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.