Enhancing Legal Judgment Retrieval by Re-Ranking Based on Contrastive Implicit Factors

Peng-Yi Lin, Yao-Chung Fan, Fang-Yie Leu

Abstract: Legal case retrieval has always been a crucial issue in the field of law. When handling new cases, legal experts need to refer to similar past cases to gain meaningful insights and possible verdicts. However, retrieving similar cases consumes a significant amount of time. Current retrieval methods mainly focus on case scenarios, but actual verdicts show that cases with similar scenarios do not always have similar judgments. This indicates that, apart from the explicitly defined factors in the law, judges are influenced by Implicit Factors that are not clearly defined. To address this, we propose a prompt strategy based on Summary Steered Reasoning with Distinguishing (SSRD), which uses LLM to analyze legal cases and extract these Implicit Factors. It then calculates similarity scores between retrieved cases and the Implicit Factors to reorder the retrieval results. We validated our approach using the Taiwan judgment document dataset Lawsnote and the Chinese legal case retrieval dataset LeCaRD. Experimental results show that our method outperforms the baseline: in the Lawsnote dataset, our method reduces RMSE@5 by 20% compared to the state-of-the-art method; in the LeCaRD dataset, our method improves NDCG@5 from 0.787 to 0.794. This demonstrates that our approach can significantly improve the effectiveness of existing legal case retrieval systems.

External IDs:dblp:conf/bwcca/LinFL24