id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2203.15190-zh-4
|
2203.15190-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>Methods</th>
<th>Average</th>
<th>Plane</th>
<th>Cabinet</th>
<th>Car</th>
<th>hair</th>
<th>Sofa</th>
<th>Table</th>
<th>$Water.$</th>
<th>Bed</th>
<th>ch Sh$\epsilon$</th>
<th>$lf$ Bus</th>
<th>Guitar</th>
<th>Motor.</th>
<th>Pistol</th>
<th>Skat$\epsilon$</th>
</tr>
<tr>
<td>$PCN$ $\begin{bmatrix}54\\0\end{bmatrix}$</td>
<td>9.80</td>
<td>4.22</td>
<td>8.92</td>
<td>6.49</td>
<td>2.46</td>
<td>9.92</td>
<td>12.45</td>
<td>8.78</td>
<td>19.0</td>
<td>$\textcircled{1}$ 13.:</td>
<td>9 5.1</td>
<td>1.87</td>
<td>6.03</td>
<td>6.04</td>
<td>4.70</td>
</tr>
<tr>
<td>$\operatorname{TopNet}$ $\begin{bmatrix}3.3\\3.3\end{bmatrix}$</td>
<td>10.34</td>
<td>4.09</td>
<td>9.71</td>
<td>7.36</td>
<td>3.46</td>
<td>11.21</td>
<td>12.46</td>
<td>8.50</td>
<td>18.98</td>
<td>.8 15.</td>
<td>5 5.47</td>
<td>2.13</td>
<td>7.19</td>
<td>7.33</td>
<td>4.15</td>
</tr>
<tr>
<td>$MSN$ $[222^{\prime}$</td>
<td>7.98</td>
<td>2.59</td>
<td>8.86</td>
<td>6.54</td>
<td>0.22</td>
<td>9.08</td>
<td>9.69</td>
<td>7.08</td>
<td>15.58</td>
<td>:8 11.</td>
<td>5.2 1</td>
<td>1.37</td>
<td>4.63</td>
<td>4.72</td>
<td>3.06</td>
</tr>
<tr>
<td>$CRN$ $139^{\circ}$</td>
<td>7.34</td>
<td>2.45</td>
<td>8.62</td>
<td>5.97</td>
<td>3.95</td>
<td>8.63</td>
<td>9.30</td>
<td>6.43</td>
<td>14.93</td>
<td>1 10.</td>
<td>9 4.97</td>
<td>1.67</td>
<td>4.33</td>
<td>4.47</td>
<td>3.39</td>
</tr>
<tr>
<td>VRCNet $[29]^{*}$</td>
<td>5.96</td>
<td>2.17</td>
<td>7.83</td>
<td>5.52</td>
<td>.31</td>
<td>7.42</td>
<td>7.07</td>
<td>5.15</td>
<td>11.18</td>
<td>$'6$ 7.0</td>
<td>4.4( 3</td>
<td>1.15</td>
<td>3.75</td>
<td>3.54</td>
<td>2.31</td>
</tr>
<tr>
<td>PMPNet [47</td>
<td>6.24</td>
<td>1.99</td>
<td>8.84</td>
<td>6.36</td>
<td>.77</td>
<td>8.72</td>
<td>7.71</td>
<td>5.19</td>
<td>11.77</td>
<td>8.3 17</td>
<td>5.27 1</td>
<td>1.27</td>
<td>3.95</td>
<td>3.57</td>
<td>2.35</td>
</tr>
<tr>
<td>SnowfakeNet [49]</td>
<td>5.86</td>
<td>2.04</td>
<td>7.76</td>
<td>5.61</td>
<td>.07</td>
<td>6.92</td>
<td>7.13</td>
<td>5.05</td>
<td>11.32</td>
<td>:7 7.7</td>
<td>4.4 2</td>
<td>1.16</td>
<td>3.94</td>
<td>3.52</td>
<td>3.64</td>
</tr>
<tr>
<td>3DAtriFlow(Ours)</td>
<td>5.06</td>
<td>1.59</td>
<td>7.40</td>
<td>5.44</td>
<td>$i.05$</td>
<td>6.81</td>
<td>6.14</td>
<td>4.25</td>
<td>10.62</td>
<td>3 6.5</td>
<td>4.3( 3</td>
<td>0.95</td>
<td>3.27</td>
<td>2.78</td>
<td>1.78</td>
</tr>
</tbody>
</table>
<FigureHere>
Figure 5. Visual comparison of point cloud completion re-
sults with different methods under MVP dataset.
<FigureHere>
Figure 4. Visual comparison of 2D-to-3D reconstruction
特别是，PSGN [4] 和 AtlasNet [5] 是基于点云的方法它们与 3DAttriFlow 最相关。然而，与这两种方法相比，3DAttriFlow 实现了超过 25 个% 性能增益。正如我们在 SEC 中讨论的那样。1,上述两种方法采用了2D 到 3D 重建的典型范例，其中 AtlasNet[5] 基于全局特征的隐式输入直接解码整个形状，而 PSGN[4]则利用编码器和解码器之间的特征通道来引人各种级别的语义。这些实践都不能从图像中学习显式语义特征， 而只能尝试从编码器的隐式全局特征或中间层解码形状。相比之下，3DAttriFlow 可以利用从图像中学习的隐式和显式语义属性，分别通过几何子管道和语义子管道。因此，3DAttriFlow 能够基于来自显式语义属性的更明确的指导来预测 3D 形状的细节，并且比它的同行获得更好的性能
定性比较。2D 到 3D 重建的视觉比较如图 4 所示。请注意，对于 AtlasNet,我们遵循其原始可视化设置来展示重建的网格而不是点云。与其他方法相比， 3DAttriFlow 在广泛的对象类别上重建了更好的细节。例如，在椅子类别(图 4的 5$^th$行)上，在 OccNet 的预测中缺少腿，而 PSGN 和 AtlasNet 的椅子预测则模棱两可且充满噪音。至于飞机类别，PSGN 和 AtlasNet 都无法重建 1$^{st}$和 2$^{nd}$行中发动机的详细形状，而OccNet 无法稳定地对发动机进行正确的预测$(2^{nd}$行中的故障)
results with different methods under ShapeNet dataset.
4.1. 2D-to-3D Reconstruction on ShapeNet dataset
## $4. 2. 3\mathbf{D} \textbf{Completion on MVP dataset}$
数据集摘要和评估度量。我们按照 OccNet [26] 的实验设置来评估 ShapeNet 数据集 [1] 上的 3DAttriFlow。整个数据集由 43,783 个网格对象组成，包含 13 个类别，将按照 OccNet [26] 的相同策略分为训练，验证和测试。由于我们的方法侧重于从 2D 图像重建 3D 点云，因此我们遵循 AtlasNet[5]在 3D 对象的网格表面上均匀采样 30k 点，作为训练的地面实况。遵循先前的方法[26,37] ,我们使用由公式 (6)描述的 L1 倒角距离作为评估指标。为了与其他重建 3D 体素或网格的方法进行比较，我们遵循 OccNet [26] 从其输出表面采样 2,048 个点，然后计算 L1 与地面真相的倒角距离。对于基于体素的方法，我们还将其体素输出传输到网格中，然后在网格表面上应用点云采样
定量比较。2D 到 3D 重建的结果如表 1 所示，其中 3DAttriFlow 实现了优于其他比较对应方法的性能。
数据集简介和评估度量。我们按照 VRCNet[29] 的实验设置来评估我们在 MVP 数据集 [29] 上的 3DAttriFlow。该数据集由从 ShapeNet 中选择的模型生成的16 类不完整/完整点云组成，然后分为训练集(62,400个形状对)和测试集(41,600 个形状对)。遵循以前的方法[29,33,47] ,我们使用 L2 倒角距离作为评估指标
定量比较。点云完成情况对比见表 2。与当前最先进的完成方法 SnowflakeNet [49]相比，3DAttriFlow 将
|
[
[
0.10296630859375,
0.0889892578125,
0.86767578125,
0.0889892578125,
0.86767578125,
0.10418701171875,
0.10296630859375,
0.10418701171875
],
[
0.0882568359375,
0.1060791015625,
0.134033203125,
0.1060791015625,
0.134033203125,
0.117431640625,
0.0882568359375,
0.117431640625
],
[
0.281005859375,
0.1060791015625,
0.32275390625,
0.1060791015625,
0.32275390625,
0.117431640625,
0.281005859375,
0.117431640625
],
[
0.365234375,
0.1060791015625,
0.39794921875,
0.1060791015625,
0.39794921875,
0.117431640625,
0.365234375,
0.117431640625
],
[
0.404296875,
0.10479736328125,
0.435546875,
0.106689453125,
0.4345703125,
0.11871337890625,
0.40283203125,
0.11614990234375
],
[
0.44775390625,
0.1060791015625,
0.47216796875,
0.1060791015625,
0.47216796875,
0.117431640625,
0.44775390625,
0.117431640625
],
[
0.484375,
0.1060791015625,
0.51318359375,
0.1060791015625,
0.51318359375,
0.117431640625,
0.484375,
0.117431640625
],
[
0.52294921875,
0.1060791015625,
0.5556640625,
0.1060791015625,
0.5556640625,
0.117431640625,
0.52294921875,
0.117431640625
],
[
0.57177734375,
0.1060791015625,
0.5947265625,
0.1060791015625,
0.5947265625,
0.117431640625,
0.57177734375,
0.117431640625
],
[
0.60546875,
0.1060791015625,
0.6396484375,
0.1060791015625,
0.6396484375,
0.117431640625,
0.60546875,
0.117431640625
],
[
0.650390625,
0.1060791015625,
0.6787109375,
0.1060791015625,
0.6787109375,
0.117431640625,
0.650390625,
0.117431640625
],
[
0.7197265625,
0.1060791015625,
0.75244140625,
0.1060791015625,
0.75244140625,
0.117431640625,
0.7197265625,
0.117431640625
],
[
0.7646484375,
0.1060791015625,
0.7958984375,
0.1060791015625,
0.7958984375,
0.117431640625,
0.7646484375,
0.117431640625
],
[
0.80810546875,
0.1060791015625,
0.8388671875,
0.1060791015625,
0.8388671875,
0.117431640625,
0.80810546875,
0.117431640625
],
[
0.3349609375,
0.106689453125,
0.35302734375,
0.106689453125,
0.35302734375,
0.11614990234375,
0.3349609375,
0.11614990234375
],
[
0.1871337890625,
0.106689453125,
0.27294921875,
0.106689453125,
0.27294921875,
0.117431640625,
0.1871337890625,
0.117431640625
],
[
0.68896484375,
0.106689453125,
0.708984375,
0.106689453125,
0.708984375,
0.117431640625,
0.68896484375,
0.117431640625
],
[
0.8486328125,
0.10797119140625,
0.8798828125,
0.10797119140625,
0.8798828125,
0.11614990234375,
0.8486328125,
0.11614990234375
],
[
0.08905029296875,
0.11932373046875,
0.1356201171875,
0.11932373046875,
0.1356201171875,
0.1307373046875,
0.08905029296875,
0.1307373046875
],
[
0.291748046875,
0.11932373046875,
0.314453125,
0.11932373046875,
0.314453125,
0.1307373046875,
0.291748046875,
0.1307373046875
],
[
0.33251953125,
0.11932373046875,
0.35693359375,
0.11932373046875,
0.35693359375,
0.1319580078125,
0.33251953125,
0.1319580078125
],
[
0.36767578125,
0.11932373046875,
0.396240234375,
0.11932373046875,
0.396240234375,
0.1307373046875,
0.36767578125,
0.1307373046875
],
[
0.406005859375,
0.11932373046875,
0.4345703125,
0.11932373046875,
0.4345703125,
0.1307373046875,
0.406005859375,
0.1307373046875
],
[
0.449462890625,
0.11932373046875,
0.47216796875,
0.11932373046875,
0.47216796875,
0.1307373046875,
0.449462890625,
0.1307373046875
],
[
0.486083984375,
0.11932373046875,
0.5146484375,
0.11932373046875,
0.5146484375,
0.1319580078125,
0.486083984375,
0.1319580078125
],
[
0.529296875,
0.11932373046875,
0.55224609375,
0.11932373046875,
0.55224609375,
0.1319580078125,
0.529296875,
0.1319580078125
],
[
0.57177734375,
0.11932373046875,
0.59326171875,
0.11932373046875,
0.59326171875,
0.1319580078125,
0.57177734375,
0.1319580078125
],
[
0.650390625,
0.11932373046875,
0.67724609375,
0.11932373046875,
0.67724609375,
0.1307373046875,
0.650390625,
0.1307373046875
],
[
0.685546875,
0.11932373046875,
0.7099609375,
0.11932373046875,
0.7099609375,
0.1319580078125,
0.685546875,
0.1319580078125
],
[
0.72607421875,
0.11932373046875,
0.74853515625,
0.11932373046875,
0.74853515625,
0.1319580078125,
0.72607421875,
0.1319580078125
],
[
0.76953125,
0.11932373046875,
0.79345703125,
0.11932373046875,
0.79345703125,
0.1307373046875,
0.76953125,
0.1307373046875
],
[
0.81396484375,
0.11932373046875,
0.8349609375,
0.11932373046875,
0.8349609375,
0.1319580078125,
0.81396484375,
0.1319580078125
],
[
0.85693359375,
0.11871337890625,
0.8798828125,
0.121826171875,
0.8759765625,
0.1331787109375,
0.85302734375,
0.1300048828125
],
[
0.61376953125,
0.12060546875,
0.6298828125,
0.12060546875,
0.6298828125,
0.1300048828125,
0.61376953125,
0.1300048828125
],
[
0.2017822265625,
0.121826171875,
0.22216796875,
0.121826171875,
0.22216796875,
0.1300048828125,
0.2017822265625,
0.1300048828125
],
[
0.2467041015625,
0.12060546875,
0.26953125,
0.12060546875,
0.26953125,
0.1319580078125,
0.2467041015625,
0.1319580078125
],
[
0.366943359375,
0.1287841796875,
0.39794921875,
0.1287841796875,
0.39794921875,
0.14453125,
0.366943359375,
0.14453125
],
[
0.68310546875,
0.1275634765625,
0.71337890625,
0.1275634765625,
0.71337890625,
0.154052734375,
0.68310546875,
0.154052734375
],
[
0.81201171875,
0.1287841796875,
0.83837890625,
0.1287841796875,
0.83837890625,
0.154052734375,
0.81201171875,
0.154052734375
],
[
0.2900390625,
0.1300048828125,
0.314453125,
0.1300048828125,
0.314453125,
0.143310546875,
0.2900390625,
0.143310546875
],
[
0.200927734375,
0.1300048828125,
0.226318359375,
0.1300048828125,
0.226318359375,
0.1427001953125,
0.200927734375,
0.1427001953125
],
[
0.334228515625,
0.1300048828125,
0.35693359375,
0.1300048828125,
0.35693359375,
0.143310546875,
0.334228515625,
0.143310546875
],
[
0.44677734375,
0.1300048828125,
0.473876953125,
0.1300048828125,
0.473876953125,
0.1427001953125,
0.44677734375,
0.1427001953125
],
[
0.484375,
0.1300048828125,
0.51318359375,
0.1300048828125,
0.51318359375,
0.1427001953125,
0.484375,
0.1427001953125
],
[
0.529296875,
0.1300048828125,
0.55224609375,
0.1300048828125,
0.55224609375,
0.143310546875,
0.529296875,
0.143310546875
],
[
0.56884765625,
0.1300048828125,
0.59716796875,
0.1300048828125,
0.59716796875,
0.14453125,
0.56884765625,
0.14453125
],
[
0.611328125,
0.1300048828125,
0.6337890625,
0.1300048828125,
0.6337890625,
0.143310546875,
0.611328125,
0.143310546875
],
[
0.650390625,
0.1300048828125,
0.67724609375,
0.1300048828125,
0.67724609375,
0.1427001953125,
0.650390625,
0.1427001953125
],
[
0.7255859375,
0.1300048828125,
0.74853515625,
0.1300048828125,
0.74853515625,
0.143310546875,
0.7255859375,
0.143310546875
],
[
0.76953125,
0.1300048828125,
0.79345703125,
0.1300048828125,
0.79345703125,
0.143310546875,
0.76953125,
0.143310546875
],
[
0.85546875,
0.1300048828125,
0.87744140625,
0.1300048828125,
0.87744140625,
0.143310546875,
0.85546875,
0.143310546875
],
[
0.2459716796875,
0.1307373046875,
0.26953125,
0.1307373046875,
0.26953125,
0.1427001953125,
0.2459716796875,
0.1427001953125
],
[
0.0906982421875,
0.1307373046875,
0.148681640625,
0.1307373046875,
0.148681640625,
0.1427001953125,
0.0906982421875,
0.1427001953125
],
[
0.406005859375,
0.1307373046875,
0.4345703125,
0.1287841796875,
0.435546875,
0.1427001953125,
0.406982421875,
0.14453125
],
[
0.2459716796875,
0.1395263671875,
0.26953125,
0.1395263671875,
0.26953125,
0.1514892578125,
0.2459716796875,
0.1514892578125
],
[
0.291748046875,
0.1395263671875,
0.314453125,
0.1395263671875,
0.314453125,
0.1514892578125,
0.291748046875,
0.1514892578125
],
[
0.33251953125,
0.1395263671875,
0.35693359375,
0.1395263671875,
0.35693359375,
0.1514892578125,
0.33251953125,
0.1514892578125
],
[
0.366943359375,
0.1395263671875,
0.39453125,
0.1395263671875,
0.39453125,
0.1514892578125,
0.366943359375,
0.1514892578125
],
[
0.406982421875,
0.1395263671875,
0.435546875,
0.1395263671875,
0.435546875,
0.1514892578125,
0.406982421875,
0.1514892578125
],
[
0.44775390625,
0.1395263671875,
0.47216796875,
0.1395263671875,
0.47216796875,
0.1514892578125,
0.44775390625,
0.1514892578125
],
[
0.486083984375,
0.1383056640625,
0.51318359375,
0.141357421875,
0.5107421875,
0.15283203125,
0.483642578125,
0.149658203125
],
[
0.529296875,
0.1395263671875,
0.55224609375,
0.1395263671875,
0.55224609375,
0.1514892578125,
0.529296875,
0.1514892578125
],
[
0.56884765625,
0.1395263671875,
0.59619140625,
0.1395263671875,
0.59619140625,
0.1514892578125,
0.56884765625,
0.1514892578125
],
[
0.611328125,
0.1395263671875,
0.6337890625,
0.1395263671875,
0.6337890625,
0.1514892578125,
0.611328125,
0.1514892578125
],
[
0.64892578125,
0.1395263671875,
0.67626953125,
0.1395263671875,
0.67626953125,
0.1514892578125,
0.64892578125,
0.1514892578125
],
[
0.7255859375,
0.1395263671875,
0.74853515625,
0.1395263671875,
0.74853515625,
0.1514892578125,
0.7255859375,
0.1514892578125
],
[
0.76953125,
0.1395263671875,
0.79345703125,
0.1395263671875,
0.79345703125,
0.1514892578125,
0.76953125,
0.1514892578125
],
[
0.8544921875,
0.1395263671875,
0.87744140625,
0.1395263671875,
0.87744140625,
0.1514892578125,
0.8544921875,
0.1514892578125
],
[
0.200927734375,
0.14013671875,
0.2254638671875,
0.14013671875,
0.2254638671875,
0.1514892578125,
0.200927734375,
0.1514892578125
],
[
0.0906982421875,
0.14013671875,
0.1356201171875,
0.14013671875,
0.1356201171875,
0.1514892578125,
0.0906982421875,
0.1514892578125
],
[
0.330810546875,
0.146484375,
0.358642578125,
0.146484375,
0.358642578125,
0.173583984375,
0.330810546875,
0.173583984375
],
[
0.76953125,
0.147705078125,
0.7939453125,
0.15087890625,
0.79150390625,
0.1610107421875,
0.76708984375,
0.1578369140625
],
[
0.81396484375,
0.147705078125,
0.8349609375,
0.147705078125,
0.8349609375,
0.160400390625,
0.81396484375,
0.160400390625
],
[
0.85302734375,
0.1458740234375,
0.88134765625,
0.147705078125,
0.87744140625,
0.1749267578125,
0.8486328125,
0.1717529296875
],
[
0.2467041015625,
0.1483154296875,
0.26953125,
0.1483154296875,
0.26953125,
0.1610107421875,
0.2467041015625,
0.1610107421875
],
[
0.2017822265625,
0.1483154296875,
0.223876953125,
0.1483154296875,
0.223876953125,
0.1622314453125,
0.2017822265625,
0.1622314453125
],
[
0.291748046875,
0.1483154296875,
0.314453125,
0.1483154296875,
0.314453125,
0.1610107421875,
0.291748046875,
0.1610107421875
],
[
0.369384765625,
0.1483154296875,
0.39208984375,
0.1483154296875,
0.39208984375,
0.1610107421875,
0.369384765625,
0.1610107421875
],
[
0.406982421875,
0.1483154296875,
0.4345703125,
0.1483154296875,
0.4345703125,
0.1610107421875,
0.406982421875,
0.1610107421875
],
[
0.449462890625,
0.1483154296875,
0.47216796875,
0.1483154296875,
0.47216796875,
0.1622314453125,
0.449462890625,
0.1622314453125
],
[
0.48779296875,
0.1483154296875,
0.5107421875,
0.1483154296875,
0.5107421875,
0.1610107421875,
0.48779296875,
0.1610107421875
],
[
0.529296875,
0.1483154296875,
0.55224609375,
0.1483154296875,
0.55224609375,
0.1610107421875,
0.529296875,
0.1610107421875
],
[
0.56884765625,
0.1483154296875,
0.59619140625,
0.1483154296875,
0.59619140625,
0.1610107421875,
0.56884765625,
0.1610107421875
],
[
0.611328125,
0.1483154296875,
0.6337890625,
0.1483154296875,
0.6337890625,
0.1622314453125,
0.611328125,
0.1622314453125
],
[
0.650390625,
0.1483154296875,
0.67626953125,
0.1483154296875,
0.67626953125,
0.1610107421875,
0.650390625,
0.1610107421875
],
[
0.68701171875,
0.1483154296875,
0.7099609375,
0.1483154296875,
0.7099609375,
0.1622314453125,
0.68701171875,
0.1622314453125
],
[
0.72607421875,
0.1483154296875,
0.74853515625,
0.1483154296875,
0.74853515625,
0.1610107421875,
0.72607421875,
0.1610107421875
],
[
0.2034912109375,
0.1558837890625,
0.226318359375,
0.1590576171875,
0.22216796875,
0.17236328125,
0.1993408203125,
0.169189453125
],
[
0.0906982421875,
0.149658203125,
0.1356201171875,
0.149658203125,
0.1356201171875,
0.160400390625,
0.0906982421875,
0.160400390625
],
[
0.370849609375,
0.1578369140625,
0.39208984375,
0.1578369140625,
0.39208984375,
0.17236328125,
0.370849609375,
0.17236328125
],
[
0.529296875,
0.1572265625,
0.55224609375,
0.1572265625,
0.55224609375,
0.1717529296875,
0.529296875,
0.1717529296875
],
[
0.2467041015625,
0.1590576171875,
0.26953125,
0.1590576171875,
0.26953125,
0.1717529296875,
0.2467041015625,
0.1717529296875
],
[
0.0906982421875,
0.1590576171875,
0.1502685546875,
0.1590576171875,
0.1502685546875,
0.169189453125,
0.0906982421875,
0.169189453125
],
[
0.291748046875,
0.1590576171875,
0.316162109375,
0.1590576171875,
0.316162109375,
0.1717529296875,
0.291748046875,
0.1717529296875
],
[
0.41015625,
0.1590576171875,
0.43310546875,
0.1590576171875,
0.43310546875,
0.1717529296875,
0.41015625,
0.1717529296875
],
[
0.449462890625,
0.1590576171875,
0.47216796875,
0.1590576171875,
0.47216796875,
0.1717529296875,
0.449462890625,
0.1717529296875
],
[
0.48779296875,
0.1590576171875,
0.5107421875,
0.1590576171875,
0.5107421875,
0.1717529296875,
0.48779296875,
0.1717529296875
],
[
0.56884765625,
0.1590576171875,
0.59619140625,
0.1590576171875,
0.59619140625,
0.1717529296875,
0.56884765625,
0.1717529296875
],
[
0.61279296875,
0.1590576171875,
0.6357421875,
0.1590576171875,
0.6357421875,
0.1717529296875,
0.61279296875,
0.1717529296875
],
[
0.650390625,
0.1590576171875,
0.6748046875,
0.1590576171875,
0.6748046875,
0.1717529296875,
0.650390625,
0.1717529296875
],
[
0.68701171875,
0.1590576171875,
0.7099609375,
0.1590576171875,
0.7099609375,
0.1717529296875,
0.68701171875,
0.1717529296875
],
[
0.72607421875,
0.1590576171875,
0.74853515625,
0.1590576171875,
0.74853515625,
0.1717529296875,
0.72607421875,
0.1717529296875
],
[
0.76953125,
0.1590576171875,
0.79345703125,
0.1590576171875,
0.79345703125,
0.1717529296875,
0.76953125,
0.1717529296875
],
[
0.81396484375,
0.1590576171875,
0.8349609375,
0.1590576171875,
0.8349609375,
0.1717529296875,
0.81396484375,
0.1717529296875
],
[
0.08905029296875,
0.1673583984375,
0.1544189453125,
0.169189453125,
0.1544189453125,
0.18115234375,
0.08905029296875,
0.1787109375
],
[
0.200927734375,
0.1685791015625,
0.2254638671875,
0.1685791015625,
0.2254638671875,
0.18115234375,
0.200927734375,
0.18115234375
],
[
0.2459716796875,
0.1685791015625,
0.271240234375,
0.1685791015625,
0.271240234375,
0.18115234375,
0.2459716796875,
0.18115234375
],
[
0.291748046875,
0.1685791015625,
0.316162109375,
0.1685791015625,
0.316162109375,
0.18115234375,
0.291748046875,
0.18115234375
],
[
0.33251953125,
0.1685791015625,
0.35693359375,
0.1685791015625,
0.35693359375,
0.18115234375,
0.33251953125,
0.18115234375
],
[
0.369384765625,
0.1685791015625,
0.39208984375,
0.1685791015625,
0.39208984375,
0.18115234375,
0.369384765625,
0.18115234375
],
[
0.408447265625,
0.1685791015625,
0.43310546875,
0.1685791015625,
0.43310546875,
0.18115234375,
0.408447265625,
0.18115234375
],
[
0.449462890625,
0.1673583984375,
0.47216796875,
0.1673583984375,
0.47216796875,
0.18115234375,
0.449462890625,
0.18115234375
],
[
0.48779296875,
0.1685791015625,
0.5107421875,
0.1685791015625,
0.5107421875,
0.18115234375,
0.48779296875,
0.18115234375
],
[
0.529296875,
0.1685791015625,
0.5537109375,
0.1685791015625,
0.5537109375,
0.18115234375,
0.529296875,
0.18115234375
],
[
0.56884765625,
0.1685791015625,
0.59716796875,
0.1685791015625,
0.59716796875,
0.18115234375,
0.56884765625,
0.18115234375
],
[
0.611328125,
0.1685791015625,
0.6357421875,
0.1685791015625,
0.6357421875,
0.18115234375,
0.611328125,
0.18115234375
],
[
0.650390625,
0.1685791015625,
0.6748046875,
0.1685791015625,
0.6748046875,
0.18115234375,
0.650390625,
0.18115234375
],
[
0.685546875,
0.1685791015625,
0.7099609375,
0.1685791015625,
0.7099609375,
0.18115234375,
0.685546875,
0.18115234375
],
[
0.7255859375,
0.1685791015625,
0.7490234375,
0.1685791015625,
0.7490234375,
0.18115234375,
0.7255859375,
0.18115234375
],
[
0.76953125,
0.1685791015625,
0.79345703125,
0.1685791015625,
0.79345703125,
0.18115234375,
0.76953125,
0.18115234375
],
[
0.81201171875,
0.1685791015625,
0.8349609375,
0.1685791015625,
0.8349609375,
0.18115234375,
0.81201171875,
0.18115234375
],
[
0.8544921875,
0.1685791015625,
0.87744140625,
0.1685791015625,
0.87744140625,
0.18115234375,
0.8544921875,
0.18115234375
],
[
0.3349609375,
0.175537109375,
0.35693359375,
0.179931640625,
0.35302734375,
0.1925048828125,
0.330810546875,
0.1881103515625
],
[
0.410888671875,
0.175537109375,
0.4345703125,
0.1787109375,
0.4306640625,
0.1925048828125,
0.406982421875,
0.189453125
],
[
0.53125,
0.175537109375,
0.55224609375,
0.1787109375,
0.5498046875,
0.1925048828125,
0.52685546875,
0.189453125
],
[
0.68896484375,
0.175537109375,
0.71142578125,
0.1787109375,
0.70751953125,
0.1925048828125,
0.6845703125,
0.189453125
],
[
0.85693359375,
0.175537109375,
0.8798828125,
0.1787109375,
0.8759765625,
0.1925048828125,
0.85302734375,
0.189453125
],
[
0.369384765625,
0.177978515625,
0.39208984375,
0.177978515625,
0.39208984375,
0.1900634765625,
0.369384765625,
0.1900634765625
],
[
0.08905029296875,
0.177978515625,
0.1732177734375,
0.1787109375,
0.1732177734375,
0.1900634765625,
0.08905029296875,
0.189453125
],
[
0.2017822265625,
0.1787109375,
0.2254638671875,
0.1787109375,
0.2254638671875,
0.1900634765625,
0.2017822265625,
0.1900634765625
],
[
0.2467041015625,
0.1787109375,
0.26953125,
0.1787109375,
0.26953125,
0.1912841796875,
0.2467041015625,
0.1912841796875
],
[
0.291748046875,
0.1787109375,
0.314453125,
0.1787109375,
0.314453125,
0.1912841796875,
0.291748046875,
0.1912841796875
],
[
0.449462890625,
0.1787109375,
0.47216796875,
0.1787109375,
0.47216796875,
0.1900634765625,
0.449462890625,
0.1900634765625
],
[
0.48779296875,
0.1787109375,
0.5107421875,
0.1787109375,
0.5107421875,
0.1900634765625,
0.48779296875,
0.1900634765625
],
[
0.56884765625,
0.1787109375,
0.59619140625,
0.1787109375,
0.59619140625,
0.1900634765625,
0.56884765625,
0.1900634765625
],
[
0.650390625,
0.1787109375,
0.67333984375,
0.1787109375,
0.67333984375,
0.1912841796875,
0.650390625,
0.1912841796875
],
[
0.76953125,
0.1787109375,
0.79345703125,
0.1787109375,
0.79345703125,
0.1900634765625,
0.76953125,
0.1900634765625
],
[
0.81396484375,
0.177978515625,
0.8349609375,
0.177978515625,
0.8349609375,
0.1900634765625,
0.81396484375,
0.1900634765625
],
[
0.72802734375,
0.179931640625,
0.74853515625,
0.179931640625,
0.74853515625,
0.1900634765625,
0.72802734375,
0.1900634765625
],
[
0.61376953125,
0.179931640625,
0.6337890625,
0.179931640625,
0.6337890625,
0.1900634765625,
0.61376953125,
0.1900634765625
],
[
0.450927734375,
0.1900634765625,
0.47216796875,
0.1944580078125,
0.468017578125,
0.205810546875,
0.44677734375,
0.2020263671875
],
[
0.488525390625,
0.189453125,
0.51123046875,
0.1932373046875,
0.5087890625,
0.205810546875,
0.484375,
0.20263671875
],
[
0.2017822265625,
0.1925048828125,
0.2254638671875,
0.1925048828125,
0.2254638671875,
0.203857421875,
0.2017822265625,
0.203857421875
],
[
0.08905029296875,
0.1925048828125,
0.1846923828125,
0.1925048828125,
0.1846923828125,
0.203857421875,
0.08905029296875,
0.203857421875
],
[
0.2459716796875,
0.1925048828125,
0.26953125,
0.1925048828125,
0.26953125,
0.203857421875,
0.2459716796875,
0.203857421875
],
[
0.291748046875,
0.1925048828125,
0.314453125,
0.1925048828125,
0.314453125,
0.2052001953125,
0.291748046875,
0.2052001953125
],
[
0.334228515625,
0.1925048828125,
0.35693359375,
0.1925048828125,
0.35693359375,
0.203857421875,
0.334228515625,
0.203857421875
],
[
0.369384765625,
0.1925048828125,
0.39208984375,
0.1925048828125,
0.39208984375,
0.203857421875,
0.369384765625,
0.203857421875
],
[
0.56884765625,
0.1925048828125,
0.59619140625,
0.1925048828125,
0.59619140625,
0.203857421875,
0.56884765625,
0.203857421875
],
[
0.611328125,
0.1925048828125,
0.6337890625,
0.1925048828125,
0.6337890625,
0.203857421875,
0.611328125,
0.203857421875
],
[
0.650390625,
0.1925048828125,
0.67333984375,
0.1925048828125,
0.67333984375,
0.203857421875,
0.650390625,
0.203857421875
],
[
0.76953125,
0.1925048828125,
0.79150390625,
0.1925048828125,
0.79150390625,
0.2052001953125,
0.76953125,
0.2052001953125
],
[
0.81396484375,
0.1925048828125,
0.8349609375,
0.1925048828125,
0.8349609375,
0.2052001953125,
0.81396484375,
0.2052001953125
],
[
0.85546875,
0.1925048828125,
0.87841796875,
0.1925048828125,
0.87841796875,
0.2052001953125,
0.85546875,
0.2052001953125
],
[
0.53173828125,
0.1932373046875,
0.5498046875,
0.1932373046875,
0.5498046875,
0.20263671875,
0.53173828125,
0.20263671875
],
[
0.410888671875,
0.1932373046875,
0.431396484375,
0.1932373046875,
0.431396484375,
0.203857421875,
0.410888671875,
0.203857421875
],
[
0.689453125,
0.1932373046875,
0.7099609375,
0.1932373046875,
0.7099609375,
0.203857421875,
0.689453125,
0.203857421875
],
[
0.72607421875,
0.1932373046875,
0.74658203125,
0.1932373046875,
0.74658203125,
0.203857421875,
0.72607421875,
0.203857421875
],
[
0.328369140625,
0.21337890625,
0.441162109375,
0.21337890625,
0.441162109375,
0.2266845703125,
0.328369140625,
0.2266845703125
],
[
0.609375,
0.212158203125,
0.6396484375,
0.212158203125,
0.6396484375,
0.2235107421875,
0.609375,
0.2235107421875
],
[
0.66015625,
0.21337890625,
0.68896484375,
0.21337890625,
0.68896484375,
0.2215576171875,
0.66015625,
0.2215576171875
],
[
0.16259765625,
0.2139892578125,
0.1968994140625,
0.2139892578125,
0.1968994140625,
0.2259521484375,
0.16259765625,
0.2259521484375
],
[
0.11273193359375,
0.2139892578125,
0.1446533203125,
0.2139892578125,
0.1446533203125,
0.2259521484375,
0.11273193359375,
0.2259521484375
],
[
0.2181396484375,
0.2139892578125,
0.25244140625,
0.2139892578125,
0.25244140625,
0.2259521484375,
0.2181396484375,
0.2259521484375
],
[
0.26953125,
0.2139892578125,
0.3203125,
0.2139892578125,
0.3203125,
0.2259521484375,
0.26953125,
0.2259521484375
],
[
0.515625,
0.2139892578125,
0.5400390625,
0.2139892578125,
0.5400390625,
0.2259521484375,
0.515625,
0.2259521484375
],
[
0.7060546875,
0.2139892578125,
0.740234375,
0.2139892578125,
0.740234375,
0.2215576171875,
0.7060546875,
0.2215576171875
],
[
0.771484375,
0.2139892578125,
0.8447265625,
0.2139892578125,
0.8447265625,
0.2215576171875,
0.771484375,
0.2215576171875
],
[
0.66259765625,
0.2310791015625,
0.67236328125,
0.2310791015625,
0.67236328125,
0.23486328125,
0.66259765625,
0.23486328125
],
[
0.560546875,
0.216552734375,
0.57666015625,
0.216552734375,
0.57666015625,
0.2247314453125,
0.560546875,
0.2247314453125
],
[
0.3798828125,
0.2374267578125,
0.41259765625,
0.2374267578125,
0.41259765625,
0.25244140625,
0.3798828125,
0.25244140625
],
[
0.7138671875,
0.2374267578125,
0.736328125,
0.24365234375,
0.73046875,
0.255615234375,
0.708984375,
0.25
],
[
0.80810546875,
0.2392578125,
0.8408203125,
0.2392578125,
0.8408203125,
0.264404296875,
0.80810546875,
0.264404296875
],
[
0.82666015625,
0.236083984375,
0.8349609375,
0.236083984375,
0.8349609375,
0.2392578125,
0.82666015625,
0.2392578125
],
[
0.6640625,
0.2412109375,
0.68896484375,
0.2412109375,
0.68896484375,
0.25390625,
0.6640625,
0.25390625
],
[
0.3447265625,
0.240478515625,
0.369384765625,
0.240478515625,
0.369384765625,
0.251953125,
0.3447265625,
0.251953125
],
[
0.51318359375,
0.242431640625,
0.5361328125,
0.242431640625,
0.5361328125,
0.255615234375,
0.51318359375,
0.255615234375
],
[
0.85693359375,
0.2412109375,
0.8837890625,
0.2412109375,
0.8837890625,
0.25830078125,
0.85693359375,
0.25830078125
],
[
0.76123046875,
0.24560546875,
0.77294921875,
0.24560546875,
0.77294921875,
0.25,
0.76123046875,
0.25
],
[
0.28759765625,
0.24365234375,
0.314453125,
0.24365234375,
0.314453125,
0.251953125,
0.28759765625,
0.251953125
],
[
0.81201171875,
0.25244140625,
0.81884765625,
0.251953125,
0.81884765625,
0.255126953125,
0.81396484375,
0.255615234375
],
[
0.66650390625,
0.26318359375,
0.6787109375,
0.26318359375,
0.6787109375,
0.271484375,
0.66650390625,
0.271484375
],
[
0.85888671875,
0.261962890625,
0.88134765625,
0.261962890625,
0.88134765625,
0.285888671875,
0.85888671875,
0.285888671875
],
[
0.814453125,
0.26513671875,
0.8203125,
0.26513671875,
0.8203125,
0.270751953125,
0.814453125,
0.270751953125
],
[
0.515625,
0.264404296875,
0.52783203125,
0.264404296875,
0.52783203125,
0.27587890625,
0.515625,
0.27587890625
],
[
0.50341796875,
0.34521484375,
0.890625,
0.34521484375,
0.890625,
0.359130859375,
0.50341796875,
0.359130859375
],
[
0.50244140625,
0.35791015625,
0.830078125,
0.359130859375,
0.830078125,
0.373779296875,
0.50244140625,
0.37255859375
],
[
0.5048828125,
0.380615234375,
0.890625,
0.380615234375,
0.890625,
0.395751953125,
0.5048828125,
0.395751953125
],
[
0.50341796875,
0.395751953125,
0.890625,
0.396484375,
0.890625,
0.410888671875,
0.50341796875,
0.410400390625
],
[
0.5048828125,
0.412353515625,
0.89208984375,
0.412353515625,
0.89208984375,
0.425537109375,
0.5048828125,
0.425537109375
],
[
0.5048828125,
0.4267578125,
0.88818359375,
0.4267578125,
0.88818359375,
0.43994140625,
0.5048828125,
0.43994140625
],
[
0.50244140625,
0.442626953125,
0.88818359375,
0.442626953125,
0.88818359375,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.50341796875,
0.45703125,
0.8896484375,
0.455810546875,
0.8896484375,
0.47021484375,
0.50341796875,
0.4716796875
],
[
0.5048828125,
0.472900390625,
0.890625,
0.472900390625,
0.890625,
0.486083984375,
0.5048828125,
0.486083984375
],
[
0.50244140625,
0.486083984375,
0.890625,
0.4873046875,
0.890625,
0.501953125,
0.50244140625,
0.50048828125
],
[
0.5048828125,
0.50244140625,
0.890625,
0.50244140625,
0.890625,
0.5166015625,
0.5048828125,
0.5166015625
],
[
0.5048828125,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.5048828125,
0.53076171875
],
[
0.50341796875,
0.5322265625,
0.890625,
0.5322265625,
0.890625,
0.54541015625,
0.50341796875,
0.54541015625
],
[
0.5048828125,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.5048828125,
0.56103515625
],
[
0.50341796875,
0.56298828125,
0.89208984375,
0.56298828125,
0.89208984375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.50341796875,
0.57763671875,
0.6357421875,
0.57763671875,
0.6357421875,
0.59130859375,
0.50341796875,
0.59130859375
],
[
0.52294921875,
0.59326171875,
0.89208984375,
0.59228515625,
0.89208984375,
0.60693359375,
0.52294921875,
0.60791015625
],
[
0.080078125,
0.6015625,
0.4697265625,
0.6015625,
0.4697265625,
0.61572265625,
0.080078125,
0.61572265625
],
[
0.50244140625,
0.609375,
0.89208984375,
0.60791015625,
0.89208984375,
0.62255859375,
0.50244140625,
0.62353515625
],
[
0.080078125,
0.61572265625,
0.44677734375,
0.61572265625,
0.44677734375,
0.62890625,
0.080078125,
0.62890625
],
[
0.50244140625,
0.62255859375,
0.89208984375,
0.62353515625,
0.89208984375,
0.63818359375,
0.50244140625,
0.63720703125
],
[
0.08087158203125,
0.63818359375,
0.4697265625,
0.63818359375,
0.4697265625,
0.6513671875,
0.08087158203125,
0.6513671875
],
[
0.50341796875,
0.6396484375,
0.89208984375,
0.6396484375,
0.89208984375,
0.65283203125,
0.50341796875,
0.65283203125
],
[
0.5048828125,
0.65380859375,
0.89208984375,
0.65380859375,
0.89208984375,
0.66748046875,
0.5048828125,
0.66748046875
],
[
0.101318359375,
0.66748046875,
0.46728515625,
0.66748046875,
0.46728515625,
0.68115234375,
0.101318359375,
0.68115234375
],
[
0.50341796875,
0.66943359375,
0.890625,
0.66943359375,
0.890625,
0.68310546875,
0.50341796875,
0.68310546875
],
[
0.08087158203125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.08087158203125,
0.69775390625
],
[
0.50244140625,
0.6845703125,
0.890625,
0.68310546875,
0.890625,
0.69775390625,
0.50244140625,
0.69873046875
],
[
0.08087158203125,
0.69873046875,
0.468017578125,
0.69873046875,
0.468017578125,
0.7119140625,
0.08087158203125,
0.7119140625
],
[
0.50341796875,
0.69873046875,
0.890625,
0.69873046875,
0.890625,
0.71337890625,
0.50341796875,
0.71337890625
],
[
0.08087158203125,
0.7119140625,
0.468017578125,
0.71337890625,
0.468017578125,
0.72802734375,
0.08087158203125,
0.7265625
],
[
0.50244140625,
0.71337890625,
0.890625,
0.71337890625,
0.890625,
0.72802734375,
0.50244140625,
0.72802734375
],
[
0.08087158203125,
0.72900390625,
0.468017578125,
0.72900390625,
0.468017578125,
0.7421875,
0.08087158203125,
0.7421875
],
[
0.50244140625,
0.72802734375,
0.580078125,
0.72998046875,
0.580078125,
0.74560546875,
0.50244140625,
0.74365234375
],
[
0.08251953125,
0.74365234375,
0.468017578125,
0.74365234375,
0.468017578125,
0.7568359375,
0.08251953125,
0.7568359375
],
[
0.0849609375,
0.7587890625,
0.468017578125,
0.7587890625,
0.468017578125,
0.77294921875,
0.0849609375,
0.77294921875
],
[
0.50341796875,
0.7568359375,
0.7861328125,
0.7568359375,
0.7861328125,
0.77099609375,
0.50341796875,
0.77099609375
],
[
0.08087158203125,
0.7734375,
0.46728515625,
0.7734375,
0.46728515625,
0.787109375,
0.08087158203125,
0.787109375
],
[
0.52294921875,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79443359375,
0.52294921875,
0.79443359375
],
[
0.08087158203125,
0.7890625,
0.468017578125,
0.7890625,
0.468017578125,
0.80322265625,
0.08087158203125,
0.80322265625
],
[
0.50341796875,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.8095703125,
0.50341796875,
0.8095703125
],
[
0.08251953125,
0.8037109375,
0.468017578125,
0.8037109375,
0.468017578125,
0.8173828125,
0.08251953125,
0.8173828125
],
[
0.50341796875,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.08251953125,
0.8193359375,
0.468017578125,
0.8193359375,
0.468017578125,
0.83349609375,
0.08251953125,
0.83349609375
],
[
0.50244140625,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.08087158203125,
0.833984375,
0.468017578125,
0.83349609375,
0.468017578125,
0.84765625,
0.08087158203125,
0.8486328125
],
[
0.50341796875,
0.8388671875,
0.890625,
0.84033203125,
0.890625,
0.85498046875,
0.50341796875,
0.85400390625
],
[
0.08087158203125,
0.849609375,
0.375732421875,
0.849609375,
0.375732421875,
0.86279296875,
0.08087158203125,
0.86279296875
],
[
0.50341796875,
0.85498046875,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.50341796875,
0.869140625
],
[
0.1004638671875,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.1004638671875,
0.88525390625
],
[
0.52294921875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.52294921875,
0.88623046875
],
[
0.08087158203125,
0.88525390625,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.08087158203125,
0.8994140625
],
[
0.50341796875,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"Table 2. Point cloud completion on MVP dataset in terms of per-point L2 Chamfer distance ×104 (lower is better).",
"Methods",
"Cabinet",
"Chair",
"Lamp",
"Sofa",
"Table",
"Water",
"Bed",
"Bench",
"Shelf",
"Guitar",
"Motor",
"Pistol",
"Car",
"AveragePlane",
"Bus",
"Skate.",
"PCN[54]",
"8.92",
"6.49",
"12.46",
"19.54",
"9.92",
"12.45",
"8.78",
"19.0",
"13.39",
"5.15",
"1.87",
"6.03",
"6.04",
"4.70",
"9.0",
"9.80",
"4.22",
"13.46",
"84",
"13",
"9.71",
"10.34",
"7.36",
"11.21",
"12.46",
"8.50",
"18.98",
"8.58",
"15.15",
"2.13",
"7.19",
"4.15",
"4.09",
"TopNet [33]",
"20.53",
"2.59",
"8.86",
"6.54",
"10.22",
"12.64",
"9.08",
"9.69",
"7.08",
"15.58",
"6.38",
"11.31",
"1.37",
"4.63",
"3.06",
"7.98",
"MSN [22]",
"",
"4.33",
"4.47",
"",
"2.45",
"7.34",
"8.62",
"8.95",
"11.16",
"8.63",
"9.30",
"6.43",
"14.93",
"6.11",
"10.39",
"4.97",
"1.67",
"5.96",
"CRN30",
"7.31",
"5.15",
"2.17",
"VRCNet[29",
"7.83",
"8.29",
"7.42",
"7.07",
"11.18",
"4.76",
"7.03",
"4.40",
"1.15",
"3.75",
"3.54",
"PMPNet[17]",
"6.24",
"1.99",
"8.84",
"6.36",
"7.77",
"6.18",
"8.72",
"7.71",
"5.19",
"11.77",
"5.07",
"8.34",
"5.27",
"1.27",
"3.95",
"3.57",
"2.35",
"5.61",
"7.42",
"5.05",
"4.46",
"3.64",
"7.07",
"SnowflakeNet [19]",
"5.86",
"2.04",
"7.76",
"6.92",
"7.13",
"11.32",
"7.72",
"3.94",
"3.52",
"1.16",
"4.87",
"6.81",
"6.14",
"5.06",
"3DAttriFlow(Ours)",
"1.59",
"7.40",
"5.44",
"6.05",
"10.62",
"3.73",
"6.53",
"3.27",
"2.78",
"1.78",
"4.25",
"5.01",
"4.30",
"0.95",
"OccNet 3DAttriFlow",
"TopNet",
"VRCNet",
"R2N2",
"Input",
"PSGN",
"AtlasNet",
"Input",
"PMP-Net",
"lakeNet3DAttriFlov",
"",
"PCN",
"",
"n",
"真",
"",
"旺",
"",
"",
"",
"",
"",
"",
"",
"用",
"1",
"工",
"Figure 5. Visual comparison of point cloud completion re-",
"sults with different methods under MVP dataset.",
"特别是，PSGN[4]和AtlasNet[5]是基于点云的方法，",
"它们与3DAttriFlow最相关。然而，与这两种方法相",
"比，3DAttriFlow实现了超过25个%性能增益。正如",
"我们在SEC中讨论的那样。1上述两种方法采用了",
"2D到3D重建的典型范例，其中AtlasNet[5]基于全",
"局特征的隐式输入直接解码整个形状，而PSGN「4]则",
"利用编码器和解码器之间的特征通道来引入各种级别",
"的语义。这些实践都不能从图像中学习显式语义特征，",
"而只能尝试从编码器的隐式全局特征或中间层解码形",
"状。相比之下，3DAttriFlow可以利用从图像中学习的",
"隐式和显式语义属性，分别通过几何子管道和语义子",
"管道。因此，3DAttriFlow能够基于来自显式语义属性",
"的更明确的指导来预测3D形状的细节，并且比它的同",
"行获得更好的性能",
"定性比较。2D到3D重建的视觉比较如图4所",
"Figure 4.Visual comparison of 2D-to-3D reconstruction",
"示。请注意，对于AtlasNet，我们遵循其原始可视化",
"results with different methods under ShapeNet dataset.",
"设置来展示重建的网格而不是点云。与其他方法相比，",
"4.1.2D-to-3DReconstructionon ShapeNetdataset",
"3DAttriFlow在广泛的对象类别上重建了更好的细节。",
"例如，在椅子类别（图4的5th行）上，在OccNet的",
"数据集摘要和评估度量。我们按照OccNet[26]的实",
"预测中缺少腿，而PSGN和AtlasNet的椅子预测则模",
"验设置来评估ShapeNet数据集[]上的3DAttriFlow。",
"棱两可且充满噪音。至于飞机类别，PSGN和AtlasNet",
"整个数据集由43,783个网格对象组成，包含13个类",
"都无法重建1st和2nd行中发动机的详细形状，而",
"别，将按照OccNet[26]的相同策略分为训练，验证和",
"OccNet无法稳定地对发动机进行正确的预测（2nd行",
"测试。由于我们的方法侧重于从2D图像重建3D点",
"中的故障)",
"云，因此我们遵循AtlasNet[5]在3D对象的网格表面",
"上均匀采样30k点，作为训练的地面实况。遵循先前",
"4.2. 3D Completion on MVP dataset",
"的方法[26,37，我们使用由公式（6）描述的L1倒",
"数据集简介和评估度量。我们按照VRCNet[29]的",
"角距离作为评估指标。为了与其他重建3D体素或网",
"实验设置来评估我们在MVP数据集[29]上的3DAt-",
"格的方法进行比较，我们遵循OccNet[26】从其输出表",
"triFlow。该数据集由从ShapeNet中选择的模型生成的",
"面采样2,048个点，然后计算L1与地面真相的倒角距",
"16类不完整/完整点云组成，然后分为训练集（62,400",
"离。对于基于体素的方法，我们还将其体素输出传输到",
"个形状对）和测试集（41,600个形状对）。遵循以前的",
"网格中，然后在网格表面上应用点云采样",
"方法[29,33，47]，我们使用L2倒角距离作为评估指标",
"定量比较。2D到3D重建的结果如表1所示，其",
"定量比较。点云完成情况对比见表2。与当前最先进",
"中3DAttriFlow实现了优于其他比较对应方法的性能。",
"的完成方法SnowflakeNet【49]相比，3DAttriFlow将"
] |
{
"arxiv_id": "2203.15190",
"language": "zh",
"page_idx": 4
}
|
||
2310.12680-zh-0
|
2310.12680-zh-0.jpg
|
/burl@stx null def /BU.S /burl@stx null def def /BU.SS currentpoint /burl@lly exch def /burl@llx exch def burl@stx
# On the Optimization and Generalization of Multi-head Attention
Puneesh Deora$^1\dagger$ Rouzbeh Ghaderi$^1\dagger$ Hossein Taheri$^2\dagger$ Christos Thrampoulidis$^1$
$^{1}$University of British Columbia, Vancouver $^{2}$University of California, Santa Barbara
$\mathbf{Abstract}$
The training and generalization dynamics of the Transformer's core mechanism, namely the Attention mechanism, remain under-explored. Besides, existing analyses primarily focus on singlehead attention. Inspired by the demonstrated benefits of overparameterization when training fullyconnected networks, we investigate the potential optimization and generalization advantages of using multiple attention heads. Towards this goal, we derive convergence and generalization guarantees $\hat{\text{for gradient-descent training of a single-layer multi-head self-attention model, under a suitable realiz-}}$ ability condition on the data. We then establish primitive conditions on the initialization that ensure realizability holds. Finally, we demonstrate that these conditions are satisfied for a simple tokenizedmixture model. We expect the analysis can be extended to various data-model and architecture variations.
# 1 $\textbf{Introduction}$
变形金刚已经成为深度学习中一个很有前途的范例，这主要归功于它们独特的自我注意机制。受模型在自然语言处理 (Devlin et al., 2019; Brown et al., 2020; Raffel et al., 2020) 和计算机视觉 (Dosovitskiy et al., 2021; Radford et al., 2021; Touvron et al., 2021) 中的最新性能的推动，注意力机制的理论研究最近引起了人们的极大兴趣。许多研究已经探索了注意力的表达能力，例如。g。(Baldi and Vershynin, 2022; Dong et al., 2021; Yun et al., 2020a,b; Sanford et al., 2023; Bietti et al., 2023) , 最近在 (Baldi and Vershynin, 2022; Dong et al., 2021; Yun et al., 2020a,b; Mahdavi et al., 2023) 中研究了有关内存容量的初步发现。为了理解训练注意模型的优化方面，Sahiner et al. (2022); Ergen et al. (2022) 研究了凸松弛， 而 Tarzanagh et al.(2023) 研究了模型的隐式偏差。此外，Edelman et al.(2021) 还提出了基于容量和Rademacher 复杂度的自注意泛化范围。然而，对于训练注意力模型的梯度下降(GD)的有限时间优化和泛化动力学的探索在很大程度上仍然是一个悬而未决的问题
最近在这个方向上的贡献，作为我们工作的动力，包括研究 (Jelassi et al., 2022) , (Li et al., 2023a) 和 (Oymak et al., 2023) 。这三部作品集中于具有单个注意力头的单层注意力模型。此外，尽管对数据进行了必要的简化假设，但所有三种情况下的分析都相当复杂，并且似乎高度专注于个人注意力和数据模型。我们将这些作品的更多细节推迟到 7 节。这些直接和高度专业化的分析提出了一定的挑战。首先， 仍然不确定它们是否可以包含在更广泛的框架中，该框架可以扩展到更复杂的注意力体系结构和多样化的数据模型。其次，它们似乎与近年来蓬勃发展的传统架构 (如全连接和卷积神经网络) 的现有框架脱节。。g。,(Jacot et al., 2018; Ji and Telgarsky, 2020; Richards and Rabbat, 2021; Banerjee et al., 2022;mis $=\frac{1}{2}$ mans $=\frac{1}{2}$ mans pons puris pond Telis and ,Richards and Rabbat, 2021; Banerjee et al., 2022; Taheri and Thrampoulidis, 2023) 。因此，还不清楚注意力的引入如何改变分析格局。。在这项工作中， 我们研究了通过梯度方法训练的多头注意机制的优化和泛化特性。我们的方法特别利用了多个注意力的利用。尽管注意模型中的注意头和 MLP 中的隐藏节点之间存在操作差异，但我们从分析的角度证明， 这种并行性使得能够利用为后者开发的框架来研究前者。特别是对于泛化分析，我们利用算法稳定性框架在超参数化 MLP (Richards and Kuzborskij, 2021; Taheri and Thrampoulidis, 2023) 中的应用方面的最新进展贡献。具体来说，我们的贡献如下。我们研究了在二进制分类设置中具有$H$头的多头注意力(MHA) 层的梯度下降优化的训练和泛化。对于此设置，在 2 节中进行了详细说明，我们分析了具有逻辑损失的训练，包括注意力权重(参数化 softmaxlogits)以及将输出令牌转换为标签预测的线性解码器
$\overline{^\dagger\text{These authors contributed equally. Alphabetical ordering used.}}$
Emails: $\{$ puneeshdeora, rghaderi , cthrampo $\}$ @ece.ubc.ca, hossein@ucsb.edu. This work is supported by NSERC Discovery Grant RGPIN-2021-03677 and NSF Grant CCF-2009030. The authors also acknowledge use of the Sockeye cluster by ÜBC Advanced Research Computing. PD thanks Bhavya Vasudeva for the helpful discussions.
1
|
[
[
0.1234130859375,
0.07659912109375,
0.99658203125,
0.07659912109375,
0.99658203125,
0.0872802734375,
0.1234130859375,
0.0872802734375
],
[
0.264404296875,
0.133056640625,
0.73046875,
0.133056640625,
0.73046875,
0.1478271484375,
0.264404296875,
0.1478271484375
],
[
0.359375,
0.157958984375,
0.63623046875,
0.1602783203125,
0.63623046875,
0.174560546875,
0.359375,
0.1728515625
],
[
0.1318359375,
0.19482421875,
0.85791015625,
0.19482421875,
0.85791015625,
0.2120361328125,
0.1318359375,
0.2120361328125
],
[
0.310546875,
0.21728515625,
0.68603515625,
0.2197265625,
0.68603515625,
0.2340087890625,
0.310546875,
0.232177734375
],
[
0.32421875,
0.2340087890625,
0.67236328125,
0.235107421875,
0.67236328125,
0.2493896484375,
0.32421875,
0.248779296875
],
[
0.466796875,
0.292236328125,
0.5341796875,
0.292236328125,
0.5341796875,
0.303955078125,
0.466796875,
0.303955078125
],
[
0.183837890625,
0.30810546875,
0.8388671875,
0.30810546875,
0.8388671875,
0.32177734375,
0.183837890625,
0.32177734375
],
[
0.162109375,
0.32177734375,
0.84033203125,
0.32177734375,
0.84033203125,
0.3349609375,
0.162109375,
0.3349609375
],
[
0.160400390625,
0.3349609375,
0.8388671875,
0.3349609375,
0.8388671875,
0.3486328125,
0.160400390625,
0.3486328125
],
[
0.162109375,
0.347412109375,
0.8388671875,
0.347412109375,
0.8388671875,
0.361083984375,
0.162109375,
0.361083984375
],
[
0.162109375,
0.359375,
0.83447265625,
0.359375,
0.83447265625,
0.372314453125,
0.162109375,
0.372314453125
],
[
0.160400390625,
0.372314453125,
0.8388671875,
0.372314453125,
0.8388671875,
0.385986328125,
0.160400390625,
0.385986328125
],
[
0.160400390625,
0.385498046875,
0.8369140625,
0.385498046875,
0.8369140625,
0.3984375,
0.160400390625,
0.3984375
],
[
0.162841796875,
0.40087890625,
0.8369140625,
0.40087890625,
0.8369140625,
0.410888671875,
0.162841796875,
0.410888671875
],
[
0.162841796875,
0.41455078125,
0.83642578125,
0.41455078125,
0.83642578125,
0.424560546875,
0.162841796875,
0.424560546875
],
[
0.160400390625,
0.424560546875,
0.231689453125,
0.427001953125,
0.231689453125,
0.438232421875,
0.160400390625,
0.435791015625
],
[
0.1175537109375,
0.46728515625,
0.31982421875,
0.46728515625,
0.31982421875,
0.482177734375,
0.1175537109375,
0.482177734375
],
[
0.12005615234375,
0.49755859375,
0.87744140625,
0.49755859375,
0.87744140625,
0.51123046875,
0.12005615234375,
0.51123046875
],
[
0.12005615234375,
0.5126953125,
0.87744140625,
0.5126953125,
0.87744140625,
0.525390625,
0.12005615234375,
0.525390625
],
[
0.12005615234375,
0.52685546875,
0.87890625,
0.52685546875,
0.87890625,
0.54052734375,
0.12005615234375,
0.54052734375
],
[
0.12005615234375,
0.54150390625,
0.87158203125,
0.54150390625,
0.87158203125,
0.5546875,
0.12005615234375,
0.5546875
],
[
0.12005615234375,
0.5546875,
0.87890625,
0.5546875,
0.87890625,
0.568359375,
0.12005615234375,
0.568359375
],
[
0.12005615234375,
0.5693359375,
0.87744140625,
0.5693359375,
0.87744140625,
0.5830078125,
0.12005615234375,
0.5830078125
],
[
0.12005615234375,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.118408203125,
0.595703125,
0.87890625,
0.5966796875,
0.87890625,
0.61083984375,
0.118408203125,
0.60986328125
],
[
0.12005615234375,
0.61328125,
0.87890625,
0.61328125,
0.87890625,
0.62353515625,
0.12005615234375,
0.62353515625
],
[
0.12176513671875,
0.62744140625,
0.57080078125,
0.62744140625,
0.57080078125,
0.63818359375,
0.12176513671875,
0.63818359375
],
[
0.1435546875,
0.6396484375,
0.87744140625,
0.6396484375,
0.87744140625,
0.65380859375,
0.1435546875,
0.65380859375
],
[
0.12005615234375,
0.65283203125,
0.8759765625,
0.65283203125,
0.8759765625,
0.66650390625,
0.12005615234375,
0.66650390625
],
[
0.12005615234375,
0.66845703125,
0.87890625,
0.66845703125,
0.87890625,
0.681640625,
0.12005615234375,
0.681640625
],
[
0.118408203125,
0.681640625,
0.87744140625,
0.68310546875,
0.87744140625,
0.69677734375,
0.118408203125,
0.6953125
],
[
0.12176513671875,
0.69873046875,
0.87744140625,
0.69873046875,
0.87744140625,
0.708984375,
0.12176513671875,
0.708984375
],
[
0.12176513671875,
0.71337890625,
0.87744140625,
0.71337890625,
0.87744140625,
0.7236328125,
0.12176513671875,
0.7236328125
],
[
0.118408203125,
0.7255859375,
0.87890625,
0.7236328125,
0.87890625,
0.73828125,
0.118408203125,
0.74072265625
],
[
0.12005615234375,
0.7392578125,
0.87744140625,
0.7392578125,
0.87744140625,
0.7529296875,
0.12005615234375,
0.7529296875
],
[
0.12005615234375,
0.7529296875,
0.8759765625,
0.7529296875,
0.8759765625,
0.7666015625,
0.12005615234375,
0.7666015625
],
[
0.12176513671875,
0.76953125,
0.87890625,
0.76953125,
0.87890625,
0.77978515625,
0.12176513671875,
0.77978515625
],
[
0.12176513671875,
0.78369140625,
0.87744140625,
0.78369140625,
0.87744140625,
0.79443359375,
0.12176513671875,
0.79443359375
],
[
0.118408203125,
0.7958984375,
0.87890625,
0.79638671875,
0.87890625,
0.81005859375,
0.118408203125,
0.80859375
],
[
0.12005615234375,
0.81005859375,
0.87744140625,
0.81005859375,
0.87744140625,
0.82373046875,
0.12005615234375,
0.82373046875
],
[
0.1175537109375,
0.82373046875,
0.87744140625,
0.82470703125,
0.87744140625,
0.83837890625,
0.1175537109375,
0.83740234375
],
[
0.12005615234375,
0.8388671875,
0.8759765625,
0.8388671875,
0.8759765625,
0.8525390625,
0.12005615234375,
0.8525390625
],
[
0.12005615234375,
0.85791015625,
0.52294921875,
0.859375,
0.52294921875,
0.87353515625,
0.12005615234375,
0.87255859375
],
[
0.12005615234375,
0.87255859375,
0.87890625,
0.87255859375,
0.87890625,
0.8857421875,
0.12005615234375,
0.8857421875
],
[
0.12176513671875,
0.88525390625,
0.8759765625,
0.88525390625,
0.8759765625,
0.8955078125,
0.12176513671875,
0.8955078125
],
[
0.12176513671875,
0.896484375,
0.82861328125,
0.896484375,
0.82861328125,
0.90673828125,
0.12176513671875,
0.90673828125
],
[
0.48779296875,
0.939453125,
0.5029296875,
0.93505859375,
0.5087890625,
0.9453125,
0.49365234375,
0.94970703125
]
] |
[
"/burl@stxnulldef/BU.S/burl@stxnulldefdef/BU.SScurrentpoint/burl@llyexchdef/burl@llxexchdefburl@stx",
"On the Optimization and Generalization",
"of Multi-head Attention",
"Puneesh Deoralt Rouzbeh Ghaderiit Hossein Taheri2t Christos Thrampoulidisl",
"1UniversityofBritishColumbia,Vancouver",
"2University of California, Santa Barbara",
"Abstract",
"The training and generalization dynamics of the Transformer's core mechanism, namely the",
"Attention mechanism, remain under-explored. Besides, existing analyses primarily focus on single-",
"head attention. Inspired by the demonstrated benefits of overparameterization when training fully-",
"connected networks, we investigate the potential optimization and generalization advantages of using",
"multiple attention heads. Towards this goal, we derive convergence and generalization guarantee",
"for gradient-descent training of a single-layer multi-head self-attention model, under a suitable realiz-",
"ability condition on the data. We then establish primitive conditions on the initialization that ensure",
"realizabilityholds.Finally,we demonstratethatthese conditions aresatisfied for a simpletokenized-",
"mixture model.We expect the analysis can be extended to various data-model and architecture",
"variations.",
"1Introduction",
"变形金刚已经成为深度学习中一个很有前途的范例，这主要归功于它们独特的自我注意机制。受模型",
"在自然语言处理（Devlin etal.,2019;Brown etal.，2020;Raffel et al.,2020)和计算机视觉(Dosovitskiy",
"etal.，2021；Radford etal.，2021；Touvronetal.，2021）中的最新性能的推动，注意力机制的理论研究",
"最近引起了人们的极大兴趣。许多研究已经探索了注意力的表达能力，例如。g。（BaldiandVershynin",
"2022; Dong et al., 2021; Yun et al., 2020a,b; Sanford et al., 2023; Bietti et al., 2023) , 最近在 (Baldi and",
"Vershynin,2022;Dong et al.,2021;Yun et al.,2020a,b;Mahdavi et al.,2023)中研究了有关内存容量的",
"初步发现。为了理解训练注意模型的优化方面，Sahineretal.（2022)；Ergenetal.（2022）研究了凸松弛，",
"而Tarzanagh etal.（2023）研究了模型的隐式偏差。此外，Edelman etal.（2021）还提出了基于容量和",
"Rademacher复杂度的自注意泛化范围。然而，对于训练注意力模型的梯度下降（GD）的有限时间优化",
"和泛化动力学的探索在很大程度上仍然是一个悬而未决的问题",
"最近在这个方向上的贡献，作为我们工作的动力，包括研究（Jelassietal.，2022），（Lietal.，2023a)",
"和（Oymaketal.，2023）。这三部作品集中于具有单个注意力头的单层注意力模型。此外，尽管对数据",
"进行了必要的简化假设，但所有三种情况下的分析都相当复杂，并且似乎高度专注于个人注意力和数据",
"模型。我们将这些作品的更多细节推迟到7节。这些直接和高度专业化的分析提出了一定的挑战。首先，",
"仍然不确定它们是否可以包含在更广泛的框架中，该框架可以扩展到更复杂的注意力体系结构和多样化",
"的数据模型。其次，它们似乎与近年来蓬勃发展的传统架构（如全连接和卷积神经网络）的现有框架脱",
"节。。 g。, (Jacot et al., 2018; Ji and Telgarsky, 2020; Richards and Rabbat, 2021; Banerjee et al., 2022;",
"TaheriandThrampoulidis，2023）。因此，还不清楚注意力的引人如何改变分析格局。。在这项工作中",
"我们研究了通过梯度方法训练的多头注意机制的优化和泛化特性。我们的方法特别利用了多个注意力的",
"利用。尽管注意模型中的注意头和MLP中的隐藏节点之间存在操作差异，但我们从分析的角度证明：",
"这种并行性使得能够利用为后者开发的框架来研究前者。特别是对于泛化分析，我们利用算法稳定性框",
"架在超参数化MLP（RichardsandKuzborskij，2021；TaheriandThrampoulidis，2023）中的应用方面的",
"最新进展贡献。具体来说，我们的贡献如下。我们研究了在二进制分类设置中具有H头的多头注意力",
"（MHA）层的梯度下降优化的训练和泛化。对于此设置，在2节中进行了详细说明，我们分析了具有逻",
"辑损失的训练，包括注意力权重（参数化softmaxlogits）以及将输出令牌转换为标签预测的线性解码器",
"tThese authors contributed equally. Alphabetical ordering used.",
"Emails:{ puneeshdeora, rghaderi, cthrampo } @ece.ubc.ca, hossein@ucsb.edu. This work is supported",
"byNSERCDiscoveryGrantRGPIN-2021-03677 andNSFGrantCCF-2009030.Theauthorsalsoacknowledgeuseofthe",
"Sockeye cluster by UBCAdvanced Research Computing-PD thanks Bhavya Vasudeva forthe helpful discussions",
"1"
] |
{
"arxiv_id": "2310.12680",
"language": "zh",
"page_idx": 0
}
|
||
2310.12680-zh-1
|
2310.12680-zh-1.jpg
|
首先，在 3 节中，我们描述了经验损失$\widehat{L}$的关键属性，特别是确定它是自有界的并且满足关键的自有界弱凸属性，i。e。$\lambda_{\min}(\nabla^2\widehat{L}(\boldsymbol{\theta}))\gtrsim-\frac\kappa{\sqrt{H}}\widehat{L}(\boldsymbol{\theta})$,用于仅轻度依赖于参数向量$\boldsymbol{\theta}$的参数$\kappa$ 。建立这些属性(以及量化$\kappa$ )涉及仔细计算和限制 MHA 层的梯度和 Hessian, 这些计算可能超出我们论文的范围
其次，在 4.1-4.2 节中，我们以最一般的形式介绍了我们的训练和泛化边界。根据经验损失$\widehat{L}(\boldsymbol{\theta})$和距离$\left\|\boldsymbol{\theta}-\boldsymbol{\theta}_0\right\|$到适当选择的目标向量 $\boldsymbol{\theta}$ 的初始化 $\boldsymbol{\theta}_0$ 给出界限。到初始化的距离还控制边界保持所需的最小头数$H\gtrsim\|\boldsymbol{\theta}-\boldsymbol{\theta}_0\|^6$ 。选择使边界紧密的适当参数$\boldsymbol{\theta}$通常特定于数据设置和所选的初始化。为了指导这样的选择，在 4.3 节中，我们将初始化$\boldsymbol{\theta}_{0}$上的原始和直接检查条件形式化，以确保可以找到适当的$\boldsymbol{\theta}$。简而言之，假设初始化时的模型输出在训练集大小$n$上是对数的，并且数据相对于具有恒定裕度$\gamma$ 的 MHA 模型的神经切线核 (NTK) 特征是可分离的，则推论 2 表明，对于步长$\eta=\widetilde{O}(1)$和$\|\boldsymbol{\theta}-\boldsymbol{\theta}_0\|$ 的对数，仅由梯度 XMATHX 下降步长提供。。我们注意到，上述 NTK 可分离性假设虽然与相关，但与标准 NTK 分析不同。此外，虽然这个假设足以应用我们的一般界限，但它不是必要条件
在 5 节中，我们研究了 MHA 在具有标签- (ir) 相关令牌的令牌化混合数据模型上的性能。对于此模型，我们表明，从零初始化开始经过一个随机梯度步骤后，MHA 模型的 NTK 特征将数据与边距$\gamma_{\star}$ 分开。因此，应用 4.1 节中的一般分析，我们建立了如上所述的训练和泛化边界，用于对数数量的头部。为了评估这些边界的最优性，我们证明了 MHA 具有足够的表现力，可以实现优于$\gamma_\star$的边界$\gamma_\mathrm{attn}$ 。达到$\gamma_\mathrm{attn}$的机制涉及选择足够大的范数的键查询权重，这通过抑制与标签无关的令牌来使 softmax 非线性饱和。我们将大型规范要求确定为在我们的理论框架中选择这些权重作为目标参数的潜在瓶颈，并讨论了有关将分析框架扩展到该特定制度的开放性问题
纸张的剩余部分组织如下。我们的主要训练和泛化边界的证明草图在 6 节中给出。在 7 节中进一步讨论了最密切相关的工作。本文在 8 节中总结了我们的发现的含义以及它们引起的开放性问题。。附录中包含详细的证明，我们还提供了对合成数据进行的实验的数值结果
# 2 $\textbf{Preliminaries}$
表示法。$\varphi(\cdot):\mathbb{R}^T\to\mathbb{R}^T$表示 softmax 映射，$\varphi^\prime(\boldsymbol{v}):=\nabla\boldsymbol{\varphi}(\boldsymbol{v})=\operatorname{diag}(\boldsymbol{\varphi}(\boldsymbol{v}))-\boldsymbol{v}\boldsymbol{v}^\intercal$表示$\boldsymbol{v}\in\mathbb{R}^T$处的梯度。对于$t\in[T]$ , 我们为$\boldsymbol{\varphi}(\boldsymbol{v})\in\mathbb{R}^T$的第$t$条目编写$\boldsymbol{\varphi}_t(\boldsymbol{v})$ 。对于$\boldsymbol{A}\in\mathbb{R}^n\times m$ ,我们将$A_i$,表示为第$i$行，将$\boldsymbol{A}_:,j$表示为第$j$列。回想一下诱导矩阵范数符号$\|\boldsymbol{A}\|_p,q=\max_{\|\boldsymbol{v}\|_p=1}\|\boldsymbol{A}\boldsymbol{v}\|_q$ ,尤其是以下内容：$\|\boldsymbol{A}\|_{2,\infty}=\max_{i\in[n]}\|\boldsymbol{A}_{i,:}\|,\|\boldsymbol{A}\|_{1,2}=\max_{j\in[m]}\|\boldsymbol{A}_{:,j}\|$,和$\|\boldsymbol{A}\|_{1,\infty}=\max_{j\in[m]}\|\boldsymbol{A}_{:,j}\|\boldsymbol{A}_{:,j}\|_{\infty}$。为简单起见$\|A\|,\|\boldsymbol{v}\|$表示欧几里得范数，$\lambda_\min(\boldsymbol{A})$表示最小特征值。我们让$a\wedge b=\min\{a,b\}$和$a\vee b=\max_{\:\prime}\{a,b\}$ 。我们用 concat 表示向量级联运算。所有对数都是自然对数(以$e$为基数)。我们将$w_1,\boldsymbol{w}_2\in\mathbb{R}^{d^{\prime}}$之间的线段表示为$[\boldsymbol{w}_1,\boldsymbol{w}_2]=\{\boldsymbol{w}:\boldsymbol{w}=\alpha\boldsymbol{w}_1+(1-\alpha)\boldsymbol{w}_2,\alpha\in[0,1]\}$ 。最后 , 为了简化说明 ,在主体中使用隐藏绝对常数的“》”或“s”表示法表示某些不等式。我们也偶尔使用标准符号$\mathcal{O},\Omega$和$\widetilde{\mathcal{O}},\widetilde{\Omega}$来隐藏多对数因子。除非另有说明，否则这些顺序符号相对于训练集大小$n$ 。无论何时使用，都在附录中指定精确的常量
单头自我关注。由键、查询和值矩阵$W_Q,W_K\in\mathbb{R}^{dxd_h}$ and$W_V\in\mathbb{R}^{dxd_v}$参数化的具有上下文大小$T$和
维度$d$的单层自关注头 ATTN$:\mathbb{R}^T\times d\to\mathbb{R}^{\breve{T}\times d}$由下式给出：
$$\mathrm{ATTN}(\boldsymbol{X};\boldsymbol{W}_{Q},\boldsymbol{W}_{K},\boldsymbol{W}_{V}):=\boldsymbol{\varphi}(\boldsymbol{X}\boldsymbol{W}_{Q}\boldsymbol{W}_{K}^{\intercal}\boldsymbol{X}^{\intercal})\boldsymbol{X}\boldsymbol{W}_{V}\:.$$
这里，$X=[x_1,x_2,\ldots,x_T]^{\intercal}\in\mathbb{R}^{T\times d}$是输入令牌矩阵，$\varphi(\boldsymbol{XW}_Q\boldsymbol{W}_K^{\intercal}\boldsymbol{X}^{\intercal})\in\mathbb{R}^{T\times T}$是关注矩阵，$\boldsymbol{\varphi}(\cdot):\mathbb{R}^T\to$ $\mathbb{R}^T$是。应用于矩阵的 Softmax 按行操作。要在预测标签中转换 Attention 输出，我们使用线性投影头( 又名解码器)组成 ATTN 。因此，模型的输出是$^{1}$
$$\Phi(\boldsymbol{X};\boldsymbol{W},\boldsymbol{U}):=\langle\boldsymbol{U},\boldsymbol{\varphi}(\boldsymbol{X}\boldsymbol{W}\boldsymbol{X}^{\intercal})\boldsymbol{X}\rangle\:.$$
(1)
注意，我们将权重矩阵$\boldsymbol{W}_V$的值吸收到投影仪$U=[\boldsymbol{u}_1,\ldots,\boldsymbol{u}_T]^{\intercal}\in\mathbb{R}^{T\times d}$中。此外，我们将整个密钥查
询乘积矩阵参数化为$\boldsymbol{W}:=\boldsymbol{W}_Q\boldsymbol{W}_K^\top.$
多头自我关注。我们的重点是带有$^\mathrm{\Lambda}H$ 头的多头注意力 (MHA) 模型。对于输出矩阵 $W_{Oh}\in\mathbb{R}^{d_v\times d}$,
MHA 模型的输出令牌由
$$\sum_{h\in[H]}\mathrm{ATTN}(\boldsymbol{X};\boldsymbol{W}_{Qh},\boldsymbol{W}_{Kh},\boldsymbol{W}_{Vh})\boldsymbol{W}_{Oh},$$
给出。与单头关注类似，我们将这些令牌投影到标量预测值。将$W_{Vh}W_{Oh}$ 吸收到投影层中并对$\overline{^{\dagger}\text{While we focus on (i) Full-projection: trainable}}$ matrix $\boldsymbol{U}\in\mathbb{R}^T\times d$, our results also apply to (ii) Pooling: $\boldsymbol{U}=\boldsymbol{u}1_T^{\intercal}$ with
trainable $\boldsymbol{u}\in\mathbb{R}^d$, and (iii) Last-token output: $\boldsymbol{U}=\begin{bmatrix}0_{d\times(T-1)}&\boldsymbol{u}\end{bmatrix}^{\dagger}$ with trainable $\boldsymbol{u}\in\mathbb{R}^d$
2
|
[
[
0.145263671875,
0.08612060546875,
0.875,
0.08612060546875,
0.875,
0.09979248046875,
0.145263671875,
0.09979248046875
],
[
0.12005615234375,
0.1015625,
0.87744140625,
0.1015625,
0.87744140625,
0.11639404296875,
0.12005615234375,
0.11639404296875
],
[
0.12005615234375,
0.1181640625,
0.873046875,
0.1181640625,
0.873046875,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.145263671875,
0.1318359375,
0.87744140625,
0.1318359375,
0.87744140625,
0.1455078125,
0.145263671875,
0.1455078125
],
[
0.12005615234375,
0.146728515625,
0.87744140625,
0.146728515625,
0.87744140625,
0.1602783203125,
0.12005615234375,
0.1602783203125
],
[
0.118408203125,
0.1602783203125,
0.87744140625,
0.160888671875,
0.87744140625,
0.174560546875,
0.118408203125,
0.17333984375
],
[
0.1234130859375,
0.177001953125,
0.8759765625,
0.177001953125,
0.8759765625,
0.18701171875,
0.1234130859375,
0.18701171875
],
[
0.12176513671875,
0.188232421875,
0.875,
0.188232421875,
0.875,
0.201904296875,
0.12176513671875,
0.201904296875
],
[
0.12005615234375,
0.2037353515625,
0.875,
0.2037353515625,
0.875,
0.21728515625,
0.12005615234375,
0.21728515625
],
[
0.12005615234375,
0.218505859375,
0.87890625,
0.218505859375,
0.87890625,
0.232177734375,
0.12005615234375,
0.232177734375
],
[
0.118408203125,
0.23095703125,
0.771484375,
0.232177734375,
0.771484375,
0.2452392578125,
0.118408203125,
0.24462890625
],
[
0.1435546875,
0.2464599609375,
0.880859375,
0.2464599609375,
0.880859375,
0.260009765625,
0.1435546875,
0.260009765625
],
[
0.118408203125,
0.260009765625,
0.87744140625,
0.26123046875,
0.87744140625,
0.27490234375,
0.118408203125,
0.273681640625
],
[
0.1234130859375,
0.276611328125,
0.8759765625,
0.276611328125,
0.8759765625,
0.287353515625,
0.1234130859375,
0.287353515625
],
[
0.12176513671875,
0.289306640625,
0.87890625,
0.289306640625,
0.87890625,
0.303955078125,
0.12176513671875,
0.303955078125
],
[
0.12005615234375,
0.303955078125,
0.87744140625,
0.303955078125,
0.87744140625,
0.317626953125,
0.12005615234375,
0.317626953125
],
[
0.1234130859375,
0.319580078125,
0.8759765625,
0.319580078125,
0.8759765625,
0.330078125,
0.1234130859375,
0.330078125
],
[
0.1234130859375,
0.3349609375,
0.5029296875,
0.3349609375,
0.5029296875,
0.342529296875,
0.1234130859375,
0.342529296875
],
[
0.14697265625,
0.347412109375,
0.8759765625,
0.347412109375,
0.8759765625,
0.358154296875,
0.14697265625,
0.358154296875
],
[
0.12005615234375,
0.35986328125,
0.87890625,
0.35986328125,
0.87890625,
0.37353515625,
0.12005615234375,
0.37353515625
],
[
0.1234130859375,
0.3759765625,
0.623046875,
0.3759765625,
0.623046875,
0.385986328125,
0.1234130859375,
0.385986328125
],
[
0.118408203125,
0.410888671875,
0.32568359375,
0.410888671875,
0.32568359375,
0.42578125,
0.118408203125,
0.42578125
],
[
0.1175537109375,
0.440673828125,
0.87890625,
0.439453125,
0.87890625,
0.457275390625,
0.1175537109375,
0.457763671875
],
[
0.12005615234375,
0.457275390625,
0.8759765625,
0.457275390625,
0.8759765625,
0.47021484375,
0.12005615234375,
0.47021484375
],
[
0.12005615234375,
0.471435546875,
0.87744140625,
0.471435546875,
0.87744140625,
0.486328125,
0.12005615234375,
0.486328125
],
[
0.12176513671875,
0.486328125,
0.880859375,
0.486328125,
0.880859375,
0.50048828125,
0.12176513671875,
0.50048828125
],
[
0.12005615234375,
0.49951171875,
0.8759765625,
0.49951171875,
0.8759765625,
0.51318359375,
0.12005615234375,
0.51318359375
],
[
0.1368408203125,
0.515625,
0.87744140625,
0.515625,
0.87744140625,
0.529296875,
0.1368408203125,
0.529296875
],
[
0.12005615234375,
0.52978515625,
0.87744140625,
0.52978515625,
0.87744140625,
0.54345703125,
0.12005615234375,
0.54345703125
],
[
0.12176513671875,
0.5419921875,
0.8759765625,
0.5419921875,
0.8759765625,
0.5556640625,
0.12176513671875,
0.5556640625
],
[
0.12005615234375,
0.55810546875,
0.87890625,
0.55810546875,
0.87890625,
0.57177734375,
0.12005615234375,
0.57177734375
],
[
0.12176513671875,
0.57373046875,
0.189697265625,
0.57373046875,
0.189697265625,
0.58447265625,
0.12176513671875,
0.58447265625
],
[
0.12176513671875,
0.59130859375,
0.87890625,
0.59130859375,
0.87890625,
0.60546875,
0.12176513671875,
0.60546875
],
[
0.12005615234375,
0.60498046875,
0.54833984375,
0.60498046875,
0.54833984375,
0.619140625,
0.12005615234375,
0.619140625
],
[
0.300537109375,
0.6318359375,
0.69873046875,
0.6318359375,
0.69873046875,
0.646484375,
0.300537109375,
0.646484375
],
[
0.12005615234375,
0.6591796875,
0.26025390625,
0.66015625,
0.26025390625,
0.6728515625,
0.12005615234375,
0.67138671875
],
[
0.26611328125,
0.65673828125,
0.87890625,
0.65673828125,
0.87890625,
0.6748046875,
0.26611328125,
0.6748046875
],
[
0.118408203125,
0.6728515625,
0.87890625,
0.67333984375,
0.87890625,
0.6884765625,
0.118408203125,
0.68701171875
],
[
0.118408203125,
0.6884765625,
0.497802734375,
0.6884765625,
0.497802734375,
0.70166015625,
0.118408203125,
0.70166015625
],
[
0.364501953125,
0.71435546875,
0.63232421875,
0.71435546875,
0.63232421875,
0.72802734375,
0.364501953125,
0.72802734375
],
[
0.85546875,
0.71435546875,
0.880859375,
0.71435546875,
0.880859375,
0.728515625,
0.85546875,
0.728515625
],
[
0.12176513671875,
0.7412109375,
0.87890625,
0.7412109375,
0.87890625,
0.755859375,
0.12176513671875,
0.755859375
],
[
0.118408203125,
0.75439453125,
0.38623046875,
0.7548828125,
0.38623046875,
0.76953125,
0.118408203125,
0.7685546875
],
[
0.12176513671875,
0.76953125,
0.880859375,
0.76953125,
0.880859375,
0.78369140625,
0.12176513671875,
0.78369140625
],
[
0.118408203125,
0.783203125,
0.300537109375,
0.78515625,
0.300537109375,
0.798828125,
0.118408203125,
0.79638671875
],
[
0.35693359375,
0.79638671875,
0.64990234375,
0.7998046875,
0.64990234375,
0.81396484375,
0.35693359375,
0.81103515625
],
[
0.349365234375,
0.81103515625,
0.38623046875,
0.81103515625,
0.38623046875,
0.82666015625,
0.349365234375,
0.82666015625
],
[
0.12176513671875,
0.83447265625,
0.8759765625,
0.83447265625,
0.8759765625,
0.84423828125,
0.12176513671875,
0.84423828125
],
[
0.1175537109375,
0.87060546875,
0.87890625,
0.87255859375,
0.87890625,
0.89013671875,
0.1175537109375,
0.88818359375
],
[
0.1234130859375,
0.88818359375,
0.53662109375,
0.89111328125,
0.53662109375,
0.9091796875,
0.1234130859375,
0.90625
],
[
0.5322265625,
0.892578125,
0.6962890625,
0.892578125,
0.6962890625,
0.90673828125,
0.5322265625,
0.90673828125
],
[
0.489501953125,
0.93603515625,
0.5087890625,
0.93603515625,
0.5087890625,
0.94970703125,
0.489501953125,
0.94970703125
]
] |
[
"首先，在3节中，我们描述了经验损失工的关键属性，特别是确定它是自有界的并且满足关键的自",
"有界弱凸属性，i。e。^min(2L(の))≥-L(の），用于仅轻度依赖于参数向量的参数。建立这些属",
"性（以及量化K）涉及仔细计算和限制MHA层的梯度和Hessian，这些计算可能超出我们论文的范围",
"其次，在4.1－4.2节中，我们以最一般的形式介绍了我们的训练和泛化边界。根据经验损失L0）和",
"距离α-9ol到适当选择的目标向量的初始化o给出界限。到初始化的距离还控制边界保持所需的",
"最小头数H≥-6。选择使边界紧密的适当参数θ通常特定于数据设置和所选的初始化。为了指",
"导这样的选择，在4.3节中，我们将初始化0上的原始和直接检查条件形式化，以确保可以找到适当的",
"9。简而言之，假设初始化时的模型输出在训练集大小n上是对数的，并且数据相对于具有恒定裕度",
"的MHA模型的神经切线核（NTK）特征是可分离的，则推论2表明，对于步长n=O(1和θ-6o",
"的对数，仅由梯度XMATHX下降步长提供。。我们注意到，上述NTK可分离性假设虽然与相关，但与",
"标准NTK分析不同。此外，虽然这个假设足以应用我们的一般界限，但它不是必要条件",
"在5节中，我们研究了MHA在具有标签-（ir）相关令牌的令牌化混合数据模型上的性能。对于此",
"模型，我们表明，从零初始化开始经过一个随机梯度步骤后，MHA模型的NTK特征将数据与边距",
"分开。因此，应用4.1节中的一般分析，我们建立了如上所述的训练和泛化边界，用于对数数量的头部，",
"为了评估这些边界的最优性，我们证明了MHA具有足够的表现力，可以实现优于的边界attn。达",
"到attn的机制涉及选择足够大的范数的键查询权重，这通过抑制与标签无关的令牌来使softmax非线",
"性饱和。我们将大型规范要求确定为在我们的理论框架中选择这些权重作为目标参数的潜在瓶颈，并讨",
"企有关将分析框架扩展到该特定制度的升放性题",
"纸张的剩余部分组织如下。我们的主要训练和泛化边界的证明草图在6节中给出。在7节中进一步",
"讨论了最密切相关的工作。本文在8节中总结了我们的发现的含义以及它们引起的开放性问题。。附录",
"中包含详细的证明，我们还提供了对合成数据进行的实验的数值结果",
"2Preliminaries",
"表示法。 p():RT→ RT表示 softmax映射，():=Vp(u)=diag(sp(u))-uT表示 e RT处的",
"梯度。对于tε[T]，我们为u)eRT的第t条目编写Pt(u。对于AeRnxm，我们将Ai：表示为",
"第i行，将A.表示为第j列。回想一下诱导矩阵范数符号Alp,q=max=iAug，尤其是以下",
"内容：A2,o=maxie[n]A,:ll,A|1,2=maxje[m]A,il,和A1,=maxje[m]|A.,il。为简单起见,",
"A,ull表示欧几里得范数，入min(A）表示最小特征值。我们让a^b=min(a,b)和avb=maxa,b)",
"我们用concat表示向量级联运算。所有对数都是自然对数（以e为基数）。我们将W1，w2Rd之间",
"的线段表示为[w1,w2]=(w：w=αw1+(1-α)w2,αe[0,1]。最后，为了简化说明，在主体中使用隐",
"藏绝对常数的“≥”或“≤”表示法表示某些不等式。我们也偶尔使用标准符号O.2和,来隐藏多",
"对数因子。除非另有说明，否则这些顺序符号相对于训练集大小n。无论何时使用，都在附录中指定精",
"确的常量",
"单头自我关注。由键、查询和值矩阵WQ，WkeRdxdhandWvRdxd参数化的具有上下文大小T和",
"维度d的单层自关注头ATTN：RTxd→RTxd由下式给出：",
"ATTN(X; WQ, Wk,Wv) := β(XWQWKX)XWv:",
"这里，X=[a1,a2,",
",&TJTRT×d是输人令牌矩阵，(XWQWEXT)eRTxT是关注矩阵，()：RT→",
"RT是。应用于矩阵的Softmax按行操作。要在预测标签中转换Attention输出，我们使用线性投影头",
"（又名解码器）组成ATTN。因此，模型的输出是t",
"(X; W,U) :=(U,p(XWX)X).",
"(1)",
"注意，我们将权重矩阵Wv的值吸收到投影仪U=[u1..,uT]TeRTxd中。此外，我们将整个密钥查",
"询乘积矩阵参数化为W：=WoWK",
"多头自我关注。我们的重点是带有H头的多头注意力（MHA）模型。对于输出矩阵WohεRdxd",
"MHA模型的输出令牌由",
"E ATTN(X;Weh,Wkh, Wvh)Woh:",
"he[H]",
"给出。与单头关注类似，我们将这些令牌投影到标量预测值。将WvhWoh吸收到投影层中并对",
"fwhile we focus on (i) Ful-projection: trainable matrix U e RTxd , our results also apply to (i) Pooling: U= ulI with",
"trainable ue Rd , and (ii) Last-token output: U-[odx(T-1)u",
"with trainable u e Rd.",
"2"
] |
{
"arxiv_id": "2310.12680",
"language": "zh",
"page_idx": 1
}
|
||
2310.12680-zh-2
|
2310.12680-zh-2.jpg
|
$W_{h}:=W_{Q_{h}}W_{K_{h}}^{\top\ddagger}$进行参数化，我们得出以下 MHA 模型
(2)
$$\widetilde{\Phi}(\boldsymbol{X};\widetilde{\boldsymbol{W}},\widetilde{\boldsymbol{U}}):=\frac{1}{\sqrt{H}}\sum_{h\epsilon[H]}\Phi(\boldsymbol{X};\boldsymbol{W}_{h},\boldsymbol{U}_{h})=\frac{1}{\sqrt{H}}\sum_{h\epsilon[H]}\langle\boldsymbol{U}_{h},\boldsymbol{\varphi}(\boldsymbol{X}\boldsymbol{W}_{h}\boldsymbol{X}^{\intercal})\boldsymbol{X}\rangle\:,$$
由$\widetilde{W}:=$concat$\left[\{\boldsymbol W_h\}_{h\in[H]}\right[$和$\widetilde{\boldsymbol{U}}:=$concat$\left[\{\boldsymbol U_h\}_{h\in[H]}\right[$.参数化 1$/\sqrt{H}$ 缩放类似于 MLP 文献中的归一化。g。(Du et al., 2019; Ji and Telgarsky, 2021;Richards and Kuzborskij, 2021) , 确 保 $U_h$初始化$\mathcal{O}_H(1)$时模型方差为常数阶
在整个过程中，我们将使用$\boldsymbol{\theta}_h:=$ concat$(\boldsymbol U_h,\boldsymbol W_h)\in\mathbb{R}^{dT+d^2}$,表示$h$ -注意头的可训练参数，并将$\widetilde{\boldsymbol{\theta}}:=$concat$(\{\boldsymbol{\theta}_h\}_{h\in[H]})\in\mathbb{R}^{H(dT+d^2)}$表示为整体模型的可训练参数。更一般地说，我们对与多头模型有关的数量使用“二”表示法的约定。最后，通过一些轻微的符号滥用，我们定义：$\|\widetilde{\boldsymbol{\theta}}\|_{2,\infty}:=\max_{h\in[H]}\|\boldsymbol{\theta}_h\|.$ 梯度下降训练。$\operatorname{E\text{给定训练集 }}(\boldsymbol{X}_i,y_i)_{i\in[n]}$,使用$n$ IID 样本，我们将基于逻辑损失的经验风险最小化
$$\widehat{L}(\widetilde{\boldsymbol{\theta}}):=\frac{1}{n}\sum\limits_{i\in[n]}\ell(y_{i}\widetilde{\Phi}(\boldsymbol{X}_{i};\widetilde{\boldsymbol{\theta}})):=\frac{1}{n}\sum\limits_{i\in[n]}\log(1+e^{-y_{i}\widetilde{\Phi}(\boldsymbol{X}_{i};\widetilde{\boldsymbol{\theta}})})\:.$$
Our analysis extends to any convex, smooth, Lipschitz and self-bounded loss.$^{8}$经验风险被最小化为定义为测试损失的近似值。形式上，在$\widetilde{\boldsymbol{\theta}}^{(0)}$处初始化并配备步长$\eta>0$ ,在每次迭代$k\geq0$时，GD 执行以下更新：
$$\widetilde{\boldsymbol{\theta}}^{(k+1)}=\widetilde{\boldsymbol{\theta}}^{(k)}-\eta\nabla\widehat{L}(\widetilde{\boldsymbol{\theta}}^{(k)})\:.$$
3 Gradient and Hessian bounds of soft-max attention
本节建立了在多头注意力模型上评估的逻辑经验风险$\widehat{L}(.)$的梯度和 Hessian 的界限。为此，我们首先推导了自注意模型的梯度和 Hessian 的欧儿里得范数和谱范数的界限。为了简化符号，我们在这里说明单头模型的边界 (请参阅 App。用于多头模型的 A.1 ): $\Phi(\boldsymbol{X};\boldsymbol{\theta}):=\Phi(\boldsymbol{X};\boldsymbol{W},\boldsymbol{U})=\langle\boldsymbol{U},\boldsymbol{\varphi}(\boldsymbol{XWX}^\top)\boldsymbol{X}\rangle.$
$$\begin{aligned}&\bullet\quad\nabla_{\boldsymbol{U}}\Phi(\boldsymbol{X};\boldsymbol{\theta})=\boldsymbol{\varphi}(\boldsymbol{X}\boldsymbol{W}\boldsymbol{X}^{\mathrm{T}})\boldsymbol{X},\quad and\quad\nabla_{\boldsymbol{W}}\Phi(\boldsymbol{X};\boldsymbol{\theta})=\sum_{t=1}^{T}\boldsymbol{x}_{t}\boldsymbol{u}_{t}^{\mathrm{T}}\boldsymbol{X}^{\mathrm{T}}\boldsymbol{\varphi}^{\prime}(\boldsymbol{X}\boldsymbol{W}^{\mathrm{T}}\boldsymbol{x}_{t})\boldsymbol{X}.\\&\bullet\quad\nabla_{\boldsymbol{W}}\langle\boldsymbol{a},\nabla_{\boldsymbol{U}}\Phi(\boldsymbol{X};\boldsymbol{\theta})\boldsymbol{b}\rangle=\sum_{t=1}^{T}\boldsymbol{x}_{t}a_{t}b^{\mathrm{T}}\boldsymbol{X}^{\mathrm{T}}\boldsymbol{\varphi}^{\prime}(\boldsymbol{X}\boldsymbol{W}^{\mathrm{T}}\boldsymbol{x}_{t})\boldsymbol{X},\quad and\\&\nabla_{\boldsymbol{W}}\langle\boldsymbol{c},\nabla_{\boldsymbol{W}}\Phi(\boldsymbol{X};\boldsymbol{\theta})\boldsymbol{b}\rangle=\sum_{t=1}^{T}(\boldsymbol{c}^{\mathrm{T}}\boldsymbol{x}_{t})\boldsymbol{x}_{t}\boldsymbol{d}^{\mathrm{T}}\boldsymbol{\varphi}^{\prime}(\boldsymbol{X}\boldsymbol{W}^{\mathrm{T}}\boldsymbol{x}_{t})\boldsymbol{X}\\&where\:\boldsymbol{d}:=\mathrm{diag}(\boldsymbol{X}\boldsymbol{b})\boldsymbol{X}\boldsymbol{u}_{t}-\boldsymbol{X}\boldsymbol{u}_{t}\boldsymbol{b}^{\mathrm{T}}\boldsymbol{X}^{\mathrm{T}}\boldsymbol{\varphi}(\boldsymbol{X}\boldsymbol{W}^{\mathrm{T}}\boldsymbol{x}_{t})-\boldsymbol{X}\boldsymbol{b}\boldsymbol{u}_{t}^{\mathrm{T}}\boldsymbol{X}^{\mathrm{T}}\boldsymbol{\varphi}(\boldsymbol{X}\boldsymbol{W}^{\mathrm{T}}\boldsymbol{x}_{t})\:.\end{aligned}$$
这些计算意味着以下有用的界限。
$\textbf{Proposition 1 ( Model Gradient/ Hessian bounds) . The Euclidean norm of the gradient and the spectral}$
$$\begin{aligned}&\|\nabla_{\boldsymbol{\theta}}\Phi(\boldsymbol{X};\boldsymbol{\theta})\|\leq2\|\boldsymbol{X}\|_{2,\infty}^{2}\sum_{t=1}^{T}\|\boldsymbol{X}\boldsymbol{u}_{t}\|_{\infty}+\sqrt{T}\|\boldsymbol{X}\|_{2,\infty}\:.\\&\|\nabla_{\boldsymbol{\theta}}^{2}\Phi(\boldsymbol{X};\boldsymbol{\theta})\|\leq6\:d\:\|\boldsymbol{X}\|_{2,\infty}^{2}\:\|\boldsymbol{X}\|_{1,\infty}^{2}\sum_{t=1}^{T}\|\boldsymbol{X}\boldsymbol{u}_{t}\|_{\infty}+2\:\sqrt{T\:d}\:\|\boldsymbol{X}\|_{2,\infty}^{2}\:\|\boldsymbol{X}\|_{1,\infty}\:.\end{aligned}$$
接下来，我们关注经验损失$\widehat{L}$ 。为了得出其梯度和 Hessian 的界限，我们利用了命题 1 的模型界限
以及逻辑损失是自界的事实，即。e。,$|\ell^{\prime}(t)|\leq\ell(t)$ 。为了提供具体的陈述，我们首先引人一个温和的有
界性假设。
Assumption 1 (Bounded data). $Data\left(\boldsymbol{X},y\right)\in\mathbb{R}^{T\times d}\times\mathbb{R}$ satisfy the following conditions almost surely:
$y\in \{ \pm 1\}$ , and for some $R\geq 1$ , $it$ holds for all $t\in [ T]$ that $\| \boldsymbol{x}_{t}\| \leq R$ .
$^{\dagger}$These relaxations sacrifice some generality since it is common practice to set values for $d_h$ and $d_v$ such that $d_v=d/H<d$ ,thus imposing low-rank restrictions on the product matrices $\dot{W}_{Q_h}\boldsymbol{W}_{K_h}^{\intercal},\boldsymbol{W}_{V_h}\boldsymbol{W}_{O_h}.$ We defer a treatment of these to future work.
$^{6}$A function $\ell:\mathbb{R}\to\mathbb{R}$ is self-bounded if $\exists C>0$ such that $|\ell^\prime(t)|\leq C\ell(t)$ (Ji and Telgarsky, 2018; Schliserman and Koren,
2022) .
3
|
[
[
0.12005615234375,
0.08612060546875,
0.5517578125,
0.08489990234375,
0.5517578125,
0.09979248046875,
0.12005615234375,
0.1009521484375
],
[
0.3359375,
0.1134033203125,
0.3525390625,
0.1134033203125,
0.3525390625,
0.1217041015625,
0.3359375,
0.1217041015625
],
[
0.5556640625,
0.11102294921875,
0.57080078125,
0.11639404296875,
0.56494140625,
0.12353515625,
0.5498046875,
0.1181640625
],
[
0.53662109375,
0.1181640625,
0.7958984375,
0.10809326171875,
0.796875,
0.133056640625,
0.5380859375,
0.1436767578125
],
[
0.85546875,
0.1181640625,
0.87890625,
0.1181640625,
0.87890625,
0.1312255859375,
0.85546875,
0.1312255859375
],
[
0.364501953125,
0.119384765625,
0.3779296875,
0.119384765625,
0.3779296875,
0.12353515625,
0.364501953125,
0.12353515625
],
[
0.208251953125,
0.11639404296875,
0.318115234375,
0.11639404296875,
0.318115234375,
0.1312255859375,
0.208251953125,
0.1312255859375
],
[
0.32080078125,
0.12054443359375,
0.5400390625,
0.10986328125,
0.54248046875,
0.133056640625,
0.322509765625,
0.1436767578125
],
[
0.118408203125,
0.149658203125,
0.880859375,
0.149658203125,
0.880859375,
0.16748046875,
0.118408203125,
0.16748046875
],
[
0.1292724609375,
0.1650390625,
0.87890625,
0.1644287109375,
0.87890625,
0.1787109375,
0.1292724609375,
0.179931640625
],
[
0.12005615234375,
0.1787109375,
0.32568359375,
0.1787109375,
0.32568359375,
0.1923828125,
0.12005615234375,
0.1923828125
],
[
0.1435546875,
0.1923828125,
0.880859375,
0.1923828125,
0.880859375,
0.210205078125,
0.1435546875,
0.210205078125
],
[
0.1158447265625,
0.2078857421875,
0.8818359375,
0.207275390625,
0.8818359375,
0.22802734375,
0.1158447265625,
0.2286376953125
],
[
0.1175537109375,
0.2244873046875,
0.87890625,
0.226806640625,
0.87890625,
0.24462890625,
0.1175537109375,
0.2423095703125
],
[
0.1175537109375,
0.240478515625,
0.86376953125,
0.2410888671875,
0.86376953125,
0.2587890625,
0.1175537109375,
0.25830078125
],
[
0.2822265625,
0.265380859375,
0.7109375,
0.2666015625,
0.7109375,
0.29150390625,
0.2822265625,
0.290283203125
],
[
0.34765625,
0.281982421875,
0.393798828125,
0.290283203125,
0.3896484375,
0.3017578125,
0.343505859375,
0.29345703125
],
[
0.5087890625,
0.27978515625,
0.56689453125,
0.290283203125,
0.5615234375,
0.302734375,
0.50390625,
0.292236328125
],
[
0.124267578125,
0.311279296875,
0.87744140625,
0.311279296875,
0.87744140625,
0.325927734375,
0.124267578125,
0.325927734375
],
[
0.12005615234375,
0.32470703125,
0.87890625,
0.325927734375,
0.87890625,
0.34033203125,
0.12005615234375,
0.339111328125
],
[
0.1234130859375,
0.341552734375,
0.17626953125,
0.341552734375,
0.17626953125,
0.352783203125,
0.1234130859375,
0.352783203125
],
[
0.39794921875,
0.35986328125,
0.59765625,
0.3623046875,
0.59765625,
0.37890625,
0.39794921875,
0.377197265625
],
[
0.118408203125,
0.40087890625,
0.779296875,
0.40087890625,
0.779296875,
0.4150390625,
0.118408203125,
0.4150390625
],
[
0.12005615234375,
0.432861328125,
0.87744140625,
0.432861328125,
0.87744140625,
0.446533203125,
0.12005615234375,
0.446533203125
],
[
0.12005615234375,
0.44775390625,
0.8759765625,
0.44775390625,
0.8759765625,
0.46142578125,
0.12005615234375,
0.46142578125
],
[
0.12005615234375,
0.4619140625,
0.861328125,
0.4619140625,
0.861328125,
0.4755859375,
0.12005615234375,
0.4755859375
],
[
0.1175537109375,
0.482666015625,
0.82275390625,
0.48388671875,
0.82275390625,
0.498779296875,
0.1175537109375,
0.49755859375
],
[
0.118408203125,
0.50830078125,
0.424072265625,
0.50830078125,
0.424072265625,
0.521484375,
0.118408203125,
0.521484375
],
[
0.57421875,
0.5029296875,
0.5810546875,
0.5029296875,
0.5810546875,
0.51025390625,
0.57421875,
0.51025390625
],
[
0.450927734375,
0.5087890625,
0.55517578125,
0.5087890625,
0.55517578125,
0.52099609375,
0.450927734375,
0.52099609375
],
[
0.57421875,
0.50830078125,
0.76904296875,
0.50830078125,
0.76904296875,
0.521484375,
0.57421875,
0.521484375
],
[
0.5634765625,
0.51953125,
0.58837890625,
0.51953125,
0.58837890625,
0.53076171875,
0.5634765625,
0.53076171875
],
[
0.4306640625,
0.51123046875,
0.453369140625,
0.51123046875,
0.453369140625,
0.5185546875,
0.4306640625,
0.5185546875
],
[
0.33154296875,
0.5322265625,
0.339111328125,
0.5322265625,
0.339111328125,
0.53759765625,
0.33154296875,
0.53759765625
],
[
0.1175537109375,
0.54052734375,
0.322509765625,
0.5390625,
0.322509765625,
0.55419921875,
0.1175537109375,
0.5546875
],
[
0.330078125,
0.5419921875,
0.343505859375,
0.5419921875,
0.343505859375,
0.54638671875,
0.330078125,
0.54638671875
],
[
0.339111328125,
0.54052734375,
0.59375,
0.54052734375,
0.59375,
0.55419921875,
0.339111328125,
0.55419921875
],
[
0.32568359375,
0.55419921875,
0.344970703125,
0.55419921875,
0.344970703125,
0.5625,
0.32568359375,
0.5625
],
[
0.118408203125,
0.57470703125,
0.297119140625,
0.572265625,
0.297119140625,
0.5859375,
0.118408203125,
0.58837890625
],
[
0.31494140625,
0.572265625,
0.51904296875,
0.572265625,
0.51904296875,
0.58740234375,
0.31494140625,
0.58740234375
],
[
0.301513671875,
0.58837890625,
0.318115234375,
0.58837890625,
0.318115234375,
0.595703125,
0.301513671875,
0.595703125
],
[
0.2216796875,
0.60107421875,
0.77099609375,
0.60107421875,
0.77099609375,
0.6142578125,
0.2216796875,
0.6142578125
],
[
0.1435546875,
0.62353515625,
0.387939453125,
0.62353515625,
0.387939453125,
0.63720703125,
0.1435546875,
0.63720703125
],
[
0.12005615234375,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.6650390625,
0.12005615234375,
0.6650390625
],
[
0.12005615234375,
0.66650390625,
0.72705078125,
0.66650390625,
0.72705078125,
0.6796875,
0.12005615234375,
0.6796875
],
[
0.12005615234375,
0.68701171875,
0.318115234375,
0.689453125,
0.318115234375,
0.70361328125,
0.12005615234375,
0.70166015625
],
[
0.330078125,
0.68603515625,
0.5322265625,
0.68701171875,
0.5322265625,
0.705078125,
0.330078125,
0.70361328125
],
[
0.313232421875,
0.69677734375,
0.31982421875,
0.69677734375,
0.31982421875,
0.70166015625,
0.313232421875,
0.70166015625
],
[
0.33154296875,
0.70361328125,
0.3525390625,
0.70361328125,
0.3525390625,
0.7119140625,
0.33154296875,
0.7119140625
],
[
0.12005615234375,
0.72021484375,
0.393798828125,
0.72265625,
0.393798828125,
0.73681640625,
0.12005615234375,
0.7353515625
],
[
0.4013671875,
0.72021484375,
0.6767578125,
0.7216796875,
0.6767578125,
0.736328125,
0.4013671875,
0.7353515625
],
[
0.403076171875,
0.73828125,
0.4189453125,
0.73828125,
0.4189453125,
0.7421875,
0.403076171875,
0.7421875
],
[
0.1427001953125,
0.75439453125,
0.87744140625,
0.7548828125,
0.87744140625,
0.76953125,
0.1427001953125,
0.7685546875
],
[
0.118408203125,
0.76708984375,
0.87744140625,
0.7685546875,
0.87744140625,
0.783203125,
0.118408203125,
0.7822265625
],
[
0.12005615234375,
0.78515625,
0.19482421875,
0.78515625,
0.19482421875,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.118408203125,
0.80859375,
0.87890625,
0.80859375,
0.87890625,
0.82666015625,
0.118408203125,
0.82666015625
],
[
0.118408203125,
0.82666015625,
0.62451171875,
0.82666015625,
0.62451171875,
0.84033203125,
0.118408203125,
0.84033203125
],
[
0.118408203125,
0.845703125,
0.880859375,
0.84716796875,
0.880859375,
0.8623046875,
0.118408203125,
0.85986328125
],
[
0.1275634765625,
0.85986328125,
0.880859375,
0.85986328125,
0.880859375,
0.87353515625,
0.1275634765625,
0.87353515625
],
[
0.124267578125,
0.87060546875,
0.204833984375,
0.87158203125,
0.204833984375,
0.8828125,
0.124267578125,
0.8818359375
],
[
0.1175537109375,
0.88134765625,
0.87890625,
0.8828125,
0.87890625,
0.896484375,
0.1175537109375,
0.89453125
],
[
0.1259765625,
0.8955078125,
0.17041015625,
0.8955078125,
0.17041015625,
0.90673828125,
0.1259765625,
0.90673828125
],
[
0.490234375,
0.93505859375,
0.5087890625,
0.93505859375,
0.5087890625,
0.94970703125,
0.490234375,
0.94970703125
]
] |
[
"Wh：=WeWkT进行参数化，我们得出以下MHA模型",
"1",
"1",
"Zhe[m(Uh, 9(XWhXT)X),",
"(2)",
"",
"(X; W,U) :-",
"VZhe[m) (X; Wn, Un) =",
"由 W;=concat[(Wh}he[)[和 U:= concat[{Un]he[H)[·参数化 1/VH缩放类似于 MLP文献中的归",
"-化。g。(Du et al.,2019;Ji and Telgarsky,2021;Richards and Kuzborskij,2021），确保Uh初始化",
"OH(1)时模型方差为常数阶",
"在整个过程中，我们将使用Oh:=concat(Uh,Wh）eRdT+d²，表示h-注意头的可训练参数，并将",
"：=concat((on)he[H)eIRH(d+d-）表示为整体模型的可训练参数。更一般地说，我们对与多头模型有关",
"的数量使用“～”表示法的约定。最后，通过一些轻微的符号滥用，我们定义：[2,=maxhe[H]0nl",
"梯度下降训练。E给定训练集（X，yi)ie[n)，使用n IID样本，我们将基于逻辑损失的经验风险最小化",
"L(0) := 1 Z (y;(X:;0) :=1 Z log(1 +e-v:(x:0).",
"nien]",
"mien)",
"Our analysis extends to any convex, smooth, Lipschitz and self-bounded loss.s经验风险被最小化为定",
"义为测试损失的近似值。形式上，在(0）处初始化并配备步长n>0，在每次迭代≥0时，GD执行以",
"下更新：",
"可(k+1) = (k) - nVL(0(k)) .",
"3 Gradient and Hessian bounds of soft-max attention",
"本节建立了在多头注意力模型上评估的逻辑经验风险L(）的梯度和Hessian的界限。为此，我们首先推",
"导了自注意模型的梯度和Hessian的欧几里得范数和谱范数的界限。为了简化符号，我们在这里说明单",
"头模型的边界（请参阅ApP。用于多头模型的A.1)：Φ(X;0)=Φ(XW,U)=(U,(XWXT)X)",
"Lemma 1 (Gradient/Hessian formulas). For all a e RT , b,c e Rd the model's gradients satisfy:",
". VuΦ(X;0)=β(XWx)X,and",
"1",
"Vwd(X;0) :",
" tut x p'(Xw at)X.",
"t=1",
"",
"T",
". Vw(a, VuΦ(X;0)b) =",
"",
"atatbTXT'(XWTat)X,and",
"=1",
"Vw(c, Vw4(X; 0)b) =",
"(c'at) atd '(Xwat)X",
"t=",
"where d := diag(Xb)Xut - Xutb X p(XW a&t) - Xbut X p(XWTat)",
"这些计算意味着以下有用的界限。",
"Proposition 1 (Model Gradient/Hessian bounds). The Euclidean norm of the gradient and the spectral",
"norm of the Hessian of the single-head Attention model (1) are bounded as follows:",
". /Ved(X;0)≤2|X2.。",
"Z IXut l。 + VTX2,00 .",
"o",
"t=1",
".v(X;0)≤6dX2.X/1.",
"Xut+2VTdX2.0X1,0",
"",
"接下来，我们关注经验损失L。为了得出其梯度和Hessian的界限，我们利用了命题1的模型界限",
"以及逻辑损失是自界的事实，即。e。，le(t)<l(t）。为了提供具体的陈述，我们首先引人一个温和的有",
"界性假设。",
"Assumption 1 (Bounded data). Data (X,y) e RTxd x R satisfy the following conditions almost surely",
"y e(+1) , and for some R≥1, it holds for all t e[T] that atll <R.",
"+These relaxations sacrifice some generality since it is common practice to set values for dh and dy such that du = d/H<d",
", thus imposing low-rank restrictions on the product matrices Woh Wk h , Wvh Woh - We defer a treatment of these to",
"future work.",
"SA function l : R -→ R is self-bounded if 3 C > 0 such that [e'(t)l ≤ Ce(t) (Ji and Telgarsky, 2018; Schliserman and Koren,",
"2022)",
"3"
] |
{
"arxiv_id": "2310.12680",
"language": "zh",
"page_idx": 2
}
|
||
2310.12680-zh-3
|
2310.12680-zh-3.jpg
|
$$\begin{aligned}&ounds{:}^{\P}\\&(1)\quad\|\nabla\widehat{L}(\widetilde{\boldsymbol{\theta}})\|\leq\beta_{1}(\widetilde{\boldsymbol{\theta}})\:\widehat{L}(\widetilde{\boldsymbol{\theta}})\:,&&\beta_{1}(\widetilde{\boldsymbol{\theta}}):=\sqrt{T}\:R\left[2\:R^{2}\:\|\widetilde{\boldsymbol{\theta}}\|_{2,\infty}+1\right[\:.\\&(2)\quad\|\nabla^{2}\widehat{L}(\widetilde{\boldsymbol{\theta}})\|\leq\beta_{2}(\widetilde{\boldsymbol{\theta}})\:,&&\beta_{2}(\widetilde{\boldsymbol{\theta}}):=\frac{1}{\sqrt{H}}\:\beta_{3}(\widetilde{\boldsymbol{\theta}})+\frac{1}{4}\:\beta_{1}(\widetilde{\boldsymbol{\theta}})^{2}.\\&(3)\:\lambda_{\mathrm{min}}(\nabla^{2}\widehat{L}(\widetilde{\boldsymbol{\theta}}))\geq-\frac{\beta_{3}(\widetilde{\boldsymbol{\theta}})}{\sqrt{H}}\widehat{L}(\widetilde{\boldsymbol{\theta}})&&\beta_{3}(\widetilde{\boldsymbol{\theta}}):=2\:\sqrt{T\:d}\:R^{3}\left[3\:\sqrt{d}\:R^{2}\:\|\widetilde{\boldsymbol{\theta}}\|_{2,\infty}+1\right[\:.\end{aligned}$$
Corollary 1 (Loss properties). UnderAssumption 1, the objective's gradient and Hessian satisfy the
上述损失属性对于下一节中的训练和泛化分析至关重要。性质(1)建立了经验损失的自有界性。这将用于在泛化分析中分析 GD 更新的稳定性。属性(2)将用于建立梯度更新的下降，以适当选择步长$\eta$ 。请注意，平滑度上限取决于$\widetilde{\boldsymbol{\theta}}$ ,因此要显示下降，我们还需要保证更新的有界性。最后，性质(3)建立了损失的自界弱凸性质，这对于训练和泛化分析都是至关重要的。具体来说，随着磁头$H$的数量增加， 最小特征值变得不那么正，表明朝着凸状行为的方向发展
# 4 Main results
在本节中，我们将介绍多头注意力的训练和泛化范围。
## 4.1 Training bounds
我们在以下定理中陈述了有关火车损失收敛的主要结果。请参阅应用程序。B 用于精确常量和详细证明
Theorem 1 (Training loss). Fix iteration horizon K ≥ 1 and any $\widetilde{\boldsymbol{\theta}}\in\mathbb{R}^{H(dT+d^2)}andH$ satisfying
$$\sqrt{H}\gtrsim dT^{1/2}R^{5}\|\widetilde{\boldsymbol{\theta}}\|_{2,\infty}\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|^{3}.$$
(3)
$Fix\textit{ step- size}$
$$\eta\leq1\:\wedge\:1/\rho(\widetilde{\boldsymbol{\theta}})\:\wedge\:\frac{\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|^{2}}{K\widehat{L}(\widetilde{\boldsymbol{\theta}})}\:\wedge\:\frac{\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|^{2}}{\widetilde{L}(\widetilde{\boldsymbol{\theta}}^{(0)})},\quad with\quad\rho(\widetilde{\boldsymbol{\theta}})\lesssim d^{3/2}\:T^{3/2}\:R^{13}\:\|\widetilde{\boldsymbol{\theta}}\|_{2,\infty}^{2}\:\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|^{2}.$$
$Then$, the following bounds hold for the training loss and the $weights’ norm$ $at$ iteration $K$ $of$ $GD:$
(4)
$$\begin{aligned}&\widehat{L}(\widetilde{\boldsymbol{\theta}}^{(K)})\leq\frac{1}{K}\sum_{k=1}^{K}\widehat{L}(\widetilde{\boldsymbol{\theta}}_{k})\leq2\widehat{L}(\widetilde{\boldsymbol{\theta}})+\frac{5\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|^{2}}{4\eta K},\\&\|\widetilde{\boldsymbol{\theta}}^{(K)}-\widetilde{\boldsymbol{\theta}}^{(0)}\|\leq4\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|.\end{aligned}$$
要应用定理并得出具体的火车损失界限，需要一组适当的目标参数$\widetilde{\boldsymbol{\theta}}$ ,以最小化 (4)中的界限。因此，$\widetilde{\boldsymbol{\theta}}$应同时实现小损耗 ($\widehat{L}(\widetilde{\boldsymbol{\theta}})$ )和初始化距离 ($\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|$ )。在下面的假设 2 中形式化了此需求。到初始化的距离以及$\|\widetilde{\boldsymbol{\theta}}\|_{2,\infty}$决定了我们的边界需要多少个头才能保持。此外，鉴于 (4)中的界限，对 $\widetilde{\boldsymbol{\theta}}$ 的适当选择获得与$\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|^2/K.$相同阶数的$\widehat{L}(\widetilde{\boldsymbol{\theta}})$是合理的。因此，定理对步长的限制受损失的逆局部平滑度控制：$\eta \lesssim 1/ \rho ( \widetilde{\boldsymbol{\theta }} )$ .
## 4.2 Generalization bounds
接下来，我们绑定了预期的泛化间隙。期望是关于 (w。r。t)列车集的随机性。请参阅应用程序。C 用
于详细证明，它基于算法稳定性
Theorem 2 (Generalization loss). $Fix\textit{any K}\geq 1$ , $any\widetilde\boldsymbol{\theta}\textit{and H satisfying (3) , and any step-size \eta}$
satisfying the conditions $of$ $Thm.$ 1 . Then the expected generalization gap $at$ iteration $K$ $satisfles$,
$$\mathbb{E}\Big[L(\widetilde{\boldsymbol{\theta}}^{(K)})-\widehat{L}(\widetilde{\boldsymbol{\theta}}^{(K)})\Big[\leq\frac{4}{n}\:\mathbb{E}\Big[2\:K\:\widehat{L}(\widetilde{\boldsymbol{\theta}})+\frac{9\|\widetilde{\boldsymbol{\theta}}-\widetilde{\boldsymbol{\theta}}^{(0)}\|^{2}}{4\eta}\Big[\:.$$
(5)
关于磁头数量的条件与定理 1 中相应条件的常数相同。此外，通过与 Thm 结合，泛化-间隙界限转换为测试-损失界限。1。最后，类似于 Thm。1 ,我们可以在可实现性假设下得到具体的界限；见 Cor。
$\P_{\text{In all the bounds in this paper involving }\|\widetilde{\boldsymbol{\theta}}\|_{2,\infty},\text{ it is possible to substitute this term with max}_{h\in[H]}\|\boldsymbol{U}_h\|.\mathrm{~However,~for}}$
the sake of notation simplicity, we opt for a slightly looser bound $\max_{h\in[H]}\|\boldsymbol{U}_{h}\|\leq\max_{h\epsilon[H]}\|\boldsymbol{\theta}_{h}\|=:\|\widetilde{\boldsymbol{\theta}}\|_{2,\infty}$
4
|
[
[
0.12005615234375,
0.0872802734375,
0.87890625,
0.0872802734375,
0.87890625,
0.1015625,
0.12005615234375,
0.1015625
],
[
0.12176513671875,
0.1015625,
0.1822509765625,
0.1015625,
0.1822509765625,
0.1134033203125,
0.12176513671875,
0.1134033203125
],
[
0.118408203125,
0.114013671875,
0.32421875,
0.114013671875,
0.32421875,
0.12890625,
0.118408203125,
0.12890625
],
[
0.39794921875,
0.114013671875,
0.64404296875,
0.114013671875,
0.64404296875,
0.12890625,
0.39794921875,
0.12890625
],
[
0.118408203125,
0.1318359375,
0.291259765625,
0.1318359375,
0.291259765625,
0.146728515625,
0.118408203125,
0.146728515625
],
[
0.4013671875,
0.1300048828125,
0.634765625,
0.1300048828125,
0.634765625,
0.15087890625,
0.4013671875,
0.15087890625
],
[
0.11419677734375,
0.146728515625,
0.68994140625,
0.146728515625,
0.68994140625,
0.17041015625,
0.11419677734375,
0.17041015625
],
[
0.287109375,
0.1650390625,
0.295654296875,
0.163330078125,
0.297119140625,
0.16748046875,
0.2880859375,
0.168701171875
],
[
0.14697265625,
0.179931640625,
0.87744140625,
0.179931640625,
0.87744140625,
0.193603515625,
0.14697265625,
0.193603515625
],
[
0.12176513671875,
0.193603515625,
0.875,
0.193603515625,
0.875,
0.207275390625,
0.12176513671875,
0.207275390625
],
[
0.12176513671875,
0.208984375,
0.8759765625,
0.208984375,
0.8759765625,
0.22265625,
0.12176513671875,
0.22265625
],
[
0.1234130859375,
0.22265625,
0.87890625,
0.22265625,
0.87890625,
0.236328125,
0.1234130859375,
0.236328125
],
[
0.118408203125,
0.2369384765625,
0.5380859375,
0.2369384765625,
0.5380859375,
0.25048828125,
0.118408203125,
0.25048828125
],
[
0.118408203125,
0.273681640625,
0.318115234375,
0.273681640625,
0.318115234375,
0.2880859375,
0.118408203125,
0.2880859375
],
[
0.118408203125,
0.303955078125,
0.509765625,
0.303955078125,
0.509765625,
0.317626953125,
0.118408203125,
0.317626953125
],
[
0.118408203125,
0.33740234375,
0.340087890625,
0.33740234375,
0.340087890625,
0.350830078125,
0.118408203125,
0.350830078125
],
[
0.12005615234375,
0.361083984375,
0.873046875,
0.361083984375,
0.873046875,
0.374755859375,
0.12005615234375,
0.374755859375
],
[
0.1158447265625,
0.37890625,
0.84033203125,
0.38134765625,
0.84033203125,
0.402587890625,
0.1158447265625,
0.40087890625
],
[
0.38037109375,
0.412109375,
0.62060546875,
0.412109375,
0.62060546875,
0.427001953125,
0.38037109375,
0.427001953125
],
[
0.85546875,
0.41455078125,
0.880859375,
0.41455078125,
0.880859375,
0.427490234375,
0.85546875,
0.427490234375
],
[
0.118408203125,
0.438232421875,
0.21240234375,
0.440673828125,
0.21240234375,
0.454833984375,
0.118408203125,
0.453125
],
[
0.29296875,
0.463134765625,
0.483642578125,
0.46142578125,
0.483642578125,
0.478515625,
0.29296875,
0.48095703125
],
[
0.1695556640625,
0.47265625,
0.291259765625,
0.47021484375,
0.291259765625,
0.486328125,
0.1695556640625,
0.488037109375
],
[
0.50146484375,
0.47265625,
0.8271484375,
0.47265625,
0.8271484375,
0.48681640625,
0.50146484375,
0.48681640625
],
[
0.310546875,
0.482666015625,
0.36865234375,
0.482666015625,
0.36865234375,
0.49755859375,
0.310546875,
0.49755859375
],
[
0.405517578125,
0.482666015625,
0.478515625,
0.482666015625,
0.478515625,
0.49755859375,
0.405517578125,
0.49755859375
],
[
0.491943359375,
0.478515625,
0.49951171875,
0.478515625,
0.49951171875,
0.482666015625,
0.491943359375,
0.482666015625
],
[
0.397216796875,
0.48681640625,
0.40478515625,
0.48681640625,
0.40478515625,
0.4912109375,
0.397216796875,
0.4912109375
],
[
0.40478515625,
0.486328125,
0.41064453125,
0.486328125,
0.41064453125,
0.4912109375,
0.40478515625,
0.4912109375
],
[
0.12005615234375,
0.50830078125,
0.8330078125,
0.50830078125,
0.8330078125,
0.521484375,
0.12005615234375,
0.521484375
],
[
0.32421875,
0.5390625,
0.450927734375,
0.53076171875,
0.453369140625,
0.55029296875,
0.32666015625,
0.55859375
],
[
0.43408203125,
0.54052734375,
0.66748046875,
0.525390625,
0.6708984375,
0.55029296875,
0.4375,
0.5654296875
],
[
0.39208984375,
0.54638671875,
0.453369140625,
0.5517578125,
0.450927734375,
0.568359375,
0.3896484375,
0.5634765625
],
[
0.85546875,
0.54345703125,
0.880859375,
0.54345703125,
0.880859375,
0.55712890625,
0.85546875,
0.55712890625
],
[
0.60205078125,
0.5517578125,
0.63623046875,
0.5517578125,
0.63623046875,
0.56396484375,
0.60205078125,
0.56396484375
],
[
0.33154296875,
0.5693359375,
0.52294921875,
0.5693359375,
0.52294921875,
0.58447265625,
0.33154296875,
0.58447265625
],
[
0.145263671875,
0.5986328125,
0.87744140625,
0.5986328125,
0.87744140625,
0.61328125,
0.145263671875,
0.61328125
],
[
0.12005615234375,
0.61083984375,
0.8759765625,
0.61083984375,
0.8759765625,
0.6259765625,
0.12005615234375,
0.6259765625
],
[
0.12176513671875,
0.626953125,
0.8759765625,
0.626953125,
0.8759765625,
0.64013671875,
0.12176513671875,
0.64013671875
],
[
0.12176513671875,
0.64111328125,
0.8759765625,
0.64111328125,
0.8759765625,
0.65625,
0.12176513671875,
0.65625
],
[
0.12005615234375,
0.65771484375,
0.297119140625,
0.65771484375,
0.297119140625,
0.6728515625,
0.12005615234375,
0.6728515625
],
[
0.12176513671875,
0.69140625,
0.181396484375,
0.69140625,
0.181396484375,
0.703125,
0.12176513671875,
0.703125
],
[
0.17626953125,
0.6904296875,
0.4013671875,
0.6904296875,
0.4013671875,
0.70361328125,
0.17626953125,
0.70361328125
],
[
0.12005615234375,
0.71337890625,
0.8759765625,
0.71337890625,
0.8759765625,
0.72705078125,
0.12005615234375,
0.72705078125
],
[
0.12005615234375,
0.728515625,
0.35693359375,
0.728515625,
0.35693359375,
0.7421875,
0.12005615234375,
0.7421875
],
[
0.118408203125,
0.7490234375,
0.8818359375,
0.75,
0.8818359375,
0.76708984375,
0.118408203125,
0.7666015625
],
[
0.12176513671875,
0.7666015625,
0.8388671875,
0.7666015625,
0.8388671875,
0.77978515625,
0.12176513671875,
0.77978515625
],
[
0.295654296875,
0.79345703125,
0.7021484375,
0.78515625,
0.70361328125,
0.81640625,
0.297119140625,
0.82470703125
],
[
0.85546875,
0.80029296875,
0.87890625,
0.80029296875,
0.87890625,
0.8154296875,
0.85546875,
0.8154296875
],
[
0.623046875,
0.80810546875,
0.69775390625,
0.81005859375,
0.69775390625,
0.82470703125,
0.623046875,
0.822265625
],
[
0.14697265625,
0.83447265625,
0.87744140625,
0.83447265625,
0.87744140625,
0.84716796875,
0.14697265625,
0.84716796875
],
[
0.12005615234375,
0.84716796875,
0.8759765625,
0.84716796875,
0.8759765625,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.1175537109375,
0.876953125,
0.87890625,
0.87890625,
0.87890625,
0.896484375,
0.1175537109375,
0.89453125
],
[
0.1275634765625,
0.8935546875,
0.80615234375,
0.89453125,
0.80615234375,
0.9091796875,
0.1275634765625,
0.90771484375
],
[
0.490234375,
0.93603515625,
0.5087890625,
0.93603515625,
0.5087890625,
0.95068359375,
0.490234375,
0.95068359375
]
] |
[
"Corollary 1 (Loss properties). Under Assumption 1 , the objective's gradient and Hessian satisfy the",
"bounds:",
"(1)[VL(0)≤β1(0)L(0) ,",
"B1(0) := VT R[2 R202,0 + 1",
"(2)V2L(0) ≤β2(0) ,",
"β2(0) := βa(0) + 1 β1(0)2.",
"(3) min(v2L(0) ≥ -BL(0) β3(0) := 2 Td R? [3 VaR2 02, +1[ .",
"",
"上述损失属性对于下一节中的训练和泛化分析至关重要。性质（1）建立了经验损失的自有界性。这将",
"用于在泛化分析中分析GD更新的稳定性。属性（2）将用于建立梯度更新的下降，以适当选择步长n",
"请注意，平滑度上限取决于，因此要显示下降，我们还需要保证更新的有界性。最后，性质（3）建立",
"了损失的自界弱凸性质，这对于训练和泛化分析都是至关重要的。具体来说，随着磁头H的数量增加，",
"最小特征值变得不那么正，表明朝着凸状行为的方向发展",
"4 Main results",
"在本节中，我们将介绍多头注意力的训练和泛化范围。",
"4.1 Training bounds",
"我们在以下定理中陈述了有关火车损失收敛的主要结果。请参阅应用程序。B用于精确常量和详细证明",
"Theorem 1 (Training loss). Fir iteration horizon K ≥1 and any e RH(dT+d) and H satisfying",
"VH≥dT1/2R502,0-0(0)|3",
"(3)",
"Fir step-size",
"-(0)2-(0)2",
"≤1 1/p(0) ",
"withp(0)≤d3/2 T3/2 R1302.0-0(0)2",
"KL(O)",
"L(0(0))",
"",
"",
"",
"Then, the following bounds hold for the training loss and the weights' norm at iteration K of GD",
"L(G(K)≤1",
"2 (0)≤22(0) + 518- 0(0)2",
"“Ki",
"(4)",
"4nK",
"(K) (0)≤ 4 - (0) ",
"要应用定理并得出具体的火车损失界限，需要一组适当的目标参数，以最小化（4）中的界限。因",
"此，应同时实现小损耗（L(））和初始化距离（-（0)）。在下面的假设2中形式化了此需求。到",
"初始化的距离以及02决定了我们的边界需要多少个头才能保持。此外，鉴于（4）中的界限，对θ的",
"适当选择获得与-(0)2/K.相同阶数的L(）是合理的。因此，定理对步长的限制受损失的逆局部平",
"滑度控制：n≤1/p(@)。",
"4.2",
"Generalization bounds",
"接下来，我们绑定了预期的泛化间隙。期望是关于（w。r。t）列车集的随机性。请参阅应用程序。C用",
"于详细证明，它基于算法稳定性",
"Theorem 2 (Generalization loss). Fia any K ≥1 , any and H satisfying (3) , and any step-size n",
"satisfying the conditions of Thm. 1 . Then the erpected generalization gap at iteration K satisfies",
"E[(6) - (0)[[2 2(0)+ 910-90[",
"(5)",
"4n",
"关于磁头数量的条件与定理1中相应条件的常数相同。此外，通过与Thm结合，泛化-间隙界限转",
"换为测试-损失界限。1。最后，类似于Thm。1，我们可以在可实现性假设下得到具体的界限；见Cor",
"In all the bounds in this paper involving 2,, it is possible to substitute this term with maxhe[Hj UlHowever, for",
"the sake of notation simplicity, we opt for a slightly looser bound maxhe[H] |Uh ≤ maxhe[H) enl =: 2,",
"4"
] |
{
"arxiv_id": "2310.12680",
"language": "zh",
"page_idx": 3
}
|
||
2310.12680-zh-4
|
2310.12680-zh-4.jpg
|
4 在应用程序。C.2 。对于泛化分析，我们要求可实现性假设几乎肯定适用于从数据分布采样的所有训
练集
我们注意到优化和泛化的界限与逻辑回归 (Soudry et al., 2018; Ji and Telgarsky, 2018; Shamir, 202$\underline{1})$ 的类似界限相同。然而，为了使这些界限有效，我们需要足够多的头以及存在一组适当的目标参数$\widetilde{\theta}^{\prime}$ 如定理条件中所述。也就是说，这些条件与可实现性条件有关，这保证了在初始化附近获得较小的训练误差。下一个假设将这些条件形式化。
$\mathbf{Assumption~2~(Realizability).~There~exist~non-increasing~functions~g~:~\mathbb{R}_+\to\mathbb{R}_+~and~g_0~:~\mathbb{R}_+\to\mathbb{R}_+~.}$ such that $\forall \epsilon > 0$ , there exists model parameters $\widetilde{\boldsymbol{\theta }} _{( \epsilon ) }\in \mathbb{R} ^{H( dT+ d^2) }$ for $which:$ $( i)$ the empirical loss over $n$ data samples satisfies $\widehat{L} ( \widetilde{\boldsymbol{\theta }} _{( \varepsilon ) }) \leq \varepsilon$ , $( ii)$ $\| \widetilde{\boldsymbol{\theta }} _{( \varepsilon ) }- \widetilde{\boldsymbol{\theta }} ^{( 0) }\| \leq g_{0}( \varepsilon )$ , $and$, $( iii)$ $\| \widetilde{\boldsymbol{\theta }} _{( \varepsilon ) }\| _{2, \infty }\leq g( \varepsilon ) .$
通过此假设，我们可以将上述定理的结果专门用于特定的数据设置；请参阅应用程序中的推论 3 和 4
。B.5 和 C.2 。在下一节中，我们将进一步展示如何满足可实现性假设。
4.3 Primitive conditions for checking realizability
在这里，我们在数据和初始化上引入了一组更原始，更直接的检查条件，以确保可实现性假设 2 成立
$\textbf{Definition 1 }($Good initialization).$We\textit{say }\widetilde{\boldsymbol{\theta }} ^{( 0) }=$concat$(\boldsymbol{\theta}_1^{(0)},\ldots,\boldsymbol{\theta}_H^{(0)})$ is $a$ good initialization with
respect $to$ training data $( \boldsymbol{X}_i, y_i) _{i\in [ n] }$ provided the following three properties $hold.$
$P1.$ Parameter $L_{2, \infty }$ $- bound:$ There exists parameter $B_{2}\geq 1$ such that $\forall h\in [ H]$ $it$ holds $\| \boldsymbol{\theta }_{h}^{( 0) }\| _{2}\leq B_{2}.$
$P2. Modelbound: There$ exists parameter $B_{\Phi }\geq 1$ such that $\forall i\in [ n]$ $it$ holds $\left | \widetilde{\Phi } ( \boldsymbol{X}_{i}; \widetilde{\boldsymbol{\theta }} ^{( 0) }) \right | \leq B_{\Phi }$ .
$P3. NTKseparability. There$ exists $\widetilde{\boldsymbol{\theta }} _{\star }\in \mathbb{R} ^{H( dT+ d^{2}) }$ and $\gamma > 0$ such that $\| \widetilde{\boldsymbol{\theta }} _{\star }\| = \sqrt {2}$ and $\forall i\in [ n]$ $, it$ holds
$y_{i}\left\langle\nabla\widetilde{\Phi}\left[\tilde{\boldsymbol{X}}_{i};\widetilde{\boldsymbol{\theta}}^{(0)}\right]\tilde{\boldsymbol{\theta}}_{\star}\right\rangle\geq\gamma.$
道具。附录中的 7 表明，从 good 初始化开始，我们总能找到满足可实现性的$\widetilde{\theta}_{(\epsilon)}$假设 2 提供了足
够多的磁头。因此，给定 good 初始化，我们可以立即应用定理 1 和 2 来获得以下具体界限
Corollary 2 (General bounds under good initialization). Suppose good $initialization\widetilde{\boldsymbol{\theta}}^{(0)}andlet$
$$\sqrt{H}\gtrsim d\:T^{1/2}\:R^{5}\:B_{2}^{2}\:g_{0}(\frac{1}{K})^{3},$$
$where g_{0}( \frac 1K) = \frac {2B_{\Phi }+ \log ( K) }\gamma$ . Further fax any $step- size$ $\eta \leq$ 1 $\wedge$ $1/ \rho ( K)$ $\wedge$ $\frac {4B_{\Phi }^{2}}{\gamma ^{2}\log ( 1+ e^{B_{\Phi }}) }$ where $\rho ( K)$ $\gtrsim$
$d^{3/ 2}T^{3/ 2}R^{13}g_{0}( \frac 1K) ^{4}.$ $Then$, $it$ holds that
$$\begin{aligned}&\widehat{L}(\widetilde{\boldsymbol{\theta}}^{(K)})\leq\frac{2}{K}+\frac{5\left(2B_{\Phi}+\log(K)\right)^{2}}{4\gamma^{2}\:\eta\:K},\\&\mathbb{E}\Big[L(\widetilde{\boldsymbol{\theta}}^{(K)})-\widehat{L}(\widetilde{\boldsymbol{\theta}}^{(K)})\Big[\leq\frac{17\left(2B_{\Phi}+\log(K)\right)^{2}}{\gamma^{2}\:\eta\:n}\:.\end{aligned}$$
考虑$K$ GD 步骤后的训练损失：假设$B_\Phi=\widetilde{\mathcal{O}}_K(1)$和$\gamma=\mathcal{O}_K(1)$ ,然后选择$\eta=\widetilde{\mathcal{O}}_K(1)$ ,则推论保证训练损失为$\widetilde{\mathcal{O}}_K(\frac1K)$ ,前提是磁头的对数$H=\Omega(\log^6(K))$ 。此外，在$K\approx n$ GD 步骤之后，预期测试损耗为$\mathcal{O}(1/n)$
$\textbf{Remark 1. The last two conditions }( P2, P3) for$ good initialization are similar to the conditions needed $in$ $(Taheri$ and Thrampoulidis, 2023; Ji and Telgarsky, 2020; Nitanda et al., 2019) for analysis of two- $layerMLPs.Comparedto\left(Jiand\right.Telgarsky,2020;Nitandaetal.,2019)which$ assume random Gaussian $initialization~\widetilde{\boldsymbol{\theta}}^{(0)},~and~similar~to~Taheri~and~Thrampoulidis~(2023)~the~NTK~separability~assumption$ $( P3) \textit{ can potentially accommodate deterministic }\widetilde{\boldsymbol{\theta }} ^{( 0) }. \textit{ Condition ( P1) appears because we allow training}$ $both\textit{ layers of the model. Specifically the L}_{2, \infty }\textit{ norm originates from the Hessian bounds in Corollary 1 . }$
5 Application to tokenized-mixture model
我们现在通过一个示例演示我们的结果如何应用于特定的数据模型
5
|
[
[
0.12005615234375,
0.08612060546875,
0.87744140625,
0.0872802734375,
0.87744140625,
0.1009521484375,
0.12005615234375,
0.09979248046875
],
[
0.12176513671875,
0.1015625,
0.154541015625,
0.1015625,
0.154541015625,
0.1134033203125,
0.12176513671875,
0.1134033203125
],
[
0.14697265625,
0.1175537109375,
0.8759765625,
0.1175537109375,
0.8759765625,
0.127685546875,
0.14697265625,
0.127685546875
],
[
0.12005615234375,
0.12890625,
0.873046875,
0.12890625,
0.873046875,
0.142578125,
0.12005615234375,
0.142578125
],
[
0.12176513671875,
0.144287109375,
0.87890625,
0.144287109375,
0.87890625,
0.157958984375,
0.12176513671875,
0.157958984375
],
[
0.118408203125,
0.15673828125,
0.397216796875,
0.157958984375,
0.397216796875,
0.1728515625,
0.118408203125,
0.171630859375
],
[
0.12005615234375,
0.18115234375,
0.87744140625,
0.1817626953125,
0.87744140625,
0.196533203125,
0.12005615234375,
0.1953125
],
[
0.12005615234375,
0.1953125,
0.87890625,
0.1953125,
0.87890625,
0.213134765625,
0.12005615234375,
0.213134765625
],
[
0.118408203125,
0.2120361328125,
0.75244140625,
0.2120361328125,
0.75244140625,
0.2298583984375,
0.118408203125,
0.2298583984375
],
[
0.145263671875,
0.2381591796875,
0.880859375,
0.2381591796875,
0.880859375,
0.251708984375,
0.145263671875,
0.251708984375
],
[
0.118408203125,
0.2529296875,
0.6484375,
0.2529296875,
0.6484375,
0.2666015625,
0.118408203125,
0.2666015625
],
[
0.12005615234375,
0.283935546875,
0.62744140625,
0.283935546875,
0.62744140625,
0.298583984375,
0.12005615234375,
0.298583984375
],
[
0.12176513671875,
0.308837890625,
0.85791015625,
0.308837890625,
0.85791015625,
0.322509765625,
0.12176513671875,
0.322509765625
],
[
0.118408203125,
0.330078125,
0.87890625,
0.330078125,
0.87890625,
0.347412109375,
0.118408203125,
0.347412109375
],
[
0.12176513671875,
0.3486328125,
0.69873046875,
0.3486328125,
0.69873046875,
0.363525390625,
0.12176513671875,
0.363525390625
],
[
0.1292724609375,
0.374755859375,
0.8564453125,
0.374755859375,
0.8564453125,
0.388427734375,
0.1292724609375,
0.388427734375
],
[
0.1292724609375,
0.40087890625,
0.8212890625,
0.40087890625,
0.8212890625,
0.4150390625,
0.1292724609375,
0.4150390625
],
[
0.1259765625,
0.424560546875,
0.87890625,
0.424560546875,
0.87890625,
0.4423828125,
0.1259765625,
0.4423828125
],
[
0.162109375,
0.4423828125,
0.3525390625,
0.4423828125,
0.3525390625,
0.457275390625,
0.162109375,
0.457275390625
],
[
0.1435546875,
0.468505859375,
0.8759765625,
0.468505859375,
0.8759765625,
0.482177734375,
0.1435546875,
0.482177734375
],
[
0.12005615234375,
0.482666015625,
0.79248046875,
0.482666015625,
0.79248046875,
0.496337890625,
0.12005615234375,
0.496337890625
],
[
0.1175537109375,
0.5048828125,
0.8388671875,
0.5029296875,
0.8388671875,
0.52099609375,
0.1175537109375,
0.5224609375
],
[
0.393798828125,
0.52783203125,
0.587890625,
0.53515625,
0.5859375,
0.56005859375,
0.39208984375,
0.552734375
],
[
0.5810546875,
0.54150390625,
0.59033203125,
0.54150390625,
0.59033203125,
0.5458984375,
0.5810546875,
0.5458984375
],
[
0.11419677734375,
0.5634765625,
0.884765625,
0.57080078125,
0.884765625,
0.59716796875,
0.11419677734375,
0.59033203125
],
[
0.7119140625,
0.572265625,
0.73046875,
0.572265625,
0.73046875,
0.57666015625,
0.7119140625,
0.57666015625
],
[
0.118408203125,
0.58837890625,
0.424072265625,
0.5888671875,
0.424072265625,
0.60693359375,
0.118408203125,
0.60546875
],
[
0.408935546875,
0.619140625,
0.58447265625,
0.619140625,
0.58447265625,
0.63427734375,
0.408935546875,
0.63427734375
],
[
0.3359375,
0.626953125,
0.4072265625,
0.626953125,
0.4072265625,
0.64111328125,
0.3359375,
0.64111328125
],
[
0.399658203125,
0.6298828125,
0.441650390625,
0.63818359375,
0.435791015625,
0.65380859375,
0.393798828125,
0.6455078125
],
[
0.48779296875,
0.63818359375,
0.5458984375,
0.63818359375,
0.5458984375,
0.6494140625,
0.48779296875,
0.6494140625
],
[
0.33154296875,
0.66015625,
0.6591796875,
0.64794921875,
0.66015625,
0.6728515625,
0.333251953125,
0.68505859375
],
[
0.5576171875,
0.6728515625,
0.60693359375,
0.6728515625,
0.60693359375,
0.6884765625,
0.5576171875,
0.6884765625
],
[
0.1435546875,
0.69775390625,
0.8759765625,
0.69677734375,
0.8759765625,
0.71142578125,
0.1435546875,
0.7119140625
],
[
0.118408203125,
0.71044921875,
0.87744140625,
0.71142578125,
0.87744140625,
0.728515625,
0.118408203125,
0.72802734375
],
[
0.12005615234375,
0.72705078125,
0.2626953125,
0.72705078125,
0.2626953125,
0.74072265625,
0.12005615234375,
0.74072265625
],
[
0.12005615234375,
0.75048828125,
0.880859375,
0.75048828125,
0.880859375,
0.76416015625,
0.12005615234375,
0.76416015625
],
[
0.12005615234375,
0.76416015625,
0.87744140625,
0.76416015625,
0.87744140625,
0.77783203125,
0.12005615234375,
0.77783203125
],
[
0.12005615234375,
0.779296875,
0.87890625,
0.779296875,
0.87890625,
0.7919921875,
0.12005615234375,
0.7919921875
],
[
0.118408203125,
0.79150390625,
0.87890625,
0.79345703125,
0.87890625,
0.80810546875,
0.118408203125,
0.8056640625
],
[
0.1175537109375,
0.80419921875,
0.880859375,
0.8046875,
0.880859375,
0.822265625,
0.1175537109375,
0.82177734375
],
[
0.12005615234375,
0.82177734375,
0.8759765625,
0.82177734375,
0.8759765625,
0.8359375,
0.12005615234375,
0.8359375
],
[
0.12005615234375,
0.85791015625,
0.646484375,
0.85791015625,
0.646484375,
0.87255859375,
0.12005615234375,
0.87255859375
],
[
0.12176513671875,
0.8896484375,
0.60546875,
0.8896484375,
0.60546875,
0.90283203125,
0.12176513671875,
0.90283203125
],
[
0.490234375,
0.93603515625,
0.5087890625,
0.93603515625,
0.5087890625,
0.95068359375,
0.490234375,
0.95068359375
]
] |
[
"4在应用程序。C.2。对于泛化分析，我们要求可实现性假设几乎肯定适用于从数据分布采样的所有训",
"练集",
"我们注意到优化和泛化的界限与逻辑回归（Soudryetal.，2018：JiandTelgarsky，2018：Shamir，2021",
"的类似界限相同。然而，为了使这些界限有效，我们需要足够多的头以及存在一组适当的目标参数",
"如定理条件中所述。也就是说，这些条件与可实现性条件有关，这保证了在初始化附近获得较小的训练",
"误差。下一个假设将这些条件形式化。",
"Assumption 2 (Realizability).There erist non-increasing functions g : R++R+ and go :R+→R+",
"such that Ve > O , there erists model parameters S(e) E RH(dT+d’) for which: (i) the empirical loss over n",
"data samples satisfies L(0(e)≤e, (i) [0(e) -(0)| ≤go(e) , and, (iu) [0(e) 2,o ≤g(e).",
"通过此假设，我们可以将上述定理的结果专门用于特定的数据设置；请参阅应用程序中的推论3和4",
"。B.5和C.2。在下一节中，我们将进一步展示如何满足可实现性假设。",
"4.3Primitive conditions for checking realizability",
"在这里，我们在数据和初始化上引人了一组更原始，更直接的检查条件，以确保可实现性假设2成立",
"Definition 1 (Good initialization). We say (0) = concat(o(0),...,o(0) is a good initialization with",
"respect to training data (X,, yi)ien) provided the following three properties hold",
"P1. Parameter L2,oo -bound: There erists parameter B2 ≥1 such that Vh e [H] it holds o(02 ≤ B2",
"P2. Model bound: There eaists parameter B ≥1 such that Vie[n] it holds (X;@(0)≤ Ba",
"P3. NTK separability: There erists , e RH(dT+d°) and >0 such that [5ll = V2 and Vie [n] , it holds",
"Yi (v4[X,; 0(0)[,0.) ≥%-",
"道具。附录中的7表明，从good初始化开始，我们总能找到满足可实现性的θ(e）假设2提供了足",
"够多的磁头。因此，给定good初始化，我们可以立即应用定理1和2来获得以下具体界限",
"Corollary 2 (General bounds under good initialization). Suppose good initialization (o) and let",
"VH≥dT1/2 R B& go()",
"",
"where go(K) = 2Bet)ps(K) . Purther fia any step-size n ≤1 ^ 1/p(K) ^ 1og(ie) where p(K) ≥",
"4",
"d3/2 T3/2 R13 go()4. Then, it holds that",
"25(2B+log(K))2",
"L(0(K)) <",
"≤K",
"42nK",
"E[ (@(K) - L(G(K) ≤17(2Bs + log(K)2",
"nn",
"考虑KGD步骤后的训练损失：假设B=Ok(1)和=Ok(1)，然后选择n=Ok(1)，则推论",
"保证训练损失为k(），前提是磁头的对数H=2(log（K)。此外，在K~nGD步骤之后，预期",
"测试损耗为O(1/n)",
"Remark 1. The last two conditions (P2,P3) for good initialization are similar to the conditions needed",
"in (Taheri and Thrampoulidis, 2023; Ji and Telgarsky, 2020; Nitanda et al., 2019) for analysis of two",
"layer MLPs. Compared to (Ji and Telgarsky, 2020; Nitanda et al., 2019) which assume random Gaussian",
"initialization (o) , and similar to Taheri and Thrampoulidis (2023) the NTK separability assumption",
"(P3) can potentially accommodate deterministic (o) . Condition (Pi) appears because we allow training",
"both layers of the model. Specifically the L2,oo norm originates from the Hessian bounds in Corollary 1",
"5Application to tokenized-mixture model",
"我们现在通过一个示例演示我们的结果如何应用于特定的数据模型",
"5"
] |
{
"arxiv_id": "2310.12680",
"language": "zh",
"page_idx": 4
}
|
||
1509.04783-zh-3
|
1509.04783-zh-3.jpg
|
<table>
<tbody>
<tr>
<td>Algorithm Pairwise based 1learn deco ition</td>
<td>learning ased 112</td>
<td> </td>
</tr>
<tr>
<td>Input$:\{\phi(\mathcal{X}_{m_{i},m_{j}})\}_{\forall m_{i}\neq m_{j}\in\{1,\ldots,M\}},\{y_{m_{i}}\}_{\forall m_{i}\in1}$</td>
<td>$y_{m_{i}}\}_{\forall m_{i}\in1,\ldots,M}$</td>
<td> </td>
</tr>
<tr>
<td>Output: $\{ \boldsymbol{w}_{m_{i}, m}$ ${}_{j}\}_{m_{i}\neq m_{j}\in\{1,\ldots,M\}},\boldsymbol{w}_{h},\beta$ 1</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Initialize $\beta \leftarrow 1$ , $w_{h}\leftarrow 1$ , $w_{m_{i}, m_{j}}\leftarrow 1$ ;</td>
<td> </td>
<td> </td>
</tr>
</tbody>
</table>
方法中需要学习的参数数量从$\left(\prod_{m=1}^M|z_m|+|h|\right)$到
$\left(\sum_{m_i\neq m_j}|z_{m_i}||z_{m_j}|+|h|\right)$大大减少了
$\begin{array}{rcl}\text{令}&\phi(\mathcal{X}_{m_i,m_j})&\stackrel{\Delta}{\equiv}&p(z_{m_i}|\mathcal{X}_{m_i},h)p(z_{m_j}|\mathcal{X}_{m_j},h)&\in\end{array}$ $\mathbb{R}^{(|z_{m_{i}}||z_{m_{j}}|)\times|h|}$为视图$m_i$ 和$m_j$ 之间的成对视觉单词矩阵，其中 $\mathcal{X}_m_i,m_j=\{\mathcal{X}_{m_i},\mathcal{X}_{m_j}\}$。还让$\boldsymbol w_{m_i,m_j}\triangleq$ $p( y_{m_i}$ = $y_{m_j}| z_{m_i}, z_{m_j}$ ) $\in$ $\mathbb{R} ^{| z_{m_i}| | z_{m_j}| }$ 和 $\beta$ $\triangleq$ $p( y_1$ = $F( gm_{i}$ $gm_{j}| • m_{i}, \cdotp m_{j}) \subset \mathbb{R} ^{| z_{m_{i}}| | z_{m_{j}}| }$ $\cdots = y_{M}| y_{m_{i}}= y_{m_{j}}) \in \mathbb{R} ^{| z_{m_{i}}| | z_{m_{j}}| }$。然 后 基 于 方 程 5, 我们可以重写方程 4 如下：
$$\tilde{f}(\mathcal{X}_{1,\cdots,M})=\sum_{m_i\neq m_j}\beta_{m_i,m_j}\boldsymbol{w}_{m_i,m_j}^T\phi(\mathcal{X}_{m_i,m_j})\boldsymbol{w}_h,$$
其中$\beta_{m_i,m_j}$表示视图对的$\beta$中的条目。
为了学习我们在方程式 6 中的模型参数，我们提出了如下两种学习方法，即 multi-view training 和double-view training :
Multi-view training:
$$\operatorname*{min}_{\{\boldsymbol{w}_{m_{i},m_{j}}\},}\frac{\lambda_{1}}{2}\sum_{m_{i}\neq m_{j}}\|\boldsymbol{w}_{m_{i},m_{j}}\|_{2}^{2}+\frac{\lambda_{2}}{2}\|\boldsymbol{w}_{h}\|_{2}^{2}+\frac{\lambda_{3}}{2}\|\boldsymbol{\beta}\|_{2}^{2}\\+\sum_{k=1}^{N}\ell\left(y_{1,\cdots,M}^{(k)},\tilde{f}(\mathcal{X}_{1,\cdots,M}^{(k)})\right),\:\mathrm{s.t.}\:\boldsymbol{\beta}\geq\boldsymbol{0},\quad(7)$$
Double-view training:
$$\min_{\{\boldsymbol{w}_{m_{i},m_{j}}\},}\frac{\lambda_{1}}{2}\sum_{m_{i}\neq m_{j}}\|\boldsymbol{w}_{m_{i},m_{j}}\|_{2}^{2}+\frac{\lambda_{2}}{2}\|\boldsymbol{w}_{h}\|_{2}^{2}+\frac{\lambda_{3}}{2}\|\boldsymbol{\beta}\|_{2}^{2}\\+\sum_{k=1}^{N}\sum_{m_{i}\neq m_{j}}\ell\left(y_{m_{i},m_{j}}^{(k)},\tilde{f}(\mathcal{X}_{m_{i},m_{j}}^{(k)})\right),\:\mathrm{s.t.}\:\boldsymbol{\beta}\geq\boldsymbol{0},\:(8)$$
其中$\forall k,y_{m_i,m_j}^{(k)}=1$如果在$k$组中两个人的标签$y_m_i=$ $y_{m_j}$ 成立；否则，0。这里，$\geq$表示元素式$\geq$运算符。两种训练都可以使用标准 SVM 求解器的交替优化来完成。仍然保证局部最优。对于两视图场景，两种训练方法本质上是相同的，并且通常与视图数量成二次方
比例。如果我们将所有视图组织为循环图，则线性可伸缩性也是可能的。这两种训练方法的差异来自损失函数，其中在多视图训练中《测量组(i.e.多视图)损失， 而在双视图训练中$\ell$测量成对视图损失。我们的算法总结在 Alg。1
# 3. Experiments
我们评估了我们在人 Re-ID 和亲属关系验证任务上的方法以及在基准数据集上的最新方法。所有实验均使用标准培训/测试协议。对于每种比较方法，我们要么引用论文的原始结果(在表中由($\cdot)^{*} 表示),要么从$ 发布的代码中计算。我们的结果报告为 3 次试验的平均值
对于每个实验，我们选择与其他方法相同或相似的低级特征(请参阅小节中的详细信息)进行公平比较。我们对图像进行密集采样，以生成每个像素的低级局部特征。然后，我们使用 K-Means 来构建每个视图中具有约$2\times10^4$个随机选择的功能的可视词汇表。此外，基于欧几里得距离，每个局部特征都被量化为这些视觉单词之一。请注意，可以采用更复杂的特征选择方法来产生更好的性能，但是为了计算效率和泛化能力， 我们不会对该组件进行微调
我们将方程式 3 和 LIBLINEAR [7] 的棋盘距离用作具有铰链损失的 SVM 求解器。我们随机生成关于$3\times10^4$的训练样本来学习模型参数$w$的。正则化参数由交叉验证确定。
## $3. 1. \textbf{ Person Re- identification}$
对于性能度量，我们采用标准的累积匹配特征(CMC)曲线，该曲线将识别率显示为等级的函数。rank- $r$的识别率是与 rank- $r$或更高的相应图库实体正确匹配的查询的比例
对于具有多个摄像机视图的任务，我们遵循[5]比较两个摄像机视图下的结果。将来自多个视图的结果视为高维张量，每个视图一个维度。为了从多视图结果中预测成对匹配(e.g.从视图 1、2 和 3 的关节的预测结果中识别摄像机视图 1 和视图 2 之间的匹配),我们可以求和或找到额外维度上的最大值。交叉验证用于为每个数据集选择更好的方法。
3.1.1 Two Camera Views
两个视图之间的人 Re-ID 是最简单的方案。我们在VIPeR [13] 和 CUHKCampus [35] 数据集上测试了我们的方法。我们从图像中的每个 5 Re-ID 像素补丁中提取 672-dimColor+SIFT 1 向量作为低级特征。我们遵循[35]中的两个数据集的实验设置
我们的比较结果列于表 1 。正如我们所看到的，在VIPeR上，来自[36]的“Mid-level+LADF”是当前最好的方法，它利用更具判别性的中级过滤器作为特征和强大的分类器，来自[31]的“SCNCD $_{final}$”是第$\overline{{^{1}\mathrm{We}\:\mathrm{downloaded}\:\mathrm{th}}}$e code from https://github.com/ Robert0812/salience\_match .
|
[
[
0.08087158203125,
0.091552734375,
0.455810546875,
0.0921630859375,
0.455810546875,
0.106689453125,
0.08087158203125,
0.1060791015625
],
[
0.5048828125,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.5048828125,
0.1060791015625
],
[
0.0767822265625,
0.106689453125,
0.4697265625,
0.10986328125,
0.4697265625,
0.1275634765625,
0.0767822265625,
0.125
],
[
0.5048828125,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.5048828125,
0.121826171875
],
[
0.1396484375,
0.125,
0.2279052734375,
0.125,
0.2279052734375,
0.1363525390625,
0.1396484375,
0.1363525390625
],
[
0.5048828125,
0.125,
0.8896484375,
0.125,
0.8896484375,
0.1351318359375,
0.5048828125,
0.1351318359375
],
[
0.0767822265625,
0.136962890625,
0.351318359375,
0.1383056640625,
0.351318359375,
0.1558837890625,
0.0767822265625,
0.1546630859375
],
[
0.5048828125,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.50341796875,
0.1514892578125,
0.609375,
0.1514892578125,
0.609375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.080078125,
0.1610107421875,
0.36279296875,
0.1610107421875,
0.36279296875,
0.175537109375,
0.080078125,
0.175537109375
],
[
0.08087158203125,
0.1767578125,
0.1217041015625,
0.1767578125,
0.1217041015625,
0.1875,
0.08087158203125,
0.1875
],
[
0.50244140625,
0.1817626953125,
0.6396484375,
0.1817626953125,
0.6396484375,
0.1956787109375,
0.50244140625,
0.1956787109375
],
[
0.10455322265625,
0.189453125,
0.46728515625,
0.189453125,
0.46728515625,
0.20263671875,
0.10455322265625,
0.20263671875
],
[
0.10455322265625,
0.20263671875,
0.41259765625,
0.20263671875,
0.41259765625,
0.216552734375,
0.10455322265625,
0.216552734375
],
[
0.52392578125,
0.2083740234375,
0.88818359375,
0.2083740234375,
0.88818359375,
0.2183837890625,
0.52392578125,
0.2183837890625
],
[
0.10455322265625,
0.21533203125,
0.468017578125,
0.2171630859375,
0.468017578125,
0.2310791015625,
0.10455322265625,
0.2291259765625
],
[
0.50341796875,
0.2215576171875,
0.890625,
0.2215576171875,
0.890625,
0.23486328125,
0.50341796875,
0.23486328125
],
[
0.10296630859375,
0.2291259765625,
0.46728515625,
0.2310791015625,
0.46728515625,
0.24560546875,
0.10296630859375,
0.24365234375
],
[
0.50634765625,
0.238037109375,
0.8896484375,
0.238037109375,
0.8896484375,
0.248779296875,
0.50634765625,
0.248779296875
],
[
0.080078125,
0.246826171875,
0.1829833984375,
0.2474365234375,
0.1829833984375,
0.261962890625,
0.080078125,
0.261474609375
],
[
0.5048828125,
0.251953125,
0.890625,
0.251953125,
0.890625,
0.26513671875,
0.5048828125,
0.26513671875
],
[
0.08087158203125,
0.261962890625,
0.340576171875,
0.261962890625,
0.340576171875,
0.27978515625,
0.08087158203125,
0.27978515625
],
[
0.50341796875,
0.26513671875,
0.8896484375,
0.266357421875,
0.8896484375,
0.281005859375,
0.50341796875,
0.27978515625
],
[
0.50341796875,
0.2822265625,
0.53759765625,
0.2822265625,
0.53759765625,
0.294189453125,
0.50341796875,
0.294189453125
],
[
0.52392578125,
0.296630859375,
0.890625,
0.296630859375,
0.890625,
0.31005859375,
0.52392578125,
0.31005859375
],
[
0.08251953125,
0.30810546875,
0.46728515625,
0.30810546875,
0.46728515625,
0.32568359375,
0.08251953125,
0.32568359375
],
[
0.5048828125,
0.3125,
0.88818359375,
0.3125,
0.88818359375,
0.32568359375,
0.5048828125,
0.32568359375
],
[
0.08416748046875,
0.330810546875,
0.361083984375,
0.328857421875,
0.361083984375,
0.349853515625,
0.08416748046875,
0.3515625
],
[
0.5048828125,
0.326904296875,
0.890625,
0.326904296875,
0.890625,
0.34033203125,
0.5048828125,
0.34033203125
],
[
0.5048828125,
0.34228515625,
0.8896484375,
0.34228515625,
0.8896484375,
0.35595703125,
0.5048828125,
0.35595703125
],
[
0.09722900390625,
0.3515625,
0.4453125,
0.354248046875,
0.4453125,
0.375,
0.09722900390625,
0.37255859375
],
[
0.43701171875,
0.35986328125,
0.465576171875,
0.35986328125,
0.465576171875,
0.36865234375,
0.43701171875,
0.36865234375
],
[
0.50341796875,
0.356689453125,
0.890625,
0.356689453125,
0.890625,
0.37060546875,
0.50341796875,
0.37060546875
],
[
0.0767822265625,
0.367431640625,
0.471435546875,
0.371826171875,
0.471435546875,
0.392578125,
0.0767822265625,
0.38818359375
],
[
0.50341796875,
0.37255859375,
0.890625,
0.37255859375,
0.890625,
0.386474609375,
0.50341796875,
0.386474609375
],
[
0.0767822265625,
0.391357421875,
0.47216796875,
0.386474609375,
0.47216796875,
0.410888671875,
0.0767822265625,
0.416748046875
],
[
0.50244140625,
0.386474609375,
0.890625,
0.386962890625,
0.890625,
0.401611328125,
0.50244140625,
0.40087890625
],
[
0.0767822265625,
0.40771484375,
0.47216796875,
0.40771484375,
0.47216796875,
0.431884765625,
0.0767822265625,
0.431884765625
],
[
0.50341796875,
0.40283203125,
0.890625,
0.40283203125,
0.890625,
0.416748046875,
0.50341796875,
0.416748046875
],
[
0.5048828125,
0.419189453125,
0.69775390625,
0.419189453125,
0.69775390625,
0.429931640625,
0.5048828125,
0.429931640625
],
[
0.10296630859375,
0.427978515625,
0.47216796875,
0.4267578125,
0.47216796875,
0.448974609375,
0.10296630859375,
0.449462890625
],
[
0.52392578125,
0.43310546875,
0.890625,
0.43310546875,
0.890625,
0.4462890625,
0.52392578125,
0.4462890625
],
[
0.08087158203125,
0.447509765625,
0.271240234375,
0.447509765625,
0.271240234375,
0.4609375,
0.08087158203125,
0.4609375
],
[
0.50634765625,
0.449462890625,
0.88818359375,
0.449462890625,
0.88818359375,
0.460205078125,
0.50634765625,
0.460205078125
],
[
0.50341796875,
0.462158203125,
0.89208984375,
0.462158203125,
0.89208984375,
0.47607421875,
0.50341796875,
0.47607421875
],
[
0.08905029296875,
0.472900390625,
0.468017578125,
0.474853515625,
0.468017578125,
0.49560546875,
0.08905029296875,
0.49365234375
],
[
0.50244140625,
0.4765625,
0.6259765625,
0.477783203125,
0.6259765625,
0.492431640625,
0.50244140625,
0.4912109375
],
[
0.1854248046875,
0.4912109375,
0.236083984375,
0.49365234375,
0.2344970703125,
0.50830078125,
0.1846923828125,
0.505859375
],
[
0.50244140625,
0.501953125,
0.7255859375,
0.501953125,
0.7255859375,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.080078125,
0.517578125,
0.363525390625,
0.51806640625,
0.363525390625,
0.53271484375,
0.080078125,
0.5322265625
],
[
0.525390625,
0.5263671875,
0.8896484375,
0.5263671875,
0.8896484375,
0.537109375,
0.525390625,
0.537109375
],
[
0.101318359375,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.101318359375,
0.54736328125
],
[
0.5009765625,
0.53857421875,
0.890625,
0.5390625,
0.890625,
0.5537109375,
0.5009765625,
0.55322265625
],
[
0.08087158203125,
0.54736328125,
0.468017578125,
0.548828125,
0.468017578125,
0.56298828125,
0.08087158203125,
0.5625
],
[
0.50244140625,
0.5537109375,
0.89208984375,
0.5546875,
0.89208984375,
0.5693359375,
0.50244140625,
0.568359375
],
[
0.07843017578125,
0.5625,
0.2484130859375,
0.564453125,
0.2484130859375,
0.5791015625,
0.07843017578125,
0.5771484375
],
[
0.50341796875,
0.568359375,
0.6650390625,
0.568359375,
0.6650390625,
0.58203125,
0.50341796875,
0.58203125
],
[
0.0906982421875,
0.59033203125,
0.2344970703125,
0.59130859375,
0.2344970703125,
0.60595703125,
0.0906982421875,
0.60498046875
],
[
0.52392578125,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.52392578125,
0.5986328125
],
[
0.158447265625,
0.60693359375,
0.330810546875,
0.6142578125,
0.330078125,
0.6396484375,
0.1568603515625,
0.6318359375
],
[
0.326904296875,
0.60693359375,
0.416748046875,
0.6171875,
0.41259765625,
0.6396484375,
0.32275390625,
0.62890625
],
[
0.5048828125,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.6142578125,
0.5048828125,
0.6142578125
],
[
0.416748046875,
0.609375,
0.4755859375,
0.61865234375,
0.4697265625,
0.63818359375,
0.410888671875,
0.62890625
],
[
0.11273193359375,
0.619140625,
0.1478271484375,
0.619140625,
0.1478271484375,
0.630859375,
0.11273193359375,
0.630859375
],
[
0.5048828125,
0.61572265625,
0.8896484375,
0.61572265625,
0.8896484375,
0.62890625,
0.5048828125,
0.62890625
],
[
0.0931396484375,
0.62744140625,
0.1846923828125,
0.625,
0.1854248046875,
0.642578125,
0.09478759765625,
0.64599609375
],
[
0.2034912109375,
0.62744140625,
0.2115478515625,
0.62744140625,
0.2115478515625,
0.6318359375,
0.2034912109375,
0.6318359375
],
[
0.339111328125,
0.62255859375,
0.351318359375,
0.6318359375,
0.339111328125,
0.6416015625,
0.326904296875,
0.6318359375
],
[
0.416748046875,
0.62744140625,
0.43310546875,
0.625,
0.43701171875,
0.63623046875,
0.420654296875,
0.64013671875
],
[
0.178955078125,
0.63525390625,
0.2320556640625,
0.63525390625,
0.2320556640625,
0.6494140625,
0.178955078125,
0.6494140625
],
[
0.5048828125,
0.630859375,
0.890625,
0.630859375,
0.890625,
0.64453125,
0.5048828125,
0.64453125
],
[
0.111083984375,
0.6416015625,
0.146240234375,
0.6416015625,
0.146240234375,
0.65283203125,
0.111083984375,
0.65283203125
],
[
0.5048828125,
0.64599609375,
0.8896484375,
0.64599609375,
0.8896484375,
0.6591796875,
0.5048828125,
0.6591796875
],
[
0.11767578125,
0.6572265625,
0.13232421875,
0.6572265625,
0.13232421875,
0.66748046875,
0.11767578125,
0.66748046875
],
[
0.0931396484375,
0.66650390625,
0.2646484375,
0.66845703125,
0.2646484375,
0.68994140625,
0.0931396484375,
0.68798828125
],
[
0.50341796875,
0.6611328125,
0.890625,
0.6611328125,
0.890625,
0.6748046875,
0.50341796875,
0.6748046875
],
[
0.15283203125,
0.67041015625,
0.1829833984375,
0.67041015625,
0.1829833984375,
0.6748046875,
0.15283203125,
0.6748046875
],
[
0.29736328125,
0.671875,
0.39208984375,
0.671875,
0.39208984375,
0.68310546875,
0.29736328125,
0.68310546875
],
[
0.44677734375,
0.67041015625,
0.468017578125,
0.67041015625,
0.468017578125,
0.68505859375,
0.44677734375,
0.68505859375
],
[
0.50341796875,
0.67529296875,
0.70751953125,
0.67529296875,
0.70751953125,
0.689453125,
0.50341796875,
0.689453125
],
[
0.11273193359375,
0.68603515625,
0.136474609375,
0.68603515625,
0.136474609375,
0.69775390625,
0.11273193359375,
0.69775390625
],
[
0.25732421875,
0.6748046875,
0.28759765625,
0.6748046875,
0.28759765625,
0.681640625,
0.25732421875,
0.681640625
],
[
0.0906982421875,
0.70068359375,
0.2467041015625,
0.70263671875,
0.2467041015625,
0.71728515625,
0.0906982421875,
0.71533203125
],
[
0.50341796875,
0.70849609375,
0.69921875,
0.70849609375,
0.69921875,
0.7216796875,
0.50341796875,
0.7216796875
],
[
0.328369140625,
0.71826171875,
0.4150390625,
0.72607421875,
0.41259765625,
0.74853515625,
0.324462890625,
0.7412109375
],
[
0.416748046875,
0.7197265625,
0.476318359375,
0.72900390625,
0.471435546875,
0.74755859375,
0.410888671875,
0.73779296875
],
[
0.188720703125,
0.72265625,
0.33251953125,
0.724609375,
0.33251953125,
0.74755859375,
0.188720703125,
0.74560546875
],
[
0.111083984375,
0.72802734375,
0.1405029296875,
0.72998046875,
0.1396484375,
0.7421875,
0.1094970703125,
0.74072265625
],
[
0.1641845703125,
0.72265625,
0.1829833984375,
0.72265625,
0.1829833984375,
0.73486328125,
0.1641845703125,
0.73486328125
],
[
0.5048828125,
0.73095703125,
0.890625,
0.73095703125,
0.890625,
0.74560546875,
0.5048828125,
0.74560546875
],
[
0.09234619140625,
0.73779296875,
0.1829833984375,
0.734375,
0.1846923828125,
0.75244140625,
0.0931396484375,
0.7568359375
],
[
0.3349609375,
0.73779296875,
0.34716796875,
0.73779296875,
0.34716796875,
0.751953125,
0.3349609375,
0.751953125
],
[
0.420654296875,
0.7392578125,
0.43310546875,
0.7392578125,
0.43310546875,
0.75048828125,
0.420654296875,
0.75048828125
],
[
0.5048828125,
0.7470703125,
0.890625,
0.7470703125,
0.890625,
0.76025390625,
0.5048828125,
0.76025390625
],
[
0.111083984375,
0.751953125,
0.146240234375,
0.751953125,
0.146240234375,
0.76318359375,
0.111083984375,
0.76318359375
],
[
0.5048828125,
0.76220703125,
0.8896484375,
0.76220703125,
0.8896484375,
0.77587890625,
0.5048828125,
0.77587890625
],
[
0.1192626953125,
0.76806640625,
0.131591796875,
0.76953125,
0.1298828125,
0.77783203125,
0.1168212890625,
0.77587890625
],
[
0.50341796875,
0.7763671875,
0.89208984375,
0.7763671875,
0.89208984375,
0.79052734375,
0.50341796875,
0.79052734375
],
[
0.1446533203125,
0.77783203125,
0.336669921875,
0.77783203125,
0.336669921875,
0.798828125,
0.1446533203125,
0.798828125
],
[
0.1829833984375,
0.74365234375,
0.2296142578125,
0.74560546875,
0.2279052734375,
0.76025390625,
0.181396484375,
0.75830078125
],
[
0.09722900390625,
0.78076171875,
0.143798828125,
0.78076171875,
0.143798828125,
0.79541015625,
0.09722900390625,
0.79541015625
],
[
0.336669921875,
0.78076171875,
0.46728515625,
0.78076171875,
0.46728515625,
0.79443359375,
0.336669921875,
0.79443359375
],
[
0.111083984375,
0.796875,
0.1805419921875,
0.7998046875,
0.1805419921875,
0.81103515625,
0.111083984375,
0.80810546875
],
[
0.50341796875,
0.79248046875,
0.76953125,
0.79248046875,
0.76953125,
0.80615234375,
0.50341796875,
0.80615234375
],
[
0.521484375,
0.80615234375,
0.89208984375,
0.806640625,
0.89208984375,
0.82080078125,
0.521484375,
0.8193359375
],
[
0.080078125,
0.82470703125,
0.461669921875,
0.8271484375,
0.461669921875,
0.84130859375,
0.080078125,
0.8388671875
],
[
0.5048828125,
0.8212890625,
0.89208984375,
0.8212890625,
0.89208984375,
0.83544921875,
0.5048828125,
0.83544921875
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84033203125,
0.468017578125,
0.85498046875,
0.080078125,
0.85595703125
],
[
0.50341796875,
0.8369140625,
0.890625,
0.8369140625,
0.890625,
0.85107421875,
0.50341796875,
0.85107421875
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.50244140625,
0.85107421875,
0.89208984375,
0.85107421875,
0.89208984375,
0.86865234375,
0.50244140625,
0.86865234375
],
[
0.08251953125,
0.87353515625,
0.46728515625,
0.87353515625,
0.46728515625,
0.8837890625,
0.08251953125,
0.8837890625
],
[
0.5185546875,
0.87353515625,
0.8896484375,
0.876953125,
0.8896484375,
0.89013671875,
0.5185546875,
0.88720703125
],
[
0.08087158203125,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.6845703125,
0.88818359375,
0.6845703125,
0.8994140625,
0.50244140625,
0.8984375
]
] |
[
"Algorithm 1Pairwise decomposition based learning",
"比例。如果我们将所有视图组织为循环图，则线性可伸",
"Input: [o(Xm,.m,))vmitmje[1..M),(ym,)vmie...M,",
"缩性也是可能的。这两种训练方法的差异来自损失函",
"1, 2,3 ≥ 0",
"数，其中在多视图训练中l测量组（i.e.多视图）损失",
"Output: [wmim,Imi+mje{,..M),Wh,β",
"而在双视图训练中l测量成对视图损失。我们的算法",
"总结在Alg。1",
"Initialize β←1, Wh←1, Wmi,m, ←1;",
"repeat",
"3.Experiments",
"Solve [wmi,m) in Eq. 7 (multi-view training) or Eq.",
"8 (double-view training) by fixing β and wh ;",
"我们评估了我们在人Re-ID和亲属关系验证任务上",
"Solve wh in Eq. 7 or Eq. 8 by fixing β and [wmi,m,] ;",
"的方法以及在基准数据集上的最新方法。所有实验均",
"Solve β in Eq. 7 or Eq. 8 by fixing [wmim,) and wh .",
"使用标准培训/测试协议。对于每种比较方法，我们要",
"until Converge;",
"么引用论文的原始结果（在表中由（)*表示），要么从",
"return [Wm.m,Im#mje{...M),Wh, β",
"发布的代码中计算。我们的结果报告为3次试验的平",
"均值",
"对于每个实验，我们选择与其他方法相同或相似的",
"方法中需要学习的参数数量从（1Im=11zml+[hl）到",
"低级特征（请参阅小节中的详细信息）进行公平比较",
"(Zmzmg|zmllzmsl+[hl)大大减少了",
"我们对图像进行密集采样，以生成每个像素的低级局",
"部特征。然后，我们使用K-Means来构建每个视图中",
"令p(Xm,m）≤p(zm:lm,h)p(zm,lXmg,h)",
"e",
"具有约2×104个随机选择的功能的可视词汇表。此",
"R(1zmllam,D)x/nl为视图m；和m之间的成对视觉单",
"外，基于欧几里得距离，每个局部特征都被量化为这些",
"词矩阵，其中Xmm，=[m，m]。还让wmm，会",
"视觉单词之一。请注意，可以采用更复杂的特征选择方",
"p(ym = Umjlzma,zm)) eRl2mllzmjl 和 β p(y1 =",
"法来产生更好的性能，但是为了计算效率和泛化能力，",
"我们不会对该组件进行微调",
"=yMlym，=9m）=Rl2mllzmjl。然后基于方程5，",
"我们将方程式3和LIBLINEAR[7]的棋盘距离用",
"我们可以重写方程4如下：",
"作具有铰链损失的SVM求解器。我们随机生成关于",
"3×104的训练样本来学习模型参数的。正则化参数",
"F(X...M) = Bmmj wmm,o(Xmm,)wh, (6)",
"由交叉验证确定。",
"mitmj",
"3.1.Person Re-identification",
"其中βmi.m，表示视图对的β中的条目。",
"对于性能度量，我们采用标准的累积匹配特征",
"为了学习我们在方程式6中的模型参数，我们提",
"（CMC）曲线，该曲线将识别率显示为等级的函数。",
"出了如下两种学习方法，即multi-viewtraining和",
"rank-r的识别率是与rank-r或更高的相应图库实体",
"double-view training:",
"正确匹配的查询的比例",
"Multi-view training:",
"对于具有多个摄像机视图的任务，我们遵循[5]比较",
"[wm,ml+",
"+",
"两个摄像机视图下的结果。将来自多个视图的结果视",
"",
"min",
"为高维张量，每个视图一个维度。为了从多视图结果中",
"(wmamjh, 2",
"",
"2",
"2",
"mitmj",
"预测成对匹配（e.g.从视图1、2和3的关节的预测结",
"Wh,β",
"果中识别摄像机视图1和视图2之间的匹配），我们可",
"N",
"+Ee (... (a)",
"以求和或找到额外维度上的最大值。交叉验证用于为",
"",
" .s.t.β>0",
"(7)",
"每个数据集选择更好的方法。",
"k=1",
"xr",
"Double-view training:",
"3.1.1Two Camera Views",
"兴w+",
"",
"Z Iwm,m,ll2 +",
"min",
"11",
"两个视图之间的人Re-ID是最简单的方案。我们在",
"(um.m), 2",
"2",
"2",
"VIPeR[13]和CUHKCampus[35]数据集上测试了我",
"Wh,β",
"们的方法。我们从图像中的每个5Re-ID像素补丁中",
"N",
"提取672-dimColor+SIFT1向量作为低级特征。我们",
"E e (um)m, F(xm)m)",
"mitm",
"+>",
") ,s.t. β≥0, (8)",
"k=lm+m",
"遵循[35]中的两个数据集的实验设置",
"我们的比较结果列于表1。正如我们所看到的，在",
"其中Vk,ymmz=1如果在k组中两个人的标签ym",
"VIPeR上，来自[36]的“Mid-level+LADF”是当前最",
"m，成立；否则，0。这里，≥表示元素式≥运算符。",
"好的方法，它利用更具判别性的中级过滤器作为特征",
"两种训练都可以使用标准SVM求解器的交替优化来",
"和强大的分类器，来自[31]的“SCNCD final”是第",
"完成。仍然保证局部最优。对于两视图场景，两种训练",
"IWedownloadedthecodefromhttps://github.com/",
"方法本质上是相同的，并且通常与视图数量成二次方",
"Robert0812/saliencematch."
] |
{
"arxiv_id": "1509.04783",
"language": "zh",
"page_idx": 3
}
|
||
2104.11116-zh-0
|
2104.11116-zh-0.jpg
|
# $\textbf{Pose- Controllable Talking Face Generation by}$ Implicitly Modularized Audio-Visual Representation
Hang Zhou $^{1}$,Yasheng Sun $^{2,3}$, Wayne Wu $^{2,4}$,Chen Change Loy $^{4}$ ,Xiaogang Wang $^{1}$ ,Ziwei Liu $^{4\boxtimes}$
$^{1}$ CUHK - SenseTime Joint Lab, The Chinese University of Hong Kong $^{2}$ SenseTime Research
$^{3}$ Tokyo Institute of Technology $^{4}$ S-Lab, Nanyang Technological University
{ zhouhang@link,xgwang@ee} .cuhk.edu.hk, wuwenyan@sensetime.com,{ ccloy,ziwei.liu{ @ntu.edu.sg
<FigureHere>
Figure 1: Illustration of Pose-Controllable Audio-Visual System (PC-AVS). Our approach takes one frame as identity reference and generates audio-driven talking faces with pose controlled by another pose source video. The mouth shapes of the generated frames are matched with the first row (synced video with audio) while the pose is matched with the bottom row (pose source).
# Abstract
pose code will be complementarily learned in a modulated $convolution-based$ reconstruction framework. Extensive experiments show that our method generates accurately lipsynced talking faces whose poses are controllable by other videos. Moreover, our model has multiple advanced capabilities including extreme view robustness and talking face frontalization.
## $1. \textbf{ Introduction}$
While accurate lip synchronization has been achieved for arbitrary-subject audio-driven talking face generation, the $problemofhowtoefficientlydrivetheheadposeremains.$ Previous methods rely on pre-estimated structural information such as landmarks and 3D parameters, aiming to generate personalized rhythmic movements. However, the inaccuracy of such estimated information under extreme conditions would lead to degradation problems. In this paper, we propose a clean yet effective framework to generate posecontrollable talking faces. We operate on non-aligned raw face images, using only a single photo as an identity reference. The key is to modularize audio-visual representations by devising an implicit low-dimension pose code. Substantially, both speech content and head pose information lie in a joint non- identity embedding space. While speech content information can be defined by learning the intrinsic synchronization between audio-visual modalities, we identify that a
用音频驱动静态肖像对于娱乐领域的各种应用非常重要，例如数字人类动画，电影中的视觉配音以及短视频的快速创建。借助深度学习，以前的研究人员采用两种不同的方法来分析音频驱动的会说话的人脸：1)通过纯潜在特征学习和图像重建$\mathbb{E}[14,77,9,72,50,55,44]$ ,以及 2) 借用结构中间表示的帮助，例如 2D 地标$\overline{\mathrm{E}}$
$^{\mathrm{l}}$Code, models, and demo videos are available at https://hangz-
nju-cuhk.github.io/projects/PC-AVS .
|
[
[
0.256591796875,
0.1337890625,
0.7138671875,
0.1337890625,
0.7138671875,
0.1483154296875,
0.256591796875,
0.1483154296875
],
[
0.223876953125,
0.1572265625,
0.74853515625,
0.1572265625,
0.74853515625,
0.1717529296875,
0.223876953125,
0.1717529296875
],
[
0.11517333984375,
0.2020263671875,
0.9345703125,
0.199462890625,
0.9345703125,
0.2203369140625,
0.11517333984375,
0.222900390625
],
[
0.1446533203125,
0.2215576171875,
0.7099609375,
0.222900390625,
0.7099609375,
0.2374267578125,
0.1446533203125,
0.236083984375
],
[
0.72802734375,
0.222900390625,
0.904296875,
0.222900390625,
0.904296875,
0.2374267578125,
0.72802734375,
0.2374267578125
],
[
0.2099609375,
0.2374267578125,
0.83251953125,
0.2392578125,
0.83251953125,
0.25830078125,
0.2099609375,
0.255615234375
],
[
0.09234619140625,
0.260009765625,
0.6611328125,
0.260009765625,
0.6611328125,
0.2734375,
0.09234619140625,
0.2734375
],
[
0.6845703125,
0.260009765625,
0.9453125,
0.260009765625,
0.9453125,
0.2734375,
0.6845703125,
0.2734375
],
[
0.79150390625,
0.304931640625,
0.84619140625,
0.304931640625,
0.84619140625,
0.335205078125,
0.79150390625,
0.335205078125
],
[
0.2467041015625,
0.313232421875,
0.2548828125,
0.313232421875,
0.2548828125,
0.31640625,
0.2467041015625,
0.31640625
],
[
0.488525390625,
0.311279296875,
0.52392578125,
0.30810546875,
0.52685546875,
0.322509765625,
0.491943359375,
0.32568359375
],
[
0.79736328125,
0.363037109375,
0.8388671875,
0.363037109375,
0.8388671875,
0.373779296875,
0.79736328125,
0.373779296875
],
[
0.810546875,
0.371826171875,
0.84228515625,
0.373779296875,
0.8408203125,
0.385009765625,
0.8095703125,
0.38330078125
],
[
0.7939453125,
0.417236328125,
0.85302734375,
0.417236328125,
0.85302734375,
0.427978515625,
0.7939453125,
0.427978515625
],
[
0.7958984375,
0.427978515625,
0.8486328125,
0.427978515625,
0.8486328125,
0.438232421875,
0.7958984375,
0.438232421875
],
[
0.13232421875,
0.466552734375,
0.181396484375,
0.468505859375,
0.181396484375,
0.479248046875,
0.13232421875,
0.4765625
],
[
0.1282958984375,
0.477783203125,
0.1650390625,
0.477783203125,
0.1650390625,
0.488037109375,
0.1282958984375,
0.488037109375
],
[
0.80224609375,
0.4892578125,
0.833984375,
0.4892578125,
0.833984375,
0.50048828125,
0.80224609375,
0.50048828125
],
[
0.81396484375,
0.50390625,
0.83837890625,
0.50390625,
0.83837890625,
0.50830078125,
0.81396484375,
0.50830078125
],
[
0.080078125,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.564453125,
0.080078125,
0.564453125
],
[
0.080078125,
0.56640625,
0.494384765625,
0.56494140625,
0.494384765625,
0.5791015625,
0.080078125,
0.580078125
],
[
0.58642578125,
0.564453125,
0.8896484375,
0.56640625,
0.8896484375,
0.580078125,
0.58642578125,
0.57763671875
],
[
0.53125,
0.5712890625,
0.55615234375,
0.5712890625,
0.55615234375,
0.57470703125,
0.53125,
0.57470703125
],
[
0.57421875,
0.57080078125,
0.5908203125,
0.57080078125,
0.5908203125,
0.57470703125,
0.57421875,
0.57470703125
],
[
0.07843017578125,
0.580078125,
0.7587890625,
0.580078125,
0.7587890625,
0.59326171875,
0.07843017578125,
0.59326171875
],
[
0.23779296875,
0.60595703125,
0.310546875,
0.60595703125,
0.310546875,
0.62060546875,
0.23779296875,
0.62060546875
],
[
0.5009765625,
0.60791015625,
0.890625,
0.60791015625,
0.890625,
0.6220703125,
0.5009765625,
0.6220703125
],
[
0.5048828125,
0.62548828125,
0.890625,
0.62548828125,
0.890625,
0.63623046875,
0.5048828125,
0.63623046875
],
[
0.101318359375,
0.64306640625,
0.46728515625,
0.64306640625,
0.46728515625,
0.6572265625,
0.101318359375,
0.6572265625
],
[
0.50244140625,
0.6396484375,
0.89208984375,
0.6396484375,
0.89208984375,
0.65283203125,
0.50244140625,
0.65283203125
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.65478515625,
0.89208984375,
0.65478515625,
0.89208984375,
0.66845703125,
0.50244140625,
0.66845703125
],
[
0.08087158203125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.08087158203125,
0.68798828125
],
[
0.50341796875,
0.66943359375,
0.89208984375,
0.66943359375,
0.89208984375,
0.68310546875,
0.50341796875,
0.68310546875
],
[
0.50244140625,
0.68310546875,
0.89208984375,
0.6845703125,
0.89208984375,
0.69873046875,
0.50244140625,
0.69775390625
],
[
0.08087158203125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.08087158203125,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.4697265625,
0.70556640625,
0.4697265625,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50244140625,
0.70068359375,
0.60107421875,
0.70068359375,
0.60107421875,
0.7119140625,
0.50244140625,
0.7119140625
],
[
0.07843017578125,
0.72021484375,
0.468017578125,
0.7197265625,
0.468017578125,
0.734375,
0.07843017578125,
0.73486328125
],
[
0.080078125,
0.736328125,
0.468017578125,
0.734375,
0.468017578125,
0.74755859375,
0.080078125,
0.75
],
[
0.08087158203125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.08087158203125,
0.7646484375
],
[
0.50341796875,
0.7470703125,
0.63623046875,
0.7470703125,
0.63623046875,
0.76025390625,
0.50341796875,
0.76025390625
],
[
0.080078125,
0.7666015625,
0.4697265625,
0.7666015625,
0.4697265625,
0.77978515625,
0.080078125,
0.77978515625
],
[
0.52392578125,
0.77294921875,
0.890625,
0.77294921875,
0.890625,
0.7861328125,
0.52392578125,
0.7861328125
],
[
0.08087158203125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.08087158203125,
0.79443359375
],
[
0.5048828125,
0.7861328125,
0.890625,
0.7861328125,
0.890625,
0.7998046875,
0.5048828125,
0.7998046875
],
[
0.080078125,
0.796875,
0.4697265625,
0.796875,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50634765625,
0.8037109375,
0.88818359375,
0.8037109375,
0.88818359375,
0.814453125,
0.50634765625,
0.814453125
],
[
0.08087158203125,
0.81103515625,
0.4697265625,
0.81103515625,
0.4697265625,
0.82470703125,
0.08087158203125,
0.82470703125
],
[
0.50341796875,
0.81640625,
0.8896484375,
0.8173828125,
0.8896484375,
0.83203125,
0.50341796875,
0.83056640625
],
[
0.08087158203125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50341796875,
0.83349609375,
0.88232421875,
0.83349609375,
0.88232421875,
0.8466796875,
0.50341796875,
0.8466796875
],
[
0.08087158203125,
0.84130859375,
0.4697265625,
0.84130859375,
0.4697265625,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.87841796875,
0.8369140625,
0.88232421875,
0.8369140625,
0.88232421875,
0.84228515625,
0.87841796875,
0.84228515625
],
[
0.5048828125,
0.84765625,
0.890625,
0.8466796875,
0.890625,
0.861328125,
0.5048828125,
0.8623046875
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85498046875,
0.4697265625,
0.869140625,
0.080078125,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.47216796875,
0.8720703125,
0.47216796875,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.5185546875,
0.87353515625,
0.89208984375,
0.87548828125,
0.89208984375,
0.8896484375,
0.5185546875,
0.88720703125
],
[
0.080078125,
0.88720703125,
0.471435546875,
0.88720703125,
0.471435546875,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.8896484375,
0.78125,
0.8896484375,
0.78125,
0.8994140625,
0.50341796875,
0.8994140625
]
] |
[
"Pose-Controllable Talking Face Generation bv",
"ImplicitlyModularizedAudio-VisualRepresentation",
"Hang Zhou 1 , Yasheng Sun 2,3 , Wayne Wu 2,4 , Chen Change Loy 4 , Xiaogang Wang 1 , Ziwei Liu 4",
"1CUHK-SenseTime Joint Lab,TheChineseUniversity of HongKong",
"2 SenseTime Research",
"3 Tokyo Institute of Technology4 S-Lab, Nanyang Technological University",
"f zhouhang@link,xgwang@ee } .cuhk.edu.hk,wuwenyan@sensetime.com,",
"ccloy,ziwei.liu } @ntu.edu.sg",
"vnacad",
"",
"",
"Audio",
"purce",
"Generated",
"Results",
"Identity",
"Refere",
"Pose",
"rCE",
"Figure 1:Illustration of Pose-Controllable Audio-Visual System (PC-AVS). Our approach takes one frame as identity reference and",
"generates audio-driven talking faces with pose controlled by another",
"leo.Themouthshapesof thegenerated frames are",
"011",
"a",
"matched with the first row (synced video with audio) while the pose is matched with the bottom row (pose source).",
"Abstract",
"pose code will be complementarily learned in a modulated",
"convolution-based reconstruction framework.Extensive ex",
"While accurate lip synchronization has been achieved for",
"periments show that our method generates accurately lip-",
"arbitrary-subject audio-driven talking face generation, the",
"synced talking faces whose poses are controllable by other",
"problem of how to efficiently drive the head pose remains.",
"videos. Moreover, our model has multiple advanced capa-",
"bilities including extreme view robustness and talking face",
"Previous methods rely on pre-estimated structural informa-",
"tion such as landmarks and 3D parameters, aiming to gen-",
"frontalization.",
"erate personalized rhythmic movements. However, the inac",
"curacy of such estimated information under extreme condi-",
"tions would lead to degradation problems. In this paper, we",
"1. Introduction",
"propose a clean yet effective framework to generate pose-",
"用音频驱动静态肖像对于娱乐领域的各种应用非常",
"controllable talking faces. We operate on non-aligned raw",
"重要，例如数字人类动画，电影中的视觉配音以及短视",
"face images, using only a single photo as an identity refer-",
"频的快速创建。借助深度学习，以前的研究人员采用两",
"ence. The key is to modularize audio-visual representations",
"种不同的方法来分析音频驱动的会说话的人脸：1)通过",
"by devising an implicit low-dimension pose code. Substan-",
"纯潜在特征学习和图像重建E[14,77，9,72,50,55，44",
"tially, both speech content and head pose information lie in",
"-",
"，以及2）借用结构中间表示的帮助，例如2D地标E",
"a joint non-identity embedding space. While speech content",
"information can be defined by learning the intrinsic synchro-",
"ICode,models,and demo videos are availableat https://hangz",
"nization between audio-visual modalities, we identify that a",
"nju-cuhk.github.io/projects/pc-Avs."
] |
{
"arxiv_id": "2104.11116",
"language": "zh",
"page_idx": 0
}
|
||
2104.11116-zh-1
|
2104.11116-zh-1.jpg
|
[51,10,18] 或 3D 表示$\mathbb{E}[1,52,49,8,75,46,28]$。尽管在生成准确的嘴部运动方面取得了很大进展，但以前的大多数方法都无法对头部姿势进行建模，这是说话面部看起来自然的关键因素之一
在生成带有音频的口型同步视频时控制头部姿势非常具有挑战性。1)一方面，很少可以从音频中推断出姿势信息。虽然大多数以前的作品选择在视频中保留原始姿势，但最近，一些作品已经解决了从音频生成个性化有节奏的头部运动的问题[8,75,66]。但是，他们依靠一小段视频来学习个人节奏$\overline{\mathbb{E}}[66,8]$ ,这对于各种场景可能都是无法访问的。2) 另一方面，上述所有方法都依赖于 3D 结构中间表示$\mathbb{E}[66,8,75]$。姿态信息与面部运动固有地耦合在一起，这会影响基于重建的$\mathbb{E}$ [14,72] 和基于 2D 地标的方法$\underline{\mathbb{E}}[10,18]$。因此，最合理的方法是利用 3D 模型$\boxed{\mathrm{F}}[33,52,8,71]$ ,其中明确定义了姿势参数和表达式参数$\overline{\mathrm{E}}$[2]。然而，通常需要长期视频来学习 3D 模型的特定于人的渲染器。更重要的是，这种表示在极端情况下是不准确的，例如大姿势或弱光条件
在这项工作中，我们提出了姿势可控视听系统(PCAVS),该系统在驱动带有音频的任意说话面孔时实现了自由姿势控制。我们不是从音频中学习姿势运动，而是利用另一个姿势源视频来仅补偿头部运动，如图所示。1 。具体来说，在我们的工作中不需要结构信息。关键是设计一个不含嘴形或身份信息的隐式低维姿势代码。通过这种方式，视听表示被模块化为三个关键因素的空间：语音内容，头部姿势和身份信息
特别是，我们通过数据增强从视觉领域识别非身份潜在嵌人的存在。直觉上，互补的语音内容和姿势信息应该源于它。提取视觉和音频表示之间的共享信息可以通过同步两种模式来产生语音内容空间，这也被证明对各种下游视听任务都是有益的[17,40,73,36,25] 。但是，如果没有精确识别的结构信息，就没有明确的方法对姿势进行建模。在这里，我们利用 3D 姿势参数的先验知识，即仅仅 12 维的向量，包括 3D 旋转矩阵， 2D 位置偏移偏差和比例因子，就足以表示头部姿势。因此，我们定义了从非身份空间到低维度代码的映射， 该映射隐式地代表姿势。值得注意的是，我们不使用其他 3D 先验来建模不同姿势之间的转换。然后，通过额外的身份监督，整个说话面部表示的模块化已经完成
最后一个关键是视频帧之间的跨帧重建，其中所有表示都是互补学习的。还设计了一个生成器，其卷积内核由嵌人特征调制。具体来说，我们从模块化空间组装功能，并使用它们来缩放$\mathbb{E}[31]$ 中提出的内核权重。权重调制的表达能力也以隐式的方式将我们的模块化强加于视听表示，即。e。,为了保证较低的重构损失，在语音内容嵌入照顾嘴巴的同时，低维码自动控制 pose。在推理过程中，我们可以通过一段音频片段驱动任意面部，头部运动由另一个视频控制。还可以实现多个高级属性，例如极端视图鲁棒性和正面说话的面孔
我们的贡献总结如下：1) 我们建议通过设计受 3D 启发的低维姿势代码，将说话的人脸表示分别模块化为语音内容，头部姿势和身份的空间。2)在具有调制
卷积的基于构造的框架中隐式和互补地学习模块化。3) 我们的模型生成具有精确嘴唇同步的姿势可控的说话面孔。4) 由于我们的系统中没有使用结构中间信息， 因此我们的模型几乎不需要预处理，并且对输人视图具有鲁棒性
# $2. \textbf{ Related Work}$
音频驱动说话人脸生成。通过音频输入驱动说话的面孔$\mathbb{E}[7,78]$ 长期以来一直是计算机视觉和图形学领域的重要研究兴趣，其中结构信息和拼接技术起着至关重要的作用$\boxed{\mathbb{E}}[4,3,57,56,76]$。例如，Bregler 等人。[4]重写嘴轮廓。随着深度学习的发展，已经提出了不同的方法来通过时间建模生成地标[20,21,51]。然而，以前的方法大多是特定于扬声器的。最近，随着端到端交叉视听生成$\mathbb{E}[11,70,24,69,64,74,23,54,65,22]$的发展，研究人员探索了与说话者无关的设置$\mathbb{E}[14,50,45,72,75]$ ,该设置寻求一种通用模型，该模型处理具有一个或几个帧参考的所有身份。在 Chung 等人之后。[14]首先提出了一种基于端到端重构的网络，Zhou 等人。$\mathbb{E}[72]$ 通过对抗性表示学习进一步从单词中分离出身份。基于重建的方法的关键思想是确定音频和视频之间的同步[17,72,45,44],其中 Prajwal 等人。[44]带音频的同步口，用于基于修补的重建。然而，由于难以将头部姿势与面部运动解耦，因此这些基于重建的作品通常会忽略头部运动。
作为更紧凑和易于学习的目标，结构信息被用作基于GAN 的重建管道中的中间表示 [10,18,52,49,75,66] 。Chenetal。[10] 和 Das 等人。$\mathbb{E}[18]$都使用两阶段网络首先预测 2D 地标并生成人脸。但是姿势和嘴巴也纠缠在 2D 地标内。另一方面，3D 工具$\overline{\mathbb{E}}[2,19,5,29]$ 充当强大的中间表示。Zhouetal。特别是具有个性化运动的 3D 对话地标模型，可产生当前最自然的结果。然而， 在他们的模型中没有实现自由姿势控制。Chen etal。和Yi etal。两者都利用 3D 模型来学习自然姿势。然而，他们的方法不能在“一次性”条件下呈现良好的质量。此外，在极端条件下，三维拟合精度会显著下降。在这里，我们建议在不依赖任何中间结构信息的情况下实现一个参考系的自由姿态控制。面重演。我们的工作还涉及视觉驱动的说话面孔，这是在面部再现领域进行研究的。。与音频驱动的作品类似，大多数面部重演作品取决于结构信息，例如地标$\mathbb{E}[62,26,68,67]$, 分割图$\mathbb{E}[12,6]$ 和 3D 模型$\mathbb{E}[53,33,32]$。具体来说，某些作品$\overline{\mathbb{E}}[61,47,59]$学习像素之间的翘曲没有定义的结构信息。虽然相当多的论文[67,6,18]在几个帧上采用元学习过程来获得新的身份，但我们的方法不需要这个过程
# $3. \textbf{ Our Approach}$
我们介绍了可控制姿势的视听系统(PC-AVS),该系统旨在实现自由姿势控制，同时驱动静态照片与音频交谈。整个管道如图 3 所示。在本节中，我们首先通过识别非身份空间(Sec。3.1),然后我们提供视听
2
|
[
[
0.08251953125,
0.0947265625,
0.46728515625,
0.0947265625,
0.46728515625,
0.10479736328125,
0.08251953125,
0.10479736328125
],
[
0.50634765625,
0.0947265625,
0.890625,
0.0947265625,
0.890625,
0.10479736328125,
0.50634765625,
0.10479736328125
],
[
0.08251953125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08251953125,
0.121826171875
],
[
0.50341796875,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.08251953125,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08251953125,
0.1351318359375
],
[
0.50341796875,
0.12249755859375,
0.88818359375,
0.12371826171875,
0.88818359375,
0.136962890625,
0.50341796875,
0.1363525390625
],
[
0.08251953125,
0.14013671875,
0.295654296875,
0.14013671875,
0.295654296875,
0.15087890625,
0.08251953125,
0.15087890625
],
[
0.5048828125,
0.1383056640625,
0.8896484375,
0.1383056640625,
0.8896484375,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.098876953125,
0.1546630859375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1685791015625,
0.098876953125,
0.169189453125
],
[
0.50634765625,
0.1546630859375,
0.5849609375,
0.1546630859375,
0.5849609375,
0.1666259765625,
0.50634765625,
0.1666259765625
],
[
0.08251953125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.18310546875,
0.08251953125,
0.18310546875
],
[
0.08251953125,
0.1849365234375,
0.46728515625,
0.1849365234375,
0.46728515625,
0.1988525390625,
0.08251953125,
0.1988525390625
],
[
0.50244140625,
0.1849365234375,
0.646484375,
0.1849365234375,
0.646484375,
0.1988525390625,
0.50244140625,
0.1988525390625
],
[
0.08416748046875,
0.2020263671875,
0.464111328125,
0.2020263671875,
0.464111328125,
0.212158203125,
0.08416748046875,
0.212158203125
],
[
0.5048828125,
0.210205078125,
0.890625,
0.210205078125,
0.890625,
0.2235107421875,
0.5048828125,
0.2235107421875
],
[
0.08416748046875,
0.2171630859375,
0.46728515625,
0.2171630859375,
0.46728515625,
0.2279052734375,
0.08416748046875,
0.2279052734375
],
[
0.08087158203125,
0.2291259765625,
0.468017578125,
0.2298583984375,
0.468017578125,
0.2442626953125,
0.08087158203125,
0.24365234375
],
[
0.50244140625,
0.2259521484375,
0.8896484375,
0.2259521484375,
0.8896484375,
0.2392578125,
0.50244140625,
0.2392578125
],
[
0.50244140625,
0.2392578125,
0.8896484375,
0.240478515625,
0.8896484375,
0.25390625,
0.50244140625,
0.25244140625
],
[
0.08251953125,
0.24560546875,
0.465576171875,
0.24560546875,
0.465576171875,
0.2587890625,
0.08251953125,
0.2587890625
],
[
0.08087158203125,
0.261474609375,
0.468017578125,
0.261474609375,
0.468017578125,
0.274658203125,
0.08087158203125,
0.274658203125
],
[
0.50341796875,
0.255615234375,
0.890625,
0.255615234375,
0.890625,
0.26953125,
0.50341796875,
0.26953125
],
[
0.08416748046875,
0.27587890625,
0.468017578125,
0.27587890625,
0.468017578125,
0.2890625,
0.08416748046875,
0.2890625
],
[
0.50341796875,
0.270751953125,
0.8896484375,
0.270751953125,
0.8896484375,
0.2841796875,
0.50341796875,
0.2841796875
],
[
0.08251953125,
0.291748046875,
0.468017578125,
0.291748046875,
0.468017578125,
0.304931640625,
0.08251953125,
0.304931640625
],
[
0.50341796875,
0.285888671875,
0.8896484375,
0.285888671875,
0.8896484375,
0.2998046875,
0.50341796875,
0.2998046875
],
[
0.08087158203125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.3193359375,
0.08087158203125,
0.3193359375
],
[
0.5048828125,
0.300537109375,
0.890625,
0.300537109375,
0.890625,
0.314453125,
0.5048828125,
0.314453125
],
[
0.08416748046875,
0.3212890625,
0.468017578125,
0.3212890625,
0.468017578125,
0.335205078125,
0.08416748046875,
0.335205078125
],
[
0.5048828125,
0.317626953125,
0.8896484375,
0.317626953125,
0.8896484375,
0.32763671875,
0.5048828125,
0.32763671875
],
[
0.50244140625,
0.330078125,
0.8896484375,
0.328857421875,
0.8896484375,
0.343505859375,
0.50244140625,
0.3447265625
],
[
0.08251953125,
0.337158203125,
0.46728515625,
0.337158203125,
0.46728515625,
0.347900390625,
0.08251953125,
0.347900390625
],
[
0.5048828125,
0.34521484375,
0.890625,
0.34521484375,
0.890625,
0.359130859375,
0.5048828125,
0.359130859375
],
[
0.08416748046875,
0.352783203125,
0.46728515625,
0.352783203125,
0.46728515625,
0.363037109375,
0.08416748046875,
0.363037109375
],
[
0.50341796875,
0.359130859375,
0.890625,
0.361083984375,
0.890625,
0.375732421875,
0.50341796875,
0.373779296875
],
[
0.08251953125,
0.367431640625,
0.1446533203125,
0.367431640625,
0.1446533203125,
0.37890625,
0.08251953125,
0.37890625
],
[
0.50341796875,
0.375732421875,
0.890625,
0.375732421875,
0.890625,
0.39013671875,
0.50341796875,
0.39013671875
],
[
0.1004638671875,
0.3818359375,
0.468017578125,
0.3818359375,
0.468017578125,
0.395751953125,
0.1004638671875,
0.395751953125
],
[
0.5048828125,
0.391357421875,
0.890625,
0.391357421875,
0.890625,
0.40478515625,
0.5048828125,
0.40478515625
],
[
0.08087158203125,
0.3984375,
0.468017578125,
0.3984375,
0.468017578125,
0.412353515625,
0.08087158203125,
0.412353515625
],
[
0.50634765625,
0.40771484375,
0.8896484375,
0.40771484375,
0.8896484375,
0.41845703125,
0.50634765625,
0.41845703125
],
[
0.08660888671875,
0.415283203125,
0.465576171875,
0.415283203125,
0.465576171875,
0.425537109375,
0.08660888671875,
0.425537109375
],
[
0.50341796875,
0.421630859375,
0.890625,
0.421630859375,
0.890625,
0.43505859375,
0.50341796875,
0.43505859375
],
[
0.08251953125,
0.4287109375,
0.468017578125,
0.4287109375,
0.468017578125,
0.442626953125,
0.08251953125,
0.442626953125
],
[
0.08087158203125,
0.443115234375,
0.4697265625,
0.442626953125,
0.4697265625,
0.45703125,
0.08087158203125,
0.458251953125
],
[
0.50634765625,
0.438232421875,
0.890625,
0.438232421875,
0.890625,
0.448974609375,
0.50634765625,
0.448974609375
],
[
0.50341796875,
0.451904296875,
0.890625,
0.451904296875,
0.890625,
0.46533203125,
0.50341796875,
0.46533203125
],
[
0.08251953125,
0.458984375,
0.46728515625,
0.458984375,
0.46728515625,
0.472900390625,
0.08251953125,
0.472900390625
],
[
0.50341796875,
0.466552734375,
0.625,
0.466552734375,
0.625,
0.479736328125,
0.50341796875,
0.479736328125
],
[
0.08251953125,
0.474853515625,
0.46728515625,
0.474853515625,
0.46728515625,
0.488037109375,
0.08251953125,
0.488037109375
],
[
0.52392578125,
0.48291015625,
0.890625,
0.48291015625,
0.890625,
0.496826171875,
0.52392578125,
0.496826171875
],
[
0.08087158203125,
0.4892578125,
0.3896484375,
0.4892578125,
0.3896484375,
0.50244140625,
0.08087158203125,
0.50244140625
],
[
0.50244140625,
0.496826171875,
0.8896484375,
0.498779296875,
0.8896484375,
0.51220703125,
0.50244140625,
0.51025390625
],
[
0.098876953125,
0.505859375,
0.46728515625,
0.505859375,
0.46728515625,
0.51953125,
0.098876953125,
0.51953125
],
[
0.51123046875,
0.51220703125,
0.88818359375,
0.51220703125,
0.88818359375,
0.52587890625,
0.51123046875,
0.52587890625
],
[
0.08087158203125,
0.52099609375,
0.46728515625,
0.52099609375,
0.46728515625,
0.5341796875,
0.08087158203125,
0.5341796875
],
[
0.50341796875,
0.52783203125,
0.8896484375,
0.52783203125,
0.8896484375,
0.54150390625,
0.50341796875,
0.54150390625
],
[
0.08087158203125,
0.5361328125,
0.46728515625,
0.5361328125,
0.46728515625,
0.5498046875,
0.08087158203125,
0.5498046875
],
[
0.50341796875,
0.54248046875,
0.890625,
0.54248046875,
0.890625,
0.556640625,
0.50341796875,
0.556640625
],
[
0.08251953125,
0.5517578125,
0.468017578125,
0.5517578125,
0.468017578125,
0.56494140625,
0.08251953125,
0.56494140625
],
[
0.50341796875,
0.556640625,
0.88818359375,
0.556640625,
0.88818359375,
0.57080078125,
0.50341796875,
0.57080078125
],
[
0.08087158203125,
0.56640625,
0.46728515625,
0.56640625,
0.46728515625,
0.580078125,
0.08087158203125,
0.580078125
],
[
0.5048828125,
0.57470703125,
0.8896484375,
0.57470703125,
0.8896484375,
0.58544921875,
0.5048828125,
0.58544921875
],
[
0.09478759765625,
0.58349609375,
0.465576171875,
0.58349609375,
0.465576171875,
0.59326171875,
0.09478759765625,
0.59326171875
],
[
0.50341796875,
0.58837890625,
0.890625,
0.58837890625,
0.890625,
0.6015625,
0.50341796875,
0.6015625
],
[
0.08251953125,
0.5966796875,
0.468017578125,
0.5966796875,
0.468017578125,
0.60986328125,
0.08251953125,
0.60986328125
],
[
0.5048828125,
0.60498046875,
0.8896484375,
0.60498046875,
0.8896484375,
0.61572265625,
0.5048828125,
0.61572265625
],
[
0.08251953125,
0.61279296875,
0.46728515625,
0.61279296875,
0.46728515625,
0.62353515625,
0.08251953125,
0.62353515625
],
[
0.08087158203125,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.08087158203125,
0.6396484375
],
[
0.50634765625,
0.62060546875,
0.890625,
0.62060546875,
0.890625,
0.630859375,
0.50634765625,
0.630859375
],
[
0.5048828125,
0.6337890625,
0.890625,
0.6337890625,
0.890625,
0.64794921875,
0.5048828125,
0.64794921875
],
[
0.08251953125,
0.642578125,
0.468017578125,
0.642578125,
0.468017578125,
0.65576171875,
0.08251953125,
0.65576171875
],
[
0.50244140625,
0.64794921875,
0.890625,
0.6494140625,
0.890625,
0.66357421875,
0.50244140625,
0.6611328125
],
[
0.08251953125,
0.65771484375,
0.46728515625,
0.65771484375,
0.46728515625,
0.66845703125,
0.08251953125,
0.66845703125
],
[
0.50341796875,
0.66357421875,
0.890625,
0.6640625,
0.890625,
0.67822265625,
0.50341796875,
0.6767578125
],
[
0.08251953125,
0.671875,
0.468017578125,
0.671875,
0.468017578125,
0.68505859375,
0.08251953125,
0.68505859375
],
[
0.50341796875,
0.6787109375,
0.890625,
0.6787109375,
0.890625,
0.6923828125,
0.50341796875,
0.6923828125
],
[
0.08251953125,
0.6875,
0.45751953125,
0.6875,
0.45751953125,
0.70068359375,
0.08251953125,
0.70068359375
],
[
0.50634765625,
0.69580078125,
0.88818359375,
0.69580078125,
0.88818359375,
0.70556640625,
0.50634765625,
0.70556640625
],
[
0.101318359375,
0.7041015625,
0.46728515625,
0.7041015625,
0.46728515625,
0.71728515625,
0.101318359375,
0.71728515625
],
[
0.50634765625,
0.7109375,
0.8896484375,
0.7109375,
0.8896484375,
0.7216796875,
0.50634765625,
0.7216796875
],
[
0.08416748046875,
0.72021484375,
0.465576171875,
0.72021484375,
0.465576171875,
0.73095703125,
0.08416748046875,
0.73095703125
],
[
0.50341796875,
0.7236328125,
0.8896484375,
0.72265625,
0.8896484375,
0.7373046875,
0.50341796875,
0.73779296875
],
[
0.08416748046875,
0.736328125,
0.465576171875,
0.736328125,
0.465576171875,
0.7470703125,
0.08416748046875,
0.7470703125
],
[
0.5048828125,
0.7392578125,
0.8896484375,
0.7392578125,
0.8896484375,
0.75244140625,
0.5048828125,
0.75244140625
],
[
0.08251953125,
0.74853515625,
0.465576171875,
0.74853515625,
0.465576171875,
0.76220703125,
0.08251953125,
0.76220703125
],
[
0.50341796875,
0.75390625,
0.8896484375,
0.75390625,
0.8896484375,
0.767578125,
0.50341796875,
0.767578125
],
[
0.08251953125,
0.7646484375,
0.468017578125,
0.7646484375,
0.468017578125,
0.77783203125,
0.08251953125,
0.77783203125
],
[
0.5048828125,
0.76953125,
0.8896484375,
0.76953125,
0.8896484375,
0.78271484375,
0.5048828125,
0.78271484375
],
[
0.08087158203125,
0.77978515625,
0.468017578125,
0.77978515625,
0.468017578125,
0.79345703125,
0.08087158203125,
0.79345703125
],
[
0.50341796875,
0.7841796875,
0.5537109375,
0.7841796875,
0.5537109375,
0.798828125,
0.50341796875,
0.798828125
],
[
0.08251953125,
0.79443359375,
0.46728515625,
0.79443359375,
0.46728515625,
0.80810546875,
0.08251953125,
0.80810546875
],
[
0.08251953125,
0.81005859375,
0.46728515625,
0.81005859375,
0.46728515625,
0.82373046875,
0.08251953125,
0.82373046875
],
[
0.50244140625,
0.81640625,
0.65185546875,
0.81640625,
0.65185546875,
0.83056640625,
0.50244140625,
0.83056640625
],
[
0.08251953125,
0.82470703125,
0.468017578125,
0.82470703125,
0.468017578125,
0.83837890625,
0.08251953125,
0.83837890625
],
[
0.08087158203125,
0.8388671875,
0.42236328125,
0.8388671875,
0.42236328125,
0.85302734375,
0.08087158203125,
0.85302734375
],
[
0.52392578125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.52392578125,
0.85498046875
],
[
0.101318359375,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.101318359375,
0.87060546875
],
[
0.5048828125,
0.85693359375,
0.8896484375,
0.85693359375,
0.8896484375,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.08251953125,
0.87353515625,
0.46728515625,
0.87353515625,
0.46728515625,
0.8837890625,
0.08251953125,
0.8837890625
],
[
0.50244140625,
0.87060546875,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.50244140625,
0.88525390625
],
[
0.08251953125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08251953125,
0.90087890625
],
[
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
]
] |
[
"[51，10.181或3D表示F1.52,49，8.75，46.281。尽管",
"卷积的基于构造的框架中隐式和互补地学习模块化。3）",
"在生成准确的嘴部运动方面取得了很大进展，但以前",
"我们的模型生成具有精确嘴唇同步的姿势可控的说话",
"的大多数方法都无法对头部姿势进行建模，这是说话",
"面孔。4）由于我们的系统中没有使用结构中间信息",
"面部看起来自然的关键因素之",
"因此我们的模型几乎不需要预处理，并且对输入视图",
"在生成带有音频的口型同步视频时控制头部姿势非",
"具有鲁棒性",
"常具有挑战性。1）一方面，很少可以从音频中推断出",
"姿势信息。虽然大多数以前的作品选择在视频中保留",
"2.Related Work",
"原始姿势，但最近，一些作品已经解决了从音频生成个",
"音频驱动说话人脸生成。通过音频输人驱动说话的面",
"性化有节奏的头部运动的问题[8，75，66]。但是，他们",
"依靠一小段视频来学习个人节奏E[66,8]，这对于各种",
"孔E[7，78]长期以来一直是计算机视觉和图形学领域",
"的重要研究兴趣，其中结构信息和拼接技术起着至关重",
"场景可能都是无法访问的。2）另一方面，上述所有方",
"法都依赖于3D结构中间表示E[66,8，75]。姿态信息",
"要的作用E[4,3,57,56,76]。例如，Bregler等人。[4]重",
"与面部运动固有地耦合在一起，这会影响基于重建的E",
"写嘴轮廓。随着深度学习的发展，已经提出了不同的方",
"[14，72]和基于2D地标的方法E[10,18]。因此，最合",
"法来通过时间建模生成地标[20,21,51]。然而，以前的",
"理的方法是利用3D模型E[33,52,8，71]，其中明确定",
"方法大多是特定于扬声器的。最近，随着端到端交叉视",
"义了姿势参数和表达式参数E[2]。然而，通常需要长",
"听生成F11，70.24,69.64，74，23，54，65.221的发展，研",
"究人员探索了与说话者无关的设置E[14,50,45，72,75]",
"期视频来学习3D模型的特定于人的染器。更重要的",
"，该设置寻求一种通用模型，该模型处理具有一个或几",
"是，这种表示在极端情况下是不准确的，例如大姿势或",
"个顿参考的所有身份。在Chung等人之后。[14]首先",
"弱光条件",
"提出了一种基于端到端重构的网络，Zhou等人。E[72]",
"在这项工作中，我们提出了姿势可控视听系统（PC",
"通过对抗性表示学习进一步从单词中分离出身份。基",
"AVS），该系统在驱动带有音频的任意说话面孔时实现",
"于重建的方法的关键思想是确定音频和视频之间的同",
"了自由姿势控制。我们不是从音频中学习姿势运动，而",
"步[17，72,45，44]，其中Prajwal等人。[44]带音频的",
"是利用另一个姿势源视频来仅补偿头部运动，如图所",
"示。1。具体来说，在我们的工作中不需要结构信息。",
"司步口，用于基于修补的重建。然而，由于难以将头部",
"姿势与面部运动解耦，因此这些基于重建的作品通常",
"关键是设计一个不含嘴形或身份信息的隐式低维姿势",
"会忽略头部运动。",
"代码。通过这种方式，视听表示被模块化为三个关键因",
"作为更紧凑和易于学习的目标，结构信息被用作基于",
"素的空间：语音内容，头部姿势和身份信息",
"GAN的重建管道中的中间表示[10,18，52，49,75,66]",
"特别是，我们通过数据增强从视觉领域识别非身份",
"Chenetal。[10]和Das等人。E[18]都使用两阶段网",
"潜在嵌人的存在。直觉上，互补的语音内容和姿势信息",
"络首先预测2D地标并生成人脸。但是姿势和嘴巴也纠",
"应该源于它。提取视觉和音频表示之间的共享信息可",
"缠在2D地标内。另一方面，3D工具E[2,19,5,29]充",
"以通过同步两种模式来产生语音内容空间，这也被证",
"当强大的中间表示。Zhouetal。特别是具有个性化运动",
"明对各种下游视听任务都是有益的[17，40,73，36，25]",
"的3D对话地标模型，可产生当前最自然的结果。然而，",
"但是，如果没有精确识别的结构信息，就没有明确的",
"在他们的模型中没有实现自由姿势控制。Chenetal。和",
"方法对姿势进行建模。在这里，我们利用3D姿势参数",
"Yietal。两者都利用3D模型来学习自然姿势。然而，他",
"的先验知识，即仅仅12维的向量，包括3D旋转矩阵，",
"2D位置偏移偏差和比例因子，就足以表示头部姿势。",
"们的方法不能在“一次性”条件下呈现良好的质量。此",
"外，在极端条件下，三维拟合精度会显著下降。在这",
"因此，我们定义了从非身份空间到低维度代码的映射，",
"里，我们建议在不依赖任何中间结构信息的情况下实",
"该映射隐式地代表姿势。值得注意的是，我们不使用其",
"现一个参考系的自由姿态控制。面重演。我们的工作还",
"他3D先验来建模不同姿势之间的转换。然后，通过额",
"涉及视觉驱动的说话面孔，这是在面部再现领域进行",
"外的身份监督，整个说话面部表示的模块化已经完成",
"研究的。。与音频驱动的作品类似，大多数面部重演作",
"最后一个关键是视频帧之间的跨帧重建，其中所有",
"品取决于结构信息，例如地标F[62，26，68，67]，分割",
"表示都是互补学习的。还设计了一个生成器，其卷积内",
"图E[12,6]和3D模型E[53，33,32]。具体来说，某些",
"核由嵌人特征调制。具体来说，我们从模块化空间组装",
"作品E[61,47,59]学习像素之间的翘曲没有定义的结",
"功能，并使用它们来缩放E[31]中提出的内核权重。权",
"构信息。虽然相当多的论文[67，6,18]在几个顿上采用",
"重调制的表达能力也以隐式的方式将我们的模块化强",
"元学习过程来获得新的身份，但我们的方法不需要这",
"加于视听表示，即。e。，为了保证较低的重构损失，在",
"个过程",
"语音内容嵌入照顾嘴巴的同时，低维码自动控制pose。",
"在推理过程中，我们可以通过一段音频片段驱动任意",
"3. Our Approach",
"面部，头部运动由另一个视频控制。还可以实现多个高",
"级属性，例如极端视图鲁棒性和正面说话的面孔",
"我们介绍了可控制姿势的视听系统（PC-AVS），该",
"我们的贡献总结如下：1）我们建议通过设计受3D",
"系统旨在实现自由姿势控制，同时驱动静态照片与音",
"启发的低维姿势代码，将说话的人脸表示分别模块化",
"频交谈。整个管道如图3所示。在本节中，我们首先",
"为语音内容，头部姿势和身份的空间。2）在具有调制",
"通过识别非身份空间（Sec。3.1），然后我们提供视听",
"2"
] |
{
"arxiv_id": "2104.11116",
"language": "zh",
"page_idx": 1
}
|
||
2104.11116-zh-3
|
2104.11116-zh-3.jpg
|
<FigureHere>
Figure 3: The overall pipeline of our Pose-Controllable Audio-Visual System(PC-AVS) framework. The identity reference $I_{(ref)}$ is $\begin{array}{l}\text{encoded by E}_{i}\text{ to the identity space( red). Encoder E}_{n}\text{ encodes video clip}V\text{to }F_{n}\text{ in the non-identity space(grey).Then it is mapped to}\\F_{c}^{\upsilon}\text{in the speech content space(blue), which it shares with }F_{c}^{\alpha}\text{ encoded by E}_{c}^{\alpha}\end{array}$ $,\mathcal{F}_{c(k)}^{v-}$from two negative examples. Their features in pose and identity spaces are also shown. Specifically, we map $F_n$ to pose features $F_p=f_{p(1:k)}$ in the pose space (yellow). Though motivated by 3D priors, the pose features are not supervised by or necessarily represent the traditional $[R,t,s]$ 3D parameters. Finally, a pair of features $\{f_i(i),f_{p(k)},f_{c(k)}^a\}$ are assembled together and sent to generator G .
这个想法与使用 3D 先验进行无监督 3D 表示学习的论文有一些相似之处[38,48]。不同的是，我们只在所需数据的最小维度上使用先验知识。虽然隐式学习的特征不可能拥有与真实 3D 姿态参数相同的值，但这种直觉对我们的设计很重要。具有较大尺寸的姿势代码可能包含不需要的其他信息。Sec 中描述了定义的姿势代码起作用的原因。3.3
标识空间编码。以前的研究已经很好地解决了身份空间的学习问题[14,72,75]。当我们在名人[15]的视频上训练我们的网络时，身份标签自然存在。我们的身份空间$F_i=\mathrm{E}_i(V)=\{f_{i(1)},\ldots,f_{i(K)}\}$可以在具有softmax 交叉熵损失$\mathcal{L}_i$的身份分类上学习
内核的权重。在每个卷积块处，学习多层感知器以将$f_{cat(k)}$映射到调制向量$\mathcal{M}$,该调制向量具有与输入特征的通道相同的维度。对于卷积核权重$w$中的每个值$w_{xyz}$ , 其中 $x$ 是其在输入特征通道上的位置$,y$ 与输出通道号有关，而$z$表示空间位置，在给定$x$的$\mathcal{M}$的小值的情况下对其进行调制和归一化，以避免以下情况：。
网络培训。最后，通过图像重建联合训练特征空间模块化和生成器。我们直接借用[42]中应用的相同损失函数。生成的图像和地面实况图像被发送到具有$N_D$ 层的多尺度鉴别器 D。鉴别器用于计算其层内的特征图$L_1$距离以及对抗性生成学习。还使用了依赖于具有$N_P$层的预训练 VGG 网络的感知损失。所有损失函数都可以简单表示为：
# 3.3. Talking Face Generation
$$\begin{aligned}&\mathcal{L}_{\mathrm{GAN}}=\operatorname*{minmax}_{\mathrm{G}}\sum_{\mathrm{D}}^{N_{D}}(\mathbb{E}_{I_{(k)}}[\log\mathrm{D}_{n}(I_{(k)})]\\&&\text{(4}\\&+\mathbb{E}_{f_{cat(k)}}[\log(1-\mathrm{D}_{n}(\mathrm{G}(f_{cat(k)})]),\\&\mathcal{L}_{L_{1}}=\sum_{i=1}^{N_{D}}\|\mathrm{D}_{n}(\mathrm{I}_{(k)})-\mathrm{D}_{n}(\mathrm{G}(f_{cat(k)}))\|_{1},&\text{(}\\&\mathcal{L}_{\mathrm{vgg}}=\sum_{i=1}^{N_{P}}\|\mathrm{VGG}_{n}(I_{(k)})-\mathrm{VGG}_{n}(\mathrm{G}(f_{cat(k)}))\|_{1}.&\text{(}\end{aligned}$$
嵌入在三个模块化空间中的特征组成，用于最终重建目标帧$V$。对于特定情况 ,我们将分别从 $I_( ref)$ $, S_{( k) }$ 和$I_{(k)}^{\prime}$编码的$f_i( ref)$ , $f_{c( k) }^{a}$和$f_p(k)$连接起来，并通过生成器 G 生成$I_{(k)}$
生成器设计。许多以前的研究[14,72,45,44]利用跳过连接来更好地保留输入标识。然而，这样的范例进一步限制了改变姿势的可能性，因为通过连接保留的低级信息极大地影响了生成结果。与它们的结构不同，我们直接将特征的空间维度编码为一个。随着最近生成模型结构的发展，基于样式的生成器在图像生成领域取得了巨大的成功[30,31]。他们在恢复细节和风格操纵方面的表达能力也是我们框架的重要组成部分
在本文中，我们根据经验提出了通过调制卷积生成人脸的方法，该方法已被证明在图像到图像的转换中有效$\mathbb{E}[41]$。详细地说，级联特征 $f_cat(k)=$ $\{f_{i(ref)},f_{c(k)}^a,f_{p(k)}\}$用作潜在代码，以调制生成器卷积
此重建训练不仅将特征映射到图像空间，而且还以互补的方式隐式地确保表示模块化。虽然内容空间$F_c$ 中的特征是没有姿势信息的同步视听表示，但为了抑制重建损失，低维$f_p$会自动补偿姿势信息。此外，在大多数以前的方法中，生成的图像与其监督之间的姿势不匹配，这损害了学习过程。与我们的设置不同，由
4
|
[
[
0.2548828125,
0.10479736328125,
0.30224609375,
0.10479736328125,
0.30224609375,
0.11932373046875,
0.2548828125,
0.11932373046875
],
[
0.4248046875,
0.106689453125,
0.439453125,
0.106689453125,
0.439453125,
0.117431640625,
0.4248046875,
0.117431640625
],
[
0.50244140625,
0.1060791015625,
0.5146484375,
0.1060791015625,
0.5146484375,
0.117431640625,
0.50244140625,
0.117431640625
],
[
0.564453125,
0.10797119140625,
0.609375,
0.10986328125,
0.60888671875,
0.125,
0.56396484375,
0.12249755859375
],
[
0.1895751953125,
0.11614990234375,
0.2099609375,
0.11614990234375,
0.2099609375,
0.1307373046875,
0.1895751953125,
0.1307373046875
],
[
0.31201171875,
0.10919189453125,
0.325927734375,
0.10919189453125,
0.325927734375,
0.117431640625,
0.31201171875,
0.117431640625
],
[
0.79736328125,
0.11614990234375,
0.83837890625,
0.1275634765625,
0.830078125,
0.1458740234375,
0.7890625,
0.1331787109375
],
[
0.11767578125,
0.143310546875,
0.16015625,
0.1483154296875,
0.1568603515625,
0.1641845703125,
0.11358642578125,
0.1590576171875
],
[
0.5849609375,
0.14453125,
0.64892578125,
0.146484375,
0.64892578125,
0.1610107421875,
0.5849609375,
0.1590576171875
],
[
0.71142578125,
0.146484375,
0.72802734375,
0.146484375,
0.72802734375,
0.1622314453125,
0.71142578125,
0.1622314453125
],
[
0.7939453125,
0.15283203125,
0.84619140625,
0.1546630859375,
0.8447265625,
0.173583984375,
0.79345703125,
0.1717529296875
],
[
0.25732421875,
0.1787109375,
0.289306640625,
0.1787109375,
0.289306640625,
0.1875,
0.25732421875,
0.1875
],
[
0.330810546875,
0.1817626953125,
0.414306640625,
0.1900634765625,
0.41015625,
0.208984375,
0.326904296875,
0.2008056640625
],
[
0.250732421875,
0.189453125,
0.2998046875,
0.189453125,
0.2998046875,
0.199462890625,
0.250732421875,
0.199462890625
],
[
0.56884765625,
0.1900634765625,
0.5888671875,
0.1932373046875,
0.58642578125,
0.2008056640625,
0.56640625,
0.1976318359375
],
[
0.423095703125,
0.1932373046875,
0.484375,
0.1932373046875,
0.484375,
0.2083740234375,
0.423095703125,
0.2083740234375
],
[
0.193603515625,
0.1932373046875,
0.226318359375,
0.212158203125,
0.2099609375,
0.2291259765625,
0.17724609375,
0.210205078125
],
[
0.50244140625,
0.1988525390625,
0.521484375,
0.1988525390625,
0.521484375,
0.20263671875,
0.50244140625,
0.20263671875
],
[
0.1422119140625,
0.2083740234375,
0.1544189453125,
0.2083740234375,
0.1544189453125,
0.2139892578125,
0.1422119140625,
0.2139892578125
],
[
0.35693359375,
0.2083740234375,
0.365234375,
0.2083740234375,
0.365234375,
0.212158203125,
0.35693359375,
0.212158203125
],
[
0.366943359375,
0.205810546875,
0.3857421875,
0.205810546875,
0.3857421875,
0.2139892578125,
0.366943359375,
0.2139892578125
],
[
0.101318359375,
0.2197265625,
0.125,
0.2197265625,
0.125,
0.2279052734375,
0.101318359375,
0.2279052734375
],
[
0.26708984375,
0.216552734375,
0.28759765625,
0.216552734375,
0.28759765625,
0.2291259765625,
0.26708984375,
0.2291259765625
],
[
0.098876953125,
0.2298583984375,
0.17724609375,
0.2342529296875,
0.176513671875,
0.251953125,
0.09722900390625,
0.2474365234375
],
[
0.67724609375,
0.236083984375,
0.708984375,
0.236083984375,
0.708984375,
0.248779296875,
0.67724609375,
0.248779296875
],
[
0.81201171875,
0.2392578125,
0.82666015625,
0.2392578125,
0.82666015625,
0.2474365234375,
0.81201171875,
0.2474365234375
],
[
0.69775390625,
0.255126953125,
0.7294921875,
0.255126953125,
0.7294921875,
0.267578125,
0.69775390625,
0.267578125
],
[
0.318603515625,
0.240478515625,
0.328369140625,
0.240478515625,
0.328369140625,
0.2474365234375,
0.318603515625,
0.2474365234375
],
[
0.459228515625,
0.261962890625,
0.49267578125,
0.268310546875,
0.48779296875,
0.285888671875,
0.453369140625,
0.27978515625
],
[
0.529296875,
0.264404296875,
0.57666015625,
0.272705078125,
0.57177734375,
0.28857421875,
0.52392578125,
0.27978515625
],
[
0.7900390625,
0.266357421875,
0.87109375,
0.266357421875,
0.87109375,
0.2841796875,
0.7900390625,
0.2841796875
],
[
0.381591796875,
0.267578125,
0.404296875,
0.267578125,
0.404296875,
0.27978515625,
0.381591796875,
0.27978515625
],
[
0.2998046875,
0.266357421875,
0.3203125,
0.266357421875,
0.3203125,
0.279052734375,
0.2998046875,
0.279052734375
],
[
0.71826171875,
0.26953125,
0.74609375,
0.26953125,
0.74609375,
0.281005859375,
0.71826171875,
0.281005859375
],
[
0.134033203125,
0.281005859375,
0.1405029296875,
0.281005859375,
0.1405029296875,
0.2841796875,
0.134033203125,
0.2841796875
],
[
0.338134765625,
0.28271484375,
0.3857421875,
0.28271484375,
0.3857421875,
0.294189453125,
0.338134765625,
0.294189453125
],
[
0.58642578125,
0.285888671875,
0.5986328125,
0.285888671875,
0.5986328125,
0.290283203125,
0.58642578125,
0.290283203125
],
[
0.273681640625,
0.296630859375,
0.289306640625,
0.296630859375,
0.289306640625,
0.304931640625,
0.273681640625,
0.304931640625
],
[
0.101318359375,
0.28857421875,
0.176513671875,
0.29345703125,
0.1748046875,
0.30810546875,
0.1004638671875,
0.302978515625
],
[
0.45751953125,
0.3017578125,
0.5087890625,
0.3017578125,
0.5087890625,
0.3125,
0.45751953125,
0.3125
],
[
0.68115234375,
0.29736328125,
0.69921875,
0.29736328125,
0.69921875,
0.304931640625,
0.68115234375,
0.304931640625
],
[
0.10455322265625,
0.311279296875,
0.1641845703125,
0.311279296875,
0.1641845703125,
0.322509765625,
0.10455322265625,
0.322509765625
],
[
0.07843017578125,
0.328857421875,
0.89208984375,
0.330078125,
0.89208984375,
0.3447265625,
0.07843017578125,
0.343505859375
],
[
0.080078125,
0.3447265625,
0.89208984375,
0.3447265625,
0.89208984375,
0.35791015625,
0.080078125,
0.35791015625
],
[
0.07843017578125,
0.354248046875,
0.890625,
0.356689453125,
0.890625,
0.375732421875,
0.07843017578125,
0.37255859375
],
[
0.0767822265625,
0.37060546875,
0.89453125,
0.37255859375,
0.89453125,
0.393310546875,
0.0767822265625,
0.391357421875
],
[
0.08087158203125,
0.38818359375,
0.89208984375,
0.38818359375,
0.89208984375,
0.40283203125,
0.08087158203125,
0.40283203125
],
[
0.0767822265625,
0.40087890625,
0.86279296875,
0.399658203125,
0.86279296875,
0.417236328125,
0.0767822265625,
0.41845703125
],
[
0.098876953125,
0.434326171875,
0.468017578125,
0.434326171875,
0.468017578125,
0.447509765625,
0.098876953125,
0.447509765625
],
[
0.50341796875,
0.434326171875,
0.890625,
0.434326171875,
0.890625,
0.447509765625,
0.50341796875,
0.447509765625
],
[
0.080078125,
0.447509765625,
0.468017578125,
0.448974609375,
0.468017578125,
0.46337890625,
0.080078125,
0.462158203125
],
[
0.50341796875,
0.449462890625,
0.890625,
0.448974609375,
0.890625,
0.46337890625,
0.50341796875,
0.464111328125
],
[
0.08251953125,
0.464111328125,
0.468017578125,
0.464111328125,
0.468017578125,
0.477783203125,
0.08251953125,
0.477783203125
],
[
0.50341796875,
0.464111328125,
0.89208984375,
0.464111328125,
0.89208984375,
0.477783203125,
0.50341796875,
0.477783203125
],
[
0.08087158203125,
0.479736328125,
0.468017578125,
0.479736328125,
0.468017578125,
0.49365234375,
0.08087158203125,
0.49365234375
],
[
0.5,
0.479248046875,
0.89208984375,
0.4765625,
0.89208984375,
0.494384765625,
0.5,
0.496826171875
],
[
0.08251953125,
0.494384765625,
0.468017578125,
0.494384765625,
0.468017578125,
0.50830078125,
0.08251953125,
0.50830078125
],
[
0.50341796875,
0.494384765625,
0.890625,
0.494384765625,
0.890625,
0.50830078125,
0.50341796875,
0.50830078125
],
[
0.08251953125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.08251953125,
0.52294921875
],
[
0.50341796875,
0.5087890625,
0.890625,
0.5087890625,
0.890625,
0.52294921875,
0.50341796875,
0.52294921875
],
[
0.08087158203125,
0.5234375,
0.2607421875,
0.52294921875,
0.2607421875,
0.537109375,
0.08087158203125,
0.53857421875
],
[
0.5048828125,
0.5244140625,
0.53515625,
0.5244140625,
0.53515625,
0.537109375,
0.5048828125,
0.537109375
],
[
0.08251953125,
0.54052734375,
0.46728515625,
0.54052734375,
0.46728515625,
0.5537109375,
0.08251953125,
0.5537109375
],
[
0.5048828125,
0.5390625,
0.890625,
0.5390625,
0.890625,
0.55322265625,
0.5048828125,
0.55322265625
],
[
0.08087158203125,
0.55615234375,
0.468017578125,
0.55615234375,
0.468017578125,
0.5693359375,
0.08087158203125,
0.5693359375
],
[
0.50341796875,
0.5537109375,
0.890625,
0.5537109375,
0.890625,
0.568359375,
0.50341796875,
0.568359375
],
[
0.08087158203125,
0.5712890625,
0.468017578125,
0.5712890625,
0.468017578125,
0.58544921875,
0.08087158203125,
0.58544921875
],
[
0.50341796875,
0.568359375,
0.890625,
0.5693359375,
0.890625,
0.583984375,
0.50341796875,
0.58349609375
],
[
0.08087158203125,
0.58544921875,
0.468017578125,
0.58544921875,
0.468017578125,
0.60302734375,
0.08087158203125,
0.60302734375
],
[
0.5048828125,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.5048828125,
0.5986328125
],
[
0.0767822265625,
0.60107421875,
0.373291015625,
0.599609375,
0.373291015625,
0.6142578125,
0.0767822265625,
0.61572265625
],
[
0.5048828125,
0.60107421875,
0.8896484375,
0.60107421875,
0.8896484375,
0.6142578125,
0.5048828125,
0.6142578125
],
[
0.50341796875,
0.61572265625,
0.890625,
0.61572265625,
0.890625,
0.62890625,
0.50341796875,
0.62890625
],
[
0.080078125,
0.630859375,
0.3095703125,
0.630859375,
0.3095703125,
0.64453125,
0.080078125,
0.64453125
],
[
0.50341796875,
0.630859375,
0.64208984375,
0.630859375,
0.64208984375,
0.64453125,
0.50341796875,
0.64453125
],
[
0.1004638671875,
0.65478515625,
0.468017578125,
0.65478515625,
0.468017578125,
0.66845703125,
0.1004638671875,
0.66845703125
],
[
0.63623046875,
0.65087890625,
0.66015625,
0.65087890625,
0.66015625,
0.6640625,
0.63623046875,
0.6640625
],
[
0.07843017578125,
0.66748046875,
0.4697265625,
0.66943359375,
0.4697265625,
0.6875,
0.07843017578125,
0.68505859375
],
[
0.51953125,
0.66357421875,
0.79150390625,
0.66357421875,
0.79150390625,
0.68115234375,
0.51953125,
0.68115234375
],
[
0.61376953125,
0.6796875,
0.62744140625,
0.6796875,
0.62744140625,
0.6845703125,
0.61376953125,
0.6845703125
],
[
0.869140625,
0.67529296875,
0.8935546875,
0.67529296875,
0.8935546875,
0.689453125,
0.869140625,
0.689453125
],
[
0.08087158203125,
0.68310546875,
0.468017578125,
0.68310546875,
0.468017578125,
0.70068359375,
0.08087158203125,
0.70068359375
],
[
0.63818359375,
0.68310546875,
0.66015625,
0.68310546875,
0.66015625,
0.69140625,
0.63818359375,
0.69140625
],
[
0.57861328125,
0.6943359375,
0.81640625,
0.69580078125,
0.81640625,
0.71337890625,
0.57861328125,
0.7119140625
],
[
0.08087158203125,
0.70068359375,
0.2115478515625,
0.7021484375,
0.2115478515625,
0.71630859375,
0.08087158203125,
0.71533203125
],
[
0.08087158203125,
0.71728515625,
0.468017578125,
0.71826171875,
0.468017578125,
0.73291015625,
0.08087158203125,
0.73193359375
],
[
0.57275390625,
0.71728515625,
0.59326171875,
0.71728515625,
0.59326171875,
0.72998046875,
0.57275390625,
0.72998046875
],
[
0.52392578125,
0.72998046875,
0.55615234375,
0.734375,
0.5537109375,
0.7470703125,
0.521484375,
0.7421875
],
[
0.5908203125,
0.72802734375,
0.8056640625,
0.73095703125,
0.8056640625,
0.74755859375,
0.5908203125,
0.744140625
],
[
0.08251953125,
0.734375,
0.468017578125,
0.734375,
0.468017578125,
0.74853515625,
0.08251953125,
0.74853515625
],
[
0.869140625,
0.73095703125,
0.8935546875,
0.73095703125,
0.8935546875,
0.744140625,
0.869140625,
0.744140625
],
[
0.08087158203125,
0.75,
0.46728515625,
0.75,
0.46728515625,
0.76318359375,
0.08087158203125,
0.76318359375
],
[
0.57177734375,
0.7470703125,
0.5947265625,
0.7470703125,
0.5947265625,
0.7587890625,
0.57177734375,
0.7587890625
],
[
0.08087158203125,
0.7646484375,
0.468017578125,
0.7646484375,
0.468017578125,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.57421875,
0.76025390625,
0.5947265625,
0.76025390625,
0.5947265625,
0.7734375,
0.57421875,
0.7734375
],
[
0.521484375,
0.7734375,
0.580078125,
0.77587890625,
0.57861328125,
0.79150390625,
0.51953125,
0.7890625
],
[
0.56787109375,
0.76953125,
0.89453125,
0.771484375,
0.89453125,
0.79248046875,
0.56787109375,
0.79052734375
],
[
0.08251953125,
0.77978515625,
0.468017578125,
0.77978515625,
0.468017578125,
0.79345703125,
0.08251953125,
0.79345703125
],
[
0.08087158203125,
0.79345703125,
0.465576171875,
0.79345703125,
0.465576171875,
0.806640625,
0.08087158203125,
0.806640625
],
[
0.57421875,
0.79248046875,
0.5947265625,
0.79248046875,
0.5947265625,
0.80029296875,
0.57421875,
0.80029296875
],
[
0.08087158203125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82373046875,
0.08087158203125,
0.82373046875
],
[
0.52392578125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.52392578125,
0.82470703125
],
[
0.08416748046875,
0.8271484375,
0.404296875,
0.8271484375,
0.404296875,
0.8369140625,
0.08416748046875,
0.8369140625
],
[
0.5048828125,
0.82568359375,
0.8896484375,
0.82568359375,
0.8896484375,
0.8388671875,
0.5048828125,
0.8388671875
],
[
0.1004638671875,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.1004638671875,
0.85498046875
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.080078125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.07843017578125,
0.86865234375,
0.451904296875,
0.87060546875,
0.451904296875,
0.88818359375,
0.07843017578125,
0.88623046875
],
[
0.5048828125,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.8837890625,
0.4697265625,
0.90087890625,
0.080078125,
0.90380859375
],
[
0.50341796875,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.93603515625,
0.48046875,
0.93603515625
]
] |
[
"fi(ref)",
"F",
"F",
"fi(ref)",
"Ei",
"L",
"Lvgg",
"(ref)",
"fcat(1:K)",
"G",
"/VGG",
"Non-",
"(Fp=f p(1:K)",
"Identity",
"fpo",
"[R,t,s]?",
"Er",
"Pno",
"-",
"",
"",
"",
"Fn",
"V= I'(1:K)",
"Mod",
"D",
"Mod",
"n.",
"FCO",
"Fc(k)",
"LLLGAN",
"F!",
"Fa",
"Mod",
"",
"positive",
"fL",
"L",
"A = S(1:K)",
"hegative",
"mlp",
"",
"Figure 3:The overall pipeline of our Pose-Controllable Audio-Visual System(PC-AVS) framework.The identity reference I(ref) is",
"encoded by E; to the identity space (red). Encoder En encodes video clip V to Fn in the non-identity space (grey). Then it is mapped to",
"Fg in the speech content space (blue), which it shares with F° encoded by Eg from audio spectrograms A. We also draw encodings Fe(G)",
"fromwonativempeTheifaturesieandidntitypasealoshownicalywmaFfatu",
"F, = fp(1:k) in the pose space (yellow). Though motivated by 3D priors, the pose features are not supervised by or necessarily represent",
"the traditional [R, t, s] 3D parameters. Finally, a pair of features (fi(), fp(k), fe(k)} are assembled together and sent to generator G .",
"这个想法与使用3D先验进行无监督3D表示学习的",
"内核的权重。在每个卷积块处，学习多层感知器以将",
"论文有一些相似之处[38，48]。不同的是，我们只在所",
"fcat(k)映射到调制向量M，该调制向量具有与输入特",
"需数据的最小维度上使用先验知识。虽然隐式学习的",
"征的通道相同的维度。对于卷积核权重w中的每个值",
"特征不可能拥有与真实3D姿态参数相同的值，但这种",
"wayz，其中a是其在输人特征通道上的位置，y与输",
"直觉对我们的设计很重要。具有较大尺寸的姿势代码",
"出通道号有关，而表示空间位置，在给定&的M的",
"可能包含不需要的其他信息。Sec中描述了定义的姿势",
"小值的情况下对其进行调制和归一化，以避免以下情",
"代码起作用的原因。3.3",
"况：。",
"标识空间编码。以前的研究已经很好地解决了身份空",
"网络培训。最后，通过图像重建联合训练特征空间模",
"间的学习问题【14，72,75]。当我们在名人[15]的视",
"块化和生成器。我们直接借用【42]中应用的相同损失",
"频上训练我们的网络时，身份标签自然存在。我们的",
"函数。生成的图像和地面实况图像被发送到具有ND",
"身份空间F,=E;(V)=（fi(1)，,fi(K))可以在具有",
"层的多尺度鉴别器D。鉴别器用于计算其层内的特征",
"softmax交叉炳损失，的身份分类上学习",
"图L1距离以及对抗性生成学习。还使用了依赖于具有",
"Np层的预训练VGG网络的感知损失。所有损失函数",
"3.3. Talking Face Generation",
"都可以简单表示为：",
"嵌人在三个模块化空间中的特征组成，用于最终重",
"ND",
"建目标顿V。对于特定情况，我们将分别从I(ref)，S(k)",
"CGAN = minmax (EI() [log Dn(I(k)",
"",
"(4)",
"和 I(k)编码的 fi(ref)，fa(k)和 fp(k)连接起来，并通过",
"n-1",
"+ Efcat() [log(1 -Dn(G(feat(h)),",
"生成器G生成I(k)",
"生成器设计。许多以前的研究[14,72，45，441利用跳过",
"ND",
"CLi",
"JDn(I()) -Dn(G(feat(k)1,",
"连接来更好地保留输人标识。然而，这样的范例进一步",
"(5)",
"限制了改变姿势的可能性，因为通过连接保留的低级",
"i=1",
"信息极大地影响了生成结果。与它们的结构不同，我们",
"NP",
"Cvgg=",
" IIVGGn((k) - VGG,(G(feat(k)I1. (6)",
"直接将特征的空间维度编码为一个。随着最近生成模",
"型结构的发展，基于样式的生成器在图像生成领域取",
"-1",
"得了巨大的成功[30,31]。他们在恢复细节和风格操纵",
"此重建训练不仅将特征映射到图像空间，而且还以",
"方面的表达能力也是我们框架的重要组成部分",
"互补的方式隐式地确保表示模块化。虽然内容空间F。",
"在本文中，我们根据经验提出了通过调制卷积",
"中的特征是没有姿势信息的同步视听表示，但为了抑",
"生成人脸的方法，该方法已被证明在图像到图像的",
"制重建损失，低维f会自动补偿姿势信息。此外，在",
"转换中有效E[41]。详细地说，级联特征fcat(k)",
"大多数以前的方法中，生成的图像与其监督之间的姿",
"(fi(ref)，f(k)，fp(k)}用作潜在代码，以调制生成器卷积",
"势不匹配，这损害了学习过程。与我们的设置不同，由",
"4"
] |
{
"arxiv_id": "2104.11116",
"language": "zh",
"page_idx": 3
}
|
||
2104.11116-zh-4
|
2104.11116-zh-4.jpg
|
Table 1: The quantitative results on LRW [16] and VoxCeleb2 [15]. All methods are compared under the four metrics. For LMD the lower the better, and the higher the better for other metrics. $^{\dagger}$ Note that we directly evaluate the authors' generated samples on VoxCeleb2 under their setting. They have not provided examples on LRW.
<table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th colspan="4">LRW[16]</th>
<th colspan="4">$\operatorname{VoxCeleb2}[15]$</th>
</tr>
<tr>
<th>$SSIM\uparrow$</th>
<th>CPBD$\uparrow$</th>
<th>$LMD\downarrow$</th>
<th>$\mathrm{Sync}_{conf}$ 个</th>
<th>SSIM$\uparrow$</th>
<th>CPBD$\uparrow$</th>
<th>$LMD$ $)\downarrow$</th>
<th>$Sync.$ $\dot{'}conf$ 11</th>
</tr>
<tr>
<td>ATVG$\left [ 10\right ]$</td>
<td>0.810</td>
<td>0.102</td>
<td>5.25</td>
<td>4.1</td>
<td>0.826</td>
<td>0.061</td>
<td>6.49</td>
<td>4.3</td>
</tr>
<tr>
<td>Wav2Lip 1 [44]</td>
<td>0.862</td>
<td>0.152</td>
<td>5.73</td>
<td>6.9</td>
<td>0.846</td>
<td>0.078</td>
<td>12.26</td>
<td>4.5</td>
</tr>
<tr>
<td>MakeitTalk [75]</td>
<td>0.796</td>
<td>0.161</td>
<td>7.13</td>
<td>3.1</td>
<td>0.817</td>
<td>0.068</td>
<td>31.44</td>
<td>2.8</td>
</tr>
<tr>
<td>$\mathrm{RhvthmicHead}^{\dagger}$ $18^{*}$</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>0.779</td>
<td>0.802</td>
<td>14.76</td>
<td>3.8</td>
</tr>
<tr>
<td>Ground Truth</td>
<td>1.000</td>
<td>0.173</td>
<td>0.00</td>
<td>6.5</td>
<td>1.000</td>
<td>0.090</td>
<td>0.00</td>
<td>5.9</td>
</tr>
<tr>
<td>Ours- FixPose</td>
<td>0.815</td>
<td>0.180</td>
<td>6.14</td>
<td>6.3</td>
<td>0.820</td>
<td>0.084</td>
<td>7.68</td>
<td>5.8</td>
</tr>
<tr>
<td>PC- AVS ( Ours) </td>
<td>0.861</td>
<td>0.185</td>
<td>3.93</td>
<td>6.4</td>
<td>0.886</td>
<td>0.083</td>
<td>6.88</td>
<td>5.9</td>
</tr>
</tbody>
</table>
于生成的姿势通过我们的姿势代码$f_p$与地面实况对齐， 因此语音内容特征的学习可以进一步受益于重建损失。这导致更准确的嘴唇同步
整个系统的总体学习目标制定如下：
$$\mathcal{L}_{total}=\mathcal{L}_{\mathrm{GAN}}+\lambda_{1}\mathcal{L}_{L_{1}}+\lambda_{v}\mathcal{L}_{\mathrm{vgg}}+\lambda_{c}\mathcal{L}_{c}+\lambda_{i}\mathcal{L}_{i},$$
,其中$\lambda$是平衡系数
4. Experiments
间的 2s 梅尔频谱图作为条件进行采样。根据经验将$\lambda$ 设置为 1。我们的模型在带有八个 16GBTeslaV100GPU 的 PyTorch$\boxed{\mathbb{E}}[43]$上实现。请注意，我们的身份编码器是在带有$\mathcal{L}_i$的 Voxceleb2$\mathbb{E}[$15] 数据集中提供的标签上预先训练的。语音内容空间也首先进行了预训练，损失为$\mathcal{L}_c$。然后将这些模型加载到生成器和姿势空间学习的整体框架中
比较方法。我们将我们的方法与目前支持任意主题说话人脸生成的最佳模型进行比较。它们是：AVTGE[10] , 基于 2D 地标的方法的代表；Wav2Lip$\mathbb{E}[44]$ ,一种基于重建的方法，声称具有最先进的口型同步结果。MakeitTalk$\underline{\mathbb{E}}[75]$,它利用 3D 地标并根据驾驶音频生成个性化的头部动作。RhythmicHead[8],在不同的设置下产生有节奏的头部运动。请注意，它的代码在本文的最终版本之前是无法运行的，因此我们在图中只显示了它的两个一次性结果。4 , 它是在作者的帮助下生成的。表 1 中的数值比较是在作者提供的 VoxCeleb2样本的设置下进行的，以供参考。具体来说，我们还直接在 GroundTruth 上显示评估。
# 4.1. Experimental Settings
数据集。我们利用了两个在以前的研究中广泛使用的视听数据集，VoxCeleb2$\underline{\mathbb{E}}[$15] 和 LRW$\underline{\mathbb{F}}[$16]。两个数据集都提供检测到的和裁剪的人脸。
## 4.2. Quantitative Evaluation
$\bullet$VoxCeleb2[15].There are a total of 6,112 celebri-
ties in VoxCeleb2 covering over 1 million utterances. While 5,994 speakers lie in the training set, 118 are in the test set. The qualities of the videos differ largely. There are extreme cases with large head pose movements, low-light conditions, and different extents of blurry. No test identity has been seen during training. $\bullet$ Lip Reading in the Wild (LRW) [16]. This dataset is originally proposed for lip reading. It contains over 1000 utterances of 500 different words within each 1- second video. Compared to VoxCeleb2, the videos in this dataset are mostly clean with high-quality and near-frontal faces of BBC news. Thus most of the utterances and identities in the test set are seen during training.
实现细节。$\mathbb{E}_i$的结构是 ResNeXt50[63]。$\mathbb{E}_n$是从[72] 借来的，$\mathbb{E}_c^a$是从[13] 借来的 ResNetSE34。生成器由6 块调制卷积组成。与以前的某些作品不同$\mathbb{E}[9,72,8]$ ,我们不为每个帧对齐面部关键点。所有图像的大小为$224\times224$ 。音频被预处理到 16kHz,然后转换为具有 FFT 窗口大小 1280, 跳长 160 和 80Mel 滤波器组的梅尔频谱图。对于每一帧，0。以目标帧时间步长在中
评估度量。我们对以前在说话人脸生成领域涉及的指标进行定量评估。我们使用 SSIM$\mathbb{E}[60]$来说明生成质量，并从$\mathbb{E}[55]$ 中采用累积概率模糊检测(CPBD)度量$\mathbb{E}[37]$ 来评估结果的清晰度。然后，我们使用嘴周围的地标距离(LMD)$\operatorname{E}[10]$和 SyncNet$\boxed{\mathrm{E}}[16]$中提出的置信度得分($Sync_{conf}$)来说明嘴形和嘴唇同步的准确性。尽管我们不使用地标来生成视频，但我们专门检测地标以进行评估。补充材料中显示了信息较少的指标的结果，例如用于图像质量的 PSNR 和用于身份保留的 CSIM $\mathbb{E}[8,67]$
图像生成范例。我们在自驱动设置下进行实验，选择测试集中每个视频的第一张图像作为身份参考。然后将音频用作驾驶条件，以生成其随附的整个视频。评估是在所有生成的帧和地面真相之间进行的。当不为我们的方法给出姿势代码时，我们可以将其固定为与输入相同的角度，从而使头部保持静止。我们将在此设置下生成的结果称为 Ours-FixPose。由于我们的方法可以由
5
|
[
[
0.080078125,
0.0889892578125,
0.890625,
0.0889892578125,
0.890625,
0.1029052734375,
0.080078125,
0.1029052734375
],
[
0.080078125,
0.10418701171875,
0.89208984375,
0.10418701171875,
0.89208984375,
0.117431640625,
0.080078125,
0.117431640625
],
[
0.080078125,
0.117431640625,
0.450927734375,
0.117431640625,
0.450927734375,
0.1307373046875,
0.080078125,
0.1307373046875
],
[
0.371826171875,
0.146484375,
0.443603515625,
0.146484375,
0.443603515625,
0.1610107421875,
0.371826171875,
0.1610107421875
],
[
0.67333984375,
0.146484375,
0.78125,
0.146484375,
0.78125,
0.1610107421875,
0.67333984375,
0.1610107421875
],
[
0.25732421875,
0.1654052734375,
0.313720703125,
0.1673583984375,
0.31201171875,
0.18310546875,
0.256591796875,
0.18115234375
],
[
0.476318359375,
0.1666259765625,
0.560546875,
0.1685791015625,
0.560546875,
0.1875,
0.476318359375,
0.1849365234375
],
[
0.79736328125,
0.1666259765625,
0.8798828125,
0.169189453125,
0.87841796875,
0.1875,
0.7958984375,
0.184326171875
],
[
0.330810546875,
0.1673583984375,
0.3896484375,
0.1673583984375,
0.3896484375,
0.1817626953125,
0.330810546875,
0.1817626953125
],
[
0.0931396484375,
0.1673583984375,
0.1502685546875,
0.1673583984375,
0.1502685546875,
0.1817626953125,
0.0931396484375,
0.1817626953125
],
[
0.406982421875,
0.1673583984375,
0.463134765625,
0.1673583984375,
0.463134765625,
0.1817626953125,
0.406982421875,
0.1817626953125
],
[
0.576171875,
0.1673583984375,
0.63232421875,
0.1673583984375,
0.63232421875,
0.1817626953125,
0.576171875,
0.1817626953125
],
[
0.650390625,
0.1673583984375,
0.7099609375,
0.1673583984375,
0.7099609375,
0.1817626953125,
0.650390625,
0.1817626953125
],
[
0.72802734375,
0.1673583984375,
0.78173828125,
0.1673583984375,
0.78173828125,
0.1817626953125,
0.72802734375,
0.1817626953125
],
[
0.5048828125,
0.1881103515625,
0.53173828125,
0.1881103515625,
0.53173828125,
0.2052001953125,
0.5048828125,
0.2052001953125
],
[
0.82421875,
0.1881103515625,
0.85302734375,
0.1881103515625,
0.85302734375,
0.2052001953125,
0.82421875,
0.2052001953125
],
[
0.339111328125,
0.189453125,
0.381591796875,
0.189453125,
0.381591796875,
0.2052001953125,
0.339111328125,
0.2052001953125
],
[
0.26318359375,
0.189453125,
0.306396484375,
0.189453125,
0.306396484375,
0.203857421875,
0.26318359375,
0.203857421875
],
[
0.416748046875,
0.189453125,
0.453369140625,
0.189453125,
0.453369140625,
0.2052001953125,
0.416748046875,
0.2052001953125
],
[
0.58251953125,
0.189453125,
0.625,
0.189453125,
0.625,
0.2052001953125,
0.58251953125,
0.2052001953125
],
[
0.65869140625,
0.189453125,
0.69921875,
0.189453125,
0.69921875,
0.2052001953125,
0.65869140625,
0.2052001953125
],
[
0.73779296875,
0.189453125,
0.77294921875,
0.189453125,
0.77294921875,
0.2052001953125,
0.73779296875,
0.2052001953125
],
[
0.26318359375,
0.20263671875,
0.306396484375,
0.2052001953125,
0.3056640625,
0.2203369140625,
0.261474609375,
0.2183837890625
],
[
0.09478759765625,
0.1900634765625,
0.17236328125,
0.1900634765625,
0.17236328125,
0.2052001953125,
0.09478759765625,
0.2052001953125
],
[
0.416748046875,
0.203857421875,
0.453369140625,
0.203857421875,
0.453369140625,
0.2197265625,
0.416748046875,
0.2197265625
],
[
0.5048828125,
0.203857421875,
0.53173828125,
0.203857421875,
0.53173828125,
0.2203369140625,
0.5048828125,
0.2203369140625
],
[
0.58251953125,
0.20263671875,
0.6259765625,
0.2052001953125,
0.625,
0.2203369140625,
0.5810546875,
0.2183837890625
],
[
0.82421875,
0.203857421875,
0.85302734375,
0.203857421875,
0.85302734375,
0.2203369140625,
0.82421875,
0.2203369140625
],
[
0.339111328125,
0.2052001953125,
0.381591796875,
0.2052001953125,
0.381591796875,
0.2197265625,
0.339111328125,
0.2197265625
],
[
0.09478759765625,
0.2052001953125,
0.1895751953125,
0.2052001953125,
0.1895751953125,
0.2197265625,
0.09478759765625,
0.2197265625
],
[
0.65869140625,
0.2052001953125,
0.701171875,
0.2052001953125,
0.701171875,
0.2197265625,
0.65869140625,
0.2197265625
],
[
0.734375,
0.2052001953125,
0.775390625,
0.2052001953125,
0.775390625,
0.2197265625,
0.734375,
0.2197265625
],
[
0.583984375,
0.2183837890625,
0.625,
0.2203369140625,
0.62353515625,
0.236083984375,
0.58251953125,
0.2342529296875
],
[
0.82421875,
0.2183837890625,
0.85302734375,
0.2183837890625,
0.85302734375,
0.23486328125,
0.82421875,
0.23486328125
],
[
0.339111328125,
0.2197265625,
0.3798828125,
0.2197265625,
0.3798828125,
0.23486328125,
0.339111328125,
0.23486328125
],
[
0.26318359375,
0.2197265625,
0.306396484375,
0.2197265625,
0.306396484375,
0.2342529296875,
0.26318359375,
0.2342529296875
],
[
0.418212890625,
0.2197265625,
0.453369140625,
0.2197265625,
0.453369140625,
0.23486328125,
0.418212890625,
0.23486328125
],
[
0.65869140625,
0.2197265625,
0.701171875,
0.2197265625,
0.701171875,
0.2342529296875,
0.65869140625,
0.2342529296875
],
[
0.73388671875,
0.2197265625,
0.775390625,
0.2197265625,
0.775390625,
0.23486328125,
0.73388671875,
0.23486328125
],
[
0.50634765625,
0.2203369140625,
0.53125,
0.2203369140625,
0.53125,
0.2342529296875,
0.50634765625,
0.2342529296875
],
[
0.09478759765625,
0.2203369140625,
0.2034912109375,
0.2203369140625,
0.2034912109375,
0.2342529296875,
0.09478759765625,
0.2342529296875
],
[
0.0931396484375,
0.2342529296875,
0.238525390625,
0.23486328125,
0.238525390625,
0.25,
0.0931396484375,
0.248779296875
],
[
0.58251953125,
0.23486328125,
0.625,
0.23486328125,
0.625,
0.25,
0.58251953125,
0.25
],
[
0.65869140625,
0.23486328125,
0.701171875,
0.23486328125,
0.701171875,
0.25,
0.65869140625,
0.25
],
[
0.734375,
0.23486328125,
0.775390625,
0.23486328125,
0.775390625,
0.25,
0.734375,
0.25
],
[
0.82421875,
0.2342529296875,
0.85302734375,
0.2342529296875,
0.85302734375,
0.250732421875,
0.82421875,
0.250732421875
],
[
0.50341796875,
0.248779296875,
0.53173828125,
0.248779296875,
0.53173828125,
0.266357421875,
0.50341796875,
0.266357421875
],
[
0.82421875,
0.248779296875,
0.85302734375,
0.248779296875,
0.85302734375,
0.26513671875,
0.82421875,
0.26513671875
],
[
0.339111328125,
0.25,
0.381591796875,
0.25,
0.381591796875,
0.26513671875,
0.339111328125,
0.26513671875
],
[
0.2646484375,
0.25,
0.306396484375,
0.25,
0.306396484375,
0.26513671875,
0.2646484375,
0.26513671875
],
[
0.416748046875,
0.25,
0.453369140625,
0.25,
0.453369140625,
0.26513671875,
0.416748046875,
0.26513671875
],
[
0.583984375,
0.25,
0.6259765625,
0.25,
0.6259765625,
0.26513671875,
0.583984375,
0.26513671875
],
[
0.65673828125,
0.25,
0.701171875,
0.25,
0.701171875,
0.26513671875,
0.65673828125,
0.26513671875
],
[
0.736328125,
0.25,
0.77294921875,
0.25,
0.77294921875,
0.26513671875,
0.736328125,
0.26513671875
],
[
0.82421875,
0.264404296875,
0.85302734375,
0.264404296875,
0.85302734375,
0.281005859375,
0.82421875,
0.281005859375
],
[
0.0931396484375,
0.250732421875,
0.188720703125,
0.250732421875,
0.188720703125,
0.26513671875,
0.0931396484375,
0.26513671875
],
[
0.26318359375,
0.26513671875,
0.3056640625,
0.26513671875,
0.3056640625,
0.27978515625,
0.26318359375,
0.27978515625
],
[
0.339111328125,
0.26513671875,
0.381591796875,
0.26513671875,
0.381591796875,
0.27978515625,
0.339111328125,
0.27978515625
],
[
0.416748046875,
0.26513671875,
0.453369140625,
0.26513671875,
0.453369140625,
0.281005859375,
0.416748046875,
0.281005859375
],
[
0.583984375,
0.26513671875,
0.625,
0.26513671875,
0.625,
0.27978515625,
0.583984375,
0.27978515625
],
[
0.65869140625,
0.26513671875,
0.701171875,
0.26513671875,
0.701171875,
0.27978515625,
0.65869140625,
0.27978515625
],
[
0.73779296875,
0.26513671875,
0.771484375,
0.26513671875,
0.771484375,
0.281005859375,
0.73779296875,
0.281005859375
],
[
0.50634765625,
0.266357421875,
0.53173828125,
0.266357421875,
0.53173828125,
0.27978515625,
0.50634765625,
0.27978515625
],
[
0.09478759765625,
0.267578125,
0.1895751953125,
0.267578125,
0.1895751953125,
0.279052734375,
0.09478759765625,
0.279052734375
],
[
0.09234619140625,
0.27978515625,
0.1993408203125,
0.281005859375,
0.1993408203125,
0.29541015625,
0.09234619140625,
0.294189453125
],
[
0.26318359375,
0.27978515625,
0.3056640625,
0.27978515625,
0.3056640625,
0.29541015625,
0.26318359375,
0.29541015625
],
[
0.339111328125,
0.27978515625,
0.381591796875,
0.27978515625,
0.381591796875,
0.29541015625,
0.339111328125,
0.29541015625
],
[
0.418212890625,
0.27978515625,
0.453369140625,
0.27978515625,
0.453369140625,
0.29541015625,
0.418212890625,
0.29541015625
],
[
0.58251953125,
0.27978515625,
0.625,
0.27978515625,
0.625,
0.29541015625,
0.58251953125,
0.29541015625
],
[
0.65869140625,
0.27978515625,
0.701171875,
0.27978515625,
0.701171875,
0.29541015625,
0.65869140625,
0.29541015625
],
[
0.73779296875,
0.27978515625,
0.771484375,
0.27978515625,
0.771484375,
0.29541015625,
0.73779296875,
0.29541015625
],
[
0.82421875,
0.27978515625,
0.85302734375,
0.27978515625,
0.85302734375,
0.296630859375,
0.82421875,
0.296630859375
],
[
0.08251953125,
0.323974609375,
0.4697265625,
0.324462890625,
0.4697265625,
0.339111328125,
0.08251953125,
0.33837890625
],
[
0.50634765625,
0.281005859375,
0.53173828125,
0.281005859375,
0.53173828125,
0.294189453125,
0.50634765625,
0.294189453125
],
[
0.5048828125,
0.324462890625,
0.89208984375,
0.324462890625,
0.89208984375,
0.33837890625,
0.5048828125,
0.33837890625
],
[
0.08087158203125,
0.339111328125,
0.471435546875,
0.339111328125,
0.471435546875,
0.354248046875,
0.08087158203125,
0.354248046875
],
[
0.50244140625,
0.33837890625,
0.8896484375,
0.33837890625,
0.8896484375,
0.3515625,
0.50244140625,
0.3515625
],
[
0.08087158203125,
0.354736328125,
0.2607421875,
0.354736328125,
0.2607421875,
0.36865234375,
0.08087158203125,
0.36865234375
],
[
0.50244140625,
0.352783203125,
0.890625,
0.354248046875,
0.890625,
0.36865234375,
0.50244140625,
0.367431640625
],
[
0.1004638671875,
0.369384765625,
0.35302734375,
0.369384765625,
0.35302734375,
0.38330078125,
0.1004638671875,
0.38330078125
],
[
0.5048828125,
0.369384765625,
0.890625,
0.369384765625,
0.890625,
0.38330078125,
0.5048828125,
0.38330078125
],
[
0.50634765625,
0.385009765625,
0.890625,
0.385009765625,
0.890625,
0.3984375,
0.50634765625,
0.3984375
],
[
0.096435546875,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.410400390625,
0.096435546875,
0.410400390625
],
[
0.50341796875,
0.399658203125,
0.890625,
0.399658203125,
0.890625,
0.4140625,
0.50341796875,
0.4140625
],
[
0.5048828125,
0.415283203125,
0.61767578125,
0.415283203125,
0.61767578125,
0.4287109375,
0.5048828125,
0.4287109375
],
[
0.07843017578125,
0.421630859375,
0.2303466796875,
0.42041015625,
0.2303466796875,
0.43505859375,
0.07843017578125,
0.436279296875
],
[
0.5048828125,
0.43115234375,
0.890625,
0.43115234375,
0.890625,
0.4443359375,
0.5048828125,
0.4443359375
],
[
0.08087158203125,
0.45068359375,
0.214111328125,
0.45068359375,
0.214111328125,
0.464111328125,
0.08087158203125,
0.464111328125
],
[
0.50341796875,
0.44580078125,
0.890625,
0.44580078125,
0.890625,
0.458984375,
0.50341796875,
0.458984375
],
[
0.51953125,
0.460205078125,
0.890625,
0.460205078125,
0.890625,
0.473388671875,
0.51953125,
0.473388671875
],
[
0.080078125,
0.473388671875,
0.289306640625,
0.474853515625,
0.289306640625,
0.4892578125,
0.080078125,
0.488037109375
],
[
0.50341796875,
0.47607421875,
0.89208984375,
0.47607421875,
0.89208984375,
0.490478515625,
0.50341796875,
0.490478515625
],
[
0.50341796875,
0.4912109375,
0.8896484375,
0.4912109375,
0.8896484375,
0.5048828125,
0.50341796875,
0.5048828125
],
[
0.08087158203125,
0.49755859375,
0.46728515625,
0.49755859375,
0.46728515625,
0.51123046875,
0.08087158203125,
0.51123046875
],
[
0.50341796875,
0.505859375,
0.89208984375,
0.505859375,
0.89208984375,
0.51953125,
0.50341796875,
0.51953125
],
[
0.08087158203125,
0.51220703125,
0.468017578125,
0.51220703125,
0.468017578125,
0.5263671875,
0.08087158203125,
0.5263671875
],
[
0.50341796875,
0.51953125,
0.890625,
0.52099609375,
0.890625,
0.53515625,
0.50341796875,
0.5341796875
],
[
0.080078125,
0.5263671875,
0.330810546875,
0.52783203125,
0.330810546875,
0.54248046875,
0.080078125,
0.54150390625
],
[
0.5048828125,
0.5361328125,
0.890625,
0.5361328125,
0.890625,
0.5498046875,
0.5048828125,
0.5498046875
],
[
0.098876953125,
0.5517578125,
0.4697265625,
0.5517578125,
0.4697265625,
0.56494140625,
0.098876953125,
0.56494140625
],
[
0.50244140625,
0.55029296875,
0.8896484375,
0.5498046875,
0.8896484375,
0.564453125,
0.50244140625,
0.56494140625
],
[
0.11273193359375,
0.5673828125,
0.46728515625,
0.5673828125,
0.46728515625,
0.58056640625,
0.11273193359375,
0.58056640625
],
[
0.50341796875,
0.56640625,
0.89208984375,
0.56640625,
0.89208984375,
0.580078125,
0.50341796875,
0.580078125
],
[
0.11273193359375,
0.58203125,
0.468017578125,
0.58203125,
0.468017578125,
0.59521484375,
0.11273193359375,
0.59521484375
],
[
0.50341796875,
0.58056640625,
0.890625,
0.58056640625,
0.890625,
0.5947265625,
0.50341796875,
0.5947265625
],
[
0.11273193359375,
0.59765625,
0.468017578125,
0.59765625,
0.468017578125,
0.611328125,
0.11273193359375,
0.611328125
],
[
0.50244140625,
0.59521484375,
0.72119140625,
0.5966796875,
0.72119140625,
0.611328125,
0.50244140625,
0.60986328125
],
[
0.11273193359375,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.11273193359375,
0.62548828125
],
[
0.50341796875,
0.6220703125,
0.72607421875,
0.6220703125,
0.72607421875,
0.63623046875,
0.50341796875,
0.63623046875
],
[
0.11358642578125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.6396484375,
0.11358642578125,
0.6396484375
],
[
0.111083984375,
0.6416015625,
0.461669921875,
0.642578125,
0.461669921875,
0.6572265625,
0.111083984375,
0.65576171875
],
[
0.5048828125,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.5048828125,
0.65771484375
],
[
0.5048828125,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.5048828125,
0.673828125
],
[
0.1004638671875,
0.66748046875,
0.4697265625,
0.66748046875,
0.4697265625,
0.68115234375,
0.1004638671875,
0.68115234375
],
[
0.5048828125,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.5048828125,
0.689453125
],
[
0.111083984375,
0.681640625,
0.4697265625,
0.681640625,
0.4697265625,
0.69580078125,
0.111083984375,
0.69580078125
],
[
0.50341796875,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.11273193359375,
0.6962890625,
0.471435546875,
0.6962890625,
0.471435546875,
0.71044921875,
0.11273193359375,
0.71044921875
],
[
0.11273193359375,
0.7109375,
0.468017578125,
0.7109375,
0.468017578125,
0.724609375,
0.11273193359375,
0.724609375
],
[
0.50341796875,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.50341796875,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.73486328125,
0.50341796875,
0.73486328125
],
[
0.11273193359375,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.11273193359375,
0.74072265625
],
[
0.50341796875,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.50341796875,
0.75
],
[
0.111083984375,
0.7421875,
0.468017578125,
0.7412109375,
0.468017578125,
0.7548828125,
0.111083984375,
0.755859375
],
[
0.5048828125,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.76513671875,
0.5048828125,
0.76513671875
],
[
0.11273193359375,
0.75830078125,
0.468017578125,
0.75830078125,
0.468017578125,
0.771484375,
0.11273193359375,
0.771484375
],
[
0.111083984375,
0.77099609375,
0.1690673828125,
0.77294921875,
0.1683349609375,
0.787109375,
0.1094970703125,
0.78515625
],
[
0.50244140625,
0.76513671875,
0.89208984375,
0.7646484375,
0.89208984375,
0.77880859375,
0.50244140625,
0.77978515625
],
[
0.5048828125,
0.77978515625,
0.650390625,
0.77978515625,
0.650390625,
0.79443359375,
0.5048828125,
0.79443359375
],
[
0.08087158203125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.5048828125,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5048828125,
0.81005859375
],
[
0.08251953125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.08251953125,
0.82470703125
],
[
0.50341796875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08087158203125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.08087158203125,
0.8388671875
],
[
0.5048828125,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.08251953125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08251953125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.080078125,
0.869140625
],
[
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
],
[
0.50341796875,
0.8720703125,
0.8896484375,
0.8720703125,
0.8896484375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.46728515625,
0.88720703125,
0.46728515625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.939453125,
0.47802734375,
0.939453125
]
] |
[
"Table 1: The quantitative results on LRW [16] and VoxCeleb2 [15] : All methods are compared under the four metrics. For LMD the",
"lower the better, and the higher the better for other metrics. Note that we directly evaluate the authors' generated samples on VoxCeleb2",
"under their setting. They have not provided examples on LRW.",
"LRW [16]",
"VoxCeleb2 [15]",
"SSIM个",
"SynCcong ↑",
"SynCconf ↑",
"CPBD 个",
"Method",
"LMD +",
"SSIM个",
"CPBD 个",
"LMD +",
"4.1",
"4.3",
"0.102",
"0.810",
"5.25",
"0.826",
"0.061",
"6.49",
"0.862",
"ATVG [10]",
"5.73",
"6.9",
"0.846",
"4.5",
"0.152",
"Wav2Lip [44]",
"0.078",
"12.26",
"0.817",
"2.8",
"0.161",
"0.796",
"7.13",
"0.068",
"31.44",
"3.1",
"MakeitTalk [75]",
"Rhythmic Head t [8]",
"0.779",
"0.802",
"14.76",
"3.8",
"6.5",
"5.9",
"0.173",
"1.000",
"0.00",
"1.000",
"0.090",
"0.00",
"5.8",
"Ground Truth",
"0.815",
"0.180",
"6.14",
"0.820",
"0.084",
"7.68",
"6.3",
"Ours-Fix Pose",
"PC-AVS (Ours)",
"0.861",
"0.185",
"3.93",
"0.886",
"0.083",
"6.88",
"5.9",
"于生成的姿势通过我们的姿势代码fp与地面实况对齐，",
"6.4",
"间的2s梅尔频谱图作为条件进行采样。根据经验将入",
"因此语音内容特征的学习可以进一步受益于重建损失。",
"设置为1。我们的模型在带有八个16GBTeslaV100GPU",
"这导致更准确的嘴唇同步",
"的PyTorchE[43]上实现。请注意，我们的身份编码器",
"整个系统的总体学习目标制定如下：",
"是在带有C;的Voxceleb2E[15]数据集中提供的标签",
"上预先训练的。语音内容空间也首先进行了预训练，损",
"Ctotal=LGAN+AiCLi+AuLvgg+C+A,Ci,(7)",
"失为Ce。然后将这些模型加载到生成器和姿势空间学",
"习的整体框架中",
"，其中入是平衡系数",
"比较方法。我们将我们的方法与目前支持任意主题说",
"4.Experiments",
"话人脸生成的最佳模型进行比较。它们是：AVTGE[10]",
"基于2D地标的方法的代表；Wav2LipE[44]，一种",
"4.1.Experimental Settings",
"基于重建的方法，声称具有最先进的口型同步结果。",
"MakeitTalkE[75]，它利用3D地标并根据驾驶音频生",
"数据集。我们利用了两个在以前的研究中广泛使用的",
"成个性化的头部动作。RhythmicHead[8］，在不同的设",
"视听数据集，VoxCeleb2E[15]和LRWE[16]。两个数",
"置下产生有节奏的头部运动。请注意，它的代码在本文",
"据集都提供检测到的和裁剪的人脸。",
"的最终版本之前是无法运行的，因此我们在图中只显",
". VoxCeleb2 [15], There are a total of 6,112 celebri-",
"示了它的两个一次性结果。4，它是在作者的帮助下",
"ties in VoxCeleb2 covering over 1 million utterances.",
"生成的。表1中的数值比较是在作者提供的VoxCeleb2",
"While 5,994 speakers lie in the training set, 118 are in",
"样本的设置下进行的，以供参考。具体来说，我们还直",
"the test set. The qualities of the videos differ largely.",
"接在GroundTruth上显示评估。",
"There are extreme cases with large head pose move",
"4.2.QuantitativeEvaluation",
"ments,low-lightconditions,anddifferent extentsof",
"blurry. No test identity has been seen during training.",
"评估度量。我们对以前在说话人脸生成领域涉及的指",
"标进行定量评估。我们使用SSIME[60]来说明生成质",
"• Lip Reading in the Wild (LRW) [16]. This dataset",
"量，并从E[55]中采用累积概率模糊检测（CPBD）度",
"is originally proposed for lip reading. It contains over",
"量E[37]来评估结果的清晰度。然后，我们使用嘴周围",
"1000 utterances of 500 different words within each 1-",
"second video. Compared to VoxCeleb2, the videos",
"的地标距离（LMD)E[10]和SyncNetE[16]中提出的",
"置信度得分（Syncconf）来说明嘴形和嘴唇同步的准",
"in this dataset are mostly clean with high-quality and",
"确性。尽管我们不使用地标来生成视频，但我们专门检",
"near-frontal faces of BBC news.Thus most of theut",
"测地标以进行评估。补充材料中显示了信息较少的指",
"terances and identities in the test set are seen during",
"training.",
"标的结果，例如用于图像质量的PSNR和用于身份保",
"留的CSIME[8,67]",
"实现细节。E；的结构是ResNeXt50[63]。En是从【72]",
"图像生成范例。我们在自驱动设置下进行实验，选择测",
"借来的，E°是从[13]借来的ResNetSE34。生成器由",
"试集中每个视频的第一张图像作为身份参考。然后将",
"6块调制卷积组成。与以前的某些作品不同E[9,72.8]",
"音频用作驾驶条件，以生成其随附的整个视频。评估是",
"，我们不为每个顿对齐面部关键点。所有图像的大小",
"在所有生成的顿和地面真相之间进行的。当不为我们",
"为224×224。音频被预处理到16kHz，然后转换为具",
"的方法给出姿势代码时，我们可以将其固定为与输人",
"有FFT窗口大小1280，跳长160和80Mel滤波器组的",
"相同的角度，从而使头部保持静止。我们将在此设置下",
"梅尔频谱图。对于每一顿，0。以目标顿时间步长在中",
"生成的结果称为Ours-FixPose。由于我们的方法可以由",
"5"
] |
{
"arxiv_id": "2104.11116",
"language": "zh",
"page_idx": 4
}
|
||
2403.15667-zh-1
|
2403.15667-zh-1.jpg
|
$\textbf{QueryExplorer: 用 于 搜 索 和 探 索 的 交 互 式 查 询 生 成 助 手 }$
## Kaustubh D. Dhole
## $\textbf{Eugene Agichtein}$
Shivam BajajRamraj Chandradevan
Department of Computer Science
Emory University, Atlanta
eugene.agichtein@emory.edu
Department of Computer Science
Emory University, Atlanta kaustubh. dhole@emory. edu
Cox Communication$\Phi$epartment of Computer Science
Atlanta
Emory University, Atlanta rchan31@emory.edu
shivam.bajaj@cox.com
## Abstract
Formulating effective search queries remains a challenging task, particularly when users lack expertise in a specific domain or are not proficient in the language of the content. Providing example documents of interest might be easier for a user. However, such query-by-example scenarios are prone to concept drift, and the retrieval effectiveness is highly sensitive to the query generation method, without a clear way to incorporate user feedback. To enable exploration and to support Human-In-The-Loop experiments we propose QueryExplorer $^{1}-$ an interactive query generation, reformulation, and retrieval interface with support for HuggingFace generation models and PyTerrier's retrieval pipelines and datasets, and extensive logging of human feedback. To allow users to create and modify effective queries, our demo $^2$supports complementary approaches of using LLMs interactively, assisting the user with edits and feedback at multiple stages of the query formulation process. With support for recording fine-grained interactions and user annotations, QueryExplorer can serve as a valuable experimental and research platform for annotation, qualitative evaluation, and conducting Human-in-the-Loop (HITL) experiments for complex search tasks where users struggle to formulate queries.
# 1 Introduction
随着 Internet 越来越多地提供对多种语言和领域的信息的访问，能够以多种语言检索文档变得至关重要。但是，为跨语言和多语言检索创建有效的搜索查询对于用户来说可能是一项艰巨的任务。首先，用户可能不熟悉包含所需信息的文档语言，甚至可能不知道这些信息，因此很难进行有效的查询。其次，大多数人不熟悉其他领域或领域使用的词汇和行话，这可能会阻碍他们制定良好搜索查询的能力。考虑一
$^{\text{l}}$https://github.com/emory-irlab/query-exploren
$^{2}$Demonstration Video of QueryExplorer
种情况，其中用户的任务是识别与法律纠纷相关的文档。他们可能不熟悉专业术语，但拥有相关特定文档的示例
“示例查询”(QBE)是解决此类挑战的一种解决方案。它允许用户通过指定他们正在搜索的示例文档(而不是显式查询)来浏览文档集合。尽管在示例查询领域已经取得了相当大的进步，但在最近使用神经 IR 技术$\overline{\mathrm{E}}(2???)$ 时， 缺乏有效且易于配置的搜索界面工具来探索和注释示例查询实验，尤其是在交互式设置。
虽然一个好的 QBE 界面可能是有价值的，但一个可以促进使用 Human-In-The-Loop(HITL) 设置生成查询的工具可以导致更有效的搜索。大量的先前工作表明，通过提供域内和世界知识$\mathbb{E}(??)$ 以及通过搜索结果的相关性或伪相关性反馈$\overline {\mathrm{E} }( 2? ? ? )$ ,可以通过搜索者的输入$\overline{\mathrm{F}}$ (?)来改进自动生成的查询，并且该过程可以迭代多次，直到搜索者对所提供的搜索结果感到满意为止。研究搜索者行为并收集相关注释数据的研究人员可能需要迭代不同的查询生成器，多个提示和训练策略，多个检索管道，记录各种合并隐式反馈的方式以及严格数量的超参数。例如，搜索引擎分析师可能对搜索者如何编辑查询感兴趣
为了研究这些急需的功能，我们提出了一个可配置的交互式搜索界面工具 QueryExplorer
该工具支持自动和交互式查询生成以及单语言和多语言交互式搜索的重构。QueryExplorer 可以促进和记录查询生成过程和搜索者的交互-从纯 QBE 设置中的查询公式到跟踪结果查询对检索到的搜索结果的影响以及用户的探索。据我们所知，QueryExplorer 是第一个与研究社区共享的交互式查询探索界面
具体来说，QueryExplorer 演示了以下新功
能：
$\bullet$ A simple document search functionality with natural support for multi-lingual and crosslingual search, making it easy for searchers to navigate and analyze search results.
·Support for both automated query genera-
|
[
[
0.202392578125,
0.0955810546875,
0.7958984375,
0.0955810546875,
0.7958984375,
0.10986328125,
0.202392578125,
0.10986328125
],
[
0.118408203125,
0.1383056640625,
0.28369140625,
0.1383056640625,
0.28369140625,
0.1519775390625,
0.118408203125,
0.1519775390625
],
[
0.339111328125,
0.1383056640625,
0.65576171875,
0.1383056640625,
0.65576171875,
0.1519775390625,
0.339111328125,
0.1519775390625
],
[
0.72314453125,
0.1383056640625,
0.87744140625,
0.1383056640625,
0.87744140625,
0.1519775390625,
0.72314453125,
0.1519775390625
],
[
0.0965576171875,
0.155029296875,
0.30322265625,
0.155029296875,
0.30322265625,
0.168701171875,
0.0965576171875,
0.168701171875
],
[
0.330078125,
0.1561279296875,
0.6591796875,
0.1561279296875,
0.6591796875,
0.169189453125,
0.330078125,
0.169189453125
],
[
0.73779296875,
0.1561279296875,
0.94482421875,
0.1561279296875,
0.94482421875,
0.169189453125,
0.73779296875,
0.169189453125
],
[
0.12005615234375,
0.171630859375,
0.2822265625,
0.171630859375,
0.2822265625,
0.185302734375,
0.12005615234375,
0.185302734375
],
[
0.37451171875,
0.171630859375,
0.4248046875,
0.171630859375,
0.4248046875,
0.1859130859375,
0.37451171875,
0.1859130859375
],
[
0.51708984375,
0.171630859375,
0.68408203125,
0.171630859375,
0.68408203125,
0.185302734375,
0.51708984375,
0.185302734375
],
[
0.71875,
0.171630859375,
0.8818359375,
0.171630859375,
0.8818359375,
0.185302734375,
0.71875,
0.185302734375
],
[
0.32080078125,
0.18701171875,
0.476806640625,
0.188232421875,
0.476806640625,
0.203125,
0.32080078125,
0.201904296875
],
[
0.5322265625,
0.18701171875,
0.66748046875,
0.189453125,
0.66748046875,
0.203125,
0.5322265625,
0.20068359375
],
[
0.7001953125,
0.189453125,
0.89990234375,
0.189453125,
0.89990234375,
0.203125,
0.7001953125,
0.203125
],
[
0.2626953125,
0.262451171875,
0.341796875,
0.262451171875,
0.341796875,
0.276611328125,
0.2626953125,
0.276611328125
],
[
0.10662841796875,
0.188232421875,
0.29296875,
0.188232421875,
0.29296875,
0.201904296875,
0.10662841796875,
0.201904296875
],
[
0.5146484375,
0.26318359375,
0.880859375,
0.26318359375,
0.880859375,
0.276611328125,
0.5146484375,
0.276611328125
],
[
0.51318359375,
0.279052734375,
0.87890625,
0.279052734375,
0.87890625,
0.292236328125,
0.51318359375,
0.292236328125
],
[
0.14697265625,
0.292236328125,
0.456787109375,
0.292236328125,
0.456787109375,
0.302734375,
0.14697265625,
0.302734375
],
[
0.5146484375,
0.295654296875,
0.67919921875,
0.295654296875,
0.67919921875,
0.308837890625,
0.5146484375,
0.308837890625
],
[
0.145263671875,
0.3046875,
0.45751953125,
0.3046875,
0.45751953125,
0.319580078125,
0.145263671875,
0.319580078125
],
[
0.52880859375,
0.31005859375,
0.87890625,
0.31005859375,
0.87890625,
0.32373046875,
0.52880859375,
0.32373046875
],
[
0.145263671875,
0.319580078125,
0.459228515625,
0.319580078125,
0.459228515625,
0.333251953125,
0.145263671875,
0.333251953125
],
[
0.5146484375,
0.327880859375,
0.880859375,
0.327880859375,
0.880859375,
0.341552734375,
0.5146484375,
0.341552734375
],
[
0.145263671875,
0.334228515625,
0.45751953125,
0.334228515625,
0.45751953125,
0.347412109375,
0.145263671875,
0.347412109375
],
[
0.51318359375,
0.34326171875,
0.880859375,
0.34326171875,
0.880859375,
0.35693359375,
0.51318359375,
0.35693359375
],
[
0.14697265625,
0.349853515625,
0.456787109375,
0.349853515625,
0.456787109375,
0.35986328125,
0.14697265625,
0.35986328125
],
[
0.14697265625,
0.364013671875,
0.456787109375,
0.364013671875,
0.456787109375,
0.374755859375,
0.14697265625,
0.374755859375
],
[
0.5146484375,
0.35986328125,
0.880859375,
0.35986328125,
0.880859375,
0.37353515625,
0.5146484375,
0.37353515625
],
[
0.14697265625,
0.3759765625,
0.45751953125,
0.3759765625,
0.45751953125,
0.3896484375,
0.14697265625,
0.3896484375
],
[
0.51318359375,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.3896484375,
0.51318359375,
0.3896484375
],
[
0.14697265625,
0.391357421875,
0.456787109375,
0.391357421875,
0.456787109375,
0.402099609375,
0.14697265625,
0.402099609375
],
[
0.51123046875,
0.391357421875,
0.87890625,
0.391357421875,
0.87890625,
0.405029296875,
0.51123046875,
0.405029296875
],
[
0.145263671875,
0.405029296875,
0.456787109375,
0.405029296875,
0.456787109375,
0.418701171875,
0.145263671875,
0.418701171875
],
[
0.51123046875,
0.4091796875,
0.84619140625,
0.4091796875,
0.84619140625,
0.4228515625,
0.51123046875,
0.4228515625
],
[
0.145263671875,
0.418701171875,
0.45751953125,
0.418701171875,
0.45751953125,
0.431640625,
0.145263671875,
0.431640625
],
[
0.5322265625,
0.424560546875,
0.880859375,
0.424560546875,
0.880859375,
0.438232421875,
0.5322265625,
0.438232421875
],
[
0.145263671875,
0.432861328125,
0.45751953125,
0.432861328125,
0.45751953125,
0.446533203125,
0.145263671875,
0.446533203125
],
[
0.5263671875,
0.439453125,
0.8818359375,
0.440673828125,
0.8818359375,
0.454833984375,
0.5263671875,
0.45361328125
],
[
0.145263671875,
0.44775390625,
0.424072265625,
0.44775390625,
0.424072265625,
0.46142578125,
0.145263671875,
0.46142578125
],
[
0.145263671875,
0.46142578125,
0.456787109375,
0.460205078125,
0.456787109375,
0.474365234375,
0.145263671875,
0.4755859375
],
[
0.5146484375,
0.457275390625,
0.880859375,
0.457275390625,
0.880859375,
0.47021484375,
0.5146484375,
0.47021484375
],
[
0.145263671875,
0.4755859375,
0.45751953125,
0.4755859375,
0.45751953125,
0.4892578125,
0.145263671875,
0.4892578125
],
[
0.51318359375,
0.47265625,
0.87890625,
0.47265625,
0.87890625,
0.486328125,
0.51318359375,
0.486328125
],
[
0.145263671875,
0.490478515625,
0.45751953125,
0.488037109375,
0.45751953125,
0.501953125,
0.145263671875,
0.50439453125
],
[
0.51318359375,
0.48681640625,
0.880859375,
0.488037109375,
0.880859375,
0.5029296875,
0.51318359375,
0.501953125
],
[
0.1485595703125,
0.505859375,
0.456787109375,
0.505859375,
0.456787109375,
0.5166015625,
0.1485595703125,
0.5166015625
],
[
0.5146484375,
0.5048828125,
0.8818359375,
0.5048828125,
0.8818359375,
0.5185546875,
0.5146484375,
0.5185546875
],
[
0.14697265625,
0.5185546875,
0.45751953125,
0.5185546875,
0.45751953125,
0.5322265625,
0.14697265625,
0.5322265625
],
[
0.51123046875,
0.52099609375,
0.87890625,
0.52099609375,
0.87890625,
0.53369140625,
0.51123046875,
0.53369140625
],
[
0.14697265625,
0.53369140625,
0.456787109375,
0.53369140625,
0.456787109375,
0.54443359375,
0.14697265625,
0.54443359375
],
[
0.5146484375,
0.5380859375,
0.87890625,
0.5380859375,
0.87890625,
0.548828125,
0.5146484375,
0.548828125
],
[
0.14697265625,
0.5458984375,
0.456787109375,
0.5458984375,
0.456787109375,
0.55859375,
0.14697265625,
0.55859375
],
[
0.51318359375,
0.55419921875,
0.87744140625,
0.55419921875,
0.87744140625,
0.56396484375,
0.51318359375,
0.56396484375
],
[
0.14697265625,
0.56103515625,
0.455078125,
0.56103515625,
0.455078125,
0.57177734375,
0.14697265625,
0.57177734375
],
[
0.51318359375,
0.5693359375,
0.8818359375,
0.5693359375,
0.8818359375,
0.5830078125,
0.51318359375,
0.5830078125
],
[
0.145263671875,
0.576171875,
0.456787109375,
0.576171875,
0.456787109375,
0.5888671875,
0.145263671875,
0.5888671875
],
[
0.145263671875,
0.5888671875,
0.459228515625,
0.5888671875,
0.459228515625,
0.6025390625,
0.145263671875,
0.6025390625
],
[
0.51318359375,
0.58447265625,
0.8818359375,
0.5849609375,
0.8818359375,
0.599609375,
0.51318359375,
0.5986328125
],
[
0.1485595703125,
0.60498046875,
0.456787109375,
0.60498046875,
0.456787109375,
0.615234375,
0.1485595703125,
0.615234375
],
[
0.51318359375,
0.6015625,
0.880859375,
0.6015625,
0.880859375,
0.615234375,
0.51318359375,
0.615234375
],
[
0.14697265625,
0.619140625,
0.456787109375,
0.619140625,
0.456787109375,
0.6298828125,
0.14697265625,
0.6298828125
],
[
0.51123046875,
0.6162109375,
0.8818359375,
0.61767578125,
0.8818359375,
0.6318359375,
0.51123046875,
0.63134765625
],
[
0.145263671875,
0.6318359375,
0.45751953125,
0.6318359375,
0.45751953125,
0.6455078125,
0.145263671875,
0.6455078125
],
[
0.51318359375,
0.63427734375,
0.66162109375,
0.63427734375,
0.66162109375,
0.64794921875,
0.51318359375,
0.64794921875
],
[
0.14697265625,
0.646484375,
0.45751953125,
0.646484375,
0.45751953125,
0.66015625,
0.14697265625,
0.66015625
],
[
0.5322265625,
0.6494140625,
0.880859375,
0.6494140625,
0.880859375,
0.6630859375,
0.5322265625,
0.6630859375
],
[
0.14697265625,
0.662109375,
0.456787109375,
0.662109375,
0.456787109375,
0.6728515625,
0.14697265625,
0.6728515625
],
[
0.5146484375,
0.66455078125,
0.8818359375,
0.6650390625,
0.8818359375,
0.6796875,
0.5146484375,
0.6787109375
],
[
0.145263671875,
0.6748046875,
0.45751953125,
0.6748046875,
0.45751953125,
0.6884765625,
0.145263671875,
0.6884765625
],
[
0.53076171875,
0.681640625,
0.880859375,
0.681640625,
0.880859375,
0.6953125,
0.53076171875,
0.6953125
],
[
0.145263671875,
0.6884765625,
0.27197265625,
0.6884765625,
0.27197265625,
0.70166015625,
0.145263671875,
0.70166015625
],
[
0.5146484375,
0.69873046875,
0.880859375,
0.69873046875,
0.880859375,
0.7119140625,
0.5146484375,
0.7119140625
],
[
0.1502685546875,
0.71630859375,
0.264404296875,
0.71630859375,
0.264404296875,
0.73095703125,
0.1502685546875,
0.73095703125
],
[
0.5146484375,
0.71435546875,
0.880859375,
0.71435546875,
0.880859375,
0.72802734375,
0.5146484375,
0.72802734375
],
[
0.1351318359375,
0.72021484375,
0.1427001953125,
0.72021484375,
0.1427001953125,
0.7255859375,
0.1351318359375,
0.7255859375
],
[
0.515625,
0.732421875,
0.87890625,
0.732421875,
0.87890625,
0.7421875,
0.515625,
0.7421875
],
[
0.1175537109375,
0.71875,
0.133544921875,
0.71875,
0.133544921875,
0.728515625,
0.1175537109375,
0.728515625
],
[
0.1175537109375,
0.7412109375,
0.484375,
0.7412109375,
0.484375,
0.7548828125,
0.1175537109375,
0.7548828125
],
[
0.51318359375,
0.74658203125,
0.8818359375,
0.74658203125,
0.8818359375,
0.76025390625,
0.51318359375,
0.76025390625
],
[
0.118408203125,
0.7587890625,
0.484375,
0.7587890625,
0.484375,
0.7724609375,
0.118408203125,
0.7724609375
],
[
0.51123046875,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.77490234375,
0.51123046875,
0.77490234375
],
[
0.118408203125,
0.77490234375,
0.484375,
0.77490234375,
0.484375,
0.7880859375,
0.118408203125,
0.7880859375
],
[
0.51318359375,
0.779296875,
0.78857421875,
0.779296875,
0.78857421875,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.118408203125,
0.79150390625,
0.484375,
0.79150390625,
0.484375,
0.8046875,
0.118408203125,
0.8046875
],
[
0.53076171875,
0.79443359375,
0.880859375,
0.79443359375,
0.880859375,
0.80810546875,
0.53076171875,
0.80810546875
],
[
0.118408203125,
0.80712890625,
0.484375,
0.80712890625,
0.484375,
0.82080078125,
0.118408203125,
0.82080078125
],
[
0.51318359375,
0.81103515625,
0.5400390625,
0.81103515625,
0.5400390625,
0.82373046875,
0.51318359375,
0.82373046875
],
[
0.1175537109375,
0.822265625,
0.484375,
0.82177734375,
0.484375,
0.8359375,
0.1175537109375,
0.83740234375
],
[
0.118408203125,
0.8388671875,
0.486083984375,
0.8388671875,
0.486083984375,
0.8525390625,
0.118408203125,
0.8525390625
],
[
0.5341796875,
0.83447265625,
0.880859375,
0.83447265625,
0.880859375,
0.84716796875,
0.5341796875,
0.84716796875
],
[
0.1175537109375,
0.85400390625,
0.486083984375,
0.85498046875,
0.486083984375,
0.869140625,
0.1175537109375,
0.8681640625
],
[
0.5458984375,
0.849609375,
0.8818359375,
0.8486328125,
0.8818359375,
0.86328125,
0.5458984375,
0.8642578125
],
[
0.118408203125,
0.87158203125,
0.486083984375,
0.87158203125,
0.486083984375,
0.88525390625,
0.118408203125,
0.88525390625
],
[
0.54833984375,
0.8662109375,
0.880859375,
0.8662109375,
0.880859375,
0.8798828125,
0.54833984375,
0.8798828125
],
[
0.54736328125,
0.8818359375,
0.82958984375,
0.88134765625,
0.82958984375,
0.8955078125,
0.54736328125,
0.896484375
],
[
0.139404296875,
0.8955078125,
0.4248046875,
0.8955078125,
0.4248046875,
0.9091796875,
0.139404296875,
0.9091796875
],
[
0.139404296875,
0.90673828125,
0.38623046875,
0.90771484375,
0.38623046875,
0.92138671875,
0.139404296875,
0.92041015625
],
[
0.5322265625,
0.90625,
0.8818359375,
0.90771484375,
0.8818359375,
0.9228515625,
0.5322265625,
0.92041015625
]
] |
[
"QueryExplorer：用于搜索和探索的交互式香询生成助手",
"Kaustubh D. Dhole",
"Shivam BajajRamraj Chandradevan",
"Eugene Agichtein",
"Department of Computer Science",
"Cox CommunicatiorBepartment of Computer Science",
"Department of Computer Science",
"Emory University, Atlanta",
"Atlanta",
"Emory University, Atlanta",
"Emory University, Atlanta",
"shivam.bajaj@cox.com",
"rchan31@emory.edu",
"eugene.agichtein@emory.edu",
"Abstract",
"kaustubh.dhole@emory.edu",
"种情况，其中用户的任务是识别与法律纠纷相",
"关的文档。他们可能不熟悉专业术语，但拥有",
"Formulatingeffectivesearchqueriesremainsa",
"相关特定文档的示例",
"challenging task, particularly when users lack",
"“示例查询”（OBE）是解决此类挑战的一种",
"expertise in a specific domain or are not profi-",
"解决方案。它允许用户通过指定他们正在搜索",
"cient in the language of the content. Providing",
"的示例文档（而不是显式查询）来浏览文档集",
"exampledocumentsofinterestmightbeeasier",
"for a user.However, such query-by-example",
"合。尽管在示例查询领域已经取得了相当大的",
"scenarios are prone to concept drift, and the re",
"进步，但在最近使用神经IR技术E(????时，",
"trieval effectiveness is highly sensitive to the",
"缺乏有效且易于配置的搜索界面工具来探索和",
"query generation method, without a clear way",
"注释示例查询实验，尤其是在交互式设置",
"to incorporate user feedback. To enable ex-",
"虽然一个好的QBE界面可能是有价值的，但",
"ploration and to support Human-In-The-Loop",
"个可以促进使用Human-In-The-Loop(HITL)",
"experiments we propose QueryExplorer",
"an interactive query generation, reformulation,",
"设置生成查询的工具可以导致更有效的搜索。",
"and retrieval interface with support for Hug.",
"大量的先前工作表明，通过提供域内和世界知",
"gingFace generation models and PyTerrier's",
"识E(??）以及通过搜索结果的相关性或伪相",
"retrievalpipelines anddatasets,andextensive",
"关性反馈E（???？），可以通过搜索者的输入E",
"logging of human feedback. To allow users to",
"（?）来改进自动生成的查询，并且该过程可以",
"createandmodifyeffectivequeries,ourdemo",
"选代多次，直到搜索者对所提供的搜索结果感",
"2 supports complementary approaches of using",
"到满意为止。研究搜索者行为并收集相关注释",
"LLMs interactively, assisting the user with ed",
"数据的研究人员可能需要迭代不同的查询生成",
"its and feedback at multiple stages of the query",
"formulation process. With support for record-",
"器，多个提示和训练策略，多个检索管道，记",
"ingfine-grainedinteractions anduserannota",
"录各种合并隐式反馈的方式以及严格数量的超",
"tions,QueryExplorercanserveasavaluable",
"参数。例如，搜索引擎分析师可能对搜索者如",
"experimental and research platform for anno-",
"何编辑查询感兴趣",
"tation, qualitative evaluation, and conducting",
"为了研究这些急需的功能，我们提出了一个",
"Human-in-the-Loop (HITL)experiments for",
"可配置的交互式搜索界面工具QueryExplorer",
"complex search tasks where users struggle to",
"该工具支持自动和交互式查询生成以及单语",
"formulate queries.",
"言和多语言交互式搜索的重构。QueryExplorer",
"Introduction",
"可以促进和记录查询生成过程和搜索者的交",
"",
"互-从纯OBE设置中的香询公式到跟踪结果香",
"1",
"随着Internet越来越多地提供对多种语言和领",
"询对检索到的搜索结果的影响以及用户的探",
"域的信息的访问，能够以多种语言检索文档变",
"索。据我们所知，QueryExplorer是第一个与研",
"得至关重要。但是，为跨语言和多语言检索创",
"究社区共享的交互式查询探索界面",
"建有效的搜索查询对于用户来说可能是一项艰",
"具体来说，QueryExplorer演示了以下新功",
"巨的任务。首先，用户可能不熟悉包含所需信",
"能:",
"息的文档语言，甚至可能不知道这些信息，因",
"此很难进行有效的查询。其次，大多数人不熟",
": A simple document search functionality with",
"悉其他领域或领域使用的词汇和行话，这可能",
"natural support for multi-lingual and cross-",
"会阻碍他们制定良好搜索查询的能力。考虑一",
"lingual search, making it easy for searchers",
"to navigate and analyze search results.",
"https://github.com/emory-irlab/query-explorer",
"2DemonstrationVideoof QueryExplorer",
". Support for both automated query genera-"
] |
{
"arxiv_id": "2403.15667",
"language": "zh",
"page_idx": 1
}
|
||
2403.15667-zh-3
|
2403.15667-zh-3.jpg
|
## 3 $\textbf{QueryExplorer}$
QueryExplorer 接口由 2 个选项卡组成--“查询生成”选项卡和“设置”选项卡。它们都描述如下。“查询生成”选项卡显示给最终用户或搜索者和注释者，“设置”选项卡保留给研究人员$^{5}$,他们希望通过调查不同的设置来收集用于查询生成和 IR 研究的数据。完整的界面是使用 HuggingFace 的 Gradio 平台构建的。Gradio (?)是一个开源 Python 包，用于快速创建易于使用、可配置的 UI 组件，并已被广泛用于机器学习模型
# 3.1 Searcher's Tab: Query Generation
此查询生成选项卡可作为搜索者和注释者的简单界面，允许端到端查询生成(QG)-i)临时QG: 用户可以自己编写搜索查询 ii) 通过示例查询：用户可以通过提示 HuggingFace$\boxed{\mathrm{F}}(?)$模型生成查询，并选择适当的检索，通过 XIRrier 重新生成文档-通过有用的关键字。我们显示前 k 个相关文档及其源语言翻译 (如果适用)
每个生成的查询都可以单独使用或组合使用以检索文档。用户可以进一步编辑查询，并从查询重组器的输出中获得帮助。用户可以进一步与检索到的文档或段落进行交互，并为每个文档提供相关性注释
我们现在描述为演示的上述每个设置提供的默认模型。通过对配置设置或代码进行少量修改，可以轻松地用研究人员的选择替换其中的每一个
3.1.1 Query-By-Example Generation (QBE)
我们使用|E|flan-t5-xxl 模型|E|(?),它是经过指令调整的E(?) 版本的文本到文本转换器T5$\boxed{\mathrm{E}}(?)$。它已经在大量任务上进行了微调，使其方便$\overline{\mathbb{E}}(?)$ 用于学习新任务。向用户显示的默认版本是$\overline{\mathrm{E}}$flan-t5-xxl 上的一些包装器-它接受用户的示例文档或段落，并在$\overline{\mathrm{EGenerate}}$ a query given the following document 之前加上一条指令，以及来自 MSMarco 的 3 个文档查询对作为样例。
## 3.1.2 Query Reformulation (QR)
我们使用零拍方法为给定查询生成关键字。flan-t5-xxl 被传递指令 Improve the search effectiveness by suggesting expansion terms for the query(?)以及原始查询作为输入。零拍查询重构(???)最近流行扩展查询，通过大型语言模型的零拍提示来提高检索
5We use the term searchers and researchers to differentiate between the higher level goals of the two tabs but both could encompass analysts/annotators/testers, etc.
效率。面向用户的界面可以提供减轻不良重新
配方的机会(?)(请参阅附录列表 1)
# 3.1.3 Retrieving Documents
对于检索，我们使用 PyTerrier 检索器作为默认。PyTerrier 的体系结构固有地设计为支持对检索器和重新排序器的操作，以构建端到端检索管道，并且一直是信息检索研究人员中检索引擎的流行选择。在 QueryExplorer 中，研究人员也可以在下面的字典中添加自己的自定义PyTerrier 管道 (请参阅附录列表 2)。这也将反映在“设置”选项卡的下拉列表中
3.1.4 $\underset {( \mathrm{RF}) }{\operatorname* { \operatorname* { Incorporating} } }$ Relevance Feedback
我们为搜索者提供了通过利用他们选择的检索文档来改进当前查询的能力。我们使用零拍方法将用户选择的文档合并为$\mathbb{E}??$ 的样式。提示用户选择的文档和查询通过指令区Based on the given context information $C$ ,generate keywords for the following query 重新生成关键字，其中$C$ 是用户选择的文档(请参阅附录列表$\mathbb{E}$3 )
# 3.2 PyTerrier and HuggingFace support
为了加快 IR 研究人员进行各种实验的原型制作过程，QueryExplorer 集成了 PyTerrier$[\mathbb{E}(?)$ 支持。这种集成可以利用 QueryExplorer 接口中通过 PyTerrier 创建的检索管道，从而增强其跨不同检索和重新排序算法的注释和定性分析功能。在演示过程中，我们展示了该界面使用BM25 管道显示搜索结果的功能，突出了根据需要用其他自定义管道替换该界面的灵活性。此功能本质上为 PyTerrier 检索管道添加了一层定性分析。此外，为了扩大 PyTerrier 在处理IR 数据集中的实用性，QueryExplorer 旨在促进对这些数据集中的文档进行索引，从而实现跨多个基准和数据集的定性实验
认识到 HuggingFace 的$\mathbb{E}(?)$ 模型在研究社区中的广泛采用，QueryExplorer 利用这些模型进行查询生成，重构和合并反馈。这允许对各种大型语言模型的搜索功能进行全面评估
虽然我们在设计时考虑到 PyTerrier 和 HuggingFace 生态系统的受欢迎程度，但使用其他软件包的数据集和模型也可以在其系统上利用QueryExplorer 接口
## 3.3 Relevance Annotations
我们允许对每个文档进行相关性注释，以帮助研究人员通过滑块组件收集相关性注释。注释立即保存在单独的 JSON 文件中
|
[
[
0.11248779296875,
0.08489990234375,
0.285400390625,
0.08612060546875,
0.285400390625,
0.1009521484375,
0.11248779296875,
0.09979248046875
],
[
0.51318359375,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.5146484375,
0.10272216796875,
0.8212890625,
0.10272216796875,
0.8212890625,
0.11639404296875,
0.5146484375,
0.11639404296875
],
[
0.118408203125,
0.1134033203125,
0.486083984375,
0.1134033203125,
0.486083984375,
0.1270751953125,
0.118408203125,
0.1270751953125
],
[
0.118408203125,
0.1300048828125,
0.483642578125,
0.1300048828125,
0.483642578125,
0.14013671875,
0.118408203125,
0.14013671875
],
[
0.5087890625,
0.1300048828125,
0.748046875,
0.1312255859375,
0.748046875,
0.1455078125,
0.5087890625,
0.144287109375
],
[
0.12005615234375,
0.1455078125,
0.486083984375,
0.1455078125,
0.486083984375,
0.1591796875,
0.12005615234375,
0.1591796875
],
[
0.51318359375,
0.152587890625,
0.880859375,
0.152587890625,
0.880859375,
0.166259765625,
0.51318359375,
0.166259765625
],
[
0.12005615234375,
0.163330078125,
0.484375,
0.163330078125,
0.484375,
0.17333984375,
0.12005615234375,
0.17333984375
],
[
0.51318359375,
0.169189453125,
0.87744140625,
0.169189453125,
0.87744140625,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.12005615234375,
0.1787109375,
0.484375,
0.1787109375,
0.484375,
0.189453125,
0.12005615234375,
0.189453125
],
[
0.51318359375,
0.185302734375,
0.880859375,
0.185302734375,
0.880859375,
0.198974609375,
0.51318359375,
0.198974609375
],
[
0.12005615234375,
0.1953125,
0.484375,
0.1953125,
0.484375,
0.2060546875,
0.12005615234375,
0.2060546875
],
[
0.5146484375,
0.203125,
0.87890625,
0.203125,
0.87890625,
0.213134765625,
0.5146484375,
0.213134765625
],
[
0.118408203125,
0.210205078125,
0.48779296875,
0.210205078125,
0.48779296875,
0.223876953125,
0.118408203125,
0.223876953125
],
[
0.51123046875,
0.21728515625,
0.880859375,
0.21728515625,
0.880859375,
0.23095703125,
0.51123046875,
0.23095703125
],
[
0.118408203125,
0.2257080078125,
0.486083984375,
0.2257080078125,
0.486083984375,
0.2392578125,
0.118408203125,
0.2392578125
],
[
0.51318359375,
0.2327880859375,
0.87890625,
0.2327880859375,
0.87890625,
0.2464599609375,
0.51318359375,
0.2464599609375
],
[
0.1175537109375,
0.2423095703125,
0.486083984375,
0.2423095703125,
0.486083984375,
0.255859375,
0.1175537109375,
0.255859375
],
[
0.51318359375,
0.2493896484375,
0.87890625,
0.2493896484375,
0.87890625,
0.26318359375,
0.51318359375,
0.26318359375
],
[
0.118408203125,
0.2587890625,
0.26611328125,
0.2587890625,
0.26611328125,
0.2724609375,
0.118408203125,
0.2724609375
],
[
0.5146484375,
0.267333984375,
0.80419921875,
0.267333984375,
0.80419921875,
0.27783203125,
0.5146484375,
0.27783203125
],
[
0.11419677734375,
0.2861328125,
0.435791015625,
0.2880859375,
0.435791015625,
0.3017578125,
0.11419677734375,
0.2998046875
],
[
0.51318359375,
0.29443359375,
0.57421875,
0.29443359375,
0.57421875,
0.305908203125,
0.51318359375,
0.305908203125
],
[
0.56689453125,
0.29443359375,
0.84619140625,
0.29443359375,
0.84619140625,
0.30810546875,
0.56689453125,
0.30810546875
],
[
0.118408203125,
0.31005859375,
0.48193359375,
0.31005859375,
0.48193359375,
0.320556640625,
0.118408203125,
0.320556640625
],
[
0.568359375,
0.311279296875,
0.60546875,
0.311279296875,
0.60546875,
0.32373046875,
0.568359375,
0.32373046875
],
[
0.12005615234375,
0.32666015625,
0.484375,
0.32666015625,
0.484375,
0.34033203125,
0.12005615234375,
0.34033203125
],
[
0.51318359375,
0.33203125,
0.880859375,
0.33203125,
0.880859375,
0.345703125,
0.51318359375,
0.345703125
],
[
0.118408203125,
0.342529296875,
0.486083984375,
0.342529296875,
0.486083984375,
0.355712890625,
0.118408203125,
0.355712890625
],
[
0.51123046875,
0.347412109375,
0.880859375,
0.347412109375,
0.880859375,
0.361083984375,
0.51123046875,
0.361083984375
],
[
0.1175537109375,
0.358154296875,
0.48779296875,
0.35693359375,
0.48779296875,
0.371826171875,
0.1175537109375,
0.372314453125
],
[
0.51123046875,
0.363525390625,
0.880859375,
0.363525390625,
0.880859375,
0.377197265625,
0.51123046875,
0.377197265625
],
[
0.1175537109375,
0.372314453125,
0.486083984375,
0.37353515625,
0.486083984375,
0.388427734375,
0.1175537109375,
0.38720703125
],
[
0.51318359375,
0.380126953125,
0.8818359375,
0.380126953125,
0.8818359375,
0.393798828125,
0.51318359375,
0.393798828125
],
[
0.118408203125,
0.3896484375,
0.484375,
0.3896484375,
0.484375,
0.402587890625,
0.118408203125,
0.402587890625
],
[
0.51123046875,
0.394287109375,
0.880859375,
0.3955078125,
0.880859375,
0.4091796875,
0.51123046875,
0.407958984375
],
[
0.12005615234375,
0.40673828125,
0.47607421875,
0.40673828125,
0.47607421875,
0.42041015625,
0.12005615234375,
0.42041015625
],
[
0.509765625,
0.412109375,
0.87890625,
0.412109375,
0.87890625,
0.42578125,
0.509765625,
0.42578125
],
[
0.1376953125,
0.424560546875,
0.483642578125,
0.424560546875,
0.483642578125,
0.435302734375,
0.1376953125,
0.435302734375
],
[
0.51123046875,
0.427490234375,
0.87890625,
0.427490234375,
0.87890625,
0.441162109375,
0.51123046875,
0.441162109375
],
[
0.12005615234375,
0.441162109375,
0.486083984375,
0.441162109375,
0.486083984375,
0.451904296875,
0.12005615234375,
0.451904296875
],
[
0.51318359375,
0.44482421875,
0.73779296875,
0.44482421875,
0.73779296875,
0.457763671875,
0.51318359375,
0.457763671875
],
[
0.118408203125,
0.4560546875,
0.47265625,
0.4560546875,
0.47265625,
0.4697265625,
0.118408203125,
0.4697265625
],
[
0.118408203125,
0.471435546875,
0.48779296875,
0.471435546875,
0.48779296875,
0.485107421875,
0.118408203125,
0.485107421875
],
[
0.51123046875,
0.474365234375,
0.5615234375,
0.474365234375,
0.5615234375,
0.486328125,
0.51123046875,
0.486328125
],
[
0.55517578125,
0.474365234375,
0.8447265625,
0.474365234375,
0.8447265625,
0.488037109375,
0.55517578125,
0.488037109375
],
[
0.12005615234375,
0.4892578125,
0.28369140625,
0.4892578125,
0.28369140625,
0.49951171875,
0.12005615234375,
0.49951171875
],
[
0.51123046875,
0.49755859375,
0.87890625,
0.49755859375,
0.87890625,
0.51123046875,
0.51123046875,
0.51123046875
],
[
0.1368408203125,
0.50439453125,
0.486083984375,
0.50439453125,
0.486083984375,
0.51708984375,
0.1368408203125,
0.51708984375
],
[
0.51123046875,
0.51318359375,
0.880859375,
0.5126953125,
0.880859375,
0.52685546875,
0.51123046875,
0.52783203125
],
[
0.118408203125,
0.521484375,
0.483642578125,
0.521484375,
0.483642578125,
0.5322265625,
0.118408203125,
0.5322265625
],
[
0.1175537109375,
0.53515625,
0.483642578125,
0.53515625,
0.483642578125,
0.548828125,
0.1175537109375,
0.548828125
],
[
0.5146484375,
0.52978515625,
0.880859375,
0.52978515625,
0.880859375,
0.54345703125,
0.5146484375,
0.54345703125
],
[
0.5146484375,
0.5458984375,
0.880859375,
0.5458984375,
0.880859375,
0.55859375,
0.5146484375,
0.55859375
],
[
0.12005615234375,
0.55419921875,
0.172119140625,
0.55419921875,
0.172119140625,
0.5654296875,
0.12005615234375,
0.5654296875
],
[
0.5146484375,
0.5634765625,
0.87890625,
0.5634765625,
0.87890625,
0.57373046875,
0.5146484375,
0.57373046875
],
[
0.1175537109375,
0.5791015625,
0.480224609375,
0.5791015625,
0.480224609375,
0.59326171875,
0.1175537109375,
0.59326171875
],
[
0.51123046875,
0.57666015625,
0.87744140625,
0.57666015625,
0.87744140625,
0.59033203125,
0.51123046875,
0.59033203125
],
[
0.118408203125,
0.60107421875,
0.486083984375,
0.60107421875,
0.486083984375,
0.6142578125,
0.118408203125,
0.6142578125
],
[
0.5146484375,
0.595703125,
0.87890625,
0.595703125,
0.87890625,
0.60546875,
0.5146484375,
0.60546875
],
[
0.51123046875,
0.609375,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.51123046875,
0.62255859375
],
[
0.118408203125,
0.61767578125,
0.486083984375,
0.61767578125,
0.486083984375,
0.63134765625,
0.118408203125,
0.63134765625
],
[
0.51318359375,
0.626953125,
0.87158203125,
0.626953125,
0.87158203125,
0.64013671875,
0.51318359375,
0.64013671875
],
[
0.1175537109375,
0.6328125,
0.486083984375,
0.6328125,
0.486083984375,
0.646484375,
0.1175537109375,
0.646484375
],
[
0.5146484375,
0.64453125,
0.87744140625,
0.64453125,
0.87744140625,
0.65478515625,
0.5146484375,
0.65478515625
],
[
0.12005615234375,
0.65087890625,
0.483642578125,
0.65087890625,
0.483642578125,
0.6611328125,
0.12005615234375,
0.6611328125
],
[
0.51123046875,
0.6591796875,
0.880859375,
0.6591796875,
0.880859375,
0.6728515625,
0.51123046875,
0.6728515625
],
[
0.118408203125,
0.6650390625,
0.486083984375,
0.6650390625,
0.486083984375,
0.6787109375,
0.118408203125,
0.6787109375
],
[
0.51318359375,
0.6748046875,
0.880859375,
0.6748046875,
0.880859375,
0.6884765625,
0.51318359375,
0.6884765625
],
[
0.118408203125,
0.68115234375,
0.486083984375,
0.68115234375,
0.486083984375,
0.69482421875,
0.118408203125,
0.69482421875
],
[
0.51123046875,
0.69140625,
0.767578125,
0.69140625,
0.767578125,
0.70166015625,
0.51123046875,
0.70166015625
],
[
0.1158447265625,
0.69775390625,
0.48779296875,
0.69677734375,
0.48779296875,
0.71142578125,
0.1158447265625,
0.7119140625
],
[
0.118408203125,
0.71337890625,
0.48779296875,
0.71337890625,
0.48779296875,
0.72705078125,
0.118408203125,
0.72705078125
],
[
0.53076171875,
0.7080078125,
0.880859375,
0.7080078125,
0.880859375,
0.72265625,
0.53076171875,
0.72265625
],
[
0.51123046875,
0.72265625,
0.880859375,
0.7236328125,
0.880859375,
0.73828125,
0.51123046875,
0.73681640625
],
[
0.118408203125,
0.72998046875,
0.273681640625,
0.72998046875,
0.273681640625,
0.74365234375,
0.118408203125,
0.74365234375
],
[
0.51318359375,
0.7392578125,
0.87744140625,
0.7392578125,
0.87744140625,
0.7529296875,
0.51318359375,
0.7529296875
],
[
0.173828125,
0.75732421875,
0.391357421875,
0.75732421875,
0.391357421875,
0.77099609375,
0.173828125,
0.77099609375
],
[
0.51318359375,
0.755859375,
0.8447265625,
0.755859375,
0.8447265625,
0.76953125,
0.51318359375,
0.76953125
],
[
0.53076171875,
0.77392578125,
0.880859375,
0.77392578125,
0.880859375,
0.78759765625,
0.53076171875,
0.78759765625
],
[
0.118408203125,
0.75830078125,
0.181396484375,
0.75830078125,
0.181396484375,
0.76953125,
0.118408203125,
0.76953125
],
[
0.12005615234375,
0.779296875,
0.484375,
0.779296875,
0.484375,
0.7890625,
0.12005615234375,
0.7890625
],
[
0.118408203125,
0.79345703125,
0.48779296875,
0.79345703125,
0.48779296875,
0.80712890625,
0.118408203125,
0.80712890625
],
[
0.51123046875,
0.7890625,
0.880859375,
0.7880859375,
0.880859375,
0.802734375,
0.51123046875,
0.80419921875
],
[
0.118408203125,
0.81005859375,
0.486083984375,
0.81005859375,
0.486083984375,
0.82373046875,
0.118408203125,
0.82373046875
],
[
0.51318359375,
0.8056640625,
0.880859375,
0.8056640625,
0.880859375,
0.8193359375,
0.51318359375,
0.8193359375
],
[
0.118408203125,
0.826171875,
0.48779296875,
0.826171875,
0.48779296875,
0.8388671875,
0.118408203125,
0.8388671875
],
[
0.509765625,
0.822265625,
0.6650390625,
0.82080078125,
0.6650390625,
0.83447265625,
0.509765625,
0.8359375
],
[
0.118408203125,
0.8427734375,
0.486083984375,
0.8427734375,
0.486083984375,
0.85546875,
0.118408203125,
0.85546875
],
[
0.1175537109375,
0.85791015625,
0.486083984375,
0.85791015625,
0.486083984375,
0.87158203125,
0.1175537109375,
0.87158203125
],
[
0.51123046875,
0.85400390625,
0.73779296875,
0.85400390625,
0.73779296875,
0.8642578125,
0.51123046875,
0.8642578125
],
[
0.51318359375,
0.87451171875,
0.87744140625,
0.87451171875,
0.87744140625,
0.88818359375,
0.51318359375,
0.88818359375
],
[
0.1368408203125,
0.8818359375,
0.484375,
0.88427734375,
0.484375,
0.8984375,
0.1368408203125,
0.896484375
],
[
0.118408203125,
0.896484375,
0.486083984375,
0.896484375,
0.486083984375,
0.91015625,
0.118408203125,
0.91015625
],
[
0.51123046875,
0.89111328125,
0.87890625,
0.89111328125,
0.87890625,
0.90478515625,
0.51123046875,
0.90478515625
],
[
0.118408203125,
0.90673828125,
0.376220703125,
0.9091796875,
0.376220703125,
0.9228515625,
0.118408203125,
0.92041015625
],
[
0.515625,
0.9091796875,
0.76318359375,
0.9091796875,
0.76318359375,
0.91943359375,
0.515625,
0.91943359375
]
] |
[
"3QueryExplorer",
"效率。面向用户的界面可以提供减轻不良重新",
"配方的机会（?）（请参阅附录列表1）",
"QueryExplorer接口由2个选项卡组成--“查询",
"生成”选项卡和“设置”选项卡。它们都描述",
"3.1.3 Retrieving Documents",
"如下。“查询生成”选项卡显示给最终用户或",
"对于检索，我们使用PyTerrier检索器作为默",
"搜索者和注释者：“设置”选项卡保留给研究",
"认。PyTerrier的体系结构固有地设计为支持对",
"人员5，他们希望通过调查不同的设置来收",
"检索器和重新排序器的操作，以构建端到端检",
"集用于查询生成和IR研究的数据。完整的界",
"索管道，并且一直是信息检索研究人员中检索",
"面是使用HuggingFace的Gradio平台构建的。",
"引擎的流行选择。在QueryExplorer中，研究",
"Gradio（?）是一个开源Python包，用于快速创",
"人员也可以在下面的字典中添加自己的自定义",
"建易于使用、可配置的UI组件，并已被广泛",
"PyTerrier管道（请参阅附录列表2）。这也将",
"用于机器学习模型",
"反映在“设置”选项卡的下拉列表中",
"3.1Searcher's Tab: Query Generation",
"3.1.4",
"Incorporating RelevanceFeedback",
"此查询生成选项卡可作为搜索者和注释者的简",
"(RF)",
"单界面，允许端到端查询生成（QG）-i）临时",
"我们为搜索者提供了通过利用他们选择的检",
"QG：用户可以自已编写搜索查询ii）通过示例",
"索文档来改进当前查询的能力。我们使用零",
"查询：用户可以通过提示HuggingFaceE（?）模",
"拍方法将用户选择的文档合并为E?？的样",
"型生成查询，并选择适当的检索，通过XIRrier",
"式。提示用户选择的文档和查询通过指令E",
"重新生成文档-通过有用的关键字。我们显示",
"Based on the given context information",
"前k个相关文档及其源语言翻译（如果适用）",
"C , generate keywords for the following",
"每个生成的查询都可以单独使用或组合使用",
"query重新生成关键字，其中C是用户选择的",
"以检索文档。用户可以进一步编辑查询，并从",
"文档（请参阅附录列表E3）",
"查询重组器的输出中获得帮助。用户可以进",
"步与检索到的文档或段落进行交互，并为每个",
"3.2",
"PyTerrier and HuggingFace support",
"文档提供相关性注释",
"为了加快IR研究人员进行各种实验的原型制",
"我们现在描述为演示的上述每个设置提供的",
"作过程，QueryExplorer集成了PyTerrierE（？）",
"默认模型。通过对配置设置或代码进行少量修",
"改，可以轻松地用研究人员的选择替换其中的",
"支持。这种集成可以利用QueryExplorer接口",
"中通过PyTerrier创建的检索管道，从而增强其",
"每一个",
"跨不同检索和重新排序算法的注释和定性分析",
"3.1.1Query-By-Example Generation (QBE)",
"功能。在演示过程中，我们展示了该界面使用",
"我们使用Eflan-t5-xxl模型E(?），它是经",
"BM25管道显示搜索结果的功能，突出了根据",
"需要用其他自定义管道替换该界面的灵活性。",
"过指令调整的E(?）版本的文本到文本转换器",
"此功能本质上为PyTerrier检索管道添加了",
"T5E（？）。它已经在大量任务上进行了微调，使",
"层定性分析。此外，为了扩大PyTerrier在处理",
"其方便E（？）用于学习新任务。向用户显示的",
"IR数据集中的实用性，QueryExplorer旨在促",
"默认版本是Eflan-t5-xxl上的一些包装器-它",
"进对这些数据集中的文档进行索引，从而实现",
"接受用户的示例文档或段落，并在EGenerate",
"跨多个基准和数据集的定性实验",
"a query given thefollowing document之前",
"加上一条指令，以及来自MSMarco的3个文",
"认识到HuggingFace的E(？）模型在研究社",
"区中的广泛采用，QueryExplorer利用这些模型",
"档查询对作为样例。",
"进行查询生成，重构和合并反馈。这允许对各",
"Query Reformulation (QR)",
"种大型语言模型的搜索功能进行全面评估",
"虽然我们在设计时考虑到PyTerrier和Hug",
"3.1.2",
"我们使用零拍方法为给定查询生成关键字。",
"flan-t5-xxl被传递指令Improvethesearch",
"gingFace生态系统的受欢迎程度，但使用其他",
"effectiveness by suggesting expansion",
"软件包的数据集和模型也可以在其系统上利用",
"termsforthequery（?）以及原始查询作",
"QueryExplorer接口",
"为输入。零拍查询重构（??？）最近流行扩展查",
"询，通过大型语言模型的零拍提示来提高检索",
"3.3RelevanceAnnotations",
"我们允许对每个文档进行相关性注释，以帮助",
"5we use the term searchers and researchers to differentiate",
"between the higher level goals of the two tabs but both could",
"研究人员通过滑块组件收集相关性注释。注释",
"encompass analysts/annotators/testers, etc.",
"立即保存在单独的JSON文件中"
] |
{
"arxiv_id": "2403.15667",
"language": "zh",
"page_idx": 3
}
|
||
2403.15667-zh-4
|
2403.15667-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>Functions</th>
<th>Actions Annotators) $I.VP$ ann</th>
<th>Contiourahle Settind s (Researchers)</th>
</tr>
<tr>
<td>query generator append\_keywords document retriever send feedback</td>
<td>Provide Example Documents Edit Generated Queries Edit Reformulations to Create Better Query Annotate Relevance of Document to Query Select Documents for Providing Relevance Feedback</td>
<td>Choice of Domain, Example Documents 0-shot/Few-shot QG, Prompt, Exemplars, HF model 0-shot/Few-shot QG, Prompt, Exemplars, HF model Retrieval PyTerrier Pipeline, Index, Documents, 0-shot/Few-shot QG, Prompt, Exemplars, Number of Documents</td>
</tr>
<tr>
<td>send feedback Table 1: The differ $\operatorname{ing}$ the actions (in shc therioh</td>
<td>Select Documents for Providing Relevance Feedback ent functions (on the left) that searchers ar the center). Each of them can be configur ov researcherc</td>
<td>0-shot/Few-shot QG, Prompt, Exemplars, Number of Documents ${\mathrm{d~annotators~can~take~assistance~from~while~perfor}}$ ed through the Settings tab along various paramete</td>
</tr>
</tbody>
</table>
<FigureHere>
Figure 2: The User Interface tab: The user provides an example document related to cricket, uses the query generator to generate multiple queries, selects one of them, and uses the reformulator to further improve the query. In this case, the reformulator has suggested a useful term “cricket”to increase the retrieval effectiveness of the initial query.
|
[
[
0.1376953125,
0.152587890625,
0.19482421875,
0.155029296875,
0.19482421875,
0.166259765625,
0.1376953125,
0.1644287109375
],
[
0.2476806640625,
0.152587890625,
0.43408203125,
0.15380859375,
0.43408203125,
0.1650390625,
0.2476806640625,
0.1644287109375
],
[
0.52490234375,
0.152587890625,
0.71728515625,
0.15380859375,
0.71728515625,
0.1650390625,
0.52490234375,
0.1644287109375
],
[
0.2476806640625,
0.1644287109375,
0.403076171875,
0.1650390625,
0.403076171875,
0.177001953125,
0.2476806640625,
0.17578125
],
[
0.5263671875,
0.1644287109375,
0.73779296875,
0.166259765625,
0.73779296875,
0.177001953125,
0.5263671875,
0.174560546875
],
[
0.139404296875,
0.177001953125,
0.2275390625,
0.17578125,
0.2275390625,
0.18701171875,
0.139404296875,
0.188232421875
],
[
0.2493896484375,
0.177001953125,
0.37451171875,
0.177001953125,
0.37451171875,
0.18701171875,
0.2493896484375,
0.18701171875
],
[
0.52880859375,
0.177001953125,
0.796875,
0.177001953125,
0.796875,
0.18701171875,
0.52880859375,
0.18701171875
],
[
0.246826171875,
0.185302734375,
0.474365234375,
0.18701171875,
0.474365234375,
0.20068359375,
0.246826171875,
0.198974609375
],
[
0.5263671875,
0.185302734375,
0.79833984375,
0.1859130859375,
0.79833984375,
0.1995849609375,
0.5263671875,
0.198974609375
],
[
0.52880859375,
0.19775390625,
0.7734375,
0.198974609375,
0.7734375,
0.210205078125,
0.52880859375,
0.208984375
],
[
0.14111328125,
0.188232421875,
0.235107421875,
0.188232421875,
0.235107421875,
0.1995849609375,
0.14111328125,
0.1995849609375
],
[
0.14111328125,
0.20068359375,
0.240966796875,
0.20068359375,
0.240966796875,
0.2078857421875,
0.14111328125,
0.2078857421875
],
[
0.2493896484375,
0.198974609375,
0.470947265625,
0.198974609375,
0.470947265625,
0.208984375,
0.2493896484375,
0.208984375
],
[
0.14111328125,
0.210205078125,
0.21826171875,
0.210205078125,
0.21826171875,
0.2203369140625,
0.14111328125,
0.2203369140625
],
[
0.2493896484375,
0.2078857421875,
0.861328125,
0.2078857421875,
0.861328125,
0.221435546875,
0.2493896484375,
0.221435546875
],
[
0.1158447265625,
0.232177734375,
0.8818359375,
0.232177734375,
0.8818359375,
0.2452392578125,
0.1158447265625,
0.2452392578125
],
[
0.11419677734375,
0.2452392578125,
0.8818359375,
0.2464599609375,
0.8818359375,
0.26123046875,
0.11419677734375,
0.260009765625
],
[
0.118408203125,
0.26318359375,
0.358642578125,
0.26318359375,
0.358642578125,
0.273681640625,
0.118408203125,
0.273681640625
],
[
0.51318359375,
0.45361328125,
0.5498046875,
0.45361328125,
0.5498046875,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.76171875,
0.52978515625,
0.77099609375,
0.52978515625,
0.77099609375,
0.533203125,
0.76171875,
0.533203125
],
[
0.7958984375,
0.6923828125,
0.8037109375,
0.6923828125,
0.8037109375,
0.6953125,
0.7958984375,
0.6953125
],
[
0.6455078125,
0.75,
0.6826171875,
0.75048828125,
0.6826171875,
0.7548828125,
0.6455078125,
0.75439453125
],
[
0.7060546875,
0.751953125,
0.7314453125,
0.751953125,
0.7314453125,
0.755859375,
0.7060546875,
0.755859375
],
[
0.759765625,
0.75048828125,
0.8134765625,
0.75048828125,
0.8134765625,
0.7548828125,
0.759765625,
0.7548828125
],
[
0.1158447265625,
0.79052734375,
0.880859375,
0.79052734375,
0.880859375,
0.80419921875,
0.1158447265625,
0.80419921875
],
[
0.1158447265625,
0.8046875,
0.880859375,
0.8046875,
0.880859375,
0.818359375,
0.1158447265625,
0.818359375
],
[
0.1175537109375,
0.818359375,
0.880859375,
0.818359375,
0.880859375,
0.83203125,
0.1175537109375,
0.83203125
],
[
0.1175537109375,
0.83447265625,
0.2032470703125,
0.83447265625,
0.2032470703125,
0.845703125,
0.1175537109375,
0.845703125
]
] |
[
"Functions",
"Actions (Searchers and Annotators)",
"[Configurable Settings (Researchers)",
"Provide Example Documents",
"ChoiceofDomain,ExampleDocuments",
"query_generator",
"Edit Generated Queries",
"O-shot/Few-shot QG, Prompt, Exemplars, HF model",
"Edit Reformulations to Create Better Query",
"O-shot/Fewshot QG, Prompt,Exemplars,HF model",
"RetrievalPyTerrierPipeline, Index,Documents.",
"append_keywords",
"document retriever",
"AnnotateRelevanceofDocumenttoQuery",
"send_feedback",
"Select Documents for Providing Relevance FeedbackO-shot/Few-shot QG, Prompt, Exemplars, Number of Documents",
"Table 1: The different functions (on the left) that searchers and annotators can take assistance from while perform-",
"ing the actions (in the center). Each of them can be configured through the Settings tab along various parameters",
"(shown ontheright)byresearchers.",
"earch and Re",
"",
"1",
"",
"201",
"",
"Figure 2: The User Interface tab: The user provides an example document related to cricket, uses the query",
"generator to generate multiple queries, selects one of them, and uses the reformulator to further improve the query.",
"In this case, the reformulator has suggested a useful term “cricket\" to increase the retrieval effectiveness of the",
"initial query."
] |
{
"arxiv_id": "2403.15667",
"language": "zh",
"page_idx": 4
}
|
||
2101.05036-zh-0
|
2101.05036-zh-0.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
估计和评估深度对象检测器中的回归预测不确定性
Steven L. Waslander
Institute for Aerospace Studies
University of Toronto
stevenw@utias.utoronto.ca
Ali Harakeh
Institute for Aerospace Studies
University of Toronto
ali.harakeh@utoronto.ca
# Abstract
Predictive uncertainty estimation is an essential next step for the reliable deployment of deep object detectors in safety-critical tasks. In this work, we focus on estimating predictive distributions for bounding box regression output with variance networks. We show that in the context of object detection, training variance networks with negative log likelihood (NLL) can lead to high entropy predictive distributions regardless of the correctness of the output mean. We propose to use the energy score as a non-local proper scoring rule and find that when used for training, the energy score leads to better calibrated and lower entropy predictive distributions than NLL. We also address the widespread use of non-proper scoring metrics for evaluating predictive distributions from deep object detectors by proposing an alternate evaluation approach founded on proper scoring rules. Using the proposed evaluation tools, we show that although variance networks can be used to produce high quality predictive distributions, ad-hoc approaches used by seminal object detectors for choosing regression targets during training do not provide wide enough data support for reliable variance learning. We hope that our work helps shift evaluation in probabilis tic object detection to better align with predictive uncertainty evaluation in other machine learning domains. Code for all models, evaluation, and datasets is available at: $\bar{\text{https://github.com/asharakeh/probdet.git}}$
## 1 Introduction
深度对象检测器越来越多地被部署为安全关键型机器人和自动化应用中的感知组件。为了可靠和安全的操作，使用探测器作为传感器的后续任务需要与其输出相关的有意义的预测不确定性估计。例如，过度自信的错误预测可能导致计划任务中的非最佳决策，而不自信的正确预测可能导致传感器融合中的信息利用不足。本文研究了概率对象检测器，标准对象检测器的扩展，该扩展同时估计输出类别和边界框的预测分布
本文旨在确定最新趋势所遵循的最新趋势的缺点最先进的概率对象检测器，并为已发现的问题提供理论上可行的解决方案。具体来说，我们观察到大多数最先进的概率对象检测器方法$\mathbb{E}($Feng et al., 2018a; Le et al., 2018; Feng et al., 2018b; He et al., 2019; Kraus & Dietmayer, 2019; Meyer et al., 2019; Choi et al., 2019; Feng et al., 2020; He & Wang, 2020; Harakeh et al., 2020; Lee et al., 2020) 都建立在确定性对象检测后端上，以通过使用方差网络$\boxed{\mathrm{E}}($Detlefsen et al., 2019) 修改此类后端来估计边界框预测分布。然后使用负对数似然(NLL)学习使用方差网络估计的边界框预测分布的均值和方差。在评估其输出预测分布的质量时，这些方法也经常使用不适当的评分规则，例如平均平均精度(mAP)
NLL 的陷阱我们表明，在普通对象检测器使用的标准训练程序下，使用 NLL 作为最小化目标会导致方差网络输出高熵分布，而不管输出边界框的正确性如何。我们通过使用能量得分 (Gneiting&Raftery,2007) 来解决这个问题，这是一种基于能量统计 (Székely&Rizzo, 2013)的距离敏感的适当评分规则，作为训练方差网络的替代方法。我们表明，当使用适当的评分规则进行评估时，使用能量评分学习的预测分布具有较低的熵，更好的校准和更高的质量
评估的陷阱我们通过提供基于完善的适当评分规则E(Gneiting &Raftery, 2007) 的评估工具来解决用于评估概率对象检测器的非适当评分规则的广泛使用，该评估工具仅在估计的
1
|
[
[
0.176513671875,
0.034088134765625,
0.50634765625,
0.034088134765625,
0.50634765625,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.17724609375,
0.10162353515625,
0.7958984375,
0.10162353515625,
0.7958984375,
0.1256103515625,
0.17724609375,
0.1256103515625
],
[
0.1854248046875,
0.154052734375,
0.271240234375,
0.154052734375,
0.271240234375,
0.1654052734375,
0.1854248046875,
0.1654052734375
],
[
0.5087890625,
0.15283203125,
0.65625,
0.15283203125,
0.65625,
0.1666259765625,
0.5087890625,
0.1666259765625
],
[
0.1846923828125,
0.1673583984375,
0.406982421875,
0.1673583984375,
0.406982421875,
0.18115234375,
0.1846923828125,
0.18115234375
],
[
0.50634765625,
0.1666259765625,
0.7294921875,
0.1666259765625,
0.7294921875,
0.179931640625,
0.50634765625,
0.179931640625
],
[
0.1846923828125,
0.179931640625,
0.340576171875,
0.179931640625,
0.340576171875,
0.1932373046875,
0.1846923828125,
0.1932373046875
],
[
0.50732421875,
0.18115234375,
0.6640625,
0.18115234375,
0.6640625,
0.1944580078125,
0.50732421875,
0.1944580078125
],
[
0.1829833984375,
0.1932373046875,
0.359375,
0.1944580078125,
0.359375,
0.208984375,
0.1829833984375,
0.2083740234375
],
[
0.50732421875,
0.1944580078125,
0.701171875,
0.1956787109375,
0.701171875,
0.208984375,
0.50732421875,
0.2083740234375
],
[
0.461669921875,
0.24365234375,
0.53759765625,
0.24560546875,
0.53759765625,
0.261474609375,
0.461669921875,
0.2587890625
],
[
0.233642578125,
0.276611328125,
0.765625,
0.276611328125,
0.765625,
0.290283203125,
0.233642578125,
0.290283203125
],
[
0.2344970703125,
0.292236328125,
0.76123046875,
0.292236328125,
0.76123046875,
0.302978515625,
0.2344970703125,
0.302978515625
],
[
0.233642578125,
0.304931640625,
0.7646484375,
0.304931640625,
0.7646484375,
0.3193359375,
0.233642578125,
0.3193359375
],
[
0.233642578125,
0.318115234375,
0.76318359375,
0.318115234375,
0.76318359375,
0.33203125,
0.233642578125,
0.33203125
],
[
0.233642578125,
0.333251953125,
0.7646484375,
0.333251953125,
0.7646484375,
0.3466796875,
0.233642578125,
0.3466796875
],
[
0.233642578125,
0.3466796875,
0.7646484375,
0.3466796875,
0.7646484375,
0.361083984375,
0.233642578125,
0.361083984375
],
[
0.2344970703125,
0.3623046875,
0.3857421875,
0.3623046875,
0.3857421875,
0.37255859375,
0.2344970703125,
0.37255859375
],
[
0.37744140625,
0.359130859375,
0.765625,
0.35986328125,
0.765625,
0.375,
0.37744140625,
0.373779296875
],
[
0.2344970703125,
0.375,
0.765625,
0.375,
0.765625,
0.38818359375,
0.2344970703125,
0.38818359375
],
[
0.2344970703125,
0.38818359375,
0.7646484375,
0.38818359375,
0.7646484375,
0.40283203125,
0.2344970703125,
0.40283203125
],
[
0.233642578125,
0.401611328125,
0.7646484375,
0.401611328125,
0.7646484375,
0.415283203125,
0.233642578125,
0.415283203125
],
[
0.2320556640625,
0.4140625,
0.765625,
0.415283203125,
0.765625,
0.429931640625,
0.2320556640625,
0.4287109375
],
[
0.2344970703125,
0.429931640625,
0.765625,
0.429931640625,
0.765625,
0.443115234375,
0.2344970703125,
0.443115234375
],
[
0.233642578125,
0.442626953125,
0.765625,
0.44140625,
0.765625,
0.455810546875,
0.233642578125,
0.45703125
],
[
0.2320556640625,
0.455810546875,
0.76318359375,
0.45703125,
0.76318359375,
0.4716796875,
0.2320556640625,
0.47021484375
],
[
0.233642578125,
0.4716796875,
0.765625,
0.4716796875,
0.765625,
0.48486328125,
0.233642578125,
0.48486328125
],
[
0.2320556640625,
0.484130859375,
0.76318359375,
0.484130859375,
0.76318359375,
0.498779296875,
0.2320556640625,
0.498779296875
],
[
0.233642578125,
0.498779296875,
0.76318359375,
0.498779296875,
0.76318359375,
0.51220703125,
0.233642578125,
0.51220703125
],
[
0.233642578125,
0.51318359375,
0.76318359375,
0.51318359375,
0.76318359375,
0.5263671875,
0.233642578125,
0.5263671875
],
[
0.2320556640625,
0.52587890625,
0.7646484375,
0.5263671875,
0.7646484375,
0.54052734375,
0.2320556640625,
0.5390625
],
[
0.2320556640625,
0.5390625,
0.73388671875,
0.5390625,
0.73388671875,
0.55322265625,
0.2320556640625,
0.55322265625
],
[
0.1707763671875,
0.5791015625,
0.316162109375,
0.58203125,
0.316162109375,
0.5966796875,
0.1707763671875,
0.59326171875
],
[
0.176513671875,
0.61279296875,
0.8203125,
0.61279296875,
0.8203125,
0.62353515625,
0.176513671875,
0.62353515625
],
[
0.176513671875,
0.62548828125,
0.82177734375,
0.62548828125,
0.82177734375,
0.6396484375,
0.176513671875,
0.6396484375
],
[
0.1732177734375,
0.63818359375,
0.82177734375,
0.63818359375,
0.82177734375,
0.6513671875,
0.1732177734375,
0.6513671875
],
[
0.17724609375,
0.65478515625,
0.82177734375,
0.65478515625,
0.82177734375,
0.66552734375,
0.17724609375,
0.66552734375
],
[
0.1748046875,
0.66748046875,
0.61669921875,
0.66748046875,
0.61669921875,
0.68115234375,
0.1748046875,
0.68115234375
],
[
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
],
[
0.176513671875,
0.72998046875,
0.82275390625,
0.72998046875,
0.82275390625,
0.744140625,
0.176513671875,
0.744140625
],
[
0.176513671875,
0.744140625,
0.8203125,
0.744140625,
0.8203125,
0.7548828125,
0.176513671875,
0.7548828125
],
[
0.17724609375,
0.75830078125,
0.82177734375,
0.75830078125,
0.82177734375,
0.771484375,
0.17724609375,
0.771484375
],
[
0.17724609375,
0.771484375,
0.82275390625,
0.771484375,
0.82275390625,
0.78515625,
0.17724609375,
0.78515625
],
[
0.17724609375,
0.7861328125,
0.7255859375,
0.7861328125,
0.7255859375,
0.796875,
0.17724609375,
0.796875
],
[
0.1748046875,
0.80615234375,
0.8203125,
0.80615234375,
0.8203125,
0.8193359375,
0.1748046875,
0.8193359375
],
[
0.176513671875,
0.82080078125,
0.82275390625,
0.82080078125,
0.82275390625,
0.833984375,
0.176513671875,
0.833984375
],
[
0.1748046875,
0.83349609375,
0.82177734375,
0.833984375,
0.82177734375,
0.8486328125,
0.1748046875,
0.84765625
],
[
0.1748046875,
0.84765625,
0.82177734375,
0.84765625,
0.82177734375,
0.8623046875,
0.1748046875,
0.8623046875
],
[
0.176513671875,
0.8623046875,
0.82177734375,
0.8623046875,
0.82177734375,
0.87548828125,
0.176513671875,
0.87548828125
],
[
0.176513671875,
0.87744140625,
0.2254638671875,
0.87744140625,
0.2254638671875,
0.8896484375,
0.176513671875,
0.8896484375
],
[
0.17724609375,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.17724609375,
0.9111328125
],
[
0.176513671875,
0.91357421875,
0.82177734375,
0.91357421875,
0.82177734375,
0.923828125,
0.176513671875,
0.923828125
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2021",
"估计和评估深度对象检测器中的回归预测不确定性",
"Ali Harakeh",
"Steven L. Waslander",
"Institute for Aerospace Studies",
"Institute for Aerospace Studies",
"University of Toronto",
"University of Toronto",
"ali.harakeh@utoronto.ca",
"stevenw@utias.utoronto.ca",
"Abstract",
"Predictive uncertainty estimation is an essential next step for the reliable",
"deployment of deep object detectors in safety-critical tasks.Inthis work",
"we focus on estimating predictive distributions for bounding box regression",
"output with variance networks: We show that in the context of object de-",
"tection, training variance networks with negative log likelihood (NLL) can",
"lead to high entropy predictive distributions regardless of the correctness",
"of the output mean.",
".Weproposetouse theenergyscore asa non-local",
"proper scoring rule and find that when used for training, the energy score",
"leads to better calibrated and lower entropy predictive distributions than",
"NLL. We also address the widespread use of non-proper scoring metrics for",
"evaluating predictive distributions from deep object detectors by proposing",
"an alternate evaluation approach founded on proper scoring rules. Using",
"the proposed evaluation tools, we show that although variance networks",
"can be used to produce high quality predictive distributions, ad-hoc ap-",
"proaches used by seminal object detectors for choosing regression targets",
"during training do not provide wide enough data support for reliable vari-",
"ance learning. We hope that our work helps shift evaluation in probabilis-",
"tic object detection to better align with predictive uncertainty evaluation",
"in othermachinelearning domains.Codefor all models,evaluation,and",
"datasets is available at: https://github.com/asharakeh/probdet.git ",
"1Introduction",
"深度对象检测器越来越多地被部署为安全关键型机器人和自动化应用中的感知组件。为了",
"可靠和安全的操作，使用探测器作为传感器的后续任务需要与其输出相关的有意义的预测",
"不确定性估计。例如，过度自信的错误预测可能导致计划任务中的非最佳决策，而不自信的",
"正确预测可能导致传感器融合中的信息利用不足。本文研究了概率对象检测器，标准对象",
"检测器的扩展，该扩展同时估计输出类别和边界框的预测分布",
"本文旨在确定最新趋势所遵循的最新趋势的缺点最先进的概率对象检测器，并为已发现",
"的问题提供理论上可行的解决方案。具体来说，我们观察到大多数最先进的概率对象检测",
"器方法E(Feng et al.,2018a;Le et al.,2018; Feng et al.,2018b;He et al., 2019;Kraus &",
"Dietmayer, 2019; Meyer et al., 2019; Choi et al., 2019; Feng et al., 2020; He & Wang, 2020",
"Harakehetal.，2020:Leeetal.，2020）都建立在确定性对象检测后端上，以通过使用方差",
"网络E(Detlefsenetal.，2019）修改此类后端来估计边界框预测分布。然后使用负对数似然",
"（NLL）学习使用方差网络估计的边界框预测分布的均值和方差。在评估其输出预测分布的",
"质量时，这些方法也经常使用不适当的评分规则，例如平均平均精度（mAP）",
"NLL的陷阱我们表明，在普通对象检测器使用的标准训练程序下，使用NLL作为最小化目",
"标会导致方差网络输出高分布，而不管输出边界框的正确性如何。我们通过使用能量得",
"分（Gneiting&Raftery，2007）来解决这个问题，这是一种基于能量统计（Szekely&Rizzo，",
"2013）的距离敏感的适当评分规则，作为训练方差网络的替代方法。我们表明，当使用适当",
"的评分规则进行评估时，使用能量评分学习的预测分布具有较低的摘，更好的校准和更高",
"的质量",
"评估的陷阱我们通过提供基于完善的适当评分规则E（Gneiting&Raftery，2007）的评估工",
"具来解决用于评估概率对象检测器的非适当评分规则的广泛使用，该评估工具仅在估计的",
"1"
] |
{
"arxiv_id": "2101.05036",
"language": "zh",
"page_idx": 0
}
|
||
2101.05036-zh-1
|
2101.05036-zh-1.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
预测分布等于真实目标分布时才最小化分类和回归。使用建议的工具，我们在分布内，移位和分布外数据上对三种常见对象检测体系结构的概率扩展进行了基准测试。我们的结果表明，方差网络可以使用其预测熵来区分分布内，移位和分布外数据。我们发现，开创性对象检测器在训练期间选择回归目标所使用的 ad-hoc 方法无法为方差网络中的可靠学习提供足够广泛的数据支持。最后，我们在 SEC 中提供了明确的建议。5 ,以避免上述陷阱
# 2 Related Work
用深度神经网络估计预测分布长期以来一直是研究界感兴趣的话题。贝叶斯神经网络(BNN)$\operatorname{E}($MacKay, 1992) 通过在给定预定义先验分布的情况下近似一组网络参数上的后验分布来量化预测不确定性。方差网络E(Nix &Weigend, 1994) 通过单独的神经网络分支估计每个输出的均值和方差来捕获预测不确定性$\boxed{\mathrm{E}}($Kendall &Gal,2017) ,并且通常使用最大似然估计$\operatorname{E}($Detlefsen et al.,2019)进行训练。深度集成(Lakshminarayanan et al. 2017) 从不同的网络初始化中训练方差网络的多个副本，以估计输出样本集的预测分布。MonteCarlo(MC)Dropout (Gal $\&$ Ghahramani, 2016) 基于在测试时激活脱落层生成的输出样本提供预测不确定性估计。我们请读者参考 Detlefsen et al.(2019)的工作，深人比较方差网络、BNN、集成和 MC 在回归任务上的性能。我们发现方差网络是这些方法中最具可扩展性的对象检测任务。最后，我们不像在$\boxed{\mathrm{F}}$Kendall&Gal (2017) 中所做的那样区分偶然不确定性和认知不确定性，而是将重点放在反映两种类型的预测不确定性$\overline{\mathrm{E}}($Detlefsen et al., 2019) 上
最先进的概率对象检测器通过将 Kendall & Gal (2017) 的工作适应于最先进的对象检测器来对预测不确定性进行建模。标准检测器使用方差网络进行扩展，通常称为方差回归头，以及均值边界框回归头，并使用 NLLL$\underline{\mathrm{E}}($Feng et al., 2018a; Le et al., 2018; He et al., 2019; Lee et al., 2020; Feng et al., 2020; He& Wang, 2020) 训练所得网络。某些方法将方差网络与 Dropout$\boxed{\mathrm{E}}($Feng et al., 2018b; Kraus & Dietmayer, 2019) 结合在一起，并在测试时使用蒙特卡洛采样。其他 (Meyer et al., 2019; Choi et al., 2019; Harakeh et al., 2020) 通过修改非最大抑制后处理阶段，利用输出预测方差。这样的修改与论文的范围是正交的。重要的是要注意，现有的概率对象检测器的很大一部分集中在非适当的评分指标上，例如 mAP 和校准误差，以评估其预测分布的质量。最近的方法E(Harakeh et al., 2020; He &Wang, 2020) 使用 Hall et al.(2020) 提出的基于概率的检测质量 (PDQ) 来评估概率对象检测器，这也可以证明是不正确的$\overline{\mathbb{E}}($请参阅附录 D.3 )。相反，我们将$\boxed{\mathbb{E}}$Hoiem et al. (2012) 提出的错误分解与完善的适当评分规则$\overline{\mathrm{E}}($Gneiting &Raftery,2007)相结合，以评估概率对象检测器
3 Learning Bounding Box Distributions With Proper Scoring Rules
$\begin{array}{c}\mathrm{y\in\{1,\ldots,K\}~。给定~N~i~的训练数据集~}\mathcal{D}=\{(\boldsymbol{x}_n,y_n,\boldsymbol{z}_n)\}_{n=1}^N\text{。i。d 来自真实联合条}\\\mathrm{y\in\{1,\ldots,K\}~}_{\text{如物率分析}}\:{\mathrm{A}}_{\text{为一个}}^{*}(\boldsymbol{x}_n,\boldsymbol{x}_n)_{\text{为}}^{*}(\boldsymbol{x}_n,\boldsymbol{x}_n)\}_{n=1}^{N}\text{。}\mathrm{i。d~}\text{来自真实联合条}\end{array}$ 件概率分布$p^*($y$,\boldsymbol z|\boldsymbol{x})=p^*(\boldsymbol{z}|\mathbf{y},\boldsymbol{x})p^*($y$|\boldsymbol{x})$的样本，我们使用具有参数向量$\boldsymbol{\theta}$的神经网络对$p_{\boldsymbol{\theta}}(\boldsymbol{z}|\mathbf{y},\boldsymbol{x})$和$p_{\boldsymbol{\theta}}($y$|\boldsymbol{x})$进行建模。$p_{\boldsymbol{\theta}}(\boldsymbol{z}|\mathbf{y},\boldsymbol{x})$固定为多元高斯分布$\mathcal{N}(\boldsymbol{\mu}(\boldsymbol{x},\boldsymbol{\theta}),\boldsymbol{\Sigma}(\boldsymbol{x},\boldsymbol{\theta}))$ , $p_{\boldsymbol{\theta}}($y$|\boldsymbol{x})$固定为分类分布 Cat$(p_1(\boldsymbol{x},\boldsymbol{\theta}),\ldots,p_K(\boldsymbol{x},\boldsymbol{\theta}))$ 。除非另有说明，否则$z\in\mathbb{R}^4$和表示为$\left(u_{\mathrm{min}},v_{\mathrm{min}},u_{\mathrm{max}},v_{\mathrm{max}}\right)$,其中$\left(u_{\mathrm{min}},v_{\mathrm{min}}\right),\left(u_{\mathrm{max}},v_{\mathrm{max}}\right)$分别是左上方和右下方边界框角的像素坐标。在整个工作中，我们用粗体字符表示随机变量，并且在数据集$\mathcal{D}$中实现的随机变量的相关地面实况实例是斜体的
3.1 Proper Scoring Rules
令$a$为 y 或$z$。评分规则是一个函数$S(p_{\boldsymbol{\theta}},(\boldsymbol{a},\boldsymbol{x}))$ ,在给定实现$a|x\sim p^*(\boldsymbol{a}|x)$的实际事件的情况下，该函数为预测分布$p_{\boldsymbol{\theta}}(a|x)$的质量分配一个数值，其中较低的值表示更好的质量。稍微滥用表示法，让$S(p_{\boldsymbol{\theta}},p^{*})$也引用$S(p_{\boldsymbol{\theta}},.)$的期望值，那么如果$S(p^*,p^*)\leq S(p_{\boldsymbol{\theta}},p^*)$ ,则认为评分规则是正确的，当且仅当$p_{\boldsymbol{\theta}}=p^*$相等，这意味着实际数据生成分布被分配了最低可能的得分值E$( Gneiting \& Raftery, 2007)$。我们在附录 D.2 中提供了适当评分规则的更正式的定义。除了适当的概念外，评分规则还可以进一步分为本地规则和非本地规则。局部评分规则仅根据其在真实目标处的值来评估预测分布，而非局部规则则考虑了预测分布的其他特征。例如，对距离敏感的非局部适当评分规则会奖励将概率质量分配给真实目标附近的预测分布，即使不完全在该目标处。Lakshminarayanan et al.(2017) 注意到使用适当的评分规则作为神经网络损失函数来学习预测分布的效用
2
|
[
[
0.176513671875,
0.034088134765625,
0.50634765625,
0.034088134765625,
0.50634765625,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.17724609375,
0.106689453125,
0.8203125,
0.106689453125,
0.8203125,
0.117431640625,
0.17724609375,
0.117431640625
],
[
0.1748046875,
0.11932373046875,
0.8203125,
0.11932373046875,
0.8203125,
0.1331787109375,
0.1748046875,
0.1331787109375
],
[
0.176513671875,
0.1331787109375,
0.82177734375,
0.1331787109375,
0.82177734375,
0.146484375,
0.176513671875,
0.146484375
],
[
0.17724609375,
0.146484375,
0.82177734375,
0.146484375,
0.82177734375,
0.160400390625,
0.17724609375,
0.160400390625
],
[
0.176513671875,
0.1610107421875,
0.77783203125,
0.1610107421875,
0.77783203125,
0.1717529296875,
0.176513671875,
0.1717529296875
],
[
0.17236328125,
0.2020263671875,
0.325927734375,
0.2020263671875,
0.325927734375,
0.21533203125,
0.17236328125,
0.21533203125
],
[
0.178955078125,
0.238037109375,
0.82177734375,
0.238037109375,
0.82177734375,
0.248779296875,
0.178955078125,
0.248779296875
],
[
0.176513671875,
0.248779296875,
0.82177734375,
0.25,
0.82177734375,
0.264404296875,
0.176513671875,
0.26318359375
],
[
0.176513671875,
0.264404296875,
0.82177734375,
0.264404296875,
0.82177734375,
0.27783203125,
0.176513671875,
0.27783203125
],
[
0.178955078125,
0.27783203125,
0.82275390625,
0.27783203125,
0.82275390625,
0.291748046875,
0.178955078125,
0.291748046875
],
[
0.178955078125,
0.291748046875,
0.8203125,
0.292236328125,
0.8203125,
0.30615234375,
0.178955078125,
0.304931640625
],
[
0.176513671875,
0.30615234375,
0.82177734375,
0.30615234375,
0.82177734375,
0.3193359375,
0.176513671875,
0.3193359375
],
[
0.176513671875,
0.3193359375,
0.82177734375,
0.3193359375,
0.82177734375,
0.333251953125,
0.176513671875,
0.333251953125
],
[
0.178955078125,
0.335205078125,
0.8203125,
0.335205078125,
0.8203125,
0.34521484375,
0.178955078125,
0.34521484375
],
[
0.17724609375,
0.3466796875,
0.82177734375,
0.3466796875,
0.82177734375,
0.35986328125,
0.17724609375,
0.35986328125
],
[
0.17724609375,
0.35986328125,
0.82177734375,
0.35986328125,
0.82177734375,
0.373779296875,
0.17724609375,
0.373779296875
],
[
0.17724609375,
0.375,
0.82275390625,
0.375,
0.82275390625,
0.38818359375,
0.17724609375,
0.38818359375
],
[
0.176513671875,
0.39013671875,
0.283447265625,
0.39013671875,
0.283447265625,
0.40087890625,
0.176513671875,
0.40087890625
],
[
0.176513671875,
0.410400390625,
0.82275390625,
0.410400390625,
0.82275390625,
0.423583984375,
0.176513671875,
0.423583984375
],
[
0.17724609375,
0.425537109375,
0.82177734375,
0.425537109375,
0.82177734375,
0.436279296875,
0.17724609375,
0.436279296875
],
[
0.17724609375,
0.4375,
0.82177734375,
0.4375,
0.82177734375,
0.451904296875,
0.17724609375,
0.451904296875
],
[
0.176513671875,
0.451904296875,
0.82177734375,
0.451904296875,
0.82177734375,
0.46533203125,
0.176513671875,
0.46533203125
],
[
0.17724609375,
0.46533203125,
0.82177734375,
0.46533203125,
0.82177734375,
0.479248046875,
0.17724609375,
0.479248046875
],
[
0.17724609375,
0.48095703125,
0.81884765625,
0.48095703125,
0.81884765625,
0.4912109375,
0.17724609375,
0.4912109375
],
[
0.178955078125,
0.494384765625,
0.8203125,
0.494384765625,
0.8203125,
0.5048828125,
0.178955078125,
0.5048828125
],
[
0.17724609375,
0.505859375,
0.82177734375,
0.505859375,
0.82177734375,
0.51953125,
0.17724609375,
0.51953125
],
[
0.1748046875,
0.52099609375,
0.8203125,
0.52099609375,
0.8203125,
0.5341796875,
0.1748046875,
0.5341796875
],
[
0.176513671875,
0.53271484375,
0.82177734375,
0.5341796875,
0.82177734375,
0.54736328125,
0.176513671875,
0.546875
],
[
0.17724609375,
0.54736328125,
0.82177734375,
0.54736328125,
0.82177734375,
0.56103515625,
0.17724609375,
0.56103515625
],
[
0.176513671875,
0.5625,
0.81640625,
0.5625,
0.81640625,
0.57568359375,
0.176513671875,
0.57568359375
],
[
0.1707763671875,
0.6015625,
0.7587890625,
0.60302734375,
0.7587890625,
0.6171875,
0.1707763671875,
0.6162109375
],
[
0.1748046875,
0.63720703125,
0.82177734375,
0.63623046875,
0.82177734375,
0.65087890625,
0.1748046875,
0.6513671875
],
[
0.1748046875,
0.65087890625,
0.82177734375,
0.6494140625,
0.82177734375,
0.6640625,
0.1748046875,
0.66552734375
],
[
0.1748046875,
0.6640625,
0.81640625,
0.6640625,
0.81640625,
0.6787109375,
0.1748046875,
0.6787109375
],
[
0.176513671875,
0.6796875,
0.82177734375,
0.6796875,
0.82177734375,
0.693359375,
0.176513671875,
0.693359375
],
[
0.1748046875,
0.6923828125,
0.351318359375,
0.69580078125,
0.351318359375,
0.71044921875,
0.1748046875,
0.70703125
],
[
0.343017578125,
0.68994140625,
0.55810546875,
0.693359375,
0.55810546875,
0.71044921875,
0.343017578125,
0.70703125
],
[
0.59228515625,
0.693359375,
0.82177734375,
0.693359375,
0.82177734375,
0.70703125,
0.59228515625,
0.70703125
],
[
0.1748046875,
0.705078125,
0.8203125,
0.705078125,
0.8203125,
0.7197265625,
0.1748046875,
0.7197265625
],
[
0.55224609375,
0.69970703125,
0.58837890625,
0.69775390625,
0.58837890625,
0.705078125,
0.55224609375,
0.70703125
],
[
0.17724609375,
0.72265625,
0.451904296875,
0.72265625,
0.451904296875,
0.73291015625,
0.17724609375,
0.73291015625
],
[
0.1748046875,
0.75830078125,
0.370849609375,
0.75830078125,
0.370849609375,
0.771484375,
0.1748046875,
0.771484375
],
[
0.176513671875,
0.787109375,
0.81884765625,
0.787109375,
0.81884765625,
0.79736328125,
0.176513671875,
0.79736328125
],
[
0.17724609375,
0.7998046875,
0.82177734375,
0.7998046875,
0.82177734375,
0.81298828125,
0.17724609375,
0.81298828125
],
[
0.17724609375,
0.814453125,
0.82275390625,
0.814453125,
0.82275390625,
0.82763671875,
0.17724609375,
0.82763671875
],
[
0.1895751953125,
0.82763671875,
0.82177734375,
0.82763671875,
0.82177734375,
0.84228515625,
0.1895751953125,
0.84228515625
],
[
0.17724609375,
0.84130859375,
0.82177734375,
0.84130859375,
0.82177734375,
0.85498046875,
0.17724609375,
0.85498046875
],
[
0.17724609375,
0.85693359375,
0.8203125,
0.85693359375,
0.8203125,
0.8671875,
0.17724609375,
0.8671875
],
[
0.17724609375,
0.869140625,
0.82177734375,
0.869140625,
0.82177734375,
0.88330078125,
0.17724609375,
0.88330078125
],
[
0.178955078125,
0.88525390625,
0.81884765625,
0.88525390625,
0.81884765625,
0.89599609375,
0.178955078125,
0.89599609375
],
[
0.17724609375,
0.89794921875,
0.82177734375,
0.89794921875,
0.82177734375,
0.9111328125,
0.17724609375,
0.9111328125
],
[
0.176513671875,
0.9111328125,
0.61376953125,
0.9111328125,
0.61376953125,
0.9248046875,
0.176513671875,
0.9248046875
],
[
0.49267578125,
0.94775390625,
0.50732421875,
0.95068359375,
0.50341796875,
0.9619140625,
0.488525390625,
0.958984375
]
] |
[
"Published as a conference paper at ICLR 2021",
"预测分布等于真实目标分布时才最小化分类和回归。使用建议的工具，我们在分布内，移位",
"和分布外数据上对三种常见对象检测体系结构的概率扩展进行了基准测试。我们的结果表",
"明，方差网络可以使用其预测嫡来区分分布内，移位和分布外数据。我们发现，开创性对象",
"检测器在训练期间选择回归目标所使用的ad-hoc方法无法为方差网络中的可靠学习提供足",
"够广泛的数据支持。最后：我们在SEC中提供了明确的建议。5：以避免上述陷阱",
"2Related Work",
"用深度神经网络估计预测分布长期以来一直是研究界感兴趣的话题。贝叶斯神经网络",
"（BNN）E(MacKay，1992）通过在给定预定义先验分布的情况下近似一组网络参数上的后",
"验分布来量化预测不确定性。方差网络E(Nix&Weigend，1994）通过单独的神经网络分",
"支估计每个输出的均值和方差来捕获预测不确定性E(Kendall&Gal，2017），并且通常使",
"用最大似然估计E（Detlefsenetal.，2019）进行训练。深度集成（Lakshminarayananetal.",
"2017）从不同的网络初始化中训练方差网络的多个副本，以估计输出样本集的预测分布。",
"MonteCarlo(MC)Dropout（Gal&Ghahramani，2016）基于在测试时激活脱落层生成的输",
"出样本提供预测不确定性估计。我们请读者参考Detlefsenetal.（2019）的工作，深入比较",
"方差网络、BNN、集成和MC在回归任务上的性能。我们发现方差网络是这些方法中最具",
"可扩展性的对象检测任务。最后，我们不像在EKendall&Gal（2017）中所做的那样区分",
"偶然不确定性和认知不确定性，而是将重点放在反映两种类型的预测不确定性E(Detlefsen",
"etal.,2019)上",
"最先进的概率对象检测器通过将Kendall&Gal（2017）的工作适应于最先进的对象检测器",
"来对预测不确定性进行建模。标准检测器使用方差网络进行扩展，通常称为方差回归头，以",
"及均值边界框回归头，并使用NLLE(Fengetal.,2018a;Leetal.,2018;Heetal.,2019",
"Leeetal.,2020;Fengetal.，2020;He&Wang，2020）训练所得网络。某些方法将方差网络",
"与DropoutE(Fengetal.，2018b；Kraus&Dietmayer，2019）结合在一起，并在测试时使用",
"蒙特卡洛采样。其他（Meyeretal.2019:Choietal.，2019：Harakehetal..2020）通过修改",
"非最大抑制后处理阶段，利用输出预测方差。这样的修改与论文的范围是正交的。重要的是",
"要注意，现有的概率对象检测器的很大一部分集中在非适当的评分指标上，例如mAP和校",
"准误差，以评估其预测分布的质量。最近的方法E(Harakehetal.，2020；He&Wang，2020）",
"使用Halletal.（2020）提出的基于概率的检测质量（PDQ）来评估概率对象检测器，这也",
"可以证明是不正确的E（请参阅附录D.3）。相反，我们将EHoiemetal.（2012)提出的错误",
"分解与完善的适当评分规则E(Gneiting&Raftery，2007）相结合，以评估概率对象检测器",
"3Learning Bounding BoxDistributions With Proper ScoringRules",
"yE[1,...,K}。给定Ni的训练数据集D=(an，Un，zn))N1。i。d来自真实联合条",
"件概率分布p*(y,za）=p*(zy,a)p*(ya）的样本，我们使用具有参数向量的神经网络",
"对 pe(zly,a)和 pe(yla）进行建模。pe(zly,a）固定为多元高斯分布 N(μu(a,の),(a,の))",
"pe(ya）固定为分类分布Cat(pi(a,の)，...,pk(a,の)）。除非另有说明，否则zER和表示",
"为（umin,Umin,Umax,Uma",
"ax），其中（umin，Umin），（umax",
"x）分别是左上方和右下方边界框",
"角的像素坐标。在整个工作中，我们用粗体字符表示随机变量，并且在数据集D中实现的",
"x,Um",
"随机变量的相关地面实况实例是斜体的",
"3.1Proper Scoring Rules",
"今a为y或z。评分规则是一个函数S(pe：(a.)）：在给定实现ac~p*（a）的实际事",
"件的情况下，该函数为预测分布pe(aα）的质量分配一个数值，其中较低的值表示更好的质",
"量。稍微滥用表示法，让S(pep*）也引用S(pe,）的期望值，那么如果S(p*,p*）≤S(po,p*)",
"则认为评分规则是正确的，当且仅当Pe=p*相等，这意味着实际数据生成分布被分配了",
"最低可能的得分值E（Gneiting&Raftery，2007）。我们在附录D.2中提供了适当评分规则",
"的更正式的定义。除了适当的概念外，评分规则还可以进一步分为本地规则和非本地规则",
"局部评分规则仅根据其在真实目标处的值来评估预测分布，而非局部规则则考虑了预测分",
"布的其他特征。例如，对距离敏感的非局部适当评分规则会奖励将概率质量分配给真实自",
"标附近的预测分布，即使不完全在该目标处。Lakshminarayananetal.（2017）注意到使用",
"适当的评分规则作为神经网络损失函数来学习预测分布的效用",
"2"
] |
{
"arxiv_id": "2101.05036",
"language": "zh",
"page_idx": 1
}
|
||
2101.05036-zh-2
|
2101.05036-zh-2.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
<FigureHere>
<FigureHere>
Figure 1: A toy example showing values of NLL (Blue), ES (Green) and DMM (Orange) plotted against the parameter $\sigma$ of an isotropic covariance matrix $\sigma\boldsymbol{I}$, when assigned to low error (Solid Lines) and high error (Dashed Lines) detection outputs from DETR.
分类任务的预测不确定性在最近的文献中得到了广泛的研究 (Ovadia et al., 2019; Ashukha et al., 2020) 。我们发现常用的适当评分规则，例如 NLL 和 Brier 评分E(Brier, 1950) ,对于学习和评估概率对象检测器的分类预测分布是令人满意的。另一方面，回归任务绝大多数依赖于单个适当的评分规则，即负对数似然$\underline{\mathrm{E}}($Kendall &Gal, 2017; Lakshminarayanan et al., 2017; Detlefsen et al., 2019)。NLL 是一个局部评分规则，如果用于评估纯推理问题， 应该是令人满意的 (Bernardo,1979)。此外，选择适当的评分规则不应该是渐进的，$p_{\boldsymbol{\theta}}$的真实参数应该通过在训练期间最小化任何适当的评分规则 (Gneiting &Raftery,2007) 来恢复。不幸的是，对象检测不符合这些理想化的假设，我们将在下一节中展示使用 NLL 学习和评估边界框预测分布的陷阱。我们还探索了能量得分$\mathbb{E}($Gneiting &Raftery,2007),这是一种适当的非局部评分规则，可作为学习和评估多元高斯预测分布的替代方法
3.2 Negative Log Likelihood as a Scoring Rule
对于多元高斯，NLL 可以写为：
$$\mathrm{NLL}=\frac{1}{2N}\sum_{n=1}^{N}(z_{n}-\boldsymbol{\mu}(\boldsymbol{x}_{n},\boldsymbol{\theta}))^{\intercal}\boldsymbol{\Sigma}(\boldsymbol{x}_{n},\boldsymbol{\theta})^{-1}(\boldsymbol{z}_{n}-\boldsymbol{\mu}(\boldsymbol{x}_{n},\boldsymbol{\theta}))+\log\det\boldsymbol{\Sigma}(\boldsymbol{x}_{n},\boldsymbol{\theta}),$$
其中$N$是数据集$\mathcal{D}$的大小。NLL 是唯一正确的评分规则，也是局部的$\mathbb{E}($Bernardo,1979) , 较高的值意味着在真实目标值下预测密度质量较差。当用作最小化损失时，NLL 的第一项鼓励增加预测分布的熵，因为平均估计值与真实目标值不同。对数行列式正则化项具有对比效应，惩罚高熵分布，并防止在所有数据点上无限高不确定性预测的零损失 (Kendall $\&$Gal,2017)。
Machete (2013)已经表明，NLL 更喜欢信息较少的预测密度，即使概率质量集中在可能的结果上，也会惩罚过度自信。我们通过一个简单的玩具示例演示了此属性与对象检测的相关性。来自最先进的对象检测器 DETRE(Carion et al.,2020)的边界框结果，经过训练以在 COCO 验证拆分上实现 42% 的竞争性 mAP,并分配了单个模拟多元高斯概率分布$\mathcal{N}(\boldsymbol{\mu}(\boldsymbol{x}_n,\boldsymbol{\theta}),\sigma\boldsymbol{I})$ ,其中$\sigma\boldsymbol{I}$是$4\times4$各向同性协方差矩阵，其中$\sigma$作为可变参数。我们将检测结果分为 IOU <=0.5的高误差集和 IOU>0.5的低误差集。其中 IOU 被确定为场景中任何地面实况边界框的最大 IOU。我们在图 1 中绘制了高错误和低错误检测实例的 NLL 损失值，其中 NLL 是在$[10^{-2},10^5]$之间的$\sigma$值处估计的。正如预期的那样，对于低错误检测，使 NLL 最小化的方差值比高错误检测低两个数量级。更有趣的是，对于低错误和高错误检测，NLL 的行为都远离其最小值。与较高的熵分布相比，NLL 对较低的熵分布 ($\sigma$的值较小)的惩罚更为严重，这一特性被证明对E4 节中的训练方差网络有害
3
|
[
[
0.176513671875,
0.034088134765625,
0.50634765625,
0.034088134765625,
0.50634765625,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.200927734375,
0.10101318359375,
0.233642578125,
0.12249755859375,
0.2132568359375,
0.141357421875,
0.1805419921875,
0.11932373046875
],
[
0.2091064453125,
0.10418701171875,
0.223876953125,
0.10418701171875,
0.223876953125,
0.111083984375,
0.2091064453125,
0.111083984375
],
[
0.30224609375,
0.106689453125,
0.351318359375,
0.106689453125,
0.351318359375,
0.117431640625,
0.30224609375,
0.117431640625
],
[
0.689453125,
0.106689453125,
0.73779296875,
0.106689453125,
0.73779296875,
0.117431640625,
0.689453125,
0.117431640625
],
[
0.1690673828125,
0.117431640625,
0.233642578125,
0.111083984375,
0.25244140625,
0.2235107421875,
0.188720703125,
0.2298583984375
],
[
0.2900390625,
0.117431640625,
0.351318359375,
0.117431640625,
0.351318359375,
0.1275634765625,
0.2900390625,
0.1275634765625
],
[
0.685546875,
0.117431640625,
0.73779296875,
0.117431640625,
0.73779296875,
0.1275634765625,
0.685546875,
0.1275634765625
],
[
0.6298828125,
0.125,
0.6640625,
0.1287841796875,
0.65673828125,
0.160400390625,
0.62353515625,
0.1558837890625
],
[
0.63134765625,
0.149658203125,
0.64453125,
0.15087890625,
0.64208984375,
0.1817626953125,
0.62841796875,
0.18115234375
],
[
0.64453125,
0.177978515625,
0.65869140625,
0.18310546875,
0.65185546875,
0.1932373046875,
0.63818359375,
0.1881103515625
],
[
0.348876953125,
0.203857421875,
0.369384765625,
0.203857421875,
0.369384765625,
0.216552734375,
0.348876953125,
0.216552734375
],
[
0.775390625,
0.2052001953125,
0.80224609375,
0.2052001953125,
0.80224609375,
0.216552734375,
0.775390625,
0.216552734375
],
[
0.3056640625,
0.205810546875,
0.32275390625,
0.205810546875,
0.32275390625,
0.21533203125,
0.3056640625,
0.21533203125
],
[
0.26318359375,
0.205810546875,
0.27783203125,
0.205810546875,
0.27783203125,
0.2139892578125,
0.26318359375,
0.2139892578125
],
[
0.64794921875,
0.205810546875,
0.6689453125,
0.205810546875,
0.6689453125,
0.2139892578125,
0.64794921875,
0.2139892578125
],
[
0.69384765625,
0.205810546875,
0.71142578125,
0.205810546875,
0.71142578125,
0.2139892578125,
0.69384765625,
0.2139892578125
],
[
0.74169921875,
0.205810546875,
0.7646484375,
0.205810546875,
0.7646484375,
0.2139892578125,
0.74169921875,
0.2139892578125
],
[
0.1871337890625,
0.2139892578125,
0.23779296875,
0.2139892578125,
0.23779296875,
0.30810546875,
0.1871337890625,
0.30810546875
],
[
0.291748046875,
0.2108154296875,
0.2998046875,
0.2171630859375,
0.291748046875,
0.2235107421875,
0.283447265625,
0.2171630859375
],
[
0.72119140625,
0.21533203125,
0.73046875,
0.21533203125,
0.73046875,
0.2197265625,
0.72119140625,
0.2197265625
],
[
0.25732421875,
0.2259521484375,
0.303955078125,
0.2279052734375,
0.30224609375,
0.2392578125,
0.256591796875,
0.2374267578125
],
[
0.2548828125,
0.238037109375,
0.303955078125,
0.238037109375,
0.303955078125,
0.248779296875,
0.2548828125,
0.248779296875
],
[
0.191162109375,
0.292236328125,
0.2303466796875,
0.292236328125,
0.2303466796875,
0.330810546875,
0.191162109375,
0.330810546875
],
[
0.2132568359375,
0.323974609375,
0.240234375,
0.323974609375,
0.240234375,
0.3359375,
0.2132568359375,
0.3359375
],
[
0.25732421875,
0.323974609375,
0.289306640625,
0.32568359375,
0.28759765625,
0.337158203125,
0.256591796875,
0.335205078125
],
[
0.289306640625,
0.32568359375,
0.33251953125,
0.32568359375,
0.33251953125,
0.3359375,
0.289306640625,
0.3359375
],
[
0.348876953125,
0.32568359375,
0.366943359375,
0.32568359375,
0.366943359375,
0.335205078125,
0.348876953125,
0.335205078125
],
[
0.1748046875,
0.352783203125,
0.82177734375,
0.354736328125,
0.82177734375,
0.369384765625,
0.1748046875,
0.367431640625
],
[
0.176513671875,
0.36865234375,
0.82177734375,
0.36865234375,
0.82177734375,
0.3818359375,
0.176513671875,
0.3818359375
],
[
0.176513671875,
0.38330078125,
0.7548828125,
0.38330078125,
0.7548828125,
0.396484375,
0.176513671875,
0.396484375
],
[
0.176513671875,
0.4287109375,
0.82275390625,
0.4287109375,
0.82275390625,
0.442626953125,
0.176513671875,
0.442626953125
],
[
0.1732177734375,
0.442626953125,
0.82177734375,
0.442626953125,
0.82177734375,
0.455810546875,
0.1732177734375,
0.455810546875
],
[
0.178955078125,
0.458251953125,
0.82177734375,
0.458251953125,
0.82177734375,
0.468505859375,
0.178955078125,
0.468505859375
],
[
0.176513671875,
0.47021484375,
0.82275390625,
0.47021484375,
0.82275390625,
0.484130859375,
0.176513671875,
0.484130859375
],
[
0.176513671875,
0.48291015625,
0.8203125,
0.48486328125,
0.8203125,
0.49755859375,
0.176513671875,
0.49560546875
],
[
0.176513671875,
0.49755859375,
0.82275390625,
0.49755859375,
0.82275390625,
0.51220703125,
0.176513671875,
0.51220703125
],
[
0.17724609375,
0.51123046875,
0.82177734375,
0.51123046875,
0.82177734375,
0.5244140625,
0.17724609375,
0.5244140625
],
[
0.17724609375,
0.52587890625,
0.82177734375,
0.52587890625,
0.82177734375,
0.5390625,
0.17724609375,
0.5390625
],
[
0.176513671875,
0.5390625,
0.82275390625,
0.5390625,
0.82275390625,
0.55322265625,
0.176513671875,
0.55322265625
],
[
0.176513671875,
0.5546875,
0.74609375,
0.5546875,
0.74609375,
0.56494140625,
0.176513671875,
0.56494140625
],
[
0.1748046875,
0.58837890625,
0.52294921875,
0.58837890625,
0.52294921875,
0.6015625,
0.1748046875,
0.6015625
],
[
0.176513671875,
0.6162109375,
0.400390625,
0.6162109375,
0.400390625,
0.6298828125,
0.176513671875,
0.6298828125
],
[
0.273681640625,
0.6396484375,
0.8310546875,
0.6494140625,
0.830078125,
0.68505859375,
0.27294921875,
0.6748046875
],
[
0.30810546875,
0.64453125,
0.32177734375,
0.64013671875,
0.325927734375,
0.6494140625,
0.31201171875,
0.65283203125
],
[
0.22216796875,
0.65576171875,
0.26953125,
0.65576171875,
0.26953125,
0.66748046875,
0.22216796875,
0.66748046875
],
[
0.275390625,
0.6640625,
0.316162109375,
0.6640625,
0.316162109375,
0.67529296875,
0.275390625,
0.67529296875
],
[
0.30224609375,
0.67236328125,
0.330078125,
0.67236328125,
0.330078125,
0.6796875,
0.30224609375,
0.6796875
],
[
0.176513671875,
0.69580078125,
0.82177734375,
0.69580078125,
0.82177734375,
0.708984375,
0.176513671875,
0.708984375
],
[
0.191162109375,
0.71044921875,
0.8095703125,
0.71044921875,
0.8095703125,
0.72021484375,
0.191162109375,
0.72021484375
],
[
0.176513671875,
0.7236328125,
0.82177734375,
0.7236328125,
0.82177734375,
0.7373046875,
0.176513671875,
0.7373046875
],
[
0.176513671875,
0.7373046875,
0.82421875,
0.7373046875,
0.82421875,
0.751953125,
0.176513671875,
0.751953125
],
[
0.176513671875,
0.75244140625,
0.281005859375,
0.75244140625,
0.281005859375,
0.76318359375,
0.176513671875,
0.76318359375
],
[
0.1871337890625,
0.77294921875,
0.82275390625,
0.77294921875,
0.82275390625,
0.7861328125,
0.1871337890625,
0.7861328125
],
[
0.176513671875,
0.78515625,
0.82177734375,
0.78515625,
0.82177734375,
0.798828125,
0.176513671875,
0.798828125
],
[
0.176513671875,
0.7998046875,
0.82275390625,
0.7998046875,
0.82275390625,
0.81298828125,
0.176513671875,
0.81298828125
],
[
0.1748046875,
0.8125,
0.82275390625,
0.81298828125,
0.82275390625,
0.82763671875,
0.1748046875,
0.8271484375
],
[
0.1748046875,
0.82763671875,
0.82275390625,
0.82763671875,
0.82275390625,
0.84130859375,
0.1748046875,
0.84130859375
],
[
0.1748046875,
0.84130859375,
0.82177734375,
0.84130859375,
0.82177734375,
0.85498046875,
0.1748046875,
0.85498046875
],
[
0.17724609375,
0.85498046875,
0.82275390625,
0.85498046875,
0.82275390625,
0.869140625,
0.17724609375,
0.869140625
],
[
0.1748046875,
0.8671875,
0.82177734375,
0.86865234375,
0.82177734375,
0.88330078125,
0.1748046875,
0.8818359375
],
[
0.1748046875,
0.8818359375,
0.82177734375,
0.88330078125,
0.82177734375,
0.896484375,
0.1748046875,
0.89599609375
],
[
0.176513671875,
0.89599609375,
0.82275390625,
0.896484375,
0.82275390625,
0.9111328125,
0.176513671875,
0.91015625
],
[
0.176513671875,
0.9111328125,
0.7099609375,
0.9111328125,
0.7099609375,
0.9248046875,
0.176513671875,
0.9248046875
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.9609375,
0.49267578125,
0.9609375
]
] |
[
"Published as a conference paper at ICLR 2021",
"Poi",
"10°",
"1OU>0.5",
"IOU>0.5",
"",
"-IOU<0.5",
"1OU<0.5",
"0",
"",
"10",
"106",
"106",
"103",
"100",
"103",
"100",
"103",
"",
"0",
"",
"1OU>0.5",
"1OU<0.5",
"",
"10-3",
"10%",
"103",
"106",
"Figure l: A toy example showing values of NLL (Blue), ES (Green) and DMM (Orange)",
"plotted against the parameter of an isotropic covariance matrix o , when assigned to low",
"error (Solid Lines) and high error (Dashed Lines) detection outputs from DETR.",
"分类任务的预测不确定性在最近的文献中得到了广泛的研究（Ovadiaetal.，2019；Ashukha",
"etal.，2020）。我们发现常用的适当评分规则，例如NLL和Brier评分E(Brier，1950)，对",
"于学习和评估概率对象检测器的分类预测分布是令人满意的。另一方面，回归任务绝大多",
"数依赖于单个适当的评分规则，即负对数似然E(Kendall&Gal，2017；Lakshminarayanan",
"etal.2017：Detlefsenetal.2019）。NLL是一个局部评分规则，如果用于评估纯推理问题",
"应该是令人满意的（Bernardo，1979）。此外，选择适当的评分规则不应该是渐进的，pe的",
"真实参数应该通过在训练期间最小化任何适当的评分规则（Gneiting&Raftery，2007）来恢",
"复。不幸的是，对象检测不符合这些理想化的假设，我们将在下一节中展示使用NLL学习",
"和评估边界框预测分布的陷阱。我们还探索了能量得分E（Gneiting&Raftery，2007），这",
"是一种适当的非局部评分规则，可作为学习和评估多元高斯预测分布的替代方法",
"3.2Negative Log Likelihood as a Scoring Rule",
"对于多元高斯，NLL可以写为：",
"之(cn -(m,)2(μm,0)-(2n -(μm,0) ogdt (μm,0), (1)",
"",
"NLL :",
"2NZ",
"0-1",
"其中N是数据集D的大小。NLL是唯一正确的评分规则，也是局部的E(Bernardo，1979)",
"较高的值意味着在直实目标值下预测密度质量较差。当用作最小化损失时，NLL的第",
"项鼓励增加预测分布的摘，因为平均估计值与真实目标值不同。对数行列式正则化项具有",
"对比效应，惩罚高熵分布，并防止在所有数据点上无限高不确定性预测的零损失（Kendall",
"&Gal,2017)。",
"Machete（2013）已经表明，NLL更喜欢信息较少的预测密度，即使概率质量集中在可能",
"的结果上，也会惩罚过度自信。我们通过一个简单的玩具示例演示了此属性与对象检测的",
"相关性。来自最先进的对象检测器DETRE(Carionetal.，2020）的边界框结果，经过训练",
"以在COCO验证拆分上实现42%的竞争性mAP，并分配了单个模拟多元高斯概率分布",
"Nμuαn，の),I），其中I是4×4各向同性协方差矩阵，其中α作为可变参数。我们将",
"检测结果分为IOU<=0.5的高误差集和IOU>0.5的低误差集，其中IOU被确定为场景",
"中任何地面实况边界框的最大IOU。我们在图1中绘制了高错误和低错误检测实例的NLL",
"损失值，其中NLL是在[10-2,1051之间的α值处估计的。正如预期的那样，对于低错误检",
"测，使NLL最小化的方差值比高错误检测低两个数量级。更有趣的是，对于低错误和高错",
"误检测，NLL的行为都远离其最小值。与较高的分布相比，NLL对较低的摘分布（α的",
"值较小）的惩罚更为严重，这一特性被证明对E4节中的训练方差网络有害",
"3"
] |
{
"arxiv_id": "2101.05036",
"language": "zh",
"page_idx": 2
}
|
||
2101.05036-zh-3
|
2101.05036-zh-3.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
3.3 The Energy Score (ES)
能量得分是严格正确且非局部的$\mathbb{E}($Gneiting et al.,2008)评分规则，用于评估多变量的概
率预测。能量得分，可以写成：
$$\mathrm{ES}=\dfrac{1}{N}\sum\limits_{n=1}^N\left(\dfrac{1}{M}\sum\limits_{i=1}^M||\boldsymbol{z}_{n,i}-\boldsymbol{z}_n||-\dfrac{1}{2M^2}\sum\limits_{i=1}^M\sum\limits_{j=1}^M||\boldsymbol{z}_{n,i}-\boldsymbol{z}_{n,j}||\right),$$
(2)
其中$z_n$是地面实况边界框，而$z_{n,i}$是 i$^\mathrm{th}$ i。i。d 样品来自$\mathcal{N}(\boldsymbol{\mu}(x_n,\boldsymbol{\theta}),\boldsymbol{\Sigma}(x_n,\boldsymbol{\theta}))$
能量得分来自能量距离$\overline{\mathrm{E}}($Rizzo &Székely, 2016) $^1$, 这是最大平均差异度量$\overline{\mathrm{E}}($Sejdinovic et al., 2013) , 用于测量随机向量分布之间的距离。除了其理论上的吸引力之外，能量得分对于多元高斯分布具有有效的蒙特卡洛近似$\boxed{\mathrm{E}}($Gneiting et al.,2008) ,写为：
1, (3)
$$\mathrm E\mathrm S=\dfrac{1}{N}\sum\limits_{n=1}^N\left(\dfrac{1}{M}\sum\limits_{i=1}^M||z_{n,i}-z_n||-\dfrac{1}{2(M-1)}\sum\limits_{i=1}^{M-1}||z_{n,i}-z_{n,i+1}||\right),$$
销来计算公式$\overline{\mathrm{E}}$3 中的近似值。当使用能量分数作为评估指标时，我们还将 1000 的值用于
$M$
,它只需要从$\mathcal{N}(\boldsymbol{\mu}(x_n,\boldsymbol{\theta}),\boldsymbol{\Sigma}(x_n,\boldsymbol{\theta}))$中提取一组$M$样本，用于小批量中的每个对象实例。为了训练我们的对象检测器，我们发现将$M$设置为等于 1000 可以使我们以很少的计算开能量得分先前已成功用作 DISCONets$\underline{\mathrm{E}}($Bouchacourt et al.,2016)中的优化目标，以训练从后验概率分布输出样本的神经网络，并训练生成对抗网络$\boxed{\mathrm{E}}($Bellemare et al.,2017)。与DISCO 网络不同，我们使用能量得分来学习具有可微分采样的参数分布
由于能量距离是非局部的，因此它倾向于将概率质量放置在地面实况目标值附近的分布，即使不完全位于该目标。回到图 1 中的玩具示例，ES 显示为在与 NLL 相似的值下最小化， 这并不是一个令人惊讶的观察结果，因为两者都是正确的评分规则。与 NLL 不同，ES 对高熵分布的惩罚比低熵分布更严重。在下一节中，我们将观察到使用能量得分作为损失时， 此属性的影响，与 NLL 相比，导致更好的校准，更低的熵预测分布
3.4 Direct Moment Matching (DMM)
我们在学习预测分布的实验中考虑的最终评分规则是直接矩匹配(DMM)$\operatorname{E}(\operatorname{Feng}\operatorname{et}\operatorname{al}.$,
2020):
$$\mathrm{DMM}=\frac{1}{N}\sum_{n=1}^{N}||z_{n}-\boldsymbol{\mu}(\boldsymbol{x}_{n},\boldsymbol{\theta})||_{p}+||\boldsymbol{\Sigma}(\boldsymbol{x}_{n},\boldsymbol{\theta})-(\boldsymbol{z}_{n}-\boldsymbol{\mu}(\boldsymbol{x}_{n},\boldsymbol{\theta}))(\boldsymbol{z}_{n}-\boldsymbol{\mu}(\boldsymbol{x}_{n},\boldsymbol{\theta}))^{\mathsf{T}}||_{p},$$
其中$\|.\|_p$是$p$范数。DMM 以前是由$\overline{\mathrm{E}}$Feng et al.(2020)提出的，作为使用多阶段训练程序校准$\dot{\text{3D 物体检测器的辅助损耗。DMM 将预测分布的均值和协方差矩阵与使用预测值}}$ 和真实目标值获得的样本统计信息进行匹配。DMM 不是一个适当的评分规则，我们用一个例子展示这个属性。如果$(\boldsymbol z_n-\boldsymbol{\mu}(\boldsymbol{x}_n,\boldsymbol{\theta}))$是零向量，则只有当$\Sigma(x_n,\boldsymbol{\theta})$的所有条目都是 0时。DMM 才会最小化。而人管数据生成分布的实际协方差如何。我们使用使出 DMM 训练的方差网络的结果来讨论仅依赖距离敏感的适当评分规则进行评估的陷阱
4 Experiments And Results
对于我们的实验，我们使用方差网络扩展了三种常见的对象检测方法 DETR$\underline{\mathrm{E}}($Carion et al. 2020) , RetinaNet$\boxed{\mathrm{E}}($Lin et al., 2017) 和 FasterRCNN$\boxed{\mathrm{E}}($Ren et al., 2015) , 以估计多元高斯边界框预测分布的参数。我们使用交叉熵来学习分类预测分布，使用 NLL、DMM 或 ES 来学习边界框预测分布，从而得到 9 架构/回归损失组合。所有网络都有开放源代码和完善的超参数，尽可能使用。我们的概率扩展与确定性对应物的原始实现中报告的 mAP 紧密匹配(参见图。13)。式 equation 4 中用于计算 DMM 的$p$的值是根据每个确定性后端用于边界框回归的原始范数损失确定的。对于 FasterRCNN 和 RetinaNet, 我们使用平滑的 L1损失，而对于 DETR, 我们使用 L1 范数。有关模型训练和推理实现的更多详细信息，请参见附录 A
$^{1}$See AppendixE.
|
[
[
0.176513671875,
0.034088134765625,
0.50634765625,
0.034088134765625,
0.50634765625,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.1748046875,
0.1060791015625,
0.38232421875,
0.1060791015625,
0.38232421875,
0.11932373046875,
0.1748046875,
0.11932373046875
],
[
0.1748046875,
0.1300048828125,
0.82275390625,
0.1300048828125,
0.82275390625,
0.143310546875,
0.1748046875,
0.143310546875
],
[
0.17724609375,
0.146484375,
0.393798828125,
0.146484375,
0.393798828125,
0.1572265625,
0.17724609375,
0.1572265625
],
[
0.33251953125,
0.17041015625,
0.358642578125,
0.1685791015625,
0.359375,
0.175537109375,
0.334228515625,
0.177978515625
],
[
0.398681640625,
0.1685791015625,
0.414306640625,
0.1685791015625,
0.414306640625,
0.1767578125,
0.398681640625,
0.1767578125
],
[
0.583984375,
0.1717529296875,
0.609375,
0.1717529296875,
0.609375,
0.1787109375,
0.583984375,
0.1787109375
],
[
0.3798828125,
0.17236328125,
0.7158203125,
0.173583984375,
0.7158203125,
0.205810546875,
0.3798828125,
0.2052001953125
],
[
0.31787109375,
0.173583984375,
0.33251953125,
0.173583984375,
0.33251953125,
0.18310546875,
0.31787109375,
0.18310546875
],
[
0.378173828125,
0.1749267578125,
0.39208984375,
0.177978515625,
0.390625,
0.184326171875,
0.375732421875,
0.179931640625
],
[
0.53515625,
0.175537109375,
0.55615234375,
0.175537109375,
0.55615234375,
0.18310546875,
0.53515625,
0.18310546875
],
[
0.56884765625,
0.173583984375,
0.57861328125,
0.173583984375,
0.57861328125,
0.1767578125,
0.56884765625,
0.1767578125
],
[
0.3349609375,
0.179931640625,
0.35546875,
0.18310546875,
0.35302734375,
0.1900634765625,
0.334228515625,
0.1881103515625
],
[
0.80224609375,
0.1817626953125,
0.82275390625,
0.1817626953125,
0.82275390625,
0.1956787109375,
0.80224609375,
0.1956787109375
],
[
0.366943359375,
0.184326171875,
0.420654296875,
0.1944580078125,
0.416748046875,
0.208984375,
0.36279296875,
0.1988525390625
],
[
0.56787109375,
0.1956787109375,
0.61279296875,
0.1956787109375,
0.61279296875,
0.20703125,
0.56787109375,
0.20703125
],
[
0.1748046875,
0.2183837890625,
0.771484375,
0.2183837890625,
0.771484375,
0.23291015625,
0.1748046875,
0.23291015625
],
[
0.17724609375,
0.24365234375,
0.82177734375,
0.24365234375,
0.82177734375,
0.25390625,
0.17724609375,
0.25390625
],
[
0.176513671875,
0.255615234375,
0.82177734375,
0.255615234375,
0.82177734375,
0.26953125,
0.176513671875,
0.26953125
],
[
0.1748046875,
0.268310546875,
0.72216796875,
0.26953125,
0.72216796875,
0.2841796875,
0.1748046875,
0.28271484375
],
[
0.28759765625,
0.285888671875,
0.351318359375,
0.296630859375,
0.340576171875,
0.335205078125,
0.27685546875,
0.324462890625
],
[
0.316162109375,
0.290283203125,
0.404296875,
0.28857421875,
0.406005859375,
0.31005859375,
0.31787109375,
0.3125
],
[
0.5849609375,
0.29345703125,
0.609375,
0.29345703125,
0.609375,
0.300537109375,
0.5849609375,
0.300537109375
],
[
0.53759765625,
0.296630859375,
0.57666015625,
0.296630859375,
0.57666015625,
0.30810546875,
0.53759765625,
0.30810546875
],
[
0.268798828125,
0.1767578125,
0.35693359375,
0.186279296875,
0.35302734375,
0.2052001953125,
0.2646484375,
0.1956787109375
],
[
0.256591796875,
0.3037109375,
0.294189453125,
0.3037109375,
0.294189453125,
0.31494140625,
0.256591796875,
0.31494140625
],
[
0.339111328125,
0.3017578125,
0.5048828125,
0.290283203125,
0.5087890625,
0.3193359375,
0.343017578125,
0.330810546875
],
[
0.5888671875,
0.304931640625,
0.72119140625,
0.304931640625,
0.72119140625,
0.318115234375,
0.5888671875,
0.318115234375
],
[
0.80224609375,
0.304931640625,
0.82275390625,
0.304931640625,
0.82275390625,
0.318115234375,
0.80224609375,
0.318115234375
],
[
0.5009765625,
0.314453125,
0.51318359375,
0.314453125,
0.51318359375,
0.318115234375,
0.5009765625,
0.318115234375
],
[
0.50732421875,
0.30615234375,
0.611328125,
0.314453125,
0.60888671875,
0.333251953125,
0.5048828125,
0.324462890625
],
[
0.324462890625,
0.3193359375,
0.3544921875,
0.318115234375,
0.3544921875,
0.326904296875,
0.324462890625,
0.32763671875
],
[
0.38232421875,
0.3212890625,
0.40283203125,
0.3212890625,
0.40283203125,
0.330078125,
0.38232421875,
0.330078125
],
[
0.191162109375,
0.337158203125,
0.8203125,
0.337158203125,
0.8203125,
0.35107421875,
0.191162109375,
0.35107421875
],
[
0.17724609375,
0.3515625,
0.82177734375,
0.3515625,
0.82177734375,
0.365478515625,
0.17724609375,
0.365478515625
],
[
0.17724609375,
0.367431640625,
0.8203125,
0.367431640625,
0.8203125,
0.37744140625,
0.17724609375,
0.37744140625
],
[
0.1748046875,
0.380126953125,
0.193603515625,
0.380126953125,
0.193603515625,
0.392578125,
0.1748046875,
0.392578125
],
[
0.176513671875,
0.40087890625,
0.82275390625,
0.40087890625,
0.82275390625,
0.4140625,
0.176513671875,
0.4140625
],
[
0.176513671875,
0.41357421875,
0.82275390625,
0.41357421875,
0.82275390625,
0.427978515625,
0.176513671875,
0.427978515625
],
[
0.1732177734375,
0.4267578125,
0.689453125,
0.427978515625,
0.689453125,
0.442626953125,
0.1732177734375,
0.44140625
],
[
0.17724609375,
0.449462890625,
0.82275390625,
0.449462890625,
0.82275390625,
0.46337890625,
0.17724609375,
0.46337890625
],
[
0.17724609375,
0.46337890625,
0.82177734375,
0.46337890625,
0.82177734375,
0.4765625,
0.17724609375,
0.4765625
],
[
0.176513671875,
0.4765625,
0.82275390625,
0.4765625,
0.82275390625,
0.490478515625,
0.176513671875,
0.490478515625
],
[
0.176513671875,
0.4912109375,
0.82421875,
0.4912109375,
0.82421875,
0.5048828125,
0.176513671875,
0.5048828125
],
[
0.176513671875,
0.5048828125,
0.66015625,
0.5048828125,
0.66015625,
0.51806640625,
0.176513671875,
0.51806640625
],
[
0.17236328125,
0.53515625,
0.45751953125,
0.53515625,
0.45751953125,
0.548828125,
0.17236328125,
0.548828125
],
[
0.176513671875,
0.56103515625,
0.82275390625,
0.56103515625,
0.82275390625,
0.57470703125,
0.176513671875,
0.57470703125
],
[
0.176513671875,
0.5771484375,
0.2303466796875,
0.5771484375,
0.2303466796875,
0.58837890625,
0.176513671875,
0.58837890625
],
[
0.238525390625,
0.5966796875,
0.83251953125,
0.58837890625,
0.833984375,
0.630859375,
0.240234375,
0.6396484375
],
[
0.2607421875,
0.60107421875,
0.281005859375,
0.60693359375,
0.275390625,
0.6162109375,
0.2548828125,
0.60986328125
],
[
0.279296875,
0.5966796875,
0.293212890625,
0.5966796875,
0.293212890625,
0.60595703125,
0.279296875,
0.60595703125
],
[
0.188720703125,
0.60693359375,
0.259033203125,
0.60986328125,
0.25732421875,
0.62353515625,
0.1871337890625,
0.62060546875
],
[
0.275390625,
0.626953125,
0.298095703125,
0.626953125,
0.298095703125,
0.63525390625,
0.275390625,
0.63525390625
],
[
0.191162109375,
0.64306640625,
0.82177734375,
0.64306640625,
0.82177734375,
0.6572265625,
0.191162109375,
0.6572265625
],
[
0.176513671875,
0.65771484375,
0.82177734375,
0.65771484375,
0.82177734375,
0.671875,
0.176513671875,
0.671875
],
[
0.176513671875,
0.671875,
0.82177734375,
0.671875,
0.82177734375,
0.68505859375,
0.176513671875,
0.68505859375
],
[
0.176513671875,
0.68505859375,
0.82421875,
0.68505859375,
0.82421875,
0.69970703125,
0.176513671875,
0.69970703125
],
[
0.176513671875,
0.69873046875,
0.82421875,
0.69873046875,
0.82421875,
0.7119140625,
0.176513671875,
0.7119140625
],
[
0.176513671875,
0.71337890625,
0.69921875,
0.71337890625,
0.69921875,
0.7265625,
0.176513671875,
0.7265625
],
[
0.17236328125,
0.7470703125,
0.426513671875,
0.74755859375,
0.426513671875,
0.76220703125,
0.17236328125,
0.76123046875
],
[
0.176513671875,
0.77783203125,
0.8203125,
0.77783203125,
0.8203125,
0.79150390625,
0.176513671875,
0.79150390625
],
[
0.176513671875,
0.79248046875,
0.81884765625,
0.79248046875,
0.81884765625,
0.80322265625,
0.176513671875,
0.80322265625
],
[
0.1748046875,
0.8046875,
0.82275390625,
0.8037109375,
0.82275390625,
0.818359375,
0.1748046875,
0.8193359375
],
[
0.176513671875,
0.8193359375,
0.82177734375,
0.8193359375,
0.82177734375,
0.83349609375,
0.176513671875,
0.83349609375
],
[
0.17724609375,
0.83544921875,
0.8203125,
0.83544921875,
0.8203125,
0.84521484375,
0.17724609375,
0.84521484375
],
[
0.176513671875,
0.8466796875,
0.82177734375,
0.8466796875,
0.82177734375,
0.85986328125,
0.176513671875,
0.85986328125
],
[
0.1748046875,
0.85986328125,
0.8203125,
0.85986328125,
0.8203125,
0.87353515625,
0.1748046875,
0.87353515625
],
[
0.176513671875,
0.875,
0.82275390625,
0.875,
0.82275390625,
0.88818359375,
0.176513671875,
0.88818359375
],
[
0.17724609375,
0.89013671875,
0.2484130859375,
0.89013671875,
0.2484130859375,
0.90087890625,
0.17724609375,
0.90087890625
],
[
0.1968994140625,
0.9111328125,
0.31787109375,
0.9111328125,
0.31787109375,
0.9248046875,
0.1968994140625,
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 2021",
"3.3The Energy Score (ES)",
"能量得分是严格正确且非局部的E(Gneitingetal.，2008）评分规则，用于评估多变量的概",
"率预测。能量得分，可以写成：",
"N",
"M",
"M",
"2 11m -2 l - 2211 2n ll",
"1",
"",
"1",
"",
"",
"(2)",
"M",
"-11",
"其中zn是地面实况边界框，而zn,是ithi。i。d样品来自N(u(an,の),(an,の)）",
"能量得分来自能量距离E(Rizzo&Szekely2016）1这是最大平均差异度量l（Sejdinovic",
"etal.，2013），用于测量随机向量分布之间的距离。除了其理论上的吸引力之外，能量得分",
"对于多元高斯分布具有有效的蒙特卡洛近似E（Gneitingetal，2008），写为：",
"",
"(1M",
"M-",
"1",
"ES=Z",
"ES=",
"( I2mi - 2nll -",
"I|zn,i - zn,i+1l",
"(3)",
"",
"2(M -1) 1",
"-",
"2=1",
"它只需要从N(μ(an,0),(an,の))中提取一组M样本，用于小批量中的每个对象实例",
"为了训练我们的对象检测器，我们发现将M设置为等于1000可以使我们以很少的计算开",
"销来计算公式F3中的近似值。当使用能量分数作为评估指标时，我们还将1000的值用于",
"M",
"能量得分先前已成功用作DISCONetsE(Bouchacourtetal.，2016）中的优化目标，以训练",
"从后验概率分布输出样本的神经网络，并训练生成对抗网络E(Bellemareetal.，2017）。与",
"DISCO网络不同，我们使用能量得分来学习具有可微分采样的参数分布",
"由于能量距离是非局部的，因此它倾向于将概率质量放置在地面实况目标值附近的分布，即",
"使不完全位于该目标。回到图1中的玩具示例，ES显示为在与NLL相似的值下最小化",
"这并不是一个令人惊诉的观察结果，因为两者都是正确的评分规则。与NLL不同，ES对",
"高摘分布的惩罚比低摘分布更严重。在下一节中，我们将观察到使用能量得分作为损失时，",
"此属性的影响，与NLL相比，导致更好的校准，更低的熵预测分布",
"3.4Direct Moment Matching (DMM)",
"我们在学习预测分布的实验中考虑的最终评分规则是直接矩匹配（DMM)E(Fengetal.,",
"2020):",
")+() )()",
"1",
"N",
"DMM= ",
"n=1",
"其中lp是p范数。DMM以前是由EFengetal.（2020)提出的，作为使用多阶段训练",
"程序校准3D物体检测器的辅助损耗。DMM将预测分布的均值和协方差矩阵与使用预测值",
"和真实目标值获得的样本统计信息进行匹配。DMM不是一个适当的评分规则，我们用一个",
"例子展示这个属性。如果（zn一μ(an,の)）是零向量，则只有当Z(an,の）的所有条目都是0",
"时，DMM才会最小化，而不管数据生成分布的实际协方差如何。我们使用使用DMM训练",
"的方差网络的结果来讨论仅依赖距离敏感的适当评分规则进行评估的陷阱",
"4ExperimentsAndResults",
"对于我们的实验，我们使用方差网络扩展了三种常见的对象检测方法DETRE(Carionetal.",
"2020）RetinaNetF(Linetal.,2017）和FasterRCNNF(Renetal..2015）以估计多元高",
"斯边界框预测分布的参数。我们使用交叉摘来学习分类预测分布，使用NLL、DMM或ES",
"来学习边界框预测分布，从而得到9架构/回归损失组合。所有网络都有开放源代码和完善",
"的超参数，尽可能使用。我们的概率扩展与确定性对应物的原始实现中报告的mAP紧密匹",
"配（参见图。13）。式equation4中用于计算DMM的p的值是根据每个确定性后端用于",
"边界框回归的原始范数损失确定的。对于FasterRCNN和RetinaNet，我们使用平滑的L1",
"损失，而对于DETR，我们使用LI范数。有关模型训练和推理实现的更多详细信息，请参",
"见附录A",
"SeeAppendixE",
"4"
] |
{
"arxiv_id": "2101.05036",
"language": "zh",
"page_idx": 3
}
|
||
2101.05036-zh-4
|
2101.05036-zh-4.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
<FigureHere>
Figure 2: Average over 80 classification categories of NLL, ES, and MSE for bounding box predictive distributions estimates from probabilistic detectors with DETR, RetinaNet, and $\dot{\text{FasterRCNN backends on in-distribution }}($COCO),artificially shifted (C1-C5), and naturally shifted (OpenIm) datasets. Error bars represent the 95% confidence intervals around the mean.
(c)
(a)
(b)
(f)
(e)
(g)
True Positives Loc. Errors False Positives True Positives Loc. Errors False Positives
$\begin{array}{c}{{\frac{\textcircled{5}}{6}8.00}}\\{{\frac{\textcircled{2}}{\textcircled{4}}}}\\{{\frac{\textcircled{4}}{9}}6.00}\end{array}$
$r/4$
$\stackrel{-1}{\textcircled{3}}_{4.00}$ $\sum_{m=1}^{q}z.00$
<FigureHere>
Figure 3: Average over 80 classification categories of NLL and Brier score for classification predictive distributions generated using DETR. Error bars represent the 95% confidence $\bar{\text{intervals around the mean. Similar trends are seen for RetinaNet and FasterRCNN backends}}$ in Figures 10,11.
所有概率对象检测器都在 COCO (Lin et al., 2014) 训练数据分割上训练。对于测试，COCO 验证数据集用作分发内数据。根据最近关于评估预测不确定性估计$\mathbb{E}($Ovadia et al., 2019) 的质量的建议，我们还在偏移的数据分布上测试了我们的概率对象检测器。我们通过应用18 来构建 COCO 验证数据集 (C1, C3 和 C5) 的 3 失真版本，该 18 是由 Hendrycks & Dietterich (2019)在增加的强度级别[1,3,5]下引人的不同图像损坏。为了测试自然数据集偏移，我们使用 OpenImages 数据 (Kuznetsova et al.,2020)创建一个与 COCO 具有相同类别的偏移数据集，以及一个不包含 COCO 中没有的 80 类别的分发数据集。有关这些数据集及其结构的更多详细信息，请参见附录 B
选择三个确定性后端来表示一级(RetinaNet),两级(FasterRCNN)和最近提出的基于集合 (DETR)的对象检测器。此外，DETR $^{2}$、RetinaNet 和 FasterRCNN 模型的实现是在Detectron2 (Wu et al., 2019) 对象检测框架下公开的，超参数经过优化，可以为 COCO 数据集产生最佳的检测结果
$^{2}$https://github.com/facebookresearch/detr/tree/master/d2
5
|
[
[
0.1732177734375,
0.0328369140625,
0.50732421875,
0.032196044921875,
0.50732421875,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.2091064453125,
0.111083984375,
0.2197265625,
0.111083984375,
0.2197265625,
0.12060546875,
0.2091064453125,
0.12060546875
],
[
0.2181396484375,
0.10797119140625,
0.25732421875,
0.10797119140625,
0.25732421875,
0.11871337890625,
0.2181396484375,
0.11871337890625
],
[
0.226318359375,
0.11553955078125,
0.2607421875,
0.11553955078125,
0.2607421875,
0.12249755859375,
0.226318359375,
0.12249755859375
],
[
0.310546875,
0.121826171875,
0.326904296875,
0.121826171875,
0.326904296875,
0.1287841796875,
0.310546875,
0.1287841796875
],
[
0.1829833984375,
0.11297607421875,
0.2017822265625,
0.11297607421875,
0.2017822265625,
0.12060546875,
0.1829833984375,
0.12060546875
],
[
0.181396484375,
0.12371826171875,
0.2034912109375,
0.12371826171875,
0.2034912109375,
0.1383056640625,
0.181396484375,
0.1383056640625
],
[
0.2197265625,
0.12249755859375,
0.253173828125,
0.12249755859375,
0.253173828125,
0.1300048828125,
0.2197265625,
0.1300048828125
],
[
0.178955078125,
0.1307373046875,
0.205078125,
0.1319580078125,
0.2017822265625,
0.17041015625,
0.176513671875,
0.169189453125
],
[
0.53759765625,
0.1331787109375,
0.70166015625,
0.11236572265625,
0.7158203125,
0.173583984375,
0.55126953125,
0.1956787109375
],
[
0.68115234375,
0.1307373046875,
0.7255859375,
0.1307373046875,
0.7255859375,
0.1578369140625,
0.68115234375,
0.1578369140625
],
[
0.705078125,
0.1363525390625,
0.7138671875,
0.1363525390625,
0.7138671875,
0.141357421875,
0.705078125,
0.141357421875
],
[
0.52685546875,
0.1483154296875,
0.5595703125,
0.1427001953125,
0.56396484375,
0.154052734375,
0.53125,
0.160400390625
],
[
0.701171875,
0.15283203125,
0.76953125,
0.1558837890625,
0.7685546875,
0.1717529296875,
0.69921875,
0.1685791015625
],
[
0.5703125,
0.1634521484375,
0.57666015625,
0.1634521484375,
0.57666015625,
0.1666259765625,
0.5703125,
0.1666259765625
],
[
0.51953125,
0.154052734375,
0.5361328125,
0.15087890625,
0.5400390625,
0.160400390625,
0.52392578125,
0.1634521484375
],
[
0.60107421875,
0.1634521484375,
0.619140625,
0.1634521484375,
0.619140625,
0.1666259765625,
0.60107421875,
0.1666259765625
],
[
0.7099609375,
0.1654052734375,
0.79833984375,
0.1717529296875,
0.7958984375,
0.1881103515625,
0.70751953125,
0.1817626953125
],
[
0.59228515625,
0.1685791015625,
0.705078125,
0.1610107421875,
0.70751953125,
0.184326171875,
0.5947265625,
0.1912841796875
],
[
0.2017822265625,
0.169189453125,
0.2303466796875,
0.169189453125,
0.2303466796875,
0.18115234375,
0.2017822265625,
0.18115234375
],
[
0.451904296875,
0.1685791015625,
0.52392578125,
0.173583984375,
0.52294921875,
0.1849365234375,
0.450927734375,
0.179931640625
],
[
0.521484375,
0.169189453125,
0.56396484375,
0.1673583984375,
0.564453125,
0.177978515625,
0.52294921875,
0.179931640625
],
[
0.328369140625,
0.1717529296875,
0.348876953125,
0.1717529296875,
0.348876953125,
0.179931640625,
0.328369140625,
0.179931640625
],
[
0.273681640625,
0.1717529296875,
0.2998046875,
0.1717529296875,
0.2998046875,
0.1787109375,
0.273681640625,
0.1787109375
],
[
0.35693359375,
0.175537109375,
0.400390625,
0.175537109375,
0.400390625,
0.18310546875,
0.35693359375,
0.18310546875
],
[
0.40185546875,
0.17236328125,
0.431396484375,
0.169189453125,
0.43310546875,
0.1767578125,
0.40283203125,
0.179931640625
],
[
0.1895751953125,
0.189453125,
0.19775390625,
0.189453125,
0.19775390625,
0.1932373046875,
0.1895751953125,
0.1932373046875
],
[
0.2484130859375,
0.177978515625,
0.26318359375,
0.177978515625,
0.26318359375,
0.1817626953125,
0.2484130859375,
0.1817626953125
],
[
0.181396484375,
0.20703125,
0.2017822265625,
0.20703125,
0.2017822265625,
0.2215576171875,
0.181396484375,
0.2215576171875
],
[
0.223876953125,
0.203857421875,
0.27294921875,
0.203857421875,
0.27294921875,
0.2139892578125,
0.223876953125,
0.2139892578125
],
[
0.181396484375,
0.2171630859375,
0.200927734375,
0.2171630859375,
0.200927734375,
0.2342529296875,
0.181396484375,
0.2342529296875
],
[
0.181396484375,
0.2310791015625,
0.200927734375,
0.2310791015625,
0.200927734375,
0.2442626953125,
0.181396484375,
0.2442626953125
],
[
0.708984375,
0.24560546875,
0.740234375,
0.24560546875,
0.740234375,
0.25244140625,
0.708984375,
0.25244140625
],
[
0.2034912109375,
0.255615234375,
0.2296142578125,
0.255615234375,
0.2296142578125,
0.26318359375,
0.2034912109375,
0.26318359375
],
[
0.65625,
0.2568359375,
0.68701171875,
0.2568359375,
0.68701171875,
0.261474609375,
0.65625,
0.261474609375
],
[
0.7158203125,
0.25244140625,
0.81396484375,
0.255615234375,
0.81396484375,
0.271484375,
0.7158203125,
0.268310546875
],
[
0.373291015625,
0.260009765625,
0.400390625,
0.260009765625,
0.400390625,
0.267578125,
0.373291015625,
0.267578125
],
[
0.2181396484375,
0.287353515625,
0.25732421875,
0.287353515625,
0.25732421875,
0.29736328125,
0.2181396484375,
0.29736328125
],
[
0.701171875,
0.28857421875,
0.71728515625,
0.28857421875,
0.71728515625,
0.296630859375,
0.701171875,
0.296630859375
],
[
0.310546875,
0.3037109375,
0.31787109375,
0.300537109375,
0.32177734375,
0.30615234375,
0.316162109375,
0.309326171875
],
[
0.438720703125,
0.306884765625,
0.451904296875,
0.306884765625,
0.451904296875,
0.31494140625,
0.438720703125,
0.31494140625
],
[
0.314453125,
0.31494140625,
0.325927734375,
0.31494140625,
0.325927734375,
0.3193359375,
0.314453125,
0.3193359375
],
[
0.69775390625,
0.31640625,
0.71728515625,
0.31640625,
0.71728515625,
0.324462890625,
0.69775390625,
0.324462890625
],
[
0.1854248046875,
0.324462890625,
0.1968994140625,
0.3212890625,
0.19775390625,
0.32568359375,
0.188720703125,
0.32763671875
],
[
0.336669921875,
0.32763671875,
0.361083984375,
0.32763671875,
0.361083984375,
0.33203125,
0.336669921875,
0.33203125
],
[
0.580078125,
0.33203125,
0.59228515625,
0.33203125,
0.59228515625,
0.3359375,
0.580078125,
0.3359375
],
[
0.70361328125,
0.33203125,
0.7294921875,
0.33203125,
0.7294921875,
0.34033203125,
0.70361328125,
0.34033203125
],
[
0.7646484375,
0.32763671875,
0.771484375,
0.32763671875,
0.771484375,
0.330810546875,
0.7646484375,
0.330810546875
],
[
0.19775390625,
0.337158203125,
0.306396484375,
0.335205078125,
0.306396484375,
0.352783203125,
0.19775390625,
0.354736328125
],
[
0.5849609375,
0.337158203125,
0.701171875,
0.335205078125,
0.701171875,
0.352783203125,
0.5849609375,
0.354736328125
],
[
0.7138671875,
0.33837890625,
0.82421875,
0.33837890625,
0.82421875,
0.352783203125,
0.7138671875,
0.352783203125
],
[
0.455810546875,
0.339111328125,
0.52685546875,
0.341552734375,
0.52685546875,
0.3515625,
0.455810546875,
0.349853515625
],
[
0.330810546875,
0.341552734375,
0.346435546875,
0.341552734375,
0.346435546875,
0.34521484375,
0.330810546875,
0.34521484375
],
[
0.34716796875,
0.341552734375,
0.435546875,
0.337158203125,
0.43701171875,
0.35107421875,
0.348876953125,
0.354736328125
],
[
0.525390625,
0.34228515625,
0.54345703125,
0.34228515625,
0.54345703125,
0.3466796875,
0.525390625,
0.3466796875
],
[
0.310546875,
0.359130859375,
0.431396484375,
0.359130859375,
0.431396484375,
0.37255859375,
0.310546875,
0.37255859375
],
[
0.62109375,
0.359130859375,
0.775390625,
0.359130859375,
0.775390625,
0.37255859375,
0.62109375,
0.37255859375
],
[
0.1732177734375,
0.3837890625,
0.82275390625,
0.3837890625,
0.82275390625,
0.397705078125,
0.1732177734375,
0.397705078125
],
[
0.1732177734375,
0.3984375,
0.82421875,
0.397705078125,
0.82421875,
0.412353515625,
0.1732177734375,
0.41357421875
],
[
0.176513671875,
0.4140625,
0.82177734375,
0.4140625,
0.82177734375,
0.4248046875,
0.176513671875,
0.4248046875
],
[
0.1732177734375,
0.4267578125,
0.826171875,
0.4267578125,
0.826171875,
0.43994140625,
0.1732177734375,
0.43994140625
],
[
0.1748046875,
0.442626953125,
0.2197265625,
0.442626953125,
0.2197265625,
0.453857421875,
0.1748046875,
0.453857421875
],
[
0.2156982421875,
0.479736328125,
0.238525390625,
0.479736328125,
0.238525390625,
0.494384765625,
0.2156982421875,
0.494384765625
],
[
0.32275390625,
0.479736328125,
0.34716796875,
0.479736328125,
0.34716796875,
0.49365234375,
0.32275390625,
0.49365234375
],
[
0.431396484375,
0.479736328125,
0.453369140625,
0.479736328125,
0.453369140625,
0.494384765625,
0.431396484375,
0.494384765625
],
[
0.5400390625,
0.48095703125,
0.56396484375,
0.48095703125,
0.56396484375,
0.494384765625,
0.5400390625,
0.494384765625
],
[
0.64892578125,
0.479736328125,
0.6708984375,
0.479736328125,
0.6708984375,
0.494384765625,
0.64892578125,
0.494384765625
],
[
0.75634765625,
0.48095703125,
0.779296875,
0.48095703125,
0.779296875,
0.49560546875,
0.75634765625,
0.49560546875
],
[
0.178955078125,
0.5,
0.281005859375,
0.5,
0.281005859375,
0.51123046875,
0.178955078125,
0.51123046875
],
[
0.294189453125,
0.5,
0.381591796875,
0.5,
0.381591796875,
0.51123046875,
0.294189453125,
0.51123046875
],
[
0.3896484375,
0.498779296875,
0.82421875,
0.5,
0.82421875,
0.5146484375,
0.3896484375,
0.51318359375
],
[
0.1805419921875,
0.51220703125,
0.1968994140625,
0.51220703125,
0.1968994140625,
0.51953125,
0.1805419921875,
0.51953125
],
[
0.28515625,
0.51123046875,
0.3056640625,
0.51123046875,
0.3056640625,
0.51953125,
0.28515625,
0.51953125
],
[
0.39453125,
0.51220703125,
0.414306640625,
0.51220703125,
0.414306640625,
0.52099609375,
0.39453125,
0.52099609375
],
[
0.56640625,
0.5146484375,
0.6005859375,
0.51806640625,
0.59716796875,
0.52880859375,
0.56396484375,
0.5244140625
],
[
0.60693359375,
0.51220703125,
0.6708984375,
0.5146484375,
0.6689453125,
0.5341796875,
0.60546875,
0.5322265625
],
[
0.50634765625,
0.5166015625,
0.52685546875,
0.5166015625,
0.52685546875,
0.5244140625,
0.50634765625,
0.5244140625
],
[
0.240234375,
0.517578125,
0.271240234375,
0.517578125,
0.271240234375,
0.52978515625,
0.240234375,
0.52978515625
],
[
0.310546875,
0.5166015625,
0.34228515625,
0.5166015625,
0.34228515625,
0.52880859375,
0.310546875,
0.52880859375
],
[
0.390625,
0.51806640625,
0.418212890625,
0.51806640625,
0.418212890625,
0.53515625,
0.390625,
0.53515625
],
[
0.455810546875,
0.517578125,
0.488525390625,
0.517578125,
0.488525390625,
0.52880859375,
0.455810546875,
0.52880859375
],
[
0.73046875,
0.51806640625,
0.73876953125,
0.51806640625,
0.73876953125,
0.5234375,
0.73046875,
0.5234375
],
[
0.283447265625,
0.52294921875,
0.31201171875,
0.52294921875,
0.31201171875,
0.5341796875,
0.283447265625,
0.5341796875
],
[
0.17724609375,
0.521484375,
0.200927734375,
0.521484375,
0.200927734375,
0.5341796875,
0.17724609375,
0.5341796875
],
[
0.45751953125,
0.52783203125,
0.484375,
0.52783203125,
0.484375,
0.53515625,
0.45751953125,
0.53515625
],
[
0.80224609375,
0.52294921875,
0.810546875,
0.52294921875,
0.810546875,
0.5263671875,
0.80224609375,
0.5263671875
],
[
0.2442626953125,
0.52880859375,
0.268798828125,
0.52880859375,
0.268798828125,
0.53271484375,
0.2442626953125,
0.53271484375
],
[
0.50341796875,
0.5322265625,
0.525390625,
0.5322265625,
0.525390625,
0.54052734375,
0.50341796875,
0.54052734375
],
[
0.611328125,
0.5322265625,
0.6357421875,
0.5322265625,
0.6357421875,
0.54052734375,
0.611328125,
0.54052734375
],
[
0.6953125,
0.53076171875,
0.70166015625,
0.53076171875,
0.70166015625,
0.5361328125,
0.6953125,
0.5361328125
],
[
0.178955078125,
0.53857421875,
0.1993408203125,
0.53857421875,
0.1993408203125,
0.548828125,
0.178955078125,
0.548828125
],
[
0.39453125,
0.53857421875,
0.4150390625,
0.53857421875,
0.4150390625,
0.548828125,
0.39453125,
0.548828125
],
[
0.72119140625,
0.53271484375,
0.740234375,
0.53271484375,
0.740234375,
0.54052734375,
0.72119140625,
0.54052734375
],
[
0.285888671875,
0.5390625,
0.3056640625,
0.5390625,
0.3056640625,
0.546875,
0.285888671875,
0.546875
],
[
0.6298828125,
0.54248046875,
0.6748046875,
0.53515625,
0.67724609375,
0.546875,
0.63232421875,
0.55322265625
],
[
0.72119140625,
0.54248046875,
0.74658203125,
0.5498046875,
0.74169921875,
0.5625,
0.7158203125,
0.5537109375
],
[
0.611328125,
0.54345703125,
0.6396484375,
0.54345703125,
0.6396484375,
0.55615234375,
0.611328125,
0.55615234375
],
[
0.3056640625,
0.548828125,
0.330078125,
0.548828125,
0.330078125,
0.55322265625,
0.3056640625,
0.55322265625
],
[
0.50732421875,
0.54345703125,
0.515625,
0.55322265625,
0.50732421875,
0.55810546875,
0.5,
0.548828125
],
[
0.178955078125,
0.5498046875,
0.1993408203125,
0.5498046875,
0.1993408203125,
0.56298828125,
0.178955078125,
0.56298828125
],
[
0.39453125,
0.5498046875,
0.408447265625,
0.5498046875,
0.408447265625,
0.56005859375,
0.39453125,
0.56005859375
],
[
0.744140625,
0.5498046875,
0.77783203125,
0.5498046875,
0.77783203125,
0.5673828125,
0.744140625,
0.5673828125
],
[
0.5048828125,
0.56103515625,
0.60302734375,
0.564453125,
0.60107421875,
0.58203125,
0.50341796875,
0.5791015625
],
[
0.2900390625,
0.5517578125,
0.294189453125,
0.556640625,
0.291748046875,
0.5595703125,
0.285888671875,
0.5537109375
],
[
0.61376953125,
0.56103515625,
0.7236328125,
0.564453125,
0.72216796875,
0.58349609375,
0.61279296875,
0.580078125
],
[
0.72119140625,
0.56103515625,
0.82177734375,
0.56494140625,
0.8203125,
0.58349609375,
0.7197265625,
0.5791015625
],
[
0.3896484375,
0.564453125,
0.491943359375,
0.5673828125,
0.491943359375,
0.58203125,
0.3896484375,
0.5791015625
],
[
0.1748046875,
0.5625,
0.259033203125,
0.5673828125,
0.25732421875,
0.58203125,
0.1732177734375,
0.5771484375
],
[
0.285888671875,
0.5673828125,
0.303955078125,
0.568359375,
0.303955078125,
0.57275390625,
0.285888671875,
0.5712890625
],
[
0.32275390625,
0.5693359375,
0.374267578125,
0.5693359375,
0.374267578125,
0.580078125,
0.32275390625,
0.580078125
],
[
0.30810546875,
0.58544921875,
0.365234375,
0.58544921875,
0.365234375,
0.599609375,
0.30810546875,
0.599609375
],
[
0.61376953125,
0.5859375,
0.71337890625,
0.5859375,
0.71337890625,
0.59765625,
0.61376953125,
0.59765625
],
[
0.1748046875,
0.6123046875,
0.82421875,
0.6123046875,
0.82421875,
0.62548828125,
0.1748046875,
0.62548828125
],
[
0.1732177734375,
0.625,
0.82275390625,
0.625,
0.82275390625,
0.6396484375,
0.1732177734375,
0.6396484375
],
[
0.1748046875,
0.6396484375,
0.82421875,
0.6396484375,
0.82421875,
0.65283203125,
0.1748046875,
0.65283203125
],
[
0.17236328125,
0.65380859375,
0.325927734375,
0.6513671875,
0.325927734375,
0.66552734375,
0.17236328125,
0.66748046875
],
[
0.176513671875,
0.70556640625,
0.82421875,
0.70556640625,
0.82421875,
0.7197265625,
0.176513671875,
0.7197265625
],
[
0.176513671875,
0.7197265625,
0.82177734375,
0.7197265625,
0.82177734375,
0.734375,
0.176513671875,
0.734375
],
[
0.176513671875,
0.73291015625,
0.82177734375,
0.73291015625,
0.82177734375,
0.7470703125,
0.176513671875,
0.7470703125
],
[
0.1748046875,
0.7470703125,
0.82275390625,
0.74560546875,
0.82275390625,
0.76025390625,
0.1748046875,
0.76123046875
],
[
0.176513671875,
0.76123046875,
0.82177734375,
0.76123046875,
0.82177734375,
0.7744140625,
0.176513671875,
0.7744140625
],
[
0.176513671875,
0.7744140625,
0.82177734375,
0.7744140625,
0.82177734375,
0.7890625,
0.176513671875,
0.7890625
],
[
0.1748046875,
0.78857421875,
0.82275390625,
0.78857421875,
0.82275390625,
0.80322265625,
0.1748046875,
0.80322265625
],
[
0.1748046875,
0.80322265625,
0.5107421875,
0.80322265625,
0.5107421875,
0.81640625,
0.1748046875,
0.81640625
],
[
0.176513671875,
0.82470703125,
0.82177734375,
0.82470703125,
0.82177734375,
0.83837890625,
0.176513671875,
0.83837890625
],
[
0.176513671875,
0.8369140625,
0.82275390625,
0.8369140625,
0.82275390625,
0.8515625,
0.176513671875,
0.8515625
],
[
0.1748046875,
0.8515625,
0.82421875,
0.8515625,
0.82421875,
0.86572265625,
0.1748046875,
0.86572265625
],
[
0.1748046875,
0.8662109375,
0.35546875,
0.8662109375,
0.35546875,
0.8798828125,
0.1748046875,
0.8798828125
],
[
0.19287109375,
0.9091796875,
0.59228515625,
0.91015625,
0.59228515625,
0.9248046875,
0.19287109375,
0.923828125
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2021",
".",
"DETRNLL",
"DETRDMM",
"060",
"2400",
"300",
"DETRES",
"Suensue",
"",
"",
"2",
"",
"0",
"",
"",
"",
"Coco Cest Dataset",
"coco....m",
"coco",
"coCo ClestDatas",
"Cs Openlm",
"COCO",
"5Open",
"TestDatase",
"Openie",
"50",
"",
"400",
"RetinaNet ES",
"0300",
"200",
"00",
"COCO",
"c5.oen",
"coco Test Datase oper",
"Datase",
"RCNNES",
"250",
"o",
"15",
"U14",
"150",
"0",
"",
"200",
"100",
"-",
"coco Test Batase Openim",
"cocotestBatasesopenm",
"coco Test Datases openlm",
"cocoCletCatasc",
"ncr",
"Test Datascs Openlm",
"5.00",
"(a) True Positives",
"(b) Localization Errors",
"Figure 2: Average over 80 classification categories of NLL, ES, and MSE for bounding box",
"predictive distributions estimates from probabilistic detectors with DETR, RetinaNet, and",
"FasterRCNNbackendsonin-distribution(COCO),artificiallyshifted(C1-C5),andnaturally",
"shifted (OpenIm) datasets. Error bars represent the 95% confidence intervals around the",
"mean.",
"(a)",
"(b)",
"(c)",
"(e)",
"(f)",
"(g)",
"True Positives",
"Loc. Errors",
"FalsePositives True PositivesLoc.Errors_False Positives",
"8.00",
".8.00",
"88.00",
"oem ",
"150",
"1.50",
":CEMU",
"CETL",
"06.00",
":DETR MAL",
".56",
"壹6.00",
"06.00",
"+DETE",
"",
"R",
"8.1.00",
"81.00",
"",
"g4.00",
"g'4.00",
"81.00",
"g4.00",
"",
"0.50",
"10.0.50",
"",
"uO",
"2.00",
"22",
"",
"o.cocofestBatase.omm",
"",
"o.oco eae ataen",
"o.cocofeataseoom",
"o.ococoFesttaetoen",
"10.00cocofestBata",
"900",
"TestDatasetor",
"(d) NLL",
"(h) Brier Score",
"Figure 3: Average over 80 classification categories of NLL and Brier score for classification",
"predictive distributions generated using DETR. Error bars represent the 95% confidence",
"intervals around the mean. Similar trends are seen for RetinaNet and FasterRCNN backends",
"in Figures 10,11.",
"所有概率对象检测器都在COCO（Linetal.2014)训练数据分割上训练。对于测试，COCO",
"验证数据集用作分发内数据。根据最近关于评估预测不确定性估计E(Ovadiaetal.，2019)",
"的质量的建议，我们还在偏移的数据分布上测试了我们的概率对象检测器。我们通过应用",
"18来构建COCO验证数据集（C1，C3和C5）的3失真版本，该18是由Hendrycks&",
"Dietterich（2019）在增加的强度级别[1,3,5]下引人的不同图像损坏。为了测试自然数据集",
"偏移，我们使用OpenImages数据（Kuznetsova etal.，2020）创建一个与COCO具有相同",
"类别的偏移数据集，以及一个不包含COCO中没有的80类别的分发数据集。有关这些数",
"据集及其结构的更多详细信息，请参见附录B",
"选择三个确定性后端来表示一级（RetinaNet），两级（FasterRCNN）和最近提出的基于集",
"合（DETR）的对象检测器。此外，DETR2、RetinaNet和FasterRCNN模型的实现是在",
"Detectron2（Wuetal.，2019）对象检测框架下公开的，超参数经过优化，可以为COCO数",
"据集产生最佳的检测结果",
"2https://github.com/facebookresearch/detr/tree/master/d2",
"5"
] |
{
"arxiv_id": "2101.05036",
"language": "zh",
"page_idx": 4
}
|
||
2205.09991-zh-3
|
2205.09991-zh-3.jpg
|
Title Suppressed Due to Excessive Size
<FigureHere>
Figure 3.(Properties of diffusion planners) (a) Learned long-horizon planning: Diffuser's learned planning procedure does not suffer from the myopic failure modes common to shooting algorithms and is able to plan over long horizons with sparse reward. (b) Temporal compositionality: Even though the model is not Markovian, it generates trajectories via iterated refinements to local consistency. As a result, it exhibits the types of generalization usually associated with Markovian models, with the ability to stitch together snippets of trajectories from the training data to generate novel plan. (c) Variable-length plans: Despite being a trajectory-level model, Diffuser's planning horizon is not determined by its architecture. The horizon can be updated after training by changing the dimensionality of the input noise. (d) Task compositionality: Diffuser can be composed with new reward functions to plan for tasks unseen during training. In all subfigures, $\mathbf{O}$ denotes a starting state and $\Theta$ denotes a goal state.
轨迹$\tau\sim p(\boldsymbol{\tau}\mid\mathcal{O}_{1:T}=1)$的第一个动作，然后在标准的后退范围控制回路中再次开始计划过程。算法 1 中给出了引导规划方法的伪代码
步长$t$的最优性，并带有$p(\mathcal{O}_t=1)=\exp(r(\boldsymbol{s}_t,\boldsymbol{a}_t))$ 。我们可以通过在公式 1 中设置$h(\boldsymbol{\tau})=p(\mathcal{O}_{1:T}\mid\boldsymbol{\tau})$来从一组最佳轨迹中采样：
$$\tilde{p}_{\theta}(\tau)=p(\tau\mid\mathcal{O}_{1:T}=1)\propto p(\tau)p(\mathcal{O}_{1:T}=1\mid\tau).$$
3.3 Goal-Conditioned RL as Inpainting
我们已经将强化学习问题交换为条件采样之一。值得庆幸的是，之前已经有很多关于扩散模型的条件抽样的工作。虽然很难从该分布中进行精确采样，但是当$p(\mathcal{O}_{1:T}\mid\tau^i)$足够平滑时，反向扩散过程过渡可以近似为高斯 (Sohl-Dickstein et al., 2015) :
$$p_{\theta}(\tau^{i-1}\mid\tau^{i},\mathcal{O}_{1:T})\approx\mathcal{N}(\tau^{i-1};\mu+\Sigma g,\Sigma)$$
(3)
其中$\mu,\Sigma$是原始反向过程转换$p_\theta(\tau^{i-1}\mid\tau^{i})$和
一些规划问题更自然地表现为约束满足，而不是奖励最大化。在这些设置中，目标是产生满足一组约束的任何可行轨迹，例如在目标位置终止。吸引由公式$\overline{\mathrm{E}}$2描述的轨迹的二维数组表示，此设置可以转换为修复问题，其中状态和动作约束的作用类似于图像(SohlDickstein et al., 2015) 中观察到的像素。阵列中所有未观察到的位置必须由扩散模型以与观察到的约束一致的方式填充
此任务所需的扰动函数是观测值的狄拉克增量和其他地方的常数。具体来说，如果$c_t$是时间步长$t$的状态约束，则
$$h(\boldsymbol{\tau})=\delta_{\boldsymbol{c}_t}(\boldsymbol{s}_0,\boldsymbol{a}_0,\dots,\boldsymbol{s}_T,\boldsymbol{a}_T)=\begin{cases}+\infty&\text{if}\boldsymbol{c}_t=\boldsymbol{s}_t\\\quad0&\text{otherwise}\end{cases}$$
动作约束的定义是相同的。实际上，这可以通过从未扰动的反向过程$\tau^i-1\sim p_\theta(\boldsymbol{\tau}^{i-1}\mid\boldsymbol{\tau}^i)$采样并在所有扩散时间步长$i\in\{0,1,\ldots,N\}$之后用条件值$c_t$替换采样值来实现
即使是奖励最大化问题也需要通过修复进行调节，因为所有采样轨迹都应在当前状态下开始。此条件由算法 1 中的第 10 行描述
$$\begin{aligned}\text{g}&=\nabla_{\tau}\log p(\mathcal{O}_{1:T}\mid\tau)|_{\tau=\mu}\\&=\sum_{t=0}^{T}\nabla_{\boldsymbol{s}_{t},\boldsymbol{a}_{t}}r(\boldsymbol{s}_{t},\boldsymbol{a}_{t})|_{(\boldsymbol{s}_{t},\boldsymbol{a}_{t})=\mu_{t}}=\nabla\mathcal{J}(\mu).\end{aligned}$$
$\text{鲍䶮絭提供了用于生成类条件图像(Dhariwal }\$ enter ender enter Nichol, 2021) 的分类器引导采样与强化学习问题设置之间的直接转换。我们首先在所有可用轨迹数据的状态和动作上训练扩散模型$p_\theta(\tau)$ 。然后，我们训练一个单独的模型$\mathcal{J}_\phi$来预测轨迹样本$\tau^i$的累积奖励。根据公式 3 ,通过修改反向过程的均值$\mu$ ,使用$\mathcal{J}_\phi$ 的梯度来指导轨迹采样过程。可以在环境中执行采样
|
[
[
0.359375,
0.059356689453125,
0.615234375,
0.059356689453125,
0.615234375,
0.0694580078125,
0.359375,
0.0694580078125
],
[
0.35302734375,
0.09344482421875,
0.369384765625,
0.09344482421875,
0.369384765625,
0.09661865234375,
0.35302734375,
0.09661865234375
],
[
0.1298828125,
0.10101318359375,
0.136474609375,
0.0947265625,
0.143798828125,
0.0997314453125,
0.1356201171875,
0.10479736328125
],
[
0.406005859375,
0.11553955078125,
0.4150390625,
0.11614990234375,
0.414306640625,
0.121826171875,
0.404296875,
0.12060546875
],
[
0.279296875,
0.125,
0.293212890625,
0.125,
0.293212890625,
0.1331787109375,
0.279296875,
0.1331787109375
],
[
0.35302734375,
0.125,
0.373291015625,
0.12060546875,
0.37744140625,
0.1331787109375,
0.35693359375,
0.136962890625
],
[
0.404296875,
0.125,
0.418212890625,
0.125,
0.418212890625,
0.1351318359375,
0.404296875,
0.1351318359375
],
[
0.484375,
0.12249755859375,
0.50341796875,
0.12249755859375,
0.50341796875,
0.1363525390625,
0.484375,
0.1363525390625
],
[
0.53125,
0.1256103515625,
0.54345703125,
0.1256103515625,
0.54345703125,
0.1337890625,
0.53125,
0.1337890625
],
[
0.6337890625,
0.18115234375,
0.6689453125,
0.18115234375,
0.6689453125,
0.1932373046875,
0.6337890625,
0.1932373046875
],
[
0.23779296875,
0.126953125,
0.2467041015625,
0.126953125,
0.2467041015625,
0.1337890625,
0.23779296875,
0.1337890625
],
[
0.2344970703125,
0.18310546875,
0.301513671875,
0.18310546875,
0.301513671875,
0.1944580078125,
0.2344970703125,
0.1944580078125
],
[
0.7587890625,
0.1817626953125,
0.7939453125,
0.1817626953125,
0.7939453125,
0.1944580078125,
0.7587890625,
0.1944580078125
],
[
0.375732421875,
0.2197265625,
0.386474609375,
0.2183837890625,
0.390625,
0.2474365234375,
0.3798828125,
0.248779296875
],
[
0.74169921875,
0.222900390625,
0.7490234375,
0.222900390625,
0.7490234375,
0.2266845703125,
0.74169921875,
0.2266845703125
],
[
0.3857421875,
0.2392578125,
0.390625,
0.240478515625,
0.38818359375,
0.248779296875,
0.38232421875,
0.2474365234375
],
[
0.13232421875,
0.2235107421875,
0.1396484375,
0.2235107421875,
0.1396484375,
0.2291259765625,
0.13232421875,
0.2291259765625
],
[
0.28759765625,
0.255615234375,
0.3056640625,
0.255615234375,
0.3056640625,
0.26318359375,
0.28759765625,
0.26318359375
],
[
0.1846923828125,
0.271484375,
0.191162109375,
0.27587890625,
0.1846923828125,
0.28271484375,
0.17724609375,
0.279052734375
],
[
0.59326171875,
0.30810546875,
0.64453125,
0.30810546875,
0.64453125,
0.318115234375,
0.59326171875,
0.318115234375
],
[
0.736328125,
0.30810546875,
0.7509765625,
0.30810546875,
0.7509765625,
0.31494140625,
0.736328125,
0.31494140625
],
[
0.7939453125,
0.30810546875,
0.82861328125,
0.30810546875,
0.82861328125,
0.32080078125,
0.7939453125,
0.32080078125
],
[
0.0882568359375,
0.330078125,
0.8857421875,
0.330810546875,
0.8857421875,
0.34521484375,
0.0882568359375,
0.3447265625
],
[
0.71728515625,
0.31005859375,
0.73046875,
0.31005859375,
0.73046875,
0.314453125,
0.71728515625,
0.314453125
],
[
0.08905029296875,
0.3447265625,
0.8837890625,
0.3447265625,
0.8837890625,
0.35791015625,
0.08905029296875,
0.35791015625
],
[
0.08905029296875,
0.359130859375,
0.8857421875,
0.359130859375,
0.8857421875,
0.37255859375,
0.08905029296875,
0.37255859375
],
[
0.0906982421875,
0.375,
0.8837890625,
0.375,
0.8837890625,
0.385009765625,
0.0906982421875,
0.385009765625
],
[
0.08905029296875,
0.386474609375,
0.8837890625,
0.386962890625,
0.8837890625,
0.40087890625,
0.08905029296875,
0.399658203125
],
[
0.08905029296875,
0.40087890625,
0.8857421875,
0.40087890625,
0.8857421875,
0.4140625,
0.08905029296875,
0.4140625
],
[
0.0906982421875,
0.416748046875,
0.8837890625,
0.416748046875,
0.8837890625,
0.4267578125,
0.0906982421875,
0.4267578125
],
[
0.08905029296875,
0.4287109375,
0.88623046875,
0.4287109375,
0.88623046875,
0.442626953125,
0.08905029296875,
0.442626953125
],
[
0.0882568359375,
0.442626953125,
0.53173828125,
0.442626953125,
0.53173828125,
0.455810546875,
0.0882568359375,
0.455810546875
],
[
0.0906982421875,
0.481689453125,
0.48046875,
0.481689453125,
0.48046875,
0.496826171875,
0.0906982421875,
0.496826171875
],
[
0.5,
0.481689453125,
0.8857421875,
0.481689453125,
0.8857421875,
0.49560546875,
0.5,
0.49560546875
],
[
0.0906982421875,
0.49755859375,
0.4755859375,
0.49755859375,
0.4755859375,
0.51123046875,
0.0906982421875,
0.51123046875
],
[
0.5,
0.49755859375,
0.88232421875,
0.49755859375,
0.88232421875,
0.51123046875,
0.5,
0.51123046875
],
[
0.08905029296875,
0.51220703125,
0.2607421875,
0.51220703125,
0.2607421875,
0.52587890625,
0.08905029296875,
0.52587890625
],
[
0.5009765625,
0.51318359375,
0.7158203125,
0.51318359375,
0.7158203125,
0.5263671875,
0.5009765625,
0.5263671875
],
[
0.111083984375,
0.54052734375,
0.450927734375,
0.54052734375,
0.450927734375,
0.5537109375,
0.111083984375,
0.5537109375
],
[
0.498291015625,
0.54248046875,
0.79736328125,
0.54443359375,
0.79736328125,
0.55810546875,
0.498291015625,
0.55615234375
],
[
0.5107421875,
0.56640625,
0.8837890625,
0.56640625,
0.8837890625,
0.580078125,
0.5107421875,
0.580078125
],
[
0.0906982421875,
0.57470703125,
0.47216796875,
0.57470703125,
0.47216796875,
0.58837890625,
0.0906982421875,
0.58837890625
],
[
0.0906982421875,
0.5888671875,
0.47216796875,
0.5888671875,
0.47216796875,
0.60302734375,
0.0906982421875,
0.60302734375
],
[
0.50244140625,
0.58349609375,
0.8837890625,
0.58349609375,
0.8837890625,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.5009765625,
0.5966796875,
0.8837890625,
0.5966796875,
0.8837890625,
0.60986328125,
0.5009765625,
0.60986328125
],
[
0.0906982421875,
0.60400390625,
0.47216796875,
0.60400390625,
0.47216796875,
0.6171875,
0.0906982421875,
0.6171875
],
[
0.50244140625,
0.61279296875,
0.8837890625,
0.61279296875,
0.8837890625,
0.626953125,
0.50244140625,
0.626953125
],
[
0.08905029296875,
0.619140625,
0.47216796875,
0.619140625,
0.47216796875,
0.63330078125,
0.08905029296875,
0.63330078125
],
[
0.5009765625,
0.626953125,
0.8837890625,
0.626953125,
0.8837890625,
0.64013671875,
0.5009765625,
0.64013671875
],
[
0.08905029296875,
0.6337890625,
0.35693359375,
0.63330078125,
0.35693359375,
0.64794921875,
0.08905029296875,
0.6484375
],
[
0.50244140625,
0.64306640625,
0.8837890625,
0.64306640625,
0.8837890625,
0.6572265625,
0.50244140625,
0.6572265625
],
[
0.134033203125,
0.6611328125,
0.428955078125,
0.6611328125,
0.428955078125,
0.67529296875,
0.134033203125,
0.67529296875
],
[
0.50341796875,
0.65771484375,
0.875,
0.65771484375,
0.875,
0.671875,
0.50341796875,
0.671875
],
[
0.451904296875,
0.6640625,
0.47216796875,
0.6640625,
0.47216796875,
0.6748046875,
0.451904296875,
0.6748046875
],
[
0.50341796875,
0.6748046875,
0.6005859375,
0.6748046875,
0.6005859375,
0.68603515625,
0.50341796875,
0.68603515625
],
[
0.0906982421875,
0.69580078125,
0.426513671875,
0.69580078125,
0.426513671875,
0.71044921875,
0.0906982421875,
0.71044921875
],
[
0.50341796875,
0.69580078125,
0.8837890625,
0.69580078125,
0.8837890625,
0.708984375,
0.50341796875,
0.708984375
],
[
0.50244140625,
0.7109375,
0.8857421875,
0.7109375,
0.8857421875,
0.724609375,
0.50244140625,
0.724609375
],
[
0.12744140625,
0.72265625,
0.32177734375,
0.7236328125,
0.32177734375,
0.73779296875,
0.12744140625,
0.7373046875
],
[
0.5009765625,
0.72607421875,
0.5703125,
0.72607421875,
0.5703125,
0.74072265625,
0.5009765625,
0.74072265625
],
[
0.1666259765625,
0.744140625,
0.176513671875,
0.744140625,
0.176513671875,
0.75,
0.1666259765625,
0.75
],
[
0.136474609375,
0.751953125,
0.4345703125,
0.75390625,
0.4345703125,
0.7744140625,
0.136474609375,
0.77294921875
],
[
0.75634765625,
0.75244140625,
0.7939453125,
0.75244140625,
0.7939453125,
0.76318359375,
0.75634765625,
0.76318359375
],
[
0.79150390625,
0.75,
0.87353515625,
0.751953125,
0.87353515625,
0.7666015625,
0.79150390625,
0.7646484375
],
[
0.50634765625,
0.7568359375,
0.73046875,
0.75830078125,
0.73046875,
0.77587890625,
0.50634765625,
0.7744140625
],
[
0.158447265625,
0.771484375,
0.1846923828125,
0.771484375,
0.1846923828125,
0.78271484375,
0.158447265625,
0.78271484375
],
[
0.779296875,
0.76953125,
0.875,
0.76953125,
0.875,
0.78076171875,
0.779296875,
0.78076171875
],
[
0.09234619140625,
0.79345703125,
0.476318359375,
0.798828125,
0.476318359375,
0.82080078125,
0.09234619140625,
0.81494140625
],
[
0.50341796875,
0.79541015625,
0.88232421875,
0.79541015625,
0.88232421875,
0.80615234375,
0.50341796875,
0.80615234375
],
[
0.098876953125,
0.806640625,
0.1168212890625,
0.806640625,
0.1168212890625,
0.81005859375,
0.098876953125,
0.81005859375
],
[
0.50244140625,
0.80810546875,
0.88134765625,
0.80810546875,
0.88134765625,
0.82275390625,
0.50244140625,
0.82275390625
],
[
0.0906982421875,
0.81640625,
0.473876953125,
0.81640625,
0.473876953125,
0.830078125,
0.0906982421875,
0.830078125
],
[
0.50244140625,
0.82470703125,
0.8857421875,
0.82470703125,
0.8857421875,
0.83837890625,
0.50244140625,
0.83837890625
],
[
0.0906982421875,
0.83056640625,
0.471435546875,
0.83056640625,
0.471435546875,
0.8447265625,
0.0906982421875,
0.8447265625
],
[
0.50244140625,
0.84033203125,
0.58251953125,
0.84033203125,
0.58251953125,
0.8515625,
0.50244140625,
0.8515625
],
[
0.08905029296875,
0.8466796875,
0.473876953125,
0.8466796875,
0.473876953125,
0.85986328125,
0.08905029296875,
0.85986328125
],
[
0.108642578125,
0.861328125,
0.473876953125,
0.861328125,
0.473876953125,
0.875,
0.108642578125,
0.875
],
[
0.50244140625,
0.861328125,
0.8837890625,
0.861328125,
0.8837890625,
0.875,
0.50244140625,
0.875
],
[
0.0906982421875,
0.876953125,
0.47216796875,
0.876953125,
0.47216796875,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.50341796875,
0.876953125,
0.8837890625,
0.876953125,
0.8837890625,
0.89013671875,
0.50341796875,
0.89013671875
],
[
0.09234619140625,
0.89306640625,
0.471435546875,
0.89306640625,
0.471435546875,
0.90380859375,
0.09234619140625,
0.90380859375
],
[
0.5009765625,
0.892578125,
0.6689453125,
0.892578125,
0.6689453125,
0.90576171875,
0.5009765625,
0.90576171875
]
] |
[
"Title Suppressed Due to ExcessiveSize",
"",
"2",
"A",
"+",
"0",
"?",
"O",
"☆",
"data",
"?",
"denoising",
"plan",
"",
"",
"",
"C",
"O",
"LA",
"reward",
"+",
"plan",
"Figure 3. (Properties of diffusion planners) (a) Learned long-horizon planning: Diffuser 's learned planning procedure",
"",
"does not suffer from the myopic failure modes common to shooting algorithms and is able to plan over long horizons",
"with sparse reward. (b) Temporal compositionality: Even though the model is not Markovian, it generates trajectories",
"via iterated refinements to local consistency.As a result,it exhibits the types of generalization usually associated with",
"Markovian models, with the ability to stitch together snippets of trajectories from the training data to generate novel",
"plan. (c) Variable-length plans: Despite being a trajectory-level model, Diffuser's planning horizon is not determined by",
"its architecture.The horizon can be updated aftertraining bychanging the dimensionality of the input noise.(d)Task",
"compositionality: Diffuser can be composed with new reward functions to plan for tasks unseen during training. In all",
"subfigures, O denotes a starting state and @ denotes a goal state.",
"步长t的最优性，并带有p(Ot=1)=exp(r(st,at))。",
"轨迹T～P(T1O1:T=1)的第一个动作，然后在标准",
"我们可以通过在公式 1中设置h(-)=p(O1:T|）来",
"的后退范围控制回路中再次开始计划过程。算法1中",
"从一组最佳轨迹中采样：",
"给出了引导规划方法的伪代码",
"pe(T) =p( | O1:T = 1) α p(T)p(O1:T = 1 [ T).",
"3.3 Goal-Conditioned RL as Inpainting",
"一些规划问题更自然地表现为约束满足，而不是奖励",
"我们已经将强化学习问题交换为条件采样之一。值得",
"庆幸的是，之前已经有很多关于扩散模型的条件抽样",
"最大化。在这些设置中，目标是产生满足一组约束的",
"任何可行轨迹，例如在目标位置终止。吸引由公式E2",
"的工作。虽然很难从该分布中进行精确采样，但是当",
"描述的轨迹的二维数组表示，此设置可以转换为修复",
"p(O1:T「-\"）足够平滑时，反向扩散过程过渡可以近似",
"问题，其中状态和动作约束的作用类似于图像（Sohl",
"为高斯（Sohl-Dickstein et al.，2015）：",
"Dicksteinetal.,2015）中观察到的像素。阵列中所有",
"pe(-i-1 [T, O1:T) ~ N(-i-l;μ+Eg,)",
"未观察到的位置必须由扩散模型以与观察到的约束一",
"(3)",
"致的方式填充",
"其中μ,是原始反向过程转换pe(-i-1|）和",
"此任务所需的扰动函数是观测值的狄拉克增量和其他",
"地方的常数。具体来说，如果ct是时间步长t的状态",
"g =V-logp(O1:T T)IT=μ",
"约束，则",
"T",
"=Vst,ar(st, at)l(st,ai)=μt = VJ(μ),.",
"+8",
"if ct=St",
"h(T) = Se(80,a0,.., ST,aT) ",
"t=0",
"0otherwise",
"跑奏繁提供了用于生成类条件图像（Dhariwal&",
"动作约束的定义是相同的。实际上，这可以通过从未",
"",
"扰动的反向过程i-1～pe(i-1|\"）采样并在所有扩",
"Nichol,2021）的分类器引导采样与强化学习问题设",
"散时间步长iE[0,1,...,N】之后用条件值ct替换采",
"置之间的直接转换。我们首先在所有可用轨迹数据的",
"样值来实现",
"状态和动作上训练扩散模型pe(）。然后，我们训练",
"个单独的模型J来预测轨迹样本-的累积奖励。",
"即使是奖励最大化问题也需要通过修复进行调节，因",
"根据公式3，通过修改反向过程的均值μ，使用",
"为所有采样轨迹都应在当前状态下开始。此条件由算",
"的梯度来指导轨迹采样过程。可以在环境中执行采样",
"法1中的第10行描述"
] |
{
"arxiv_id": "2205.09991",
"language": "zh",
"page_idx": 3
}
|
||
2208.10906-zh-0
|
2208.10906-zh-0.jpg
|
DualSmoke: Sketch-Based Smoke Illustration Design with Two-Stage
# Generative Model
Haoran Xie$^*$
JAIST
Keisuke Arihara
JAIST
Syuhei Sato
Hosei University
Kazunori Miyata
JAIST
Abstract
The dynamic effects of smoke are impressive in illustration design, but it is a troublesome and challenging issue for common users to design the smoke effect without domain knowledge of fluid simulations. In this work, we propose DualSmoke , two stage global-to-local generation framework for the interactive smoke illustration design. For the global stage, the proposed approach utilizes fluid patterns to generate Lagrangian coherent structure from the user's hand-drawn sketches. For the local stage, the detailed flow patterns are obtained from the generated coherent structure. Finally, we apply the guiding force feld to the smoke simulator to design the desired smoke illustration. To construct the training dataset, DualSmoke generates flow patterns using the fnite-time Lyapunov exponents of the velocity fields. The synthetic sketch data is generated from the flow patterns by skeleton extraction. From our user study, it is verified that the proposed design interface can provide various smoke illustration designs with good user usability. Our code is available at: https://github.com/shasph/DualSmoke
## 1. Introduction
Nowadays, due to the rapid prevalence of social media and personal computing devices, it becomes an emerging and growing topic to help amateur users achieve high-quality content creation. The dynamic effects of 2D illustration play an important role to increase the attraction of various media such as e-books, education materials, and art design. Especially, the design of complex flow design is difficult for common users without the specific domain knowledge and design skills, which are required in commercial design tools, such as PhotoShop and Blender. Previous works
$^*$Japan Advanced Institute of Science and Technology, 1-1
Nomi, Ishikawa, 9231292, Japan. E-mail: xie@jaist.ac.jp
<FigureHere>
Figure 1. The proposed DualSmoke design framework utilizes fuid patterns to generate velocity field in the generative model
adopted kinetic textures [23] and elemental dynamics [43] to author the dynamical effects, and sketchbased approaches to design fluid motions [18,46]. However, these approaches may fail to provide the turbulent details in flow motions, especially for smoke design. To solve this issue, we aim to provide an efficient design tool for smoke illustration design.
用于流体控制的常规方法主要集中在形状约束[29] 上。然而 ,由于非线性流，具有专业知识的用户需要耗时的过程和繁琐的参数调整。对于基于关键帧的方法 [40] ,设计了优化过程来探索适当的外力，通常计算量很大。此外，对于烟雾模拟，以交互方式调整参数和边界条件是不实际的。To efficiently design global motions of the fluid, low-resolution simulations were used as input for controlling high-resolution simulation [27,32]. Lowresolution simulations can be executed at low costs, so users can easily repeat the simulation multiple times to design the motions. However, designing operations of fluid motions are not still intuitive, because users need to adjust simulation parameters. For more intuitive design, paths or sketches were introduced as inputs to generate the fluid flows [24,25] . We follow this idea ${\mathrm{and~our~smoke~design~system~is~based~on~sketching~op-}}$ erations.
In this work, we propose DualSmoke, a two-stage generation framework with conditional generative adversarial networks (cGAN) for smoke illustration design to meet the users' design intention ,如图$\color{#c00}{\mathrm{E}}$1所示。For the phase of data training, 所提出的框架利
1
|
[
[
0.13232421875,
0.1307373046875,
0.83837890625,
0.1331787109375,
0.83837890625,
0.15087890625,
0.13232421875,
0.1483154296875
],
[
0.39794921875,
0.1572265625,
0.576171875,
0.1572265625,
0.576171875,
0.1717529296875,
0.39794921875,
0.1717529296875
],
[
0.15283203125,
0.20263671875,
0.259033203125,
0.20263671875,
0.259033203125,
0.2203369140625,
0.15283203125,
0.2203369140625
],
[
0.2998046875,
0.2052001953125,
0.43701171875,
0.2052001953125,
0.43701171875,
0.2197265625,
0.2998046875,
0.2197265625
],
[
0.50341796875,
0.2052001953125,
0.60888671875,
0.2052001953125,
0.60888671875,
0.2197265625,
0.50341796875,
0.2197265625
],
[
0.67333984375,
0.2052001953125,
0.81640625,
0.2052001953125,
0.81640625,
0.2197265625,
0.67333984375,
0.2197265625
],
[
0.1732177734375,
0.2215576171875,
0.2320556640625,
0.2215576171875,
0.2320556640625,
0.2374267578125,
0.1732177734375,
0.2374267578125
],
[
0.339111328125,
0.222900390625,
0.400390625,
0.222900390625,
0.400390625,
0.2374267578125,
0.339111328125,
0.2374267578125
],
[
0.483642578125,
0.2215576171875,
0.62353515625,
0.2235107421875,
0.62353515625,
0.238037109375,
0.483642578125,
0.236083984375
],
[
0.71142578125,
0.222900390625,
0.771484375,
0.222900390625,
0.771484375,
0.2374267578125,
0.71142578125,
0.2374267578125
],
[
0.236083984375,
0.272705078125,
0.314453125,
0.272705078125,
0.314453125,
0.290283203125,
0.236083984375,
0.290283203125
],
[
0.8056640625,
0.27587890625,
0.8505859375,
0.27587890625,
0.8505859375,
0.28271484375,
0.8056640625,
0.28271484375
],
[
0.098876953125,
0.30615234375,
0.465576171875,
0.30615234375,
0.465576171875,
0.3193359375,
0.098876953125,
0.3193359375
],
[
0.080078125,
0.322509765625,
0.468017578125,
0.322509765625,
0.468017578125,
0.3359375,
0.080078125,
0.3359375
],
[
0.080078125,
0.33837890625,
0.4697265625,
0.33837890625,
0.4697265625,
0.3515625,
0.080078125,
0.3515625
],
[
0.5810546875,
0.34033203125,
0.60302734375,
0.34033203125,
0.60302734375,
0.3447265625,
0.5810546875,
0.3447265625
],
[
0.68896484375,
0.343505859375,
0.705078125,
0.343505859375,
0.705078125,
0.347900390625,
0.68896484375,
0.347900390625
],
[
0.08087158203125,
0.352783203125,
0.468017578125,
0.352783203125,
0.468017578125,
0.3662109375,
0.08087158203125,
0.3662109375
],
[
0.09234619140625,
0.3662109375,
0.4697265625,
0.367431640625,
0.4697265625,
0.3818359375,
0.09234619140625,
0.380615234375
],
[
0.5185546875,
0.3662109375,
0.57421875,
0.3662109375,
0.57421875,
0.37744140625,
0.5185546875,
0.37744140625
],
[
0.611328125,
0.3662109375,
0.6953125,
0.367431640625,
0.6953125,
0.37890625,
0.611328125,
0.37744140625
],
[
0.71826171875,
0.367431640625,
0.77685546875,
0.367431640625,
0.77685546875,
0.37744140625,
0.71826171875,
0.37744140625
],
[
0.8095703125,
0.371826171875,
0.8310546875,
0.371826171875,
0.8310546875,
0.375,
0.8095703125,
0.375
],
[
0.08087158203125,
0.38330078125,
0.468017578125,
0.38330078125,
0.468017578125,
0.396484375,
0.08087158203125,
0.396484375
],
[
0.07843017578125,
0.396484375,
0.463134765625,
0.397705078125,
0.463134765625,
0.412353515625,
0.07843017578125,
0.410888671875
],
[
0.50244140625,
0.391357421875,
0.890625,
0.391357421875,
0.890625,
0.40478515625,
0.50244140625,
0.40478515625
],
[
0.5009765625,
0.404052734375,
0.890625,
0.40478515625,
0.890625,
0.419189453125,
0.5009765625,
0.41845703125
],
[
0.080078125,
0.41357421875,
0.468017578125,
0.41357421875,
0.468017578125,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50341796875,
0.42041015625,
0.57861328125,
0.42041015625,
0.57861328125,
0.431884765625,
0.50341796875,
0.431884765625
],
[
0.08087158203125,
0.429931640625,
0.468017578125,
0.429931640625,
0.468017578125,
0.43994140625,
0.08087158203125,
0.43994140625
],
[
0.080078125,
0.443115234375,
0.46728515625,
0.443115234375,
0.46728515625,
0.45703125,
0.080078125,
0.45703125
],
[
0.080078125,
0.458984375,
0.4697265625,
0.458984375,
0.4697265625,
0.472900390625,
0.080078125,
0.472900390625
],
[
0.50341796875,
0.462158203125,
0.890625,
0.462158203125,
0.890625,
0.47607421875,
0.50341796875,
0.47607421875
],
[
0.080078125,
0.473388671875,
0.4697265625,
0.473388671875,
0.4697265625,
0.488037109375,
0.080078125,
0.488037109375
],
[
0.50244140625,
0.477783203125,
0.890625,
0.477783203125,
0.890625,
0.4912109375,
0.50244140625,
0.4912109375
],
[
0.07843017578125,
0.4873046875,
0.468017578125,
0.488037109375,
0.468017578125,
0.50244140625,
0.07843017578125,
0.501953125
],
[
0.5009765625,
0.492431640625,
0.890625,
0.492431640625,
0.890625,
0.5068359375,
0.5009765625,
0.5068359375
],
[
0.07843017578125,
0.50244140625,
0.4697265625,
0.50390625,
0.4697265625,
0.51806640625,
0.07843017578125,
0.517578125
],
[
0.50341796875,
0.50830078125,
0.890625,
0.50830078125,
0.890625,
0.521484375,
0.50341796875,
0.521484375
],
[
0.08087158203125,
0.51953125,
0.468017578125,
0.51953125,
0.468017578125,
0.53271484375,
0.08087158203125,
0.53271484375
],
[
0.50244140625,
0.52294921875,
0.89208984375,
0.52294921875,
0.89208984375,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.080078125,
0.5341796875,
0.468017578125,
0.5341796875,
0.468017578125,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.50244140625,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.080078125,
0.548828125,
0.468017578125,
0.548828125,
0.468017578125,
0.5625,
0.080078125,
0.5625
],
[
0.5009765625,
0.5517578125,
0.79345703125,
0.55322265625,
0.79345703125,
0.56640625,
0.5009765625,
0.56494140625
],
[
0.080078125,
0.564453125,
0.4697265625,
0.564453125,
0.4697265625,
0.57763671875,
0.080078125,
0.57763671875
],
[
0.52685546875,
0.57080078125,
0.88623046875,
0.57080078125,
0.88623046875,
0.58056640625,
0.52685546875,
0.58056640625
],
[
0.080078125,
0.5791015625,
0.468017578125,
0.5791015625,
0.468017578125,
0.59228515625,
0.080078125,
0.59228515625
],
[
0.50732421875,
0.583984375,
0.890625,
0.583984375,
0.890625,
0.59765625,
0.50732421875,
0.59765625
],
[
0.080078125,
0.5947265625,
0.468017578125,
0.5947265625,
0.468017578125,
0.60791015625,
0.080078125,
0.60791015625
],
[
0.50341796875,
0.5986328125,
0.88818359375,
0.5986328125,
0.88818359375,
0.6123046875,
0.50341796875,
0.6123046875
],
[
0.0882568359375,
0.611328125,
0.46728515625,
0.611328125,
0.46728515625,
0.6220703125,
0.0882568359375,
0.6220703125
],
[
0.51123046875,
0.6142578125,
0.8896484375,
0.6142578125,
0.8896484375,
0.62744140625,
0.51123046875,
0.62744140625
],
[
0.080078125,
0.625,
0.37744140625,
0.625,
0.37744140625,
0.6396484375,
0.080078125,
0.6396484375
],
[
0.5048828125,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.5048828125,
0.642578125
],
[
0.50341796875,
0.64306640625,
0.8896484375,
0.64306640625,
0.8896484375,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.50341796875,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.50341796875,
0.67236328125
],
[
0.080078125,
0.67041015625,
0.21728515625,
0.671875,
0.21728515625,
0.68603515625,
0.080078125,
0.68505859375
],
[
0.50341796875,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.50244140625,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50244140625,
0.7041015625
],
[
0.098876953125,
0.69775390625,
0.46728515625,
0.69775390625,
0.46728515625,
0.7109375,
0.098876953125,
0.7109375
],
[
0.50244140625,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.50244140625,
0.71826171875
],
[
0.080078125,
0.71337890625,
0.4697265625,
0.71337890625,
0.4697265625,
0.7265625,
0.080078125,
0.7265625
],
[
0.50244140625,
0.71826171875,
0.89208984375,
0.7197265625,
0.89208984375,
0.734375,
0.50244140625,
0.73291015625
],
[
0.07843017578125,
0.72900390625,
0.4697265625,
0.7265625,
0.4697265625,
0.7412109375,
0.07843017578125,
0.74365234375
],
[
0.50244140625,
0.73486328125,
0.89208984375,
0.73486328125,
0.89208984375,
0.74853515625,
0.50244140625,
0.74853515625
],
[
0.07843017578125,
0.7421875,
0.4697265625,
0.7412109375,
0.4697265625,
0.755859375,
0.07843017578125,
0.7568359375
],
[
0.50244140625,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76220703125,
0.50244140625,
0.76220703125
],
[
0.07843017578125,
0.755859375,
0.468017578125,
0.7568359375,
0.468017578125,
0.771484375,
0.07843017578125,
0.77099609375
],
[
0.50244140625,
0.76513671875,
0.89208984375,
0.76513671875,
0.89208984375,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.08087158203125,
0.7744140625,
0.46728515625,
0.7744140625,
0.46728515625,
0.78515625,
0.08087158203125,
0.78515625
],
[
0.50244140625,
0.77978515625,
0.89208984375,
0.77978515625,
0.89208984375,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.080078125,
0.78857421875,
0.4697265625,
0.78857421875,
0.4697265625,
0.8017578125,
0.080078125,
0.8017578125
],
[
0.50244140625,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.8095703125,
0.50244140625,
0.8095703125
],
[
0.080078125,
0.8037109375,
0.468017578125,
0.8037109375,
0.468017578125,
0.8173828125,
0.080078125,
0.8173828125
],
[
0.50244140625,
0.81103515625,
0.564453125,
0.81103515625,
0.564453125,
0.82275390625,
0.50244140625,
0.82275390625
],
[
0.080078125,
0.818359375,
0.468017578125,
0.818359375,
0.468017578125,
0.83203125,
0.080078125,
0.83203125
],
[
0.521484375,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.84033203125,
0.521484375,
0.84033203125
],
[
0.080078125,
0.833984375,
0.468017578125,
0.833984375,
0.468017578125,
0.84765625,
0.080078125,
0.84765625
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.8486328125,
0.471435546875,
0.8486328125,
0.471435546875,
0.8623046875,
0.080078125,
0.8623046875
],
[
0.50341796875,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.869140625,
0.50341796875,
0.869140625
],
[
0.09722900390625,
0.87548828125,
0.468017578125,
0.87548828125,
0.468017578125,
0.8896484375,
0.09722900390625,
0.8896484375
],
[
0.50244140625,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.07843017578125,
0.88623046875,
0.42724609375,
0.88720703125,
0.42724609375,
0.9013671875,
0.07843017578125,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.8896484375,
0.88623046875,
0.8896484375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.92529296875,
0.491943359375,
0.92529296875,
0.491943359375,
0.93505859375,
0.48046875,
0.93505859375
]
] |
[
"DualSmoke: Sketch-Based Smoke Illustration Design with Two-Stage",
"Generative Model",
"Haoran Xie*",
"Keisuke Arihara",
"Syuhei Sato",
"Kazunori Miyata",
"JAIST",
"JAIST",
"HoseiUniversity",
"JAIST",
"Abstract",
"idingForce",
"The dynamic effects of smoke are impressive in il.",
"lustration design, but it is a troublesome and chal-",
"lenging issue for common users to design the smoke",
"CGAN",
"GAN",
"effect without domain knowledge of fluid simulations.",
"In this work, we propose DualSmoke, two stage",
"User Sketch",
"Fluid Pattern (LCS)",
"Velocity Field",
"a",
"global-to-local generation framework for the interac-",
"tive smoke illustration design. For the global stage,",
"Figure 1. The proposed DualSmoke design framework uti-",
"lizes fuid patterns to generate velocity field in the genera-",
"the proposed approach utilizes fluid patterns to gen-",
"tive model.",
"erate Lagrangian coherent structure from the user's",
"hand-drawn sketches. For the local stage, the de-",
"tailed flow patterns are obtained from the generated",
"adopted kinetic textures [23] and elemental dymam-",
"coherent structure. Finally, we apply the guiding",
"ics [43l to author the dynamical effects, and sketch-",
"force field to the smoke simulator to design the de-",
"based approaches to design fuid motions [18, 46]",
"sired smoke illustration. To construct the training",
"However, these approaches may fail to provide the tur-",
"dataset, DualSmoke generates flow patterns using",
"bulent details in flow motions, especially for smoke de-",
"the finite-time Lyapunov exponents of the velocity",
"sign. To solve this issue, we aim to provide an efficient",
"fields. The synthetic sketch data is generated from",
"design tool for smoke illustration design.",
"the flow patterns by skeleton extraction. From our",
"用于流体控制的常规方法主要集中在形状束29",
"user study, it is verified that the proposed design in-",
"上。然而，由于非线性流，具有专业知识的用户需要耗时",
"terface can provide various smoke illustration designs",
"的过程和繁琐的参数调整。对于基于关键顿的方法「40",
"with good user usability.Our code is available at:",
"设计了优化过程来探索适当的外力，通常计算量很大。",
"https://github.com/shasph/DualSmoke:",
"此外，对于烟雾模拟，以交互方式调整参数和边界条件",
"是不实际的。To efficiently design global motions of the",
"fluid, low-resolution simulations were used as input for",
"1.Introduction",
"controlling high-resolution simulation [27,32] Low-",
"resolution simulations can be executed at low costs, so",
"Nowadays, due to the rapid prevalence of social me-",
"users can easily repeat the simulation multiple times to",
"dia and personal computing devices, it becomes an",
"design the motions. However, designing operations of",
"emerging and growing topic to help amateur users",
"fluid motions are not still intuitive, because users need",
"achieve high-quality content creation.The dynamic",
"to adjust simulation parameters.For more intuitive",
"effects of 2D illustration play an important role to in-",
"design, paths or sketches were introduced as inputs to",
"creasetheattraction of various mediasuch ase-books.",
"generate the fuid flows [24, 25l. We follow this idea",
"education materials, and art design. Especially, the",
"and our smoke design system is based on sketching op-",
"design of complex flow design is difficult for common",
"erations.",
"users without the specific domain knowledge and de-",
"In this work, we propose DualSmoke, a two-stage",
"sign skills, which are required in commercial design",
"generation framework with conditional generative ad-",
"tools, such as PhotoShop and Blender: Previous works",
"versarial networks (cGAN) for smoke illustration de-",
"*Japan Advanced Institute of Science and Technology, 1-1",
"sign to meet the users'design intention，如图E1所",
"Nomi, Ishikawa, 9231292, Japan. E-mail: xiejaist.ac.jp",
"示。Forthephase ofdatatraining，所提出的框架利",
"1"
] |
{
"arxiv_id": "2208.10906",
"language": "zh",
"page_idx": 0
}
|
||
2208.10906-zh-1
|
2208.10906-zh-1.jpg
|
用给定速度场的拉格朗日相干结构(LCS)来表示 velocity 场中的主流，该主流是作为有限时间 Lyapunov 指数(FTLE)场的脊而获得的。然后，我们使用热方程从 LCS 中提取合成草图。For the inference phase, we can generate the LCS domain and the velocity field
然后，
from the input of the user's freehand sketch.
我们从 the generated LCS 计算导向力和速度场 and add the force into the smoke simulations. Note that we set the flow direction as upward in the simulations
为了评估提出的系统，
as common design scenarios.
我们进行了评估实验，以验证系统的可用性和主观评估。在评估实验中，要求参与者在有和没有设计约束的情况下使用所提出的系统执行任务
这项工作的主要贡献如下。
$\cdot$ We propose DualSmoke , a two-stage generative model for smoke design considering the flow patterns as structure information in the learning proceSS.
$\bullet$ We constructed the paired training dataset with a pattern-guided smoke simulation where the FTLE and LCS fields are calculated to obtain the synthetic sketches.
$\bullet$ We develop a user-friendly design interface with the proposed system for various design scenarios and verify the proposed user interface with subjective assessment and user experience.
## 2. Related Work
## 2.1. Sketch-based Content Creation
Freehand sketching is a naive and efficient way to describe objects and motions for content creation in computer graphics and user interface [4]. For geometry modelling, sketch-based interfaces have been proposed for freeform modeling [21], terrain authoring [15], and character modeling [5]. For physical simulations, sketch input has been adopted in crystal [30] and character animations [14]. For more intuitive design, paths or sketches were introduced as inputs to generate the fluid flows [25,41]. In this work, we follow this sketchbased idea and adopt the deep learning based approach to guide the smoke design with the rough sketch input to achieve high-quality flow details.
Recently, the deep learning based approaches have been explored largely in sketch-based content creation. Hu 等人。$\mathbb{E}[18]$引人了用于流体设计的交互式用户界面，并基于 cGAN 获得的速度场生成了流体模拟。在地形建模中使用了类似的方法[15] 。Yan 等人。[44]在沉浸式虚拟现实环境中使用 3D 草图生成带有cGAN 的液体的飞溅表示。For these previous works, it is difficult to achieve high-quality design results. A
recent work adopted volumetric density fields in 3D smoke reconstruction using CNN networks [24]. However, the sketch input may require fine details of viewpoint and density information, rather than rough handdrawn sketches. To address these issues, we adopt the global-to-local two-stage deep generative modeling for smoke design. A similar strategy has been utilized in facial image synthesis [20] and character motion planning [28] . This work is the frst trial to apply the strategy in smoke design with pattern-guided flow simulations. With the pre-trained generative models, the proposed deep learning based framework can achieve an interactive system without the expensive computation on the fluid simulations.
2.2. Control of Fluid and Smoke Simulations
对于流体模拟，提出了一种稳定的流体求解器$\overline{\mathbb{E}}[36]$ 它是无条件稳定的，并已广泛用于研究和产品社区。
从那以后，求解器得到了其他研究人员的改进[10,34] 。但是，用户需要反复调整参数并运行仿真以生成所需的流量
To address this problem, many fuid control methods have been proposed. Treuille 等人。[40]使用准牛顿法计算最佳外力，以实现用户指定的关键帧。Rasmussen 等人。[29] 提出了一种通过控制粒子控制流体速度的方法。Dobashi et al. [9] controlled the cloud simulation to ft the cloud shape into an input curve. Simulation parameters are automatically adjusted according to the difference between the curve and the cloud shape. Wu et al. [41] introduced the pattern image and computed the driving force based on the image for locally controlling the fluid shape. Tang 等人。[39] 解决了高分辨率模拟控制力优化中缺乏参数空间约束和高维的问题。In these methods, users can specify the target shape of the fluid via input curves, images, or 3D meshes, but the flow motion cannot be designed intuitively.
To efficiently design the fluid flow, some researchers used velocity felds obtained from a low-resolution fluid simulation as an input. Nielsen et al. [27] proposed an optimization method to guide the high-resolution simulation according to the low-resolution input. Yuan 等人。[45] 提出了一种通过从低分辨率流中提取 LCS 来控制基于低分辨率流的高分辨率仿真流的方法。Huang et al. [19] also proposed a control method using the low-resolution simulation as a preview. This method can control any properties including velocity felds as well as densities, temperatures, and so on. Sato et al. [32] formulated the optimization problem in stream function space to reduce the problem size of Nielsen's method [27]. Our system also uses a velocity field and LCS to control the smoke simulation [45],but the velocity field and LCS are obtained via user sketches.
2
|
[
[
0.08251953125,
0.0921630859375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.1060791015625,
0.08251953125,
0.1060791015625
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08087158203125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08087158203125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.1363525390625,
0.08087158203125,
0.1363525390625
],
[
0.5009765625,
0.12249755859375,
0.890625,
0.121826171875,
0.890625,
0.1363525390625,
0.5009765625,
0.136962890625
],
[
0.08087158203125,
0.1383056640625,
0.46728515625,
0.1383056640625,
0.46728515625,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.5009765625,
0.136962890625,
0.8896484375,
0.136962890625,
0.8896484375,
0.15087890625,
0.5009765625,
0.15087890625
],
[
0.080078125,
0.1514892578125,
0.46728515625,
0.1514892578125,
0.46728515625,
0.1654052734375,
0.080078125,
0.1654052734375
],
[
0.50341796875,
0.15283203125,
0.89208984375,
0.15283203125,
0.89208984375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.07843017578125,
0.1666259765625,
0.468017578125,
0.1673583984375,
0.468017578125,
0.1817626953125,
0.07843017578125,
0.18115234375
],
[
0.5009765625,
0.1666259765625,
0.8896484375,
0.1666259765625,
0.8896484375,
0.18115234375,
0.5009765625,
0.18115234375
],
[
0.08087158203125,
0.18310546875,
0.468017578125,
0.18310546875,
0.468017578125,
0.1962890625,
0.08087158203125,
0.1962890625
],
[
0.5009765625,
0.1817626953125,
0.890625,
0.1817626953125,
0.890625,
0.1956787109375,
0.5009765625,
0.1956787109375
],
[
0.080078125,
0.1976318359375,
0.4697265625,
0.1976318359375,
0.4697265625,
0.2108154296875,
0.080078125,
0.2108154296875
],
[
0.50244140625,
0.1976318359375,
0.890625,
0.1976318359375,
0.890625,
0.2108154296875,
0.50244140625,
0.2108154296875
],
[
0.08087158203125,
0.21337890625,
0.4697265625,
0.21337890625,
0.4697265625,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.50244140625,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2279052734375,
0.50244140625,
0.2279052734375
],
[
0.306396484375,
0.2279052734375,
0.46728515625,
0.2279052734375,
0.46728515625,
0.2412109375,
0.306396484375,
0.2412109375
],
[
0.50244140625,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.08087158203125,
0.24365234375,
0.468017578125,
0.24365234375,
0.468017578125,
0.2568359375,
0.08087158203125,
0.2568359375
],
[
0.08087158203125,
0.2298583984375,
0.31201171875,
0.2298583984375,
0.31201171875,
0.240478515625,
0.08087158203125,
0.240478515625
],
[
0.50244140625,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.50244140625,
0.2568359375
],
[
0.08087158203125,
0.2568359375,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.08087158203125,
0.270751953125
],
[
0.50244140625,
0.2587890625,
0.89208984375,
0.2587890625,
0.89208984375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.080078125,
0.2734375,
0.325927734375,
0.272705078125,
0.325927734375,
0.287353515625,
0.080078125,
0.28857421875
],
[
0.50341796875,
0.2734375,
0.890625,
0.2734375,
0.890625,
0.287353515625,
0.50341796875,
0.287353515625
],
[
0.1004638671875,
0.2890625,
0.289306640625,
0.2890625,
0.289306640625,
0.302978515625,
0.1004638671875,
0.302978515625
],
[
0.50244140625,
0.2890625,
0.71337890625,
0.2890625,
0.71337890625,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.096435546875,
0.31494140625,
0.471435546875,
0.31494140625,
0.471435546875,
0.328857421875,
0.096435546875,
0.328857421875
],
[
0.50244140625,
0.31494140625,
0.8505859375,
0.31494140625,
0.8505859375,
0.328857421875,
0.50244140625,
0.328857421875
],
[
0.111083984375,
0.330810546875,
0.468017578125,
0.330810546875,
0.468017578125,
0.3447265625,
0.111083984375,
0.3447265625
],
[
0.111083984375,
0.3466796875,
0.468017578125,
0.3466796875,
0.468017578125,
0.35986328125,
0.111083984375,
0.35986328125
],
[
0.52392578125,
0.341552734375,
0.88818359375,
0.341552734375,
0.88818359375,
0.354736328125,
0.52392578125,
0.354736328125
],
[
0.111083984375,
0.35986328125,
0.1478271484375,
0.3623046875,
0.146240234375,
0.375,
0.1094970703125,
0.37255859375
],
[
0.51953125,
0.356689453125,
0.8896484375,
0.356689453125,
0.8896484375,
0.37060546875,
0.51953125,
0.37060546875
],
[
0.5048828125,
0.37255859375,
0.88818359375,
0.37255859375,
0.88818359375,
0.38330078125,
0.5048828125,
0.38330078125
],
[
0.09478759765625,
0.386962890625,
0.468017578125,
0.386962890625,
0.468017578125,
0.40087890625,
0.09478759765625,
0.40087890625
],
[
0.50341796875,
0.3837890625,
0.890625,
0.385009765625,
0.890625,
0.399658203125,
0.50341796875,
0.3984375
],
[
0.1094970703125,
0.40283203125,
0.4697265625,
0.40087890625,
0.4697265625,
0.415283203125,
0.1094970703125,
0.417236328125
],
[
0.50244140625,
0.40087890625,
0.5556640625,
0.40087890625,
0.5556640625,
0.415283203125,
0.50244140625,
0.415283203125
],
[
0.111083984375,
0.416748046875,
0.46728515625,
0.417236328125,
0.46728515625,
0.431884765625,
0.111083984375,
0.43115234375
],
[
0.52392578125,
0.417236328125,
0.890625,
0.417236328125,
0.890625,
0.43115234375,
0.52392578125,
0.43115234375
],
[
0.11273193359375,
0.43310546875,
0.2254638671875,
0.43310546875,
0.2254638671875,
0.4462890625,
0.11273193359375,
0.4462890625
],
[
0.5048828125,
0.43310546875,
0.8896484375,
0.43310546875,
0.8896484375,
0.4462890625,
0.5048828125,
0.4462890625
],
[
0.50341796875,
0.447509765625,
0.8896484375,
0.447509765625,
0.8896484375,
0.4609375,
0.50341796875,
0.4609375
],
[
0.09478759765625,
0.458984375,
0.468017578125,
0.458984375,
0.468017578125,
0.473388671875,
0.09478759765625,
0.473388671875
],
[
0.50244140625,
0.4609375,
0.89208984375,
0.46337890625,
0.89208984375,
0.477783203125,
0.50244140625,
0.47607421875
],
[
0.111083984375,
0.47607421875,
0.4697265625,
0.47607421875,
0.4697265625,
0.4892578125,
0.111083984375,
0.4892578125
],
[
0.50341796875,
0.477783203125,
0.89208984375,
0.477783203125,
0.89208984375,
0.4912109375,
0.50341796875,
0.4912109375
],
[
0.1094970703125,
0.490478515625,
0.46728515625,
0.490478515625,
0.46728515625,
0.50390625,
0.1094970703125,
0.50390625
],
[
0.50244140625,
0.49365234375,
0.8896484375,
0.49365234375,
0.8896484375,
0.5068359375,
0.50244140625,
0.5068359375
],
[
0.111083984375,
0.505859375,
0.393798828125,
0.505859375,
0.393798828125,
0.51953125,
0.111083984375,
0.51953125
],
[
0.50341796875,
0.50830078125,
0.89208984375,
0.50830078125,
0.89208984375,
0.521484375,
0.50341796875,
0.521484375
],
[
0.50244140625,
0.52294921875,
0.89208984375,
0.52294921875,
0.89208984375,
0.5361328125,
0.50244140625,
0.5361328125
],
[
0.07843017578125,
0.53515625,
0.2296142578125,
0.53515625,
0.2296142578125,
0.5498046875,
0.07843017578125,
0.5498046875
],
[
0.50244140625,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.5517578125,
0.50244140625,
0.5517578125
],
[
0.50244140625,
0.55322265625,
0.89208984375,
0.5517578125,
0.89208984375,
0.56640625,
0.50244140625,
0.5673828125
],
[
0.07843017578125,
0.5595703125,
0.359375,
0.56005859375,
0.359375,
0.57470703125,
0.07843017578125,
0.57373046875
],
[
0.50341796875,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58203125,
0.50341796875,
0.58203125
],
[
0.098876953125,
0.583984375,
0.46728515625,
0.583984375,
0.46728515625,
0.59765625,
0.098876953125,
0.59765625
],
[
0.5048828125,
0.58349609375,
0.8896484375,
0.58349609375,
0.8896484375,
0.5966796875,
0.5048828125,
0.5966796875
],
[
0.08087158203125,
0.599609375,
0.46728515625,
0.599609375,
0.46728515625,
0.61279296875,
0.08087158203125,
0.61279296875
],
[
0.50341796875,
0.5986328125,
0.89208984375,
0.5986328125,
0.89208984375,
0.6123046875,
0.50341796875,
0.6123046875
],
[
0.080078125,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.080078125,
0.62890625
],
[
0.50244140625,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.08087158203125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64306640625,
0.08087158203125,
0.64306640625
],
[
0.5009765625,
0.62890625,
0.89208984375,
0.62744140625,
0.89208984375,
0.642578125,
0.5009765625,
0.64306640625
],
[
0.07843017578125,
0.64306640625,
0.4697265625,
0.64453125,
0.4697265625,
0.6591796875,
0.07843017578125,
0.65771484375
],
[
0.50244140625,
0.64453125,
0.6435546875,
0.64453125,
0.6435546875,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.521484375,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.521484375,
0.673828125
],
[
0.08087158203125,
0.6767578125,
0.46728515625,
0.6767578125,
0.46728515625,
0.6875,
0.08087158203125,
0.6875
],
[
0.50341796875,
0.6767578125,
0.890625,
0.6767578125,
0.890625,
0.6875,
0.50341796875,
0.6875
],
[
0.080078125,
0.68994140625,
0.4697265625,
0.68994140625,
0.4697265625,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.5009765625,
0.689453125,
0.89208984375,
0.68994140625,
0.89208984375,
0.705078125,
0.5009765625,
0.7041015625
],
[
0.080078125,
0.705078125,
0.4697265625,
0.705078125,
0.4697265625,
0.71826171875,
0.080078125,
0.71826171875
],
[
0.50244140625,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.50244140625,
0.7197265625
],
[
0.080078125,
0.7197265625,
0.46728515625,
0.7197265625,
0.46728515625,
0.73291015625,
0.080078125,
0.73291015625
],
[
0.50341796875,
0.72021484375,
0.89208984375,
0.72021484375,
0.89208984375,
0.734375,
0.50341796875,
0.734375
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.5048828125,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.5048828125,
0.74853515625
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.7646484375,
0.080078125,
0.7646484375
],
[
0.50244140625,
0.75,
0.89208984375,
0.75048828125,
0.89208984375,
0.76513671875,
0.50244140625,
0.7646484375
],
[
0.08087158203125,
0.767578125,
0.334228515625,
0.767578125,
0.334228515625,
0.77783203125,
0.08087158203125,
0.77783203125
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.098876953125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.098876953125,
0.79443359375
],
[
0.50244140625,
0.78076171875,
0.89208984375,
0.78076171875,
0.89208984375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.080078125,
0.796875,
0.46728515625,
0.796875,
0.46728515625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.796875,
0.89208984375,
0.796875,
0.89208984375,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.09234619140625,
0.8095703125,
0.46728515625,
0.81005859375,
0.46728515625,
0.82470703125,
0.09234619140625,
0.82373046875
],
[
0.5009765625,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5009765625,
0.82470703125
],
[
0.08251953125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.08087158203125,
0.85595703125,
0.46728515625,
0.85595703125,
0.46728515625,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.50244140625,
0.87060546875
],
[
0.080078125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88623046875,
0.080078125,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50244140625,
0.88525390625
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88623046875,
0.8837890625,
0.88623046875,
0.8837890625,
0.8994140625,
0.50341796875,
0.8994140625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"用给定速度场的拉格朗日相干结构（LCS）来表示ve-",
"recent work adopted volumetric density fields in 3D",
"locity场中的主流，该主流是作为有限时间Lyapunov",
"smoke reconstruction using CNN networks [24] . How-",
"指数（FTLE）场的脊而获得的。然后，我们使用热方",
"ever, the sketch input may require fine details of view-",
"程从LCS中提取合成草图。Fortheinferencephase,",
"point and density information, rather than rough hand-",
"we can generate the LCS domain and the velocity field",
"drawn sketches. To address these issues, we adopt the",
"from the input of the user's freehand sketch.然后，",
"global-to-local two-stage deep generative modeling for",
"我们从thegeneratedLCS计算导向力和速度场and",
"smoke design. A similar strategy has been utilized in",
"add the force into the smoke simulations.Note that",
"facial image synthesis [20l and character motion plan-",
"we set the flow direction as upward in the simulations",
"ning [28] . This work is the first trial to apply the",
"为了评估提出的系统，",
"strategy in smoke design with pattern-guided flow sim-",
"我们进行了评估实验，以验证系统的可用性和主观评",
"as commondesignscenarios.",
"ulations. With the pre-trained generative models, the",
"估。在评估实验中，要求参与者在有和没有设计约束的",
"proposed deep learning based framework can achieve",
"情况下使用所提出的系统执行任务",
"an interactive system without the expensive computa-",
"这项工作的主要贡献如下。",
"tion on the fluid simulations.",
": We propose DualSmoke , a two-stage generative",
"2.2. Control of Fluid and Smoke Simulations",
"model for smoke design considering the flow pat-",
"terns as structure information in the learning pro-",
"对于流体模拟，提出了一种稳定的流体求解器E[36",
"cess.",
"它是无条件稳定的，并已广泛用于研究和产品社区。",
"从那以后，求解器得到了其他研究人员的改进「10.34",
". We constructed the paired training dataset with a",
"。但是，用户需要反复调整参数并运行仿真以生成所需",
" pattern-guided smoke simulation where the FTLE",
"的流量",
"and LCS fields are calculated to obtain the syn-",
"To address this problem, many fuid control meth-",
"thetic sketches.",
"odshavebeenproposed.Treuille等人。[4ol使用",
"准牛顿法计算最佳外力，以实现用户指定的关键顿。",
": We develop a user-friendly design interface with",
"Rasmussen等人。[29]提出了一种通过控制粒子控制流",
"the proposed system for various design scenarios",
"体速度的方法。Dobashietal.[9]controlled the cloud",
"and verify the proposed user interface with sub-",
"simulation to fit the cloud shape into an input curve.",
"jective assessment and user experience.",
"Simulation parameters are automatically adjusted ac-",
"cording to the difference between the curve and the",
"2. Related Work",
"cloud shape. Wu et al. [4ll introduced the pattern",
"image and computed the driving force based on the",
"2.1. Sketch-based Content Creation",
"image for locally controlling the fluid shape.Tang 等",
"Freehand sketching is a naive and efficient way to de-",
"人。[39]解决了高分辨率模拟控制力优化中缺乏参数",
"scribe objects and motions for content creation in com-",
"空间约束和高维的问题。In these methods，users can",
"puter graphics and user interface [4] . For geometry",
"specify the target shape of the fluid via input curves,",
"modelling, sketch-based interfaces have been proposed",
"images, or 3D meshes, but the flow motion cannot be",
"for freeform modeling [21, terrain authoring [15],",
"designed intuitively",
"and character modeling [5l . For physical simulations,",
"To efficiently design the fluid flow, some researchers",
"sketch input has been adopted in crystal 30 and char-",
"used velocityfields obtained fromalow-resolutionfluid",
"acter animations [14l . For more intuitive design, paths",
"simulation as an input. Nielsen et al. [27] proposed an",
"or sketches were introduced as inputs to generate the",
"optimization method to guide the high-resolution sim-",
"fluid flows 25,41l . In this work, we follow this sketch-",
"ulation according to the low-resolution input.Yuan 等",
"based idea and adopt the deep learning based approach",
"人。「45]提出了一种通过从低分辨率流中提取LCS来",
"to guide the smoke design with the rough sketch input",
"控制基于低分辨率流的高分辨率仿真流的方法。Huang",
"to achieve high-quality flow details.",
"et al.[19] also proposed a control method using the",
"Recently, the deep learning based approaches have",
"low-resolution simulation as a preview. This method",
"been explored largely in sketch-based content creation.",
"can control any properties including velocity fields as",
"Hu等人。E[181引人了用于流体设计的交互式用",
"well as densities, temperatures, and so on. Sato et",
"户界面，并基于cGAN获得的速度场生成了流体模",
"al. [32 formulated the optimization problem in stream",
"拟。在地形建模中使用了类似的方法[15]。Yan等人。",
"function space to reduce the problem size of Nielsen's",
"[44]在沉浸式虚拟现实环境中使用3D草图生成带有",
"method [27] : Our system also uses a velocity field",
"cGAN的液体的飞溅表示。For these previous works,",
"and LCS to control the smoke simulation [45l , but the",
"it is difficult to achieve high-quality design results. A",
"velocity field and LCS are obtained via user sketches.",
"2"
] |
{
"arxiv_id": "2208.10906",
"language": "zh",
"page_idx": 1
}
|
||
2205.15319-zh-0
|
2205.15319-zh-0.jpg
|
# AdaProp: Learning Adaptive Propagation for Graph Neural Network based Knowledge Graph Reasoning
Yongqi Zhang"
zhangyongqi@4paradigm.com
4Paradigm Inc. Beijing, China
Zhanke Zhou$^{*}$
cszkzhou@comp.hkbu.edu.hk Department of Computer Science, Hong Kong Baptist University
Hong Kong, China
Quanming Yao$^{\dagger}$
qyaoaa@tsinghua.edu.cn Department of Electronic
Engineering, Tsinghua University
Beijing, China
Xiaowen Chu xwchu@ust.hk
Data Science and Analytics Thrust,
HKUST (Guangzhou) Guangzhou, China
Bo Han
bhanml@comp.hkbu.edu.hk
Department of Computer Science, Hong Kong Baptist University
Hong Kong, China
# $\mathbf{ABSTRACT}$
ACM Reference Format:
Yongqi Zhang, Zhanke Zhou, Quanming Yao, Xiaowen Chu, and Bo Han. 2023. AdaProp: Learning Adaptive Propagation for Graph Neural Network based Knowledge Graph Reasoning .In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '23), August 6-10, 2023, Long Beach, CA, USA. ACM, New York, NY, USA, 16 pages. https://doi.org/10.1145/3580305.3599404
# 1 INTRODUCTION
Due to the popularity of Graph Neural Networks (GNNs), various GNN-based methods have been designed to reason on knowledge graphs (KGs). An important design component of GNN-based KG reasoning methods is called the propagation path, which contains a set of involved entities in each propagation step. Existing methods use hand-designed propagation paths, ignoring the correlation between the entities and the query relation. In addition, the number of involved entities will explosively grow at larger propagation steps. In this work, we are motivated to learn an adaptive propagation path in order to flter out irrelevant entities while preserving promising targets. First, we design an incremental sampling mechanism where the nearby targets and layer-wise connections can be preserved with linear complexity. Second, we design a learning-based sampling distribution to identify the semantically related entities. Extensive experiments show that our method is powerful, efficient and semantic-aware. The code is available at https://github.com/LARS-research/AdaProp .
# $\csc\cos$coNCEPTS $\cdot$ Computing methodologies $\to$Knowledge representation and reasoning.
$\mathbf{KEYWORDS}$
知识图(KG)是一个语义图，表示现实世界实体[15,35]之间的关系。基于 KGs 的推理旨在根据现有事实推导给定查询的缺失答案[1,6,18] 。已广泛应用于药物相互作用预测[44] 、个性化推荐$\left[4,36\right]$、问答$\left[11,16\right]$。带有查询 (LeBron, lives\_in,?) 和答案 LosAngeles 的 KG 推理示例如图$\boxed{\mathrm{E}}$1 所示。从符号上讲推理问题可以表示为查询$(e_q,r_q,?)$,其中查询实体$e_q$和查询关系$r_q$。KG 推理的目的是通过从给定 KG 中学习的本地证据来找到$^{\prime}(e_q,r_q,?)$的目标答案实体$e_a$
用于 KG 推理的代表性方法可以分为三类：(i) 基于三元组的模型[3,9,35,45,48]通过学习的实体和关系嵌人直接对每个答案实体进行评分；(ii) 基于路径的方法$\left[7,8,28,29,47\right]$学习逻辑规则，以从 $e_q$ 开始生成关系路径，并通过基于 MAMTHXXIGX5的实体传播方法探索哪个实体更可能是目标答案。在这三个类别中，基于 GNN 的方法实现了最先进的性能[42,46,49] 。具体来说，R-GCN[30] 和 CompGCN[34] 在所有实体的整个邻域中传播。与一般 GCN 相同，它们很容易遭受过度平滑问题[26]。NBFNet[49]和RED-GNN[46]在查询实体$e_q$周围的本地邻域中传播。通过选择传播范围，它们具有比 R-GCN 或CompGCN 更强的推理能力。但是，它们的本地邻域主要依赖于查询实体，并且大多数实体将在更大的传播步骤中涉及
现有的基于 GNN 的方法在所有实体上或围绕查询实体的所有邻域之间传播，忽略了局部邻域与查询关系之间的语义相关性。直观地说，我们需要的本地信息取决于查询实体和查询关系。例如，与图 1 中的查询最相关的本地证据包括 (LeBron, 部分\_of, 湖人队)&(湖人队，位于\_in, L。A。)和(勒布朗，萨凡·纳的配偶)&(萨凡·纳，住在，L。A。)。如果查询更改为(勒布朗，父亲，?),则最相关的信息将更改为(勒布朗 JR,儿子，勒布朗)。此外，现有的基于GNN 的方法不可避免地会涉及太多的无关实体和事实，增加了学习难度和计算成本，尤其是在大规模的 KG 上
Knowledge graph, Graph embedding, Knowledge graph reasoning,
Graph sampling, Graph neural network
$^{\cdot}$This work was partially performed when Z. Zhou was an intern in 4Paradigm. Y. Zhang and Z. Zhou have equal contribution, and correspondence is to Q. Yao
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. $KDD'23,August~6-10,~2023,~Long~Beach,~CA,~USA$
$\odot$ 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 979-8-4007-0103-0/23/08....815.00
https://doi.org/10.1145/3580305.3599404
|
[
[
0.1217041015625,
0.10418701171875,
0.87744140625,
0.10418701171875,
0.87744140625,
0.125,
0.1217041015625,
0.125
],
[
0.2099609375,
0.126953125,
0.78759765625,
0.1300048828125,
0.78759765625,
0.1514892578125,
0.2099609375,
0.1483154296875
],
[
0.1690673828125,
0.169189453125,
0.289306640625,
0.1673583984375,
0.289306640625,
0.1849365234375,
0.1690673828125,
0.1875
],
[
0.441162109375,
0.169189453125,
0.55615234375,
0.169189453125,
0.55615234375,
0.184326171875,
0.441162109375,
0.184326171875
],
[
0.69921875,
0.1685791015625,
0.833984375,
0.1622314453125,
0.8349609375,
0.184326171875,
0.701171875,
0.1900634765625
],
[
0.12744140625,
0.1849365234375,
0.336669921875,
0.186279296875,
0.336669921875,
0.2008056640625,
0.12744140625,
0.199462890625
],
[
0.400390625,
0.186279296875,
0.6005859375,
0.186279296875,
0.6005859375,
0.199462890625,
0.400390625,
0.199462890625
],
[
0.68310546875,
0.186279296875,
0.85302734375,
0.1849365234375,
0.85302734375,
0.199462890625,
0.68310546875,
0.2008056640625
],
[
0.1805419921875,
0.199462890625,
0.281982421875,
0.2020263671875,
0.281982421875,
0.216552734375,
0.1805419921875,
0.2139892578125
],
[
0.3857421875,
0.2020263671875,
0.61376953125,
0.2020263671875,
0.61376953125,
0.21533203125,
0.3857421875,
0.21533203125
],
[
0.68310546875,
0.2020263671875,
0.8544921875,
0.2020263671875,
0.8544921875,
0.21533203125,
0.68310546875,
0.21533203125
],
[
0.1829833984375,
0.216552734375,
0.281005859375,
0.216552734375,
0.281005859375,
0.2310791015625,
0.1829833984375,
0.2310791015625
],
[
0.396240234375,
0.21533203125,
0.60107421875,
0.216552734375,
0.60107421875,
0.2310791015625,
0.396240234375,
0.2298583984375
],
[
0.654296875,
0.216552734375,
0.88134765625,
0.216552734375,
0.88134765625,
0.2298583984375,
0.654296875,
0.2298583984375
],
[
0.435546875,
0.2310791015625,
0.56396484375,
0.2310791015625,
0.56396484375,
0.2442626953125,
0.435546875,
0.2442626953125
],
[
0.71826171875,
0.2310791015625,
0.81787109375,
0.2310791015625,
0.81787109375,
0.24560546875,
0.71826171875,
0.24560546875
],
[
0.30810546875,
0.2587890625,
0.420654296875,
0.2587890625,
0.420654296875,
0.2734375,
0.30810546875,
0.2734375
],
[
0.60107421875,
0.255615234375,
0.66650390625,
0.2587890625,
0.6650390625,
0.274658203125,
0.6005859375,
0.271484375
],
[
0.31201171875,
0.27587890625,
0.418212890625,
0.2734375,
0.418212890625,
0.28857421875,
0.31201171875,
0.290283203125
],
[
0.53759765625,
0.274658203125,
0.7294921875,
0.274658203125,
0.7294921875,
0.2890625,
0.53759765625,
0.2890625
],
[
0.2467041015625,
0.290283203125,
0.48193359375,
0.290283203125,
0.48193359375,
0.3037109375,
0.2467041015625,
0.3037109375
],
[
0.5185546875,
0.290283203125,
0.74658203125,
0.290283203125,
0.74658203125,
0.3037109375,
0.5185546875,
0.3037109375
],
[
0.291748046875,
0.3037109375,
0.438720703125,
0.304931640625,
0.438720703125,
0.3193359375,
0.291748046875,
0.318115234375
],
[
0.53125,
0.30615234375,
0.734375,
0.30615234375,
0.734375,
0.3193359375,
0.53125,
0.3193359375
],
[
0.301513671875,
0.32080078125,
0.4306640625,
0.32080078125,
0.4306640625,
0.333984375,
0.301513671875,
0.333984375
],
[
0.56884765625,
0.32080078125,
0.69775390625,
0.32080078125,
0.69775390625,
0.333984375,
0.56884765625,
0.333984375
],
[
0.0882568359375,
0.34521484375,
0.181396484375,
0.34521484375,
0.181396484375,
0.356689453125,
0.0882568359375,
0.356689453125
],
[
0.51953125,
0.347900390625,
0.66259765625,
0.347900390625,
0.66259765625,
0.35791015625,
0.51953125,
0.35791015625
],
[
0.0849609375,
0.3623046875,
0.48046875,
0.3623046875,
0.48046875,
0.375732421875,
0.0849609375,
0.375732421875
],
[
0.51708984375,
0.359130859375,
0.91259765625,
0.359130859375,
0.91259765625,
0.37255859375,
0.51708984375,
0.37255859375
],
[
0.08416748046875,
0.375732421875,
0.48046875,
0.37744140625,
0.48046875,
0.391357421875,
0.08416748046875,
0.389404296875
],
[
0.51708984375,
0.371826171875,
0.9140625,
0.371826171875,
0.9140625,
0.385009765625,
0.51708984375,
0.385009765625
],
[
0.51708984375,
0.3837890625,
0.9140625,
0.3837890625,
0.9140625,
0.397705078125,
0.51708984375,
0.397705078125
],
[
0.0849609375,
0.391357421875,
0.48193359375,
0.391357421875,
0.48193359375,
0.40478515625,
0.0849609375,
0.40478515625
],
[
0.51708984375,
0.395751953125,
0.91259765625,
0.396484375,
0.91259765625,
0.410400390625,
0.51708984375,
0.4091796875
],
[
0.0849609375,
0.406005859375,
0.48193359375,
0.406005859375,
0.48193359375,
0.419189453125,
0.0849609375,
0.419189453125
],
[
0.51708984375,
0.4091796875,
0.9140625,
0.4091796875,
0.9140625,
0.42236328125,
0.51708984375,
0.42236328125
],
[
0.0849609375,
0.419189453125,
0.48193359375,
0.419189453125,
0.48193359375,
0.43310546875,
0.0849609375,
0.43310546875
],
[
0.515625,
0.42236328125,
0.7861328125,
0.42041015625,
0.7861328125,
0.434326171875,
0.515625,
0.436279296875
],
[
0.08660888671875,
0.434326171875,
0.4794921875,
0.434326171875,
0.4794921875,
0.4443359375,
0.08660888671875,
0.4443359375
],
[
0.0849609375,
0.4462890625,
0.48193359375,
0.4462890625,
0.48193359375,
0.460205078125,
0.0849609375,
0.460205078125
],
[
0.0849609375,
0.460205078125,
0.483642578125,
0.460205078125,
0.483642578125,
0.473388671875,
0.0849609375,
0.473388671875
],
[
0.5185546875,
0.46533203125,
0.53173828125,
0.462158203125,
0.5361328125,
0.4697265625,
0.52294921875,
0.473388671875
],
[
0.5361328125,
0.466552734375,
0.54345703125,
0.466552734375,
0.54345703125,
0.47021484375,
0.5361328125,
0.47021484375
],
[
0.54833984375,
0.46337890625,
0.69140625,
0.46337890625,
0.69140625,
0.473388671875,
0.54833984375,
0.473388671875
],
[
0.08416748046875,
0.474853515625,
0.48193359375,
0.473388671875,
0.48193359375,
0.4873046875,
0.08416748046875,
0.488037109375
],
[
0.51708984375,
0.479248046875,
0.9140625,
0.479736328125,
0.9140625,
0.494384765625,
0.51708984375,
0.49365234375
],
[
0.08416748046875,
0.488037109375,
0.48046875,
0.4873046875,
0.48046875,
0.50048828125,
0.08416748046875,
0.501953125
],
[
0.5185546875,
0.494384765625,
0.91259765625,
0.494384765625,
0.91259765625,
0.50830078125,
0.5185546875,
0.50830078125
],
[
0.08660888671875,
0.50390625,
0.48046875,
0.50390625,
0.48046875,
0.5146484375,
0.08660888671875,
0.5146484375
],
[
0.51953125,
0.51025390625,
0.91015625,
0.51025390625,
0.91015625,
0.52099609375,
0.51953125,
0.52099609375
],
[
0.08660888671875,
0.517578125,
0.48046875,
0.517578125,
0.48046875,
0.52783203125,
0.08660888671875,
0.52783203125
],
[
0.5185546875,
0.521484375,
0.9140625,
0.521484375,
0.9140625,
0.53515625,
0.5185546875,
0.53515625
],
[
0.08660888671875,
0.52978515625,
0.476318359375,
0.52978515625,
0.476318359375,
0.54052734375,
0.08660888671875,
0.54052734375
],
[
0.5185546875,
0.53515625,
0.91259765625,
0.53515625,
0.91259765625,
0.548828125,
0.5185546875,
0.548828125
],
[
0.0849609375,
0.54345703125,
0.48046875,
0.54345703125,
0.48046875,
0.556640625,
0.0849609375,
0.556640625
],
[
0.5185546875,
0.5498046875,
0.91259765625,
0.5498046875,
0.91259765625,
0.564453125,
0.5185546875,
0.564453125
],
[
0.08660888671875,
0.5595703125,
0.4794921875,
0.5595703125,
0.4794921875,
0.5693359375,
0.08660888671875,
0.5693359375
],
[
0.51953125,
0.56298828125,
0.9140625,
0.56298828125,
0.9140625,
0.57763671875,
0.51953125,
0.57763671875
],
[
0.08660888671875,
0.5712890625,
0.48046875,
0.5712890625,
0.48046875,
0.58203125,
0.08660888671875,
0.58203125
],
[
0.08416748046875,
0.58349609375,
0.426513671875,
0.58544921875,
0.426513671875,
0.599609375,
0.08416748046875,
0.59765625
],
[
0.5185546875,
0.57763671875,
0.74609375,
0.57763671875,
0.74609375,
0.59228515625,
0.5185546875,
0.59228515625
],
[
0.53369140625,
0.59033203125,
0.9140625,
0.59130859375,
0.9140625,
0.60595703125,
0.53369140625,
0.60498046875
],
[
0.5185546875,
0.60595703125,
0.91259765625,
0.60595703125,
0.91259765625,
0.619140625,
0.5185546875,
0.619140625
],
[
0.08660888671875,
0.61279296875,
0.221435546875,
0.61279296875,
0.221435546875,
0.626953125,
0.08660888671875,
0.626953125
],
[
0.5185546875,
0.619140625,
0.9140625,
0.619140625,
0.9140625,
0.63330078125,
0.5185546875,
0.63330078125
],
[
0.0849609375,
0.630859375,
0.48193359375,
0.630859375,
0.48193359375,
0.64453125,
0.0849609375,
0.64453125
],
[
0.51953125,
0.6337890625,
0.9140625,
0.6337890625,
0.9140625,
0.64794921875,
0.51953125,
0.64794921875
],
[
0.08416748046875,
0.642578125,
0.1854248046875,
0.64599609375,
0.1854248046875,
0.66015625,
0.08416748046875,
0.6572265625
],
[
0.5185546875,
0.64794921875,
0.91259765625,
0.64794921875,
0.91259765625,
0.6611328125,
0.5185546875,
0.6611328125
],
[
0.51953125,
0.6611328125,
0.9140625,
0.6611328125,
0.9140625,
0.6748046875,
0.51953125,
0.6748046875
],
[
0.08660888671875,
0.673828125,
0.188720703125,
0.673828125,
0.188720703125,
0.68505859375,
0.08660888671875,
0.68505859375
],
[
0.521484375,
0.6767578125,
0.9111328125,
0.6767578125,
0.9111328125,
0.6875,
0.521484375,
0.6875
],
[
0.08416748046875,
0.689453125,
0.48046875,
0.69140625,
0.48046875,
0.70556640625,
0.08416748046875,
0.7041015625
],
[
0.5185546875,
0.68798828125,
0.9140625,
0.68798828125,
0.9140625,
0.7021484375,
0.5185546875,
0.7021484375
],
[
0.0849609375,
0.7041015625,
0.328369140625,
0.7041015625,
0.328369140625,
0.71728515625,
0.0849609375,
0.71728515625
],
[
0.5185546875,
0.70263671875,
0.91259765625,
0.70263671875,
0.91259765625,
0.71630859375,
0.5185546875,
0.71630859375
],
[
0.51953125,
0.71826171875,
0.9111328125,
0.71826171875,
0.9111328125,
0.72900390625,
0.51953125,
0.72900390625
],
[
0.08416748046875,
0.736328125,
0.455078125,
0.7373046875,
0.455078125,
0.751953125,
0.08416748046875,
0.75048828125
],
[
0.51708984375,
0.73095703125,
0.91259765625,
0.73095703125,
0.91259765625,
0.744140625,
0.51708984375,
0.744140625
],
[
0.08251953125,
0.74755859375,
0.459228515625,
0.74853515625,
0.459228515625,
0.76220703125,
0.08251953125,
0.76123046875
],
[
0.51953125,
0.744140625,
0.8857421875,
0.744140625,
0.8857421875,
0.75830078125,
0.51953125,
0.75830078125
],
[
0.53515625,
0.75830078125,
0.91259765625,
0.75830078125,
0.91259765625,
0.771484375,
0.53515625,
0.771484375
],
[
0.5185546875,
0.771484375,
0.91259765625,
0.771484375,
0.91259765625,
0.78515625,
0.5185546875,
0.78515625
],
[
0.08416748046875,
0.77880859375,
0.48046875,
0.77978515625,
0.48046875,
0.79345703125,
0.08416748046875,
0.79248046875
],
[
0.08660888671875,
0.79248046875,
0.4794921875,
0.79248046875,
0.4794921875,
0.80322265625,
0.08660888671875,
0.80322265625
],
[
0.51953125,
0.787109375,
0.9111328125,
0.787109375,
0.9111328125,
0.79736328125,
0.51953125,
0.79736328125
],
[
0.08660888671875,
0.80322265625,
0.48046875,
0.80322265625,
0.48046875,
0.81298828125,
0.08660888671875,
0.81298828125
],
[
0.51953125,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.51953125,
0.81298828125
],
[
0.08660888671875,
0.81298828125,
0.48046875,
0.81298828125,
0.48046875,
0.82373046875,
0.08660888671875,
0.82373046875
],
[
0.51708984375,
0.81298828125,
0.9140625,
0.8125,
0.9140625,
0.82568359375,
0.51708984375,
0.8271484375
],
[
0.08416748046875,
0.8212890625,
0.48193359375,
0.8212890625,
0.48193359375,
0.83544921875,
0.08416748046875,
0.83544921875
],
[
0.08660888671875,
0.83349609375,
0.47802734375,
0.83349609375,
0.47802734375,
0.84326171875,
0.08660888671875,
0.84326171875
],
[
0.521484375,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.521484375,
0.8388671875
],
[
0.0849609375,
0.84130859375,
0.471435546875,
0.84326171875,
0.471435546875,
0.85400390625,
0.0849609375,
0.8515625
],
[
0.5185546875,
0.84130859375,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.5185546875,
0.85498046875
],
[
0.08660888671875,
0.85400390625,
0.316162109375,
0.85400390625,
0.316162109375,
0.8642578125,
0.08660888671875,
0.8642578125
],
[
0.521484375,
0.85498046875,
0.91259765625,
0.85498046875,
0.91259765625,
0.86865234375,
0.521484375,
0.86865234375
],
[
0.0849609375,
0.86279296875,
0.473876953125,
0.8623046875,
0.473876953125,
0.87548828125,
0.0849609375,
0.876953125
],
[
0.5185546875,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.5185546875,
0.8818359375
],
[
0.08660888671875,
0.875,
0.293212890625,
0.875,
0.293212890625,
0.88525390625,
0.08660888671875,
0.88525390625
],
[
0.0849609375,
0.88525390625,
0.273681640625,
0.88525390625,
0.273681640625,
0.89599609375,
0.0849609375,
0.89599609375
],
[
0.5185546875,
0.88330078125,
0.650390625,
0.88330078125,
0.650390625,
0.896484375,
0.5185546875,
0.896484375
]
] |
[
"AdaProp: Learning Adaptive Propagation for Graph Neural",
"Network based Knowledge Graph Reasoning",
"Yongqi Zhang*",
"Zhanke Zhou*",
"Quanming Yao",
"zhangyongqi@4paradigm.com",
"cszkzhou@comp.hkbu.edu.hk",
"qyaoaa@tsinghua.edu.cn",
"4Paradigm Inc.",
"Department of Computer Science,",
"Department of Electronic",
"Beijing, China",
"Hong Kong Baptist University",
"Engineering, Tsinghua University",
"Hong Kong, China",
"Beijing, China",
"Xiaowen Chu",
"Bo Han",
"xwchu@ust.hk",
"bhanml@comp.hkbu.edu.hk",
"Data Science and Analytics Thrust,",
"Department of Computer Science,",
"HKUST (Guangzhou)",
"Hong Kong Baptist University",
"Guangzhou, China",
"Hong Kong, China",
"ABSTRACT",
"ACMReferenceFormat:",
"Due to the popularity of Graph Neural Networks (GNNs), various",
"Yongqi Zhang, Zhanke Zhou, Quanming Yao, Xiaowen Chu, and Bo Han.",
"GNN-based methods have been designed to reason on knowledge",
"2023. AdaProp: Learning Adaptive Propagation for Graph Neural Network",
"based Knowledge Graph Reasoning: In Proceedings of the 29th ACM",
"graphs (KGs). An important design component of GNN-based KG",
"SIGKDD Conference on Knowledge Discovery and Data Mining (KDD'23)",
"reasoning methods is called the propagation path, which contains a",
"August 6-10, 2023, Long Beach, CA, USA. ACM, New York, NY, USA,",
"set of involved entities in each propagation step. Existing methods",
"16pages.https://doi.org/10.1145/3580305.3599404",
"use hand-designed propagation paths,ignoring the correlation",
"between the entities and the query relation. In addition, the",
"number of involved entities will explosively grow at larger propa-",
"1",
"",
"INTRODUCTION",
"gation steps.In this work, we are motivated to learn an adaptive",
"知识图（KG）是一个语义图，表示现实世界实体[15,35]之间",
"propagation path in order to filter out irrelevant entities while",
"的关系。基于KGs的推理旨在根据现有事实推导给定查询的",
"preserving promising targets. First, we design an incremental",
"缺失答案[1,6，18]。已广泛应用于药物相互作用预测【44]、个",
"sampling mechanism where the nearby targets and layer-wise",
"性化推荐[4,36]、问答[11,16]。带有查询（LeBron，lives_in，？",
"connections can be preserved with linear complexity.Second",
"和答案LosAngeles的KG推理示例如图E1所示。从符号上讲，",
"we design a learning-based sampling distribution to identify the",
"推理问题可以表示为查询（eg,rg,？），其中查询实体eg和查询",
"semanticallyrelated entities.Extensiveexperiments showthatour",
"关系rg。KG推理的目的是通过从给定KG中学习的本地证据",
"method is powerful.efficient and semantic-aware.The code is",
"available at https://github.com/LARS-research/AdaProp",
"来找到（eg,rg,？)的目标答案实体ea",
"用于KG推理的代表性方法可以分为三类：（i)基于三元组的",
"模型[3,9,35,45,48]通过学习的实体和关系嵌人直接对每个答",
"CCS CONCEPTS",
"案实体进行评分；（ii基于路径的方法[7,8,28,29,47]学习逻辑",
".Computing methodologies →Knowledge representation",
"规则，以从eg开始生成关系路径，并通过基于MAMTHXXIGX5",
"and reasoning.",
"的实体传播方法探索哪个实体更可能是目标答案。在这三个",
"类别中，基于GNN的方法实现了最先进的性能[42，46,49】。",
"KEYWORDS",
"具体来说，R-GCN[30]和CompGCN[34]在所有实体的整个",
"Knowledge graph, Graph embedding, Knowledge graph reasoning",
"邻域中传播。与一般GCN相同，它们很容易遭受过度平滑问",
"Graph sampling, Graph neural network",
"题[26]。NBFNet[49]和RED-GNN[46]在查询实体eg周围的",
"本地邻域中传播。通过选择传播范围，它们具有比R-GCN或",
"-This work was partially performed when Z. Zhou was an intern in 4Paradigm.",
"CompGCN更强的推理能力。但是，它们的本地邻域主要依赖",
"Y.Zhang and z.Zhou have equal contribution, and correspondence is to Q. Yao.",
"于查询实体，并且大多数实体将在更大的传播步骤中涉及",
"现有的基于GNN的方法在所有实体上或围绕查询实体的所",
"有邻域之间传播，忽略了局部邻域与查询关系之间的语义相",
"Permission to make digital or hard copies of allor part of this work for personal or",
"classroom use is granted without fee provided that copies are not made or distributed",
"关性。直观地说，我们需要的本地信息取决于查询实体和查询",
"for profit or commercial advantage and that copies bear this notice and the full",
"关系。例如，与图1中的查询最相关的本地证据包括(LeBron，",
"citation on the first page. Copyrights for components of this work owned by others",
"部分_of，湖人队)&(湖人队，位于_in，L。A。）和（勒布朗，萨",
"than the author(s) must be honored. Abstracting with credit is permitted To copy",
"otherwise,or republish, to post on servers or to redistribute to lists, requires prio",
"凡·纳的配偶）&（萨凡纳，住在，L。A。）。如果查询更改为",
"specificpermission and/orafee.Request permissions frompermissions@acm.org.",
"（勒布朗，父亲，？），则最相关的信息将更改为（勒布朗JR，儿",
"KDD23,August 6-10,2023, LongBeach,CA,USA",
"子，勒布朗）。此外，现有的基于GNN的方法不可避免地会涉",
"@ 2023 Copyright held by the owner/author(s).Publication rights licensed to ACM.",
"及太多的无关实体和事实，增加了学习难度和计算成本，尤其",
"ACMISBN979-8-4007-0103-0/23/08..$15.00",
"https://doi.org/10.1145/3580305.3599404",
"是在大规模的KG上"
] |
{
"arxiv_id": "2205.15319",
"language": "zh",
"page_idx": 0
}
|
||
2205.15319-zh-1
|
2205.15319-zh-1.jpg
|
KDD'23, August 6-10, 2023, Long Beach, CA, USA
<FigureHere>
Figure 1: An example of KG reasoning. The query (Lebron, lives\_in,?) asks which city Lebron lives in. Dashed lines are irrelevant facts to this query while solid lines are relevant.
$px$
受上述示例的启发，我们建议在传播过程中自适应地对语义相关的实体进行采样。在设计采样算法时存在三个挑战：(i) KG 的结构复杂且大小较大；(ü) KG 是多关系的，其中边缘和查询具有不同种类的关系，表示不同的语义含义；(iii)没有直接的监督和简单的启发式度量来指示实体与给定查询的语义相关性。为了解决这些挑战，我们提出了 AdaProp,这是一种具有自适应传播路径的基于 GNN 的方法。关键思想是减少涉及的实体数量，同时保留传播路径中的相关实体。这是通过增量采样机制和适应不同查询的基于学习的采样分布来实现的。通过这种方式，我们的方法学习了一种自适应传播路径，该路径保留了逐层连接，选择了语义相关的实体，而减少了有希望的目标。主要贡献总结如下：
$\bullet$ We propose a connection-preserving sampling scheme, called
incremental sampling, which only has linear complexity with regard to the propagation steps and can preserve the layer-wise connections between sampled entities.
$\bullet$ We design a semantic-aware Gumble top- $k$ distribution which adaptively selects local neighborhoods relevant to the query relation and is learned by a straight through estimator.
$\bullet$ Experiments show that the proposed method achieves the stateof-the-art performance in both transductive and inductive KG reasoning settings. It can preserve more target entities than the other samplers. And the case study shows that the learned sampler is query-dependent and semantic-aware.
令$G=(\mathcal{V},\mathcal{R},\mathcal{E},Q)$是 KG 的一个实例，其中$\mathcal{V},\mathcal{R}$是实体和关系的集合，$\mathcal{E} = \{ ( e_s, r, e_0) | e_s, e_o\in \mathcal{V} , r\in \mathcal{R} \}$ 是事实三元组的集合$,Q=\{(e_{q},r_{q},e_{a})|e_{q},e_{a}\in\mathcal{V},r\in\mathcal{R}\}$ 是具有$^1$查询$(e_q,r_{q},?)$ 的查询三元组的集合。推理任务是学习一个函数，该函数根据事实三元组$\mathcal{E}$预测每个查询$(e_q,r_q,?)$的答案$e_a\in\mathcal{V}$
2 RELATED WORKS
2.1 GNN for KG reasoning
由于图形神经网络 (GNNs)$\boxed{\mathbb{E}}[13,20]$在建模图形结构数据方面的成功，最近的作品[33,34,46,49]尝试利用 GNNs 的力量进行
$^{1}$Note that predicting missing head in KG can also be formulated in this way by adding
inverse relations (details in Appendix B.1 ).
Yongqi Zhang, Zhanke Zhou, Quanming Yao, Xiaowen Chu, and Bo Han
KG 推理。它们通常遵循消息传播框架[13,20]在实体之间传播消息。在第$l$个传播步骤中，具体而言，首先基于实体表示$h_e_s^{t-1}$ 和关系表示$h_r^t$,在具有消息函数$m_{(e_s,r,e_o)}^t:=$ MESS$\left(\boldsymbol h_{e_s}^{\ell-1},\boldsymbol{h}_r^\ell\right)$ 的边$\mathcal{E}^\ell=\{(e_s,r,e_0)\in\mathcal{E}|e_s\in\mathcal{V}^{\ell-1}),r\in\mathcal{R},e_o\in\mathcal{V}^\ell\}$上计算消息。然后通过$h_{e_{o}}^{\ell}:=\delta\left(\mathrm{AGG}(m_{(e_{s},r,e_{o})}^{\ell},(e_{s},r,e_{o})\in\mathcal{E}^{\ell})\right)$将消息从上一步$(e_s\in\mathcal{V}^{t-1})$的实体传播到当前步骤$(e_o\in\mathcal{V}^l)$的实体，并具有激活函数$\delta(\cdot)$ 。附录 A.1 中提供了针对不同方法的指定函数 MESS(·) 和 AGG(·)。在$L$步骤传播之后，获得实体表示$h_{e_o}^L$以测量每个实体$e_o\in\mathcal{V}^L$的合理性
对于基于 GNN 的 KG 推理方法，参与传播的实体集是不同的。根据传播范围的设计，现有的基于 GNN 的方法通常可以分为三类：px
$\bullet$ Full propagation methods, e.g., R-GCN[30] and CompGCN[34] ,propagate among all the entities, i.e., $V^t=\mathcal{V}.$ They are limited to small propagation steps due to the large memory cost and the over-smoothing problem of GNNs [26] over full neighbors.
$\bullet$ Progressive propagation methods (RED-GNN[46] and NBFNet[49]
),propagate from the query entity $e_q$ and progressively to the $L$ -hop neighborhood of $e_q$,i.e., $\mathcal{V}^0=\{e_q\}$ and $\mathcal{V}^\ell=\bigcup_{e\in\mathcal{V}^{\ell-1}}\mathcal{N}(e)$ ,where $N(e)$ contains the 1-hop neighbors of entity $e.$
$\bullet$ Constrained propagation methods, e.g., GraIL [33] and CoMPILE[23] , propagate within a constrained range, i.e, $\mathcal{V}^{\ell}=$ $\mathcal{V} _{e_{q, e_{a}}}^{L}$ , where $\mathcal{V} _{e_{q, e_{a}}}^{L}\subset \mathcal{V}$ is the enclosing subgraph for $e_{q}$ and $e_{a}$ .Since $\mathcal{V} _{e_{q, e_{a}}}^{L}$ are different for different pairs of $( e_{q}, e_{a})$ , these methods are extremely expensive, especia scale KGs (see discussion in Appendix A.2 ).
我们在图$\boxed{\mathrm{E}}$2 的中间提供了三个类别的传播方案的图形说明。其中，渐进式传播方法实现了最先进的推理性能。它们比受约束的传播方法有效得多，用于传播的实体依赖于查询实体，过滤掉许多不相关的实体。但是，当$L$ 变大时，渐进式方法的$\mathcal{V}^L$也将涉及大多数甚至所有实体
## 2.2 $\textbf{Sampling methods for GNN}$
采样方法可以在传播过程中控制和选择实体。引人了几种方法来对齐次图上的节点进行采样，目的是提高 GNN 的可伸缩性。它们可以分为三类：(i)节点式采样方法，例如 GraphSAGEEE [14]和 PASS$\boxed{\mathrm{F}}[43]$中的每个局部 XMAowX 子 XMAX\_1; (ii) 逐层采样方法，例如 FastGCN$\boxed{\mathbb{E}}[5]$, Adaptive-SAMATHX[E] XMAXXXXMAXXXXXXXXXXXX。所有这些方法都是为同构图设计的，目的是解决可伸缩性问题
对于 KG 推理，还提出了一些抽样方法。RS-GCN[12]提出了一种节点采样方法来选择 R-GCN/CompGCN 的邻居。DPMPN [40]包含一个完整的传播 GNN 来学习全局特征，并以分层采样的方式修剪另一个 GNN 进行推理。这些采样方法降低了在较大 KG 上传播的内存成本，但与基本模型相比，它们的经验性能并没有获得太多收益
# 3 ТНЕ РROРОSED МЕТНОФ 3.1 Problem formulation
我们将用于 KG 推理的 GNN 模型表示为$F(\mathbf{w},\widehat{G}^L)$ ,模型参数为w。$\widehat{G}^L$在这里称为传播路径，包含每个传播步骤中涉及的实体集，即。e。,$\widehat{G} ^{L}= \{ V^{0}, \mathcal{V} ^{1}, . . . , \mathcal{V} ^{L}\}$ ,其中$\mathcal{V}^{\ell},\ell=1\ldots L$ 是第$\ell$个传播步骤中涉及的实体集。实体的表示从$\mathcal{V}^0$传播到$\widehat{G}^{L}$中的$\mathcal{V}^L$,用于多个步骤
|
[
[
0.0849609375,
0.07574462890625,
0.325927734375,
0.07574462890625,
0.325927734375,
0.0889892578125,
0.0849609375,
0.0889892578125
],
[
0.5703125,
0.07574462890625,
0.91259765625,
0.07574462890625,
0.91259765625,
0.0889892578125,
0.5703125,
0.0889892578125
],
[
0.408447265625,
0.10986328125,
0.44775390625,
0.10986328125,
0.44775390625,
0.117431640625,
0.408447265625,
0.117431640625
],
[
0.51953125,
0.10919189453125,
0.91015625,
0.10919189453125,
0.91015625,
0.11932373046875,
0.51953125,
0.11932373046875
],
[
0.406005859375,
0.11932373046875,
0.453369140625,
0.11932373046875,
0.453369140625,
0.1307373046875,
0.406005859375,
0.1307373046875
],
[
0.134033203125,
0.1256103515625,
0.1405029296875,
0.1256103515625,
0.1405029296875,
0.1287841796875,
0.134033203125,
0.1287841796875
],
[
0.5185546875,
0.121826171875,
0.90869140625,
0.121826171875,
0.90869140625,
0.1351318359375,
0.5185546875,
0.1351318359375
],
[
0.1070556640625,
0.1337890625,
0.1666259765625,
0.1337890625,
0.1666259765625,
0.14453125,
0.1070556640625,
0.14453125
],
[
0.2467041015625,
0.1351318359375,
0.3095703125,
0.1351318359375,
0.3095703125,
0.1458740234375,
0.2467041015625,
0.1458740234375
],
[
0.393798828125,
0.14013671875,
0.451904296875,
0.14013671875,
0.451904296875,
0.1514892578125,
0.393798828125,
0.1514892578125
],
[
0.51318359375,
0.1363525390625,
0.91259765625,
0.1351318359375,
0.91259765625,
0.1558837890625,
0.51318359375,
0.1572265625
],
[
0.1192626953125,
0.1514892578125,
0.158447265625,
0.1514892578125,
0.158447265625,
0.1634521484375,
0.1192626953125,
0.1634521484375
],
[
0.2099609375,
0.1483154296875,
0.21728515625,
0.1483154296875,
0.21728515625,
0.1514892578125,
0.2099609375,
0.1514892578125
],
[
0.5185546875,
0.1546630859375,
0.91259765625,
0.1546630859375,
0.91259765625,
0.169189453125,
0.5185546875,
0.169189453125
],
[
0.1732177734375,
0.1673583984375,
0.1854248046875,
0.1673583984375,
0.1854248046875,
0.173583984375,
0.1732177734375,
0.173583984375
],
[
0.39453125,
0.1654052734375,
0.40283203125,
0.1654052734375,
0.40283203125,
0.169189453125,
0.39453125,
0.169189453125
],
[
0.406005859375,
0.169189453125,
0.414306640625,
0.17236328125,
0.410888671875,
0.175537109375,
0.40283203125,
0.17236328125
],
[
0.51708984375,
0.17041015625,
0.9140625,
0.17041015625,
0.9140625,
0.1881103515625,
0.51708984375,
0.1881103515625
],
[
0.1396484375,
0.179931640625,
0.146240234375,
0.179931640625,
0.146240234375,
0.1875,
0.1396484375,
0.1875
],
[
0.1854248046875,
0.18310546875,
0.2418212890625,
0.18310546875,
0.2418212890625,
0.1944580078125,
0.1854248046875,
0.1944580078125
],
[
0.39208984375,
0.177978515625,
0.44775390625,
0.177978515625,
0.44775390625,
0.1881103515625,
0.39208984375,
0.1881103515625
],
[
0.23779296875,
0.186279296875,
0.26708984375,
0.189453125,
0.265625,
0.1962890625,
0.236083984375,
0.1932373046875
],
[
0.26953125,
0.186279296875,
0.273681640625,
0.186279296875,
0.273681640625,
0.1925048828125,
0.26953125,
0.1925048828125
],
[
0.521484375,
0.1875,
0.91259765625,
0.1875,
0.91259765625,
0.2008056640625,
0.521484375,
0.2008056640625
],
[
0.358642578125,
0.1956787109375,
0.40185546875,
0.1900634765625,
0.40283203125,
0.1976318359375,
0.359375,
0.203857421875
],
[
0.5185546875,
0.2020263671875,
0.91259765625,
0.2020263671875,
0.91259765625,
0.21533203125,
0.5185546875,
0.21533203125
],
[
0.136474609375,
0.1900634765625,
0.1478271484375,
0.1900634765625,
0.1478271484375,
0.2008056640625,
0.136474609375,
0.2008056640625
],
[
0.12091064453125,
0.2052001953125,
0.16259765625,
0.2083740234375,
0.160888671875,
0.2197265625,
0.1192626953125,
0.216552734375
],
[
0.193603515625,
0.20703125,
0.236083984375,
0.20703125,
0.236083984375,
0.2183837890625,
0.193603515625,
0.2183837890625
],
[
0.37744140625,
0.2083740234375,
0.393798828125,
0.2083740234375,
0.393798828125,
0.216552734375,
0.37744140625,
0.216552734375
],
[
0.1396484375,
0.2215576171875,
0.15283203125,
0.2215576171875,
0.15283203125,
0.2298583984375,
0.1396484375,
0.2298583984375
],
[
0.365234375,
0.2215576171875,
0.374267578125,
0.2183837890625,
0.378173828125,
0.2247314453125,
0.369384765625,
0.2279052734375
],
[
0.51953125,
0.216552734375,
0.9111328125,
0.216552734375,
0.9111328125,
0.2266845703125,
0.51953125,
0.2266845703125
],
[
0.351318359375,
0.2266845703125,
0.361083984375,
0.2215576171875,
0.365234375,
0.2259521484375,
0.35546875,
0.2310791015625
],
[
0.51708984375,
0.2291259765625,
0.80126953125,
0.2291259765625,
0.80126953125,
0.242431640625,
0.51708984375,
0.242431640625
],
[
0.404296875,
0.236083984375,
0.451904296875,
0.236083984375,
0.451904296875,
0.2474365234375,
0.404296875,
0.2474365234375
],
[
0.53515625,
0.242431640625,
0.91259765625,
0.242431640625,
0.91259765625,
0.255615234375,
0.53515625,
0.255615234375
],
[
0.11517333984375,
0.25390625,
0.178955078125,
0.251953125,
0.178955078125,
0.26318359375,
0.11517333984375,
0.26513671875
],
[
0.256591796875,
0.248779296875,
0.29736328125,
0.248779296875,
0.29736328125,
0.2587890625,
0.256591796875,
0.2587890625
],
[
0.51953125,
0.25830078125,
0.9111328125,
0.25830078125,
0.9111328125,
0.268310546875,
0.51953125,
0.268310546875
],
[
0.32275390625,
0.26953125,
0.375732421875,
0.26953125,
0.375732421875,
0.27978515625,
0.32275390625,
0.27978515625
],
[
0.5185546875,
0.271484375,
0.609375,
0.272705078125,
0.609375,
0.2841796875,
0.5185546875,
0.28271484375
],
[
0.51953125,
0.287353515625,
0.91259765625,
0.287353515625,
0.91259765625,
0.300537109375,
0.51953125,
0.300537109375
],
[
0.11767578125,
0.274658203125,
0.1641845703125,
0.274658203125,
0.1641845703125,
0.290283203125,
0.11767578125,
0.290283203125
],
[
0.25,
0.29345703125,
0.298095703125,
0.29345703125,
0.298095703125,
0.304931640625,
0.25,
0.304931640625
],
[
0.408447265625,
0.292236328125,
0.449462890625,
0.294189453125,
0.44775390625,
0.30615234375,
0.406982421875,
0.3037109375
],
[
0.275390625,
0.304931640625,
0.291748046875,
0.304931640625,
0.291748046875,
0.314453125,
0.275390625,
0.314453125
],
[
0.53515625,
0.300537109375,
0.9140625,
0.2998046875,
0.9140625,
0.314453125,
0.53515625,
0.31494140625
],
[
0.08660888671875,
0.31640625,
0.4794921875,
0.31640625,
0.4794921875,
0.330078125,
0.08660888671875,
0.330078125
],
[
0.53369140625,
0.31494140625,
0.91259765625,
0.31494140625,
0.91259765625,
0.328857421875,
0.53369140625,
0.328857421875
],
[
0.0849609375,
0.330078125,
0.48046875,
0.330078125,
0.48046875,
0.343505859375,
0.0849609375,
0.343505859375
],
[
0.53173828125,
0.328857421875,
0.88818359375,
0.32763671875,
0.88818359375,
0.34228515625,
0.53173828125,
0.343505859375
],
[
0.0849609375,
0.3447265625,
0.4755859375,
0.3447265625,
0.4755859375,
0.35791015625,
0.0849609375,
0.35791015625
],
[
0.5185546875,
0.34228515625,
0.9306640625,
0.343505859375,
0.9306640625,
0.356689453125,
0.5185546875,
0.35595703125
],
[
0.273681640625,
0.361083984375,
0.291748046875,
0.361083984375,
0.291748046875,
0.37060546875,
0.273681640625,
0.37060546875
],
[
0.53369140625,
0.35791015625,
0.91259765625,
0.35791015625,
0.91259765625,
0.371826171875,
0.53369140625,
0.371826171875
],
[
0.53173828125,
0.37060546875,
0.9140625,
0.371826171875,
0.9140625,
0.386474609375,
0.53173828125,
0.385009765625
],
[
0.10296630859375,
0.380126953125,
0.48046875,
0.380126953125,
0.48046875,
0.393310546875,
0.10296630859375,
0.393310546875
],
[
0.53125,
0.3837890625,
0.87109375,
0.3837890625,
0.87109375,
0.397705078125,
0.53125,
0.397705078125
],
[
0.08905029296875,
0.395751953125,
0.48046875,
0.395751953125,
0.48046875,
0.406005859375,
0.08905029296875,
0.406005859375
],
[
0.521484375,
0.40087890625,
0.9140625,
0.40087890625,
0.9140625,
0.4140625,
0.521484375,
0.4140625
],
[
0.0849609375,
0.40771484375,
0.48046875,
0.40771484375,
0.48046875,
0.421630859375,
0.0849609375,
0.421630859375
],
[
0.53369140625,
0.41357421875,
0.91259765625,
0.41357421875,
0.91259765625,
0.427978515625,
0.53369140625,
0.427978515625
],
[
0.08660888671875,
0.421630859375,
0.48046875,
0.421630859375,
0.48046875,
0.43505859375,
0.08660888671875,
0.43505859375
],
[
0.53369140625,
0.4267578125,
0.9140625,
0.4267578125,
0.9140625,
0.4443359375,
0.53369140625,
0.4443359375
],
[
0.08660888671875,
0.4375,
0.4794921875,
0.4375,
0.4794921875,
0.447509765625,
0.08660888671875,
0.447509765625
],
[
0.53125,
0.44140625,
0.91259765625,
0.44140625,
0.91259765625,
0.458984375,
0.53125,
0.458984375
],
[
0.08660888671875,
0.449462890625,
0.48046875,
0.449462890625,
0.48046875,
0.46337890625,
0.08660888671875,
0.46337890625
],
[
0.53515625,
0.45703125,
0.91259765625,
0.458251953125,
0.91259765625,
0.4716796875,
0.53515625,
0.47021484375
],
[
0.08660888671875,
0.46337890625,
0.48046875,
0.46337890625,
0.48046875,
0.4765625,
0.08660888671875,
0.4765625
],
[
0.53369140625,
0.4716796875,
0.7998046875,
0.4716796875,
0.7998046875,
0.48486328125,
0.53369140625,
0.48486328125
],
[
0.0882568359375,
0.477783203125,
0.48046875,
0.477783203125,
0.48046875,
0.4912109375,
0.0882568359375,
0.4912109375
],
[
0.0882568359375,
0.4912109375,
0.4794921875,
0.4912109375,
0.4794921875,
0.5048828125,
0.0882568359375,
0.5048828125
],
[
0.53515625,
0.4873046875,
0.91259765625,
0.4873046875,
0.91259765625,
0.50048828125,
0.53515625,
0.50048828125
],
[
0.08660888671875,
0.5048828125,
0.48046875,
0.5048828125,
0.48046875,
0.51806640625,
0.08660888671875,
0.51806640625
],
[
0.5185546875,
0.501953125,
0.91259765625,
0.501953125,
0.91259765625,
0.51513671875,
0.5185546875,
0.51513671875
],
[
0.08660888671875,
0.51806640625,
0.48046875,
0.51806640625,
0.48046875,
0.5322265625,
0.08660888671875,
0.5322265625
],
[
0.5185546875,
0.51513671875,
0.91259765625,
0.51513671875,
0.91259765625,
0.52880859375,
0.5185546875,
0.52880859375
],
[
0.0849609375,
0.53271484375,
0.26953125,
0.53271484375,
0.26953125,
0.546875,
0.0849609375,
0.546875
],
[
0.5185546875,
0.52978515625,
0.91259765625,
0.52978515625,
0.91259765625,
0.54345703125,
0.5185546875,
0.54345703125
],
[
0.51708984375,
0.54248046875,
0.7685546875,
0.54248046875,
0.7685546875,
0.556640625,
0.51708984375,
0.556640625
],
[
0.09234619140625,
0.5498046875,
0.48193359375,
0.5498046875,
0.48193359375,
0.56298828125,
0.09234619140625,
0.56298828125
],
[
0.10455322265625,
0.56298828125,
0.48193359375,
0.56298828125,
0.48193359375,
0.5771484375,
0.10455322265625,
0.5771484375
],
[
0.5185546875,
0.57470703125,
0.56201171875,
0.57470703125,
0.56201171875,
0.5859375,
0.5185546875,
0.5859375
],
[
0.560546875,
0.57373046875,
0.7890625,
0.57373046875,
0.7890625,
0.5869140625,
0.560546875,
0.5869140625
],
[
0.51953125,
0.59130859375,
0.91259765625,
0.59130859375,
0.91259765625,
0.60498046875,
0.51953125,
0.60498046875
],
[
0.10540771484375,
0.57763671875,
0.48046875,
0.57763671875,
0.48046875,
0.59130859375,
0.10540771484375,
0.59130859375
],
[
0.10540771484375,
0.59228515625,
0.339111328125,
0.59228515625,
0.339111328125,
0.60302734375,
0.10540771484375,
0.60302734375
],
[
0.09234619140625,
0.60498046875,
0.48046875,
0.60498046875,
0.48046875,
0.61865234375,
0.09234619140625,
0.61865234375
],
[
0.51953125,
0.60498046875,
0.91259765625,
0.60498046875,
0.91259765625,
0.61865234375,
0.51953125,
0.61865234375
],
[
0.10455322265625,
0.61865234375,
0.48046875,
0.61865234375,
0.48046875,
0.6318359375,
0.10455322265625,
0.6318359375
],
[
0.521484375,
0.62060546875,
0.9111328125,
0.62060546875,
0.9111328125,
0.630859375,
0.521484375,
0.630859375
],
[
0.10455322265625,
0.63330078125,
0.431396484375,
0.63330078125,
0.431396484375,
0.64306640625,
0.10455322265625,
0.64306640625
],
[
0.51953125,
0.63330078125,
0.9140625,
0.63330078125,
0.9140625,
0.646484375,
0.51953125,
0.646484375
],
[
0.09234619140625,
0.646484375,
0.483642578125,
0.646484375,
0.483642578125,
0.66015625,
0.09234619140625,
0.66015625
],
[
0.5185546875,
0.646484375,
0.91259765625,
0.646484375,
0.91259765625,
0.66015625,
0.5185546875,
0.66015625
],
[
0.10455322265625,
0.66015625,
0.48193359375,
0.66015625,
0.48193359375,
0.673828125,
0.10455322265625,
0.673828125
],
[
0.515625,
0.66015625,
0.9140625,
0.66015625,
0.9140625,
0.673828125,
0.515625,
0.673828125
],
[
0.10540771484375,
0.67529296875,
0.48046875,
0.67529296875,
0.48046875,
0.68603515625,
0.10540771484375,
0.68603515625
],
[
0.5185546875,
0.6748046875,
0.7548828125,
0.6748046875,
0.7548828125,
0.68798828125,
0.5185546875,
0.68798828125
],
[
0.10540771484375,
0.6875,
0.48046875,
0.6875,
0.48046875,
0.70068359375,
0.10540771484375,
0.70068359375
],
[
0.53515625,
0.68798828125,
0.91015625,
0.68798828125,
0.91015625,
0.7021484375,
0.53515625,
0.7021484375
],
[
0.10540771484375,
0.7021484375,
0.40185546875,
0.7021484375,
0.40185546875,
0.71533203125,
0.10540771484375,
0.71533203125
],
[
0.53173828125,
0.7021484375,
0.91259765625,
0.7021484375,
0.91259765625,
0.71533203125,
0.53173828125,
0.71533203125
],
[
0.10540771484375,
0.7197265625,
0.4794921875,
0.7197265625,
0.4794921875,
0.72998046875,
0.10540771484375,
0.72998046875
],
[
0.5185546875,
0.71533203125,
0.91259765625,
0.71533203125,
0.91259765625,
0.72900390625,
0.5185546875,
0.72900390625
],
[
0.08660888671875,
0.73193359375,
0.48046875,
0.73193359375,
0.48046875,
0.74560546875,
0.08660888671875,
0.74560546875
],
[
0.51953125,
0.73095703125,
0.9111328125,
0.73095703125,
0.9111328125,
0.7412109375,
0.51953125,
0.7412109375
],
[
0.08660888671875,
0.74560546875,
0.4794921875,
0.74560546875,
0.4794921875,
0.76025390625,
0.08660888671875,
0.76025390625
],
[
0.5185546875,
0.74365234375,
0.91259765625,
0.74365234375,
0.91259765625,
0.7568359375,
0.5185546875,
0.7568359375
],
[
0.08660888671875,
0.76025390625,
0.48046875,
0.76025390625,
0.48046875,
0.7734375,
0.08660888671875,
0.7734375
],
[
0.5185546875,
0.755859375,
0.68115234375,
0.755859375,
0.68115234375,
0.76953125,
0.5185546875,
0.76953125
],
[
0.08660888671875,
0.7734375,
0.420654296875,
0.7734375,
0.420654296875,
0.78857421875,
0.08660888671875,
0.78857421875
],
[
0.521484375,
0.7861328125,
0.5361328125,
0.79248046875,
0.529296875,
0.8017578125,
0.5146484375,
0.79541015625
],
[
0.5361328125,
0.79150390625,
0.54345703125,
0.79150390625,
0.54345703125,
0.796875,
0.5361328125,
0.796875
],
[
0.54736328125,
0.787109375,
0.7685546875,
0.787109375,
0.7685546875,
0.80029296875,
0.54736328125,
0.80029296875
],
[
0.0849609375,
0.80029296875,
0.1168212890625,
0.80029296875,
0.1168212890625,
0.8125,
0.0849609375,
0.8125
],
[
0.1094970703125,
0.7998046875,
0.27294921875,
0.7998046875,
0.27294921875,
0.81298828125,
0.1094970703125,
0.81298828125
],
[
0.515625,
0.80615234375,
0.73779296875,
0.806640625,
0.73779296875,
0.8212890625,
0.515625,
0.82080078125
],
[
0.08660888671875,
0.82080078125,
0.13232421875,
0.82080078125,
0.13232421875,
0.83203125,
0.08660888671875,
0.83203125
],
[
0.125,
0.818359375,
0.316162109375,
0.8193359375,
0.316162109375,
0.833984375,
0.125,
0.83349609375
],
[
0.5185546875,
0.82470703125,
0.91259765625,
0.82470703125,
0.91259765625,
0.83837890625,
0.5185546875,
0.83837890625
],
[
0.08905029296875,
0.8388671875,
0.48046875,
0.8388671875,
0.48046875,
0.849609375,
0.08905029296875,
0.849609375
],
[
0.5185546875,
0.8388671875,
0.91259765625,
0.8388671875,
0.91259765625,
0.85302734375,
0.5185546875,
0.85302734375
],
[
0.0849609375,
0.8515625,
0.48046875,
0.8515625,
0.48046875,
0.86572265625,
0.0849609375,
0.86572265625
],
[
0.5185546875,
0.85400390625,
0.91259765625,
0.85400390625,
0.91259765625,
0.8671875,
0.5185546875,
0.8671875
],
[
0.0849609375,
0.87060546875,
0.48193359375,
0.87255859375,
0.48193359375,
0.88623046875,
0.0849609375,
0.8837890625
],
[
0.51953125,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.51953125,
0.8818359375
],
[
0.08416748046875,
0.8837890625,
0.2900390625,
0.8837890625,
0.2900390625,
0.89794921875,
0.08416748046875,
0.89794921875
],
[
0.515625,
0.880859375,
0.705078125,
0.8818359375,
0.705078125,
0.896484375,
0.515625,
0.89599609375
]
] |
[
"KDD 23, August 6-10, 2023, Long Beach, CA, USA",
"Yongqi Zhang, Zhanke Zhou, Quanming Yao, Xiaowen Chu, and Bo Han",
"",
"KG推理。它们通常遵循消息传播框架「13.20]在实体之间传播",
"OO",
"",
"消息。在第1个传播步骤中，具体而言，首先基于实体表示hc-",
"LeBron.JR",
"Savan Nah",
"located in",
"和关系表示h，在具有消息函数mesr,ea）=MESs(nch)",
"sonlof",
"",
"的边 l = [(es, r,eo) e les et-1),r R,e。 V) 上计算消",
"S",
"",
"",
"息。然后通过he。=(AGG(m(es,r,ea),(es,r,eo)e8t)将消息从",
"",
"lives_in",
"LOSANGELE",
"",
"",
"上一步（esEvt-1）的实体传播到当前步骤（eEt）的实",
"inz",
"体，并具有激活函数8(）。附录A.1中提供了针对不同方法的",
"福",
"LeBron",
"part_of",
"",
"",
"C",
"指定函数MESS(）和AGG(）。在L步骤传播之后，获得实体",
"ix",
"表示he。以测量每个实体e。EVL的合理性",
"lives_in",
"对于基于GNN的KG推理方法，参与传播的实体集是不同",
"sporisor_of",
"partLof",
"的。根据传播范围的设计，现有的基于GNN的方法通常可以",
"spouse_of",
"分为三类：px",
". Full propagation methods, e.g., R-GCN [30] and CompGCN [34]",
"NIKE",
"A.Davis",
"Britney",
"px",
", propagate among all the entities, i.e.,=.They are limited",
"Figure 1: An example of KG reasoning. The query (Lebron,",
"to small propagation steps due to the large memory cost and the",
"lives_in,?) asks which city Lebron lives in. Dashed lines are",
"over-smoothing problem of GNNs [26] over full neighbors.",
"irrelevant facts to this query while solid lines are relevant.",
".Progressive propagation methods (RED-GNN[46] and NBFNet[49]",
"px",
"), propagate from the query entity eq and progressively to the L",
"-hop neighborhood of eg,ie., v'=(eq) and '=Ueeve--N(e)",
"受上述示例的启发，我们建议在传播过程中自适应地对语",
", where N(e) contains the 1-hop neighbors of entity e.",
"义相关的实体进行采样。在设计采样算法时存在三个挑战：（i）",
": Constrained propagation methods, e.g., GralL [33] and CoM-",
"KG的结构复杂且大小较大；（ii）KG是多关系的，其中边缘和",
"PILE [23] . propagate within a constrained range, ie., yt -",
"查询具有不同种类的关系，表示不同的语义含义；（i）没有直",
"Vegea , where Veg,ea c is the enclosing subgraph for eq",
"接的监督和简单的启发式度量来指示实体与给定查询的语义",
"and ea. Since Veg.ea are different for different pairs of (eq,ea)",
"相关性。为了解决这些挑战，我们提出了AdaProp，这是一种",
", these methods are extremely expensive, especially on large-",
"具有自适应传播路径的基于GNN的方法。关键思想是减少涉",
"scale KGs (see discussion in Appendix A.2 ).",
"及的实体数量，同时保留传播路径中的相关实体。这是通过增",
"量采样机制和适应不同查询的基于学习的采样分布来实现的。",
"我们在图E2的中间提供了三个类别的传播方案的图形说",
"通过这种方式，我们的方法学习了一种自适应传播路径，该路",
"明。其中，渐进式传播方法实现了最先进的推理性能。它们比",
"径保留了逐层连接，选择了语义相关的实体，而减少了有希望",
"受约束的传播方法有效得多，用于传播的实体依赖于查询实",
"的目标。主要贡献总结如下：",
"体，过滤掉许多不相关的实体。但是，当L变大时，渐进式方",
"法的也将涉及大多数甚至所有实体",
". We propose a connection-preserving sampling scheme, called",
"incremental sampling, which only has linear complexity with",
"2.2",
"Sampling methods for GNN",
"采样方法可以在传播过程中控制和选择实体。引人了几种方法",
"regard to the propagation steps and can preserve the layer-wise",
"connectionsbetween sampled entities.",
" We design a semantic-aware Gumble top- k distribution which",
"来对齐次图上的节点进行采样，目的是提高GNN的可伸缩性。",
"adaptively selects local neighborhoods relevant to the query",
"它们可以分为三类：i）节点式采样方法，例如GraphSAGEE",
"relation and is learned byastraightthrough estimator",
"[14]和PASSE[43]中的每个局部XMAowX子XMAX_1；（i)",
".Experiments show that the proposed method achieves the state-",
"逐层采样方法，例如FastGCNE[5]，Adaptive-SAMATHXE",
"of-the-art performance in both transductive and inductive KG",
"XMAXXXXMAXXXXXXXXXXXX。所有这些方法都是为同构",
"reasoning settings. It can preserve more target entities than",
"图设计的，目的是解决可伸缩性问题",
"the other samplers. And the case study shows that the learned",
"对于KG推理，还提出了一些抽样方法。RS-GCN[12]提出了",
"sampler is query-dependent and semantic-aware.",
"种节点采样方法来选择R-GCN/CompGCN的邻居。DPMPN",
"令G=VR8,Q）是KG的一个实例，其中V，R是实体",
"[40]包含一个完整的传播GNN来学习全局特征，并以分层采",
"和关系的集合，&={(es,reo)les,eoEV,reR）是事实三元组的",
"样的方式修剪另一个GNN进行推理。这些采样方法降低了在",
"集合，Q=((e,rq,ea)leq,eaeV,reR）是具有1查询(eg,rg?",
"较大KG上传播的内存成本，但与基本模型相比，它们的经验",
"的查询三元组的集合。推理任务是学习一个函数，该函数根据",
"性能并没有获得太多收益",
"事实三元组8预测每个查询（eg,rg.？)的答案eaE",
"3",
"",
"THEPROPOSED METHOD",
"2",
"RELATED WORKS",
"3.1Problemformulation",
"2.1",
"GNN forKGreasoning",
"我们将用于KG推理的GNN模型表示为F(w,GL），模型参数",
"由于图形神经网络（GNNs)E[13,20]在建模图形结构数据方面",
"为W。GL在这里称为传播路径，包含每个传播步骤中涉及的",
"的成功，最近的作品[33,34,46,49]尝试利用GNNs的力量进行",
"实体集，即。e。，GL=(o,Vl,..VL)，其中Vt,l=1..L",
"INote that predicting missing head in KG can also be formulated in this way by adding",
"是第个传播步骤中涉及的实体集。实体的表示从Vo传播到",
"inverse relations (details in Appendix B.1 ).",
"GL中的VL，用于多个步骤"
] |
{
"arxiv_id": "2205.15319",
"language": "zh",
"page_idx": 1
}
|
||
2205.15319-zh-3
|
2205.15319-zh-3.jpg
|
KDD'23, August 6-10, 2023, Long Beach, CA, USA
线性增长 w。r。t。传播步骤$L$ 。其次，可以基于 Prop 保留逐层连接。2 , 它显示了优于逐层采样方法的优势。第三，如$\overline{\mathrm{E}}$ 4.2.1节所示，与具有相同数量的涉及实体的其他采样机制相比，保留目标答案实体的可能性更大
PROPOSITION 1. The number of involved entities in the propaga- $tionpath(\left|\bigcup_{f=0...L}\mathcal{V}_{e_{q},r_{q}}^{\ell}\right|)ofincrementalsamplingisboundedby$ $O( LK)$ .
PROPOSITION $2.Foralltheentitiese\in\mathcal{V}_{eq,r_q}^{\ell-1}withincremental$ $sampling, there$ exists $at$ least one entity $e^{\prime }$ $\in$ $V_{eq, r_q}^{t}$ and relation $r\in \mathcal{R}$ such that $( e, r, e^{\prime }) \in \mathcal{E}$ .
3.3 Learning semantic-aware distribution
增量采样机制缓解了减少目标候选项的问题。但是，从$\overline{v}_{e_{q,r_q}}^t$ 中随机抽取$K$实体，并没有考虑不同实体对不同查询的语义相关性。像波束搜索[38] 这样的基于贪婪的搜索算法在这里不适用，因为没有直接的启发式方法来测量语义相关性。为了解决这个问题，我们为每个传播步骤$\ell$,i用参数 $\theta^\ell$ 参数化采样器$S(\cdot)$。e。, $V_{eq,r_{q}}^{t}=S(V_{eq,r_{q}}^{t-1};\theta^{t})$。在以下两部分中，我们介绍了采样分布以及如何优化采样参数
3.3.1 Parameterized sampling distribution .对于基于 GNN 的KG 推理方法，最终传播步骤中的实体表示$h_{e_o}^L$用于测量实体
$e_{o}$作为目标答案实体的合理性。换句话说，学习实体表示以指示它们与给定查询$(e_q,r_q,?)$的相关性。为了分享这些知识， 我们在每个层$\ell=1\ldots L$中引人一个带有参数$\theta^\ell\in\mathbb{R}^d$的线性映射函数$g(\boldsymbol{h}_{e_o}^\ell;\boldsymbol{\theta}^\ell)=\boldsymbol{\theta}^{\ell^\top}\boldsymbol{h}_{e_o}^\ell$,并根据概率分布
$$p^{\ell}(e):=\exp\left(g(\boldsymbol{h}_{e}^{\ell};\boldsymbol{\theta}^{\ell})/\tau\right)\Big/\sum_{e^{\prime}\in\overline{V}_{eq,rq}}\exp\left(g(\boldsymbol{h}_{e^{\prime}}^{\ell};\boldsymbol{\theta}^{\ell})/\tau\right),$$
进行采样，温度值$\tau>0$
在不从$p^\ell(e)$替换的情况下对$K$实体进行采样需要顺序采样并更新每个样本的分布，这可能很昂贵。我们遵循 Gumbeltopk 技巧 [22,39] 来解决这个问题。Gumbel-trick 首先从均匀分布中采样$|\overline{V}_e_q,r_q^t|$ 个独立噪声，即。e。$,U_e\sim$Uniform(0,1) 为每$\uparrow e\in\overline{V}_{eq,rq}^{\ell}$形成Gumbel日志$G_e= g( \boldsymbol{h}_e^{\ell }; \boldsymbol{\theta }^{\ell }) - \log ( - \log U_e)$ 。然后根据 Gumbellogits $G_e$的值收集$\overline{V}_{e_q,r_q}^{\iota}$中的顶级$K$实体正如在[22,39] 中证明的那样，此过程等效于$p^\ell(e)$中没有替换的样本
3.3.2 Learning strategy.我们将$\boldsymbol\theta=\{\boldsymbol{\theta}^1,\ldots,\boldsymbol{\theta}^L\}$表示为采样参数，将参数化的传播路径表示为$\widehat{G}_{e_q,r_q}^L(\theta)$。由于我们在(2) 中的采样分布与 GNN 表示密切相关，因此联合优化模型参数和采样器参数可以更好地在两个部分之间共享知识。具体来说，我们设计了以下目标
$\cdot$ (3)
$$w^{*},\theta^{*}=\arg\min_{w,\theta}\sum_{(e_{q},r_{q},e_{a})\in Q_{\mathrm{tra}}}\mathcal{L}\big(F(w,\widehat{G}_{e_{q},r_{q}}^{L}(\theta)),e_{a}\big).$$
每个实例上的损失函数$\mathcal{L}$ 是 所 有 实 体 $e_o\in \mathcal{V} _{eq, ra}^L$上的二进制
交叉熵损失，i。e。
$$\mathcal{L}(F(\boldsymbol{w},\widehat{G}^{L}(\boldsymbol{\theta})),e_{a})=-\sum_{e_{o}\in V_{eq,ra}^{L}}y_{e_{o}}\log(\phi_{e_{o}})+(1-y_{e_{o}})\log(1-\phi_{e_{o}}),$$
,其中实体$e_o$的似然分数$\phi_{e_o}=f(\boldsymbol{h}_{e_o}^L;\boldsymbol{w}_\phi^\top)\in[0,1]$表示$e_o$作为目标答案实体的合理性，如果$e_o=e_a$否则为0，则标签$y_{e_o}=1$
重新参数化技巧通常用于 Gumbel 分布 [17]。但是，它不做显式采样，因此仍然需要很高的计算成本。相反 ,我们选择直通(ST)估计[2,17],它可以近似估计离散变量的梯度。ST 估计器
Yongqi Zhang, Zhanke Zhou, Quanming Yao, Xiaowen Chu, and Bo Han
<table>
<tbody>
<tr>
<td>$\overline{\mathrm{Alg}}$</td>
<td>orithm AdaProp adaptıve e propagation path</td>
</tr>
<tr>
<td>Req 1: 2:</td>
<td>${\mathrm{uire:~query~}(e_{q},r_{q},?)\:,\:\mathcal{V}_{e_{q},r_{q}}^{0}}$ $_{q}= \{ e_{q}\}$ , steps $L$ , number of sampled entities $K$, and functions MESS(·) and AGG(·) . $\textbf{for}\ell = 1\ldots L$do get the neighboring entities $N(\mathcal{V}_e_{q,r_{c}}^{\ell-1}$ , $) = \cup _{e\in \mathcal{V} _{a}^{t- 1}}$ $\mathcal{N} ( e)$ , $e_{a,r_{a}}$</td>
</tr>
<tr>
<td> </td>
<td>the newly-visited entities $\overline{V}$ and edges ${\mathcal{E} }^{\ell }= \{ ( e_{s}, r, e_{o}) | e_{s}\in$ $a,r$, $1,a$ F $N$</td>
</tr>
<tr>
<td>3:</td>
<td>obtain $m$ $\mathbf{s},r,e_{o}$ $\xi$ $(e_{s},r,e_{0})$ $F$</td>
</tr>
<tr>
<td>4:</td>
<td>obtalr Ilutics $e_{f}$ $\mathcal{N} ( \mathcal{V} _{eq, r_{a}}^{t- 1})$ ; $a.r_{c}$</td>
</tr>
<tr>
<td>5:</td>
<td>logits computation: obtain the Gumbel logits H $T_{\theta_{n}}$ $\log(-\log U_{\mathrm{a}})$ with $U_\mathrm{a}\sim$I Iniform(0,1) for ent</td>
</tr>
<tr>
<td>6:</td>
<td>candidate sampling: obtain sampled entities $\int_{n\boldsymbol{r}\boldsymbol{\sigma}}\tan^{-1}$ 一 $a_{1}$, P</td>
</tr>
<tr>
<td>7:</td>
<td>straight-through: $\boldsymbol{h}_e^{\ell}:=\left(1\mathrm{-no}\_\mathrm{grad}(p^{\ell}(e))+p^{\ell}(e)\right)\cdot\boldsymbol{h}_e^{\ell}$ for entitie $\rho\in V^{r}$</td>
</tr>
<tr>
<td>8:</td>
<td>11 11011215</td>
</tr>
<tr>
<td>10: 1</td>
<td>return $f(\boldsymbol{h}_{e_{\mathrm{o}}}^{L};\boldsymbol{w}^{|})$ for each $\boldsymbol e_o\in$ $\vdots q,r_q$ $px$</td>
</tr>
</tbody>
</table>
的关键思想是在采样信号中反向传播，就好像它是恒等函数一样。具体来说，我们使用$\boldsymbol{h}_e^\ell : = \left ( 1\mathrm{- no\_ grad}( p^\ell ( e) ) + p^\ell ( e) \right ) \cdot \boldsymbol{h}_e^\ell$ ,而不是直接使用实体表示$h_e^\ell$来计算消息，其中 no\_grad$(p^\ell(e))$ 表示反向传播信号不会经过此术语。在这种方法中，前向计算不会受到影响，而后向信号可以通过概率乘法器$p^\ell(e)$ 。估计离散采样分布上的梯度的另一种选择是 REINFORCE 技术$\overline{\mathrm{E}}$ [25,37]。然而，已知 REINFORCE 梯度具有高方差$\overline{\mathbb{E}}[32]$ ,并且方差将在传播步骤中累积，因此这里不使用。Adam 优化器[19]同时更新模型参数$w$和采样器参数$\theta$
总的来说，学习语义感知分布的目的是根据查询关系自适应地保留有前途的目标。如第 4.3 节所示。通过与实体表示共享知识，采样传播路径中的选定信息在语义上与查询关系相关。另一方面，与未学习的版本相比，学习的分布保证了目标实体的更高覆盖率
# 3.4 The full algorithm
AdaProp 的完整过程如算法 1 所示。给定第$(\ell-1)$步中的实体$\mathcal{V} _{eq, r_{q}}^{\ell - 1}$ ,我们首先获得$\mathcal{V}_e_q,r_q^{\ell-1}$的相邻实体$\mathcal{N} ( \mathcal{V} _eq, r_q^{\ell - 1})$ ,并生成新访问的实体$\overline{V}_{e_{q},r_{q}}^{t}=N(\mathcal{V}^{t-1})\setminus\mathcal{V}_{e_{q},r_{q}}^{t-1}$作为行回2 中的候选者。然后，在行$\overline{\mathbb{E}}$3 中的边 $\mathcal{E}^\ell$ 上计算与查询相关的消息并将其传播到行$\overline{\mathrm{E}}$4 中的 N$(\mathcal{V}_e_{q,r_q}^{\ell-1})$ 中的实体。根据在 5 行中计算的 Gumbel 对数，在6行中对$\overline{V}_{e_{q},r_{q}}^{\ell}$中的顶级$K$实体进行采样。至于 ST 估计器，我们修改了7 行中$\widetilde{v}_{e_q,r_q}^t$中实体的隐藏表示。最后，将采样实体集$\widetilde{v}_{e_{q,r_{q}}^{\ell}}$与$V_{e_{q,r_{q}}^{\ell-1}}$连接，以在$\overline{\mathrm{E}}$ 8 行中形成第$\ell$步实体$\mathcal{V}_e_q,r_q^t$集。在传播$L$步骤之后，返回所有涉及实体的最终步骤表示$h_{e_o}^L$。将$v_{e_q,r_q}^L$中$f(\boldsymbol{h}_{e_o}^L;\boldsymbol{w}_\phi)$评估的得分最高的实体作为预测目标
与现有的基于 GNN 的 KG 推理方法相比，关键的区别在于， 在消息传播之前，不再给出和固定算法$\boxed{\mathrm{E}}1$ 中的传播路径。相反，它根据实体表示进行自适应调整，并在每个传播步骤中动态更新。此外，算法 1 还享有比全传播和渐进传播方法更高的
|
[
[
0.0849609375,
0.07574462890625,
0.32275390625,
0.07574462890625,
0.32275390625,
0.08587646484375,
0.0849609375,
0.08587646484375
],
[
0.5703125,
0.07574462890625,
0.91259765625,
0.07574462890625,
0.91259765625,
0.0889892578125,
0.5703125,
0.0889892578125
],
[
0.0882568359375,
0.10919189453125,
0.4794921875,
0.10919189453125,
0.4794921875,
0.11932373046875,
0.0882568359375,
0.11932373046875
],
[
0.51708984375,
0.10479736328125,
0.87744140625,
0.1060791015625,
0.87744140625,
0.12060546875,
0.51708984375,
0.11932373046875
],
[
0.08660888671875,
0.121826171875,
0.48046875,
0.121826171875,
0.48046875,
0.1351318359375,
0.08660888671875,
0.1351318359375
],
[
0.515625,
0.125,
0.9140625,
0.125,
0.9140625,
0.1427001953125,
0.515625,
0.1427001953125
],
[
0.08660888671875,
0.1351318359375,
0.48046875,
0.1351318359375,
0.48046875,
0.1483154296875,
0.08660888671875,
0.1483154296875
],
[
0.5390625,
0.1395263671875,
0.8037109375,
0.1395263671875,
0.8037109375,
0.15283203125,
0.5390625,
0.15283203125
],
[
0.08660888671875,
0.149658203125,
0.32177734375,
0.149658203125,
0.32177734375,
0.160400390625,
0.08660888671875,
0.160400390625
],
[
0.52392578125,
0.1514892578125,
0.64208984375,
0.1514892578125,
0.64208984375,
0.1654052734375,
0.52392578125,
0.1654052734375
],
[
0.5498046875,
0.1622314453125,
0.8759765625,
0.1641845703125,
0.8759765625,
0.18310546875,
0.5498046875,
0.18115234375
],
[
0.10296630859375,
0.1685791015625,
0.4794921875,
0.169189453125,
0.4794921875,
0.18310546875,
0.10296630859375,
0.1817626953125
],
[
0.08660888671875,
0.184326171875,
0.48193359375,
0.184326171875,
0.48193359375,
0.1988525390625,
0.08660888671875,
0.1988525390625
],
[
0.55224609375,
0.1817626953125,
0.8896484375,
0.18310546875,
0.8896484375,
0.2008056640625,
0.55224609375,
0.199462890625
],
[
0.08660888671875,
0.1988525390625,
0.134033203125,
0.1988525390625,
0.134033203125,
0.210205078125,
0.08660888671875,
0.210205078125
],
[
0.55224609375,
0.1976318359375,
0.861328125,
0.1976318359375,
0.861328125,
0.2183837890625,
0.55224609375,
0.2183837890625
],
[
0.55126953125,
0.208984375,
0.66015625,
0.2171630859375,
0.65673828125,
0.236083984375,
0.54833984375,
0.2279052734375
],
[
0.72802734375,
0.205810546875,
0.744140625,
0.205810546875,
0.744140625,
0.210205078125,
0.72802734375,
0.210205078125
],
[
0.098876953125,
0.21533203125,
0.484375,
0.216552734375,
0.484375,
0.2342529296875,
0.098876953125,
0.23291015625
],
[
0.525390625,
0.2183837890625,
0.5400390625,
0.2183837890625,
0.5400390625,
0.2266845703125,
0.525390625,
0.2266845703125
],
[
0.66845703125,
0.21533203125,
0.91259765625,
0.21533203125,
0.91259765625,
0.2298583984375,
0.66845703125,
0.2298583984375
],
[
0.08416748046875,
0.2310791015625,
0.48193359375,
0.23291015625,
0.48193359375,
0.250732421875,
0.08416748046875,
0.248779296875
],
[
0.55615234375,
0.231689453125,
0.646484375,
0.231689453125,
0.646484375,
0.24365234375,
0.55615234375,
0.24365234375
],
[
0.525390625,
0.24560546875,
0.53759765625,
0.24560546875,
0.53759765625,
0.255126953125,
0.525390625,
0.255126953125
],
[
0.55224609375,
0.2412109375,
0.9150390625,
0.2412109375,
0.9150390625,
0.261962890625,
0.55224609375,
0.261962890625
],
[
0.08416748046875,
0.2474365234375,
0.26708984375,
0.2474365234375,
0.26708984375,
0.261474609375,
0.08416748046875,
0.261474609375
],
[
0.5556640625,
0.2587890625,
0.62353515625,
0.2587890625,
0.62353515625,
0.2734375,
0.5556640625,
0.2734375
],
[
0.08416748046875,
0.272705078125,
0.438720703125,
0.2734375,
0.438720703125,
0.28857421875,
0.08416748046875,
0.287353515625
],
[
0.525390625,
0.274658203125,
0.5390625,
0.274658203125,
0.5390625,
0.285400390625,
0.525390625,
0.285400390625
],
[
0.5537109375,
0.2734375,
0.90673828125,
0.2734375,
0.90673828125,
0.28857421875,
0.5537109375,
0.28857421875
],
[
0.5498046875,
0.2841796875,
0.90625,
0.28857421875,
0.90625,
0.309326171875,
0.5498046875,
0.304931640625
],
[
0.08660888671875,
0.29345703125,
0.463134765625,
0.29345703125,
0.463134765625,
0.306884765625,
0.08660888671875,
0.306884765625
],
[
0.5498046875,
0.298583984375,
0.8837890625,
0.302978515625,
0.8837890625,
0.324462890625,
0.5498046875,
0.32080078125
],
[
0.0882568359375,
0.309326171875,
0.4794921875,
0.309326171875,
0.4794921875,
0.322509765625,
0.0882568359375,
0.322509765625
],
[
0.525390625,
0.30810546875,
0.5361328125,
0.30810546875,
0.5361328125,
0.317626953125,
0.525390625,
0.317626953125
],
[
0.08660888671875,
0.322509765625,
0.48046875,
0.322509765625,
0.48046875,
0.3359375,
0.08660888671875,
0.3359375
],
[
0.55224609375,
0.318115234375,
0.7236328125,
0.3212890625,
0.7236328125,
0.34033203125,
0.55224609375,
0.337158203125
],
[
0.0882568359375,
0.33837890625,
0.47802734375,
0.33837890625,
0.47802734375,
0.348388671875,
0.0882568359375,
0.348388671875
],
[
0.52685546875,
0.34228515625,
0.5361328125,
0.34228515625,
0.5361328125,
0.349853515625,
0.52685546875,
0.349853515625
],
[
0.5537109375,
0.33837890625,
0.91259765625,
0.33837890625,
0.91259765625,
0.3515625,
0.5537109375,
0.3515625
],
[
0.08660888671875,
0.35107421875,
0.4794921875,
0.35107421875,
0.4794921875,
0.3642578125,
0.08660888671875,
0.3642578125
],
[
0.55224609375,
0.3466796875,
0.6259765625,
0.354248046875,
0.62353515625,
0.371826171875,
0.5498046875,
0.3642578125
],
[
0.08660888671875,
0.365478515625,
0.48046875,
0.365478515625,
0.48046875,
0.380126953125,
0.08660888671875,
0.380126953125
],
[
0.55126953125,
0.365478515625,
0.8798828125,
0.36865234375,
0.8798828125,
0.386474609375,
0.55126953125,
0.38330078125
],
[
0.525390625,
0.371826171875,
0.53759765625,
0.371826171875,
0.53759765625,
0.380126953125,
0.525390625,
0.380126953125
],
[
0.08660888671875,
0.380126953125,
0.338134765625,
0.380126953125,
0.338134765625,
0.393310546875,
0.08660888671875,
0.393310546875
],
[
0.525390625,
0.38330078125,
0.58251953125,
0.38330078125,
0.58251953125,
0.39453125,
0.525390625,
0.39453125
],
[
0.5185546875,
0.39013671875,
0.765625,
0.393310546875,
0.765625,
0.412353515625,
0.5185546875,
0.4091796875
],
[
0.0882568359375,
0.40087890625,
0.4794921875,
0.40087890625,
0.4794921875,
0.4140625,
0.0882568359375,
0.4140625
],
[
0.0849609375,
0.4140625,
0.47802734375,
0.4140625,
0.47802734375,
0.427978515625,
0.0849609375,
0.427978515625
],
[
0.70751953125,
0.412353515625,
0.7236328125,
0.412353515625,
0.7236328125,
0.421630859375,
0.70751953125,
0.421630859375
],
[
0.0882568359375,
0.429931640625,
0.4794921875,
0.429931640625,
0.4794921875,
0.43994140625,
0.0882568359375,
0.43994140625
],
[
0.08660888671875,
0.442626953125,
0.4794921875,
0.442626953125,
0.4794921875,
0.45703125,
0.08660888671875,
0.45703125
],
[
0.5185546875,
0.443115234375,
0.9111328125,
0.443115234375,
0.9111328125,
0.45703125,
0.5185546875,
0.45703125
],
[
0.08660888671875,
0.458251953125,
0.48046875,
0.458251953125,
0.48046875,
0.4716796875,
0.08660888671875,
0.4716796875
],
[
0.5185546875,
0.458251953125,
0.91259765625,
0.458251953125,
0.91259765625,
0.4716796875,
0.5185546875,
0.4716796875
],
[
0.08660888671875,
0.472900390625,
0.373291015625,
0.472900390625,
0.373291015625,
0.486083984375,
0.08660888671875,
0.486083984375
],
[
0.51953125,
0.472900390625,
0.91259765625,
0.472900390625,
0.91259765625,
0.486083984375,
0.51953125,
0.486083984375
],
[
0.09234619140625,
0.490478515625,
0.486083984375,
0.4912109375,
0.486083984375,
0.51953125,
0.09234619140625,
0.51806640625
],
[
0.5185546875,
0.486083984375,
0.91015625,
0.486083984375,
0.91015625,
0.5,
0.5185546875,
0.5
],
[
0.51953125,
0.501953125,
0.91259765625,
0.501953125,
0.91259765625,
0.51220703125,
0.51953125,
0.51220703125
],
[
0.51953125,
0.5146484375,
0.9140625,
0.5146484375,
0.9140625,
0.52783203125,
0.51953125,
0.52783203125
],
[
0.08660888671875,
0.52294921875,
0.238525390625,
0.52294921875,
0.238525390625,
0.5361328125,
0.08660888671875,
0.5361328125
],
[
0.5185546875,
0.52783203125,
0.91259765625,
0.52783203125,
0.91259765625,
0.54150390625,
0.5185546875,
0.54150390625
],
[
0.10296630859375,
0.53515625,
0.4794921875,
0.53515625,
0.4794921875,
0.548828125,
0.10296630859375,
0.548828125
],
[
0.521484375,
0.54345703125,
0.9111328125,
0.54345703125,
0.9111328125,
0.5537109375,
0.521484375,
0.5537109375
],
[
0.0882568359375,
0.5498046875,
0.4794921875,
0.5498046875,
0.4794921875,
0.56298828125,
0.0882568359375,
0.56298828125
],
[
0.5185546875,
0.55615234375,
0.78369140625,
0.55615234375,
0.78369140625,
0.5693359375,
0.5185546875,
0.5693359375
],
[
0.08660888671875,
0.56298828125,
0.48046875,
0.56298828125,
0.48046875,
0.5771484375,
0.08660888671875,
0.5771484375
],
[
0.5361328125,
0.57080078125,
0.9111328125,
0.57080078125,
0.9111328125,
0.58056640625,
0.5361328125,
0.58056640625
],
[
0.0882568359375,
0.5791015625,
0.48046875,
0.5791015625,
0.48046875,
0.59326171875,
0.0882568359375,
0.59326171875
],
[
0.51953125,
0.58349609375,
0.91259765625,
0.58349609375,
0.91259765625,
0.5966796875,
0.51953125,
0.5966796875
],
[
0.0849609375,
0.59326171875,
0.486083984375,
0.59326171875,
0.486083984375,
0.6142578125,
0.0849609375,
0.6142578125
],
[
0.51953125,
0.5966796875,
0.91259765625,
0.5966796875,
0.91259765625,
0.60986328125,
0.51953125,
0.60986328125
],
[
0.0849609375,
0.6123046875,
0.48193359375,
0.6123046875,
0.48193359375,
0.6298828125,
0.0849609375,
0.6298828125
],
[
0.51953125,
0.611328125,
0.9140625,
0.611328125,
0.9140625,
0.625,
0.51953125,
0.625
],
[
0.0849609375,
0.62890625,
0.48046875,
0.6298828125,
0.48046875,
0.64453125,
0.0849609375,
0.64306640625
],
[
0.51953125,
0.626953125,
0.63818359375,
0.626953125,
0.63818359375,
0.63720703125,
0.51953125,
0.63720703125
],
[
0.08660888671875,
0.64599609375,
0.1478271484375,
0.64599609375,
0.1478271484375,
0.6572265625,
0.08660888671875,
0.6572265625
],
[
0.515625,
0.65087890625,
0.71337890625,
0.6494140625,
0.71337890625,
0.6640625,
0.515625,
0.66552734375
],
[
0.08660888671875,
0.66357421875,
0.4794921875,
0.66357421875,
0.4794921875,
0.67822265625,
0.08660888671875,
0.67822265625
],
[
0.53369140625,
0.66748046875,
0.91259765625,
0.66845703125,
0.91259765625,
0.68310546875,
0.53369140625,
0.681640625
],
[
0.0849609375,
0.67529296875,
0.48193359375,
0.6767578125,
0.48193359375,
0.6943359375,
0.0849609375,
0.693359375
],
[
0.5185546875,
0.6845703125,
0.9111328125,
0.6845703125,
0.9111328125,
0.69873046875,
0.5185546875,
0.69873046875
],
[
0.08660888671875,
0.693359375,
0.48046875,
0.693359375,
0.48046875,
0.70703125,
0.08660888671875,
0.70703125
],
[
0.51953125,
0.70068359375,
0.91259765625,
0.70068359375,
0.91259765625,
0.71533203125,
0.51953125,
0.71533203125
],
[
0.0849609375,
0.70703125,
0.48046875,
0.70849609375,
0.48046875,
0.7216796875,
0.0849609375,
0.72021484375
],
[
0.51953125,
0.71630859375,
0.91015625,
0.71630859375,
0.91015625,
0.73095703125,
0.51953125,
0.73095703125
],
[
0.08660888671875,
0.72265625,
0.250732421875,
0.72265625,
0.250732421875,
0.736328125,
0.08660888671875,
0.736328125
],
[
0.51953125,
0.73193359375,
0.9111328125,
0.73193359375,
0.9111328125,
0.74560546875,
0.51953125,
0.74560546875
],
[
0.1168212890625,
0.74072265625,
0.23779296875,
0.7392578125,
0.23779296875,
0.75390625,
0.1168212890625,
0.7548828125
],
[
0.3095703125,
0.7392578125,
0.39453125,
0.74365234375,
0.393798828125,
0.75830078125,
0.30810546875,
0.75390625
],
[
0.378173828125,
0.74365234375,
0.48046875,
0.74365234375,
0.48046875,
0.7548828125,
0.378173828125,
0.7548828125
],
[
0.515625,
0.7470703125,
0.9150390625,
0.7470703125,
0.9150390625,
0.767578125,
0.515625,
0.767578125
],
[
0.5146484375,
0.76123046875,
0.9140625,
0.76318359375,
0.9140625,
0.78076171875,
0.5146484375,
0.77880859375
],
[
0.1953125,
0.74853515625,
0.318603515625,
0.74853515625,
0.318603515625,
0.76318359375,
0.1953125,
0.76318359375
],
[
0.0849609375,
0.7666015625,
0.48046875,
0.76806640625,
0.48046875,
0.7822265625,
0.0849609375,
0.77978515625
],
[
0.0882568359375,
0.7841796875,
0.2034912109375,
0.7841796875,
0.2034912109375,
0.79443359375,
0.0882568359375,
0.79443359375
],
[
0.5185546875,
0.78076171875,
0.9140625,
0.78076171875,
0.9140625,
0.79541015625,
0.5185546875,
0.79541015625
],
[
0.08660888671875,
0.80029296875,
0.2197265625,
0.80029296875,
0.2197265625,
0.814453125,
0.08660888671875,
0.814453125
],
[
0.273681640625,
0.8017578125,
0.47802734375,
0.8017578125,
0.47802734375,
0.81494140625,
0.273681640625,
0.81494140625
],
[
0.51708984375,
0.796875,
0.91259765625,
0.796875,
0.91259765625,
0.81103515625,
0.51708984375,
0.81103515625
],
[
0.2091064453125,
0.806640625,
0.281005859375,
0.806640625,
0.281005859375,
0.818359375,
0.2091064453125,
0.818359375
],
[
0.5185546875,
0.81103515625,
0.9140625,
0.81103515625,
0.9140625,
0.8291015625,
0.5185546875,
0.8291015625
],
[
0.0906982421875,
0.82568359375,
0.4794921875,
0.82568359375,
0.4794921875,
0.8388671875,
0.0906982421875,
0.8388671875
],
[
0.51953125,
0.8271484375,
0.740234375,
0.8271484375,
0.740234375,
0.84033203125,
0.51953125,
0.84033203125
],
[
0.0882568359375,
0.84130859375,
0.476318359375,
0.84130859375,
0.476318359375,
0.85498046875,
0.0882568359375,
0.85498046875
],
[
0.5361328125,
0.84130859375,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.5361328125,
0.85498046875
],
[
0.10296630859375,
0.85498046875,
0.4794921875,
0.85498046875,
0.4794921875,
0.86865234375,
0.10296630859375,
0.86865234375
],
[
0.5185546875,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.5185546875,
0.86865234375
],
[
0.08905029296875,
0.87060546875,
0.4794921875,
0.87060546875,
0.4794921875,
0.880859375,
0.08905029296875,
0.880859375
],
[
0.5185546875,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.8818359375,
0.5185546875,
0.8818359375
],
[
0.08660888671875,
0.88330078125,
0.48046875,
0.88330078125,
0.48046875,
0.896484375,
0.08660888671875,
0.896484375
],
[
0.5185546875,
0.88330078125,
0.91259765625,
0.88330078125,
0.91259765625,
0.896484375,
0.5185546875,
0.896484375
]
] |
[
"KDD23,August6-10,2023.LongBeach.CA.USA",
"Yongqi Zhang, Zhanke Zhou, Quanming Yao, Xiaowen Chu, and Bo Han",
"线性增长w。r。t。传播步骤L。其次，可以基于Prop保留逐",
"Algorithm 1 AdaProp: learning adaptive propagation path.",
"层连接。2，它显示了优于逐层采样方法的优势。第三，如E",
"Require: query (eq,rq,?),Vegrq=(eq), steps L,number of sampled",
"4.2.1节所示，与具有相同数量的涉及实体的其他采样机制相",
"entities K, and functions MESS() and AGG() .",
"比，保留目标答案实体的可能性更大",
"I: for f = 1...L do",
"get the neighboring entities N(Veiq) =Ueevegi, N(e),",
"PROPOSITION 1.The number of involved entities in the propaga",
"tion path ( [Ue=0... Vegrgl )of incremental sampling is bounded by",
"the newly-visited entities egrg = N(Veglg)Vegi,and",
"O(LK)",
"edges e' = (es r,eo)les e Vegra,eo e N(Vegr,)):",
"obtain m(es.r,eo)",
"",
"PROPOSITION 2. For all the entities e e Ve,, with incremental",
"3:",
":=MESS(he,l,he,l.hf.hfg) for edges",
"sampling, there exists at least one entity e'e vegrg and relation",
"(es,r,eo) Eef :",
"4:",
"obtain h=(AGG(m(esr,e) (eseo))forentities ",
"reR such that (e,r,e')e8.",
"N(Veg-rg) :",
"3.3Learning semantic-awaredistribution",
"5:",
"logits computation: obtain the Gumbel logits Geo= g(he:et)-",
"lgg) withUenifom(01)forntities",
"增量采样机制缓解了减少目标候选项的问题。但是，从。",
"candidate sampling: obtain sampled entties Vegra",
"中随机抽取K实体，并没有考虑不同实体对不同查询的语义",
"6:",
"相关性。像波束搜索[38]这样的基于贪婪的搜索算法在这里",
"(arg topkGeo,eo e Veg,rg):",
"不适用，因为没有直接的启发式方法来测量语义相关性。为了",
"7:",
"straight-through: h :=(1-no_grad(p'(e))+p'(e))-h, for entities",
"解决这个问题，我们为每个传播步骤，i用参数o参数化采",
"eeVeara:",
"样器S()。e。，Veg.rg=S(Vegrge\"）。在以下两部分中，我们",
"update propagation path: update Vegrg = Vegrq Ueqrg",
"8:",
"介绍了采样分布以及如何优化采样参数",
"9: end for",
"10: return f(ht.: wT) for each eoe Veg.r",
"3.3.1Parameterized samplingdistribution.对于基于GNN的",
"KG推理方法，最终传播步骤中的实体表示hc。用于测量实体",
"px",
"eo作为目标答案实体的合理性。换句话说：学习实体表示以",
"指示它们与给定查询（eqrg.）的相关性。为了分享这些知识，",
"的关键思想是在采样信号中反向传播，就好像它是恒等函数一",
"我们在每个层（=1...L中引人一个带有参数o'eRd的线性映",
"样。具体来说，我们使用h：=(1-no_grad(p(e))+p(e)·h，而",
"射函数g(he;0)=ohe。，并根据概率分布",
"不是直接使用实体表示he来计算消息，其中no_grad(p(e))",
"p(e) = exp(h:0)/t)/Zeeag exp(g(hg:0/)/t), (2)",
"表示反向传播信号不会经过此术语。在这种方法中，前向计",
"算不会受到影响，而后向信号可以通过概率乘法器p（e）。估",
"计离散采样分布上的梯度的另一种选择是REINFORCE技术E",
"进行采样，温度值t>0",
"[25,37]。然而，已知REINFORCE梯度具有高方差E[32]，并",
"在不从p'(e)替换的情况下对K实体进行采样需要顺序采样",
"且方差将在传播步骤中累积，因此这里不使用。Adam优化器",
"并更新每个样本的分布，这可能很昂贵。我们遵循Gumbeltop",
"[19]同时更新模型参数w和采样器参数0",
"k技巧[22,39]来解决这个问题。Gumbel-trick首先从均匀分布",
"总的来说，学习语义感知分布的目的是根据查询关系自适",
"中采样eg,rgl个独立噪声，即。e。，Ue～Uniform(0,1)为每",
"应地保留有前途的目标。如第4.3节所示。通过与实体表示共",
"个e e Vegrg 形成 Gumbel 日志 Ge=g(he:of)-log(-log Ue)。",
"享知识，采样传播路径中的选定信息在语义上与查询关系相",
"然后根据Gumbellogits Ge的值收集Vegrg中的顶级K实体。",
"关。另一方面，与未学习的版本相比，学习的分布保证了目标",
"正如在[22,39]中证明的那样，此过程等效于p(e）中没有替",
"实体的更高覆盖率",
"换的样本",
"3.4The full algorithm",
"3.3.2Learningstrategy.我们将={010L）表示为采样",
"AdaProp的完整过程如算法1所示。给定第（e-1)步中的实",
"参数，将参数化的传播路径表示为Gcg.rg(0）。由于我们在(2)",
"体Veq-rq，我们首先获得Veq-rg的相邻实体N(Veg-rg），并",
"中的采样分布与GNN表示密切相关，因此联合优化模型参数",
"生成新访问的实体egr。=N(Vt-1)|Veor。作为行E2中的",
"和采样器参数可以更好地在两个部分之间共享知识。具体来",
"候选者。然后，在行E3中的边8上计算与查询相关的消息",
"说，我们设计了以下目标",
"并将其传播到行E4中的N(Veal）中的实体。根据在5行中",
"w*,0*=argmin",
"L(F(w,Geg,rg",
"t.(0)),ea)(3)",
"计算的Gumbel 对数，在6行中对Vegrg，中的顶级K实体进",
"行采样。至于ST估计器，我们修改了7行中Vegrg中实体的",
"w,eL(eqrq.ea)eQra",
"每个实例上的损失函数L是所有实体eoEvea,ra上的二进制",
"交叉炳损失，i。e",
"隐藏表示。最后，将采样实体集cara与Veol。连接，以在E",
"L(F(w, G(0)),ea)= -",
"ye,log(peo)+(1-ye.)log(1-pe.),",
"8行中形成第（步实体Vea.rg集。在传播L步骤之后，返回所",
"Leoevara",
"有涉及实体的最终步骤表示 he。。将Vegrq中f(heo;we）评",
"，其中实体eo的似然分数de。=f(he:wT)[0,1]表示eo作为",
"估的得分最高的实体作为预测目标",
"目标答案实体的合理性，如果eo=ea否则为0，则标签ye。=1",
"与现有的基于GNN的KG推理方法相比，关键的区别在于，",
"重新参数化技巧通常用于Gumbel分布[17]。但是，它不做",
"在消息传播之前，不再给出和固定算法E1中的传播路径。相",
"显式采样，因此仍然需要很高的计算成本。相反，我们选择直通",
"反，它根据实体表示进行自适应调整，并在每个传播步骤中动",
"(ST)估计[2,17]，它可以近似估计离散变量的梯度。ST估计器",
"态更新。此外，算法1还享有比全传播和渐进传播方法更高的"
] |
{
"arxiv_id": "2205.15319",
"language": "zh",
"page_idx": 3
}
|
||
2309.15427-zh-3
|
2309.15427-zh-3.jpg
|
<FigureHere>
Figure 2: The overall framework. Given a multiple choice question, we frst retrieve subgraphs from the knowledge graph based on the entities in the question and options. We then develop Graph Neural Prompting (GNP) to encode the pertinent factual knowledge and structural information to obtain the Graph Neural Prompt. GNP contains various designs including a GNN, a cross-modality pooling module, a domain projector, and a self-supervised link prediction objective. Later, the obtained Graph Neural Prompt is sent into LLM for inference along with the input text embedding. We utilize the standard maximum likelihood objective for downstream task adaptation, while LLM is kept frozen or tuned depending on different experimental settings.
Prompting LLMs for Question Answering
提示是引起来自 LLM (Liu et al. 2023) 的响应的事实上的方法。提示 LLM 进行多项选择问题回答的典型方法很简单。给定一个问题$Q$,可选的上下文$C$和答案选项$A$,我们首先将$C,\dot{Q},A$的连接标记为一系列输人文本标记$X$ 。然后，我们设计了一系列提示令牌$P$ ,并将其添加到输人文本令牌$X$中，后者后来被视为LLM 模型的输人以生成预测$y^\prime=f([P,X])$ 。可以使用标准的最大似然损失，使用老师强制 (Williams and Zipser 1989) 和交叉熵损失来训练 LLM 模型以适应下游任务：
$$\mathcal L_{llm}=-\log p(y|X,\Theta),$$
(1)
其中$p$是模型参数化的概率分布。提示$P$可以是文本输入形式的硬提示，也可以是可学习嵌入向量形式的软提示
与仅使用文本字符串作为硬提示的现有方法不同，我们的 GraphNeuralPrompting 方法将知识图$\mathcal{G}$中包含的结构和事实信息编码为软提示$P$ ,这是一个可训练的向量序列，可以与$X$的令牌嵌人串联。鼓励学习$P$ ,以提供来自$\mathcal{G}$ 的丰富结构信息和知识，以及每个数据实例的任务指令
Subgraph Retrieval
为了在语义上将输人文本标记$X$与具有数百万个节点的海量知识图$\mathcal{G}$对齐，我们检索$\mathcal{G}$的子图，其中包含$X$中标记的相关实体。特别是，对于每个答案选项$a_k$ 及其相应的上下文$C$和问题$Q$ ,我们首先通过实体链接获得一组匹配的实体$\mathcal{E}_match$ ,以将$X$中的令牌与$\mathcal{G}$ 中的实体进行匹配。然后，我们基于$\mathcal{E}_match$中的实体
检索子图$G^{\prime}$ ,包括它们的两跳邻居和连接它们的关系(Yasunaga et al. 2022)。检索到的子图包含必要的内容和知识，以帮助模型回答$Q$
Graph Neural Prompting
图形神经提示包含各种设计，包括嵌人知识图的 GNN 编码器，确定相关节点嵌入的跨模态池模块，桥接图形和文本之间差异的域投影仪，以及鼓励模型识别结构信息的自监督链接预测目标
GNN 编码器。尽管检索到的子图$\mathcal{G}^\prime$包含有关问题和答案选择的丰富上下文信息，但某些实体和关系与实际问题无关。在$\mathcal{G}^\prime$中直接馈送每个事实三元组会引人噪声并防止 LLM 模型集中在关键信息上。因此，我们引人GNN 来编码最相关的知识，并进一步整合实体之间的复杂关系。特别是，我们首先使用预先训练的实体嵌入(Feng et al. 2020; Yasunaga, Leskovec, and Liang 2022) 初始化节点嵌入。接下来，我们使用标准的图关注网络$( \text{Veli\v {c}kovi\' {c}etal. 2018) 作 为 检 索 到 的 子 图 }\mathcal{G} ^{\prime }$的 GNN 编码器。编码过程如下：
$$H_{1}=f_{GNN}(\mathcal{G}'),$$
(2)
其中$H_1\in\mathbb{R}^{d_g}$表示 GNN 为$\mathcal{G}^\prime$中的每个节点学习的
节点嵌人，$d_g$表示 GNN 编码器的输出维度
跨模态池。为了确定与问题相关的最相关节点，并将节点嵌入合并为整体图级表示以供后续使用，我们设计了跨模态池模块。特别是，我们首先引人一个自我关注层，利用内部图特征和节点之间的隐式交互来动态识别节点重要性：
$$H_2=\text{Self-Attn}(H_1),$$
(3)
|
[
[
0.12744140625,
0.080810546875,
0.1396484375,
0.07763671875,
0.1446533203125,
0.0889892578125,
0.13232421875,
0.0921630859375
],
[
0.34716796875,
0.07763671875,
0.406005859375,
0.0802001953125,
0.406005859375,
0.0921630859375,
0.34716796875,
0.09027099609375
],
[
0.50244140625,
0.08209228515625,
0.685546875,
0.08209228515625,
0.685546875,
0.0953369140625,
0.50244140625,
0.0953369140625
],
[
0.1478271484375,
0.0953369140625,
0.16259765625,
0.0953369140625,
0.16259765625,
0.10479736328125,
0.1478271484375,
0.10479736328125
],
[
0.334228515625,
0.09027099609375,
0.393798828125,
0.0921630859375,
0.393798828125,
0.10418701171875,
0.334228515625,
0.10162353515625
],
[
0.32275390625,
0.10418701171875,
0.378173828125,
0.10418701171875,
0.378173828125,
0.11553955078125,
0.32275390625,
0.11553955078125
],
[
0.59619140625,
0.10797119140625,
0.66259765625,
0.10797119140625,
0.66259765625,
0.1319580078125,
0.59619140625,
0.1319580078125
],
[
0.7138671875,
0.1029052734375,
0.7958984375,
0.10418701171875,
0.7958984375,
0.11553955078125,
0.7138671875,
0.1142578125
],
[
0.1070556640625,
0.10919189453125,
0.191162109375,
0.10919189453125,
0.191162109375,
0.12060546875,
0.1070556640625,
0.12060546875
],
[
0.3544921875,
0.1142578125,
0.386474609375,
0.11871337890625,
0.384033203125,
0.1319580078125,
0.351318359375,
0.1275634765625
],
[
0.71728515625,
0.11553955078125,
0.7939453125,
0.11553955078125,
0.7939453125,
0.126953125,
0.71728515625,
0.126953125
],
[
0.82177734375,
0.10986328125,
0.83837890625,
0.10986328125,
0.83837890625,
0.117431640625,
0.82177734375,
0.117431640625
],
[
0.10455322265625,
0.121826171875,
0.1953125,
0.121826171875,
0.1953125,
0.1319580078125,
0.10455322265625,
0.1319580078125
],
[
0.400390625,
0.11614990234375,
0.4150390625,
0.11614990234375,
0.4150390625,
0.1256103515625,
0.400390625,
0.1256103515625
],
[
0.205078125,
0.1300048828125,
0.2132568359375,
0.1300048828125,
0.2132568359375,
0.1351318359375,
0.205078125,
0.1351318359375
],
[
0.336669921875,
0.1337890625,
0.3447265625,
0.1363525390625,
0.34228515625,
0.1395263671875,
0.3349609375,
0.1383056640625
],
[
0.420654296875,
0.136962890625,
0.431396484375,
0.136962890625,
0.431396484375,
0.146484375,
0.420654296875,
0.146484375
],
[
0.366943359375,
0.143310546875,
0.378173828125,
0.143310546875,
0.378173828125,
0.1514892578125,
0.366943359375,
0.1514892578125
],
[
0.59716796875,
0.14453125,
0.66650390625,
0.14453125,
0.66650390625,
0.1673583984375,
0.59716796875,
0.1673583984375
],
[
0.2320556640625,
0.1514892578125,
0.283447265625,
0.1514892578125,
0.283447265625,
0.1634521484375,
0.2320556640625,
0.1634521484375
],
[
0.41015625,
0.1514892578125,
0.426513671875,
0.1514892578125,
0.426513671875,
0.160400390625,
0.41015625,
0.160400390625
],
[
0.779296875,
0.1483154296875,
0.79736328125,
0.1483154296875,
0.79736328125,
0.1673583984375,
0.779296875,
0.1673583984375
],
[
0.2344970703125,
0.1634521484375,
0.281982421875,
0.1634521484375,
0.281982421875,
0.1749267578125,
0.2344970703125,
0.1749267578125
],
[
0.334228515625,
0.1634521484375,
0.34228515625,
0.1634521484375,
0.34228515625,
0.1673583984375,
0.334228515625,
0.1673583984375
],
[
0.41259765625,
0.169189453125,
0.42724609375,
0.1717529296875,
0.4248046875,
0.18115234375,
0.41015625,
0.1787109375
],
[
0.70361328125,
0.169189453125,
0.71728515625,
0.169189453125,
0.71728515625,
0.177978515625,
0.70361328125,
0.177978515625
],
[
0.370849609375,
0.17041015625,
0.3857421875,
0.17041015625,
0.3857421875,
0.1787109375,
0.370849609375,
0.1787109375
],
[
0.09722900390625,
0.173583984375,
0.19775390625,
0.173583984375,
0.19775390625,
0.1849365234375,
0.09722900390625,
0.1849365234375
],
[
0.340576171875,
0.173583984375,
0.348876953125,
0.173583984375,
0.348876953125,
0.1787109375,
0.340576171875,
0.1787109375
],
[
0.525390625,
0.1749267578125,
0.5537109375,
0.1749267578125,
0.5537109375,
0.186279296875,
0.525390625,
0.186279296875
],
[
0.7099609375,
0.177978515625,
0.73046875,
0.177978515625,
0.73046875,
0.1925048828125,
0.7099609375,
0.1925048828125
],
[
0.8388671875,
0.175537109375,
0.85302734375,
0.175537109375,
0.85302734375,
0.1787109375,
0.8388671875,
0.1787109375
],
[
0.78369140625,
0.18115234375,
0.7939453125,
0.18115234375,
0.7939453125,
0.1956787109375,
0.78369140625,
0.1956787109375
],
[
0.10296630859375,
0.1849365234375,
0.1953125,
0.1849365234375,
0.1953125,
0.1962890625,
0.10296630859375,
0.1962890625
],
[
0.5947265625,
0.18115234375,
0.6640625,
0.18115234375,
0.6640625,
0.2083740234375,
0.5947265625,
0.2083740234375
],
[
0.330810546875,
0.1912841796875,
0.435546875,
0.1912841796875,
0.435546875,
0.2020263671875,
0.330810546875,
0.2020263671875
],
[
0.515625,
0.1875,
0.560546875,
0.1875,
0.560546875,
0.1976318359375,
0.515625,
0.1976318359375
],
[
0.70166015625,
0.205810546875,
0.736328125,
0.205810546875,
0.736328125,
0.2171630859375,
0.70166015625,
0.2171630859375
],
[
0.77294921875,
0.20703125,
0.8056640625,
0.20703125,
0.8056640625,
0.2310791015625,
0.77294921875,
0.2310791015625
],
[
0.8349609375,
0.210205078125,
0.8798828125,
0.212158203125,
0.87841796875,
0.2247314453125,
0.833984375,
0.222900390625
],
[
0.6005859375,
0.2108154296875,
0.6611328125,
0.2108154296875,
0.6611328125,
0.222900390625,
0.6005859375,
0.222900390625
],
[
0.11273193359375,
0.21337890625,
0.1683349609375,
0.21337890625,
0.1683349609375,
0.2235107421875,
0.11273193359375,
0.2235107421875
],
[
0.240234375,
0.2215576171875,
0.261474609375,
0.2215576171875,
0.261474609375,
0.2342529296875,
0.240234375,
0.2342529296875
],
[
0.69384765625,
0.2171630859375,
0.744140625,
0.2171630859375,
0.744140625,
0.2291259765625,
0.69384765625,
0.2291259765625
],
[
0.281982421875,
0.2235107421875,
0.3095703125,
0.2235107421875,
0.3095703125,
0.236083984375,
0.281982421875,
0.236083984375
],
[
0.35693359375,
0.2235107421875,
0.37744140625,
0.2266845703125,
0.374267578125,
0.2342529296875,
0.35546875,
0.231689453125
],
[
0.83251953125,
0.2247314453125,
0.8837890625,
0.2247314453125,
0.8837890625,
0.236083984375,
0.83251953125,
0.236083984375
],
[
0.69775390625,
0.2279052734375,
0.740234375,
0.2298583984375,
0.73876953125,
0.2412109375,
0.69677734375,
0.2392578125
],
[
0.11273193359375,
0.2247314453125,
0.17724609375,
0.2247314453125,
0.17724609375,
0.236083984375,
0.11273193359375,
0.236083984375
],
[
0.265625,
0.236083984375,
0.324462890625,
0.236083984375,
0.324462890625,
0.2474365234375,
0.265625,
0.2474365234375
],
[
0.35546875,
0.2374267578125,
0.375732421875,
0.2374267578125,
0.375732421875,
0.2442626953125,
0.35546875,
0.2442626953125
],
[
0.75732421875,
0.2374267578125,
0.82421875,
0.2374267578125,
0.82421875,
0.248779296875,
0.75732421875,
0.248779296875
],
[
0.11273193359375,
0.2442626953125,
0.160888671875,
0.2442626953125,
0.160888671875,
0.255126953125,
0.11273193359375,
0.255126953125
],
[
0.11273193359375,
0.255615234375,
0.176513671875,
0.255615234375,
0.176513671875,
0.267578125,
0.11273193359375,
0.267578125
],
[
0.11273193359375,
0.268310546875,
0.160888671875,
0.268310546875,
0.160888671875,
0.27978515625,
0.11273193359375,
0.27978515625
],
[
0.473876953125,
0.26318359375,
0.619140625,
0.26318359375,
0.619140625,
0.2734375,
0.473876953125,
0.2734375
],
[
0.11273193359375,
0.281005859375,
0.1854248046875,
0.281005859375,
0.1854248046875,
0.291748046875,
0.11273193359375,
0.291748046875
],
[
0.7998046875,
0.27783203125,
0.8310546875,
0.27783203125,
0.8310546875,
0.300537109375,
0.7998046875,
0.300537109375
],
[
0.25244140625,
0.285400390625,
0.298095703125,
0.285400390625,
0.298095703125,
0.296630859375,
0.25244140625,
0.296630859375
],
[
0.56201171875,
0.285888671875,
0.5810546875,
0.285888671875,
0.5810546875,
0.290283203125,
0.56201171875,
0.290283203125
],
[
0.62353515625,
0.285888671875,
0.65625,
0.285888671875,
0.65625,
0.29736328125,
0.62353515625,
0.29736328125
],
[
0.72802734375,
0.2890625,
0.775390625,
0.2890625,
0.775390625,
0.2998046875,
0.72802734375,
0.2998046875
],
[
0.85693359375,
0.2890625,
0.89892578125,
0.2890625,
0.89892578125,
0.2998046875,
0.85693359375,
0.2998046875
],
[
0.34228515625,
0.290283203125,
0.346435546875,
0.296630859375,
0.34228515625,
0.298583984375,
0.338134765625,
0.292236328125
],
[
0.29736328125,
0.290283203125,
0.310546875,
0.290283203125,
0.310546875,
0.298583984375,
0.29736328125,
0.298583984375
],
[
0.3203125,
0.292236328125,
0.325927734375,
0.292236328125,
0.325927734375,
0.298583984375,
0.3203125,
0.298583984375
],
[
0.384033203125,
0.29345703125,
0.40283203125,
0.29345703125,
0.40283203125,
0.300537109375,
0.384033203125,
0.300537109375
],
[
0.611328125,
0.296630859375,
0.6640625,
0.298583984375,
0.6640625,
0.31005859375,
0.611328125,
0.30810546875
],
[
0.73046875,
0.2998046875,
0.76953125,
0.2998046875,
0.76953125,
0.311279296875,
0.73046875,
0.311279296875
],
[
0.78369140625,
0.2998046875,
0.8408203125,
0.3017578125,
0.8388671875,
0.31640625,
0.78173828125,
0.314453125
],
[
0.85546875,
0.2998046875,
0.89892578125,
0.2998046875,
0.89892578125,
0.311279296875,
0.85546875,
0.311279296875
],
[
0.2344970703125,
0.296630859375,
0.2900390625,
0.296630859375,
0.2900390625,
0.30810546875,
0.2344970703125,
0.30810546875
],
[
0.1094970703125,
0.300537109375,
0.19287109375,
0.300537109375,
0.19287109375,
0.311279296875,
0.1094970703125,
0.311279296875
],
[
0.38232421875,
0.3037109375,
0.408447265625,
0.3037109375,
0.408447265625,
0.311279296875,
0.38232421875,
0.311279296875
],
[
0.90625,
0.304931640625,
0.91015625,
0.304931640625,
0.91015625,
0.31005859375,
0.90625,
0.31005859375
],
[
0.61767578125,
0.309326171875,
0.66015625,
0.309326171875,
0.66015625,
0.32080078125,
0.61767578125,
0.32080078125
],
[
0.2484130859375,
0.30810546875,
0.2900390625,
0.30810546875,
0.2900390625,
0.318115234375,
0.2484130859375,
0.318115234375
],
[
0.7236328125,
0.311279296875,
0.779296875,
0.311279296875,
0.779296875,
0.322509765625,
0.7236328125,
0.322509765625
],
[
0.859375,
0.311279296875,
0.8935546875,
0.311279296875,
0.8935546875,
0.322509765625,
0.859375,
0.322509765625
],
[
0.79345703125,
0.3125,
0.83837890625,
0.3125,
0.83837890625,
0.322509765625,
0.79345703125,
0.322509765625
],
[
0.12744140625,
0.3125,
0.17724609375,
0.3125,
0.17724609375,
0.323974609375,
0.12744140625,
0.323974609375
],
[
0.08660888671875,
0.34228515625,
0.91259765625,
0.34228515625,
0.91259765625,
0.35595703125,
0.08660888671875,
0.35595703125
],
[
0.0849609375,
0.35595703125,
0.91259765625,
0.354736328125,
0.91259765625,
0.369384765625,
0.0849609375,
0.37060546875
],
[
0.08660888671875,
0.37060546875,
0.9140625,
0.37060546875,
0.9140625,
0.3837890625,
0.08660888671875,
0.3837890625
],
[
0.08660888671875,
0.3837890625,
0.91259765625,
0.3837890625,
0.91259765625,
0.397705078125,
0.08660888671875,
0.397705078125
],
[
0.0849609375,
0.396484375,
0.9111328125,
0.397705078125,
0.9111328125,
0.412353515625,
0.0849609375,
0.410888671875
],
[
0.08660888671875,
0.412353515625,
0.9111328125,
0.412353515625,
0.9111328125,
0.425537109375,
0.08660888671875,
0.425537109375
],
[
0.08416748046875,
0.4248046875,
0.576171875,
0.4267578125,
0.576171875,
0.43994140625,
0.08416748046875,
0.438232421875
],
[
0.0931396484375,
0.46728515625,
0.420654296875,
0.46728515625,
0.420654296875,
0.48095703125,
0.0931396484375,
0.48095703125
],
[
0.521484375,
0.46728515625,
0.9111328125,
0.46728515625,
0.9111328125,
0.48095703125,
0.521484375,
0.48095703125
],
[
0.08660888671875,
0.486083984375,
0.47802734375,
0.48486328125,
0.47802734375,
0.5,
0.08660888671875,
0.50048828125
],
[
0.52294921875,
0.48095703125,
0.9111328125,
0.48095703125,
0.9111328125,
0.494384765625,
0.52294921875,
0.494384765625
],
[
0.0906982421875,
0.50048828125,
0.476318359375,
0.50048828125,
0.476318359375,
0.5146484375,
0.0906982421875,
0.5146484375
],
[
0.521484375,
0.49560546875,
0.72119140625,
0.49560546875,
0.72119140625,
0.5087890625,
0.521484375,
0.5087890625
],
[
0.08660888671875,
0.5146484375,
0.47802734375,
0.5146484375,
0.47802734375,
0.52783203125,
0.08660888671875,
0.52783203125
],
[
0.08660888671875,
0.52783203125,
0.47802734375,
0.52783203125,
0.47802734375,
0.54150390625,
0.08660888671875,
0.54150390625
],
[
0.525390625,
0.52294921875,
0.72607421875,
0.5234375,
0.72607421875,
0.53857421875,
0.525390625,
0.537109375
],
[
0.08660888671875,
0.54150390625,
0.4794921875,
0.54052734375,
0.4794921875,
0.5546875,
0.08660888671875,
0.55615234375
],
[
0.521484375,
0.54443359375,
0.91259765625,
0.54443359375,
0.91259765625,
0.55810546875,
0.521484375,
0.55810546875
],
[
0.10296630859375,
0.55615234375,
0.47802734375,
0.55615234375,
0.47802734375,
0.5693359375,
0.10296630859375,
0.5693359375
],
[
0.521484375,
0.55810546875,
0.91259765625,
0.55810546875,
0.91259765625,
0.5712890625,
0.521484375,
0.5712890625
],
[
0.0882568359375,
0.5693359375,
0.47802734375,
0.5693359375,
0.47802734375,
0.58349609375,
0.0882568359375,
0.58349609375
],
[
0.51953125,
0.57080078125,
0.91259765625,
0.5712890625,
0.91259765625,
0.5859375,
0.51953125,
0.58544921875
],
[
0.08660888671875,
0.58349609375,
0.4794921875,
0.58203125,
0.4794921875,
0.5966796875,
0.08660888671875,
0.59765625
],
[
0.521484375,
0.58544921875,
0.70166015625,
0.58544921875,
0.70166015625,
0.5986328125,
0.521484375,
0.5986328125
],
[
0.08660888671875,
0.59765625,
0.476318359375,
0.59765625,
0.476318359375,
0.611328125,
0.08660888671875,
0.611328125
],
[
0.521484375,
0.599609375,
0.91259765625,
0.599609375,
0.91259765625,
0.61279296875,
0.521484375,
0.61279296875
],
[
0.08660888671875,
0.611328125,
0.1446533203125,
0.611328125,
0.1446533203125,
0.62548828125,
0.08660888671875,
0.62548828125
],
[
0.521484375,
0.61279296875,
0.9111328125,
0.6142578125,
0.9111328125,
0.62890625,
0.521484375,
0.62744140625
],
[
0.200927734375,
0.626953125,
0.365234375,
0.626953125,
0.365234375,
0.6416015625,
0.200927734375,
0.6416015625
],
[
0.455810546875,
0.626953125,
0.47802734375,
0.626953125,
0.47802734375,
0.64013671875,
0.455810546875,
0.64013671875
],
[
0.52294921875,
0.62744140625,
0.9111328125,
0.62744140625,
0.9111328125,
0.6416015625,
0.52294921875,
0.6416015625
],
[
0.0882568359375,
0.646484375,
0.47802734375,
0.646484375,
0.47802734375,
0.66015625,
0.0882568359375,
0.66015625
],
[
0.52294921875,
0.642578125,
0.9111328125,
0.642578125,
0.9111328125,
0.65576171875,
0.52294921875,
0.65576171875
],
[
0.0882568359375,
0.6611328125,
0.47802734375,
0.6611328125,
0.47802734375,
0.6748046875,
0.0882568359375,
0.6748046875
],
[
0.521484375,
0.65576171875,
0.9111328125,
0.65576171875,
0.9111328125,
0.66943359375,
0.521484375,
0.66943359375
],
[
0.521484375,
0.66845703125,
0.91015625,
0.66943359375,
0.91015625,
0.6845703125,
0.521484375,
0.68310546875
],
[
0.0882568359375,
0.6767578125,
0.1217041015625,
0.6767578125,
0.1217041015625,
0.68798828125,
0.0882568359375,
0.68798828125
],
[
0.10540771484375,
0.689453125,
0.47802734375,
0.689453125,
0.47802734375,
0.70263671875,
0.10540771484375,
0.70263671875
],
[
0.52294921875,
0.6845703125,
0.9111328125,
0.6845703125,
0.9111328125,
0.69775390625,
0.52294921875,
0.69775390625
],
[
0.0882568359375,
0.70263671875,
0.476318359375,
0.70263671875,
0.476318359375,
0.71728515625,
0.0882568359375,
0.71728515625
],
[
0.521484375,
0.69775390625,
0.9111328125,
0.69775390625,
0.9111328125,
0.7109375,
0.521484375,
0.7109375
],
[
0.52294921875,
0.7109375,
0.91259765625,
0.7109375,
0.91259765625,
0.72607421875,
0.52294921875,
0.72607421875
],
[
0.0882568359375,
0.71728515625,
0.47802734375,
0.71728515625,
0.47802734375,
0.73095703125,
0.0882568359375,
0.73095703125
],
[
0.08660888671875,
0.73095703125,
0.4794921875,
0.73095703125,
0.4794921875,
0.744140625,
0.08660888671875,
0.744140625
],
[
0.521484375,
0.72607421875,
0.69384765625,
0.72607421875,
0.69384765625,
0.7392578125,
0.521484375,
0.7392578125
],
[
0.09234619140625,
0.74560546875,
0.47802734375,
0.74560546875,
0.47802734375,
0.7587890625,
0.09234619140625,
0.7587890625
],
[
0.65673828125,
0.75,
0.775390625,
0.75,
0.775390625,
0.7646484375,
0.65673828125,
0.7646484375
],
[
0.890625,
0.75048828125,
0.91259765625,
0.75048828125,
0.91259765625,
0.7646484375,
0.890625,
0.7646484375
],
[
0.0849609375,
0.75830078125,
0.200927734375,
0.75830078125,
0.200927734375,
0.771484375,
0.0849609375,
0.771484375
],
[
0.52294921875,
0.7734375,
0.9111328125,
0.7734375,
0.9111328125,
0.787109375,
0.52294921875,
0.787109375
],
[
0.09234619140625,
0.7861328125,
0.2459716796875,
0.7861328125,
0.2459716796875,
0.80029296875,
0.09234619140625,
0.80029296875
],
[
0.521484375,
0.78857421875,
0.83642578125,
0.78857421875,
0.83642578125,
0.8017578125,
0.521484375,
0.8017578125
],
[
0.08905029296875,
0.8046875,
0.476318359375,
0.8046875,
0.476318359375,
0.8193359375,
0.08905029296875,
0.8193359375
],
[
0.521484375,
0.80322265625,
0.9111328125,
0.80322265625,
0.9111328125,
0.81640625,
0.521484375,
0.81640625
],
[
0.0882568359375,
0.8193359375,
0.476318359375,
0.8193359375,
0.476318359375,
0.83349609375,
0.0882568359375,
0.83349609375
],
[
0.52294921875,
0.8173828125,
0.9111328125,
0.8173828125,
0.9111328125,
0.83056640625,
0.52294921875,
0.83056640625
],
[
0.08660888671875,
0.83203125,
0.476318359375,
0.83349609375,
0.476318359375,
0.84765625,
0.08660888671875,
0.8466796875
],
[
0.525390625,
0.83056640625,
0.91259765625,
0.83056640625,
0.91259765625,
0.8447265625,
0.525390625,
0.8447265625
],
[
0.08905029296875,
0.84765625,
0.47802734375,
0.84765625,
0.47802734375,
0.861328125,
0.08905029296875,
0.861328125
],
[
0.52294921875,
0.84521484375,
0.91259765625,
0.84521484375,
0.91259765625,
0.859375,
0.52294921875,
0.859375
],
[
0.0849609375,
0.85986328125,
0.4794921875,
0.85986328125,
0.4794921875,
0.87744140625,
0.0849609375,
0.87744140625
],
[
0.52392578125,
0.85986328125,
0.611328125,
0.85986328125,
0.611328125,
0.8720703125,
0.52392578125,
0.8720703125
],
[
0.0882568359375,
0.87548828125,
0.47802734375,
0.87548828125,
0.47802734375,
0.89013671875,
0.0882568359375,
0.89013671875
],
[
0.64208984375,
0.87353515625,
0.7890625,
0.875,
0.7890625,
0.8896484375,
0.64208984375,
0.88818359375
],
[
0.8896484375,
0.87548828125,
0.9140625,
0.87548828125,
0.9140625,
0.8896484375,
0.8896484375,
0.8896484375
]
] |
[
"O",
"Option (c)",
"GraphNeural Prompting (GNP)",
"®",
"Option (b)",
"Option (a)",
"8",
"Self-supervised",
"General Domain",
"PA3",
"LinkPrediction",
"h",
"KnowledgeGraph",
"O",
"or",
"",
"C",
"01",
"80日",
"Subgraph",
"A2",
"日",
"Retrieval",
"A:",
"",
"2",
"Q3",
"Biomedical Domain",
"S",
"GNN",
"尚",
"",
".0",
"KnowledgeGraph",
"08",
"ExtractedSubgraphs",
"Encoder",
"Cross",
"0",
"Domain",
"NodeEmbs",
"Question:",
"A-Z",
"Modality",
"LLM",
"Text",
"Projector",
"Pooling",
"Q1Q2Q3Q4",
"Dictionary",
"Emb",
"GraphEmbs",
"Options:",
"(a) A1A2A3",
"(b) B1B2",
"ObtainGraphNeuralPrompt",
"(c)C1C2C3C4",
"A",
"Graph",
"9",
"Large",
"A1A2'A3",
"A1A2A3",
"",
"",
".",
"Text",
"Language",
"Model",
"Maximum",
"Ground",
"Neural",
"Multiple Choice",
"Emb",
"",
"Models",
"Prompt",
"Prediction",
"Truth",
"Likelihood",
"Question",
"Figure 2: The overall framework. Given a multiple choice question, we first retrieve subgraphs from the knowledge",
"graph based on the entities in the question and options. We then develop Graph Neural Prompting (GNP) to encode",
"the pertinent factual knowledge and structural information to obtain the Graph Neural Prompt. GNP contains",
"various designs including a GNN, a cross-modality pooling module, a domain projector, and a self-supervised link",
"prediction objective. Later, the obtained Graph Neural Prompt is sent into LLM for inference along with the input",
"text embedding. We utilize the standard maximum likelihood objective for downstream task adaptation, while LLM",
"iskeptfrozenortuneddependingondifferentexperimental settings.",
"Prompting LLMs for QuestionAnswering",
"检索子图g'：包括它们的两跳邻居和连接它们的关系",
"提示是引起来自LLM（Liuetal.2023）的响应的事实",
"Yasunagaetal.2022）。检索到的子图包含必要的内容",
"上的方法。提示LLM进行多项选择问题回答的典型方",
"和知识，以帮助模型回答Q",
"法很简单。给定一个问题Q，可选的上下文C和答案",
"选项A，我们首先将C,Q,A的连接标记为一系列输",
"Graph Neural Prompting",
"人文本标记X。然后，我们设计了一系列提示令牌P",
"图形神经提示包含各种设计，包括嵌入知识图的GNN",
"并将其添加到输文本令牌X中，后者后来被视为",
"编码器，确定相关节点嵌入的跨模态池模块，桥接图形",
"LLM模型的输人以生成预测y'=f([P,X）。可以使",
"和文本之间差异的域投影仪，以及鼓励模型识别结构信",
"用标准的最大似然损失，使用老师强制（Williamsand",
"息的自监督链接预测目标",
"Zipser1989）和交叉损失来训练LLM模型以适应下",
"GNN编码器。尽管检索到的子图G'包含有关问题和答",
"游任务：",
"案选择的丰富上下文信息，但某些实体和关系与实际问",
"Cum = - logp(ylX,),",
"(1)",
"题无关。在'中直接馈送每个事实三元组会引人噪声",
"其中p是模型参数化的概率分布。提示P可以是文本",
"并防止LLM模型集中在关键信息上。因此，我们引人",
"输人形式的硬提示，也可以是可学习嵌人向量形式的软",
"GNN来编码最相关的知识，并进一步整合实体之间的",
"复杂关系。特别是，我们首先使用预先训练的实体嵌人",
"提示",
"与仅使用文本字符串作为硬提示的现有方法不同，我",
"(Feng et al. 2020; Yasunaga, Leskovec, and Liang 2022)",
"们的GraphNeuralPrompting方法将知识图g中包含",
"初始化节点嵌人。接下来，我们使用标准的图关注网络",
"(Velickovicetal.2018）作为检索到的子图g'的GNN",
"的结构和事实信息编码为软提示P，这是一个可训练",
"的向量序列，可以与X的令牌嵌人串联。鼓励学习P",
"编码器。编码过程如下：",
"以提供来自9的丰富结构信息和知识，以及每个数据",
"H1 = fGNN(g'),",
"(2)",
"实例的任务指令",
"其中HiERdg表示GNN为G'中的每个节点学习的",
"Subgraph Retrieval",
"节点嵌入，dg表示GNN编码器的输出维度",
"为了在语义上将输人文本标记X与具有数百万个节点",
"跨模态池。为了确定与问题相关的最相关节点，并将节",
"的海量知识图9对齐，我们检索9的子图，其中包含",
"点嵌入合并为整体图级表示以供后续使用，我们设计",
"X中标记的相关实体。特别是，对于每个答案选项ak",
"了跨模态池模块。特别是，我们首先引人一个自我关注",
"及其相应的上下文C和问题Q，我们首先通过实体链",
"层，利用内部图特征和节点之间的隐式交互来动态识别",
"接获得一组匹配的实体&match，以将X中的令牌与g",
"节点重要性：",
"中的实体进行匹配。然后，我们基于&match中的实体",
"H2 = Self-Attn(H1),",
"(3)"
] |
{
"arxiv_id": "2309.15427",
"language": "zh",
"page_idx": 3
}
|
||
1903.08855-zh-0
|
1903.08855-zh-0.jpg
|
# 语言知识和上下文表示的可传递性
Nelson F. Liu$^{\bullet \heartsuit * }$ Matt Gardner* Yonatan Belinkov$^{\textcircled{\mathrm{o}}}$
Matthew E. Peters* Noah A. Smith$^{\spadesuit\clubsuit}$
$^{\bullet}$Paul G. Allen School of Computer Science &Engineering,
University of Washington, Seattle, WA, USA
$^{\textcircled{\circ}}$Department of Linguistics, University of Washington, Seattle, WA, USA
${}^{{}}$Allen Institute for Artificial Intelligence, Seattle, WA, USA
$^{\diamond}$Harvard John A. Paulson School of Engineering and Applied Sciences and
MIT Computer Science and Artificial Intelligence Laboratory, Cambridge, MA, USA
$\{$ nfliu,nasmith $\}$ @cs.washington.edu
belinkov@seas.harvard.edu
$\{{\mathrm{~mattg,matthewp~}}\}{\mathrm{~@allenai.org,}}$
Abstract
<FigureHere>
Figure 1: An illustration of the probing model setup used to study the linguistic knowledge within contextual word representations.
Contextual word representations derived from large-scale neural language models are successful across a diverse set of NLP tasks, suggesting that they encode useful and transferable features of language. To shed light on the linguistic knowledge they capture, we study the representations produced by several recent pretrained contextualizers (variants of ELMo, the OpenAI transformer language model, and BERT) with a suite of seventeen diverse probing tasks. We find that linear models $\bar{\text{trained on top of frozen contextual rep-}}$ resentations are competitive with state-ofthe-art task-specific models in many cases, but fail on tasks requiring fine-grained linguistic knowledge $\dot{\text{(e.g., conjunct identifi-}}$ cation). To investigate the transferability of contextual word representations, we quantify differences in the transferability of individual layers within contextualizers, especially between recurrent neural networks ( RNNs) and transformers. For in- stance, higher layers of RNNs are more task-specific, while transformer layers do not exhibit the same monotonic trend. In addition, to better understand what makes contextual word representations transferable, we compare language model pretraining with eleven supervised pretraining tasks. For any given task, pretraining on a closely related task yields better performance than language model pretraining (which is better on average) when the pretraining dataset is fixed. However, language model pretraining on more data gives the best results.
Introduction
型的关键组成部分。传统上，这些单词向量是静态的-为每个单词分配一个向量。最近的工作已经探索了上下文单词表示(此后：cwrs), 它为每个单词分配一个向量，该向量是整个输人序列的函数；这使他们能够在上下文中对单词的使用进行建模。cwrs 通常是在具有大数据集的任务上训练的神经网络(我们称为上下文化器)的输出，例如机器翻译 (McCann et al., 2017)和语言建模 (Peters et al., 2018a) 。cwrs 非常有效-在最新模型中使用它们代替传统的静态单词向量可以在各种 NLP 任务中获得巨大收益
cwrs 的广泛成功表明，它们编码了有用的， 可转移的语言功能。然而，他们的语言知识和可转移性还没有得到很好的理解
最近的工作探索了语言模型和神经机器翻译系统捕获的语言知识，但是这些研究通常集中在一个现象上，例如。g。,了解分层语法 (Blevins et al., 2018) 或形态学 (Belinkov et al., 2017a) 。我们通过研究 cwrs 来扩展先前的工作，其中包括 17 个不同的探测任务，旨在评估各种现象，例如共指，语义关系知识和实体信息等。结果是 cwrs 中编码的语言知识的更广泛视图
关于可移植性，在语言建模任务上对上下文化器进行预训练在经验上取得了最大的成功， 但是我们也可以考虑对具有其他监督目标的上
## 1
预训练单词表示 (Mikolov et al., 2013; Pennington et al., 2014) 是最先进的神经 NLP 模
$^*$Work done while at the Allen Institute for Artificial
Intelligence.
|
[
[
0.322509765625,
0.0736083984375,
0.68115234375,
0.0736083984375,
0.68115234375,
0.0914306640625,
0.322509765625,
0.0914306640625
],
[
0.26025390625,
0.097412109375,
0.568359375,
0.097412109375,
0.568359375,
0.11517333984375,
0.26025390625,
0.11517333984375
],
[
0.58447265625,
0.09979248046875,
0.75146484375,
0.09979248046875,
0.75146484375,
0.114013671875,
0.58447265625,
0.114013671875
],
[
0.322509765625,
0.114013671875,
0.5087890625,
0.1134033203125,
0.5087890625,
0.1312255859375,
0.322509765625,
0.1318359375
],
[
0.5087890625,
0.119384765625,
0.52490234375,
0.119384765625,
0.52490234375,
0.1270751953125,
0.5087890625,
0.1270751953125
],
[
0.51708984375,
0.11517333984375,
0.68408203125,
0.11224365234375,
0.68408203125,
0.1300048828125,
0.51708984375,
0.133056640625
],
[
0.2451171875,
0.1300048828125,
0.76171875,
0.1318359375,
0.76171875,
0.149658203125,
0.2451171875,
0.1478271484375
],
[
0.307373046875,
0.1484375,
0.69873046875,
0.149658203125,
0.69873046875,
0.1644287109375,
0.307373046875,
0.163330078125
],
[
0.183837890625,
0.163330078125,
0.82373046875,
0.1650390625,
0.82373046875,
0.182861328125,
0.183837890625,
0.18115234375
],
[
0.2392578125,
0.179931640625,
0.767578125,
0.1817626953125,
0.767578125,
0.1995849609375,
0.2392578125,
0.19775390625
],
[
0.172119140625,
0.196533203125,
0.83642578125,
0.198974609375,
0.83642578125,
0.2161865234375,
0.172119140625,
0.21435546875
],
[
0.139404296875,
0.21728515625,
0.869140625,
0.21728515625,
0.869140625,
0.23095703125,
0.139404296875,
0.23095703125
],
[
0.344970703125,
0.2340087890625,
0.66748046875,
0.2340087890625,
0.66748046875,
0.2476806640625,
0.344970703125,
0.2476806640625
],
[
0.2275390625,
0.2493896484375,
0.5205078125,
0.25048828125,
0.5205078125,
0.265380859375,
0.2275390625,
0.26416015625
],
[
0.54736328125,
0.248779296875,
0.78271484375,
0.2493896484375,
0.78271484375,
0.26416015625,
0.54736328125,
0.26318359375
],
[
0.26611328125,
0.267333984375,
0.343505859375,
0.267333984375,
0.343505859375,
0.281982421875,
0.26611328125,
0.281982421875
],
[
0.5146484375,
0.265380859375,
0.59423828125,
0.265380859375,
0.59423828125,
0.276611328125,
0.5146484375,
0.276611328125
],
[
0.59619140625,
0.26953125,
0.63232421875,
0.26953125,
0.63232421875,
0.281982421875,
0.59619140625,
0.281982421875
],
[
0.6767578125,
0.267333984375,
0.7109375,
0.267333984375,
0.7109375,
0.283203125,
0.6767578125,
0.283203125
],
[
0.75390625,
0.26953125,
0.78271484375,
0.26953125,
0.78271484375,
0.283203125,
0.75390625,
0.283203125
],
[
0.82958984375,
0.26953125,
0.861328125,
0.26953125,
0.861328125,
0.281982421875,
0.82958984375,
0.281982421875
],
[
0.5146484375,
0.275634765625,
0.58837890625,
0.275634765625,
0.58837890625,
0.2861328125,
0.5146484375,
0.2861328125
],
[
0.67919921875,
0.27978515625,
0.724609375,
0.27978515625,
0.724609375,
0.29638671875,
0.67919921875,
0.29638671875
],
[
0.8330078125,
0.27978515625,
0.861328125,
0.27978515625,
0.861328125,
0.297607421875,
0.8330078125,
0.297607421875
],
[
0.76416015625,
0.28076171875,
0.77685546875,
0.283935546875,
0.77099609375,
0.29443359375,
0.759765625,
0.29150390625
],
[
0.611328125,
0.28076171875,
0.63037109375,
0.28076171875,
0.63037109375,
0.29443359375,
0.611328125,
0.29443359375
],
[
0.145263671875,
0.287353515625,
0.4609375,
0.2880859375,
0.4609375,
0.302734375,
0.145263671875,
0.3017578125
],
[
0.60693359375,
0.29443359375,
0.63232421875,
0.29443359375,
0.63232421875,
0.3154296875,
0.60693359375,
0.3154296875
],
[
0.67919921875,
0.29443359375,
0.70947265625,
0.29443359375,
0.70947265625,
0.314208984375,
0.67919921875,
0.314208984375
],
[
0.755859375,
0.295654296875,
0.7841796875,
0.295654296875,
0.7841796875,
0.3134765625,
0.755859375,
0.3134765625
],
[
0.82861328125,
0.29345703125,
0.86572265625,
0.29345703125,
0.86572265625,
0.3154296875,
0.82861328125,
0.3154296875
],
[
0.14697265625,
0.303955078125,
0.4609375,
0.303955078125,
0.4609375,
0.317626953125,
0.14697265625,
0.317626953125
],
[
0.51318359375,
0.298583984375,
0.58447265625,
0.298583984375,
0.58447265625,
0.308837890625,
0.51318359375,
0.308837890625
],
[
0.14697265625,
0.317626953125,
0.459228515625,
0.317626953125,
0.459228515625,
0.330810546875,
0.14697265625,
0.330810546875
],
[
0.51318359375,
0.32177734375,
0.59423828125,
0.322509765625,
0.59423828125,
0.334228515625,
0.51318359375,
0.333251953125
],
[
0.8486328125,
0.317626953125,
0.85986328125,
0.317626953125,
0.85986328125,
0.32177734375,
0.8486328125,
0.32177734375
],
[
0.67529296875,
0.32470703125,
0.7119140625,
0.32470703125,
0.7119140625,
0.34033203125,
0.67529296875,
0.34033203125
],
[
0.75,
0.32470703125,
0.78662109375,
0.32470703125,
0.78662109375,
0.34033203125,
0.75,
0.34033203125
],
[
0.8271484375,
0.32470703125,
0.8623046875,
0.32470703125,
0.8623046875,
0.34033203125,
0.8271484375,
0.34033203125
],
[
0.145263671875,
0.33203125,
0.4609375,
0.33203125,
0.4609375,
0.345703125,
0.145263671875,
0.345703125
],
[
0.51318359375,
0.33203125,
0.591796875,
0.33203125,
0.591796875,
0.342529296875,
0.51318359375,
0.342529296875
],
[
0.60205078125,
0.32666015625,
0.6337890625,
0.32666015625,
0.6337890625,
0.339111328125,
0.60205078125,
0.339111328125
],
[
0.68994140625,
0.339111328125,
0.6962890625,
0.339111328125,
0.6962890625,
0.344482421875,
0.68994140625,
0.344482421875
],
[
0.765625,
0.34033203125,
0.7734375,
0.34033203125,
0.7734375,
0.345703125,
0.765625,
0.345703125
],
[
0.14697265625,
0.3466796875,
0.4609375,
0.3466796875,
0.4609375,
0.35986328125,
0.14697265625,
0.35986328125
],
[
0.6591796875,
0.3486328125,
0.8017578125,
0.3486328125,
0.8017578125,
0.3623046875,
0.6591796875,
0.3623046875
],
[
0.14697265625,
0.35986328125,
0.459228515625,
0.35986328125,
0.459228515625,
0.37353515625,
0.14697265625,
0.37353515625
],
[
0.611328125,
0.364013671875,
0.62451171875,
0.364013671875,
0.62451171875,
0.377197265625,
0.611328125,
0.377197265625
],
[
0.68701171875,
0.364013671875,
0.7001953125,
0.364013671875,
0.7001953125,
0.377197265625,
0.68701171875,
0.377197265625
],
[
0.759765625,
0.363525390625,
0.77734375,
0.363525390625,
0.77734375,
0.377685546875,
0.759765625,
0.377685546875
],
[
0.8388671875,
0.364013671875,
0.85205078125,
0.364013671875,
0.85205078125,
0.3759765625,
0.8388671875,
0.3759765625
],
[
0.14697265625,
0.374755859375,
0.459228515625,
0.374755859375,
0.459228515625,
0.388427734375,
0.14697265625,
0.388427734375
],
[
0.51318359375,
0.377197265625,
0.57666015625,
0.377197265625,
0.57666015625,
0.388427734375,
0.51318359375,
0.388427734375
],
[
0.67529296875,
0.377197265625,
0.7109375,
0.37890625,
0.70947265625,
0.391357421875,
0.67333984375,
0.3896484375
],
[
0.75146484375,
0.377197265625,
0.78662109375,
0.37890625,
0.78515625,
0.391357421875,
0.75,
0.3896484375
],
[
0.82373046875,
0.3759765625,
0.86376953125,
0.377685546875,
0.8623046875,
0.39013671875,
0.82275390625,
0.388427734375
],
[
0.14697265625,
0.3896484375,
0.459228515625,
0.3896484375,
0.459228515625,
0.402587890625,
0.14697265625,
0.402587890625
],
[
0.603515625,
0.377685546875,
0.62890625,
0.377685546875,
0.62890625,
0.39013671875,
0.603515625,
0.39013671875
],
[
0.14697265625,
0.402587890625,
0.459228515625,
0.402587890625,
0.459228515625,
0.416259765625,
0.14697265625,
0.416259765625
],
[
0.5146484375,
0.40380859375,
0.884765625,
0.40380859375,
0.884765625,
0.41748046875,
0.5146484375,
0.41748046875
],
[
0.14697265625,
0.41748046875,
0.459228515625,
0.41748046875,
0.459228515625,
0.431640625,
0.14697265625,
0.431640625
],
[
0.51123046875,
0.418701171875,
0.88330078125,
0.418701171875,
0.88330078125,
0.431640625,
0.51123046875,
0.431640625
],
[
0.14697265625,
0.43115234375,
0.4609375,
0.43115234375,
0.4609375,
0.44482421875,
0.14697265625,
0.44482421875
],
[
0.5146484375,
0.43408203125,
0.75390625,
0.43408203125,
0.75390625,
0.44482421875,
0.5146484375,
0.44482421875
],
[
0.145263671875,
0.446533203125,
0.240966796875,
0.4453125,
0.240966796875,
0.457275390625,
0.145263671875,
0.457763671875
],
[
0.235107421875,
0.446533203125,
0.459228515625,
0.446533203125,
0.459228515625,
0.457275390625,
0.235107421875,
0.457275390625
],
[
0.145263671875,
0.460205078125,
0.45751953125,
0.460205078125,
0.45751953125,
0.473876953125,
0.145263671875,
0.473876953125
],
[
0.145263671875,
0.473876953125,
0.459228515625,
0.47265625,
0.459228515625,
0.486328125,
0.145263671875,
0.48681640625
],
[
0.5146484375,
0.47265625,
0.88330078125,
0.473876953125,
0.88330078125,
0.488037109375,
0.5146484375,
0.48681640625
],
[
0.14697265625,
0.488037109375,
0.459228515625,
0.488037109375,
0.459228515625,
0.501953125,
0.14697265625,
0.501953125
],
[
0.5146484375,
0.4892578125,
0.8818359375,
0.490478515625,
0.8818359375,
0.5048828125,
0.5146484375,
0.50439453125
],
[
0.14697265625,
0.501953125,
0.456787109375,
0.501953125,
0.456787109375,
0.515625,
0.14697265625,
0.515625
],
[
0.515625,
0.50732421875,
0.88330078125,
0.50732421875,
0.88330078125,
0.52099609375,
0.515625,
0.52099609375
],
[
0.145263671875,
0.5166015625,
0.459228515625,
0.515625,
0.459228515625,
0.52978515625,
0.145263671875,
0.53076171875
],
[
0.515625,
0.5224609375,
0.8818359375,
0.5224609375,
0.8818359375,
0.5361328125,
0.515625,
0.5361328125
],
[
0.14697265625,
0.52978515625,
0.45751953125,
0.52978515625,
0.45751953125,
0.54345703125,
0.14697265625,
0.54345703125
],
[
0.51708984375,
0.5380859375,
0.8818359375,
0.5380859375,
0.8818359375,
0.552734375,
0.51708984375,
0.552734375
],
[
0.14697265625,
0.5458984375,
0.4609375,
0.5458984375,
0.4609375,
0.55859375,
0.14697265625,
0.55859375
],
[
0.14697265625,
0.55859375,
0.459228515625,
0.55859375,
0.459228515625,
0.572265625,
0.14697265625,
0.572265625
],
[
0.5146484375,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.56787109375,
0.5146484375,
0.56787109375
],
[
0.14697265625,
0.572265625,
0.459228515625,
0.57373046875,
0.459228515625,
0.58740234375,
0.14697265625,
0.5859375
],
[
0.515625,
0.57080078125,
0.88330078125,
0.57080078125,
0.88330078125,
0.58447265625,
0.515625,
0.58447265625
],
[
0.1435546875,
0.58740234375,
0.459228515625,
0.5859375,
0.459228515625,
0.599609375,
0.1435546875,
0.60107421875
],
[
0.515625,
0.5859375,
0.88330078125,
0.5859375,
0.88330078125,
0.60107421875,
0.515625,
0.60107421875
],
[
0.14697265625,
0.6015625,
0.459228515625,
0.6015625,
0.459228515625,
0.615234375,
0.14697265625,
0.615234375
],
[
0.5146484375,
0.6025390625,
0.884765625,
0.6025390625,
0.884765625,
0.6162109375,
0.5146484375,
0.6162109375
],
[
0.145263671875,
0.615234375,
0.4609375,
0.615234375,
0.4609375,
0.62890625,
0.145263671875,
0.62890625
],
[
0.5146484375,
0.61767578125,
0.880859375,
0.61767578125,
0.880859375,
0.63134765625,
0.5146484375,
0.63134765625
],
[
0.145263671875,
0.6298828125,
0.462646484375,
0.6298828125,
0.462646484375,
0.6435546875,
0.145263671875,
0.6435546875
],
[
0.5146484375,
0.63427734375,
0.8818359375,
0.63427734375,
0.8818359375,
0.6484375,
0.5146484375,
0.6484375
],
[
0.1485595703125,
0.6455078125,
0.459228515625,
0.6455078125,
0.459228515625,
0.65625,
0.1485595703125,
0.65625
],
[
0.5146484375,
0.65087890625,
0.57275390625,
0.65087890625,
0.57275390625,
0.6650390625,
0.5146484375,
0.6650390625
],
[
0.145263671875,
0.65771484375,
0.4609375,
0.65771484375,
0.4609375,
0.67138671875,
0.145263671875,
0.67138671875
],
[
0.14697265625,
0.6728515625,
0.459228515625,
0.6728515625,
0.459228515625,
0.68603515625,
0.14697265625,
0.68603515625
],
[
0.52880859375,
0.66748046875,
0.87890625,
0.66748046875,
0.87890625,
0.68115234375,
0.52880859375,
0.68115234375
],
[
0.14697265625,
0.68701171875,
0.459228515625,
0.68701171875,
0.459228515625,
0.70068359375,
0.14697265625,
0.70068359375
],
[
0.515625,
0.68408203125,
0.88330078125,
0.68408203125,
0.88330078125,
0.69873046875,
0.515625,
0.69873046875
],
[
0.145263671875,
0.70068359375,
0.45751953125,
0.70068359375,
0.45751953125,
0.71435546875,
0.145263671875,
0.71435546875
],
[
0.515625,
0.70068359375,
0.7734375,
0.70068359375,
0.7734375,
0.71435546875,
0.515625,
0.71435546875
],
[
0.145263671875,
0.71533203125,
0.459228515625,
0.71533203125,
0.459228515625,
0.728515625,
0.145263671875,
0.728515625
],
[
0.5322265625,
0.71728515625,
0.88330078125,
0.71728515625,
0.88330078125,
0.732421875,
0.5322265625,
0.732421875
],
[
0.145263671875,
0.72998046875,
0.459228515625,
0.72998046875,
0.459228515625,
0.74365234375,
0.145263671875,
0.74365234375
],
[
0.5146484375,
0.732421875,
0.8818359375,
0.73291015625,
0.8818359375,
0.74755859375,
0.5146484375,
0.74658203125
],
[
0.14697265625,
0.74365234375,
0.459228515625,
0.74365234375,
0.459228515625,
0.75732421875,
0.14697265625,
0.75732421875
],
[
0.5146484375,
0.74755859375,
0.8818359375,
0.75,
0.8818359375,
0.76416015625,
0.5146484375,
0.7626953125
],
[
0.14697265625,
0.75830078125,
0.4609375,
0.75830078125,
0.4609375,
0.771484375,
0.14697265625,
0.771484375
],
[
0.14697265625,
0.77099609375,
0.459228515625,
0.77099609375,
0.459228515625,
0.78515625,
0.14697265625,
0.78515625
],
[
0.5146484375,
0.765625,
0.8818359375,
0.765625,
0.8818359375,
0.77978515625,
0.5146484375,
0.77978515625
],
[
0.14697265625,
0.7861328125,
0.4609375,
0.7861328125,
0.4609375,
0.7998046875,
0.14697265625,
0.7998046875
],
[
0.51318359375,
0.7822265625,
0.8818359375,
0.7822265625,
0.8818359375,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.14697265625,
0.80029296875,
0.307373046875,
0.80029296875,
0.307373046875,
0.81396484375,
0.14697265625,
0.81396484375
],
[
0.515625,
0.79736328125,
0.8818359375,
0.79736328125,
0.8818359375,
0.81103515625,
0.515625,
0.81103515625
],
[
0.151123046875,
0.8193359375,
0.26611328125,
0.82177734375,
0.26611328125,
0.8359375,
0.151123046875,
0.83447265625
],
[
0.515625,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.515625,
0.82763671875
],
[
0.1259765625,
0.82177734375,
0.133544921875,
0.8291015625,
0.1259765625,
0.8330078125,
0.118408203125,
0.826171875
],
[
0.515625,
0.830078125,
0.884765625,
0.830078125,
0.884765625,
0.84326171875,
0.515625,
0.84326171875
],
[
0.12005615234375,
0.845703125,
0.489501953125,
0.845703125,
0.489501953125,
0.859375,
0.12005615234375,
0.859375
],
[
0.515625,
0.8466796875,
0.62646484375,
0.8466796875,
0.62646484375,
0.85791015625,
0.515625,
0.85791015625
],
[
0.1175537109375,
0.8623046875,
0.489501953125,
0.85986328125,
0.489501953125,
0.87451171875,
0.1175537109375,
0.876953125
],
[
0.5341796875,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.5341796875,
0.876953125
],
[
0.1376953125,
0.8798828125,
0.490234375,
0.8818359375,
0.490234375,
0.8994140625,
0.1376953125,
0.89794921875
],
[
0.515625,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.892578125,
0.515625,
0.892578125
],
[
0.1175537109375,
0.89453125,
0.204833984375,
0.896484375,
0.204833984375,
0.91015625,
0.1175537109375,
0.90771484375
],
[
0.5146484375,
0.8935546875,
0.8818359375,
0.89453125,
0.8818359375,
0.9091796875,
0.5146484375,
0.90771484375
]
] |
[
"语言知识和上下文表示的可传递性",
"Nelson F. Liu* Matt Gardner*",
"Yonatan Belinkov",
"Matthew E. Peters*",
"",
"Noah A. Smith**",
"*Paul G. Allen School of Computer Science & Engineering,",
"University of Washington, Seattle, WA, USA",
"Department of Linguistics, University of Washington, Seattle, WA, USA",
"*Allen Institute for Artificial Intelligence, Seattle, WA, USA",
"Harvard John A. Paulson School of Engineering and Applied Sciences and",
"MIT Computer Science and Artificial Intelligence Laboratory, Cambridge, MA, USA",
"f nfliu,nasmith 1 @cs.washington.edu",
" mattg,matthewp @allenai.org,",
"belinkov@seas.harvard.edu",
"Abstract",
"Predicted Labels",
"NNP",
"NNP",
"VBZ",
"NNP",
"(e.g., Pos tags)",
".",
"1",
"4",
"+",
"Contextual word representations derived",
"中",
"中",
"D",
"中",
"from large-scale neural language models",
"Probing Model",
"are successful across a diverse set of NLP",
"Contextual Word",
"二",
"8",
"8",
"8",
"tasks, suggesting that they encode useful",
"Representations",
"",
"4",
"4",
"and transferable features of language. To",
"Pretrained Contextualizer",
"shed light on the linguistic knowledge they",
"1",
"1",
"1",
"4",
"capture, we study the representations pro-",
"Input Tokens",
"Haag",
"plays",
"Elianti",
"duced by several recent pretrained contex-",
"Ms.",
"tualizers (variants of ELMo, the OpenAI",
"Figure l: An illustration of the probing model",
"transformer language model, and BERT)",
"setup used to study the linguistic knowledge within",
"with a suite of seventeen diverse prob-",
"contextual word representations.",
"ingtasks.",
"We find that linear models",
"trained on top of frozen contextual rep-",
"resentations are competitive with state-of-",
"型的关键组成部分。传统上，这些单词向量是",
"the-art task-specific models in many cases,",
"静态的-为每个单词分配一个向量。最近的工",
"but fail on tasks requiring fine-grained lin",
"作已经探索了上下文单词表示（此后：cwrs），",
"guistic knowledge (e.g., conjunct identifi-",
"它为每个单词分配一个向量，该向量是整个输",
"cation). To investigate the transferabil",
"人序列的函数；这使他们能够在上下文中对单",
"ity of contextual word representations, we",
"quantify differences in the transferability",
"词的使用进行建模。cwrs通常是在具有大数据",
"of individual layers within contextualizers,",
"集的任务上训练的神经网络（我们称为上下文",
"especially between recurrent neural net-",
"化器）的输出，例如机器翻译（McCannetal.,",
"works (RNNs) and transformers. For in-",
"2017）和语言建模（Petersetal.,2018a）。cwrs",
"stance, higher layers of RNNs are more",
"非常有效-在最新模型中使用它们代替传统的",
"task-specific, while transformer layers do",
"静态单词向量可以在各种NLP任务中获得巨",
"not exhibit the same monotonic trend. In",
"大收益",
"addition, to better understand what makes",
"contextual word representations transfer-",
"cwrs的广泛成功表明，它们编码了有用的",
"able, we compare language model pre-",
"可转移的语言功能。然而，他们的语言知识和",
"training with eleven supervised pretrain-",
"可转移性还没有得到很好的理解",
"ing tasks. For any given task, pretrain-",
"最近的工作探索了语言模型和神经机器翻",
"ing on a closely related task yields bet-",
"译系统捕获的语言知识，但是这些研究通常",
"ter performance than language model pre-",
"集中在一个现象上，例如。g。，了解分层语",
"training (which is better on average) when",
"the pretraining dataset is fixed. However,",
"法（Blevins et al.，2018）或形态学（Belinkov",
"language model pretraining on more data",
"etal.，2017a）。我们通过研究cwrs来扩展先",
"gives the best results.",
"前的工作，其中包括17个不同的探测任务，旨",
"Introduction",
"在评估各种现象，例如共指，语义关系知识和",
"1",
"实体信息等。结果是cwrs中编码的语言知识",
"预训练单词表示（Mikolovetal.，2013；Pen-",
"的更广泛视图",
"ningtonetal.，2014）是最先进的神经NLP模",
"关于可移植性，在语言建模任务上对上下文",
"*Work done while at the Allen Institute for Artificial",
"化器进行预训练在经验上取得了最大的成功，",
"Intelligence.",
"但是我们也可以考虑对具有其他监督目标的上"
] |
{
"arxiv_id": "1903.08855",
"language": "zh",
"page_idx": 0
}
|
||
1903.08855-zh-1
|
1903.08855-zh-1.jpg
|
下文化器进行预训练，并探索其语言知识。我们研究了预培训任务如何影响学习的语言知识，考虑了十二个预培训任务并评估了九个目标任务的可转移性
更好地理解 cwrs 的语言知识和可移植性对于通过新的编码器架构和预训练任务进行原则性增强是必要的。本文提问并回答：
$1.{\mathrm{~What~features~of~language~do~these~vec-}}$
tors capture, and what do they miss? (
84 )
2.How and why does transferability vary
across representation layers in contextu-
alizers? $(\begin{array}{c}{85}\end{array})$
3. How does the choice of pretraining
task affect the vectors' learned linguistic
knowledge and transferability? ( §6 )
我们使用探测模型$^1$(Shi et al., 2016b; Adi et al., 2017; Hupkes et al., 2018; Belinkov and Glass, 2019) 来分析 cwrs 中的语言信息。具体来说，我们从预训练的上下文化器生成单词的特征，并训练一个模型，仅从这些特征进行预测( Figure 1)。如果可以训练一个简单的模型来预测关于一个单词的语言信息(e。g。,其词性标签)或一对单词(e。g。,它们的语义关${\text{系)仅从 cwr(s),我们就可以合理地得出结论}}$, cwr (s) 对这些信息进行编码
我们的分析揭示了有趣的见解，例如：
1. Linear models trained on top of frozen
cwrs are competitive with state-of-the-art task-specific models in many cases, but fail on tasks requiring fne-grained linguistic knowledge. In these cases, we show that task-trained contextual features greatly help with encoding the requisite knowledge.
2. The first layer output of long short-term
memory (LSTM) recurrent neural networks is consistently the most transferable, whereas it is the middle layers for ${\mathrm{transformers.}}$
3. Higher layers in LSTMs are more task-
specific (and thus less general), while the transformer layers do not exhibit this same monotonic increase in taskspecificity.
$^{1}$Sometimes called auxiliary or diagnostic classifiers.
4. Language model pretraining yields repre-
sentations that are more transferable in general than eleven other candidate pretraining tasks, though pretraining on related tasks yields the strongest results for individual end tasks.
### 2 Probing Tasks
我们构建了一套 17 种不同的英语探测任务， 并用它来更好地理解 cwrs 中包含的语言知识。与以前分析句子嵌人(Adi et al., 2017; Conneau et al., 2018, inter alia ) 的属性和任务性能的研究相反，我们特别关注理解单个或成对单词的 cwrs 。我们发布此分析工具包，以支持未来在探测表示内容方面的工作。$^{2}$有关任务设置的详细信息，请参见 Appendix ??
## 2.1 Token Labeling
过去探索神经模型内部表示的大部分工作都检查了各种令牌标记任务，其中每个令牌(Be- $\operatorname*{linkov~et~al.,~2017a,b;~Blevins~et~al.,~2018,~in-}$ ter alia )都是独立做出决定的。我们综合了这些不同的研究，并在此基础上提出了额外的探测任务
词性标记 (POS)任务评估 cwrs 是否捕获基本语法。我们使用两个标准数据集进行实验：PennTreebank (PTB; Marcus et al., 1993 )和通用依赖英语 WebTreebank (UD-EWT; Silveira et al., 2014 )
CCG 超标记 (CCG) 任务评估向量关于上下文中单词的句法角色的细粒度信息。它被认为是“almost parsing”(Bangalore and Joshi, 1999) , 因为一系列超级标签将一个句子映射到一小组可能的解析。我们使用 CCGbank (Hockenmaier and Steedman, 2007),将 PTB 转换为 CCG 派生
语法选区祖先标记任务旨在探查向量对分层语法的了解。对于给定的单词，将对探查模型进行训练，以预测短语结构树中其父(Parent), 祖父母(GPParent)或曾祖父(GGPParent) 的组成标签(来自PTB)
在语义标记任务(ST)中，为令牌分配了标签，这些标签反映了它们在上下文中的语义角色。这些语义标签评估词汇语义，它们对冗余的 POS 区别进行抽象，并消除 POS 标签中有用案例的歧义。我们使用 Bjerva et al. (2016)的数据集；标记集已被开发为并行含义库 (Abzianidze et al., 2017) 的一部分
$^{2}$http://nelsonliu.me/papers/
contextual-repr-analysis
|
[
[
0.12176513671875,
0.07598876953125,
0.48779296875,
0.07598876953125,
0.48779296875,
0.08905029296875,
0.12176513671875,
0.08905029296875
],
[
0.5283203125,
0.07781982421875,
0.880859375,
0.07781982421875,
0.880859375,
0.0914306640625,
0.5283203125,
0.0914306640625
],
[
0.12005615234375,
0.09320068359375,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.5517578125,
0.09320068359375,
0.88330078125,
0.09320068359375,
0.88330078125,
0.10687255859375,
0.5517578125,
0.10687255859375
],
[
0.1234130859375,
0.11102294921875,
0.484375,
0.11102294921875,
0.484375,
0.1217041015625,
0.1234130859375,
0.1217041015625
],
[
0.5517578125,
0.1092529296875,
0.8818359375,
0.10809326171875,
0.8818359375,
0.122314453125,
0.5517578125,
0.12353515625
],
[
0.12005615234375,
0.1258544921875,
0.270263671875,
0.1258544921875,
0.270263671875,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.5517578125,
0.1258544921875,
0.880859375,
0.1258544921875,
0.880859375,
0.1395263671875,
0.5517578125,
0.1395263671875
],
[
0.139404296875,
0.142578125,
0.48779296875,
0.142578125,
0.48779296875,
0.1561279296875,
0.139404296875,
0.1561279296875
],
[
0.5498046875,
0.141357421875,
0.8818359375,
0.141357421875,
0.8818359375,
0.155029296875,
0.5498046875,
0.155029296875
],
[
0.12005615234375,
0.1591796875,
0.48779296875,
0.1591796875,
0.48779296875,
0.1728515625,
0.12005615234375,
0.1728515625
],
[
0.54833984375,
0.1561279296875,
0.7197265625,
0.15673828125,
0.7197265625,
0.171630859375,
0.54833984375,
0.17041015625
],
[
0.12005615234375,
0.174560546875,
0.403076171875,
0.174560546875,
0.403076171875,
0.188232421875,
0.12005615234375,
0.188232421875
],
[
0.5146484375,
0.190673828125,
0.5322265625,
0.190673828125,
0.5322265625,
0.1995849609375,
0.5146484375,
0.1995849609375
],
[
0.54736328125,
0.18701171875,
0.67529296875,
0.188232421875,
0.67529296875,
0.203125,
0.54736328125,
0.201904296875
],
[
0.53076171875,
0.193603515625,
0.54150390625,
0.193603515625,
0.54150390625,
0.196533203125,
0.53076171875,
0.196533203125
],
[
0.1318359375,
0.2060546875,
0.483642578125,
0.2060546875,
0.483642578125,
0.2203369140625,
0.1318359375,
0.2203369140625
],
[
0.51318359375,
0.2161865234375,
0.8818359375,
0.21728515625,
0.8818359375,
0.232177734375,
0.51318359375,
0.23095703125
],
[
0.15283203125,
0.22265625,
0.490234375,
0.221435546875,
0.490234375,
0.236328125,
0.15283203125,
0.2369384765625
],
[
0.5146484375,
0.232177734375,
0.8818359375,
0.2327880859375,
0.8818359375,
0.2476806640625,
0.5146484375,
0.2464599609375
],
[
0.1561279296875,
0.240478515625,
0.189697265625,
0.240478515625,
0.189697265625,
0.2529296875,
0.1561279296875,
0.2529296875
],
[
0.51708984375,
0.2493896484375,
0.8818359375,
0.2493896484375,
0.8818359375,
0.26318359375,
0.51708984375,
0.26318359375
],
[
0.130126953125,
0.262451171875,
0.48779296875,
0.26318359375,
0.48779296875,
0.27783203125,
0.130126953125,
0.276611328125
],
[
0.5146484375,
0.265380859375,
0.8818359375,
0.265380859375,
0.8818359375,
0.279052734375,
0.5146484375,
0.279052734375
],
[
0.1561279296875,
0.279052734375,
0.48779296875,
0.279052734375,
0.48779296875,
0.292236328125,
0.1561279296875,
0.292236328125
],
[
0.515625,
0.281982421875,
0.87890625,
0.281982421875,
0.87890625,
0.292236328125,
0.515625,
0.292236328125
],
[
0.154541015625,
0.29345703125,
0.27197265625,
0.29638671875,
0.27197265625,
0.311279296875,
0.154541015625,
0.30810546875
],
[
0.515625,
0.297607421875,
0.8818359375,
0.297607421875,
0.8818359375,
0.311279296875,
0.515625,
0.311279296875
],
[
0.1292724609375,
0.317626953125,
0.489501953125,
0.319580078125,
0.489501953125,
0.33740234375,
0.1292724609375,
0.3349609375
],
[
0.51708984375,
0.3134765625,
0.8818359375,
0.3134765625,
0.8818359375,
0.327880859375,
0.51708984375,
0.327880859375
],
[
0.5146484375,
0.328857421875,
0.861328125,
0.328857421875,
0.861328125,
0.342529296875,
0.5146484375,
0.342529296875
],
[
0.1561279296875,
0.33740234375,
0.48779296875,
0.33740234375,
0.48779296875,
0.350830078125,
0.1561279296875,
0.350830078125
],
[
0.154541015625,
0.3515625,
0.456787109375,
0.35400390625,
0.456787109375,
0.3681640625,
0.154541015625,
0.366455078125
],
[
0.51123046875,
0.359375,
0.6884765625,
0.35986328125,
0.6884765625,
0.374755859375,
0.51123046875,
0.37353515625
],
[
0.1376953125,
0.385498046875,
0.489501953125,
0.385498046875,
0.489501953125,
0.3984375,
0.1376953125,
0.3984375
],
[
0.5146484375,
0.38427734375,
0.8818359375,
0.38427734375,
0.8818359375,
0.39794921875,
0.5146484375,
0.39794921875
],
[
0.12005615234375,
0.402099609375,
0.489501953125,
0.402099609375,
0.489501953125,
0.4150390625,
0.12005615234375,
0.4150390625
],
[
0.5146484375,
0.399658203125,
0.880859375,
0.399658203125,
0.880859375,
0.413330078125,
0.5146484375,
0.413330078125
],
[
0.12005615234375,
0.41748046875,
0.48779296875,
0.41748046875,
0.48779296875,
0.43115234375,
0.12005615234375,
0.43115234375
],
[
0.5146484375,
0.4150390625,
0.880859375,
0.4150390625,
0.880859375,
0.4287109375,
0.5146484375,
0.4287109375
],
[
0.12176513671875,
0.435302734375,
0.484375,
0.435302734375,
0.484375,
0.4453125,
0.12176513671875,
0.4453125
],
[
0.51318359375,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.51318359375,
0.4453125
],
[
0.12005615234375,
0.449462890625,
0.48779296875,
0.449462890625,
0.48779296875,
0.463134765625,
0.12005615234375,
0.463134765625
],
[
0.5146484375,
0.44775390625,
0.88330078125,
0.44775390625,
0.88330078125,
0.46142578125,
0.5146484375,
0.46142578125
],
[
0.118408203125,
0.465576171875,
0.486083984375,
0.465576171875,
0.486083984375,
0.478515625,
0.118408203125,
0.478515625
],
[
0.5146484375,
0.465576171875,
0.5693359375,
0.465576171875,
0.5693359375,
0.476806640625,
0.5146484375,
0.476806640625
],
[
0.12176513671875,
0.482177734375,
0.489501953125,
0.482177734375,
0.489501953125,
0.496337890625,
0.12176513671875,
0.496337890625
],
[
0.5341796875,
0.48095703125,
0.88330078125,
0.48095703125,
0.88330078125,
0.49462890625,
0.5341796875,
0.49462890625
],
[
0.12176513671875,
0.498779296875,
0.489501953125,
0.498779296875,
0.489501953125,
0.51318359375,
0.12176513671875,
0.51318359375
],
[
0.515625,
0.49755859375,
0.88330078125,
0.49755859375,
0.88330078125,
0.51123046875,
0.515625,
0.51123046875
],
[
0.12005615234375,
0.51416015625,
0.491943359375,
0.51416015625,
0.491943359375,
0.52783203125,
0.12005615234375,
0.52783203125
],
[
0.5146484375,
0.51318359375,
0.88330078125,
0.51318359375,
0.88330078125,
0.52685546875,
0.5146484375,
0.52685546875
],
[
0.12005615234375,
0.52978515625,
0.349365234375,
0.52978515625,
0.349365234375,
0.54443359375,
0.12005615234375,
0.54443359375
],
[
0.515625,
0.52978515625,
0.88330078125,
0.52978515625,
0.88330078125,
0.54345703125,
0.515625,
0.54345703125
],
[
0.1376953125,
0.54736328125,
0.43994140625,
0.54736328125,
0.43994140625,
0.56103515625,
0.1376953125,
0.56103515625
],
[
0.5146484375,
0.5458984375,
0.68603515625,
0.5458984375,
0.68603515625,
0.55859375,
0.5146484375,
0.55859375
],
[
0.5341796875,
0.5625,
0.8818359375,
0.5625,
0.8818359375,
0.57666015625,
0.5341796875,
0.57666015625
],
[
0.1318359375,
0.57763671875,
0.489501953125,
0.57763671875,
0.489501953125,
0.59130859375,
0.1318359375,
0.59130859375
],
[
0.515625,
0.5791015625,
0.8818359375,
0.5791015625,
0.8818359375,
0.5927734375,
0.515625,
0.5927734375
],
[
0.154541015625,
0.59423828125,
0.489501953125,
0.59326171875,
0.489501953125,
0.60693359375,
0.154541015625,
0.60791015625
],
[
0.5146484375,
0.59326171875,
0.8818359375,
0.59423828125,
0.8818359375,
0.609375,
0.5146484375,
0.60791015625
],
[
0.1561279296875,
0.60986328125,
0.489501953125,
0.60986328125,
0.489501953125,
0.62353515625,
0.1561279296875,
0.62353515625
],
[
0.515625,
0.60986328125,
0.8818359375,
0.60986328125,
0.8818359375,
0.62451171875,
0.515625,
0.62451171875
],
[
0.1561279296875,
0.6259765625,
0.490234375,
0.6259765625,
0.490234375,
0.6396484375,
0.1561279296875,
0.6396484375
],
[
0.5146484375,
0.6259765625,
0.8818359375,
0.6259765625,
0.8818359375,
0.6396484375,
0.5146484375,
0.6396484375
],
[
0.1561279296875,
0.64111328125,
0.48779296875,
0.64111328125,
0.48779296875,
0.65478515625,
0.1561279296875,
0.65478515625
],
[
0.515625,
0.6435546875,
0.88330078125,
0.6435546875,
0.88330078125,
0.65771484375,
0.515625,
0.65771484375
],
[
0.1561279296875,
0.65771484375,
0.489501953125,
0.65771484375,
0.489501953125,
0.67138671875,
0.1561279296875,
0.67138671875
],
[
0.5146484375,
0.6591796875,
0.66015625,
0.6591796875,
0.66015625,
0.6728515625,
0.5146484375,
0.6728515625
],
[
0.1561279296875,
0.6748046875,
0.486083984375,
0.6748046875,
0.486083984375,
0.689453125,
0.1561279296875,
0.689453125
],
[
0.5322265625,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.5322265625,
0.6904296875
],
[
0.154541015625,
0.689453125,
0.2939453125,
0.6904296875,
0.2939453125,
0.705078125,
0.154541015625,
0.70361328125
],
[
0.5146484375,
0.6923828125,
0.880859375,
0.6923828125,
0.880859375,
0.7060546875,
0.5146484375,
0.7060546875
],
[
0.515625,
0.708984375,
0.8818359375,
0.708984375,
0.8818359375,
0.72265625,
0.515625,
0.72265625
],
[
0.1318359375,
0.71533203125,
0.48779296875,
0.71533203125,
0.48779296875,
0.72998046875,
0.1318359375,
0.72998046875
],
[
0.515625,
0.7255859375,
0.88330078125,
0.7255859375,
0.88330078125,
0.7392578125,
0.515625,
0.7392578125
],
[
0.1561279296875,
0.732421875,
0.486083984375,
0.732421875,
0.486083984375,
0.7451171875,
0.1561279296875,
0.7451171875
],
[
0.51318359375,
0.74072265625,
0.7109375,
0.74072265625,
0.7109375,
0.75439453125,
0.51318359375,
0.75439453125
],
[
0.1561279296875,
0.74755859375,
0.489501953125,
0.74755859375,
0.489501953125,
0.76123046875,
0.1561279296875,
0.76123046875
],
[
0.154541015625,
0.76318359375,
0.48779296875,
0.76318359375,
0.48779296875,
0.77783203125,
0.154541015625,
0.77783203125
],
[
0.53076171875,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.7724609375,
0.53076171875,
0.7724609375
],
[
0.154541015625,
0.77783203125,
0.2626953125,
0.77978515625,
0.2626953125,
0.79443359375,
0.154541015625,
0.7919921875
],
[
0.51318359375,
0.77392578125,
0.8818359375,
0.77490234375,
0.8818359375,
0.7890625,
0.51318359375,
0.7880859375
],
[
0.51904296875,
0.79150390625,
0.8818359375,
0.79150390625,
0.8818359375,
0.8046875,
0.51904296875,
0.8046875
],
[
0.130126953125,
0.80419921875,
0.489501953125,
0.802734375,
0.489501953125,
0.81787109375,
0.130126953125,
0.818359375
],
[
0.515625,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.515625,
0.82080078125
],
[
0.1561279296875,
0.82080078125,
0.489501953125,
0.82080078125,
0.489501953125,
0.8349609375,
0.1561279296875,
0.8349609375
],
[
0.515625,
0.822265625,
0.8818359375,
0.822265625,
0.8818359375,
0.8359375,
0.515625,
0.8359375
],
[
0.1561279296875,
0.83740234375,
0.489501953125,
0.83740234375,
0.489501953125,
0.85107421875,
0.1561279296875,
0.85107421875
],
[
0.5146484375,
0.8388671875,
0.880859375,
0.8388671875,
0.880859375,
0.8525390625,
0.5146484375,
0.8525390625
],
[
0.1561279296875,
0.8525390625,
0.489501953125,
0.8525390625,
0.489501953125,
0.8662109375,
0.1561279296875,
0.8662109375
],
[
0.5146484375,
0.85400390625,
0.81689453125,
0.85498046875,
0.81689453125,
0.869140625,
0.5146484375,
0.8681640625
],
[
0.1561279296875,
0.869140625,
0.240966796875,
0.869140625,
0.240966796875,
0.88427734375,
0.1561279296875,
0.88427734375
],
[
0.5322265625,
0.8818359375,
0.73291015625,
0.88427734375,
0.73291015625,
0.8984375,
0.5322265625,
0.896484375
],
[
0.14111328125,
0.89453125,
0.489501953125,
0.8955078125,
0.489501953125,
0.91015625,
0.14111328125,
0.9091796875
],
[
0.515625,
0.8984375,
0.67919921875,
0.8984375,
0.67919921875,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"下文化器进行预训练，并探索其语言知识。我",
"4. Language model pretraining yields repre-",
"们研究了预培训任务如何影响学习的语言知",
"sentations that are more transferable in",
"识，考虑了十二个预培训任务并评估了九个自",
"general than eleven other candidate pre-",
"标任务的可转移性",
"training tasks, though pretraining on re-",
"更好地理解cwrs的语言知识和可移植性对",
"lated tasks yields the strongest results for",
"于通过新的编码器架构和预训练任务进行原则",
"individual end tasks.",
"性增强是必要的。本文提问并回答：",
"2",
"Probing Tasks",
"",
"1. What features of language do these vec",
"我们构建了一套17种不同的英语探测任务，",
"tors capture, and what do they miss? (",
"并用它来更好地理解cwrs中包含的语言知",
"$4)",
"识。与以前分析句子嵌入（Adi etal.，2017;",
"2. How and why does transferability vary",
"Conneauetal.，2018，interalia）的属性和任",
"across representation layers in contextu-",
"务性能的研究相反，我们特别关注理解单个或",
"alizers?（$5)",
"成对单词的cwrs。我们发布此分析工具包，以",
"3. How does the choice of pretraining",
"支持未来在探测表示内容方面的工作。2有关",
"任务设置的详细信息，请参见Appendix？？",
"task affect the vectors' learned linguistic",
"knowledge and transferability?($6)",
"2.1Token Labeling",
"我们使用探测模型1（Shietal2016b；Adi",
"过去探索神经模型内部表示的大部分工作都检",
"et al., 2017; Hupkes et al., 2018; Belinkov and",
"查了各种令牌标记任务，其中每个令牌（Be-",
"Glass，2019）来分析cwrs中的语言信息。具体",
"linkov et al., 2017a,b; Blevins et al., 2018, in-",
"来说：我们从预训练的上下文化器生成单词的",
"teralia）都是独立做出决定的。我们综合了这",
"特征，并训练一个模型，仅从这些特征进行预",
"些不同的研究，并在此基础上提出了额外的探",
"测（Figure1）。如果可以训练一个简单的模",
"测任务",
"型来预测关于一个单词的语言信息（e。g。，其",
"词性标记（POS）任务评估cwrs是否捕获",
"词性标签）或一对单词e。g。，它们的语义关",
"基本语法。我们使用两个标准数据集进行实",
"系）仅从cwr（s)，我们就可以合理地得出结论，",
"验: PennTreebank (PTB; Marcus et al., 1993",
"cwr（s）对这些信息进行编码",
"）和通用依赖英语WebTreebank（UD-EWT;",
"我们的分析揭示了有趣的见解，例如：",
"Silveira et al., 2014)",
"CCG超标记（CCG）任务评估向量关于上下",
"1. Linear models trained on top of frozen",
"文中单词的句法角色的细粒度信息。它被认为",
"cwrs are competitive with state-of-the-art",
"是“almost parsing”(Bangalore and Joshi,",
"task-specific models in many cases, but",
"1999），因为一系列超级标签将一个句子映",
"fail on tasks requiring fine-grained lin-",
"射到一小组可能的解析。我们使用CCGbank",
"guistic knowledge. In these cases, we",
"(Hockenmaier and Steedman, 2007)，将PTB",
"show that task-trained contextual fea-",
"转换为CCG派生",
"tures greatly help with encoding the req-",
"语法选区祖先标记任务旨在探查向量对分层",
"uisite knowledge.",
"语法的了解。对于给定的单词，将对探查模型",
"进行训练，以预测短语结构树中其父（Parent），",
"2. The first layer output of long short-term",
"祖父母（GPParent）或曾祖父（GGPParent）",
"memory (LSTM) recurrent neural net-",
"的组成标签（来自PTB",
"works is consistently the most transfer-",
"able, whereas it is the middle layers for",
"在语义标记任务（ST）中，为令牌分配了",
"transformers.",
"标签，这些标签反映了它们在上下文中的语",
"义角色。这些语义标签评估词汇语义，它们对",
"3. Higher layers in LSTMs are more task-",
"穴余的POS区别进行抽象，并消除POS标",
"specific (and thus less general), while",
"签中有用案例的歧义。我们使用Bjervaetal.",
"the transformer layers do not exhibit",
"(2016）的数据集；标记集已被开发为并行含义",
"this same monotonic increase in task-",
"库（Abzianidzeetal.，2017）的一部分",
"specificity.",
"2http://nelsonliu.me/papers/",
"ISometimes called auxiliary or diagnostic classifiers.",
"contextual-repr-analysis"
] |
{
"arxiv_id": "1903.08855",
"language": "zh",
"page_idx": 1
}
|
||
1903.08855-zh-2
|
1903.08855-zh-2.jpg
|
(1) I was booked for/DURATION 2 nights at/Locus this
hotel in/TIME Oct 2007 .
(2) I went to/GOAL ohm after/EXPLANATION\c{c}TIME
reading some of/QUANTITY~WHOLE the reviews .
(3) It was very upsetting to see this kind of/SPECIES
behavior especially in\_front\_of/LOCUS
$my/SOCIALREL\leadsto GESTALT four year\_old~.$
Figure 2:
Annotated sentences from the
STREUSLE 4.0 corpus, used in the preposition supersense disambiguation task. Prepositions are marked by boldface, immediately followed by their labeled function. If applicable, $\leadsto$ precedes the preposition's labeled role. Figure reproduced from Schneider et al. (2018) .
介词超语义消歧是对介词的词汇语义贡献(功能；PS-fxn)及其介导的语义角色或关系(角色；PS 角色)进行分类的任务。此任务是一种专门的词义消歧，并检查了词汇语义知识的一个方面。与上面的标记任务相反，该模型是在单令牌介词上训练和评估的(而不是为序列中的每个令牌做出决定)。我们使用 STREUSLE4。0 语料库 (Schneider et al., 2018);例句出现在 Figure 2中
事件真实性(EF) 任务涉及用它们描述$( \text{Saur\'{ı } and Pustejovsky, 2009, 2012; de Marn- }$ effe et al., 2012) 的事件的真实性来标记短语。例如，在从 Rudinger et al. (2018) 复制的以下示例中，“(1a)表示没有发生离开，而表面上相似的(1b)则没有”
乔不记得离开。。乔不记得离开了
我们使用通用分解语义它发生了 v2 数据集(Rudinger et al.,2018),并训练模型以预测[-3,3]范围内的(非)事实值。与上面的标记任务不同，此任务被视为回归问题，其中仅对与事件相对应的令牌(而不是序列中的每个令牌)进行预测。性能是使用皮尔逊相关性($r$) 来衡量的；我们报告($r\times100$),因此所有任务的指标都在 0 到 100 之间
## 2.2 Segmentation
我们的几个探测任务涉及使用 BIO 或 IO 标签进行分段。在这里，模型被训练成仅从单个单词的 cwr 预测标签
语法分块(Chunk)测试 cwrs 是否包含跨度和边界的概念；任务是将文本分割为浅组成块。我们使用 CoNLL2000 共享任务数据集(Tjong Kim Sang and Buchholz, 2000)
命名实体识别 (NER) 检查 cwrs 是否编码有关实体类型的信息。我们使用 CoNLL2003 共享任务数据集 (Tjong Kim Sang and De Meul- $\operatorname{der,}2003)$
语法错误检测(GED)是识别需要编辑以产生语法正确的句子的标记的任务。鉴于cwrs 是从在大量语法文本上训练的模型中提取的，因此此任务评估嵌入是否对指示其输入异常的特征进行编码(在这种情况下为非语法性)。我们使用第一个英文证书 (Yannakoudakis et al.,2011)数据集，由 Rei and Yannakoudakis (2016)转换为序列标记格式
连词识别 (Conj) 任务挑战模型，以识别包含协调构造中的连词的令牌。这样做需要非常具体的语法知识。数据来自 Ficler and Goldberg (2016)的协调注释 PTB
## 2.3 Pairwise Relations
我们还设计了探测任务，用于检查单词之间的关系是否在 cwrs 中编码。在这些任务中，给定一个单词对$w_1$ , $w_2$ ,我们将$[w_1,w_2,w_1\odot w_2]$ 输入到探测模型中；它经过训练以预测有关标记 (Belinkov, 2018) 之间关系的信息
我们区分电弧预测和电弧分类任务。弧预测是一个二进制分类任务，其中模型被训练以识别两个令牌之间是否存在关系。弧分类是一个多类分类任务，其中模型具有两个令牌，这些令牌通过某种关系链接并经过训练以确定它们是如何相关的
例如，在句法依赖弧预测任务中，模型被赋予两个标记$(w_a,w_b)$的表示形式，并被训练以预测句子的句法依赖解析是否包含以$w_a$为头部，以$w_b$为修饰符的依赖弧。句法依赖弧分类任务向模型呈现两个标记$(w_\mathrm{head},w_\mathrm{mod})$的表示，其中$w_\mathrm{mod}$是$w_\mathrm{head}$的修饰符，并且模型被训练以预测链接它们的句法关系的类型(该依赖弧上的标签)。我们使用 PTB(转换为 UD) 和 UD-EWT
同样，语义依赖弧预测训练模型以预测两个令牌是否通过语义依赖弧连接，而语义依赖弧分类任务训练模型以对令牌之间的语义关系进行分类。我们将 SemEval2015 共享任务(Oepen et al.,2015)中的数据集与 DELPHINMRS 派生语义依赖 (DM) 目标表示一起使用
句法和语义依赖性弧预测和分类任务与用于语义和句法依赖性解析的最新模型密切相关， 该模型对 cwrs 的对进行评分以做出头部附着和弧标记决策 (Dozat and Manning, 2016, 2018)
为了生成依赖弧预测任务的反例，我们取每个正例$(w_\mathrm{head},w_\mathrm{mod})$并生成一个新的反例$(w_\mathrm{rand},w_\mathrm{mod})$。$w_\mathrm{rand}$是句子中的一个随机标记，不是$w_\mathrm{mod}$的头。因此，这些任务中使用的数据集是平衡的
|
[
[
0.12005615234375,
0.0748291015625,
0.151123046875,
0.0748291015625,
0.151123046875,
0.08612060546875,
0.12005615234375,
0.08612060546875
],
[
0.1427001953125,
0.07183837890625,
0.468505859375,
0.07244873046875,
0.468505859375,
0.08612060546875,
0.1427001953125,
0.08489990234375
],
[
0.5322265625,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.5322265625,
0.09027099609375
],
[
0.15283203125,
0.08612060546875,
0.301513671875,
0.08612060546875,
0.301513671875,
0.0968017578125,
0.15283203125,
0.0968017578125
],
[
0.515625,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.515625,
0.10687255859375
],
[
0.12176513671875,
0.1051025390625,
0.15283203125,
0.1051025390625,
0.15283203125,
0.1134033203125,
0.12176513671875,
0.1134033203125
],
[
0.151123046875,
0.10272216796875,
0.456787109375,
0.10272216796875,
0.456787109375,
0.1134033203125,
0.151123046875,
0.1134033203125
],
[
0.151123046875,
0.1134033203125,
0.466796875,
0.1134033203125,
0.466796875,
0.1270751953125,
0.151123046875,
0.1270751953125
],
[
0.5146484375,
0.10986328125,
0.8818359375,
0.10986328125,
0.8818359375,
0.12353515625,
0.5146484375,
0.12353515625
],
[
0.1175537109375,
0.1300048828125,
0.447509765625,
0.12890625,
0.447509765625,
0.1436767578125,
0.1175537109375,
0.144287109375
],
[
0.515625,
0.1258544921875,
0.88330078125,
0.1258544921875,
0.88330078125,
0.1395263671875,
0.515625,
0.1395263671875
],
[
0.15283203125,
0.142578125,
0.391357421875,
0.142578125,
0.391357421875,
0.1561279296875,
0.15283203125,
0.1561279296875
],
[
0.515625,
0.142578125,
0.87890625,
0.142578125,
0.87890625,
0.152587890625,
0.515625,
0.152587890625
],
[
0.1502685546875,
0.15380859375,
0.418212890625,
0.152587890625,
0.418212890625,
0.166259765625,
0.1502685546875,
0.16748046875
],
[
0.51904296875,
0.1591796875,
0.87890625,
0.1591796875,
0.87890625,
0.169189453125,
0.51904296875,
0.169189453125
],
[
0.22412109375,
0.1776123046875,
0.489501953125,
0.1776123046875,
0.489501953125,
0.191162109375,
0.22412109375,
0.191162109375
],
[
0.51318359375,
0.1728515625,
0.8818359375,
0.1728515625,
0.8818359375,
0.1859130859375,
0.51318359375,
0.1859130859375
],
[
0.12005615234375,
0.1787109375,
0.2032470703125,
0.1787109375,
0.2032470703125,
0.190673828125,
0.12005615234375,
0.190673828125
],
[
0.12005615234375,
0.1923828125,
0.48779296875,
0.1923828125,
0.48779296875,
0.2060546875,
0.12005615234375,
0.2060546875
],
[
0.51318359375,
0.188232421875,
0.86376953125,
0.189453125,
0.86376953125,
0.203125,
0.51318359375,
0.201904296875
],
[
0.118408203125,
0.2060546875,
0.48779296875,
0.2060546875,
0.48779296875,
0.2197265625,
0.118408203125,
0.2197265625
],
[
0.5341796875,
0.207275390625,
0.8818359375,
0.207275390625,
0.8818359375,
0.2203369140625,
0.5341796875,
0.2203369140625
],
[
0.118408203125,
0.2197265625,
0.48779296875,
0.2203369140625,
0.48779296875,
0.2340087890625,
0.118408203125,
0.2327880859375
],
[
0.5146484375,
0.223876953125,
0.8818359375,
0.223876953125,
0.8818359375,
0.2369384765625,
0.5146484375,
0.2369384765625
],
[
0.12005615234375,
0.236328125,
0.484375,
0.236328125,
0.484375,
0.2464599609375,
0.12005615234375,
0.2464599609375
],
[
0.5146484375,
0.2381591796875,
0.88330078125,
0.2392578125,
0.88330078125,
0.254150390625,
0.5146484375,
0.2529296875
],
[
0.12005615234375,
0.248779296875,
0.48779296875,
0.248779296875,
0.48779296875,
0.262451171875,
0.12005615234375,
0.262451171875
],
[
0.515625,
0.255859375,
0.7109375,
0.255859375,
0.7109375,
0.26953125,
0.515625,
0.26953125
],
[
0.12005615234375,
0.262451171875,
0.334228515625,
0.26416015625,
0.334228515625,
0.27783203125,
0.12005615234375,
0.275634765625
],
[
0.51123046875,
0.28515625,
0.7109375,
0.28515625,
0.7109375,
0.298583984375,
0.51123046875,
0.298583984375
],
[
0.139404296875,
0.3046875,
0.48779296875,
0.3046875,
0.48779296875,
0.318359375,
0.139404296875,
0.318359375
],
[
0.5146484375,
0.308837890625,
0.8818359375,
0.308837890625,
0.8818359375,
0.32373046875,
0.5146484375,
0.32373046875
],
[
0.1175537109375,
0.320556640625,
0.489501953125,
0.322509765625,
0.489501953125,
0.336181640625,
0.1175537109375,
0.334228515625
],
[
0.5146484375,
0.325927734375,
0.8818359375,
0.325927734375,
0.8818359375,
0.339111328125,
0.5146484375,
0.339111328125
],
[
0.118408203125,
0.33740234375,
0.48779296875,
0.33740234375,
0.48779296875,
0.3515625,
0.118408203125,
0.3515625
],
[
0.51708984375,
0.34033203125,
0.880859375,
0.341552734375,
0.880859375,
0.355712890625,
0.51708984375,
0.355224609375
],
[
0.1351318359375,
0.35400390625,
0.489501953125,
0.35400390625,
0.489501953125,
0.36767578125,
0.1351318359375,
0.36767578125
],
[
0.515625,
0.359375,
0.880859375,
0.359375,
0.880859375,
0.369384765625,
0.515625,
0.369384765625
],
[
0.12005615234375,
0.37060546875,
0.486083984375,
0.37060546875,
0.486083984375,
0.38134765625,
0.12005615234375,
0.38134765625
],
[
0.5146484375,
0.37353515625,
0.80419921875,
0.37353515625,
0.80419921875,
0.38720703125,
0.5146484375,
0.38720703125
],
[
0.12005615234375,
0.385986328125,
0.48779296875,
0.385986328125,
0.48779296875,
0.399658203125,
0.12005615234375,
0.399658203125
],
[
0.5322265625,
0.391357421875,
0.8818359375,
0.391357421875,
0.8818359375,
0.405029296875,
0.5322265625,
0.405029296875
],
[
0.12005615234375,
0.402099609375,
0.48779296875,
0.402099609375,
0.48779296875,
0.4150390625,
0.12005615234375,
0.4150390625
],
[
0.5146484375,
0.40625,
0.880859375,
0.40673828125,
0.880859375,
0.42041015625,
0.5146484375,
0.419189453125
],
[
0.118408203125,
0.418701171875,
0.486083984375,
0.418701171875,
0.486083984375,
0.431640625,
0.118408203125,
0.431640625
],
[
0.515625,
0.42333984375,
0.8818359375,
0.42333984375,
0.8818359375,
0.43701171875,
0.515625,
0.43701171875
],
[
0.12176513671875,
0.435791015625,
0.376220703125,
0.435791015625,
0.376220703125,
0.446533203125,
0.12176513671875,
0.446533203125
],
[
0.515625,
0.439453125,
0.88330078125,
0.439453125,
0.88330078125,
0.453125,
0.515625,
0.453125
],
[
0.139404296875,
0.45361328125,
0.48779296875,
0.45361328125,
0.48779296875,
0.46435546875,
0.139404296875,
0.46435546875
],
[
0.515625,
0.4560546875,
0.88330078125,
0.4560546875,
0.88330078125,
0.4697265625,
0.515625,
0.4697265625
],
[
0.12005615234375,
0.4697265625,
0.48193359375,
0.4697265625,
0.48193359375,
0.479736328125,
0.12005615234375,
0.479736328125
],
[
0.515625,
0.47265625,
0.62451171875,
0.47265625,
0.62451171875,
0.48388671875,
0.515625,
0.48388671875
],
[
0.12176513671875,
0.48388671875,
0.48779296875,
0.48388671875,
0.48779296875,
0.49755859375,
0.12176513671875,
0.49755859375
],
[
0.5322265625,
0.4892578125,
0.88330078125,
0.4892578125,
0.88330078125,
0.5029296875,
0.5322265625,
0.5029296875
],
[
0.12176513671875,
0.50048828125,
0.486083984375,
0.50048828125,
0.486083984375,
0.51416015625,
0.12176513671875,
0.51416015625
],
[
0.51318359375,
0.50439453125,
0.88330078125,
0.5048828125,
0.88330078125,
0.51953125,
0.51318359375,
0.5185546875
],
[
0.12005615234375,
0.5166015625,
0.489501953125,
0.5166015625,
0.489501953125,
0.52978515625,
0.12005615234375,
0.52978515625
],
[
0.515625,
0.521484375,
0.8818359375,
0.521484375,
0.8818359375,
0.53515625,
0.515625,
0.53515625
],
[
0.12005615234375,
0.5322265625,
0.291259765625,
0.5322265625,
0.291259765625,
0.5458984375,
0.12005615234375,
0.5458984375
],
[
0.5146484375,
0.5361328125,
0.880859375,
0.5361328125,
0.880859375,
0.5498046875,
0.5146484375,
0.5498046875
],
[
0.139404296875,
0.5498046875,
0.408935546875,
0.5498046875,
0.408935546875,
0.5634765625,
0.139404296875,
0.5634765625
],
[
0.515625,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.568359375,
0.515625,
0.568359375
],
[
0.1376953125,
0.56640625,
0.48779296875,
0.56640625,
0.48779296875,
0.580078125,
0.1376953125,
0.580078125
],
[
0.5146484375,
0.5693359375,
0.8818359375,
0.5693359375,
0.8818359375,
0.58447265625,
0.5146484375,
0.58447265625
],
[
0.12176513671875,
0.5830078125,
0.48779296875,
0.5830078125,
0.48779296875,
0.5966796875,
0.12176513671875,
0.5966796875
],
[
0.5146484375,
0.58447265625,
0.8818359375,
0.5849609375,
0.8818359375,
0.599609375,
0.5146484375,
0.5986328125
],
[
0.12176513671875,
0.599609375,
0.48779296875,
0.599609375,
0.48779296875,
0.61328125,
0.12176513671875,
0.61328125
],
[
0.5146484375,
0.6015625,
0.88330078125,
0.6015625,
0.88330078125,
0.615234375,
0.5146484375,
0.615234375
],
[
0.12176513671875,
0.615234375,
0.48779296875,
0.615234375,
0.48779296875,
0.62890625,
0.12176513671875,
0.62890625
],
[
0.515625,
0.6181640625,
0.6171875,
0.6181640625,
0.6171875,
0.6298828125,
0.515625,
0.6298828125
],
[
0.12176513671875,
0.6298828125,
0.48779296875,
0.6318359375,
0.48779296875,
0.6455078125,
0.12176513671875,
0.6435546875
],
[
0.5341796875,
0.63525390625,
0.8818359375,
0.63525390625,
0.8818359375,
0.6484375,
0.5341796875,
0.6484375
],
[
0.118408203125,
0.646484375,
0.48779296875,
0.646484375,
0.48779296875,
0.66015625,
0.118408203125,
0.66015625
],
[
0.51318359375,
0.6494140625,
0.8818359375,
0.65087890625,
0.8818359375,
0.6650390625,
0.51318359375,
0.66455078125
],
[
0.118408203125,
0.662109375,
0.48779296875,
0.662109375,
0.48779296875,
0.67578125,
0.118408203125,
0.67578125
],
[
0.515625,
0.66748046875,
0.88330078125,
0.66748046875,
0.88330078125,
0.68115234375,
0.515625,
0.68115234375
],
[
0.12005615234375,
0.6796875,
0.345947265625,
0.6796875,
0.345947265625,
0.693359375,
0.12005615234375,
0.693359375
],
[
0.51708984375,
0.68408203125,
0.8818359375,
0.68408203125,
0.8818359375,
0.69775390625,
0.51708984375,
0.69775390625
],
[
0.51708984375,
0.69970703125,
0.880859375,
0.69970703125,
0.880859375,
0.71337890625,
0.51708984375,
0.71337890625
],
[
0.118408203125,
0.7080078125,
0.27783203125,
0.7080078125,
0.27783203125,
0.72265625,
0.118408203125,
0.72265625
],
[
0.5146484375,
0.71533203125,
0.8818359375,
0.71533203125,
0.8818359375,
0.728515625,
0.5146484375,
0.728515625
],
[
0.118408203125,
0.73095703125,
0.48779296875,
0.73095703125,
0.48779296875,
0.74462890625,
0.118408203125,
0.74462890625
],
[
0.497802734375,
0.736328125,
0.50537109375,
0.736328125,
0.50537109375,
0.74072265625,
0.497802734375,
0.74072265625
],
[
0.51318359375,
0.73388671875,
0.53564453125,
0.73388671875,
0.53564453125,
0.74462890625,
0.51318359375,
0.74462890625
],
[
0.12005615234375,
0.74755859375,
0.48779296875,
0.74755859375,
0.48779296875,
0.76123046875,
0.12005615234375,
0.76123046875
],
[
0.5322265625,
0.75,
0.8818359375,
0.75,
0.8818359375,
0.76318359375,
0.5322265625,
0.76318359375
],
[
0.12005615234375,
0.76416015625,
0.291259765625,
0.76416015625,
0.291259765625,
0.77783203125,
0.12005615234375,
0.77783203125
],
[
0.5146484375,
0.76416015625,
0.87890625,
0.76416015625,
0.87890625,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.1376953125,
0.78076171875,
0.489501953125,
0.78076171875,
0.489501953125,
0.7958984375,
0.1376953125,
0.7958984375
],
[
0.5146484375,
0.7822265625,
0.87890625,
0.7822265625,
0.87890625,
0.7919921875,
0.5146484375,
0.7919921875
],
[
0.12176513671875,
0.79736328125,
0.48779296875,
0.79736328125,
0.48779296875,
0.81103515625,
0.12176513671875,
0.81103515625
],
[
0.51318359375,
0.79736328125,
0.880859375,
0.79736328125,
0.880859375,
0.81103515625,
0.51318359375,
0.81103515625
],
[
0.12005615234375,
0.8134765625,
0.48779296875,
0.8134765625,
0.48779296875,
0.82666015625,
0.12005615234375,
0.82666015625
],
[
0.51123046875,
0.8134765625,
0.56103515625,
0.8134765625,
0.56103515625,
0.82763671875,
0.51123046875,
0.82763671875
],
[
0.12005615234375,
0.830078125,
0.4306640625,
0.830078125,
0.4306640625,
0.84326171875,
0.12005615234375,
0.84326171875
],
[
0.5341796875,
0.830078125,
0.8818359375,
0.83056640625,
0.8818359375,
0.845703125,
0.5341796875,
0.84423828125
],
[
0.1376953125,
0.84716796875,
0.486083984375,
0.84716796875,
0.486083984375,
0.86083984375,
0.1376953125,
0.86083984375
],
[
0.5146484375,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.8623046875,
0.5146484375,
0.8623046875
],
[
0.12176513671875,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.12176513671875,
0.876953125
],
[
0.5146484375,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.87744140625,
0.5146484375,
0.87744140625
],
[
0.12005615234375,
0.8798828125,
0.489501953125,
0.8798828125,
0.489501953125,
0.8935546875,
0.12005615234375,
0.8935546875
],
[
0.5146484375,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.118408203125,
0.8935546875,
0.202392578125,
0.8955078125,
0.202392578125,
0.91015625,
0.118408203125,
0.90771484375
],
[
0.515625,
0.8955078125,
0.66162109375,
0.8955078125,
0.66162109375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"(1)",
"Iwasbookedfor/DURATION2nights at/LOCUS this",
"语法错误检测（GED）是识别需要编辑以",
"hotelin/TIMEOct2007",
"产生语法正确的句子的标记的任务。鉴于",
"(2)",
"Iwentto/GOALohmafter/ExPLANATION+TIME",
"reading some ot/QUANTITY+WHOLE the reviews",
"cwrs是从在大量语法文本上训练的模型中",
"(3)It was veryupsetting to see thiskind ot/SPECIES",
"提取的，因此此任务评估嵌入是否对指示其",
"behaviorespeciallyin_front_of/LocUs",
"输入异常的特征进行编码（在这种情况下为",
"my/SoCIALRELGESTALTfouryear_old",
"非语法性）。我们使用第一个英文证书（Yan",
"Annotated sentencesfrom the",
"nakoudakisetal.,2011）数据集，由Reiand",
"Figure2:",
"STREUSLE 4.0 corpus, used in the preposi-",
"Yannakoudakis（2016）转换为序列标记格式",
"tion supersense disambiguation task. Prepositions",
"连词识别（Coni）任务挑战模型，以识别包含",
"aremarkedbyboldface, immediately followed by",
"协调构造中的连词的令牌。这样做需要非常具",
"their labeled function.If applicable,precedes",
"体的语法知识。数据来自Ficlerand Goldberg",
"the preposition's labeled role. Figure reproduced",
"(2016）的协调注释PTB",
"from Schneider et al.(2018)",
"2.3Pairwise Relations",
"介词超语义消歧是对介词的词汇语义贡献",
"我们还设计了探测任务，用于检查单词之间的",
"（功能；PS-fxn）及其介导的语义角色或关系",
"关系是否在cwrs中编码。在这些任务中，给定",
"（角色；PS角色）进行分类的任务。此任务是",
"一个单词对w1，w2，我们将[w1,w2,1w2]",
"种专门的词义消歧，并检查了词汇语义知",
"输入到探测模型中；它经过训练以预测有关标",
"识的一个方面。与上面的标记任务相反，该",
"记（Belinkov，2018）之间关系的信息",
"模型是在单令牌介词上训练和评估的（而不",
"我们区分电弧预测和电弧分类任务。弧预测",
"是为序列中的每个令牌做出决定）。我们使",
"是一个二进制分类任务，其中模型被训练以识",
"用STREUSLE4。0语料库(Schneideretal.,",
"别两个令牌之间是否存在关系。弧分类是一个",
"2018）：例句出现在Figure2中",
"多类分类任务，其中模型具有两个令牌，这些",
"事件真实性（EF）任务涉及用它们描述",
"令牌通过某种关系链接并经过训练以确定它们",
"(Sauri and Pustejovsky.2009,2012:de Marn",
"是如何相关的",
"effeetal.，2012）的事件的真实性来标记短语。",
"例如，在句法依赖弧预测任务中，模型被赋",
"例如，在从Rudingeretal.（2018）复制的以下",
"予两个标记（wa，wb）的表示形式，并被训练以",
"示例中，“（1a）表示没有发生离开，而表面上",
"预测句子的句法依赖解析是否包含以wa为头",
"相似的（1b）则没有\"",
"部，以W6为修饰符的依赖弧。句法依赖弧分类",
"乔不记得离开。。_乔不记得离开了",
"任务向模型呈现两个标记（Whead，Wmod）的表",
"我们使用通用分解语义它发生了v2数据集",
"示，其中Wmod是whead的修饰符，并且模型被",
"(Rudingeretal.，2018），并训练模型以预测",
"训练以预测链接它们的句法关系的类型（该依",
"[-3,3]范围内的（非）事实值。与上面的标记",
"赖弧上的标签)。我们使用PTB（转换为UD）",
"任务不同，此任务被视为回归问题，其中仅对",
"和 UD-EWT",
"与事件相对应的令牌（而不是序列中的每个令",
"同样，语义依赖弧预测训练模型以预测两个",
"牌）进行预测。性能是使用皮尔逊相关性（r）",
"令牌是否通过语义依赖弧连接，而语义依赖",
"来衡量的；我们报告（r×100），因此所有任",
"弧分类任务训练模型以对令牌之间的语义关",
"务的指标都在0到100之间",
"系进行分类。我们将SemEval2015共享任务",
"(Oepenetal.，2015）中的数据集与DELPH-",
"2.2Segmentation",
"INMRS派生语义依赖（DM）目标表示一起使",
"我们的几个探测任务涉及使用BIO或IO标",
"",
"用",
"签进行分段。在这里，模型被训练成仅从单个",
"句法和语义依赖性弧预测和分类任务与用于",
"单词的cwr预测标签",
"语义和句法依赖性解析的最新模型密切相关",
"语法分块（Chunk）测试cwrs是否包含跨",
"该模型对cwrs的对进行评分以做出头部附",
"度和边界的概念；任务是将文本分割为浅组",
"着和弧标记决策（Dozatand Manning，2016",
"成块。我们使用CoNLL2000共享任务数据集",
"2018)",
"(Tjong Kim Sang and Buchholz, 2000)",
"为了生成依赖弧预测任务的反例，我们取",
"命名实体识别（NER）检查cwrs是否编码有",
"每个正例（Whead，Wmod）并生成一个新的反例",
"关实体类型的信息。我们使用CoNLL2003共",
"（wrand，Wmod）。Wrand是句子中的一个随机标",
"享任务数据集（TjongKimSangandDeMeul-",
"记，不是Wmod的头。因此，这些任务中使用",
"der, 2003)",
"的数据集是平衡的"
] |
{
"arxiv_id": "1903.08855",
"language": "zh",
"page_idx": 2
}
|
||
1903.08855-zh-3
|
1903.08855-zh-3.jpg
|
共同引用。我们使用来自 CoNLL2012 共享任务 (Pradhan et al.,2012) 的数据集。为了生成负面示例，我们遵循与依赖弧预测任务类似的过程：给出一个正面示例$(w_a,w_b)$ ,其中$w_b$在$w_a$之后发生，并且两个令牌共享一个共同引用集群，我们创建一个负面示例$( w_{\text{random\_ entity}}, w_b)$ ,其中$w_\text{random\_entity}$是在$w_b$之前发生的令牌，并且属于不同的共同引用集群
## 3 Models
Probing Model 我们使用线性模型作为我们的探测模型；限制其容量使我们能够专注于可以从 cwrs 中轻松提取哪些信息。有关探 基线。实际上，我们经常看到探测模型可以与
测模型训练超参数和其他详细信息，请参见
$\operatorname{Appendix}~??$
Contextualizers 我们研究了六种公开可用
的英语语境化单词表示模型
ELMo (Peters et al., 2018a) 连接在双向语言建模 (biLM) 任务上独立训练的两个上下文化器的输出。ELMo(原始)使用 2层 LSTM 进行上下文化。我们还研究了 Peters et al. (2018b)的两个变体：ELMo(4层)使用4层LSTM,而 ELMo (变压器)使用 6 层变压器(Vaswani et al.,2017)。这些模型中的每一个都在 800M 的句子改组新闻专线文本令牌上进行了训练(10 亿个单词基准；Chelba et al., 2014
OpenAI 转换器 (Radford et al., 2018) 是一个从左到右的 12 层转换器语言模型，训练了来自 7,000 多本未出版的独特书籍 (BookCorpus; Zhu et al., 2015 )的 800M 连续文本标记
BERT ( Devlin et al. , 2018) 使用双向转换器，该转换器在掩码语言建模任务和下一个句子预测任务上进行了联合训练。该模型是在BookCorpus 和英文维基百科上训练的，总共约有 3300M 令牌。我们研究使用 12 层变压器的 BERT(基础，套管)和使用 24 层变压器的 BERT (大，套管)
# Pretrained Contextualizer Comparison
为了更好地理解由预训练的上下文化器捕获的语言知识，我们使用我们的一组探测任务来分析它们的每个层。这些语境化者在许多方面有所不同，控制它们之间的所有差异不在这项工作的范围之内。我们专注于探索可用的模型， 将更系统的培训制度和模型架构比较留给未来的工作
我们还考虑了一个共同引用弧预测任务，其 4.1 Experimental Setup
中模型被训练以预测两个实体是否从其 cwrs 我们的探测模型是根据每个上下文化器的各
个层生成的表示进行训练的。我们还将其与在非上下文向量上训练的线性探测模型进行比较(在有条件的共同爬行上训练的 300 维 GloVe; Pennington et al., 2014 ), 以评估上下文化的收益
## 4.2 Results and Discussion
Table 1 将每个上下文化器的最佳性能探测模型与 GloVe 基线和任务的先前技术水平进行$\overline{\text{比较(不包括使用预训练 cwrs 的方法)。}}^{3,4}$
只需一个线性模型，我们就可以轻松提取各种 NLP 任务的高性能所需的大量信息。在所有情况下，cwrs 的性能都明显优于非上下文最先进的模型(通常经过仔细调整和特定于任务)的性能相媲美或超过其性能。特别是，线性探测模型在语法错误检测和介词超感官识别(角色和功能)方面超过了已发表的最新技术。
比较基于 ELMo 的上下文化器，我们看到ELMo (4 层)和 ELMo (原始) 基本上是均匀的，尽管两种递归模型都优于 ELMo(变压器)。我们还看到，OpenAI 变压器的性能明显低于 ELMo 模型和 BERT。鉴于它也是唯一以单向(从左到右)方式训练的模型，这重申了双向性是最高质量上下文化器(Devlin et al., 2018) 的关键组件。此外，OpenAI 转换器是唯一在小写文本上训练的模型，这阻碍了它在 NER 等任务上的性能。BERT 显著改进了 ELMo 和 OpenAI 模型
我们的探测任务结果表明，cwr 的当前方法并未在其输入中捕获有关实体和共引用现象的许多可传递信息 (e。g。, NER 结果为 Table 1 ,共参考弧预测结果为 Appendix ?? )。为了缓解这一弱点，未来的工作可以使用显式实体表示 (Ji et al., 2017; Yang et al., 2017; Bosselut et al.,2017) 来增强预训练的上下文化器
Probing Failures 虽然探测模型在许多任务中处于或接近最先进的性能，但它们在其他几个任务中的性能也不佳，包括 NER,语法错误检测和联合识别。发生这种情况的原因可能是：(1) cwr 根本不对相关信息或任何预测相关性进行编码，或者(2)探测模型没有从向量中提取信息或预测相关性所需的能力。在前一种情况下，学习特定于任务的上下文特征对于
${\frac{^3\text{See Appen}}{\mathrm{dix~??~for~references~to~the~previous~state}}}$
of the art (without pretraining).
$^{4}$For brevity, in this section we omit probing tasks that cannot be compared to prior work. See Appendix ?? for pretrained contextualizer performance for all layers and all tasks.
|
[
[
0.1376953125,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.1376953125,
0.09027099609375
],
[
0.486083984375,
0.07598876953125,
0.72314453125,
0.07781982421875,
0.72314453125,
0.0926513671875,
0.486083984375,
0.09027099609375
],
[
0.118408203125,
0.0926513671875,
0.48779296875,
0.0926513671875,
0.48779296875,
0.105712890625,
0.118408203125,
0.105712890625
],
[
0.5146484375,
0.097412109375,
0.8818359375,
0.097412109375,
0.8818359375,
0.11102294921875,
0.5146484375,
0.11102294921875
],
[
0.118408203125,
0.1092529296875,
0.48779296875,
0.1092529296875,
0.48779296875,
0.122314453125,
0.118408203125,
0.122314453125
],
[
0.5146484375,
0.1134033203125,
0.8818359375,
0.1134033203125,
0.8818359375,
0.1270751953125,
0.5146484375,
0.1270751953125
],
[
0.12005615234375,
0.1258544921875,
0.489501953125,
0.1258544921875,
0.489501953125,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.515625,
0.1300048828125,
0.8818359375,
0.1300048828125,
0.8818359375,
0.1436767578125,
0.515625,
0.1436767578125
],
[
0.12176513671875,
0.142578125,
0.486083984375,
0.142578125,
0.486083984375,
0.152587890625,
0.12176513671875,
0.152587890625
],
[
0.51318359375,
0.146728515625,
0.884765625,
0.146728515625,
0.884765625,
0.1602783203125,
0.51318359375,
0.1602783203125
],
[
0.12176513671875,
0.157958984375,
0.489501953125,
0.157958984375,
0.489501953125,
0.171630859375,
0.12176513671875,
0.171630859375
],
[
0.51318359375,
0.160888671875,
0.8818359375,
0.162109375,
0.8818359375,
0.177001953125,
0.51318359375,
0.17578125
],
[
0.118408203125,
0.1728515625,
0.486083984375,
0.1728515625,
0.486083984375,
0.1859130859375,
0.118408203125,
0.1859130859375
],
[
0.5146484375,
0.1787109375,
0.55322265625,
0.1787109375,
0.55322265625,
0.191162109375,
0.5146484375,
0.191162109375
],
[
0.1259765625,
0.190673828125,
0.484375,
0.190673828125,
0.484375,
0.20068359375,
0.1259765625,
0.20068359375
],
[
0.118408203125,
0.204833984375,
0.490234375,
0.204833984375,
0.490234375,
0.22265625,
0.118408203125,
0.22265625
],
[
0.515625,
0.2060546875,
0.74462890625,
0.2060546875,
0.74462890625,
0.2161865234375,
0.515625,
0.2161865234375
],
[
0.12005615234375,
0.22265625,
0.48779296875,
0.22265625,
0.48779296875,
0.236328125,
0.12005615234375,
0.236328125
],
[
0.51318359375,
0.223876953125,
0.8818359375,
0.2257080078125,
0.8818359375,
0.2392578125,
0.51318359375,
0.2369384765625
],
[
0.12005615234375,
0.2392578125,
0.17626953125,
0.2392578125,
0.17626953125,
0.25048828125,
0.12005615234375,
0.25048828125
],
[
0.515625,
0.2410888671875,
0.880859375,
0.2410888671875,
0.880859375,
0.254638671875,
0.515625,
0.254638671875
],
[
0.51318359375,
0.254638671875,
0.8818359375,
0.2529296875,
0.8818359375,
0.270751953125,
0.51318359375,
0.2724609375
],
[
0.118408203125,
0.265380859375,
0.21826171875,
0.265380859375,
0.21826171875,
0.276611328125,
0.118408203125,
0.276611328125
],
[
0.53564453125,
0.273681640625,
0.8818359375,
0.273681640625,
0.8818359375,
0.287353515625,
0.53564453125,
0.287353515625
],
[
0.1234130859375,
0.2880859375,
0.48779296875,
0.2880859375,
0.48779296875,
0.3017578125,
0.1234130859375,
0.3017578125
],
[
0.5146484375,
0.289306640625,
0.8818359375,
0.289306640625,
0.8818359375,
0.302734375,
0.5146484375,
0.302734375
],
[
0.12005615234375,
0.305908203125,
0.486083984375,
0.305908203125,
0.486083984375,
0.31640625,
0.12005615234375,
0.31640625
],
[
0.5146484375,
0.305908203125,
0.88330078125,
0.305908203125,
0.88330078125,
0.319580078125,
0.5146484375,
0.319580078125
],
[
0.12176513671875,
0.32177734375,
0.489501953125,
0.32177734375,
0.489501953125,
0.3349609375,
0.12176513671875,
0.3349609375
],
[
0.51318359375,
0.319580078125,
0.8818359375,
0.320556640625,
0.8818359375,
0.3349609375,
0.51318359375,
0.334228515625
],
[
0.12176513671875,
0.33740234375,
0.48779296875,
0.33740234375,
0.48779296875,
0.350830078125,
0.12176513671875,
0.350830078125
],
[
0.51318359375,
0.336181640625,
0.880859375,
0.336181640625,
0.880859375,
0.349853515625,
0.51318359375,
0.349853515625
],
[
0.12005615234375,
0.35400390625,
0.22412109375,
0.3515625,
0.22412109375,
0.366455078125,
0.12005615234375,
0.3681640625
],
[
0.5146484375,
0.352783203125,
0.880859375,
0.352783203125,
0.880859375,
0.366455078125,
0.5146484375,
0.366455078125
],
[
0.51318359375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.51318359375,
0.3818359375
],
[
0.1259765625,
0.3759765625,
0.486083984375,
0.377197265625,
0.486083984375,
0.391357421875,
0.1259765625,
0.39013671875
],
[
0.51318359375,
0.385986328125,
0.8759765625,
0.385986328125,
0.8759765625,
0.399658203125,
0.51318359375,
0.399658203125
],
[
0.1175537109375,
0.392578125,
0.339111328125,
0.392578125,
0.339111328125,
0.40625,
0.1175537109375,
0.40625
],
[
0.53564453125,
0.402099609375,
0.880859375,
0.402099609375,
0.880859375,
0.4150390625,
0.53564453125,
0.4150390625
],
[
0.1368408203125,
0.4091796875,
0.48779296875,
0.4091796875,
0.48779296875,
0.4228515625,
0.1368408203125,
0.4228515625
],
[
0.51318359375,
0.416259765625,
0.8818359375,
0.41748046875,
0.8818359375,
0.431640625,
0.51318359375,
0.43115234375
],
[
0.12005615234375,
0.424560546875,
0.489501953125,
0.42578125,
0.489501953125,
0.440673828125,
0.12005615234375,
0.439453125
],
[
0.5146484375,
0.43408203125,
0.880859375,
0.43408203125,
0.880859375,
0.44775390625,
0.5146484375,
0.44775390625
],
[
0.12176513671875,
0.443603515625,
0.484375,
0.443603515625,
0.484375,
0.45361328125,
0.12176513671875,
0.45361328125
],
[
0.515625,
0.45068359375,
0.8818359375,
0.45068359375,
0.8818359375,
0.46435546875,
0.515625,
0.46435546875
],
[
0.12005615234375,
0.457763671875,
0.489501953125,
0.457763671875,
0.489501953125,
0.471435546875,
0.12005615234375,
0.471435546875
],
[
0.515625,
0.466064453125,
0.8818359375,
0.466064453125,
0.8818359375,
0.479736328125,
0.515625,
0.479736328125
],
[
0.12005615234375,
0.473876953125,
0.489501953125,
0.473876953125,
0.489501953125,
0.48681640625,
0.12005615234375,
0.48681640625
],
[
0.51904296875,
0.482177734375,
0.8818359375,
0.482177734375,
0.8818359375,
0.495361328125,
0.51904296875,
0.495361328125
],
[
0.12005615234375,
0.490478515625,
0.48779296875,
0.490478515625,
0.48779296875,
0.50439453125,
0.12005615234375,
0.50439453125
],
[
0.515625,
0.498779296875,
0.8818359375,
0.498779296875,
0.8818359375,
0.5126953125,
0.515625,
0.5126953125
],
[
0.12176513671875,
0.505859375,
0.48779296875,
0.505859375,
0.48779296875,
0.51953125,
0.12176513671875,
0.51953125
],
[
0.51318359375,
0.515625,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.51318359375,
0.529296875
],
[
0.12005615234375,
0.521484375,
0.489501953125,
0.521484375,
0.489501953125,
0.53515625,
0.12005615234375,
0.53515625
],
[
0.5146484375,
0.53076171875,
0.880859375,
0.53076171875,
0.880859375,
0.54443359375,
0.5146484375,
0.54443359375
],
[
0.12005615234375,
0.5380859375,
0.489501953125,
0.5380859375,
0.489501953125,
0.5517578125,
0.12005615234375,
0.5517578125
],
[
0.5146484375,
0.5458984375,
0.88330078125,
0.54638671875,
0.88330078125,
0.56103515625,
0.5146484375,
0.56005859375
],
[
0.12005615234375,
0.5546875,
0.172119140625,
0.5546875,
0.172119140625,
0.56640625,
0.12005615234375,
0.56640625
],
[
0.51708984375,
0.5634765625,
0.73046875,
0.5634765625,
0.73046875,
0.57666015625,
0.51708984375,
0.57666015625
],
[
0.1376953125,
0.5693359375,
0.47265625,
0.5693359375,
0.47265625,
0.5830078125,
0.1376953125,
0.5830078125
],
[
0.466796875,
0.572265625,
0.484375,
0.572265625,
0.484375,
0.580078125,
0.466796875,
0.580078125
],
[
0.5322265625,
0.5791015625,
0.8818359375,
0.5791015625,
0.8818359375,
0.5927734375,
0.5322265625,
0.5927734375
],
[
0.12005615234375,
0.58740234375,
0.484375,
0.58740234375,
0.484375,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.5146484375,
0.59326171875,
0.880859375,
0.59326171875,
0.880859375,
0.60693359375,
0.5146484375,
0.60693359375
],
[
0.12005615234375,
0.6015625,
0.48779296875,
0.6015625,
0.48779296875,
0.615234375,
0.12005615234375,
0.615234375
],
[
0.515625,
0.61083984375,
0.8818359375,
0.61083984375,
0.8818359375,
0.6259765625,
0.515625,
0.6259765625
],
[
0.12005615234375,
0.6181640625,
0.456787109375,
0.6181640625,
0.456787109375,
0.6318359375,
0.12005615234375,
0.6318359375
],
[
0.52294921875,
0.6259765625,
0.8818359375,
0.626953125,
0.8818359375,
0.64111328125,
0.52294921875,
0.64013671875
],
[
0.1351318359375,
0.63427734375,
0.486083984375,
0.63427734375,
0.486083984375,
0.64794921875,
0.1351318359375,
0.64794921875
],
[
0.5205078125,
0.6318359375,
0.52880859375,
0.6318359375,
0.52880859375,
0.63720703125,
0.5205078125,
0.63720703125
],
[
0.5146484375,
0.6435546875,
0.88330078125,
0.6435546875,
0.88330078125,
0.65673828125,
0.5146484375,
0.65673828125
],
[
0.12176513671875,
0.65283203125,
0.486083984375,
0.65283203125,
0.486083984375,
0.6630859375,
0.12176513671875,
0.6630859375
],
[
0.5146484375,
0.6591796875,
0.8818359375,
0.6591796875,
0.8818359375,
0.6728515625,
0.5146484375,
0.6728515625
],
[
0.12176513671875,
0.66748046875,
0.489501953125,
0.66748046875,
0.489501953125,
0.68115234375,
0.12176513671875,
0.68115234375
],
[
0.51318359375,
0.67578125,
0.8369140625,
0.67578125,
0.8369140625,
0.689453125,
0.51318359375,
0.689453125
],
[
0.12005615234375,
0.68310546875,
0.48779296875,
0.68310546875,
0.48779296875,
0.69677734375,
0.12005615234375,
0.69677734375
],
[
0.118408203125,
0.69775390625,
0.486083984375,
0.69775390625,
0.486083984375,
0.71142578125,
0.118408203125,
0.71142578125
],
[
0.5205078125,
0.70068359375,
0.880859375,
0.70068359375,
0.880859375,
0.71435546875,
0.5205078125,
0.71435546875
],
[
0.118408203125,
0.71337890625,
0.489501953125,
0.71435546875,
0.489501953125,
0.728515625,
0.118408203125,
0.72802734375
],
[
0.5146484375,
0.71630859375,
0.8818359375,
0.71630859375,
0.8818359375,
0.72998046875,
0.5146484375,
0.72998046875
],
[
0.1175537109375,
0.72998046875,
0.295654296875,
0.73095703125,
0.295654296875,
0.7451171875,
0.1175537109375,
0.74462890625
],
[
0.515625,
0.732421875,
0.8818359375,
0.732421875,
0.8818359375,
0.7451171875,
0.515625,
0.7451171875
],
[
0.51708984375,
0.7490234375,
0.88330078125,
0.7490234375,
0.88330078125,
0.7626953125,
0.51708984375,
0.7626953125
],
[
0.12005615234375,
0.75830078125,
0.151123046875,
0.75830078125,
0.151123046875,
0.76953125,
0.12005615234375,
0.76953125
],
[
0.1561279296875,
0.75830078125,
0.38037109375,
0.75830078125,
0.38037109375,
0.771484375,
0.1561279296875,
0.771484375
],
[
0.515625,
0.76416015625,
0.8818359375,
0.76416015625,
0.8818359375,
0.77783203125,
0.515625,
0.77783203125
],
[
0.1485595703125,
0.7724609375,
0.255126953125,
0.77490234375,
0.255126953125,
0.7890625,
0.1485595703125,
0.78759765625
],
[
0.51708984375,
0.78076171875,
0.88330078125,
0.78076171875,
0.88330078125,
0.79443359375,
0.51708984375,
0.79443359375
],
[
0.118408203125,
0.798828125,
0.486083984375,
0.798828125,
0.486083984375,
0.8125,
0.118408203125,
0.8125
],
[
0.515625,
0.79736328125,
0.8818359375,
0.79736328125,
0.8818359375,
0.81103515625,
0.515625,
0.81103515625
],
[
0.12005615234375,
0.8154296875,
0.48779296875,
0.8154296875,
0.48779296875,
0.8291015625,
0.12005615234375,
0.8291015625
],
[
0.51318359375,
0.8134765625,
0.880859375,
0.8134765625,
0.880859375,
0.82666015625,
0.51318359375,
0.82666015625
],
[
0.12005615234375,
0.83056640625,
0.486083984375,
0.83056640625,
0.486083984375,
0.845703125,
0.12005615234375,
0.845703125
],
[
0.5322265625,
0.83447265625,
0.88330078125,
0.8359375,
0.88330078125,
0.85107421875,
0.5322265625,
0.8486328125
],
[
0.12005615234375,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.509765625,
0.845703125,
0.73291015625,
0.84716796875,
0.73291015625,
0.8623046875,
0.509765625,
0.85986328125
],
[
0.12005615234375,
0.8642578125,
0.489501953125,
0.8642578125,
0.489501953125,
0.87744140625,
0.12005615234375,
0.87744140625
],
[
0.5341796875,
0.859375,
0.88330078125,
0.85986328125,
0.88330078125,
0.87353515625,
0.5341796875,
0.87255859375
],
[
0.51318359375,
0.87060546875,
0.81201171875,
0.87255859375,
0.81201171875,
0.88720703125,
0.51318359375,
0.88525390625
],
[
0.80615234375,
0.87451171875,
0.82275390625,
0.87451171875,
0.82275390625,
0.8818359375,
0.80615234375,
0.8818359375
],
[
0.81689453125,
0.87158203125,
0.880859375,
0.87353515625,
0.880859375,
0.88525390625,
0.81689453125,
0.8828125
],
[
0.118408203125,
0.87744140625,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.118408203125,
0.892578125
],
[
0.515625,
0.88525390625,
0.88330078125,
0.88525390625,
0.88330078125,
0.8984375,
0.515625,
0.8984375
],
[
0.118408203125,
0.8955078125,
0.17626953125,
0.8955078125,
0.17626953125,
0.91015625,
0.118408203125,
0.91015625
],
[
0.5146484375,
0.896484375,
0.6708984375,
0.896484375,
0.6708984375,
0.91015625,
0.5146484375,
0.91015625
]
] |
[
"我们还考虑了一个共同引用弧预测任务，其",
"4.1ExperimentalSetup",
"中模型被训练以预测两个实体是否从其cwrs",
"我们的探测模型是根据每个上下文化器的各",
"共同引用。我们使用来自CoNLL2012共享任",
"个层生成的表示进行训练的。我们还将其与在",
"务（Pradhanetal.，2012）的数据集。为了生",
"非上下文向量上训练的线性探测模型进行比较",
"成负面示例，我们遵循与依赖弧预测任务类",
"（在有条件的共同爬行上训练的300维GloVe；",
"似的过程：给出一个正面示例（wa,Wb），其",
"Penningtonetal.，2014），以评估上下文化的",
"中Wh在Wa之后发生，并且两个令牌共享",
"收益",
"一个共同弓用集群，我们创建一个负面示例",
"(Wrandom_entity,Wb)，其中 Wrandom_entity 是在",
"4.2ResultsandDiscussion",
"Wb之前发生的令牌，并且属于不同的共同引",
"Table1将每个上下文化器的最佳性能探测模",
"用集群",
"型与GloVe基线和任务的先前技术水平进行",
"比较（不包括使用预训练cwrs的方法）。3，4",
"3Models",
"只需一个线性模型，我们就可以轻松提取各",
"ProbingModel我们使用线性模型作为我们",
"种NLP任务的高性能所需的大量信息。在所",
"的探测模型：限制其容量使我们能够专注于",
"有情况下，cwrs的性能都明显优于非上下文",
"可以从cwrs中轻松提取哪些信息。有关探",
"基线。实际上，我们经常看到探测模型可以与",
"测模型训练超参数和其他详细信息，请参见",
"最先进的模型（通常经过仔细调整和特定于任",
"Appendix ??",
"务）的性能相媲美或超过其性能。特别是，线",
"性探测模型在语法错误检测和介词超感官识别",
"Contextualizers我们研究了六种公开可用",
"（角色和功能）方面超过了已发表的最新技术。",
"的英语语境化单词表示模型",
"比较基于ELMo的上下文化器，我们看到",
"ELMo（Petersetal.，2018a）连接在双向语",
"ELMo（4层）和ELMo（原始）基本上是均",
"言建模（biLM）任务上独立训练的两个上下文",
"匀的，尽管两种递归模型都优于ELMo（变压",
"化器的输出。ELMo（原始）使用2层LSTM",
"器）。我们还看到，OpenAI变压器的性能明",
"进行上下文化。我们还研究了Peterset al.",
"显低于ELMo模型和BERT。鉴于它也是唯",
"(2018b）的两个变体：ELMo（4层）使用4层",
"一以单向（从左到右）方式训练的模型，这",
"LSTM，而ELMo（变压器）使用6层变压器",
"重申了双向性是最高质量上下文化器（Devlin",
"(Vaswanietal.，2017）。这些模型中的每一个",
"etal.，2018）的关键组件。此外，OpenAI转换",
"都在800M的句子改组新闻专线文本令牌上",
"器是唯一在小写文本上训练的模型，这阻碍了",
"进行了训练（10亿个单词基准；Chelbaetal.,",
"它在NER等任务上的性能。BERT显著改进",
"2014",
"了ELMo和 OpenAI模型",
"OpenAI转换器（Radfordetal.,2018）是",
"一",
"我们的探测任务结果表明，cwr的当前方法",
"个从左到右的12层转换器语言模型，训练了来",
"并未在其输入中捕获有关实体和共引用现象的",
"自7,000多本未出版的独特书籍（BookCorpus;",
"许多可传递信息（e。g。，NER结果为Table1",
"Zhuetal.，2015）的800M连续文本标记",
"共参考弧预测结果为Appendix？?？）。为了缓",
"BERT（Devlinetal.，2018）使用双向转换",
"",
"解这一弱点，未来的工作可以使用显式实体表",
"器，该转换器在掩码语言建模任务和下一个",
" (Ji et al., 2017; Yang et al., 2017; Bosselut",
"句子预测任务上进行了联合训练。该模型是在",
"etal.,2017）来增强预训练的上下文化器",
"BookCorpus和英文维基百科上训练的，总共",
"约有3300M令牌。我们研究使用12层变压器",
"ProbingFailures虽然探测模型在许多任务",
"的BERT（基础，套管）和使用24层变压器",
"中处于或接近最先进的性能，但它们在其他几",
"的BERT（大，套管）",
"个任务中的性能也不佳，包括NER，语法错",
"误检测和联合识别。发生这种情况的原因可能",
"4",
"Pretrained Contextualizer",
"是：（1）cwr根本不对相关信息或任何预测相",
"Comparison",
"关性进行编码，或者（2）探测模型没有从向量",
"为了更好地理解由预训练的上下文化器捕获的",
"中提取信息或预测相关性所需的能力。在前一",
"语言知识，我们使用我们的一组探测任务来分",
"种情况下，学习特定于任务的上下文特征对于",
"析它们的每个层。这些语境化者在许多方面有",
"3See Appendix?? for references to the previous state",
"所不同，控制它们之间的所有差异不在这项工",
"of the art (without pretraining).",
"作的范围之内。我们专注于探索可用的模型，",
"4For brevity, in this section we omit probing tasks",
"that cannot be compared to prior work.",
"",
"See Ap-",
"将更系统的培训制度和模型架构比较留给未来",
"pendix ?? for pretrained contextualizer performance for",
"的工作",
"all layers and all tasks."
] |
{
"arxiv_id": "1903.08855",
"language": "zh",
"page_idx": 3
}
|
||
1903.09588-zh-0
|
1903.09588-zh-0.jpg
|
通过构造辅助句子利用 BERT 进行基于方面的情感分析
Chi Sun, Luyao Huang, Xipeng Qiu$^*$
Shanghai Key Laboratory of Intelligent Information Processing, Fudan University
School of Computer Science, Fudan University
$825{\mathrm{~Zhangheng~Road,~Shanghai,~China}}$
$\{$ sunc17,lyhuang18,xpqiu $\}$ @fudan.edu.cn
$\operatorname{Abstract}$
Aspect-based sentiment analysis (ABSA), which aims to identify fne-grained opinion polarity towards a specific aspect, is a challenging subtask of sentiment analysis (SA). In this paper, we construct an auxiliary sentence from the aspect and convert ABSA to a sentence-pair classification task, such as question answering (QA) and natural language inference (NLI). We finetune the pre-trained model from BERT and achieve new state-of-the-art results on SentiHood and SemEval-2014 Task 4 datasets. 1
### 1 Introduction
情感分析(SA)是自然语言处理中的重要任务。它解决了意见，情感和主观性的计算处界而且在工业界都受到了广泛关注，通过亚马逊等网站上的在线评论提供实时反馈，可以利用客户对特定产品或服务的意见。此任务的基本假设是整个文本具有总体极性。但是， 用户的评论可能包含不同的方面，例如：“这本书是精装版，但价格有点高。“,外观”的极性是正的，而’价格’的极性是负的。基于方面的情感分析(ABSA)(JoandOh,2011; Pontiki et al., 2014, 2015, 2016) 旨在识别针对特定方面的细粒度极性。此任务允许用户评估给定产品或服务的每个方面的汇总情绪，并更详细地了解其质量。SA 和 ABSA 都是句子级或文档级任务，但是一个注释可能引用多个对象，并且句子级任务无法处理具有多个目标的句子。因此，Saeidi et al.(2016)引人了基于目标方面的情感分析(TABSA)的任务，该任务旨在识别与给定目标相关的特定方面的细粒度意见极性。任务可以分为两个步骤：(1)第一步是确定与每个目标相关的方
$^{*}$Corresponding author.
$^{1}$The source codes are available at https://github.
com/HSLCY/ABSA-BERT-pair
面；(2)第二步是将方面的极性解析为给定目标(T) ABSA 最早的工作主要依赖于特征工程 (Wagner et al., 2014; Kiritchenko et al., 2014) ,随后基于神经网络的方法 (Nguyen and Shirai, 2015; Wang et al., 2016; Tang et al., 2015, 2016; Wang et al., 2017) 实现了更高的精度。最近，Ma et al. (2018) 将有用的常识知识纳人深度神经网络，以进一步增强模型的结果。Liu et al. (2018) 优化内存网络并将其应用于其模型以更好地捕获语言结构
最近，预训练的语言模型，如 ELMo (Peters et al., 2018) 、OpenAIGPT (Radford et al., 2018)和 BERT (Devlin et al., 2018),已经显示示出它们在减轻特征工程工作方面的有效性。特别是，BERT 在 QA 和 NLI 方面取得了优异的成绩。但是，直接使用预训练的 BERT 模型在(T) ABSA 任务中没有太大改进(请参见
理-情感被收集，分析和总结。它不仅在学术 表 3 )。我们认为这是由于预先训练的 BERT
模型使用不当所致
由于 BERT 的输入表示既可以表示单个文本句子，也可以表示一对文本句子，因此我们可以将(T) ABSA 转换为句子对分类任务，并微调预训练的 BERT。在本文中，我们研究了几种构造辅助句子的方法，并将(T) ABSA 转换为句子对分类任务。我们微调来自 BERT 的预训练模型，并在 (T)ABSA 任务上实现新的最新结果。我们还进行了比较实验，以验证基于句子对的分类优于具有微调 BERT 的单句分类，这意味着改进不仅来自 BERT,而且来自我们的方法。特别是，我们的贡献是双重的： 1。我们提出了一种新的解决方案(T)ABSA 通过将其转换为句子对分类任务。2。我们微调预训练的 BERT 模型，并在 SentiHood 和SemEval-2014 任务 4 数据集上获得最新的结果。
### 2 Methodologv
在本节中，我们将详细描述我们的方法
|
[
[
0.210693359375,
0.08612060546875,
0.791015625,
0.08612060546875,
0.791015625,
0.1009521484375,
0.210693359375,
0.1009521484375
],
[
0.35107421875,
0.133056640625,
0.66650390625,
0.133056640625,
0.66650390625,
0.1478271484375,
0.35107421875,
0.1478271484375
],
[
0.1502685546875,
0.1484375,
0.8564453125,
0.149658203125,
0.8564453125,
0.1644287109375,
0.1502685546875,
0.163330078125
],
[
0.30322265625,
0.1650390625,
0.70361328125,
0.1650390625,
0.70361328125,
0.179931640625,
0.30322265625,
0.179931640625
],
[
0.334228515625,
0.1817626953125,
0.6708984375,
0.1817626953125,
0.6708984375,
0.1953125,
0.334228515625,
0.1953125
],
[
0.31982421875,
0.198974609375,
0.6884765625,
0.19775390625,
0.6884765625,
0.2120361328125,
0.31982421875,
0.213134765625
],
[
0.26611328125,
0.265380859375,
0.343505859375,
0.265380859375,
0.343505859375,
0.27978515625,
0.26611328125,
0.27978515625
],
[
0.51708984375,
0.2666015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.28076171875,
0.51708984375,
0.28076171875
],
[
0.51708984375,
0.283203125,
0.8818359375,
0.283203125,
0.8818359375,
0.29638671875,
0.51708984375,
0.29638671875
],
[
0.1502685546875,
0.29150390625,
0.459228515625,
0.29150390625,
0.459228515625,
0.3046875,
0.1502685546875,
0.3046875
],
[
0.5146484375,
0.297607421875,
0.87890625,
0.297607421875,
0.87890625,
0.311279296875,
0.5146484375,
0.311279296875
],
[
0.1485595703125,
0.305908203125,
0.45751953125,
0.305908203125,
0.45751953125,
0.319580078125,
0.1485595703125,
0.319580078125
],
[
0.51318359375,
0.3134765625,
0.8818359375,
0.314208984375,
0.8818359375,
0.328857421875,
0.51318359375,
0.327880859375
],
[
0.1485595703125,
0.32177734375,
0.456787109375,
0.32177734375,
0.456787109375,
0.33203125,
0.1485595703125,
0.33203125
],
[
0.51318359375,
0.330078125,
0.880859375,
0.330078125,
0.880859375,
0.34326171875,
0.51318359375,
0.34326171875
],
[
0.1502685546875,
0.336181640625,
0.45751953125,
0.336181640625,
0.45751953125,
0.3466796875,
0.1502685546875,
0.3466796875
],
[
0.14697265625,
0.347412109375,
0.45166015625,
0.3486328125,
0.45166015625,
0.3623046875,
0.14697265625,
0.361083984375
],
[
0.5146484375,
0.3466796875,
0.8818359375,
0.3466796875,
0.8818359375,
0.35986328125,
0.5146484375,
0.35986328125
],
[
0.1485595703125,
0.363525390625,
0.45751953125,
0.363525390625,
0.45751953125,
0.377197265625,
0.1485595703125,
0.377197265625
],
[
0.5146484375,
0.363525390625,
0.8818359375,
0.363525390625,
0.8818359375,
0.377197265625,
0.5146484375,
0.377197265625
],
[
0.14697265625,
0.3759765625,
0.45751953125,
0.377197265625,
0.45751953125,
0.391357421875,
0.14697265625,
0.39013671875
],
[
0.5146484375,
0.37890625,
0.8818359375,
0.377685546875,
0.8818359375,
0.392578125,
0.5146484375,
0.393798828125
],
[
0.1502685546875,
0.392578125,
0.31494140625,
0.392578125,
0.31494140625,
0.402587890625,
0.1502685546875,
0.402587890625
],
[
0.307373046875,
0.392578125,
0.456787109375,
0.392578125,
0.456787109375,
0.402587890625,
0.307373046875,
0.402587890625
],
[
0.515625,
0.3955078125,
0.8818359375,
0.3955078125,
0.8818359375,
0.4091796875,
0.515625,
0.4091796875
],
[
0.14697265625,
0.405029296875,
0.456787109375,
0.40380859375,
0.456787109375,
0.418701171875,
0.14697265625,
0.419189453125
],
[
0.51708984375,
0.412109375,
0.791015625,
0.412109375,
0.791015625,
0.42578125,
0.51708984375,
0.42578125
],
[
0.145263671875,
0.418701171875,
0.459228515625,
0.41748046875,
0.459228515625,
0.431640625,
0.145263671875,
0.432861328125
],
[
0.1502685546875,
0.435302734375,
0.45751953125,
0.435302734375,
0.45751953125,
0.4453125,
0.1502685546875,
0.4453125
],
[
0.5341796875,
0.43115234375,
0.8818359375,
0.43115234375,
0.8818359375,
0.44482421875,
0.5341796875,
0.44482421875
],
[
0.1485595703125,
0.44775390625,
0.4609375,
0.44775390625,
0.4609375,
0.46142578125,
0.1485595703125,
0.46142578125
],
[
0.51123046875,
0.446533203125,
0.880859375,
0.4453125,
0.880859375,
0.460205078125,
0.51123046875,
0.46142578125
],
[
0.1485595703125,
0.463134765625,
0.2301025390625,
0.463134765625,
0.2301025390625,
0.474365234375,
0.1485595703125,
0.474365234375
],
[
0.5146484375,
0.4619140625,
0.8818359375,
0.463134765625,
0.8818359375,
0.47802734375,
0.5146484375,
0.476806640625
],
[
0.515625,
0.479736328125,
0.880859375,
0.479736328125,
0.880859375,
0.493408203125,
0.515625,
0.493408203125
],
[
0.118408203125,
0.488037109375,
0.156982421875,
0.488037109375,
0.156982421875,
0.49951171875,
0.118408203125,
0.49951171875
],
[
0.1502685546875,
0.48681640625,
0.264404296875,
0.48681640625,
0.264404296875,
0.501953125,
0.1502685546875,
0.501953125
],
[
0.5146484375,
0.495361328125,
0.8818359375,
0.495361328125,
0.8818359375,
0.5087890625,
0.5146484375,
0.5087890625
],
[
0.118408203125,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.525390625,
0.118408203125,
0.525390625
],
[
0.515625,
0.51318359375,
0.87890625,
0.51318359375,
0.87890625,
0.52392578125,
0.515625,
0.52392578125
],
[
0.12005615234375,
0.529296875,
0.48779296875,
0.529296875,
0.48779296875,
0.5419921875,
0.12005615234375,
0.5419921875
],
[
0.5146484375,
0.52783203125,
0.8818359375,
0.52783203125,
0.8818359375,
0.54150390625,
0.5146484375,
0.54150390625
],
[
0.12005615234375,
0.54443359375,
0.489501953125,
0.54443359375,
0.489501953125,
0.55810546875,
0.12005615234375,
0.55810546875
],
[
0.515625,
0.54443359375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55810546875,
0.515625,
0.55810546875
],
[
0.12176513671875,
0.56103515625,
0.486083984375,
0.56103515625,
0.486083984375,
0.57177734375,
0.12176513671875,
0.57177734375
],
[
0.51318359375,
0.56005859375,
0.6630859375,
0.56005859375,
0.6630859375,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.12005615234375,
0.57666015625,
0.48779296875,
0.57666015625,
0.48779296875,
0.59033203125,
0.12005615234375,
0.59033203125
],
[
0.5341796875,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59423828125,
0.5341796875,
0.59423828125
],
[
0.12005615234375,
0.59326171875,
0.48779296875,
0.59326171875,
0.48779296875,
0.60693359375,
0.12005615234375,
0.60693359375
],
[
0.51318359375,
0.59423828125,
0.8818359375,
0.595703125,
0.8818359375,
0.60986328125,
0.51318359375,
0.609375
],
[
0.12176513671875,
0.609375,
0.489501953125,
0.609375,
0.489501953125,
0.62255859375,
0.12176513671875,
0.62255859375
],
[
0.515625,
0.6123046875,
0.8818359375,
0.6123046875,
0.8818359375,
0.6259765625,
0.515625,
0.6259765625
],
[
0.12176513671875,
0.6259765625,
0.489501953125,
0.6259765625,
0.489501953125,
0.64013671875,
0.12176513671875,
0.64013671875
],
[
0.5146484375,
0.62744140625,
0.8818359375,
0.62744140625,
0.8818359375,
0.64111328125,
0.5146484375,
0.64111328125
],
[
0.12005615234375,
0.64013671875,
0.489501953125,
0.64013671875,
0.489501953125,
0.65478515625,
0.12005615234375,
0.65478515625
],
[
0.515625,
0.64453125,
0.8818359375,
0.64453125,
0.8818359375,
0.65771484375,
0.515625,
0.65771484375
],
[
0.12005615234375,
0.65771484375,
0.48779296875,
0.65771484375,
0.48779296875,
0.67138671875,
0.12005615234375,
0.67138671875
],
[
0.5146484375,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.5146484375,
0.67333984375
],
[
0.12005615234375,
0.67333984375,
0.489501953125,
0.67333984375,
0.489501953125,
0.68701171875,
0.12005615234375,
0.68701171875
],
[
0.5146484375,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.5146484375,
0.6904296875
],
[
0.12005615234375,
0.689453125,
0.48779296875,
0.689453125,
0.48779296875,
0.703125,
0.12005615234375,
0.703125
],
[
0.5146484375,
0.693359375,
0.8818359375,
0.693359375,
0.8818359375,
0.70703125,
0.5146484375,
0.70703125
],
[
0.12176513671875,
0.705078125,
0.48779296875,
0.705078125,
0.48779296875,
0.71875,
0.12176513671875,
0.71875
],
[
0.515625,
0.708984375,
0.87890625,
0.708984375,
0.87890625,
0.7197265625,
0.515625,
0.7197265625
],
[
0.12005615234375,
0.7216796875,
0.489501953125,
0.7216796875,
0.489501953125,
0.7353515625,
0.12005615234375,
0.7353515625
],
[
0.515625,
0.7255859375,
0.87890625,
0.7255859375,
0.87890625,
0.736328125,
0.515625,
0.736328125
],
[
0.12176513671875,
0.73681640625,
0.489501953125,
0.73681640625,
0.489501953125,
0.75048828125,
0.12176513671875,
0.75048828125
],
[
0.51708984375,
0.7412109375,
0.884765625,
0.7412109375,
0.884765625,
0.7548828125,
0.51708984375,
0.7548828125
],
[
0.12005615234375,
0.75439453125,
0.489501953125,
0.75439453125,
0.489501953125,
0.76708984375,
0.12005615234375,
0.76708984375
],
[
0.5146484375,
0.75830078125,
0.8818359375,
0.75830078125,
0.8818359375,
0.771484375,
0.5146484375,
0.771484375
],
[
0.118408203125,
0.7685546875,
0.484375,
0.7685546875,
0.484375,
0.7822265625,
0.118408203125,
0.7822265625
],
[
0.515625,
0.77392578125,
0.880859375,
0.77392578125,
0.880859375,
0.78369140625,
0.515625,
0.78369140625
],
[
0.12176513671875,
0.7861328125,
0.48779296875,
0.7861328125,
0.48779296875,
0.7998046875,
0.12176513671875,
0.7998046875
],
[
0.5146484375,
0.7890625,
0.8818359375,
0.7890625,
0.8818359375,
0.802734375,
0.5146484375,
0.802734375
],
[
0.1234130859375,
0.8017578125,
0.489501953125,
0.8017578125,
0.489501953125,
0.8154296875,
0.1234130859375,
0.8154296875
],
[
0.51123046875,
0.80419921875,
0.8818359375,
0.8046875,
0.8818359375,
0.8193359375,
0.51123046875,
0.818359375
],
[
0.12005615234375,
0.818359375,
0.48779296875,
0.818359375,
0.48779296875,
0.83203125,
0.12005615234375,
0.83203125
],
[
0.5146484375,
0.822265625,
0.5400390625,
0.822265625,
0.5400390625,
0.83447265625,
0.5146484375,
0.83447265625
],
[
0.12005615234375,
0.8349609375,
0.486083984375,
0.8349609375,
0.486083984375,
0.8486328125,
0.12005615234375,
0.8486328125
],
[
0.12005615234375,
0.85107421875,
0.486083984375,
0.85107421875,
0.486083984375,
0.8642578125,
0.12005615234375,
0.8642578125
],
[
0.509765625,
0.85791015625,
0.6630859375,
0.86083984375,
0.6630859375,
0.876953125,
0.509765625,
0.87353515625
],
[
0.15283203125,
0.87158203125,
0.30908203125,
0.87158203125,
0.30908203125,
0.88525390625,
0.15283203125,
0.88525390625
],
[
0.139404296875,
0.8818359375,
0.486083984375,
0.88427734375,
0.486083984375,
0.8984375,
0.139404296875,
0.896484375
],
[
0.12005615234375,
0.896484375,
0.341796875,
0.896484375,
0.341796875,
0.91015625,
0.12005615234375,
0.91015625
],
[
0.515625,
0.8955078125,
0.8271484375,
0.8955078125,
0.8271484375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"通过构造辅助句子利用BERT进行基于方面的情感分析",
"Chi Sun, Luyao Huang, Xipeng Qiu*",
"Shanghai Key Laboratory of Intelligent Information Processing, Fudan University",
"School of Computer Science, Fudan University",
"825 Zhangheng Road, Shanghai, China",
"[ sunc17,lyhuang18,xpqiu 1 @fudan.edu.cn",
"Abstract",
"面；（2）第二步是将方面的极性解析为给定",
"目标（T）ABSA最早的工作主要依赖于特征",
"Aspect-based sentiment analysis (ABSA),",
"工程 (Wagner et al., 2014; Kiritchenko et al.",
"which aims to identify fine-grained opin-",
"2014），随后基于神经网络的方法（Nguyenand",
"ion polarity towards aspecific aspect,is a",
"Shirai, 2015: Wang et al., 2016; Tang et al..",
"challenging subtask of sentiment analysis",
"(SA). In this paper, we construct an aux",
"2015，2016；Wangetal，2017）实现了更高的",
"iliary sentence from the aspect and con-",
"精度。最近，Maetal.（2018）将有用的常识知",
"vert ABSA to a sentence-pair classification",
"识纳入深度神经网络，以进一步增强模型的结",
"task,suchas question",
"answering(QA)and",
"果。Liuetal.（2018）优化内存网络并将其应",
"natural language inference (NLI). We fine-",
"用于其模型以更好地捕获语言结构",
"tune the pre-trained model from BERT",
"and achieve new state-of-the-art results",
"最近，预训练的语言模型，如ELMo（Peters",
"on SentiHood and SemEval-2014 Task 4",
"et al., 2018)、OpenAIGPT (Radford et al.",
"datasets. 1",
"2018）和BERT（Devlinetal.，2018），已经显",
"示出它们在减轻特征工程工作方面的有效性。",
"1",
"Introduction",
"特别是，BERT在QA和NLI方面取得了优",
"情感分析（SA）是自然语言处理中的重要任",
"异的成绩。但是，直接使用预训练的BERT模",
"务。它解决了意见，情感和主观性的计算处",
"型在（T）ABSA任务中没有太大改进（请参见",
"理-情感被收集，分析和总结。它不仅在学术",
"表3）。我们认为这是由于预先训练的BERT",
"界而月在工业界都受到了广泛关注，通过亚",
"模型使用不当所致",
"马逊等网站上的在线评论提供实时反馈，可",
"由于BERT的输入表示既可以表示单个文",
"以利用客户对特定产品或服务的意见。此任务",
"本句子，也可以表示一对文本句子，因此我们",
"的基本假设是整个文本具有总体极性。但是，",
"可以将（T）ABSA转换为句子对分类任务，并",
"用户的评论可能包含不同的方面，例如：“这",
"微调预训练的BERT。在本文中，我们研究了",
"本书是精装版，但价格有点高。“’外观”的",
"几种构造辅助句子的方法，并将（T）ABSA转",
"极性是正的，而”价格”的极性是负的。基于",
"换为句子对分类任务。我们微调来自BERT的",
"方面的情感分析（ABSA）（JoandOh，2011;",
"预训练模型，并在（T)ABSA任务上实现新的",
"Pontiki et al.，2014，2015，2016）旨在识别针",
"最新结果。我们还进行了比较实验，以验证基",
"对特定方面的细粒度极性。此任务允许用户评",
"于句子对的分类优于具有微调BERT的单右",
"估给定产品或服务的每个方面的汇总情绪，并",
"分类，这意味着改进不仅来自BERT，而目来",
"更详细地了解其质量。SA和ABSA都是句子",
"自我们的方法。特别是，我们的贡献是双重的：",
"级或文档级任务，但是一个注释可能引用多",
"1。我们提出了一种新的解决方案(T)ABSA",
"个对象，并且句子级任务无法处理具有多个",
"通过将其转换为句子对分类任务。2。我们微",
"目标的句子。因此，Saeidietal.（2016）引人",
"调预训练的BERT模型，并在SentiHood和",
"了基于目标方面的情感分析（TABSA）的任",
"SemEval-2014任务4数据集上获得最新的结",
"务，该任务旨在识别与给定目标相关的特定",
"果。",
"方面的细粒度意见极性。任务可以分为两个步",
"骤：（1）第一步是确定与每个目标相关的方",
"2Methodology",
"Corresponding author",
"IThe source codes are available at https://github.",
"com/HSLCY/ABSA-BERT-pair",
"在本节中，我们将详细描述我们的方法"
] |
{
"arxiv_id": "1903.09588",
"language": "zh",
"page_idx": 0
}
|
||
1903.09588-zh-1
|
1903.09588-zh-1.jpg
|
### 2.1 Task description
TABSA 在 TABSA 中，一个句子$s$通常由一系列单词组成$:\{w_1,\cdots,w_m\}$ ,其中一些单词$\{w_{i_1},\cdots,w_{i_k}\}$是预先识别的目标$\{t_1,\cdots,t_k\}$ ,在 Saeidi et al.(2016)之后，我们将任务设置为 3 类分类问题：给定句子 $s,-$组XMATHX 实体。正如我们在表 1 中看到的那样，(LOCATION2,价格)的黄金标准极性为负，而(LOCATION1,价格)的极性为无
ABSA 在 ABSA 中，目标-方面对$\{t,a\}$仅成为方面$a$ 。此设置等效于同时学习 SemEval- 2014 任务$4^{2}$的子任务 3(方面类别检测)和子任务 4(方面类别极性)
对于目标-方面对(LOCATION1,安全)的集合，我们生成的句子是“您如何看待位置-1 的安全性？"
Sentences for NLI-M 对于 NLI 任务，我们在生成句子时设置的条件不那么严格，形式也简单得多。此时创建的句子不是标准句子，而是简单的伪句子，以(LOCATION1,safety) 对为例：辅助句子是：“位置-1-safety"。
Sentences for QA-B 对于 QA-B ,我们添加标签信息，并暂时将 TABSA 转换为二进制分类问题$(label\in\{yes,no\})$,以获得概率分布。此时，每个目标-方面对将生成三个序列，例如“位置-1 的方面安全的极性为正”,“位置-1 的方面安全的极性为负”,“位置-1 的方面安全的极性为无”。我们使用yes的概率值作为匹配分数。对于生成三个序列( $positive,negative,none$ ) 的目标-方面对，我们将预测类别的匹配得分最高的序列类别。
Sentences for NLI-B NLI-B 和 QA-B 之间的区别在于，辅助句子从问题变为伪句子。辅助句子是：“位置-1-安全-阳性”,“位置-1-安全-阴性”和“位置-1-安全-无”在我们构造辅助句子之后，我们可以将 TABSA 任务从单个句子分类任务转换为句子对分类任务。如表3所示，这是一项必要的操作，可以显着改善TABSA 任务的实验结果。
$\operatorname{Example:}$
LOCATION2 is central London so extremely expensive, LOCATION1 is often considered the coolest area of London.
<table>
<tbody>
<tr>
<th>Target</th>
<th>$\operatorname{Aspect}$</th>
<th>Sentiment</th>
</tr>
<tr>
<td>$LOC1$</td>
<td>general</td>
<td>Positive</td>
</tr>
<tr>
<td>$LOC1$</td>
<td>price</td>
<td>None</td>
</tr>
<tr>
<td>LOC1</td>
<td>safety</td>
<td>None</td>
</tr>
<tr>
<td>$LWVL$ LOC2</td>
<td>$p11ct$ safety</td>
<td>$\operatorname{lVegaulve}$ None</td>
</tr>
<tr>
<td>LOC2</td>
<td>transit- location</td>
<td>Positive</td>
</tr>
</tbody>
</table>
2.3 Fine-tuning pre-trained BERT
Table 1: An example of SentiHood dataset.
2.2 Construction of the auxiliary sentence
BERT (Devlin et al., 2018) 是一种新的语言表示模型，它使用双向转换器对大型语料库进行预训练，并在其他任务上对预训练的模型进行微调。我们在 TABSA 任务上微调预训练的BERT 模型。让我们简要看看输入表示和微调过程。
为简单起见，我们主要以 TABSA 为例描述我们的方法。我们考虑以下四种方法将 TABSA 任务转换为句子对分类任务：
## 2.3.1 Input representation
<table>
<tbody>
<tr>
<th>Methods</th>
<th>Output</th>
<th>AuxiliarvS Sentence BERT 的输入表示可以 显式表示今牌仔列中</th>
</tr>
<tr>
<td>QA- M</td>
<td>S. P. </td>
<td>Ouestio VH $\mathbb{R}-\mathbb{X}$计文$\mathbb{Z}$ $x_{1}^{-2}$ 给定的今牌、耳输入表</td>
</tr>
<tr>
<td>NLI- M</td>
<td>S. P. </td>
<td>$J_{\mathbf{c}\boldsymbol{\rho}}$ 机位置嵌入求机采</td>
</tr>
<tr>
<td>$QA-B$</td>
<td>yes, no</td>
<td>$\max/$ $SP$ 构造的。对十分类件举 ${\text{舟个怪列的第}}$ =-//中</td>
</tr>
<tr>
<td>NLI- B</td>
<td>ves, no</td>
<td>$\mathrm{P}_{\mathrm{Selld}0}$ $\left(\begin{array}{c}{1:1}\\{1:1}\end{array}\right)$</td>
</tr>
</tbody>
</table>
${\overline {2. 3}}. 2$ Fine-tuning procedure
Table 2: The construction methods. Due to limited space, we use the following abbreviations: S.P. for sentiment polarity , w/o for without , and w/ for with .
BERT 微调非常简单。为了获得输入序列的固定维池化表示，我们使用最终的隐藏状态(i。e。,变压器的输出)的第一个令牌作为输入。我们将向量表示为$C\in\mathbb{R}^H$ 。然后我们添加一个分类层，其参数矩阵为$W\in\mathbb{R}^K\times H$, 其中$K$是类别数。最后，通过 softmax 函数$P=softmax(CW^\mathrm{T})$计算每个类别$P$的概率。
Sentences for QA-M 我们要从目标-方面对生成的句子是一个问题，格式需要相同。例如
$^{2}$http://alt.qcri.org/semeval2014/task4/
|
[
[
0.118408203125,
0.07659912109375,
0.301513671875,
0.07659912109375,
0.301513671875,
0.09027099609375,
0.118408203125,
0.09027099609375
],
[
0.515625,
0.07598876953125,
0.8818359375,
0.07598876953125,
0.8818359375,
0.09027099609375,
0.515625,
0.09027099609375
],
[
0.1234130859375,
0.0955810546875,
0.48779296875,
0.0968017578125,
0.48779296875,
0.11102294921875,
0.1234130859375,
0.10986328125
],
[
0.515625,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.515625,
0.10687255859375
],
[
0.12005615234375,
0.11224365234375,
0.490234375,
0.11224365234375,
0.490234375,
0.1300048828125,
0.12005615234375,
0.1300048828125
],
[
0.515625,
0.10986328125,
0.5859375,
0.10986328125,
0.5859375,
0.1217041015625,
0.515625,
0.1217041015625
],
[
0.12005615234375,
0.1300048828125,
0.48779296875,
0.1300048828125,
0.48779296875,
0.144287109375,
0.12005615234375,
0.144287109375
],
[
0.51904296875,
0.1318359375,
0.8818359375,
0.133056640625,
0.8818359375,
0.1478271484375,
0.51904296875,
0.146728515625
],
[
0.1351318359375,
0.1455078125,
0.486083984375,
0.1455078125,
0.486083984375,
0.1591796875,
0.1351318359375,
0.1591796875
],
[
0.5146484375,
0.149658203125,
0.880859375,
0.149658203125,
0.880859375,
0.163330078125,
0.5146484375,
0.163330078125
],
[
0.12176513671875,
0.160888671875,
0.48779296875,
0.160888671875,
0.48779296875,
0.174560546875,
0.12176513671875,
0.174560546875
],
[
0.1275634765625,
0.15087890625,
0.1368408203125,
0.15087890625,
0.1368408203125,
0.155029296875,
0.1275634765625,
0.155029296875
],
[
0.5146484375,
0.1650390625,
0.880859375,
0.1650390625,
0.880859375,
0.1787109375,
0.5146484375,
0.1787109375
],
[
0.118408203125,
0.17578125,
0.489501953125,
0.177001953125,
0.489501953125,
0.191162109375,
0.118408203125,
0.190673828125
],
[
0.5146484375,
0.179931640625,
0.8818359375,
0.18115234375,
0.8818359375,
0.1953125,
0.5146484375,
0.19482421875
],
[
0.118408203125,
0.1923828125,
0.486083984375,
0.1923828125,
0.486083984375,
0.2060546875,
0.118408203125,
0.2060546875
],
[
0.515625,
0.19775390625,
0.82958984375,
0.19775390625,
0.82958984375,
0.21142578125,
0.515625,
0.21142578125
],
[
0.12005615234375,
0.208984375,
0.466796875,
0.208984375,
0.466796875,
0.223876953125,
0.12005615234375,
0.223876953125
],
[
0.51708984375,
0.2203369140625,
0.8818359375,
0.22265625,
0.8818359375,
0.2369384765625,
0.51708984375,
0.235107421875
],
[
0.1275634765625,
0.235107421875,
0.489501953125,
0.235107421875,
0.489501953125,
0.248779296875,
0.1275634765625,
0.248779296875
],
[
0.515625,
0.2381591796875,
0.88330078125,
0.2381591796875,
0.88330078125,
0.2529296875,
0.515625,
0.2529296875
],
[
0.12005615234375,
0.25048828125,
0.48779296875,
0.2493896484375,
0.48779296875,
0.26416015625,
0.12005615234375,
0.265380859375
],
[
0.5146484375,
0.2529296875,
0.88330078125,
0.254150390625,
0.88330078125,
0.268310546875,
0.5146484375,
0.267333984375
],
[
0.12005615234375,
0.267333984375,
0.489501953125,
0.267333984375,
0.489501953125,
0.28076171875,
0.12005615234375,
0.28076171875
],
[
0.5146484375,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.5146484375,
0.283935546875
],
[
0.12176513671875,
0.283935546875,
0.328369140625,
0.283935546875,
0.328369140625,
0.297607421875,
0.12176513671875,
0.297607421875
],
[
0.515625,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.515625,
0.2998046875
],
[
0.51708984375,
0.302734375,
0.8818359375,
0.302734375,
0.8818359375,
0.31640625,
0.51708984375,
0.31640625
],
[
0.1275634765625,
0.31640625,
0.204833984375,
0.31640625,
0.204833984375,
0.330810546875,
0.1275634765625,
0.330810546875
],
[
0.515625,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.33203125,
0.515625,
0.33203125
],
[
0.1275634765625,
0.330810546875,
0.49951171875,
0.33203125,
0.49951171875,
0.3466796875,
0.1275634765625,
0.345703125
],
[
0.5146484375,
0.3349609375,
0.880859375,
0.3349609375,
0.880859375,
0.3486328125,
0.5146484375,
0.3486328125
],
[
0.1292724609375,
0.3486328125,
0.49951171875,
0.3486328125,
0.49951171875,
0.3623046875,
0.1292724609375,
0.3623046875
],
[
0.50390625,
0.355224609375,
0.51123046875,
0.355224609375,
0.51123046875,
0.358154296875,
0.50390625,
0.358154296875
],
[
0.5146484375,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.364013671875,
0.5146484375,
0.364013671875
],
[
0.1275634765625,
0.363525390625,
0.345947265625,
0.363525390625,
0.345947265625,
0.377685546875,
0.1275634765625,
0.377685546875
],
[
0.515625,
0.36767578125,
0.6513671875,
0.36767578125,
0.6513671875,
0.38134765625,
0.515625,
0.38134765625
],
[
0.1275634765625,
0.39013671875,
0.185546875,
0.39013671875,
0.185546875,
0.405029296875,
0.1275634765625,
0.405029296875
],
[
0.258544921875,
0.39013671875,
0.316650390625,
0.39013671875,
0.316650390625,
0.405029296875,
0.258544921875,
0.405029296875
],
[
0.391357421875,
0.39013671875,
0.47265625,
0.39013671875,
0.47265625,
0.405029296875,
0.391357421875,
0.405029296875
],
[
0.5224609375,
0.391357421875,
0.8818359375,
0.391357421875,
0.8818359375,
0.405029296875,
0.5224609375,
0.405029296875
],
[
0.1234130859375,
0.40625,
0.181396484375,
0.40625,
0.181396484375,
0.42041015625,
0.1234130859375,
0.42041015625
],
[
0.25439453125,
0.40625,
0.318115234375,
0.40380859375,
0.31982421875,
0.419189453125,
0.255126953125,
0.421630859375
],
[
0.39794921875,
0.40625,
0.466796875,
0.40625,
0.466796875,
0.42041015625,
0.39794921875,
0.42041015625
],
[
0.515625,
0.407958984375,
0.88330078125,
0.407958984375,
0.88330078125,
0.421630859375,
0.515625,
0.421630859375
],
[
0.1275634765625,
0.421630859375,
0.181396484375,
0.421630859375,
0.181396484375,
0.435791015625,
0.1275634765625,
0.435791015625
],
[
0.264404296875,
0.4228515625,
0.310546875,
0.4228515625,
0.310546875,
0.43701171875,
0.264404296875,
0.43701171875
],
[
0.41064453125,
0.419189453125,
0.456787109375,
0.421630859375,
0.455078125,
0.43701171875,
0.408935546875,
0.435302734375
],
[
0.5146484375,
0.42333984375,
0.8818359375,
0.42333984375,
0.8818359375,
0.43701171875,
0.5146484375,
0.43701171875
],
[
0.1275634765625,
0.43701171875,
0.181396484375,
0.43701171875,
0.181396484375,
0.453125,
0.1275634765625,
0.453125
],
[
0.26123046875,
0.438232421875,
0.312255859375,
0.438232421875,
0.312255859375,
0.453125,
0.26123046875,
0.453125
],
[
0.408935546875,
0.43701171875,
0.45751953125,
0.43701171875,
0.45751953125,
0.453125,
0.408935546875,
0.453125
],
[
0.515625,
0.440673828125,
0.88330078125,
0.440673828125,
0.88330078125,
0.45361328125,
0.515625,
0.45361328125
],
[
0.2225341796875,
0.453125,
0.349365234375,
0.45361328125,
0.349365234375,
0.468505859375,
0.2225341796875,
0.46728515625
],
[
0.408935546875,
0.451904296875,
0.45751953125,
0.45361328125,
0.456787109375,
0.4697265625,
0.4072265625,
0.46728515625
],
[
0.515625,
0.4560546875,
0.8818359375,
0.4560546875,
0.8818359375,
0.4697265625,
0.515625,
0.4697265625
],
[
0.408935546875,
0.468505859375,
0.456787109375,
0.47021484375,
0.455078125,
0.486328125,
0.4072265625,
0.48388671875
],
[
0.1275634765625,
0.45361328125,
0.181396484375,
0.45361328125,
0.181396484375,
0.468505859375,
0.1275634765625,
0.468505859375
],
[
0.1259765625,
0.47021484375,
0.181396484375,
0.47021484375,
0.181396484375,
0.485107421875,
0.1259765625,
0.485107421875
],
[
0.25439453125,
0.47021484375,
0.316650390625,
0.468505859375,
0.318115234375,
0.48388671875,
0.255126953125,
0.486328125
],
[
0.5146484375,
0.47021484375,
0.88330078125,
0.47265625,
0.88330078125,
0.486328125,
0.5146484375,
0.48388671875
],
[
0.1275634765625,
0.486328125,
0.181396484375,
0.486328125,
0.181396484375,
0.50048828125,
0.1275634765625,
0.50048828125
],
[
0.2626953125,
0.48681640625,
0.310546875,
0.48681640625,
0.310546875,
0.5029296875,
0.2626953125,
0.5029296875
],
[
0.393798828125,
0.485107421875,
0.470947265625,
0.48681640625,
0.470947265625,
0.5029296875,
0.393798828125,
0.50048828125
],
[
0.5146484375,
0.488037109375,
0.8818359375,
0.488037109375,
0.8818359375,
0.501953125,
0.5146484375,
0.501953125
],
[
0.1275634765625,
0.49951171875,
0.1822509765625,
0.5029296875,
0.181396484375,
0.5185546875,
0.1259765625,
0.515625
],
[
0.26123046875,
0.50048828125,
0.313232421875,
0.5029296875,
0.312255859375,
0.5185546875,
0.26025390625,
0.5166015625
],
[
0.408935546875,
0.501953125,
0.456787109375,
0.501953125,
0.456787109375,
0.51708984375,
0.408935546875,
0.51708984375
],
[
0.51318359375,
0.501953125,
0.72119140625,
0.5029296875,
0.72119140625,
0.51708984375,
0.51318359375,
0.5166015625
],
[
0.1275634765625,
0.5185546875,
0.181396484375,
0.5185546875,
0.181396484375,
0.533203125,
0.1275634765625,
0.533203125
],
[
0.2225341796875,
0.51953125,
0.349365234375,
0.51953125,
0.349365234375,
0.533203125,
0.2225341796875,
0.533203125
],
[
0.39794921875,
0.5185546875,
0.466796875,
0.5185546875,
0.466796875,
0.533203125,
0.39794921875,
0.533203125
],
[
0.5146484375,
0.53076171875,
0.5615234375,
0.53076171875,
0.5615234375,
0.5419921875,
0.5146484375,
0.5419921875
],
[
0.55517578125,
0.52978515625,
0.8076171875,
0.52978515625,
0.8076171875,
0.54443359375,
0.55517578125,
0.54443359375
],
[
0.1427001953125,
0.5498046875,
0.2166748046875,
0.5498046875,
0.2166748046875,
0.56103515625,
0.1427001953125,
0.56103515625
],
[
0.2091064453125,
0.5498046875,
0.465087890625,
0.5498046875,
0.465087890625,
0.5634765625,
0.2091064453125,
0.5634765625
],
[
0.51318359375,
0.548828125,
0.8818359375,
0.55029296875,
0.8818359375,
0.5654296875,
0.51318359375,
0.5634765625
],
[
0.51708984375,
0.56787109375,
0.8818359375,
0.56787109375,
0.8818359375,
0.58056640625,
0.51708984375,
0.58056640625
],
[
0.515625,
0.5830078125,
0.8818359375,
0.5830078125,
0.8818359375,
0.5966796875,
0.515625,
0.5966796875
],
[
0.12176513671875,
0.5927734375,
0.1695556640625,
0.5927734375,
0.1695556640625,
0.60400390625,
0.12176513671875,
0.60400390625
],
[
0.162109375,
0.59130859375,
0.47607421875,
0.59130859375,
0.47607421875,
0.60498046875,
0.162109375,
0.60498046875
],
[
0.515625,
0.5986328125,
0.8818359375,
0.5986328125,
0.8818359375,
0.6123046875,
0.515625,
0.6123046875
],
[
0.12176513671875,
0.6123046875,
0.48779296875,
0.6123046875,
0.48779296875,
0.6259765625,
0.12176513671875,
0.6259765625
],
[
0.5146484375,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.5146484375,
0.62890625
],
[
0.12176513671875,
0.62890625,
0.489501953125,
0.62890625,
0.489501953125,
0.642578125,
0.12176513671875,
0.642578125
],
[
0.51318359375,
0.63134765625,
0.5634765625,
0.63134765625,
0.5634765625,
0.646484375,
0.51318359375,
0.646484375
],
[
0.12005615234375,
0.6455078125,
0.34765625,
0.6455078125,
0.34765625,
0.6591796875,
0.12005615234375,
0.6591796875
],
[
0.5146484375,
0.65478515625,
0.5751953125,
0.65478515625,
0.5751953125,
0.66943359375,
0.5146484375,
0.66943359375
],
[
0.57080078125,
0.66015625,
0.57666015625,
0.66015625,
0.57666015625,
0.6650390625,
0.57080078125,
0.6650390625
],
[
0.57275390625,
0.65625,
0.7421875,
0.65625,
0.7421875,
0.66943359375,
0.57275390625,
0.66943359375
],
[
0.2275390625,
0.6748046875,
0.291259765625,
0.6767578125,
0.28955078125,
0.6923828125,
0.225830078125,
0.6904296875
],
[
0.51123046875,
0.67333984375,
0.87890625,
0.67333984375,
0.87890625,
0.68701171875,
0.51123046875,
0.68701171875
],
[
0.34765625,
0.6767578125,
0.51904296875,
0.6767578125,
0.51904296875,
0.69140625,
0.34765625,
0.69140625
],
[
0.1259765625,
0.6767578125,
0.2032470703125,
0.6767578125,
0.2032470703125,
0.69140625,
0.1259765625,
0.69140625
],
[
0.2392578125,
0.6923828125,
0.279541015625,
0.6923828125,
0.279541015625,
0.7080078125,
0.2392578125,
0.7080078125
],
[
0.344970703125,
0.69140625,
0.88330078125,
0.68701171875,
0.88330078125,
0.705078125,
0.344970703125,
0.708984375
],
[
0.31494140625,
0.70703125,
0.88330078125,
0.703125,
0.88330078125,
0.7216796875,
0.31494140625,
0.7255859375
],
[
0.1368408203125,
0.693359375,
0.193115234375,
0.693359375,
0.193115234375,
0.7080078125,
0.1368408203125,
0.7080078125
],
[
0.1351318359375,
0.708984375,
0.19482421875,
0.708984375,
0.19482421875,
0.7236328125,
0.1351318359375,
0.7236328125
],
[
0.240966796875,
0.7080078125,
0.27783203125,
0.7080078125,
0.27783203125,
0.724609375,
0.240966796875,
0.724609375
],
[
0.2149658203125,
0.7236328125,
0.30322265625,
0.7255859375,
0.30322265625,
0.74365234375,
0.2149658203125,
0.7412109375
],
[
0.51123046875,
0.7216796875,
0.880859375,
0.7216796875,
0.880859375,
0.7353515625,
0.51123046875,
0.7353515625
],
[
0.3525390625,
0.7255859375,
0.49365234375,
0.7255859375,
0.49365234375,
0.7392578125,
0.3525390625,
0.7392578125
],
[
0.139404296875,
0.724609375,
0.193115234375,
0.724609375,
0.193115234375,
0.74072265625,
0.139404296875,
0.74072265625
],
[
0.1368408203125,
0.74072265625,
0.193115234375,
0.74072265625,
0.193115234375,
0.755859375,
0.1368408203125,
0.755859375
],
[
0.219970703125,
0.7421875,
0.301513671875,
0.7421875,
0.301513671875,
0.75732421875,
0.219970703125,
0.75732421875
],
[
0.31982421875,
0.74072265625,
0.765625,
0.7353515625,
0.765625,
0.7529296875,
0.31982421875,
0.75830078125
],
[
0.5146484375,
0.76318359375,
0.75830078125,
0.76318359375,
0.75830078125,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.1175537109375,
0.771484375,
0.489501953125,
0.77099609375,
0.489501953125,
0.78515625,
0.1175537109375,
0.7861328125
],
[
0.51318359375,
0.78076171875,
0.8818359375,
0.7822265625,
0.8818359375,
0.79638671875,
0.51318359375,
0.7958984375
],
[
0.118408203125,
0.78759765625,
0.489501953125,
0.78759765625,
0.489501953125,
0.80029296875,
0.118408203125,
0.80029296875
],
[
0.1175537109375,
0.7998046875,
0.489501953125,
0.80029296875,
0.489501953125,
0.8154296875,
0.1175537109375,
0.81396484375
],
[
0.51708984375,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.51708984375,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.16455078125,
0.81640625,
0.162841796875,
0.82763671875,
0.118408203125,
0.826171875
],
[
0.51708984375,
0.81640625,
0.880859375,
0.81640625,
0.880859375,
0.82666015625,
0.51708984375,
0.82666015625
],
[
0.51708984375,
0.830078125,
0.8818359375,
0.830078125,
0.8818359375,
0.84423828125,
0.51708984375,
0.84423828125
],
[
0.51318359375,
0.845703125,
0.87158203125,
0.8427734375,
0.87158203125,
0.85986328125,
0.51318359375,
0.86328125
],
[
0.1259765625,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.1259765625,
0.869140625
],
[
0.515625,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.87744140625,
0.515625,
0.87744140625
],
[
0.12005615234375,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.12005615234375,
0.8857421875
],
[
0.515625,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.515625,
0.8935546875
],
[
0.1376953125,
0.89453125,
0.411376953125,
0.8955078125,
0.411376953125,
0.91015625,
0.1376953125,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.5458984375,
0.8955078125,
0.5458984375,
0.91015625,
0.5146484375,
0.91015625
]
] |
[
"2.1Task description",
"对于目标-方面对（LOCATION1，安全）的集",
"TABSA在TABSA中，一个句子S通常由一",
"合，我们生成的句子是“您如何看待位置-1的",
"系列单词组成：【w1,，wm】，其中一些单词",
"安全性？”",
"[wi，.，wi］是预先识别的目标{ti，··，t]",
"Sentences forNLI-M对于NLI任务，我们",
"在Saeidietal.（2016）之后，我们将任务",
"在生成句子时设置的条件不那么严格，形式也",
"设置为3类分类问题：给定句子s，一组",
"",
"简单得多。此时创建的句子不是标准句子，而",
"XMATHX实体。正如我们在表1中看到的那",
"是简单的伪句子，以（LOCATION1，safety）",
"样，（LOCATION2，价格）的黄金标准极性为",
"对为例：辅助句子是：“位置-1-safety\"。",
"负，而（LOCATION1，价格）的极性为无",
"SentencesforQA-B对于QA-B，我们添加标",
"ABSA在ABSA中，目标-方面对{t,a}仅",
"签信息，并暂时将TABSA转换为二进制分类",
"成为方面α。此设置等效于同时学习SemEval-",
"问题（labelE{yes,no}），以获得概率分布。此",
"2014任务42的子任务3（方面类别检测）和",
"时，每个目标-方面对将生成三个序列，例如“位",
"子任务4（方面类别极性）",
"置-1的方面安全的极性为正”，“位置-1的方面",
"安全的极性为负”，“位置-1的方面安全的极性",
"Example:",
"为无”。我们使用yes的概率值作为匹配分数。",
"LOCATION2 is central London so extremely",
"对于生成三个序列（positive，negative,none）",
"expensive, LOCATION1 is often considered",
"",
"的目标-方面对，我们将预测类别的匹配得分",
"the coolest area of London",
"最高的序列类别。",
"Target",
"Aspect",
"Sentiment",
"Sentences for NLI-B NLI-B 和 QA-B 之间",
"LOC1",
"general",
"Positive",
"的区别在于，辅助句子从问题变为伪句子。辅",
"LOC1",
"price",
"None",
"助句子是：“位置-1-安全-阳性”，“位置-1-安",
"LOC1",
"safety",
"None",
"全-阴性”和“位置-1-安全-无”在我们构造辅",
"transit-location",
"None",
"助句子之后，我们可以将TABSA任务从单个",
"None",
"LOC1",
"LOC2",
"general",
"句子分类任务转换为句子对分类任务。如表",
"LOC2",
"price",
"Negative",
"3所示，这是一项必要的操作，可以显着改善",
"LOC2",
"safety",
"None",
"TABSA任务的实验结果。",
"LOC2",
"transit-location",
"Positive",
"2.3",
"Fine-tuning pre-trained BERT",
"Table 1:",
"An example of SentiHood dataset.",
"BERT（Devlinetal.，2018）是一种新的语言",
"表示模型，它使用双向转换器对大型语料库进",
"行预训练，并在其他任务上对预训练的模型进",
"2.2",
"Construction of the auxiliary sentence",
"行微调。我们在TABSA任务上微调预训练的",
"为简单起见，我们主要以TABSA为例描述我",
"BERT模型。让我们简要看看输表示和微调",
"们的方法。我们考虑以下四种方法将TABSA",
"过程。",
"任务转换为句子对分类任务：",
"2.3.1",
"",
"Input representation",
"Output",
"BERT的输入表示可以显式表示令牌序列中",
"Auxiliary Sentence",
" Methods",
"S.P.",
"QuestionwoS.P.的一对文本句子。对于给定的令牌，其输人表",
"Pseudo-sentencew/o S.P示是通过对相应的令牌，段和位置嵌人求和来",
"QA-M",
"NLI-M",
"S.P.",
"[ yes,no ]",
"构造的。对于分类任务，每个序列的第一个单",
"Question w/ S.P.",
"QA-B",
"NLI-B",
"[ yes,no ]",
"Pseudo-sentencew/S.P词是唯一的分类嵌人（[CLS]）。",
"2.3.2Fine-tuning procedure",
"Table 2: The construction methods. Due to limited",
"BERT微调非常简单。为了获得输入序列的",
"space, we use the following abbreviations: S.P. for",
"sentiment polarity , w/o for without , and w/ for",
"固定维池化表示，我们使用最终的隐藏状态",
"with.",
"(i。e。：变压器的输出）的第一个令牌作为输",
"人。我们将向量表示为CERH。然后我们",
"添加一个分类层，其参数矩阵为WERKxH",
"Sentences forQA-M我们要从目标-方面对",
"其中K是类别数。最后，通过 softmax函数",
"生成的句子是一个问题，格式需要相同。例如，",
"P=softmaa(CWT）计算每个类别P的概",
"2http://alt.qcri.org/semeval2014/task4/",
"率。"
] |
{
"arxiv_id": "1903.09588",
"language": "zh",
"page_idx": 1
}
|
||
1903.09588-zh-2
|
1903.09588-zh-2.jpg
|
2.3.3 BERT-single and BERT-pair
BERT-single for (T)ABSA 用于单句分类任务的 BERT。假设目标类别的数量为$n_t$, 方面类别为$n_a$ 。我们将 TABSA 视为$n_t\cdot n_a$ 目标方面相关的情感分类问题的组合，首先对每个情感分类问题进行分类，然后对获得的结果进行总结。对于 ABSA, 我们微调预训练的BERT 模型来训练$n_a$分类器的各个方面，然后总结结果。
用于句子对分类任务的
BERT-pair for (T)ABSA BERT。基于 2.2节中构造的辅助句子，我们使用句子对分类方法来求解(T) ABSA。对应于四种构造句子的方式，我们将模型命名为：BERT-pair-QAM, BERT- pair- NLI- M, BERT- pair- QA- B 和BERT- pair- NLI- B。
## 3 Experiments
3.1 Datasets
我们在 SentiHood (Saeidi et al., 2016) 数据集上评估我们的方法，该数据集包含 5,215 个句子，3,。每个句子都包含一个具有情绪极性$y$ 的目标-方面对$\{t,a\}$的列表。最终，给定句子$s$和句子中的目标$t$ ,我们需要：(1)检测到目标$t$的方面$a$的提及；(2)确定检测到的目标方面对的正面或负面情绪极性$y$ -。我们还评估了我们在 SemEval-2014Task4 (Pontiki et al.,2014)数据集 $^3$上进行基于方面的情感分析的方法。与 SentiHood 的唯一区别是目标-方面对$\{t,a\}$仅成为方面$a$ 。此设置允许我们联合评估子任务 3(方面类别检测)和子任务 4(方面类别极性)。
## 3.2 Hyperparameters
我们使用预训练的无案例 BERT 基础模型$^{4}$ 进行微调。Transformer 块的数量为 12, 隐藏层大小为 768, 自关注头的数量为 12,预训练模型的参数总数为 110M。微调时，我们将退出概率保持在0。1,将纪元数设置为 4。初始学习率为 2e-5, 批量大小为 24。
## 3.3 Exp-I: TABSA
我们将我们的模型与以下模型进行比较：
$\bullet$ LR (Saeidi et al., 2016): a logistic re-
gression classifier with n-gram and pos-
$\operatorname{tag}$features.
$\bullet$ LSTM-Final (Saeidi et al., 2016): a biL-
STM model with the fnal state as a rep-
resentation.
$\bullet$ LSTM-Loc (Saeidi et al., 2016): a biL-
STM model with the state associated with
the target position as a representation.
$\bullet$LSTM+TA+SA(Maet al.,2018):a
biLSTM model which introduces complex target-level and sentence-level attention mechanisms.
$\bullet$ SenticLSTM (Ma et al., 2018): an up-
graded version of the LSTM+TA+SA model which introduces external information from SenticNet (Cambria et al., 2016) .
$\bullet$Dmu-Entnet (Liu et al., 2018): a bi-
directional EntNet (Henaff et al., 2016) with external“memory chains”with a delayed memory update mechanism to track entities.
在评估 SentiHood 期间，在 Saeidi et al. (2016) 之后，我们仅考虑最常见的四个方面(一般， 价格，运输-位置，安全性)。在评估方面检测时，在 Ma et al. (2018)之后，我们使用严格的准确性和 Macro-F1,并且我们还报告了AUC。在情感分类中，我们使用准确性和宏观平均 AUC 作为评价指标。
3.3.1 Results
在 SentiHood 上的结果如表 3 所示。BERTsingle 模型在方面检测上的结果优于 DmuEntnet, 但情感分类的准确性远低于 SenticLstm 和 Dmu-Entnet, 相差 3。8 和 5。5 分别。然而，BERT-pair 在方面检测和情感分析方面的表现明显优于其他模型，获得 9。4 宏观平均 F1 和 2。比 Dmu-Entnet 提高了 6 个精度。总体而言，四种 BERT 对模型的性能接近。值得注意的是，BERT-pair-NLI 模型在方面检测上表现相对更好，而 BERT-pair-QA 模型在情感分类上表现更好。此外，与其他模型相比，BERT-pair-QA-B 和 BERT-pair-NLI-B 模型在情感分类上可以获得更好的 AUC 值。
## 3.4 Exp-II: ABSA
SemEval-2014 任务 4 的基准测试是 Pontiki et al. (2014)和 ATAE-LSTM (Wang et al., 2016) 中性能最好的两个系统。在评估 SemEval-2014 任务 4 子任务 3 和子任务 4
$\begin{array}{ll}{3}\text{http://alt.qcri.org/semeval2014/task4/}&{\text{由 borninevar-1月L}0\text{和}0\text{相}1\text{L}/\text{D}0\:1\text{P}0\text{ Mill}}\\{4}\text{https://storage.googleapis.com/bert_models/2018_10}\text{时}_{8/}\text{在 Pontiki et al.(2014) 之后，我们分别使}\\{4}&{4}\\{\text{https://storage.googleapis.com/bert_models/2$
${\text{用 Micro-F1 和准确性。SemEval-2014 上的}}$
uncased$\_$L- 12$\_$H- 768$\_$A- 12. zip
|
[
[
0.1175537109375,
0.07598876953125,
0.4130859375,
0.07598876953125,
0.4130859375,
0.08905029296875,
0.1175537109375,
0.08905029296875
],
[
0.53076171875,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.53076171875,
0.09027099609375
],
[
0.139404296875,
0.0955810546875,
0.486083984375,
0.0955810546875,
0.486083984375,
0.10986328125,
0.139404296875,
0.10986328125
],
[
0.54833984375,
0.0914306640625,
0.880859375,
0.09320068359375,
0.880859375,
0.10809326171875,
0.54833984375,
0.105712890625
],
[
0.12005615234375,
0.11102294921875,
0.490234375,
0.11224365234375,
0.490234375,
0.1270751953125,
0.12005615234375,
0.1258544921875
],
[
0.5517578125,
0.10986328125,
0.64404296875,
0.10986328125,
0.64404296875,
0.1217041015625,
0.5517578125,
0.1217041015625
],
[
0.12005615234375,
0.127685546875,
0.484375,
0.12890625,
0.484375,
0.1436767578125,
0.12005615234375,
0.142578125
],
[
0.5322265625,
0.133056640625,
0.8818359375,
0.133056640625,
0.8818359375,
0.146728515625,
0.5322265625,
0.146728515625
],
[
0.12005615234375,
0.144287109375,
0.486083984375,
0.1436767578125,
0.486083984375,
0.157958984375,
0.12005615234375,
0.1591796875
],
[
0.5517578125,
0.1484375,
0.88330078125,
0.1484375,
0.88330078125,
0.162109375,
0.5517578125,
0.162109375
],
[
0.12005615234375,
0.160888671875,
0.486083984375,
0.160888671875,
0.486083984375,
0.174560546875,
0.12005615234375,
0.174560546875
],
[
0.54833984375,
0.1644287109375,
0.86376953125,
0.1650390625,
0.86376953125,
0.179931640625,
0.54833984375,
0.1787109375
],
[
0.12005615234375,
0.177001953125,
0.48779296875,
0.177001953125,
0.48779296875,
0.190673828125,
0.12005615234375,
0.190673828125
],
[
0.118408203125,
0.191162109375,
0.48779296875,
0.1923828125,
0.48779296875,
0.207275390625,
0.118408203125,
0.2060546875
],
[
0.53076171875,
0.188232421875,
0.88330078125,
0.189453125,
0.88330078125,
0.2037353515625,
0.53076171875,
0.203125
],
[
0.12005615234375,
0.208984375,
0.219970703125,
0.208984375,
0.219970703125,
0.223876953125,
0.12005615234375,
0.223876953125
],
[
0.54833984375,
0.204833984375,
0.88330078125,
0.2060546875,
0.88330078125,
0.2203369140625,
0.54833984375,
0.2197265625
],
[
0.1376953125,
0.2257080078125,
0.322509765625,
0.2257080078125,
0.322509765625,
0.240478515625,
0.1376953125,
0.240478515625
],
[
0.5517578125,
0.22265625,
0.88330078125,
0.22265625,
0.88330078125,
0.236328125,
0.5517578125,
0.236328125
],
[
0.5498046875,
0.2381591796875,
0.654296875,
0.2381591796875,
0.654296875,
0.2529296875,
0.5498046875,
0.2529296875
],
[
0.1376953125,
0.251708984375,
0.489501953125,
0.251708984375,
0.489501953125,
0.2666015625,
0.1376953125,
0.2666015625
],
[
0.53076171875,
0.26123046875,
0.880859375,
0.262451171875,
0.880859375,
0.276611328125,
0.53076171875,
0.275634765625
],
[
0.118408203125,
0.267333984375,
0.484375,
0.267333984375,
0.484375,
0.28076171875,
0.118408203125,
0.28076171875
],
[
0.12005615234375,
0.283935546875,
0.486083984375,
0.283935546875,
0.486083984375,
0.298583984375,
0.12005615234375,
0.298583984375
],
[
0.54833984375,
0.279052734375,
0.8818359375,
0.276611328125,
0.8818359375,
0.29150390625,
0.54833984375,
0.29345703125
],
[
0.5498046875,
0.29443359375,
0.88330078125,
0.29345703125,
0.88330078125,
0.30810546875,
0.5498046875,
0.308837890625
],
[
0.12005615234375,
0.300537109375,
0.483642578125,
0.300537109375,
0.483642578125,
0.314208984375,
0.12005615234375,
0.314208984375
],
[
0.5498046875,
0.31005859375,
0.8818359375,
0.31005859375,
0.8818359375,
0.32373046875,
0.5498046875,
0.32373046875
],
[
0.12005615234375,
0.31640625,
0.486083984375,
0.31640625,
0.486083984375,
0.330078125,
0.12005615234375,
0.330078125
],
[
0.5498046875,
0.32666015625,
0.611328125,
0.32666015625,
0.611328125,
0.341552734375,
0.5498046875,
0.341552734375
],
[
0.118408203125,
0.333251953125,
0.273681640625,
0.333251953125,
0.273681640625,
0.3466796875,
0.118408203125,
0.3466796875
],
[
0.53076171875,
0.350830078125,
0.880859375,
0.350830078125,
0.880859375,
0.364013671875,
0.53076171875,
0.364013671875
],
[
0.1175537109375,
0.361083984375,
0.15869140625,
0.361083984375,
0.15869140625,
0.377197265625,
0.1175537109375,
0.377197265625
],
[
0.15283203125,
0.3623046875,
0.267822265625,
0.3623046875,
0.267822265625,
0.377197265625,
0.15283203125,
0.377197265625
],
[
0.5517578125,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.5517578125,
0.38134765625
],
[
0.162841796875,
0.388427734375,
0.2392578125,
0.388427734375,
0.2392578125,
0.402587890625,
0.162841796875,
0.402587890625
],
[
0.5517578125,
0.38427734375,
0.880859375,
0.38427734375,
0.880859375,
0.39794921875,
0.5517578125,
0.39794921875
],
[
0.160400390625,
0.392578125,
0.166259765625,
0.392578125,
0.166259765625,
0.3984375,
0.160400390625,
0.3984375
],
[
0.54833984375,
0.3984375,
0.88330078125,
0.399658203125,
0.88330078125,
0.41455078125,
0.54833984375,
0.413330078125
],
[
0.12005615234375,
0.3896484375,
0.154541015625,
0.3896484375,
0.154541015625,
0.402099609375,
0.12005615234375,
0.402099609375
],
[
0.118408203125,
0.4091796875,
0.48779296875,
0.410400390625,
0.48779296875,
0.424560546875,
0.118408203125,
0.42333984375
],
[
0.55322265625,
0.41748046875,
0.61376953125,
0.41748046875,
0.61376953125,
0.4287109375,
0.55322265625,
0.4287109375
],
[
0.12176513671875,
0.427001953125,
0.48779296875,
0.427001953125,
0.48779296875,
0.440673828125,
0.12176513671875,
0.440673828125
],
[
0.5146484375,
0.43701171875,
0.880859375,
0.43701171875,
0.880859375,
0.45068359375,
0.5146484375,
0.45068359375
],
[
0.12005615234375,
0.443603515625,
0.489501953125,
0.443603515625,
0.489501953125,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.5146484375,
0.453125,
0.880859375,
0.453125,
0.880859375,
0.46728515625,
0.5146484375,
0.46728515625
],
[
0.12005615234375,
0.458984375,
0.486083984375,
0.458984375,
0.486083984375,
0.47265625,
0.12005615234375,
0.47265625
],
[
0.12005615234375,
0.474365234375,
0.486083984375,
0.474365234375,
0.486083984375,
0.488037109375,
0.12005615234375,
0.488037109375
],
[
0.5146484375,
0.4697265625,
0.8818359375,
0.4697265625,
0.8818359375,
0.482666015625,
0.5146484375,
0.482666015625
],
[
0.515625,
0.485107421875,
0.8818359375,
0.485107421875,
0.8818359375,
0.498779296875,
0.515625,
0.498779296875
],
[
0.12176513671875,
0.4912109375,
0.48779296875,
0.4912109375,
0.48779296875,
0.5048828125,
0.12176513671875,
0.5048828125
],
[
0.51318359375,
0.49951171875,
0.880859375,
0.50048828125,
0.880859375,
0.515625,
0.51318359375,
0.51416015625
],
[
0.12176513671875,
0.50732421875,
0.48779296875,
0.50732421875,
0.48779296875,
0.52099609375,
0.12176513671875,
0.52099609375
],
[
0.5146484375,
0.51708984375,
0.8818359375,
0.51708984375,
0.8818359375,
0.53076171875,
0.5146484375,
0.53076171875
],
[
0.12005615234375,
0.52392578125,
0.489501953125,
0.52392578125,
0.489501953125,
0.53759765625,
0.12005615234375,
0.53759765625
],
[
0.12005615234375,
0.5390625,
0.48779296875,
0.5390625,
0.48779296875,
0.552734375,
0.12005615234375,
0.552734375
],
[
0.515625,
0.53369140625,
0.72119140625,
0.53369140625,
0.72119140625,
0.54736328125,
0.515625,
0.54736328125
],
[
0.12005615234375,
0.5546875,
0.48779296875,
0.5546875,
0.48779296875,
0.5693359375,
0.12005615234375,
0.5693359375
],
[
0.51318359375,
0.55712890625,
0.63818359375,
0.55712890625,
0.63818359375,
0.57177734375,
0.51318359375,
0.57177734375
],
[
0.118408203125,
0.57080078125,
0.486083984375,
0.57080078125,
0.486083984375,
0.58447265625,
0.118408203125,
0.58447265625
],
[
0.5146484375,
0.576171875,
0.8818359375,
0.576171875,
0.8818359375,
0.5888671875,
0.5146484375,
0.5888671875
],
[
0.118408203125,
0.5859375,
0.48779296875,
0.58740234375,
0.48779296875,
0.6015625,
0.118408203125,
0.60107421875
],
[
0.5146484375,
0.5927734375,
0.8818359375,
0.5927734375,
0.8818359375,
0.60546875,
0.5146484375,
0.60546875
],
[
0.118408203125,
0.6025390625,
0.318115234375,
0.6025390625,
0.318115234375,
0.6162109375,
0.118408203125,
0.6162109375
],
[
0.51318359375,
0.60791015625,
0.8818359375,
0.60693359375,
0.8818359375,
0.62158203125,
0.51318359375,
0.62255859375
],
[
0.5146484375,
0.62353515625,
0.880859375,
0.62353515625,
0.880859375,
0.63720703125,
0.5146484375,
0.63720703125
],
[
0.118408203125,
0.6328125,
0.1695556640625,
0.6328125,
0.1695556640625,
0.64453125,
0.118408203125,
0.64453125
],
[
0.162109375,
0.6328125,
0.30908203125,
0.6328125,
0.30908203125,
0.646484375,
0.162109375,
0.646484375
],
[
0.515625,
0.64013671875,
0.88330078125,
0.64013671875,
0.88330078125,
0.65380859375,
0.515625,
0.65380859375
],
[
0.118408203125,
0.65283203125,
0.490234375,
0.65185546875,
0.490234375,
0.66650390625,
0.118408203125,
0.66748046875
],
[
0.515625,
0.65673828125,
0.88330078125,
0.65673828125,
0.88330078125,
0.67041015625,
0.515625,
0.67041015625
],
[
0.12005615234375,
0.66943359375,
0.489501953125,
0.66943359375,
0.489501953125,
0.68310546875,
0.12005615234375,
0.68310546875
],
[
0.5146484375,
0.67138671875,
0.8818359375,
0.67138671875,
0.8818359375,
0.68603515625,
0.5146484375,
0.68603515625
],
[
0.1234130859375,
0.68701171875,
0.486083984375,
0.68701171875,
0.486083984375,
0.69775390625,
0.1234130859375,
0.69775390625
],
[
0.51318359375,
0.6884765625,
0.88330078125,
0.68701171875,
0.88330078125,
0.70166015625,
0.51318359375,
0.703125
],
[
0.12005615234375,
0.70068359375,
0.48779296875,
0.70068359375,
0.48779296875,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.515625,
0.70361328125,
0.8818359375,
0.70361328125,
0.8818359375,
0.71875,
0.515625,
0.71875
],
[
0.12005615234375,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.12005615234375,
0.732421875
],
[
0.5146484375,
0.7197265625,
0.880859375,
0.7197265625,
0.880859375,
0.73291015625,
0.5146484375,
0.73291015625
],
[
0.12005615234375,
0.73388671875,
0.37451171875,
0.73388671875,
0.37451171875,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.5146484375,
0.7353515625,
0.880859375,
0.7353515625,
0.880859375,
0.7490234375,
0.5146484375,
0.7490234375
],
[
0.515625,
0.7529296875,
0.88330078125,
0.7529296875,
0.88330078125,
0.7666015625,
0.515625,
0.7666015625
],
[
0.118408203125,
0.76123046875,
0.287109375,
0.76123046875,
0.287109375,
0.775390625,
0.118408203125,
0.775390625
],
[
0.51318359375,
0.7685546875,
0.86572265625,
0.7685546875,
0.86572265625,
0.7822265625,
0.51318359375,
0.7822265625
],
[
0.12005615234375,
0.78369140625,
0.43994140625,
0.78369140625,
0.43994140625,
0.79736328125,
0.12005615234375,
0.79736328125
],
[
0.51318359375,
0.79345703125,
0.67822265625,
0.79345703125,
0.67822265625,
0.80810546875,
0.51318359375,
0.80810546875
],
[
0.1351318359375,
0.8134765625,
0.490234375,
0.8134765625,
0.490234375,
0.82666015625,
0.1351318359375,
0.82666015625
],
[
0.51318359375,
0.81396484375,
0.880859375,
0.81396484375,
0.880859375,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.1561279296875,
0.830078125,
0.49365234375,
0.830078125,
0.49365234375,
0.84423828125,
0.1561279296875,
0.84423828125
],
[
0.51123046875,
0.830078125,
0.8818359375,
0.83056640625,
0.8818359375,
0.845703125,
0.51123046875,
0.84423828125
],
[
0.154541015625,
0.84326171875,
0.258544921875,
0.845703125,
0.258544921875,
0.85986328125,
0.154541015625,
0.85791015625
],
[
0.51318359375,
0.845703125,
0.8818359375,
0.8466796875,
0.8818359375,
0.86083984375,
0.51318359375,
0.85986328125
],
[
0.5146484375,
0.8623046875,
0.8818359375,
0.8623046875,
0.8818359375,
0.876953125,
0.5146484375,
0.876953125
],
[
0.139404296875,
0.87060546875,
0.411376953125,
0.87060546875,
0.411376953125,
0.88525390625,
0.139404296875,
0.88525390625
],
[
0.1368408203125,
0.87890625,
0.884765625,
0.87255859375,
0.884765625,
0.896484375,
0.1368408203125,
0.90283203125
],
[
0.1175537109375,
0.89453125,
0.3359375,
0.8955078125,
0.3359375,
0.91015625,
0.1175537109375,
0.9091796875
],
[
0.515625,
0.8955078125,
0.86572265625,
0.8955078125,
0.86572265625,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"2.3.3BERT-single and BERT-pair",
": LSTM-Final (Saeidi et al., 2016) : a biL-",
"BERT-singlefor(T)ABSA用于单句分类",
"STM model with the final state as a rep-",
"任务的BERT。假设目标类别的数量为nt，",
"resentation.",
"方面类别为na。我们将TABSA视为nt·na",
": LSTM-Loc (Saeidi et al., 2016) : a biL-",
"目标方面相关的情感分类问题的组合，首先对",
"STM model with the state associated with",
"每个情感分类问题进行分类，然后对获得的结",
"the target position as a representation.",
"果进行总结。对于ABSA，我们微调预训练的",
"BERT模型来训练na分类器的各个方面，然",
".LSTM+TA+SA (Ma et al., 2018) : a",
"后总结结果。",
"biLSTM model which introduces complex",
"用于句子对分类任务的",
"target-level and sentence-level attention",
"mechanisms.",
"BERT-pair for (T)ABSA BERT。基于 2.2",
": SenticLSTM (Ma et al., 2018) : an up-",
"节中构造的辅助句子，我们使用句子对分类方",
"法来求解（T）ABSA。对应于四种构造句子",
"graded version of the LSTM+TA+SA",
"model which introduces external infor-",
"的方式，我们将模型命名为：BERT-pair-QA",
"mation from SenticNet (Cambria et al.,",
"M,BERT-pair-NLI-M,BERT-pair-QA-B 和",
"2016)",
"BERT-pair-NLI-B。",
". Dmu-Entnet (Liu et al. 2018) : a bi-",
"3",
"Experiments",
"directional EntNet (Henaff et al., 2016)",
"Datasets",
"with external “memory chains\"with a de-",
"",
"layed memory update mechanism to track",
"3.1",
"我们在SentiHood（Saeidietal，2016）数据集",
"entities.",
"上评估我们的方法，该数据集包含5,215个句",
"在评估SentiHood期间，在Saeidietal.（2016）",
"子，3,。每个句子都包含一个具有情绪极性y",
"之后，我们仅考虑最常见的四个方面（一般",
"的目标-方面对{t,α}的列表。最终，给定句子",
"s和句子中的目标t，我们需要：（1）检测到",
"价格，运输-位置，安全性）。在评估方面检",
"测时，在Maetal.（2018）之后，我们使用严",
"目标t的方面aα的提及；（2）确定检测到的",
"格的准确性和Macro-F1，并且我们还报告了",
"目标方面对的正面或负面情绪极性y-。我们",
"AUC。在情感分类中，我们使用准确性和宏观",
"还评估了我们在SemEval-2014Task4（Pontiki",
"etal.，2014）数据集3上进行基于方面的情感",
"平均AUC作为评价指标。",
"分析的方法。与 SentiHood 的唯一区别是目",
"3.3.1Results",
"标-方面对{t,α}仅成为方面α。此设置允许",
"在SentiHood上的结果如表3所示。BERT-",
"我们联合评估子任务3（方面类别检测）和子",
"single模型在方面检测上的结果优于Dmu-",
"任务4（方面类别极性）。",
"Entnet，但情感分类的准确性远低于Senti-",
"cLstm和Dmu-Entnet，相差3。8和5。5分",
"3.2",
"Hyperparameters",
"别。然而，BERT-pair在方面检测和情感分析",
"我们使用预训练的无案例BERT基础模型4",
"方面的表现明显优于其他模型，获得9。4宏",
"进行微调。Transformer块的数量为12，隐藏",
"观平均F1和2。比Dmu-Entnet提高了6个",
"层大小为768，自关注头的数量为12，预训练",
"精度。总体而言，四种BERT对模型的性能接",
"模型的参数总数为110M。微调时，我们将退",
"近。值得注意的是，BERT-pair-NLI模型在方",
"出概率保持在0。1，将纪元数设置为4。初始",
"面检测上表现相对更好，而BERT-pair-QA模",
"学习率为2e-5，批量大小为24。",
"型在情感分类上表现更好。此外，与其他模型",
"相比，BERT-pair-QA-B和BERT-pair-NLI-B",
"3.3Exp-I: TABSA",
"模型在情感分类上可以获得更好的AUC值",
"我们将我们的模型与以下模型进行比较：",
"3.4Exp-II: ABSA",
": LR (Saeidi et al., 2016) : a logistic re-",
"SemEval-2014任务4 的基准测试是Pon-",
"gression classifier with n-gram and pos-",
"tiki et al.(2014)和 ATAE-LSTM (Wang",
"tag features.",
"etal.，2016）中性能最好的两个系统。在评",
"估SemEval-2014任务4子任务3和子任务4",
"3http://alt.qcri.org/semeval2014/task4/",
"https/storage.googleapis.com/bertmodels/2018_1时s/在Pontikietal（2014）之后，我们分别使",
"uncased_L-12_H-768_A-12.zip",
"用Micro-F1和准确性。SemEval-2014上的"
] |
{
"arxiv_id": "1903.09588",
"language": "zh",
"page_idx": 2
}
|
||
1903.09588-zh-3
|
1903.09588-zh-3.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">Model</th>
<th colspan="3">Aspect</th>
<th colspan="2">Sentiment</th>
</tr>
<tr>
<th>$Acc.$</th>
<th>$\overline{F_1}$</th>
<th>$AUC$</th>
<th>$Acc.$</th>
<th>$AUC$</th>
</tr>
<tr>
<td>LR (Saeidi et al., 2016)</td>
<td> </td>
<td>39.3</td>
<td>92.4</td>
<td>87.5</td>
<td>90.5</td>
</tr>
<tr>
<td>(Saeidi et al., 2016 LSTM- Final</td>
<td>-</td>
<td>68.9</td>
<td>89.8</td>
<td>82.0</td>
<td>85.4</td>
</tr>
<tr>
<td>LSTM- Loc Saeidi et al. , 2016</td>
<td>-</td>
<td>69.3</td>
<td>89.7</td>
<td>81.9</td>
<td>83.9</td>
</tr>
<tr>
<td>LSTM+ TA+ SA (Ma et al., 2018</td>
<td>66.4</td>
<td>76.7</td>
<td> </td>
<td>86.8</td>
<td>-</td>
</tr>
<tr>
<td>SenticLSTM (Ma et al., 2018</td>
<td>67.4</td>
<td>78.2</td>
<td>-</td>
<td>89.3</td>
<td>-</td>
</tr>
<tr>
<td>Dmu- Entnet (Liu et al., 2018)</td>
<td>73.5</td>
<td>78.5</td>
<td>94.4</td>
<td>91.0</td>
<td>94.8</td>
</tr>
<tr>
<td>BERT- single</td>
<td>73.7</td>
<td>81.0</td>
<td>96.4</td>
<td>85.5</td>
<td>84.2</td>
</tr>
<tr>
<td>BERT- pair- QA- M</td>
<td>79.4</td>
<td>86.4</td>
<td>97.0</td>
<td>93.6</td>
<td>96.4</td>
</tr>
<tr>
<td>BERT- pair- NLI- M</td>
<td>78.3</td>
<td>87.0</td>
<td>97.5</td>
<td>92.1</td>
<td>96.5</td>
</tr>
<tr>
<td>BERT- pair- QA- B</td>
<td>79.2</td>
<td>87.9</td>
<td>97.1</td>
<td>93.3</td>
<td>97.0</td>
</tr>
<tr>
<td>BERT-pair-NLI-B</td>
<td>79.8</td>
<td>87.5</td>
<td>96.6</td>
<td>92.8</td>
<td>96.9</td>
</tr>
</tbody>
</table>
Table 3: Performance on SentiHood dataset. We boldface the score with the best performance across
all models. We use the results reported in Saeidi et al. (2016) ,Ma et al. (2018) and Liu et al. (2018)
$“\_”$ means not reported.
<table>
<tbody>
<tr>
<th>Models</th>
<th>$P$</th>
<th>$R$</th>
<th>F1</th>
</tr>
<tr>
<td>XRCE</td>
<td>83.23</td>
<td>81.37</td>
<td>82.29</td>
</tr>
<tr>
<td>NRC-Canada</td>
<td>91.04</td>
<td>86.24</td>
<td>88.58</td>
</tr>
<tr>
<td>BERT- single</td>
<td>92.78</td>
<td>89.07</td>
<td>90.89</td>
</tr>
<tr>
<td>BERT- pair- QA- M</td>
<td>$I$ 92.87</td>
<td>90.24</td>
<td>91.54</td>
</tr>
<tr>
<td>BERT- pair- NLI- N</td>
<td>$VI$ 93.15</td>
<td>90.24</td>
<td>91.67</td>
</tr>
<tr>
<td>BERT- pair- QA- B</td>
<td>93.04</td>
<td>89.95</td>
<td>91.47</td>
</tr>
<tr>
<td>BERT- pair- NLI- E</td>
<td>3 93.57</td>
<td>90.83</td>
<td>92.18</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<th>Models</th>
<th>4-way</th>
<th>$3-way$</th>
<th>Binary</th>
</tr>
<tr>
<td>XRCE</td>
<td>78.1</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>NRC-Canada</td>
<td>82.9</td>
<td> </td>
<td>-</td>
</tr>
<tr>
<td>LSTM</td>
<td>一</td>
<td>82.0</td>
<td>88.3</td>
</tr>
<tr>
<td>ATAE- LSTM</td>
<td> </td>
<td>84.0</td>
<td>89.9</td>
</tr>
<tr>
<td>BERT- single</td>
<td>83.7</td>
<td>86.9</td>
<td>93.3</td>
</tr>
<tr>
<td>BERT- pair- QA- M</td>
<td>85.2</td>
<td>89.3</td>
<td>95.4</td>
</tr>
<tr>
<td>BERT- pair- NLI- M</td>
<td>85.1</td>
<td>88.7</td>
<td>94.4</td>
</tr>
<tr>
<td>BERT- pair- QA- B</td>
<td>85.9</td>
<td>89.9</td>
<td>95.6</td>
</tr>
<tr>
<td>BERT- pair- NLI- B</td>
<td>84.6</td>
<td>88.7</td>
<td>95.1</td>
</tr>
</tbody>
</table>
Table 4: Test set results for Semeval-2014 task 4 Subtask 3: Aspect Category Detection. We use the results reported in XRCE (Brun et al., 2014) and NRC-Canada (Kiritchenko et al., 2014) .
Table 5: Test set accuracy $(\begin{array}{c}\%\end{array})$ for Semeval- 2014 task 4 Subtask 4: Aspect Category Polarity. We use the results reported in XRCE (Brun et al. 2014) , NRC- Canada ( Kiritchenko et al. , 2014) and ATAE-LSTM (Wang et al., 2016) .“-”means not reported.
3.4.1 Results
结果如表 4 和表 5 所示。我们发现 BERTsingle 在这两个子任务上取得了更好的效果， 并且 BERT-pair 比 BERT-single 取得了进一步的改进。BERT-pair-NLI-B 模型实现了方面类别检测的最佳性能。对于方面类别极性， BERT-pair-QA-B 在所有 4 路、3 路和二进制设置上表现最佳。
# 4 Discussion
一个句子预测任务。由于附加的目标和方面信息，TABSA 比 SA 更复杂。在 TABSA 上直接微调预先训练的 BERT 无法实现性能增长。但是，当我们将目标和方面分开以形成辅助句子并将 TABSA 转换为句子对分类任务时，场景类似于 QA 和 NLI, 然后可以充分利用预训练的 BERT 模型的优势。我们的方法不限于TABSA,这种构造方法可以用于其他类似的任务。对于 ABSA, 我们可以使用相同的方法来构造仅包含方面的辅助句子在 BERT-pair 模型中，BERT-pair-QA-B 和 BERT-pair-NLI-B 在情感分类上实现了更好的 AUC 值，这可能是由于标签信息的建模。
将哪个方面的目标语料库。原始数据集中的一个句子$s_i$将扩展为句子对分类任务中的$\left(s_i,t_1,a_1\right),\cdots,\left(s_i,t_1,a_{n_a}\right),\cdots,\left(s_i,t_{n_t},a_{n_a}\right)$ 。另一方面，从 BERT 模型对 QA 和 NLI 任务 (Devlin et al., 2018) 的惊人改进可以看出， BERT 模型在处理句子对分类任务方面具有优势。此优势来自无监督掩码语言模型和下
|
[
[
0.5498046875,
0.0789794921875,
0.609375,
0.0789794921875,
0.609375,
0.09320068359375,
0.5498046875,
0.09320068359375
],
[
0.6962890625,
0.07659912109375,
0.7841796875,
0.0789794921875,
0.7841796875,
0.09320068359375,
0.6962890625,
0.0914306640625
],
[
0.210693359375,
0.08612060546875,
0.268798828125,
0.08612060546875,
0.268798828125,
0.1009521484375,
0.210693359375,
0.1009521484375
],
[
0.49951171875,
0.0955810546875,
0.5458984375,
0.0955810546875,
0.5458984375,
0.10986328125,
0.49951171875,
0.10986328125
],
[
0.609375,
0.09442138671875,
0.6572265625,
0.09442138671875,
0.6572265625,
0.10986328125,
0.609375,
0.10986328125
],
[
0.74462890625,
0.0955810546875,
0.79248046875,
0.0955810546875,
0.79248046875,
0.10986328125,
0.74462890625,
0.10986328125
],
[
0.69189453125,
0.0968017578125,
0.73046875,
0.0968017578125,
0.73046875,
0.1092529296875,
0.69189453125,
0.1092529296875
],
[
0.5634765625,
0.0968017578125,
0.587890625,
0.0968017578125,
0.587890625,
0.10986328125,
0.5634765625,
0.10986328125
],
[
0.21240234375,
0.1175537109375,
0.403076171875,
0.1175537109375,
0.403076171875,
0.1318359375,
0.21240234375,
0.1318359375
],
[
0.5556640625,
0.1175537109375,
0.59619140625,
0.1175537109375,
0.59619140625,
0.133056640625,
0.5556640625,
0.133056640625
],
[
0.611328125,
0.1175537109375,
0.6533203125,
0.1175537109375,
0.6533203125,
0.133056640625,
0.611328125,
0.133056640625
],
[
0.68994140625,
0.1175537109375,
0.73046875,
0.1175537109375,
0.73046875,
0.133056640625,
0.68994140625,
0.133056640625
],
[
0.74658203125,
0.1175537109375,
0.78857421875,
0.1175537109375,
0.78857421875,
0.133056640625,
0.74658203125,
0.133056640625
],
[
0.5556640625,
0.133056640625,
0.59619140625,
0.133056640625,
0.59619140625,
0.1484375,
0.5556640625,
0.1484375
],
[
0.611328125,
0.133056640625,
0.6533203125,
0.133056640625,
0.6533203125,
0.1484375,
0.611328125,
0.1484375
],
[
0.68994140625,
0.133056640625,
0.73046875,
0.133056640625,
0.73046875,
0.1484375,
0.68994140625,
0.1484375
],
[
0.748046875,
0.133056640625,
0.78857421875,
0.133056640625,
0.78857421875,
0.1484375,
0.748046875,
0.1484375
],
[
0.5556640625,
0.1484375,
0.59765625,
0.1484375,
0.59765625,
0.1644287109375,
0.5556640625,
0.1644287109375
],
[
0.21240234375,
0.1341552734375,
0.476806640625,
0.1341552734375,
0.476806640625,
0.1478271484375,
0.21240234375,
0.1478271484375
],
[
0.21240234375,
0.149658203125,
0.46337890625,
0.149658203125,
0.46337890625,
0.163330078125,
0.21240234375,
0.163330078125
],
[
0.61279296875,
0.15087890625,
0.64990234375,
0.15087890625,
0.64990234375,
0.163330078125,
0.61279296875,
0.163330078125
],
[
0.68994140625,
0.1484375,
0.73046875,
0.1484375,
0.73046875,
0.1644287109375,
0.68994140625,
0.1644287109375
],
[
0.748046875,
0.15087890625,
0.78662109375,
0.15087890625,
0.78662109375,
0.163330078125,
0.748046875,
0.163330078125
],
[
0.55517578125,
0.1650390625,
0.59619140625,
0.1650390625,
0.59619140625,
0.18115234375,
0.55517578125,
0.18115234375
],
[
0.68994140625,
0.1650390625,
0.73046875,
0.1650390625,
0.73046875,
0.18115234375,
0.68994140625,
0.18115234375
],
[
0.50390625,
0.16748046875,
0.54150390625,
0.16748046875,
0.54150390625,
0.1787109375,
0.50390625,
0.1787109375
],
[
0.21240234375,
0.166259765625,
0.484375,
0.166259765625,
0.484375,
0.179931640625,
0.21240234375,
0.179931640625
],
[
0.63232421875,
0.169189453125,
0.6396484375,
0.174560546875,
0.63232421875,
0.1787109375,
0.62451171875,
0.17333984375
],
[
0.55517578125,
0.18115234375,
0.59619140625,
0.18115234375,
0.59619140625,
0.19775390625,
0.55517578125,
0.19775390625
],
[
0.68994140625,
0.18115234375,
0.73046875,
0.18115234375,
0.73046875,
0.196533203125,
0.68994140625,
0.196533203125
],
[
0.50390625,
0.182861328125,
0.54248046875,
0.182861328125,
0.54248046875,
0.1953125,
0.50390625,
0.1953125
],
[
0.214111328125,
0.1817626953125,
0.453369140625,
0.1817626953125,
0.453369140625,
0.1953125,
0.214111328125,
0.1953125
],
[
0.63037109375,
0.18701171875,
0.634765625,
0.18701171875,
0.634765625,
0.1923828125,
0.63037109375,
0.1923828125
],
[
0.76318359375,
0.1859130859375,
0.77099609375,
0.1859130859375,
0.77099609375,
0.1923828125,
0.76318359375,
0.1923828125
],
[
0.214111328125,
0.19775390625,
0.455078125,
0.19775390625,
0.455078125,
0.2120361328125,
0.214111328125,
0.2120361328125
],
[
0.5029296875,
0.196533203125,
0.54248046875,
0.196533203125,
0.54248046875,
0.2120361328125,
0.5029296875,
0.2120361328125
],
[
0.5556640625,
0.19775390625,
0.59619140625,
0.19775390625,
0.59619140625,
0.213134765625,
0.5556640625,
0.213134765625
],
[
0.61279296875,
0.198974609375,
0.6513671875,
0.198974609375,
0.6513671875,
0.21142578125,
0.61279296875,
0.21142578125
],
[
0.69189453125,
0.198974609375,
0.72900390625,
0.198974609375,
0.72900390625,
0.21142578125,
0.69189453125,
0.21142578125
],
[
0.748046875,
0.19775390625,
0.78857421875,
0.19775390625,
0.78857421875,
0.213134765625,
0.748046875,
0.213134765625
],
[
0.210693359375,
0.218505859375,
0.31982421875,
0.2203369140625,
0.31982421875,
0.235107421875,
0.210693359375,
0.2327880859375
],
[
0.50390625,
0.221435546875,
0.54150390625,
0.221435546875,
0.54150390625,
0.2340087890625,
0.50390625,
0.2340087890625
],
[
0.5576171875,
0.221435546875,
0.59619140625,
0.221435546875,
0.59619140625,
0.2340087890625,
0.5576171875,
0.2340087890625
],
[
0.61279296875,
0.221435546875,
0.6513671875,
0.221435546875,
0.6513671875,
0.2340087890625,
0.61279296875,
0.2340087890625
],
[
0.69189453125,
0.221435546875,
0.72900390625,
0.221435546875,
0.72900390625,
0.2340087890625,
0.69189453125,
0.2340087890625
],
[
0.75,
0.221435546875,
0.78662109375,
0.221435546875,
0.78662109375,
0.2340087890625,
0.75,
0.2340087890625
],
[
0.21240234375,
0.235107421875,
0.361083984375,
0.236328125,
0.361083984375,
0.25048828125,
0.21240234375,
0.2493896484375
],
[
0.5556640625,
0.2340087890625,
0.59619140625,
0.236328125,
0.59423828125,
0.251708984375,
0.55517578125,
0.2493896484375
],
[
0.611328125,
0.236328125,
0.6533203125,
0.236328125,
0.6533203125,
0.251708984375,
0.611328125,
0.251708984375
],
[
0.68994140625,
0.236328125,
0.73046875,
0.236328125,
0.73046875,
0.251708984375,
0.68994140625,
0.251708984375
],
[
0.748046875,
0.236328125,
0.78857421875,
0.236328125,
0.78857421875,
0.251708984375,
0.748046875,
0.251708984375
],
[
0.5029296875,
0.251708984375,
0.5439453125,
0.251708984375,
0.5439453125,
0.267333984375,
0.5029296875,
0.267333984375
],
[
0.50390625,
0.2369384765625,
0.54248046875,
0.2369384765625,
0.54248046875,
0.2493896484375,
0.50390625,
0.2493896484375
],
[
0.5556640625,
0.251708984375,
0.59619140625,
0.251708984375,
0.59619140625,
0.267333984375,
0.5556640625,
0.267333984375
],
[
0.611328125,
0.251708984375,
0.6533203125,
0.251708984375,
0.6533203125,
0.267333984375,
0.611328125,
0.267333984375
],
[
0.68994140625,
0.251708984375,
0.73046875,
0.251708984375,
0.73046875,
0.267333984375,
0.68994140625,
0.267333984375
],
[
0.748046875,
0.251708984375,
0.78857421875,
0.251708984375,
0.78857421875,
0.267333984375,
0.748046875,
0.267333984375
],
[
0.5029296875,
0.267333984375,
0.5439453125,
0.267333984375,
0.5439453125,
0.283203125,
0.5029296875,
0.283203125
],
[
0.21240234375,
0.2529296875,
0.366943359375,
0.2529296875,
0.366943359375,
0.2666015625,
0.21240234375,
0.2666015625
],
[
0.21240234375,
0.268310546875,
0.358642578125,
0.268310546875,
0.358642578125,
0.281982421875,
0.21240234375,
0.281982421875
],
[
0.5576171875,
0.26953125,
0.59423828125,
0.26953125,
0.59423828125,
0.281982421875,
0.5576171875,
0.281982421875
],
[
0.61376953125,
0.26953125,
0.6513671875,
0.26953125,
0.6513671875,
0.281982421875,
0.61376953125,
0.281982421875
],
[
0.68994140625,
0.267333984375,
0.7314453125,
0.267333984375,
0.7314453125,
0.283203125,
0.68994140625,
0.283203125
],
[
0.75,
0.26953125,
0.78662109375,
0.26953125,
0.78662109375,
0.281982421875,
0.75,
0.281982421875
],
[
0.5029296875,
0.283935546875,
0.54248046875,
0.283935546875,
0.54248046875,
0.2998046875,
0.5029296875,
0.2998046875
],
[
0.68994140625,
0.283935546875,
0.7314453125,
0.283935546875,
0.7314453125,
0.2998046875,
0.68994140625,
0.2998046875
],
[
0.748046875,
0.28515625,
0.7900390625,
0.28515625,
0.7900390625,
0.300537109375,
0.748046875,
0.300537109375
],
[
0.5576171875,
0.2861328125,
0.59423828125,
0.2861328125,
0.59423828125,
0.298583984375,
0.5576171875,
0.298583984375
],
[
0.21240234375,
0.28515625,
0.361083984375,
0.28515625,
0.361083984375,
0.298583984375,
0.21240234375,
0.298583984375
],
[
0.61279296875,
0.2861328125,
0.6513671875,
0.2861328125,
0.6513671875,
0.298583984375,
0.61279296875,
0.298583984375
],
[
0.118408203125,
0.314208984375,
0.88330078125,
0.314208984375,
0.88330078125,
0.327880859375,
0.118408203125,
0.327880859375
],
[
0.12005615234375,
0.328857421875,
0.880859375,
0.328857421875,
0.880859375,
0.342529296875,
0.12005615234375,
0.342529296875
],
[
0.118408203125,
0.342529296875,
0.295654296875,
0.34326171875,
0.295654296875,
0.358154296875,
0.118408203125,
0.35693359375
],
[
0.69287109375,
0.385498046875,
0.74658203125,
0.38720703125,
0.74462890625,
0.402587890625,
0.69189453125,
0.40087890625
],
[
0.82373046875,
0.385498046875,
0.884765625,
0.388427734375,
0.88330078125,
0.40380859375,
0.82275390625,
0.40087890625
],
[
0.75732421875,
0.385986328125,
0.8095703125,
0.388427734375,
0.8076171875,
0.402587890625,
0.755859375,
0.40087890625
],
[
0.1318359375,
0.38720703125,
0.197265625,
0.38720703125,
0.197265625,
0.402099609375,
0.1318359375,
0.402099609375
],
[
0.31494140625,
0.385986328125,
0.337646484375,
0.385986328125,
0.337646484375,
0.402099609375,
0.31494140625,
0.402099609375
],
[
0.3779296875,
0.38720703125,
0.399658203125,
0.38720703125,
0.399658203125,
0.402587890625,
0.3779296875,
0.402587890625
],
[
0.4375,
0.38720703125,
0.462646484375,
0.38720703125,
0.462646484375,
0.402099609375,
0.4375,
0.402099609375
],
[
0.52294921875,
0.38720703125,
0.587890625,
0.38720703125,
0.587890625,
0.402099609375,
0.52294921875,
0.402099609375
],
[
0.133544921875,
0.4091796875,
0.19140625,
0.4091796875,
0.19140625,
0.424560546875,
0.133544921875,
0.424560546875
],
[
0.364501953125,
0.4091796875,
0.411376953125,
0.4091796875,
0.411376953125,
0.424560546875,
0.364501953125,
0.424560546875
],
[
0.4248046875,
0.40673828125,
0.474365234375,
0.4091796875,
0.47265625,
0.424560546875,
0.424072265625,
0.4228515625
],
[
0.5224609375,
0.4091796875,
0.5810546875,
0.4091796875,
0.5810546875,
0.42333984375,
0.5224609375,
0.42333984375
],
[
0.69873046875,
0.4091796875,
0.73779296875,
0.4091796875,
0.73779296875,
0.424560546875,
0.69873046875,
0.424560546875
],
[
0.78076171875,
0.4150390625,
0.78662109375,
0.4150390625,
0.78662109375,
0.4228515625,
0.78076171875,
0.4228515625
],
[
0.85205078125,
0.41748046875,
0.85986328125,
0.419189453125,
0.8564453125,
0.42333984375,
0.8486328125,
0.42041015625
],
[
0.30322265625,
0.410400390625,
0.35107421875,
0.410400390625,
0.35107421875,
0.424560546875,
0.30322265625,
0.424560546875
],
[
0.1318359375,
0.42578125,
0.2435302734375,
0.42578125,
0.2435302734375,
0.440673828125,
0.1318359375,
0.440673828125
],
[
0.301513671875,
0.42578125,
0.3525390625,
0.42578125,
0.3525390625,
0.440673828125,
0.301513671875,
0.440673828125
],
[
0.364501953125,
0.42578125,
0.4130859375,
0.42578125,
0.4130859375,
0.440673828125,
0.364501953125,
0.440673828125
],
[
0.4248046875,
0.42578125,
0.474365234375,
0.42578125,
0.474365234375,
0.441162109375,
0.4248046875,
0.441162109375
],
[
0.52294921875,
0.42578125,
0.634765625,
0.42578125,
0.634765625,
0.440673828125,
0.52294921875,
0.440673828125
],
[
0.7001953125,
0.427001953125,
0.736328125,
0.427001953125,
0.736328125,
0.439453125,
0.7001953125,
0.439453125
],
[
0.52294921875,
0.441162109375,
0.5810546875,
0.441162109375,
0.5810546875,
0.4560546875,
0.52294921875,
0.4560546875
],
[
0.76416015625,
0.441162109375,
0.80419921875,
0.441162109375,
0.80419921875,
0.457275390625,
0.76416015625,
0.457275390625
],
[
0.83642578125,
0.441162109375,
0.8759765625,
0.441162109375,
0.8759765625,
0.457275390625,
0.83642578125,
0.457275390625
],
[
0.301513671875,
0.44775390625,
0.35107421875,
0.44775390625,
0.35107421875,
0.463134765625,
0.301513671875,
0.463134765625
],
[
0.4248046875,
0.44775390625,
0.474365234375,
0.44775390625,
0.474365234375,
0.463134765625,
0.4248046875,
0.463134765625
],
[
0.364501953125,
0.448974609375,
0.411376953125,
0.448974609375,
0.411376953125,
0.463134765625,
0.364501953125,
0.463134765625
],
[
0.133544921875,
0.448974609375,
0.240966796875,
0.448974609375,
0.240966796875,
0.463134765625,
0.133544921875,
0.463134765625
],
[
0.52490234375,
0.458984375,
0.634765625,
0.458984375,
0.634765625,
0.47021484375,
0.52490234375,
0.47021484375
],
[
0.765625,
0.458984375,
0.8037109375,
0.458984375,
0.8037109375,
0.471435546875,
0.765625,
0.471435546875
],
[
0.8369140625,
0.458984375,
0.873046875,
0.458984375,
0.873046875,
0.471435546875,
0.8369140625,
0.471435546875
],
[
0.301513671875,
0.46435546875,
0.3525390625,
0.46435546875,
0.3525390625,
0.478515625,
0.301513671875,
0.478515625
],
[
0.133544921875,
0.46435546875,
0.28369140625,
0.46435546875,
0.28369140625,
0.478515625,
0.133544921875,
0.478515625
],
[
0.364501953125,
0.46435546875,
0.4130859375,
0.46435546875,
0.4130859375,
0.478515625,
0.364501953125,
0.478515625
],
[
0.4248046875,
0.46435546875,
0.474365234375,
0.46435546875,
0.474365234375,
0.478515625,
0.4248046875,
0.478515625
],
[
0.71533203125,
0.46435546875,
0.72119140625,
0.46435546875,
0.72119140625,
0.47021484375,
0.71533203125,
0.47021484375
],
[
0.301513671875,
0.479736328125,
0.3525390625,
0.479736328125,
0.3525390625,
0.495361328125,
0.301513671875,
0.495361328125
],
[
0.69775390625,
0.479736328125,
0.73779296875,
0.479736328125,
0.73779296875,
0.495361328125,
0.69775390625,
0.495361328125
],
[
0.76416015625,
0.479736328125,
0.80419921875,
0.479736328125,
0.80419921875,
0.495361328125,
0.76416015625,
0.495361328125
],
[
0.83642578125,
0.479736328125,
0.8759765625,
0.479736328125,
0.8759765625,
0.495361328125,
0.83642578125,
0.495361328125
],
[
0.364501953125,
0.48095703125,
0.4130859375,
0.48095703125,
0.4130859375,
0.495361328125,
0.364501953125,
0.495361328125
],
[
0.1351318359375,
0.48095703125,
0.2880859375,
0.48095703125,
0.2880859375,
0.49462890625,
0.1351318359375,
0.49462890625
],
[
0.4248046875,
0.48095703125,
0.474365234375,
0.48095703125,
0.474365234375,
0.495361328125,
0.4248046875,
0.495361328125
],
[
0.52294921875,
0.48095703125,
0.62890625,
0.48095703125,
0.62890625,
0.495361328125,
0.52294921875,
0.495361328125
],
[
0.1351318359375,
0.496337890625,
0.280517578125,
0.496337890625,
0.280517578125,
0.51123046875,
0.1351318359375,
0.51123046875
],
[
0.301513671875,
0.496337890625,
0.3525390625,
0.496337890625,
0.3525390625,
0.51123046875,
0.301513671875,
0.51123046875
],
[
0.364501953125,
0.496337890625,
0.4130859375,
0.496337890625,
0.4130859375,
0.51123046875,
0.364501953125,
0.51123046875
],
[
0.4248046875,
0.496337890625,
0.474365234375,
0.496337890625,
0.474365234375,
0.51123046875,
0.4248046875,
0.51123046875
],
[
0.52490234375,
0.496337890625,
0.67333984375,
0.496337890625,
0.67333984375,
0.51123046875,
0.52490234375,
0.51123046875
],
[
0.69873046875,
0.496337890625,
0.73779296875,
0.496337890625,
0.73779296875,
0.5126953125,
0.69873046875,
0.5126953125
],
[
0.765625,
0.49755859375,
0.8037109375,
0.49755859375,
0.8037109375,
0.51025390625,
0.765625,
0.51025390625
],
[
0.83642578125,
0.495361328125,
0.8759765625,
0.495361328125,
0.8759765625,
0.51123046875,
0.83642578125,
0.51123046875
],
[
0.301513671875,
0.5126953125,
0.35107421875,
0.5126953125,
0.35107421875,
0.52685546875,
0.301513671875,
0.52685546875
],
[
0.364501953125,
0.5126953125,
0.4130859375,
0.5126953125,
0.4130859375,
0.52685546875,
0.364501953125,
0.52685546875
],
[
0.4248046875,
0.5126953125,
0.474365234375,
0.5126953125,
0.474365234375,
0.52685546875,
0.4248046875,
0.52685546875
],
[
0.69775390625,
0.5126953125,
0.73779296875,
0.5126953125,
0.73779296875,
0.52783203125,
0.69775390625,
0.52783203125
],
[
0.76416015625,
0.5126953125,
0.80419921875,
0.5126953125,
0.80419921875,
0.52783203125,
0.76416015625,
0.52783203125
],
[
0.83642578125,
0.5126953125,
0.8759765625,
0.5126953125,
0.8759765625,
0.52783203125,
0.83642578125,
0.52783203125
],
[
0.52490234375,
0.51318359375,
0.6767578125,
0.51318359375,
0.6767578125,
0.52685546875,
0.52490234375,
0.52685546875
],
[
0.133544921875,
0.51318359375,
0.28369140625,
0.51318359375,
0.28369140625,
0.52685546875,
0.133544921875,
0.52685546875
],
[
0.52294921875,
0.529296875,
0.66748046875,
0.529296875,
0.66748046875,
0.54345703125,
0.52294921875,
0.54345703125
],
[
0.7001953125,
0.52978515625,
0.73779296875,
0.52978515625,
0.73779296875,
0.5419921875,
0.7001953125,
0.5419921875
],
[
0.76416015625,
0.52783203125,
0.80419921875,
0.52783203125,
0.80419921875,
0.54345703125,
0.76416015625,
0.54345703125
],
[
0.8369140625,
0.52978515625,
0.875,
0.52978515625,
0.875,
0.5419921875,
0.8369140625,
0.5419921875
],
[
0.118408203125,
0.54345703125,
0.486083984375,
0.54345703125,
0.486083984375,
0.55419921875,
0.118408203125,
0.55419921875
],
[
0.52490234375,
0.54443359375,
0.67333984375,
0.54443359375,
0.67333984375,
0.55859375,
0.52490234375,
0.55859375
],
[
0.7001953125,
0.5458984375,
0.736328125,
0.5458984375,
0.736328125,
0.55810546875,
0.7001953125,
0.55810546875
],
[
0.765625,
0.5458984375,
0.8037109375,
0.5458984375,
0.8037109375,
0.55810546875,
0.765625,
0.55810546875
],
[
0.83642578125,
0.54345703125,
0.875,
0.54345703125,
0.875,
0.55859375,
0.83642578125,
0.55859375
],
[
0.118408203125,
0.55712890625,
0.489501953125,
0.5556640625,
0.489501953125,
0.57080078125,
0.118408203125,
0.57177734375
],
[
0.1175537109375,
0.57080078125,
0.48779296875,
0.57080078125,
0.48779296875,
0.58447265625,
0.1175537109375,
0.58447265625
],
[
0.515625,
0.576171875,
0.58447265625,
0.576171875,
0.58447265625,
0.58740234375,
0.515625,
0.58740234375
],
[
0.587890625,
0.580078125,
0.59423828125,
0.580078125,
0.59423828125,
0.58447265625,
0.587890625,
0.58447265625
],
[
0.59033203125,
0.576171875,
0.88330078125,
0.576171875,
0.88330078125,
0.59033203125,
0.59033203125,
0.59033203125
],
[
0.118408203125,
0.5859375,
0.44921875,
0.5859375,
0.44921875,
0.599609375,
0.118408203125,
0.599609375
],
[
0.51318359375,
0.5888671875,
0.8818359375,
0.5888671875,
0.8818359375,
0.6025390625,
0.51318359375,
0.6025390625
],
[
0.5146484375,
0.60400390625,
0.880859375,
0.60400390625,
0.880859375,
0.61767578125,
0.5146484375,
0.61767578125
],
[
0.5146484375,
0.6181640625,
0.88330078125,
0.6181640625,
0.88330078125,
0.6318359375,
0.5146484375,
0.6318359375
],
[
0.118408203125,
0.626953125,
0.162841796875,
0.626953125,
0.162841796875,
0.64111328125,
0.118408203125,
0.64111328125
],
[
0.1796875,
0.62890625,
0.2418212890625,
0.62890625,
0.2418212890625,
0.64013671875,
0.1796875,
0.64013671875
],
[
0.5146484375,
0.6318359375,
0.884765625,
0.6318359375,
0.884765625,
0.646484375,
0.5146484375,
0.646484375
],
[
0.12005615234375,
0.646484375,
0.489501953125,
0.646484375,
0.489501953125,
0.6611328125,
0.12005615234375,
0.6611328125
],
[
0.509765625,
0.64794921875,
0.58251953125,
0.6455078125,
0.58251953125,
0.6591796875,
0.509765625,
0.6611328125
],
[
0.1175537109375,
0.6630859375,
0.489501953125,
0.662109375,
0.489501953125,
0.6767578125,
0.1175537109375,
0.677734375
],
[
0.118408203125,
0.677734375,
0.47265625,
0.6787109375,
0.47265625,
0.693359375,
0.118408203125,
0.6923828125
],
[
0.466796875,
0.68408203125,
0.48193359375,
0.681640625,
0.483642578125,
0.6884765625,
0.468505859375,
0.6904296875
],
[
0.12176513671875,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.12176513671875,
0.708984375
],
[
0.5283203125,
0.70166015625,
0.8818359375,
0.70166015625,
0.8818359375,
0.71533203125,
0.5283203125,
0.71533203125
],
[
0.12005615234375,
0.71142578125,
0.489501953125,
0.71142578125,
0.489501953125,
0.7255859375,
0.12005615234375,
0.7255859375
],
[
0.5146484375,
0.71630859375,
0.8818359375,
0.71630859375,
0.8818359375,
0.73095703125,
0.5146484375,
0.73095703125
],
[
0.118408203125,
0.7255859375,
0.48779296875,
0.72705078125,
0.48779296875,
0.7412109375,
0.118408203125,
0.74072265625
],
[
0.5146484375,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.5146484375,
0.74755859375
],
[
0.118408203125,
0.7421875,
0.258544921875,
0.74365234375,
0.258544921875,
0.75830078125,
0.118408203125,
0.75732421875
],
[
0.5146484375,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.15283203125,
0.77099609375,
0.2493896484375,
0.7724609375,
0.2493896484375,
0.78759765625,
0.15283203125,
0.78515625
],
[
0.51904296875,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.51904296875,
0.77978515625
],
[
0.12005615234375,
0.77392578125,
0.133544921875,
0.77392578125,
0.133544921875,
0.78369140625,
0.12005615234375,
0.78369140625
],
[
0.5146484375,
0.78076171875,
0.8818359375,
0.7822265625,
0.8818359375,
0.79638671875,
0.5146484375,
0.7958984375
],
[
0.12005615234375,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.12005615234375,
0.8125
],
[
0.5146484375,
0.79736328125,
0.8818359375,
0.79736328125,
0.8818359375,
0.8125,
0.5146484375,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.8291015625,
0.12005615234375,
0.8291015625
],
[
0.51318359375,
0.81396484375,
0.88330078125,
0.81396484375,
0.88330078125,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.118408203125,
0.830078125,
0.48779296875,
0.83056640625,
0.48779296875,
0.845703125,
0.118408203125,
0.84423828125
],
[
0.5146484375,
0.83056640625,
0.880859375,
0.83056640625,
0.880859375,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.133544921875,
0.8466796875,
0.489501953125,
0.845703125,
0.489501953125,
0.85986328125,
0.133544921875,
0.86083984375
],
[
0.5146484375,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.1275634765625,
0.8525390625,
0.1368408203125,
0.8525390625,
0.1368408203125,
0.85693359375,
0.1275634765625,
0.85693359375
],
[
0.12005615234375,
0.8623046875,
0.489501953125,
0.8623046875,
0.489501953125,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.515625,
0.8642578125,
0.88330078125,
0.8642578125,
0.88330078125,
0.87451171875,
0.515625,
0.87451171875
],
[
0.12005615234375,
0.87890625,
0.48779296875,
0.87890625,
0.48779296875,
0.892578125,
0.12005615234375,
0.892578125
],
[
0.5146484375,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.12005615234375,
0.89453125,
0.48779296875,
0.89453125,
0.48779296875,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.515625,
0.89453125,
0.705078125,
0.89453125,
0.705078125,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"Aspect",
"Sentiment",
"Model",
"Acc.",
"AUC",
"AUC",
"Acc.",
"F1",
"LR (Saeidi et al., 2016)",
"39.3",
"92.4",
"87.5",
"90.5",
"68.9",
"89.8",
"82.0",
"85.4",
"69.3",
"LSTM-Final (Saeidi et al., 2016)",
"LSTM-Loc (Saeidi et al., 2016)",
"89.7",
"81.9",
"83.9",
"76.7",
"86.8",
"66.4",
"LSTM+TA+SA (Ma et al., 2018)",
".",
"78.2",
"89.3",
"67.4",
"SenticLSTM (Ma et al., 2018)",
"",
"-",
"Dmu-Entnet (Liu et al., 2018)",
"73.5",
"78.5",
"94.4",
"91.0",
"94.8",
"BERT-single",
"73.7",
"81.0",
"96.4",
"85.5",
"84.2",
"BERT-pair-QA-M",
"86.4",
"97.0",
"93.6",
"96.4",
"78.3",
"79.4",
"87.0",
"97.5",
"92.1",
"96.5",
"79.2",
"BERT-pair-NLI-M",
"BERT-pair-QA-B",
"87.9",
"97.1",
"93.3",
"97.0",
"79.8",
"92.8",
"96.9",
"87.5",
"BERT-pair-NLI-B",
"96.6",
"Table 3:Performance on SentiHood dataset. We boldface the score with the best performance across",
"all models. We use the results reported in Saeidi et al. (2016) , Ma et al. (2018) and Liu et al. (2018)",
"\"_\" means not reported.",
"4-way",
"Binary",
"3-way",
" Models",
"P",
"R",
"F1",
"Models",
"XRCE",
"81.37",
"82.29",
"XRCE",
"78.1",
"",
"2",
"83.23",
" NRC-Canada",
"91.04",
"86.24",
"88.58",
"NRC-Canada",
"82.9",
"LSTM",
"82.0",
"88.3",
"92.78",
"90.89",
"89.07",
"BERT-single",
"ATAE-LSTM",
"84.0",
"89.9",
"92.87",
"BERT-pair-QA-M",
"90.24",
"91.54",
"-",
"93.15",
"83.7",
"86.9",
"93.3",
"90.24",
"BERT-pair-NLI-M",
"91.67",
"BERT-single",
"BERT-pair-QA-B",
"93.04",
"89.95",
"91.47",
"BERT-pair-QA-M",
"85.2",
"89.3",
"95.4",
"93.57",
"90.83",
"92.18",
"85.1",
"88.7",
"94.4",
"BERT-pair-NLI-M",
"BERT-pair-NLI-B",
"BERT-pair-QA-B",
"85.9",
"89.9",
"95.6",
"Table 4:Test setresultsforSemeval-2014task",
"BERT-pair-NLI-B",
"84.6",
"88.7",
"95.1",
"4 Subtask 3: Aspect Category Detection. We use",
"the results reported in XRCE (Brun et al., 2014)",
"Table 5:",
"",
"Test set accuracy （% for Semeval-",
"and NRC-Canada (Kiritchenko et al., 2014)",
"2014 task 4 Subtask 4: Aspect Category Polarity.",
"We use the results reported in XRCE (Brun et al..",
"2014) , NRC-Canada (Kiritchenko et al., 2014) and",
"3.4.1",
"Results",
"ATAE-LSTM (Wang et al., 2016) . \"~\" means not",
"结果如表4和表5所示。我们发现BERT-",
"reported.",
"single在这两个子任务上取得了更好的效果，",
"并且BERT-pair比BERT-single取得了进",
"",
"步的改进。BERT-pair-NLI-B模型实现了方",
"个句子预测任务。由于附加的目标和方面信",
"面类别检测的最佳性能。对于方面类别极性，",
"息，TABSA 比SA更复杂。在TABSA 上直",
"BERT-pair-QA-B在所有4路、3路和二进制",
"接微调预先训练的BERT无法实现性能增长。",
"设置上表现最佳。",
"但是，当我们将目标和方面分开以形成辅助句",
"Discussion",
"子并将TABSA转换为句子对分类任务时，场",
"4",
"景类似于QA和NLI，然后可以充分利用预训",
"将哪个方面的目标语料库。原始数据集中",
"练的BERT模型的优势。我们的方法不限于",
"的一个句子si将扩展为句子对分类任务中",
"TABSA，这种构造方法可以用于其他类似的任",
"的(si,ti,ai),..,(si,ti,ana),..,(si,tnt,ana)",
"务。对于ABSA，我们可以使用相同的方法来",
"另一方面，从BERT模型对QA和NLI任",
"构造仅包含方面的辅助句子在BERT-pair模",
"",
"务（Devlinetal.，2018）的惊人改进可以看出，",
"型中，BERT-pair-QA-B和BERT-pair-NLI-B",
"BERT模型在处理句子对分类任务方面具有",
"在情感分类上实现了更好的AUC值，这可能",
"优势。此优势来自无监督掩码语言模型和下",
"是由于标签信息的建模。"
] |
{
"arxiv_id": "1903.09588",
"language": "zh",
"page_idx": 3
}
|
||
2304.04496-zh-0
|
2304.04496-zh-0.jpg
|
DeFeeNet: Consecutive 3D Human Motion Prediction with Deviation Feedback
Xiaoning Sun$^1,\text{Huaijiang Sun}^{1\boxtimes},\text{Bin Li}^{2},\text{Dong Wei}^{1},\text{Weiqing Li}^{1},\text{Jianfeng Lu}^{1}$
$^{1}$Nanjing University of Science and Technology, China
$^{2}$Tianjin AiForward Science and Technology Co., Ltd., China
$\{$ sunxiaoning,sunhuaijiang $\}$ @njust.edu.cn, libin@aiforward.com
<FigureHere>
Figure 1. (a) Current human motion prediction task: a one-off “observe then predict" process within the isolated unit. (b) Consecutive human motion prediction with deviation feedback: covering multiple“observe then predict” rounds, where the prediction deviation (i.e., mistake) in previous round could constrain the model to make better prediction in the following round.
Abstract
into the existing predictor to enable a deviation-aware prediction manner, which, for the first time, allows for information transmit across adjacent prediction units. We design two versions of DeFeeNet as MLP-based and GRU-based, respectively. On Human3.6M and more complicated BABEL, experimental results indicate that our proposed network improves consecutive human motion prediction performance regardless of the basic model.
## 1. Introduction
Let us rethink the real-world scenarios that require human motion prediction techniques, such as humanrobot collaboration. Current works simplify the task of predicting human motions into a one-off process of forecasting a short future sequence (usually no longer than 1 second) based on a historical observed one. However, such simplification may fail to meet practical needs due to the neglect of the fact that motion prediction in real applications is not an isolated“observe then predict" unit, but a consecutive process composed of many rounds of such unit, semi-overlapped along the entire sequence. As time goes on, the predicted part of previous round has its corresponding ground truth observable in the new round, but their deviation inbetween is neither exploited nor able to be captured by existing isolated learning fashion. In this paper, we propose DeFeeNet, a simple yet effective network that can be added on existing one-off prediction models to realize deviation perception and feedback when applied to consecutive motion prediction task. At each prediction round, the deviation generated by previous unit is first encoded by our DeFeeNet, and then incorporated
在智能时代，人类必须在持续的一段时间内与机器人，机器或自主系统共享一个公共空间，例如在人类运动协作[33] ,运动跟踪[10]和自动驾驶[38]场景中。为了满足人类的需求，同时保持其安全性，确定性人体运动预测旨在预测给定历史观测的未来人体姿态序列，已成为研究热点，并已形成相对完整的实现程序。随着时间的流逝，人类的姿势可能变得难以预测，因此当前的工作将实际需求抽象为学习“观察几帧，然后预测以下帧”的简化任务，预测长度主要设置为[33] 秒$[4,7–9,16,20,21,24,27,28,34,40]$
从本质上讲，这种简化的预测任务可以被视为限制
|
[
[
0.0882568359375,
0.1337890625,
0.8857421875,
0.1337890625,
0.8857421875,
0.1483154296875,
0.0882568359375,
0.1483154296875
],
[
0.1094970703125,
0.177978515625,
0.859375,
0.177978515625,
0.859375,
0.1988525390625,
0.1094970703125,
0.1988525390625
],
[
0.2548828125,
0.1962890625,
0.72119140625,
0.1988525390625,
0.72119140625,
0.216552734375,
0.2548828125,
0.2139892578125
],
[
0.223876953125,
0.212158203125,
0.7509765625,
0.21533203125,
0.7509765625,
0.236083984375,
0.223876953125,
0.23291015625
],
[
0.26708984375,
0.2374267578125,
0.69775390625,
0.2374267578125,
0.69775390625,
0.250732421875,
0.26708984375,
0.250732421875
],
[
0.12744140625,
0.285888671875,
0.3056640625,
0.285888671875,
0.3056640625,
0.296630859375,
0.12744140625,
0.296630859375
],
[
0.50341796875,
0.2890625,
0.72216796875,
0.290283203125,
0.72216796875,
0.3017578125,
0.50341796875,
0.300537109375
],
[
0.406982421875,
0.31640625,
0.43701171875,
0.31640625,
0.43701171875,
0.32080078125,
0.406982421875,
0.32080078125
],
[
0.5595703125,
0.3193359375,
0.865234375,
0.3193359375,
0.865234375,
0.3515625,
0.5595703125,
0.3515625
],
[
0.108642578125,
0.32763671875,
0.214111328125,
0.32763671875,
0.214111328125,
0.354248046875,
0.108642578125,
0.354248046875
],
[
0.4794921875,
0.330078125,
0.57666015625,
0.32763671875,
0.57666015625,
0.349853515625,
0.4794921875,
0.3515625
],
[
0.72119140625,
0.373779296875,
0.73388671875,
0.373779296875,
0.73388671875,
0.376953125,
0.72119140625,
0.376953125
],
[
0.09722900390625,
0.399658203125,
0.324462890625,
0.399658203125,
0.324462890625,
0.443115234375,
0.09722900390625,
0.443115234375
],
[
0.5009765625,
0.40087890625,
0.86376953125,
0.399658203125,
0.86376953125,
0.43994140625,
0.5009765625,
0.44140625
],
[
0.3798828125,
0.41845703125,
0.488525390625,
0.41845703125,
0.488525390625,
0.436279296875,
0.3798828125,
0.436279296875
],
[
0.08087158203125,
0.468505859375,
0.890625,
0.468505859375,
0.890625,
0.481689453125,
0.08087158203125,
0.481689453125
],
[
0.080078125,
0.481689453125,
0.89208984375,
0.481689453125,
0.89208984375,
0.49560546875,
0.080078125,
0.49560546875
],
[
0.080078125,
0.496826171875,
0.8935546875,
0.496826171875,
0.8935546875,
0.51025390625,
0.080078125,
0.51025390625
],
[
0.07843017578125,
0.51123046875,
0.125,
0.51123046875,
0.125,
0.52294921875,
0.07843017578125,
0.52294921875
],
[
0.236083984375,
0.53515625,
0.314453125,
0.53515625,
0.314453125,
0.55322265625,
0.236083984375,
0.55322265625
],
[
0.50244140625,
0.53857421875,
0.89208984375,
0.53857421875,
0.89208984375,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.50244140625,
0.5537109375,
0.89208984375,
0.5537109375,
0.89208984375,
0.5673828125,
0.50244140625,
0.5673828125
],
[
0.098876953125,
0.5693359375,
0.468017578125,
0.5693359375,
0.468017578125,
0.58349609375,
0.098876953125,
0.58349609375
],
[
0.50244140625,
0.568359375,
0.88818359375,
0.568359375,
0.88818359375,
0.58203125,
0.50244140625,
0.58203125
],
[
0.07843017578125,
0.58349609375,
0.468017578125,
0.583984375,
0.468017578125,
0.5986328125,
0.07843017578125,
0.59765625
],
[
0.50341796875,
0.583984375,
0.89208984375,
0.583984375,
0.89208984375,
0.59765625,
0.50341796875,
0.59765625
],
[
0.080078125,
0.599609375,
0.4697265625,
0.599609375,
0.4697265625,
0.61279296875,
0.080078125,
0.61279296875
],
[
0.5048828125,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.611328125,
0.5048828125,
0.611328125
],
[
0.08087158203125,
0.6142578125,
0.46728515625,
0.6142578125,
0.46728515625,
0.62744140625,
0.08087158203125,
0.62744140625
],
[
0.50244140625,
0.6142578125,
0.89208984375,
0.6142578125,
0.89208984375,
0.62744140625,
0.50244140625,
0.62744140625
],
[
0.080078125,
0.6298828125,
0.4697265625,
0.6298828125,
0.4697265625,
0.64453125,
0.080078125,
0.64453125
],
[
0.50244140625,
0.6298828125,
0.89208984375,
0.6298828125,
0.89208984375,
0.64306640625,
0.50244140625,
0.64306640625
],
[
0.08087158203125,
0.646484375,
0.468017578125,
0.646484375,
0.468017578125,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.50244140625,
0.64453125,
0.89208984375,
0.64453125,
0.89208984375,
0.65771484375,
0.50244140625,
0.65771484375
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.6591796875,
0.59326171875,
0.6591796875,
0.59326171875,
0.67041015625,
0.50341796875,
0.67041015625
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.07843017578125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.07843017578125,
0.7041015625
],
[
0.080078125,
0.70556640625,
0.471435546875,
0.70556640625,
0.471435546875,
0.7197265625,
0.080078125,
0.7197265625
],
[
0.50341796875,
0.70849609375,
0.6396484375,
0.70849609375,
0.6396484375,
0.72265625,
0.50341796875,
0.72265625
],
[
0.080078125,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.080078125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.75,
0.080078125,
0.75
],
[
0.52392578125,
0.734375,
0.890625,
0.734375,
0.890625,
0.74755859375,
0.52392578125,
0.74755859375
],
[
0.080078125,
0.75048828125,
0.468017578125,
0.75048828125,
0.468017578125,
0.76513671875,
0.080078125,
0.76513671875
],
[
0.50634765625,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.50634765625,
0.76318359375
],
[
0.08251953125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.08251953125,
0.77783203125
],
[
0.50341796875,
0.7646484375,
0.890625,
0.76318359375,
0.890625,
0.77783203125,
0.50341796875,
0.77880859375
],
[
0.080078125,
0.78076171875,
0.471435546875,
0.78076171875,
0.471435546875,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.5048828125,
0.77880859375,
0.8896484375,
0.77880859375,
0.8896484375,
0.79248046875,
0.5048828125,
0.79248046875
],
[
0.080078125,
0.796875,
0.471435546875,
0.796875,
0.471435546875,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50244140625,
0.79345703125,
0.8896484375,
0.79345703125,
0.8896484375,
0.806640625,
0.50244140625,
0.806640625
],
[
0.080078125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.080078125,
0.82470703125
],
[
0.50341796875,
0.81005859375,
0.88818359375,
0.81005859375,
0.88818359375,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.080078125,
0.82568359375,
0.4697265625,
0.82568359375,
0.4697265625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.50341796875,
0.82470703125,
0.890625,
0.82470703125,
0.890625,
0.83837890625,
0.50341796875,
0.83837890625
],
[
0.080078125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84033203125,
0.890625,
0.84033203125,
0.890625,
0.85400390625,
0.50341796875,
0.85400390625
],
[
0.080078125,
0.85595703125,
0.46728515625,
0.85595703125,
0.46728515625,
0.869140625,
0.080078125,
0.869140625
],
[
0.50341796875,
0.85498046875,
0.8896484375,
0.85400390625,
0.8896484375,
0.86865234375,
0.50341796875,
0.869140625
],
[
0.080078125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50341796875,
0.87060546875,
0.75732421875,
0.87060546875,
0.75732421875,
0.8837890625,
0.50341796875,
0.8837890625
],
[
0.080078125,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.8994140625,
0.080078125,
0.8994140625
],
[
0.52294921875,
0.88525390625,
0.890625,
0.88623046875,
0.890625,
0.90087890625,
0.52294921875,
0.8994140625
]
] |
[
"DeFeeNet: Consecutive 3D Human Motion Prediction with Deviation Feedback",
"Xiaoning Sun 1, Huaijiang Sun 1 , Bin Li ?, Dong Wei I, Weiqing Li 1, Jianfeng Lu 1",
"1 Nanjing University of Science and Technology, China",
"2 Tianjin AiForward Science and Technology Co., Ltd., China",
"{ sunxiaoning,sunhuaijiang 1 @njust.edu.cn, libin@aiforward.com",
"(a)One-offprediction:isolatedunit",
"(b)Consecutiveprediction:roundbyround",
"raund.t",
"nnRkk",
"",
"力",
"atia",
"专共大人夫",
"ApPPERAAR",
"",
"Figure 1. (a) Current human motion prediction task: a one-off “observe then predict\" process within the isolated unit. (b)",
"Consecutive human motion prediction with deviation feedback: covering multiple “observe then predict\" rounds, where the",
"prediction deviation (i.e., mistake) in previous round could constrain the model to make better prediction in the following",
"round.",
"Abstract",
"into the existing predictor to enable a deviation-aware",
"prediction manner, which, for the first time, allows for",
"Let us rethink the real-world scenarios that require",
"information transmit across adjacent prediction units",
"human motion prediction techniques, such as human-",
"We design two versions of DeFeeNet as MLP-based",
"robot collaboration: Current works simplify the task of",
"andGRU-based.respectively.OnHuman3.6Mand",
"predicting human motions into a one-off process of fore-",
"more complicated BABEL, experimental results indi-",
"casting a short future sequence (usually no longer than",
"cate that our proposed network improves consecutive",
"1 second)based on ahistorical observed one.However,",
"human motion prediction performance regardless of the",
"such simplification may fail to meet practical needs",
"basic model.",
"due to the neglect of the fact that motion prediction",
"in real applications is not an isolated “observe then",
"predict\" unit, but a consecutive process composed of",
"1.Introduction",
"many rounds of such unit, semi-overlapped along the",
"entire sequence. As time goes on, the predicted part",
"在智能时代，人类必须在持续的一段时间内与机器",
"of previous round has its corresponding ground truth",
"人，机器或自主系统共享一个公共空间，例如在人类运",
"observable in the new round.but their deviation in-",
"动协作[33]，运动跟踪[10]和自动驾驶[38]场景中。",
"between is neither exploited nor able to be captured",
"为了满足人类的需求，同时保持其安全性，确定性人",
"by existing isolated learning fashion. In this paper, we",
"体运动预测旨在预测给定历史观测的未来人体姿态序",
"propose DeFeeNet, a simple yet effective network that",
"列，已成为研究热点，并已形成相对完整的实现程序",
"can be added on existing one-off prediction models to",
"随着时间的流逝，人类的姿势可能变得难以预测，因",
"realize deviation perception and feedback when applied",
"此当前的工作将实际需求抽象为学习“观察几顿，然",
"to consecutive motion prediction task. At each predic-",
"后预测以下顿”的简化任务，预测长度主要设置为[33",
"tion round, the deviation generated by previous unit is",
"秒[4,7-9,16,20,21,24,27,28,34,40]",
"first encoded by our DeFeeNet, and then incorporated",
"从本质上讲，这种简化的预测任务可以被视为限制"
] |
{
"arxiv_id": "2304.04496",
"language": "zh",
"page_idx": 0
}
|
||
2304.04496-zh-1
|
2304.04496-zh-1.jpg
|
在短的，孤立的“观察然后预测”单元内的一次性获取目标。然而，该单元实际上不适用于需要在长时间的人-机器人/机器共存期间对人类进行连续观察和预测的现实。虽然直观上，沿着时间逐轮滑动这样的单元可以大致满足连续预测的需要，但一个被忽视的事实是， 每轮预测单元都被安排在半重叠结构中(见图 1)。随着时间的推移，以前预测的东西现在可以观察到相应的地面实况，但是它们之间的偏差(i。e。,错误)仍未探索
我们的主要见解在于，机器人/机器应该能够检测并从它们所犯的错误中学习。由于人体运动数据固有的连续性和一致性，这种信息将非常强大，可以提高未来的预测精度。半重叠，多轮单元结构提供了一种自然的方式来跨相邻预测单元传输偏差反馈，但是当前的一次性单元预测策略无法实现
基于这种情况，在本文中，我们提出了 DeFeeNet, 这是一个简单而有效的网络，可以将其添加到现有的一次性人体运动预测模型上，以在应用于连续预测时实现偏差感知和费用回退。通过挖掘以前工作中忽略的元素“偏差”,我们引人了归纳偏差，DeFeeNet 可以学习从过去的偏差中得出某些“模式”,然后约束模型以做出更好的预测。具体来说，我们的 DeFeeNet 有两个版本：基于 MLP 的版本包含一个时间混合 MLP 和一个空间混合 MLP;基于 GRU 的版本包含一层GRU,其中全连接层仅用于尺寸对齐。在每个预测轮次中，DeFeeNet 充当插入到现有预测器中的附加分支， 该分支将前一个单元的偏差编码为潜在表示，然后将其传输到主管道中。总之，我们的贡献如下：
$\bullet$ We mine the element“deviation" that neglected in
existing human motion prediction works, extending current within-unit research horizon to crossunit, which, for the frst time, allows for information transmit across adjacent units.
$\bullet$ We propose DeFeeNet, a simple yet effective net-
work that can be added on existing motion prediction models to enable consecutive prediction (i) with flexible round number and (ii) in a deviationaware manner. It can learn to derive certain patterns from past mistakes and constrain the model to make better predictions.
$\bullet$ Our DeFeeNet is agnostic to its basic models, and
capable of yielding stably improved performance on Human3.6M [15] , as well as a more recent and challenging dataset BABEL [32] that contains samples with different actions and their transitions.
# 2. Related Work
2.1. Human Motion Prediction
模型设计范例。人体运动预测模型设计的两个主流是基于顺序网络[7,9,16,28,30]和前馈网络[1,5,8,20,21,27,
33,34,40] 。对于前者，RNNs [7,16] , LSTMs [9,28]和GRUs [?,28,30]用于提取运动序列的时间特征，而对于后者，$\dot{\mathrm{CNNs[20]~,~GNNs~[21]~,~GCNs~[8,27,33-35,40]}}$ 和转换器[1,5]学习空间依赖性或时空信息，以更好地描绘人类骨骼结构和时空关节。所有这些当前的设计都将人类运动预测简化为一个简短的，孤立的“观察然后预测”单元，其预测长度大部分为[7,9,16,28,30]秒， 并且不考虑其上下文，因此无法满足实际需求，例如需要连续观察和预测人类行为的人机协作场景
样本选择协议。目前在这个领域，样本是以一种统一的方式从已知动作类别的长运动序列中切割小块。每个样件(i。e。,单元) 由作为观察的前几帧和作为预测目标的其他帧组成，其中一块中的姿势都属于同一动作类别。因此，这种长度短但单一动作的采样约定确定了当前经过训练的预测器无法处理涵盖多个动作类别和动作变化周期的现实世界预测任务
受上述激励，我们从连续的角度重新制定了人类运动预测任务，并认为以前的预测偏差应由机器人/机器检测和学习，以帮助进行以下预测。因此，我们提出了一种重新采样方法，将多轮灵活的半重叠单元绘制为一个具有相对较长范围的样本(详细信息为$\color{red}{3.4}$ 节),从而实现了(i)我们的偏差感知连续预测以及(ii)首次尝试在确定性人体运动预测中考虑动作变化。请注意，在本文中，我们仅指旨在仅预测一个未来序列的确定性运动预测。随机预测(预测多个可能的序列)[2,25,26,37,39]或运动合成(生成没有观测值的序列) [11,31]不在我们讨论的范围内
## $2. 2. \textbf{ Deviation- Aware Prediction}$
关于偏差感知预测的其他看似相似的想法包括轨迹预测[41]中的本地行为感知 (LBA)框架和节点回归[17],节点分类[14]和流量预测[18,22]中的残差校正框架。对于前者，LBA [41] 是指涉及对象当前位置的所有历史轨迹的集合，以帮助缩小未来轨迹的搜索空间。尽管“其他物体的历史轨迹”概念在某种程度上类似于我们的“前一轮预测”,但它与偏差的使用完全无关。。对于后者，残差被定义为 GT 和预测之间的差异，就像我们的偏差一样，但是所有这些工作首先生成一个“原始”预测，然后将其直接与估计的残差相加，以便可以在一定程度上校正原始预测(请注意，估) 计的残差是根据训练顶点[14,17]的残差或基于历史残差[18,22]计算的)。换句话说，它们的残差校正用作后处理步骤，并且残差表达式完全独立于原始管道。相反，我们采用“编码器分离而解码器统一”的策略，该策略允许对先前的偏差进行单独编码，同时对当前观测值和先前偏差的信息进行联合解码。因此，我们的预测自然会得到改进，而不是通过后处理进行校正，这就是为什么我们将其视为偏差感知预测而不是预测校正的原因。由于人体运动数据更加精细，无法通过简单地添加残差/偏差向量来校正，因此我们的联合解码策略可以利用原始预测器的解码能力，仅在插入轻量级网络分支的情况下才能实现偏差感知预测
|
[
[
0.08087158203125,
0.09344482421875,
0.468017578125,
0.09344482421875,
0.468017578125,
0.106689453125,
0.08087158203125,
0.106689453125
],
[
0.50341796875,
0.09344482421875,
0.890625,
0.09344482421875,
0.890625,
0.106689453125,
0.50341796875,
0.106689453125
],
[
0.0849609375,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.0849609375,
0.12060546875
],
[
0.50341796875,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.08416748046875,
0.125,
0.465576171875,
0.125,
0.465576171875,
0.1351318359375,
0.08416748046875,
0.1351318359375
],
[
0.50244140625,
0.12371826171875,
0.88818359375,
0.12249755859375,
0.88818359375,
0.1363525390625,
0.50244140625,
0.136962890625
],
[
0.08087158203125,
0.1383056640625,
0.46728515625,
0.1383056640625,
0.46728515625,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.5048828125,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.08087158203125,
0.15283203125,
0.46728515625,
0.154052734375,
0.46728515625,
0.1673583984375,
0.08087158203125,
0.1666259765625
],
[
0.50341796875,
0.154052734375,
0.8896484375,
0.154052734375,
0.8896484375,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.08251953125,
0.169189453125,
0.46728515625,
0.169189453125,
0.46728515625,
0.179931640625,
0.08251953125,
0.179931640625
],
[
0.50341796875,
0.1685791015625,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.08416748046875,
0.1849365234375,
0.46728515625,
0.1849365234375,
0.46728515625,
0.1956787109375,
0.08416748046875,
0.1956787109375
],
[
0.50341796875,
0.18310546875,
0.8896484375,
0.18310546875,
0.8896484375,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.08087158203125,
0.1988525390625,
0.468017578125,
0.1988525390625,
0.468017578125,
0.212158203125,
0.08087158203125,
0.212158203125
],
[
0.50341796875,
0.1988525390625,
0.890625,
0.1988525390625,
0.890625,
0.212158203125,
0.50341796875,
0.212158203125
],
[
0.08087158203125,
0.2139892578125,
0.11358642578125,
0.2139892578125,
0.11358642578125,
0.2266845703125,
0.08087158203125,
0.2266845703125
],
[
0.5048828125,
0.2139892578125,
0.81640625,
0.2139892578125,
0.81640625,
0.2279052734375,
0.5048828125,
0.2279052734375
],
[
0.1004638671875,
0.2291259765625,
0.46728515625,
0.2291259765625,
0.46728515625,
0.242431640625,
0.1004638671875,
0.242431640625
],
[
0.5048828125,
0.2342529296875,
0.8896484375,
0.2342529296875,
0.8896484375,
0.2442626953125,
0.5048828125,
0.2442626953125
],
[
0.08251953125,
0.24560546875,
0.46728515625,
0.24560546875,
0.46728515625,
0.255615234375,
0.08251953125,
0.255615234375
],
[
0.50341796875,
0.2474365234375,
0.890625,
0.2474365234375,
0.890625,
0.261474609375,
0.50341796875,
0.261474609375
],
[
0.08251953125,
0.261474609375,
0.46728515625,
0.261474609375,
0.46728515625,
0.271484375,
0.08251953125,
0.271484375
],
[
0.50341796875,
0.26318359375,
0.8896484375,
0.26318359375,
0.8896484375,
0.276611328125,
0.50341796875,
0.276611328125
],
[
0.08251953125,
0.27587890625,
0.46728515625,
0.27587890625,
0.46728515625,
0.285888671875,
0.08251953125,
0.285888671875
],
[
0.50341796875,
0.27783203125,
0.890625,
0.27783203125,
0.890625,
0.291748046875,
0.50341796875,
0.291748046875
],
[
0.08087158203125,
0.2890625,
0.453369140625,
0.2890625,
0.453369140625,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.455078125,
0.29345703125,
0.463134765625,
0.29345703125,
0.463134765625,
0.29736328125,
0.455078125,
0.29736328125
],
[
0.5048828125,
0.29345703125,
0.8896484375,
0.29345703125,
0.8896484375,
0.306884765625,
0.5048828125,
0.306884765625
],
[
0.08087158203125,
0.304931640625,
0.279296875,
0.304931640625,
0.279296875,
0.318115234375,
0.08087158203125,
0.318115234375
],
[
0.5048828125,
0.30810546875,
0.890625,
0.30810546875,
0.890625,
0.3212890625,
0.5048828125,
0.3212890625
],
[
0.101318359375,
0.3193359375,
0.468017578125,
0.3193359375,
0.468017578125,
0.333251953125,
0.101318359375,
0.333251953125
],
[
0.5048828125,
0.324462890625,
0.73046875,
0.324462890625,
0.73046875,
0.335205078125,
0.5048828125,
0.335205078125
],
[
0.08087158203125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.08087158203125,
0.347900390625
],
[
0.52294921875,
0.34033203125,
0.890625,
0.341552734375,
0.890625,
0.35595703125,
0.52294921875,
0.354736328125
],
[
0.08660888671875,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.363037109375,
0.08660888671875,
0.363037109375
],
[
0.50341796875,
0.356689453125,
0.890625,
0.356689453125,
0.890625,
0.37060546875,
0.50341796875,
0.37060546875
],
[
0.08087158203125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.08087158203125,
0.37744140625
],
[
0.50341796875,
0.37255859375,
0.890625,
0.37255859375,
0.890625,
0.386474609375,
0.50341796875,
0.386474609375
],
[
0.08087158203125,
0.380126953125,
0.46728515625,
0.380126953125,
0.46728515625,
0.393310546875,
0.08087158203125,
0.393310546875
],
[
0.50341796875,
0.386962890625,
0.890625,
0.386962890625,
0.890625,
0.40087890625,
0.50341796875,
0.40087890625
],
[
0.08251953125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.08251953125,
0.40771484375
],
[
0.5048828125,
0.404052734375,
0.8896484375,
0.404052734375,
0.8896484375,
0.4140625,
0.5048828125,
0.4140625
],
[
0.08087158203125,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.5048828125,
0.41845703125,
0.890625,
0.41845703125,
0.890625,
0.431884765625,
0.5048828125,
0.431884765625
],
[
0.08251953125,
0.4267578125,
0.46728515625,
0.4267578125,
0.46728515625,
0.4375,
0.08251953125,
0.4375
],
[
0.50244140625,
0.43310546875,
0.890625,
0.43310546875,
0.890625,
0.4462890625,
0.50244140625,
0.4462890625
],
[
0.08087158203125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.08087158203125,
0.453857421875
],
[
0.50341796875,
0.447509765625,
0.890625,
0.447509765625,
0.890625,
0.4609375,
0.50341796875,
0.4609375
],
[
0.080078125,
0.453857421875,
0.468017578125,
0.455078125,
0.468017578125,
0.4697265625,
0.080078125,
0.468505859375
],
[
0.5048828125,
0.462158203125,
0.890625,
0.462158203125,
0.890625,
0.47607421875,
0.5048828125,
0.47607421875
],
[
0.08251953125,
0.47021484375,
0.46728515625,
0.47021484375,
0.46728515625,
0.484130859375,
0.08251953125,
0.484130859375
],
[
0.50341796875,
0.477783203125,
0.8896484375,
0.477783203125,
0.8896484375,
0.4912109375,
0.50341796875,
0.4912109375
],
[
0.08251953125,
0.486083984375,
0.468017578125,
0.486083984375,
0.468017578125,
0.5,
0.08251953125,
0.5
],
[
0.50244140625,
0.4912109375,
0.75244140625,
0.492431640625,
0.75244140625,
0.5068359375,
0.50244140625,
0.505859375
],
[
0.08087158203125,
0.501953125,
0.398681640625,
0.501953125,
0.398681640625,
0.51513671875,
0.08087158203125,
0.51513671875
],
[
0.096435546875,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.096435546875,
0.53857421875
],
[
0.50341796875,
0.52880859375,
0.75341796875,
0.52880859375,
0.75341796875,
0.54248046875,
0.50341796875,
0.54248046875
],
[
0.111083984375,
0.54052734375,
0.46728515625,
0.5390625,
0.46728515625,
0.55322265625,
0.111083984375,
0.5537109375
],
[
0.111083984375,
0.5546875,
0.46728515625,
0.5537109375,
0.46728515625,
0.568359375,
0.111083984375,
0.5693359375
],
[
0.52392578125,
0.55615234375,
0.890625,
0.55615234375,
0.890625,
0.5693359375,
0.52392578125,
0.5693359375
],
[
0.11273193359375,
0.5693359375,
0.46728515625,
0.5693359375,
0.46728515625,
0.58349609375,
0.11273193359375,
0.58349609375
],
[
0.50341796875,
0.57080078125,
0.8896484375,
0.57080078125,
0.8896484375,
0.583984375,
0.50341796875,
0.583984375
],
[
0.11517333984375,
0.5869140625,
0.365234375,
0.5869140625,
0.365234375,
0.59765625,
0.11517333984375,
0.59765625
],
[
0.5048828125,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.5048828125,
0.5986328125
],
[
0.5048828125,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.6142578125,
0.5048828125,
0.6142578125
],
[
0.09478759765625,
0.609375,
0.46728515625,
0.609375,
0.46728515625,
0.62255859375,
0.09478759765625,
0.62255859375
],
[
0.50634765625,
0.6171875,
0.8896484375,
0.6171875,
0.8896484375,
0.62744140625,
0.50634765625,
0.62744140625
],
[
0.11358642578125,
0.62548828125,
0.463134765625,
0.62548828125,
0.463134765625,
0.63623046875,
0.11358642578125,
0.63623046875
],
[
0.50341796875,
0.630859375,
0.890625,
0.630859375,
0.890625,
0.64453125,
0.50341796875,
0.64453125
],
[
0.11358642578125,
0.6396484375,
0.468017578125,
0.6396484375,
0.468017578125,
0.65283203125,
0.11358642578125,
0.65283203125
],
[
0.5048828125,
0.64599609375,
0.890625,
0.64599609375,
0.890625,
0.6591796875,
0.5048828125,
0.6591796875
],
[
0.11517333984375,
0.65576171875,
0.465576171875,
0.65576171875,
0.465576171875,
0.66650390625,
0.11517333984375,
0.66650390625
],
[
0.50341796875,
0.6611328125,
0.890625,
0.6611328125,
0.890625,
0.6748046875,
0.50341796875,
0.6748046875
],
[
0.11358642578125,
0.66943359375,
0.46728515625,
0.66943359375,
0.46728515625,
0.68310546875,
0.11358642578125,
0.68310546875
],
[
0.5048828125,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.5048828125,
0.689453125
],
[
0.11517333984375,
0.68603515625,
0.46728515625,
0.68603515625,
0.46728515625,
0.6962890625,
0.11517333984375,
0.6962890625
],
[
0.5048828125,
0.6923828125,
0.8896484375,
0.6923828125,
0.8896484375,
0.70263671875,
0.5048828125,
0.70263671875
],
[
0.11358642578125,
0.7021484375,
0.303955078125,
0.7021484375,
0.303955078125,
0.7119140625,
0.11358642578125,
0.7119140625
],
[
0.50341796875,
0.70556640625,
0.89208984375,
0.70556640625,
0.89208984375,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.09478759765625,
0.72265625,
0.468017578125,
0.72265625,
0.468017578125,
0.7373046875,
0.09478759765625,
0.7373046875
],
[
0.50341796875,
0.7216796875,
0.89208984375,
0.7216796875,
0.89208984375,
0.73486328125,
0.50341796875,
0.73486328125
],
[
0.11273193359375,
0.73779296875,
0.468017578125,
0.73779296875,
0.468017578125,
0.751953125,
0.11273193359375,
0.751953125
],
[
0.50634765625,
0.73779296875,
0.8896484375,
0.73779296875,
0.8896484375,
0.74853515625,
0.50634765625,
0.74853515625
],
[
0.11358642578125,
0.75244140625,
0.465576171875,
0.75390625,
0.465576171875,
0.767578125,
0.11358642578125,
0.7666015625
],
[
0.5048828125,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.5048828125,
0.76513671875
],
[
0.11358642578125,
0.76806640625,
0.468017578125,
0.76806640625,
0.468017578125,
0.7822265625,
0.11358642578125,
0.7822265625
],
[
0.5048828125,
0.767578125,
0.89208984375,
0.767578125,
0.89208984375,
0.78076171875,
0.5048828125,
0.78076171875
],
[
0.11273193359375,
0.7841796875,
0.46728515625,
0.7841796875,
0.46728515625,
0.79736328125,
0.11273193359375,
0.79736328125
],
[
0.50244140625,
0.78076171875,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.50244140625,
0.79443359375
],
[
0.111083984375,
0.798828125,
0.1544189453125,
0.798828125,
0.1544189453125,
0.81298828125,
0.111083984375,
0.81298828125
],
[
0.5048828125,
0.79736328125,
0.8896484375,
0.79736328125,
0.8896484375,
0.81103515625,
0.5048828125,
0.81103515625
],
[
0.50634765625,
0.81298828125,
0.8896484375,
0.81298828125,
0.8896484375,
0.82373046875,
0.50634765625,
0.82373046875
],
[
0.080078125,
0.82470703125,
0.2279052734375,
0.82470703125,
0.2279052734375,
0.83837890625,
0.080078125,
0.83837890625
],
[
0.5048828125,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.5048828125,
0.84228515625,
0.8896484375,
0.84228515625,
0.8896484375,
0.85595703125,
0.5048828125,
0.85595703125
],
[
0.080078125,
0.8486328125,
0.3203125,
0.8486328125,
0.3203125,
0.8623046875,
0.080078125,
0.8623046875
],
[
0.5048828125,
0.859375,
0.8896484375,
0.859375,
0.8896484375,
0.869140625,
0.5048828125,
0.869140625
],
[
0.08087158203125,
0.87255859375,
0.468017578125,
0.87255859375,
0.468017578125,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.5048828125,
0.87255859375,
0.8896484375,
0.87255859375,
0.8896484375,
0.88623046875,
0.5048828125,
0.88623046875
],
[
0.08087158203125,
0.88623046875,
0.4697265625,
0.88720703125,
0.4697265625,
0.9013671875,
0.08087158203125,
0.90087890625
],
[
0.5048828125,
0.8896484375,
0.779296875,
0.8896484375,
0.779296875,
0.8994140625,
0.5048828125,
0.8994140625
]
] |
[
"在短的，孤立的“观察然后预测”单元内的一次性获取",
"33,34,40]。对于前者，RNNs[7,16]，LSTMs[9,28]和",
"自标。然而，该单元实际上不适用于需要在长时间的",
"GRUs[?,28,30]用于提取运动序列的时间特征，而对于",
"人-机器人/机器共存期间对人类进行连续观察和预测",
"后者，CNNs[20]，GNNs[21]，GCNs8,27,33-35,40",
"的现实。虽然直观上，沿着时间逐轮滑动这样的单元可",
"和转换器「1,5】学习空间依赖性或时空信息，以更好地",
"以大致满足连续预测的需要，但一个被忽视的事实是，",
"描绘人类骨骼结构和时空关节。所有这些当前的设计",
"每轮预测单元都被安排在半重叠结构中（见图1）。随",
"都将人类运动预测简化为一个简短的，孤立的“观察然",
"着时间的推移，以前预测的东西现在可以观察到相应",
"后预测”单元，其预测长度大部分为「7,9,16,28,30]秒",
"的地面实况，但是它们之间的偏差i。e。，错误）仍未",
"并且不考虑其上下文，因此无法满足实际需求，例如需",
"探索",
"要连续观察和预测人类行为的人机协作场景",
"我们的主要见解在于，机器人/机器应该能够检测并",
"样本选择协议。自前在这个领域，样本是以一种统一的",
"从它们所犯的错误中学习。由于人体运动数据固有的",
"方式从已知动作类别的长运动序列中切割小块。每个",
"连续性和一致性，这种信息将非常强大，可以提高未来",
"样件(i。e。，单元）由作为观察的前几顿和作为预测目",
"的预测精度。半重叠，多轮单元结构提供了一种自然的",
"标的其他顿组成，其中一块中的姿势都属于同一动作类",
"方式来跨相邻预测单元传输偏差反馈，但是当前的",
"",
"别。因此，这种长度短但单一动作的采样约定确定了当",
"次性单元预测策略无法实现",
"前经过训练的预测器无法处理涵盖多个动作类别和动",
"基于这种情况，在本文中，我们提出了DeFeeNet，",
"作变化周期的现实世界预测任务",
"这是一个简单而有效的网络，可以将其添加到现有的",
"受上述激励，我们从连续的角度重新制定了人类运",
"一次性人体运动预测模型上，以在应用于连续预测时",
"动预测任务，并认为以前的预测偏差应由机器人/机",
"实现偏差感知和费用回退。通过挖掘以前工作中忽略",
"器检测和学习，以帮助进行以下预测。因此，我们提",
"的元素“偏差”，我们引人了归纳偏差，DeFeeNet可",
"出了一种重新采样方法，将多轮灵活的半重叠单元绘",
"以学习从过去的偏差中得出某些“模式”，然后约束模",
"制为一个具有相对较长范围的样本（详细信息为3.4",
"型以做出更好的预测。具体来说，我们的DeFeeNet有",
"节），从而实现了（i）我们的偏差感知连续预测以及",
"两个版本：基于MLP的版本包含一个时间混合MLP",
"（ii）首次尝试在确定性人体运动预测中考虑动作变化。",
"和一个空间混合MLP；基于GRU的版本包含一层",
"请注意，在本文中，我们仅指旨在仅预测一个未来序",
"GRU，其中全连接层仅用于尺寸对齐。在每个预测轮",
"列的确定性运动预测。随机预测（预测多个可能的序",
"次中，DeFeeNet充当插人到现有预测器中的附加分支",
"列）[2,25,26,37,39]或运动合成（生成没有观测值的序",
"该分支将前一个单元的偏差编码为潜在表示，然后将",
"列）【11,31]不在我们讨论的范围内",
"其传输到主管道中。总之，我们的贡献如下：",
": We mine the element \"deviation\" that neglected in",
"2.2.Deviation-AwarePrediction",
"existing human motion prediction works, extend-",
"ing current within-unit research horizon to cross-",
"关于偏差感知预测的其他看似相似的想法包括轨",
"unit, which, for the first time, allows for informa-",
"迹预测【41]中的本地行为感知（LBA）框架和节点回",
"tion transmit across adjacent units.",
"归[17]，节点分类「14]和流量预测[18,22]中的残差",
"校正框架。对于前者，LBA【4I]是指涉及对象当前位",
": We propose DeFeeNet, a simple yet effective net-",
"置的所有历史轨迹的集合，以帮助缩小未来轨迹的搜",
"work that can be added on existing motion pre",
"索空间。尽管“其他物体的历史轨迹”概念在某种程度",
"diction models to enable consecutive prediction (i)",
"上类似于我们的“前一轮预测”，但它与偏差的使用完",
"withflexibleround numberand (ii)in adeviation-",
"全无关。。对于后者，残差被定义为GT和预测之间的",
"aware manner. It can learn to derive certain pat-",
"差异，就像我们的偏差一样，但是所有这些工作首先",
"terns from past mistakes and constrain the model",
"生成一个“原始”预测，然后将其直接与估计的残差相",
"tomakebetterpredictions",
"加，以便可以在一定程度上校正原始预测（请注意，估",
": Our DeFeeNet is agnostic to its basic models, and",
"计的残差是根据训练顶点[14,17]的残差或基于历史残",
"capable of yielding stably improved performance",
"差18，22计算的）。换句话说，它们的残差校正用作",
"on Human3.6M [15] , as well as a more recent",
"后处理步骤，并且残差表达式完全独立于原始管道。相",
"and challenging dataset BABEL [32] that contains",
"反，我们采用“编码器分离而解码器统一”的策略，该",
"samples with different actions and their transi-",
"策略允许对先前的偏差进行单独编码，同时对当前观",
"tions.",
"测值和先前偏差的信息进行联合解码。因此，我们的预",
"测自然会得到改进，而不是通过后处理进行校正，这就",
"2.Related Work",
"是为什么我们将其视为偏差感知预测而不是预测校正",
"的原因。由于人体运动数据更加精细，无法通过简单地",
"2.1. Human Motion Prediction",
"添加残差/偏差向量来校正，因此我们的联合解码策略",
"模型设计范例。人体运动预测模型设计的两个主流是基",
"可以利用原始预测器的解码能力，仅在插入轻量级网",
"于顺序网络[7,9,16,28,30]和前馈网络[1,5,8,20,21,27,",
"络分支的情况下才能实现偏差感知预测"
] |
{
"arxiv_id": "2304.04496",
"language": "zh",
"page_idx": 1
}
|
||
2304.04496-zh-2
|
2304.04496-zh-2.jpg
|
## 3. Proposed Method
在大多数情况下，3D 人体结构由基于人体关节坐标的纯姿势参数或 SMPL 模型 [23] 表示，该模型描绘了具有姿势和形状的 3D 人体网格。根据任务需求，我们遵循[26,31]丢弃形状项，只预测姿势项。当前，人体运动预测[4,27,34]旨在根据观察到的$N$ -帧$X=[x_1,x_2,\cdots,x_N]$序列，预测具有$T$帧的未来序列$\hat{Y} = [ \hat{y} _{1}, \hat{y} _{2}, \cdots , \hat{y} _{T}]$ ,其中每个帧(i。e。,pose)分别用$\hat{y}_i\in\mathbb{R}^K$和$x_i\in\mathbb{R}^K$表示。我们定义了以$N$帧观测为单位预测$T$帧的过程。然而，在这个孤立单元内的研究完全忽略了“预测错误”元素对连续运动预测的重要性，并且无法实现偏差感知和反馈以提高预测精度
3.1. Reformulation from Consecutive Perspective
我们介绍我们重新制定的人体运动预测任务。连续预测涉及多轮单元，其中$N+T$帧沿整个序列半重叠。假设一个长度为$L$的序列$S=[s_1,s_2,\cdots,s_L]$足够长， 我们将第一轮定义为$R_1=s_{1:N+T}$,接下来的每一轮都是$T$帧，i。e。,第$r$轮$R_r=s_1+(r-1)T:N+T+(r-1)T$ ,其中$r\geqslant 1$ ,在本文中我们设置$N\geqslant T$ 。因此，如果我们定义所有运动回合$R=[R_1,R_2,\cdots,R_r,\cdots]$并重$\begin{array}{l}\text{写 }R_{r}=[X_{r},Y_{r}]=[x_{r,1},\cdots,x_{r,N},y_{r,1},\cdots,y_{r,T}]\:,&\text{则}\\\text{可以推断出 }x_{r,N-T:N}=y_{r-1,1:T}\:,&\text{并且在回合 }R_{r-1}\end{array}$ 中产生的偏差表示为$D_r- 1= f_d( x_{r, N- T: N}, \hat{y} _{r- 1, 1: T})$ , 以用于提高回合$R_r$的预测精度。当$r=1$时，$D_0$为空，因为不能使用偏差。基本上，我们将连续运动预测的任务视为逐轮实现$R$
# 3.2. Deviation-Aware Prediction
为了详细表达 3.1 节中提到的偏差$D_{r-1}=$ $f_{d}( x_{r, N- T: N}, \hat{y} _{r- 1, 1: T})$ ,我们从[4,12]中汲取灵感， 该[4,12]基于两个相邻姿势的关节位移来设计其预测损失函数(i。e。,速度)而不是关节位置。我们不是直接在$x_{r,N-T:N}$和$\hat{y}_{r-1,1:T}$之间进行减法，而是首先将它们的速度计算为$v( x_r, N- T: N)$ $, v( \hat{y} _{r- 1, 1: T}) \in \mathbb{R} ^{K\times ( T- 1) }$ ,然后获得：
$$\begin{aligned}D_{r-1}&=f_d(x_{r,N-T:N},\hat{y}_{r-1,1:T})\\&=v(x_{r,N-T:N})-v(\hat{y}_{r-1,1:T})\in\mathbb{R}^{K\times(T-1)}.\end{aligned}$$
(1)
这样，与专注于学习孤立单元中观测$X$和目标$Y$ 之间的预测函数$Y=\varphi_o(X)$的现有工作相比，我们的偏差感知连续预测旨在为每个预测回合找到一个包含偏差反馈的函数：
$$Y_r=\varphi(X_r,D_{r-1}),$$
(2)
其中$\varphi:\mathbb{R}^K\times N\times\mathbb{R}^{K\times(T-1)}\mapsto\mathbb{R}^{K\times T}$ 。注意$D_0=0$
当$r=1$
在图 2 中，我们提供了 DeFeeNet 如何嵌人到现有预测框架中的概述，并有助于在相对较长的周期内逐轮实现偏差感知的运动预测。在$r$轮，我们将当前一轮观测$X_{r}=x_{r,1:N}$送人主管道( i。e。, 现 有 的 预 测 器 ) 。
同时，我们的 DeFeeNet 可以检测到刚刚发生的预测偏差，如$r-1$轮的预测(i。e。, $\hat{Y}_{r-1}=\hat{y}_{r-1,1:T}$)现在可以观察到它的地面实况。因此，通过编码和合并此偏差$D_{r- 1}$ ,可以自然地改善当前预测$\hat{Y}_r=\hat{y}_{r,1:T}$
## 3.3. Network Architecture
我们基于 MLP 和 GRU 构建了两个版本的 DeFeeNet (请参见图 3),这两个版本都很简单但很有效，因为我们不鼓励 DeFeeNet 在原始预测模型上增加过多的计算负担
受 MLP-Mixer 结构 [36] 在计算机视觉中的成功及其在下游人体运动建模任务 [4,6,12] 中的变体的启发， 我们提出了基于 MLP 的 DeFeeNet 版本，但它更轻巧但足够有效。与在[4,6,12,36]中使用多个 MLP 混合器块相比，我们的设计丢弃了块堆叠，并且仅包含一个时间混合 MLP,一个空间混合 MLP 和跳过连接(请参见左图 3 )。两个 MLP 混音器模块均由两个完全连接的层组成，GELU 激活函数 [13] 和层归一化操作 [3]
对于基于 GRU 的版本，如右图 3 所示，我们采用了一层 GRU, 在时间维度上附加了一个全连接层，在空间维度上附加了一个用于维度对齐的层，以插人到原始管道中进行以下计算
具体来说，首先向原始预测器馈送$R_1$中的观察到的片段$X_1=x_{1,1:N}$,并在时间步长 $T$ 处获得其第一轮预测$\hat{Y}_1=\hat{y}_{1,1:T}$ ,其中 DeFeeNet 未激活且不涉及偏差信息。然而，在接下来的几轮中，$X_r$变得可观察到$,v(\hat{y}_{r-1,1:T})$和$v(x_{r,N-T:N})$之间的偏差出现，并被馈送到我们的 DeFeeNet 中，以捕获其时间和空间信息，并将其添加到主管道中，从而实现相邻预测单元之间的偏差反馈
3.4. Two-Round Training & Multi-Round Testing
我们的 DeFeeNet 的目标是通过对预测偏差的连续感知和反馈来提高原始预测精度，以及在持续的时间段内保持这种优先级。由于半重叠结构为我们提供了一种学习预测偏差的自然方法，因此我们相应地介绍了我们的重采样方法和损失函数
重采样。与现有的将“先观察后预测”单元作为样本的预测模型相比，我们将两轮单元作为一个训练样本，多轮单元作为一个测试样本，i。e。、$R_{\mathrm{train}}=[R_1,R_2]$和$R_{\mathrm{test}}=[R_1,R_2,\cdots,R_{\mathrm{max}(r)}]$。对于训练阶段，我们只是让 DeFeeNet 学习根据其感知的先前相邻偏差来产生有用的信息，因此只有两轮预测(包含一个重叠区域)就足够了。在测试阶段，需要具有多轮但灵活的预测单元的样本，因为我们的最终目标是评估 DeFeeNet 是否能够提前对未知持续时间的连续观察和预测产生稳定的改进
损失函数。在当前作品[4,8,12,27,34]之后，我们的损失基于平均每关节位置误差(MPJPE)[15],但不同的是，我们计算两轮预测的损失。根据$\color{red}{3.1}$节中的表示
|
[
[
0.07843017578125,
0.091552734375,
0.259033203125,
0.09027099609375,
0.259033203125,
0.10479736328125,
0.07843017578125,
0.1060791015625
],
[
0.50341796875,
0.091552734375,
0.890625,
0.0921630859375,
0.890625,
0.106689453125,
0.50341796875,
0.1060791015625
],
[
0.5048828125,
0.10919189453125,
0.8896484375,
0.10919189453125,
0.8896484375,
0.12371826171875,
0.5048828125,
0.12371826171875
],
[
0.101318359375,
0.11614990234375,
0.468017578125,
0.11614990234375,
0.468017578125,
0.1300048828125,
0.101318359375,
0.1300048828125
],
[
0.50341796875,
0.121826171875,
0.890625,
0.12371826171875,
0.890625,
0.136962890625,
0.50341796875,
0.1351318359375
],
[
0.08087158203125,
0.1319580078125,
0.468017578125,
0.1319580078125,
0.468017578125,
0.1458740234375,
0.08087158203125,
0.1458740234375
],
[
0.5009765625,
0.1383056640625,
0.8447265625,
0.1395263671875,
0.8447265625,
0.154052734375,
0.5009765625,
0.15283203125
],
[
0.08251953125,
0.1483154296875,
0.465576171875,
0.1483154296875,
0.465576171875,
0.1590576171875,
0.08251953125,
0.1590576171875
],
[
0.08087158203125,
0.1622314453125,
0.468017578125,
0.1622314453125,
0.468017578125,
0.175537109375,
0.08087158203125,
0.175537109375
],
[
0.08087158203125,
0.1767578125,
0.468017578125,
0.175537109375,
0.468017578125,
0.1900634765625,
0.08087158203125,
0.1912841796875
],
[
0.5009765625,
0.17236328125,
0.7099609375,
0.173583984375,
0.7099609375,
0.1881103515625,
0.5009765625,
0.1875
],
[
0.08087158203125,
0.1925048828125,
0.468017578125,
0.1925048828125,
0.468017578125,
0.20703125,
0.08087158203125,
0.20703125
],
[
0.07843017578125,
0.203857421875,
0.468017578125,
0.205810546875,
0.468017578125,
0.2235107421875,
0.07843017578125,
0.2215576171875
],
[
0.52294921875,
0.2008056640625,
0.890625,
0.199462890625,
0.890625,
0.2139892578125,
0.52294921875,
0.21533203125
],
[
0.08087158203125,
0.2215576171875,
0.468017578125,
0.2215576171875,
0.468017578125,
0.236083984375,
0.08087158203125,
0.236083984375
],
[
0.5009765625,
0.216552734375,
0.88818359375,
0.216552734375,
0.88818359375,
0.2298583984375,
0.5009765625,
0.2298583984375
],
[
0.50341796875,
0.2310791015625,
0.890625,
0.2310791015625,
0.890625,
0.2442626953125,
0.50341796875,
0.2442626953125
],
[
0.08087158203125,
0.2374267578125,
0.468017578125,
0.2374267578125,
0.468017578125,
0.250732421875,
0.08087158203125,
0.250732421875
],
[
0.08087158203125,
0.25244140625,
0.468017578125,
0.25244140625,
0.468017578125,
0.266357421875,
0.08087158203125,
0.266357421875
],
[
0.5048828125,
0.2474365234375,
0.6337890625,
0.2474365234375,
0.6337890625,
0.25830078125,
0.5048828125,
0.25830078125
],
[
0.08251953125,
0.26953125,
0.455078125,
0.26953125,
0.455078125,
0.27978515625,
0.08251953125,
0.27978515625
],
[
0.52294921875,
0.264404296875,
0.890625,
0.264404296875,
0.890625,
0.27783203125,
0.52294921875,
0.27783203125
],
[
0.5048828125,
0.27978515625,
0.890625,
0.27978515625,
0.890625,
0.29345703125,
0.5048828125,
0.29345703125
],
[
0.080078125,
0.292236328125,
0.463134765625,
0.292236328125,
0.463134765625,
0.30615234375,
0.080078125,
0.30615234375
],
[
0.50341796875,
0.294189453125,
0.890625,
0.29345703125,
0.890625,
0.30810546875,
0.50341796875,
0.309326171875
],
[
0.50341796875,
0.31005859375,
0.890625,
0.31005859375,
0.890625,
0.323974609375,
0.50341796875,
0.323974609375
],
[
0.101318359375,
0.317626953125,
0.46728515625,
0.317626953125,
0.46728515625,
0.32763671875,
0.101318359375,
0.32763671875
],
[
0.08251953125,
0.330078125,
0.4697265625,
0.330078125,
0.4697265625,
0.343505859375,
0.08251953125,
0.343505859375
],
[
0.50341796875,
0.32568359375,
0.8896484375,
0.32568359375,
0.8896484375,
0.339111328125,
0.50341796875,
0.339111328125
],
[
0.08087158203125,
0.34521484375,
0.468017578125,
0.34521484375,
0.468017578125,
0.35986328125,
0.08087158203125,
0.35986328125
],
[
0.50244140625,
0.34033203125,
0.890625,
0.34033203125,
0.890625,
0.354248046875,
0.50244140625,
0.354248046875
],
[
0.07843017578125,
0.359130859375,
0.468017578125,
0.35791015625,
0.468017578125,
0.375732421875,
0.07843017578125,
0.376953125
],
[
0.50341796875,
0.354248046875,
0.890625,
0.354736328125,
0.890625,
0.369384765625,
0.50341796875,
0.36865234375
],
[
0.07843017578125,
0.371826171875,
0.46728515625,
0.376953125,
0.46728515625,
0.39453125,
0.07843017578125,
0.389404296875
],
[
0.50341796875,
0.37060546875,
0.890625,
0.37060546875,
0.890625,
0.3837890625,
0.50341796875,
0.3837890625
],
[
0.08660888671875,
0.392578125,
0.46728515625,
0.392578125,
0.46728515625,
0.40283203125,
0.08660888671875,
0.40283203125
],
[
0.52294921875,
0.389404296875,
0.890625,
0.389404296875,
0.890625,
0.40283203125,
0.52294921875,
0.40283203125
],
[
0.08087158203125,
0.406005859375,
0.468017578125,
0.406005859375,
0.468017578125,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.5048828125,
0.404052734375,
0.890625,
0.404052734375,
0.890625,
0.417236328125,
0.5048828125,
0.417236328125
],
[
0.08087158203125,
0.421630859375,
0.468017578125,
0.421630859375,
0.468017578125,
0.436279296875,
0.08087158203125,
0.436279296875
],
[
0.50341796875,
0.419189453125,
0.8896484375,
0.419189453125,
0.8896484375,
0.43310546875,
0.50341796875,
0.43310546875
],
[
0.08251953125,
0.436279296875,
0.46728515625,
0.436279296875,
0.46728515625,
0.45068359375,
0.08251953125,
0.45068359375
],
[
0.5048828125,
0.436279296875,
0.68310546875,
0.436279296875,
0.68310546875,
0.4462890625,
0.5048828125,
0.4462890625
],
[
0.08087158203125,
0.45068359375,
0.45751953125,
0.45263671875,
0.45751953125,
0.466552734375,
0.08087158203125,
0.464111328125
],
[
0.455078125,
0.45703125,
0.464111328125,
0.45703125,
0.464111328125,
0.462158203125,
0.455078125,
0.462158203125
],
[
0.52294921875,
0.451904296875,
0.890625,
0.45263671875,
0.890625,
0.46728515625,
0.52294921875,
0.466552734375
],
[
0.08087158203125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.08087158203125,
0.479736328125
],
[
0.50244140625,
0.46728515625,
0.87744140625,
0.46728515625,
0.87744140625,
0.48486328125,
0.50244140625,
0.48486328125
],
[
0.08087158203125,
0.481689453125,
0.46728515625,
0.481689453125,
0.46728515625,
0.49560546875,
0.08087158203125,
0.49560546875
],
[
0.50341796875,
0.48486328125,
0.8896484375,
0.48486328125,
0.8896484375,
0.5,
0.50341796875,
0.5
],
[
0.08087158203125,
0.496826171875,
0.2467041015625,
0.496826171875,
0.2467041015625,
0.51025390625,
0.08087158203125,
0.51025390625
],
[
0.50634765625,
0.501953125,
0.8896484375,
0.501953125,
0.8896484375,
0.51220703125,
0.50634765625,
0.51220703125
],
[
0.08087158203125,
0.521484375,
0.330810546875,
0.521484375,
0.330810546875,
0.53515625,
0.08087158203125,
0.53515625
],
[
0.5048828125,
0.5166015625,
0.890625,
0.5166015625,
0.890625,
0.52978515625,
0.5048828125,
0.52978515625
],
[
0.50341796875,
0.52880859375,
0.890625,
0.52978515625,
0.890625,
0.54443359375,
0.50341796875,
0.54345703125
],
[
0.101318359375,
0.54345703125,
0.46728515625,
0.54345703125,
0.46728515625,
0.556640625,
0.101318359375,
0.556640625
],
[
0.5048828125,
0.54541015625,
0.88818359375,
0.54541015625,
0.88818359375,
0.5595703125,
0.5048828125,
0.5595703125
],
[
0.080078125,
0.56005859375,
0.468017578125,
0.55810546875,
0.468017578125,
0.57275390625,
0.080078125,
0.57470703125
],
[
0.5048828125,
0.5625,
0.60107421875,
0.5625,
0.60107421875,
0.57373046875,
0.5048828125,
0.57373046875
],
[
0.08087158203125,
0.57470703125,
0.46728515625,
0.57470703125,
0.46728515625,
0.58837890625,
0.08087158203125,
0.58837890625
],
[
0.08087158203125,
0.5888671875,
0.46728515625,
0.5888671875,
0.46728515625,
0.60302734375,
0.08087158203125,
0.60302734375
],
[
0.50244140625,
0.5947265625,
0.88818359375,
0.5966796875,
0.88818359375,
0.611328125,
0.50244140625,
0.609375
],
[
0.08087158203125,
0.60498046875,
0.465576171875,
0.60498046875,
0.465576171875,
0.619140625,
0.08087158203125,
0.619140625
],
[
0.0767822265625,
0.6162109375,
0.468017578125,
0.6123046875,
0.468017578125,
0.63623046875,
0.0767822265625,
0.64013671875
],
[
0.52294921875,
0.62353515625,
0.890625,
0.62353515625,
0.890625,
0.63720703125,
0.52294921875,
0.63720703125
],
[
0.07843017578125,
0.63525390625,
0.1707763671875,
0.63330078125,
0.1707763671875,
0.64794921875,
0.07843017578125,
0.6494140625
],
[
0.5048828125,
0.63818359375,
0.8896484375,
0.63818359375,
0.8896484375,
0.6513671875,
0.5048828125,
0.6513671875
],
[
0.50341796875,
0.65380859375,
0.88818359375,
0.65380859375,
0.88818359375,
0.66748046875,
0.50341796875,
0.66748046875
],
[
0.0931396484375,
0.6611328125,
0.310546875,
0.662109375,
0.310546875,
0.6767578125,
0.0931396484375,
0.67529296875
],
[
0.44775390625,
0.67529296875,
0.464111328125,
0.67529296875,
0.464111328125,
0.68505859375,
0.44775390625,
0.68505859375
],
[
0.50732421875,
0.67041015625,
0.8896484375,
0.67041015625,
0.8896484375,
0.68115234375,
0.50732421875,
0.68115234375
],
[
0.13232421875,
0.681640625,
0.438720703125,
0.6767578125,
0.438720703125,
0.69775390625,
0.13232421875,
0.70263671875
],
[
0.4306640625,
0.68310546875,
0.443603515625,
0.6796875,
0.4453125,
0.68310546875,
0.431396484375,
0.68603515625
],
[
0.50634765625,
0.6845703125,
0.73193359375,
0.6845703125,
0.73193359375,
0.69775390625,
0.50634765625,
0.69775390625
],
[
0.50634765625,
0.7041015625,
0.8896484375,
0.7041015625,
0.8896484375,
0.7138671875,
0.50634765625,
0.7138671875
],
[
0.1004638671875,
0.71044921875,
0.465576171875,
0.71044921875,
0.465576171875,
0.7236328125,
0.1004638671875,
0.7236328125
],
[
0.50244140625,
0.71630859375,
0.890625,
0.71728515625,
0.890625,
0.73095703125,
0.50244140625,
0.72998046875
],
[
0.08087158203125,
0.7265625,
0.468017578125,
0.7265625,
0.468017578125,
0.74072265625,
0.08087158203125,
0.74072265625
],
[
0.50341796875,
0.73291015625,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.50341796875,
0.7470703125
],
[
0.08251953125,
0.7412109375,
0.468017578125,
0.7412109375,
0.468017578125,
0.7548828125,
0.08251953125,
0.7548828125
],
[
0.50341796875,
0.7470703125,
0.890625,
0.74755859375,
0.890625,
0.76220703125,
0.50341796875,
0.76123046875
],
[
0.08251953125,
0.75830078125,
0.2034912109375,
0.75830078125,
0.2034912109375,
0.76806640625,
0.08251953125,
0.76806640625
],
[
0.5048828125,
0.76220703125,
0.890625,
0.76220703125,
0.890625,
0.77587890625,
0.5048828125,
0.77587890625
],
[
0.2059326171875,
0.78271484375,
0.339111328125,
0.7841796875,
0.339111328125,
0.798828125,
0.2059326171875,
0.79736328125
],
[
0.50341796875,
0.77783203125,
0.890625,
0.77783203125,
0.890625,
0.79150390625,
0.50341796875,
0.79150390625
],
[
0.44775390625,
0.787109375,
0.468017578125,
0.787109375,
0.468017578125,
0.79736328125,
0.44775390625,
0.79736328125
],
[
0.50341796875,
0.79345703125,
0.890625,
0.79345703125,
0.890625,
0.806640625,
0.50341796875,
0.806640625
],
[
0.080078125,
0.80810546875,
0.4697265625,
0.80810546875,
0.4697265625,
0.82568359375,
0.080078125,
0.82568359375
],
[
0.50341796875,
0.80810546875,
0.890625,
0.80810546875,
0.890625,
0.8212890625,
0.50341796875,
0.8212890625
],
[
0.08251953125,
0.82763671875,
0.1422119140625,
0.82763671875,
0.1422119140625,
0.8388671875,
0.08251953125,
0.8388671875
],
[
0.50341796875,
0.82373046875,
0.8896484375,
0.82373046875,
0.8896484375,
0.8369140625,
0.50341796875,
0.8369140625
],
[
0.1004638671875,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.1004638671875,
0.85498046875
],
[
0.5048828125,
0.8388671875,
0.58642578125,
0.8388671875,
0.58642578125,
0.85107421875,
0.5048828125,
0.85107421875
],
[
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,
0.87255859375,
0.46728515625,
0.87255859375,
0.46728515625,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.080078125,
0.88623046875,
0.4697265625,
0.88720703125,
0.4697265625,
0.9013671875,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"3.Proposed Method",
"同时，我们的DeFeeNet可以检测到刚刚发生的预测偏",
"差，如r-1轮的预测（i。e。，Yr-1=9r-1,1:T）现在",
"在大多数情况下，3D人体结构由基于人体关节坐",
"可以观察到它的地面实况。因此，通过编码和合并此偏",
"标的纯姿势参数或SMPL模型[23]表示，该模型描",
"差Dr-1，可以自然地改善当前预测Y=9r,1:T",
"绘了具有姿势和形状的3D人体网格。根据任务需",
"求，我们遵循[26,31]丢弃形状项，只预测姿势项。当",
"前，人体运动预测【4,27,34]旨在根据观察到的N-顿",
"3.3.Network Architecture",
"X=[1,2，·,N]序列，预测具有T顿的未来序列",
"Y=[91,92，,9T]，其中每个顿（i。e。，pose)分别",
"我们基于MLP和GRU构建了两个版本的De-",
"用9iERK和aiERK表示。我们定义了以N顿观测",
"FeeNet（请参见图3），这两个版本都很简单但很",
"有效，因为我们不鼓励DeFeeNet在原始预测模型上增",
"为单位预测T顿的过程。然而，在这个孤立单元内的",
"研究完全忽略了“预测错误”元素对连续运动预测的重",
"加过多的计算负担",
"要性，并且无法实现偏差感知和反馈以提高预测精度",
"受MLP-Mixer结构[36]在计算机视觉中的成功及",
"其在下游人体运动建模任务【4,6,12】中的变体的启发，",
"3.1. Reformulation from Consecutive Perspective",
"我们提出了基于MLP的DeFeeNet版本，但它更轻巧",
"但足够有效。与在[4,6,12,36]中使用多个MLP混合",
"我们介绍我们重新制定的人体运动预测任务。连续",
"预测涉及多轮单元，其中N+T顿沿整个序列半重叠。",
"器块相比，我们的设计丢弃了块堆叠，并且仅包含一个",
"假设一个长度为L的序列S=[S1,S2,·.，SL]足够长，",
"时间混合MLP，一个空间混合MLP和跳过连接（请",
"我们将第一轮定义为R1=S1:N+T，接下来的每一轮",
"参见左图3）。两个MLP混音器模块均由两个完全连",
"都是T顿，i。e。，第轮R,=S1+(r-1)T:N+T+(r-1)T",
"接的层组成，GELU激活函数13]和层归一化操作[3]",
"其中r≥1，在本文中我们设置N≥T。因此，如果",
"对于基于GRU的版本，如右图3所示，我们采用",
"我们定义所有运动回合R=R1,R2，，Rr，..并重",
"了一层GRU，在时间维度上附加了一个全连接层，在",
"写 R =[X,Y,] =[r,1,.,ar,N,Yr,1,.,Yr,T] , 则",
"空间维度上附加了一个用于维度对齐的层，以插人到",
"可以推断出r,N-T:N=r-1,1:T，并且在回合R-1",
"原始管道中进行以下计算",
"中产生的偏差表示为Dr-1=fa(ar,N-T:N,9r-1,1:T)",
"",
"具体来说，首先向原始预测器馈送R1中的观察到",
"以用于提高回合Rr的预测精度。当r=1时，Do为",
"的片段Xi=a1,1:N，并在时间步长 T处获得其第",
"空，因为不能使用偏差。基本上，我们将连续运动预测",
"轮预测Yi=91,1:T，其中DeFeeNet未激活且不涉及",
"的任务视为逐轮实现R",
"偏差信息。然而，在接下来的儿轮中，X，变得可观察",
"3.2. Deviation-Aware Prediction",
"到，(yr-1,1:T）和(Cr,N-T:N）之间的偏差出现，并被",
"馈送到我们的DeFeeNet中，以捕获其时间和空间信",
"为了详细表达3.1节中提到的偏差Dr-1、=",
"息，并将其添加到主管道中，从而实现相邻预测单元之",
"fa(r,N-T:N，9r-1,1:T），我们从【4,12]中汲取灵感，",
"间的偏差反馈",
"该【4,12]基于两个相邻姿势的关节位移来设计其预测",
"损失函数（i。e。，速度）而不是关节位置。我们不是直接",
"3.4. Two-Round Training & Multi-Round Testing",
"在ar,N-T:N和9r-1,1:T之间进行减法，而是首先将它",
"们的速度计算为(aT.N-TN)，(9r-1.1:T）ERK(-",
"我们的DeFeeNet的目标是通过对预测偏差的连续",
"，然后获得：",
"感知和反馈来提高原始预测精度，以及在持续的时间",
"段内保持这种优先级。由于半重叠结构为我们提供了",
"Dr-1= fa(cr,N-T:N,9r-1,1:T)",
"(1",
"一种学习预测偏差的自然方法，因此我们相应地介绍",
"= U(Tr,N-T:N) - U(9r-1,1T) e RKX(T-1).",
"",
"了我们的重采样方法和损失函数",
"重采样。与现有的将“先观察后预测”单元作为样本的",
"这样，与专注于学习孤立单元中观测X和目标Y",
"预测模型相比，我们将两轮单元作为一个训练样本，多",
"之间的预测函数Y=。(X）的现有工作相比，我们的",
"轮单元作为一个测试样本，i。e。、Rtrain=[R1,R2]和",
"偏差感知连续预测旨在为每个预测回合找到一个包含",
"Rtest=[R1,R2，.·，Rmax(r)]。对于训练阶段，我们只",
"偏差反馈的函数：",
"是让DeFeeNet学习根据其感知的先前相邻偏差来产",
"Y, = p(Xr, Dr-1),",
"生有用的信息，因此只有两轮预测（包含一个重叠区",
"(2)",
"域）就足够了。在测试阶段，需要具有多轮但灵活的预",
"其中:RK×N× RK×(T-1)→RKxT。注意 Do=0",
"测单元的样本，因为我们的最终目标是评估DeFeeNet",
"当r=1",
"是否能够提前对未知持续时间的连续观察和预测产生",
"在图2中，我们提供了DeFeeNet如何嵌入到现有",
"稳定的改进",
"预测框架中的概述，并有助于在相对较长的周期内逐",
"损失函数。在当前作品[4,8,12,27,34]之后，我们的损",
"轮实现偏差感知的运动预测。在r轮，我们将当前一轮",
"失基于平均每关节位置误差（MPJPE）【15]，但不同",
"观测X,=r,1:N送人主管道(i。e。，现有的预测器)。",
"的是，我们计算两轮预测的损失。根据3.1节中的表示"
] |
{
"arxiv_id": "2304.04496",
"language": "zh",
"page_idx": 2
}
|
||
2304.04496-zh-3
|
2304.04496-zh-3.jpg
|
<FigureHere>
Figure 2. (a) Existing baselines that simplify human motion prediction as a one-off process of predicting a $T$ -pose sequence based on $N$ -pose observation (i.e, the isolated prediction unit). (b) DeFeeNet allows for human motion prediction round by round with a deviation-aware manner. Round $r$ is marked in mint green while $r-1$ in orange. During the prediction of round $r$ , our DeFeeNet detects the deviation between current observation and previous round of prediction, to realize deviation feedback across adjacent units.
<table>
<tbody>
<tr>
<th rowspan="2">frame num.</th>
<th colspan="4">walking</th>
<th colspan="5">eating</th>
<th colspan="4">smoking</th>
<th> </th>
<th colspan="3">discussiom</th>
<th colspan="4">directions</th>
</tr>
<tr>
<th>2</th>
<th>4</th>
<th>8</th>
<th>10</th>
<th>2</th>
<th>4</th>
<th>4</th>
<th>8</th>
<th>10</th>
<th>2</th>
<th>4</th>
<th>8</th>
<th>10</th>
<th>0 2</th>
<th>2</th>
<th>8</th>
<th>10</th>
<th>2</th>
<th>4</th>
<th>8</th>
<th>10</th>
</tr>
<tr>
<td>LTD- GCN [27</td>
<td>12.3</td>
<td>23.0</td>
<td>39.8</td>
<td>46.1</td>
<td>$\overline{8.4}$</td>
<td>$\overline{16.9}$</td>
<td>.6.9 3</td>
<td>33.2</td>
<td>40.7</td>
<td>$\overline{8.0}$</td>
<td>16.2</td>
<td>31.9</td>
<td>38.9 9 1</td>
<td>.9 12</td>
<td>2.5 2</td>
<td>$\overline{58.5}$ 1</td>
<td>$\overline{71.7}$ 5</td>
<td>9.0</td>
<td>$\overline{19.9}$</td>
<td>43.4</td>
<td>$\overline{53.7}$</td>
</tr>
<tr>
<td>LTD- DeFee( MLP) - rl</td>
<td>$\overline{11.7}$</td>
<td>$\overline{21.2}$</td>
<td>36.4</td>
<td>43.5</td>
<td>$\overline{8.1}$</td>
<td>16.2</td>
<td>.6.2 3</td>
<td>31.9</td>
<td>39.8</td>
<td>$\overline{7.7}$</td>
<td>15.3</td>
<td>29.9</td>
<td>36.8 9</td>
<td>.8 12</td>
<td>2.7 2</td>
<td>$\overline{56.4}$ 3</td>
<td>69.4 4</td>
<td>9.1</td>
<td>$\overline{19.3}$</td>
<td>41.7</td>
<td>$\overline{52.0}$</td>
</tr>
<tr>
<td>LTD-DeFee(MLP)-r2</td>
<td>10.2</td>
<td>19.9</td>
<td>35.5</td>
<td>42.5</td>
<td>6.7</td>
<td>14.8</td>
<td>.4.8 3</td>
<td>30.7</td>
<td>38.3</td>
<td>6.5</td>
<td>14.0</td>
<td>28.8</td>
<td>35.6 8</td>
<td>10 .6</td>
<td>0.0 2</td>
<td>54.5 $\frac{3}{2}$</td>
<td>68.1 5</td>
<td>6.9</td>
<td>16.7</td>
<td>39.6</td>
<td>50.2</td>
</tr>
<tr>
<td>LTD- DeFee( GRU) - r1</td>
<td>11.8</td>
<td>22.2</td>
<td>36.8</td>
<td>44.0</td>
<td>8.2</td>
<td>16.9</td>
<td>.6.9 3</td>
<td>32.2</td>
<td>40.2</td>
<td>7.8</td>
<td>15.8</td>
<td>30.2</td>
<td>37.5 2</td>
<td>.5 12</td>
<td>2.5 2</td>
<td>57.1</td>
<td>70.4 l</td>
<td>8.8</td>
<td>19.5</td>
<td>42.4</td>
<td>52.8</td>
</tr>
<tr>
<td rowspan="2">LTD- DeFee( GRU) - r2</td>
<td>10.4</td>
<td>20.0</td>
<td>34.7</td>
<td>42.2</td>
<td>7.0</td>
<td>15.2</td>
<td>.5.2 3</td>
<td>31.4</td>
<td>38.4</td>
<td>6.8</td>
<td>14.5</td>
<td>29.0</td>
<td>35.8 0</td>
<td>.8 11</td>
<td>1.. 1 2 $\frac{1}{2}=\frac{1}{2}$</td>
<td>55.8 1</td>
<td>3 68.2</td>
<td>7.0</td>
<td>17.0</td>
<td>40.0</td>
<td>50.9</td>
</tr>
<tr>
<td colspan="4">greeting</td>
<td> </td>
<td>phc</td>
<td>phoni 1</td>
<td>ning</td>
<td> </td>
<td> </td>
<td colspan="5">posing purchas</td>
<td>chases</td>
<td>S</td>
<td> </td>
<td>sit</td>
<td>ting</td>
<td> </td>
</tr>
<tr>
<td>frame s mum.</td>
<td>2</td>
<td>4</td>
<td>8</td>
<td>10</td>
<td>2</td>
<td>4</td>
<td>4</td>
<td>8</td>
<td>10</td>
<td>2</td>
<td>4</td>
<td>8</td>
<td>10</td>
<td>0 2</td>
<td>2</td>
<td>8</td>
<td>10</td>
<td>2</td>
<td>4</td>
<td>8</td>
<td>10</td>
</tr>
<tr>
<td>STS- GCN [34]</td>
<td>$\overline{18.7}$</td>
<td>$\overline{34.9}$</td>
<td>$\overline{71.6}$</td>
<td>86.4</td>
<td>$\overline{13.7}$</td>
<td>22.4</td>
<td>22.4 4</td>
<td>43.6</td>
<td>53.8</td>
<td>16.4</td>
<td>$\overline{30.4}$</td>
<td>67.6</td>
<td>84.7 6</td>
<td>.7 19</td>
<td>9.1 3</td>
<td>70.2 3</td>
<td>83.1 2</td>
<td>15.2</td>
<td>$\overline{25.1}$</td>
<td>49.8</td>
<td>60.8</td>
</tr>
<tr>
<td>STS- DeFee( MLP) - rl</td>
<td>18.4</td>
<td>$\overline{34.2}$</td>
<td>$\overline{71.2}$</td>
<td>85.4</td>
<td>$\overline{13.5}$</td>
<td>21.9</td>
<td>21.9 4</td>
<td>42.1</td>
<td>52.5</td>
<td>16.6</td>
<td>$\overline{30.4}$</td>
<td>66.2</td>
<td>83.2 2</td>
<td>2 19</td>
<td>9.5 3</td>
<td>69.4 5</td>
<td>82.3 4</td>
<td>15.0</td>
<td>24.8</td>
<td>49.6</td>
<td>$\overline{59.8}$</td>
</tr>
<tr>
<td>STS- DeFee( MLP) - r2</td>
<td>16.4</td>
<td>32.8</td>
<td>68.8</td>
<td>82.5</td>
<td>11.4</td>
<td>19.8</td>
<td>.9.8 4</td>
<td>40.8</td>
<td>49.7</td>
<td>14.8</td>
<td>28.3</td>
<td>64.8</td>
<td>80.5 8</td>
<td>.5 16</td>
<td>6.6 3</td>
<td>$67.\epsilon$ $j$</td>
<td>80.6 6</td>
<td>14.0</td>
<td>23.3</td>
<td>47.5</td>
<td>58.7</td>
</tr>
<tr>
<td>STS- DeFee( GRU) - rl</td>
<td>18.5</td>
<td>34.5</td>
<td>71.4</td>
<td>85.8</td>
<td>13.5</td>
<td>22.1</td>
<td>22.1 4</td>
<td>42.5</td>
<td>53.0</td>
<td>16.2</td>
<td>30.2</td>
<td>66.7</td>
<td>83.5 7</td>
<td>.5 18</td>
<td>8.9 3</td>
<td>69.4 2</td>
<td>82.4 4</td>
<td>15.1</td>
<td>24.9</td>
<td>49.5</td>
<td>59.6</td>
</tr>
<tr>
<td rowspan="2">STS-DeFee(GRU)-r2</td>
<td>16.8</td>
<td>33.0</td>
<td>68.5</td>
<td>83.2</td>
<td>11.6</td>
<td>19.9</td>
<td>.9.9 4</td>
<td>41.0</td>
<td>50.1</td>
<td>14.7</td>
<td>28.3</td>
<td>65.0</td>
<td>81.1 0</td>
<td>16 1</td>
<td>6.8 3</td>
<td>67.0</td>
<td>80.3 9</td>
<td>14.2</td>
<td>23.6</td>
<td>47.7</td>
<td>58.7</td>
</tr>
<tr>
<td colspan="7">taking] sittingdown</td>
<td colspan="2">$\operatorname{takingphoto}$</td>
<td colspan="4"> </td>
<td> </td>
<td colspan="3"> </td>
<td colspan="4">walkingtogether</td>
</tr>
<tr>
<td>frame ${\mathrm{mum.}}$</td>
<td>2</td>
<td>4</td>
<td>8</td>
<td>10</td>
<td>2</td>
<td>4</td>
<td>4</td>
<td>8</td>
<td>10</td>
<td>2</td>
<td>4</td>
<td>8</td>
<td>10</td>
<td>0 2</td>
<td>2</td>
<td>8</td>
<td>10</td>
<td>2</td>
<td>4</td>
<td>8</td>
<td>10</td>
</tr>
<tr>
<td>MotionMixer</td>
<td>12.0</td>
<td>$\overline{31.4}$</td>
<td>64.4</td>
<td>74.5</td>
<td>9.0</td>
<td>$\overline{18.9}$</td>
<td>.8.9 4</td>
<td>41.0</td>
<td>51.6</td>
<td>10.2</td>
<td>$\overline{21.1}$</td>
<td>45.2</td>
<td>56.4 2</td>
<td>4 20</td>
<td>$\overline{0.5}$ 4</td>
<td>$\overline{75.\epsilon}$ 3</td>
<td>$\overline{87.8}$ 5</td>
<td>10.5</td>
<td>20.6</td>
<td>38.7</td>
<td>$\overline{43.5}$</td>
</tr>
<tr>
<td>MotMix- DeFee( MLP) - rl</td>
<td>12.0</td>
<td>30.6</td>
<td>63.7</td>
<td>72.6</td>
<td>9.1</td>
<td>$\overline{18.2}$</td>
<td>.8.2 3</td>
<td>39.7</td>
<td>50.4</td>
<td>10.4</td>
<td>20.5</td>
<td>$\overline{44.1}$</td>
<td>$\overline{55.6}$ 1</td>
<td>.6 19</td>
<td>9.5 4</td>
<td>$\overline{74.4}$ 1</td>
<td>86.9 4</td>
<td>10.2</td>
<td>$\overline{19.8}$</td>
<td>37.8</td>
<td>42.7</td>
</tr>
<tr>
<td>MotMix- DeFee( MLP) - r2</td>
<td>9.8</td>
<td>29.1</td>
<td>61.8</td>
<td>70.2</td>
<td>7.8</td>
<td>16.9</td>
<td>.6.9 3</td>
<td>37.1</td>
<td>47.7</td>
<td>9.3</td>
<td>19.5</td>
<td>42.0</td>
<td>53.4 0</td>
<td>17 4</td>
<td>7.3 4</td>
<td>72.8 $j$</td>
<td>84.4 3</td>
<td>8.3</td>
<td>19.1</td>
<td>35.9</td>
<td>41.5 $\frac{1}{2}=\frac{1}{2}$</td>
</tr>
<tr>
<td>MotMix-DeFee(GRU)-rl</td>
<td>11.7</td>
<td>30.8</td>
<td>63.6</td>
<td>73.7</td>
<td>8.8</td>
<td>18.5</td>
<td>.8.5 4</td>
<td>40.2</td>
<td>50.8</td>
<td>10.1</td>
<td>20.8</td>
<td>44.5</td>
<td>55.4 5</td>
<td>4 19</td>
<td>9.5 4</td>
<td>74.4</td>
<td>87.1 4</td>
<td>10.4</td>
<td>19.8</td>
<td>38.0</td>
<td>43.1</td>
</tr>
<tr>
<td>MotMix- DeFee( GRU) - r2</td>
<td>10.1</td>
<td>29.4</td>
<td>62.0</td>
<td>70.8</td>
<td>7.8</td>
<td>16.9</td>
<td>6.9 3</td>
<td>38.3</td>
<td>47.9</td>
<td>9.6</td>
<td>19.8</td>
<td>42.3</td>
<td>53.6 3</td>
<td>17 .6</td>
<td>7.6 4</td>
<td>72.7</td>
<td>84.9 7</td>
<td>8.8</td>
<td>19.0</td>
<td>36.1</td>
<td>41.8</td>
</tr>
</tbody>
</table>
Table 1. Top to bottom: Prediction errors produced by the original baselines (isolated), baselines with MLP-based DeFeeNet inserted (round 1 and 2), and with GRU-based DeFeeNet inserted (round 1 and 2). Values in bold indicate lower errors and prove the deviation feedback is valid. For each baseline, we present the performance on 5 actions out of 15 in Human3.6M. We present prediction errors at frame 2,4,8,and 10.
# 4. Experiments
法，我们将$r(r=1,2)$回合的损失表示为
$$\mathcal{L}_r=\dfrac{1}{T}\sum\limits_{t=1}^T\|\hat{y}_{r,t}-y_{r,t}\|_2,$$
(3)
在本节中，我们将在连续人体运动预测的任务上评估我们的 DeFeeNet。两个数据集 (Human3。6M[15] 和较新的，具有挑战性的 BABEL [32])进行了预处理以满足任务要求，我们在插人 DeFeeNet 的情况下在三个不同的运动预测基线上进行了实验。我们还提供消融研究以供进一步分析
,其中$\hat{y}_{r,t}$表示$r$回合$t$帧的预测姿态$,y_r,t$表示相应的地面实况。请注意，为了公平比较，当我们的 DeFeeNet 在不同的基准模型上添加并在不同的数据集上进行评估时，方程 3 在具体细节上会有一些轻微的变化 (可能参考补充材料)
然后可以将总损失计算为：
## $4. 1. \textbf{ Datasets}$
(4)
$$\mathcal{L}_{\mathrm{total}}=\mathcal{L}_{1}+\mathcal{L}_{2}.$$
Human3。6M [15] 是人类运动建模中使用最广泛的数据集之一，其中包括七个演员在 15 个类别下执行动作， 例如步行，进食，吸烟和讨论，下采样率为 25fps 以生成序列数据。每个人体骨骼姿势由 32 个关节表示，我们遵循现有的作品[4,8,12,27,34,40] ,对其中 22 个关节进行评估。目前，他们的目标是制作 10 或 25 帧(i。e。, 400 毫秒短期或 1000 毫秒长期预测) 基于 10 帧的
除了尽可能准确地监督两轮预测外，$\mathcal{L}_1$还可以进一步确保包含 DeFeeNet 不会过多地阻碍主要管道 (因为已经训练了模型参数以适应 DeFeeNet\_的存在，但是在第 1 轮中没有偏差可以保持有效的偏差),而
|
[
[
0.61376953125,
0.0997314453125,
0.76708984375,
0.0997314453125,
0.76708984375,
0.10986328125,
0.61376953125,
0.10986328125
],
[
0.1422119140625,
0.10797119140625,
0.1953125,
0.10797119140625,
0.1953125,
0.11553955078125,
0.1422119140625,
0.11553955078125
],
[
0.5146484375,
0.111083984375,
0.560546875,
0.111083984375,
0.560546875,
0.1256103515625,
0.5146484375,
0.1256103515625
],
[
0.63134765625,
0.10919189453125,
0.75244140625,
0.10919189453125,
0.75244140625,
0.11614990234375,
0.63134765625,
0.11614990234375
],
[
0.814453125,
0.11553955078125,
0.85107421875,
0.11553955078125,
0.85107421875,
0.1256103515625,
0.814453125,
0.1256103515625
],
[
0.1396484375,
0.11932373046875,
0.279296875,
0.121826171875,
0.279296875,
0.1395263671875,
0.1396484375,
0.136962890625
],
[
0.44677734375,
0.12249755859375,
0.476318359375,
0.12249755859375,
0.476318359375,
0.1351318359375,
0.44677734375,
0.1351318359375
],
[
0.6650390625,
0.12371826171875,
0.69677734375,
0.12371826171875,
0.69677734375,
0.1319580078125,
0.6650390625,
0.1319580078125
],
[
0.5390625,
0.1256103515625,
0.580078125,
0.1256103515625,
0.580078125,
0.14013671875,
0.5390625,
0.14013671875
],
[
0.3896484375,
0.11236572265625,
0.44677734375,
0.121826171875,
0.44287109375,
0.136962890625,
0.3857421875,
0.1275634765625
],
[
0.39453125,
0.1300048828125,
0.41259765625,
0.1300048828125,
0.41259765625,
0.1331787109375,
0.39453125,
0.1331787109375
],
[
0.60888671875,
0.1287841796875,
0.65185546875,
0.1337890625,
0.64794921875,
0.15283203125,
0.6044921875,
0.147705078125
],
[
0.6640625,
0.1337890625,
0.79736328125,
0.1331787109375,
0.79736328125,
0.14453125,
0.6640625,
0.1458740234375
],
[
0.7998046875,
0.1307373046875,
0.83837890625,
0.136962890625,
0.8349609375,
0.147705078125,
0.79736328125,
0.141357421875
],
[
0.373291015625,
0.14013671875,
0.418212890625,
0.14013671875,
0.418212890625,
0.1514892578125,
0.373291015625,
0.1514892578125
],
[
0.63232421875,
0.14013671875,
0.65283203125,
0.14013671875,
0.65283203125,
0.147705078125,
0.63232421875,
0.147705078125
],
[
0.83251953125,
0.1351318359375,
0.85693359375,
0.1351318359375,
0.85693359375,
0.147705078125,
0.83251953125,
0.147705078125
],
[
0.60888671875,
0.1483154296875,
0.6435546875,
0.15087890625,
0.64208984375,
0.1634521484375,
0.60693359375,
0.1610107421875
],
[
0.81884765625,
0.1514892578125,
0.85302734375,
0.1514892578125,
0.85302734375,
0.1641845703125,
0.81884765625,
0.1641845703125
],
[
0.293212890625,
0.1578369140625,
0.334228515625,
0.1578369140625,
0.334228515625,
0.173583984375,
0.293212890625,
0.173583984375
],
[
0.221435546875,
0.1590576171875,
0.27294921875,
0.1590576171875,
0.27294921875,
0.169189453125,
0.221435546875,
0.169189453125
],
[
0.0931396484375,
0.154052734375,
0.2320556640625,
0.1558837890625,
0.2320556640625,
0.1749267578125,
0.0931396484375,
0.17236328125
],
[
0.418212890625,
0.1622314453125,
0.53759765625,
0.1654052734375,
0.53759765625,
0.18310546875,
0.418212890625,
0.179931640625
],
[
0.54736328125,
0.1654052734375,
0.56787109375,
0.1654052734375,
0.56787109375,
0.1685791015625,
0.54736328125,
0.1685791015625
],
[
0.11517333984375,
0.1717529296875,
0.1422119140625,
0.173583984375,
0.1405029296875,
0.186279296875,
0.11358642578125,
0.18310546875
],
[
0.29736328125,
0.1749267578125,
0.330078125,
0.1749267578125,
0.330078125,
0.1900634765625,
0.29736328125,
0.1900634765625
],
[
0.61376953125,
0.1881103515625,
0.62353515625,
0.1881103515625,
0.62353515625,
0.1912841796875,
0.61376953125,
0.1912841796875
],
[
0.07843017578125,
0.208984375,
0.89208984375,
0.210205078125,
0.89208984375,
0.2235107421875,
0.07843017578125,
0.222900390625
],
[
0.08087158203125,
0.2247314453125,
0.890625,
0.2247314453125,
0.890625,
0.23486328125,
0.08087158203125,
0.23486328125
],
[
0.080078125,
0.2374267578125,
0.89208984375,
0.2374267578125,
0.89208984375,
0.250732421875,
0.080078125,
0.250732421875
],
[
0.08087158203125,
0.25244140625,
0.890625,
0.25244140625,
0.890625,
0.26318359375,
0.08087158203125,
0.26318359375
],
[
0.07843017578125,
0.26513671875,
0.351318359375,
0.26513671875,
0.351318359375,
0.279052734375,
0.07843017578125,
0.279052734375
],
[
0.310546875,
0.29345703125,
0.35302734375,
0.29541015625,
0.351318359375,
0.306884765625,
0.3095703125,
0.304931640625
],
[
0.423095703125,
0.29345703125,
0.461669921875,
0.29541015625,
0.4599609375,
0.30810546875,
0.42236328125,
0.30615234375
],
[
0.53125,
0.29541015625,
0.576171875,
0.29541015625,
0.576171875,
0.306884765625,
0.53125,
0.306884765625
],
[
0.63623046875,
0.294189453125,
0.69140625,
0.294189453125,
0.69140625,
0.30615234375,
0.63623046875,
0.30615234375
],
[
0.74658203125,
0.29541015625,
0.80126953125,
0.29541015625,
0.80126953125,
0.30615234375,
0.74658203125,
0.30615234375
],
[
0.1805419921875,
0.3037109375,
0.240234375,
0.30615234375,
0.240234375,
0.317626953125,
0.1805419921875,
0.31494140625
],
[
0.334228515625,
0.306884765625,
0.351318359375,
0.306884765625,
0.351318359375,
0.31494140625,
0.334228515625,
0.31494140625
],
[
0.365234375,
0.306884765625,
0.38232421875,
0.306884765625,
0.38232421875,
0.31494140625,
0.365234375,
0.31494140625
],
[
0.44287109375,
0.306884765625,
0.461669921875,
0.306884765625,
0.461669921875,
0.31494140625,
0.44287109375,
0.31494140625
],
[
0.47802734375,
0.306884765625,
0.49267578125,
0.306884765625,
0.49267578125,
0.31494140625,
0.47802734375,
0.31494140625
],
[
0.53369140625,
0.306884765625,
0.54736328125,
0.306884765625,
0.54736328125,
0.31494140625,
0.53369140625,
0.31494140625
],
[
0.55615234375,
0.306884765625,
0.57421875,
0.306884765625,
0.57421875,
0.31494140625,
0.55615234375,
0.31494140625
],
[
0.58642578125,
0.306884765625,
0.6044921875,
0.306884765625,
0.6044921875,
0.31494140625,
0.58642578125,
0.31494140625
],
[
0.6435546875,
0.306884765625,
0.66015625,
0.306884765625,
0.66015625,
0.31494140625,
0.6435546875,
0.31494140625
],
[
0.6650390625,
0.306884765625,
0.685546875,
0.306884765625,
0.685546875,
0.31494140625,
0.6650390625,
0.31494140625
],
[
0.69775390625,
0.306884765625,
0.71337890625,
0.306884765625,
0.71337890625,
0.31494140625,
0.69775390625,
0.31494140625
],
[
0.75244140625,
0.306884765625,
0.76708984375,
0.306884765625,
0.76708984375,
0.314453125,
0.75244140625,
0.314453125
],
[
0.77392578125,
0.306884765625,
0.79736328125,
0.306884765625,
0.79736328125,
0.31494140625,
0.77392578125,
0.31494140625
],
[
0.8056640625,
0.306884765625,
0.82421875,
0.306884765625,
0.82421875,
0.31494140625,
0.8056640625,
0.31494140625
],
[
0.420654296875,
0.30810546875,
0.4345703125,
0.3037109375,
0.438720703125,
0.3125,
0.4248046875,
0.31640625
],
[
0.313720703125,
0.30810546875,
0.325927734375,
0.30810546875,
0.325927734375,
0.314453125,
0.313720703125,
0.314453125
],
[
0.396240234375,
0.31005859375,
0.40283203125,
0.31005859375,
0.40283203125,
0.313232421875,
0.396240234375,
0.313232421875
],
[
0.7255859375,
0.309326171875,
0.736328125,
0.306884765625,
0.73779296875,
0.3125,
0.72607421875,
0.314453125
],
[
0.1732177734375,
0.317626953125,
0.2467041015625,
0.317626953125,
0.2467041015625,
0.328857421875,
0.1732177734375,
0.328857421875
],
[
0.27783203125,
0.318115234375,
0.496826171875,
0.318115234375,
0.496826171875,
0.328857421875,
0.27783203125,
0.328857421875
],
[
0.498291015625,
0.318115234375,
0.82861328125,
0.318115234375,
0.82861328125,
0.328857421875,
0.498291015625,
0.328857421875
],
[
0.1568603515625,
0.330078125,
0.265625,
0.330078125,
0.265625,
0.34033203125,
0.1568603515625,
0.34033203125
],
[
0.27685546875,
0.330810546875,
0.498291015625,
0.330810546875,
0.498291015625,
0.341552734375,
0.27685546875,
0.341552734375
],
[
0.496826171875,
0.330810546875,
0.52685546875,
0.330810546875,
0.52685546875,
0.339111328125,
0.496826171875,
0.339111328125
],
[
0.52294921875,
0.32763671875,
0.830078125,
0.32763671875,
0.830078125,
0.341552734375,
0.52294921875,
0.341552734375
],
[
0.1502685546875,
0.34228515625,
0.26708984375,
0.34228515625,
0.26708984375,
0.352783203125,
0.1502685546875,
0.352783203125
],
[
0.273681640625,
0.34033203125,
0.386474609375,
0.341552734375,
0.386474609375,
0.352783203125,
0.273681640625,
0.3515625
],
[
0.381591796875,
0.34228515625,
0.420654296875,
0.34228515625,
0.420654296875,
0.352783203125,
0.381591796875,
0.352783203125
],
[
0.4150390625,
0.34228515625,
0.498291015625,
0.34228515625,
0.498291015625,
0.352783203125,
0.4150390625,
0.352783203125
],
[
0.52392578125,
0.34228515625,
0.60888671875,
0.34228515625,
0.60888671875,
0.352783203125,
0.52392578125,
0.352783203125
],
[
0.6044921875,
0.34228515625,
0.830078125,
0.34228515625,
0.830078125,
0.352783203125,
0.6044921875,
0.352783203125
],
[
0.74609375,
0.352783203125,
0.82666015625,
0.352783203125,
0.82666015625,
0.363037109375,
0.74609375,
0.363037109375
],
[
0.498291015625,
0.343505859375,
0.52783203125,
0.343505859375,
0.52783203125,
0.3515625,
0.498291015625,
0.3515625
],
[
0.156005859375,
0.354248046875,
0.2646484375,
0.354248046875,
0.2646484375,
0.3642578125,
0.156005859375,
0.3642578125
],
[
0.27783203125,
0.354248046875,
0.386474609375,
0.354248046875,
0.386474609375,
0.3642578125,
0.27783203125,
0.3642578125
],
[
0.4150390625,
0.354248046875,
0.495849609375,
0.354248046875,
0.495849609375,
0.3642578125,
0.4150390625,
0.3642578125
],
[
0.52685546875,
0.354248046875,
0.60888671875,
0.354248046875,
0.60888671875,
0.3642578125,
0.52685546875,
0.3642578125
],
[
0.609375,
0.354248046875,
0.7197265625,
0.354248046875,
0.7197265625,
0.3642578125,
0.609375,
0.3642578125
],
[
0.50341796875,
0.354736328125,
0.521484375,
0.354736328125,
0.521484375,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.393798828125,
0.354736328125,
0.408447265625,
0.354736328125,
0.408447265625,
0.363037109375,
0.393798828125,
0.363037109375
],
[
0.7255859375,
0.354736328125,
0.74267578125,
0.354736328125,
0.74267578125,
0.3623046875,
0.7255859375,
0.3623046875
],
[
0.74267578125,
0.356689453125,
0.7490234375,
0.356689453125,
0.7490234375,
0.35986328125,
0.74267578125,
0.35986328125
],
[
0.1544189453125,
0.365478515625,
0.26708984375,
0.365478515625,
0.26708984375,
0.375732421875,
0.1544189453125,
0.375732421875
],
[
0.27685546875,
0.365478515625,
0.386474609375,
0.365478515625,
0.386474609375,
0.375732421875,
0.27685546875,
0.375732421875
],
[
0.38818359375,
0.363037109375,
0.5009765625,
0.363037109375,
0.5009765625,
0.376953125,
0.38818359375,
0.376953125
],
[
0.495849609375,
0.365478515625,
0.830078125,
0.365478515625,
0.830078125,
0.375732421875,
0.495849609375,
0.375732421875
],
[
0.3095703125,
0.380126953125,
0.3544921875,
0.380126953125,
0.3544921875,
0.391357421875,
0.3095703125,
0.391357421875
],
[
0.420654296875,
0.380126953125,
0.465576171875,
0.380126953125,
0.465576171875,
0.391357421875,
0.420654296875,
0.391357421875
],
[
0.63818359375,
0.380126953125,
0.689453125,
0.380126953125,
0.689453125,
0.391357421875,
0.63818359375,
0.391357421875
],
[
0.75634765625,
0.380126953125,
0.7939453125,
0.380126953125,
0.7939453125,
0.391357421875,
0.75634765625,
0.391357421875
],
[
0.66845703125,
0.38818359375,
0.685546875,
0.393310546875,
0.68115234375,
0.401611328125,
0.6640625,
0.396484375
],
[
0.53515625,
0.380615234375,
0.57177734375,
0.380615234375,
0.57177734375,
0.392578125,
0.53515625,
0.392578125
],
[
0.1805419921875,
0.39013671875,
0.238525390625,
0.39013671875,
0.238525390625,
0.401611328125,
0.1805419921875,
0.401611328125
],
[
0.366943359375,
0.391357421875,
0.381591796875,
0.391357421875,
0.381591796875,
0.399658203125,
0.366943359375,
0.399658203125
],
[
0.476318359375,
0.391357421875,
0.49267578125,
0.391357421875,
0.49267578125,
0.399658203125,
0.476318359375,
0.399658203125
],
[
0.58837890625,
0.391357421875,
0.6044921875,
0.391357421875,
0.6044921875,
0.40087890625,
0.58837890625,
0.40087890625
],
[
0.69775390625,
0.391357421875,
0.71337890625,
0.391357421875,
0.71337890625,
0.399658203125,
0.69775390625,
0.399658203125
],
[
0.80615234375,
0.391357421875,
0.82275390625,
0.391357421875,
0.82275390625,
0.40087890625,
0.80615234375,
0.40087890625
],
[
0.44287109375,
0.392578125,
0.461669921875,
0.392578125,
0.461669921875,
0.399658203125,
0.44287109375,
0.399658203125
],
[
0.313720703125,
0.393310546875,
0.32275390625,
0.393310546875,
0.32275390625,
0.3984375,
0.313720703125,
0.3984375
],
[
0.334228515625,
0.392578125,
0.351318359375,
0.392578125,
0.351318359375,
0.399658203125,
0.334228515625,
0.399658203125
],
[
0.39453125,
0.393310546875,
0.40283203125,
0.393310546875,
0.40283203125,
0.397705078125,
0.39453125,
0.397705078125
],
[
0.423095703125,
0.393310546875,
0.4345703125,
0.393310546875,
0.4345703125,
0.397705078125,
0.423095703125,
0.397705078125
],
[
0.64453125,
0.393310546875,
0.65673828125,
0.393310546875,
0.65673828125,
0.399658203125,
0.64453125,
0.399658203125
],
[
0.7587890625,
0.393310546875,
0.765625,
0.393310546875,
0.765625,
0.397705078125,
0.7587890625,
0.397705078125
],
[
0.78173828125,
0.393310546875,
0.79150390625,
0.393310546875,
0.79150390625,
0.3984375,
0.78173828125,
0.3984375
],
[
0.55615234375,
0.395751953125,
0.56787109375,
0.389404296875,
0.57275390625,
0.39453125,
0.56201171875,
0.40087890625
],
[
0.53369140625,
0.39453125,
0.54150390625,
0.389404296875,
0.54736328125,
0.393310546875,
0.5400390625,
0.399658203125
],
[
0.1732177734375,
0.40283203125,
0.2442626953125,
0.40283203125,
0.2442626953125,
0.41357421875,
0.1732177734375,
0.41357421875
],
[
0.27685546875,
0.40283203125,
0.63623046875,
0.40283203125,
0.63623046875,
0.41357421875,
0.27685546875,
0.41357421875
],
[
0.63134765625,
0.40283203125,
0.830078125,
0.40283203125,
0.830078125,
0.41357421875,
0.63134765625,
0.41357421875
],
[
0.1544189453125,
0.4140625,
0.2646484375,
0.4140625,
0.2646484375,
0.4248046875,
0.1544189453125,
0.4248046875
],
[
0.27685546875,
0.4140625,
0.830078125,
0.4140625,
0.830078125,
0.4248046875,
0.27685546875,
0.4248046875
],
[
0.1544189453125,
0.4267578125,
0.265625,
0.4267578125,
0.265625,
0.4375,
0.1544189453125,
0.4375
],
[
0.27783203125,
0.4267578125,
0.82861328125,
0.4267578125,
0.82861328125,
0.4375,
0.27783203125,
0.4375
],
[
0.1568603515625,
0.438232421875,
0.2646484375,
0.438232421875,
0.2646484375,
0.448974609375,
0.1568603515625,
0.448974609375
],
[
0.279296875,
0.438232421875,
0.60693359375,
0.438232421875,
0.60693359375,
0.448974609375,
0.279296875,
0.448974609375
],
[
0.61279296875,
0.439453125,
0.7236328125,
0.439453125,
0.7236328125,
0.4462890625,
0.61279296875,
0.4462890625
],
[
0.7197265625,
0.438232421875,
0.82861328125,
0.438232421875,
0.82861328125,
0.448974609375,
0.7197265625,
0.448974609375
],
[
0.156005859375,
0.449462890625,
0.265625,
0.449462890625,
0.265625,
0.460205078125,
0.156005859375,
0.460205078125
],
[
0.27685546875,
0.449462890625,
0.82861328125,
0.449462890625,
0.82861328125,
0.460205078125,
0.27685546875,
0.460205078125
],
[
0.2998046875,
0.464111328125,
0.36279296875,
0.464111328125,
0.36279296875,
0.47607421875,
0.2998046875,
0.47607421875
],
[
0.410888671875,
0.464111328125,
0.473876953125,
0.464111328125,
0.473876953125,
0.47607421875,
0.410888671875,
0.47607421875
],
[
0.53369140625,
0.464111328125,
0.57421875,
0.464111328125,
0.57421875,
0.47607421875,
0.53369140625,
0.47607421875
],
[
0.63232421875,
0.462158203125,
0.69287109375,
0.464111328125,
0.69287109375,
0.47607421875,
0.63232421875,
0.473388671875
],
[
0.73388671875,
0.464111328125,
0.81787109375,
0.464111328125,
0.81787109375,
0.474853515625,
0.73388671875,
0.474853515625
],
[
0.1805419921875,
0.474853515625,
0.238525390625,
0.474853515625,
0.238525390625,
0.486083984375,
0.1805419921875,
0.486083984375
],
[
0.4794921875,
0.474853515625,
0.49267578125,
0.477783203125,
0.488525390625,
0.486083984375,
0.4755859375,
0.481689453125
],
[
0.5888671875,
0.47607421875,
0.60302734375,
0.47607421875,
0.60302734375,
0.484130859375,
0.5888671875,
0.484130859375
],
[
0.69775390625,
0.4765625,
0.71337890625,
0.4765625,
0.71337890625,
0.484130859375,
0.69775390625,
0.484130859375
],
[
0.8037109375,
0.4765625,
0.82177734375,
0.4765625,
0.82177734375,
0.48486328125,
0.8037109375,
0.48486328125
],
[
0.4248046875,
0.477783203125,
0.4345703125,
0.477783203125,
0.4345703125,
0.48291015625,
0.4248046875,
0.48291015625
],
[
0.36767578125,
0.477783203125,
0.3798828125,
0.477783203125,
0.3798828125,
0.484130859375,
0.36767578125,
0.484130859375
],
[
0.451904296875,
0.477783203125,
0.4599609375,
0.477783203125,
0.4599609375,
0.48291015625,
0.451904296875,
0.48291015625
],
[
0.5361328125,
0.477783203125,
0.54345703125,
0.477783203125,
0.54345703125,
0.48291015625,
0.5361328125,
0.48291015625
],
[
0.56396484375,
0.477783203125,
0.5703125,
0.477783203125,
0.5703125,
0.48291015625,
0.56396484375,
0.48291015625
],
[
0.73046875,
0.477783203125,
0.73779296875,
0.479736328125,
0.734375,
0.484130859375,
0.72802734375,
0.48095703125
],
[
0.78369140625,
0.477783203125,
0.7958984375,
0.477783203125,
0.7958984375,
0.484130859375,
0.78369140625,
0.484130859375
],
[
0.646484375,
0.479248046875,
0.65625,
0.479248046875,
0.65625,
0.48291015625,
0.646484375,
0.48291015625
],
[
0.316162109375,
0.479248046875,
0.324462890625,
0.479248046875,
0.324462890625,
0.48291015625,
0.316162109375,
0.48291015625
],
[
0.67236328125,
0.479248046875,
0.68310546875,
0.479248046875,
0.68310546875,
0.48291015625,
0.67236328125,
0.48291015625
],
[
0.75732421875,
0.479248046875,
0.76708984375,
0.479248046875,
0.76708984375,
0.48291015625,
0.75732421875,
0.48291015625
],
[
0.1683349609375,
0.486083984375,
0.250732421875,
0.4873046875,
0.250732421875,
0.498779296875,
0.1683349609375,
0.49755859375
],
[
0.338134765625,
0.48095703125,
0.3447265625,
0.47607421875,
0.3505859375,
0.479736328125,
0.343017578125,
0.48486328125
],
[
0.27783203125,
0.4873046875,
0.386474609375,
0.4873046875,
0.386474609375,
0.49755859375,
0.27783203125,
0.49755859375
],
[
0.3896484375,
0.4892578125,
0.408447265625,
0.4892578125,
0.408447265625,
0.496826171875,
0.3896484375,
0.496826171875
],
[
0.418212890625,
0.4892578125,
0.7138671875,
0.4892578125,
0.7138671875,
0.496826171875,
0.418212890625,
0.496826171875
],
[
0.71142578125,
0.486083984375,
0.830078125,
0.486083984375,
0.830078125,
0.5,
0.71142578125,
0.5
],
[
0.143798828125,
0.496826171875,
0.82861328125,
0.49755859375,
0.82861328125,
0.51123046875,
0.143798828125,
0.51025390625
],
[
0.410888671875,
0.4912109375,
0.418212890625,
0.4912109375,
0.418212890625,
0.494384765625,
0.410888671875,
0.494384765625
],
[
0.146240234375,
0.5087890625,
0.830078125,
0.5087890625,
0.830078125,
0.52294921875,
0.146240234375,
0.52294921875
],
[
0.1478271484375,
0.52294921875,
0.826171875,
0.52294921875,
0.826171875,
0.53271484375,
0.1478271484375,
0.53271484375
],
[
0.1478271484375,
0.53271484375,
0.82666015625,
0.53271484375,
0.82666015625,
0.54345703125,
0.1478271484375,
0.54345703125
],
[
0.08087158203125,
0.56005859375,
0.890625,
0.56005859375,
0.890625,
0.57080078125,
0.08087158203125,
0.57080078125
],
[
0.07843017578125,
0.57275390625,
0.89208984375,
0.57275390625,
0.89208984375,
0.5859375,
0.07843017578125,
0.5859375
],
[
0.080078125,
0.5859375,
0.890625,
0.5859375,
0.890625,
0.599609375,
0.080078125,
0.599609375
],
[
0.08087158203125,
0.599609375,
0.4345703125,
0.599609375,
0.4345703125,
0.61279296875,
0.08087158203125,
0.61279296875
],
[
0.08087158203125,
0.6416015625,
0.390625,
0.6416015625,
0.390625,
0.65478515625,
0.08087158203125,
0.65478515625
],
[
0.50341796875,
0.6416015625,
0.53515625,
0.6416015625,
0.53515625,
0.65283203125,
0.50341796875,
0.65283203125
],
[
0.52783203125,
0.6416015625,
0.640625,
0.6416015625,
0.640625,
0.65478515625,
0.52783203125,
0.65478515625
],
[
0.1666259765625,
0.6611328125,
0.373291015625,
0.6494140625,
0.378173828125,
0.7021484375,
0.17236328125,
0.71337890625
],
[
0.2467041015625,
0.66552734375,
0.261474609375,
0.66552734375,
0.261474609375,
0.67529296875,
0.2467041015625,
0.67529296875
],
[
0.52392578125,
0.66748046875,
0.890625,
0.66748046875,
0.890625,
0.68115234375,
0.52392578125,
0.68115234375
],
[
0.44677734375,
0.6796875,
0.468017578125,
0.6796875,
0.468017578125,
0.6923828125,
0.44677734375,
0.6923828125
],
[
0.50341796875,
0.681640625,
0.8896484375,
0.681640625,
0.8896484375,
0.69580078125,
0.50341796875,
0.69580078125
],
[
0.24267578125,
0.69580078125,
0.268798828125,
0.69580078125,
0.268798828125,
0.70703125,
0.24267578125,
0.70703125
],
[
0.50244140625,
0.69580078125,
0.8896484375,
0.69580078125,
0.8896484375,
0.71044921875,
0.50244140625,
0.71044921875
],
[
0.08660888671875,
0.71533203125,
0.468017578125,
0.71533203125,
0.468017578125,
0.72998046875,
0.08660888671875,
0.72998046875
],
[
0.50341796875,
0.7119140625,
0.8896484375,
0.7119140625,
0.8896484375,
0.72607421875,
0.50341796875,
0.72607421875
],
[
0.080078125,
0.72998046875,
0.468017578125,
0.72900390625,
0.468017578125,
0.74365234375,
0.080078125,
0.744140625
],
[
0.5048828125,
0.72802734375,
0.890625,
0.72802734375,
0.890625,
0.7412109375,
0.5048828125,
0.7412109375
],
[
0.08087158203125,
0.74560546875,
0.468017578125,
0.74560546875,
0.468017578125,
0.7587890625,
0.08087158203125,
0.7587890625
],
[
0.50244140625,
0.7412109375,
0.66650390625,
0.7412109375,
0.66650390625,
0.7548828125,
0.50244140625,
0.7548828125
],
[
0.08087158203125,
0.76123046875,
0.46728515625,
0.76123046875,
0.46728515625,
0.7744140625,
0.08087158203125,
0.7744140625
],
[
0.50244140625,
0.76953125,
0.609375,
0.77099609375,
0.609375,
0.78515625,
0.50244140625,
0.7841796875
],
[
0.08087158203125,
0.77587890625,
0.205078125,
0.77587890625,
0.205078125,
0.7890625,
0.08087158203125,
0.7890625
],
[
0.1004638671875,
0.79150390625,
0.289306640625,
0.79150390625,
0.289306640625,
0.8046875,
0.1004638671875,
0.8046875
],
[
0.50244140625,
0.79443359375,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.50244140625,
0.8095703125
],
[
0.2099609375,
0.814453125,
0.336669921875,
0.81640625,
0.336669921875,
0.833984375,
0.2099609375,
0.83203125
],
[
0.50341796875,
0.81005859375,
0.88818359375,
0.81005859375,
0.88818359375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.4453125,
0.8173828125,
0.468017578125,
0.8173828125,
0.468017578125,
0.830078125,
0.4453125,
0.830078125
],
[
0.5048828125,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.5048828125,
0.8388671875
],
[
0.08087158203125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.080078125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.87060546875,
0.080078125,
0.869140625
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.080078125,
0.88525390625,
0.40283203125,
0.88623046875,
0.40283203125,
0.90087890625,
0.080078125,
0.8994140625
],
[
0.50244140625,
0.88623046875,
0.890625,
0.88623046875,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"(b) Deviation-Aware Prediction Framework",
"[a]ExistingOne",
"prediadiong",
"forConsecutiveMotionPrediction",
"r.17",
"unit",
"",
"Baseline",
"fdeviation",
"永意型",
"",
"#",
"EncoderDecoder",
"业",
"round (r-1):",
"",
"",
"XI:N",
"ym1:T",
"贺材",
"Decoder",
"Enede",
"店",
"NT",
"XEN",
"YI:T",
"",
"Figure 2. (a) Existing baselines that simplify human motion prediction as a one-off process of predicting a T -pose sequence",
"based on N-poseobservation (i.e.,theisolated prediction unit).(b)DeFeeNetallowsfor human motion prediction round",
"by round with a deviation-aware manner. Round r is marked in mint green while r -1 in orange. During the prediction",
"of round r our DeFeeNet detects the deviation between current observation and previous round of prediction, to realize",
"deviation feedback across adjacent units.",
"walking",
"eating",
"smoking",
"discussion",
"directions",
"frame num.",
"8",
"10",
"8",
"10",
"4",
"8",
"10",
"4",
"8",
"10",
"4",
"8",
"10",
"4",
"4.",
"·",
"5",
"LTD-GCN27",
"12.323.039.846.18.416.933.240.7",
"8.016.231.938.912.527.458.571.79.019.943.453.7",
"LTD-DeFee(MLP)-r1",
"11.721.236.443.58.116.231.9 39.8",
"77.",
"15.329.9 36.812.726.856.469.49.119.341.752.0",
"LTD-DeFee(MLP)-r2",
"10.219.935.542.5",
"6.7",
"14.830.738.3",
"14.028.835.6",
"10.023.854.568.16.916.739.650.2",
"19.542.4 52.8",
"6.5",
"LTD-DeFee(GRU)-rl",
"11.822.236.8 44.0",
"16.932.240.2",
"15.830.237.5",
"12.526.757.170.4",
"7.8",
"8.2",
"8.8",
"",
"LTD-DeFee(GRU)-r2)",
"10.420.034.7:42.2",
"7.015.2 31.4 38.4",
"6.814.529.035.811.125.455.868.27.017.040.050.9",
"greeting",
"phoning",
"purchases",
"sitting",
"8",
"posing",
"frame num",
"10",
"10",
"10",
"10",
"10",
"8",
"4",
"8",
"5",
"",
"4",
"4",
"8",
"0",
"",
"STS-GCNB4",
"18.734.971.686.413.722.443.653.816.430.467.684.719.1",
"35.870.283.115.225.149.860.8",
"STS-DeFee(MLP)-rl",
"18.434.271.285.413.521.942.152.516.630.466.283.219.535.569.482.315.024.849.659.8",
"STS-DeFee(MLP)-r2",
"16.432.868.882.511.419.840.849.714.828.364.880.516.632.567.680.614.023.347.558.7",
"STS-DeFee(GRU)-rl",
"18.534.571.485.813.522.142.553.016.230.266.783.5",
"18.935.269.4.82.4",
"15.124.949.5 59.6",
"STS-DeFee(GRU)-r2",
"16.833.068.583.211.619.941.050.114.728.365.081.116.832.767.980.314.223.647.758.7",
"sittingdown",
"takingphoto",
"waiting",
"walkingdog",
"walkingtogether",
"frame num",
"10",
"10",
"10",
"10",
"4",
"10",
"8",
"4",
"8",
"",
"8",
"",
"a",
"X",
"4",
"MotionMixer[4",
"",
"12.0 31.4 64.4 74.5",
"9.0",
"18.9/41.051.610.221.145.256.420.542.875.687.8",
"810.5 20.6 38.7 43.5",
"MotMix-DeFee(MLP)-r112.030.663.772.69.118.239.750.410.420.544.155.619.542.474.4 86.910.219.837.842.7",
"",
"MotMix-DeFee(MLP)-r29.8 29.1 61.8 70.2 7.8 16.9 37.1 47.7 9.3 19.5 42.0 53.4 17.3 40.9 72.8 84.48.3 19.1 35.9 41.5",
"MotMix-DeFee(GRU)-r111.730.8 63.6 73.78.818.5 40.2 50.810.1 20.8 44.5 55.419.5 42.1 74.4 87.110.4 19.8 38.0 43.1",
"MotMix-DeFee(GRU)-r210.1 29.4 62.0 70.87.816.9 38.347.99.619.8 42.353.617.6 41.172.784.98.819.036.141.8",
"Table1.Top tobottom:Prediction errors produced by the original baselines (isolated),baselines with MLP-based DeFeeNet",
"inserted (round 1 and 2), and with GRU-based DeFeeNet inserted (round 1 and 2). Values in bold indicate lower errors and",
"prove the deviation feedback is valid. For each baseline, we present the performance on 5 actions out of 15 in Human3.6M.",
"We present prediction errors at frame 2, 4, 8, and 10.",
"法，我们将r（r=1,2）回合的损失表示为：",
"4.",
"Experiments",
"C - o - l",
"T",
"在本节中，我们将在连续人体运动预测的任务上评",
"(3)",
"估我们的DeFeeNet。两个数据集（Human3。6M「15)",
"t=1",
"和较新的，具有挑战性的BABEL[321）进行了预处理",
"其中9r.t表示r回合t顿的预测姿态，yr.t表示相应的",
"以满足任务要求，我们在插人DeFeeNet的情况下在三",
"地面实况。请注意，为了公平比较，当我们的DeFeeNet",
"个不同的运动预测基线上进行了实验。我们还提供消",
"在不同的基准模型上添加并在不同的数据集上进行评",
"融研究以供进一步分析",
"估时，方程3在具体细节上会有一些轻微的变化（可",
"4.1. Datasets",
"能参考补充材料）",
"然后可以将总损失计算为：",
"Human3。6M「15]是人类运动建模中使用最广泛的数",
"Ctotal = Li + L2.",
"据集之一，其中包括七个演员在15个类别下执行动作",
"(4)",
"例如步行，进食，吸烟和讨论，下采样率为25fps以生",
"除了尽可能准确地监督两轮预测外，L1还可以进一步",
"成序列数据。每个人体骨骼姿势由32个关节表示，我",
"确保包含DeFeeNet不会过多地阻碍主要管道（因为已",
"们遵循现有的作品4,8,12,27,34,40]，对其中22个关",
"经训练了模型参数以适应DeFeeNet的存在，但是在",
"节进行评估。目前，他们的目标是制作10或25帧（1。",
"第1轮中没有偏差可以保持有效的偏差），而",
"e。，400毫秒短期或1000毫秒长期预测）基于10顿的"
] |
{
"arxiv_id": "2304.04496",
"language": "zh",
"page_idx": 3
}
|
||
2304.04496-zh-4
|
2304.04496-zh-4.jpg
|
<FigureHere>
Figure 3. Architectures of MLP-based and GRU-based De-
FeeNet.
观察，在孤立的“单元”内
对于连续预测任务 ,我们使用我们的重采样方法(3.4节)进行进一步的预处理。为了公平比较，我们还将每个预测回合设置为观察到 10 个帧，并预测以下 10 个帧。训练样本由两个相邻的回合组成，其中第二轮是10帧，与第一轮相比。测试样本首先安排 2 轮，以证明相邻单元之间的偏差反馈是有效的，然后我们安排10 轮评估 DeFeeNet 在相对较长的范围内的稳定性
<table>
<tbody>
<tr>
<th rowspan="4">frame mum. isolated</th>
<th colspan="4">LTD- GCN [27]</th>
<th colspan="6">LTD- GCN [27] avg 38.64</th>
<th colspan="6">STS-GCN [3/1] avg 41.16</th>
<th colspan="6">MotionMixer $\begin{bmatrix}4\\1\end{bmatrix}$ avg 35.52</th>
</tr>
<tr>
<th rowspan="3">2 12.69</th>
<th rowspan="2">26.06 -DeFee</th>
<th rowspan="3">52.28 $(\mathrm{MLP}^{\circ}$</th>
<th>10</th>
</tr>
<tr>
<th>63.53</th>
</tr>
<tr>
<th rowspan="1">$.\Pi_{\mathrm{P}}$</th>
<th> </th>
<th>$- D($MLP)</th>
<th>$-D(GRU)$</th>
<th>$- D($MLP)</th>
<th>$-D(GRU)$</th>
<th>$- D($MLP)</th>
<th>$-D(GRU$</th>
</tr>
<tr>
<td>rl</td>
<td>12.59</td>
<td>25.16</td>
<td>50.02</td>
<td>61.47</td>
<td>37.31</td>
<td>37.65</td>
<td>40.76</td>
<td>40.89</td>
<td>34.12</td>
<td>34.35</td>
</tr>
<tr>
<td>$r2$</td>
<td>10.30</td>
<td>22.60</td>
<td>48.17</td>
<td>59.79</td>
<td>35.21</td>
<td>35.80</td>
<td>38.15</td>
<td>38.77</td>
<td>32.38</td>
<td>32.69</td>
</tr>
<tr>
<td>$r3$</td>
<td>10.43</td>
<td>22.74</td>
<td>48.44</td>
<td>60.14</td>
<td>35.44</td>
<td>35.67</td>
<td>38.33</td>
<td>38.67</td>
<td>32.16</td>
<td>32.74</td>
</tr>
<tr>
<td>r4</td>
<td>10.39</td>
<td>22.69</td>
<td>48.31</td>
<td>59.95</td>
<td>35.33</td>
<td>35.66</td>
<td>38.46</td>
<td>38.72</td>
<td>32.05</td>
<td>32.75</td>
</tr>
<tr>
<td>$r5$</td>
<td>10.38</td>
<td>22.68</td>
<td>48.27</td>
<td>59.91</td>
<td>35.31</td>
<td>35.91</td>
<td>38.45</td>
<td>38.90</td>
<td>32.17</td>
<td>32.59</td>
</tr>
<tr>
<td>r6</td>
<td>10.38</td>
<td>22.67</td>
<td>48.25</td>
<td>59.87</td>
<td>35.30</td>
<td>36.10</td>
<td>38.37</td>
<td>39.05</td>
<td>32.34</td>
<td>32.69</td>
</tr>
<tr>
<td>$r7$</td>
<td>10.37</td>
<td>22.66</td>
<td>48.24</td>
<td>59.88</td>
<td>35.29</td>
<td>35.88</td>
<td>38.41</td>
<td>38.95</td>
<td>32.33</td>
<td>32.66</td>
</tr>
<tr>
<td>r8</td>
<td>10.39</td>
<td>22.68</td>
<td>48.27</td>
<td>59.89</td>
<td>35.31</td>
<td>35.66</td>
<td>38.45</td>
<td>38.89</td>
<td>32.47</td>
<td>32.77</td>
</tr>
<tr>
<td>$r9$</td>
<td>10.40</td>
<td>22.71</td>
<td>48.30</td>
<td>59.92</td>
<td>35.33</td>
<td>35.75</td>
<td>38.38</td>
<td>38.93</td>
<td>32.41</td>
<td>32.71</td>
</tr>
<tr>
<td>rl0</td>
<td>10.41</td>
<td>22.72</td>
<td>48.33</td>
<td>59.96</td>
<td>35.35</td>
<td>35.70</td>
<td>38.49</td>
<td>38.78</td>
<td>32.44</td>
<td>32.75</td>
</tr>
</tbody>
</table>
Table 2. Consecutive 10 rounds of prediction errors on Human3.6M. Values in bold from round 2 to 10 indicate that deviation-aware prediction stably yields improved performance. Detailed errors of LTD-DeFee(MLP) at frame 2, 4,8,10 are on the left. Average errors (avg) of these four testpoints produced by other baselines with DeFeeNet (abbreviated as D ) are on the right.
BABEL[32] 是最近提出的数据集，其中包含有关操作的语言标签。与其他用于人体运动预测的数据集不同，
其中每个序列中包含的动作属于同一类别，BABEL 允许具有多种动作的序列。目前，最适合我们任务的预处理方法是 [26] , 它导出两个数据集文件：(a)20 个动作标签下的单动作序列 (如站立、行走、步进和伸展); (b) 包含这 20 个动作的两类动作及其中间转换的序列， 两者都是 30fps, 只有姿势参数用于评估，而 SMPL 形状参数被丢弃 (可以参考 XMPX2 中提供的链接)
由于尚未在 BABEL 上进行确定性运动预测的实验， 因此我们首先从上述序列文件中剪切孤立的单位样本以训练三个基线，观察到的和预测的帧均设置为 10。然后，类似于 Human3。6M,我们用 2 轮重新排列我们的训练样本，同时用 2 轮进行有效性验证和 5 轮进行稳定性验证。此外，我们删除了少于 200 个序列样本的动作，并留下了 11 个动作类别的样本进行评估
评估度量。在[4,8,12,24,34,40]之后，我们通过 3D 人体关节坐标上的平均每关节位置误差(MPJPE)来评估我们的网络，该网络计算了预测与相应地面实况之间不同关节差异的平均$L_{2}$ -范数
# $4. 2. \textbf{ Baselines}$
我们的 DeFeeNet 对其基本预测模型是不可知的，并且可以添加到许多现有网络上，以实现连续的人体运动预测。在这里，我们为实验介绍三个基线
LTD-GCN [27] 是一个基于 GCN 的模型，由 12 个GCN 块组成，描述了不同接头尺寸与残余连接的空间相关性。离散余弦变换(DCT)用于提取时间特征
STS-GCN [34] 设计了允许时空串扰的单图结构，该结构将图分解为可分离的空间和时间矩阵，以充分学习关节和时间相互作用
MotionMixer [4]提出了一种有效的运动预测模型，该模型仅通过顺序混合空间和时间依赖性来采用 MLP, 并通过增强的挤压和激励块
# 4.3. Implementation Details
我们在单个 NVIDIARTX2080Ti 上使用 Adam 优化器 [19] 在 Pytorch [29] 框架下进行实验。在基于 MLP 的版本中，时间混合和空间混合的隐藏层大小设置为与基本管道的相应尺寸相同。在基于 GRU 的版本中， 无论基本模型如何，隐藏大小都设置为 256,而完全连接的图层用于尺寸对齐。详细信息在补充材料中提供。对于 LTD-GCN、STS-GCN 和 MotionMixer,我们使用 DeFeeNet 联合训练每个基线，学习率分别为0.0005 、0.01和0.01 。我们在 Human3 上训练 50 个纪元。BABEL 上的 6M 和 100
# 4.4. Results on Human3.6M
表 1 显示了在 Human3 上插人和不插人 DeFeeNet 的模型的预测误差。6M 数据集。在没有 DeFeeNet 的情况下，当前的基线观察和预测孤立的短单元内的人体运动 (i。e。, 10 帧的长度为 400 毫秒)。但是，我们联合训练了两轮预测过程，其中在第 1 轮(r1)中生成的预测偏差由我们插人的 DeFeeNet 检测和学习，以便协助原始管道以先前的偏差作为反馈进行预测，以提
|
[
[
0.1544189453125,
0.09344482421875,
0.2254638671875,
0.09344482421875,
0.2254638671875,
0.10479736328125,
0.1544189453125,
0.10479736328125
],
[
0.325927734375,
0.09344482421875,
0.39453125,
0.09344482421875,
0.39453125,
0.10479736328125,
0.325927734375,
0.10479736328125
],
[
0.50341796875,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50341796875,
0.1060791015625
],
[
0.375732421875,
0.10797119140625,
0.42724609375,
0.111083984375,
0.426513671875,
0.12249755859375,
0.374267578125,
0.11932373046875
],
[
0.50341796875,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.2132568359375,
0.10986328125,
0.221435546875,
0.11614990234375,
0.214111328125,
0.121826171875,
0.2059326171875,
0.11553955078125
],
[
0.11767578125,
0.10986328125,
0.17724609375,
0.11236572265625,
0.17724609375,
0.12371826171875,
0.11767578125,
0.121826171875
],
[
0.2197265625,
0.11297607421875,
0.25732421875,
0.11297607421875,
0.25732421875,
0.12060546875,
0.2197265625,
0.12060546875
],
[
0.28759765625,
0.10986328125,
0.3505859375,
0.11236572265625,
0.3505859375,
0.12371826171875,
0.28759765625,
0.121826171875
],
[
0.5048828125,
0.12249755859375,
0.890625,
0.12249755859375,
0.890625,
0.1363525390625,
0.5048828125,
0.1363525390625
],
[
0.5048828125,
0.1383056640625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.310546875,
0.1558837890625,
0.326904296875,
0.1558837890625,
0.326904296875,
0.18115234375,
0.310546875,
0.18115234375
],
[
0.50341796875,
0.15283203125,
0.88818359375,
0.15283203125,
0.88818359375,
0.1666259765625,
0.50341796875,
0.1666259765625
],
[
0.1446533203125,
0.1590576171875,
0.15283203125,
0.1590576171875,
0.15283203125,
0.1685791015625,
0.1446533203125,
0.1685791015625
],
[
0.2254638671875,
0.1572265625,
0.2344970703125,
0.1572265625,
0.2344970703125,
0.1685791015625,
0.2254638671875,
0.1685791015625
],
[
0.39208984375,
0.1590576171875,
0.408447265625,
0.1590576171875,
0.408447265625,
0.18115234375,
0.39208984375,
0.18115234375
],
[
0.1422119140625,
0.1685791015625,
0.156005859375,
0.1685791015625,
0.156005859375,
0.179931640625,
0.1422119140625,
0.179931640625
],
[
0.22216796875,
0.1685791015625,
0.236083984375,
0.1685791015625,
0.236083984375,
0.179931640625,
0.22216796875,
0.179931640625
],
[
0.5048828125,
0.1673583984375,
0.89208984375,
0.1673583984375,
0.89208984375,
0.18115234375,
0.5048828125,
0.18115234375
],
[
0.181396484375,
0.1817626953125,
0.1968994140625,
0.1817626953125,
0.1968994140625,
0.1900634765625,
0.181396484375,
0.1900634765625
],
[
0.3505859375,
0.18310546875,
0.370849609375,
0.18310546875,
0.370849609375,
0.1900634765625,
0.3505859375,
0.1900634765625
],
[
0.50341796875,
0.18310546875,
0.85107421875,
0.18310546875,
0.85107421875,
0.1962890625,
0.50341796875,
0.1962890625
],
[
0.181396484375,
0.1912841796875,
0.191162109375,
0.189453125,
0.193603515625,
0.1956787109375,
0.1846923828125,
0.1988525390625
],
[
0.1502685546875,
0.2008056640625,
0.226318359375,
0.2008056640625,
0.226318359375,
0.2108154296875,
0.1502685546875,
0.2108154296875
],
[
0.52392578125,
0.1976318359375,
0.88818359375,
0.1976318359375,
0.88818359375,
0.212158203125,
0.52392578125,
0.212158203125
],
[
0.340576171875,
0.21337890625,
0.375732421875,
0.21337890625,
0.375732421875,
0.2259521484375,
0.340576171875,
0.2259521484375
],
[
0.5048828125,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.5048828125,
0.2266845703125
],
[
0.1666259765625,
0.2197265625,
0.2115478515625,
0.2197265625,
0.2115478515625,
0.2342529296875,
0.1666259765625,
0.2342529296875
],
[
0.181396484375,
0.21533203125,
0.188720703125,
0.2108154296875,
0.193603515625,
0.216552734375,
0.1871337890625,
0.2203369140625
],
[
0.5048828125,
0.2279052734375,
0.890625,
0.2279052734375,
0.890625,
0.2412109375,
0.5048828125,
0.2412109375
],
[
0.35693359375,
0.236083984375,
0.361083984375,
0.236083984375,
0.361083984375,
0.2442626953125,
0.35693359375,
0.2442626953125
],
[
0.146240234375,
0.24365234375,
0.2320556640625,
0.24365234375,
0.2320556640625,
0.255126953125,
0.146240234375,
0.255126953125
],
[
0.330078125,
0.2474365234375,
0.3896484375,
0.2474365234375,
0.3896484375,
0.2587890625,
0.330078125,
0.2587890625
],
[
0.50341796875,
0.24365234375,
0.89208984375,
0.24365234375,
0.89208984375,
0.2568359375,
0.50341796875,
0.2568359375
],
[
0.1748046875,
0.251953125,
0.2034912109375,
0.25390625,
0.2017822265625,
0.264404296875,
0.1732177734375,
0.261962890625
],
[
0.3544921875,
0.260009765625,
0.363525390625,
0.260009765625,
0.363525390625,
0.271484375,
0.3544921875,
0.271484375
],
[
0.5048828125,
0.2587890625,
0.890625,
0.2587890625,
0.890625,
0.272705078125,
0.5048828125,
0.272705078125
],
[
0.3203125,
0.272705078125,
0.39794921875,
0.272705078125,
0.39794921875,
0.2841796875,
0.3203125,
0.2841796875
],
[
0.188720703125,
0.264404296875,
0.193603515625,
0.267578125,
0.188720703125,
0.2734375,
0.1829833984375,
0.270751953125
],
[
0.156005859375,
0.2734375,
0.221435546875,
0.2734375,
0.221435546875,
0.285400390625,
0.156005859375,
0.285400390625
],
[
0.50341796875,
0.2734375,
0.890625,
0.2734375,
0.890625,
0.287353515625,
0.50341796875,
0.287353515625
],
[
0.178955078125,
0.2841796875,
0.2017822265625,
0.2841796875,
0.2017822265625,
0.31005859375,
0.178955078125,
0.31005859375
],
[
0.35546875,
0.28857421875,
0.366943359375,
0.290283203125,
0.36279296875,
0.298583984375,
0.351318359375,
0.296630859375
],
[
0.50244140625,
0.28857421875,
0.87353515625,
0.28857421875,
0.87353515625,
0.3017578125,
0.50244140625,
0.3017578125
],
[
0.338134765625,
0.2998046875,
0.378173828125,
0.2998046875,
0.378173828125,
0.311279296875,
0.338134765625,
0.311279296875
],
[
0.1829833984375,
0.30615234375,
0.1953125,
0.30615234375,
0.1953125,
0.3193359375,
0.1829833984375,
0.3193359375
],
[
0.50341796875,
0.3037109375,
0.890625,
0.3037109375,
0.890625,
0.317626953125,
0.50341796875,
0.317626953125
],
[
0.158447265625,
0.3193359375,
0.2181396484375,
0.3193359375,
0.2181396484375,
0.330810546875,
0.158447265625,
0.330810546875
],
[
0.35693359375,
0.3193359375,
0.361083984375,
0.3193359375,
0.361083984375,
0.324462890625,
0.35693359375,
0.324462890625
],
[
0.5048828125,
0.318115234375,
0.890625,
0.318115234375,
0.890625,
0.33203125,
0.5048828125,
0.33203125
],
[
0.181396484375,
0.330810546875,
0.193603515625,
0.330810546875,
0.193603515625,
0.34033203125,
0.181396484375,
0.34033203125
],
[
0.330078125,
0.32763671875,
0.386474609375,
0.32763671875,
0.386474609375,
0.339111328125,
0.330078125,
0.339111328125
],
[
0.1502685546875,
0.339111328125,
0.2254638671875,
0.34033203125,
0.2254638671875,
0.3515625,
0.1502685546875,
0.35107421875
],
[
0.35302734375,
0.339111328125,
0.365234375,
0.339111328125,
0.365234375,
0.35107421875,
0.35302734375,
0.35107421875
],
[
0.5048828125,
0.333984375,
0.8896484375,
0.333984375,
0.8896484375,
0.347900390625,
0.5048828125,
0.347900390625
],
[
0.318603515625,
0.35107421875,
0.398681640625,
0.35107421875,
0.398681640625,
0.365478515625,
0.318603515625,
0.365478515625
],
[
0.50341796875,
0.349853515625,
0.736328125,
0.349853515625,
0.736328125,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.3544921875,
0.36865234375,
0.365234375,
0.36865234375,
0.365234375,
0.37744140625,
0.3544921875,
0.37744140625
],
[
0.176513671875,
0.3515625,
0.1993408203125,
0.3515625,
0.1993408203125,
0.35986328125,
0.176513671875,
0.35986328125
],
[
0.156005859375,
0.37255859375,
0.221435546875,
0.37255859375,
0.221435546875,
0.3837890625,
0.156005859375,
0.3837890625
],
[
0.336669921875,
0.37744140625,
0.381591796875,
0.37744140625,
0.381591796875,
0.392578125,
0.336669921875,
0.392578125
],
[
0.50244140625,
0.37255859375,
0.615234375,
0.373779296875,
0.615234375,
0.38818359375,
0.50244140625,
0.386962890625
],
[
0.178955078125,
0.385009765625,
0.1968994140625,
0.385009765625,
0.1968994140625,
0.40478515625,
0.178955078125,
0.40478515625
],
[
0.2197265625,
0.3837890625,
0.30810546875,
0.3837890625,
0.30810546875,
0.39453125,
0.2197265625,
0.39453125
],
[
0.2197265625,
0.395751953125,
0.330810546875,
0.395751953125,
0.330810546875,
0.406005859375,
0.2197265625,
0.406005859375
],
[
0.51953125,
0.39453125,
0.89208984375,
0.395751953125,
0.89208984375,
0.41357421875,
0.51953125,
0.412353515625
],
[
0.2197265625,
0.40478515625,
0.324462890625,
0.40478515625,
0.324462890625,
0.417236328125,
0.2197265625,
0.417236328125
],
[
0.35546875,
0.40283203125,
0.363525390625,
0.40283203125,
0.363525390625,
0.412353515625,
0.35546875,
0.412353515625
],
[
0.50341796875,
0.410888671875,
0.89208984375,
0.412353515625,
0.89208984375,
0.4267578125,
0.50341796875,
0.425537109375
],
[
0.50341796875,
0.427978515625,
0.81396484375,
0.427978515625,
0.81396484375,
0.44140625,
0.50341796875,
0.44140625
],
[
0.1829833984375,
0.406005859375,
0.193603515625,
0.406005859375,
0.193603515625,
0.415283203125,
0.1829833984375,
0.415283203125
],
[
0.07843017578125,
0.43115234375,
0.46728515625,
0.429931640625,
0.46728515625,
0.4443359375,
0.07843017578125,
0.44580078125
],
[
0.08087158203125,
0.4462890625,
0.13232421875,
0.4462890625,
0.13232421875,
0.458251953125,
0.08087158203125,
0.458251953125
],
[
0.50244140625,
0.44140625,
0.890625,
0.442626953125,
0.890625,
0.45703125,
0.50244140625,
0.455810546875
],
[
0.50341796875,
0.458251953125,
0.890625,
0.458251953125,
0.890625,
0.4716796875,
0.50341796875,
0.4716796875
],
[
0.50341796875,
0.473388671875,
0.85546875,
0.473388671875,
0.85546875,
0.4873046875,
0.50341796875,
0.4873046875
],
[
0.080078125,
0.484130859375,
0.27783203125,
0.48291015625,
0.27783203125,
0.49755859375,
0.080078125,
0.498779296875
],
[
0.50341796875,
0.488037109375,
0.890625,
0.488037109375,
0.890625,
0.501953125,
0.50341796875,
0.501953125
],
[
0.1004638671875,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.1004638671875,
0.51318359375
],
[
0.50341796875,
0.501953125,
0.890625,
0.50244140625,
0.890625,
0.517578125,
0.50341796875,
0.5166015625
],
[
0.08251953125,
0.51513671875,
0.468017578125,
0.51513671875,
0.468017578125,
0.52880859375,
0.08251953125,
0.52880859375
],
[
0.5048828125,
0.51806640625,
0.650390625,
0.51806640625,
0.650390625,
0.5322265625,
0.5048828125,
0.5322265625
],
[
0.08251953125,
0.52978515625,
0.468017578125,
0.52978515625,
0.468017578125,
0.54345703125,
0.08251953125,
0.54345703125
],
[
0.50244140625,
0.5322265625,
0.89208984375,
0.53271484375,
0.89208984375,
0.54736328125,
0.50244140625,
0.546875
],
[
0.08087158203125,
0.54443359375,
0.468017578125,
0.54443359375,
0.468017578125,
0.55810546875,
0.08087158203125,
0.55810546875
],
[
0.50341796875,
0.548828125,
0.89208984375,
0.548828125,
0.89208984375,
0.5625,
0.50341796875,
0.5625
],
[
0.08087158203125,
0.5595703125,
0.46728515625,
0.5595703125,
0.46728515625,
0.57275390625,
0.08087158203125,
0.57275390625
],
[
0.50341796875,
0.564453125,
0.701171875,
0.564453125,
0.701171875,
0.57763671875,
0.50341796875,
0.57763671875
],
[
0.08087158203125,
0.57568359375,
0.468017578125,
0.57568359375,
0.468017578125,
0.5888671875,
0.08087158203125,
0.5888671875
],
[
0.080078125,
0.5888671875,
0.4453125,
0.59130859375,
0.4453125,
0.60498046875,
0.080078125,
0.60302734375
],
[
0.50341796875,
0.5869140625,
0.72216796875,
0.58837890625,
0.72216796875,
0.60302734375,
0.50341796875,
0.6015625
],
[
0.33251953125,
0.61279296875,
0.396240234375,
0.61572265625,
0.396240234375,
0.626953125,
0.33251953125,
0.625
],
[
0.52392578125,
0.611328125,
0.890625,
0.611328125,
0.890625,
0.62548828125,
0.52392578125,
0.62548828125
],
[
0.416748046875,
0.61572265625,
0.484375,
0.61572265625,
0.484375,
0.626953125,
0.416748046875,
0.626953125
],
[
0.156005859375,
0.6162109375,
0.2181396484375,
0.6162109375,
0.2181396484375,
0.626953125,
0.156005859375,
0.626953125
],
[
0.2467041015625,
0.61572265625,
0.314453125,
0.61572265625,
0.314453125,
0.626953125,
0.2467041015625,
0.626953125
],
[
0.2646484375,
0.62353515625,
0.295654296875,
0.62353515625,
0.295654296875,
0.646484375,
0.2646484375,
0.646484375
],
[
0.351318359375,
0.625,
0.37744140625,
0.625,
0.37744140625,
0.64599609375,
0.351318359375,
0.64599609375
],
[
0.4345703125,
0.626953125,
0.463134765625,
0.626953125,
0.463134765625,
0.64599609375,
0.4345703125,
0.64599609375
],
[
0.50341796875,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.6416015625,
0.50341796875,
0.6416015625
],
[
0.143798828125,
0.62744140625,
0.1519775390625,
0.62744140625,
0.1519775390625,
0.6318359375,
0.143798828125,
0.6318359375
],
[
0.08660888671875,
0.626953125,
0.13232421875,
0.626953125,
0.13232421875,
0.6337890625,
0.08660888671875,
0.6337890625
],
[
0.2099609375,
0.62744140625,
0.21728515625,
0.62744140625,
0.21728515625,
0.6318359375,
0.2099609375,
0.6318359375
],
[
0.21728515625,
0.62744140625,
0.2303466796875,
0.62744140625,
0.2303466796875,
0.6318359375,
0.21728515625,
0.6318359375
],
[
0.1895751953125,
0.62890625,
0.19775390625,
0.62890625,
0.19775390625,
0.6318359375,
0.1895751953125,
0.6318359375
],
[
0.08905029296875,
0.63330078125,
0.238525390625,
0.63330078125,
0.238525390625,
0.646484375,
0.08905029296875,
0.646484375
],
[
0.238525390625,
0.64306640625,
0.49267578125,
0.64306640625,
0.49267578125,
0.6572265625,
0.238525390625,
0.6572265625
],
[
0.50341796875,
0.642578125,
0.890625,
0.642578125,
0.890625,
0.65576171875,
0.50341796875,
0.65576171875
],
[
0.25,
0.65380859375,
0.271240234375,
0.65576171875,
0.26953125,
0.6640625,
0.2484130859375,
0.662109375
],
[
0.158447265625,
0.646484375,
0.2132568359375,
0.646484375,
0.2132568359375,
0.65380859375,
0.158447265625,
0.65380859375
],
[
0.1380615234375,
0.65576171875,
0.236083984375,
0.65576171875,
0.236083984375,
0.66357421875,
0.1380615234375,
0.66357421875
],
[
0.2900390625,
0.65576171875,
0.313720703125,
0.65576171875,
0.313720703125,
0.66357421875,
0.2900390625,
0.66357421875
],
[
0.3349609375,
0.65576171875,
0.3544921875,
0.65576171875,
0.3544921875,
0.66357421875,
0.3349609375,
0.66357421875
],
[
0.375732421875,
0.65576171875,
0.396240234375,
0.65576171875,
0.396240234375,
0.66357421875,
0.375732421875,
0.66357421875
],
[
0.420654296875,
0.65576171875,
0.43701171875,
0.65576171875,
0.43701171875,
0.66357421875,
0.420654296875,
0.66357421875
],
[
0.5048828125,
0.6572265625,
0.88818359375,
0.65576171875,
0.88818359375,
0.67041015625,
0.5048828125,
0.671875
],
[
0.1356201171875,
0.662109375,
0.236083984375,
0.6640625,
0.236083984375,
0.6748046875,
0.1356201171875,
0.67236328125
],
[
0.1070556640625,
0.65771484375,
0.111083984375,
0.65771484375,
0.111083984375,
0.66357421875,
0.1070556640625,
0.66357421875
],
[
0.2484130859375,
0.66357421875,
0.271240234375,
0.66357421875,
0.271240234375,
0.6748046875,
0.2484130859375,
0.6748046875
],
[
0.33251953125,
0.66357421875,
0.35546875,
0.66357421875,
0.35546875,
0.6748046875,
0.33251953125,
0.6748046875
],
[
0.374267578125,
0.662109375,
0.398681640625,
0.66552734375,
0.396240234375,
0.6767578125,
0.371826171875,
0.673828125
],
[
0.416748046875,
0.66357421875,
0.439453125,
0.66357421875,
0.439453125,
0.6748046875,
0.416748046875,
0.6748046875
],
[
0.459228515625,
0.6640625,
0.4794921875,
0.6640625,
0.4794921875,
0.67236328125,
0.459228515625,
0.67236328125
],
[
0.285888671875,
0.6640625,
0.313720703125,
0.6611328125,
0.316162109375,
0.67236328125,
0.289306640625,
0.67529296875
],
[
0.101318359375,
0.66748046875,
0.111083984375,
0.6640625,
0.11517333984375,
0.671875,
0.10540771484375,
0.6748046875
],
[
0.10540771484375,
0.67041015625,
0.11767578125,
0.6748046875,
0.11273193359375,
0.6845703125,
0.1004638671875,
0.68115234375
],
[
0.136474609375,
0.671875,
0.23779296875,
0.671875,
0.23779296875,
0.681640625,
0.136474609375,
0.681640625
],
[
0.2484130859375,
0.671875,
0.27294921875,
0.671875,
0.27294921875,
0.6845703125,
0.2484130859375,
0.6845703125
],
[
0.289306640625,
0.67041015625,
0.316162109375,
0.673828125,
0.313720703125,
0.68505859375,
0.285888671875,
0.681640625
],
[
0.373291015625,
0.671875,
0.39794921875,
0.671875,
0.39794921875,
0.6845703125,
0.373291015625,
0.6845703125
],
[
0.416748046875,
0.671875,
0.441162109375,
0.671875,
0.441162109375,
0.68505859375,
0.416748046875,
0.68505859375
],
[
0.45751953125,
0.67236328125,
0.483642578125,
0.67236328125,
0.483642578125,
0.6845703125,
0.45751953125,
0.6845703125
],
[
0.5048828125,
0.67236328125,
0.89208984375,
0.67236328125,
0.89208984375,
0.68603515625,
0.5048828125,
0.68603515625
],
[
0.1356201171875,
0.68115234375,
0.23779296875,
0.68115234375,
0.23779296875,
0.69140625,
0.1356201171875,
0.69140625
],
[
0.33251953125,
0.67236328125,
0.35693359375,
0.67236328125,
0.35693359375,
0.6845703125,
0.33251953125,
0.6845703125
],
[
0.2484130859375,
0.68115234375,
0.27294921875,
0.68115234375,
0.27294921875,
0.693359375,
0.2484130859375,
0.693359375
],
[
0.289306640625,
0.68115234375,
0.314453125,
0.68115234375,
0.314453125,
0.693359375,
0.289306640625,
0.693359375
],
[
0.330810546875,
0.68115234375,
0.35693359375,
0.68115234375,
0.35693359375,
0.693359375,
0.330810546875,
0.693359375
],
[
0.373291015625,
0.68115234375,
0.398681640625,
0.68115234375,
0.398681640625,
0.693359375,
0.373291015625,
0.693359375
],
[
0.416748046875,
0.68115234375,
0.441162109375,
0.68115234375,
0.441162109375,
0.693359375,
0.416748046875,
0.693359375
],
[
0.45751953125,
0.68115234375,
0.48193359375,
0.68115234375,
0.48193359375,
0.693359375,
0.45751953125,
0.693359375
],
[
0.418212890625,
0.68798828125,
0.441162109375,
0.69140625,
0.438720703125,
0.70263671875,
0.4150390625,
0.69970703125
],
[
0.50341796875,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.289306640625,
0.689453125,
0.314453125,
0.689453125,
0.314453125,
0.7021484375,
0.289306640625,
0.7021484375
],
[
0.098876953125,
0.6845703125,
0.1094970703125,
0.6787109375,
0.11767578125,
0.689453125,
0.1070556640625,
0.6943359375
],
[
0.373291015625,
0.689453125,
0.398681640625,
0.689453125,
0.398681640625,
0.7021484375,
0.373291015625,
0.7021484375
],
[
0.45751953125,
0.689453125,
0.48193359375,
0.689453125,
0.48193359375,
0.70068359375,
0.45751953125,
0.70068359375
],
[
0.2484130859375,
0.68994140625,
0.27294921875,
0.68994140625,
0.27294921875,
0.7021484375,
0.2484130859375,
0.7021484375
],
[
0.136474609375,
0.68994140625,
0.23779296875,
0.68994140625,
0.23779296875,
0.7021484375,
0.136474609375,
0.7021484375
],
[
0.330810546875,
0.68994140625,
0.35693359375,
0.68994140625,
0.35693359375,
0.7021484375,
0.330810546875,
0.7021484375
],
[
0.330810546875,
0.69775390625,
0.35546875,
0.69775390625,
0.35546875,
0.71044921875,
0.330810546875,
0.71044921875
],
[
0.373291015625,
0.69775390625,
0.39794921875,
0.69775390625,
0.39794921875,
0.71044921875,
0.373291015625,
0.71044921875
],
[
0.416748046875,
0.69775390625,
0.441162109375,
0.69775390625,
0.441162109375,
0.71044921875,
0.416748046875,
0.71044921875
],
[
0.1356201171875,
0.69873046875,
0.236083984375,
0.69873046875,
0.236083984375,
0.708984375,
0.1356201171875,
0.708984375
],
[
0.2484130859375,
0.69873046875,
0.27294921875,
0.69873046875,
0.27294921875,
0.71044921875,
0.2484130859375,
0.71044921875
],
[
0.289306640625,
0.69873046875,
0.314453125,
0.69873046875,
0.314453125,
0.71044921875,
0.289306640625,
0.71044921875
],
[
0.45751953125,
0.69873046875,
0.48193359375,
0.69873046875,
0.48193359375,
0.71044921875,
0.45751953125,
0.71044921875
],
[
0.1004638671875,
0.70068359375,
0.11273193359375,
0.69775390625,
0.1168212890625,
0.70556640625,
0.10455322265625,
0.71044921875
],
[
0.1004638671875,
0.6923828125,
0.11273193359375,
0.689453125,
0.1168212890625,
0.69873046875,
0.10455322265625,
0.7021484375
],
[
0.369384765625,
0.7041015625,
0.40185546875,
0.7041015625,
0.40185546875,
0.7412109375,
0.369384765625,
0.7412109375
],
[
0.41259765625,
0.7041015625,
0.4453125,
0.7041015625,
0.4453125,
0.74072265625,
0.41259765625,
0.74072265625
],
[
0.50341796875,
0.7021484375,
0.89208984375,
0.7021484375,
0.89208984375,
0.71533203125,
0.50341796875,
0.71533203125
],
[
0.285888671875,
0.705078125,
0.316162109375,
0.705078125,
0.316162109375,
0.73095703125,
0.285888671875,
0.73095703125
],
[
0.2459716796875,
0.705078125,
0.275390625,
0.705078125,
0.275390625,
0.73095703125,
0.2459716796875,
0.73095703125
],
[
0.328369140625,
0.705078125,
0.358642578125,
0.705078125,
0.358642578125,
0.73095703125,
0.328369140625,
0.73095703125
],
[
0.4599609375,
0.70849609375,
0.48046875,
0.70849609375,
0.48046875,
0.71826171875,
0.4599609375,
0.71826171875
],
[
0.136474609375,
0.70849609375,
0.236083984375,
0.70849609375,
0.236083984375,
0.71826171875,
0.136474609375,
0.71826171875
],
[
0.10296630859375,
0.708984375,
0.11358642578125,
0.708984375,
0.11358642578125,
0.72900390625,
0.10296630859375,
0.72900390625
],
[
0.20751953125,
0.7138671875,
0.236083984375,
0.71630859375,
0.2344970703125,
0.72802734375,
0.2059326171875,
0.72607421875
],
[
0.45751953125,
0.71533203125,
0.48046875,
0.71533203125,
0.48046875,
0.72802734375,
0.45751953125,
0.72802734375
],
[
0.5048828125,
0.71728515625,
0.89208984375,
0.71728515625,
0.89208984375,
0.73095703125,
0.5048828125,
0.73095703125
],
[
0.2484130859375,
0.7216796875,
0.273681640625,
0.724609375,
0.271240234375,
0.73779296875,
0.2459716796875,
0.73486328125
],
[
0.1380615234375,
0.71826171875,
0.2099609375,
0.71826171875,
0.2099609375,
0.72607421875,
0.1380615234375,
0.72607421875
],
[
0.1356201171875,
0.724609375,
0.2115478515625,
0.724609375,
0.2115478515625,
0.73486328125,
0.1356201171875,
0.73486328125
],
[
0.289306640625,
0.72607421875,
0.314453125,
0.72607421875,
0.314453125,
0.73779296875,
0.289306640625,
0.73779296875
],
[
0.330810546875,
0.72607421875,
0.35546875,
0.72607421875,
0.35546875,
0.73779296875,
0.330810546875,
0.73779296875
],
[
0.45751953125,
0.72607421875,
0.48046875,
0.72607421875,
0.48046875,
0.7373046875,
0.45751953125,
0.7373046875
],
[
0.2091064453125,
0.7265625,
0.2344970703125,
0.7265625,
0.2344970703125,
0.73486328125,
0.2091064453125,
0.73486328125
],
[
0.10296630859375,
0.7265625,
0.11517333984375,
0.7265625,
0.11517333984375,
0.73486328125,
0.10296630859375,
0.73486328125
],
[
0.371826171875,
0.734375,
0.400390625,
0.734375,
0.400390625,
0.74560546875,
0.371826171875,
0.74560546875
],
[
0.5185546875,
0.73291015625,
0.89208984375,
0.73291015625,
0.89208984375,
0.7470703125,
0.5185546875,
0.7470703125
],
[
0.2467041015625,
0.73486328125,
0.271240234375,
0.736328125,
0.271240234375,
0.744140625,
0.2467041015625,
0.74365234375
],
[
0.13232421875,
0.73486328125,
0.238525390625,
0.73486328125,
0.238525390625,
0.74560546875,
0.13232421875,
0.74560546875
],
[
0.2900390625,
0.73486328125,
0.316162109375,
0.73486328125,
0.316162109375,
0.74365234375,
0.2900390625,
0.74365234375
],
[
0.330078125,
0.73486328125,
0.35546875,
0.73486328125,
0.35546875,
0.74365234375,
0.330078125,
0.74365234375
],
[
0.4150390625,
0.73486328125,
0.439453125,
0.736328125,
0.439453125,
0.744140625,
0.4150390625,
0.74365234375
],
[
0.459228515625,
0.736328125,
0.48193359375,
0.736328125,
0.48193359375,
0.744140625,
0.459228515625,
0.744140625
],
[
0.50244140625,
0.7470703125,
0.6787109375,
0.7470703125,
0.6787109375,
0.76025390625,
0.50244140625,
0.76025390625
],
[
0.1004638671875,
0.736328125,
0.11767578125,
0.736328125,
0.11767578125,
0.74365234375,
0.1004638671875,
0.74365234375
],
[
0.08251953125,
0.7587890625,
0.465576171875,
0.7587890625,
0.465576171875,
0.76953125,
0.08251953125,
0.76953125
],
[
0.080078125,
0.771484375,
0.468017578125,
0.771484375,
0.468017578125,
0.78515625,
0.080078125,
0.78515625
],
[
0.50341796875,
0.7734375,
0.7236328125,
0.7734375,
0.7236328125,
0.787109375,
0.50341796875,
0.787109375
],
[
0.08087158203125,
0.787109375,
0.465576171875,
0.787109375,
0.465576171875,
0.79736328125,
0.08087158203125,
0.79736328125
],
[
0.08087158203125,
0.80029296875,
0.465576171875,
0.80029296875,
0.465576171875,
0.81103515625,
0.08087158203125,
0.81103515625
],
[
0.52294921875,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.52294921875,
0.81005859375
],
[
0.080078125,
0.81298828125,
0.468017578125,
0.81298828125,
0.468017578125,
0.8271484375,
0.080078125,
0.8271484375
],
[
0.50341796875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.080078125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.080078125,
0.84033203125
],
[
0.5048828125,
0.82763671875,
0.8896484375,
0.82763671875,
0.8896484375,
0.83837890625,
0.5048828125,
0.83837890625
],
[
0.080078125,
0.84130859375,
0.301513671875,
0.84130859375,
0.301513671875,
0.85498046875,
0.080078125,
0.85498046875
],
[
0.50341796875,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.5048828125,
0.85791015625,
0.890625,
0.85791015625,
0.890625,
0.86865234375,
0.5048828125,
0.86865234375
],
[
0.08087158203125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88525390625,
0.4697265625,
0.88623046875,
0.4697265625,
0.90087890625,
0.080078125,
0.8994140625
],
[
0.50341796875,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"MLP-based",
"GRU-based",
"其中每个序列中包含的动作属于同一类别，BABEL允",
"Yr-LIT",
"许具有多种动作的序列。目前，最适合我们任务的预",
"",
"XN-T:N",
"1:T",
"XN-T:N",
"处理方法是[26]，它导出两个数据集文件：(a)20个动",
"作标签下的单动作序列（如站立、行走、步进和伸展）；",
"HO",
"(b）包含这20个动作的两类动作及其中间转换的序列",
"→",
"+",
"+?",
"?",
"?",
"两者都是30fps，只有姿势参数用于评估，而SMPL形",
"0",
"0",
"状参数被丢弃（可以参考XMPX2中提供的链接）",
"Y",
"Fully-connected",
"由于尚未在BABEL上进行确定性运动预测的实验",
"GRU",
"因此我们首先从上述序列文件中剪切孤立的单位样本",
"GELU",
"2",
"以训练三个基线，观察到的和预测的顿均设置为10。",
"",
"Temporal-mixing",
"Transpose",
"然后，类似于Human3。6M，我们用2轮重新排列我",
"MLP",
"→",
"们的训练样本，同时用2轮进行有效性验证和5轮进",
"Fully-connected",
"",
"Layer Norm",
"行稳定性验证。此外，我们删除了少于200个序列样",
"苗",
"?",
"本的动作，并留下了11个动作类别的样本进行评估",
"GELU",
"1",
"评估度量。在「4,8,12,24,34,401之后，我们通过3D人",
"Transpose",
"-",
"体关节坐标上的平均每关节位置误差（MPJPE）来评",
"4",
"Transpose",
"Spatial-mixing",
"1",
"估我们的网络，该网络计算了预测与相应地面实况之",
"Fully-connected",
"间不同关节差异的平均L2-范数",
"t",
"MLP",
"LayerNorm",
"GELU",
"4.2. Baselines",
"$",
"velocity calculation",
"element-wise subtraction",
"我们的DeFeeNet对其基本预测模型是不可知的，并",
"@element-wise addition",
"+",
"且可以添加到许多现有网络上，以实现连续的人体运",
"动预测。在这里，我们为实验介绍三个基线",
"1",
"Figure 3. Architectures of MLP-based and GRU-based De-",
"FeeNet.",
"LTD-GCN「27]是一个基于GCN的模型，由12个",
"GCN块组成，描述了不同接头尺寸与残余连接的空间",
"相关性。离散余弦变换（DCT）用于提取时间特征",
"观察，在孤立的“单元”内",
"STS-GCN[34]设计了允许时空串扰的单图结构，该结",
"对于连续预测任务，我们使用我们的重采样方法（3.4",
"构将图分解为可分离的空间和时间矩阵，以充分学习",
"节）进行进一步的预处理。为了公平比较，我们还将每",
"关节和时间相互作用",
"个预测回合设置为观察到10个顿，并预测以下10个",
"MotionMixer4]提出了一种有效的运动预测模型，该",
"顿。训练样本由两个相邻的回合组成，其中第二轮是",
"模型仅通过顺序混合空间和时间依赖性来采用MLP，",
"10顿，与第一轮相比。测试样本首先安排2轮，以证",
"并通过增强的挤压和激励块",
"明相邻单元之间的偏差反馈是有效的，然后我们安排",
"10轮评估DeFeeNet在相对较长的范围内的稳定性",
"4.3.Implementation Details",
"STS-GCN [3]",
"我们在单个NVIDIARTX2080Ti上使用Adam优化",
"MotionMixer []",
"LTD-GCN[27]",
"LTD-GCN [2]",
"R",
"",
"35.52",
"器[19]在Pytorch[29]框架下进行实验。在基于MLP",
"",
"frame num.",
"",
"10",
"",
"igolated12.69 26.06 52.28 63.53",
"D(MLP) -D(GRU)-D(MLP) -D(GRU)-D(MLP) -D(GRU)",
"的版本中，时间混合和空间混合的隐藏层大小设置为",
"37.31",
"-DeFee(MLP)",
"12.5925.1650.0261.47",
"37.65",
"40.76",
"40.89",
"84.12",
"与基本管道的相应尺寸相同。在基于GRU的版本中",
"10.30 22.60 48.17 59.79",
"",
"35.21",
"38.15",
"38.77",
"32.38",
"32.69",
"35.80",
"12",
"3",
"10.43 22.74 48.44 60.14",
"35.44",
"35.67",
"38.67",
"32.16",
"32.74",
"无论基本模型如何，隐藏大小都设置为256，而完全连",
"10.39 22.69 48.31 59.95",
"38.33",
"35.33",
"35.66",
"38.46",
"38.72",
"32.05",
"32.75",
"32.17",
"接的图层用于尺寸对齐。详细信息在补充材料中提供。",
"35.91",
"t",
"38.90",
"32.59",
"35.31",
"10.38 22.68 48.27 59.91",
"38.45",
"38.37",
"39.05",
"32.31",
"10.38 22.67 48.25 59.87",
"35.30",
"36.10",
"32.69",
"te",
"t5",
"",
"",
"对于LTD-GCN、STS-GCN和MotionMixer，我们使",
"",
"",
"",
"32.66",
"10.37 22.66 48.24 59.88",
"",
"59.89",
"32.77",
"用DeFeeNet联合训练每个基线，学习率分别为0.0005",
"35.33",
"10.3922.6848.27",
"10.40 22.71 48.30",
"35.75",
"38.38",
"32.71",
"59.92",
"r9",
"38.78",
"0.01和0.01。我们在Human3上训练50个纪元。",
"35.35",
"10.41 22.72 48.33 59.96",
"35.70",
"38.49",
"32.44",
"32.75",
"BABEL上的6M和100",
"r10",
"Table2.Consecutive 1oroundsof prediction errors onHu-",
"man3.6M. Values in bold from round 2 to 10 indicate that",
"4.4. Results on Human3.6M",
"deviation-aware prediction stably yields improved perfor-",
"mance.Detailed errors of LTD-DeFee(MLP) at frame 2",
"表1显示了在Human3上插人和不插人DeFeeNet",
"4, 8, 10 are on the left. Average errors (avg) of these four",
"的模型的预测误差。6M数据集。在没有DeFeeNet的",
"testpoints produced by other baselines with DeFeeNet (ab-",
"情况下，当前的基线观察和预测孤立的短单元内的人",
"breviated as D ) are on the right.",
"体运动(i。e。，10顿的长度为400毫秒）。但是，我们",
"联合训练了两轮预测过程，其中在第1轮（r1）中生成",
"BABEL「32]是最近提出的数据集，其中包含有关操作",
"的预测偏差由我们插入的DeFeeNet检测和学习，以便",
"的语言标签。与其他用于人体运动预测的数据集不同，",
"协助原始管道以先前的偏差作为反馈进行预测，以提"
] |
{
"arxiv_id": "2304.04496",
"language": "zh",
"page_idx": 4
}
|
||
1805.06533-zh-0
|
1805.06533-zh-0.jpg
|
## 简单常识故事中角色的朴素心理建模
Hannah Rashkin$^\dagger$, Antoine Bosselut$^\dagger$, Maarten Sap$^\dagger$, Kevin Knight$^\dagger$ and Yejin Choi$^{\dagger9}$ $^{\dagger}$Paul G. Allen School of Computer Science & Engineering, University of Washington
$^{6}$Allen Institute for Artificial Intelligence
$\{\begin{array}{c}{\mathrm{hrashkin,msap,antoineb,yejin~}}\\{\mathrm{@cs.washington.edu}}\\\end{array}$
$^{\ddagger}$ Information Sciences Institute & Computer Science, University of Southern California
$\operatorname{knight@isi.edu}$
## $\operatorname{Abstract}$
Understanding a narrative requires reading between the lines and reasoning about the unspoken but obvious implications about events and people's mental states - a capability that is trivial for humans but remarkably hard for machines. To facilitate research addressing this challenge, we introduce a new annotation framework to explain naive psychology of story characters as fully-specified chains of mental states with respect to motivations and emotional reactions . Our work presents a new large-scale dataset with rich lowlevel annotations and establishes baseline performance on several new tasks, suggesting avenues for future research.
### 1 Introduction
理解一个故事需要推理故事中的事件和人物的心理状态之间的因果关系，即使这些关系没有明确说明。正如常识性故事完形填空共享任务(Mostafazadeh et al., 2017) 所示，这种推理对于统计和神经机器读者来说都非常困难-尽管对人类来说微不足道。人类和机器之间的这种明显的性能差距并不奇怪，因为大多数强大的语言模型都是为了有效地学习本地流利模式而设计的。因此，他们通常缺乏从文本中的表面模式中抽象出来以模拟更复杂的隐含动态的能力，例如直观角色的心理状态或预测其合理的下一步行动
在本文中，我们构建了一种新的注释形式主义，以根据角色的心理状态对常识性短篇小说(Mostafazadeh et al., 2016) 进行密集标记。生成的数据集提供了三个独特的属性。首先，如图$\boxed{\mathrm{E}}1$ 所示，该数据集提供了每个故事角色的动机和情绪反应的完整指定链，作为事件的前后条件。其次，注释包括实体的状态变化，即
<FigureHere>
Figure 1: A story example with partial annotations for motivations (dashed) and emotional reactions (solid). Open text explanations are in black (e.g.,“frustrated”) and formal theory labels are in blue with brackets (e.g.$,“[$esteem$]”).$
|
[
[
0.310546875,
0.08612060546875,
0.69287109375,
0.08612060546875,
0.69287109375,
0.09979248046875,
0.310546875,
0.09979248046875
],
[
0.1368408203125,
0.12469482421875,
0.87744140625,
0.12353515625,
0.87744140625,
0.1383056640625,
0.1368408203125,
0.1395263671875
],
[
0.1368408203125,
0.14013671875,
0.873046875,
0.1436767578125,
0.873046875,
0.157958984375,
0.1368408203125,
0.155029296875
],
[
0.32666015625,
0.157958984375,
0.6826171875,
0.1591796875,
0.6826171875,
0.17333984375,
0.32666015625,
0.1728515625
],
[
0.274658203125,
0.17578125,
0.73486328125,
0.17578125,
0.73486328125,
0.190673828125,
0.274658203125,
0.190673828125
],
[
0.1234130859375,
0.189453125,
0.884765625,
0.190673828125,
0.884765625,
0.2078857421875,
0.1234130859375,
0.207275390625
],
[
0.435791015625,
0.208984375,
0.56689453125,
0.2078857421875,
0.56689453125,
0.22265625,
0.435791015625,
0.223876953125
],
[
0.2626953125,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.28076171875,
0.2626953125,
0.279052734375
],
[
0.14697265625,
0.29345703125,
0.462646484375,
0.29443359375,
0.462646484375,
0.308837890625,
0.14697265625,
0.30810546875
],
[
0.145263671875,
0.31005859375,
0.45751953125,
0.311279296875,
0.45751953125,
0.32470703125,
0.145263671875,
0.32373046875
],
[
0.14697265625,
0.32666015625,
0.4609375,
0.32666015625,
0.4609375,
0.34033203125,
0.14697265625,
0.34033203125
],
[
0.14697265625,
0.34326171875,
0.4609375,
0.34326171875,
0.4609375,
0.35693359375,
0.14697265625,
0.35693359375
],
[
0.14697265625,
0.359375,
0.4609375,
0.359375,
0.4609375,
0.372314453125,
0.14697265625,
0.372314453125
],
[
0.14697265625,
0.374755859375,
0.459228515625,
0.374755859375,
0.459228515625,
0.388427734375,
0.14697265625,
0.388427734375
],
[
0.1485595703125,
0.392578125,
0.45751953125,
0.392578125,
0.45751953125,
0.402587890625,
0.1485595703125,
0.402587890625
],
[
0.14697265625,
0.40673828125,
0.45751953125,
0.40673828125,
0.45751953125,
0.42041015625,
0.14697265625,
0.42041015625
],
[
0.51708984375,
0.407958984375,
0.59375,
0.410400390625,
0.59375,
0.424560546875,
0.51708984375,
0.4228515625
],
[
0.8134765625,
0.407958984375,
0.875,
0.407958984375,
0.875,
0.4228515625,
0.8134765625,
0.4228515625
],
[
0.54150390625,
0.4228515625,
0.56494140625,
0.4228515625,
0.56494140625,
0.435791015625,
0.54150390625,
0.435791015625
],
[
0.8310546875,
0.4228515625,
0.85546875,
0.4228515625,
0.85546875,
0.435791015625,
0.8310546875,
0.435791015625
],
[
0.5400390625,
0.432861328125,
0.60546875,
0.435791015625,
0.603515625,
0.449462890625,
0.5380859375,
0.446533203125
],
[
0.14697265625,
0.42333984375,
0.459228515625,
0.42333984375,
0.459228515625,
0.43701171875,
0.14697265625,
0.43701171875
],
[
0.14697265625,
0.439453125,
0.4609375,
0.439453125,
0.4609375,
0.453125,
0.14697265625,
0.453125
],
[
0.8310546875,
0.43408203125,
0.85546875,
0.43408203125,
0.85546875,
0.446533203125,
0.8310546875,
0.446533203125
],
[
0.61376953125,
0.4453125,
0.77734375,
0.4453125,
0.77734375,
0.4560546875,
0.61376953125,
0.4560546875
],
[
0.14697265625,
0.454833984375,
0.4609375,
0.454833984375,
0.4609375,
0.4697265625,
0.14697265625,
0.4697265625
],
[
0.634765625,
0.45361328125,
0.759765625,
0.454833984375,
0.759765625,
0.466064453125,
0.634765625,
0.465576171875
],
[
0.54150390625,
0.465576171875,
0.56494140625,
0.4697265625,
0.5615234375,
0.47802734375,
0.5400390625,
0.474365234375
],
[
0.57080078125,
0.468505859375,
0.61376953125,
0.4619140625,
0.6171875,
0.468505859375,
0.57421875,
0.474365234375
],
[
0.82958984375,
0.46435546875,
0.85546875,
0.46435546875,
0.85546875,
0.476806640625,
0.82958984375,
0.476806640625
],
[
0.14697265625,
0.471435546875,
0.459228515625,
0.471435546875,
0.459228515625,
0.485107421875,
0.14697265625,
0.485107421875
],
[
0.54248046875,
0.47802734375,
0.55908203125,
0.47802734375,
0.55908203125,
0.486328125,
0.54248046875,
0.486328125
],
[
0.8310546875,
0.474365234375,
0.85546875,
0.474365234375,
0.85546875,
0.488037109375,
0.8310546875,
0.488037109375
],
[
0.14697265625,
0.488037109375,
0.4609375,
0.488037109375,
0.4609375,
0.501953125,
0.14697265625,
0.501953125
],
[
0.61376953125,
0.48681640625,
0.77685546875,
0.48681640625,
0.77685546875,
0.49755859375,
0.61376953125,
0.49755859375
],
[
0.62060546875,
0.496337890625,
0.77685546875,
0.496337890625,
0.77685546875,
0.50732421875,
0.62060546875,
0.50732421875
],
[
0.14697265625,
0.50439453125,
0.459228515625,
0.50439453125,
0.459228515625,
0.51708984375,
0.14697265625,
0.51708984375
],
[
0.5400390625,
0.505859375,
0.5634765625,
0.505859375,
0.5634765625,
0.5185546875,
0.5400390625,
0.5185546875
],
[
0.8310546875,
0.5087890625,
0.85546875,
0.5087890625,
0.85546875,
0.5224609375,
0.8310546875,
0.5224609375
],
[
0.54150390625,
0.51708984375,
0.5634765625,
0.51708984375,
0.5634765625,
0.52978515625,
0.54150390625,
0.52978515625
],
[
0.8330078125,
0.5185546875,
0.85400390625,
0.5185546875,
0.85400390625,
0.53076171875,
0.8330078125,
0.53076171875
],
[
0.62060546875,
0.529296875,
0.78271484375,
0.529296875,
0.78271484375,
0.5390625,
0.62060546875,
0.5390625
],
[
0.14697265625,
0.51953125,
0.45751953125,
0.51953125,
0.45751953125,
0.533203125,
0.14697265625,
0.533203125
],
[
0.14697265625,
0.5361328125,
0.45751953125,
0.5361328125,
0.45751953125,
0.5498046875,
0.14697265625,
0.5498046875
],
[
0.791015625,
0.533203125,
0.8388671875,
0.52978515625,
0.84033203125,
0.54150390625,
0.79248046875,
0.54443359375
],
[
0.609375,
0.5390625,
0.77734375,
0.5390625,
0.77734375,
0.5498046875,
0.609375,
0.5498046875
],
[
0.57861328125,
0.54736328125,
0.61279296875,
0.54443359375,
0.61376953125,
0.5517578125,
0.580078125,
0.5546875
],
[
0.82861328125,
0.54736328125,
0.85791015625,
0.54736328125,
0.85791015625,
0.57470703125,
0.82861328125,
0.57470703125
],
[
0.54248046875,
0.5517578125,
0.56103515625,
0.5517578125,
0.56103515625,
0.56005859375,
0.54248046875,
0.56005859375
],
[
0.145263671875,
0.5517578125,
0.45751953125,
0.55029296875,
0.45751953125,
0.5654296875,
0.145263671875,
0.56640625
],
[
0.5458984375,
0.5634765625,
0.5576171875,
0.5634765625,
0.5576171875,
0.56787109375,
0.5458984375,
0.56787109375
],
[
0.6171875,
0.57373046875,
0.62451171875,
0.57373046875,
0.62451171875,
0.57666015625,
0.6171875,
0.57666015625
],
[
0.623046875,
0.57080078125,
0.77490234375,
0.57080078125,
0.77490234375,
0.58056640625,
0.623046875,
0.58056640625
],
[
0.1259765625,
0.580078125,
0.1368408203125,
0.58740234375,
0.1275634765625,
0.5927734375,
0.118408203125,
0.5859375
],
[
0.151123046875,
0.5791015625,
0.26611328125,
0.580078125,
0.26611328125,
0.59423828125,
0.151123046875,
0.59326171875
],
[
0.640625,
0.58056640625,
0.755859375,
0.58056640625,
0.755859375,
0.59130859375,
0.640625,
0.59130859375
],
[
0.80029296875,
0.5859375,
0.81787109375,
0.5859375,
0.81787109375,
0.59326171875,
0.80029296875,
0.59326171875
],
[
0.5400390625,
0.59326171875,
0.55908203125,
0.59033203125,
0.5615234375,
0.6015625,
0.5439453125,
0.60498046875
],
[
0.587890625,
0.59326171875,
0.60107421875,
0.59326171875,
0.60107421875,
0.5966796875,
0.587890625,
0.5966796875
],
[
0.8251953125,
0.59033203125,
0.85791015625,
0.59033203125,
0.85791015625,
0.6162109375,
0.8251953125,
0.6162109375
],
[
0.12005615234375,
0.60546875,
0.48779296875,
0.60546875,
0.48779296875,
0.619140625,
0.12005615234375,
0.619140625
],
[
0.623046875,
0.6123046875,
0.7734375,
0.6123046875,
0.7734375,
0.62255859375,
0.623046875,
0.62255859375
],
[
0.779296875,
0.61767578125,
0.78857421875,
0.62060546875,
0.78515625,
0.62353515625,
0.77685546875,
0.62060546875
],
[
0.78662109375,
0.6181640625,
0.84228515625,
0.62255859375,
0.84033203125,
0.6328125,
0.78515625,
0.62890625
],
[
0.62890625,
0.62158203125,
0.77099609375,
0.62158203125,
0.77099609375,
0.63525390625,
0.62890625,
0.63525390625
],
[
0.12005615234375,
0.62158203125,
0.48779296875,
0.62158203125,
0.48779296875,
0.63525390625,
0.12005615234375,
0.63525390625
],
[
0.66650390625,
0.63134765625,
0.72900390625,
0.6328125,
0.72900390625,
0.64453125,
0.66650390625,
0.642578125
],
[
0.8251953125,
0.6318359375,
0.85791015625,
0.6318359375,
0.85791015625,
0.6591796875,
0.8251953125,
0.6591796875
],
[
0.5380859375,
0.6328125,
0.56689453125,
0.6328125,
0.56689453125,
0.65771484375,
0.5380859375,
0.65771484375
],
[
0.118408203125,
0.6357421875,
0.484375,
0.6357421875,
0.484375,
0.6494140625,
0.118408203125,
0.6494140625
],
[
0.118408203125,
0.65283203125,
0.48779296875,
0.65185546875,
0.48779296875,
0.66650390625,
0.118408203125,
0.66748046875
],
[
0.12176513671875,
0.66943359375,
0.48779296875,
0.66943359375,
0.48779296875,
0.68310546875,
0.12176513671875,
0.68310546875
],
[
0.51318359375,
0.66650390625,
0.8818359375,
0.66650390625,
0.8818359375,
0.6796875,
0.51318359375,
0.6796875
],
[
0.12005615234375,
0.68603515625,
0.486083984375,
0.68603515625,
0.486083984375,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.51123046875,
0.68115234375,
0.87890625,
0.68115234375,
0.87890625,
0.69482421875,
0.51123046875,
0.69482421875
],
[
0.12005615234375,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.51318359375,
0.69873046875,
0.880859375,
0.69873046875,
0.880859375,
0.7119140625,
0.51318359375,
0.7119140625
],
[
0.118408203125,
0.71728515625,
0.486083984375,
0.71728515625,
0.486083984375,
0.73095703125,
0.118408203125,
0.73095703125
],
[
0.51318359375,
0.71435546875,
0.8818359375,
0.71435546875,
0.8818359375,
0.728515625,
0.51318359375,
0.728515625
],
[
0.118408203125,
0.73291015625,
0.48779296875,
0.73388671875,
0.48779296875,
0.7490234375,
0.118408203125,
0.74755859375
],
[
0.51318359375,
0.72998046875,
0.884765625,
0.72998046875,
0.884765625,
0.74462890625,
0.51318359375,
0.74462890625
],
[
0.12005615234375,
0.75048828125,
0.48779296875,
0.75048828125,
0.48779296875,
0.76416015625,
0.12005615234375,
0.76416015625
],
[
0.5146484375,
0.74658203125,
0.65576171875,
0.74658203125,
0.65576171875,
0.76123046875,
0.5146484375,
0.76123046875
],
[
0.12176513671875,
0.7666015625,
0.48779296875,
0.7666015625,
0.48779296875,
0.78076171875,
0.12176513671875,
0.78076171875
],
[
0.12176513671875,
0.7822265625,
0.214111328125,
0.7822265625,
0.214111328125,
0.79638671875,
0.12176513671875,
0.79638671875
],
[
0.1368408203125,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.1368408203125,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.486083984375,
0.81396484375,
0.486083984375,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.12005615234375,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.12176513671875,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.12176513671875,
0.86083984375
],
[
0.12176513671875,
0.86328125,
0.48779296875,
0.86328125,
0.48779296875,
0.876953125,
0.12176513671875,
0.876953125
],
[
0.118408203125,
0.87890625,
0.484375,
0.87890625,
0.484375,
0.892578125,
0.118408203125,
0.892578125
],
[
0.12005615234375,
0.89453125,
0.48779296875,
0.8935546875,
0.48779296875,
0.90771484375,
0.12005615234375,
0.9091796875
]
] |
[
"简单常识故事中角色的朴素心理建模",
"Hannah Rashkint, Antoine Bosselutt, Maarten Sapt, Kevin Knightt and Yejin Choits",
"tPaul G.Allen School of Computer Science & Engineering, University of Washington",
"Allen Institute forArtificial Intelligence",
"[hrashkin,msap,antoineb,yejin @cs.washington.edu",
"+ Information Sciences Institute & Computer Science, University of Southern California",
"knight@isi.edu",
"Abstract",
"Understanding a narrative requires",
"reading between the lines and reason-",
"ing about the unspoken but obvious",
"implications about events and people's",
"mental states a capability that is",
"trivial for humans but remarkably hard",
"formachines.Tofacilitateresearchad-",
"dressing this challenge, we introduce a",
"Instructor",
"Players",
"E",
"E",
"MIstabilly",
"new annotation framework to explain",
"naive psychology of story characters as",
"M",
"Theband instructortoldthe",
"fully-specified chains of mental states",
"bandto start playing.",
"E",
"confiden",
"E",
"with respect to motivations and emo-",
"M",
"M",
"tional reactions : Our work presents a",
"Heoftenstoppedthemusic",
"when players were off-tone.",
"new large-scale dataset with rich low-",
"E",
"E",
"M",
"M",
"Theygrewtiredand started",
"level annotations and establishes base-",
"line performance on several new tasks,",
"need'rest",
"playingworseafterawhile.",
"ustrated",
"日",
"E",
"suggesting avenues for future research.",
"A",
"",
"The instructor was furious",
"1",
"Introduction",
"and threw his chair.",
"aid",
"E",
"O",
"日",
"理解一个故事需要推理故事中的事件和人物的",
"He cancelled practiceand",
"e",
"fisgustfear",
"expected us to perform",
"心理状态之间的因果关系，即使这些关系没有",
"tomorrow.",
"日",
"日",
"明确说明。正如常识性故事完形填空共享任务",
"(Mostafazadehetal.，2017）所示，这种推理对",
"于统计和神经机器读者来说都非常困难-尽管",
"Figure l:A story example with partial an-",
"对人类来说微不足道。人类和机器之间的这种",
"notations for motivations (dashed) and emo",
"明显的性能差距并不奇怪，因为大多数强大的",
"tional reactions (solid).Open text explana-",
"语言模型都是为了有效地学习本地流利模式而",
"tions are in black (e.g., “frustrated\") and for-",
"设计的。因此，他们通常缺乏从文本中的表面",
"mal theory labels are in blue with brackets",
"模式中抽象出来以模拟更复杂的隐含动态的能",
"(e.g., \"[esteem]\").",
"力，例如直观角色的心理状态或预测其合理的",
"下一步行动",
"在本文中，我们构建了一种新的注释形式主",
"义，以根据角色的心理状态对常识性短篇小说",
"(Mostafazadehetal.，2016）进行密集标记。生",
"成的数据集提供了三个独特的属性。首先，如",
"图E1所示，该数据集提供了每个故事角色的",
"动机和情绪反应的完整指定链，作为事件的前",
"后条件。其次，注释包括实体的状态变化，即"
] |
{
"arxiv_id": "1805.06533",
"language": "zh",
"page_idx": 0
}
|
||
1805.06533-zh-2
|
1805.06533-zh-2.jpg
|
<FigureHere>
Figure 3: The annotation pipeline for the fine-grained annotations with an example story.
Ouyang and McKeown, 2015; Lukin et al., 2016) 的相关主题上已经开发了各种注释资源，但我们的研究是第一个注释故事角色的心理状态效应的完整链。这带来了几个独特的挑战，因为注释需要(1)解释话语(2)理解隐含的因果效应，以及(3)理解形式心理学理论类别。在先前的文献中，这种复杂性的注释通常由专家 (Deng and Wiebe, 2015; Ouyang and McKeown, 2015) 执行。虽然可靠，但这些注释的扩展成本过高。因此，我们引入了一个新的注释框架，该框架将一组较小的孤立任务流水线化，如图 3 所示。所有注释都是使用来自AmazonMechanicalTurk 的众包工作人员收集的。
## 3.1 Annotation Pipeline
(e。g。“亚历克斯旅行”)不被视为角色可能有潜在动机的意志行为。对于每一行和故事字符对，我们获得 4 个注释。因为在下一阶段的注释中仍然可以过滤掉对，所以我们选择一个慷慨的阈值，其中只有 2 个注释者必须投票认为句子发生了有意的动作，以用作动机注释任务的输人(3a)
( 2b) Affect Resolution 此任务旨在识别故事角色$j$有情绪反应的所有行。重要的是，即使字符未明确出现在句子$k$中，通常也可以推断出字符$j$的情绪反应。例如，在图 3中，我们想要注释叙述者对第 2 行的反应，即使他们没有被提及，因为他们的情绪反应是可推断的。我们每行每个字符获得 4 个注释。至少有 2 个注释者投票的行用作下一个任务的输人：(3b) 情绪反应
(3a)Motivation 我们使用动作解决阶段(2a) 的输出，要求工作人员在有意发起事件的行中注释字符动机。我们为故事中的一行，前几行和特定字符提供 3 个注释器。他们被要求产生一个自由响应句子，描述是什么原因导致角色在该行中的行为，并选择最相关的马斯洛类别和赖斯子类别。在图$\overline{\mathrm{E}}$3 中，注释者将第 1 行中的叙述者的动机描述为希望“有陪伴”,然后选择了爱 (Maslow)和家庭(Reiss)作为分类标签。由于许多注释者不熟悉激励理论，因此我们要求他们在第一次尝试任务时完成教程( 3b) Emotional Reaction 同时，我们使用来自影响解决阶段(2b)的输出来询问工作人员， 在他们受到影响的行之后，角色的情绪反应是什么。与动机一样，我们从故事，其先前的上下文和特定的角色中给 3 个注释者一行。我们要求他们在开放文本中描述角色在句子中的事件后的感受(最多三种情绪)。作为后续行动， 我们要求员工使用 3 点 likert 评分将他们的免费回复与 Plutchik 类别进行比较。在图 3 中， 我们在第 1 行中包含了对叙述者情绪反应的响应。尽管在该行中没有直接提到叙述者，但
我们描述了下图$\overline{\mathrm{E}}3$ 所示的完整注释管道的组件和工作流程。图中的示例故事用于说明管道中各个步骤的输出(此示例的完整注释在附录中)
$(1){\mathrm{~Entity~Resolution~}}$管道中的第一个任务旨在发现 (1)每个故事$i$中的字符集$E_i$和 (2) 句子集$S_{ij}$ ,其中明确提到了特定字符$j\in E_i$ 。例如，在图 3 中的故事中，注释者识别的字符是“I/me”和“Mycousin”,它们分别出现在句子$\{1,4,5\}$和$\{1,2,3,4,5\}$中
我们使用$S_{ij}$通过修剪未绑定到字符的句子的未来任务来控制管道后续部分的工作流。由于$S_{ij}$用于修剪后续任务，因此我们采用高召回率策略来包括至少一个注释者选择的所有句子
( 2a) Action Resolution 下一个任务确定出现在句子$k$中的字符$j$是否正在采取可以归因于动机的任何动作。我们仅对句子$k\in S_{ij}$ 执行动作解析。在运行示例中，我们想知道第 2 行中的表亲没有做任何有意的事情，从而允许我们在下一个管道阶段(3a)中省略此行，其中注释了角色的动机。状态描述(e。g。,“亚历克斯感到忧郁”)或被动的活动参与
|
[
[
0.5751953125,
0.07598876953125,
0.6826171875,
0.07598876953125,
0.6826171875,
0.08612060546875,
0.5751953125,
0.08612060546875
],
[
0.74072265625,
0.07659912109375,
0.79833984375,
0.07659912109375,
0.79833984375,
0.0872802734375,
0.74072265625,
0.0872802734375
],
[
0.59765625,
0.0872802734375,
0.66162109375,
0.0872802734375,
0.66162109375,
0.097412109375,
0.59765625,
0.097412109375
],
[
0.7421875,
0.0885009765625,
0.80615234375,
0.0885009765625,
0.80615234375,
0.0955810546875,
0.7421875,
0.0955810546875
],
[
0.59423828125,
0.0955810546875,
0.6650390625,
0.0955810546875,
0.6650390625,
0.10687255859375,
0.59423828125,
0.10687255859375
],
[
0.7421875,
0.0955810546875,
0.79833984375,
0.0955810546875,
0.79833984375,
0.10687255859375,
0.7421875,
0.10687255859375
],
[
0.26025390625,
0.10394287109375,
0.298828125,
0.105712890625,
0.297119140625,
0.1175537109375,
0.258544921875,
0.11517333984375
],
[
0.5341796875,
0.114013671875,
0.60693359375,
0.114013671875,
0.60693359375,
0.1258544921875,
0.5341796875,
0.1258544921875
],
[
0.6171875,
0.10986328125,
0.62646484375,
0.10986328125,
0.62646484375,
0.11517333984375,
0.6171875,
0.11517333984375
],
[
0.1822509765625,
0.1181640625,
0.364501953125,
0.1181640625,
0.364501953125,
0.1258544921875,
0.1822509765625,
0.1258544921875
],
[
0.5283203125,
0.12353515625,
0.60693359375,
0.1258544921875,
0.60693359375,
0.14013671875,
0.5283203125,
0.1383056640625
],
[
0.6630859375,
0.12054443359375,
0.755859375,
0.12054443359375,
0.755859375,
0.1318359375,
0.6630859375,
0.1318359375
],
[
0.181396484375,
0.1300048828125,
0.197265625,
0.1300048828125,
0.197265625,
0.1383056640625,
0.181396484375,
0.1383056640625
],
[
0.2166748046875,
0.127685546875,
0.345947265625,
0.127685546875,
0.345947265625,
0.1383056640625,
0.2166748046875,
0.1383056640625
],
[
0.2149658203125,
0.1383056640625,
0.355224609375,
0.1383056640625,
0.355224609375,
0.1519775390625,
0.2149658203125,
0.1519775390625
],
[
0.4013671875,
0.1395263671875,
0.468505859375,
0.1395263671875,
0.468505859375,
0.15087890625,
0.4013671875,
0.15087890625
],
[
0.1796875,
0.149658203125,
0.364501953125,
0.149658203125,
0.364501953125,
0.163330078125,
0.1796875,
0.163330078125
],
[
0.181396484375,
0.141357421875,
0.2091064453125,
0.141357421875,
0.2091064453125,
0.149658203125,
0.181396484375,
0.149658203125
],
[
0.4013671875,
0.1519775390625,
0.478515625,
0.1519775390625,
0.478515625,
0.163330078125,
0.4013671875,
0.163330078125
],
[
0.208251953125,
0.160888671875,
0.36279296875,
0.163330078125,
0.36279296875,
0.17333984375,
0.208251953125,
0.171630859375
],
[
0.5341796875,
0.162109375,
0.603515625,
0.162109375,
0.603515625,
0.1728515625,
0.5341796875,
0.1728515625
],
[
0.6591796875,
0.1602783203125,
0.75732421875,
0.1602783203125,
0.75732421875,
0.171630859375,
0.6591796875,
0.171630859375
],
[
0.20068359375,
0.1650390625,
0.208251953125,
0.1650390625,
0.208251953125,
0.169189453125,
0.20068359375,
0.169189453125
],
[
0.183837890625,
0.1650390625,
0.197265625,
0.1650390625,
0.197265625,
0.171630859375,
0.183837890625,
0.171630859375
],
[
0.2032470703125,
0.171630859375,
0.36865234375,
0.171630859375,
0.36865234375,
0.185302734375,
0.2032470703125,
0.185302734375
],
[
0.53564453125,
0.17333984375,
0.615234375,
0.17333984375,
0.615234375,
0.18408203125,
0.53564453125,
0.18408203125
],
[
0.67822265625,
0.17333984375,
0.73779296875,
0.17333984375,
0.73779296875,
0.18408203125,
0.67822265625,
0.18408203125
],
[
0.270263671875,
0.18408203125,
0.297119140625,
0.18701171875,
0.2939453125,
0.198974609375,
0.267822265625,
0.1953125
],
[
0.75390625,
0.1787109375,
0.76416015625,
0.1787109375,
0.76416015625,
0.182861328125,
0.75390625,
0.182861328125
],
[
0.486083984375,
0.1859130859375,
0.49365234375,
0.18115234375,
0.49951171875,
0.185302734375,
0.491943359375,
0.190673828125
],
[
0.447509765625,
0.1953125,
0.54248046875,
0.1953125,
0.54248046875,
0.2060546875,
0.447509765625,
0.2060546875
],
[
0.57275390625,
0.19482421875,
0.6826171875,
0.1953125,
0.6826171875,
0.2060546875,
0.57275390625,
0.204833984375
],
[
0.71533203125,
0.1953125,
0.8134765625,
0.1953125,
0.8134765625,
0.2060546875,
0.71533203125,
0.2060546875
],
[
0.453369140625,
0.2060546875,
0.5400390625,
0.2060546875,
0.5400390625,
0.213134765625,
0.453369140625,
0.213134765625
],
[
0.60546875,
0.2060546875,
0.654296875,
0.2060546875,
0.654296875,
0.213134765625,
0.60546875,
0.213134765625
],
[
0.736328125,
0.2060546875,
0.80419921875,
0.2060546875,
0.80419921875,
0.213134765625,
0.736328125,
0.213134765625
],
[
0.45166015625,
0.21435546875,
0.54150390625,
0.21435546875,
0.54150390625,
0.2244873046875,
0.45166015625,
0.2244873046875
],
[
0.599609375,
0.21435546875,
0.66015625,
0.21435546875,
0.66015625,
0.2244873046875,
0.599609375,
0.2244873046875
],
[
0.71533203125,
0.213134765625,
0.81201171875,
0.215576171875,
0.81201171875,
0.226806640625,
0.71533203125,
0.2244873046875
],
[
0.130126953125,
0.2340087890625,
0.861328125,
0.235107421875,
0.861328125,
0.248779296875,
0.130126953125,
0.2476806640625
],
[
0.1234130859375,
0.271484375,
0.486083984375,
0.271484375,
0.486083984375,
0.281982421875,
0.1234130859375,
0.281982421875
],
[
0.51708984375,
0.2724609375,
0.57421875,
0.2724609375,
0.57421875,
0.283203125,
0.51708984375,
0.283203125
],
[
0.5859375,
0.26953125,
0.8818359375,
0.26953125,
0.8818359375,
0.283203125,
0.5859375,
0.283203125
],
[
0.12005615234375,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.12005615234375,
0.2998046875
],
[
0.515625,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.515625,
0.2998046875
],
[
0.118408203125,
0.300537109375,
0.486083984375,
0.300537109375,
0.486083984375,
0.314208984375,
0.118408203125,
0.314208984375
],
[
0.515625,
0.3017578125,
0.8818359375,
0.3017578125,
0.8818359375,
0.3154296875,
0.515625,
0.3154296875
],
[
0.12005615234375,
0.318359375,
0.48779296875,
0.318359375,
0.48779296875,
0.33203125,
0.12005615234375,
0.33203125
],
[
0.515625,
0.318359375,
0.86962890625,
0.318359375,
0.86962890625,
0.33203125,
0.515625,
0.33203125
],
[
0.8671875,
0.322509765625,
0.8759765625,
0.322509765625,
0.8759765625,
0.32666015625,
0.8671875,
0.32666015625
],
[
0.12005615234375,
0.3349609375,
0.48779296875,
0.3349609375,
0.48779296875,
0.3486328125,
0.12005615234375,
0.3486328125
],
[
0.515625,
0.3349609375,
0.8818359375,
0.3349609375,
0.8818359375,
0.3486328125,
0.515625,
0.3486328125
],
[
0.12176513671875,
0.352783203125,
0.486083984375,
0.352783203125,
0.486083984375,
0.363525390625,
0.12176513671875,
0.363525390625
],
[
0.5146484375,
0.349853515625,
0.880859375,
0.349853515625,
0.880859375,
0.363525390625,
0.5146484375,
0.363525390625
],
[
0.12176513671875,
0.36767578125,
0.486083984375,
0.36767578125,
0.486083984375,
0.38134765625,
0.12176513671875,
0.38134765625
],
[
0.51318359375,
0.366455078125,
0.634765625,
0.365234375,
0.634765625,
0.380126953125,
0.51318359375,
0.38134765625
],
[
0.12005615234375,
0.3818359375,
0.486083984375,
0.3818359375,
0.486083984375,
0.3955078125,
0.12005615234375,
0.3955078125
],
[
0.12005615234375,
0.3984375,
0.48779296875,
0.3984375,
0.48779296875,
0.412109375,
0.12005615234375,
0.412109375
],
[
0.5205078125,
0.393798828125,
0.8818359375,
0.393798828125,
0.8818359375,
0.40673828125,
0.5205078125,
0.40673828125
],
[
0.515625,
0.4091796875,
0.8818359375,
0.4091796875,
0.8818359375,
0.42333984375,
0.515625,
0.42333984375
],
[
0.12005615234375,
0.4150390625,
0.48779296875,
0.4150390625,
0.48779296875,
0.4287109375,
0.12005615234375,
0.4287109375
],
[
0.118408203125,
0.43115234375,
0.48779296875,
0.43115234375,
0.48779296875,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.5146484375,
0.427001953125,
0.87890625,
0.427001953125,
0.87890625,
0.43701171875,
0.5146484375,
0.43701171875
],
[
0.515625,
0.441162109375,
0.8818359375,
0.441162109375,
0.8818359375,
0.454833984375,
0.515625,
0.454833984375
],
[
0.12005615234375,
0.44775390625,
0.486083984375,
0.44775390625,
0.486083984375,
0.46142578125,
0.12005615234375,
0.46142578125
],
[
0.12005615234375,
0.463134765625,
0.48779296875,
0.463134765625,
0.48779296875,
0.476806640625,
0.12005615234375,
0.476806640625
],
[
0.515625,
0.458984375,
0.880859375,
0.458984375,
0.880859375,
0.4697265625,
0.515625,
0.4697265625
],
[
0.515625,
0.473876953125,
0.8818359375,
0.473876953125,
0.8818359375,
0.48681640625,
0.515625,
0.48681640625
],
[
0.12005615234375,
0.479736328125,
0.14697265625,
0.479736328125,
0.14697265625,
0.493408203125,
0.12005615234375,
0.493408203125
],
[
0.515625,
0.4892578125,
0.8818359375,
0.4892578125,
0.8818359375,
0.5029296875,
0.515625,
0.5029296875
],
[
0.162109375,
0.5048828125,
0.33154296875,
0.5048828125,
0.33154296875,
0.5185546875,
0.162109375,
0.5185546875
],
[
0.515625,
0.5048828125,
0.88330078125,
0.5048828125,
0.88330078125,
0.51953125,
0.515625,
0.51953125
],
[
0.51318359375,
0.5224609375,
0.587890625,
0.5224609375,
0.587890625,
0.53369140625,
0.51318359375,
0.53369140625
],
[
0.12005615234375,
0.505859375,
0.1695556640625,
0.505859375,
0.1695556640625,
0.51708984375,
0.12005615234375,
0.51708984375
],
[
0.12005615234375,
0.525390625,
0.486083984375,
0.525390625,
0.486083984375,
0.5390625,
0.12005615234375,
0.5390625
],
[
0.12176513671875,
0.54150390625,
0.48779296875,
0.54150390625,
0.48779296875,
0.5546875,
0.12176513671875,
0.5546875
],
[
0.51904296875,
0.54638671875,
0.8818359375,
0.54638671875,
0.8818359375,
0.56005859375,
0.51904296875,
0.56005859375
],
[
0.12176513671875,
0.55859375,
0.486083984375,
0.55859375,
0.486083984375,
0.5693359375,
0.12176513671875,
0.5693359375
],
[
0.515625,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.515625,
0.57666015625
],
[
0.12005615234375,
0.57470703125,
0.1485595703125,
0.57470703125,
0.1485595703125,
0.58740234375,
0.12005615234375,
0.58740234375
],
[
0.515625,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59326171875,
0.515625,
0.59326171875
],
[
0.1259765625,
0.59716796875,
0.486083984375,
0.59716796875,
0.486083984375,
0.61083984375,
0.1259765625,
0.61083984375
],
[
0.515625,
0.595703125,
0.8818359375,
0.595703125,
0.8818359375,
0.609375,
0.515625,
0.609375
],
[
0.12176513671875,
0.6142578125,
0.48779296875,
0.6142578125,
0.48779296875,
0.62744140625,
0.12176513671875,
0.62744140625
],
[
0.5146484375,
0.6123046875,
0.880859375,
0.6123046875,
0.880859375,
0.6259765625,
0.5146484375,
0.6259765625
],
[
0.12005615234375,
0.6298828125,
0.48779296875,
0.6298828125,
0.48779296875,
0.64453125,
0.12005615234375,
0.64453125
],
[
0.515625,
0.62744140625,
0.8818359375,
0.62744140625,
0.8818359375,
0.64111328125,
0.515625,
0.64111328125
],
[
0.1318359375,
0.6455078125,
0.48779296875,
0.6455078125,
0.48779296875,
0.6591796875,
0.1318359375,
0.6591796875
],
[
0.515625,
0.64453125,
0.8818359375,
0.64453125,
0.8818359375,
0.65771484375,
0.515625,
0.65771484375
],
[
0.118408203125,
0.66015625,
0.489501953125,
0.6611328125,
0.489501953125,
0.67578125,
0.118408203125,
0.6748046875
],
[
0.1292724609375,
0.6494140625,
0.1368408203125,
0.6494140625,
0.1368408203125,
0.65478515625,
0.1292724609375,
0.65478515625
],
[
0.515625,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.515625,
0.67333984375
],
[
0.118408203125,
0.6767578125,
0.393798828125,
0.6767578125,
0.393798828125,
0.6904296875,
0.118408203125,
0.6904296875
],
[
0.51708984375,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.51708984375,
0.6904296875
],
[
0.139404296875,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.139404296875,
0.7080078125
],
[
0.515625,
0.6923828125,
0.8818359375,
0.6923828125,
0.8818359375,
0.7060546875,
0.515625,
0.7060546875
],
[
0.12176513671875,
0.7119140625,
0.484375,
0.7119140625,
0.484375,
0.72265625,
0.12176513671875,
0.72265625
],
[
0.51708984375,
0.71044921875,
0.880859375,
0.71044921875,
0.880859375,
0.72021484375,
0.51708984375,
0.72021484375
],
[
0.12176513671875,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.74072265625,
0.12176513671875,
0.74072265625
],
[
0.5205078125,
0.73291015625,
0.8818359375,
0.73388671875,
0.8818359375,
0.7490234375,
0.5205078125,
0.74755859375
],
[
0.12005615234375,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75439453125,
0.12005615234375,
0.75439453125
],
[
0.515625,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.515625,
0.76318359375
],
[
0.12005615234375,
0.7587890625,
0.139404296875,
0.7587890625,
0.139404296875,
0.7724609375,
0.12005615234375,
0.7724609375
],
[
0.515625,
0.76708984375,
0.8818359375,
0.76708984375,
0.8818359375,
0.78076171875,
0.515625,
0.78076171875
],
[
0.1275634765625,
0.78076171875,
0.486083984375,
0.78076171875,
0.486083984375,
0.7958984375,
0.1275634765625,
0.7958984375
],
[
0.5146484375,
0.7822265625,
0.87890625,
0.7822265625,
0.87890625,
0.7958984375,
0.5146484375,
0.7958984375
],
[
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,
0.8154296875,
0.489501953125,
0.8291015625,
0.118408203125,
0.82763671875
],
[
0.515625,
0.81640625,
0.880859375,
0.81640625,
0.880859375,
0.82666015625,
0.515625,
0.82666015625
],
[
0.118408203125,
0.83056640625,
0.48779296875,
0.830078125,
0.48779296875,
0.84423828125,
0.118408203125,
0.845703125
],
[
0.51318359375,
0.830078125,
0.880859375,
0.83056640625,
0.880859375,
0.845703125,
0.51318359375,
0.84423828125
],
[
0.12005615234375,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.5146484375,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.12176513671875,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.12176513671875,
0.876953125
],
[
0.515625,
0.86328125,
0.880859375,
0.86328125,
0.880859375,
0.876953125,
0.515625,
0.876953125
],
[
0.12176513671875,
0.8798828125,
0.491943359375,
0.8798828125,
0.491943359375,
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,
0.489501953125,
0.89453125,
0.489501953125,
0.9091796875,
0.118408203125,
0.91015625
],
[
0.5146484375,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"Character action lines:",
"Motivation:",
"I, me: 1,4, 5",
"Line1,I,me",
"Cousin: 3, 4, 5",
"love/family",
"Story:",
"(2a)Action",
"T",
"(1)Iletmycousinstaywithme",
"Resolution",
"(3a)Motivation",
"(2)",
"Hehad nowheretogo",
"However, he was a slob.",
"(1)Entity",
"(4) I was about to kick him out.",
"(3)",
"Resolution",
"Whenhe cooked mea huge",
"(2b) Affect",
"(3b)Emotional",
"",
"5)",
"breakfast, I decided he could",
"Resolution",
"Reaction",
"stay",
"",
".",
"Character mentions:",
"Character affect lines:",
"Emotional Reaction",
"1.me(lines 1.4,5).",
"I.me:2-5",
"Line3.Lme:",
"Cousin (lines 1-5)",
"Cousin: 2,5",
"sad/disgusted/angry",
"Figure 3: The annotation pipeline for the fine-grained annotations with an example story.",
"Ouyang and McKeown,2015:Lukin et al..",
"(e. g..",
"“亚历克斯旅行”）不被视为角色可能",
"2016）的相关主题上已经开发了各种注释资",
"有潜在动机的意志行为。对于每一行和故事字",
"源，但我们的研究是第一个注释故事角色的心",
"符对，我们获得4个注释。因为在下一阶段的",
"理状态效应的完整链。这带来了几个独特的挑",
"注释中仍然可以过滤掉对，所以我们选择一",
"",
"战，因为注释需要（1）解释话语（2）理解隐含",
"慨的阈值，其中只有2个注释者必须投票认",
"的因果效应，以及（3）理解形式心理学理论类",
"为句子发生了有意的动作，以用作动机注释任",
"别。在先前的文献中，这种复杂性的注释通常",
"务的输入（3a)",
"由专家 (Deng and Wiebe,2015:Ouyang and",
"McKeown，2015）执行。虽然可靠，但这些注",
"(2b）AffectResolution此任务旨在识别故事",
"角色有情绪反应的所有行。重要的是，即使",
"释的扩展成本过高。因此，我们引人了一个新",
"的注释框架，该框架将一组较小的孤立任务流",
"字符未明确出现在子k中：通常也可以推断",
"出字符的情绪反应。例如，在图3中，我们",
"水线化，如图3所示。所有注释都是使用来自",
"AmazonMechanicalTurk的众包工作人员收集",
"想要注释叙述者对第2行的反应，即使他们没",
"有被提及，因为他们的情绪反应是可推断的。",
"的。",
"我们每行每个字符获得4个注释。至少有2个",
"Annotation Pipeline",
"注释者投票的行用作下一个任务的输人：（3b）",
"情绪反应",
"3.1",
"我们描述了下图E3所示的完整注释管道的组",
"件和工作流程。图中的示例故事用于说明管道",
"(3a）Motivation我们使用动作解决阶段（2a）",
"中各个步骤的输出（此示例的完整注释在附录",
"的输出，要求工作人员在有意发起事件的行中",
"中)",
"注释字符动机。我们为故事中的一行，前几行",
"(1）EntityResolution管道中的第一个任务",
"和特定字符提供3个注释器。他们被要求产生",
"旨在发现（1）每个故事i中的字符集E和（2）",
"一个自由响应句子，描述是什么原因导致角色",
"句子集Sii，其中明确提到了特定字符jEE",
"在该行中的行为，并选择最相关的马斯洛类别",
"例如，在图3中的故事中，注释者识别的字",
"和赖斯子类别。在图E3中，注释者将第1行",
"符是“I/me”和“Mycousin”，它们分别出现",
"",
"中的叙述者的动机描述为希望“有陪伴”，然",
"在句子[1,4,5]和[1,2,3,4,5}中",
"后选择了爱（Maslow）和家庭（Reiss）作为分",
"我们使用Sii通过修剪未绑定到字符的句子",
"类标签。由于许多注释者不熟悉激励理论，因",
"的未来任务来控制管道后续部分的工作流。由",
"此我们要求他们在第一次尝试任务时完成教程",
"于Si用于修剪后续任务，因此我们采用高召",
"(3b）EmotionalReaction同时，我们使用来",
"回率策略来包括至少一个注释者选择的所有右",
"自影响解决阶段（2b）的输出来询问工作人员，",
"子",
"在他们受到影响的行之后，角色的情绪反应是",
"(2a）ActionResolution下一个任务确定出",
"什么。与动机一样，我们从故事，其先前的上",
"现在句子k中的字符是否正在采取可以归",
"下文和特定的角色中给3个注释者一行。我们",
"因于动机的任何动作。我们仅对句子kESi",
"要求他们在开放文本中描述角色在句子中的事",
"执行动作解析。在运行示例中，我们想知道",
"件后的感受（最多三种情绪）。作为后续行动",
"第2行中的表亲没有做任何有意的事情，从",
"我们要求员工使用3点likert评分将他们的免",
"而允许我们在下一个管道阶段（3a）中省略",
"费回复与Plutchik类别进行比较。在图3中",
"此行，其中注释了角色的动机。状态描述（e。",
"我们在第1行中包含了对叙述者情绪反应的",
"g。，“亚历克斯感到忧郁”）或被动的活动参与",
"响应。尽管在该行中没有直接提到叙述者，但"
] |
{
"arxiv_id": "1805.06533",
"language": "zh",
"page_idx": 2
}
|
||
1805.06533-zh-3
|
1805.06533-zh-3.jpg
|
$\#$Unique
% Annotations
% Annotations
$\#$Unique
Text
<table>
<tbody>
<tr>
<th>become experienced</th>
<th>8740</th>
<th>$Spir.$ . qrowth</th>
<th>0.17</th>
<th> $\begin{array}{c}{{\mathrm{i}}}\\{{\mathrm{i}}}\end{array}$ $\vdots$ $\begin{array}{c}{:}\\{:}\end{array}$ i 1</th>
</tr>
<tr>
<td>meet goal; to look nice</td>
<td>9676</td>
<td>Esteem</td>
<td>0.</td>
<td>222</td>
</tr>
<tr>
<td>to support his friends</td>
<td>19768</td>
<td>Love</td>
<td> </td>
<td>0</td>
</tr>
<tr>
<td>be employed; stay dry</td>
<td>11645</td>
<td>Stability</td>
<td> </td>
<td>0</td>
</tr>
<tr>
<td>rest more; food</td>
<td>4953</td>
<td>$Phys.$</td>
<td>0.13</td>
<td>****</td>
</tr>
</tbody>
</table>
Open Text
where selected
$\underline{2055}$
outraged; picky
disgust 0.14
1888888888
dismayed; disoriented 3205
surprise $\begin{array}{cc}\text{surprise}&0.33\\\text{anger}&0.16\end{array}$
$\underline{2016}$
enraged; provoked
anger
25
touched; appreciated 2964
where selected
trust
$\underline{2371}$
0.23
sadness
excluded; bleak
0.49
future oriented; uptight $\frac{3868}{213c}$ anticipation
$\underline{3136}$
0.51
joy
happier; jubiliation
12
fear
frozen in fear; fatalistic
Figure 4: Examples of open-text explanations that annotators provided corresponding with the categories they selected. The bars on the right of the categories represent the percentage of lines where annotators selected that category (out of those character-line pairs with positive motivation/emotional reaction).
<table>
<tbody>
<tr>
<th> </th>
<th rowspan="2">train</th>
<th colspan="2">Fine-grained</th>
<th rowspan="2">Label Type</th>
<th> </th>
<th rowspan="2">$PPA$</th>
<th rowspan="2">KA</th>
<th rowspan="2">$\%$ Agree $W$ Maj. Lbl</th>
</tr>
<tr>
<th> </th>
<th>dev</th>
<th>test</th>
<th> </th>
</tr>
<tr>
<td>$\#$ annotated stories</td>
<td>10000</td>
<td>2500</td>
<td>2500</td>
<td> </td>
<td>1</td>
<td>7</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\#$ characters story</td>
<td>2.03</td>
<td>2.02</td>
<td>1.82</td>
<td>Maslow</td>
<td> </td>
<td>777</td>
<td>$VV$ 11</td>
<td> </td>
</tr>
<tr>
<td>77 元L 110 1</td>
<td>40104</td>
<td>${\mathcal{O}}/0{\mathcal{L}}$</td>
<td>$000\bot$ 7</td>
<td rowspan="2">Reiss</td>
<td>Dev</td>
<td>.91</td>
<td>24</td>
<td>0.95</td>
</tr>
<tr>
<td>$\mathbf{vv}/$</td>
<td>$VV\bot V$</td>
<td>${\mathbf{L}}{\mathbf{L}}{\mathbf{V}}{\mathbf{V}}{\mathbf{L}}$</td>
<td>LVIVW 4</td>
<td>Test</td>
<td>.91</td>
<td>24</td>
<td>0.95</td>
</tr>
<tr>
<td>able 1: Annotated da</td>
<td>,ta stati:</td>
<td>stics for</td>
<td>each</td>
<td> </td>
<td>Dev</td>
<td>.71</td>
<td>.32</td>
<td>0.84</td>
</tr>
<tr>
<td>ataset</td>
<td> </td>
<td> </td>
<td> </td>
<td>Plutcmk</td>
<td>Test</td>
<td>.70</td>
<td>.29</td>
<td>0.83</td>
</tr>
</tbody>
</table>
一位注释者记录说，他们会对表弟感到“烦恼” 而感到 slob,并选择了 Plutchik 类别以表示悲伤，厌恶和愤怒
Table 2: Agreement Statistics (PPA=Pairwise percent agreement of worker responses per binary category, KA= Krippendorff's Alpha)
3.2 Dataset Statistics and Insights
平均每个类别的协议来比较注释器的性能。对于 Plutchik 分数，我们将“中度关联”评级视为与“高度关联”评级一致。一致性百分比和Krippendorff 的 如表 2 所示。我们还计算单个注释和大多数标签之间的一致性百分比。$^2$
<NEXT\_LINE
为了更清楚地了解此数据集中的一致性，我们为注释者对创建聚合混淆矩阵。首先，我们对所有配对注释之间的答案组合计数进行求和 (不包括 none 标签)。如果注释者选择了多个类别，我们将在所选类别之间均匀地拆分计数。我们在总混淆矩阵上计算 NPMI。在图 5中，我们显示了动机类别的 NPMI 混淆矩阵
在动机注释中，我们在对角线上找到最高分(i。e。, Reiss 协议), 大多数混淆发生在同一马斯洛类别的 Reiss 动机之间 (图$\overline{\mathrm{E}}5$ 中的黑色轮廓)。其他分歧通常涉及主题相似的 Reiss
Cost 与理论类别分配相对应的任务是管道中最困难和最昂贵的任务(每个故事$\sim\$4)$。因此，我们只获得三分之一的注释故事的理论类别标签，我们将其分配给开发和测试集。训练数据用缩短的管道进行注释，其中只有两个工人的动机和情绪反应的开放文本描述(每个故事$\sim\$1)$
Scale 到目前为止，我们的数据集包括 15,000个故事(从 ROC 故事训练集中随机选择)中的动机和情感的总共 300k 低级注释。它涵盖了超过 150,000 个字符行对，其中 56k 个字符行对具有带注释的动机，而 105k 个字符行对具有带注释的情感变化(i。e。none 以外的标签)。表 1 显示了跨训练、开发和测试拆分的细分。图 4 显示了在有积极变化的情况下，为动机和情感类别选择不同标签的频率
Agreements 为了进行质量控制，我们删除了始终生产低质量工作的工人，如附录中所述。在分类集 (Maslow, Reiss 和 Plutchik) 中，我们通过将每个单独的类别视为二进制标签(如果他们在其响应集中包含该类别，则为 1)并
$^{2}$Majority label for the motivation categories is what was agreed upon by at least two annotators per category. For emotion categories, we averaged the pointwise ratings and counted a category if the average rat- $\operatorname{ing}$ was$\geq 2$ .
|
[
[
0.2509765625,
0.0736083984375,
0.30322265625,
0.0736083984375,
0.30322265625,
0.08489990234375,
0.2509765625,
0.08489990234375
],
[
0.37451171875,
0.0748291015625,
0.456787109375,
0.0748291015625,
0.456787109375,
0.08489990234375,
0.37451171875,
0.08489990234375
],
[
0.64404296875,
0.0748291015625,
0.69873046875,
0.07659912109375,
0.69873046875,
0.0885009765625,
0.64404296875,
0.08612060546875
],
[
0.77099609375,
0.0748291015625,
0.85205078125,
0.07598876953125,
0.85205078125,
0.0872802734375,
0.77099609375,
0.08612060546875
],
[
0.2476806640625,
0.0843505859375,
0.306396484375,
0.0843505859375,
0.306396484375,
0.09442138671875,
0.2476806640625,
0.09442138671875
],
[
0.372802734375,
0.0843505859375,
0.45751953125,
0.0843505859375,
0.45751953125,
0.09442138671875,
0.372802734375,
0.09442138671875
],
[
0.64208984375,
0.08612060546875,
0.7021484375,
0.08612060546875,
0.7021484375,
0.097412109375,
0.64208984375,
0.097412109375
],
[
0.77099609375,
0.0872802734375,
0.85205078125,
0.0872802734375,
0.85205078125,
0.09442138671875,
0.77099609375,
0.09442138671875
],
[
0.1435546875,
0.1009521484375,
0.2283935546875,
0.1009521484375,
0.2283935546875,
0.11102294921875,
0.1435546875,
0.11102294921875
],
[
0.2509765625,
0.097412109375,
0.2822265625,
0.097412109375,
0.2822265625,
0.1092529296875,
0.2509765625,
0.1092529296875
],
[
0.312255859375,
0.1015625,
0.364501953125,
0.1015625,
0.364501953125,
0.11224365234375,
0.312255859375,
0.11224365234375
],
[
0.372802734375,
0.1009521484375,
0.41064453125,
0.1009521484375,
0.41064453125,
0.11224365234375,
0.372802734375,
0.11224365234375
],
[
0.64990234375,
0.1051025390625,
0.67822265625,
0.1051025390625,
0.67822265625,
0.1134033203125,
0.64990234375,
0.1134033203125
],
[
0.7001953125,
0.105712890625,
0.779296875,
0.105712890625,
0.779296875,
0.1175537109375,
0.7001953125,
0.1175537109375
],
[
0.8037109375,
0.10394287109375,
0.84033203125,
0.10394287109375,
0.84033203125,
0.12054443359375,
0.8037109375,
0.12054443359375
],
[
0.2451171875,
0.11102294921875,
0.279541015625,
0.11102294921875,
0.279541015625,
0.119384765625,
0.2451171875,
0.119384765625
],
[
0.5283203125,
0.10687255859375,
0.640625,
0.10687255859375,
0.640625,
0.1175537109375,
0.5283203125,
0.1175537109375
],
[
0.124267578125,
0.1134033203125,
0.2493896484375,
0.1134033203125,
0.2493896484375,
0.12353515625,
0.124267578125,
0.12353515625
],
[
0.30908203125,
0.1134033203125,
0.364501953125,
0.1134033203125,
0.364501953125,
0.12469482421875,
0.30908203125,
0.12469482421875
],
[
0.411376953125,
0.1134033203125,
0.447509765625,
0.1134033203125,
0.447509765625,
0.12469482421875,
0.411376953125,
0.12469482421875
],
[
0.6513671875,
0.1217041015625,
0.6767578125,
0.1217041015625,
0.6767578125,
0.1300048828125,
0.6513671875,
0.1300048828125
],
[
0.82275390625,
0.1217041015625,
0.8505859375,
0.1217041015625,
0.8505859375,
0.1353759765625,
0.82275390625,
0.1353759765625
],
[
0.72900390625,
0.12353515625,
0.779296875,
0.12353515625,
0.779296875,
0.1353759765625,
0.72900390625,
0.1353759765625
],
[
0.1351318359375,
0.12469482421875,
0.2376708984375,
0.12353515625,
0.2376708984375,
0.1353759765625,
0.1351318359375,
0.135986328125
],
[
0.2509765625,
0.122314453125,
0.280517578125,
0.122314453125,
0.280517578125,
0.1341552734375,
0.2509765625,
0.1341552734375
],
[
0.32421875,
0.12469482421875,
0.364501953125,
0.12469482421875,
0.364501953125,
0.135986328125,
0.32421875,
0.135986328125
],
[
0.37451171875,
0.1258544921875,
0.408935546875,
0.1258544921875,
0.408935546875,
0.1341552734375,
0.37451171875,
0.1341552734375
],
[
0.5205078125,
0.12469482421875,
0.6455078125,
0.12353515625,
0.6455078125,
0.1353759765625,
0.5205078125,
0.135986328125
],
[
0.2509765625,
0.133056640625,
0.279541015625,
0.133056640625,
0.279541015625,
0.144287109375,
0.2509765625,
0.144287109375
],
[
0.328369140625,
0.1353759765625,
0.364501953125,
0.13720703125,
0.36279296875,
0.149658203125,
0.32666015625,
0.1478271484375
],
[
0.397216796875,
0.1383056640625,
0.4248046875,
0.1383056640625,
0.4248046875,
0.1455078125,
0.397216796875,
0.1455078125
],
[
0.744140625,
0.1395263671875,
0.779296875,
0.141357421875,
0.77734375,
0.15380859375,
0.7421875,
0.1519775390625
],
[
0.6513671875,
0.14013671875,
0.6826171875,
0.14013671875,
0.6826171875,
0.1484375,
0.6513671875,
0.1484375
],
[
0.12176513671875,
0.135986328125,
0.2476806640625,
0.135986328125,
0.2476806640625,
0.149658203125,
0.12176513671875,
0.149658203125
],
[
0.252685546875,
0.146728515625,
0.279541015625,
0.146728515625,
0.279541015625,
0.157958984375,
0.252685546875,
0.157958984375
],
[
0.5283203125,
0.1436767578125,
0.640625,
0.142578125,
0.640625,
0.15380859375,
0.5283203125,
0.155029296875
],
[
0.861328125,
0.146728515625,
0.86962890625,
0.146728515625,
0.86962890625,
0.1519775390625,
0.861328125,
0.1519775390625
],
[
0.307373046875,
0.1484375,
0.364501953125,
0.1484375,
0.364501953125,
0.1602783203125,
0.307373046875,
0.1602783203125
],
[
0.1435546875,
0.149658203125,
0.2283935546875,
0.149658203125,
0.2283935546875,
0.1602783203125,
0.1435546875,
0.1602783203125
],
[
0.393798828125,
0.149658203125,
0.4248046875,
0.149658203125,
0.4248046875,
0.157958984375,
0.393798828125,
0.157958984375
],
[
0.2493896484375,
0.1561279296875,
0.365234375,
0.1591796875,
0.365234375,
0.17333984375,
0.2493896484375,
0.17041015625
],
[
0.5224609375,
0.1591796875,
0.646484375,
0.1591796875,
0.646484375,
0.1728515625,
0.5224609375,
0.1728515625
],
[
0.646484375,
0.1591796875,
0.68115234375,
0.1591796875,
0.68115234375,
0.16748046875,
0.646484375,
0.16748046875
],
[
0.724609375,
0.1602783203125,
0.779296875,
0.1602783203125,
0.779296875,
0.171630859375,
0.724609375,
0.171630859375
],
[
0.44921875,
0.1602783203125,
0.483642578125,
0.1602783203125,
0.483642578125,
0.171630859375,
0.44921875,
0.171630859375
],
[
0.12176513671875,
0.160888671875,
0.246826171875,
0.162109375,
0.246826171875,
0.17333984375,
0.12176513671875,
0.1728515625
],
[
0.85791015625,
0.160888671875,
0.87158203125,
0.160888671875,
0.87158203125,
0.16748046875,
0.85791015625,
0.16748046875
],
[
0.252685546875,
0.169189453125,
0.2822265625,
0.169189453125,
0.2822265625,
0.18115234375,
0.252685546875,
0.18115234375
],
[
0.339111328125,
0.1728515625,
0.364501953125,
0.1728515625,
0.364501953125,
0.185302734375,
0.339111328125,
0.185302734375
],
[
0.44921875,
0.17333984375,
0.483642578125,
0.17333984375,
0.483642578125,
0.1817626953125,
0.44921875,
0.1817626953125
],
[
0.73779296875,
0.17333984375,
0.779296875,
0.17578125,
0.77734375,
0.191162109375,
0.736328125,
0.189453125
],
[
0.6513671875,
0.17578125,
0.6767578125,
0.17578125,
0.6767578125,
0.18408203125,
0.6513671875,
0.18408203125
],
[
0.1368408203125,
0.174560546875,
0.235107421875,
0.174560546875,
0.235107421875,
0.185302734375,
0.1368408203125,
0.185302734375
],
[
0.5439453125,
0.1787109375,
0.62451171875,
0.1787109375,
0.62451171875,
0.189453125,
0.5439453125,
0.189453125
],
[
0.79248046875,
0.177001953125,
0.82861328125,
0.177001953125,
0.82861328125,
0.189453125,
0.79248046875,
0.189453125
],
[
0.252685546875,
0.182861328125,
0.2822265625,
0.182861328125,
0.2822265625,
0.191162109375,
0.252685546875,
0.191162109375
],
[
0.334228515625,
0.18408203125,
0.365234375,
0.18408203125,
0.365234375,
0.1953125,
0.334228515625,
0.1953125
],
[
0.39794921875,
0.1859130859375,
0.418212890625,
0.1859130859375,
0.418212890625,
0.193603515625,
0.39794921875,
0.193603515625
],
[
0.210693359375,
0.213134765625,
0.88330078125,
0.21435546875,
0.88330078125,
0.2286376953125,
0.210693359375,
0.22802734375
],
[
0.1234130859375,
0.18701171875,
0.2476806640625,
0.18701171875,
0.2476806640625,
0.19775390625,
0.1234130859375,
0.19775390625
],
[
0.12005615234375,
0.2161865234375,
0.210693359375,
0.2161865234375,
0.210693359375,
0.22802734375,
0.12005615234375,
0.22802734375
],
[
0.1175537109375,
0.2298583984375,
0.8818359375,
0.23095703125,
0.8818359375,
0.2452392578125,
0.1175537109375,
0.24462890625
],
[
0.118408203125,
0.2476806640625,
0.884765625,
0.2476806640625,
0.884765625,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.1175537109375,
0.262451171875,
0.3779296875,
0.26318359375,
0.3779296875,
0.27783203125,
0.1175537109375,
0.276611328125
],
[
0.397216796875,
0.2998046875,
0.5029296875,
0.3017578125,
0.5029296875,
0.317626953125,
0.397216796875,
0.3154296875
],
[
0.78515625,
0.305908203125,
0.884765625,
0.305908203125,
0.884765625,
0.320556640625,
0.78515625,
0.320556640625
],
[
0.51318359375,
0.312255859375,
0.609375,
0.314208984375,
0.609375,
0.328857421875,
0.51318359375,
0.32666015625
],
[
0.68115234375,
0.314208984375,
0.72314453125,
0.314208984375,
0.72314453125,
0.327880859375,
0.68115234375,
0.327880859375
],
[
0.736328125,
0.314208984375,
0.767578125,
0.314208984375,
0.767578125,
0.32666015625,
0.736328125,
0.32666015625
],
[
0.4072265625,
0.322509765625,
0.444091796875,
0.322509765625,
0.444091796875,
0.33837890625,
0.4072265625,
0.33837890625
],
[
0.47265625,
0.322509765625,
0.509765625,
0.322509765625,
0.509765625,
0.33837890625,
0.47265625,
0.33837890625
],
[
0.8076171875,
0.320556640625,
0.88818359375,
0.320556640625,
0.88818359375,
0.33837890625,
0.8076171875,
0.33837890625
],
[
0.1292724609375,
0.339111328125,
0.291259765625,
0.33837890625,
0.291259765625,
0.352783203125,
0.1292724609375,
0.35400390625
],
[
0.33154296875,
0.32373046875,
0.3779296875,
0.32373046875,
0.3779296875,
0.33837890625,
0.33154296875,
0.33837890625
],
[
0.32568359375,
0.339111328125,
0.3779296875,
0.339111328125,
0.3779296875,
0.35400390625,
0.32568359375,
0.35400390625
],
[
0.403076171875,
0.34033203125,
0.444091796875,
0.34033203125,
0.444091796875,
0.35400390625,
0.403076171875,
0.35400390625
],
[
0.468505859375,
0.34033203125,
0.509765625,
0.34033203125,
0.509765625,
0.35400390625,
0.468505859375,
0.35400390625
],
[
0.623046875,
0.34326171875,
0.6630859375,
0.34326171875,
0.6630859375,
0.359375,
0.623046875,
0.359375
],
[
0.69287109375,
0.34326171875,
0.72314453125,
0.34326171875,
0.72314453125,
0.35986328125,
0.69287109375,
0.35986328125
],
[
0.744140625,
0.344482421875,
0.76904296875,
0.344482421875,
0.76904296875,
0.358154296875,
0.744140625,
0.358154296875
],
[
0.84814453125,
0.34326171875,
0.88671875,
0.34326171875,
0.88671875,
0.359375,
0.84814453125,
0.359375
],
[
0.337646484375,
0.355224609375,
0.3779296875,
0.355224609375,
0.3779296875,
0.37060546875,
0.337646484375,
0.37060546875
],
[
0.52880859375,
0.35400390625,
0.591796875,
0.35400390625,
0.591796875,
0.365234375,
0.52880859375,
0.365234375
],
[
0.40478515625,
0.35693359375,
0.443359375,
0.35693359375,
0.443359375,
0.369384765625,
0.40478515625,
0.369384765625
],
[
0.1292724609375,
0.355712890625,
0.2939453125,
0.355712890625,
0.2939453125,
0.37060546875,
0.1292724609375,
0.37060546875
],
[
0.470947265625,
0.35693359375,
0.5087890625,
0.35693359375,
0.5087890625,
0.369384765625,
0.470947265625,
0.369384765625
],
[
0.623046875,
0.358154296875,
0.6650390625,
0.35986328125,
0.6630859375,
0.3759765625,
0.62109375,
0.37353515625
],
[
0.69287109375,
0.359375,
0.72314453125,
0.359375,
0.72314453125,
0.3759765625,
0.69287109375,
0.3759765625
],
[
0.744140625,
0.35986328125,
0.76904296875,
0.35986328125,
0.76904296875,
0.374755859375,
0.744140625,
0.374755859375
],
[
0.84814453125,
0.35986328125,
0.88671875,
0.35986328125,
0.88671875,
0.3759765625,
0.84814453125,
0.3759765625
],
[
0.1275634765625,
0.37060546875,
0.30908203125,
0.37060546875,
0.30908203125,
0.385498046875,
0.1275634765625,
0.385498046875
],
[
0.32568359375,
0.371826171875,
0.3779296875,
0.371826171875,
0.3779296875,
0.385986328125,
0.32568359375,
0.385986328125
],
[
0.399658203125,
0.371826171875,
0.444091796875,
0.371826171875,
0.444091796875,
0.38720703125,
0.399658203125,
0.38720703125
],
[
0.466796875,
0.372314453125,
0.5087890625,
0.372314453125,
0.5087890625,
0.385986328125,
0.466796875,
0.385986328125
],
[
0.69287109375,
0.3818359375,
0.72314453125,
0.3818359375,
0.72314453125,
0.3984375,
0.69287109375,
0.3984375
],
[
0.7421875,
0.3818359375,
0.771484375,
0.3818359375,
0.771484375,
0.3984375,
0.7421875,
0.3984375
],
[
0.84814453125,
0.38427734375,
0.884765625,
0.38427734375,
0.884765625,
0.396728515625,
0.84814453125,
0.396728515625
],
[
0.62451171875,
0.38427734375,
0.66162109375,
0.38427734375,
0.66162109375,
0.396728515625,
0.62451171875,
0.396728515625
],
[
0.32568359375,
0.385498046875,
0.378662109375,
0.38720703125,
0.3779296875,
0.402587890625,
0.32421875,
0.40087890625
],
[
0.45751953125,
0.38720703125,
0.509765625,
0.38720703125,
0.509765625,
0.402587890625,
0.45751953125,
0.402587890625
],
[
0.39208984375,
0.388427734375,
0.444091796875,
0.388427734375,
0.444091796875,
0.402587890625,
0.39208984375,
0.402587890625
],
[
0.1292724609375,
0.388427734375,
0.3046875,
0.388427734375,
0.3046875,
0.402587890625,
0.1292724609375,
0.402587890625
],
[
0.5380859375,
0.3896484375,
0.5859375,
0.391357421875,
0.58447265625,
0.40625,
0.53662109375,
0.40380859375
],
[
0.62451171875,
0.3984375,
0.6650390625,
0.3984375,
0.6650390625,
0.41455078125,
0.62451171875,
0.41455078125
],
[
0.7421875,
0.39794921875,
0.771484375,
0.39794921875,
0.771484375,
0.41455078125,
0.7421875,
0.41455078125
],
[
0.84814453125,
0.3984375,
0.88671875,
0.3984375,
0.88671875,
0.41455078125,
0.84814453125,
0.41455078125
],
[
0.20654296875,
0.421630859375,
0.489501953125,
0.421630859375,
0.489501953125,
0.435791015625,
0.20654296875,
0.435791015625
],
[
0.6943359375,
0.399658203125,
0.7197265625,
0.399658203125,
0.7197265625,
0.413330078125,
0.6943359375,
0.413330078125
],
[
0.12005615234375,
0.4228515625,
0.19140625,
0.4228515625,
0.19140625,
0.43408203125,
0.12005615234375,
0.43408203125
],
[
0.62451171875,
0.421630859375,
0.66162109375,
0.421630859375,
0.66162109375,
0.43701171875,
0.62451171875,
0.43701171875
],
[
0.6962890625,
0.4228515625,
0.72119140625,
0.4228515625,
0.72119140625,
0.435791015625,
0.6962890625,
0.435791015625
],
[
0.744140625,
0.4228515625,
0.77099609375,
0.4228515625,
0.77099609375,
0.435791015625,
0.744140625,
0.435791015625
],
[
0.84814453125,
0.4228515625,
0.884765625,
0.4228515625,
0.884765625,
0.435302734375,
0.84814453125,
0.435302734375
],
[
0.5263671875,
0.427001953125,
0.59765625,
0.4287109375,
0.59619140625,
0.44482421875,
0.52490234375,
0.4423828125
],
[
0.19482421875,
0.427001953125,
0.202392578125,
0.427001953125,
0.202392578125,
0.43115234375,
0.19482421875,
0.43115234375
],
[
0.62451171875,
0.43701171875,
0.6650390625,
0.43701171875,
0.6650390625,
0.453125,
0.62451171875,
0.453125
],
[
0.74072265625,
0.435791015625,
0.77099609375,
0.435791015625,
0.77099609375,
0.453125,
0.74072265625,
0.453125
],
[
0.84619140625,
0.43701171875,
0.88671875,
0.43701171875,
0.88671875,
0.453125,
0.84619140625,
0.453125
],
[
0.6943359375,
0.438232421875,
0.72314453125,
0.438232421875,
0.72314453125,
0.451904296875,
0.6943359375,
0.451904296875
],
[
0.118408203125,
0.438232421875,
0.1822509765625,
0.438232421875,
0.1822509765625,
0.453125,
0.118408203125,
0.453125
],
[
0.124267578125,
0.4755859375,
0.48779296875,
0.4755859375,
0.48779296875,
0.4892578125,
0.124267578125,
0.4892578125
],
[
0.5146484375,
0.4755859375,
0.8818359375,
0.4755859375,
0.8818359375,
0.4892578125,
0.5146484375,
0.4892578125
],
[
0.12005615234375,
0.4921875,
0.48779296875,
0.4921875,
0.48779296875,
0.505859375,
0.12005615234375,
0.505859375
],
[
0.51318359375,
0.4912109375,
0.88330078125,
0.490478515625,
0.88330078125,
0.5048828125,
0.51318359375,
0.505859375
],
[
0.1175537109375,
0.50732421875,
0.268798828125,
0.50830078125,
0.268798828125,
0.5224609375,
0.1175537109375,
0.521484375
],
[
0.51318359375,
0.50732421875,
0.8818359375,
0.505859375,
0.8818359375,
0.52099609375,
0.51318359375,
0.521484375
],
[
0.51318359375,
0.52490234375,
0.55517578125,
0.52490234375,
0.55517578125,
0.5380859375,
0.51318359375,
0.5380859375
],
[
0.12005615234375,
0.53369140625,
0.15283203125,
0.53369140625,
0.15283203125,
0.5458984375,
0.12005615234375,
0.5458984375
],
[
0.160400390625,
0.53759765625,
0.166259765625,
0.53759765625,
0.166259765625,
0.5419921875,
0.160400390625,
0.5419921875
],
[
0.162841796875,
0.53515625,
0.414794921875,
0.53515625,
0.414794921875,
0.5458984375,
0.162841796875,
0.5458984375
],
[
0.1259765625,
0.5546875,
0.48779296875,
0.5546875,
0.48779296875,
0.5693359375,
0.1259765625,
0.5693359375
],
[
0.15283203125,
0.5380859375,
0.160400390625,
0.5380859375,
0.160400390625,
0.5419921875,
0.15283203125,
0.5419921875
],
[
0.5146484375,
0.5634765625,
0.8818359375,
0.56396484375,
0.8818359375,
0.5791015625,
0.5146484375,
0.57763671875
],
[
0.12176513671875,
0.57177734375,
0.489501953125,
0.57177734375,
0.489501953125,
0.5849609375,
0.12176513671875,
0.5849609375
],
[
0.515625,
0.58056640625,
0.8818359375,
0.58056640625,
0.8818359375,
0.59423828125,
0.515625,
0.59423828125
],
[
0.12005615234375,
0.58740234375,
0.48779296875,
0.58740234375,
0.48779296875,
0.60107421875,
0.12005615234375,
0.60107421875
],
[
0.12005615234375,
0.6015625,
0.489501953125,
0.6025390625,
0.489501953125,
0.61767578125,
0.12005615234375,
0.6162109375
],
[
0.515625,
0.5966796875,
0.88330078125,
0.5966796875,
0.88330078125,
0.61083984375,
0.515625,
0.61083984375
],
[
0.118408203125,
0.6181640625,
0.484375,
0.6181640625,
0.484375,
0.6318359375,
0.118408203125,
0.6318359375
],
[
0.5146484375,
0.61328125,
0.8818359375,
0.61328125,
0.8818359375,
0.626953125,
0.5146484375,
0.626953125
],
[
0.515625,
0.62744140625,
0.873046875,
0.626953125,
0.873046875,
0.64111328125,
0.515625,
0.642578125
],
[
0.118408203125,
0.63525390625,
0.484375,
0.63525390625,
0.484375,
0.6484375,
0.118408203125,
0.6484375
],
[
0.5341796875,
0.6455078125,
0.66162109375,
0.6455078125,
0.66162109375,
0.6591796875,
0.5341796875,
0.6591796875
],
[
0.12005615234375,
0.65283203125,
0.210693359375,
0.65283203125,
0.210693359375,
0.66455078125,
0.12005615234375,
0.66455078125
],
[
0.53564453125,
0.662109375,
0.88330078125,
0.662109375,
0.88330078125,
0.67578125,
0.53564453125,
0.67578125
],
[
0.12005615234375,
0.677734375,
0.48779296875,
0.677734375,
0.48779296875,
0.69140625,
0.12005615234375,
0.69140625
],
[
0.5146484375,
0.6767578125,
0.88330078125,
0.677734375,
0.88330078125,
0.6923828125,
0.5146484375,
0.69140625
],
[
0.12005615234375,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.12005615234375,
0.7080078125
],
[
0.5146484375,
0.69482421875,
0.8818359375,
0.69482421875,
0.8818359375,
0.7080078125,
0.5146484375,
0.7080078125
],
[
0.12005615234375,
0.71142578125,
0.48779296875,
0.71142578125,
0.48779296875,
0.724609375,
0.12005615234375,
0.724609375
],
[
0.515625,
0.71044921875,
0.8818359375,
0.71044921875,
0.8818359375,
0.7236328125,
0.515625,
0.7236328125
],
[
0.1234130859375,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.74072265625,
0.1234130859375,
0.74072265625
],
[
0.5146484375,
0.724609375,
0.8818359375,
0.7255859375,
0.8818359375,
0.74072265625,
0.5146484375,
0.7392578125
],
[
0.118408203125,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.118408203125,
0.755859375
],
[
0.51318359375,
0.7412109375,
0.88330078125,
0.7412109375,
0.88330078125,
0.7548828125,
0.51318359375,
0.7548828125
],
[
0.12176513671875,
0.7587890625,
0.489501953125,
0.7587890625,
0.489501953125,
0.7724609375,
0.12176513671875,
0.7724609375
],
[
0.5146484375,
0.75830078125,
0.873046875,
0.75830078125,
0.873046875,
0.7724609375,
0.5146484375,
0.7724609375
],
[
0.12005615234375,
0.77392578125,
0.48779296875,
0.77490234375,
0.48779296875,
0.7890625,
0.12005615234375,
0.7880859375
],
[
0.5322265625,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.5322265625,
0.7890625
],
[
0.12176513671875,
0.79150390625,
0.489501953125,
0.79150390625,
0.489501953125,
0.8046875,
0.12176513671875,
0.8046875
],
[
0.51318359375,
0.79052734375,
0.8759765625,
0.79052734375,
0.8759765625,
0.80419921875,
0.51318359375,
0.80419921875
],
[
0.118408203125,
0.80712890625,
0.4130859375,
0.80712890625,
0.4130859375,
0.82080078125,
0.118408203125,
0.82080078125
],
[
0.51708984375,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.51708984375,
0.82080078125
],
[
0.51708984375,
0.822265625,
0.88330078125,
0.822265625,
0.88330078125,
0.83740234375,
0.51708984375,
0.83740234375
],
[
0.1275634765625,
0.83056640625,
0.486083984375,
0.83056640625,
0.486083984375,
0.84423828125,
0.1275634765625,
0.84423828125
],
[
0.1234130859375,
0.8466796875,
0.489501953125,
0.8466796875,
0.489501953125,
0.86083984375,
0.1234130859375,
0.86083984375
],
[
0.53076171875,
0.8466796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.8623046875,
0.53076171875,
0.86083984375
],
[
0.12005615234375,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.51318359375,
0.85986328125,
0.8818359375,
0.85986328125,
0.8818359375,
0.87353515625,
0.51318359375,
0.87353515625
],
[
0.51123046875,
0.87158203125,
0.88330078125,
0.87060546875,
0.88330078125,
0.88525390625,
0.51123046875,
0.8857421875
],
[
0.12005615234375,
0.87890625,
0.489501953125,
0.87890625,
0.489501953125,
0.892578125,
0.12005615234375,
0.892578125
],
[
0.51318359375,
0.88427734375,
0.884765625,
0.88427734375,
0.884765625,
0.89794921875,
0.51318359375,
0.89794921875
],
[
0.118408203125,
0.89453125,
0.486083984375,
0.89453125,
0.486083984375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.5146484375,
0.89794921875,
0.603515625,
0.89794921875,
0.603515625,
0.90771484375,
0.5146484375,
0.90771484375
]
] |
[
"#Unique",
"%Annotations",
"#Unique",
"%Annotations",
"OpenText",
"where selected",
"OpenText",
"whereselected",
"outraged; picky",
"2055",
"disgust",
"0.14",
"8740",
"Spir.growth",
"0.17",
"3205",
"become experienced",
"dismayed,disoriented",
"surprise",
"0.33",
"9676",
"0.22",
"Esteem",
"enraged;provoked",
"2016",
"anger",
"0.16",
"meetgoal;to looknice",
"2964",
"trust",
"0.25",
"Love",
"19768",
"touched: appreciated",
"2371",
"to support his friends",
"",
"sadness",
"excluded; bleak",
"0.23",
"3868 anticipation",
"be employed; stay dry",
"11645",
"Stability",
"0.49",
"future oriented; uptight",
"C",
"3136",
"joy",
"0.51",
"Phys.",
"4953",
"happier: jubiliation",
"rest more; food",
"0.13",
"2376",
"fear",
"0.2",
"Examples of open-text explanations that annotatorsprovided corresponding with",
"frozen infear;fatalistic",
"Figure 4:",
"the categories they selected. The bars on the right of the categories represent the percentage",
"of lines where annotators selected that category (out of those character-line pairs with positive",
"motivation/emotional reaction).",
"Fine-grained",
"% Agree w/",
"Label Type",
"PPA",
"KA",
"dev",
"test",
"Maj. Lbl",
"# annotated stories",
"train",
"10000",
"2500",
"2500",
"Dev",
".77",
".30",
"0.88",
"2.03",
"Maslow",
"2.02",
"# characters / story",
"1.82",
"Test",
".77",
".31",
"0.89",
"# char-lines w/ motiv",
"40154",
"8762",
"6831",
".91",
".24",
"0.95",
"Dev",
"76613",
"13785",
"14532",
"# char-lines w/ emot",
"Reiss",
"Test",
".24",
"0.95",
"Annotated data statistics for each",
".91",
"Table 1:",
"Dev",
".71",
".32",
"0.84",
"Plutchik",
"",
"Test",
".29",
"0.83",
".70",
"dataset",
"一位注释者记录说，他们会对表弟感到“烦恼”",
"Table 2: Agreement Statistics (PPA = Pair-",
"而感到slob，并选择了Plutchik类别以表示",
"wise percent agreement of worker responses",
"悲伤，厌恶和愤怒",
"per binary category, KA-Krippendorff's Al-",
"pha)",
"3.2",
"",
"Dataset Statistics and Insights",
"Cost与理论类别分配相对应的任务是管道",
"",
"平均每个类别的协议来比较注释器的性能。对",
"中最困难和最昂贵的任务（每个故事～$4）。",
"于Plutchik分数，我们将“中度关联”评级视",
"因此，我们只获得三分之一的注释故事的理论",
"类别标签，我们将其分配给开发和测试集。训",
"为与“高度关联”评级一致。一致性百分比和",
"练数据用缩短的管道进行注释，其中只有两个",
"Krippendorf的如表2所示。我们还计算单",
"个注释和大多数标签之间的一致性百分比。2",
"工人的动机和情绪反应的开放文本描述（每个",
"<NEXT LINE",
"故事～$1)",
"为了更清楚地了解此数据集中的一致性，我",
"Scale到目前为止，我们的数据集包括15.000",
"们为注释者对创建聚合混淆矩阵。首先，我们",
"个故事（从ROC故事训练集中随机选择）中",
"对所有配对注释之间的答案组合计数进行求",
"的动机和情感的总共300k低级注释。。它涵盖",
"和（不包括none标签）。如果注释者选择了多",
"了超过150,000个字符行对，其中56k个字符",
"个类别，我们将在所选类别之间均匀地拆分计",
"行对具有带注释的动机，而105k个字符行对",
"数。我们在总混淆矩阵上计算NPMI。在图5",
"具有带注释的情感变化（i。e。none以外的标",
"中，我们显示了动机类别的NPMI混淆矩阵",
"签）。表1显示了跨训练、开发和测试拆分的",
"在动机注释中，我们在对角线上找到最高分",
"细分。图4显示了在有积极变化的情况下，为",
"(i。e。，Reiss协议），大多数混淆发生在同",
"动机和情感类别选择不同标签的频率",
"马斯洛类别的Reiss动机之间（图E5中的黑",
"色轮廓）。其他分歧通常涉及主题相似的Reiss",
"Agreements为了进行质量控制，我们删除",
"了始终生产低质量工作的工人，如附录中所述。",
"2Majority label for the motivation categories is what",
"在分类集（Maslow，Reiss和Plutchik）中，我",
"was agreed upon by at least two annotators per cate-",
"gory. For emotion categories, we averaged the point-",
"们通过将每个单独的类别视为二进制标签（如",
"wise ratings and counted a category if the average rat-",
"果他们在其响应集中包含该类别，则为1）并",
"ing was > 2 ."
] |
{
"arxiv_id": "1805.06533",
"language": "zh",
"page_idx": 3
}
|
||
1805.06533-zh-4
|
1805.06533-zh-4.jpg
|
Annotator
$NPMI$
<FigureHere>
Figure 5: NPMI confusion matrix on motivational categories for all annotator pairs with color scaling for legibility. The highest values are generally along diagonal or within Maslow categories (outlined in black). We highlight a few common points of disagreement between thematically similar categories.
<FigureHere>
及可选的前面的上下文线) 和一个角色，并预测动机(或情绪反应)。为每个马斯洛需求，赖斯动机或 Plutchik 类别预测一个二进制标签。Annotation Classification 由于开发集和测试集包含成对的分类标签和自由文本解释，因此我们提出了三个任务，其中模型必须在给出情绪反应或动机解释的情况下预测正确的Maslow/ Reiss/ Plutchik 标 签 。
Explanation Generation 最后，我们可以使用自由文本解释来训练模型来描述自由文本中角色的心理状态(图$\overline{\mathrm{E}}$4 中的示例)。这些解释允许两个条件生成任务，其中模型必须生成描述角色的情绪反应或动机的单词
General model architectures for
Figure 6:
three new task types
### 5 Baseline Models
子类别，例如宁静(精神放松)和休息(身体放松)。。我们在附录中提供了 Plutchik 类别的分析，发现对角线上的高分，通常在“积极情绪”集群(喜悦，信任)或“消极情绪”集群(愤怒，厌恶，悲伤)中的类别之间发生分歧。
## 4 Tasks
三个任务的通用模型体系结构如图 6 所示。我们在下面描述每个模型组件。状态分类和解释生成模型可以单独训练，也可以在多任务设置中训练
在状态分类和解释生成任务中，将从故事$x^\mathrm{s}$ 中给出一个模型，其中包含来自词汇表$V$ 的$N$单词$\{w_0^s,w_1^s,\ldots,w_N^s\}$ ,该故事$e_j\in E$中的一个字符，其中$E$是故事中的字符集，以及(可选)故事$V$中的前面的。角色心理状态的表示形式编码为：
此新数据集中的注释所涵盖的多种模式允许探索多个新任务。我们在下面概述了三种任务类型，总共涵盖了要评估的八个任务。图 6 总结了任务类型输入和输出之间的差异
State Classification 三个主要任务涉及对每个标签集(Maslow,Reiss,Plutchik)的故事角色的心理状态进行分类，这些标签集是为我们的数据集的开发和测试拆分而收集的。在每个分类任务中，模型都会给出一个故事线(以
$$h^e=\text{Encoder}(x^s,C[e_j])$$
(1)
,其中$C[e_j]$对应于$C$中出现$e_j$的句子的连
接子集
|
[
[
0.1956787109375,
0.114013671875,
0.2032470703125,
0.114013671875,
0.2032470703125,
0.1181640625,
0.1956787109375,
0.1181640625
],
[
0.81689453125,
0.114013671875,
0.85400390625,
0.114013671875,
0.85400390625,
0.1258544921875,
0.81689453125,
0.1258544921875
],
[
0.68994140625,
0.1181640625,
0.72119140625,
0.1181640625,
0.72119140625,
0.1270751953125,
0.68994140625,
0.1270751953125
],
[
0.187255859375,
0.1181640625,
0.22412109375,
0.1181640625,
0.22412109375,
0.17578125,
0.187255859375,
0.17578125
],
[
0.73291015625,
0.1181640625,
0.77734375,
0.1181640625,
0.77734375,
0.12890625,
0.73291015625,
0.12890625
],
[
0.2418212890625,
0.1258544921875,
0.316650390625,
0.1258544921875,
0.316650390625,
0.133056640625,
0.2418212890625,
0.133056640625
],
[
0.450927734375,
0.1258544921875,
0.56689453125,
0.1258544921875,
0.56689453125,
0.133056640625,
0.450927734375,
0.133056640625
],
[
0.8017578125,
0.1270751953125,
0.82861328125,
0.1270751953125,
0.82861328125,
0.1395263671875,
0.8017578125,
0.1395263671875
],
[
0.340087890625,
0.127685546875,
0.37451171875,
0.127685546875,
0.37451171875,
0.1312255859375,
0.340087890625,
0.1312255859375
],
[
0.60693359375,
0.127685546875,
0.67822265625,
0.127685546875,
0.67822265625,
0.1318359375,
0.60693359375,
0.1318359375
],
[
0.69287109375,
0.1300048828125,
0.77734375,
0.1312255859375,
0.77734375,
0.142578125,
0.69287109375,
0.141357421875
],
[
0.2301025390625,
0.133056640625,
0.2376708984375,
0.133056640625,
0.2376708984375,
0.1395263671875,
0.2301025390625,
0.1395263671875
],
[
0.156982421875,
0.1217041015625,
0.177978515625,
0.12353515625,
0.17626953125,
0.1341552734375,
0.154541015625,
0.1312255859375
],
[
0.5322265625,
0.1353759765625,
0.62451171875,
0.1353759765625,
0.62451171875,
0.1383056640625,
0.5322265625,
0.1383056640625
],
[
0.57080078125,
0.1484375,
0.62451171875,
0.1484375,
0.62451171875,
0.1519775390625,
0.57080078125,
0.1519775390625
],
[
0.7109375,
0.149658203125,
0.76171875,
0.149658203125,
0.76171875,
0.1602783203125,
0.7109375,
0.1602783203125
],
[
0.474365234375,
0.15673828125,
0.55908203125,
0.15673828125,
0.55908203125,
0.1602783203125,
0.474365234375,
0.1602783203125
],
[
0.62890625,
0.1561279296875,
0.67822265625,
0.1561279296875,
0.67822265625,
0.1602783203125,
0.62890625,
0.1602783203125
],
[
0.71533203125,
0.15673828125,
0.75732421875,
0.1591796875,
0.75732421875,
0.166259765625,
0.71533203125,
0.1644287109375
],
[
0.491943359375,
0.1644287109375,
0.56494140625,
0.1644287109375,
0.56494140625,
0.168701171875,
0.491943359375,
0.168701171875
],
[
0.58251953125,
0.1644287109375,
0.60693359375,
0.1644287109375,
0.60693359375,
0.168701171875,
0.58251953125,
0.168701171875
],
[
0.6455078125,
0.1644287109375,
0.65576171875,
0.162109375,
0.6572265625,
0.166259765625,
0.646484375,
0.168701171875
],
[
0.19482421875,
0.169189453125,
0.21240234375,
0.169189453125,
0.21240234375,
0.1787109375,
0.19482421875,
0.1787109375
],
[
0.231689453125,
0.17041015625,
0.2376708984375,
0.17041015625,
0.2376708984375,
0.1787109375,
0.231689453125,
0.1787109375
],
[
0.58251953125,
0.1776123046875,
0.63037109375,
0.1776123046875,
0.63037109375,
0.18115234375,
0.58251953125,
0.18115234375
],
[
0.63037109375,
0.1776123046875,
0.66015625,
0.1776123046875,
0.66015625,
0.1817626953125,
0.63037109375,
0.1817626953125
],
[
0.145263671875,
0.179931640625,
0.15283203125,
0.179931640625,
0.15283203125,
0.18701171875,
0.145263671875,
0.18701171875
],
[
0.235107421875,
0.1602783203125,
0.2418212890625,
0.1650390625,
0.235107421875,
0.17041015625,
0.2275390625,
0.1650390625
],
[
0.2301025390625,
0.1817626953125,
0.26025390625,
0.1817626953125,
0.26025390625,
0.190673828125,
0.2301025390625,
0.190673828125
],
[
0.53076171875,
0.182861328125,
0.6826171875,
0.18408203125,
0.6826171875,
0.191162109375,
0.53076171875,
0.190673828125
],
[
0.14697265625,
0.191162109375,
0.151123046875,
0.191162109375,
0.151123046875,
0.196533203125,
0.14697265625,
0.196533203125
],
[
0.19482421875,
0.18408203125,
0.2032470703125,
0.18115234375,
0.208251953125,
0.18701171875,
0.198974609375,
0.190673828125
],
[
0.8076171875,
0.193603515625,
0.8251953125,
0.193603515625,
0.8251953125,
0.201904296875,
0.8076171875,
0.201904296875
],
[
0.193115234375,
0.20068359375,
0.2216796875,
0.2037353515625,
0.21826171875,
0.21435546875,
0.189697265625,
0.21142578125
],
[
0.28955078125,
0.207275390625,
0.318115234375,
0.207275390625,
0.318115234375,
0.210205078125,
0.28955078125,
0.210205078125
],
[
0.31982421875,
0.207275390625,
0.333251953125,
0.207275390625,
0.333251953125,
0.210205078125,
0.31982421875,
0.210205078125
],
[
0.453369140625,
0.2037353515625,
0.6826171875,
0.201904296875,
0.6826171875,
0.2120361328125,
0.453369140625,
0.21435546875
],
[
0.2301025390625,
0.2120361328125,
0.2435302734375,
0.2120361328125,
0.2435302734375,
0.2203369140625,
0.2301025390625,
0.2203369140625
],
[
0.611328125,
0.213134765625,
0.6826171875,
0.213134765625,
0.6826171875,
0.2203369140625,
0.611328125,
0.2203369140625
],
[
0.2476806640625,
0.21435546875,
0.26123046875,
0.21435546875,
0.26123046875,
0.218505859375,
0.2476806640625,
0.218505859375
],
[
0.26611328125,
0.21435546875,
0.4072265625,
0.21435546875,
0.4072265625,
0.218505859375,
0.26611328125,
0.218505859375
],
[
0.474365234375,
0.21435546875,
0.50537109375,
0.21435546875,
0.50537109375,
0.218505859375,
0.474365234375,
0.218505859375
],
[
0.68603515625,
0.22802734375,
0.77490234375,
0.22802734375,
0.77490234375,
0.2381591796875,
0.68603515625,
0.2381591796875
],
[
0.2275390625,
0.23095703125,
0.5263671875,
0.23095703125,
0.5263671875,
0.2410888671875,
0.2275390625,
0.2410888671875
],
[
0.19482421875,
0.2078857421875,
0.2149658203125,
0.213134765625,
0.2091064453125,
0.2257080078125,
0.18896484375,
0.2203369140625
],
[
0.19482421875,
0.232177734375,
0.214111328125,
0.232177734375,
0.214111328125,
0.2410888671875,
0.19482421875,
0.2410888671875
],
[
0.5498046875,
0.2340087890625,
0.67919921875,
0.2340087890625,
0.67919921875,
0.2410888671875,
0.5498046875,
0.2410888671875
],
[
0.19482421875,
0.2410888671875,
0.21240234375,
0.2410888671875,
0.21240234375,
0.2493896484375,
0.19482421875,
0.2493896484375
],
[
0.2283935546875,
0.2410888671875,
0.2568359375,
0.2410888671875,
0.2568359375,
0.2493896484375,
0.2283935546875,
0.2493896484375
],
[
0.5517578125,
0.2423095703125,
0.60107421875,
0.2423095703125,
0.60107421875,
0.2493896484375,
0.5517578125,
0.2493896484375
],
[
0.611328125,
0.243408203125,
0.646484375,
0.243408203125,
0.646484375,
0.2476806640625,
0.611328125,
0.2476806640625
],
[
0.6484375,
0.243408203125,
0.6689453125,
0.2410888671875,
0.6708984375,
0.2476806640625,
0.64990234375,
0.2493896484375
],
[
0.6708984375,
0.243408203125,
0.6826171875,
0.243408203125,
0.6826171875,
0.2476806640625,
0.6708984375,
0.2476806640625
],
[
0.68701171875,
0.2381591796875,
0.7734375,
0.2381591796875,
0.7734375,
0.2452392578125,
0.68701171875,
0.2452392578125
],
[
0.31982421875,
0.24462890625,
0.333251953125,
0.24462890625,
0.333251953125,
0.2476806640625,
0.31982421875,
0.2476806640625
],
[
0.43994140625,
0.24462890625,
0.5087890625,
0.24462890625,
0.5087890625,
0.248779296875,
0.43994140625,
0.248779296875
],
[
0.187255859375,
0.254150390625,
0.202392578125,
0.243408203125,
0.2149658203125,
0.2529296875,
0.20068359375,
0.26318359375
],
[
0.5322265625,
0.254150390625,
0.62646484375,
0.254150390625,
0.62646484375,
0.26123046875,
0.5322265625,
0.26123046875
],
[
0.80615234375,
0.2493896484375,
0.82861328125,
0.2493896484375,
0.82861328125,
0.25830078125,
0.80615234375,
0.25830078125
],
[
0.25439453125,
0.257080078125,
0.279541015625,
0.257080078125,
0.279541015625,
0.260009765625,
0.25439453125,
0.260009765625
],
[
0.3046875,
0.257080078125,
0.31494140625,
0.257080078125,
0.31494140625,
0.260009765625,
0.3046875,
0.260009765625
],
[
0.344970703125,
0.257080078125,
0.37451171875,
0.257080078125,
0.37451171875,
0.260009765625,
0.344970703125,
0.260009765625
],
[
0.37451171875,
0.257080078125,
0.40478515625,
0.257080078125,
0.40478515625,
0.260009765625,
0.37451171875,
0.260009765625
],
[
0.424072265625,
0.257080078125,
0.50537109375,
0.257080078125,
0.50537109375,
0.260009765625,
0.424072265625,
0.260009765625
],
[
0.2283935546875,
0.262451171875,
0.68603515625,
0.262451171875,
0.68603515625,
0.2724609375,
0.2283935546875,
0.2724609375
],
[
0.19482421875,
0.262451171875,
0.2091064453125,
0.262451171875,
0.2091064453125,
0.270751953125,
0.19482421875,
0.270751953125
],
[
0.156982421875,
0.268310546875,
0.210693359375,
0.268310546875,
0.210693359375,
0.27978515625,
0.156982421875,
0.27978515625
],
[
0.2275390625,
0.270751953125,
0.29296875,
0.270751953125,
0.29296875,
0.27783203125,
0.2275390625,
0.27783203125
],
[
0.300537109375,
0.270751953125,
0.6826171875,
0.270751953125,
0.6826171875,
0.27783203125,
0.300537109375,
0.27783203125
],
[
0.273681640625,
0.279052734375,
0.38623046875,
0.279052734375,
0.38623046875,
0.283203125,
0.273681640625,
0.283203125
],
[
0.6650390625,
0.279052734375,
0.68115234375,
0.279052734375,
0.68115234375,
0.283203125,
0.6650390625,
0.283203125
],
[
0.81201171875,
0.275634765625,
0.82861328125,
0.275634765625,
0.82861328125,
0.283203125,
0.81201171875,
0.283203125
],
[
0.450927734375,
0.27978515625,
0.50537109375,
0.27978515625,
0.50537109375,
0.283203125,
0.450927734375,
0.283203125
],
[
0.2301025390625,
0.27978515625,
0.273681640625,
0.27978515625,
0.273681640625,
0.283203125,
0.2301025390625,
0.283203125
],
[
0.53076171875,
0.27978515625,
0.56494140625,
0.27978515625,
0.56494140625,
0.283203125,
0.53076171875,
0.283203125
],
[
0.6533203125,
0.27978515625,
0.6650390625,
0.27978515625,
0.6650390625,
0.283203125,
0.6533203125,
0.283203125
],
[
0.12005615234375,
0.2998046875,
0.88330078125,
0.2998046875,
0.88330078125,
0.3134765625,
0.12005615234375,
0.3134765625
],
[
0.118408203125,
0.3154296875,
0.88330078125,
0.3154296875,
0.88330078125,
0.328857421875,
0.118408203125,
0.328857421875
],
[
0.118408203125,
0.330078125,
0.880859375,
0.330078125,
0.880859375,
0.34326171875,
0.118408203125,
0.34326171875
],
[
0.118408203125,
0.3486328125,
0.26611328125,
0.3486328125,
0.26611328125,
0.3623046875,
0.118408203125,
0.3623046875
],
[
0.185546875,
0.38427734375,
0.2225341796875,
0.38427734375,
0.2225341796875,
0.3955078125,
0.185546875,
0.3955078125
],
[
0.274658203125,
0.38427734375,
0.34765625,
0.38427734375,
0.34765625,
0.394287109375,
0.274658203125,
0.394287109375
],
[
0.39794921875,
0.385498046875,
0.468505859375,
0.385498046875,
0.468505859375,
0.3955078125,
0.39794921875,
0.3955078125
],
[
0.51708984375,
0.38720703125,
0.88330078125,
0.38720703125,
0.88330078125,
0.40087890625,
0.51708984375,
0.40087890625
],
[
0.160400390625,
0.393798828125,
0.246826171875,
0.394287109375,
0.246826171875,
0.40625,
0.160400390625,
0.405029296875
],
[
0.124267578125,
0.385986328125,
0.14111328125,
0.385986328125,
0.14111328125,
0.40380859375,
0.124267578125,
0.40380859375
],
[
0.27783203125,
0.3955078125,
0.34765625,
0.3955078125,
0.34765625,
0.40625,
0.27783203125,
0.40625
],
[
0.39208984375,
0.394287109375,
0.47607421875,
0.394287109375,
0.47607421875,
0.40625,
0.39208984375,
0.40625
],
[
0.515625,
0.402587890625,
0.8818359375,
0.402587890625,
0.8818359375,
0.416259765625,
0.515625,
0.416259765625
],
[
0.181396484375,
0.410400390625,
0.2166748046875,
0.407958984375,
0.21826171875,
0.42041015625,
0.1822509765625,
0.4228515625
],
[
0.411376953125,
0.4091796875,
0.44921875,
0.4091796875,
0.44921875,
0.421630859375,
0.411376953125,
0.421630859375
],
[
0.270263671875,
0.413330078125,
0.339111328125,
0.413330078125,
0.339111328125,
0.42333984375,
0.270263671875,
0.42333984375
],
[
0.124267578125,
0.410888671875,
0.14111328125,
0.410888671875,
0.14111328125,
0.43701171875,
0.124267578125,
0.43701171875
],
[
0.17626953125,
0.42041015625,
0.2283935546875,
0.42041015625,
0.2283935546875,
0.43115234375,
0.17626953125,
0.43115234375
],
[
0.40478515625,
0.42041015625,
0.45751953125,
0.42041015625,
0.45751953125,
0.43115234375,
0.40478515625,
0.43115234375
],
[
0.51318359375,
0.418701171875,
0.880859375,
0.418701171875,
0.880859375,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.1275634765625,
0.441162109375,
0.139404296875,
0.441162109375,
0.1376953125,
0.471435546875,
0.1259765625,
0.471435546875
],
[
0.280517578125,
0.4228515625,
0.341796875,
0.4228515625,
0.341796875,
0.432861328125,
0.280517578125,
0.432861328125
],
[
0.189697265625,
0.45068359375,
0.219970703125,
0.45361328125,
0.21826171875,
0.46142578125,
0.18896484375,
0.457763671875
],
[
0.29296875,
0.451904296875,
0.33154296875,
0.451904296875,
0.33154296875,
0.463134765625,
0.29296875,
0.463134765625
],
[
0.4189453125,
0.454833984375,
0.45166015625,
0.454833984375,
0.45166015625,
0.463134765625,
0.4189453125,
0.463134765625
],
[
0.5224609375,
0.460205078125,
0.880859375,
0.460205078125,
0.880859375,
0.473876953125,
0.5224609375,
0.473876953125
],
[
0.5146484375,
0.4755859375,
0.880859375,
0.4755859375,
0.880859375,
0.4892578125,
0.5146484375,
0.4892578125
],
[
0.187255859375,
0.482666015625,
0.333251953125,
0.482666015625,
0.333251953125,
0.496337890625,
0.187255859375,
0.496337890625
],
[
0.1259765625,
0.493408203125,
0.1435546875,
0.49462890625,
0.14111328125,
0.525390625,
0.1234130859375,
0.52490234375
],
[
0.18896484375,
0.495361328125,
0.26025390625,
0.49755859375,
0.26025390625,
0.5126953125,
0.18896484375,
0.51025390625
],
[
0.5146484375,
0.4921875,
0.880859375,
0.4921875,
0.880859375,
0.505859375,
0.5146484375,
0.505859375
],
[
0.41064453125,
0.49951171875,
0.455078125,
0.49951171875,
0.455078125,
0.51416015625,
0.41064453125,
0.51416015625
],
[
0.515625,
0.5087890625,
0.8818359375,
0.5087890625,
0.8818359375,
0.5224609375,
0.515625,
0.5224609375
],
[
0.2359619140625,
0.52099609375,
0.2822265625,
0.52099609375,
0.2822265625,
0.53515625,
0.2359619140625,
0.53515625
],
[
0.51318359375,
0.5224609375,
0.75244140625,
0.52392578125,
0.75244140625,
0.5380859375,
0.51318359375,
0.53759765625
],
[
0.124267578125,
0.548828125,
0.1427001953125,
0.548828125,
0.1427001953125,
0.56640625,
0.124267578125,
0.56640625
],
[
0.399658203125,
0.54736328125,
0.4609375,
0.54736328125,
0.4609375,
0.55859375,
0.399658203125,
0.55859375
],
[
0.5224609375,
0.5498046875,
0.8818359375,
0.5498046875,
0.8818359375,
0.5634765625,
0.5224609375,
0.5634765625
],
[
0.4013671875,
0.55712890625,
0.465087890625,
0.55712890625,
0.465087890625,
0.56787109375,
0.4013671875,
0.56787109375
],
[
0.18896484375,
0.55029296875,
0.330078125,
0.55029296875,
0.330078125,
0.56396484375,
0.18896484375,
0.56396484375
],
[
0.51708984375,
0.5654296875,
0.8818359375,
0.5654296875,
0.8818359375,
0.5791015625,
0.51708984375,
0.5791015625
],
[
0.5146484375,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59326171875,
0.5146484375,
0.59326171875
],
[
0.12005615234375,
0.58837890625,
0.197265625,
0.58837890625,
0.197265625,
0.599609375,
0.12005615234375,
0.599609375
],
[
0.2225341796875,
0.5859375,
0.489501953125,
0.5859375,
0.489501953125,
0.599609375,
0.2225341796875,
0.599609375
],
[
0.515625,
0.595703125,
0.88330078125,
0.5966796875,
0.88330078125,
0.61083984375,
0.515625,
0.60986328125
],
[
0.1175537109375,
0.6015625,
0.28955078125,
0.60400390625,
0.28955078125,
0.61767578125,
0.1175537109375,
0.615234375
],
[
0.5146484375,
0.61328125,
0.7734375,
0.61328125,
0.7734375,
0.626953125,
0.5146484375,
0.626953125
],
[
0.12005615234375,
0.64013671875,
0.486083984375,
0.64013671875,
0.486083984375,
0.65380859375,
0.12005615234375,
0.65380859375
],
[
0.5146484375,
0.646484375,
0.52880859375,
0.646484375,
0.52880859375,
0.65478515625,
0.5146484375,
0.65478515625
],
[
0.5263671875,
0.646484375,
0.54833984375,
0.646484375,
0.54833984375,
0.65478515625,
0.5263671875,
0.65478515625
],
[
0.5439453125,
0.6435546875,
0.69287109375,
0.6435546875,
0.69287109375,
0.65673828125,
0.5439453125,
0.65673828125
],
[
0.12005615234375,
0.65673828125,
0.48779296875,
0.65673828125,
0.48779296875,
0.67138671875,
0.12005615234375,
0.67138671875
],
[
0.12176513671875,
0.6748046875,
0.48779296875,
0.6748046875,
0.48779296875,
0.6884765625,
0.12176513671875,
0.6884765625
],
[
0.51318359375,
0.66943359375,
0.8818359375,
0.67041015625,
0.8818359375,
0.68505859375,
0.51318359375,
0.68408203125
],
[
0.12005615234375,
0.6904296875,
0.48779296875,
0.6904296875,
0.48779296875,
0.70361328125,
0.12005615234375,
0.70361328125
],
[
0.51318359375,
0.68505859375,
0.880859375,
0.68505859375,
0.880859375,
0.69970703125,
0.51318359375,
0.69970703125
],
[
0.118408203125,
0.7060546875,
0.484375,
0.7060546875,
0.484375,
0.7197265625,
0.118408203125,
0.7197265625
],
[
0.5146484375,
0.703125,
0.88330078125,
0.703125,
0.88330078125,
0.71728515625,
0.5146484375,
0.71728515625
],
[
0.515625,
0.71875,
0.59033203125,
0.71875,
0.59033203125,
0.73291015625,
0.515625,
0.73291015625
],
[
0.12005615234375,
0.7353515625,
0.1368408203125,
0.7353515625,
0.1368408203125,
0.74462890625,
0.12005615234375,
0.74462890625
],
[
0.15283203125,
0.73388671875,
0.208251953125,
0.73388671875,
0.208251953125,
0.7451171875,
0.15283203125,
0.7451171875
],
[
0.5341796875,
0.736328125,
0.8818359375,
0.736328125,
0.8818359375,
0.75,
0.5341796875,
0.75
],
[
0.5146484375,
0.75048828125,
0.88330078125,
0.751953125,
0.88330078125,
0.7666015625,
0.5146484375,
0.765625
],
[
0.14111328125,
0.73681640625,
0.1485595703125,
0.73681640625,
0.1485595703125,
0.7421875,
0.14111328125,
0.7421875
],
[
0.12176513671875,
0.75830078125,
0.48779296875,
0.75830078125,
0.48779296875,
0.771484375,
0.12176513671875,
0.771484375
],
[
0.12005615234375,
0.7724609375,
0.486083984375,
0.7724609375,
0.486083984375,
0.7861328125,
0.12005615234375,
0.7861328125
],
[
0.5146484375,
0.7685546875,
0.8818359375,
0.7685546875,
0.8818359375,
0.783203125,
0.5146484375,
0.783203125
],
[
0.51318359375,
0.783203125,
0.88330078125,
0.78369140625,
0.88330078125,
0.798828125,
0.51318359375,
0.79736328125
],
[
0.12005615234375,
0.79052734375,
0.48779296875,
0.79052734375,
0.48779296875,
0.80419921875,
0.12005615234375,
0.80419921875
],
[
0.118408203125,
0.8046875,
0.4130859375,
0.8056640625,
0.4130859375,
0.82080078125,
0.118408203125,
0.8193359375
],
[
0.5146484375,
0.7998046875,
0.88330078125,
0.80029296875,
0.88330078125,
0.8154296875,
0.5146484375,
0.81396484375
],
[
0.515625,
0.81787109375,
0.6708984375,
0.81787109375,
0.6708984375,
0.83056640625,
0.515625,
0.83056640625
],
[
0.124267578125,
0.830078125,
0.486083984375,
0.830078125,
0.486083984375,
0.84423828125,
0.124267578125,
0.84423828125
],
[
0.12005615234375,
0.845703125,
0.486083984375,
0.8466796875,
0.486083984375,
0.86083984375,
0.12005615234375,
0.85986328125
],
[
0.60107421875,
0.845703125,
0.7958984375,
0.845703125,
0.7958984375,
0.85986328125,
0.60107421875,
0.85986328125
],
[
0.85791015625,
0.845703125,
0.88330078125,
0.845703125,
0.88330078125,
0.85986328125,
0.85791015625,
0.85986328125
],
[
0.12005615234375,
0.86328125,
0.48779296875,
0.86328125,
0.48779296875,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.118408203125,
0.87890625,
0.484375,
0.87890625,
0.484375,
0.892578125,
0.118408203125,
0.892578125
],
[
0.52880859375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.52880859375,
0.8935546875
],
[
0.5205078125,
0.88525390625,
0.52880859375,
0.88525390625,
0.52880859375,
0.8896484375,
0.5205078125,
0.8896484375
],
[
0.118408203125,
0.8935546875,
0.489501953125,
0.89453125,
0.489501953125,
0.9091796875,
0.118408203125,
0.90771484375
],
[
0.51318359375,
0.89453125,
0.57080078125,
0.89453125,
0.57080078125,
0.9091796875,
0.51318359375,
0.9091796875
]
] |
[
"ol",
"NPMI",
"deall",
"I",
"vs.Honor",
"-0.010.180.100.1",
"00.150.06-0.04-0.10-0.070.0",
"0.68",
"0000",
"001007-014-0",
"Serenity vs.Rest",
"0.0",
"Sprowe",
"0070002000",
"nT0",
"Approval vs.",
"SOOR",
"0",
"Belonging",
"0.07-0.140.08-0.0",
"2-0.06",
"",
"Powe",
"",
"000",
"018.0",
"2",
"00",
"0.000.13",
"0.03-0.07-0.050.01 0.010.04-0.14-0.10",
"",
"o",
"0.34",
"Contact",
"1000",
"0",
"0.040.390.10.020.02-0.170.170.140.190.070.150.1",
"0.06",
"-0.06-0.05-0.110.12",
"0:02",
"001:0:01001:002007007-0000",
"0100",
"Health (Stability)vs.",
"0.03-0.04-0.06-0.02-0.13-0.08-0.05-0.06-0.09-0.10-0.05-0.07-0.18-0.20-0.10-0.0",
"Belan",
"Health",
"0.450.150.020.160.140.0900",
"Tranc",
"0.07-0.0",
"0.100.420.",
"-008.009",
"-0.18",
"O0",
"Physiological needs",
"",
"0.15-0.190.160.0",
"",
"0.09-0.14-0.05,0.090.45",
"0.10",
"nn",
"0",
"nnnn",
"on",
"0002000",
"1.00-0.02-0.03-0.02-0.010.130.02-0.02-0.140.01-0.140.05-0.16-0.04-1.00-0.11-1.000.120.07:0.09-0.01-1.000.040.10",
"Othe",
"PhysiologFood",
"0.03-0.11-0.10-0",
"-0.14-1.00-0.17-0.22-0.18-0.20-0.14-0.15-0.20-0.18-0.11-0.090.200.000.130.13-0.12-0.100.67-0.08",
"05015008003022013",
"05",
"<0",
"401601120",
"00013:0",
"m00",
"005",
"Figure 5: NPMI confusion matrix on motivational categories for all annotator pairs with color",
"scaling for legibility. The highest values are generally along diagonal or within Maslow categories",
"(outlined in black). We highlight a few common points of disagreement between thematically",
"similar categories.",
"State",
"Explanation",
"Annotation",
"及可选的前面的上下文线）和一个角色，并预",
"Classification",
"ysel",
"Generation",
"Classification",
"测动机（或情绪反应）。为每个马斯洛需求，赖",
"",
"",
"Open-text",
"ndino",
"Category",
"Category",
"斯动机或Plutchik类别预测一个二进制标签",
"upooec",
"Explanation",
"ogRes",
"LSTM",
"ogReg",
"AnnotationClassification由于开发集和测",
"试集包含成对的分类标签和自由文本解释，因",
"000000000",
"poou",
"i = fene(s char)",
"此我们提出了三个任务，其中模型必须在给",
"TFIDF",
"出情绪反应或动机解释的情况下预测正确的",
"Encoder",
"Maslow/Reiss/Plutchik标签。",
"",
"Open-text",
"ExplanationGeneration最后，我们可以使",
"Explanation",
"Character, Story contex, Lin",
"用自由文本解释来训练模型来描述自由文本中",
"角色的心理状态（图E4中的示例）。这些解释",
"Figure 6:",
"General model architectures for",
"允许两个条件生成任务，其中模型必须生成描",
"three new task types",
"述角色的情绪反应或动机的单词",
"子类别，例如宁静（精神放松）和休息（身体",
"5",
"",
" Baseline Models",
"放松)。。我们在附录中提供了Plutchik类别的",
"分析，发现对角线上的高分，通常在“积极情",
"三个任务的通用模型体系结构如图6所示。",
"绪”集群（喜悦，信任）或“消极情绪”集群",
"我们在下面描述每个模型组件。状态分类和解",
"（愤怒，厌恶，悲伤）中的类别之间发生分歧。",
"释生成模型可以单独训练，也可以在多任务设",
"置中训练",
"4.",
"Tasks",
"在状态分类和解释生成任务中，将从故事as",
"中给出一个模型，其中包含来自词汇表V的",
"",
"此新数据集中的注释所涵盖的多种模式允许探",
"索多个新任务。我们在下面概述了三种任务类",
"N单词[wo,wi,..,w}，该故事ejEE中",
"的一个字符，其中E是故事中的字符集，以",
"型，总共涵盖了要评估的八个任务。图6总",
"结了任务类型输和输出之间的差异",
"及（可选）故事V中的前面的。角色心理状态",
"的表示形式编码为：",
"StateClassification三个主要任务涉及对每",
"个标签集（Maslow，Reiss，Plutchik）的故事",
"he = Encoder(α',C[eil)",
"(1)",
"角色的心理状态进行分类，这些标签集是为我",
"们的数据集的开发和测试拆分而收集的。在每",
"其中C[e]对应于C中出现ei的句子的连",
"",
"个分类任务中，模型都会给出一个故事线（以",
"接子集"
] |
{
"arxiv_id": "1805.06533",
"language": "zh",
"page_idx": 4
}
|
||
2210.12186-zh-1
|
2210.12186-zh-1.jpg
|
通过语义奖励提高放射学报告生成的事实正确性
Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen,
Emily Tsai, Omar Almusa, Curtis P. Langlotz
## Stanford University jbdel@stanford.edu
<FigureHere>
Figure 1: Overview of our radiology report generation pipeline. First, a neural network generates a radiology report given a chest X-ray image. We then leverage RadGraph to create semantic annotations of the output used to design reinforcement learning rewards.
## $\mathbf{Abstract}$
Neural image-to-text radiology report generation systems offer the potential to improve radiology reporting by reducing the repetitive process of report drafting and identifying possible medical errors. These systems have achieved promising performance as measured by widely used NLG metrics such as BLEU and CIDEr. However, the current systems face important limitations. First, they present an increased complexity in architecture that offers only marginal improvements on NLG metrics. Secondly, these systems that achieve high performance on these metrics are not always factually complete or consistent due to both inadequate training and evaluation. Recent studies have shown the systems can be substantially improved by using new methods encouraging 1) the generation of domain entities consistent with the reference and 2) describing these entities in inferentially consistent ways. So far, these methods rely on weakly-supervised approaches (rule-based) and named entity recognition systems that are not specific to the chest X-ray domain. To overcome this limitation, we propose a new method, the RadGraph reward, to further improve the factual completeness and correctness of generated radiology reports. More precisely, we leverage the RadGraph dataset containing annotated chest X-ray reports with entities and relations between enti- ${\hat{\text{ties. On two open radiology report datasets,}}}$ our system substantially improves the scores up to 14.2 % and 25.3 % on metrics evaluating the factual correctness and completeness of reports.
## 1 Introduction
自然语言生成(NLG)的一个重要医学应用是构建辅助系统，该系统可以拍摄患者的 X 射线图像并生成描述图像中临床观察结果的文本报告$\overline{\mathrm{E}}($Jing et al., 2018; Li et al., 2018; Chen et al., 2020; Miura et al., 2021)。这是一项临床上重要的任务，有可能减少放射科医生的重复性工
作，并普遍改善临床沟通 (Kahn Jr et al.,2009) 。最近，人们对新的体系结构$\operatorname{E}($Chen et al., 2020, 2021; Alfarghaly et al., 2021)以及如何将数据结构输人系统$\overline{\mathrm{E}}($Liu et al.,2021a) 给予了很多关注。通过广泛使用的 NLG 指标(如 BLEU$\boxed{\mathrm{F}}$(Papineni et al.,2002)和 CIDEr$\boxed{\mathrm{E}}$ (Vedantam et al.,2015)) 衡量，这些系统取得了有希望的性能。然而，这些研究面临着重要的局限性。首先，它们增加了架构的复杂性，只提供了对 NLG 指标的微小改进。其次， 由于对这些系统的培训和评估不足，这些在NLG 指标上实现高性能的系统并不总是事实完整或一致的。$\overline{\mathrm{E}}$Miura et al.(2021)表明，现有系统在事实完整性和一致性方面不足，并且可以通过用“面向事实”的方法代替广泛使用的 NLG 指标来大大改善图像到文本放射学报告生成(RRG)系统，鼓励 1)生成与参考一致的域实体，2)以推理一致的方式描述这些实体。到目前为止，这些新方法依赖于弱监督方法(基于规则)来构建用于放射学报告和生物医学命名实体识别系统$\overline{\mathrm{E}}($Zhang et al., 2021)
|
[
[
0.246826171875,
0.09320068359375,
0.75146484375,
0.09320068359375,
0.75146484375,
0.11102294921875,
0.246826171875,
0.11102294921875
],
[
0.2476806640625,
0.14013671875,
0.75390625,
0.14013671875,
0.75390625,
0.15380859375,
0.2476806640625,
0.15380859375
],
[
0.316650390625,
0.15673828125,
0.68603515625,
0.15673828125,
0.68603515625,
0.17041015625,
0.316650390625,
0.17041015625
],
[
0.420654296875,
0.17333984375,
0.5810546875,
0.17333984375,
0.5810546875,
0.18701171875,
0.420654296875,
0.18701171875
],
[
0.3955078125,
0.190673828125,
0.611328125,
0.190673828125,
0.611328125,
0.2037353515625,
0.3955078125,
0.2037353515625
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.6650390625,
0.271484375,
0.6826171875,
0.271484375,
0.6826171875,
0.29638671875,
0.6650390625,
0.29638671875
],
[
0.145263671875,
0.283203125,
0.45751953125,
0.283203125,
0.45751953125,
0.29638671875,
0.145263671875,
0.29638671875
],
[
0.580078125,
0.28515625,
0.59619140625,
0.279052734375,
0.603515625,
0.2880859375,
0.587890625,
0.29443359375
],
[
0.14697265625,
0.2998046875,
0.45751953125,
0.2998046875,
0.45751953125,
0.31005859375,
0.14697265625,
0.31005859375
],
[
0.145263671875,
0.3134765625,
0.459228515625,
0.3134765625,
0.459228515625,
0.32666015625,
0.145263671875,
0.32666015625
],
[
0.6689453125,
0.320556640625,
0.7138671875,
0.320556640625,
0.7138671875,
0.3349609375,
0.6689453125,
0.3349609375
],
[
0.145263671875,
0.32666015625,
0.459228515625,
0.32666015625,
0.459228515625,
0.34033203125,
0.145263671875,
0.34033203125
],
[
0.6884765625,
0.339111328125,
0.6962890625,
0.339111328125,
0.6962890625,
0.345703125,
0.6884765625,
0.345703125
],
[
0.75244140625,
0.34033203125,
0.77685546875,
0.34033203125,
0.77685546875,
0.3486328125,
0.75244140625,
0.3486328125
],
[
0.298828125,
0.342529296875,
0.45751953125,
0.342529296875,
0.45751953125,
0.352783203125,
0.298828125,
0.352783203125
],
[
0.14697265625,
0.341552734375,
0.301513671875,
0.341552734375,
0.301513671875,
0.3515625,
0.14697265625,
0.3515625
],
[
0.145263671875,
0.355224609375,
0.459228515625,
0.355224609375,
0.459228515625,
0.3681640625,
0.145263671875,
0.3681640625
],
[
0.56494140625,
0.363525390625,
0.58837890625,
0.363525390625,
0.58837890625,
0.36767578125,
0.56494140625,
0.36767578125
],
[
0.14697265625,
0.37060546875,
0.456787109375,
0.37060546875,
0.456787109375,
0.38134765625,
0.14697265625,
0.38134765625
],
[
0.8251953125,
0.365234375,
0.83642578125,
0.365234375,
0.83642578125,
0.372314453125,
0.8251953125,
0.372314453125
],
[
0.5146484375,
0.3759765625,
0.63232421875,
0.3759765625,
0.63232421875,
0.385986328125,
0.5146484375,
0.385986328125
],
[
0.62744140625,
0.377685546875,
0.68603515625,
0.377685546875,
0.68603515625,
0.385498046875,
0.62744140625,
0.385498046875
],
[
0.69287109375,
0.377685546875,
0.87744140625,
0.377685546875,
0.87744140625,
0.385498046875,
0.69287109375,
0.385498046875
],
[
0.6513671875,
0.38720703125,
0.68603515625,
0.38720703125,
0.68603515625,
0.394287109375,
0.6513671875,
0.394287109375
],
[
0.145263671875,
0.39794921875,
0.459228515625,
0.39794921875,
0.459228515625,
0.410888671875,
0.145263671875,
0.410888671875
],
[
0.1435546875,
0.3818359375,
0.45751953125,
0.383056640625,
0.45751953125,
0.396728515625,
0.1435546875,
0.3955078125
],
[
0.51123046875,
0.40625,
0.6767578125,
0.40625,
0.6767578125,
0.419189453125,
0.51123046875,
0.419189453125
],
[
0.6708984375,
0.4091796875,
0.74462890625,
0.4091796875,
0.74462890625,
0.416259765625,
0.6708984375,
0.416259765625
],
[
0.74072265625,
0.407958984375,
0.861328125,
0.407958984375,
0.861328125,
0.418701171875,
0.74072265625,
0.418701171875
],
[
0.145263671875,
0.42578125,
0.456787109375,
0.42578125,
0.456787109375,
0.439453125,
0.145263671875,
0.439453125
],
[
0.145263671875,
0.413330078125,
0.45751953125,
0.413330078125,
0.45751953125,
0.42333984375,
0.145263671875,
0.42333984375
],
[
0.145263671875,
0.440673828125,
0.459228515625,
0.440673828125,
0.459228515625,
0.45361328125,
0.145263671875,
0.45361328125
],
[
0.64990234375,
0.435791015625,
0.7314453125,
0.435791015625,
0.7314453125,
0.44775390625,
0.64990234375,
0.44775390625
],
[
0.14697265625,
0.4560546875,
0.45751953125,
0.4560546875,
0.45751953125,
0.466064453125,
0.14697265625,
0.466064453125
],
[
0.51318359375,
0.4619140625,
0.8818359375,
0.4619140625,
0.8818359375,
0.4755859375,
0.51318359375,
0.4755859375
],
[
0.145263671875,
0.468505859375,
0.459228515625,
0.468505859375,
0.459228515625,
0.482177734375,
0.145263671875,
0.482177734375
],
[
0.51123046875,
0.476806640625,
0.87890625,
0.476806640625,
0.87890625,
0.490478515625,
0.51123046875,
0.490478515625
],
[
0.1435546875,
0.482666015625,
0.45751953125,
0.482177734375,
0.45751953125,
0.495361328125,
0.1435546875,
0.496337890625
],
[
0.509765625,
0.490478515625,
0.87890625,
0.490478515625,
0.87890625,
0.50439453125,
0.509765625,
0.50439453125
],
[
0.145263671875,
0.49755859375,
0.45751953125,
0.49755859375,
0.45751953125,
0.51123046875,
0.145263671875,
0.51123046875
],
[
0.145263671875,
0.51123046875,
0.459228515625,
0.5126953125,
0.459228515625,
0.525390625,
0.145263671875,
0.52490234375
],
[
0.51318359375,
0.505859375,
0.880859375,
0.505859375,
0.880859375,
0.5166015625,
0.51318359375,
0.5166015625
],
[
0.51123046875,
0.5185546875,
0.8271484375,
0.5185546875,
0.8271484375,
0.5322265625,
0.51123046875,
0.5322265625
],
[
0.145263671875,
0.525390625,
0.459228515625,
0.525390625,
0.459228515625,
0.5390625,
0.145263671875,
0.5390625
],
[
0.145263671875,
0.5390625,
0.459228515625,
0.5390625,
0.459228515625,
0.552734375,
0.145263671875,
0.552734375
],
[
0.14697265625,
0.55419921875,
0.45751953125,
0.55419921875,
0.45751953125,
0.56787109375,
0.14697265625,
0.56787109375
],
[
0.145263671875,
0.56787109375,
0.459228515625,
0.56787109375,
0.459228515625,
0.58056640625,
0.145263671875,
0.58056640625
],
[
0.5146484375,
0.56396484375,
0.88330078125,
0.56396484375,
0.88330078125,
0.57763671875,
0.5146484375,
0.57763671875
],
[
0.145263671875,
0.58203125,
0.459228515625,
0.58203125,
0.459228515625,
0.595703125,
0.145263671875,
0.595703125
],
[
0.145263671875,
0.5966796875,
0.459228515625,
0.5966796875,
0.459228515625,
0.60986328125,
0.145263671875,
0.60986328125
],
[
0.51318359375,
0.5986328125,
0.880859375,
0.5986328125,
0.880859375,
0.6123046875,
0.51318359375,
0.6123046875
],
[
0.145263671875,
0.60986328125,
0.45751953125,
0.60986328125,
0.45751953125,
0.62353515625,
0.145263671875,
0.62353515625
],
[
0.51123046875,
0.61328125,
0.87890625,
0.61328125,
0.87890625,
0.626953125,
0.51123046875,
0.626953125
],
[
0.145263671875,
0.62353515625,
0.455078125,
0.62353515625,
0.455078125,
0.63720703125,
0.145263671875,
0.63720703125
],
[
0.51123046875,
0.6298828125,
0.87890625,
0.6298828125,
0.87890625,
0.6435546875,
0.51123046875,
0.6435546875
],
[
0.145263671875,
0.6396484375,
0.45751953125,
0.6396484375,
0.45751953125,
0.65283203125,
0.145263671875,
0.65283203125
],
[
0.51318359375,
0.6455078125,
0.87890625,
0.6455078125,
0.87890625,
0.66015625,
0.51318359375,
0.66015625
],
[
0.1485595703125,
0.65478515625,
0.456787109375,
0.65478515625,
0.456787109375,
0.6650390625,
0.1485595703125,
0.6650390625
],
[
0.145263671875,
0.66650390625,
0.45751953125,
0.66650390625,
0.45751953125,
0.6796875,
0.145263671875,
0.6796875
],
[
0.51123046875,
0.6630859375,
0.8818359375,
0.6630859375,
0.8818359375,
0.6767578125,
0.51123046875,
0.6767578125
],
[
0.145263671875,
0.681640625,
0.45751953125,
0.681640625,
0.45751953125,
0.6953125,
0.145263671875,
0.6953125
],
[
0.509765625,
0.677734375,
0.880859375,
0.6787109375,
0.880859375,
0.693359375,
0.509765625,
0.6923828125
],
[
0.145263671875,
0.69482421875,
0.45751953125,
0.69482421875,
0.45751953125,
0.7080078125,
0.145263671875,
0.7080078125
],
[
0.515625,
0.693359375,
0.87744140625,
0.693359375,
0.87744140625,
0.7080078125,
0.515625,
0.7080078125
],
[
0.145263671875,
0.71044921875,
0.45751953125,
0.71044921875,
0.45751953125,
0.7236328125,
0.145263671875,
0.7236328125
],
[
0.51318359375,
0.71044921875,
0.87890625,
0.71044921875,
0.87890625,
0.7236328125,
0.51318359375,
0.7236328125
],
[
0.145263671875,
0.7236328125,
0.45751953125,
0.7236328125,
0.45751953125,
0.73681640625,
0.145263671875,
0.73681640625
],
[
0.51318359375,
0.7255859375,
0.87744140625,
0.72705078125,
0.87744140625,
0.7412109375,
0.51318359375,
0.74072265625
],
[
0.145263671875,
0.73828125,
0.45751953125,
0.73828125,
0.45751953125,
0.751953125,
0.145263671875,
0.751953125
],
[
0.5146484375,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.5146484375,
0.75732421875
],
[
0.145263671875,
0.7529296875,
0.459228515625,
0.7529296875,
0.459228515625,
0.7666015625,
0.145263671875,
0.7666015625
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.1435546875,
0.7685546875,
0.187255859375,
0.7685546875,
0.187255859375,
0.77978515625,
0.1435546875,
0.77978515625
],
[
0.5146484375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.5146484375,
0.7890625
],
[
0.1485595703125,
0.7958984375,
0.264404296875,
0.79638671875,
0.264404296875,
0.81103515625,
0.1485595703125,
0.81005859375
],
[
0.5146484375,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.5146484375,
0.8056640625
],
[
0.118408203125,
0.798828125,
0.1318359375,
0.798828125,
0.1318359375,
0.80810546875,
0.118408203125,
0.80810546875
],
[
0.5146484375,
0.80810546875,
0.87890625,
0.80810546875,
0.87890625,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.12005615234375,
0.82373046875,
0.486083984375,
0.822265625,
0.486083984375,
0.83740234375,
0.12005615234375,
0.83837890625
],
[
0.51123046875,
0.82470703125,
0.87744140625,
0.82470703125,
0.87744140625,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.118408203125,
0.84033203125,
0.486083984375,
0.84033203125,
0.486083984375,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.51318359375,
0.84033203125,
0.86572265625,
0.84033203125,
0.86572265625,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.12005615234375,
0.85791015625,
0.483642578125,
0.85791015625,
0.483642578125,
0.8681640625,
0.12005615234375,
0.8681640625
],
[
0.51123046875,
0.85546875,
0.8818359375,
0.85498046875,
0.8818359375,
0.869140625,
0.51123046875,
0.87060546875
],
[
0.12005615234375,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.12005615234375,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.484375,
0.88818359375,
0.484375,
0.90185546875,
0.1175537109375,
0.90185546875
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"通过语义奖励提高放射学报告生成的事实正确性",
"Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen,",
"Emily Tsai, Omar Almusa, Curtis P. Langlotz",
"Stanford University",
"jbdel@stanford.edu",
"Abstract",
"CNN",
"Neural image-to-text radiology report genera-",
"福",
"tion systemsofferthepotentialtoimprovera-",
"diology reporting by reducing the repetitive",
"BERT",
"process of report drafting and identifying pos-",
"+",
"modirty",
"Thesesystems have",
"sible medical errors.",
"achieved promising performance as measured",
"modif",
"by widely used NLG metrics such as BLEU",
"",
"moderateoBs-opcardiome",
"egalyOBs.DP",
"mildOBs-DP fluidOBs-DPoverloadOBs-DP",
"ocatedat",
"important limitations. First, they present an in-",
"and CIDEr. However, the current systems face",
"but no overt B5-DApulmonary ANAT",
"T-opedemaoBs-",
"DA.nopneumoniaOBS-DA",
"only marginal improvements on NLG metrics.",
"creased complexity in architecturethat offers",
"Secondly, these systems that achieve high per-",
"RadGraph",
"formance onthese metrics arenot always fac-",
"Figure l: Overview of our radiology report generation",
"tually complete or consistent due to both inad-",
"pipeline. First, a neural network generates a radiology",
"equate training and evaluation. Recent studies",
"report given a chest X-ray image. We then leverage",
"have shown the systems can be substantially",
"improved by using new methods encouraging",
"RadGraphto create semantic annotations of the output",
"used to design reinforcement learning rewards.",
"I) the generation of domain entities consistent",
"with the reference and 2) describing these en-",
"tities in inferentially consistent ways. So far,",
"these methods rely on weakly-supervised ap-",
"作，并普遍改善临床沟通（KahnJretal.，2009）。",
"proaches (rule-based) and named entity recog-",
"nition systems that are not specific to the chest",
"最近，人们对新的体系结构E(Chen etal.,",
"X-ray domain. To overcome this limitation, we",
"2020,2021;Alfarghalyetal.,2021）以及如何",
"propose a new method, the RadGraph reward",
"将数据结构输入系统E(Liuetal.,2021la)给",
"to further improve the factual completeness",
"予了很多关注。通过广泛使用的NLG指标",
"andcorrectnessofgeneratedradiologyreports.",
"More precisely, we leverage the RadGraph",
"(如 BLEUE(Papineni et al.,2002)和 CIDErE",
"dataset containing annotated chest X-ray re-",
"（Vedantametal.2015））衡量，这些系统取得",
"ports with entities and relations between enti",
"了有希望的性能。然而，这些研究面临着重",
"ties. On two open radiology report datasets,",
"要的局限性。首先，它们增加了架构的复杂",
"our system substantially improves the scores",
"性，只提供了对NLG指标的微小改进。其次",
"up to 14.2 % and 25.3 % on metrics evaluating",
"由于对这些系统的培训和评估不足，这些在",
"the factual correctness and completeness of re-",
"NLG指标上实现高性能的系统并不总是事实",
"ports.",
"完整或一致的。EMiuraetal.（2021）表明，现",
"Introduction",
"有系统在事实完整性和一致性方面不足，并且",
"1",
"可以通过用“面向事实”的方法代替广泛使用",
"自然语言生成（NLG）的一个重要医学应用是",
"的NLG指标来大大改善图像到文本放射学报",
"构建辅助系统，该系统可以拍摄患者的X射线",
"告生成（RRG）系统，鼓励1）生成与参考",
"图像并生成描述图像中临床观察结果的文本报",
"致的域实体，2）以推理一致的方式描述这些",
"告E(Jing et al., 2018; Li et al., 2018; Chen et al.,",
"实体。到目前为止，这些新方法依赖于弱监督",
"2020Miuraetal.，2021）。这是一项临床上重",
"方法（基于规则）来构建用于放射学报告和生",
"要的任务，有可能减少放射科医生的重复性工",
"物医学命名实体识别系统E(Zhangetal.,2021)"
] |
{
"arxiv_id": "2210.12186",
"language": "zh",
"page_idx": 1
}
|
||
2210.12186-zh-2
|
2210.12186-zh-2.jpg
|
的 NLI 模型，这些模型不特定于胸部 X 射线。尽管这些“以事实为导向”的方法受到弱监督或仅限于通用生物医学实体，但它们的使用在广泛的指标和委员会认证的放射科医生的评估中显示出实质性的改进。这些发现促使我们提出一种新方法，以进一步提高生成的放射学报告的事实完整性和正确性。更准确地说，我们利用 RadGraph$\boxed{\mathrm{F}}$(Jain et al.,2021),这是一个由放射科医生注释的数据集，其中包含胸部X 射线放射学报告以及注释的实体和关系。这些注释允许我们创建两个语义图，一个用于生成的，一个用于参考报告。然后，我们引入三个简单的奖励，根据实体和关系对两个图之间的差异进行评分。这些奖励可以使用强化学习 (RL) 直接优化，以进一步提高我们系统生成的报告的质量。通过这样做，我们在两个流行的胸部 X 射线数据集上展示了我们的模型能够最大化定义的奖励，但也优于以前在各种NLG 和面向事实的指标上的工作
总之，我们的贡献是：
$\bullet$ We propose a simple RRG architecture that 1) is fast to train and suitable for a RL setup and 2) performs equally well as the previous and more complex architectures proposed in the literature.
$\cdot$ We leverage the RadGraph dataset and the associated fine-tuned model to design semanticbased rewards that qualitatively evaluate the factual correctness and completeness of the generated reports.
$\bullet$ We show on two datasets that directly opti-
mizing these rewards outperforms previous approaches that prioritize traditional NLG metrics.
该论文的结构如下：首先，我们描述了我们面向事实的基于图的奖励(§2 )。更准确地说，我们首先检查 RadGraph 数据集 (§E2.1)以及我们如何利用注释来创建奖励(§E2.2)。然后， 我们解释用于生成报告的模型(§$\mathbb{E}$3)的体系结构，以及如何使用负对数似然(NLL)和RL 对其进行训练。随后的部分专门介绍用于实验的数据集(§E4)和用于评估报告生成的度量$(\S$E5 )。后一部分分为两组：经典 NLG 度量$(\S$E5.1)和面向事实的度量(§$\overline{\mathrm{E}}$5.2)。最后， 我们给出了结果(§6),并以一节寻址相关作品(§7)结束本文。
2 Factually-oriented Graph-Based
# Reward
在本节中，我们提出了一种新的基于语义图的奖励，称为 RadGraph 奖励，在我们的实验中使用。我们首先在 2.1 节中解释 RadGraph 数据集，以及我们如何获得塑造我们奖励的注释。在 2.2 节中，我们解释了如何构造 RadGraph 奖励及其不同的变体
### 2.1 RadGraph
<FigureHere>
Figure 2: An example of a report annotated with enti-
ties and relations in the RadGraph dataset
RadGraph$\underline{\mathrm{E}}($Jain et al., 2021)是基于旨在构建放射学报告的新颖信息提取模式的全文胸部 X 射线放射学报告中的实体和关系的数据集。该数据集包含来自 MIMIC-CXR 数据集|E|(Johnson et al.,2019)的 500 份放射学报告的委员会认证放射科医生注释，总共对应于 14,579个实体和 10,889 个关系。此外，RadGraph 还包括 100 份放射学报告的测试数据集，分为两组独立的委员会认证放射科医生注释，分别来自 MIMIC-CXR 和 CheXpert$\underline{\mathrm{E}}($Smit et al., 2020)数据集(各50 份报告)
Entities 实体被定义为可以包括一个或多个相邻单词的文本的连续跨度。RadGraph 中的实体围绕两个概念：解剖学和观察。观察存在三个不确定性水平，导致四个不同的实体：解剖学(ANAT-DP),观察：明确存在(OBS-DP), 观察：不确定(OBS-U)和观察：明确不存在(OBS-DA)。解剖学是指在放射学报告中出现的解剖学身体部位，例如“肺”。观察是指与视觉特征，可识别的病理生理过程或诊断疾病分类相关的单词。例如，观察结果可以是“积液”或更一般的短语，如“增加”
Relations 关系定义为两个实体之间的有向边。存在三个层次：暗示(。,。),位于(。,。), 并修改(。,。)。(观察，观察)的建议是两个观察实体之间的关系，表明第二个观察的存在是从第一个观察的存在推断出的。位于(观察， 解剖学) 是观察实体和解剖学实体之间的关系，表示观察与解剖学相关。虽然位于通常指的是位置，它也可以用来描述观察和解剖学之
|
[
[
0.1175537109375,
0.08489990234375,
0.484375,
0.08612060546875,
0.484375,
0.1009521484375,
0.1175537109375,
0.09979248046875
],
[
0.5087890625,
0.08612060546875,
0.8271484375,
0.08612060546875,
0.8271484375,
0.09979248046875,
0.5087890625,
0.09979248046875
],
[
0.5439453125,
0.10394287109375,
0.61279296875,
0.10394287109375,
0.61279296875,
0.11517333984375,
0.5439453125,
0.11517333984375
],
[
0.118408203125,
0.12054443359375,
0.484375,
0.12054443359375,
0.484375,
0.1341552734375,
0.118408203125,
0.1341552734375
],
[
0.51318359375,
0.12890625,
0.87890625,
0.12890625,
0.87890625,
0.142578125,
0.51318359375,
0.142578125
],
[
0.1175537109375,
0.1353759765625,
0.483642578125,
0.1353759765625,
0.483642578125,
0.1484375,
0.1175537109375,
0.1484375
],
[
0.51318359375,
0.1455078125,
0.87890625,
0.1455078125,
0.87890625,
0.1591796875,
0.51318359375,
0.1591796875
],
[
0.12005615234375,
0.15380859375,
0.483642578125,
0.15380859375,
0.483642578125,
0.1644287109375,
0.12005615234375,
0.1644287109375
],
[
0.5146484375,
0.160888671875,
0.880859375,
0.160888671875,
0.880859375,
0.174560546875,
0.5146484375,
0.174560546875
],
[
0.1175537109375,
0.16748046875,
0.484375,
0.16748046875,
0.484375,
0.18115234375,
0.1175537109375,
0.18115234375
],
[
0.51318359375,
0.1776123046875,
0.880859375,
0.1776123046875,
0.880859375,
0.191162109375,
0.51318359375,
0.191162109375
],
[
0.118408203125,
0.185302734375,
0.484375,
0.185302734375,
0.484375,
0.198974609375,
0.118408203125,
0.198974609375
],
[
0.51318359375,
0.193603515625,
0.8818359375,
0.193603515625,
0.8818359375,
0.207275390625,
0.51318359375,
0.207275390625
],
[
0.118408203125,
0.201904296875,
0.486083984375,
0.201904296875,
0.486083984375,
0.215576171875,
0.118408203125,
0.215576171875
],
[
0.51318359375,
0.210205078125,
0.66162109375,
0.210205078125,
0.66162109375,
0.223876953125,
0.51318359375,
0.223876953125
],
[
0.118408203125,
0.21728515625,
0.47021484375,
0.21728515625,
0.47021484375,
0.23095703125,
0.118408203125,
0.23095703125
],
[
0.1175537109375,
0.232177734375,
0.486083984375,
0.2327880859375,
0.486083984375,
0.2476806640625,
0.1175537109375,
0.2464599609375
],
[
0.51318359375,
0.2381591796875,
0.54150390625,
0.2381591796875,
0.54150390625,
0.2493896484375,
0.51318359375,
0.2493896484375
],
[
0.5400390625,
0.2410888671875,
0.5498046875,
0.2410888671875,
0.5498046875,
0.2464599609375,
0.5400390625,
0.2464599609375
],
[
0.5556640625,
0.2381591796875,
0.64404296875,
0.2381591796875,
0.64404296875,
0.2529296875,
0.5556640625,
0.2529296875
],
[
0.1175537109375,
0.2493896484375,
0.484375,
0.2493896484375,
0.484375,
0.26318359375,
0.1175537109375,
0.26318359375
],
[
0.1175537109375,
0.26416015625,
0.483642578125,
0.2666015625,
0.483642578125,
0.27978515625,
0.1175537109375,
0.27783203125
],
[
0.1175537109375,
0.27978515625,
0.484375,
0.28076171875,
0.484375,
0.295654296875,
0.1175537109375,
0.29443359375
],
[
0.118408203125,
0.29638671875,
0.483642578125,
0.29638671875,
0.483642578125,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.52294921875,
0.30712890625,
0.5693359375,
0.30712890625,
0.5693359375,
0.311279296875,
0.52294921875,
0.311279296875
],
[
0.58837890625,
0.305908203125,
0.60791015625,
0.305908203125,
0.60791015625,
0.3134765625,
0.58837890625,
0.3134765625
],
[
0.7060546875,
0.30712890625,
0.73046875,
0.30712890625,
0.73046875,
0.311279296875,
0.7060546875,
0.311279296875
],
[
0.8369140625,
0.30712890625,
0.861328125,
0.30712890625,
0.861328125,
0.311279296875,
0.8369140625,
0.311279296875
],
[
0.118408203125,
0.3134765625,
0.484375,
0.3134765625,
0.484375,
0.327880859375,
0.118408203125,
0.327880859375
],
[
0.118408203125,
0.330078125,
0.484375,
0.330078125,
0.484375,
0.34326171875,
0.118408203125,
0.34326171875
],
[
0.6630859375,
0.336181640625,
0.7119140625,
0.336181640625,
0.7119140625,
0.34033203125,
0.6630859375,
0.34033203125
],
[
0.12005615234375,
0.3466796875,
0.484375,
0.3466796875,
0.484375,
0.35986328125,
0.12005615234375,
0.35986328125
],
[
0.66015625,
0.344482421875,
0.7109375,
0.344482421875,
0.7109375,
0.355224609375,
0.66015625,
0.355224609375
],
[
0.1175537109375,
0.361083984375,
0.483642578125,
0.361083984375,
0.483642578125,
0.374755859375,
0.1175537109375,
0.374755859375
],
[
0.5146484375,
0.37060546875,
0.8818359375,
0.37060546875,
0.8818359375,
0.38427734375,
0.5146484375,
0.38427734375
],
[
0.1175537109375,
0.377197265625,
0.486083984375,
0.377685546875,
0.486083984375,
0.392578125,
0.1175537109375,
0.391357421875
],
[
0.51123046875,
0.383056640625,
0.79248046875,
0.383056640625,
0.79248046875,
0.396728515625,
0.51123046875,
0.396728515625
],
[
0.1158447265625,
0.392578125,
0.382080078125,
0.393798828125,
0.382080078125,
0.407958984375,
0.1158447265625,
0.40673828125
],
[
0.1175537109375,
0.410888671875,
0.291259765625,
0.412109375,
0.291259765625,
0.427001953125,
0.1175537109375,
0.42578125
],
[
0.51318359375,
0.41748046875,
0.880859375,
0.41748046875,
0.880859375,
0.43115234375,
0.51318359375,
0.43115234375
],
[
0.51318359375,
0.432861328125,
0.880859375,
0.432861328125,
0.880859375,
0.446533203125,
0.51318359375,
0.446533203125
],
[
0.1376953125,
0.446533203125,
0.48779296875,
0.446533203125,
0.48779296875,
0.46142578125,
0.1376953125,
0.46142578125
],
[
0.5146484375,
0.449462890625,
0.880859375,
0.449462890625,
0.880859375,
0.463134765625,
0.5146484375,
0.463134765625
],
[
0.151123046875,
0.4619140625,
0.48779296875,
0.463134765625,
0.48779296875,
0.47802734375,
0.151123046875,
0.476806640625
],
[
0.51123046875,
0.465576171875,
0.87890625,
0.465576171875,
0.87890625,
0.478515625,
0.51123046875,
0.478515625
],
[
0.154541015625,
0.479736328125,
0.486083984375,
0.479736328125,
0.486083984375,
0.493408203125,
0.154541015625,
0.493408203125
],
[
0.51123046875,
0.482177734375,
0.880859375,
0.48095703125,
0.880859375,
0.49462890625,
0.51123046875,
0.495361328125
],
[
0.15283203125,
0.49462890625,
0.48779296875,
0.495361328125,
0.48779296875,
0.51025390625,
0.15283203125,
0.5087890625
],
[
0.5146484375,
0.49755859375,
0.880859375,
0.49755859375,
0.880859375,
0.51123046875,
0.5146484375,
0.51123046875
],
[
0.154541015625,
0.5126953125,
0.2509765625,
0.5126953125,
0.2509765625,
0.52392578125,
0.154541015625,
0.52392578125
],
[
0.5146484375,
0.51318359375,
0.880859375,
0.51318359375,
0.880859375,
0.52783203125,
0.5146484375,
0.52783203125
],
[
0.51318359375,
0.529296875,
0.87890625,
0.529296875,
0.87890625,
0.5419921875,
0.51318359375,
0.5419921875
],
[
0.1376953125,
0.5458984375,
0.496337890625,
0.5458984375,
0.496337890625,
0.56005859375,
0.1376953125,
0.56005859375
],
[
0.51123046875,
0.54443359375,
0.880859375,
0.5458984375,
0.880859375,
0.56005859375,
0.51123046875,
0.55859375
],
[
0.154541015625,
0.5625,
0.49365234375,
0.5625,
0.49365234375,
0.57666015625,
0.154541015625,
0.57666015625
],
[
0.51318359375,
0.56005859375,
0.880859375,
0.56103515625,
0.880859375,
0.576171875,
0.51318359375,
0.57470703125
],
[
0.154541015625,
0.5791015625,
0.489501953125,
0.5791015625,
0.489501953125,
0.5927734375,
0.154541015625,
0.5927734375
],
[
0.51123046875,
0.57763671875,
0.7421875,
0.57763671875,
0.7421875,
0.59130859375,
0.51123046875,
0.59130859375
],
[
0.154541015625,
0.59326171875,
0.483642578125,
0.59326171875,
0.483642578125,
0.60693359375,
0.154541015625,
0.60693359375
],
[
0.51904296875,
0.60400390625,
0.87890625,
0.60400390625,
0.87890625,
0.6181640625,
0.51904296875,
0.6181640625
],
[
0.15283203125,
0.61083984375,
0.287109375,
0.60986328125,
0.287109375,
0.62451171875,
0.15283203125,
0.6259765625
],
[
0.5146484375,
0.62060546875,
0.880859375,
0.62060546875,
0.880859375,
0.63427734375,
0.5146484375,
0.63427734375
],
[
0.51318359375,
0.6357421875,
0.87744140625,
0.6357421875,
0.87744140625,
0.6494140625,
0.51318359375,
0.6494140625
],
[
0.1368408203125,
0.6435546875,
0.490234375,
0.64453125,
0.490234375,
0.6591796875,
0.1368408203125,
0.65771484375
],
[
0.51318359375,
0.65283203125,
0.880859375,
0.65283203125,
0.880859375,
0.66650390625,
0.51318359375,
0.66650390625
],
[
0.15283203125,
0.6611328125,
0.486083984375,
0.6611328125,
0.486083984375,
0.6748046875,
0.15283203125,
0.6748046875
],
[
0.51318359375,
0.66943359375,
0.880859375,
0.66943359375,
0.880859375,
0.68310546875,
0.51318359375,
0.68310546875
],
[
0.151123046875,
0.6767578125,
0.486083984375,
0.6748046875,
0.486083984375,
0.689453125,
0.151123046875,
0.69140625
],
[
0.5146484375,
0.68505859375,
0.880859375,
0.68505859375,
0.880859375,
0.69873046875,
0.5146484375,
0.69873046875
],
[
0.154541015625,
0.6953125,
0.214111328125,
0.6953125,
0.214111328125,
0.70703125,
0.154541015625,
0.70703125
],
[
0.51318359375,
0.703125,
0.87890625,
0.703125,
0.87890625,
0.71337890625,
0.51318359375,
0.71337890625
],
[
0.51318359375,
0.71630859375,
0.87890625,
0.71630859375,
0.87890625,
0.72998046875,
0.51318359375,
0.72998046875
],
[
0.118408203125,
0.72802734375,
0.484375,
0.72802734375,
0.484375,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.51123046875,
0.732421875,
0.880859375,
0.73291015625,
0.880859375,
0.74755859375,
0.51123046875,
0.74658203125
],
[
0.12176513671875,
0.7451171875,
0.486083984375,
0.7451171875,
0.486083984375,
0.755859375,
0.12176513671875,
0.755859375
],
[
0.51123046875,
0.74755859375,
0.880859375,
0.7490234375,
0.880859375,
0.76318359375,
0.51123046875,
0.7626953125
],
[
0.1175537109375,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.1175537109375,
0.7724609375
],
[
0.509765625,
0.76416015625,
0.77734375,
0.76416015625,
0.77734375,
0.77783203125,
0.509765625,
0.77783203125
],
[
0.1158447265625,
0.7724609375,
0.486083984375,
0.77490234375,
0.486083984375,
0.7919921875,
0.1158447265625,
0.79052734375
],
[
0.118408203125,
0.7919921875,
0.484375,
0.7919921875,
0.484375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.515625,
0.79052734375,
0.880859375,
0.7919921875,
0.880859375,
0.80712890625,
0.515625,
0.8046875
],
[
0.118408203125,
0.80810546875,
0.489501953125,
0.80810546875,
0.489501953125,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.51123046875,
0.80712890625,
0.8818359375,
0.80810546875,
0.8818359375,
0.822265625,
0.51123046875,
0.82177734375
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51318359375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51123046875,
0.84130859375,
0.87890625,
0.84130859375,
0.87890625,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.1175537109375,
0.87158203125,
0.489501953125,
0.87255859375,
0.489501953125,
0.88720703125,
0.1175537109375,
0.8857421875
],
[
0.51318359375,
0.87353515625,
0.880859375,
0.87353515625,
0.880859375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.1175537109375,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.274658203125,
0.90478515625,
0.274658203125,
0.91943359375,
0.1175537109375,
0.91943359375
],
[
0.51318359375,
0.90478515625,
0.87890625,
0.90478515625,
0.87890625,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"的NLI模型，这些模型不特定于胸部X射线。",
"2 Factually-oriented Graph-Based",
"Reward",
"尽管这些“以事实为导向”的方法受到弱监督",
"在本节中，我们提出了一种新的基于语义图的",
"或仅限于通用生物医学实体，但它们的使用在",
"奖励，称为RadGraph奖励，在我们的实验中",
"广泛的指标和委员会认证的放射科医生的评",
"使用。我们首先在2.1节中解释RadGraph数据",
"估中显示出实质性的改进。这些发现促使我们",
"集，以及我们如何获得塑造我们奖励的注释。",
"提出一种新方法，以进一步提高生成的放射学",
"在2.2节中，我们解释了如何构造RadGraph奖",
"报告的事实完整性和正确性。更准确地说，我",
"励及其不同的变体",
"们利用RadGraphE(Jainetal，2021），这是",
"个由放射科医生注释的数据集，其中包含胸部",
"2.1",
"",
"RadGraph",
"X射线放射学报告以及注释的实体和关系。这",
"些注释允许我们创建两个语义图，一个用于",
"生成的，一个用于参考报告。然后，我们引人",
"三个简单的奖励，根据实体和关系对两个图之",
".increasec",
"right",
"pacity",
"ecuor",
"间的差异进行评分。这些奖励可以使用强化学",
"习（RL）直接优化，以进一步提高我们系统生",
"ououora",
"成的报告的质量。通过这样做，我们在两个流",
"Definey Absent",
"行的胸部X射线数据集上展示了我们的模型",
"Figure 2: An example of a report annotated with enti-",
"能够最大化定义的奖励，但也优于以前在各种",
"ties and relations in the RadGraph dataset",
"NLG和面向事实的指标上的工作",
"总之，我们的贡献是：",
"RadGraphE(Jain etal.，2021）是基于旨在构建",
"放射学报告的新颖信息提取模式的全文胸部X",
": We propose a simple RRG architecture that",
"射线放射学报告中的实体和关系的数据集。该",
"1) is fast to train and suitable for a RL setup",
"数据集包含来自MIMIC-CXR数据集EJohn",
"and 2) performs equally well as the previous",
"sonetal.，2019）的500份放射学报告的委员",
"and more complex architectures proposed in",
"会认证放射科医生注释，总共对应于14,579",
"the literature",
"个实体和10,889个关系。此外，RadGraph还",
"包括100份放射学报告的测试数据集，分为",
". We leverage the RadGraph dataset and the as-",
"两组独立的委员会认证放射科医生注释，分",
"sociated fine-tuned model to design semantic-",
"别来自 MIMIC-CXR和 CheXpertE(Smit et al.,",
"based rewards that qualitatively evaluate the",
"2020）数据集（各50份报告）",
"factual correctness and completeness of the",
"Entities实体被定义为可以包括一个或多个",
"generated reports.",
"相邻单词的文本的连续跨度。RadGraph中的",
"实体围绕两个概念：解学和观察。观察存在",
": We show on two datasets that directly opti-",
"三个不确定性水平，导致四个不同的实体：解",
"mizing these rewards outperforms previous",
"剖学（ANAT-DP），观察：明确存在（OBS-DP）",
"approaches that prioritize traditional NLG",
"观察：不确定（OBS-U）和观察：明确不存在",
"metrics.",
"（OBS-DA）。解部学是指在放射学报告中出现",
"的解剖学身体部位，例如“肺”。观察是指与",
"该论文的结构如下：首先，我们描述了我们面",
"视觉特征，可识别的病理生理过程或诊断疾病",
"向事实的基于图的奖励（82）。更准确地说，我",
"分类相关的单词。例如，观察结果可以是“积",
"们首先检查RadGraph数据集（sE2.1）以及我",
"液”或更一般的短语，如“增加”",
"们如何利用注释来创建奖励（sE2.2）。然后，",
"我们解释用于生成报告的模型（sE3）的体系",
"Relations关系定义为两个实体之间的有向",
"结构，以及如何使用负对数似然（NLL）和RL",
"边。存在三个层次：暗示（。，。），位于（。，。），",
"对其进行训练。随后的部分专门介绍用于实验",
"并修改（。，。）。（观察，观察）的建议是两个观",
"的数据集（sE4）和用于评估报告生成的度量",
"察实体之间的关系，表明第二个观察的存在是",
"（sE5）。后一部分分为两组：经典NLG度量",
"从第一个观察的存在推断出的。位于（观察，",
"（sE5.1）和面向事实的度量（sE5.2）。最后，",
"解剖学）是观察实体和解学实体之间的关",
"我们给出了结果（86），并以一节寻址相关作",
"系，表示观察与解剖学相关。虽然位于通常指",
"品（$7）结束本文。",
"的是位置，它也可以用来描述观察和解剖学之"
] |
{
"arxiv_id": "2210.12186",
"language": "zh",
"page_idx": 2
}
|
||
2210.12186-zh-3
|
2210.12186-zh-3.jpg
|
间的其他关系，如形状或颜色。修改(观察， 观察)或修改(解剖，解剖)是两个观察实体或两个解剖实体之间的关系，指示第一个实体修改第二个实体的范围或量化第二个实体的程度。
作者还发布了一个在 RadGraph 数据集上微调的 PubMedBERT (Gu et al., 2021) 模型。我们利用这个经过训练的模型为实验中使用的数据集创建注释。我们将在下面的内容中将此模型称为 RadGraph 模型
## 2.2 RadGraph reward
<FigureHere>
Figure 3: Graph view of the RadGraph annotations for
the report in Figure 2 .
使用 RadGraph 注释方案和模型，我们设计了 F-score 风格的奖励，用于衡量生成的放射学报告与参考报告相比的一致性和完整性。我们的每个奖励都利用了 RadGraph 上发布的微调 PubMedBERT 模型的输出，即生成的报告及其引用上的实体和关系。报告的 RadGraph 注释可以表示为图形$\mathcal{G}(V,E)$ ,其中节点集$V=\{v_1,v_2,\ldots,v_{|V|}\}$包含实体，而边集$E=\left\{e_1,e_2,\ldots,e_{|E|}\right\}$表示实体对之间的关系。该图是有向的，这意味着边缘$e=(v_1,v_2)\neq(v_2,v_1)$。图 3 中描述了一个示例。图的每个节点或边也有一个标签，我们将其表示为实体$i$的$v_{i_L}$ (例如“OBS-DP”或“ANAT”)和关系$e=(v_i,v_j)$的$e_ij_L$(例如“已修改”或“位于”)。现在我们开始描述我们的三个奖励。
$\mathrm{RG}_{\mathrm{E} }$ 此奖励仅关注节点$V$。对于生成的报告$y$,我们创建了一组新的节点标签对$\bar{V} _y= \{ ( v_i, v_{i_L}) \} _{i\in [ 1. . | V| ] }$ ,其中包括所有实体及其相应的标签。我们继续为参考报告$\hat{y}$构造相
同的集合，并将该集合表示为$\bar{V}_{\hat{y}}$ 。
$\mathrm{RG}_{\mathrm{ER}}$ 此奖励重点关注节点$V$ 以及节点在 $E$ 中是否具有关系。对于生成的报告$y$ ,我们创建一组新的三元组$\bar{V}_y=\{(v_i,v_{i_L},\epsilon_i)\}_{i\in[1..|V|]}$。如果$v_i$在$E$中具有关系，则$\epsilon_i$的值为1，否则为 0。我们继续为参考报告$\hat{y}$构造相同的集合，并将该集合表示为$\bar{V}_{\hat{y}}$ 。
$\mathrm{RG}_{\overline {\mathrm{ER}}}$ 此奖励关注$E$中的节点$V$及其关系。对于生成的报告$y$ ,我们创建一组新的元组$\bar{V} _{y}= \{ ( v_{i}, v_{i_{L}}, ( v_{i}, v_{j}) , e_{ij_{L}}) \mid i\in [ 1. . | V| ] $, $j\in$ $[ 1. . | V| ] $, $j\neq i$, $( v_{i}, \dot{v_{j}} ) \in E\}$。此外，对于所有没有关系的节点$v_i$ , 我 们 在 $\bar{V} _y$中包含一个元组$(v_i,v_{i_L})$ 。我们继续为参考报告$\hat{y}$构造相同的集合，并将此集合表示为$\bar{V}_{\hat{y}}$ 。
最后，RG$_\mathrm{E}$, RG$_\mathrm{ER}$和 RG$_\mathrm{\overline{ER}}$被定义为它们各自的集合$\bar{V}_{\hat{y}}$ 和$\bar{V}_y$ 之间的精度和召回率的调和平均值。
作为说明，我们在附录EB 中提供了图E3 中
的图$\mathcal{G}$的集 $\bar{V}$
### 3 $\mathbf{Model}$
体系结构 为了对 X 射线图像进行编码，我们使用 Densenet-121$\boxed{\mathrm{F}}($Huang et al.,2017)提取大小为$49\times1024$的卷积视觉特征$D$ 。为了生成语言，我们使用了一个单层的 BERT(Vaswani et al., 2017; Devlin et al., 2019) , 在视觉特征上具有交叉注意力。更正式地说，变压器层的交叉注意力被写成：
$$\text{Cross-Attention}(Q,K,V)=\mathrm{softmax}(\frac{QK^{\top}}{\sqrt{d}})V$$
其中$Q$ 是大小为$d$ 的 BERT 隐藏状态，$K$ 和
$V$是视觉特征$D$ 。模型的全部细节可以在附
录 D 中找到。
训练 如果我们将$\theta$表示为模型参数，则通过最大化观察到的报告$Y=\left(y_1,y_2,\cdots,y_n\right)$的似然或换句话说，通过最小化负对数似然来学习$\theta$。目标函数由下式给出：
(2)
$$\mathcal{L}(\boldsymbol{\theta})=-\sum_{t=1}^n\log p_{\boldsymbol{\theta}}(\boldsymbol{y}_t|\boldsymbol{y}<t,\boldsymbol{D})$$
在 NLL 训练之后，我们开始 RL 训练，以优化我们面向事实的奖励之一。方程 2 中的损失函数现在由下式给出：
|
[
[
0.118408203125,
0.08489990234375,
0.489501953125,
0.08489990234375,
0.489501953125,
0.09857177734375,
0.118408203125,
0.09857177734375
],
[
0.51318359375,
0.08612060546875,
0.78662109375,
0.08612060546875,
0.78662109375,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.118408203125,
0.1015625,
0.484375,
0.1015625,
0.484375,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.12005615234375,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.51318359375,
0.1181640625,
0.5615234375,
0.12054443359375,
0.56103515625,
0.1318359375,
0.51123046875,
0.1300048828125
],
[
0.580078125,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.580078125,
0.1312255859375
],
[
0.1175537109375,
0.133056640625,
0.48779296875,
0.1341552734375,
0.48779296875,
0.1484375,
0.1175537109375,
0.1478271484375
],
[
0.51318359375,
0.1341552734375,
0.87744140625,
0.1341552734375,
0.87744140625,
0.1478271484375,
0.51318359375,
0.1478271484375
],
[
0.1175537109375,
0.1484375,
0.14697265625,
0.1484375,
0.14697265625,
0.1650390625,
0.1175537109375,
0.1650390625
],
[
0.51123046875,
0.1478271484375,
0.8818359375,
0.1519775390625,
0.8818359375,
0.168701171875,
0.51123046875,
0.1644287109375
],
[
0.5146484375,
0.16748046875,
0.880859375,
0.16748046875,
0.880859375,
0.18115234375,
0.5146484375,
0.18115234375
],
[
0.118408203125,
0.182861328125,
0.486083984375,
0.182861328125,
0.486083984375,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.5146484375,
0.18408203125,
0.87890625,
0.18408203125,
0.87890625,
0.19775390625,
0.5146484375,
0.19775390625
],
[
0.118408203125,
0.1995849609375,
0.486083984375,
0.1995849609375,
0.486083984375,
0.213134765625,
0.118408203125,
0.213134765625
],
[
0.51123046875,
0.196533203125,
0.736328125,
0.198974609375,
0.736328125,
0.2161865234375,
0.51123046875,
0.21435546875
],
[
0.118408203125,
0.215576171875,
0.483642578125,
0.215576171875,
0.483642578125,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.1175537109375,
0.2298583984375,
0.48779296875,
0.23095703125,
0.48779296875,
0.2452392578125,
0.1175537109375,
0.24462890625
],
[
0.51318359375,
0.23095703125,
0.5810546875,
0.2340087890625,
0.580078125,
0.248779296875,
0.51123046875,
0.2452392578125
],
[
0.57275390625,
0.23095703125,
0.880859375,
0.232177734375,
0.880859375,
0.2464599609375,
0.57275390625,
0.2452392578125
],
[
0.1175537109375,
0.2476806640625,
0.276123046875,
0.2476806640625,
0.276123046875,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.5146484375,
0.248779296875,
0.880859375,
0.248779296875,
0.880859375,
0.262451171875,
0.5146484375,
0.262451171875
],
[
0.51318359375,
0.26318359375,
0.8818359375,
0.26318359375,
0.8818359375,
0.28076171875,
0.51318359375,
0.28076171875
],
[
0.1158447265625,
0.276611328125,
0.16455078125,
0.276611328125,
0.16455078125,
0.29150390625,
0.1158447265625,
0.29150390625
],
[
0.15869140625,
0.27783203125,
0.310546875,
0.27783203125,
0.310546875,
0.292236328125,
0.15869140625,
0.292236328125
],
[
0.51318359375,
0.279052734375,
0.87890625,
0.28076171875,
0.87890625,
0.295654296875,
0.51318359375,
0.29345703125
],
[
0.5146484375,
0.297607421875,
0.880859375,
0.297607421875,
0.880859375,
0.311279296875,
0.5146484375,
0.311279296875
],
[
0.12005615234375,
0.3134765625,
0.16455078125,
0.3134765625,
0.16455078125,
0.32470703125,
0.12005615234375,
0.32470703125
],
[
0.51318359375,
0.3134765625,
0.87890625,
0.3134765625,
0.87890625,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.26611328125,
0.32470703125,
0.33154296875,
0.32470703125,
0.33154296875,
0.339111328125,
0.26611328125,
0.339111328125
],
[
0.51123046875,
0.32666015625,
0.77099609375,
0.327880859375,
0.77099609375,
0.345703125,
0.51123046875,
0.344482421875
],
[
0.274658203125,
0.339111328125,
0.322509765625,
0.339111328125,
0.322509765625,
0.3466796875,
0.274658203125,
0.3466796875
],
[
0.151123046875,
0.347412109375,
0.19482421875,
0.349853515625,
0.193115234375,
0.361083984375,
0.1502685546875,
0.359375
],
[
0.15869140625,
0.358154296875,
0.189697265625,
0.3623046875,
0.185546875,
0.37353515625,
0.154541015625,
0.3681640625
],
[
0.411376953125,
0.35986328125,
0.462646484375,
0.35986328125,
0.462646484375,
0.371826171875,
0.411376953125,
0.371826171875
],
[
0.51123046875,
0.359375,
0.880859375,
0.35986328125,
0.880859375,
0.377685546875,
0.51123046875,
0.377197265625
],
[
0.411376953125,
0.371826171875,
0.459228515625,
0.371826171875,
0.459228515625,
0.3818359375,
0.411376953125,
0.3818359375
],
[
0.515625,
0.377685546875,
0.880859375,
0.377685546875,
0.880859375,
0.392578125,
0.515625,
0.392578125
],
[
0.51318359375,
0.393798828125,
0.59765625,
0.393798828125,
0.59765625,
0.407958984375,
0.51318359375,
0.407958984375
],
[
0.1796875,
0.410400390625,
0.2149658203125,
0.410400390625,
0.2149658203125,
0.4228515625,
0.1796875,
0.4228515625
],
[
0.1796875,
0.4228515625,
0.210693359375,
0.4228515625,
0.210693359375,
0.435302734375,
0.1796875,
0.435302734375
],
[
0.5146484375,
0.427001953125,
0.87890625,
0.427001953125,
0.87890625,
0.440673828125,
0.5146484375,
0.440673828125
],
[
0.34765625,
0.440673828125,
0.414794921875,
0.440673828125,
0.414794921875,
0.451904296875,
0.34765625,
0.451904296875
],
[
0.51123046875,
0.4423828125,
0.62646484375,
0.4423828125,
0.62646484375,
0.457275390625,
0.51123046875,
0.457275390625
],
[
0.359375,
0.454833984375,
0.399658203125,
0.454833984375,
0.399658203125,
0.4619140625,
0.359375,
0.4619140625
],
[
0.160400390625,
0.46142578125,
0.2568359375,
0.463134765625,
0.2568359375,
0.47802734375,
0.160400390625,
0.4755859375
],
[
0.51904296875,
0.471435546875,
0.53076171875,
0.478515625,
0.51904296875,
0.4892578125,
0.50732421875,
0.482177734375
],
[
0.54248046875,
0.47265625,
0.60546875,
0.47265625,
0.60546875,
0.48681640625,
0.54248046875,
0.48681640625
],
[
0.5341796875,
0.47802734375,
0.54150390625,
0.47802734375,
0.54150390625,
0.482177734375,
0.5341796875,
0.482177734375
],
[
0.18896484375,
0.4755859375,
0.2333984375,
0.4755859375,
0.2333984375,
0.48681640625,
0.18896484375,
0.48681640625
],
[
0.5146484375,
0.50048828125,
0.880859375,
0.50048828125,
0.880859375,
0.51416015625,
0.5146484375,
0.51416015625
],
[
0.1175537109375,
0.5185546875,
0.484375,
0.5185546875,
0.484375,
0.529296875,
0.1175537109375,
0.529296875
],
[
0.5146484375,
0.5166015625,
0.880859375,
0.5166015625,
0.880859375,
0.53076171875,
0.5146484375,
0.53076171875
],
[
0.118408203125,
0.533203125,
0.273681640625,
0.533203125,
0.273681640625,
0.54345703125,
0.118408203125,
0.54345703125
],
[
0.515625,
0.533203125,
0.880859375,
0.533203125,
0.880859375,
0.54638671875,
0.515625,
0.54638671875
],
[
0.51318359375,
0.548828125,
0.880859375,
0.54736328125,
0.880859375,
0.5625,
0.51318359375,
0.5634765625
],
[
0.118408203125,
0.5634765625,
0.486083984375,
0.56396484375,
0.486083984375,
0.5791015625,
0.118408203125,
0.57763671875
],
[
0.51123046875,
0.56396484375,
0.87744140625,
0.56396484375,
0.87744140625,
0.57763671875,
0.51123046875,
0.57763671875
],
[
0.12005615234375,
0.58056640625,
0.48779296875,
0.58056640625,
0.48779296875,
0.59423828125,
0.12005615234375,
0.59423828125
],
[
0.51318359375,
0.58203125,
0.880859375,
0.58203125,
0.880859375,
0.595703125,
0.51318359375,
0.595703125
],
[
0.118408203125,
0.59716796875,
0.48779296875,
0.59716796875,
0.48779296875,
0.61083984375,
0.118408203125,
0.61083984375
],
[
0.51318359375,
0.59716796875,
0.64990234375,
0.59716796875,
0.64990234375,
0.61083984375,
0.51318359375,
0.61083984375
],
[
0.1175537109375,
0.6123046875,
0.48779296875,
0.61083984375,
0.48779296875,
0.6259765625,
0.1175537109375,
0.626953125
],
[
0.1175537109375,
0.62890625,
0.484375,
0.62890625,
0.484375,
0.642578125,
0.1175537109375,
0.642578125
],
[
0.796875,
0.62744140625,
0.84619140625,
0.62744140625,
0.84619140625,
0.642578125,
0.796875,
0.642578125
],
[
0.515625,
0.63427734375,
0.80615234375,
0.6328125,
0.80615234375,
0.65087890625,
0.515625,
0.65185546875
],
[
0.1175537109375,
0.6435546875,
0.484375,
0.64453125,
0.484375,
0.6591796875,
0.1175537109375,
0.65771484375
],
[
0.8154296875,
0.6435546875,
0.8486328125,
0.6435546875,
0.8486328125,
0.66015625,
0.8154296875,
0.66015625
],
[
0.8486328125,
0.6396484375,
0.869140625,
0.6396484375,
0.869140625,
0.64794921875,
0.8486328125,
0.64794921875
],
[
0.85791015625,
0.6591796875,
0.880859375,
0.6591796875,
0.880859375,
0.67138671875,
0.85791015625,
0.67138671875
],
[
0.118408203125,
0.6767578125,
0.486083984375,
0.6767578125,
0.486083984375,
0.69140625,
0.118408203125,
0.69140625
],
[
0.5146484375,
0.6748046875,
0.880859375,
0.6748046875,
0.880859375,
0.6884765625,
0.5146484375,
0.6884765625
],
[
0.1368408203125,
0.69482421875,
0.486083984375,
0.69482421875,
0.486083984375,
0.708984375,
0.1368408203125,
0.708984375
],
[
0.51123046875,
0.6904296875,
0.880859375,
0.6904296875,
0.880859375,
0.70361328125,
0.51123046875,
0.70361328125
],
[
0.1259765625,
0.69970703125,
0.1368408203125,
0.69970703125,
0.1368408203125,
0.705078125,
0.1259765625,
0.705078125
],
[
0.1175537109375,
0.708984375,
0.486083984375,
0.708984375,
0.486083984375,
0.72705078125,
0.1175537109375,
0.72705078125
],
[
0.51123046875,
0.7060546875,
0.623046875,
0.7060546875,
0.623046875,
0.72021484375,
0.51123046875,
0.72021484375
],
[
0.118408203125,
0.72705078125,
0.484375,
0.72705078125,
0.484375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.1158447265625,
0.7421875,
0.486083984375,
0.7412109375,
0.486083984375,
0.755859375,
0.1158447265625,
0.75732421875
],
[
0.51318359375,
0.73828125,
0.87890625,
0.73828125,
0.87890625,
0.751953125,
0.51318359375,
0.751953125
],
[
0.509765625,
0.751953125,
0.8818359375,
0.75439453125,
0.8818359375,
0.771484375,
0.509765625,
0.76953125
],
[
0.12005615234375,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.12005615234375,
0.7724609375
],
[
0.1158447265625,
0.77392578125,
0.48779296875,
0.7724609375,
0.48779296875,
0.79052734375,
0.1158447265625,
0.79150390625
],
[
0.51318359375,
0.771484375,
0.8818359375,
0.771484375,
0.8818359375,
0.78515625,
0.51318359375,
0.78515625
],
[
0.11419677734375,
0.7880859375,
0.484375,
0.7890625,
0.484375,
0.80712890625,
0.11419677734375,
0.8056640625
],
[
0.509765625,
0.78759765625,
0.7421875,
0.78759765625,
0.7421875,
0.80029296875,
0.509765625,
0.80029296875
],
[
0.1175537109375,
0.8056640625,
0.484375,
0.80712890625,
0.484375,
0.82177734375,
0.1175537109375,
0.82080078125
],
[
0.118408203125,
0.822265625,
0.2376708984375,
0.822265625,
0.2376708984375,
0.83740234375,
0.118408203125,
0.83740234375
],
[
0.6484375,
0.81787109375,
0.6630859375,
0.8125,
0.67236328125,
0.82470703125,
0.6572265625,
0.830078125
],
[
0.568359375,
0.82763671875,
0.63232421875,
0.82763671875,
0.63232421875,
0.8427734375,
0.568359375,
0.8427734375
],
[
0.6455078125,
0.82470703125,
0.82958984375,
0.82470703125,
0.82958984375,
0.845703125,
0.6455078125,
0.845703125
],
[
0.85791015625,
0.8291015625,
0.880859375,
0.8291015625,
0.880859375,
0.8427734375,
0.85791015625,
0.8427734375
],
[
0.6484375,
0.84033203125,
0.67529296875,
0.84326171875,
0.6708984375,
0.85791015625,
0.64404296875,
0.85498046875
],
[
0.1175537109375,
0.85546875,
0.484375,
0.85546875,
0.484375,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.88720703125,
0.118408203125,
0.88720703125
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.12005615234375,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90380859375,
0.12005615234375,
0.90380859375
],
[
0.51123046875,
0.88818359375,
0.880859375,
0.88720703125,
0.880859375,
0.90185546875,
0.51123046875,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.6689453125,
0.90478515625,
0.6689453125,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"间的其他关系，如形状或颜色。修改（观察，",
"同的集合，并将该集合表示为Vg。",
"观察）或修改（解剖，解剖）是两个观察实体",
"或两个解剖实体之间的关系，指示第一个实体",
"RGER",
"此奖励重点关注节点V以及节点在E",
"修改第二个实体的范围或量化第二个实体的程",
"中是否具有关系。对于生成的报告y，我们创",
"度。",
"建一组新的三元组V=【(ui，Ui，e))ic[1VI。",
"如果ui在E中具有关系，则E的值为1，否",
"作者还发布了一个在RadGraph数据集上微调",
"则为0。我们继续为参考报告9构造相同的集",
"的PubMedBERT（Guetal.，2021）模型。我们",
"合，并将该集合表示为Vg。",
"利用这个经过训练的模型为实验中使用的数据",
"集创建注释。我们将在下面的内容中将此模型",
"RGER",
"此奖励关注E中的节点V及其关系。",
"称为RadGraph模型",
"对于生成的报告y，我们创建一组新的元组",
"Vy = [(ui, Uit,(vi,vj),eiL) [ i e [1.,]Vll, j e",
"2.2",
" RadGraph reward",
"[1..|VIl,j≠i（ui,u)EE。此外，对于所有",
"没有关系的节点i，我们在Vy中包含一个元",
"GRAPH",
"组（Ui，UiL）。我们继续为参考报告9构造相同",
"infection",
"的集合，并将此集合表示为Vg。",
"OBS-DP",
"lower",
"ANAT",
"opacity",
"最后，RGe，RGER和RGE被定义为它们各",
"OBS-DP",
"自的集合V和Vu之间的精度和召回率的调",
"和平均值。",
"right",
"ANAT",
"作为说明，我们在附录EB中提供了图E3中",
"increased",
"的图g的集√",
"OBS-DF",
"pneumothorax",
"3",
" Model",
"",
"OBS-DA",
"体系结构为了对X射线图像进行编码，我们",
"Figure3:Graphviewof theRadGraphannotationsfor",
"使用Densenet-121E(Huangetal.,2017）提取大",
"the report in Figure 2.",
"小为49×1024的卷积视觉特征D。为了生成",
"语言，我们使用了一个单层的BERT（Vaswani",
"使用RadGraph注释方案和模型，我们设",
"etal.,2017:Devlinetal.,2019），在视觉特征上",
"计了F-score风格的奖励，用于衡量生成",
"具有交叉注意力。更正式地说，变压器层的交",
"的放射学报告与参考报告相比的一致性",
"叉注意力被写成：",
"和完整性。我们的每个奖励都利用了Rad-",
"Graph上发布的微调PubMedBERT模型的",
".QKT",
"Cross-Attention(Q, K, V) = softmax(",
"输出，即生成的报告及其引用上的实体和关系。",
"Va",
"V",
"(1)",
"报告的RadGraph注释可以表示为图形G(V,E)",
"其中Q是大小为d的BERT隐藏状态，K和",
"其中节点集V={ui,U2，·,Uv}包含实",
"V是视觉特征D。模型的全部细节可以在附",
"",
"体，而边集E=[e1,e2..,eE}表示实体对",
"录 D 中找到。",
"之间的关系。该图是有向的，这意味着边缘",
"e=（Ui,V2）≠（V2,U1）。图3中描述了一个示",
"训练如果我们将表示为模型参数，则通过",
"最大化观察到的报告Y=（y1,y2，，9n）的",
"例。图的每个节点或边也有一个标签，我们将",
"其表示为实体i的Ui（例如“OBS-DP”或",
"似然或换句话说，通过最小化负对数似然来学",
"“ANAT”）和关系e=（ui,i）的eiL（例如",
"习θ。目标函数由下式给出：",
"“已修改”或“位于”）。现在我们开始描述我",
"们的三个奖励。",
"n",
"C(0) =",
" log pe(ytly < t, D)",
"(2)",
"=1",
"RGe此奖励仅关注节点V。对于生成的",
"报告y，我们创建了一组新的节点标签对",
"在NLL训练之后，我们开始RL训练，以优化",
"Vy={(iUiL))ie[1.|Vil，其中包括所有实体及",
"我们面向事实的奖励之一。方程2中的损失函",
"其相应的标签。我们继续为参考报告9构造相",
"数现在由下式给出："
] |
{
"arxiv_id": "2210.12186",
"language": "zh",
"page_idx": 3
}
|
||
2210.12186-zh-4
|
2210.12186-zh-4.jpg
|
(3)
$$\mathcal{L}(\boldsymbol{\theta})=-\mathbb{E}_{\boldsymbol{Y}\sim p_{\boldsymbol{\theta}}}r(\boldsymbol{Y})$$
其中$r(Y)$是对生成的报告的奖励。我们使用SCST 算法 (Rennie et al., 2017) 来近似我们的不可微奖励函数的期望梯度。表达式变为：
$$\nabla_{\boldsymbol{\theta}}\mathcal{L}(\boldsymbol{\theta})\approx-(r(\boldsymbol{Y})-r(\bar{\boldsymbol{Y}}))\nabla_{\boldsymbol{\theta}}\log_{p\boldsymbol{\theta}}(\boldsymbol{Y})$$
这里 $r(\bar{Y})$ 充当基线$\overline{\mathbb{E}}($Sutton et al.,1998),以减少$r(Y)$的方差。在我们的例子中，$r(\bar{Y})$是在训练期间从模型中采样的预期奖励。
超参数 我们的模型由 1 个尺寸为 768 的变压器块和尺寸为 3072 的前馈层组成。作为优化器，我们选择 Adam (Kingma and Ba,2015) , 学习率为$3e^{- 4}$ ,小批量大小为 128。我们用大小为 3 的波束搜索进行解码
## 4 Datasets
为了进行我们的实验，我们使用了两个胸部X 射线数据集：MIMIC-CXR$\boxed{\mathrm{E}}($Johnson et al., 2019)和 Open-i 胸部 X 射线 (印第安纳大学， Demner-Fushman et al. (2012) )。对于这两个数据集，我们使用官方拆分，但我们丢弃不包含结果部分的报告。因此，MIMIC-CXR 由152,173训练样本、用于验证的1，196和用于测试的2，347组成；同样，如果我们不计算每个研究中图像的多重性，并且如果我们丢弃。由于这使得数据集太小，无法训练和测试RRG 模型，因此 Open-i 数据集仅用于测试目的。
对于两个数据集的每个样本，我们通过在地面真相发现部分上运行 CheXbert 标签器$\overline{\mathrm{E}}($Smit et al., 2020) 来生成我们认为的地面真相放射学诊断。这为每个地面实况报告创建了相关的诊断标签，该标签描述了14个可能的观察结果的存在程度(例如。g。报告可能为阳性、阴性、 不确定或未指定的实变或水肿)。此标签用于计算 F1CheXbert 分数(请参阅 5.2 节)
## 5 $\textbf{Metrics}$
射学报告的事实正确性和完整性。
## 5.1 $\mathbf{NLG- oriented}$
我们报告 BLEU (Papineni et al.,2002)和CIDEr (Vedantam et al.,2015) 指标以评估世代。为了与以前的工作$\overline{\mathrm{E}}($Chen et al.,2021)保持一致，我们还报告了 ROUGE-L 度量$\overline{\mathrm{E}}($Lin, 2004)
## 5.2 Factually-oriented
以下提供的度量以不同的方式评估生成的结果的事实正确性和完整性。我们继续描述它们及其差异。
fact$_{\mathrm{ENT}}$ (Miura et al., 2021) 将命名实体识别器 (NER) 应用于生成的报告$\hat{y}$和相应的引用$y$ ,分别给出两组提取的实体$\mathbb{E}_\hat{y}$和$\mathbb{E}_y$ 。fact$_\mathrm{ENT}$ 定义为两组$\mathbb{E}_{\hat{y}}$和$\mathbb{E}_y$之间的精度和召回率的调和平均值。Stanza (Qi et al.,2020) 的临床模型用作 NER。
fact$_\mathrm{ENTNLI}$ (Miura et al., 2021) 此分数是fact$_\mathrm{ENT}$与自然语言推理(NLI)的扩展。在这里，如果$\mathbb{E}_y$中存在$\mathbb{E}_\hat{y}$的实体$e$ ,则不会自动将其视为正确。要被视为有效，包含实体$e$的句子$s_{\hat{y}}$不得与参考报告中的对应句子$s_y$相矛盾。参考报告中对应的句子$s_y$是 BERTScore 最高的句子 (Zhang et al., 2019) 对$s_{\hat{y}}$ 。NLI 模型输出句子$s_y$是否与$s_\hat{y}$矛盾。我们使用Miura et al. (2021)的 NLI 模型权重，它依赖于BERT 架构。
F $_1$ CheXbert$\underline{\mathrm{E}}$(Zhang et al.,2020) 此分数使用 CheXbert$\boxed{\mathrm{E}}($Smit et al.,2020),这是一种基于 Transformer 的模型，经过训练可在放射学报告作为输入的情况下输出胸部 X 射线的异常 (十四个类别)。F $_1$ CheXbert 是 CheXbert 对生成的报告$\hat{y}$和相应的参考$y$的预测之间的F1 分数。为了与以前的工作一致，分数是根据5 个观察值计算的：肺不张，心脏肥大，实变， 水肿和胸腔积液。
RG 奖励 我们使用 2.2 节中解释的奖励作为
评估分数
5.3 RadGraph vs fact$_\mathrm{ENT}$ and fact$_\mathrm{ENTNLI}$
理论上，RG$_\mathrm{E}$封装了 fact$_\mathrm{ENT}$和 fact$_\mathrm{ENTNLI}$概念。事实上，RG$_{\mathrm{E}}$专注于拥有正确的实体和正
在本节中，我们将继续报告用于评估模型生成 确的实体标签。鉴于实体的标签包含解剖或观的结果与人工编辑的结果的所有指标。我们将 察的概念，前者根据定义始终存在，而后者具指标分为两类，NLG 文献中广泛报道的 NLG 有一定程度的存在 (“存在”,“不存在”或“不指标，以及面向事实的指标，专门用于评估放 确定”), RGE 可以以与 fact$_\mathrm{ENTNLI}$相同的方式
惩罚与参考相矛盾的报告。
|
[
[
0.509765625,
0.08489990234375,
0.6689453125,
0.08489990234375,
0.6689453125,
0.09857177734375,
0.509765625,
0.09857177734375
],
[
0.2149658203125,
0.105712890625,
0.385498046875,
0.105712890625,
0.385498046875,
0.12054443359375,
0.2149658203125,
0.12054443359375
],
[
0.462646484375,
0.105712890625,
0.509765625,
0.105712890625,
0.509765625,
0.12054443359375,
0.462646484375,
0.12054443359375
],
[
0.51318359375,
0.1092529296875,
0.880859375,
0.1092529296875,
0.880859375,
0.122314453125,
0.51318359375,
0.122314453125
],
[
0.51318359375,
0.12469482421875,
0.880859375,
0.12469482421875,
0.880859375,
0.1383056640625,
0.51318359375,
0.1383056640625
],
[
0.12005615234375,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.51318359375,
0.141357421875,
0.880859375,
0.141357421875,
0.880859375,
0.155029296875,
0.51318359375,
0.155029296875
],
[
0.118408203125,
0.1519775390625,
0.48779296875,
0.1519775390625,
0.48779296875,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.509765625,
0.1561279296875,
0.87890625,
0.1561279296875,
0.87890625,
0.169189453125,
0.509765625,
0.169189453125
],
[
0.1175537109375,
0.166259765625,
0.455078125,
0.16748046875,
0.455078125,
0.18115234375,
0.1175537109375,
0.179931640625
],
[
0.51123046875,
0.17041015625,
0.55908203125,
0.1728515625,
0.5576171875,
0.18701171875,
0.509765625,
0.185302734375
],
[
0.51318359375,
0.201904296875,
0.70361328125,
0.201904296875,
0.70361328125,
0.215576171875,
0.51318359375,
0.215576171875
],
[
0.130126953125,
0.2161865234375,
0.48779296875,
0.21728515625,
0.48779296875,
0.232177734375,
0.130126953125,
0.23095703125
],
[
0.51318359375,
0.2244873046875,
0.87890625,
0.2244873046875,
0.87890625,
0.2381591796875,
0.51318359375,
0.2381591796875
],
[
0.51318359375,
0.2423095703125,
0.8759765625,
0.2423095703125,
0.8759765625,
0.2529296875,
0.51318359375,
0.2529296875
],
[
0.1175537109375,
0.2493896484375,
0.489501953125,
0.25048828125,
0.489501953125,
0.265380859375,
0.1175537109375,
0.26416015625
],
[
0.5087890625,
0.254638671875,
0.580078125,
0.257080078125,
0.580078125,
0.271484375,
0.5087890625,
0.26953125
],
[
0.118408203125,
0.267333984375,
0.486083984375,
0.267333984375,
0.486083984375,
0.28076171875,
0.118408203125,
0.28076171875
],
[
0.118408203125,
0.283203125,
0.420654296875,
0.283203125,
0.420654296875,
0.29638671875,
0.118408203125,
0.29638671875
],
[
0.51318359375,
0.289306640625,
0.880859375,
0.289306640625,
0.880859375,
0.303955078125,
0.51318359375,
0.303955078125
],
[
0.51123046875,
0.303955078125,
0.8818359375,
0.305908203125,
0.8818359375,
0.319580078125,
0.51123046875,
0.317626953125
],
[
0.1175537109375,
0.31640625,
0.484375,
0.31640625,
0.484375,
0.330078125,
0.1175537109375,
0.330078125
],
[
0.5146484375,
0.32470703125,
0.53076171875,
0.32470703125,
0.53076171875,
0.33203125,
0.5146484375,
0.33203125
],
[
0.52294921875,
0.320556640625,
0.87890625,
0.32177734375,
0.87890625,
0.336181640625,
0.52294921875,
0.3349609375
],
[
0.12005615234375,
0.3349609375,
0.486083984375,
0.3349609375,
0.486083984375,
0.345703125,
0.12005615234375,
0.345703125
],
[
0.51318359375,
0.33740234375,
0.880859375,
0.33740234375,
0.880859375,
0.3515625,
0.51318359375,
0.3515625
],
[
0.118408203125,
0.349853515625,
0.490234375,
0.349853515625,
0.490234375,
0.363525390625,
0.118408203125,
0.363525390625
],
[
0.51123046875,
0.3515625,
0.880859375,
0.352783203125,
0.880859375,
0.36767578125,
0.51123046875,
0.366455078125
],
[
0.118408203125,
0.365234375,
0.48779296875,
0.365234375,
0.48779296875,
0.37890625,
0.118408203125,
0.37890625
],
[
0.51318359375,
0.369384765625,
0.62109375,
0.369384765625,
0.62109375,
0.38427734375,
0.51318359375,
0.38427734375
],
[
0.1175537109375,
0.380126953125,
0.339111328125,
0.38134765625,
0.339111328125,
0.3955078125,
0.1175537109375,
0.394287109375
],
[
0.51123046875,
0.402099609375,
0.6171875,
0.40380859375,
0.6171875,
0.419189453125,
0.51123046875,
0.41748046875
],
[
0.609375,
0.402587890625,
0.880859375,
0.402587890625,
0.880859375,
0.416259765625,
0.609375,
0.416259765625
],
[
0.11419677734375,
0.412109375,
0.2275390625,
0.413330078125,
0.2275390625,
0.427490234375,
0.11419677734375,
0.427001953125
],
[
0.51123046875,
0.418701171875,
0.880859375,
0.41748046875,
0.880859375,
0.431640625,
0.51123046875,
0.432861328125
],
[
0.51318359375,
0.435302734375,
0.880859375,
0.435302734375,
0.880859375,
0.448974609375,
0.51318359375,
0.448974609375
],
[
0.118408203125,
0.441162109375,
0.489501953125,
0.441162109375,
0.489501953125,
0.454833984375,
0.118408203125,
0.454833984375
],
[
0.51318359375,
0.45068359375,
0.880859375,
0.45068359375,
0.880859375,
0.46435546875,
0.51318359375,
0.46435546875
],
[
0.1158447265625,
0.457275390625,
0.48779296875,
0.457275390625,
0.48779296875,
0.47021484375,
0.1158447265625,
0.47021484375
],
[
0.5146484375,
0.46728515625,
0.87890625,
0.46728515625,
0.87890625,
0.48095703125,
0.5146484375,
0.48095703125
],
[
0.1175537109375,
0.473876953125,
0.489501953125,
0.473876953125,
0.489501953125,
0.48681640625,
0.1175537109375,
0.48681640625
],
[
0.5146484375,
0.482666015625,
0.880859375,
0.482666015625,
0.880859375,
0.496337890625,
0.5146484375,
0.496337890625
],
[
0.1175537109375,
0.490478515625,
0.48779296875,
0.490478515625,
0.48779296875,
0.50439453125,
0.1175537109375,
0.50439453125
],
[
0.51318359375,
0.49951171875,
0.880859375,
0.49951171875,
0.880859375,
0.51318359375,
0.51318359375,
0.51318359375
],
[
0.118408203125,
0.505859375,
0.48779296875,
0.505859375,
0.48779296875,
0.51953125,
0.118408203125,
0.51953125
],
[
0.51123046875,
0.51416015625,
0.87890625,
0.515625,
0.87890625,
0.52978515625,
0.51123046875,
0.529296875
],
[
0.118408203125,
0.521484375,
0.484375,
0.521484375,
0.484375,
0.5361328125,
0.118408203125,
0.5361328125
],
[
0.509765625,
0.52978515625,
0.880859375,
0.53076171875,
0.880859375,
0.5458984375,
0.509765625,
0.54443359375
],
[
0.1175537109375,
0.53759765625,
0.483642578125,
0.53759765625,
0.483642578125,
0.55029296875,
0.1175537109375,
0.55029296875
],
[
0.51123046875,
0.5458984375,
0.611328125,
0.54736328125,
0.611328125,
0.5625,
0.51123046875,
0.56005859375
],
[
0.118408203125,
0.55419921875,
0.484375,
0.55419921875,
0.484375,
0.56787109375,
0.118408203125,
0.56787109375
],
[
0.12005615234375,
0.57177734375,
0.483642578125,
0.57177734375,
0.483642578125,
0.58203125,
0.12005615234375,
0.58203125
],
[
0.1175537109375,
0.5849609375,
0.486083984375,
0.5849609375,
0.486083984375,
0.5986328125,
0.1175537109375,
0.5986328125
],
[
0.51318359375,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59326171875,
0.51318359375,
0.59326171875
],
[
0.51318359375,
0.59423828125,
0.880859375,
0.595703125,
0.880859375,
0.60986328125,
0.51318359375,
0.609375
],
[
0.1158447265625,
0.60107421875,
0.483642578125,
0.6015625,
0.483642578125,
0.6162109375,
0.1158447265625,
0.615234375
],
[
0.5146484375,
0.6123046875,
0.87890625,
0.6123046875,
0.87890625,
0.6259765625,
0.5146484375,
0.6259765625
],
[
0.118408203125,
0.6181640625,
0.145263671875,
0.6181640625,
0.145263671875,
0.6318359375,
0.118408203125,
0.6318359375
],
[
0.51123046875,
0.62744140625,
0.87890625,
0.62744140625,
0.87890625,
0.64111328125,
0.51123046875,
0.64111328125
],
[
0.51123046875,
0.6435546875,
0.87890625,
0.6435546875,
0.87890625,
0.65673828125,
0.51123046875,
0.65673828125
],
[
0.118408203125,
0.65283203125,
0.48193359375,
0.65283203125,
0.48193359375,
0.6630859375,
0.118408203125,
0.6630859375
],
[
0.51318359375,
0.6611328125,
0.87890625,
0.6611328125,
0.87890625,
0.6748046875,
0.51318359375,
0.6748046875
],
[
0.118408203125,
0.66845703125,
0.486083984375,
0.66845703125,
0.486083984375,
0.681640625,
0.118408203125,
0.681640625
],
[
0.51318359375,
0.6767578125,
0.880859375,
0.6767578125,
0.880859375,
0.6904296875,
0.51318359375,
0.6904296875
],
[
0.118408203125,
0.68408203125,
0.486083984375,
0.68408203125,
0.486083984375,
0.69775390625,
0.118408203125,
0.69775390625
],
[
0.51123046875,
0.693359375,
0.88330078125,
0.693359375,
0.88330078125,
0.70703125,
0.51123046875,
0.70703125
],
[
0.118408203125,
0.70068359375,
0.484375,
0.70068359375,
0.484375,
0.71435546875,
0.118408203125,
0.71435546875
],
[
0.51318359375,
0.708984375,
0.6484375,
0.708984375,
0.6484375,
0.72265625,
0.51318359375,
0.72265625
],
[
0.118408203125,
0.71630859375,
0.484375,
0.71630859375,
0.484375,
0.72998046875,
0.118408203125,
0.72998046875
],
[
0.509765625,
0.7236328125,
0.880859375,
0.724609375,
0.880859375,
0.7392578125,
0.509765625,
0.73828125
],
[
0.118408203125,
0.73291015625,
0.48779296875,
0.73291015625,
0.48779296875,
0.74658203125,
0.118408203125,
0.74658203125
],
[
0.51318359375,
0.7412109375,
0.5859375,
0.7412109375,
0.5859375,
0.7529296875,
0.51318359375,
0.7529296875
],
[
0.118408203125,
0.7490234375,
0.484375,
0.7490234375,
0.484375,
0.7626953125,
0.118408203125,
0.7626953125
],
[
0.1175537109375,
0.76416015625,
0.424072265625,
0.76416015625,
0.424072265625,
0.77783203125,
0.1175537109375,
0.77783203125
],
[
0.51123046875,
0.7685546875,
0.84814453125,
0.771484375,
0.84814453125,
0.7861328125,
0.51123046875,
0.783203125
],
[
0.11419677734375,
0.79443359375,
0.219970703125,
0.79443359375,
0.219970703125,
0.80859375,
0.11419677734375,
0.80859375
],
[
0.51318359375,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.80712890625,
0.51318359375,
0.80712890625
],
[
0.5146484375,
0.81005859375,
0.8759765625,
0.81005859375,
0.8759765625,
0.82080078125,
0.5146484375,
0.82080078125
],
[
0.118408203125,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51123046875,
0.82470703125,
0.87890625,
0.82470703125,
0.87890625,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.118408203125,
0.87353515625,
0.489501953125,
0.87353515625,
0.489501953125,
0.88720703125,
0.118408203125,
0.88720703125
],
[
0.51318359375,
0.87353515625,
0.8818359375,
0.87353515625,
0.8818359375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.118408203125,
0.8896484375,
0.385498046875,
0.8896484375,
0.385498046875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.724609375,
0.8896484375,
0.724609375,
0.90283203125,
0.51318359375,
0.90283203125
]
] |
[
"5.1NLG-oriented",
"L() = -Ey~per(Y)",
"(3)",
"我们报告BLEU（Papineniet al.，2002）和",
"CIDEr（Vedantam etal.，2015）指标以评估世",
"其中r(Y）是对生成的报告的奖励。我们使用",
"代。为了与以前的工作E(Chenetal.,2021)保",
"SCST算法（Rennieetal.，2017）来近似我们的",
"持一致，我们还报告了ROUGE-L度量ELin",
"不可微奖励函数的期望梯度。表达式变为：",
"2004)",
"5.2Factually-oriented",
"VeL(0) ~-(r(Y) -r(Y))Vlogpe(Y) (4)",
"以下提供的度量以不同的方式评估生成的结果",
"的事实正确性和完整性。我们继续描述它们及",
"这里r(Y）充当基线(Suttonetal.，1998），以",
"其差异。",
"减少r(Y）的方差。在我们的例子中，r(Y）是",
"在训练期间从模型中采样的预期奖励。",
"facteNT（Miuraetal.,2021）将命名实体识别",
"器（NER）应用于生成的报告9和相应的引用y",
"超参数我们的模型由1个尺寸为768的变",
"",
"分别给出两组提取的实体E，和Ey。factENT",
"压器块和尺寸为3072的前馈层组成。作为优",
"定义为两组E和E之间的精度和召回率的",
"化器，我们选择Adam（KingmaandBa，2015），",
"调和平均值。Stanza（Qietal.，2020）的临床模",
"学习率为3e-4，小批量大小为128。我们用大",
"型用作NER。",
"小为3的波束搜索进行解码",
"factENTNLI",
"（Miura et al.，2021）此分数是",
"4Datasets",
"factENT与自然语言推理（NLI）的扩展。在这",
"里，如果Ey中存在E，的实体e，则不会自动",
"为了进行我们的实验，我们使用了两个胸部",
"将其视为正确。要被视为有效，包含实体e的",
"X射线数据集：MIMIC-CXRE(Johnsonetal.,",
"句子sg不得与参考报告中的对应句子su相矛",
"2019）和Open-i胸部X射线（印第安纳大学，",
"盾。参考报告中对应的句子su是BERTScore",
"Demner-Fushmanetal.（2012））。对于这两个",
"最高的句子（Zhangetal.，2019）对sg。NLI",
"数据集，我们使用官方拆分，但我们丢弃不",
"模型输出句子su是否与sg矛盾。我们使用",
"包含结果部分的报告。因此，MIMIC-CXR由",
"Miuraetal.（2021）的NLI模型权重，它依赖于",
"152,173训练样本、用于验证的1,196和用于",
"BERT架构。",
"测试的2,347组成；同样，如果我们不计算",
"每个研究中图像的多重性，并且如果我们丢",
"弃。由于这使得数据集太小，无法训练和测试",
"F1CheXbertE(Zhangetal.,2020)此分数使",
"用CheXbertE（Smitetal.，2020），这是一种基",
"RRG模型，因此Open-i数据集仅用于测试目",
"于Transformer的模型，经过训练可在放射学",
"的。",
"报告作为输入的情况下输出胸部X射线的异",
"常（十四个类别）。F，CheXbert是CheXber",
"对于两个数据集的每个样本，我们通过在地面",
"对生成的报告9和相应的参考y的预测之间的",
"真相发现部分上运行CheXbert标签器E(Smit",
"F1分数。为了与以前的工作一致，分数是根据",
"etal.,2020)来生成我们认为的地面真相放射学",
"5个观察值计算的：肺不张，心脏肥大，实变，",
"诊断。这为每个地面实况报告创建了相关的诊",
"水肿和胸腔积液。",
"断标签，该标签描述了14个可能的观察结果",
"RG奖励我们使用2.2节中解释的奖励作为",
"的存在程度（例如。g。报告可能为阳性、阴性、",
"评估分数",
"不确定或未指定的实变或水肿)。此标签用于",
"计算F1CheXbert分数（请参阅5.2节）",
"5.3 RadGraph vs factENT and factENTNLI",
"5Metrics",
"理论上，RGe封装了factENT和factENTNLI概",
"念。事实上，RGE专注于拥有正确的实体和正",
"在本节中，我们将继续报告用于评估模型生成",
"确的实体标签。鉴于实体的标签包含解剖或观",
"的结果与人工编辑的结果的所有指标。我们将",
"察的概念，前者根据定义始终存在，而后者具",
"指标分为两类，NLG文献中广泛报道的NLG",
"有一定程度的存在（“存在”，“不存在”或“不",
"指标，以及面向事实的指标，专门用于评估放",
"确定”），RGe可以以与factENTNLI相同的方式",
"射学报告的事实正确性和完整性。",
"惩罚与参考相矛盾的报告。"
] |
{
"arxiv_id": "2210.12186",
"language": "zh",
"page_idx": 4
}
|
||
1804.05945-zh-1
|
1804.05945-zh-1.jpg
|
# 使用混合机器翻译在语法错误校正中接近人类水平的性能
## Roman Grundkiewicz
## $\textbf{Marcin Junczys- Dowmunt}$ Microsoft
University of Edinburgh
10 Crichton St, Edinburgh EH8 9AB, Scotland
Redmond, WA 98052, USA
rgrundki@inf.ed.ac.uk
marcinjd@microsoft.com
## Abstract
We combine two of the most popular approaches to automated Grammatical Error Correction (GEC): GEC based on Statistical Machine Translation (SMT) and GEC based on Neural Machine Translation (NMT). The hybrid system achieves new state-of-the-art results on the CoNLL- 2014 and JFLEG benchmarks. This GEC system preserves the accuracy of SMT output and, at the same time, generates more fluent sentences as it typical for NMT. Our analysis shows that the created systems are closer to reaching human-level performance than any other GEC system reported so far.
## 1 Introduction
当前，最有效的 GEC 系统基于基于短语的统计机器翻译 (Rozovskaya and Roth, 2016; JunczysDowmunt and Grundkiewicz, 2016; Chollampatt and Ng,2017)。根据自动评估指标，依赖于神经机器翻译 (Yuan and Briscoe, 2016; Xie et al., 2016; Schmaltz et al., 2017; Ji et al., 2017) 的系统尚无法实现与 SMT 系统一样高的性能(有关在 CoNLL-2014 测试集上的比较，请参见表$\operatorname{E1}$ )。然而，已经表明，神经方法可以产生更流畅的输出，这可能是人类评估者 (Napoles et al.,2017) 所希望的。在这项工作中，我们在混合 GEC 系统中结合了两种 MT 口味。这样的 GEC 系统保留了 SMT 输出的准确性，同时生成了更流畅的句子，在两个不同的基准上实现了最新的结果：基于注释的 CoNLL-2014 和基于流利度的 JFLEG 基准。此外，与人类黄金标准的比较表明，所创建的系统比迄今为止文献中描述的任何其他 GEC 系统都更接近达到人类水平的性能
使用一致的训练数据和预处理(§2),我们
首先创建强 SMT(§3)和 NMT(§4)基线
<FigureHere>
Figure 1: Comparison of SMT, NMT and hybrid GEC systems on the CoNLL-2014 test set ( M$^{2}).$
系统。然后，我们通过流水线和重新排序(§5 )对系统组合进行实验。最后，我们将性能与人类注释进行比较，并确定当前最先进的系统$(\S6)$的问题
### 2 $\textbf{Data and preprocessing}$
我们的主要训练数据是 NUCLE (Dahlmeier et al., 2013) 。来自公开的 Lang-8 语料库 (Mizumoto et al., 2012) 的英语句子可作为附加训练数据
我们使用来自 2013年和 2014 年 (Ng et al., 2013,2014)的两个 CoNLL 共享任务的官方测试集作为开发和测试数据，并使用 M$^2$ (Dahlmeier and Ng, 2012) 进行评估。我们还报告了使用 GLEU 度量 (Napoles et al., 2015) 的JFLEG (Napoles et al., 2017) 的结果。数据集提供了开发和测试集拆分。所有数据集在表 1 中列出
我们使用 NLTK 标记器 (Bird and Loper, 2004)对 Lang-8 进行预处理，并在 NUCLE 和JFLEG 中保留原始的标记化。句子是用摩西
|
[
[
0.32421875,
0.08489990234375,
0.67919921875,
0.08489990234375,
0.67919921875,
0.09979248046875,
0.32421875,
0.09979248046875
],
[
0.382080078125,
0.1051025390625,
0.62109375,
0.1051025390625,
0.62109375,
0.119384765625,
0.382080078125,
0.119384765625
],
[
0.240966796875,
0.149658203125,
0.432373046875,
0.149658203125,
0.432373046875,
0.163330078125,
0.240966796875,
0.163330078125
],
[
0.55322265625,
0.149658203125,
0.7841796875,
0.149658203125,
0.7841796875,
0.1644287109375,
0.55322265625,
0.1644287109375
],
[
0.2376708984375,
0.166259765625,
0.435791015625,
0.166259765625,
0.435791015625,
0.18115234375,
0.2376708984375,
0.18115234375
],
[
0.62646484375,
0.1650390625,
0.7109375,
0.1650390625,
0.7109375,
0.179931640625,
0.62646484375,
0.179931640625
],
[
0.1485595703125,
0.182861328125,
0.5283203125,
0.182861328125,
0.5283203125,
0.196533203125,
0.1485595703125,
0.196533203125
],
[
0.55322265625,
0.182861328125,
0.78271484375,
0.182861328125,
0.78271484375,
0.196533203125,
0.55322265625,
0.196533203125
],
[
0.2091064453125,
0.20068359375,
0.46337890625,
0.20068359375,
0.46337890625,
0.21435546875,
0.2091064453125,
0.21435546875
],
[
0.5341796875,
0.201904296875,
0.8037109375,
0.201904296875,
0.8037109375,
0.2120361328125,
0.5341796875,
0.2120361328125
],
[
0.599609375,
0.26416015625,
0.63623046875,
0.26416015625,
0.63623046875,
0.276611328125,
0.599609375,
0.276611328125
],
[
0.69189453125,
0.26416015625,
0.73291015625,
0.26416015625,
0.73291015625,
0.27783203125,
0.69189453125,
0.27783203125
],
[
0.77734375,
0.26318359375,
0.82861328125,
0.26318359375,
0.82861328125,
0.27783203125,
0.77734375,
0.27783203125
],
[
0.54150390625,
0.28076171875,
0.5634765625,
0.28076171875,
0.5634765625,
0.29443359375,
0.54150390625,
0.29443359375
],
[
0.26611328125,
0.265380859375,
0.343505859375,
0.265380859375,
0.343505859375,
0.27978515625,
0.26611328125,
0.27978515625
],
[
0.145263671875,
0.292236328125,
0.4609375,
0.29345703125,
0.4609375,
0.30810546875,
0.145263671875,
0.30712890625
],
[
0.8330078125,
0.3017578125,
0.8447265625,
0.3017578125,
0.8447265625,
0.325927734375,
0.8330078125,
0.325927734375
],
[
0.145263671875,
0.31005859375,
0.4609375,
0.308837890625,
0.4609375,
0.32373046875,
0.145263671875,
0.32470703125
],
[
0.54248046875,
0.30810546875,
0.56689453125,
0.30810546875,
0.56689453125,
0.32177734375,
0.54248046875,
0.32177734375
],
[
0.14697265625,
0.32666015625,
0.4609375,
0.32666015625,
0.4609375,
0.34033203125,
0.14697265625,
0.34033203125
],
[
0.8095703125,
0.322509765625,
0.8193359375,
0.322509765625,
0.8193359375,
0.33837890625,
0.8095703125,
0.33837890625
],
[
0.14697265625,
0.342529296875,
0.459228515625,
0.342529296875,
0.459228515625,
0.355712890625,
0.14697265625,
0.355712890625
],
[
0.5439453125,
0.341552734375,
0.56689453125,
0.341552734375,
0.56689453125,
0.355224609375,
0.5439453125,
0.355224609375
],
[
0.634765625,
0.344482421875,
0.64404296875,
0.344482421875,
0.64404296875,
0.35400390625,
0.634765625,
0.35400390625
],
[
0.7841796875,
0.3466796875,
0.79443359375,
0.3466796875,
0.79443359375,
0.35400390625,
0.7841796875,
0.35400390625
],
[
0.611328125,
0.352783203125,
0.6171875,
0.352783203125,
0.6171875,
0.359375,
0.611328125,
0.359375
],
[
0.1485595703125,
0.35986328125,
0.459228515625,
0.35986328125,
0.459228515625,
0.37060546875,
0.1485595703125,
0.37060546875
],
[
0.634765625,
0.355712890625,
0.640625,
0.355712890625,
0.640625,
0.361083984375,
0.634765625,
0.361083984375
],
[
0.14697265625,
0.374755859375,
0.462646484375,
0.374755859375,
0.462646484375,
0.388427734375,
0.14697265625,
0.388427734375
],
[
0.5439453125,
0.3759765625,
0.56689453125,
0.3759765625,
0.56689453125,
0.3896484375,
0.5439453125,
0.3896484375
],
[
0.14697265625,
0.39013671875,
0.4609375,
0.39013671875,
0.4609375,
0.40380859375,
0.14697265625,
0.40380859375
],
[
0.14697265625,
0.40673828125,
0.4609375,
0.40673828125,
0.4609375,
0.42041015625,
0.14697265625,
0.42041015625
],
[
0.5458984375,
0.4091796875,
0.56689453125,
0.4091796875,
0.56689453125,
0.4228515625,
0.5458984375,
0.4228515625
],
[
0.14697265625,
0.42333984375,
0.4609375,
0.42333984375,
0.4609375,
0.43701171875,
0.14697265625,
0.43701171875
],
[
0.14697265625,
0.439453125,
0.4609375,
0.439453125,
0.4609375,
0.453125,
0.14697265625,
0.453125
],
[
0.580078125,
0.45068359375,
0.59619140625,
0.45068359375,
0.59619140625,
0.457763671875,
0.580078125,
0.457763671875
],
[
0.67822265625,
0.453125,
0.68994140625,
0.453125,
0.68994140625,
0.4560546875,
0.67822265625,
0.4560546875
],
[
0.58837890625,
0.46142578125,
0.60693359375,
0.46142578125,
0.60693359375,
0.465576171875,
0.58837890625,
0.465576171875
],
[
0.14697265625,
0.454833984375,
0.459228515625,
0.454833984375,
0.459228515625,
0.4697265625,
0.14697265625,
0.4697265625
],
[
0.6708984375,
0.4619140625,
0.7119140625,
0.4619140625,
0.7119140625,
0.466064453125,
0.6708984375,
0.466064453125
],
[
0.14697265625,
0.471435546875,
0.4609375,
0.471435546875,
0.4609375,
0.485107421875,
0.14697265625,
0.485107421875
],
[
0.5498046875,
0.4755859375,
0.55908203125,
0.4755859375,
0.55908203125,
0.48095703125,
0.5498046875,
0.48095703125
],
[
0.1485595703125,
0.4892578125,
0.459228515625,
0.4892578125,
0.459228515625,
0.49951171875,
0.1485595703125,
0.49951171875
],
[
0.14697265625,
0.50439453125,
0.4609375,
0.50439453125,
0.4609375,
0.51708984375,
0.14697265625,
0.51708984375
],
[
0.5146484375,
0.5126953125,
0.88330078125,
0.5126953125,
0.88330078125,
0.525390625,
0.5146484375,
0.525390625
],
[
0.145263671875,
0.51953125,
0.2626953125,
0.5185546875,
0.2626953125,
0.533203125,
0.145263671875,
0.53369140625
],
[
0.5146484375,
0.52783203125,
0.873046875,
0.52783203125,
0.873046875,
0.54150390625,
0.5146484375,
0.54150390625
],
[
0.12005615234375,
0.548828125,
0.133544921875,
0.548828125,
0.133544921875,
0.55859375,
0.12005615234375,
0.55859375
],
[
0.15283203125,
0.548828125,
0.264404296875,
0.548828125,
0.264404296875,
0.56005859375,
0.15283203125,
0.56005859375
],
[
0.12005615234375,
0.572265625,
0.48779296875,
0.57177734375,
0.48779296875,
0.5859375,
0.12005615234375,
0.58740234375
],
[
0.51708984375,
0.5693359375,
0.88330078125,
0.5693359375,
0.88330078125,
0.5830078125,
0.51708984375,
0.5830078125
],
[
0.118408203125,
0.58837890625,
0.489501953125,
0.58837890625,
0.489501953125,
0.6015625,
0.118408203125,
0.6015625
],
[
0.51318359375,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.51318359375,
0.59716796875
],
[
0.1175537109375,
0.60400390625,
0.489501953125,
0.60498046875,
0.489501953125,
0.619140625,
0.1175537109375,
0.6181640625
],
[
0.51708984375,
0.60107421875,
0.8818359375,
0.60107421875,
0.8818359375,
0.615234375,
0.51708984375,
0.615234375
],
[
0.118408203125,
0.62158203125,
0.486083984375,
0.62158203125,
0.486083984375,
0.63525390625,
0.118408203125,
0.63525390625
],
[
0.5146484375,
0.6181640625,
0.623046875,
0.6181640625,
0.623046875,
0.6298828125,
0.5146484375,
0.6298828125
],
[
0.118408203125,
0.63525390625,
0.489501953125,
0.6357421875,
0.489501953125,
0.65087890625,
0.118408203125,
0.6494140625
],
[
0.515625,
0.646484375,
0.52880859375,
0.646484375,
0.52880859375,
0.65478515625,
0.515625,
0.65478515625
],
[
0.5439453125,
0.6435546875,
0.75390625,
0.64453125,
0.75390625,
0.6591796875,
0.5439453125,
0.65771484375
],
[
0.12005615234375,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.66650390625,
0.12005615234375,
0.66650390625
],
[
0.12005615234375,
0.66845703125,
0.486083984375,
0.66845703125,
0.486083984375,
0.681640625,
0.12005615234375,
0.681640625
],
[
0.5146484375,
0.66943359375,
0.88330078125,
0.66943359375,
0.88330078125,
0.68310546875,
0.5146484375,
0.68310546875
],
[
0.12176513671875,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69970703125,
0.12176513671875,
0.69970703125
],
[
0.51318359375,
0.68505859375,
0.88330078125,
0.68505859375,
0.88330078125,
0.69970703125,
0.51318359375,
0.69970703125
],
[
0.12176513671875,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.12176513671875,
0.71533203125
],
[
0.5146484375,
0.70166015625,
0.88330078125,
0.70166015625,
0.88330078125,
0.71533203125,
0.5146484375,
0.71533203125
],
[
0.12176513671875,
0.71875,
0.489501953125,
0.71875,
0.489501953125,
0.732421875,
0.12176513671875,
0.732421875
],
[
0.515625,
0.71875,
0.55322265625,
0.71875,
0.55322265625,
0.73095703125,
0.515625,
0.73095703125
],
[
0.12005615234375,
0.73388671875,
0.48779296875,
0.73388671875,
0.48779296875,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.5322265625,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.5322265625,
0.74755859375
],
[
0.12176513671875,
0.75,
0.48779296875,
0.75,
0.48779296875,
0.76416015625,
0.12176513671875,
0.76416015625
],
[
0.51318359375,
0.7490234375,
0.8818359375,
0.75,
0.8818359375,
0.76416015625,
0.51318359375,
0.76318359375
],
[
0.118408203125,
0.765625,
0.486083984375,
0.765625,
0.486083984375,
0.779296875,
0.118408203125,
0.779296875
],
[
0.5146484375,
0.765625,
0.88330078125,
0.76416015625,
0.88330078125,
0.779296875,
0.5146484375,
0.77978515625
],
[
0.118408203125,
0.78076171875,
0.486083984375,
0.78076171875,
0.486083984375,
0.79443359375,
0.118408203125,
0.79443359375
],
[
0.51123046875,
0.78076171875,
0.880859375,
0.78076171875,
0.880859375,
0.79443359375,
0.51123046875,
0.79443359375
],
[
0.118408203125,
0.79736328125,
0.489501953125,
0.79638671875,
0.489501953125,
0.81103515625,
0.118408203125,
0.8125
],
[
0.51318359375,
0.79638671875,
0.8818359375,
0.79736328125,
0.8818359375,
0.8125,
0.51318359375,
0.81103515625
],
[
0.118408203125,
0.81396484375,
0.48779296875,
0.81396484375,
0.48779296875,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.51318359375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51318359375,
0.82763671875
],
[
0.12005615234375,
0.830078125,
0.489501953125,
0.830078125,
0.489501953125,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.5146484375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.118408203125,
0.8466796875,
0.48779296875,
0.845703125,
0.48779296875,
0.85986328125,
0.118408203125,
0.86083984375
],
[
0.5146484375,
0.84716796875,
0.5517578125,
0.84716796875,
0.5517578125,
0.85986328125,
0.5146484375,
0.85986328125
],
[
0.12005615234375,
0.86328125,
0.2509765625,
0.86328125,
0.2509765625,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.5322265625,
0.8623046875,
0.88330078125,
0.8623046875,
0.88330078125,
0.876953125,
0.5322265625,
0.876953125
],
[
0.1376953125,
0.87744140625,
0.486083984375,
0.87744140625,
0.486083984375,
0.892578125,
0.1376953125,
0.892578125
],
[
0.5146484375,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.892578125,
0.5146484375,
0.892578125
],
[
0.12176513671875,
0.89453125,
0.489501953125,
0.89453125,
0.489501953125,
0.9091796875,
0.12176513671875,
0.9091796875
],
[
0.51123046875,
0.8935546875,
0.88330078125,
0.89453125,
0.88330078125,
0.9091796875,
0.51123046875,
0.90771484375
]
] |
[
"使用混合机器翻译在语法错误校正",
"中接近人类水平的性能",
"Roman Grundkiewicz",
"Marcin Junczys-Dowmunt",
"University of Edinburgh",
"Microsoft",
"10 Crichton St, Edinburgh EH8 9AB, Scotland",
"Redmond, WA 98052, USA",
"rgrundki@inf.ed.ac.uk",
"marcinid@microsoft.com",
"SMT",
"NMT",
"Hybrid",
"M?",
"Abstract",
"We combine two of the most popular ap-",
"..",
"proaches to automated Grammatical Er-",
"55",
"ror Correction (GEC): GEC based on Sta-",
"0C",
"tistical Machine Translation (SMT) and",
"50",
"",
"O",
"0",
"GEC based on Neural Machine Transla",
"6",
"tion (NMT). The hybrid system achieves",
"45",
"new state-of-the-art results on the CoNLL-",
"2014 and JFLEG benchmarks. This GEC",
"40",
"system preserves the accuracy of SMT out-",
"put and, at the same time, generates more",
"16",
"",
"",
"fluent sentences as it typical for NMT.",
"Q",
"Our analysis shows that the created sys-",
"",
"tems are closer to reaching human-level",
"performance than any other GEC system",
"Figure 1: Comparison of SMT, NMT and hybrid",
"reported so far.",
"GEC systems on the CoNLL-2014 test set (M?).",
"1",
"Introduction",
"当前，最有效的GEC系统基于基于短语的统计",
"系统。然后，我们通过流水线和重新排序（S5",
"机器翻译（RozovskayaandRoth，2016:Junczys-",
"）对系统组合进行实验。最后，我们将性能与",
"Dowmunt and Grundkiewicz, 2016; Chollampatt",
"人类注释进行比较，并确定当前最先进的系统",
"andNg，2017）。根据自动评估指标，依赖于神",
"（86）的问题",
"经机器翻译（YuanandBriscoe,2016;Xieetal.,",
"2",
"Data andpreprocessing",
"2016; Schmaltz et al., 2017; Ji et al., 2017)的系",
"统尚无法实现与SMT系统一样高的性能（有",
"我们的主要训练数据是NUCLE（Dahlmeier",
"关在CoNLL-2014测试集上的比较，请参见",
"etal.,2013）。来自公开的Lang-8语料库（Mizu",
"表E1）。然而，已经表明，神经方法可以产生",
"motoetal.,2012）的英语句子可作为附加训练",
"更流畅的输出，这可能是人类评估者(Napoles",
"数据",
"etal.,2017)所希望的。在这项工作中，我们在",
"我们使用来自2013年和2014年（Ngetal.,",
"混合GEC系统中结合了两种MT口味。这样",
"2013，2014）的两个CoNLL共享任务的官方",
"的GEC系统保留了SMT输出的准确性，同时",
"测试集作为开发和测试数据，并使用M?",
"生成了更流畅的句子，在两个不同的基准上实",
"(DahlmeierandNg，2012）进行评估。我们还报",
"现了最新的结果：基于注释的CoNLL-2014和",
"告了使用GLEU度量(Napolesetal.,2015）的",
"基于流利度的JFLEG基准。此外，与人类黄金",
"JFLEG(Napolesetal.，2017）的结果。数据集提",
"标准的比较表明，所创建的系统比迄今为止文",
"供了开发和测试集拆分。所有数据集在表1中",
"献中描述的任何其他GEC系统都更接近达到",
"列出",
"人类水平的性能",
"我们使用NLTK标记器（Bird andLoper,",
"使用一致的训练数据和预处理（s2），我们",
"2004）对Lang-8进行预处理，并在NUCLE和",
"首先创建强SMT（$3）和NMT（84）基线",
"JFLEG中保留原始的标记化。句子是用摩西"
] |
{
"arxiv_id": "1804.05945",
"language": "zh",
"page_idx": 1
}
|
||
1804.05945-zh-2
|
1804.05945-zh-2.jpg
|
<table>
<tbody>
<tr>
<th>Corpus</th>
<th>Sentences</th>
<th>Tokens</th>
</tr>
<tr>
<td>NUCLE</td>
<td>57,151</td>
<td>$1,162\mathbf{K}$</td>
</tr>
<tr>
<td>Lang- 8NAIST</td>
<td>1.943,901</td>
<td>25,026K</td>
</tr>
<tr>
<td>CoNLL- 2013 ( dev) </td>
<td>1,381</td>
<td>$29K$</td>
</tr>
<tr>
<td>CoNLL- 2014 $-(test)$</td>
<td>1,312</td>
<td>$30K$</td>
</tr>
<tr>
<td>$\mathbf{JFLEG}$ Dev</td>
<td>754</td>
<td>$14K$</td>
</tr>
<tr>
<td>JFLEG iTest </td>
<td>747</td>
<td>$13K$</td>
</tr>
</tbody>
</table>
Table 1: Statistics for training and testing data sets.
(Koehn et al., 2007) 的脚本编写的。为了处理词汇表外的单词，我们使用 Sennrich et al. (2016b) 的字节对编码(BPE)将令牌拆分为50k个子单词单元。BPE 代码仅从 Lang-8 和 NUCLE 的正确句子中提取
# 3 SMT systems
对于我们基于 SMT 的系统，我们遵循 JunczysDowmunt and Grundkiewicz (2016) 提出的配方. 并使用基于短语的 SMT 系统，该系统具有特定于任务的特征的对数线性组合。我们使用单词级 Levenshtein 距离和编辑操作计数作为密集特征(密集),并将单词类(WC)上具有一个单词左/右上下文的单词的校正模式作为稀疏特征(稀疏)。我们还试验了额外的字符级密集特征 (Char。ops)。所有系统都使用从训练数据的目标侧估计的 5 克语言模型 (LM) 和OSM (Durrani et al., 2011) , 以及在通用爬网数据(Buck et al., 2014) 上训练的 5 克 LM 和 9 克$WCLM$
Experiment settings 使用 Moses (Koehn et al., 2007)训练翻译模型，使用 MGIZA++(Gao and Vogel,2008) 生成单词对齐模型，并且不使用重新排序模型。语言模型是使用 KenLM (Heafield, 2011)构建的，而单词类是使用 word2vec$^{1}$训练的
我们分别针对 M$^{2}$和 GLEU 指标调整系统。MERT (Och, 2003) 用于调整密集特征，BatchMira (Cherry and Foster, 2012) 用于调整稀疏特征。对于$\dot{\mathbf{M}^{2}}$调谐，我们遵循 NUCLE 上的4 倍交叉验证，并采用 Junczys-Dowmunt and Grundkiewicz (2016) 推荐的自适应错误率。在 GLEU 上评估的模型在 JFLEGDev 上使用 GLEU 评分器进行优化，我们将其添加到 Moses。我们使用 4 次调谐运行的平均特征权重报告模型的结果
Results 在其他条件相同的情况下，使用原始标记化，应用子字单元以及扩展基于编辑
的功能会导致与 Junczys-Dowmunt and Grund-
$^{1}$https: / / github. com/ dav/ word2vec
<table>
<tbody>
<tr>
<th>System</th>
<th>$P$</th>
<th>CoNLL $R$</th>
<th>$M^{2}$</th>
<th>JFLEG GLEU</th>
</tr>
<tr>
<th>System</th>
<th>$P$</th>
<th>CoNLL $R$</th>
<th>$M^{2}$</th>
<th>JFLEG GLEU</th>
</tr>
<tr>
<td>$SMT$ Dense </td>
<td>56.91</td>
<td>30.25</td>
<td>48.38</td>
<td>54.68</td>
</tr>
<tr>
<td>Sparse +</td>
<td>60.28</td>
<td>29.40</td>
<td>49.82</td>
<td>55.25</td>
</tr>
<tr>
<td>$Char.$ + $0ps$</td>
<td>60.27</td>
<td>30.21</td>
<td>50.27</td>
<td>55.79</td>
</tr>
</tbody>
</table>
Table 2: Results for SMT baseline systems on the CoNLL-2014( M^2) and JFLEG Test(GLEU) sets.
kiewicz (2016):49类似的系统。82vs49。49 M$^2$
(表 2)
基于短语的 SMT 系统不能很好地处理正交错误 (Napoles et al., 2017) --如果在训练语料库中未看到源词，则很可能将其复制为目标词。子字单元可以帮助部分解决这个问题。添加基于字符级编辑计数的功能会增加两个测试集上的结果
结果为 55。79JFLEG 测试上的 GLEU 已经比 Sakaguchi et al. (2017)的 GLEU 调整的 NMT 系统好 2 分，并且比 Chollampatt and Ng (2017) 使用其 M$^{2}$调整的 SMT 系统的最佳报告结果差 1 分，即使此时未使用其他拼写校正。我们在后面的部分中尝试使用专门的拼写检查方法
### 4 $\mathbf{NMT}$ svstems
我们为基于 NMT 的系统选择的模型架构是具有双向单层编码器和解码器的引人注目的编码器-解码器模型，两者均使用 GRU 作为其 RNN 变体 (Sennrich et al., 2017) 。Sakaguchi et al. (2017) 已经针对 GEC 任务测试了类似的体系结构，但是我们使用了不同的超参数
为了提高我们的 NMT 模型的性能，与 Xie et al.(2016) 和 Ji et al.(2017)类似，我们将它们与附加的大规模语言模型结合在一起。与以前使用$n$ -gram 概率 LM 的研究相反，我们使用 GRU 单元构建了一个 2 层递归神经网络语言模型 (RNNLM),并在英语通用爬行数据(Buck et al., 2014) 上再次对其进行了训练
Experimental settings 我们使用 Marian 工具包 (Junczys-Dowmunt et al., 2018)在用于 SMT 基线的相同数据上进行训练，即。e。NUCLE 和 Lang-8。RNN 隐藏状态大小设置为 1024, 嵌人大小为 512。源和目标词汇表以及子词单元相同
使用 Adam (Kingma and Ba, 2014) 和安装在4GBGPU 内存中的小批量大小进行优化。我们用比例丢失(Gal and Ghahramani, 2016) 正则化模型，丢失概率为 0。2 在所有 RNN 输入和状态。除此之外，我们放弃了概率为 0 的整个源词和目标词。2 和0。1 分别。基于 CoNLL-2013测试集上的交叉熵成本，我们以 10 的耐心使
|
[
[
0.39208984375,
0.07659912109375,
0.443359375,
0.07659912109375,
0.443359375,
0.0914306640625,
0.39208984375,
0.0914306640625
],
[
0.7021484375,
0.07781982421875,
0.75244140625,
0.07781982421875,
0.75244140625,
0.08905029296875,
0.7021484375,
0.08905029296875
],
[
0.8017578125,
0.07781982421875,
0.8486328125,
0.07781982421875,
0.8486328125,
0.08905029296875,
0.8017578125,
0.08905029296875
],
[
0.3046875,
0.0789794921875,
0.366943359375,
0.0789794921875,
0.366943359375,
0.09027099609375,
0.3046875,
0.09027099609375
],
[
0.166259765625,
0.0789794921875,
0.214111328125,
0.0789794921875,
0.214111328125,
0.09027099609375,
0.166259765625,
0.09027099609375
],
[
0.54736328125,
0.0885009765625,
0.59765625,
0.09027099609375,
0.59619140625,
0.1051025390625,
0.5458984375,
0.10272216796875
],
[
0.75830078125,
0.09027099609375,
0.7841796875,
0.09027099609375,
0.7841796875,
0.10272216796875,
0.75830078125,
0.10272216796875
],
[
0.80029296875,
0.08905029296875,
0.84814453125,
0.08905029296875,
0.84814453125,
0.10394287109375,
0.80029296875,
0.10394287109375
],
[
0.7197265625,
0.0926513671875,
0.736328125,
0.0926513671875,
0.736328125,
0.1015625,
0.7197265625,
0.1015625
],
[
0.166259765625,
0.09320068359375,
0.2216796875,
0.0955810546875,
0.219970703125,
0.1092529296875,
0.16455078125,
0.10687255859375
],
[
0.32421875,
0.09320068359375,
0.365234375,
0.0955810546875,
0.364501953125,
0.10809326171875,
0.322509765625,
0.105712890625
],
[
0.391357421875,
0.0968017578125,
0.43994140625,
0.0968017578125,
0.43994140625,
0.10809326171875,
0.391357421875,
0.10809326171875
],
[
0.6767578125,
0.0914306640625,
0.69189453125,
0.0914306640625,
0.69189453125,
0.1015625,
0.6767578125,
0.1015625
],
[
0.5458984375,
0.105712890625,
0.62646484375,
0.10809326171875,
0.62646484375,
0.122314453125,
0.5458984375,
0.12054443359375
],
[
0.79833984375,
0.10687255859375,
0.8447265625,
0.10687255859375,
0.8447265625,
0.1217041015625,
0.79833984375,
0.1217041015625
],
[
0.3837890625,
0.10809326171875,
0.43994140625,
0.10809326171875,
0.43994140625,
0.119384765625,
0.3837890625,
0.119384765625
],
[
0.306396484375,
0.10809326171875,
0.365234375,
0.10809326171875,
0.365234375,
0.119384765625,
0.306396484375,
0.119384765625
],
[
0.6689453125,
0.10809326171875,
0.79833984375,
0.10809326171875,
0.79833984375,
0.1217041015625,
0.6689453125,
0.1217041015625
],
[
0.791015625,
0.114013671875,
0.79833984375,
0.1092529296875,
0.80419921875,
0.1134033203125,
0.796875,
0.1181640625
],
[
0.166259765625,
0.1181640625,
0.287109375,
0.1181640625,
0.287109375,
0.1318359375,
0.166259765625,
0.1318359375
],
[
0.330078125,
0.119384765625,
0.366943359375,
0.119384765625,
0.366943359375,
0.1318359375,
0.330078125,
0.1318359375
],
[
0.41064453125,
0.119384765625,
0.43994140625,
0.119384765625,
0.43994140625,
0.1318359375,
0.41064453125,
0.1318359375
],
[
0.6708984375,
0.1217041015625,
0.75146484375,
0.1217041015625,
0.75146484375,
0.133056640625,
0.6708984375,
0.133056640625
],
[
0.75146484375,
0.1217041015625,
0.79443359375,
0.1217041015625,
0.79443359375,
0.133056640625,
0.75146484375,
0.133056640625
],
[
0.80615234375,
0.1217041015625,
0.84228515625,
0.1217041015625,
0.84228515625,
0.133056640625,
0.80615234375,
0.133056640625
],
[
0.56689453125,
0.122314453125,
0.62060546875,
0.122314453125,
0.62060546875,
0.133056640625,
0.56689453125,
0.133056640625
],
[
0.16796875,
0.1092529296875,
0.26025390625,
0.1092529296875,
0.26025390625,
0.119384765625,
0.16796875,
0.119384765625
],
[
0.166259765625,
0.1312255859375,
0.285400390625,
0.1318359375,
0.285400390625,
0.1436767578125,
0.166259765625,
0.142578125
],
[
0.330078125,
0.1300048828125,
0.36865234375,
0.1318359375,
0.366943359375,
0.144287109375,
0.328369140625,
0.142578125
],
[
0.41064453125,
0.1312255859375,
0.43994140625,
0.1312255859375,
0.43994140625,
0.1436767578125,
0.41064453125,
0.1436767578125
],
[
0.580078125,
0.133056640625,
0.6533203125,
0.133056640625,
0.6533203125,
0.144287109375,
0.580078125,
0.144287109375
],
[
0.6689453125,
0.133056640625,
0.7958984375,
0.133056640625,
0.7958984375,
0.1436767578125,
0.6689453125,
0.1436767578125
],
[
0.80615234375,
0.1318359375,
0.8427734375,
0.1318359375,
0.8427734375,
0.1436767578125,
0.80615234375,
0.1436767578125
],
[
0.16796875,
0.1436767578125,
0.2418212890625,
0.1436767578125,
0.2418212890625,
0.155029296875,
0.16796875,
0.155029296875
],
[
0.340087890625,
0.1436767578125,
0.36865234375,
0.1436767578125,
0.36865234375,
0.155029296875,
0.340087890625,
0.155029296875
],
[
0.411376953125,
0.1436767578125,
0.43994140625,
0.1436767578125,
0.43994140625,
0.155029296875,
0.411376953125,
0.155029296875
],
[
0.341796875,
0.155029296875,
0.36865234375,
0.155029296875,
0.36865234375,
0.16748046875,
0.341796875,
0.16748046875
],
[
0.411376953125,
0.155029296875,
0.441650390625,
0.155029296875,
0.441650390625,
0.16748046875,
0.411376953125,
0.16748046875
],
[
0.51123046875,
0.157958984375,
0.8818359375,
0.1591796875,
0.8818359375,
0.177001953125,
0.51123046875,
0.17578125
],
[
0.515625,
0.1787109375,
0.87890625,
0.1787109375,
0.87890625,
0.189453125,
0.515625,
0.189453125
],
[
0.166259765625,
0.1561279296875,
0.2418212890625,
0.1561279296875,
0.2418212890625,
0.16748046875,
0.166259765625,
0.16748046875
],
[
0.118408203125,
0.18408203125,
0.484375,
0.18408203125,
0.484375,
0.19775390625,
0.118408203125,
0.19775390625
],
[
0.51318359375,
0.218505859375,
0.8818359375,
0.21728515625,
0.8818359375,
0.232177734375,
0.51318359375,
0.2327880859375
],
[
0.118408203125,
0.226806640625,
0.48779296875,
0.226806640625,
0.48779296875,
0.240478515625,
0.118408203125,
0.240478515625
],
[
0.12005615234375,
0.2423095703125,
0.48779296875,
0.2423095703125,
0.48779296875,
0.255859375,
0.12005615234375,
0.255859375
],
[
0.5146484375,
0.2369384765625,
0.5693359375,
0.2369384765625,
0.5693359375,
0.248779296875,
0.5146484375,
0.248779296875
],
[
0.5341796875,
0.251708984375,
0.8818359375,
0.251708984375,
0.8818359375,
0.265380859375,
0.5341796875,
0.265380859375
],
[
0.12005615234375,
0.2587890625,
0.48779296875,
0.2587890625,
0.48779296875,
0.2724609375,
0.12005615234375,
0.2724609375
],
[
0.5146484375,
0.2666015625,
0.880859375,
0.2666015625,
0.880859375,
0.27978515625,
0.5146484375,
0.27978515625
],
[
0.12176513671875,
0.275634765625,
0.48779296875,
0.275634765625,
0.48779296875,
0.289306640625,
0.12176513671875,
0.289306640625
],
[
0.515625,
0.283203125,
0.880859375,
0.283203125,
0.880859375,
0.29638671875,
0.515625,
0.29638671875
],
[
0.12005615234375,
0.29150390625,
0.2493896484375,
0.29150390625,
0.2493896484375,
0.3046875,
0.12005615234375,
0.3046875
],
[
0.51708984375,
0.2998046875,
0.88330078125,
0.2998046875,
0.88330078125,
0.3134765625,
0.51708984375,
0.3134765625
],
[
0.1175537109375,
0.319580078125,
0.270263671875,
0.320556640625,
0.270263671875,
0.3349609375,
0.1175537109375,
0.334228515625
],
[
0.515625,
0.314208984375,
0.8818359375,
0.31640625,
0.8818359375,
0.330078125,
0.515625,
0.327880859375
],
[
0.5146484375,
0.33203125,
0.57080078125,
0.33203125,
0.57080078125,
0.3466796875,
0.5146484375,
0.3466796875
],
[
0.12005615234375,
0.347412109375,
0.489501953125,
0.347412109375,
0.489501953125,
0.3623046875,
0.12005615234375,
0.3623046875
],
[
0.5322265625,
0.347412109375,
0.88330078125,
0.347412109375,
0.88330078125,
0.3623046875,
0.5322265625,
0.3623046875
],
[
0.12005615234375,
0.364013671875,
0.490234375,
0.364013671875,
0.490234375,
0.377685546875,
0.12005615234375,
0.377685546875
],
[
0.5146484375,
0.363525390625,
0.8818359375,
0.363525390625,
0.8818359375,
0.377197265625,
0.5146484375,
0.377197265625
],
[
0.12005615234375,
0.38134765625,
0.48779296875,
0.38134765625,
0.48779296875,
0.394287109375,
0.12005615234375,
0.394287109375
],
[
0.515625,
0.380126953125,
0.8818359375,
0.380126953125,
0.8818359375,
0.394287109375,
0.515625,
0.394287109375
],
[
0.12176513671875,
0.396728515625,
0.48779296875,
0.396728515625,
0.48779296875,
0.410400390625,
0.12176513671875,
0.410400390625
],
[
0.5146484375,
0.396728515625,
0.880859375,
0.396728515625,
0.880859375,
0.410400390625,
0.5146484375,
0.410400390625
],
[
0.12176513671875,
0.412109375,
0.489501953125,
0.412109375,
0.489501953125,
0.42578125,
0.12176513671875,
0.42578125
],
[
0.51318359375,
0.410888671875,
0.8818359375,
0.412109375,
0.8818359375,
0.427001953125,
0.51318359375,
0.42578125
],
[
0.12176513671875,
0.427490234375,
0.47265625,
0.427490234375,
0.47265625,
0.441162109375,
0.12176513671875,
0.441162109375
],
[
0.466796875,
0.431640625,
0.486083984375,
0.431640625,
0.486083984375,
0.439453125,
0.466796875,
0.439453125
],
[
0.5146484375,
0.4287109375,
0.8818359375,
0.4287109375,
0.8818359375,
0.4423828125,
0.5146484375,
0.4423828125
],
[
0.12005615234375,
0.44482421875,
0.486083984375,
0.44482421875,
0.486083984375,
0.457763671875,
0.12005615234375,
0.457763671875
],
[
0.118408203125,
0.460205078125,
0.486083984375,
0.460205078125,
0.486083984375,
0.473876953125,
0.118408203125,
0.473876953125
],
[
0.515625,
0.460205078125,
0.52880859375,
0.460205078125,
0.52880859375,
0.468505859375,
0.515625,
0.468505859375
],
[
0.5263671875,
0.460205078125,
0.54833984375,
0.460205078125,
0.54833984375,
0.46728515625,
0.5263671875,
0.46728515625
],
[
0.54736328125,
0.4560546875,
0.66748046875,
0.457275390625,
0.66748046875,
0.471435546875,
0.54736328125,
0.47021484375
],
[
0.12005615234375,
0.47802734375,
0.48779296875,
0.47802734375,
0.48779296875,
0.4912109375,
0.12005615234375,
0.4912109375
],
[
0.5146484375,
0.482666015625,
0.880859375,
0.482666015625,
0.880859375,
0.496337890625,
0.5146484375,
0.496337890625
],
[
0.12005615234375,
0.493408203125,
0.489501953125,
0.493408203125,
0.489501953125,
0.50732421875,
0.12005615234375,
0.50732421875
],
[
0.515625,
0.49951171875,
0.8818359375,
0.49951171875,
0.8818359375,
0.51318359375,
0.515625,
0.51318359375
],
[
0.12005615234375,
0.5087890625,
0.489501953125,
0.5087890625,
0.489501953125,
0.5224609375,
0.12005615234375,
0.5224609375
],
[
0.5146484375,
0.515625,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.5146484375,
0.529296875
],
[
0.12005615234375,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5390625,
0.12005615234375,
0.5390625
],
[
0.51318359375,
0.52978515625,
0.88330078125,
0.5322265625,
0.88330078125,
0.5458984375,
0.51318359375,
0.54345703125
],
[
0.12005615234375,
0.54150390625,
0.17626953125,
0.54150390625,
0.17626953125,
0.552734375,
0.12005615234375,
0.552734375
],
[
0.5146484375,
0.54736328125,
0.8818359375,
0.54736328125,
0.8818359375,
0.56103515625,
0.5146484375,
0.56103515625
],
[
0.1234130859375,
0.568359375,
0.48779296875,
0.568359375,
0.48779296875,
0.58203125,
0.1234130859375,
0.58203125
],
[
0.5146484375,
0.5634765625,
0.84619140625,
0.5634765625,
0.84619140625,
0.57666015625,
0.5146484375,
0.57666015625
],
[
0.12005615234375,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.5341796875,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59326171875,
0.5341796875,
0.59326171875
],
[
0.12005615234375,
0.60107421875,
0.48779296875,
0.60107421875,
0.48779296875,
0.6142578125,
0.12005615234375,
0.6142578125
],
[
0.5146484375,
0.595703125,
0.8818359375,
0.595703125,
0.8818359375,
0.609375,
0.5146484375,
0.609375
],
[
0.12176513671875,
0.6162109375,
0.48779296875,
0.6162109375,
0.48779296875,
0.6298828125,
0.12176513671875,
0.6298828125
],
[
0.515625,
0.61083984375,
0.8818359375,
0.61083984375,
0.8818359375,
0.6259765625,
0.515625,
0.6259765625
],
[
0.12005615234375,
0.6328125,
0.489501953125,
0.6328125,
0.489501953125,
0.646484375,
0.12005615234375,
0.646484375
],
[
0.515625,
0.62744140625,
0.8818359375,
0.62744140625,
0.8818359375,
0.642578125,
0.515625,
0.642578125
],
[
0.5146484375,
0.642578125,
0.8818359375,
0.6435546875,
0.8818359375,
0.65771484375,
0.5146484375,
0.65673828125
],
[
0.12176513671875,
0.6494140625,
0.156982421875,
0.6494140625,
0.156982421875,
0.662109375,
0.12176513671875,
0.662109375
],
[
0.515625,
0.65771484375,
0.880859375,
0.65771484375,
0.880859375,
0.6728515625,
0.515625,
0.6728515625
],
[
0.1376953125,
0.6650390625,
0.489501953125,
0.6650390625,
0.489501953125,
0.6787109375,
0.1376953125,
0.6787109375
],
[
0.51318359375,
0.6748046875,
0.84033203125,
0.67578125,
0.84033203125,
0.6904296875,
0.51318359375,
0.689453125
],
[
0.118408203125,
0.681640625,
0.490234375,
0.681640625,
0.490234375,
0.6953125,
0.118408203125,
0.6953125
],
[
0.12005615234375,
0.69873046875,
0.48779296875,
0.69873046875,
0.48779296875,
0.7119140625,
0.12005615234375,
0.7119140625
],
[
0.51904296875,
0.70166015625,
0.8818359375,
0.70166015625,
0.8818359375,
0.71533203125,
0.51904296875,
0.71533203125
],
[
0.118408203125,
0.71337890625,
0.48779296875,
0.71337890625,
0.48779296875,
0.72705078125,
0.118408203125,
0.72705078125
],
[
0.515625,
0.71728515625,
0.8818359375,
0.71728515625,
0.8818359375,
0.732421875,
0.515625,
0.732421875
],
[
0.118408203125,
0.728515625,
0.48779296875,
0.728515625,
0.48779296875,
0.7421875,
0.118408203125,
0.7421875
],
[
0.515625,
0.73388671875,
0.88330078125,
0.73388671875,
0.88330078125,
0.74755859375,
0.515625,
0.74755859375
],
[
0.12005615234375,
0.7451171875,
0.48779296875,
0.7451171875,
0.48779296875,
0.76025390625,
0.12005615234375,
0.76025390625
],
[
0.5146484375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.1234130859375,
0.7626953125,
0.48779296875,
0.7626953125,
0.48779296875,
0.775390625,
0.1234130859375,
0.775390625
],
[
0.515625,
0.7666015625,
0.88330078125,
0.7666015625,
0.88330078125,
0.77978515625,
0.515625,
0.77978515625
],
[
0.12005615234375,
0.77685546875,
0.486083984375,
0.77685546875,
0.486083984375,
0.79052734375,
0.12005615234375,
0.79052734375
],
[
0.5146484375,
0.783203125,
0.55322265625,
0.783203125,
0.55322265625,
0.7958984375,
0.5146484375,
0.7958984375
],
[
0.12005615234375,
0.7919921875,
0.48779296875,
0.79345703125,
0.48779296875,
0.80810546875,
0.12005615234375,
0.80712890625
],
[
0.5341796875,
0.798828125,
0.88330078125,
0.798828125,
0.88330078125,
0.8125,
0.5341796875,
0.8125
],
[
0.118408203125,
0.80859375,
0.160400390625,
0.80859375,
0.160400390625,
0.82470703125,
0.118408203125,
0.82470703125
],
[
0.5146484375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.5146484375,
0.82763671875
],
[
0.124267578125,
0.8359375,
0.489501953125,
0.83740234375,
0.489501953125,
0.8515625,
0.124267578125,
0.85107421875
],
[
0.51708984375,
0.83056640625,
0.88330078125,
0.83056640625,
0.88330078125,
0.84423828125,
0.51708984375,
0.84423828125
],
[
0.5146484375,
0.84716796875,
0.88330078125,
0.84716796875,
0.88330078125,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.12176513671875,
0.85400390625,
0.48779296875,
0.85400390625,
0.48779296875,
0.86767578125,
0.12176513671875,
0.86767578125
],
[
0.5146484375,
0.86328125,
0.880859375,
0.86328125,
0.880859375,
0.876953125,
0.5146484375,
0.876953125
],
[
0.12005615234375,
0.869140625,
0.48779296875,
0.869140625,
0.48779296875,
0.88427734375,
0.12005615234375,
0.88427734375
],
[
0.51708984375,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.892578125,
0.51708984375,
0.892578125
],
[
0.139404296875,
0.89453125,
0.429931640625,
0.8955078125,
0.429931640625,
0.91015625,
0.139404296875,
0.9091796875
],
[
0.51708984375,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.51708984375,
0.9091796875
]
] |
[
"Tokens",
"CoNLL",
"JFLEG",
"Sentences",
"Corpus",
"System",
"M?",
"GLEU",
"R",
"NUCLE",
"57.151",
"1,162K",
"P",
"SMT Dense",
"54.68",
"25,026K",
"1,943,901",
"56.91 30.25 48.38",
"",
"CoNLL-2013 (dev)",
"1,381",
"29K",
"60.2829.40",
"49.82",
"55.25",
"+ Sparse",
"Lang-8 NAIST",
"CoNLL-2014 (test)",
"1,312",
"30K",
"+ Char. ops",
"60.2730.2150.27",
"55.79",
"JFLEGDev",
"754",
"14K",
"747",
"13K",
"Table 2: Results for SMT baseline systems on the",
"CoNLL-2014 (M) and JFLEG Test (GLEU) sets",
"JFLEGTest",
"Table 1: Statistics for training and testing data sets",
"kiewicz(2016）：49类似的系统。82vs49。49M2",
"(Koehnetal.,2007)的脚本编写的。为了处理词",
"汇表外的单词，我们使用Sennrichetal.（2016b）",
"(表 2)",
"基于短语的SMT系统不能很好地处理正交",
"的字节对编码（BPE）将令牌拆分为50k个子",
"错误（Napolesetal..2017）---如果在训练语料库",
"单词单元。BPE代码仅从Lang-8和NUCLE的",
"中未看到源词，则很可能将其复制为目标词。",
"正确句子中提取",
"子字单元可以帮助部分解决这个问题。添加基",
"3SMT systems",
"于字符级编辑计数的功能会增加两个测试集上",
"的结果",
"对于我们基于SMT的系统，我们遵循Junczys-",
"结果为55。79JFLEG测试上的GLEU已经",
"DowmuntandGrundkiewicz(2016）提出的配方，",
"比Sakaguchietal.(2017)的GLEU调整的NMT",
"并使用基于短语的SMT系统，该系统具有特",
"系统好2分，并且比ChollampattandNg（2017）",
"定于任务的特征的对数线性组合。我们使用单",
"使用其M?调整的SMT系统的最佳报告结果",
"词级Levenshtein距离和编辑操作计数作为密",
"差1分，即使此时未使用其他拼写校正。我们",
"集特征（密集），并将单词类（WC）上具有",
"一",
"在后面的部分中尝试使用专门的拼写检查方法",
"个单词左/右上下文的单词的校正模式作为稀",
"疏特征（稀疏）。我们还试验了额外的字符级",
"4.",
"",
"NMT systems",
"密集特征（Char。ops)。所有系统都使用从训练",
"我们为基于NMT的系统选择的模型架构是",
"数据的目标侧估计的5克语言模型（LM）和",
"具有双向单层编码器和解码器的引人注目的",
"OSM(Durranietal.，2011），以及在通用爬网数",
"编码器-解码器模型，两者均使用GRU作为",
"据（Bucketal.,2014）上训练的5克LM和9克",
"其RNN变体(Sennrich et al.,2017)。Sakaguchi",
"WCLM",
"etal.（2017)已经针对GEC任务测试了类似的",
"Experiment settings使用 Moses (Koehn et al.,",
"体系结构，但是我们使用了不同的超参数",
"2007）训练翻译模型，使用MGIZA++（Gaoand",
"为了提高我们的NMT模型的性能，与Xie",
"Vogel,2008)生成单词对齐模型，并且不使用重",
"etal.（2016)和Jietal.（2017）类似，我们将它",
"新排序模型。语言模型是使用KenLM(Heafield,",
"们与附加的大规模语言模型结合在一起。与以",
"2011）构建的，而单词类是使用word2vec1训",
"前使用n-gram概率LM的研究相反，我们使",
"用GRU单元构建了一个2层递归神经网络语",
"练的",
"言模型（RNNLM），并在英语通用爬行数据",
"我们分别针对M2和GLEU指标调整系统。",
"(Bucketal.,2014）上再次对其进行了训练",
"MERT（Och，2003）用于调整密集特征，Batch-",
"Mira（CherryandFoster，2012）用于调整稀疏特",
"Experimental settings我们使用Marian工具",
"征。对于M?调谐，我们遵循NUCLE上的4倍",
"包（Junczys-Dowmuntet al.，2018）在用于SMT",
"交叉验证，并采用Junczys-Dowmunt andGrund-",
"基线的相同数据上进行训练，即。e。NUCLE",
"kiewicz(2016)推荐的自适应错误率。在GLEU",
"和Lang-8。RNN隐藏状态大小设置为1024，嵌",
"上评估的模型在JFLEGDeV上使用GLEU评",
"入大小为512。源和目标词汇表以及子词单元",
"分器进行优化，我们将其添加到Moses。我们",
"相同",
"使用4次调谐运行的平均特征权重报告模型的",
"使用Adam（KingmaandBa，2014）和安装在",
"结果",
"4GBGPU内存中的小批量大小进行优化。我们",
"Results在其他条件相同的情况下，使用原",
"用比例丢失（GalandGhahramani，2016）正则化",
"模型，丢失概率为0。2在所有RNN输入和状",
"始标记化，应用子字单元以及扩展基于编辑",
"态。除此之外，我们放弃了概率为0的整个源",
"的功能会导致与Junczys-DowmuntandGrund",
"词和目标词。2和0。1分别。基于CoNLL-2013",
"'https://github.com/dav/word2vec",
"测试集上的交叉摘成本，我们以10的耐心使"
] |
{
"arxiv_id": "1804.05945",
"language": "zh",
"page_idx": 2
}
|
||
1804.05945-zh-3
|
1804.05945-zh-3.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th> </th>
<th>ZoNLL</th>
<th> </th>
<th>JFLEG</th>
</tr>
<tr>
<th>System</th>
<th>$P$</th>
<th>CoNLL $R$</th>
<th>$M^{2}$</th>
<th>JFLEG GLEU</th>
</tr>
<tr>
<th>System</th>
<th>$P$</th>
<th>POLTED $R$</th>
<th>$M^{2}$</th>
<th>L GLEU</th>
</tr>
<tr>
<td>$NMT$</td>
<td>66.61</td>
<td>17.58</td>
<td>42.76</td>
<td>50.08</td>
</tr>
<tr>
<td>NMT+ RNN- LM</td>
<td>61.05</td>
<td>26.71</td>
<td>48.56</td>
<td>56.04</td>
</tr>
<tr>
<td>NMT$\times 4$</td>
<td>71.10</td>
<td>15.42</td>
<td>41.29</td>
<td>50.30</td>
</tr>
<tr>
<td>NMT$\times 4+$RNN- LM</td>
<td>60.27</td>
<td>30.08</td>
<td>50.19</td>
<td>56.74</td>
</tr>
</tbody>
</table>
Table 3: Results for NMT systems on the CoNLL-
2014( M$^2$) and JFLEG Test (GLEU) sets.
用早期停止。每 10,000个小批量验证并保存模型。作为最终模型，我们根据 M$^{2}$或 GLEU 指标在最后十个模型检查点中选择一个在开发集上性能最佳的模型
RNNLM 的 RNN 隐藏状态和嵌入、目标词汇表和优化选项的大小与用于我们的 NMT 模型的大小相同。通过波束大小为 12 的波束搜索进行解码。我们按句子长度对每个假设的分数进行归一化
Results 单个 NMT 模型实现的性能低于SMT 基线(表 3)。然而，M$^2$得分为 42。CoNLL-2014 的 76 已经高于 41 的最佳公布结果。53 M$^2$,用于未通过其他语言模型增强的Ji et al. (2017) 的严格神经 GEC 系统
我们的 RNNLM 通过集成解码 (Sennrich et al., 2016a) 与 NMT 模型集成。与 Ji et al. (2017)类似，我们在开发集$^{2}$上使用网格搜索来选择语言模型的权重。这大大提高了召回率，从而显著提高了两个测试集上的结果(+5。$8\textbf{M}^{2}$ 和+5。96GLEU)
另一方面，由四个独立训练的模型 3 $(\mathbf{NMT???})$组成的集合以召回为代价提高了精度，甚至可能导致性能下降。将 RNNLM 添加到该集合中可以平衡这种负面影响，从而产生 50。19 M$^{2}$。这些是迄今为止在纯神经 GEC 系统的两个基准测试中报告的最高结果
Comparison to SMT systems 通过模型集成， 神经系统实现了类似于 SMT 基线的性能(图 2 )。没有 CCLM 的精简版 SMT 系统令人惊讶地在 JFLEG 上给出了比 NMT 系统更好的结果， 而 CoNLL-2014 则相反。JFLEG 性能较低的原因可能是大量的拼写错误，SMT 系统使用子字单元可以更有效地纠正这些错误
如果两个系统都通过大规模语言模型增强， 则神经系统在 JFLEG 上的性能优于 SMT 系统，并且与 CoNLL-2014 上的 SMT 系统竞争。但是，不知道如果将 NMT 模型与概率$n$ -gramLM 结合使用，结果是否会保留，就像
$^2$Used weights are 0.2 and 0.25 for M$^2$ and GLEU evalua-
tion, respectively.
$^{3}$Each model is weighted equally during decoding
$\mathrm{CoNLL}_{\mathrm{M}^2}$
$\mathrm{JFLEG}_{\mathrm{GLEU}}$
<FigureHere>
Figure 2: Contribution of a language model (LM)
for SMT and NMT GEC systems.
以前的著作(Xie et al., 2016; Ji et al., 2017) 中提
出的那样
### 5 Hybrid SMT-NMT systems
我们尝试了流水线和重新评分方法，以便将我
们最好的 SMT 和 NMTGEC 系统结合起来$^{4}$
SMT-NMT pipelines 由 SMT 系统校正的输出作为输入传递到带有或不带有 RNNLM$^{5}$的NMT 集成。在这种情况下，NMT 系统用作自动后期编辑系统。流水线通过提高召回率改善了两个测试集的结果(表 4)。由于没有 RNNLM 的 NMT 系统的性能远低于单独的 SMT 系统的性能，这意味着两种方法都会产生互补的校正
Rescoring with NMT 由一个系统从另一个系统获得的 n 个最佳列表的重新评分是GEC 中常用的技术，它允许组合多个不同的系统甚至不同的方法 (Hoang et al., 2016; Yannakoudakis et al., 2017; Chollampatt and Ng, 2017; Ji et al., 2017)。在我们的实验中，我们使用 SMT 系统生成 1000个 n-best 列表，并从每个神经元件添加单独的分数。NMT 模型和RNNLM 的分数以负对数空间中概率的形式添加。重新计分的权重是从开发集上的 BatchMira 算法 (Cherry and Foster, 2012) 的单次运行
${^4\text{The best sys}}$tem combinations are chosen again based on
the development sets, i.e. CoNLL-2013 and JFLEG Dev. We
omit these results as they are highly overestimated.
$^{5}$We did not observed any improvements if the order of
the systems is reversed.
|
[
[
0.5810546875,
0.07244873046875,
0.6533203125,
0.0748291015625,
0.6533203125,
0.08905029296875,
0.5810546875,
0.0872802734375
],
[
0.73779296875,
0.0736083984375,
0.8193359375,
0.07598876953125,
0.8193359375,
0.09027099609375,
0.73779296875,
0.0885009765625
],
[
0.4248046875,
0.07781982421875,
0.47265625,
0.07781982421875,
0.47265625,
0.08905029296875,
0.4248046875,
0.08905029296875
],
[
0.322509765625,
0.0748291015625,
0.376220703125,
0.07659912109375,
0.37451171875,
0.09027099609375,
0.32080078125,
0.0885009765625
],
[
0.38037109375,
0.09027099609375,
0.405517578125,
0.09027099609375,
0.405517578125,
0.10272216796875,
0.38037109375,
0.10272216796875
],
[
0.4248046875,
0.09027099609375,
0.47021484375,
0.09027099609375,
0.47021484375,
0.1015625,
0.4248046875,
0.1015625
],
[
0.300537109375,
0.0914306640625,
0.313232421875,
0.0914306640625,
0.313232421875,
0.1009521484375,
0.300537109375,
0.1009521484375
],
[
0.1351318359375,
0.0914306640625,
0.183837890625,
0.0914306640625,
0.183837890625,
0.10272216796875,
0.1351318359375,
0.10272216796875
],
[
0.341796875,
0.0914306640625,
0.358642578125,
0.08905029296875,
0.359375,
0.1009521484375,
0.343505859375,
0.1015625
],
[
0.1351318359375,
0.10809326171875,
0.17626953125,
0.10809326171875,
0.17626953125,
0.1217041015625,
0.1351318359375,
0.1217041015625
],
[
0.29296875,
0.1092529296875,
0.418212890625,
0.1092529296875,
0.418212890625,
0.119384765625,
0.29296875,
0.119384765625
],
[
0.418212890625,
0.1092529296875,
0.466796875,
0.1092529296875,
0.466796875,
0.12054443359375,
0.418212890625,
0.12054443359375
],
[
0.291259765625,
0.119384765625,
0.418212890625,
0.119384765625,
0.418212890625,
0.133056640625,
0.291259765625,
0.133056640625
],
[
0.62109375,
0.119384765625,
0.6884765625,
0.119384765625,
0.6884765625,
0.1383056640625,
0.62109375,
0.1383056640625
],
[
0.85791015625,
0.11639404296875,
0.87890625,
0.11639404296875,
0.87890625,
0.1300048828125,
0.85791015625,
0.1300048828125
],
[
0.42822265625,
0.12054443359375,
0.465087890625,
0.12054443359375,
0.465087890625,
0.1318359375,
0.42822265625,
0.1318359375
],
[
0.1368408203125,
0.1217041015625,
0.2476806640625,
0.1217041015625,
0.2476806640625,
0.133056640625,
0.1368408203125,
0.133056640625
],
[
0.291259765625,
0.1312255859375,
0.4189453125,
0.1312255859375,
0.4189453125,
0.144287109375,
0.291259765625,
0.144287109375
],
[
0.51904296875,
0.1300048828125,
0.54150390625,
0.1300048828125,
0.54150390625,
0.142578125,
0.51904296875,
0.142578125
],
[
0.429931640625,
0.1318359375,
0.466796875,
0.1318359375,
0.466796875,
0.1436767578125,
0.429931640625,
0.1436767578125
],
[
0.1368408203125,
0.1318359375,
0.20068359375,
0.1318359375,
0.20068359375,
0.1436767578125,
0.1368408203125,
0.1436767578125
],
[
0.420654296875,
0.1353759765625,
0.42822265625,
0.1353759765625,
0.42822265625,
0.1395263671875,
0.420654296875,
0.1395263671875
],
[
0.81689453125,
0.1318359375,
0.8486328125,
0.1318359375,
0.8486328125,
0.149658203125,
0.81689453125,
0.149658203125
],
[
0.1351318359375,
0.1436767578125,
0.279541015625,
0.1436767578125,
0.279541015625,
0.15673828125,
0.1351318359375,
0.15673828125
],
[
0.28955078125,
0.142578125,
0.4189453125,
0.142578125,
0.4189453125,
0.15673828125,
0.28955078125,
0.15673828125
],
[
0.42822265625,
0.1436767578125,
0.466796875,
0.1436767578125,
0.466796875,
0.1561279296875,
0.42822265625,
0.1561279296875
],
[
0.7841796875,
0.146728515625,
0.84814453125,
0.15087890625,
0.8447265625,
0.169189453125,
0.78076171875,
0.1650390625
],
[
0.6572265625,
0.149658203125,
0.68603515625,
0.149658203125,
0.68603515625,
0.174560546875,
0.6572265625,
0.174560546875
],
[
0.420654296875,
0.1478271484375,
0.42822265625,
0.1478271484375,
0.42822265625,
0.1519775390625,
0.420654296875,
0.1519775390625
],
[
0.8486328125,
0.15673828125,
0.875,
0.1484375,
0.88330078125,
0.162109375,
0.85791015625,
0.17041015625
],
[
0.118408203125,
0.1728515625,
0.48779296875,
0.1728515625,
0.48779296875,
0.18701171875,
0.118408203125,
0.18701171875
],
[
0.118408203125,
0.18701171875,
0.42822265625,
0.188232421875,
0.42822265625,
0.203125,
0.118408203125,
0.201904296875
],
[
0.51904296875,
0.18408203125,
0.54150390625,
0.18408203125,
0.54150390625,
0.196533203125,
0.51904296875,
0.196533203125
],
[
0.85791015625,
0.188232421875,
0.880859375,
0.188232421875,
0.880859375,
0.203125,
0.85791015625,
0.203125
],
[
0.7119140625,
0.208984375,
0.748046875,
0.208984375,
0.748046875,
0.23095703125,
0.7119140625,
0.23095703125
],
[
0.587890625,
0.2197265625,
0.61865234375,
0.21728515625,
0.62060546875,
0.22802734375,
0.59033203125,
0.23095703125
],
[
0.12005615234375,
0.23095703125,
0.484375,
0.23095703125,
0.484375,
0.24462890625,
0.12005615234375,
0.24462890625
],
[
0.85791015625,
0.2257080078125,
0.880859375,
0.2257080078125,
0.880859375,
0.2392578125,
0.85791015625,
0.2392578125
],
[
0.5205078125,
0.240478515625,
0.53662109375,
0.240478515625,
0.53662109375,
0.2493896484375,
0.5205078125,
0.2493896484375
],
[
0.12176513671875,
0.2464599609375,
0.48779296875,
0.2464599609375,
0.48779296875,
0.26123046875,
0.12176513671875,
0.26123046875
],
[
0.5556640625,
0.248779296875,
0.5859375,
0.251708984375,
0.58251953125,
0.26318359375,
0.55517578125,
0.260009765625
],
[
0.118408203125,
0.262451171875,
0.486083984375,
0.262451171875,
0.486083984375,
0.275634765625,
0.118408203125,
0.275634765625
],
[
0.80419921875,
0.262451171875,
0.84619140625,
0.262451171875,
0.84619140625,
0.2724609375,
0.80419921875,
0.2724609375
],
[
0.84033203125,
0.262451171875,
0.87890625,
0.262451171875,
0.87890625,
0.27490234375,
0.84033203125,
0.27490234375
],
[
0.8037109375,
0.273681640625,
0.8447265625,
0.273681640625,
0.8447265625,
0.28515625,
0.8037109375,
0.28515625
],
[
0.75146484375,
0.26318359375,
0.78515625,
0.26318359375,
0.78515625,
0.279052734375,
0.75146484375,
0.279052734375
],
[
0.1234130859375,
0.28076171875,
0.26611328125,
0.28076171875,
0.26611328125,
0.29150390625,
0.1234130859375,
0.29150390625
],
[
0.1351318359375,
0.29345703125,
0.48779296875,
0.29443359375,
0.48779296875,
0.308837890625,
0.1351318359375,
0.30810546875
],
[
0.55322265625,
0.298583984375,
0.67919921875,
0.297607421875,
0.67919921875,
0.31640625,
0.55322265625,
0.317626953125
],
[
0.7109375,
0.300537109375,
0.771484375,
0.298583984375,
0.7734375,
0.3134765625,
0.7119140625,
0.3154296875
],
[
0.7841796875,
0.3046875,
0.80419921875,
0.3017578125,
0.8076171875,
0.308837890625,
0.78515625,
0.311279296875
],
[
0.12005615234375,
0.311279296875,
0.48779296875,
0.311279296875,
0.48779296875,
0.325927734375,
0.12005615234375,
0.325927734375
],
[
0.6533203125,
0.311279296875,
0.66650390625,
0.311279296875,
0.66650390625,
0.314208984375,
0.6533203125,
0.314208984375
],
[
0.8193359375,
0.30712890625,
0.82861328125,
0.300537109375,
0.83642578125,
0.305908203125,
0.8271484375,
0.312255859375
],
[
0.603515625,
0.322509765625,
0.64208984375,
0.32666015625,
0.6396484375,
0.33740234375,
0.60107421875,
0.333251953125
],
[
0.634765625,
0.32177734375,
0.6455078125,
0.32177734375,
0.6455078125,
0.32666015625,
0.634765625,
0.32666015625
],
[
0.75390625,
0.325927734375,
0.76904296875,
0.325927734375,
0.76904296875,
0.3349609375,
0.75390625,
0.3349609375
],
[
0.78857421875,
0.32666015625,
0.8017578125,
0.32666015625,
0.8017578125,
0.3349609375,
0.78857421875,
0.3349609375
],
[
0.59033203125,
0.327880859375,
0.615234375,
0.319580078125,
0.62109375,
0.328857421875,
0.59765625,
0.33740234375
],
[
0.12005615234375,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.12005615234375,
0.341552734375
],
[
0.118408203125,
0.342529296875,
0.486083984375,
0.34326171875,
0.486083984375,
0.358154296875,
0.118408203125,
0.35693359375
],
[
0.51318359375,
0.355224609375,
0.8818359375,
0.355712890625,
0.8818359375,
0.37060546875,
0.51318359375,
0.369384765625
],
[
0.12176513671875,
0.361083984375,
0.2301025390625,
0.361083984375,
0.2301025390625,
0.372314453125,
0.12176513671875,
0.372314453125
],
[
0.5146484375,
0.372314453125,
0.76318359375,
0.372314453125,
0.76318359375,
0.385986328125,
0.5146484375,
0.385986328125
],
[
0.1275634765625,
0.385498046875,
0.484375,
0.385498046875,
0.484375,
0.3984375,
0.1275634765625,
0.3984375
],
[
0.12005615234375,
0.40087890625,
0.48779296875,
0.40087890625,
0.48779296875,
0.4150390625,
0.12005615234375,
0.4150390625
],
[
0.12005615234375,
0.41748046875,
0.48779296875,
0.41748046875,
0.48779296875,
0.431640625,
0.12005615234375,
0.431640625
],
[
0.515625,
0.41455078125,
0.8818359375,
0.41455078125,
0.8818359375,
0.427490234375,
0.515625,
0.427490234375
],
[
0.12176513671875,
0.43408203125,
0.48779296875,
0.43408203125,
0.48779296875,
0.44775390625,
0.12176513671875,
0.44775390625
],
[
0.5146484375,
0.429931640625,
0.591796875,
0.429931640625,
0.591796875,
0.44482421875,
0.5146484375,
0.44482421875
],
[
0.118408203125,
0.449462890625,
0.405517578125,
0.449462890625,
0.405517578125,
0.463134765625,
0.118408203125,
0.463134765625
],
[
0.51123046875,
0.4619140625,
0.5263671875,
0.457763671875,
0.53076171875,
0.466064453125,
0.515625,
0.47021484375
],
[
0.54150390625,
0.4619140625,
0.54736328125,
0.4619140625,
0.54736328125,
0.46728515625,
0.54150390625,
0.46728515625
],
[
0.54736328125,
0.457275390625,
0.78662109375,
0.457763671875,
0.78662109375,
0.47265625,
0.54736328125,
0.471435546875
],
[
0.1376953125,
0.466064453125,
0.48779296875,
0.466064453125,
0.48779296875,
0.479736328125,
0.1376953125,
0.479736328125
],
[
0.118408203125,
0.482177734375,
0.486083984375,
0.482177734375,
0.486083984375,
0.495361328125,
0.118408203125,
0.495361328125
],
[
0.515625,
0.48388671875,
0.88330078125,
0.48388671875,
0.88330078125,
0.49755859375,
0.515625,
0.49755859375
],
[
0.118408203125,
0.49755859375,
0.48779296875,
0.49755859375,
0.48779296875,
0.5126953125,
0.118408203125,
0.5126953125
],
[
0.515625,
0.498779296875,
0.861328125,
0.498779296875,
0.861328125,
0.51318359375,
0.515625,
0.51318359375
],
[
0.12005615234375,
0.51416015625,
0.48779296875,
0.51416015625,
0.48779296875,
0.52783203125,
0.12005615234375,
0.52783203125
],
[
0.5205078125,
0.52490234375,
0.880859375,
0.52490234375,
0.880859375,
0.5380859375,
0.5205078125,
0.5380859375
],
[
0.12005615234375,
0.53076171875,
0.489501953125,
0.53076171875,
0.489501953125,
0.54443359375,
0.12005615234375,
0.54443359375
],
[
0.118408203125,
0.54638671875,
0.298828125,
0.54638671875,
0.298828125,
0.56005859375,
0.118408203125,
0.56005859375
],
[
0.51708984375,
0.54150390625,
0.8818359375,
0.54150390625,
0.8818359375,
0.5546875,
0.51708984375,
0.5546875
],
[
0.1368408203125,
0.56103515625,
0.491943359375,
0.55859375,
0.491943359375,
0.57666015625,
0.1368408203125,
0.5791015625
],
[
0.5146484375,
0.55712890625,
0.8818359375,
0.55712890625,
0.8818359375,
0.57080078125,
0.5146484375,
0.57080078125
],
[
0.118408203125,
0.5791015625,
0.486083984375,
0.5791015625,
0.486083984375,
0.5927734375,
0.118408203125,
0.5927734375
],
[
0.515625,
0.57373046875,
0.880859375,
0.57373046875,
0.880859375,
0.58740234375,
0.515625,
0.58740234375
],
[
0.12005615234375,
0.59326171875,
0.486083984375,
0.59326171875,
0.486083984375,
0.60693359375,
0.12005615234375,
0.60693359375
],
[
0.515625,
0.5888671875,
0.88330078125,
0.5888671875,
0.88330078125,
0.6025390625,
0.515625,
0.6025390625
],
[
0.5146484375,
0.60400390625,
0.88330078125,
0.60498046875,
0.88330078125,
0.619140625,
0.5146484375,
0.6181640625
],
[
0.12176513671875,
0.61083984375,
0.48779296875,
0.61083984375,
0.48779296875,
0.62451171875,
0.12176513671875,
0.62451171875
],
[
0.5146484375,
0.62060546875,
0.8818359375,
0.62060546875,
0.8818359375,
0.63427734375,
0.5146484375,
0.63427734375
],
[
0.12005615234375,
0.626953125,
0.48779296875,
0.626953125,
0.48779296875,
0.64013671875,
0.12005615234375,
0.64013671875
],
[
0.12005615234375,
0.6435546875,
0.432373046875,
0.6435546875,
0.432373046875,
0.65673828125,
0.12005615234375,
0.65673828125
],
[
0.515625,
0.64013671875,
0.5341796875,
0.64013671875,
0.5341796875,
0.6494140625,
0.515625,
0.6494140625
],
[
0.5224609375,
0.6611328125,
0.86376953125,
0.6611328125,
0.86376953125,
0.67578125,
0.5224609375,
0.67578125
],
[
0.124267578125,
0.66943359375,
0.484375,
0.66943359375,
0.484375,
0.68310546875,
0.124267578125,
0.68310546875
],
[
0.515625,
0.6787109375,
0.8818359375,
0.6787109375,
0.8818359375,
0.6923828125,
0.515625,
0.6923828125
],
[
0.12176513671875,
0.68505859375,
0.490234375,
0.68505859375,
0.490234375,
0.69873046875,
0.12176513671875,
0.69873046875
],
[
0.5146484375,
0.69482421875,
0.8818359375,
0.69482421875,
0.8818359375,
0.7080078125,
0.5146484375,
0.7080078125
],
[
0.12005615234375,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.12005615234375,
0.71533203125
],
[
0.515625,
0.71142578125,
0.88330078125,
0.71142578125,
0.88330078125,
0.724609375,
0.515625,
0.724609375
],
[
0.12005615234375,
0.71728515625,
0.484375,
0.71728515625,
0.484375,
0.73095703125,
0.12005615234375,
0.73095703125
],
[
0.515625,
0.72705078125,
0.880859375,
0.72705078125,
0.880859375,
0.74072265625,
0.515625,
0.74072265625
],
[
0.12005615234375,
0.73388671875,
0.48779296875,
0.73388671875,
0.48779296875,
0.74755859375,
0.12005615234375,
0.74755859375
],
[
0.5146484375,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.5146484375,
0.75732421875
],
[
0.12176513671875,
0.75,
0.48779296875,
0.75,
0.48779296875,
0.76318359375,
0.12176513671875,
0.76318359375
],
[
0.515625,
0.7587890625,
0.8818359375,
0.7587890625,
0.8818359375,
0.7724609375,
0.515625,
0.7724609375
],
[
0.12176513671875,
0.7666015625,
0.3955078125,
0.7666015625,
0.3955078125,
0.77978515625,
0.12176513671875,
0.77978515625
],
[
0.1368408203125,
0.7822265625,
0.484375,
0.78076171875,
0.484375,
0.7958984375,
0.1368408203125,
0.79638671875
],
[
0.51708984375,
0.77685546875,
0.880859375,
0.77685546875,
0.880859375,
0.78759765625,
0.51708984375,
0.78759765625
],
[
0.5146484375,
0.79150390625,
0.88330078125,
0.79150390625,
0.88330078125,
0.8046875,
0.5146484375,
0.8046875
],
[
0.12005615234375,
0.79736328125,
0.48779296875,
0.79736328125,
0.48779296875,
0.8125,
0.12005615234375,
0.8125
],
[
0.51318359375,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.51318359375,
0.82080078125
],
[
0.12005615234375,
0.81396484375,
0.489501953125,
0.81396484375,
0.489501953125,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.51123046875,
0.82373046875,
0.8818359375,
0.82373046875,
0.8818359375,
0.83740234375,
0.51123046875,
0.83740234375
],
[
0.12005615234375,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.1175537109375,
0.8466796875,
0.48779296875,
0.845703125,
0.48779296875,
0.85986328125,
0.1175537109375,
0.86083984375
],
[
0.5341796875,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.85986328125,
0.5341796875,
0.85986328125
],
[
0.5146484375,
0.859375,
0.884765625,
0.859375,
0.884765625,
0.87255859375,
0.5146484375,
0.87255859375
],
[
0.1376953125,
0.8681640625,
0.490234375,
0.87060546875,
0.490234375,
0.88525390625,
0.1376953125,
0.8828125
],
[
0.515625,
0.87255859375,
0.82275390625,
0.87255859375,
0.82275390625,
0.8828125,
0.515625,
0.8828125
],
[
0.12005615234375,
0.88427734375,
0.225830078125,
0.88427734375,
0.225830078125,
0.8955078125,
0.12005615234375,
0.8955078125
],
[
0.5341796875,
0.8818359375,
0.884765625,
0.88427734375,
0.884765625,
0.8984375,
0.5341796875,
0.896484375
],
[
0.14111328125,
0.89453125,
0.44580078125,
0.896484375,
0.44580078125,
0.91015625,
0.14111328125,
0.90771484375
],
[
0.5146484375,
0.89794921875,
0.6572265625,
0.89794921875,
0.6572265625,
0.90771484375,
0.5146484375,
0.90771484375
]
] |
[
"CoNLL M?",
"JFLEG GLEU",
"JFLEG",
"CoNLL",
"M?",
"GLEU",
"P",
"System",
"R",
"NMT",
"66.61 17.58 42.76",
"50.08",
"61.05 26.71 48.56",
"59.3 50.2",
"58",
"56.04",
"NMT + RNN-LM",
"71.10 15.42 41.29",
"50",
"50.30",
"NMT × 4",
"",
"56.7",
"NMT × 4 + RNN-LM",
"60.27 30.08 50.19",
"56.74",
"55.8 ±",
"王",
"",
"156",
"Table 3: Results for NMT systems on the CoNLL",
"2014 (M2) and JFLEG Test (GLEU) sets.",
"45",
"54",
"525",
"41.3",
"用早期停止。每10,000个小批量验证并保存模",
"52",
"40",
"型。作为最终模型，我们根据M?或GLEU指",
"38.3",
"标在最后十个模型检查点中选择一个在开发集",
"-Avg",
"50",
"×4",
"A5",
"上性能最佳的模型",
"RNNLM的RNN隐藏状态和嵌人、目标词",
"MM",
"MM",
"N",
"汇表和优化选项的大小与用于我们的NMT模",
"",
"N",
"",
"A",
"N",
"",
"SMI",
"型的大小相同。通过波束大小为12的波束搜",
"索进行解码。我们按句子长度对每个假设的分",
"Figure 2: Contribution of a language model (LM)",
"数进行归一化",
"for SMT and NMT GEC systems.",
"Results单个NMT模型实现的性能低于",
"SMT基线（表3）。然而，M²得分为42。",
"CoNLL-2014的76已经高于41的最佳公布结",
"以前的著作（Xieetal.2016;Jietal.,2017）中提",
"果。53M²，用于未通过其他语言模型增强的",
"出的那样",
"Jietal.(2017)的严格神经GEC系统",
"5",
"",
"Hybrid SMT-NMT systems",
"我们的RNNLM通过集成解码（Sennrich",
"etal.,2016a）与NMT模型集成。与Jietal",
"我们尝试了流水线和重新评分方法，以便将我",
"(2017）类似，我们在开发集2上使用网格搜",
"们最好的SMT和NMTGEC系统结合起来4",
"索来选择语言模型的权重。这大大提高了召回",
"SMT-NMTpipelines由SMT系统校正的输",
"率，从而显著提高了两个测试集上的结果（+5。",
"8 M2 和 +5。96GLEU)",
"出作为输人传递到带有或不带有RNNLM5的",
"另一方面，由四个独立训练的模型3",
"NMT集成。在这种情况下，NMT系统用作自动",
"（NMT???）组成的集合以召回为代价提高了",
"后期编辑系统。流水线通过提高召回率改善了",
"精度，甚至可能导致性能下降。将RNNLM添",
"两个测试集的结果（表4）。由于没有RNNLM",
"的NMT系统的性能远低于单独的SMT系统",
"加到该集合中可以平衡这种负面影响，从而产",
"的性能，这意味着两种方法都会产生互补的校",
"生50。19M?。这些是迄今为止在纯神经GEC",
"系统的两个基准测试中报告的最高结果",
"正",
"RescoringwithNMT由一个系统从另",
"ComparisontoSMTsystems通过模型集成",
"个系统获得的n个最佳列表的重新评分是",
"神经系统实现了类似于SMT基线的性能（图2",
"GEC中常用的技术，它允许组合多个不同",
")。没有CCLM的精简版SMT系统令人惊地",
"的系统甚至不同的方法（Hoang etal,2016;",
"在JFLEG上给出了比NMT系统更好的结果",
"Yannakoudakis et al., 2017; Chollampatt and Ng",
"而CoNLL-2014则相反。JFLEG性能较低的原",
"2017；Jietal.，2017）。在我们的实验中，我们",
"因可能是大量的拼写错误，SMT系统使用子",
"使用SMT系统生成1000个n-best列表，并从",
"字单元可以更有效地纠正这些错误",
"如果两个系统都通过大规模语言模型增强",
"每个神经元件添加单独的分数。NMT模型和",
"RNNLM的分数以负对数空间中概率的形式",
"则神经系统在JFLEG上的性能优于SMT系",
"添加。重新计分的权重是从开发集上的Batch",
"统，并且与CoNLL-2014上的SMT系统竞",
"Mira算法（CherryandFoster，2012）的单次运行",
"争。但是，不知道如果将NMT模型与概率",
"n-gramLM结合使用，结果是否会保留，就像",
"+The best system combinations are chosen again based on",
"the development sets, i.e. CoNLL-2013 and JFLEG Dev. We",
"2Used weights are 0.2 and 0.25 for M? and GLEU evalua-",
"omit these results as they are highly overestimated.",
"tion, respectively.",
"5we did not observed any improvements if the order of",
"3Each model is weighted equally during decoding.",
"the systems is reversed."
] |
{
"arxiv_id": "1804.05945",
"language": "zh",
"page_idx": 3
}
|
||
1804.05945-zh-4
|
1804.05945-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>System</th>
<th>$P$</th>
<th>CoNLL $R$</th>
<th>$M^{2}$</th>
<th>JFLEG GLEU</th>
</tr>
<tr>
<td>Best SMT</td>
<td>60.27</td>
<td>30.21</td>
<td>50.27</td>
<td>55.79</td>
</tr>
<tr>
<td>$\rightarrow$Pip. NMT</td>
<td>60.25</td>
<td>34.80</td>
<td>52.56</td>
<td>57.21</td>
</tr>
<tr>
<td>Pip. NMT+ LM $\rightarrow$</td>
<td>58.87</td>
<td>39.23</td>
<td>53.51</td>
<td>58.83</td>
</tr>
<tr>
<td>+Res. RNN-LM</td>
<td>70.97</td>
<td>24.86</td>
<td>51.77</td>
<td>56.97</td>
</tr>
<tr>
<td>+Res. NMT</td>
<td>70.40</td>
<td>26.69</td>
<td>53.03</td>
<td>57.21</td>
</tr>
<tr>
<td>+Res. NMT+LM</td>
<td>71.40</td>
<td>28.60</td>
<td>54.95</td>
<td>57.53</td>
</tr>
<tr>
<td>$\to$Pip. NMT+LM</td>
<td>65.73</td>
<td>33.36</td>
<td>55.05</td>
<td>58.83</td>
</tr>
<tr>
<td>+Spell SMT</td>
<td>70.80</td>
<td>30.57</td>
<td>56.05</td>
<td>60.09</td>
</tr>
<tr>
<td>$\rightarrow$Pip. NMT+LM</td>
<td>66.77</td>
<td>34.49</td>
<td>56.25</td>
<td>61.50</td>
</tr>
</tbody>
</table>
Table 4: Results for hybrid SMT-NMT systems on the CoNLL-2014( M$^{2}$)and JFLEG Test(GLEU) sets.
# 中获得的
与流水线相反，重新计分以召回为代价提高了精度，并且对于CoNLL 数据更有效，导致多达54。95 M$^{2}$。在 JFLEG 上，仅使用 RNNLM 进行重新评分会产生与使用 NMT 集成进行重新评分相似的结果。但是，在该测试集上，重新计分的最佳结果低于流水线的最佳结果。似乎SMT 系统无法在 n 个最佳列表中产生与 NMT 集合所产生的一样多的校正
Spelling correction and final results 对 NMTrescoredSMT 系统和 NMT 系统进行流水线操作可进一步改进。我们认为，这可以通过两种方法对精度和召回权衡的不同贡献来解释，类似于 NMT 集成和我们的 RNNLM 组合所观察到的效果
在我们最终的混合系统之上，我们添加了一个拼写检查组件，该组件在流水线之前运行。我们使用 Chollampatt and Ng (2017)之后的字符级 SMT 系统，他们在基于单词的 SMT 系统中部署未知单词。由于我们基于 BPE 的SMT 并没有真正遭受未知单词的困扰，因此我们对将由 BPE 算法分段的单词运行拼写检查组件。最后一个系统实现了本文报告的最佳结果：56。25 M$^{2}$ onCoNLL-2014and61。50GLEUonJFLEGTest
## 6 Analysis and future work
对于这两个基准，我们的系统都接近自动评估结果，这些结果已被声称与 CoNLL-2014 测试集和 JFLEG 测试中的人类水平表现相对应
Example outputs 表 5 显示了来自 JFLEG 测试语料库的示例源句子的系统输出，说明了统计和神经方法的互补性。SMT 和 NMT 系统产生不同的校正。重新评分能够生成唯一的校正$( \textit{ is change}\to has\textit{changed }) $,但它无法从神经系
统生成一些校正，例如。g。拼写错误( becom 和dificullty)。另一方面，流水线可能无法改善SMT 系统 ( is changed ) 进行的局部校正。这两种方法的组合产生的输出，这是最相似的引用
Comparison with human annotations Bryant and Ng (2015) 创建了 CoNLL-2014 测试集的扩展，总共有 10 个注释器，JFLEG 已经合并了来自 4 个注释器的更正。M$^2$和 GLEU 的人类水平结果是通过平均每个注释者分别相对于其余9个(CoNLL)或3个(JFLEG)注释者的得分来计算的
图$\overline{\mathrm{E}}$3 包含人类水平得分，我们的结果以及Chollampatt and Ng (2017) 以前报告的最佳结果。根据 M$^2$,我们最好的系统达到了人类平均得分的近 100 个%,而 GLEU 的近 99 个% 比以前的作品$^{6}$更接近该界限
但是，进一步检查发现，与人为校正相比， 自动系统的精度/召回权衡表明覆盖率较低-较低的召回率由高精度$^{7}$补偿。例如，自动系统可能会错过一些明显的错误纠正，因此很容易与人类参考区分开来。未来的工作将需要人类的评估工作来得出更多的结论
7
Acknowledgments
这项工作部分由 Facebook 资助。本文所包含的观点和结论是作者的观点和结论，不应被解释为必然代表 Facebook 的官方政策或认可， 无论是明示的还是暗示的。
# References
Steven Bird and Edward Loper. 2004. NLTK: the nat-
ural language toolkit. In Proceedings of the ACL 2004 on Interactive poster and demonstration sessions. Association for Computational Linguistics, page 31.
Christopher Bryant and Hwee Tou Ng.2015. How
far are we from fully automatic high quality grammatical error correction? In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Vol- $ume~l:~Long~Papers).~Association~for~Computa-$ $^{6}$During the camera-ready preparation, Chollampatt and
Ng (2018) have published a GEC system based on a multilayer convolutional encoder-decoder neural network with a character-based spell-checking module improving the previous best result to $\dot{5} 4. 79$ $M^{2}$ on $\dot{\text{CoNLL- 2014 and 57. 47 GLEU}}$ on JFLEG Test.
A similar imbalance between precision and recall is visi-
ble on JFLEG when the M$^{2}$ metric is used.
|
[
[
0.328369140625,
0.07781982421875,
0.3779296875,
0.07781982421875,
0.3779296875,
0.08905029296875,
0.328369140625,
0.08905029296875
],
[
0.42822265625,
0.07781982421875,
0.474365234375,
0.07781982421875,
0.474365234375,
0.08905029296875,
0.42822265625,
0.08905029296875
],
[
0.5146484375,
0.07659912109375,
0.88330078125,
0.07659912109375,
0.88330078125,
0.0914306640625,
0.5146484375,
0.0914306640625
],
[
0.130126953125,
0.09027099609375,
0.1822509765625,
0.09027099609375,
0.1822509765625,
0.1051025390625,
0.130126953125,
0.1051025390625
],
[
0.38037109375,
0.09027099609375,
0.41064453125,
0.09027099609375,
0.41064453125,
0.10272216796875,
0.38037109375,
0.10272216796875
],
[
0.42822265625,
0.08905029296875,
0.474365234375,
0.08905029296875,
0.474365234375,
0.10394287109375,
0.42822265625,
0.10394287109375
],
[
0.3359375,
0.09320068359375,
0.349365234375,
0.0885009765625,
0.358642578125,
0.09979248046875,
0.344970703125,
0.1051025390625
],
[
0.2939453125,
0.0914306640625,
0.30908203125,
0.0914306640625,
0.30908203125,
0.1015625,
0.2939453125,
0.1015625
],
[
0.515625,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.515625,
0.10687255859375
],
[
0.1318359375,
0.1092529296875,
0.197265625,
0.1092529296875,
0.197265625,
0.12054443359375,
0.1318359375,
0.12054443359375
],
[
0.2939453125,
0.10809326171875,
0.424072265625,
0.10809326171875,
0.424072265625,
0.1217041015625,
0.2939453125,
0.1217041015625
],
[
0.4306640625,
0.1092529296875,
0.47021484375,
0.1092529296875,
0.47021484375,
0.12054443359375,
0.4306640625,
0.12054443359375
],
[
0.5146484375,
0.1092529296875,
0.8818359375,
0.1092529296875,
0.8818359375,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.133544921875,
0.1217041015625,
0.219970703125,
0.119384765625,
0.219970703125,
0.1341552734375,
0.133544921875,
0.135986328125
],
[
0.295654296875,
0.1217041015625,
0.337646484375,
0.1217041015625,
0.337646484375,
0.133056640625,
0.295654296875,
0.133056640625
],
[
0.333251953125,
0.1217041015625,
0.378662109375,
0.1217041015625,
0.378662109375,
0.133056640625,
0.333251953125,
0.133056640625
],
[
0.376220703125,
0.1217041015625,
0.42236328125,
0.1217041015625,
0.42236328125,
0.133056640625,
0.376220703125,
0.133056640625
],
[
0.4306640625,
0.1217041015625,
0.468505859375,
0.1217041015625,
0.468505859375,
0.133056640625,
0.4306640625,
0.133056640625
],
[
0.2939453125,
0.1312255859375,
0.42236328125,
0.1312255859375,
0.42236328125,
0.1455078125,
0.2939453125,
0.1455078125
],
[
0.515625,
0.1258544921875,
0.8818359375,
0.1258544921875,
0.8818359375,
0.1395263671875,
0.515625,
0.1395263671875
],
[
0.1368408203125,
0.133056640625,
0.246826171875,
0.133056640625,
0.246826171875,
0.144287109375,
0.1368408203125,
0.144287109375
],
[
0.4306640625,
0.133056640625,
0.468505859375,
0.133056640625,
0.468505859375,
0.144287109375,
0.4306640625,
0.144287109375
],
[
0.515625,
0.141357421875,
0.53662109375,
0.141357421875,
0.53662109375,
0.155029296875,
0.515625,
0.155029296875
],
[
0.133544921875,
0.15087890625,
0.2392578125,
0.15087890625,
0.2392578125,
0.162109375,
0.133544921875,
0.162109375
],
[
0.2939453125,
0.149658203125,
0.4248046875,
0.149658203125,
0.4248046875,
0.163330078125,
0.2939453125,
0.163330078125
],
[
0.4306640625,
0.15087890625,
0.47021484375,
0.15087890625,
0.47021484375,
0.162109375,
0.4306640625,
0.162109375
],
[
0.133544921875,
0.163330078125,
0.210693359375,
0.163330078125,
0.210693359375,
0.174560546875,
0.133544921875,
0.174560546875
],
[
0.295654296875,
0.162109375,
0.424072265625,
0.162109375,
0.424072265625,
0.17578125,
0.295654296875,
0.17578125
],
[
0.4306640625,
0.163330078125,
0.468505859375,
0.163330078125,
0.468505859375,
0.174560546875,
0.4306640625,
0.174560546875
],
[
0.51904296875,
0.16748046875,
0.8818359375,
0.16748046875,
0.8818359375,
0.18115234375,
0.51904296875,
0.18115234375
],
[
0.1318359375,
0.17578125,
0.2435302734375,
0.17578125,
0.2435302734375,
0.18701171875,
0.1318359375,
0.18701171875
],
[
0.295654296875,
0.17333984375,
0.42236328125,
0.17333984375,
0.42236328125,
0.188232421875,
0.295654296875,
0.188232421875
],
[
0.4306640625,
0.17578125,
0.468505859375,
0.17578125,
0.468505859375,
0.18701171875,
0.4306640625,
0.18701171875
],
[
0.5146484375,
0.18408203125,
0.880859375,
0.18408203125,
0.880859375,
0.19775390625,
0.5146484375,
0.19775390625
],
[
0.151123046875,
0.193603515625,
0.2626953125,
0.193603515625,
0.2626953125,
0.204833984375,
0.151123046875,
0.204833984375
],
[
0.2939453125,
0.191162109375,
0.424072265625,
0.191162109375,
0.424072265625,
0.204833984375,
0.2939453125,
0.204833984375
],
[
0.4306640625,
0.193603515625,
0.47021484375,
0.193603515625,
0.47021484375,
0.204833984375,
0.4306640625,
0.204833984375
],
[
0.4306640625,
0.203125,
0.47021484375,
0.204833984375,
0.468505859375,
0.21728515625,
0.429931640625,
0.215576171875
],
[
0.515625,
0.1995849609375,
0.880859375,
0.1995849609375,
0.880859375,
0.213134765625,
0.515625,
0.213134765625
],
[
0.1502685546875,
0.2060546875,
0.2283935546875,
0.2060546875,
0.2283935546875,
0.21728515625,
0.1502685546875,
0.21728515625
],
[
0.295654296875,
0.2037353515625,
0.424072265625,
0.2037353515625,
0.424072265625,
0.218505859375,
0.295654296875,
0.218505859375
],
[
0.162841796875,
0.2161865234375,
0.280517578125,
0.2161865234375,
0.280517578125,
0.2298583984375,
0.162841796875,
0.2298583984375
],
[
0.29296875,
0.215576171875,
0.4248046875,
0.21435546875,
0.4248046875,
0.2286376953125,
0.29296875,
0.2298583984375
],
[
0.4306640625,
0.21728515625,
0.47021484375,
0.21728515625,
0.47021484375,
0.2286376953125,
0.4306640625,
0.2286376953125
],
[
0.515625,
0.2161865234375,
0.88330078125,
0.2161865234375,
0.88330078125,
0.2298583984375,
0.515625,
0.2298583984375
],
[
0.515625,
0.232177734375,
0.88330078125,
0.232177734375,
0.88330078125,
0.2452392578125,
0.515625,
0.2452392578125
],
[
0.1175537109375,
0.243408203125,
0.489501953125,
0.2452392578125,
0.489501953125,
0.2587890625,
0.1175537109375,
0.257080078125
],
[
0.515625,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.515625,
0.26123046875
],
[
0.118408203125,
0.26123046875,
0.489501953125,
0.26123046875,
0.489501953125,
0.27490234375,
0.118408203125,
0.27490234375
],
[
0.515625,
0.265380859375,
0.62451171875,
0.265380859375,
0.62451171875,
0.276611328125,
0.515625,
0.276611328125
],
[
0.118408203125,
0.279052734375,
0.15283203125,
0.279052734375,
0.15283203125,
0.29150390625,
0.118408203125,
0.29150390625
],
[
0.5341796875,
0.28076171875,
0.8818359375,
0.28076171875,
0.8818359375,
0.29443359375,
0.5341796875,
0.29443359375
],
[
0.51318359375,
0.295654296875,
0.88330078125,
0.29638671875,
0.88330078125,
0.311279296875,
0.51318359375,
0.31005859375
],
[
0.515625,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.32666015625,
0.515625,
0.32666015625
],
[
0.12005615234375,
0.320556640625,
0.193115234375,
0.320556640625,
0.193115234375,
0.33203125,
0.12005615234375,
0.33203125
],
[
0.5146484375,
0.327880859375,
0.884765625,
0.328857421875,
0.884765625,
0.34326171875,
0.5146484375,
0.342529296875
],
[
0.139404296875,
0.336181640625,
0.489501953125,
0.336181640625,
0.489501953125,
0.349853515625,
0.139404296875,
0.349853515625
],
[
0.515625,
0.344482421875,
0.75244140625,
0.344482421875,
0.75244140625,
0.359375,
0.515625,
0.359375
],
[
0.12176513671875,
0.352783203125,
0.48779296875,
0.352783203125,
0.48779296875,
0.366455078125,
0.12176513671875,
0.366455078125
],
[
0.5322265625,
0.35986328125,
0.8818359375,
0.361083984375,
0.8818359375,
0.3759765625,
0.5322265625,
0.374755859375
],
[
0.12005615234375,
0.3681640625,
0.48779296875,
0.3681640625,
0.48779296875,
0.3818359375,
0.12005615234375,
0.3818359375
],
[
0.51708984375,
0.377685546875,
0.8818359375,
0.377685546875,
0.8818359375,
0.391357421875,
0.51708984375,
0.391357421875
],
[
0.12176513671875,
0.385986328125,
0.486083984375,
0.385986328125,
0.486083984375,
0.396728515625,
0.12176513671875,
0.396728515625
],
[
0.5146484375,
0.392578125,
0.880859375,
0.392578125,
0.880859375,
0.40625,
0.5146484375,
0.40625
],
[
0.12005615234375,
0.40087890625,
0.48779296875,
0.40087890625,
0.48779296875,
0.41455078125,
0.12005615234375,
0.41455078125
],
[
0.515625,
0.410400390625,
0.8818359375,
0.410400390625,
0.8818359375,
0.42333984375,
0.515625,
0.42333984375
],
[
0.12176513671875,
0.416259765625,
0.486083984375,
0.416259765625,
0.486083984375,
0.429931640625,
0.12176513671875,
0.429931640625
],
[
0.5146484375,
0.424560546875,
0.8818359375,
0.42578125,
0.8818359375,
0.440673828125,
0.5146484375,
0.439453125
],
[
0.12005615234375,
0.432861328125,
0.48779296875,
0.432861328125,
0.48779296875,
0.446533203125,
0.12005615234375,
0.446533203125
],
[
0.515625,
0.4423828125,
0.771484375,
0.4423828125,
0.771484375,
0.4560546875,
0.515625,
0.4560546875
],
[
0.12005615234375,
0.448974609375,
0.340087890625,
0.448974609375,
0.340087890625,
0.4619140625,
0.12005615234375,
0.4619140625
],
[
0.12176513671875,
0.474365234375,
0.490234375,
0.473876953125,
0.490234375,
0.488037109375,
0.12176513671875,
0.4892578125
],
[
0.5146484375,
0.473876953125,
0.5263671875,
0.473876953125,
0.5263671875,
0.48388671875,
0.5146484375,
0.48388671875
],
[
0.1175537109375,
0.4912109375,
0.486083984375,
0.4912109375,
0.486083984375,
0.5048828125,
0.1175537109375,
0.5048828125
],
[
0.51318359375,
0.498779296875,
0.6533203125,
0.498779296875,
0.6533203125,
0.5126953125,
0.51318359375,
0.5126953125
],
[
0.12005615234375,
0.5087890625,
0.484375,
0.5087890625,
0.484375,
0.51953125,
0.12005615234375,
0.51953125
],
[
0.5322265625,
0.515625,
0.8818359375,
0.515625,
0.8818359375,
0.529296875,
0.5322265625,
0.529296875
],
[
0.118408203125,
0.52392578125,
0.48779296875,
0.52392578125,
0.48779296875,
0.53759765625,
0.118408203125,
0.53759765625
],
[
0.5146484375,
0.53076171875,
0.8818359375,
0.52978515625,
0.8818359375,
0.54443359375,
0.5146484375,
0.5458984375
],
[
0.118408203125,
0.5390625,
0.48779296875,
0.5380859375,
0.48779296875,
0.552734375,
0.118408203125,
0.55419921875
],
[
0.51318359375,
0.54638671875,
0.88330078125,
0.54638671875,
0.88330078125,
0.56005859375,
0.51318359375,
0.56005859375
],
[
0.118408203125,
0.5556640625,
0.193115234375,
0.5556640625,
0.193115234375,
0.56787109375,
0.118408203125,
0.56787109375
],
[
0.5146484375,
0.5625,
0.72314453125,
0.5634765625,
0.72314453125,
0.57763671875,
0.5146484375,
0.57666015625
],
[
0.1368408203125,
0.572265625,
0.486083984375,
0.572265625,
0.486083984375,
0.5859375,
0.1368408203125,
0.5859375
],
[
0.130126953125,
0.58740234375,
0.48779296875,
0.58837890625,
0.48779296875,
0.6025390625,
0.130126953125,
0.6015625
],
[
0.12005615234375,
0.60498046875,
0.48779296875,
0.60498046875,
0.48779296875,
0.6181640625,
0.12005615234375,
0.6181640625
],
[
0.51123046875,
0.60693359375,
0.611328125,
0.609375,
0.611328125,
0.62255859375,
0.51123046875,
0.62060546875
],
[
0.12005615234375,
0.62060546875,
0.486083984375,
0.62060546875,
0.486083984375,
0.63134765625,
0.12005615234375,
0.63134765625
],
[
0.12176513671875,
0.6357421875,
0.48779296875,
0.6357421875,
0.48779296875,
0.6494140625,
0.12176513671875,
0.6494140625
],
[
0.51318359375,
0.63134765625,
0.88330078125,
0.63134765625,
0.88330078125,
0.64453125,
0.51318359375,
0.64453125
],
[
0.12005615234375,
0.65185546875,
0.48779296875,
0.65185546875,
0.48779296875,
0.6650390625,
0.12005615234375,
0.6650390625
],
[
0.5341796875,
0.646484375,
0.8818359375,
0.646484375,
0.8818359375,
0.65673828125,
0.5341796875,
0.65673828125
],
[
0.53076171875,
0.65771484375,
0.8818359375,
0.6591796875,
0.8818359375,
0.67041015625,
0.53076171875,
0.66943359375
],
[
0.12176513671875,
0.66845703125,
0.48779296875,
0.66845703125,
0.48779296875,
0.681640625,
0.12176513671875,
0.681640625
],
[
0.53076171875,
0.67138671875,
0.88330078125,
0.67138671875,
0.88330078125,
0.68505859375,
0.53076171875,
0.68505859375
],
[
0.12005615234375,
0.68408203125,
0.48779296875,
0.68408203125,
0.48779296875,
0.69775390625,
0.12005615234375,
0.69775390625
],
[
0.5322265625,
0.68505859375,
0.591796875,
0.68505859375,
0.591796875,
0.69677734375,
0.5322265625,
0.69677734375
],
[
0.118408203125,
0.69970703125,
0.489501953125,
0.70068359375,
0.489501953125,
0.71533203125,
0.118408203125,
0.71435546875
],
[
0.5146484375,
0.708984375,
0.8818359375,
0.708984375,
0.8818359375,
0.72265625,
0.5146484375,
0.72265625
],
[
0.118408203125,
0.71875,
0.204833984375,
0.71875,
0.204833984375,
0.72998046875,
0.118408203125,
0.72998046875
],
[
0.53076171875,
0.7216796875,
0.8818359375,
0.7216796875,
0.8818359375,
0.7353515625,
0.53076171875,
0.7353515625
],
[
0.5322265625,
0.7353515625,
0.7138671875,
0.73388671875,
0.7138671875,
0.74755859375,
0.5322265625,
0.7490234375
],
[
0.70361328125,
0.73388671875,
0.88330078125,
0.7353515625,
0.88330078125,
0.75,
0.70361328125,
0.7490234375
],
[
0.118408203125,
0.7451171875,
0.37451171875,
0.7451171875,
0.37451171875,
0.76025390625,
0.118408203125,
0.76025390625
],
[
0.53076171875,
0.74755859375,
0.8818359375,
0.74755859375,
0.8818359375,
0.76123046875,
0.53076171875,
0.76123046875
],
[
0.5322265625,
0.7626953125,
0.88330078125,
0.7626953125,
0.88330078125,
0.775390625,
0.5322265625,
0.775390625
],
[
0.12005615234375,
0.7724609375,
0.48779296875,
0.7724609375,
0.48779296875,
0.7861328125,
0.12005615234375,
0.7861328125
],
[
0.5322265625,
0.77490234375,
0.884765625,
0.77490234375,
0.884765625,
0.7880859375,
0.5322265625,
0.7880859375
],
[
0.118408203125,
0.78759765625,
0.48779296875,
0.78759765625,
0.48779296875,
0.80029296875,
0.118408203125,
0.80029296875
],
[
0.5322265625,
0.7890625,
0.88330078125,
0.7890625,
0.88330078125,
0.7998046875,
0.5322265625,
0.7998046875
],
[
0.12005615234375,
0.8046875,
0.459228515625,
0.8046875,
0.459228515625,
0.818359375,
0.12005615234375,
0.818359375
],
[
0.5322265625,
0.81103515625,
0.884765625,
0.8125,
0.884765625,
0.826171875,
0.5322265625,
0.82470703125
],
[
0.1234130859375,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.845703125,
0.1234130859375,
0.845703125
],
[
0.5146484375,
0.826171875,
0.8818359375,
0.826171875,
0.8818359375,
0.8359375,
0.5146484375,
0.8359375
],
[
0.5146484375,
0.83837890625,
0.88330078125,
0.83837890625,
0.88330078125,
0.8486328125,
0.5146484375,
0.8486328125
],
[
0.12176513671875,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.12176513671875,
0.86083984375
],
[
0.5146484375,
0.849609375,
0.8818359375,
0.849609375,
0.8818359375,
0.85986328125,
0.5146484375,
0.85986328125
],
[
0.12176513671875,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.12176513671875,
0.876953125
],
[
0.51318359375,
0.859375,
0.88330078125,
0.859375,
0.88330078125,
0.87255859375,
0.51318359375,
0.87255859375
],
[
0.51318359375,
0.87255859375,
0.611328125,
0.87255859375,
0.611328125,
0.88427734375,
0.51318359375,
0.88427734375
],
[
0.118408203125,
0.87890625,
0.48779296875,
0.87744140625,
0.48779296875,
0.892578125,
0.118408203125,
0.8935546875
],
[
0.5341796875,
0.8828125,
0.884765625,
0.88427734375,
0.884765625,
0.89794921875,
0.5341796875,
0.896484375
],
[
0.118408203125,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.51318359375,
0.89794921875,
0.7734375,
0.89794921875,
0.7734375,
0.90771484375,
0.51318359375,
0.90771484375
]
] |
[
"CoNLL",
"JFLEG",
"统生成一些校正，例如。g。拼写错误（becom",
"System",
"M?",
"GLEU",
"R",
"P",
"和dificullty）。另一方面，流水线可能无法改善",
"Best SMT",
"60.27 30.21 50.27",
"55.79",
"SMT系统（ischanged）进行的局部校正。。这",
"→ Pip. NMT",
"60.25",
"34.80",
"52.56",
"57.21",
"58.87 39.23 53.51",
"两种方法的组合产生的输出，这是最相似的引",
"→ Pip. NMT+LM",
"58.83",
"用",
"+ Res. RNN-LM",
"70.97 24.86 51.77",
"56.97",
"+ Res. NMT",
"70.40 26.69 53.03",
"57.21",
"Comparison with human annotationsBryant",
"+ Res.NMT+LM",
"71.40 28.60 54.95",
"57.53",
"andNg（2015）创建了CoNLL-2014测试集的扩",
"→ Pip. NMT+LM",
"65.73 33.3655.05",
"58.83",
"60.09",
"展，总共有10个注释器，JFLEG已经合并了",
"+ Spell SMT",
"70.80 30.57 56.05",
"→ Pip. NMT+LM",
"66.77 34.49 56.25",
"61.50",
"来自4个注释器的更正。M?和GLEU的人类",
"水平结果是通过平均每个注释者分别相对于其",
"Table 4: Results for hybrid SMT-NMT systems on",
"余9个（CoNLL）或3个（JFLEG）注释者的",
"the CoNLL-2014 ( M2) and JFLEG Test (GLEU)",
"得分来计算的",
"sets.",
"图E3包含人类水平得分，我们的结果以及",
"ChollampattandNg（2017)以前报告的最佳结",
"果。根据M²，我们最好的系统达到了人类平",
"中获得的",
"均得分的近100个%，而GLEU的近99个%",
"与流水线相反，重新计分以召回为代价提高",
"比以前的作品６更接近该界限",
"了精度，并且对于CoNLL数据更有效，导致多",
"但是，进一步检查发现，与人为校正相比，",
"达54。95M²。在JFLEG上，仅使用RNNLM",
"自动系统的精度/召回权衡表明覆盖率较低-较",
"进行重新评分会产生与使用NMT集成进行重",
"低的召回率由高精度7补偿。例如，自动系统",
"新评分相似的结果。但是，在该测试集上，重新",
"可能会错过一些明显的错误纠正，因此很容易",
"计分的最佳结果低于流水线的最佳结果。似乎",
"与人类参考区分开来。。未来的工作将需要人",
"SMT系统无法在n个最佳列表中产生与NMT",
"类的评估工作来得出更多的结论",
"集合所产生的一样多的校正",
"Spelling correction and final results 对 NMT-",
"7",
"rescoredSMT系统和NMT系统进行流水线操",
"Acknowledgments",
"作可进一步改进。我们认为，这可以通过两种",
"这项工作部分由Facebook资助。本文所包",
"方法对精度和召回权衡的不同贡献来解释，类",
"含的观点和结论是作者的观点和结论，不应被",
"似于NMT集成和我们的RNNLM组合所观察",
"解释为必然代表Facebook的官方政策或认可，",
"到的效果",
"无论是明示的还是暗示的。",
"在我们最终的混合系统之上，我们添加了",
"一个拼写检查组件，该组件在流水线之前运",
"行。我们使用ChollampattandNg（2017）之后",
"References",
"的字符级SMT系统：他们在基于单词的SMT",
"系统中部署未知单词。由于我们基于BPE的",
"Steven Bird and Edward Loper. 2004. NLTK: the nat",
"SMT并没有真正遭受未知单词的困扰，因此我",
"ural language toolkit.In Proceedingsof the ACL",
"2004 on Interactive poster and demonstration ses",
"们对将由BPE算法分段的单词运行拼写检查",
"sions. Association for Computational Linguistics,",
"组件。最后一个系统实现了本文报告的最佳结",
"page 31.",
"果:56。25M?onCoNLL-2014and61.50GLEUon-",
"Christopher Bryant and Hwee Tou Ng. 2015. How",
"JFLEGTest",
"far are we from fully automatic high quality gram",
"matical error correction?",
"In Proceedings of the",
"6Analysis and future work",
"53rd Annual Meeting of the Association for Compu",
"tational Linguistics and the 7th International Joint",
"对于这两个基准，我们的系统都接近自动评估",
"Conference on Natural Language Processing (Vol-",
"结果，这些结果已被声称与CoNLL-2014测试",
"ume l:Long Papers). Association for Computa-",
"集和JFLEG测试中的人类水平表现相对应",
"‘During the camera-ready preparation, Chollampatt and",
"Exampleoutputs表5显示了来自JFLEG测",
"Ng (2018) have published a GEC system based on a multi",
"layer convolutional encoder-decoder neural network with a",
"试语料库的示例源句子的系统输出，说明了统",
"character-based spell-checking module improving the previ",
"计和神经方法的互补性。SMT和NMT系统产",
"ous best result to 54.79 M? on CoNLL-2014 and 57.47GLEU",
"on JFLEG Test.",
"生不同的校正。重新评分能够生成唯一的校正",
"7A similar imbalance between precision and recall is visi-",
"（ischange一→haschanged），但它无法从神经系",
"bleonJFLEGwhentheM?metricisused."
] |
{
"arxiv_id": "1804.05945",
"language": "zh",
"page_idx": 4
}
|
||
2106.00793-zh-0
|
2106.00793-zh-0.jpg
|
# CoRI:与数据增强集成以进行开放式信息提取
$\begin{array}{c}{\mathrm{Zhengbao~Jiang^{1*},~Jialong~Han^{2},~Bunyamin~Sisman^{2},~Xin~Luna~Dong^{2}}}\\{\mathrm{Language~Technologies~Institute,~Carnegie~Mellon~University^{1}}}\\{\mathrm{Amazon^{2}}}\end{array}$
zhengbaj@cs.cmu.edu
$\{$ jialongh,bunyamis,lunadong $\}$ @amazon.com
$\operatorname{Abstract}$
Integrating extracted knowledge from the Web to knowledge graphs (KGs) can facilitate tasks like question answering. We study relation integration that aims to align free-text relations in subject-relationobject extractions to relations in a target KG. To address the challenge that free-text relations are ambiguous, previous methods exploit neighbor entities and relations for additional context. However, the predictions are made independently, which can be mutually inconsistent. We propose a two-stage Co llective R elation I ntegration (CoRI) model, where the frst stage independently makes candidate predictions, and the second stage employs a collective model that accesses all candidate predictions to make globally coherent predictions. We further improve the collective model with augmented data from the portion of the target KG that is otherwise unused. Experiment results on two datasets show that CoRI can significantly outperform the baselines, improving AUC from .677 to .748 and from .716 to .780, respectively.
## 1 Introduction
由于其庞大的容量，Web 已成为知识提取的主要资源。开放式信息提取(openIE; Sekine 2006; Banko et al.2007 ) 是一种突出的方法， 它可以在不假定一组预定义关系的情况下在自由文本中获取主题-关系-对象提取。增强下游应用程序(如问答)的一种方法是将这些自由文本提取集成到知识图(KG)e.g.,Freebase 中。。关系集成是集成这些提取的第一步，其中它们的自由文本关系(i.e.,源关系)被规范化为目标 KG (i.e.,目标关系)中的关系。只有在关系集成之后，实体链接才能继续将自由文本主题和对象解析为目标 KG 中的规范实体
${^*{\mathrm{~This~work~was~performed~while~at~Amazon.}}}$
<FigureHere>
Figure 1: A motivating example. Trained on parallel data , a local model may suffer from sparse context for a new entity pair Nell-Marie at inference, wrongly disambiguating“parent”to father instead of mother .
自然语言处理 (NLP) 社区研究了 Local Approaches. 关系集成。通过在源图和目标 KG 中的实体名称之间以文字形式进行精确匹配， 以前的方法可以像 Fig. 1 一样在两个图之间获得并行数据 i.e.,公共实体对。源图中实体对(e.g., Malia-Barack) 的特征及其在目标 KG ( e. g. , father ) 中的关系用于训练模型以预测未来提取的目标关系。一个常见的挑战是源关系的模糊性，e.g.,“父”可能在不同的上下文中对应于 father 或 mother 。先前的方法利用了上下文特征，包括嵌入可见实体(e.g.,“Malia”; Riedel et al. 2013 ), (e.g.,“parent";因Riedel et al. 2013; Toutanova et al. 2015; Verga et al. 2017,2016; Weston et al. 2013 ) 之间的中间关系以及实体对周围的邻居关系(e.g.,“性别”;$\mathbb{E}$Zhang et al. 2019 )
假设丰富的上下文来解决歧义挑战，以前的方法可能会在源图的不断发展和不完整的性质下不足。例如，在 Fig.1 的下部，新兴实体可能来自具有稀疏上下文信息的新提取。对于Nell-Marie 对，在并行数据上学习的传统模型可能既看不到实体，也看不到要依赖的邻域信息 (e.g.,“性别”),因此无法消除“父”的歧义并错误地预测 father 。由于先前方法的局
|
[
[
0.399658203125,
0.08197021484375,
0.60107421875,
0.0843505859375,
0.60107421875,
0.1015625,
0.399658203125,
0.09979248046875
],
[
0.359375,
0.1051025390625,
0.64208984375,
0.1051025390625,
0.64208984375,
0.119384765625,
0.359375,
0.119384765625
],
[
0.1796875,
0.144287109375,
0.8330078125,
0.144287109375,
0.8330078125,
0.1591796875,
0.1796875,
0.1591796875
],
[
0.2376708984375,
0.160888671875,
0.767578125,
0.160888671875,
0.767578125,
0.17578125,
0.2376708984375,
0.17578125
],
[
0.46337890625,
0.1776123046875,
0.5439453125,
0.17578125,
0.5439453125,
0.190673828125,
0.46337890625,
0.1923828125
],
[
0.41064453125,
0.19482421875,
0.59765625,
0.19482421875,
0.59765625,
0.208984375,
0.41064453125,
0.208984375
],
[
0.30908203125,
0.21142578125,
0.70947265625,
0.21142578125,
0.70947265625,
0.2257080078125,
0.30908203125,
0.2257080078125
],
[
0.264404296875,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.28076171875,
0.264404296875,
0.279052734375
],
[
0.6650390625,
0.267333984375,
0.72705078125,
0.267333984375,
0.72705078125,
0.27490234375,
0.6650390625,
0.27490234375
],
[
0.75830078125,
0.26416015625,
0.8369140625,
0.26416015625,
0.8369140625,
0.275634765625,
0.75830078125,
0.275634765625
],
[
0.51904296875,
0.27490234375,
0.57080078125,
0.27490234375,
0.57080078125,
0.28515625,
0.51904296875,
0.28515625
],
[
0.5615234375,
0.2861328125,
0.58447265625,
0.2861328125,
0.58447265625,
0.29345703125,
0.5615234375,
0.29345703125
],
[
0.6572265625,
0.283935546875,
0.6962890625,
0.283935546875,
0.6962890625,
0.29150390625,
0.6572265625,
0.29150390625
],
[
0.14697265625,
0.29150390625,
0.459228515625,
0.29150390625,
0.459228515625,
0.3046875,
0.14697265625,
0.3046875
],
[
0.58447265625,
0.29150390625,
0.615234375,
0.29150390625,
0.615234375,
0.303955078125,
0.58447265625,
0.303955078125
],
[
0.68408203125,
0.290283203125,
0.705078125,
0.295654296875,
0.7001953125,
0.303955078125,
0.67919921875,
0.298583984375
],
[
0.791015625,
0.297607421875,
0.8154296875,
0.297607421875,
0.8154296875,
0.3017578125,
0.791015625,
0.3017578125
],
[
0.145263671875,
0.303955078125,
0.4609375,
0.3046875,
0.4609375,
0.319580078125,
0.145263671875,
0.318359375
],
[
0.5693359375,
0.305908203125,
0.60205078125,
0.305908203125,
0.60205078125,
0.314208984375,
0.5693359375,
0.314208984375
],
[
0.6708984375,
0.31005859375,
0.68408203125,
0.31005859375,
0.68408203125,
0.3134765625,
0.6708984375,
0.3134765625
],
[
0.6171875,
0.3154296875,
0.646484375,
0.3154296875,
0.646484375,
0.318359375,
0.6171875,
0.318359375
],
[
0.5322265625,
0.311279296875,
0.56494140625,
0.311279296875,
0.56494140625,
0.319580078125,
0.5322265625,
0.319580078125
],
[
0.5498046875,
0.318359375,
0.623046875,
0.320556640625,
0.623046875,
0.33203125,
0.5498046875,
0.330078125
],
[
0.79248046875,
0.31640625,
0.82861328125,
0.31640625,
0.82861328125,
0.32373046875,
0.79248046875,
0.32373046875
],
[
0.63232421875,
0.322509765625,
0.66748046875,
0.32177734375,
0.66748046875,
0.328857421875,
0.63232421875,
0.330078125
],
[
0.84814453125,
0.33203125,
0.875,
0.33203125,
0.875,
0.34033203125,
0.84814453125,
0.34033203125
],
[
0.14697265625,
0.320556640625,
0.4609375,
0.320556640625,
0.4609375,
0.334228515625,
0.14697265625,
0.334228515625
],
[
0.14697265625,
0.334228515625,
0.4609375,
0.334228515625,
0.4609375,
0.347412109375,
0.14697265625,
0.347412109375
],
[
0.82861328125,
0.341552734375,
0.86962890625,
0.341552734375,
0.86962890625,
0.352783203125,
0.82861328125,
0.352783203125
],
[
0.1485595703125,
0.347412109375,
0.459228515625,
0.3486328125,
0.459228515625,
0.3623046875,
0.1485595703125,
0.361083984375
],
[
0.51904296875,
0.3515625,
0.54833984375,
0.3515625,
0.54833984375,
0.35986328125,
0.51904296875,
0.35986328125
],
[
0.5751953125,
0.350830078125,
0.61279296875,
0.350830078125,
0.61279296875,
0.35986328125,
0.5751953125,
0.35986328125
],
[
0.646484375,
0.352783203125,
0.66748046875,
0.352783203125,
0.66748046875,
0.35693359375,
0.646484375,
0.35693359375
],
[
0.705078125,
0.352783203125,
0.73486328125,
0.352783203125,
0.73486328125,
0.35693359375,
0.705078125,
0.35693359375
],
[
0.57861328125,
0.359375,
0.62451171875,
0.361083984375,
0.623046875,
0.372314453125,
0.57666015625,
0.37060546875
],
[
0.6337890625,
0.3623046875,
0.67529296875,
0.3623046875,
0.67529296875,
0.369384765625,
0.6337890625,
0.369384765625
],
[
0.72119140625,
0.35986328125,
0.76416015625,
0.35986328125,
0.76416015625,
0.37060546875,
0.72119140625,
0.37060546875
],
[
0.771484375,
0.3623046875,
0.8505859375,
0.3623046875,
0.8505859375,
0.369384765625,
0.771484375,
0.369384765625
],
[
0.5263671875,
0.364013671875,
0.568359375,
0.364013671875,
0.568359375,
0.3759765625,
0.5263671875,
0.3759765625
],
[
0.1502685546875,
0.363525390625,
0.456787109375,
0.363525390625,
0.456787109375,
0.37353515625,
0.1502685546875,
0.37353515625
],
[
0.56103515625,
0.371826171875,
0.603515625,
0.371826171875,
0.603515625,
0.3818359375,
0.56103515625,
0.3818359375
],
[
0.69189453125,
0.3759765625,
0.72705078125,
0.37353515625,
0.72705078125,
0.38134765625,
0.69287109375,
0.383056640625
],
[
0.77099609375,
0.371826171875,
0.86962890625,
0.371826171875,
0.86962890625,
0.37890625,
0.77099609375,
0.37890625
],
[
0.78662109375,
0.37890625,
0.85986328125,
0.37890625,
0.85986328125,
0.3896484375,
0.78662109375,
0.3896484375
],
[
0.72314453125,
0.38427734375,
0.74462890625,
0.385986328125,
0.744140625,
0.3955078125,
0.72119140625,
0.393798828125
],
[
0.55517578125,
0.385498046875,
0.57666015625,
0.385498046875,
0.57666015625,
0.39794921875,
0.55517578125,
0.39794921875
],
[
0.14697265625,
0.377197265625,
0.459228515625,
0.377197265625,
0.459228515625,
0.39013671875,
0.14697265625,
0.39013671875
],
[
0.59619140625,
0.385498046875,
0.623046875,
0.385498046875,
0.623046875,
0.396728515625,
0.59619140625,
0.396728515625
],
[
0.68408203125,
0.388427734375,
0.69775390625,
0.388427734375,
0.69775390625,
0.394287109375,
0.68408203125,
0.394287109375
],
[
0.78271484375,
0.388427734375,
0.8818359375,
0.388427734375,
0.8818359375,
0.3984375,
0.78271484375,
0.3984375
],
[
0.54736328125,
0.39794921875,
0.615234375,
0.399658203125,
0.615234375,
0.410888671875,
0.54736328125,
0.4091796875
],
[
0.14697265625,
0.391357421875,
0.459228515625,
0.391357421875,
0.459228515625,
0.405029296875,
0.14697265625,
0.405029296875
],
[
0.14697265625,
0.405029296875,
0.459228515625,
0.405029296875,
0.459228515625,
0.418701171875,
0.14697265625,
0.418701171875
],
[
0.6533203125,
0.399658203125,
0.76416015625,
0.399658203125,
0.76416015625,
0.410400390625,
0.6533203125,
0.410400390625
],
[
0.1485595703125,
0.42041015625,
0.45751953125,
0.42041015625,
0.45751953125,
0.43115234375,
0.1485595703125,
0.43115234375
],
[
0.5146484375,
0.421630859375,
0.88330078125,
0.421630859375,
0.88330078125,
0.435302734375,
0.5146484375,
0.435302734375
],
[
0.14697265625,
0.43408203125,
0.459228515625,
0.43408203125,
0.459228515625,
0.44775390625,
0.14697265625,
0.44775390625
],
[
0.5146484375,
0.43701171875,
0.880859375,
0.43701171875,
0.880859375,
0.44775390625,
0.5146484375,
0.44775390625
],
[
0.14697265625,
0.44775390625,
0.459228515625,
0.44775390625,
0.459228515625,
0.46142578125,
0.14697265625,
0.46142578125
],
[
0.5146484375,
0.45068359375,
0.884765625,
0.45068359375,
0.884765625,
0.46435546875,
0.5146484375,
0.46435546875
],
[
0.14697265625,
0.46142578125,
0.4609375,
0.4619140625,
0.4609375,
0.4755859375,
0.14697265625,
0.474365234375
],
[
0.5146484375,
0.46435546875,
0.87744140625,
0.46435546875,
0.87744140625,
0.478515625,
0.5146484375,
0.478515625
],
[
0.145263671875,
0.474365234375,
0.4609375,
0.4755859375,
0.4609375,
0.490478515625,
0.145263671875,
0.4892578125
],
[
0.5146484375,
0.478515625,
0.6513671875,
0.479736328125,
0.6513671875,
0.4912109375,
0.5146484375,
0.490478515625
],
[
0.14697265625,
0.490478515625,
0.459228515625,
0.490478515625,
0.459228515625,
0.50439453125,
0.14697265625,
0.50439453125
],
[
0.1485595703125,
0.5048828125,
0.4609375,
0.5048828125,
0.4609375,
0.5185546875,
0.1485595703125,
0.5185546875
],
[
0.14697265625,
0.5185546875,
0.459228515625,
0.5185546875,
0.459228515625,
0.5322265625,
0.14697265625,
0.5322265625
],
[
0.53564453125,
0.521484375,
0.8818359375,
0.5224609375,
0.8818359375,
0.53759765625,
0.53564453125,
0.5361328125
],
[
0.1485595703125,
0.533203125,
0.459228515625,
0.533203125,
0.459228515625,
0.54638671875,
0.1485595703125,
0.54638671875
],
[
0.51318359375,
0.5380859375,
0.8818359375,
0.5380859375,
0.8818359375,
0.5517578125,
0.51318359375,
0.5517578125
],
[
0.14697265625,
0.54638671875,
0.4609375,
0.54638671875,
0.4609375,
0.56005859375,
0.14697265625,
0.56005859375
],
[
0.515625,
0.55419921875,
0.8818359375,
0.5546875,
0.8818359375,
0.5693359375,
0.515625,
0.568359375
],
[
0.14697265625,
0.56103515625,
0.459228515625,
0.56103515625,
0.459228515625,
0.57470703125,
0.14697265625,
0.57470703125
],
[
0.14697265625,
0.576171875,
0.459228515625,
0.576171875,
0.459228515625,
0.5888671875,
0.14697265625,
0.5888671875
],
[
0.515625,
0.57177734375,
0.8818359375,
0.57177734375,
0.8818359375,
0.5849609375,
0.515625,
0.5849609375
],
[
0.1485595703125,
0.59130859375,
0.459228515625,
0.59130859375,
0.459228515625,
0.6015625,
0.1485595703125,
0.6015625
],
[
0.5146484375,
0.5859375,
0.880859375,
0.5859375,
0.880859375,
0.599609375,
0.5146484375,
0.599609375
],
[
0.1485595703125,
0.60400390625,
0.4609375,
0.60400390625,
0.4609375,
0.61767578125,
0.1485595703125,
0.61767578125
],
[
0.5146484375,
0.6025390625,
0.88330078125,
0.6015625,
0.88330078125,
0.6162109375,
0.5146484375,
0.61767578125
],
[
0.14697265625,
0.6181640625,
0.459228515625,
0.6181640625,
0.459228515625,
0.6318359375,
0.14697265625,
0.6318359375
],
[
0.5146484375,
0.619140625,
0.8818359375,
0.619140625,
0.8818359375,
0.6328125,
0.5146484375,
0.6328125
],
[
0.14697265625,
0.63134765625,
0.459228515625,
0.6328125,
0.459228515625,
0.646484375,
0.14697265625,
0.64453125
],
[
0.51708984375,
0.63525390625,
0.8818359375,
0.63525390625,
0.8818359375,
0.6494140625,
0.51708984375,
0.6494140625
],
[
0.145263671875,
0.646484375,
0.1956787109375,
0.646484375,
0.1956787109375,
0.6611328125,
0.145263671875,
0.6611328125
],
[
0.51708984375,
0.65185546875,
0.8818359375,
0.65185546875,
0.8818359375,
0.66650390625,
0.51708984375,
0.66650390625
],
[
0.154541015625,
0.67138671875,
0.26611328125,
0.67138671875,
0.26611328125,
0.68603515625,
0.154541015625,
0.68603515625
],
[
0.5146484375,
0.66650390625,
0.8818359375,
0.66748046875,
0.8818359375,
0.681640625,
0.5146484375,
0.68115234375
],
[
0.12005615234375,
0.67333984375,
0.1351318359375,
0.67578125,
0.1318359375,
0.68603515625,
0.1175537109375,
0.68408203125
],
[
0.51904296875,
0.68408203125,
0.88671875,
0.68408203125,
0.88671875,
0.69873046875,
0.51904296875,
0.69873046875
],
[
0.118408203125,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.71044921875,
0.118408203125,
0.71044921875
],
[
0.51318359375,
0.69873046875,
0.884765625,
0.69873046875,
0.884765625,
0.71630859375,
0.51318359375,
0.71630859375
],
[
0.12005615234375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.7255859375,
0.12005615234375,
0.7255859375
],
[
0.5146484375,
0.71630859375,
0.880859375,
0.71630859375,
0.880859375,
0.72998046875,
0.5146484375,
0.72998046875
],
[
0.118408203125,
0.728515625,
0.489501953125,
0.72998046875,
0.489501953125,
0.74462890625,
0.118408203125,
0.74365234375
],
[
0.51318359375,
0.73095703125,
0.8818359375,
0.732421875,
0.8818359375,
0.74658203125,
0.51318359375,
0.7451171875
],
[
0.12005615234375,
0.74462890625,
0.486083984375,
0.7451171875,
0.486083984375,
0.76025390625,
0.12005615234375,
0.7587890625
],
[
0.51708984375,
0.7490234375,
0.88330078125,
0.7490234375,
0.88330078125,
0.76318359375,
0.51708984375,
0.76318359375
],
[
0.12176513671875,
0.7626953125,
0.48779296875,
0.7626953125,
0.48779296875,
0.775390625,
0.12176513671875,
0.775390625
],
[
0.515625,
0.76416015625,
0.72705078125,
0.76416015625,
0.72705078125,
0.779296875,
0.515625,
0.779296875
],
[
0.12005615234375,
0.77685546875,
0.484375,
0.77685546875,
0.484375,
0.79052734375,
0.12005615234375,
0.79052734375
],
[
0.5322265625,
0.78076171875,
0.8818359375,
0.7822265625,
0.8818359375,
0.79638671875,
0.5322265625,
0.7958984375
],
[
0.12176513671875,
0.79443359375,
0.48779296875,
0.79443359375,
0.48779296875,
0.80810546875,
0.12176513671875,
0.80810546875
],
[
0.515625,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.515625,
0.8125
],
[
0.12176513671875,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82373046875,
0.12176513671875,
0.82373046875
],
[
0.51904296875,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51904296875,
0.82763671875
],
[
0.12005615234375,
0.82666015625,
0.486083984375,
0.82666015625,
0.486083984375,
0.84033203125,
0.12005615234375,
0.84033203125
],
[
0.5146484375,
0.83056640625,
0.880859375,
0.83056640625,
0.880859375,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.12005615234375,
0.8427734375,
0.48779296875,
0.8427734375,
0.48779296875,
0.85546875,
0.12005615234375,
0.85546875
],
[
0.51123046875,
0.845703125,
0.8818359375,
0.8466796875,
0.8818359375,
0.86083984375,
0.51123046875,
0.85986328125
],
[
0.12005615234375,
0.85693359375,
0.489501953125,
0.85791015625,
0.489501953125,
0.87255859375,
0.12005615234375,
0.87158203125
],
[
0.515625,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.515625,
0.876953125
],
[
0.12005615234375,
0.87451171875,
0.47021484375,
0.87451171875,
0.47021484375,
0.88818359375,
0.12005615234375,
0.88818359375
],
[
0.515625,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.515625,
0.8935546875
],
[
0.1485595703125,
0.8955078125,
0.453369140625,
0.8955078125,
0.453369140625,
0.9091796875,
0.1485595703125,
0.9091796875
],
[
0.51708984375,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.51708984375,
0.9091796875
]
] |
[
"CoRI：与数据增强",
"集成以进行开放式信息提取",
"Zhengbao Jiangl*Jialong Han?,Bunyamin Sisman?,Xin Luna Dong",
"Language Technologies Institute, Carnegie Mellon University",
"Amazon?",
"zhengbaj@cs.cmu.edu",
" jialongh,bunyamis,lunadong } @amazon.com",
"Abstract",
"Paralleldata",
"UnmatchedKG",
"Michelle",
"bare",
"mothe",
"Integrating extracted knowledge from the",
"Mali",
"",
"aoth",
"Web to knowledge graphs (KGs) can facil-",
"parent",
"a+",
"ar",
"Barack",
"gender-male",
"father",
"qende",
"Train",
"itate tasks like question answering. We",
"study relation integration that aims to",
"Inferenc",
"align free-text relations in subject-relation-",
"Marie",
"Burton",
"var",
"utor",
"ffather",
"fathe",
"father",
"sourceentity",
"parent",
"object extractions to relations in a target",
"parent",
"ather",
"target entity",
"source relation",
"Billy",
"Nell",
"KG. To address the challenge that free-text",
"Billy",
"Jel",
"target relation",
"Source graph",
"relations are ambiguous, previous methods",
"exploit neighbor entities and relations for",
"Predicted target graph",
"additional context.However,the predic",
"Figure 1: A motivating example. Trained on par-",
"tions are made independently, which can",
"allel dataa local model may suffer from sparse",
"be mutually inconsistent.We propose a",
"context for a new entity pair Nell-Marie at infer-",
"two-stage Co llective R elation I ntegra-",
"ence, wrongly disambiguating “parent\" to father",
"tion (CoRI) model, where the first stage in-",
"insteadof mother",
"dependently makes candidate predictions,",
"and the second stage employs a collec-",
"tive model that accesses all candidate pre",
"自然语言处理（NLP）社区研究了LocalAp",
"dictions to make globally coherent predic-",
"proaches.关系集成。通过在源图和目标KG",
"tions.We further improve the collective",
"中的实体名称之间以文字形式进行精确匹配，",
"model with augmented data from the por",
"tion of the target KG that is otherwise un-",
"以前的方法可以像Fig.1一样在两个图之间",
"used.Experiment results on two datasets",
"获得并行数据i.e..公共实体对。源图中实体",
"show that CoRI can significantly outper-",
"对（e.g.，Malia-Barack）的特征及其在目标KG",
"form the baselines, improving AUC from",
"（e.g.，father）中的关系用于训练模型以预测未",
".677to.748and from.716to.780,respec-",
"来提取的目标关系。一个常见的挑战是源关系",
"tively.",
"的模糊性，e.g.，“父”可能在不同的上下文中对",
"Introduction",
"应于father或mother。先前的方法利用了上",
"1",
"下文特征，包括嵌人可见实体（e.g，“Malia\"；",
"由于其庞大的容量，Web已成为知识提取的",
"Riedel et al. 2013 ), (e.g., “\"parent\";ERiedel",
"主要资源。开放式信息提取（openIE：Sekine",
"et al. 2013; Toutanova et al. 2015; Verga et al.",
"2006；Bankoetal.2007）是一种突出的方法，",
"2017，2016；Westonetal.2013）之间的中间",
"它可以在不假定一组预定义关系的情况下在自",
"关系以及实体对周围的邻居关系（e.g.，“性",
"由文本中获取主题-关系-对象提取。增强下游",
"别\";EZhang et al.2019)",
"应用程序（如问答）的一种方法是将这些自由",
"假设丰富的上下文来解决歧义挑战，以前的",
"文本提取集成到知识图（KG）e.g.，Freebase",
"方法可能会在源图的不断发展和不完整的性质",
"中。。关系集成是集成这些提取的第一步，其中",
"下不足。例如，在Fig.1的下部，新兴实体",
"它们的自由文本关系（i.e.，源关系）被规范化",
"可能来自具有稀疏上下文信息的新提取。对于",
"为目标KG（i.e.，目标关系）中的关系。只有",
"Nell-Marie对，在并行数据上学习的传统模型",
"在关系集成之后，实体链接才能继续将自由文",
"可能既看不到实体，也看不到要依赖的邻域信",
"本主题和对象解析为目标KG中的规范实体",
"息（e.g，“性别\"），因此无法消除“父”的歧",
"* This work was performed while at Amazon.",
"义并错误地预测father。由于先前方法的局"
] |
{
"arxiv_id": "2106.00793",
"language": "zh",
"page_idx": 0
}
|
||
2106.00793-zh-1
|
2106.00793-zh-1.jpg
|
Table 1: Comparisons between CoRI and baselines.
部性，不同实体对的 i.e., 预测是彼此独立进行的，因此模型不知道“Nell”在最终预测中有两个父亲。这样的预测在常识上是不连贯的， 一个人更有可能有一个父亲和一个母亲，这由并行数据的目标 KG 部分中 Malia 周围的图形结构表示
1.1 Our Collective Approach
为了缓解局部方法的不相干预测问题，我们提出了协作关系集成 (CoRI),它利用相邻实体对之间的预测依赖性来增强全局一致性
具体来说，我们遵循两个阶段，i.e.,候选生成和集体推理。在候选生成中，我们只需使用局部模型作为候选进行独立预测，Fig.1 下部的所有三对 e.g., father 。在集体推理中，我们采用一个集体模型，该模型知道目标图的常见子结构 e.g., Malia 。集体模型不仅通过将所有上下文特征作为本地模型的输入，而且还通过将当前和所有邻居对的候选预测作为输入来进行预测。对于 Nell-Marie 对，集体模型将可以访问 Nell-Burton 的候选预测 father ,这有助于将其最终预测翻转为正确的 mother 。Tab. 1 从四个方面总结了 CoRI 和代表性的先前工作。据我们所知，CoRI 是第一个集体执行关系集成而不是本地的
负责进行全局一致的预测，需要对集体模型进行训练，以对目标 KG 的通用结构进行编码，e.g., Malia 在 Fig. 1 的并行数据中只有一个父亲/母亲。为此，我们以堆叠的方式训练集体模型 (Wolpert, 1992)。我们首先在并行数据上训练第一阶段局部模型，然后通过对第一阶段的邻居实体对的候选预测进行条件化(e.g., father 代表 Malia-Barrack)来训练第二阶段集体模型，以做出全局一致的预测 (e.g., mother 代表 Malia-Michelle)
Parallel Data Augmentation. 并行数据可能受到精确名称匹配的低召回率或开放 IE 系统生成的有限提取的限制。我们观察到，即使没有对应物提取，目标图的 unmatched 部分(如在 Fig.1 中) 也可能具有丰富的通用结构来指导集体模型的训练。为此，我们建议通过从unmatched KG 中采样子图来增强并行数据， 并通过合成其提取来创建伪并行数据，因此集
总而言之，我们的贡献有三个方面：(1)我
<table>
<tbody>
<tr>
<th>Methods</th>
<th>Middle relation</th>
<th>entity $No$ param.</th>
<th>Neighbor relation</th>
<th>r Collective体模型 $\operatorname{inference}$ 数据口</th>
</tr>
</tbody>
</table>
们提出了 CoRI, 这是一个两阶段框架，通过进行具有全局一致性的集体预测来改进最先{}进的方法。(2)我们建议使用 unmatched 目标
KG 来增加训练数据。(3)在两个数据集上的实验结果表明了我们方法的优越性，从。677至。748 和。716 至。780, 分别
### 2 Preliminaries
在本节中，我们首先制定关系集成的任务，然后通过使用最先进的方法 OpenKIE(Zhang et al., 2019) 举例说明本地方法
2.1 Relation Integration
我们将来自开放 IE 系统的主题-关系-对象提取视为源图$\mathcal{K}(\mathcal{E},\mathcal{R})=\{(s,r,o)\mid s,o\in$ $\mathcal{E} , r\in \mathcal{R} \}$ ,其中$\mathcal{E}$表示提取的文本实体，e.g. “BarackObama”, $\mathcal{R}$表示提取的源关系，e.g., “parent”。我们用$(s,o)$表示源实体对。对于$( s, o)$ , ${\mathcal{K} }_{s, o}= \{ r$ | $( s, r, o) \in {\mathcal{K} }\}$表示它们之间的所有源关系。类似地$,\mathcal{K}_r=\{s,o\mid(s,r,o)\in$ $\mathcal{K}\}$表示之间具有关系$r$的所有实体对。我们使用联合$\mathcal{K}_{\mathcal{R}}=\bigcup_{r\in\mathcal{R}}\mathcal{K}_r$来引用所有提取的实体对
Definition 1( Relation Integration ). Given a source graph $\mathcal{K}$ and a target KG $\mathcal{K}^\prime(\mathcal{E}^\prime,\mathcal{R}^\prime)$ with target entities $\mathcal{E}^\prime$ and target relations $\mathcal{R}^\prime$, the task of relation integration is to predict all applicable target relations for each extracted entity pair in $\mathcal{K}_{\mathcal{R}}:$
$$\Gamma\subseteq\mathcal{K}_{\mathcal{R}}\times\mathcal{R}^{\prime},$$
where $(s,r^{\prime},o)\in\Gamma$ is an integrated extraction indicating that a target relation $r^\prime$ holds for $( s, o)$ .
为了训练关系集成模型，所有方法都采用如
下形式的并行数据：
Definition 2 ( Parallel Data ). Parallel data are common entity pairs shared between $\mathcal{K}_{\mathcal{R}}$ and $\mathcal{K}_{\mathcal{R}^{\prime}}^{\prime}$ and their ground truth target relations in $\mathcal{K} ^{\prime }: \mathcal{T} = \{ \langle ( s, o) , \mathcal{K} _{s, o}^{\prime }\rangle \mid ( s, o) \in \mathcal{K} _{\mathcal{R} }\cap \mathcal{K} _{\mathcal{R} ^{\prime }}^{\prime }\}$ . For example, $\langle($Malia,Barack$),\{$father$\}\rangle$ is an instance of parallel data in Fig. 1 .
为了获得并行数据，一种广泛使用的方法是通过精确名称匹配找到$\mathcal{E}$和$\mathcal{E}^\prime$共享的实体， 然后生成公共实体对及其基本事实
|
[
[
0.124267578125,
0.08074951171875,
0.18896484375,
0.0831298828125,
0.187255859375,
0.097412109375,
0.1234130859375,
0.0955810546875
],
[
0.2626953125,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.2626953125,
0.09027099609375
],
[
0.26025390625,
0.0885009765625,
0.32421875,
0.09027099609375,
0.32421875,
0.1015625,
0.26025390625,
0.09979248046875
],
[
0.3837890625,
0.0843505859375,
0.609375,
0.0914306640625,
0.60791015625,
0.1092529296875,
0.382080078125,
0.1015625
],
[
0.32080078125,
0.0914306640625,
0.3720703125,
0.0914306640625,
0.3720703125,
0.1015625,
0.32080078125,
0.1015625
],
[
0.12176513671875,
0.10687255859375,
0.2568359375,
0.10687255859375,
0.2568359375,
0.12054443359375,
0.12176513671875,
0.12054443359375
],
[
0.274658203125,
0.11102294921875,
0.2880859375,
0.10687255859375,
0.295654296875,
0.1175537109375,
0.2822265625,
0.1217041015625
],
[
0.5341796875,
0.10986328125,
0.8818359375,
0.10986328125,
0.8818359375,
0.12469482421875,
0.5341796875,
0.12469482421875
],
[
0.12005615234375,
0.119384765625,
0.25439453125,
0.119384765625,
0.25439453125,
0.133056640625,
0.12005615234375,
0.133056640625
],
[
0.2822265625,
0.1181640625,
0.295654296875,
0.12353515625,
0.2880859375,
0.1341552734375,
0.273681640625,
0.127685546875
],
[
0.341796875,
0.12054443359375,
0.355224609375,
0.12054443359375,
0.355224609375,
0.1312255859375,
0.341796875,
0.1312255859375
],
[
0.12176513671875,
0.1312255859375,
0.2568359375,
0.1312255859375,
0.2568359375,
0.144287109375,
0.12176513671875,
0.144287109375
],
[
0.5146484375,
0.1270751953125,
0.880859375,
0.1270751953125,
0.880859375,
0.14013671875,
0.5146484375,
0.14013671875
],
[
0.28369140625,
0.1353759765625,
0.28955078125,
0.135986328125,
0.2880859375,
0.142578125,
0.2822265625,
0.141357421875
],
[
0.12005615234375,
0.141357421875,
0.204833984375,
0.1436767578125,
0.204833984375,
0.157958984375,
0.12005615234375,
0.1561279296875
],
[
0.344970703125,
0.1455078125,
0.35107421875,
0.1455078125,
0.35107421875,
0.152587890625,
0.344970703125,
0.152587890625
],
[
0.51318359375,
0.141357421875,
0.8818359375,
0.142578125,
0.8818359375,
0.15673828125,
0.51318359375,
0.1561279296875
],
[
0.27783203125,
0.1484375,
0.2880859375,
0.1436767578125,
0.2939453125,
0.1484375,
0.28369140625,
0.15380859375
],
[
0.515625,
0.1591796875,
0.8818359375,
0.1591796875,
0.8818359375,
0.1728515625,
0.515625,
0.1728515625
],
[
0.118408203125,
0.171630859375,
0.489501953125,
0.171630859375,
0.489501953125,
0.185302734375,
0.118408203125,
0.185302734375
],
[
0.5146484375,
0.174560546875,
0.8818359375,
0.174560546875,
0.8818359375,
0.188232421875,
0.5146484375,
0.188232421875
],
[
0.51708984375,
0.1923828125,
0.8818359375,
0.1923828125,
0.8818359375,
0.203125,
0.51708984375,
0.203125
],
[
0.515625,
0.207275390625,
0.77685546875,
0.207275390625,
0.77685546875,
0.2203369140625,
0.515625,
0.2203369140625
],
[
0.12176513671875,
0.21435546875,
0.486083984375,
0.21435546875,
0.486083984375,
0.22802734375,
0.12176513671875,
0.22802734375
],
[
0.12005615234375,
0.2286376953125,
0.486083984375,
0.2286376953125,
0.486083984375,
0.243408203125,
0.12005615234375,
0.243408203125
],
[
0.5205078125,
0.2381591796875,
0.53076171875,
0.2452392578125,
0.5224609375,
0.25048828125,
0.51318359375,
0.24462890625
],
[
0.5322265625,
0.2423095703125,
0.54736328125,
0.2423095703125,
0.54736328125,
0.2464599609375,
0.5322265625,
0.2464599609375
],
[
0.54833984375,
0.2381591796875,
0.66650390625,
0.2381591796875,
0.66650390625,
0.251708984375,
0.54833984375,
0.251708984375
],
[
0.12005615234375,
0.2452392578125,
0.484375,
0.2452392578125,
0.484375,
0.2587890625,
0.12005615234375,
0.2587890625
],
[
0.1318359375,
0.260009765625,
0.48779296875,
0.262451171875,
0.48779296875,
0.275634765625,
0.1318359375,
0.273681640625
],
[
0.5146484375,
0.2666015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.27978515625,
0.5146484375,
0.27978515625
],
[
0.12005615234375,
0.27783203125,
0.48779296875,
0.27783203125,
0.48779296875,
0.29150390625,
0.12005615234375,
0.29150390625
],
[
0.51318359375,
0.28076171875,
0.8818359375,
0.281982421875,
0.8818359375,
0.29638671875,
0.51318359375,
0.295654296875
],
[
0.12176513671875,
0.295654296875,
0.21240234375,
0.295654296875,
0.21240234375,
0.30712890625,
0.12176513671875,
0.30712890625
],
[
0.51318359375,
0.298583984375,
0.76416015625,
0.298583984375,
0.76416015625,
0.312255859375,
0.51318359375,
0.312255859375
],
[
0.12005615234375,
0.32470703125,
0.1695556640625,
0.32470703125,
0.1695556640625,
0.336181640625,
0.12005615234375,
0.336181640625
],
[
0.166259765625,
0.322509765625,
0.36865234375,
0.322509765625,
0.36865234375,
0.336181640625,
0.166259765625,
0.336181640625
],
[
0.51123046875,
0.32666015625,
0.7255859375,
0.327880859375,
0.7255859375,
0.342529296875,
0.51123046875,
0.341552734375
],
[
0.12176513671875,
0.345703125,
0.489501953125,
0.345703125,
0.489501953125,
0.359375,
0.12176513671875,
0.359375
],
[
0.5146484375,
0.350830078125,
0.8818359375,
0.350830078125,
0.8818359375,
0.364013671875,
0.5146484375,
0.364013671875
],
[
0.12005615234375,
0.361083984375,
0.489501953125,
0.361083984375,
0.489501953125,
0.374755859375,
0.12005615234375,
0.374755859375
],
[
0.5146484375,
0.36767578125,
0.8818359375,
0.36767578125,
0.8818359375,
0.38134765625,
0.5146484375,
0.38134765625
],
[
0.12005615234375,
0.377685546875,
0.43408203125,
0.377685546875,
0.43408203125,
0.391357421875,
0.12005615234375,
0.391357421875
],
[
0.51318359375,
0.3818359375,
0.8818359375,
0.383056640625,
0.8818359375,
0.39794921875,
0.51318359375,
0.396728515625
],
[
0.139404296875,
0.394287109375,
0.48779296875,
0.394287109375,
0.48779296875,
0.407958984375,
0.139404296875,
0.407958984375
],
[
0.51123046875,
0.3984375,
0.880859375,
0.40087890625,
0.880859375,
0.41455078125,
0.51123046875,
0.412109375
],
[
0.12176513671875,
0.410888671875,
0.48779296875,
0.410888671875,
0.48779296875,
0.424560546875,
0.12176513671875,
0.424560546875
],
[
0.51123046875,
0.41455078125,
0.8818359375,
0.4150390625,
0.8818359375,
0.429931640625,
0.51123046875,
0.4287109375
],
[
0.12005615234375,
0.427001953125,
0.489501953125,
0.427001953125,
0.489501953125,
0.440673828125,
0.12005615234375,
0.440673828125
],
[
0.51318359375,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.51318359375,
0.4453125
],
[
0.12005615234375,
0.443603515625,
0.489501953125,
0.443603515625,
0.489501953125,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.51318359375,
0.44775390625,
0.8818359375,
0.44775390625,
0.8818359375,
0.46142578125,
0.51318359375,
0.46142578125
],
[
0.12005615234375,
0.458984375,
0.48779296875,
0.458984375,
0.48779296875,
0.47265625,
0.12005615234375,
0.47265625
],
[
0.51123046875,
0.46435546875,
0.880859375,
0.46435546875,
0.880859375,
0.47802734375,
0.51123046875,
0.47802734375
],
[
0.12005615234375,
0.474365234375,
0.48779296875,
0.473876953125,
0.48779296875,
0.488037109375,
0.12005615234375,
0.4892578125
],
[
0.51123046875,
0.47802734375,
0.88330078125,
0.478515625,
0.88330078125,
0.496337890625,
0.51123046875,
0.495361328125
],
[
0.12176513671875,
0.4912109375,
0.48779296875,
0.4912109375,
0.48779296875,
0.5048828125,
0.12176513671875,
0.5048828125
],
[
0.5146484375,
0.495361328125,
0.57421875,
0.495361328125,
0.57421875,
0.51025390625,
0.5146484375,
0.51025390625
],
[
0.12005615234375,
0.50830078125,
0.486083984375,
0.50830078125,
0.486083984375,
0.521484375,
0.12005615234375,
0.521484375
],
[
0.51123046875,
0.51708984375,
0.8818359375,
0.5185546875,
0.8818359375,
0.533203125,
0.51123046875,
0.5322265625
],
[
0.12005615234375,
0.52392578125,
0.48779296875,
0.52392578125,
0.48779296875,
0.53759765625,
0.12005615234375,
0.53759765625
],
[
0.118408203125,
0.5390625,
0.489501953125,
0.5380859375,
0.489501953125,
0.552734375,
0.118408203125,
0.55419921875
],
[
0.51123046875,
0.53369140625,
0.8818359375,
0.53369140625,
0.8818359375,
0.54736328125,
0.51123046875,
0.54736328125
],
[
0.12005615234375,
0.5556640625,
0.468505859375,
0.5556640625,
0.468505859375,
0.5693359375,
0.12005615234375,
0.5693359375
],
[
0.5146484375,
0.55029296875,
0.8818359375,
0.55029296875,
0.8818359375,
0.56396484375,
0.5146484375,
0.56396484375
],
[
0.12005615234375,
0.57177734375,
0.489501953125,
0.57177734375,
0.489501953125,
0.5859375,
0.12005615234375,
0.5859375
],
[
0.5146484375,
0.568359375,
0.880859375,
0.568359375,
0.880859375,
0.5791015625,
0.5146484375,
0.5791015625
],
[
0.12176513671875,
0.58740234375,
0.489501953125,
0.58740234375,
0.489501953125,
0.6015625,
0.12176513671875,
0.6015625
],
[
0.51318359375,
0.58447265625,
0.88330078125,
0.58447265625,
0.88330078125,
0.59716796875,
0.51318359375,
0.59716796875
],
[
0.118408203125,
0.6025390625,
0.32080078125,
0.6025390625,
0.32080078125,
0.6162109375,
0.118408203125,
0.6162109375
],
[
0.5146484375,
0.60107421875,
0.6689453125,
0.60107421875,
0.6689453125,
0.6142578125,
0.5146484375,
0.6142578125
],
[
0.1376953125,
0.619140625,
0.48779296875,
0.62060546875,
0.48779296875,
0.63525390625,
0.1376953125,
0.63427734375
],
[
0.118408203125,
0.6357421875,
0.486083984375,
0.6357421875,
0.486083984375,
0.6494140625,
0.118408203125,
0.6494140625
],
[
0.64208984375,
0.6318359375,
0.755859375,
0.6318359375,
0.755859375,
0.646484375,
0.64208984375,
0.646484375
],
[
0.12005615234375,
0.65185546875,
0.486083984375,
0.65185546875,
0.486083984375,
0.6650390625,
0.12005615234375,
0.6650390625
],
[
0.1376953125,
0.67041015625,
0.486083984375,
0.67041015625,
0.486083984375,
0.68115234375,
0.1376953125,
0.68115234375
],
[
0.51318359375,
0.6650390625,
0.880859375,
0.6650390625,
0.880859375,
0.6787109375,
0.51318359375,
0.6787109375
],
[
0.12005615234375,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69873046875,
0.12005615234375,
0.69873046875
],
[
0.51318359375,
0.68115234375,
0.8818359375,
0.68115234375,
0.8818359375,
0.6953125,
0.51318359375,
0.6953125
],
[
0.12176513671875,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.12176513671875,
0.71533203125
],
[
0.51123046875,
0.69775390625,
0.56494140625,
0.69775390625,
0.56494140625,
0.7119140625,
0.51123046875,
0.7119140625
],
[
0.12005615234375,
0.71630859375,
0.48779296875,
0.71728515625,
0.48779296875,
0.73095703125,
0.12005615234375,
0.72998046875
],
[
0.5341796875,
0.72021484375,
0.8818359375,
0.72021484375,
0.8818359375,
0.73388671875,
0.5341796875,
0.73388671875
],
[
0.1175537109375,
0.73291015625,
0.48779296875,
0.732421875,
0.48779296875,
0.74658203125,
0.1175537109375,
0.74755859375
],
[
0.51708984375,
0.736328125,
0.6708984375,
0.736328125,
0.6708984375,
0.75,
0.51708984375,
0.75
],
[
0.12005615234375,
0.7490234375,
0.486083984375,
0.75048828125,
0.486083984375,
0.765625,
0.12005615234375,
0.76318359375
],
[
0.51318359375,
0.75830078125,
0.8818359375,
0.7587890625,
0.8818359375,
0.77392578125,
0.51318359375,
0.7724609375
],
[
0.118408203125,
0.765625,
0.35107421875,
0.76416015625,
0.35107421875,
0.779296875,
0.118408203125,
0.77978515625
],
[
0.51318359375,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.51318359375,
0.7880859375
],
[
0.12005615234375,
0.783203125,
0.48779296875,
0.783203125,
0.48779296875,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.51318359375,
0.79052734375,
0.880859375,
0.79052734375,
0.880859375,
0.80419921875,
0.51318359375,
0.80419921875
],
[
0.12005615234375,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.12005615234375,
0.8125
],
[
0.509765625,
0.8056640625,
0.88330078125,
0.80810546875,
0.88330078125,
0.822265625,
0.509765625,
0.82080078125
],
[
0.12176513671875,
0.81640625,
0.484375,
0.81640625,
0.484375,
0.82666015625,
0.12176513671875,
0.82666015625
],
[
0.51318359375,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.12005615234375,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.51318359375,
0.84033203125,
0.79248046875,
0.84033203125,
0.79248046875,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.12005615234375,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.12005615234375,
0.86083984375
],
[
0.12176513671875,
0.865234375,
0.486083984375,
0.865234375,
0.486083984375,
0.8759765625,
0.12176513671875,
0.8759765625
],
[
0.5341796875,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.5341796875,
0.876953125
],
[
0.12005615234375,
0.87890625,
0.491943359375,
0.87890625,
0.491943359375,
0.8935546875,
0.12005615234375,
0.8935546875
],
[
0.515625,
0.87890625,
0.88330078125,
0.87890625,
0.88330078125,
0.8935546875,
0.515625,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51318359375,
0.89453125,
0.79248046875,
0.8955078125,
0.79248046875,
0.91015625,
0.51318359375,
0.9091796875
]
] |
[
"Methods",
"MiddleNoentityNeighborCollective体模型可以从表征所需全局相干性的其他训练",
"relation",
"relation inference数据中受益",
"param.",
"(Riedel et al., 2013)",
"V",
"总而言之，我们的贡献有三个方面：（1）我",
"(Verga et al., 2017)",
"1",
"V",
"(Zhang et al., 2019)",
"们提出了CoRI，这是一个两阶段框架，通过",
"V",
"CoRI (ours)",
"",
"进行具有全局一致性的集体预测来改进最先",
"",
"进的方法。（2）我们建议使用unmatched目标",
"Table 1: Comparisons between CoRI and baselines.",
"KG来增加训练数据。(3）在两个数据集上的",
"实验结果表明了我们方法的优越性，从。677",
"至。748和。716至。780，分别",
"部性，不同实体对的i.e.，预测是彼此独立进行",
"的，因此模型不知道“Nell”在最终预测中有",
"2",
"",
"Preliminaries",
"两个父亲。这样的预测在常识上是不连贯的",
"个人更有可能有一个父亲和一个母亲，这由",
"在本节中，我们首先制定关系集成的任务，然",
"并行数据的目标KG部分中Malia周围的图",
"后通过使用最先进的方法OpenKIE(Zhang",
"形结构表示",
"etal，2019）举例说明本地方法",
"1.1",
"Our CollectiveApproach",
"2.1Relation Integration",
"为了缓解局部方法的不相干预测问题，我们提",
"我们将来自开放IE系统的主题-关系-对象",
"出了协作关系集成（CoRI），它利用相邻实体",
"提取视为源图 K(8,R) =[(s,r,o)1s, E",
"对之间的预测依赖性来增强全局一致性",
"8,rER)，其中&表示提取的文本实体，e.g.",
"具体来说，我们遵循两个阶段，i.e.，候选生",
"“BarackObama”，R表示提取的源关系，e.g.",
"成和集体推理。在候选生成中，我们只需使用",
"“parent”。我们用（s,o）表示源实体对。对于",
"局部模型作为候选进行独立预测，Fig.1下部",
"(s,o)，Ks.0=(r|(s,r,o) EK) 表示它们之间",
"的所有三对e.g.，father。在集体推理中，我",
"的所有源关系。类似地，K=(s,o(s,r,o)E",
"们采用一个集体模型，该模型知道目标图的常",
"K}表示之间具有关系r的所有实体对。我们",
"见子结构e.g.，Malia。集体模型不仅通过将",
"使用联合KR=UreRK,来引用所有提取的",
"所有上下文特征作为本地模型的输入，而且还",
"实体对",
"通过将当前和所有邻居对的候选预测作为输入",
"Definition 1 (Relation Integration ). Given",
"来进行预测。对于Nell-Marie对，集体模型将",
"可以访问Nell-Burton的候选预测father，这",
"a source graph K and a target KG K'(&', R')",
"有助于将其最终预测翻转为正确的mother",
"with target entities &' and target relations R'",
"Tab.1从四个方面总结了CoRI和代表性的先",
"the task of relation integration is to predict al",
"前工作。据我们所知，CoRI是第一个集体执",
"applicable target relations for each extracted",
"行关系集成而不是本地的",
"entity pair in KR :",
"负责进行全局一致的预测，需要对集体模型",
"进行训练，以对目标KG的通用结构进行编",
"T CKR × R',",
"码：e.g.，Malia在Fig.1的并行数据中只有",
"个父亲/母亲。为此，我们以堆叠的方式训",
"where (s, r',o) e F is an integrated extraction",
"练集体模型（Wolpert，1992）。我们首先在并",
"indicating that a target relation r' holds for",
"行数据上训练第一阶段局部模型，然后通过对",
"(s,0) .",
"第一阶段的邻居实体对的候选预测进行条件化",
"为了训练关系集成模型，所有方法都采用如",
"（e.g.，father代表Malia-Barrack）来训练第二",
"下形式的并行数据：",
"阶段集体模型，以做出全局一致的预测（e.g.",
"Definition 2(Parallel Data).Parallel data are",
"mother代表Malia-Michelle)",
"common entity pairs shared between Kand",
"ParallelDataAugmentation.并行数据可能受",
"Cp, and their ground truth target relations in",
"到精确名称匹配的低召回率或开放IE系统生",
"K': T = [((s, 0),Kso) I(s,0) EKR nKR) .",
"成的有限提取的限制。我们观察到，即使没有",
"For example, ((Malia, Barack), [father)) is an",
"对应物提取，目标图的unmatched部分（如",
"instance of parallel data in Fig. 1",
"在Fig.1中）也可能具有丰富的通用结构来",
"指导集体模型的训练。为此，我们建议通过从",
"为了获得并行数据，一种广泛使用的方法是",
"unmatchedKG中采样子图来增强并行数据，",
"通过精确名称匹配找到8和8'共享的实体，",
"并通过合成其提取来创建伪并行数据，因此集",
"然后生成公共实体对及其基本事实"
] |
{
"arxiv_id": "2106.00793",
"language": "zh",
"page_idx": 1
}
|
||
2106.00793-zh-2
|
2106.00793-zh-2.jpg
|
2.2 Local Approaches
以前的本地方法通过假设它们的独立性来对潜
在的集成提取进行评分：
$$P(\Gamma\mid\mathcal{K})=\prod_{(s,r',o)\in\mathcal{K}_{\mathcal{R}}\times\mathcal{R}'}P_{\theta}(s,r',o\mid\mathcal{K}),$$
其中$\theta$是本地模型的参数。实现最先进性能的一个代表性本地模型是 OpenKI (Zhang et al., 2019) 。它通过将源关系的嵌入分组和平均为三个部分来编码$\mathcal{K}$中$(s,o)$的邻域。令$\mathcal{K}_s,.$是$s$ 与$o$以外的邻居实体之间的源关系集，对于$\mathcal{K}_{.,o}$也是如此。OpenKI 通过将三个平均嵌人连接到局部表示$t_l$中来表示$(s,o):$
$$t_l=[A(\mathcal{K}_{s,o});A(\mathcal{K}_{s,\cdot});A(\mathcal{K}_{\cdot,o})],$$
(2)
,其中$l$ 代表局部，而 $A(.)$ 接受一组关系并输出其嵌入的平均值。然后通过以下方式对每个集成提取进行评分：
$$P_\theta(s,r',o\mid\mathcal{K})=\sigma(MLP_l(t_l))_{r'},$$
(3)
其中$MLP_l$是多层感知器，而$\sigma$是 S 形函数。给定一个并行数据$\mathcal{T} = \left \{ \langle ( s, o) , \mathcal{K} _{s, o}^{\prime }\rangle \right \}$ , 每个训练示例的损失函数在最大化正目标关系的概率和最小化负目标关系的概率之间进行交易：
$$\begin{aligned}&L\big((s,o),\mathcal{K}_{s,o}^{\prime}\big)=-\frac{1}{|\mathcal{K}_{s,o}^{\prime}|}\sum_{r^{\prime}\in\mathcal{K}_{s,o}^{\prime}}\log P_{\theta}(s,r^{\prime},o\mid\mathcal{K}\\&+\frac{\gamma}{|\mathcal{R}^{\prime}\setminus\mathcal{K}_{s,o}^{\prime}|}\sum_{r^{\prime}\in\mathcal{R}^{\prime}\setminus\mathcal{K}_{s,o}^{\prime}}\log P_{\theta}(s,r^{\prime},o\mid\mathcal{K}),\\&&(4)\end{aligned}$$
其中$\gamma$是一个超参数，用于说明正负关系之间的矛盾，因为后者通常超过前者。最后的损失是所有例子的总和
3 Collective Relation Integration
如 § 1 中所述，局部方法的缺点是对不同实体对的预测是独立进行的。忽视他们的依赖性可能会导致预测彼此不一致
为了解决这个问题，我们提出了一种集体方法 CoRI, 它通过两个阶段实现集体关系整合： 候选生成和集体推理。在本节中，我们将演示两个阶段的输入和输出，以及我们当前的实现
## 3.1 Candidate Generation
如 § 1.1 中所述，候选生成的责任是将候选预测提供给集体推理阶段。形式上，候选预测$\Gamma^l$
<FigureHere>
Figure 2: Input of both stages on the Nell-Marie case. Solid edges are features for Nell-Marie. Additional edges in the lower part are predicted candidate target relations $\Gamma^l.$
$(l$表示局部)是通过在源图$\kappa$上执行局部模
型来生成的：
(5)
$$\Gamma^l=\mathop{\mathrm{argmax}}_{\Gamma}P_{\theta}(\Gamma\mid\mathcal{K}).$$
$\Gamma^l$中的候选预测可能部分错误，但是其他正确的预测可以在集体模型的指导下，在集体推理阶段帮助调整其相邻实体对的错误预测
例如，在 Fig. 2 的上部，我们有一个具有三个实体对的源图$\mathcal{K}$ 。候选生成的输入是整个$\mathcal{K}$。在应用本地模型(在我们的例子中为OpenKI) 之后，我们在 Fig. 2 的下部有三个额外的边作为输出$\Gamma^l$ 。请注意，Nell-Marie 的候选预测 father (用黑色轮廓表示)是不正确的，因为$\mathcal{K}$, i.e.,中的邻域信息不足，实体对之间和周围的关系(用实线表示)都是模糊的“父”
幸运的是，实体对 Nell-Burton 对于本地模型而言相对容易预测为 father ,因为它可以利用 Billy-Burton 之间的邻居关系“父亲”。此类正确的候选预测包含在$\Gamma^l$中，作为附加信号提供给集体推理阶段，以便以后纠正错误的预测，例如 Nell-Marie 的 father
### 3.2 Collective Inference
集体推理的责任是对目标图的结构进行编码并使用这些信息通过加强它们之间的一致性来优化候选预测$\Gamma^l$ 。为此，集体模型$P_\mathrm{\beta}$ (具有参数$\beta$ )同时将源图$\mathcal{K}$ 和 候 选 预 测 $\Gamma ^l$作为输入，并输出最终预测$\Gamma:$
$$P(\Gamma\mid\mathcal{K})=P_{\beta}(\Gamma\mid\mathcal{K},\Gamma^{l}).$$
(6)
在 Fig. 2 的 Nell-Marie 情况下，在进行最终预测时，使用其自身的候选预测和相邻实体对的预测(Fig.2中下部的$\Gamma^l$中的实心边)来利用它们之间的依赖性。我们将候选预测的嵌人连接到第一阶段获得的局部表示$t_l$ ,并将
|
[
[
0.12005615234375,
0.07659912109375,
0.312255859375,
0.07659912109375,
0.312255859375,
0.0914306640625,
0.12005615234375,
0.0914306640625
],
[
0.53564453125,
0.08197021484375,
0.54736328125,
0.08197021484375,
0.54736328125,
0.08612060546875,
0.53564453125,
0.08612060546875
],
[
0.60205078125,
0.07781982421875,
0.640625,
0.08013916015625,
0.640625,
0.0885009765625,
0.60205078125,
0.08612060546875
],
[
0.118408203125,
0.09857177734375,
0.484375,
0.09857177734375,
0.484375,
0.11224365234375,
0.118408203125,
0.11224365234375
],
[
0.53564453125,
0.09979248046875,
0.591796875,
0.09979248046875,
0.591796875,
0.10986328125,
0.53564453125,
0.10986328125
],
[
0.58447265625,
0.09442138671875,
0.68115234375,
0.0914306640625,
0.6826171875,
0.10687255859375,
0.5859375,
0.10986328125
],
[
0.12005615234375,
0.114013671875,
0.312255859375,
0.114013671875,
0.312255859375,
0.127685546875,
0.12005615234375,
0.127685546875
],
[
0.580078125,
0.1181640625,
0.60205078125,
0.1181640625,
0.60205078125,
0.1312255859375,
0.580078125,
0.1312255859375
],
[
0.654296875,
0.1181640625,
0.68115234375,
0.1181640625,
0.68115234375,
0.1312255859375,
0.654296875,
0.1312255859375
],
[
0.724609375,
0.12054443359375,
0.8671875,
0.1217041015625,
0.8671875,
0.133056640625,
0.724609375,
0.1318359375
],
[
0.599609375,
0.13720703125,
0.646484375,
0.14013671875,
0.6455078125,
0.1519775390625,
0.59765625,
0.1484375
],
[
0.7060546875,
0.1383056640625,
0.869140625,
0.1383056640625,
0.869140625,
0.1519775390625,
0.7060546875,
0.1519775390625
],
[
0.5341796875,
0.141357421875,
0.568359375,
0.141357421875,
0.568359375,
0.149658203125,
0.5341796875,
0.149658203125
],
[
0.133544921875,
0.1436767578125,
0.48779296875,
0.1436767578125,
0.48779296875,
0.160888671875,
0.133544921875,
0.160888671875
],
[
0.58251953125,
0.149658203125,
0.67529296875,
0.15673828125,
0.67236328125,
0.17041015625,
0.580078125,
0.163330078125
],
[
0.6650390625,
0.1519775390625,
0.68994140625,
0.1519775390625,
0.68994140625,
0.1561279296875,
0.6650390625,
0.1561279296875
],
[
0.7021484375,
0.152587890625,
0.79833984375,
0.155029296875,
0.79833984375,
0.169189453125,
0.7021484375,
0.16748046875
],
[
0.2275390625,
0.1644287109375,
0.330078125,
0.1644287109375,
0.330078125,
0.17578125,
0.2275390625,
0.17578125
],
[
0.55322265625,
0.1591796875,
0.65576171875,
0.1650390625,
0.654296875,
0.1787109375,
0.5517578125,
0.1728515625
],
[
0.5322265625,
0.174560546875,
0.56689453125,
0.174560546875,
0.56689453125,
0.1817626953125,
0.5322265625,
0.1817626953125
],
[
0.75,
0.171630859375,
0.85400390625,
0.17041015625,
0.85400390625,
0.1817626953125,
0.75,
0.182861328125
],
[
0.5810546875,
0.182861328125,
0.60205078125,
0.182861328125,
0.60205078125,
0.191162109375,
0.5810546875,
0.191162109375
],
[
0.6533203125,
0.1817626953125,
0.68115234375,
0.1817626953125,
0.68115234375,
0.193603515625,
0.6533203125,
0.193603515625
],
[
0.12005615234375,
0.189453125,
0.486083984375,
0.189453125,
0.486083984375,
0.203125,
0.12005615234375,
0.203125
],
[
0.77490234375,
0.185302734375,
0.869140625,
0.18408203125,
0.869140625,
0.19482421875,
0.77490234375,
0.1953125
],
[
0.1318359375,
0.207275390625,
0.486083984375,
0.207275390625,
0.486083984375,
0.2203369140625,
0.1318359375,
0.2203369140625
],
[
0.5146484375,
0.213134765625,
0.8818359375,
0.213134765625,
0.8818359375,
0.226806640625,
0.5146484375,
0.226806640625
],
[
0.12005615234375,
0.223876953125,
0.48779296875,
0.223876953125,
0.48779296875,
0.2369384765625,
0.12005615234375,
0.2369384765625
],
[
0.5146484375,
0.226806640625,
0.884765625,
0.226806640625,
0.884765625,
0.240478515625,
0.5146484375,
0.240478515625
],
[
0.12176513671875,
0.2392578125,
0.48779296875,
0.2392578125,
0.48779296875,
0.2529296875,
0.12176513671875,
0.2529296875
],
[
0.5146484375,
0.240478515625,
0.880859375,
0.2410888671875,
0.880859375,
0.254638671875,
0.5146484375,
0.254150390625
],
[
0.12005615234375,
0.255859375,
0.489501953125,
0.255859375,
0.489501953125,
0.26953125,
0.12005615234375,
0.26953125
],
[
0.51318359375,
0.254638671875,
0.705078125,
0.254638671875,
0.705078125,
0.268310546875,
0.51318359375,
0.268310546875
],
[
0.118408203125,
0.270751953125,
0.486083984375,
0.270751953125,
0.486083984375,
0.283935546875,
0.118408203125,
0.283935546875
],
[
0.12005615234375,
0.2861328125,
0.41064453125,
0.287353515625,
0.41064453125,
0.3017578125,
0.12005615234375,
0.300537109375
],
[
0.5146484375,
0.29638671875,
0.8818359375,
0.29638671875,
0.8818359375,
0.31005859375,
0.5146484375,
0.31005859375
],
[
0.181396484375,
0.318359375,
0.468505859375,
0.318359375,
0.468505859375,
0.33203125,
0.181396484375,
0.33203125
],
[
0.4609375,
0.319580078125,
0.48779296875,
0.319580078125,
0.48779296875,
0.330810546875,
0.4609375,
0.330810546875
],
[
0.515625,
0.314208984375,
0.615234375,
0.314208984375,
0.615234375,
0.325927734375,
0.515625,
0.325927734375
],
[
0.60205078125,
0.339111328125,
0.79248046875,
0.341552734375,
0.79248046875,
0.359375,
0.60205078125,
0.35693359375
],
[
0.1318359375,
0.347412109375,
0.486083984375,
0.3486328125,
0.486083984375,
0.363525390625,
0.1318359375,
0.3623046875
],
[
0.85791015625,
0.34326171875,
0.8818359375,
0.34326171875,
0.8818359375,
0.358154296875,
0.85791015625,
0.358154296875
],
[
0.67529296875,
0.355224609375,
0.68408203125,
0.361083984375,
0.6767578125,
0.366455078125,
0.66748046875,
0.35986328125
],
[
0.12005615234375,
0.364013671875,
0.48779296875,
0.364013671875,
0.48779296875,
0.377685546875,
0.12005615234375,
0.377685546875
],
[
0.12005615234375,
0.38134765625,
0.276123046875,
0.38134765625,
0.276123046875,
0.394287109375,
0.12005615234375,
0.394287109375
],
[
0.51318359375,
0.380126953125,
0.8818359375,
0.383056640625,
0.8818359375,
0.39794921875,
0.51318359375,
0.394287109375
],
[
0.515625,
0.3984375,
0.88330078125,
0.3984375,
0.88330078125,
0.412109375,
0.515625,
0.412109375
],
[
0.172119140625,
0.410400390625,
0.489501953125,
0.410888671875,
0.489501953125,
0.42578125,
0.172119140625,
0.424560546875
],
[
0.5146484375,
0.41455078125,
0.8271484375,
0.41455078125,
0.8271484375,
0.427490234375,
0.5146484375,
0.427490234375
],
[
0.5341796875,
0.43115234375,
0.8818359375,
0.43115234375,
0.8818359375,
0.44482421875,
0.5341796875,
0.44482421875
],
[
0.12176513671875,
0.441162109375,
0.48779296875,
0.441162109375,
0.48779296875,
0.4560546875,
0.12176513671875,
0.4560546875
],
[
0.515625,
0.446533203125,
0.8818359375,
0.446533203125,
0.8818359375,
0.460205078125,
0.515625,
0.460205078125
],
[
0.12176513671875,
0.457763671875,
0.491943359375,
0.457763671875,
0.491943359375,
0.471435546875,
0.12176513671875,
0.471435546875
],
[
0.5146484375,
0.46142578125,
0.8818359375,
0.4619140625,
0.8818359375,
0.476806640625,
0.5146484375,
0.4755859375
],
[
0.118408203125,
0.473876953125,
0.486083984375,
0.473876953125,
0.486083984375,
0.48681640625,
0.118408203125,
0.48681640625
],
[
0.515625,
0.478515625,
0.8818359375,
0.478515625,
0.8818359375,
0.4921875,
0.515625,
0.4921875
],
[
0.12005615234375,
0.4912109375,
0.484375,
0.4912109375,
0.484375,
0.501953125,
0.12005615234375,
0.501953125
],
[
0.515625,
0.49462890625,
0.8818359375,
0.49462890625,
0.8818359375,
0.50830078125,
0.515625,
0.50830078125
],
[
0.12005615234375,
0.50732421875,
0.14697265625,
0.50732421875,
0.14697265625,
0.52099609375,
0.12005615234375,
0.52099609375
],
[
0.5146484375,
0.51123046875,
0.8818359375,
0.51025390625,
0.8818359375,
0.52490234375,
0.5146484375,
0.525390625
],
[
0.276123046875,
0.52685546875,
0.890625,
0.5390625,
0.88916015625,
0.57177734375,
0.274658203125,
0.55859375
],
[
0.515625,
0.52783203125,
0.8818359375,
0.52783203125,
0.8818359375,
0.54150390625,
0.515625,
0.54150390625
],
[
0.12176513671875,
0.54150390625,
0.252685546875,
0.54150390625,
0.252685546875,
0.5556640625,
0.12176513671875,
0.5556640625
],
[
0.268798828125,
0.54443359375,
0.3720703125,
0.55810546875,
0.366943359375,
0.57666015625,
0.264404296875,
0.5634765625
],
[
0.2568359375,
0.552734375,
0.267822265625,
0.552734375,
0.267822265625,
0.55810546875,
0.2568359375,
0.55810546875
],
[
0.5146484375,
0.55859375,
0.57666015625,
0.55859375,
0.57666015625,
0.57373046875,
0.5146484375,
0.57373046875
],
[
0.177978515625,
0.57373046875,
0.189697265625,
0.57373046875,
0.189697265625,
0.5859375,
0.177978515625,
0.5859375
],
[
0.2476806640625,
0.57666015625,
0.450927734375,
0.57666015625,
0.450927734375,
0.59716796875,
0.2476806640625,
0.59716796875
],
[
0.5341796875,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59033203125,
0.5341796875,
0.59033203125
],
[
0.124267578125,
0.58447265625,
0.231689453125,
0.5888671875,
0.2301025390625,
0.60791015625,
0.1234130859375,
0.6025390625
],
[
0.219970703125,
0.595703125,
0.301513671875,
0.599609375,
0.300537109375,
0.6142578125,
0.21826171875,
0.60986328125
],
[
0.5146484375,
0.59130859375,
0.8818359375,
0.5927734375,
0.8818359375,
0.60693359375,
0.5146484375,
0.60546875
],
[
0.515625,
0.609375,
0.8818359375,
0.609375,
0.8818359375,
0.62255859375,
0.515625,
0.62255859375
],
[
0.46337890625,
0.6162109375,
0.489501953125,
0.6162109375,
0.489501953125,
0.62744140625,
0.46337890625,
0.62744140625
],
[
0.5146484375,
0.62353515625,
0.880859375,
0.62353515625,
0.880859375,
0.63720703125,
0.5146484375,
0.63720703125
],
[
0.12176513671875,
0.64453125,
0.486083984375,
0.64453125,
0.486083984375,
0.65771484375,
0.12176513671875,
0.65771484375
],
[
0.51708984375,
0.642578125,
0.87890625,
0.642578125,
0.87890625,
0.65283203125,
0.51708984375,
0.65283203125
],
[
0.12005615234375,
0.662109375,
0.48779296875,
0.662109375,
0.48779296875,
0.67578125,
0.12005615234375,
0.67578125
],
[
0.5146484375,
0.65673828125,
0.78271484375,
0.65673828125,
0.78271484375,
0.67041015625,
0.5146484375,
0.67041015625
],
[
0.12005615234375,
0.6787109375,
0.267822265625,
0.6787109375,
0.267822265625,
0.6923828125,
0.12005615234375,
0.6923828125
],
[
0.5146484375,
0.68408203125,
0.72119140625,
0.68408203125,
0.72119140625,
0.69775390625,
0.5146484375,
0.69775390625
],
[
0.15283203125,
0.70703125,
0.420654296875,
0.70703125,
0.420654296875,
0.72021484375,
0.15283203125,
0.72021484375
],
[
0.515625,
0.7060546875,
0.8818359375,
0.7060546875,
0.8818359375,
0.7197265625,
0.515625,
0.7197265625
],
[
0.515625,
0.7216796875,
0.8818359375,
0.7216796875,
0.8818359375,
0.7353515625,
0.515625,
0.7353515625
],
[
0.118408203125,
0.7080078125,
0.156982421875,
0.7080078125,
0.156982421875,
0.7197265625,
0.118408203125,
0.7197265625
],
[
0.12005615234375,
0.73291015625,
0.489501953125,
0.73291015625,
0.489501953125,
0.74658203125,
0.12005615234375,
0.74658203125
],
[
0.515625,
0.73681640625,
0.8818359375,
0.73681640625,
0.8818359375,
0.751953125,
0.515625,
0.751953125
],
[
0.118408203125,
0.7490234375,
0.484375,
0.7490234375,
0.484375,
0.7626953125,
0.118408203125,
0.7626953125
],
[
0.5146484375,
0.7529296875,
0.880859375,
0.7529296875,
0.880859375,
0.7666015625,
0.5146484375,
0.7666015625
],
[
0.118408203125,
0.76416015625,
0.32080078125,
0.76416015625,
0.32080078125,
0.77783203125,
0.118408203125,
0.77783203125
],
[
0.51708984375,
0.77099609375,
0.71728515625,
0.77099609375,
0.71728515625,
0.78369140625,
0.51708984375,
0.78369140625
],
[
0.139404296875,
0.7822265625,
0.486083984375,
0.7822265625,
0.486083984375,
0.7958984375,
0.139404296875,
0.7958984375
],
[
0.118408203125,
0.79638671875,
0.48779296875,
0.79736328125,
0.48779296875,
0.8125,
0.118408203125,
0.81103515625
],
[
0.59619140625,
0.7998046875,
0.80029296875,
0.798828125,
0.80029296875,
0.8134765625,
0.59619140625,
0.81396484375
],
[
0.85400390625,
0.8017578125,
0.8818359375,
0.8017578125,
0.8818359375,
0.81396484375,
0.85400390625,
0.81396484375
],
[
0.12005615234375,
0.81396484375,
0.486083984375,
0.81396484375,
0.486083984375,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.12005615234375,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.53076171875,
0.830078125,
0.8818359375,
0.83056640625,
0.8818359375,
0.845703125,
0.53076171875,
0.84423828125
],
[
0.5146484375,
0.84716796875,
0.880859375,
0.84716796875,
0.880859375,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.12005615234375,
0.85791015625,
0.345947265625,
0.85791015625,
0.345947265625,
0.87158203125,
0.12005615234375,
0.87158203125
],
[
0.515625,
0.8623046875,
0.8818359375,
0.8623046875,
0.8818359375,
0.876953125,
0.515625,
0.876953125
],
[
0.12005615234375,
0.87744140625,
0.48779296875,
0.87890625,
0.48779296875,
0.8935546875,
0.12005615234375,
0.892578125
],
[
0.515625,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.515625,
0.8935546875
],
[
0.12005615234375,
0.8955078125,
0.48779296875,
0.8935546875,
0.48779296875,
0.90771484375,
0.12005615234375,
0.91015625
],
[
0.515625,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"2.2Local Approaches",
"VC",
"Burto",
"以前的本地方法通过假设它们的独立性来对潜",
"parent",
"parent father",
"在的集成提取进行评分：",
"Ne",
"Billy",
"inputtothefirststage",
"Burton",
"input to the second stage",
"Mari",
"P(I)= Pe(s,r',o), (1)",
"parent father",
"Le",
"OOentity",
"(s,r\",o)EKRxR",
"parentlther",
"cathe",
"sourcerelation",
"Lell",
"Billy",
"其中θ是本地模型的参数。实现最先进性能的",
"rget relation",
"个代表性本地模型是OpenKI（Zhangetal.",
"Figure 2: Input of both stages on the Nell-Marie",
"2019）。它通过将源关系的嵌入分组和平均为",
"case. Solid edges are features for Nell-Marie. Ad-",
"三个部分来编码K中（s,o）的邻域。令Ks：是",
"ditional edges in the lower part are predicted can",
"s与o以外的邻居实体之间的源关系集，对于",
"didate target relations P!",
"K.。也是如此。OpenKI通过将三个平均嵌人",
"连接到局部表示ti中来表示（s,o）：",
"（l表示局部）是通过在源图K上执行局部模",
"ti = [A(Ks,); A(Ks.); A(K.,)],(",
"(2)",
"型来生成的：",
"I' = argmax Pe(T I K).",
"其中1代表局部，而A(.）接受一组关系并输",
"(5)",
"",
"出其嵌人的平均值。然后通过以下方式对每个",
"集成提取进行评分：",
"Fl中的候选预测可能部分错误，但是其他正确",
"的预测可以在集体模型的指导下，在集体推理",
"Pe(s, r,o / K) = o(MLPi(ti)r\",(3)",
"阶段帮助调整其相邻实体对的错误预测",
"例如，在Fig.2的上部，我们有一个具有",
"其中MLPI是多层感知器，而是S形函",
"三个实体对的源图K。候选生成的输入是整",
"数。给定一个并行数据T=《(s,o),K)，",
"个K。在应用本地模型（在我们的例子中为",
"每个训练示例的损失函数在最大化正目标关系",
"OpenKI）之后，我们在Fig.2的下部有三个",
"的概率和最小化负目标关系的概率之间进行交",
"额外的边作为输出T。请注意，Nell-Marie的",
"易:",
"候选预测father（用黑色轮廓表示）是不正",
"Zlog P(s,r, |K）对之间和周围的关系（用实线表示）都是模糊",
"确的，因为K，i.e.，中的邻域信息不足，实体",
"L((s,0),Ks.o) =",
"IK's,ol reK's,o",
"",
"的“父”",
"",
"log Pe(s, r',o | K),",
"幸运的是，实体对Nell-Burton对于本地模",
"+R\\Xs,ol",
"rER\\Ks,o",
"型而言相对容易预测为father，因为它可以利",
"用Billy-Burton之间的邻居关系“父亲”。此",
"(4)",
"类正确的候选预测包含在F中，作为附加信",
"其中是一个超参数，用于说明正负关系之间",
"号提供给集体推理阶段：以便以后纠正错误的",
"的矛盾，因为后者通常超过前者。最后的损失",
"预测，例如Nell-Marie的father",
"是所有例子的总和",
"3.2Collective Inference",
"Collective Relation Integration",
"集体推理的责任是对目标图的结构进行编码，",
"并使用这些信息通过加强它们之间的一致性来",
"3",
"如S1中所述，局部方法的缺点是对不同实体",
"优化候选预测Tl。为此，集体模型P（具有",
"对的预测是独立进行的。忽视他们的依赖性可",
"参数β）同时将源图K和候选预测作为输",
"能会导致预测彼此不一致",
"入，并输出最终预测I：",
"为了解决这个问题，我们提出了一种集体方",
"法CoRI，它通过两个阶段实现集体关系整合：",
"P(T [K) = PB(T IK,T).",
"(6)",
"候选生成和集体推理。在本节中，我们将演示",
"两个阶段的输入和输出，以及我们当前的实现",
"在Fig.2的Nell-Marie情况下，在进行最",
"终预测时，使用其自身的候选预测和相邻实体",
"3.1Candidate Generation",
"对的预测（Fig.2中下部的F'中的实心边）来",
"如S1.1中所述，候选生成的责任是将候选预",
"利用它们之间的依赖性。我们将候选预测的嵌",
"测提供给集体推理阶段。形式上，候选预测T",
"人连接到第一阶段获得的局部表示t，并将"
] |
{
"arxiv_id": "2106.00793",
"language": "zh",
"page_idx": 2
}
|
||
2106.00793-zh-4
|
2106.00793-zh-4.jpg
|
<table>
<tbody>
<tr>
<td>Al $pr$</td>
<td>orithm 2: Our augmentation ap- ach. </td>
</tr>
<tr>
<td>Al Pr $F$</td>
<td>rithm$ 2\nobreak {: } $ Our augmentation ap- ach. sult: Collective model $\beta$ with data</td>
</tr>
<tr>
<td>$fc$</td>
<td>entity pair$\in\mathcal{K}_{\mathcal R}\cap\mathcal{K}_{\mathcal{R}^{\prime}}^{\prime}$ $do$ $S\leftarrow$its top $K$ similar entity pairs in $\underline {{\mathcal{K} }}_{\mathcal{R} }^{p}\cap \underline {{\mathcal{K} }}_{\mathcal{R} ^{\prime }}^{\prime }$ ; $\overline {{\mathcal{T} }^{p}}\leftarrow \overline {{\mathcal{T} }^{p}}\cup \{ \langle ( s, o) , {\mathcal{K} }_{s. o}^{\prime }\rangle \mid ( s, o) \in S\}$ ;</td>
</tr>
<tr>
<td>e</td>
<td>$\operatorname { \text{Train on}} \mathcal{T} \cup \overline {\mathcal{T} ^{p}}\operatorname { \mathrm{with~Alg. ~1; }}$</td>
</tr>
</tbody>
</table>
$P_\beta$专注于源图的域。为了缓解此问题，我们仅使用类似于$\mathcal{T}$的$\mathcal{T}^p$的子集，如 Fig.3 中的黑色轮廓部分所示。具体来说，我们将每个实体对$(s,o)$表示为一个虚拟文档，周围的目标关系$\mathcal{K}_{s,o}^{\prime}\cup\mathcal{K}_{s,\cdot}^{\prime}\cup\mathcal{K}_{\cdot,o}^{\prime}$作为“令牌”。对于来自并行数据$\mathcal{T}$的每个实体对，我们使用 BM25E (Robertson and Zaragoza, 2009) 从$\mathcal{T}^p$中检索其顶部的$K$最相似的实体对，并将其添加到选定的伪并行数据$\overline{\mathcal{T}^p}$中进行训练，如 Alg.2中所述
5 Experimental Settings
5.1 Datasets and Evaluation
我们使用 ReVerb 数据集 (Fader et al., 2011) 作为源图，并分别使用 Freebase $^{1}$和 Wikidata $^2$作为目标 KG。我们遵循 Zhang et al. (2019) 中的同名匹配方法来获取并行数据。为了模拟在有限的标记数据上训练模型但将其应用于大型测试集的真实场景，我们在并行数据中使用 20 个实体对的% 进行训练，并使用其他 80个% 进行测试，并且没有重叠。我们还比较了$\S6.3$中其他比率下的性能。Tab.2 中列出了数据集统计信息
<table>
<tbody>
<tr>
<th>Datasets</th>
<th>$\#$Train</th>
<th>$\#$ Test</th>
<th>$|\mathcal{R}|$</th>
</tr>
<tr>
<td>ReVerb+Freebase</td>
<td>12,344</td>
<td>49,629</td>
<td>97,196</td>
</tr>
<tr>
<td>ReVerb+Wikidata</td>
<td>8,447</td>
<td>33,849</td>
<td>182,407</td>
</tr>
</tbody>
</table>
Table 2: Dataset statistics. We follow Zhang et al. (2019) to use the most frequent 250 target relations.
$^{1}$https://developers.google.com/freebase
$^{2}$https://www.wikidata.org
我们通过根据其概率对所有集成提取进行排名来进行评估，并报告曲线下面积(AUC)。考虑到我们希望在保持高精度的同时集成尽可能多的提取的真实场景，我们还报告了当精度为0时的 Recall 和 F 1 。8,0。9 或 0。95
## 5.2 Compared Methods
我们在实验中比较了以下方法
Relation Translation 是一种简单的方法，它将源关系映射到具有条件概率$P(r^{\prime}\mid r)$的目标关系，类似于 Eq.9。对于实体对$(s,o)$,预测的目标关系为$\{\arg\max_{r^{\prime}}P(r^{\prime}|r)\mid r\in\mathcal{K}_{s,o}\}$ Universal Schema (E-model) ${\mathrm{E}}($Riedel et al., 2013) 通过矩阵分解来学习实体和关系嵌入， 无法推广到看不见的实体。它是一个本地模型，可对每个集成提取进行独立评分
Rowless Universal Schema $\boxed{\mathrm{F}}($Verga et al., 2017)是一个本地模型，它通过消除特定于实体的参数来改进 E 模型，从而推广到看不见的实体
OpenKI (Zhang et al., 2019) 是一个本地模型，它通过将邻居关系用于更多上下文来解决源关系的歧义
CoRI 是我们用 Alg.1 训练的集体两阶段关系
集成模型
CoRI+DA 是我们的模型，其中使用 Alg. 2通过伪并行数据增强了训练数据。为了验证基于检索的伪数据选择的必要性，我们还与随机DA 基线进行了比较，其中我们选择了$K$ 随机实体对
CoRI + KGE 是另一种利用 unmatched KG 的方法，它以无监督的方式在整个目标 KG 上训练了 KG 嵌入(KGE)。我们用在目标图上训练的 TransE (Bordes et al.,2013) 嵌入初始化 Eq.7 中由$A(.)$平均的目标关系的嵌入
### 5.3 Implementation Details
我们对所有关系统一使用 32 维嵌人 ,并使用学习率为 0 的 AdamWE(Loshchilov and Hutter. 2019) 优化器。01 和 Epsilon$10^{-8}$。Eq.4 中的比率$\gamma$设置为 10。我们最多对 30 个邻居源关系进行采样，以处理具有太多邻居关系的实体对。我们在 Alg.1 中使用$T=5$折叠来训练我们的集体模型。我们在伪数据选择中检索顶级$K=5$实体对，分别向 Tab. 2 中的两个数据集添加约 20K 和 12K 实体对。在伪数据选择中，我们使用 ElasticSearch $^{3}$中的 BM25E (Robertson and Zaragoza, 2009) 实现。我们使用 OpenKE 发布的 KGE。$^{4}$我们的模型是用
$^3$https://www.elastic.co/
$^{4}$https://github.com/thunlp/OpenKE
|
[
[
0.2493896484375,
0.0789794921875,
0.43994140625,
0.0789794921875,
0.43994140625,
0.0926513671875,
0.2493896484375,
0.0926513671875
],
[
0.5322265625,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.5322265625,
0.09027099609375
],
[
0.515625,
0.0914306640625,
0.880859375,
0.0914306640625,
0.880859375,
0.1051025390625,
0.515625,
0.1051025390625
],
[
0.1292724609375,
0.08074951171875,
0.2376708984375,
0.08074951171875,
0.2376708984375,
0.0926513671875,
0.1292724609375,
0.0926513671875
],
[
0.124267578125,
0.0968017578125,
0.187255859375,
0.09442138671875,
0.18896484375,
0.1092529296875,
0.1259765625,
0.11102294921875
],
[
0.1368408203125,
0.1134033203125,
0.39208984375,
0.1134033203125,
0.39208984375,
0.12353515625,
0.1368408203125,
0.12353515625
],
[
0.515625,
0.1092529296875,
0.8818359375,
0.1092529296875,
0.8818359375,
0.122314453125,
0.515625,
0.122314453125
],
[
0.193115234375,
0.1270751953125,
0.287109375,
0.1270751953125,
0.287109375,
0.1341552734375,
0.193115234375,
0.1341552734375
],
[
0.515625,
0.12469482421875,
0.8818359375,
0.12469482421875,
0.8818359375,
0.1383056640625,
0.515625,
0.1383056640625
],
[
0.1368408203125,
0.1353759765625,
0.408935546875,
0.1353759765625,
0.408935546875,
0.1484375,
0.1368408203125,
0.1484375
],
[
0.51123046875,
0.1395263671875,
0.8330078125,
0.14013671875,
0.8330078125,
0.155029296875,
0.51123046875,
0.15380859375
],
[
0.1368408203125,
0.1478271484375,
0.267822265625,
0.1484375,
0.267822265625,
0.163330078125,
0.1368408203125,
0.162109375
],
[
0.1368408203125,
0.160888671875,
0.450927734375,
0.160888671875,
0.450927734375,
0.174560546875,
0.1368408203125,
0.174560546875
],
[
0.5146484375,
0.169189453125,
0.5615234375,
0.169189453125,
0.5615234375,
0.18115234375,
0.5146484375,
0.18115234375
],
[
0.55517578125,
0.169189453125,
0.724609375,
0.169189453125,
0.724609375,
0.182861328125,
0.55517578125,
0.182861328125
],
[
0.162841796875,
0.18408203125,
0.301513671875,
0.18408203125,
0.301513671875,
0.19775390625,
0.162841796875,
0.19775390625
],
[
0.1485595703125,
0.174560546875,
0.16796875,
0.174560546875,
0.16796875,
0.18408203125,
0.1485595703125,
0.18408203125
],
[
0.166259765625,
0.196533203125,
0.307373046875,
0.196533203125,
0.307373046875,
0.210205078125,
0.166259765625,
0.210205078125
],
[
0.515625,
0.191162109375,
0.75390625,
0.191162109375,
0.75390625,
0.204833984375,
0.515625,
0.204833984375
],
[
0.162841796875,
0.207275390625,
0.322509765625,
0.2078857421875,
0.322509765625,
0.22265625,
0.162841796875,
0.221435546875
],
[
0.51318359375,
0.2078857421875,
0.880859375,
0.2078857421875,
0.880859375,
0.221435546875,
0.51318359375,
0.221435546875
],
[
0.1376953125,
0.221435546875,
0.16455078125,
0.221435546875,
0.16455078125,
0.2327880859375,
0.1376953125,
0.2327880859375
],
[
0.5146484375,
0.223876953125,
0.88330078125,
0.22265625,
0.88330078125,
0.2369384765625,
0.5146484375,
0.2381591796875
],
[
0.1376953125,
0.2369384765625,
0.316650390625,
0.2381591796875,
0.316650390625,
0.251708984375,
0.1376953125,
0.25048828125
],
[
0.51708984375,
0.240478515625,
0.88330078125,
0.240478515625,
0.88330078125,
0.254150390625,
0.51708984375,
0.254150390625
],
[
0.1351318359375,
0.248779296875,
0.341796875,
0.248779296875,
0.341796875,
0.262451171875,
0.1351318359375,
0.262451171875
],
[
0.16455078125,
0.260009765625,
0.418212890625,
0.260009765625,
0.418212890625,
0.273681640625,
0.16455078125,
0.273681640625
],
[
0.515625,
0.255859375,
0.880859375,
0.255859375,
0.880859375,
0.270751953125,
0.515625,
0.270751953125
],
[
0.172119140625,
0.26953125,
0.25439453125,
0.271484375,
0.25439453125,
0.287353515625,
0.172119140625,
0.28515625
],
[
0.5146484375,
0.271484375,
0.8818359375,
0.2724609375,
0.8818359375,
0.287353515625,
0.5146484375,
0.2861328125
],
[
0.16455078125,
0.283935546875,
0.432373046875,
0.2861328125,
0.432373046875,
0.300537109375,
0.16455078125,
0.298583984375
],
[
0.51318359375,
0.287353515625,
0.88330078125,
0.2880859375,
0.88330078125,
0.302734375,
0.51318359375,
0.3017578125
],
[
0.1376953125,
0.2998046875,
0.16455078125,
0.2998046875,
0.16455078125,
0.311279296875,
0.1376953125,
0.311279296875
],
[
0.51708984375,
0.3046875,
0.88330078125,
0.3046875,
0.88330078125,
0.318359375,
0.51708984375,
0.318359375
],
[
0.1368408203125,
0.312255859375,
0.37451171875,
0.312255859375,
0.37451171875,
0.325927734375,
0.1368408203125,
0.325927734375
],
[
0.515625,
0.320556640625,
0.8076171875,
0.320556640625,
0.8076171875,
0.334228515625,
0.515625,
0.334228515625
],
[
0.51318359375,
0.336181640625,
0.8818359375,
0.33740234375,
0.8818359375,
0.3515625,
0.51318359375,
0.350830078125
],
[
0.118408203125,
0.359375,
0.486083984375,
0.359375,
0.486083984375,
0.372314453125,
0.118408203125,
0.372314453125
],
[
0.5146484375,
0.35400390625,
0.8818359375,
0.35400390625,
0.8818359375,
0.36767578125,
0.5146484375,
0.36767578125
],
[
0.51318359375,
0.3681640625,
0.884765625,
0.369384765625,
0.884765625,
0.38427734375,
0.51318359375,
0.383056640625
],
[
0.12176513671875,
0.377197265625,
0.48779296875,
0.377197265625,
0.48779296875,
0.39013671875,
0.12176513671875,
0.39013671875
],
[
0.51318359375,
0.38427734375,
0.59033203125,
0.385986328125,
0.59033203125,
0.40087890625,
0.51318359375,
0.3984375
],
[
0.12005615234375,
0.392578125,
0.48779296875,
0.392578125,
0.48779296875,
0.40625,
0.12005615234375,
0.40625
],
[
0.5146484375,
0.402587890625,
0.79833984375,
0.402587890625,
0.79833984375,
0.416259765625,
0.5146484375,
0.416259765625
],
[
0.7900390625,
0.40087890625,
0.88330078125,
0.402587890625,
0.88330078125,
0.416259765625,
0.7900390625,
0.41455078125
],
[
0.12176513671875,
0.4091796875,
0.489501953125,
0.4091796875,
0.489501953125,
0.4228515625,
0.12176513671875,
0.4228515625
],
[
0.51318359375,
0.418701171875,
0.8818359375,
0.418701171875,
0.8818359375,
0.431640625,
0.51318359375,
0.431640625
],
[
0.12176513671875,
0.424560546875,
0.486083984375,
0.424560546875,
0.486083984375,
0.438232421875,
0.12176513671875,
0.438232421875
],
[
0.12005615234375,
0.440673828125,
0.489501953125,
0.440673828125,
0.489501953125,
0.454833984375,
0.12005615234375,
0.454833984375
],
[
0.515625,
0.435791015625,
0.62451171875,
0.435791015625,
0.62451171875,
0.44775390625,
0.515625,
0.44775390625
],
[
0.51123046875,
0.449462890625,
0.88330078125,
0.45068359375,
0.88330078125,
0.465576171875,
0.51123046875,
0.46435546875
],
[
0.12005615234375,
0.457275390625,
0.48779296875,
0.457275390625,
0.48779296875,
0.47021484375,
0.12005615234375,
0.47021484375
],
[
0.118408203125,
0.473876953125,
0.486083984375,
0.473876953125,
0.486083984375,
0.48681640625,
0.118408203125,
0.48681640625
],
[
0.515625,
0.468505859375,
0.587890625,
0.468505859375,
0.587890625,
0.479736328125,
0.515625,
0.479736328125
],
[
0.51318359375,
0.482666015625,
0.884765625,
0.48388671875,
0.884765625,
0.498779296875,
0.51318359375,
0.49755859375
],
[
0.12005615234375,
0.4892578125,
0.490234375,
0.4892578125,
0.490234375,
0.5029296875,
0.12005615234375,
0.5029296875
],
[
0.12005615234375,
0.5048828125,
0.177978515625,
0.5048828125,
0.177978515625,
0.51953125,
0.12005615234375,
0.51953125
],
[
0.515625,
0.50048828125,
0.8818359375,
0.50048828125,
0.8818359375,
0.51416015625,
0.515625,
0.51416015625
],
[
0.515625,
0.51708984375,
0.8818359375,
0.51708984375,
0.8818359375,
0.53076171875,
0.515625,
0.53076171875
],
[
0.151123046875,
0.5361328125,
0.349365234375,
0.53759765625,
0.349365234375,
0.5517578125,
0.151123046875,
0.55029296875
],
[
0.51318359375,
0.53076171875,
0.88330078125,
0.5322265625,
0.88330078125,
0.54638671875,
0.51318359375,
0.5458984375
],
[
0.118408203125,
0.53759765625,
0.156982421875,
0.53759765625,
0.156982421875,
0.548828125,
0.118408203125,
0.548828125
],
[
0.5146484375,
0.548828125,
0.587890625,
0.548828125,
0.587890625,
0.56005859375,
0.5146484375,
0.56005859375
],
[
0.16455078125,
0.56396484375,
0.366943359375,
0.56396484375,
0.366943359375,
0.57763671875,
0.16455078125,
0.57763671875
],
[
0.5146484375,
0.5654296875,
0.88330078125,
0.5654296875,
0.88330078125,
0.5791015625,
0.5146484375,
0.5791015625
],
[
0.515625,
0.58203125,
0.88330078125,
0.58203125,
0.88330078125,
0.595703125,
0.515625,
0.595703125
],
[
0.12005615234375,
0.56640625,
0.1695556640625,
0.56640625,
0.1695556640625,
0.57763671875,
0.12005615234375,
0.57763671875
],
[
0.118408203125,
0.58837890625,
0.48779296875,
0.58740234375,
0.48779296875,
0.6015625,
0.118408203125,
0.6025390625
],
[
0.5146484375,
0.5966796875,
0.880859375,
0.5966796875,
0.880859375,
0.60986328125,
0.5146484375,
0.60986328125
],
[
0.118408203125,
0.60400390625,
0.489501953125,
0.6025390625,
0.489501953125,
0.61767578125,
0.118408203125,
0.6181640625
],
[
0.1175537109375,
0.6181640625,
0.486083984375,
0.619140625,
0.486083984375,
0.63427734375,
0.1175537109375,
0.6328125
],
[
0.515625,
0.61328125,
0.8818359375,
0.61328125,
0.8818359375,
0.62744140625,
0.515625,
0.62744140625
],
[
0.118408203125,
0.63525390625,
0.486083984375,
0.63525390625,
0.486083984375,
0.6484375,
0.118408203125,
0.6484375
],
[
0.5146484375,
0.6298828125,
0.873046875,
0.6298828125,
0.873046875,
0.64453125,
0.5146484375,
0.64453125
],
[
0.118408203125,
0.65185546875,
0.484375,
0.65185546875,
0.484375,
0.6650390625,
0.118408203125,
0.6650390625
],
[
0.5146484375,
0.6591796875,
0.56494140625,
0.6591796875,
0.56494140625,
0.67041015625,
0.5146484375,
0.67041015625
],
[
0.5576171875,
0.65771484375,
0.75390625,
0.65771484375,
0.75390625,
0.67138671875,
0.5576171875,
0.67138671875
],
[
0.12176513671875,
0.66845703125,
0.48779296875,
0.66845703125,
0.48779296875,
0.681640625,
0.12176513671875,
0.681640625
],
[
0.12176513671875,
0.68505859375,
0.489501953125,
0.68505859375,
0.489501953125,
0.69873046875,
0.12176513671875,
0.69873046875
],
[
0.515625,
0.6796875,
0.88330078125,
0.6796875,
0.88330078125,
0.693359375,
0.515625,
0.693359375
],
[
0.5146484375,
0.69482421875,
0.880859375,
0.6953125,
0.880859375,
0.71044921875,
0.5146484375,
0.708984375
],
[
0.12176513671875,
0.70068359375,
0.486083984375,
0.70068359375,
0.486083984375,
0.71435546875,
0.12176513671875,
0.71435546875
],
[
0.5146484375,
0.71044921875,
0.880859375,
0.71044921875,
0.880859375,
0.724609375,
0.5146484375,
0.724609375
],
[
0.12005615234375,
0.71728515625,
0.484375,
0.71728515625,
0.484375,
0.73095703125,
0.12005615234375,
0.73095703125
],
[
0.118408203125,
0.732421875,
0.2476806640625,
0.73291015625,
0.2476806640625,
0.74755859375,
0.118408203125,
0.74658203125
],
[
0.51708984375,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.51708984375,
0.7412109375
],
[
0.51708984375,
0.74462890625,
0.87890625,
0.74462890625,
0.87890625,
0.7548828125,
0.51708984375,
0.7548828125
],
[
0.1292724609375,
0.76416015625,
0.197265625,
0.7666015625,
0.197265625,
0.78076171875,
0.1292724609375,
0.779296875
],
[
0.515625,
0.76025390625,
0.88330078125,
0.76025390625,
0.88330078125,
0.77490234375,
0.515625,
0.77490234375
],
[
0.295654296875,
0.7666015625,
0.355224609375,
0.7666015625,
0.355224609375,
0.78076171875,
0.295654296875,
0.78076171875
],
[
0.3779296875,
0.7666015625,
0.4306640625,
0.7666015625,
0.4306640625,
0.78076171875,
0.3779296875,
0.78076171875
],
[
0.47021484375,
0.76708984375,
0.496337890625,
0.76708984375,
0.496337890625,
0.779296875,
0.47021484375,
0.779296875
],
[
0.51318359375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.133544921875,
0.78369140625,
0.264404296875,
0.78369140625,
0.264404296875,
0.79736328125,
0.133544921875,
0.79736328125
],
[
0.301513671875,
0.78515625,
0.345947265625,
0.78515625,
0.345947265625,
0.79638671875,
0.301513671875,
0.79638671875
],
[
0.378662109375,
0.78369140625,
0.42822265625,
0.78369140625,
0.42822265625,
0.798828125,
0.378662109375,
0.798828125
],
[
0.45751953125,
0.78515625,
0.50732421875,
0.78515625,
0.50732421875,
0.79638671875,
0.45751953125,
0.79638671875
],
[
0.378662109375,
0.79443359375,
0.429931640625,
0.79638671875,
0.42822265625,
0.81103515625,
0.3779296875,
0.80859375
],
[
0.45166015625,
0.7919921875,
0.884765625,
0.7890625,
0.884765625,
0.80712890625,
0.45166015625,
0.81005859375
],
[
0.30322265625,
0.79736328125,
0.344970703125,
0.79736328125,
0.344970703125,
0.80859375,
0.30322265625,
0.80859375
],
[
0.133544921875,
0.79638671875,
0.268798828125,
0.79736328125,
0.268798828125,
0.80859375,
0.133544921875,
0.80810546875
],
[
0.515625,
0.80859375,
0.8818359375,
0.80859375,
0.8818359375,
0.822265625,
0.515625,
0.822265625
],
[
0.118408203125,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51318359375,
0.822265625,
0.884765625,
0.82177734375,
0.884765625,
0.8388671875,
0.51318359375,
0.84033203125
],
[
0.118408203125,
0.8388671875,
0.490234375,
0.8388671875,
0.490234375,
0.8525390625,
0.118408203125,
0.8525390625
],
[
0.5146484375,
0.84130859375,
0.88330078125,
0.84130859375,
0.88330078125,
0.85498046875,
0.5146484375,
0.85498046875
],
[
0.12005615234375,
0.8525390625,
0.160400390625,
0.85498046875,
0.15869140625,
0.86767578125,
0.118408203125,
0.865234375
],
[
0.515625,
0.85693359375,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.515625,
0.87060546875
],
[
0.139404296875,
0.88134765625,
0.4130859375,
0.8818359375,
0.4130859375,
0.896484375,
0.139404296875,
0.8955078125
],
[
0.5341796875,
0.8818359375,
0.705078125,
0.8818359375,
0.705078125,
0.896484375,
0.5341796875,
0.896484375
],
[
0.1376953125,
0.89453125,
0.328369140625,
0.8955078125,
0.328369140625,
0.91015625,
0.1376953125,
0.9091796875
],
[
0.53564453125,
0.8955078125,
0.79443359375,
0.8955078125,
0.79443359375,
0.91015625,
0.53564453125,
0.91015625
]
] |
[
"Our augmentation ap",
"我们通过根据其概率对所有集成提取进行排",
"名来进行评估，并报告曲线下面积（AUC）。考",
"Algorithm 2:",
"proach.",
"Result: Collective model β with data",
"虑到我们希望在保持高精度的同时集成尽可能",
"augmentation",
"多的提取的真实场景，我们还报告了当精度为",
"(1) Synthesizing Pseudo Extractions XP",
"0时的Recall和F1。8,0。9或0。95",
"KP0:TP←0:",
"for (s',r',o') EK\", where (s',o')eKR, /KR",
"5.2",
" Compared Methods",
"$←$ando←o;",
"do",
"Sample r ~ P(rlr') ;",
"我们在实验中比较了以下方法",
"PKPu((s,r,)) ;",
"RelationTranslation是一种简单的方法，它将",
"end",
"源关系映射到具有条件概率P(r|r）的目标",
"(2) Pseudo Data Selection",
"关系，类似于Eq.9。对于实体对（s,o），预",
"for entity pair e KR nKR, do",
"S its top K similar entity pairs in",
"测的目标关系为argmaxrP(r'|r）「rEKs,o}",
"KRnKR;",
"Universal Schema (E-model) E(Riedel et al.,",
"Tp- TPU(((s, 0),Ks,o) [ (s,0) e S) ;",
"2013）通过矩阵分解来学习实体和关系嵌人，",
"end",
"无法推广到看不见的实体。它是一个本地模",
"β + Train on TUTp with Alg. 1;",
"型，可对每个集成提取进行独立评分",
"Rowless Universal Schema E(Verga et al..",
"Ps专注于源图的域。为了缓解此问题，我们仅",
"2017）是一个本地模型，它通过消除特定于",
"实体的参数来改进E模型，从而推广到看不",
"使用类似于T的TP的子集，如Fig.3中的",
"见的实体",
"黑色轮廓部分所示。具体来说，我们将每个实",
"OpenKI(Zhang et al.,2019)是-",
"一个本地模",
"体对（s,o）表示为一个虚拟文档，周围的目标",
"型，它通过将邻居关系用于更多上下文来解决",
"关系。UKUK。作为“令牌”。对于来自",
"并行数据T的每个实体对，我们使用BM25E",
"源关系的歧义",
"CoRI是我们用Alg.1训练的集体两阶段关系",
"(RobertsonandZaragoza,2009)从Tp中检索",
"其顶部的K最相似的实体对，并将其添加到",
"集成模型",
"CoRI+DA是我们的模型，其中使用Alg.2",
"选定的伪并行数据Tp中进行训练，如Alg.2",
"中所述",
"通过伪并行数据增强了训练数据。为了验证基",
"于检索的伪数据选择的必要性，我们还与随机",
"Experimental Settings",
"DA基线进行了比较，其中我们选择了K随",
"5",
"机实体对",
"Datasets and Evaluation",
"CoRI+KGE是另一种利用unmatchedKG",
"的方法，它以无监督的方式在整个目标KG上",
"5.1",
"我们使用ReVerb数据集（Faderetal.，2011)",
"训练了KG嵌入（KGE）。我们用在目标图上",
"作为源图，并分别使用Freebase1和Wikidata",
"2作为目标KG。我们遵循Zhangetal.（2019)",
"训练的TransE（Bordesetal.，2013）嵌人初",
"中的同名匹配方法来获取并行数据。为了模拟",
"始化Eq.7中由A(）平均的目标关系的嵌人",
"在有限的标记数据上训练模型但将其应用于",
"5.3",
" Implementation Details",
"大型测试集的真实场景，我们在并行数据中使",
"用20个实体对的%进行训练，并使用其他80",
"我们对所有关系统一使用32维嵌入，并使用学",
"习率为0的AdamWE(LoshchilovandHutter",
"个%进行测试，并且没有重叠。我们还比较了",
"2019）优化器。01和Epsilon10-8。Eq.4中的",
"S6.3中其他比率下的性能。Tab.2中列出了",
"数据集统计信息",
"比率设置为10。我们最多对30个邻居源关",
"系进行采样，以处理具有太多邻居关系的实体",
"Datasets",
"对。我们在Alg.1中使用T=5折叠来训练",
"# Train",
"# Test",
"R]",
"我们的集体模型。我们在伪数据选择中检索顶",
"ReVerb + Freebase",
"12.344",
"49,629",
"97,196",
"33,849",
"182,407级K=5实体对，分别向Tab.2中的两个数",
"8,447",
"ReVerb + Wikidata",
"据集添加约20K和12K实体对。在伪数据选",
"Table 2: Dataset statistics. We follow Zhang et al.",
"择中，我们使用ElasticSearch3中的BM25",
"(2019) to use the most frequent 250 target rela-",
"(RobertsonandZaragoza，2009）实现。我们使",
"tions.",
"用OpenKE发布的KGE。4我们的模型是用",
"Ihttps://developers.google.com/freebase",
"3https://www.elastic.co/",
"2https://www.wikidata.org",
"*https://github.com/thunlp/OpenKE"
] |
{
"arxiv_id": "2106.00793",
"language": "zh",
"page_idx": 4
}
|
||
2207.06544-zh-0
|
2207.06544-zh-0.jpg
|
# Volatility Based Kernels and Moving Average Means for Accurate Forecasting with Gaussian Processes
。3in
$\operatorname{Abstract}$
A broad class of stochastic volatility models are defined by systems of stochastic differential equations. While these models have seen widespread success in domains such as fnance and statistical climatology, they typically lack an ability to condition on historical data to produce a true posterior distribution. To address this fundamental limitation, we show how to re-cast a class of stochastic volatility models as a hierarchical Gaussian process (GP) model with specialized covariance functions. This GP model retains the inductive biases of the stochastic volatility model while providing the posterior predictive distribution given by GP inference. Within this framework, we take inspiration from well studied domains to introduce a new class of models, Volt and Magpie , that significantly outperform baselines in stock and wind speed forecasting, and naturally extend to the multitask setting.
## 1. Introduction
高斯过程 (GP) 在时间序列建模方面取得了巨大成功使其成为建模和预测时间相关的金融和气候数据的挑战性任务的有力候选者。但是，构建 GP 历来依赖于选择开箱即用的内核和均值函数，这些假设并不适用于所有情况；例如，Matérn 和 RBF 内核假定数据的平稳性，而常量或线性均值假定数据的基本趋势不是时变的。虽然对构建更灵活的 GP 的追求导致了对诸如频谱混合和深度内核之类的内核函数的大量研究，但这些内核主要是针对一般问题而开发的，而不是根据有关手头任务的先验知识来构建内核。在这项工作中， 我们通过基于领域知识来构建 GP 建模，以构建一组具有归纳偏差的新颖核函数和均值函数，非常适合于金融和气候学等领域的预测，其中数据随时间随机演
$^*$Equal contribution$^1$New York University. Correspon-
dence to: Greg Benton < gwb260@nyu.edu >.
Proceedings of the $39^{th}$ International Conference on Machine Learning, Baltimore, Maryland, USA, PMLR 162, 2022. Copyright 2022 by the author(s).
Gregory Benton $^*^{1}$ Wesley J. Maddox$^*^{1}$ Andrew Gordon Wilson$^{1}$
$\overline{9}$
<FigureHere>
Figure 1. An overview of stochastic volatility, Volt, and forecasting. Top: the observed data over and the corresponding volatility path. Middle: the learned volatility from the data, and volatility forecasts. Bottom: the data over 1 year and forecasts, with each sample path corresponding to a distinct sample from the volatility forecast.
### 变
金融和气候时间序列都是非平稳的，其特征是具有时变和随机波动率或变化程度。在随机波动率 (SV) 模型的激励下，我们开发了一个具有专门内核的分层 GP 模型，称为模型 Volt Bollerslev (1986); Wilson and Ghahramani (2010);Hagan et al.(2002)。Volt 使用波动率预测来指定未来数据观测的协方差结构。通过不仅考虑单个波动率预测，而且考虑波动率预测的分布，我们在数据域中推导出协方差函数的分布。考虑波动性的不确定性并将其传播到我们的数据预测中，得出的预测分布已根据数据进行了很好的校准，为理解风险水平和模拟潜在结果提供了关键工具。有关随机波动率的更多信息，请参见附录 A.1
图 1 提供了 Volt 描述的分层 GP 模型的图形表示。给定一组观察值(顶行),我们推断出这些收益的波动率路径(中间行),并形成一个分层 GP 模型，其中第一
|
[
[
0.22216796875,
0.11236572265625,
0.75244140625,
0.11297607421875,
0.75244140625,
0.1307373046875,
0.22216796875,
0.1300048828125
],
[
0.238525390625,
0.136962890625,
0.73779296875,
0.136962890625,
0.73779296875,
0.1514892578125,
0.238525390625,
0.1514892578125
],
[
0.233642578125,
0.2020263671875,
0.73779296875,
0.2020263671875,
0.73779296875,
0.216552734375,
0.233642578125,
0.216552734375
],
[
0.0882568359375,
0.2183837890625,
0.1298828125,
0.216552734375,
0.131591796875,
0.231689453125,
0.08905029296875,
0.2342529296875
],
[
0.24267578125,
0.2298583984375,
0.318603515625,
0.231689453125,
0.318603515625,
0.246826171875,
0.24267578125,
0.2442626953125
],
[
0.7900390625,
0.240478515625,
0.81640625,
0.240478515625,
0.81640625,
0.248779296875,
0.7900390625,
0.248779296875
],
[
0.1217041015625,
0.25244140625,
0.441162109375,
0.25244140625,
0.441162109375,
0.266357421875,
0.1217041015625,
0.266357421875
],
[
0.53515625,
0.25244140625,
0.55810546875,
0.25244140625,
0.55810546875,
0.264404296875,
0.53515625,
0.264404296875
],
[
0.78515625,
0.251953125,
0.85693359375,
0.24560546875,
0.859375,
0.260009765625,
0.78759765625,
0.266357421875
],
[
0.8544921875,
0.261474609375,
0.87109375,
0.261474609375,
0.87109375,
0.279052734375,
0.8544921875,
0.279052734375
],
[
0.1217041015625,
0.268310546875,
0.441162109375,
0.268310546875,
0.441162109375,
0.2822265625,
0.1217041015625,
0.2822265625
],
[
0.1217041015625,
0.28271484375,
0.441162109375,
0.28271484375,
0.441162109375,
0.296630859375,
0.1217041015625,
0.296630859375
],
[
0.82666015625,
0.27783203125,
0.85693359375,
0.27783203125,
0.85693359375,
0.290283203125,
0.82666015625,
0.290283203125
],
[
0.12091064453125,
0.29736328125,
0.439453125,
0.298583984375,
0.439453125,
0.3125,
0.12091064453125,
0.311279296875
],
[
0.5458984375,
0.29345703125,
0.55810546875,
0.29345703125,
0.55810546875,
0.3037109375,
0.5458984375,
0.3037109375
],
[
0.5458984375,
0.30615234375,
0.57177734375,
0.30615234375,
0.57177734375,
0.318115234375,
0.5458984375,
0.318115234375
],
[
0.6357421875,
0.306884765625,
0.6611328125,
0.306884765625,
0.6611328125,
0.318115234375,
0.6357421875,
0.318115234375
],
[
0.7255859375,
0.306884765625,
0.7509765625,
0.306884765625,
0.7509765625,
0.318115234375,
0.7255859375,
0.318115234375
],
[
0.81396484375,
0.30615234375,
0.8388671875,
0.30615234375,
0.8388671875,
0.3193359375,
0.81396484375,
0.3193359375
],
[
0.689453125,
0.323974609375,
0.736328125,
0.323974609375,
0.736328125,
0.333984375,
0.689453125,
0.333984375
],
[
0.12091064453125,
0.311279296875,
0.439453125,
0.313232421875,
0.439453125,
0.32763671875,
0.12091064453125,
0.32568359375
],
[
0.1217041015625,
0.32763671875,
0.441162109375,
0.32763671875,
0.441162109375,
0.341552734375,
0.1217041015625,
0.341552734375
],
[
0.53125,
0.335205078125,
0.55810546875,
0.339111328125,
0.5556640625,
0.3515625,
0.52783203125,
0.348388671875
],
[
0.68896484375,
0.3359375,
0.7548828125,
0.3359375,
0.7548828125,
0.3466796875,
0.68896484375,
0.3466796875
],
[
0.8037109375,
0.330078125,
0.82666015625,
0.330078125,
0.82666015625,
0.341552734375,
0.8037109375,
0.341552734375
],
[
0.1217041015625,
0.343505859375,
0.441162109375,
0.343505859375,
0.441162109375,
0.356689453125,
0.1217041015625,
0.356689453125
],
[
0.51318359375,
0.3515625,
0.529296875,
0.3515625,
0.529296875,
0.37060546875,
0.51318359375,
0.37060546875
],
[
0.68896484375,
0.347900390625,
0.76318359375,
0.347900390625,
0.76318359375,
0.359130859375,
0.68896484375,
0.359130859375
],
[
0.1192626953125,
0.356689453125,
0.441162109375,
0.35791015625,
0.441162109375,
0.37255859375,
0.1192626953125,
0.371826171875
],
[
0.52783203125,
0.365478515625,
0.5537109375,
0.365478515625,
0.5537109375,
0.376953125,
0.52783203125,
0.376953125
],
[
0.1192626953125,
0.37255859375,
0.441162109375,
0.373779296875,
0.441162109375,
0.386962890625,
0.1192626953125,
0.386474609375
],
[
0.1192626953125,
0.386962890625,
0.439453125,
0.386962890625,
0.439453125,
0.40087890625,
0.1192626953125,
0.40087890625
],
[
0.5458984375,
0.397705078125,
0.57177734375,
0.397705078125,
0.57177734375,
0.410400390625,
0.5458984375,
0.410400390625
],
[
0.6357421875,
0.397705078125,
0.6611328125,
0.397705078125,
0.6611328125,
0.410400390625,
0.6357421875,
0.410400390625
],
[
0.7255859375,
0.397705078125,
0.7509765625,
0.397705078125,
0.7509765625,
0.410400390625,
0.7255859375,
0.410400390625
],
[
0.814453125,
0.396484375,
0.8388671875,
0.396484375,
0.8388671875,
0.410400390625,
0.814453125,
0.410400390625
],
[
0.12091064453125,
0.404052734375,
0.441162109375,
0.404052734375,
0.441162109375,
0.417236328125,
0.12091064453125,
0.417236328125
],
[
0.1192626953125,
0.41845703125,
0.439453125,
0.41845703125,
0.439453125,
0.431884765625,
0.1192626953125,
0.431884765625
],
[
0.62109375,
0.4140625,
0.65185546875,
0.4140625,
0.65185546875,
0.425537109375,
0.62109375,
0.425537109375
],
[
0.53515625,
0.43115234375,
0.55615234375,
0.43115234375,
0.55615234375,
0.443115234375,
0.53515625,
0.443115234375
],
[
0.62353515625,
0.425537109375,
0.6708984375,
0.427978515625,
0.6708984375,
0.438232421875,
0.62353515625,
0.436279296875
],
[
0.1217041015625,
0.434326171875,
0.441162109375,
0.434326171875,
0.441162109375,
0.447509765625,
0.1217041015625,
0.447509765625
],
[
0.12091064453125,
0.449462890625,
0.441162109375,
0.449462890625,
0.441162109375,
0.46337890625,
0.12091064453125,
0.46337890625
],
[
0.54345703125,
0.45703125,
0.5556640625,
0.45703125,
0.5556640625,
0.468505859375,
0.54345703125,
0.468505859375
],
[
0.12091064453125,
0.464111328125,
0.439453125,
0.464111328125,
0.439453125,
0.477783203125,
0.12091064453125,
0.477783203125
],
[
0.1217041015625,
0.479248046875,
0.441162109375,
0.479248046875,
0.441162109375,
0.492431640625,
0.1217041015625,
0.492431640625
],
[
0.5458984375,
0.488037109375,
0.57177734375,
0.488037109375,
0.57177734375,
0.501953125,
0.5458984375,
0.501953125
],
[
0.6357421875,
0.4892578125,
0.66259765625,
0.4892578125,
0.66259765625,
0.501953125,
0.6357421875,
0.501953125
],
[
0.7255859375,
0.4892578125,
0.7490234375,
0.4892578125,
0.7490234375,
0.501953125,
0.7255859375,
0.501953125
],
[
0.814453125,
0.488037109375,
0.8388671875,
0.488037109375,
0.8388671875,
0.501953125,
0.814453125,
0.501953125
],
[
0.68701171875,
0.501953125,
0.701171875,
0.501953125,
0.701171875,
0.51220703125,
0.68701171875,
0.51220703125
],
[
0.1192626953125,
0.492431640625,
0.441162109375,
0.49365234375,
0.441162109375,
0.50830078125,
0.1192626953125,
0.5068359375
],
[
0.1192626953125,
0.50830078125,
0.441162109375,
0.5087890625,
0.441162109375,
0.5234375,
0.1192626953125,
0.52294921875
],
[
0.1217041015625,
0.5244140625,
0.441162109375,
0.5244140625,
0.441162109375,
0.53857421875,
0.1217041015625,
0.53857421875
],
[
0.5009765625,
0.53271484375,
0.88623046875,
0.53271484375,
0.88623046875,
0.546875,
0.5009765625,
0.546875
],
[
0.1192626953125,
0.5390625,
0.44287109375,
0.53857421875,
0.44287109375,
0.55322265625,
0.1192626953125,
0.5537109375
],
[
0.5,
0.546875,
0.88623046875,
0.546875,
0.88623046875,
0.56103515625,
0.5,
0.56103515625
],
[
0.1192626953125,
0.55322265625,
0.295654296875,
0.5546875,
0.295654296875,
0.5693359375,
0.1192626953125,
0.5673828125
],
[
0.5,
0.56005859375,
0.8857421875,
0.5595703125,
0.8857421875,
0.57373046875,
0.5,
0.57470703125
],
[
0.5,
0.57373046875,
0.8837890625,
0.57373046875,
0.8837890625,
0.5869140625,
0.5,
0.5869140625
],
[
0.5,
0.58837890625,
0.88623046875,
0.58837890625,
0.88623046875,
0.6015625,
0.5,
0.6015625
],
[
0.08905029296875,
0.59765625,
0.22216796875,
0.59765625,
0.22216796875,
0.6123046875,
0.08905029296875,
0.6123046875
],
[
0.5009765625,
0.60302734375,
0.87744140625,
0.60302734375,
0.87744140625,
0.6162109375,
0.5009765625,
0.6162109375
],
[
0.0906982421875,
0.62353515625,
0.4697265625,
0.62353515625,
0.4697265625,
0.63720703125,
0.0906982421875,
0.63720703125
],
[
0.08905029296875,
0.63720703125,
0.473876953125,
0.63818359375,
0.473876953125,
0.65283203125,
0.08905029296875,
0.6513671875
],
[
0.0906982421875,
0.65380859375,
0.473876953125,
0.65380859375,
0.473876953125,
0.66748046875,
0.0906982421875,
0.66748046875
],
[
0.5009765625,
0.65087890625,
0.51953125,
0.65087890625,
0.51953125,
0.6640625,
0.5009765625,
0.6640625
],
[
0.0906982421875,
0.66845703125,
0.47216796875,
0.66845703125,
0.47216796875,
0.681640625,
0.0906982421875,
0.681640625
],
[
0.50244140625,
0.67236328125,
0.8837890625,
0.67236328125,
0.8837890625,
0.68603515625,
0.50244140625,
0.68603515625
],
[
0.0906982421875,
0.6845703125,
0.47216796875,
0.6845703125,
0.47216796875,
0.69775390625,
0.0906982421875,
0.69775390625
],
[
0.50244140625,
0.68798828125,
0.8837890625,
0.68798828125,
0.8837890625,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.0906982421875,
0.69970703125,
0.473876953125,
0.69970703125,
0.473876953125,
0.71337890625,
0.0906982421875,
0.71337890625
],
[
0.50244140625,
0.70263671875,
0.8837890625,
0.70263671875,
0.8837890625,
0.71630859375,
0.50244140625,
0.71630859375
],
[
0.0906982421875,
0.7138671875,
0.473876953125,
0.7138671875,
0.473876953125,
0.72802734375,
0.0906982421875,
0.72802734375
],
[
0.5,
0.71630859375,
0.8857421875,
0.71728515625,
0.8857421875,
0.73193359375,
0.5,
0.73095703125
],
[
0.08905029296875,
0.72998046875,
0.473876953125,
0.72998046875,
0.473876953125,
0.74365234375,
0.08905029296875,
0.74365234375
],
[
0.5009765625,
0.73291015625,
0.8837890625,
0.73291015625,
0.8837890625,
0.7470703125,
0.5009765625,
0.7470703125
],
[
0.08905029296875,
0.744140625,
0.473876953125,
0.744140625,
0.473876953125,
0.75830078125,
0.08905029296875,
0.75830078125
],
[
0.5009765625,
0.74853515625,
0.8837890625,
0.74853515625,
0.8837890625,
0.76220703125,
0.5009765625,
0.76220703125
],
[
0.0906982421875,
0.7587890625,
0.47216796875,
0.7587890625,
0.47216796875,
0.77294921875,
0.0906982421875,
0.77294921875
],
[
0.5009765625,
0.76220703125,
0.8837890625,
0.76318359375,
0.8837890625,
0.77783203125,
0.5009765625,
0.7763671875
],
[
0.0906982421875,
0.7744140625,
0.473876953125,
0.7744140625,
0.473876953125,
0.78857421875,
0.0906982421875,
0.78857421875
],
[
0.5009765625,
0.77783203125,
0.88232421875,
0.77783203125,
0.88232421875,
0.79150390625,
0.5009765625,
0.79150390625
],
[
0.0906982421875,
0.79052734375,
0.47216796875,
0.79052734375,
0.47216796875,
0.8037109375,
0.0906982421875,
0.8037109375
],
[
0.50244140625,
0.79541015625,
0.88232421875,
0.79541015625,
0.88232421875,
0.80615234375,
0.50244140625,
0.80615234375
],
[
0.08905029296875,
0.8046875,
0.473876953125,
0.8046875,
0.473876953125,
0.818359375,
0.08905029296875,
0.818359375
],
[
0.50244140625,
0.8095703125,
0.8857421875,
0.8095703125,
0.8857421875,
0.82275390625,
0.50244140625,
0.82275390625
],
[
0.10455322265625,
0.8271484375,
0.473876953125,
0.82763671875,
0.473876953125,
0.84228515625,
0.10455322265625,
0.84130859375
],
[
0.50244140625,
0.82470703125,
0.8837890625,
0.82470703125,
0.8837890625,
0.83837890625,
0.50244140625,
0.83837890625
],
[
0.08905029296875,
0.84130859375,
0.396240234375,
0.84130859375,
0.396240234375,
0.85595703125,
0.08905029296875,
0.85595703125
],
[
0.50244140625,
0.8388671875,
0.7646484375,
0.8388671875,
0.7646484375,
0.85302734375,
0.50244140625,
0.85302734375
],
[
0.08905029296875,
0.86572265625,
0.473876953125,
0.86572265625,
0.473876953125,
0.8798828125,
0.08905029296875,
0.8798828125
],
[
0.50244140625,
0.861328125,
0.8857421875,
0.861328125,
0.8857421875,
0.875,
0.50244140625,
0.875
],
[
0.08905029296875,
0.8798828125,
0.473876953125,
0.8798828125,
0.473876953125,
0.89306640625,
0.08905029296875,
0.89306640625
],
[
0.5009765625,
0.876953125,
0.8837890625,
0.876953125,
0.8837890625,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.08905029296875,
0.892578125,
0.348876953125,
0.892578125,
0.348876953125,
0.90576171875,
0.08905029296875,
0.90576171875
],
[
0.5009765625,
0.892578125,
0.87744140625,
0.892578125,
0.87744140625,
0.90576171875,
0.5009765625,
0.90576171875
]
] |
[
"Volatility Based Kernels and Moving Average Means",
"for Accurate Forecasting with Gaussian Processes",
"Gregory Benton *1 Wesley J. Maddox * 1 Andrew Gordon Wilsonl",
"。3in",
"Abstract",
"Data",
"A broad class of stochastic volatility models",
"10-",
"Volatilit0.2",
"0",
"are defined by systems of stochastic differ-",
"ential equations. While these models have",
"-0.1",
"seen widespread success in domains such as",
"9",
"0.0",
"0.5",
"1.0",
"1.5",
"TrueVol.",
"finance and statistical climatology, they typ",
"ically lack an ability to condition on histori-",
"0.2",
"LearnedVol",
"M",
"cal data to produce a true posterior distribu-",
"10A",
"PredictedVol",
"tion. To address this fundamental limitation,",
"0.1",
"we show how to re-cast a class of stochas-",
"tic volatility models as a hierarchical Gaus",
"0.0",
"0.5",
"1.0",
"1.5",
"sian process (GP) model with specialized co-",
"variance functions.This GP model retains",
"Data",
"10",
"Predictio",
"the inductive biases of the stochastic volatil-",
"ity model while providing the posterior pre-",
"9",
"dictive distribution given by GP inference.",
"Within this framework, we take inspiration",
"0.0",
"0.5",
"1.0",
"1.5",
"X",
"from well studied domains to introduce a new",
"class of models, Volt and Magpie , that sig-",
"nificantly outperform baselines in stock and",
"Figure 1. An overview of stochastic volatility, Volt, and",
"wind speed forecasting, and naturally extend",
"forecasting.Top: the observed data over and the corre-",
"to the multitask setting.",
"sponding volatility path.Middle: the learned volatility",
"from the data, and volatility forecasts. Bottom: the data",
"over 1 year and forecasts, with each sample path corre-",
"1.Introduction",
"sponding to a distinct sample from the volatility forecast.",
"高斯过程（GP）在时间序列建模方面取得了巨大成功",
"使其成为建模和预测时间相关的金融和气候数据的挑",
"战性任务的有力候选者。但是，构建GP历来依赖于选",
"变",
"择开箱即用的内核和均值函数，这些假设并不适用于",
"金融和气候时间序列都是非平稳的，其特征是具有时",
"所有情况；例如，Matern和RBF内核假定数据的平",
"变和随机波动率或变化程度。在随机波动率（SV）模型",
"稳性，而常量或线性均值假定数据的基本趋势不是时",
"的激励下，我们开发了一个具有专门内核的分层GP",
"变的。虽然对构建更灵活的GP的追求导致了对诸如",
"模型，称为模型VoltBollerslev（1986)；Wilsonand",
"频谱混合和深度内核之类的内核函数的大量研究，但",
"Ghahramani(2010);Haganet al.(2002)。Volt使用",
"这些内核主要是针对一般问题而开发的，而不是根据",
"波动率预测来指定未来数据观测的协方差结构。通过",
"有关手头任务的先验知识来构建内核。在这项工作中，",
"不仅考虑单个波动率预测，而且考虑波动率预测的分",
"我们通过基于领域知识来构建GP建模，以构建一组",
"布，我们在数据域中推导出协方差函数的分布。考虑波",
"具有归纳偏差的新颖核函数和均值函数，非常适合于",
"动性的不确定性并将其传播到我们的数据预测中，得",
"金融和气候学等领域的预测，其中数据随时间随机演",
"出的预测分布已根据数据进行了很好的校准，为理解",
"*Equal contribution 1 New York University.Correspon-",
"风险水平和模拟潜在结果提供了关键工具。有关随机",
"dence to: Greg Benton < gwb260@nyu.edu >.",
"波动率的更多信息，请参见附录A.1",
"Proceedings of the 3gth International Conference on Ma-",
"图1提供了Volt描述的分层GP模型的图形表示。给",
"chine Learning, Baltimore, Maryland, USA, PMLR 162,",
"定一组观察值（顶行），我们推断出这些收益的波动率",
"2022. Copyright 2022 by the author(s).",
"路径（中间行），并形成一个分层GP模型，其中第"
] |
{
"arxiv_id": "2207.06544",
"language": "zh",
"page_idx": 0
}
|
||
2207.06544-zh-1
|
2207.06544-zh-1.jpg
|
Title Suppressed Due to Excessive Size
个 GP 对波动率进行建模，第二个 GP 用于预测给定
样本数据的分布来自波动率 GP(底行)
Volt 描述的协方差结构提供了预测中不确定性的忠实表示，但忽略了数据空间 GP 的均值函数，这是捕获数据趋势的强大工具。为此，我们共同介绍了移动平均高斯过程或 Magpie, 其中我们用移动平均值替换了GP 中的标准参数均值函数。移动平均线是气候学和金融学等领域广泛使用的技术 (Nau, 2014) 。通过将移动平均线的趋势拟合功能与 GP 的概率框架相结合， 我们可以生成既准确又具有校准不确定性的预测
虽然 Volt 和 Magpie 可以单独使用，但我们将它们作为一个单独的作品呈现，因为它们的组合特别解决了具有挑战性的预测问题。在股票价格或风速等时间演变领域，过程的固有随机性使我们无法在遥远的未来产生准确的点估计，我们只需要准确性，但不确定的是忠实于数据的随机性。出于这个原因，人们既需要喜鹊提供的准确趋势捕捉，也需要伏特提供的准确不确定性表示
我们的主要贡献如下：
高斯过程 copula 波动率模型 (GPCV) 对波动率进行参数化，而 Wu et al.(2014) 使用 GP 状态空间模型和粒子滤波器对波动率进行估计。与 Wilson and Ghahramani (2010) 类似，Lázaro-Gredilla and Titsias (2011)使用高斯过程对具有指数联系的波动率模型进行参数化，但使用高度结构化的变分近似进行推断。Liu et al. (2020)使用多任务高斯过程来预测未来的波动性，将其模型应用于外汇货币收益，再次对价格进行一步预测。至关重要的是，仅预测波动性确实可以直接预测数据，这是我们使用 Volt 的主要目标。此外， 伏特建立在 GPCV 的基础上，但也可以使用其他波动率估计方法，如这里描述的方法
诸如 Heston 模型 (Heston, 1993) 和 SABR (Hagan et al., 2002) 之类的随机波动率模型将证券价格和相关波动率的演变视为 SDE 的耦合系统。此类 SDE 通常用作金融衍生品定价的方法。与我们的观点不同，这些模型通常用于在风险中性措施下对股票期权进行定价，Volt 和 Magpie 我们专注于通过对观察结果进行调节来进行预测推断
Särkkä and Solin (2019)对高斯过程和 SDE 之间的联系进行了广泛的研究，他们提出了基于卡尔曼滤波的方法来估计 SDE 启发的 GP 模型中的 GP 超参数， 我们在这里不考虑，而是更喜欢基于边际似然的估计。线性微分方程系统先前已通过常微分方程 (Alvarez) et al., 2009) 和偏微分方程 (Särkkä, 2011) 的潜力模型集成到 GP 模型中。为了执行推理，Alvarez et al. (2009)得出与微分算子在特定内核上的线性投影相对应的协方差，而 Särkkä $( 2011) ; \text{ S\"{a} rkk\"{a} and Solin}$ (2019)则显式使用投影算子来开发内核函数以仿真SDE系统。同样，Zhu and Dunson (2013) 使用 SDE 导出嵌套的 GP , 但是他们的方法产生了具有非深但结构化的协方差函数的标准 GP。Gonzalvez et al. (2019) 中已经研究了 GP 的自回归均值函数，但是在他们的方法中，他们使用自回归特征作为 GP 模型的输入，而不是指定先验函数的方法
虽然上面的许多参考文献都专门针对金融，但 Volt 和Magpie 适用于广泛的领域，包括气候建模。自回归和波动率模型已成功应用于 Mehdizadeh et al. (2020); Liu et al. (2011)和 Tian et al. (2018)中的风和降水预报等领域
高斯过程自回归模型 (Requeima et al., 2019) 将不同任务的高斯过程进行堆叠，将一个任务的 GP 作为下一个任务的均值函数。因此，它与我们的移动平均线或多任务方法只有轻微的相似之处。此外，许多经过充分研究的自回归模型，例如。g。AR(p)族，可以写成高斯过程 (Williams, 2010) 。作为开发特定于域的内核功能的替代方法，可以替代地构造通用内核的手动组合，这需要像？, Ch5 中那样进行大量的手动调整。4、或求解离散优化问题 (Lloyd et al., 2014; Sun et al., 2018) 。由于我们希望高效简洁地开发我们的模型，我们也不考虑这些模型
$\bullet$ Deriving a hierarchical GP model, Volt, inspired
by stochastic volatility models that produces calibrated forecasts of stochastic time series (Section 3 ).
$\bullet$ Describing a simple but powerful mean function,
Magpie, that enables Gaussian process models to
accurately forecast trends (Section 3 ).
$\bullet$Using Volt and Magpie to produce highly cali-
brated forecasts in financial and climatological do-
$\operatorname*{mains}($Section 4).
$\bullet$Extending our procedure to multitask problems
by accounting for correlations in both volatility and price across different fnancial assets and different spatial locations (Section 5 ).
$\bullet$ We make code available here .
# 2. Related Work
早期的自回归方法来建模时间序列收益的波动率，例如 GARCH, 已经取得了广泛的成功 (Bollerslev, 1986) 。这些方法通常将波动率过程视为一个随时间变化的序列，对于推断和预测波动率是有效的，但通常不会像我们在 Volt 中那样直接与数据上的模型接口
波动率模型已扩展为使用神经网络或高斯过程作为其基本组件。例如，Cao et al. (2020) 使用多层感知器来估计波动率表面，而 Luo et al. (2018) 使用具有推出功能的 RNN 来预测未来的波动率，但只考虑了一步的前瞻性价格预测。Wilson and Ghahramani (2010) 使用高斯过程通过拉普拉斯近似和 MCMC 抽样引人
|
[
[
0.358642578125,
0.058074951171875,
0.615234375,
0.058074951171875,
0.615234375,
0.07135009765625,
0.358642578125,
0.07135009765625
],
[
0.08905029296875,
0.085205078125,
0.4755859375,
0.08709716796875,
0.4755859375,
0.10101318359375,
0.08905029296875,
0.0985107421875
],
[
0.5,
0.08587646484375,
0.8857421875,
0.08587646484375,
0.8857421875,
0.0997314453125,
0.5,
0.0997314453125
],
[
0.08905029296875,
0.10162353515625,
0.370849609375,
0.10162353515625,
0.370849609375,
0.11553955078125,
0.08905029296875,
0.11553955078125
],
[
0.50244140625,
0.10162353515625,
0.8837890625,
0.10162353515625,
0.8837890625,
0.11553955078125,
0.50244140625,
0.11553955078125
],
[
0.50244140625,
0.117431640625,
0.8837890625,
0.117431640625,
0.8837890625,
0.1307373046875,
0.50244140625,
0.1307373046875
],
[
0.08905029296875,
0.125,
0.473876953125,
0.125,
0.473876953125,
0.1383056640625,
0.08905029296875,
0.1383056640625
],
[
0.5009765625,
0.1319580078125,
0.88232421875,
0.1319580078125,
0.88232421875,
0.1458740234375,
0.5009765625,
0.1458740234375
],
[
0.09234619140625,
0.141357421875,
0.47216796875,
0.141357421875,
0.47216796875,
0.1514892578125,
0.09234619140625,
0.1514892578125
],
[
0.5009765625,
0.147705078125,
0.8857421875,
0.147705078125,
0.8857421875,
0.1610107421875,
0.5009765625,
0.1610107421875
],
[
0.0906982421875,
0.1546630859375,
0.473876953125,
0.1546630859375,
0.473876953125,
0.1685791015625,
0.0906982421875,
0.1685791015625
],
[
0.0882568359375,
0.169189453125,
0.47216796875,
0.1685791015625,
0.47216796875,
0.18310546875,
0.0882568359375,
0.184326171875
],
[
0.50341796875,
0.1641845703125,
0.88232421875,
0.1641845703125,
0.88232421875,
0.1749267578125,
0.50341796875,
0.1749267578125
],
[
0.5009765625,
0.177978515625,
0.8857421875,
0.177978515625,
0.8857421875,
0.1912841796875,
0.5009765625,
0.1912841796875
],
[
0.09234619140625,
0.186279296875,
0.47216796875,
0.186279296875,
0.47216796875,
0.1962890625,
0.09234619140625,
0.1962890625
],
[
0.50244140625,
0.1925048828125,
0.8857421875,
0.1925048828125,
0.8857421875,
0.205810546875,
0.50244140625,
0.205810546875
],
[
0.09234619140625,
0.2020263671875,
0.47216796875,
0.2020263671875,
0.47216796875,
0.212158203125,
0.09234619140625,
0.212158203125
],
[
0.08905029296875,
0.21337890625,
0.4755859375,
0.2139892578125,
0.4755859375,
0.2291259765625,
0.08905029296875,
0.2279052734375
],
[
0.50244140625,
0.208984375,
0.88232421875,
0.208984375,
0.88232421875,
0.2197265625,
0.50244140625,
0.2197265625
],
[
0.50244140625,
0.222900390625,
0.8857421875,
0.222900390625,
0.8857421875,
0.236083984375,
0.50244140625,
0.236083984375
],
[
0.09234619140625,
0.231689453125,
0.4306640625,
0.231689453125,
0.4306640625,
0.242431640625,
0.09234619140625,
0.242431640625
],
[
0.50244140625,
0.238037109375,
0.8857421875,
0.238037109375,
0.8857421875,
0.251953125,
0.50244140625,
0.251953125
],
[
0.0906982421875,
0.25244140625,
0.473876953125,
0.25244140625,
0.473876953125,
0.266357421875,
0.0906982421875,
0.266357421875
],
[
0.5,
0.25244140625,
0.73046875,
0.25244140625,
0.73046875,
0.266357421875,
0.5,
0.266357421875
],
[
0.09234619140625,
0.268310546875,
0.47216796875,
0.268310546875,
0.47216796875,
0.2822265625,
0.09234619140625,
0.2822265625
],
[
0.50244140625,
0.27587890625,
0.8837890625,
0.27587890625,
0.8837890625,
0.2890625,
0.50244140625,
0.2890625
],
[
0.0931396484375,
0.285400390625,
0.47216796875,
0.285400390625,
0.47216796875,
0.29541015625,
0.0931396484375,
0.29541015625
],
[
0.50244140625,
0.292236328125,
0.88232421875,
0.292236328125,
0.88232421875,
0.302978515625,
0.50244140625,
0.302978515625
],
[
0.09234619140625,
0.2998046875,
0.47216796875,
0.2998046875,
0.47216796875,
0.31005859375,
0.09234619140625,
0.31005859375
],
[
0.5009765625,
0.30615234375,
0.8837890625,
0.30615234375,
0.8837890625,
0.3193359375,
0.5009765625,
0.3193359375
],
[
0.0906982421875,
0.3125,
0.473876953125,
0.313232421875,
0.473876953125,
0.326904296875,
0.0906982421875,
0.32568359375
],
[
0.50341796875,
0.322509765625,
0.8837890625,
0.322509765625,
0.8837890625,
0.333251953125,
0.50341796875,
0.333251953125
],
[
0.09234619140625,
0.328857421875,
0.473876953125,
0.328857421875,
0.473876953125,
0.34228515625,
0.09234619140625,
0.34228515625
],
[
0.50244140625,
0.3359375,
0.8837890625,
0.3359375,
0.8837890625,
0.349853515625,
0.50244140625,
0.349853515625
],
[
0.09234619140625,
0.34521484375,
0.47216796875,
0.34521484375,
0.47216796875,
0.35595703125,
0.09234619140625,
0.35595703125
],
[
0.50244140625,
0.35107421875,
0.8837890625,
0.35107421875,
0.8837890625,
0.3642578125,
0.50244140625,
0.3642578125
],
[
0.09234619140625,
0.35986328125,
0.1707763671875,
0.35986328125,
0.1707763671875,
0.371826171875,
0.09234619140625,
0.371826171875
],
[
0.50341796875,
0.367431640625,
0.646484375,
0.367431640625,
0.646484375,
0.37744140625,
0.50341796875,
0.37744140625
],
[
0.08905029296875,
0.380615234375,
0.2442626953125,
0.380615234375,
0.2442626953125,
0.39453125,
0.08905029296875,
0.39453125
],
[
0.5,
0.386474609375,
0.8857421875,
0.38818359375,
0.8857421875,
0.40283203125,
0.5,
0.40087890625
],
[
0.50244140625,
0.404052734375,
0.8837890625,
0.404052734375,
0.8837890625,
0.417236328125,
0.50244140625,
0.417236328125
],
[
0.10296630859375,
0.415283203125,
0.473876953125,
0.415283203125,
0.473876953125,
0.4287109375,
0.10296630859375,
0.4287109375
],
[
0.50244140625,
0.41845703125,
0.8837890625,
0.41845703125,
0.8837890625,
0.431884765625,
0.50244140625,
0.431884765625
],
[
0.1217041015625,
0.429931640625,
0.473876953125,
0.429931640625,
0.473876953125,
0.443115234375,
0.1217041015625,
0.443115234375
],
[
0.50244140625,
0.434326171875,
0.8837890625,
0.434326171875,
0.8837890625,
0.447509765625,
0.50244140625,
0.447509765625
],
[
0.12091064453125,
0.4443359375,
0.471435546875,
0.4443359375,
0.471435546875,
0.458251953125,
0.12091064453125,
0.458251953125
],
[
0.5009765625,
0.449462890625,
0.88232421875,
0.449462890625,
0.88232421875,
0.46337890625,
0.5009765625,
0.46337890625
],
[
0.12091064453125,
0.458984375,
0.1519775390625,
0.46337890625,
0.148681640625,
0.47607421875,
0.11767578125,
0.4716796875
],
[
0.5,
0.464111328125,
0.8837890625,
0.464111328125,
0.8837890625,
0.477783203125,
0.5,
0.477783203125
],
[
0.50244140625,
0.479248046875,
0.88232421875,
0.479248046875,
0.88232421875,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.10455322265625,
0.4873046875,
0.47216796875,
0.4873046875,
0.47216796875,
0.50048828125,
0.10455322265625,
0.50048828125
],
[
0.5009765625,
0.49365234375,
0.8857421875,
0.494384765625,
0.8857421875,
0.50830078125,
0.5009765625,
0.5068359375
],
[
0.1217041015625,
0.501953125,
0.47216796875,
0.501953125,
0.47216796875,
0.5166015625,
0.1217041015625,
0.5166015625
],
[
0.50244140625,
0.51025390625,
0.8837890625,
0.51025390625,
0.8837890625,
0.5234375,
0.50244140625,
0.5234375
],
[
0.1217041015625,
0.517578125,
0.398681640625,
0.517578125,
0.398681640625,
0.53076171875,
0.1217041015625,
0.53076171875
],
[
0.50341796875,
0.52587890625,
0.8837890625,
0.52587890625,
0.8837890625,
0.5361328125,
0.50341796875,
0.5361328125
],
[
0.10296630859375,
0.54248046875,
0.47216796875,
0.54248046875,
0.47216796875,
0.55615234375,
0.10296630859375,
0.55615234375
],
[
0.50244140625,
0.5390625,
0.8857421875,
0.5390625,
0.8857421875,
0.55322265625,
0.50244140625,
0.55322265625
],
[
0.1217041015625,
0.5595703125,
0.47216796875,
0.5595703125,
0.47216796875,
0.57275390625,
0.1217041015625,
0.57275390625
],
[
0.50244140625,
0.5546875,
0.8857421875,
0.5546875,
0.8857421875,
0.568359375,
0.50244140625,
0.568359375
],
[
0.12335205078125,
0.57470703125,
0.25732421875,
0.57470703125,
0.25732421875,
0.58837890625,
0.12335205078125,
0.58837890625
],
[
0.50244140625,
0.5693359375,
0.8837890625,
0.5693359375,
0.8837890625,
0.58349609375,
0.50244140625,
0.58349609375
],
[
0.50244140625,
0.58544921875,
0.8837890625,
0.58544921875,
0.8837890625,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.10296630859375,
0.60107421875,
0.471435546875,
0.60107421875,
0.471435546875,
0.6142578125,
0.10296630859375,
0.6142578125
],
[
0.5009765625,
0.60107421875,
0.8837890625,
0.60107421875,
0.8837890625,
0.6142578125,
0.5009765625,
0.6142578125
],
[
0.1217041015625,
0.6171875,
0.47216796875,
0.6171875,
0.47216796875,
0.630859375,
0.1217041015625,
0.630859375
],
[
0.5009765625,
0.61572265625,
0.68115234375,
0.61572265625,
0.68115234375,
0.62890625,
0.5009765625,
0.62890625
],
[
0.12335205078125,
0.6318359375,
0.473876953125,
0.6318359375,
0.473876953125,
0.64599609375,
0.12335205078125,
0.64599609375
],
[
0.5048828125,
0.6396484375,
0.8837890625,
0.6396484375,
0.8837890625,
0.6494140625,
0.5048828125,
0.6494140625
],
[
0.12091064453125,
0.64794921875,
0.37744140625,
0.64794921875,
0.37744140625,
0.6611328125,
0.12091064453125,
0.6611328125
],
[
0.50244140625,
0.65478515625,
0.8837890625,
0.65478515625,
0.8837890625,
0.66552734375,
0.50244140625,
0.66552734375
],
[
0.101318359375,
0.673828125,
0.340576171875,
0.673828125,
0.340576171875,
0.6875,
0.101318359375,
0.6875
],
[
0.50244140625,
0.66845703125,
0.8857421875,
0.66845703125,
0.8857421875,
0.681640625,
0.50244140625,
0.681640625
],
[
0.50244140625,
0.6845703125,
0.8837890625,
0.6845703125,
0.8837890625,
0.69775390625,
0.50244140625,
0.69775390625
],
[
0.50244140625,
0.69873046875,
0.583984375,
0.69873046875,
0.583984375,
0.71044921875,
0.50244140625,
0.71044921875
],
[
0.08905029296875,
0.70849609375,
0.2344970703125,
0.70849609375,
0.2344970703125,
0.7216796875,
0.08905029296875,
0.7216796875
],
[
0.50244140625,
0.72021484375,
0.8857421875,
0.72021484375,
0.8857421875,
0.734375,
0.50244140625,
0.734375
],
[
0.0906982421875,
0.734375,
0.47216796875,
0.734375,
0.47216796875,
0.74755859375,
0.0906982421875,
0.74755859375
],
[
0.50244140625,
0.7373046875,
0.88134765625,
0.7373046875,
0.88134765625,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.09234619140625,
0.75048828125,
0.471435546875,
0.75048828125,
0.471435546875,
0.76123046875,
0.09234619140625,
0.76123046875
],
[
0.5048828125,
0.75244140625,
0.8837890625,
0.75244140625,
0.8837890625,
0.76318359375,
0.5048828125,
0.76318359375
],
[
0.098876953125,
0.7646484375,
0.47216796875,
0.7646484375,
0.47216796875,
0.77783203125,
0.098876953125,
0.77783203125
],
[
0.5009765625,
0.7666015625,
0.8857421875,
0.7666015625,
0.8857421875,
0.77978515625,
0.5009765625,
0.77978515625
],
[
0.0906982421875,
0.77880859375,
0.473876953125,
0.77880859375,
0.473876953125,
0.79248046875,
0.0906982421875,
0.79248046875
],
[
0.096435546875,
0.76953125,
0.10296630859375,
0.76953125,
0.10296630859375,
0.77294921875,
0.096435546875,
0.77294921875
],
[
0.50341796875,
0.78271484375,
0.8837890625,
0.78271484375,
0.8837890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.08905029296875,
0.79345703125,
0.42236328125,
0.79345703125,
0.42236328125,
0.806640625,
0.08905029296875,
0.806640625
],
[
0.50244140625,
0.796875,
0.8857421875,
0.796875,
0.8857421875,
0.81005859375,
0.50244140625,
0.81005859375
],
[
0.0906982421875,
0.8173828125,
0.473876953125,
0.8173828125,
0.473876953125,
0.83056640625,
0.0906982421875,
0.83056640625
],
[
0.50341796875,
0.81298828125,
0.88134765625,
0.81298828125,
0.88134765625,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.0906982421875,
0.83203125,
0.473876953125,
0.83203125,
0.473876953125,
0.84521484375,
0.0906982421875,
0.84521484375
],
[
0.5009765625,
0.8271484375,
0.8857421875,
0.8271484375,
0.8857421875,
0.84033203125,
0.5009765625,
0.84033203125
],
[
0.0906982421875,
0.8466796875,
0.47216796875,
0.8466796875,
0.47216796875,
0.85986328125,
0.0906982421875,
0.85986328125
],
[
0.5009765625,
0.84228515625,
0.88232421875,
0.84228515625,
0.88232421875,
0.85595703125,
0.5009765625,
0.85595703125
],
[
0.5,
0.85498046875,
0.8857421875,
0.85595703125,
0.8857421875,
0.87060546875,
0.5,
0.869140625
],
[
0.08905029296875,
0.8623046875,
0.47216796875,
0.8623046875,
0.47216796875,
0.87548828125,
0.08905029296875,
0.87548828125
],
[
0.0906982421875,
0.876953125,
0.471435546875,
0.876953125,
0.471435546875,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.5009765625,
0.8720703125,
0.69775390625,
0.8720703125,
0.69775390625,
0.88525390625,
0.5009765625,
0.88525390625
],
[
0.0906982421875,
0.892578125,
0.47216796875,
0.892578125,
0.47216796875,
0.90576171875,
0.0906982421875,
0.90576171875
]
] |
[
"Title Suppressed Due to Excessive Size",
"个GP对波动率进行建模，第二个GP用于预测给定",
"高斯过程copula波动率模型（GPCV）对波动率进",
"样本数据的分布来自波动率GP（底行）",
"行参数化，而Wuetal.（2014）使用GP状态空间模",
"型和粒子滤波器对波动率进行估计。与Wilsonand",
"Volt描述的协方差结构提供了预测中不确定性的忠实",
"Ghahramani (2010)类似,Lazaro-Gredilla and Titsias",
"表示，但忽略了数据空间GP的均值函数，这是捕获",
"(2011）使用高斯过程对具有指数联系的波动率模型进",
"数据趋势的强大工具。为此，我们共同介绍了移动平",
"均高斯过程或Magpie，其中我们用移动平均值替换了",
"行参数化，但使用高度结构化的变分近似进行推断",
"Liuetal.（2020）使用多任务高斯过程来预测未来的波",
"GP中的标准参数均值函数。移动平均线是气候学和",
"动性，将其模型应用于外汇货币收益，再次对价格进",
"金融学等领域广泛使用的技术（Nau，2014）。通过将",
"移动平均线的趋势拟合功能与GP的概率框架相结合，",
"行一步预测。至关重要的是，仅预测波动性确实可以",
"直接预测数据，这是我们使用Volt的主要目标。此外，",
"我们可以生成既准确又具有校准不确定性的预测",
"伏特建立在GPCV的基础上，但也可以使用其他波动",
"虽然Volt和Magpie可以单独使用，但我们将它们作",
"率估计方法，如这里描述的方法",
"为一个单独的作品呈现，因为它们的组合特别解决了",
"诸如Heston模型（Heston，1993）和SABR（Hagan",
"具有挑战性的预测问题。在股票价格或风速等时间演",
"etal.，2002）之类的随机波动率模型将证券价格和相关",
"变领域，过程的固有随机性使我们无法在遥远的未来",
"波动率的演变视为SDE的耦合系统。此类 SDE通常",
"产生准确的点估计，我们只需要准确性，但不确定的",
"用作金融衍生品定价的方法。与我们的观点不同，这",
"是忠实于数据的随机性。出于这个原因，人们既需要",
"些模型通常用于在风险中性措施下对股票期权进行定",
"喜提供的准确趋势捕提，也需要伏特提供的准确不",
"价，Volt和Magpie我们专注于通过对观察结果进行",
"确定性表示",
"调节来进行预测推断",
"我们的主要贡献如下：",
"Sarkka and Solin（2019）对高斯过程和SDE之间的",
"联系进行了广泛的研究，他们提出了基于卡尔曼滤波",
": Deriving a hierarchical GP model, Volt, inspired",
"的方法来估计SDE启发的GP模型中的GP超参数",
"by stochastic volatility models that produces cali-",
"我们在这里不考虑，而是更喜欢基于边际似然的估计。",
"brated forecasts of stochastic time series (Section",
"线性微分方程系统先前已通过常微分方程（Alvarez",
"3).",
"etal.，2009）和偏微分方程（Sarkka，2011）的潜力模",
"型集成到GP模型中。为了执行推理，Alvarezetal",
": Describing a simple but powerful mean function",
"(2009）得出与微分算子在特定内核上的线性投影相",
"Magpie, that enables Gaussian process models to",
"对应的协方差，而Sarkka（2011)；SarkkaandSolin",
"accurately forecast trends (Section 3).",
"（2019）则显式使用投影算子来开发内核函数以仿真",
". Using Volt and Magpie to produce highly cali",
"SDE系统。同样，ZhuandDunson（2013）使用SDE",
"brated forecasts in financial and climatological do",
"导出嵌套的GP，但是他们的方法产生了具有非深但结",
"mains (Section 4 ).",
"构化的协方差函数的标准GP。Gonzalvezetal.(2019)",
"中已经研究了GP的自回归均值函数，但是在他们的",
". Extending our procedure to multitask problems",
"方法中，他们使用自回归特征作为GP模型的输入，而",
"by accounting for correlations in both volatility",
"不是指定先验函数的方法",
"and price across different financial assets and dif-",
"虽然上面的许多参考文献都专门针对金融，但Volt和",
"ferent spatial locations (Section 5).",
"Magpie适用于广泛的领域，包括气候建模。自回归",
": We make code available here",
"和波动率模型已成功应用于Mehdizadehetal.（2020)；",
"Liuetal.（2011）和Tianetal.（2018）中的风和降水",
"预报等领域",
"2. Related Work",
"高斯过程自回归模型（Requeimaetal.，2019）将不同",
"早期的自回归方法来建模时间序列收益的波动率，例",
"任务的高斯过程进行堆叠：将一个任务的GP作为下",
"如GARCH，已经取得了广泛的成功（Bollerslev，1986）",
"一个任务的均值函数。因此，它与我们的移动平均线",
"这些方法通常将波动率过程视为一个随时间变化的",
"或多任务方法只有轻微的相似之处。此外，许多经过",
"序列，对于推断和预测波动率是有效的，但通常不会",
"",
"充分研究的自回归模型，例如。g。AR(p）族，可以写",
"像我们在Volt中那样直接与数据上的模型接口",
"成高斯过程（Williams，2010）。作为开发特定于域的",
"波动率模型已扩展为使用神经网络或高斯过程作为其",
"内核功能的替代方法，可以替代地构造通用内核的手",
"基本组件。例如，Cao.etal.（2020）使用多层感知器来",
"动组合，这需要像？，Ch5中那样进行大量的手动调",
"估计波动率表面，而Luoetal.（2018）使用具有推出",
"整。4、或求解离散优化问题（Lloydetal.，2014；Sun",
"etal.，2018）。由于我们希望高效简洁地开发我们的模",
"功能的RNN来预测未来的波动率，但只考虑了一步",
"的前瞻性价格预测。Wilson andGhahramani（2010)",
"型，我们也不考虑这些模型",
"使用高斯过程通过拉普拉斯近似和MCMC抽样引入"
] |
{
"arxiv_id": "2207.06544",
"language": "zh",
"page_idx": 1
}
|
||
2207.06544-zh-3
|
2207.06544-zh-3.jpg
|
Title Suppressed Due to Excessive Size
据上产生对数正态分布的混合。对后验进行采样需要在测试输入上采样$N_v$对数波动率路径$v^*$,对于其中的每个，我们生成一个内核$K_s(t,t^{\prime},V^*)$ ,并采样$N_s$ 数据路径$S^*=\exp(s^*)$,生成$N_v\times N_s$样本
在 3.1 节的高斯过程观点中，SDE 过程的标准蒙特卡洛模拟等效于从方程 (5)的先验分布到时间$T$而不是后验分布 (Sauer, 2012) 的采样对数波动率路径$v^*=v(\boldsymbol{t}^*)$,。使用 $S_T^*=\exp(s_T^*)$,的先前样本，我们可以对价格上的未来分布进行蒙特卡洛估计。但是，这种方法与我们使用 Volt 进行采样的方法之间的区别在于，Volt 从波动率的后验分布和以观察为条件的数据中进行采样，而基于 SDE 的方法则从波动率的先验分布中进行采样
3.2. Inference
在这里，我们概述了使用一系列价格观察来训练方程5 中 GP 的超参数并形成相关的后验预测分布的过程。通常，训练过程可以被认为是一个三步过程：a) 使用高斯过程 Copula 波动率 (GPCV) 模型来推断波动率路径$V$,给定一系列观测值$S$,b)通过最大化观察到的 GPCV 相对于 GPCV 推断的波动率的边际对数似然(MLL)来学习 GP 在对数波动率空间中的超参数。。请注意，我们使用 GPCV 来估计波动率是一种建模选择，我们也可以使用任何其他波动率估计模型， 如 GARCH
Inferring Volatility from Training Data 公式 (5)中概述的模型的一个挑战是，对于某些范围的训练观察， 需要同时具有数据和波动率观察。为了估计波动率，我们使用 Wilson and Ghahramani (2010) 首次提出的高斯 Copula 过程波动率 (GPCV) 模型的变体。我们的GPCV 模型使用扭曲的高斯过程来根据以下公式对响应$w(t)$的可变性进行建模：
$$\begin{aligned}f(t)&\sim\mathcal{GP}(c,K_v(t,t'))\\\gamma(f(t))&=\exp\{f(t)\}\\w(t)&\sim\mathcal{N}(0,\gamma^2(f(t))).\end{aligned}$$
(7)
我们使用从方程 (5)中的对数波动率 SDE 得出的内核来推断潜在函数$f(t),$,并使用变分推断 (Hensman et al., 2013; 2015) 来训练模型。有关更多详细信息，请参见附录 B.3
遵循 Wilson and Ghahramani (2010) 的价格估计方法，我们将响应视为数据的对数收益，即：$w(t_i)=$ $\log S(t_i)-\log S(t_{i-1}).$我们通过从$f(t)$中绘制后验样本并将其传递给翘曲函数$\sigma(\cdot);$来构造$0,\cdots,t-1$上的波动率预测，因此我们对$V(t)$的估计是 THMAT Training the Gaussian Processes 给定与使用 GPCV 学习的训练数据相关的波动率路径，我们假设根据方程 (5) 对对数波动率和对数数据进行高斯过程先验。给定训练数据上的波动率，对数波动率模型的单个超参数是描述 volvol 的$\sigma^2$项。对数数据模型的超参数只是方程 (5) 中均值的参数，如果我们使用像喜鹊这样的非参数均值，则没有。为了训练，我们使用基于梯度的优化 (Rasmussen and Williams, 2008, Chapter 5)最大化模型相对于其超参数的 MLL。无论使用喜鹊均值如何，Volt 中推理的总计算成本只是在均匀间隔的数据上训练一个变分 GP 和两个标准 GP 模型的成本，这可以通过利用数据(Wilson and Nickisch, 2015) 的 (Toeplitz) 结构来有效地完成
3.3. Predictions
在 Volt 中，我们将对数波动率 GP 设置在由 GPCV
Rollout Predictions 喜鹊意味着只允许提前一步进行预测，因此我们以推出的方式进行预测。也就是说，我们使用观测值$s_0,\ldots,s_t$从 GP 后验$p(s_t+1|s_0,\ldots,s_t)$ 中采样$\hat{s}_{t+1}$,然后在$\hat{s}_t+1$上条件我们的 GP (和喜鹊均值),以便从更新的 GP 后验$p(s_t+2|s_0,\ldots,s_t,\hat{s}_{t+1})$ 中采样$\hat{s} _{t+ 2}$ ,依此类推。这些推出预测对喜鹊框架至关重要。通过对价格预测进行顺序采样并使用每个观察值更新 GP,我们可以以其他 GP 无法实现的方式在移动平均均值中进行趋势反转。对于传统方法，转出是不必要的，因为每个时间步长上的条件方法都会分解为单个多元高斯分布
# 4. Forecasting
在金融和气候应用中，我们都将数据视为随机发展的， 因此有兴趣预测结果的分布，而不是点估计。因此，我们使用校准和负对数似然作为我们感兴趣的主要度量， 而不是像均方误差这样的精度度量
我们通过计算真实观测值小于预测分布的经验计算的$p^{th}$分位数的频率来计算百分位$p$的校准。更具体地说，对于在时间$T$和百分位$p$的价格股票$S$的预测， 我们计算预测 $q_T$ 的经验分位数，其中$\hat{\mathbb{P}}(S_T<q_T)=p$ 。然后，我们可以通过计算$C_p=\frac1K\sum_k\mathbb{I}_{\{S_{T_k}<q_{T_k}\}}$作为$S_T<q_T$在$p^th$不同预测上的平均频率，将$p$的校准计算为事件$S_T<q_T$的经验观察频率。如果我们的预测得到了很好的校准，那么对于$p$的每个值，该经验频率将接近$p$ ;因此，通过在一定的百分位数$p$范围内计算我们的预测的校准，我们可以确定预测分布的总体校准。这样的校准指标类似于在 Kuleshov et al. (2018)中探索的回归指标
请注意，为了在此设置中进行准确的校准，我们的预测分布必须与所有分位数的经验观察结果匹配。例如，
推断的对数路径上，并将对数数据 GP 设置在对数价 我们不能仅仅预测价格上涨的时间与观察到的价格上格的历史观察值上，并从后验分布中抽取样本，在数 涨频率相匹配的固定百分比，并期望实现准确的校准
|
[
[
0.358642578125,
0.058074951171875,
0.615234375,
0.058074951171875,
0.615234375,
0.07135009765625,
0.358642578125,
0.07135009765625
],
[
0.08905029296875,
0.08837890625,
0.1895751953125,
0.08837890625,
0.1895751953125,
0.0997314453125,
0.08905029296875,
0.0997314453125
],
[
0.50244140625,
0.08837890625,
0.8837890625,
0.08837890625,
0.8837890625,
0.0985107421875,
0.50244140625,
0.0985107421875
],
[
0.5009765625,
0.10162353515625,
0.88232421875,
0.10162353515625,
0.88232421875,
0.11553955078125,
0.5009765625,
0.11553955078125
],
[
0.08905029296875,
0.10986328125,
0.473876953125,
0.10986328125,
0.473876953125,
0.12371826171875,
0.08905029296875,
0.12371826171875
],
[
0.50244140625,
0.117431640625,
0.88232421875,
0.117431640625,
0.88232421875,
0.1307373046875,
0.50244140625,
0.1307373046875
],
[
0.08905029296875,
0.1256103515625,
0.473876953125,
0.1256103515625,
0.473876953125,
0.1395263671875,
0.08905029296875,
0.1395263671875
],
[
0.5,
0.1307373046875,
0.82666015625,
0.1307373046875,
0.82666015625,
0.14453125,
0.5,
0.14453125
],
[
0.0906982421875,
0.1427001953125,
0.471435546875,
0.1427001953125,
0.471435546875,
0.15283203125,
0.0906982421875,
0.15283203125
],
[
0.08905029296875,
0.1546630859375,
0.473876953125,
0.154052734375,
0.473876953125,
0.1685791015625,
0.08905029296875,
0.169189453125
],
[
0.5009765625,
0.1546630859375,
0.8837890625,
0.1546630859375,
0.8837890625,
0.1685791015625,
0.5009765625,
0.1685791015625
],
[
0.0906982421875,
0.17041015625,
0.473876953125,
0.17041015625,
0.473876953125,
0.184326171875,
0.0906982421875,
0.184326171875
],
[
0.5048828125,
0.1717529296875,
0.88232421875,
0.1717529296875,
0.88232421875,
0.1817626953125,
0.5048828125,
0.1817626953125
],
[
0.0906982421875,
0.1849365234375,
0.473876953125,
0.1849365234375,
0.473876953125,
0.1988525390625,
0.0906982421875,
0.1988525390625
],
[
0.5009765625,
0.1849365234375,
0.8837890625,
0.1849365234375,
0.8837890625,
0.1988525390625,
0.5009765625,
0.1988525390625
],
[
0.0906982421875,
0.2008056640625,
0.47216796875,
0.2008056640625,
0.47216796875,
0.2139892578125,
0.0906982421875,
0.2139892578125
],
[
0.5009765625,
0.2008056640625,
0.88232421875,
0.2008056640625,
0.88232421875,
0.2139892578125,
0.5009765625,
0.2139892578125
],
[
0.0906982421875,
0.216552734375,
0.47216796875,
0.216552734375,
0.47216796875,
0.2298583984375,
0.0906982421875,
0.2298583984375
],
[
0.50341796875,
0.216552734375,
0.8837890625,
0.216552734375,
0.8837890625,
0.2266845703125,
0.50341796875,
0.2266845703125
],
[
0.0906982421875,
0.2310791015625,
0.4794921875,
0.2310791015625,
0.4794921875,
0.2442626953125,
0.0906982421875,
0.2442626953125
],
[
0.5009765625,
0.2298583984375,
0.8837890625,
0.2298583984375,
0.8837890625,
0.24365234375,
0.5009765625,
0.24365234375
],
[
0.09234619140625,
0.2474365234375,
0.1732177734375,
0.2474365234375,
0.1732177734375,
0.2587890625,
0.09234619140625,
0.2587890625
],
[
0.5009765625,
0.24365234375,
0.8837890625,
0.2442626953125,
0.8837890625,
0.2587890625,
0.5009765625,
0.25830078125
],
[
0.50244140625,
0.260009765625,
0.88232421875,
0.260009765625,
0.88232421875,
0.2734375,
0.50244140625,
0.2734375
],
[
0.08905029296875,
0.276611328125,
0.47216796875,
0.276611328125,
0.47216796875,
0.290283203125,
0.08905029296875,
0.290283203125
],
[
0.50244140625,
0.276611328125,
0.5986328125,
0.276611328125,
0.5986328125,
0.28857421875,
0.50244140625,
0.28857421875
],
[
0.0906982421875,
0.29345703125,
0.47216796875,
0.29345703125,
0.47216796875,
0.306884765625,
0.0906982421875,
0.306884765625
],
[
0.09234619140625,
0.31005859375,
0.47216796875,
0.31005859375,
0.47216796875,
0.32080078125,
0.09234619140625,
0.32080078125
],
[
0.0906982421875,
0.322509765625,
0.47216796875,
0.323974609375,
0.47216796875,
0.337158203125,
0.0906982421875,
0.3359375
],
[
0.0882568359375,
0.337158203125,
0.47216796875,
0.33837890625,
0.47216796875,
0.352783203125,
0.0882568359375,
0.3515625
],
[
0.08905029296875,
0.3515625,
0.471435546875,
0.3515625,
0.471435546875,
0.365478515625,
0.08905029296875,
0.365478515625
],
[
0.0906982421875,
0.36865234375,
0.28515625,
0.36865234375,
0.28515625,
0.3818359375,
0.0906982421875,
0.3818359375
],
[
0.5009765625,
0.375732421875,
0.8837890625,
0.376953125,
0.8837890625,
0.39013671875,
0.5009765625,
0.389404296875
],
[
0.5009765625,
0.391357421875,
0.8857421875,
0.392578125,
0.8857421875,
0.406005859375,
0.5009765625,
0.40478515625
],
[
0.214111328125,
0.397705078125,
0.370849609375,
0.397705078125,
0.370849609375,
0.410888671875,
0.214111328125,
0.410888671875
],
[
0.50244140625,
0.4072265625,
0.8857421875,
0.4072265625,
0.8857421875,
0.421630859375,
0.50244140625,
0.421630859375
],
[
0.191162109375,
0.416748046875,
0.33251953125,
0.416748046875,
0.33251953125,
0.429931640625,
0.191162109375,
0.429931640625
],
[
0.453369140625,
0.419189453125,
0.473876953125,
0.419189453125,
0.473876953125,
0.429931640625,
0.453369140625,
0.429931640625
],
[
0.5009765625,
0.42041015625,
0.8857421875,
0.421630859375,
0.8857421875,
0.436279296875,
0.5009765625,
0.43505859375
],
[
0.2099609375,
0.436279296875,
0.366943359375,
0.436279296875,
0.366943359375,
0.45068359375,
0.2099609375,
0.45068359375
],
[
0.5009765625,
0.43505859375,
0.88623046875,
0.436279296875,
0.88623046875,
0.45068359375,
0.5009765625,
0.449462890625
],
[
0.50244140625,
0.451904296875,
0.88232421875,
0.451904296875,
0.88232421875,
0.46533203125,
0.50244140625,
0.46533203125
],
[
0.0906982421875,
0.46533203125,
0.47216796875,
0.46533203125,
0.47216796875,
0.479248046875,
0.0906982421875,
0.479248046875
],
[
0.50341796875,
0.468505859375,
0.8837890625,
0.468505859375,
0.8837890625,
0.479248046875,
0.50341796875,
0.479248046875
],
[
0.08905029296875,
0.48095703125,
0.471435546875,
0.48095703125,
0.471435546875,
0.494384765625,
0.08905029296875,
0.494384765625
],
[
0.5009765625,
0.48095703125,
0.8857421875,
0.481689453125,
0.8857421875,
0.49560546875,
0.5009765625,
0.494384765625
],
[
0.08905029296875,
0.49560546875,
0.473876953125,
0.49560546875,
0.473876953125,
0.5087890625,
0.08905029296875,
0.5087890625
],
[
0.5009765625,
0.49755859375,
0.8837890625,
0.49755859375,
0.8837890625,
0.51123046875,
0.5009765625,
0.51123046875
],
[
0.0906982421875,
0.51123046875,
0.1854248046875,
0.51220703125,
0.1854248046875,
0.5234375,
0.0906982421875,
0.52294921875
],
[
0.50244140625,
0.51220703125,
0.8837890625,
0.51220703125,
0.8837890625,
0.52587890625,
0.50244140625,
0.52587890625
],
[
0.5009765625,
0.52783203125,
0.68115234375,
0.52783203125,
0.68115234375,
0.54150390625,
0.5009765625,
0.54150390625
],
[
0.09234619140625,
0.53515625,
0.47216796875,
0.53515625,
0.47216796875,
0.54541015625,
0.09234619140625,
0.54541015625
],
[
0.0906982421875,
0.5498046875,
0.471435546875,
0.5498046875,
0.471435546875,
0.56298828125,
0.0906982421875,
0.56298828125
],
[
0.08905029296875,
0.564453125,
0.47216796875,
0.564453125,
0.47216796875,
0.57763671875,
0.08905029296875,
0.57763671875
],
[
0.498291015625,
0.56103515625,
0.6259765625,
0.56298828125,
0.6259765625,
0.57763671875,
0.498291015625,
0.57568359375
],
[
0.0906982421875,
0.5791015625,
0.473876953125,
0.5791015625,
0.473876953125,
0.59228515625,
0.0906982421875,
0.59228515625
],
[
0.5,
0.5859375,
0.88232421875,
0.5869140625,
0.88232421875,
0.6015625,
0.5,
0.60107421875
],
[
0.0906982421875,
0.59326171875,
0.463134765625,
0.59326171875,
0.463134765625,
0.60693359375,
0.0906982421875,
0.60693359375
],
[
0.50244140625,
0.60302734375,
0.8857421875,
0.60302734375,
0.8857421875,
0.6162109375,
0.50244140625,
0.6162109375
],
[
0.50244140625,
0.61865234375,
0.8857421875,
0.61865234375,
0.8857421875,
0.6318359375,
0.50244140625,
0.6318359375
],
[
0.08905029296875,
0.62548828125,
0.340576171875,
0.62548828125,
0.340576171875,
0.6396484375,
0.08905029296875,
0.6396484375
],
[
0.33251953125,
0.62548828125,
0.47216796875,
0.62548828125,
0.47216796875,
0.6396484375,
0.33251953125,
0.6396484375
],
[
0.5009765625,
0.63330078125,
0.74658203125,
0.6337890625,
0.74658203125,
0.64794921875,
0.5009765625,
0.646484375
],
[
0.08905029296875,
0.6416015625,
0.47216796875,
0.6416015625,
0.47216796875,
0.65478515625,
0.08905029296875,
0.65478515625
],
[
0.08905029296875,
0.65576171875,
0.4755859375,
0.65576171875,
0.4755859375,
0.66943359375,
0.08905029296875,
0.66943359375
],
[
0.5009765625,
0.65576171875,
0.8857421875,
0.65576171875,
0.8857421875,
0.66943359375,
0.5009765625,
0.66943359375
],
[
0.0906982421875,
0.671875,
0.473876953125,
0.671875,
0.473876953125,
0.68505859375,
0.0906982421875,
0.68505859375
],
[
0.5,
0.66943359375,
0.8837890625,
0.66943359375,
0.8837890625,
0.68310546875,
0.5,
0.68310546875
],
[
0.0906982421875,
0.68505859375,
0.473876953125,
0.68505859375,
0.473876953125,
0.69873046875,
0.0906982421875,
0.69873046875
],
[
0.5009765625,
0.68603515625,
0.88232421875,
0.68603515625,
0.88232421875,
0.69970703125,
0.5009765625,
0.69970703125
],
[
0.0931396484375,
0.70263671875,
0.471435546875,
0.70263671875,
0.471435546875,
0.71337890625,
0.0931396484375,
0.71337890625
],
[
0.5009765625,
0.7021484375,
0.8857421875,
0.7021484375,
0.8857421875,
0.71533203125,
0.5009765625,
0.71533203125
],
[
0.0906982421875,
0.71630859375,
0.471435546875,
0.71630859375,
0.471435546875,
0.72998046875,
0.0906982421875,
0.72998046875
],
[
0.498291015625,
0.7138671875,
0.88818359375,
0.7138671875,
0.88818359375,
0.73486328125,
0.498291015625,
0.73486328125
],
[
0.08905029296875,
0.73095703125,
0.47216796875,
0.73095703125,
0.47216796875,
0.744140625,
0.08905029296875,
0.744140625
],
[
0.5009765625,
0.734375,
0.8857421875,
0.734375,
0.8857421875,
0.74755859375,
0.5009765625,
0.74755859375
],
[
0.08905029296875,
0.7470703125,
0.473876953125,
0.7470703125,
0.473876953125,
0.76025390625,
0.08905029296875,
0.76025390625
],
[
0.5009765625,
0.75,
0.8837890625,
0.75,
0.8837890625,
0.76318359375,
0.5009765625,
0.76318359375
],
[
0.08905029296875,
0.76123046875,
0.47216796875,
0.76123046875,
0.47216796875,
0.7744140625,
0.08905029296875,
0.7744140625
],
[
0.50244140625,
0.7666015625,
0.88232421875,
0.7666015625,
0.88232421875,
0.7763671875,
0.50244140625,
0.7763671875
],
[
0.0906982421875,
0.7763671875,
0.473876953125,
0.7763671875,
0.473876953125,
0.79052734375,
0.0906982421875,
0.79052734375
],
[
0.50244140625,
0.78076171875,
0.88232421875,
0.78076171875,
0.88232421875,
0.79150390625,
0.50244140625,
0.79150390625
],
[
0.09234619140625,
0.79345703125,
0.47216796875,
0.79345703125,
0.47216796875,
0.8037109375,
0.09234619140625,
0.8037109375
],
[
0.50244140625,
0.79443359375,
0.8837890625,
0.79443359375,
0.8837890625,
0.80810546875,
0.50244140625,
0.80810546875
],
[
0.08905029296875,
0.806640625,
0.318603515625,
0.806640625,
0.318603515625,
0.82080078125,
0.08905029296875,
0.82080078125
],
[
0.50341796875,
0.81103515625,
0.88134765625,
0.81103515625,
0.88134765625,
0.8212890625,
0.50341796875,
0.8212890625
],
[
0.50244140625,
0.82568359375,
0.6845703125,
0.82568359375,
0.6845703125,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.0882568359375,
0.83837890625,
0.20751953125,
0.83837890625,
0.20751953125,
0.8515625,
0.0882568359375,
0.8515625
],
[
0.50341796875,
0.8466796875,
0.8857421875,
0.8466796875,
0.8857421875,
0.85986328125,
0.50341796875,
0.85986328125
],
[
0.08905029296875,
0.8623046875,
0.473876953125,
0.8623046875,
0.473876953125,
0.87548828125,
0.08905029296875,
0.87548828125
],
[
0.50244140625,
0.8623046875,
0.8857421875,
0.8623046875,
0.8857421875,
0.87548828125,
0.50244140625,
0.87548828125
],
[
0.0906982421875,
0.87890625,
0.471435546875,
0.87890625,
0.471435546875,
0.8896484375,
0.0906982421875,
0.8896484375
],
[
0.50244140625,
0.87890625,
0.8837890625,
0.87890625,
0.8837890625,
0.8896484375,
0.50244140625,
0.8896484375
],
[
0.0882568359375,
0.8916015625,
0.473876953125,
0.892578125,
0.473876953125,
0.90576171875,
0.0882568359375,
0.90478515625
],
[
0.50244140625,
0.89453125,
0.8759765625,
0.89453125,
0.8759765625,
0.90478515625,
0.50244140625,
0.90478515625
]
] |
[
"Title Suppressed Due to Excessive Size",
"3.2.Inference",
"据上产生对数正态分布的混合。对后验进行采样需要",
"在测试输入上采样N，对数波动率路径*，对于其中",
"在这里，我们概述了使用一系列价格观察来训练方程",
"的每个，我们生成一个内核Ks（t,t,V*），并采样N",
"5中GP的超参数并形成相关的后验预测分布的过程。",
"数据路径S*=exp(s*)，生成N×N，样本",
"通常，训练过程可以被认为是一个三步过程：a）使用",
"高斯过程Copula波动率（GPCV）模型来推断波动率",
"在3.1节的高斯过程观点中，SDE过程的标准蒙特",
"路径V，给定一系列观测值S，b）通过最大化观察",
"卡洛模拟等效于从方程（5）的先验分布到时间T而",
"到的GPCV相对于GPCV推断的波动率的边际对数",
"不是后验分布（Sauer，2012）的采样对数波动率路径",
"似然（MLL）来学习GP在对数波动率空间中的超参",
"u*=u(t*)），。使用S=exp(s），的先前样本，我们可",
"数。。请注意，我们使用GPCV来估计波动率是一种",
"以对价格上的未来分布进行蒙特卡洛估计。但是，这",
"建模选择，我们也可以使用任何其他波动率估计模型，",
"种方法与我们使用Volt进行采样的方法之间的区别在",
"如GARCH",
"于，Volt从波动率的后验分布和以观察为条件的数据",
"中进行采样，而基于SDE的方法则从波动率的先验分",
"InferringVolatilityfromTrainingData公式(5)中",
"布中进行采样",
"概述的模型的一个挑战是，对于某些范围的训练观察，",
"需要同时具有数据和波动率观察。为了估计波动率，我",
"们使用WilsonandGhahramani（2010）首次提出的高",
"斯Copula过程波动率（GPCV）模型的变体。我们的",
"GPCV模型使用扭曲的高斯过程来根据以下公式对响",
"应w(t)的可变性进行建模：",
"RolloutPredictions喜鹊意味着只允许提前一步进行",
"预测，因此我们以推出的方式进行预测。也就是说，我",
"f(t) ~ GP(c, K,(t,t))",
"们使用观测值S0,···,8t从GP后验p(st+1so,·.·,8t）",
"(f(t)) =exp(f(t))",
"(7)",
"中采样st+1，然后在St+1上条件我们的GP（和喜",
"w(t) ~ N(0, (f(t)))",
"均值），以便从更新的GP后验p(st+2s0，...，St，$t+1)",
"中采样St+2，依此类推。这些推出预测对喜鹊框架至",
"我们使用从方程（5）中的对数波动率SDE得出的内",
"关重要。通过对价格预测进行顺序采样并使用每个观",
"核来推断潜在函数f(t)，，并使用变分推断(Hensman",
"察值更新GP，我们可以以其他GP无法实现的方式",
"etal.，2013；2015）来训练模型。有关更多详细信息，请",
"在移动平均均值中进行趋势反转。对于传统方法，转",
"参见附录B.3",
"出是不必要的，因为每个时间步长上的条件方法都会",
"分解为单个多元高斯分布",
"遵循wilsonandGhahramani（2010）的价格估计方",
"法，我们将响应视为数据的对数收益，即：w(ti）=",
"logS(t)-logS(ti-1).我们通过从f(t)中绘制后验样",
"4.Forecasting",
"本并将其传递给翘曲函数α();来构造0,.·,t-1上",
"在金融和气候应用中，我们都将数据视为随机发展的",
"的波动率预测，因此我们对V(t）的估计是THMAT",
"因此有兴趣预测结果的分布，而不是点估计。因此，我",
"们使用校准和负对数似然作为我们感兴趣的主要度量，",
"Training the Gaussian Processes",
"给定与使用GPCV",
"而不是像均方误差这样的精度度量",
"学习的训练数据相关的波动率路径，我们假设根据方",
"程（5）对对数波动率和对数数据进行高斯过程先验。",
"我们通过计算真实观测值小于预测分布的经验计算的",
"给定训练数据上的波动率，对数波动率模型的单个超",
"pth分位数的频率来计算百分位p的校准。更具体地",
"参数是描述volvol 的α2项。对数数据模型的超参数",
"说，对于在时间T和百分位p的价格股票S的预测",
"只是方程（5）中均值的参数，如果我们使用像喜鹊这",
"我们计算预测qT的经验分位数，其中P(S<qT）=p",
"样的非参数均值，则没有。为了训练，我们使用基于",
"。然后，我们可以通过计算Cp=I(ST<qT）作",
"梯度的优化（Rasmussen and Williams，2008，Chapter",
"为ST<qT在pth不同预测上的平均频率，将p的校",
"5）最大化模型相对于其超参数的MLL。无论使用喜鹊",
"准计算为事件ST<qT的经验观察频率。如果我们的",
"均值如何，Volt中推理的总计算成本只是在均匀间隔",
"预测得到了很好的校准，那么对于P的每个值，该经",
"的数据上训练一个变分GP和两个标准GP模型的成",
"验频率将接近p；因此，通过在一定的百分位数p范",
"本，这可以通过利用数据（wilsonandNickisch，2015）",
"围内计算我们的预测的校准，我们可以确定预测分布",
"的（Toeplitz）结构来有效地完成",
"的总体校准。这样的校准指标类似于在Kuleshovetal",
"(2018）中探索的回归指标",
"3.3. Predictions",
"请注意，为了在此设置中进行准确的校准，我们的预",
"在Volt中，我们将对数波动率GP设置在由GPCV",
"测分布必须与所有分位数的经验观察结果匹配。例如，",
"推断的对数路径上，并将对数数据GP设置在对数价",
"我们不能仅仅预测价格上涨的时间与观察到的价格上",
"格的历史观察值上，并从后验分布中抽取样本，在数",
"涨率相匹配的固定百分比，并期望实现准确的校准"
] |
{
"arxiv_id": "2207.06544",
"language": "zh",
"page_idx": 3
}
|
||
2305.16806-zh-1
|
2305.16806-zh-1.jpg
|
$\mathbf{GPT}$是否产生较少的文字翻译？
Vikas Raunak Arul Menezes
$\mathbf{MattPost}$
Hany Hassan Awadalla
Microsoft Azure AI Redmond, Washington
$\{$ viraunak,arulm,mpost,hanyh $\}$ @microsoft.com
# Abstract
Large Language Models (LLMs) such as GPT- 3 have emerged as general-purpose language models capable of addressing many natural language generation or understanding tasks. On the task of Machine Translation (MT), multiple works have investigated few-shot prompting mechanisms to elicit better translations from LLMs. However, there has been relatively little investigation on how such translations differ qualitatively from the translations generated by standard Neural Machine Translation (NMT) models. In this work, we investigate these differences in terms of the literalness of translations produced by the two systems. Using literalness measures involving word alignment and monotonicity, we find that translations out of English (E$\to\mathbb{X}$) from GPTs tend to be less literal, while exhibiting similar or better scores on MT quality metrics. We demonstrate that this finding is borne out in human evaluations as well. We then show that these differences are especially pronounced when translating sentences that contain idiomatic expressions.
## 1 $\textbf{Introduction}$
尽管仅在语言建模目标上进行了培训，但在对齐的并行数据 (Briakou et al., 2023) 上没有明确的监督，但诸如 GPT-3 或 PaLM (Brown et al., 2020; Chowdhery et al., 2022) 之类的 LLM 在少量提示下实现了接近最先进的翻译性能 (Vilar et al., 2022; Hendy et al., 2023)。调查这些模型的输出的工作已经注意到，当使用较旧的基于表面的指标(如 BLEU(Papineni et al., 2002a) ) 时，性能的提高是不可见的，这通常显示出相对于 NMT 系统的巨大损失。这就提出了一个问题：这些 LLM 翻译在质量上与传统 NMT 系统有何不同？
我们使用翻译文字的属性来探讨这个问题机器翻译系统长期以来一直被认为倾向于产生过于直译的 (Dankers et al.,2022b),我们已经观察到 LLM 似乎不太容易受到这个问题的影
<table>
<tbody>
<tr>
<th>SOurce</th>
<th>He survived by</th>
<th>the skin</th>
<th>skin his $of1$ $\frac{1}{2}$ 1 1</th>
<th>teeth</th>
</tr>
<tr>
<td>$NMT$</td>
<td>Il a</td>
<td>survéc</td>
<td>survécu</td>
<td>par</td>
</tr>
<tr>
<td>GPT- 3</td>
<td>$Il$ a: survécu de</td>
<td>1ustesse</td>
<td>stesse</td>
<td> </td>
</tr>
</tbody>
</table>
Table 1: An example where GPT-3 produces a more natural (non-literal) translation of an English idiom. When word-aligning these sentences, the source word skin remains unaligned for the GPT-3 translation.
响(表 1)。我们调查这些观察结果是否可以定量验证。首先，我们使用基于单词对齐和单调性的度量来量化 LLM 是否比 NMT 系统产生更少的文字翻译，并将这些数字用于人类评估$(\S$E2)。接下来，我们专门研究成语，比较它们在自然和合成数据设置下的翻译方式(§E3
我们的调查重点是英语和德语，中文和俄语这三种类型不同的语言之间的翻译。我们的发现总结如下：(1)我们发现，当用英语翻译时， 来自 GPT 系列 LLM 的两个 LLM 的翻译确实确实比 NMT 同行的翻译少，并且(2)在具有惯用表达的句子的情况下尤其如此
# 2 Quantifying Translation Literalness
我们将最先进的 NMT 系统与最有能力的可公开访问的 GPT 模型 (在撰写本文时) 进行了比较，这些模型旨在捕捉翻译字面量的差异。我们进行基于指标的自动评估和人工评估。我们在下面解释评估和实验细节
Datasets 我们使用官方的 WMT21En-De, DeEn, En-Ru 和 Ru-En 新闻翻译测试集来评估(Barrault et al., 2021)
Measures of Quality 我们使用 COMET-QE $^{1}$(Rei et al.,2020)作为质量估计(QE)度量(Fomicheva et al., 2020) 来量化翻译的流畅性和充分性。使用 QE 作为度量的优点是，它排除了任何参考偏差的存在，这已被证明在相关序列转换任务 (Goyal et al., 2022) 中估计 LLM 输
$^{1}$wmt20-comet-qe-da
|
[
[
0.333251953125,
0.0955810546875,
0.6689453125,
0.0955810546875,
0.6689453125,
0.10986328125,
0.333251953125,
0.10986328125
],
[
0.172119140625,
0.1341552734375,
0.295654296875,
0.1341552734375,
0.295654296875,
0.1484375,
0.172119140625,
0.1484375
],
[
0.3359375,
0.133056640625,
0.459228515625,
0.1341552734375,
0.459228515625,
0.1484375,
0.3359375,
0.1478271484375
],
[
0.50146484375,
0.1341552734375,
0.59033203125,
0.1341552734375,
0.59033203125,
0.1484375,
0.50146484375,
0.1484375
],
[
0.63037109375,
0.1353759765625,
0.8310546875,
0.1353759765625,
0.8310546875,
0.1484375,
0.63037109375,
0.1484375
],
[
0.420654296875,
0.16748046875,
0.58447265625,
0.16748046875,
0.58447265625,
0.1817626953125,
0.420654296875,
0.1817626953125
],
[
0.4072265625,
0.182861328125,
0.59423828125,
0.18408203125,
0.59423828125,
0.198974609375,
0.4072265625,
0.19775390625
],
[
0.279541015625,
0.20068359375,
0.73291015625,
0.1995849609375,
0.73291015625,
0.21435546875,
0.279541015625,
0.215576171875
],
[
0.264404296875,
0.254638671875,
0.340087890625,
0.254638671875,
0.340087890625,
0.2666015625,
0.264404296875,
0.2666015625
],
[
0.51904296875,
0.25830078125,
0.57861328125,
0.25830078125,
0.57861328125,
0.2724609375,
0.51904296875,
0.2724609375
],
[
0.59033203125,
0.25830078125,
0.85986328125,
0.25830078125,
0.85986328125,
0.2724609375,
0.59033203125,
0.2724609375
],
[
0.145263671875,
0.281982421875,
0.459228515625,
0.281982421875,
0.459228515625,
0.295654296875,
0.145263671875,
0.295654296875
],
[
0.51904296875,
0.28076171875,
0.568359375,
0.28076171875,
0.568359375,
0.295654296875,
0.51904296875,
0.295654296875
],
[
0.591796875,
0.283203125,
0.60546875,
0.283203125,
0.60546875,
0.29345703125,
0.591796875,
0.29345703125
],
[
0.6572265625,
0.283203125,
0.67236328125,
0.283203125,
0.67236328125,
0.29443359375,
0.6572265625,
0.29443359375
],
[
0.7255859375,
0.283203125,
0.7841796875,
0.283203125,
0.7841796875,
0.29443359375,
0.7255859375,
0.29443359375
],
[
0.8388671875,
0.283935546875,
0.8671875,
0.283935546875,
0.8671875,
0.29638671875,
0.8388671875,
0.29638671875
],
[
0.1435546875,
0.295654296875,
0.45751953125,
0.29638671875,
0.45751953125,
0.311279296875,
0.1435546875,
0.31005859375
],
[
0.59423828125,
0.297607421875,
0.748046875,
0.297607421875,
0.748046875,
0.312255859375,
0.59423828125,
0.312255859375
],
[
0.14697265625,
0.311279296875,
0.45751953125,
0.311279296875,
0.45751953125,
0.32470703125,
0.14697265625,
0.32470703125
],
[
0.5224609375,
0.317626953125,
0.5751953125,
0.317626953125,
0.5751953125,
0.328857421875,
0.5224609375,
0.328857421875
],
[
0.587890625,
0.31640625,
0.771484375,
0.31640625,
0.771484375,
0.330078125,
0.587890625,
0.330078125
],
[
0.145263671875,
0.325927734375,
0.45751953125,
0.325927734375,
0.45751953125,
0.339111328125,
0.145263671875,
0.339111328125
],
[
0.145263671875,
0.33837890625,
0.459228515625,
0.33837890625,
0.459228515625,
0.3515625,
0.145263671875,
0.3515625
],
[
0.51123046875,
0.347412109375,
0.8818359375,
0.347412109375,
0.8818359375,
0.361083984375,
0.51123046875,
0.361083984375
],
[
0.1485595703125,
0.355224609375,
0.45751953125,
0.355224609375,
0.45751953125,
0.365234375,
0.1485595703125,
0.365234375
],
[
0.14697265625,
0.36767578125,
0.45751953125,
0.36767578125,
0.45751953125,
0.38134765625,
0.14697265625,
0.38134765625
],
[
0.51318359375,
0.3623046875,
0.87890625,
0.3623046875,
0.87890625,
0.3759765625,
0.51318359375,
0.3759765625
],
[
0.145263671875,
0.3818359375,
0.459228515625,
0.3818359375,
0.459228515625,
0.3955078125,
0.145263671875,
0.3955078125
],
[
0.51318359375,
0.377685546875,
0.880859375,
0.377685546875,
0.880859375,
0.388427734375,
0.51318359375,
0.388427734375
],
[
0.51123046875,
0.39013671875,
0.8447265625,
0.39013671875,
0.8447265625,
0.40380859375,
0.51123046875,
0.40380859375
],
[
0.1485595703125,
0.39794921875,
0.45751953125,
0.39794921875,
0.45751953125,
0.407958984375,
0.1485595703125,
0.407958984375
],
[
0.145263671875,
0.410400390625,
0.456787109375,
0.410400390625,
0.456787109375,
0.42333984375,
0.145263671875,
0.42333984375
],
[
0.51123046875,
0.410400390625,
0.87890625,
0.410400390625,
0.87890625,
0.42333984375,
0.51123046875,
0.42333984375
],
[
0.1485595703125,
0.42578125,
0.456787109375,
0.42578125,
0.456787109375,
0.435791015625,
0.1485595703125,
0.435791015625
],
[
0.5146484375,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.440673828125,
0.5146484375,
0.440673828125
],
[
0.145263671875,
0.438232421875,
0.45751953125,
0.438232421875,
0.45751953125,
0.451904296875,
0.145263671875,
0.451904296875
],
[
0.51318359375,
0.44482421875,
0.87744140625,
0.44482421875,
0.87744140625,
0.454833984375,
0.51318359375,
0.454833984375
],
[
0.14697265625,
0.453125,
0.45751953125,
0.453125,
0.45751953125,
0.466064453125,
0.14697265625,
0.466064453125
],
[
0.5146484375,
0.458984375,
0.880859375,
0.458984375,
0.880859375,
0.47265625,
0.5146484375,
0.47265625
],
[
0.145263671875,
0.46728515625,
0.456787109375,
0.46728515625,
0.456787109375,
0.48095703125,
0.145263671875,
0.48095703125
],
[
0.509765625,
0.473876953125,
0.87890625,
0.474365234375,
0.87890625,
0.4892578125,
0.509765625,
0.488037109375
],
[
0.1485595703125,
0.482666015625,
0.214111328125,
0.482666015625,
0.214111328125,
0.493408203125,
0.1485595703125,
0.493408203125
],
[
0.208251953125,
0.482666015625,
0.456787109375,
0.482666015625,
0.456787109375,
0.493408203125,
0.208251953125,
0.493408203125
],
[
0.14697265625,
0.495361328125,
0.45751953125,
0.495361328125,
0.45751953125,
0.5087890625,
0.14697265625,
0.5087890625
],
[
0.51318359375,
0.490478515625,
0.880859375,
0.4892578125,
0.880859375,
0.50439453125,
0.51318359375,
0.5048828125
],
[
0.14697265625,
0.51123046875,
0.456787109375,
0.51123046875,
0.456787109375,
0.521484375,
0.14697265625,
0.521484375
],
[
0.145263671875,
0.52392578125,
0.45751953125,
0.52392578125,
0.45751953125,
0.53759765625,
0.145263671875,
0.53759765625
],
[
0.53076171875,
0.52392578125,
0.8818359375,
0.52392578125,
0.8818359375,
0.53759765625,
0.53076171875,
0.53759765625
],
[
0.1485595703125,
0.5390625,
0.45751953125,
0.5390625,
0.45751953125,
0.5498046875,
0.1485595703125,
0.5498046875
],
[
0.51123046875,
0.5380859375,
0.880859375,
0.5390625,
0.880859375,
0.55419921875,
0.51123046875,
0.552734375
],
[
0.145263671875,
0.55029296875,
0.455078125,
0.55029296875,
0.455078125,
0.56396484375,
0.145263671875,
0.56396484375
],
[
0.51123046875,
0.5546875,
0.880859375,
0.5546875,
0.880859375,
0.568359375,
0.51123046875,
0.568359375
],
[
0.145263671875,
0.56640625,
0.45751953125,
0.56640625,
0.45751953125,
0.580078125,
0.145263671875,
0.580078125
],
[
0.51318359375,
0.57177734375,
0.880859375,
0.57177734375,
0.880859375,
0.5859375,
0.51318359375,
0.5859375
],
[
0.145263671875,
0.58056640625,
0.45751953125,
0.58056640625,
0.45751953125,
0.59423828125,
0.145263671875,
0.59423828125
],
[
0.5146484375,
0.58740234375,
0.880859375,
0.58740234375,
0.880859375,
0.6015625,
0.5146484375,
0.6015625
],
[
0.1485595703125,
0.5966796875,
0.45751953125,
0.5966796875,
0.45751953125,
0.60693359375,
0.1485595703125,
0.60693359375
],
[
0.145263671875,
0.60791015625,
0.32666015625,
0.60791015625,
0.32666015625,
0.62158203125,
0.145263671875,
0.62158203125
],
[
0.51318359375,
0.60400390625,
0.7900390625,
0.60400390625,
0.7900390625,
0.61767578125,
0.51318359375,
0.61767578125
],
[
0.51318359375,
0.6318359375,
0.5498046875,
0.6318359375,
0.5498046875,
0.6435546875,
0.51318359375,
0.6435546875
],
[
0.54248046875,
0.63134765625,
0.8564453125,
0.63134765625,
0.8564453125,
0.6455078125,
0.54248046875,
0.6455078125
],
[
0.1158447265625,
0.63818359375,
0.151123046875,
0.63818359375,
0.151123046875,
0.6494140625,
0.1158447265625,
0.6494140625
],
[
0.151123046875,
0.63818359375,
0.26123046875,
0.63818359375,
0.26123046875,
0.6494140625,
0.151123046875,
0.6494140625
],
[
0.51123046875,
0.65478515625,
0.87744140625,
0.65478515625,
0.87744140625,
0.66845703125,
0.51123046875,
0.66845703125
],
[
0.118408203125,
0.6630859375,
0.486083984375,
0.6630859375,
0.486083984375,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.1175537109375,
0.677734375,
0.484375,
0.677734375,
0.484375,
0.69140625,
0.1175537109375,
0.69140625
],
[
0.51318359375,
0.6728515625,
0.8818359375,
0.6728515625,
0.8818359375,
0.68603515625,
0.51318359375,
0.68603515625
],
[
0.51123046875,
0.6884765625,
0.880859375,
0.6884765625,
0.880859375,
0.70166015625,
0.51123046875,
0.70166015625
],
[
0.1175537109375,
0.69482421875,
0.484375,
0.693359375,
0.484375,
0.7080078125,
0.1175537109375,
0.708984375
],
[
0.51318359375,
0.705078125,
0.880859375,
0.705078125,
0.880859375,
0.71875,
0.51318359375,
0.71875
],
[
0.1175537109375,
0.71142578125,
0.486083984375,
0.71142578125,
0.486083984375,
0.724609375,
0.1175537109375,
0.724609375
],
[
0.12005615234375,
0.72802734375,
0.486083984375,
0.72802734375,
0.486083984375,
0.7412109375,
0.12005615234375,
0.7412109375
],
[
0.5146484375,
0.72265625,
0.7314453125,
0.72265625,
0.7314453125,
0.73291015625,
0.5146484375,
0.73291015625
],
[
0.1175537109375,
0.74365234375,
0.486083984375,
0.74365234375,
0.486083984375,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.515625,
0.74462890625,
0.8818359375,
0.74462890625,
0.8818359375,
0.7587890625,
0.515625,
0.7587890625
],
[
0.118408203125,
0.76025390625,
0.484375,
0.76025390625,
0.484375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51123046875,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.77490234375,
0.51123046875,
0.77490234375
],
[
0.118408203125,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.7890625,
0.118408203125,
0.7890625
],
[
0.51123046875,
0.77783203125,
0.67236328125,
0.77783203125,
0.67236328125,
0.79150390625,
0.51123046875,
0.79150390625
],
[
0.1175537109375,
0.79150390625,
0.483642578125,
0.79150390625,
0.483642578125,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51904296875,
0.8017578125,
0.8818359375,
0.8017578125,
0.8818359375,
0.81640625,
0.51904296875,
0.81640625
],
[
0.118408203125,
0.80810546875,
0.484375,
0.80810546875,
0.484375,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.51123046875,
0.81640625,
0.880859375,
0.81787109375,
0.880859375,
0.83203125,
0.51123046875,
0.83056640625
],
[
0.118408203125,
0.82373046875,
0.483642578125,
0.82373046875,
0.483642578125,
0.83740234375,
0.118408203125,
0.83740234375
],
[
0.51318359375,
0.83447265625,
0.87890625,
0.83447265625,
0.87890625,
0.84716796875,
0.51318359375,
0.84716796875
],
[
0.118408203125,
0.84130859375,
0.219970703125,
0.84130859375,
0.219970703125,
0.8525390625,
0.118408203125,
0.8525390625
],
[
0.5146484375,
0.85107421875,
0.87890625,
0.85107421875,
0.87890625,
0.8642578125,
0.5146484375,
0.8642578125
],
[
0.1351318359375,
0.85693359375,
0.486083984375,
0.85693359375,
0.486083984375,
0.87060546875,
0.1351318359375,
0.87060546875
],
[
0.515625,
0.8662109375,
0.87890625,
0.8662109375,
0.87890625,
0.8798828125,
0.515625,
0.8798828125
],
[
0.118408203125,
0.87255859375,
0.484375,
0.87255859375,
0.484375,
0.88720703125,
0.118408203125,
0.88720703125
],
[
0.51318359375,
0.8828125,
0.880859375,
0.8828125,
0.880859375,
0.896484375,
0.51318359375,
0.896484375
],
[
0.118408203125,
0.8896484375,
0.484375,
0.8896484375,
0.484375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.484375,
0.90478515625,
0.484375,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.5341796875,
0.90380859375,
0.6650390625,
0.90478515625,
0.6650390625,
0.91943359375,
0.5341796875,
0.91845703125
]
] |
[
"GPT是否产生较少的文字翻译？",
"Vikas Raunak",
"Arul Menezes",
"Matt Post",
"Hany Hassan Awadalla",
"Microsoft Azure Al",
"Redmond, Washington",
"( viraunak,arulm,mpost,hanyh } @microsoft.com",
"Abstract",
"source",
"He survived by the skin of his teeth",
"Large Language Models (LLMs) such as GPT-",
"NMT",
"I1",
"a",
"survécu",
"par",
"3 have emerged as general-purpose language",
"la peau de ses dents",
"models capable of addressing many natural",
"GPT-3",
"Il a survécu de justesse",
"language generation or understanding tasks.",
"On the task of Machine Translation (MT), mul-",
"Table l: An example where GPT-3 produces a more",
"tipleworkshaveinvestigatedfew-shotprompt-",
"ing mechanisms to elicit better translations",
"natural (non-literal) translation of an English idiom.",
"from LLMs. However, there has been rela-",
"When word-aligningthese sentences,the source word",
"skin remains unaligned for the GPT-3 translation.",
"tivelylittleinvestigationonhow suchtransla-",
"tions differ qualitatively from the translations",
"响（表1）。我们调查这些观察结果是否可以定",
"generatedbystandardNeuralMachineTrans",
"量验证。首先，我们使用基于单词对齐和单调",
"lation (NMT) models. In this work, we in",
"性的度量来量化LLM是否比NMT系统产生",
"vestigate these differences in terms of the lit-",
"更少的文字翻译，并将这些数字用于人类评估",
"eralness of translations produced by the two",
"（sE2）。接下来，我们专门研究成语，比较它",
"systems..",
"Using literalness measures involv",
"ing word alignment and monotonicity, we find",
"们在自然和合成数据设置下的翻译方式（sE3",
"thattranslationsoutofEnglish(E→X)from",
"GPTs tend to be less literal, while exhibiting",
"我们的调查重点是英语和德语，中文和俄语",
"similarorbetter scoresonMTqualitymet-",
"这三种类型不同的语言之间的翻译。我们的发",
"rics. We demonstrate that this finding is borne",
"现总结如下：（1）我们发现，当用英语翻译时，",
"out in human evaluations as well. We then",
"来自GPT系列LLM的两个LLM的翻译确实",
"show that these differences are especially pro",
"确实比NMT同行的翻译少，并且（2）在具有",
"nouncedwhentranslatingsentencesthatcon-",
"tain idiomatic expressions.",
"惯用表达的句子的情况下尤其如此",
"2",
" Quantifying Translation Literalness",
"1",
"Introduction",
"我们将最先进的NMT系统与最有能力的可公",
"尽管仅在语言建模目标上进行了培训，但在对",
"齐的并行数据(Briakouetal.,2023)上没有明确",
"开访问的GPT模型（在撰写本文时）进行了比",
"较，这些模型旨在捕捉翻译字面量的差异。我",
"的监督，但诸如GPT-3或PaLM(Brownetal.",
"们进行基于指标的自动评估和人工评估。我们",
"2020;Chowdhery et al.,2022）之类的LLM在少",
"量提示下实现了接近最先进的翻译性能（Vilar",
"在下面解释评估和实验细节",
"etal,2022;Hendyetal.,2023）。调查这些模型",
"Datasets我们使用官方的WMT21En-De，De-",
"的输出的工作已经注意到，当使用较旧的基于",
"En，En-Ru和Ru-En新闻翻译测试集来评估",
"表面的指标（如BLEU（Papinenietal.，2002a））",
"(Barrault et al., 2021)",
"时，性能的提高是不可见的，这通常显示出相",
"MeasuresofQuality我们使用COMET-QE",
"对于NMT系统的巨大损失。这就提出了一个",
"1（Reietal.，2020）作为质量估计（QE）度量",
"问题：这些LLM翻译在质量上与传统NMT系",
"(Fomichevaetal.,2020）来量化翻译的流畅性和",
"统有何不同？",
"充分性。使用QE作为度量的优点是，它排除",
"我们使用翻译文字的属性来探讨这个问题。",
"了任何参考偏差的存在，这已被证明在相关序",
"机器翻译系统长期以来一直被认为倾向于产生",
"列转换任务（Goyaletal.,2022）中估计LLM输",
"过于直译的（Dankersetal.，2022b），我们已经",
"观察到LLM似乎不太容易受到这个问题的影",
"'wmt20-comet-qe-da"
] |
{
"arxiv_id": "2305.16806",
"language": "zh",
"page_idx": 1
}
|
||
2305.16806-zh-4
|
2305.16806-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>$MT$ System</th>
<th>$\mathbf{C-QE}\uparrow$</th>
<th>$USW\downarrow$</th>
<th>$NM$</th>
</tr>
<tr>
<td>$\mathbf{MS-Translator}$</td>
<td>21.46</td>
<td>13.70</td>
<td>9.63</td>
</tr>
<tr>
<td>$\mathbf{WMT}^{\prime }21$ SOTA</td>
<td>23.25</td>
<td>14.47</td>
<td>10.21</td>
</tr>
<tr>
<td>text- davinci- 002</td>
<td>23.67</td>
<td>18.08</td>
<td>11.39</td>
</tr>
</tbody>
</table>
Table 5: Natural Idiomatic Sentences: Combined Re-
sults over MAGPIE, EPIE, PIE (5,712 sentences).
文字(释义)表达或通过目标语言中的等效习语。请注意，更大的非文字性并不意味着更好的比喻性。翻译中的非文学性可能是由与所需的比喻翻译不同的翻译变化产生的
3.1 Translation with Idiomatic Datasets
在本节中，我们量化了传统 NMT 系统和 GPT 模型之间带有成语的句子翻译的差异。没有任何以英语为中心的平行语料库专门用于带有成语的句子。因此，我们尝试使用带有成语的单语 (英语) 句子。在 2 节中使用相同的提示生成翻译。下面列举了带有自然惯用语的数据集：
$\cdot$MAGPIE (Haagsma et al., 2020) contains a set of sentences annotated with their idiomaticity, alongside a confidence score. We use the sentences pertaining to the news domain which are marked as idiomatic with cent percent annotator confidence (totalling 3,666 sentences).
$\bullet$ EPIE (Saxena and Paul, 2020) contains idioms and example sentences demonstrating their usage. We use the sentences available for static idioms (totalling 1,046 sentences).
$\bullet$ The PIE dataset (Zhou et al., 2021) contains idioms along with their usage. We randomly sample 1K sentences from the corpus.
Results 结果如表 5 所示。我们发现 textdavinci-002 比 WMT'21SOTA 系统产生更好的翻译质量，具有更多的未对齐单词以及更高的非单调性
Further Analysis 请注意，将翻译质量的提高直接归因于更好的习语翻译是具有挑战性的。此外，基于相似性的质量度量(例如 COMETQE 本身) 可能会惩罚非文字性，即使它们比表面级别的度量(例如 BLEU 或 ChrF(Papineni et al., 2002b; Popović, 2015) ) 不太可能这样做。因此，尽管自然的单语数据集提供了一个有用的测试平台来研究比喻的组成能力，但很难对系统之间的比喻组成进行明确的比较。因此，
<table>
<tbody>
<tr>
<th>$\operatorname{Expression}$</th>
<th>$\mathbf{C-QE}\uparrow$</th>
<th>$\overline{\mathrm{USW}\downarrow}$</th>
<th>$NM$</th>
</tr>
<tr>
<td>Random Phrases</td>
<td>-2.45</td>
<td>+1.62</td>
<td>+0.14</td>
</tr>
<tr>
<td>Named Entities</td>
<td>-1.50</td>
<td>+0.81</td>
<td>+0.39</td>
</tr>
<tr>
<td>Idioms</td>
<td>+5.90</td>
<td>+2.82</td>
<td>+1.95</td>
</tr>
</tbody>
</table>
Table 6: Synthetic sentences with Idioms vs Synthetic sentences containing other expressions: The difference between GPT (text-davinci-002) performance and NMT performance (Microsoft Translator) is reported.
我们还对合成数据进行了实验，其中我们明确控制了输入句子的细粒度属性。我们通过将输人句子中的大部分变化分配给合成数据生成中的某些组成表达式来做到这一点
## 3.2 Synthetic Experiments
对于我们的下一个实验，我们生成合成英语句子，每个句子包含特定类型的表达：(i)名称，(ii)随机描述性短语和(iii)习语。我们以零拍的方式提示 text-davinci-002,要求它生成一个句子，其中每个类型的不同实例化(详细信息在附录 B 中)。然后，我们使用不同的系统翻译这些句子，以调查系统之间和类型之间对我们的字面度量的相对影响。在每个对照实验中，我们将合成英语句子翻译成德语
Synthetic Dataset 1 如上所述，我们生成包含三种类型的表达式的句子，即命名实体(e。g。,杰西卡·阿尔巴), 随机描述性短语(e。g。, 盘子上的大蛋糕)和成语(e。g。,黑暗中的一枪)。表达式来源以及进一步的数据生成细节在附录 B 中给出。结果见表 6
<table>
<tbody>
<tr>
<th>Num Idioms</th>
<th>1</th>
<th>2</th>
<th>3</th>
<th>4</th>
</tr>
<tr>
<td>$USW$</td>
<td>17.58</td>
<td>18.39</td>
<td>18.28</td>
<td>18.99</td>
</tr>
</tbody>
</table>
Table 7: Synthetic sentences with multiple idioms (1- 4): Increasing the number of idioms increases the number of unaligned source words in text-davinci-002 translations.
Synthetic Dataset 2 我们生成包含多个成语(从1到4不等)的句子。附录 B 中提供了提示&示例。结果见表 7
Results 表6 显示，在成语的情况下，未对齐的源词的百分比最高，其次是随机描述性短语&命名实体。结果与以下假设一致：所探索的GPT 模型产生的$\mathbb{E}\to\mathbb{X}$文字翻译较少，因为句子中的命名实体或描述性短语会允许更多的文字翻译为可接受的，这与带有成语的句子不同。。Davinci-002 在翻译带有成语的句子时获得了更高的 COMET-QE 分数，但获得了更高比例的未对齐源词。同样，对于习语，非单调性得分的差异也要高得多。。这些结果提供了
|
[
[
0.54736328125,
0.0831298828125,
0.6337890625,
0.08489990234375,
0.6337890625,
0.09979248046875,
0.54736328125,
0.097412109375
],
[
0.66650390625,
0.0843505859375,
0.73876953125,
0.0843505859375,
0.73876953125,
0.09857177734375,
0.66650390625,
0.09857177734375
],
[
0.73779296875,
0.0843505859375,
0.8154296875,
0.0843505859375,
0.8154296875,
0.09857177734375,
0.73779296875,
0.09857177734375
],
[
0.8134765625,
0.08197021484375,
0.875,
0.0843505859375,
0.873046875,
0.09979248046875,
0.81201171875,
0.097412109375
],
[
0.15283203125,
0.09027099609375,
0.2418212890625,
0.09027099609375,
0.2418212890625,
0.1051025390625,
0.15283203125,
0.1051025390625
],
[
0.270263671875,
0.09027099609375,
0.344970703125,
0.09027099609375,
0.344970703125,
0.1051025390625,
0.270263671875,
0.1051025390625
],
[
0.34765625,
0.0914306640625,
0.414794921875,
0.0914306640625,
0.414794921875,
0.10272216796875,
0.34765625,
0.10272216796875
],
[
0.418212890625,
0.09027099609375,
0.474365234375,
0.09027099609375,
0.474365234375,
0.1051025390625,
0.418212890625,
0.1051025390625
],
[
0.52490234375,
0.09979248046875,
0.65576171875,
0.1015625,
0.65576171875,
0.11639404296875,
0.52490234375,
0.114013671875
],
[
0.67919921875,
0.1009521484375,
0.724609375,
0.1009521484375,
0.724609375,
0.11639404296875,
0.67919921875,
0.11639404296875
],
[
0.75146484375,
0.1009521484375,
0.80029296875,
0.1009521484375,
0.80029296875,
0.11639404296875,
0.75146484375,
0.11639404296875
],
[
0.8212890625,
0.1009521484375,
0.87158203125,
0.1009521484375,
0.87158203125,
0.11639404296875,
0.8212890625,
0.11639404296875
],
[
0.2822265625,
0.11224365234375,
0.328369140625,
0.11224365234375,
0.328369140625,
0.1270751953125,
0.2822265625,
0.1270751953125
],
[
0.358642578125,
0.11224365234375,
0.40478515625,
0.11224365234375,
0.40478515625,
0.1270751953125,
0.358642578125,
0.1270751953125
],
[
0.42822265625,
0.1134033203125,
0.47021484375,
0.1134033203125,
0.47021484375,
0.1270751953125,
0.42822265625,
0.1270751953125
],
[
0.75244140625,
0.11639404296875,
0.79833984375,
0.11639404296875,
0.79833984375,
0.1312255859375,
0.75244140625,
0.1312255859375
],
[
0.82275390625,
0.11639404296875,
0.86962890625,
0.11639404296875,
0.86962890625,
0.1318359375,
0.82275390625,
0.1318359375
],
[
0.5322265625,
0.1175537109375,
0.64990234375,
0.1175537109375,
0.64990234375,
0.1312255859375,
0.5322265625,
0.1312255859375
],
[
0.1435546875,
0.114013671875,
0.2509765625,
0.114013671875,
0.2509765625,
0.1258544921875,
0.1435546875,
0.1258544921875
],
[
0.68115234375,
0.1175537109375,
0.72314453125,
0.1175537109375,
0.72314453125,
0.1312255859375,
0.68115234375,
0.1312255859375
],
[
0.1351318359375,
0.127685546875,
0.2568359375,
0.12890625,
0.2568359375,
0.1436767578125,
0.1351318359375,
0.142578125
],
[
0.2822265625,
0.12890625,
0.328369140625,
0.12890625,
0.328369140625,
0.1436767578125,
0.2822265625,
0.1436767578125
],
[
0.359375,
0.127685546875,
0.40478515625,
0.127685546875,
0.40478515625,
0.1436767578125,
0.359375,
0.1436767578125
],
[
0.426513671875,
0.12890625,
0.470947265625,
0.12890625,
0.470947265625,
0.1436767578125,
0.426513671875,
0.1436767578125
],
[
0.56494140625,
0.133056640625,
0.62060546875,
0.133056640625,
0.62060546875,
0.1478271484375,
0.56494140625,
0.1478271484375
],
[
0.67822265625,
0.133056640625,
0.72705078125,
0.133056640625,
0.72705078125,
0.1478271484375,
0.67822265625,
0.1478271484375
],
[
0.75146484375,
0.133056640625,
0.80029296875,
0.133056640625,
0.80029296875,
0.1478271484375,
0.75146484375,
0.1478271484375
],
[
0.8212890625,
0.133056640625,
0.87158203125,
0.133056640625,
0.87158203125,
0.1478271484375,
0.8212890625,
0.1478271484375
],
[
0.2822265625,
0.144287109375,
0.328369140625,
0.144287109375,
0.328369140625,
0.1591796875,
0.2822265625,
0.1591796875
],
[
0.359375,
0.144287109375,
0.40478515625,
0.144287109375,
0.40478515625,
0.1591796875,
0.359375,
0.1591796875
],
[
0.424072265625,
0.144287109375,
0.47265625,
0.144287109375,
0.47265625,
0.1591796875,
0.424072265625,
0.1591796875
],
[
0.51123046875,
0.1602783203125,
0.8818359375,
0.1602783203125,
0.8818359375,
0.17333984375,
0.51123046875,
0.17333984375
],
[
0.133544921875,
0.1455078125,
0.258544921875,
0.1455078125,
0.258544921875,
0.1591796875,
0.133544921875,
0.1591796875
],
[
0.1158447265625,
0.177001953125,
0.486083984375,
0.177001953125,
0.486083984375,
0.190673828125,
0.1158447265625,
0.190673828125
],
[
0.5146484375,
0.177001953125,
0.880859375,
0.177001953125,
0.880859375,
0.18701171875,
0.5146484375,
0.18701171875
],
[
0.1158447265625,
0.191162109375,
0.453369140625,
0.1923828125,
0.453369140625,
0.207275390625,
0.1158447265625,
0.2060546875
],
[
0.51123046875,
0.188232421875,
0.880859375,
0.188232421875,
0.880859375,
0.201904296875,
0.51123046875,
0.201904296875
],
[
0.51123046875,
0.203125,
0.875,
0.203125,
0.875,
0.2161865234375,
0.51123046875,
0.2161865234375
],
[
0.12005615234375,
0.2203369140625,
0.484375,
0.2203369140625,
0.484375,
0.2340087890625,
0.12005615234375,
0.2340087890625
],
[
0.51123046875,
0.2257080078125,
0.87744140625,
0.2257080078125,
0.87744140625,
0.2392578125,
0.51123046875,
0.2392578125
],
[
0.118408203125,
0.2369384765625,
0.484375,
0.2369384765625,
0.484375,
0.25048828125,
0.118408203125,
0.25048828125
],
[
0.51123046875,
0.240478515625,
0.87890625,
0.2423095703125,
0.87890625,
0.255859375,
0.51123046875,
0.254150390625
],
[
0.118408203125,
0.2529296875,
0.48779296875,
0.2529296875,
0.48779296875,
0.2666015625,
0.118408203125,
0.2666015625
],
[
0.51123046875,
0.255859375,
0.87890625,
0.257080078125,
0.87890625,
0.271484375,
0.51123046875,
0.270751953125
],
[
0.1175537109375,
0.26953125,
0.393798828125,
0.26953125,
0.393798828125,
0.283203125,
0.1175537109375,
0.283203125
],
[
0.51123046875,
0.2724609375,
0.77099609375,
0.273681640625,
0.77099609375,
0.2880859375,
0.51123046875,
0.287353515625
],
[
0.1158447265625,
0.29443359375,
0.44580078125,
0.295654296875,
0.44580078125,
0.31005859375,
0.1158447265625,
0.308837890625
],
[
0.51318359375,
0.302734375,
0.5615234375,
0.302734375,
0.5615234375,
0.314208984375,
0.51318359375,
0.314208984375
],
[
0.55517578125,
0.3017578125,
0.73779296875,
0.3017578125,
0.73779296875,
0.3154296875,
0.55517578125,
0.3154296875
],
[
0.118408203125,
0.31640625,
0.486083984375,
0.31640625,
0.486083984375,
0.330078125,
0.118408203125,
0.330078125
],
[
0.51318359375,
0.32470703125,
0.880859375,
0.32470703125,
0.880859375,
0.33837890625,
0.51318359375,
0.33837890625
],
[
0.118408203125,
0.333251953125,
0.484375,
0.333251953125,
0.484375,
0.3466796875,
0.118408203125,
0.3466796875
],
[
0.51123046875,
0.339111328125,
0.87890625,
0.339111328125,
0.87890625,
0.352783203125,
0.51123046875,
0.352783203125
],
[
0.118408203125,
0.349853515625,
0.486083984375,
0.349853515625,
0.486083984375,
0.363525390625,
0.118408203125,
0.363525390625
],
[
0.51318359375,
0.355712890625,
0.880859375,
0.355712890625,
0.880859375,
0.369384765625,
0.51318359375,
0.369384765625
],
[
0.118408203125,
0.364013671875,
0.484375,
0.364013671875,
0.484375,
0.377685546875,
0.118408203125,
0.377685546875
],
[
0.51318359375,
0.372314453125,
0.880859375,
0.372314453125,
0.880859375,
0.385986328125,
0.51318359375,
0.385986328125
],
[
0.1175537109375,
0.38134765625,
0.484375,
0.38134765625,
0.484375,
0.394287109375,
0.1175537109375,
0.394287109375
],
[
0.5146484375,
0.388427734375,
0.8818359375,
0.388427734375,
0.8818359375,
0.402099609375,
0.5146484375,
0.402099609375
],
[
0.1175537109375,
0.396728515625,
0.486083984375,
0.3984375,
0.486083984375,
0.412109375,
0.1175537109375,
0.410400390625
],
[
0.51318359375,
0.40380859375,
0.87890625,
0.40380859375,
0.87890625,
0.41748046875,
0.51318359375,
0.41748046875
],
[
0.1175537109375,
0.41455078125,
0.145263671875,
0.41455078125,
0.145263671875,
0.427490234375,
0.1175537109375,
0.427490234375
],
[
0.51318359375,
0.419189453125,
0.87744140625,
0.418701171875,
0.87744140625,
0.432861328125,
0.51318359375,
0.43408203125
],
[
0.5146484375,
0.435791015625,
0.880859375,
0.435791015625,
0.880859375,
0.449462890625,
0.5146484375,
0.449462890625
],
[
0.1376953125,
0.4423828125,
0.48779296875,
0.4423828125,
0.48779296875,
0.4560546875,
0.1376953125,
0.4560546875
],
[
0.51318359375,
0.453125,
0.8427734375,
0.453125,
0.8427734375,
0.466064453125,
0.51318359375,
0.466064453125
],
[
0.151123046875,
0.458984375,
0.489501953125,
0.458984375,
0.489501953125,
0.47265625,
0.151123046875,
0.47265625
],
[
0.15283203125,
0.474365234375,
0.48779296875,
0.474365234375,
0.48779296875,
0.488037109375,
0.15283203125,
0.488037109375
],
[
0.51708984375,
0.478515625,
0.880859375,
0.478515625,
0.880859375,
0.4921875,
0.51708984375,
0.4921875
],
[
0.151123046875,
0.490478515625,
0.489501953125,
0.4892578125,
0.489501953125,
0.50439453125,
0.151123046875,
0.5048828125
],
[
0.515625,
0.496337890625,
0.880859375,
0.496337890625,
0.880859375,
0.50732421875,
0.515625,
0.50732421875
],
[
0.154541015625,
0.50732421875,
0.48779296875,
0.50732421875,
0.48779296875,
0.52099609375,
0.154541015625,
0.52099609375
],
[
0.51318359375,
0.51123046875,
0.8818359375,
0.51123046875,
0.8818359375,
0.525390625,
0.51318359375,
0.525390625
],
[
0.154541015625,
0.5224609375,
0.48779296875,
0.5224609375,
0.48779296875,
0.53759765625,
0.154541015625,
0.53759765625
],
[
0.51123046875,
0.525390625,
0.86376953125,
0.52685546875,
0.86376953125,
0.54150390625,
0.51123046875,
0.54052734375
],
[
0.154541015625,
0.5390625,
0.2822265625,
0.5390625,
0.2822265625,
0.552734375,
0.154541015625,
0.552734375
],
[
0.51123046875,
0.54345703125,
0.880859375,
0.5419921875,
0.880859375,
0.55712890625,
0.51123046875,
0.55810546875
],
[
0.51318359375,
0.56005859375,
0.76318359375,
0.56005859375,
0.76318359375,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.1376953125,
0.56640625,
0.489501953125,
0.56640625,
0.489501953125,
0.580078125,
0.1376953125,
0.580078125
],
[
0.151123046875,
0.58203125,
0.486083984375,
0.5830078125,
0.486083984375,
0.59716796875,
0.151123046875,
0.5966796875
],
[
0.53076171875,
0.5859375,
0.63037109375,
0.58837890625,
0.63037109375,
0.6025390625,
0.53076171875,
0.60107421875
],
[
0.71728515625,
0.59033203125,
0.73046875,
0.59033203125,
0.73046875,
0.60107421875,
0.71728515625,
0.60107421875
],
[
0.77685546875,
0.59033203125,
0.7900390625,
0.59033203125,
0.7900390625,
0.60107421875,
0.77685546875,
0.60107421875
],
[
0.82958984375,
0.58837890625,
0.8486328125,
0.58837890625,
0.8486328125,
0.6015625,
0.82958984375,
0.6015625
],
[
0.15283203125,
0.5986328125,
0.486083984375,
0.5986328125,
0.486083984375,
0.6123046875,
0.15283203125,
0.6123046875
],
[
0.6572265625,
0.59130859375,
0.6708984375,
0.59130859375,
0.6708984375,
0.599609375,
0.6572265625,
0.599609375
],
[
0.55908203125,
0.6025390625,
0.603515625,
0.6025390625,
0.603515625,
0.61767578125,
0.55908203125,
0.61767578125
],
[
0.63623046875,
0.60400390625,
0.86572265625,
0.60400390625,
0.86572265625,
0.61767578125,
0.63623046875,
0.61767578125
],
[
0.151123046875,
0.61328125,
0.476806640625,
0.6142578125,
0.476806640625,
0.62890625,
0.151123046875,
0.62744140625
],
[
0.509765625,
0.6298828125,
0.880859375,
0.6298828125,
0.880859375,
0.6435546875,
0.509765625,
0.6435546875
],
[
0.1376953125,
0.64111328125,
0.486083984375,
0.64111328125,
0.486083984375,
0.65478515625,
0.1376953125,
0.65478515625
],
[
0.5146484375,
0.64794921875,
0.87890625,
0.64794921875,
0.87890625,
0.65771484375,
0.5146484375,
0.65771484375
],
[
0.15283203125,
0.65771484375,
0.486083984375,
0.65771484375,
0.486083984375,
0.6728515625,
0.15283203125,
0.6728515625
],
[
0.51123046875,
0.6591796875,
0.8818359375,
0.6591796875,
0.8818359375,
0.6728515625,
0.51123046875,
0.6728515625
],
[
0.154541015625,
0.67333984375,
0.43408203125,
0.67333984375,
0.43408203125,
0.6884765625,
0.154541015625,
0.6884765625
],
[
0.51318359375,
0.6748046875,
0.56103515625,
0.6748046875,
0.56103515625,
0.68603515625,
0.51318359375,
0.68603515625
],
[
0.197265625,
0.703125,
0.489501953125,
0.703125,
0.489501953125,
0.71630859375,
0.197265625,
0.71630859375
],
[
0.515625,
0.69970703125,
0.880859375,
0.70068359375,
0.880859375,
0.71533203125,
0.515625,
0.71435546875
],
[
0.51123046875,
0.71728515625,
0.880859375,
0.71728515625,
0.880859375,
0.73095703125,
0.51123046875,
0.73095703125
],
[
0.124267578125,
0.70361328125,
0.204833984375,
0.70361328125,
0.204833984375,
0.71533203125,
0.124267578125,
0.71533203125
],
[
0.1175537109375,
0.71875,
0.486083984375,
0.71875,
0.486083984375,
0.732421875,
0.1175537109375,
0.732421875
],
[
0.118408203125,
0.7353515625,
0.484375,
0.7353515625,
0.484375,
0.7490234375,
0.118408203125,
0.7490234375
],
[
0.51318359375,
0.73291015625,
0.70361328125,
0.73291015625,
0.70361328125,
0.74658203125,
0.51318359375,
0.74658203125
],
[
0.118408203125,
0.751953125,
0.193115234375,
0.751953125,
0.193115234375,
0.76318359375,
0.118408203125,
0.76318359375
],
[
0.515625,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.515625,
0.77392578125
],
[
0.12005615234375,
0.77490234375,
0.484375,
0.775390625,
0.484375,
0.79052734375,
0.12005615234375,
0.7890625
],
[
0.51123046875,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.12005615234375,
0.7919921875,
0.484375,
0.7919921875,
0.484375,
0.8056640625,
0.12005615234375,
0.8056640625
],
[
0.51123046875,
0.79150390625,
0.87890625,
0.79150390625,
0.87890625,
0.8046875,
0.51123046875,
0.8046875
],
[
0.118408203125,
0.80810546875,
0.48779296875,
0.80712890625,
0.48779296875,
0.82080078125,
0.118408203125,
0.82177734375
],
[
0.51123046875,
0.80712890625,
0.880859375,
0.80810546875,
0.880859375,
0.822265625,
0.51123046875,
0.82177734375
],
[
0.1175537109375,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.1175537109375,
0.83837890625
],
[
0.5146484375,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.486083984375,
0.84130859375,
0.486083984375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.515625,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.515625,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.483642578125,
0.87255859375,
0.483642578125,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.51123046875,
0.87158203125,
0.880859375,
0.87255859375,
0.880859375,
0.88720703125,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.5146484375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.5146484375,
0.90283203125
],
[
0.12176513671875,
0.90673828125,
0.48779296875,
0.90673828125,
0.48779296875,
0.91748046875,
0.12176513671875,
0.91748046875
],
[
0.51318359375,
0.90478515625,
0.87890625,
0.90478515625,
0.87890625,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"Expression",
"C-QE个",
"USW",
"NMI",
"MT System",
"C-QE个",
"USW+",
"NM",
"RandomPhrases",
"-2.45",
"+1.62",
"+0.14",
"21.46",
"13.70",
"9.63",
"+0.81",
"+0.39",
"Named Entities",
"MS-Translator",
"-1.50",
"WMT'21 SOTA",
"23.25",
"14.47",
"10.21",
"Idioms",
"+5.90",
"+2.82",
"+1.95",
"23.67",
"18.08",
"11.39",
"Table 6: Synthetic sentences with Idioms vs Synthetic",
"text-davinci-002",
"Table 5: Natural Idiomatic Sentences: Combined Re-",
"sentences containing other expressions:The differ-",
"sults over MAGPIE, EPIE, PIE (5,712 sentences).",
"ence between GPT (text-davinci-002) performance and",
"NMT performance (Microsoft Translator) is reported.",
"文字（释义）表达或通过目标语言中的等效习",
"我们还对合成数据进行了实验，其中我们明确",
"语。请注意，更大的非文字性并不意味着更好",
"控制了输入句子的细粒度属性。我们通过将输",
"的比喻性。翻译中的非文学性可能是由与所需",
"入句子中的大部分变化分配给合成数据生成中",
"的比喻翻译不同的翻译变化产生的",
"的某些组成表达式来做到这一点",
"3.1Translation with Idiomatic Datasets",
"3.2",
"Synthetic Experiments",
"在本节中，我们量化了传统NMT系统和GPT",
"对于我们的下一个实验，我们生成合成英语",
"模型之间带有成语的句子翻译的差异。没有任",
"句子，每个句子包含特定类型的表达：（i）名",
"何以英语为中心的平行语料库专门用于带有成",
"称，（ii）随机描述性短语和（ii）习语。我们",
"语的句子。因此，我们尝试使用带有成语的单",
"以零拍的方式提示text-davinci-002，要求它生",
"语（英语）句子。在2节中使用相同的提示",
"成一个句子，其中每个类型的不同实例化（详",
"生成翻译。下面列举了带有自然惯用语的数据",
"细信息在附录B中）。然后，我们使用不同的",
"集：",
"系统翻译这些句子，以调查系统之间和类型之",
"间对我们的字面度量的相对影响。在每个对照",
": MAGPIE (Haagsma et al., 2020) contains",
"实验中，我们将合成英语句子翻译成德语",
"a set of sentences annotated with their id-",
"iomaticity, alongside a confidence score. We",
"SyntheticDataset1如上所述，我们生成包",
"use the sentences pertaining to the news do-",
"含三种类型的表达式的句子，即命名实体（e。",
"main which are marked as idiomatic with",
"g。，杰西卡阿尔巴），随机描述性短语（e。g。，",
"cent percent annotator confidence (totalling",
"盘子上的大蛋糕）和成语(e。g。，黑暗中的",
"3,666 sentences).",
"枪)。表达式来源以及进一步的数据生成细节",
"在附录B中给出。结果见表6",
": EPIE (Saxena and Paul, 2020) contains id-",
"ioms and example sentences demonstrating",
"Num Idioms",
"2",
"3",
"4",
"their usage. We use the sentences available",
"1",
"USW",
"17.5818.3918.2818.99",
"for static idioms (totalling 1,046 sentences).",
"Table 7: Synthetic sentences with multiple idioms (1)",
": The PIE dataset (Zhou et al., 2021) contains",
"4):Increasingthenumberofidiomsincreasesthenum",
"idioms along with their usage. We randomly",
"ber of unaligned source words in text-davinci-002 trans-",
"sample 1K sentences from the corpus.",
"lations",
"结果如表5所示。我们发现text-",
"SyntheticDataset2我们生成包含多个成语",
"（从1到4不等）的句子。附录B中提供了提",
"Results",
"davinci-002比WMT'21SOTA系统产生更好的",
"翻译质量，具有更多的未对齐单词以及更高的",
"示&示例。结果见表7",
"非单调性",
"Results表6显示，在成语的情况下，未对齐",
"FurtherAnalysis请注意，将翻译质量的提高",
"的源词的百分比最高，其次是随机描述性短语",
"直接归因于更好的习语翻译是具有挑战性的。",
"&命名实体。结果与以下假设一致：所探索的",
"此外，基于相似性的质量度量（例如COMET-",
"GPT模型产生的E→X文字翻译较少，因为",
"QE本身）可能会惩罚非文字性，即使它们比表",
"句子中的命名实体或描述性短语会允许更多的",
"面级别的度量（例如BLEU或ChrF（Papineni",
"文字翻译为可接受的，这与带有成语的句子不",
"etal.，2002b;Popovic，2015））不太可能这样做。",
"同。。Davinci-002在翻译带有成语的句子时获",
"因此，尽管自然的单语数据集提供了一个有用",
"得了更高的COMET-QE分数，但获得了更高",
"的测试平台来研究比喻的组成能力，但很难对",
"比例的未对齐源词。同样，对于习语，非单调",
"系统之间的比喻组成进行明确的比较。。因此，",
"性得分的差异也要高得多。。这些结果提供了"
] |
{
"arxiv_id": "2305.16806",
"language": "zh",
"page_idx": 4
}
|
||
1612.01928-zh-0
|
1612.01928-zh-0.jpg
|
# 用于嘈杂语音识别的不变表示
Philémon Brakel
$MILA$
Université de Montréal
${\mathbf{D}}{\mathbf{mitriy}}{\mathbf{Serdyuk}}^{*}$
MILA, Université de Montréal
Montr$\' {e}$al, QCH3T1J4
serdyuk@iro.umontreal.ca
## Bhuvana Ramabhadran IBM Watson, Yorktown Heights, NY
Kartik Audhkhasi
IBM Watson,
Yorktown Heights, NY
Samuel Thomas
IBM Watson, Yorktown Heights, NY
Yoshua Bengio
MILA, CIFAR Fellow, Université de Montréal
# Abstract
Modern automatic speech recognition (ASR) systems need to be robust under acoustic variability arising from environmental, speaker, channel, and recording conditions. Ensuring such robustness to variability is a challenge in modern day neural network-based ASR systems, especially when all types of variability are not seen during training. We attempt to address this problem by encouraging the neural network acoustic model to learn invariant feature representations. We use ideas from recent research on image generation using Generative Adversarial Networks and domain adaptation ideas extending adversarial gradient-based training. A recent work from Ganin et al. proposes to use adversarial training for image domain adaptation by using an intermediate representation from the main target classification network to deteriorate the domain classifier performance through a separate neural network. Our work focuses on investigating neural architectures which produce representations invariant to noise conditions for ASR. We evaluate the proposed architecture on the Aurora-4 task, a popular benchmark for noise robust ASR. We show that our method generalizes better than the standard multicondition training especially when only a few noise categories are seen during training.
### 1 Introduction
自动语音识别 (ASR) 最具挑战性的方面之一是训练和测试声学条件之间的不匹配。在测试过程中，系统可能会遇到新的录制条件、麦克风类型、扬声器、口音和背景噪声类型。此外， 即使在训练期间看到测试场景，其统计信息也可能存在显着差异。因此，开发对看不见的声学条件不变的 ASR 系统很重要。已经提出了几种基于模型和特征的自适应方法，例如最大似然线性回归(MLLR),基于特征的 MLLR 和 iVectors$\boxed{\mathrm{F}}($Saon $et\textit{al. , 2013) 来 处 理 说 话 }$ 人的可变性；和噪声自适应训练 (NAT; ?) 和矢量泰勒级数 (?, VTS;) 随着用于 ASR (Hinton $etal.$ ,2012; Seide $etal.$ ,2011; Sainath $etal.$ ,2011)的深度神经网络(DNN) 声学模型的日益成功，人们提出了端到端系统$\overline{\mathrm{E}}($Miao et al. ,2015; Sainath et al. ,2015)来对单个网络中的声学条件进行建模。这使我们能够利用网络学习高度非线性特征变换的能力，在构建促进噪声不变表示学习的训练目标函数时具有更大的灵活性。这项工作的主要思想是迫使声学模型学习对噪声条件不变的表示，而不是明确使用噪声鲁棒的声学特征(第EB 节)。这种类型的噪声不变训练仅在训练期间需要噪声条件标签。它与生成对抗网络(GAN)的思
$^{*}$Dmitriy Serdyuk performed the work during an internship at IBM Watson.
30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain
|
[
[
0.31787109375,
0.125,
0.6845703125,
0.126953125,
0.6845703125,
0.1514892578125,
0.31787109375,
0.149658203125
],
[
0.250732421875,
0.203857421875,
0.375732421875,
0.203857421875,
0.375732421875,
0.2171630859375,
0.250732421875,
0.2171630859375
],
[
0.463134765625,
0.203857421875,
0.59228515625,
0.203857421875,
0.59228515625,
0.2171630859375,
0.463134765625,
0.2171630859375
],
[
0.65673828125,
0.203857421875,
0.77685546875,
0.203857421875,
0.77685546875,
0.2171630859375,
0.65673828125,
0.2171630859375
],
[
0.48046875,
0.21533203125,
0.57177734375,
0.2171630859375,
0.57177734375,
0.231689453125,
0.48046875,
0.2298583984375
],
[
0.69384765625,
0.216552734375,
0.740234375,
0.216552734375,
0.740234375,
0.2310791015625,
0.69384765625,
0.2310791015625
],
[
0.2320556640625,
0.2310791015625,
0.3857421875,
0.2310791015625,
0.3857421875,
0.2442626953125,
0.2320556640625,
0.2442626953125
],
[
0.20751953125,
0.2171630859375,
0.410888671875,
0.2171630859375,
0.410888671875,
0.2310791015625,
0.20751953125,
0.2310791015625
],
[
0.450927734375,
0.2310791015625,
0.60302734375,
0.2310791015625,
0.60302734375,
0.24560546875,
0.450927734375,
0.24560546875
],
[
0.640625,
0.2310791015625,
0.7939453125,
0.2310791015625,
0.7939453125,
0.2442626953125,
0.640625,
0.2442626953125
],
[
0.205078125,
0.24560546875,
0.41259765625,
0.24560546875,
0.41259765625,
0.2587890625,
0.205078125,
0.2587890625
],
[
0.2646484375,
0.27978515625,
0.43310546875,
0.27978515625,
0.43310546875,
0.294189453125,
0.2646484375,
0.294189453125
],
[
0.59326171875,
0.27978515625,
0.70751953125,
0.27978515625,
0.70751953125,
0.294189453125,
0.59326171875,
0.294189453125
],
[
0.2254638671875,
0.294189453125,
0.468017578125,
0.294189453125,
0.468017578125,
0.30810546875,
0.2254638671875,
0.30810546875
],
[
0.52783203125,
0.294189453125,
0.771484375,
0.294189453125,
0.771484375,
0.30810546875,
0.52783203125,
0.30810546875
],
[
0.44677734375,
0.32763671875,
0.55126953125,
0.328857421875,
0.55126953125,
0.343505859375,
0.44677734375,
0.34228515625
],
[
0.346435546875,
0.343505859375,
0.65283203125,
0.343505859375,
0.65283203125,
0.356689453125,
0.346435546875,
0.356689453125
],
[
0.463134765625,
0.39453125,
0.53759765625,
0.39453125,
0.53759765625,
0.4091796875,
0.463134765625,
0.4091796875
],
[
0.233642578125,
0.4248046875,
0.765625,
0.4248046875,
0.765625,
0.438232421875,
0.233642578125,
0.438232421875
],
[
0.233642578125,
0.438232421875,
0.765625,
0.438232421875,
0.765625,
0.45263671875,
0.233642578125,
0.45263671875
],
[
0.233642578125,
0.45263671875,
0.7646484375,
0.45263671875,
0.7646484375,
0.466552734375,
0.233642578125,
0.466552734375
],
[
0.233642578125,
0.46728515625,
0.765625,
0.46728515625,
0.765625,
0.48095703125,
0.233642578125,
0.48095703125
],
[
0.233642578125,
0.479736328125,
0.765625,
0.479736328125,
0.765625,
0.49365234375,
0.233642578125,
0.49365234375
],
[
0.233642578125,
0.494384765625,
0.765625,
0.494384765625,
0.765625,
0.50830078125,
0.233642578125,
0.50830078125
],
[
0.233642578125,
0.50830078125,
0.76708984375,
0.50830078125,
0.76708984375,
0.521484375,
0.233642578125,
0.521484375
],
[
0.2320556640625,
0.52099609375,
0.7646484375,
0.521484375,
0.7646484375,
0.5361328125,
0.2320556640625,
0.53515625
],
[
0.2320556640625,
0.5341796875,
0.765625,
0.53515625,
0.765625,
0.5498046875,
0.2320556640625,
0.548828125
],
[
0.233642578125,
0.5498046875,
0.76708984375,
0.5498046875,
0.76708984375,
0.56298828125,
0.233642578125,
0.56298828125
],
[
0.233642578125,
0.56298828125,
0.76708984375,
0.56298828125,
0.76708984375,
0.5771484375,
0.233642578125,
0.5771484375
],
[
0.2320556640625,
0.57568359375,
0.7646484375,
0.57568359375,
0.7646484375,
0.5888671875,
0.2320556640625,
0.5888671875
],
[
0.233642578125,
0.59033203125,
0.765625,
0.59033203125,
0.765625,
0.60400390625,
0.233642578125,
0.60400390625
],
[
0.233642578125,
0.60498046875,
0.765625,
0.60498046875,
0.765625,
0.61865234375,
0.233642578125,
0.61865234375
],
[
0.2344970703125,
0.619140625,
0.7646484375,
0.619140625,
0.7646484375,
0.6298828125,
0.2344970703125,
0.6298828125
],
[
0.233642578125,
0.6318359375,
0.7646484375,
0.6318359375,
0.7646484375,
0.64599609375,
0.233642578125,
0.64599609375
],
[
0.233642578125,
0.64306640625,
0.2900390625,
0.646484375,
0.289306640625,
0.6611328125,
0.2320556640625,
0.65771484375
],
[
0.1748046875,
0.68505859375,
0.2091064453125,
0.68505859375,
0.2091064453125,
0.6962890625,
0.1748046875,
0.6962890625
],
[
0.2017822265625,
0.68505859375,
0.31787109375,
0.68505859375,
0.31787109375,
0.69873046875,
0.2017822265625,
0.69873046875
],
[
0.17724609375,
0.7138671875,
0.82421875,
0.7138671875,
0.82421875,
0.72802734375,
0.17724609375,
0.72802734375
],
[
0.1748046875,
0.72802734375,
0.82177734375,
0.72802734375,
0.82177734375,
0.7412109375,
0.1748046875,
0.7412109375
],
[
0.176513671875,
0.7421875,
0.82275390625,
0.7421875,
0.82275390625,
0.755859375,
0.176513671875,
0.755859375
],
[
0.1732177734375,
0.7548828125,
0.82275390625,
0.755859375,
0.82275390625,
0.77099609375,
0.1732177734375,
0.76953125
],
[
0.176513671875,
0.77099609375,
0.82275390625,
0.77099609375,
0.82275390625,
0.7841796875,
0.176513671875,
0.7841796875
],
[
0.17724609375,
0.7841796875,
0.82275390625,
0.7841796875,
0.82275390625,
0.79736328125,
0.17724609375,
0.79736328125
],
[
0.1748046875,
0.79736328125,
0.82177734375,
0.79736328125,
0.82177734375,
0.81103515625,
0.1748046875,
0.81103515625
],
[
0.176513671875,
0.81103515625,
0.82275390625,
0.81103515625,
0.82275390625,
0.82470703125,
0.176513671875,
0.82470703125
],
[
0.176513671875,
0.82568359375,
0.82275390625,
0.82568359375,
0.82275390625,
0.8388671875,
0.176513671875,
0.8388671875
],
[
0.176513671875,
0.8388671875,
0.82421875,
0.8388671875,
0.82421875,
0.85302734375,
0.176513671875,
0.85302734375
],
[
0.1748046875,
0.85107421875,
0.82275390625,
0.8515625,
0.82275390625,
0.8662109375,
0.1748046875,
0.86572265625
],
[
0.176513671875,
0.8662109375,
0.82275390625,
0.8662109375,
0.82275390625,
0.8798828125,
0.176513671875,
0.8798828125
],
[
0.1968994140625,
0.8896484375,
0.63818359375,
0.8896484375,
0.63818359375,
0.90283203125,
0.1968994140625,
0.90283203125
],
[
0.176513671875,
0.92529296875,
0.7158203125,
0.92529296875,
0.7158203125,
0.93603515625,
0.176513671875,
0.93603515625
]
] |
[
"用于噜杂语音识别的不变表示",
"Dmitriy Serdyuk*",
"Kartik Audhkhasi",
"Philemon Brakel",
"IBM Watson,",
"MILA",
"Montréal, QC H3T 1J4",
"MILA, Université de Montréal",
"Yorktown Heights, NY",
"Université de Montreal",
"serdyuk@iro.umontreal.ca",
"Bhuvana Ramabhadran",
"Samuel Thomas",
"IBM Watson, Yorktown Heights, NY",
"IBM Watson, Yorktown Heights, NY",
"Yoshua Bengio",
"MILA, CIFAR Fellow, Université de Montréal",
"Abstract",
"Modern automatic speech recognition (ASR) systems need to be robust under",
"acoustic variability arising from environmental, speaker, channel, and recording",
"conditions. Ensuring such robustness to variability is a challenge in modern day",
"neural network-based ASR systems, especially when all types of variability are",
"not seen during training. We attempt to address this problem by encouraging the",
"neural network acoustic model to learn invariant feature representations. We use",
"ideas from recent research on image generation using Generative Adversarial Net-",
"works and domain adaptation ideas extending adversarial gradient-based training.",
"A recent work from Ganin et al. proposes to use adversarial training for image",
"domain adaptation by using an intermediate representation from the main target",
"classification network to deteriorate the domain classifier performance through a",
"separate neural network. Our work focuses on investigating neural architectures",
"which produce representations invariant to noise conditions for ASR. We evaluate",
"the proposed architecture on the Aurora-4 task, a popular benchmark for noise",
"robustASR.We showthat our methodgeneralizes betterthanthe standard multi-",
"condition training especially when only a few noise categories are seen during",
"training.",
"1",
"Introduction",
"自动语音识别（ASR）最具挑战性的方面之一是训练和测试声学条件之间的不匹配。在测试",
"过程中，系统可能会遇到新的录制条件、麦克风类型、扬声器、口音和背景噪声类型。此外",
"即使在训练期间看到测试场景，其统计信息也可能存在显着差异。因此，开发对看不见的",
"声学条件不变的ASR系统很重要。已经提出了几种基于模型和特征的自适应方法，例如最",
"大似然线性回归（MLLR），基于特征的MLLR和iVectorsE(Saonetal.，2013）来处理说话",
"人的可变性；和噪声自适应训练（NAT？）和矢量泰勒级数（?，VTS：）随着用于ASR（Hinton",
"etal.，2012;Seideetal.，2011;Sainathetal.，2011)的深度神经网络（DNN）声学模型的日",
"益成功，人们提出了端到端系统E(Miaoetal.，2015；Sainathetal.，2015）来对单个网络中",
"的声学条件进行建模。这使我们能够利用网络学习高度非线性特征变换的能力，在构建促",
"进噪声不变表示学习的训练目标函数时具有更大的灵活性。这项工作的主要思想是迫使声",
"学模型学习对噪声条件不变的表示，而不是明确使用噪声鲁棒的声学特征（第E3节）。这",
"种类型的噪声不变训练仅在训练期间需要噪声条件标签。它与生成对抗网络（GAN）的思",
"*Dmitriy Serdyuk performed the work during an internship at IBM Watson.",
"3OthConferenceonNeuralInformationProcessingSystems(NIPS2016),Barcelona,Spain"
] |
{
"arxiv_id": "1612.01928",
"language": "zh",
"page_idx": 0
}
|
||
1612.01928-zh-1
|
1612.01928-zh-1.jpg
|
想以及分别在$\underline{\mathrm{E}}$Goodfellow et al.(2014)和$\overline{\mathrm{E}}$Ganin&Lempitsky (2014)中提出的梯度反向方法有关 (第$\mathbb{E}$2节)。我们在 4节中介绍了 Aurora-4 语音识别任务的结果，并在 5节中总结了我们的发现。
## 2 $\textbf{Related Work}$
生成对抗网络由两个网络组成：生成器和鉴别器。生成器网络$G$具有随机生成的特征向量的输人，并被要求生成一个样本，即。g。图像，类似于训练集中的图像。鉴别器网络$D$可以接收来自生成器$G$的生成图像或来自训练集的图像。它的任务是区分“假”生成的图像和从数据集中获取的“真实”图像。因此，鉴别器只是一个具有 S 形输出层的分类器网络， 可以使用梯度反向传播进行训练。此梯度可以进一步传播到发电机网络
GAN 设置中的两个网络相互竞争：生成器试图欺骗鉴别器网络，而鉴别器则尽力识别是否
存在欺骗，类似于对抗性博弈论设置。形式上，GAN 训练的目标函数为
$$\min_{G}\max_{D}V(D,G)=\mathbb{E}_{\boldsymbol{x}\sim p_{\mathrm{data}}(\boldsymbol{x})}[\log D(\boldsymbol{x})]+\mathbb{E}_{\boldsymbol{z}\sim p_{\boldsymbol{z}}(\boldsymbol{z})}[\log(1-D(G(\boldsymbol{z})))].$$
鉴别器$D$上的最大化形成了通常的交叉熵目标，相对于$D$的参数计算梯度。使用通过第二项传播的梯度来最小化$G$的参数。$G$上的最小化使其能够产生$D$分类器作为训练分类器的示例
在 Radford et al. (2015) 中提出了优化 GAN 的几个实用指南，并在 Salimans et al. (2016) 中进行了进一步的探索。Ganin& Lempitsky (2014) 的先前工作提出了一种训练可适应新域的网络的方法。训练数据由用感兴趣的类和单独的域(图像背景)标签标记的图像组成。该网络具有类似$Y$的结构：图像被馈送到第一个网络，该网络产生一个隐藏的表示 $h$。然后将此表示$h$输入到两个单独的网络：域分类器网络(D)和目标分类器网络(R)。训练的目标是学习对域标签不变的隐藏表示，并在目标分类任务上表现良好，以便域信息在测试时不会干扰目标分类器。与 GAN 目标类似，该目标迫使生成分布接近数据分布，梯度反向方法使域分布彼此相似
网络训练有三个目标：隐藏表示$h$应该对目标分类器有帮助，对域分类器有害，并且域分类
器应该具有良好的分类精度。更正式地说，作者将损失函数定义为
$$L=L_1(\hat{y},y;\theta_R,\theta_E)+\alpha L_2(\hat{d},d;\theta_D)-\beta L_3(\hat{d},d;\theta_E),$$
, (1)
,其中$y$是基本真值类，$d$是域标签，相应的帽子变量是网络预测，而$\theta_E,\theta_R$和$\theta_D$分别是
编码器，识别器和域分类器网络的参数子集。超参数$\alpha$和$\beta$表示损失函数项的相对影响
Yu et al. (2013) 中讨论了神经网络产生的表示对内部降噪的影响，这项工作为 Aurora-4 数据集的实验设定了基线。最近，在[E]Shunohara(2016)中，以对抗性的方式对多层 S 形网络进行了内部转录任务的训练，该任务被噪声破坏。
3 $\textbf{Invariant Representations for Speech Recognitior}$
大多数 ASR 系统是 DNN-HMM 混合系统。上下文相关(CD)HMM 状态 (声学模型) 是感兴趣的类标签。录音条件、说话者身份或性别代表 GAN 中的域。任务是使 HMM 状态分类器网络的隐藏层表示相对于 这 些 域 不 变。WehypothesizethatthisadversarialmethodoftraininghelpstheHMMstateclassifnertogeneralizebettertounseendomainconditionsandrequiresonlyasmalladditionalamountofsupervision,i。
e。域标签
图$\boxed{\mathrm{E}}$la 描绘了模型，该模型与梯度反向方法的模型相同。它是一个经过训练的前馈神经网络，用于预测 CDHMM 状态，其分支预测域 (噪声条件)。此分支在测试阶段被丢弃。在我们的实验中，我们使用噪声条件作为域标签，将所有噪声类型合并到一个标签中，并作为另一个标签进行清洁。我们的训练损失函数是方程 1 , 其中$L_3$设置为$d\log(1-\hat{d})+(1-d)\log(\hat{d})$ ,用于训练期间的稳定性。$L_3$项最大限度地提高了错误域分类的概率，而梯度反向则最大限度地降低了正确分类的概率。术语$L_1$和$L_2$是规则的交叉熵，使用相应的参数$\theta_E$和$\theta_D$ 将其最小化。为简单起见，我们只使用一个超参数——第三项的权重。
2
|
[
[
0.17724609375,
0.0947265625,
0.8203125,
0.0947265625,
0.8203125,
0.10479736328125,
0.17724609375,
0.10479736328125
],
[
0.176513671875,
0.106689453125,
0.82275390625,
0.106689453125,
0.82275390625,
0.12060546875,
0.176513671875,
0.12060546875
],
[
0.178955078125,
0.12249755859375,
0.281005859375,
0.12249755859375,
0.281005859375,
0.1337890625,
0.178955078125,
0.1337890625
],
[
0.1732177734375,
0.1622314453125,
0.326904296875,
0.1622314453125,
0.326904296875,
0.175537109375,
0.1732177734375,
0.175537109375
],
[
0.17724609375,
0.1976318359375,
0.8203125,
0.1976318359375,
0.8203125,
0.2083740234375,
0.17724609375,
0.2083740234375
],
[
0.17724609375,
0.210205078125,
0.82177734375,
0.210205078125,
0.82177734375,
0.2235107421875,
0.17724609375,
0.2235107421875
],
[
0.17724609375,
0.2235107421875,
0.82177734375,
0.2235107421875,
0.82177734375,
0.2374267578125,
0.17724609375,
0.2374267578125
],
[
0.17724609375,
0.238037109375,
0.82275390625,
0.238037109375,
0.82275390625,
0.251953125,
0.17724609375,
0.251953125
],
[
0.178955078125,
0.25244140625,
0.6806640625,
0.25244140625,
0.6806640625,
0.26318359375,
0.178955078125,
0.26318359375
],
[
0.176513671875,
0.271484375,
0.82275390625,
0.271484375,
0.82275390625,
0.285400390625,
0.176513671875,
0.285400390625
],
[
0.17724609375,
0.287353515625,
0.685546875,
0.287353515625,
0.685546875,
0.29736328125,
0.17724609375,
0.29736328125
],
[
0.25,
0.314453125,
0.7490234375,
0.314453125,
0.7490234375,
0.328857421875,
0.25,
0.328857421875
],
[
0.176513671875,
0.349853515625,
0.82275390625,
0.349853515625,
0.82275390625,
0.363037109375,
0.176513671875,
0.363037109375
],
[
0.176513671875,
0.363037109375,
0.82275390625,
0.363037109375,
0.82275390625,
0.376953125,
0.176513671875,
0.376953125
],
[
0.176513671875,
0.37744140625,
0.2091064453125,
0.37744140625,
0.2091064453125,
0.389404296875,
0.176513671875,
0.389404296875
],
[
0.176513671875,
0.397705078125,
0.82177734375,
0.397705078125,
0.82177734375,
0.410888671875,
0.176513671875,
0.410888671875
],
[
0.17724609375,
0.410888671875,
0.82275390625,
0.410888671875,
0.82275390625,
0.4248046875,
0.17724609375,
0.4248046875
],
[
0.176513671875,
0.425537109375,
0.82275390625,
0.425537109375,
0.82275390625,
0.439453125,
0.176513671875,
0.439453125
],
[
0.17724609375,
0.439453125,
0.82177734375,
0.439453125,
0.82177734375,
0.45263671875,
0.17724609375,
0.45263671875
],
[
0.1748046875,
0.45068359375,
0.82177734375,
0.451904296875,
0.82177734375,
0.46533203125,
0.1748046875,
0.464111328125
],
[
0.176513671875,
0.466552734375,
0.82177734375,
0.466552734375,
0.82177734375,
0.479736328125,
0.176513671875,
0.479736328125
],
[
0.17724609375,
0.48095703125,
0.82177734375,
0.48095703125,
0.82177734375,
0.494384765625,
0.17724609375,
0.494384765625
],
[
0.17724609375,
0.49560546875,
0.324462890625,
0.49560546875,
0.324462890625,
0.505859375,
0.17724609375,
0.505859375
],
[
0.17724609375,
0.5146484375,
0.82275390625,
0.5146484375,
0.82275390625,
0.52783203125,
0.17724609375,
0.52783203125
],
[
0.17724609375,
0.52978515625,
0.646484375,
0.52978515625,
0.646484375,
0.54052734375,
0.17724609375,
0.54052734375
],
[
0.314453125,
0.55810546875,
0.6845703125,
0.556640625,
0.6845703125,
0.5712890625,
0.314453125,
0.57275390625
],
[
0.8037109375,
0.5595703125,
0.826171875,
0.5595703125,
0.826171875,
0.57275390625,
0.8037109375,
0.57275390625
],
[
0.191162109375,
0.5869140625,
0.82275390625,
0.5869140625,
0.82275390625,
0.60107421875,
0.191162109375,
0.60107421875
],
[
0.1748046875,
0.599609375,
0.7998046875,
0.6015625,
0.7998046875,
0.61572265625,
0.1748046875,
0.61279296875
],
[
0.1846923828125,
0.6220703125,
0.82177734375,
0.6220703125,
0.82177734375,
0.63525390625,
0.1846923828125,
0.63525390625
],
[
0.176513671875,
0.63525390625,
0.82275390625,
0.63525390625,
0.82275390625,
0.6484375,
0.176513671875,
0.6484375
],
[
0.17724609375,
0.6494140625,
0.52685546875,
0.6494140625,
0.52685546875,
0.66015625,
0.17724609375,
0.66015625
],
[
0.17236328125,
0.689453125,
0.62353515625,
0.68994140625,
0.62353515625,
0.705078125,
0.17236328125,
0.7041015625
],
[
0.176513671875,
0.724609375,
0.82275390625,
0.724609375,
0.82275390625,
0.73779296875,
0.176513671875,
0.73779296875
],
[
0.176513671875,
0.73779296875,
0.82275390625,
0.73779296875,
0.82275390625,
0.751953125,
0.176513671875,
0.751953125
],
[
0.176513671875,
0.751953125,
0.82275390625,
0.751953125,
0.82275390625,
0.76513671875,
0.176513671875,
0.76513671875
],
[
0.1748046875,
0.7646484375,
0.58251953125,
0.7666015625,
0.58251953125,
0.77978515625,
0.1748046875,
0.77783203125
],
[
0.57666015625,
0.76806640625,
0.82177734375,
0.76806640625,
0.82177734375,
0.77880859375,
0.57666015625,
0.77880859375
],
[
0.176513671875,
0.77978515625,
0.91845703125,
0.77978515625,
0.91845703125,
0.79443359375,
0.176513671875,
0.79443359375
],
[
0.176513671875,
0.79443359375,
0.2484130859375,
0.79443359375,
0.2484130859375,
0.80615234375,
0.176513671875,
0.80615234375
],
[
0.17724609375,
0.814453125,
0.82177734375,
0.814453125,
0.82177734375,
0.82763671875,
0.17724609375,
0.82763671875
],
[
0.1748046875,
0.8271484375,
0.82177734375,
0.82763671875,
0.82177734375,
0.84228515625,
0.1748046875,
0.84130859375
],
[
0.17724609375,
0.84326171875,
0.80810546875,
0.84326171875,
0.80810546875,
0.85400390625,
0.17724609375,
0.85400390625
],
[
0.176513671875,
0.85791015625,
0.82275390625,
0.85791015625,
0.82275390625,
0.8720703125,
0.176513671875,
0.8720703125
],
[
0.193603515625,
0.87255859375,
0.8203125,
0.87255859375,
0.8203125,
0.88330078125,
0.193603515625,
0.88330078125
],
[
0.176513671875,
0.8837890625,
0.82177734375,
0.88525390625,
0.82177734375,
0.8984375,
0.176513671875,
0.89794921875
],
[
0.176513671875,
0.8984375,
0.6708984375,
0.8984375,
0.6708984375,
0.912109375,
0.176513671875,
0.912109375
],
[
0.490234375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.94970703125,
0.490234375,
0.94970703125
]
] |
[
"想以及分别在FGoodfellowetal.（2014）和EGanin&Lempitsky（2014）中提出的梯度反向方",
"法有关（第E2节）。我们在4节中介绍了Aurora-4语音识别任务的结果，并在5节中总结",
"了我们的发现。",
"2Related Work",
"生成对抗网络由两个网络组成：生成器和鉴别器。生成器网络G具有随机生成的特征向量",
"的输人，并被要求生成一个样本，即。g。图像，类似于训练集中的图像。鉴别器网络D可",
"以接收来自生成器G的生成图像或来自训练集的图像。它的任务是区分“假”生成的图像",
"和从数据集中获取的“真实”图像。因此，鉴别器只是一个具有S形输出层的分类器网络",
"可以使用梯度反向传播进行训练。此梯度可以进一步传播到发电机网络",
"GAN设置中的两个网络相互竞争：生成器试图欺骗鉴别器网络，而鉴别器则尽力识别是否",
"存在欺骗，类似于对抗性博奔论设置。形式上，GAN训练的目标函数为",
"min max V(D, G) = Ea~pata(a)[log D(α)] + Ez~p(z)[log(1 - D(G(z))]",
"鉴别器D上的最大化形成了通常的交叉炳目标，相对于D的参数计算梯度。使用通过第二",
"项传播的梯度来最小化G的参数。G上的最小化使其能够产生D分类器作为训练分类器的",
"示例",
"在Radfordetal（2015）中提出了优化GAN的几个实用指南，并在Salimansetal.（2016）",
"中进行了进一步的探索。Ganin&Lempitsky(2014)的先前工作提出了一种训练可适应新域",
"的网络的方法。训练数据由用感兴趣的类和单独的域（图像背景）标签标记的图像组成。该",
"网络具有类似Y的结构：图像被馈送到第一个网络，该网络产生一个隐藏的表示h。然后",
"将此表示h输人到两个单独的网络：域分类器网络（D）和目标分类器网络（R）。训练的目",
"标是学习对域标签不变的隐藏表示，并在目标分类任务上表现良好，以便域信息在测试时",
"不会干扰目标分类器。与GAN目标类似，该目标迫使生成分布接近数据分布，梯度反向方",
"法使域分布彼此相似",
"网络训练有三个目标：隐藏表示h应该对目标分类器有帮助，对域分类器有害，并且域分类",
"器应该具有良好的分类精度。更正式地说，作者将损失函数定义为",
"L= Li(9,y; OR,Oe)+QL2(d,d;OD)-βL3(d,d;0e),",
"(1)",
"其中y是基本真值类，d是域标签，相应的帽子变量是网络预测，而E，R和D分别是",
"编码器，识别器和域分类器网络的参数子集。超参数α和β表示损失函数项的相对影响",
"Yuetal.（2013）中讨论了神经网络产生的表示对内部降噪的影响，这项工作为Aurora-4数",
"据集的实验设定了基线。最近，在EShunohara（2016）中，以对抗性的方式对多层S形网络",
"进行了内部转录任务的训练，该任务被噪声破坏。",
"3InvariantRepresentationsforSpeechRecognition",
"大多数ASR系统是DNN-HMM混合系统。上下文相关（CD）HMM状",
"态（声学模型）是感兴趣的类标签。录音条件、说话者身份或性别代",
"表GAN中的域。任务是使HMM状态分类器网络的隐藏层表示相对",
"于这些域不变。Wehypothesizethatthisadvers",
"sarialmethodoftraininghelpstheHMM",
"stateclassifiertogeneralizebettertounseendomainconditionsandrequiresonlyasmalladditionalamountofsupervision,i。",
"e。域标签",
"图E1a描绘了模型，该模型与梯度反向方法的模型相同。它是一个经过训练的前馈神经网",
"络，用于预测CDHMM状态，其分支预测域（噪声条件）。此分支在测试阶段被丢弃。在我们",
"的实验中，我们使用噪声条件作为域标签，将所有噪声类型合并到一个标签中，并作为另",
"个标签进行清洁。我们的训练损失函数是方程1，其中L3设置为dlog(1一d)+(1-d)log(d)",
"用于训练期间的稳定性。L3项最大限度地提高了错误域分类的概率，而梯度反向则最大",
"限度地降低了正确分类的概率。术语L1和L2是规则的交叉，使用相应的参数QE和θD",
"将其最小化。为简单起见，我们只使用一个超参数一一第三项的权重。",
"2"
] |
{
"arxiv_id": "1612.01928",
"language": "zh",
"page_idx": 1
}
|
||
1612.01928-zh-2
|
1612.01928-zh-2.jpg
|
<FigureHere>
(a) The model consists of three neural networks. The encoder $E$ produces the intermediate representation $h$ which used in the recognizer $R$ and in the domain discriminator $D$ . The hidden representation $h$ is trained to improve the recognition and minimize the domain discriminator accuracy. The domain discriminator is a classifier trained to maximize its accuracy on the noise type clas-
<FigureHere>
model varying with the number of noise conditions used for training. Bot-
tom: Average performance on seen versus unseen noise conditions. Testing
was performed on all wv1 conditions (Sennheiser microphone).
Figure 1: Model structure for invariant training and ASR results.
3
|
[
[
0.455810546875,
0.11553955078125,
0.5009765625,
0.11553955078125,
0.5009765625,
0.126953125,
0.455810546875,
0.126953125
],
[
0.521484375,
0.11297607421875,
0.5888671875,
0.11297607421875,
0.5888671875,
0.1275634765625,
0.521484375,
0.1275634765625
],
[
0.473876953125,
0.1307373046875,
0.484375,
0.1307373046875,
0.484375,
0.143310546875,
0.473876953125,
0.143310546875
],
[
0.5498046875,
0.1307373046875,
0.560546875,
0.1307373046875,
0.560546875,
0.1427001953125,
0.5498046875,
0.1427001953125
],
[
0.4697265625,
0.15087890625,
0.48779296875,
0.15087890625,
0.48779296875,
0.1673583984375,
0.4697265625,
0.1673583984375
],
[
0.54833984375,
0.154052734375,
0.560546875,
0.154052734375,
0.560546875,
0.1654052734375,
0.54833984375,
0.1654052734375
],
[
0.529296875,
0.1962890625,
0.54833984375,
0.1962890625,
0.54833984375,
0.210205078125,
0.529296875,
0.210205078125
],
[
0.4697265625,
0.23486328125,
0.488525390625,
0.23486328125,
0.488525390625,
0.25,
0.4697265625,
0.25
],
[
0.416748046875,
0.2822265625,
0.54345703125,
0.2822265625,
0.54345703125,
0.3037109375,
0.416748046875,
0.3037109375
],
[
0.41259765625,
0.311279296875,
0.5986328125,
0.311279296875,
0.5986328125,
0.324462890625,
0.41259765625,
0.324462890625
],
[
0.414306640625,
0.323974609375,
0.5986328125,
0.323974609375,
0.5986328125,
0.337158203125,
0.414306640625,
0.337158203125
],
[
0.416748046875,
0.33837890625,
0.59716796875,
0.33837890625,
0.59716796875,
0.348388671875,
0.416748046875,
0.348388671875
],
[
0.41259765625,
0.35107421875,
0.59716796875,
0.349853515625,
0.59716796875,
0.361083984375,
0.41259765625,
0.3623046875
],
[
0.41259765625,
0.3623046875,
0.6005859375,
0.3623046875,
0.6005859375,
0.375732421875,
0.41259765625,
0.375732421875
],
[
0.414306640625,
0.375,
0.5908203125,
0.375,
0.5908203125,
0.385009765625,
0.414306640625,
0.385009765625
],
[
0.414306640625,
0.386962890625,
0.5986328125,
0.386962890625,
0.5986328125,
0.40087890625,
0.414306640625,
0.40087890625
],
[
0.41259765625,
0.399658203125,
0.6005859375,
0.399658203125,
0.6005859375,
0.41357421875,
0.41259765625,
0.41357421875
],
[
0.414306640625,
0.41357421875,
0.5986328125,
0.41357421875,
0.5986328125,
0.423583984375,
0.414306640625,
0.423583984375
],
[
0.414306640625,
0.425537109375,
0.564453125,
0.425537109375,
0.564453125,
0.436279296875,
0.414306640625,
0.436279296875
],
[
0.5703125,
0.4267578125,
0.59619140625,
0.4267578125,
0.59619140625,
0.43505859375,
0.5703125,
0.43505859375
],
[
0.414306640625,
0.438232421875,
0.5986328125,
0.438232421875,
0.5986328125,
0.448974609375,
0.414306640625,
0.448974609375
],
[
0.414306640625,
0.45068359375,
0.59716796875,
0.45068359375,
0.59716796875,
0.4609375,
0.414306640625,
0.4609375
],
[
0.41259765625,
0.462158203125,
0.60107421875,
0.4609375,
0.60107421875,
0.474853515625,
0.41259765625,
0.47607421875
],
[
0.41259765625,
0.47607421875,
0.5,
0.47607421875,
0.5,
0.4873046875,
0.41259765625,
0.4873046875
],
[
0.609375,
0.5087890625,
0.67626953125,
0.5087890625,
0.67626953125,
0.52099609375,
0.609375,
0.52099609375
],
[
0.60888671875,
0.52099609375,
0.64892578125,
0.52099609375,
0.64892578125,
0.53076171875,
0.60888671875,
0.53076171875
],
[
0.30224609375,
0.5498046875,
0.330810546875,
0.5498046875,
0.330810546875,
0.58203125,
0.30224609375,
0.58203125
],
[
0.3203125,
0.5888671875,
0.326904296875,
0.5888671875,
0.326904296875,
0.59521484375,
0.3203125,
0.59521484375
],
[
0.4599609375,
0.6337890625,
0.5498046875,
0.6337890625,
0.5498046875,
0.64453125,
0.4599609375,
0.64453125
],
[
0.56396484375,
0.66943359375,
0.63232421875,
0.66943359375,
0.63232421875,
0.6796875,
0.56396484375,
0.6796875
],
[
0.56396484375,
0.67822265625,
0.60693359375,
0.6796875,
0.60546875,
0.69140625,
0.56201171875,
0.689453125
],
[
0.56396484375,
0.68994140625,
0.6748046875,
0.68994140625,
0.6748046875,
0.70068359375,
0.56396484375,
0.70068359375
],
[
0.56640625,
0.70263671875,
0.646484375,
0.70263671875,
0.646484375,
0.71044921875,
0.56640625,
0.71044921875
],
[
0.30224609375,
0.708984375,
0.330810546875,
0.708984375,
0.330810546875,
0.7412109375,
0.30224609375,
0.7412109375
],
[
0.463134765625,
0.79443359375,
0.5458984375,
0.79443359375,
0.5458984375,
0.8017578125,
0.463134765625,
0.8017578125
],
[
0.28515625,
0.8125,
0.72802734375,
0.8125,
0.72802734375,
0.82568359375,
0.28515625,
0.82568359375
],
[
0.28515625,
0.82568359375,
0.7294921875,
0.82568359375,
0.7294921875,
0.8388671875,
0.28515625,
0.8388671875
],
[
0.28515625,
0.83837890625,
0.7294921875,
0.83837890625,
0.7294921875,
0.8515625,
0.28515625,
0.8515625
],
[
0.28515625,
0.85107421875,
0.66015625,
0.85107421875,
0.66015625,
0.8642578125,
0.28515625,
0.8642578125
],
[
0.28759765625,
0.8720703125,
0.7099609375,
0.8720703125,
0.7099609375,
0.88525390625,
0.28759765625,
0.88525390625
],
[
0.494384765625,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94775390625,
0.494384765625,
0.94775390625
]
] |
[
"phone",
"condition",
"个",
"个",
"R",
"D",
"Jh",
"E",
"",
"(a) The model consists of three",
"neural networks: The encoder",
"E produces the intermediate",
"representation h which used",
"in the recognizer R and in",
"the domain discriminator D",
"The hidden representation h is",
"trained to improve the recogni-",
"tion and minimize the domain",
"discriminator accuracy.",
"The",
"domain discriminator is a clas-",
"sifier trained to maximize its",
"accuracy on the noise type clas-",
"sificationtask",
"Baseline avg",
"Invavg",
"",
"14",
"#noiseconditions",
"Baseline seen",
"Inv seen",
"Baseline unseen (wv1)",
"Invunseen(wv1)",
"",
"#noisecondition",
"(b) Up: Average performance of the baseline multi-condition and invariance",
"model varying with the number of noise conditions used for training. Bot-",
"tom: Average performance on seen versus unseen noise conditions. Testing",
"was performed on all wv1 conditions (Sennheiser microphone).",
"Figure l: Model structure for invariant training and ASR results.",
"3"
] |
{
"arxiv_id": "1612.01928",
"language": "zh",
"page_idx": 2
}
|
||
1612.01928-zh-3
|
1612.01928-zh-3.jpg
|
Table 1: Average word error rate (WER $\%$ ) on Aurora-4 dataset on all test conditions, including seen and unseen noise and unseen microphone. First column is the number of noise conditions used for the training. The last row is a preliminary experiment with layer-wise pre-training close to stateof-the-art model and a corresponding invariance training starting with a pretrained model
<table>
<tbody>
<tr>
<th rowspan="2">Noise</th>
<th rowspan="2">Inv</th>
<th rowspan="2">$BL$</th>
<th colspan="2">A</th>
<th colspan="2">$B$</th>
<th colspan="2">$C$</th>
<th colspan="2">$D$</th>
</tr>
<tr>
<th>Inv</th>
<th>$BL$</th>
<th>Inv</th>
<th>$BL$</th>
<th>Inv</th>
<th>$BL$</th>
<th>Inv</th>
<th>$BL$</th>
</tr>
<tr>
<td>1</td>
<td>$\overline{16.36}$</td>
<td>18.14</td>
<td>$\overline{6.54}$</td>
<td>$\overline{7.57}$</td>
<td>12.71</td>
<td>$\overline{14.09}$</td>
<td>$\overline{11.45}$</td>
<td>$\overline{13.10}$</td>
<td>22.47</td>
<td>$\overline{24.80}$</td>
</tr>
<tr>
<td>2</td>
<td>15.56</td>
<td>17.39</td>
<td>5.90</td>
<td>6.58</td>
<td>11.69</td>
<td>13.28</td>
<td>11.12</td>
<td>13.51</td>
<td>21.79</td>
<td>23.96</td>
</tr>
<tr>
<td>3</td>
<td>14.24</td>
<td>14.67</td>
<td>5.45</td>
<td>5.08</td>
<td>10.76</td>
<td>12.44</td>
<td>9.75</td>
<td>9.84</td>
<td>19.93</td>
<td>19.30</td>
</tr>
<tr>
<td>4</td>
<td>13.61</td>
<td>13.84</td>
<td>5.08</td>
<td>5.29</td>
<td>9.73</td>
<td>9.97</td>
<td>9.49</td>
<td>9.56</td>
<td>19.49</td>
<td>19.90</td>
</tr>
<tr>
<td>5</td>
<td>13.41</td>
<td>13.02</td>
<td>5.12</td>
<td>5.34</td>
<td>9.52</td>
<td>9.42</td>
<td>9.55</td>
<td>8.67</td>
<td>19.33</td>
<td>18.65</td>
</tr>
<tr>
<td>6</td>
<td>12.62</td>
<td>12.60</td>
<td>4.80</td>
<td>4.61</td>
<td>9.04</td>
<td>8.86</td>
<td>8.76</td>
<td>8.59</td>
<td>18.16</td>
<td>18.21</td>
</tr>
<tr>
<td>6米 </td>
<td>$\overline{11.85}$</td>
<td>$\overline{11.99}$</td>
<td>$\overline{4.52}$</td>
<td>$\overline{4.76}$</td>
<td>$\overline{8.76}$</td>
<td>$\overline{8.76}$</td>
<td>$\overline{7.79}$</td>
<td>$\overline{8.57}$</td>
<td>$\overline{16.84}$</td>
<td>$\overline{16.99}$</td>
</tr>
</tbody>
</table>
## 4 $\textbf{Experiments}$
我们通过实验评估了我们在基准良好的 Aurora-4 (Parihar & Picone, 2002) 嘈杂语音识别任务上的方法。Aurora-4 基于华尔街日报语料库 (WSJ0)。它包含六个类别的噪声，这些噪声被添加到干净的数据中。每一个干净和嘈杂的话语被过滤，以模拟频率特性。训练数据包含每个噪声条件下的 4400 个干净的话语和 446 个话语，即。e。总共 2676 个嘈杂的话语。测试集包括干净的数据，被 6 种噪声类型损坏的数据，以及在干净和嘈杂的情况下使用不同麦克风记录的数据
对于干净和嘈杂的数据，我们提取 40 维 Mel-filterbank 特征，其增量和增量-增量在？?? 帧上拼接，从而产生 1320 个输入特征，这些特征随后被均值和方差归一化。基线声学模型是一个6层 DNN, 每层有 2048 个整流线性单元。它是使用动量加速的随机梯度下降训练 15 个时期的新鲍勃退火$\operatorname{E}$(如 Morgan & Bourlard, 1995; ?)
为了评估我们的方法对未见噪声的泛化的影响，我们对不同的可见噪声集进行了6次实验。这些网络在干净的数据上进行训练，每个噪声条件按以下顺序逐一添加：机场，杂音，汽车， 餐厅，街道和火车。最后一个训练组包含所有噪声，因此与标准多条件训练设置匹配。对于每个训练组，我们训练了基线和不变性模型，我们在$4^{th}$层分支到一个二进制分类器，预测干净与嘈杂的数据。由于干净和嘈杂的话语数量之间的不平衡，我们不得不对嘈杂的帧进行过采样，以确保每个小批量包含相等数量的干净和嘈杂的语音帧
表 1总结了结果。图 1b 可视化了基线多条件训练和不变性训练的单词错误率，因为看到的噪声类型的数量不同。我们得出的结论是，在训练过程中提供少量噪声类型时，可以实现最佳性能增益。可以看出，与多条件训练相比，不变性训练能够更好地推广到看不见的噪声类型
我们注意到，我们的实验没有使用通常用于小型数据集的分层预训练。报告的基线 WER 非常接近最新技术。当使用所有噪声类型(表$\overline{\mathrm{E}}$1 的最后一行)进行训练时，我们在预训练网络(整体 WER 更好) 上的初步实验显示出与非预训练网络相同的趋势
# 5 Discussion
本文介绍了生成对抗网络和不变性训练在噪声鲁棒语音识别中的应用。我们表明，不变性训练有助于 ASR 系统更好地推广到看不见的噪声条件，并在训练期间看到少量噪声类型时提高字错误率。我们的实验表明，与图像识别任务相反，在语音识别中，域自适应网络遭受欠拟合。因此，方程 1 中$L_3$项的梯度是不可靠的，并且有噪声。未来的研究包括对域自适应网络的增强，同时探索替代网络架构和不变性促进损失函数
# 6
Acknowledgments
我们要感谢 YaroslavGanin, DavidWarde-Farley 的深人讨论，Theano$\boxed{\mathrm{F}}$Theano Development Team (2016) , Blocks 和 Fuel$\boxed{\mathrm{E}}$van Merriënboer et al. (2015) 的开发人员提供了出色的工具包。
4
|
[
[
0.17236328125,
0.09783935546875,
0.82275390625,
0.0985107421875,
0.82275390625,
0.11297607421875,
0.17236328125,
0.11236572265625
],
[
0.1748046875,
0.11297607421875,
0.82421875,
0.11297607421875,
0.82421875,
0.126953125,
0.1748046875,
0.126953125
],
[
0.1748046875,
0.126953125,
0.826171875,
0.126953125,
0.826171875,
0.141357421875,
0.1748046875,
0.141357421875
],
[
0.1748046875,
0.14013671875,
0.7587890625,
0.14013671875,
0.7587890625,
0.1546630859375,
0.1748046875,
0.1546630859375
],
[
0.205078125,
0.154052734375,
0.2484130859375,
0.1558837890625,
0.2467041015625,
0.1673583984375,
0.2034912109375,
0.1654052734375
],
[
0.268798828125,
0.1546630859375,
0.294189453125,
0.1546630859375,
0.294189453125,
0.1673583984375,
0.268798828125,
0.1673583984375
],
[
0.325927734375,
0.1546630859375,
0.3505859375,
0.1546630859375,
0.3505859375,
0.1673583984375,
0.325927734375,
0.1673583984375
],
[
0.410888671875,
0.1558837890625,
0.423095703125,
0.1558837890625,
0.423095703125,
0.1673583984375,
0.410888671875,
0.1673583984375
],
[
0.5146484375,
0.1558837890625,
0.52783203125,
0.1558837890625,
0.52783203125,
0.1666259765625,
0.5146484375,
0.1666259765625
],
[
0.62744140625,
0.1558837890625,
0.6396484375,
0.1558837890625,
0.6396484375,
0.1666259765625,
0.62744140625,
0.1666259765625
],
[
0.740234375,
0.1558837890625,
0.75341796875,
0.1558837890625,
0.75341796875,
0.1666259765625,
0.740234375,
0.1666259765625
],
[
0.381591796875,
0.1654052734375,
0.41015625,
0.17041015625,
0.406005859375,
0.18310546875,
0.37744140625,
0.177978515625
],
[
0.53515625,
0.1673583984375,
0.560546875,
0.1673583984375,
0.560546875,
0.179931640625,
0.53515625,
0.179931640625
],
[
0.59326171875,
0.1666259765625,
0.619140625,
0.169189453125,
0.61767578125,
0.18115234375,
0.5908203125,
0.177978515625
],
[
0.64892578125,
0.1673583984375,
0.67333984375,
0.1673583984375,
0.67333984375,
0.179931640625,
0.64892578125,
0.179931640625
],
[
0.76123046875,
0.1673583984375,
0.7861328125,
0.1673583984375,
0.7861328125,
0.179931640625,
0.76123046875,
0.179931640625
],
[
0.48046875,
0.1685791015625,
0.50732421875,
0.1685791015625,
0.50732421875,
0.18115234375,
0.48046875,
0.18115234375
],
[
0.426513671875,
0.1685791015625,
0.453369140625,
0.1685791015625,
0.453369140625,
0.179931640625,
0.426513671875,
0.179931640625
],
[
0.7060546875,
0.1685791015625,
0.736328125,
0.1685791015625,
0.736328125,
0.18115234375,
0.7060546875,
0.18115234375
],
[
0.25732421875,
0.1817626953125,
0.303955078125,
0.1817626953125,
0.303955078125,
0.1962890625,
0.25732421875,
0.1962890625
],
[
0.41015625,
0.1717529296875,
0.423095703125,
0.1717529296875,
0.423095703125,
0.1749267578125,
0.41015625,
0.1749267578125
],
[
0.31787109375,
0.18310546875,
0.36279296875,
0.18310546875,
0.36279296875,
0.1944580078125,
0.31787109375,
0.1944580078125
],
[
0.375732421875,
0.18310546875,
0.414306640625,
0.18310546875,
0.414306640625,
0.1962890625,
0.375732421875,
0.1962890625
],
[
0.423095703125,
0.18310546875,
0.45751953125,
0.18310546875,
0.45751953125,
0.1944580078125,
0.423095703125,
0.1944580078125
],
[
0.47216796875,
0.18310546875,
0.5146484375,
0.18310546875,
0.5146484375,
0.1944580078125,
0.47216796875,
0.1944580078125
],
[
0.529296875,
0.18310546875,
0.57177734375,
0.18310546875,
0.57177734375,
0.1944580078125,
0.529296875,
0.1944580078125
],
[
0.58251953125,
0.18310546875,
0.6259765625,
0.18310546875,
0.6259765625,
0.1944580078125,
0.58251953125,
0.1944580078125
],
[
0.640625,
0.18310546875,
0.6845703125,
0.18310546875,
0.6845703125,
0.1944580078125,
0.640625,
0.1944580078125
],
[
0.6953125,
0.18310546875,
0.73876953125,
0.18310546875,
0.73876953125,
0.1944580078125,
0.6953125,
0.1944580078125
],
[
0.74853515625,
0.18310546875,
0.79736328125,
0.18310546875,
0.79736328125,
0.1944580078125,
0.74853515625,
0.1944580078125
],
[
0.41259765625,
0.1849365234375,
0.418212890625,
0.189453125,
0.4150390625,
0.1912841796875,
0.41015625,
0.1875
],
[
0.2344970703125,
0.1849365234375,
0.25,
0.1849365234375,
0.25,
0.1932373046875,
0.2344970703125,
0.1932373046875
],
[
0.6845703125,
0.1875,
0.69287109375,
0.18310546875,
0.69677734375,
0.1881103515625,
0.689453125,
0.1925048828125
],
[
0.2607421875,
0.1956787109375,
0.30224609375,
0.1956787109375,
0.30224609375,
0.210205078125,
0.2607421875,
0.210205078125
],
[
0.31787109375,
0.1956787109375,
0.359375,
0.1956787109375,
0.359375,
0.210205078125,
0.31787109375,
0.210205078125
],
[
0.471435546875,
0.1956787109375,
0.515625,
0.1956787109375,
0.515625,
0.210205078125,
0.471435546875,
0.210205078125
],
[
0.52783203125,
0.1956787109375,
0.57177734375,
0.1956787109375,
0.57177734375,
0.210205078125,
0.52783203125,
0.210205078125
],
[
0.5849609375,
0.1956787109375,
0.62744140625,
0.1956787109375,
0.62744140625,
0.210205078125,
0.5849609375,
0.210205078125
],
[
0.64208984375,
0.1956787109375,
0.68310546875,
0.1956787109375,
0.68310546875,
0.210205078125,
0.64208984375,
0.210205078125
],
[
0.69677734375,
0.1956787109375,
0.740234375,
0.1956787109375,
0.740234375,
0.210205078125,
0.69677734375,
0.210205078125
],
[
0.75244140625,
0.1956787109375,
0.7958984375,
0.1956787109375,
0.7958984375,
0.210205078125,
0.75244140625,
0.210205078125
],
[
0.4248046875,
0.1962890625,
0.45751953125,
0.1962890625,
0.45751953125,
0.208984375,
0.4248046875,
0.208984375
],
[
0.236083984375,
0.1976318359375,
0.2467041015625,
0.1976318359375,
0.2467041015625,
0.20703125,
0.236083984375,
0.20703125
],
[
0.37744140625,
0.1962890625,
0.410888671875,
0.1962890625,
0.410888671875,
0.208984375,
0.37744140625,
0.208984375
],
[
0.2607421875,
0.208984375,
0.303955078125,
0.208984375,
0.303955078125,
0.2235107421875,
0.2607421875,
0.2235107421875
],
[
0.31787109375,
0.208984375,
0.359375,
0.208984375,
0.359375,
0.2235107421875,
0.31787109375,
0.2235107421875
],
[
0.47216796875,
0.208984375,
0.515625,
0.208984375,
0.515625,
0.2235107421875,
0.47216796875,
0.2235107421875
],
[
0.529296875,
0.208984375,
0.57177734375,
0.208984375,
0.57177734375,
0.2235107421875,
0.529296875,
0.2235107421875
],
[
0.75341796875,
0.20703125,
0.7958984375,
0.208984375,
0.7939453125,
0.2235107421875,
0.75244140625,
0.2215576171875
],
[
0.4248046875,
0.210205078125,
0.45751953125,
0.210205078125,
0.45751953125,
0.222900390625,
0.4248046875,
0.222900390625
],
[
0.37744140625,
0.210205078125,
0.41015625,
0.210205078125,
0.41015625,
0.222900390625,
0.37744140625,
0.222900390625
],
[
0.5888671875,
0.210205078125,
0.62353515625,
0.210205078125,
0.62353515625,
0.222900390625,
0.5888671875,
0.222900390625
],
[
0.64453125,
0.210205078125,
0.6787109375,
0.210205078125,
0.6787109375,
0.222900390625,
0.64453125,
0.222900390625
],
[
0.69921875,
0.210205078125,
0.73779296875,
0.210205078125,
0.73779296875,
0.2215576171875,
0.69921875,
0.2215576171875
],
[
0.31787109375,
0.222900390625,
0.359375,
0.222900390625,
0.359375,
0.2374267578125,
0.31787109375,
0.2374267578125
],
[
0.236083984375,
0.2108154296875,
0.2467041015625,
0.2108154296875,
0.2467041015625,
0.2215576171875,
0.236083984375,
0.2215576171875
],
[
0.261474609375,
0.2235107421875,
0.301513671875,
0.2235107421875,
0.301513671875,
0.23486328125,
0.261474609375,
0.23486328125
],
[
0.37744140625,
0.2235107421875,
0.41015625,
0.2235107421875,
0.41015625,
0.236083984375,
0.37744140625,
0.236083984375
],
[
0.4248046875,
0.2235107421875,
0.45751953125,
0.2235107421875,
0.45751953125,
0.23486328125,
0.4248046875,
0.23486328125
],
[
0.476318359375,
0.2235107421875,
0.5107421875,
0.2235107421875,
0.5107421875,
0.236083984375,
0.476318359375,
0.236083984375
],
[
0.53369140625,
0.2235107421875,
0.56640625,
0.2235107421875,
0.56640625,
0.236083984375,
0.53369140625,
0.236083984375
],
[
0.5888671875,
0.2235107421875,
0.62158203125,
0.2235107421875,
0.62158203125,
0.23486328125,
0.5888671875,
0.23486328125
],
[
0.64453125,
0.2235107421875,
0.6787109375,
0.2235107421875,
0.6787109375,
0.236083984375,
0.64453125,
0.236083984375
],
[
0.69775390625,
0.2235107421875,
0.73876953125,
0.2235107421875,
0.73876953125,
0.23486328125,
0.69775390625,
0.23486328125
],
[
0.7548828125,
0.2235107421875,
0.7939453125,
0.2235107421875,
0.7939453125,
0.23486328125,
0.7548828125,
0.23486328125
],
[
0.261474609375,
0.236083984375,
0.30224609375,
0.236083984375,
0.30224609375,
0.250732421875,
0.261474609375,
0.250732421875
],
[
0.236083984375,
0.2247314453125,
0.2467041015625,
0.2247314453125,
0.2467041015625,
0.23486328125,
0.236083984375,
0.23486328125
],
[
0.31787109375,
0.236083984375,
0.359375,
0.236083984375,
0.359375,
0.250732421875,
0.31787109375,
0.250732421875
],
[
0.64453125,
0.236083984375,
0.6787109375,
0.236083984375,
0.6787109375,
0.251953125,
0.64453125,
0.251953125
],
[
0.69677734375,
0.236083984375,
0.73876953125,
0.236083984375,
0.73876953125,
0.250732421875,
0.69677734375,
0.250732421875
],
[
0.75341796875,
0.236083984375,
0.7939453125,
0.236083984375,
0.7939453125,
0.250732421875,
0.75341796875,
0.250732421875
],
[
0.37744140625,
0.2374267578125,
0.41015625,
0.2374267578125,
0.41015625,
0.25,
0.37744140625,
0.25
],
[
0.240234375,
0.2374267578125,
0.2484130859375,
0.24365234375,
0.240234375,
0.25,
0.2320556640625,
0.24365234375
],
[
0.4248046875,
0.2374267578125,
0.45751953125,
0.2374267578125,
0.45751953125,
0.25,
0.4248046875,
0.25
],
[
0.47802734375,
0.2374267578125,
0.5087890625,
0.2374267578125,
0.5087890625,
0.25,
0.47802734375,
0.25
],
[
0.53369140625,
0.2374267578125,
0.56640625,
0.2374267578125,
0.56640625,
0.25,
0.53369140625,
0.25
],
[
0.5888671875,
0.2374267578125,
0.62158203125,
0.2374267578125,
0.62158203125,
0.25,
0.5888671875,
0.25
],
[
0.240234375,
0.248779296875,
0.25244140625,
0.25830078125,
0.2418212890625,
0.266357421875,
0.2296142578125,
0.2568359375
],
[
0.2607421875,
0.251953125,
0.301513671875,
0.251953125,
0.301513671875,
0.26318359375,
0.2607421875,
0.26318359375
],
[
0.318603515625,
0.251953125,
0.359375,
0.251953125,
0.359375,
0.26318359375,
0.318603515625,
0.26318359375
],
[
0.37744140625,
0.251953125,
0.41015625,
0.251953125,
0.41015625,
0.26318359375,
0.37744140625,
0.26318359375
],
[
0.423095703125,
0.251953125,
0.455810546875,
0.251953125,
0.455810546875,
0.26318359375,
0.423095703125,
0.26318359375
],
[
0.476318359375,
0.251953125,
0.5107421875,
0.251953125,
0.5107421875,
0.26318359375,
0.476318359375,
0.26318359375
],
[
0.53369140625,
0.251953125,
0.56640625,
0.251953125,
0.56640625,
0.26318359375,
0.53369140625,
0.26318359375
],
[
0.5888671875,
0.251953125,
0.62158203125,
0.251953125,
0.62158203125,
0.264404296875,
0.5888671875,
0.264404296875
],
[
0.646484375,
0.251953125,
0.6787109375,
0.251953125,
0.6787109375,
0.264404296875,
0.646484375,
0.264404296875
],
[
0.69677734375,
0.251953125,
0.73876953125,
0.251953125,
0.73876953125,
0.26318359375,
0.69677734375,
0.26318359375
],
[
0.75341796875,
0.250732421875,
0.7939453125,
0.250732421875,
0.7939453125,
0.26513671875,
0.75341796875,
0.26513671875
],
[
0.223876953125,
0.266357421875,
0.301513671875,
0.266357421875,
0.301513671875,
0.281005859375,
0.223876953125,
0.281005859375
],
[
0.41015625,
0.266357421875,
0.5107421875,
0.266357421875,
0.5107421875,
0.281005859375,
0.41015625,
0.281005859375
],
[
0.529296875,
0.266357421875,
0.62353515625,
0.268310546875,
0.62353515625,
0.28271484375,
0.529296875,
0.281005859375
],
[
0.67333984375,
0.267578125,
0.74169921875,
0.267578125,
0.74169921875,
0.2822265625,
0.67333984375,
0.2822265625
],
[
0.75244140625,
0.268310546875,
0.79736328125,
0.268310546875,
0.79736328125,
0.27978515625,
0.75244140625,
0.27978515625
],
[
0.64453125,
0.26953125,
0.6806640625,
0.26953125,
0.6806640625,
0.281005859375,
0.64453125,
0.281005859375
],
[
0.316162109375,
0.267578125,
0.414306640625,
0.267578125,
0.414306640625,
0.2822265625,
0.316162109375,
0.2822265625
],
[
0.1707763671875,
0.3037109375,
0.31787109375,
0.304931640625,
0.31787109375,
0.322509765625,
0.1707763671875,
0.3212890625
],
[
0.176513671875,
0.3359375,
0.82177734375,
0.3359375,
0.82177734375,
0.349853515625,
0.176513671875,
0.349853515625
],
[
0.17724609375,
0.349853515625,
0.82275390625,
0.349853515625,
0.82275390625,
0.363037109375,
0.17724609375,
0.363037109375
],
[
0.1748046875,
0.3642578125,
0.82275390625,
0.3642578125,
0.82275390625,
0.37744140625,
0.1748046875,
0.37744140625
],
[
0.1748046875,
0.376953125,
0.82177734375,
0.376953125,
0.82177734375,
0.39013671875,
0.1748046875,
0.39013671875
],
[
0.176513671875,
0.392578125,
0.82275390625,
0.392578125,
0.82275390625,
0.406005859375,
0.176513671875,
0.406005859375
],
[
0.176513671875,
0.40478515625,
0.275390625,
0.406005859375,
0.275390625,
0.417236328125,
0.176513671875,
0.416748046875
],
[
0.176513671875,
0.425537109375,
0.82421875,
0.425537109375,
0.82421875,
0.439453125,
0.176513671875,
0.439453125
],
[
0.1748046875,
0.439453125,
0.810546875,
0.439453125,
0.810546875,
0.453857421875,
0.1748046875,
0.453857421875
],
[
0.1748046875,
0.453857421875,
0.82275390625,
0.453857421875,
0.82275390625,
0.46728515625,
0.1748046875,
0.46728515625
],
[
0.1732177734375,
0.466552734375,
0.5400390625,
0.46728515625,
0.5400390625,
0.48095703125,
0.1732177734375,
0.479736328125
],
[
0.178955078125,
0.4892578125,
0.8203125,
0.4892578125,
0.8203125,
0.5,
0.178955078125,
0.5
],
[
0.176513671875,
0.501953125,
0.82177734375,
0.501953125,
0.82177734375,
0.51513671875,
0.176513671875,
0.51513671875
],
[
0.17724609375,
0.51513671875,
0.82177734375,
0.51513671875,
0.82177734375,
0.52880859375,
0.17724609375,
0.52880859375
],
[
0.176513671875,
0.52978515625,
0.82177734375,
0.52978515625,
0.82177734375,
0.54345703125,
0.176513671875,
0.54345703125
],
[
0.17724609375,
0.54345703125,
0.82421875,
0.54345703125,
0.82421875,
0.556640625,
0.17724609375,
0.556640625
],
[
0.176513671875,
0.556640625,
0.65185546875,
0.556640625,
0.65185546875,
0.57080078125,
0.176513671875,
0.57080078125
],
[
0.176513671875,
0.5771484375,
0.82177734375,
0.5771484375,
0.82177734375,
0.59033203125,
0.176513671875,
0.59033203125
],
[
0.176513671875,
0.59130859375,
0.82275390625,
0.59130859375,
0.82275390625,
0.60498046875,
0.176513671875,
0.60498046875
],
[
0.17724609375,
0.60693359375,
0.82275390625,
0.60693359375,
0.82275390625,
0.6171875,
0.17724609375,
0.6171875
],
[
0.176513671875,
0.62060546875,
0.191162109375,
0.62060546875,
0.191162109375,
0.6298828125,
0.176513671875,
0.6298828125
],
[
0.176513671875,
0.6396484375,
0.82275390625,
0.6396484375,
0.82275390625,
0.65283203125,
0.176513671875,
0.65283203125
],
[
0.176513671875,
0.65283203125,
0.82275390625,
0.65283203125,
0.82275390625,
0.66650390625,
0.176513671875,
0.66650390625
],
[
0.176513671875,
0.66845703125,
0.6748046875,
0.66845703125,
0.6748046875,
0.6787109375,
0.176513671875,
0.6787109375
],
[
0.17236328125,
0.69970703125,
0.2998046875,
0.70068359375,
0.2998046875,
0.71533203125,
0.17236328125,
0.7138671875
],
[
0.176513671875,
0.73095703125,
0.82275390625,
0.73095703125,
0.82275390625,
0.744140625,
0.176513671875,
0.744140625
],
[
0.176513671875,
0.74560546875,
0.82275390625,
0.74560546875,
0.82275390625,
0.7587890625,
0.176513671875,
0.7587890625
],
[
0.1732177734375,
0.75830078125,
0.82275390625,
0.7568359375,
0.82275390625,
0.771484375,
0.1732177734375,
0.77294921875
],
[
0.176513671875,
0.77294921875,
0.82275390625,
0.77294921875,
0.82275390625,
0.7861328125,
0.176513671875,
0.7861328125
],
[
0.17724609375,
0.78857421875,
0.615234375,
0.78857421875,
0.615234375,
0.798828125,
0.17724609375,
0.798828125
],
[
0.1732177734375,
0.8193359375,
0.1895751953125,
0.8193359375,
0.1895751953125,
0.833984375,
0.1732177734375,
0.833984375
],
[
0.1748046875,
0.85107421875,
0.298095703125,
0.85107421875,
0.298095703125,
0.8642578125,
0.1748046875,
0.8642578125
],
[
0.1748046875,
0.869140625,
0.82275390625,
0.87060546875,
0.82275390625,
0.88525390625,
0.1748046875,
0.8837890625
],
[
0.176513671875,
0.8837890625,
0.82275390625,
0.8837890625,
0.82275390625,
0.89794921875,
0.176513671875,
0.89794921875
],
[
0.176513671875,
0.8994140625,
0.2017822265625,
0.8994140625,
0.2017822265625,
0.912109375,
0.176513671875,
0.912109375
],
[
0.490234375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.94970703125,
0.490234375,
0.94970703125
]
] |
[
"Table l: Average word error rate (WER % ) on Aurora-4 dataset on all test conditions, including",
"seen and unseen noise and unseen microphone. First column is the number of noise conditions used",
"for the training. The last row is a preliminary experiment with layer-wise pre-training close to state-",
"of-the-art model and a corresponding invariance training starting with a pretrained model",
"Noise",
"Inv",
"BL",
"A",
"B",
"C",
"D",
"Inv",
"BL",
"Inv",
"BL",
"BL",
"Inv",
"BL",
"Inv",
"16.36",
"",
"18.14",
"6.54",
"7.57",
"12.71",
"14.09",
"11.45",
"13.10",
"22.47",
"24.80",
"",
"1",
"/",
"15.56",
"17.39",
"11.69",
"13.28",
"11.12",
"13.51",
"21.79",
"23.96",
"6.58",
"2",
"5.90",
"14.24",
"14.67",
"10.76",
"12.44",
"19.30",
"5.08",
"5.45",
"9.75",
"9.84",
"19.93",
"13.84",
"3",
"13.61",
"5.08",
"5.29",
"9.73",
"9.97",
"9.49",
"9.56",
"19.49",
"19.90",
"13.41",
"4",
"13.02",
"8.67",
"19.33",
"18.65",
"5.12",
"5",
"5.34",
"9.52",
"9.42",
"9.55",
"6",
"12.62",
"12.60",
"4.80",
"4.61",
"9.04",
"8.86",
"8.76",
"8.59",
"18.16",
"18.21",
"6*11.85",
"4.768.76",
"8.767.79",
"716.84",
"16.99",
"8.57",
"11.994.52",
"4Experiments",
"我们通过实验评估了我们在基准良好的Aurora-4（(Parihar&Picone,2002）杂语音识别任务",
"上的方法。Aurora-4基于华尔街日报语料库（WSJO)。它包含六个类别的噪声，这些噪声被",
"添加到干净的数据中。每一个干净和噜杂的话语被过滤，以模拟频率特性。训练数据包含每",
"个噪声条件下的4400个干净的话语和446个话语，即。e。总共2676个噜杂的话语。测试",
"集包括干净的数据，被6种噪声类型损坏的数据，以及在干净和噜杂的情况下使用不同麦克",
"风记录的数据",
"对于干净和杂的数据，我们提取40维Mel-filterbank特征，其增量和增量-增量在???顿上",
"拼接，从而产生1320个输人特征，这些特征随后被均值和方差归一化。基线声学模型是",
"个6层DNN，每层有2048个整流线性单元。它是使用动量加速的随机梯度下降训练15个",
"时期的新鲍勃退火E（如Morgan&Bourlard，1995;?）",
"为了评估我们的方法对未见噪声的泛化的影响，我们对不同的可见噪声集进行了6次实验",
"这些网络在干净的数据上进行训练，每个噪声条件按以下顺序逐一添加：机场，杂音，汽车，",
"餐厅，街道和火车。最后一个训练组包含所有噪声，因此与标准多条件训练设置匹配。对于",
"每个训练组，我们训练了基线和不变性模型，我们在4th层分支到一个二进制分类器，预测",
"干净与杂的数据。由于干净和噜杂的话语数量之间的不平衡，我们不得不对杂的顿进",
"行过采样，以确保每个小批量包含相等数量的干净和噜杂的语音顿",
"表1总结了结果。图1b可视化了基线多条件训练和不变性训练的单词错误率，因为看到的",
"噪声类型的数量不同。我们得出的结论是，在训练过程中提供少量噪声类型时，可以实现最",
"佳性能增益。可以看出，与多条件训练相比，不变性训练能够更好地推广到看不见的噪声类",
"型",
"我们注意到，我们的实验没有使用通常用于小型数据集的分层预训练。报告的基线WER非",
"常接近最新技术。当使用所有噪声类型（表E1的最后一行）进行训练时，我们在预训练网",
"络（整体WER更好）上的初步实验显示出与非预训练网络相同的趋势",
"5Discussion",
"本文介绍了生成对抗网络和不变性训练在噪声鲁棒语音识别中的应用。我们表明，不变性",
"训练有助于ASR系统更好地推广到看不见的噪声条件，并在训练期间看到少量噪声类型时",
"提高字错误率。我们的实验表明，与图像识别任务相反，在语音识别中，域自适应网络遭受",
"欠拟合。因此，方程1中L3项的梯度是不可靠的，并且有噪声。未来的研究包括对域自适",
"应网络的增强，同时探索替代网络架构和不变性促进损失函数",
"6",
"Acknowledgments",
"我们要感谢YaroslavGanin，DavidWarde-Farley的深人讨论，TheanoETheanoDevelopment",
"Team（2016），Blocks和FuelEvanMerrienboeretal.（2015）的开发人员提供了出色的工具",
"包。",
"4"
] |
{
"arxiv_id": "1612.01928",
"language": "zh",
"page_idx": 3
}
|
||
1612.01928-zh-4
|
1612.01928-zh-4.jpg
|
# References
Ganin, Yaroslav, & Lempitsky, Victor. 2014. Unsupervised Domain Adaptation by Backpropaga-
tion. $ArXi\nu e$-prints, Sept.
Goodfellow, Ian, Pouget-Abadie, Jean, Mirza, Mehdi, Xu, Bing, Warde-Farley, David, Ozair, Sher-
jil, Courville, Aaron, & Bengio, Yoshua. 2014. Generative Adversarial Nets. Pages $2672-2680of:$ Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N. D., & Weinberger, K. Q. (eds), Advances in Neural Information Processing Systems 27. Curran Associates, Inc.
Hinton, Geoffrey, Deng, Li, Yu, Dong, Dahl, George E, Mohamed, Abdel-rahman, Jaitly, Navdeep,
Senior, Andrew, Vanhoucke, Vincent, Nguyen, Patrick, Sainath, Tara N, $etal..2012.$ Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal Processing Magazine, 29(6), 82-97.
Kalinli, Ozlem, Seltzer, Michael L, Droppo, Jasha, & Alex. 2010. Noise adaptive training
for robust automatic speech recognition. $IEEE$ Transactions on Audio, Speech, and Language
$Processing,\mathbf{18}(8),1889-1901.$
Miao, Yajie, Gowayyed, Mohammad, & Metze, Florian. 2015. EESEN: End-to-end speech recog-
nition using deep RNN models and WFST-based decoding. Pages 167-174 of: 2015 IEEE Work-
shop on Automatic Speech Recognition and Understanding (ASRU). IEEE.
Morgan, Nelson, & Bourlard, Herve. 1995. Continuous speech recognition. $IEEE$ signal processing
$magazine,12(3),24-42.$
Parihar, N, & & Picone, J. 2002. Aurora working group: DSR front end LVCSR evaluation AU/385/02.
$Inst.forSignalandInformationProcess,MississippiStateUniversity,Tech.Rep,40,94.$
Radford, Alec, Metz, Luke, & Chintala, Soumith. 2015. Unsupervised representation learning with
deep convolutional generative adversarial networks. $arXi\nu\hat{p}reprintar\hat{X}iv:1511.06434.$
Sainath, Tara N, Kingsbury, Brian, Ramabhadran, Bhuvana, Fousek, Petr, Novak, Petr, & Mohamed, Abdel-rahman. 2011. Making deep belief networks effective for large vocabulary continuous speech recognition. Pages 30-35 of: Automatic Speech Recognition and Understanding (ASRU), $2011\textit{IEEE Workshop on. IEEE}.$
Sainath, Tara N, Weiss, Ron J, Senior, Andrew, Wilson, Kevin W, & Vinyals, Oriol. 2015. Learning
the speech front-end with raw waveform CLDNNs. In: Proc. Interspeech.
Salimans, Tim, Goodfellow, Ian, Zaremba, Wojciech, Cheung, Vicki, Radford, Alec, & Chen, Xi.
2016. Improved techniques for training GANs. arXiv preprint arXiv:1606.03498.
Saon, George, Soltau, Hagen, Nahamoo, David, & Picheny, Michael. 2013. Speaker adaptation of
neural network acoustic models using i-vectors. Pages 55-59 of: ASRU.
Seide, Frank, Li, Gang, & Yu, Dong. 2011. Conversational Speech Transcription Using Context-
Dependent Deep Neural Networks. Pages 437-440 of: Interspeech.
Shunohara, Yusuke. 2016. Adversarial Multi-task Learning of Deep Neural Networks for Robust
Speech Recognition. Interspeech 2016, 2369-2372.
Theano Development Team. 2016. Theano: A Python framework for fast computation of mathemat-
ical expressions. arXive-prints, abs/1605.02688(May).
Un, Chong Kwan, Kim, Nam Soo, et al. .1998. Speech recognition in noisy environments using
first-order vector Taylor series. Speech Communication, 24(1), 39-49.
van Merriënboer, Bart, Bahdanau, Dzmitry, Dumoulin, Vincent, Serdyuk, Dmitriy, Warde-Farley, David, Chorowski, Jan, & Bengio, Yoshua. 2015. Blocks and Fuel: Frameworks for deep learning. $CoRR,\mathbf{abs/1506.00619}.$
Yu, Dong, Seltzer, Michael L, Li, Jinyu, Huang, Jui-Ting,& Seide, Frank. 2013. Feature learning in deep neural networks- studies on speech recognition tasks. $arXi\nu \textit{preprint arXiv: 1301. 3605. }$
5
|
[
[
0.176513671875,
0.091552734375,
0.268798828125,
0.091552734375,
0.268798828125,
0.1060791015625,
0.176513671875,
0.1060791015625
],
[
0.17236328125,
0.11297607421875,
0.826171875,
0.1142578125,
0.826171875,
0.1287841796875,
0.17236328125,
0.1275634765625
],
[
0.1895751953125,
0.1300048828125,
0.359375,
0.1300048828125,
0.359375,
0.14013671875,
0.1895751953125,
0.14013671875
],
[
0.176513671875,
0.154052734375,
0.82177734375,
0.154052734375,
0.82177734375,
0.1641845703125,
0.176513671875,
0.1641845703125
],
[
0.1895751953125,
0.1666259765625,
0.82177734375,
0.1654052734375,
0.82177734375,
0.179931640625,
0.1895751953125,
0.18115234375
],
[
0.191162109375,
0.179931640625,
0.82275390625,
0.179931640625,
0.82275390625,
0.1932373046875,
0.191162109375,
0.1932373046875
],
[
0.1895751953125,
0.1944580078125,
0.65185546875,
0.1944580078125,
0.65185546875,
0.2083740234375,
0.1895751953125,
0.2083740234375
],
[
0.1748046875,
0.2171630859375,
0.82177734375,
0.2171630859375,
0.82177734375,
0.231689453125,
0.1748046875,
0.231689453125
],
[
0.1895751953125,
0.231689453125,
0.82177734375,
0.231689453125,
0.82177734375,
0.24560546875,
0.1895751953125,
0.24560546875
],
[
0.193603515625,
0.2474365234375,
0.8203125,
0.2474365234375,
0.8203125,
0.25830078125,
0.193603515625,
0.25830078125
],
[
0.191162109375,
0.2587890625,
0.515625,
0.2587890625,
0.515625,
0.2734375,
0.191162109375,
0.2734375
],
[
0.1748046875,
0.28271484375,
0.82177734375,
0.28271484375,
0.82177734375,
0.296630859375,
0.1748046875,
0.296630859375
],
[
0.191162109375,
0.296630859375,
0.82177734375,
0.296630859375,
0.82177734375,
0.311279296875,
0.191162109375,
0.311279296875
],
[
0.1895751953125,
0.31005859375,
0.39453125,
0.309326171875,
0.39453125,
0.323974609375,
0.1895751953125,
0.324462890625
],
[
0.1748046875,
0.333984375,
0.82421875,
0.333984375,
0.82421875,
0.348388671875,
0.1748046875,
0.348388671875
],
[
0.191162109375,
0.347900390625,
0.826171875,
0.347900390625,
0.826171875,
0.361083984375,
0.191162109375,
0.361083984375
],
[
0.1895751953125,
0.361083984375,
0.6806640625,
0.361083984375,
0.6806640625,
0.375,
0.1895751953125,
0.375
],
[
0.17236328125,
0.3837890625,
0.82275390625,
0.386474609375,
0.82275390625,
0.399658203125,
0.17236328125,
0.397705078125
],
[
0.188720703125,
0.40087890625,
0.351318359375,
0.3984375,
0.351318359375,
0.412353515625,
0.188720703125,
0.4140625
],
[
0.17236328125,
0.421630859375,
0.82177734375,
0.42236328125,
0.82177734375,
0.4375,
0.17236328125,
0.436279296875
],
[
0.1895751953125,
0.4375,
0.76953125,
0.4375,
0.76953125,
0.45068359375,
0.1895751953125,
0.45068359375
],
[
0.17236328125,
0.458984375,
0.82275390625,
0.460205078125,
0.82275390625,
0.474853515625,
0.17236328125,
0.473388671875
],
[
0.1895751953125,
0.474853515625,
0.75732421875,
0.474853515625,
0.75732421875,
0.488037109375,
0.1895751953125,
0.488037109375
],
[
0.1732177734375,
0.498779296875,
0.82275390625,
0.498779296875,
0.82275390625,
0.51220703125,
0.1732177734375,
0.51220703125
],
[
0.191162109375,
0.51318359375,
0.82177734375,
0.51318359375,
0.82177734375,
0.5263671875,
0.191162109375,
0.5263671875
],
[
0.191162109375,
0.5244140625,
0.82275390625,
0.52587890625,
0.82275390625,
0.54052734375,
0.191162109375,
0.5390625
],
[
0.191162109375,
0.5390625,
0.404296875,
0.5390625,
0.404296875,
0.5537109375,
0.191162109375,
0.5537109375
],
[
0.17236328125,
0.5625,
0.82275390625,
0.56298828125,
0.82275390625,
0.57763671875,
0.17236328125,
0.5771484375
],
[
0.1895751953125,
0.57763671875,
0.67724609375,
0.57763671875,
0.67724609375,
0.59130859375,
0.1895751953125,
0.59130859375
],
[
0.1732177734375,
0.6015625,
0.82275390625,
0.6015625,
0.82275390625,
0.61572265625,
0.1732177734375,
0.61572265625
],
[
0.1895751953125,
0.61572265625,
0.7255859375,
0.61572265625,
0.7255859375,
0.62890625,
0.1895751953125,
0.62890625
],
[
0.1732177734375,
0.63720703125,
0.826171875,
0.63818359375,
0.826171875,
0.65283203125,
0.1732177734375,
0.6513671875
],
[
0.1895751953125,
0.65283203125,
0.6650390625,
0.65283203125,
0.6650390625,
0.66650390625,
0.1895751953125,
0.66650390625
],
[
0.1748046875,
0.6767578125,
0.826171875,
0.6767578125,
0.826171875,
0.68994140625,
0.1748046875,
0.68994140625
],
[
0.191162109375,
0.68994140625,
0.6357421875,
0.68994140625,
0.6357421875,
0.7041015625,
0.191162109375,
0.7041015625
],
[
0.1732177734375,
0.7119140625,
0.826171875,
0.71337890625,
0.826171875,
0.72802734375,
0.1732177734375,
0.7265625
],
[
0.191162109375,
0.72802734375,
0.53173828125,
0.72802734375,
0.53173828125,
0.7421875,
0.191162109375,
0.7421875
],
[
0.1748046875,
0.751953125,
0.826171875,
0.751953125,
0.826171875,
0.76513671875,
0.1748046875,
0.76513671875
],
[
0.1895751953125,
0.7666015625,
0.55615234375,
0.7666015625,
0.55615234375,
0.77978515625,
0.1895751953125,
0.77978515625
],
[
0.1748046875,
0.78857421875,
0.82275390625,
0.78857421875,
0.82275390625,
0.80322265625,
0.1748046875,
0.80322265625
],
[
0.188720703125,
0.80322265625,
0.65185546875,
0.8017578125,
0.65185546875,
0.81640625,
0.188720703125,
0.8173828125
],
[
0.1748046875,
0.8271484375,
0.82177734375,
0.8271484375,
0.82177734375,
0.84033203125,
0.1748046875,
0.84033203125
],
[
0.191162109375,
0.84130859375,
0.82177734375,
0.84130859375,
0.82177734375,
0.85498046875,
0.191162109375,
0.85498046875
],
[
0.191162109375,
0.85498046875,
0.351318359375,
0.85498046875,
0.351318359375,
0.86865234375,
0.191162109375,
0.86865234375
],
[
0.1732177734375,
0.876953125,
0.82177734375,
0.87744140625,
0.82177734375,
0.892578125,
0.1732177734375,
0.8916015625
],
[
0.188720703125,
0.892578125,
0.81640625,
0.892578125,
0.81640625,
0.90576171875,
0.188720703125,
0.90576171875
],
[
0.49267578125,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"References",
"Ganin, Yaroslav, & Lempitsky, Victor. 2014. Unsupervised Domain Adaptation by Backpropaga-",
"tion.ArXive-prints.Sept",
"Goodfellow,Ian.Pouget-Abadie.Jean.Mirza,Mehdi,Xu.Bing,Warde-Farley.David.Ozair.Sher",
"jil, Courville, Aaron, & Bengio, Yoshua.2014. Generative Adversarial Nets. Pages 2672-2680 of.",
"Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N. D., & Weinberger, K. Q. (eds), Advances",
"in Neural Information Processing Systems 27.Curran Associates, Inc.",
"Hinton, Geoffrey, Deng, Li, Yu, Dong, Dahl, George E, Mohamed, Abdel-rahman, Jaitly, Navdeep.",
"Senior, Andrew, Vanhoucke, Vincent, Nguyen, Patrick, Sainath, Tara N, et al. . 2012. Deep neural",
"networksforacousticmodelinginspeechrecognition:Thesharedviewsoffourresearchgroups",
"IEEE Signal Processing Magazine, 29(6), 82-97.",
"Kalinli, Ozlem, Seltzer, Michael L, Droppo, Jasha, & Acero, Alex. 2010. Noise adaptive training",
"for robust automatic speech recognition. IEEE Transactions on Audio, Speech, and Language",
"Processing,18(8), 1889-1901.",
"Miao, Yajie, Gowayyed, Mohammad, & Metze, Florian. 2015. EESEN: End-to-end speech recog",
"nition using deep RNN models and WFST-based decoding. Pages 167-174 of: 2015 IEEE Work-",
"shop on Automatic Speech Recognition and Understanding (ASRU). IEEE.",
"Morgan,Nelson,&Bourlard,Herve.1995.Continuous speechrecognition.IEEE signal processing",
"magazine,12(3),2442.",
"Parihar,N, & Picone, J.2002. Aurora working group: DSR front end LVCSR evaluation AU/385/02.",
"Inst. for Signal and Information Process, Mississippi State University, Tech. Rep, 40, 94.",
"Radford, Alec,Metz, Luke, & Chintala, Soumith.2015.Unsupervised representation learning with",
"deep convolutional generative adversarial networks: arXiv preprint arXiv:1511.06434.",
"Sainath, Tara N, Kingsbury, Brian, Ramabhadran, Bhuvana, Fousek, Petr, Novak, Petr, & Mohamed,",
"Abdel-rahman.20ll.Making deep belief networks effective for large vocabulary continuous",
"speech recognition.Pages 30-35 of:Automatic Speech Recognition and Understanding (ASRU),",
"2011 IEEE Workshop on. IEEE.",
"Sainath, Tara N, Weiss, Ron J, Senior, Andrew, Wilson, Kevin W, & Vinyals, Oriol. 2015. Learning",
"the speech front-end with raw waveform CLDNNs. In: Proc. Interspeech.",
"Salimans, Tim, Goodfellow, Ian, Zaremba, Wojciech, Cheung, Vicki, Radford, Alec, & Chen, Xi.",
"2016. Improved techniques for training GANs. arXiv preprint arXiv:1606.03498.",
"Saon, George, Soltau,Hagen, Nahamoo, David, & Picheny, Michael. 2013. Speaker adaptation of",
"neural network acoustic models using i-vectors. Pages 55-59 of: ASRU.",
"Seide, Frank, Li, Gang, & Yu, Dong. 20ll., Conversational Speech Transcription Using Context-",
"Dependent Deep Neural Networks. Pages 437-440 of: Interspeech.",
"Shunohara, Yusuke.2016.Adversarial Multi-task Learning of Deep Neural Networks for Robust",
"Speech Recognition. Interspeech 2016, 2369-2372.",
"Theano Development Team. 2016. Theano: A Python framework for fast computation of mathemat-",
"ical expressions. arXiv e-prints, abs/1605.02688(May).",
"Un, Chong Kwan, Kim, Nam Soo, et al. - 1998. Speech recognition in noisy environments using",
"first-order vector Taylor series. Speech Communication, 24(1), 39-49.",
"van Merrienboer, Bart, Bahdanau, Dzmitry, Dumoulin, Vincent, Serdyuk, Dmitriy, Warde-Farley",
"David, Chorowski, Jan, & Bengio, Yoshua.2015.Blocks and Fuel: Frameworks for deep learning.",
"CoRR, abs/1506.00619.",
"Yu, Dong,Seltzer,Michael L, Li, Jinyu,Huang,Jui-Ting,& Seide, Frank.2013.Feature learning",
"in deep neural networks - studies on speech recognition tasks. arXiv preprint arXiv:1301.3605.",
"5"
] |
{
"arxiv_id": "1612.01928",
"language": "zh",
"page_idx": 4
}
|
||
2304.08158-zh-1
|
2304.08158-zh-1.jpg
|
SIGIR'23, July 23-27, 2023, Taipei, Taiwan
(1)“Long-term”preferences: this term designates intrinsic user
preferences, independent of the current context [40]
(2)“Short-term”preferences: on the contrary, this term refers
to how the context affects preferences, including how recent interactions alter the perception of a recommendation [1,8]
近年来，已经提出了几个 SR 系统来对这两个方面进行建模， 通常建立在深度学习的基础上用于序列建模$\overline{\mathbb{E}}[8,9,12,13,16$, 19,30,37,50,51,53,54] 。本文主要研究用于 SR 的变压器 [38] 。使用注意力机制$\mathbb{E}[38]$ 进行序列建模，他们最近在 SR 问题$\overline{\mathrm{F}}$ [5,7,14,16,33,41,43,45,48,51]上显示出具有竞争力的结果
2.2.1 Transformers for SR. SASRec [16] 是第一个利用自我关
注来识别 SR 时间序列中相关项目的系统。BERT4Rec[33] 随后实现了双向自注意技术。为了更仔细地结合长期和短期偏好， AttRec [51] 采用了多任务度量学习框架，而 SSE-PT [41] 将用户“嵌人”向量表示与序列中的每个项目。最近，FISSA$\underline{\mathrm{E}}[14]$ 结合了 Factored Item Similarity Model (FISM) [15] 的专心版本， 以满足长期用户的喜好，并使用基于项目相似性的门控机制来合并短期和长期表示，并具有良好的性能。最近的其他工作还将 SR 的注意力机制与图神经网络$\mathbb{E}[27,42]$ 或基于对比学习$\mathbb{E}$ [7,20]的模型相结合，研究了基于公平意识的注意力 SR $\boxed{\mathrm{E}}[18]$ , 并提出了更有效的训练策略$\mathbb{E}[26]$
2.2.2 Transformers for Time-Aware SR. 尽管取得了成功，但这
些变形金刚也因没有充分利用时间因素[6,10,19,44,49]而受到批评。虽然他们的注意力机制利用交互序列中的相对位置进行偏好建模 ,但他们忽略了每个交互的时间上下文，这些上下文由$\mathbb{E}$2.1 节中的$C_ui$ 序列捕获。然而，如引言中所示，与时间相关的因素，如当前月，日或小时，可以显着修改用户需求和偏好[4,10,49]
尝试解决此问题，TiSASRec$\boxed{\mathrm{E}}[19]$通过使用自我注意操作分析交互之间的时间间隔来增强 SASRec。TASER[49]在同一模型中组合成对相对和逐点绝对时间模式。MEANTIME[6] 利用多个自我关注的头来学习几种嵌人表示，对不同的时间模式进行编码。TimelyRec [44]通过处理分层时间粒度对用户行为中的异构时间模式进行建模。最近，CARCA[F[29] 依靠通过交叉注意计算的上下文相关性来提供时间感知建议。尽管如此我们认为时间感知 SR 仍然没有得到完全解决。虽然现有的时间感知转换器使用简单的连接或添加来收集项目和时间表示， 但这种方法过于简化了用户偏好和时间上下文之间的复杂关系。此外，除了 MEANTIME$\boxed{\mathrm{E}}[6]$之外，上述方法还存在多余的注意力问题[2,22,39] 。本文的其余部分将强调，克服这些限制可以更好地模型偏好并从经验上改善$\overline{\mathrm{E}}$时间感知ESR
# з А МIХТURE SYSTEM FOR TIME-AWARE SR 3.1 General Overview
本节 3个绍 MOJITO。如图 1 所示 , 我们的系统涉及两个组件。第一个模型在 3.2 节中进行了描述，它对受相互作用和时间上下文影响的短期意图进行了建模。第二个在 3.3 节中介绍，它捕获了长期偏好。每个组件返回一个相关性分数，分别表示为$r_{L+1}^\mathrm{short}(v)$和$r_L+1^\mathrm{long}(v)$。他们估计每个项目$v\in\mathcal{V}$作为观察到的相互作用序列的扩展的可能性。为了选择要推荐的项目，我们将分数组合如下，对于一些超参数$\lambda\in[0,1]:$
$$r_{L+1}(v)=\lambda r_{L+1}^{\mathrm{short}}(v)+(1-\lambda)r_{L+1}^{\mathrm{long}}(v).$$
(1)
VA. Tran, et al.
3.2 $\textbf{Short- Term Representation}$
我们利用自我注意机制[16,38]进行时间感知的短期表示学习。由于已知这种方法会遇到冗余头$^{2}$问题，其中不同的注意力头实际上会学习相同的信息[2,22,39],因此我们以 Nguyen 等人最近的努力为基础。[23] ,使我们的注意力能够通过共享的全球注意力矩阵池进行交互。Nguyen 等人。[23] 使用混合模型[3] 对每个头部的局部注意力矩阵进行采样。在我们的工作中，我们通过引人不同的归纳语义偏差来指导它们的学习，从而增强了全局矩阵表示的多样性，如下所述
3.2.1 Embedding Layer.此第一个块处理来自节 2.1 的$S_{ui}$和$C_{ui}$序列。我们用$M^I\in\mathbb{R}^{|\mathcal{V}|\times d}$表示一个可学习的项目嵌入矩
阵，其中的行是表示每个项目的$d$ 维嵌人向量，对于某些维度$d\in\mathbb{N}^*$。使用这种表示法，可以将序列 $S_ui$ 表示为嵌入矩阵， 如 下 所 示 : $E_{S_{ui}}^{I}= [ \boldsymbol{m}_{v_{1}}^{I}, \boldsymbol{m}_{v_{2}}^{I}, \ldots , \boldsymbol{m}_{v_{L}}^{I}] ^{\intercal }\in \mathbb{R} ^{L\times d}$ 如 下 所 示 : $E_{S_{ui}}^{I}= [ \boldsymbol{m}_{v_{1}}^{I}, \boldsymbol{m}_{v_{2}}^{I}, \ldots , \boldsymbol{m}_{v_{L}}^{I}] ^{\intercal }\in \mathbb{R} ^{L\times d}$
我们利用 Xu 等人的平移不变时间内核。[47]学习$d$维兼容自注意操作的时间上下文嵌人向量。具体来说，使用 Mercer 内核函数[46]将每个$c_ti\in\boldsymbol{c}_t$ 映射到向量 $\boldsymbol m_{c_{ti}}^C\in\mathbb{R}^d$。我们将这些向量连接并注入到线性层函数$g(\cdot)$中，以获得每个上下文元组$\boldsymbol c_t$的嵌入表示 $\boldsymbol m_{\mathbf{c}}_t^C=g([\boldsymbol{m}_{c_{t1}}^{C};\boldsymbol{m}_{c_{t2}}^{C};\ldots;\boldsymbol{m}_{c_{tC}}^{C}])\in\mathbb{R}^{d}$。我们获得了一个项目/上下文嵌人矩阵，该矩阵结合了项目和上下 文 嵌 人 : $E_{S_{ui}, C_{ui}}= [ [ \boldsymbol{m}_{v_{1}}^{I}; \boldsymbol{m}_{c_{1}}^{C}] ; \ldots ; [ \boldsymbol{m}_{v_{L}}^{I}; \boldsymbol{m}_{c_{L}}^{C}] ] ^{\intercal }\in \mathbb{R} ^{L\times 2d}$ 下 文 嵌 入 : $E_{S_{ui}, C_{ui}}= [ [ \boldsymbol{m}_{v_{1}}^{I}; \boldsymbol{m}_{c_{1}}^{C}] ; \ldots ; [ \boldsymbol{m}_{v_{L}}^{I}; \boldsymbol{m}_{c_{L}}^{C}] ] ] ^{\intercal }\in \mathbb{R} ^{L\times 2d}$
最 后 , 为 了 对 序 列 中 位 置 的 影 响 进 行 建 模 , 我 们 使 用 可 学 的 位 置 嵌 入 $P= [ \boldsymbol{n}_{s}- \boldsymbol{n}_{s}] ^{\mathrm{T} }\in \mathbb{R} ^{L\times 2d}$ 主 富 了 $E_{s}\subset C$ 习 的 位 置 嵌 入 $P= [ \boldsymbol{p}_{1}, \ldots , \boldsymbol{p}_{L}] ^{\mathrm{T} }\in \mathbb{R} ^{L\times 2d}$ 丰 富 了 $E_{Sui, C_{ui}}$, 获 得 了 最 终 的 输 人 矩 阵 $X^{( 0) }= [ x_{1}^{( 0) }, \ldots , x_{L}^{( 0) }] ^{\mathrm{T} }\in \mathbb{R} ^{L\times 2d}$, 其 中$x_{l}^{(0)}=[m_{vl}^{I};m_{cl}^{C}]+p_{l},\forall l\in\{1,\ldots,L\}$。我们旨在区分同一项目在不同位置的不同影响
3.2.2 Attention Mixtures.我们将$X^{(0)}$传递给$B\in\mathbb{N}^*$堆叠式
自关注块(SAB)$\operatorname{E}[38]$。对于$b\in\{1,\ldots,B\}$ , $b^\mathrm{~th}$ 块的输出为$X^{(b)}=$SAB$^(b)(X^{(b-1)})$。SAB 包含一个带有$H\in\mathbb{N}^*$头的自我关注层 SAL(·),混合了上下文和项目嵌入表示，然后是前馈层$\boxed{\mathrm{FFFL}}(\cdot):$
$$\mathrm{SAL}(X)=\mathrm{MultiHead}(\{X_{j}^{\mathrm{Att}}\}_{j=1}^{H})=\mathrm{Concat}(X_{1}^{\mathrm{Att}},\ldots,X_{H}^{\mathrm{Att}})W^{O},\\\mathrm{SAB}(X)=\mathrm{FFL}(\mathrm{SAL}(X))=ReLU(X^{\mathrm{Att}}W_{1}+b_{1})W_{2}+b_{2},\quad(2)$$
其中$X\in\mathbb{R}^L\times2d$是每个块的输入$,W^O\in\mathbb{R}^{2Hd\times2d}$是输出的投影矩阵，$W_1,W_2\in\mathbb{R}^{2d\times2d}$和$b_1,\boldsymbol{b}_2\in\mathbb{R}^{1\times2d}$是 FFL 网络两层的权重和偏置。$X_j^\mathrm{Att}=$softmax$(A_j/\sqrt{d})V$是每个磁头的输出。在这项工作中，我们从高斯混合概率模型中对$A_j$进行采样， 自动学习时间上下文和先前项目交互的相对重要性以进行顺序建模：
$$A_{j}\sim\sum_{k\in\{it,c\}}p_{kj}N(A;Q_{k}K_{k}^{\intercal},\sigma_{k}^{2}),\sum_{k\in\{it,c\}}p_{kj}=1,p_{kj}\geq0,\:(3)$$
其中$Q_k= X_kW_{Q}^{k}$ , $K_{k}= X_{k}W_{K}^{k}$ 和$V=XW_V$,以及$W_Q^{k},W_{K}^{k}\in$ $\mathbb{R}^{d\times d}$和$W_V\in\tilde{\mathbb{R}}^{\tilde{2}d\times2d}$是投影值矩阵。由模型学习的方差项$\sigma_k^2$ 按元素方式应用于相应注意力矩阵的每个元素
3.2.3 Prediction.我们使用潜在因子模型$\mathbb{E}[16,19]$ 来估计在观察到的上下文$c_{L+1}$中任何$v\in\mathcal{V}$扩展$S_{ui}$的相关性。我们
$^{2}$We refer, for instance, to the experiments of Kang and McAuley [16], showing that increasing the number of heads in SASRec does not improve sequential recommenda tion.
|
[
[
0.08660888671875,
0.07574462890625,
0.283447265625,
0.0770263671875,
0.283447265625,
0.08837890625,
0.08660888671875,
0.08709716796875
],
[
0.83642578125,
0.0770263671875,
0.90673828125,
0.0770263671875,
0.90673828125,
0.08837890625,
0.83642578125,
0.08837890625
],
[
0.10455322265625,
0.10797119140625,
0.48193359375,
0.10797119140625,
0.48193359375,
0.121826171875,
0.10455322265625,
0.121826171875
],
[
0.51708984375,
0.106689453125,
0.78515625,
0.106689453125,
0.78515625,
0.12060546875,
0.51708984375,
0.12060546875
],
[
0.1258544921875,
0.121826171875,
0.450927734375,
0.121826171875,
0.450927734375,
0.1351318359375,
0.1258544921875,
0.1351318359375
],
[
0.51953125,
0.126953125,
0.9111328125,
0.126953125,
0.9111328125,
0.136962890625,
0.51953125,
0.136962890625
],
[
0.10540771484375,
0.136962890625,
0.48046875,
0.136962890625,
0.48046875,
0.147705078125,
0.10540771484375,
0.147705078125
],
[
0.51953125,
0.141357421875,
0.91015625,
0.141357421875,
0.91015625,
0.1514892578125,
0.51953125,
0.1514892578125
],
[
0.1258544921875,
0.149658203125,
0.48046875,
0.1483154296875,
0.48046875,
0.1622314453125,
0.1258544921875,
0.1634521484375
],
[
0.51953125,
0.1546630859375,
0.9111328125,
0.1546630859375,
0.9111328125,
0.1654052734375,
0.51953125,
0.1654052734375
],
[
0.125,
0.1641845703125,
0.48193359375,
0.1641845703125,
0.48193359375,
0.177978515625,
0.125,
0.177978515625
],
[
0.51953125,
0.1666259765625,
0.91259765625,
0.1666259765625,
0.91259765625,
0.179931640625,
0.51953125,
0.179931640625
],
[
0.51953125,
0.179931640625,
0.91259765625,
0.179931640625,
0.91259765625,
0.1932373046875,
0.51953125,
0.1932373046875
],
[
0.0882568359375,
0.1962890625,
0.47802734375,
0.1962890625,
0.47802734375,
0.20703125,
0.0882568359375,
0.20703125
],
[
0.51953125,
0.1956787109375,
0.9111328125,
0.1956787109375,
0.9111328125,
0.205810546875,
0.51953125,
0.205810546875
],
[
0.0882568359375,
0.210205078125,
0.47802734375,
0.210205078125,
0.47802734375,
0.2203369140625,
0.0882568359375,
0.2203369140625
],
[
0.51708984375,
0.20703125,
0.91259765625,
0.208984375,
0.91259765625,
0.222900390625,
0.51708984375,
0.2203369140625
],
[
0.08660888671875,
0.222900390625,
0.48193359375,
0.222900390625,
0.48193359375,
0.236083984375,
0.08660888671875,
0.236083984375
],
[
0.51953125,
0.2235107421875,
0.79833984375,
0.2235107421875,
0.79833984375,
0.2342529296875,
0.51953125,
0.2342529296875
],
[
0.0882568359375,
0.238037109375,
0.48193359375,
0.238037109375,
0.48193359375,
0.248779296875,
0.0882568359375,
0.248779296875
],
[
0.5185546875,
0.2442626953125,
0.91259765625,
0.2442626953125,
0.91259765625,
0.25830078125,
0.5185546875,
0.25830078125
],
[
0.08660888671875,
0.251953125,
0.461669921875,
0.251953125,
0.461669921875,
0.261962890625,
0.08660888671875,
0.261962890625
],
[
0.51708984375,
0.255615234375,
0.9140625,
0.255615234375,
0.9140625,
0.2734375,
0.51708984375,
0.2734375
],
[
0.08660888671875,
0.272705078125,
0.48046875,
0.272705078125,
0.48046875,
0.285888671875,
0.08660888671875,
0.285888671875
],
[
0.5185546875,
0.272705078125,
0.91259765625,
0.272705078125,
0.91259765625,
0.285888671875,
0.5185546875,
0.285888671875
],
[
0.0849609375,
0.285400390625,
0.48046875,
0.285888671875,
0.48046875,
0.2998046875,
0.0849609375,
0.298583984375
],
[
0.51708984375,
0.287353515625,
0.9140625,
0.287353515625,
0.9140625,
0.300537109375,
0.51708984375,
0.300537109375
],
[
0.08660888671875,
0.2998046875,
0.4794921875,
0.2998046875,
0.4794921875,
0.313232421875,
0.08660888671875,
0.313232421875
],
[
0.515625,
0.2998046875,
0.814453125,
0.29736328125,
0.814453125,
0.31494140625,
0.515625,
0.317626953125
],
[
0.08660888671875,
0.313232421875,
0.4794921875,
0.313232421875,
0.4794921875,
0.326904296875,
0.08660888671875,
0.326904296875
],
[
0.53173828125,
0.31640625,
0.91259765625,
0.317626953125,
0.91259765625,
0.330810546875,
0.53173828125,
0.330078125
],
[
0.08660888671875,
0.32568359375,
0.48046875,
0.326904296875,
0.48046875,
0.341552734375,
0.08660888671875,
0.34033203125
],
[
0.51953125,
0.330810546875,
0.91259765625,
0.330810546875,
0.91259765625,
0.3447265625,
0.51953125,
0.3447265625
],
[
0.08660888671875,
0.341552734375,
0.48046875,
0.341552734375,
0.48046875,
0.354736328125,
0.08660888671875,
0.354736328125
],
[
0.5185546875,
0.3447265625,
0.91259765625,
0.343505859375,
0.91259765625,
0.35791015625,
0.5185546875,
0.359130859375
],
[
0.08660888671875,
0.35595703125,
0.48046875,
0.35595703125,
0.48046875,
0.369384765625,
0.08660888671875,
0.369384765625
],
[
0.5185546875,
0.359130859375,
0.91259765625,
0.359130859375,
0.91259765625,
0.37255859375,
0.5185546875,
0.37255859375
],
[
0.08905029296875,
0.37060546875,
0.48046875,
0.37060546875,
0.48046875,
0.380615234375,
0.08905029296875,
0.380615234375
],
[
0.51953125,
0.373779296875,
0.91259765625,
0.373779296875,
0.91259765625,
0.386962890625,
0.51953125,
0.386962890625
],
[
0.08660888671875,
0.38330078125,
0.48046875,
0.38330078125,
0.48046875,
0.396484375,
0.08660888671875,
0.396484375
],
[
0.5185546875,
0.38818359375,
0.91259765625,
0.38818359375,
0.91259765625,
0.401611328125,
0.5185546875,
0.401611328125
],
[
0.0882568359375,
0.3984375,
0.4794921875,
0.3984375,
0.4794921875,
0.4091796875,
0.0882568359375,
0.4091796875
],
[
0.51708984375,
0.40087890625,
0.89892578125,
0.397705078125,
0.89892578125,
0.416748046875,
0.51708984375,
0.419189453125
],
[
0.08416748046875,
0.410400390625,
0.31787109375,
0.410888671875,
0.31787109375,
0.4248046875,
0.08416748046875,
0.423583984375
],
[
0.53515625,
0.416748046875,
0.91259765625,
0.416748046875,
0.91259765625,
0.429931640625,
0.53515625,
0.429931640625
],
[
0.515625,
0.4248046875,
0.91650390625,
0.4267578125,
0.91650390625,
0.448974609375,
0.515625,
0.4462890625
],
[
0.08416748046875,
0.431884765625,
0.48046875,
0.43310546875,
0.48046875,
0.4462890625,
0.08416748046875,
0.44580078125
],
[
0.08660888671875,
0.4462890625,
0.48046875,
0.4462890625,
0.48046875,
0.460205078125,
0.08660888671875,
0.460205078125
],
[
0.5146484375,
0.44140625,
0.91650390625,
0.44140625,
0.91650390625,
0.46533203125,
0.5146484375,
0.46533203125
],
[
0.77392578125,
0.449462890625,
0.779296875,
0.453857421875,
0.77392578125,
0.458251953125,
0.7685546875,
0.453857421875
],
[
0.0882568359375,
0.462158203125,
0.4794921875,
0.462158203125,
0.4794921875,
0.472900390625,
0.0882568359375,
0.472900390625
],
[
0.5146484375,
0.458984375,
0.9140625,
0.46337890625,
0.9140625,
0.484130859375,
0.5146484375,
0.479736328125
],
[
0.08660888671875,
0.474853515625,
0.48046875,
0.474853515625,
0.48046875,
0.488037109375,
0.08660888671875,
0.488037109375
],
[
0.51708984375,
0.48095703125,
0.66845703125,
0.48095703125,
0.66845703125,
0.494384765625,
0.51708984375,
0.494384765625
],
[
0.0882568359375,
0.488037109375,
0.48046875,
0.488037109375,
0.48046875,
0.501953125,
0.0882568359375,
0.501953125
],
[
0.08660888671875,
0.501953125,
0.48046875,
0.501953125,
0.48046875,
0.51513671875,
0.08660888671875,
0.51513671875
],
[
0.5185546875,
0.50390625,
0.91259765625,
0.50390625,
0.91259765625,
0.517578125,
0.5185546875,
0.517578125
],
[
0.0882568359375,
0.517578125,
0.19287109375,
0.517578125,
0.19287109375,
0.52783203125,
0.0882568359375,
0.52783203125
],
[
0.521484375,
0.517578125,
0.91259765625,
0.517578125,
0.91259765625,
0.53076171875,
0.521484375,
0.53076171875
],
[
0.101318359375,
0.52880859375,
0.4794921875,
0.52783203125,
0.4794921875,
0.54248046875,
0.101318359375,
0.54345703125
],
[
0.515625,
0.52783203125,
0.9140625,
0.52978515625,
0.9140625,
0.54736328125,
0.515625,
0.54541015625
],
[
0.08660888671875,
0.54345703125,
0.48046875,
0.54345703125,
0.48046875,
0.556640625,
0.08660888671875,
0.556640625
],
[
0.5185546875,
0.546875,
0.91259765625,
0.546875,
0.91259765625,
0.56005859375,
0.5185546875,
0.56005859375
],
[
0.08660888671875,
0.556640625,
0.48046875,
0.556640625,
0.48046875,
0.57080078125,
0.08660888671875,
0.57080078125
],
[
0.51953125,
0.56103515625,
0.59716796875,
0.56103515625,
0.59716796875,
0.57275390625,
0.51953125,
0.57275390625
],
[
0.0882568359375,
0.5712890625,
0.48193359375,
0.5712890625,
0.48193359375,
0.58544921875,
0.0882568359375,
0.58544921875
],
[
0.08660888671875,
0.58544921875,
0.48046875,
0.58544921875,
0.48046875,
0.5986328125,
0.08660888671875,
0.5986328125
],
[
0.52294921875,
0.58056640625,
0.90625,
0.58056640625,
0.90625,
0.59521484375,
0.52294921875,
0.59521484375
],
[
0.69384765625,
0.58544921875,
0.701171875,
0.59033203125,
0.6953125,
0.59521484375,
0.68896484375,
0.59033203125
],
[
0.08905029296875,
0.60107421875,
0.4794921875,
0.60107421875,
0.4794921875,
0.611328125,
0.08905029296875,
0.611328125
],
[
0.5185546875,
0.599609375,
0.896484375,
0.60107421875,
0.896484375,
0.61865234375,
0.5185546875,
0.6171875
],
[
0.89453125,
0.60595703125,
0.91015625,
0.60595703125,
0.91015625,
0.61572265625,
0.89453125,
0.61572265625
],
[
0.0882568359375,
0.61279296875,
0.4794921875,
0.61279296875,
0.4794921875,
0.626953125,
0.0882568359375,
0.626953125
],
[
0.0882568359375,
0.62744140625,
0.4794921875,
0.62744140625,
0.4794921875,
0.63818359375,
0.0882568359375,
0.63818359375
],
[
0.51953125,
0.626953125,
0.91259765625,
0.626953125,
0.91259765625,
0.6416015625,
0.51953125,
0.6416015625
],
[
0.0882568359375,
0.642578125,
0.4794921875,
0.642578125,
0.4794921875,
0.65283203125,
0.0882568359375,
0.65283203125
],
[
0.515625,
0.63818359375,
0.9140625,
0.6396484375,
0.9140625,
0.65771484375,
0.515625,
0.6572265625
],
[
0.0882568359375,
0.65478515625,
0.48046875,
0.65478515625,
0.48046875,
0.66845703125,
0.0882568359375,
0.66845703125
],
[
0.5185546875,
0.65576171875,
0.91259765625,
0.65576171875,
0.91259765625,
0.67041015625,
0.5185546875,
0.67041015625
],
[
0.0882568359375,
0.66748046875,
0.48046875,
0.66748046875,
0.48046875,
0.68115234375,
0.0882568359375,
0.68115234375
],
[
0.51708984375,
0.67236328125,
0.91015625,
0.673828125,
0.91015625,
0.6875,
0.51708984375,
0.68603515625
],
[
0.0849609375,
0.68115234375,
0.48193359375,
0.6796875,
0.48193359375,
0.693359375,
0.0849609375,
0.6943359375
],
[
0.51953125,
0.6875,
0.9140625,
0.6875,
0.9140625,
0.70068359375,
0.51953125,
0.70068359375
],
[
0.08660888671875,
0.69775390625,
0.455810546875,
0.69775390625,
0.455810546875,
0.70849609375,
0.08660888671875,
0.70849609375
],
[
0.5185546875,
0.7021484375,
0.56884765625,
0.7021484375,
0.56884765625,
0.71337890625,
0.5185546875,
0.71337890625
],
[
0.1168212890625,
0.724609375,
0.47802734375,
0.724609375,
0.47802734375,
0.73486328125,
0.1168212890625,
0.73486328125
],
[
0.529296875,
0.724609375,
0.5703125,
0.724609375,
0.5703125,
0.7392578125,
0.529296875,
0.7392578125
],
[
0.57666015625,
0.724609375,
0.59228515625,
0.7197265625,
0.59716796875,
0.72900390625,
0.58251953125,
0.734375
],
[
0.60693359375,
0.724609375,
0.75341796875,
0.724609375,
0.75341796875,
0.7392578125,
0.60693359375,
0.7392578125
],
[
0.74853515625,
0.72607421875,
0.765625,
0.72021484375,
0.76953125,
0.72998046875,
0.75244140625,
0.734375
],
[
0.77783203125,
0.72607421875,
0.9111328125,
0.724609375,
0.9111328125,
0.73779296875,
0.77783203125,
0.7392578125
],
[
0.60107421875,
0.72900390625,
0.60888671875,
0.72900390625,
0.60888671875,
0.73486328125,
0.60107421875,
0.73486328125
],
[
0.08660888671875,
0.72607421875,
0.11517333984375,
0.72607421875,
0.11517333984375,
0.734375,
0.08660888671875,
0.734375
],
[
0.580078125,
0.73095703125,
0.59716796875,
0.73095703125,
0.59716796875,
0.7392578125,
0.580078125,
0.7392578125
],
[
0.75244140625,
0.73095703125,
0.76953125,
0.73095703125,
0.76953125,
0.7392578125,
0.75244140625,
0.7392578125
],
[
0.775390625,
0.73095703125,
0.78173828125,
0.73095703125,
0.78173828125,
0.736328125,
0.775390625,
0.736328125
],
[
0.56201171875,
0.73779296875,
0.611328125,
0.74072265625,
0.609375,
0.7548828125,
0.560546875,
0.75244140625
],
[
0.08660888671875,
0.744140625,
0.275390625,
0.744140625,
0.275390625,
0.7548828125,
0.08660888671875,
0.7548828125
],
[
0.73876953125,
0.7392578125,
0.78515625,
0.7412109375,
0.78369140625,
0.75244140625,
0.73779296875,
0.75048828125
],
[
0.08660888671875,
0.76123046875,
0.48046875,
0.76123046875,
0.48046875,
0.7744140625,
0.08660888671875,
0.7744140625
],
[
0.5185546875,
0.76220703125,
0.91259765625,
0.76220703125,
0.91259765625,
0.7763671875,
0.5185546875,
0.7763671875
],
[
0.0882568359375,
0.7763671875,
0.4794921875,
0.7763671875,
0.4794921875,
0.787109375,
0.0882568359375,
0.787109375
],
[
0.5146484375,
0.7763671875,
0.91259765625,
0.77783203125,
0.91259765625,
0.79541015625,
0.5146484375,
0.79443359375
],
[
0.0882568359375,
0.78857421875,
0.48046875,
0.78857421875,
0.48046875,
0.8017578125,
0.0882568359375,
0.8017578125
],
[
0.5185546875,
0.79736328125,
0.81201171875,
0.79736328125,
0.81201171875,
0.80810546875,
0.5185546875,
0.80810546875
],
[
0.08660888671875,
0.8037109375,
0.4794921875,
0.8037109375,
0.4794921875,
0.814453125,
0.08660888671875,
0.814453125
],
[
0.0882568359375,
0.8173828125,
0.4794921875,
0.8173828125,
0.4794921875,
0.83203125,
0.0882568359375,
0.83203125
],
[
0.5185546875,
0.818359375,
0.91259765625,
0.818359375,
0.91259765625,
0.83203125,
0.5185546875,
0.83203125
],
[
0.0882568359375,
0.83349609375,
0.48046875,
0.83349609375,
0.48046875,
0.8466796875,
0.0882568359375,
0.8466796875
],
[
0.5185546875,
0.83349609375,
0.91259765625,
0.83349609375,
0.91259765625,
0.8466796875,
0.5185546875,
0.8466796875
],
[
0.08660888671875,
0.84765625,
0.378173828125,
0.84765625,
0.378173828125,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.1683349609375,
0.8662109375,
0.398681640625,
0.8662109375,
0.398681640625,
0.88818359375,
0.1683349609375,
0.88818359375
],
[
0.51708984375,
0.861328125,
0.91259765625,
0.8623046875,
0.91259765625,
0.876953125,
0.51708984375,
0.87548828125
],
[
0.463134765625,
0.8720703125,
0.483642578125,
0.8720703125,
0.483642578125,
0.8837890625,
0.463134765625,
0.8837890625
],
[
0.5185546875,
0.875,
0.9111328125,
0.875,
0.9111328125,
0.88525390625,
0.5185546875,
0.88525390625
],
[
0.51708984375,
0.88525390625,
0.5439453125,
0.88525390625,
0.5439453125,
0.896484375,
0.51708984375,
0.896484375
]
] |
[
"SIGIR ‘23, July 2327, 2023, Taipei, Taiwan",
"V.A. Tran, et al",
"(1) \"Long-term\"preferences: this term designates intrinsic user",
"3.2Short-TermRepresentation",
"preferences, independent of the current context [40] .",
"我们利用自我注意机制「16，381进行时间感知的短期表示学习。",
"(2)\"Short-term\"preferences: on the contrary, this term refers",
"由于已知这种方法会遇到元余头2问题，其中不同的注意力头",
"to how the context affects preferences, including how recent",
"实际上会学习相同的信息[2,22,39]，因此我们以Nguyen等",
"interactions alter the perception of a recommendation [1, 8]",
"人最近的努力为基础。[23]，使我们的注意力能够通过共享的",
"全球注意力矩阵池进行交互。Nguyen等人。[23]使用混合模",
"近年来，已经提出了几个SR系统来对这两个方面进行建模",
"型[3]对每个头部的局部注意力矩阵进行采样。在我们的工作",
"通常建立在深度学习的基础上用于序列建模F[8.9,12.13.16",
"中，我们通过引入不同的归纳语义偏差来指导它们的学习，从",
"19,30,37,50,51,53,54]。本文主要研究用于SR的变压器[38】。",
"而增强了全局矩阵表示的多样性，如下所述",
"使用注意力机制E[38]进行序列建模，他们最近在SR问题E",
"3.2.1EmbeddingLayer。此第一个块处理来自节2.1的Sui和",
"[5,7，14，16,33,41，43，45,48,51】上显示出具有竞争力的结果",
"Cui序列。我们用MIeR/VIxd表示一个可学习的项目嵌人矩",
"2.2.1TransformersforSR.SASRec[16]是第一个利用自我关",
"阵，其中的行是表示每个项目的d维嵌人向量，对于某些维度",
"注来识别SR时间序列中相关项目的系统。BERT4Rec[33]随后",
"dEN*。使用这种表示法，可以将序列Sui表示为嵌人矩阵，",
"实现了双向自注意技术。为了更仔细地结合长期和短期偏好，",
"如下所示：ES=[m,m...m]TeRLxd",
"AttRec[51]采用了多任务度量学习框架，而SSE-PT[41]将用",
"我们利用Xu等人的平移不变时间内核。[47]学习d维兼容",
"户“嵌人”向量表示与序列中的每个项目。最近，FISSAE[14]",
"自注意操作的时间上下文嵌人向量。具体来说，使用Mercer",
"结合了FactoredItemSimilarityModel(FISM)[15]的专心版本，",
"内核函数[46]将每个ctiEct映射到向量mc,ERd。我们将",
"以满足长期用户的喜好，并使用基于项目相似性的门控机制来",
"这些向量连接并注人到线性层函数9(）中，以获得每个上下文",
"合并短期和长期表示，并具有良好的性能。最近的其他工作还",
"元组ct的嵌人表示mc=g（[mc;m,...;mcc）eRd。我",
"将SR的注意力机制与图神经网络国[27,42]或基于对比学习E",
"们获得了一个项目/上下文嵌人矩阵，该矩阵结合了项目和上",
"[7.20]的模型相结合，研究了基于公平意识的注意力SRE[18]",
"下文嵌人：Esu.Cu.=[mo;me]...[murm&,]TeRLx2d",
"，并提出了更有效的训练策略E[26]",
"最后，为了对序列中位置的影响进行建模，我们使用可学",
"习的位置嵌人P=[P1.L]TeRLx2d丰富了Esu.Cu，获",
"2.2.2TransformersforTime-AwareSR，尽管取得了成功，但这",
"些变形金刚也因没有充分利用时间因素[6,10,19,44,49]而受",
"得了最终的输人矩阵x(0)=[x(..(]Te RLx2a，其中",
"I",
"到批评。虽然他们的注意力机制利用交互序列中的相对位置",
"x()=[mm)]+Pi,VLe(1.,L)。我们旨在区分同一项目",
"进行偏好建模，但他们忽略了每个交互的时间上下文，这些上",
"在不同位置的不同影响",
"下文由E2.1节中的Cui序列捕获。然而，如引言中所示，与时",
"间相关的因素，如当前月，日或小时，可以显着修改用户需求",
"3.2.2AttentionMixtures.我们将X(0）传递给BeN*堆叠式",
"和偏好【4,10,49]",
"自关注块（SAB)E[38]。对于bE(1....,B)，bth块的输出为",
"尝试解决此问题，TiSASRecE[19]通过使用自我注意操作分",
"X(b)=SAB(b)(X(b-1)）。SAB包含一个带有HEN头的自我",
"析交互之间的时间间隔来增强SASRec。TASER[49]在同一模",
"关注层SAL(），混合了上下文和项目嵌入表示，然后是前馈",
"型中组合成对相对和逐点绝对时间模式。MEANTIME[6]利用",
"层FFL()：",
"多个自我关注的头来学习几种嵌人表示，对不同的时间模式",
"进行编码。TimelyRec[44]通过处理分层时间粒度对用户行为",
"SAL(X) =MultiHead((xAtt)H-,) = Concat(xAt..,XAtt)wO",
"",
"中的异构时间模式进行建模。最近，CARCAE[29]依靠通过交",
"SAB(X) = FFL(SAL(X) = ReLU(XAW1 +b)W2 + b2,",
"(2)",
"叉注意计算的上下文相关性来提供时间感知建议。尽管如此，",
"我们认为时间感知SR仍然没有得到完全解决。虽然现有的时",
"其中XERL×2d是每个块的输人，WOER2Hd×2d是输出的投",
"间感知转换器使用简单的连接或添加来收集项目和时间表示，",
"影矩阵，W1，W2eR2dx2d和bi,b2R1x2d是FFL网络两层",
"但这种方法过于简化了用户偏好和时间上下文之间的复杂关",
"的权重和偏置。xAtt=softmax(A;/V@)V是每个磁头的输出。",
"系。此外，除了MEANTIMEE[6]之外，上述方法还存在多余",
"在这项工作中，我们从高斯混合概率模型中对A，进行采样",
"的注意力问题[2,22,39]。本文的其余部分将强调，克服这些",
"自动学习时间上下文和先前项目交互的相对重要性以进行顺",
"限制可以更好地模型偏好并从经验上改善E时间感知ESR",
"序建模：",
"AMIXTURESYSTEMFORTIME-AWARESR",
"Aj~",
"T",
"PkjN(A;QkKT,o),",
"",
"pkj = 1, Pkj ≥ 0,(3)",
"",
"3",
"",
"1",
"",
"ke(it,c)",
"3.1GeneralOverview",
"ke(it,c)",
"本节3介绍MOJITO。如图1所示，我们的系统涉及两个组件。",
"其中Qk=X,k，K=Xwk和V=XWv以及wkwk",
"第一个模型在3.2节中进行了描述，它对受相互作用和时间上",
"Rdxd和WvE R2dx2d是投影值矩阵。由模型学习的方差项o？",
"下文影响的短期意图进行了建模。第二个在3.3节中介绍，它",
"按元素方式应用于相应注意力矩阵的每个元素",
"捕获了长期偏好。每个组件返回一个相关性分数，分别表示为",
"rshort(u）和rlong(u）。他们估计每个项目uE作为观察到的",
"3.2.3Prediction，我们使用潜在因子模型国[16,19]来估计在",
"相互作用序列的扩展的可能性。为了选择要推荐的项目，我们",
"观察到的上下文cL+1中任何ueV扩展Sui的相关性。我们",
"将分数组合如下，对于一些超参数入E[0,1]：",
"rL+1(0) = Arhor(0) + (1 - a)rlong(0).",
"2werfer,for instance, to the experiments ofKang and McAuley [16],showing that",
"(1)",
"increasing the number of heads in SASRec does not improve sequential recommenda",
"tion."
] |
{
"arxiv_id": "2304.08158",
"language": "zh",
"page_idx": 1
}
|
||
1311.7662-zh-1
|
1311.7662-zh-1.jpg
|
# 二进制哈希中不对称的力量
Behnam Neyshabur Payman Yadollahpour Yury Makarychev
Toyota Technological Institute at Chicago
[btavakoli,pyadolla,yury$]@$ttic.edu
Ruslan Salakhutdinov
Departments of Statistics and Computer Science
University of Toronto rsalakhu@cs.toronto.edu
Nathan Srebro
Toyota Technological Institute at Chicago
and Technion, Haifa, Israel
nati@ttic.edu
Abstract
When approximating binary similarity using the hamming distance between short binary hashes, we show that even if the similarity is symmetric, we can have shorter and more accurate hashes by using two distinct code maps. I.e. by approximating the similarity between $x$ and $x^\prime$ as the hamming distance between $f(x)$ and $g(x^\prime)$ , for two distinct binary codes $f,g$ , rather than as the hamming distance between $f(x)$ and $f( x^{\prime })$ .
## 1 Introduction
使用短二进制哈希对高维对象进行编码可用于快速近似相似度计算和最近邻搜索。计算两个短二进制字符串之间的汉明距离是一种非常便宜的计算操作，并且发送此类哈希字符串以在服务器上进行查找的通信成本(e。g。在移动设备上拍摄的图像中发送所有特征或补丁的哈希值)很低。此外，还可以在填充的哈希表中快速查找附近的哈希字符串。事实上，从包含数十亿个数据点的语料库中检索类似项目的短列表只需几分之一秒，这在图像、视频、音频和文档检索任务中很重要[11,9,10,13] 。此外，紧凑的二进制代码具有显着的存储效率，并允许将大量数据集存储在内存中。因此，希望找到与某些目标相似性概念很好地对应的短二进制哈希。位置敏感哈希的开创性工作使用随机线性阈值来获取哈希[1]的位。后来的工作建议学习哈希函数适应数据的分布[15,11,5,7,3] 。最近的工作集中在学习哈希函数，以便优化与特定数据集的目标相似性度量的一致性[14,8,9,6] 。获取准确和 short 哈希值很重要一计算和通信成本随哈希值的长度线性缩放，更重要的是，哈希表的内存成本可以随长度呈指数缩放
在所有上述方法中，两个对象之间的相似性$S(x,x^{\prime})$由同一哈希函数的输出之间的汉明距离近似，i。e。在$f(x)$和$f(x^\prime)$之间，对于一些$f\in\{\pm1\}^k$ 。这里的重点是相同的哈希函数应用于$x$和$x^\prime$ (在像 LSH 这样的方法中，可以使用多个哈希来提高准确性，但比较仍然在相同函数的输出之间)
我们知道的唯一例外是使用对象到分数向量$\tilde{f}(x)\in[-1,1]^k$的单个映射，在数据库中使用其阈值$f( x) = \tilde{f} ( x) \in \{ \pm 1\} ^{k}$ ,并且使用$f(x),\tilde{f}(x^{\prime})$近似$x$和$x^\prime$之间的相似性。这已被称为“非对称哈希”[2,4] , 但是即使具有这种对称性，两个映射也都基于相同的分数映射$\tilde{f}(\cdot)$ 。也就是说，不对称性在于比较的一侧被阈值化，而另一侧是分数，但不是在实际映射中
2
|
[
[
0.334228515625,
0.14013671875,
0.6650390625,
0.14013671875,
0.6650390625,
0.1578369140625,
0.334228515625,
0.1578369140625
],
[
0.2099609375,
0.1787109375,
0.38232421875,
0.18115234375,
0.38232421875,
0.1956787109375,
0.2099609375,
0.1932373046875
],
[
0.41015625,
0.179931640625,
0.60107421875,
0.179931640625,
0.60107421875,
0.1944580078125,
0.41015625,
0.1944580078125
],
[
0.6396484375,
0.18115234375,
0.7861328125,
0.18115234375,
0.7861328125,
0.1956787109375,
0.6396484375,
0.1956787109375
],
[
0.32275390625,
0.1962890625,
0.67626953125,
0.1976318359375,
0.67626953125,
0.21533203125,
0.32275390625,
0.2139892578125
],
[
0.38232421875,
0.184326171875,
0.408447265625,
0.184326171875,
0.408447265625,
0.1925048828125,
0.38232421875,
0.1925048828125
],
[
0.359375,
0.2171630859375,
0.64208984375,
0.2171630859375,
0.64208984375,
0.2310791015625,
0.359375,
0.2310791015625
],
[
0.408447265625,
0.2587890625,
0.5908203125,
0.2587890625,
0.5908203125,
0.272705078125,
0.408447265625,
0.272705078125
],
[
0.295654296875,
0.276611328125,
0.70166015625,
0.276611328125,
0.70166015625,
0.291748046875,
0.295654296875,
0.291748046875
],
[
0.41015625,
0.294189453125,
0.5888671875,
0.294189453125,
0.5888671875,
0.309326171875,
0.41015625,
0.309326171875
],
[
0.39794921875,
0.3125,
0.6005859375,
0.3125,
0.6005859375,
0.326904296875,
0.39794921875,
0.326904296875
],
[
0.43701171875,
0.354248046875,
0.56201171875,
0.354248046875,
0.56201171875,
0.36865234375,
0.43701171875,
0.36865234375
],
[
0.32275390625,
0.37060546875,
0.67333984375,
0.37255859375,
0.67333984375,
0.386962890625,
0.32275390625,
0.385009765625
],
[
0.3857421875,
0.39013671875,
0.61279296875,
0.39013671875,
0.61279296875,
0.404052734375,
0.3857421875,
0.404052734375
],
[
0.44287109375,
0.40771484375,
0.5556640625,
0.40771484375,
0.5556640625,
0.421630859375,
0.44287109375,
0.421630859375
],
[
0.4697265625,
0.455078125,
0.529296875,
0.455078125,
0.529296875,
0.4697265625,
0.4697265625,
0.4697265625
],
[
0.178955078125,
0.474853515625,
0.8408203125,
0.4765625,
0.8408203125,
0.490478515625,
0.178955078125,
0.488037109375
],
[
0.156005859375,
0.4892578125,
0.84228515625,
0.490478515625,
0.84228515625,
0.50390625,
0.156005859375,
0.50244140625
],
[
0.1568603515625,
0.50390625,
0.8388671875,
0.50390625,
0.8388671875,
0.517578125,
0.1568603515625,
0.517578125
],
[
0.156005859375,
0.5166015625,
0.84228515625,
0.517578125,
0.84228515625,
0.5322265625,
0.156005859375,
0.53076171875
],
[
0.1568603515625,
0.5322265625,
0.26708984375,
0.5322265625,
0.26708984375,
0.54541015625,
0.1568603515625,
0.54541015625
],
[
0.11273193359375,
0.5693359375,
0.28515625,
0.5712890625,
0.28515625,
0.5859375,
0.11273193359375,
0.583984375
],
[
0.1192626953125,
0.6015625,
0.8798828125,
0.6015625,
0.8798828125,
0.6123046875,
0.1192626953125,
0.6123046875
],
[
0.11767578125,
0.61572265625,
0.88134765625,
0.61572265625,
0.88134765625,
0.62890625,
0.11767578125,
0.62890625
],
[
0.1168212890625,
0.630859375,
0.88232421875,
0.630859375,
0.88232421875,
0.64453125,
0.1168212890625,
0.64453125
],
[
0.1168212890625,
0.64599609375,
0.86669921875,
0.64599609375,
0.86669921875,
0.6591796875,
0.1168212890625,
0.6591796875
],
[
0.11767578125,
0.662109375,
0.8798828125,
0.662109375,
0.8798828125,
0.67236328125,
0.11767578125,
0.67236328125
],
[
0.11767578125,
0.67529296875,
0.88134765625,
0.67529296875,
0.88134765625,
0.689453125,
0.11767578125,
0.689453125
],
[
0.1192626953125,
0.69140625,
0.88134765625,
0.69140625,
0.88134765625,
0.705078125,
0.1192626953125,
0.705078125
],
[
0.11767578125,
0.70556640625,
0.88134765625,
0.70556640625,
0.88134765625,
0.7197265625,
0.11767578125,
0.7197265625
],
[
0.11767578125,
0.72021484375,
0.88134765625,
0.72021484375,
0.88134765625,
0.734375,
0.11767578125,
0.734375
],
[
0.11767578125,
0.73486328125,
0.491943359375,
0.73486328125,
0.491943359375,
0.74853515625,
0.11767578125,
0.74853515625
],
[
0.1422119140625,
0.751953125,
0.88232421875,
0.751953125,
0.88232421875,
0.76513671875,
0.1422119140625,
0.76513671875
],
[
0.1168212890625,
0.7666015625,
0.88134765625,
0.7666015625,
0.88134765625,
0.77978515625,
0.1168212890625,
0.77978515625
],
[
0.1168212890625,
0.78076171875,
0.7099609375,
0.78076171875,
0.7099609375,
0.79443359375,
0.1168212890625,
0.79443359375
],
[
0.1422119140625,
0.796875,
0.88134765625,
0.796875,
0.88134765625,
0.81005859375,
0.1422119140625,
0.81005859375
],
[
0.1192626953125,
0.81103515625,
0.88134765625,
0.81103515625,
0.88134765625,
0.82470703125,
0.1192626953125,
0.82470703125
],
[
0.11767578125,
0.8271484375,
0.8798828125,
0.8271484375,
0.8798828125,
0.84033203125,
0.11767578125,
0.84033203125
],
[
0.1168212890625,
0.84228515625,
0.4755859375,
0.84228515625,
0.4755859375,
0.85595703125,
0.1168212890625,
0.85595703125
],
[
0.491943359375,
0.89794921875,
0.50732421875,
0.89794921875,
0.50732421875,
0.912109375,
0.491943359375,
0.912109375
]
] |
[
"二进制哈希中不对称的力量",
"Behnam Neyshabur",
" Payman Yadollahpour",
"Yury Makarychev",
"Toyota Technological Institute at Chicago",
"",
"btavakoli,pyadolla,yury@ttic.edu",
"Ruslan Salakhutdinov",
"Departments of Statistics and Computer Science",
"University of Toronto",
"rsalakhu@cs.toronto.edu",
"Nathan Srebro",
"Toyota Technological Institute at Chicago",
"and Technion, Haifa, Israel",
"nati@ttic.edu",
"Abstract",
"When approximatingbinary similarity using the hamming distancebetween short binaryhashes, we",
"show that even if the similarity is symmetric, we can have shorter and more accurate hashes by using",
"two distinct code maps. I.e. by approximating the similarity between and r' as the hamming distance",
"between f(r) and g(r'), for two distinct binary codes f,g, rather than as the hamming distance between",
"f(r) and f(r) :",
"1Introduction",
"使用短二进制哈希对高维对象进行编码可用于快速近似相似度计算和最近邻搜索。计算两个短二进制字符串",
"之间的汉明距离是一种非常便宜的计算操作，并且发送此类哈希字符串以在服务器上进行查找的通信成本",
"(e。g。在移动设备上拍摄的图像中发送所有特征或补丁的哈希值）很低。此外，还可以在填充的哈希表中快",
"速查找附近的哈希字符串。事实上，从包含数十亿个数据点的语料库中检索类似项目的短列表只需几分之",
"秒，这在图像、视频、音频和文档检索任务中很重要11，9.10，13。此外，紧凑的二进制代码具有显着的",
"存储效率，并允许将大量数据集存储在内存中。因此，希望找到与某些目标相似性概念很好地对应的短二进",
"制哈希。位置敏感哈希的开创性工作使用随机线性阈值来获取哈希[1]的位。后来的工作建议学习哈希函数",
"适应数据的分布[15，11，5，7,3]。最近的工作集中在学习哈希函数，以便优化与特定数据集的目标相似性度",
"量的一致性[14，8，9，6]。获取准确和short哈希值很重要一计算和通信成本随哈希值的长度线性缩放，更",
"重要的是，哈希表的内存成本可以随长度呈指数缩放",
"在所有上述方法中，两个对象之间的相似性S(,a）由同一哈希函数的输出之间的汉明距离近似，i。e。",
"在f(αc）和f(α）之间，对于一些fE（士1)。这里的重点是相同的哈希函数应用于r和α（在像LSH这",
"样的方法中，可以使用多个哈希来提高准确性，但比较仍然在相同函数的输出之间）",
"我们知道的唯一例外是使用对象到分数向量f(α）E［-1,1]的单个映射，在数据库中使用其值",
"f(α)=f(α)E{±1)k，并且使用f(α),f(a'）近似α和α'之间的相似性。这已被称为“非对称哈希”[2,4]",
"但是即使具有这种对称性，两个映射也都基于相同的分数映射f（）。也就是说，不对称性在于比较的一侧",
"被阈值化，而另一侧是分数，但不是在实际映射中",
"2"
] |
{
"arxiv_id": "1311.7662",
"language": "zh",
"page_idx": 1
}
|
||
1311.7662-zh-2
|
1311.7662-zh-2.jpg
|
在本文中，我们建议使用两个 distinct 映射$f(x),g(x)\in\{\pm1\}^k$ ,并通过$f(x)$和$g(x^\prime)$之间的汉明距离来近似相似度$S(x,x^{\prime})$ 。我们将这种哈希方案称为“非对称”。我们的主要结果是，即使目标相似度函数是对称的并且“表现良好”(e。g。,即使它基于对象之间的欧几里得距离),使用非对称二进制哈希可以强大得多，并允许以较短的代码长度更好地近似目标相似性。特别是，我们显示了欧几里得空间中的点集合的极端示例，其中可以使用长度为$O(r)$位的非对称二进制哈希 (基于一对不同的函数)来实现邻域相似性$S(x,x^{\prime})$ ,但是对称哈希 (基于单个函数) 将需要至少$\Omega(2^r)$位。尽管实际数据并不极端，但我们在真实数据集上的实验结果表明，使用非对称二进制哈希值具有显着优势
非对称哈希可以在几乎所有通常使用对称哈希的地方便用，通常无需任何额外的存储或计算成本。考虑将所有对象的哈希向量存储在数据库中，然后通过计算查询的哈希及其与存储的数据库哈希的汉明距离来计算与查询的相似性的典型应用。使用非对称哈希意味着对数据库和查询使用不同的哈希函数。这既不会增加数据库表示的大小，也不会增加填充数据库或执行查询的计算或通信成本，因为需要完全相同的操作。事实上，当哈希整个数据库时，非对称哈希提供了更多的改进机会。我们认为，使用两个不同的哈希函数对数据库对象和查询进行编码可以在选择数据库哈希时提供更大的灵活性。与查询哈希不同，查询哈希必须紧凑地存储并在查询出现时对其进行有效评估，如果数据库是固定的，则可以使用数据库对象到位串的任意映射。我们证明，这确实可以提高相似性精度，同时减少所需的位长。
2 Minimum Code Lengths and the Power of Asymmetry
令$S:\mathcal{X}\times\mathcal{X}\to\{\pm1\}$是一组对象$\mathcal{X}$上的二进制相似性函数，其中我们可以将$S(x,x^\prime)$解释为表示$x$和$x^\prime$ 是“相似”还是“不相似”,或者指示它们是否是“邻居”。$\mathcal{X}$的对称二进制编码是映射$f:\mathcal{X}\to\{\pm1\}^k$ ,其中$k$是代码的位长度。我们有兴趣构造代码，使得$f(x)$和$f(x^\prime)$之间的汉明距离对应于相似度$S(x,x^\prime)$ 。也就是说，对于某个阈值$\theta\in,S(x,x^{\prime})\approx(f(x),f(x^{\prime})-\theta)$ 。虽然讨论了汉明距离，但我们使用内积$u,v$更方便，这相当于自$u,v=(k-2d_h(u,v))$以来$u,v\in\{\pm1\}^k$的汉明距离$d_h(u,v)$
在本节中，我们将考虑使用任意二进制代码捕获给定相似性的问题。也就是说，我们给出了整个相似性映射$S$ , e。g。$\mathbb{E}$作为$n$对象的有限域$\mathcal{X}=\{x_1,\ldots,x_n\}$上的矩阵$S\in\{\pm1\}^n\times n$ ,带有$S_ij=S(x_i,x_j)$ 。我们要求每个对象$x_i\in\mathcal{X}$的编码$u_i=f(x_i)\in\{\pm1\}^k$和阈值$\theta$,使得$S_ij=(u_i,u_j-\theta)$,或者至少使得尽可能多的对$(i,j)$相等。需要强调的是，这里的目标纯粹是使用短二进制码来近似给定的矩阵$S$一不存在样本外泛化(还)
我们现在问：允许不对称编码是否能够以更短的代码长度近似对称相似性矩阵$S$ ?
为其列包含码字$u_i$的矩阵表示$U\in\{\pm1\}^n\times k$ ,然后通过以下矩阵分解问题给出允许精确表示$S$的最小
二进制码长度：
(1)
$$\begin{aligned}k_{s}(S)=&\min_{k,U,\theta}k\quad\mathrm{s.t}\quad U\in\{\pm1\}^{k\times n}\quad\theta\in\\&YU^{\top}U-\theta1_{n}\\&\forall_{ij}\:S_{ij}Y_{ij}>0\end{aligned}$$
其中$1_n$是$n\times n$矩阵
我们开始通过考虑上述问题的 asymmetric 变体来证明不对称的力量。也就是说，即使$S$是对称的，我们也允许将每个对象$x_i$关联两个不同的二进制码字$u_i\in\{\pm1\}^k$和$v_i\in\{\pm1\}^k$ (我们可以将其视为具有两个任意映射$u_i=f(x_i)$和$v_i=g(x_i)$ ),以便$S_{ij}=(u_i,v_j-\theta)$ 。然后，最小非对称二进制码长度由下式给出：
(2)
$$\begin{aligned}k_{a}(S)&=\quad\min_{k,U,V,\theta}k\quad\mathrm{s.t}\quad U,V\in\{\pm1\}^{k\times n}\quad\theta\in\\&YU^{\top}V-\theta\mathbb{1}_{n}\\&\forall_{ij}\:S_{ij}Y_{ij}>0\end{aligned}$$
将二进制编码问题写为矩阵分解问题对于理解我们可以通过不对称获得的能力很有用：即使$S$是对称的， 即使我们寻求对称的$Y$ , 坚持将 $Y$ 写为二进制矩阵的平方可能是一个艰难的约束。这在以下定理中得到了体现，该定理确定即使矩阵$S$是对称的并且表现良好，最小不对称二进制代码长度和最小对称代码长度之间也可能存在指数差距：
For any $r$, there exists a set of $n=2^r$ points in Euclidean space, with similarity matrix $S_{ij}=$ $\begin{cases}1&\text{if }x_{i}-x_{j}\leq1\\-1&\text{if }r_{i}-r_{i}>1&\text{,such that }k_{a}(S)\leq2r\text{ but }k_{s}(S)\geq2^{r}/2&\text{Let }I_{1}=\{1,\ldots,n/2\}\text{ and }I_{2}=\{n/2+\end{cases}$
3
|
[
[
0.1405029296875,
0.09344482421875,
0.88134765625,
0.09344482421875,
0.88134765625,
0.106689453125,
0.1405029296875,
0.106689453125
],
[
0.1192626953125,
0.10919189453125,
0.88134765625,
0.10919189453125,
0.88134765625,
0.12249755859375,
0.1192626953125,
0.12249755859375
],
[
0.11767578125,
0.12371826171875,
0.88134765625,
0.12371826171875,
0.88134765625,
0.1383056640625,
0.11767578125,
0.1383056640625
],
[
0.11767578125,
0.1383056640625,
0.8798828125,
0.1383056640625,
0.8798828125,
0.1514892578125,
0.11767578125,
0.1514892578125
],
[
0.11767578125,
0.154052734375,
0.8798828125,
0.154052734375,
0.8798828125,
0.1673583984375,
0.11767578125,
0.1673583984375
],
[
0.131591796875,
0.1685791015625,
0.88134765625,
0.1685791015625,
0.88134765625,
0.1817626953125,
0.131591796875,
0.1817626953125
],
[
0.11767578125,
0.184326171875,
0.50732421875,
0.184326171875,
0.50732421875,
0.1976318359375,
0.11767578125,
0.1976318359375
],
[
0.1446533203125,
0.2008056640625,
0.8798828125,
0.2008056640625,
0.8798828125,
0.2108154296875,
0.1446533203125,
0.2108154296875
],
[
0.1168212890625,
0.212158203125,
0.88134765625,
0.21337890625,
0.88134765625,
0.2279052734375,
0.1168212890625,
0.2266845703125
],
[
0.11767578125,
0.2291259765625,
0.88134765625,
0.2291259765625,
0.88134765625,
0.242431640625,
0.11767578125,
0.242431640625
],
[
0.11767578125,
0.24560546875,
0.8798828125,
0.24560546875,
0.8798828125,
0.255615234375,
0.11767578125,
0.255615234375
],
[
0.11767578125,
0.260009765625,
0.88134765625,
0.260009765625,
0.88134765625,
0.2734375,
0.11767578125,
0.2734375
],
[
0.11767578125,
0.274658203125,
0.88134765625,
0.274658203125,
0.88134765625,
0.28857421875,
0.11767578125,
0.28857421875
],
[
0.11767578125,
0.291748046875,
0.88134765625,
0.291748046875,
0.88134765625,
0.3017578125,
0.11767578125,
0.3017578125
],
[
0.11517333984375,
0.3037109375,
0.56396484375,
0.304931640625,
0.56396484375,
0.3193359375,
0.11517333984375,
0.318115234375
],
[
0.11358642578125,
0.341552734375,
0.70361328125,
0.343505859375,
0.70361328125,
0.35986328125,
0.11358642578125,
0.35791015625
],
[
0.11767578125,
0.373779296875,
0.88134765625,
0.373779296875,
0.88134765625,
0.386962890625,
0.11767578125,
0.386962890625
],
[
0.1168212890625,
0.386474609375,
0.88134765625,
0.386962890625,
0.88134765625,
0.401611328125,
0.1168212890625,
0.40087890625
],
[
0.1168212890625,
0.401611328125,
0.88232421875,
0.40283203125,
0.88232421875,
0.417236328125,
0.1168212890625,
0.416748046875
],
[
0.11767578125,
0.41845703125,
0.88134765625,
0.41845703125,
0.88134765625,
0.431884765625,
0.11767578125,
0.431884765625
],
[
0.11767578125,
0.434326171875,
0.6806640625,
0.434326171875,
0.6806640625,
0.447509765625,
0.11767578125,
0.447509765625
],
[
0.143798828125,
0.449462890625,
0.8798828125,
0.449462890625,
0.8798828125,
0.460205078125,
0.143798828125,
0.460205078125
],
[
0.1168212890625,
0.462158203125,
0.8798828125,
0.46337890625,
0.8798828125,
0.477783203125,
0.1168212890625,
0.4765625
],
[
0.11517333984375,
0.4765625,
0.88134765625,
0.477783203125,
0.88134765625,
0.492431640625,
0.11517333984375,
0.4912109375
],
[
0.1168212890625,
0.49365234375,
0.88134765625,
0.49365234375,
0.88134765625,
0.5068359375,
0.1168212890625,
0.5068359375
],
[
0.11517333984375,
0.5068359375,
0.223876953125,
0.50830078125,
0.223876953125,
0.52294921875,
0.11517333984375,
0.521484375
],
[
0.1422119140625,
0.52587890625,
0.7255859375,
0.52587890625,
0.7255859375,
0.5361328125,
0.1422119140625,
0.5361328125
],
[
0.1422119140625,
0.5390625,
0.88134765625,
0.5390625,
0.88134765625,
0.5537109375,
0.1422119140625,
0.5537109375
],
[
0.11358642578125,
0.55322265625,
0.22216796875,
0.55322265625,
0.22216796875,
0.56640625,
0.11358642578125,
0.56640625
],
[
0.336669921875,
0.5673828125,
0.66259765625,
0.56298828125,
0.66259765625,
0.58837890625,
0.336669921875,
0.59228515625
],
[
0.859375,
0.5693359375,
0.88232421875,
0.5693359375,
0.88232421875,
0.583984375,
0.859375,
0.583984375
],
[
0.5,
0.5869140625,
0.60302734375,
0.59033203125,
0.60302734375,
0.60498046875,
0.5,
0.6015625
],
[
0.50341796875,
0.609375,
0.6044921875,
0.609375,
0.6044921875,
0.62353515625,
0.50341796875,
0.62353515625
],
[
0.1168212890625,
0.62890625,
0.279296875,
0.62890625,
0.279296875,
0.64306640625,
0.1168212890625,
0.64306640625
],
[
0.1422119140625,
0.64453125,
0.88134765625,
0.64453125,
0.88134765625,
0.65771484375,
0.1422119140625,
0.65771484375
],
[
0.1168212890625,
0.65771484375,
0.88134765625,
0.6591796875,
0.88134765625,
0.673828125,
0.1168212890625,
0.67236328125
],
[
0.1168212890625,
0.6748046875,
0.88134765625,
0.6748046875,
0.88134765625,
0.689453125,
0.1168212890625,
0.689453125
],
[
0.11767578125,
0.68994140625,
0.1405029296875,
0.68994140625,
0.1405029296875,
0.7021484375,
0.11767578125,
0.7021484375
],
[
0.318603515625,
0.70068359375,
0.6806640625,
0.69775390625,
0.6806640625,
0.7265625,
0.318603515625,
0.72998046875
],
[
0.861328125,
0.705078125,
0.88232421875,
0.705078125,
0.88232421875,
0.7197265625,
0.861328125,
0.7197265625
],
[
0.5,
0.7216796875,
0.60107421875,
0.724609375,
0.60107421875,
0.74072265625,
0.5,
0.7373046875
],
[
0.5009765625,
0.744140625,
0.60107421875,
0.744140625,
0.60107421875,
0.7587890625,
0.5009765625,
0.7587890625
],
[
0.1396484375,
0.76318359375,
0.88134765625,
0.7646484375,
0.88134765625,
0.77880859375,
0.1396484375,
0.77783203125
],
[
0.11767578125,
0.77783203125,
0.8798828125,
0.77880859375,
0.8798828125,
0.79345703125,
0.11767578125,
0.79248046875
],
[
0.11767578125,
0.79443359375,
0.88134765625,
0.79443359375,
0.88134765625,
0.80810546875,
0.11767578125,
0.80810546875
],
[
0.1192626953125,
0.81005859375,
0.28759765625,
0.81005859375,
0.28759765625,
0.82373046875,
0.1192626953125,
0.82373046875
],
[
0.1405029296875,
0.82568359375,
0.88134765625,
0.82568359375,
0.88134765625,
0.8388671875,
0.1405029296875,
0.8388671875
],
[
0.12091064453125,
0.84130859375,
0.281005859375,
0.84326171875,
0.281005859375,
0.859375,
0.12091064453125,
0.85693359375
],
[
0.27294921875,
0.8515625,
0.88134765625,
0.8515625,
0.88134765625,
0.8662109375,
0.27294921875,
0.8662109375
],
[
0.1192626953125,
0.859375,
0.27685546875,
0.85791015625,
0.27685546875,
0.87548828125,
0.1192626953125,
0.876953125
],
[
0.491943359375,
0.8984375,
0.5087890625,
0.8984375,
0.5087890625,
0.91357421875,
0.491943359375,
0.91357421875
]
] |
[
"在本文中，我们建议使用两个distinct映射f(α),g(a）E（士1)k，并通过f(c）和g(r'）之间的汉明距离",
"来近似相似度S(&，'）。我们将这种哈希方案称为“非对称”。我们的主要结果是，即使目标相似度函数是",
"对称的并且“表现良好”（e。g。，即使它基于对象之间的欧几里得距离），使用非对称二进制哈希可以强大得",
"多，并允许以较短的代码长度更好地近似目标相似性。特别是，我们显示了欧几里得空间中的点集合的极端",
"示例，其中可以使用长度为O(r）位的非对称二进制哈希（基于一对不同的函数）来实现邻域相似性S(c，a）",
"但是对称哈希（基于单个函数）将需要至少2(2r）位。尽管实际数据并不极端，但我们在真实数据集上的",
"实验结果表明，使用非对称二进制哈希值具有显着优势",
"非对称哈希可以在儿乎所有通常使用对称哈希的地方使用，通常无需任何额外的存储或计算成本。考虑",
"将所有对象的哈希向量存储在数据库中，然后通过计算查询的哈希及其与存储的数据库哈希的汉明距离来计",
"算与查询的相似性的典型应用。使用非对称哈希意味着对数据库和查询使用不同的哈希函数。这既不会增加",
"数据库表示的大小，也不会增加填充数据库或执行香询的计算或通信成本，因为需要完全相同的操作。事实",
"上，当哈希整个数据库时，非对称哈希提供了更多的改进机会。我们认为，使用两个不同的哈希函数对数据",
"库对象和查询进行编码可以在选择数据库哈希时提供更大的灵活性。与查询哈希不同，查询哈希必须紧地",
"存储并在查询出现时对其进行有效评估，如果数据库是固定的，则可以使用数据库对象到位串的任意映射。",
"我们证明，这确实可以提高相似性精度，同时减少所需的位长。",
"2Minimum Code Lengths and the Power of Asymmetry",
"令S：×→（土1)是一组对象上的二进制相似性函数，其中我们可以将S(c,）解释为表示和c",
"是“相似”还是“不相似”，或者指示它们是否是“邻居”。的对称二进制编码是映射f：→{士1}，其",
"中k是代码的位长度。我们有兴趣构造代码，使得f(a）和f(r）之间的汉明距离对应于相似度S(，a'）。",
"也就是说，对于某个阈值E，S(α,a)～（f(α),f(α）-の）。虽然讨论了汉明距离，但我们使用内积u,更",
"方便，这相当于自u,=k-2dh(u,u)）以来u,E（士1)*的汉明距离dh(,v",
"在本节中，我们将考虑使用任意二进制代码捕获给定相似性的问题。也就是说，我们给出了整个相似性",
"映射S，e。g。E作为n对象的有限域=(ai...,an)上的矩阵SE(±1)nxn，带有Si=S(ci,ai）",
"我们要求每个对象iα的编码ui=f(αi)E【1)k和阈值，使得Sii=（u，ui一の），或者至少使得尽",
"可能多的对（i，）相等。需要强调的是，这里的目标纯粹是使用短二进制码来近似给定的矩阵S一不存在样",
"本外泛化 (还)",
"我们现在问：充许不对称编码是否能够以更短的代码长度近似对称相似性矩阵S？",
"为其列包含码字ui的矩阵表示UE{土1)nxk，然后通过以下矩阵分解问题给出允许精确表示S的最小",
"二进制码长度：",
"ks(S)= mink s.t Ue(±1)axne",
"(1)",
"YUTU-01n",
"Vij SijYij > 0",
"其中1n是n×n矩阵",
"我们开始通过考虑上述问题的asymmetric变体来证明不对称的力量。也就是说，即使S是对称的，我",
"们也允许将每个对象c关联两个不同的二进制码字iE【土1)和E1}（我们可以将其视为具有两",
"个任意映射ui=f(ci）和vi=g(ai）），以便S=（ui,us一の）。然后，最小非对称二进制码长度由下式给",
"出：",
"ka(s)=wiv。k s.t U.Ve(+1)xn0e",
"(2)",
"YUTV-01n",
"Vii SijYi > 0",
"将二进制编码问题写为矩阵分解问题对于理解我们可以通过不对称获得的能力很有用：即使S是对称的，",
"即使我们寻求对称的Y，坚持将Y写为二进制矩阵的平方可能是一个艰难的约束。这在以下定理中得到了",
"体现，该定理确定即使矩阵S是对称的并且表现良好，最小不对称二进制代码长度和最小对称代码长度之",
"间也可能存在指数差距：",
"For any r ,there exists a set of n = 2\" points in Euclidean space, with similarity matrix Sij =",
"[1 if ai-a,≤l",
", such that ka(S)≤ 2r but ks(S) ≥2/2Let I =[1,...,n/2] and I2=(n/2+",
"{-1 if ai-aj>\"",
"3"
] |
{
"arxiv_id": "1311.7662",
"language": "zh",
"page_idx": 2
}
|
||
1311.7662-zh-3
|
1311.7662-zh-3.jpg
|
$1, \ldots , n\}$ . Consider the matrix $G$ defined by $G_ii=1/2$ ,$G_{ij}=-1/(2n)$ if $i,j\in I_1$ or $i,j\in I_2$ ,and $G_{ij}=1/(2n)$ otherwise. Matrix $G$ is diagonally dominant. By the Gershgorin circle theorem, $G$ is positive definite. Therefore, there exist vectors $x_1,\ldots,x_n$ such that $x_i,x_j=G_{ij}$ (for every $i$ and $j).$ Define
$$S_{ij}=\begin{cases}1&\text{if}x_i-x_j\le1\\-1&\text{if}x_i-x_j>1\end{cases}.$$
$\begin{array}{rcl}{\mathrm{Note~that~if~}i=j\mathrm{~then~}S_{ij}=1;\mathrm{~if~}i\neq j\mathrm{~and~}(i,j)\in I_{1}\times I_{1}\cup I_{2}\times I_{2}\mathrm{~then~}x_{i}-x_{j}^{2}=G_{ii}+G_{jj}-}\\{2G_{ij}=1+1/n>1\mathrm{~and~therefore~}S_{ij}=-1.\mathrm{~Finally,~if~}i\neq j\mathrm{~and~}(i,j)\in I_{1}\times I_{2}\cup I_{2}\times I_{1}\mathrm{~then}}\end{array}$ $x_{i}-x_{j}^{2}=G_{ii}+G_{jj}-2G_{ij}=1+1/n<1$ and therefore $S_ij=1.$ We show that $k_a(S)\leq2r.$ Let $B$ be an $r\times n$ matrix whose column vectors are the vertices of the cube $\{\pm1\}^r\left(\text{in any order}\right);$let $C$ be an $r\times n$ matrix lefined by $C_{ij}=1$ if $j\in I_{1}$ and $C_{ij}=-1$ if $j\in I_{2}$. Let $U=\begin{bmatrix}B\\C\end{bmatrix}$ and $V=\begin{bmatrix}B\\-C\end{bmatrix}\cdot$For $Y=U^\top V-\theta\mathbf{1}$ where threshold $\theta=-1$ , we have that $Y_{ij}\geq1$ if $S_{ij}=1$ and $Y_{ij}\leq-1$ if $S_{ij}=-1$ .Therefore, $k_a(S)\leq2r$ . Now we show that $k_s=k_s(S)\geq n/2$ . Consider $Y,U$ and $\theta$ as in (1). Let $Y^\prime=(U^\top U).$ Note that $Y_{ij}^{\prime}\in[-k_{s},k_{s}]$ and thus $\theta\in[-k_{s}+1,k_{s}-1]$ . Let $q=[1,\ldots,1,-1,\ldots,-1]^{\top}(n/2$ ones followed by $n/2$ ${\mathrm{minus~ones}}).{\mathrm{~We~have,}}$
$$\begin{aligned}0\leq q^{\top}Y^{\prime}q&=\sum_{i=1}^nY_{ii}^{\prime}+\sum_{i,j:S_{ij}=-1}Y_{ij}^{\prime}-\sum_{i,j:S_{ij}=1,i\neq j}Y_{ij}^{\prime}\\&\leq\sum_{i=1}^nk_s+\sum_{i,j:S_{ij}=-1}(\theta-1)-\sum_{i,j:S_{ij}=1,i\neq j}(\theta+1)\\&=nk_s+(0.5n^2-n)(\theta-1)-0.5n^2(\theta+1)\\&=nk_s-n^2-n(\theta-1)\\&\leq2nk_s-n^2.\end{aligned}$$
We conclude that $k_s\geq n/2$
定理 2 的构造表明，对于某些 exists 数据集，非对称二进制哈希可能比对称哈希短得多。这是一个重要的观察，因为它表明不对称哈希可能更强大，并且应该促使我们考虑它们而不是对称哈希。定理 2 的精确构造当然是相当极端的 (事实上，可能是最极端的构造),我们不会期望实际数据集具有这种确切的结构，但我们将在以后的实际数据集上显示出明显的差距。
3 Approximate Binary Codes
当我们转向真实数据集时，我们还需要从寻求 exactly 捕获相似性矩阵的二进制编码出发。相反，我们通常只满足于近似$S$ ,并且对于任何固定的代码长度$k$ ,都寻求 “最佳捕获”相似性矩阵$S$的 (对称或非对称) $k$位代码。这是由以下优化问题捕获的：
(3)
$$\min\limits_{U,V,\theta}\:L(Y;S)\beta\sum\limits_{i,j:S_{ij}=1}\ell(Y_{ij})+(1-\beta)\sum\limits_{i,j:S_{ij}=-1}\ell(-Y_{ij})\:\mathrm{s.t.}\:U,V\in\{\pm1\}^{k\times n}\:\theta\in\\YU^\top V-\theta\mathbf{1}_n$$
其中$\ell(z)=\mathbf{1}_{z\leq0}$是零一错误，$\beta$是一个参数，它允许我们以不同的方式对正负错误进行加权。这样的加权可以补偿$\bar{S_{ij}\text{ 的不平衡 (通常更多的点对是非相似的，而不是相似的),并允许我们在精度和召回率之间获}}$ 得不同的平衡
优化问题 (3) 是一个离散，不连续且高度非凸的问题。在我们的实验中，我们将零一损失$\ell(\cdot)$替换为连续损失，并通过贪婪地更新单个比特来执行局部搜索，以改善这一目标。尽管即使$\ell(z)$是凸的，产生的目标 (更不用说离散优化问题了)仍然不是凸的，但我们发现在$z<0$上使用不平坦的损失函数是有益的，从而鼓励朝着正确的符号移动。在我们的实验中，我们使用了逻辑损失的平方根，$\ell(z)=\log^{1/2}(1+e^{-z})$
4
|
[
[
0.11767578125,
0.0921630859375,
0.88134765625,
0.0921630859375,
0.88134765625,
0.106689453125,
0.11767578125,
0.106689453125
],
[
0.1168212890625,
0.10797119140625,
0.88134765625,
0.10797119140625,
0.88134765625,
0.121826171875,
0.1168212890625,
0.121826171875
],
[
0.11358642578125,
0.12060546875,
0.82666015625,
0.121826171875,
0.82666015625,
0.1395263671875,
0.11358642578125,
0.1383056640625
],
[
0.443603515625,
0.15087890625,
0.59326171875,
0.154052734375,
0.59326171875,
0.1717529296875,
0.443603515625,
0.1685791015625
],
[
0.400390625,
0.1622314453125,
0.441162109375,
0.1622314453125,
0.441162109375,
0.177978515625,
0.400390625,
0.177978515625
],
[
0.4453125,
0.17041015625,
0.6005859375,
0.17041015625,
0.6005859375,
0.1881103515625,
0.4453125,
0.1881103515625
],
[
0.11517333984375,
0.2008056640625,
0.87158203125,
0.2020263671875,
0.87158203125,
0.216552734375,
0.11517333984375,
0.21533203125
],
[
0.1168212890625,
0.2171630859375,
0.8798828125,
0.2171630859375,
0.8798828125,
0.231689453125,
0.1168212890625,
0.231689453125
],
[
0.1168212890625,
0.231689453125,
0.8798828125,
0.231689453125,
0.8798828125,
0.246826171875,
0.1168212890625,
0.246826171875
],
[
0.1168212890625,
0.2474365234375,
0.88134765625,
0.2474365234375,
0.88134765625,
0.261962890625,
0.1168212890625,
0.261962890625
],
[
0.5703125,
0.26513671875,
0.5888671875,
0.26318359375,
0.5908203125,
0.272705078125,
0.57177734375,
0.274658203125
],
[
0.67236328125,
0.264404296875,
0.70361328125,
0.261962890625,
0.705078125,
0.274658203125,
0.67333984375,
0.276611328125
],
[
0.5986328125,
0.26953125,
0.66845703125,
0.26953125,
0.66845703125,
0.2841796875,
0.5986328125,
0.2841796875
],
[
0.11517333984375,
0.26953125,
0.56201171875,
0.268310546875,
0.56201171875,
0.28271484375,
0.11517333984375,
0.2841796875
],
[
0.7197265625,
0.26953125,
0.88134765625,
0.26953125,
0.88134765625,
0.2841796875,
0.7197265625,
0.2841796875
],
[
0.57177734375,
0.279052734375,
0.59228515625,
0.279052734375,
0.59228515625,
0.291748046875,
0.57177734375,
0.291748046875
],
[
0.6806640625,
0.279052734375,
0.708984375,
0.279052734375,
0.708984375,
0.290283203125,
0.6806640625,
0.290283203125
],
[
0.11358642578125,
0.290283203125,
0.88232421875,
0.291748046875,
0.88232421875,
0.309326171875,
0.11358642578125,
0.30810546875
],
[
0.125,
0.30810546875,
0.88134765625,
0.30810546875,
0.88134765625,
0.322509765625,
0.125,
0.322509765625
],
[
0.1168212890625,
0.323974609375,
0.88232421875,
0.323974609375,
0.88232421875,
0.337158203125,
0.1168212890625,
0.337158203125
],
[
0.11517333984375,
0.337158203125,
0.279296875,
0.33837890625,
0.279296875,
0.352783203125,
0.11517333984375,
0.3515625
],
[
0.46728515625,
0.36865234375,
0.5888671875,
0.371826171875,
0.58837890625,
0.396484375,
0.465576171875,
0.393310546875
],
[
0.5810546875,
0.369384765625,
0.6640625,
0.371826171875,
0.66259765625,
0.39453125,
0.580078125,
0.392578125
],
[
0.3857421875,
0.389404296875,
0.414306640625,
0.389404296875,
0.414306640625,
0.406005859375,
0.3857421875,
0.406005859375
],
[
0.28759765625,
0.371826171875,
0.473876953125,
0.371826171875,
0.473876953125,
0.393310546875,
0.28759765625,
0.393310546875
],
[
0.443603515625,
0.393310546875,
0.5185546875,
0.393310546875,
0.5185546875,
0.40478515625,
0.443603515625,
0.40478515625
],
[
0.5556640625,
0.393310546875,
0.6396484375,
0.393310546875,
0.6396484375,
0.40478515625,
0.5556640625,
0.40478515625
],
[
0.396240234375,
0.4091796875,
0.40283203125,
0.4091796875,
0.40283203125,
0.412353515625,
0.396240234375,
0.412353515625
],
[
0.426513671875,
0.396484375,
0.4453125,
0.395751953125,
0.4453125,
0.399658203125,
0.426513671875,
0.40087890625
],
[
0.461669921875,
0.41845703125,
0.5849609375,
0.41845703125,
0.5849609375,
0.436279296875,
0.461669921875,
0.436279296875
],
[
0.60546875,
0.41845703125,
0.64208984375,
0.41845703125,
0.64208984375,
0.4375,
0.60546875,
0.4375
],
[
0.65673828125,
0.41845703125,
0.70751953125,
0.42041015625,
0.7060546875,
0.43505859375,
0.65625,
0.43310546875
],
[
0.65185546875,
0.423583984375,
0.65673828125,
0.423583984375,
0.65673828125,
0.4287109375,
0.65185546875,
0.4287109375
],
[
0.36767578125,
0.419189453125,
0.471435546875,
0.417236328125,
0.471435546875,
0.434326171875,
0.36767578125,
0.436279296875
],
[
0.64208984375,
0.4248046875,
0.65185546875,
0.4248046875,
0.65185546875,
0.4287109375,
0.64208984375,
0.4287109375
],
[
0.3857421875,
0.436279296875,
0.5087890625,
0.436279296875,
0.5087890625,
0.45068359375,
0.3857421875,
0.45068359375
],
[
0.580078125,
0.436279296875,
0.65869140625,
0.436279296875,
0.65869140625,
0.45068359375,
0.580078125,
0.45068359375
],
[
0.369384765625,
0.455078125,
0.66845703125,
0.453857421875,
0.66845703125,
0.468505859375,
0.369384765625,
0.4697265625
],
[
0.369384765625,
0.474853515625,
0.52685546875,
0.474853515625,
0.52685546875,
0.4892578125,
0.369384765625,
0.4892578125
],
[
0.369384765625,
0.49560546875,
0.463134765625,
0.49560546875,
0.463134765625,
0.51025390625,
0.369384765625,
0.51025390625
],
[
0.11767578125,
0.52587890625,
0.314453125,
0.52587890625,
0.314453125,
0.5361328125,
0.11767578125,
0.5361328125
],
[
0.1422119140625,
0.5390625,
0.8798828125,
0.5390625,
0.8798828125,
0.55322265625,
0.1422119140625,
0.55322265625
],
[
0.11767578125,
0.5537109375,
0.8798828125,
0.5537109375,
0.8798828125,
0.5673828125,
0.11767578125,
0.5673828125
],
[
0.1192626953125,
0.57080078125,
0.8798828125,
0.57080078125,
0.8798828125,
0.58056640625,
0.1192626953125,
0.58056640625
],
[
0.1168212890625,
0.58544921875,
0.465576171875,
0.58544921875,
0.465576171875,
0.5986328125,
0.1168212890625,
0.5986328125
],
[
0.11358642578125,
0.6220703125,
0.435546875,
0.62255859375,
0.435546875,
0.64013671875,
0.11358642578125,
0.6396484375
],
[
0.1192626953125,
0.65380859375,
0.88134765625,
0.65380859375,
0.88134765625,
0.66748046875,
0.1192626953125,
0.66748046875
],
[
0.1192626953125,
0.67041015625,
0.88232421875,
0.67041015625,
0.88232421875,
0.68115234375,
0.1192626953125,
0.68115234375
],
[
0.11517333984375,
0.6845703125,
0.398681640625,
0.6845703125,
0.398681640625,
0.69775390625,
0.11517333984375,
0.69775390625
],
[
0.2320556640625,
0.71044921875,
0.76708984375,
0.7021484375,
0.76708984375,
0.72802734375,
0.2320556640625,
0.736328125
],
[
0.859375,
0.71337890625,
0.88232421875,
0.71337890625,
0.88232421875,
0.72802734375,
0.859375,
0.72802734375
],
[
0.330078125,
0.72900390625,
0.3857421875,
0.72900390625,
0.3857421875,
0.74365234375,
0.330078125,
0.74365234375
],
[
0.4599609375,
0.72900390625,
0.52783203125,
0.72900390625,
0.52783203125,
0.74365234375,
0.4599609375,
0.74365234375
],
[
0.6044921875,
0.73095703125,
0.7060546875,
0.73291015625,
0.7060546875,
0.74755859375,
0.6044921875,
0.74560546875
],
[
0.11767578125,
0.76220703125,
0.88134765625,
0.76220703125,
0.88134765625,
0.77587890625,
0.11767578125,
0.77587890625
],
[
0.11767578125,
0.77783203125,
0.88232421875,
0.77783203125,
0.88232421875,
0.79150390625,
0.11767578125,
0.79150390625
],
[
0.11767578125,
0.79345703125,
0.214111328125,
0.79345703125,
0.214111328125,
0.8046875,
0.11767578125,
0.8046875
],
[
0.1422119140625,
0.8095703125,
0.8798828125,
0.8095703125,
0.8798828125,
0.8193359375,
0.1422119140625,
0.8193359375
],
[
0.1168212890625,
0.82275390625,
0.8798828125,
0.82275390625,
0.8798828125,
0.8359375,
0.1168212890625,
0.8359375
],
[
0.11517333984375,
0.8359375,
0.88134765625,
0.8369140625,
0.88134765625,
0.8515625,
0.11517333984375,
0.85107421875
],
[
0.11767578125,
0.85302734375,
0.84228515625,
0.85302734375,
0.84228515625,
0.8662109375,
0.11767578125,
0.8662109375
],
[
0.48779296875,
0.9013671875,
0.50341796875,
0.89453125,
0.51318359375,
0.90771484375,
0.495849609375,
0.9140625
]
] |
[
"1,...,n) . Consider the matrix G defined by Gi = 1/2, Gij = -1/(2n) if i,j e Ii or i,j E I2 , and",
"Gij =1/(2n) otherwise. Matrix G is diagonally dominant. By the Gershgorin circle theorem, G is positive",
"definite. Therefore, there exist vectors i,., an such that ri, aj = Gii (for every i and j ). Define",
"Jiif ri-aj≤l",
"Si=",
"[-1 if i-j>1",
"Notethat if i=jthen Sij=1;if ijand (ii)eIi×IiUI2×I2thenai-a,2=Gi+Gjj",
"2Gi=1+1/n> 1 and therefore Sii =-1Finally, if i j and (i,)I × I2U I2 × I then",
"Ci -c,2=Gi +Gjj-2Gij =1+1/n <1 and therefore Sij =1. We show that ka(S)≤2r:Let B be an",
"r x n matrix whose column vectors are the vertices of the cube (+1jr (in any order); let C be an r × n matrix",
"B",
"B",
"and V =",
"defined by Cii=l if jeIiand Cii=-lif jEI2.LetU =",
".ForY=UTV-01n",
"C",
"-C",
"where threshold = -1, we have that Yij ≥1 if Sij = 1 and Yi, ≤-1 if Sij =-1, Therefore, ka(S) ≤2r",
"Now we show that ks = ks(S) ≥ n/2 : Consider Y, U and as in (1). Let Y' = (U U) : Note that",
"Y, E [-ks, ks] and thus E [-ks + 1, ks - 1] . Let q = [1,..., 1, -1,...,-1] (n/2 ones followed by n/2",
"minus ones). We have,",
"Z Yij-",
"ZYs",
"i=1",
"≤q'q-Y+",
"j:Si5=--1",
"ij:Su=l,j",
"",
"",
"(0-1)-",
"Z",
"(0 + 1)",
"",
"≤ks+",
"",
"i,j:Sj---",
"i,j:Su=1,详",
"= nks+(0.5n²-n)(0-1)-0.5n²(0+1)",
"=nks-n²-n(0-1)",
"≤2nks - n².",
"Weconcludethatks>n/2",
"定理2的构造表明，对于某些exists数据集，非对称二进制哈希可能比对称哈希短得多。这是一个重要",
"的观察，因为它表明不对称哈希可能更强大，并且应该促使我们考虑它们而不是对称哈希。定理2的精确构",
"造当然是相当极端的（事实上，可能是最极端的构造），我们不会期望实际数据集具有这种确切的结构：但",
"我们将在以后的实际数据集上显示出明显的差距。",
"3Approximate Binary Codes",
"当我们转向真实数据集时，我们还需要从寻求exactly捕获相似性矩阵的二进制编码出发。相反，我们通常",
"只满足于近似S，并且对于任何固定的代码长度k，都寻求“最佳捕获”相似性矩阵S的（对称或非对称",
"k位代码。这是由以下优化问题捕获的：",
"i L(Y;S)βEe(Ya) +(1-β)e(-Yi) s.t. U, Ve (±1)xn 0 e",
"(3)",
"ij:Sj=1",
"ijiSuj=-1",
"YUTV-01n",
"其中(2）=1z<0是零一错误，β是一个参数，它允许我们以不同的方式对正负错误进行加权。这样的加权",
"可以补偿Si的不平衡（通常更多的点对是非相似的，而不是相似的），并允许我们在精度和召回率之间获",
"得不同的平衡",
"优化问题（3）是一个离散，不连续且高度非凸的问题。在我们的实验中，我们将零一损失（·）替换为连",
"续损失，并通过贪婪地更新单个比特来执行局部搜索，以改善这一目标。尽管即使（）是凸的，产生的目",
"标（更不用说离散优化问题了）仍然不是凸的，但我们发现在之<0上使用不平坦的损失函数是有益的，从",
"而鼓励朝着正确的符号移动。在我们的实验中，我们使用了逻辑损失的平方根，l(z）=log/2(1+e-2)",
"4"
] |
{
"arxiv_id": "1311.7662",
"language": "zh",
"page_idx": 3
}
|
||
1311.7662-zh-4
|
1311.7662-zh-4.jpg
|
在进行样本外概括之前，我们简要报告了根据经验找到具有对称和非对称代码的实际相似性矩阵的良好近似值所需的位数。我们尝试了几个数据集，试图用对称和非对称代码来拟合它们，然后通过改变阈值$\theta$来计算平均精度 (同时保持$U$和$V$固定)。两个相似度矩阵的结果，一个基于均匀分布在超立方体中的点之间的欧几里得距离，另一个基于图像之间的语义相似度，如图 ?? 所示
4 Out of Sample Generalization: Learning a Mapping
到目前为止，我们专注于通过将任意代码字与每个对象相关联并完全忽略对象$x_i$的输入表示来学习一组固定对象上的二进制代码。我们只讨论了如何以及二进制哈希可以 approximate 的相似性，但没有考虑generalizing 额外的新对象。然而，在大多数应用程序中，我们希望能够有这样一个样本外的泛化。也就是说，我们想仅使用有限的对象训练集来学习无限域$\chi$上的映射$f:\mathcal{X}\to\{\pm1\}^k$ ,然后应用该映射以获得要遇到的未来对象的二进制码$f(x)$ ,例如$S(x,x^\prime)\approx(f(x),f(x^{\prime})-\theta)$ 。因此，映射$f:\mathcal{X}\to\{\pm1\}^k$通常仅限于某些受约束的参数类，因此我们可以在新对象上有效地表示和评估它，并确保良好的泛化。例如，当$\mathcal{X}=^d$时，我们可以考虑线性阈值映射$f_W(x)=(Wx)$,其中$W\in^k\times d$和 (·)按元素操作，如在最小损失哈希[8]中。或者，我们也可以考虑更复杂的类，例如多层网络[11,9]
我们已经看到，非对称二进制码可以使用较短的代码进行更好的近似，因此在这里寻求非对称代码也很自然。也就是说，不是学习单个参数映射$f(x)$ ,而是可以学习一对映射$f:\mathcal{X}\to\{\pm1\}^k$和$g:\mathcal{X}\to\{\pm1\}^k$ ,它们都约束为某个参数类，以及阈值$\theta$,例如$S(x,x^\prime)\approx(f(x),g(x^{\prime})-\theta)$ 。这有可能允许更好地近似相似性，从而用更短的代码获得更好的总体精度(尽管由于参数数量的增加可能会稍微更难概括)
事实上，在一个典型的应用程序中，对象数据库被哈希化以在未来的查询中进行相似性搜索，不对称性使我们能够走得更远。考虑以下设置：我们从一些无限域$\chi$中获得$n$对象$x_1,\ldots,x_n\in\mathcal{X}$ ,以及这些对象之间的相似性$S(x_i,x_j)$ 。我们的目标是使用短二进制代码来哈希这些对象，这将使我们能够快速计算这些对象 (“数据库”) 和末来对象$x$ (“查询”) 之间的近似相似性。也就是说，我们希望为数据库中的对象生成并存储紧凑的二进制代码。然后，给定一个新的查询对象，我们希望有效地计算给定查询的紧凑二进制代码， 并通过在数据库中查找与查询二进制代码相距较近的二进制代码，非常快速地检索数据库中的类似项目。回想一下，重要的是要确保哈希的位长度很小，因为短代码允许非常快速的汉明距离计算和低通信成本，如果代码需要远程发送。更重要的是，如果我们想将数据库存储在允许立即查找的哈希表中，哈希表的大小在代码长度上是指数级的
对称二进制哈希方法 (e。g。$\operatorname{E}[8]$ ),将是找到单个参数映射$f:\mathcal{X}\to\{\pm1\}^k$ ,以便$S(x,x_i)\approx$ $(f(x),f(x_i)-\theta)$用于将来的查询$x$和数据库对象$x_i$,计算所有数据库对象$x_i$的$f(x_i)$ ,并存储这些哈希 (可能在哈希表中允许在短的汉明距离内快速检索代码)。上面描述的非对称方法是找到两个参数映射$f:\mathcal{X}\to\{\pm1\}^k$和$g:\mathcal{X}\to\{\pm1\}^k$,使得$S(x,x_i)\approx(f(x),g(x_i)-\theta)$ ,然后计算并存储$g(x_i)$
但是，如果数据库是固定的，我们可以走得更远。实际上，$g(\cdot)$不需要位于受约束的参数类中，因为我们不需要将$g(\cdot)$推广到未来的对象，也不需要高效地计算它，也不需要将$g(x)$传递到数据库。因此，我们可以考虑允许数据库哈希函数$g(\cdot)$是任意映射。也就是说，我们的目标是为数据库中的每个$x_1,\ldots,x_n$找到一个简单的参数映射$f:\mathcal{X}\to\{\pm1\}^k$和$n$ arbitrary codewords $v_1,\ldots,v_n\in\{\pm1\}^k$,以便$S(x,x_i)\approx(f(x),v_i-\theta)$ 用于将来的查询$x$和数据库中的对象$x_i,\ldots,x_n$ 。这种形式的不对称可以使我们获得更大的近似能力，从而在不增加额外的计算或存储成本的情况下，通过更短的代码获得更好的精度
在 6 节中，我们对上述两种非对称策略进行了经验评估，并证明了它们的好处。但是在此之前，在下一
节中，我们将讨论一种局部搜索方法，用于查找映射$f,g$或映射$f$和代码$v_1,\ldots,v_n$ 。
# 5 Optimization
我们专注于$x\in\mathcal{X}\subset^d$和$f(x)=(Wx)$形式的线性阈值哈希映射，其中$W\in^k\times d$。给定训练点$x_1,\ldots,x_n$ ,我们再次考虑上面讨论的两个模型：$v_1,\ldots,v_n$ 。也就是说，我们通过使用附加约束$U=(W_qX)$以及$V= ( W_{d}X)$ (对于 Lin:Lin )优化问题 (3)来学习，其中 $X=[x_1\ldots x_n]\in^{d\times n}$
我们通过交替更新$W_q$ 的行和 $W_d$ (对于 Lin:Lin )或$V$ (对于 Lin:V)的行来优化这些问题。要了解
这些更新，让我们首先返回到(3)(具有不受约束的$U,V$ ),并考虑更新$U$的行$u^(t)\in^n$。表示
$$Y^{(t)}=U^\top V-\theta\mathbb{1}_n-u^{(t)}^\top v^{(t)},$$
5
|
[
[
0.1422119140625,
0.09344482421875,
0.88134765625,
0.09344482421875,
0.88134765625,
0.106689453125,
0.1422119140625,
0.106689453125
],
[
0.11767578125,
0.10919189453125,
0.88134765625,
0.10919189453125,
0.88134765625,
0.12249755859375,
0.11767578125,
0.12249755859375
],
[
0.11767578125,
0.12371826171875,
0.8798828125,
0.12371826171875,
0.8798828125,
0.136962890625,
0.11767578125,
0.136962890625
],
[
0.1192626953125,
0.1395263671875,
0.615234375,
0.1395263671875,
0.615234375,
0.15283203125,
0.1192626953125,
0.15283203125
],
[
0.11358642578125,
0.1767578125,
0.6806640625,
0.177978515625,
0.6806640625,
0.1956787109375,
0.11358642578125,
0.1944580078125
],
[
0.1192626953125,
0.2083740234375,
0.88134765625,
0.2083740234375,
0.88134765625,
0.2215576171875,
0.1192626953125,
0.2215576171875
],
[
0.1192626953125,
0.222900390625,
0.88134765625,
0.222900390625,
0.88134765625,
0.236083984375,
0.1192626953125,
0.236083984375
],
[
0.1168212890625,
0.238037109375,
0.88134765625,
0.2374267578125,
0.88134765625,
0.250732421875,
0.1168212890625,
0.251953125
],
[
0.11767578125,
0.25244140625,
0.88134765625,
0.25244140625,
0.88134765625,
0.266357421875,
0.11767578125,
0.266357421875
],
[
0.11767578125,
0.268310546875,
0.88134765625,
0.268310546875,
0.88134765625,
0.2822265625,
0.11767578125,
0.2822265625
],
[
0.12091064453125,
0.285400390625,
0.8798828125,
0.285400390625,
0.8798828125,
0.29541015625,
0.12091064453125,
0.29541015625
],
[
0.1168212890625,
0.29736328125,
0.88134765625,
0.29736328125,
0.88134765625,
0.3125,
0.1168212890625,
0.3125
],
[
0.11517333984375,
0.313232421875,
0.60107421875,
0.3125,
0.60107421875,
0.326904296875,
0.11517333984375,
0.32763671875
],
[
0.1422119140625,
0.328857421875,
0.88134765625,
0.328857421875,
0.88134765625,
0.34228515625,
0.1422119140625,
0.34228515625
],
[
0.1192626953125,
0.343505859375,
0.8798828125,
0.343505859375,
0.8798828125,
0.356689453125,
0.1192626953125,
0.356689453125
],
[
0.131591796875,
0.35791015625,
0.8798828125,
0.35791015625,
0.8798828125,
0.371826171875,
0.131591796875,
0.371826171875
],
[
0.1192626953125,
0.373779296875,
0.79345703125,
0.373779296875,
0.79345703125,
0.386962890625,
0.1192626953125,
0.386962890625
],
[
0.1422119140625,
0.38818359375,
0.88134765625,
0.38818359375,
0.88134765625,
0.40283203125,
0.1422119140625,
0.40283203125
],
[
0.11767578125,
0.404052734375,
0.88134765625,
0.404052734375,
0.88134765625,
0.417236328125,
0.11767578125,
0.417236328125
],
[
0.1168212890625,
0.419189453125,
0.88134765625,
0.419189453125,
0.88134765625,
0.43310546875,
0.1168212890625,
0.43310546875
],
[
0.1192626953125,
0.436279296875,
0.8798828125,
0.436279296875,
0.8798828125,
0.4462890625,
0.1192626953125,
0.4462890625
],
[
0.1192626953125,
0.449462890625,
0.8798828125,
0.449462890625,
0.8798828125,
0.46337890625,
0.1192626953125,
0.46337890625
],
[
0.11767578125,
0.464111328125,
0.88134765625,
0.464111328125,
0.88134765625,
0.477783203125,
0.11767578125,
0.477783203125
],
[
0.1192626953125,
0.48095703125,
0.8798828125,
0.48095703125,
0.8798828125,
0.4912109375,
0.1192626953125,
0.4912109375
],
[
0.11767578125,
0.49560546875,
0.88134765625,
0.49560546875,
0.88134765625,
0.5087890625,
0.11767578125,
0.5087890625
],
[
0.11767578125,
0.5087890625,
0.2646484375,
0.5087890625,
0.2646484375,
0.52294921875,
0.11767578125,
0.52294921875
],
[
0.1405029296875,
0.5244140625,
0.88134765625,
0.5244140625,
0.88134765625,
0.53857421875,
0.1405029296875,
0.53857421875
],
[
0.1192626953125,
0.54052734375,
0.88134765625,
0.54052734375,
0.88134765625,
0.5537109375,
0.1192626953125,
0.5537109375
],
[
0.1168212890625,
0.5546875,
0.88134765625,
0.5546875,
0.88134765625,
0.568359375,
0.1168212890625,
0.568359375
],
[
0.11517333984375,
0.568359375,
0.78759765625,
0.5693359375,
0.78759765625,
0.583984375,
0.11517333984375,
0.58349609375
],
[
0.1422119140625,
0.58544921875,
0.88134765625,
0.58544921875,
0.88134765625,
0.5986328125,
0.1422119140625,
0.5986328125
],
[
0.11767578125,
0.60107421875,
0.8798828125,
0.60107421875,
0.8798828125,
0.6142578125,
0.11767578125,
0.6142578125
],
[
0.11767578125,
0.61572265625,
0.75341796875,
0.61572265625,
0.75341796875,
0.62890625,
0.11767578125,
0.62890625
],
[
0.7548828125,
0.6162109375,
0.87744140625,
0.6162109375,
0.87744140625,
0.626953125,
0.7548828125,
0.626953125
],
[
0.11767578125,
0.6298828125,
0.53759765625,
0.6298828125,
0.53759765625,
0.64306640625,
0.11767578125,
0.64306640625
],
[
0.54345703125,
0.6298828125,
0.88134765625,
0.6298828125,
0.88134765625,
0.64453125,
0.54345703125,
0.64453125
],
[
0.1192626953125,
0.64599609375,
0.88134765625,
0.64599609375,
0.88134765625,
0.6591796875,
0.1192626953125,
0.6591796875
],
[
0.11767578125,
0.662109375,
0.65283203125,
0.662109375,
0.65283203125,
0.67236328125,
0.11767578125,
0.67236328125
],
[
0.1405029296875,
0.67529296875,
0.88232421875,
0.67529296875,
0.88232421875,
0.689453125,
0.1405029296875,
0.689453125
],
[
0.1168212890625,
0.689453125,
0.73193359375,
0.69140625,
0.73193359375,
0.70556640625,
0.1168212890625,
0.7041015625
],
[
0.11358642578125,
0.72900390625,
0.291748046875,
0.72998046875,
0.291748046875,
0.74560546875,
0.11358642578125,
0.744140625
],
[
0.11273193359375,
0.7548828125,
0.88232421875,
0.755859375,
0.88232421875,
0.7763671875,
0.11273193359375,
0.77587890625
],
[
0.12091064453125,
0.77587890625,
0.43310546875,
0.77587890625,
0.43310546875,
0.7890625,
0.12091064453125,
0.7890625
],
[
0.44287109375,
0.7744140625,
0.88134765625,
0.7744140625,
0.88134765625,
0.7890625,
0.44287109375,
0.7890625
],
[
0.11517333984375,
0.7890625,
0.6806640625,
0.7890625,
0.6806640625,
0.8037109375,
0.11517333984375,
0.8037109375
],
[
0.1405029296875,
0.8046875,
0.88134765625,
0.8046875,
0.88134765625,
0.8193359375,
0.1405029296875,
0.8193359375
],
[
0.11767578125,
0.8193359375,
0.80615234375,
0.8193359375,
0.80615234375,
0.833984375,
0.11767578125,
0.833984375
],
[
0.38818359375,
0.84765625,
0.611328125,
0.84765625,
0.611328125,
0.86572265625,
0.38818359375,
0.86572265625
],
[
0.491943359375,
0.8984375,
0.50732421875,
0.8984375,
0.50732421875,
0.91357421875,
0.491943359375,
0.91357421875
]
] |
[
"在进行样本外概括之前，我们简要报告了根据经验找到具有对称和非对称代码的实际相似性矩阵的良好",
"近似值所需的位数。我们尝试了几个数据集，试图用对称和非对称代码来拟合它们，然后通过改变阅值来",
"计算平均精度（同时保持U和V固定）。两个相似度矩阵的结果，一个基于均匀分布在超立方体中的点之",
"间的欧几里得距离，另一个基于图像之间的语义相似度，如图？？所示",
"4 Out of Sample Generalization: Learning a Mapping",
"到目前为止，我们专注于通过将任意代码字与每个对象相关联并完全忽略对象的输入表示来学习一组",
"固定对象上的二进制代码。我们只讨论了如何以及二进制哈希可以approximate的相似性，但没有考虑",
"generalizing额外的新对象。然而，在大多数应用程序中，我们希望能够有这样一个样本外的泛化。也就是",
"说，我们想仅使用有限的对象训练集来学习无限域上的映射f：→{士1}，然后应用该映射以获得要",
"遇到的未来对象的二进制码f(α），例如S(,a')～（f(α),f(a'））。因此，映射f：→{士1)通常仅",
"限于某些受约束的参数类，因此我们可以在新对象上有效地表示和评估它，并确保良好的泛化。例如，当",
"=d时，我们可以考虑线性阈值映射fw(a）=（Wr），其中Wekxd和（)按元素操作，如在最小损失哈",
"希[8]中。或者，我们也可以考虑更复杂的类，例如多层网络[11,9]",
"我们已经看到，非对称二进制码可以使用较短的代码进行更好的近似，因此在这里寻求非对称代码也很",
"自然。也就是说，不是学习单个参数映射f(c），而是可以学习一对映射f：→【士1)k和g：→（士1]k",
"它们都约束为某个参数类，以及阈值，例如S(α,r'）～（f(),g(α）-の）。这有可能允许更好地近似相",
"似性，从而用更短的代码获得更好的总体精度（尽管由于参数数量的增加可能会稍微更难概括）",
"事实上，在一个典型的应用程序中，对象数据库被哈希化以在未来的查询中进行相似性搜索，不对称性",
"使我们能够走得更远。考虑以下设置：我们从一些无限域中获得n对象a1，.，anE，以及这些对象",
"之间的相似性S(ci，&,）。我们的目标是使用短二进制代码来哈希这些对象，这将使我们能够快速计算这些",
"对象（“数据库”）和未来对象3（“查询”）之间的近似相似性。也就是说，我们希望为数据库中的对象生成",
"并存储紧凑的二进制代码。然后，给定一个新的查询对象，我们希望有效地计算给定查询的紧凑二进制代码",
"并通过在数据库中查找与查询二进制代码相距较近的二进制代码，非常快速地检索数据库中的类似项目。回",
"想一下，重要的是要确保哈希的位长度很小，因为短代码充许非常快速的汉明距离计算和低通信成本，如果",
"代码需要远程发送。更重要的是，如果我们想将数据库存储在充许立即查找的哈希表中，哈希表的大小在代",
"码长度上是指数级的",
"对称二进制哈希方法（e。g。E[8]），将是找到单个参数映射f：α→{士1)，以便S(a,i）～~",
"(f(α),f(ci）一の）用于将来的查询和数据库对象ci，计算所有数据库对象i的f(ci），并存储这些哈",
"希（可能在哈希表中允许在短的汉明距离内快速检索代码）。上面描述的非对称方法是找到两个参数映射",
"f:→{±1)和g:→(±1)h，使得 S(r,ai)~(f(r),g(ci)-の)，然后计算并存储g(ai)",
"但是，如果数据库是固定的，我们可以走得更远。实际上，9(）不需要位于受约束的参数类中，因为我们不",
"需要将g(）推广到未来的对象，也不需要高效地计算它，也不需要将g(）传递到数据库。因此，我们可以考",
"虑允许数据库哈希函数g）是任意映射。也就是说，我们的目标是为数据库中的每个1，",
"：，3n找到一个简",
"单的参数映射f：X→{±1}和narbitrarycodewordsu",
"...,Un E (1)h,以便 S(c,ri) ~ (f(r), Vi-0)",
"用于将来的查询a和数据库中的对象i，，n。这种形式的不对称可以使我们获得更大的近似能力，从而",
"在不增加额外的计算或存储成本的情况下，通过更短的代码获得更好的精度",
"在6节中，我们对上述两种非对称策略进行了经验评估，并证明了它们的好处。但是在此之前，在下一",
"节中，我们将讨论一种局部搜索方法，用于查找映射f,g或映射f和代码1，..，Un。",
"5Optimization",
"我们专注于aecd和f(r)=（Wa）形式的线性阈值哈希映射，其中Wekxd。给定训练点a1,.,an",
"，我们再次考虑上面讨论的两个模型：U1，",
"，Un。也就是说，我们通过使用附加约束U=（W.X）以及",
"V=（WaX）（对于Lin:Lin）优化问题（3）来学习，其中X=[ri...an]edxn",
"我们通过交替更新W的行和Wa（对于Lin:Lin）或V（对于Lin:V）的行来优化这些问题。要了解",
"这些更新，让我们首先返回到（3）（具有不受约束的U,V），并考虑更新U的行u(t）En。表示",
"y(t) =UTV-01n-u(t)v(t),",
"5"
] |
{
"arxiv_id": "1311.7662",
"language": "zh",
"page_idx": 4
}
|
||
2403.13749-zh-0
|
2403.13749-zh-0.jpg
|
# $\textbf{Weisfeiler and Leman Go Loopy: }$ A New Hierarchy for Graph Representational Learning
Raffaele Paolino $^{*12}$ Sohir Maskey $^{*1}$ Pascal Welke$^{3}$ Gitta Kutyniok$^{12}$
???。3in
Abstract
We introduce $r$ -loopy Weisfeiler-Leman $(r-$ $\ell$WL ), a novel hierarchy of graph isomorphism tests and a corresponding GNN framework, $r$ - $\ell$MPNN , that can count cycles up to length $r+ 2. \textbf{Most notably, we show that }r- \ell$WL can count homomorphisms of cactus graphs. This strictly extends classical 1-WL, which can only count homomorphisms of trees and, in fact, is incomparable to $k$ -WL for any fixed $k.$ We empirically validate the expressive and counting power of the proposed $r-\ell$MPNN on several synthetic datasets and present state-of-the-art predictive performance on various real-world datasets. The code is available online .
子结构以及设计超越 MPNN 有限功能的体系结构至关
重要
已经提出了许多模型来匹配或超越 WL 的表达能力。几个人从 WL 算法 (Morris et al., 2019) 的高阶变体中汲取灵感，使他们能够计算更广泛的子结构。例如，设计用于模拟 3 -WL 算法的 GNN 可以对长度为 7 的周期进行计数。然而，这种增强的表达能力是以计算成本为代价的，限制了它们的适用范围。值得注意的是， 3 -WL 算法已经构造了一个大的计算图，导致了立方向前的复杂性。因此，迫切需要设计尊重现实世界图的固有稀疏性的富有表现力的 GNN
为了响应对更具表现力和可扩展性的神经体系结构(Morris et al., 2023) 的需求，我们引人了一类新颖的颜色细化算法，称为$r$ -loopyWeisfeiler-Leman 测试$(r$ - $\ell$WL ),并构造了相应的 GNN 类，称为$r$ -loopy 图。基本思想是不仅从相邻节点收集消息，而且从连接任何两个不同相邻节点的路径收集消息。这种轻微的修改提升了架构的能力，可以对长度为$r+2$的周期进行计数，超越了经典的$k$ -WL 层次结构。具体来说，对于每个$k$ ,都存在一个 $r$ ,使得 $r-\ell$WL 的功能不低于$k$ -WL。我们在$r$ - $\ell$WL 测试中建立了层次结构，证明增加$r$严格增强了表达能力。此外，我们在我们提出的算法和仙人掌图(树的严格泛化)之间建立了联系， 证明$r-\ell$WL 可以同态计数任何仙人掌图。为了证实我们的理论结果，我们对合成数据集进行了全面的评估。最后，我们将提出的模型应用于现实世界的任务， 展示了它在基准数据集上的竞争表现
## $1. \textbf{ Introduction}$
图形神经网络 (GNNs) (Scarselli et al.,2009; Bronstein et al., 2017) 已成为处理图形结构数据的流行体系结构， 对各种应用科学做出了重大贡献。值得注意的应用包括新抗生素 (Stokes et al.,2020)的发现，社交推荐系统(Fan et al., 2019) 的进步以及假新闻检测(Monti et al., 2019)的改进。在各种类型的 GNN 中，消息传递神经网络(MPNN) (Gilmer et al., 2017)在实践中被广泛使用。基于消息传递范式，MPNN 利用本地邻域聚合来计算图表示
尽管 MPNN 取得了广泛的成功，但其表示能力仍受到Weisfeiler-Leman (WL)测试的限制，Weisfeiler-Leman (WL) 测试是用于图同构测试(Xu et al., 2019; Morris et al., 2019) 的经典算法。这种限制阻碍了 MPNN 识别甚至基本的子结构，如循环(Chen et al.,2020)。然而， 特定的子结构在许多应用中可能是至关重要的。例如， 在有机化学或生物信息学中，不同类型的循环会影响基础分子 (Deshpande et al., 2002; Koyutürk et al., 2004) 的各种化学性质。因此，研究 GNN 是否可以计算某些$^{*}$Equal contribution$^{1}$Department of Mathematics, Ludwig-
## $2. \textbf{ Related Work}$
在有关表现力的早期工作中，Xu et al.(2019) and Morris et al.(2019) 证明了 MPNN 的表达能力受 1-WL 限制。后续作品 (Maron et al., 2018; Morris et al., 2019) 引人了与$k$ -WL (Geerts et al., 2022) 具有相同表达能力的高阶 GNN。尽管这些网络是通用的 (Maron et al., 2019b; Keriven et al., 2019) , 但它们在$k$中的指数时间和空间复杂性使它们不切实际。为了解决这些限制，引入了$k$ -WL 的本地变体和相应的本地$k$ -GNN (Morris et al., 2020)。此外，提出了$k$ -跳 GNNs (Abboud et al., 2022) ,增强了1-WL以外但在3-WL(Fengetal.,2022)内的表达能力。另一条工作线通过唯一节点标识符 (Vignac et al., 2020)、随机特征(Abboud et al., 2021; Sato et al., 2021) 或特征向量 (Lim et al., 2022; Maskey et al., 2022)
$\dot{\text{Maximilians-Universität München, Munich, Germany}}^2$Munich Center for Machine Learning (MCML), Munich, Germany $^{3}$TU Wien, Vienna, Austria. Correspondence to: Raffaele Paolino< paolino@ math.lmu.de>, Sohir $\dot{\text{Maskey}}<$maskey@ math.lmu.de >.
1
|
[
[
0.3203125,
0.11236572265625,
0.654296875,
0.1142578125,
0.654296875,
0.1319580078125,
0.3203125,
0.1300048828125
],
[
0.2091064453125,
0.1363525390625,
0.7646484375,
0.136962890625,
0.7646484375,
0.1546630859375,
0.2091064453125,
0.154052734375
],
[
0.2254638671875,
0.2008056640625,
0.74658203125,
0.2008056640625,
0.74658203125,
0.21533203125,
0.2254638671875,
0.21533203125
],
[
0.08905029296875,
0.2197265625,
0.1502685546875,
0.2197265625,
0.1502685546875,
0.2310791015625,
0.08905029296875,
0.2310791015625
],
[
0.24267578125,
0.2291259765625,
0.3203125,
0.2310791015625,
0.3203125,
0.246826171875,
0.24267578125,
0.2442626953125
],
[
0.50341796875,
0.231689453125,
0.8837890625,
0.231689453125,
0.8837890625,
0.24560546875,
0.50341796875,
0.24560546875
],
[
0.5009765625,
0.2442626953125,
0.53759765625,
0.2474365234375,
0.53515625,
0.261962890625,
0.5,
0.260009765625
],
[
0.1217041015625,
0.251953125,
0.438720703125,
0.251953125,
0.438720703125,
0.26513671875,
0.1217041015625,
0.26513671875
],
[
0.12335205078125,
0.267578125,
0.439453125,
0.267578125,
0.439453125,
0.281005859375,
0.12335205078125,
0.281005859375
],
[
0.50244140625,
0.270751953125,
0.88232421875,
0.270751953125,
0.88232421875,
0.2841796875,
0.50244140625,
0.2841796875
],
[
0.12091064453125,
0.2822265625,
0.439453125,
0.2822265625,
0.439453125,
0.29541015625,
0.12091064453125,
0.29541015625
],
[
0.50244140625,
0.285400390625,
0.8837890625,
0.285400390625,
0.8837890625,
0.298583984375,
0.50244140625,
0.298583984375
],
[
0.1192626953125,
0.29736328125,
0.441162109375,
0.29736328125,
0.441162109375,
0.311279296875,
0.1192626953125,
0.311279296875
],
[
0.50244140625,
0.300537109375,
0.8837890625,
0.300537109375,
0.8837890625,
0.314453125,
0.50244140625,
0.314453125
],
[
0.12091064453125,
0.3125,
0.441162109375,
0.3125,
0.441162109375,
0.32568359375,
0.12091064453125,
0.32568359375
],
[
0.50244140625,
0.31494140625,
0.8857421875,
0.31494140625,
0.8857421875,
0.328857421875,
0.50244140625,
0.328857421875
],
[
0.1217041015625,
0.32763671875,
0.441162109375,
0.32763671875,
0.441162109375,
0.341552734375,
0.1217041015625,
0.341552734375
],
[
0.50244140625,
0.330810546875,
0.8857421875,
0.330810546875,
0.8857421875,
0.3447265625,
0.50244140625,
0.3447265625
],
[
0.1217041015625,
0.34228515625,
0.439453125,
0.34228515625,
0.439453125,
0.35595703125,
0.1217041015625,
0.35595703125
],
[
0.50244140625,
0.34521484375,
0.8857421875,
0.34521484375,
0.8857421875,
0.359130859375,
0.50244140625,
0.359130859375
],
[
0.12335205078125,
0.35791015625,
0.441162109375,
0.35791015625,
0.441162109375,
0.371826171875,
0.12335205078125,
0.371826171875
],
[
0.5009765625,
0.35986328125,
0.8837890625,
0.35986328125,
0.8837890625,
0.373779296875,
0.5009765625,
0.373779296875
],
[
0.12091064453125,
0.37255859375,
0.441162109375,
0.37255859375,
0.441162109375,
0.386474609375,
0.12091064453125,
0.386474609375
],
[
0.50341796875,
0.375732421875,
0.8857421875,
0.375732421875,
0.8857421875,
0.389404296875,
0.50341796875,
0.389404296875
],
[
0.12091064453125,
0.386962890625,
0.441162109375,
0.38818359375,
0.441162109375,
0.40283203125,
0.12091064453125,
0.401611328125
],
[
0.5009765625,
0.389404296875,
0.75244140625,
0.389404296875,
0.75244140625,
0.40283203125,
0.5009765625,
0.40283203125
],
[
0.1217041015625,
0.40283203125,
0.441162109375,
0.40283203125,
0.441162109375,
0.416748046875,
0.1217041015625,
0.416748046875
],
[
0.12091064453125,
0.416748046875,
0.441162109375,
0.417236328125,
0.441162109375,
0.43115234375,
0.12091064453125,
0.429931640625
],
[
0.50244140625,
0.41357421875,
0.8857421875,
0.41357421875,
0.8857421875,
0.4267578125,
0.50244140625,
0.4267578125
],
[
0.1217041015625,
0.43310546875,
0.438720703125,
0.43310546875,
0.438720703125,
0.4462890625,
0.1217041015625,
0.4462890625
],
[
0.50341796875,
0.429931640625,
0.8837890625,
0.429931640625,
0.8837890625,
0.43994140625,
0.50341796875,
0.43994140625
],
[
0.1192626953125,
0.4462890625,
0.313720703125,
0.448974609375,
0.313720703125,
0.462158203125,
0.1192626953125,
0.460205078125
],
[
0.50341796875,
0.443115234375,
0.88232421875,
0.443115234375,
0.88232421875,
0.45703125,
0.50341796875,
0.45703125
],
[
0.50244140625,
0.458984375,
0.8857421875,
0.458984375,
0.8857421875,
0.472900390625,
0.50244140625,
0.472900390625
],
[
0.50341796875,
0.47607421875,
0.8837890625,
0.47607421875,
0.8837890625,
0.486083984375,
0.50341796875,
0.486083984375
],
[
0.0882568359375,
0.4892578125,
0.221435546875,
0.4912109375,
0.221435546875,
0.505859375,
0.0882568359375,
0.50390625
],
[
0.50244140625,
0.4892578125,
0.8857421875,
0.4892578125,
0.8857421875,
0.50244140625,
0.50244140625,
0.50244140625
],
[
0.50244140625,
0.50390625,
0.88232421875,
0.50390625,
0.88232421875,
0.517578125,
0.50244140625,
0.517578125
],
[
0.0906982421875,
0.5166015625,
0.473876953125,
0.5166015625,
0.473876953125,
0.52978515625,
0.0906982421875,
0.52978515625
],
[
0.5009765625,
0.51806640625,
0.88232421875,
0.51806640625,
0.88232421875,
0.5322265625,
0.5009765625,
0.5322265625
],
[
0.08905029296875,
0.5322265625,
0.473876953125,
0.5322265625,
0.473876953125,
0.54541015625,
0.08905029296875,
0.54541015625
],
[
0.5,
0.53271484375,
0.88134765625,
0.53271484375,
0.88134765625,
0.546875,
0.5,
0.546875
],
[
0.08905029296875,
0.546875,
0.473876953125,
0.546875,
0.473876953125,
0.56005859375,
0.08905029296875,
0.56005859375
],
[
0.5009765625,
0.548828125,
0.8857421875,
0.548828125,
0.8857421875,
0.5625,
0.5009765625,
0.5625
],
[
0.08905029296875,
0.56103515625,
0.473876953125,
0.56103515625,
0.473876953125,
0.57470703125,
0.08905029296875,
0.57470703125
],
[
0.5009765625,
0.564453125,
0.88232421875,
0.564453125,
0.88232421875,
0.57763671875,
0.5009765625,
0.57763671875
],
[
0.08905029296875,
0.5771484375,
0.473876953125,
0.5771484375,
0.473876953125,
0.59033203125,
0.08905029296875,
0.59033203125
],
[
0.50341796875,
0.58056640625,
0.88134765625,
0.58056640625,
0.88134765625,
0.59130859375,
0.50341796875,
0.59130859375
],
[
0.0906982421875,
0.59130859375,
0.473876953125,
0.59130859375,
0.473876953125,
0.60498046875,
0.0906982421875,
0.60498046875
],
[
0.5009765625,
0.5947265625,
0.8857421875,
0.5947265625,
0.8857421875,
0.60791015625,
0.5009765625,
0.60791015625
],
[
0.09234619140625,
0.60693359375,
0.473876953125,
0.60693359375,
0.473876953125,
0.62060546875,
0.09234619140625,
0.62060546875
],
[
0.5009765625,
0.60986328125,
0.8837890625,
0.60986328125,
0.8837890625,
0.62353515625,
0.5009765625,
0.62353515625
],
[
0.0906982421875,
0.6220703125,
0.473876953125,
0.6220703125,
0.473876953125,
0.63623046875,
0.0906982421875,
0.63623046875
],
[
0.50244140625,
0.625,
0.8857421875,
0.625,
0.8857421875,
0.63818359375,
0.50244140625,
0.63818359375
],
[
0.0882568359375,
0.63525390625,
0.17236328125,
0.63720703125,
0.17236328125,
0.6513671875,
0.0882568359375,
0.6494140625
],
[
0.5009765625,
0.64013671875,
0.76318359375,
0.64013671875,
0.76318359375,
0.65380859375,
0.5009765625,
0.65380859375
],
[
0.0906982421875,
0.6591796875,
0.471435546875,
0.6591796875,
0.471435546875,
0.67236328125,
0.0906982421875,
0.67236328125
],
[
0.0882568359375,
0.673828125,
0.47216796875,
0.6748046875,
0.47216796875,
0.68798828125,
0.0882568359375,
0.6875
],
[
0.5009765625,
0.6748046875,
0.6435546875,
0.6748046875,
0.6435546875,
0.68798828125,
0.5009765625,
0.68798828125
],
[
0.08660888671875,
0.689453125,
0.471435546875,
0.68994140625,
0.471435546875,
0.7041015625,
0.08660888671875,
0.70263671875
],
[
0.0882568359375,
0.7041015625,
0.473876953125,
0.705078125,
0.473876953125,
0.7197265625,
0.0882568359375,
0.71826171875
],
[
0.5009765625,
0.70068359375,
0.8857421875,
0.70068359375,
0.8857421875,
0.7138671875,
0.5009765625,
0.7138671875
],
[
0.08905029296875,
0.72021484375,
0.473876953125,
0.72021484375,
0.473876953125,
0.734375,
0.08905029296875,
0.734375
],
[
0.5009765625,
0.71533203125,
0.8857421875,
0.71533203125,
0.8857421875,
0.72900390625,
0.5009765625,
0.72900390625
],
[
0.08905029296875,
0.73486328125,
0.4755859375,
0.73486328125,
0.4755859375,
0.74853515625,
0.08905029296875,
0.74853515625
],
[
0.50244140625,
0.72998046875,
0.8837890625,
0.72998046875,
0.8837890625,
0.74365234375,
0.50244140625,
0.74365234375
],
[
0.5009765625,
0.744140625,
0.8857421875,
0.74560546875,
0.8857421875,
0.7587890625,
0.5009765625,
0.75830078125
],
[
0.08905029296875,
0.75048828125,
0.473876953125,
0.75048828125,
0.473876953125,
0.7646484375,
0.08905029296875,
0.7646484375
],
[
0.0906982421875,
0.76513671875,
0.473876953125,
0.76513671875,
0.473876953125,
0.77880859375,
0.0906982421875,
0.77880859375
],
[
0.5009765625,
0.76025390625,
0.8857421875,
0.76025390625,
0.8857421875,
0.7734375,
0.5009765625,
0.7734375
],
[
0.5009765625,
0.7744140625,
0.8837890625,
0.7744140625,
0.8837890625,
0.78857421875,
0.5009765625,
0.78857421875
],
[
0.0906982421875,
0.78076171875,
0.47216796875,
0.78076171875,
0.47216796875,
0.79443359375,
0.0906982421875,
0.79443359375
],
[
0.50341796875,
0.79248046875,
0.88134765625,
0.79248046875,
0.88134765625,
0.80322265625,
0.50341796875,
0.80322265625
],
[
0.108642578125,
0.8037109375,
0.471435546875,
0.8037109375,
0.471435546875,
0.818359375,
0.108642578125,
0.818359375
],
[
0.5009765625,
0.80615234375,
0.8857421875,
0.80615234375,
0.8857421875,
0.8193359375,
0.5009765625,
0.8193359375
],
[
0.08905029296875,
0.8173828125,
0.47216796875,
0.8173828125,
0.47216796875,
0.83056640625,
0.08905029296875,
0.83056640625
],
[
0.5009765625,
0.82080078125,
0.8857421875,
0.82080078125,
0.8857421875,
0.833984375,
0.5009765625,
0.833984375
],
[
0.08905029296875,
0.830078125,
0.471435546875,
0.8291015625,
0.471435546875,
0.84228515625,
0.08905029296875,
0.84326171875
],
[
0.5087890625,
0.84033203125,
0.515625,
0.84033203125,
0.515625,
0.84521484375,
0.5087890625,
0.84521484375
],
[
0.51123046875,
0.8359375,
0.8837890625,
0.8359375,
0.8837890625,
0.849609375,
0.51123046875,
0.849609375
],
[
0.08905029296875,
0.84228515625,
0.4697265625,
0.84228515625,
0.4697265625,
0.85595703125,
0.08905029296875,
0.85595703125
],
[
0.08905029296875,
0.85498046875,
0.473876953125,
0.85498046875,
0.473876953125,
0.86865234375,
0.08905029296875,
0.86865234375
],
[
0.5009765625,
0.8515625,
0.8837890625,
0.8515625,
0.8837890625,
0.86572265625,
0.5009765625,
0.86572265625
],
[
0.0906982421875,
0.8720703125,
0.1004638671875,
0.8720703125,
0.1004638671875,
0.87890625,
0.0906982421875,
0.87890625
],
[
0.50244140625,
0.8662109375,
0.88232421875,
0.8662109375,
0.88232421875,
0.8798828125,
0.50244140625,
0.8798828125
],
[
0.5009765625,
0.8818359375,
0.8857421875,
0.8818359375,
0.8857421875,
0.89599609375,
0.5009765625,
0.89599609375
],
[
0.48046875,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93505859375,
0.48046875,
0.93505859375
]
] |
[
"Weisfeiler and Leman Go Loopy:",
"A New Hierarchy for Graph Representational Learning",
"Raffaele Paolino * 12 Sohir Maskey * 1 Pascal Welke3 Gitta Kutyniok12",
"???。3in",
"Abstract",
"子结构以及设计超越MPNN有限功能的体系结构至关",
"重要",
"We introduce r -loopy Weisfeiler-Leman ( r",
"(WL),a novel hierarchy of graph isomorphism",
"已经提出了许多模型来匹配或超越WL的表达能力，",
"tests and a corresponding GNN framework, r",
"几个人从WL算法（Morrisetal.，2019）的高阶变体中汲",
"- IMPNN , that can count cycles up to length",
"取灵感，使他们能够计算更广泛的子结构。例如，设",
"r + 2. Most notably, we show that r - IWL can",
"计用于模拟3-WL算法的GNN可以对长度为7的周",
"count homomorphisms of cactus graphs. This",
"期进行计数。然而，这种增强的表达能力是以计算成",
"strictly extends classical 1-WL, which can only",
"本为代价的，限制了它们的适用范围。值得注意的是，",
"count homomorphisms of trees and, in fact, is",
"3-WL算法已经构造了一个大的计算图，导致了立方",
"incomparable to k -WL for any fixed k : We",
"向前的复杂性。因此，迫切需要设计尊重现实世界图",
"empirically validate the expressive and counting",
"的固有稀疏性的富有表现力的GNN",
"power of the proposed r - IMPNN on several syn-",
"thetic datasets and present state-of-the-art predic-",
"为了响应对更具表现力和可扩展性的神经体系结构",
"tive performance on various real-world datasets.",
"（Morrisetal.，2023）的需求，我们弓人了一类新颖的颜",
"Thecodeisavailableonline",
"色细化算法，称为r-loopyWeisfeiler-Leman测试（r",
"(WL），并构造了相应的GNN类，称为r-loopy图。基",
"本思想是不仅从相邻节点收集消息，而且从连接任何",
"1.Introduction",
"两个不同相邻节点的路径收集消息。这种轻微的修改",
"提升了架构的能力，可以对长度为r+2的周期进行计",
"图形神经网络（GNNs）（Scarsellietal.，2009;Bronstein",
"数，超越了经典的k-WL层次结构。具体来说，对于",
"etal.，2017）已成为处理图形结构数据的流行体系结构，",
"每个k，都存在一个，使得r-EWL的功能不低于",
"对各种应用科学做出了重大贡献。值得注意的应用包",
"k-WL。我们在r-LWL测试中建立了层次结构，证明",
"括新抗生素（Stokesetal.，2020）的发现，社交推荐系统",
"增加r严格增强了表达能力。此外，我们在我们提出",
"(Fanetal.，2019）的进步以及假新闻检测（Montietal.,",
"的算法和仙人掌图（树的严格泛化）之间建立了联系",
"2019）的改进。在各种类型的GNN中，消息传递神经",
"证明r-WL可以同态计数任何仙人掌图。为了证实",
"网络（MPNN）（Gilmeretal.，2017）在实践中被广泛使",
"我们的理论结果，我们对合成数据集进行了全面的评",
"用。基于消息传递范式，MPNN利用本地邻域聚合来",
"估。最后，我们将提出的模型应用于现实世界的任务，",
"计算图表示",
"展示了它在基准数据集上的竞争表现",
"尽管MPNN取得了广泛的成功，但其表示能力仍受到",
"Weisfeiler-Leman（WL）测试的限制，Weisfeiler-Leman",
"2. Related Work",
"（WL）测试是用于图同构测试（Xuetal.，2019Morris",
"etal.，2019）的经典算法。这种限制阻碍了MPNN识别",
"在有关表现力的早期工作中，Xuetal.（2019)andMorris",
"甚至基本的子结构，如循环（Chenetal.，2020）。然而，",
"etal.（2019）证明了MPNN的表达能力受1-WL限制。",
"特定的子结构在许多应用中可能是至关重要的。例如，",
"后续作品（Maronetal.，2018;Morrisetal.，2019）引人",
"了与k-WL（Geertsetal.,2022）具有相同表达能力的高",
"在有机化学或生物信息学中，不同类型的循环会影响",
"基础分子(Deshpande et al.,2002;Koyutirk etal.,2004)",
"阶GNN。尽管这些网络是通用的（Maronetal.，2019b;",
"Kerivenetal.,2019），但它们在k中的指数时间和空间",
"的各种化学性质。因此，研究GNN是否可以计算某些",
"复杂性使它们不切实际。为了解决这些限制，引入了",
"Equal contribution Department of Mathematics, Ludwig",
"k-WL的本地变体和相应的本地k-GNN(Morrisetal.,",
"Maximilians-Universitat Munchen, Munich, Germany 2 Munich",
"2020）。此外，提出了k-跳GNNs（Abboudetal.,2022)",
"Center for Machine Learning (MCML), Munich, Germany 3 TU",
"",
"增强了1-WL以外但在3-WL(Fengetal.,2022)内的",
"Wien, Vienna, Austria . Correspondence to: Raffaele Paolino ",
"paolino@math.lmu.de >, Sohir Maskey < maskey@math.lmu.de",
"表达能力。另一条工作线通过唯一节点标识符（Vignac",
">",
"etal.,2020）、随机特征（Abboud etal.,2021;Satoet al.",
"2021）或特征向量（Limetal.,2022;Maskeyetal.,2022)",
"1"
] |
{
"arxiv_id": "2403.13749",
"language": "zh",
"page_idx": 0
}
|
||
2403.13749-zh-1
|
2403.13749-zh-1.jpg
|
利用位置编码来增强标准 MPNN 的表达能力
子图 GNNs (Bevilacqua et al., 2021; You et al., 2021; Frasca et al., 2022; Huang et al., 2022) 通过将标准 MPNN 分别应用于从初始输入图中选择的一袋子图来计算图表示。已知子图 GNN 比 1-WL 更具表现力，但上限为3-WL (Frasca et al., 2022) 。B. Zhang et al. (2022) 证明了ESAN (Bevilacqua et al., 2021) 能够计算图的双连接性， 这是一种相对容易计算的图统计量。此外，Michel et al. (2023) 探索处理图中路径以增强其表达能力的 GNN
Lovász (1967) 表明同态计数 (请参阅 Defnition 7) 可作为完整的图不变式，这意味着当且仅当它们的同态计数相同时，两个图才是同构的。在此完整性属性的基础上，Nguyen et al.(2020) and Welke et al.(2023) 利用同态计数来开发可表达的 GNN。Tinhofer (1986, 1991) 的理论贡献建立了1-WL 与树宽为 1 的图的同态计数能力之间的等价性。将这些结果扩展到$k$ -WL, Dell et al.(2018)证明了$k$-WL与从具有树宽度$k$的图中计数同态的能力之间的等效性。最近，B. Zhang et al. (2024)引人了一个定量框架，该框架完全表征了GNN 的表示能力。扩展到$k$ -WL 层次结构之外，他们主张将同态计数作为表达率的完整度量，因为 GNN 体系结构已被证明可以对特定的基序家族进行同态计数
在我们的方法中，我们首先通过将我们的体系结构与$k$ -WL 层次结构相关联，从更常规的角度呈现结果。这为后续探索子结构奠定了基础，我们可以有效地进行同态计数
# 3. Preliminaries
让$G$是一个简单且无向的图。我们用$V( G)$ := $\{ 1, 2, \ldots , N\}$ 表示节点集，用$E( G)$ $\subset$ $\{ e\mid e\in 2^{V( G) }, | e| = 2\}$ 表示边缘集。对于给定的节点$v\in V(G)$,直接邻域定义为
$$\mathcal{N}(v):=\{u\in V(G)\mid\{v,u\}\in E(G)\}\:.$$
我们用$\mathcal{G}$ 表示所有图的集合
Definition 1 (Homomorphism). Let F and G be two graphs. A homomorphism from F to G is a map h : $V( F) \to V( G) \textit{ such that}$
$$\{u,v\}\in E(F)\implies\{h(u),h(v)\}\in E(G).$$
The set $of$ homomorphisms from $F$ $to$ $G$ $is$ $de- noted$ $by$ Hom$( F, G)$ , and its cardinality $by$
$\hom(F,G):=\left|\operatorname{Hom}(F,G)\right|.$
直观上，同态是边缘保留映射，i。e。,它将$F$的相邻
顶点映射到$G$的相邻顶点。
Definition 2 (Subgraph Isomorphism). A subgraph isomorphism is an injective homomorphism. The set of subgraph isomorphismsfromFtoGisdenotedby Sub$(F,G),and$ $its\textit{ cardinality by sub}( F, G) = |$Sub$(F,G)|.$
<FigureHere>
Figure 1: Examples of non-injective homomorphism (row 1), subgraph isomorphism (row 2), bijective homomorphism with non-homomorphic inverse (row 3), and isomorphism (row 4). For better clarity, the mappings $h:$ $V(\bar{F})\to V(G)$ are visually represented with colors, where $F$ is consistently on the left, and $G$ is on the right in each $row.$
子图同构$h$将不同的节点映射到不同的节点因此，它
还将不同的边映射到不同的边。
Definition 3 (Isomorphism). An isomorphism is a bijective homomorphism whose inverse is also a homomorphism. If $F\textit{is isomorphic to G, we write F}\cong G. The\textit{set of iso- }$ morphisms from F to G is denoted by Ind(F,G) , and its cardinality $by$ ind$(F,G)=|$Ind$( F, G) |$ .
同构是节点之间的一对一对应关系。对逆是同态的要求也保证了边之间的一对一对应。我们参考 Figure 1来说明这些概念之间的区别。值得注意的是，当且仅当$F$ 与$G$ 的子图同构时，才存在从$F$ 到$G$ 的子图同构-因此名称
## 3.1. Graph Invariants
为了为不同的表达能力度量提供一个统一的框架，我们引人了节点和图不变量(Dimitrov et al.,2023)的概念。
Definition 4 (Node and Graph Invariant). Let P be a des-
ignated set, referred to as the palette .
$i) \textit{ A node invariant }_{( \cdot ) }is\textit{ a mapping that assigns to each}$
graph $G$ $\in$ $\mathcal{G}$ $a$ function $\zeta _G$ : $V( G)$ $\to$ $P$ , which
satisfies
$$\forall v\in V(G),\:\zeta_G(v)=\zeta_H(h(v)),$$
where H is any graph isomorphic to G and
$h\in$Ind$( G, H)$ .
2
|
[
[
0.08905029296875,
0.085205078125,
0.404296875,
0.085205078125,
0.404296875,
0.0985107421875,
0.08905029296875,
0.0985107421875
],
[
0.09234619140625,
0.10919189453125,
0.473876953125,
0.10919189453125,
0.473876953125,
0.12249755859375,
0.09234619140625,
0.12249755859375
],
[
0.7294921875,
0.10479736328125,
0.83837890625,
0.10479736328125,
0.83837890625,
0.11871337890625,
0.7294921875,
0.11871337890625
],
[
0.08905029296875,
0.125,
0.473876953125,
0.125,
0.473876953125,
0.1383056640625,
0.08905029296875,
0.1383056640625
],
[
0.09234619140625,
0.141357421875,
0.471435546875,
0.141357421875,
0.471435546875,
0.1514892578125,
0.09234619140625,
0.1514892578125
],
[
0.0906982421875,
0.1546630859375,
0.47216796875,
0.1546630859375,
0.47216796875,
0.1685791015625,
0.0906982421875,
0.1685791015625
],
[
0.74853515625,
0.1558837890625,
0.81640625,
0.1578369140625,
0.81640625,
0.173583984375,
0.74853515625,
0.1717529296875
],
[
0.08905029296875,
0.169189453125,
0.4697265625,
0.169189453125,
0.4697265625,
0.18310546875,
0.08905029296875,
0.18310546875
],
[
0.73779296875,
0.1717529296875,
0.82861328125,
0.173583984375,
0.82861328125,
0.1881103515625,
0.73779296875,
0.186279296875
],
[
0.0882568359375,
0.18310546875,
0.47216796875,
0.184326171875,
0.47216796875,
0.1988525390625,
0.0882568359375,
0.1976318359375
],
[
0.0906982421875,
0.199462890625,
0.471435546875,
0.199462890625,
0.471435546875,
0.21337890625,
0.0906982421875,
0.21337890625
],
[
0.08905029296875,
0.2139892578125,
0.453369140625,
0.2139892578125,
0.453369140625,
0.2279052734375,
0.08905029296875,
0.2279052734375
],
[
0.66650390625,
0.2108154296875,
0.67333984375,
0.2108154296875,
0.67333984375,
0.2139892578125,
0.66650390625,
0.2139892578125
],
[
0.6435546875,
0.2235107421875,
0.65185546875,
0.2235107421875,
0.65185546875,
0.2266845703125,
0.6435546875,
0.2266845703125
],
[
0.75244140625,
0.2183837890625,
0.814453125,
0.2183837890625,
0.814453125,
0.23291015625,
0.75244140625,
0.23291015625
],
[
0.0882568359375,
0.236083984375,
0.47216796875,
0.23486328125,
0.47216796875,
0.25,
0.0882568359375,
0.250732421875
],
[
0.72802734375,
0.2342529296875,
0.83837890625,
0.2342529296875,
0.83837890625,
0.2474365234375,
0.72802734375,
0.2474365234375
],
[
0.68896484375,
0.2442626953125,
0.701171875,
0.2442626953125,
0.701171875,
0.2474365234375,
0.68896484375,
0.2474365234375
],
[
0.0906982421875,
0.25244140625,
0.47216796875,
0.25244140625,
0.47216796875,
0.266357421875,
0.0906982421875,
0.266357421875
],
[
0.0906982421875,
0.268310546875,
0.47216796875,
0.268310546875,
0.47216796875,
0.2822265625,
0.0906982421875,
0.2822265625
],
[
0.0906982421875,
0.28271484375,
0.471435546875,
0.28271484375,
0.471435546875,
0.296630859375,
0.0906982421875,
0.296630859375
],
[
0.740234375,
0.287353515625,
0.82666015625,
0.287353515625,
0.82666015625,
0.3017578125,
0.740234375,
0.3017578125
],
[
0.09234619140625,
0.29736328125,
0.47216796875,
0.29736328125,
0.47216796875,
0.311279296875,
0.09234619140625,
0.311279296875
],
[
0.0906982421875,
0.313232421875,
0.473876953125,
0.313232421875,
0.473876953125,
0.326904296875,
0.0906982421875,
0.326904296875
],
[
0.0906982421875,
0.32763671875,
0.473876953125,
0.32763671875,
0.473876953125,
0.341552734375,
0.0906982421875,
0.341552734375
],
[
0.5009765625,
0.335205078125,
0.88232421875,
0.335205078125,
0.88232421875,
0.348388671875,
0.5009765625,
0.348388671875
],
[
0.08905029296875,
0.343505859375,
0.47216796875,
0.343505859375,
0.47216796875,
0.356689453125,
0.08905029296875,
0.356689453125
],
[
0.5009765625,
0.349853515625,
0.8837890625,
0.349853515625,
0.8837890625,
0.363037109375,
0.5009765625,
0.363037109375
],
[
0.0906982421875,
0.35791015625,
0.47216796875,
0.35791015625,
0.47216796875,
0.371826171875,
0.0906982421875,
0.371826171875
],
[
0.5009765625,
0.365478515625,
0.8837890625,
0.365478515625,
0.8837890625,
0.37890625,
0.5009765625,
0.37890625
],
[
0.0882568359375,
0.37255859375,
0.47216796875,
0.371826171875,
0.47216796875,
0.386474609375,
0.0882568359375,
0.386962890625
],
[
0.5,
0.37890625,
0.88818359375,
0.380126953125,
0.88818359375,
0.39453125,
0.5,
0.393310546875
],
[
0.0906982421875,
0.38818359375,
0.473876953125,
0.38818359375,
0.473876953125,
0.401611328125,
0.0906982421875,
0.401611328125
],
[
0.5009765625,
0.393310546875,
0.8857421875,
0.395751953125,
0.8857421875,
0.410400390625,
0.5009765625,
0.40771484375
],
[
0.08905029296875,
0.404052734375,
0.473876953125,
0.404052734375,
0.473876953125,
0.417236328125,
0.08905029296875,
0.417236328125
],
[
0.50244140625,
0.410400390625,
0.8857421875,
0.410400390625,
0.8857421875,
0.423583984375,
0.50244140625,
0.423583984375
],
[
0.0906982421875,
0.41845703125,
0.431396484375,
0.41845703125,
0.431396484375,
0.431884765625,
0.0906982421875,
0.431884765625
],
[
0.5009765625,
0.425537109375,
0.53125,
0.427978515625,
0.529296875,
0.439453125,
0.5,
0.4375
],
[
0.0906982421875,
0.44140625,
0.471435546875,
0.44140625,
0.471435546875,
0.455078125,
0.0906982421875,
0.455078125
],
[
0.0882568359375,
0.455078125,
0.47216796875,
0.455810546875,
0.47216796875,
0.4697265625,
0.0882568359375,
0.468505859375
],
[
0.50244140625,
0.46533203125,
0.8837890625,
0.46533203125,
0.8837890625,
0.479248046875,
0.50244140625,
0.479248046875
],
[
0.0906982421875,
0.4716796875,
0.47216796875,
0.4716796875,
0.47216796875,
0.48486328125,
0.0906982421875,
0.48486328125
],
[
0.5009765625,
0.48095703125,
0.72119140625,
0.48095703125,
0.72119140625,
0.494384765625,
0.5009765625,
0.494384765625
],
[
0.09234619140625,
0.488037109375,
0.156005859375,
0.488037109375,
0.156005859375,
0.5,
0.09234619140625,
0.5
],
[
0.5009765625,
0.5,
0.8857421875,
0.5,
0.8857421875,
0.51318359375,
0.5009765625,
0.51318359375
],
[
0.5009765625,
0.5146484375,
0.88623046875,
0.5146484375,
0.88623046875,
0.52783203125,
0.5009765625,
0.52783203125
],
[
0.0882568359375,
0.521484375,
0.2279052734375,
0.521484375,
0.2279052734375,
0.53515625,
0.0882568359375,
0.53515625
],
[
0.5009765625,
0.52978515625,
0.8857421875,
0.52978515625,
0.8857421875,
0.54345703125,
0.5009765625,
0.54345703125
],
[
0.0882568359375,
0.54541015625,
0.47216796875,
0.546875,
0.47216796875,
0.56103515625,
0.0882568359375,
0.56005859375
],
[
0.498291015625,
0.54443359375,
0.8857421875,
0.54345703125,
0.8857421875,
0.55810546875,
0.498291015625,
0.5595703125
],
[
0.2034912109375,
0.5625,
0.47216796875,
0.5625,
0.47216796875,
0.57568359375,
0.2034912109375,
0.57568359375
],
[
0.498291015625,
0.5595703125,
0.76953125,
0.56005859375,
0.76953125,
0.57470703125,
0.498291015625,
0.57373046875
],
[
0.1380615234375,
0.56494140625,
0.178955078125,
0.56494140625,
0.178955078125,
0.57275390625,
0.1380615234375,
0.57275390625
],
[
0.0906982421875,
0.56298828125,
0.134033203125,
0.56298828125,
0.134033203125,
0.57470703125,
0.0906982421875,
0.57470703125
],
[
0.08905029296875,
0.57373046875,
0.473876953125,
0.57373046875,
0.473876953125,
0.59130859375,
0.08905029296875,
0.59130859375
],
[
0.50244140625,
0.5859375,
0.8857421875,
0.5859375,
0.8857421875,
0.599609375,
0.50244140625,
0.599609375
],
[
0.0906982421875,
0.59228515625,
0.393798828125,
0.59228515625,
0.393798828125,
0.60595703125,
0.0906982421875,
0.60595703125
],
[
0.5009765625,
0.6015625,
0.8837890625,
0.6015625,
0.8837890625,
0.61572265625,
0.5009765625,
0.61572265625
],
[
0.1446533203125,
0.61865234375,
0.4150390625,
0.619140625,
0.4150390625,
0.6337890625,
0.1446533203125,
0.63330078125
],
[
0.50244140625,
0.6162109375,
0.8837890625,
0.6162109375,
0.8837890625,
0.6298828125,
0.50244140625,
0.6298828125
],
[
0.50244140625,
0.6318359375,
0.8857421875,
0.6318359375,
0.8857421875,
0.64599609375,
0.50244140625,
0.64599609375
],
[
0.08905029296875,
0.64794921875,
0.289306640625,
0.64794921875,
0.289306640625,
0.6611328125,
0.08905029296875,
0.6611328125
],
[
0.50244140625,
0.6484375,
0.5888671875,
0.6484375,
0.5888671875,
0.66015625,
0.50244140625,
0.66015625
],
[
0.0882568359375,
0.66748046875,
0.47216796875,
0.66650390625,
0.47216796875,
0.68115234375,
0.0882568359375,
0.681640625
],
[
0.5,
0.6767578125,
0.65673828125,
0.67822265625,
0.65673828125,
0.6923828125,
0.5,
0.69140625
],
[
0.0882568359375,
0.68310546875,
0.4755859375,
0.681640625,
0.4755859375,
0.6962890625,
0.0882568359375,
0.69775390625
],
[
0.08905029296875,
0.69873046875,
0.25732421875,
0.69873046875,
0.25732421875,
0.7119140625,
0.08905029296875,
0.7119140625
],
[
0.50244140625,
0.70068359375,
0.8857421875,
0.7021484375,
0.8857421875,
0.71630859375,
0.50244140625,
0.71533203125
],
[
0.50244140625,
0.71728515625,
0.8857421875,
0.71728515625,
0.8857421875,
0.73095703125,
0.50244140625,
0.73095703125
],
[
0.136474609375,
0.72607421875,
0.4248046875,
0.72607421875,
0.4248046875,
0.74072265625,
0.136474609375,
0.74072265625
],
[
0.5009765625,
0.73291015625,
0.52685546875,
0.73291015625,
0.52685546875,
0.74560546875,
0.5009765625,
0.74560546875
],
[
0.0882568359375,
0.75390625,
0.393798828125,
0.75390625,
0.393798828125,
0.767578125,
0.0882568359375,
0.767578125
],
[
0.386474609375,
0.7548828125,
0.473876953125,
0.7548828125,
0.473876953125,
0.76513671875,
0.386474609375,
0.76513671875
],
[
0.5009765625,
0.75048828125,
0.88623046875,
0.75048828125,
0.88623046875,
0.7646484375,
0.5009765625,
0.7646484375
],
[
0.08905029296875,
0.76953125,
0.473876953125,
0.76953125,
0.473876953125,
0.78271484375,
0.08905029296875,
0.78271484375
],
[
0.50244140625,
0.76806640625,
0.7490234375,
0.76806640625,
0.7490234375,
0.77880859375,
0.50244140625,
0.77880859375
],
[
0.08905029296875,
0.7841796875,
0.285888671875,
0.7841796875,
0.285888671875,
0.79736328125,
0.08905029296875,
0.79736328125
],
[
0.5146484375,
0.79736328125,
0.8857421875,
0.79736328125,
0.8857421875,
0.81103515625,
0.5146484375,
0.81103515625
],
[
0.09234619140625,
0.8125,
0.471435546875,
0.8125,
0.471435546875,
0.82568359375,
0.09234619140625,
0.82568359375
],
[
0.53173828125,
0.8125,
0.8857421875,
0.8125,
0.8857421875,
0.8271484375,
0.53173828125,
0.8271484375
],
[
0.08905029296875,
0.82470703125,
0.281982421875,
0.8271484375,
0.281982421875,
0.84130859375,
0.08905029296875,
0.8388671875
],
[
0.53369140625,
0.8291015625,
0.58837890625,
0.8291015625,
0.58837890625,
0.84033203125,
0.53369140625,
0.84033203125
],
[
0.0882568359375,
0.8447265625,
0.47216796875,
0.84521484375,
0.47216796875,
0.85986328125,
0.0882568359375,
0.859375
],
[
0.59716796875,
0.849609375,
0.81787109375,
0.85107421875,
0.81787109375,
0.86865234375,
0.59716796875,
0.8671875
],
[
0.08905029296875,
0.8623046875,
0.473876953125,
0.8623046875,
0.473876953125,
0.87548828125,
0.08905029296875,
0.87548828125
],
[
0.08905029296875,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.08905029296875,
0.89013671875
],
[
0.53369140625,
0.876953125,
0.88623046875,
0.876953125,
0.88623046875,
0.8916015625,
0.53369140625,
0.8916015625
],
[
0.0882568359375,
0.892578125,
0.38232421875,
0.892578125,
0.38232421875,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.53369140625,
0.892578125,
0.64453125,
0.892578125,
0.64453125,
0.90576171875,
0.53369140625,
0.90576171875
],
[
0.4794921875,
0.92236328125,
0.495849609375,
0.92236328125,
0.495849609375,
0.9375,
0.4794921875,
0.9375
]
] |
[
"利用位置编码来增强标准MPNN的表达能力",
"子图 GNNs (Bevilacqua et al., 2021;You et al., 2021;",
"Homomorphism",
"Frascaetal.,2022;Huangetal.,2022)通过将标准MPNN",
"分别应用于从初始输入图中选择的一袋子图来计算图",
"表示。已知子图GNN比1-WL更具表现力，但上限为",
"Subgraph",
"3-WL (Frasca et al., 2022)。B.Zhang et al.(2022)证明了",
"Isomorphism",
"ESAN（Bevilacquaetal.，2021）能够计算图的双连接性",
"这是一种相对容易计算的图统计量。此外，Micheletal.",
"(2023)探索处理图中路径以增强其表达能力的GNN",
"",
"",
"Bijective",
"Lovasz（1967）表明同态计数（请参阅Definition7）可",
"Homomorphism",
"",
"作为完整的图不变式，这意味着当且仅当它们的同态",
"计数相同时，两个图才是同构的。在此完整性属性的基",
"础上，Nguyen et al.(2020)and Welke et al.(2023)利用",
"Isomorphism",
"同态计数来开发可表达的GNN。Tinhofer（1986，1991）",
"的理论贡献建立了1-WL与树宽为1的图的同态计数",
"能力之间的等价性。将这些结果扩展到k-WL，Dell",
"Figure l: Examples of non-injective homomorphism (row",
"etal.（2018）证明了k-WL与从具有树宽度k的图中",
"1), subgraph isomorphism (row 2),bijective homomor",
"计数同态的能力之间的等效性。最近，B.Zhangetal.",
"phism with non-homomorphic inverse (row 3), and iso",
"(2024）引人了一个定量框架，该框架完全表征了GNN",
"morphism (row 4).For better clarity,the mappings h :",
"的表示能力。扩展到k-WL层次结构之外，他们主张",
"V(F) → V(G) are visually represented with colors, where",
"将同态计数作为表达率的完整度量，因为GNN体系结",
"F is consistently on the left, and G is on the right in each",
"构已被证明可以对特定的基序家族进行同态计数",
"row.",
"在我们的方法中，我们首先通过将我们的体系结构与",
"k-WL层次结构相关联，从更常规的角度呈现结果。这",
"子图同构h将不同的节点映射到不同的节点因此，它",
"为后续探索子结构奠定了基础，我们可以有效地进行",
"还将不同的边映射到不同的边。",
"同态计数",
"Definition 3 (Isomorphism). An isomorphism is a bijective",
"homomorphism whose inverse is also a homomorphism. If",
"3. Preliminaries",
"F is isomorphic to G, we write F = G: The set of iso-",
"让G是一个简单且无向的图。我们用",
"morphisms from F to G is denoted by Ind(F,G), and its",
"[1,2,....,N]表示节点集，用",
"cardinality by ind(F, G) = [Ind(F,G)] .",
"",
"V(G)",
"E(G)C (eleE2V(G),lel=2)表示边缘集。对",
"同构是节点之间的一对一对应关系。对逆是同态的要",
"于给定的节点EV(G)，直接邻域定义为",
"求也保证了边之间的一对一对应。我们参考Figure1",
"N(u) := (ue V(G) / (v, u) EE(G)) ",
"来说明这些概念之间的区别。值得注意的是，当且仅",
"当F与G的子图同构时，才存在从F到G的子图同",
"我们用9表示所有图的集合",
"构-因此名称",
"Definition 1 (Homomorphism). Let F and G be two",
"3.1. Graph Invariants",
"graphs.Ahomomorphism fromFto G is a map h:",
"V(F) → V(G) such that",
"为了为不同的表达能力度量提供一个统一的框架，我",
"们引人了节点和图不变量（Dimitrovetal.，2023）的概",
"[u,u) EE(F) [h(u),h(u)) E E(G)",
"念。",
"The set of homomorphisms fromF to",
"Gisde-",
"Definition 4 (Node and Graph Invariant). Let P be a des-",
"noted byHom(F,G) ，, and its cardinalityby",
"ignated set, referred to as the palette.",
"hom(F,G) := |Hom(F,G)l :",
"i) A node invariant S() is a mapping that assigns to each",
"直观上，同态是边缘保留映射，i。e。，它将F的相邻",
"graph G E G a function SG : V(G) → P, which",
"顶点映射到G的相邻顶点。",
"satisfies",
"Definition 2 (Subgraph Isomorphism). A subgraph isomor",
"Vu E V(G), SG(u) = SH(h(u),",
"phism is an injective homomorphism. The set of subgraph",
"isomorphisms from F to G is denoted by Sub(F,G) , and",
"where H is any graph isomorphic to G and",
"its cardinality by sub(F, G) = Sub(F, G)/ .",
"h E Ind(G, H) .",
"2"
] |
{
"arxiv_id": "2403.13749",
"language": "zh",
"page_idx": 1
}
|
||
2403.13749-zh-2
|
2403.13749-zh-2.jpg
|
其中 HASH 是内射函数。当实现稳定的着色时，算法终止，这意味着颜色的数量在一次迭代中不会改变。$t$ 迭代后的图输出由
$$c^{(t)}(G)=\text{HASH}\left(\left\{\left\{c^{(t)}(v)\mid v\in V(G)\right\}\right\}\right),$$
给出，稳定着色由$c(G)$表示。特别是，$c^(t)$和$c$的颜色是节点(分别为。graph)不变量，并且，由于符号的轻微滥用，被称为1-WL
Xu et al. (2019)已经证明，如果对于每个$t\in\{1,\ldots,T\}$ ,函数 AGGR$^(t)$,UPD$^(t)$和 READ 是各自域上的内射函数，则通过
$$\begin{aligned}m^{(t+1)}(v)&=\mathrm{AGGR}^{(t+1)}\left(\left\{\left\{h^{(t)}(u)\mid u\in\mathcal{N}(v)\right\}\right\}\right),\\h^{(t+1)}(v)&=\mathrm{UPD}^{(t+1)}\left(h^{(t)}(u),m^{(t+1)}(u)\right),\\h(G)&=\mathrm{READ}\left(\left\{\left\{h^{(T)}(v)\mid v\in V(G)\right\}\right\}\right),\end{aligned}$$
更新节点功能的 MPNN 与 1 -WL 一样强大
$ii) \textit{ A graph invariant is a function }\zeta : \mathcal{G} \to P$ such that
$\zeta ( G) = \zeta ( H) \textit{for all isomorphic pairs of graphs G}, H$
请注意，每个节点不变量通过收集图中每个节点的节点不变量的多集来诱导图不变量。例如，节点的度是节点不变量，度序列是图不变量
直观上，图不变量在同构变换下不会改变，i。e。,它不依赖于节点的顺序，而是依赖于图的结构。除了已经提到的度序列之外，图不变量的示例还包括连接组件的数量，图的直径等等。不是图不变量的函数的一个例子是边交点的数量，因为它取决于图在平面上的绘制方式 (参见，e。g。,Figure 1 的最后一行)
Definition 5 (Distinguishable Pair). Given two graphs $G,H$ and a graph invariant $\zeta:\mathcal{G}\to P$ , we say that $\zeta$ $distinguishes~G~and~H~if~\zeta(G)\neq\zeta(H)~.$
虽然很明显，如果$\zeta$区分$G$ 和$H$,那么它们就不是同构的，反之亦然：可能存在$\zeta$无法区分的非同构图对。例如，有一些具有相同程度序列的图不是同构的(参见，e。g。,Figure 2)。如果 $\zeta$可以区分所有的非同构图，则称$\zeta$为完全图不变量。完整的图不变量很重要，因为它们具有最大的表达能力。但是，已知没有完整的图不变量可以在多项式时间内解决，这为长期研究主要数学问题之一-图同构问题 (Grohe et al., 2020) 奠定了基础。
为了比较不同图不变量的表达能力，我们引人以下定
义。
Definition 6 (Power of Graph Invariants). Let $\gamma,\zeta$ be two $graphinvariants.Wesaythat\gamma ismorepowerfulthan\zeta(in)$ symbols $\gamma\subseteq\zeta$ )iffor every pair of graphs $H,G$
$$\gamma(H)=\gamma(G)\implies\zeta(H)=\zeta(G).$$
We say that $\gamma$ is strictly more powerful than $\zeta$ if $\gamma\subseteq\zeta$ and there exists a pair of graphs F,G such that $\gamma(H)\neq\gamma(G)$ and $\zeta ( H) = \zeta ( G)$ .
松散地说，如果$\gamma$可以区分$\zeta$区分的所有对，则它比$\zeta$ 更强大，如果它可以区分比$\zeta$ 更多的对，则它严格地更强大。Definition 6 允许我们测量某些算法的表达能力，特别是图形着色和 GNN 架构
# 3.2. Weisfeiler-Leman and Graph Isomorphism Networks
Weisfeiler-Leman (WL) 算法是一种完善的图不变量(Weisfeiler et al., 1968) , 具有线性复杂度。尽管效率很高，但 1-WL 无法区分成对的图形的简单而常见的示例。该算法通过以下方案迭代更新每个节点$v\in V(G)$ 的颜色：
$$c^{(t+1)}(v)\leftarrow\mathrm{HASH}\left(c^{(t)}(v),\left\{\left\{c^{(t)}(u)\mid u\in\mathcal{N}(v)\right\}\right\}\right),\\(1)$$
3.3. Higher-Order Weisfeiler-Leman Tests
可以通过考虑高阶相互作用来提升 WL 的表达能力。WL 最简单的高阶变体是$k$维 Weisfeiler-Leman 检验用$k$ -WL 表示。给定一个具有节点$V(G)$和边$E(G)$的图$G$ ,该算法生成一个新的图 $H$ ,其中每个节点都是$V(G)$
$$V(H)=\left\{\boldsymbol{v}=\left\{v_i\right\}_{i=1}^k\mid v_i\in V(G)\right\}=V(G)^k,$$
和边$E(H)$的元素的$k$元组，其中$k$中的 XMATHX。该算法根据诱导子图$G[v]$的同构类型为每个节点$v\in$ $V(H)$分配初始颜色。颜色细化方案正是应用于$H$的(1)。虽然$H$可以通过简单的算法生成，但随着$k$中节点和边的数量呈指数增长，该方法很快变得不切实际
# 3.4. Homomorphism and Subgraph Counting $\mathbf{Expressivity}$
考虑到主题$F$的出现，可以构建更细微的图形不变性。Definition 7 (Homomorphism and Subgraph Counts). Given a graph F , a graph invariant ζ can homomorphism- $count~F~iffor~all~pairs~of~graphs~G,\stackrel{\mathrm{o}}{H}$
$$\zeta(G)=\zeta(H)\implies\hom(F,G)=\hom(F,H).$$
$By$ analogy, $\zeta$ can subgraph-count F if for all pairs of
graphs $G,H$
$$\zeta(G)=\zeta(H)\implies\sinh(F,G)=\sinh(F,H).$$
如果$\mathcal{F}$是图族，我们说$\zeta$可以同态- (resp。子图-)如果$\zeta$可以同态-(resp。子图-)每$F\in\mathcal{F}$计数。我们用
$$\hom(\mathcal{F},G)=\left(\hom(F,G)\right)_{F\in\mathcal{F}},$$
3
|
[
[
0.098876953125,
0.08587646484375,
0.47216796875,
0.08587646484375,
0.47216796875,
0.0997314453125,
0.098876953125,
0.0997314453125
],
[
0.5009765625,
0.08709716796875,
0.8857421875,
0.08709716796875,
0.8857421875,
0.10101318359375,
0.5009765625,
0.10101318359375
],
[
0.1217041015625,
0.1029052734375,
0.473876953125,
0.1029052734375,
0.473876953125,
0.11614990234375,
0.1217041015625,
0.11614990234375
],
[
0.5009765625,
0.10162353515625,
0.88623046875,
0.10162353515625,
0.88623046875,
0.11553955078125,
0.5009765625,
0.11553955078125
],
[
0.5009765625,
0.117431640625,
0.6298828125,
0.117431640625,
0.6298828125,
0.1307373046875,
0.5009765625,
0.1307373046875
],
[
0.53173828125,
0.14013671875,
0.8447265625,
0.14013671875,
0.8447265625,
0.1610107421875,
0.53173828125,
0.1610107421875
],
[
0.09234619140625,
0.149658203125,
0.471435546875,
0.149658203125,
0.471435546875,
0.160400390625,
0.09234619140625,
0.160400390625
],
[
0.09234619140625,
0.1634521484375,
0.473876953125,
0.1634521484375,
0.473876953125,
0.1767578125,
0.09234619140625,
0.1767578125
],
[
0.5009765625,
0.17236328125,
0.8837890625,
0.1717529296875,
0.8837890625,
0.186279296875,
0.5009765625,
0.1875
],
[
0.09234619140625,
0.179931640625,
0.31787109375,
0.179931640625,
0.31787109375,
0.1900634765625,
0.09234619140625,
0.1900634765625
],
[
0.5009765625,
0.1875,
0.8837890625,
0.1881103515625,
0.8837890625,
0.20263671875,
0.5009765625,
0.2020263671875
],
[
0.0906982421875,
0.2020263671875,
0.471435546875,
0.2020263671875,
0.471435546875,
0.21533203125,
0.0906982421875,
0.21533203125
],
[
0.5,
0.203857421875,
0.68115234375,
0.203857421875,
0.68115234375,
0.2171630859375,
0.5,
0.2171630859375
],
[
0.0906982421875,
0.216552734375,
0.4697265625,
0.216552734375,
0.4697265625,
0.2298583984375,
0.0906982421875,
0.2298583984375
],
[
0.0906982421875,
0.2310791015625,
0.473876953125,
0.2310791015625,
0.473876953125,
0.2442626953125,
0.0906982421875,
0.2442626953125
],
[
0.5,
0.2259521484375,
0.8837890625,
0.2259521484375,
0.8837890625,
0.240478515625,
0.5,
0.240478515625
],
[
0.08905029296875,
0.246826171875,
0.464111328125,
0.246826171875,
0.464111328125,
0.260009765625,
0.08905029296875,
0.260009765625
],
[
0.5146484375,
0.242431640625,
0.8837890625,
0.242431640625,
0.8837890625,
0.2568359375,
0.5146484375,
0.2568359375
],
[
0.4599609375,
0.250732421875,
0.468017578125,
0.250732421875,
0.468017578125,
0.255126953125,
0.4599609375,
0.255126953125
],
[
0.5087890625,
0.248779296875,
0.515625,
0.248779296875,
0.515625,
0.25244140625,
0.5087890625,
0.25244140625
],
[
0.0906982421875,
0.261962890625,
0.47216796875,
0.261962890625,
0.47216796875,
0.27587890625,
0.0906982421875,
0.27587890625
],
[
0.50244140625,
0.25830078125,
0.6005859375,
0.25830078125,
0.6005859375,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.08905029296875,
0.27783203125,
0.416748046875,
0.27783203125,
0.416748046875,
0.291748046875,
0.08905029296875,
0.291748046875
],
[
0.50244140625,
0.279052734375,
0.87109375,
0.2822265625,
0.87109375,
0.3037109375,
0.50244140625,
0.300537109375
],
[
0.08905029296875,
0.294189453125,
0.4697265625,
0.294189453125,
0.4697265625,
0.309326171875,
0.08905029296875,
0.309326171875
],
[
0.0882568359375,
0.31005859375,
0.473876953125,
0.311279296875,
0.473876953125,
0.32568359375,
0.0882568359375,
0.324462890625
],
[
0.50732421875,
0.30615234375,
0.81787109375,
0.309326171875,
0.81787109375,
0.330078125,
0.50732421875,
0.326904296875
],
[
0.08660888671875,
0.324462890625,
0.358642578125,
0.32568359375,
0.358642578125,
0.34033203125,
0.08660888671875,
0.339111328125
],
[
0.5361328125,
0.335205078125,
0.83837890625,
0.3359375,
0.83837890625,
0.35791015625,
0.5361328125,
0.356689453125
],
[
0.09234619140625,
0.354736328125,
0.471435546875,
0.354736328125,
0.471435546875,
0.365478515625,
0.09234619140625,
0.365478515625
],
[
0.09234619140625,
0.36865234375,
0.47216796875,
0.36865234375,
0.47216796875,
0.3818359375,
0.09234619140625,
0.3818359375
],
[
0.5009765625,
0.36865234375,
0.8037109375,
0.36865234375,
0.8037109375,
0.3818359375,
0.5009765625,
0.3818359375
],
[
0.0906982421875,
0.3837890625,
0.47216796875,
0.3837890625,
0.47216796875,
0.397705078125,
0.0906982421875,
0.397705078125
],
[
0.08905029296875,
0.3984375,
0.471435546875,
0.3984375,
0.471435546875,
0.412353515625,
0.08905029296875,
0.412353515625
],
[
0.5009765625,
0.40087890625,
0.79833984375,
0.40087890625,
0.79833984375,
0.4140625,
0.5009765625,
0.4140625
],
[
0.0906982421875,
0.41357421875,
0.47216796875,
0.41357421875,
0.47216796875,
0.4267578125,
0.0906982421875,
0.4267578125
],
[
0.09234619140625,
0.429931640625,
0.471435546875,
0.429931640625,
0.471435546875,
0.43994140625,
0.09234619140625,
0.43994140625
],
[
0.5009765625,
0.4248046875,
0.88134765625,
0.4248046875,
0.88134765625,
0.438232421875,
0.5009765625,
0.438232421875
],
[
0.0906982421875,
0.443115234375,
0.47216796875,
0.443115234375,
0.47216796875,
0.45703125,
0.0906982421875,
0.45703125
],
[
0.5,
0.4375,
0.8837890625,
0.438232421875,
0.8837890625,
0.45263671875,
0.5,
0.451904296875
],
[
0.5009765625,
0.45263671875,
0.8837890625,
0.453857421875,
0.8837890625,
0.468505859375,
0.5009765625,
0.46728515625
],
[
0.0906982421875,
0.458984375,
0.47216796875,
0.458984375,
0.47216796875,
0.472900390625,
0.0906982421875,
0.472900390625
],
[
0.5009765625,
0.468505859375,
0.8837890625,
0.4697265625,
0.8837890625,
0.484130859375,
0.5009765625,
0.48291015625
],
[
0.09234619140625,
0.47607421875,
0.1805419921875,
0.47607421875,
0.1805419921875,
0.4873046875,
0.09234619140625,
0.4873046875
],
[
0.50341796875,
0.486083984375,
0.54150390625,
0.486083984375,
0.54150390625,
0.5,
0.50341796875,
0.5
],
[
0.0906982421875,
0.496826171875,
0.47216796875,
0.496826171875,
0.47216796875,
0.51025390625,
0.0906982421875,
0.51025390625
],
[
0.0906982421875,
0.51220703125,
0.11358642578125,
0.51220703125,
0.11358642578125,
0.5244140625,
0.0906982421875,
0.5244140625
],
[
0.52783203125,
0.51025390625,
0.85546875,
0.51025390625,
0.85546875,
0.52783203125,
0.52783203125,
0.52783203125
],
[
0.0882568359375,
0.52978515625,
0.4697265625,
0.53076171875,
0.4697265625,
0.54541015625,
0.0882568359375,
0.54443359375
],
[
0.08905029296875,
0.54541015625,
0.473876953125,
0.54541015625,
0.473876953125,
0.56005859375,
0.08905029296875,
0.56005859375
],
[
0.50341796875,
0.54052734375,
0.88232421875,
0.54052734375,
0.88232421875,
0.5537109375,
0.50341796875,
0.5537109375
],
[
0.50244140625,
0.5537109375,
0.8837890625,
0.5546875,
0.8837890625,
0.5693359375,
0.50244140625,
0.568359375
],
[
0.0882568359375,
0.56103515625,
0.406982421875,
0.56005859375,
0.406982421875,
0.57470703125,
0.0882568359375,
0.57568359375
],
[
0.50244140625,
0.57080078125,
0.8837890625,
0.57080078125,
0.8837890625,
0.583984375,
0.50244140625,
0.583984375
],
[
0.1641845703125,
0.5869140625,
0.39794921875,
0.5869140625,
0.39794921875,
0.60107421875,
0.1641845703125,
0.60107421875
],
[
0.50244140625,
0.5869140625,
0.88134765625,
0.5869140625,
0.88134765625,
0.59765625,
0.50244140625,
0.59765625
],
[
0.50244140625,
0.60107421875,
0.8759765625,
0.60107421875,
0.8759765625,
0.6142578125,
0.50244140625,
0.6142578125
],
[
0.08905029296875,
0.6123046875,
0.47216796875,
0.6123046875,
0.47216796875,
0.626953125,
0.08905029296875,
0.626953125
],
[
0.0882568359375,
0.62744140625,
0.473876953125,
0.626953125,
0.473876953125,
0.6416015625,
0.0882568359375,
0.642578125
],
[
0.498291015625,
0.630859375,
0.82177734375,
0.63330078125,
0.82177734375,
0.64794921875,
0.498291015625,
0.64599609375
],
[
0.08905029296875,
0.64306640625,
0.2181396484375,
0.64306640625,
0.2181396484375,
0.6572265625,
0.08905029296875,
0.6572265625
],
[
0.53125,
0.6484375,
0.615234375,
0.6484375,
0.615234375,
0.66015625,
0.53125,
0.66015625
],
[
0.0906982421875,
0.671875,
0.47216796875,
0.671875,
0.47216796875,
0.681640625,
0.0906982421875,
0.681640625
],
[
0.5009765625,
0.671875,
0.8837890625,
0.671875,
0.8837890625,
0.68505859375,
0.5009765625,
0.68505859375
],
[
0.08905029296875,
0.68505859375,
0.47216796875,
0.68505859375,
0.47216796875,
0.69873046875,
0.08905029296875,
0.69873046875
],
[
0.08905029296875,
0.69873046875,
0.473876953125,
0.69970703125,
0.473876953125,
0.7138671875,
0.08905029296875,
0.71337890625
],
[
0.5,
0.70263671875,
0.8857421875,
0.7041015625,
0.8857421875,
0.71826171875,
0.5,
0.71728515625
],
[
0.0906982421875,
0.71533203125,
0.330078125,
0.71533203125,
0.330078125,
0.72900390625,
0.0906982421875,
0.72900390625
],
[
0.5009765625,
0.7197265625,
0.88232421875,
0.7197265625,
0.88232421875,
0.73291015625,
0.5009765625,
0.73291015625
],
[
0.5,
0.734375,
0.75634765625,
0.73291015625,
0.75634765625,
0.74755859375,
0.5,
0.74853515625
],
[
0.0882568359375,
0.74560546875,
0.42236328125,
0.7470703125,
0.42236328125,
0.76123046875,
0.0882568359375,
0.76025390625
],
[
0.1168212890625,
0.76220703125,
0.188720703125,
0.76220703125,
0.188720703125,
0.7763671875,
0.1168212890625,
0.7763671875
],
[
0.5361328125,
0.7587890625,
0.8486328125,
0.7587890625,
0.8486328125,
0.7734375,
0.5361328125,
0.7734375
],
[
0.0882568359375,
0.78515625,
0.473876953125,
0.78515625,
0.473876953125,
0.798828125,
0.0882568359375,
0.798828125
],
[
0.5,
0.78271484375,
0.8857421875,
0.78271484375,
0.8857421875,
0.79736328125,
0.5,
0.79736328125
],
[
0.0906982421875,
0.80322265625,
0.47216796875,
0.80322265625,
0.47216796875,
0.81298828125,
0.0906982421875,
0.81298828125
],
[
0.5,
0.798828125,
0.5849609375,
0.79541015625,
0.58642578125,
0.81005859375,
0.5009765625,
0.81298828125
],
[
0.0906982421875,
0.81640625,
0.47216796875,
0.81640625,
0.47216796875,
0.830078125,
0.0906982421875,
0.830078125
],
[
0.54345703125,
0.8212890625,
0.84228515625,
0.8212890625,
0.84228515625,
0.8359375,
0.54345703125,
0.8359375
],
[
0.08905029296875,
0.830078125,
0.47216796875,
0.83056640625,
0.47216796875,
0.84521484375,
0.08905029296875,
0.8447265625
],
[
0.0906982421875,
0.84765625,
0.1478271484375,
0.84765625,
0.1478271484375,
0.859375,
0.0906982421875,
0.859375
],
[
0.5009765625,
0.85400390625,
0.8837890625,
0.85400390625,
0.8837890625,
0.8671875,
0.5009765625,
0.8671875
],
[
0.08905029296875,
0.87060546875,
0.46728515625,
0.87060546875,
0.46728515625,
0.89599609375,
0.08905029296875,
0.89599609375
],
[
0.5,
0.86865234375,
0.869140625,
0.8671875,
0.869140625,
0.8818359375,
0.5,
0.88330078125
],
[
0.449462890625,
0.89306640625,
0.47216796875,
0.89306640625,
0.47216796875,
0.9072265625,
0.449462890625,
0.9072265625
],
[
0.57666015625,
0.8896484375,
0.80615234375,
0.892578125,
0.80615234375,
0.91015625,
0.57666015625,
0.9072265625
],
[
0.4794921875,
0.92236328125,
0.495849609375,
0.92236328125,
0.495849609375,
0.93798828125,
0.4794921875,
0.93798828125
]
] |
[
"ii) A graph invariant is a function S : g -→ P such that",
"其中HASH是内射函数。当实现稳定的着色时，算法",
"C(G) = S(H)for all isomorphic pairs of graphs G, H",
"终止，这意味着颜色的数量在一次送代中不会改变。t",
"迭代后的图输出由",
"c(t)(G) = HASH ([c()(u) I v E V(G))))",
"请注意，每个节点不变量通过收集图中每个节点的节",
"点不变量的多集来诱导图不变量。例如，节点的度是",
"给出，稳定着色由c(G)表示。特别是，c(t)和c的颜",
"节点不变量，度序列是图不变量",
"色是节点（分别为。graph）不变量，并且，由于符号的",
"直观上，图不变量在同构变换下不会改变，i。e。，它",
"轻微滥用，被称为1-WL",
"不依赖于节点的顺序，而是依赖于图的结构。除了已",
"经提到的度序列之外，图不变量的示例还包括连接组",
"Xuetal.(2019)已经证明，如果对于每个tE{1..,T}",
"件的数量，图的直径等等。不是图不变量的函数的",
"函数AGGR(t),UPD(t）和READ是各自域上的内射",
"-",
"",
"个例子是边交点的数量，因为它取决于图在平面上的",
"函数，则通过",
"绘制方式（参见，e。g。，Figure1的最后一行)",
"m(++1)(v) = AGGR(++1) ([[h()(u) [ uEN(o)))",
"Definition 5 (Distinguishable Pair). Given two graph",
"G, H and a graph invariant : g -→ P, we say that (",
"h(t++1)() = UPD(t+1) (h(t)(u), m(t+1)(u)),",
"distinguishes G and H if c(G) ≠ c(H).",
"h(G) = READ ((h(T)(0) / u E V(G))",
"虽然很明显，如果区分G和H，那么它们就不是",
"同构的，反之亦然：可能存在‘无法区分的非同构图",
"更新节点功能的MPNN与1-WL一样强大",
"对。例如，有一些具有相同程度序列的图不是同构的",
"（参见，e。g。，Figure2）。如果（可以区分所有的非同",
"3.3. Higher-Order Weisfeiler-Leman Tests",
"构图，则称为完全图不变量。完整的图不变量很重",
"要，因为它们具有最大的表达能力。但是，已知没有",
"可以通过考虑高阶相互作用来提升WL的表达能力",
"完整的图不变量可以在多项式时间内解决，这为长期",
"WL最简单的高阶变体是k维Weisfeiler-Leman检验",
"用k-WL表示。给定一个具有节点V(G)和边E(G)的",
"研究主要数学问题之一-图同构问题（Groheetal.,2020)",
"图G，该算法生成一个新的图H，其中每个节点都是",
"奠定了基础。",
"V(G)",
"为了比较不同图不变量的表达能力，我们引入以下定",
"义。",
"V(H) =U= [u)h-1 I UiE V(G)(= V(G)h",
"Definition 6 (Power of Graph Invariants). Let , be twc",
"graph invariants. We say that is more powerful than C (in",
"和边E(H）的元素的k元组，其中k中的XMATHX",
"该算法根据诱导子图G[u]的同构类型为每个节点E",
"symbols S) if for every pair of graphs H,G",
"V(H）分配初始颜色。颜色细化方案正是应用于H的",
"(H) =(G) → ((H) = ((G).",
"（1）。虽然H可以通过简单的算法生成，但随着k中节",
"点和边的数量呈指数增长，该方法很快变得不切实际",
"We say that is strictly more powerful than if and",
"there exists a pair of graphs F, G such that (H) + (G)",
"3.4. Homomorphism and Subgraph Counting",
"and C(H) = C(G).",
"Expressivity",
"松散地说，如果可以区分（区分的所有对，则它比",
"考虑到主题F的出现，可以构建更细微的图形不变性。",
"S更强大，如果它可以区分比（更多的对，则它严格地",
"更强大。Definition6允许我们测量某些算法的表达能",
"Definition 7 (Homomorphism and Subgraph Counts).",
"力，特别是图形着色和GNN架构",
"Given a graph F, a graph invariant can homomorphism",
"count F if for all pairs of graphs G, H",
"3.2. Weisfeiler-Leman and Graph Isomorphism",
"Networks",
"((G) =((H) → hom(F,G) =hom(F,H).",
"Weisfeiler-Leman（WL）算法是一种完善的图不变量",
"By analogy, can subgraph-count F if for all pairs of",
"(Weisfeileretal.，1968），具有线性复杂度。尽管效率很",
"graphs G, H",
"高，但1-WL无法区分成对的图形的简单而常见的示",
"((G) =(H) → sub(F,G) = sub(F,H).",
"例。该算法通过以下方案迭代更新每个节点EV(G",
"的颜色：",
"如果F是图族，我们说可以同态-（resp。子图-）如",
"et+1)() - HASH (c((0), [[c((u) uE N()))",
"果（可以同态-(resp。子图-)每FEF计数。我们用",
"(1)",
"hom(F,G) = (hom(F,G))FeF ;",
"3"
] |
{
"arxiv_id": "2403.13749",
"language": "zh",
"page_idx": 2
}
|
||
2403.13749-zh-4
|
2403.13749-zh-4.jpg
|
<FigureHere>
Figure 2: Example of two non-isomorphic graphs that are $r$ - $\ell$WL equivalent but not $(r+1)$ - $\ell$WL equivalent: a chordal
cycle (left) and a cactus graph (right).
$\textbf{Definition 12( r- MPNN) . For }t\in \{ 0, \ldots , T- 1\}$ and $k\in \{ 0, \ldots , r\}$ , $r$ $- \ell MPNN$ applies the following $message$, update and readout functions
Defnition 11 (Cactus Graph). A cactus graph is a graph where every edge lies on at most one simple cycle. For $r\geq2,an~r~-cactus~graph~is~a~cactus~where~every~simple$ cycle has at most r vertices. We denote by M the set of all cactus graphs, and by M' the set comprising all q -cactus graphs for $q\leq r$ .
$$\begin{aligned}&m_{k}^{(t+1)}(v)=\mathrm{AGGR}_{k}^{(t+1)}\left(\left\{\left\{h_{k}^{(t)}(\boldsymbol{p})\mid\boldsymbol{p}\in\mathcal{N}_{k}(v)\right\}\right\}\right)\\&h_{r}^{(t+1)}(v)=\mathrm{UPD}^{(t+1)}\left(h_{r}^{(t)}(v),m_{0}^{(t+1)}(v)\right.\\&h_{r}^{(T)}(G)=\mathrm{READ}\left(\left\{\left\{h_{r}^{(T)}(v)\mid v\in V(G)\right\}\right\}\right).\end{aligned}$$
我们参考 Figure 2 来说明两个小仙人掌图。我们现在
准备介绍我们的主要结果。1
Theorem 2 (Homomorphism Counting Power of $r$ -$\ell$WL). Let $r\geq 0$ . $Then$, $r$ - $\ell$ $WL$ can $homomorphism- count$ $\mathcal{M} ^{r+ 2}$
我们参考 Appendix D 来获得 Theorem 2 的详细证明该证明相当复杂，需要定义仙人掌图的规范树分解和$r-\ell$WL 的展开树。演示它们的强连接，然后我们按照(Dell et al., 2018; B. Zhang et al., 2024) 中的方法分解仙人掌图的同态计数
类$\mathcal{M}^2$仅包含林；因此，Theorem 2 暗示了 1-WL 计数林能力的标准结果。由于森林是 1-WL 可以计数的唯一一类图，因此 Theorem 2 暗示$r-\ell$WL 始终比 1-WL 严格地更强大，从而证实了 Proposition 1 中的主张
Theorem 2 的含义是深刻的：它建立了$r$ - $\ell$WL 可以同态计数一大类图。此结果表明，即使是最简单的变体 1 - $\ell$WL ,其功能也不亚于子图 GNN, 后者仅限于端点共享 NED (B. Zhang et al., 2024) 的同态计数图。重要的是，Theorem 2 陈述了$r$ - $\ell$WL 的同态表达率的宽松下界。这一观察结果为未来研究探索$r-\ell$WL 的同态表达率的严格下界或上界开辟了道路。
函数$h_r^{(t)}:V(G)\to C$映射$v\mapsto h_r^{(t)}(v)$是一个节点Invarant。在以下结果中，我们建立了(2)中定义的$h_r^{(t)}$ 和 $c_r^{(t)}$之间的关系
Theorem 3 (GNN emulating $r$ -$\ell$WL). For fxed $t,r$ 20 , $we$ have $c_{r}^{( t) }$ $\sqsubseteq$ $h_{r}^{( t) }$ . $Moreover$, $h_{r}^{( t) }$ $\sqsubseteq$ $c_{r}^{( t) }$ $if$ the message and update functions in Definition 12 are injective. In particular, t iterations of r - lWL are equally expressive as a t -layered r - lMPNN.
与 1 -WL 情况 (Xu et al., 2019; Morris et al., 2019) 类似先前的结果得出了$r-\ell$MPNN 与 $r$ -$\ell$WL 一样表达的条件
为了在实践中实现$r-\ell$MPNN,我们为$\mathrm{AGGR}_{k}^{(t)},\mathrm{UPD}^{(t)}$和(3)中的 READ 选择合适的神经层。作为(Xu et al., 2019, Lemma 5)的结果，对于合适的函数$f,g$ ,聚合函数可以写成
## $6. \textbf{ Loopy Message Passing}$
在本节中，我们构建$r$的神经版本- $\ell$WL
$$\mathrm{AGGR}_k^{(t+1)}(\mathcal{N}_k(v))=f\left(\sum_{\boldsymbol{p}\in\mathcal{N}_k(v)}g(\boldsymbol{p})\right)$$
$^{1}$For technical reasons, in the remainder of this subsection, we augment node features of $p\in\mathcal{N}_r(v)$ by adding atomic types. This indicates for every $p\in\boldsymbol{p}$ if $\{p,v\}\in E(\bar{G})$ ,see Definition 15 and (7) . Note that this does not increase the computationa complexity while being more powerful than the $r-\ell$WL given in Definition 10 .
。由于 1\_WL 在森林上是内射的，因此在路径 (Arvind et al.,2015)上，并且由于 GIN 可以近似 1-WL (Xu et al.,
5
|
[
[
0.564453125,
0.08709716796875,
0.62744140625,
0.08709716796875,
0.62744140625,
0.10162353515625,
0.564453125,
0.10162353515625
],
[
0.67724609375,
0.08837890625,
0.7138671875,
0.08837890625,
0.7138671875,
0.0997314453125,
0.67724609375,
0.0997314453125
],
[
0.76123046875,
0.083984375,
0.81201171875,
0.08709716796875,
0.810546875,
0.1029052734375,
0.7607421875,
0.0997314453125
],
[
0.27294921875,
0.2052001953125,
0.330078125,
0.2052001953125,
0.330078125,
0.216552734375,
0.27294921875,
0.216552734375
],
[
0.64794921875,
0.2183837890625,
0.68115234375,
0.2183837890625,
0.68115234375,
0.2279052734375,
0.64794921875,
0.2279052734375
],
[
0.08905029296875,
0.2412109375,
0.8857421875,
0.2412109375,
0.8857421875,
0.255615234375,
0.08905029296875,
0.255615234375
],
[
0.08905029296875,
0.25830078125,
0.336669921875,
0.25830078125,
0.336669921875,
0.271484375,
0.08905029296875,
0.271484375
],
[
0.08905029296875,
0.2998046875,
0.4755859375,
0.2998046875,
0.4755859375,
0.314453125,
0.08905029296875,
0.314453125
],
[
0.498291015625,
0.298583984375,
0.8857421875,
0.298583984375,
0.8857421875,
0.3125,
0.498291015625,
0.3125
],
[
0.08905029296875,
0.31494140625,
0.473876953125,
0.31494140625,
0.473876953125,
0.328857421875,
0.08905029296875,
0.328857421875
],
[
0.5,
0.31494140625,
0.88232421875,
0.31494140625,
0.88232421875,
0.330078125,
0.5,
0.330078125
],
[
0.0882568359375,
0.330078125,
0.473876953125,
0.330078125,
0.473876953125,
0.3447265625,
0.0882568359375,
0.3447265625
],
[
0.5,
0.330810546875,
0.6953125,
0.330810546875,
0.6953125,
0.3447265625,
0.5,
0.3447265625
],
[
0.08905029296875,
0.3447265625,
0.473876953125,
0.3447265625,
0.473876953125,
0.35791015625,
0.08905029296875,
0.35791015625
],
[
0.0882568359375,
0.35986328125,
0.473876953125,
0.35986328125,
0.473876953125,
0.373779296875,
0.0882568359375,
0.373779296875
],
[
0.5009765625,
0.35986328125,
0.87158203125,
0.35986328125,
0.87158203125,
0.3837890625,
0.5009765625,
0.3837890625
],
[
0.0882568359375,
0.375732421875,
0.20751953125,
0.375732421875,
0.20751953125,
0.389404296875,
0.0882568359375,
0.389404296875
],
[
0.5087890625,
0.38818359375,
0.79736328125,
0.389404296875,
0.79736328125,
0.410888671875,
0.5087890625,
0.410400390625
],
[
0.0906982421875,
0.40478515625,
0.47216796875,
0.40478515625,
0.47216796875,
0.41845703125,
0.0906982421875,
0.41845703125
],
[
0.08905029296875,
0.42041015625,
0.283447265625,
0.42041015625,
0.283447265625,
0.434326171875,
0.08905029296875,
0.434326171875
],
[
0.736328125,
0.4267578125,
0.74853515625,
0.4267578125,
0.74853515625,
0.4375,
0.736328125,
0.4375
],
[
0.0882568359375,
0.44140625,
0.47216796875,
0.442626953125,
0.47216796875,
0.45703125,
0.0882568359375,
0.455810546875
],
[
0.72216796875,
0.44580078125,
0.80810546875,
0.44580078125,
0.80810546875,
0.46728515625,
0.72216796875,
0.46728515625
],
[
0.08660888671875,
0.45703125,
0.4755859375,
0.453857421875,
0.4755859375,
0.4716796875,
0.08660888671875,
0.474853515625
],
[
0.5185546875,
0.474853515625,
0.84228515625,
0.474853515625,
0.84228515625,
0.49560546875,
0.5185546875,
0.49560546875
],
[
0.86376953125,
0.496826171875,
0.8857421875,
0.496826171875,
0.8857421875,
0.51025390625,
0.86376953125,
0.51025390625
],
[
0.0906982421875,
0.50390625,
0.4697265625,
0.50390625,
0.4697265625,
0.517578125,
0.0906982421875,
0.517578125
],
[
0.0906982421875,
0.51806640625,
0.47216796875,
0.51806640625,
0.47216796875,
0.5322265625,
0.0906982421875,
0.5322265625
],
[
0.5009765625,
0.52587890625,
0.8837890625,
0.52587890625,
0.8837890625,
0.54345703125,
0.5009765625,
0.54345703125
],
[
0.0906982421875,
0.53271484375,
0.471435546875,
0.53271484375,
0.471435546875,
0.54736328125,
0.0906982421875,
0.54736328125
],
[
0.08905029296875,
0.548828125,
0.47216796875,
0.548828125,
0.47216796875,
0.5625,
0.08905029296875,
0.5625
],
[
0.498291015625,
0.54345703125,
0.88134765625,
0.54248046875,
0.88134765625,
0.56005859375,
0.498291015625,
0.56103515625
],
[
0.08905029296875,
0.5625,
0.22216796875,
0.56298828125,
0.22216796875,
0.57763671875,
0.08905029296875,
0.5771484375
],
[
0.498291015625,
0.55810546875,
0.6357421875,
0.56103515625,
0.6357421875,
0.5791015625,
0.498291015625,
0.57568359375
],
[
0.08905029296875,
0.583984375,
0.47216796875,
0.58544921875,
0.47216796875,
0.599609375,
0.08905029296875,
0.5986328125
],
[
0.5009765625,
0.5859375,
0.8857421875,
0.5859375,
0.8857421875,
0.599609375,
0.5009765625,
0.599609375
],
[
0.0906982421875,
0.6015625,
0.47216796875,
0.6015625,
0.47216796875,
0.61572265625,
0.0906982421875,
0.61572265625
],
[
0.5,
0.6015625,
0.8857421875,
0.6015625,
0.8857421875,
0.6162109375,
0.5,
0.6162109375
],
[
0.1192626953125,
0.6162109375,
0.47216796875,
0.6162109375,
0.47216796875,
0.6298828125,
0.1192626953125,
0.6298828125
],
[
0.5009765625,
0.61865234375,
0.88232421875,
0.61865234375,
0.88232421875,
0.6318359375,
0.5009765625,
0.6318359375
],
[
0.08905029296875,
0.6298828125,
0.44677734375,
0.630859375,
0.44677734375,
0.64599609375,
0.08905029296875,
0.64453125
],
[
0.5,
0.63330078125,
0.8837890625,
0.63330078125,
0.8837890625,
0.646484375,
0.5,
0.646484375
],
[
0.08905029296875,
0.65380859375,
0.47216796875,
0.65380859375,
0.47216796875,
0.66748046875,
0.08905029296875,
0.66748046875
],
[
0.5,
0.6494140625,
0.6787109375,
0.64794921875,
0.6787109375,
0.6611328125,
0.5,
0.66357421875
],
[
0.08905029296875,
0.66845703125,
0.47216796875,
0.66845703125,
0.47216796875,
0.681640625,
0.08905029296875,
0.681640625
],
[
0.0882568359375,
0.681640625,
0.47216796875,
0.6845703125,
0.47216796875,
0.69873046875,
0.0882568359375,
0.6962890625
],
[
0.50244140625,
0.6787109375,
0.88232421875,
0.6787109375,
0.88232421875,
0.6923828125,
0.50244140625,
0.6923828125
],
[
0.50244140625,
0.693359375,
0.8857421875,
0.693359375,
0.8857421875,
0.70849609375,
0.50244140625,
0.70849609375
],
[
0.0906982421875,
0.69970703125,
0.47216796875,
0.69970703125,
0.47216796875,
0.71337890625,
0.0906982421875,
0.71337890625
],
[
0.0906982421875,
0.7138671875,
0.47216796875,
0.7138671875,
0.47216796875,
0.72802734375,
0.0906982421875,
0.72802734375
],
[
0.5009765625,
0.708984375,
0.5361328125,
0.708984375,
0.5361328125,
0.724609375,
0.5009765625,
0.724609375
],
[
0.09234619140625,
0.72998046875,
0.473876953125,
0.72998046875,
0.473876953125,
0.74365234375,
0.09234619140625,
0.74365234375
],
[
0.5009765625,
0.73095703125,
0.8857421875,
0.73193359375,
0.8857421875,
0.7470703125,
0.5009765625,
0.74560546875
],
[
0.08905029296875,
0.74365234375,
0.361083984375,
0.744140625,
0.361083984375,
0.7587890625,
0.08905029296875,
0.75830078125
],
[
0.5009765625,
0.74755859375,
0.8837890625,
0.74755859375,
0.8837890625,
0.76220703125,
0.5009765625,
0.76220703125
],
[
0.5,
0.76220703125,
0.88623046875,
0.76123046875,
0.88623046875,
0.77880859375,
0.5,
0.77978515625
],
[
0.0882568359375,
0.77978515625,
0.310546875,
0.77978515625,
0.310546875,
0.79443359375,
0.0882568359375,
0.79443359375
],
[
0.50244140625,
0.77978515625,
0.73193359375,
0.77978515625,
0.73193359375,
0.79443359375,
0.50244140625,
0.79443359375
],
[
0.08905029296875,
0.8046875,
0.375732421875,
0.8046875,
0.375732421875,
0.8193359375,
0.08905029296875,
0.8193359375
],
[
0.5498046875,
0.8271484375,
0.611328125,
0.8271484375,
0.611328125,
0.84130859375,
0.5498046875,
0.84130859375
],
[
0.60302734375,
0.82373046875,
0.71826171875,
0.8271484375,
0.71826171875,
0.8447265625,
0.60302734375,
0.84130859375
],
[
0.7607421875,
0.82763671875,
0.8203125,
0.82763671875,
0.8203125,
0.84228515625,
0.7607421875,
0.84228515625
],
[
0.08905029296875,
0.84130859375,
0.471435546875,
0.84130859375,
0.471435546875,
0.85498046875,
0.08905029296875,
0.85498046875
],
[
0.108642578125,
0.8291015625,
0.47216796875,
0.830078125,
0.47216796875,
0.84130859375,
0.108642578125,
0.84033203125
],
[
0.73193359375,
0.84521484375,
0.7900390625,
0.84521484375,
0.7900390625,
0.85693359375,
0.73193359375,
0.85693359375
],
[
0.08905029296875,
0.85498046875,
0.473876953125,
0.85498046875,
0.473876953125,
0.86865234375,
0.08905029296875,
0.86865234375
],
[
0.0882568359375,
0.8671875,
0.473876953125,
0.8671875,
0.473876953125,
0.880859375,
0.0882568359375,
0.880859375
],
[
0.08905029296875,
0.8798828125,
0.473876953125,
0.8798828125,
0.473876953125,
0.89306640625,
0.08905029296875,
0.89306640625
],
[
0.5009765625,
0.87548828125,
0.88623046875,
0.87548828125,
0.88623046875,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.08905029296875,
0.89306640625,
0.176513671875,
0.89306640625,
0.176513671875,
0.90478515625,
0.08905029296875,
0.90478515625
],
[
0.5009765625,
0.892578125,
0.8857421875,
0.892578125,
0.8857421875,
0.90576171875,
0.5009765625,
0.90576171875
],
[
0.48193359375,
0.9248046875,
0.494384765625,
0.9248046875,
0.494384765625,
0.93505859375,
0.48193359375,
0.93505859375
]
] |
[
"No(v)",
"Ni(",
".N2(u)",
"(a) r = 0",
"(b)r:",
"Figure 2: Example of two non-isomorphic graphs that are r - WL equivalent but not (r + 1) - WL equivalent: a chordal",
"cycle (left) and a cactus graph (right).",
"Definition 11 (Cactus Graph). A cactus graph is a graph",
"Definition 12 (r -MPNN). For t E {0,...,T - 1) and",
"where every edge lies on at most one simple cycle. For",
"k E [O, ..,r] , r -lMPNN applies the following message",
"r ≥ 2, an r -cactus graph is a cactus where every simple",
"update and readout functions",
"cyclehas at most r vertices.Wedenote byM thesetofall",
"cactus graphs, and by Mr the set comprising all q -cactus",
"m(t+1)(0) = AGGR(+1) ( h(P) I pE Ns())",
"graphs for q ≤r.",
"h(t+1)() = UPD(++1) (h((u), mf++1)(0)",
"我们参考Figure2来说明两个小仙人掌图。我们现在",
"准备介绍我们的主要结果。",
":",
"Theorem 2 (Homomorphism Counting Power of r -WL)",
"m(t+1)())",
"Let r ≥ 0. Then, r -lWL can homomorphism-count Mr+2",
"h(T)(G) = READ ({[h(T)(u) / u E V(G)/)",
"(3)",
"我们参考AppendixD来获得Theorem2的详细证明",
"该证明相当复杂，需要定义仙人掌图的规范树分解和",
"函数htt)：V(G)→C映射ht(u)是一个节点",
"r-EWL的展开树。演示它们的强连接，然后我们按照",
"(Dell etal.,2018;B.Zhangetal.,2024）中的方法分解仙",
"Invarant。在以下结果中，我们建立了（2）中定义的hkt)",
"人掌图的同态计数",
"和c(t)之间的关系",
"类M2仅包含林；因此，Theorem2暗示了1-WL计数",
"Theorem 3 (GNN emulating r -lWL). For fixed t, r ≥",
"林能力的标准结果。由于森林是1-WL可以计数的唯",
"O,wehavec(t)hit).Moreover,hit)c(t)ifthe",
"类图，因此Theorem2暗示r-eWL始终比1-WL",
"message and update functions in Definition 12 are injective.",
"严格地更强大，从而证实了Proposition1中的主张",
"In particular, t iterations of r -lwL are equally expressive",
"Theorem2的含义是深刻的：它建立了r-WL可以同",
"as at -layered r -lMPNN.",
"态计数一大类图。此结果表明，即使是最简单的变体1",
"IWL，其功能也不亚于子图GNN，后者仅限于端点",
"与1-WL情况(Xuetal.,2019;Morrisetal.,2019)类似",
"先前的结果得出了-EMPNN与-(WL一样表达的",
"共享NED(B.Zhangetal.，2024）的同态计数图。重要",
"的是，Theorem2陈述了r-WL的同态表达率的宽松",
"条件",
"下界。这一观察结果为未来研究探索r-LWL的同态",
"为了在实践中实现r-MPNN，我们为",
"表达率的严格下界或上界开辟了道路。",
"AGGR(t),UPD(t）和（3）中的READ选择合适的神经",
"层。作为(Xuetal.，2019,Lemma5)的结果，对于合适",
"6. Loopy Message Passing",
"的函数f,9，聚合函数可以写成",
"在本节中，我们构建r的神经版本-eWL",
"AGGR",
"(t+1)(Ns(u) = f",
"g(p)",
"augment node features of p e N.(u) by adding atomic types.",
"'Fortechnical reasons,intheremainderofthissubsection,we",
"(PEN(u)",
"This indicates for every p E p if (p,v) e E(G) , see Defini-",
"tion 15 and (7). Note that this does not increase the computational",
"complexity whilebeing more powerful than ther - (WLgiven in",
"。由于1_WL在森林上是内射的，因此在路径（Arvind",
"Definition 10.",
"etal.,2015）上，并且由于GIN可以近似1-WL(Xuetal.",
"5"
] |
{
"arxiv_id": "2403.13749",
"language": "zh",
"page_idx": 4
}
|
||
1805.02785-zh-1
|
1805.02785-zh-1.jpg
|
# 具有稀疏奖励的确定性 MDP 的快速在线精确解决方案
Joshua R. Bertram Xuxi Yang Peng Wei
Iowa State University
Ames, IA 50011
$\{$bertram1, xuxiyang, pwei $\}$ @iastate.edu
## $\mathbf{Abstract}$
Markov Decision Processes (MDPs) are a mathematical framework for modeling sequential decision making under uncertainty. The classical approaches for solving MDPs are well known and have been widely studied, some of which rely on approximation techniques to solve MDPs with large state space and/or action space. However, most of these classical solution approaches and their approximation techniques still take much computation time to converge and usually must be re-computed if the reward function is changed. This paper introduces a novel alternative approach for exactly and efficiently solving deterministic, continuous MDPs with sparse reward sources. When the environment is such that the"distance" between states can be determined in constant time, e.g. grid world, our algorithm offers $O(|R|^2\times|A|^2\times|S|)$ , where $|R|$ is the number of reward sources, $\left|A\right|$is the number of actions, and $\left|S\right|$ is the number of states. Memory complexity for the algorithm is $O(|S|+|R|\times|A|).$ This new approach opens new avenues for boosting computational performance for certain classes of MDPs and is of tremendous value for MDP applications such as robotics and unmanned systems. This paper describes the algorithm and presents numerical experiment results to demonstrate its powerful computational performance. We also provide rigorous mathematical description of the approach.
1 $\textbf{Introduction}$
马尔可夫决策过程 (MDP) 是一种用于对代理的决策过程进行建模的方法，其中代理通过采取行动从当前状态转换到下一个状态。下一个状态的值(以及因此的动作)是通过奖励函数确定的，该奖励函数描述状态空间内的哪些状态提供奖励(或惩罚。)MDP 的解决方案，称为最优策略，描述了从任何状态采取的最优行动，并最大化 the expected cumulative 奖励。从任何初始状态遵循最优策略将导致最优 sequence of actions 通过最大化 cumulative 奖励的状态空间。虽然 MDP 是一种强大而通用的技术，但它们很难扩展到大型状态或操作空间。用于解决 MDP 的原始动态编程算法 such as value iteration and policy iteration 需要将状态 (或状态动作) 空间的完整表示存储在内存中，从而使许多复杂的 MDP 问题难以解决。此外，大多数 MDP 算法都是为固定的 MDP 设计的，其中环境和奖励系统是固定的，并且如果奖励发生变化，则需要重新计算解决方案。考虑到求解任何相当大的 MDP 所需的计算时间，这通常会阻止 MDP 在非平稳奖励系统中使用，除非可以重新制定 MDP 以捕获奖励函数中的变化作为状态本身的一部分本文提出了一种替代方法 for 有效地精确求解确定性连续 MDP, 并为单个奖励源提供了严格的数学证明。本文还提出了一种精确求解具有多个奖励源的确定性连续 MDP 的算法，并给出了数值实验结果以验证该算法并展示其强大的计算性能。我们还证明了该算法的改进性能允许在适合实时系统的执行时间内在线计算具有变化的奖励函数的 MDP, 这对于机器人，无人系统和其他具有结构化状态的领域中的 MDP 应用将具有巨大的价值空间和稀疏的奖励函数。我们的算法性能与标准值迭代算法进行了
Preprint. Work in progress.
|
[
[
0.1846923828125,
0.1287841796875,
0.81787109375,
0.1287841796875,
0.81787109375,
0.146484375,
0.1846923828125,
0.146484375
],
[
0.484375,
0.15283203125,
0.515625,
0.15283203125,
0.515625,
0.175537109375,
0.484375,
0.175537109375
],
[
0.3447265625,
0.2291259765625,
0.57666015625,
0.2291259765625,
0.57666015625,
0.242431640625,
0.3447265625,
0.242431640625
],
[
0.57275390625,
0.2310791015625,
0.58837890625,
0.2310791015625,
0.58837890625,
0.238037109375,
0.57275390625,
0.238037109375
],
[
0.583984375,
0.2298583984375,
0.65673828125,
0.2298583984375,
0.65673828125,
0.2412109375,
0.583984375,
0.2412109375
],
[
0.4248046875,
0.2412109375,
0.5703125,
0.242431640625,
0.5703125,
0.2568359375,
0.4248046875,
0.255615234375
],
[
0.443603515625,
0.2568359375,
0.5537109375,
0.2568359375,
0.5537109375,
0.270751953125,
0.443603515625,
0.270751953125
],
[
0.326904296875,
0.270751953125,
0.6806640625,
0.270751953125,
0.6806640625,
0.285400390625,
0.326904296875,
0.285400390625
],
[
0.463134765625,
0.32080078125,
0.53759765625,
0.32080078125,
0.53759765625,
0.335205078125,
0.463134765625,
0.335205078125
],
[
0.233642578125,
0.352783203125,
0.7646484375,
0.352783203125,
0.7646484375,
0.3662109375,
0.233642578125,
0.3662109375
],
[
0.233642578125,
0.367431640625,
0.76318359375,
0.367431640625,
0.76318359375,
0.380615234375,
0.233642578125,
0.380615234375
],
[
0.233642578125,
0.380126953125,
0.7646484375,
0.380126953125,
0.7646484375,
0.39453125,
0.233642578125,
0.39453125
],
[
0.2344970703125,
0.39453125,
0.7646484375,
0.39453125,
0.7646484375,
0.4091796875,
0.2344970703125,
0.4091796875
],
[
0.2344970703125,
0.40771484375,
0.765625,
0.40771484375,
0.765625,
0.421630859375,
0.2344970703125,
0.421630859375
],
[
0.233642578125,
0.421630859375,
0.765625,
0.421630859375,
0.765625,
0.43505859375,
0.233642578125,
0.43505859375
],
[
0.233642578125,
0.436279296875,
0.765625,
0.436279296875,
0.765625,
0.449462890625,
0.233642578125,
0.449462890625
],
[
0.2344970703125,
0.449462890625,
0.76318359375,
0.449462890625,
0.76318359375,
0.46337890625,
0.2344970703125,
0.46337890625
],
[
0.233642578125,
0.46337890625,
0.765625,
0.46337890625,
0.765625,
0.477783203125,
0.233642578125,
0.477783203125
],
[
0.2320556640625,
0.47607421875,
0.76708984375,
0.4765625,
0.76708984375,
0.4912109375,
0.2320556640625,
0.490478515625
],
[
0.233642578125,
0.4892578125,
0.7646484375,
0.490478515625,
0.7646484375,
0.5048828125,
0.233642578125,
0.50390625
],
[
0.236083984375,
0.50390625,
0.7646484375,
0.5048828125,
0.7646484375,
0.51806640625,
0.236083984375,
0.517578125
],
[
0.233642578125,
0.51806640625,
0.76318359375,
0.51806640625,
0.76318359375,
0.5322265625,
0.233642578125,
0.5322265625
],
[
0.233642578125,
0.5322265625,
0.765625,
0.5322265625,
0.765625,
0.546875,
0.233642578125,
0.546875
],
[
0.233642578125,
0.54541015625,
0.7607421875,
0.54541015625,
0.7607421875,
0.5595703125,
0.233642578125,
0.5595703125
],
[
0.233642578125,
0.55810546875,
0.7646484375,
0.5595703125,
0.7646484375,
0.57275390625,
0.233642578125,
0.5712890625
],
[
0.233642578125,
0.57373046875,
0.7646484375,
0.57373046875,
0.7646484375,
0.5869140625,
0.233642578125,
0.5869140625
],
[
0.2320556640625,
0.5859375,
0.5087890625,
0.5869140625,
0.5087890625,
0.6015625,
0.2320556640625,
0.60107421875
],
[
0.1748046875,
0.6298828125,
0.2091064453125,
0.6298828125,
0.2091064453125,
0.642578125,
0.1748046875,
0.642578125
],
[
0.2034912109375,
0.62890625,
0.31787109375,
0.6298828125,
0.31787109375,
0.64453125,
0.2034912109375,
0.64306640625
],
[
0.176513671875,
0.66650390625,
0.82275390625,
0.66650390625,
0.82275390625,
0.6796875,
0.176513671875,
0.6796875
],
[
0.1748046875,
0.6787109375,
0.82275390625,
0.6796875,
0.82275390625,
0.6943359375,
0.1748046875,
0.693359375
],
[
0.176513671875,
0.6943359375,
0.82177734375,
0.6943359375,
0.82177734375,
0.70849609375,
0.176513671875,
0.70849609375
],
[
0.1732177734375,
0.70849609375,
0.81884765625,
0.70849609375,
0.81884765625,
0.7216796875,
0.1732177734375,
0.7216796875
],
[
0.176513671875,
0.72265625,
0.8203125,
0.72265625,
0.8203125,
0.736328125,
0.176513671875,
0.736328125
],
[
0.176513671875,
0.73486328125,
0.82275390625,
0.73486328125,
0.82275390625,
0.74853515625,
0.176513671875,
0.74853515625
],
[
0.1748046875,
0.74853515625,
0.82275390625,
0.75,
0.82275390625,
0.7646484375,
0.1748046875,
0.76318359375
],
[
0.1748046875,
0.76318359375,
0.82177734375,
0.76318359375,
0.82177734375,
0.7763671875,
0.1748046875,
0.7763671875
],
[
0.176513671875,
0.7763671875,
0.82177734375,
0.7763671875,
0.82177734375,
0.79052734375,
0.176513671875,
0.79052734375
],
[
0.176513671875,
0.79150390625,
0.82177734375,
0.79150390625,
0.82177734375,
0.8046875,
0.176513671875,
0.8046875
],
[
0.1748046875,
0.8046875,
0.82275390625,
0.8046875,
0.82275390625,
0.818359375,
0.1748046875,
0.818359375
],
[
0.176513671875,
0.818359375,
0.82275390625,
0.818359375,
0.82275390625,
0.83203125,
0.176513671875,
0.83203125
],
[
0.176513671875,
0.83203125,
0.82275390625,
0.83203125,
0.82275390625,
0.84521484375,
0.176513671875,
0.84521484375
],
[
0.176513671875,
0.8466796875,
0.82275390625,
0.8466796875,
0.82275390625,
0.85986328125,
0.176513671875,
0.85986328125
],
[
0.1748046875,
0.85986328125,
0.8203125,
0.85986328125,
0.8203125,
0.87353515625,
0.1748046875,
0.87353515625
],
[
0.1748046875,
0.87255859375,
0.82275390625,
0.87353515625,
0.82275390625,
0.88818359375,
0.1748046875,
0.88720703125
],
[
0.176513671875,
0.88818359375,
0.82177734375,
0.88818359375,
0.82177734375,
0.9013671875,
0.176513671875,
0.9013671875
],
[
0.1748046875,
0.923828125,
0.33251953125,
0.9248046875,
0.33251953125,
0.93798828125,
0.1748046875,
0.9375
]
] |
[
"具有稀疏奖励的确定性MIDP的快速在线精确解决方",
"案",
"Joshua R.BertramXuxi Yang",
"",
"Peng Wei",
"lowa StateUniversity",
"Ames, IA.50011",
"fbertraml,xuxiyang,pweiJ @iastate.edu",
"Abstract",
"Markov Decision Processes (MDPs) are a mathematical framework for modeling",
"sequential decision making under uncertainty. The classical approaches for solv",
"ing MDPs are well known and have been widely studied, some of which rely",
"on approximation techniques to solve MDPs with large state space and/or action",
"space. However, most of these classical solution approaches and their approxi",
"mation techniques still take much computation time to converge and usually must",
"be re-computed if the reward function is changed. This paper introduces a novel",
"alternative approach for exactly and efficiently solving deterministic, continuous",
"MDPs with sparse reward sources. When the environment is such that the \"dis-",
"tance\" between states can be determined in constant time, e.g. grid world, our al-",
"gorithm offers O(/R|2 ×A|2 ×SI),whereR| is the number of reward sources,",
"A is the number of actions, and S|is the number of states.Memory complexity",
"for the algorithm is O(|S+R|× Al). This new approach opens new avenues",
"for boosting computational performance for certain classes of MDPs and is of",
"tremendous value for MDP applications such as robotics and unmanned systems",
"This paper describes the algorithm and presents numerical experiment results to",
"demonstrate its powerful computational performance.We also provide rigorous",
"mathematical description of the approach.",
"1",
"Introduction",
"马尔可夫决策过程（MDP）是一种用于对代理的决策过程进行建模的方法，其中代理通过",
"采取行动从当前状态转换到下一个状态。下一个状态的值（以及因此的动作）是通过奖励",
"函数确定的，该奖励函数描述状态空间内的哪些状态提供奖励（或惩罚。)MDP的解决方",
"案，称为最优策略，描述了从任何状态采取的最优行动，并最大化theexpectedcumulative",
"奖励。从任何初始状态遵循最优策略将导致最优sequenceofactions通过最大化cumulative",
"奖励的状态空间。虽然MDP是一种强大而通用的技术，但它们很难扩展到大型状态或操作",
"空间。用于解决MDP的原始动态编程算法suchasvalueiterationandpolicyiteration需要将",
"状态（或状态动作）空间的完整表示存储在内存中，从而使许多复杂的MDP问题难以解决",
"此外，大多数MDP算法都是为固定的MDP设计的，其中环境和奖励系统是固定的，并且",
"如果奖励发生变化，则需要重新计算解决方案。考虑到求解任何相当大的MDP所需的计算",
"时间，这通常会阻止MDP在非平稳奖励系统中使用，除非可以重新制定MDP以捕获奖励",
"函数中的变化作为状态本身的一部分本文提出了一种替代方法for有效地精确求解确定性",
"连续MDP，并为单个奖励源提供了严格的数学证明。本文还提出了一种精确求解具有多个",
"奖励源的确定性连续MDP的算法，并给出了数值实验结果以验证该算法并展示其强大的计",
"算性能。我们还证明了该算法的改进性能允许在适合实时系统的执行时间内在线计算具有",
"变化的奖励函数的MDP，这对于机器人，无人系统和其他具有结构化状态的领域中的MDP",
"应用将具有巨大的价值空间和稀疏的奖励函数。我们的算法性能与标准值迭代算法进行了",
"Preprint. Work in progress"
] |
{
"arxiv_id": "1805.02785",
"language": "zh",
"page_idx": 1
}
|
||
1805.02785-zh-2
|
1805.02785-zh-2.jpg
|
比较，以 demonstrate the substantial performance improvement 。在本文中，我们将自己限制在具有非负实际奖励的确定性连续 MDP 中，以限制论文的长度，并将尝试在以后的工作中推广到更广泛的 MDP 类别。
## 2 $\textbf{Background}$
## 2.1 Markov Decision Processes
将 MDP 描述为具有状态空间$S$ 、an 动作空间$A$ 、a 转换函数 XMATHX\_MA6。从给定的状态$s\in S$开始，操作$a\in A$会导致下一个状态$s^\prime\in S$ ,概率为$T(s,a,s^\prime)$ 。Reward functions 通常定义为从状态$s$采取操作$a$后获得的立即奖励$R(s,a)$ (尽管在本文中仅考虑仅取决于状态的奖励$R(s)$ 。) MDP 假设马尔可夫 property 适用于该问题，该问题指出未来状态$s^\prime$仅取决于当前状态$s$和采取的操作$a$ 。A policy $\pi \textbf{ describes an action }a$ that should be performed at each state $s;$ an optimal policy $\pi^*$ is a policy that generates maximum cumulative reward for every state. Solving an MDP is equivalent to finding the optimal policy $\pi^*.$
在这篇初始论文中，我们将自己限制在确定性的连续 MDP 上；为了便于说明，在不失一般性的情况下，我们使用二维网格世界作为示例环境。在确定性 MDP 中，指定的操作(e。g。, 上，下，左，右)发生无故障或错误的概率为 1.0 。连续的 MDP一旦达到目标状态，就没有序列的自然结束；它通过目标状态并继续累积奖励而没有结束。In a grid world,the state corresponds to a cell of the grid. At each state, the agent has four actions to choose from: up, down, left, and right, except for boundary state, where the agent is not allowed to take actions which lead the agent to the boudary. Thus in boundary state, the agent has two or three actions to choose from. 请注意，在本文中，我们将使用以下约定来区分时间$t$与带有上标和括号的$s^{(t)}$的状态以及带有下标的特定状态$s_k\in S$ 。因此，在时间$t$的状态$s_k$将表示为$s_k^(t)$ 。类似地，在时间$t$的动作$a_k$和奖励$r_k$将分别表示为$a_k^(t)$和$r_k^(t)$。上标本身表示提升为幂，例如在$\gamma^t$中将折扣因子$\gamma$提升为$t$的幂。状态$s$可以指状态$s\in S$或“当前”状态，其中状态$s^\prime$总是指下一个状态
## 2.2 $\textbf{Value Iteration}$
求解 MDP 的最基本方法之一是值迭代。值迭代是一种求解 MDP 的动态规划方法，它迭代地确定每个状态的值。In an infnite horizon problem with a discount factor of $0.0<\gamma<1.0$ the expected cumulative reward (or value) at time step $t$ associated with a sequence of immediate rewards $r_t$ is:
$$V(s)=\sum_{t=0}^\infty\gamma^tr^{(t)}(s^{(t)},a^{(t)})$$
(1)
More generally, this is expressed recursively using the Bellman equation [1].
(2)
$$V_n(s^{(t)})=\max_a\Big[r(s^{(t)},a)+\gamma\sum_{s^{(t+1)}}T(s^{(t+1)}|s^{(t)},a)V_{n-1}(s^{(t+1)})\Big],$$
where $n$ is the current iteration of the value iteration algorithm. Value iteration obtains the optimal value when the policy and value function become stationary with respect to the Bellman operator $L$ satisfying the equation $V^*=LV^*$ .(See Chapter 1 of [2] for more information on this important topic as well as Bellman's original treatment[1].)
检查值迭代的运行时复杂性，从[2,3]开始，值迭代算法的每次迭代都采用$O( | A| \times | S| ^2)$ ,
并且该算法所需的总体最大迭代次数是$\left|S\right|,\left|A\right|$和$\frac1{1-\gamma}\log\frac1{1-\gamma}$中的多项式。
### 3 $\textbf{Related Work}$
存在许多解决 MDP 的 classical 方法 (请参阅优秀的文本[4] ,[2] ,[5]和[6])。The value iteration algorithm requires that the entire state space be held in memory as a table; exceedingly large state and action spaces will exhaust the memory available to even large, powerful computers. Thus at some level of complexity, value iteration becomes intractable due to the explosion in stateaction space. 一些算法通过利用问题的结构来更紧凑地表示 MDP[7],[8] 来解决状态操作
3
|
[
[
0.17724609375,
0.0921630859375,
0.60546875,
0.0921630859375,
0.60546875,
0.1060791015625,
0.17724609375,
0.1060791015625
],
[
0.5986328125,
0.0947265625,
0.82177734375,
0.0947265625,
0.82177734375,
0.10479736328125,
0.5986328125,
0.10479736328125
],
[
0.176513671875,
0.106689453125,
0.82177734375,
0.106689453125,
0.82177734375,
0.12060546875,
0.176513671875,
0.12060546875
],
[
0.1748046875,
0.121826171875,
0.373291015625,
0.121826171875,
0.373291015625,
0.1351318359375,
0.1748046875,
0.1351318359375
],
[
0.1732177734375,
0.1546630859375,
0.2099609375,
0.1546630859375,
0.2099609375,
0.17041015625,
0.1732177734375,
0.17041015625
],
[
0.20751953125,
0.1558837890625,
0.31201171875,
0.1558837890625,
0.31201171875,
0.17041015625,
0.20751953125,
0.17041015625
],
[
0.1732177734375,
0.1925048828125,
0.406005859375,
0.1925048828125,
0.406005859375,
0.20703125,
0.1732177734375,
0.20703125
],
[
0.1748046875,
0.2197265625,
0.82177734375,
0.2197265625,
0.82177734375,
0.23291015625,
0.1748046875,
0.23291015625
],
[
0.176513671875,
0.2342529296875,
0.8203125,
0.2342529296875,
0.8203125,
0.2474365234375,
0.176513671875,
0.2474365234375
],
[
0.176513671875,
0.2474365234375,
0.82177734375,
0.2474365234375,
0.82177734375,
0.261474609375,
0.176513671875,
0.261474609375
],
[
0.1748046875,
0.261962890625,
0.82177734375,
0.261962890625,
0.82177734375,
0.27587890625,
0.1748046875,
0.27587890625
],
[
0.176513671875,
0.274658203125,
0.82177734375,
0.274658203125,
0.82177734375,
0.28857421875,
0.176513671875,
0.28857421875
],
[
0.176513671875,
0.2890625,
0.82177734375,
0.2890625,
0.82177734375,
0.302978515625,
0.176513671875,
0.302978515625
],
[
0.1748046875,
0.302978515625,
0.63134765625,
0.302978515625,
0.63134765625,
0.31640625,
0.1748046875,
0.31640625
],
[
0.176513671875,
0.323974609375,
0.82421875,
0.323974609375,
0.82421875,
0.337158203125,
0.176513671875,
0.337158203125
],
[
0.176513671875,
0.337158203125,
0.8203125,
0.337158203125,
0.8203125,
0.3515625,
0.176513671875,
0.3515625
],
[
0.176513671875,
0.348388671875,
0.82177734375,
0.349853515625,
0.82177734375,
0.3642578125,
0.176513671875,
0.363037109375
],
[
0.1748046875,
0.363037109375,
0.82177734375,
0.363037109375,
0.82177734375,
0.376953125,
0.1748046875,
0.376953125
],
[
0.176513671875,
0.37890625,
0.8203125,
0.37890625,
0.8203125,
0.392578125,
0.176513671875,
0.392578125
],
[
0.176513671875,
0.392578125,
0.82177734375,
0.392578125,
0.82177734375,
0.406005859375,
0.176513671875,
0.406005859375
],
[
0.1748046875,
0.406005859375,
0.8203125,
0.406005859375,
0.8203125,
0.419189453125,
0.1748046875,
0.419189453125
],
[
0.181396484375,
0.42041015625,
0.82177734375,
0.42041015625,
0.82177734375,
0.43505859375,
0.181396484375,
0.43505859375
],
[
0.176513671875,
0.436279296875,
0.82275390625,
0.4375,
0.82275390625,
0.451904296875,
0.176513671875,
0.45068359375
],
[
0.176513671875,
0.455078125,
0.82275390625,
0.455078125,
0.82275390625,
0.4697265625,
0.176513671875,
0.4697265625
],
[
0.1748046875,
0.4697265625,
0.81640625,
0.468505859375,
0.81640625,
0.48291015625,
0.1748046875,
0.484130859375
],
[
0.176513671875,
0.48486328125,
0.2254638671875,
0.48486328125,
0.2254638671875,
0.496826171875,
0.176513671875,
0.496826171875
],
[
0.1732177734375,
0.51318359375,
0.32275390625,
0.51318359375,
0.32275390625,
0.52783203125,
0.1732177734375,
0.52783203125
],
[
0.176513671875,
0.54052734375,
0.82177734375,
0.54052734375,
0.82177734375,
0.5537109375,
0.176513671875,
0.5537109375
],
[
0.176513671875,
0.5537109375,
0.82275390625,
0.5537109375,
0.82275390625,
0.5673828125,
0.176513671875,
0.5673828125
],
[
0.1748046875,
0.568359375,
0.82275390625,
0.568359375,
0.82275390625,
0.58203125,
0.1748046875,
0.58203125
],
[
0.176513671875,
0.58349609375,
0.2607421875,
0.58349609375,
0.2607421875,
0.59326171875,
0.176513671875,
0.59326171875
],
[
0.40185546875,
0.6015625,
0.5947265625,
0.5986328125,
0.5947265625,
0.619140625,
0.40185546875,
0.62255859375
],
[
0.80224609375,
0.60595703125,
0.826171875,
0.60595703125,
0.826171875,
0.62060546875,
0.80224609375,
0.62060546875
],
[
0.463134765625,
0.62353515625,
0.48193359375,
0.62353515625,
0.48193359375,
0.630859375,
0.463134765625,
0.630859375
],
[
0.1748046875,
0.63525390625,
0.67626953125,
0.63525390625,
0.67626953125,
0.6484375,
0.1748046875,
0.6484375
],
[
0.271240234375,
0.65478515625,
0.7197265625,
0.65478515625,
0.7197265625,
0.67529296875,
0.271240234375,
0.67529296875
],
[
0.8037109375,
0.6591796875,
0.82421875,
0.6591796875,
0.82421875,
0.673828125,
0.8037109375,
0.673828125
],
[
0.48193359375,
0.67529296875,
0.51318359375,
0.67529296875,
0.51318359375,
0.6875,
0.48193359375,
0.6875
],
[
0.176513671875,
0.6923828125,
0.82177734375,
0.6923828125,
0.82177734375,
0.70556640625,
0.176513671875,
0.70556640625
],
[
0.1732177734375,
0.705078125,
0.82177734375,
0.705078125,
0.82177734375,
0.71826171875,
0.1732177734375,
0.71826171875
],
[
0.1748046875,
0.71826171875,
0.82177734375,
0.71826171875,
0.82177734375,
0.73291015625,
0.1748046875,
0.73291015625
],
[
0.1748046875,
0.734375,
0.5048828125,
0.734375,
0.5048828125,
0.74755859375,
0.1748046875,
0.74755859375
],
[
0.176513671875,
0.7548828125,
0.826171875,
0.7548828125,
0.826171875,
0.76806640625,
0.176513671875,
0.76806640625
],
[
0.1748046875,
0.76806640625,
0.71728515625,
0.76806640625,
0.71728515625,
0.78271484375,
0.1748046875,
0.78271484375
],
[
0.1732177734375,
0.8046875,
0.2099609375,
0.8046875,
0.2099609375,
0.82080078125,
0.1732177734375,
0.82080078125
],
[
0.205078125,
0.806640625,
0.326904296875,
0.806640625,
0.326904296875,
0.82080078125,
0.205078125,
0.82080078125
],
[
0.1748046875,
0.84326171875,
0.82177734375,
0.84326171875,
0.82177734375,
0.85693359375,
0.1748046875,
0.85693359375
],
[
0.176513671875,
0.85791015625,
0.82177734375,
0.85791015625,
0.82177734375,
0.8720703125,
0.176513671875,
0.8720703125
],
[
0.176513671875,
0.8720703125,
0.8203125,
0.8720703125,
0.8203125,
0.88525390625,
0.176513671875,
0.88525390625
],
[
0.176513671875,
0.88525390625,
0.82177734375,
0.88525390625,
0.82177734375,
0.8984375,
0.176513671875,
0.8984375
],
[
0.1748046875,
0.8984375,
0.82275390625,
0.8984375,
0.82275390625,
0.912109375,
0.1748046875,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"比较，以demonstratethesubstantialperformanceimprovement",
"。在本文中，我们将自已限制",
"在具有非负实际奖励的确定性连续MDP中，以限制论文的长度，并将尝试在以后的工作中",
"推广到更广泛的MDP类别。",
"2",
"Background",
"2.1Markov Decision Processes",
"将MDP描述为具有状态空间S、an动作空间A、a转换函数XMATHX_MA6。从给定的状",
"态sES开始，操作aEA会导致下一个状态sES，概率为T(s,a，s）。Rewardfunctions",
"通常定义为从状态s采取操作α后获得的立即奖励R(s,α）（尽管在本文中仅考虑仅取决于",
"状态的奖励R(s）。）MDP假设马尔可夫property适用于该问题，该问题指出未来状态s仅",
"取决于当前状态s和采取的操作a。Apolicyπdescribesanactionathatshouldbeperformedat",
"each state s ; an optimal policy * is a policy that generates maximum cumulative reward for every",
"state. Solving an MDP is equivalent to finding the optimal policy *.",
"在这篇初始论文中，我们将自已限制在确定性的连续MDP上；为了便于说明，在不失一般",
"性的情况下，我们使用二维网格世界作为示例环境。在确定性MDP中，指定的操作(e。g。",
"上，下，左，右）发生无故障或错误的概率为1.0。连续的MDP一旦达到目标状态，就没",
"有序列的自然结束；它通过目标状态并继续累积奖励而没有结束。Inagridworld,thestate",
"corresponds to a cell of the grid. At each state, the agent has four actions to choose from: up, down",
"left, and right, except for boundary state, where the agent is not allowed to take actions which lead",
"the agent to the boudary. Thus in boundary state, the agent has two or three actions to choose from",
"请注意，在本文中，我们将使用以下约定来区分时间t与带有上标和括号的s(t）的状态以",
"及带有下标的特定状态SkES。因此，在时间t的状态sk将表示为st。类似地，在时间",
"t的动作a和奖励rk将分别表示为at）和rt）。上标本身表示提升为幂，例如在中将折",
"扣因子提升为t的幂。状态s可以指状态sES或“当前”状态，其中状态s总是指下",
"个状态",
"2.2Value Iteration",
"求解MDP的最基本方法之一是值迭代。值迭代是一种求解MDP的动态规划方法，它迭",
"代地确定每个状态的值。In an infinite horizon problem with a discount factor of 0.0<<1.0",
"the expected cumulative reward (or value) at time step t associated with a sequence of immediate",
"rewards rt is",
"V(s) =Er(t)(s(t),a(t)",
"(1)",
"t-0",
"More generally, this is expressed recursively using the Bellman equation [1]",
"Vn(s(t) =max [r(s(t),a) + T(s(t+1)]s(t),a)Vn-1(s(t+1)]",
"(2)",
"s(t+1)",
"where n is the current iteration of the value iteration algorithm. Value iteration obtains the optimal",
"value when the policy and value function become stationary with respect to the Bellman operator L",
"satisfying the equation V* = LV* . (See Chapter 1 of [2j for more information on this important",
"topic as well as Bellman's original treatment [1] .]",
"检查值迭代的运行时复杂性，从[2，3]开始，值选代算法的每次迭代都采用O(IA|×S2），",
"并且该算法所需的总体最大选代次数是SI，IA|和一,log一中的多项式。",
"3",
"Related Work",
"存在许多解决MDP的classical方法（请参阅优秀的文本[4]，[2]，[5]和[6]。Thevalue",
"iteration algorithm requires that the entire state space be held in memory as a table; exceedingly",
"large state and action spaces will exhaust the memory available to even large, powerful computers",
"Thus at some level of complexity, value iteration becomes intractable due to the explosion in state",
"action space.一些算法通过利用问题的结构来更紧凑地表示MDP[7]，[8]来解决状态操作",
"3"
] |
{
"arxiv_id": "1805.02785",
"language": "zh",
"page_idx": 2
}
|
||
1805.02785-zh-3
|
1805.02785-zh-3.jpg
|
空间大小的爆炸，这可以提高性能。另一个主要目的是创建使用值函数近似或估计技术的方法，以避免必须在每个状态[9] 的内存中维护一个表。尝试使用神经网络[10]和蒙特卡洛树搜索 (MCTS)[11]扩展到高维状态空间，代表了最近处理状态空间爆炸的尝试，尤其是MCTS 在 Go[12, 13] 游戏中显示出令人鼓舞的结果已经尝试提高价值迭代的性能。异步值迭代是值迭代的一种变体，它在每次迭代期间仅处理某些状态，从而提高了内存使用率，并且在某些情况下可以比值迭代更快地收敛[5]。优先扫描是另一种策略，它通过某种度量对状态的处理进行排序，并在更新后反向传播到已处理状态[14,15]的前身。[16]提供了这方面研究的一个很好的总结。本文特别感兴趣的是，[17]基于[18]中的工作，使用基于 Dijkstra 算法的方法 (也讨论了确定性 MDP) 来检查随机最短路径问题，其中表明确定性 MDP可以简化为 Dijkstra 算法，该算法比值迭代具有一些性能优势。在[19]中，讨论了消除通常需要的优先级队列的方法。特别值得注意的是一个反向值迭代算法，它从一个终止的目标状态计算值迭代，考虑到目标状态可达到的状态的问题，并从那里向后工作。虽然它们确实消除了优先级队列的开销，但保留了 FIFO 队列。他们同样提出了一种前向值迭代算法，该算法考虑了从初始状态可达到的状态并从那里向前工作，他们指出这等效于[20]中的 LAO* 算法在[21] 中已经研究了改变环境的限制，即对转换函数$T$的更改受一些小值$\epsilon$的限制。在[22] 中，在基于遗憾的框架中，允许奖励在时间步骤之间任意变化，该框架依赖于在每个步骤中求解线性程序。[23] 中的环境和奖励都使用强大的动态编程方法而变化，该方法最终也依赖于每个时间步的线性规划。在[24]中，在提供$N$黑匣子“专家”的随机设置中，允许奖励在每个时间步长处任意更改 (可能是对抗性方式);在产生的策略更改上提供了相对于固定范围和预期遗憾的收敛范围，并且相对于$N$专家和$|A|$动作的性能显示为多项式，尽管它本身并不依赖于专家的范围(。在[25]中，将奖励函数分解为更简单的 MDP, 每个 MDP都使用类似于 DQN 的基于神经网络的方法进行求解，并且将更简单的 MDPQ 值函数的结果聚合为最终的全局 Q 值函数，但是该方法不会导致对如何从较小的 MDP 组成值函数有更基本的了解。[26] 中与时间相关的 MDP (TMDP) 用于计算具有连续时间维度的 MDP,根据分段线性时间步长要求精确解，但仍依赖于值迭代来近似真实(精确)解。据作者所知，没有现有的研究以本文所述的方式为这类 (或任何其他类) MDP 提供精确值函数的直接解决方案
4 $\textbf{Methodology}$
在本文中，我们根据图来描述 MDP。稍后将很清楚，我们采用此约定是为了利用将出现的属性来得出一种方法，该方法可以计算具有仅依赖于状态$s$的奖励函数$R(s)$的 MDP 的精确解。
4.1 MDP Transition Graphs
MDP 理论上可以允许通过动作$a$从任何状态$s$转换到任何其他状态$s^\prime$,该动作由转换函数$T(s,a,s^{\prime})$定义。给定$s,a,s^\prime$的零值表示不可能进行转换，否则给出(0,1]的概率，并且状态$s$和$s^\prime$ is defined in this paper as connected 。从任何状态$s$到所有可能的下一个状态$(\bar{\text{包括状态 }s\text{ 本身) 的转换概率必须总计 }1.0\text{ 。可以定义确定性 MDP 的转换图，其中图的每}}$ 个节点是状态$s$ ,并且图的每个边是可能的动作$a$ 。转换图是一个有向图，它可能是循环的。(请注意，这只是过渡函数$T$中包含的信息的图形表示。)有关示例转换图，请参见附录 1。另请注意，对于确定性 MDP, [3] 中提供了类似的描述。对于确定性转换图，距离定义为从给定状态$s_0$移动到所需状态$s_k$所需的动作 (或转换)的最小正数
形式上，假设 MDP 具有$n$状态$\mathcal{S}=\{s_1,s_2,\cdots,s_n\}$ 。在每个状态下，都有$m$操作可供选择：$\mathcal{A}=\{a_1,a_2,\cdots,a_m\}$。在时间$t$ , 状态表示为$s^(t)\in\mathcal{S}$和动作$a^(t)\in\mathcal{A}$。由于此 MDP 是确定性的，因此给定当前状态和当前动作的下一个状态可以表示为$s^{( t+ 1) }= T( s^{( t) }, a^{( t) })$ , 其中$s^{(t)}$ 和 $a^{(t)}$ 是当前状态和当前动作，而映射$T:\mathcal{S}\times\mathcal{A}\to\mathcal{S}$ 是根据当前状态和动作的下$\overset{(1)}{\operatorname*{\operatorname*{\operatorname*{\operatorname*{\rightharpoonup}}}}}$个状态 $s^{(t+1)}$
假设初始状态为$s^(0)$ ,在采取动作$a^(0)$之后，下一个状态为$s^(1)=T(s^(0),a^{(0)})$ 。采取另一个操作$a^(1)$后，第三个状态将是$T(T(s^{(0)},a^{(0)}),a^{(1)})$。为方便起见，我们将其表示为$T(s^{(0)},a^{(0)},a^{(1)})$ 。更一般地说，如果初始状态为$s^{(0)}$ ,则在采取顺序操作$a^{(0)},a^{(1)},a^{(2)},\cdots,a^{(t)}$ 之后，代理将处于状态$T(s^(0),a^{(0)},a^{(1)},a^{(2)},\cdots,a^{(t)})$
4
|
[
[
0.176513671875,
0.0921630859375,
0.82275390625,
0.0921630859375,
0.82275390625,
0.1060791015625,
0.176513671875,
0.1060791015625
],
[
0.17724609375,
0.106689453125,
0.82275390625,
0.106689453125,
0.82275390625,
0.12060546875,
0.17724609375,
0.12060546875
],
[
0.1748046875,
0.12060546875,
0.8203125,
0.12060546875,
0.8203125,
0.1337890625,
0.1748046875,
0.1337890625
],
[
0.176513671875,
0.1337890625,
0.82275390625,
0.1337890625,
0.82275390625,
0.147705078125,
0.176513671875,
0.147705078125
],
[
0.17724609375,
0.147705078125,
0.82177734375,
0.147705078125,
0.82177734375,
0.1610107421875,
0.17724609375,
0.1610107421875
],
[
0.176513671875,
0.1610107421875,
0.82275390625,
0.1610107421875,
0.82275390625,
0.1749267578125,
0.176513671875,
0.1749267578125
],
[
0.176513671875,
0.175537109375,
0.82177734375,
0.175537109375,
0.82177734375,
0.189453125,
0.176513671875,
0.189453125
],
[
0.176513671875,
0.1881103515625,
0.82177734375,
0.1881103515625,
0.82177734375,
0.2020263671875,
0.176513671875,
0.2020263671875
],
[
0.176513671875,
0.20263671875,
0.82177734375,
0.20263671875,
0.82177734375,
0.216552734375,
0.176513671875,
0.216552734375
],
[
0.176513671875,
0.216552734375,
0.8203125,
0.216552734375,
0.8203125,
0.2298583984375,
0.176513671875,
0.2298583984375
],
[
0.17724609375,
0.2298583984375,
0.82177734375,
0.2298583984375,
0.82177734375,
0.2442626953125,
0.17724609375,
0.2442626953125
],
[
0.176513671875,
0.2442626953125,
0.8203125,
0.2442626953125,
0.8203125,
0.25830078125,
0.176513671875,
0.25830078125
],
[
0.17724609375,
0.25830078125,
0.82177734375,
0.25830078125,
0.82177734375,
0.271484375,
0.17724609375,
0.271484375
],
[
0.176513671875,
0.272705078125,
0.81201171875,
0.272705078125,
0.81201171875,
0.28271484375,
0.176513671875,
0.28271484375
],
[
0.176513671875,
0.292236328125,
0.82275390625,
0.292236328125,
0.82275390625,
0.30615234375,
0.176513671875,
0.30615234375
],
[
0.17724609375,
0.30615234375,
0.82177734375,
0.30615234375,
0.82177734375,
0.3193359375,
0.17724609375,
0.3193359375
],
[
0.17724609375,
0.3212890625,
0.82177734375,
0.3212890625,
0.82177734375,
0.33203125,
0.17724609375,
0.33203125
],
[
0.17724609375,
0.333984375,
0.82275390625,
0.333984375,
0.82275390625,
0.347900390625,
0.17724609375,
0.347900390625
],
[
0.176513671875,
0.347900390625,
0.82275390625,
0.347900390625,
0.82275390625,
0.361083984375,
0.176513671875,
0.361083984375
],
[
0.176513671875,
0.361083984375,
0.82177734375,
0.361083984375,
0.82177734375,
0.375,
0.176513671875,
0.375
],
[
0.17724609375,
0.375,
0.82421875,
0.375,
0.82421875,
0.38818359375,
0.17724609375,
0.38818359375
],
[
0.178955078125,
0.39013671875,
0.82177734375,
0.39013671875,
0.82177734375,
0.40087890625,
0.178955078125,
0.40087890625
],
[
0.176513671875,
0.40283203125,
0.82177734375,
0.40283203125,
0.82177734375,
0.416748046875,
0.176513671875,
0.416748046875
],
[
0.17724609375,
0.416748046875,
0.82177734375,
0.416748046875,
0.82177734375,
0.429931640625,
0.17724609375,
0.429931640625
],
[
0.17724609375,
0.43115234375,
0.82275390625,
0.43115234375,
0.82275390625,
0.4443359375,
0.17724609375,
0.4443359375
],
[
0.176513671875,
0.443115234375,
0.82275390625,
0.443115234375,
0.82275390625,
0.45703125,
0.176513671875,
0.45703125
],
[
0.1748046875,
0.458984375,
0.2091064453125,
0.458984375,
0.2091064453125,
0.4716796875,
0.1748046875,
0.4716796875
],
[
0.1707763671875,
0.490478515625,
0.3203125,
0.49365234375,
0.3203125,
0.51123046875,
0.1707763671875,
0.50830078125
],
[
0.1748046875,
0.52978515625,
0.8203125,
0.52978515625,
0.8203125,
0.54345703125,
0.1748046875,
0.54345703125
],
[
0.1748046875,
0.54443359375,
0.82177734375,
0.54345703125,
0.82177734375,
0.55810546875,
0.1748046875,
0.5595703125
],
[
0.176513671875,
0.556640625,
0.2156982421875,
0.5595703125,
0.214111328125,
0.5712890625,
0.1748046875,
0.5693359375
],
[
0.1748046875,
0.6162109375,
0.386474609375,
0.6162109375,
0.386474609375,
0.6298828125,
0.1748046875,
0.6298828125
],
[
0.17724609375,
0.65478515625,
0.82177734375,
0.65478515625,
0.82177734375,
0.66552734375,
0.17724609375,
0.66552734375
],
[
0.17724609375,
0.66748046875,
0.82177734375,
0.66748046875,
0.82177734375,
0.68115234375,
0.17724609375,
0.68115234375
],
[
0.178955078125,
0.68115234375,
0.82275390625,
0.68115234375,
0.82275390625,
0.69580078125,
0.178955078125,
0.69580078125
],
[
0.176513671875,
0.6943359375,
0.82177734375,
0.6943359375,
0.82177734375,
0.70849609375,
0.176513671875,
0.70849609375
],
[
0.17724609375,
0.70849609375,
0.82177734375,
0.70849609375,
0.82177734375,
0.7216796875,
0.17724609375,
0.7216796875
],
[
0.17724609375,
0.7236328125,
0.8203125,
0.7236328125,
0.8203125,
0.734375,
0.17724609375,
0.734375
],
[
0.176513671875,
0.736328125,
0.82275390625,
0.736328125,
0.82275390625,
0.75,
0.176513671875,
0.75
],
[
0.17724609375,
0.75,
0.6953125,
0.75,
0.6953125,
0.76318359375,
0.17724609375,
0.76318359375
],
[
0.176513671875,
0.77099609375,
0.82275390625,
0.77099609375,
0.82275390625,
0.7841796875,
0.176513671875,
0.7841796875
],
[
0.176513671875,
0.78515625,
0.82275390625,
0.78515625,
0.82275390625,
0.7998046875,
0.176513671875,
0.7998046875
],
[
0.1748046875,
0.798828125,
0.82177734375,
0.798828125,
0.82177734375,
0.81640625,
0.1748046875,
0.81640625
],
[
0.1748046875,
0.81494140625,
0.82177734375,
0.81640625,
0.82177734375,
0.83056640625,
0.1748046875,
0.830078125
],
[
0.181396484375,
0.83056640625,
0.28515625,
0.8271484375,
0.285888671875,
0.84130859375,
0.1829833984375,
0.84521484375
],
[
0.176513671875,
0.85107421875,
0.82275390625,
0.85107421875,
0.82275390625,
0.86572265625,
0.176513671875,
0.86572265625
],
[
0.176513671875,
0.8662109375,
0.82177734375,
0.8662109375,
0.82177734375,
0.880859375,
0.176513671875,
0.880859375
],
[
0.176513671875,
0.8818359375,
0.82275390625,
0.8818359375,
0.82275390625,
0.89599609375,
0.176513671875,
0.89599609375
],
[
0.1748046875,
0.896484375,
0.7158203125,
0.89794921875,
0.7158203125,
0.912109375,
0.1748046875,
0.9111328125
],
[
0.491943359375,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94775390625,
0.491943359375,
0.94775390625
]
] |
[
"空间大小的爆炸，这可以提高性能。另一个主要目的是创建使用值函数近似或估计技术的",
"方法，以避免必须在每个状态[9]的内存中维护一个表。尝试使用神经网络[10]和蒙特卡洛",
"树搜索（MCTS）[11]扩展到高维状态空间，代表了最近处理状态空间爆炸的尝试，尤其是",
"MCTS在Go[12,13]游戏中显示出令人鼓舞的结果已经尝试提高价值迭代的性能。异步值迭",
"代是值迭代的一种变体，它在每次迭代期间仅处理某些状态，从而提高了内存使用率，并且",
"在某些情况下可以比值迭代更快地收敛[5]。优先扫描是另一种策略，它通过某种度量对状",
"态的处理进行排序，并在更新后反向传播到已处理状态[14,15]的前身。[16]提供了这方面",
"研究的一个很好的总结。本文特别感兴趣的是，[17]基于[18]中的工作，使用基于Dijkstra",
"算法的方法（也讨论了确定性MDP）来检查随机最短路径问题，其中表明确定性MDP可以",
"简化为Dijkstra算法，该算法比值迭代具有一些性能优势。在[19]中，讨论了消除通常需要",
"的优先级队列的方法。特别值得注意的是一个反向值迭代算法，它从一个终止的目标状态计",
"算值迭代，考虑到目标状态可达到的状态的问题，并从那里向后工作。虽然它们确实消除了",
"优先级队列的开销，但保留了FIFO队列。他们同样提出了一种前向值迭代算法，该算法考",
"虑了从初始状态可达到的状态并从那里向前工作他们指出这等效于[20I中的LAO*算法",
"在[21]中已经研究了改变环境的限制，即对转换函数T的更改受一些小值e的限制。在[22]",
"中，在基于遗憾的框架中，允许奖励在时间步骤之间任意变化，该框架依赖于在每个步骤中",
"求解线性程序。[231中的环境和奖励都使用强大的动态编程方法而变化，该方法最终也依赖",
"于每个时间步的线性规划。在[24]中，在提供N黑匣子“专家”的随机设置中，允许奖励",
"在每个时间步长处任意更改（可能是对抗性方式）；在产生的策略更改上提供了相对于固定",
"范围和预期遗撼的收敛范围，并且相对于N专家和A动作的性能显示为多项式，尽管它",
"本身并不依赖于专家的范围（。在[25]中，将奖励函数分解为更简单的MDP，每个MDP都",
"使用类似于DON的基于神经网络的方法进行求解，并且将更简单的MDPQ值函数的结果",
"聚合为最终的全局Q值函数，但是该方法不会导致对如何从较小的MDP组成值函数有更基",
"本的了解。[26]中与时间相关的MDP（TMDP）用于计算具有连续时间维度的MDP，根据",
"分段线性时间步长要求精确解，但仍依赖于值迭代来近似真实（精确）解。据作者所知，没",
"有现有的研究以本文所述的方式为这类（或任何其他类）MDP提供精确值函数的直接解决",
"方案",
"4Methodology",
"在本文中，我们根据图来描述MDP。稍后将很清楚，我们采用此约定是为了利用将出现的",
"属性来得出一种方法，该方法可以计算具有仅依赖于状态s的奖励函数R(s)的MDP的精",
"确解。",
"4.1 MDP Transition Graphs",
"MDP理论上可以允许通过动作a从任何状态s转换到任何其他状态s'，该动作由转换函",
"数T(s,a,s）定义。给定s，α，s的零值表示不可能进行转换，否则给出(O,1)的概率，并",
"且状态s和sisdefined inthispaperas connected。从任何状态s到所有可能的下一个状态",
"（包括状态s本身）的转换概率必须总计1.0。可以定义确定性MDP的转换图，其中图的每",
"个节点是状态s，并且图的每个边是可能的动作α。转换图是一个有向图，它可能是循环",
"的。（请注意，这只是过渡函数T中包含的信息的图形表示。）有关示例转换图，请参见附",
"录1。另请注意，对于确定性MDP，[3]中提供了类似的描述。对于确定性转换图，距离定",
"义为从给定状态so移动到所需状态Sk所需的动作（或转换）的最小正数",
"形式上，假设MDP具有n状态S=(s1,S2，，sn）。在每个状态下，都有m操作可供选",
"择：A=a1,a2，.,am)。在时间t，状态表示为s(t)ES和动作a(t)EA。由于此MDP",
"是确定性的，因此给定当前状态和当前动作的下一个状态可以表示为s(t+1)=T(s(t),a(t))",
"其中s(t)和α(t)是当前状态和当前动作，而映射T：S×A→S是根据当前状态和动作的下",
"一个状态s(t+1)",
"假设初始状态为s(0)，在采取动作α(0)之后，下一个状态为 s(1)=T(s(0),α()）。采",
"取另一个操作α(1)后，第三个状态将是T(T(s(0),a(0)),α(1)）。为方便起见，我们将",
"其表示为T(s(0),a(0),a(1)）。更一般地说，如果初始状态为s(0)，则在采取顺序操作",
"a(0),a(1),a(2),...,a(t)之后，代理将处于状态T(s(0),a(0),a(1),a(2),.,a(t)",
"4"
] |
{
"arxiv_id": "1805.02785",
"language": "zh",
"page_idx": 3
}
|
||
1805.02785-zh-4
|
1805.02785-zh-4.jpg
|
$\textbf{Definition 1. For a deterministic MDP with finite states, if from state s }_{i},after$ taking finite actions,
the agent can reach state s $_k,then$ the distance from s to s is defined as:
$$\delta(s_i,s_k)=\min_t\{t|T(s_i,a^{(1)},a^{(2)},\cdots,a^{(t)})=s_k\}$$
(3)
$Iffromstates_{i},nomatterwhatsequenceofactionstheagenttakes,itcannotreachstates_{k},then$
the distance from si to sk is defined to be:
$$\delta(s_i,s_k)=\infty $$
$$e\:to\:itself\:\delta(s,s)=0\:for\:any\:s\in\mathcal{S}\:.$$
(4)
请注意，对于二维网格世界 MDP, 从一个状态到另一个状态的距离只是相对于网格单元的
笛卡尔坐标的曼哈顿距离
$\textbf{Definition 2. An MDP problem is said to be a fully connected MDP if all states can be reached from}$
$all\textit{ other states: }\forall s, s^{\prime }\in S, \delta ( s, s^{\prime }) < \infty$ .
通过完全连接的 MDP 的定义，我们希望检查 MDP,在这些 MDP 中，代理可以从任何给定
的初始状态到达任何状态(也就是说，所有状态都可能通过某些操作集可到达。)
4.2 Exact Solutions for Single Reward Sources
给定 MDP transition graph 的定义，我们现在描述单个奖励源的确定性连续 MDP 的确切解
决方案。我们首先定义发生在连续 MDP 中的周期的概念，并推导出精确的值函数
Definition 3. The cycle of a state s, which denoted as C(s) , is an ordered sequence of states: $s^{(1)},s^{(2)},\cdots,s^{(t)}$ where the states in this sequence satisfy the following condition: There exists a $sequence~of~action~a^{(1)},a^{(2)},\cdots,a^{(t+1)}~such~that:$
$$\begin{array}{c}T(s,a^{(1)})=s^{(1)}\\T(s,a^{(1)},a^{(2)})=s^{(2)}\end{array}$$
$$T(s,a^{(1)},a^{(2)},\cdots,a^{(t)},a^{(t+1)})=s$$
The length ofthe cycle, d(C(s)) is the number of actions in the sequence $( t+ 1) \textit{that causes a return}$
$tos.$
请注意，如果状态$s$具有多个周期，则始终存在一个具有有限距离的周期。如果存在一个动作$a\in\mathcal{A}$ ,例如$T(s,a)=s$ ,这也是一个距离为 1 的循环。请注意，状态$s$可能没有周期。还要注意，状态$s$可以具有多个周期，并且这些周期中的状态不需要是不同的。(给定周期的某些状态可能与其他周期共享。)
Definition 4. Suppose a state s has p cycles $\mathcal{C}^1(s),\cdots,\mathcal{C}^p(s)$, where $p$ can be finite or infinite.
The minimum cycle of state s , which denoted as C' (s) , is a cycle with minimum distance:
$$\mathcal{C}^*=\{\mathcal{C}^i|d(\mathcal{C}^i)\leq d(\mathcal{C}^j),\forall j\in\{1,\cdots,p\}\}$$
(5)
请注意，当且仅当状态$s$没有周期时，状态$s$才能没有最小周期。当所有周期中有多个周期
具有相同的最小距离时，状态$s$也可以有多个最小周期
我们将状态$s$的最小周期的距离表示为$\phi(s)$
我们现在描述如何计算给定最小周期定义的值函数
Theorem 1. Every deterministic continuous fully connected MDP with a single reward source has
$a\textit{ minimum cycle}.$
附录 3 中提供了证明
Theorem 2. For the deterministic continuous fully connected MDP model defined above, suppose
it has reward function as follows:
(6)
$$r(s)=\left\{\begin{matrix}r_g>0&if\:s=s_g\\0&otherwise.\end{matrix}\right.$$
5
|
[
[
0.1748046875,
0.091552734375,
0.82177734375,
0.0921630859375,
0.82177734375,
0.106689453125,
0.1748046875,
0.1060791015625
],
[
0.1748046875,
0.10797119140625,
0.65869140625,
0.10797119140625,
0.65869140625,
0.121826171875,
0.1748046875,
0.121826171875
],
[
0.328369140625,
0.1256103515625,
0.6708984375,
0.1256103515625,
0.6708984375,
0.146484375,
0.328369140625,
0.146484375
],
[
0.8037109375,
0.1300048828125,
0.826171875,
0.1300048828125,
0.826171875,
0.14453125,
0.8037109375,
0.14453125
],
[
0.176513671875,
0.1558837890625,
0.82275390625,
0.1558837890625,
0.82275390625,
0.169189453125,
0.176513671875,
0.169189453125
],
[
0.1748046875,
0.169189453125,
0.450927734375,
0.169189453125,
0.450927734375,
0.18310546875,
0.1748046875,
0.18310546875
],
[
0.453369140625,
0.2052001953125,
0.5458984375,
0.2052001953125,
0.5458984375,
0.216552734375,
0.453369140625,
0.216552734375
],
[
0.80224609375,
0.203857421875,
0.826171875,
0.203857421875,
0.826171875,
0.2183837890625,
0.80224609375,
0.2183837890625
],
[
0.1748046875,
0.2215576171875,
0.68115234375,
0.2215576171875,
0.68115234375,
0.23486328125,
0.1748046875,
0.23486328125
],
[
0.176513671875,
0.24560546875,
0.82177734375,
0.24560546875,
0.82177734375,
0.2587890625,
0.176513671875,
0.2587890625
],
[
0.176513671875,
0.260009765625,
0.35693359375,
0.260009765625,
0.35693359375,
0.2734375,
0.176513671875,
0.2734375
],
[
0.1748046875,
0.27783203125,
0.82275390625,
0.27783203125,
0.82275390625,
0.292236328125,
0.1748046875,
0.292236328125
],
[
0.1748046875,
0.292236328125,
0.44775390625,
0.292236328125,
0.44775390625,
0.30615234375,
0.1748046875,
0.30615234375
],
[
0.176513671875,
0.31640625,
0.82275390625,
0.31640625,
0.82275390625,
0.330078125,
0.176513671875,
0.330078125
],
[
0.1748046875,
0.330810546875,
0.75341796875,
0.330810546875,
0.75341796875,
0.3447265625,
0.1748046875,
0.3447265625
],
[
0.1732177734375,
0.359130859375,
0.51318359375,
0.35986328125,
0.51318359375,
0.375,
0.1732177734375,
0.373779296875
],
[
0.176513671875,
0.38818359375,
0.82275390625,
0.38818359375,
0.82275390625,
0.401611328125,
0.176513671875,
0.401611328125
],
[
0.1748046875,
0.40087890625,
0.75634765625,
0.40087890625,
0.75634765625,
0.415283203125,
0.1748046875,
0.415283203125
],
[
0.1748046875,
0.419189453125,
0.8203125,
0.419189453125,
0.8203125,
0.43310546875,
0.1748046875,
0.43310546875
],
[
0.1748046875,
0.43310546875,
0.82177734375,
0.43505859375,
0.82177734375,
0.449462890625,
0.1748046875,
0.447509765625
],
[
0.176513671875,
0.448974609375,
0.51123046875,
0.448974609375,
0.51123046875,
0.46337890625,
0.176513671875,
0.46337890625
],
[
0.439453125,
0.468505859375,
0.55615234375,
0.466552734375,
0.55615234375,
0.484130859375,
0.439453125,
0.486083984375
],
[
0.423095703125,
0.4892578125,
0.57177734375,
0.4873046875,
0.57177734375,
0.5048828125,
0.423095703125,
0.5068359375
],
[
0.375732421875,
0.52783203125,
0.615234375,
0.5263671875,
0.615234375,
0.54443359375,
0.375732421875,
0.54541015625
],
[
0.1748046875,
0.5517578125,
0.82275390625,
0.5517578125,
0.82275390625,
0.56494140625,
0.1748046875,
0.56494140625
],
[
0.17724609375,
0.568359375,
0.200927734375,
0.568359375,
0.200927734375,
0.5771484375,
0.17724609375,
0.5771484375
],
[
0.178955078125,
0.59228515625,
0.8203125,
0.59228515625,
0.8203125,
0.60302734375,
0.178955078125,
0.60302734375
],
[
0.1748046875,
0.60302734375,
0.82177734375,
0.60400390625,
0.82177734375,
0.61865234375,
0.1748046875,
0.6171875
],
[
0.1748046875,
0.6162109375,
0.82275390625,
0.6171875,
0.82275390625,
0.6318359375,
0.1748046875,
0.630859375
],
[
0.176513671875,
0.6318359375,
0.428955078125,
0.6318359375,
0.428955078125,
0.64599609375,
0.176513671875,
0.64599609375
],
[
0.1748046875,
0.6484375,
0.814453125,
0.6494140625,
0.814453125,
0.6640625,
0.1748046875,
0.66357421875
],
[
0.1748046875,
0.6640625,
0.76318359375,
0.6640625,
0.76318359375,
0.67822265625,
0.1748046875,
0.67822265625
],
[
0.35302734375,
0.68994140625,
0.640625,
0.69140625,
0.640625,
0.70556640625,
0.35302734375,
0.705078125
],
[
0.8037109375,
0.69140625,
0.82421875,
0.69140625,
0.82421875,
0.70556640625,
0.8037109375,
0.70556640625
],
[
0.176513671875,
0.71630859375,
0.82275390625,
0.71630859375,
0.82275390625,
0.72998046875,
0.176513671875,
0.72998046875
],
[
0.176513671875,
0.72998046875,
0.56787109375,
0.72998046875,
0.56787109375,
0.74365234375,
0.176513671875,
0.74365234375
],
[
0.176513671875,
0.75048828125,
0.486083984375,
0.75048828125,
0.486083984375,
0.7646484375,
0.176513671875,
0.7646484375
],
[
0.17724609375,
0.77294921875,
0.53173828125,
0.77294921875,
0.53173828125,
0.78271484375,
0.17724609375,
0.78271484375
],
[
0.1748046875,
0.7890625,
0.82177734375,
0.7890625,
0.82177734375,
0.80322265625,
0.1748046875,
0.80322265625
],
[
0.1748046875,
0.8037109375,
0.289306640625,
0.8037109375,
0.289306640625,
0.8173828125,
0.1748046875,
0.8173828125
],
[
0.1748046875,
0.82763671875,
0.3203125,
0.82763671875,
0.3203125,
0.84130859375,
0.1748046875,
0.84130859375
],
[
0.1748046875,
0.84521484375,
0.82275390625,
0.8466796875,
0.82275390625,
0.861328125,
0.1748046875,
0.85986328125
],
[
0.1748046875,
0.85986328125,
0.393798828125,
0.85986328125,
0.393798828125,
0.87353515625,
0.1748046875,
0.87353515625
],
[
0.451904296875,
0.87890625,
0.58642578125,
0.880859375,
0.58642578125,
0.8994140625,
0.451904296875,
0.89794921875
],
[
0.404296875,
0.892578125,
0.44775390625,
0.892578125,
0.44775390625,
0.90380859375,
0.404296875,
0.90380859375
],
[
0.80224609375,
0.89013671875,
0.82421875,
0.89013671875,
0.82421875,
0.90478515625,
0.80224609375,
0.90478515625
],
[
0.451904296875,
0.8984375,
0.47802734375,
0.8984375,
0.47802734375,
0.912109375,
0.451904296875,
0.912109375
],
[
0.525390625,
0.8984375,
0.59326171875,
0.8984375,
0.59326171875,
0.91357421875,
0.525390625,
0.91357421875
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Definition 1.For a deterministic MDP with finite states, if from state si,after taking finite actions",
"the agent can reach state Sk , then the distance from Si to Sk is defined as:",
"8(ss, sk) =min(tT(si,a(1),a(2), ,a(t) = 8h)",
"(3)",
"If from state si, no matter what sequence of actions the agent takes, it cannot reach state sk , then",
"the distance from si to sh is defined to be:",
"o(si, Sk) = 00",
"(4)",
"Finally, we define the distance from a state to itself o(s, s) = O for any s E S",
"请注意，对于二维网格世界MDP，从一个状态到另一个状态的距离只是相对于网格单元的",
"笛卡尔坐标的曼哈顿距离",
"Definition 2. An MDP problem is said to be a fully connected MDP if all states can be reached from",
"all other states: Vs, s' E S, d(s, s') < oo",
"通过完全连接的MDP的定义，我们希望检查MDP，在这些MDP中，代理可以从任何给定",
"的初始状态到达任何状态（也就是说，所有状态都可能通过某些操作集可到达。）",
"4.2ExactSolutionsforSingleRewardSources",
"给定MDPtransitiongraph的定义，我们现在描述单个奖励源的确定性连续MDP的确切解",
"决方案。我们首先定义发生在连续MDP中的周期的概念，并推导出精确的值函数",
"Definition 3. The cycle of a state s , which denoted as C(s) , is an ordered sequence of states",
"s(1), s(2), ..., s(t) where the states in this sequence satisfy the following condition: There exists a",
"sequence of action a(1),a(2),. ,a(t+1) such that:",
"T(s,a(1) = s(1)",
"T(s,a(1),a(2)) = s(2)",
"T(s,a(1),a(2), ,a(t),a(t+1) =",
"The length of the cycle, d(C(s)) is the number of actions in the sequence (t+1 ) that causes a return",
"tos",
"请注意，如果状态s具有多个周期，则始终存在一个具有有限距离的周期。如果存在一个动",
"作aEA，例如T(s,a）=s，这也是一个距离为1的循环。请注意，状态s可能没有周期",
"还要注意，状态s可以具有多个周期，并且这些周期中的状态不需要是不同的。（给定周期",
"的某些状态可能与其他周期共享。）",
"Definition 4. Suppose a state s has p cycles Cl(s), .:, CP(s) , where p can be finite or infinite.",
"The minimum cycle of state s, which denoted as C*(s), is a cycle with minimum distance.",
"C* = [ci]d(ci) ≤ d(c)), Vj E (1, .*,p])",
"(5)",
"请注意，当且仅当状态s没有周期时，状态s才能没有最小周期。当所有周期中有多个周期",
"具有相同的最小距离时，状态s也可以有多个最小周期",
"我们将状态s的最小周期的距离表示为Φ(s)",
"我们现在描述如何计算给定最小周期定义的值函数",
"Theorem 1. Every deterministic continuous fully connected MDP with a single reward source has",
"a minimum cycle.",
"附录3中提供了证明",
"Theorem 2. For the deterministic continuous fully connected MDP model defined above, suppose",
"it has reward function as follows:",
"Jrg >0 ifs=sg",
"r(s) =",
"(6)",
"10",
"otherwise.",
"5"
] |
{
"arxiv_id": "1805.02785",
"language": "zh",
"page_idx": 4
}
|
||
2305.10699-zh-0
|
2305.10699-zh-0.jpg
|
Dirichlet Diffusion Score Model for Biological Sequence Generation
Pavel Avdeyev$^1$Chenlai Shi$^1$Yuhao Tan$^1$Kseniia Dudnyk$^1$Jian Zhou$^1$
Abstract
连接进行似然评估以及采样方法的灵活性。然而，最初
。3in
提出的生成式 SDE 并不直接适用于离散数据建模。最近的工作提出了使扩散模型适应离散数据 (Campbell et al., 2022; Chen et al., 2022b; Sun et al., 2022; Austin et al., 2021; Hoogeboom et al., 2021b;a) 的方法，包括离散空间 (Campbell et al., 2022) 中的连续时间扩散， 但是除了基于量化的方法外，没有在连续时间 SDE 扩散框架 (Song et al.,2020) 内制定任何方法。在本手稿中，我们提出了一种将这种方法扩展到离散数据的一般机制，同时允许概率单纯形空间中的连续时间扩散。具体来说，为了利用 Dirichlet 分布与离散数据之间的自然联系，我们考虑了平稳分布为 Dirichlet 分布的概率单纯形内的连续时间扩散。离散数据的正向扩散(数据到噪声)从概率单纯形空间的顶点开始，并在同一空间中连续扩散，并且可以通过潜在变量解释来桥接连续-离散空间间隙
虽然我们的预期应用是在生物序列生成中，但我们在一系列离散数据生成任务上评估了我们的 Dirichlet 扩$散得分模型(DDSM)^{1},以更好地了解其性能。除$ 了在小型基准数据集上展示有竞争力的性能之外，我们还将其应用于生成数独谜题，以测试其生成具有强约束的高度结构化数据的能力。该模型不仅可以生成， 还可以解决包括硬谜题在内的数独谜题，这是第一次通过纯生成建模方法实现。最后，我们将 DDSM 应用于生物序列生成的实际应用。具体来说，我们开发了第一个用于设计驱动基因表达的人类启动子 DNA 序列的模型，并证明它设计了与人类基因组启动子序列相当的不同序列
Designing biological sequences is an important challenge that requires satisfying complex constraints and thus is a natural problem to address with deep generative modeling. Diffusion generative models have achieved considerable success in many applications. Score-based generative stochastic differential equations (SDE) model is a continuous-time diffusion model framework that enjoys many benefits, but the originally proposed SDEs are not naturally designed for modeling discrete data. To develop generative SDE models for discrete data such as biological sequences, here we introduce a diffusion process defined in the probability simplex space with stationary distribution being the Dirichlet distribution. This makes diffusion in continuous space natural for modeling discrete data. We refer to this approach as Dirchlet diffusion score model. We demonstrate that this technique can generate samples that satisfy hard constraints using a Sudoku generation task. This generative model can also solve Sudoku, including hard puzzles, without additional training. Finally, we applied this approach to develop the first human promoter DNA sequence design model and showed that designed sequences share similar properties with natural promoter sequences.
## 2. Background
$2.1.{\mathrm{~Score-based~Generative~Modeling~with~SDE}}$
ItÔ 扩散过程定义为
1. Introduction
$$\mathrm d\boldsymbol{x}=\boldsymbol{f}(\boldsymbol{x},t)\mathrm dt+\boldsymbol{G}(\boldsymbol{x},t)\mathrm d\boldsymbol{w},$$
扩散概率模型是一系列模型，这些模型进行反向扩散处理以从噪声中生成数据。基于分数的生成随机微分方程 (SDE) 是一种连续时间扩散模型，具有许多理想的属性，例如允许通过与概率流常微分方程(ODE)的
${\frac{1}{\mathrm{Lyda~Hill~Department~of~Bioinformatics,~University}}}$ of Texas Southwestern Medical Center, USA . Correspondence to: Jian Zhou<jian.zhou@utsouthwestern.edu>.
Proceedings of the $40^{th}$ International Conference on Machine Learning, Honolulu, Hawaii, USA. PMLR 202, 2023. Copyright 2023 by the author(s).
,其中$\boldsymbol{w}$是标准的维纳过程(a。k。a。,布朗运动),漂移系数$f(\cdot,t):\mathbb{R}^n\to\mathbb{R}^n$和扩散系数$G(\cdot,t):\mathbb{R}^n\to\mathbb{R}^n$ 是$x^t$的向量值函数。Song et al. 2020 利用了 Anderson 1982的显着结果，即可以通过以下 SDE 获得该扩散
$^{1}$Code available at https://github.com/jzhoulab/ddsm
1
|
[
[
0.148681640625,
0.11236572265625,
0.82275390625,
0.11297607421875,
0.82275390625,
0.1307373046875,
0.148681640625,
0.1300048828125
],
[
0.2099609375,
0.1787109375,
0.76123046875,
0.1787109375,
0.76123046875,
0.1925048828125,
0.2099609375,
0.1925048828125
],
[
0.0882568359375,
0.1944580078125,
0.13232421875,
0.1944580078125,
0.13232421875,
0.210205078125,
0.0882568359375,
0.210205078125
],
[
0.2442626953125,
0.2083740234375,
0.3203125,
0.2083740234375,
0.3203125,
0.222900390625,
0.2442626953125,
0.222900390625
],
[
0.5009765625,
0.2083740234375,
0.8837890625,
0.208984375,
0.8837890625,
0.2235107421875,
0.5009765625,
0.222900390625
],
[
0.50244140625,
0.2247314453125,
0.8857421875,
0.2247314453125,
0.8857421875,
0.238037109375,
0.50244140625,
0.238037109375
],
[
0.1217041015625,
0.236083984375,
0.441162109375,
0.236083984375,
0.441162109375,
0.25,
0.1217041015625,
0.25
],
[
0.5009765625,
0.2392578125,
0.8857421875,
0.2392578125,
0.8857421875,
0.25244140625,
0.5009765625,
0.25244140625
],
[
0.1217041015625,
0.250732421875,
0.441162109375,
0.250732421875,
0.441162109375,
0.264404296875,
0.1217041015625,
0.264404296875
],
[
0.50244140625,
0.255615234375,
0.88232421875,
0.255615234375,
0.88232421875,
0.266357421875,
0.50244140625,
0.266357421875
],
[
0.1217041015625,
0.264404296875,
0.439453125,
0.264404296875,
0.439453125,
0.27783203125,
0.1217041015625,
0.27783203125
],
[
0.5009765625,
0.26953125,
0.8857421875,
0.26953125,
0.8857421875,
0.28271484375,
0.5009765625,
0.28271484375
],
[
0.12091064453125,
0.279052734375,
0.44287109375,
0.27978515625,
0.44287109375,
0.294189453125,
0.12091064453125,
0.29345703125
],
[
0.5009765625,
0.2841796875,
0.88134765625,
0.2841796875,
0.88134765625,
0.29736328125,
0.5009765625,
0.29736328125
],
[
0.1217041015625,
0.296630859375,
0.181396484375,
0.296630859375,
0.181396484375,
0.30810546875,
0.1217041015625,
0.30810546875
],
[
0.1829833984375,
0.29541015625,
0.441162109375,
0.29541015625,
0.441162109375,
0.309326171875,
0.1829833984375,
0.309326171875
],
[
0.50244140625,
0.2998046875,
0.88232421875,
0.2998046875,
0.88232421875,
0.313232421875,
0.50244140625,
0.313232421875
],
[
0.12091064453125,
0.309326171875,
0.441162109375,
0.31005859375,
0.441162109375,
0.324462890625,
0.12091064453125,
0.323974609375
],
[
0.5009765625,
0.314453125,
0.88232421875,
0.314453125,
0.88232421875,
0.32763671875,
0.5009765625,
0.32763671875
],
[
0.1217041015625,
0.32568359375,
0.441162109375,
0.32568359375,
0.441162109375,
0.339111328125,
0.1217041015625,
0.339111328125
],
[
0.50244140625,
0.330078125,
0.8837890625,
0.330078125,
0.8837890625,
0.343505859375,
0.50244140625,
0.343505859375
],
[
0.1217041015625,
0.341552734375,
0.441162109375,
0.341552734375,
0.441162109375,
0.354736328125,
0.1217041015625,
0.354736328125
],
[
0.5146484375,
0.34521484375,
0.88134765625,
0.34521484375,
0.88134765625,
0.359130859375,
0.5146484375,
0.359130859375
],
[
0.1217041015625,
0.354736328125,
0.441162109375,
0.354736328125,
0.441162109375,
0.36865234375,
0.1217041015625,
0.36865234375
],
[
0.50244140625,
0.35986328125,
0.88232421875,
0.35986328125,
0.88232421875,
0.373779296875,
0.50244140625,
0.373779296875
],
[
0.1217041015625,
0.371826171875,
0.441162109375,
0.371826171875,
0.441162109375,
0.385009765625,
0.1217041015625,
0.385009765625
],
[
0.50244140625,
0.375732421875,
0.8837890625,
0.375732421875,
0.8837890625,
0.389404296875,
0.50244140625,
0.389404296875
],
[
0.12091064453125,
0.386474609375,
0.441162109375,
0.385009765625,
0.441162109375,
0.399658203125,
0.12091064453125,
0.40087890625
],
[
0.50341796875,
0.392578125,
0.88134765625,
0.392578125,
0.88134765625,
0.40283203125,
0.50341796875,
0.40283203125
],
[
0.1217041015625,
0.401611328125,
0.441162109375,
0.401611328125,
0.441162109375,
0.415283203125,
0.1217041015625,
0.415283203125
],
[
0.5009765625,
0.40478515625,
0.8857421875,
0.404052734375,
0.8857421875,
0.41845703125,
0.5009765625,
0.419189453125
],
[
0.1217041015625,
0.416748046875,
0.441162109375,
0.416748046875,
0.441162109375,
0.429931640625,
0.1217041015625,
0.429931640625
],
[
0.50244140625,
0.42041015625,
0.88623046875,
0.42041015625,
0.88623046875,
0.434326171875,
0.50244140625,
0.434326171875
],
[
0.1217041015625,
0.43115234375,
0.44287109375,
0.43115234375,
0.44287109375,
0.4443359375,
0.1217041015625,
0.4443359375
],
[
0.5009765625,
0.434326171875,
0.68701171875,
0.43505859375,
0.68701171875,
0.449462890625,
0.5009765625,
0.448974609375
],
[
0.1217041015625,
0.4462890625,
0.441162109375,
0.4462890625,
0.441162109375,
0.460205078125,
0.1217041015625,
0.460205078125
],
[
0.12091064453125,
0.460205078125,
0.441162109375,
0.4609375,
0.441162109375,
0.47607421875,
0.12091064453125,
0.474853515625
],
[
0.5009765625,
0.45703125,
0.8857421875,
0.458251953125,
0.8857421875,
0.472900390625,
0.5009765625,
0.4716796875
],
[
0.1217041015625,
0.4765625,
0.441162109375,
0.4765625,
0.441162109375,
0.490478515625,
0.1217041015625,
0.490478515625
],
[
0.51123046875,
0.473388671875,
0.88232421875,
0.472900390625,
0.88232421875,
0.486083984375,
0.51123046875,
0.4873046875
],
[
0.1217041015625,
0.492431640625,
0.441162109375,
0.492431640625,
0.441162109375,
0.505859375,
0.1217041015625,
0.505859375
],
[
0.5009765625,
0.4873046875,
0.8857421875,
0.488037109375,
0.8857421875,
0.50244140625,
0.5009765625,
0.501953125
],
[
0.12091064453125,
0.5068359375,
0.439453125,
0.5068359375,
0.439453125,
0.52099609375,
0.12091064453125,
0.52099609375
],
[
0.50341796875,
0.50390625,
0.8857421875,
0.50390625,
0.8857421875,
0.517578125,
0.50341796875,
0.517578125
],
[
0.12091064453125,
0.521484375,
0.441162109375,
0.521484375,
0.441162109375,
0.5361328125,
0.12091064453125,
0.5361328125
],
[
0.50244140625,
0.51806640625,
0.8837890625,
0.51806640625,
0.8837890625,
0.5322265625,
0.50244140625,
0.5322265625
],
[
0.12091064453125,
0.537109375,
0.441162109375,
0.537109375,
0.441162109375,
0.55029296875,
0.12091064453125,
0.55029296875
],
[
0.50244140625,
0.5341796875,
0.88232421875,
0.5341796875,
0.88232421875,
0.54736328125,
0.50244140625,
0.54736328125
],
[
0.1217041015625,
0.55322265625,
0.441162109375,
0.55322265625,
0.441162109375,
0.56640625,
0.1217041015625,
0.56640625
],
[
0.5009765625,
0.548828125,
0.88134765625,
0.548828125,
0.88134765625,
0.5625,
0.5009765625,
0.5625
],
[
0.12091064453125,
0.5673828125,
0.441162109375,
0.5673828125,
0.441162109375,
0.58056640625,
0.12091064453125,
0.58056640625
],
[
0.50244140625,
0.56298828125,
0.8837890625,
0.56298828125,
0.8837890625,
0.57763671875,
0.50244140625,
0.57763671875
],
[
0.1217041015625,
0.58349609375,
0.439453125,
0.58349609375,
0.439453125,
0.5966796875,
0.1217041015625,
0.5966796875
],
[
0.5009765625,
0.57763671875,
0.88134765625,
0.57763671875,
0.88134765625,
0.59130859375,
0.5009765625,
0.59130859375
],
[
0.12091064453125,
0.59765625,
0.439453125,
0.59765625,
0.439453125,
0.611328125,
0.12091064453125,
0.611328125
],
[
0.50244140625,
0.59521484375,
0.8837890625,
0.59521484375,
0.8837890625,
0.60595703125,
0.50244140625,
0.60595703125
],
[
0.12091064453125,
0.61279296875,
0.441162109375,
0.61279296875,
0.441162109375,
0.626953125,
0.12091064453125,
0.626953125
],
[
0.50244140625,
0.609375,
0.8857421875,
0.609375,
0.8857421875,
0.62255859375,
0.50244140625,
0.62255859375
],
[
0.1217041015625,
0.62744140625,
0.441162109375,
0.62744140625,
0.441162109375,
0.6416015625,
0.1217041015625,
0.6416015625
],
[
0.50244140625,
0.625,
0.61669921875,
0.625,
0.61669921875,
0.63818359375,
0.50244140625,
0.63818359375
],
[
0.1192626953125,
0.6416015625,
0.441162109375,
0.642578125,
0.441162109375,
0.6572265625,
0.1192626953125,
0.65576171875
],
[
0.1217041015625,
0.66015625,
0.439453125,
0.66015625,
0.439453125,
0.67041015625,
0.1217041015625,
0.67041015625
],
[
0.5009765625,
0.6591796875,
0.63134765625,
0.6591796875,
0.63134765625,
0.673828125,
0.5009765625,
0.673828125
],
[
0.1217041015625,
0.67529296875,
0.181396484375,
0.67529296875,
0.181396484375,
0.6875,
0.1217041015625,
0.6875
],
[
0.5009765625,
0.68505859375,
0.8505859375,
0.68505859375,
0.8505859375,
0.69873046875,
0.5009765625,
0.69873046875
],
[
0.5,
0.70556640625,
0.64794921875,
0.70849609375,
0.64794921875,
0.72265625,
0.5,
0.72021484375
],
[
0.0882568359375,
0.72021484375,
0.22216796875,
0.72265625,
0.22216796875,
0.7373046875,
0.0882568359375,
0.73486328125
],
[
0.08905029296875,
0.7470703125,
0.473876953125,
0.7470703125,
0.473876953125,
0.76123046875,
0.08905029296875,
0.76123046875
],
[
0.0906982421875,
0.76220703125,
0.47216796875,
0.76220703125,
0.47216796875,
0.77587890625,
0.0906982421875,
0.77587890625
],
[
0.5908203125,
0.7666015625,
0.7939453125,
0.7666015625,
0.7939453125,
0.77978515625,
0.5908203125,
0.77978515625
],
[
0.09234619140625,
0.77880859375,
0.471435546875,
0.77880859375,
0.471435546875,
0.7890625,
0.09234619140625,
0.7890625
],
[
0.0906982421875,
0.79345703125,
0.473876953125,
0.79345703125,
0.473876953125,
0.806640625,
0.0906982421875,
0.806640625
],
[
0.1094970703125,
0.81494140625,
0.473876953125,
0.81640625,
0.473876953125,
0.83056640625,
0.1094970703125,
0.830078125
],
[
0.0882568359375,
0.82763671875,
0.473876953125,
0.830078125,
0.473876953125,
0.84326171875,
0.0882568359375,
0.84130859375
],
[
0.50732421875,
0.82470703125,
0.8837890625,
0.82470703125,
0.8837890625,
0.83544921875,
0.50732421875,
0.83544921875
],
[
0.08905029296875,
0.84228515625,
0.465576171875,
0.84228515625,
0.465576171875,
0.85595703125,
0.08905029296875,
0.85595703125
],
[
0.50244140625,
0.83837890625,
0.88232421875,
0.83837890625,
0.88232421875,
0.8515625,
0.50244140625,
0.8515625
],
[
0.5009765625,
0.85302734375,
0.8837890625,
0.85400390625,
0.8837890625,
0.8671875,
0.5009765625,
0.8662109375
],
[
0.0882568359375,
0.8662109375,
0.473876953125,
0.86572265625,
0.473876953125,
0.8798828125,
0.0882568359375,
0.880859375
],
[
0.50244140625,
0.86865234375,
0.8857421875,
0.86865234375,
0.8857421875,
0.8818359375,
0.50244140625,
0.8818359375
],
[
0.08905029296875,
0.8798828125,
0.47216796875,
0.8798828125,
0.47216796875,
0.89306640625,
0.08905029296875,
0.89306640625
],
[
0.08660888671875,
0.8916015625,
0.3095703125,
0.892578125,
0.3095703125,
0.9072265625,
0.08660888671875,
0.90576171875
],
[
0.521484375,
0.89013671875,
0.88134765625,
0.8916015625,
0.88134765625,
0.90576171875,
0.521484375,
0.90478515625
],
[
0.48193359375,
0.9248046875,
0.494384765625,
0.9248046875,
0.494384765625,
0.93505859375,
0.48193359375,
0.93505859375
]
] |
[
"Dirichlet Diffusion Score Model for Biological Sequence Generation",
"Pavel Avdeyev1 Chenlai Shi Yuhao Tan1 Kseniia Dudnyk1 Jian Zhoul",
"。3in",
"Abstract",
"连接进行似然评估以及采样方法的灵活性。然而，最初",
"提出的生成式SDE并不直接适用于离散数据建模。最",
"Designing biological sequences is an impor-",
"近的工作提出了使扩散模型适应离散数据（Campbell",
"tant challenge that requires satisfying com-",
"et al.,2022:Chenetal.,2022b:Sun etal.,2022:Austin",
"plex constraints and thus is a natural prob",
"etal.,2021;Hoogeboometal.,2021b;a）的方法，包括",
"lem to address with deep generative mod-",
"离散空间（Campbelletal.，2022）中的连续时间扩散",
"eling.",
"Diffusion generative models have",
"但是除了基于量化的方法外，没有在连续时间SDE扩",
"achieved considerable success in many ap-",
"散框架（Songetal.，2020）内制定任何方法。在本手",
"plications. Score-based generative stochas-",
"稿中，我们提出了一种将这种方法扩展到离散数据的",
"tic differential equations (SDE) model is a",
"般机制，同时允许概率单纯形空间中的连续时间扩",
"continuous-time diffusion model framework",
"散。具体来说，为了利用Dirichlet分布与离散数据之",
"that enjoys many benefits, but the originally",
"间的自然联系，我们考虑了平稳分布为Dirichlet分布",
"proposed SDEs are not naturally designed for",
"的概率单纯形内的连续时间扩散。离散数据的正向扩",
"modeling discrete data. To develop gener-",
"散（数据到噪声）从概率单纯形空间的顶点开始，并",
"ative SDE models for discrete data such as",
"在同一空间中连续扩散，并且可以通过潜在变量解释",
"biological sequences, here we introduce a dif-",
"来桥接连续-离散空间间隙",
"fusion process defined in the probability sim-",
"plex space with stationary distribution being",
"虽然我们的预期应用是在生物序列生成中，但我们在",
"the Dirichlet distribution. This makes diffu-",
"-系列离散数据生成任务上评估了我们的Dirichlet扩",
"sion in continuous space natural for modeling",
"散得分模型（DDSM）1，以更好地了解其性能。除",
"discrete data. We refer to this approach as",
"了在小型基准数据集上展示有竞争力的性能之外，我",
"Dirchlet diffusion score model. We demon-",
"们还将其应用于生成数独谜题，以测试其生成具有强",
"strate that this technique can generate sam-",
"约束的高度结构化数据的能力。该模型不仅可以生成",
"ples that satisfy hard constraints using a Su-",
"还可以解决包括硬谜题在内的数独谜题，这是第一次",
"doku generation task. This generative model",
"通过纯生成建模方法实现。最后，我们将DDSM应用",
"can also solve Sudoku, including hard puz",
"于生物序列生成的实际应用。具体来说，我们开发了",
"zles, without additional training. Finally, we",
"第一个用于设计驱动基因表达的人类启动子DNA序",
"applied this approach to develop the first hu-",
"列的模型，并证明它设计了与人类基因组启动子序列",
"man promoter DNA sequence design model",
"相当的不同序列",
"and showed that designed sequences share",
"similar propertieswith natural promoterse",
"2. Background",
"quences.",
"2.1. Score-based Generative Modeling with SDE",
"ItO扩散过程定义为",
"1.Introduction",
"扩散概率模型是一系列模型，这些模型进行反向扩散",
"处理以从噪声中生成数据。基于分数的生成随机微分",
"da = f(α,t)dt +G(α,t)dw,",
"方程（SDE）是一种连续时间扩散模型，具有许多理想",
"的属性，例如允许通过与概率流常微分方程（ODE）的",
"1 Lyda Hill Department of Bioinformatics, University",
"of Texas Southwestern Medical Center, USA.Correspon-",
"其中w是标准的维纳过程（a。k。a。布朗运动），漂移",
"dence to: Jian Zhou < jian.zhou@utsouthwestern.edu >.",
"系数f(,t):Rn→Rn和扩散系数G(,t)：Rn→Rn",
"是at的向量值函数。Songetal.2020利用了Anderson",
"Proceedings of the 4oth International Conference on Ma-",
"1982的显着结果，即可以通过以下SDE获得该扩散",
"chine Learning, Honolulu, Hawaii, USA. PMLR 202, 2023.",
"Copyright 2023 by the author(s).",
"1Code available at https://github.com/izhoulab/ddsm",
"1"
] |
{
"arxiv_id": "2305.10699",
"language": "zh",
"page_idx": 0
}
|
||
2305.10699-zh-2
|
2305.10699-zh-2.jpg
|
Title Suppressed Due to Excessive Size
<FigureHere>
Figure 1. Schematic overview of forward and reverse diffusion SDEs for Dirichlet diffusion score model. Forward and
reverse diffusion SDEs for 2-category data (a) and k-category data (b) by stick-breaking construction are shown.
通过经典的破棒构造
程 (Feng & Wang, 2007) 。建议的过程是 GEM 过程的有限维版本。我们注意到存在固定分布为 Dirichlet 分布的其他形式的扩散过程(请参见 Steinrücken et al. 2013b; Bakosi & Ristorcelli 2013 )。但是，由于它们不能分解为独立的单变量扩散过程，因此用作正向扩散过程的计算成本要高得多。
3.3. Score Matching Training for $k$ -Category
# Discrete Data
$x_{1}^{t}= \boldsymbol{v}_{1}^{t}$, $\boldsymbol{x}_{2}^{t}= ( 1- \boldsymbol{v}_{1}^{t}) \boldsymbol{v}_{2}^{t}$, $\boldsymbol{x}_{3}^{t}= ( 1- \boldsymbol{v}_{1}^{t}) ( 1- \boldsymbol{v}_{2}^{t}) \boldsymbol{v}_{3}^{t}$, $\cdots $, 收敛到具有$k-1$独立单变量 Jacobi 扩散过程的Dirichlet 分布，其中$v_1^t,v_2^t,\ldots,v_{k-1}^t$来自时间$t$的独立 Jacobi 扩散过程。因此，我们使用$x_1^t,x_2^t,\ldots,x_k^t$ 获得了具有 Dirichlet 平稳分布的多元扩散过程
为了简化表示法，我们将使用$v$和$x$分别表示手稿其余部分的$k-1$和$k$维表示。$v$和$x$之间的转换是通过破棒变换及其逆
为了获得任何 Dirichlet 平稳分布，我们将 Jacobi 扩散过程参数化。例如，对于平稳分布为平坦 Dirichlet 分布$Dir(1,1,\ldots,1)$ (i。e。,在概率单纯形上的均匀分布), 我们需要选择具有平稳分布的 Jacbobi 扩散过程$Beta(1,k-1),Beta(1,k-2),\ldots,Beta(1,1)$作为破棒结构的$v_1,v_2,\ldots,v_{k-1}$ 。这可以简单地通过将方程 5中的参数$a,b$选择为$(1,k-1),(1,k-2),\ldots,(1,1)$来实现
我们将这种多变量扩散过程称为通过破棒构造的多变量 Jacobi 扩散过程，并将这种扩散过程的生成建模方法称为 Dirichlet 扩散得分模型。已经提出了具有更一般分布族的这种扩散过程的无限维形式作为 GEM 过
使用通过破棒构造的多变量 Jacobi 扩散作为正向扩散过程，使我们能够训练$k$类别离散数据的基于分数的扩散模型
将$x$空间中扩散的初始值设置为$k$维一热编码表示的离散数据，例如$(0,0,1,\ldots,0)$ 。为了从正向扩散中采样，我们首先通过逆破棒变换将$x$的初始值映射到$v$ 空间。对于第一个 1 之后的$v$的所有尺寸，在给定$x$ 的情况下，无法通过逆破棒变换确定，我们认为它们是从相应的平稳 Beta 分布中得出的。不需要显式绘制初始值的 Beta 样本，因为密度保持不变，并且任何时候的样本和分数都是直接从 Beta 分布计算得出的
$v$空间中任何时间$t$的正向扩散样本均来自 Jacobi 扩
3
|
[
[
0.35693359375,
0.05682373046875,
0.61767578125,
0.05682373046875,
0.61767578125,
0.0706787109375,
0.35693359375,
0.0706787109375
],
[
0.148681640625,
0.08331298828125,
0.431396484375,
0.08331298828125,
0.431396484375,
0.09661865234375,
0.148681640625,
0.09661865234375
],
[
0.369384765625,
0.09783935546875,
0.52783203125,
0.09783935546875,
0.52783203125,
0.111083984375,
0.369384765625,
0.111083984375
],
[
0.67333984375,
0.09661865234375,
0.708984375,
0.0985107421875,
0.70751953125,
0.11297607421875,
0.67236328125,
0.111083984375
],
[
0.1871337890625,
0.11297607421875,
0.200927734375,
0.11297607421875,
0.200927734375,
0.121826171875,
0.1871337890625,
0.121826171875
],
[
0.1748046875,
0.0997314453125,
0.20751953125,
0.0997314453125,
0.20751953125,
0.11236572265625,
0.1748046875,
0.11236572265625
],
[
0.343017578125,
0.11297607421875,
0.55615234375,
0.11297607421875,
0.55615234375,
0.1275634765625,
0.343017578125,
0.1275634765625
],
[
0.6787109375,
0.11297607421875,
0.6953125,
0.11297607421875,
0.6953125,
0.12249755859375,
0.6787109375,
0.12249755859375
],
[
0.70751953125,
0.1275634765625,
0.74658203125,
0.1275634765625,
0.74658203125,
0.1383056640625,
0.70751953125,
0.1383056640625
],
[
0.43310546875,
0.14013671875,
0.455810546875,
0.14013671875,
0.455810546875,
0.143310546875,
0.43310546875,
0.143310546875
],
[
0.7060546875,
0.1383056640625,
0.736328125,
0.1383056640625,
0.736328125,
0.1458740234375,
0.7060546875,
0.1458740234375
],
[
0.5947265625,
0.14453125,
0.619140625,
0.146484375,
0.61767578125,
0.15283203125,
0.59326171875,
0.15087890625
],
[
0.744140625,
0.14013671875,
0.779296875,
0.14013671875,
0.779296875,
0.1514892578125,
0.744140625,
0.1514892578125
],
[
0.59326171875,
0.1546630859375,
0.62353515625,
0.1546630859375,
0.62353515625,
0.1622314453125,
0.59326171875,
0.1622314453125
],
[
0.74609375,
0.15283203125,
0.81884765625,
0.15283203125,
0.81884765625,
0.1641845703125,
0.74609375,
0.1641845703125
],
[
0.6611328125,
0.1610107421875,
0.67626953125,
0.1610107421875,
0.67626953125,
0.1654052734375,
0.6611328125,
0.1654052734375
],
[
0.26953125,
0.1717529296875,
0.289306640625,
0.1717529296875,
0.289306640625,
0.1787109375,
0.26953125,
0.1787109375
],
[
0.30810546875,
0.1685791015625,
0.62841796875,
0.1685791015625,
0.62841796875,
0.1817626953125,
0.30810546875,
0.1817626953125
],
[
0.371826171875,
0.1849365234375,
0.53759765625,
0.1849365234375,
0.53759765625,
0.1956787109375,
0.371826171875,
0.1956787109375
],
[
0.148681640625,
0.2083740234375,
0.4306640625,
0.2083740234375,
0.4306640625,
0.2215576171875,
0.148681640625,
0.2215576171875
],
[
0.17724609375,
0.2279052734375,
0.2091064453125,
0.2279052734375,
0.2091064453125,
0.2392578125,
0.17724609375,
0.2392578125
],
[
0.236083984375,
0.2279052734375,
0.66259765625,
0.2279052734375,
0.66259765625,
0.2412109375,
0.236083984375,
0.2412109375
],
[
0.6748046875,
0.2279052734375,
0.708984375,
0.2279052734375,
0.708984375,
0.2392578125,
0.6748046875,
0.2392578125
],
[
0.178955078125,
0.24560546875,
0.193603515625,
0.2342529296875,
0.2059326171875,
0.24365234375,
0.191162109375,
0.255126953125
],
[
0.67626953125,
0.242431640625,
0.68896484375,
0.2392578125,
0.69140625,
0.2442626953125,
0.6787109375,
0.248779296875
],
[
0.181396484375,
0.255126953125,
0.1993408203125,
0.255126953125,
0.1993408203125,
0.268310546875,
0.181396484375,
0.268310546875
],
[
0.74609375,
0.29541015625,
0.8056640625,
0.29541015625,
0.8056640625,
0.306884765625,
0.74609375,
0.306884765625
],
[
0.744140625,
0.309326171875,
0.81884765625,
0.309326171875,
0.81884765625,
0.32080078125,
0.744140625,
0.32080078125
],
[
0.365234375,
0.356689453125,
0.53369140625,
0.356689453125,
0.53369140625,
0.37060546875,
0.365234375,
0.37060546875
],
[
0.4697265625,
0.37255859375,
0.5009765625,
0.37255859375,
0.5009765625,
0.380615234375,
0.4697265625,
0.380615234375
],
[
0.279296875,
0.386474609375,
0.443603515625,
0.386474609375,
0.443603515625,
0.399658203125,
0.279296875,
0.399658203125
],
[
0.46728515625,
0.3818359375,
0.5361328125,
0.3818359375,
0.5361328125,
0.393310546875,
0.46728515625,
0.393310546875
],
[
0.46728515625,
0.39453125,
0.56884765625,
0.39453125,
0.56884765625,
0.40478515625,
0.46728515625,
0.40478515625
],
[
0.46728515625,
0.406005859375,
0.59228515625,
0.406005859375,
0.59228515625,
0.416748046875,
0.46728515625,
0.416748046875
],
[
0.369384765625,
0.42041015625,
0.54150390625,
0.421630859375,
0.54150390625,
0.43505859375,
0.369384765625,
0.434326171875
],
[
0.188720703125,
0.427978515625,
0.1993408203125,
0.427978515625,
0.1993408203125,
0.44140625,
0.188720703125,
0.44140625
],
[
0.68310546875,
0.427978515625,
0.69677734375,
0.427978515625,
0.69677734375,
0.438232421875,
0.68310546875,
0.438232421875
],
[
0.181396484375,
0.44140625,
0.205078125,
0.44140625,
0.205078125,
0.460205078125,
0.181396484375,
0.460205078125
],
[
0.67236328125,
0.44140625,
0.810546875,
0.44140625,
0.810546875,
0.455078125,
0.67236328125,
0.455078125
],
[
0.1829833984375,
0.458984375,
0.2034912109375,
0.458984375,
0.2034912109375,
0.474853515625,
0.1829833984375,
0.474853515625
],
[
0.6689453125,
0.45703125,
0.810546875,
0.45703125,
0.810546875,
0.47021484375,
0.6689453125,
0.47021484375
],
[
0.181396484375,
0.472900390625,
0.205078125,
0.472900390625,
0.205078125,
0.492431640625,
0.181396484375,
0.492431640625
],
[
0.6689453125,
0.472900390625,
0.810546875,
0.472900390625,
0.810546875,
0.486083984375,
0.6689453125,
0.486083984375
],
[
0.365234375,
0.49365234375,
0.53369140625,
0.49365234375,
0.53369140625,
0.5068359375,
0.365234375,
0.5068359375
],
[
0.08660888671875,
0.5087890625,
0.88623046875,
0.51025390625,
0.88623046875,
0.5244140625,
0.08660888671875,
0.5234375
],
[
0.0882568359375,
0.5244140625,
0.83251953125,
0.5244140625,
0.83251953125,
0.53857421875,
0.0882568359375,
0.53857421875
],
[
0.0906982421875,
0.5517578125,
0.23779296875,
0.5517578125,
0.23779296875,
0.56494140625,
0.0906982421875,
0.56494140625
],
[
0.5009765625,
0.5517578125,
0.8837890625,
0.5517578125,
0.8837890625,
0.56494140625,
0.5009765625,
0.56494140625
],
[
0.50244140625,
0.56640625,
0.88232421875,
0.56640625,
0.88232421875,
0.580078125,
0.50244140625,
0.580078125
],
[
0.08905029296875,
0.57373046875,
0.4697265625,
0.57373046875,
0.4697265625,
0.58837890625,
0.08905029296875,
0.58837890625
],
[
0.50244140625,
0.58349609375,
0.88134765625,
0.58349609375,
0.88134765625,
0.59326171875,
0.50244140625,
0.59326171875
],
[
0.0906982421875,
0.5966796875,
0.47216796875,
0.5966796875,
0.47216796875,
0.611328125,
0.0906982421875,
0.611328125
],
[
0.498291015625,
0.59521484375,
0.8857421875,
0.5966796875,
0.8857421875,
0.611328125,
0.498291015625,
0.60986328125
],
[
0.0882568359375,
0.611328125,
0.47216796875,
0.6123046875,
0.47216796875,
0.626953125,
0.0882568359375,
0.62548828125
],
[
0.5009765625,
0.60986328125,
0.88134765625,
0.6123046875,
0.88134765625,
0.626953125,
0.5009765625,
0.625
],
[
0.08905029296875,
0.62548828125,
0.47216796875,
0.62744140625,
0.47216796875,
0.642578125,
0.08905029296875,
0.64013671875
],
[
0.50244140625,
0.62548828125,
0.7060546875,
0.62548828125,
0.7060546875,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.08905029296875,
0.6416015625,
0.42236328125,
0.6416015625,
0.42236328125,
0.65576171875,
0.08905029296875,
0.65576171875
],
[
0.498291015625,
0.6572265625,
0.830078125,
0.6591796875,
0.830078125,
0.673828125,
0.498291015625,
0.671875
],
[
0.09234619140625,
0.66552734375,
0.473876953125,
0.66552734375,
0.473876953125,
0.6787109375,
0.09234619140625,
0.6787109375
],
[
0.08905029296875,
0.6787109375,
0.47216796875,
0.6796875,
0.47216796875,
0.6943359375,
0.08905029296875,
0.693359375
],
[
0.53173828125,
0.6748046875,
0.63232421875,
0.6748046875,
0.63232421875,
0.68603515625,
0.53173828125,
0.68603515625
],
[
0.0882568359375,
0.6943359375,
0.2197265625,
0.6943359375,
0.2197265625,
0.70849609375,
0.0882568359375,
0.70849609375
],
[
0.50341796875,
0.6962890625,
0.8857421875,
0.6962890625,
0.8857421875,
0.7109375,
0.50341796875,
0.7109375
],
[
0.0906982421875,
0.71728515625,
0.4697265625,
0.71630859375,
0.4697265625,
0.73095703125,
0.0906982421875,
0.73193359375
],
[
0.50244140625,
0.7119140625,
0.8857421875,
0.7119140625,
0.8857421875,
0.72607421875,
0.50244140625,
0.72607421875
],
[
0.0906982421875,
0.73291015625,
0.473876953125,
0.73291015625,
0.473876953125,
0.7470703125,
0.0906982421875,
0.7470703125
],
[
0.50244140625,
0.72900390625,
0.56787109375,
0.72900390625,
0.56787109375,
0.74072265625,
0.50244140625,
0.74072265625
],
[
0.0906982421875,
0.74755859375,
0.471435546875,
0.74755859375,
0.471435546875,
0.76220703125,
0.0906982421875,
0.76220703125
],
[
0.5009765625,
0.75,
0.8837890625,
0.75,
0.8837890625,
0.76318359375,
0.5009765625,
0.76318359375
],
[
0.08905029296875,
0.76318359375,
0.473876953125,
0.76220703125,
0.473876953125,
0.7763671875,
0.08905029296875,
0.77783203125
],
[
0.50244140625,
0.76513671875,
0.8837890625,
0.76513671875,
0.8837890625,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.0906982421875,
0.77880859375,
0.473876953125,
0.77880859375,
0.473876953125,
0.79248046875,
0.0906982421875,
0.79248046875
],
[
0.50244140625,
0.77978515625,
0.88623046875,
0.77978515625,
0.88623046875,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.08905029296875,
0.79345703125,
0.4755859375,
0.79248046875,
0.4755859375,
0.806640625,
0.08905029296875,
0.80810546875
],
[
0.5009765625,
0.79541015625,
0.88623046875,
0.79541015625,
0.88623046875,
0.8095703125,
0.5009765625,
0.8095703125
],
[
0.0906982421875,
0.8095703125,
0.473876953125,
0.8095703125,
0.473876953125,
0.82373046875,
0.0906982421875,
0.82373046875
],
[
0.50244140625,
0.81005859375,
0.8857421875,
0.81005859375,
0.8857421875,
0.82373046875,
0.50244140625,
0.82373046875
],
[
0.0906982421875,
0.82470703125,
0.12335205078125,
0.82470703125,
0.12335205078125,
0.8369140625,
0.0906982421875,
0.8369140625
],
[
0.50244140625,
0.82568359375,
0.8837890625,
0.82568359375,
0.8837890625,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.0906982421875,
0.84521484375,
0.47216796875,
0.84521484375,
0.47216796875,
0.85986328125,
0.0906982421875,
0.85986328125
],
[
0.5009765625,
0.84033203125,
0.8857421875,
0.8388671875,
0.8857421875,
0.85400390625,
0.5009765625,
0.85498046875
],
[
0.0906982421875,
0.861328125,
0.47216796875,
0.861328125,
0.47216796875,
0.875,
0.0906982421875,
0.875
],
[
0.5009765625,
0.85595703125,
0.85546875,
0.85595703125,
0.85546875,
0.869140625,
0.5009765625,
0.869140625
],
[
0.0906982421875,
0.876953125,
0.45751953125,
0.876953125,
0.45751953125,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.50244140625,
0.87890625,
0.88232421875,
0.87890625,
0.88232421875,
0.892578125,
0.50244140625,
0.892578125
],
[
0.0906982421875,
0.89306640625,
0.47216796875,
0.89306640625,
0.47216796875,
0.90380859375,
0.0906982421875,
0.90380859375
],
[
0.48046875,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93505859375,
0.48046875,
0.93505859375
]
] |
[
"Title Suppressed Due to Excessive Size",
"a.Diffusionprocesses for2-categorydata",
"JacobiDiffusion Process",
"Prior",
"xo",
"Data",
"dx = [a(1 - x) -bx)dt + Vx(1 -x)dw",
"xT",
"probability",
"",
"density",
"",
"Beta",
"",
"distribution",
"1.0",
"dx:",
"[a(1 x) bx) (1 - 2x) x(1 x)Vx log pt(x) dt + Vx(1 -x)dw",
"ReverseDiffusionProcess",
"b. Diffusion processes for k-category data",
"Data",
"Mutilivariate Jacobi Diffusion Process by stick-breaking construction",
"Prior",
"+o",
"+",
"0",
"Dirichlet",
"distribution",
"Reverse Diffusion Process",
"xi=Vi",
"Stick-breaking transform",
"x-(1-vi)v2",
"x-(1-vi)(1-vi)vs",
"x=(1-viD)(1-v)(1-v)",
"Jacobi DiffusionProcesses",
"",
"VT",
"O",
"Beta (1, 3) distribution",
"O",
" Beta (1, 2) distribution",
"O",
"+ Beta (1, 1) distribution",
"Reverse Diffusion Process",
"Figure 1.Schematic overview of forward and reverse diffusion SDEs for Dirichlet diffusion score model. Forward and",
"reverse diffusion SDEs for 2-category data (a) and k-category data (b) by stick-breaking construction are shown",
"通过经典的破棒构造",
"程（Feng&Wang，2007）。建议的过程是GEM过程的",
"有限维版本。我们注意到存在固定分布为Dirichlet分",
"ci = vt, cz =(1-v)v2, ct =(1-vt)(1-v)v, -.:",
"布的其他形式的扩散过程（请参见Steimriickenetal",
"收敛到具有k-1独立单变量Jacobi扩散过程的",
"2013b；Bakosi&Ristorcelli2013）。但是，由于它们",
"Dirichlet分布，其中wt,w,...,u-1来自时间t的",
"不能分解为独立的单变量扩散过程，因此用作正向扩",
"独立Jacobi扩散过程。因此，我们使用ai,a，.，a",
"散过程的计算成本要高得多。",
"获得了具有Dirichlet平稳分布的多元扩散过程",
"3.3.Score Matching Training fork -Category",
"为了简化表示法，我们将使用和&分别表示手稿其",
"余部分的k-1和k维表示。和α之间的转换是通",
"Discrete Data",
"过破棒变换及其逆",
"使用通过破棒构造的多变量Jacobi扩散作为正向扩散",
"为了获得任何Dirichlet平稳分布，我们将Jacobi扩",
"过程，使我们能够训练k类别离散数据的基于分数的",
"散过程参数化。例如，对于平稳分布为平坦Dirichlet",
"扩散模型",
"分布Dir(1,1,..,1）（i。e。，在概率单纯形上的均匀分",
"将&空间中扩散的初始值设置为k维一热编码表示的",
"布)，我们需要选择具有平稳分布的Jacbobi扩散过程",
"离散数据，例如0,0,1，...,0）。为了从正向扩散中采",
"Beta(1,k-1),Beta(1,k-2),..,Beta(1,1)作为破棒",
"样，我们首先通过逆破棒变换将&的初始值映射到",
"结构的U1,2，...，Uk-1。这可以简单地通过将方程5",
"空间。对于第一个1之后的的所有尺寸，在给定α",
"中的参数a,b选择为(1,k-1),(1,k-2)，...,(1,1)来",
"的情况下，无法通过逆破棒变换确定，我们认为它们",
"实现",
"是从相应的平稳Beta分布中得出的。不需要显式绘制",
"我们将这种多变量扩散过程称为通过破棒构造的多变",
"初始值的Beta样本，因为密度保持不变，并且任何时",
"量Jacobi扩散过程，并将这种扩散过程的生成建模方",
"候的样本和分数都是直接从Beta分布计算得出的",
"法称为Dirichlet扩散得分模型。已经提出了具有更",
"u空间中任何时间t的正向扩散样本均来自Jacobi扩",
"般分布族的这种扩散过程的无限维形式作为GEM过",
"3"
] |
{
"arxiv_id": "2305.10699",
"language": "zh",
"page_idx": 2
}
|
||
2305.10699-zh-3
|
2305.10699-zh-3.jpg
|
Title Suppressed Due to Excessive Size
散过程 (对于具有确定性初始值的维度)和平稳 Beta 分布 (对于具有未确定初始值的维度)。然后，根据相应的 Jacobi 扩散过渡密度函数和 Beta 密度函数，计算过渡密度函数在去噪分数匹配损失(方程 4)中的分数
通过应用变量变化转换，我们可以在$v$空间或$x$空间中等效地执行分数匹配，因为我们可以在$x$的分数到$v$的分数之间进行转换：
$$\frac{\partial\log p_{\boldsymbol{x}}(\boldsymbol{x})}{\partial\boldsymbol{x}}=\left(\frac{\partial\log p_{\boldsymbol{v}}(\boldsymbol{v})}{\partial\boldsymbol{v}}+\frac{\partial\log\lvert\det\frac{\partial\boldsymbol{v}}{\partial\boldsymbol{x}}\rvert}{\partial\boldsymbol{v}}\right)\frac{\partial\boldsymbol{v}}{\partial\boldsymbol{x}},$$
$$\frac{\partial\log p_{\boldsymbol{v}}(\boldsymbol{v})}{\partial\boldsymbol{v}}=\frac{\partial\log p_{\boldsymbol{x}}(\boldsymbol{x})}{\partial\boldsymbol{x}}\frac{\partial\boldsymbol{x}}{\partial\boldsymbol{v}}-\frac{\partial\log\lvert\det\frac{\partial\boldsymbol{v}}{\partial\boldsymbol{x}}\rvert}{\partial\boldsymbol{v}}.$$
分数模型$s_{\boldsymbol{\theta}}\left(x^t,t\right)$更自然地被公式化为$x$的函数，并且由于扩散系数的对角形式，我们选择在$v$空间中计算分数匹配损失
一旦学习了分数模型，从$v$空间中的反向扩散过程采样几乎与从等式 7 中的多个单变量反向扩散过程采样相同，除了分数模型将$v$的所有维度作为输入(转换为 $x$ 空间后)
3.4. Weighting Function for Score Matching Loss of
### General SDE
先前已经研究了分数匹配损失中加权函数$\lambda(t)$的选择(等式 3 和 4 ) (Song et al., 2021; Huang et al., 2021) 。假设正向扩散过程的标量扩散系数$g(\boldsymbol{v},t)$不取决于被扩散的值$\boldsymbol{v}$,则$\lambda(t)=g(t)^2$显示为似然加权 (Song et al., 2021) 。使用此权重最小化损失函数等效于最大化 ELBO (Huang et al., 2021)。但是，关于$g(\boldsymbol{v},t)$的假设不适用于 Jacobi 扩散过程
在这里，我们将
(8)
$$\begin{aligned}L(\boldsymbol{v},t)&=\left\|\frac{\partial\log p_{\boldsymbol{v}}(\boldsymbol{v})}{\partial\boldsymbol{v}}-\frac{\partial\log q_{\boldsymbol{v}}(\boldsymbol{v})}{\partial\boldsymbol{v}}\right\|_{\boldsymbol{G}\boldsymbol{G}^{T}}^{2}\\&=\left(\frac{\partial\log p_{\boldsymbol{v}}(\boldsymbol{v})}{\partial\boldsymbol{v}}-\frac{\partial\log q_{\boldsymbol{v}}(\boldsymbol{v})}{\partial\boldsymbol{v}}\right)\boldsymbol{G}(\boldsymbol{v},t)\\&\boldsymbol{G}(\boldsymbol{v},t)^{T}\left(\frac{\partial\mathrm{log}p_{\boldsymbol{v}}(\boldsymbol{v})}{\partial\boldsymbol{v}}-\frac{\partial\mathrm{log}q_{\boldsymbol{v}}(\boldsymbol{v})}{\partial\boldsymbol{v}}\right)^{T}\end{aligned}$$
的使用激发为具有矩阵形式扩散系数$G(\boldsymbol{v},t)$的任何SDE 的加权得分匹配损失的一般形式，其论点是损失函数应满足变量变化下的不变性，这是由似然函数满足的。在附录 A.3 中，我们证明了该损失函数对于任何双射可微函数$\boldsymbol{x}=h(\boldsymbol{v})$的变量变化都是不变的，而未加权损失则不是。如果$G$是标量或对角线并且不依赖于$v$ ,则我们从 Song et al.2020 恢复似然加权
## 3.5. Likelihood Computation
使用分数匹配训练模型后，我们可以使用概率流 ODE 估计似然(请参见公式$\mathbb{E}$2)。我们的公式既可以计算
概率单纯形上连续分布的可能性，也可以计算离散数
据可能性的变化下限
3.5.1.
Likelihood Computation for Continuous Variable in
Probability Simplex
我们将首先估计$v$空间中的似然，可以使用变量的概率变化公式$\mathbb{E}($Chen et al., 2018) 轻松地将其转换为 $x$ 空间中的似然
在 Song et al. 2021 之后，雅可比扩散过程的概率流
ODE 为：
$$\begin{aligned}d\boldsymbol{v}=&\left\{\frac{s}{2}[a(1-\boldsymbol{v})-b\boldsymbol{v}]\right.\\&-\frac{s}{2}(1-2\boldsymbol{v})-\frac{s}{2}\boldsymbol{v}(1-\boldsymbol{v})\boldsymbol{s}_{\theta}(\boldsymbol{v},t)\biggr\}dt=\tilde{f}dt\end{aligned}$$
通过瞬时变量变化公式，我们可以得到：
$$p_0(\boldsymbol{v^0})=e^{\int_0^t\operatorname{tr}\left(\nabla_v\tilde{\boldsymbol{f}}(v^t)\right)dt}p_t(\boldsymbol{v^t}).$$
$\begin{array}{l}\text{Jacobian 的轨迹可以由 Hutchinson 的估计量无偏地}\\\text{近似}\end{array}$
$$\operatorname{tr}\left(\nabla_{\boldsymbol{v}}\tilde{\boldsymbol{f}}(v^t)\right)=\boldsymbol{E}_{\boldsymbol{\epsilon}\sim\mathcal{N}(\boldsymbol{0},\boldsymbol{I})}[\boldsymbol{\epsilon}^T\nabla_{\boldsymbol{v}}\tilde{\boldsymbol{f}}(v^t)\boldsymbol{\epsilon}]$$
和$p_t(v^t)$是用平稳分布计算的。通过应用变量变化公
式将$p_0(\boldsymbol{v^0})$转换为$p_0(\boldsymbol{x^0})$,以获得可能性
由于此连续空间似然不能与离散空间似然直接比较， 因此接下来我们将得出一个证据下界 (ELBO),该证据下界允许与离散数据空间中的似然进行直接比较
3.5.2.
Bounding Discrete Data Likelihood with Variational
Lowerbound
为了获得离散似然的变分下界，我们将从反向扩散过程中提取的概率单纯形空间中的连续变量$x$视为直接参数化的类别分布。从这些分类分布中提取离散数据$y$。因此，我们通过对$x$ $p(\boldsymbol{y})=\int p^\mathrm{Cat}(\boldsymbol{y}|\boldsymbol{x})p(\boldsymbol{x})d\boldsymbol{x}$进行氻缘化来获得离散似然
虽然这在计算上通常很棘手，但我们使用变分下界
$ELBO$
$$\begin{aligned}\log p\left(\boldsymbol{y}\right)\geq\mathbb{E}_{q^{\mathrm{Diff}}(\boldsymbol{x}|\boldsymbol{y})}[-\log q^{\mathrm{Diff}}(\boldsymbol{x}|\boldsymbol{y})+\log p^{\mathrm{Cat}}(\boldsymbol{y}|\boldsymbol{x})\\&+\log p^{\mathrm{ODE}}(\boldsymbol{x})],\end{aligned}$$
其中$q^\mathrm{Diff}(x|y)$是$y$在时间$t_\tilde{0}$的正向扩散密度，$t_\tilde{0}$ 选择为接近0。$p^\mathrm{Cat}(\boldsymbol{y}|\boldsymbol{x})$是分类分布似然。$p^\mathrm{ODE}(\boldsymbol{x})$ 是上一小节中描述的概率流 ODE 的连续空间似然，但时间的下限是$t_{\tilde{0}}$而不是 0。这种期望是通过从正向扩
4
|
[
[
0.359375,
0.059356689453125,
0.615234375,
0.059356689453125,
0.615234375,
0.0694580078125,
0.359375,
0.0694580078125
],
[
0.08905029296875,
0.085205078125,
0.473876953125,
0.08587646484375,
0.473876953125,
0.0997314453125,
0.08905029296875,
0.0985107421875
],
[
0.50244140625,
0.08837890625,
0.8837890625,
0.08837890625,
0.8837890625,
0.0985107421875,
0.50244140625,
0.0985107421875
],
[
0.0906982421875,
0.10162353515625,
0.47216796875,
0.10162353515625,
0.47216796875,
0.11553955078125,
0.0906982421875,
0.11553955078125
],
[
0.5009765625,
0.10162353515625,
0.64892578125,
0.10162353515625,
0.64892578125,
0.11553955078125,
0.5009765625,
0.11553955078125
],
[
0.0906982421875,
0.11614990234375,
0.47216796875,
0.11614990234375,
0.47216796875,
0.1300048828125,
0.0906982421875,
0.1300048828125
],
[
0.0906982421875,
0.1319580078125,
0.47216796875,
0.1319580078125,
0.47216796875,
0.1458740234375,
0.0906982421875,
0.1458740234375
],
[
0.5,
0.1307373046875,
0.54150390625,
0.1307373046875,
0.54150390625,
0.1458740234375,
0.5,
0.1458740234375
],
[
0.0906982421875,
0.147705078125,
0.12335205078125,
0.147705078125,
0.12335205078125,
0.160400390625,
0.0906982421875,
0.160400390625
],
[
0.5009765625,
0.1546630859375,
0.88623046875,
0.1546630859375,
0.88623046875,
0.1685791015625,
0.5009765625,
0.1685791015625
],
[
0.0906982421875,
0.1717529296875,
0.471435546875,
0.1717529296875,
0.471435546875,
0.1817626953125,
0.0906982421875,
0.1817626953125
],
[
0.5009765625,
0.17041015625,
0.64794921875,
0.17041015625,
0.64794921875,
0.184326171875,
0.5009765625,
0.184326171875
],
[
0.0906982421875,
0.184326171875,
0.47216796875,
0.184326171875,
0.47216796875,
0.1976318359375,
0.0906982421875,
0.1976318359375
],
[
0.50244140625,
0.1925048828125,
0.8857421875,
0.1925048828125,
0.8857421875,
0.205810546875,
0.50244140625,
0.205810546875
],
[
0.0882568359375,
0.199462890625,
0.25732421875,
0.199462890625,
0.25732421875,
0.21337890625,
0.0882568359375,
0.21337890625
],
[
0.50244140625,
0.2083740234375,
0.88623046875,
0.2083740234375,
0.88623046875,
0.2215576171875,
0.50244140625,
0.2215576171875
],
[
0.2132568359375,
0.2197265625,
0.410888671875,
0.2197265625,
0.410888671875,
0.2442626953125,
0.2132568359375,
0.2442626953125
],
[
0.42724609375,
0.2235107421875,
0.44775390625,
0.2235107421875,
0.44775390625,
0.238037109375,
0.42724609375,
0.238037109375
],
[
0.50244140625,
0.2235107421875,
0.6005859375,
0.2235107421875,
0.6005859375,
0.23486328125,
0.50244140625,
0.23486328125
],
[
0.42236328125,
0.2291259765625,
0.428955078125,
0.2291259765625,
0.428955078125,
0.23291015625,
0.42236328125,
0.23291015625
],
[
0.108642578125,
0.2247314453125,
0.1854248046875,
0.2247314453125,
0.1854248046875,
0.2392578125,
0.108642578125,
0.2392578125
],
[
0.351318359375,
0.2374267578125,
0.381591796875,
0.2412109375,
0.37744140625,
0.25830078125,
0.34716796875,
0.25390625
],
[
0.419189453125,
0.2392578125,
0.455078125,
0.2392578125,
0.455078125,
0.255615234375,
0.419189453125,
0.255615234375
],
[
0.2467041015625,
0.2412109375,
0.268798828125,
0.2412109375,
0.268798828125,
0.255615234375,
0.2467041015625,
0.255615234375
],
[
0.13232421875,
0.240478515625,
0.16015625,
0.240478515625,
0.16015625,
0.2568359375,
0.13232421875,
0.2568359375
],
[
0.5009765625,
0.24365234375,
0.8857421875,
0.24560546875,
0.8857421875,
0.260009765625,
0.5009765625,
0.25830078125
],
[
0.5,
0.25830078125,
0.56884765625,
0.260009765625,
0.56884765625,
0.274658203125,
0.5,
0.272705078125
],
[
0.2197265625,
0.270751953125,
0.32275390625,
0.270751953125,
0.32275390625,
0.285400390625,
0.2197265625,
0.285400390625
],
[
0.339111328125,
0.267578125,
0.435546875,
0.267578125,
0.435546875,
0.285888671875,
0.339111328125,
0.285888671875
],
[
0.32275390625,
0.2734375,
0.343017578125,
0.2734375,
0.343017578125,
0.281005859375,
0.32275390625,
0.281005859375
],
[
0.1258544921875,
0.271484375,
0.200927734375,
0.271484375,
0.200927734375,
0.28271484375,
0.1258544921875,
0.28271484375
],
[
0.330078125,
0.27587890625,
0.34228515625,
0.27587890625,
0.34228515625,
0.27978515625,
0.330078125,
0.27978515625
],
[
0.148681640625,
0.287353515625,
0.1707763671875,
0.287353515625,
0.1707763671875,
0.300537109375,
0.148681640625,
0.300537109375
],
[
0.2484130859375,
0.287353515625,
0.27294921875,
0.287353515625,
0.27294921875,
0.2998046875,
0.2484130859375,
0.2998046875
],
[
0.298095703125,
0.287353515625,
0.32177734375,
0.287353515625,
0.32177734375,
0.2998046875,
0.298095703125,
0.2998046875
],
[
0.378173828125,
0.287353515625,
0.400390625,
0.287353515625,
0.400390625,
0.300537109375,
0.378173828125,
0.300537109375
],
[
0.55615234375,
0.2822265625,
0.69287109375,
0.292236328125,
0.68896484375,
0.318115234375,
0.55224609375,
0.30810546875
],
[
0.52294921875,
0.294189453125,
0.55615234375,
0.294189453125,
0.55615234375,
0.306884765625,
0.52294921875,
0.306884765625
],
[
0.560546875,
0.3037109375,
0.583984375,
0.3037109375,
0.583984375,
0.31640625,
0.560546875,
0.31640625
],
[
0.0906982421875,
0.3125,
0.47216796875,
0.3125,
0.47216796875,
0.32568359375,
0.0906982421875,
0.32568359375
],
[
0.0906982421875,
0.326904296875,
0.47216796875,
0.326904296875,
0.47216796875,
0.34033203125,
0.0906982421875,
0.34033203125
],
[
0.57177734375,
0.326904296875,
0.7939453125,
0.324462890625,
0.7939453125,
0.35107421875,
0.57177734375,
0.352783203125
],
[
0.7998046875,
0.328857421875,
0.861328125,
0.328857421875,
0.861328125,
0.343505859375,
0.7998046875,
0.343505859375
],
[
0.08905029296875,
0.34228515625,
0.205078125,
0.34228515625,
0.205078125,
0.35595703125,
0.08905029296875,
0.35595703125
],
[
0.64892578125,
0.33837890625,
0.65869140625,
0.33837890625,
0.65869140625,
0.34228515625,
0.64892578125,
0.34228515625
],
[
0.098876953125,
0.3642578125,
0.473876953125,
0.3642578125,
0.473876953125,
0.37744140625,
0.098876953125,
0.37744140625
],
[
0.50244140625,
0.37255859375,
0.78369140625,
0.37255859375,
0.78369140625,
0.386474609375,
0.50244140625,
0.386474609375
],
[
0.0906982421875,
0.37890625,
0.47216796875,
0.37890625,
0.47216796875,
0.392578125,
0.0906982421875,
0.392578125
],
[
0.0906982421875,
0.39453125,
0.47216796875,
0.39453125,
0.47216796875,
0.40771484375,
0.0906982421875,
0.40771484375
],
[
0.576171875,
0.395751953125,
0.80810546875,
0.393310546875,
0.80810546875,
0.41845703125,
0.576171875,
0.42041015625
],
[
0.08905029296875,
0.410888671875,
0.181396484375,
0.410888671875,
0.181396484375,
0.42236328125,
0.08905029296875,
0.42236328125
],
[
0.0882568359375,
0.44140625,
0.464111328125,
0.44140625,
0.464111328125,
0.455078125,
0.0882568359375,
0.455078125
],
[
0.5,
0.436279296875,
0.8837890625,
0.4375,
0.8837890625,
0.451904296875,
0.5,
0.45068359375
],
[
0.12091064453125,
0.45703125,
0.2156982421875,
0.45703125,
0.2156982421875,
0.468505859375,
0.12091064453125,
0.468505859375
],
[
0.50244140625,
0.451904296875,
0.53515625,
0.453857421875,
0.53369140625,
0.466552734375,
0.5009765625,
0.464111328125
],
[
0.09234619140625,
0.48095703125,
0.47216796875,
0.48095703125,
0.47216796875,
0.494384765625,
0.09234619140625,
0.494384765625
],
[
0.5498046875,
0.479736328125,
0.8349609375,
0.479736328125,
0.8349609375,
0.50048828125,
0.5498046875,
0.50048828125
],
[
0.08905029296875,
0.49560546875,
0.47216796875,
0.49560546875,
0.47216796875,
0.5087890625,
0.08905029296875,
0.5087890625
],
[
0.0906982421875,
0.51025390625,
0.473876953125,
0.51025390625,
0.473876953125,
0.5234375,
0.0906982421875,
0.5234375
],
[
0.50244140625,
0.5146484375,
0.88232421875,
0.5146484375,
0.88232421875,
0.52783203125,
0.50244140625,
0.52783203125
],
[
0.08905029296875,
0.52587890625,
0.473876953125,
0.52587890625,
0.473876953125,
0.5390625,
0.08905029296875,
0.5390625
],
[
0.5,
0.52880859375,
0.81396484375,
0.52978515625,
0.81396484375,
0.54443359375,
0.5,
0.54345703125
],
[
0.08905029296875,
0.54052734375,
0.47216796875,
0.54052734375,
0.47216796875,
0.5537109375,
0.08905029296875,
0.5537109375
],
[
0.08905029296875,
0.5537109375,
0.47216796875,
0.5546875,
0.47216796875,
0.5693359375,
0.08905029296875,
0.568359375
],
[
0.50341796875,
0.5537109375,
0.8798828125,
0.5537109375,
0.8798828125,
0.564453125,
0.50341796875,
0.564453125
],
[
0.0906982421875,
0.57080078125,
0.310546875,
0.57080078125,
0.310546875,
0.583984375,
0.0906982421875,
0.583984375
],
[
0.50244140625,
0.5673828125,
0.8837890625,
0.5673828125,
0.8837890625,
0.58056640625,
0.50244140625,
0.58056640625
],
[
0.50244140625,
0.583984375,
0.85888671875,
0.583984375,
0.85888671875,
0.5947265625,
0.50244140625,
0.5947265625
],
[
0.08905029296875,
0.59326171875,
0.205078125,
0.59326171875,
0.205078125,
0.60693359375,
0.08905029296875,
0.60693359375
],
[
0.5,
0.611328125,
0.5400390625,
0.611328125,
0.5400390625,
0.62548828125,
0.5,
0.62548828125
],
[
0.176513671875,
0.6171875,
0.363525390625,
0.6171875,
0.363525390625,
0.6318359375,
0.176513671875,
0.6318359375
],
[
0.10455322265625,
0.62548828125,
0.1666259765625,
0.62548828125,
0.1666259765625,
0.64013671875,
0.10455322265625,
0.64013671875
],
[
0.20751953125,
0.6337890625,
0.2296142578125,
0.6337890625,
0.2296142578125,
0.64794921875,
0.20751953125,
0.64794921875
],
[
0.30224609375,
0.6337890625,
0.325927734375,
0.6337890625,
0.325927734375,
0.64794921875,
0.30224609375,
0.64794921875
],
[
0.35693359375,
0.6416015625,
0.38818359375,
0.6416015625,
0.38818359375,
0.6484375,
0.35693359375,
0.6484375
],
[
0.50244140625,
0.63720703125,
0.8837890625,
0.63720703125,
0.8837890625,
0.64794921875,
0.50244140625,
0.64794921875
],
[
0.176513671875,
0.65380859375,
0.285888671875,
0.65380859375,
0.285888671875,
0.66845703125,
0.176513671875,
0.66845703125
],
[
0.27783203125,
0.65380859375,
0.36279296875,
0.65380859375,
0.36279296875,
0.66845703125,
0.27783203125,
0.66845703125
],
[
0.50244140625,
0.6513671875,
0.5888671875,
0.6513671875,
0.5888671875,
0.66357421875,
0.50244140625,
0.66357421875
],
[
0.363525390625,
0.6591796875,
0.42236328125,
0.6611328125,
0.420654296875,
0.67529296875,
0.36279296875,
0.673828125
],
[
0.451904296875,
0.66357421875,
0.471435546875,
0.66357421875,
0.471435546875,
0.67529296875,
0.451904296875,
0.67529296875
],
[
0.2099609375,
0.67041015625,
0.2320556640625,
0.67041015625,
0.2320556640625,
0.68310546875,
0.2099609375,
0.68310546875
],
[
0.3056640625,
0.67041015625,
0.328369140625,
0.67041015625,
0.328369140625,
0.6845703125,
0.3056640625,
0.6845703125
],
[
0.5048828125,
0.67529296875,
0.88232421875,
0.67529296875,
0.88232421875,
0.68603515625,
0.5048828125,
0.68603515625
],
[
0.236083984375,
0.69140625,
0.423095703125,
0.689453125,
0.423095703125,
0.70556640625,
0.236083984375,
0.70849609375
],
[
0.50244140625,
0.68798828125,
0.8837890625,
0.68798828125,
0.8837890625,
0.7021484375,
0.50244140625,
0.7021484375
],
[
0.1690673828125,
0.69970703125,
0.2197265625,
0.69970703125,
0.2197265625,
0.7138671875,
0.1690673828125,
0.7138671875
],
[
0.268798828125,
0.708984375,
0.291748046875,
0.708984375,
0.291748046875,
0.7216796875,
0.268798828125,
0.7216796875
],
[
0.366943359375,
0.708984375,
0.3896484375,
0.708984375,
0.3896484375,
0.72265625,
0.366943359375,
0.72265625
],
[
0.50244140625,
0.7041015625,
0.8837890625,
0.7041015625,
0.8837890625,
0.71728515625,
0.50244140625,
0.71728515625
],
[
0.5,
0.7197265625,
0.8857421875,
0.71728515625,
0.8857421875,
0.73193359375,
0.5,
0.734375
],
[
0.09234619140625,
0.73095703125,
0.47216796875,
0.73095703125,
0.47216796875,
0.744140625,
0.09234619140625,
0.744140625
],
[
0.5009765625,
0.73291015625,
0.68310546875,
0.73291015625,
0.68310546875,
0.7470703125,
0.5009765625,
0.7470703125
],
[
0.0882568359375,
0.744140625,
0.47216796875,
0.74560546875,
0.47216796875,
0.76025390625,
0.0882568359375,
0.7587890625
],
[
0.09234619140625,
0.76123046875,
0.47216796875,
0.76123046875,
0.47216796875,
0.7744140625,
0.09234619140625,
0.7744140625
],
[
0.50244140625,
0.75830078125,
0.88232421875,
0.75830078125,
0.88232421875,
0.76806640625,
0.50244140625,
0.76806640625
],
[
0.09234619140625,
0.7763671875,
0.47216796875,
0.7763671875,
0.47216796875,
0.79052734375,
0.09234619140625,
0.79052734375
],
[
0.50244140625,
0.771484375,
0.54833984375,
0.771484375,
0.54833984375,
0.78271484375,
0.50244140625,
0.78271484375
],
[
0.09234619140625,
0.79345703125,
0.471435546875,
0.79345703125,
0.471435546875,
0.8037109375,
0.09234619140625,
0.8037109375
],
[
0.50341796875,
0.79541015625,
0.87744140625,
0.79443359375,
0.87744140625,
0.81298828125,
0.50341796875,
0.814453125
],
[
0.0906982421875,
0.806640625,
0.473876953125,
0.806640625,
0.473876953125,
0.82080078125,
0.0906982421875,
0.82080078125
],
[
0.0882568359375,
0.82275390625,
0.44677734375,
0.82275390625,
0.44677734375,
0.8359375,
0.0882568359375,
0.8359375
],
[
0.76708984375,
0.8173828125,
0.8798828125,
0.8173828125,
0.8798828125,
0.83203125,
0.76708984375,
0.83203125
],
[
0.51318359375,
0.84326171875,
0.88623046875,
0.84521484375,
0.88623046875,
0.86279296875,
0.51318359375,
0.861328125
],
[
0.0882568359375,
0.85400390625,
0.301513671875,
0.85400390625,
0.301513671875,
0.8671875,
0.0882568359375,
0.8671875
],
[
0.5087890625,
0.85302734375,
0.515625,
0.85302734375,
0.515625,
0.85595703125,
0.5087890625,
0.85595703125
],
[
0.5009765625,
0.859375,
0.88623046875,
0.859375,
0.88623046875,
0.876953125,
0.5009765625,
0.876953125
],
[
0.0906982421875,
0.876953125,
0.4755859375,
0.876953125,
0.4755859375,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.5009765625,
0.876953125,
0.8837890625,
0.876953125,
0.8837890625,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.0906982421875,
0.892578125,
0.473876953125,
0.892578125,
0.473876953125,
0.90576171875,
0.0906982421875,
0.90576171875
],
[
0.498291015625,
0.892578125,
0.8798828125,
0.892578125,
0.8798828125,
0.90576171875,
0.498291015625,
0.90576171875
],
[
0.48193359375,
0.9248046875,
0.494384765625,
0.9248046875,
0.494384765625,
0.93505859375,
0.48193359375,
0.93505859375
]
] |
[
"Title Suppressed Due to ExcessiveSize",
"散过程（对于具有确定性初始值的维度）和平稳Beta",
"概率单纯形上连续分布的可能性，也可以计算离散数",
"分布（对于具有未确定初始值的维度）。然后，根据相",
"据可能性的变化下限",
"应的Jacobi扩散过渡密度函数和Beta密度函数，计",
"算过渡密度函数在去噪分数匹配损失（方程4）中的",
"3.5.1.",
"分数",
"Likelihood Computation for Continuous Variable in",
"通过应用变量变化转换：我们可以在?空间或3空间",
"Probability Simplex",
"中等效地执行分数匹配，因为我们可以在&的分数到",
"我们将首先估计空间中的似然，可以使用变量的概",
"u的分数之间进行转换：",
"率变化公式E（Chenetal.，2018）轻松地将其转换为a",
"(alog Pu(α) + logldet |",
"au",
"空间中的似然",
"",
"alog pr (α)",
"0",
"dr",
"av",
"aα",
"在Songetal.2021之后，雅可比扩散过程的概率流",
"ODE为：",
"Olog pa() Oa",
"0log|det ]",
"",
"alog pu(u)",
"",
"av",
"Oa",
"v",
"dv",
"[[a(1 - ~) - bl]",
"d:",
"[2 ",
"分数模型se（αt,t）更自然地被公式化为a的函数，并",
"且由于扩散系数的对角形式，我们选择在空间中计",
"2(1 - 20) -0(1 -0)so(u, t)",
"dt=fdi",
"算分数匹配损失",
"",
"一旦学习了分数模型，从空间中的反向扩散过程采",
"通过瞬时变量变化公式，我们可以得到",
"样几乎与从等式7中的多个单变量反向扩散过程采样",
"相同，除了分数模型将的所有维度作为输入（转换",
"Po(0) =el& t(V(o)ap(ut).",
"为&空间后",
"3.4. Weighting Function for Score Matching Loss of",
"Jacobian的轨迹可以由Hutchinson的估计量无偏地",
"General SDE",
"近似",
"先前已经研究了分数匹配损失中加权函数入（t）的选择",
"tr (Vf(ut)) = Ee~N(0,n)[eTVf(ut)e)",
"(等式3和4）(Songet al.,2021;Huangetal.,2021)。",
"假设正向扩散过程的标量扩散系数g(u,t)不取决于被",
"和Pt(ut）是用平稳分布计算的。通过应用变量变化公",
"扩散的值，则入(t)=g(t)2显示为似然加权(Song",
"式将pouo）转换为po(a），以获得可能性",
"etal.，2021）。使用此权重最小化损失函数等效于最大",
"化ELBO（Huangetal.,2021）。但是，关于g(u,t）的",
"由于此连续空间似然不能与离散空间似然直接比较",
"假设不适用于Jacobi扩散过程",
"因此接下来我们将得出一个证据下界（ELBO），该证",
"据下界充许与离散数据空间中的似然进行直接比较",
"在这里，我们将",
"3.5.2.",
"alog p (u)log q(u) I",
"L(u,t) =",
"av",
"dv",
"IGG",
"Bounding Discrete Data Likelihood with Variational",
"(alog pu(v) ",
"alog qu(v)",
"Lowerbound",
"G(u,t)",
"(8)",
"av",
"av",
"为了获得离散似然的变分下界，我们将从反向扩散过",
"(logpo(u)log qo(u))",
"程中提取的概率单纯形空间中的连续变量&视为直接",
"G(u,t)",
"av",
"du",
"参数化的类别分布。从这些分类分布中提取离散数据",
"y。因此，我们通过对ap(y)=fpCat(ya)p(a)da进",
"的使用激发为具有矩阵形式扩散系数G(,t)的任何",
"行边缘化来获得离散似然",
"SDE的加权得分匹配损失的一般形式，其论点是损失",
"函数应满足变量变化下的不变性，这是由似然函数满",
"虽然这在计算上通常很棘手，但我们使用变分下界",
"足的。在附录A.3中，我们证明了该损失函数对于任",
"ELBO",
"何双射可微函数=h（の）的变量变化都是不变的，而",
"log p(y) ≥ Egpif(mly)[-log qDif(aly) + log pCat(yla)",
"未加权损失则不是。如果G是标量或对角线并且不依",
"赖于，则我们从Songetal.2020恢复似然加权",
"+ log pODE(α),",
"其中qDif(aly）是y在时间t的正向扩散密度，to",
"3.5. Likelihood Computation",
"",
"选择为接近 0。pCat(yla）是分类分布似然。pODE(ac)",
"使用分数匹配训练模型后，我们可以使用概率流ODE",
"是上一小节中描述的概率流ODE的连续空间似然，但",
"估计似然（请参见公式E2）。我们的公式既可以计算",
"时间的下限是而不是0。这种期望是通过从正向扩",
"4"
] |
{
"arxiv_id": "2305.10699",
"language": "zh",
"page_idx": 3
}
|
||
1707.08347-zh-0
|
1707.08347-zh-0.jpg
|
RankIQA: Learning from Rankings for No-reference Image Quality Assessment
Xialei Liu
Computer Vision Center
Barcelona, Spain
xialei@cvc.uab.es
Joost van de Weijer
Computer Vision Center
Barcelona, Spain
joost@ cvc. uab. es
Andrew D. Bagdanov
MICC, University of Florence
Florence, Italy
andrew.bagdanov@unifi.it
## $\operatorname{Abstract}$
We propose a no-reference image quality assessment (NR-IQA) approach that learns from rankings $({\mathrm{~RankIQA~).~To~address~the~problem~of~limited~IQA}}$ dataset size, we train a Siamese Network to rank images in terms of image quality by using synthetically generated distortions for which relative image quality is known. These ranked image sets can be automatically generated without laborious human labeling. We then use fine-tuning to transfer the knowledge represented in the trained Siamese Network to a traditional CNN that estimates absolute image quality from single images. We demonstrate how our approach can be made significantly more efficient than traditional Siamese Networks by forward propagating a batch of images through a single network and backpropagating gradients derived from all pairs of images in the batch. Experiments on the TID2013 benchmark show that we improve the state- of- the- art by over 5$\%$. Furthermore, on the LIVE benchmark we show that our approach is superior to existing NR-IQA techniques and that we even outperform the state-of-the-art in full-reference IQA (FR-IQA) methods without having to resort to high-quality reference images to infer IQA.
## 1. Introduction
图像在我们的生活中无处不在。不幸的是，它们经常被获取，传输，存储和外部条件(如相机运动)的过程所扭曲。图像质量评估 (IQA) [34] 是一种自动预测图像感知质量的技术。IQA 估计值应与一系列非常多的人类评估人员进行的质量评估高度相关(通常称为平均意见得分 (MOS)$\mathsf{E}[27,22]$ )。IQA 已广泛应用于图像质量至关重要的问题，例如图像恢复$\mathbb{E}[13]$ ,图像超分辨率$\mathbb{E}[32]$ 和图像检索$\mathbb{E}[37]$
IQA 方法通常根据未失真图像(称为参考图像)或有关其的信息是否可用而分为三类：全参考 IQA (FRIQA),缩减参考 IQA(RR-IQA)和无参考 IQA(NRIQA)。研究主要集中在 NR-IQA 的更逼真的场景上，
<FigureHere>
Figure 1. The classical approach trains a deep CNN regressor directly on the ground-truth. Our approach trains a network from an image ranking dataset. These ranked images can be easily generated by applying distortions of varying intensities. The network parameters are then transferred to the regression network for fine-tuning. This allows for the training of deeper and wider networks.
其中必须估计没有任何参考图像的图像质量。在 NRIQA 中，许多方法都集中在特定的失真$\overline{\mathbb{E}}[38,7]$上，这限制了这些方法的适用性。其他方法考虑了一系列失真[20,23,18,19]
卷积神经网络(CNN) 正在对计算机视觉研究和实践产生巨大影响。虽然它们已经存在了几十年[16],但直到 2012 年，当 Krizhevsky 等人。[14]在 ImageNet 竞赛中通过 CNN 取得了惊人的成绩，在更广泛的计算机视觉社区中获得了广泛的关注和采用。相对于原来的 AlexNet, 网络的架构越来越深，ResNet 就是一个很深的网络架构的例子 [8] 。这种趋势的结果是，像AlexNet 和 ResNet 这样的最先进的 CNN 拥有数以亿计的参数，并且需要大量的数据来从头开始训练(而不会过度拟合)
CNN 的成功鼓励研究探索其在 NR-IQA 问题中的潜在应用。与以前的手工方法相比，这项研究带来了显著的改进[17,11,12] 。这些论文必须解决的主要问题是缺乏 IQA 的大型数据集。特别是随着网络越来越深人和广泛，参数的数量急剧增加。因此，训练需要越来
1
|
[
[
0.0882568359375,
0.1319580078125,
0.88232421875,
0.1331787109375,
0.88232421875,
0.15087890625,
0.0882568359375,
0.149658203125
],
[
0.1732177734375,
0.18310546875,
0.259033203125,
0.18310546875,
0.259033203125,
0.1976318359375,
0.1732177734375,
0.1976318359375
],
[
0.3798828125,
0.18310546875,
0.54833984375,
0.18310546875,
0.54833984375,
0.1976318359375,
0.3798828125,
0.1976318359375
],
[
0.6435546875,
0.18310546875,
0.82421875,
0.18310546875,
0.82421875,
0.1976318359375,
0.6435546875,
0.1976318359375
],
[
0.11358642578125,
0.2008056640625,
0.3203125,
0.2008056640625,
0.3203125,
0.2139892578125,
0.11358642578125,
0.2139892578125
],
[
0.361083984375,
0.2008056640625,
0.56787109375,
0.2008056640625,
0.56787109375,
0.2139892578125,
0.361083984375,
0.2139892578125
],
[
0.609375,
0.2008056640625,
0.859375,
0.2008056640625,
0.859375,
0.2139892578125,
0.609375,
0.2139892578125
],
[
0.1422119140625,
0.21533203125,
0.289306640625,
0.216552734375,
0.289306640625,
0.2342529296875,
0.1422119140625,
0.23291015625
],
[
0.390625,
0.2139892578125,
0.5361328125,
0.216552734375,
0.5361328125,
0.2342529296875,
0.390625,
0.231689453125
],
[
0.6708984375,
0.21533203125,
0.7939453125,
0.2171630859375,
0.7939453125,
0.23486328125,
0.6708984375,
0.23291015625
],
[
0.158447265625,
0.2374267578125,
0.273681640625,
0.2374267578125,
0.273681640625,
0.250732421875,
0.158447265625,
0.250732421875
],
[
0.406005859375,
0.2374267578125,
0.52294921875,
0.2374267578125,
0.52294921875,
0.250732421875,
0.406005859375,
0.250732421875
],
[
0.64794921875,
0.236083984375,
0.8203125,
0.236083984375,
0.8203125,
0.250732421875,
0.64794921875,
0.250732421875
],
[
0.238525390625,
0.287353515625,
0.31201171875,
0.287353515625,
0.31201171875,
0.3017578125,
0.238525390625,
0.3017578125
],
[
0.098876953125,
0.317626953125,
0.46728515625,
0.318115234375,
0.46728515625,
0.333251953125,
0.098876953125,
0.33203125
],
[
0.07843017578125,
0.333251953125,
0.468017578125,
0.333984375,
0.468017578125,
0.348388671875,
0.07843017578125,
0.347900390625
],
[
0.08087158203125,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.363037109375,
0.08087158203125,
0.363037109375
],
[
0.080078125,
0.3642578125,
0.46728515625,
0.3642578125,
0.46728515625,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.080078125,
0.380126953125,
0.46728515625,
0.380126953125,
0.46728515625,
0.393310546875,
0.080078125,
0.393310546875
],
[
0.08087158203125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.4091796875,
0.08087158203125,
0.4091796875
],
[
0.080078125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.5537109375,
0.410888671875,
0.5810546875,
0.410888671875,
0.5810546875,
0.4140625,
0.5537109375,
0.4140625
],
[
0.080078125,
0.4248046875,
0.468017578125,
0.423583984375,
0.468017578125,
0.438232421875,
0.080078125,
0.439453125
],
[
0.080078125,
0.43994140625,
0.465576171875,
0.43994140625,
0.465576171875,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.08087158203125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.08087158203125,
0.468505859375
],
[
0.080078125,
0.468505859375,
0.46728515625,
0.4697265625,
0.46728515625,
0.484130859375,
0.080078125,
0.48291015625
],
[
0.080078125,
0.484130859375,
0.468017578125,
0.48486328125,
0.468017578125,
0.5,
0.080078125,
0.498779296875
],
[
0.50244140625,
0.4912109375,
0.8896484375,
0.4912109375,
0.8896484375,
0.505859375,
0.50244140625,
0.505859375
],
[
0.08087158203125,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.08087158203125,
0.5146484375
],
[
0.50244140625,
0.505859375,
0.890625,
0.505859375,
0.890625,
0.51953125,
0.50244140625,
0.51953125
],
[
0.080078125,
0.5146484375,
0.4697265625,
0.51513671875,
0.4697265625,
0.52978515625,
0.080078125,
0.52880859375
],
[
0.50341796875,
0.52099609375,
0.8896484375,
0.51953125,
0.8896484375,
0.53076171875,
0.50341796875,
0.5322265625
],
[
0.080078125,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54443359375,
0.080078125,
0.54443359375
],
[
0.50244140625,
0.5341796875,
0.89208984375,
0.5341796875,
0.89208984375,
0.54736328125,
0.50244140625,
0.54736328125
],
[
0.08087158203125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.56005859375,
0.08087158203125,
0.56005859375
],
[
0.50341796875,
0.54736328125,
0.8896484375,
0.54736328125,
0.8896484375,
0.56103515625,
0.50341796875,
0.56103515625
],
[
0.080078125,
0.5595703125,
0.4697265625,
0.56005859375,
0.4697265625,
0.57470703125,
0.080078125,
0.57373046875
],
[
0.5048828125,
0.5625,
0.890625,
0.5625,
0.890625,
0.57275390625,
0.5048828125,
0.57275390625
],
[
0.07843017578125,
0.57470703125,
0.46728515625,
0.57568359375,
0.46728515625,
0.59033203125,
0.07843017578125,
0.5888671875
],
[
0.50244140625,
0.57470703125,
0.8095703125,
0.57568359375,
0.8095703125,
0.5869140625,
0.50244140625,
0.5859375
],
[
0.080078125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.080078125,
0.60498046875
],
[
0.08087158203125,
0.60791015625,
0.468017578125,
0.60791015625,
0.468017578125,
0.61865234375,
0.08087158203125,
0.61865234375
],
[
0.5048828125,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.5048828125,
0.62744140625
],
[
0.08087158203125,
0.62255859375,
0.46728515625,
0.62255859375,
0.46728515625,
0.63330078125,
0.08087158203125,
0.63330078125
],
[
0.50244140625,
0.62744140625,
0.890625,
0.62890625,
0.890625,
0.64306640625,
0.50244140625,
0.642578125
],
[
0.080078125,
0.63525390625,
0.4697265625,
0.63623046875,
0.4697265625,
0.6494140625,
0.080078125,
0.6484375
],
[
0.07843017578125,
0.6494140625,
0.3857421875,
0.65087890625,
0.3857421875,
0.66552734375,
0.07843017578125,
0.6640625
],
[
0.5048828125,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.6591796875,
0.5048828125,
0.6591796875
],
[
0.50341796875,
0.66015625,
0.63623046875,
0.66015625,
0.63623046875,
0.673828125,
0.50341796875,
0.673828125
],
[
0.52392578125,
0.6748046875,
0.89208984375,
0.6748046875,
0.89208984375,
0.68798828125,
0.52392578125,
0.68798828125
],
[
0.080078125,
0.6943359375,
0.21728515625,
0.69580078125,
0.21728515625,
0.71044921875,
0.080078125,
0.708984375
],
[
0.5048828125,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.5048828125,
0.7041015625
],
[
0.5048828125,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.7197265625,
0.5048828125,
0.7197265625
],
[
0.10296630859375,
0.72021484375,
0.468017578125,
0.72021484375,
0.468017578125,
0.734375,
0.10296630859375,
0.734375
],
[
0.5048828125,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.5048828125,
0.734375
],
[
0.08087158203125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.08087158203125,
0.75
],
[
0.5048828125,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.5048828125,
0.74853515625
],
[
0.08251953125,
0.75048828125,
0.46728515625,
0.75048828125,
0.46728515625,
0.7646484375,
0.08251953125,
0.7646484375
],
[
0.5048828125,
0.75048828125,
0.8896484375,
0.75048828125,
0.8896484375,
0.7646484375,
0.5048828125,
0.7646484375
],
[
0.08416748046875,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.08416748046875,
0.77978515625
],
[
0.5048828125,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.5048828125,
0.77978515625
],
[
0.08251953125,
0.78271484375,
0.465576171875,
0.78271484375,
0.465576171875,
0.79345703125,
0.08251953125,
0.79345703125
],
[
0.50244140625,
0.77880859375,
0.890625,
0.77978515625,
0.890625,
0.79443359375,
0.50244140625,
0.79345703125
],
[
0.08087158203125,
0.79443359375,
0.465576171875,
0.79541015625,
0.465576171875,
0.8095703125,
0.08087158203125,
0.80810546875
],
[
0.50341796875,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.8095703125,
0.50341796875,
0.8095703125
],
[
0.08251953125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.08251953125,
0.82470703125
],
[
0.50244140625,
0.81005859375,
0.59228515625,
0.81005859375,
0.59228515625,
0.82470703125,
0.50244140625,
0.82470703125
],
[
0.08087158203125,
0.82568359375,
0.316162109375,
0.82568359375,
0.316162109375,
0.8388671875,
0.08087158203125,
0.8388671875
],
[
0.52294921875,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.52294921875,
0.8388671875
],
[
0.1004638671875,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.1004638671875,
0.85498046875
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.46728515625,
0.85498046875,
0.46728515625,
0.869140625,
0.080078125,
0.87060546875
],
[
0.50634765625,
0.85791015625,
0.88818359375,
0.85791015625,
0.88818359375,
0.86865234375,
0.50634765625,
0.86865234375
],
[
0.08087158203125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.5048828125,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5048828125,
0.90087890625
],
[
0.483642578125,
0.9267578125,
0.490234375,
0.9267578125,
0.490234375,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"RankIQA: Learning from Rankings for No-reference Image Quality Assessment",
"Xialei Liu",
"Joost van de Weijer",
"Andrew D.Bagdanov",
"Computer Vision Center",
"Computer Vision Center",
"MICC, University of Florence",
"Barcelona, Spain",
"Barcelona, Spain",
"Florence, Italy",
"xialei@cvc.uab.es",
"joost@cvc.uab.es",
"andrew.bagdanov@unifi.it",
"Abstract",
"We propose a no-reference image quality assess-",
"ment (NR-IQA) approach that learns from rankings",
"(RankIQA ). To address the problem of limited IQA",
"dataset size, we train a Siamese Network to rank im-",
"ages in terms of image quality by using synthetically",
"generated distortions for which relative image quality",
"is known. These ranked image sets can be automati-",
"aoproac",
"cally generated without laborious human labeling. We",
"then use fine-tuning to transfer the knowledge repre-",
"sented in the trained Siamese Network to a traditional",
"CNN that estimates absolute image quality from sin-",
"gle images.We demonstrate how our approach can",
"Figure 1.The classical approach trains a deep CNN re-",
"be made significantly more efficient than traditional",
"gressor directly on the ground-truth. Our approach trains",
"Siamese Networks by forward propagating a batch of",
"a network from an image ranking dataset.These ranked",
"images through a single network and backpropagating",
"images can be easily generated by applying distortions of",
"gradients derived from all pairs of images in the batch.",
"varying intensities. The network parameters are then trans-",
"Experiments on the TID2013 benchmark show that we",
"ferredtotheregressionnetworkforfine-tuning.Thisallows",
"improve the state-of-the-art by over 5%.Furthermore,",
"for the training of deeper and wider networks.",
"on the LIVE benchmark we show that our approach is",
"superiorto existingNR-IQA techniques andthat we",
"其中必须估计没有任何参考图像的图像质量。在NR-",
"even outperformthe state-of-the-art infull-reference",
"IQA中，许多方法都集中在特定的失真E38，7]上，这",
"IQA (FR-IQA)methodswithouthaving toresort to",
"high-quality reference images to infer IQA",
"限制了这些方法的适用性。其他方法考虑了一系列失",
"真[20,23,18,19]",
"卷积神经网络（CNN）正在对计算机视觉研究和实",
"1. Introduction",
"践产生巨大影响。虽然它们已经存在了几十年「16]，但",
"直到2012年，当Krizhevsky等人。[14]在ImageNet",
"图像在我们的生活中无处不在。不幸的是，它们经",
"竞赛中通过CNN取得了惊人的成绩，在更广泛的计算",
"常被获取，传输，存储和外部条件（如相机运动）的过",
"机视觉社区中获得了广泛的关注和采用。相对于原来",
"程所扭曲。图像质量评估（IQA）【34是一种自动预测",
"的AlexNet，网络的架构越来越深，ResNet就是一个",
"图像感知质量的技术。IQA估计值应与一系列非常多",
"很深的网络架构的例子8］。这种趋势的结果是，像",
"的人类评估人员进行的质量评估高度相关（通常称为",
"AlexNet和ResNet这样的最先进的CNN拥有数以亿",
"平均意见得分（MOS)E[27.221）。IQA已广泛应用于",
"计的参数，并且需要大量的数据来从头开始训练（而不",
"图像质量至关重要的问题，例如图像恢复E13]，图像",
"会过度拟合）",
"超分辨率E[32]和图像检索E[37]",
"CNN的成功鼓励研究探索其在NR-IQA问题中的",
"IQA方法通常根据未失真图像（称为参考图像）或",
"潜在应用。与以前的手工方法相比，这项研究带来了显",
"有关其的信息是否可用而分为三类：全参考IQA（FR-",
"著的改进[17，11.12]。这些论文必须解决的主要问题",
"IQA），缩减参考IQA（RR-IQA）和无参考IQA（NR-",
"是缺乏IQA的大型数据集。特别是随着网络越来越深",
"IQA）。研究主要集中在NR-IQA的更逼真的场景上，",
"人和广泛，参数的数量急剧增加。因此，训练需要越来",
"1"
] |
{
"arxiv_id": "1707.08347",
"language": "zh",
"page_idx": 0
}
|
||
1707.08347-zh-1
|
1707.08347-zh-1.jpg
|
越大的注释数据集。然而，IQA 图像数据集的注释过程需要每个图像的多个人工注释，因此收集过程是非常劳动密集型和昂贵的。因此，大多数可用的 IQA 数据集太小，无法有效地训练 CNN。我们提出了一种解决大型数据集缺失的方法。主要思想(见图。$\color{red}{\mathrm{E1}}$)是， 尽管难以获得带有人类注释的 IQA 数据，但很容易生成根据其图像质量进行排名的图像。也就是说，我们可以生成图像集，其中，尽管我们没有每个生成的图像的绝对质量度量，但对于我们知道的任何一对图像，哪个图像的质量更高。我们将这种从排名中学习的方法称为 RankIQA, 并通过它学习使用 SiameseNetworks 对图像的质量进行排名，然后将从排名图像中学到的知识转移到对 IQA 数据进行微调的传统 CNN 中，以提高 IQA 的准确性。Zhang 等人提出了从失真参考图像中学习 IQA 特征的想法。专利[44]。在本文中，我们超越了这项专利，因为我们提供了我们的方法的详细描述，并通过实验验证了使用排名数据集的预训练网络的有用性
作为第二个贡献，我们提出了一种在暹罗网络中进行有效反向传播的方法。该方法通过单个网络转发一批图像，然后反向传播从批次中的所有对导出的梯度。在对已建立的 IQA 数据集进行的广泛实验中，我们表明从排名中学习可以显着改善结果，并且我们高效的反向传播算法可以比其他训练协议(例如硬负挖掘)更好，更快地训练这些网络。补充材料和项目页面可在https://github.com/xialeiliu/RankIQA
# 2. Related work
我们简要回顾了与我们的方法相关的文献。我们专注于失真通用 NR-IQA,因为它比其他 IQA 研究线更普遍适用。Traditional NR-IQA approaches. 大多数传统的 NR-IQA 可以分为自然场景统计 (NSS)方法和基于学习的方法。在 NSS 方法中，假设不同质量的图像在对特定滤波器的响应统计中有所不同。小波 [20] 、DCT [23] 和 Curvelets [18] 常用于提取不同子带的特征。这些特征分布是参数化的，例如使用广义高斯分布$\mathbb{E}[26]$ 。这些方法的目的是估计分布参数，从中可以推断出质量评估。「19] 的作者建议在空间域中提取NSS 特征，以获得显著的加速。在基于学习的方法中， 使用例如支持机器回归或神经网络$\mathbb{E}[4]$ 提取局部特征并将其映射到 MOS。码本方法[40,39]结合了不同的特征，而不是直接使用局部特征。没有 MOS 的数据集可以通过无监督学习来构建码本，这一点特别重要，因为现有数据集的大小很小。显著性图[43]可用于对人类视觉系统进行建模，并提高这些方法的精度
Deep learning for NR-IQA. 近年来，一些作品将深度学习用于 NR-IQA$\mathbb{E}[1,11,12]$ 。深度网络的主要缺点之一是需要大型标记数据集，而这些数据集目前不适用于 NR-IQA 研究。为了解决这个问题，Kang 等人。[11]考虑小的$32\times32$补丁而不是图像，从而大大增加了训练示例的数量。[2,12]的作者遵循相同的管道。在$\mathbb{E}[12]$ 中，作者设计了一个多任务 CNN,以同时学
习失真类型和图像质量。Biancoatal。[1] 建议使用预先训练的网络来缓解训练数据的缺乏。他们从在 IQA 数据集上微调的预训练模型中提取特征。然后使用这些特征来训练 SVR 模型，以将特征映射到 IQA 分数
在我们的论文中，我们提出了一种完全不同的方法来解决缺乏训练数据的问题：我们使用大量自动生成的图像质量排名来训练深度网络。这使我们能够训练比 NR-IQA 中直接在绝对 IQA 数据上训练的其他方法更深人和更广泛的网络。Learning to rank. 这些方法通过最小化排名损失来从地面实况排名中学习排名功能$\mathbb{E}[3]$ 。然后可以将此函数应用于对测试对象进行排名。[25] 的作者采用随机梯度下降方法进行成对学习来排名。这已成功应用于大型数据集。结合排名和 CNN 的思想，暹罗网络架构在人脸验证问题$\overline{\mathbb{E}}[5]$ 以及比较图像补丁$\operatorname{lE}[41]$ 方面取得了巨大的成功。在NR-IQA 上下文中应用排名的唯一其他工作是$\mathbb{E} [ 6]$ , 其中它们结合了不同的手工特征来表示来自 IQA 数据集的图像对
我们的方法是不同的，因为我们主要不是为了学习排名。相反，我们使用从排名中学习作为数据增强技术：我们使用易于获得的排名图像数据集来训练大型网络，然后针对 NR-IQA 的任务进行微调
Hard-negative mining for Siamese network training. 众所周知，对 Siamese 网络进行采样对的朴素方法是次优的。为了解决这个问题，已经提出了几种硬负挖掘的方法。在$\mathbb{E}[$29] 中，他们提出了一种硬正和硬负挖掘策略，以正向传播一组对，并通过反向传播对最高损失对进行采样。然而，硬挖掘带来了很高的计算成本 (他们报告增加了高达 80% 的总计算成本)。在 [24] 中，他们提出了半硬对选择，认为选择最难的对会导致糟糕的局部极小值。使用的批量大小约为 1800 个示例， 这再次导致计算成本的显着增加。在[33]中，作者以一批对作为输入。开在小批量中洗择则个最难的价相本。为了解决一个糟糕的局部最优，[31] 优化了一个平滑的上限损失函数，以利用小批量中所有可能的对
与这些工作相反，我们提出了一种有效的连体反向传播方法，该方法不依赖于硬负选择。相反，它考虑了小批量中所有可能的对。这样做的优点是，可以最佳地利用训练深度网络中的主要计算瓶颈，即图像通过网络的前向传播
3. Learning from rankings for NR-IQA
在本节中，我们将描述我们利用合成生成的 NRIQA 排名的方法。我们首先为我们的方法制定一个通用框架，然后描述我们如何使用暹罗网络架构从排名中学习。最后，在E3.3 节中，我们展示了如何使从排序样本中训练暹罗网络的反向传播效率大大提高
# 3.1. Overview of our approach
缺乏大型 IQA 数据集促使我们提出一种新策略，以利用大型，未标记的数据库，从中可以生成按图像质量排序的图像。我们的方法基于以下观察：给定一组任意
|
[
[
0.08251953125,
0.0947265625,
0.46728515625,
0.0947265625,
0.46728515625,
0.10479736328125,
0.08251953125,
0.10479736328125
],
[
0.50341796875,
0.09344482421875,
0.890625,
0.09344482421875,
0.890625,
0.106689453125,
0.50341796875,
0.106689453125
],
[
0.08416748046875,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.08416748046875,
0.12060546875
],
[
0.5048828125,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.5048828125,
0.121826171875
],
[
0.08416748046875,
0.125,
0.46728515625,
0.125,
0.46728515625,
0.1351318359375,
0.08416748046875,
0.1351318359375
],
[
0.5048828125,
0.125,
0.8896484375,
0.125,
0.8896484375,
0.1351318359375,
0.5048828125,
0.1351318359375
],
[
0.08087158203125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50341796875,
0.1383056640625,
0.8837890625,
0.1383056640625,
0.8837890625,
0.1514892578125,
0.50341796875,
0.1514892578125
],
[
0.08251953125,
0.15283203125,
0.46728515625,
0.15283203125,
0.46728515625,
0.1666259765625,
0.08251953125,
0.1666259765625
],
[
0.525390625,
0.1546630859375,
0.8896484375,
0.1546630859375,
0.8896484375,
0.1654052734375,
0.525390625,
0.1654052734375
],
[
0.08416748046875,
0.169189453125,
0.465576171875,
0.169189453125,
0.465576171875,
0.179931640625,
0.08416748046875,
0.179931640625
],
[
0.50341796875,
0.1685791015625,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.08416748046875,
0.1849365234375,
0.465576171875,
0.1849365234375,
0.465576171875,
0.1956787109375,
0.08416748046875,
0.1956787109375
],
[
0.50341796875,
0.184326171875,
0.890625,
0.184326171875,
0.890625,
0.1976318359375,
0.50341796875,
0.1976318359375
],
[
0.08416748046875,
0.199462890625,
0.465576171875,
0.199462890625,
0.465576171875,
0.210205078125,
0.08416748046875,
0.210205078125
],
[
0.5048828125,
0.2008056640625,
0.88623046875,
0.2008056640625,
0.88623046875,
0.2108154296875,
0.5048828125,
0.2108154296875
],
[
0.080078125,
0.212158203125,
0.468017578125,
0.21337890625,
0.468017578125,
0.2266845703125,
0.080078125,
0.2259521484375
],
[
0.50341796875,
0.2139892578125,
0.890625,
0.2139892578125,
0.890625,
0.2279052734375,
0.50341796875,
0.2279052734375
],
[
0.08251953125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.08251953125,
0.242431640625
],
[
0.50341796875,
0.23291015625,
0.890625,
0.23291015625,
0.890625,
0.246826171875,
0.50341796875,
0.246826171875
],
[
0.08416748046875,
0.24560546875,
0.46728515625,
0.24560546875,
0.46728515625,
0.255615234375,
0.08416748046875,
0.255615234375
],
[
0.50634765625,
0.25,
0.8896484375,
0.25,
0.8896484375,
0.260009765625,
0.50634765625,
0.260009765625
],
[
0.08087158203125,
0.25830078125,
0.46728515625,
0.2587890625,
0.46728515625,
0.272705078125,
0.08087158203125,
0.271484375
],
[
0.5048828125,
0.261962890625,
0.890625,
0.261962890625,
0.890625,
0.27587890625,
0.5048828125,
0.27587890625
],
[
0.08251953125,
0.274658203125,
0.46728515625,
0.274658203125,
0.46728515625,
0.28857421875,
0.08251953125,
0.28857421875
],
[
0.50341796875,
0.27783203125,
0.8896484375,
0.27783203125,
0.8896484375,
0.291748046875,
0.50341796875,
0.291748046875
],
[
0.08087158203125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.50341796875,
0.292236328125,
0.88623046875,
0.292236328125,
0.88623046875,
0.30615234375,
0.50341796875,
0.30615234375
],
[
0.08251953125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.08251953125,
0.317626953125
],
[
0.50341796875,
0.30810546875,
0.890625,
0.30810546875,
0.890625,
0.3212890625,
0.50341796875,
0.3212890625
],
[
0.08087158203125,
0.3193359375,
0.46728515625,
0.3193359375,
0.46728515625,
0.333251953125,
0.08087158203125,
0.333251953125
],
[
0.50341796875,
0.323974609375,
0.8798828125,
0.323974609375,
0.8798828125,
0.337158203125,
0.50341796875,
0.337158203125
],
[
0.08087158203125,
0.333984375,
0.46728515625,
0.333984375,
0.46728515625,
0.347900390625,
0.08087158203125,
0.347900390625
],
[
0.50341796875,
0.337158203125,
0.890625,
0.33837890625,
0.890625,
0.352783203125,
0.50341796875,
0.3515625
],
[
0.08251953125,
0.35107421875,
0.16259765625,
0.35107421875,
0.16259765625,
0.3623046875,
0.08251953125,
0.3623046875
],
[
0.50634765625,
0.354736328125,
0.5849609375,
0.354736328125,
0.5849609375,
0.3662109375,
0.50634765625,
0.3662109375
],
[
0.1004638671875,
0.3662109375,
0.46728515625,
0.3662109375,
0.46728515625,
0.380126953125,
0.1004638671875,
0.380126953125
],
[
0.52392578125,
0.369384765625,
0.8896484375,
0.369384765625,
0.8896484375,
0.38330078125,
0.52392578125,
0.38330078125
],
[
0.08087158203125,
0.380615234375,
0.464111328125,
0.380615234375,
0.464111328125,
0.39453125,
0.08087158203125,
0.39453125
],
[
0.50244140625,
0.38330078125,
0.88818359375,
0.3837890625,
0.88818359375,
0.3984375,
0.50244140625,
0.397705078125
],
[
0.08087158203125,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.4091796875,
0.08087158203125,
0.4091796875
],
[
0.5048828125,
0.40087890625,
0.88818359375,
0.40087890625,
0.88818359375,
0.410888671875,
0.5048828125,
0.410888671875
],
[
0.08251953125,
0.410888671875,
0.468017578125,
0.410888671875,
0.468017578125,
0.4248046875,
0.08251953125,
0.4248046875
],
[
0.5048828125,
0.4140625,
0.8037109375,
0.4140625,
0.8037109375,
0.427978515625,
0.5048828125,
0.427978515625
],
[
0.08416748046875,
0.427978515625,
0.46728515625,
0.427978515625,
0.46728515625,
0.438232421875,
0.08416748046875,
0.438232421875
],
[
0.50341796875,
0.434326171875,
0.88818359375,
0.434326171875,
0.88818359375,
0.447509765625,
0.50341796875,
0.447509765625
],
[
0.08416748046875,
0.44140625,
0.46728515625,
0.44140625,
0.46728515625,
0.455078125,
0.08416748046875,
0.455078125
],
[
0.5048828125,
0.45068359375,
0.88818359375,
0.45068359375,
0.88818359375,
0.4609375,
0.5048828125,
0.4609375
],
[
0.08251953125,
0.458251953125,
0.46728515625,
0.458251953125,
0.46728515625,
0.468505859375,
0.08251953125,
0.468505859375
],
[
0.5048828125,
0.464111328125,
0.890625,
0.464111328125,
0.890625,
0.477783203125,
0.5048828125,
0.477783203125
],
[
0.080078125,
0.472900390625,
0.358642578125,
0.472900390625,
0.358642578125,
0.486083984375,
0.080078125,
0.486083984375
],
[
0.5048828125,
0.479248046875,
0.890625,
0.479248046875,
0.890625,
0.492431640625,
0.5048828125,
0.492431640625
],
[
0.50341796875,
0.492431640625,
0.8896484375,
0.494384765625,
0.8896484375,
0.50830078125,
0.50341796875,
0.505859375
],
[
0.080078125,
0.5048828125,
0.22216796875,
0.5048828125,
0.22216796875,
0.51806640625,
0.080078125,
0.51806640625
],
[
0.5048828125,
0.5087890625,
0.890625,
0.5087890625,
0.890625,
0.52294921875,
0.5048828125,
0.52294921875
],
[
0.50244140625,
0.5234375,
0.88818359375,
0.5244140625,
0.88818359375,
0.53857421875,
0.50244140625,
0.537109375
],
[
0.101318359375,
0.53076171875,
0.46728515625,
0.53076171875,
0.46728515625,
0.54443359375,
0.101318359375,
0.54443359375
],
[
0.50341796875,
0.53857421875,
0.890625,
0.5390625,
0.890625,
0.55322265625,
0.50341796875,
0.5517578125
],
[
0.08087158203125,
0.546875,
0.468017578125,
0.546875,
0.468017578125,
0.56005859375,
0.08087158203125,
0.56005859375
],
[
0.08251953125,
0.56103515625,
0.468017578125,
0.56103515625,
0.468017578125,
0.57470703125,
0.08251953125,
0.57470703125
],
[
0.50634765625,
0.55615234375,
0.88623046875,
0.55615234375,
0.88623046875,
0.56640625,
0.50634765625,
0.56640625
],
[
0.5048828125,
0.5712890625,
0.88818359375,
0.5712890625,
0.88818359375,
0.58203125,
0.5048828125,
0.58203125
],
[
0.08251953125,
0.580078125,
0.46728515625,
0.580078125,
0.46728515625,
0.59326171875,
0.08251953125,
0.59326171875
],
[
0.50341796875,
0.583984375,
0.88818359375,
0.583984375,
0.88818359375,
0.59765625,
0.50341796875,
0.59765625
],
[
0.08251953125,
0.59521484375,
0.46728515625,
0.59521484375,
0.46728515625,
0.609375,
0.08251953125,
0.609375
],
[
0.5048828125,
0.60107421875,
0.88818359375,
0.60107421875,
0.88818359375,
0.6142578125,
0.5048828125,
0.6142578125
],
[
0.08416748046875,
0.6123046875,
0.465576171875,
0.6123046875,
0.465576171875,
0.62255859375,
0.08416748046875,
0.62255859375
],
[
0.5048828125,
0.6162109375,
0.861328125,
0.6162109375,
0.861328125,
0.626953125,
0.5048828125,
0.626953125
],
[
0.0931396484375,
0.62548828125,
0.468017578125,
0.62548828125,
0.468017578125,
0.6396484375,
0.0931396484375,
0.6396484375
],
[
0.525390625,
0.6298828125,
0.8896484375,
0.6298828125,
0.8896484375,
0.64306640625,
0.525390625,
0.64306640625
],
[
0.08087158203125,
0.64013671875,
0.46728515625,
0.64013671875,
0.46728515625,
0.65380859375,
0.08087158203125,
0.65380859375
],
[
0.5048828125,
0.64599609375,
0.88623046875,
0.64599609375,
0.88623046875,
0.6591796875,
0.5048828125,
0.6591796875
],
[
0.08251953125,
0.65576171875,
0.46728515625,
0.65576171875,
0.46728515625,
0.66943359375,
0.08251953125,
0.66943359375
],
[
0.5048828125,
0.66015625,
0.8896484375,
0.66015625,
0.8896484375,
0.673828125,
0.5048828125,
0.673828125
],
[
0.08251953125,
0.67041015625,
0.46728515625,
0.67041015625,
0.46728515625,
0.6845703125,
0.08251953125,
0.6845703125
],
[
0.50341796875,
0.6748046875,
0.8896484375,
0.67529296875,
0.8896484375,
0.689453125,
0.50341796875,
0.68798828125
],
[
0.08251953125,
0.6875,
0.46728515625,
0.6875,
0.46728515625,
0.69775390625,
0.08251953125,
0.69775390625
],
[
0.5048828125,
0.6923828125,
0.60107421875,
0.6923828125,
0.60107421875,
0.7041015625,
0.5048828125,
0.7041015625
],
[
0.08416748046875,
0.70263671875,
0.465576171875,
0.70263671875,
0.465576171875,
0.71337890625,
0.08416748046875,
0.71337890625
],
[
0.08251953125,
0.71630859375,
0.46728515625,
0.71630859375,
0.46728515625,
0.72998046875,
0.08251953125,
0.72998046875
],
[
0.50341796875,
0.7216796875,
0.8349609375,
0.7216796875,
0.8349609375,
0.736328125,
0.50341796875,
0.736328125
],
[
0.08087158203125,
0.73095703125,
0.46728515625,
0.73095703125,
0.46728515625,
0.744140625,
0.08087158203125,
0.744140625
],
[
0.08251953125,
0.7470703125,
0.468017578125,
0.7470703125,
0.468017578125,
0.76025390625,
0.08251953125,
0.76025390625
],
[
0.521484375,
0.7470703125,
0.890625,
0.74560546875,
0.890625,
0.7587890625,
0.521484375,
0.76025390625
],
[
0.08251953125,
0.76123046875,
0.468017578125,
0.76123046875,
0.468017578125,
0.7744140625,
0.08251953125,
0.7744140625
],
[
0.50341796875,
0.76220703125,
0.8896484375,
0.76220703125,
0.8896484375,
0.77587890625,
0.50341796875,
0.77587890625
],
[
0.08087158203125,
0.7763671875,
0.406982421875,
0.7763671875,
0.406982421875,
0.79052734375,
0.08087158203125,
0.79052734375
],
[
0.50634765625,
0.77880859375,
0.88818359375,
0.77880859375,
0.88818359375,
0.7890625,
0.50634765625,
0.7890625
],
[
0.50341796875,
0.79052734375,
0.8896484375,
0.79150390625,
0.8896484375,
0.80615234375,
0.50341796875,
0.8046875
],
[
0.08087158203125,
0.796875,
0.46728515625,
0.796875,
0.46728515625,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.08087158203125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82568359375,
0.08087158203125,
0.82568359375
],
[
0.50244140625,
0.806640625,
0.84619140625,
0.806640625,
0.84619140625,
0.82080078125,
0.50244140625,
0.82080078125
],
[
0.08251953125,
0.8271484375,
0.465576171875,
0.8271484375,
0.465576171875,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.50341796875,
0.833984375,
0.74267578125,
0.833984375,
0.74267578125,
0.84765625,
0.50341796875,
0.84765625
],
[
0.08251953125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08251953125,
0.85498046875
],
[
0.08251953125,
0.85693359375,
0.46728515625,
0.85693359375,
0.46728515625,
0.87060546875,
0.08251953125,
0.87060546875
],
[
0.52392578125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.52392578125,
0.87060546875
],
[
0.0849609375,
0.87255859375,
0.4697265625,
0.87255859375,
0.4697265625,
0.88623046875,
0.0849609375,
0.88623046875
],
[
0.5048828125,
0.87255859375,
0.8896484375,
0.87255859375,
0.8896484375,
0.88623046875,
0.5048828125,
0.88623046875
],
[
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
]
] |
[
"越大的注释数据集。然而，IQA图像数据集的注释过",
"习失真类型和图像质量。Biancoatal。I]建议使用预",
"程需要每个图像的多个人工注释，因此收集过程是非",
"先训练的网络来缓解训练数据的缺乏。他们从在IQA",
"常劳动密集型和昂贵的。因此，大多数可用的IQA数",
"数据集上微调的预训练模型中提取特征。然后使用这",
"据集太小，无法有效地训练CNN。我们提出了一种解",
"些特征来训练SVR模型，以将特征映射到IQA分数",
"决大型数据集缺失的方法。主要思想（见图。E1）是，",
"在我们的论文中，我们提出了一种完全不同的方法",
"尽管难以获得带有人类注释的IQA数据，但很容易生",
"来解决缺乏训练数据的问题：我们使用大量自动生成",
"成根据其图像质量进行排名的图像。也就是说，我们可",
"的图像质量排名来训练深度网络。这使我们能够训练",
"以生成图像集，其中，尽管我们没有每个生成的图像的",
"比NR-IQA中直接在绝对IQA数据上训练的其他方",
"绝对质量度量，但对于我们知道的任何一对图像，哪个",
"法更深人和更广泛的网络。Learningtorank．这些",
"图像的质量更高。我们将这种从排名中学习的方法称",
"方法通过最小化排名损失来从地面实况排名中学习排",
"为RankIQA，并通过它学使用SiameseNetworks对",
"名功能E3。然后可以将此函数应用于对测试对象进",
"图像的质量进行排名，然后将从排名图像中学到的知",
"行排名。[25]的作者采用随机梯度下降方法进行成对",
"识转移到对IQA数据进行微调的传统CNN中，以提",
"学习来排名。这已成功应用于大型数据集。结合排名",
"高IQA的准确性。Zhang等人提出了从失真参考图像",
"和CNN的思想，逼罗网络架构在人脸验证问题E[5",
"中学习IQA特征的想法。专利「44。在本文中，我们",
"以及比较图像补丁E[41]方面取得了巨大的成功。在",
"超越了这项专利，因为我们提供了我们的方法的详细",
"NR-IQA上下文中应用排名的唯一其他工作是E[6]",
"描述，并通过实验验证了使用排名数据集的预训练网",
"其中它们结合了不同的手工特征来表示来自IQA数据",
"络的有用性",
"集的图像对",
"作为第二个贡献，我们提出了一种在遥罗网络中进",
"我们的方法是不同的，因为我们主要不是为了学习",
"行有效反向传播的方法。该方法通过单个网络转发一",
"排名。相反，我们使用从排名中学习作为数据增强技",
"批图像，然后反向传播从批次中的所有对导出的梯度。",
"术：我们使用易于获得的排名图像数据集来训练大型",
"在对已建立的IQA数据集进行的广泛实验中，我们表",
"网络，然后针对NR-IQA的任务进行微调",
"明从排名中学习可以显着改善结果，并且我们高效的",
"Hard-negative mining for Siamese network training",
"反向传播算法可以比其他训练协议（例如硬负挖掘）更",
"众所周知，对Siamese网络进行采样对的朴素方法",
"好，更快地训练这些网络。。补充材料和项目页面可在",
"是次优的。为了解决这个问题，已经提出了几种硬负挖",
"https://github.com/xialeiliu/RankIQA",
"掘的方法。在E[29]中，他们提出了一种硬正和硬负挖",
"掘策略，以正向传播一组对，并通过反向传播对最高",
"2.Related work",
"损失对进行采样。然而，硬挖掘带来了很高的计算成",
"本（他们报告增加了高达80%的总计算成本。在[24]",
"我们简要回顾了与我们的方法相关的文献。我们专",
"中，他们提出了半硬对选择，认为选择最难的对会导致",
"注于失真通用NR-IQA，因为它比其他IQA研究线更",
"普遍适用。TraditionalNR-IQAapproaches.大多数",
"槽糕的局部极小值。使用的批量小约为1800个示例",
"这再次导致计算成本的显着增加。在33中，作者以",
"传统的NR-IQA可以分为自然场景统计（NSS）方法和",
"一批对作为输入，并在小批量中选择四个最难的负样",
"基于学习的方法。在NSS方法中，假设不同质量的图",
"本。为了解决一个糟糕的局部最优，31]优化了一个平",
"像在对特定滤波器的响应统计中有所不同。小波20",
"滑的上限损失函数，以利用小批量中所有可能的对",
"DCT[23]和Curvelets[18]常用于提取不同子带的",
"与这些工作相反，我们提出了一种有效的连体反向",
"特征。这些特征分布是参数化的，例如使用广义高斯",
"传播方法，该方法不依赖于硬负选择。相反，它考虑了",
"分布[26]。这些方法的目的是估计分布参数，从中可",
"小批量中所有可能的对。这样做的优点是，可以最佳地",
"以推断出质量评估。［19]的作者建议在空间域中提取",
"利用训练深度网络中的主要计算瓶颈，即图像通过网",
"NSS特征，以获得显著的加速。在基于学习的方法中",
"络的前向传播",
"使用例如支持机器回归或神经网络E4提取局部特征",
"并将其映射到MOS。码本方法【40，39]结合了不同的",
"3. Learning from rankings for NR-IQA",
"特征，而不是直接使用局部特征。没有MOS的数据集",
"可以通过无监督学习来构建码本，这一点特别重要，因",
"在本节中，我们将描述我们利用合成生成的NR-",
"为现有数据集的大小很小。显著性图「43]可用于对人",
"IQA排名的方法。我们首先为我们的方法制定一个通",
"类视觉系统进行建模，并提高这些方法的精度",
"用框架，然后描述我们如何使用逼罗网络架构从排名",
"中学习。最后，在E3.3节中，我们展示了如何使从排",
"DeeplearningforNR-IQA近年来，一些作品将深度",
"学习用于NR-IQAE[1,11,12]。深度网络的主要缺点",
"序样本中训练暹罗网络的反向传播效率大大提高",
"之一是需要大型标记数据集，而这些数据集目前不适",
"3.1. Overview of our approach",
"用于NR-IQA研究。为了解决这个问题，Kang等人。",
"[11]考虑小的32×32补丁而不是图像，从而大大增加",
"缺乏大型IQA数据集促使我们提出一种新策略，以",
"了训练示例的数量。[2，12]的作者遵循相同的管道。",
"利用大型，未标记的数据库，从中可以生成按图像质量",
"在E[12]中，作者设计了一个多任务CNN，以同时学",
"排序的图像。我们的方法基于以下观察：给定一组任意"
] |
{
"arxiv_id": "1707.08347",
"language": "zh",
"page_idx": 1
}
|
||
1707.08347-zh-2
|
1707.08347-zh-2.jpg
|
参考图像，很容易应用图像失真来生成排名图像数据集。例如，给定参考图像，我们可以应用各种级别的高斯模糊。这样生成的图像集可以很容易地排序，因为我们知道添加高斯模糊(或任何其他失真)总是会降低质量分数。请注意，在这样的一组排名图像中，我们没有任何图像的绝对 IQA 分数-但我们知道任何一对图像的质量更高
在学习了这些排名图像之后，我们可以对小图像质量数据集进行微调，以解决 IQA 问题。我们的方法与简单的经典方法 [12] 之间的区别如图所示。1。标准方法直接在 IQA 数据集上训练一个浅层网络，从图像中估计 IQA 分数。由于数据有限，只能使用很少的层， 这限制了准确性。由于我们可以访问具有排名图像的更大数据集，因此我们现在可以训练更深更宽的网络来学习距离嵌人。接下来，我们通过微调域适应绝对IQA 问题来遵循这一点。我们的方法的整个管道是：
1. Synthesize ranked images. Using an arbitrary set of
images, we synthetically generate deformations of these images over a range of distortion intensities. The absolute distortion amount for each image is not used in subsequent steps, but within each deformation type we know, for any pair of images, which is of higher quality. See section 4.1 for a description of the datasets used for generating ranked images and the distortions applied.
2. Train Siamese network for ranking. Using the
set of ranked images, we train a Siamese network described in the next section using the efficient Siamese backpropagation technique proposed in section 3.3. The result is a Siamese network that ranks images by image quality.
3. Fine-tune on IQA data. Finally, we extract a single
branch of the Siamese network (we are interested at this point in the representation learned in the network, and not in the ranking itself), and fine-tune it on available IQA data. This effectively calibrates the network to output IQA measurements.
# 3.2. Siamese networks for ranking
这里我们介绍暹罗网络$\mathbb{E}[5]$ ,以从图像排名中学习这是一个具有两个相同网络分支和一个丢失模块的网络。两个分支在训练期间共享权重。成对的图像和标签是网络的输入，产生两个输出，这些输出被传递到丢失模块。通过反向传播计算损失函数相对于所有模型参数的梯度，并用随机梯度法更新。具体来说，给定图像$x$作为网络的输入，通过捕获最后一层中的激活来获得$x$的输出特征表示，由$f(x;\theta)$表示。这里$\theta$是网络参数，我们将使用$y$表示图像的真实值，对于图像质量评估，它是其质量分数。因此，在我们的暹罗网络中， 最终层的输出始终是我们想要指示图像质量的单个标量。由于我们的目标是对图像进行排名，因此我们应用
成对排名铰链损失：
(1)
$$L(x_1,x_2;\theta)=\max{(0,f(x_2;\theta)-f(x_1;\theta)+\varepsilon)}$$
,其中$\varepsilon$是边距。在这里，我们不失一般性地假设$x_1$
的等级高于$x_2$ 。方程 1 中的损耗梯度由下式给出：
$$\nabla_\theta L=\begin{cases}0&\text{if}f\left(x_2;\theta\right)-f\left(x_1;\theta\right)+\varepsilon\leq0\\\nabla_\theta f\left(x_2;\theta\right)-\nabla_\theta f\left(x_1;\theta\right)&\text{otherwise.}\end{cases}$$
(2)
换句话说，当网络的结果与排名一致时，梯度为零。当网络的结果不一致时，我们降低较高分数的梯度并添加较低分数的梯度。给定$L$相对于模型参数$\theta$的梯度， 我们可以使用随机梯度下降(SGD)训练暹罗网络
# 3.3. Efficient Siamese backpropagation
暹罗网络的一个缺点是冗余计算。考虑从三个图像构造的所有可能的图像对。在标准实现中，所有三个图像都通过网络传递两次，因为它们每个都出现在两对。由于暹罗网络的两个分支是相同的，我们基本上做了两倍的工作，因为任何图像只需要通过网络传递一次。我们正是利用这个想法来提高反向传播的效率。事实上，没有什么能阻止我们考虑小批量中所有可能的对， 几乎没有任何额外的计算。我们在网络中添加一个新层，在计算损失之前，在网络末尾的小批量中生成所有可能的对。这消除了对选择的问题，并提高了效率
要估计有效的暹罗反向传播的速度，请考虑以下因素。如果我们的训练集中有一个参考图像因$n$失真水平而失真，那么对于暹罗网络的传统实现，我们必须通过网络传递总共$n^2-n$图像-这是您可以使用$n$图像生成的对数量的两倍。相反，我们建议仅一次传递所有图像，并仅在损失计算层中考虑所有可能的对。这将计算减少到仅$n$通过网络。因此，加速等于：$\frac n2-nn=n-1$ 。在最佳方案$n=M$中，其中$M$等于小批量中的图像数量，因此此方法的加速将按小批量大小的顺序进行。由于小批量中所有对的集合之间的高度相关性，我们预计收敛的最终加速比将更低
为了简化下面的表示法，我们让$\hat{y}_i=f(x_i;\theta)$ ,其中$f(x_i;\theta)$是输入$x_i$上的暹罗网络的单个分支的输出。现在，对于一对输入，损失为：
$$L(x_1,x_2,l_{12};\theta)=g(\hat y_1,\hat y_2,l_{12}),$$
(3)
其中$l_{12}$是一个标签，指示图像 1 和 2 之间的关系 (例如，对于排名，它指示$x_1$的等级是否高于$x_2$ )。为简单起见，我们省略了$g$中的$\theta$依赖项。此损失函数相对于模型参数$\theta$的梯度为：
$$\nabla_\theta L=\frac{\partial g(\hat{y}_1,\hat{y}_2,l_{12})}{\partial\hat{y}_1}\nabla_\theta\hat{y}_1+\frac{\partial g(\hat{y}_1,\hat{y}_2,l_{12})}{\partial\hat{y}_2}\nabla_\theta\hat{y}_2.$$
上述$L$的梯度是总和，因为模型参数在暹罗网络的两个分支之间共享，并且$\hat{y}_1$和$\hat{y}_2$使用完全相同的参数计算
|
[
[
0.08251953125,
0.09344482421875,
0.468017578125,
0.09344482421875,
0.468017578125,
0.106689453125,
0.08251953125,
0.106689453125
],
[
0.5048828125,
0.0947265625,
0.6396484375,
0.0947265625,
0.6396484375,
0.10479736328125,
0.5048828125,
0.10479736328125
],
[
0.08251953125,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.08251953125,
0.12060546875
],
[
0.08087158203125,
0.12371826171875,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.08087158203125,
0.136962890625
],
[
0.52783203125,
0.121826171875,
0.890625,
0.121826171875,
0.890625,
0.1363525390625,
0.52783203125,
0.1363525390625
],
[
0.08251953125,
0.14013671875,
0.46728515625,
0.14013671875,
0.46728515625,
0.15087890625,
0.08251953125,
0.15087890625
],
[
0.08251953125,
0.154052734375,
0.46728515625,
0.154052734375,
0.46728515625,
0.1673583984375,
0.08251953125,
0.1673583984375
],
[
0.515625,
0.149658203125,
0.8896484375,
0.15087890625,
0.8896484375,
0.1654052734375,
0.515625,
0.1641845703125
],
[
0.5087890625,
0.1572265625,
0.5185546875,
0.1572265625,
0.5185546875,
0.160400390625,
0.5087890625,
0.160400390625
],
[
0.08087158203125,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.08087158203125,
0.1817626953125
],
[
0.50244140625,
0.1654052734375,
0.87158203125,
0.1654052734375,
0.87158203125,
0.1787109375,
0.50244140625,
0.1787109375
],
[
0.08251953125,
0.1849365234375,
0.16259765625,
0.1849365234375,
0.16259765625,
0.1962890625,
0.08251953125,
0.1962890625
],
[
0.580078125,
0.1956787109375,
0.5947265625,
0.1956787109375,
0.5947265625,
0.205810546875,
0.580078125,
0.205810546875
],
[
0.6611328125,
0.1932373046875,
0.82666015625,
0.1932373046875,
0.82666015625,
0.20703125,
0.6611328125,
0.20703125
],
[
0.576171875,
0.1988525390625,
0.5810546875,
0.1988525390625,
0.5810546875,
0.203857421875,
0.576171875,
0.203857421875
],
[
0.101318359375,
0.2020263671875,
0.46728515625,
0.2020263671875,
0.46728515625,
0.212158203125,
0.101318359375,
0.212158203125
],
[
0.5361328125,
0.1962890625,
0.58837890625,
0.2052001953125,
0.583984375,
0.2203369140625,
0.53173828125,
0.212158203125
],
[
0.86767578125,
0.2020263671875,
0.890625,
0.2020263671875,
0.890625,
0.2139892578125,
0.86767578125,
0.2139892578125
],
[
0.57666015625,
0.208984375,
0.79150390625,
0.208984375,
0.79150390625,
0.222900390625,
0.57666015625,
0.222900390625
],
[
0.08416748046875,
0.216552734375,
0.46728515625,
0.216552734375,
0.46728515625,
0.2266845703125,
0.08416748046875,
0.2266845703125
],
[
0.08087158203125,
0.2298583984375,
0.468017578125,
0.2298583984375,
0.468017578125,
0.24365234375,
0.08087158203125,
0.24365234375
],
[
0.50341796875,
0.2374267578125,
0.890625,
0.2374267578125,
0.890625,
0.250732421875,
0.50341796875,
0.250732421875
],
[
0.08087158203125,
0.2442626953125,
0.468017578125,
0.24365234375,
0.468017578125,
0.25830078125,
0.08087158203125,
0.2587890625
],
[
0.5048828125,
0.251953125,
0.890625,
0.251953125,
0.890625,
0.26513671875,
0.5048828125,
0.26513671875
],
[
0.08251953125,
0.260009765625,
0.468017578125,
0.260009765625,
0.468017578125,
0.2734375,
0.08251953125,
0.2734375
],
[
0.5048828125,
0.267578125,
0.890625,
0.267578125,
0.890625,
0.281005859375,
0.5048828125,
0.281005859375
],
[
0.08087158203125,
0.27587890625,
0.468017578125,
0.27587890625,
0.468017578125,
0.2890625,
0.08087158203125,
0.2890625
],
[
0.50244140625,
0.2822265625,
0.86376953125,
0.281005859375,
0.86376953125,
0.29541015625,
0.50244140625,
0.296630859375
],
[
0.08416748046875,
0.292236328125,
0.46728515625,
0.292236328125,
0.46728515625,
0.302978515625,
0.08416748046875,
0.302978515625
],
[
0.08087158203125,
0.30615234375,
0.468017578125,
0.30615234375,
0.468017578125,
0.3193359375,
0.08087158203125,
0.3193359375
],
[
0.50244140625,
0.306884765625,
0.80810546875,
0.309326171875,
0.80810546875,
0.323974609375,
0.50244140625,
0.3212890625
],
[
0.08087158203125,
0.3212890625,
0.449462890625,
0.3212890625,
0.449462890625,
0.335205078125,
0.08087158203125,
0.335205078125
],
[
0.52392578125,
0.333251953125,
0.890625,
0.333251953125,
0.890625,
0.3466796875,
0.52392578125,
0.3466796875
],
[
0.080078125,
0.349853515625,
0.4697265625,
0.349853515625,
0.4697265625,
0.3642578125,
0.080078125,
0.3642578125
],
[
0.50634765625,
0.349853515625,
0.8896484375,
0.349853515625,
0.8896484375,
0.35986328125,
0.50634765625,
0.35986328125
],
[
0.098876953125,
0.365478515625,
0.471435546875,
0.3642578125,
0.471435546875,
0.37890625,
0.098876953125,
0.380126953125
],
[
0.50341796875,
0.363037109375,
0.890625,
0.363037109375,
0.890625,
0.376953125,
0.50341796875,
0.376953125
],
[
0.101318359375,
0.380126953125,
0.46728515625,
0.37890625,
0.46728515625,
0.393310546875,
0.101318359375,
0.39453125
],
[
0.50341796875,
0.37744140625,
0.88818359375,
0.37744140625,
0.88818359375,
0.391357421875,
0.50341796875,
0.391357421875
],
[
0.10296630859375,
0.395751953125,
0.468017578125,
0.395751953125,
0.468017578125,
0.4091796875,
0.10296630859375,
0.4091796875
],
[
0.5048828125,
0.393310546875,
0.890625,
0.393310546875,
0.890625,
0.4072265625,
0.5048828125,
0.4072265625
],
[
0.101318359375,
0.410888671875,
0.465576171875,
0.410888671875,
0.465576171875,
0.4248046875,
0.101318359375,
0.4248046875
],
[
0.50634765625,
0.410400390625,
0.8896484375,
0.410400390625,
0.8896484375,
0.42041015625,
0.50634765625,
0.42041015625
],
[
0.101318359375,
0.425537109375,
0.468017578125,
0.425537109375,
0.468017578125,
0.439453125,
0.101318359375,
0.439453125
],
[
0.50634765625,
0.4248046875,
0.8896484375,
0.4248046875,
0.8896484375,
0.43505859375,
0.50634765625,
0.43505859375
],
[
0.101318359375,
0.44140625,
0.46728515625,
0.44140625,
0.46728515625,
0.455078125,
0.101318359375,
0.455078125
],
[
0.5048828125,
0.438232421875,
0.890625,
0.438232421875,
0.890625,
0.451904296875,
0.5048828125,
0.451904296875
],
[
0.1004638671875,
0.455078125,
0.46728515625,
0.45703125,
0.46728515625,
0.47021484375,
0.1004638671875,
0.468505859375
],
[
0.5048828125,
0.453857421875,
0.8896484375,
0.453857421875,
0.8896484375,
0.46728515625,
0.5048828125,
0.46728515625
],
[
0.101318359375,
0.4716796875,
0.301513671875,
0.4716796875,
0.301513671875,
0.48486328125,
0.101318359375,
0.48486328125
],
[
0.50341796875,
0.46728515625,
0.86279296875,
0.468505859375,
0.86279296875,
0.481689453125,
0.50341796875,
0.48095703125
],
[
0.52392578125,
0.48486328125,
0.890625,
0.48486328125,
0.890625,
0.498779296875,
0.52392578125,
0.498779296875
],
[
0.0767822265625,
0.492431640625,
0.39453125,
0.49365234375,
0.39453125,
0.50830078125,
0.0767822265625,
0.5068359375
],
[
0.38818359375,
0.49560546875,
0.46728515625,
0.49560546875,
0.46728515625,
0.5068359375,
0.38818359375,
0.5068359375
],
[
0.5048828125,
0.5,
0.8896484375,
0.5,
0.8896484375,
0.51318359375,
0.5048828125,
0.51318359375
],
[
0.101318359375,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.101318359375,
0.52294921875
],
[
0.5048828125,
0.5146484375,
0.890625,
0.5146484375,
0.890625,
0.52783203125,
0.5048828125,
0.52783203125
],
[
0.10296630859375,
0.5244140625,
0.46728515625,
0.5244140625,
0.46728515625,
0.53857421875,
0.10296630859375,
0.53857421875
],
[
0.5048828125,
0.52880859375,
0.890625,
0.52880859375,
0.890625,
0.54345703125,
0.5048828125,
0.54345703125
],
[
0.1004638671875,
0.53857421875,
0.46728515625,
0.5390625,
0.46728515625,
0.5537109375,
0.1004638671875,
0.55322265625
],
[
0.5048828125,
0.54443359375,
0.890625,
0.54443359375,
0.890625,
0.55810546875,
0.5048828125,
0.55810546875
],
[
0.10296630859375,
0.55615234375,
0.46728515625,
0.55615234375,
0.46728515625,
0.56640625,
0.10296630859375,
0.56640625
],
[
0.5048828125,
0.56005859375,
0.890625,
0.56005859375,
0.890625,
0.57373046875,
0.5048828125,
0.57373046875
],
[
0.1004638671875,
0.568359375,
0.27783203125,
0.5693359375,
0.27783203125,
0.58349609375,
0.1004638671875,
0.58203125
],
[
0.50244140625,
0.57568359375,
0.890625,
0.57373046875,
0.890625,
0.58837890625,
0.50244140625,
0.59033203125
],
[
0.080078125,
0.59228515625,
0.46728515625,
0.59228515625,
0.46728515625,
0.60595703125,
0.080078125,
0.60595703125
],
[
0.51708984375,
0.59033203125,
0.890625,
0.59033203125,
0.890625,
0.60498046875,
0.51708984375,
0.60498046875
],
[
0.1004638671875,
0.60791015625,
0.468017578125,
0.60791015625,
0.468017578125,
0.6220703125,
0.1004638671875,
0.6220703125
],
[
0.5048828125,
0.60595703125,
0.890625,
0.60595703125,
0.890625,
0.619140625,
0.5048828125,
0.619140625
],
[
0.101318359375,
0.62255859375,
0.468017578125,
0.62255859375,
0.468017578125,
0.63623046875,
0.101318359375,
0.63623046875
],
[
0.50341796875,
0.6220703125,
0.89208984375,
0.6220703125,
0.89208984375,
0.63525390625,
0.50341796875,
0.63525390625
],
[
0.10296630859375,
0.63818359375,
0.468017578125,
0.63818359375,
0.468017578125,
0.6513671875,
0.10296630859375,
0.6513671875
],
[
0.50341796875,
0.63623046875,
0.73193359375,
0.63623046875,
0.73193359375,
0.6494140625,
0.50341796875,
0.6494140625
],
[
0.1004638671875,
0.6513671875,
0.468017578125,
0.65283203125,
0.468017578125,
0.66650390625,
0.1004638671875,
0.66552734375
],
[
0.52392578125,
0.65283203125,
0.89208984375,
0.65283203125,
0.89208984375,
0.66650390625,
0.52392578125,
0.66650390625
],
[
0.101318359375,
0.66845703125,
0.404296875,
0.66845703125,
0.404296875,
0.681640625,
0.101318359375,
0.681640625
],
[
0.50341796875,
0.66650390625,
0.890625,
0.66748046875,
0.890625,
0.681640625,
0.50341796875,
0.68115234375
],
[
0.5048828125,
0.68310546875,
0.7236328125,
0.68310546875,
0.7236328125,
0.6962890625,
0.5048828125,
0.6962890625
],
[
0.07843017578125,
0.69580078125,
0.348876953125,
0.6962890625,
0.348876953125,
0.7109375,
0.07843017578125,
0.71044921875
],
[
0.5888671875,
0.7119140625,
0.80615234375,
0.7119140625,
0.80615234375,
0.72607421875,
0.5888671875,
0.72607421875
],
[
0.81787109375,
0.71533203125,
0.8486328125,
0.71533203125,
0.8486328125,
0.7236328125,
0.81787109375,
0.7236328125
],
[
0.86767578125,
0.7119140625,
0.890625,
0.7119140625,
0.890625,
0.72607421875,
0.86767578125,
0.72607421875
],
[
0.1004638671875,
0.72021484375,
0.46728515625,
0.72021484375,
0.46728515625,
0.734375,
0.1004638671875,
0.734375
],
[
0.08087158203125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.08087158203125,
0.75
],
[
0.50244140625,
0.7392578125,
0.88818359375,
0.7392578125,
0.88818359375,
0.75244140625,
0.50244140625,
0.75244140625
],
[
0.08087158203125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.08087158203125,
0.76513671875
],
[
0.50244140625,
0.75390625,
0.890625,
0.7548828125,
0.890625,
0.76953125,
0.50244140625,
0.76806640625
],
[
0.08251953125,
0.76806640625,
0.46728515625,
0.76806640625,
0.46728515625,
0.77880859375,
0.08251953125,
0.77880859375
],
[
0.50244140625,
0.76953125,
0.88818359375,
0.76953125,
0.88818359375,
0.78271484375,
0.50244140625,
0.78271484375
],
[
0.08087158203125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.08087158203125,
0.79541015625
],
[
0.5048828125,
0.787109375,
0.69140625,
0.787109375,
0.69140625,
0.79736328125,
0.5048828125,
0.79736328125
],
[
0.080078125,
0.79541015625,
0.46728515625,
0.796875,
0.46728515625,
0.81103515625,
0.080078125,
0.81005859375
],
[
0.08087158203125,
0.8125,
0.468017578125,
0.8125,
0.468017578125,
0.82568359375,
0.08087158203125,
0.82568359375
],
[
0.5595703125,
0.80810546875,
0.89453125,
0.81005859375,
0.89453125,
0.8388671875,
0.5595703125,
0.8369140625
],
[
0.5107421875,
0.82080078125,
0.57275390625,
0.814453125,
0.576171875,
0.830078125,
0.51318359375,
0.8359375
],
[
0.08087158203125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.60107421875,
0.83056640625,
0.6298828125,
0.83056640625,
0.6298828125,
0.84521484375,
0.60107421875,
0.84521484375
],
[
0.75634765625,
0.83056640625,
0.78515625,
0.83056640625,
0.78515625,
0.84521484375,
0.75634765625,
0.84521484375
],
[
0.08087158203125,
0.84228515625,
0.468017578125,
0.84228515625,
0.468017578125,
0.85595703125,
0.08087158203125,
0.85595703125
],
[
0.08251953125,
0.859375,
0.46728515625,
0.859375,
0.46728515625,
0.869140625,
0.08251953125,
0.869140625
],
[
0.50634765625,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50634765625,
0.87060546875
],
[
0.08087158203125,
0.87255859375,
0.468017578125,
0.87255859375,
0.468017578125,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.50244140625,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50244140625,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88818359375,
0.5361328125,
0.88818359375,
0.5361328125,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"参考图像，很容易应用图像失真来生成排名图像数据",
"成对排名铰链损失：",
"集。例如，给定参考图像，我们可以应用各种级别的高",
"斯模糊。这样生成的图像集可以很容易地排序，因为我",
"L(1,c2;0) =max(0, f(c2;0)-f(a1;0) +)(1)",
"们知道添加高斯模糊（或任何其他失真）总是会降低质",
"量分数。请注意，在这样的一组排名图像中，我们没有",
"其中是边距。在这里，我们不失一般性地假设1",
"",
"任何图像的绝对IQA分数-但我们知道任何一对图像",
"的等级高于a2。方程1中的损耗梯度由下式给出：",
"的质量更高",
"10",
"if f(2;0)-f(1;0)+=≤0",
"",
"在学习了这些排名图像之后，我们可以对小图像质",
"VeL=",
"(2)",
"[Vef(r2;0)-Vef(ri;0)otherwise",
"量数据集进行微调，以解决IQA问题。我们的方法与",
"简单的经典方法[12]之间的区别如图所示。1。标准",
"换句话说，当网络的结果与排名一致时，梯度为零。当",
"方法直接在IQA数据集上训练一个浅层网络，从图像",
"网络的结果不一致时，我们降低较高分数的梯度并添",
"中估计IQA分数。由于数据有限，只能使用很少的层，",
"加较低分数的梯度。给定L相对于模型参数θ的梯度，",
"这限制了准确性。由于我们可以访问具有排名图像的",
"我们可以使用随机梯度下降（SGD）训练逼罗网络",
"更大数据集，因此我们现在可以训练更深更宽的网络",
"来学习距离嵌入。接下来，我们通过微调域适应绝对",
"3.3.Efficient Siamese backpropagation",
"IQA问题来遵循这一点。我们的方法的整个管道是：",
"暹罗网络的一个缺点是穴余计算。考虑从三个图像",
"1. Synthesize ranked images. Using an arbitrary set of",
"构造的所有可能的图像对。在标准实现中，所有三个图",
"images, we synthetically generate deformations of",
"像都通过网络传递两次，因为它们每个都出现在两对。",
"these images over a range of distortion intensities.",
"由于暹罗网络的两个分支是相同的，我们基本上做了",
"The absolute distortion amount for each image is",
"两倍的工作，因为任何图像只需要通过网络传递一次。",
"not used in subsequent steps, but within each defor-",
"我们正是利用这个想法来提高反向传播的效率。事实",
"mation type we know, for any pair of images, which",
"上，没有什么能阻正我们考虑小批量中所有可能的对，",
"is of higher quality. See section 4.1 for a description",
"几乎没有任何额外的计算。我们在网络中添加一个新",
"of the datasets used for generating ranked images",
"层，在计算损失之前，在网络末尾的小批量中生成所有",
"and the distortions applied.",
"可能的对。这消除了对选择的问题，并提高了效率",
"要估计有效的逼罗反向传播的速度，请考虑以下因",
"2.Train Siamese network for ranking.",
"Using the",
"素。如果我们的训练集中有一个参考图像因n失真水",
"set of ranked images, we train a Siamese network",
"平而失真，那么对于暹罗网络的传统实现，我们必须通",
"described in the next section using the efficient",
"过网络传递总共n?-n图像-这是您可以使用n图像生",
"Siamese backpropagation technique proposed in sec-",
"成的对数量的两倍。相反，我们建议仅一次传递所有图",
"tion 3.3.Theresult isaSiamesenetworkthatranks",
"像，并仅在损失计算层中考虑所有可能的对。这将计算",
"images by image quality.",
"减少到仅n通过网络。因此，加速等于：n-n=n-1",
"3. Fine-tune on IQA data. Finally, we extract a single",
"在最佳方案n=M中，其中M等于小批量中的图",
"branch of the Siamese network (we are interested",
"像数量，因此此方法的加速将按小批量大小的顺序进",
"at this point in the representation learned in the",
"行。由于小批量中所有对的集合之间的高度相关性，我",
"network, and not in the ranking itself), and fine-tune",
"们预计收敛的最终加速比将更低",
"it on available IQA data. This effectively calibrates",
"为了简化下面的表示法，我们让9i=f(ai；の），其",
"the network to output IQA measurements",
"中f(i0）是输人ci上的暹罗网络的单个分支的输出。",
"现在，对于一对输入，损失为：",
"3.2. Siamese networks for ranking",
"L(1, 2,l12; 0)= g(91,92,l12),",
"",
"(3)",
"这里我们介绍暹罗网络E51，以从图像排名中学习",
"这是一个具有两个相同网络分支和一个丢失模块的网",
"其中112是一个标签，指示图像1和2之间的关系（例",
"络。两个分支在训练期间共享权重。成对的图像和标签",
"如，对于排名，它指示a1的等级是否高于a2）。为简",
"是网络的输入，产生两个输出，这些输出被传递到丢失",
"单起见，我们省略了9中的θ依赖项。此损失函数相",
"模块。通过反向传播计算损失函数相对于所有模型参",
"对于模型参数9的梯度为：",
"数的梯度，并用随机梯度法更新。具体来说，给定图像",
"作为网络的输入，通过捕获最后一层中的激活来获得",
"0g(01,02 12) V091 + 9(01 02, 12) o02. (4)",
"VeL=",
"的输出特征表示，由f(;の）表示。这里是网络参",
"0g1",
"092",
"数，我们将使用9表示图像的真实值，对于图像质量",
"评估，它是其质量分数。因此，在我们的逼罗网络中，",
"上述L的梯度是总和，因为模型参数在暹罗网络的两",
"最终层的输出始终是我们想要指示图像质量的单个标",
"个分支之间共享，并且91和92使用完全相同的参数",
"量。由于我们的目标是对图像进行排名，因此我们应用",
"计算"
] |
{
"arxiv_id": "1707.08347",
"language": "zh",
"page_idx": 2
}
|
||
1707.08347-zh-3
|
1707.08347-zh-3.jpg
|
考虑到小批量 Xall 中的所有对，损失为：
(5)
$$L=\sum\limits_{i=1}^M\sum\limits_{j>i}^Mg(\hat{y}_i,\hat{y}_j,l_{ij})$$
小批量损失相对于参数$\theta$的梯度可以写为：
$$\nabla_{\theta}L=\sum_{i=1}^{M}\sum_{j>i}^{M}\frac{\partial g(\hat y_{i},\hat y_{j},l_{ij})}{\partial\hat y_{i}}\nabla_{\theta}\hat y_{i}+\frac{\partial g(\hat y_{i},\hat y_{j},l_{ij})}{\partial\hat y_{j}}\nabla_{\theta}\hat y_{j}$$
我们现在可以用矩阵形式将小批量的损失函数的梯度
表示为：<NEXXX>。对于标准的单分支网络，我们将平均所有批次样本的梯度，以获得小批量的梯度。这等效于将$P$设置为上面公式$\mathbb{E}??$中的单位矩阵。对于考虑小批量中所有对的暹罗网络，我们通过将$P$设置为：
$$P=\begin{bmatrix}0&\frac{\partial g(\hat{y}_1,\hat{y}_2,l_{12})}{\partial\hat{y}_1}&\cdots&\frac{\partial g(\hat{y}_1,\hat{y}_M,l_{1M})}{\partial\hat{y}_1}\\\frac{\partial g(\hat{y}_1,\hat{y}_2,l_{12})}{\partial\hat{y}_2}&0&\cdots&\frac{\partial g(\hat{y}_2,\hat{y}_M,l_{2M})}{\partial\hat{y}_2}\\\vdots&\vdots&\ddots&\vdots\\\frac{\partial g(\hat{y}_1,\hat{y}_M,l_{1M})}{\partial\hat{y}_M}&\cdots&\cdots&0\end{bmatrix}$$
(7)
来获得方程式$\color{red}{\mathrm{E6~。}}$对于排名铰链丢失的特定情况
将公式$\mathbb{E}$7 中的 $g$ 替换为公式$\color{red}{\mathrm{E}}\color{blue}{\mathrm{1}}$ ,并获得：
(8)
$$P=\begin{bmatrix}0&a_{12}&\cdots&a_{1M}\\a_{21}&0&\cdots&a_{2M}\\\vdots&\vdots&\ddots&\vdots\\a_{M1}&\cdots&a_{M(M-1)}&0\end{bmatrix},$$
,其中
$$a_{ij}=\begin{cases}0&\text{if}l_{ij}\left(\hat{y}_j-\hat{y}_i\right)+\varepsilon\leq0\\l_{ij}&\text{otherwise}\end{cases}$$
(9)
和$l_{ij}\in\{-1,0,1\}$ ,其中1 (-1)表示 XMATH。上述情况考虑了单个失真。相反，假设我们在训练集中有$D$类型的失真。然后，我们可以使用块对角矩阵将损失梯度计算为：
$$\nabla_\theta L=\left[\begin{array}{cccc}\nabla_\theta\hat{Y}^1\dots\nabla_\theta\hat{Y}^D\end{array}\right]\left[\begin{array}{cccc}A^1&0&\cdots&0\\0&A^2&\cdots&0\\\vdots&\vdots&\ddots&\vdots\\0&0&\cdots&A^d\end{array}\right]1_M$$
其中
$$\begin{aligned}A^{d}&=\quad\begin{bmatrix}0&a_{12}^d&\cdots&a_{1M^d}^d\\a_{21}^d&0&\cdots&a_{2M^d}^d\\\vdots&\vdots&\ddots&\vdots\\a_{M^d1}^d&\cdots&a_{M^d(M^d-1)}^d&0\end{bmatrix},\text{ar}\\\nabla_{\theta}\hat{Y}^{d}&=\quad\begin{bmatrix}\nabla_{\theta}\hat{y}_{d1}&\cdots&\nabla_{\theta}\hat{y}_{dM}\end{bmatrix}.\end{aligned}$$
其中$\hat{y}_{mn}$是指具有$m^{th}$失真的$n^{th}$图像的网络输出$d\in\{1,2,\ldots,D\}$和 XMATHLINX\_1。在上述$A^d$的定义中，$a_{ij}^d$ 是公式$\mathbb{E}$9 中的梯度系数
## $3. 4. \textbf{ Fine- tuning for NR- IQA}$
在训练一个暹罗网络对失真图像进行排序后，我们然后从网络中提取一个分支进行微调。给定具有人类IQA 测量的小批量$M$图像，我们将第$i$图像的地面实况质量得分表示为$y_i$ , 并且来自网络的预测得分为$\hat{y}_i$ ,如上所述。我们使用平方欧几里得距离作为损失函数来代替用于暹罗网络的排名损失来微调网络：
(12)
$$L(y_i,\hat{y}_i)=\frac{1}{M}\sum_{i=1}^M(y_i-\hat{y}_i)^2$$
4. Experimental results
在本节中，我们报告了一些旨在评估我们的方法在
IQA 基线和最新技术方面的性能的实验
## 4.1. Datasets
我们在实验中使用两种类型的数据集：通用的非IQA 数据集，用于生成排名图像以训练暹罗网络；IQA 数据集，用于微调和评估
IQA datasets. 我们在两个标准 IQA 基准数据集上进行实验。LIVE$\underline{\mathrm{E}}[28]$包含从 29 个原始图像生成的808 个图像，这些图像通过五种类型的失真进行失真： 高斯模糊(GB),高斯噪声(GN), JPEG 压缩 (JPEG), JPEG2000 压缩 (JP2K) 和快速衰落 (FF)。每个图像的地面实况平均意见得分在[0,100]范围内，并使用161 位人类注释者的注释进行估计。TID2013 [22] 数据集由 25 个参考图像组成，其中包含来自 24 种不同失真类型的 3000 个失真图像，具有 5 个退化级别。平均意见得分在[0,9]范围内。失真类型包括一系列噪声、压缩和传输伪影。有关特定失真类型的列表，请参见原始出版物
Datasets for generating ranked pairs. 为了在 LIVE 数据库上进行测试，我们生成了四种广泛使用且常见的失真类型：GB, GN, JPEG 和 JP2K。为了在 TID2013上进行测试，我们生成了总共 24 个失真中的 17 个(除了$\#3,\#4,\#12,\#13,\#20,\#21,\#24)$。对于我们无法产生的失真，我们从其他失真训练的网络中应用微调。这被发现产生令人满意的结果。我们使用两个数据集来生成排名图像对
滑铁卢数据集由 4,744 个从互联网上精心挑选的高质量自然图像组成。此外，我们使用 356 个场景类别的 Places2E[45]数据集的验证集。验证集中每个类别有 100 个图像，总共 36500 个图像。失真后，我们有更多的失真图像用于学习图像质量排名嵌入。使用此数据集的目的是证明可以使用不是专门为 IQA 问题设计的数据集来学习高质量的排名嵌入
|
[
[
0.101318359375,
0.0921630859375,
0.39208984375,
0.0921630859375,
0.39208984375,
0.1060791015625,
0.101318359375,
0.1060791015625
],
[
0.5009765625,
0.09027099609375,
0.890625,
0.0889892578125,
0.890625,
0.106689453125,
0.5009765625,
0.10797119140625
],
[
0.50341796875,
0.106689453125,
0.890625,
0.106689453125,
0.890625,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.226318359375,
0.1142578125,
0.27294921875,
0.1142578125,
0.27294921875,
0.1287841796875,
0.226318359375,
0.1287841796875
],
[
0.1871337890625,
0.126953125,
0.358642578125,
0.1300048828125,
0.358642578125,
0.147705078125,
0.1871337890625,
0.14453125
],
[
0.50244140625,
0.12249755859375,
0.77294921875,
0.121826171875,
0.77294921875,
0.1363525390625,
0.50244140625,
0.136962890625
],
[
0.44677734375,
0.1307373046875,
0.4697265625,
0.1307373046875,
0.4697265625,
0.14453125,
0.44677734375,
0.14453125
],
[
0.22216796875,
0.14453125,
0.275390625,
0.146484375,
0.273681640625,
0.1610107421875,
0.221435546875,
0.1590576171875
],
[
0.50244140625,
0.15283203125,
0.72802734375,
0.15283203125,
0.72802734375,
0.1666259765625,
0.50244140625,
0.1666259765625
],
[
0.08416748046875,
0.17236328125,
0.38232421875,
0.17236328125,
0.38232421875,
0.18310546875,
0.08416748046875,
0.18310546875
],
[
0.52392578125,
0.1767578125,
0.890625,
0.1767578125,
0.890625,
0.1900634765625,
0.52392578125,
0.1900634765625
],
[
0.136474609375,
0.186279296875,
0.339111328125,
0.20263671875,
0.3349609375,
0.2279052734375,
0.13232421875,
0.2108154296875
],
[
0.3349609375,
0.1925048828125,
0.468017578125,
0.20263671875,
0.464111328125,
0.2259521484375,
0.330810546875,
0.21533203125
],
[
0.5048828125,
0.1925048828125,
0.890625,
0.1925048828125,
0.890625,
0.205810546875,
0.5048828125,
0.205810546875
],
[
0.131591796875,
0.20263671875,
0.1854248046875,
0.20263671875,
0.1854248046875,
0.2266845703125,
0.131591796875,
0.2266845703125
],
[
0.08251953125,
0.208984375,
0.1380615234375,
0.208984375,
0.1380615234375,
0.2203369140625,
0.08251953125,
0.2203369140625
],
[
0.50341796875,
0.20703125,
0.890625,
0.20703125,
0.890625,
0.2203369140625,
0.50341796875,
0.2203369140625
],
[
0.22216796875,
0.2171630859375,
0.2484130859375,
0.2171630859375,
0.2484130859375,
0.2310791015625,
0.22216796875,
0.2310791015625
],
[
0.369384765625,
0.216552734375,
0.39794921875,
0.216552734375,
0.39794921875,
0.23291015625,
0.369384765625,
0.23291015625
],
[
0.136474609375,
0.2247314453125,
0.1871337890625,
0.2247314453125,
0.1871337890625,
0.236083984375,
0.136474609375,
0.236083984375
],
[
0.50244140625,
0.2215576171875,
0.8896484375,
0.222900390625,
0.8896484375,
0.2374267578125,
0.50244140625,
0.236083984375
],
[
0.44677734375,
0.2374267578125,
0.468017578125,
0.2374267578125,
0.468017578125,
0.25,
0.44677734375,
0.25
],
[
0.5107421875,
0.238037109375,
0.890625,
0.238037109375,
0.890625,
0.251953125,
0.5107421875,
0.251953125
],
[
0.08251953125,
0.25390625,
0.465576171875,
0.25390625,
0.465576171875,
0.264404296875,
0.08251953125,
0.264404296875
],
[
0.50244140625,
0.25244140625,
0.81884765625,
0.25244140625,
0.81884765625,
0.266357421875,
0.50244140625,
0.266357421875
],
[
0.08251953125,
0.266357421875,
0.46728515625,
0.266357421875,
0.46728515625,
0.27978515625,
0.08251953125,
0.27978515625
],
[
0.67724609375,
0.2734375,
0.80615234375,
0.2890625,
0.79833984375,
0.326904296875,
0.6689453125,
0.311279296875
],
[
0.0849609375,
0.28271484375,
0.465576171875,
0.28271484375,
0.465576171875,
0.29345703125,
0.0849609375,
0.29345703125
],
[
0.7158203125,
0.27783203125,
0.7294921875,
0.28857421875,
0.7158203125,
0.298583984375,
0.70166015625,
0.28857421875
],
[
0.5986328125,
0.29345703125,
0.67333984375,
0.29541015625,
0.67333984375,
0.311279296875,
0.5986328125,
0.309326171875
],
[
0.859375,
0.294189453125,
0.89208984375,
0.294189453125,
0.89208984375,
0.31005859375,
0.859375,
0.31005859375
],
[
0.68115234375,
0.296630859375,
0.73388671875,
0.311279296875,
0.7255859375,
0.32763671875,
0.67333984375,
0.313232421875
],
[
0.08087158203125,
0.296630859375,
0.46728515625,
0.296630859375,
0.46728515625,
0.31005859375,
0.08087158203125,
0.31005859375
],
[
0.08087158203125,
0.3125,
0.468017578125,
0.3125,
0.468017578125,
0.32568359375,
0.08087158203125,
0.32568359375
],
[
0.240234375,
0.3359375,
0.3447265625,
0.337158203125,
0.3447265625,
0.348388671875,
0.240234375,
0.347900390625
],
[
0.34716796875,
0.337158203125,
0.441162109375,
0.337158203125,
0.441162109375,
0.348388671875,
0.34716796875,
0.348388671875
],
[
0.50244140625,
0.34033203125,
0.708984375,
0.34033203125,
0.708984375,
0.354248046875,
0.50244140625,
0.354248046875
],
[
0.265625,
0.3466796875,
0.2900390625,
0.3466796875,
0.2900390625,
0.359130859375,
0.265625,
0.359130859375
],
[
0.178955078125,
0.339111328125,
0.19775390625,
0.339111328125,
0.19775390625,
0.352783203125,
0.178955078125,
0.352783203125
],
[
0.381591796875,
0.3466796875,
0.406982421875,
0.3466796875,
0.406982421875,
0.35791015625,
0.381591796875,
0.35791015625
],
[
0.148681640625,
0.354248046875,
0.2296142578125,
0.354248046875,
0.2296142578125,
0.36865234375,
0.148681640625,
0.36865234375
],
[
0.31787109375,
0.352783203125,
0.441162109375,
0.35107421875,
0.441162109375,
0.369384765625,
0.31787109375,
0.371826171875
],
[
0.271240234375,
0.35986328125,
0.28515625,
0.35986328125,
0.28515625,
0.371826171875,
0.271240234375,
0.371826171875
],
[
0.17724609375,
0.3662109375,
0.200927734375,
0.3662109375,
0.200927734375,
0.37744140625,
0.17724609375,
0.37744140625
],
[
0.381591796875,
0.3662109375,
0.406005859375,
0.3662109375,
0.406005859375,
0.37744140625,
0.381591796875,
0.37744140625
],
[
0.52392578125,
0.3662109375,
0.890625,
0.3662109375,
0.890625,
0.380126953125,
0.52392578125,
0.380126953125
],
[
0.271240234375,
0.380126953125,
0.285888671875,
0.380126953125,
0.285888671875,
0.397705078125,
0.271240234375,
0.397705078125
],
[
0.098876953125,
0.369384765625,
0.1298828125,
0.369384765625,
0.1298828125,
0.3818359375,
0.098876953125,
0.3818359375
],
[
0.1829833984375,
0.3818359375,
0.1968994140625,
0.3818359375,
0.1968994140625,
0.392578125,
0.1829833984375,
0.392578125
],
[
0.324462890625,
0.380126953125,
0.3447265625,
0.385009765625,
0.340576171875,
0.396484375,
0.3203125,
0.391357421875
],
[
0.3896484375,
0.38330078125,
0.400390625,
0.38330078125,
0.400390625,
0.395751953125,
0.3896484375,
0.395751953125
],
[
0.50341796875,
0.3818359375,
0.78515625,
0.3818359375,
0.78515625,
0.395751953125,
0.50341796875,
0.395751953125
],
[
0.1446533203125,
0.397705078125,
0.233642578125,
0.397705078125,
0.233642578125,
0.4091796875,
0.1446533203125,
0.4091796875
],
[
0.26708984375,
0.404052734375,
0.285888671875,
0.404052734375,
0.285888671875,
0.412353515625,
0.26708984375,
0.412353515625
],
[
0.32275390625,
0.404052734375,
0.34228515625,
0.404052734375,
0.34228515625,
0.412353515625,
0.32275390625,
0.412353515625
],
[
0.3857421875,
0.40087890625,
0.40283203125,
0.40087890625,
0.40283203125,
0.4140625,
0.3857421875,
0.4140625
],
[
0.1748046875,
0.4091796875,
0.2034912109375,
0.4091796875,
0.2034912109375,
0.42041015625,
0.1748046875,
0.42041015625
],
[
0.50244140625,
0.410400390625,
0.60888671875,
0.410888671875,
0.60888671875,
0.425537109375,
0.50244140625,
0.4248046875
],
[
0.4453125,
0.417236328125,
0.468017578125,
0.417236328125,
0.468017578125,
0.431884765625,
0.4453125,
0.431884765625
],
[
0.08416748046875,
0.43310546875,
0.2484130859375,
0.43310546875,
0.2484130859375,
0.4462890625,
0.08416748046875,
0.4462890625
],
[
0.240234375,
0.43310546875,
0.4697265625,
0.43310546875,
0.4697265625,
0.4462890625,
0.240234375,
0.4462890625
],
[
0.52294921875,
0.436279296875,
0.890625,
0.436279296875,
0.890625,
0.449462890625,
0.52294921875,
0.449462890625
],
[
0.08087158203125,
0.447509765625,
0.39453125,
0.447509765625,
0.39453125,
0.4609375,
0.08087158203125,
0.4609375
],
[
0.50341796875,
0.451904296875,
0.890625,
0.451904296875,
0.890625,
0.46533203125,
0.50341796875,
0.46533203125
],
[
0.191162109375,
0.472900390625,
0.214111328125,
0.472900390625,
0.214111328125,
0.486083984375,
0.191162109375,
0.486083984375
],
[
0.50341796875,
0.46728515625,
0.6845703125,
0.46728515625,
0.6845703125,
0.48095703125,
0.50341796875,
0.48095703125
],
[
0.2344970703125,
0.474853515625,
0.2607421875,
0.474853515625,
0.2607421875,
0.488037109375,
0.2344970703125,
0.488037109375
],
[
0.35302734375,
0.473388671875,
0.390625,
0.473388671875,
0.390625,
0.4892578125,
0.35302734375,
0.4892578125
],
[
0.23779296875,
0.488037109375,
0.256591796875,
0.488037109375,
0.256591796875,
0.501953125,
0.23779296875,
0.501953125
],
[
0.50634765625,
0.488037109375,
0.60693359375,
0.488037109375,
0.60693359375,
0.498779296875,
0.50634765625,
0.498779296875
],
[
0.6298828125,
0.4873046875,
0.890625,
0.4873046875,
0.890625,
0.50048828125,
0.6298828125,
0.50048828125
],
[
0.3544921875,
0.490478515625,
0.3857421875,
0.490478515625,
0.3857421875,
0.50244140625,
0.3544921875,
0.50244140625
],
[
0.19287109375,
0.4892578125,
0.221435546875,
0.4892578125,
0.221435546875,
0.50390625,
0.19287109375,
0.50390625
],
[
0.295654296875,
0.4912109375,
0.316162109375,
0.4912109375,
0.316162109375,
0.498779296875,
0.295654296875,
0.498779296875
],
[
0.143798828125,
0.49755859375,
0.176513671875,
0.49755859375,
0.176513671875,
0.51318359375,
0.143798828125,
0.51318359375
],
[
0.44677734375,
0.50048828125,
0.4697265625,
0.50048828125,
0.4697265625,
0.5146484375,
0.44677734375,
0.5146484375
],
[
0.50341796875,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50341796875,
0.51513671875
],
[
0.23779296875,
0.5087890625,
0.2548828125,
0.5087890625,
0.2548828125,
0.5244140625,
0.23779296875,
0.5244140625
],
[
0.295654296875,
0.5087890625,
0.318603515625,
0.5087890625,
0.318603515625,
0.5234375,
0.295654296875,
0.5234375
],
[
0.365234375,
0.51123046875,
0.375732421875,
0.51123046875,
0.375732421875,
0.5234375,
0.365234375,
0.5234375
],
[
0.50341796875,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.53076171875,
0.50341796875,
0.53076171875
],
[
0.200927734375,
0.51123046875,
0.2132568359375,
0.51123046875,
0.2132568359375,
0.5234375,
0.200927734375,
0.5234375
],
[
0.27294921875,
0.52783203125,
0.34228515625,
0.52978515625,
0.34228515625,
0.54443359375,
0.27294921875,
0.54248046875
],
[
0.363525390625,
0.5263671875,
0.386474609375,
0.5263671875,
0.386474609375,
0.5390625,
0.363525390625,
0.5390625
],
[
0.2344970703125,
0.53076171875,
0.256591796875,
0.53076171875,
0.256591796875,
0.53857421875,
0.2344970703125,
0.53857421875
],
[
0.1854248046875,
0.52880859375,
0.223876953125,
0.52880859375,
0.223876953125,
0.54248046875,
0.1854248046875,
0.54248046875
],
[
0.50634765625,
0.5341796875,
0.8896484375,
0.5341796875,
0.8896484375,
0.54443359375,
0.50634765625,
0.54443359375
],
[
0.07843017578125,
0.55322265625,
0.131591796875,
0.55322265625,
0.131591796875,
0.568359375,
0.07843017578125,
0.568359375
],
[
0.50341796875,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.50341796875,
0.56103515625
],
[
0.50341796875,
0.56298828125,
0.8896484375,
0.56298828125,
0.8896484375,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.233642578125,
0.580078125,
0.396240234375,
0.58056640625,
0.396240234375,
0.5986328125,
0.233642578125,
0.59765625
],
[
0.5048828125,
0.57763671875,
0.890625,
0.57763671875,
0.890625,
0.59130859375,
0.5048828125,
0.59130859375
],
[
0.44677734375,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60400390625,
0.44677734375,
0.60400390625
],
[
0.200927734375,
0.58056640625,
0.22216796875,
0.58056640625,
0.22216796875,
0.5947265625,
0.200927734375,
0.5947265625
],
[
0.15283203125,
0.59326171875,
0.1871337890625,
0.59326171875,
0.1871337890625,
0.60595703125,
0.15283203125,
0.60595703125
],
[
0.50341796875,
0.59228515625,
0.890625,
0.59228515625,
0.890625,
0.60595703125,
0.50341796875,
0.60595703125
],
[
0.2017822265625,
0.6015625,
0.2303466796875,
0.6015625,
0.2303466796875,
0.6142578125,
0.2017822265625,
0.6142578125
],
[
0.2279052734375,
0.60107421875,
0.3095703125,
0.60107421875,
0.3095703125,
0.6123046875,
0.2279052734375,
0.6123046875
],
[
0.5048828125,
0.60791015625,
0.8896484375,
0.60791015625,
0.8896484375,
0.6220703125,
0.5048828125,
0.6220703125
],
[
0.50341796875,
0.62255859375,
0.890625,
0.62255859375,
0.890625,
0.63623046875,
0.50341796875,
0.63623046875
],
[
0.08251953125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.642578125,
0.08251953125,
0.642578125
],
[
0.50244140625,
0.63623046875,
0.890625,
0.63818359375,
0.890625,
0.65283203125,
0.50244140625,
0.65087890625
],
[
0.08087158203125,
0.64453125,
0.46728515625,
0.64453125,
0.46728515625,
0.65771484375,
0.08087158203125,
0.65771484375
],
[
0.08087158203125,
0.6591796875,
0.468017578125,
0.6591796875,
0.468017578125,
0.673828125,
0.08087158203125,
0.673828125
],
[
0.50634765625,
0.65380859375,
0.60107421875,
0.65380859375,
0.60107421875,
0.66552734375,
0.50634765625,
0.66552734375
],
[
0.08251953125,
0.67529296875,
0.1871337890625,
0.67529296875,
0.1871337890625,
0.68798828125,
0.08251953125,
0.68798828125
],
[
0.50244140625,
0.67236328125,
0.779296875,
0.673828125,
0.779296875,
0.68798828125,
0.50244140625,
0.6875
],
[
0.7939453125,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.68603515625,
0.7939453125,
0.68603515625
],
[
0.50341796875,
0.68798828125,
0.890625,
0.68798828125,
0.890625,
0.7021484375,
0.50341796875,
0.7021484375
],
[
0.283447265625,
0.69775390625,
0.3203125,
0.69775390625,
0.3203125,
0.7138671875,
0.283447265625,
0.7138671875
],
[
0.3349609375,
0.69970703125,
0.35302734375,
0.69970703125,
0.35302734375,
0.71337890625,
0.3349609375,
0.71337890625
],
[
0.36767578125,
0.7021484375,
0.38818359375,
0.7021484375,
0.38818359375,
0.71044921875,
0.36767578125,
0.71044921875
],
[
0.41015625,
0.69970703125,
0.42724609375,
0.69970703125,
0.42724609375,
0.71337890625,
0.41015625,
0.71337890625
],
[
0.50244140625,
0.70263671875,
0.89208984375,
0.7021484375,
0.89208984375,
0.71630859375,
0.50244140625,
0.71728515625
],
[
0.33251953125,
0.7138671875,
0.35693359375,
0.7138671875,
0.35693359375,
0.7265625,
0.33251953125,
0.7265625
],
[
0.365234375,
0.71337890625,
0.390625,
0.71630859375,
0.38818359375,
0.7265625,
0.36279296875,
0.7236328125
],
[
0.408447265625,
0.7138671875,
0.426513671875,
0.7138671875,
0.426513671875,
0.72802734375,
0.408447265625,
0.72802734375
],
[
0.50634765625,
0.71826171875,
0.890625,
0.71826171875,
0.890625,
0.73193359375,
0.50634765625,
0.73193359375
],
[
0.07843017578125,
0.724609375,
0.27294921875,
0.7236328125,
0.27294921875,
0.7412109375,
0.07843017578125,
0.7421875
],
[
0.298095703125,
0.71630859375,
0.313720703125,
0.71337890625,
0.31787109375,
0.72607421875,
0.30224609375,
0.72900390625
],
[
0.44775390625,
0.72802734375,
0.47216796875,
0.72802734375,
0.47216796875,
0.7412109375,
0.44775390625,
0.7412109375
],
[
0.301513671875,
0.73486328125,
0.31787109375,
0.73486328125,
0.31787109375,
0.751953125,
0.301513671875,
0.751953125
],
[
0.410888671875,
0.736328125,
0.42236328125,
0.736328125,
0.42236328125,
0.74853515625,
0.410888671875,
0.74853515625
],
[
0.50634765625,
0.734375,
0.890625,
0.734375,
0.890625,
0.74755859375,
0.50634765625,
0.74755859375
],
[
0.369384765625,
0.7373046875,
0.39453125,
0.7373046875,
0.39453125,
0.74853515625,
0.369384765625,
0.74853515625
],
[
0.338134765625,
0.73779296875,
0.351318359375,
0.73779296875,
0.351318359375,
0.75,
0.338134765625,
0.75
],
[
0.404296875,
0.751953125,
0.4306640625,
0.751953125,
0.4306640625,
0.76513671875,
0.404296875,
0.76513671875
],
[
0.5048828125,
0.75,
0.890625,
0.75,
0.890625,
0.76318359375,
0.5048828125,
0.76318359375
],
[
0.338134765625,
0.75390625,
0.351318359375,
0.75390625,
0.351318359375,
0.7646484375,
0.338134765625,
0.7646484375
],
[
0.301513671875,
0.75390625,
0.314453125,
0.75390625,
0.314453125,
0.76513671875,
0.301513671875,
0.76513671875
],
[
0.369384765625,
0.7548828125,
0.39208984375,
0.7548828125,
0.39208984375,
0.76318359375,
0.369384765625,
0.76318359375
],
[
0.43701171875,
0.76953125,
0.46728515625,
0.76953125,
0.46728515625,
0.7822265625,
0.43701171875,
0.7822265625
],
[
0.5048828125,
0.7646484375,
0.890625,
0.7646484375,
0.890625,
0.77783203125,
0.5048828125,
0.77783203125
],
[
0.080078125,
0.78271484375,
0.1168212890625,
0.78271484375,
0.1168212890625,
0.798828125,
0.080078125,
0.798828125
],
[
0.5048828125,
0.77978515625,
0.6845703125,
0.77978515625,
0.6845703125,
0.79345703125,
0.5048828125,
0.79345703125
],
[
0.52294921875,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.52294921875,
0.81005859375
],
[
0.2034912109375,
0.80615234375,
0.2156982421875,
0.81494140625,
0.2034912109375,
0.82470703125,
0.191162109375,
0.81494140625
],
[
0.2344970703125,
0.80810546875,
0.26318359375,
0.80810546875,
0.26318359375,
0.82470703125,
0.2344970703125,
0.82470703125
],
[
0.303955078125,
0.81005859375,
0.326904296875,
0.81005859375,
0.326904296875,
0.8212890625,
0.303955078125,
0.8212890625
],
[
0.36767578125,
0.80615234375,
0.416748046875,
0.80810546875,
0.4150390625,
0.8291015625,
0.366943359375,
0.8271484375
],
[
0.50341796875,
0.81103515625,
0.8896484375,
0.81103515625,
0.8896484375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.1895751953125,
0.82275390625,
0.2181396484375,
0.82275390625,
0.2181396484375,
0.8388671875,
0.1895751953125,
0.8388671875
],
[
0.240234375,
0.82470703125,
0.259033203125,
0.82470703125,
0.259033203125,
0.83837890625,
0.240234375,
0.83837890625
],
[
0.303955078125,
0.82470703125,
0.326904296875,
0.82470703125,
0.326904296875,
0.8359375,
0.303955078125,
0.8359375
],
[
0.369384765625,
0.82373046875,
0.408447265625,
0.82568359375,
0.406982421875,
0.84228515625,
0.36767578125,
0.84033203125
],
[
0.50244140625,
0.82470703125,
0.8896484375,
0.82568359375,
0.8896484375,
0.84033203125,
0.50244140625,
0.8388671875
],
[
0.108642578125,
0.83544921875,
0.16015625,
0.83349609375,
0.160888671875,
0.8486328125,
0.1094970703125,
0.85107421875
],
[
0.4248046875,
0.83837890625,
0.463134765625,
0.83837890625,
0.463134765625,
0.8515625,
0.4248046875,
0.8515625
],
[
0.50341796875,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.3056640625,
0.84326171875,
0.330078125,
0.84326171875,
0.330078125,
0.861328125,
0.3056640625,
0.861328125
],
[
0.1968994140625,
0.8447265625,
0.214111328125,
0.8447265625,
0.214111328125,
0.861328125,
0.1968994140625,
0.861328125
],
[
0.2442626953125,
0.84521484375,
0.256591796875,
0.84521484375,
0.256591796875,
0.859375,
0.2442626953125,
0.859375
],
[
0.1748046875,
0.85302734375,
0.2303466796875,
0.85498046875,
0.2296142578125,
0.88623046875,
0.17236328125,
0.88330078125
],
[
0.384033203125,
0.84765625,
0.39453125,
0.84765625,
0.39453125,
0.85986328125,
0.384033203125,
0.85986328125
],
[
0.275390625,
0.85986328125,
0.359375,
0.86865234375,
0.35693359375,
0.88720703125,
0.27294921875,
0.87890625
],
[
0.50244140625,
0.85498046875,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50244140625,
0.86865234375
],
[
0.2418212890625,
0.8662109375,
0.256591796875,
0.8662109375,
0.256591796875,
0.875,
0.2418212890625,
0.875
],
[
0.384033203125,
0.8642578125,
0.396240234375,
0.8642578125,
0.396240234375,
0.87548828125,
0.384033203125,
0.87548828125
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.08660888671875,
0.8837890625,
0.16259765625,
0.88330078125,
0.16259765625,
0.8984375,
0.08660888671875,
0.8994140625
],
[
0.17236328125,
0.8837890625,
0.2279052734375,
0.88720703125,
0.226318359375,
0.90283203125,
0.1707763671875,
0.8994140625
],
[
0.27783203125,
0.88525390625,
0.34228515625,
0.88720703125,
0.340576171875,
0.90283203125,
0.27685546875,
0.90087890625
],
[
0.43701171875,
0.88623046875,
0.4697265625,
0.88623046875,
0.4697265625,
0.9013671875,
0.43701171875,
0.9013671875
],
[
0.50341796875,
0.88720703125,
0.7646484375,
0.88720703125,
0.7646484375,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.240234375,
0.8896484375,
0.2607421875,
0.8896484375,
0.2607421875,
0.89794921875,
0.240234375,
0.89794921875
]
] |
[
"考虑到小批量Xall中的所有对，损失为：",
"其中9mn是指具有mth失真的nth图像的网络输出",
"d E{1,2,..,D) 和 XMATHLINX_1。在上述Ad 的",
"MM",
"L=g(9i,9j,li)",
"定义中，α,是公式E9中的梯度系数",
"(5)",
"i=lj>i",
"3.4. Fine-tuning for NR-IQA",
"小批量损失相对于参数0的梯度可以写为",
"在训练一个暹罗网络对失真图像进行排序后，我们",
"Mag(gila)o9i+",
"0g(9i,9,la)09",
"然后从网络中提取一个分支进行微调。给定具有人类",
"Z2",
"VeL=",
"IQA测量的小批量M图像，我们将第i图像的地面实",
"Oyi",
"og",
"=l j>i",
"况质量得分表示为9元，并且来自网络的预测得分为9元",
"(6)",
"如上所述。我们使用平方欧几里得距离作为损失函数",
"我们现在可以用矩阵形式将小批量的损失函数的梯度",
"来代替用于逼罗网络的排名损失来微调网络：",
"表示为：<NEXXX>。对于标准的单分支网络，我们将",
"-0)",
"平均所有批次样本的梯度，以获得小批量的梯度。这等",
"M",
"L(yi, 9i) ",
"(12)",
"ME",
"效于将P设置为上面公式E?？中的单位矩阵。对于考",
"虑小批量中所有对的暹罗网络，我们通过将P设置为",
"8g(91.92,l12)",
"Og(91,9M,l1m)",
"4.Experimental results",
"o91",
"0",
"091",
"0g(91.92.l12)",
"...g(02M.2m)",
"0",
"092",
"092",
"在本节中，我们报告了一些旨在评估我们的方法在",
"",
"P=",
":",
"..",
"..",
"IQA基线和最新技术方面的性能的实验",
"Og(91.9M,l1M)",
"...",
"...",
"0",
"Oum",
"4.1. Datasets",
"(7)",
"来获得方程式E6。。）",
"对于排名铰链丢失的特定情况，",
"我们在实验中使用两种类型的数据集：通用的非",
"将公式E7中的g替换为公式E1，并获得：",
"IQA数据集，用于生成排名图像以训练遥罗网络：IQA",
"0",
"数据集，用于微调和评估",
"a12",
"a1m",
"0",
"IQA datasets.",
"我们在两个标准IQA基准数据集上",
"a2M",
"a21",
"....",
"P =",
"(8)",
"进行实验。LIVEE[281包含从29个原始图像生成的",
":",
"...",
":",
"808个图像，这些图像通过五种类型的失真进行失真",
"：",
"aM(M-1)",
"0",
"..",
"[aMi",
"高斯模糊（GB），高斯噪声（GN），JPEG压缩（JPEG）",
"，其中",
"JPEG2000压缩（JP2K）和快速衰落（FF）。每个图像",
"的地面实况平均意见得分在[0，100]范围内，并使用",
"if li(9-9a)+e≤0",
"161位人类注释者的注释进行估计。TID2013[22]数",
"(9)",
"[o",
"aij:",
"据集由25个参考图像组成，其中包含来自24种不同",
"lii",
"otherwise",
"失真类型的3000个失真图像，具有5个退化级别。平",
"均意见得分在[0，9]范围内。失真类型包括一系列噪",
"和liE（-1,0,1)，其中1（-1）表示XMATH。上",
"声、压缩和传输伪影。有关特定失真类型的列表，请参",
"述情况考虑了单个失真。相反，假设我们在训练集中有",
"D类型的失真。然后，我们可以使用块对角矩阵将损",
"见原始出版物",
"失梯度计算为：",
"Datasets for generating ranked pairs.",
"为了在LIVE",
"数据库上进行测试，我们生成了四种广泛使用且常见的",
"[A1",
"0",
"...",
"0",
"失真类型：GB，GN，JPEG和JP2K。为了在TID2013",
"A2",
"...",
"0",
"上进行测试，我们生成了总共24个失真中的17个（除",
"VeL=[VeYl...VoYD",
"0",
"1M",
":",
":",
"了#3，#4，#12，#13，#20，#21，#24）。对于我",
"'",
"：",
"Ad",
"们无法产生的失真，我们从其他失真训练的网络中应",
"0",
"0",
"..",
"(10)",
"用微调。这被发现产生令人满意的结果。我们使用两个",
"其中",
"数据集来生成排名图像对",
"滑铁卢数据集由4,744个从互联网上精心挑选的高",
"0",
"af2",
"N",
"agma",
"质量自然图像组成。此外，我们使用356个场景类别",
"azi",
"0",
".",
"aaMd",
"的Places2E[45]数据集的验证集。验证集中每个类别",
"Ad =",
"and",
"有100个图像，总共36500个图像。失真后，我们有",
".",
"1",
":",
"Lamrai",
":",
"aMa(Md-1)",
"更多的失真图像用于学习图像质量排名嵌人。使用此",
".",
"0",
"数据集的目的是证明可以使用不是专门为IQA问题设",
"VeYd=",
"[Veyd]",
"Veydm]",
"(11)",
"计的数据集来学习高质量的排名嵌人",
".::"
] |
{
"arxiv_id": "1707.08347",
"language": "zh",
"page_idx": 3
}
|
||
1707.08347-zh-4
|
1707.08347-zh-4.jpg
|
<FigureHere>
Figure 2. Siamese network output for JPEG distortion
considering 6 levels. This graphs illustrate the fact that the Siamese network successfully manages to separate the different distortion levels.
## 4.2. Experimental protocols
我们调查了许多网络架构，并使用标准 IQA 指标来
评估性能
Network architectures.
我们评估了三种从浅到深的
典型网络架构。我们将它们称为：Shallow, AlexNet$\boxed{\mathrm{F}}$ [14]和 VGG-16$\mathbb{E}[30]$。浅层网络有四个卷积层和一个全连接层。对于 AlexNet 和 VGG-16,我们只改变输出的数量，因为我们的目标是为每个失真图像分配一个分数
Strategy for training and testing.
我们从原始高分
辨率图像中随机采样子图像。我们这样做而不是缩放， 以避免引入由插值或滤波引起的失真。采样图像的大小由每个网络确定。但是，输入图像的大尺寸很重要， 因为输入子图像应至少为原始图像的1/3,以便捕获上下文信息。这是补丁采样方法的严重限制，该方法从原始图像中采样非常小的 32$\times32$补丁。在我们的实验中，我们对$227\times227$和$224\times224$像素图像进行采样， 具体取决于网络。我们使用 Caffe$\underline{\mathbb{E}}[10]$框架，并使用小批量随机梯度下降(SGD)进行训练，初始学习率为1e-4,用于有效的暹罗网络训练，1e-6 用于微调。培训率降低了0倍。每 10K 次迭代 1次，总共 50K 次迭代。对于两个训练阶段，我们使用$\ell_{2}$权重衰减(权重5e-4)。在训练期间，我们从每个时期的每个训练图像中采样单个子图像
测试时，我们按照$\mathbb{E}[1]$ 中的建议从原始图像中随机采样 30 个子图像。子区域输出的平均值是每个图像的最终分数
Evaluation protocols. 传统上使用两个评估指标来评估 IQA 算法的性能：线性相关系数(LCC)和斯皮尔曼等级相关系数(SROCC)。LCC 是地面实况和预测质量分数之间的线性相关的度量。给定$N$失真图像， 第$i$图像的地面实况由$y_i$表示，网络的预测得分为$\hat{y}_i$ 。LCC 的计算公式为：
(13)
$$LCC=\frac{\sum_{i=1}^N(y_i-\overline{y})(\hat{y}_i-\overline{\hat{y}})}{\sqrt{\sum_i^N(y_i-\overline{y})^2}\sqrt{\sum_i^N(\hat{y}_i-\overline{\hat{y}})^2}}$$
<FigureHere>
Figure 3. Convergence of ranking loss on JPEG distortion for our approach versus standard Siamese and hardnegative mining.
<table>
<tbody>
<tr>
<th> </th>
<th>Shallow</th>
<th>AlexNet</th>
<th>$VGG-16$</th>
</tr>
<tr>
<td>$\overline{\mathrm{LCC}}$</td>
<td>0.930</td>
<td>0.945</td>
<td>0.973</td>
</tr>
<tr>
<td>SROCC</td>
<td>0.937</td>
<td>0.949</td>
<td>0.974</td>
</tr>
</tbody>
</table>
Table 1. SROCC and LCC for our approach on LIVE using
different networks.
其中$\bar{y}$和$\overline{\hat{y}}$分别是地面实况和预测质量得分的平均
值
给定$N$失真图像，SROCC 计算如下：
$$SROCC=1-\frac{6\sum_{i=1}^{N}\left(v_i-p_i\right)^2}{N\left(N^2-1\right)},$$
(14)
其中$v_i$是地面实况 IQA 分数$y_i$在地面实况分数中的排名，而$p_i$是$\hat{y}_i$在所有$N$的输出分数中的排名。SROCC 测量地面实况与估计的 IQA 之间的单调关系
# 4.3. Learning NR-IQA from rankings
我们进行了许多实验，以评估暹罗网络学习从大型图像质量排名数据集中捕获图像失真的能力。此外，我们测量了3.3节中描述的有效连体反向传播方法的影响
Siamese networks and IQA discrimination. 为了展示我们的排名网络区分图像质量的能力，我们在 Places2验证集上训练了我们的 Siamese 网络(没有对 IQA 数据进行微调),五个级别的单一失真。然后，我们使用该网络来预测来自滑铁卢数据集的合成失真图像的图像质量，这些图像使用相同的五个级别的相同失真。网络输出绘制为图中的直方图。2 用于 JPEG 失真$^{1}$。在图中，我们根据真实的失真水平(由直方图的颜色表示)划分观察值。即使两个数据集的获取过程和场景完全不同，该模型也可以区分滑铁卢上的不同程度的失真
Efficient Siamese backpropagation. 本实验的目的是评估我们的暹罗反向传播方法的效率。我们将我们的方法与标准随机对抽样和类似于 [29]的硬负挖掘方法
$^{1}$Graphs for the other distortions are in the supplementary
material.
|
[
[
0.419189453125,
0.10479736328125,
0.4248046875,
0.10479736328125,
0.4248046875,
0.121826171875,
0.419189453125,
0.121826171875
],
[
0.080078125,
0.2374267578125,
0.468017578125,
0.236083984375,
0.468017578125,
0.25,
0.080078125,
0.250732421875
],
[
0.50341796875,
0.24365234375,
0.583984375,
0.24365234375,
0.583984375,
0.255126953125,
0.50341796875,
0.255126953125
],
[
0.57666015625,
0.242431640625,
0.88818359375,
0.242431640625,
0.88818359375,
0.255615234375,
0.57666015625,
0.255615234375
],
[
0.080078125,
0.250732421875,
0.468017578125,
0.250732421875,
0.468017578125,
0.264404296875,
0.080078125,
0.264404296875
],
[
0.5009765625,
0.255615234375,
0.890625,
0.255126953125,
0.890625,
0.26953125,
0.5009765625,
0.270751953125
],
[
0.07843017578125,
0.26318359375,
0.471435546875,
0.264404296875,
0.471435546875,
0.279052734375,
0.07843017578125,
0.27783203125
],
[
0.50244140625,
0.270751953125,
0.61279296875,
0.271484375,
0.61279296875,
0.285400390625,
0.50244140625,
0.2841796875
],
[
0.08087158203125,
0.27978515625,
0.25,
0.27978515625,
0.25,
0.290283203125,
0.08087158203125,
0.290283203125
],
[
0.6337890625,
0.298583984375,
0.69140625,
0.298583984375,
0.69140625,
0.31005859375,
0.6337890625,
0.31005859375
],
[
0.705078125,
0.298583984375,
0.75732421875,
0.298583984375,
0.75732421875,
0.31005859375,
0.705078125,
0.31005859375
],
[
0.771484375,
0.298583984375,
0.830078125,
0.298583984375,
0.830078125,
0.31005859375,
0.771484375,
0.31005859375
],
[
0.5810546875,
0.3125,
0.61767578125,
0.3125,
0.61767578125,
0.322509765625,
0.5810546875,
0.322509765625
],
[
0.646484375,
0.3125,
0.6845703125,
0.3125,
0.6845703125,
0.323974609375,
0.646484375,
0.323974609375
],
[
0.71142578125,
0.3125,
0.7490234375,
0.3125,
0.7490234375,
0.323974609375,
0.71142578125,
0.323974609375
],
[
0.78125,
0.3125,
0.81884765625,
0.3125,
0.81884765625,
0.32568359375,
0.78125,
0.32568359375
],
[
0.57275390625,
0.324462890625,
0.6259765625,
0.324462890625,
0.6259765625,
0.3359375,
0.57275390625,
0.3359375
],
[
0.080078125,
0.317626953125,
0.2998046875,
0.318115234375,
0.2998046875,
0.333251953125,
0.080078125,
0.33203125
],
[
0.64794921875,
0.324462890625,
0.68115234375,
0.324462890625,
0.68115234375,
0.337158203125,
0.64794921875,
0.337158203125
],
[
0.71337890625,
0.324462890625,
0.74853515625,
0.324462890625,
0.74853515625,
0.337158203125,
0.71337890625,
0.337158203125
],
[
0.78125,
0.324462890625,
0.81884765625,
0.324462890625,
0.81884765625,
0.33837890625,
0.78125,
0.33837890625
],
[
0.1004638671875,
0.3447265625,
0.468017578125,
0.3447265625,
0.468017578125,
0.35791015625,
0.1004638671875,
0.35791015625
],
[
0.50244140625,
0.348388671875,
0.89208984375,
0.349853515625,
0.89208984375,
0.3642578125,
0.50244140625,
0.363037109375
],
[
0.08087158203125,
0.359130859375,
0.1502685546875,
0.359130859375,
0.1502685546875,
0.373779296875,
0.08087158203125,
0.373779296875
],
[
0.50244140625,
0.363037109375,
0.62744140625,
0.363037109375,
0.62744140625,
0.376953125,
0.50244140625,
0.376953125
],
[
0.08087158203125,
0.37890625,
0.2459716796875,
0.37890625,
0.2459716796875,
0.392578125,
0.08087158203125,
0.392578125
],
[
0.26708984375,
0.37744140625,
0.468017578125,
0.37890625,
0.468017578125,
0.393310546875,
0.26708984375,
0.392578125
],
[
0.08251953125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.4091796875,
0.08251953125,
0.4091796875
],
[
0.5185546875,
0.40087890625,
0.890625,
0.401611328125,
0.890625,
0.416748046875,
0.5185546875,
0.415283203125
],
[
0.080078125,
0.4091796875,
0.468017578125,
0.40771484375,
0.468017578125,
0.42236328125,
0.080078125,
0.423583984375
],
[
0.5048828125,
0.417236328125,
0.52294921875,
0.417236328125,
0.52294921875,
0.43115234375,
0.5048828125,
0.43115234375
],
[
0.08087158203125,
0.4248046875,
0.468017578125,
0.423583984375,
0.468017578125,
0.438232421875,
0.08087158203125,
0.439453125
],
[
0.52294921875,
0.431884765625,
0.79833984375,
0.431884765625,
0.79833984375,
0.4462890625,
0.52294921875,
0.4462890625
],
[
0.08087158203125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.08087158203125,
0.453857421875
],
[
0.08087158203125,
0.455078125,
0.134033203125,
0.455078125,
0.134033203125,
0.4697265625,
0.08087158203125,
0.4697265625
],
[
0.6845703125,
0.455810546875,
0.81640625,
0.455810546875,
0.81640625,
0.48095703125,
0.6845703125,
0.48095703125
],
[
0.57421875,
0.4697265625,
0.6845703125,
0.46728515625,
0.6845703125,
0.481689453125,
0.57421875,
0.484130859375
],
[
0.71337890625,
0.46533203125,
0.73193359375,
0.46533203125,
0.73193359375,
0.472900390625,
0.71337890625,
0.472900390625
],
[
0.85888671875,
0.4697265625,
0.89208984375,
0.4697265625,
0.89208984375,
0.48486328125,
0.85888671875,
0.48486328125
],
[
0.080078125,
0.47607421875,
0.32177734375,
0.47607421875,
0.32177734375,
0.4892578125,
0.080078125,
0.4892578125
],
[
0.348876953125,
0.47607421875,
0.46728515625,
0.47607421875,
0.46728515625,
0.4892578125,
0.348876953125,
0.4892578125
],
[
0.708984375,
0.4765625,
0.79345703125,
0.479248046875,
0.79345703125,
0.49365234375,
0.708984375,
0.4912109375
],
[
0.08087158203125,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.08087158203125,
0.5048828125
],
[
0.08087158203125,
0.505859375,
0.468017578125,
0.505859375,
0.468017578125,
0.51953125,
0.08087158203125,
0.51953125
],
[
0.50244140625,
0.5068359375,
0.88818359375,
0.5068359375,
0.88818359375,
0.52099609375,
0.50244140625,
0.52099609375
],
[
0.08251953125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.53515625,
0.08251953125,
0.53515625
],
[
0.50341796875,
0.521484375,
0.890625,
0.521484375,
0.890625,
0.5361328125,
0.50341796875,
0.5361328125
],
[
0.08087158203125,
0.5361328125,
0.46728515625,
0.5361328125,
0.46728515625,
0.5498046875,
0.08087158203125,
0.5498046875
],
[
0.50244140625,
0.537109375,
0.8857421875,
0.537109375,
0.8857421875,
0.55029296875,
0.50244140625,
0.55029296875
],
[
0.08251953125,
0.5517578125,
0.46728515625,
0.5517578125,
0.46728515625,
0.56494140625,
0.08251953125,
0.56494140625
],
[
0.08251953125,
0.56640625,
0.468017578125,
0.56640625,
0.468017578125,
0.580078125,
0.08251953125,
0.580078125
],
[
0.50244140625,
0.56103515625,
0.79736328125,
0.56298828125,
0.79736328125,
0.57763671875,
0.50244140625,
0.57568359375
],
[
0.08251953125,
0.58203125,
0.46728515625,
0.58203125,
0.46728515625,
0.59521484375,
0.08251953125,
0.59521484375
],
[
0.52294921875,
0.5859375,
0.890625,
0.5859375,
0.890625,
0.599609375,
0.52294921875,
0.599609375
],
[
0.08251953125,
0.5966796875,
0.468017578125,
0.5966796875,
0.468017578125,
0.611328125,
0.08251953125,
0.611328125
],
[
0.50341796875,
0.599609375,
0.890625,
0.60107421875,
0.890625,
0.61572265625,
0.50341796875,
0.6142578125
],
[
0.08251953125,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.08251953125,
0.62548828125
],
[
0.5048828125,
0.6162109375,
0.8896484375,
0.6162109375,
0.8896484375,
0.6298828125,
0.5048828125,
0.6298828125
],
[
0.08087158203125,
0.626953125,
0.46728515625,
0.626953125,
0.46728515625,
0.64013671875,
0.08087158203125,
0.64013671875
],
[
0.50341796875,
0.6298828125,
0.52392578125,
0.63330078125,
0.51953125,
0.646484375,
0.5,
0.64306640625
],
[
0.08251953125,
0.64013671875,
0.46728515625,
0.64013671875,
0.46728515625,
0.65478515625,
0.08251953125,
0.65478515625
],
[
0.5048828125,
0.6513671875,
0.8056640625,
0.6513671875,
0.8056640625,
0.662109375,
0.5048828125,
0.662109375
],
[
0.82421875,
0.6513671875,
0.8896484375,
0.6513671875,
0.8896484375,
0.66357421875,
0.82421875,
0.66357421875
],
[
0.50634765625,
0.66748046875,
0.890625,
0.66748046875,
0.890625,
0.67822265625,
0.50634765625,
0.67822265625
],
[
0.08087158203125,
0.6572265625,
0.46728515625,
0.6572265625,
0.46728515625,
0.67041015625,
0.08087158203125,
0.67041015625
],
[
0.080078125,
0.671875,
0.468017578125,
0.671875,
0.468017578125,
0.68603515625,
0.080078125,
0.68603515625
],
[
0.5048828125,
0.68115234375,
0.890625,
0.68115234375,
0.890625,
0.6943359375,
0.5048828125,
0.6943359375
],
[
0.08251953125,
0.6875,
0.2132568359375,
0.6875,
0.2132568359375,
0.70068359375,
0.08251953125,
0.70068359375
],
[
0.50341796875,
0.69580078125,
0.8896484375,
0.69580078125,
0.8896484375,
0.708984375,
0.50341796875,
0.708984375
],
[
0.1004638671875,
0.7041015625,
0.468017578125,
0.7041015625,
0.468017578125,
0.71728515625,
0.1004638671875,
0.71728515625
],
[
0.50341796875,
0.71044921875,
0.8896484375,
0.71044921875,
0.8896484375,
0.7236328125,
0.50341796875,
0.7236328125
],
[
0.08087158203125,
0.71826171875,
0.468017578125,
0.71826171875,
0.468017578125,
0.73291015625,
0.08087158203125,
0.73291015625
],
[
0.5048828125,
0.72607421875,
0.8896484375,
0.72607421875,
0.8896484375,
0.7392578125,
0.5048828125,
0.7392578125
],
[
0.080078125,
0.734375,
0.148681640625,
0.734375,
0.148681640625,
0.74853515625,
0.080078125,
0.74853515625
],
[
0.50341796875,
0.7412109375,
0.875,
0.7412109375,
0.875,
0.7548828125,
0.50341796875,
0.7548828125
],
[
0.080078125,
0.75390625,
0.24267578125,
0.75390625,
0.24267578125,
0.76806640625,
0.080078125,
0.76806640625
],
[
0.261474609375,
0.7548828125,
0.468017578125,
0.7548828125,
0.468017578125,
0.76806640625,
0.261474609375,
0.76806640625
],
[
0.5048828125,
0.755859375,
0.89208984375,
0.755859375,
0.89208984375,
0.76953125,
0.5048828125,
0.76953125
],
[
0.08087158203125,
0.76953125,
0.468017578125,
0.76953125,
0.468017578125,
0.78271484375,
0.08087158203125,
0.78271484375
],
[
0.50341796875,
0.771484375,
0.89208984375,
0.771484375,
0.89208984375,
0.78515625,
0.50341796875,
0.78515625
],
[
0.08251953125,
0.78515625,
0.468017578125,
0.78515625,
0.468017578125,
0.798828125,
0.08251953125,
0.798828125
],
[
0.50341796875,
0.7861328125,
0.890625,
0.7861328125,
0.890625,
0.7998046875,
0.50341796875,
0.7998046875
],
[
0.08087158203125,
0.80029296875,
0.468017578125,
0.80029296875,
0.468017578125,
0.814453125,
0.08087158203125,
0.814453125
],
[
0.50341796875,
0.80029296875,
0.52294921875,
0.80029296875,
0.52294921875,
0.814453125,
0.50341796875,
0.814453125
],
[
0.08087158203125,
0.81494140625,
0.46728515625,
0.81494140625,
0.46728515625,
0.830078125,
0.08087158203125,
0.830078125
],
[
0.50244140625,
0.818359375,
0.75341796875,
0.8193359375,
0.75341796875,
0.833984375,
0.50244140625,
0.83349609375
],
[
0.77392578125,
0.818359375,
0.890625,
0.8193359375,
0.890625,
0.833984375,
0.77392578125,
0.83349609375
],
[
0.08251953125,
0.83056640625,
0.2418212890625,
0.83056640625,
0.2418212890625,
0.8447265625,
0.08251953125,
0.8447265625
],
[
0.5048828125,
0.83544921875,
0.890625,
0.83544921875,
0.890625,
0.8486328125,
0.5048828125,
0.8486328125
],
[
0.50634765625,
0.85107421875,
0.89208984375,
0.85107421875,
0.89208984375,
0.8642578125,
0.50634765625,
0.8642578125
],
[
0.2132568359375,
0.85791015625,
0.365234375,
0.85791015625,
0.365234375,
0.8798828125,
0.2132568359375,
0.8798828125
],
[
0.1192626953125,
0.869140625,
0.1690673828125,
0.869140625,
0.1690673828125,
0.8837890625,
0.1192626953125,
0.8837890625
],
[
0.236083984375,
0.86865234375,
0.2459716796875,
0.86865234375,
0.2459716796875,
0.87255859375,
0.236083984375,
0.87255859375
],
[
0.435546875,
0.87060546875,
0.4697265625,
0.87060546875,
0.4697265625,
0.88623046875,
0.435546875,
0.88623046875
],
[
0.5185546875,
0.87353515625,
0.89208984375,
0.87548828125,
0.89208984375,
0.89013671875,
0.5185546875,
0.88818359375
],
[
0.181396484375,
0.8818359375,
0.39794921875,
0.8818359375,
0.39794921875,
0.90380859375,
0.181396484375,
0.90380859375
],
[
0.5048828125,
0.8896484375,
0.55810546875,
0.8896484375,
0.55810546875,
0.8994140625,
0.5048828125,
0.8994140625
]
] |
[
"HN",
"Figure 2.Siamese network output for JPEGdistortion",
"Figure 3.",
"Convergence of ranking loss on JPEG distor",
"considering 6 levels. This graphs illustrate the fact that",
"tion for our approach versus standard Siamese and hard-",
"the Siamese network successfully manages to separate the",
"negative mining.",
"different distortionlevels.",
"Shallow",
"AlexNet",
"VGG-16",
"LCC",
"0.930",
"0.945",
"0.973",
"SROCC",
"4.2.Experimentalprotocols",
"0.937",
"0.949",
"0.974",
"我们调查了许多网络架构，并使用标准IQA指标来",
"Table 1.SROCC and LCC for our approach on LIVE using",
"评估性能",
"different networks.",
"Network architectures.",
"我们评估了三种从浅到深的",
"典型网络架构。我们将它们称为：Shallow，AlexNetE",
"其中立和分别是地面实况和预测质量得分的平均",
"[14]和VGG-16E[30]。浅层网络有四个卷积层和一个",
"值",
"全连接层。对于AlexNet和VGG-16，我们只改变输",
"给定N失真图像，SROCC计算如下：",
"出的数量，因为我们的目标是为每个失真图像分配一",
"个分数",
"6En (; - p,)2",
"SROCC=1-",
"",
"(14)",
"Strategy for training and testing.",
"我们从原始高分",
"N(N2-1)",
"辨率图像中随机采样子图像。我们这样做而不是缩放，",
"以避免引人由插值或滤波引起的失真。采样图像的大",
"其中U是地面实况IQA分数i在地面实况分数中",
"小由每个网络确定。但是，输入图像的大尺寸很重要，",
"的排名，而pi是9i在所有N的输出分数中的排名。",
"因为输人子图像应至少为原始图像的1/3，以便捕获上",
"SROCC测量地面实况与估计的IQA之间的单调关系",
"下文信息。这是补丁采样方法的严重限制，该方法从",
"原始图像中采样非常小的32×32补丁。在我们的实验",
"4.3. Learning NR-IQA from rankings",
"中，我们对227×227和224×224像素图像进行采样",
"我们进行了许多实验，以评估逼罗网络学习从大型",
"具体取决于网络。我们使用CaffeE[10]框架，并使用",
"图像质量排名数据集中捕获图像失真的能力。此外，我",
"小批量随机梯度下降（SGD）进行训练，初始学习率为",
"们测量了3.3节中描述的有效连体反向传播方法的影",
"1e-4，用于有效的暹罗网络训练，1e-6用于微调。培训",
"响",
"率降低了0倍。每10K次选代1次，总共50K次迭",
"Siamese networks and IQA discrimination.",
"为了展示",
"我们的排名网络区分图像质量的能力，我们在Places2",
"代。对于两个训练阶段，我们使用22权重衰减（权重",
"5e-4）。在训练期间，我们从每个时期的每个训练图像",
"验证集上训练了我们的Siamese网络（没有对IQA数",
"中采样单个子图像",
"据进行微调），五个级别的单一失真。然后，我们使用",
"测试时，我们按照1I中的建议从原始图像中随机",
"该网络来预测来自滑铁卢数据集的合成失真图像的图",
"采样30个子图像。子区域输出的平均值是每个图像的",
"像质量，这些图像使用相同的五个级别的相同失真。网",
"最终分数",
"络输出绘制为图中的直方图。2用于JPEG失真",
"Evaluation protocols.",
"传统上使用两个评估指标来",
"在图中，我们根据真实的失真水平（由直方图的颜色表",
"评估IQA算法的性能：线性相关系数（LCC）和斯皮",
"示）划分观察值。即使两个数据集的获取过程和场景完",
"尔曼等级相关系数（SROCC）。LCC是地面实况和预",
"全不同，该模型也可以区分滑铁卢上的不同程度的失",
"测质量分数之间的线性相关的度量。给定N失真图像，",
"真",
"第i图像的地面实况由yi表示，网络的预测得分为9i",
"Efficient Siamesebackpropagation.",
"本实验的目的是",
"。LCC的计算公式为：",
"评估我们的逼罗反向传播方法的效率。我们将我们的",
"方法与标准随机对抽样和类似于【29]的硬负挖掘方法",
"N(; - )(0 - )",
"LCC ",
"",
"(13)",
"1Graphs for the other distortions are in the supplementary",
"VEN()E(-)2",
"material."
] |
{
"arxiv_id": "1707.08347",
"language": "zh",
"page_idx": 4
}
|
||
2112.08702-zh-0
|
2112.08702-zh-0.jpg
|
$\pi$。对于合作 MARL,已经提出了许多集中训练和分散执行(CTDE)的方法，例如 COMA (Foerster et al., 2018) , QMIX (Rashid et al., 2018) , QPLEX (Wang et al., 2021) 和 FOP (Zhang et al., 2021)。但是，这些方法存在过度泛化的问题：由于代理有时会选择不协调的操作，因此使用的值函数无法很好地估计，因此无法学习最佳策略(Castellini et al.,2019)。此外，由于集中学习(Qu et al., 2020a) , 它们可能不容易随着代理数量的增加而扩展。在许多 MARL 应用程序中，有大量代理被部署为部分连接的网络，并协同做出决策以优化全局平均收益例如通信网络 (Kim et al., 2019) 和交通信号控制 (Wei et al., 2019) 。为了处理这种情况，制定了网络 MARL 以将所有代理之间的依赖关系分解为仅邻居之间的依赖关系。为避免决策信息不足，允许代理通过网络与邻居交换消息。在这样的设置中，代理学习以分散的方式做出决策是可行的 (Zhang et al., 2018; Qu et al., 2020b)。然而，如果有人试图独立做出决定，仍然存在依赖的困难，例如。g。、囚徒困境与公地悲剧(Pérolat et al.,2017)。现有方法通过一致更新值函数(Zhang et al., 2018) , 信用分配(Wang et al., 2020) 或奖励整形(Chu et al., 2020) 来解决这些问题。但是，这些方法依赖于对全局状态和联合动作 (Zhang et al.,2018) 的访问或手工制作的奖励函数(Wang et al., 2020; Chu et al., 2020) 受共享在人类的合作学习 (Eisenberg and Mussen, 1989) 中起着关键作用这一事实的启发，我们提出了学习共享 (LToS),这是一种用于网络 MARL 的分层分散学习框架。LToS 使代理能够学习与邻居动态共享奖励，从而协作优化全局目标。高级别政策通过确定如何分享回报，将全球目标分解为地方目标，而低级别政策则优化了高级别政策引起的地方目标。LToS 以分散的方式学习，我们证明了高层策略是联合高层策略的均值场近似。此外，高层和低层政策形成了双层优化，并交替学习优化全局目标。LToS 是用于联网 MARL 的通用分层框架，可以通过 RL 算法的多种组合轻松实现。我们目前通过 DDPG (Lillicrap et al., 2016) 作为高级策略和 DGN (Jiang et al., 2020) 作为低级策略实现 LToS。我们通过经验证明，在社会困境和网络 MARL 场景中，LToS 优于现有的网络 MARL 方法。
## 1 $\textbf{Related Work}$
最近有许多关于协作 MARL 的研究。大多数采用 CTDE (Foerster et al., 2018; Rashid et al. 2018; Wang et al., 2021; Zhang et al., 2021; Su and Lu, 2022)。其中许多是在分解联合 Q 函数$(\bar{\text{Rashid et al., 2018; Wang et al., 2021)的基础上构造的。然而，这些分解的方法遭受过度泛化}}$ 问题 (Castellini et al.,2019) 。其他研究更侧重于分散培训，我们的工作与之联系更密切，总结如下。联网 MARL Zhang et al.(2018) 和 Qu et al.(2019) 建议对局部值函数进行共识更新， 其中每个代理保留全局值函数的局部副本，但假定具有全局信息。Qu et al.(2020a)提出了代理之间的意图传播，其中每个代理都根据其他代理共享的意图更新其策略，但是由于通过网络传播的意图，策略可能会收敛缓慢。Qu et al.(2020b) 和 Lin et al.(2020) 研究了指数衰减特性，即。e。,代理彼此之间的影响在图距离上呈指数衰减，而 Chu et al.(2020)引人了空间折扣因子来捕捉代理之间的影响，该因子仍然是手动调整的。然而，这些研究都没有提供一个明确的机制来解决网络 MARL 中的社会困境。
奖励设计 Hostallero et al. (2020) 旨在最大化社会福利，但他们只是使用时间差异误差进行奖励塑造。由于深度 RL 中的时间差误差几乎不收敛到零，因此它仍然会使优化目标产生偏差。Mguni et al.(2019) 在原始奖励中添加了一个额外的部分，作为基于非潜在的奖励整形并使用贝叶斯优化来诱导收敛到代理之间的理想平衡。但是，额外的部分在情节中保持固定，这使其处理动态环境的能力降低。此外，奖励整形改变了原始的优化问题。Hughes et al (2018) 提出了不公平厌恶模型来平衡代理人的私欲和社会公平。Wang et al.(2020) 考虑学习Shapley 值作为学分分配。然而，这些方法仍然依赖于手工制作的奖励设计。Lupu and Precup $(20\bar{2}0)$在原始 MDP 中添加了礼物作为额外的动作，以修改 MDP 和目标。Yang et al.(2020) 建议每个代理学习一个激励函数，并根据其他代理给予的奖励和激励来优化策略。显然，这两种方法都改变了优化的原始目标。与现有工作不同，LToS 使代理能够学习与其他代理动态共享奖励，而不会产生优化目标的偏差，以便他们可以协作优化网络 MARL 中的全局目标。
# 2 $\textbf{Background}$
2.1 Networked Multi-Agent Reinforcement Learning
假定$N$ 代理与环境交互。令$\mathcal{V}=\{1,2,\cdots,N\}$ 为代理集。将多代理系统建模为无向图$\mathcal{G}(\mathcal{V},\mathcal{E})$ ,其中每个代理$i$充当顶点$i$,而$\mathcal{E}\subseteq\mathcal{V}\times\mathcal{V}$是所有边的集合。当且仅当$e_ij=(i,j)\in\mathcal{E}$时， 两个代理$i,j\in\mathcal{V}$可以相互通信。我们将代理$i$及其在图中的所有邻居一起表示为集合$\mathcal{N}_i$ 。环境$s\in\mathcal{S}$的状态根据转移概率$\mathcal{P}_a:\mathcal{S}\times\mathcal{A}\times\mathcal{S}\to[0,1]$在联合动作$\boldsymbol a\in\mathcal{A}$上转移，其中联
1
|
[
[
0.1732177734375,
0.0921630859375,
0.82275390625,
0.091552734375,
0.82275390625,
0.1060791015625,
0.1732177734375,
0.106689453125
],
[
0.1748046875,
0.106689453125,
0.82275390625,
0.106689453125,
0.82275390625,
0.12060546875,
0.1748046875,
0.12060546875
],
[
0.176513671875,
0.12060546875,
0.82275390625,
0.12060546875,
0.82275390625,
0.1337890625,
0.176513671875,
0.1337890625
],
[
0.176513671875,
0.1337890625,
0.82275390625,
0.1337890625,
0.82275390625,
0.147705078125,
0.176513671875,
0.147705078125
],
[
0.17724609375,
0.146484375,
0.38818359375,
0.147705078125,
0.38818359375,
0.1610107421875,
0.17724609375,
0.160400390625
],
[
0.3798828125,
0.147705078125,
0.82275390625,
0.147705078125,
0.82275390625,
0.1610107421875,
0.3798828125,
0.1610107421875
],
[
0.176513671875,
0.1610107421875,
0.82177734375,
0.1610107421875,
0.82177734375,
0.1749267578125,
0.176513671875,
0.1749267578125
],
[
0.17724609375,
0.175537109375,
0.82275390625,
0.175537109375,
0.82275390625,
0.189453125,
0.17724609375,
0.189453125
],
[
0.178955078125,
0.1875,
0.82275390625,
0.1881103515625,
0.82275390625,
0.20263671875,
0.178955078125,
0.2020263671875
],
[
0.176513671875,
0.20263671875,
0.82177734375,
0.20263671875,
0.82177734375,
0.216552734375,
0.176513671875,
0.216552734375
],
[
0.176513671875,
0.216552734375,
0.82275390625,
0.216552734375,
0.82275390625,
0.2310791015625,
0.176513671875,
0.2310791015625
],
[
0.17724609375,
0.2298583984375,
0.82275390625,
0.2298583984375,
0.82275390625,
0.2442626953125,
0.17724609375,
0.2442626953125
],
[
0.17724609375,
0.24365234375,
0.82275390625,
0.24365234375,
0.82275390625,
0.25830078125,
0.17724609375,
0.25830078125
],
[
0.176513671875,
0.25830078125,
0.82275390625,
0.25830078125,
0.82275390625,
0.271484375,
0.176513671875,
0.271484375
],
[
0.176513671875,
0.270751953125,
0.82421875,
0.271484375,
0.82421875,
0.285888671875,
0.176513671875,
0.285400390625
],
[
0.1748046875,
0.285400390625,
0.810546875,
0.285400390625,
0.810546875,
0.298583984375,
0.1748046875,
0.298583984375
],
[
0.1748046875,
0.29736328125,
0.82275390625,
0.298583984375,
0.82275390625,
0.313232421875,
0.1748046875,
0.3125
],
[
0.176513671875,
0.3125,
0.82177734375,
0.3125,
0.82177734375,
0.32568359375,
0.176513671875,
0.32568359375
],
[
0.176513671875,
0.326904296875,
0.82275390625,
0.326904296875,
0.82275390625,
0.34033203125,
0.176513671875,
0.34033203125
],
[
0.176513671875,
0.34033203125,
0.8203125,
0.34033203125,
0.8203125,
0.354248046875,
0.176513671875,
0.354248046875
],
[
0.176513671875,
0.354248046875,
0.82275390625,
0.354248046875,
0.82275390625,
0.367431640625,
0.176513671875,
0.367431640625
],
[
0.1748046875,
0.367431640625,
0.82275390625,
0.3662109375,
0.82275390625,
0.380615234375,
0.1748046875,
0.3818359375
],
[
0.17724609375,
0.3818359375,
0.82275390625,
0.3818359375,
0.82275390625,
0.395751953125,
0.17724609375,
0.395751953125
],
[
0.176513671875,
0.396484375,
0.336669921875,
0.396484375,
0.336669921875,
0.410400390625,
0.176513671875,
0.410400390625
],
[
0.17236328125,
0.427978515625,
0.326904296875,
0.4287109375,
0.326904296875,
0.443115234375,
0.17236328125,
0.442626953125
],
[
0.176513671875,
0.458984375,
0.82275390625,
0.458984375,
0.82275390625,
0.472900390625,
0.176513671875,
0.472900390625
],
[
0.1748046875,
0.472900390625,
0.82275390625,
0.472900390625,
0.82275390625,
0.4873046875,
0.1748046875,
0.4873046875
],
[
0.1748046875,
0.484130859375,
0.82275390625,
0.48486328125,
0.82275390625,
0.5,
0.1748046875,
0.498779296875
],
[
0.17724609375,
0.5,
0.82177734375,
0.5,
0.82177734375,
0.51318359375,
0.17724609375,
0.51318359375
],
[
0.176513671875,
0.5146484375,
0.82177734375,
0.5146484375,
0.82177734375,
0.52783203125,
0.176513671875,
0.52783203125
],
[
0.176513671875,
0.52783203125,
0.82275390625,
0.52783203125,
0.82275390625,
0.54150390625,
0.176513671875,
0.54150390625
],
[
0.176513671875,
0.54150390625,
0.82177734375,
0.54150390625,
0.82177734375,
0.5546875,
0.176513671875,
0.5546875
],
[
0.176513671875,
0.5537109375,
0.82275390625,
0.5546875,
0.82275390625,
0.5693359375,
0.176513671875,
0.568359375
],
[
0.176513671875,
0.568359375,
0.82275390625,
0.568359375,
0.82275390625,
0.58203125,
0.176513671875,
0.58203125
],
[
0.17724609375,
0.583984375,
0.82275390625,
0.583984375,
0.82275390625,
0.5947265625,
0.17724609375,
0.5947265625
],
[
0.188720703125,
0.5966796875,
0.53125,
0.5966796875,
0.53125,
0.60986328125,
0.188720703125,
0.60986328125
],
[
0.176513671875,
0.6171875,
0.82275390625,
0.6171875,
0.82275390625,
0.630859375,
0.176513671875,
0.630859375
],
[
0.176513671875,
0.630859375,
0.82275390625,
0.630859375,
0.82275390625,
0.64453125,
0.176513671875,
0.64453125
],
[
0.178955078125,
0.64794921875,
0.20751953125,
0.64794921875,
0.20751953125,
0.65478515625,
0.178955078125,
0.65478515625
],
[
0.2017822265625,
0.64453125,
0.82177734375,
0.64453125,
0.82177734375,
0.65771484375,
0.2017822265625,
0.65771484375
],
[
0.176513671875,
0.65771484375,
0.82177734375,
0.65771484375,
0.82177734375,
0.671875,
0.176513671875,
0.671875
],
[
0.176513671875,
0.67236328125,
0.82177734375,
0.67236328125,
0.82177734375,
0.68603515625,
0.176513671875,
0.68603515625
],
[
0.176513671875,
0.68603515625,
0.8203125,
0.68603515625,
0.8203125,
0.69970703125,
0.176513671875,
0.69970703125
],
[
0.1748046875,
0.69873046875,
0.82177734375,
0.69970703125,
0.82177734375,
0.7138671875,
0.1748046875,
0.71337890625
],
[
0.176513671875,
0.71337890625,
0.82275390625,
0.71337890625,
0.82275390625,
0.7265625,
0.176513671875,
0.7265625
],
[
0.17724609375,
0.7265625,
0.82177734375,
0.7265625,
0.82177734375,
0.74072265625,
0.17724609375,
0.74072265625
],
[
0.17724609375,
0.7412109375,
0.82177734375,
0.7412109375,
0.82177734375,
0.7548828125,
0.17724609375,
0.7548828125
],
[
0.176513671875,
0.7548828125,
0.82177734375,
0.7548828125,
0.82177734375,
0.76806640625,
0.176513671875,
0.76806640625
],
[
0.1748046875,
0.76953125,
0.19775390625,
0.76953125,
0.19775390625,
0.7822265625,
0.1748046875,
0.7822265625
],
[
0.1707763671875,
0.7998046875,
0.314453125,
0.8017578125,
0.314453125,
0.8193359375,
0.1707763671875,
0.8173828125
],
[
0.17236328125,
0.830078125,
0.5595703125,
0.83203125,
0.5595703125,
0.8466796875,
0.17236328125,
0.8447265625
],
[
0.176513671875,
0.85791015625,
0.82421875,
0.85791015625,
0.82421875,
0.8720703125,
0.176513671875,
0.8720703125
],
[
0.1871337890625,
0.869140625,
0.82177734375,
0.87060546875,
0.82177734375,
0.88525390625,
0.1871337890625,
0.8837890625
],
[
0.176513671875,
0.8837890625,
0.81640625,
0.8837890625,
0.81640625,
0.89794921875,
0.176513671875,
0.89794921875
],
[
0.1748046875,
0.8984375,
0.82421875,
0.8984375,
0.82421875,
0.912109375,
0.1748046875,
0.912109375
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.9462890625,
0.49267578125,
0.9462890625
]
] |
[
"元。对于合作MARL，已经提出了许多集中训练和分散执行（CTDE）的方法，例如COMA",
"(Foerster et al., 2018), QMIX (Rashid et al., 2018), QPLEX (Wang et al., 2021) 和 FOP (Zhang",
"etal.，2021）。但是，这些方法存在过度泛化的问题：由于代理有时会选择不协调的操作，因",
"此使用的值函数无法很好地估计，因此无法学习最佳策略（Castellinietal.，2019）。此外，由",
"于集中学习（Quetal.,2020a），",
"，它们可能不容易随着代理数量的增加而扩展。在许多MARL",
"应用程序中，有大量代理被部署为部分连接的网络，并协同做出决策以优化全局平均收益",
"例如通信网络（Kimetal.，2019）和交通信号控制（Weietal.,2019）。为了处理这种情况，制定",
"了网络MARL以将所有代理之间的依赖关系分解为仅邻居之间的依赖关系。为避免决策信",
"息不足，允许代理通过网络与邻居交换消息。在这样的设置中，代理学习以分散的方式做出",
"决策是可行的（Zhangetal.,2018;Quetal.,2020b）。然而，如果有人试图独立做出决定，仍然",
"存在依赖的困难，例如。g。、徒困境与公地悲剧(Pérolatetal.,2017）。现有方法通过一致",
"更新值函数（Zhangetal.,2018），信用分配（Wangetal.,2020）或奖励整形（Chuetal.,2020）来",
"解决这些问题。但是，这些方法依赖于对全局状态和联合动作（Zhangetal.,2018）的访问或",
"手工制作的奖励函数（Wangetal.,2020;Chuetal.,2020）受共享在人类的合作学习（Eisenberg",
"andMussen，1989）中起着关键作用这一事实的启发，我们提出了学习共享（LToS），这是",
"种用于网络MARL的分层分散学习框架。LToS使代理能够学习与邻居动态共享奖励，从而",
"协作优化全局目标。高级别政策通过确定如何分享回报，将全球目标分解为地方目标，而低",
"级别政策则优化了高级别政策引起的地方目标。LToS以分散的方式学习，我们证明了高层",
"策略是联合高层策略的均值场近似。此外，高层和低层政策形成了双层优化，并交替学习优",
"化全局目标。LToS是用于联网MARL的通用分层框架，可以通过RL算法的多种组合轻松",
"实现。我们目前通过DDPG（Lillicrapetal.,2016）作为高级策略和DGN（Jiangetal.2020）作",
"为低级策略实现LToS。我们通过经验证明，在社会困境和网络MARL场景中，LToS优于现",
"有的网络MARL方法。",
"1Related Work",
"最近有许多关于协作MARL的研究。大多数采用CTDE（Foersteretal.，2018;Rashidetal.，",
"2018Wangetal.,2021;Zhangetal.,2021;SuandLu,2022）。其中许多是在分解联合Q函数",
"(Rashidetal.2018；Wangetal.2021）的基础上构造的。然而，这些分解的方法遭受过度泛化",
"问题（Castellinietal.,2019）。其他研究更侧重于分散培训，我们的工作与之联系更密切，总",
"结如下。联网MARLZhangetal.（2018）和Quetal.（2019）建议对局部值函数进行共识更新",
"其中每个代理保留全局值函数的局部副本，但假定具有全局信息。Quetal.（2020a）提出了代",
"理之间的意图传播，其中每个代理都根据其他代理共享的意图更新其策略，但是由于通过",
"网络传播的意图，策略可能会收敛缓慢。Quetal.(2020b）和Linetal.（2020)研究了指数衰减",
"特性，即。e。，代理被此之间的影响在图距离上呈指数衰减，而Chuetal.(2020)引人了空",
"间折扣因子来捕捉代理之间的影响，该因子仍然是手动调整的。然而，这些研究都没有提供",
"个明确的机制来解决网络MARL中的社会困境。",
"奖励设计Hostalleroetal.(2020）旨在最大化社会福利，但他们只是使用时间差异误差进行奖",
"励塑造。由于深度RL中的时间差误差几乎不收敛到零，因此它仍然会使优化目标产生偏",
"君。",
"Mgunietal.(2019)在原始奖励中添加了一个额外的部分，作为基于非潜在的奖励整形",
"并使用贝叶斯优化来诱导收敛到代理之间的理想平衡。但是，额外的部分在情节中保持固",
"定，这使其处理动态环境的能力降低。此外，奖励整形改变了原始的优化问题。Hughesetal.",
"(2018）提出了不公平厌恶模型来平衡代理人的私欲和社会公平。Wangetal.(2020)考虑学习",
"Shapley值作为学分分配。然而，这些方法仍然依赖于手工制作的奖励设计。LupuandPrecup",
"(2020）在原始MDP中添加了礼物作为额外的动作，以修改MDP和目标。Yangetal.(2020)",
"建议每个代理学习一个激励函数，并根据其他代理给予的奖励和激励来优化策略。显然，这",
"两种方法都改变了优化的原始目标。与现有工作不同，LToS使代理能够学习与其他代理动",
"态共享奖励，而不会产生优化目标的偏差，以便他们可以协作优化网络MARL中的全局目",
"标。",
"2Background",
"2.1Networked Multi-Agent Reinforcement Learning",
"假定N代理与环境交互。令V=1,2,..，N)为代理集。将多代理系统建模为无向图G(V,)",
"其中每个代理i充当顶点，而CV×V是所有边的集合。当且仅当eii=（i,j）Eε时",
"两个代理i，iEV可以相互通信。我们将代理i及其在图中的所有邻居一起表示为集合N",
"环境sES的状态根据转移概率Pa:S×A×S→[O,1]在联合动作aEA上转移，其中联",
"1"
] |
{
"arxiv_id": "2112.08702",
"language": "zh",
"page_idx": 0
}
|
||
2112.08702-zh-1
|
2112.08702-zh-1.jpg
|
合动作集$\mathcal{A}=\times_i\in\mathcal{V}\mathcal{A}_i$ 。每个代理$i$都有一个策略$\pi_i\in\Pi_i:\mathcal{S}\times\mathcal{A}_i\to[0,1]$ ,我们将所有代理的联合策略表示为$\pi\in\Pi=\times_i\in\mathcal{V}\Pi_i$ (Zhang et al., 2018)。对于联网的 MARL, 一个常见且现实的假设是，每个代理$i$的奖励仅取决于其动作及其邻居(Qu et al.,2020a)的动作，即。e。,$r_i(s,\boldsymbol{a})=r_i(s,a_{\mathcal{N}_i})$。此外，每个代理$i$可能仅获得部分观测值$o_i\in\mathcal{O}_i$,但可以通过$\mathcal{N}_i$ (Jiang et al., 2020) 的观测值或观测历史 (Chu et al., 2020) 来近似状态，为简单起见，它们均由$o_i$表示。全球目标是最大化所有代理的累积奖励之和，即。e。,$\sum_t=0^\infty\sum_{i=1}^N\gamma^tr_i^t$ 。
## 2.2 Markov Game
在这样的设置中，每个代理可以单独最大化自己的预期收益，这被称为马尔可夫博弈。这可能会导致稳定的结果或纳什均衡，但就全球目标而言，这通常是次优的。给定$\pi$ ,代理$i$的值函数由
(1)
$$v_{i}^{\boldsymbol{\pi}}(s)=\sum_{\boldsymbol{a}}\pi(\boldsymbol{a}|s)\sum_{s'}p_{a}(s'|s,\boldsymbol{a})[r_{i}+\gamma v_{i}^{\boldsymbol{\pi}}(s')],$$
给出，其中$p_a\in\mathcal{P}_a$描述状态转换。纳什均衡定义为 (Mguni et al., 2019)
$$v_i^{(\pi_i,\pi_{-i})}(s)\geq v_i^{(\pi_i',\pi_{-i})}(s),\quad\forall\pi_i'\in\Pi_i,\forall s\in\mathcal{S},\forall i\in\mathcal{V},$$
,其中$\pi _- i= \times _{j\in \mathcal{V} \setminus \{ i\} }\pi _j$ 。
## 3 $\textbf{Method}$
LToS 是一个分散的层次结构。在每个代理上，高层策略基于低层策略确定奖励共享的权重而低层策略直接与环境交互以优化高层策略引起的本地目标。因此，它们形成了双层优化， 并交替向全局目标学习。
## 3.1 Reward Sharing
奖励共享的直觉是，如果代理与他人共享奖励，则每个代理都必须考虑其行为对他人的后果，从而促进了合作。在网络 MARL 中，由于假定代理的奖励取决于邻居的行为，因此我们仅允许在相邻代理之间共享奖励。这是因为邻居的行为变化直接影响奖励，而邻居外的代理只能通过状态分布的变化间接影响代理的回报。此外，这也符合邻居之间通信受限的网络MARL 的设置。对于$\mathcal{V}$的图，我们还定义了一组由$\mathcal{E}$构造的有向边$\mathcal{D}$ 。具体来说，我们为每个代理$i$添加一个循环$d_ii\in\mathcal{D}$ ,并将每个无向边$e_ij\in\mathcal{E}$拆分为两个有向边$:d_ij=(i,j)$和$d_{ji}=(j,i)\in\mathcal{D}$。每个代理$i$根据约束$\sum_j\in\mathcal{N}_iw_{ij}=1$为每个有向边$d_ij,\forall j\in\check{N}_i$确定权重$w_{ij}\in [ 0, 1]$ ,以便$w_ij$代理$i$的环境奖励$r_i$的比例将共享给代理$j$ 。令$\boldsymbol w\in\mathcal{W}=\times_d_{ij}\in\mathcal{D}w_{ij}$ 为图的权重。因此，每个代理$i$共享后的成形奖励定义为
(2)
$$r_i^{\boldsymbol{w}}=\sum_{j\in\mathcal{N}_i}w_{ji}r_j.$$
## 3.2 Hierarchy
假设有一个联合高级策略$\phi\in\Phi:\mathcal{S}\times\mathcal{W}\to[0,1]$来确定$\boldsymbol{w}$ 。给定$\phi$和$\boldsymbol{w}$ ,我们可以根据
(1)将每个代理$i$处的$\pi$的值函数定义为
(3)
$$v_{i}^{\boldsymbol{\pi}}(s;\phi)=\sum_{\boldsymbol{w}}\phi(\boldsymbol{w}|s)\sum_{\boldsymbol{a}}\boldsymbol{\pi}(\boldsymbol{a}|s,\boldsymbol{w})\sum_{s^{\prime}}p_{a}(s^{\prime}|s,\boldsymbol{a})[r_{i}^{\boldsymbol{w}}+\gamma v_{i}^{\boldsymbol{\pi}}(s^{\prime};\boldsymbol{\phi})]\\v_{i}^{\boldsymbol{\pi}}(s;\boldsymbol{w},\boldsymbol{\phi})=\sum_{\boldsymbol{a}}\boldsymbol{\pi}(\boldsymbol{a}|s,\boldsymbol{w})\sum_{s^{\prime}}p_{a}(s^{\prime}|s,\boldsymbol{a})[r_{i}^{\boldsymbol{w}}+\gamma v_{i}^{\boldsymbol{\pi}}(s^{\prime};\boldsymbol{\phi})].$$
值得注意的是，$\boldsymbol{w}$是分配方案的多维操作，而不是概率分布。在我们的推导中，为了简单起见，我们将$w$表示为离散动作。只要我们将所有求和更改为积分，它也适用于连续操作。设$V_{\mathcal{V}}^{\phi}(s;\boldsymbol{\pi})\doteq\sum_{i\in\mathcal{V}}v_{i}^{\pi}(s;\phi)$ 和 $Q_{\mathcal{V} }^{\phi }( s, \boldsymbol{w}; \boldsymbol{\pi }) \doteq \sum _{i\in \mathcal{V} }v_{i}^{{\boldsymbol{\pi }}}( s; \boldsymbol{w}, \boldsymbol{\phi })$ 。
$\textbf{Proposition 3.1. Given ,V}_{\mathcal{V}}^{\Phi}(s;\boldsymbol{\pi})$ and $Q_{\mathcal{V}}^{\Phi}(s,\boldsymbol{w};\boldsymbol{\pi})$ are respectively the value function and
action-value function of $\phi$ .
口
$Proof.$ The proof is deferred to Appendix ??.
2
|
[
[
0.17724609375,
0.0921630859375,
0.82275390625,
0.0921630859375,
0.82275390625,
0.1060791015625,
0.17724609375,
0.1060791015625
],
[
0.176513671875,
0.106689453125,
0.7509765625,
0.106689453125,
0.7509765625,
0.12060546875,
0.176513671875,
0.12060546875
],
[
0.76318359375,
0.10986328125,
0.81884765625,
0.10986328125,
0.81884765625,
0.117431640625,
0.76318359375,
0.117431640625
],
[
0.17724609375,
0.12060546875,
0.82177734375,
0.12060546875,
0.82177734375,
0.1337890625,
0.17724609375,
0.1337890625
],
[
0.1748046875,
0.1337890625,
0.82177734375,
0.1331787109375,
0.82177734375,
0.147705078125,
0.1748046875,
0.1483154296875
],
[
0.1748046875,
0.147705078125,
0.82275390625,
0.147705078125,
0.82275390625,
0.1610107421875,
0.1748046875,
0.1610107421875
],
[
0.1748046875,
0.1641845703125,
0.79345703125,
0.1641845703125,
0.79345703125,
0.1787109375,
0.1748046875,
0.1787109375
],
[
0.1748046875,
0.1944580078125,
0.318603515625,
0.1944580078125,
0.318603515625,
0.2083740234375,
0.1748046875,
0.2083740234375
],
[
0.176513671875,
0.2203369140625,
0.82177734375,
0.2203369140625,
0.82177734375,
0.2342529296875,
0.176513671875,
0.2342529296875
],
[
0.176513671875,
0.2342529296875,
0.82177734375,
0.2342529296875,
0.82177734375,
0.2474365234375,
0.176513671875,
0.2474365234375
],
[
0.17724609375,
0.248779296875,
0.240234375,
0.248779296875,
0.240234375,
0.260009765625,
0.17724609375,
0.260009765625
],
[
0.334228515625,
0.26318359375,
0.6640625,
0.26318359375,
0.6640625,
0.281005859375,
0.334228515625,
0.281005859375
],
[
0.80224609375,
0.26513671875,
0.826171875,
0.26513671875,
0.826171875,
0.279052734375,
0.80224609375,
0.279052734375
],
[
0.176513671875,
0.29541015625,
0.68310546875,
0.29541015625,
0.68310546875,
0.31005859375,
0.176513671875,
0.31005859375
],
[
0.30810546875,
0.31494140625,
0.68896484375,
0.317626953125,
0.68896484375,
0.33837890625,
0.30810546875,
0.3359375
],
[
0.1732177734375,
0.339111328125,
0.365234375,
0.3447265625,
0.365234375,
0.359130859375,
0.1732177734375,
0.354248046875
],
[
0.1732177734375,
0.375,
0.27685546875,
0.375,
0.27685546875,
0.389404296875,
0.1732177734375,
0.389404296875
],
[
0.17724609375,
0.4072265625,
0.82177734375,
0.4072265625,
0.82177734375,
0.417236328125,
0.17724609375,
0.417236328125
],
[
0.176513671875,
0.419189453125,
0.82177734375,
0.419189453125,
0.82177734375,
0.43310546875,
0.176513671875,
0.43310546875
],
[
0.176513671875,
0.43310546875,
0.34716796875,
0.43310546875,
0.34716796875,
0.4462890625,
0.176513671875,
0.4462890625
],
[
0.1732177734375,
0.4609375,
0.33251953125,
0.46337890625,
0.33251953125,
0.477783203125,
0.1732177734375,
0.47607421875
],
[
0.176513671875,
0.4892578125,
0.82275390625,
0.4892578125,
0.82275390625,
0.50244140625,
0.176513671875,
0.50244140625
],
[
0.17724609375,
0.50244140625,
0.82275390625,
0.50244140625,
0.82275390625,
0.5166015625,
0.17724609375,
0.5166015625
],
[
0.176513671875,
0.5166015625,
0.82177734375,
0.5166015625,
0.82177734375,
0.52978515625,
0.176513671875,
0.52978515625
],
[
0.176513671875,
0.53076171875,
0.81884765625,
0.53076171875,
0.81884765625,
0.54150390625,
0.176513671875,
0.54150390625
],
[
0.1748046875,
0.54345703125,
0.82177734375,
0.54345703125,
0.82177734375,
0.556640625,
0.1748046875,
0.556640625
],
[
0.176513671875,
0.556640625,
0.8203125,
0.556640625,
0.8203125,
0.57080078125,
0.176513671875,
0.57080078125
],
[
0.1748046875,
0.57080078125,
0.82421875,
0.57080078125,
0.82421875,
0.58837890625,
0.1748046875,
0.58837890625
],
[
0.1748046875,
0.58544921875,
0.82421875,
0.5869140625,
0.82421875,
0.60498046875,
0.1748046875,
0.60302734375
],
[
0.176513671875,
0.60302734375,
0.57861328125,
0.60302734375,
0.57861328125,
0.6162109375,
0.176513671875,
0.6162109375
],
[
0.43701171875,
0.62353515625,
0.55810546875,
0.626953125,
0.55810546875,
0.64453125,
0.43701171875,
0.6416015625
],
[
0.8037109375,
0.62744140625,
0.82421875,
0.62744140625,
0.82421875,
0.6416015625,
0.8037109375,
0.6416015625
],
[
0.48046875,
0.64306640625,
0.5146484375,
0.64306640625,
0.5146484375,
0.65576171875,
0.48046875,
0.65576171875
],
[
0.1732177734375,
0.66845703125,
0.285888671875,
0.671875,
0.285888671875,
0.68603515625,
0.1732177734375,
0.68310546875
],
[
0.1748046875,
0.69580078125,
0.82275390625,
0.6962890625,
0.82275390625,
0.7109375,
0.1748046875,
0.71044921875
],
[
0.1732177734375,
0.71044921875,
0.47216796875,
0.7109375,
0.47216796875,
0.724609375,
0.1732177734375,
0.7236328125
],
[
0.259033203125,
0.73193359375,
0.740234375,
0.734375,
0.740234375,
0.75048828125,
0.259033203125,
0.74853515625
],
[
0.80224609375,
0.7392578125,
0.82421875,
0.7392578125,
0.82421875,
0.75390625,
0.80224609375,
0.75390625
],
[
0.3447265625,
0.751953125,
0.359375,
0.751953125,
0.359375,
0.76220703125,
0.3447265625,
0.76220703125
],
[
0.42724609375,
0.7548828125,
0.435546875,
0.7548828125,
0.435546875,
0.7587890625,
0.42724609375,
0.7587890625
],
[
0.2900390625,
0.76953125,
0.7060546875,
0.76953125,
0.7060546875,
0.7841796875,
0.2900390625,
0.7841796875
],
[
0.39453125,
0.787109375,
0.41015625,
0.787109375,
0.41015625,
0.79443359375,
0.39453125,
0.79443359375
],
[
0.17724609375,
0.8037109375,
0.82177734375,
0.8037109375,
0.82177734375,
0.8173828125,
0.17724609375,
0.8173828125
],
[
0.1748046875,
0.81640625,
0.82177734375,
0.8173828125,
0.82177734375,
0.83203125,
0.1748046875,
0.83056640625
],
[
0.176513671875,
0.833984375,
0.64453125,
0.833984375,
0.64453125,
0.8486328125,
0.176513671875,
0.8486328125
],
[
0.1748046875,
0.85498046875,
0.82421875,
0.85498046875,
0.82421875,
0.869140625,
0.1748046875,
0.869140625
],
[
0.1732177734375,
0.869140625,
0.35546875,
0.869140625,
0.35546875,
0.88330078125,
0.1732177734375,
0.88330078125
],
[
0.1748046875,
0.8984375,
0.4755859375,
0.8984375,
0.4755859375,
0.912109375,
0.1748046875,
0.912109375
],
[
0.80810546875,
0.89794921875,
0.826171875,
0.89794921875,
0.826171875,
0.9111328125,
0.80810546875,
0.9111328125
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"合动作集A=×iVA。每个代理i都有一个策略元EI，：S×A,→[0,1]，我们将所有代",
"理的联合策略表示为元EII=×ievII，（Zhangetal.,2018）。对于联网的MARL，",
"一个常贝",
"且现实的假设是，每个代理i的奖励仅取决于其动作及其邻居（Quetal.,2020a）的动作，即。",
"e。，r;(s,a)=ri(s,an.）。此外，每个代理i可能仅获得部分观测值oiEOi，但可以通过",
"N，（Jiangetal.,2020）的观测值或观测历史（Chuetal.，2020）来近似状态，为简单起见，它们",
"均由oi表示。全球目标是最大化所有代理的累积奖励之和，即。e。，乙t=。Zi=1tr。",
"2.2Markov Game",
"在这样的设置中，每个代理可以单独最大化自已的预期收益，这被称为马尔可夫博奔。这可",
"能会导致稳定的结果或纳什均衡，但就全球目标而言，这通常是次优的。给定元，代理i的",
"值函数由",
"uT(s) =Zπ(als) pa(s [s,a)[r; + ut(s')]",
"(1)",
"给出，其中paEPa描述状态转换。纳什均衡定义为(Mgunietal.,2019)",
"(,-)(S) ≥u(\",-\")(s), V; eII, VseS, Vie V,",
"，其中-=xe。",
"3Method",
"LToS是一个分散的层次结构。在每个代理上，高层策略基于低层策略确定奖励共享的权重",
"而低层策略直接与环境交互以优化高层策略引起的本地目标。因此，它们形成了双层优化",
"并交替向全局目标学习。",
"3.1 Reward Sharing",
"奖励共享的直觉是，如果代理与他人共享奖励，则每个代理都必须考虑其行为对他人的后",
"果，从而促进了合作。在网络MARL中，由于假定代理的奖励取决于邻居的行为，因此我",
"们仅允许在相邻代理之间共享奖励。这是因为邻居的行为变化直接影响奖励，而邻居外的代",
"理只能通过状态分布的变化间接影响代理的回报。此外，这也符合邻居之间通信受限的网络",
"MARL的设置。对于V的图，我们还定义了一组由构造的有向边D。具体来说，我们为每",
"个代理i添加一个循环diiED，并将每个无向边eiiEε拆分为两个有向边：dii=（i,i）和",
"dji=（ji）D。每个代理i根据约束jeNwi=1为每个有向边di,ViN,确定权重",
"wiE[0,1]，以便wi代理i的环境奖励ri的比例将共享给代理。令wEW=XdjeDwi",
"为图的权重。因此，每个代理i共享后的成形奖励定义为",
"rw-wjirj.",
"(2)",
"jeN,",
"3.2Hierarchy",
"假设有一个联合高级策略ΦEΦ：S×W→[0,1]来确定w。给定Φ和w，我们可以根据",
"（1）将每个代理i处的元的值函数定义为",
"v(s; Φ) =d(ws) (a|s, w) pa(s|s,a)[rw + u(s'; Φ)],",
"(3)",
"w",
"a",
"v(s; w, Φ) = 元(a|s, w) > pa(s'|s,a)[r +v(s'; Φ)]",
"a",
"值得注意的是，W是分配方案的多维操作，而不是概率分布。在我们的推导中，为了简单",
"起见，我们将w表示为离散动作。只要我们将所有求和更改为积分，它也适用于连续操作。",
"设 V(s; ) = ZivvT(s; Φ) 和 Q(s, W; π) = ZievT(s; W,Φ) 。",
"Proposition 3.1. Given π , VΦ(s; ) and Q$(s, w; ) are respectively the value function and",
"action-value function of p.",
"Proof. The proof is deferred to Appendix ?? .",
"口",
"2"
] |
{
"arxiv_id": "2112.08702",
"language": "zh",
"page_idx": 1
}
|
||
2112.08702-zh-2
|
2112.08702-zh-2.jpg
|
命题 3.1 暗示$\phi$通过生成$w$直接优化全局目标，给定$\pi$ 。与现有的分层 RL 方法不同，我们可以在每个代理处基于$\pi$的值函数直接构造$\phi$的值函数和动作值函数。由于$\phi$优化了给定$\pi$的全局目标，而$\pi_i$在给定$\phi$的每个代理上分别优化了成形奖励(假设$\pi$收敛于纳什均衡或稳定结果，表示为$\lim)$,它们形成了双层优化。令$J_\phi(\pi)$和$J_{\pi}(\phi)$分别表示$\phi$和$\pi$ 的目标。双层优化可以表述如下，
$$\begin{array}{rl}\max_{\phi}&J_{\phi}(\pi^{*}(\phi))\\s.t.&\pi^{*}(\phi)=\arg\lim_{\pi}J_{\pi}(\phi).\end{array}$$
(4)
## 33 Decentralized Learning
我们从集体学习开始，实现平均奖励的全局优化。到目前为止，联合高层政策仍处于集中形式。请注意，场景需要一个分散的方法，每个代理都有自己的奖励。现在我们转向以分散的方式学习联合高层政策。设$w_i^\mathrm{out}\doteq\{w_{ij}|j\in\mathcal{N}_i\}$和$w_i^\mathrm{in}\doteq\{w_{ji}|j\in\mathcal{N}_i\}$ 。以下命题证明了每个代理在高层上的独立性。
Proposition $3. 2. \textit{The joint high level policy }\phi$ can be learned in a decentralized manner, and the
decentralized high-level policies of all agents form a mean-field approximation of $\phi$
口
$Proof.$ The proof is deferred to Appendix ??.
命题$\overline{\mathrm{E}}$3.1 和 3.2 表示，对于每个代理$i$ ,低级策略仅学习本地$\pi_i(a_i|s,w_i^\mathrm{in})$以优化$r_i^\mathbf{w}$的累积奖励，因为$r_i^{\boldsymbol{w}}$完全由$w_i^\mathrm{in}$根据$(i)\_1$。而高层策略$\phi_i$只需要在本地确定$w_i^\mathrm{out}$来优化$r_\nu^{\phi}$ 的累积奖励。
因此，对于分散式学习，(4)可以针对每个代理$i$在本地分解为
$$\begin{aligned}&\max_{\phi_{i}}\quad J_{\phi_{i}}(\phi_{-i},\pi_{1}^{*}(\phi),\cdots,\pi_{N}^{*}(\phi))\\&s.t.\quad\pi_{i}^{*}(\phi)=\arg\max_{\pi_{i}}J_{\pi_{i}}(\pi_{-i},\phi_{1}(\pi),\cdot\cdot,\phi_{N}(\pi)).\end{aligned}$$
(5)
现在我们使用 max 而不是$\lim$中的。我们滥用表示法，让$\phi$和$\pi$也分别表示它们的参数化。
为了解决优化，我们有
$$\begin{aligned}\nabla_{\phi_{i}}J_{\phi_{i}}(\phi_{-i},\pi_{1}^{*}(\phi),\cdots,\pi_{N}^{*}(\phi))\\&\approx\nabla_{\phi_{i}}J_{\phi_{i}}(\phi_{-i},\pi_{1}+\alpha\nabla_{\pi_{1}}J_{\pi_{1}}(\phi),\cdots,\pi_{N}+\alpha\nabla_{\pi_{N}}J_{\pi_{N}}(\phi)),\end{aligned}$$
(6)
,其中$\alpha$是低级策略的学习率。令$\pi_i^\prime$表示$\pi_i+\alpha\nabla_{\pi_i}J_{\pi_i}(\phi)$,我们有
$\nabla_{\phi_i}J_{\phi_i}(\phi_{-i},\pi_1^*(\phi),\cdots,\pi_N^*(\phi))$
$$\begin{aligned}&\approx\nabla_{\phi_{i}}J_{\phi_{i}}(\phi_{-i},\pi_{1}^{\prime},\cdots,\pi_{N}^{\prime})+\alpha\sum_{j=1}^{N}\nabla_{\phi_{i},\pi_{j}}^{2}J_{\pi_{j}}(\phi)\nabla_{\pi_{j}^{\prime}}J_{\phi_{i}}(\phi_{-i},\pi_{1}^{\prime},\cdots,\pi_{N}^{\prime}).\end{aligned}$$
由于计算复杂度高，二阶导数被忽略，而不会导致显著的性能下降，例如在元学习(Finn et al., 2017) 和神经架构搜索 (Liu et al., 2019) 中。不同的是，我们的低级政策需要一个以上的梯度步骤，直到收敛。同样，我们有
$$\nabla_{\pi_{i}}J_{\pi_{i}}(\pi_{-i},\phi_{1}^{*}(\pi),\cdots,\phi_{N}^{*}(\pi))\\\approx\nabla_{\pi_{i}}J_{\pi_{i}}(\pi_{-i},\phi_{1}+\beta\nabla_{\phi_{1}}J_{\phi_{1}}(\pi),\cdots,\phi_{N}+\beta\nabla_{\phi_{N}}J_{\phi_{N}}(\pi)),$$
,其中$\beta$是高级策略的学习率。因此，我们可以通过一阶近似以分散的方式求解双层优化(
4 )。对于每个代理$i$、$\phi_i$和$\pi_i$交替更新。
在分布式学习中，由于每个代理$i$通常无法访问状态，因此我们分别通过$\phi_i(w_i^\mathrm{out}|o_i)$和$\pi_i(a_i|o_i,w_i^\mathrm{in})$进一步近似$\phi_i(w_i^\mathrm{out}|s)$和$\pi_i(a_i|s,w_i^\mathrm{in})$。此外，在网络MARL 中 , 由 于 每 个 代 理 $i$ 都与相邻代理密切相关， 因此(5)可以进一步视为$\pi_i$在给定$\phi_{\mathcal{N}_i}$的情况下最大 化 了 $r_i^w$ 的 累 计 折 扣 奖 励 , 其 中 $\phi _{\mathcal{N} _i}= \times _{j\in \mathcal{N} _i}\phi _j$ 等 工 VAATH $= \text{ TH}_i\text{ H}_i\text{ HI}_i\text{ HI}_i\text{ HI}_i\text{ HI}_i$ MI 于 XMATH。在训练过程中，$\pi_{\mathcal{N}_{i}}$和$\phi_{\mathcal{N}_{i}}$分别被$w_{i}^\mathrm{ou}$
<FigureHere>
3
|
[
[
0.17724609375,
0.0921630859375,
0.82275390625,
0.0921630859375,
0.82275390625,
0.1060791015625,
0.17724609375,
0.1060791015625
],
[
0.17724609375,
0.106689453125,
0.82275390625,
0.106689453125,
0.82275390625,
0.12060546875,
0.17724609375,
0.12060546875
],
[
0.17724609375,
0.121826171875,
0.82177734375,
0.121826171875,
0.82177734375,
0.1351318359375,
0.17724609375,
0.1351318359375
],
[
0.176513671875,
0.1337890625,
0.82275390625,
0.1337890625,
0.82275390625,
0.147705078125,
0.176513671875,
0.147705078125
],
[
0.176513671875,
0.147705078125,
0.41015625,
0.147705078125,
0.41015625,
0.1610107421875,
0.176513671875,
0.1610107421875
],
[
0.39453125,
0.17041015625,
0.515625,
0.17041015625,
0.515625,
0.1849365234375,
0.39453125,
0.1849365234375
],
[
0.398681640625,
0.18310546875,
0.416748046875,
0.177978515625,
0.420654296875,
0.1875,
0.404296875,
0.1925048828125
],
[
0.80224609375,
0.186279296875,
0.826171875,
0.186279296875,
0.826171875,
0.199462890625,
0.80224609375,
0.199462890625
],
[
0.400390625,
0.1962890625,
0.6044921875,
0.1962890625,
0.6044921875,
0.210205078125,
0.400390625,
0.210205078125
],
[
0.17236328125,
0.2298583984375,
0.381591796875,
0.231689453125,
0.381591796875,
0.246826171875,
0.17236328125,
0.2442626953125
],
[
0.17724609375,
0.2587890625,
0.82177734375,
0.2587890625,
0.82177734375,
0.26953125,
0.17724609375,
0.26953125
],
[
0.1748046875,
0.270751953125,
0.82177734375,
0.26953125,
0.82177734375,
0.2841796875,
0.1748046875,
0.285400390625
],
[
0.1748046875,
0.28271484375,
0.82275390625,
0.2841796875,
0.82275390625,
0.298583984375,
0.1748046875,
0.29736328125
],
[
0.176513671875,
0.29736328125,
0.363525390625,
0.29736328125,
0.363525390625,
0.311279296875,
0.176513671875,
0.311279296875
],
[
0.176513671875,
0.317626953125,
0.82275390625,
0.317626953125,
0.82275390625,
0.330810546875,
0.176513671875,
0.330810546875
],
[
0.1748046875,
0.330810546875,
0.73046875,
0.330810546875,
0.73046875,
0.3447265625,
0.1748046875,
0.3447265625
],
[
0.1748046875,
0.361083984375,
0.4755859375,
0.361083984375,
0.4755859375,
0.375,
0.1748046875,
0.375
],
[
0.80810546875,
0.35986328125,
0.82666015625,
0.35986328125,
0.82666015625,
0.373779296875,
0.80810546875,
0.373779296875
],
[
0.176513671875,
0.389404296875,
0.82177734375,
0.389404296875,
0.82177734375,
0.40283203125,
0.176513671875,
0.40283203125
],
[
0.1748046875,
0.40478515625,
0.82177734375,
0.406005859375,
0.82177734375,
0.42041015625,
0.1748046875,
0.419189453125
],
[
0.17724609375,
0.421630859375,
0.2646484375,
0.421630859375,
0.2646484375,
0.43310546875,
0.17724609375,
0.43310546875
],
[
0.1748046875,
0.439453125,
0.63232421875,
0.43994140625,
0.63232421875,
0.455078125,
0.1748046875,
0.453857421875
],
[
0.32275390625,
0.46337890625,
0.564453125,
0.46337890625,
0.564453125,
0.477783203125,
0.32275390625,
0.477783203125
],
[
0.334228515625,
0.477783203125,
0.346435546875,
0.477783203125,
0.346435546875,
0.481689453125,
0.334228515625,
0.481689453125
],
[
0.80224609375,
0.479248046875,
0.826171875,
0.479248046875,
0.826171875,
0.49365234375,
0.80224609375,
0.49365234375
],
[
0.330078125,
0.4892578125,
0.67333984375,
0.4892578125,
0.67333984375,
0.50244140625,
0.330078125,
0.50244140625
],
[
0.1748046875,
0.51953125,
0.8203125,
0.51953125,
0.8203125,
0.53271484375,
0.1748046875,
0.53271484375
],
[
0.17724609375,
0.53271484375,
0.338134765625,
0.53271484375,
0.338134765625,
0.546875,
0.17724609375,
0.546875
],
[
0.2459716796875,
0.5537109375,
0.468017578125,
0.5546875,
0.468017578125,
0.5693359375,
0.2459716796875,
0.568359375
],
[
0.80224609375,
0.564453125,
0.82421875,
0.564453125,
0.82421875,
0.57763671875,
0.80224609375,
0.57763671875
],
[
0.3349609375,
0.57275390625,
0.75244140625,
0.57275390625,
0.75244140625,
0.5859375,
0.3349609375,
0.5859375
],
[
0.176513671875,
0.5966796875,
0.6689453125,
0.5966796875,
0.6689453125,
0.60986328125,
0.176513671875,
0.60986328125
],
[
0.2132568359375,
0.619140625,
0.43310546875,
0.619140625,
0.43310546875,
0.63330078125,
0.2132568359375,
0.63330078125
],
[
0.498291015625,
0.6337890625,
0.5146484375,
0.64013671875,
0.50732421875,
0.65087890625,
0.491943359375,
0.64453125
],
[
0.261474609375,
0.646484375,
0.78515625,
0.6484375,
0.78515625,
0.66943359375,
0.261474609375,
0.66748046875
],
[
0.53515625,
0.65576171875,
0.55126953125,
0.65576171875,
0.55126953125,
0.66015625,
0.53515625,
0.66015625
],
[
0.490234375,
0.66650390625,
0.5146484375,
0.66650390625,
0.5146484375,
0.6787109375,
0.490234375,
0.6787109375
],
[
0.17724609375,
0.689453125,
0.82177734375,
0.689453125,
0.82177734375,
0.70263671875,
0.17724609375,
0.70263671875
],
[
0.176513671875,
0.70263671875,
0.82275390625,
0.70263671875,
0.82275390625,
0.71630859375,
0.176513671875,
0.71630859375
],
[
0.176513671875,
0.71630859375,
0.451904296875,
0.71630859375,
0.451904296875,
0.72998046875,
0.176513671875,
0.72998046875
],
[
0.2467041015625,
0.7392578125,
0.468017578125,
0.7392578125,
0.468017578125,
0.75244140625,
0.2467041015625,
0.75244140625
],
[
0.3349609375,
0.755859375,
0.7509765625,
0.755859375,
0.7509765625,
0.76953125,
0.3349609375,
0.76953125
],
[
0.178955078125,
0.77978515625,
0.82275390625,
0.77978515625,
0.82275390625,
0.79345703125,
0.178955078125,
0.79345703125
],
[
0.17724609375,
0.79345703125,
0.4697265625,
0.79345703125,
0.4697265625,
0.806640625,
0.17724609375,
0.806640625
],
[
0.176513671875,
0.814453125,
0.54833984375,
0.814453125,
0.54833984375,
0.82763671875,
0.176513671875,
0.82763671875
],
[
0.75634765625,
0.8193359375,
0.7900390625,
0.8193359375,
0.7900390625,
0.83203125,
0.75634765625,
0.83203125
],
[
0.1732177734375,
0.82763671875,
0.54833984375,
0.8271484375,
0.54833984375,
0.84130859375,
0.1732177734375,
0.84228515625
],
[
0.188720703125,
0.84326171875,
0.54833984375,
0.84326171875,
0.54833984375,
0.85693359375,
0.188720703125,
0.85693359375
],
[
0.5537109375,
0.8447265625,
0.60888671875,
0.8447265625,
0.60888671875,
0.85498046875,
0.5537109375,
0.85498046875
],
[
0.6953125,
0.84326171875,
0.75732421875,
0.84326171875,
0.75732421875,
0.85791015625,
0.6953125,
0.85791015625
],
[
0.1732177734375,
0.85595703125,
0.5439453125,
0.85791015625,
0.5439453125,
0.8720703125,
0.1732177734375,
0.869140625
],
[
0.60888671875,
0.85400390625,
0.6650390625,
0.85400390625,
0.6650390625,
0.86572265625,
0.60888671875,
0.86572265625
],
[
0.176513671875,
0.87060546875,
0.54736328125,
0.87060546875,
0.54736328125,
0.88525390625,
0.176513671875,
0.88525390625
],
[
0.1732177734375,
0.880859375,
0.5498046875,
0.88330078125,
0.5498046875,
0.90087890625,
0.1732177734375,
0.8984375
],
[
0.17724609375,
0.8984375,
0.5439453125,
0.8984375,
0.5439453125,
0.91357421875,
0.17724609375,
0.91357421875
],
[
0.640625,
0.9013671875,
0.65283203125,
0.9013671875,
0.65283203125,
0.90771484375,
0.640625,
0.90771484375
],
[
0.689453125,
0.90087890625,
0.7607421875,
0.90283203125,
0.7587890625,
0.92041015625,
0.68896484375,
0.91845703125
],
[
0.61767578125,
0.91552734375,
0.62841796875,
0.91552734375,
0.62841796875,
0.91845703125,
0.61767578125,
0.91845703125
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
],
[
0.736328125,
0.93115234375,
0.74853515625,
0.93115234375,
0.74853515625,
0.93994140625,
0.736328125,
0.93994140625
],
[
0.7060546875,
0.94873046875,
0.71142578125,
0.94873046875,
0.71142578125,
0.9541015625,
0.7060546875,
0.9541015625
],
[
0.73779296875,
0.943359375,
0.74853515625,
0.943359375,
0.74853515625,
0.958984375,
0.73779296875,
0.958984375
],
[
0.59619140625,
0.9736328125,
0.640625,
0.9736328125,
0.640625,
0.984375,
0.59619140625,
0.984375
],
[
0.62353515625,
0.9697265625,
0.64892578125,
0.9697265625,
0.64892578125,
0.97265625,
0.62353515625,
0.97265625
],
[
0.6953125,
0.974609375,
0.75634765625,
0.974609375,
0.75634765625,
0.98486328125,
0.6953125,
0.98486328125
]
] |
[
"命题3.1暗示Φ通过生成W直接优化全局目标，给定元。与现有的分层RL方法不同，我",
"们可以在每个代理处基于元的值函数直接构造Φ的值函数和动作值函数。由于Φ优化了给",
"定元的全局目标，而元在给定Φ的每个代理上分别优化了成形奖励（假设元收敛于纳什",
"均衡或稳定结果，表示为lim），它们形成了双层优化。令J（元）和J元（Φ）分别表示Φ和元",
"的目标。双层优化可以表述如下，",
"maxJs(π*(Φ)",
"",
"(4)",
"s.t. π*(Φ) =arglim J(Φ).",
"3.3Decentralized Learning",
"我们从集体学寸开始，实现平均奖励的全局优化。到自前为止，联合高层政策仍处于集中形",
"式。请注意，场景需要一个分散的方法，每个代理都有自己的奖励。现在我们转向以分散的",
"方式学习联合高层政策。设wout=（wilieN)和wn=【wiiliEN。以下命题证明了每",
"个代理在高层上的独立性。",
"Proposition 3.2. The joint high level policy can be learned in a decentralized manner, and the",
"decentralized high-level policies of all agents form a mean-field approximation of p .",
"Proof. The proof is deferred to Appendix ??.",
"口",
"命题E3.1和3.2表示，对于每个代理i，低级策略仅学习本地π（aiswin）以优化r\"的累",
"积奖励，因为rw完全由win根据（i）_1。而高层策略Φ只需要在本地确定wout来优化r中",
"的累积奖励。",
"因此，对于分散式学习，（4）可以针对每个代理i在本地分解为",
"maxJs, (Φ-,元(Φ), ,元(Φ))",
"4",
"(5)",
"s.t,(Φ)=argmaxJ(元i,1(元),,Φ(元))",
"现在我们使用max而不是lim中的。我们滥用表示法，让Φ和π也分别表示它们的参数化",
"为了解决优化，我们有",
"VJ(@-i,i(), ..,(Φ)",
"(6)",
"~ VJ(,i+QJ(),..:,T+VJ())",
"，其中α是低级策略的学习率。令元，表示元十α√元，J元（Φ），我们有",
"V Jo(Φ-i, (Φ), -*, 元(Φ))",
"N",
"~V, Jo(, ,,)+a, J,() o (, i,,)",
"",
"3=1",
"由于计算复杂度高，二阶导数被忽略，而不会导致显著的性能下降，例如在元学习（Finn",
"etal.,2017)和神经架构搜索（Liuetal.，2019）中。不同的是，我们的低级政策需要一个以上",
"的梯度步骤，直到收敛。同样，我们有",
"V元J元,(_i,Φ(元), ...,(元))",
"~V,J (i,Φ1+βVJ(),,Φ+βVJ()",
"，其中β是高级策略的学习率。因此，我们可以通过一阶近似以分散的方式求解双层优化（",
"4）。对于每个代理i、中和元交替更新。",
"在分布式学习中，由于每个代理通常无法访问状",
"agent i",
"态，因此我们分别通过中（woutoi）和（aioi，win）进",
"步近似Φ(wouts）和(ais,wi）。此外，在网络",
"neighbors",
"p:(wputlo.)",
"MARL中，由于每个代理i都与相邻代理密切相关",
"wputw",
"因此（5）可以进一步视为元在给定ΦN的情况下最",
"化了ru的累计折扣奖励，其中n=ie,等",
"于XMATH。在训练过程中，元N和N分别被wou",
"w",
";(ai|0i,up)",
".",
"3",
"",
"Z",
"",
"data flow",
"EO",
"environment"
] |
{
"arxiv_id": "2112.08702",
"language": "zh",
"page_idx": 2
}
|
||
2112.08702-zh-3
|
2112.08702-zh-3.jpg
|
<FigureHere>
jungle
Figure 2: Three experimental scenarios: (a) prisoner, (b) jungle , and (c) traffic
和$w_i^\mathrm{in}$的交互隐式考虑。LToS 的体系结构如图 1 所示。在每个时间步，每个代理$i$的高层策略根据观察结果决定动作$w_i^\mathrm{out}$作为奖励共享的权重。然后，低级策略将观察和$w_i^\mathrm{in}$作为输人，并输出操作。代理$i$根据$w_i^\mathrm{in}$获得高级和低级策略的成形奖励。梯度沿紫色虚线反向传播。
此外，从命题 3.1 ,我们有：$q_i^{\phi_i}(s,w_i^\mathrm{out};\pi_{\mathcal{N}_i})=$ $v_i^{\pi _i}( s; w_i^\mathrm{in}, \phi _{\mathcal{N} _i})$ ,其中$q_i^{\phi_i}$是$\phi_i$的动作值函数，给定$\pi_{\mathcal{N}_i},v_i^{\pi_i}$是 $\pi_i$ 的值函数，给定 $\phi_\mathcal{N}_i$ 并以 $w_i^\mathrm{in}$。如前所述，我们大约有 $q_i^\phi_i(o_i,w_i^\mathrm{out})=v_i^{\pi_i}(o_i;w_i^\mathrm{in})$ 。我们可以看到，$\phi_i$的动作值函数等同于$\pi_i$的值函数。也就是说，我们可以使用单个网络同时近似这两个函数。对于确定性的低层策略，高层和低层策略可以共享相同的行动价值函数。在 LToS 的当前实例化中，我们将 DGN(Jiang et al.,2020) (Q 学习)用于低级策略， 将 DDPG (Lillicrap et al., 2016) 用于高级策略。因此，DGN 的 Q 网络也是 DDPG 的批评者$w_i^\mathrm{in}$的梯度是根据$a_i$的最大 Q 值计算的。为了完整性，算法 1 (见附录 6 )给出了基于 DDPG 和 DGN 的 LToS 训练过程。附录 7 中也提供了有关培训 LToS 的更多讨论。
## 4 $\textbf{Experiments}$
为了进行实验，我们采用了图$\overline{\mathrm{E}}$2 中描述的三种场景：囚犯，丛林和交通，其中囚犯和丛林 (Jiang et al., 2020) 是关于社会困境的网格游戏，可以轻松衡量代理的合作，而交通是网络 MARL 的现实场景。我们遵循网络 MARL 的原则，只允许在 Zhang et al.(2018) 和 Chu et al. (2020) 附近进行通信。为了说明每个代理学习的奖励共享方案，我们使用一个简单的指标：自私，即代理选择为自己保留的奖励比例。对于消融，我们保持每个代理的共享权重固定，称为固定 LToS。在整个实验过程中，我们还与包括 DQN 和 DGN 在内的基线进行了比较，其中 DGN 还可以在没有奖励共享的情况下消融 LToS,因为 DGN 是 LToS 的低级策略。为了通过集中学习直接最大化全球回报，我们在三个场景中使用 QMIX (Rashid et al., 2018)作为基线，在囚犯中使用另外两个场景。此外，由于 LToS 旨在通过网络 MARL 中的奖励共享带来和谐的合作，我们将 LToS 与网络 MARL 的三种方法进行了比较，即。e。, ConseNet (Zhang et al., 2018) , NeurComm (Chu et al., 2020) 和 IntentionPropagation (缩写为IP) (Qu et al., 2020a) 和 LIO (Yang et al., 2020) 用于激励学习，所有这些都使用递归神经网络$(\mathbf{RNN})$或图形神经网络 (GNN) 用于部分可观察的环境。有关超参数的更多详细信息，请参见附录 8 。
## 4.1 Prisoner
我们使用来自 Sodomka et al.(2013) 的著名矩阵博弈囚徒困境的网格博弈版本来实证证明LToS 能够学习合作策略以实现全局最优(i。e。,最大化全球平均回报)。如图$w\in\mathcal{W}=$ $\times_{d_{ij}\in\mathcal{D}}w_{ij}$所示，有两个代理$A$和$B$,它们分别从网格走廊中间的两侧开始，进行全面观察。在每个时间步，每个代理选择一个左或右的动作，并移动到相应的相邻网格，并且每个动作都会产生一个成本-0.01 。有三个目标，两端两个目标，中间一个目标。agent 达到目标获得奖励+1 。一旦某个代理达到目标或两个代理同时达到不同的目标，游戏就结束了。这个游戏类似于囚徒困境：追求中间目标(“缺陷”)将带来比更远的目标(“合作”)更多的
4
|
[
[
0.44677734375,
0.1060791015625,
0.45751953125,
0.1060791015625,
0.45751953125,
0.11553955078125,
0.44677734375,
0.11553955078125
],
[
0.5439453125,
0.10418701171875,
0.5556640625,
0.10418701171875,
0.5556640625,
0.11553955078125,
0.5439453125,
0.11553955078125
],
[
0.69287109375,
0.10797119140625,
0.7158203125,
0.111083984375,
0.71142578125,
0.12249755859375,
0.68896484375,
0.11932373046875
],
[
0.463134765625,
0.1363525390625,
0.5361328125,
0.1363525390625,
0.5361328125,
0.147705078125,
0.463134765625,
0.147705078125
],
[
0.285888671875,
0.1610107421875,
0.298095703125,
0.1610107421875,
0.298095703125,
0.1654052734375,
0.285888671875,
0.1654052734375
],
[
0.398681640625,
0.184326171875,
0.414306640625,
0.184326171875,
0.414306640625,
0.1912841796875,
0.398681640625,
0.1912841796875
],
[
0.306396484375,
0.20703125,
0.330810546875,
0.20703125,
0.330810546875,
0.21533203125,
0.306396484375,
0.21533203125
],
[
0.6337890625,
0.2139892578125,
0.646484375,
0.2139892578125,
0.646484375,
0.2183837890625,
0.6337890625,
0.2183837890625
],
[
0.625,
0.2215576171875,
0.6845703125,
0.2215576171875,
0.6845703125,
0.23291015625,
0.625,
0.23291015625
],
[
0.6298828125,
0.2342529296875,
0.71728515625,
0.236083984375,
0.71728515625,
0.246826171875,
0.6298828125,
0.2442626953125
],
[
0.3896484375,
0.24365234375,
0.410888671875,
0.242431640625,
0.410888671875,
0.246826171875,
0.3896484375,
0.2474365234375
],
[
0.343017578125,
0.255126953125,
0.40185546875,
0.255126953125,
0.40185546875,
0.26513671875,
0.343017578125,
0.26513671875
],
[
0.5888671875,
0.251953125,
0.64794921875,
0.251953125,
0.64794921875,
0.26318359375,
0.5888671875,
0.26318359375
],
[
0.236083984375,
0.272705078125,
0.7646484375,
0.272705078125,
0.7646484375,
0.285888671875,
0.236083984375,
0.285888671875
],
[
0.176513671875,
0.302978515625,
0.54833984375,
0.302978515625,
0.54833984375,
0.31640625,
0.176513671875,
0.31640625
],
[
0.1748046875,
0.31494140625,
0.54736328125,
0.31640625,
0.54736328125,
0.330810546875,
0.1748046875,
0.330078125
],
[
0.1748046875,
0.328857421875,
0.54736328125,
0.330078125,
0.54736328125,
0.3447265625,
0.1748046875,
0.343505859375
],
[
0.176513671875,
0.34521484375,
0.54833984375,
0.34521484375,
0.54833984375,
0.359130859375,
0.176513671875,
0.359130859375
],
[
0.1748046875,
0.35986328125,
0.5458984375,
0.35986328125,
0.5458984375,
0.373779296875,
0.1748046875,
0.373779296875
],
[
0.17724609375,
0.375732421875,
0.279296875,
0.375732421875,
0.279296875,
0.386962890625,
0.17724609375,
0.386962890625
],
[
0.1748046875,
0.396484375,
0.5458984375,
0.397705078125,
0.5458984375,
0.412353515625,
0.1748046875,
0.410888671875
],
[
0.176513671875,
0.4140625,
0.54833984375,
0.4140625,
0.54833984375,
0.427978515625,
0.176513671875,
0.427978515625
],
[
0.1748046875,
0.4287109375,
0.82275390625,
0.4287109375,
0.82275390625,
0.4462890625,
0.1748046875,
0.4462890625
],
[
0.188720703125,
0.4443359375,
0.82177734375,
0.443115234375,
0.82177734375,
0.458251953125,
0.188720703125,
0.458984375
],
[
0.17724609375,
0.458984375,
0.82177734375,
0.458984375,
0.82177734375,
0.472900390625,
0.17724609375,
0.472900390625
],
[
0.17724609375,
0.472900390625,
0.82177734375,
0.472900390625,
0.82177734375,
0.486083984375,
0.17724609375,
0.486083984375
],
[
0.1748046875,
0.486083984375,
0.8203125,
0.486083984375,
0.8203125,
0.50048828125,
0.1748046875,
0.50048828125
],
[
0.176513671875,
0.5,
0.82421875,
0.5,
0.82421875,
0.5146484375,
0.176513671875,
0.5146484375
],
[
0.1748046875,
0.5146484375,
0.7060546875,
0.5146484375,
0.7060546875,
0.52783203125,
0.1748046875,
0.52783203125
],
[
0.17236328125,
0.54736328125,
0.31787109375,
0.548828125,
0.31787109375,
0.56298828125,
0.17236328125,
0.5625
],
[
0.176513671875,
0.57763671875,
0.8203125,
0.57763671875,
0.8203125,
0.59130859375,
0.176513671875,
0.59130859375
],
[
0.1748046875,
0.59130859375,
0.8203125,
0.59130859375,
0.8203125,
0.60498046875,
0.1748046875,
0.60498046875
],
[
0.1748046875,
0.60595703125,
0.82177734375,
0.60595703125,
0.82177734375,
0.619140625,
0.1748046875,
0.619140625
],
[
0.1748046875,
0.619140625,
0.82177734375,
0.619140625,
0.82177734375,
0.6337890625,
0.1748046875,
0.6337890625
],
[
0.176513671875,
0.6337890625,
0.82177734375,
0.6337890625,
0.82177734375,
0.64794921875,
0.176513671875,
0.64794921875
],
[
0.176513671875,
0.646484375,
0.82177734375,
0.64599609375,
0.82177734375,
0.66015625,
0.176513671875,
0.6611328125
],
[
0.17724609375,
0.6611328125,
0.82177734375,
0.6611328125,
0.82177734375,
0.6748046875,
0.17724609375,
0.6748046875
],
[
0.176513671875,
0.673828125,
0.82177734375,
0.6748046875,
0.82177734375,
0.689453125,
0.176513671875,
0.68798828125
],
[
0.176513671875,
0.68798828125,
0.82275390625,
0.68798828125,
0.82275390625,
0.7021484375,
0.176513671875,
0.7021484375
],
[
0.1748046875,
0.70068359375,
0.82275390625,
0.70263671875,
0.82275390625,
0.71728515625,
0.1748046875,
0.71533203125
],
[
0.176513671875,
0.71630859375,
0.82177734375,
0.71630859375,
0.82177734375,
0.72998046875,
0.176513671875,
0.72998046875
],
[
0.17724609375,
0.72998046875,
0.82177734375,
0.72998046875,
0.82177734375,
0.74365234375,
0.17724609375,
0.74365234375
],
[
0.1748046875,
0.7421875,
0.82275390625,
0.74365234375,
0.82275390625,
0.75830078125,
0.1748046875,
0.7568359375
],
[
0.17724609375,
0.7587890625,
0.26953125,
0.7587890625,
0.26953125,
0.77099609375,
0.17724609375,
0.77099609375
],
[
0.1732177734375,
0.7861328125,
0.27685546875,
0.787109375,
0.27685546875,
0.8017578125,
0.1732177734375,
0.80029296875
],
[
0.176513671875,
0.81298828125,
0.82275390625,
0.81298828125,
0.82275390625,
0.8271484375,
0.176513671875,
0.8271484375
],
[
0.1748046875,
0.82763671875,
0.82177734375,
0.82763671875,
0.82177734375,
0.84130859375,
0.1748046875,
0.84130859375
],
[
0.1748046875,
0.84033203125,
0.82421875,
0.83837890625,
0.82421875,
0.85595703125,
0.1748046875,
0.85791015625
],
[
0.17724609375,
0.859375,
0.8203125,
0.859375,
0.8203125,
0.869140625,
0.17724609375,
0.869140625
],
[
0.176513671875,
0.87060546875,
0.82177734375,
0.87060546875,
0.82177734375,
0.8837890625,
0.176513671875,
0.8837890625
],
[
0.176513671875,
0.8837890625,
0.82177734375,
0.8837890625,
0.82177734375,
0.89794921875,
0.176513671875,
0.89794921875
],
[
0.1748046875,
0.8984375,
0.82275390625,
0.8984375,
0.82275390625,
0.912109375,
0.1748046875,
0.912109375
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"A",
"B",
"",
"(a) prisoner",
"Bwe",
"十",
"酒",
"",
"内口口",
"OOOOn",
"ingTox",
"(b)jungle",
"(c) traffic",
"Figure 2: Three experimental scenarios: (a) prisoner, (b) jungle, and (c) traffic :",
"和win的交互隐式考虑。LToS的体系结构如图1所",
"示。在每个时间步，每个代理i的高层策略根据观察",
"结果决定动作wgut作为奖励共享的权重。然后，低级",
"策略将观察和win作为输入，并输出操作。代理i根",
"据wi获得高级和低级策略的成形奖励。梯度沿紫色",
"虚线反向传播。",
"此外，从命题3.1，我们有：q（s,wou；元）=",
"u元（s;win，n），其中qe是Φi的动作值函数，给定",
"元，是元的值函数，给定中并以win。如前所述，我们大约有q（oi，wour）=（oi；win）",
"我们可以看到，Φ的动作值函数等同于元，的值函数。也就是说，我们可以使用单个网络",
"同时近似这两个函数。对于确定性的低层策略，高层和低层策略可以共享相同的行动价值",
"函数。在LToS的当前实例化中，我们将DGN（Jiangetal.,2020）（Q学习）用于低级策略",
"将DDPG(Lillicrapetal.，2016)用于高级策略。因此，DGN的Q网络也是DDPG的批评者",
"win的梯度是根据ai的最大Q值计算的。为了完整性，算法1（见附录6）给出了基于DDPG",
"和DGN的LToS训练过程。附录7中也提供了有关培训LToS的更多讨论。",
"4Experiments",
"为了进行实验，我们采用了图E2中描述的三种场景：囚犯，丛林和交通，其中囚犯和丛",
"林（Jiangetal.，2020）是关于社会困境的网格游戏，可以轻松衡量代理的合作，而交通是网",
"络MARL的现实场景。我们遵循网络MARL的原则，只允许在Zhangetal.（2018）和Chu",
"etal.（2020)附近进行通信。为了说明每个代理学习的奖励共享方案，我们使用一个简单的",
"指标：自私，即代理选择为自己保留的奖励比例。对于消融，我们保持每个代理的共享权",
"重固定，称为固定LToS。在整个实验过程中，我们还与包括DQN和DGN在内的基线进行",
"了比较，其中DGN还可以在没有奖励共享的情况下消融LToS，因为DGN是LToS的低级",
"策略。为了通过集中学习直接最大化全球回报，我们在三个场景中使用QMIX（Rashidetal.",
"2018）作为基线，在囚犯中使用另外两个场景。此外，由于LToS旨在通过网络MARL中",
"的奖励共享带来和谐的合作，我们将LToS与网络MARL的三种方法进行了比较，即。e。",
"ConseNet (Zhang et al.,2018)，NeurComm(Chu et al.,2020)和 IntentionPropagation（缩写为",
"IP）（Quetal.,2020a）和LIO（Yangetal.2020)用于激励学习，所有这些都使用递归神经网络",
"（RNN）或图形神经网络（GNN）用于部分可观察的环境。有关超参数的更多详细信息，请",
"参见附录8。",
"4.1Prisoner",
"我们使用来自Sodomkaetal.（2013）的著名矩阵博奔囚徒困境的网格博奔版本来实证证明",
"LToS能够学习合作策略以实现全局最优(i。e。，最大化全球平均回报)。如图wEW",
"XdieDwi所示，有两个代理A和B，它们分别从网格走廊中间的两侧开始，进行全面观",
"察。在每个时间步，每个代理选择一个左或右的动作，并移动到相应的相网格，并且每个",
"动作都会产生一个成本一0.01。有三个目标，两端两个目标，中间一个目标。agent达到目",
"标获得奖励+1。一旦某个代理达到目标或两个代理同时达到不同的目标，游戏就结束了。",
"这个游戏类似于囚徒困境：追求中间目标（“缺陷”）将带来比更远的目标（“合作”）更多的",
"4"
] |
{
"arxiv_id": "2112.08702",
"language": "zh",
"page_idx": 3
}
|
||
2112.08702-zh-4
|
2112.08702-zh-4.jpg
|
<FigureHere>
Figure 3: Learning curves in (a) prisoner, (b) jungle, and (c) traffic. All the curves are plotted using 5 training runs with different random seeds, where the solid line is the mean and the shadowed area is enclosed by the min and max value.
回报，但是如果两个代理都采用这种方式，则会发生碰撞，并且只有一个代理以相等的概率赢得目标。相反，如果两个代理商都“合作”,则可以获得更高的回报，尽管这需要更多的步骤。可能的最高回报是 1 图 3a 从平均收益的角度说明了所有方法的学习曲线。请注意， 对于所有三种情况，我们用实线表示不同随机种子的 5 次训练运行的平均值，用阴影区域表示最小$V$最大值。作为自利优化的结果，DQN 收敛到“缺陷/缺陷”纳什均衡，其中每个代理都收到关于 0.5 的预期奖励。DGN 也是如此，因为它的目的只是利用邻居的观察结果，而囚犯已经是一个完全可观察的环境。给定一个手动调整的奖励塑造因素来指导代理以最大化平均回报，NeurComm 和固定 LToS 代理最终能够合作。ConseNet 和 QMIX 也是如此。然而，它们收敛缓慢。相比之下，IP代理的学习速度较慢，其性能仅比 0.5 高一点。LIO 特工在开始时很快就进行了合作，但他们无法形成稳定的合作，也许需要更长时间的培训才能摆脱这种不稳定 JointDQN、Coco-Q (Sodomka et al.,2013)和 LToS 的性能类似，优于其他方法。JointDQN 是一个集中式 DQN, 它控制代理$A$和$B$的联合行动，因此应该能够实现最佳性能，但即使在这样一个简单的双代理场景中，仍然需要时间来收敛。作为一种改进的表格 Q 学习方法，Coco-Q 引人 coco 值(Kalai and Kalai, 2010) 作为 Bellman 方程中期望收益的替代，并将差值视为转移奖励。但是，它是专门为某些游戏设计的，并且很难扩展到双人游戏之外。LToS 可以学习奖励共享方案，其中一个代理首先将所有奖励给予另一个代理，以便防止它们都出现“缺陷”,从而快速实现最佳平均回报，如实验中所观察到的。通过囚犯， 我们验证了 LToS 可以通过学习分享奖励来逃避局部最优。
# 4.2 Jungle
丛林是 Jiang et al. (2020) 基于 MAgent (Zheng et al., 2017) 提出的关于道德困境的场景。如图 2b 所示，有$N$剂和$L$固定食品。在每个时间步，每个代理可以攻击或移动到一个相邻的网格。吃 (攻击食物)带来正奖励+1,但攻击其他代理获得更高的奖励+2。然而，受害者遭受负奖励-4,这使得代理之间的每次攻击都是负和动作。此外，攻击空白网格会得到一个小的负奖励-0.01 (抑制过度攻击)。我们遵循 Jiang et al.(2020)的原始设置：地图大小$=30\times30$ 网格，$N= 20$ , $L= 12$ , 观察值由一个人的坐标和附近的 $11\times11$ 网格的字段组成。每个代理都有 3 个最接近的代理作为其邻居。与囚犯相比，丛林有更多的特工，因此 JointDQN 和 Coco-Q 在这种情况下被忽略。丛林的另一个挑战是网络拓扑是动态的，因为每个代理总是可以移动。幸运的是，拓扑变化缓慢且可预测，算法可能会将时变邻居集$\mathcal{N}_i$作为输人的一部分。因此，仍然有可能很好地估计成形的奖励和价值函数。
Table 1: Average reward per step of all the methods in jungle .
<table>
<tbody>
<tr>
<th>$DQN$</th>
<th>$DGN$</th>
<th>fixed LToS</th>
<th>LToS</th>
<th>NeurComm</th>
<th>ConseNet</th>
<th>$LIO$</th>
<th>IP</th>
<th>$QMIX$</th>
<th>bound $\operatorname{upper}$</th>
</tr>
<tr>
<td>0.24</td>
<td>0.66</td>
<td>0.71</td>
<td>0.86</td>
<td>-0.05</td>
<td>-0.04</td>
<td>0.00</td>
<td>0.63</td>
<td>0.04</td>
<td>0.95</td>
</tr>
</tbody>
</table>
图 3b 说明了所有方法的学习曲线，它们在收敛后的性能也总结在表 1 中。NeurComm, ConseNet 和 QMIX 在此任务中表现不佳。在 NeurComm 中，每个代理都会获得“延迟的全局信息”。但是，当通过动态拓扑进行通信时，无法形成延迟的全局信息的稳定模式。ConseNet 是在充分观察的前提下构建的，当输入不仅是部分的，而且在顺序和内容上变化很大时，它很难学习。LIO 代理也表现不佳，因为它们在动态拓扑中无法相互区分，因此无法学习适当的激励函数。此外，LIO 需要对手建模，但很难在动态环境中同时对所有其他代理进行建模。QMIX 没有这些问题。在针对全局优化时，如 LIO, 它意识到攻击通常意味着负和动作，
5
|
[
[
0.418212890625,
0.1142578125,
0.4248046875,
0.11932373046875,
0.418212890625,
0.125,
0.410888671875,
0.11932373046875
],
[
0.529296875,
0.18310546875,
0.53759765625,
0.18310546875,
0.53759765625,
0.186279296875,
0.529296875,
0.186279296875
],
[
0.4345703125,
0.1925048828125,
0.529296875,
0.1956787109375,
0.529296875,
0.20703125,
0.4345703125,
0.203857421875
],
[
0.53173828125,
0.1962890625,
0.59228515625,
0.1962890625,
0.59228515625,
0.203857421875,
0.53173828125,
0.203857421875
],
[
0.5908203125,
0.1976318359375,
0.6044921875,
0.1976318359375,
0.6044921875,
0.2020263671875,
0.5908203125,
0.2020263671875
],
[
0.67626953125,
0.1976318359375,
0.6845703125,
0.1976318359375,
0.6845703125,
0.20263671875,
0.67626953125,
0.20263671875
],
[
0.7099609375,
0.1988525390625,
0.72802734375,
0.1988525390625,
0.72802734375,
0.20263671875,
0.7099609375,
0.20263671875
],
[
0.74267578125,
0.1988525390625,
0.7490234375,
0.1988525390625,
0.7490234375,
0.2020263671875,
0.74267578125,
0.2020263671875
],
[
0.775390625,
0.1956787109375,
0.7861328125,
0.1976318359375,
0.78515625,
0.20263671875,
0.77392578125,
0.199462890625
],
[
0.80810546875,
0.1988525390625,
0.81640625,
0.1962890625,
0.81787109375,
0.199462890625,
0.8095703125,
0.2020263671875
],
[
0.473876953125,
0.21337890625,
0.53369140625,
0.21337890625,
0.53369140625,
0.2247314453125,
0.473876953125,
0.2247314453125
],
[
0.689453125,
0.21337890625,
0.74853515625,
0.21337890625,
0.74853515625,
0.2247314453125,
0.689453125,
0.2247314453125
],
[
0.1748046875,
0.2342529296875,
0.82421875,
0.2342529296875,
0.82421875,
0.2474365234375,
0.1748046875,
0.2474365234375
],
[
0.250732421875,
0.216552734375,
0.324462890625,
0.216552734375,
0.324462890625,
0.2266845703125,
0.250732421875,
0.2266845703125
],
[
0.1748046875,
0.2474365234375,
0.82421875,
0.2474365234375,
0.82421875,
0.261474609375,
0.1748046875,
0.261474609375
],
[
0.176513671875,
0.26318359375,
0.453369140625,
0.26318359375,
0.453369140625,
0.2734375,
0.176513671875,
0.2734375
],
[
0.176513671875,
0.296630859375,
0.82421875,
0.296630859375,
0.82421875,
0.31005859375,
0.176513671875,
0.31005859375
],
[
0.1748046875,
0.309326171875,
0.498291015625,
0.30810546875,
0.498291015625,
0.322509765625,
0.1748046875,
0.323974609375
],
[
0.50341796875,
0.314453125,
0.5107421875,
0.314453125,
0.5107421875,
0.317626953125,
0.50341796875,
0.317626953125
],
[
0.50634765625,
0.31005859375,
0.8203125,
0.31005859375,
0.8203125,
0.32080078125,
0.50634765625,
0.32080078125
],
[
0.1748046875,
0.322509765625,
0.8203125,
0.322509765625,
0.8203125,
0.3359375,
0.1748046875,
0.3359375
],
[
0.176513671875,
0.337158203125,
0.82421875,
0.337158203125,
0.82421875,
0.35107421875,
0.176513671875,
0.35107421875
],
[
0.176513671875,
0.35107421875,
0.82275390625,
0.35107421875,
0.82275390625,
0.3642578125,
0.176513671875,
0.3642578125
],
[
0.176513671875,
0.365478515625,
0.82275390625,
0.365478515625,
0.82275390625,
0.37890625,
0.176513671875,
0.37890625
],
[
0.176513671875,
0.37744140625,
0.82177734375,
0.37744140625,
0.82177734375,
0.392578125,
0.176513671875,
0.392578125
],
[
0.176513671875,
0.391357421875,
0.82275390625,
0.391357421875,
0.82275390625,
0.40478515625,
0.176513671875,
0.40478515625
],
[
0.176513671875,
0.406005859375,
0.82275390625,
0.406005859375,
0.82275390625,
0.42041015625,
0.176513671875,
0.42041015625
],
[
0.176513671875,
0.42041015625,
0.82421875,
0.42041015625,
0.82421875,
0.434326171875,
0.176513671875,
0.434326171875
],
[
0.176513671875,
0.434326171875,
0.82275390625,
0.434326171875,
0.82275390625,
0.447509765625,
0.176513671875,
0.447509765625
],
[
0.176513671875,
0.447509765625,
0.82275390625,
0.447509765625,
0.82275390625,
0.4609375,
0.176513671875,
0.4609375
],
[
0.17724609375,
0.4609375,
0.82275390625,
0.4609375,
0.82275390625,
0.474853515625,
0.17724609375,
0.474853515625
],
[
0.1748046875,
0.473388671875,
0.82275390625,
0.474853515625,
0.82275390625,
0.4892578125,
0.1748046875,
0.488037109375
],
[
0.176513671875,
0.4892578125,
0.82275390625,
0.4892578125,
0.82275390625,
0.50244140625,
0.176513671875,
0.50244140625
],
[
0.176513671875,
0.50244140625,
0.82275390625,
0.50244140625,
0.82275390625,
0.5166015625,
0.176513671875,
0.5166015625
],
[
0.17724609375,
0.5166015625,
0.82177734375,
0.5166015625,
0.82177734375,
0.53076171875,
0.17724609375,
0.53076171875
],
[
0.1748046875,
0.52978515625,
0.583984375,
0.52978515625,
0.583984375,
0.54345703125,
0.1748046875,
0.54345703125
],
[
0.1732177734375,
0.56005859375,
0.2646484375,
0.56298828125,
0.2646484375,
0.5771484375,
0.1732177734375,
0.57373046875
],
[
0.17724609375,
0.5869140625,
0.82275390625,
0.5869140625,
0.82275390625,
0.60107421875,
0.17724609375,
0.60107421875
],
[
0.176513671875,
0.599609375,
0.82275390625,
0.60107421875,
0.82275390625,
0.61572265625,
0.176513671875,
0.6142578125
],
[
0.176513671875,
0.61572265625,
0.82275390625,
0.61572265625,
0.82275390625,
0.62890625,
0.176513671875,
0.62890625
],
[
0.17724609375,
0.62890625,
0.82177734375,
0.62890625,
0.82177734375,
0.642578125,
0.17724609375,
0.642578125
],
[
0.176513671875,
0.642578125,
0.82177734375,
0.642578125,
0.82177734375,
0.65576171875,
0.176513671875,
0.65576171875
],
[
0.176513671875,
0.65478515625,
0.82177734375,
0.65576171875,
0.82177734375,
0.66943359375,
0.176513671875,
0.66845703125
],
[
0.176513671875,
0.67041015625,
0.82177734375,
0.67041015625,
0.82177734375,
0.6845703125,
0.176513671875,
0.6845703125
],
[
0.176513671875,
0.6845703125,
0.82275390625,
0.6845703125,
0.82275390625,
0.69775390625,
0.176513671875,
0.69775390625
],
[
0.1748046875,
0.6962890625,
0.82275390625,
0.69775390625,
0.82275390625,
0.7119140625,
0.1748046875,
0.7109375
],
[
0.1732177734375,
0.71044921875,
0.7236328125,
0.71044921875,
0.7236328125,
0.7236328125,
0.1732177734375,
0.7236328125
],
[
0.294189453125,
0.74072265625,
0.69775390625,
0.7412109375,
0.69775390625,
0.7548828125,
0.294189453125,
0.75390625
],
[
0.2099609375,
0.76025390625,
0.78515625,
0.76123046875,
0.78515625,
0.77880859375,
0.2099609375,
0.77783203125
],
[
0.2156982421875,
0.78271484375,
0.291748046875,
0.78271484375,
0.291748046875,
0.79443359375,
0.2156982421875,
0.79443359375
],
[
0.3203125,
0.78271484375,
0.3544921875,
0.78271484375,
0.3544921875,
0.79443359375,
0.3203125,
0.79443359375
],
[
0.378173828125,
0.78271484375,
0.419189453125,
0.78271484375,
0.419189453125,
0.79443359375,
0.378173828125,
0.79443359375
],
[
0.439453125,
0.78271484375,
0.476318359375,
0.78271484375,
0.476318359375,
0.79443359375,
0.439453125,
0.79443359375
],
[
0.51953125,
0.78271484375,
0.55810546875,
0.78271484375,
0.55810546875,
0.79443359375,
0.51953125,
0.79443359375
],
[
0.576171875,
0.7822265625,
0.66015625,
0.7822265625,
0.66015625,
0.79248046875,
0.576171875,
0.79248046875
],
[
0.654296875,
0.78271484375,
0.70751953125,
0.78271484375,
0.70751953125,
0.79443359375,
0.654296875,
0.79443359375
],
[
0.7294921875,
0.78271484375,
0.75732421875,
0.78271484375,
0.75732421875,
0.79443359375,
0.7294921875,
0.79443359375
],
[
0.420654296875,
0.78515625,
0.441162109375,
0.78515625,
0.441162109375,
0.79345703125,
0.420654296875,
0.79345703125
],
[
0.3505859375,
0.78515625,
0.375732421875,
0.78515625,
0.375732421875,
0.79345703125,
0.3505859375,
0.79345703125
],
[
0.55615234375,
0.78515625,
0.57177734375,
0.78515625,
0.57177734375,
0.79248046875,
0.55615234375,
0.79248046875
],
[
0.1748046875,
0.814453125,
0.82177734375,
0.81640625,
0.82177734375,
0.830078125,
0.1748046875,
0.82763671875
],
[
0.30224609375,
0.7861328125,
0.3095703125,
0.7861328125,
0.3095703125,
0.79150390625,
0.30224609375,
0.79150390625
],
[
0.17724609375,
0.83056640625,
0.82275390625,
0.83056640625,
0.82275390625,
0.84130859375,
0.17724609375,
0.84130859375
],
[
0.176513671875,
0.84326171875,
0.82275390625,
0.84326171875,
0.82275390625,
0.85693359375,
0.176513671875,
0.85693359375
],
[
0.17724609375,
0.85693359375,
0.82275390625,
0.85693359375,
0.82275390625,
0.87060546875,
0.17724609375,
0.87060546875
],
[
0.1748046875,
0.869140625,
0.82275390625,
0.86865234375,
0.82275390625,
0.88330078125,
0.1748046875,
0.8837890625
],
[
0.176513671875,
0.8837890625,
0.82275390625,
0.8837890625,
0.82275390625,
0.89794921875,
0.176513671875,
0.89794921875
],
[
0.176513671875,
0.8984375,
0.82275390625,
0.8984375,
0.82275390625,
0.912109375,
0.176513671875,
0.912109375
],
[
0.494384765625,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94775390625,
0.494384765625,
0.94775390625
]
] |
[
"",
"LT",
"52505007501000",
"125015001750",
"200",
"80",
"160.",
"24",
"",
"",
"(b)jungle",
"()traffic",
"Figure 3: Learning curves in (a) prisoner, (b) jungle, and (c) traffic . All the curves are plotted",
"(a)prisoner",
"using 5 training runs with different random seeds, where the solid line is the mean and the shadowed",
"area is enclosedbytheminand max value",
"回报，但是如果两个代理都采用这种方式，则会发生碰撞，并且只有一个代理以相等的概率",
"赢得目标。相反，如果两个代理商都“合作”",
"",
"则可以获得更高的回报，尽管这需要更多的",
"步骤。可能的最高回报是1图3a从平均收益的角度说明了所有方法的学习曲线。请注意",
"对于所有三种情况，我们用实线表示不同随机种子的5次训练运行的平均值，用阴影区域表",
"示最小/最大值。作为自利优化的结果，DQN收敛到“缺陷/缺陷”纳什均衡，其中每个代理",
"都收到关于0.5的预期奖励。DGN也是如此，因为它的目的只是利用邻居的观察结果，而",
"囚犯已经是一个完全可观察的环境。给定一个手动调整的奖励塑造因素来指导代理以最大",
"化平均回报，NeurComm和固定LToS代理最终能够合作。ConseNet和OMIX也是如此。然",
"而，它们收敛缓慢。相比之下，IP代理的学习速度较慢，其性能仅比0.5高一点。LIO特工",
"在开始时很快就进行了合作，但他们无法形成稳定的合作，也许需要更长时间的培训才能",
"摆脱这种不稳定JointDQN、Coco-Q（Sodomkaetal.,2013）和LToS的性能类似，优于其他方",
"法。JointDQN是一个集中式DQN，它控制代理A和B的联合行动，因此应该能够实现最",
"佳性能，但即使在这样一个简单的双代理场景中，仍然需要时间来收敛。作为一种改进的表",
"格Q学习方法，Coco-Q引人coco值（KalaiandKalai20i0）作为Bellman方程中期望收益的",
"替代，并将差值视为转移奖励。但是，它是专门为某些游戏设计的，并且很难扩展到双人游",
"戏之外。LToS可以学习奖励共享方案，其中一个代理首先将所有奖励给予另一个代理，以",
"便防止它们都出现“缺陷”，从而快速实现最佳平均回报，如实验中所观察到的。通过囚犯",
"我们验证了LToS可以通过学习分享奖励来逃避局部最优。",
"4.2Jungle",
"丛林是Jiangetal.（2020）基于MAgent（Zhengetal.,2017）提出的关于道德困境的场景。如",
"图2b所示，有N剂和L固定食品。在每个时间步，每个代理可以攻击或移动到一个相邻",
"的网格。吃（攻击食物）带来正奖励+1，但攻击其他代理获得更高的奖励+2。然而，受",
"害者遭受负奖励－4，这使得代理之间的每次攻击都是负和动作。此外，攻击空白网格会得",
"到一个小的负奖励-0.01（抑制过度攻击）。我们遵循Jiangetal.（2020）的原始设置：地图",
"大小=30×30网格，N=20，L=12，观察值由一个人的坐标和附近的11×11网格的字",
"段组成。每个代理都有3个最接近的代理作为其邻居。与囚犯相比，丛林有更多的特工，因",
"此JointDQN和Coco-Q在这种情况下被忽略。丛林的另一个挑战是网络拓扑是动态的，因",
"为每个代理总是可以移动。幸运的是，拓扑变化缓慢且可预测，算法可能会将时变邻居集",
"人，作为输人的一部分。因此，仍然有可能很好地估计成形的奖励和价值函数。",
"Tablel:Averagereward perstepofall themethods in jungle",
"DQN DGN fixedLToS LToS NeurComm ConseNet LIOIPQMIXupper bound",
"0.240.66",
"0.71",
"0.86",
"-0.05",
"-0.04",
"0.000.63",
"0.041",
"0.95",
"",
"",
"",
"图3b说明了所有方法的学习曲线，它们在收敛后的性能也总结在表1中。NeurComm",
"",
"ConseNet和QMIX在此任务中表现不佳。在NeurComm中，每个代理都会获得“延迟的全局",
"信息”。但是，当通过动态拓扑进行通信时，无法形成延迟的全局信息的稳定模式。ConseNet",
"是在充分观察的前提下构建的，当输人不仅是部分的，而且在顺序和内容上变化很大时，它",
"很难学习。LIO代理也表现不佳，因为它们在动态拓扑中无法相互区分，因此无法学习适当",
"的激励函数。此外，LIO需要对手建模，但很难在动态环境中同时对所有其他代理进行建",
"模。QMIX没有这些问题。在针对全局优化时，如LIO，它意识到攻击通常意味着负和动作",
"5"
] |
{
"arxiv_id": "2112.08702",
"language": "zh",
"page_idx": 4
}
|
||
2112.08772-zh-2
|
2112.08772-zh-2.jpg
|
$\delta$ -SAM)。具体来说，我们首先通过在模型权重上用随机高斯噪声扰动损失来估计每个实例的 Hessian 和梯度范数。接下来，$\delta$ -SAM 动态重新加权每批训练实例内的损失，然后计算一个共享权重扰动，使重新加权的批次损失最大化。最后，我们在原始的(未加权)批次上更新扰动模型。与 SAM 相比，$\delta$ -SAM 在估计梯度变化率时只需要额外的计算成本，这可以使用三个额外的前向通过来有效地执行
我们评估$\delta$ -SAM 对微调预训练的语言模型 (PLMs)。在标准 GLUE 基准 (Wang et al., 2018),自监督语义文本相似性(STS)和抽象摘要任务上的实验表明，除了显着优于基本模型外，$\delta$ -SAM 还始终优于 SAM,平均仅需$18\%$额外的计算成本
本文的主要贡献有三个方面。首先，我们分析了每个实例权重扰动的训练目标，并发现在某些假设下，它可以通过加权批量扰动来近似，其中实例根据其引起的梯度变化率有效地重新加权。其次，我们建议使用随机扰动作为估计，以有效地实现加权方案。第三，我们在一组不同的数据集上评估$\delta$ -SAM,并在所有方面找到一致的改进
### 2 $\textbf{Related Work}$
Model Generalization. 深度神经网络经常被过度参数化，并且可能会遭受较差的泛化E (Zhang et al., 2021)。为了提高神经模型的泛化性，人们付出了很多努力，导致了包括数据增强$\boxed{\mathrm{F}}($Sennrich et al., 2016; Wei and Zou, 2019; Sun et al., 2020; Kumar et al., 2020; Thakur et al., 2021), 正则化$\overline{\mathrm{E}}($Loshchilov and Hutter, 2018; Xuhong et al., 2018; Liang et al., 2021) 和改进的优化过程$\overline{\mathrm{E}}($Izmailov et al., 2018; Mobahi et al., 2020; Heo et al., 2021)在内的方法。这些方法考虑了泛化的不同方面，可以结合起来以实现更好的性能。其中，对抗性训练(Goodfellow et al., 2014) 已经证明了其在不需要任何额外数据或外部知识的情况下提高模型泛化的有效性，并已被广泛尝试增强 NLP 模型 (Zhu et al., 2019; Jiang et al., 2020a; Pereira et al., 2021; Li and Qiu, 2021)。对抗性训练的工作原理是对输入嵌入进行对抗性干扰，要么将对抗性风险降至最低，要么将风险变化调整为较小。具体来说，FreeLB$\boxed{\mathrm{F}}($Zhu et al., 2019) 使用投影梯度下降$\overline{\mathbb{E}}\left(\mathrm{PGD;Madry~et~al.2018~)~在输入嵌入}\right)$ 上产生对抗性扰动，并在相邻步骤$\overline{\mathrm{E}}$(Shafahi et al., 2019) 中更新模型参数时回收计算出的梯度，以降低计算成本。TAT (Pereira et al.,2021) 通过在扰动计算中优先考虑最常预测错误的类来改进 FreeLB。TAVAT (Li and Qiu, 2021) 使
<table>
<tbody>
<tr>
<td>$NI$ T 112 $a_{2}=\frac{1}{2}$</td>
<td>rithm 1: SAM and $\delta$ -SAM</td>
</tr>
<tr>
<td> </td>
<td>Rewe1gh $\mathcal{B}$by $Eq.$ $.D$ Compute gradient $\nabla l_{\mathcal{B}}(\boldsymbol{w})$ of the $^{\prime}$reweiohted) hat $ab$ $^{\prime}G$ cal risk</td>
</tr>
<tr>
<td> </td>
<td>(reweıghted) batch's empırıcal rısk. Perturb the model weights by $=\rho\nabla l_{\mathcal{R}}(\boldsymbol{w})/\|\nabla l_{\mathcal{R}}(\boldsymbol{w})\|$</td>
</tr>
<tr>
<td> </td>
<td>Update $w$ w.r.t. the unweighted $\frac{1}{\mathrm{v}}\sum_{j=1}^{N}l_{j}(\boldsymbol{w}+\epsilon^{*})$ empirical risk $<\frac{1}{N}$ $1/\Lambda$ with the optimizer $h$ .</td>
</tr>
</tbody>
</table>
用令牌级累积扰动词汇表来指导 PGD 中的初始化。然而，这些工作只考虑输入特征表示的鲁棒性，而我们考虑所有模型权重的鲁棒性
Sharpness-Aware Minimization. Foret et al. (2020) 利用平坦最小值和更好的模型泛化之间的相关性，并提出 SAM 用于训练对对抗权重扰动具有鲁棒性的深度神经模型。它在视觉 (Chen et al.,2022; Zheng et al., 2021)和语言 (Bahri et al., 2022) 模式的任务中都证明了有效性。已经提出了 SAM 的几种变体来提高其效率或有效性。为了提高效率，Brock et al. (2021)建议通过在批处理中扰动较少的实例来加快 SAM; Du et al.(2022)引人了随机权重扰动和对清晰度敏感的数据选择，以减少计算开销。为了提高效率；Kwon et al.(2021)建议自适应地设置 SAM 的半径，以使其对参数的尺度不变；Zhuang et al. (2022) 在扰动模型的正交方向上引入了梯度上升优化步骤，以实现更好的平整度。与上述研究相反，本文首次探讨了如何缩小每例和每批重量扰动的差距。因此， 我们提出了一种对每个实例权重扰动的有效近似，它显示了几个 NLP 任务的改进结果，同时不会带来太多的计算开销
# 3 Sharpness-Aware $\textbf{Minimization}\left ( \mathbf{SAM}\right )$
在本节中，我们简要回顾了 SAM 的原理并讨论了其局限性。文献已经观察到平坦极小值和更好的模型泛化之间的直接相关性，无论是经验上还是理论上 (Keskar et al., 2016; Dziugaite and Roy, 2017; Li et al., 2018; Jiang et al., 2020b)
|
[
[
0.1175537109375,
0.0872802734375,
0.484375,
0.0872802734375,
0.484375,
0.1009521484375,
0.1175537109375,
0.1009521484375
],
[
0.5205078125,
0.08905029296875,
0.76904296875,
0.08905029296875,
0.76904296875,
0.10272216796875,
0.5205078125,
0.10272216796875
],
[
0.12005615234375,
0.10272216796875,
0.484375,
0.10272216796875,
0.484375,
0.11639404296875,
0.12005615234375,
0.11639404296875
],
[
0.52880859375,
0.10687255859375,
0.75146484375,
0.10687255859375,
0.75146484375,
0.12054443359375,
0.52880859375,
0.12054443359375
],
[
0.1175537109375,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.53662109375,
0.1217041015625,
0.78515625,
0.122314453125,
0.78515625,
0.141357421875,
0.53662109375,
0.14013671875
],
[
0.118408203125,
0.1341552734375,
0.465087890625,
0.1341552734375,
0.465087890625,
0.1478271484375,
0.118408203125,
0.1478271484375
],
[
0.5400390625,
0.14013671875,
0.8271484375,
0.14013671875,
0.8271484375,
0.15380859375,
0.5400390625,
0.15380859375
],
[
0.12005615234375,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.1650390625,
0.12005615234375,
0.1650390625
],
[
0.5380859375,
0.1561279296875,
0.8447265625,
0.1561279296875,
0.8447265625,
0.169189453125,
0.5380859375,
0.169189453125
],
[
0.1175537109375,
0.166259765625,
0.483642578125,
0.166259765625,
0.483642578125,
0.179931640625,
0.1175537109375,
0.179931640625
],
[
0.52880859375,
0.171630859375,
0.72119140625,
0.171630859375,
0.72119140625,
0.185302734375,
0.52880859375,
0.185302734375
],
[
0.118408203125,
0.182861328125,
0.483642578125,
0.182861328125,
0.483642578125,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.52880859375,
0.189453125,
0.73779296875,
0.189453125,
0.73779296875,
0.203125,
0.52880859375,
0.203125
],
[
0.118408203125,
0.1995849609375,
0.483642578125,
0.1995849609375,
0.483642578125,
0.213134765625,
0.118408203125,
0.213134765625
],
[
0.53076171875,
0.204833984375,
0.69775390625,
0.204833984375,
0.69775390625,
0.218505859375,
0.53076171875,
0.218505859375
],
[
0.12005615234375,
0.215576171875,
0.393798828125,
0.215576171875,
0.393798828125,
0.2286376953125,
0.12005615234375,
0.2286376953125
],
[
0.5576171875,
0.2203369140625,
0.81787109375,
0.2203369140625,
0.81787109375,
0.235107421875,
0.5576171875,
0.235107421875
],
[
0.1368408203125,
0.232177734375,
0.484375,
0.232177734375,
0.484375,
0.2452392578125,
0.1368408203125,
0.2452392578125
],
[
0.5576171875,
0.2369384765625,
0.65576171875,
0.235107421875,
0.65576171875,
0.248779296875,
0.5576171875,
0.25048828125
],
[
0.1175537109375,
0.2464599609375,
0.484375,
0.2476806640625,
0.484375,
0.262451171875,
0.1175537109375,
0.26123046875
],
[
0.5576171875,
0.2529296875,
0.71728515625,
0.25048828125,
0.71728515625,
0.265380859375,
0.5576171875,
0.267333984375
],
[
0.1175537109375,
0.26318359375,
0.484375,
0.26318359375,
0.484375,
0.276611328125,
0.1175537109375,
0.276611328125
],
[
0.55908203125,
0.26953125,
0.80419921875,
0.26953125,
0.80419921875,
0.283203125,
0.55908203125,
0.283203125
],
[
0.118408203125,
0.28076171875,
0.484375,
0.28076171875,
0.484375,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.56689453125,
0.2861328125,
0.8271484375,
0.2861328125,
0.8271484375,
0.2998046875,
0.56689453125,
0.2998046875
],
[
0.118408203125,
0.295654296875,
0.486083984375,
0.295654296875,
0.486083984375,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.5576171875,
0.3017578125,
0.77490234375,
0.3017578125,
0.77490234375,
0.3154296875,
0.5576171875,
0.3154296875
],
[
0.1175537109375,
0.312255859375,
0.28369140625,
0.312255859375,
0.28369140625,
0.325927734375,
0.1175537109375,
0.325927734375
],
[
0.56689453125,
0.317626953125,
0.78857421875,
0.317626953125,
0.78857421875,
0.330810546875,
0.56689453125,
0.330810546875
],
[
0.1368408203125,
0.328857421875,
0.484375,
0.328857421875,
0.484375,
0.342529296875,
0.1368408203125,
0.342529296875
],
[
0.5576171875,
0.333251953125,
0.791015625,
0.333251953125,
0.791015625,
0.3466796875,
0.5576171875,
0.3466796875
],
[
0.118408203125,
0.345703125,
0.48193359375,
0.345703125,
0.48193359375,
0.355712890625,
0.118408203125,
0.355712890625
],
[
0.56689453125,
0.349853515625,
0.8251953125,
0.349853515625,
0.8251953125,
0.364013671875,
0.56689453125,
0.364013671875
],
[
0.1175537109375,
0.35986328125,
0.483642578125,
0.35986328125,
0.483642578125,
0.37353515625,
0.1175537109375,
0.37353515625
],
[
0.568359375,
0.366455078125,
0.724609375,
0.366455078125,
0.724609375,
0.380126953125,
0.568359375,
0.380126953125
],
[
0.1175537109375,
0.377197265625,
0.483642578125,
0.377197265625,
0.483642578125,
0.39013671875,
0.1175537109375,
0.39013671875
],
[
0.118408203125,
0.394287109375,
0.48193359375,
0.394287109375,
0.48193359375,
0.405029296875,
0.118408203125,
0.405029296875
],
[
0.1175537109375,
0.410400390625,
0.486083984375,
0.410400390625,
0.486083984375,
0.42333984375,
0.1175537109375,
0.42333984375
],
[
0.51318359375,
0.412109375,
0.87744140625,
0.412109375,
0.87744140625,
0.42578125,
0.51318359375,
0.42578125
],
[
0.124267578125,
0.42578125,
0.48779296875,
0.42578125,
0.48779296875,
0.439453125,
0.124267578125,
0.439453125
],
[
0.51318359375,
0.4287109375,
0.87890625,
0.4287109375,
0.87890625,
0.4423828125,
0.51318359375,
0.4423828125
],
[
0.118408203125,
0.4423828125,
0.285400390625,
0.4423828125,
0.285400390625,
0.4560546875,
0.118408203125,
0.4560546875
],
[
0.5146484375,
0.44482421875,
0.8623046875,
0.44482421875,
0.8623046875,
0.457763671875,
0.5146484375,
0.457763671875
],
[
0.51318359375,
0.46435546875,
0.880859375,
0.46435546875,
0.880859375,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.1158447265625,
0.471435546875,
0.151123046875,
0.471435546875,
0.151123046875,
0.48388671875,
0.1158447265625,
0.48388671875
],
[
0.151123046875,
0.471435546875,
0.274658203125,
0.471435546875,
0.274658203125,
0.485107421875,
0.151123046875,
0.485107421875
],
[
0.51318359375,
0.48095703125,
0.880859375,
0.48095703125,
0.880859375,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.51318359375,
0.496337890625,
0.880859375,
0.496337890625,
0.880859375,
0.51123046875,
0.51318359375,
0.51123046875
],
[
0.1175537109375,
0.5029296875,
0.486083984375,
0.5029296875,
0.486083984375,
0.5166015625,
0.1175537109375,
0.5166015625
],
[
0.118408203125,
0.5185546875,
0.48779296875,
0.5185546875,
0.48779296875,
0.5322265625,
0.118408203125,
0.5322265625
],
[
0.5146484375,
0.51318359375,
0.880859375,
0.51318359375,
0.880859375,
0.52685546875,
0.5146484375,
0.52685546875
],
[
0.51123046875,
0.529296875,
0.880859375,
0.529296875,
0.880859375,
0.5419921875,
0.51123046875,
0.5419921875
],
[
0.118408203125,
0.53515625,
0.484375,
0.53515625,
0.484375,
0.548828125,
0.118408203125,
0.548828125
],
[
0.5146484375,
0.54443359375,
0.87744140625,
0.54443359375,
0.87744140625,
0.55810546875,
0.5146484375,
0.55810546875
],
[
0.12005615234375,
0.5517578125,
0.48779296875,
0.5517578125,
0.48779296875,
0.5654296875,
0.12005615234375,
0.5654296875
],
[
0.51318359375,
0.56103515625,
0.880859375,
0.56103515625,
0.880859375,
0.57470703125,
0.51318359375,
0.57470703125
],
[
0.12005615234375,
0.56787109375,
0.48779296875,
0.56787109375,
0.48779296875,
0.58056640625,
0.12005615234375,
0.58056640625
],
[
0.51123046875,
0.576171875,
0.880859375,
0.57666015625,
0.880859375,
0.59033203125,
0.51123046875,
0.5888671875
],
[
0.118408203125,
0.5830078125,
0.484375,
0.5830078125,
0.484375,
0.5966796875,
0.118408203125,
0.5966796875
],
[
0.1175537109375,
0.59716796875,
0.484375,
0.5986328125,
0.484375,
0.6123046875,
0.1175537109375,
0.61083984375
],
[
0.51123046875,
0.59326171875,
0.880859375,
0.59326171875,
0.880859375,
0.60693359375,
0.51123046875,
0.60693359375
],
[
0.51123046875,
0.60791015625,
0.8818359375,
0.609375,
0.8818359375,
0.62353515625,
0.51123046875,
0.62255859375
],
[
0.118408203125,
0.615234375,
0.484375,
0.615234375,
0.484375,
0.62890625,
0.118408203125,
0.62890625
],
[
0.1175537109375,
0.6298828125,
0.486083984375,
0.6298828125,
0.486083984375,
0.6435546875,
0.1175537109375,
0.6435546875
],
[
0.51318359375,
0.6259765625,
0.880859375,
0.6259765625,
0.880859375,
0.6396484375,
0.51318359375,
0.6396484375
],
[
0.1175537109375,
0.64794921875,
0.48779296875,
0.64794921875,
0.48779296875,
0.6611328125,
0.1175537109375,
0.6611328125
],
[
0.51318359375,
0.642578125,
0.87890625,
0.642578125,
0.87890625,
0.65625,
0.51318359375,
0.65625
],
[
0.51318359375,
0.65771484375,
0.87890625,
0.65771484375,
0.87890625,
0.67138671875,
0.51318359375,
0.67138671875
],
[
0.118408203125,
0.66455078125,
0.48779296875,
0.66455078125,
0.48779296875,
0.677734375,
0.118408203125,
0.677734375
],
[
0.118408203125,
0.6796875,
0.484375,
0.6796875,
0.484375,
0.693359375,
0.118408203125,
0.693359375
],
[
0.5146484375,
0.6748046875,
0.880859375,
0.6748046875,
0.880859375,
0.6884765625,
0.5146484375,
0.6884765625
],
[
0.1175537109375,
0.69482421875,
0.483642578125,
0.69482421875,
0.483642578125,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.51123046875,
0.689453125,
0.87744140625,
0.689453125,
0.87744140625,
0.703125,
0.51123046875,
0.703125
],
[
0.118408203125,
0.7119140625,
0.486083984375,
0.7119140625,
0.486083984375,
0.7255859375,
0.118408203125,
0.7255859375
],
[
0.51318359375,
0.70703125,
0.875,
0.70703125,
0.875,
0.71728515625,
0.51318359375,
0.71728515625
],
[
0.118408203125,
0.72802734375,
0.484375,
0.72802734375,
0.484375,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.51318359375,
0.72265625,
0.8818359375,
0.72265625,
0.8818359375,
0.736328125,
0.51318359375,
0.736328125
],
[
0.51318359375,
0.73828125,
0.880859375,
0.73828125,
0.880859375,
0.751953125,
0.51318359375,
0.751953125
],
[
0.1175537109375,
0.74462890625,
0.486083984375,
0.74462890625,
0.486083984375,
0.75830078125,
0.1175537109375,
0.75830078125
],
[
0.51123046875,
0.7529296875,
0.87890625,
0.7529296875,
0.87890625,
0.7666015625,
0.51123046875,
0.7666015625
],
[
0.1175537109375,
0.76025390625,
0.484375,
0.76025390625,
0.484375,
0.77392578125,
0.1175537109375,
0.77392578125
],
[
0.51123046875,
0.76953125,
0.7138671875,
0.76953125,
0.7138671875,
0.783203125,
0.51123046875,
0.783203125
],
[
0.118408203125,
0.77685546875,
0.484375,
0.77685546875,
0.484375,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.118408203125,
0.7919921875,
0.484375,
0.7919921875,
0.484375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.509765625,
0.798828125,
0.5498046875,
0.798828125,
0.5498046875,
0.81005859375,
0.509765625,
0.81005859375
],
[
0.54248046875,
0.798828125,
0.69873046875,
0.798828125,
0.69873046875,
0.8134765625,
0.54248046875,
0.8134765625
],
[
0.1175537109375,
0.80712890625,
0.484375,
0.80810546875,
0.484375,
0.822265625,
0.1175537109375,
0.82177734375
],
[
0.5400390625,
0.8134765625,
0.72119140625,
0.81396484375,
0.72119140625,
0.8291015625,
0.5400390625,
0.82763671875
],
[
0.12005615234375,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.12005615234375,
0.83837890625
],
[
0.12005615234375,
0.84130859375,
0.486083984375,
0.84130859375,
0.486083984375,
0.85498046875,
0.12005615234375,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.486083984375,
0.85693359375,
0.486083984375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.515625,
0.87451171875,
0.87890625,
0.87451171875,
0.87890625,
0.88525390625,
0.515625,
0.88525390625
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.51123046875,
0.90625,
0.880859375,
0.90625,
0.880859375,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"8-SAM)。具体来说，我们首先通过在模型权",
"Algorithm 1: SAM and 8 -SAM",
"重上用随机高斯噪声扰动损失来估计每个实例",
"Input: model fw , training set",
"的Hessian和梯度范数。接下来，8-SAM动态",
"S(aci, yi))-, loss function",
"重新加权每批训练实例内的损失，然后计算",
"I : Wx X x y-→ R+, batch size N,",
"个共享权重扰动，使重新加权的批次损失最大",
"neighborhood size p E R+, optimizer h",
"化。最后，我们在原始的（未加权）批次上更新",
"Output: a flat solution w",
"扰动模型。与SAM相比，-SAM在估计梯度",
"Initialize model weights w .",
"变化率时只需要额外的计算成本，这可以使用",
"while not converge do",
"三个额外的前向通过来有效地执行",
"Sample a batch B = [(αcj, yi)-]",
"我们评估8-SAM对微调预训练的语言模",
"Foro-SAM:",
"型(PLMs)。在标准GLUE基准（Wang et al.,",
"Reweigh B by Eq.5",
"2018）：自监督语义文本相似性（STS）和抽",
"Compute gradient Vl(w) of the",
"象摘要任务上的实验表明，除了显着优于基本",
"(reweighted) batch's empirical risk.",
"模型外，S-SAM还始终优于SAM，平均仅需",
"Perturb the model weights by",
"18%额外的计算成本",
"e* = pVl(w)/ l/Vlr(w)ll2 ",
"本文的主要贡献有三个方面。首先，我们分",
"Update w w.r.t. the unweighted",
"析了每个实例权重扰动的训练目标，并发现",
"empirical risk i=il,(w + e*)",
"在某些假设下，它可以通过加权批量扰动来近",
"with the optimizer h",
"似，其中实例根据其引起的梯度变化率有效地",
"重新加权。其次，我们建议使用随机扰动作头",
"估计，以有效地实现加权方案。第三，我们在",
"用令牌级累积扰动词汇表来指导PGD中的初",
"一组不同的数据集上评估-SAM，并在所有",
"始化。然而，这些工作只考虑输入特征表示的",
"方面找到一致的改进",
"鲁棒性，而我们考虑所有模型权重的鲁棒性",
"Sharpness-Aware Minimization. Foret et al.",
"2",
"Related Work",
"(2020)利用平坦最小值和更好的模型泛化之",
"间的相关性，并提出SAM用于训练对对抗权",
"ModelGeneralization.深度神经网络经常被",
"过度参数化，并且可能会遭受较差的泛化日",
"重扰动具有鲁棒性的深度神经模型。它在视",
"觉(Chen etal.,2022;Zhenget al.,2021)和语",
"(Zhangetal.，2021）。为了提高神经模型的泛",
"言（Bahrietal.，2022）模式的任务中都证明了",
"化性，人们付出了很多努力，导致了包括数据",
"有效性。已经提出了SAM的几种变体来提高",
"增强E(Sennrich et al.,2016;Wei and Zou, 2019;",
"其效率或有效性。为了提高效率，Brocketal.",
"Sun et al., 2020; Kumar et al., 2020; Thakur et al..",
"2021），正则化E(LoshchilovandHutter,2018",
"(2021)建议通过在批处理中扰动较少的实例来",
"加快SAM；Duetal.(2022)引入了随机权重扰",
"Xuhong et al.,2018;Liang et al.,2021)和改进的",
"优化过程E(Izmailov et al., 2018;Mobahi et al.,",
"动和对清晰度敏感的数据选择，以减少计算开",
"2020;He0etal.,2021）在内的方法。。这些方法",
"销。为了提高效率；Kwonetal.(2021）建议自",
"适应地设置SAM的半径，以使其对参数的尺",
"考虑了泛化的不同方面，可以结合起来以实现",
"更好的性能。其中，对抗性训练（Goodfellow",
"度不变；Zhuangetal.(2022）在扰动模型的正交",
"etal.,2014)已经证明了其在不需要任何额外",
"方向上引入了梯度上升优化步骤，以实现更好",
"数据或外部知识的情况下提高模型泛化的有效",
"的平整度。与上述研究相反，本文首次探讨了",
"性，并已被广泛尝试增强NLP模型（Zhuetal.",
"如何缩小每例和每批重量扰动的差距。因此，",
"我们提出了一种对每个实例权重扰动的有效近",
"2019; Jiang et al., 2020a; Pereira et al., 2021; Li",
"似，它显示了几个NLP任务的改进结果，同时",
"andQiu,2021）。对抗性训练的工作原理是对",
"不会带来太多的计算开销",
"输入嵌入进行对抗性干扰，要么将对抗性风险",
"降至最低，要么将风险变化调整为较小。具体",
"3",
" Sharpness-Aware",
"来说，FreeLB(Zhuetal.2019）使用投影梯度",
"Minimization (SAM)",
"下降E（PGD；Madryetal.2018）在输入嵌入",
"上产生对抗性扰动，并在相邻步骤E(Shafahi",
"在本节中，我们简要回顾了SAM的原理并讨",
"etal.,2019)中更新模型参数时回收计算出的梯",
"论了其局限性。文献已经观察到平坦极小值和",
"度，以降低计算成本。TAT(Pereiraetal.,2021)",
"更好的模型泛化之间的直接相关性，无论是经",
"通过在扰动计算中优先考虑最常预测错误的类",
"验上还是理论上（Keskaretal.，2016;Dziugaite",
"来改进FreeLB。TAVAT(Liand Qiu,2021）使",
"and Roy, 2017; Li et al., 2018; Jiang et al., 2020b)"
] |
{
"arxiv_id": "2112.08772",
"language": "zh",
"page_idx": 2
}
|
||
2112.08772-zh-3
|
2112.08772-zh-3.jpg
|
。为了找到一个平坦的损失景观，SAME(Foret 重扰动会削弱对抗性训练，因此我们提出了对et al., 2020) 对模型权重进行了反向扰动，并在 SAM $\delta$ -SAME (具有动态重新加权的 SAM) 一批大小为$N$的批次上优化了以下最小-最大 的简单而有效的修改，该修改可以近似每个实目标：
$$\min_{\boldsymbol{w}}\max_{\boldsymbol{\epsilon}:\left\|\boldsymbol{\epsilon}\right\|_2\leq\rho}\frac{1}{N}\sum_{i=1}^{N}l_i(\boldsymbol{w}+\boldsymbol{\epsilon}),$$
(1)
,其中给定模型权重$w$,内部最大化寻求最小化$\epsilon$的扰动，并使$L_2$ -4 经验范数。此训练目标旨在找到其邻域具有均匀低训练损失的模型参数。由于找到$\epsilon$的精确解是 NP 困难的，因此 SAM 通过对批次的经验风险进行单步梯度下降来估计内部最大化的解$\epsilon^*:$
$$\begin{aligned}l(\boldsymbol{w})&=\frac{1}{N}\sum_{i=1}^{N}l_{i}(\boldsymbol{w}),\\\boldsymbol{\epsilon}^{*}&\approx\arg\max_{\boldsymbol{\epsilon}:\|\boldsymbol{\epsilon}\|_{2}\leq\rho}l(\boldsymbol{w})+\boldsymbol{\epsilon}^{\intercal}\nabla l(\boldsymbol{w})\\&=\rho\nabla l(\boldsymbol{w})/\left\|\nabla l(\boldsymbol{w})\right\|_{2}.\end{aligned}$$
外部最小化可以通过独立的优化器来执行$\mathbb{E}($e。g。,Adam; Kingma and Ba 2015)。SAM 将训练模型的计算成本大致提高了一倍，每批需要两次向前和两次向后传递。Alg.1 中概述了 SAM 算法
除了按批次进行扰动外，还可以对单个实例
执行权重扰动：
(2)
$$\min_{\boldsymbol{w}}\frac{1}{N}\sum_{i=1}^{N}\max_{\boldsymbol{\epsilon}_{i}:\left\|\boldsymbol{\epsilon}_{i}\right\|_{2}\leq\rho}l_{i}(\boldsymbol{w}+\boldsymbol{\epsilon}_{i}),$$
其中$\epsilon_i$是通过对单个实例进行单步梯度下降来计算的。这种方法类似于 NLP 中的许多对抗训练方法，例如 VATE(Miyato et al., 2018) 和 FreeLB$\boxed{\mathrm{F}}($Zhu et al., 2019) , 不同之处在于扰动是根据模型权重而不是仅输入嵌人来计算的。我们将 Eq.1 和 Eq.2 的目标分别称为每批重量扰动和每例重量扰动。Foret et al.(2020) 在同一篇论文中指出，每个实例的权重扰动会产生较小的测试误差，并且是模型泛化的更好预测指标。
尽管有效，但每个实例的权重扰动会大大增加计算和内存成本，需要 2$N$向前和 2$N$向后传递一批大小的$N$ 。由于每个实例的权重扰动会独立修改所有模型权重，因此需要在不同的模型副本上计算每个单独实例的扰动。因此，对于大规模训练，每个实例的权重扰动在计算上可能负担不起
# 4 SAM with Dynamic Reweighting
例的权重扰动，而无需太多额外的计算成本。我们的重新加权方法的动机是对每个实例的权重扰动进行近似的理论分析，以证明其卓越的效率。基于这种动机，我们然后说明如何在实现中实现$\delta$ -SAM
### 4.1 Theoretical Motivations
在本小节中，我们通过正式分析每个实例扰动提出的训练目标来激励我们的动态重新加权方法，并表明它可以用加权批量扰动来近似，这激发了我们的$\delta$ -SAM 算法
Preliminary.我们从 SAM 中的锐度的角度来激励我们的方法，该方法将损失景观的平坦度量化为模型权重附近区域中损失的增加。每批和每实例权重扰动的锐度定义为：
$$\mathcal{R}_{\mathrm{batch}}=\max_{\boldsymbol{\epsilon}:\left\|\boldsymbol{\epsilon}\right\|_{2}\leq\rho}\frac{1}{N}\sum_{i=1}^{N}\left(l_{i}(\boldsymbol{w}+\boldsymbol{\epsilon})-l_{i}(\boldsymbol{w})\right),\\\mathcal{R}_{\mathrm{inst}}=\frac{1}{N}\sum_{i=1}^{N}\max_{\boldsymbol{\epsilon}_{i}:\left\|\boldsymbol{\epsilon}_{i}\right\|_{2}\leq\rho}\left(l_{i}(\boldsymbol{w}+\boldsymbol{\epsilon}_{i})-l_{i}(\boldsymbol{w})\right).$$
由于非共享$\epsilon _i$ , $\mathcal{R} _\mathrm{inst}\geq \mathcal{R} _\mathrm{batch}$ ,表明 $\mathcal{R}_\mathrm{inst}$的正则化效果更强。但是，$\mathcal{R}_\mathrm{inst}$的计算成本很高，因为$N$内部最大化问题中的$\epsilon_i$必须通过$N$单个实例上的梯度下降来计算，为此需要$O(N)$向后通过网络。在下面的分析中，我们展示了如何用加权的每批权重扰动来近似更强的$\mathcal{R}_\mathrm{inst}$
我们首先考虑一般经验风险$l_i$的二阶展开，
例如$i:$
$$l_i(\boldsymbol{w}+\boldsymbol{\epsilon})=l_i(\boldsymbol{w})+\nabla l_i(\boldsymbol{w})^\intercal\boldsymbol{\epsilon}+\frac{1}{2}\boldsymbol{\epsilon}^\intercal\boldsymbol{H}_i(\boldsymbol{w})\boldsymbol{\epsilon}.$$
为了进行易于处理的理论分析，我们假设 Hessian 是低秩的正定矩阵$\boldsymbol{H}_i(w)=$ $a_i\nabla l_i(\boldsymbol{w})\nabla l_i(\boldsymbol{w})^\intercal,(a_i>0)$。然后，我们以一步梯度下降的封闭形式在二阶近似下获得$\mathcal{R}_\mathrm{batch}$ 和$\mathcal{R}_\mathrm{inst}$中的扰动，以与 SAM 的实践保持一致：
$$\epsilon=\rho\nabla l(\boldsymbol{w})/\left\|\nabla l(\boldsymbol{w})\right\|_{2},\\\epsilon_{i}=\rho\nabla l_{i}(\boldsymbol{w})/\left\|\nabla l_{i}(\boldsymbol{w})\right\|_{2},$$
其中$\nabla l(\boldsymbol w)=\frac1N\sum_{i=1}^N\nabla l_i(\boldsymbol w)$是批次的平均
梯度
Training Objective. 获得$\mathcal{R}_\mathrm{batch}$或$\mathcal{R}_\mathrm{inst}$后SAM 将计算$\frac\partial{\partial w}\mathcal{R}_\mathrm{batch}$或$\frac\partial{\partial w}\mathcal{R}_\mathrm{inst}$并更新模型
在本文中，我们寻求在权重扰动方面有更好 权重以最大程度地减少损失。为了实现更有效的对手来改进 SAM。由于 SAM 采用的每批权 的扰动，我们寻求与梯度$\mathcal{R}_\mathrm{inst}$保持一致，该
|
[
[
0.1234130859375,
0.08612060546875,
0.48779296875,
0.08612060546875,
0.48779296875,
0.09979248046875,
0.1234130859375,
0.09979248046875
],
[
0.5146484375,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.5146484375,
0.1009521484375
],
[
0.118408203125,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.8818359375,
0.10272216796875,
0.8818359375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.118408203125,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.5146484375,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.118408203125,
0.1341552734375,
0.16455078125,
0.1341552734375,
0.16455078125,
0.1484375,
0.118408203125,
0.1484375
],
[
0.515625,
0.135986328125,
0.87890625,
0.135986328125,
0.87890625,
0.146728515625,
0.515625,
0.146728515625
],
[
0.51318359375,
0.15087890625,
0.880859375,
0.15087890625,
0.880859375,
0.1644287109375,
0.51318359375,
0.1644287109375
],
[
0.310546875,
0.15673828125,
0.32666015625,
0.15673828125,
0.32666015625,
0.16748046875,
0.310546875,
0.16748046875
],
[
0.2880859375,
0.163330078125,
0.30322265625,
0.163330078125,
0.30322265625,
0.1728515625,
0.2880859375,
0.1728515625
],
[
0.301513671875,
0.16748046875,
0.30908203125,
0.16748046875,
0.30908203125,
0.17333984375,
0.301513671875,
0.17333984375
],
[
0.5146484375,
0.16748046875,
0.8759765625,
0.16748046875,
0.8759765625,
0.1776123046875,
0.5146484375,
0.1776123046875
],
[
0.462646484375,
0.171630859375,
0.489501953125,
0.171630859375,
0.489501953125,
0.182861328125,
0.462646484375,
0.182861328125
],
[
0.18896484375,
0.17041015625,
0.42236328125,
0.152587890625,
0.42822265625,
0.189453125,
0.19482421875,
0.207275390625
],
[
0.19140625,
0.1728515625,
0.268798828125,
0.1728515625,
0.268798828125,
0.18408203125,
0.19140625,
0.18408203125
],
[
0.51318359375,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.307373046875,
0.189453125,
0.330078125,
0.189453125,
0.330078125,
0.196533203125,
0.307373046875,
0.196533203125
],
[
0.51318359375,
0.198974609375,
0.6533203125,
0.198974609375,
0.6533203125,
0.2120361328125,
0.51318359375,
0.2120361328125
],
[
0.1318359375,
0.207275390625,
0.486083984375,
0.207275390625,
0.486083984375,
0.2203369140625,
0.1318359375,
0.2203369140625
],
[
0.118408203125,
0.2244873046875,
0.48193359375,
0.2244873046875,
0.48193359375,
0.235107421875,
0.118408203125,
0.235107421875
],
[
0.51123046875,
0.2244873046875,
0.75,
0.2244873046875,
0.75,
0.2381591796875,
0.51123046875,
0.2381591796875
],
[
0.1175537109375,
0.2392578125,
0.484375,
0.2392578125,
0.484375,
0.2529296875,
0.1175537109375,
0.2529296875
],
[
0.51318359375,
0.2464599609375,
0.880859375,
0.2464599609375,
0.880859375,
0.260009765625,
0.51318359375,
0.260009765625
],
[
0.12005615234375,
0.257080078125,
0.484375,
0.257080078125,
0.484375,
0.267333984375,
0.12005615234375,
0.267333984375
],
[
0.51318359375,
0.262451171875,
0.87890625,
0.262451171875,
0.87890625,
0.275634765625,
0.51318359375,
0.275634765625
],
[
0.118408203125,
0.271484375,
0.486083984375,
0.271484375,
0.486083984375,
0.28515625,
0.118408203125,
0.28515625
],
[
0.51123046875,
0.276611328125,
0.880859375,
0.27783203125,
0.880859375,
0.292236328125,
0.51123046875,
0.29150390625
],
[
0.12176513671875,
0.287353515625,
0.3720703125,
0.287353515625,
0.3720703125,
0.300537109375,
0.12176513671875,
0.300537109375
],
[
0.51123046875,
0.29345703125,
0.72900390625,
0.29345703125,
0.72900390625,
0.30712890625,
0.51123046875,
0.30712890625
],
[
0.2275390625,
0.308837890625,
0.330078125,
0.32373046875,
0.32421875,
0.3466796875,
0.2216796875,
0.33203125
],
[
0.2451171875,
0.31005859375,
0.27197265625,
0.31005859375,
0.27197265625,
0.32177734375,
0.2451171875,
0.32177734375
],
[
0.51318359375,
0.314208984375,
0.880859375,
0.314208984375,
0.880859375,
0.327880859375,
0.51318359375,
0.327880859375
],
[
0.166259765625,
0.32470703125,
0.2091064453125,
0.32470703125,
0.2091064453125,
0.336181640625,
0.166259765625,
0.336181640625
],
[
0.20654296875,
0.328857421875,
0.2166748046875,
0.328857421875,
0.2166748046875,
0.334228515625,
0.20654296875,
0.334228515625
],
[
0.2275390625,
0.327880859375,
0.279541015625,
0.33837890625,
0.273681640625,
0.35400390625,
0.2216796875,
0.34326171875
],
[
0.51318359375,
0.330810546875,
0.880859375,
0.330810546875,
0.880859375,
0.344482421875,
0.51318359375,
0.344482421875
],
[
0.5146484375,
0.3466796875,
0.880859375,
0.3466796875,
0.880859375,
0.35986328125,
0.5146484375,
0.35986328125
],
[
0.181396484375,
0.35400390625,
0.31494140625,
0.355712890625,
0.31494140625,
0.37353515625,
0.181396484375,
0.371826171875
],
[
0.307373046875,
0.355224609375,
0.435791015625,
0.35693359375,
0.435791015625,
0.37060546875,
0.307373046875,
0.3681640625
],
[
0.252685546875,
0.3681640625,
0.307373046875,
0.3681640625,
0.307373046875,
0.380126953125,
0.252685546875,
0.380126953125
],
[
0.51318359375,
0.363525390625,
0.77734375,
0.363525390625,
0.77734375,
0.377197265625,
0.51318359375,
0.377197265625
],
[
0.20654296875,
0.385986328125,
0.3779296875,
0.385986328125,
0.3779296875,
0.399658203125,
0.20654296875,
0.399658203125
],
[
0.65576171875,
0.392578125,
0.6767578125,
0.399658203125,
0.6689453125,
0.410400390625,
0.6484375,
0.402587890625
],
[
0.67529296875,
0.38720703125,
0.69189453125,
0.38720703125,
0.69189453125,
0.39794921875,
0.67529296875,
0.39794921875
],
[
0.51708984375,
0.402587890625,
0.59033203125,
0.405029296875,
0.59033203125,
0.416259765625,
0.51708984375,
0.41455078125
],
[
0.58447265625,
0.40380859375,
0.64404296875,
0.40380859375,
0.64404296875,
0.41455078125,
0.58447265625,
0.41455078125
],
[
0.6689453125,
0.3984375,
0.85986328125,
0.402099609375,
0.85986328125,
0.416259765625,
0.6689453125,
0.413330078125
],
[
0.580078125,
0.410400390625,
0.68603515625,
0.405029296875,
0.68701171875,
0.42333984375,
0.58251953125,
0.4287109375
],
[
0.640625,
0.4091796875,
0.6484375,
0.4091796875,
0.6484375,
0.412109375,
0.640625,
0.412109375
],
[
0.6689453125,
0.42041015625,
0.69287109375,
0.4150390625,
0.69873046875,
0.427490234375,
0.67529296875,
0.432861328125
],
[
0.12005615234375,
0.410888671875,
0.484375,
0.410888671875,
0.484375,
0.424560546875,
0.12005615234375,
0.424560546875
],
[
0.118408203125,
0.427001953125,
0.484375,
0.427001953125,
0.484375,
0.441162109375,
0.118408203125,
0.441162109375
],
[
0.587890625,
0.43701171875,
0.63232421875,
0.4287109375,
0.63818359375,
0.4453125,
0.59375,
0.45361328125
],
[
0.12005615234375,
0.44482421875,
0.484375,
0.44482421875,
0.484375,
0.454833984375,
0.12005615234375,
0.454833984375
],
[
0.51708984375,
0.443603515625,
0.640625,
0.438232421875,
0.64404296875,
0.4697265625,
0.5205078125,
0.474365234375
],
[
0.6572265625,
0.448974609375,
0.7109375,
0.448974609375,
0.7109375,
0.460205078125,
0.6572265625,
0.460205078125
],
[
0.7060546875,
0.44775390625,
0.8759765625,
0.44775390625,
0.8759765625,
0.46142578125,
0.7060546875,
0.46142578125
],
[
0.118408203125,
0.458984375,
0.486083984375,
0.458984375,
0.486083984375,
0.47265625,
0.118408203125,
0.47265625
],
[
0.60693359375,
0.458984375,
0.70947265625,
0.4560546875,
0.7109375,
0.471435546875,
0.60791015625,
0.474365234375
],
[
0.12176513671875,
0.4755859375,
0.2216796875,
0.4755859375,
0.2216796875,
0.48681640625,
0.12176513671875,
0.48681640625
],
[
0.1368408203125,
0.4912109375,
0.48779296875,
0.4912109375,
0.48779296875,
0.5048828125,
0.1368408203125,
0.5048828125
],
[
0.5146484375,
0.48681640625,
0.880859375,
0.48681640625,
0.880859375,
0.50048828125,
0.5146484375,
0.50048828125
],
[
0.118408203125,
0.50732421875,
0.2376708984375,
0.50732421875,
0.2376708984375,
0.52099609375,
0.118408203125,
0.52099609375
],
[
0.51318359375,
0.5029296875,
0.87890625,
0.5029296875,
0.87890625,
0.5166015625,
0.51318359375,
0.5166015625
],
[
0.51318359375,
0.5185546875,
0.880859375,
0.51708984375,
0.880859375,
0.5322265625,
0.51318359375,
0.533203125
],
[
0.2216796875,
0.52783203125,
0.264404296875,
0.54150390625,
0.255126953125,
0.55712890625,
0.210693359375,
0.54345703125
],
[
0.240966796875,
0.52978515625,
0.26025390625,
0.52978515625,
0.26025390625,
0.5390625,
0.240966796875,
0.5390625
],
[
0.17626953125,
0.53759765625,
0.345947265625,
0.552734375,
0.340087890625,
0.5830078125,
0.17041015625,
0.56787109375
],
[
0.51318359375,
0.53515625,
0.880859375,
0.53515625,
0.880859375,
0.548828125,
0.51318359375,
0.548828125
],
[
0.333251953125,
0.5419921875,
0.4248046875,
0.54443359375,
0.4248046875,
0.55810546875,
0.333251953125,
0.5556640625
],
[
0.4609375,
0.54443359375,
0.48779296875,
0.54443359375,
0.48779296875,
0.5556640625,
0.4609375,
0.5556640625
],
[
0.32568359375,
0.548828125,
0.339111328125,
0.548828125,
0.339111328125,
0.552734375,
0.32568359375,
0.552734375
],
[
0.27783203125,
0.5458984375,
0.32080078125,
0.5458984375,
0.32080078125,
0.5556640625,
0.27783203125,
0.5556640625
],
[
0.51318359375,
0.55029296875,
0.880859375,
0.55029296875,
0.880859375,
0.56396484375,
0.51318359375,
0.56396484375
],
[
0.5146484375,
0.56787109375,
0.880859375,
0.56787109375,
0.880859375,
0.58056640625,
0.5146484375,
0.58056640625
],
[
0.118408203125,
0.580078125,
0.484375,
0.580078125,
0.484375,
0.59326171875,
0.118408203125,
0.59326171875
],
[
0.5146484375,
0.58203125,
0.5693359375,
0.58447265625,
0.568359375,
0.59716796875,
0.51318359375,
0.595703125
],
[
0.12005615234375,
0.59716796875,
0.483642578125,
0.59716796875,
0.483642578125,
0.60791015625,
0.12005615234375,
0.60791015625
],
[
0.53076171875,
0.599609375,
0.8818359375,
0.599609375,
0.8818359375,
0.61328125,
0.53076171875,
0.61328125
],
[
0.118408203125,
0.6123046875,
0.486083984375,
0.6123046875,
0.486083984375,
0.6259765625,
0.118408203125,
0.6259765625
],
[
0.5146484375,
0.61767578125,
0.57421875,
0.61767578125,
0.57421875,
0.62744140625,
0.5146484375,
0.62744140625
],
[
0.118408203125,
0.62744140625,
0.484375,
0.62744140625,
0.484375,
0.64111328125,
0.118408203125,
0.64111328125
],
[
0.118408203125,
0.64453125,
0.484375,
0.64453125,
0.484375,
0.65771484375,
0.118408203125,
0.65771484375
],
[
0.51123046875,
0.642578125,
0.87890625,
0.642578125,
0.87890625,
0.67333984375,
0.51123046875,
0.67333984375
],
[
0.78076171875,
0.64453125,
0.79443359375,
0.64453125,
0.79443359375,
0.6484375,
0.78076171875,
0.6484375
],
[
0.118408203125,
0.66015625,
0.484375,
0.66015625,
0.484375,
0.67333984375,
0.118408203125,
0.67333984375
],
[
0.118408203125,
0.6767578125,
0.483642578125,
0.6767578125,
0.483642578125,
0.68701171875,
0.118408203125,
0.68701171875
],
[
0.51318359375,
0.677734375,
0.880859375,
0.6787109375,
0.880859375,
0.693359375,
0.51318359375,
0.6923828125
],
[
0.1175537109375,
0.69140625,
0.484375,
0.69140625,
0.484375,
0.705078125,
0.1175537109375,
0.705078125
],
[
0.51123046875,
0.69482421875,
0.87890625,
0.6953125,
0.87890625,
0.708984375,
0.51123046875,
0.7080078125
],
[
0.118408203125,
0.7080078125,
0.486083984375,
0.7080078125,
0.486083984375,
0.7216796875,
0.118408203125,
0.7216796875
],
[
0.51123046875,
0.71142578125,
0.87890625,
0.71142578125,
0.87890625,
0.724609375,
0.51123046875,
0.724609375
],
[
0.12005615234375,
0.7255859375,
0.20068359375,
0.7255859375,
0.20068359375,
0.73681640625,
0.12005615234375,
0.73681640625
],
[
0.51318359375,
0.7255859375,
0.880859375,
0.72802734375,
0.880859375,
0.7421875,
0.51318359375,
0.74072265625
],
[
0.1368408203125,
0.7412109375,
0.486083984375,
0.7412109375,
0.486083984375,
0.7548828125,
0.1368408203125,
0.7548828125
],
[
0.51318359375,
0.74365234375,
0.8818359375,
0.74365234375,
0.8818359375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.118408203125,
0.75732421875,
0.484375,
0.75732421875,
0.484375,
0.77099609375,
0.118408203125,
0.77099609375
],
[
0.118408203125,
0.77392578125,
0.483642578125,
0.77392578125,
0.483642578125,
0.78369140625,
0.118408203125,
0.78369140625
],
[
0.59423828125,
0.76953125,
0.79248046875,
0.77099609375,
0.79248046875,
0.7880859375,
0.59423828125,
0.78759765625
],
[
0.118408203125,
0.7890625,
0.486083984375,
0.7890625,
0.486083984375,
0.802734375,
0.118408203125,
0.802734375
],
[
0.59033203125,
0.79052734375,
0.8017578125,
0.79150390625,
0.8017578125,
0.8056640625,
0.59033203125,
0.8046875
],
[
0.12005615234375,
0.80712890625,
0.484375,
0.80712890625,
0.484375,
0.81787109375,
0.12005615234375,
0.81787109375
],
[
0.118408203125,
0.82177734375,
0.48779296875,
0.82177734375,
0.48779296875,
0.8349609375,
0.118408203125,
0.8349609375
],
[
0.51123046875,
0.818359375,
0.880859375,
0.818359375,
0.880859375,
0.8359375,
0.51123046875,
0.8359375
],
[
0.118408203125,
0.83740234375,
0.28369140625,
0.83740234375,
0.28369140625,
0.85107421875,
0.118408203125,
0.85107421875
],
[
0.5146484375,
0.83740234375,
0.5498046875,
0.83740234375,
0.5498046875,
0.8486328125,
0.5146484375,
0.8486328125
],
[
0.51123046875,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1158447265625,
0.86328125,
0.435791015625,
0.8642578125,
0.435791015625,
0.87890625,
0.1158447265625,
0.87744140625
],
[
0.51123046875,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.88720703125,
0.51123046875,
0.88720703125
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"。为了找到一个平坦的损失景观，SAME(Foret",
"重扰动会削弱对抗性训练，因此我们提出了对",
"etal.,2020)对模型权重进行了反向扰动，并在",
"SAM-SAME（具有动态重新加权的SAM)",
"一批大小为N的批次上优化了以下最小-最大",
"的简单而有效的修改，该修改可以近似每个实",
"目标：",
"例的权重扰动，而无需太多额外的计算成本。",
"我们的重新加权方法的动机是对每个实例的权",
"N",
"1",
"",
"重扰动进行近似的理论分析，以证明其卓越的",
"(1)",
"(+),",
"minmax",
"效率。基于这种动机，我们然后说明如何在实",
"2=1",
"现中实现S-SAM",
"其中给定模型权重w，内部最大化寻求最小",
"化E的扰动，并使L2-4经验范数。此训练目",
"4.1Theoretical Motivations",
"标旨在找到其邻域具有均匀低训练损失的模型",
"在本小节中，我们通过正式分析每个实例扰动",
"参数。由于找到E的精确解是NP困难的，因",
"提出的训练目标来激励我们的动态重新加权方",
"此SAM通过对批次的经验风险进行单步梯度",
"法，并表明它可以用加权批量扰动来近似，这",
"下降来估计内部最大化的解e*：",
"激发了我们的-SAM算法",
"1t(w),",
"N",
"Preliminary.我们从SAM中的锐度的角度来",
"l(w) ",
"=",
"N",
"激励我们的方法，该方法将损失景观的平坦度",
"量化为模型权重附近区域中损失的增加。每批",
"e* ~ arg max.",
"I(w)+eVl(w)",
"e:lell2≤p",
"和每实例权重扰动的锐度定义为：",
"= pVI(w)/ J/Vl(w)ll2",
"1.",
"N",
"Rbatch =",
"max",
" (l;(w+e) -l;(w),",
"e:l/",
"",
"2=1",
"外部最小化可以通过独立的优化器来执",
"行E(e。g。,Adam;Kingma andBa2015)。SAM",
"1N",
"将训练模型的计算成本大致提高了一倍，每批",
"Rnst=2",
"max",
"(li(w +ei) - li(w)) .",
"需要两次向前和两次向后传递。Alg.1中概述",
"- e:lill2≤p",
"了SAM算法",
"除了按批次进行扰动外，还可以对单个实例",
"由于非共享Ei，Rinst≥Rbatch，表明Rinst的",
"执行权重扰动：",
"正则化效果更强。但是，Rinst的计算成本很",
"高，因为N内部最大化问题中的Ei必须通过",
"",
"N",
"li",
"N单个实例上的梯度下降来计算，为此需要",
"li(w +ei),",
"(2)",
"",
"max",
"O(N）向后通过网络。在下面的分析中，我们",
"展示了如何用加权的每批权重扰动来近似更强",
"其中Ei是通过对单个实例进行单步梯度下降",
"的Rins",
"来计算的。这种方法类似于NLP中的许多对",
"我们首先考虑一般经验风险li的二阶展开，",
"抗训练方法，例如VATE（Miyatoetal.，2018)",
"例如：",
"和FreeLBE(Zhuetal.，2019），不同之处在于",
"扰动是根据模型权重而不是仅输入嵌入来计算",
"(w +e) = (w) +V(w)e+eTH:(w)e.",
"",
"的。我们将Eq.1和Eq.2的目标分别称为每",
"批重量扰动和每例重量扰动。Foretetal.（(2020)",
"为了进行易于处理的理论分析，我们假",
"在同一篇论文中指出，每个实例的权重扰动会",
"设Hessian是低秩的正定矩阵H;(w）=",
"产生较小的测试误差，并且是模型泛化的更好",
"aiVli(w)l;(w),(ai>O)。然后，我们以一步",
"预测指标。",
"梯度下降的封闭形式在二阶近似下获得Rbatch",
"尽管有效，但每个实例的权重扰动会大大增",
"和Rinst中的扰动，以与SAM的实践保持一致：",
"加计算和内存成本，需要2N向前和2N向后",
"传递一批大小的N。由于每个实例的权重扰",
"E= pVl(w) / IIVl(w)Il2 ,",
"动会独立修改所有模型权重，因此需要在不",
"Ei = pVl;(w)/ l/Vli(w)Il2 ",
"同的模型副本上计算每个单独实例的扰动。因",
"此，对于大规模训练，每个实例的权重扰动在",
"其中l(w)=Vl;(w)是批次的平均",
"计算上可能负担不起",
"梯度",
"TrainingObjective.获得Rbatch或Rinst后，",
"4SAM withDynamicReweighting",
"SAM将计算%Rbatch或%Rinst并更新模型",
"在本文中，我们寻求在权重扰动方面有更好",
"权重以最大程度地减少损失。为了实现更有效",
"的对手来改进SAM。由于SAM采用的每批权",
"的扰动，我们寻求与梯度Rinst保持一致，该"
] |
{
"arxiv_id": "2112.08772",
"language": "zh",
"page_idx": 3
}
|
||
2112.08772-zh-4
|
2112.08772-zh-4.jpg
|
梯度确定在强大的每实例对手下如何更新模型 不可计算，但回Eq.4 的关键观察是，可以通权重。我们希望以与每个实例对手“类似”的 过使用批处理中所有实例共享的扰动来优化每方式更新模型权重，而不是显式计算昂贵的术 个实例的权重扰动。因此，我们尝试仅使用$\epsilon^{\prime}$ 语$\mathcal{R}_\mathrm{inst}$。这里的“相似”是指$\mathcal{R}_\mathrm{inst}$的梯度与 的(粗略)估计来对模型进行扰动。现在下一我们的新目标之间的余弦相似性为正。$^{1}$
因此，我们首先导出每个实例权重扰动的梯度。具体来说，在内最大化步骤中计算$\epsilon_i$之后$(\epsilon_i$ 在外最小化中没有微分),每实例扰动的梯 次上使用一次反向传播来计算$\frac\partial{\partial w}\mathcal{R}_\mathrm{inst}$。要了
度为：
$$\begin{aligned}\frac{\partial}{\partial\boldsymbol{w}}\mathcal{R}_{\mathrm{inst}}&=\frac{1}{N}\sum_{i=1}^{N}\left(\nabla l_{i}(\boldsymbol{w}+\boldsymbol{\epsilon}_{i})-\nabla l_{i}(\boldsymbol{w})\right)\\&=\frac{1}{N}\sum_{i=1}^{N}\boldsymbol{H}_{i}(\boldsymbol{w})\boldsymbol{\epsilon}_{i}\\&=\frac{1}{N}\sum_{i=1}^{N}\rho a_{i}\left\|\nabla l_{i}(\boldsymbol{w})\right\|_{2}\nabla l_{i}(\boldsymbol{w}).\:(\end{aligned}$$
在本文中，我们的目标是找到一个每批扰动$\epsilon^{\prime}$,它产生的梯度的方向与每个实例的梯度对齐，因此将使用基于梯度的优化器类似地更新模型权重。具体来说，对于定义为$\mathcal{R}:=\frac1N\sum_{i=1}^{N}\left(l_{i}(\boldsymbol{w}+\boldsymbol{\epsilon}^{\prime})-l_{i}(\boldsymbol{w})\right)$的共享扰动$\epsilon^{\prime}$和$\mathcal{R}$,其梯度为：
$$\begin{aligned}\frac{\partial}{\partial\boldsymbol{w}}\mathcal{R}&=\frac{1}{N}\sum_{i=1}^{N}\left(\nabla l_{i}(\boldsymbol{w}+\boldsymbol{\epsilon}^{\prime})-\nabla l_{i}(\boldsymbol{w})\right)\\&=\left(\frac1N\sum_{i=1}^N\boldsymbol{H}_i(\boldsymbol{w})\right)\boldsymbol{\epsilon}^{\prime}.\end{aligned}$$
与普通的$\mathcal{R}_\mathrm{batch}$相比，这里我们建议使用不同的$\epsilon^{\prime}$。我们的目标是优化$\mathcal{R}$也会导致较小的$\mathcal{R}_{\mathrm{inst}};$也就是说，我们试图找到一个$\epsilon^\prime$,使得$\left(\frac{\partial}{\partial w}\mathcal{R}\right)^{\mathsf{T}}\frac{\partial}{\partial w}\mathcal{R}_{\mathrm{inst}}>0$。一个简单的选择是：
(4)
$$\epsilon'=\rho\cdot\left(\frac{\partial}{\partial\boldsymbol{w}}\mathcal{R}_{\mathrm{inst}}\right)\Big/\left\|\frac{\partial}{\partial\boldsymbol{w}}\mathcal{R}_{\mathrm{inst}}\right\|_{2}.$$
因为在我们的假设下每个$H_i$都是正定的，所以$\frac{1}{N}\sum_{i=1}^{N}\boldsymbol{H}_{i}(\boldsymbol{w})$ 也是正定的，那么我们有
$$\begin{aligned}&\left(\frac{\partial}{\partial\boldsymbol{w}}\mathcal{R}\right)^{\mathsf{T}}\frac{\partial}{\partial\boldsymbol{w}}\mathcal{R}_{\mathrm{inst}}\\&\propto\left(\frac{\partial}{\partial\boldsymbol{w}}\mathcal{R}_{\mathrm{inst}}\right)^{\mathsf{T}}\left(\frac{1}{N}\sum_{i=1}^{N}\boldsymbol{H}_{i}(\boldsymbol{w})\right)\left(\frac{\partial}{\partial\boldsymbol{w}}\mathcal{R}_{\mathrm{inst}}\right)\\&>0.\end{aligned}$$
因此，在此特定 $\epsilon^\prime=\frac\partial{\partial w}\mathcal{R}_{\mathrm{inst}}$下$,\epsilon^\prime$\_。尽管$\frac{\partial}{\partial w}\mathcal{R}_{\mathrm{inst}}$是我们旨在近似的梯度，并且在这里
If the cosine similarity is positive, then optimizing $\mathcal{R}$
also leads to optimized $\mathcal{R}_\mathrm{inst}.$
个挑战是如何有效地得出这样的估计。一个重要的观察结果是，在我们对$H_i(w)$的假设和二阶近似的使用下，可以通过在重新加权的批解这一事实，我们定义权重$g_i=a_i\left\|\nabla l_i(\boldsymbol{w})\right\|_2$ ,重新加权的批次为：
$$l_{\text{reweighted}}=\frac{1}{N}\sum_{i=1}^{N}g_{i}l_{i}(\boldsymbol{w}).$$
然后，将$g_i$视为常数$\frac\partial{\partial w}l_\text{reweighted}\propto\frac\partial{\partial w}\mathcal{R}_{\mathrm{inst~}}$, 如公式$\overline{\mathrm{E}}$3 中定义。与每批 SAM 相比，这种重新加权在计算实例权重$g_i$时只需要很小的额外计算成本。我们将在下一节中介绍如何有效地估计这些权重
## 4.2 Implementation
在$\delta$ -SAM 中，实现中的关键问题是如何有效地估计实例权重$g_i$ 。我们通过对随机扰动进行采样来解决此问题。具体来说，对于随机扰动$\underline{r}$,其中 $r_i$ 遵循高斯分布 $\mathcal{N}(0,\sigma\boldsymbol{I})$, 在与第$\boxed{\mathrm{F}}4.1$ 节中相同的假设下，我们有：
$$\begin{aligned}&E[(l_{i}(\boldsymbol{w}+\boldsymbol{r})-l_{i}(\boldsymbol{w}-\boldsymbol{r}))^{2}]\\&=E[\nabla l_{i}(\boldsymbol{w})^{\mathsf{T}}\boldsymbol{r}\cdot\nabla l_{i}(\boldsymbol{w})^{\mathsf{T}}\boldsymbol{r}]\\&=\sigma^{2}\left\|\nabla l_{i}(\boldsymbol{w})\right\|_{2}^{2}.\\&E[l_{i}(\boldsymbol{w}+\boldsymbol{r})+l_{i}(\boldsymbol{w}-\boldsymbol{r})-2l_{i}(\boldsymbol{w})]\\&=E[\boldsymbol{r}^{\intercal}\boldsymbol{H}_{\boldsymbol{i}}\boldsymbol{r}]\\&=a_{i}\sigma^{2}\left\|\nabla l_{i}(\boldsymbol{w})\right\|_{2}^{2}.\end{aligned}$$
因此，通过对随机扰动进行采样并取期望， 我们可以得到$a_i$ 的无偏估计和范数梯度XMATHX。每个估计都需要三个前向传递来计算$l_i( \boldsymbol{w})$ , $l_i( \boldsymbol{w}+ \boldsymbol{r})$和$l_i(\boldsymbol{w}-\boldsymbol{r})$ 。由于我们不需要保存用于反向传播的中间状态(PyTorch 中的 no\_grad ),因此这些正向传递比正常传递更快。在$\delta$ -SAM 中，为了算法的效率，我们在$\delta$ -SAM 中为每个批次仅采样一个 (共享) $r\sim\mathcal{N}(0,\sigma\boldsymbol{I})$ ,然后通过以下方式计算实例权重$g_i:$
$$g_i=\frac{|l_i(\boldsymbol{w}+\boldsymbol{r})+l_i(\boldsymbol{w}-\boldsymbol{r})-2l_i(\boldsymbol{w})|}{\max(|l_i(\boldsymbol{w}+\boldsymbol{r})-l_i(\boldsymbol{w}-\boldsymbol{r})|\:,\eta)},$$
(5)
其中$\eta$是避免除以零的超参数。导出实例权重后，加权批量权重扰动可以通过以下方式计
|
[
[
0.12176513671875,
0.0885009765625,
0.484375,
0.0885009765625,
0.484375,
0.09857177734375,
0.12176513671875,
0.09857177734375
],
[
0.51318359375,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.118408203125,
0.1015625,
0.486083984375,
0.1015625,
0.486083984375,
0.11517333984375,
0.118408203125,
0.11517333984375
],
[
0.51318359375,
0.10394287109375,
0.87744140625,
0.10394287109375,
0.87744140625,
0.114013671875,
0.51318359375,
0.114013671875
],
[
0.118408203125,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.51318359375,
0.1181640625,
0.87890625,
0.1181640625,
0.87890625,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.86376953125,
0.1353759765625,
0.86376953125,
0.1484375,
0.51318359375,
0.1484375
],
[
0.118408203125,
0.15087890625,
0.426513671875,
0.15087890625,
0.426513671875,
0.1644287109375,
0.118408203125,
0.1644287109375
],
[
0.515625,
0.15087890625,
0.87890625,
0.15087890625,
0.87890625,
0.1644287109375,
0.515625,
0.1644287109375
],
[
0.1368408203125,
0.166259765625,
0.484375,
0.166259765625,
0.484375,
0.179931640625,
0.1368408203125,
0.179931640625
],
[
0.51318359375,
0.16748046875,
0.8759765625,
0.16748046875,
0.8759765625,
0.1776123046875,
0.51318359375,
0.1776123046875
],
[
0.12005615234375,
0.182861328125,
0.486083984375,
0.182861328125,
0.486083984375,
0.196533203125,
0.12005615234375,
0.196533203125
],
[
0.5205078125,
0.18701171875,
0.52880859375,
0.18701171875,
0.52880859375,
0.191162109375,
0.5205078125,
0.191162109375
],
[
0.52490234375,
0.182861328125,
0.87890625,
0.182861328125,
0.87890625,
0.196533203125,
0.52490234375,
0.196533203125
],
[
0.1175537109375,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.1175537109375,
0.2120361328125
],
[
0.509765625,
0.19775390625,
0.87890625,
0.19775390625,
0.87890625,
0.215576171875,
0.509765625,
0.215576171875
],
[
0.118408203125,
0.21435546875,
0.16455078125,
0.21435546875,
0.16455078125,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.51318359375,
0.213134765625,
0.880859375,
0.215576171875,
0.880859375,
0.2298583984375,
0.51318359375,
0.22802734375
],
[
0.5146484375,
0.23095703125,
0.68701171875,
0.23095703125,
0.68701171875,
0.2452392578125,
0.5146484375,
0.2452392578125
],
[
0.2476806640625,
0.2369384765625,
0.267822265625,
0.2369384765625,
0.267822265625,
0.25048828125,
0.2476806640625,
0.25048828125
],
[
0.14111328125,
0.243408203125,
0.162841796875,
0.243408203125,
0.162841796875,
0.257080078125,
0.14111328125,
0.257080078125
],
[
0.15869140625,
0.248779296875,
0.2216796875,
0.255859375,
0.2166748046875,
0.270751953125,
0.154541015625,
0.26318359375
],
[
0.2149658203125,
0.25048828125,
0.465087890625,
0.2423095703125,
0.466796875,
0.268310546875,
0.2166748046875,
0.276611328125
],
[
0.2301025390625,
0.2464599609375,
0.2376708984375,
0.2464599609375,
0.2376708984375,
0.254150390625,
0.2301025390625,
0.254150390625
],
[
0.66015625,
0.254150390625,
0.80419921875,
0.240478515625,
0.8154296875,
0.295654296875,
0.6708984375,
0.308837890625
],
[
0.133544921875,
0.26123046875,
0.162841796875,
0.26123046875,
0.162841796875,
0.2724609375,
0.133544921875,
0.2724609375
],
[
0.246826171875,
0.26953125,
0.27197265625,
0.26953125,
0.27197265625,
0.281982421875,
0.246826171875,
0.281982421875
],
[
0.59375,
0.2724609375,
0.6884765625,
0.279052734375,
0.68701171875,
0.29443359375,
0.591796875,
0.2880859375
],
[
0.2509765625,
0.27978515625,
0.270263671875,
0.283203125,
0.26611328125,
0.29638671875,
0.246826171875,
0.29345703125
],
[
0.2333984375,
0.2861328125,
0.246826171875,
0.295654296875,
0.235107421875,
0.303955078125,
0.2216796875,
0.29443359375
],
[
0.2451171875,
0.29443359375,
0.343505859375,
0.298583984375,
0.341796875,
0.3134765625,
0.2435302734375,
0.308837890625
],
[
0.705078125,
0.290283203125,
0.73486328125,
0.290283203125,
0.73486328125,
0.302734375,
0.705078125,
0.302734375
],
[
0.22412109375,
0.305908203125,
0.252685546875,
0.305908203125,
0.252685546875,
0.318359375,
0.22412109375,
0.318359375
],
[
0.246826171875,
0.3134765625,
0.27197265625,
0.3134765625,
0.27197265625,
0.325927734375,
0.246826171875,
0.325927734375
],
[
0.198974609375,
0.32373046875,
0.49951171875,
0.3134765625,
0.50390625,
0.36767578125,
0.2032470703125,
0.377685546875
],
[
0.51123046875,
0.319580078125,
0.880859375,
0.319580078125,
0.880859375,
0.33740234375,
0.51123046875,
0.33740234375
],
[
0.252685546875,
0.327880859375,
0.26611328125,
0.327880859375,
0.26611328125,
0.341552734375,
0.252685546875,
0.341552734375
],
[
0.5146484375,
0.3349609375,
0.87890625,
0.3349609375,
0.87890625,
0.3486328125,
0.5146484375,
0.3486328125
],
[
0.51123046875,
0.350830078125,
0.87890625,
0.350830078125,
0.87890625,
0.364013671875,
0.51123046875,
0.364013671875
],
[
0.246826171875,
0.359375,
0.27197265625,
0.359375,
0.27197265625,
0.371826171875,
0.246826171875,
0.371826171875
],
[
0.5146484375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.5146484375,
0.3818359375
],
[
0.1368408203125,
0.38134765625,
0.486083984375,
0.38134765625,
0.486083984375,
0.3955078125,
0.1368408203125,
0.3955078125
],
[
0.51123046875,
0.385498046875,
0.64208984375,
0.385498046875,
0.64208984375,
0.3984375,
0.51123046875,
0.3984375
],
[
0.1175537109375,
0.3955078125,
0.484375,
0.396728515625,
0.484375,
0.410888671875,
0.1175537109375,
0.410400390625
],
[
0.118408203125,
0.413330078125,
0.48779296875,
0.413330078125,
0.48779296875,
0.427001953125,
0.118408203125,
0.427001953125
],
[
0.509765625,
0.413330078125,
0.68603515625,
0.413330078125,
0.68603515625,
0.427490234375,
0.509765625,
0.427490234375
],
[
0.118408203125,
0.429931640625,
0.486083984375,
0.429931640625,
0.486083984375,
0.443603515625,
0.118408203125,
0.443603515625
],
[
0.51123046875,
0.435302734375,
0.880859375,
0.435791015625,
0.880859375,
0.449462890625,
0.51123046875,
0.448974609375
],
[
0.1158447265625,
0.443603515625,
0.48779296875,
0.443603515625,
0.48779296875,
0.46142578125,
0.1158447265625,
0.46142578125
],
[
0.51318359375,
0.453125,
0.87890625,
0.453125,
0.87890625,
0.466064453125,
0.51318359375,
0.466064453125
],
[
0.1158447265625,
0.460205078125,
0.280517578125,
0.46142578125,
0.280517578125,
0.4755859375,
0.1158447265625,
0.474365234375
],
[
0.51318359375,
0.468505859375,
0.87890625,
0.468505859375,
0.87890625,
0.482177734375,
0.51318359375,
0.482177734375
],
[
0.21240234375,
0.48388671875,
0.46337890625,
0.495361328125,
0.4609375,
0.52685546875,
0.2091064453125,
0.515625
],
[
0.2435302734375,
0.486328125,
0.26025390625,
0.486328125,
0.26025390625,
0.495361328125,
0.2435302734375,
0.495361328125
],
[
0.5146484375,
0.485107421875,
0.880859375,
0.485107421875,
0.880859375,
0.498779296875,
0.5146484375,
0.498779296875
],
[
0.1485595703125,
0.49462890625,
0.162841796875,
0.488037109375,
0.175537109375,
0.49951171875,
0.160400390625,
0.50732421875
],
[
0.173828125,
0.50048828125,
0.204833984375,
0.50048828125,
0.204833984375,
0.5126953125,
0.173828125,
0.5126953125
],
[
0.51123046875,
0.49951171875,
0.80615234375,
0.50048828125,
0.80615234375,
0.515625,
0.51123046875,
0.51416015625
],
[
0.1485595703125,
0.50830078125,
0.185546875,
0.50830078125,
0.185546875,
0.52099609375,
0.1485595703125,
0.52099609375
],
[
0.20068359375,
0.50732421875,
0.258544921875,
0.5048828125,
0.26025390625,
0.52099609375,
0.202392578125,
0.5224609375
],
[
0.240966796875,
0.51708984375,
0.2626953125,
0.51708984375,
0.2626953125,
0.525390625,
0.240966796875,
0.525390625
],
[
0.258544921875,
0.5322265625,
0.274658203125,
0.5322265625,
0.274658203125,
0.5419921875,
0.258544921875,
0.5419921875
],
[
0.56689453125,
0.533203125,
0.78271484375,
0.533203125,
0.78271484375,
0.54638671875,
0.56689453125,
0.54638671875
],
[
0.219970703125,
0.54443359375,
0.344970703125,
0.5361328125,
0.34765625,
0.56005859375,
0.2225341796875,
0.568359375
],
[
0.2333984375,
0.53759765625,
0.258544921875,
0.53759765625,
0.258544921875,
0.548828125,
0.2333984375,
0.548828125
],
[
0.5634765625,
0.5517578125,
0.771484375,
0.5517578125,
0.771484375,
0.56640625,
0.5634765625,
0.56640625
],
[
0.2493896484375,
0.56103515625,
0.2822265625,
0.56103515625,
0.2822265625,
0.572265625,
0.2493896484375,
0.572265625
],
[
0.5634765625,
0.568359375,
0.6943359375,
0.57177734375,
0.6943359375,
0.59033203125,
0.5634765625,
0.58740234375
],
[
0.12005615234375,
0.58447265625,
0.484375,
0.58447265625,
0.484375,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.56689453125,
0.59326171875,
0.8330078125,
0.59326171875,
0.8330078125,
0.60693359375,
0.56689453125,
0.60693359375
],
[
0.118408203125,
0.599609375,
0.484375,
0.599609375,
0.484375,
0.61328125,
0.118408203125,
0.61328125
],
[
0.56103515625,
0.609375,
0.66650390625,
0.61083984375,
0.66650390625,
0.62890625,
0.56103515625,
0.626953125
],
[
0.1175537109375,
0.615234375,
0.484375,
0.6142578125,
0.484375,
0.62890625,
0.1175537109375,
0.6298828125
],
[
0.12176513671875,
0.63134765625,
0.45751953125,
0.63134765625,
0.45751953125,
0.6484375,
0.12176513671875,
0.6484375
],
[
0.5634765625,
0.6328125,
0.70947265625,
0.6328125,
0.70947265625,
0.65087890625,
0.5634765625,
0.65087890625
],
[
0.240966796875,
0.6611328125,
0.2626953125,
0.6611328125,
0.2626953125,
0.66943359375,
0.240966796875,
0.66943359375
],
[
0.26123046875,
0.66455078125,
0.313232421875,
0.66748046875,
0.312255859375,
0.681640625,
0.258544921875,
0.6787109375
],
[
0.355224609375,
0.66015625,
0.36865234375,
0.66015625,
0.36865234375,
0.66845703125,
0.355224609375,
0.66845703125
],
[
0.37451171875,
0.66455078125,
0.42236328125,
0.66845703125,
0.4189453125,
0.681640625,
0.3720703125,
0.677734375
],
[
0.462646484375,
0.66650390625,
0.486083984375,
0.66650390625,
0.486083984375,
0.6787109375,
0.462646484375,
0.6787109375
],
[
0.51318359375,
0.66650390625,
0.880859375,
0.66650390625,
0.880859375,
0.6796875,
0.51318359375,
0.6796875
],
[
0.225830078125,
0.6728515625,
0.270263671875,
0.6728515625,
0.270263671875,
0.68701171875,
0.225830078125,
0.68701171875
],
[
0.162841796875,
0.6650390625,
0.2283935546875,
0.6650390625,
0.2283935546875,
0.68115234375,
0.162841796875,
0.68115234375
],
[
0.344970703125,
0.6748046875,
0.38037109375,
0.6748046875,
0.38037109375,
0.68603515625,
0.344970703125,
0.68603515625
],
[
0.4130859375,
0.6787109375,
0.432373046875,
0.6787109375,
0.432373046875,
0.6884765625,
0.4130859375,
0.6884765625
],
[
0.51318359375,
0.68310546875,
0.87890625,
0.68310546875,
0.87890625,
0.69677734375,
0.51318359375,
0.69677734375
],
[
0.118408203125,
0.69970703125,
0.48779296875,
0.69970703125,
0.48779296875,
0.71337890625,
0.118408203125,
0.71337890625
],
[
0.5087890625,
0.69775390625,
0.87890625,
0.69775390625,
0.87890625,
0.71142578125,
0.5087890625,
0.71142578125
],
[
0.1175537109375,
0.71435546875,
0.476806640625,
0.71435546875,
0.476806640625,
0.732421875,
0.1175537109375,
0.732421875
],
[
0.51318359375,
0.71435546875,
0.87890625,
0.71435546875,
0.87890625,
0.72802734375,
0.51318359375,
0.72802734375
],
[
0.515625,
0.73291015625,
0.87890625,
0.73291015625,
0.87890625,
0.74365234375,
0.515625,
0.74365234375
],
[
0.14697265625,
0.74462890625,
0.16455078125,
0.7412109375,
0.1695556640625,
0.751953125,
0.151123046875,
0.7548828125
],
[
0.2216796875,
0.74365234375,
0.235107421875,
0.74365234375,
0.235107421875,
0.75439453125,
0.2216796875,
0.75439453125
],
[
0.2435302734375,
0.751953125,
0.279541015625,
0.751953125,
0.279541015625,
0.76416015625,
0.2435302734375,
0.76416015625
],
[
0.5146484375,
0.74755859375,
0.880859375,
0.74755859375,
0.880859375,
0.76123046875,
0.5146484375,
0.76123046875
],
[
0.1376953125,
0.7587890625,
0.175537109375,
0.7587890625,
0.175537109375,
0.771484375,
0.1376953125,
0.771484375
],
[
0.175537109375,
0.75439453125,
0.19140625,
0.75439453125,
0.19140625,
0.76123046875,
0.175537109375,
0.76123046875
],
[
0.2149658203125,
0.7587890625,
0.2435302734375,
0.7587890625,
0.2435302734375,
0.771484375,
0.2149658203125,
0.771484375
],
[
0.51318359375,
0.7626953125,
0.87890625,
0.7626953125,
0.87890625,
0.775390625,
0.51318359375,
0.775390625
],
[
0.2939453125,
0.779296875,
0.310546875,
0.779296875,
0.310546875,
0.7880859375,
0.2939453125,
0.7880859375
],
[
0.51318359375,
0.779296875,
0.880859375,
0.779296875,
0.880859375,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.41064453125,
0.78515625,
0.426513671875,
0.78515625,
0.426513671875,
0.79443359375,
0.41064453125,
0.79443359375
],
[
0.27197265625,
0.7861328125,
0.279541015625,
0.7861328125,
0.279541015625,
0.7919921875,
0.27197265625,
0.7919921875
],
[
0.172119140625,
0.78076171875,
0.183837890625,
0.7890625,
0.172119140625,
0.79736328125,
0.160400390625,
0.7890625
],
[
0.187255859375,
0.7890625,
0.22412109375,
0.79345703125,
0.2216796875,
0.80712890625,
0.183837890625,
0.802734375
],
[
0.280517578125,
0.7880859375,
0.312255859375,
0.7880859375,
0.312255859375,
0.80029296875,
0.280517578125,
0.80029296875
],
[
0.306396484375,
0.79150390625,
0.3720703125,
0.79150390625,
0.3720703125,
0.8056640625,
0.306396484375,
0.8056640625
],
[
0.429931640625,
0.7890625,
0.47265625,
0.79443359375,
0.47021484375,
0.80810546875,
0.426513671875,
0.802734375
],
[
0.1259765625,
0.798828125,
0.1368408203125,
0.79638671875,
0.139404296875,
0.8017578125,
0.1292724609375,
0.80419921875
],
[
0.2626953125,
0.798828125,
0.307373046875,
0.798828125,
0.307373046875,
0.8134765625,
0.2626953125,
0.8134765625
],
[
0.51318359375,
0.7958984375,
0.880859375,
0.7958984375,
0.880859375,
0.80859375,
0.51318359375,
0.80859375
],
[
0.3955078125,
0.80029296875,
0.4375,
0.80029296875,
0.4375,
0.81396484375,
0.3955078125,
0.81396484375
],
[
0.15283203125,
0.80029296875,
0.189697265625,
0.80029296875,
0.189697265625,
0.8134765625,
0.15283203125,
0.8134765625
],
[
0.2880859375,
0.80810546875,
0.31494140625,
0.80810546875,
0.31494140625,
0.82080078125,
0.2880859375,
0.82080078125
],
[
0.5146484375,
0.8125,
0.56689453125,
0.8125,
0.56689453125,
0.82666015625,
0.5146484375,
0.82666015625
],
[
0.1259765625,
0.82470703125,
0.160400390625,
0.82470703125,
0.160400390625,
0.83740234375,
0.1259765625,
0.83740234375
],
[
0.5751953125,
0.84130859375,
0.82958984375,
0.84130859375,
0.82958984375,
0.85498046875,
0.5751953125,
0.85498046875
],
[
0.118408203125,
0.8515625,
0.484375,
0.8515625,
0.484375,
0.8662109375,
0.118408203125,
0.8662109375
],
[
0.5341796875,
0.8525390625,
0.568359375,
0.8525390625,
0.568359375,
0.8642578125,
0.5341796875,
0.8642578125
],
[
0.8564453125,
0.85107421875,
0.880859375,
0.85107421875,
0.880859375,
0.86328125,
0.8564453125,
0.86328125
],
[
0.57421875,
0.85693359375,
0.82958984375,
0.85791015625,
0.82958984375,
0.87255859375,
0.57421875,
0.87158203125
],
[
0.1158447265625,
0.86767578125,
0.48779296875,
0.86328125,
0.48779296875,
0.88427734375,
0.1158447265625,
0.88818359375
],
[
0.1368408203125,
0.89111328125,
0.486083984375,
0.892578125,
0.486083984375,
0.90673828125,
0.1368408203125,
0.90625
],
[
0.5146484375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.5146484375,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.295654296875,
0.90478515625,
0.295654296875,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.5146484375,
0.90478515625,
0.87890625,
0.90478515625,
0.87890625,
0.91845703125,
0.5146484375,
0.91845703125
]
] |
[
"梯度确定在强大的每实例对手下如何更新模型",
"不可计算，但EEq.4的关键观察是，可以通",
"权重。我们希望以与每个实例对手“类似”的",
"过使用批处理中所有实例共享的扰动来优化每",
"方式更新模型权重，而不是显式计算昂贵的术",
"个实例的权重扰动。因此，我们尝试仅使用e",
"语Rinst。这里的“相似”是指Rinst 的梯度与",
"的（粗略）估计来对模型进行扰动。现在下",
"我们的新目标之间的余弦相似性为正。",
"个挑战是如何有效地得出这样的估计。一个重",
"因此，我们首先导出每个实例权重扰动的梯",
"要的观察结果是，在我们对H（W）的假设和",
"度。具体来说，在内最大化步骤中计算E之后",
"",
"二阶近似的使用下，可以通过在重新加权的批",
"（e，在外最小化中没有微分），每实例扰动的梯",
"次上使用一次反向传播来计算%Rinst。要了",
"度为：",
"解这一事实，我们定义权重gi=ai/Vl;(w)Il2",
"，重新加权的批次为：",
"N",
"8",
"Rinst =",
"(Vi(w +e:) - V;(w)",
"1",
"a",
"dw",
"i-1",
"lreweighted =",
"N",
"1",
"厂Hi(w)e",
"i1",
"N",
"i=1",
"-/ ( i (o 3 ",
"然后，将gi视为常数晶lreweightedα晶Rinst",
"N",
"如公式E3中定义。与每批SAM相比，这种重",
"新加权在计算实例权重9i时只需要很小的额",
"i-1",
"外计算成本。我们将在下二节中介绍如何有效",
"在本文中，我们的目标是找到一个每批扰",
"地估计这些权重",
"动ε，它产生的梯度的方向与每个实例的",
"梯度对齐，因此将使用基于梯度的优化器类",
"4.2Implementation",
"似地更新模型权重。具体来说，对于定义为",
"在8-SAM中，实现中的关键问题是如何有效",
"R:= i(l;(w+e')-l;(w)的共享扰动",
"地估计实例权重9元。我们通过对随机扰动进",
"e'和R，其梯度为：",
"行采样来解决此问题。具体来说，对于随机扰",
"(Vi(w+e)-V()",
"N",
"动r，其中ri遵循高斯分布N(O,oI），在与",
"O",
"R=",
"第E4.1节中相同的假设下，我们有：",
"w",
"N",
"i=1",
"N",
"E[(li(w +r) -li(w -r))2i",
"(H(w)",
"1 ",
"= E[Vli(w)Tr . Vli;(w)Tr",
"i=1",
"= 2 //Vl;(w)2",
"与普通的Rbatch相比，这里我们建议使用不同",
"E[l;(w +r) +li(w -r) - 2l;(w))",
"的e’。我们的目标是优化R也会导致较小的",
"= E[rTH;r]",
"Rinst；也就是说，我们试图找到一个e'，使得",
"(%R）%Rinst>0。一个简单的选择是：",
"= aio2 I/Vl;(w)I2",
"0",
"-Rinst",
"0",
"Rinst",
"(4)",
"因此，通过对随机扰动进行采样并取期望，",
"(ow",
"=p.",
"Ow",
"12",
"我们可以得到ai的无偏估计和范数梯度",
"因为在我们的假设下每个H，都是正定的，所",
"XMATHX。每个估计都需要三个前向传递来",
"以H;(w)也是正定的，那么我们有：",
"计算l(w)，l(w+r)和l(w-r)。由于我们",
"不需要保存用于反向传播的中间状态（PyTorch",
"0",
"0",
"Rinst",
"中的no_grad），因此这些正向传递比正常传",
"aw",
"R",
"ow",
"递更快。在8-SAM中，为了算法的效率，我",
"N",
"们在8-SAM中为每个批次仅采样一个（共享）",
"0",
"1",
"0",
"Rins",
"1",
"H;(w)",
"Rinst",
"X",
"NZ",
"r～N(O,αI），然后通过以下方式计算实例权",
"Ow",
"Ow",
"i=1",
"重gi：",
"> 0.",
"[l;(w +r) +l;(w -r) - 2l;(w)l",
"因此，在此特定e=Rinst下，e-。尽管",
"9i=",
"(5)",
"max(lli(w +r) - li(w-r)l,n)",
"嘉Rinst 是我们旨在近似的梯度，并且在这里",
"'if the cosine similarity is positive, then optimizing R",
"其中Ⅱ是避免除以零的超参数。导出实例权",
"also leads to optimized Rinst ",
"重后，加权批量权重扰动可以通过以下方式计"
] |
{
"arxiv_id": "2112.08772",
"language": "zh",
"page_idx": 4
}
|
||
2403.09712-zh-0
|
2403.09712-zh-0.jpg
|
<table>
<tbody>
<tr>
<th>Question</th>
<th>What is the period of the author of $OffonaComet?$</th>
</tr>
<tr>
<td>Answer</td>
<td>1828-1905</td>
</tr>
<tr>
<td> </td>
<td>Knowledge Graph</td>
</tr>
<tr>
<td>Around $the\textit{Moon}$</td>
<td>Jules Verne Entity uthor period 1828-1905 Attribute ıthor occupation</td>
</tr>
<tr>
<td>$Off.$ $On$ $a$ Comet</td>
<td>novelist playwright</td>
</tr>
</tbody>
</table>
Figure 1: A toy example of KBQA, which requires complex reasoning marked in red
<NEXT\_THLINX<LIN
# 1 INTRODUCTION
基于知识的问答(KBQA)是自然语言处理(NLP)和数据挖掘研究$\overline{\mathbb{E}}[33]$ 中的一项关键任务，它可以充当访问和处理 Web 数据和知识的方法，并导致有用的应用程序，例如智能语音助手和搜索引擎，尤其是大型语言模型(LLM)$\mathbb{E}[30]$。如图$\mathbb{E}$1所示，KBQA 旨在基于背景知识以自然语言回答问题，背景知识通常被格式化为知识图(KGs)[E][19,45,49]。因此，KBQA需要自然语言理解(NLU)和知识推理的能力，使其成为相关领域的一项具有挑战性的任务
在文献中，由于在其他 NLP 任务$\overline{\mathbb{E}}[12,15,18]$ 上的显著成果，研究人员提出了许多基于深度学习的 KBQA$\underline{\mathbb{E}}[28,33,49]$ 解决方案，其中预训练语言模型 (LM) 因其强大的 NLU 能力$\mathbb{E}[4,6,29,30]$ 而成为最有希望的。不幸的是，包括 LLM 在内的 LM 在知识应用$\mathbb{E}[21,23]$ 中效果不佳，这阻碍了其在 KBQA 中的应用。因此，研究人员已经尝试了很大的努力，以增强与知识相关的任务的LMs (将知识事实输人到 LMs 中，或者使用基于知识的任务$\mathbb{E}[21,31,34,36,37,46,47,50,53]$ 对 LMs 进行预训练),这极大地改善了知识相关任务中的 LMs。但是，这些方法通常将 KG 作为额外的预训练语料库$\mathbb{E}[21,50]$ 的补充来学习，这不能覆盖整个 KG,并且可能会忽略某些在某些任务中有用的知识，从而导致知识学习不完整。在这一点上，一个简单的解决方案是基于 KG 生成预训练语料库。尽管已经沿着这条路线开发了许多方法[2,3,20,48] ,但它们通常依赖于特定的技术或资源来有效地生成语料库(例如。g。,需要预先训练的生成模型来生成句子，或者以固定格式生成句子),这在实践中可能无法使用，从而限制了其应用。因此，在本文中，我们希望设计一个通用框架来生成以 KG 为中心的语料库，以对 LM 进行全面的知识预训练，该框架不仅限于特定技术， 还可以使用不同的详细实现来实现灵活的应用程序
然而，沿着这条线，存在几个不平凡的技术挑战。首先，有许多解决方案可以根据给定的 KG 针对不同的需求生成句子(例如g。,预训练生成 LM[2],固定句子模板[20])。而且，尽管大多数 KG 都将知识三元组与实体 ID 一起存储，但是一些高质量的 KG 还包含附加的属性信息，这些信息以各种形式存储 (e。g。、文本、数字和日期),并且需要不同的处理。如何统一和概括这些不同的技术和数据形式仍然是开放的。其次，生成的句子与自然句子不同，甚至可能看起来失真，这可能会误导LM 并损害 LM 在预训练中的自然语言理解能力$\mathbb{E}[2,20]$。现有方法根据其生成方法使用特定技术来解决此问题(e。g。,用复杂的生成式 LMs[2]生成更类似于自然的句子，或者采用专门设计的句子模板来减少负面影响[20]),但是对于一般框架中的任意生成方法，如何克服这一缺点是一个不平凡的问题。最后，现有的用 KGs 增强 LMs 的方法更侧重于利用 K-BERT [21] 和 ERNIE [50] 等相关知识来提高语言理解，而很少考虑类似人类的复杂推理能力。人类可以按照特定的模式对多个知识事实进行推理，这在 KBQA 任务中也是广泛要求的。例如，在图$\overline{\mathrm{E}}$1 中，要得出答案，LM 首先需要找到彗星上的 Off 的作者是儒勒·凡尔纳，然后儒勒·凡尔纳的时期是1828-1905 年。如何使 LM 具有如此复杂的推理能力是一个具有挑战性的问题
为此，在本文中，我们提出了一个通用的 Knowledge-InjectedCurprocoursP 再培训框架 (KICP),以实现对 KBQA 的全面 KG 学习和开发，该框架由知识注人(KI),知识适应(KA)和课程推理(CR)。具体来说，KI模块将KG 三元组转换为句子，以构建用于完整知识学习的预训练语料库，并将该过程概括为三个关键步骤，即。e。,文本表征，句子构造和掩蔽，可以通过不同的详细技术和各种数据形式来实现，以实现灵活的应用。接下来，为了减少生成语料库和自然语料库之间的差异对 LM 预训练带来的负面影响，KA 模块修复了原始 LM 以保持其 NLU 能力，并通过与 LM 一起工作的可训练适配器从生成的语料库中学习知识。最后，
|
[
[
0.1690673828125,
0.14453125,
0.25732421875,
0.14453125,
0.25732421875,
0.1622314453125,
0.1690673828125,
0.1622314453125
],
[
0.3349609375,
0.1458740234375,
0.82861328125,
0.1458740234375,
0.82861328125,
0.160400390625,
0.3349609375,
0.160400390625
],
[
0.1732177734375,
0.17236328125,
0.253173828125,
0.17236328125,
0.253173828125,
0.1900634765625,
0.1732177734375,
0.1900634765625
],
[
0.52685546875,
0.173583984375,
0.63818359375,
0.173583984375,
0.63818359375,
0.1881103515625,
0.52685546875,
0.1881103515625
],
[
0.414306640625,
0.2020263671875,
0.58642578125,
0.2020263671875,
0.58642578125,
0.216552734375,
0.414306640625,
0.216552734375
],
[
0.37744140625,
0.2259521484375,
0.490234375,
0.2279052734375,
0.490234375,
0.242431640625,
0.37744140625,
0.240478515625
],
[
0.77392578125,
0.231689453125,
0.83642578125,
0.23486328125,
0.8349609375,
0.25390625,
0.77294921875,
0.250732421875
],
[
0.1478271484375,
0.242431640625,
0.226318359375,
0.242431640625,
0.226318359375,
0.260009765625,
0.1478271484375,
0.260009765625
],
[
0.293212890625,
0.24365234375,
0.35546875,
0.24560546875,
0.3544921875,
0.261474609375,
0.291748046875,
0.2587890625
],
[
0.495849609375,
0.24560546875,
0.56396484375,
0.24365234375,
0.56396484375,
0.2587890625,
0.495849609375,
0.261474609375
],
[
0.58642578125,
0.25390625,
0.69921875,
0.25390625,
0.69921875,
0.271484375,
0.58642578125,
0.271484375
],
[
0.1380615234375,
0.264404296875,
0.2320556640625,
0.264404296875,
0.2320556640625,
0.2822265625,
0.1380615234375,
0.2822265625
],
[
0.76123046875,
0.260009765625,
0.85888671875,
0.261962890625,
0.85888671875,
0.27978515625,
0.76123046875,
0.27783203125
],
[
0.285888671875,
0.281005859375,
0.346435546875,
0.279052734375,
0.34716796875,
0.294189453125,
0.28759765625,
0.296630859375
],
[
0.5361328125,
0.28857421875,
0.64208984375,
0.28857421875,
0.64208984375,
0.30615234375,
0.5361328125,
0.30615234375
],
[
0.39208984375,
0.298583984375,
0.50244140625,
0.296630859375,
0.50244140625,
0.311279296875,
0.39208984375,
0.313232421875
],
[
0.143798828125,
0.30810546875,
0.221435546875,
0.30810546875,
0.221435546875,
0.32568359375,
0.143798828125,
0.32568359375
],
[
0.3349609375,
0.32568359375,
0.414306640625,
0.32568359375,
0.414306640625,
0.3447265625,
0.3349609375,
0.3447265625
],
[
0.53515625,
0.324462890625,
0.6435546875,
0.324462890625,
0.6435546875,
0.3466796875,
0.53515625,
0.3466796875
],
[
0.1519775390625,
0.33203125,
0.214111328125,
0.33203125,
0.214111328125,
0.3466796875,
0.1519775390625,
0.3466796875
],
[
0.221435546875,
0.406005859375,
0.77783203125,
0.406005859375,
0.77783203125,
0.419189453125,
0.221435546875,
0.419189453125
],
[
0.10455322265625,
0.448974609375,
0.23779296875,
0.448974609375,
0.23779296875,
0.462158203125,
0.10455322265625,
0.462158203125
],
[
0.09234619140625,
0.48095703125,
0.09722900390625,
0.48095703125,
0.09722900390625,
0.4873046875,
0.09234619140625,
0.4873046875
],
[
0.101318359375,
0.481689453125,
0.11358642578125,
0.481689453125,
0.11358642578125,
0.486083984375,
0.101318359375,
0.486083984375
],
[
0.11517333984375,
0.4765625,
0.259033203125,
0.4765625,
0.259033203125,
0.490478515625,
0.11517333984375,
0.490478515625
],
[
0.08905029296875,
0.496826171875,
0.91015625,
0.496826171875,
0.91015625,
0.5068359375,
0.08905029296875,
0.5068359375
],
[
0.0882568359375,
0.5087890625,
0.91259765625,
0.5087890625,
0.91259765625,
0.52294921875,
0.0882568359375,
0.52294921875
],
[
0.0882568359375,
0.52294921875,
0.91259765625,
0.52294921875,
0.91259765625,
0.5361328125,
0.0882568359375,
0.5361328125
],
[
0.08660888671875,
0.537109375,
0.5439453125,
0.537109375,
0.5439453125,
0.55029296875,
0.08660888671875,
0.55029296875
],
[
0.10296630859375,
0.55029296875,
0.9111328125,
0.55029296875,
0.9111328125,
0.564453125,
0.10296630859375,
0.564453125
],
[
0.0882568359375,
0.56640625,
0.9111328125,
0.56640625,
0.9111328125,
0.5771484375,
0.0882568359375,
0.5771484375
],
[
0.0882568359375,
0.57763671875,
0.91015625,
0.57763671875,
0.91015625,
0.59130859375,
0.0882568359375,
0.59130859375
],
[
0.08660888671875,
0.59130859375,
0.91259765625,
0.59228515625,
0.91259765625,
0.60595703125,
0.08660888671875,
0.60498046875
],
[
0.08905029296875,
0.60595703125,
0.91259765625,
0.60595703125,
0.91259765625,
0.619140625,
0.08905029296875,
0.619140625
],
[
0.0882568359375,
0.619140625,
0.91259765625,
0.619140625,
0.91259765625,
0.63330078125,
0.0882568359375,
0.63330078125
],
[
0.0882568359375,
0.63525390625,
0.91015625,
0.63525390625,
0.91015625,
0.64599609375,
0.0882568359375,
0.64599609375
],
[
0.11358642578125,
0.64794921875,
0.91259765625,
0.64794921875,
0.91259765625,
0.6611328125,
0.11358642578125,
0.6611328125
],
[
0.08905029296875,
0.66357421875,
0.91259765625,
0.66357421875,
0.91259765625,
0.673828125,
0.08905029296875,
0.673828125
],
[
0.08660888671875,
0.67529296875,
0.406982421875,
0.67529296875,
0.406982421875,
0.68603515625,
0.08660888671875,
0.68603515625
],
[
0.10296630859375,
0.689453125,
0.9111328125,
0.689453125,
0.9111328125,
0.70263671875,
0.10296630859375,
0.70263671875
],
[
0.08660888671875,
0.70263671875,
0.91259765625,
0.7021484375,
0.91259765625,
0.71630859375,
0.08660888671875,
0.71728515625
],
[
0.0882568359375,
0.71728515625,
0.9111328125,
0.71728515625,
0.9111328125,
0.73095703125,
0.0882568359375,
0.73095703125
],
[
0.0882568359375,
0.73095703125,
0.91259765625,
0.73095703125,
0.91259765625,
0.744140625,
0.0882568359375,
0.744140625
],
[
0.0906982421875,
0.7470703125,
0.90869140625,
0.7470703125,
0.90869140625,
0.7568359375,
0.0906982421875,
0.7568359375
],
[
0.08905029296875,
0.75830078125,
0.91259765625,
0.75830078125,
0.91259765625,
0.771484375,
0.08905029296875,
0.771484375
],
[
0.08660888671875,
0.771484375,
0.91259765625,
0.771484375,
0.91259765625,
0.78515625,
0.08660888671875,
0.78515625
],
[
0.08905029296875,
0.787109375,
0.90673828125,
0.787109375,
0.90673828125,
0.79736328125,
0.08905029296875,
0.79736328125
],
[
0.0882568359375,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.0882568359375,
0.81298828125
],
[
0.0882568359375,
0.814453125,
0.57275390625,
0.814453125,
0.57275390625,
0.82763671875,
0.0882568359375,
0.82763671875
],
[
0.10540771484375,
0.8291015625,
0.91015625,
0.8291015625,
0.91015625,
0.8388671875,
0.10540771484375,
0.8388671875
],
[
0.0882568359375,
0.84130859375,
0.91015625,
0.84130859375,
0.91015625,
0.85498046875,
0.0882568359375,
0.85498046875
],
[
0.0906982421875,
0.85693359375,
0.56787109375,
0.85693359375,
0.56787109375,
0.8671875,
0.0906982421875,
0.8671875
],
[
0.58251953125,
0.85693359375,
0.90869140625,
0.85693359375,
0.90869140625,
0.8671875,
0.58251953125,
0.8671875
],
[
0.0882568359375,
0.87060546875,
0.9111328125,
0.87060546875,
0.9111328125,
0.880859375,
0.0882568359375,
0.880859375
],
[
0.08905029296875,
0.88525390625,
0.9111328125,
0.88525390625,
0.9111328125,
0.89599609375,
0.08905029296875,
0.89599609375
]
] |
[
"Question",
"What is the period of the author of Off on a Comet ?",
"Answer",
"1828 -1905",
"KnowledgeGraph",
"JulesVerne",
"Entity",
"Around",
"author",
"period",
"1828 - 1905",
"the Moon",
"Attribute",
"author",
"occupation",
"/occupation",
"Off on a",
"novelist",
"playwright",
"Comet",
"Figure l: A toy example of KBQA, which requires complex reasoning marked in red.",
"<NEXT_THLINX<LIN",
"",
"",
"INTRODUCTION",
"基于知识的问答（KBOA）是自然语言处理（NLP）和数据挖掘研究E[33]中的一项关键任务，它可以充当访问和处理Web数据和",
"知识的方法，并导致有用的应用程序，例如智能语音助手和搜索引擎，尤其是大型语言模型（LLM)E[30】。如图E1所示，KBQA",
"旨在基于背景知识以自然语言回答问题，背景知识通常被格式化为知识图（KGs)E[19,45,49]。因此，KBQA需要自然语言理解",
"（NLU）和知识推理的能力，使其成为相关领域的一项具有挑战性的任务",
"在文献中，由于在其他NLP任务E[12,15,18]上的显著成果，研究人员提出了许多基于深度学习的KBQAE[28,33,49]解决方",
"案，其中预训练语言模型（LM）因其强大的NLU能力E[4,6,29,30]而成为最有希望的。不幸的是，包括LLM在内的LM在知识",
"应用E[21,23]中效果不佳，这阻碍了其在KBQA中的应用。因此，研究人员已经尝试了很大的努力，以增强与知识相关的任务的",
"LMs（将知识事实输人到LMs中，或者使用基于知识的任务E[21,31,34,36,37，46,47,50,53]对LMs进行预训练），这极大地改善",
"了知识相关任务中的LMs。但是，这些方法通常将KG作为额外的预训练语料库E[21,50]的补充来学习，这不能覆盖整个KG，并",
"且可能会忽略某些在某些任务中有用的知识，从而导致知识学习不完整。在这一点上，一个简单的解决方案是基于KG生成预训",
"练语料库。尽管已经沿着这条路线开发了许多方法[2，3，20，48]，但它们通常依赖于特定的技术或资源来有效地生成语料库（例如。",
"需要预先训练的生成模型来生成句子，或者以固定格式生成句子)，这在实践中可能无法使用，从而限制了其应用。因此，在",
"本文中，我们希望设计一个通用框架来生成以KG为中心的语料库，以对LM进行全面的知识预训练，该框架不仅限于特定技术，",
"还可以使用不同的详细实现来实现灵活的应用程序",
"然而，沿着这条线，存在几个不平凡的技术挑战。首先，有许多解决方案可以根据给定的KG针对不同的需求生成句子（例如。",
"g。，预训练生成LM[2]，固定句子模板[20]）。而且，尽管大多数KG都将知识三元组与实体ID一起存储，但是一些高质量的KG",
"还包含附加的属性信息，这些信息以各种形式存储（e。g。、文本、数字和日期)，并且需要不同的处理。如何统一和概括这些不同",
"的技术和数据形式仍然是开放的。其次，生成的句子与自然句子不同，甚至可能看起来失真，这可能会误导LM并损害LM在预",
"训练中的自然语言理解能力E[2.20]。现有方法根据其生成方法使用特定技术来解决此问题（e。g。，用复杂的生成式LMs[2]生",
"成更类似于自然的句子，或者采用专门设计的句子模板来减少负面影响[20]），但是对于一般框架中的任意生成方法，如何克服",
"这一缺点是一个不平凡的问题。最后，现有的用KGs增强LMs的方法更侧重于利用K-BERT[21]和ERNIE[50]等相关知识来提高",
"语言理解，而很少考虑类似人类的复杂推理能力。人类可以按照特定的模式对多个知识事实进行推理，这在KBOA任务中也是广",
"泛要求的。例如，在图E1中，要得出答案，LM首先需要找到彗星上的Of的作者是儒勒凡尔纳，然后儒勒·凡尔纳的时期是",
"1828-1905年。如何使LM具有如此复杂的推理能力是一个具有挑战性的问题",
"为此，在本文中，我们提出了一个通用的Knowledge-InjectedCurprocoursP再培训框架（KICP），以实现对KBOA的全面KG学",
"习和开发，该框架由知识注入（KI)，知识适应（KA）和课程推理（CR)。具体来说，KI模块将KG三元组转换为句子，以构建用",
"于完整知识学习的预训练语料库，并将该过程概括为三个关键步骤，即。e",
"文本表征，句子构造和掩蔽，可以通过不同的详细",
"技术和各种数据形式来实现，以实现灵活的应用。接下来，为了减少生成语料库和自然语料库之间的差异对LM预训练带来的负",
"面影响，KA模块修复了原始LM以保持其NLU能力，并通过与LM一起工作的可训练适配器从生成的语料库中学习知识。最后，"
] |
{
"arxiv_id": "2403.09712",
"language": "zh",
"page_idx": 0
}
|
||
2403.09712-zh-1
|
2403.09712-zh-1.jpg
|
WWW'24, May 13-17, 2024, Singapore, Singapore
Xin Lin et al.
<FigureHere>
Figure 2: The architecture of the proposed KICP framework. (a) The overview of KICP. (b) The knowledge injection module (KI) converts KG triples into sentences. (c) The knowledge adaptation module (KA) works with the LM to keep NLU ability and learn knowledge. (d) The curriculum reasoning module (CR) constructs easy-to-hard reasoning-required pretraining corpora.
为了预训练具有复杂推理能力的 LM, CR 模块遵循人类常见的推理模式，并构建需要复杂知识推理的语料库。此外，CR 模块将复杂的语料库排列成难度增加的三节课，并按照课程学习方式从易到难对 LM 进行训练，以降低预训练难度。最后，我们提供了通用框架的实现，并在四个实词数据集上进行了广泛的实验，以评估 KICP。结果表明，我们的框架可以实现更高的性能，并很好地推广到其他 QA 任务
# 2 RELATED WORK
Knowledge-Based Question Answering. 在文献中，关于 KBQA 的研究大致可以分为知识增强的 LM (稍后介绍),以及基于 KG 的推理，包括基于路径的$\underline{\mathrm{E}[27]}$ ,基于嵌人的$\underline{\mathrm{E}[11,33]}$ 和基于图的方法$\overline{\mathrm{E}[10,14,\underline{28},44,45,49]}$。基于路径的方法将问题映射到实体和关系中，以便在 KGE[27] 上进行推理。基于嵌人的方法，例如 EmbedKGQA$\mathbb{E}[33]$ ,将问题和 KG 表示在同一潜在空间中，并通过简单的向量计算来推断答案。基于图的方法[10,14,28,44,45,49]从 KG 中采样一个子图，并用神经网络对子图进行推理。基于图的方法由于在可解释性和性能之间的良好权衡而被广泛应用于复杂推理中，但是子图内的知识建模不足可能导致有限的鲁棒性。此外，大型语言模型 (LLM) 最近已成为 KBQA 任务中很有前途的方法$\overline{\mathbb{E}}[1,8]$。研究人员提出了几种先进的技术来提高其知识推理能力，包括思想链提示$\boxed{\mathbb{E}}[39]$, 问题分解$\boxed{\mathbb{E}}[52]$ 和检索增强生成$\boxed{\mathbb{E}}[43]$
Knowledge-Enhanced Language Model. 由于预训练的 LM 在基于知识的任务$\overline{\mathbb{E}[21,23]}$ 上显示出其弱点，因此研究人员尝试了许多努力来使用 KG 的知识来增强 LM, 包括显式方法$\overline{\mathbb{E}}[21,26,31,50]$ 和隐式方法$\overline{\mathbb{E}}[9,20,34,36,37,42]$。显式方法将知识事实或嵌人作为附加输入馈送到 LM 中，以利用相关知识。例如，K-BERT[21] 将知识三元组注人句子中，作为 LM 的输人。Zhang 等人。[50] 开发了一个聚合器网络，将 KG 实体嵌人到 LM 中。隐式方法设计特殊的预训练任务，以使用 LM 从 KG 和语料库中学习知识。Sun 等人。[34] 引人了用于预训练的实体屏蔽策略，Wang 等人。[37] 将 LM 训练为具有实体描述的知识嵌入模型。为了更好地利用 KG 三元组，Liu 等人。[20] 用 KG 三元组和 Agarwal 等人生成多语言综合预训练语料库。[2] 采用生成 LM 来合成更自然的语料库。总之，显式方法更直接地利用知识，但需要额外的知识注释作为输人，而隐式方法可以很容易地应用于下游任务，但需要大量的预训练
我们的工作与以前的方法不同，如下所示。首先，将 KG 转换为语料库的现有方法通常仅限于特定的技术和资源，而我们的方法是一个通用框架，可以针对不同的情况使用不同的详细实现。其次，现有的方法更侧重于利用相关知识提高语言理解，而我们的方法进一步使具有复杂推理能力的 LM 具有专门设计的预训练任务
3 KICP: KNOWLEDGE-INIECTED CURRICULUM PRETRAINING
3.1 $\textbf{Problem Definition}$
基于知识的问答(KBQA)由知识图$\mathcal{K}G$和问答对$(Q,Y)$组成。我们假设 KG 包含有关两个实体之间的关系以及每个实体的属性的知识三元组，其中属性值具有可以转换为文本的多种形式(文本在词汇表$V$上定义为$V^+)$。因此，KG 可以定义为$\mathcal{X}G=(\mathbb{E},\mathbb{R},\Sigma)$ ,其中$\mathbb{E}$是实体集，$\mathbb{R}$是关系和属性集，$\Sigma$表示知识三元组。每个三元组$(h,r,t)\in\Sigma\left(h,t\in\mathbb{E},r\in\mathbb{R}\right)$表示实体$h$和$t$具有关系$r$ (e。g。,“JulesVerne”是图$\mathbb{E}1$ 中“彗星上的关闭”的“作者”),而$(h,r,t)\in\sum\left(h\in\mathbb{E},r\in\mathbb{R},t\in V^+\right)$表示实体$r$的属性$r$,其中$V^+$属性为。g。,“儒勒·凡尔纳”的“时期”为“1828-1905”) 此外，每个实体 $e\in\mathbb{E}$都分配有多个名称 $N_e=\{n_e1,n_e_2,\ldots,n_e_k\}$
|
[
[
0.0882568359375,
0.07574462890625,
0.32275390625,
0.0770263671875,
0.32275390625,
0.08837890625,
0.0882568359375,
0.08709716796875
],
[
0.84716796875,
0.0770263671875,
0.90869140625,
0.0770263671875,
0.90869140625,
0.08837890625,
0.84716796875,
0.08837890625
],
[
0.134033203125,
0.12371826171875,
0.261474609375,
0.12371826171875,
0.261474609375,
0.136962890625,
0.134033203125,
0.136962890625
],
[
0.2998046875,
0.12371826171875,
0.38818359375,
0.12371826171875,
0.38818359375,
0.1351318359375,
0.2998046875,
0.1351318359375
],
[
0.4345703125,
0.12371826171875,
0.576171875,
0.12371826171875,
0.576171875,
0.136962890625,
0.4345703125,
0.136962890625
],
[
0.73046875,
0.12371826171875,
0.8857421875,
0.12371826171875,
0.8857421875,
0.136962890625,
0.73046875,
0.136962890625
],
[
0.11517333984375,
0.136962890625,
0.191162109375,
0.1383056640625,
0.191162109375,
0.149658203125,
0.11517333984375,
0.1483154296875
],
[
0.31201171875,
0.1427001953125,
0.41015625,
0.1427001953125,
0.41015625,
0.154052734375,
0.31201171875,
0.154052734375
],
[
0.6005859375,
0.136962890625,
0.70361328125,
0.1383056640625,
0.70361328125,
0.1483154296875,
0.6005859375,
0.147705078125
],
[
0.4306640625,
0.143310546875,
0.5888671875,
0.143310546875,
0.5888671875,
0.154052734375,
0.4306640625,
0.154052734375
],
[
0.76123046875,
0.143310546875,
0.85302734375,
0.143310546875,
0.85302734375,
0.1546630859375,
0.76123046875,
0.1546630859375
],
[
0.62744140625,
0.149658203125,
0.685546875,
0.149658203125,
0.685546875,
0.1572265625,
0.62744140625,
0.1572265625
],
[
0.2034912109375,
0.1483154296875,
0.273681640625,
0.1483154296875,
0.273681640625,
0.160400390625,
0.2034912109375,
0.160400390625
],
[
0.1217041015625,
0.1514892578125,
0.19775390625,
0.1546630859375,
0.1968994140625,
0.169189453125,
0.12091064453125,
0.1666259765625
],
[
0.775390625,
0.154052734375,
0.8388671875,
0.154052734375,
0.8388671875,
0.1641845703125,
0.775390625,
0.1641845703125
],
[
0.205078125,
0.160400390625,
0.275390625,
0.1634521484375,
0.275390625,
0.1749267578125,
0.205078125,
0.1717529296875
],
[
0.5556640625,
0.169189453125,
0.71337890625,
0.17041015625,
0.71337890625,
0.1849365234375,
0.5556640625,
0.184326171875
],
[
0.79736328125,
0.1666259765625,
0.84716796875,
0.1666259765625,
0.84716796875,
0.177978515625,
0.79736328125,
0.177978515625
],
[
0.73388671875,
0.1787109375,
0.7587890625,
0.1787109375,
0.7587890625,
0.1875,
0.73388671875,
0.1875
],
[
0.7900390625,
0.177978515625,
0.84326171875,
0.177978515625,
0.84326171875,
0.189453125,
0.7900390625,
0.189453125
],
[
0.1217041015625,
0.184326171875,
0.188720703125,
0.184326171875,
0.188720703125,
0.1956787109375,
0.1217041015625,
0.1956787109375
],
[
0.301513671875,
0.179931640625,
0.371826171875,
0.1817626953125,
0.371826171875,
0.1925048828125,
0.301513671875,
0.1900634765625
],
[
0.73388671875,
0.1875,
0.75732421875,
0.1875,
0.75732421875,
0.1988525390625,
0.73388671875,
0.1988525390625
],
[
0.79345703125,
0.1849365234375,
0.83837890625,
0.1875,
0.83642578125,
0.2020263671875,
0.79150390625,
0.199462890625
],
[
0.3056640625,
0.1912841796875,
0.414306640625,
0.1912841796875,
0.414306640625,
0.2052001953125,
0.3056640625,
0.2052001953125
],
[
0.451904296875,
0.1932373046875,
0.5703125,
0.1932373046875,
0.5703125,
0.2008056640625,
0.451904296875,
0.2008056640625
],
[
0.60693359375,
0.1925048828125,
0.70361328125,
0.1925048828125,
0.70361328125,
0.20263671875,
0.60693359375,
0.20263671875
],
[
0.453369140625,
0.2020263671875,
0.56640625,
0.20263671875,
0.56640625,
0.2139892578125,
0.453369140625,
0.21337890625
],
[
0.301513671875,
0.20263671875,
0.4248046875,
0.20263671875,
0.4248046875,
0.21337890625,
0.301513671875,
0.21337890625
],
[
0.12091064453125,
0.1976318359375,
0.188720703125,
0.1976318359375,
0.188720703125,
0.208984375,
0.12091064453125,
0.208984375
],
[
0.6044921875,
0.20263671875,
0.7060546875,
0.20263671875,
0.7060546875,
0.21337890625,
0.6044921875,
0.21337890625
],
[
0.74609375,
0.2108154296875,
0.8056640625,
0.2108154296875,
0.8056640625,
0.2215576171875,
0.74609375,
0.2215576171875
],
[
0.12091064453125,
0.2215576171875,
0.259033203125,
0.2215576171875,
0.259033203125,
0.23486328125,
0.12091064453125,
0.23486328125
],
[
0.431396484375,
0.2259521484375,
0.58642578125,
0.2259521484375,
0.58642578125,
0.2392578125,
0.431396484375,
0.2392578125
],
[
0.76123046875,
0.2203369140625,
0.79345703125,
0.2203369140625,
0.79345703125,
0.23291015625,
0.76123046875,
0.23291015625
],
[
0.1641845703125,
0.23291015625,
0.2181396484375,
0.23291015625,
0.2181396484375,
0.24365234375,
0.1641845703125,
0.24365234375
],
[
0.3095703125,
0.238037109375,
0.451904296875,
0.2392578125,
0.451904296875,
0.250732421875,
0.3095703125,
0.25
],
[
0.76708984375,
0.23486328125,
0.78125,
0.2310791015625,
0.78515625,
0.240478515625,
0.771484375,
0.24365234375
],
[
0.1871337890625,
0.246826171875,
0.193603515625,
0.24365234375,
0.1968994140625,
0.246826171875,
0.1895751953125,
0.25
],
[
0.74609375,
0.246826171875,
0.8056640625,
0.246826171875,
0.8056640625,
0.2568359375,
0.74609375,
0.2568359375
],
[
0.30810546875,
0.25,
0.7138671875,
0.25,
0.7138671875,
0.26318359375,
0.30810546875,
0.26318359375
],
[
0.11358642578125,
0.248779296875,
0.148681640625,
0.250732421875,
0.1478271484375,
0.261962890625,
0.11273193359375,
0.260009765625
],
[
0.156005859375,
0.25390625,
0.2254638671875,
0.255615234375,
0.2254638671875,
0.270751953125,
0.156005859375,
0.268310546875
],
[
0.3095703125,
0.26318359375,
0.4248046875,
0.26318359375,
0.4248046875,
0.276611328125,
0.3095703125,
0.276611328125
],
[
0.7607421875,
0.25830078125,
0.79150390625,
0.25830078125,
0.79150390625,
0.266357421875,
0.7607421875,
0.266357421875
],
[
0.50634765625,
0.26953125,
0.57421875,
0.26953125,
0.57421875,
0.276611328125,
0.50634765625,
0.276611328125
],
[
0.6396484375,
0.26953125,
0.708984375,
0.26953125,
0.708984375,
0.276611328125,
0.6396484375,
0.276611328125
],
[
0.310546875,
0.27587890625,
0.486083984375,
0.274658203125,
0.486083984375,
0.285888671875,
0.310546875,
0.287353515625
],
[
0.51318359375,
0.27978515625,
0.68310546875,
0.27978515625,
0.68310546875,
0.287353515625,
0.51318359375,
0.287353515625
],
[
0.1666259765625,
0.270751953125,
0.221435546875,
0.270751953125,
0.221435546875,
0.2822265625,
0.1666259765625,
0.2822265625
],
[
0.12091064453125,
0.281005859375,
0.13232421875,
0.281005859375,
0.13232421875,
0.2890625,
0.12091064453125,
0.2890625
],
[
0.316162109375,
0.28857421875,
0.48193359375,
0.28857421875,
0.48193359375,
0.3017578125,
0.316162109375,
0.3017578125
],
[
0.51123046875,
0.28857421875,
0.70751953125,
0.28857421875,
0.70751953125,
0.298583984375,
0.51123046875,
0.298583984375
],
[
0.74853515625,
0.2841796875,
0.80224609375,
0.2841796875,
0.80224609375,
0.291748046875,
0.74853515625,
0.291748046875
],
[
0.76123046875,
0.291748046875,
0.79150390625,
0.29345703125,
0.7900390625,
0.3037109375,
0.7607421875,
0.3017578125
],
[
0.56884765625,
0.302978515625,
0.58642578125,
0.302978515625,
0.58642578125,
0.30615234375,
0.56884765625,
0.30615234375
],
[
0.5947265625,
0.300537109375,
0.66650390625,
0.300537109375,
0.66650390625,
0.30810546875,
0.5947265625,
0.30810546875
],
[
0.408447265625,
0.30810546875,
0.439453125,
0.30810546875,
0.439453125,
0.3125,
0.408447265625,
0.3125
],
[
0.734375,
0.313232421875,
0.76953125,
0.313232421875,
0.76953125,
0.32080078125,
0.734375,
0.32080078125
],
[
0.34228515625,
0.30615234375,
0.381591796875,
0.30615234375,
0.381591796875,
0.313232421875,
0.34228515625,
0.313232421875
],
[
0.38232421875,
0.317626953125,
0.4794921875,
0.317626953125,
0.4794921875,
0.324462890625,
0.38232421875,
0.324462890625
],
[
0.51123046875,
0.317626953125,
0.69140625,
0.317626953125,
0.69140625,
0.32763671875,
0.51123046875,
0.32763671875
],
[
0.346435546875,
0.3193359375,
0.37744140625,
0.3193359375,
0.37744140625,
0.322509765625,
0.346435546875,
0.322509765625
],
[
0.313720703125,
0.318115234375,
0.3447265625,
0.318115234375,
0.3447265625,
0.322509765625,
0.313720703125,
0.322509765625
],
[
0.740234375,
0.3212890625,
0.875,
0.3212890625,
0.875,
0.335205078125,
0.740234375,
0.335205078125
],
[
0.406982421875,
0.330810546875,
0.4794921875,
0.330810546875,
0.4794921875,
0.33837890625,
0.406982421875,
0.33837890625
],
[
0.57421875,
0.330810546875,
0.6357421875,
0.330810546875,
0.6357421875,
0.335205078125,
0.57421875,
0.335205078125
],
[
0.77783203125,
0.333251953125,
0.8349609375,
0.333251953125,
0.8349609375,
0.343505859375,
0.77783203125,
0.343505859375
],
[
0.0882568359375,
0.37890625,
0.91015625,
0.37890625,
0.91015625,
0.392578125,
0.0882568359375,
0.392578125
],
[
0.0882568359375,
0.392578125,
0.91015625,
0.392578125,
0.91015625,
0.406005859375,
0.0882568359375,
0.406005859375
],
[
0.08660888671875,
0.4072265625,
0.90869140625,
0.4072265625,
0.90869140625,
0.42041015625,
0.08660888671875,
0.42041015625
],
[
0.08905029296875,
0.4443359375,
0.91015625,
0.4443359375,
0.91015625,
0.455078125,
0.08905029296875,
0.455078125
],
[
0.0882568359375,
0.45703125,
0.9111328125,
0.45703125,
0.9111328125,
0.47021484375,
0.0882568359375,
0.47021484375
],
[
0.0882568359375,
0.47021484375,
0.91259765625,
0.47021484375,
0.91259765625,
0.484130859375,
0.0882568359375,
0.484130859375
],
[
0.08660888671875,
0.484130859375,
0.21728515625,
0.484130859375,
0.21728515625,
0.49755859375,
0.08660888671875,
0.49755859375
],
[
0.08660888671875,
0.51318359375,
0.11517333984375,
0.51318359375,
0.11517333984375,
0.521484375,
0.08660888671875,
0.521484375
],
[
0.111083984375,
0.51025390625,
0.261474609375,
0.51025390625,
0.261474609375,
0.5234375,
0.111083984375,
0.5234375
],
[
0.08905029296875,
0.52783203125,
0.91259765625,
0.52783203125,
0.91259765625,
0.54150390625,
0.08905029296875,
0.54150390625
],
[
0.0882568359375,
0.54248046875,
0.91259765625,
0.54248046875,
0.91259765625,
0.55615234375,
0.0882568359375,
0.55615234375
],
[
0.08660888671875,
0.55615234375,
0.91259765625,
0.55615234375,
0.91259765625,
0.5693359375,
0.08660888671875,
0.5693359375
],
[
0.08905029296875,
0.5712890625,
0.91259765625,
0.5712890625,
0.91259765625,
0.58203125,
0.08905029296875,
0.58203125
],
[
0.08905029296875,
0.58544921875,
0.9111328125,
0.58544921875,
0.9111328125,
0.59521484375,
0.08905029296875,
0.59521484375
],
[
0.0882568359375,
0.59765625,
0.91259765625,
0.59765625,
0.91259765625,
0.611328125,
0.0882568359375,
0.611328125
],
[
0.08905029296875,
0.61279296875,
0.57666015625,
0.61279296875,
0.57666015625,
0.62353515625,
0.08905029296875,
0.62353515625
],
[
0.0906982421875,
0.62548828125,
0.91259765625,
0.62548828125,
0.91259765625,
0.6396484375,
0.0906982421875,
0.6396484375
],
[
0.0882568359375,
0.6396484375,
0.91015625,
0.6396484375,
0.91015625,
0.65283203125,
0.0882568359375,
0.65283203125
],
[
0.08660888671875,
0.6513671875,
0.9111328125,
0.65283203125,
0.9111328125,
0.66650390625,
0.08660888671875,
0.66552734375
],
[
0.0882568359375,
0.66650390625,
0.91259765625,
0.66650390625,
0.91259765625,
0.6796875,
0.0882568359375,
0.6796875
],
[
0.08905029296875,
0.681640625,
0.9111328125,
0.681640625,
0.9111328125,
0.6923828125,
0.08905029296875,
0.6923828125
],
[
0.0882568359375,
0.6943359375,
0.91259765625,
0.6943359375,
0.91259765625,
0.70849609375,
0.0882568359375,
0.70849609375
],
[
0.08660888671875,
0.70703125,
0.91259765625,
0.70849609375,
0.91259765625,
0.7216796875,
0.08660888671875,
0.72021484375
],
[
0.0882568359375,
0.72265625,
0.1168212890625,
0.72265625,
0.1168212890625,
0.734375,
0.0882568359375,
0.734375
],
[
0.10296630859375,
0.736328125,
0.9111328125,
0.7373046875,
0.9111328125,
0.74853515625,
0.10296630859375,
0.74755859375
],
[
0.08660888671875,
0.75,
0.91259765625,
0.75,
0.91259765625,
0.76318359375,
0.08660888671875,
0.76318359375
],
[
0.08660888671875,
0.76513671875,
0.52685546875,
0.76513671875,
0.52685546875,
0.77587890625,
0.08660888671875,
0.77587890625
],
[
0.08416748046875,
0.78857421875,
0.6337890625,
0.78857421875,
0.6337890625,
0.8017578125,
0.08416748046875,
0.8017578125
],
[
0.0849609375,
0.8095703125,
0.28759765625,
0.8095703125,
0.28759765625,
0.82275390625,
0.0849609375,
0.82275390625
],
[
0.08905029296875,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.08905029296875,
0.8388671875
],
[
0.08660888671875,
0.84033203125,
0.91015625,
0.84130859375,
0.91015625,
0.85498046875,
0.08660888671875,
0.85400390625
],
[
0.096435546875,
0.861328125,
0.10296630859375,
0.861328125,
0.10296630859375,
0.8642578125,
0.096435546875,
0.8642578125
],
[
0.1004638671875,
0.85693359375,
0.90869140625,
0.85693359375,
0.90869140625,
0.8671875,
0.1004638671875,
0.8671875
],
[
0.08660888671875,
0.86865234375,
0.91259765625,
0.86865234375,
0.91259765625,
0.88330078125,
0.08660888671875,
0.88330078125
],
[
0.0882568359375,
0.88330078125,
0.9111328125,
0.88330078125,
0.9111328125,
0.89794921875,
0.0882568359375,
0.89794921875
]
] |
[
"www24,May13172024,SingaporeSingapore",
"Xin Lin et al.",
"(a) KICP Framework",
"Knowledge Triplet",
"(b)Knowledge Injection",
"(c) Knowledge Adaptation",
"Pretraining Task",
"(el,rl, 1828-1905)",
"(JulesVeme,period",
"TextCharacterization (Txt)",
"Knowledge-Enhanced",
"1828-1905)",
"Curriculum",
"MLMLOSs",
"Representation",
"Reasoning",
"Sentence Construction ()",
"Tunable",
"Fixed",
"Knowledge",
"Knowledge",
"MaskedCorpus",
"LM",
"Adapter",
"The period of [MASK]",
"EntityMasking/Whole",
"The period of Jules",
"Word Masking (Msk)",
"MASKJis1828-1905",
"Adaptation",
"Verne is 1828-1905.",
"Transformer",
"The period of Jules Verme is",
"(d) Curriculum Reasoning",
"Block",
"1828-1905",
"Lesson I: Knowledge Learning",
"T",
"",
"Transformer",
"(Off ona Comel,author,Jules Veme)Theauthor of Ofon a Comet is JulesVeme",
"Corpus",
"Knowledge",
"Lesson 2: CoT Learning",
"Block",
"Theauthorof",
"is Jules Veme.",
"(Offon a Comet,author,JulesVeme)",
"[SEPTheperiodofJulesVemeis18",
"Injection",
"KC",
"(JulesVeme.period 1828-1905)",
"1905.[SEP]Theperiodof theauthorofO",
"Transforme",
"Block",
"",
"is1828-1905",
"neat",
"Sentenc",
"13:Com",
".author.JulesVerne",
"Theperiodof theauthorof OffonaCor",
"",
"",
"The period of Jules Veme is",
"d.1828-1905)",
"1S18X-100",
"1828-1905",
"Figure 2: The architecture of the proposed KICP framework. (a) The overview of KICP. (b) The knowledge injection module",
"(KI) converts KG triples into sentences. (c) The knowledge adaptation module (KA) works with the LM to keep NLU ability and",
"learn knowledge. (d) The curriculum reasoning module (CR) constructs easy-to-hard reasoning-required pretraining corpora",
"为了预训练具有复杂推理能力的LM，CR模块遵循人类常见的推理模式，并构建需要复杂知识推理的语料库。此外，CR模块将复",
"杂的语料库排列成难度增加的三节课，并按照课程学习方式从易到难对LM进行训练，以降低预训练难度。最后，我们提供了通",
"用框架的实现，并在四个实词数据集上进行了广泛的实验，以评估KICP。结果表明，我们的框架可以实现更高的性能，并很好地",
"推广到其他QA任务",
"2",
"RELATED WORK",
"Knowledge-BasedQuestionAnswering.在文献中，关于KBQA的研究大致可以分为知识增强的LM（稍后介绍），以及基于KG的",
"推理，包括基于路径的E[27]，基于嵌人的E[11,33]和基于图的方法E[10,14,28,44,45,49]。基于路径的方法将问题映射到实体",
"和关系中，以便在KG国[27]上进行推理。基于嵌人的方法，例如EmbedKGQAE[33]，将问题和KG表示在同一潜在空间中，并",
"通过简单的向量计算来推断答案。基于图的方法[10,14,28，44,45，49]从KG中采样一个子图，并用神经网络对子图进行推理。基",
"于图的方法由于在可解释性和性能之间的良好权衡而被广泛应用于复杂推理中，但是子图内的知识建模不足可能导致有限的鲁棒",
"性。此外，大型语言模型（LLM）最近已成为KBQA任务中很有前途的方法E[1,8】。研究人员提出了几种先进的技术来提高其知",
"识推理能力，包括思想链提示E[39]，问题分解E[52]和检索增强生成E[43]",
"Knowledge-EnhancedLanguageModel.由于预训练的LM在基于知识的任务E[21,23]上显示出其弱点，因此研究人员尝试了许多",
"努力来使用KG的知识来增强LM，包括显式方法E[21,26,31,50]和隐式方法E[9,20,34,36,37,42]。显式方法将知识事实或嵌入",
"作为附加输入馈送到LM中，以利用相关知识。例如，K-BERT[21]将知识三元组注人句子中，作为LM的输人。Zhang等人。[50]",
"开发了一个聚合器网络，将KG实体嵌入到LM中。隐式方法设计特殊的预训练任务，以使用LM从KG和语料库中学习知识。Sun",
"等人。[34]引入了用于预训练的实体屏蔽策略，Wang等人。[37]将LM训练为具有实体描述的知识嵌入模型。为了更好地利用KG",
"三元组，Liu等人。[20]用KG三元组和Agarwal等人生成多语言综合预训练语料库。[2]采用生成LM来合成更自然的语料库。总",
"之，显式方法更直接地利用知识，但需要额外的知识注释作为输入，而隐式方法可以很容易地应用于下游任务，但需要大量的预",
"训练",
"我们的工作与以前的方法不同，如下所示。首先，将KG转换为语料库的现有方法通常仅限于特定的技术和资源，而我们的方",
"法是一个通用框架，可以针对不同的情况使用不同的详细实现。其次，现有的方法更侧重于利用相关知识提高语言理解，而我们",
"的方法进一步使具有复杂推理能力的LM具有专门设计的预训练任务",
"3KICP:KNOWLEDGE-INJECTEDCURRICULUMPRETRAINING",
"3.1ProblemDefinition",
"基于知识的问答（KBOA）由知识图KG和问答对（O.Y）组成。我们假设KG包含有关两个实体之间的关系以及每个实体的属性的",
"知识三元组，其中属性值具有可以转换为文本的多种形式（文本在词汇表V上定义为V+）。因此，KG可以定义为KG=（B,R,Z）",
"",
"其中E是实体集，R是关系和属性集，Z表示知识三元组。每个三元组（h，rt）EZ（h,tEE，rER）表示实体h和t具有关系",
"(e。g。，“JulesVerne”是图国1中“彗星上的关闭”的“作者\")，而（h,r,t)e（heB，rER，teV+）表示实体r的属性r，其",
"中V+属性为。g。，“儒勒·凡尔纳”的“时期”为“1828-1905”）此外，每个实体eEB都分配有多个名称Ne=（ne1,ne2.nek)"
] |
{
"arxiv_id": "2403.09712",
"language": "zh",
"page_idx": 1
}
|
||
2403.09712-zh-2
|
2403.09712-zh-2.jpg
|
A Knowledge-Injected Curriculum Pretraining Framework for Question Answering
WWW'24, May 13-17, 2024, Singapore, Singapore
(每个名称$n_e_i\in V^+)。\mathbb{R}$分配有类似的名称。在问答对$(Q,Y)$中 , $Q= \{ q_1, q_2, \ldots , q_n\} \in V^+$ $( q_i\in V)$是自然语言中的问题，而 Y 是在$\mathcal{K}G$下推断的$Q$的答案，其形式取决于任务 (e。g。, KBQA 经常从$\mathcal{K}G$中选择一个实体或属性值，而生成 QA 则从自然文本或数学表达式等某些词汇中生成形式语言$\mathbb{E}[16,17]$ )
给定知识图$KG$和问答对$(Q,Y)$, KBQA 的目标是训练一个模型$M:(\mathcal{R}G,Q)\to Y$,以预测$\mathcal{K}G$下问题$Q$的答案 Y。在本文中我们首先用$\kappa g$预训练一个语言模型$\mathcal{L}\mathcal{M}$ ,然后在$M$中使用它来预测答案$Y$到$Q$ 。我们期望$\mathcal{L}\mathcal{M}$可以全面地从$\kappa g$中学习知识，并很好地处理复杂的推理
# 3.2 Method
我们提出了一个通用的 Knowledge-InjectedC 过程 P 再训练框架 (KICP) 来对$\mathcal{L}\mathcal{M}$进行预训练，以进行全面的知识学习和复杂的推理，该框架不仅限于特定的技术，而且可以轻松地与不同的实现一起使用以实现灵活的应用程序。如图E2 E(a)所示，KICP 由三个关键组件组成，i。e。知识注人(KI)、知识适应(KA) 和课程推理 (CR)。具体而言，KI 通过将 KG 三元组转换为句子来构建预训练语料库，从而将 KG 中的知识完全注人 LM 中，并将各种生成技术概括为三个关键步骤。为了减少生成语料库与自然语料库之间的差距带来的负面影响，KA 修复了原始 LM 以保持其 NLU 能力，并为框架配备了可训练的知识适配器，以从生成的语料库中学习知识。为了对具有复杂推理能力的 LM 进行预训练，CR 遵循人类推理的常见模式，构建了几个具有不同难度的推理所需的语料库，并以课程的方式对 LM 进行了从易到难的训练，以促进模型学习
3.2.1 Knowledge Injection.为了克服使用 KG 作为外部语料库的补充所带来的知识学习不足，我们直接将 KG 三元组转换为句子作为预训练语料库，以将知识注人 LM。此外，在文献$\mathbb{E}[2,20]$ 中存在几种针对不同要求的有效句子生成技术，并且 KG 包含需要不同处理的多种形式的数据(例如。g。、ID、文本、数字和日期)。因此，为了将这些详细技术概括为一个通用框架，该框架不限于在各种情况下灵活应用的特定技术，如图$\overline{\mathrm{E}}$2 E( b) 所示，我们将句子生成过程抽象为三个关键步骤，i。e。、文本表征、句子构造和掩蔽
Text Characterization. 给定从$\mathcal{K}G$采样的一个三元组$k=(h,r,t)\in\Sigma$ , KI 首先将三元组的所有字段表征为文本$(Txt)$,这些文本用作句子的主干元素以生成。对于存储在 ID 中的实体和关系，我们映射无意义的 ID(e。g。, e1) 到一个有意义的名称(JulesVerne), 该名称在每次迭代中从关联的名称集中动态采样，以增加语料库多样性。更多的采样策略也可以在这里应用于其他需求[20]。对于各种形式的属性值(e。g、数字、日期和文本),我们使用它们的文字描述，因为它们总是可以用文字表达，尽管原始形式。这样，我们可以统一实体、关系和属性值的不同处理
Sentence Construction. 获得文本元素后，KI 应用句子构造策略$\tau$将这些元素组装成一个完整的句子，包括对元素进行重新排序和
转换以及添加辅助词。策略$\tau$可以用不同的现有技术来实现，如句子模板、基于语法的规则和生成的 LMs [2,20]
Masking.最后一步是对生成的句子进行掩码，以进行掩码语言模型(MLM)预训练。为了强制知识学习并匹配实体和属性值之间的差异，我们更喜欢对句子中的知识元素(从三元组转换而来的)赋予更多的权重，并对实体和属性值应用不同的屏蔽策略 Msk 。例如，我们在实体上应用实体掩码[34],它掩码整个实体名称以强制学习关系知识，而不是记忆实体名称，并在属性值上应用全字掩码(WWM)[5],因为这些值可能包含太多信息(例如。g。,传记),如果都被掩盖，很难恢复。WWM 的工作方式也类似于短值上的实体掩码 (例如。g。, 数字) 通过屏蔽整个单词。这里可以使用更多的掩码技术作为 Msk
总的来说，句子生成过程如下：
$$KI(k)=Msk(\tau(Txt(h),Txt(r),Txt(t))),\:k=(h,r,t)\in\sum.$$
(1)
知识注人的语料库由句子 KI(k) 组成，这些句子是从预训练中从 KG 采样的三元组动态生成的。这样，KI 将整个 KG 转换为语料库，从而隐式地将来自 KG 的所有信息存储在语料库中，例如结构信息。与将 KG 重写为语料库的现有方法相比，KI 不依赖于特定的技术或资源，因此可以针对各种应用程序需求使用不同的实现
3.2.2 Knowledge Adaptation.显然，KI 生成的语料库与自然语料库不同，因为句子可能不严格遵循语法(特别是对于一些简单的$\tau$),并且语料库的多样性受到限制。在语料库上对 LM 进行预训练可能会损害 NLU 的能力，并且在白然文本上效果不佳。此外， 由于拟议的通用框架中的句子生成技术是任意的，因此我们不能像现有研究那样使用与特定生成技术相关的方法来解决该问题回[2,20] 。因此，在知识适应 (KA) 中，\_我们转向在知识预训练期间保持 LM 的 NLU 能力
如图El2 E(c) 所示，遵循 LM 调优$\mathbb{E}[7,36]$ 中的适配器范例，我们固定 LM 参数并在原始 LM $LM$ 上方添加可训练的知识适配器模块$Ad$。$Ad$使用$LM$的语义输出作为输入，并输出知识增强的表示。此外，为了深人改善语义和知识的融合，使用了LM 中所有层的语义输出。KA 的计算公式如下：
$$KA(x)=Ad(LM(x)),$$
(2)
,其中$x$是输人句子。$Ad$可以用任何神经网络实现，预计它有一个适当的大小，以包含足够的空间进行知识学习，并避免大大增
加计算复杂性
在预训练中，对$Ad$的参数进行了训练，以从构建的语料库中学习知识，而原始 LM 是固定的。由于原始 LM 不受$Ad$的影响，
因此尽可能保留了 NLU 能力，以减少生成的语料库与自然语料库之间的差距所带来的负面影响
3.2.3 Curriculum Reasoning.使用 KI 和 KA, KICP 可以有效地将 KG 注人 LM, 但仍然缺乏现实世界 KBQA 任务中所要求的对多个知识事实的复杂推理能力。为了使 LM 具有这种能力，课程推理模块(CR)在需要复杂推理的语料库上对 LM 进行预训练，如图$\mathbb{E}$2$\mathbb{E}($d)所示
很难为所有 KG 收集足够的推理所需的语料库，因此我们也基于 KG 构建了语料库。人类经常按照特定的模式进行复杂的推理(例如。g。,多顶推理),这对参与的三元组(e。g。,链状三元组)。因此，我们按照这些模式构建语料库(例如。g。,“《彗星上的起
|
[
[
0.0882568359375,
0.0770263671875,
0.47216796875,
0.07763671875,
0.47216796875,
0.0889892578125,
0.0882568359375,
0.08837890625
],
[
0.6748046875,
0.07574462890625,
0.9111328125,
0.0770263671875,
0.9111328125,
0.09027099609375,
0.6748046875,
0.0889892578125
],
[
0.08660888671875,
0.106689453125,
0.91259765625,
0.10797119140625,
0.91259765625,
0.12249755859375,
0.08660888671875,
0.121826171875
],
[
0.0882568359375,
0.12249755859375,
0.91259765625,
0.12249755859375,
0.91259765625,
0.136962890625,
0.0882568359375,
0.136962890625
],
[
0.08905029296875,
0.136962890625,
0.420654296875,
0.136962890625,
0.420654296875,
0.147705078125,
0.08905029296875,
0.147705078125
],
[
0.101318359375,
0.149658203125,
0.91015625,
0.1483154296875,
0.91015625,
0.1634521484375,
0.101318359375,
0.1641845703125
],
[
0.08660888671875,
0.1641845703125,
0.91259765625,
0.1641845703125,
0.91259765625,
0.177978515625,
0.08660888671875,
0.177978515625
],
[
0.0882568359375,
0.1787109375,
0.27783203125,
0.1787109375,
0.27783203125,
0.189453125,
0.0882568359375,
0.189453125
],
[
0.0849609375,
0.203857421875,
0.1953125,
0.203857421875,
0.1953125,
0.2171630859375,
0.0849609375,
0.2171630859375
],
[
0.0882568359375,
0.2235107421875,
0.91015625,
0.2235107421875,
0.91015625,
0.2342529296875,
0.0882568359375,
0.2342529296875
],
[
0.08905029296875,
0.2374267578125,
0.91015625,
0.2374267578125,
0.91015625,
0.2474365234375,
0.08905029296875,
0.2474365234375
],
[
0.08660888671875,
0.248779296875,
0.9111328125,
0.25,
0.9111328125,
0.26318359375,
0.08660888671875,
0.261962890625
],
[
0.08905029296875,
0.26513671875,
0.91015625,
0.26513671875,
0.91015625,
0.27587890625,
0.08905029296875,
0.27587890625
],
[
0.08660888671875,
0.276611328125,
0.9111328125,
0.27783203125,
0.9111328125,
0.291748046875,
0.08660888671875,
0.290283203125
],
[
0.0882568359375,
0.291748046875,
0.91259765625,
0.291748046875,
0.91259765625,
0.304931640625,
0.0882568359375,
0.304931640625
],
[
0.0882568359375,
0.306884765625,
0.56787109375,
0.306884765625,
0.56787109375,
0.317626953125,
0.0882568359375,
0.317626953125
],
[
0.08905029296875,
0.32763671875,
0.91015625,
0.32763671875,
0.91015625,
0.33837890625,
0.08905029296875,
0.33837890625
],
[
0.0849609375,
0.339111328125,
0.9111328125,
0.34033203125,
0.9111328125,
0.354248046875,
0.0849609375,
0.352783203125
],
[
0.08660888671875,
0.352783203125,
0.9111328125,
0.352783203125,
0.9111328125,
0.3662109375,
0.08660888671875,
0.3662109375
],
[
0.0882568359375,
0.36865234375,
0.9111328125,
0.36865234375,
0.9111328125,
0.3818359375,
0.0882568359375,
0.3818359375
],
[
0.0882568359375,
0.3818359375,
0.16015625,
0.3818359375,
0.16015625,
0.393310546875,
0.0882568359375,
0.393310546875
],
[
0.08905029296875,
0.39453125,
0.91259765625,
0.39453125,
0.91259765625,
0.40771484375,
0.08905029296875,
0.40771484375
],
[
0.08660888671875,
0.4072265625,
0.654296875,
0.4091796875,
0.654296875,
0.423583984375,
0.08660888671875,
0.421630859375
],
[
0.646484375,
0.410400390625,
0.9111328125,
0.4091796875,
0.9111328125,
0.42041015625,
0.646484375,
0.421630859375
],
[
0.0882568359375,
0.42236328125,
0.9140625,
0.42236328125,
0.9140625,
0.436279296875,
0.0882568359375,
0.436279296875
],
[
0.0882568359375,
0.4375,
0.9140625,
0.4375,
0.9140625,
0.45068359375,
0.0882568359375,
0.45068359375
],
[
0.0882568359375,
0.451904296875,
0.406982421875,
0.451904296875,
0.406982421875,
0.462158203125,
0.0882568359375,
0.462158203125
],
[
0.0906982421875,
0.464111328125,
0.91259765625,
0.464111328125,
0.91259765625,
0.477783203125,
0.0906982421875,
0.477783203125
],
[
0.08905029296875,
0.479736328125,
0.7998046875,
0.479736328125,
0.7998046875,
0.490478515625,
0.08905029296875,
0.490478515625
],
[
0.0906982421875,
0.492431640625,
0.91259765625,
0.492431640625,
0.91259765625,
0.505859375,
0.0906982421875,
0.505859375
],
[
0.0882568359375,
0.505859375,
0.90869140625,
0.505859375,
0.90869140625,
0.51953125,
0.0882568359375,
0.51953125
],
[
0.0882568359375,
0.51953125,
0.91259765625,
0.51953125,
0.91259765625,
0.53271484375,
0.0882568359375,
0.53271484375
],
[
0.0882568359375,
0.5341796875,
0.91259765625,
0.5341796875,
0.91259765625,
0.54736328125,
0.0882568359375,
0.54736328125
],
[
0.08905029296875,
0.548828125,
0.689453125,
0.548828125,
0.689453125,
0.5595703125,
0.08905029296875,
0.5595703125
],
[
0.10455322265625,
0.56103515625,
0.301513671875,
0.56103515625,
0.301513671875,
0.57470703125,
0.10455322265625,
0.57470703125
],
[
0.32275390625,
0.58203125,
0.67626953125,
0.58203125,
0.67626953125,
0.5966796875,
0.32275390625,
0.5966796875
],
[
0.89453125,
0.58203125,
0.9140625,
0.58203125,
0.9140625,
0.59521484375,
0.89453125,
0.59521484375
],
[
0.08660888671875,
0.6015625,
0.91259765625,
0.60302734375,
0.91259765625,
0.6162109375,
0.08660888671875,
0.61572265625
],
[
0.0882568359375,
0.6162109375,
0.91259765625,
0.6162109375,
0.91259765625,
0.6298828125,
0.0882568359375,
0.6298828125
],
[
0.0849609375,
0.6298828125,
0.5146484375,
0.630859375,
0.5146484375,
0.64453125,
0.0849609375,
0.64306640625
],
[
0.08905029296875,
0.65283203125,
0.9111328125,
0.65283203125,
0.9111328125,
0.66357421875,
0.08905029296875,
0.66357421875
],
[
0.0882568359375,
0.66650390625,
0.90869140625,
0.66650390625,
0.90869140625,
0.6767578125,
0.0882568359375,
0.6767578125
],
[
0.08660888671875,
0.6787109375,
0.9140625,
0.67822265625,
0.9140625,
0.69140625,
0.08660888671875,
0.6923828125
],
[
0.0882568359375,
0.6943359375,
0.650390625,
0.6943359375,
0.650390625,
0.705078125,
0.0882568359375,
0.705078125
],
[
0.10455322265625,
0.70556640625,
0.9111328125,
0.70556640625,
0.9111328125,
0.7197265625,
0.10455322265625,
0.7197265625
],
[
0.0849609375,
0.7197265625,
0.91259765625,
0.72021484375,
0.91259765625,
0.734375,
0.0849609375,
0.73291015625
],
[
0.0882568359375,
0.736328125,
0.351318359375,
0.736328125,
0.351318359375,
0.7470703125,
0.0882568359375,
0.7470703125
],
[
0.4345703125,
0.74853515625,
0.56396484375,
0.75,
0.56396484375,
0.76123046875,
0.4345703125,
0.76025390625
],
[
0.89453125,
0.74853515625,
0.91259765625,
0.74853515625,
0.91259765625,
0.76220703125,
0.89453125,
0.76220703125
],
[
0.0906982421875,
0.76513671875,
0.91259765625,
0.76513671875,
0.91259765625,
0.77880859375,
0.0906982421875,
0.77880859375
],
[
0.08660888671875,
0.77880859375,
0.1748046875,
0.78076171875,
0.1748046875,
0.79248046875,
0.08660888671875,
0.79052734375
],
[
0.10296630859375,
0.79345703125,
0.91259765625,
0.79345703125,
0.91259765625,
0.806640625,
0.10296630859375,
0.806640625
],
[
0.0882568359375,
0.806640625,
0.69677734375,
0.806640625,
0.69677734375,
0.82080078125,
0.0882568359375,
0.82080078125
],
[
0.0882568359375,
0.82763671875,
0.91259765625,
0.82763671875,
0.91259765625,
0.84130859375,
0.0882568359375,
0.84130859375
],
[
0.08660888671875,
0.84033203125,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.08660888671875,
0.85400390625
],
[
0.08905029296875,
0.85693359375,
0.1993408203125,
0.85693359375,
0.1993408203125,
0.8671875,
0.08905029296875,
0.8671875
],
[
0.10455322265625,
0.869140625,
0.91259765625,
0.869140625,
0.91259765625,
0.88330078125,
0.10455322265625,
0.88330078125
],
[
0.08660888671875,
0.8837890625,
0.91259765625,
0.8837890625,
0.91259765625,
0.8984375,
0.08660888671875,
0.8984375
]
] |
[
"A Knowledge-Injected Curriculum Pretraining Framework for Question Answering",
"wwW 24, May 13-17,2024, Singapore, Singapore",
"（每个名称neiEV+）。R分配有类似的名称。在问答对（Q,Y）中，Q=(q1,q2,,qn)V+（qiEV）是自然语言中的问题，而Y",
"是在KG下推断的Q的答案，其形式取决于任务（e。g。，KBQA经常从KG中选择一个实体或属性值，而生成QA则从自然文本",
"或数学表达式等某些词汇中生成形式语言E[16,17]",
"给定知识图KG和问答对（Q,Y），KBQA的目标是训练一个模型M:(KG.Q)一Y，以预测KG下问题Q的答案Y。在本文中",
"我们首先用KG预训练一个语言模型LM，然后在M中使用它来预测答案Y到Q。我们期望LM可以全面地从KG中学习知",
"识，并很好地处理复杂的推理",
"3.2Method",
"我们提出了一个通用的Knowledge-InjectedC过程P再训练框架（KICP）来对LM进行预训练，以进行全面的知识学习和复杂的",
"推理，该框架不仅限于特定的技术，而且可以轻松地与不同的实现一起使用以实现灵活的应用程序。如图E2E（a）所示，KICF",
"由三个关键组件组成，i。e。、知识注人(KI)、知识适应(KA)和课程推理(CR)。具体而言，KI通过将KG三元组转换为句子来构建",
"预训练语料库，从而将KG中的知识完全注入LM中，并将各种生成技术概括为三个关键步骤。为了减少生成语料库与自然语料",
"库之间的差距带来的负面影响，KA修复了原始LM以保持其NLU能力，并为框架配备了可训练的知识适配器，以从生成的语料",
"库中学习知识。为了对具有复杂推理能力的LM进行预训练，CR遵循人类推理的常见模式，构建了几个具有不同难度的推理所需",
"的语料库，并以课程的方式对LM进行了从易到难的训练，以促进模型学习",
"3.2.1KnowledgeInjection.为了克服使用KG作为外部语料库的补充所带来的知识学习不足，我们直接将KG三元组转换为句子",
"作为预训练语料库，以将知识注人LM。此外，在文献E[2,20]中存在几种针对不同要求的有效句子生成技术，并且KG包含需要",
"不同处理的多种形式的数据（例如。g。、ID、文本、数字和日期)。因此，为了将这些详细技术概括为一个通用框架，该框架不限",
"于在各种情况下灵活应用的特定技术，如图E2E（b）所示，我们将句子生成过程抽象为三个关键步骤，i。e。、文本表征、句子",
"构造和掩蔽",
"TextCharacterization.给定从KG采样的一个三元组k=（h,r,t）EZ，KI首先将三元组的所有字段表征为文本（Txt），这些文本用",
"作句子的主干元素以生成。对于存储在ID中的实体和关系，我们映射无意义的ID(e。g。",
"，e1）到一个有意义的名称（JulesVerne）",
"该名称在每次迭代中从关联的名称集中动态采样，以增加语料库多样性。更多的采样策略也可以在这里应用于其他需求[20]。对",
"于各种形式的属性值（e。g、数字、日期和文本)，我们使用它们的文字描述，因为它们总是可以用文字表达，尽管原始形式。这",
"样，我们可以统一实体、关系和属性值的不同处理",
"SentenceConstruction.获得文本元素后，KI应用句子构造策略t将这些元素组装成一个完整的句子，包括对元素进行重新排序和",
"转换以及添加辅助词。策略t可以用不同的现有技术来实现，如句子模板、基于语法的规则和生成的LMs[2，20",
"Masking.最后一步是对生成的句子进行掩码，以进行掩码语言模型（MLM）预训练。为了强制知识学习并匹配实体和属性值之间",
"的差异，我们更喜欢对句子中的知识元素（从三元组转换而来的）赋予更多的权重，并对实体和属性值应用不同的屏蔽策略Msk",
"例如，我们在实体上应用实体掩码[34]】，它掩码整个实体名称以强制学习关系知识，而不是记忆实体名称，并在属性值上应用全",
"字掩码（WWM)[5]，因为这些值可能包含太多信息（例如。g。，传记)，如果都被掩盖，很难恢复。WWM的工作方式也类似于短",
"值上的实体掩码（例如。g。，数字）通过屏蔽整个单词。这里可以使用更多的掩码技术作为Msk",
"总的来说，句子生成过程如下：",
"KI(k) = Msk(t(Txt(h), Txt(r), Txt(t), k = (h, r, t) E Z",
"(1)",
"知识注人的语料库由句子KI(k）组成，这些句子是从预训练中从KG采样的三元组动态生成的。这样，KI将整个KG转换为语料",
"库，从而隐式地将来自KG的所有信息存储在语料库中，例如结构信息。与将KG重写为语料库的现有方法相比，KI不依赖于特",
"定的技术或资源，因此可以针对各种应用程序需求使用不同的实现",
"3.2.2KnowledgeAdaptation，显然，KI生成的语料库与自然语料库不同，因为句子可能不严格遵循语法（特别是对于一些简单的",
"t），并且语料库的多样性受到限制。在语料库上对LM进行预训练可能会损害NLU的能力，并且在自然文本上效果不佳。此外",
"由于拟议的通用框架中的句子生成技术是任意的，因此我们不能像现有研究那样使用与特定生成技术相关的方法来解决该问题回",
"[2,20】。因此，在知识适应（KA）中，我们转向在知识预训练期保持LM的NLU能力",
"如图E2E（c）所示，遵循LM调优国[7,36]中的适配器范例，我们固定LM参数并在原始LMLM上方添加可训练的知识适配",
"器模块Ad。Ad使用LM的语义输出作为输人，并输出知识增强的表示。此外，为了深人改善语义和知识的融合，使用了LM中",
"所有层的语义输出。KA的计算公式如下：",
"KA(x) = Ad(LM(x)),",
"(2)",
"。其中x是输入句子。Ad可以用任何神经网络实现，预计它有一个适当的大小，以包含足够的空间进行知识学习，并避免大大增",
"加计算复杂性",
"在预训练中，对Ad的参数进行了训练，以从构建的语料库中学习知识，而原始LM是固定的。由于原始LM不受Ad的影响，",
"因此尽可能保留了NLU能力，以减少生成的语料库与自然语料库之间的差距所带来的负面影响",
"3.2.3CurriculumReasoning.使用KI和KA，KICP可以有效地将KG注人LM，但仍然缺乏现实世界KBQA任务中所要求的对多",
"个知识事实的复杂推理能力。为了使LM具有这种能力，课程推理模块（CR）在需要复杂推理的语料库上对LM进行预训练，如",
"图E2E（d）所示",
"很难为所有KG收集足够的推理所需的语料库，因此我们也基于KG构建了语料库。人类经常按照特定的模式进行复杂的推理",
"（例如。g。，多顶推理)，这对参与的三元组(e。g。，链状三元组)。因此，我们按照这些模式构建语料库（例如。g。，“《彗星上的起"
] |
{
"arxiv_id": "2403.09712",
"language": "zh",
"page_idx": 2
}
|
||
2403.09712-zh-3
|
2403.09712-zh-3.jpg
|
WWW'24, May 13-17, 2024, Singapore, Singapore
Xin Lin et al
Table 1: Overall Results of All Methods on Four Datasets
<table>
<tbody>
<tr>
<th rowspan="2">Dataset Metric</th>
<th colspan="2">CN- QA</th>
<th colspan="2">ComplexWebOuestions</th>
<th rowspan="2">FreebaseQA $ACC$</th>
<th rowspan="2">Math23K $ACC$</th>
</tr>
<tr>
<th>F1</th>
<th>$EM$</th>
<th>F1</th>
<th>$EM$</th>
</tr>
<tr>
<td>$GPT4$</td>
<td>0.459</td>
<td>0.358</td>
<td>0.802</td>
<td>0.721</td>
<td>0.918</td>
<td>/</td>
</tr>
<tr>
<td>ChatGLM2- 6B</td>
<td>0.389</td>
<td>0.274</td>
<td>0.494</td>
<td>0.432</td>
<td>0.610</td>
<td>/</td>
</tr>
<tr>
<td>EmbedKGQA</td>
<td>0.417</td>
<td>0.303</td>
<td>0.760</td>
<td>0.730</td>
<td>0.707</td>
<td>/</td>
</tr>
<tr>
<td>$BERT$</td>
<td>0.607</td>
<td>0.458</td>
<td>0.856</td>
<td>0.763</td>
<td>0.896</td>
<td>0.801</td>
</tr>
<tr>
<td>$\operatorname{RoBERTa}$</td>
<td>0.610</td>
<td>0.456</td>
<td>0.863</td>
<td>0.779</td>
<td>0.892</td>
<td>0.803</td>
</tr>
<tr>
<td>ERNIE</td>
<td>0.614</td>
<td>0.459</td>
<td>0.861</td>
<td>0.772</td>
<td>0.901</td>
<td>0.796</td>
</tr>
<tr>
<td>K- BERT</td>
<td>0.620</td>
<td>0.462</td>
<td>0.866</td>
<td>0.774</td>
<td>0.896</td>
<td>0.799</td>
</tr>
<tr>
<td>KEPLER</td>
<td>0.628</td>
<td>0.467</td>
<td>0.868</td>
<td>0.785</td>
<td>0.906</td>
<td>/</td>
</tr>
<tr>
<td>K- Adapter</td>
<td>0.612</td>
<td>0.462</td>
<td>0.866</td>
<td>0.802</td>
<td>0.905</td>
<td>/</td>
</tr>
<tr>
<td>KICP- KA</td>
<td>0.633</td>
<td>0.469</td>
<td>0.871</td>
<td>0.809</td>
<td>0.903</td>
<td>0.797</td>
</tr>
<tr>
<td>KICP- ATT</td>
<td>0.629</td>
<td>0.466</td>
<td>1/</td>
<td>1 1</td>
<td>1/</td>
<td>/</td>
</tr>
<tr>
<td>$KICP$</td>
<td>$0.639^{*}$ $\neq$</td>
<td>$0.480^{\star}$</td>
<td>$0.880^{*}$</td>
<td>$0.819^{*}$</td>
<td>0.911</td>
<td>$0.809^{\star}$</td>
</tr>
</tbody>
</table>
(3)
飞》的作者时期是 1828-1905 年”)。我们首先对几个三元组$\{k_1,\ldots,k_n\}$进行采样，并与 KG 的限制相匹配，例如链状三元组{(彗星上的 Off, 作者，JulesVerne) , (JulesVerne, 时期，1828-1905) }进行多跳推理，然后将它们转换为具有类似于 KI 的管道Comp 的复杂组合，其中：<NEXMATHX。这样，复杂的语料库与人类的推理相匹配，并明确地利用了KG 的结构信息。CR 模块可以支持更多的推理模式
复杂的组成通常会丢弃一些信息以从知识中推断出来，因此很难直接对 LM 进行预训练 (e。g。,在前面的示例中，“JulesVerne” 被丢弃，这使得在没有相关知识的情况下很难理解)。因此，如图E2 E( d) 所示，我们将预训练分为三节课，并生成语料库，从易到难跟随课程学习$\operatorname{\overline{E}[51]}$ ,以促进模型学习
Lesson 1: Knowledge Learning. 我们首先对 KG 中的单个三元组进行预训练 LM。我们用 KI 基于每个句子的一个三元组$k$构建这个
语料库，并预先训练LM(i。e。, KA) 在 MLM 任务上记忆知识事实如下：
$$\min_{\theta_{Ad},\:\theta_{MLM}}L_{1}(k)=MLM(KA(KI(k))),$$
其中$\theta_{Ad}$ 和$\theta_{MLM}$ 表示 KA 和 MLM 头中知识适配器 $Ad$ 的可训练参数
Lesson 2: CoT Learning. 从 KG 学习了基本知识事实之后，接下来我们教 LM 如何使用相关知识事实进行复杂的推理。受思想链(CoT)$\overline{\mathbb{E}}[25,40]$的启发，我们通过Comp为某些推理模式组装了具有复杂成分的每个句子，并通过$KI$将所有相关知识作为基于三元组$\{k_1,\ldots,k_n\}$的推理步骤。为了避免信息泄露，我们屏蔽了相同的元素(e。g。,实体)在最终的合成和推理步骤中，并在 MLM 任务上预训练 LM, 如下所示：
$$\begin{aligned}\min_{\theta_{Ad},\:\theta_{MLM}}L_{2}(k_{1},\ldots,k_{n})=&MLM(KA([KI(k_{1}),\ldots,\\&KI(k_{n}),Comp(k_{1},\ldots,k_{n})])),\end{aligned}$$
(4)
其中[,]表示文本连接，而$\{k_1,\ldots,k_n\}$与 Comp 的推理模式匹配
Lesson 3: Composition Learning. 在最困难的课程中，我们预训练 LM 以将记忆的知识作为现实世界的 QA 任务进行推理，在该任务中，我们仅提供最终作品，而没有相关的推理步骤。因此，我们通过 Comp 构建具有复杂成分的语料库，并在 MLM 任务上对LM 进行预训练，如下所示：
(5)
$$\min_{\theta_{Ad},\:\theta_{MLM}}L_{3}(k_{1},\ldots,k_{n})=MLM(KA(Comp(k_{1},\ldots,k_{n}))).$$
在预训练中，使用随机采样的三元组动态生成语料库。我们在附录 D 中的三个课程中演示了一些语料库样本。通过三堂预训练课，
明确使 LM 具备 KBQA 任务中所需的类人复杂推理能力，并以课程学习降低预训练难度
3.2.4 QA Fine-Tuning.在 KG 上预训练后，LM 可以轻松地应用于不同的下游 QA 任务中，而无需额外的注释或外部知识输入。具体来说，LM(i。e。, KA ) 读取问题$Q$ 作为输人，并输出知识增强向量，该向量被馈送到与任务相关的预测头 Pred 以生成答案 Y。整个系统$(KA$和 Pred 中的$LM$和$Ad$ )可以根据任务相关目标函数$\mathcal{L}$对不同的 QA 任务进行微调，如下所示：
(6)
$$\min_{\theta_{LM},\:\theta_{Ad},\:\theta_{Pred}}L_{QA}(Q,Y)=\mathcal{L}(Pred(KA(Q)),Y),$$
其中$\theta_{LM},\theta_{Ad}$和$Q$这些模块的参数是
|
[
[
0.08660888671875,
0.07574462890625,
0.324462890625,
0.07763671875,
0.324462890625,
0.0889892578125,
0.08660888671875,
0.08709716796875
],
[
0.84619140625,
0.0770263671875,
0.90869140625,
0.0770263671875,
0.90869140625,
0.08837890625,
0.84619140625,
0.08837890625
],
[
0.313720703125,
0.106689453125,
0.685546875,
0.106689453125,
0.685546875,
0.117431640625,
0.313720703125,
0.117431640625
],
[
0.2646484375,
0.1383056640625,
0.314453125,
0.1383056640625,
0.314453125,
0.15283203125,
0.2646484375,
0.15283203125
],
[
0.36767578125,
0.1383056640625,
0.418212890625,
0.1383056640625,
0.418212890625,
0.15283203125,
0.36767578125,
0.15283203125
],
[
0.44775390625,
0.136962890625,
0.5986328125,
0.1383056640625,
0.5986328125,
0.15283203125,
0.44775390625,
0.1514892578125
],
[
0.59716796875,
0.1383056640625,
0.685546875,
0.1383056640625,
0.685546875,
0.15283203125,
0.59716796875,
0.15283203125
],
[
0.685546875,
0.1395263671875,
0.75341796875,
0.1395263671875,
0.75341796875,
0.15087890625,
0.685546875,
0.15087890625
],
[
0.26708984375,
0.149658203125,
0.31201171875,
0.1514892578125,
0.310546875,
0.1673583984375,
0.265625,
0.1654052734375
],
[
0.4599609375,
0.1514892578125,
0.48046875,
0.1514892578125,
0.48046875,
0.1666259765625,
0.4599609375,
0.1666259765625
],
[
0.53759765625,
0.149658203125,
0.56201171875,
0.15283203125,
0.5595703125,
0.1673583984375,
0.53515625,
0.1641845703125
],
[
0.70751953125,
0.1514892578125,
0.74267578125,
0.1514892578125,
0.74267578125,
0.1673583984375,
0.70751953125,
0.1673583984375
],
[
0.40185546875,
0.15283203125,
0.431396484375,
0.15283203125,
0.431396484375,
0.1654052734375,
0.40185546875,
0.1654052734375
],
[
0.35693359375,
0.15283203125,
0.37744140625,
0.15283203125,
0.37744140625,
0.1666259765625,
0.35693359375,
0.1666259765625
],
[
0.6298828125,
0.15283203125,
0.66259765625,
0.15283203125,
0.66259765625,
0.1654052734375,
0.6298828125,
0.1654052734375
],
[
0.346435546875,
0.1717529296875,
0.38818359375,
0.1717529296875,
0.38818359375,
0.186279296875,
0.346435546875,
0.186279296875
],
[
0.398681640625,
0.1717529296875,
0.439453125,
0.1717529296875,
0.439453125,
0.186279296875,
0.398681640625,
0.186279296875
],
[
0.53125,
0.1717529296875,
0.56787109375,
0.1717529296875,
0.56787109375,
0.1875,
0.53125,
0.1875
],
[
0.451904296875,
0.17236328125,
0.490234375,
0.17236328125,
0.490234375,
0.186279296875,
0.451904296875,
0.186279296875
],
[
0.27294921875,
0.17236328125,
0.30810546875,
0.17236328125,
0.30810546875,
0.184326171875,
0.27294921875,
0.184326171875
],
[
0.62841796875,
0.17236328125,
0.66259765625,
0.17236328125,
0.66259765625,
0.1849365234375,
0.62841796875,
0.1849365234375
],
[
0.7197265625,
0.17236328125,
0.73046875,
0.17236328125,
0.73046875,
0.1849365234375,
0.7197265625,
0.1849365234375
],
[
0.346435546875,
0.1849365234375,
0.386474609375,
0.1849365234375,
0.386474609375,
0.199462890625,
0.346435546875,
0.199462890625
],
[
0.53125,
0.1849365234375,
0.56884765625,
0.1849365234375,
0.56884765625,
0.2008056640625,
0.53125,
0.2008056640625
],
[
0.40185546875,
0.186279296875,
0.439453125,
0.186279296875,
0.439453125,
0.199462890625,
0.40185546875,
0.199462890625
],
[
0.2459716796875,
0.186279296875,
0.3349609375,
0.186279296875,
0.3349609375,
0.1976318359375,
0.2459716796875,
0.1976318359375
],
[
0.453369140625,
0.186279296875,
0.488525390625,
0.186279296875,
0.488525390625,
0.1988525390625,
0.453369140625,
0.1988525390625
],
[
0.62841796875,
0.186279296875,
0.66259765625,
0.186279296875,
0.66259765625,
0.1988525390625,
0.62841796875,
0.1988525390625
],
[
0.7197265625,
0.186279296875,
0.7294921875,
0.186279296875,
0.7294921875,
0.1976318359375,
0.7197265625,
0.1976318359375
],
[
0.2459716796875,
0.20263671875,
0.33251953125,
0.2052001953125,
0.33251953125,
0.2197265625,
0.2459716796875,
0.2171630859375
],
[
0.34228515625,
0.2052001953125,
0.3857421875,
0.2052001953125,
0.3857421875,
0.2197265625,
0.34228515625,
0.2197265625
],
[
0.400390625,
0.2052001953125,
0.44287109375,
0.2052001953125,
0.44287109375,
0.2203369140625,
0.400390625,
0.2203369140625
],
[
0.449462890625,
0.2052001953125,
0.490234375,
0.2052001953125,
0.490234375,
0.2197265625,
0.449462890625,
0.2197265625
],
[
0.53173828125,
0.205810546875,
0.56787109375,
0.205810546875,
0.56787109375,
0.2183837890625,
0.53173828125,
0.2183837890625
],
[
0.62841796875,
0.205810546875,
0.66259765625,
0.205810546875,
0.66259765625,
0.2183837890625,
0.62841796875,
0.2183837890625
],
[
0.7197265625,
0.205810546875,
0.73046875,
0.205810546875,
0.73046875,
0.2171630859375,
0.7197265625,
0.2171630859375
],
[
0.438720703125,
0.210205078125,
0.44287109375,
0.210205078125,
0.44287109375,
0.21533203125,
0.438720703125,
0.21533203125
],
[
0.3349609375,
0.2108154296875,
0.34228515625,
0.2108154296875,
0.34228515625,
0.2139892578125,
0.3349609375,
0.2139892578125
],
[
0.27294921875,
0.2247314453125,
0.30810546875,
0.2247314453125,
0.30810546875,
0.2374267578125,
0.27294921875,
0.2374267578125
],
[
0.3505859375,
0.2259521484375,
0.384033203125,
0.2259521484375,
0.384033203125,
0.238037109375,
0.3505859375,
0.238037109375
],
[
0.40185546875,
0.2259521484375,
0.439453125,
0.2259521484375,
0.439453125,
0.2392578125,
0.40185546875,
0.2392578125
],
[
0.451904296875,
0.2259521484375,
0.490234375,
0.2259521484375,
0.490234375,
0.2392578125,
0.451904296875,
0.2392578125
],
[
0.53173828125,
0.2259521484375,
0.56787109375,
0.2259521484375,
0.56787109375,
0.238037109375,
0.53173828125,
0.238037109375
],
[
0.62841796875,
0.2259521484375,
0.66259765625,
0.2259521484375,
0.66259765625,
0.238037109375,
0.62841796875,
0.238037109375
],
[
0.7060546875,
0.2235107421875,
0.74267578125,
0.2235107421875,
0.74267578125,
0.2392578125,
0.7060546875,
0.2392578125
],
[
0.259033203125,
0.238037109375,
0.3203125,
0.238037109375,
0.3203125,
0.25244140625,
0.259033203125,
0.25244140625
],
[
0.400390625,
0.2374267578125,
0.439453125,
0.2392578125,
0.438720703125,
0.255126953125,
0.398681640625,
0.25244140625
],
[
0.7060546875,
0.238037109375,
0.744140625,
0.238037109375,
0.744140625,
0.25390625,
0.7060546875,
0.25390625
],
[
0.453369140625,
0.2392578125,
0.488525390625,
0.2392578125,
0.488525390625,
0.251953125,
0.453369140625,
0.251953125
],
[
0.3505859375,
0.2392578125,
0.3857421875,
0.2392578125,
0.3857421875,
0.251953125,
0.3505859375,
0.251953125
],
[
0.53173828125,
0.2392578125,
0.56787109375,
0.2392578125,
0.56787109375,
0.251953125,
0.53173828125,
0.251953125
],
[
0.62841796875,
0.2392578125,
0.66259765625,
0.2392578125,
0.66259765625,
0.251953125,
0.62841796875,
0.251953125
],
[
0.26708984375,
0.2568359375,
0.31201171875,
0.2568359375,
0.31201171875,
0.271484375,
0.26708984375,
0.271484375
],
[
0.346435546875,
0.2568359375,
0.386474609375,
0.2568359375,
0.386474609375,
0.271484375,
0.346435546875,
0.271484375
],
[
0.40185546875,
0.25830078125,
0.439453125,
0.25830078125,
0.439453125,
0.271484375,
0.40185546875,
0.271484375
],
[
0.451904296875,
0.25830078125,
0.490234375,
0.25830078125,
0.490234375,
0.271484375,
0.451904296875,
0.271484375
],
[
0.53173828125,
0.25830078125,
0.56787109375,
0.25830078125,
0.56787109375,
0.270751953125,
0.53173828125,
0.270751953125
],
[
0.62744140625,
0.25830078125,
0.66259765625,
0.25830078125,
0.66259765625,
0.270751953125,
0.62744140625,
0.270751953125
],
[
0.70751953125,
0.25830078125,
0.74267578125,
0.25830078125,
0.74267578125,
0.270751953125,
0.70751953125,
0.270751953125
],
[
0.26318359375,
0.270751953125,
0.316162109375,
0.270751953125,
0.316162109375,
0.285400390625,
0.26318359375,
0.285400390625
],
[
0.3505859375,
0.272705078125,
0.3857421875,
0.272705078125,
0.3857421875,
0.285400390625,
0.3505859375,
0.285400390625
],
[
0.40185546875,
0.272705078125,
0.439453125,
0.272705078125,
0.439453125,
0.285888671875,
0.40185546875,
0.285888671875
],
[
0.453369140625,
0.272705078125,
0.488525390625,
0.272705078125,
0.488525390625,
0.285400390625,
0.453369140625,
0.285400390625
],
[
0.53173828125,
0.272705078125,
0.56640625,
0.272705078125,
0.56640625,
0.285400390625,
0.53173828125,
0.285400390625
],
[
0.62841796875,
0.272705078125,
0.66259765625,
0.272705078125,
0.66259765625,
0.285400390625,
0.62841796875,
0.285400390625
],
[
0.7060546875,
0.272705078125,
0.74267578125,
0.272705078125,
0.74267578125,
0.2841796875,
0.7060546875,
0.2841796875
],
[
0.398681640625,
0.285400390625,
0.439453125,
0.285400390625,
0.439453125,
0.2998046875,
0.398681640625,
0.2998046875
],
[
0.53125,
0.285400390625,
0.56884765625,
0.285400390625,
0.56884765625,
0.300537109375,
0.53125,
0.300537109375
],
[
0.3505859375,
0.285888671875,
0.3857421875,
0.285888671875,
0.3857421875,
0.298583984375,
0.3505859375,
0.298583984375
],
[
0.26318359375,
0.285888671875,
0.316162109375,
0.285888671875,
0.316162109375,
0.29736328125,
0.26318359375,
0.29736328125
],
[
0.453369140625,
0.285888671875,
0.488525390625,
0.285888671875,
0.488525390625,
0.298583984375,
0.453369140625,
0.298583984375
],
[
0.62841796875,
0.285888671875,
0.66259765625,
0.285888671875,
0.66259765625,
0.298583984375,
0.62841796875,
0.298583984375
],
[
0.71826171875,
0.287353515625,
0.73046875,
0.287353515625,
0.73046875,
0.2998046875,
0.71826171875,
0.2998046875
],
[
0.34716796875,
0.298583984375,
0.386474609375,
0.298583984375,
0.386474609375,
0.314453125,
0.34716796875,
0.314453125
],
[
0.450927734375,
0.298583984375,
0.490234375,
0.298583984375,
0.490234375,
0.314453125,
0.450927734375,
0.314453125
],
[
0.53125,
0.298583984375,
0.56787109375,
0.298583984375,
0.56787109375,
0.314453125,
0.53125,
0.314453125
],
[
0.6259765625,
0.298583984375,
0.6640625,
0.298583984375,
0.6640625,
0.314453125,
0.6259765625,
0.314453125
],
[
0.398681640625,
0.2998046875,
0.439453125,
0.2998046875,
0.439453125,
0.314453125,
0.398681640625,
0.314453125
],
[
0.256591796875,
0.300537109375,
0.32275390625,
0.300537109375,
0.32275390625,
0.3125,
0.256591796875,
0.3125
],
[
0.7197265625,
0.300537109375,
0.73046875,
0.300537109375,
0.73046875,
0.313232421875,
0.7197265625,
0.313232421875
],
[
0.259033203125,
0.318115234375,
0.3203125,
0.318115234375,
0.3203125,
0.333251953125,
0.259033203125,
0.333251953125
],
[
0.398681640625,
0.318115234375,
0.439453125,
0.318115234375,
0.439453125,
0.333251953125,
0.398681640625,
0.333251953125
],
[
0.7060546875,
0.317626953125,
0.74267578125,
0.317626953125,
0.74267578125,
0.333251953125,
0.7060546875,
0.333251953125
],
[
0.451904296875,
0.3193359375,
0.48779296875,
0.3193359375,
0.48779296875,
0.33203125,
0.451904296875,
0.33203125
],
[
0.348876953125,
0.3193359375,
0.384033203125,
0.3193359375,
0.384033203125,
0.33203125,
0.348876953125,
0.33203125
],
[
0.53173828125,
0.3193359375,
0.56787109375,
0.3193359375,
0.56787109375,
0.33203125,
0.53173828125,
0.33203125
],
[
0.62744140625,
0.3193359375,
0.66259765625,
0.3193359375,
0.66259765625,
0.33203125,
0.62744140625,
0.33203125
],
[
0.256591796875,
0.33203125,
0.324462890625,
0.33203125,
0.324462890625,
0.3466796875,
0.256591796875,
0.3466796875
],
[
0.3505859375,
0.333251953125,
0.3857421875,
0.333251953125,
0.3857421875,
0.34521484375,
0.3505859375,
0.34521484375
],
[
0.398681640625,
0.33203125,
0.439453125,
0.33203125,
0.439453125,
0.347900390625,
0.398681640625,
0.347900390625
],
[
0.5400390625,
0.333251953125,
0.5556640625,
0.33203125,
0.55615234375,
0.3447265625,
0.54150390625,
0.34521484375
],
[
0.63818359375,
0.333984375,
0.65185546875,
0.333984375,
0.65185546875,
0.3447265625,
0.63818359375,
0.3447265625
],
[
0.7197265625,
0.333984375,
0.7294921875,
0.333984375,
0.7294921875,
0.34521484375,
0.7197265625,
0.34521484375
],
[
0.343017578125,
0.3466796875,
0.38818359375,
0.3466796875,
0.38818359375,
0.361083984375,
0.343017578125,
0.361083984375
],
[
0.4697265625,
0.335205078125,
0.4755859375,
0.3359375,
0.47216796875,
0.3447265625,
0.465576171875,
0.343505859375
],
[
0.39794921875,
0.3466796875,
0.441162109375,
0.3466796875,
0.441162109375,
0.361083984375,
0.39794921875,
0.361083984375
],
[
0.52783203125,
0.3466796875,
0.5703125,
0.3466796875,
0.5703125,
0.361083984375,
0.52783203125,
0.361083984375
],
[
0.625,
0.3466796875,
0.66650390625,
0.3466796875,
0.66650390625,
0.361083984375,
0.625,
0.361083984375
],
[
0.705078125,
0.3466796875,
0.74609375,
0.3466796875,
0.74609375,
0.361083984375,
0.705078125,
0.361083984375
],
[
0.449462890625,
0.347900390625,
0.490234375,
0.347900390625,
0.490234375,
0.359130859375,
0.449462890625,
0.359130859375
],
[
0.27294921875,
0.347900390625,
0.306396484375,
0.347900390625,
0.306396484375,
0.35986328125,
0.27294921875,
0.35986328125
],
[
0.08905029296875,
0.391357421875,
0.9111328125,
0.391357421875,
0.9111328125,
0.406005859375,
0.08905029296875,
0.406005859375
],
[
0.08660888671875,
0.40478515625,
0.91259765625,
0.4072265625,
0.91259765625,
0.421630859375,
0.08660888671875,
0.419189453125
],
[
0.08660888671875,
0.42041015625,
0.91259765625,
0.42041015625,
0.91259765625,
0.434326171875,
0.08660888671875,
0.434326171875
],
[
0.08660888671875,
0.434326171875,
0.2059326171875,
0.434326171875,
0.2059326171875,
0.447509765625,
0.08660888671875,
0.447509765625
],
[
0.10296630859375,
0.448974609375,
0.91259765625,
0.448974609375,
0.91259765625,
0.462158203125,
0.10296630859375,
0.462158203125
],
[
0.08660888671875,
0.462158203125,
0.91259765625,
0.462158203125,
0.91259765625,
0.47607421875,
0.08660888671875,
0.47607421875
],
[
0.08905029296875,
0.477783203125,
0.378173828125,
0.477783203125,
0.378173828125,
0.488037109375,
0.08905029296875,
0.488037109375
],
[
0.08905029296875,
0.490478515625,
0.91259765625,
0.490478515625,
0.91259765625,
0.50390625,
0.08905029296875,
0.50390625
],
[
0.08660888671875,
0.50390625,
0.55224609375,
0.50390625,
0.55224609375,
0.517578125,
0.08660888671875,
0.517578125
],
[
0.40283203125,
0.5263671875,
0.60888671875,
0.5263671875,
0.60888671875,
0.54052734375,
0.40283203125,
0.54052734375
],
[
0.89453125,
0.5263671875,
0.9140625,
0.5263671875,
0.9140625,
0.54052734375,
0.89453125,
0.54052734375
],
[
0.3896484375,
0.5361328125,
0.4453125,
0.53857421875,
0.4453125,
0.5498046875,
0.3896484375,
0.54736328125
],
[
0.08660888671875,
0.55615234375,
0.529296875,
0.55615234375,
0.529296875,
0.5693359375,
0.08660888671875,
0.5693359375
],
[
0.08905029296875,
0.57080078125,
0.91259765625,
0.57080078125,
0.91259765625,
0.583984375,
0.08905029296875,
0.583984375
],
[
0.08660888671875,
0.583984375,
0.9111328125,
0.583984375,
0.9111328125,
0.59765625,
0.08660888671875,
0.59765625
],
[
0.0882568359375,
0.5986328125,
0.9140625,
0.5986328125,
0.9140625,
0.6123046875,
0.0882568359375,
0.6123046875
],
[
0.08660888671875,
0.6123046875,
0.281005859375,
0.6123046875,
0.281005859375,
0.62548828125,
0.08660888671875,
0.62548828125
],
[
0.35302734375,
0.63720703125,
0.378173828125,
0.63525390625,
0.3798828125,
0.64306640625,
0.3544921875,
0.64599609375
],
[
0.374267578125,
0.63330078125,
0.6259765625,
0.63330078125,
0.6259765625,
0.646484375,
0.374267578125,
0.646484375
],
[
0.340576171875,
0.642578125,
0.393798828125,
0.642578125,
0.393798828125,
0.65380859375,
0.340576171875,
0.65380859375
],
[
0.8935546875,
0.64599609375,
0.9140625,
0.64599609375,
0.9140625,
0.66015625,
0.8935546875,
0.66015625
],
[
0.48779296875,
0.65478515625,
0.65673828125,
0.65478515625,
0.65673828125,
0.66845703125,
0.48779296875,
0.66845703125
],
[
0.0882568359375,
0.68115234375,
0.495849609375,
0.68115234375,
0.495849609375,
0.69140625,
0.0882568359375,
0.69140625
],
[
0.09234619140625,
0.6943359375,
0.9111328125,
0.6943359375,
0.9111328125,
0.705078125,
0.09234619140625,
0.705078125
],
[
0.08660888671875,
0.70703125,
0.91259765625,
0.70703125,
0.91259765625,
0.72021484375,
0.08660888671875,
0.72021484375
],
[
0.0849609375,
0.7216796875,
0.268798828125,
0.7216796875,
0.268798828125,
0.73486328125,
0.0849609375,
0.73486328125
],
[
0.34228515625,
0.73779296875,
0.6650390625,
0.73779296875,
0.6650390625,
0.751953125,
0.34228515625,
0.751953125
],
[
0.89453125,
0.7392578125,
0.9140625,
0.7392578125,
0.9140625,
0.75244140625,
0.89453125,
0.75244140625
],
[
0.330078125,
0.74755859375,
0.384033203125,
0.75,
0.384033203125,
0.76123046875,
0.330078125,
0.7587890625
],
[
0.08660888671875,
0.7666015625,
0.9140625,
0.7666015625,
0.9140625,
0.77978515625,
0.08660888671875,
0.77978515625
],
[
0.0882568359375,
0.7822265625,
0.63623046875,
0.7822265625,
0.63623046875,
0.79248046875,
0.0882568359375,
0.79248046875
],
[
0.0882568359375,
0.8046875,
0.13232421875,
0.8046875,
0.13232421875,
0.81494140625,
0.0882568359375,
0.81494140625
],
[
0.12744140625,
0.8046875,
0.9111328125,
0.8046875,
0.9111328125,
0.81494140625,
0.12744140625,
0.81494140625
],
[
0.0882568359375,
0.8173828125,
0.9140625,
0.8173828125,
0.9140625,
0.83056640625,
0.0882568359375,
0.83056640625
],
[
0.0882568359375,
0.83203125,
0.79150390625,
0.83203125,
0.79150390625,
0.84228515625,
0.0882568359375,
0.84228515625
],
[
0.43310546875,
0.85400390625,
0.63623046875,
0.85400390625,
0.63623046875,
0.8671875,
0.43310546875,
0.8671875
],
[
0.89453125,
0.85400390625,
0.9140625,
0.85400390625,
0.9140625,
0.8671875,
0.89453125,
0.8671875
],
[
0.361083984375,
0.8623046875,
0.435546875,
0.8642578125,
0.435546875,
0.87548828125,
0.361083984375,
0.87353515625
],
[
0.38818359375,
0.85693359375,
0.414306640625,
0.85693359375,
0.414306640625,
0.8642578125,
0.38818359375,
0.8642578125
],
[
0.08660888671875,
0.8837890625,
0.339111328125,
0.8837890625,
0.339111328125,
0.89794921875,
0.08660888671875,
0.89794921875
]
] |
[
"www°24,May13-17,2024,Singapore,Singapore",
"Xin Lin et al.",
"Table1:OverallResultsofAllMethodsonFourDatasets",
"Dataset",
"CN-QA",
"ComplexWebQuestions",
"FreebaseQA",
"Math23K",
"Metric",
"F1",
"EM",
"ACC",
"EM",
"F1",
"ACC",
"0.459",
"0.358",
"0.721",
"0.802",
"GPT4",
"0.918",
"I",
"0.389",
"0.432",
"0.274",
"ChatGLM2-6B",
"0.494",
"0.610",
"1",
"EmbedKGQA",
"0.417",
"0.303",
"0.760",
"0.730",
"0.707",
"/",
"",
"",
"BERT",
"0.607",
"0.458",
"0.856",
"0.763",
"0.896",
"0.801",
"RoBERTa",
"0.456",
"0.803",
"0.863",
"0.610",
"0.779",
"0.892",
"ERNIE",
"0.614",
"0.459",
"0.861",
"0.772",
"0.901",
"0.796",
"K-BERT",
"0.620",
"0.462",
"0.866",
"0.774",
"0.896",
"0.799",
"0.467",
"0.785",
"0.628",
"KEPLER",
"0.868",
"0.906",
"/",
"0.612",
"0.866",
"0.802",
"0.905",
"0.462",
"K-Adapter",
"",
"KICP-KA",
"0.469",
"0.797",
"0.871",
"0.633",
"0.809",
"0.903",
"KICP-ATT",
"0.629",
"0.466",
"/",
"/",
"/",
"0.639*",
"e",
"0.480*",
"0.819*",
"0.911*",
"0.809*",
"0.880*",
"KICP",
"飞》的作者时期是1828-1905年\"）。我们首先对几个三元组（k1，..\"，kn）进行采样，并与KG的限制相匹配，例如链状三元组（（彗",
"星上的Off，作者，JulesVerne），（JulesVerne，时期，1828-1905）进行多跳推理，然后将它们转换为具有类似于KI的管道Comp",
"的复杂组合，其中：<NEXMATHX。这样，复杂的语料库与人类的推理相匹配，并明确地利用了KG的结构信息。CR模块可以支",
"持更多的推理模式",
"复杂的组成通常会丢弃一些信息以从知识中推断出来，因此很难直接对LM进行预训练（e。g。，在前面的示例中，“JulesVerne”",
"被丢弃，这使得在没有相关知识的情况下很难理解)。因此，如图国2国（d）所示，我们将预训练分为三节课，并生成语料库，从",
"易到难跟随课程学习E[51]，以促进模型学习",
"Lesson1:KnowledgeLearning.我们首先对KG中的单个三元组进行预训练LM。我们用KI基于每个句子的一个三元组k构建这个",
"语料库，并预先训练LM(i。e。，KA）在MLM任务上记忆知识事实如下：",
"min Li(k) = MLM(KA(KI(k))),",
"(3)",
"OAd, OMLM",
"其中Ad和QMLM表示KA和MLM头中知识适配器Ad的可训练参数",
"Lesson2:CoTLearning．从KG学习了基本知识事实之后，接下来我们教LM如何使用相关知识事实进行复杂的推理。受思想链",
"(CoT)E[25,40]的启发，我们通过Comp为某些推理模式组装了具有复杂成分的每个句子，并通过KI将所有相关知识作为基于三",
"元组(ki,...，kn)的推理步骤。为了避免信息泄露，我们屏蔽了相同的元素（e。g。，实体）在最终的合成和推理步骤中，并在MLM",
"任务上预训练LM，如下所示：",
"min",
"1L2(k1,...,kn) =MLM(KA([KI(k1),...,",
"0Ad,OMLM",
"(4)",
"KI(kn), Comp(k1,...,kn))D))",
"其中「，1表示文本连接，而（ki....kn与Comp的推理模式匹配",
"Lesson3：CompositionLearning：在最困难的课程中，我们预训练LM以将记忆的知识作为现实世界的QA任务进行推理，在该任",
"务中，我们仅提供最终作品，而没有相关的推理步骤。因此，我们通过Comp构建具有复杂成分的语料库，并在MLM任务上对",
"LM进行预训练，如下所示：",
"minL(ki....,kn) = MLM(KA(Comp(ki....,kn))",
"(5)",
"0Ad,OMLM",
"在预训练中，使用随机采样的三元组动态生成语料库。我们在附录D中的三个课程中演示了一些语料库样本。通过三堂预训练课，",
"明确使LM具备KBQA任务中所需的类人复杂推理能力，并以课程学习降低预训练难度",
"3.2.4",
"QAFine-Tuning，在KG上预训练后，LM可以轻松地应用于不同的下游QA任务中，而无需额外的注释或外部知识输入。具",
"体来说，LM(i。e。，KA）读取问题Q作为输人，并输出知识增强向量，该向量被馈送到与任务相关的预测头Pred以生成答案Y。",
"整个系统（KA和Pred中的LM和Ad）可以根据任务相关目标函数L对不同的QA任务进行微调，如下所示：",
"LQA(Q,Y) = L(Pred(KA(Q),Y),",
"(6)",
"OLM,OAd,Opred",
"min",
"其中QLM，QAd和Q这些模块的参数是"
] |
{
"arxiv_id": "2403.09712",
"language": "zh",
"page_idx": 3
}
|
||
2403.09712-zh-4
|
2403.09712-zh-4.jpg
|
WWW'24,May 13-17, 2024, Singapore, Singapore
A Knowledge-Injected Curriculum Pretraining Framework for Question Answering
# 3.3 $\textbf{Implementation}$
在本节中，我们提供了通用 KICP 框架的实现。在 KI 中，我们实现了$\overline{\mathrm{E}}$3.2.1 节中提到的文本表征和掩码，并通过简单地连接所有
字段来实现$\tau$,这在我们的数据集上效果很好
在 KA 中，我们以相同的层数和减半的向量维度将知识适配器$Ad$实现为 BERT。在$Ad$的每一层中，输人 (来自$LM$相应层的语义向量)首先用线性模型投影到来自最后一层的隐藏向量的潜在空间，然后与隐藏向量相加以馈送到 BERT 层。$Ad$和 LM 的最终向量与线性层合并作为输出
在CR中，我们使用两种广泛使用的推理模式实现 Comp ,即。e。、多跳推理和多对象推理。多跳推理(e。g。,作者 OffonaComet 的时期是 1828-1905) 首先从题目中的主题实体推断一个中间实体(作者 OffonaComet 是 JulesVerne),然后用它来推断下一个中间实体，直到得出答案(JulesVerne 的时期是 1828-1905)。因此，知识三元组形成链状结构，其中一个三元组的尾部实体是下一个三元组的头部 (e。g。, 儒勒·凡尔纳)。给定这些三元组，Comp 丢弃所有中间实体并按顺序连接其他字段。多对象推理(e。g。, 儒勒凡尔纳的职业是小说家和剧作家)从一个主题实体推断出几个结果，因此知识三元组共享相同的头部实体和关系(儒勒·凡尔纳和职业)。给定三元组，Comp丢弃头和关系期望第一个，并将所有尾巴与第一个头和关系连接起来。此外，我们的框架还可以通过定义三元组的采样限制和Comp方法，以类似的方式轻松推广到其他推理模式，例如比较推理。对于每个句子，我们采样 2 到 3个匹配模式的三元组
# 4 EXPERIMENTS
# 4.1 Experimental Setup
, KEPLER[E[37] 和 K-Adapter[E[36] 。我们还包括基于嵌人的嵌人 KGQA [33] 和两个 LLMGPT4 [1] 和 ChatGLM2 [8] 作为 KBQA 数
据集的基线。我们在附录 B 中简要介绍了基线
bQuestions[E[35] 和 FreebaseQA[E[13] (均带有 Wikidata$\mathbb{E}[37]$ ),以及生成数据集 Math23K[E[38] (带有 HowNet[E[32] ),以推广到其他与知识相关的 QA。数据集的介绍和统计信息见附录 A
KBQA 使用 KG 中的实体或属性值回答问题。为了在不损失太多困难的情况下降低计算复杂性，我们在 KBQA上采样了 10 个具
有相同类型真值的硬候选答案进行预测。我们还从每个数据集的整个 KG 中采样一个子图，以加速预训练
4.1.2 $Baseline Methods. 我 们 将 KICP 与 原 始 的 LMsBERTE[ 6] 和 RoBERTa\boxed{\Gamma }[ 22] 以 及 知 识 增 强 的 LMsERNIE\boxed{\Gamma }[ 50] , K- BERT\boxed{\Gamma }[ 21]$ 4.1.3 Training Details. 我们通过 Huggingface 基于预训练的 BERT 使用 Pytorch 实现 KICP。$^{1}$对于所有方法，我们在中文数据集CN-QA 和 Math23K 上使用“bert-base-chinese” 版本作为$LM$ ,在英文数据集 ComplexWebQuestions 和 FreebaseQA 上使用“bertbase-uncased"。KA 的$Ad$的 BERT 层数为 12 (等于$LM$ ),隐藏向量的维数为 384 ($LM$的一半),输出向量的维数为768 (等于$LM$
我们使用 AdamW$\overline {\mathbb{E} }$[ 24] 预训练 3 个时期的模型。批量大小设置为 32, 学习率为0。0005,它在前 10 个% 步骤中升温，然后线性
衰减。MLM 的屏蔽概率设置为 0。第 1 课和第 3 课中的 15, 以及 0。3 在第 2 课中，因为语料库包含更多重复的信息
我们在具有两个 2 的 Linux 服务器上运行所有实验。20GHzIntelXeonE5-2650CPU 和 TeslaK80GPU。$^{2}$
# 4.2 Experimental Results
4.2.1 $Overall Results. 在 本 节 中 , 我 们 将 KICP 与 所 有 基 线 进 行 比 较 。我 们 使 用 F1 分 数 ( F1) 和 精 确 匹 配 分 数 ( EM) 作 为 多 标$ 签数据集 CN-QA 和 ComplexWebQuestions 的指标，以及单标签数据集 FreebaseQA 的准确性 (ACC)。Math23K 具有答案准确性(ACC),即。e。,如果计算出的答案等于真值，则预测表达式被视为正确
表 1 中报告了四个数据集的结果。$^{3}$我们用配对 t 检验对 KICP 相对于基线 (GPT4 除外)的改善进行了统计检验，发现$p<0.05\left(\text{标}\right)$ 记为“)的改善是显著的。我们可以得到以下观察结果。首先，KICP 优于所有基线，这清楚地表明了其在 QA 任务的知识学习和利用方面的有效性。其次，在模型相似但不同的预训练任务下，KICP 的性能优于 K-Adapter, 显示出预训练任务的显着影响。第三， LLM 的性能并不比 KBQA 上的微调方法更好。GPT4 在广泛研究的 ComplexWebQuestions 和 FreebaseQA 上实现了相当的性能，但在 CN-QA 上远远落后，而较小的 ChatGLM2 的性能甚至更差。第四，知识增强的方法在大多数情况下优于原始的 LMs,证明了知识是 QA 推理的关键要素。最后，知识注人并没有给 Math23K 带来太大的改善甚至负面影响。原因可能是 Math23K 需要 NLU 远远超过知识
4.2.2 $AblationStudy.$ 在本节中，我们进行消融实验以研究属性知识和知识适应的有效性(我们将在$\mathbb{E}$4.3 节中详细研究课程推理)。我们介绍了KICP 的两个变体：KICP-KA 删除了知识适应模块并直接训练原始 LM 的参数，而 KICP-ATT 丢弃了属性知识并仅在实体关系知识上进行预训练。两种变体的结果也报告在表 1 中。$^{4}$我们可以总结以下结论。首先，这两个变体的表现要比 KICP 差，这表明 KA 可以减少生成的语料厍的负面影响，并且属性知识在 KBQA 中也很有用。接下来，在 CN-QA 中，KICP-ATT 的表现要比 KICP-KA 差，这意味着属性知识开发在此任务上的贡献要大于知识适应。结果是合理的，因为 CN-QA 的很大一部分需要
$^1$https://huggingface.co/transformers
$^{2}$Our codes are available at https://github.com/l-xin/KICP.
$^{3}$We do not evaluate KEPLER and K-Adapter on Math23K, as pretraining the two methods requires entity descriptions, which are unavailable on HowNet.$^{4}$The results of KICP-ATT on ComplexWebQuestions, FreebaseQA and Math23K are unavailable, as Wikid
|
[
[
0.08660888671875,
0.07574462890625,
0.47216796875,
0.0770263671875,
0.47216796875,
0.09027099609375,
0.08660888671875,
0.0889892578125
],
[
0.6748046875,
0.0770263671875,
0.91015625,
0.07763671875,
0.91015625,
0.0889892578125,
0.6748046875,
0.08837890625
],
[
0.0849609375,
0.10797119140625,
0.26318359375,
0.10797119140625,
0.26318359375,
0.121826171875,
0.0849609375,
0.121826171875
],
[
0.08660888671875,
0.1256103515625,
0.91259765625,
0.1256103515625,
0.91259765625,
0.1395263671875,
0.08660888671875,
0.1395263671875
],
[
0.0882568359375,
0.141357421875,
0.3798828125,
0.141357421875,
0.3798828125,
0.1514892578125,
0.0882568359375,
0.1514892578125
],
[
0.101318359375,
0.15283203125,
0.91259765625,
0.154052734375,
0.91259765625,
0.1673583984375,
0.101318359375,
0.1666259765625
],
[
0.08660888671875,
0.1666259765625,
0.9111328125,
0.1666259765625,
0.9111328125,
0.179931640625,
0.08660888671875,
0.179931640625
],
[
0.0882568359375,
0.1817626953125,
0.27685546875,
0.1817626953125,
0.27685546875,
0.1925048828125,
0.0882568359375,
0.1925048828125
],
[
0.101318359375,
0.1925048828125,
0.9111328125,
0.1944580078125,
0.9111328125,
0.208984375,
0.101318359375,
0.20703125
],
[
0.08660888671875,
0.208984375,
0.91259765625,
0.208984375,
0.91259765625,
0.222900390625,
0.08660888671875,
0.222900390625
],
[
0.08660888671875,
0.222900390625,
0.91259765625,
0.222900390625,
0.91259765625,
0.236083984375,
0.08660888671875,
0.236083984375
],
[
0.08660888671875,
0.236083984375,
0.91259765625,
0.236083984375,
0.91259765625,
0.250732421875,
0.08660888671875,
0.250732421875
],
[
0.08660888671875,
0.248779296875,
0.9111328125,
0.250732421875,
0.9111328125,
0.264404296875,
0.08660888671875,
0.261962890625
],
[
0.08660888671875,
0.264404296875,
0.91259765625,
0.264404296875,
0.91259765625,
0.27783203125,
0.08660888671875,
0.27783203125
],
[
0.08660888671875,
0.279052734375,
0.9140625,
0.279052734375,
0.9140625,
0.292236328125,
0.08660888671875,
0.292236328125
],
[
0.08660888671875,
0.292236328125,
0.22216796875,
0.292236328125,
0.22216796875,
0.30615234375,
0.08660888671875,
0.30615234375
],
[
0.08416748046875,
0.32080078125,
0.2459716796875,
0.32080078125,
0.2459716796875,
0.333984375,
0.08416748046875,
0.333984375
],
[
0.08416748046875,
0.339111328125,
0.293212890625,
0.34033203125,
0.293212890625,
0.354736328125,
0.08416748046875,
0.354248046875
],
[
0.0882568359375,
0.35791015625,
0.9111328125,
0.35791015625,
0.9111328125,
0.371826171875,
0.0882568359375,
0.371826171875
],
[
0.08660888671875,
0.371826171875,
0.91259765625,
0.371826171875,
0.91259765625,
0.385009765625,
0.08660888671875,
0.385009765625
],
[
0.08660888671875,
0.386962890625,
0.491943359375,
0.386962890625,
0.491943359375,
0.397705078125,
0.08660888671875,
0.397705078125
],
[
0.101318359375,
0.3984375,
0.91259765625,
0.399658203125,
0.91259765625,
0.41357421875,
0.101318359375,
0.412353515625
],
[
0.0849609375,
0.412353515625,
0.7587890625,
0.41357421875,
0.7587890625,
0.4267578125,
0.0849609375,
0.425537109375
],
[
0.0882568359375,
0.436279296875,
0.9111328125,
0.436279296875,
0.9111328125,
0.449462890625,
0.0882568359375,
0.449462890625
],
[
0.08905029296875,
0.45068359375,
0.9111328125,
0.45068359375,
0.9111328125,
0.464111328125,
0.08905029296875,
0.464111328125
],
[
0.08660888671875,
0.464111328125,
0.38232421875,
0.464111328125,
0.38232421875,
0.477783203125,
0.08660888671875,
0.477783203125
],
[
0.0882568359375,
0.4873046875,
0.91259765625,
0.4873046875,
0.91259765625,
0.50048828125,
0.0882568359375,
0.50048828125
],
[
0.0882568359375,
0.50048828125,
0.91259765625,
0.50048828125,
0.91259765625,
0.5146484375,
0.0882568359375,
0.5146484375
],
[
0.0882568359375,
0.5166015625,
0.9111328125,
0.5166015625,
0.9111328125,
0.5263671875,
0.0882568359375,
0.5263671875
],
[
0.10296630859375,
0.54248046875,
0.91259765625,
0.54248046875,
0.91259765625,
0.55615234375,
0.10296630859375,
0.55615234375
],
[
0.08660888671875,
0.556640625,
0.8310546875,
0.556640625,
0.8310546875,
0.57080078125,
0.08660888671875,
0.57080078125
],
[
0.10296630859375,
0.5693359375,
0.7294921875,
0.5693359375,
0.7294921875,
0.58349609375,
0.10296630859375,
0.58349609375
],
[
0.08416748046875,
0.5986328125,
0.306396484375,
0.59765625,
0.306396484375,
0.6123046875,
0.08416748046875,
0.61279296875
],
[
0.0882568359375,
0.6162109375,
0.91259765625,
0.6162109375,
0.91259765625,
0.6298828125,
0.0882568359375,
0.6298828125
],
[
0.0882568359375,
0.6298828125,
0.9140625,
0.6298828125,
0.9140625,
0.64306640625,
0.0882568359375,
0.64306640625
],
[
0.0882568359375,
0.64599609375,
0.55224609375,
0.64599609375,
0.55224609375,
0.65576171875,
0.0882568359375,
0.65576171875
],
[
0.10455322265625,
0.65771484375,
0.91259765625,
0.65771484375,
0.91259765625,
0.671875,
0.10455322265625,
0.671875
],
[
0.08905029296875,
0.673828125,
0.9111328125,
0.673828125,
0.9111328125,
0.6845703125,
0.08905029296875,
0.6845703125
],
[
0.08905029296875,
0.6875,
0.90869140625,
0.6875,
0.90869140625,
0.69775390625,
0.08905029296875,
0.69775390625
],
[
0.08660888671875,
0.69970703125,
0.91259765625,
0.69970703125,
0.91259765625,
0.71337890625,
0.08660888671875,
0.71337890625
],
[
0.08905029296875,
0.71533203125,
0.9111328125,
0.71533203125,
0.9111328125,
0.72607421875,
0.08905029296875,
0.72607421875
],
[
0.08660888671875,
0.72802734375,
0.91259765625,
0.72802734375,
0.91259765625,
0.7412109375,
0.08660888671875,
0.7412109375
],
[
0.0882568359375,
0.74365234375,
0.160888671875,
0.74365234375,
0.160888671875,
0.75390625,
0.0882568359375,
0.75390625
],
[
0.08905029296875,
0.76513671875,
0.91015625,
0.76513671875,
0.91015625,
0.77587890625,
0.08905029296875,
0.77587890625
],
[
0.0882568359375,
0.77978515625,
0.90869140625,
0.77978515625,
0.90869140625,
0.79052734375,
0.0882568359375,
0.79052734375
],
[
0.08905029296875,
0.79345703125,
0.90869140625,
0.79345703125,
0.90869140625,
0.8037109375,
0.08905029296875,
0.8037109375
],
[
0.0849609375,
0.8046875,
0.91259765625,
0.80615234375,
0.91259765625,
0.8193359375,
0.0849609375,
0.818359375
],
[
0.0882568359375,
0.8212890625,
0.9111328125,
0.8212890625,
0.9111328125,
0.83203125,
0.0882568359375,
0.83203125
],
[
0.08416748046875,
0.849609375,
0.259033203125,
0.85107421875,
0.259033203125,
0.86572265625,
0.08416748046875,
0.8642578125
],
[
0.0882568359375,
0.86279296875,
0.35546875,
0.86279296875,
0.35546875,
0.876953125,
0.0882568359375,
0.876953125
],
[
0.0849609375,
0.8720703125,
0.371826171875,
0.875,
0.371826171875,
0.88818359375,
0.0849609375,
0.88525390625
],
[
0.37744140625,
0.876953125,
0.52783203125,
0.876953125,
0.52783203125,
0.8837890625,
0.37744140625,
0.8837890625
],
[
0.54833984375,
0.876953125,
0.7890625,
0.876953125,
0.7890625,
0.8837890625,
0.54833984375,
0.8837890625
],
[
0.08660888671875,
0.8837890625,
0.82275390625,
0.8837890625,
0.82275390625,
0.89794921875,
0.08660888671875,
0.89794921875
]
] |
[
"A Knowledge-Injected Curriculum Pretraining Framework for Question Answering",
"wWW'24,May1317,2024,Singapore,Singapore",
"3.3Implementation",
"在本节中，我们提供了通用KICP框架的实现。在KI中，我们实现了E3.2.1节中提到的文本表征和掩码，并通过简单地连接所有",
"字段来实现T，这在我们的数据集上效果很好",
"在KA中，我们以相同的层数和减半的向量维度将知识适配器Ad实现为BERT。在Ad的每一层中，输人（来自LM相应层的",
"语义向量）首先用线性模型投影到来自最后一层的隐藏向量的潜在空间，然后与隐藏向量相加以馈送到BERT层。Ad和LM的最",
"终向量与线性层合并作为输出",
"在cR中，我们使用两种广泛使用的推理模式实现Comp，即。e。、多跳推理和多对象推理。多跳推理(e。g。，作者OffonaComet",
"的时期是1828-1905）首先从题目中的主题实体推断一个中间实体（作者OffonaComet是JulesVerne)，然后用它来推断下一个中间实",
"体，直到得出答案（JulesVerne的时期是1828-1905)。因此，知识三元组形成链状结构，其中一个三元组的尾部实体是下一个三元组",
"的头部（e。g。，儒勒·凡尔纳)。给定这些三元组，Comp丢弃所有中间实体并按顺序连接其他字段。多对象推理(e。g。，儒勒",
"凡尔纳的职业是小说家和剧作家）从一个主题实体推断出几个结果，因此知识三元组共享相同的头部实体和关系（儒勒·凡尔纳和",
"职业)。给定三元组，Comp丢弃头和关系期望第一个，并将所有尾巴与第一个头和关系连接起来。此外，我们的框架还可以通过",
"定义三元组的采样限制和Comp方法，以类似的方式轻松推广到其他推理模式，例如比较推理。对于每个句子，我们采样2到3",
"个匹配模式的三元组",
"4EXPERIMENTS",
"4.1Experimental Setup",
"4.1.1Datasets：我们使用三个KBQA数据集来评估基于知识的推理的KICP，即。e。，CN-QA（以CN-KG为KG），ComplexWe",
"bQuestionsE[35]和FreebaseQAE[13]（均带有WikidataE[37]），以及生成数据集Math23KE[38]】（带有HowNet[32]），以推",
"广到其他与知识相关的QA。数据集的介绍和统计信息见附录A",
"KBQA使用KG中的实体或属性值回答问题。为了在不损失太多困难的情况下降低计算复杂性，我们在KBQA上采样了10个具",
"有相同类型真值的硬候选答案进行预测。我们还从每个数据集的整个KG中采样一个子图，以加速预训练",
"4.1.2BaselineMethods.我们将KICP与原始的LMsBERTE[6]和RoBERTaE[22]以及知识增强的LMsERNIEE[50]，K-BERTE[21]",
"，KEPLERE[37]和K-AdapterE[36]。我们还包括基于嵌人的嵌人KGQA[33]和两个LLMGPT4[1]和ChatGLM2[8]作为KBQA数",
"据集的基线。我们在附录B中简要介绍了基线",
"4.1.3TrainingDetails.我们通过Huggingface基于预训练的BERT使用Pytorch实现KICP。1对于所有方法，我们在中文数据集",
"CN-QA和Math23K上使用“bert-base-chinese”版本作为LM，在英文数据集ComplexWebQuestions和FreebaseQA上使用“bert-",
"base-uncased。KA的Ad的BERT层数为12（等于LM），隐藏向量的维数为384（LM的一半），输出向量的维数为768（等于LM",
"我们使用AdamWE[24]预训练3个时期的模型。批量大小设置为32，学习率为0。0005，它在前10个%步骤中升温，然后线性",
"衰减。MLM的屏蔽概率设置为0。第1课和第3课中的15，以及0。3在第2课中，因为语料库包含更多重复的信息",
"我们在具有两个2的Linux服务器上运行所有实验。20GHzIntelXeonE5-2650CPU和TeslaK80GPU。",
"4.2ExperimentalResults",
"4.2.1OverallResults.在本节中，我们将KICP与所有基线进行比较。我们使用F1分数（F1）和精确匹配分数（EM）作为多标",
"签数据集CN-QA和ComplexWebQuestions的指标，以及单标签数据集FreebaseQA的准确性（ACC）。Math23K具有答案准确性",
"（ACC），即。e。，如果计算出的答案等于真值，则预测表达式被视为正确",
"表1中报告了四个数据集的结果。3我们用配对t检验对KICP相对于基线(GPT4除外)的改善进行了统计检验，发现P<0.05(标",
"记为*）的改善是显著的。我们可以得到以下观察结果。首先，KICP优于所有基线，这清楚地表明了其在QA任务的知识学习和利",
"用方面的有效性。其次，在模型相似但不同的预训练任务下，KICP的性能优于K-Adapter，显示出预训练任务的显着影响。第三",
"LLM的性能并不比KBQA上的微调方法更好。GPT4在广泛研究的ComplexWebQuestions和FreebaseQA上实现了相当的性能，但",
"在CN-QA上远远落后，而较小的ChatGLM2的性能甚至更差。第四，知识增强的方法在大多数情况下优于原始的LMs，证明了知",
"识是QA推理的关键要素。最后，知识注人并没有给Math23K带来太大的改善甚至负面影响。原因可能是Math23K需要NLU远",
"远超过知识",
"4.2.2AblationStudy.在本节中，我们进行消融实验以研究属性知识和知识适应的有效性（我们将在E4.3节中详细研究课程推",
"理）。我们介绍了KICP的两个变体：KICP-KA删除了知识适应模块并直接训练原始LM的参数，而KICP-ATT丢弃了属性知识并",
"仅在实体关系知识上进行预训练。两种变体的结果也报告在表1中。4我们可以总结以下结论。首先，这两个变体的表现要比KICI",
"差，这表明KA可以减少生成的语料库的负面影响，并且属性知识在KBQA中也很有用。接下来，在CN-QA中，KICP-ATT的表",
"现要比KICP-KA差，这意味着属性知识开发在此任务上的贡献要大于知识适应。结果是合理的，因为CN-QA的很大一部分需要",
"Thttps://huggingface.co/transformers",
"2 Our codes are available at https:/github.com/l-xin/KICP",
"3we do not evaluate KEPLER and K-Adapter on Math23K, as",
"retraining thetwomethods requ",
"entitydescriptions, which areunavailable onHowN",
"*The results of KICP-ATT on ComplexWebQuestions, FreebaseQA and Math23K are unavailable, as Wikidata and HowNet do not contain attribute knowledge"
] |
{
"arxiv_id": "2403.09712",
"language": "zh",
"page_idx": 4
}
|
||
2209.13513-zh-3
|
2209.13513-zh-3.jpg
|
Title Suppressed Due to Excessive Size
根对$\Sigma_t$进行归一化来计算相关矩阵$F_t\in[-1,1]^{V\times V}$ 。通过这种方式构造动态节点特征，每个节点的特征向量的长度为$B=V$ 。这种节点特征的选择是由先前在静态脑图上的工作所激发的，其中节点的 FC 配置文件与其他特征相比显示出优越的性能 (Li et al., 2021; Kan et al., 2022; Cui et al., 2022)
## 4.2. Dynamic graph classifier
Spatial-temporal feature extraction 为了从 (4)和 (5) 中学习动态图$G_1:T=(\boldsymbol{A}_{1:T},\boldsymbol{F}_{1:T})$的时空表示，我们采用$L_C$分层递归 GNN$\underline{\mathrm{E}}($Seo et al.,2018),表示为$f_{\theta_F}(\cdot)$ 。为简单起见，我们使用门控递归单元(GRU)$\boxed{\mathrm{E}}$ (Cho et al., 2014) 和每个门作为图卷积网络 (GCN)$\boxed{\mathrm{E}}$ (Kipf $\&$ Welling, 2016) 来实现递归关系。具体来说，定义了第$l$层中的门$g\in\{r,u,c\}$的 GCN
$$\mathrm{GCN}_{g}^{(l)}(\boldsymbol{F}_{t},\boldsymbol{A}_{t})=\hat{\boldsymbol{D}}_{t}^{-1/2}\hat{A}_{t}\hat{\boldsymbol{D}}_{t}^{-1/2}\boldsymbol{F}_{t}\boldsymbol{W}_{g}^{(l)}$$
(6)
,其中$W_g^{(l)}\in\mathbb{R}^{K_C\times K_C}$ 是可训练的权重矩阵$,\hat{A}_t=$ $A_t+I_V$表示具有添加的自边缘的快照邻接矩阵，$\hat{D}_t=$ $\operatorname{diag}(\hat{\boldsymbol{A}}_t\hat{\mathbf{1}}_{1\times V}^\top)$是度矩阵。在每个快照处，GRU 的第$l$ 层被定义为
(7)
$$\boldsymbol{R}_{t}^{(l)}=\mathrm{Sigmoid}\Big(\mathrm{GCN}_{r}^{(l)}(\tilde{\boldsymbol{H}}_{t}^{(l-1)}||\tilde{\boldsymbol{H}}_{t-1}^{(l)},\boldsymbol{A}_{t})\Big)\\\boldsymbol{U}_{t}^{(l)}=\mathrm{Sigmoid}\Big(\mathrm{GCN}_{u}^{(l)}(\tilde{\boldsymbol{H}}_{t}^{(l-1)}||\tilde{\boldsymbol{H}}_{t-1}^{(l)},\boldsymbol{A}_{t})\Big)\\\boldsymbol{C}_{t}^{(l)}=\mathrm{Tanh}\Big(\mathrm{GCN}_{c}^{(l)}(\tilde{\boldsymbol{H}}_{t}^{(l-1)}||\boldsymbol{R}_{t}^{(l)}\odot\tilde{\boldsymbol{H}}_{t-1}^{(l)},\boldsymbol{A}_{t})\Big)$$
(8)
(9)
(10)
$$\tilde{\boldsymbol{H}}_t^{(l)}=\boldsymbol{U}_t^{(l)}\odot\tilde{\boldsymbol{H}}_{t-1}^{(l)}+(1-\boldsymbol{U}_t^{(l)})\odot\boldsymbol{C}_t^{(l)}$$
,其中 $R_t^{(l)},U_t^{(l)}\in\mathbb{R}^{V\times K_c}$ 分别是复位门和更新门。将初始隐藏状态$\tilde{\boldsymbol{H}}_t^{(0)}=\boldsymbol{F}_t$设置为输入节点特征，并将$\tilde{\boldsymbol{H}}_0^{(l)}\in\mathbf{0}_{V\times K_C}$初始化为零矩阵。符号$\odot$和$\parallel$分别表示Hadamard 积和特征级联算子。通过为每个快照迭代(7)-(10) ,我们获得每层输出嵌入$H_1:T^{(l)}\in\mathbb{R}^{V\times K_C\times T}$ 。这些嵌人沿着特征维度连接，以组合来自与每个节点相距$L_C$跳的邻居的信息，然后在节点维度上进行平均，以创建遵循
$$H_{1:T}^C=\phi(||_{l=1}^{L_C}H_{1:T}^{(l)})\in\mathbb{R}^{K_CL_C\times T}$$
(11)
的脑图嵌入序列，其中$\phi=\frac1V\mathbf{1}_{1\times V}$表示平均池化矩阵。虽然我们只关注 GCN 和 GRU,但将提出的方法扩展到其他 GNNs (Hamilton et al., 2017; Veličković et al., 2017) 和 RNNs (Hochreiter & Schmidhuber, 1997)是很简单的，我们将其留给未来的工作
Temporal attention readout 为了突出显示包含$\mathbb{E}$ (11)中最相关的大脑图嵌入$\boldsymbol{H}_{1:T}^C$的快照，我们合并了一个新颖的时间注意力读出层，表示为$f_{\theta _T}( \cdot )$ , 它改编自挤压和激发注意力网络$\overline{\mathrm{E}}($Hu et al., 2018) 。更正式地说，该层引入了在
$$\alpha=\mathrm{Sigmoid}(\mathrm{ReLU}(\psi H_{1:T}^CW_1)W_2)$$
(12)
之后定义的时间注意力得分矩阵$\boldsymbol\alpha\in(0,1)^{1\times T}$,其$中W_1\in\mathbb{R}^{T\times\tau T},W_2\in\mathbb{R}^{\tau T\times T}$是捕获时间依赖性的可训练权重矩阵。超参数 $\tau\in(0,1]$ 控制瓶颈， $\psi=\frac{1}{K_CL_C}\mathbf{1}_{1\times K_CL_C}$表示平均池化矩阵
Graph-level representation 使用$\boxed{\mathrm{E}}(12)$ 的时间注意力
得分$\alpha$ ,我们使用快照的加权和
(13)
$$h_{\mathcal{G}}=(\alpha\odot H_{1:T}^{C})\xi^{\top}$$
来计算图级表示$h_G\in\mathbb{R}^{L_CK_C}$,其中$\xi=\mathbf{1}_{1\times T}$表示总和池矩阵。最后，我们通过一个线性层传递图级表示，将其映射到类概率向量$p(y|\boldsymbol{X}_{1:T^{\prime}})\in\Delta^C$ 。正式
$$p(y|\boldsymbol{X}_{1:T'})=\mathrm{Softmax}(\boldsymbol{h}_{\mathcal{G}}\boldsymbol{W}_3)$$
(14)
,其中$W_3\in\mathbb{R}^{L_CL_K\times C}$ 是可训练的权重矩阵
4.3. Loss function
由于我们的任务是图分类，因此我们通过最小化交叉熵损失$\mathcal{L}_\mathrm{CE}(y,\hat{y})$以及对表示为$\mathcal{R}(G_1:T)$的学习图的先前约束的集合来训练 DynDepNet ,使得
(15)
$$\mathcal{L}(y,\{\hat{y},G_{1:T}\})=\mathcal{L}_{\mathrm{CE}}(y,\hat{y})+\mathcal{R}(G_{1:T}).$$
这里，$\mathcal{L} _{\mathrm{CE}}( y, \hat{y} ) = - \sum _{c= 1}^C\mathbb{1} ( y= c) \log p( y| \boldsymbol{X}_{1: T^{\prime }}) _c$和$p(\cdot|\cdot)_c$是来自概率输出的$c$ -THX。这鼓励 DynDepNet 学习将类差异编码到$A_{1:T}$中的任务感知动态图
Regularization constraints 由于图中的连接节点更有可能共享相似的特征E(McPherson et al.,2001) ,因此我们添加了一个正则化项，以鼓励定义为
$$\mathcal{L}_{\mathrm{FS}}(\boldsymbol{A}_{1:T},\boldsymbol{F}_{1:T})=\frac{1}{V^2}\sum_{t=1}^T\mathrm{Tr}(\boldsymbol{F}_t^\top\hat{\boldsymbol{L}}_t\boldsymbol{F}_t)$$
(16)
的特征平滑度，其中 Tr(-) 表示矩阵跟踪算子，$\hat{L}_t=$ $D_t^{-1/2}L_tD_t^{-1/2}$是(对称)归一化拉普拉斯矩阵，定义为。此外，为了阻止快照之间图结构的不稳定变化， 我们还添加了一个先验约束，鼓励时间平滑度定义为
$$\mathcal{L}_{\mathrm{TS}}(\boldsymbol{A}_{1:T})=\sum_{t=1}^{T-1}||\boldsymbol{A}_t-\boldsymbol{A}_{t+1}||_1$$
(17)
其中$\|\cdot\|_1$表示矩阵 L1 范数。此外，为了鼓励在 (4) 中学习大稀疏参数 Softmax$(\theta_P)$ ,我们进一步添加了稀疏正则化项定义的
$$\mathcal{L}_{\mathrm{SP}}(A_{1:T})=\sum_{t=1}^T||A_t||_1,$$
(18)
它与$\mathcal{L}_{\mathrm{CE}}(\cdot,\cdot)$结合使用，确保仅将最导人的特定于任务的边保留在$A_t$中。我们试图最小化的最终损失
|
[
[
0.359375,
0.059356689453125,
0.615234375,
0.059356689453125,
0.615234375,
0.0694580078125,
0.359375,
0.0694580078125
],
[
0.0882568359375,
0.085205078125,
0.4755859375,
0.08209228515625,
0.4755859375,
0.0997314453125,
0.0882568359375,
0.1029052734375
],
[
0.5009765625,
0.085205078125,
0.8857421875,
0.085205078125,
0.8857421875,
0.0997314453125,
0.5009765625,
0.0997314453125
],
[
0.10455322265625,
0.10162353515625,
0.47216796875,
0.10162353515625,
0.47216796875,
0.11553955078125,
0.10455322265625,
0.11553955078125
],
[
0.498291015625,
0.09783935546875,
0.88623046875,
0.0985107421875,
0.88623046875,
0.11614990234375,
0.498291015625,
0.11553955078125
],
[
0.09234619140625,
0.11871337890625,
0.47216796875,
0.11871337890625,
0.47216796875,
0.1287841796875,
0.09234619140625,
0.1287841796875
],
[
0.5009765625,
0.11614990234375,
0.88623046875,
0.11614990234375,
0.88623046875,
0.1300048828125,
0.5009765625,
0.1300048828125
],
[
0.0906982421875,
0.1319580078125,
0.473876953125,
0.1319580078125,
0.473876953125,
0.1458740234375,
0.0906982421875,
0.1458740234375
],
[
0.498291015625,
0.1331787109375,
0.78173828125,
0.1300048828125,
0.78173828125,
0.146484375,
0.498291015625,
0.149658203125
],
[
0.0906982421875,
0.147705078125,
0.47216796875,
0.147705078125,
0.47216796875,
0.1610107421875,
0.0906982421875,
0.1610107421875
],
[
0.0882568359375,
0.1610107421875,
0.373291015625,
0.1622314453125,
0.373291015625,
0.1767578125,
0.0882568359375,
0.175537109375
],
[
0.5009765625,
0.1654052734375,
0.8837890625,
0.1654052734375,
0.8837890625,
0.1787109375,
0.5009765625,
0.1787109375
],
[
0.5009765625,
0.179931640625,
0.736328125,
0.179931640625,
0.736328125,
0.1932373046875,
0.5009765625,
0.1932373046875
],
[
0.0882568359375,
0.1932373046875,
0.301513671875,
0.1925048828125,
0.301513671875,
0.20703125,
0.0882568359375,
0.2083740234375
],
[
0.619140625,
0.2083740234375,
0.76123046875,
0.205810546875,
0.76123046875,
0.2203369140625,
0.619140625,
0.222900390625
],
[
0.85302734375,
0.2083740234375,
0.88623046875,
0.2083740234375,
0.88623046875,
0.2235107421875,
0.85302734375,
0.2235107421875
],
[
0.08905029296875,
0.2171630859375,
0.47216796875,
0.2171630859375,
0.47216796875,
0.2310791015625,
0.08905029296875,
0.2310791015625
],
[
0.09234619140625,
0.23291015625,
0.47216796875,
0.23291015625,
0.47216796875,
0.246826171875,
0.09234619140625,
0.246826171875
],
[
0.498291015625,
0.2342529296875,
0.8837890625,
0.2342529296875,
0.8837890625,
0.251953125,
0.498291015625,
0.251953125
],
[
0.0906982421875,
0.246826171875,
0.47216796875,
0.246826171875,
0.47216796875,
0.261474609375,
0.0906982421875,
0.261474609375
],
[
0.50244140625,
0.250732421875,
0.8857421875,
0.250732421875,
0.8857421875,
0.26513671875,
0.50244140625,
0.26513671875
],
[
0.0906982421875,
0.261962890625,
0.473876953125,
0.261474609375,
0.473876953125,
0.27587890625,
0.0906982421875,
0.276611328125
],
[
0.5,
0.266357421875,
0.87744140625,
0.26513671875,
0.87744140625,
0.27978515625,
0.5,
0.281005859375
],
[
0.0882568359375,
0.276611328125,
0.473876953125,
0.27587890625,
0.473876953125,
0.290283203125,
0.0882568359375,
0.291748046875
],
[
0.0882568359375,
0.291748046875,
0.47216796875,
0.292236328125,
0.47216796875,
0.306884765625,
0.0882568359375,
0.30615234375
],
[
0.5849609375,
0.294189453125,
0.7998046875,
0.294189453125,
0.7998046875,
0.309326171875,
0.5849609375,
0.309326171875
],
[
0.8544921875,
0.29541015625,
0.8857421875,
0.29541015625,
0.8857421875,
0.30810546875,
0.8544921875,
0.30810546875
],
[
0.0906982421875,
0.306884765625,
0.373291015625,
0.30810546875,
0.373291015625,
0.322509765625,
0.0906982421875,
0.3212890625
],
[
0.5,
0.318115234375,
0.826171875,
0.317626953125,
0.826171875,
0.33837890625,
0.5,
0.339111328125
],
[
0.134033203125,
0.330078125,
0.4306640625,
0.330078125,
0.4306640625,
0.347900390625,
0.134033203125,
0.347900390625
],
[
0.435546875,
0.333984375,
0.473876953125,
0.333984375,
0.473876953125,
0.347900390625,
0.435546875,
0.347900390625
],
[
0.08905029296875,
0.356689453125,
0.4755859375,
0.356689453125,
0.4755859375,
0.375,
0.08905029296875,
0.375
],
[
0.50244140625,
0.354736328125,
0.6337890625,
0.354736328125,
0.6337890625,
0.36865234375,
0.50244140625,
0.36865234375
],
[
0.09234619140625,
0.376953125,
0.473876953125,
0.376953125,
0.473876953125,
0.39013671875,
0.09234619140625,
0.39013671875
],
[
0.50341796875,
0.380126953125,
0.88134765625,
0.380126953125,
0.88134765625,
0.39013671875,
0.50341796875,
0.39013671875
],
[
0.08905029296875,
0.392578125,
0.4755859375,
0.392578125,
0.4755859375,
0.4072265625,
0.08905029296875,
0.4072265625
],
[
0.50244140625,
0.393310546875,
0.8837890625,
0.393310546875,
0.8837890625,
0.4072265625,
0.50244140625,
0.4072265625
],
[
0.08905029296875,
0.4072265625,
0.1732177734375,
0.4072265625,
0.1732177734375,
0.421630859375,
0.08905029296875,
0.421630859375
],
[
0.50244140625,
0.4091796875,
0.810546875,
0.4091796875,
0.810546875,
0.42236328125,
0.50244140625,
0.42236328125
],
[
0.1094970703125,
0.429931640625,
0.423095703125,
0.43115234375,
0.423095703125,
0.448974609375,
0.1094970703125,
0.447509765625
],
[
0.44775390625,
0.434326171875,
0.47216796875,
0.434326171875,
0.47216796875,
0.447509765625,
0.44775390625,
0.447509765625
],
[
0.5390625,
0.436279296875,
0.8857421875,
0.43505859375,
0.8857421875,
0.449462890625,
0.5390625,
0.45068359375
],
[
0.108642578125,
0.45703125,
0.423095703125,
0.458251953125,
0.423095703125,
0.4765625,
0.108642578125,
0.47607421875
],
[
0.4453125,
0.462158203125,
0.471435546875,
0.462158203125,
0.471435546875,
0.474853515625,
0.4453125,
0.474853515625
],
[
0.498291015625,
0.4609375,
0.88818359375,
0.460205078125,
0.88818359375,
0.481689453125,
0.498291015625,
0.48291015625
],
[
0.5,
0.479248046875,
0.88623046875,
0.477783203125,
0.88623046875,
0.49560546875,
0.5,
0.496826171875
],
[
0.1094970703125,
0.486083984375,
0.451904296875,
0.486083984375,
0.451904296875,
0.50390625,
0.1094970703125,
0.50390625
],
[
0.5,
0.496826171875,
0.8408203125,
0.496826171875,
0.8408203125,
0.51025390625,
0.5,
0.51025390625
],
[
0.451904296875,
0.51025390625,
0.47216796875,
0.51025390625,
0.47216796875,
0.5234375,
0.451904296875,
0.5234375
],
[
0.1094970703125,
0.52978515625,
0.39208984375,
0.52978515625,
0.39208984375,
0.54443359375,
0.1094970703125,
0.54443359375
],
[
0.441162109375,
0.53076171875,
0.473876953125,
0.53076171875,
0.473876953125,
0.546875,
0.441162109375,
0.546875
],
[
0.5009765625,
0.52783203125,
0.8837890625,
0.52783203125,
0.8837890625,
0.54150390625,
0.5009765625,
0.54150390625
],
[
0.50244140625,
0.54345703125,
0.8857421875,
0.54345703125,
0.8857421875,
0.556640625,
0.50244140625,
0.556640625
],
[
0.404296875,
0.53271484375,
0.419189453125,
0.5341796875,
0.418212890625,
0.54150390625,
0.40283203125,
0.54052734375
],
[
0.0906982421875,
0.5537109375,
0.473876953125,
0.55615234375,
0.473876953125,
0.57373046875,
0.0906982421875,
0.5712890625
],
[
0.50244140625,
0.55810546875,
0.81201171875,
0.55810546875,
0.81201171875,
0.5712890625,
0.50244140625,
0.5712890625
],
[
0.08905029296875,
0.57275390625,
0.4755859375,
0.57275390625,
0.4755859375,
0.59033203125,
0.08905029296875,
0.59033203125
],
[
0.08905029296875,
0.59033203125,
0.473876953125,
0.59228515625,
0.473876953125,
0.60986328125,
0.08905029296875,
0.60791015625
],
[
0.6845703125,
0.59033203125,
0.70166015625,
0.59033203125,
0.70166015625,
0.599609375,
0.6845703125,
0.599609375
],
[
0.70361328125,
0.58544921875,
0.7236328125,
0.58544921875,
0.7236328125,
0.59521484375,
0.70361328125,
0.59521484375
],
[
0.5400390625,
0.59228515625,
0.6845703125,
0.59765625,
0.68310546875,
0.6162109375,
0.5390625,
0.611328125
],
[
0.66650390625,
0.5947265625,
0.82421875,
0.583984375,
0.82861328125,
0.6142578125,
0.6708984375,
0.625
],
[
0.8544921875,
0.5986328125,
0.8837890625,
0.5986328125,
0.8837890625,
0.611328125,
0.8544921875,
0.611328125
],
[
0.08905029296875,
0.609375,
0.473876953125,
0.609375,
0.473876953125,
0.62255859375,
0.08905029296875,
0.62255859375
],
[
0.69921875,
0.60791015625,
0.7294921875,
0.6123046875,
0.72607421875,
0.62548828125,
0.69677734375,
0.6220703125
],
[
0.0849609375,
0.625,
0.473876953125,
0.619140625,
0.473876953125,
0.63720703125,
0.0849609375,
0.642578125
],
[
0.10455322265625,
0.64013671875,
0.473876953125,
0.64013671875,
0.473876953125,
0.65380859375,
0.10455322265625,
0.65380859375
],
[
0.5009765625,
0.6396484375,
0.88232421875,
0.63720703125,
0.88232421875,
0.6513671875,
0.5009765625,
0.65380859375
],
[
0.08905029296875,
0.65380859375,
0.47216796875,
0.65478515625,
0.47216796875,
0.66943359375,
0.08905029296875,
0.66845703125
],
[
0.5,
0.65283203125,
0.8857421875,
0.65478515625,
0.8857421875,
0.67236328125,
0.5,
0.67041015625
],
[
0.0906982421875,
0.67041015625,
0.221435546875,
0.67041015625,
0.221435546875,
0.6845703125,
0.0906982421875,
0.6845703125
],
[
0.50341796875,
0.67236328125,
0.8837890625,
0.67236328125,
0.8837890625,
0.68603515625,
0.50341796875,
0.68603515625
],
[
0.158447265625,
0.693359375,
0.40283203125,
0.689453125,
0.40283203125,
0.70703125,
0.158447265625,
0.7109375
],
[
0.50341796875,
0.689453125,
0.8759765625,
0.689453125,
0.8759765625,
0.69970703125,
0.50341796875,
0.69970703125
],
[
0.443603515625,
0.69580078125,
0.47216796875,
0.69580078125,
0.47216796875,
0.70849609375,
0.443603515625,
0.70849609375
],
[
0.0906982421875,
0.71826171875,
0.47216796875,
0.71826171875,
0.47216796875,
0.73291015625,
0.0906982421875,
0.73291015625
],
[
0.67626953125,
0.7138671875,
0.70361328125,
0.7138671875,
0.70361328125,
0.7265625,
0.67626953125,
0.7265625
],
[
0.6748046875,
0.724609375,
0.80810546875,
0.724609375,
0.80810546875,
0.7421875,
0.6748046875,
0.7421875
],
[
0.8544921875,
0.72802734375,
0.8857421875,
0.72802734375,
0.8857421875,
0.74072265625,
0.8544921875,
0.74072265625
],
[
0.0906982421875,
0.734375,
0.473876953125,
0.734375,
0.473876953125,
0.74755859375,
0.0906982421875,
0.74755859375
],
[
0.5810546875,
0.72900390625,
0.68115234375,
0.72900390625,
0.68115234375,
0.74072265625,
0.5810546875,
0.74072265625
],
[
0.67724609375,
0.7421875,
0.70166015625,
0.7421875,
0.70166015625,
0.7548828125,
0.67724609375,
0.7548828125
],
[
0.0882568359375,
0.74853515625,
0.47216796875,
0.74755859375,
0.47216796875,
0.76123046875,
0.0882568359375,
0.76220703125
],
[
0.08905029296875,
0.76318359375,
0.473876953125,
0.76318359375,
0.473876953125,
0.7763671875,
0.08905029296875,
0.7763671875
],
[
0.50341796875,
0.767578125,
0.8837890625,
0.767578125,
0.8837890625,
0.78076171875,
0.50341796875,
0.78076171875
],
[
0.0906982421875,
0.77880859375,
0.410888671875,
0.77880859375,
0.410888671875,
0.79248046875,
0.0906982421875,
0.79248046875
],
[
0.50244140625,
0.7822265625,
0.88134765625,
0.7822265625,
0.88134765625,
0.79541015625,
0.50244140625,
0.79541015625
],
[
0.5,
0.79736328125,
0.64794921875,
0.79736328125,
0.64794921875,
0.81103515625,
0.5,
0.81103515625
],
[
0.0906982421875,
0.8095703125,
0.473876953125,
0.8095703125,
0.473876953125,
0.82275390625,
0.0906982421875,
0.82275390625
],
[
0.0882568359375,
0.82373046875,
0.47216796875,
0.82373046875,
0.47216796875,
0.8369140625,
0.0882568359375,
0.8369140625
],
[
0.708984375,
0.82373046875,
0.72216796875,
0.82373046875,
0.72216796875,
0.833984375,
0.708984375,
0.833984375
],
[
0.60546875,
0.83544921875,
0.78125,
0.83544921875,
0.78125,
0.85302734375,
0.60546875,
0.85302734375
],
[
0.85302734375,
0.8369140625,
0.88623046875,
0.8369140625,
0.88623046875,
0.85302734375,
0.85302734375,
0.85302734375
],
[
0.08905029296875,
0.85302734375,
0.476318359375,
0.85400390625,
0.476318359375,
0.86865234375,
0.08905029296875,
0.8671875
],
[
0.08905029296875,
0.83837890625,
0.4755859375,
0.8388671875,
0.4755859375,
0.85400390625,
0.08905029296875,
0.85302734375
],
[
0.70166015625,
0.85302734375,
0.72802734375,
0.85302734375,
0.72802734375,
0.8642578125,
0.70166015625,
0.8642578125
],
[
0.0906982421875,
0.869140625,
0.285888671875,
0.869140625,
0.285888671875,
0.88330078125,
0.0906982421875,
0.88330078125
],
[
0.51708984375,
0.876953125,
0.8837890625,
0.876953125,
0.8837890625,
0.8916015625,
0.51708984375,
0.8916015625
],
[
0.1519775390625,
0.892578125,
0.41259765625,
0.892578125,
0.41259765625,
0.90576171875,
0.1519775390625,
0.90576171875
],
[
0.44287109375,
0.89306640625,
0.4794921875,
0.89306640625,
0.4794921875,
0.90478515625,
0.44287109375,
0.90478515625
],
[
0.5009765625,
0.892578125,
0.8837890625,
0.892578125,
0.8837890625,
0.90576171875,
0.5009765625,
0.90576171875
],
[
0.48046875,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93505859375,
0.48046875,
0.93505859375
]
] |
[
"Title Suppressed Due to ExcessiveSize",
"根对Zt进行归一化来计算相关矩阵FtE[-1,1]VxV",
"之后定义的时间注意力得分矩阵αE(0,1)1xT，其",
"通过这种方式构造动态节点特征，每个节点的特征",
"中WIEIRTXTT，W2ERTTxT是捕获时间依赖",
"向量的长度为B三V。这种节点特征的选择是由先",
"性的可训练权重矩阵。超参数TE(O,1I控制瓶颈，",
"前在静态脑图上的工作所激发的，其中节点的FC配",
"=L11xKcL表示平均池化矩阵",
"置文件与其他特征相比显示出优越的性能（Lietal.",
"2021; Kan et al., 2022: Cui et al., 2022)",
"Graph-levelrepresentation使用E(12）的时间注意力",
"得分α，我们使用快照的加权和",
"4.2. Dynamic graph classifier",
"hg =(QOHT)T",
"(13)",
"Spatial-temporalfeatureextraction为了从（4)和(5)",
"中学习动态图G1:T=（A1:T，Fi:T）的时空表示，我们",
"来计算图级表示hgERLcKc，其中=11xT表示",
"采用Lc分层递归GNNE（Seoetal.，2018），表示为",
"总和池矩阵。最后，我们通过一个线性层传递图级表",
"fe（）。为简单起见，我们使用门控递归单元（GRU)E",
"示，将其映射到类概率向量p(yX1:T）E△C。正式",
"(Choetal.,2014）和每个门作为图卷积网络（GCN)E",
"(Kipf&Welling，2016）来实现递归关系。具体来说，定",
"p(y|Xi:T) = Softmax(hgW3)",
"(14)",
"义了第l层中的门gE(r,u,c}的GCN",
"，其中WsERLcLk×C是可训练的权重矩阵",
"GCN()(Ft, At) = D-1/2 A,D-1/2 FW(l)",
"(6)",
"，其中W）RKcxKc是可训练的权重矩阵，At=",
"4.3. Loss function",
"At+Iv表示具有添加的自边缘的快照邻接矩阵，Dt=",
"由于我们的任务是图分类，因此我们通过最小化交叉",
"diag(At1ixv）是度矩阵。在每个快照处，GRU的第l",
"熵损失LcE(y,9)以及对表示为R(G1:T）的学习图的",
"层被定义为",
"先前约束的集合来训练DynDepNet，使得",
"R() = Sigmoid(GCN((H(-1)|H(1, At)",
"(7)",
"C(y, (9, G1:T)) = CcE(y, 9) + R(G1:T).(15)",
"U() = Sigmoid(GCN((H(-1)H(1, At))",
"(8)",
"这里，Cce(g,)=-c11(y=c)logp(g|X1:T)。和",
"p(-I)。是来自概率输出的c-THX。这鼓励DynDepNet",
"C() = Tanh(GCN((H(-1)|R) H(-1, At))",
"学习将类差异编码到A1:T中的任务感知动态图",
"(9)",
"H)=U()oH, +(1-U()oC)",
"(10)",
"Regularizationconstraints由于图中的连接节点更有",
"可能共享相似的特征E（McPhersonetal.，2001），因",
"",
"，其中R(\"),()RV×Kc分别是复位门和更新门。将",
"此我们添加了一个正则化项，以鼓励定义为",
"初始隐藏状态H(0）=F 设置为输人节点特征，并将",
"HOv×K。初始化为零矩阵。符号和I分别表示",
"1 :",
"T",
"Crs(A1:T,Fi:T) = ",
"T(FTLF)",
"(16)",
"Hadamard积和特征级联算子。通过为每个快照迭代",
"E",
"（7)-（10），我们获得每层输出嵌人H(RV×Ke×T",
"这些嵌人沿着特征维度连接，以组合来自与每个节",
"的特征平滑度，其中Tr(）表示矩阵跟踪算子，L",
"点相距Lc跳的邻居的信息，然后在节点维度上进行",
"D-1/2LtD-1/2是（对称）归一化拉普拉斯矩阵，定",
"平均，以创建遵循",
"义为。此外，为了阻止快照之间图结构的不稳定变化，",
"HT=Φ(/H) RKeLe×T",
"我们还添加了一个先验约束，鼓励时间平滑度定义为",
"(11)",
"的脑图嵌入序列，其中Φ=11×V表示平均池化矩",
"T-1",
" llAt - At+1lli",
"(17)",
"阵。虽然我们只关注GCN和GRU，但将提出的方法",
"LTS(A1:T)=",
"t=1",
"扩展到其他GNNs（Hamiltonetal.2017:Velickovic",
"et al.,2017)和 RNNs(Hochreiter & Schmidhuber,",
"，其中Ⅱ·I1表示矩阵L1范数。此外，为了鼓励在（4)",
"1997）是很简单的，我们将其留给未来的工作",
"中学习大稀疏参数Softmax(Op），我们进一步添加了",
"稀疏正则化项定义的",
"Temporalattentionreadout为了突出显示包含E",
"（11）中最相关的大脑图嵌人H%T的快照，我们合",
"T",
"Csp(A1:T) =ZIAtl1,",
"(18)",
"它改编自挤压和激发注意力网络E(Huetal.，2018）。",
"并了一个新颖的时间注意力读出层，表示为fo（），",
"t=1",
"更正式地说，该层引人了在",
"它与LcE（，）结合使用，确保仅将最导人的特定于",
"α = Sigmoid(ReLU(HrW)W2)",
"(12)",
"任务的边保留在At中。我们试图最小化的最终损失",
"4"
] |
{
"arxiv_id": "2209.13513",
"language": "zh",
"page_idx": 3
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.