Skip to content
Open access

CECRA: Structured Evidence-Aware Reranking for Long-Document Legal Case Retrieval

Oct 2026 · Applied Sciences · 0 citations

Abstract

Legal case retrieval requires identifying relevant cases from lengthy judicial documents based on factual and legal similarities. Existing methods often match cases through global semantic representations. When candidates are semantically similar, global representations may underrepresent discriminative factual details and judicial reasoning. We propose CECRA, a structured evidence-aware reranking method that uses Qwen3-32B to construct Summary, Fact Evidence, and Reasoning Evidence offline. CECRA combines global Summary matching with fixed, charge-independent Fact Evidence and Reasoning Evidence via validity-masked slot-level local interaction; Agreement remains a secondary matching component with auxiliary supervision. On LeCaRDv2 under the fixed-candidate-pool evaluation protocol, mean average precision (MAP), averaged over three seeds, was 72.95 for CECRA, 70.72 for fine-tuned SAILER (SAILER-FT), and 67.32 for KELLER. Jointly removing both local views and associated Agreement calculations reduced MAP by 2.24 percentage points, consistent with a contribution from structured local evidence but not isolating Agreement’s independent effect. Pairwise diagnostics showed the largest gains for candidates with similar global semantics but different local evidence. These results indicate that local structured evidence can complement case-level semantic representations in distinguishing highly similar candidates.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.