Mitigating the Evidence-Related Factors in Automated Fact-Checking

Abstract

The rapid proliferation of digital misinformation highlights the urgent need for robust automated fact-checking systems that can accurately distinguish truth from falsehood. A persistent challenge for these systems is the occurrence of false positives, where truthful information is incorrectly flagged as misleading due to limitations in evidence assessment, including insufficient evidence, logical inconsistencies, and conflicting information. This research introduces a novel two-phase approach to address these issues. In Phase 1, relationships between claims and evidence are modeled using a graph-based mechanism to identify evidence-related shortcomings that contribute to false positives. Phase 2 enhances evidence quality by integrating domain-specific knowledge, employing pretrained language models such as BERT, RoBERTa, and BioBERT across diverse datasets like FEVER, LIAR-Plus, HoVER, and PubMed. Our findings demonstrate that addressing these evidence-related factors significantly reduces false positives, resulting in more accurate fact-checking. These results underscore the effectiveness of our enhanced evidence assessment method, providing valuable insights for developing reliable fact-checking systems adaptable across multiple domains. This research lays a foundation for future innovations in misinformation mitigation, fostering a more trustworthy digital information landscape.