Knowledge-Enhanced Graph Attention Network for Fact Verification

Fact verification aims to evaluate the authenticity of a given claim based on evidence sentences retrieved from Wikipedia articles. Existing works mainly leverage natural language inference methods model semantic interaction and evidence, or further employ graph structure capture relation features between multiple evidences. However, previous have limited representation ability in encoding complicated units evidences, thus cannot support sophisticated reasoning. In addition, amount supervisory signals lead encoder could not distinguish distinctions different structures weaken ability. To address above issues, we propose Knowledge-Enhanced Graph Attention network (KEGA) for fact verification, which introduces knowledge integration module enhance claims evidences by incorporating external knowledge. Moreover, KEGA leverages an auxiliary loss contrastive learning fine-tune attention learn discriminative graph. Comprehensive experiments conducted FEVER, large-scale benchmark dataset demonstrate superiority our proposal both multi-evidences single-evidence scenarios. findings show that background words can effectively improve performance.