Datasets:

ArXiv:
License:
kiddothe2b commited on
Commit
4c1adf8
1 Parent(s): b4013a9

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +18 -17
README.md CHANGED
@@ -101,7 +101,7 @@ task_ids:
101
 
102
  ### Dataset Summary
103
 
104
- We present a benchmark suite of four datasets for evaluating the fairness of pre-trained legal language models and the techniques used to fine-tune them for downstream tasks. Our benchmarks cover four jurisdictions (European Council, USA, Swiss, and Chinese), five languages (English, German, French, Italian and Chinese) and fairness across five attributes (gender, age, nationality/region, language, and legal area). In our experiments, we evaluate pre-trained language models using several group-robust fine-tuning techniques and show that performance group disparities are vibrant in many cases, while none of these techniques guarantee fairness, nor consistently mitigate group disparities. Furthermore, we provide a quantitative and qualitative analysis of our results, highlighting open challenges in the development of robustness methods in legal NLP.
105
 
106
  For the purpose of this work, we release four domain-specific BERT models with continued pre-training on the corpora of the examined datasets (ECtHR, SCOTUS, FSCS, CAIL). We train mini-sized BERT models with 6 Transformer blocks, 384 hidden units, and 12 attention heads. We warm-start all models from the public MiniLMv2 (Wang et al., 2021) using the distilled version of RoBERTa (Liu et al., 2019). For the English datasets (ECtHR, SCOTUS) and the one distilled from XLM-R (Conneau et al., 2021) for the rest (trilingual FSCS, and Chinese CAIL). [[Link to Models](https://huggingface.co/models?search=fairlex)]
107
 
@@ -123,19 +123,19 @@ The supported tasks are the following:
123
  The European Court of Human Rights (ECtHR) hears allegations that a state has breached human rights provisions of the European Convention of Human Rights (ECHR). We use the dataset of Chalkidis et al. (2021), which contains 11K cases from ECtHR's public database.
124
  Each case is mapped to *articles* of the ECHR that were violated (if any). This is a multi-label text classification task. Given the facts of a case, the goal is to predict the ECHR articles that were violated, if any, as decided (ruled) by the court. The cases are chronologically split into training (9k, 2001--16), development (1k, 2016--17), and test (1k, 2017--19) sets.
125
 
126
- To facilitate the study of fairness of text classifiers, we record for each case the following attributes: (a) The _defendant states_, which are the European states that allegedly violated the ECHR. The defendant states for each case is a subset of the 47 Member States of the Council of Europe; To have statistical support, we group defendant states in two groups:
127
- Central-Eastern European states, on one hand, and all other states, as classified by the EuroVoc thesaurus. (b) The _applicant's age_ at the time of the decision. We extract the birth year of the applicant from the case facts, if possible, and classify its case in an age group (<=35, <=64, or older) ; and (c) the _applicant's gender_, extracted from the facts, if possible based on pronouns, classified in two categories (male, female).
128
 
129
  #### scotus
130
 
131
- The US Supreme Court (SCOTUS) is the highest federal court in the United States of America and generally hears only the most controversial or otherwise complex cases which have not been sufficiently well solved by lower courts.
132
- We combine information from SCOTUS opinions with the Supreme Court DataBase (SCDB) (Spaeth, 2020). SCDB provides metadata (e.g., date of publication, decisions, issues, decision directions and many more) for all cases. We consider the available 14 thematic issue areas (e.g, Criminal Procedure, Civil Rights, Economic Activity, etc.). This is a single-label multi-class document classification task. Given the court opinion, the goal is to predict the issue area whose focus is on the subject matter of the controversy (dispute). SCOTUS contains a total of 9,262 cases that we split chronologically into 80% for training (7.4k, 1946--1982), 10% for development (914, 1982--1991) and 10% for testing (931, 1991--2016).
133
 
134
- From SCDB, we also use the following attributes to study fairness: (a) the _type of respondent_, which is a manual categorization of respondents (defendants) in five categories (person, public entity, organization, facility and other); and (c) the _direction of the decision_, i.e., whether the decision is liberal, or conservative, provided by SCDB.
135
 
136
  #### fscs
137
 
138
- The Federal Supreme Court of Switzerland (FSCS) is the last level of appeal in Switzerland and similarly to SCOTUS, the court generally hears only the most controversial or otherwise complex cases which have not been sufficiently well solved by lower courts. The court often focus only on small parts of previous decision, where they discuss possible wrong reasoning by the lower court. The Swiss-Judgment-Predict dataset (Niklaus et al., 2021) contains more than 85K decisions from the FSCS written in one of three languages (50K German, 31K French, 4K Italian) from the years 2000 to 2020.
139
  The dataset is not parallel, i.e., all cases are unique and decisions are written only in a single language.
140
  The dataset provides labels for a simplified binary (_approval_, _dismissal_) classification task. Given the facts of the case, the goal is to predict if the plaintiff's request is valid or partially valid. The cases are also chronologically split into training (59.7k, 2000-2014), development (8.2k, 2015-2016), and test (17.4k, 2017-2020) sets.
141
 
@@ -145,7 +145,7 @@ The dataset provides three additional attributes: (a) the _language_ of the FSCS
145
 
146
  The Supreme People's Court of China (CAIL) is the last level of appeal in China and considers cases that originated from the high people's courts concerning matters of national importance. The Chinese AI and Law challenge (CAIL) dataset (Xiao et al., 2018) is a Chinese legal NLP dataset for judgment prediction and contains over 1m criminal cases. The dataset provides labels for *relevant article of criminal code* prediction, *charge* (type of crime) prediction, imprisonment *term* (period) prediction, and monetary *penalty* prediction. The publication of the original dataset has been the topic of an active debate in the NLP community(Leins et al., 2020; Tsarapatsanis and Aletras, 2021; Bender, 2021).
147
 
148
- Recently, Wang et al. (2021) re-annotated a subset of approx. 100k cases with demographic attributes. Specifically the new dataset has been annotated with: (a) the _applicant's gender_, classified in two categories (male, female); and (b) the _region_ of the court that denotes in which out of the 7 provincial-level administrative regions was the case judged. We re-split the dataset chronologically into training (80k, 2013-2017), development (12k, 2017-2018), and test (12k, 2018) sets. In our study, we re-frame the imprisonment _term_ prediction and examine a soft version, dubbed _crime severity_ prediction task, a multi-class classification task, where given the facts of a case, the goal is to predict how severe was the committed crime with respect to the imprisonment term. We approximate crime severity by the length of imprisonment term, split in 6 clusters (0, <=12, <=36, <=60, <=120, >120 months).
149
 
150
  ### Languages
151
 
@@ -217,7 +217,7 @@ An example of 'train' looks as follows.
217
  #### scotus
218
  - `text`: a `string` feature (the court opinion).
219
  - `label`: a classification label (the relevant issue area). The issue areas are: (1, Criminal Procedure), (2, Civil Rights), (3, First Amendment), (4, Due Process), (5, Privacy), (6, Attorneys), (7, Unions), (8, Economic Activity), (9, Judicial Power), (10, Federalism), (11, Interstate Relations), (12, Federal Taxation), (13, Miscellaneous), (14, Private Action).
220
- - `respondent_type`: the type of respondent, which is a manual categorization (clustering) of respondents (defendants) in five categories (person, public entity, organization, facility and other).
221
  - `decision_direction`: the direction of the decision, i.e., whether the decision is liberal, or conservative, provided by SCDB.
222
 
223
  #### fscs
@@ -229,7 +229,7 @@ An example of 'train' looks as follows.
229
 
230
  #### cail
231
  - `text`: a `string` feature (the factual description of the case).
232
- - `label`: a classification label (crime severity derived by the inprisonment term).
233
  - `defendant_gender`: the gender of the defendant (Male or Female).
234
  - `court_region`: the region of the court that denotes in which out of the 7 provincial-level administrative regions was the case judged.
235
 
@@ -264,7 +264,7 @@ We standardize and put together four datasets: ECtHR (Chalkidis et al., 2021), S
264
 
265
  The benchmark is not a blind stapling of pre-existing resources, we augment previous datasets. In the case of ECtHR, previously unavailable demographic attributes have been released to make the original dataset amenable for fairness research. For SCOTUS, two resources (court opinions with SCDB) have been combined for the very same reason, while the authors provide a manual categorization (clustering) of respondents.
266
 
267
- All datasets, except SCOTUS, are publicly available and have been previously published. If datasets or the papers where they were introduced in were not compiled or written by the authors, the original work is referenced and authors encourage FairLex users to do so as well. In fact, this work should only be referenced, in addition to citing the original work, when jointly experimenting with multiple FairLex datasets and using the FairLex evaluation framework and infrastructure, or use any newly introduced annotations (ECtHR, SCOTUS). Otherwise only the original work should be cited.
268
 
269
  #### Who are the source language producers?
270
 
@@ -280,26 +280,27 @@ All datasets, except SCOTUS, are publicly available and have been previously pub
280
 
281
  All classification labels rely on legal decisions (ECtHR, FSCS, CAIL), or are part of archival procedures (SCOTUS).
282
 
283
- The demographic attributes and other metadata are either provided by the legal databases or have been extracted automatically from the text by means of Regular Expressions.
284
 
285
  Consider the **Dataset Description** and **Discussion of Biases** sections, and the original publication for detailed information.
286
 
287
  ### Personal and Sensitive Information
288
 
 
289
  ECtHR cases are partially annonymized by the court. Its data is processed and made public in accordance with the European Data Protection Law.
290
- SCOTUS cases may also contain personal information and the data is processed and made available by the US Supreme Court, whose proceedings are public. While this ensures compliance with US law, it is very likely that similarly to the ECtHR any processing could be justified by either implied consent or legitimate interest under European law. In FSCS, the names of the parties have been redacted by the court according to its official guidelines. To the best of our knowledge, same applies for CAIL.
291
 
292
  ## Considerations for Using the Data
293
 
294
  ### Social Impact of Dataset
295
 
296
- This work can help practitioners to build assisting technology for legal professionals - with respect to the legal framework (jurisdiction) they operate -; technology that does not only rely on performance on majority groups, but also considering minorities and the robustness of the developed models across them. This is an important application field, where more research should be conducted (Tsarapatsanis and Aletras, 2021) in order to improve legal services and democratize law, but more importantly highlight (inform the audience on) the various multi-aspect shortcomings seeking a responsible and ethical (fair) deployment of technology.
297
 
298
  ### Discussion of Biases
299
 
300
- The current version of FairLex covers a very small fraction of legal applications, jurisdictions, and protected attributes. The benchmark inevitably cannot cover "_everything in the whole wide (legal) world_" (Raji et al., 2021), but nonetheless we believe that the published resources will help critical research in the area of fairness.
301
 
302
- Some protected attributes within the datasets are extracted automatically, i.e., the gender and the age of the ECtHR dataset, by means of Regular Expressions, or manually clustered by the authors, such as the defendant state in the ECtHR dataset and the respondent attribute in the SCOTUS dataset. Those assumptions and simplifications can hold in an experimental setting only and by no means should be used in real-world applications where some simplifications, e.g., binary gender, would not be appropriate. By no means, the authors or future users have to endorse the law standards or framework of the examined datasets, to any degree rather than the publication and use of the data.
303
 
304
  ### Other Known Limitations
305
 
@@ -341,7 +342,7 @@ The benchmark is released under a [Attribution-NonCommercial-ShareAlike 4.0 Inte
341
  }
342
  ```
343
 
344
- **Note:** Please consider citing and give credits to all publications releasing the examined datasets.
345
 
346
  ### Contributions
347
 
 
101
 
102
  ### Dataset Summary
103
 
104
+ We present a benchmark suite of four datasets for evaluating the fairness of pre-trained legal language models and the techniques used to fine-tune them for downstream tasks. Our benchmarks cover four jurisdictions (European Council, USA, Swiss, and Chinese), five languages (English, German, French, Italian, and Chinese), and fairness across five attributes (gender, age, nationality/region, language, and legal area). In our experiments, we evaluate pre-trained language models using several group-robust fine-tuning techniques and show that performance group disparities are vibrant in many cases, while none of these techniques guarantee fairness, nor consistently mitigate group disparities. Furthermore, we provide a quantitative and qualitative analysis of our results, highlighting open challenges in the development of robustness methods in legal NLP.
105
 
106
  For the purpose of this work, we release four domain-specific BERT models with continued pre-training on the corpora of the examined datasets (ECtHR, SCOTUS, FSCS, CAIL). We train mini-sized BERT models with 6 Transformer blocks, 384 hidden units, and 12 attention heads. We warm-start all models from the public MiniLMv2 (Wang et al., 2021) using the distilled version of RoBERTa (Liu et al., 2019). For the English datasets (ECtHR, SCOTUS) and the one distilled from XLM-R (Conneau et al., 2021) for the rest (trilingual FSCS, and Chinese CAIL). [[Link to Models](https://huggingface.co/models?search=fairlex)]
107
 
 
123
  The European Court of Human Rights (ECtHR) hears allegations that a state has breached human rights provisions of the European Convention of Human Rights (ECHR). We use the dataset of Chalkidis et al. (2021), which contains 11K cases from ECtHR's public database.
124
  Each case is mapped to *articles* of the ECHR that were violated (if any). This is a multi-label text classification task. Given the facts of a case, the goal is to predict the ECHR articles that were violated, if any, as decided (ruled) by the court. The cases are chronologically split into training (9k, 2001--16), development (1k, 2016--17), and test (1k, 2017--19) sets.
125
 
126
+ To facilitate the study of the fairness of text classifiers, we record for each case the following attributes: (a) The _defendant states_, which are the European states that allegedly violated the ECHR. The defendant states for each case is a subset of the 47 Member States of the Council of Europe; To have statistical support, we group defendant states in two groups:
127
+ Central-Eastern European states, on one hand, and all other states, as classified by the EuroVoc thesaurus. (b) The _applicant's age_ at the time of the decision. We extract the birth year of the applicant from the case facts, if possible, and classify its case in an age group (<=35, <=64, or older); and (c) the _applicant's gender_, extracted from the facts, if possible based on pronouns, classified in two categories (male, female).
128
 
129
  #### scotus
130
 
131
+ The US Supreme Court (SCOTUS) is the highest federal court in the United States of America and generally hears only the most controversial or otherwise complex cases that have not been sufficiently well solved by lower courts.
132
+ We combine information from SCOTUS opinions with the Supreme Court DataBase (SCDB) (Spaeth, 2020). SCDB provides metadata (e.g., date of publication, decisions, issues, decision directions, and many more) for all cases. We consider the available 14 thematic issue areas (e.g, Criminal Procedure, Civil Rights, Economic Activity, etc.). This is a single-label multi-class document classification task. Given the court's opinion, the goal is to predict the issue area whose focus is on the subject matter of the controversy (dispute). SCOTUS contains a total of 9,262 cases that we split chronologically into 80% for training (7.4k, 1946--1982), 10% for development (914, 1982--1991) and 10% for testing (931, 1991--2016).
133
 
134
+ From SCDB, we also use the following attributes to study fairness: (a) the _type of respondent_, which is a manual categorization of respondents (defendants) in five categories (person, public entity, organization, facility, and other); and (c) the _direction of the decision_, i.e., whether the decision is liberal, or conservative, provided by SCDB.
135
 
136
  #### fscs
137
 
138
+ The Federal Supreme Court of Switzerland (FSCS) is the last level of appeal in Switzerland and similarly to SCOTUS, the court generally hears only the most controversial or otherwise complex cases which have not been sufficiently well solved by lower courts. The court often focuses only on small parts of the previous decision, where they discuss possible wrong reasoning by the lower court. The Swiss-Judgment-Predict dataset (Niklaus et al., 2021) contains more than 85K decisions from the FSCS written in one of three languages (50K German, 31K French, 4K Italian) from the years 2000 to 2020.
139
  The dataset is not parallel, i.e., all cases are unique and decisions are written only in a single language.
140
  The dataset provides labels for a simplified binary (_approval_, _dismissal_) classification task. Given the facts of the case, the goal is to predict if the plaintiff's request is valid or partially valid. The cases are also chronologically split into training (59.7k, 2000-2014), development (8.2k, 2015-2016), and test (17.4k, 2017-2020) sets.
141
 
 
145
 
146
  The Supreme People's Court of China (CAIL) is the last level of appeal in China and considers cases that originated from the high people's courts concerning matters of national importance. The Chinese AI and Law challenge (CAIL) dataset (Xiao et al., 2018) is a Chinese legal NLP dataset for judgment prediction and contains over 1m criminal cases. The dataset provides labels for *relevant article of criminal code* prediction, *charge* (type of crime) prediction, imprisonment *term* (period) prediction, and monetary *penalty* prediction. The publication of the original dataset has been the topic of an active debate in the NLP community(Leins et al., 2020; Tsarapatsanis and Aletras, 2021; Bender, 2021).
147
 
148
+ Recently, Wang et al. (2021) re-annotated a subset of approx. 100k cases with demographic attributes. Specifically, the new dataset has been annotated with: (a) the _applicant's gender_, classified in two categories (male, female); and (b) the _region_ of the court that denotes in which out of the 7 provincial-level administrative regions was the case judged. We re-split the dataset chronologically into training (80k, 2013-2017), development (12k, 2017-2018), and test (12k, 2018) sets. In our study, we re-frame the imprisonment _term_ prediction and examine a soft version, dubbed _crime severity_ prediction task, a multi-class classification task, where given the facts of a case, the goal is to predict how severe was the committed crime with respect to the imprisonment term. We approximate crime severity by the length of imprisonment term, split in 6 clusters (0, <=12, <=36, <=60, <=120, >120 months).
149
 
150
  ### Languages
151
 
 
217
  #### scotus
218
  - `text`: a `string` feature (the court opinion).
219
  - `label`: a classification label (the relevant issue area). The issue areas are: (1, Criminal Procedure), (2, Civil Rights), (3, First Amendment), (4, Due Process), (5, Privacy), (6, Attorneys), (7, Unions), (8, Economic Activity), (9, Judicial Power), (10, Federalism), (11, Interstate Relations), (12, Federal Taxation), (13, Miscellaneous), (14, Private Action).
220
+ - `respondent_type`: the type of respondent, which is a manual categorization (clustering) of respondents (defendants) in five categories (person, public entity, organization, facility, and other).
221
  - `decision_direction`: the direction of the decision, i.e., whether the decision is liberal, or conservative, provided by SCDB.
222
 
223
  #### fscs
 
229
 
230
  #### cail
231
  - `text`: a `string` feature (the factual description of the case).
232
+ - `label`: a classification label (crime severity derived by the imprisonment term).
233
  - `defendant_gender`: the gender of the defendant (Male or Female).
234
  - `court_region`: the region of the court that denotes in which out of the 7 provincial-level administrative regions was the case judged.
235
 
 
264
 
265
  The benchmark is not a blind stapling of pre-existing resources, we augment previous datasets. In the case of ECtHR, previously unavailable demographic attributes have been released to make the original dataset amenable for fairness research. For SCOTUS, two resources (court opinions with SCDB) have been combined for the very same reason, while the authors provide a manual categorization (clustering) of respondents.
266
 
267
+ All datasets, except SCOTUS, are publicly available and have been previously published. If datasets or the papers where they were introduced were not compiled or written by the authors, the original work is referenced and authors encourage FairLex users to do so as well. In fact, this work should only be referenced, in addition to citing the original work, when jointly experimenting with multiple FairLex datasets and using the FairLex evaluation framework and infrastructure, or using any newly introduced annotations (ECtHR, SCOTUS). Otherwise only the original work should be cited.
268
 
269
  #### Who are the source language producers?
270
 
 
280
 
281
  All classification labels rely on legal decisions (ECtHR, FSCS, CAIL), or are part of archival procedures (SCOTUS).
282
 
283
+ The demographic attributes and other metadata are either provided by the legal databases or have been extracted automatically from the text by means of Regular Expressions.
284
 
285
  Consider the **Dataset Description** and **Discussion of Biases** sections, and the original publication for detailed information.
286
 
287
  ### Personal and Sensitive Information
288
 
289
+ The data is in general partially anonymized in accordance with the applicable national law. The data is considered to be in the public sphere from a privacy perspective. This is a very sensitive matter, as the courts try to keep a balance between transparency (the public's right to know) and privacy (respect for private and family life).
290
  ECtHR cases are partially annonymized by the court. Its data is processed and made public in accordance with the European Data Protection Law.
291
+ SCOTUS cases may also contain personal information and the data is processed and made available by the US Supreme Court, whose proceedings are public. While this ensures compliance with US law, it is very likely that similarly to the ECtHR any processing could be justified by either implied consent or legitimate interest under European law. In FSCS, the names of the parties have been redacted by the courts according to the official guidelines. CAIL cases are also partially anonymized by the courts according to the courts' policy. Its data is processed and made public in accordance with Chinese Law.
292
 
293
  ## Considerations for Using the Data
294
 
295
  ### Social Impact of Dataset
296
 
297
+ This work can help practitioners to build assisting technology for legal professionals - with respect to the legal framework (jurisdiction) they operate -; technology that does not only rely on performance on majority groups but also considering minorities and the robustness of the developed models across them. This is an important application field, where more research should be conducted (Tsarapatsanis and Aletras, 2021) in order to improve legal services and democratize law, but more importantly, highlight (inform the audience on) the various multi-aspect shortcomings seeking a responsible and ethical (fair) deployment of technology.
298
 
299
  ### Discussion of Biases
300
 
301
+ The current version of FairLex covers a very small fraction of legal applications, jurisdictions, and protected attributes. The benchmark inevitably cannot cover "_everything in the whole wide (legal) world_" (Raji et al., 2021), but nonetheless, we believe that the published resources will help critical research in the area of fairness.
302
 
303
+ Some protected attributes within the datasets are extracted automatically, i.e., the gender and the age of the ECtHR dataset, by means of Regular Expressions, or manually clustered by the authors, such as the defendant state in the ECtHR dataset and the respondent attribute in the SCOTUS dataset. Those assumptions and simplifications can hold in an experimental setting only and by no means should be used in real-world applications where some simplifications, e.g., binary gender, would not be appropriate. By no means, do the authors or future users have to endorse the law standards or framework of the examined datasets, to any degree rather than the publication and use of the data.
304
 
305
  ### Other Known Limitations
306
 
 
342
  }
343
  ```
344
 
345
+ **Note:** Please consider citing and giving credits to all publications releasing the examined datasets.
346
 
347
  ### Contributions
348