MT-SAL: Multi-Task Structure-Aware Learning for Legal Document Summarization

The objective of the automatically summarizing of legal documents is to reduce the workload of legal professionals. However, current methods face two main issues: i) Error accumulation in extractive-abstractive approaches, where errors during extraction adversely impact the quality of the generated summaries, and ii) Neglect of the highly structured nature of legal documents. To tackle these problems, this paper propose a Multi-Task Structure-Aware Learning framework (MT-SAL) for long legal documents summarization. MT-SAL tackles error accumulation by adding a sentence importance classification auxiliary task to simplify long legal documents. Additionally, we introduce a document reordering auxiliary task to help the model deeply comprehend the structured nature of legal texts. Jointly training auxiliary tasks with the main summarization task enables efficiently generate high-quality summaries. Experiments shows MT-SAL surpasses SOTA baselines by significant margin.