PRecG uses GNNs and rhetorical roles for legal precedent retrieval

Legal precedent retrieval is critical for case preparation and litigation strategy, but current methods treat documents as monolithic texts, missing the rhetorical organization that gives different legal entities and concepts varying significance. A new paper from Devanshu Verma, Vasudha Bhatnagar, Vikas Kumar, and Balaji Ganesan introduces PRecG (Precedent Retrieval with Graph Neural Networks), a pipeline that addresses this gap by first segmenting legal judgments based on the rhetorical roles of sentences (e.g., facts, arguments, ratio decidendi). For each segment, a knowledge graph captures legal entities and their relationships, allowing context-sensitive representations to be learned and aggregated into segment-level embeddings. These are then combined into a unified document embedding, and similarity between precedent pairs is computed hierarchically.

Validated on a benchmark Indian legal dataset, PRecG demonstrates consistent improvements over state-of-the-art baselines, proving that incorporating rhetorical structure and graph-based entity modeling enhances retrieval accuracy. The approach is particularly powerful because legal reasoning often hinges on the role a statement plays within a judgment—not just its literal meaning. By preserving this structure, PRecG could meaningfully improve legal research tools, reduce manual review time, and help lawyers and judges find more relevant precedents. The paper (arXiv:2607.09094) is 23 pages and represents a significant step toward AI systems that truly understand legal texts.