Researchers propose PRecG pipeline for legal precedent retrieval using rhetorical role segmentation and graph neural networks

Legal precedent retrieval is a core task in legal case preparation, planning, litigation strategy, and legal research. Prior approaches typically embed entire legal documents into low-dimensional semantic spaces and compute similarity based on proximity of these representations. These methods, however, treat documents as monolithic texts and thus overlook the rhetorical organization of legal technicalities, missing nuanced legal meanings and the contextual significance of entities and concepts that vary by their rhetorical roles.

The proposed PRecG pipeline addresses this by hierarchically learning representations of legal judgments. It first decomposes each document into distinct semantic units (segments) based on the rhetorical roles of sentences. For each segment, a knowledge graph is constructed to capture legal entities and their relationships within that segment. Contextual representations of entities are learned and aggregated to produce segment-level embeddings, which are then integrated into a unified document-level representation. Finally, the semantic similarity between a pair of documents is computed.

The authors validate PRecG through extensive experiments on a benchmark Indian legal dataset, comparing it against state-of-the-art baselines to demonstrate its effectiveness. The paper positions PRecG as an advancement over prior monolithic-text retrieval methods by explicitly modeling rhetorical structure and legal entity relationships.