DiscoLQA: zero-shot discourse-based legal question answering on European Legislation

Abstract

The structures of discourse used by legal and ordinary languages share differences that foster technical issues when applying or fine-tuning general-purpose language models for open-domain question answering on legal resources. For example, longer sentences may be preferred in European laws (i.e., Brussels I bis Regulation EU 1215/2012) to reduce potential ambiguities and improve comprehensibility, distracting a language model trained on ordinary English. In this article, we investigate some mechanisms to isolate and capture the discursive patterns of legalese in order to perform zero-shot question answering, i.e., without training on legal documents. Specifically, we use pre-trained open-domain answer retrieval systems and study what happens when changing the type of information to consider for retrieval. Indeed, by selecting only the important parts of discourse (e.g., elementary units of discourse, EDU for short, or abstract representations of meaning, AMR for short), we should be able to help the answer retriever identify the elements of interest. Hence, with this paper, we publish Q4EU, a new evaluation dataset that includes more than 70 questions and 200 answers on 6 different European norms, and study what happens to a baseline system when only EDUs or AMRs are used during information retrieval. Our results show that the versions using EDUs are overall the best, leading to state-of-the-art F1, precision, NDCG and MRR scores.