New features improve Vietnamese statutory article retrieval accuracy

Researchers improved the retrieval of Vietnamese statutory articles on the DRiLL dataset by incorporating lineage and citation evidence.

The study used features regarding provision identity, local form, same-law support, and citation and lineage evidence to refine the retrieval process. By using selectors that share candidates, listwise Qwen evidence, and specific learner settings, the team increased the official F2 score from 71.77% to 72.91%. The most significant descriptive gains occurred on questions involving two articles.

The added features accounted for 18.20% of the total decision-tree split gain, primarily driven by same-law and citation evidence. The results identified cases of base provisions, amendments, and companion recovery while also noting instances of distractor removal and retrieval failures. These findings suggest that similar methods could be applied to other datasets and representations.