[Tech Report] Structuring the Law: Leveraging CRFs for Vietnamese Legal Logic Recognition

TL;DR

Researchers have developed a robust framework for automatically identifying the logical building blocks of Vietnamese legal text. By moving beyond fragile rule-based systems to Conditional Random Fields (CRFs), the study demonstrates that combining basic linguistic features with "semantic triggers" significantly improves the accuracy of identifying legal assumptions and provisions, reaching a peak F1-score of 73.12%.

Background: The Complexity of Legal Engineering

In the realm of "Legal Engineering," the first hurdle is teaching a machine to read. Legal sentences in Vietnam—much like in other jurisdictions—are notoriously long and nested. However, they possess a hidden, rigid structure. Every law article typically consists of:

• Assumption: The subject and the condition (e.g., "In the case of...").
• Provision: The rights or duties involved (e.g., "...must perform...").
• Sanction: The penalty for non-compliance.

Previous attempts to automate this used manual rules (JAPE language), which were too rigid to handle the creative ways legal draftsmen can phrase a condition.

Methodology: Sequence Labeling with Semantic Insight

The researchers treated the law as a stream of tokens, where each word is assigned a tag following the IOB2 (Inside, Outside, Beginning) format.

1. The Architecture

The system follows a classic Machine Learning pipeline: Pre-processing (handling problematic Vietnamese characters) -> Feature Engineering -> CRF Training -> Evaluation.

Fig 1. The system architecture shows the transition from raw XML legal documents to a trained CRF model.

2. Feature Engineering: The Secret Sauce

To bridge the gap between simple grammar and legal logic, the authors used four levels of features:

• WORD & POS: Standard NLP features using a 5-word sliding window.
• TRIGGER: A hand-curated list of 140 keywords. For instance, "Nếu" (If) is a strong signal for an Assumption, while "là" (is) often denotes a Definition.
• ONTOLOGY: Integrating 299 legal terms to help the model recognize specialized entities.

Key Results & Performance

The experiments conducted on a Vietnamese Business Law dataset showed that the machine learning approach is significantly more flexible than manual rules.

Fig 2. Performance comparison highlights the significant gain in recognizing Assumptions and Provisions over legacy systems.

Why did it work?

• Contextual Awareness: Unlike rules that look for exact matches, CRFs look at the probability of a tag based on the surrounding words.
• Semantic Triggers: The inclusion of trigger features acted as "highway signs" for the model, pointing it toward the start of new logical sections even when the sentence structure was non-standard.

Critical Analysis & Limitations

While the system succeeded in Term Recognition (reaching nearly 100% precision due to the highly consistent style of definitions), it struggled with Sanctions. This was largely due to data scarcity—only 3.3% of the corpus was tagged as a "Sanction."

Conclusion and Future Outlook

This work marks a significant step forward in Legal Engineering in Vietnam. It demonstrates that combining machine learning (CRFs) with domain keywords (Triggers) provides a better tool for legal analysis than rigid human-defined rules. Future research could explore:

• Deep Learning (LLMs): Moving to a Transformer-based model (e.g., LegalBERT tailored for Vietnamese) to capture even longer-range dependencies.
• Hybrid Models: Combining the precision of rules (for definitions) with the flexibility of CRFs (for complex provisions).

By Nguyen Truong Son, et al. (Legal Engineering Project, VNU-HCM).