Presentation Information
[4Yin-B-52]Extractive Summarization of Legal Decisions using Retrieval-Augmented Generation with Similar Case Search
〇Koki Okano1, Fumikatsu Anaguchi1, Takeshi Morita1 (1. Aoyama Gakuin University)
Keywords:
Extractive Summarization
In legal and tax practice, summaries have become increasingly important for quickly and accurately grasping the content of legal decisions. However, creating summaries requires specialized expertise and is highly dependent on individual skills, making human cost reduction a challenge. While automation of document summarization using large language models (LLMs) has advanced in recent years, traditional LLMs have struggled with legal documents. This is because legal documents cannot tolerate misinterpretations of factual relationships and require extractive generation that faithfully preserves the original text.Therefore, this study proposes an extractive summarization method for legal decisions based on Retrieval-Augmented Generation (RAG) with similar case search. It controls extraction accuracy and generation format by searching for similar past cases and presenting them to the LLM.To validate this method's effectiveness, we comparatively evaluated search methods, prompt strategies, and LLM selection. Results showed that a hybrid search combining vector similarity with metadata (character count, number of issues, tax category, ruling outcome) and a Chain of Thought prompt using one similar case demonstrated high extraction accuracy. Furthermore, among major LLMs, Claude Opus 4.5 demonstrated the best performance, indicating its applicability in practical use.