FairLex: A multilingual benchmark for evaluating fairness in legal text processing

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.

Ilias Chalkidis, Tommaso Passini, Sheng Zhang, Letizia Tomada, Sebastian Felix Schwemer, and Anders Søgaard. 2022. FairLex: A multilingual bench-mark for evaluating fairness in legal text processing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, Dublin, Ireland.

Pre-training details

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, SPC). 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 SPC).

Models list

Model name	Training corpora	Language
`coastalcph/fairlex-ecthr-minlm`	ECtHR	`en`
`coastalcph/fairlex-scotus-minlm`	SCOTUS	`en`
`coastalcph/fairlex-fscs-minlm`	FSCS	[`de`, `fr`, `it`]
`coastalcph/fairlex-cail-minlm`	CAIL	`zh`

Load Pretrained Model

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("coastalcph/fairlex-fscs-minlm")
model = AutoModel.from_pretrained("coastalcph/fairlex-fscs-minlm")

Evaluation on downstream tasks

Consider the experiments in the article:

Ilias Chalkidis, Tommaso Passini, Sheng Zhang, Letizia Tomada, Sebastian Felix Schwemer, and Anders Søgaard. 2022. Fairlex: A multilingual bench-mark for evaluating fairness in legal text processing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, Dublin, Ireland.

Author - Publication

@inproceedings{chalkidis-2022-fairlex,
author={Chalkidis, Ilias and Passini, Tommaso and Zhang, Sheng and
        Tomada, Letizia and Schwemer, Sebastian Felix and Søgaard, Anders},
title={FairLex: A Multilingual Benchmark for Evaluating Fairness in Legal Text Processing},
booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics},
year={2022},
address={Dublin, Ireland}
}

Ilias Chalkidis on behalf of CoAStaL NLP Group

| Github: @ilias.chalkidis | Twitter: @KiddoThe2B |