The PRecG pipeline improves legal precedent retrieval by decomposing documents into semantic units based on rhetorical roles rather than treating them as monolithic texts.

  • It constructs knowledge graphs for each rhetorical segment to capture legal entities and relationships.
  • Contextual entity representations are aggregated to create segment-level embeddings.
  • These embeddings are integrated to produce a unified document-level representation for similarity computation.
  • The approach was validated on a benchmark Indian legal dataset against state-of-the-art baselines.