Use of Retrieval-Augmented Large Language Model for COVID-19 Fact-Checking (Preprint)

Abstract

The COVID-19 pandemic has been accompanied by an “infodemic,” where the rapid spread of misinformation has exacerbated public health challenges. Traditional fact-checking methods, though effective, are time-consuming and resource-intensive, limiting their ability to combat misinformation at scale. Large language models (LLMs) like GPT-4 offer a more scalable solution, but their susceptibility to generating hallucinations—plausible yet incorrect information—compromises their reliability.

This study aims to enhance the accuracy and reliability of COVID-19 fact-checking by integrating a retrieval-augmented generation (RAG) system with LLMs, specifically addressing the limitations of hallucination and context inaccuracy inherent in standalone LLMs.

We constructed a context dataset comprising approximately 130,000 peer-reviewed articles related to COVID-19 from PubMed and Scopus. This dataset was integrated with GPT-4 to develop multiple RAG-enhanced models: the Naive RAG, LOTR-RAG, CRAG (Corrective RAG), and SRAG (Self-RAG). The RAG systems were designed to retrieve relevant external information, which was then embedded and indexed in a vector store for similarity searches. Two datasets, one real-world and one synthesized, each containing 500 claims, were used to evaluate the performance of these models. Each model’s accuracy, F1 score, precision, and sensitivity were compared to assess their effectiveness in reducing hallucinations and improving fact-checking accuracy.

The baseline GPT-4 model achieved an accuracy of 0.856 on the real-world dataset. The Naive RAG model improved this to 0.946, while the LOTR-RAG model further increased accuracy to 0.951. The CRAG and SRAG models outperformed all others, achieving accuracies of 0.972 and 0.973, respectively. The baseline GPT-4 model reached an accuracy of 0.960 on the synthesized dataset. The Naive RAG model increased this to 0.972, and the LOTR-RAG, CRAG, and SRAG models achieved an accuracy of 0.978. These findings demonstrate that the RAG-enhanced models consistently maintained high accuracy levels, closely mirroring ground-truth labels and significantly reducing hallucinations. The CRAG and SRAG models also provided more detailed and contextually accurate explanations, further establishing the superiority of agentic RAG frameworks in delivering reliable and precise fact-checking outputs across diverse datasets.

The integration of RAG systems with LLMs offers a robust approach to enhancing the accuracy and contextual relevance of automated fact-checking, making it a valuable tool in combating misinformation during public health crises. While the results are promising, the study acknowledges certain limitations, including the dependence on the quality of the external datasets and the potential biases in the retrieved documents. Future research should focus on refining these models, exploring their applicability across various domains, and ensuring ethical deployment. Policy implications include the need for guidelines and frameworks supporting AI-driven fact-checking systems while maintaining transparency, accuracy, and public trust.