Refined Question Reading for Multi-Type Legal Question Answering

Multi-type Legal Question Answering(MLQA) aims to automatically answer legal questions described in natural language. Existing MLQA models typically consist of a text reading component and an answer prediction component. However, their text reading often focuses on how to handle long legal documents, and pays few attention to the given question, potentially leading to suboptimal answering performance. Our observation indicates that although current MLQA models may perform well for single-span extraction in terms of F1, their effectiveness in multi-span extraction becomes particularly worse in terms of Exact Match(EM) where multiple answer spans are required with fixed span number and ordering. To address this issue, we propose a new, pluggable, question-refining reading module. This module is compatible with diverse answer prediction components and assists different MLQA models in better focusing on the question content. Consequently, it enhances the prediction accuracy of MLQA. Experimental results on the CAIL 2021 Law MRC dataset, the largest one in MLQA demonstrate that our module significantly outperforms the existing state-of-the-art models, by 1.2% to 10.4% in terms of EM with competitive F1 values.