Open-weight LLMs
Evaluated across models and three health-misinformation datasets.
← All publications
Under review
Health Misinformation Refutation with Dual Knowledge Graphs
A dual-path knowledge-graph framework for generating factual refutations at an appropriate explanatory depth.
The problem
Why this work matters
Health explanations differ in more than readability. They also vary in concept density, causal depth, and abstraction. Existing approaches rarely control those dimensions while maintaining strong factual grounding.
Methodology
How the approach works
01Target granularity
02Layered ExplanationGraph
03Shared GroundingGraph
04Dual-path retrieval
05Structured refutation
Contributions
What this work contributes
- 01
A multidimensional definition of explanation granularity using concept density, causal depth, and abstraction grain.
- 02
A layered ExplanationGraph spanning expert, scientific, and foundational biomedical knowledge.
- 03
A separate GroundingGraph that preserves authoritative factual evidence across audience levels.
- 04
Hierarchical dual-path retrieval that connects level-specific mechanisms with shared grounding evidence.
Evaluation
Key results
FActScore gain
Highest factual precision across the evaluated generators.
Reader clarity
Average understandability in the reader study.
Judge agreement
Quadratic weighted kappa against human granularity labels.