Structure and Lineage Evidence for Vietnamese Statutory Article Retrieval on DRiLL

Duc-Vu Nguyen, Ngan Luu-Thuy NguyenarXiv 2026 · arXiv

TL;DR

This paper enhances Vietnamese legal retrieval by incorporating "legal-inspired" features—such as statutory structure and amendment lineage—into a multi-stage ranking pipeline. By moving beyond simple text similarity to model how laws cite and update one another, the authors improved the official $F_2$ score on the DRiLL benchmark from 71.77% to 72.91%, with the most significant gains occurring on complex queries requiring multiple legal articles.

What Problem Does This Paper Address?

Legal retrieval is notoriously difficult because laws are not isolated documents; they are a dense web of interconnected rules. In the Vietnamese legal system, a single query might require a "base" Article plus its subsequent amendments, related Clauses (Khoản), or specific Points (Điểm).

Standard retrieval systems often rely on semantic similarity (how much the query "looks like" the document). However, an amendment might use very different wording than the original law it modifies, causing standard systems to miss it. The authors address this gap by asking: Can we use the inherent structure and lineage of statutes to improve retrieval beyond conventional similarity scores?

Methods

The authors build upon a staged retrieval architecture (inspired by the "EDM" framework) but introduce 27 new features derived from legal domain knowledge. Their pipeline follows three main steps:

  1. Candidate Retrieval: A hybrid approach combining BM25 (lexical) and BGE-M3 (dense) embeddings, supplemented by "HyDE" (generating a hypothetical legal passage to expand the query) and citation expansion.
  2. Ranker (LightGBM): This stage uses 22 conventional features plus the 27 new legal features. These new features are grouped into:
  3. Identity & Local Form: Matching specific Article/Clause numbers mentioned in the query.
  4. Same-Law Support: Measuring if a candidate belongs to a law that already has other high-scoring candidates in the pool.
  5. Citation/Lineage Support: Using directed links to see if a candidate is cited by or amends another relevant provision.
  6. Selector (LightGBM + LLM): A final stage that uses a listwise judgment from the Qwen2.5-72B-Instruct model to evaluate the top 10 candidates collectively, rather than in isolation.
Fig. 1. Published EDM [12], internal control and proposed method. Blue marks EDM- derived stage designs, amber simplified LLM evidence, and green added legal features. EDM also uses Qwen2.5-72B-Instruct. Arrows follow the comparison protocol. Both final selectors use the Top-10 prefix of the proposed Top-100 and shared LLM evidence.

Datasets

The study uses the DRiLL (Statutory Article Retrieval in the Expansive Legal Landscape) corpus, which consists of:

  • 59,636 Articles from Vietnamese statutes.
  • 3,129 Questions total (split into training, development, and test sets).
  • Multi-Article Focus: Approximately 24.4% of the test questions require retrieving more than one distinct Article to be considered a complete answer.
Table 1. Question counts in DRiLL. Multi-Article counts are derived using at least two distinct gold Article IDs per question.

Key Insights and Ideas

  • "Same-Law" is a Strong Signal: One of the most effective features was "Same-law maximum." If the system finds one very strong match within a specific Circular or Decree, it acts as a "magnet," increasing the likelihood that other articles from that same instrument are also relevant.
  • Lineage Recovers "Hidden" Amendments: The system successfully demonstrated "companion recovery." In one case, it found a base provision about dossier requirements and used lineage cues to recover an amendment that changed the number of required dossier sets—an amendment that the baseline system missed.
  • Structural Features Reduce Noise: Features that check for "Compatible Clause/Point cues" help the system distinguish between two articles that might share keywords but belong to different legal procedures or jurisdictions.
  • LLM Efficiency: By using a single "collective judgment" for the top 10 candidates rather than dozens of pairwise comparisons, the authors significantly reduced the number of LLM calls required compared to previous state-of-the-art methods, without sacrificing accuracy.
Table 2. Features informed by legal domain knowledge.

Key Results

  • Overall Performance: The proposed method achieved a five-run mean $F_2$ of 72.91%, outperforming the internal control (71.77%) and the reported scores of other leading systems like ducanger and unknow_123.
  • Multi-Article Gains: The largest improvement was seen in questions requiring exactly two articles, where the per-query $F_2$ jumped by 4.45 percentage points.
  • Feature Importance: The added legal features accounted for 18.20% of the total decision-tree split gain in the final selector model, confirming that these domain-specific cues are actively being used by the learner to make better decisions.

Limitations and Future Work

  • Ablation Studies: The authors have not yet performed individual ablations for every single feature, meaning we don't know exactly which of the 27 features is the "silver bullet."
  • Dataset Specificity: The evaluation is currently limited to the DRiLL dataset. Cross-dataset validation is needed to see if these Vietnamese-specific structural cues translate to other legal systems.
  • Heuristic Nature: The citation and amendment links are parsed heuristically. They do not establish absolute legal validity or track complete version histories, which would be required for a production-grade legal tool.
  • Future Directions: The authors suggest exploring graph neural networks (GNNs) to represent these legal relationships more natively, rather than as flat features in a gradient-boosted tree.

Summary updated

Abstract

We study Vietnamese statutory Article retrieval on DRiLL using features for provision identity and local form, same-law support, and citation and lineage evidence. Across five random seeds, selectors sharing candidates, listwise Qwen evidence, and learner settings improve official $F_2$ from 71.77% to 72.91%, with the largest descriptive gain on two-Article questions. Added features account for 18.20% of the proposed selector's total decision-tree split gain, measured as training-objective reduction, mainly from same-law and citation evidence. Cases reveal base provisions and amendments, companion recovery, distractor removal, and failures. These findings motivate research on other datasets, representations, and learners without establishing individual feature effects or requiring LightGBM, hand-crafted scores, or explicit graphs.
BibTeX
@misc{nguyen2026structure,
  author = {Duc-Vu Nguyen and Ngan Luu-Thuy Nguyen},
  title = {{Structure and Lineage Evidence for Vietnamese Statutory Article Retrieval on DRiLL}},
  year = {2026},
  eprint = {2611.09200},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09200},
}
Figure