PaperlayerView exact revision on arXiv

Source-linked AI summary

TrumorGPT: Graph-Based Retrieval-Augmented Large Language Model for Fact-Checking

Ching Nam Hang, Pei-Duo Yu, Chee Wei Tan

arXiv:2505.07891v2cs.CLcs.AI

TL;DR

Health misinformation spreads rapidly, while fact-checking faces limits from information volume, changing health knowledge, and LLM hallucinations. TrumorGPT combines an LLM with few-shot semantic health knowledge-graph construction and GraphRAG using updated medical information. The paper reports robust and superior performance in fact-checking health-related claims and news.

  • Problem

    Rapidly spreading health misinformation and the limitations of static LLM knowledge and conventional fact-checking motivate updated, reliable automated verification.

  • Method

    TrumorGPT combines an LLM with few-shot learning, topic-aware semantic health knowledge-graph construction, and GraphRAG over updated health information.

  • Results

    TrumorGPT demonstrates superior and robust performance in fact-checking medical knowledge, health-related news, and health-related statements.

  • Takeaways & Limitations

    The framework provides focused health fact-checking by grounding semantic reasoning in curated, updated health knowledge graphs.

Abstract

from arXiv · show

In the age of social media, the rapid spread of misinformation and rumors has led to the emergence of infodemics, where false information poses a significant threat to society. To combat this issue, we introduce TrumorGPT, a novel generative artificial intelligence solution designed for fact-checking in the health domain. TrumorGPT aims to distinguish "trumors", which are health-related rumors that turn out to be true, providing a crucial tool in differentiating between mere speculation and verified facts. This framework leverages a large language model (LLM) with few-shot learning for semantic health knowledge graph construction and semantic reasoning. TrumorGPT incorporates graph-based retrieval-augmented generation (GraphRAG) to address the hallucination issue common in LLMs and the limitations of static training data. GraphRAG involves accessing and utilizing information from regularly updated semantic health knowledge graphs that consist of the latest medical news and health information, ensuring that fact-checking by TrumorGPT is based on the most recent data. Evaluating with extensive healthcare datasets, TrumorGPT demonstrates superior performance in fact-checking for public health claims. Its ability to effectively conduct fact-checking across various platforms marks a critical step forward in the fight against health-related misinformation, enhancing trust and accuracy in the digital information age.

I. INTRODUCTION

Health misinformation spreads rapidly online, while conventional fact-checking struggles with the volume, speed, subjectivity, and contextual complexity of digital content. TrumorGPT addresses these challenges by combining an LLM, semantic health knowledge graphs, few-shot learning, topic-aware graph construction, and GraphRAG for updated health fact-checking.

  • Motivation: Infodemics spread health misinformation rapidly through social networks, creating serious consequences for individuals and public health.The paper highlights COVID-19 misinformation and the difficulty of finding credible medical advice online.
  • Motivation: Manual fact-checking cannot easily keep pace with the volume and speed of online information and may overlook framing, context manipulation, and evaluator bias.The paper also identifies limited collaboration and a lack of standardized methodologies as additional challenges.
  • Proposed Framework: TrumorGPT is a generative AI framework that identifies “trumors,” or health-related rumors that ultimately prove factual.It is designed for health fact-checking and aims to bridge skepticism and verified fact in public health discourse.
  • Proposed Framework: The framework combines an LLM with semantic health knowledge graphs, few-shot learning, topic-enhanced sentence centrality, topic-specific TextRank, and GraphRAG.GraphRAG supplies updated health knowledge for semantic reasoning, while the graph-construction methods support semantic analysis.
  • Evaluation: TrumorGPT is evaluated on medical-knowledge accuracy and the truthfulness of health-related news, with the paper reporting superior fact-checking performance.The evaluation is presented as extensive and focuses on health-related verification tasks.

III. METHODS

TrumorGPT constructs semantic health knowledge graphs from health-related input and uses them with a pretrained LLM for semantically reasoned fact-checking. The workflow combines topic-focused extraction, updated health knowledge, and graph-based retrieval to produce verified answers.

  • TrumorGPT processes health-related user queries to construct semantic health knowledge graphs and generate fact-checking answers through semantic reasoning.The framework retrieves and synthesizes health information before determining the factual accuracy of the input query.
  • The framework uses few-shot learning with topic-enhanced sentence centrality and topic-specific TextRank to extract relevant information for graph construction.These methods identify central topics and health-related keywords from large text inputs.
  • TrumorGPT applies GraphRAG to retrieve and synthesize information from an updated medical and health knowledge base.The resulting knowledge graph supports semantic reasoning for fact-checking.
  • The architecture combines an extensive medical and health knowledge base with algorithms that generate a verified semantic health knowledge graph and reasoned fact-checking answers.
  • A semantic health knowledge graph represents entities, relations, and factual triples, enabling truth assessment through graph edges or short paths between entities.A query is evaluated as a triple (h, r, t) within the graph framework.

1) Topic-Enhanced Embeddings with LDA and BERT:

The topic-enhanced embedding method combines semantic information from BERT with topic distributions from LDA. These representations support a similarity graph whose central sentences are both semantically meaningful and aligned with health topics.

  • LDA treats each sentence as a document and produces topic-distribution vectors reflecting the sentence’s probability of belonging to identified topics.The LDA model can be trained on a health-focused corpus and applied to new sentences.
  • BERT embeddings and LDA topic distributions are combined to represent each sentence using both semantic and topical information.The combined representation integrates high-dimensional BERT features with topic probabilities.
  • The method introduces a weighting parameter η to balance semantic and topical contributions in the combined sentence representation.Experiments set η = 0.7, giving slightly more weight to BERT semantic information.
  • A weighted sentence graph connects vertices according to cosine similarity between their combined representation vectors.The edge weight is defined from the dot product and norms of the corresponding vectors.
  • PageRank-style centrality identifies sentences that are both semantically meaningful and topically aligned with the health domain.The centrality computation incorporates LDA-derived domain knowledge and BERT-based semantic understanding.

C. Topic-Specific TextRank

Topic-Specific TextRank adapts graph ranking to emphasize topic-relevant vertices and transitions. Its convergence is guaranteed by topic-specific teleportation, while speed depends mainly on the original graph’s second-largest eigenvalue and the damping factor.

  • Topic-Specific TextRank modifies conventional TextRank by incorporating topic relevance into vertex selection, edge weighting, and scoring.The adaptation uses topic relevance scores to prioritize health-related content.
  • Topic-specific teleportation prioritizes relevant vertices, while the adjusted random walk also weights transitions using topic relevance scores.
  • The transition matrix P is column-stochastic, so it defines a valid Markov-chain transition process.Its entries are non-negative and each column sums to 1.
  • The topic-specific teleportation term (1 −d)E makes the Markov chain irreducible and aperiodic, guaranteeing a unique stationary distribution.The stationary distribution has strictly positive entries.
  • The convergence rate is governed by λ2(P′) = dλ2(P), so larger damping factors or larger |λ2(P)| generally slow convergence.The convergence factor approaches 1 as the damping factor increases when |λ2(P)| is close to 1.
  • With λ2(P) = 0.9899, d = 0.6 converged after 30 iterations, whereas d = 0.95 failed to converge within 100 iterations.The simulation used a Watts-Strogatz graph with (k, n) = (4, 500).
  • The implementation uses d = 0.85 and stops power iteration when the TST-vector change falls below 10^-6.

D. Large Language Model with Few-Shot Learning

TrumorGPT uses GPT-4 with few-shot learning to construct semantic health knowledge graphs from new health-related inputs. The approach helps extract central content from lengthy texts and generalize graph construction without exhaustive retraining.

  • TrumorGPT instructs GPT-4 to use topic-enhanced sentence centrality and topic-specific TextRank when constructing semantic health knowledge graphs.This addresses the difficulty of having GPT-4 learn mathematically involved graph-construction algorithms from scratch.
  • Few-shot learning adapts GPT-4 from query–knowledge-graph examples to construct graphs for unseen inputs.The objective is to generalize graph construction across new scenarios.
  • The approach extracts key phrases and sentences that represent central ideas and organize relationships within the semantic health knowledge graph.
  • Few-shot learning supports efficient generalization to diverse inputs without exhaustive retraining on large datasets.The resulting graphs remain focused and relevant for extensive source texts.

E. Graph-Based Retrieval-Augmented Generation

GraphRAG augments TrumorGPT by retrieving current semantic health knowledge graphs and using graph similarity to verify health-related queries. The workflow converts a query into a graph, retrieves relevant graphs, and produces True, False, or Undetermined outcomes.

  • TrumorGPT transforms each user query into a query knowledge graph and retrieves database graphs according to semantic similarity.
  • The framework determines truthfulness from graph similarity scores, using a threshold to classify queries as True, False, or Undetermined.
  • Its scoring function combines subgraph isomorphism and Jaccard similarity over overlapping consecutive triples, optionally weighting more important triples.
  • GraphRAG gives GPT-4 access to updated, structured health knowledge graphs to address hallucinations linked to static or incomplete training data.
  • When direct verification is unavailable, TrumorGPT can expand related knowledge graphs or provide relevant knowledge to guide judgments.

IV. PERFORMANCE EVALUATION

The evaluation examines TrumorGPT’s effectiveness in validating the truthfulness of health news content as a fact-checking application.

  • The study evaluates TrumorGPT as a framework for validating the truthfulness of health news content.

A. Experimental Setup

The experiments update TrumorGPT with English DBpedia knowledge and compare its public-health fact-checking performance against several advanced language models.

  • The knowledge base uses English RDF triples from DBpedia’s latest-core collection to incorporate recently updated structured information.
  • The evaluation compares TrumorGPT with GPT-3.5, GPT-4, LLaMA 3.2, PaLM 2, Claude 3.5 Sonnet, and Gemini 1.5.
  • Fact-checking performance is assessed using accuracy, precision, recall, and F1-score, with true and false positives and negatives defining evaluation outcomes.

B. Evaluation Approach

The evaluation focuses on health policy and political news around the 2024 U.S. presidential election, using PolitiFact claims and ratings as the fact-checking basis.

  • The study targets health policy and political news because fact-checking supports informed voter decisions and political discourse before the election.
  • The dataset centers on PolitiFact statements concerning healthcare, public health, and U.S. political policies.
  • PolitiFact evaluates claims with six categories: True, Mostly True, Half True, Mostly False, False, and an additional lowest-accuracy category.

C. Results

TrumorGPT achieved the strongest reported performance among the evaluated language models for health-related claim verification. Its GraphRAG-based, health-specialized approach also produced concise fact-checking responses.

  • 88.5% accuracy was achieved by TrumorGPT on 600 PolitiFact health-related claims, exceeding the five comparison language models.The dataset contained 300 true and 300 false statements from the “Health Care” and “Coronavirus” categories.
  • TrumorGPT led all evaluated models in precision, recall, and F1-score for health-related claim fact-checking.Across models, precision exceeded accuracy, which exceeded recall, while F1-score was lowest.
  • GraphRAG supplements GPT-4 with current semantic health knowledge graphs, addressing the limits of static training data and general-purpose model knowledge.The framework also uses health-specific enhancements to identify nuances in health and medical misinformation.
  • 2.8 sentences was TrumorGPT’s lowest average response length, indicating concise delivery of fact-checking results.The paper links shorter responses with efficient information retrieval and easier user comprehension.

D. TrumorGPT for Fact-Checking of Public Health Claims

TrumorGPT performs binary public-health fact-checking strongly, correctly identifying most true and false statements while favoring concise explanations. Its accuracy declines for six-way labels and as knowledge graphs grow more complex with longer articles.

  • 92% of false statements and 85% of true statements were correctly identified in TrumorGPT’s binary fact-checking confusion matrix.The model performed slightly better on false statements than true statements.
  • 88.5% binary accuracy contrasted with 49.3% six-category accuracy on the same 600 statements.The multi-class system used PolitiFact’s original six categories.
  • TrumorGPT generally placed true claims in “True,” “Mostly True,” or “Half True,” and false claims in the corresponding false categories.When uncertain, it often selected middle-ground labels such as “Half True” or “Mostly False.”
  • TrumorGPT’s shorter explanations were associated with higher accuracy, supporting concise fact-checking responses.The analysis examined prediction accuracy alongside explanation length across language models.
  • Knowledge graphs ranged from 6 to 18 vertices and generally grew linearly with article length, while error rates also increased.Longer articles may make key-point identification more difficult for TrumorGPT.
  • TrumorGPT can verify factual health statements by referencing current medical research and health-organization guidelines.The paper illustrates this capability with a balanced-diet claim related to diabetes management.

V. DISCUSSION

The discussion contrasts TrumorGPT’s precision-focused GraphRAG with HybridRAG and LightRAG, then proposes hybrid retrieval and incremental updates to broaden coverage while preserving reliable health fact-checking.

  • Comparison with HybridRAG: HybridRAG combines vector-based passage retrieval with graph-based entity extraction, offering broader coverage for unconventional or implicit health claims.Vector search can recover relevant information that direct graph lookup may miss, while the graph preserves precise relationships.
  • Comparison with HybridRAG: TrumorGPT’s GraphRAG delivers highly faithful, contextually precise fact-checking when claims map clearly to entities and relationships in its curated health graph.Its targeted retrieval may instead return “undetermined” when the relevant fact is not explicitly represented.
  • Comparison with LightRAG: LightRAG uses dual-level retrieval and incremental graph updates to capture specific facts, broader themes, and fresh medical evidence efficiently.This design can assemble multi-step evidence such as a drug’s effect chain from administration to outcome.
  • Comparison with LightRAG: Compared with LightRAG’s broader multi-step reasoning, TrumorGPT emphasizes precise, focused true-or-false decisions for claims represented in its curated graphs.The discussion identifies incremental updates and dual-level retrieval as potential enhancements that need not compromise fact-based verification.
  • Proposed enhancements: A proposed hybrid mechanism would combine graph-based precision with vector-based contextual search to retrieve structured facts and implicit cues from unstructured text.The discussion also proposes incremental graph updates so new medical studies and real-time health news can be integrated without extensive reindexing or downtime.
  • Framework contribution: TrumorGPT combines few-shot learning for semantic health knowledge-graph construction with retrieval-augmented generation that supplies GPT-4 with updated health information.The framework is presented as addressing hallucination in large language models while verifying health-related statements.
Loading 2505.07891v2…