LSDK-LegalSum: improving legal judgment summarization using logical structure and domain knowledge

Legal judgment summarization enhances the efficiency and accuracy of processing and retrieving similar cases by distilling key information from legal judgment documents. Legal judgments, which record the trial process and outcome, tend to have lengthy texts. However, most existing methods for legal judgment summarization struggle to effectively handle long texts and focus only on the serialized text information, neglecting the auxiliary role of the document’s logical structure in generating summaries. Additionally, legal judgment documents contain extensive domain knowledge formed by professional legal terminology, which is crucial for accurately understanding and analyzing the content of the documents. We propose a method to improve the quality of legal judgment summaries by utilizing logical structures and domain knowledge (LSDK-LegalSum). Specifically, we first use a multi-label document classification approach to divide the legal judgment into four parts: ‘Type’, ‘Claim’, ‘Fact’ and ‘Result’, then generate a summary of the ‘Type’ part, and combine it with the ‘Claim’ part to create a new summary, and so forth until the final summary is produced. Moreover, the case descriptions in legal judgments cause substantial textual noise; therefore, we employ a text-denoising method that integrates domain knowledge to further enhance the performance of the summaries. Extensive experiments on the CAIL2020 legal summarization dataset have shown that our method achieves improvements of 8.37, 16.64, and 13.07 percentage points in Rouge-1, Rouge-2, and Rouge-L scores respectively over the state-of-the-art models. This demonstrates that our method not only produces concise and semantically complete summaries but also ensures that the content is logically rigorous and well-organized, meeting practical needs more effectively.