Commit
·
733b673
1
Parent(s):
5fb0190
Update LexGLUE README.md (#4285)
Browse filesUpdate the leaderboard based on the latest results presented in the ACL 2022 version of the article.
Commit from https://github.com/huggingface/datasets/commit/746ab8cc7975b75496b9eca5d9d10cc913ba9494
README.md
CHANGED
@@ -135,19 +135,41 @@ The UNFAIR-ToS dataset contains 50 Terms of Service (ToS) from on-line platforms
|
|
135 |
The CaseHOLD (Case Holdings on Legal Decisions) dataset includes multiple choice questions about holdings of US court cases from the Harvard Law Library case law corpus. Holdings are short summaries of legal rulings accompany referenced decisions relevant for the present case. The input consists of an excerpt (or prompt) from a court decision, containing a reference to a particular case, while the holding statement is masked out. The model must identify the correct (masked) holding statement from a selection of five choices.
|
136 |
|
137 |
|
138 |
-
The current leaderboard includes several Transformer-based (Vaswaniet al., 2017) pre-trained language models, which achieve state-of-the-art performance in most NLP tasks (Bommasani et al., 2021) and NLU benchmarks (Wang et al., 2019a).
|
139 |
|
|
|
140 |
|
141 |
<table>
|
142 |
-
<tr><td>Dataset</td><td>ECtHR
|
143 |
-
<tr><td>Model</td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1</td><td>μ-F1 / m-F1 </td></tr>
|
144 |
-
<tr><td>
|
145 |
-
<tr><td
|
146 |
-
|
147 |
-
<
|
148 |
-
<
|
149 |
-
|
150 |
-
<
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
</table>
|
152 |
|
153 |
### Languages
|
@@ -369,7 +391,7 @@ An example of 'test' looks as follows.
|
|
369 |
|
370 |
*Ilias Chalkidis, Abhik Jana, Dirk Hartung, Michael Bommarito, Ion Androutsopoulos, Daniel Martin Katz, and Nikolaos Aletras.*
|
371 |
*LexGLUE: A Benchmark Dataset for Legal Language Understanding in English.*
|
372 |
-
*
|
373 |
|
374 |
|
375 |
### Licensing Information
|
@@ -380,22 +402,16 @@ An example of 'test' looks as follows.
|
|
380 |
|
381 |
[*Ilias Chalkidis, Abhik Jana, Dirk Hartung, Michael Bommarito, Ion Androutsopoulos, Daniel Martin Katz, and Nikolaos Aletras.*
|
382 |
*LexGLUE: A Benchmark Dataset for Legal Language Understanding in English.*
|
383 |
-
*
|
384 |
```
|
385 |
-
@
|
386 |
-
title={
|
387 |
-
author={Chalkidis, Ilias and
|
388 |
-
|
389 |
-
Hartung, Dirk and
|
390 |
-
Bommarito, Michael and
|
391 |
-
Androutsopoulos, Ion and
|
392 |
-
Katz, Daniel Martin and
|
393 |
Aletras, Nikolaos},
|
394 |
-
year={
|
395 |
-
|
396 |
-
|
397 |
-
primaryClass={cs.CL},
|
398 |
-
note = {arXiv: 2110.00976},
|
399 |
}
|
400 |
```
|
401 |
|
|
|
135 |
The CaseHOLD (Case Holdings on Legal Decisions) dataset includes multiple choice questions about holdings of US court cases from the Harvard Law Library case law corpus. Holdings are short summaries of legal rulings accompany referenced decisions relevant for the present case. The input consists of an excerpt (or prompt) from a court decision, containing a reference to a particular case, while the holding statement is masked out. The model must identify the correct (masked) holding statement from a selection of five choices.
|
136 |
|
137 |
|
138 |
+
The current leaderboard includes several Transformer-based (Vaswaniet al., 2017) pre-trained language models, which achieve state-of-the-art performance in most NLP tasks (Bommasani et al., 2021) and NLU benchmarks (Wang et al., 2019a). Results reported by [Chalkidis et al. (2021)](https://arxiv.org/abs/2110.00976):
|
139 |
|
140 |
+
*Task-wise Test Results*
|
141 |
|
142 |
<table>
|
143 |
+
<tr><td><b>Dataset</b></td><td><b>ECtHR A</b></td><td><b>ECtHR B</b></td><td><b>SCOTUS</b></td><td><b>EUR-LEX</b></td><td><b>LEDGAR</b></td><td><b>UNFAIR-ToS</b></td><td><b>CaseHOLD</b></td></tr>
|
144 |
+
<tr><td><b>Model</b></td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1</td><td>μ-F1 / m-F1 </td></tr>
|
145 |
+
<tr><td>TFIDF+SVM</td><td> 64.7 / 51.7 </td><td>74.6 / 65.1 </td><td> <b>78.2</b> / <b>69.5</b> </td><td>71.3 / 51.4 </td><td>87.2 / 82.4 </td><td>95.4 / 78.8</td><td>n/a </td></tr>
|
146 |
+
<tr><td colspan="8" style='text-align:center'><b>Medium-sized Models (L=12, H=768, A=12)</b></td></tr>
|
147 |
+
<td>BERT</td> <td> 71.2 / 63.6 </td> <td> 79.7 / 73.4 </td> <td> 68.3 / 58.3 </td> <td> 71.4 / 57.2 </td> <td> 87.6 / 81.8 </td> <td> 95.6 / 81.3 </td> <td> 70.8 </td> </tr>
|
148 |
+
<td>RoBERTa</td> <td> 69.2 / 59.0 </td> <td> 77.3 / 68.9 </td> <td> 71.6 / 62.0 </td> <td> 71.9 / <b>57.9</b> </td> <td> 87.9 / 82.3 </td> <td> 95.2 / 79.2 </td> <td> 71.4 </td> </tr>
|
149 |
+
<td>DeBERTa</td> <td> 70.0 / 60.8 </td> <td> 78.8 / 71.0 </td> <td> 71.1 / 62.7 </td> <td> <b>72.1</b> / 57.4 </td> <td> 88.2 / 83.1 </td> <td> 95.5 / 80.3 </td> <td> 72.6 </td> </tr>
|
150 |
+
<td>Longformer</td> <td> 69.9 / 64.7 </td> <td> 79.4 / 71.7 </td> <td> 72.9 / 64.0 </td> <td> 71.6 / 57.7 </td> <td> 88.2 / 83.0 </td> <td> 95.5 / 80.9 </td> <td> 71.9 </td> </tr>
|
151 |
+
<td>BigBird</td> <td> 70.0 / 62.9 </td> <td> 78.8 / 70.9 </td> <td> 72.8 / 62.0 </td> <td> 71.5 / 56.8 </td> <td> 87.8 / 82.6 </td> <td> 95.7 / 81.3 </td> <td> 70.8 </td> </tr>
|
152 |
+
<td>Legal-BERT</td> <td> 70.0 / 64.0 </td> <td> <b>80.4</b> / <b>74.7</b> </td> <td> 76.4 / 66.5 </td> <td> <b>72.1</b> / 57.4 </td> <td> 88.2 / 83.0 </td> <td> <b>96.0</b> / <b>83.0</b> </td> <td> 75.3 </td> </tr>
|
153 |
+
<td>CaseLaw-BERT</td> <td> 69.8 / 62.9 </td> <td> 78.8 / 70.3 </td> <td> 76.6 / 65.9 </td> <td> 70.7 / 56.6 </td> <td> 88.3 / 83.0 </td> <td> <b>96.0</b> / 82.3 </td> <td> <b>75.4</b> </td> </tr>
|
154 |
+
<tr><td colspan="8" style='text-align:center'><b>Large-sized Models (L=24, H=1024, A=18)</b></td></tr>
|
155 |
+
<tr><td>RoBERTa</td> <td> <b>73.8</b> / <b>67.6</b> </td> <td> 79.8 / 71.6 </td> <td> 75.5 / 66.3 </td> <td> 67.9 / 50.3 </td> <td> <b>88.6</b> / <b>83.6</b> </td> <td> 95.8 / 81.6 </td> <td> 74.4 </td> </tr>
|
156 |
+
</table>
|
157 |
+
|
158 |
+
*Averaged (Mean over Tasks) Test Results*
|
159 |
+
|
160 |
+
<table>
|
161 |
+
<tr><td><b>Averaging</b></td><td><b>Arithmetic</b></td><td><b>Harmonic</b></td><td><b>Geometric</b></td></tr>
|
162 |
+
<tr><td><b>Model</b></td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td><td>μ-F1 / m-F1 </td></tr>
|
163 |
+
<tr><td colspan="4" style='text-align:center'><b>Medium-sized Models (L=12, H=768, A=12)</b></td></tr>
|
164 |
+
<tr><td>BERT</td><td> 77.8 / 69.5 </td><td> 76.7 / 68.2 </td><td> 77.2 / 68.8 </td></tr>
|
165 |
+
<tr><td>RoBERTa</td><td> 77.8 / 68.7 </td><td> 76.8 / 67.5 </td><td> 77.3 / 68.1 </td></tr>
|
166 |
+
<tr><td>DeBERTa</td><td> 78.3 / 69.7 </td><td> 77.4 / 68.5 </td><td> 77.8 / 69.1 </td></tr>
|
167 |
+
<tr><td>Longformer</td><td> 78.5 / 70.5 </td><td> 77.5 / 69.5 </td><td> 78.0 / 70.0 </td></tr>
|
168 |
+
<tr><td>BigBird</td><td> 78.2 / 69.6 </td><td> 77.2 / 68.5 </td><td> 77.7 / 69.0 </td></tr>
|
169 |
+
<tr><td>Legal-BERT</td><td> <b>79.8</b> / <b>72.0</b> </td><td> <b>78.9</b> / <b>70.8</b> </td><td> <b>79.3</b> / <b>71.4</b> </td></tr>
|
170 |
+
<tr><td>CaseLaw-BERT</td><td> 79.4 / 70.9 </td><td> 78.5 / 69.7 </td><td> 78.9 / 70.3 </td></tr>
|
171 |
+
<tr><td colspan="4" style='text-align:center'><b>Large-sized Models (L=24, H=1024, A=18)</b></td></tr>
|
172 |
+
<tr><td>RoBERTa</td><td> 79.4 / 70.8 </td><td> 78.4 / 69.1 </td><td> 78.9 / 70.0 </td></tr>
|
173 |
</table>
|
174 |
|
175 |
### Languages
|
|
|
391 |
|
392 |
*Ilias Chalkidis, Abhik Jana, Dirk Hartung, Michael Bommarito, Ion Androutsopoulos, Daniel Martin Katz, and Nikolaos Aletras.*
|
393 |
*LexGLUE: A Benchmark Dataset for Legal Language Understanding in English.*
|
394 |
+
*2022. In the Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics. Dublin, Ireland.*
|
395 |
|
396 |
|
397 |
### Licensing Information
|
|
|
402 |
|
403 |
[*Ilias Chalkidis, Abhik Jana, Dirk Hartung, Michael Bommarito, Ion Androutsopoulos, Daniel Martin Katz, and Nikolaos Aletras.*
|
404 |
*LexGLUE: A Benchmark Dataset for Legal Language Understanding in English.*
|
405 |
+
*2022. In the Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics. Dublin, Ireland.*](https://arxiv.org/abs/2110.00976)
|
406 |
```
|
407 |
+
@inproceedings{chalkidis-etal-2021-lexglue,
|
408 |
+
title={LexGLUE: A Benchmark Dataset for Legal Language Understanding in English},
|
409 |
+
author={Chalkidis, Ilias and Jana, Abhik and Hartung, Dirk and
|
410 |
+
Bommarito, Michael and Androutsopoulos, Ion and Katz, Daniel Martin and
|
|
|
|
|
|
|
|
|
411 |
Aletras, Nikolaos},
|
412 |
+
year={2022},
|
413 |
+
booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics},
|
414 |
+
address={Dubln, Ireland},
|
|
|
|
|
415 |
}
|
416 |
```
|
417 |
|