Dense Retrieval Improves Cross-Lingual Legal QA for Vietnamese Labour Law
September 24, 2026
A verifier-guided pipeline for Vietnamese-English legal QA demonstrates that dense retrieval significantly outperforms learned-sparse methods (0.358 vs 0.032 R@5). The system uses claim decomposition and citation reachability checks to prevent unsupported legal assertions.
HOW THIS AFFECTS YOU
●
builderUse dense retrieval over sparse or hybrid methods when building cross-lingual legal RAG systems for low-resource languages.
●
researcherThe study provides a new bilingual evaluation suite specifically for legal reasoning phenomena.