KNowNEt:Guided Health Information Seeking from LLMs via Knowledge Graph Integration

Youfu Yan, Yu Hou, Yongkang Xiao, Rui Zhang, Qianwen Wang

TL;DR

KNOWNET is a visualization system that anchors the conversational power of Large Language Models (LLMs) to the factual bedrock of Knowledge Graphs (KGs). By extracting structured "triples" from LLM responses and validating them against peer-reviewed literature, it transforms potentially unreliable chat sessions into a guided, evidence-based exploration of complex health topics.

What Problem Does This Paper Address?

While LLMs like GPT-4 are increasingly used for health information seeking, they face two major hurdles: hallucinations and linear information delivery. In the high-stakes domain of medical advice, an LLM might confidently state a falsehood (hallucination) or provide a "wall of text" that obscures the complex, non-linear relationships between drugs, diseases, and biological processes. Furthermore, users often suffer from "exploration blindness"—they don't know what follow-up questions to ask to gain a comprehensive understanding of a topic.

Methods

KNOWNET addresses these gaps through a multi-layered integration of LLMs and KGs:

  1. Triple Extraction & Matching: When an LLM generates a response, KNOWNET identifies entities and their relations (e.g., [Vitamin E] -> [prevents] -> [Alzheimer’s]). It then uses OpenAI’s embedding models to map these entities to standardized nodes in a Knowledge Graph.
  2. Automated Validation: The system checks the KG for evidence. It classifies relations into three categories:
  3. Support (✓): Direct evidence exists in the KG.
  4. Relevant (i): No direct link, but a two-hop path or a similar relation exists (e.g., the LLM says "prevents" while the KG says "slows").
  5. Unsure (?): No supporting evidence found in the KG.
  6. Progressive Visualization: Instead of a static chart, KNOWNET uses a "focus+context" design. As the conversation progresses, it builds a node-link diagram where new information is highlighted and previous steps are faded, allowing users to track their "exploration path."
  7. Recommendation Engine: Based on the user's current position in the KG, the system suggests "next-step" queries to ensure they don't overlook critical related topics (like side effects or physiological mechanisms).

Datasets

The researchers evaluated KNOWNET using ADInt, a specialized Knowledge Graph focused on Alzheimer’s Disease and non-pharmaceutical interventions.

  • ADInt Stats: 162,212 nodes (drugs, diseases, symptoms) and over 1 million edges derived from 754,224 scientific papers.
  • Terminology: The system utilizes the UMLS (Unified Medical Language System) to ensure standardized medical vocabulary across the LLM and KG.

Key Insights and Ideas

  • The "Truth" Skeleton: The paper treats the Knowledge Graph as a "scaffold" for the LLM. The LLM provides the conversational nuance and context, while the KG provides the structural "truth" that can be audited via PMIDs (PubMed IDs).
  • Non-Binary Validation: One of the most interesting ideas is the "Relevant" label. It acknowledges that LLMs and KGs often use different granularities of language. A drug that "treats" a disease in an LLM's eyes might "coexist with" or "affect" a related pathway in a KG; KNOWNET captures these "near-miss" validations rather than just marking them as wrong.
  • Cognitive Alignment: By visualizing the search as a graph, the system aligns with symbolic cognitive modeling—the idea that humans understand concepts by building networks of related entities.

Key Results

The authors conducted 15 use cases and interviews with medical informatics experts. Key findings include:

  • Trust Calibration: Experts noted that being able to click an edge and see the specific scientific paper (PMID) behind a claim "dramatically shifted" their perspective on using AI for professional research.
  • Reduced Information Overload: The progressive graph helped users focus on the current query without losing the "big picture" of their 30-minute search session.
  • Complementarity: The system caught instances where the LLM was too cautious or vague (e.g., using "Ginkgo biloba" generally when the KG required "Ginkgo biloba extract" for a specific benefit).

Limitations and Future Work

  • KG Dependency: KNOWNET is only as good as the KG it’s connected to. If a breakthrough happened yesterday and isn't in the KG, the system might mark a true LLM statement as "Unsure."
  • Graph Scalability: For very broad searches, the graph can become a "hairball." The authors suggest future work should include hierarchical abstractions to collapse complex sub-graphs.
  • Common Sense Gaps: KGs often lack "common sense" links (e.g., "fish oil is rich in Omega-3") because researchers don't write papers about things everyone already knows. This can lead to "Relevant" tags for obvious facts.
  • User-Driven KGs: The authors aim to allow users to dynamically integrate their own knowledge or specific private datasets into the KG in future iterations.

Summary updated

Abstract

The increasing reliance on Large Language Models (LLMs) for health information seeking can pose severe risks due to the potential for misinformation and the complexity of these topics. This paper introduces KnowNet a visualization system that integrates LLMs with Knowledge Graphs (KG) to provide enhanced accuracy and structured exploration. Specifically, for enhanced accuracy, KnowNet extracts triples (e.g., entities and their relations) from LLM outputs and maps them into the validated information and supported evidence in external KGs. For structured exploration, KnowNet provides next-step recommendations based on the neighborhood of the currently explored entities in KGs, aiming to guide a comprehensive understanding without overlooking critical aspects. To enable reasoning with both the structured data in KGs and the unstructured outputs from LLMs, KnowNet conceptualizes the understanding of a subject as the gradual construction of graph visualization. A progressive graph visualization is introduced to monitor past inquiries, and bridge the current query with the exploration history and next-step recommendations. We demonstrate the effectiveness of our system via use cases and expert interviews.
BibTeX
@article{yan2024knownet,
  author = {Youfu Yan and Yu Hou and Yongkang Xiao and Rui Zhang and Qianwen Wang},
  title = {{KNowNEt:Guided Health Information Seeking from LLMs via Knowledge Graph Integration}},
  journal = {IEEE Transactions on Visualization and Computer Graphics},
  year = {2024},
  doi = {10.1109/tvcg.2024.3456364},
}
Figure