ReCaRe: A Bilingual Benchmark for Revision Candidate Retrieval in Evolving Legal Corpora
Abstract
Document corpora in organizational and public settings evolve as statutes, policies, specifications, and knowledge bases are revised. Maintaining such corpora requires identifying documents that should be inspected after a change, yet this candidate-identification step has not been directly addressed by existing benchmarks. We formalize this problem as Revision Candidate Retrieval (RCR): retrieving documents for expert inspection and possible revision. RCR consists of two complementary tasks: Rat2Rev, which retrieves articles given an amendment rationale, and Rev2Rev, which retrieves co-revised articles given an observed revision. To study RCR, we introduce ReCaRe, a bilingual benchmark built from EU and Japanese legislative amendments, containing 91, 361 EU articles and 90, 170 Japanese articles. Our analysis shows low lexical overlap, length asymmetry, and implicit cross-article revision relationships. Experiments show that BM25 remains a strong first-stage baseline, long-context dense retrievers improve recall, zero-shot listwise LLM reranking improves Rat2Rev rankings, and document expansion and dense-retriever domain adaptation provide complementary first-stage gains.