Claim Verification Leveraging In-context Learning and Retrieval Augmented Generation

The proliferation of digital information has escalated the necessity for efficient and scalable fact-checking methods to combat misinformation. Existing solutions mainly rely on customized learning models that leverage ad-hoc training data and do not fit well in different domains. Indeed, claim verification stresses the availability of an updated and open knowledge base for the designed model. This paper proposes a novel approach that integrates Retrieval-Augmented Generation (RAG) with Large Language Models (LLMs) to enhance claim verification, leveraging in-context learning to assess the veracity of input claims. The methodology involves a two-step process, the evidence retrieval—including web document summarization and claim-focused relation extraction—and claim validation mainly consisting of triple relation extraction and comparison. Evidence retrieval, by filtering information sources, guarantees the reliability of the verdict, enabling the claim verification feasibility in multiple domains. Experimental activities are conducted using the FEVER dataset, and the results demonstrate that the proposed framework significantly improves claim verification accuracy at the state-of-art.