791015625,
0.41015625,
0.62255859375,
0.1748046875,
0.62353515625
],
[
0.494384765625,
0.630859375,
0.60546875,
0.630859375,
0.60546875,
0.6416015625,
0.494384765625,
0.6416015625
],
[
0.365234375,
0.64013671875,
0.471435546875,
0.64013671875,
0.471435546875,
0.65478515625,
0.365234375,
0.65478515625
],
[
0.4697265625,
0.65087890625,
0.62841796875,
0.6484375,
0.62841796875,
0.66357421875,
0.4697265625,
0.66552734375
],
[
0.1748046875,
0.67041015625,
0.459228515625,
0.671875,
0.459228515625,
0.68603515625,
0.1748046875,
0.68505859375
],
[
0.176513671875,
0.6962890625,
0.82177734375,
0.6962890625,
0.82177734375,
0.7109375,
0.176513671875,
0.7109375
],
[
0.176513671875,
0.708984375,
0.82275390625,
0.71044921875,
0.82275390625,
0.724609375,
0.176513671875,
0.7236328125
],
[
0.176513671875,
0.724609375,
0.82177734375,
0.724609375,
0.82177734375,
0.73779296875,
0.176513671875,
0.73779296875
],
[
0.176513671875,
0.73779296875,
0.373291015625,
0.7373046875,
0.373291015625,
0.75048828125,
0.176513671875,
0.751953125
],
[
0.176513671875,
0.755859375,
0.82177734375,
0.7568359375,
0.82177734375,
0.771484375,
0.176513671875,
0.77099609375
],
[
0.176513671875,
0.77099609375,
0.7900390625,
0.77099609375,
0.7900390625,
0.7841796875,
0.176513671875,
0.7841796875
],
[
0.17236328125,
0.787109375,
0.82421875,
0.78857421875,
0.82421875,
0.8095703125,
0.17236328125,
0.80810546875
],
[
0.1732177734375,
0.80615234375,
0.82421875,
0.80810546875,
0.82421875,
0.8271484375,
0.1732177734375,
0.82470703125
],
[
0.17236328125,
0.8271484375,
0.48046875,
0.82470703125,
0.48046875,
0.84228515625,
0.17236328125,
0.8447265625
],
[
0.1707763671875,
0.8466796875,
0.57666015625,
0.8466796875,
0.57666015625,
0.87060546875,
0.1707763671875,
0.87060546875
],
[
0.176513671875,
0.87744140625,
0.271240234375,
0.87744140625,
0.271240234375,
0.892578125,
0.176513671875,
0.892578125
],
[
0.1748046875,
0.89453125,
0.82275390625,
0.896484375,
0.82275390625,
0.91357421875,
0.1748046875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.326904296875,
0.9111328125,
0.326904296875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.49267578125,
0.94775390625,
0.50732421875,
0.9521484375,
0.50244140625,
0.96337890625,
0.48779296875,
0.958984375
]
] |
[
"Published as a conference paper at ICLR 2024",
"4.1EQUALLENGTHANDMAXIMALLYEQUIANGULAR SETOFVECTORS",
"Definition 1. An Equal Length and Maximally Equiangular Set ( ELMES) is a set of non-zero",
"vectors (b==1, Φ e IRd+ for some k ≥ 0 and d > i, such that Vj + j' , llpjll ='lpgill and",
"(oj, Φj)=. Simply, all vectors in this set are equal length and maximally equiangular.",
"向量的等角集和最大等角集（ELMES）与表示理论中的等角紧顿E（Welch，1974;Fickusetal.",
"2018）以及最近的神经崩溃作品中突出显示的单纯形等角紧顿E在XX末端探索软最大层几",
"何2。我们提供了在？？中比较这些结构的其他讨论，并提供了ELMES作为浸入Rd+k中的",
"常规d-单纯形的直观视图",
"4.2LABELSYMMETRY",
"将支持类分配给数字标签的对称性是元学习算法的重要属性。例如，如果我们有支持集类（",
"塔，熊，树】，【的映射熊->1，塔->2，树->3}应该为查询点产生与不同映射【熊->2，塔->3",
"树->1相同的预测。为了探索这种对称性，我们研究了模型如何使用类嵌入",
"根据我们在Section3中的公式，我们将演示向量表示为嵌的图像和嵌入Φ的标签的串",
"联：[βΦ]。这个向量直接输人到自我注意机制中，在那里我们矩阵乘以键、查询和价值自",
"我注意头。为简单起见，只采用这些矩阵中的一个，头维k：",
"el[-1o,）(o Ta)+[e )",
"此转换的输出将是两个向量的总和：一个由图像嵌人β和可学习{}1之间的内积组",
"成，另一个由类嵌人Φ和可学习（=i组成。请注意，由于《p，I）和（Φ，i）之间的",
"添加，Equation（1）意味着CAML对于将标签分配给支持集类并不是不变的；但是，我们",
"可以将类嵌入【Φ)-的几何形状限制为原则上相对于标签对称性。专门针对i≠≠k、",
"(i，)=Φi，Φk)和llpill=Φ;ll",
"类似于卷积滤波器学习匹配图像中的模式，我们的分析假设可学习的【1",
"将收敛",
"到向量，该向量通过受某些约束的单个类嵌人来最大化内积。在此假设下，我们询问d类嵌",
"人【}-1的哪种几何形状允许可学习的,向量最明确地检测单个类嵌人。为了回答这个",
"问题，我们为d一类集上的每个定义了一个概率质量函数，以便最大化ith类的概率与最",
"大化《，）并同样最小化的（，）一致。",
"Definition 2. Let X be a discrete Random Variable taking on values in {1,2, ...,d) : For",
"learnable vector i , define probability mass function py:(X = j) as the softmax over",
"[oi, )..(od, i)] so that:",
"ell/:lllllcos(s)",
"pb(X =j)=",
"Zkellba lllsllcos(ei.k)",
"where Oi,s is the angle between pj and bi:",
"我们说学习检测类j，当1≤d的p(X=）>p(X=）与≠。通过将类嵌",
"人分配给支持类的对称性，我们可以假设学习检测类i的的数量与学习检测所有对（i，）",
"的类i的，的数量相似。我们还利用标签分配中的对称性来支持集合类，以做出以下假设。",
"每个假设的理由位于？？中。",
"Assumption 1. Suppose (bi)-1 are learnable class detectors of unit norm with at least one i",
"detecting each class 1≤i≤ d. The probability Pu, (X =)=pw(X =) for1 ≤ i, j≤ d.",
"Assumption 2. Define P:(X = z)(oii-(m+1) as the probability of wb: detecting i from the",
"set of vectors [Φ;)-1, m < d . Then the probability P,(X = j))(Φi)i=(m+1) = pb(X =",
"i)\\(oili=(m+1)for1≤i,j≤mandm ≥2.",
"Assumption 3. When b: =, P (X = i) s maximized.",
"当XMATHX。",
"Theorem 1. The set of class embeddings [Φ;]g-, Vj, 1 ≤ j ≤ d that maximizes pw, (X = j) is",
"necessarilyanELMES",
"4"
] |
{
"arxiv_id": "2310.10971",
"language": "zh",
"page_idx": 3
}
|
||
2310.10971-zh-4
|
2310.10971-zh-4.jpg
|
Published as a conference paper at ICLR 2024
Table l: MiniImageNet & CIFAR-fs mean accuracy and standard error across 10,000 test epochs. t
indicates the pre-trained image encoder backbone was frozen during training.
<table>
<tbody>
<tr>
<th rowspan="2">Method ( Backbone) </th>
<th colspan="2">CIFAR-fs</th>
<th colspan="2">MiniImageNet</th>
</tr>
<tr>
<th>$5w-1s$</th>
<th>$5w-5s$</th>
<th>$5w-1s$</th>
<th>$5w-5s$</th>
</tr>
<tr>
<td>In-Domain[Meta-Training] P>M>F Hu et al. (2022)</td>
<td>84.3</td>
<td>92.2</td>
<td>95.3</td>
<td>98.4</td>
</tr>
<tr>
<td>Universal Meta-Learning;</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>No Meta- Training or Finetuning</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>ProtoNet (Snell et al., 2017)</td>
<td>$62.9\pm.2$</td>
<td>$79.7\pm.2$</td>
<td>$92.1\pm.1$</td>
<td>$97.1\pm.0$</td>
</tr>
<tr>
<td>ProtoNet . 十</td>
<td>$57.7\pm.2$</td>
<td>$81.0\pm.2$</td>
<td>$85.3\pm.2$</td>
<td>$96.0\pm.1$</td>
</tr>
<tr>
<td>MetaOpt ( Lee et al. , 2019) </td>
<td>$53.1\pm.3$</td>
<td>$73.1\pm.2$</td>
<td>$78.5\pm.2$</td>
<td>$91.6\pm.1$</td>
</tr>
<tr>
<td>MetaOpt 1</td>
<td>$61.7\pm.2$</td>
<td>$83.1\pm.1$</td>
<td>$86.9\pm.2$</td>
<td>$96.5\pm.1$</td>
</tr>
<tr>
<td>MetaODA (Zhang et al., 2021)</td>
<td>$60.4\pm.2$</td>
<td>$83.2\pm.1$</td>
<td>$88.2\pm.2$</td>
<td>$97.4\pm.0$</td>
</tr>
<tr>
<td>(Kirsch et al., 2022) GPICL 1</td>
<td>$41.5\pm.4$</td>
<td>$78.3\pm.2$</td>
<td>$95.6\pm.1$</td>
<td>$98.2\pm.1$</td>
</tr>
<tr>
<td>SNAIL $\angle($Mishra et al.,2017)</td>
<td>$62.1\pm.3$</td>
<td>$71.1\pm.3$</td>
<td>$93.6\pm.1$</td>
<td>$98.1\pm.0$</td>
</tr>
<tr>
<td>CAML</td>
<td>$70.8\pm.2$</td>
<td>$85.5\pm.1$</td>
<td>$96.2\pm.1$</td>
<td>$98.6\pm.0$</td>
</tr>
</tbody>
</table>
或者，从信息论的角度来看，我们可以将 ELMES 解释为使$\psi_i$检测类$i$的熵最小化的类嵌
人。非正式地，ELMES 使$\psi_i$在检测类$i$时具有最小的不确定性。
Proposition 1. Let $H_{\psi_{i}}(X)$ be the entropy $ofp_{\psi_{i}}(X).AnELMES$ minimizes $H_\psi_{i}(X).$
4.3 PERMUTATION INVARIANCE.
除了标签对称之外，还希望 CAML 的输出预测不依赖于序列中演示的顺序。例如，如果我们有支持集类$\left\{\text{塔，熊，树}\right\}$,序列$\left\{\text{(熊->1),(塔->2),(树->3)}\right\}$应该产生与置换序列$\left\{\text{们有支持集类}\left\{\text{塔，熊，树}\right\},\text{序列}\left\{\text{(熊->1),(塔->2),(树->3)}\right\}\text{应该产生与置换序列}\left\{\text{个有支持集类}\right\},\text{ 模 }\left\{\text{塔->1),( 塔->2),(树->3)}\right\}\text{应该产生与置换序列}\left\{\:\right\},\text{ 应该产生与置换序列}\left\{\:\right\},\:\left\{\:\:\:\:\:\:\:\:\:\:\:\:\:\:\text{}\right\}\right\}$ (树->3),(熊->1),(塔 THX->2)MAX。在 Kossen et al.(2021); Fifty et al.(2023)的先前工作的基础上，足以证明 ELMES 标签编码器对于输入序列中的排列是等变的，以表明 CAML 对于排列是不变的。
Proposition 2. Consider an n -sequence of one-hot labels stacked into a matrix $\mathcal{S}\in\mathbb{R}^n\times w$ ,and an $\dot{E}LMES$ label encoder denoted by $W\in\mathbb{R}^{w\times d}$ with w denoting “way”and d the dimension of the label embedding. The label embedding SW is equivariant to permutations.
# 5 EXPERIMENTS
为了量化通用图像分类性能，我们评估了一组不同的 11 元学习基准，这些基准分为 4 个不
同的类别：
1. Generic Object Recognition: mini-ImageNet (Vinyals et al.,2016),tiered-ImageNet (Ren et al.,
2018), CIFAR-fs (Bertinetto et al.,2018), and Pascal VOC (Everingham et al.)
2. Fine-Grained Image Classification: CUB (Wah et al.,2011),Aircraft (Maji et al.,2013) ,
meta-iNat (Wertheimer& Hariharan, 2019), and tiered meta-iNat (Wertheimer& Hariharan,
2019)
3. Unnatural Image Classification: ChestX (Guo et al., 2020) and Paintings (Crowley&Zisserman,
2015)
4. Inter-Domain Image Classification: Pascal+Paintings (Everingham et al.; Crowley& Zisserman,
2015).
通用对象识别，细粒度图像分类和非自然图像分类是 XMATHX-。除此之外，我们通过将PascalVOC 与绘画相结合来构成一个具有挑战性的新的域间类别，以便每个类都由自然图像和绘画组成。这使我们能够评估元学习算法在同一类中跨领域进行泛化的能力。例如，“塔” 类的支持图像可能是梵高的“星夜”,而查询可能是埃菲尔铁塔的图片。人类有能力在这些领域之间概括视觉概念；然而，元学习算法在这个公式上挣扎 (Jankowski&Grąbczewski, $\overline{20}11)$。
## 5.1 BASELINES
我们评估 CAML、原型网络 (ProtoNet) (Snell et al.,2017)、MetaOpt (Lee et al.,2019) 、 MetaQDA (Zhang et al.,2021)、SNAIL (Mishra et al.,2017)和 GPICL XMATHX (Kirsch et al.,
5
|
[
[
0.1748046875,
0.034088134765625,
0.4794921875,
0.034088134765625,
0.4794921875,
0.04736328125,
0.1748046875,
0.04736328125
],
[
0.1748046875,
0.0706787109375,
0.82421875,
0.0706787109375,
0.82421875,
0.083984375,
0.1748046875,
0.083984375
],
[
0.1732177734375,
0.08331298828125,
0.6787109375,
0.083984375,
0.6787109375,
0.0985107421875,
0.1732177734375,
0.09783935546875
],
[
0.498291015625,
0.10797119140625,
0.56396484375,
0.10797119140625,
0.56396484375,
0.12249755859375,
0.498291015625,
0.12249755859375
],
[
0.2467041015625,
0.1142578125,
0.374267578125,
0.1142578125,
0.374267578125,
0.1275634765625,
0.2467041015625,
0.1275634765625
],
[
0.64453125,
0.10919189453125,
0.736328125,
0.10919189453125,
0.736328125,
0.12060546875,
0.64453125,
0.12060546875
],
[
0.468017578125,
0.126953125,
0.51123046875,
0.126953125,
0.51123046875,
0.1427001953125,
0.468017578125,
0.1427001953125
],
[
0.54833984375,
0.126953125,
0.59326171875,
0.126953125,
0.59326171875,
0.1427001953125,
0.54833984375,
0.1427001953125
],
[
0.62841796875,
0.126953125,
0.67236328125,
0.126953125,
0.67236328125,
0.1427001953125,
0.62841796875,
0.1427001953125
],
[
0.708984375,
0.126953125,
0.75341796875,
0.126953125,
0.75341796875,
0.1427001953125,
0.708984375,
0.1427001953125
],
[
0.236083984375,
0.1458740234375,
0.410888671875,
0.147705078125,
0.410888671875,
0.1622314453125,
0.236083984375,
0.160400390625
],
[
0.473876953125,
0.160400390625,
0.50732421875,
0.160400390625,
0.50732421875,
0.1767578125,
0.473876953125,
0.1767578125
],
[
0.5537109375,
0.160400390625,
0.58837890625,
0.160400390625,
0.58837890625,
0.1767578125,
0.5537109375,
0.1767578125
],
[
0.6337890625,
0.160400390625,
0.66845703125,
0.160400390625,
0.66845703125,
0.1767578125,
0.6337890625,
0.1767578125
],
[
0.71337890625,
0.160400390625,
0.74853515625,
0.160400390625,
0.74853515625,
0.1767578125,
0.71337890625,
0.1767578125
],
[
0.236083984375,
0.179931640625,
0.400390625,
0.18115234375,
0.400390625,
0.1956787109375,
0.236083984375,
0.1944580078125
],
[
0.238525390625,
0.1634521484375,
0.3896484375,
0.1634521484375,
0.3896484375,
0.173583984375,
0.238525390625,
0.173583984375
],
[
0.236083984375,
0.1932373046875,
0.439453125,
0.1956787109375,
0.439453125,
0.210205078125,
0.236083984375,
0.2083740234375
],
[
0.4599609375,
0.2083740234375,
0.51953125,
0.2083740234375,
0.51953125,
0.222900390625,
0.4599609375,
0.222900390625
],
[
0.54150390625,
0.2083740234375,
0.60107421875,
0.2083740234375,
0.60107421875,
0.222900390625,
0.54150390625,
0.222900390625
],
[
0.62109375,
0.2083740234375,
0.6806640625,
0.2083740234375,
0.6806640625,
0.222900390625,
0.62109375,
0.222900390625
],
[
0.701171875,
0.2083740234375,
0.76123046875,
0.2083740234375,
0.76123046875,
0.222900390625,
0.701171875,
0.222900390625
],
[
0.5400390625,
0.2197265625,
0.60107421875,
0.2215576171875,
0.6005859375,
0.2374267578125,
0.5390625,
0.23486328125
],
[
0.236083984375,
0.208984375,
0.414306640625,
0.208984375,
0.414306640625,
0.222900390625,
0.236083984375,
0.222900390625
],
[
0.23779296875,
0.2235107421875,
0.3056640625,
0.2235107421875,
0.3056640625,
0.23486328125,
0.23779296875,
0.23486328125
],
[
0.4599609375,
0.2215576171875,
0.521484375,
0.2215576171875,
0.521484375,
0.236083984375,
0.4599609375,
0.236083984375
],
[
0.62158203125,
0.2235107421875,
0.6806640625,
0.2235107421875,
0.6806640625,
0.23486328125,
0.62158203125,
0.23486328125
],
[
0.70166015625,
0.2235107421875,
0.7587890625,
0.2235107421875,
0.7587890625,
0.23486328125,
0.70166015625,
0.23486328125
],
[
0.236083984375,
0.236083984375,
0.404296875,
0.236083984375,
0.404296875,
0.250732421875,
0.236083984375,
0.250732421875
],
[
0.4599609375,
0.2374267578125,
0.51953125,
0.2374267578125,
0.51953125,
0.248779296875,
0.4599609375,
0.248779296875
],
[
0.54150390625,
0.2374267578125,
0.6005859375,
0.2374267578125,
0.6005859375,
0.248779296875,
0.54150390625,
0.248779296875
],
[
0.62158203125,
0.2374267578125,
0.6787109375,
0.2374267578125,
0.6787109375,
0.248779296875,
0.62158203125,
0.248779296875
],
[
0.70166015625,
0.2374267578125,
0.7587890625,
0.2374267578125,
0.7587890625,
0.248779296875,
0.70166015625,
0.248779296875
],
[
0.238525390625,
0.251953125,
0.30224609375,
0.251953125,
0.30224609375,
0.26318359375,
0.238525390625,
0.26318359375
],
[
0.461669921875,
0.251953125,
0.51953125,
0.251953125,
0.51953125,
0.26318359375,
0.461669921875,
0.26318359375
],
[
0.54150390625,
0.251953125,
0.5986328125,
0.251953125,
0.5986328125,
0.26318359375,
0.54150390625,
0.26318359375
],
[
0.62158203125,
0.251953125,
0.6787109375,
0.251953125,
0.6787109375,
0.26318359375,
0.62158203125,
0.26318359375
],
[
0.70166015625,
0.251953125,
0.7587890625,
0.251953125,
0.7587890625,
0.26318359375,
0.70166015625,
0.26318359375
],
[
0.236083984375,
0.26318359375,
0.4306640625,
0.26318359375,
0.4306640625,
0.276611328125,
0.236083984375,
0.276611328125
],
[
0.54150390625,
0.264404296875,
0.5986328125,
0.264404296875,
0.5986328125,
0.27587890625,
0.54150390625,
0.27587890625
],
[
0.70166015625,
0.264404296875,
0.7607421875,
0.264404296875,
0.7607421875,
0.27587890625,
0.70166015625,
0.27587890625
],
[
0.62158203125,
0.26513671875,
0.6806640625,
0.26513671875,
0.6806640625,
0.27587890625,
0.62158203125,
0.27587890625
],
[
0.461669921875,
0.26513671875,
0.51953125,
0.26513671875,
0.51953125,
0.276611328125,
0.461669921875,
0.276611328125
],
[
0.4599609375,
0.276611328125,
0.521484375,
0.276611328125,
0.521484375,
0.291748046875,
0.4599609375,
0.291748046875
],
[
0.5400390625,
0.276611328125,
0.60107421875,
0.276611328125,
0.60107421875,
0.291748046875,
0.5400390625,
0.291748046875
],
[
0.62109375,
0.276611328125,
0.6806640625,
0.276611328125,
0.6806640625,
0.291748046875,
0.62109375,
0.291748046875
],
[
0.701171875,
0.276611328125,
0.7587890625,
0.276611328125,
0.7587890625,
0.291748046875,
0.701171875,
0.291748046875
],
[
0.23779296875,
0.290283203125,
0.414306640625,
0.290283203125,
0.414306640625,
0.3037109375,
0.23779296875,
0.3037109375
],
[
0.238525390625,
0.27783203125,
0.41259765625,
0.27783203125,
0.41259765625,
0.291748046875,
0.238525390625,
0.291748046875
],
[
0.4599609375,
0.290283203125,
0.521484375,
0.290283203125,
0.521484375,
0.304931640625,
0.4599609375,
0.304931640625
],
[
0.5400390625,
0.290283203125,
0.60107421875,
0.290283203125,
0.60107421875,
0.304931640625,
0.5400390625,
0.304931640625
],
[
0.619140625,
0.290283203125,
0.6806640625,
0.290283203125,
0.6806640625,
0.304931640625,
0.619140625,
0.304931640625
],
[
0.70166015625,
0.291748046875,
0.7607421875,
0.291748046875,
0.7607421875,
0.302978515625,
0.70166015625,
0.302978515625
],
[
0.45751953125,
0.302978515625,
0.52294921875,
0.304931640625,
0.52294921875,
0.3193359375,
0.45751953125,
0.317626953125
],
[
0.619140625,
0.304931640625,
0.68115234375,
0.304931640625,
0.68115234375,
0.31640625,
0.619140625,
0.31640625
],
[
0.5390625,
0.30615234375,
0.6005859375,
0.30615234375,
0.6005859375,
0.317626953125,
0.5390625,
0.317626953125
],
[
0.23779296875,
0.30615234375,
0.28515625,
0.30615234375,
0.28515625,
0.31640625,
0.23779296875,
0.31640625
],
[
0.69921875,
0.30615234375,
0.76123046875,
0.30615234375,
0.76123046875,
0.317626953125,
0.69921875,
0.317626953125
],
[
0.176513671875,
0.34521484375,
0.82275390625,
0.34521484375,
0.82275390625,
0.359130859375,
0.176513671875,
0.359130859375
],
[
0.17724609375,
0.359130859375,
0.63623046875,
0.359130859375,
0.63623046875,
0.37255859375,
0.17724609375,
0.37255859375
],
[
0.1732177734375,
0.375,
0.7509765625,
0.375732421875,
0.7509765625,
0.39013671875,
0.1732177734375,
0.389404296875
],
[
0.1748046875,
0.4072265625,
0.2099609375,
0.40478515625,
0.2115478515625,
0.417236328125,
0.176513671875,
0.419189453125
],
[
0.2156982421875,
0.4072265625,
0.418212890625,
0.4072265625,
0.418212890625,
0.42041015625,
0.2156982421875,
0.42041015625
],
[
0.1748046875,
0.431884765625,
0.82177734375,
0.431884765625,
0.82177734375,
0.4462890625,
0.1748046875,
0.4462890625
],
[
0.17724609375,
0.4462890625,
0.82177734375,
0.4462890625,
0.82177734375,
0.460205078125,
0.17724609375,
0.460205078125
],
[
0.176513671875,
0.460205078125,
0.82275390625,
0.460205078125,
0.82275390625,
0.473388671875,
0.176513671875,
0.473388671875
],
[
0.1748046875,
0.473388671875,
0.82177734375,
0.472900390625,
0.82177734375,
0.4873046875,
0.1748046875,
0.488037109375
],
[
0.17724609375,
0.490478515625,
0.314453125,
0.490478515625,
0.314453125,
0.50048828125,
0.17724609375,
0.50048828125
],
[
0.1748046875,
0.5048828125,
0.82421875,
0.5048828125,
0.82421875,
0.51806640625,
0.1748046875,
0.51806640625
],
[
0.1732177734375,
0.5166015625,
0.82666015625,
0.5166015625,
0.82666015625,
0.5341796875,
0.1732177734375,
0.5341796875
],
[
0.17236328125,
0.5322265625,
0.69140625,
0.53271484375,
0.69140625,
0.54736328125,
0.17236328125,
0.546875
],
[
0.176513671875,
0.568359375,
0.1895751953125,
0.568359375,
0.1895751953125,
0.580078125,
0.176513671875,
0.580078125
],
[
0.2091064453125,
0.5673828125,
0.33251953125,
0.5673828125,
0.33251953125,
0.58056640625,
0.2091064453125,
0.58056640625
],
[
0.1748046875,
0.59521484375,
0.82275390625,
0.5966796875,
0.82275390625,
0.611328125,
0.1748046875,
0.60986328125
],
[
0.176513671875,
0.6123046875,
0.2467041015625,
0.6123046875,
0.2467041015625,
0.62353515625,
0.176513671875,
0.62353515625
],
[
0.1748046875,
0.63623046875,
0.82421875,
0.63623046875,
0.82421875,
0.6494140625,
0.1748046875,
0.6494140625
],
[
0.1968994140625,
0.6494140625,
0.7255859375,
0.6494140625,
0.7255859375,
0.66357421875,
0.1968994140625,
0.66357421875
],
[
0.1748046875,
0.66357421875,
0.82275390625,
0.66357421875,
0.82275390625,
0.67822265625,
0.1748046875,
0.67822265625
],
[
0.1953125,
0.67822265625,
0.82275390625,
0.67822265625,
0.82275390625,
0.69140625,
0.1953125,
0.69140625
],
[
0.1968994140625,
0.6923828125,
0.233642578125,
0.6923828125,
0.233642578125,
0.70263671875,
0.1968994140625,
0.70263671875
],
[
0.17236328125,
0.7041015625,
0.82275390625,
0.70556640625,
0.82275390625,
0.7197265625,
0.17236328125,
0.71728515625
],
[
0.1968994140625,
0.72021484375,
0.2344970703125,
0.72021484375,
0.2344970703125,
0.73193359375,
0.1968994140625,
0.73193359375
],
[
0.1732177734375,
0.73291015625,
0.82275390625,
0.73291015625,
0.82275390625,
0.7470703125,
0.1732177734375,
0.7470703125
],
[
0.1953125,
0.7470703125,
0.24267578125,
0.7470703125,
0.24267578125,
0.76123046875,
0.1953125,
0.76123046875
],
[
0.176513671875,
0.771484375,
0.82177734375,
0.771484375,
0.82177734375,
0.78515625,
0.176513671875,
0.78515625
],
[
0.1732177734375,
0.7841796875,
0.8203125,
0.7841796875,
0.8203125,
0.79736328125,
0.1732177734375,
0.79736328125
],
[
0.1748046875,
0.79736328125,
0.82177734375,
0.798828125,
0.82177734375,
0.81298828125,
0.1748046875,
0.8125
],
[
0.176513671875,
0.81298828125,
0.82177734375,
0.81298828125,
0.82177734375,
0.8271484375,
0.176513671875,
0.8271484375
],
[
0.1748046875,
0.82470703125,
0.82421875,
0.8271484375,
0.82421875,
0.84130859375,
0.1748046875,
0.8388671875
],
[
0.1748046875,
0.84033203125,
0.2254638671875,
0.84228515625,
0.2254638671875,
0.85400390625,
0.1748046875,
0.8515625
],
[
0.1732177734375,
0.87060546875,
0.221435546875,
0.87060546875,
0.221435546875,
0.88525390625,
0.1732177734375,
0.88525390625
],
[
0.2156982421875,
0.87255859375,
0.29736328125,
0.87255859375,
0.29736328125,
0.8837890625,
0.2156982421875,
0.8837890625
],
[
0.176513671875,
0.896484375,
0.81787109375,
0.896484375,
0.81787109375,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.1748046875,
0.9111328125,
0.82421875,
0.9111328125,
0.82421875,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2024",
"Table 1: MinilmageNet & CIFAR-fs mean accuracy and standard error across 10,000 test epochs. +",
"indicates the pre-trained image encoder backbone was frozen during training.",
"CIFAR-fs",
"Method (Backbone)",
"MinilmageNet",
"5w-1s",
"5w-5s",
"5w-1s",
"5w-5s",
"In-Domain [Meta-Training]",
"84.3",
"92.2",
"95.3",
"98.4",
"Universal Meta-Learning;",
"P>M>FHuetal.(2022)",
"No Meta-Training or Finetuning",
"62.9 ±.2",
"79.7 ±.2",
"92.1 ±.1",
"97.1 ±.0",
"81.0 ±.2",
"ProtoNet (Snell et al., 2017)",
"ProtoNet t",
"57.7±.2",
"85.3 ±.2",
"96.0 ±.1",
"MetaOpt (Lee et al., 2019)",
"53.1±.3",
"73.1 ±.2",
"78.5 ±.2",
"91.6 ±.1",
"MetaOpt",
"61.7 ±.2",
"83.1 ±.1",
"86.9±.2",
"96.5 ±.1",
"MetaQDA (Zhang etal,2021)",
"83.2 ±.1",
"97.4 ±.0",
"88.2±.2",
"60.4±.2",
"41.5 ±.4",
"78.3 ±.2",
"95.6 ±.1",
"98.2 ±.1",
"SNAIL (Mishra et al., 2017)",
"GPICL (Kirsch et al., 2022)",
"62.1 ±.3",
"71.1 ±.3",
"93.6 ±.1",
"98.1 ±.0",
"70.8±.2",
"96.2±.1",
"85.5 ±.1",
"CAML",
"98.6±.0",
"或者，从信息论的角度来看，我们可以将ELMES解释为使中；检测类i的摘最小化的类嵌",
"人。非正式地，ELMES使在检测类i时具有最小的不确定性。",
"Proposition 1. Let Hu, (X) be the entropy of pu, (X).An ELMES minimizes H,(X).",
"4.3",
"PERMUTATIONINVARIANCE.",
"除了标签对称之外，还希望CAML的输出预测不依赖于序列中演示的顺序。例如，如果我",
"们有支持集类【塔，熊，树}，序列（（熊->1），（塔->2)，（树->3）}应该产生与置换序列（",
"（树->3），（熊->1），（塔THX->2）MAX。在Kossenetal.（2021);Fiftyetal.（2023）的先前工",
"作的基础上，足以证明ELMES标签编码器对于输人序列中的排列是等变的，以表明CAMIL",
"对于排列是不变的。",
"Proposition 2. Consider an n -sequence of one-hot labels stacked into a matrix S e Rnxw, and",
"an ELMES label encoder denoted by W IRwxd with w denoting \"way\" and d the dimension of",
"the label embedding. The label embedding SW is equivariant to permutations.",
"5",
"EXPERIMENTS",
"为了量化通用图像分类性能，我们评估了一组不同的11元学习基准，这些基准分为4个不",
"同的类别：",
"1. Generic Object Recognition: mini-ImageNet (Vinyals et al., 2016), tiered-ImageNet (Ren et al.,",
"2018),CIFAR-fs (Bertinettoet al.,2018),and Pascal VOC(Everinghamet al.)",
"2. Fine-Grained Image Classification: CUB (Wah et al., 201l) , Aircraft (Maji et al., 2013),",
"meta-iNat (Wertheimer & Hariharan, 2019),and tiered meta-iNat (Wertheimer & Hariharan,",
"2019)",
"3.Unnatural ImageClassification:ChestX(Guoetal.,2020)andPaintings(Crowley&Zisserman",
"2015)",
"4.Inter-Domain Image Classification:Pascal+Paintings (Everingham etal.;Crowley &Zisserman,",
"2015).",
"通用对象识别，细粒度图像分类和非自然图像分类是XMATHX-。除此之外，我们通过将",
"PascalVOC与绘画相结合来构成一个具有挑战性的新的域间类别，以便每个类都由自然图像",
"和绘画组成。这使我们能够评估元学习算法在同一类中跨领域进行泛化的能力。例如，“塔”",
"类的支持图像可能是梵高的“星夜”，而查询可能是埃菲尔铁塔的图片。人类有能力在这",
"些领域之间概括视觉概念；然而，元学习算法在这个公式上挣扎（Jankowski&Grabczewski",
"2011)。",
"5.1",
"BASELINES",
"我们评估CAML、原型网络（ProtoNet）（Snelletal.，2017）、MetaOpt（Leeetal.，2019）",
"MetaQDA(Zhanget al.,2021)、SNAIL(Mishra et al.,2017)和GPICLXMATHX(Kirsch et al.,",
"5"
] |
{
"arxiv_id": "2310.10971",
"language": "zh",
"page_idx": 4
}
|
||
1811.12833-zh-1
|
1811.12833-zh-1.jpg
|
器，该鉴别器从分割网络中获取特征图并尝试预测输人域。分段网络试图欺骗鉴别器，从而使来自两个域的特征具有相似的分布。Hoffman et al.[15] 是第一个将对抗方法应用于 UDA 的语义分割。通过从源域传输标签统计信息，它们还具有特定于类别的适应性。在[5]中使用了类似的全局和类对齐方法，其中类对齐是通过对网格软伪标签进行对抗性训练来完成的。在[4] 中，对抗性训练用于空间感知自适应以及蒸馏损失，以专门解决合成 2 实域偏移。[16] 使用残差网使源特征图类似于使用对抗性训练的目标特征图，然后将特征图用于分割任务。在[41] 中，对抗性方法用于输出空间，以受益于跨域的结构一致性。[32,33]提出了另一种使用对抗性训练的有趣方法：他们在目标域图像上获得了两个预测，这是由两个分类器[33]或在分类器[32]中使用 dropout 完成的。给定两个预测，分类器被训练以最大化分布之间的差异，而网络的特征提取器部分被训练以最小化它
一些方法建立在生成网络上，以生成以源为条件的目标图像。Hoffman et al. [14]提出了周期一致的对抗域自适应 (CyCADA),其中它们在像素级和特征级表示上都进行了自适应。对于像素级自适应，他们使用Cycle-GAN [48] 生成以源图像为条件的目标图像。在[34] 中，学习生成模型以从特征空间重建图像。然后， 对于域自适应，使用生成器模块训练特征模块以在源特征上产生目标图像，反之亦然。在 DCAN [43] 中， 生成器和分段网络中使用了通道式特征对齐。在生成的图像上学习分割网络，其中源的内容和目标的样式将源分割图用作地面实况。[50]中的作者使用生成对抗网络 (GAN)E[11]来对齐源和目标嵌人。此外，它们用保守损失(CL)代替交叉熵损失，从而惩罚源示例的简单和困难情况。CL 方法与大多数 UDA 方法正交， 包括我们的方法：它可以使任何对源使用交叉熵的方法受益
UDA 的另一种方法是自我训练。这个想法是使用来自集成模型或模型的先前状态的预测作为未标记数据的伪标签来训练当前模型。许多半监督方法[20,39]使用自训练。在[51]中，UDA 在语义分割上采用了自训练，并进一步扩展了类平衡和空间先验。正如我们在2.1节中讨论的那样，自我训练与建议的熵最小化方法有有趣的联系
在其他一些方法中，[26]通过多次损失使用对抗性和生成技术的组合，[46]结合了用于外观适应的生成方法和用于表示适应的对抗性训练，[45]提出了一种课程风格的 UDA 学习通过在局部 (超像素级)和全局标签分布上实施一致性
熵最小化已被证明可用于半监督学习[12,38]和聚类$\left[17,18\right]$。它最近也被应用于分类任务 $\left[25\right]$的域自适应。据我们所知，我们首先成功地应用基于熵的 UDA 训练来获得语义分割任务的竞争性能
# 2. Approaches
在本节中，我们提出了两种建议的熵最小化方法，使用 (i) 无监督熵损失和 (ii)对抗训练。为了构建我们
的模型，我们从现有的语义分割框架开始，并添加一个用于域适应的额外网络分支。图 2 说明了我们的体系结构
我们的模型在源域上进行了有监督损失的训练。形式上，我们考虑一组$\mathcal{X}_s\subset\mathbb{R}^H\times W\times3$源示例以及相关的地面实况$C$类分割图$\mathcal{Y}_s\subset(1,C)^H\times W$。示例$x_s$ 是$H\times W$彩色图像，关联映射$y_s$的条目$y_s^{(h,w)}=$ $\begin{bmatrix}\boldsymbol{y}_{\mathrm{s}}^{(h,w,c)}\end{bmatrix}_{c}$提供了像素 $(h,w)$ 的标签作为独热矢量。令$\bar{F}$是一个语义分割网络，它获取图像$x$并预测$C$维的“软分割图”$F(\boldsymbol x)=\boldsymbol{P}_x=\begin{bmatrix}\boldsymbol{P}_x^{(h,w,c)}\end{bmatrix}_{h,w,c}$。凭借最终的 softmax 层，每个$C$维像素矢量$\left[\boldsymbol P_{\boldsymbol{x}}^{(h,w,c)}\right]_{c}$ 表现为在类上的离散分布。如果一个类脱颖而出，芬布是挑剔的(低熵),如果分数均匀分布，从网络的角度来看，不确定性的迹象，熵很大。学习$F$的参数$\theta_F$以最小化源样本上的分割损失$\mathcal{L}_seg(x_s,y_s)=$ $-\sum_{h=1}^{H}\sum_{w=1}^{W}\sum_{c=1}^{C}y_{s}^{(h,w,c)}\log\boldsymbol{P}_{\boldsymbol{x}_{s}}^{(h,w,c)}$。在仅在源域上进行训练而没有域自适应的情况下，优化问题仅显示为：
$$\min_{\theta_{F}}\frac{1}{|\mathcal{X}_{s}|}\sum_{\boldsymbol{x}_{s}\in\mathcal{X}_{s}}\mathcal{L}_{seg}(\boldsymbol{x}_{s},\boldsymbol{y}_{s}).$$
(1)
# 2.1. Direct entropy minimization
对于目标域，由于我们没有用于图像样本$x_t\in\mathcal{X}_t$ 的注释$y_t$ ,因此我们无法使用 (1)来学习$F$ 。一些方法使用模型的预测$\hat{y}_t$作为$y_t$的代理。此外，此代理仅用于预测足够可靠的像素。我们建议不使用高置信度代理，而是约束模型，使其产生高置信度预测。我们通过最小化预测的熵来实现这一点
我们引人熵损失$\mathcal{L}_{ent}$来直接最大化目标域中的预测确定性。在这项工作中，我们使用香农熵[36]。给定目标输人图像$x_t$ ,熵图$E_{\boldsymbol{x}_t}\in[0,1]^{H\times W}$由归一化为[0,1] 范围的独立像素熵组成：
$$\boldsymbol{E}_{\boldsymbol{x}_t}^{(h,w)}=\frac{-1}{\log(C)}\sum_{c=1}^C\boldsymbol{P}_{\boldsymbol{x}_t}^{(h,w,c)}\log\boldsymbol{P}_{\boldsymbol{x}_t}^{(h,w,c)},$$
(2)
在像素$(h,w)$处。熵图的一个例子如图 2 所示。熵损
失$\mathcal{L}_{ent}$定义为所有像素归一化熵的总和：
(3)
$$\mathcal{L}_{ent}(\boldsymbol{x}_{t})=\sum_{h,w}\boldsymbol{E}_{\boldsymbol{x}_{t}}^{(h,w)}.$$
在训练过程中，我们共同优化了源样本上的有监督分割损失$\mathcal{L}_seg$和目标样本上的无监督熵损失$\mathcal{L}_ent$ 。最终的优化问题表述如下：
$$\min_{\theta_{F}}\frac{1}{|\mathcal{X}_{s}|}\sum_{\boldsymbol{x}_{s}}\mathcal{L}_{seg}(\boldsymbol{x}_{s},\boldsymbol{y}_{s})+\frac{\lambda_{ent}}{|\mathcal{X}_{t}|}\sum_{\boldsymbol{x}_{t}}\mathcal{L}_{ent}(\boldsymbol{x}_{t}),$$
, 其中$\mathbb{E}\lambda_ent$ 作为熵项$\mathbb{E}\mathcal{L}_ent$ 的加权因子
Connection to self-training. 伪标记是一种简单而有效的半监督学习方法 [21]。最近，该方法已通过迭代自训练(ST) 过程$\mathbb{E}[51]$ 应用于语义分割任务中的 UDA。ST 方法假设目标样本的高得分像素预测集
2
|
[
[
0.08251953125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.08251953125,
0.1060791015625
],
[
0.50341796875,
0.0921630859375,
0.88818359375,
0.0921630859375,
0.88818359375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.08416748046875,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.121826171875,
0.08416748046875,
0.121826171875
],
[
0.50341796875,
0.10797119140625,
0.88818359375,
0.10797119140625,
0.88818359375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.080078125,
0.12249755859375,
0.468017578125,
0.121826171875,
0.468017578125,
0.1351318359375,
0.080078125,
0.1363525390625
],
[
0.50341796875,
0.12371826171875,
0.53759765625,
0.12371826171875,
0.53759765625,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.08251953125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08251953125,
0.1514892578125
],
[
0.52294921875,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.52294921875,
0.1514892578125
],
[
0.08251953125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1666259765625,
0.08251953125,
0.1666259765625
],
[
0.5009765625,
0.147705078125,
0.8935546875,
0.147705078125,
0.8935546875,
0.1685791015625,
0.5009765625,
0.1685791015625
],
[
0.5009765625,
0.1673583984375,
0.8095703125,
0.1634521484375,
0.8095703125,
0.18115234375,
0.5009765625,
0.1849365234375
],
[
0.7998046875,
0.1685791015625,
0.8896484375,
0.17041015625,
0.8896484375,
0.1817626953125,
0.7998046875,
0.179931640625
],
[
0.08087158203125,
0.18310546875,
0.46728515625,
0.18310546875,
0.46728515625,
0.1962890625,
0.08087158203125,
0.1962890625
],
[
0.08416748046875,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08416748046875,
0.179931640625
],
[
0.5009765625,
0.1849365234375,
0.89208984375,
0.18115234375,
0.89208984375,
0.1988525390625,
0.5009765625,
0.20263671875
],
[
0.08416748046875,
0.199462890625,
0.46728515625,
0.199462890625,
0.46728515625,
0.210205078125,
0.08416748046875,
0.210205078125
],
[
0.50341796875,
0.1956787109375,
0.8935546875,
0.1956787109375,
0.8935546875,
0.2197265625,
0.50341796875,
0.2197265625
],
[
0.08416748046875,
0.21337890625,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.08416748046875,
0.2266845703125
],
[
0.50341796875,
0.216552734375,
0.890625,
0.216552734375,
0.890625,
0.2310791015625,
0.50341796875,
0.2310791015625
],
[
0.08416748046875,
0.2298583984375,
0.46728515625,
0.2298583984375,
0.46728515625,
0.240478515625,
0.08416748046875,
0.240478515625
],
[
0.5,
0.23291015625,
0.80224609375,
0.2279052734375,
0.80224609375,
0.248779296875,
0.5,
0.25390625
],
[
0.7861328125,
0.236083984375,
0.890625,
0.231689453125,
0.89208984375,
0.25,
0.78759765625,
0.25390625
],
[
0.08251953125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.08251953125,
0.2568359375
],
[
0.5009765625,
0.25244140625,
0.8837890625,
0.248779296875,
0.8837890625,
0.270751953125,
0.5009765625,
0.274658203125
],
[
0.08087158203125,
0.2587890625,
0.453369140625,
0.2587890625,
0.453369140625,
0.272705078125,
0.08087158203125,
0.272705078125
],
[
0.08087158203125,
0.2734375,
0.468017578125,
0.2734375,
0.468017578125,
0.287353515625,
0.08087158203125,
0.287353515625
],
[
0.50244140625,
0.270751953125,
0.8896484375,
0.268310546875,
0.8896484375,
0.28271484375,
0.50244140625,
0.285400390625
],
[
0.080078125,
0.2890625,
0.468017578125,
0.28857421875,
0.468017578125,
0.3017578125,
0.080078125,
0.302978515625
],
[
0.5048828125,
0.287353515625,
0.8896484375,
0.287353515625,
0.8896484375,
0.29736328125,
0.5048828125,
0.29736328125
],
[
0.08087158203125,
0.302978515625,
0.46728515625,
0.302978515625,
0.46728515625,
0.31640625,
0.08087158203125,
0.31640625
],
[
0.5048828125,
0.302978515625,
0.8896484375,
0.302978515625,
0.8896484375,
0.313232421875,
0.5048828125,
0.313232421875
],
[
0.08251953125,
0.318115234375,
0.46728515625,
0.318115234375,
0.46728515625,
0.33203125,
0.08251953125,
0.33203125
],
[
0.50341796875,
0.31640625,
0.88232421875,
0.31640625,
0.88232421875,
0.330078125,
0.50341796875,
0.330078125
],
[
0.08087158203125,
0.333251953125,
0.24267578125,
0.333251953125,
0.24267578125,
0.3466796875,
0.08087158203125,
0.3466796875
],
[
0.5185546875,
0.333251953125,
0.56396484375,
0.335205078125,
0.56201171875,
0.349853515625,
0.51708984375,
0.347900390625
],
[
0.54345703125,
0.32763671875,
0.896484375,
0.32568359375,
0.896484375,
0.35107421875,
0.54345703125,
0.352783203125
],
[
0.11517333984375,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.363037109375,
0.11517333984375,
0.363037109375
],
[
0.50634765625,
0.348388671875,
0.890625,
0.348388671875,
0.890625,
0.363037109375,
0.50634765625,
0.363037109375
],
[
0.08251953125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.08251953125,
0.37744140625
],
[
0.50244140625,
0.3642578125,
0.54345703125,
0.3642578125,
0.54345703125,
0.37890625,
0.50244140625,
0.37890625
],
[
0.080078125,
0.37744140625,
0.468017578125,
0.37890625,
0.468017578125,
0.393310546875,
0.080078125,
0.392578125
],
[
0.5947265625,
0.37744140625,
0.7099609375,
0.389404296875,
0.705078125,
0.419189453125,
0.5888671875,
0.40771484375
],
[
0.63818359375,
0.37744140625,
0.66259765625,
0.380126953125,
0.6611328125,
0.389404296875,
0.63623046875,
0.386962890625
],
[
0.6640625,
0.380126953125,
0.7958984375,
0.38330078125,
0.7958984375,
0.40087890625,
0.6640625,
0.397705078125
],
[
0.86767578125,
0.385009765625,
0.890625,
0.385009765625,
0.890625,
0.3984375,
0.86767578125,
0.3984375
],
[
0.08251953125,
0.39453125,
0.46728515625,
0.39453125,
0.46728515625,
0.40771484375,
0.08251953125,
0.40771484375
],
[
0.08087158203125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.50244140625,
0.416748046875,
0.7587890625,
0.416748046875,
0.7587890625,
0.429931640625,
0.50244140625,
0.429931640625
],
[
0.08087158203125,
0.4248046875,
0.46728515625,
0.4248046875,
0.46728515625,
0.438232421875,
0.08087158203125,
0.438232421875
],
[
0.08251953125,
0.44140625,
0.46728515625,
0.44140625,
0.46728515625,
0.451904296875,
0.08251953125,
0.451904296875
],
[
0.52392578125,
0.438232421875,
0.890625,
0.438232421875,
0.890625,
0.451904296875,
0.52392578125,
0.451904296875
],
[
0.08087158203125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.08087158203125,
0.468505859375
],
[
0.50244140625,
0.45068359375,
0.8896484375,
0.449462890625,
0.8896484375,
0.46728515625,
0.50244140625,
0.468505859375
],
[
0.08251953125,
0.4716796875,
0.46728515625,
0.4716796875,
0.46728515625,
0.481689453125,
0.08251953125,
0.481689453125
],
[
0.50341796875,
0.468505859375,
0.890625,
0.468505859375,
0.890625,
0.481689453125,
0.50341796875,
0.481689453125
],
[
0.08087158203125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.50341796875,
0.48291015625,
0.890625,
0.484130859375,
0.890625,
0.49755859375,
0.50341796875,
0.496826171875
],
[
0.08251953125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.08251953125,
0.5146484375
],
[
0.50341796875,
0.498779296875,
0.8896484375,
0.498779296875,
0.8896484375,
0.51220703125,
0.50341796875,
0.51220703125
],
[
0.08251953125,
0.517578125,
0.465576171875,
0.517578125,
0.465576171875,
0.52783203125,
0.08251953125,
0.52783203125
],
[
0.50341796875,
0.5146484375,
0.74853515625,
0.5146484375,
0.74853515625,
0.52783203125,
0.50341796875,
0.52783203125
],
[
0.08251953125,
0.52978515625,
0.46728515625,
0.52978515625,
0.46728515625,
0.54345703125,
0.08251953125,
0.54345703125
],
[
0.52294921875,
0.52880859375,
0.8896484375,
0.52880859375,
0.8896484375,
0.54248046875,
0.52294921875,
0.54248046875
],
[
0.080078125,
0.54345703125,
0.468017578125,
0.54443359375,
0.468017578125,
0.5595703125,
0.080078125,
0.55810546875
],
[
0.50341796875,
0.54443359375,
0.8896484375,
0.54443359375,
0.8896484375,
0.55810546875,
0.50341796875,
0.55810546875
],
[
0.08251953125,
0.56005859375,
0.46728515625,
0.56005859375,
0.46728515625,
0.57373046875,
0.08251953125,
0.57373046875
],
[
0.5009765625,
0.55615234375,
0.8935546875,
0.5537109375,
0.8935546875,
0.57470703125,
0.5009765625,
0.5771484375
],
[
0.080078125,
0.57568359375,
0.131591796875,
0.57568359375,
0.131591796875,
0.59033203125,
0.080078125,
0.59033203125
],
[
0.5048828125,
0.57470703125,
0.71142578125,
0.57470703125,
0.71142578125,
0.58837890625,
0.5048828125,
0.58837890625
],
[
0.098876953125,
0.5888671875,
0.468017578125,
0.59033203125,
0.468017578125,
0.60498046875,
0.098876953125,
0.60400390625
],
[
0.08251953125,
0.60595703125,
0.468017578125,
0.60595703125,
0.468017578125,
0.619140625,
0.08251953125,
0.619140625
],
[
0.63134765625,
0.60595703125,
0.67236328125,
0.60595703125,
0.67236328125,
0.62060546875,
0.63134765625,
0.62060546875
],
[
0.66650390625,
0.60791015625,
0.85302734375,
0.60986328125,
0.85302734375,
0.6318359375,
0.66650390625,
0.6298828125
],
[
0.86669921875,
0.61572265625,
0.8896484375,
0.61572265625,
0.8896484375,
0.62744140625,
0.86669921875,
0.62744140625
],
[
0.60888671875,
0.6162109375,
0.701171875,
0.62548828125,
0.69775390625,
0.642578125,
0.60546875,
0.63330078125
],
[
0.55126953125,
0.611328125,
0.61767578125,
0.611328125,
0.61767578125,
0.6298828125,
0.55126953125,
0.6298828125
],
[
0.08087158203125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.08087158203125,
0.63525390625
],
[
0.8447265625,
0.61865234375,
0.87109375,
0.61865234375,
0.87109375,
0.62548828125,
0.8447265625,
0.62548828125
],
[
0.08251953125,
0.63623046875,
0.46728515625,
0.63623046875,
0.46728515625,
0.6494140625,
0.08251953125,
0.6494140625
],
[
0.08251953125,
0.65283203125,
0.465576171875,
0.65283203125,
0.465576171875,
0.66357421875,
0.08251953125,
0.66357421875
],
[
0.50341796875,
0.64794921875,
0.890625,
0.64794921875,
0.890625,
0.6611328125,
0.50341796875,
0.6611328125
],
[
0.08087158203125,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.08087158203125,
0.6796875
],
[
0.50341796875,
0.662109375,
0.79833984375,
0.662109375,
0.79833984375,
0.67529296875,
0.50341796875,
0.67529296875
],
[
0.08251953125,
0.68310546875,
0.17724609375,
0.68310546875,
0.17724609375,
0.6943359375,
0.08251953125,
0.6943359375
],
[
0.61279296875,
0.689453125,
0.77685546875,
0.68505859375,
0.77783203125,
0.71044921875,
0.61376953125,
0.7138671875
],
[
0.1004638671875,
0.6962890625,
0.468017578125,
0.6962890625,
0.468017578125,
0.71044921875,
0.1004638671875,
0.71044921875
],
[
0.869140625,
0.6923828125,
0.890625,
0.6923828125,
0.890625,
0.70556640625,
0.869140625,
0.70556640625
],
[
0.08251953125,
0.7119140625,
0.46728515625,
0.7119140625,
0.46728515625,
0.72607421875,
0.08251953125,
0.72607421875
],
[
0.69921875,
0.708984375,
0.7255859375,
0.708984375,
0.7255859375,
0.72021484375,
0.69921875,
0.72021484375
],
[
0.08087158203125,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.08087158203125,
0.74072265625
],
[
0.50341796875,
0.72265625,
0.8896484375,
0.72265625,
0.8896484375,
0.736328125,
0.50341796875,
0.736328125
],
[
0.50244140625,
0.73486328125,
0.890625,
0.736328125,
0.890625,
0.75048828125,
0.50244140625,
0.75
],
[
0.08251953125,
0.7412109375,
0.468017578125,
0.7412109375,
0.468017578125,
0.7548828125,
0.08251953125,
0.7548828125
],
[
0.50341796875,
0.75048828125,
0.67236328125,
0.75048828125,
0.67236328125,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.08087158203125,
0.7568359375,
0.2115478515625,
0.7568359375,
0.2115478515625,
0.77099609375,
0.08087158203125,
0.77099609375
],
[
0.1004638671875,
0.77294921875,
0.468017578125,
0.77294921875,
0.468017578125,
0.7861328125,
0.1004638671875,
0.7861328125
],
[
0.56640625,
0.77783203125,
0.58837890625,
0.78076171875,
0.58642578125,
0.79150390625,
0.56396484375,
0.787109375
],
[
0.7197265625,
0.77587890625,
0.89208984375,
0.7841796875,
0.890625,
0.8046875,
0.71826171875,
0.796875
],
[
0.583984375,
0.78076171875,
0.71826171875,
0.78515625,
0.71728515625,
0.80322265625,
0.58251953125,
0.798828125
],
[
0.525390625,
0.77880859375,
0.619140625,
0.7890625,
0.61376953125,
0.8193359375,
0.51953125,
0.8095703125
],
[
0.08087158203125,
0.787109375,
0.46728515625,
0.787109375,
0.46728515625,
0.80029296875,
0.08087158203125,
0.80029296875
],
[
0.525390625,
0.787109375,
0.56640625,
0.787109375,
0.56640625,
0.79736328125,
0.525390625,
0.79736328125
],
[
0.72216796875,
0.79150390625,
0.775390625,
0.79345703125,
0.77392578125,
0.8095703125,
0.72119140625,
0.806640625
],
[
0.08087158203125,
0.8017578125,
0.46728515625,
0.8017578125,
0.46728515625,
0.81494140625,
0.08087158203125,
0.81494140625
],
[
0.7646484375,
0.8046875,
0.77294921875,
0.8046875,
0.77294921875,
0.81005859375,
0.7646484375,
0.81005859375
],
[
0.08251953125,
0.8173828125,
0.34228515625,
0.8173828125,
0.34228515625,
0.83056640625,
0.08251953125,
0.83056640625
],
[
0.5048828125,
0.8193359375,
0.8037109375,
0.8193359375,
0.8037109375,
0.83349609375,
0.5048828125,
0.83349609375
],
[
0.5087890625,
0.84033203125,
0.890625,
0.84130859375,
0.890625,
0.85595703125,
0.5087890625,
0.85498046875
],
[
0.08087158203125,
0.84765625,
0.111083984375,
0.84765625,
0.111083984375,
0.85986328125,
0.08087158203125,
0.85986328125
],
[
0.10540771484375,
0.8486328125,
0.2099609375,
0.8486328125,
0.2099609375,
0.861328125,
0.10540771484375,
0.861328125
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.1004638671875,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.1004638671875,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.8896484375,
0.8720703125,
0.8896484375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.5009765625,
0.88525390625,
0.8896484375,
0.88623046875,
0.8896484375,
0.90087890625,
0.5009765625,
0.8994140625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"器，该鉴别器从分割网络中获取特征图并尝试预测输",
"的模型，我们从现有的语义分割框架开始，并添加一个",
"人域。分段网络试图欺骗鉴别器，从而使来自两个域",
"用于域适应的额外网络分支。图2说明了我们的体系",
"的特征具有相似的分布。Hoffmanetal.[15]是第一个",
"结构",
"将对抗方法应用于UDA的语义分割。通过从源域传输",
"我们的模型在源域上进行了有监督损失的训练。形",
"标签统计信息，它们还具有特定于类别的适应性。在",
"式上，我们考虑一组X。CRHXWx3源示例以及相关",
"的地面实况C类分割图yC（1,C)HxW",
"。示例as",
"通过对网格软伪标签进行对抗性训练来完成的。在「4",
"51中使用了类似的全局和类对齐方法，其中类对齐是",
"是H×W彩色图像，关联映射ys的条目y(h,w）=",
"中，对抗性训练用于空间感知自适应以及蒸馏损失，以",
"[ysh,u,]。提供了像素（h,w)的标签作为独热矢量。令",
"专门解决合成2实域偏移。「16]使用残差网使源特征",
"F是一个语义分割网络，它获取图像&并预测C维",
"图类似于使用对抗性训练的目标特征图，然后将特征",
"的“软分割图”F(a)=Pa=[p(h.w,)]",
"Th,w,c。凭借",
"图用于分割任务。在「41]中，对抗性方法用于输出空",
"最终的softmax层，每个C维像素矢量[Pp(h,w,c)]",
"间，以受益于跨域的结构一致性。[32，33]提出了另",
"种使用对抗性训练的有趣方法：他们在目标域图像上",
"表现为在类上的离散分布。如果一个类脱颖而出，芬",
"获得了两个预测，这是由两个分类器[33]或在分类器",
"布是挑剔的（低摘）：如果分数均匀分布，从网络的",
"[32]中使用dropout完成的。给定两个预测，分类器被",
"角度来看，不确定性的迹象，摘很大。学寸F的参",
"训练以最大化分布之间的差异，而网络的特征提取器",
"数F以最小化源样本上的分割损失Lseg（acs,ys）",
"部分被训练以最小化它",
"Zh=1",
"Wyhw)logP），在仅在源域",
"些方法建立在生成网络上，以生成以源为条件的",
"上进行训练而没有域自适应的情况下，优化问题仅显",
"目标图像。Hoffmanetal.[14]提出了周期一致的对抗",
"示为：",
"域自适应（CyCADA），其中它们在像素级和特征级表",
"inex",
"1",
" Cseg(as, ys).",
"(1)",
"示上都进行了自适应。对于像素级自适应，他们使用",
"Cycle-GAN【48]生成以源图像为条件的目标图像。在",
"2.1.Direct entropy minimization",
"[34]中，学习生成模型以从特征空间重建图像。然后，",
"对于域自适应，使用生成器模块训练特征模块以在源",
"对于目标域，由于我们没有用于图像样本tEXt",
"特征上产生目标图像，反之亦然。在DCAN「43中，",
"的注释 yt，因此我们无法使用（1）来学习F。一些方",
"生成器和分段网络中使用了通道式特征对齐。在生成",
"法使用模型的预测9t作为t的代理。此外，此代理",
"的图像上学习分割网络，其中源的内容和目标的样式",
"仅用于预测足够可靠的像素。我们建议不使用高置信",
"将源分割图用作地面实况。50】中的作者使用生成对抗",
"度代理，而是约束模型，使其产生高置信度预测。我们",
"网络（GANE11来对齐源和自标入。此外，它们",
"通过最小化预测的来实现这一点",
"用保守损失（CL）代替交叉损失，从而惩罚源示例",
"我们引入摘损失Lent来直接最大化目标域中的预测",
"的简单和困难情况。CL方法与大多数UDA方法正交，",
"确定性。在这项工作中，我们使用香农摘[36]。给定",
"包括我们的方法：它可以使任何对源使用交叉炳的方",
"自标输人图像at，摘图E，E[0,ijHxW由归一化为",
"法受益",
"[0,1]范围的独立像素熵组成：",
"UDA的另一种方法是自我训练。这个想法是使用来",
"自集成模型或模型的先前状态的预测作为未标记数据",
"-1",
"ZP(h,we) og P(h,w,c),",
"(2)",
"1og(C) 1",
"E(h.w) =",
"的伪标签来训练当前模型。许多半监督方法[20，39]使",
"",
"用自训练。在[51]中，UDA在语义分割上采用了自训",
"练，并进一步扩展了类平衡和空间先验。正如我们在",
"在像素（h,w）处。摘图的一个例子如图2所示。摘损",
"2.1节中讨论的那样，自我训练与建议的嫡最小化方法",
"失Lent定义为所有像素归一化炳的总和：",
"有有趣的联系",
"Cent(rt) =EEl(htm).",
"在其他一些方法中，26]通过多次损失使用对抗性",
"(3)",
"和生成技术的组合，【46]结合了用于外观适应的生成方",
"h,w",
"法和用于表示适应的对抗性训练，45]】提出了一种课程",
"在训练过程中，我们共同优化了源样本上的有监督分",
"割损失Lseg和目标样本上的无监督摘损失Cent。最",
"风格的UDA学习通过在局部（超像素级）和全局标签",
"终的优化问题表述如下：",
"分布上实施一致性",
"摘最小化已被证明可用于半监督学习[12，38]和聚",
"1",
"Nent Lent(at), (4)",
"Lseg(as, ys) +",
"南",
"类17,18】。它最近也被应用于分类任务[25]的域自适",
"min.",
"x",
"应。据我们所知，我们首先成功地应用基于摘的UDA",
"T.",
"训练来获得语义分割任务的竞争性能",
"，其中E入ent作为项ECent的加权因子",
"Connection to self-training：伪标记是一种简单而",
"2.",
"Approaches",
"有效的半监督学习方法[21]。最近，该方法已通过",
"在本节中，我们提出了两种建议的摘最小化方法，使",
"迭代自训练（ST）过程E[51]应用于语义分割任务中",
"用（i）无监督损失和（i）对抗训练。为了构建我们",
"的UDA。ST方法假设目标样本的高得分像素预测集",
"2"
] |
{
"arxiv_id": "1811.12833",
"language": "zh",
"page_idx": 1
}
|
||
1811.12833-zh-2
|
1811.12833-zh-2.jpg
|
<FigureHere>
Figure 2: Approach overview. The figure shows our two approaches for UDA. First, direct entropy minimization minimizes
the entropy of the target $\boldsymbol{P}_{\boldsymbol{x}_t}$, which is equivalent to minimizing the sum of weighted self-information maps $\boldsymbol{I}_{\boldsymbol{x}_t}.$ In the second, complementary approach, we use adversarial training to enforce the consistency in $\boldsymbol{I}_{\boldsymbol{x}}$ across domains. Red arrows are used for target domain and blue arrows for source. An example of entropy map is shown for illustration.
$K\subset(1,H)\times(1,W)$是正确的，且概率很高。这样的假设允许在目标预测上使用带有伪标签的交叉熵损失。在实践中，$K$是通过选择具有固定或预定阈值的高分像素来构建的。为了得出与熵最小化的联系，我们将ST 方法的熵问题写为：
$$\min_{\theta_{F}}\frac{1}{|\mathcal{X}_{s}|}\sum_{\boldsymbol{x}_{s}}\mathcal{L}_{seg}(\boldsymbol{x}_{s},\boldsymbol{y}_{s})+\frac{\lambda_{pl}}{|\mathcal{X}_{t}|}\sum_{\boldsymbol{x}_{t}}\mathcal{L}_{seg}(\boldsymbol{x}_{t},\hat{\boldsymbol{y}}_{t}),$$
的 weighted self-information distributions ,我们间接地最小化了目标预测的熵。此外，由于适应是在加权自信息空间上进行的，我们的模型利用了从源到目标的结构信息
详 细 地 说 , 给 定 一 个 按 像 素 预 测 的 类 分 数 $P_{x}^{( h, w, c) }$ 白 小 片 中 一 个 个 上 下 ,
, 自我信息或“令人惊讶”
$\mathbb{E} [ 40]$ 被定义为
$-\log\boldsymbol{P}_{\boldsymbol{x}}^{(h,w,c)}$。实际上，E(2)中的熵$\boldsymbol E_x^(\dot{h},w)$是自信息$\mathbb{E} _{c}[ - \log \boldsymbol{P}_{\boldsymbol{x}}^{( h, w, c) }]$ 的 期 望 值 。我 们 在 这 里 对 由 像 素 级 向 量 $\boldsymbol{I}_{\boldsymbol{x}}^{( h, w) }= - \boldsymbol{P}_{\boldsymbol{x}}^{( h, w) }\cdot \log \boldsymbol{P}_{\boldsymbol{x}}^{( h, w) }\cdot 2$组 成 的 加 权 自 信 息映射$I_x$执行对抗性适应。这样的向量可以看作是香农熵的解纠缠。然后，我们构造一个具有参数$\theta_D$的全卷积鉴别器网络$D$ ,以$I_{x}$为输人，并产生域分类输出i.e.,类标签1(resp。0)forthesource(resp。目标)域。与$\mathbb{E}[11]$ 类似，我们训练鉴别器以区分来自源图像和目标图像的输出，同时训练分割网络以欺骗鉴别器。详细地说，让$\mathcal{L}_D$的交叉熵域分类损失。鉴别器的训练目标是：
,其中$\hat{y}_t$是$x_t$的单热点类预测，并且具有：
$$\mathcal{L}_{seg}(\boldsymbol{x}_t,\hat{\boldsymbol{y}}_t)=-\sum_{(h,w)\in K}\sum_{c=1}^C\hat{\boldsymbol{y}}_t^{(h,w,c)}\log\boldsymbol{P}_{\boldsymbol{x}_t}^{(h,w,c)}.$$
比较方程式$\mathbb{E}(2-3)$ 和$\mathbb{E}$XMATH4 我们的。与 ST$\mathbb{E}$ [51]不同，我们基于熵的方法不需要复杂的调度过程来选择阈值。甚至，与 ST 假设相反，我们在$\color{red}{\mathrm{E3.3}}$节中显示，在某些情况下，对“硬”或“最混乱”像素进行训练会产生更好的性能
(7)
$$\min_{\theta_{D}}\frac{1}{|\mathcal{X}_{s}|}\sum_{x_{s}}\mathcal{L}_{D}(\boldsymbol{I}_{\boldsymbol{x}_{s}},1)+\frac{1}{|\mathcal{X}_{t}|}\sum_{\boldsymbol{x}_{t}}\mathcal{L}_{D}(\boldsymbol{I}_{\boldsymbol{x}_{t}},0),$$
2.2. Minimizing entropy with adversarial learning
,训练分割网络的对抗性目标是：
(8)
$$\min_{\theta_F}\frac{1}{|\mathcal{X}_t|}\sum_{\boldsymbol{x}_t}\mathcal{L}_D(\boldsymbol{I}_{\boldsymbol{x}_t},1).$$
结合$\mathbb{E}(1)$ 和$\mathbb{E}(8)$, 我们得出了优化问题
$$\min_{\theta_{F}}\frac{1}{|\mathcal{X}_{s}|}\sum_{\boldsymbol{x}_{s}}\mathcal{L}_{seg}(\boldsymbol{x}_{s},\boldsymbol{y}_{s})+\frac{\lambda_{adv}}{|\mathcal{X}_{t}|}\sum_{\boldsymbol{x}_{t}}\mathcal{L}_{D}(\boldsymbol{I}_{\boldsymbol{x}_{t}},1),$$
输入图像的熵损失在公式$\mathbb{E}(\color{red}3$ )中定义为独立的逐像素预测熵之和。因此，仅仅最小化这种损失就忽略了局部语义之间的结构依赖性。如[41]所示，对于语义分割中的 UDA, 结构化输出空间上的自适应是有益的。它基于以下事实：源域和目标域在语义布局上具有很强的相似性
在这一部分中，我们介绍了一个统一的对抗训练框架，该框架通过使目标的熵分布类似于源来间接最小化熵。这允许利用域之间的结构一致性。为此，我们将UDA 任务表述为在加权自信息空间上最小化源和目标之间的分布距离。图 2 说明了我们的对抗性学习过程。我们的对抗性方法的动机是，训练后的模型自然会对类似源的图像产生低熵预测。通过对齐目标域和源域
与加权因子 XMAversar。在训练过程中，我们使用$\overline{\mathrm{E}}$
(7)和$\mathbb{E}(9)$中的目标函数交替优化网络$D$和$F$
${^{2}\mathrm{Abusing~notations,~}^{2}\cdot}$,and$,\log$,stand for Hadamard prod-
uct and point-wise logarithm respectively.
3
|
[
[
0.5849609375,
0.10797119140625,
0.64208984375,
0.10797119140625,
0.64208984375,
0.11932373046875,
0.5849609375,
0.11932373046875
],
[
0.57275390625,
0.1275634765625,
0.75341796875,
0.1275634765625,
0.75341796875,
0.141357421875,
0.57275390625,
0.141357421875
],
[
0.1094970703125,
0.1514892578125,
0.1192626953125,
0.1514892578125,
0.1192626953125,
0.1673583984375,
0.1094970703125,
0.1673583984375
],
[
0.11273193359375,
0.1666259765625,
0.1168212890625,
0.1673583984375,
0.11273193359375,
0.1749267578125,
0.1094970703125,
0.173583984375
],
[
0.59326171875,
0.1787109375,
0.71142578125,
0.179931640625,
0.71142578125,
0.1912841796875,
0.59326171875,
0.1900634765625
],
[
0.5556640625,
0.186279296875,
0.56884765625,
0.189453125,
0.564453125,
0.1988525390625,
0.55126953125,
0.1956787109375
],
[
0.5361328125,
0.2020263671875,
0.59326171875,
0.2020263671875,
0.59326171875,
0.21337890625,
0.5361328125,
0.21337890625
],
[
0.51708984375,
0.21337890625,
0.61279296875,
0.2139892578125,
0.61279296875,
0.2259521484375,
0.51708984375,
0.2247314453125
],
[
0.36767578125,
0.2279052734375,
0.390625,
0.2310791015625,
0.386474609375,
0.246826171875,
0.363525390625,
0.24365234375
],
[
0.7509765625,
0.2279052734375,
0.83837890625,
0.2298583984375,
0.83837890625,
0.2442626953125,
0.7509765625,
0.242431640625
],
[
0.78759765625,
0.2235107421875,
0.79833984375,
0.2235107421875,
0.79833984375,
0.2279052734375,
0.78759765625,
0.2279052734375
],
[
0.1094970703125,
0.23486328125,
0.1192626953125,
0.23486328125,
0.1192626953125,
0.26318359375,
0.1094970703125,
0.26318359375
],
[
0.2099609375,
0.238037109375,
0.221435546875,
0.238037109375,
0.221435546875,
0.242431640625,
0.2099609375,
0.242431640625
],
[
0.30810546875,
0.246826171875,
0.443603515625,
0.246826171875,
0.443603515625,
0.260009765625,
0.30810546875,
0.260009765625
],
[
0.156005859375,
0.267578125,
0.1666259765625,
0.267578125,
0.1666259765625,
0.270751953125,
0.156005859375,
0.270751953125
],
[
0.56640625,
0.2734375,
0.76318359375,
0.2734375,
0.76318359375,
0.28857421875,
0.56640625,
0.28857421875
],
[
0.1422119140625,
0.290283203125,
0.1650390625,
0.290283203125,
0.1650390625,
0.302978515625,
0.1422119140625,
0.302978515625
],
[
0.5810546875,
0.290283203125,
0.66259765625,
0.290283203125,
0.66259765625,
0.304931640625,
0.5810546875,
0.304931640625
],
[
0.325927734375,
0.298583984375,
0.423095703125,
0.298583984375,
0.423095703125,
0.313232421875,
0.325927734375,
0.313232421875
],
[
0.54345703125,
0.309326171875,
0.85888671875,
0.30810546875,
0.85888671875,
0.322509765625,
0.54345703125,
0.323974609375
],
[
0.08087158203125,
0.32763671875,
0.89208984375,
0.32763671875,
0.89208984375,
0.34228515625,
0.08087158203125,
0.34228515625
],
[
0.08087158203125,
0.34228515625,
0.89208984375,
0.34228515625,
0.89208984375,
0.35595703125,
0.08087158203125,
0.35595703125
],
[
0.080078125,
0.35595703125,
0.8935546875,
0.35595703125,
0.8935546875,
0.369384765625,
0.080078125,
0.369384765625
],
[
0.07843017578125,
0.369384765625,
0.80224609375,
0.369384765625,
0.80224609375,
0.38330078125,
0.07843017578125,
0.38330078125
],
[
0.080078125,
0.392578125,
0.468017578125,
0.391357421875,
0.468017578125,
0.406005859375,
0.080078125,
0.4072265625
],
[
0.50244140625,
0.391357421875,
0.890625,
0.392578125,
0.890625,
0.4072265625,
0.50244140625,
0.406005859375
],
[
0.08087158203125,
0.40771484375,
0.468017578125,
0.40771484375,
0.468017578125,
0.421630859375,
0.08087158203125,
0.421630859375
],
[
0.50341796875,
0.40771484375,
0.8896484375,
0.40771484375,
0.8896484375,
0.421630859375,
0.50341796875,
0.421630859375
],
[
0.08087158203125,
0.423583984375,
0.46728515625,
0.423583984375,
0.46728515625,
0.4375,
0.08087158203125,
0.4375
],
[
0.50341796875,
0.423583984375,
0.890625,
0.423583984375,
0.890625,
0.4375,
0.50341796875,
0.4375
],
[
0.08087158203125,
0.438232421875,
0.468017578125,
0.438232421875,
0.468017578125,
0.451904296875,
0.08087158203125,
0.451904296875
],
[
0.50244140625,
0.438232421875,
0.57177734375,
0.438232421875,
0.57177734375,
0.45263671875,
0.50244140625,
0.45263671875
],
[
0.07843017578125,
0.451904296875,
0.2459716796875,
0.45263671875,
0.2459716796875,
0.46728515625,
0.07843017578125,
0.466552734375
],
[
0.51953125,
0.45263671875,
0.8935546875,
0.447509765625,
0.8935546875,
0.468505859375,
0.51953125,
0.473388671875
],
[
0.51708984375,
0.4697265625,
0.890625,
0.468505859375,
0.890625,
0.48291015625,
0.51708984375,
0.484130859375
],
[
0.1502685546875,
0.479736328125,
0.2900390625,
0.484130859375,
0.289306640625,
0.505859375,
0.148681640625,
0.501953125
],
[
0.28759765625,
0.47607421875,
0.47216796875,
0.484130859375,
0.471435546875,
0.5087890625,
0.285888671875,
0.50048828125
],
[
0.5009765625,
0.48095703125,
0.8935546875,
0.481689453125,
0.8935546875,
0.50390625,
0.5009765625,
0.50244140625
],
[
0.1446533203125,
0.486083984375,
0.1519775390625,
0.486083984375,
0.1519775390625,
0.4892578125,
0.1446533203125,
0.4892578125
],
[
0.1356201171875,
0.481689453125,
0.1446533203125,
0.481689453125,
0.1446533203125,
0.4892578125,
0.1356201171875,
0.4892578125
],
[
0.09234619140625,
0.488037109375,
0.12744140625,
0.488037109375,
0.12744140625,
0.5,
0.09234619140625,
0.5
],
[
0.11767578125,
0.494384765625,
0.176513671875,
0.4912109375,
0.17724609375,
0.5068359375,
0.1192626953125,
0.51025390625
],
[
0.289306640625,
0.494384765625,
0.325927734375,
0.494384765625,
0.325927734375,
0.51025390625,
0.289306640625,
0.51025390625
],
[
0.5,
0.498779296875,
0.8935546875,
0.50048828125,
0.8935546875,
0.521484375,
0.5,
0.51953125
],
[
0.16015625,
0.505859375,
0.176513671875,
0.505859375,
0.176513671875,
0.5146484375,
0.16015625,
0.5146484375
],
[
0.32275390625,
0.505859375,
0.339111328125,
0.505859375,
0.339111328125,
0.51318359375,
0.32275390625,
0.51318359375
],
[
0.5,
0.5146484375,
0.89453125,
0.51513671875,
0.89453125,
0.54052734375,
0.5,
0.5390625
],
[
0.098876953125,
0.5,
0.1258544921875,
0.5,
0.1258544921875,
0.5068359375,
0.098876953125,
0.5068359375
],
[
0.08251953125,
0.521484375,
0.40283203125,
0.521484375,
0.40283203125,
0.53515625,
0.08251953125,
0.53515625
],
[
0.50244140625,
0.5361328125,
0.890625,
0.53857421875,
0.890625,
0.55322265625,
0.50244140625,
0.55029296875
],
[
0.275390625,
0.5498046875,
0.283447265625,
0.548828125,
0.28515625,
0.5517578125,
0.27685546875,
0.5537109375
],
[
0.2998046875,
0.55322265625,
0.471435546875,
0.5546875,
0.471435546875,
0.5771484375,
0.2998046875,
0.57470703125
],
[
0.50341796875,
0.55322265625,
0.890625,
0.55322265625,
0.890625,
0.56640625,
0.50341796875,
0.56640625
],
[
0.09478759765625,
0.56005859375,
0.221435546875,
0.56005859375,
0.221435546875,
0.57470703125,
0.09478759765625,
0.57470703125
],
[
0.2279052734375,
0.5595703125,
0.256591796875,
0.5595703125,
0.256591796875,
0.57373046875,
0.2279052734375,
0.57373046875
],
[
0.27685546875,
0.56005859375,
0.310546875,
0.5625,
0.3095703125,
0.57373046875,
0.275390625,
0.5712890625
],
[
0.5048828125,
0.5673828125,
0.890625,
0.5673828125,
0.890625,
0.58203125,
0.5048828125,
0.58203125
],
[
0.2115478515625,
0.57763671875,
0.2900390625,
0.5771484375,
0.2900390625,
0.58837890625,
0.2115478515625,
0.5888671875
],
[
0.50341796875,
0.58349609375,
0.890625,
0.58349609375,
0.890625,
0.5966796875,
0.50341796875,
0.5966796875
],
[
0.08251953125,
0.5947265625,
0.46728515625,
0.5947265625,
0.46728515625,
0.609375,
0.08251953125,
0.609375
],
[
0.50341796875,
0.5966796875,
0.88818359375,
0.59765625,
0.88818359375,
0.6123046875,
0.50341796875,
0.611328125
],
[
0.08251953125,
0.61279296875,
0.465576171875,
0.61279296875,
0.465576171875,
0.62353515625,
0.08251953125,
0.62353515625
],
[
0.50341796875,
0.61279296875,
0.8896484375,
0.61279296875,
0.8896484375,
0.626953125,
0.50341796875,
0.626953125
],
[
0.08251953125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.08251953125,
0.64013671875
],
[
0.50244140625,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.6416015625,
0.50244140625,
0.6416015625
],
[
0.08251953125,
0.642578125,
0.468017578125,
0.642578125,
0.468017578125,
0.65576171875,
0.08251953125,
0.65576171875
],
[
0.5048828125,
0.64306640625,
0.529296875,
0.64306640625,
0.529296875,
0.65576171875,
0.5048828125,
0.65576171875
],
[
0.08087158203125,
0.6572265625,
0.2459716796875,
0.6572265625,
0.2459716796875,
0.67041015625,
0.08087158203125,
0.67041015625
],
[
0.564453125,
0.66357421875,
0.5908203125,
0.66650390625,
0.58837890625,
0.6767578125,
0.56201171875,
0.673828125
],
[
0.69775390625,
0.66357421875,
0.865234375,
0.6494140625,
0.87109375,
0.68994140625,
0.70361328125,
0.7041015625
],
[
0.72216796875,
0.66357421875,
0.74169921875,
0.66943359375,
0.73779296875,
0.6767578125,
0.71826171875,
0.67041015625
],
[
0.52685546875,
0.67236328125,
0.560546875,
0.67236328125,
0.560546875,
0.6845703125,
0.52685546875,
0.6845703125
],
[
0.58642578125,
0.66552734375,
0.7099609375,
0.67041015625,
0.708984375,
0.689453125,
0.5849609375,
0.6845703125
],
[
0.525390625,
0.673828125,
0.59716796875,
0.67529296875,
0.59619140625,
0.69580078125,
0.52392578125,
0.693359375
],
[
0.86767578125,
0.67236328125,
0.890625,
0.67236328125,
0.890625,
0.68505859375,
0.86767578125,
0.68505859375
],
[
0.07843017578125,
0.68115234375,
0.4697265625,
0.681640625,
0.4697265625,
0.6962890625,
0.07843017578125,
0.69580078125
],
[
0.5947265625,
0.689453125,
0.60693359375,
0.689453125,
0.60693359375,
0.69775390625,
0.5947265625,
0.69775390625
],
[
0.740234375,
0.689453125,
0.76123046875,
0.689453125,
0.76123046875,
0.69775390625,
0.740234375,
0.69775390625
],
[
0.09722900390625,
0.7041015625,
0.468017578125,
0.705078125,
0.468017578125,
0.7197265625,
0.09722900390625,
0.71826171875
],
[
0.5048828125,
0.70849609375,
0.73876953125,
0.70849609375,
0.73876953125,
0.7216796875,
0.5048828125,
0.7216796875
],
[
0.08251953125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.08251953125,
0.734375
],
[
0.60302734375,
0.73291015625,
0.78515625,
0.71728515625,
0.7900390625,
0.75390625,
0.60888671875,
0.76953125
],
[
0.65625,
0.72998046875,
0.66259765625,
0.72998046875,
0.66259765625,
0.7373046875,
0.65625,
0.7373046875
],
[
0.0849609375,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.0849609375,
0.75
],
[
0.615234375,
0.736328125,
0.64794921875,
0.736328125,
0.64794921875,
0.74853515625,
0.615234375,
0.74853515625
],
[
0.86669921875,
0.73486328125,
0.8935546875,
0.73486328125,
0.8935546875,
0.751953125,
0.86669921875,
0.751953125
],
[
0.08416748046875,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.08416748046875,
0.7646484375
],
[
0.68310546875,
0.7548828125,
0.69287109375,
0.7548828125,
0.69287109375,
0.76220703125,
0.68310546875,
0.76220703125
],
[
0.08087158203125,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.08087158203125,
0.77978515625
],
[
0.50341796875,
0.77294921875,
0.80126953125,
0.77294921875,
0.80126953125,
0.7861328125,
0.50341796875,
0.7861328125
],
[
0.08251953125,
0.78076171875,
0.17724609375,
0.78076171875,
0.17724609375,
0.79248046875,
0.08251953125,
0.79248046875
],
[
0.7099609375,
0.78857421875,
0.896484375,
0.80029296875,
0.8935546875,
0.82568359375,
0.70751953125,
0.81298828125
],
[
0.098876953125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.81005859375,
0.098876953125,
0.81005859375
],
[
0.5107421875,
0.798828125,
0.60546875,
0.798828125,
0.60546875,
0.8291015625,
0.5107421875,
0.8291015625
],
[
0.55615234375,
0.796875,
0.564453125,
0.79345703125,
0.5703125,
0.80322265625,
0.56201171875,
0.80615234375
],
[
0.583984375,
0.8017578125,
0.5986328125,
0.8017578125,
0.5986328125,
0.8095703125,
0.583984375,
0.8095703125
],
[
0.5986328125,
0.8017578125,
0.708984375,
0.8017578125,
0.708984375,
0.81640625,
0.5986328125,
0.81640625
],
[
0.5986328125,
0.8046875,
0.60302734375,
0.8046875,
0.60302734375,
0.81005859375,
0.5986328125,
0.81005859375
],
[
0.08087158203125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.7138671875,
0.80810546875,
0.76953125,
0.80615234375,
0.771484375,
0.82470703125,
0.7158203125,
0.8271484375
],
[
0.59326171875,
0.8193359375,
0.6005859375,
0.82470703125,
0.59326171875,
0.830078125,
0.58642578125,
0.82470703125
],
[
0.75244140625,
0.82080078125,
0.76953125,
0.82080078125,
0.76953125,
0.830078125,
0.75244140625,
0.830078125
],
[
0.08251953125,
0.82763671875,
0.465576171875,
0.82763671875,
0.465576171875,
0.83837890625,
0.08251953125,
0.83837890625
],
[
0.08087158203125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50634765625,
0.83837890625,
0.890625,
0.83837890625,
0.890625,
0.85302734375,
0.50634765625,
0.85302734375
],
[
0.08087158203125,
0.85595703125,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.08087158203125,
0.869140625
],
[
0.50244140625,
0.85400390625,
0.84326171875,
0.8515625,
0.84326171875,
0.8662109375,
0.50244140625,
0.86865234375
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.51708984375,
0.875,
0.890625,
0.875,
0.890625,
0.8896484375,
0.51708984375,
0.8896484375
],
[
0.08087158203125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.5048828125,
0.8896484375,
0.75732421875,
0.8896484375,
0.75732421875,
0.8994140625,
0.5048828125,
0.8994140625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"Cent(t)",
"Direct entropyminimization",
"",
"",
"Pixel-wiseaggregate",
"I.",
"Weighted",
"self-information",
"Pa",
"EntropyImage",
"",
"",
"H",
"Soft-segmentation map",
"",
"+LD(Iea,1)+L(Iat,)",
"ys",
"Lp(Iat,1)",
"+Lseg(as,ys)",
"(EntropyminimizationwithAdversarial training）",
"Figure 2: Approach overview : The figure shows our two approaches for UDA. First, direct entropy minimization minimizes",
"the entropy of the target Pat , which is equivalent to minimizing the sum of weighted self-information maps Iat : In the",
"second, complementary approach, we use adversarial training to enforce the consistency in I across domains. Red arrows",
"are used for target domain and blue arrows for source. An example of entropy map is shown for illustration.",
"KC（1,H）×（1,W）是正确的，且概率很高。这样的",
"的weighted self-informationdistributions，我们间接",
"假设允许在目标预测上使用带有伪标签的交叉摘损失。",
"地最小化了目标预测的熵。此外，由于适应是在加权自",
"在实践中，K是通过选择具有固定或预定阈值的高分",
"信息空间上进行的，我们的模型利用了从源到目标的",
"像素来构建的。为了得出与熵最小化的联系，我们将",
"结构信息",
"ST方法的熵问题写为：",
"详细地说，给定一个按像素预测的类分数P(h,w,e)",
"自我信息或“令人惊”[40]被定义为",
"Cseg(ars, ys) +",
"Cseg(at, 9t), (5)",
"-log P(h,w,e)。实际上,E(2)中的摘 E(h,)是自信息",
"",
"1",
"min",
"",
"x",
"E。[-log P(h,w,e)] 的期望值。我们在这里对由像素级向",
"Es",
"t",
"量（h）=-P(h)logP(h）2组成的加权自信息",
"OF",
"，其中yt是&t的单热点类预测，并且具有：",
"映射I执行对抗性适应。这样的向量可以看作是香农",
"",
"(h,w,c) 1og P(h,u,c)。 (6)",
"熵的解纠缠。然后，我们构造一个具有参数θD的全卷",
"Cseg(at, yt) =-",
"",
"9t",
"积鉴别器网络D，以I为输人，并产生域分类输出",
"(h,w)eKc=l",
"i.e.，类标签1(resp。0)forthesource(resp。目标）域。",
"比较方程式E(2-3）和EXMATH4我们的。与STE",
"与E[11]类似，我们训练鉴别器以区分来自源图像和目",
"51]不同，我们基于的方法不需要复杂的调度过程来",
"标图像的输出，同时训练分割网络以欺骗鉴别器。详细",
"选择阈值。甚至，与ST假设相反，我们在E3.3节中",
"地说，让CD的交叉熵域分类损失。鉴别器的训练目标",
"显示，在某些情况下，对“硬”或“最混乱”像素进行",
"是:",
"训练会产生更好的性能",
"1",
"co(a.)",
"1",
"min",
"D(Ias,1) +",
"i4",
"(7)",
"2.2. Minimizing entropy with adversarial learning",
"c",
"E,",
"输入图像的摘损失在公式国（3）中定义为独立的逐",
"，训练分割网络的对抗性目标是：",
"像素预测嫡之和。因此，仅仅最小化这种损失就忽略",
"Co(,1)",
"1",
"了局部语义之间的结构依赖性。如【41]所示，对于语",
"min",
"(8)",
"义分割中的UDA，结构化输出空间上的自适应是有益",
"C",
"的。它基于以下事实：源域和目标域在语义布局上具有",
"结合E（1）和E（8），我们得出了优化问题",
"很强的相似性",
"Aau C(I1), (9)",
"在这一部分中，我们介绍了一个统一的对抗训练框",
"",
"1",
"",
"Lseg(as, ys) +",
"",
"架，该框架通过使目标的分布类似于源来间接最小",
"IXI2",
"&",
"at",
"化熵。这充许利用域之间的结构一致性。为此，我们将",
"UDA任务表述为在加权自信息空间上最小化源和目标",
"与加权因子XMAversar。在训练过程中，我们使用E",
"之间的分布距离。图2说明了我们的对抗性学习过程。",
"（7）和E(9）中的目标函数交替优化网络D和F",
"我们的对抗性方法的动机是，训练后的模型自然会对",
"2Abusing notations, '.' and ' log ' stand for Hadamard prod-",
"类似源的图像产生低熵预测。通过对齐目标域和源域",
"uct and point-wise logarithm respectively.",
"3"
] |
{
"arxiv_id": "1811.12833",
"language": "zh",
"page_idx": 2
}
|
||
1811.12833-zh-3
|
1811.12833-zh-3.jpg
|
# 2.3. Incorporating class-ratio priors
熵最小化可能会偏向一些简单的类。因此，有时用一些先验来指导学习是有益的。为此，我们基于类在源标签上的分布使用简单的类先验。我们将类先验向量$p_s$计算为源标签上每类像素数的$\ell_1$归一化直方图。现在，基于预测的$P_{x_t}$ ,使用
$$\mathcal{L}_{cp}(\boldsymbol{x}_t)=\sum_{c=1}^C\max\big(0,\mu\boldsymbol{p}_s^{(c)}-\mathbb{E}_c(\boldsymbol{P}_{\boldsymbol{x}_t}^{(c)})\big),$$
(10)
,其中$\mu\in[0,1]$ 用于放松类先验约束，任何类的预期概率与类先验$p_s$之间的差异太大。这解决了单个目标映像上的类分布不一定接近$p_s$的事实
# 3. Experiments
在本节中，我们将介绍我们的实验结果。第 3.1节介绍了使用的数据集以及我们的训练参数。在 3.2节和 3.3 节中，我们报告并讨论了我们的主要结果。In Section 3.3 , we discuss a preliminary result on entropy-based UDA for detection.
# 3.1. Experimental details
Datasets. 为了评估我们的方法，我们使用具有挑战性的合成-2-实数无监督域自适应设置。模型在完全注释的合成数据上进行训练，并在现实世界的数据上进行验证。在这样的设置中，模型在训练期间可以访问一些未标记的真实图像。为了训练我们的模型，我们使用 GTA5 [30] 或 SYNTHIA [31] 作为源域合成数据，同时使用 Cityscapes 数据集 [6]的训练分割作为目标域数据。以前在其他作品中也使用过类似的设置[15,14,41,51]。详细说明：
$\bullet$GTA5 $\to$Cityscapes: The GTA5 dataset con-
sists of 24,966 synthesized frames captured from a video game. Images are provided with pixellevel semantic annotations of 33 classes. Similar to [15], we use the 19 classes in common with the Cityscapes dataset.
$\bullet$ SYNTHIA$\to$Cityscapes: We use the SYNTHIA-
RAND-CITYSCAPES set$^3$ with 9,400 synthesized images for training. We train our models with 16 common classes in SYNTHIA and Cityscapes. While evaluating we compare the performance on 16- and 13-class subsets following the protocol used in [51].
在这两个设置中，2,975 未标记的 Cityscapes 图像用于训练。我们使用标准的平均交点超过联合(mIoU) 度量$\overline{\mathbb{E}[9]\text{ 来衡量分割性能。对 500 验证映像进行评估}}$
$^3$A split of the SYNTHIA dataset [31] using compatible labels
with the Cityscapes dataset.
Network architectures. 我们使用 Deeplab-V2$\boxed{\mathrm{F}}[2]$ 作为基本语义分割体系结构$F$ 。为了更好地捕捉场景上下文，将 Atrous 空间金字塔池 (ASPP) 应用于最后一层的特征输出。采样率固定为$\{6,12,18,24\}$ ,类似于$\mathbb{E}[2]$ 中的 ASPP-L 模型。我们在两种不同的基础深度 CNN 架构上进行了实验：VGG-16$\underline{\mathrm{E}}[37]$和 ResNet- $101\underline{\mathrm{E}}[13]$。在$\mathbb{E}[2]$ 之后，我们修改了最后一层的步幅和膨胀率，以生成具有更大视场的更密集的特征图。为了进一步提高 ResNet-101 的性能，我们对来自 conv4和 conv5 功能的多级输出执行自适应[41]
2.2节中介绍的对抗网络$D$与 DCGAN [28] 中使用的网络具有相同的体系结构。加权自信息映射$I_x$通过 4 卷积层转发，每个卷积层都与具有固定负斜率 0.2的 leaky-ReLU 层耦合。最后，分类器层生成分类输出， 指示输入是对应于源域还是目标域
Implementation details. 我们在实现中采用 PyTorch 深度学习框架 [27] 。所有实验均在具有 11GB 内存的单个 NVIDIA1080TIGPU 上完成。我们的模型，除了$\color{red}{\mathrm{E}}\color{red}{\mathrm{2.2~}}$节中提到的对抗性鉴别器外，都是使用随机梯度下降优化器$\mathbb{E}[1]$ 进行训练的，其学习率为$2. 5\times 10^{- 4}$ , 动量为 0.9 , 权重为 $10^-4$ 。我们使用具有学习率 10$^{-4}$ 的 Adam 优化器$\mathbb{E}[19]$ 来训练鉴别器。为了调度学习速率，我们遵循 [2]中提到的多项式退火过程
熵和对抗损失的加权因子：要为$\mathcal{L}_{ent}$设置权重，训练集性能提供了重要的指示。如果$\lambda_{ent}$很大，则熵下降得太快，并且模型强烈偏向于几个类。然而，当$\lambda_{ent}$ 选择在合适的范围内时，性能更好，并且对精确值不敏感。因此，我们对所有实验使用相同的$\lambda _{ent}= 0. 001$ , 而不管网络或数据集如何。类似的参数适用于 (9)中的权重$\lambda_{adv}$ 。我们在所有实验中都修复了$\lambda_{adv}=0.001$
# 3.2. Results
我们提供了与不同基线相比的方法的实验结果。我们的模型在两个 UDA 基准测试中实现了最先进的性能。在下面的内容中，我们展示了我们的方法在不同设置，i.e. ,训练集和基本 CNN 中的不同行为
GTA5$\to$Cityscapes: 我们在表E1 中报告-在城市景观验证集上以 mIoU(%) 表示的语义分割性能。我们的第一种直接熵最小化方法，在表 1-a 中称为 MinEnt , 在基于 VGG-16 和 ResNet-101 的 CNN 上实现了与最先进基线相当的性能。MinEnt 在没有和具有类平衡$\mathbb{E}$ [51] 的情况下优于自我训练(ST)方法。与$\mathbb{E}[41]$ 相比，基于 ResNet-101 的 MinEnt 显示出相似的结果， 但没有求助于鉴别器网络的训练。熵损失的轻开销使得 MinEnt 模型的训练时间要少得多。我们的熵方法的另一个优点是易于训练。事实上，由于其不稳定性， 训练对抗网络通常被认为是一项艰巨的任务。我们观察到一个更稳定的行为训练模型与熵损失
有趣的是，我们发现在某些情况下，仅在某些范围上应用熵损失效果最好。使用基于 ResNet-101 的模型可以观察到这种现象。实际上，我们通过对每个目标样本的顶部 30% 内具有熵值的像素进行训练来获得更
4
|
[
[
0.08087158203125,
0.0921630859375,
0.361083984375,
0.0921630859375,
0.361083984375,
0.1060791015625,
0.08087158203125,
0.1060791015625
],
[
0.5087890625,
0.0921630859375,
0.6787109375,
0.0921630859375,
0.6787109375,
0.1060791015625,
0.5087890625,
0.1060791015625
],
[
0.69677734375,
0.091552734375,
0.88818359375,
0.0921630859375,
0.88818359375,
0.106689453125,
0.69677734375,
0.1060791015625
],
[
0.5048828125,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.5048828125,
0.121826171875
],
[
0.1004638671875,
0.117431640625,
0.46728515625,
0.117431640625,
0.46728515625,
0.1307373046875,
0.1004638671875,
0.1307373046875
],
[
0.50244140625,
0.12249755859375,
0.8896484375,
0.12371826171875,
0.8896484375,
0.136962890625,
0.50244140625,
0.1363525390625
],
[
0.09478759765625,
0.1307373046875,
0.468017578125,
0.1319580078125,
0.468017578125,
0.146484375,
0.09478759765625,
0.1458740234375
],
[
0.5087890625,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.5087890625,
0.1514892578125
],
[
0.080078125,
0.146484375,
0.46728515625,
0.147705078125,
0.46728515625,
0.1610107421875,
0.080078125,
0.160400390625
],
[
0.50341796875,
0.1514892578125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1673583984375,
0.50341796875,
0.1666259765625
],
[
0.080078125,
0.1634521484375,
0.468017578125,
0.1622314453125,
0.468017578125,
0.175537109375,
0.080078125,
0.1767578125
],
[
0.5048828125,
0.1685791015625,
0.8896484375,
0.1685791015625,
0.8896484375,
0.1817626953125,
0.5048828125,
0.1817626953125
],
[
0.08087158203125,
0.177978515625,
0.281982421875,
0.177978515625,
0.281982421875,
0.1912841796875,
0.08087158203125,
0.1912841796875
],
[
0.5048828125,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.5048828125,
0.1962890625
],
[
0.50634765625,
0.199462890625,
0.8896484375,
0.199462890625,
0.8896484375,
0.210205078125,
0.50634765625,
0.210205078125
],
[
0.19775390625,
0.2108154296875,
0.2099609375,
0.2108154296875,
0.2099609375,
0.2171630859375,
0.19775390625,
0.2171630859375
],
[
0.50732421875,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.50732421875,
0.2266845703125
],
[
0.11358642578125,
0.2197265625,
0.468017578125,
0.2183837890625,
0.468017578125,
0.236083984375,
0.11358642578125,
0.2374267578125
],
[
0.5048828125,
0.2291259765625,
0.80224609375,
0.2291259765625,
0.80224609375,
0.242431640625,
0.5048828125,
0.242431640625
],
[
0.191162109375,
0.236083984375,
0.2156982421875,
0.236083984375,
0.2156982421875,
0.248779296875,
0.191162109375,
0.248779296875
],
[
0.53369140625,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.53369140625,
0.2568359375
],
[
0.5048828125,
0.2587890625,
0.890625,
0.2587890625,
0.890625,
0.272705078125,
0.5048828125,
0.272705078125
],
[
0.09722900390625,
0.266357421875,
0.46728515625,
0.266357421875,
0.46728515625,
0.276611328125,
0.09722900390625,
0.276611328125
],
[
0.08087158203125,
0.27978515625,
0.468017578125,
0.27978515625,
0.468017578125,
0.294189453125,
0.08087158203125,
0.294189453125
],
[
0.5048828125,
0.27587890625,
0.890625,
0.27587890625,
0.890625,
0.285888671875,
0.5048828125,
0.285888671875
],
[
0.50244140625,
0.28857421875,
0.8935546875,
0.2890625,
0.8935546875,
0.3037109375,
0.50244140625,
0.302978515625
],
[
0.080078125,
0.29541015625,
0.3544921875,
0.29541015625,
0.3544921875,
0.309326171875,
0.080078125,
0.309326171875
],
[
0.50341796875,
0.304931640625,
0.7490234375,
0.304931640625,
0.7490234375,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.5107421875,
0.322509765625,
0.8896484375,
0.322509765625,
0.8896484375,
0.3359375,
0.5107421875,
0.3359375
],
[
0.080078125,
0.330078125,
0.2181396484375,
0.330078125,
0.2181396484375,
0.343505859375,
0.080078125,
0.343505859375
],
[
0.50341796875,
0.3359375,
0.890625,
0.337158203125,
0.890625,
0.3515625,
0.50341796875,
0.35107421875
],
[
0.098876953125,
0.35595703125,
0.468017578125,
0.354736328125,
0.468017578125,
0.36865234375,
0.098876953125,
0.369384765625
],
[
0.5048828125,
0.352783203125,
0.890625,
0.352783203125,
0.890625,
0.3662109375,
0.5048828125,
0.3662109375
],
[
0.08251953125,
0.37060546875,
0.4697265625,
0.37060546875,
0.4697265625,
0.385009765625,
0.08251953125,
0.385009765625
],
[
0.50341796875,
0.36865234375,
0.890625,
0.36865234375,
0.890625,
0.3818359375,
0.50341796875,
0.3818359375
],
[
0.08087158203125,
0.385009765625,
0.468017578125,
0.385009765625,
0.468017578125,
0.3984375,
0.08087158203125,
0.3984375
],
[
0.50341796875,
0.38330078125,
0.890625,
0.3837890625,
0.890625,
0.397705078125,
0.50341796875,
0.396484375
],
[
0.07843017578125,
0.40087890625,
0.468017578125,
0.401611328125,
0.468017578125,
0.415283203125,
0.07843017578125,
0.4140625
],
[
0.50341796875,
0.397705078125,
0.890625,
0.397705078125,
0.890625,
0.410888671875,
0.50341796875,
0.410888671875
],
[
0.080078125,
0.417236328125,
0.32177734375,
0.417236328125,
0.32177734375,
0.43115234375,
0.080078125,
0.43115234375
],
[
0.5048828125,
0.412353515625,
0.890625,
0.412353515625,
0.890625,
0.4267578125,
0.5048828125,
0.4267578125
],
[
0.5048828125,
0.4287109375,
0.890625,
0.4287109375,
0.890625,
0.442626953125,
0.5048828125,
0.442626953125
],
[
0.08087158203125,
0.4462890625,
0.27783203125,
0.4462890625,
0.27783203125,
0.460205078125,
0.08087158203125,
0.460205078125
],
[
0.5048828125,
0.4443359375,
0.55224609375,
0.4443359375,
0.55224609375,
0.455810546875,
0.5048828125,
0.455810546875
],
[
0.52294921875,
0.458984375,
0.890625,
0.458984375,
0.890625,
0.472900390625,
0.52294921875,
0.472900390625
],
[
0.0882568359375,
0.4697265625,
0.181396484375,
0.4697265625,
0.181396484375,
0.479736328125,
0.0882568359375,
0.479736328125
],
[
0.176513671875,
0.468505859375,
0.46728515625,
0.468505859375,
0.46728515625,
0.481689453125,
0.176513671875,
0.481689453125
],
[
0.50341796875,
0.473388671875,
0.8896484375,
0.473388671875,
0.8896484375,
0.4873046875,
0.50341796875,
0.4873046875
],
[
0.080078125,
0.481689453125,
0.468017578125,
0.484130859375,
0.468017578125,
0.49755859375,
0.080078125,
0.49560546875
],
[
0.50244140625,
0.488037109375,
0.88818359375,
0.4892578125,
0.88818359375,
0.50390625,
0.50244140625,
0.50244140625
],
[
0.08251953125,
0.5,
0.46728515625,
0.5,
0.46728515625,
0.51025390625,
0.08251953125,
0.51025390625
],
[
0.50341796875,
0.50390625,
0.890625,
0.50390625,
0.890625,
0.517578125,
0.50341796875,
0.517578125
],
[
0.08251953125,
0.51318359375,
0.468017578125,
0.51318359375,
0.468017578125,
0.5263671875,
0.08251953125,
0.5263671875
],
[
0.5048828125,
0.51953125,
0.88232421875,
0.51953125,
0.88232421875,
0.53271484375,
0.5048828125,
0.53271484375
],
[
0.08251953125,
0.52880859375,
0.468017578125,
0.52880859375,
0.468017578125,
0.54248046875,
0.08251953125,
0.54248046875
],
[
0.5048828125,
0.5361328125,
0.8896484375,
0.5361328125,
0.8896484375,
0.546875,
0.5048828125,
0.546875
],
[
0.08087158203125,
0.54345703125,
0.468017578125,
0.54345703125,
0.468017578125,
0.556640625,
0.08087158203125,
0.556640625
],
[
0.50341796875,
0.5498046875,
0.88818359375,
0.5498046875,
0.88818359375,
0.564453125,
0.50341796875,
0.564453125
],
[
0.08251953125,
0.5595703125,
0.468017578125,
0.5595703125,
0.468017578125,
0.57275390625,
0.08251953125,
0.57275390625
],
[
0.50244140625,
0.5693359375,
0.5986328125,
0.5712890625,
0.5986328125,
0.5859375,
0.50244140625,
0.583984375
],
[
0.08416748046875,
0.57568359375,
0.46728515625,
0.57568359375,
0.46728515625,
0.5859375,
0.08416748046875,
0.5859375
],
[
0.08087158203125,
0.5888671875,
0.30810546875,
0.5888671875,
0.30810546875,
0.60302734375,
0.08087158203125,
0.60302734375
],
[
0.52392578125,
0.59521484375,
0.890625,
0.59521484375,
0.890625,
0.609375,
0.52392578125,
0.609375
],
[
0.09478759765625,
0.61572265625,
0.465576171875,
0.61572265625,
0.465576171875,
0.6298828125,
0.09478759765625,
0.6298828125
],
[
0.5048828125,
0.60986328125,
0.890625,
0.60986328125,
0.890625,
0.62353515625,
0.5048828125,
0.62353515625
],
[
0.1094970703125,
0.6298828125,
0.465576171875,
0.6298828125,
0.465576171875,
0.64306640625,
0.1094970703125,
0.64306640625
],
[
0.5048828125,
0.62548828125,
0.89208984375,
0.62548828125,
0.89208984375,
0.6396484375,
0.5048828125,
0.6396484375
],
[
0.111083984375,
0.64599609375,
0.465576171875,
0.64599609375,
0.465576171875,
0.6591796875,
0.111083984375,
0.6591796875
],
[
0.50341796875,
0.64013671875,
0.82177734375,
0.64013671875,
0.82177734375,
0.65380859375,
0.50341796875,
0.65380859375
],
[
0.111083984375,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.111083984375,
0.673828125
],
[
0.5048828125,
0.6591796875,
0.6650390625,
0.66015625,
0.6650390625,
0.6748046875,
0.5048828125,
0.673828125
],
[
0.67626953125,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.67626953125,
0.673828125
],
[
0.11273193359375,
0.6767578125,
0.468017578125,
0.6767578125,
0.468017578125,
0.68994140625,
0.11273193359375,
0.68994140625
],
[
0.5048828125,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.5048828125,
0.689453125
],
[
0.11273193359375,
0.69140625,
0.253173828125,
0.69140625,
0.253173828125,
0.705078125,
0.11273193359375,
0.705078125
],
[
0.50341796875,
0.68994140625,
0.88818359375,
0.69140625,
0.88818359375,
0.705078125,
0.50341796875,
0.7041015625
],
[
0.096435546875,
0.70703125,
0.465576171875,
0.70703125,
0.465576171875,
0.72021484375,
0.096435546875,
0.72021484375
],
[
0.50341796875,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.111083984375,
0.72265625,
0.46728515625,
0.72265625,
0.46728515625,
0.736328125,
0.111083984375,
0.736328125
],
[
0.5048828125,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.5048828125,
0.734375
],
[
0.111083984375,
0.7373046875,
0.465576171875,
0.7373046875,
0.465576171875,
0.75048828125,
0.111083984375,
0.75048828125
],
[
0.5048828125,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.5048828125,
0.75
],
[
0.11358642578125,
0.75244140625,
0.468017578125,
0.75244140625,
0.468017578125,
0.7666015625,
0.11358642578125,
0.7666015625
],
[
0.50634765625,
0.75244140625,
0.8896484375,
0.75244140625,
0.8896484375,
0.76318359375,
0.50634765625,
0.76318359375
],
[
0.11273193359375,
0.767578125,
0.464111328125,
0.767578125,
0.464111328125,
0.78076171875,
0.11273193359375,
0.78076171875
],
[
0.50634765625,
0.76806640625,
0.8896484375,
0.76806640625,
0.8896484375,
0.77880859375,
0.50634765625,
0.77880859375
],
[
0.11358642578125,
0.7841796875,
0.465576171875,
0.7841796875,
0.465576171875,
0.79443359375,
0.11358642578125,
0.79443359375
],
[
0.50341796875,
0.7822265625,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.50341796875,
0.79541015625
],
[
0.11273193359375,
0.798828125,
0.2646484375,
0.798828125,
0.2646484375,
0.8125,
0.11273193359375,
0.8125
],
[
0.50244140625,
0.79541015625,
0.88818359375,
0.79736328125,
0.88818359375,
0.81103515625,
0.50244140625,
0.8095703125
],
[
0.5048828125,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.5048828125,
0.82470703125
],
[
0.1004638671875,
0.8291015625,
0.468017578125,
0.8291015625,
0.468017578125,
0.84228515625,
0.1004638671875,
0.84228515625
],
[
0.50341796875,
0.8271484375,
0.79736328125,
0.8271484375,
0.79736328125,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.08416748046875,
0.84521484375,
0.46728515625,
0.84521484375,
0.46728515625,
0.85595703125,
0.08416748046875,
0.85595703125
],
[
0.52294921875,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85595703125,
0.52294921875,
0.85595703125
],
[
0.08087158203125,
0.859375,
0.465576171875,
0.859375,
0.465576171875,
0.87255859375,
0.08087158203125,
0.87255859375
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.096435546875,
0.875,
0.46728515625,
0.87548828125,
0.46728515625,
0.8896484375,
0.096435546875,
0.88818359375
],
[
0.50341796875,
0.87060546875,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.50341796875,
0.88525390625
],
[
0.08087158203125,
0.88818359375,
0.253173828125,
0.8896484375,
0.253173828125,
0.90087890625,
0.08087158203125,
0.8994140625
],
[
0.50244140625,
0.88623046875,
0.890625,
0.88818359375,
0.890625,
0.9013671875,
0.50244140625,
0.8994140625
],
[
0.4794921875,
0.92529296875,
0.49267578125,
0.92529296875,
0.49267578125,
0.9375,
0.4794921875,
0.9375
]
] |
[
"2.3. Incorporating class-ratio priors",
"Network architectures.",
"我们使用Deeplab-V2日[2",
"作为基本语义分割体系结构F。为了更好地捕捉场景",
"摘最小化可能会偏向一些简单的类。因此，有时用",
"上下文，将Atrous空间金字塔池（ASPP）应用于最后",
"些先验来指导学习是有益的。为此，我们基于类在源",
"一层的特征输出。采样率固定为[6,12,18,24]，类似",
"标签上的分布使用简单的类先验。我们将类先验向量",
"于E[2]中的ASPP-L模型。我们在两种不同的基础深",
"Ps计算为源标签上每类像素数的l1归一化直方图。现",
"度CNN架构上进行了实验：VGG-16E[37]和ResNet-",
"在，基于预测的Pat，使用",
"101E[13]。在E[2]之后，我们修改了最后一层的步幅",
"和膨胀率，以生成具有更大视场的更密集的特征图。为",
"C",
"了进一步提高ResNet-101的性能，我们对来自conv4",
"Lep(at) =max (0, μp(c) -Ec(P(c), (10)",
"和conv5功能的多级输出执行自适应【41]",
"c=1",
"2.2节中介绍的对抗网络D与DCGAN28中使",
"用的网络具有相同的体系结构。加权自信息映射I通",
"其中uE0.1用于放松类先验约束，任何类的预期",
"概率与类先验P：之间的差异太大。这解决了单个目标",
"过4卷积层转发，每个卷积层都与具有固定负斜率0.2",
"的leaky-ReLU层耦合。最后，分类器层生成分类输出，",
"映像上的类分布不一定接近p。的事实",
"指示输入是对应于源域还是目标域",
"Implementationdetails：我们在实现中采用Py-",
"3.Experiments",
"Torch深度学习框架[27]。所有实验均在具有11GB",
"在本节中，我们将介绍我们的实验结果。第3.1",
"内存的单个NVIDIA1080TIGPU上完成。我们的模",
"节介绍了使用的数据集以及我们的训练参数。在3.2",
"型，除了E2.2节中提到的对抗性鉴别器外，都是使",
"节和3.3节中，我们报告并讨论了我们的主要结果。",
"用随机梯度下降优化器E[1]进行训练的，其学习率为",
"In Section3.3,wediscuss apreliminaryresult on",
"2.5×10-4，动量为0.9，权重为10-4。我们使用具",
"entropy-based UDA for detection.",
"有学习率10-4的Adam优化器E[19]来训练鉴别器。",
"为了调度学习速率，我们遵循[2]中提到的多项式退",
"3.1.Experimental details",
"火过程",
"摘和对抗损失的加权因子：要为Lent设置权重，训",
"Datasets.",
"为了评估我们的方法，我们使用具有挑",
"练集性能提供了重要的指示。如果入ent很大，则摘下",
"战性的合成-2-实数无监督域自适应设置。模型在完全",
"降得太快，并且模型强烈偏向于几个类。然而，当入ent",
"注释的合成数据上进行训练，并在现实世界的数据上",
"选择在合适的范围内时，性能更好，并且对精确值不敏",
"进行验证。在这样的设置中，模型在训练期间可以访",
"感。因此，我们对所有实验使用相同的入ent=0.001",
"问一些未标记的真实图像。为了训练我们的模型，我",
"而不管网络或数据集如何。类似的参数适用于（9）中的",
"们使用GTA5[30或SYNTHIA311作为源域合成",
"权重入adu。我们在所有实验中都修复了入adu=0.001",
"数据，同时使用Cityscapes数据集[6]的训练分割作",
"3.2. Results",
"为自标域数据。以前在其他作品中也使用过类似的设",
"置【15，14，41，51]。详细说明：",
"我们提供了与不同基线相比的方法的实验结果。我",
". GTA5 → Cityscapes: The GTA5 dataset con",
"们的模型在两个UDA基准测试中实现了最先进的性",
"sists of 24, 966 synthesized frames captured from",
"能。在下面的内容中，我们展示了我们的方法在不同设",
"a video game. Images are provided with pixel-",
"置，i.e.，训练集和基本CNN中的不同行为",
"level semantic annotations of 33 classes. Similar",
"GTA5 → Cityscapes:",
"我们在表E1中报告-在城市景",
"to [15], we use the 19 classes in common with the",
"观验证集上以mloU（%）表示的语义分割性能。我们的",
"Cityscapes dataset.",
"第一种直接熵最小化方法，在表1-a中称为MinEnt",
"• SYNTHIA -→ Cityscapes: We use the SYNTHIA-",
"在基于 VGG-16 和 ResNet-101的 CNN上实现了与最",
"RAND-CITYSCAPES set3 with 9,400 synthe-",
"先进基线相当的性能。MinEnt在没有和具有类平衡E",
"sized images for training.We train our mod-",
"[51]的情况下优于自我训练（ST）方法。与E[41]相",
"els with 16 common classes in SYNTHIA and",
"比，基于ResNet-101的MinEnt显示出相似的结果",
"Cityscapes. While evaluating we compare the per",
"但没有求助于鉴别器网络的训练。损失的轻开销使",
"formanceon16-and13-classsubsetsfollowingthe",
"得MinEnt模型的训练时间要少得多。我们的方法",
"protocol used in [51].",
"的另一个优点是易于训练。事实上，由于其不稳定性",
"训练对抗网络通常被认为是一项艰巨的任务。我们观",
"在这两个设置中，2,975未标记的Cityscapes图像",
"察到一个更稳定的行为训练模型与炳损失",
"用于训练。我们使用标准的平均交点超过联合（mIoU）",
"有趣的是，我们发现在某些情况下，仅在某些范围",
"度量E[9]来衡量分割性能。对500验证映像进行评估",
"上应用摘损失效果最好。使用基于ResNet-101的模型",
"3A split of the SYNTHIAdataset[3] using compatible labels",
"可以观察到这种现象。实际上，我们通过对每个目标",
"with the Cityscapes dataset.",
"样本的顶部30%内具有嫡值的像素进行训练来获得更",
"4"
] |
{
"arxiv_id": "1811.12833",
"language": "zh",
"page_idx": 3
}
|
||
2108.11550-zh-1
|
2108.11550-zh-1.jpg
|
The Surprising Effectiveness of Visual Odometry Techniques
for Embodied PointGoal Navigation
Xiaoming Zhao$^\dagger$,Harsh Agrawal$^\dagger$,Dhruv Batra$^{\dagger,\S}$, Alexander Schwing$^\dagger$
$^{\dagger}$ University of Illinois, Urbana-Champaign $^{\dagger}$ Georgia Institute of Technology
$^{6}$ Facebook AI Research
https://xiaoming-zhao.github.io/projects/pointnav-vo/
## 1. Introduction
在室内环境中高效，准确地导航的能力是个人机器人的基础，多年来一直是计算机视觉研究的重点$\mathbb{E}[37]$ 。为了围绕一个共同的框架和标准度量来凝聚社区。Anderson et al. [2] 提出了 PointGoal 导航的任务。在PointGoal 导航中，代理在以前看不见的环境中随机生成，并且必须导航到相对于代理的初始位置和方向指定的点目标 e.g. , '相对于 start 向北 5米，向西 3 米”。代理使用离散动作空间( e.g. , ???。25 $m$ 、turn\_left 或 turn\_right 30° 和 stop)在环境中导航。在无噪声自我中心视觉(无噪声 RGB+ 深度传感器)、无噪声驱动(e.g. , turn\_left 将始终精确地转动 30°)和使
$^{1}$https://eval.ai/web/challenges/challenge-page/580/
$leaderboard/ 1631\allowbreak ( {\mathrm{mod}} $Habitat Team).
$\operatorname{Abstract}$
It is fundamental for personal robots to reliably navigate to a specified goal. To study this task, PointGoal navigation has been introduced in simulated Embodied AI environments. Recent advances solve this PointGoal navigation task with near-perfect accuracy (99.6 % success) in photo-realistically simulated environments, assuming noiseless egocentric vision, noiseless actuation and most importantly, perfect localization. However, under realistic noise models for visual sensors and actuation, and without access to a“GPS and Compass sensor," the 99.6% -success agents for PointGoal navigation only succeed with $0.3\%.^{1}$ In this work, we demonstrate the surprising effectiveness of visual odometry for the task of PointGoal navigation in this realistic setting, i.e., with realistic noise models for perception and actuation and without access to GPS and Compass sensors. We show that integrating visual odometry techniques into navigation policies improves the state-of-the-art on the popular Habitat PointNav benchmark by a large margin, improving success from $64.5\%$ to $71.7\%$ while executing 6.4 times faster.
<FigureHere>
Figure 1: Noiseless (a) and noisy (b) PointGoal navigation. In the noisy setting, the agent observes: 1) sensor noises in egocentric observation; 2) actuation perturbations. The second column shows a histogram of orientation angle changes caused by a turn\_left action; 3) no localization information. The agent's inaccurate localization results in uncertainty about the goal location.
用 GPS+Compass 传感器的完美定位的假设下，最近的方法以近乎完美的精度解决了这一任务(99。6% 成功 [53]
然而，这些假设是不现实的。请注意，GPS 传感器通常不会在室内环境中产生精确的位置。此外，真实机器人的感知和驱动通常在很大程度上取决于环境照明和表面的摩擦系数。为了研究这种更现实的设置，在最近的基准测试$^{2}$中，更新了PointGoal 导航，以包括来自真实机器人的嘈杂驱动模型[35]。例如，对于单个 turn\_left 动作，实际转弯角度会有很大变化，如$\operatorname{Fig.~1}$的第二列所示。此外，还结合了$\operatorname{E}[9]$ 的 RGB 和深度噪声模型来模拟现实世界的相机。最重要的是，如Fig. 1 的第三列所示，代理无法访问 GPS+Compass 数据，必须仅基于自我中心的 RGB+ 深度 (RGB-D) 测量进行导航。在这样一个更现实的设置下，一个在无噪声场景中近乎完美的策略的性能 [53] 急剧下降到$0。3\%$。对其进行改进，先前的最新技术$\mathbb{E}[24]$ 将粒子SLAM 集成到视觉导航中，并实现了 64 的成功率。5这样逼真的设置下的%。相比 99。6% 在无噪声版本的任务上的成功率，具有噪声感知和驱动以及没有本地化信息的导航仍然具有挑战性为了更好地了解在这种现实环境中导航的挑战，我们研究了三种视觉里程计(VO) 技术。我们发现这些 VO 技术在这个现实的环境
${^2\mathrm{https:}//\mathrm{aihabitat.org}/\mathrm{challenge}/2020/}$
|
[
[
0.181396484375,
0.1331787109375,
0.7890625,
0.1351318359375,
0.7890625,
0.149658203125,
0.181396484375,
0.147705078125
],
[
0.30224609375,
0.1546630859375,
0.6689453125,
0.1558837890625,
0.6689453125,
0.173583984375,
0.30224609375,
0.17236328125
],
[
0.1683349609375,
0.2052001953125,
0.8037109375,
0.2052001953125,
0.8037109375,
0.2197265625,
0.1683349609375,
0.2197265625
],
[
0.1094970703125,
0.222900390625,
0.406005859375,
0.2235107421875,
0.406005859375,
0.238037109375,
0.1094970703125,
0.2374267578125
],
[
0.43310546875,
0.2247314453125,
0.66015625,
0.2247314453125,
0.66015625,
0.238037109375,
0.43310546875,
0.238037109375
],
[
0.69677734375,
0.222900390625,
0.861328125,
0.2235107421875,
0.861328125,
0.238037109375,
0.69677734375,
0.2374267578125
],
[
0.30810546875,
0.242431640625,
0.67626953125,
0.242431640625,
0.67626953125,
0.255615234375,
0.30810546875,
0.255615234375
],
[
0.54736328125,
0.2734375,
0.62158203125,
0.27587890625,
0.62158203125,
0.290283203125,
0.54736328125,
0.28857421875
],
[
0.6806640625,
0.2734375,
0.74267578125,
0.27587890625,
0.74267578125,
0.290283203125,
0.6806640625,
0.28857421875
],
[
0.79736328125,
0.276611328125,
0.869140625,
0.276611328125,
0.869140625,
0.28857421875,
0.79736328125,
0.28857421875
],
[
0.23779296875,
0.292236328125,
0.313720703125,
0.292236328125,
0.313720703125,
0.306884765625,
0.23779296875,
0.306884765625
],
[
0.5087890625,
0.29345703125,
0.52685546875,
0.29345703125,
0.52685546875,
0.300537109375,
0.5087890625,
0.300537109375
],
[
0.78369140625,
0.291748046875,
0.833984375,
0.291748046875,
0.833984375,
0.302978515625,
0.78369140625,
0.302978515625
],
[
0.85888671875,
0.29541015625,
0.87158203125,
0.29541015625,
0.87158203125,
0.304931640625,
0.85888671875,
0.304931640625
],
[
0.78515625,
0.3037109375,
0.8310546875,
0.3037109375,
0.8310546875,
0.311279296875,
0.78515625,
0.311279296875
],
[
0.7861328125,
0.311279296875,
0.83251953125,
0.311279296875,
0.83251953125,
0.322509765625,
0.7861328125,
0.322509765625
],
[
0.098876953125,
0.324462890625,
0.468017578125,
0.324462890625,
0.468017578125,
0.33837890625,
0.098876953125,
0.33837890625
],
[
0.08087158203125,
0.339111328125,
0.46728515625,
0.339111328125,
0.46728515625,
0.352783203125,
0.08087158203125,
0.352783203125
],
[
0.66015625,
0.3359375,
0.76123046875,
0.337158203125,
0.76123046875,
0.347900390625,
0.66015625,
0.3466796875
],
[
0.080078125,
0.354736328125,
0.468017578125,
0.354248046875,
0.468017578125,
0.36865234375,
0.080078125,
0.369384765625
],
[
0.5087890625,
0.354736328125,
0.529296875,
0.356689453125,
0.52783203125,
0.3662109375,
0.50732421875,
0.3642578125
],
[
0.78515625,
0.35595703125,
0.833984375,
0.35595703125,
0.833984375,
0.367431640625,
0.78515625,
0.367431640625
],
[
0.85693359375,
0.35791015625,
0.87158203125,
0.35791015625,
0.87158203125,
0.3662109375,
0.85693359375,
0.3662109375
],
[
0.78173828125,
0.365478515625,
0.833984375,
0.365478515625,
0.833984375,
0.375732421875,
0.78173828125,
0.375732421875
],
[
0.08251953125,
0.371826171875,
0.46728515625,
0.371826171875,
0.46728515625,
0.3818359375,
0.08251953125,
0.3818359375
],
[
0.8544921875,
0.37060546875,
0.86376953125,
0.367431640625,
0.86669921875,
0.371826171875,
0.85693359375,
0.373779296875
],
[
0.79833984375,
0.376953125,
0.82177734375,
0.376953125,
0.82177734375,
0.385009765625,
0.79833984375,
0.385009765625
],
[
0.08251953125,
0.386962890625,
0.46728515625,
0.386962890625,
0.46728515625,
0.397705078125,
0.08251953125,
0.397705078125
],
[
0.08087158203125,
0.40087890625,
0.46728515625,
0.40087890625,
0.46728515625,
0.4140625,
0.08087158203125,
0.4140625
],
[
0.08087158203125,
0.416748046875,
0.468017578125,
0.416748046875,
0.468017578125,
0.429931640625,
0.08087158203125,
0.429931640625
],
[
0.50341796875,
0.417236328125,
0.88818359375,
0.417236328125,
0.88818359375,
0.431884765625,
0.50341796875,
0.431884765625
],
[
0.080078125,
0.43115234375,
0.468017578125,
0.43115234375,
0.468017578125,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.50244140625,
0.431884765625,
0.890625,
0.43310546875,
0.890625,
0.4443359375,
0.50244140625,
0.443115234375
],
[
0.080078125,
0.44580078125,
0.4697265625,
0.4443359375,
0.4697265625,
0.458251953125,
0.080078125,
0.458984375
],
[
0.50341796875,
0.447509765625,
0.8896484375,
0.447509765625,
0.8896484375,
0.458251953125,
0.50341796875,
0.458251953125
],
[
0.080078125,
0.460205078125,
0.46728515625,
0.460205078125,
0.46728515625,
0.474853515625,
0.080078125,
0.474853515625
],
[
0.50341796875,
0.460205078125,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50341796875,
0.473388671875
],
[
0.080078125,
0.47607421875,
0.468017578125,
0.47607421875,
0.468017578125,
0.4892578125,
0.080078125,
0.4892578125
],
[
0.50341796875,
0.473388671875,
0.890625,
0.473388671875,
0.890625,
0.4873046875,
0.50341796875,
0.4873046875
],
[
0.07843017578125,
0.490478515625,
0.46728515625,
0.4892578125,
0.46728515625,
0.50390625,
0.07843017578125,
0.5048828125
],
[
0.50244140625,
0.486083984375,
0.890625,
0.4873046875,
0.890625,
0.50048828125,
0.50244140625,
0.5
],
[
0.07843017578125,
0.505859375,
0.468017578125,
0.5048828125,
0.468017578125,
0.51953125,
0.07843017578125,
0.52099609375
],
[
0.50244140625,
0.50048828125,
0.7060546875,
0.50048828125,
0.7060546875,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.080078125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.53515625,
0.080078125,
0.53515625
],
[
0.50341796875,
0.5263671875,
0.890625,
0.5263671875,
0.890625,
0.54052734375,
0.50341796875,
0.54052734375
],
[
0.080078125,
0.5361328125,
0.468017578125,
0.53515625,
0.468017578125,
0.5498046875,
0.080078125,
0.55029296875
],
[
0.50341796875,
0.54248046875,
0.89208984375,
0.54248046875,
0.89208984375,
0.55615234375,
0.50341796875,
0.55615234375
],
[
0.08087158203125,
0.5517578125,
0.468017578125,
0.5517578125,
0.468017578125,
0.56494140625,
0.08087158203125,
0.56494140625
],
[
0.50244140625,
0.556640625,
0.56396484375,
0.556640625,
0.56396484375,
0.5712890625,
0.50244140625,
0.5712890625
],
[
0.08251953125,
0.568359375,
0.46728515625,
0.568359375,
0.46728515625,
0.5791015625,
0.08251953125,
0.5791015625
],
[
0.52294921875,
0.5712890625,
0.890625,
0.5712890625,
0.890625,
0.5859375,
0.52294921875,
0.5859375
],
[
0.08087158203125,
0.58056640625,
0.46728515625,
0.58056640625,
0.46728515625,
0.5947265625,
0.08087158203125,
0.5947265625
],
[
0.5048828125,
0.58837890625,
0.890625,
0.58837890625,
0.890625,
0.6015625,
0.5048828125,
0.6015625
],
[
0.08087158203125,
0.59521484375,
0.46728515625,
0.59521484375,
0.46728515625,
0.609375,
0.08087158203125,
0.609375
],
[
0.50341796875,
0.60302734375,
0.890625,
0.60302734375,
0.890625,
0.6162109375,
0.50341796875,
0.6162109375
],
[
0.080078125,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.080078125,
0.62548828125
],
[
0.5048828125,
0.61865234375,
0.890625,
0.61865234375,
0.890625,
0.6318359375,
0.5048828125,
0.6318359375
],
[
0.080078125,
0.626953125,
0.43701171875,
0.626953125,
0.43701171875,
0.64013671875,
0.080078125,
0.64013671875
],
[
0.50341796875,
0.6318359375,
0.8896484375,
0.6318359375,
0.8896484375,
0.646484375,
0.50341796875,
0.646484375
],
[
0.5048828125,
0.6484375,
0.8896484375,
0.6484375,
0.8896484375,
0.662109375,
0.5048828125,
0.662109375
],
[
0.50341796875,
0.66357421875,
0.8896484375,
0.66357421875,
0.8896484375,
0.6767578125,
0.50341796875,
0.6767578125
],
[
0.080078125,
0.67236328125,
0.21728515625,
0.67236328125,
0.21728515625,
0.6875,
0.080078125,
0.6875
],
[
0.50341796875,
0.6787109375,
0.890625,
0.6787109375,
0.890625,
0.6923828125,
0.50341796875,
0.6923828125
],
[
0.1004638671875,
0.69873046875,
0.468017578125,
0.69873046875,
0.468017578125,
0.7119140625,
0.1004638671875,
0.7119140625
],
[
0.5048828125,
0.6943359375,
0.890625,
0.6943359375,
0.890625,
0.70849609375,
0.5048828125,
0.70849609375
],
[
0.08251953125,
0.71337890625,
0.46728515625,
0.71337890625,
0.46728515625,
0.72802734375,
0.08251953125,
0.72802734375
],
[
0.50341796875,
0.708984375,
0.890625,
0.708984375,
0.890625,
0.7236328125,
0.50341796875,
0.7236328125
],
[
0.0931396484375,
0.72900390625,
0.46728515625,
0.72900390625,
0.46728515625,
0.7421875,
0.0931396484375,
0.7421875
],
[
0.5048828125,
0.7236328125,
0.890625,
0.7236328125,
0.890625,
0.7373046875,
0.5048828125,
0.7373046875
],
[
0.0906982421875,
0.73486328125,
0.09722900390625,
0.73486328125,
0.09722900390625,
0.73779296875,
0.0906982421875,
0.73779296875
],
[
0.08087158203125,
0.74365234375,
0.468017578125,
0.74365234375,
0.468017578125,
0.7568359375,
0.08087158203125,
0.7568359375
],
[
0.5048828125,
0.7392578125,
0.890625,
0.7392578125,
0.890625,
0.75244140625,
0.5048828125,
0.75244140625
],
[
0.08087158203125,
0.7587890625,
0.468017578125,
0.7587890625,
0.468017578125,
0.77294921875,
0.08087158203125,
0.77294921875
],
[
0.5048828125,
0.75390625,
0.8896484375,
0.75390625,
0.8896484375,
0.767578125,
0.5048828125,
0.767578125
],
[
0.08087158203125,
0.7734375,
0.468017578125,
0.7734375,
0.468017578125,
0.787109375,
0.08087158203125,
0.787109375
],
[
0.50244140625,
0.76953125,
0.8896484375,
0.76953125,
0.8896484375,
0.78271484375,
0.50244140625,
0.78271484375
],
[
0.50244140625,
0.78271484375,
0.89208984375,
0.7841796875,
0.89208984375,
0.798828125,
0.50244140625,
0.79736328125
],
[
0.08087158203125,
0.7890625,
0.19287109375,
0.7890625,
0.19287109375,
0.80322265625,
0.08087158203125,
0.80322265625
],
[
0.200927734375,
0.7890625,
0.4697265625,
0.7890625,
0.4697265625,
0.80322265625,
0.200927734375,
0.80322265625
],
[
0.08251953125,
0.8037109375,
0.468017578125,
0.8037109375,
0.468017578125,
0.818359375,
0.08251953125,
0.818359375
],
[
0.50341796875,
0.7998046875,
0.890625,
0.7998046875,
0.890625,
0.81298828125,
0.50341796875,
0.81298828125
],
[
0.08087158203125,
0.8193359375,
0.4697265625,
0.8193359375,
0.4697265625,
0.83349609375,
0.08087158203125,
0.83349609375
],
[
0.5048828125,
0.81640625,
0.8896484375,
0.81640625,
0.8896484375,
0.8271484375,
0.5048828125,
0.8271484375
],
[
0.08416748046875,
0.833984375,
0.468017578125,
0.833984375,
0.468017578125,
0.84765625,
0.08416748046875,
0.84765625
],
[
0.5048828125,
0.830078125,
0.890625,
0.830078125,
0.890625,
0.84326171875,
0.5048828125,
0.84326171875
],
[
0.080078125,
0.849609375,
0.468017578125,
0.8486328125,
0.468017578125,
0.86279296875,
0.080078125,
0.8642578125
],
[
0.5048828125,
0.84521484375,
0.88818359375,
0.84521484375,
0.88818359375,
0.859375,
0.5048828125,
0.859375
],
[
0.50244140625,
0.85986328125,
0.89208984375,
0.85986328125,
0.89208984375,
0.87353515625,
0.50244140625,
0.87353515625
],
[
0.096435546875,
0.87548828125,
0.420654296875,
0.87548828125,
0.420654296875,
0.8896484375,
0.096435546875,
0.8896484375
],
[
0.080078125,
0.88720703125,
0.293212890625,
0.88720703125,
0.293212890625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5185546875,
0.88720703125,
0.7587890625,
0.88720703125,
0.7587890625,
0.90087890625,
0.5185546875,
0.90087890625
]
] |
[
"The Surprising Effectiveness of Visual Odometry Techniques",
"for Embodied PointGoal Navigation",
"Xiaoming Zhao t, Harsh Agrawal , Dhruv Batra t,s,Alexander Schwing t",
"t University of Illinois, Urbana-Champaign",
"+ Georgia Institute of Technology",
"$FacebookAIResearch",
"https://xiaoming-zhao.github.io/projects/pointnav-vo/",
"Observation",
"Actuation",
"Localization",
"Abstract",
"(a)",
"Point-Goal",
"O",
"coordinate",
"[0.3, 0.4] ",
"It is fundamental for personal robots to reliably nav",
"igate to a specified goal. To study this task, PointGoal",
"1520 2530 3540 4550",
"navigation has been introduced in simulated Embodied",
"(b)",
"Point-Goal",
"O",
"coordinate",
"AIenvironments.RecentadvancessolvethisPointGoal",
"",
"[2.?]",
"navigationtaskwithnear-perfectaccuracy(99.6%suc",
"cess) in photo-realistically simulated environments, as",
"suming noiseless egocentric vision, noiseless actuation",
"Figure l: Noiseless (a) and noisy (b) PointGoal navigation",
"and most importantly, perfect localization .However,",
"In the noisy setting, the agent observes: 1) sensor noises in",
"under realistic noise models for visual sensors and ac-",
"egocentricobservation;2)actuationperturbations.Thesec",
"tuation, and without access to a \"GPS and Compass",
"ond column shows a histogram of orientation angle changes",
"sensor,\" the 99.6 % -success agents for PointGoal nav-",
"caused by a turn left action; 3) no localization informa",
"igation only succeed with 0.3%.1 In this work.",
"tion. The agent's inaccurate localization results in uncer",
"we demonstrate the surprising effectiveness of visual",
"tainty about the goal location.",
"odometry for the task of PointGoal navigation in this",
"用GPS+Compass传感器的完美定位的假设下，最近",
"realistic setting, i.e., with realistic noise models for",
"的方法以近乎完美的精度解决了这一任务（99。6%成",
"perception and actuation and without access to GPS",
"功】[53]",
"and Compasssensors.Weshowthatintegratingvisual",
"然而，这些假设是不现实的。请注意，GPS传感器",
"odometry techniques into navigation policies improves",
"通常不会在室内环境中产生精确的位置。此外，真实机",
"the state-of-the-art on the popular Habitat PointNav",
"器人的感知和驱动通常在很大程度上取决于环境照明",
"benchmark by a large margin, improving success from",
"和表面的摩擦系数。为了研究这种更现实的设置，在",
"64.5 % to 71.7 % while executing 6.4 times faster.",
"最近的基准测试2中，更新了PointGoal导航，以包",
"括来自真实机器人的噜杂驱动模型[35]。例如，对于",
"单个turnleft动作，实际转弯角度会有很大变化，如",
"1.Introduction",
"Fig.1的第二列所示。此外，还结合了E[9]的RGB和",
"在室内环境中高效，准确地导航的能力是个人机器",
"深度噪声模型来模拟现实世界的相机。最重要的是，如",
"人的基础，多年来一直是计算机视觉研究的重点E37]",
"Fig.1的第三列所示，代理无法访问GPS+Compass",
"为了围绕一个共同的框架和标准度量来凝聚社区，",
"数据，必须仅基于自我中心的RGB+深度（RGB-D)",
"",
"Anderson etal.「2提出了PointGoal导航的任务。在",
"测量进行导航。在这样一个更现实的设置下，一个在",
"PointGoal导航中，代理在以前看不见的环境中随机生",
"无噪声场景中近乎完美的策略的性能「53]急剧下降到",
"成，并且必须导航到相对于代理的初始位置和方向指定",
"0。3%。对其进行改进，先前的最新技术E241将粒子",
"SLAM集成到视觉导航中，并实现了64的成功率。5",
"的点目标e.g.，",
"，相对于start向北5米，向西3米。",
"代理使用离散动作空间（e.g：，???。25m、turn_left",
"这样逼真的设置下的%。相比99。6%在无噪声版本的",
"或turn_right30°和stop）在环境中导航。在无噪声",
"任务上的成功率，具有噪声感知和驱动以及没有本地",
"自我中心视觉（无噪声RGB+深度传感器）、无噪声驱",
"化信息的导航仍然具有挑战性为了更好地了解在这种",
"动（e.g：，turn_left将始终精确地转动30°）和使",
"现实环境中导航的挑战，我们研究了三种视觉里程计",
"（VO）技术。我们发现这些VO技术在这个现实的环境",
"1https://eval.ai/web/challenges/challenge-page/580/",
"leaderboard/1631 (Habitat Team).",
"2https://aihabitat.org/challenge/2020/"
] |
{
"arxiv_id": "2108.11550",
"language": "zh",
"page_idx": 1
}
|
||
2108.11550-zh-2
|
2108.11550-zh-2.jpg
|
中对 PointGoal 导航非常有效。具体来说，我们 1) 利用视觉里程计的几何不变性；2) 结合离散化和集成以防止噪声；3) 使用深度信息的自上而下正投影作为附加信号。对于 1), 我们注意到给定对观测值的估计运动与为置换观测值估计的运动有关。两个损失项鼓励这种关系。对于 2), 我们在视觉里程计模型的最后两层中研究了 Dropout$\underline{\mathrm{E}}[46]$ ,以防止$\underline{\mathrm{E}}[25]$ 之后的自运动预测中的不确定性。我们还发现深度离散化是有效的。对于 3), 我们从每个单独步骤的深度信息推断出以自我为中心的自上而下的投影。我们发现，这样一个简单的投影，对每个步骤都是局部的，有利于自运动估计
在人居挑战 2020PointNav 基准测试中，我们证明了这三种技术的效果令人惊讶，达到了 71。7% 成功率和-52。5 个%SPL,在 64 个。5% 和 37。7 来自现有技术的%SPL(SOTA)。此外，在导航策略中使用 VO 也会执行 6。比之前的 SOTA 快 4 倍。我们进行了详尽的消融，以显示三种技术中每种技术的功效，并发现所有上述技术都有助于更准确的导航
重要的是，我们单独训练此视觉里程计模型，而不是通过策略在线学习它。使用 VO 模型替代完美的GPS+Compass ,可以重复使用通过完美的本地化信息 (i.e. ,带有 GPS+Compass 传感器)学习的导航策略，而无需进行任何昂贵的重新训练。请注意，视觉里程计模型可以使用仅几百万帧的静态数据集针对不同的环境动态进行训练。相比之下，导航策略通常使用超过十亿个帧进行训练，这些帧是使用六个月的 GPU 时间$\mathbb{E}[53]$ 收集的
总而言之，我们研究了三种用于现实 PointGoal 导航的技术：1)通过损失利用几何不变性；2)结合离散化和集成；3) 使用深度信息的自上而下投影
我们展示：仅使用几百万帧离线学习这样的视觉里程计模型，并直接替换导航策略的 GPS+Compass 输人，即可在标准 PointNav 基准上实现 SOTA 性能
# 2. Related work
包含任务的导航。最近，人们对体现人工智能领域重新产生了兴趣。该社区在 3D 环境[27,6,47,56,55] 的照片般逼真的扫描之上构建了多个室内导航模拟器[41,57,40,27] 。为了测试机器人感知、导航和与环境交互的能力，社区还引入了几个任务[57,5,45,10,52, 36,3,28,48,22,21,51,16,34,33,31,32]和基准测试。具体来说，Batra et al. [5] 引人了对象导航任务的评估详细信息，要求代理导航到给定的对象类，而不是最终的目标点。同样，RoomNavigation [36] 要求代理导航到给定的房间类型。最近，Krantz et al. $\operatorname{E}[45,28,48]$ 扩展了导航任务，以利用自然语言指令。VLN [2,28] 和 ALFRED [45]要求代理遵循一系列自然语言指令以达到指定的目标。Thomason et al. [48]介绍了需要来回通信才能到达所需位置的视觉和对话导航。Jain et al. [22,21]开发 FurnLift 和 FurnMove 研究视觉多智能体导航。虽然这些任务的设置不同，但每个任务都需要代理在环境中准确导航。为此，代理的导航策略假
定完全了解代理的位置和方向(例如，通过使用完美的GPS+Compass 传感器)。最近，为了减轻这种不切实际的假设，Datta$\boxed{\mathrm{F}}$et al. $\boxed{\mathrm{F}}[1]$建议从一对深度图中估计自运动。像他们一样，我们也从视觉观察中进行自我运动估计。然而，不同的是，我们研究提高鲁棒性的组件。如我们在 Sec. 4.3 中所示，在不提高对观察和驱动噪声的鲁棒性的情况下，该模型产生的结果较差
相机姿态估计和视觉里程计(VO)。相机姿态估计与定位估计有关。E.g. ,在上述自运动估计$\mathbb{E}[11]$ 之后，研究了直接使用卷积神经网络(CNN)估计相对相机姿态的方法$\mathbb{E}[59,30]$ 。这些模型通常不考虑鲁棒性。同时，在过去的几十年里，已经开发了许多方法来处理VO [42,14] 。管道通常由几个步骤组成，从摄像机校准、特征选择和匹配到来自对应关系的运动估计、离群点检测和束调整。最近，针对 VO 提出了各种基于深度学习的架构。例如，Wang et al. $\mathbb{E}[49]$提出了一种 CNN 递归神经网络 (RNN),用于根据 RGB 输入估算室外环境中的 VO。由于室内导航中的三个连续帧几乎没有重叠，因此我们发现使用 RNN 进行顺序训练无济于事。相反，我们使用更快的 ResNet-18$\underline{\mathrm{E}}[17]$架构从嘈杂的 RGB-D 输入对中学习 VO。王 et al.[50] 利用刚体运动的数学群性质学习室外导航的 VO 模型。同样，我们也利用几何不变性约束作为训练过程中的自监督信号。此外，我们故意利用使模型对观察噪声具有鲁棒性的表示形式
为了对代理关于其自身运动预测的不确定性进行建模，Kendall et al. $\mathbb{E}[25]$在每个卷积层和倒数第二个线性层之后使用 Dropout$\underline{\mathrm{E}}[46]$ 。在测试时，他们的模型使用 40 个随机样本来获得自运动的稳健估计。当用作导航策略的输入时，模型在每个时间步长的 40 次向前传递非常昂贵。此外，由于 VO 模型的输人已经很嘈杂，因此将 Dropout 添加到 CNN 架构中几乎没有好处。相反，我们将 Dropout 添加到模型的最后两层， 并通过缩放最后两层的参数来近似平均多个模型的预测的效果。这允许使用单个前向传递进行稳健的估计
# 3. Approach
我们研究了一种简单但有效的视觉里程计(VO)模型，该模型适用于从一对嘈杂的 RGB-D 帧中预测自运动的具体化 AI 任务。此 VO 模型仅基于经典组件，可在下游导航任务中用作完美 GPS+Compass 传感器的直接替代品。在下面，在讨论组件之前提供概述
# 3.1.0verview
该模型在 Fig. $\color{red}{2}$中进行了说明。PointGoal 导航 [2] 要求代理导航到点目标$v_t^g$,该点目标是在每个时间步$t$ 相对于代理的当前位置指定的。在第一次移动之后， 由于噪声，代理仅具有相对位置的估计$\widehat{v}_t^g$
基于估计的相对坐标$\widehat{v}_t^g$以及以自我为中心的观测值$I_{\leq t}$,直到时间$t$ , e.g. , 来自 RGB-D 传感器的测量值，代理选择朝向目标的下一个动作。为此，代理在操作空间$\mathcal{A}=\{$turn\_left,turn\_right$,\ldots\}$、i.e.、策略$\pi(\cdot|\widehat{\boldsymbol{v}}_t^g,I_{\leq t})$上计算分布。执行操作$a_t\in\mathcal{A}$后，代理
|
[
[
0.08087158203125,
0.091552734375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.106689453125,
0.08087158203125,
0.1060791015625
],
[
0.50341796875,
0.09344482421875,
0.8896484375,
0.09344482421875,
0.8896484375,
0.106689453125,
0.50341796875,
0.106689453125
],
[
0.08416748046875,
0.10797119140625,
0.46728515625,
0.10797119140625,
0.46728515625,
0.121826171875,
0.08416748046875,
0.121826171875
],
[
0.50244140625,
0.106689453125,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.50244140625,
0.12060546875
],
[
0.08251953125,
0.125,
0.465576171875,
0.125,
0.465576171875,
0.1351318359375,
0.08251953125,
0.1351318359375
],
[
0.5048828125,
0.12249755859375,
0.890625,
0.12249755859375,
0.890625,
0.1363525390625,
0.5048828125,
0.1363525390625
],
[
0.08087158203125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.5048828125,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.08416748046875,
0.1546630859375,
0.465576171875,
0.1546630859375,
0.465576171875,
0.1654052734375,
0.08416748046875,
0.1654052734375
],
[
0.50341796875,
0.154052734375,
0.8896484375,
0.154052734375,
0.8896484375,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.08087158203125,
0.1673583984375,
0.465576171875,
0.1673583984375,
0.465576171875,
0.18115234375,
0.08087158203125,
0.18115234375
],
[
0.5048828125,
0.1685791015625,
0.8896484375,
0.1685791015625,
0.8896484375,
0.1817626953125,
0.5048828125,
0.1817626953125
],
[
0.08416748046875,
0.1849365234375,
0.46728515625,
0.1849365234375,
0.46728515625,
0.1956787109375,
0.08416748046875,
0.1956787109375
],
[
0.50341796875,
0.18310546875,
0.84619140625,
0.18310546875,
0.84619140625,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.08251953125,
0.199462890625,
0.465576171875,
0.199462890625,
0.465576171875,
0.210205078125,
0.08251953125,
0.210205078125
],
[
0.50634765625,
0.2008056640625,
0.88818359375,
0.2008056640625,
0.88818359375,
0.2108154296875,
0.50634765625,
0.2108154296875
],
[
0.08087158203125,
0.21337890625,
0.465576171875,
0.21337890625,
0.465576171875,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.5048828125,
0.2139892578125,
0.890625,
0.2139892578125,
0.890625,
0.2279052734375,
0.5048828125,
0.2279052734375
],
[
0.08251953125,
0.2291259765625,
0.465576171875,
0.2291259765625,
0.465576171875,
0.242431640625,
0.08251953125,
0.242431640625
],
[
0.50341796875,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.50341796875,
0.242431640625
],
[
0.08251953125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.08251953125,
0.2568359375
],
[
0.50341796875,
0.2442626953125,
0.88818359375,
0.2442626953125,
0.88818359375,
0.25830078125,
0.50341796875,
0.25830078125
],
[
0.08251953125,
0.261474609375,
0.09722900390625,
0.261474609375,
0.09722900390625,
0.270751953125,
0.08251953125,
0.270751953125
],
[
0.5048828125,
0.260009765625,
0.890625,
0.260009765625,
0.890625,
0.2734375,
0.5048828125,
0.2734375
],
[
0.101318359375,
0.27587890625,
0.465576171875,
0.27587890625,
0.465576171875,
0.285888671875,
0.101318359375,
0.285888671875
],
[
0.50341796875,
0.274658203125,
0.8896484375,
0.274658203125,
0.8896484375,
0.28857421875,
0.50341796875,
0.28857421875
],
[
0.08087158203125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.50341796875,
0.290283203125,
0.88818359375,
0.291748046875,
0.88818359375,
0.302978515625,
0.50341796875,
0.3017578125
],
[
0.09722900390625,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.09722900390625,
0.317626953125
],
[
0.50341796875,
0.304931640625,
0.8896484375,
0.304931640625,
0.8896484375,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.08251953125,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.08251953125,
0.333251953125
],
[
0.50341796875,
0.318115234375,
0.8759765625,
0.318115234375,
0.8759765625,
0.33203125,
0.50341796875,
0.33203125
],
[
0.08251953125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.08251953125,
0.348388671875
],
[
0.50634765625,
0.3359375,
0.88818359375,
0.3359375,
0.88818359375,
0.3466796875,
0.50634765625,
0.3466796875
],
[
0.08251953125,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.363037109375,
0.08251953125,
0.363037109375
],
[
0.50341796875,
0.349853515625,
0.890625,
0.349853515625,
0.890625,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.08087158203125,
0.3642578125,
0.325927734375,
0.3642578125,
0.325927734375,
0.37744140625,
0.08087158203125,
0.37744140625
],
[
0.50634765625,
0.3662109375,
0.8896484375,
0.3662109375,
0.8896484375,
0.376953125,
0.50634765625,
0.376953125
],
[
0.10296630859375,
0.380615234375,
0.46728515625,
0.380615234375,
0.46728515625,
0.391357421875,
0.10296630859375,
0.391357421875
],
[
0.5048828125,
0.380126953125,
0.890625,
0.380126953125,
0.890625,
0.393310546875,
0.5048828125,
0.393310546875
],
[
0.08251953125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.08251953125,
0.40771484375
],
[
0.5048828125,
0.395751953125,
0.88818359375,
0.395751953125,
0.88818359375,
0.4091796875,
0.5048828125,
0.4091796875
],
[
0.08251953125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.08251953125,
0.423583984375
],
[
0.5048828125,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.423583984375,
0.5048828125,
0.423583984375
],
[
0.08087158203125,
0.4248046875,
0.468017578125,
0.4248046875,
0.468017578125,
0.438232421875,
0.08087158203125,
0.438232421875
],
[
0.5048828125,
0.425537109375,
0.890625,
0.425537109375,
0.890625,
0.439453125,
0.5048828125,
0.439453125
],
[
0.08087158203125,
0.43994140625,
0.46728515625,
0.43994140625,
0.46728515625,
0.453857421875,
0.08087158203125,
0.453857421875
],
[
0.50732421875,
0.442626953125,
0.8896484375,
0.442626953125,
0.8896484375,
0.45263671875,
0.50732421875,
0.45263671875
],
[
0.08087158203125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.08087158203125,
0.468505859375
],
[
0.50341796875,
0.455810546875,
0.65283203125,
0.455810546875,
0.65283203125,
0.4697265625,
0.50341796875,
0.4697265625
],
[
0.08087158203125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.484130859375,
0.08087158203125,
0.484130859375
],
[
0.52294921875,
0.47021484375,
0.890625,
0.4697265625,
0.890625,
0.484130859375,
0.52294921875,
0.48486328125
],
[
0.08251953125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.08251953125,
0.498779296875
],
[
0.50341796875,
0.486083984375,
0.890625,
0.486083984375,
0.890625,
0.5,
0.50341796875,
0.5
],
[
0.08087158203125,
0.50048828125,
0.193603515625,
0.50048828125,
0.193603515625,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.50341796875,
0.5,
0.8896484375,
0.50048828125,
0.8896484375,
0.51513671875,
0.50341796875,
0.5146484375
],
[
0.10296630859375,
0.517578125,
0.46728515625,
0.517578125,
0.46728515625,
0.52783203125,
0.10296630859375,
0.52783203125
],
[
0.50634765625,
0.51806640625,
0.88623046875,
0.51806640625,
0.88623046875,
0.52880859375,
0.50634765625,
0.52880859375
],
[
0.08251953125,
0.53076171875,
0.468017578125,
0.53076171875,
0.468017578125,
0.54443359375,
0.08251953125,
0.54443359375
],
[
0.50341796875,
0.53076171875,
0.88818359375,
0.5322265625,
0.88818359375,
0.54541015625,
0.50341796875,
0.54443359375
],
[
0.08087158203125,
0.54541015625,
0.39794921875,
0.54541015625,
0.39794921875,
0.5595703125,
0.08087158203125,
0.5595703125
],
[
0.50341796875,
0.546875,
0.890625,
0.54541015625,
0.890625,
0.56005859375,
0.50341796875,
0.56103515625
],
[
0.101318359375,
0.56103515625,
0.46728515625,
0.56103515625,
0.46728515625,
0.57470703125,
0.101318359375,
0.57470703125
],
[
0.50341796875,
0.5625,
0.88818359375,
0.5625,
0.88818359375,
0.57568359375,
0.50341796875,
0.57568359375
],
[
0.08087158203125,
0.57568359375,
0.46728515625,
0.57568359375,
0.46728515625,
0.5888671875,
0.08087158203125,
0.5888671875
],
[
0.50341796875,
0.5771484375,
0.88818359375,
0.5771484375,
0.88818359375,
0.59033203125,
0.50341796875,
0.59033203125
],
[
0.08087158203125,
0.59130859375,
0.44287109375,
0.59130859375,
0.44287109375,
0.60498046875,
0.08087158203125,
0.60498046875
],
[
0.50634765625,
0.59326171875,
0.88818359375,
0.59326171875,
0.88818359375,
0.60400390625,
0.50634765625,
0.60400390625
],
[
0.5048828125,
0.60693359375,
0.87744140625,
0.60693359375,
0.87744140625,
0.62060546875,
0.5048828125,
0.62060546875
],
[
0.080078125,
0.619140625,
0.221435546875,
0.619140625,
0.221435546875,
0.63330078125,
0.080078125,
0.63330078125
],
[
0.50341796875,
0.63720703125,
0.615234375,
0.63720703125,
0.615234375,
0.6513671875,
0.50341796875,
0.6513671875
],
[
0.08251953125,
0.64599609375,
0.46728515625,
0.64599609375,
0.46728515625,
0.6591796875,
0.08251953125,
0.6591796875
],
[
0.08087158203125,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.08087158203125,
0.673828125
],
[
0.52392578125,
0.66357421875,
0.890625,
0.66357421875,
0.890625,
0.6767578125,
0.52392578125,
0.6767578125
],
[
0.08087158203125,
0.6748046875,
0.465576171875,
0.6748046875,
0.465576171875,
0.68798828125,
0.08087158203125,
0.68798828125
],
[
0.50341796875,
0.67822265625,
0.8896484375,
0.67822265625,
0.8896484375,
0.69140625,
0.50341796875,
0.69140625
],
[
0.08416748046875,
0.6923828125,
0.46728515625,
0.6923828125,
0.46728515625,
0.70263671875,
0.08416748046875,
0.70263671875
],
[
0.5048828125,
0.693359375,
0.8896484375,
0.693359375,
0.8896484375,
0.70703125,
0.5048828125,
0.70703125
],
[
0.08087158203125,
0.70556640625,
0.46728515625,
0.705078125,
0.46728515625,
0.71826171875,
0.08087158203125,
0.7197265625
],
[
0.50341796875,
0.70849609375,
0.8896484375,
0.70849609375,
0.8896484375,
0.7216796875,
0.50341796875,
0.7216796875
],
[
0.08087158203125,
0.72021484375,
0.46728515625,
0.72021484375,
0.46728515625,
0.734375,
0.08087158203125,
0.734375
],
[
0.50341796875,
0.7236328125,
0.84716796875,
0.7236328125,
0.84716796875,
0.7373046875,
0.50341796875,
0.7373046875
],
[
0.08087158203125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.08087158203125,
0.75
],
[
0.50244140625,
0.744140625,
0.615234375,
0.74560546875,
0.615234375,
0.76025390625,
0.50244140625,
0.7587890625
],
[
0.08251953125,
0.751953125,
0.46728515625,
0.751953125,
0.46728515625,
0.76513671875,
0.08251953125,
0.76513671875
],
[
0.08251953125,
0.7666015625,
0.468017578125,
0.7666015625,
0.468017578125,
0.77978515625,
0.08251953125,
0.77978515625
],
[
0.52392578125,
0.76513671875,
0.8896484375,
0.76513671875,
0.8896484375,
0.77978515625,
0.52392578125,
0.77978515625
],
[
0.08251953125,
0.78271484375,
0.464111328125,
0.78271484375,
0.464111328125,
0.79345703125,
0.08251953125,
0.79345703125
],
[
0.50341796875,
0.78076171875,
0.88818359375,
0.78076171875,
0.88818359375,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.080078125,
0.796875,
0.465576171875,
0.79443359375,
0.465576171875,
0.8095703125,
0.080078125,
0.81103515625
],
[
0.50244140625,
0.79443359375,
0.89208984375,
0.79541015625,
0.89208984375,
0.81005859375,
0.50244140625,
0.8095703125
],
[
0.08087158203125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.50341796875,
0.81103515625,
0.80224609375,
0.81103515625,
0.80224609375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08251953125,
0.8271484375,
0.465576171875,
0.8271484375,
0.465576171875,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.52392578125,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84033203125,
0.52392578125,
0.84033203125
],
[
0.08251953125,
0.84130859375,
0.46728515625,
0.84033203125,
0.46728515625,
0.85498046875,
0.08251953125,
0.85595703125
],
[
0.5048828125,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85595703125,
0.5048828125,
0.85595703125
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.08251953125,
0.87255859375,
0.468017578125,
0.87255859375,
0.468017578125,
0.88623046875,
0.08251953125,
0.88623046875
],
[
0.50341796875,
0.87060546875,
0.87841796875,
0.8720703125,
0.87841796875,
0.88623046875,
0.50341796875,
0.88525390625
],
[
0.87353515625,
0.875,
0.88818359375,
0.875,
0.88818359375,
0.88330078125,
0.87353515625,
0.88330078125
],
[
0.08087158203125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"中对PointGoal导航非常有效。具体来说，我们1）利",
"定完全了解代理的位置和方向（例如，通过使用完美的",
"用视觉里程计的几何不变性；2）结合离散化和集成以",
"GPS+Compass传感器）。最近，为了减轻这种不切实",
"防止噪声：3）使用深度信息的自上而下正投影作为附",
"际的假设，DattaEletal.E[11]建议从一对深度图中估",
"加信号。对于1），我们注意到给定对观测值的估计运",
"计自运动。像他们一样，我们也从视觉观察中进行自我",
"动与为置换观测值估计的运动有关。两个损失项鼓励",
"运动估计。然而，不同的是，我们研究提高鲁棒性的组",
"这种关系。对于2），我们在视觉里程计模型的最后两",
"件。如我们在Sec.4.3中所示，在不提高对观察和驱动",
"层中研究了DropoutE46，以防止E25之后的自运",
"噪声的鲁棒性的情况下，该模型产生的结果较差",
"动预测中的不确定性。我们还发现深度离散化是有效",
"相机姿态估计和视觉单程计（VO）。相机姿态估计与定",
"的。对于3），我们从每个单独步骤的深度信息推断出",
"位估计有关。E.g.，在上述自运动估计E[11]之后，研",
"以自我为中心的自上而下的投影。我们发现，这样一个",
"究了直接使用卷积神经网络（CNN）估计相对相机姿",
"简单的投影，对每个步骤都是局部的，有利于自运动估",
"态的方法E[59，301。这些模型通常不考虑鲁棒性。同",
"计",
"时，在过去的几十年里，已经开发了许多方法来处理",
"在人居挑战2020PointNav基准测试中，我们证明了",
"VO【42，14。管道通常由几个步骤组成，从摄像机校",
"这三种技术的效果令人惊，达到了71。7%成功率和",
"准、特征选择和匹配到来自对应关系的运动估计、离",
"52。5个%SPL，在64个。5%和37。7来自现有技",
"群点检测和束调整。最近，针对VO提出了各种基于",
"术的%SPL(SOTA)。此外，在导航策略中使用VO也",
"深度学习的架构。例如，Wangetal.E[49]提出了",
"会执行6。比之前的SOTA快4倍。我们进行了详尽",
"种CNN递归神经网络（RNN），用于根据RGB输入",
"的消融，以显示三种技术中每种技术的功效，并发现所",
"估算室外环境中的VO。由于室内导航中的三个连续顿",
"有上述技术都有助于更准确的导航",
"几乎没有重叠，因此我们发现使用RNN进行顺序训练",
"重要的是，我们单独训练此视觉里程计模型，而不",
"无济于事。相反，我们使用更快的ResNet-18E[17]架",
"是通过策略在线学习它。使用VO模型替代完美的",
"构从噜杂的RGB-D输入对中学习VO。王etal.50",
"GPS+Compass，可以重复使用通过完美的本地化信",
"利用刚体运动的数学群性质学习室外导航的VO模型。",
"息（i.e.，带有GPS+Compass传感器）学习的导航策",
"同样，我们也利用几何不变性约束作为训练过程中的",
"略，而无需进行任何昂贵的重新训练。请注意，视觉里",
"自监督信号。此外，我们故意利用使模型对观察噪声具",
"程计模型可以使用仅几百万顿的静态数据集针对不同",
"有鲁棒性的表示形式",
"的环境动态进行训练。相比之下，导航策略通常使用超",
"为了对代理关于其自身运动预测的不确定性进行建",
"过十亿个顿进行训练，这些顿是使用六个月的GPU时",
"模，Kendall etal.E[25]在每个卷积层和倒数第二个",
"间E[53]收集的",
"线性层之后使用DropoutE[46]。在测试时，他们的模",
"总而言之，我们研究了三种用于现实PointGoal导",
"型使用40个随机样本来获得自运动的稳健估计。当用",
"航的技术：1）通过损失利用几何不变性；2）结合离散",
"作导航策略的输入时，模型在每个时间步长的40次向",
"化和集成；3）使用深度信息的自上而下投影",
"前传递非常昂贵。此外，由于VO模型的输入已经很",
"我们展示：仅使用几百万顿离线学习这样的视觉里",
"噜杂，因此将Dropout添加到CNN架构中几乎没有",
"程计模型，并直接替换导航策略的GPS+Compass输",
"好处。相反，我们将Dropout添加到模型的最后两层",
"人，即可在标准PointNav基准上实现SOTA性能",
"并通过缩放最后两层的参数来近似平均多个模型的预",
"测的效果。这允许使用单个前向传递进行稳健的估计",
"2. Related work",
"3.Approach",
"包含任务的导航。最近，人们对体现人工智能领域重",
"新产生了兴趣。该社区在3D环境[27，6，47，56，55",
"我们研究了一种简单但有效的视觉里程计（VO）模",
"的照片般逼真的扫描之上构建了多个室内导航模拟器",
"型，该模型适用于从一对噜杂的RGB-D顿中预测自运",
"41.57，4027。为了测试机器人感知、导航和与环境",
"动的具体化AI任务。此VO模型仅基于经典组件，可",
"交互的能力，社区还引人了几个任务「57，5，45，10，52，",
"在下游导航任务中用作完美GPS+Compass传感器的",
"36,3,28,48，22,21,51，16,34,33,31,32和基准测试。",
"直接替代品。在下面，在讨论组件之前提供概述",
"具体来说，Batraetal.[5]引入了对象导航任务的评估",
"3.1. Overview",
"详细信息，要求代理导航到给定的对象类，而不是最终",
"的目标点。同样，RoomNavigation[36]要求代理导航",
"该模型在Fig.2中进行了说明。PointGoal导航[2]",
"到给定的房间类型。最近，Krantzetal.F45，28，48",
"要求代理导航到点目标，该点目标是在每个时间步",
"扩展了导航任务，以利用自然语言指令。VLN[2，28",
"t相对于代理的当前位置指定的。在第一次移动之后，",
"和ALFRED「45]要求代理遵循一系列自然语言指令",
"由于噪声，代理仅具有相对位置的估计",
"以达到指定的目标。Thomason etal.[48]介绍了需要",
"基于估计的相对坐标以及以自我为中心的观测",
"来回通信才能到达所需位置的视觉和对话导航。Jain",
"值I<t，直到时间t，e.g.，来自RGB-D传感器的测",
"etal.[22，21]开发FurnLift和FurnMove研究视觉多",
"量值，代理选择朝向目标的下一个动作。为此，代理在",
"智能体导航。虽然这些任务的设置不同，但每个任务都",
"操作空间A={turn_left,turn_right，..}、i.e.",
"策",
"需要代理在环境中准确导航。为此，代理的导航策略假",
"略π(,I<t）上计算分布。执行操作atEA后，代理"
] |
{
"arxiv_id": "2108.11550",
"language": "zh",
"page_idx": 2
}
|
||
2108.11550-zh-4
|
2108.11550-zh-4.jpg
|
<FigureHere>
Figure 3: Steps to infer an egocentric top-down projection
from depth. Top and bottom rows show inferred top-down projections from noisy and noiseless depth image at the same location. (b,e): top-down scatter plot. (c,f): the soft top-down projection. As can be seen, after processing, (c) and (f) share more similarities than (b) and (e), making the representation more robust to depth noises.
# 3.3. Robustness to Uncertainty
除了利用几何不变性之外，我们发现进一步增加模型的$SE(2)$估计的鲁棒性很重要。这很重要，因为测量会产生噪声：1)即使由于观察噪声而使相机的位置和方向相同，视觉观察也会有所不同。这使得观察结果的处理变得脆弱；2)驱动中的扰动会影响 VO 模型的预测，因为它们会增加旋转和平移的方差。对于鲁棒性，我们使用两种经典技术：
合奏。为了提高鲁棒性，可以训练一组模型。对集合上的预测进行平均通常会减少方差。然而，基于强化学习(RL)的导航系统需要数十亿个样本来训练一个好的策略 [53] 。由于策略依赖于 VO 模型来提供代理的当前位置估计，因此提高推理速度并避免不必要的计算非常重要。因此，我们发现训练一个 CNN 架构，同时将Dropout [46]添加到最后两个全连接 (FC)层，而不是集成多个模型，这很有帮助。这在经济上类似于训练大量合奏的行为[4,18] 。在训练期间，Dropout 以$p$的概率随机禁用 FC 层中的隐藏单元，本质上是从子网集合中采样。在推理过程中，FC 层中的每个隐藏单元都使用相同的因子$p$进行缩放，以模拟来自多个子网络的预测的平均值
深度离散化。此外，我们发现深度离散化可以更可靠地表示距离传感器的自我中心观察。具体来说，使用独热编码将单通道深度图 depth 离散化为具有$N$通道的表示 d-depth 。给定图像坐标 $(x,y)$ 处的 depth 的像素， 我们通过
$$\mathrm{d-depth}_i(x,y)=\mathbb{1}\left\{\mathrm{depth}(x,y)\in[z_{i-1},z_i)\right\},$$
(7)
获得 d-depth 的$i$通道的值，其中 1$\{\cdot\}$表示指标函数， $\{z_{i-1},z_i\}$是离散化间隔的端点。直观上，这增加了深度不确定性对$\min_i\frac{|z_i-z_{i-1}|}2$的绝对容忍度，因为除非深度条目越过间隔边界，否则将生成相同的表示。根据经验，我们发现使用端点$z_i=i\cdot(z_{\max}-z_{\min})/N$可以很好地将等距离散化为$N$间隔。这里，$z_\mathrm{max}$和$z_\mathrm{min}$ 分别是最大(10 $m$)和最小深度(0 $m$)值
3.4. Top-Down Projection as Additional Signal
直观地图应进一步提高模型的鲁棒性。但是，我们设置中的关键挑战：深度传感器中的噪声相当微妙，通常几乎看不见(请参阅 Fig. 3 (a, d))。但是，一旦投影到 2D 布局，噪声就会表现为 Fig. 3 (b, e) 中明显的总偏差，孔和阻塞。为了应对这一挑战，我们使用归一化软投影。Fig.3(c,f) 中显示的归一化软投影 s-proj$_t$类似于深度图给出的房间布局。请注意，它们也比 Fig. 3 ( b, e) 中给出的投影具有更多的相似性
我们通过以下方式获得软投影：1)将深度观测值映射到 3D 点云，2) 使用 2D 自上而下的正交投影，以及 3) 相对于每个像素内的点数对投影进行归一化。提供软投影作为端到端训练的 VO 模型的输入，该模型学习如何正确使用它。附录中提供了如何计算软投影的详细信息
# 3.5. VO Model Architecture, Training Details, and Integration with Navigation Policy
模型体系结构。Eq.(??)中的视觉里程计模型$f_\phi$采用ResNet-18 [17] 主干提取视觉特征。为此，我们首先通过
$$\psi(I_t)\triangleq(\mathrm{RGB}_t,\mathrm{depth}_t,\mathrm{d-depth}_t,\mathrm{s-proj}_t).$$
(8)
从 Fig. $\color{red}2($c)中绘制的自我中心观察计算表示，然后，我们沿着通道维度堆叠$(\psi(I_t),\psi(I_{t+1}))$以获得 ResNet- $18\text{ 输入。由于 RGB}_{t}\mathrm{~,~depth}_{t}\mathrm{~,~d-depth}_{t}$和 s-proj$_t$分别具有三个，一个，$N$和一个通道，因此 ResNet-18 的输入是具有$(2N+10)$通道的张量。为了估计 $\widehat{H}_{C_t\to\mathcal{C}_{t+1}}$
我们在 ResNet-18 特征提取器的顶部使用两个带有Dropout 的完全连接 (FC) 层。这些 FC 层在 512 维特征上运行，并产生输出$\left(\hat{\xi}_{\mathcal{C}_t\to\mathcal{C}_{t+1}}^x,\hat{\xi}_{\mathcal{C}_t\to\mathcal{C}_{t+1}}^z,\hat{\theta}_{\mathcal{C}_t\to\mathcal{C}_{t+1}}\right)$ 。这里$\hat{\xi}_{\mathcal{C}_t\to\mathcal{C}_{t+1}}^z$指的是在代理的向前方向上的平移，而$\hat{\xi}_{\mathcal{C}_t\to\mathcal{C}_{t+1}}^x$指的是在垂直于地平面上的向前运动的方向上的平移
VO 训练。我们在数据集$\mathcal{D} _\mathrm{train}$ = $\left \{ \left ( ( I_{t}, I_{t+ 1}) , \xi _{\mathcal{C} _{t}\rightarrow \mathcal{C} _{t+ 1}}, \theta _{\mathcal{C} _{t}\rightarrow \mathcal{C} _{t+ 1}}\right ) \right \}$ $\overset {\triangle }{\operatorname* { = } }$ $\left \{ d_{\mathcal{C} _{t}\rightarrow \mathcal{C} _{t+ 1}}\right \}$上训练视觉里程计模型$f_\phi$ 。每个数据点由一对以自我为中心的观测值以及地面实况平移和旋转角度组成。对该模型进行了优化，以共同最小化 Eq. (3)和 Eq. (4) , i.e. 中定义的回归损失和几何不变性损失，我们解决了$\operatorname*{min}_\phi\mathcal{L}_\mathrm{VO}\triangleq$
$$\sum_{d_{\mathcal{C}_{t}\to\mathcal{C}_{t+1}}\in\mathcal{D}_{\mathrm{train}}}\left[\lambda_{\mathrm{reg}}\mathcal{L}_{\mathcal{C}_{t}\to\mathcal{C}_{t+1}}^{\mathrm{reg}}+\lambda_{\mathrm{inv}}^{\mathrm{trans}}\mathcal{L}_{\mathcal{C}_{t}\to\mathcal{C}_{t+1}}^{\mathrm{inv,~trans}}+\lambda_{\mathrm{inv}}^{\mathrm{rot}}\mathcal{L}_{\mathcal{C}_{t}\to\mathcal{C}_{t+1}}^{\mathrm{inv,~rot}}\right],$$
其中$\lambda _\mathrm{reg}$ , $\lambda _\mathrm{inv}^\mathrm{trans}$和$\lambda_\mathrm{inv}^\mathrm{rot}$是用户指定的超参数。我们在实验中将它们设置为 1.0 。我们使用$2.5\times10^{-4}$的学习速率，用 Adam [26]优化 VO 模型。训练期间的退出因子为$p=0.2$
导航策略培训。我们工作的重点是现实条件下的 PointGoal 导航，Eli. e. , 嘈杂的观测和驱动以及无法访问GPU+Compass 传感器。为了证明 VO 技术可以简单地替代地面实况 GPS+Compass 传感器，我们直接使
|
[
[
0.08660888671875,
0.0953369140625,
0.096435546875,
0.0953369140625,
0.096435546875,
0.10101318359375,
0.08660888671875,
0.10101318359375
],
[
0.50341796875,
0.0921630859375,
0.865234375,
0.0921630859375,
0.865234375,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.4599609375,
0.10101318359375,
0.47216796875,
0.10986328125,
0.461669921875,
0.11871337890625,
0.449462890625,
0.10919189453125
],
[
0.453369140625,
0.11297607421875,
0.4697265625,
0.11297607421875,
0.4697265625,
0.1307373046875,
0.453369140625,
0.1307373046875
],
[
0.52392578125,
0.11236572265625,
0.8896484375,
0.11236572265625,
0.8896484375,
0.1256103515625,
0.52392578125,
0.1256103515625
],
[
0.5048828125,
0.126953125,
0.890625,
0.126953125,
0.890625,
0.14013671875,
0.5048828125,
0.14013671875
],
[
0.5048828125,
0.1427001953125,
0.8896484375,
0.1427001953125,
0.8896484375,
0.1558837890625,
0.5048828125,
0.1558837890625
],
[
0.50341796875,
0.1572265625,
0.8896484375,
0.1558837890625,
0.8896484375,
0.17041015625,
0.50341796875,
0.1717529296875
],
[
0.08416748046875,
0.1666259765625,
0.1004638671875,
0.1666259765625,
0.1004638671875,
0.175537109375,
0.08416748046875,
0.175537109375
],
[
0.5048828125,
0.17236328125,
0.890625,
0.17236328125,
0.890625,
0.186279296875,
0.5048828125,
0.186279296875
],
[
0.4599609375,
0.189453125,
0.465576171875,
0.189453125,
0.465576171875,
0.1976318359375,
0.4599609375,
0.1976318359375
],
[
0.5048828125,
0.1881103515625,
0.890625,
0.1881103515625,
0.890625,
0.2020263671875,
0.5048828125,
0.2020263671875
],
[
0.4599609375,
0.2008056640625,
0.464111328125,
0.2008056640625,
0.464111328125,
0.205810546875,
0.4599609375,
0.205810546875
],
[
0.50244140625,
0.20263671875,
0.890625,
0.2020263671875,
0.890625,
0.216552734375,
0.50244140625,
0.2171630859375
],
[
0.5048828125,
0.2183837890625,
0.890625,
0.2183837890625,
0.890625,
0.231689453125,
0.5048828125,
0.231689453125
],
[
0.08087158203125,
0.2310791015625,
0.468017578125,
0.2310791015625,
0.468017578125,
0.2442626953125,
0.08087158203125,
0.2442626953125
],
[
0.52392578125,
0.2342529296875,
0.890625,
0.2342529296875,
0.890625,
0.2474365234375,
0.52392578125,
0.2474365234375
],
[
0.08087158203125,
0.246826171875,
0.465576171875,
0.246826171875,
0.465576171875,
0.2568359375,
0.08087158203125,
0.2568359375
],
[
0.5048828125,
0.25,
0.890625,
0.25,
0.890625,
0.26318359375,
0.5048828125,
0.26318359375
],
[
0.08087158203125,
0.25830078125,
0.46728515625,
0.25830078125,
0.46728515625,
0.271484375,
0.08087158203125,
0.271484375
],
[
0.50341796875,
0.26318359375,
0.8896484375,
0.26318359375,
0.8896484375,
0.276611328125,
0.50341796875,
0.276611328125
],
[
0.08087158203125,
0.271484375,
0.46728515625,
0.271484375,
0.46728515625,
0.285400390625,
0.08087158203125,
0.285400390625
],
[
0.50341796875,
0.27978515625,
0.890625,
0.27978515625,
0.890625,
0.29345703125,
0.50341796875,
0.29345703125
],
[
0.08087158203125,
0.28857421875,
0.465576171875,
0.28857421875,
0.465576171875,
0.298583984375,
0.08087158203125,
0.298583984375
],
[
0.08087158203125,
0.2998046875,
0.468017578125,
0.2998046875,
0.468017578125,
0.313232421875,
0.08087158203125,
0.313232421875
],
[
0.50341796875,
0.294189453125,
0.8896484375,
0.294189453125,
0.8896484375,
0.30810546875,
0.50341796875,
0.30810546875
],
[
0.080078125,
0.314453125,
0.366943359375,
0.314453125,
0.366943359375,
0.32763671875,
0.080078125,
0.32763671875
],
[
0.50341796875,
0.31005859375,
0.583984375,
0.31005859375,
0.583984375,
0.3212890625,
0.50341796875,
0.3212890625
],
[
0.50244140625,
0.330810546875,
0.890625,
0.33203125,
0.890625,
0.3466796875,
0.50244140625,
0.34521484375
],
[
0.07843017578125,
0.3466796875,
0.318603515625,
0.347900390625,
0.318603515625,
0.3623046875,
0.07843017578125,
0.361083984375
],
[
0.54345703125,
0.347900390625,
0.80615234375,
0.347900390625,
0.80615234375,
0.3623046875,
0.54345703125,
0.3623046875
],
[
0.5048828125,
0.36865234375,
0.8896484375,
0.36865234375,
0.8896484375,
0.3818359375,
0.5048828125,
0.3818359375
],
[
0.1004638671875,
0.375,
0.46728515625,
0.375,
0.46728515625,
0.38818359375,
0.1004638671875,
0.38818359375
],
[
0.50244140625,
0.3818359375,
0.890625,
0.38330078125,
0.890625,
0.397705078125,
0.50244140625,
0.396484375
],
[
0.08087158203125,
0.38818359375,
0.46728515625,
0.38818359375,
0.46728515625,
0.401611328125,
0.08087158203125,
0.401611328125
],
[
0.08251953125,
0.404052734375,
0.468017578125,
0.404052734375,
0.468017578125,
0.417236328125,
0.08251953125,
0.417236328125
],
[
0.50341796875,
0.3984375,
0.52294921875,
0.3984375,
0.52294921875,
0.412353515625,
0.50341796875,
0.412353515625
],
[
0.08087158203125,
0.419189453125,
0.46728515625,
0.419189453125,
0.46728515625,
0.43310546875,
0.08087158203125,
0.43310546875
],
[
0.55126953125,
0.42041015625,
0.890625,
0.42041015625,
0.890625,
0.434326171875,
0.55126953125,
0.434326171875
],
[
0.08087158203125,
0.43505859375,
0.46728515625,
0.43505859375,
0.46728515625,
0.448974609375,
0.08087158203125,
0.448974609375
],
[
0.5048828125,
0.442626953125,
0.89208984375,
0.442626953125,
0.89208984375,
0.455810546875,
0.5048828125,
0.455810546875
],
[
0.08251953125,
0.449462890625,
0.46728515625,
0.449462890625,
0.46728515625,
0.46337890625,
0.08251953125,
0.46337890625
],
[
0.08087158203125,
0.464111328125,
0.281982421875,
0.464111328125,
0.281982421875,
0.477783203125,
0.08087158203125,
0.477783203125
],
[
0.5048828125,
0.458251953125,
0.89208984375,
0.458251953125,
0.89208984375,
0.4716796875,
0.5048828125,
0.4716796875
],
[
0.50341796875,
0.472900390625,
0.8896484375,
0.472900390625,
0.8896484375,
0.486083984375,
0.50341796875,
0.486083984375
],
[
0.08416748046875,
0.4873046875,
0.46728515625,
0.4873046875,
0.46728515625,
0.50048828125,
0.08416748046875,
0.50048828125
],
[
0.50341796875,
0.488037109375,
0.89208984375,
0.486083984375,
0.89208984375,
0.5,
0.50341796875,
0.501953125
],
[
0.08087158203125,
0.501953125,
0.46728515625,
0.501953125,
0.46728515625,
0.51513671875,
0.08087158203125,
0.51513671875
],
[
0.50244140625,
0.50048828125,
0.87841796875,
0.50244140625,
0.87841796875,
0.517578125,
0.50244140625,
0.51513671875
],
[
0.08087158203125,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.53076171875,
0.08087158203125,
0.53076171875
],
[
0.5087890625,
0.5234375,
0.515625,
0.5234375,
0.515625,
0.5263671875,
0.5087890625,
0.5263671875
],
[
0.51953125,
0.517578125,
0.8896484375,
0.5166015625,
0.8896484375,
0.53076171875,
0.51953125,
0.5322265625
],
[
0.080078125,
0.5322265625,
0.46728515625,
0.5322265625,
0.46728515625,
0.54541015625,
0.080078125,
0.54541015625
],
[
0.50341796875,
0.53271484375,
0.8896484375,
0.53271484375,
0.8896484375,
0.546875,
0.50341796875,
0.546875
],
[
0.08416748046875,
0.548828125,
0.46728515625,
0.548828125,
0.46728515625,
0.5595703125,
0.08416748046875,
0.5595703125
],
[
0.50341796875,
0.55029296875,
0.8857421875,
0.55029296875,
0.8857421875,
0.568359375,
0.50341796875,
0.568359375
],
[
0.08251953125,
0.5625,
0.468017578125,
0.5625,
0.468017578125,
0.57568359375,
0.08251953125,
0.57568359375
],
[
0.50341796875,
0.5712890625,
0.890625,
0.5712890625,
0.890625,
0.5888671875,
0.50341796875,
0.5888671875
],
[
0.08087158203125,
0.57763671875,
0.468017578125,
0.57763671875,
0.468017578125,
0.59130859375,
0.08087158203125,
0.59130859375
],
[
0.08087158203125,
0.59228515625,
0.468017578125,
0.59228515625,
0.468017578125,
0.60595703125,
0.08087158203125,
0.60595703125
],
[
0.50244140625,
0.59130859375,
0.89208984375,
0.59130859375,
0.89208984375,
0.609375,
0.50244140625,
0.609375
],
[
0.08251953125,
0.60791015625,
0.468017578125,
0.60791015625,
0.468017578125,
0.6220703125,
0.08251953125,
0.6220703125
],
[
0.50634765625,
0.60986328125,
0.5703125,
0.60986328125,
0.5703125,
0.6220703125,
0.50634765625,
0.6220703125
],
[
0.08087158203125,
0.62255859375,
0.468017578125,
0.62255859375,
0.468017578125,
0.63623046875,
0.08087158203125,
0.63623046875
],
[
0.5009765625,
0.62255859375,
0.84228515625,
0.625,
0.84228515625,
0.6396484375,
0.5009765625,
0.63720703125
],
[
0.08087158203125,
0.63818359375,
0.468017578125,
0.63818359375,
0.468017578125,
0.6513671875,
0.08087158203125,
0.6513671875
],
[
0.50341796875,
0.63818359375,
0.89208984375,
0.63720703125,
0.89208984375,
0.65478515625,
0.50341796875,
0.65576171875
],
[
0.08087158203125,
0.65283203125,
0.468017578125,
0.65283203125,
0.468017578125,
0.66650390625,
0.08087158203125,
0.66650390625
],
[
0.5048828125,
0.65478515625,
0.890625,
0.65478515625,
0.890625,
0.66845703125,
0.5048828125,
0.66845703125
],
[
0.080078125,
0.66748046875,
0.2132568359375,
0.66748046875,
0.2132568359375,
0.68115234375,
0.080078125,
0.68115234375
],
[
0.5048828125,
0.66943359375,
0.890625,
0.66943359375,
0.890625,
0.68310546875,
0.5048828125,
0.68310546875
],
[
0.08087158203125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.08087158203125,
0.7041015625
],
[
0.5048828125,
0.68505859375,
0.890625,
0.68505859375,
0.890625,
0.69873046875,
0.5048828125,
0.69873046875
],
[
0.08087158203125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.71826171875,
0.08087158203125,
0.71826171875
],
[
0.50244140625,
0.69970703125,
0.8896484375,
0.69970703125,
0.8896484375,
0.71337890625,
0.50244140625,
0.71337890625
],
[
0.08087158203125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.08087158203125,
0.734375
],
[
0.5009765625,
0.71533203125,
0.59326171875,
0.71337890625,
0.59326171875,
0.72802734375,
0.5009765625,
0.72998046875
],
[
0.08087158203125,
0.73486328125,
0.4697265625,
0.73486328125,
0.4697265625,
0.75,
0.08087158203125,
0.75
],
[
0.529296875,
0.7392578125,
0.56396484375,
0.7392578125,
0.56396484375,
0.755859375,
0.529296875,
0.755859375
],
[
0.54345703125,
0.73486328125,
0.89794921875,
0.73779296875,
0.89794921875,
0.76318359375,
0.54345703125,
0.76025390625
],
[
0.7236328125,
0.7412109375,
0.771484375,
0.7412109375,
0.771484375,
0.74560546875,
0.7236328125,
0.74560546875
],
[
0.81787109375,
0.7392578125,
0.8759765625,
0.7392578125,
0.8759765625,
0.7470703125,
0.81787109375,
0.7470703125
],
[
0.08251953125,
0.751953125,
0.1478271484375,
0.751953125,
0.1478271484375,
0.76318359375,
0.08251953125,
0.76318359375
],
[
0.490234375,
0.75830078125,
0.5888671875,
0.75830078125,
0.5888671875,
0.77294921875,
0.490234375,
0.77294921875
],
[
0.10455322265625,
0.77099609375,
0.468017578125,
0.771484375,
0.468017578125,
0.7861328125,
0.10455322265625,
0.78515625
],
[
0.51318359375,
0.7763671875,
0.89453125,
0.7744140625,
0.89453125,
0.79541015625,
0.51318359375,
0.79736328125
],
[
0.5087890625,
0.7861328125,
0.515625,
0.7861328125,
0.515625,
0.7890625,
0.5087890625,
0.7890625
],
[
0.08251953125,
0.79443359375,
0.468017578125,
0.79443359375,
0.468017578125,
0.80810546875,
0.08251953125,
0.80810546875
],
[
0.50341796875,
0.79443359375,
0.890625,
0.79345703125,
0.890625,
0.80810546875,
0.50341796875,
0.8095703125
],
[
0.08251953125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.08251953125,
0.82373046875
],
[
0.5048828125,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82373046875,
0.5048828125,
0.82373046875
],
[
0.08087158203125,
0.82373046875,
0.4697265625,
0.82373046875,
0.4697265625,
0.84130859375,
0.08087158203125,
0.84130859375
],
[
0.50341796875,
0.82568359375,
0.64453125,
0.82568359375,
0.64453125,
0.8388671875,
0.50341796875,
0.8388671875
],
[
0.08251953125,
0.84228515625,
0.468017578125,
0.84228515625,
0.468017578125,
0.85595703125,
0.08251953125,
0.85595703125
],
[
0.5048828125,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.080078125,
0.85693359375,
0.46728515625,
0.85595703125,
0.46728515625,
0.87060546875,
0.080078125,
0.8720703125
],
[
0.50244140625,
0.85498046875,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.50244140625,
0.869140625
],
[
0.08087158203125,
0.87060546875,
0.468017578125,
0.87255859375,
0.468017578125,
0.88720703125,
0.08087158203125,
0.88525390625
],
[
0.50341796875,
0.869140625,
0.890625,
0.87060546875,
0.890625,
0.88525390625,
0.50341796875,
0.8837890625
],
[
0.08087158203125,
0.88818359375,
0.384033203125,
0.88818359375,
0.384033203125,
0.9013671875,
0.08087158203125,
0.9013671875
],
[
0.5048828125,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.5048828125,
0.90087890625
]
] |
[
"a",
"3.4.Top-Down Projection as Additional Signal",
"08",
"0.6",
"直观地图应进一步提高模型的鲁棒性。但是，我们",
"设置中的关键挑战：深度传感器中的噪声相当微妙，通",
"常几乎看不见（请参阅Fig.3（a，d））。但是，一旦",
"投影到2D布局，噪声就会表现为Fig.3（b，e）中明",
"(d)",
"显的总偏差，孔和阻塞。。为了应对这一挑战，我们使",
"0",
"用归一化软投影。Fig.3（c，f）中显示的归一化软投",
"",
"影s-projt类似于深度图给出的房间布局。请注意，它",
"们也比Fig.3（b，e）中给出的投影具有更多的相似性",
"Figure 3: Steps to infer an egocentric top-down projection",
"我们通过以下方式获得软投影：1）将深度观测值映",
"from depth.Top and bottom rows show inferred top-down",
"射到3D点云，2）使用2D自上而下的正交投影，以",
"projections from noisy and noiseless depth image at the",
"及3）相对于每个像素内的点数对投影进行归一化。提",
"same location: (b,e): top-down scatter plot. (c,f): the soft",
"供软投影作为端到端训练的VO模型的输入，该模型",
"top-down projection.As can be seen,after processing,(c)",
"and (f) share more similarities than (b) and (e), making the",
"学习如何正确使用它。附录中提供了如何计算软投影",
"representation more robust to depth noises.",
"的详细信息",
"3.5. VO Model Architecture, Training Details, and",
"3.3. Robustness to Uncertainty",
"Integration with Navigation Policy",
"模型体系结构。Eq：（??）中的视觉里程计模型f采用",
"除了利用几何不变性之外，我们发现进一步增加模",
"ResNet-18[17]主干提取视觉特征。为此，我们首先通",
"型的SE(2）估计的鲁棒性很重要。这很重要，因为测",
"量会产生噪声：1）即使由于观察噪声而使相机的位置",
"过",
"和方向相同，视觉观察也会有所不同。这使得观察结果",
"b(It)≤(RGBt,deptht,d-deptht,s-projt).(8)",
"的处理变得脆弱；2）驱动中的扰动会影响VO模型的",
"从Fig.2（c)中绘制的自我中心观察计算表示，然后，我",
"预测，因为它们会增加旋转和平移的方差。对于鲁棒",
"性，我们使用两种经典技术：",
"们沿着通道维度堆叠（(It)，(It+1)）以获得ResNet-",
"18输人。由于RGBt，deptht，d-depth,和s-projt分",
"合奏。为了提高鲁棒性，可以训练一组模型。对集合上",
"别具有三个，一个，N和一个通道，因此ResNet-18的",
"的预测进行平均通常会减少方差。然而，基于强化学习",
"输人是具有（2N+10）通道的张量。为了估计Hc.→ct",
"（RL）的导航系统需要数十亿个样本来训练一个好的策",
"",
"我们在ResNet-18特征提取器的顶部使用两个带有",
"略[53]。由于策略依赖于VO模型来提供代理的当前",
"Dropout的完全连接（FC）层。这些FC层在512维特",
"位置估计，因此提高推理速度并避免不必要的计算非",
"征上运行，并产生输出（→Ct+1，→C+1,Oct→Ct+1）",
"常重要。因此，我们发现训练一个CNN架构，同时将",
"。这里→Ct+1指的是在代理的向前方向上的平移，而",
"Dropout【46]添加到最后两个全连接（FC）层，而不是",
"集成多个模型，这很有帮助。这在经济上类似于训练大",
",→Ct+1指的是在垂直于地平面上的向前运动的方向",
"量合奏的行为[4，18]。在训练期间，Dropout以p的",
"上的平移",
"概率随机禁用FC层中的隐藏单元，本质上是从子网",
"VO训练。我们在数据集Dtrain",
"集合中采样。在推理过程中，FC层中的每个隐藏单元",
"{(It, It+1),Sc→Ct+1,ct-ct+)}≤[de,=c+} 上",
"都使用相同的因子p进行缩放，以模拟来自多个子网",
"训练视觉里程计模型f。。每个数据点由一对以自我为",
"络的预测的平均值",
"中心的观测值以及地面实况平移和旋转角度组成。对",
"深度离散化。此外，我们发现深度离散化可以更可靠地",
"该模型进行了优化，以共同最小化Eq.（3）和Eq.（4）",
"表示距离传感器的自我中心观察。具体来说，使用独热",
"i.e.中定义的回归损失和几何不变性损失，我们解决了",
"编码将单通道深度图depth离散化为具有N通道的表",
"minCvo",
"示d-depth。给定图像坐标（a,y）处的depth的像素，",
"7",
"[Areg Ce+Ans Cacan+At ceot ]",
"n.tran",
"cinv,rot",
"我们通过",
"dct-+Ci+1EDtrain",
"d-depth,(r,y) = 1 {depth(a,y) [zi-1, zi)},(7)",
"其中入reg入rans和入rt是用户指定的超参数。我",
"",
"获得d-depth的i通道的值，其中1(}表示指标函数，",
"们在实验中将它们设置为1.0。我们使用2.5×10-4的",
"【-1，z}是离散化间隔的端点。直观上，这增加了深",
"学习速率，用Adam「26】优化VO模型。训练期间的",
"度不确定性对min;2-i-l的绝对容忍度，因为除非",
"退出因子为P=0.2",
"深度条目越过间隔边界，否则将生成相同的表示。根据",
"导航策略培训。我们工作的重点是现实条件下的Point-",
"经验，我们发现使用端点zi=i·（zmax-zmin)/N可",
"Goal导航，Ei.e.，噜杂的观测和驱动以及无法访问",
"以很好地将等距离散化为N间隔。这里，zmax和zmin",
"GPU+Compass传感器。为了证明VO技术可以简单",
"分别是最大（10m）和最小深度（0m）值",
"地替代地面实况GPS+Compass传感器，我们直接使"
] |
{
"arxiv_id": "2108.11550",
"language": "zh",
"page_idx": 4
}
|
||
2110.08168-zh-1
|
2110.08168-zh-1.jpg
|
$\mathbf{D}\mathbf{YLE}:$抽象
长输人汇总的动态潜在提取
$\textbf{Ziming Mao}^{* 1}$ $\textbf{Chen Henry Wu}^{* 2}$ $\textbf{Ansong Ni}^{1}$ $\textbf{Yusen Zhang}^{3}$
Rui Zhang$^{3}$ $\dot{\textbf{Tao Yu}} ^{4}$ $\ddot{\textbf{Budhaditya Deb}} ^{5}$
$\textbf{Chenguang Zhu}^{5}$ Åhmed H. Awadallah$^5$ Dragomir Radev$^{1}$
$^{1}$ Yale University$^{2}$Carnegie Mellon University$^{3}$Penn State University
$^{4}$The University of Hong Kong$^5$ Microsoft Research ziming.mao@yale.edu, henrychenwu@cmu.edu
## Abstract
<FigureHere>
Figure 1: Overview of our approach. The input is a document $X$ (each $x\in X$ is a sentence) and an optional query $q$ , and the output is a summary $y.$
Transformer-based models have achieved state-of-the-art performance on short-input summarization. However, they still struggle with summarizing longer text. In this paper, we present DYLE, a novel dynamic latent extraction approach for abstractive long-input summarization. DYLE jointly trains an extractor and a generator and treats the extracted text snippets as the latent variable, allowing dynamic snippet-level attention weights during decoding. To provide adequate supervision, we propose simple yet effective heuristics for oracle extraction as well as a consistency loss term, which encourages the extractor to approximate the averaged dynamic weights predicted by the generator. We evaluate our method on different long-document and long-dialogue summarization tasks: GovReport, QMSum, and arXiv. Experiment results show that DYLE outperforms all existing methods on GovReport and QMSum, with gains up to 6.1 ROUGE, while yielding strong results on arXiv. Further analysis shows that the proposed dynamic weights provide interpretability of our generation process. 1
## 1 Introduction
的内存复杂性，以便它们可以处理更多令牌。其次，提取然后生成方法从输入中提取突出的文本，然后汇总提取的文本。提取器要么通过全面监督 (Zhong et al., 2021b) 进行独立训练， 要么使用强化学习 (Williams, 1992; Chen and Bansal, 2018; Bae et al., 2019; Bražinskas et al., 2021)进行优化。第三，提出了将源文本分为(Gidiotis and Tsoumakas, 2020; Wu et al., 2021; Liu et al., 2021) 部分的模型，这些部分分别进行汇总和组合以形成完整的摘要。第四，分层模型 (Rohde et al., 2021; Zhu et al., 2020) 通过捕获句子或话语级别的依赖关系来改善摘要。我们在 2 节中详细阐述了这四个方向及其局限性
我们认为，提取然后生成的方法模仿了一个人如何处理长输入摘要：首先识别文本中的重要信息，然后对其进行总结 For R1:(Kiy-) oumarsi, 2015; Sun et al., 2020) 。提取然后生成框架基于这样的假设，即对汇总有用的重要信息仅占输入的一小部分，这是一个明智的假设，因为输入长度很长。这种方法将源输入缩短到预设长度，这解决了模型无法处理超过一定限制的更长输入的主要挑战。然而，以前单独训练的提取然后生成的方法是有限的，因为它们遭受从提取器到生成器的级联误差。虽然引人了各种强化学习技术来桥接这两个步骤。但它们有明显的缺点(在 3.3 节中讨论),我们认为长输入使得这种方法次优
在本文中，我们提出了一种用于长输入摘
基于 Transformer 的 (Vaswani et al.,2017) 预训练语言模型 (PLM),如 BART (Lewis et al., 2020a) 和 T5 (Raffel et al.,2020) ,在短文本摘要上取得了最先进的性能。但是，由于完全自我关注的 (Tay et al., 2020a) 的内存复杂性很高PLM 仍然难以处理长输人 (Rohde et al., 2021) 。模型效率和摘要质量提出了一对挑战 (Huang et al.,2021): 模型需要捕获分散在长输入中的信息，同时保持较低的计算成本。
以前的模型主要以四种方式处理长输入汇总。首先，稀疏注意力 (Child et al.,2019; Beltagy et al., 2020; Tay et al., 2020b) 用于降低变形金刚
$^{*}$Equal Contributions.
$^{1}$Our code is available at: https://github.com/
Yale-LILY/DYLE
|
[
[
0.426513671875,
0.08197021484375,
0.5693359375,
0.08489990234375,
0.5693359375,
0.10394287109375,
0.426513671875,
0.1009521484375
],
[
0.358642578125,
0.105712890625,
0.64208984375,
0.105712890625,
0.64208984375,
0.12054443359375,
0.358642578125,
0.12054443359375
],
[
0.214111328125,
0.133056640625,
0.34765625,
0.1318359375,
0.34765625,
0.146728515625,
0.214111328125,
0.1478271484375
],
[
0.359375,
0.133056640625,
0.5517578125,
0.1300048828125,
0.5517578125,
0.146728515625,
0.359375,
0.149658203125
],
[
0.5439453125,
0.133056640625,
0.6708984375,
0.1318359375,
0.6708984375,
0.146728515625,
0.5439453125,
0.1478271484375
],
[
0.66748046875,
0.1341552734375,
0.791015625,
0.1341552734375,
0.791015625,
0.1478271484375,
0.66748046875,
0.1478271484375
],
[
0.54150390625,
0.149658203125,
0.69287109375,
0.1484375,
0.69287109375,
0.163330078125,
0.54150390625,
0.1644287109375
],
[
0.341796875,
0.1353759765625,
0.364501953125,
0.1353759765625,
0.364501953125,
0.142578125,
0.341796875,
0.142578125
],
[
0.32666015625,
0.15087890625,
0.429931640625,
0.15087890625,
0.429931640625,
0.162109375,
0.32666015625,
0.162109375
],
[
0.426513671875,
0.15087890625,
0.5283203125,
0.1484375,
0.5283203125,
0.163330078125,
0.426513671875,
0.1650390625
],
[
0.5341796875,
0.15380859375,
0.54150390625,
0.15380859375,
0.54150390625,
0.157958984375,
0.5341796875,
0.157958984375
],
[
0.240966796875,
0.166259765625,
0.78076171875,
0.1650390625,
0.78076171875,
0.179931640625,
0.240966796875,
0.18115234375
],
[
0.349365234375,
0.179931640625,
0.60791015625,
0.1817626953125,
0.60791015625,
0.1995849609375,
0.349365234375,
0.19775390625
],
[
0.59423828125,
0.179931640625,
0.8095703125,
0.182861328125,
0.8095703125,
0.20068359375,
0.59423828125,
0.19775390625
],
[
0.28369140625,
0.196533203125,
0.73876953125,
0.198974609375,
0.73876953125,
0.2161865234375,
0.28369140625,
0.21435546875
],
[
0.193115234375,
0.18115234375,
0.343505859375,
0.182861328125,
0.343505859375,
0.19775390625,
0.193115234375,
0.1953125
],
[
0.26025390625,
0.21728515625,
0.7421875,
0.21728515625,
0.7421875,
0.23095703125,
0.26025390625,
0.23095703125
],
[
0.615234375,
0.251708984375,
0.75390625,
0.251708984375,
0.75390625,
0.265380859375,
0.615234375,
0.265380859375
],
[
0.771484375,
0.251708984375,
0.873046875,
0.251708984375,
0.873046875,
0.262451171875,
0.771484375,
0.262451171875
],
[
0.5224609375,
0.265380859375,
0.59375,
0.265380859375,
0.59375,
0.276611328125,
0.5224609375,
0.276611328125
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.70361328125,
0.270751953125,
0.80615234375,
0.270751953125,
0.80615234375,
0.281982421875,
0.70361328125,
0.281982421875
],
[
0.14697265625,
0.283935546875,
0.45751953125,
0.283935546875,
0.45751953125,
0.29443359375,
0.14697265625,
0.29443359375
],
[
0.5634765625,
0.283203125,
0.57275390625,
0.283203125,
0.57275390625,
0.292236328125,
0.5634765625,
0.292236328125
],
[
0.82275390625,
0.2861328125,
0.82958984375,
0.2861328125,
0.82958984375,
0.29150390625,
0.82275390625,
0.29150390625
],
[
0.145263671875,
0.29638671875,
0.459228515625,
0.29638671875,
0.459228515625,
0.31005859375,
0.145263671875,
0.31005859375
],
[
0.5458984375,
0.298583984375,
0.611328125,
0.300537109375,
0.609375,
0.3154296875,
0.5439453125,
0.3134765625
],
[
0.75390625,
0.300537109375,
0.8193359375,
0.300537109375,
0.8193359375,
0.3154296875,
0.75390625,
0.3154296875
],
[
0.14697265625,
0.3134765625,
0.264404296875,
0.3134765625,
0.264404296875,
0.32373046875,
0.14697265625,
0.32373046875
],
[
0.2568359375,
0.311279296875,
0.45751953125,
0.311279296875,
0.45751953125,
0.32470703125,
0.2568359375,
0.32470703125
],
[
0.1435546875,
0.325927734375,
0.45751953125,
0.325927734375,
0.45751953125,
0.339111328125,
0.1435546875,
0.339111328125
],
[
0.79833984375,
0.3349609375,
0.83642578125,
0.3349609375,
0.83642578125,
0.3466796875,
0.79833984375,
0.3466796875
],
[
0.14697265625,
0.341552734375,
0.456787109375,
0.341552734375,
0.456787109375,
0.3515625,
0.14697265625,
0.3515625
],
[
0.654296875,
0.342529296875,
0.7109375,
0.344482421875,
0.70947265625,
0.359375,
0.6533203125,
0.35693359375
],
[
0.8017578125,
0.3466796875,
0.8330078125,
0.3466796875,
0.8330078125,
0.359375,
0.8017578125,
0.359375
],
[
0.145263671875,
0.355224609375,
0.459228515625,
0.355224609375,
0.459228515625,
0.3681640625,
0.145263671875,
0.3681640625
],
[
0.599609375,
0.361083984375,
0.6884765625,
0.361083984375,
0.6884765625,
0.372314453125,
0.599609375,
0.372314453125
],
[
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
],
[
0.51318359375,
0.385498046875,
0.5810546875,
0.385498046875,
0.5810546875,
0.396728515625,
0.51318359375,
0.396728515625
],
[
0.5810546875,
0.388427734375,
0.58837890625,
0.388427734375,
0.58837890625,
0.393798828125,
0.5810546875,
0.393798828125
],
[
0.58447265625,
0.38427734375,
0.8818359375,
0.38427734375,
0.8818359375,
0.3984375,
0.58447265625,
0.3984375
],
[
0.145263671875,
0.396728515625,
0.459228515625,
0.396728515625,
0.459228515625,
0.410888671875,
0.145263671875,
0.410888671875
],
[
0.51123046875,
0.3984375,
0.880859375,
0.3984375,
0.880859375,
0.412109375,
0.51123046875,
0.412109375
],
[
0.145263671875,
0.410400390625,
0.459228515625,
0.410400390625,
0.459228515625,
0.42333984375,
0.145263671875,
0.42333984375
],
[
0.51123046875,
0.413330078125,
0.82958984375,
0.413330078125,
0.82958984375,
0.427490234375,
0.51123046875,
0.427490234375
],
[
0.145263671875,
0.424560546875,
0.45751953125,
0.424560546875,
0.45751953125,
0.438232421875,
0.145263671875,
0.438232421875
],
[
0.145263671875,
0.439453125,
0.459228515625,
0.439453125,
0.459228515625,
0.453125,
0.145263671875,
0.453125
],
[
0.14697265625,
0.454833984375,
0.45751953125,
0.454833984375,
0.45751953125,
0.465576171875,
0.14697265625,
0.465576171875
],
[
0.5146484375,
0.45361328125,
0.880859375,
0.45361328125,
0.880859375,
0.46728515625,
0.5146484375,
0.46728515625
],
[
0.145263671875,
0.46728515625,
0.459228515625,
0.46728515625,
0.459228515625,
0.48095703125,
0.145263671875,
0.48095703125
],
[
0.51318359375,
0.4697265625,
0.880859375,
0.4697265625,
0.880859375,
0.48388671875,
0.51318359375,
0.48388671875
],
[
0.145263671875,
0.482177734375,
0.45751953125,
0.482177734375,
0.45751953125,
0.495361328125,
0.145263671875,
0.495361328125
],
[
0.5146484375,
0.486328125,
0.880859375,
0.486328125,
0.880859375,
0.49951171875,
0.5146484375,
0.49951171875
],
[
0.1427001953125,
0.495361328125,
0.459228515625,
0.49462890625,
0.459228515625,
0.5087890625,
0.1427001953125,
0.51025390625
],
[
0.5146484375,
0.501953125,
0.87744140625,
0.501953125,
0.87744140625,
0.515625,
0.5146484375,
0.515625
],
[
0.145263671875,
0.51025390625,
0.45751953125,
0.51025390625,
0.45751953125,
0.52392578125,
0.145263671875,
0.52392578125
],
[
0.51318359375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.1485595703125,
0.525390625,
0.45751953125,
0.525390625,
0.45751953125,
0.5361328125,
0.1485595703125,
0.5361328125
],
[
0.145263671875,
0.5380859375,
0.459228515625,
0.5380859375,
0.459228515625,
0.5517578125,
0.145263671875,
0.5517578125
],
[
0.51318359375,
0.53369140625,
0.880859375,
0.53369140625,
0.880859375,
0.54736328125,
0.51318359375,
0.54736328125
],
[
0.14697265625,
0.552734375,
0.459228515625,
0.552734375,
0.459228515625,
0.56640625,
0.14697265625,
0.56640625
],
[
0.51123046875,
0.548828125,
0.880859375,
0.5498046875,
0.880859375,
0.56396484375,
0.51123046875,
0.5634765625
],
[
0.145263671875,
0.5654296875,
0.456787109375,
0.5654296875,
0.456787109375,
0.5791015625,
0.145263671875,
0.5791015625
],
[
0.51318359375,
0.5654296875,
0.880859375,
0.5654296875,
0.880859375,
0.5791015625,
0.51318359375,
0.5791015625
],
[
0.145263671875,
0.58056640625,
0.45751953125,
0.58056640625,
0.45751953125,
0.59423828125,
0.145263671875,
0.59423828125
],
[
0.51123046875,
0.5830078125,
0.880859375,
0.5830078125,
0.880859375,
0.5966796875,
0.51123046875,
0.5966796875
],
[
0.14697265625,
0.5966796875,
0.45751953125,
0.5966796875,
0.45751953125,
0.60693359375,
0.14697265625,
0.60693359375
],
[
0.51123046875,
0.59716796875,
0.87890625,
0.59716796875,
0.87890625,
0.61083984375,
0.51123046875,
0.61083984375
],
[
0.145263671875,
0.609375,
0.459228515625,
0.609375,
0.459228515625,
0.62255859375,
0.145263671875,
0.62255859375
],
[
0.51123046875,
0.61328125,
0.880859375,
0.6142578125,
0.880859375,
0.62890625,
0.51123046875,
0.62744140625
],
[
0.145263671875,
0.62353515625,
0.403076171875,
0.62353515625,
0.403076171875,
0.63720703125,
0.145263671875,
0.63720703125
],
[
0.51123046875,
0.6298828125,
0.880859375,
0.6298828125,
0.880859375,
0.6435546875,
0.51123046875,
0.6435546875
],
[
0.51318359375,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.1158447265625,
0.65380859375,
0.145263671875,
0.65380859375,
0.145263671875,
0.6650390625,
0.1158447265625,
0.6650390625
],
[
0.151123046875,
0.65380859375,
0.2626953125,
0.65380859375,
0.2626953125,
0.6650390625,
0.151123046875,
0.6650390625
],
[
0.51318359375,
0.6630859375,
0.5341796875,
0.6630859375,
0.5341796875,
0.6767578125,
0.51318359375,
0.6767578125
],
[
0.1175537109375,
0.677734375,
0.48779296875,
0.6787109375,
0.48779296875,
0.693359375,
0.1175537109375,
0.6923828125
],
[
0.53076171875,
0.6787109375,
0.87890625,
0.6787109375,
0.87890625,
0.6923828125,
0.53076171875,
0.6923828125
],
[
0.118408203125,
0.69482421875,
0.48779296875,
0.693359375,
0.48779296875,
0.7080078125,
0.118408203125,
0.708984375
],
[
0.515625,
0.6953125,
0.880859375,
0.6953125,
0.880859375,
0.708984375,
0.515625,
0.708984375
],
[
0.118408203125,
0.71044921875,
0.486083984375,
0.71044921875,
0.486083984375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.51318359375,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.51318359375,
0.724609375
],
[
0.118408203125,
0.72705078125,
0.483642578125,
0.72705078125,
0.483642578125,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.509765625,
0.7255859375,
0.87890625,
0.7255859375,
0.87890625,
0.7392578125,
0.509765625,
0.7392578125
],
[
0.118408203125,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.5146484375,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.5146484375,
0.75732421875
],
[
0.1175537109375,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.1175537109375,
0.7724609375
],
[
0.5146484375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.5146484375,
0.77392578125
],
[
0.118408203125,
0.775390625,
0.486083984375,
0.775390625,
0.486083984375,
0.7890625,
0.118408203125,
0.7890625
],
[
0.5146484375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.79052734375,
0.5146484375,
0.79052734375
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51318359375,
0.7919921875,
0.87158203125,
0.7919921875,
0.87158203125,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.8671875,
0.79638671875,
0.875,
0.79638671875,
0.875,
0.80029296875,
0.8671875,
0.80029296875
],
[
0.118408203125,
0.80859375,
0.382080078125,
0.80859375,
0.382080078125,
0.822265625,
0.118408203125,
0.822265625
],
[
0.5146484375,
0.80810546875,
0.87890625,
0.80810546875,
0.87890625,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.1368408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.1368408203125,
0.83837890625
],
[
0.51123046875,
0.822265625,
0.880859375,
0.82373046875,
0.880859375,
0.83837890625,
0.51123046875,
0.83740234375
],
[
0.118408203125,
0.84130859375,
0.486083984375,
0.84130859375,
0.486083984375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.5146484375,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85400390625,
0.5146484375,
0.85400390625
],
[
0.1175537109375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51123046875,
0.85546875,
0.87890625,
0.85546875,
0.87890625,
0.869140625,
0.51123046875,
0.869140625
],
[
0.5146484375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.5146484375,
0.8857421875
],
[
0.145263671875,
0.8798828125,
0.27783203125,
0.8798828125,
0.27783203125,
0.8935546875,
0.145263671875,
0.8935546875
],
[
0.1376953125,
0.89111328125,
0.48779296875,
0.892578125,
0.48779296875,
0.90673828125,
0.1376953125,
0.90625
],
[
0.5146484375,
0.88818359375,
0.75390625,
0.88818359375,
0.75390625,
0.90283203125,
0.5146484375,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.2476806640625,
0.90478515625,
0.2476806640625,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.53076171875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.53076171875,
0.91845703125
]
] |
[
"DYLE：抽象",
"长输人汇总的动态潜在提取",
"Ziming Mao* 1",
"Chen Henry Wu*2",
"Ansong Nil",
"Yusen Zhang?",
"Budhaditya Deb",
"",
"Rui Zhang3",
" TaoYu4",
"",
"Chenguang Zhu5Ahmed H.Awadallah5Dragomir Radevl",
"2 Carnegie Mellon University",
"3 Penn State University",
"4 The University of Hong Kong 5 Microsoft Research",
"1Yale University",
"ziming.mao@yale.edu, henrychenwu@cmu.edu",
"Consistency loss",
"Generation loss",
"Oracle loss",
"Abstract",
"Dynamic weights",
"Transformer-basedmodelshaveachieved",
"1",
"口",
"state-of-the-art performance on short-input",
"Extractor",
"Generator",
"summarization.",
"However, they still struggle",
"with summarizing longer text. In this pa-",
"",
"per,wepresent DYLE,a novel dynamiclatent",
"Queryq",
"y<t",
"extraction approach for abstractive long-input",
"Document X",
"summarization. DYLE jointly trains an extrac-",
"tor and a generator and treats the extracted",
"Figure l:",
"",
"Overview of our approach. The input is a",
"text snippets as the latent variable, allowing",
"document X (each α e X is a sentence) and an op",
"dynamic snippet-level attention weights dur-",
"tional query q , and the output is a summary y .",
"ing decoding.\"To provide adequate supervi",
"sion, we propose simple yet effective heuris-",
"tics for oracle extraction as well as a con-",
"的内存复杂性，以便它们可以处理更多令牌。",
"sistency loss term, which encourages the ex-",
"其次，提取然后生成方法从输人中提取突出的",
"tractor to approximate the averaged dynamic",
"文本，然后汇总提取的文本。提取器要么通过",
"weights predicted by the generator. We eval-",
"全面监督(Zhongetal.,2021b）进行独立训练",
"uate our method on different long-document",
"要么使用强化学习（Williams，1992；Chenand",
"and long-dialogue summarization tasks: Gov-",
"Report, QMSum, and arXiv.Experiment re-",
"Bansal, 2018; Bae et al., 2019; Brazinskas et al..",
"sults show that DYLE outperforms all exist-",
"2021)进行优化。第三，提出了将源文本分为",
"ing methods on GovReport and QMSum, with",
"(Gidiotis and Tsoumakas, 2020; Wu et al., 2021;",
"gains up to 6.1 ROUGE, while yielding strong",
"Liuetal.,2021）部分的模型，这些部分分别进",
"results on arXiv.Further analysis shows that",
"行汇总和组合以形成完整的摘要。第四，分层",
"the proposed dynamic weights provide inter-",
"模型(Rohdeetal.,2021;Zhuetal.,2020)通过",
"pretability of our generation process.",
"捕获句子或话语级别的依赖关系来改善摘要。",
"我们在2节中详细阐述了这四个方向及其局限",
"1",
"Introduction",
"性",
"基于Transformer的（Vaswani et al.，2017）预",
"我们认为，提取然后生成的方法模仿了一个",
"训练语言模型(PLM)，如BART(Lewisetal.",
"人如何处理长输人摘要：首先识别文本中的",
"2020a）和T5（Raffeletal.,2020），在短文本摘",
"重要信息，然后对其进行总结ForRl：（Kiy-",
"要上取得了最先进的性能。但是，由于完全自",
"oumarsi.2015:Sun et al.2020）。提取然后生",
"我关注的（Tayetal.,2020a）的内存复杂性很高，",
"成框架基于这样的假设，即对汇总有用的重要",
"PLM仍然难以处理长输人(Rohdeetal.,2021)",
"信息仅占输入的一小部分，这是一个明智的假",
"。模型效率和摘要质量提出了一对挑战(Huang",
"设，因为输入长度很长。这种方法将源输人缩",
"etal.,2021）：模型需要捕获分散在长输人中的",
"短到预设长度，这解决了模型无法处理超过",
"-",
"信息，同时保持较低的计算成本。",
"定限制的更长输入的主要挑战。然而，以前单",
"以前的模型主要以四种方式处理长输入汇",
"独训练的提取然后生成的方法是有限的，因为",
"总。首先，稀疏注意力（Childetal.2019;Beltagy",
"它们遭受从提取器到生成器的级联误差。虽然",
"etal.,2020;Tayetal.2020b）用于降低变形金刚",
"引人了各种强化学习技术来桥接这两个步骤",
"但它们有明显的缺点（在3.3节中讨论），我们",
"*Equal Contributions.",
"'our code is available at: https://github.com/",
"认为长输入使得这种方法次优",
"Yale-LILY/DYLE",
"在本文中，我们提出了一种用于长输入摘"
] |
{
"arxiv_id": "2110.08168",
"language": "zh",
"page_idx": 1
}
|
||
2110.08168-zh-2
|
2110.08168-zh-2.jpg
|
潜伏。对于输出令牌，DYLE 分别计算以每个输入片段为条件的概率，并且通过在生成器以先前生成的令牌为条件分配的学习动态权重下对所有输入片段进行边缘化来计算其生成概率我们优化了具有两个代理损耗的提取器。首
先，我们根据参考摘要计算提取 oracle, 并在最佳 ROUGE 分数上进行贪婪搜索。这些 oracle 片段被用作提取器学习信号的目标。此外，我们提出了一致性损失，以鼓励提取器将其在代码段上的预测权重近似为生成器预测的平均动态权重
我们在三个长输入摘要数据集上进行了实验：GovReport (Huang et al., 2021) 和 arXiv (Cohan et al.,2018) 用于长文档摘要，QMSum (Zhong et al., 2021b) 用于长对话摘要。我们的方法 DYLE 在 GovReport 和 QMSum 上的表现大大优于现有方法，同时在 arXiv 上取得了很好的效果。值得注意的是，DYLE 的收益为 4。2/6。1/4。在 GovReport 上，ROUGE-1/2/L 比以前的最佳方法高出0分。这些实验证明了DYLE 对多个长输入汇总任务的通用性
我们总结我们的贡献如下：
· We introduce DYLE, a dynamic latent extrac-
tion approach for abstractive long-input summarization. DYLE better captures information in the long input and reduces computational cost;
$\cdot$ We propose multiple auxiliary optimizations for the effective training of DYLE: 1) extractive oracle as a learning signal for the extractor; 2) consistency loss that bridges extraction and generation; 3) hybrid training methods that make the extraction more robust;
$\bullet$ Experimental results show that DYLE largely outperforms the state-of-the-art on two long input summarization datasets. We also conducted a detailed analysis that shows dynamic weights improve model interpretability.
## 2 Related Work
For R3: We introduce in detail the four main categories of methods in recent work to address longinput summarization tasks.
Sparse attention mechanism 完全注意力机制具有二次记忆成本。先前的研究工作提出了
要的新方法：Dy namic L atent E xtractionforAb- 不同的稀疏注意机制来降低内存成本。LongstractiveSummarization ( DYLE )。DYLE 联合训 former (Beltagy et al., 2020) 使用块和全局注意练提取器和生成器，并使提取的文本片段保持 力模式的扩展滑动窗口。BigBird (Zaheer et al.,
2020) 采用滑动窗口和随机块。重整器 (Kitaev et al., 2020) 使用局部敏感哈希。除了优化编码器的自我关注之外，Huang et al.(2021)还提出了头向位置步幅，以降低编码器-解码器关注的成本。然而，稀疏的注意力减少了预训练的好处，并牺牲了部分接受领域
Extract-then-generate method 此方法从输
人中提取重要的文本片段，然后生成总体摘要。这些方法中的大多数都是单独训练的 (Zhang et al., 2019; Lebanoff et al., 2019; Xu and Durrett, 2019; Bajaj et al., 2021; Zhang et al., 2021b) , 当我们将提取的代码段传递给生成器时，它们会遭受信息丢失。一些方法试图通过桥接两个阶段来减少损失。Chen and Bansal (2018) 采用具有句子级策略梯度的强化学习(RL)。Bae et al. (2019)提出了摘要级策略梯度。使用 RL 会在长输入文本上遇到各种缺点，这将在 3.3 节中详细说明。DYLE 是不同的，因为我们共同训练一个提取然后生成的模型，用于使用潜在变量进行总结
Divide-and-conquer approach 长输入汇总
中的一种常见方法是分而治之(Gidiotis and Tsoumakas, 2020; Grail et al., 2021; Zhang et al., 2021a)。它将一个长的输入分成多个部分，这些部分分别汇总并组合在一起以生成最终的摘要。但是，这些模型不能捕获各部分之间的上下文依赖关系，并假设输入具有特定的结构
Hierarchical models 已经提出了各种分层模型来处理较长的输入。Cohan et al.(2018) 使用分层编码器和话语感知解码器对文档话语结构进行建模。HAT-Bart (Rohde et al., 2021) 提出了一种新的基于分层注意力转换器的体系结构，该体系结构试图捕获句子和段落级信息。HMNet (Zhu et al., 2020) 构建了一个层次结构， 其中包括话语级信息和演讲者角色。然而，这些模型主要关注模型性能，而不是减少内存和计算成本
3 Our Approach
我们的方法概述如图 1所示。在 3.1节中， 我们制定了我们的任务和提取器-生成器框架。在 3.2 节中，我们介绍了长输入提取器的参数化。在 3.3 节中，我们介绍了生成器公式和新颖的一致性损失。提取器模块同时优化了一致性损失和 oracle 损失，我们在 3.4 节中对此进行了详细说明。3.5 节总结了总体培训目标
|
[
[
0.12005615234375,
0.0872802734375,
0.490234375,
0.0872802734375,
0.490234375,
0.1009521484375,
0.12005615234375,
0.1009521484375
],
[
0.51318359375,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.118408203125,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.118408203125,
0.12054443359375,
0.484375,
0.12054443359375,
0.484375,
0.1312255859375,
0.118408203125,
0.1312255859375
],
[
0.5146484375,
0.12054443359375,
0.8759765625,
0.12054443359375,
0.8759765625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51123046875,
0.1341552734375,
0.8759765625,
0.1341552734375,
0.8759765625,
0.1478271484375,
0.51123046875,
0.1478271484375
],
[
0.1175537109375,
0.15087890625,
0.483642578125,
0.15087890625,
0.483642578125,
0.1644287109375,
0.1175537109375,
0.1644287109375
],
[
0.51123046875,
0.149658203125,
0.880859375,
0.15087890625,
0.880859375,
0.1650390625,
0.51123046875,
0.1644287109375
],
[
0.118408203125,
0.16748046875,
0.484375,
0.16748046875,
0.484375,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.87890625,
0.16748046875,
0.87890625,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.1175537109375,
0.1817626953125,
0.486083984375,
0.1817626953125,
0.486083984375,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.5146484375,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.5146484375,
0.196533203125
],
[
0.1376953125,
0.20068359375,
0.483642578125,
0.20068359375,
0.483642578125,
0.21142578125,
0.1376953125,
0.21142578125
],
[
0.51123046875,
0.20068359375,
0.87744140625,
0.20068359375,
0.87744140625,
0.21142578125,
0.51123046875,
0.21142578125
],
[
0.1175537109375,
0.215576171875,
0.486083984375,
0.215576171875,
0.486083984375,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.51123046875,
0.215576171875,
0.75244140625,
0.215576171875,
0.75244140625,
0.2286376953125,
0.51123046875,
0.2286376953125
],
[
0.1175537109375,
0.2298583984375,
0.486083984375,
0.23095703125,
0.486083984375,
0.2452392578125,
0.1175537109375,
0.24462890625
],
[
0.1175537109375,
0.2476806640625,
0.486083984375,
0.2476806640625,
0.486083984375,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.51904296875,
0.243408203125,
0.880859375,
0.243408203125,
0.880859375,
0.257080078125,
0.51904296875,
0.257080078125
],
[
0.118408203125,
0.26416015625,
0.484375,
0.26416015625,
0.484375,
0.27490234375,
0.118408203125,
0.27490234375
],
[
0.51123046875,
0.2587890625,
0.8818359375,
0.260009765625,
0.8818359375,
0.273681640625,
0.51123046875,
0.2724609375
],
[
0.1175537109375,
0.279052734375,
0.484375,
0.279052734375,
0.484375,
0.292236328125,
0.1175537109375,
0.292236328125
],
[
0.51123046875,
0.27490234375,
0.880859375,
0.275634765625,
0.880859375,
0.290283203125,
0.51123046875,
0.289306640625
],
[
0.118408203125,
0.295654296875,
0.175537109375,
0.295654296875,
0.175537109375,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.515625,
0.29345703125,
0.87890625,
0.29345703125,
0.87890625,
0.303955078125,
0.515625,
0.303955078125
],
[
0.1368408203125,
0.312255859375,
0.484375,
0.312255859375,
0.484375,
0.325927734375,
0.1368408203125,
0.325927734375
],
[
0.51318359375,
0.30810546875,
0.880859375,
0.30810546875,
0.880859375,
0.32177734375,
0.51318359375,
0.32177734375
],
[
0.118408203125,
0.327880859375,
0.486083984375,
0.327880859375,
0.486083984375,
0.341552734375,
0.118408203125,
0.341552734375
],
[
0.51318359375,
0.32470703125,
0.880859375,
0.32470703125,
0.880859375,
0.33837890625,
0.51318359375,
0.33837890625
],
[
0.118408203125,
0.344482421875,
0.486083984375,
0.344482421875,
0.486083984375,
0.358154296875,
0.118408203125,
0.358154296875
],
[
0.51318359375,
0.34033203125,
0.880859375,
0.34033203125,
0.880859375,
0.35400390625,
0.51318359375,
0.35400390625
],
[
0.118408203125,
0.35986328125,
0.486083984375,
0.35986328125,
0.486083984375,
0.37353515625,
0.118408203125,
0.37353515625
],
[
0.51318359375,
0.355712890625,
0.880859375,
0.355712890625,
0.880859375,
0.369384765625,
0.51318359375,
0.369384765625
],
[
0.118408203125,
0.377197265625,
0.486083984375,
0.377197265625,
0.486083984375,
0.39013671875,
0.118408203125,
0.39013671875
],
[
0.51318359375,
0.371826171875,
0.87890625,
0.371826171875,
0.87890625,
0.385498046875,
0.51318359375,
0.385498046875
],
[
0.118408203125,
0.392578125,
0.486083984375,
0.392578125,
0.486083984375,
0.40673828125,
0.118408203125,
0.40673828125
],
[
0.509765625,
0.38720703125,
0.880859375,
0.38720703125,
0.880859375,
0.40087890625,
0.509765625,
0.40087890625
],
[
0.1175537109375,
0.4091796875,
0.484375,
0.4091796875,
0.484375,
0.4228515625,
0.1175537109375,
0.4228515625
],
[
0.5146484375,
0.40380859375,
0.880859375,
0.40380859375,
0.880859375,
0.418701171875,
0.5146484375,
0.418701171875
],
[
0.118408203125,
0.424560546875,
0.486083984375,
0.424560546875,
0.486083984375,
0.438232421875,
0.118408203125,
0.438232421875
],
[
0.51318359375,
0.42041015625,
0.880859375,
0.42041015625,
0.880859375,
0.43408203125,
0.51318359375,
0.43408203125
],
[
0.51123046875,
0.435302734375,
0.880859375,
0.435791015625,
0.880859375,
0.45068359375,
0.51123046875,
0.449462890625
],
[
0.12005615234375,
0.4423828125,
0.483642578125,
0.4423828125,
0.483642578125,
0.453125,
0.12005615234375,
0.453125
],
[
0.1175537109375,
0.457275390625,
0.4248046875,
0.457275390625,
0.4248046875,
0.47021484375,
0.1175537109375,
0.47021484375
],
[
0.515625,
0.45361328125,
0.60546875,
0.45361328125,
0.60546875,
0.465576171875,
0.515625,
0.465576171875
],
[
0.1376953125,
0.474365234375,
0.344970703125,
0.474365234375,
0.344970703125,
0.485107421875,
0.1376953125,
0.485107421875
],
[
0.51708984375,
0.47802734375,
0.8818359375,
0.478515625,
0.8818359375,
0.496337890625,
0.51708984375,
0.495361328125
],
[
0.1376953125,
0.49951171875,
0.490234375,
0.49951171875,
0.490234375,
0.51318359375,
0.1376953125,
0.51318359375
],
[
0.51318359375,
0.496337890625,
0.880859375,
0.496337890625,
0.880859375,
0.51025390625,
0.51318359375,
0.51025390625
],
[
0.15283203125,
0.5166015625,
0.489501953125,
0.5166015625,
0.489501953125,
0.52978515625,
0.15283203125,
0.52978515625
],
[
0.51123046875,
0.5126953125,
0.87890625,
0.5126953125,
0.87890625,
0.525390625,
0.51123046875,
0.525390625
],
[
0.154541015625,
0.533203125,
0.490234375,
0.533203125,
0.490234375,
0.54638671875,
0.154541015625,
0.54638671875
],
[
0.51123046875,
0.52783203125,
0.87890625,
0.52783203125,
0.87890625,
0.54150390625,
0.51123046875,
0.54150390625
],
[
0.151123046875,
0.54736328125,
0.48779296875,
0.548828125,
0.48779296875,
0.5634765625,
0.151123046875,
0.5625
],
[
0.51318359375,
0.54345703125,
0.880859375,
0.5458984375,
0.880859375,
0.55859375,
0.51318359375,
0.55712890625
],
[
0.15283203125,
0.56396484375,
0.2392578125,
0.56396484375,
0.2392578125,
0.5791015625,
0.15283203125,
0.5791015625
],
[
0.51318359375,
0.56005859375,
0.87890625,
0.56005859375,
0.87890625,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.51708984375,
0.5791015625,
0.861328125,
0.5791015625,
0.861328125,
0.5888671875,
0.51708984375,
0.5888671875
],
[
0.1376953125,
0.59130859375,
0.484375,
0.59130859375,
0.484375,
0.60498046875,
0.1376953125,
0.60498046875
],
[
0.15283203125,
0.60693359375,
0.489501953125,
0.60693359375,
0.489501953125,
0.62060546875,
0.15283203125,
0.62060546875
],
[
0.51708984375,
0.60498046875,
0.880859375,
0.60498046875,
0.880859375,
0.6181640625,
0.51708984375,
0.6181640625
],
[
0.15283203125,
0.62353515625,
0.489501953125,
0.62353515625,
0.489501953125,
0.63720703125,
0.15283203125,
0.63720703125
],
[
0.51318359375,
0.62158203125,
0.880859375,
0.62158203125,
0.880859375,
0.63525390625,
0.51318359375,
0.63525390625
],
[
0.15283203125,
0.6396484375,
0.491943359375,
0.6396484375,
0.491943359375,
0.65283203125,
0.15283203125,
0.65283203125
],
[
0.51123046875,
0.6357421875,
0.87890625,
0.6357421875,
0.87890625,
0.6494140625,
0.51123046875,
0.6494140625
],
[
0.15283203125,
0.65478515625,
0.490234375,
0.65478515625,
0.490234375,
0.66943359375,
0.15283203125,
0.66943359375
],
[
0.51318359375,
0.65380859375,
0.87744140625,
0.65380859375,
0.87744140625,
0.66455078125,
0.51318359375,
0.66455078125
],
[
0.154541015625,
0.67138671875,
0.459228515625,
0.67138671875,
0.459228515625,
0.68505859375,
0.154541015625,
0.68505859375
],
[
0.51318359375,
0.66943359375,
0.880859375,
0.66943359375,
0.880859375,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.51123046875,
0.68505859375,
0.8818359375,
0.68603515625,
0.8818359375,
0.69970703125,
0.51123046875,
0.69873046875
],
[
0.1368408203125,
0.69775390625,
0.484375,
0.69873046875,
0.484375,
0.71337890625,
0.1368408203125,
0.7119140625
],
[
0.51123046875,
0.70068359375,
0.87890625,
0.70068359375,
0.87890625,
0.71435546875,
0.51123046875,
0.71435546875
],
[
0.154541015625,
0.71533203125,
0.48779296875,
0.71533203125,
0.48779296875,
0.728515625,
0.154541015625,
0.728515625
],
[
0.51318359375,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.732421875,
0.51318359375,
0.732421875
],
[
0.154541015625,
0.73095703125,
0.489501953125,
0.73095703125,
0.489501953125,
0.74462890625,
0.154541015625,
0.74462890625
],
[
0.51123046875,
0.73291015625,
0.880859375,
0.73388671875,
0.880859375,
0.7490234375,
0.51123046875,
0.74755859375
],
[
0.154541015625,
0.74658203125,
0.490234375,
0.74658203125,
0.490234375,
0.76025390625,
0.154541015625,
0.76025390625
],
[
0.5146484375,
0.751953125,
0.587890625,
0.751953125,
0.587890625,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.154541015625,
0.76318359375,
0.489501953125,
0.76318359375,
0.489501953125,
0.77685546875,
0.154541015625,
0.77685546875
],
[
0.151123046875,
0.77783203125,
0.181396484375,
0.77783203125,
0.181396484375,
0.79443359375,
0.151123046875,
0.79443359375
],
[
0.509765625,
0.78076171875,
0.67529296875,
0.78076171875,
0.67529296875,
0.7958984375,
0.509765625,
0.7958984375
],
[
0.1158447265625,
0.80712890625,
0.145263671875,
0.80712890625,
0.145263671875,
0.8193359375,
0.1158447265625,
0.8193359375
],
[
0.1485595703125,
0.8046875,
0.273681640625,
0.8056640625,
0.273681640625,
0.82080078125,
0.1485595703125,
0.8193359375
],
[
0.51123046875,
0.80810546875,
0.8818359375,
0.80712890625,
0.8818359375,
0.82177734375,
0.51123046875,
0.822265625
],
[
0.5146484375,
0.82666015625,
0.87890625,
0.82666015625,
0.87890625,
0.83740234375,
0.5146484375,
0.83740234375
],
[
0.1435546875,
0.81005859375,
0.1485595703125,
0.81005859375,
0.1485595703125,
0.8154296875,
0.1435546875,
0.8154296875
],
[
0.118408203125,
0.83203125,
0.48779296875,
0.83203125,
0.48779296875,
0.845703125,
0.118408203125,
0.845703125
],
[
0.509765625,
0.84130859375,
0.87890625,
0.84130859375,
0.87890625,
0.85498046875,
0.509765625,
0.85498046875
],
[
0.1175537109375,
0.8486328125,
0.489501953125,
0.8486328125,
0.489501953125,
0.8623046875,
0.1175537109375,
0.8623046875
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.1158447265625,
0.8642578125,
0.31494140625,
0.8642578125,
0.31494140625,
0.87744140625,
0.1158447265625,
0.87744140625
],
[
0.5146484375,
0.87451171875,
0.880859375,
0.87451171875,
0.880859375,
0.88525390625,
0.5146484375,
0.88525390625
],
[
0.1234130859375,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.1234130859375,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.12005615234375,
0.90625,
0.48779296875,
0.90625,
0.48779296875,
0.91943359375,
0.12005615234375,
0.91943359375
],
[
0.51123046875,
0.90478515625,
0.85205078125,
0.90478515625,
0.85205078125,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"要的新方法：DynamicLatentExtractionforAb-",
"不同的稀疏注意机制来降低内存成本。Long",
"stractiveSummarization（DYLE）。DYLE 联合训",
"former（Beltagyetal.，2020）使用块和全局注意",
"练提取器和生成器，并使提取的文本片段保持",
"力模式的扩展滑动窗口。BigBird（Zaheeretal.",
"潜伏。对于输出令牌，DYLE分别计算以每个",
"2020）采用滑动窗口和随机块。重整器（Kitaer",
"输入片段为条件的概率，并且通过在生成器以",
"etal.,2020)使用局部敏感哈希。除了优化编码",
"先前生成的令牌为条件分配的学习动态权重下",
"器的自我关注之外，Huangetal.（2021）还提出",
"对所有输入片段进行边缘化来计算其生成概率",
"了头向位置步幅，以降低编码器-解码器关注",
"我们优化了具有两个代理损耗的提取器。首",
"的成本。然而，稀蔬的注意力减少了预训练的",
"先，我们根据参考摘要计算提取oracle，并在最",
"好处，并牺牲了部分接受领域",
"佳ROUGE分数上进行贪婪搜索。这些oracle",
"片段被用作提取器学习信号的目标。此外，我",
"Extract-then-generatemethod此方法从输",
"们提出了致性损失，以鼓励提取器将其在代",
"入中提取重要的文本片段，然后生成总体摘要。",
"码段上的预测权重近似为生成器预测的平均动",
"这些方法中的大多数都是单独训练的(Zhang",
"态权重",
"et al.,2019; Lebanoff et al.,2019; Xu and Durrett.",
"我们在三个长输入摘要数据集上进行了",
"2019; Bajaj et al., 2021; Zhang et al., 2021b),当",
"实验：GovReport (Huang et al.,2021)和 arXiv",
"我们将提取的代码段传递给生成器时，它们会",
"(Cohanetal.，2018）用于长文档摘要，QMSum",
"遭受信息丢失。一些方法试图通过桥接两个阶",
"(Zhongetal.,2021b）用于长对话摘要。我们的",
"段来减少损失。ChenandBansal（2018）采用具",
"方法DYLE在GovReport和QMSum上的表现",
"有句子级策略梯度的强化学习（RL）。Baeetal.",
"大大优于现有方法，同时在arXiv上取得了很",
"(2019)提出了摘要级策略梯度。使用RL会在",
"好的效果。值得注意的是，DYLE的收益为4。",
"长输入文本上遇到各种缺点，这将在3.3节中",
"2/6。1/4。在 GovReport上，ROUGE-1/2/L 比",
"详细说明。DYLE是不同的，因为我们共同训",
"练一个提取然后生成的模型，用于使用潜在变",
"以前的最佳方法高出0分。这些实验证明了",
"DYLE对多个长输人汇总任务的通用性",
"量进行总结",
"我们总结我们的贡献如下：",
"Divide-and-conquer approach长输入汇总",
". We introduce DYLE, a dynamic latent extrac-",
"中的一种常见方法是分而治之（Gidiotis and",
"tion approach for abstractive long-input sum-",
"Tsoumakas, 2020; Grail et al., 2021; Zhang et al.",
"marization. DYLE better captures informa-",
"2021a）。它将一个长的输入分成多个部分，这",
"tion in the long input and reduces computa-",
"些部分分别汇总并组合在一起以生成最终的摘",
"tional cost;",
"要。但是，这些模型不能捕获各部分之间的上",
"下文依赖关系，并假设输入具有特定的结构",
": We propose multiple auxiliary optimizations",
"for the effective training of DYLE: 1) extrac-",
"Hierarchicalmodels已经提出了各种分层模",
"tive oracle as a learning signal for the extrac-",
"型来处理较长的输入。Cohanetal.(2018)使用",
"tor; 2) consistency loss that bridges extrac-",
"分层编码器和话语感知解码器对文档话语结",
"tion and generation; 3) hybrid training meth-",
"构进行建模。HAT-Bart（Rohdeetal..2021）提",
"ods that make the extraction more robust;",
"出了一种新的基于分层注意力转换器的体系结",
"构，该体系结构试图捕获句子和段落级信息。",
". Experimental results show that DYLE largely",
"HMNet(Zhuetal.,2020)构建了一个层次结构",
"outperforms the state-of-the-art on two long",
"其中包括话语级信息和演讲者角色。然而，这",
"input summarization datasets. We also con-",
"些模型主要关注模型性能，而不是减少内存和",
"ducted a detailed analysis that shows dy-",
"计算成本",
"namic weights improve model interpretabil-",
"ity.",
"3OurApproach",
"2",
"Related Work",
"我们的方法概述如图1所示。在3.1节中，",
"我们制定了我们的任务和提取器-生成器框架。",
"",
"For R3 : We introduce in detail the four main cat-",
"在3.2节中，我们介绍了长输人提取器的参数",
"egories of methods in recent work to address long-",
"化。在3.3节中，我们介绍了生成器公式和新",
"input summarization tasks.",
"颖的一致性损失。提取器模块同时优化了一一致",
"Sparseattentionmechanism完全注意力机",
"性损失和oracle损失，我们在3.4节中对此进",
"制具有二次记忆成本。先前的研究工作提出了",
"行了详细说明。3.5节总结了总体培训目标"
] |
{
"arxiv_id": "2110.08168",
"language": "zh",
"page_idx": 2
}
|
||
2110.08168-zh-3
|
2110.08168-zh-3.jpg
|
<FigureHere>
Figure 2: Long-input extractor. We divide the document into chunks, each containing consecutive snippets. A shared RoBERTa encodes each chunk independently.
3.1 $\textbf{Extractor- Generator Framework}$
在长输入摘要任务中，输入由$L$ 文本片段、 $X=(x_1,\ldots,x_L)$和可选查询$q$ (如果查询与摘要配对)组成。在长输入摘要中，文本片段$L$的数量可能很大。输出是长度为$T$的摘要$y$ 。对于对话摘要任务，将每个发言者的对话话语用作片段。对于文档，我们将输入标记为句子，并将每个句子用作片段。目标是学习一个模型，该模型在给定输入代码段$X$和先前生成的令牌$y_<t$的情况下生成一系列摘要令牌$y$
$$P_\theta(y|q,X)=\prod_{t=1}^TP_\theta(y_t|q,X,y_{<t})$$
提取器将查询和源文本作为输入，并为每个文本片段$x_i$输出分数$s_i=E_\eta(q,x_i)$ 。这里$\eta$ 是提取器参数。我们根据$X$的分数从文档$X$ 中提取$K$片段$X_K:$
$$X_K=\text{top-}K(E_\eta(q,x_i),x_i\in X)$$
(1)
从$X$中检索$X_K$后，提取器-生成器框架通过将$X$替换为$X_K$来对输出概率进行建模，i。$e$。,
(2)
$$\begin{aligned}P_{\theta}(y|q,X)&=P_{\theta}(y|q,X_{K})\\&=\prod_{t=1}^TP_\theta(y_t|q,X_K,y_{<t})\end{aligned}$$
请注意，方程式$\mathbb{E}($1)中的 top- $K$ 运算是不可微的，并且我们不通过 top- $K$传播梯度；相反，我们提出了在E3.3 节和E3.4 节中优化提取器的方法。
<FigureHere>
Figure 3: At each decoding time step, our generator predicts the dynamic weight and the generation probability for each extracted snippet.
# 3.2 Extractor for Long Inputs
一个有趣的研究问题是如何设计用于长输入的提取器。受 GPU 内存的限制，将所有代码片段连接起来并用大型预训练语言模型进行编码是不切实际的。如图$\overline{\mathrm{E}}$2 所示，我们将连续的代码段分组为块。我们将查询$q$与每个块连接， 并在其所属的块中独立计算每个代码段的编码向量。我们使用 MLP 将编码向量投影到标量分数$s_i=E_\eta(q,x_i)$
## 3.3 Generator with Dynamic Weights
Challenges 提取然后生成模型在长输人汇总中面临两个挑战。第一个挑战是提取操作 (方程(1)中的 top- $K$ )是不可微分的。一种方法是采用基于 RL 的优化 (Chen and Bansal, 2018; Bae et al.,2019),它有两个缺点。首先，针对大型动作空间的强化学习(i。e。,当$L$非常大时，从$L$片段中提取$K$ )具有高方差。其次，当前的方法大多使用句子级 ROUGE(Chen and Bansal, 2018) 或摘要级 ROUGE (Bae et al., 2019)作为训练奖励。使用句子级 ROUGE 可能会选择内容重叠的句子 (Narayan et al., 2018)
从而导致冗余的最终摘要。使用摘要级别的ROUGE 会导致训练信号的稀疏性，而更长的输人会使这种方法更难训练。第二个挑战是可解释性：人们可能想知道生成器是否在每个解码时间步长利用提取的信息
为了解决这些挑战，我们提出了一个生成器， 该生成器在每个时间步长为每个提取的片段动态分配权重。与提取器分数不同，提取器分数与解码时间步长无关，生成器在不同的时间步长分配不同的动态分数。动态权重使解码过程可解释，并通过降低不相关片段的权重来帮助对提取进行去噪。它还使用一致性损失为提取器提供训练信号。
Generator formulation 生成器的概述如图E 3 所示。对于每个提取的代码段$x$ ,生成器预
测此代码段上的生成概率$P_\theta(y_t|q,x,y_{<t})$和此代码段的动态权重$P_\theta(x|q,X_K,y_{<t})$ 。每个提取的代码段的独立编码节省了内存，因为代
|
[
[
0.66162109375,
0.0872802734375,
0.7197265625,
0.0872802734375,
0.7197265625,
0.097412109375,
0.66162109375,
0.097412109375
],
[
0.767578125,
0.08612060546875,
0.86376953125,
0.08489990234375,
0.86376953125,
0.0968017578125,
0.767578125,
0.097412109375
],
[
0.60546875,
0.0955810546875,
0.63037109375,
0.0955810546875,
0.63037109375,
0.10809326171875,
0.60546875,
0.10809326171875
],
[
0.16796875,
0.10394287109375,
0.279541015625,
0.105712890625,
0.279541015625,
0.1175537109375,
0.16796875,
0.11517333984375
],
[
0.654296875,
0.10809326171875,
0.73046875,
0.1092529296875,
0.73046875,
0.12054443359375,
0.654296875,
0.119384765625
],
[
0.767578125,
0.1092529296875,
0.86376953125,
0.1092529296875,
0.86376953125,
0.12054443359375,
0.767578125,
0.12054443359375
],
[
0.5576171875,
0.1181640625,
0.611328125,
0.1181640625,
0.611328125,
0.133056640625,
0.5576171875,
0.133056640625
],
[
0.300537109375,
0.1300048828125,
0.344970703125,
0.1300048828125,
0.344970703125,
0.144287109375,
0.300537109375,
0.144287109375
],
[
0.447509765625,
0.1395263671875,
0.468505859375,
0.1395263671875,
0.468505859375,
0.1519775390625,
0.447509765625,
0.1519775390625
],
[
0.59765625,
0.1383056640625,
0.6337890625,
0.1383056640625,
0.6337890625,
0.1455078125,
0.59765625,
0.1455078125
],
[
0.52490234375,
0.1484375,
0.5859375,
0.1484375,
0.5859375,
0.1591796875,
0.52490234375,
0.1591796875
],
[
0.60546875,
0.1484375,
0.63232421875,
0.1484375,
0.63232421875,
0.160888671875,
0.60546875,
0.160888671875
],
[
0.378662109375,
0.1519775390625,
0.391357421875,
0.1519775390625,
0.391357421875,
0.16748046875,
0.378662109375,
0.16748046875
],
[
0.189697265625,
0.152587890625,
0.202392578125,
0.152587890625,
0.202392578125,
0.166259765625,
0.189697265625,
0.166259765625
],
[
0.181396484375,
0.1436767578125,
0.198974609375,
0.135986328125,
0.208251953125,
0.149658203125,
0.19140625,
0.1561279296875
],
[
0.455078125,
0.15673828125,
0.4609375,
0.15673828125,
0.4609375,
0.162109375,
0.455078125,
0.162109375
],
[
0.759765625,
0.1644287109375,
0.87158203125,
0.166259765625,
0.87158203125,
0.1776123046875,
0.759765625,
0.17578125
],
[
0.1796875,
0.17333984375,
0.2376708984375,
0.17333984375,
0.2376708984375,
0.185302734375,
0.1796875,
0.185302734375
],
[
0.366943359375,
0.1728515625,
0.429931640625,
0.1728515625,
0.429931640625,
0.18701171875,
0.366943359375,
0.18701171875
],
[
0.20654296875,
0.19482421875,
0.219970703125,
0.19482421875,
0.219970703125,
0.198974609375,
0.20654296875,
0.198974609375
],
[
0.393798828125,
0.19482421875,
0.41064453125,
0.19482421875,
0.41064453125,
0.198974609375,
0.393798828125,
0.198974609375
],
[
0.51318359375,
0.193603515625,
0.8818359375,
0.193603515625,
0.8818359375,
0.207275390625,
0.51318359375,
0.207275390625
],
[
0.1275634765625,
0.2037353515625,
0.187255859375,
0.2037353515625,
0.187255859375,
0.215576171875,
0.1275634765625,
0.215576171875
],
[
0.318115234375,
0.2037353515625,
0.353515625,
0.2037353515625,
0.353515625,
0.215576171875,
0.318115234375,
0.215576171875
],
[
0.435791015625,
0.204833984375,
0.453369140625,
0.204833984375,
0.453369140625,
0.21435546875,
0.435791015625,
0.21435546875
],
[
0.5146484375,
0.207275390625,
0.8818359375,
0.207275390625,
0.8818359375,
0.2203369140625,
0.5146484375,
0.2203369140625
],
[
0.177978515625,
0.221435546875,
0.1822509765625,
0.221435546875,
0.1822509765625,
0.22802734375,
0.177978515625,
0.22802734375
],
[
0.51123046875,
0.2203369140625,
0.73046875,
0.221435546875,
0.73046875,
0.236328125,
0.51123046875,
0.235107421875
],
[
0.1956787109375,
0.25830078125,
0.2568359375,
0.25830078125,
0.2568359375,
0.268310546875,
0.1956787109375,
0.268310546875
],
[
0.509765625,
0.260009765625,
0.765625,
0.26123046875,
0.765625,
0.275634765625,
0.509765625,
0.27490234375
],
[
0.1175537109375,
0.283935546875,
0.48779296875,
0.283935546875,
0.48779296875,
0.297607421875,
0.1175537109375,
0.297607421875
],
[
0.5283203125,
0.283935546875,
0.880859375,
0.283935546875,
0.880859375,
0.297607421875,
0.5283203125,
0.297607421875
],
[
0.1175537109375,
0.298583984375,
0.489501953125,
0.298583984375,
0.489501953125,
0.312255859375,
0.1175537109375,
0.312255859375
],
[
0.51318359375,
0.298583984375,
0.87890625,
0.298583984375,
0.87890625,
0.3134765625,
0.51318359375,
0.3134765625
],
[
0.118408203125,
0.314208984375,
0.48779296875,
0.314208984375,
0.48779296875,
0.32470703125,
0.118408203125,
0.32470703125
],
[
0.51123046875,
0.314208984375,
0.87890625,
0.314208984375,
0.87890625,
0.327880859375,
0.51123046875,
0.327880859375
],
[
0.1175537109375,
0.325927734375,
0.162841796875,
0.327880859375,
0.162109375,
0.34033203125,
0.1158447265625,
0.33837890625
],
[
0.5146484375,
0.33203125,
0.87890625,
0.33203125,
0.87890625,
0.345703125,
0.5146484375,
0.345703125
],
[
0.5146484375,
0.347412109375,
0.87744140625,
0.347412109375,
0.87744140625,
0.361083984375,
0.5146484375,
0.361083984375
],
[
0.1158447265625,
0.3681640625,
0.4248046875,
0.3681640625,
0.4248046875,
0.3818359375,
0.1158447265625,
0.3818359375
],
[
0.5146484375,
0.364013671875,
0.880859375,
0.364013671875,
0.880859375,
0.377685546875,
0.5146484375,
0.377685546875
],
[
0.51318359375,
0.37890625,
0.880859375,
0.38134765625,
0.880859375,
0.394287109375,
0.51318359375,
0.392578125
],
[
0.118408203125,
0.39013671875,
0.486083984375,
0.39013671875,
0.486083984375,
0.40380859375,
0.118408203125,
0.40380859375
],
[
0.51318359375,
0.396728515625,
0.6650390625,
0.396728515625,
0.6650390625,
0.410888671875,
0.51318359375,
0.410888671875
],
[
0.1158447265625,
0.40625,
0.484375,
0.40673828125,
0.484375,
0.421630859375,
0.1158447265625,
0.42041015625
],
[
0.118408203125,
0.42333984375,
0.486083984375,
0.42333984375,
0.486083984375,
0.43701171875,
0.118408203125,
0.43701171875
],
[
0.51123046875,
0.42333984375,
0.5556640625,
0.42333984375,
0.5556640625,
0.435302734375,
0.51123046875,
0.435302734375
],
[
0.5556640625,
0.42333984375,
0.82373046875,
0.42333984375,
0.82373046875,
0.43701171875,
0.5556640625,
0.43701171875
],
[
0.1158447265625,
0.438232421875,
0.48779296875,
0.439453125,
0.48779296875,
0.45361328125,
0.1158447265625,
0.453125
],
[
0.515625,
0.4423828125,
0.880859375,
0.44482421875,
0.880859375,
0.458984375,
0.515625,
0.457275390625
],
[
0.1318359375,
0.4560546875,
0.486083984375,
0.4560546875,
0.486083984375,
0.4697265625,
0.1318359375,
0.4697265625
],
[
0.5146484375,
0.460205078125,
0.880859375,
0.460205078125,
0.880859375,
0.473876953125,
0.5146484375,
0.473876953125
],
[
0.1175537109375,
0.471435546875,
0.484375,
0.471435546875,
0.484375,
0.485107421875,
0.1175537109375,
0.485107421875
],
[
0.1275634765625,
0.46142578125,
0.1368408203125,
0.46142578125,
0.1368408203125,
0.465576171875,
0.1275634765625,
0.465576171875
],
[
0.51318359375,
0.476806640625,
0.880859375,
0.476806640625,
0.880859375,
0.490478515625,
0.51318359375,
0.490478515625
],
[
0.118408203125,
0.48681640625,
0.484375,
0.48681640625,
0.484375,
0.50048828125,
0.118408203125,
0.50048828125
],
[
0.5146484375,
0.4921875,
0.880859375,
0.4921875,
0.880859375,
0.505859375,
0.5146484375,
0.505859375
],
[
0.118408203125,
0.50439453125,
0.484375,
0.50439453125,
0.484375,
0.51708984375,
0.118408203125,
0.51708984375
],
[
0.51318359375,
0.5087890625,
0.880859375,
0.5087890625,
0.880859375,
0.5224609375,
0.51318359375,
0.5224609375
],
[
0.118408203125,
0.51953125,
0.48779296875,
0.51953125,
0.48779296875,
0.53369140625,
0.118408203125,
0.53369140625
],
[
0.5146484375,
0.525390625,
0.880859375,
0.525390625,
0.880859375,
0.5390625,
0.5146484375,
0.5390625
],
[
0.515625,
0.54150390625,
0.880859375,
0.54150390625,
0.880859375,
0.5546875,
0.515625,
0.5546875
],
[
0.28369140625,
0.54736328125,
0.30322265625,
0.54736328125,
0.30322265625,
0.56103515625,
0.28369140625,
0.56103515625
],
[
0.51123046875,
0.5556640625,
0.880859375,
0.5556640625,
0.880859375,
0.5693359375,
0.51123046875,
0.5693359375
],
[
0.173828125,
0.5625,
0.42822265625,
0.5625,
0.42822265625,
0.580078125,
0.173828125,
0.580078125
],
[
0.51318359375,
0.572265625,
0.880859375,
0.572265625,
0.880859375,
0.5859375,
0.51318359375,
0.5859375
],
[
0.279541015625,
0.580078125,
0.306396484375,
0.580078125,
0.306396484375,
0.5927734375,
0.279541015625,
0.5927734375
],
[
0.51123046875,
0.58837890625,
0.87890625,
0.58837890625,
0.87890625,
0.6015625,
0.51123046875,
0.6015625
],
[
0.1368408203125,
0.60400390625,
0.484375,
0.60400390625,
0.484375,
0.61767578125,
0.1368408203125,
0.61767578125
],
[
0.51123046875,
0.60498046875,
0.880859375,
0.60400390625,
0.880859375,
0.6181640625,
0.51123046875,
0.619140625
],
[
0.118408203125,
0.6181640625,
0.48779296875,
0.619140625,
0.48779296875,
0.63427734375,
0.118408203125,
0.6328125
],
[
0.51904296875,
0.626953125,
0.52880859375,
0.626953125,
0.52880859375,
0.63134765625,
0.51904296875,
0.63134765625
],
[
0.5283203125,
0.62158203125,
0.880859375,
0.62158203125,
0.880859375,
0.63525390625,
0.5283203125,
0.63525390625
],
[
0.118408203125,
0.6357421875,
0.484375,
0.6357421875,
0.484375,
0.6494140625,
0.118408203125,
0.6494140625
],
[
0.51318359375,
0.63720703125,
0.87890625,
0.63720703125,
0.87890625,
0.65087890625,
0.51318359375,
0.65087890625
],
[
0.118408203125,
0.65087890625,
0.28369140625,
0.65185546875,
0.28369140625,
0.66650390625,
0.118408203125,
0.6650390625
],
[
0.5146484375,
0.65283203125,
0.87890625,
0.65283203125,
0.87890625,
0.66650390625,
0.5146484375,
0.66650390625
],
[
0.5146484375,
0.66943359375,
0.8818359375,
0.66943359375,
0.8818359375,
0.68408203125,
0.5146484375,
0.68408203125
],
[
0.177978515625,
0.68408203125,
0.42822265625,
0.68408203125,
0.42822265625,
0.69873046875,
0.177978515625,
0.69873046875
],
[
0.462646484375,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69677734375,
0.462646484375,
0.69677734375
],
[
0.51123046875,
0.68505859375,
0.7314453125,
0.68505859375,
0.7314453125,
0.69873046875,
0.51123046875,
0.69873046875
],
[
0.5322265625,
0.70068359375,
0.87744140625,
0.70166015625,
0.87744140625,
0.71630859375,
0.5322265625,
0.71533203125
],
[
0.118408203125,
0.71533203125,
0.486083984375,
0.71533203125,
0.486083984375,
0.728515625,
0.118408203125,
0.728515625
],
[
0.5146484375,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.732421875,
0.5146484375,
0.732421875
],
[
0.118408203125,
0.732421875,
0.489501953125,
0.732421875,
0.489501953125,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.51123046875,
0.73388671875,
0.880859375,
0.73291015625,
0.880859375,
0.74755859375,
0.51123046875,
0.7490234375
],
[
0.1175537109375,
0.75048828125,
0.1435546875,
0.75048828125,
0.1435546875,
0.7626953125,
0.1175537109375,
0.7626953125
],
[
0.515625,
0.75048828125,
0.87890625,
0.75048828125,
0.87890625,
0.76416015625,
0.515625,
0.76416015625
],
[
0.515625,
0.7666015625,
0.87890625,
0.7666015625,
0.87890625,
0.77978515625,
0.515625,
0.77978515625
],
[
0.1695556640625,
0.779296875,
0.36865234375,
0.779296875,
0.36865234375,
0.79345703125,
0.1695556640625,
0.79345703125
],
[
0.51318359375,
0.7822265625,
0.87744140625,
0.7822265625,
0.87744140625,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.274658203125,
0.79638671875,
0.295654296875,
0.79638671875,
0.295654296875,
0.81005859375,
0.274658203125,
0.81005859375
],
[
0.46337890625,
0.80419921875,
0.486083984375,
0.80419921875,
0.486083984375,
0.81396484375,
0.46337890625,
0.81396484375
],
[
0.51123046875,
0.798828125,
0.87890625,
0.798828125,
0.87890625,
0.8125,
0.51123046875,
0.8125
],
[
0.25439453125,
0.80859375,
0.438232421875,
0.80859375,
0.438232421875,
0.830078125,
0.25439453125,
0.830078125
],
[
0.51123046875,
0.8134765625,
0.6513671875,
0.81396484375,
0.6513671875,
0.8291015625,
0.51123046875,
0.82763671875
],
[
0.27197265625,
0.8291015625,
0.300537109375,
0.8291015625,
0.300537109375,
0.84033203125,
0.27197265625,
0.84033203125
],
[
0.51708984375,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85400390625,
0.51708984375,
0.85400390625
],
[
0.118408203125,
0.85546875,
0.48779296875,
0.85498046875,
0.48779296875,
0.869140625,
0.118408203125,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.51123046875,
0.87158203125,
0.880859375,
0.87255859375,
0.880859375,
0.88720703125,
0.51123046875,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.484375,
0.88818359375,
0.484375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.12005615234375,
0.90625,
0.21826171875,
0.90625,
0.21826171875,
0.91748046875,
0.12005615234375,
0.91748046875
],
[
0.51318359375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"LMhead",
"generationprob",
"ht",
"Extracted snippets",
"Weight head",
"dynamic weight",
"Seq2Seq",
"Top-K",
"SL",
"昌",
"querya",
"y<t",
"+",
"+",
"",
"+",
"Extracted snippets",
"RoBERTa",
"RoBERTa",
"++",
"++",
"Figure 3: At each decoding time step, our generator",
"query ai",
"query",
"CL",
"predicts the dynamic weight and the generation proba-",
"",
"bility for each extracted snippet.",
"Document",
"3.2Extractor for Long Inputs",
"Figure 2: Long-input extractor. We divide the doc-",
"个有趣的研究问题是如何设计用于长输入的",
"ument into chunks, each containing consecutive snip-",
"提取器。受GPU内存的限制，将所有代码片",
"pets.A sharedRoBERTa encodes eachchunk indepen-",
"段连接起来并用大型预训练语言模型进行编码",
"dently",
"是不切实际的。如图E2所示，我们将连续的",
"代码段分组为块。我们将查询q与每个块连接",
"3.1 Extractor-Generator Framework",
"并在其所属的块中独立计算每个代码段的编码",
"向量。我们使用MLP将编码向量投影到标量",
"在长输入摘要任务中，输入由L文本片段、",
"分数si=En(q,ai)",
"X=（1....,L）和可选查询g（如果查询与",
"摘要配对）组成。在长输入摘要中，文本片段",
"3.3",
"Generator with Dynamic Weights",
"L的数量可能很大。输出是长度为T的摘要y",
"Challenges提取然后生成模型在长输人汇总",
"对于对话摘要任务，将每个发言者的对话话",
"中面临两个挑战。第一个挑战是提取操作（方",
"语用作片段。对于文档，我们将输入标记为句",
"",
"程（1）中的top-K）是不可微分的。一种方法",
"子，并将每个句子用作片段。目标是学习一个",
"是采用基于RL的优化（ChenandBansal,2018;",
"模型，该模型在给定输人代码段X和先前生",
"Baeetal.，2019），它有两个缺点。首先，针对",
"成的令牌y<t的情况下生成一系列摘要令牌y",
"大型动作空间的强化学习i。e。，当L非常",
"大时，从L片段中提取K）具有高方差。其",
"T",
"次，当前的方法大多使用句子级ROUGE（Chen",
"Pe(ylq, X) = II Pe(ytlq, X, y<t)",
"andBansal,2018）或摘要级ROUGE(Baeetal.,",
"t=1",
"2019）作为训练奖励。使用句子级ROUGE可",
"提取器将查询和源文本作为输入，并为每个",
"能会选择内容重叠的句子(Narayanetal.,2018)",
"文本片段c输出分数si=En(g,ci）。这里n",
"",
"从而导致亢余的最终摘要。使用摘要级别的",
"是提取器参数。我们根据X的分数从文档X",
"ROUGE会导致训练信号的稀疏性，而更长的",
"中提取K片段XK：",
"输入会使这种方法更难训练。第二个挑战是可",
"解释性：人们可能想知道生成器是否在每个解",
"Xk = top-K(En(q, ai), a E X)",
"(1)",
"码时间步长利用提取的信息",
"为了解决这些挑战，我们提出了一个生成器",
"从X中检索XK后，提取器-生成器框架通过",
"该生成器在每个时间步长为每个提取的片段动",
"将X替换为Xk来对输出概率进行建模，i。",
"态分配权重。与提取器分数不同，提取器分数",
"e.,",
"与解码时间步长无关，生成器在不同的时间步",
"长分配不同的动态分数。动态权重使解码过程",
"Pe(ylq, X) = Pe(ylq, Xk)",
"可解释，并通过降低不相关片段的权重来帮助",
"T",
"(2)",
"对提取进行去噪。它还使用一致性损失为提取",
"= II Pe(ytlq, Xk, y<t)",
"器提供训练信号。",
"t=1",
"Generatorformulation生成器的概述如图E",
"请注意，方程式E（1）中的top-K运算是不",
"3所示。对于每个提取的代码段α，生成器预",
"可微的，并且我们不通过top-K传播梯度；相",
"测此代码段上的生成概率Pe(ytlq,a,y<t）和此",
"反，我们提出了在E3.3节和E3.4节中优化提",
"代码段的动态权重Pe(αq,Xk，y<t）。每个提",
"取器的方法。",
"取的代码段的独立编码节省了内存，因为代"
] |
{
"arxiv_id": "2110.08168",
"language": "zh",
"page_idx": 3
}
|
||
2110.08168-zh-4
|
2110.08168-zh-4.jpg
|
码段不需要相互关注。不失一般性，我们假设$P_\theta(\cdot|q,x,y_{<t})$是通过首先将输入$(q,x,y_{<t})$ 映射到上下文化表示向量$h_t^x$来计算的。对于 Transformers (Vaswani et al.,2017)和带有关注模型 (Bahdanau et al., 2015) 的编码器-解码器，$h_t^x$通常是模型在最终语言模型头之前的输出。通过将$h_t^x$送入语言模型头计算生成概率$P_\theta(y_t|q,x,y_{<t})$ 。对于动态权重$P_\theta ( x| q, X_K, y_{< t})$ ,我们采用单独的 MLP 将每个$h_t^x$映射到标量 logit $l^x$,并且$P_\theta(\cdot|q,X,y_{<t})$ 被定义为 softmax$(\{l^x\}_{x\in X})$。我们通过对所有提取的片段进行边缘化来计算生成概率：
$$\begin{aligned}P_{\theta}(y|q,X_{K})&=\prod_{t=1}^{T}\sum_{x\in X_{K}}\\P_{\theta}(y_{t}|q,x,y_{<t})P_{\theta}(x|q,X_{K},y_{<t})\end{aligned}$$
(3)
每个解码时间步长$t$的动态权重$P_\theta(x|q,X_K,y_{<t})$允许我们解释生成器如何利用提取的片段。例如，对特定代码段的权重越大，表示该代码段对当前解码时间步长的重要性越大。发电损失定义为黄金汇总的$NLL:$
(4)
$$\mathcal{L}_{gen}^{\theta}=-\log P_{\theta}(y|q,X_{K})$$
其中$P_\theta(y|q,X_K)$在公式$\mathbb{E}(2)$中定义。在这里，我们不将$\mathcal{L}_gen^\theta$的梯度传播到提取器参数， 因为 top- $K$是不可微分的。相反，在 3.3 节和 3.4 节中描述了优化提取器的方法
Consistency loss 我们还利用动态权重为提取器提供训练信号。由于代码段的动态权重可以解释为代码段在特定时间步长的重要性，因此我们对所有解码步骤的动态权重进行平均， 并将平均权重视为代码段的总体重要性。基于这种直觉，我们提出了我们所说的一致性损失，它测量平均动态权重分布和提取器分布之间的距离。我们希望这两个分布在$X$的任意子集上接近。为简单起见，我们将$X_K$作为子集，并将一致性损失定义为
$$\begin{aligned}\mathcal{L}_{consist}^{\eta}&=\mathrm{KL}\Big[\frac{1}{T}\sum_{t=1}^{T}P_{\theta}(\cdot|q,X_{K},y_{<t})\:||\\&\mathrm{softmax}(E_{\eta}(q,x_{i}),x_{i}\in X_{K})\Big]\end{aligned}$$
(5)
请注意，一致性损失是使用提取器的参数$\eta$进行上标的，这意味着我们不计算生成器参数$\theta$ 的梯度。由于我们希望$X$的任意子集上的分布距离较小，因此我们不会通过 top- $K$运算符传播梯度
## 3.4 Leveraging Extractive Oracles
对于长输入汇总，在训练期间使用的提取片段$X_K$对于稳定优化很重要。我们建议在训练过程中利用提取神谕，而不是使用公式$\mathbb{E}(1)$ 中定义的$X_K$ 。在测试期间不使用提取神谕
Greedy search for extractive oracles Extractiveoracles 表示一组选定的文本片段，在给定黄金摘要的情况下，其串联可最大化评估指标。我们使用贪婪搜索实现提取 oracle。具体来说，我们从一个空集开始，我们迭代地从输人中选择一个片段，使得该片段和已经选择的片段的连接最大化 ROUGE-1 , ROUGE-2 和ROUGE-L 分数的平均值给出黄金摘要。我们将提取神谕表示为$X_o$
Hybrid training 我们利用提取神谕来定义训练期间使用的$X_K$ 。如果 oracles 的数量等于或超过$K$,我们将$X_K$定义为第一个$K$ oracle 片段。如果 oracles 的数量小于$K$ ,我们将$X_{K}$ 定义为$X_o$和未出现在$X_o$中的提取器排名最高的代码段的联合。这种混合训练有两个好处。首先，与等式$\mathbb{E}(1)$中定义的$X_K$相比， 它为生成器提供了更高质量的输入。其次，它减少了对 oracle 的依赖，并提高了我们模型在训练集之外的泛化能力，因为贪婪搜索中省略的其他文本片段可能有助于生成
Oracle loss 提取神谕$X_o$用作我们模型提取部分的监督信号。oracle 损失$\mathcal{L}_oracle^\eta$是根据提取器所选集合中的所有块与提取 oracle 之间的交叉熵损失计算得出的。形式上，oracle 损失计算为
$$\mathcal{L}_{oracle}^{\eta}=-\dfrac{1}{|X_o|}\sum_{x\in X_o}\log\dfrac{e^{E_\eta(q,x)}}{\sum_{x_i\in X}e^{E_\eta(q,x_i)}}$$
## 3.5 Training Objective
我们方法的总体训练目标是
$$\mathcal{L}^{\theta,\eta}=\lambda_{g}\mathcal{L}_{gen}^{\theta}+\lambda_{o}\mathcal{L}_{oracle}^{\eta}+\lambda_{c}\mathcal{L}_{consist}^{\eta}$$
(7)
,其中$\lambda_g$、$\lambda_o$ 和$\lambda_c$ 是平衡损耗分量的超参数。计算上标参数的梯度。具体来说，提取器仅使用一致性损失和 oracle 损失进行优化，而生成器仅使用生成损失进行优化
# 4 Experiment Setups
### 4.1 Datasets
我们将以下长输入抽象摘要数据集视为评估基
准：2
$^{2}$QMSum and arXiv can be accessed through Summer-
Time (Ni et al., 2021a) .
|
[
[
0.118408203125,
0.0872802734375,
0.489501953125,
0.0872802734375,
0.489501953125,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.484375,
0.08905029296875,
0.51318359375,
0.08905029296875,
0.51318359375,
0.0968017578125,
0.484375,
0.0968017578125
],
[
0.50537109375,
0.08489990234375,
0.79833984375,
0.08612060546875,
0.79833984375,
0.1009521484375,
0.50537109375,
0.09979248046875
],
[
0.118408203125,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.1175537109375,
0.118408203125,
0.1175537109375
],
[
0.51318359375,
0.10809326171875,
0.87890625,
0.10809326171875,
0.87890625,
0.1217041015625,
0.51318359375,
0.1217041015625
],
[
0.118408203125,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.51318359375,
0.12353515625,
0.880859375,
0.12353515625,
0.880859375,
0.13720703125,
0.51318359375,
0.13720703125
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51318359375,
0.1395263671875,
0.880859375,
0.1395263671875,
0.880859375,
0.152587890625,
0.51318359375,
0.152587890625
],
[
0.118408203125,
0.149658203125,
0.486083984375,
0.149658203125,
0.486083984375,
0.163330078125,
0.118408203125,
0.163330078125
],
[
0.51123046875,
0.1561279296875,
0.84619140625,
0.1561279296875,
0.84619140625,
0.169189453125,
0.51123046875,
0.169189453125
],
[
0.118408203125,
0.16748046875,
0.483642578125,
0.16748046875,
0.483642578125,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.1175537109375,
0.1817626953125,
0.483642578125,
0.1817626953125,
0.483642578125,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.51904296875,
0.179931640625,
0.8818359375,
0.179931640625,
0.8818359375,
0.193603515625,
0.51904296875,
0.193603515625
],
[
0.1175537109375,
0.19775390625,
0.48779296875,
0.198974609375,
0.48779296875,
0.213134765625,
0.1175537109375,
0.2120361328125
],
[
0.509765625,
0.19482421875,
0.87890625,
0.19482421875,
0.87890625,
0.2078857421875,
0.509765625,
0.2078857421875
],
[
0.118408203125,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2298583984375,
0.118408203125,
0.2298583984375
],
[
0.5146484375,
0.2120361328125,
0.87890625,
0.2120361328125,
0.87890625,
0.2257080078125,
0.5146484375,
0.2257080078125
],
[
0.1175537109375,
0.2298583984375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.1175537109375,
0.243408203125
],
[
0.51318359375,
0.2286376953125,
0.880859375,
0.2286376953125,
0.880859375,
0.2423095703125,
0.51318359375,
0.2423095703125
],
[
0.1175537109375,
0.2464599609375,
0.48779296875,
0.2476806640625,
0.48779296875,
0.262451171875,
0.1175537109375,
0.26123046875
],
[
0.51318359375,
0.24462890625,
0.880859375,
0.24462890625,
0.880859375,
0.25830078125,
0.51318359375,
0.25830078125
],
[
0.118408203125,
0.26416015625,
0.438232421875,
0.26416015625,
0.438232421875,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.515625,
0.262451171875,
0.880859375,
0.262451171875,
0.880859375,
0.2724609375,
0.515625,
0.2724609375
],
[
0.51318359375,
0.275634765625,
0.880859375,
0.276611328125,
0.880859375,
0.290283203125,
0.51318359375,
0.289306640625
],
[
0.267822265625,
0.295654296875,
0.28369140625,
0.295654296875,
0.28369140625,
0.303955078125,
0.267822265625,
0.303955078125
],
[
0.51123046875,
0.29150390625,
0.87890625,
0.29150390625,
0.87890625,
0.3046875,
0.51123046875,
0.3046875
],
[
0.139404296875,
0.3017578125,
0.32666015625,
0.3017578125,
0.32666015625,
0.32666015625,
0.139404296875,
0.32666015625
],
[
0.51318359375,
0.308837890625,
0.6884765625,
0.308837890625,
0.6884765625,
0.322509765625,
0.51318359375,
0.322509765625
],
[
0.462646484375,
0.318359375,
0.486083984375,
0.318359375,
0.486083984375,
0.33203125,
0.462646484375,
0.33203125
],
[
0.26123046875,
0.32470703125,
0.32666015625,
0.32470703125,
0.32666015625,
0.336181640625,
0.26123046875,
0.336181640625
],
[
0.51708984375,
0.333251953125,
0.880859375,
0.333251953125,
0.880859375,
0.3466796875,
0.51708984375,
0.3466796875
],
[
0.193115234375,
0.341552734375,
0.438232421875,
0.342529296875,
0.438232421875,
0.35693359375,
0.193115234375,
0.355712890625
],
[
0.51318359375,
0.349853515625,
0.87890625,
0.349853515625,
0.87890625,
0.363525390625,
0.51318359375,
0.363525390625
],
[
0.515625,
0.36767578125,
0.87890625,
0.36767578125,
0.87890625,
0.377685546875,
0.515625,
0.377685546875
],
[
0.12005615234375,
0.374755859375,
0.486083984375,
0.374755859375,
0.486083984375,
0.388427734375,
0.12005615234375,
0.388427734375
],
[
0.51318359375,
0.3818359375,
0.87890625,
0.3818359375,
0.87890625,
0.3955078125,
0.51318359375,
0.3955078125
],
[
0.118408203125,
0.39013671875,
0.484375,
0.39013671875,
0.484375,
0.40380859375,
0.118408203125,
0.40380859375
],
[
0.51318359375,
0.399658203125,
0.87890625,
0.399658203125,
0.87890625,
0.410400390625,
0.51318359375,
0.410400390625
],
[
0.118408203125,
0.40673828125,
0.486083984375,
0.40673828125,
0.486083984375,
0.42041015625,
0.118408203125,
0.42041015625
],
[
0.51318359375,
0.41455078125,
0.880859375,
0.41455078125,
0.880859375,
0.427490234375,
0.51318359375,
0.427490234375
],
[
0.1175537109375,
0.4228515625,
0.484375,
0.4228515625,
0.484375,
0.435791015625,
0.1175537109375,
0.435791015625
],
[
0.51318359375,
0.429931640625,
0.8818359375,
0.429931640625,
0.8818359375,
0.443603515625,
0.51318359375,
0.443603515625
],
[
0.118408203125,
0.439453125,
0.48779296875,
0.439453125,
0.48779296875,
0.453125,
0.118408203125,
0.453125
],
[
0.51318359375,
0.44482421875,
0.880859375,
0.4453125,
0.880859375,
0.458984375,
0.51318359375,
0.457763671875
],
[
0.118408203125,
0.4560546875,
0.162841796875,
0.4560546875,
0.162841796875,
0.46728515625,
0.118408203125,
0.46728515625
],
[
0.5146484375,
0.4619140625,
0.880859375,
0.4619140625,
0.880859375,
0.4755859375,
0.5146484375,
0.4755859375
],
[
0.2032470703125,
0.47265625,
0.4013671875,
0.47265625,
0.4013671875,
0.490478515625,
0.2032470703125,
0.490478515625
],
[
0.465087890625,
0.476806640625,
0.48779296875,
0.476806640625,
0.48779296875,
0.48681640625,
0.465087890625,
0.48681640625
],
[
0.51318359375,
0.47802734375,
0.87890625,
0.47802734375,
0.87890625,
0.4912109375,
0.51318359375,
0.4912109375
],
[
0.51318359375,
0.49462890625,
0.771484375,
0.49462890625,
0.771484375,
0.50830078125,
0.51318359375,
0.50830078125
],
[
0.118408203125,
0.501953125,
0.486083984375,
0.50048828125,
0.486083984375,
0.515625,
0.118408203125,
0.5166015625
],
[
0.118408203125,
0.5185546875,
0.483642578125,
0.5185546875,
0.483642578125,
0.5322265625,
0.118408203125,
0.5322265625
],
[
0.515625,
0.51708984375,
0.880859375,
0.5185546875,
0.880859375,
0.533203125,
0.515625,
0.5322265625
],
[
0.1175537109375,
0.533203125,
0.48779296875,
0.53369140625,
0.48779296875,
0.548828125,
0.1175537109375,
0.54736328125
],
[
0.51318359375,
0.53515625,
0.880859375,
0.53515625,
0.880859375,
0.5498046875,
0.51318359375,
0.5498046875
],
[
0.1175537109375,
0.5498046875,
0.411376953125,
0.5498046875,
0.411376953125,
0.5634765625,
0.1175537109375,
0.5634765625
],
[
0.51318359375,
0.5517578125,
0.87890625,
0.5517578125,
0.87890625,
0.5654296875,
0.51318359375,
0.5654296875
],
[
0.51318359375,
0.56787109375,
0.880859375,
0.56787109375,
0.880859375,
0.58056640625,
0.51318359375,
0.58056640625
],
[
0.124267578125,
0.580078125,
0.484375,
0.580078125,
0.484375,
0.59326171875,
0.124267578125,
0.59326171875
],
[
0.51318359375,
0.5830078125,
0.5693359375,
0.5830078125,
0.5693359375,
0.59716796875,
0.51318359375,
0.59716796875
],
[
0.1175537109375,
0.59423828125,
0.48193359375,
0.59423828125,
0.48193359375,
0.60791015625,
0.1175537109375,
0.60791015625
],
[
0.118408203125,
0.61083984375,
0.484375,
0.61083984375,
0.484375,
0.62451171875,
0.118408203125,
0.62451171875
],
[
0.52490234375,
0.6123046875,
0.69775390625,
0.63134765625,
0.69287109375,
0.65283203125,
0.5205078125,
0.63427734375
],
[
0.62646484375,
0.6123046875,
0.6396484375,
0.6123046875,
0.6396484375,
0.62255859375,
0.62646484375,
0.62255859375
],
[
0.75146484375,
0.60791015625,
0.8095703125,
0.60498046875,
0.81005859375,
0.62060546875,
0.75244140625,
0.62353515625
],
[
0.6572265625,
0.615234375,
0.73046875,
0.619140625,
0.72900390625,
0.63427734375,
0.65576171875,
0.6298828125
],
[
0.118408203125,
0.62744140625,
0.484375,
0.62744140625,
0.484375,
0.64111328125,
0.118408203125,
0.64111328125
],
[
0.8388671875,
0.62353515625,
0.87890625,
0.62158203125,
0.880859375,
0.63427734375,
0.84033203125,
0.6357421875
],
[
0.71875,
0.6298828125,
0.84814453125,
0.62353515625,
0.8486328125,
0.64111328125,
0.7197265625,
0.64794921875
],
[
0.118408203125,
0.6435546875,
0.484375,
0.6435546875,
0.484375,
0.65673828125,
0.118408203125,
0.65673828125
],
[
0.12005615234375,
0.66015625,
0.484375,
0.66015625,
0.484375,
0.67333984375,
0.12005615234375,
0.67333984375
],
[
0.509765625,
0.6591796875,
0.7080078125,
0.6591796875,
0.7080078125,
0.6728515625,
0.509765625,
0.6728515625
],
[
0.1175537109375,
0.67578125,
0.484375,
0.6767578125,
0.484375,
0.6904296875,
0.1175537109375,
0.689453125
],
[
0.51318359375,
0.6796875,
0.73291015625,
0.6796875,
0.73291015625,
0.693359375,
0.51318359375,
0.693359375
],
[
0.118408203125,
0.6923828125,
0.48193359375,
0.6923828125,
0.48193359375,
0.703125,
0.118408203125,
0.703125
],
[
0.54150390625,
0.70166015625,
0.88330078125,
0.70361328125,
0.88330078125,
0.724609375,
0.54150390625,
0.72265625
],
[
0.12005615234375,
0.7080078125,
0.484375,
0.7080078125,
0.484375,
0.7216796875,
0.12005615234375,
0.7216796875
],
[
0.118408203125,
0.7236328125,
0.339111328125,
0.7236328125,
0.339111328125,
0.73681640625,
0.118408203125,
0.73681640625
],
[
0.5263671875,
0.73388671875,
0.880859375,
0.73388671875,
0.880859375,
0.74755859375,
0.5263671875,
0.74755859375
],
[
0.267822265625,
0.7548828125,
0.2822265625,
0.7548828125,
0.2822265625,
0.76318359375,
0.267822265625,
0.76318359375
],
[
0.5146484375,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.240966796875,
0.7626953125,
0.2626953125,
0.76416015625,
0.26123046875,
0.7724609375,
0.2392578125,
0.77099609375
],
[
0.26025390625,
0.76416015625,
0.429931640625,
0.76708984375,
0.429931640625,
0.78515625,
0.26025390625,
0.7822265625
],
[
0.5146484375,
0.7666015625,
0.880859375,
0.7666015625,
0.880859375,
0.77978515625,
0.5146484375,
0.77978515625
],
[
0.2359619140625,
0.765625,
0.295654296875,
0.779296875,
0.285400390625,
0.802734375,
0.225830078125,
0.7890625
],
[
0.130126953125,
0.76708984375,
0.255126953125,
0.765625,
0.255126953125,
0.783203125,
0.130126953125,
0.78515625
],
[
0.258544921875,
0.76953125,
0.267822265625,
0.76953125,
0.267822265625,
0.7724609375,
0.258544921875,
0.7724609375
],
[
0.465087890625,
0.7822265625,
0.486083984375,
0.78369140625,
0.484375,
0.79345703125,
0.46337890625,
0.79150390625
],
[
0.497802734375,
0.78759765625,
0.50537109375,
0.78759765625,
0.50537109375,
0.79150390625,
0.497802734375,
0.79150390625
],
[
0.51318359375,
0.783203125,
0.73291015625,
0.783203125,
0.73291015625,
0.79638671875,
0.51318359375,
0.79638671875
],
[
0.22412109375,
0.8046875,
0.443359375,
0.8046875,
0.443359375,
0.8193359375,
0.22412109375,
0.8193359375
],
[
0.51123046875,
0.80859375,
0.7119140625,
0.81005859375,
0.7119140625,
0.82470703125,
0.51123046875,
0.82373046875
],
[
0.118408203125,
0.84033203125,
0.48779296875,
0.8388671875,
0.48779296875,
0.85400390625,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.8359375,
0.62451171875,
0.8359375,
0.62451171875,
0.84716796875,
0.51318359375,
0.84716796875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51318359375,
0.85546875,
0.880859375,
0.85546875,
0.880859375,
0.869140625,
0.51318359375,
0.869140625
],
[
0.12005615234375,
0.87451171875,
0.483642578125,
0.87451171875,
0.483642578125,
0.88525390625,
0.12005615234375,
0.88525390625
],
[
0.509765625,
0.87158203125,
0.55908203125,
0.869140625,
0.56103515625,
0.88427734375,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.484375,
0.8896484375,
0.484375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.5322265625,
0.89111328125,
0.880859375,
0.89453125,
0.880859375,
0.9091796875,
0.5322265625,
0.90625
],
[
0.1175537109375,
0.90380859375,
0.19482421875,
0.90625,
0.19482421875,
0.92041015625,
0.1175537109375,
0.91845703125
],
[
0.51123046875,
0.90380859375,
0.6572265625,
0.90478515625,
0.6572265625,
0.91943359375,
0.51123046875,
0.91845703125
]
] |
[
"码段不需要相互关注。不失一般性，我们假",
"",
"3.4Leveraging Extractive Oracles",
"设Pe(-lq,a,y<t）是通过首先将输入（q,a,y<t)",
"对于长输入汇总，在训练期间使用的提取片段",
"映射到上下文化表示向量h来计算的。对",
"Xk对于稳定优化很重要。我们建议在训练过",
"于Transformers（Vaswanietal.，2017）和带有",
"程中利用提取神谕，而不是使用公式E（1）中",
"关注模型（Bahdanauetal.，2015）的编码器-解",
"定义的XK。在测试期间不使用提取神谕",
"码器，h通常是模型在最终语言模型头之",
"前的输出。通过将h送入语言模型头计",
"Greedy search for extractive oracles Extrac-",
"算生成概率Pe(ytlq,a,y<t）。对于动态权重",
"tiveoracles表示一组选定的文本片段，在给定",
"Pe(αlq,XK，y<t），我们采用单独的MLP将每",
"黄金摘要的情况下，其串联可最大化评估指",
"个h映射到标量logitl，并且Pe(q,X,y<t)",
"标。我们使用贪婪搜索实现提取oracle。具体",
"被定义为softmax({1)rex）。我们通过对所有",
"来说，我们从一个空集开始，我们迭代地从输",
"提取的片段进行边缘化来计算生成概率：",
"入中选择一个片段，使得该片段和已经选择",
"的片段的连接最大化ROUGE-1，ROUGE-2和",
"T",
"ROUGE-L分数的平均值给出黄金摘要。我们",
"Pe(ula, Xk) =II ",
"将提取神谕表示为 X。",
"(3)",
"t-l reXk",
"Hybridtraining我们利用提取神谕来定义训",
"Pe(ytlq, , y<t)Pe(alq, X,<t)",
"练期间使用的XK。如果oracles的数量等于",
"或超过K，我们将XK定义为第一个Koracle",
"每个解码时间步长t的动态权重",
"片段。如果oracles的数量小于K，我们将Xk",
"Pe(cg,Xk,y<t）允许我们解释生成器如",
"定义为X。和未出现在X。中的提取器排名最",
"何利用提取的片段。例如，对特定代码段的权",
"高的代码段的联合。这种混合训练有两个好",
"重越大，表示该代码段对当前解码时间步长",
"处。首先，与等式E（1）中定义的XK相比，",
"的重要性越大。发电损失定义为黄金汇总的",
"它为生成器提供了更高质量的输入。其次，它",
"NLL:",
"减少了对oracle的依赖，并提高了我们模型在",
"Cgen = -log Pe(ylq, Xk)",
"(4)",
"训练集之外的泛化能力，因为贪婪搜索中省略",
"的其他文本片段可能有助于生成",
"其中Pe(ylq,Xk)在公式国(2）中定义。在这",
"里，我们不将Ceen的梯度传播到提取器参数",
"Oracleloss提取神谕X。用作我们模型提取",
"因为top-K是不可微分的。相反，在3.3节",
"部分的监督信号。oracle损失Coracle是根据提",
"和3.4节中描述了优化提取器的方法",
"取器所选集合中的所有块与提取oracle之间的",
"交叉摘损失计算得出的。形式上，oracle损失",
"Consistencyloss我们还利用动态权重为提",
"计算为",
"取器提供训练信号。由于代码段的动态权重可",
"以解释为代码段在特定时间步长的重要性，因",
"Coracle = -[Xol rex。",
"1",
"eEn(q,a)",
"log ",
"此我们对所有解码步骤的动态权重进行平均，",
"(6)",
"EriexeE,(q,ri)",
"并将平均权重视为代码段的总体重要性。基",
"于这种直觉，我们提出了我们所说的一致性损",
"3.5Training Objective",
"失，它测量平均动态权重分布和提取器分布之",
"我们方法的总体训练目标是",
"间的距离。我们希望这两个分布在X的任意",
"Con = AgCgen + AoC'racle + AeC'consist(7)",
"子集上接近。为简单起见，我们将Xk作为子",
"集，并将一致性损失定义为",
"其中入g、入。和入。是平衡损耗分量的超参数。",
"T",
"计算上标参数的梯度。具体来说，提取器仅使",
"1",
" Pe(-lq, Xk, y<t) I",
"用一致性损失和oracle损失进行优化，而生成",
"",
"Cconsist = KL",
"",
"(5)",
"",
"器仅使用生成损失进行优化",
"softmax(En(q, ai), c E Xk)",
"4Experiment Setups",
"请注意，一致性损失是使用提取器的参数进",
"4.1Datasets",
"行上标的，这意味着我们不计算生成器参数θ",
"我们将以下长输入抽象摘要数据集视为评估基",
"的梯度。由于我们希望X的任意子集上的分",
"准：2",
"布距离较小，因此我们不会通过top-K运算符",
"2QMSum and arXiv can be accessed through Summer",
"传播梯度",
"Time (Ni et al., 2021a) ,"
] |
{
"arxiv_id": "2110.08168",
"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.