ViMELD Legal Retrieval
Multi-View Vietnamese Legal Retrieval with Citation GATs & Uncertainty-Aware Ranking
A 3-stage multi-view Vietnamese legal document retrieval framework integrating lexical BM25L, Vietnamese dense embeddings, and Graph Attention Networks (GAT) over statutory citation hierarchies with uncertainty-aware meta-learning and decoupled LLM late fusion, achieving 0.9765 Recall@30.
The Technical Bottleneck
Vietnamese legal statutes form a deeply interconnected hierarchy of articles, clauses, amendments, and cross-references. Purely lexical search misses paraphrased citizen queries, while purely dense retrieval ignores statutory citation authority and suffers from bias when LLMs are used directly as black-box judges.
Why Existing Baselines Fail
Building directly on our FJCAI 2026 diagnostic study of Vietnamese legal LLMs—which revealed that models frequently fail on cross-reference resolution and statutory exceptions—ViMELD was engineered to supply high-recall, structurally grounded statutory evidence before any generative step.
Core Formulation & Insight
ViMELD combines three complementary retrieval views: (1) BM25L for exact statutory terminology, (2) Vietnamese dense bi-encoder embeddings for semantic matching, and (3) a Graph Attention Network (GAT) operating over the legal citation and hierarchy graph. An uncertainty-aware meta-learner aggregates multi-view candidates, followed by a decoupled late-fusion LLM scoring stage that prevents the LLM from acting as an uncalibrated direct judge.
Architectural Design
Three-stage multi-view retrieval pipeline uniting lexical, dense semantic, and graph-topological views with uncertainty-aware late fusion.
Motivating error analysis from our FJCAI 2026 Vietnamese Legal LLM study showing why structured retrieval is critical to prevent statutory omission and misinterpretation.
1. Tri-View Candidate Generation (BM25L + Dense + Citation GAT)
Retrieves statutory candidates simultaneously across lexical BM25L, Vietnamese dense embeddings, and a Graph Attention Network encoding cross-article citation links and legislative hierarchy.
2. Uncertainty-Aware Meta-Learner Ranking
Calibrates and fuses relevance scores and predictive uncertainties across the three views to handle queries ranging from exact article lookups to abstract layperson descriptions.
3. Decoupled Late-Fusion LLM Scoring
Integrates LLM semantic scoring in a decoupled late-fusion step rather than raw LLM-as-a-judge reranking, preserving ranking determinism and auditability.
Technical Stack & Systems Engineering
Implementation Highlights
- →Constructed a statutory citation and hierarchy graph linking Vietnamese legal codes, articles, and cross-references.
- →Designed an uncertainty-aware meta-learner and decoupled LLM late-fusion module that avoids direct LLM-judge hallucinations.
Datasets & Protocol
Datasets & Corpora
- Vietnamese Legal Document Retrieval Benchmark & Statutory Citation Corpus
Evaluation Protocol
Evaluated single-view baselines (BM25L, Dense only) against the multi-view BM25L + Dense + GAT configuration with uncertainty-aware ranking.
Tracked Evaluation Metrics
Reported Recall@30 of 0.9765 in the multi-view configuration, providing high-coverage statutory evidence for downstream Vietnamese legal reasoning.
Reported on the multi-view (BM25L + Dense + GAT) configuration
Lexical (BM25L) + Semantic (Dense) + Structural (Citation GAT)
Empirical Analysis & Interpretation
Fusing citation graph topology with lexical and dense views captures cross-referenced exceptions that standalone text encoders miss.
What This System Contributes
- 01Architected and implemented the public ViMELD-Legal-Retrieval codebase including the BM25L, dense, GAT, and uncertainty-aware fusion modules.
Boundary Conditions & Future Work
- ·Recall@30 of 0.9765 is reported for the primary multi-view configuration in the repository documentation; individual ablation splits vary by query type.
From Benchmarking to Reasoning: A Dual-Aspect, Large-Scale Evaluation of LLMs on Vietnamese Legal Text
Van-Truong Le