💬 NLP

ViHERMES: A Graph-Grounded Multihop Question Answering Benchmark and System for Vietnamese Healthcare Regulations

This paper introduces ViHERMES, a new benchmark and graph-aware retrieval system designed to address the challenges of multihop question answering over complex, interdependent Vietnamese healthcare regulations.

Original authors: Long S. T. Nguyen, Quan M. Bui, Tin T. Ngo, Quynh T. N. Vo, Dung N. H. Le, Tho T. Quan

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Long S. T. Nguyen, Quan M. Bui, Tin T. Ngo, Quynh T. N. Vo, Dung N. H. Le, Tho T. Quan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to follow a complex recipe for a traditional Vietnamese dish, but there’s a catch: the recipe book is constantly being updated. One page says "use 2 eggs," but a sticky note on a different page says "actually, use 3 eggs for all recipes written before 2023," and another page says "if you are making this for a wedding, refer to the 'Special Occasion' manual for seasoning rules."

If you just read the first page, your dish will be a disaster. You need to "hop" between different pages, follow the notes, and understand how they connect to get the right answer.

This is exactly what the researchers behind ViHERMES are tackling, but instead of recipes, they are dealing with Vietnamese healthcare laws.

The Problem: The "Legal Maze"

Healthcare regulations are not like a simple Wikipedia article where you read one paragraph and get the answer. They are a massive, interconnected web. Laws change, new rules replace old ones, and one law often points to another law for more details.

For a computer (or even a human), answering a question like "What are the current rules for hospital pricing during this transition period?" is incredibly hard because the answer isn't in one place. You have to:

Find the original rule.
Find the amendment that changed it.
Find the rule that explains how to handle the "transition."

Currently, most AI systems are like someone looking at a single snapshot of a map—they see where things are now, but they don't understand the roads connecting them or how the map has changed over time.

The Solution: Two Big Innovations

The researchers did two main things to fix this:

1. They built a "Training Ground" (The ViHERMES Dataset)

Before you can teach a student to solve complex puzzles, you need to give them high-quality puzzles. The researchers created ViHERMES, a massive collection of "multihop" questions and answers.

The "Multihop" part: These aren't simple questions. They are designed to force the AI to jump from one document to another to find the truth.
The "Grounding" part: Every answer comes with "receipts"—the exact legal text used to reach that conclusion—so we can make sure the AI isn't just making things up (hallucinating).

2. They built a "Smart Detective" (The Graph-Aware System)

Instead of a standard AI that just searches for keywords, they built a system that thinks like a legal expert. They call this a Structure-Driven Knowledge Graph.

Think of it like this:

Standard AI (Naive RAG): Is like a librarian who can find you any book if you give them a title, but doesn't know what's inside the books.
The ViHERMES System: Is like a Master Detective with a giant corkboard and red string.
- When you ask a question, the detective finds the first "clue" (the initial law).
- Then, they follow the "red strings" (the legal connections). If a law says "This replaces Law X," the detective follows that string to the new law. If a law says "See Article 5 for definitions," they follow that string too.
- Finally, they use a "Team of Specialists" (Multi-Agent System) to double-check the work: one agent finds the clues, one agent follows the strings, and one "Auditor" agent makes sure no one made a mistake before giving you the final answer.

Why does this matter?

In healthcare, being "mostly right" isn't good enough. If a hospital follows an outdated rule because an AI gave them the wrong information, people could be harmed or legal trouble could follow.

By teaching AI to understand the structure and connections of laws—rather than just reading them as flat text—this research moves us closer to digital assistants that can actually be trusted to navigate the complex world of healthcare regulations.

Technical Summary: ViHERMES

Title: ViHERMES: A Graph-Grounded Multihop Question Answering Benchmark and System for Vietnamese Healthcare Regulations

1. Problem Statement

Question Answering (QA) within regulatory domains, particularly healthcare, is exceptionally difficult due to the multihop reasoning required to navigate complex legal structures. Unlike general text, healthcare regulations are:

Hierarchically Structured: Organized into specific units (articles, clauses).
Interdependent: Frequently interconnected through formal legal relations such as amendments, replacements, supplements, and cross-references.
Dynamic: Subject to frequent revisions that change the validity of previous provisions.

Existing Retrieval-Augmented Generation (RAG) methods often treat documents as "flat" collections of text, failing to capture these formal dependencies. Furthermore, there is a significant lack of high-quality benchmarks for multihop reasoning in low-resource languages like Vietnamese, especially within the specialized intersection of healthcare and law.

2. Methodology

The authors propose a two-pronged solution: a new benchmark dataset and a specialized graph-aware system.

A. ViHERMES Dataset Construction
The authors developed a controlled pipeline to ensure high-quality, multihop reasoning:

Semantic Clustering: Using Spherical K-Means on $\ell_2$ -normalized embeddings to group regulatory contexts into coherent semantic clusters.
Graph-Inspired Sampling: Selecting representative contexts from different document titles to ensure cross-document reasoning.
LLM-based Generation: Using Large Language Models to generate question-answer pairs that require $h$ hops of reasoning (e.g., tracing an amendment chain or combining definitions from separate laws).
Structural Validation: Human annotators and experts verify the legal correctness, reasoning paths, and evidence grounding.

B. Graph-Aware Multi-Agent System
To solve the QA task, the authors introduce a system built on a Structure-driven Regulatory Knowledge Graph (SRKG):

SRKG Modeling: Unlike entity-centric graphs, the SRKG models legal units (Documents, Articles, Clauses) as nodes and formal legal relations (e.g., AMENDS, REPLACES, REFERS_TO) as typed edges.
Seeded Retrieval & Propagation: The system first identifies "seed" nodes via hybrid retrieval (BM25 + Dense). It then performs relation-aware propagation:
- Validity Tracing: Recursively following amendment chains to find the most current law.
- Contextual Supplementation: Expanding via reference relations to provide necessary legal context.
Multi-Agent Architecture:
- Interpreter: Analyzes query intent and routes the task.
- Pathfinder: Executes the graph-based retrieval and propagation.
- Auditor: Verifies the retrieved evidence against the generated claims to prevent hallucinations.
- Conductor: Orchestrates the agents and generates the final natural language response.

3. Key Contributions

ViHERMES Benchmark: The first dataset specifically designed for multihop QA over Vietnamese healthcare regulations, featuring explicit reasoning and evidence annotations.
SRKG Framework: A novel approach to representing legal documents as a structured graph based on formal drafting conventions rather than mere entity extraction.
Multi-Agent Graph-RAG System: A specialized system that integrates structural legal knowledge with agentic workflows to ensure legal validity and coherence.

4. Experimental Results

The system was evaluated against several baselines, including Naive RAG (BM25, Dense, Hybrid), reasoning-guided methods (IRCoT), and advanced graph-RAG frameworks (RAPTOR, LightRAG, HippoRAG 2).

Superior Accuracy: The proposed system outperformed all baselines across all metrics: F1 score (0.8334), LLM-as-a-Judge (0.7554), and Recall@5 (0.8461).
Efficiency: While graph-based methods are more computationally expensive than naive RAG, the proposed system maintained a competitive average latency (~14.74s), performing better than HippoRAG 2 while using fewer graph tokens.
Ablation Insights: The Interpreter (intent routing) and Auditor (verification) were found to be critical components for maintaining reasoning reliability and preventing errors.

5. Significance

This work addresses a critical gap in AI for legal and healthcare applications. By moving from "flat" retrieval to "structure-aware" reasoning, the authors demonstrate that modeling the formal logic of law is essential for building trustworthy AI in regulated industries. The methodology is highly transferable to other legal domains (e.g., tax, labor, or education) where hierarchical and interdependent document structures are the norm.