Enhanced question understanding for multi-type legal question answering

Multi-type Legal Question Answering(MLQA) aims to automatically respond to legal questions presented in natural language. Current MLQA models generally include a text reading component and an answer prediction component. However, these models often prioritize handling lengthy legal documents over closely analyzing the given question, which can lead to suboptimal answers. Our observations reveal that while existing MLQA models achieve satisfactory F1 scores in single-span extraction, their performance significantly declines in multi-span extraction, particularly in terms of Exact Match (EM) accuracy when multiple ordered answer spans are required. To address this limitation, we introduce a novel, adaptable question-refining reading module. This module, compatible with various answer prediction components, enhances MLQA models’ focus on the question content, improving answer accuracy. Experiments on general question-answering datasets demonstrate that pre-trained language model-based methods substantially enhance performance in question-answering tasks. Furthermore, experimental results on the CAIL 2021 Law MRC dataset, the largest one in MLQA, show that our module significantly outperforms the existing state-of-the-art models, by 1.2- \(-\) $-$ 10.4% in EM, with competitive F1 scores.