关键词: 大语言模型, 类案检索, 智慧司法, 文本匹配

Abstract: Abstract: With the advancement of intelligent justice initiatives, similar case retrieval technology has garnered significant attention for its crucial role in ensuring judicial fairness and efficiency. However, existing retrieval systems still face the following challenges: traditional models were easily interfered with by semantic structural similarity, making it difficult to accurately capture elements influencing judgments; pre-trained language models were constrained by input length limitations, leading to insufficient global semantic modeling of lengthy legal texts; existing aggregated similarity scoring mechanisms were susceptible to noise and lacked strong interpretability. To address these issues, a similar case retrieval model integrating legal element enhancement and multi-granularity interaction was proposed, focusing on improvements at three levels: case element extraction, deep text encoding, and score aggregation. Firstly, to tackle text redundancy and element omission in legal documents, a hierarchical case element extraction module based on large language models (LLMs) was introduced. It summarized case sub-facts according to charge classifications, effectively preserving core case facts while reducing noise interference. Secondly, to resolve the problem of deep dependency in legal text encoding, the SFA-SAILER encoding architecture was designed. The architecture captured cross-hierarchical dependencies between case facts and other sections via SAILER, and a feature-level attention mechanism was incorporated at the CLS representation point, enabling deep encoding of case information at both word and feature dimensions. Finally, the MaxSim operator was employed to aggregate similarity scores between case sub-facts. Experimental results demonstrated that compared to numerous existing models, the proposed model achieved MAP and P@3 scores of 67.45 and 60.95, respectively, on the LeCaRD dataset. NDCG@K metrics also consistently surpassed those of other models. This study provides a new approach to similar case retrieval that balances legal logic with deep semantic understanding, offering practical value for advancing judicial intelligence.

Key words: Large Language Model, Legal Case Retrieval, Intelligent Judiciary, Text Matching