Navigating Legal Case Matching via Heterogeneous Graph Learning and Matching Path Decomposition

ACL ARR 2026 January Submission3000 Authors

Keywords: Legal Case Matching, Heterogeneous Graph, Path Decomposition

Abstract: Legal case matching (LCM) endeavors to determine the relevance between query cases and target cases, which plays a pivotal role in supporting legal decisions. In legal practice, query cases typically contain only fact descriptions, while target cases, being historical cases, often include additional case analysis that provides a new perspective for the LCM task beyond semantic similarity. In statutory law systems (e.g., China), such analysis relies on law article interpretation, while in case law systems (e.g., US, UK), it relies on precedent case references. Based on these observations, we propose a relation-driven framework called RedMatch, under which target cases are intrinsically connected to one another and associated with cited laws. First, it constructs a global heterogeneous graph for all target cases to extract case-case and case-law relations. Then, a graph transformer integrates these relations in the matching prediction model to enrich the case representation. Finally, a path learning task is designed to navigate the model to decompose multiple matching paths to reach target cases by leveraging these relations. RedMatch also introduces a law article matching task via multitask learning to align LCM outcomes and enhance the method's versatility. Experiments on three publicly available datasets, including Chinese and English languages, demonstrate state-of-the-art performance of RedMatch, highlighting its effectiveness and generalizability.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: NLP Applications, Information Retrieval and Text Mining

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: Chinese, English

Submission Number: 3000