ViLegalExpert: A Large-Scale Benchmark for Vietnamese Legal Retrieval and Question Answering from Real-World Consultations

TL;DR

ViLegalExpert is a large-scale Vietnamese legal benchmark derived from real citizen-lawyer consultations, featuring over 172,000 questions across 34 legal domains, along with expert-verified answers and legal evidence. It supports multiple tasks including retrieval, extractive, and abstractive question answering, highlighting significant challenges in grounding responses in authoritative legal sources.

Why It Matters

It enables the development of trustworthy legal AI systems in Vietnamese by providing realistic, expert-validated data that reflects actual legal inquiries. This benchmark helps improve the reliability of AI tools used in legal advice, especially in regions where access to legal expertise is limited.

Key Findings

• Pretrained language models perform well on question answering tasks but struggle with generating answers grounded in legal evidence.
• Hybrid retrieval methods outperform pure neural or keyword-based approaches in locating relevant legal provisions.
• Mapping natural language legal questions to authoritative legal texts remains a major challenge, indicating a need for better semantic alignment techniques.