RIEHAN: Relevant Information Enhanced Hierarchical Attention Network for Automated Claim Verification

The spread of misinformation in online media has caused significant societal problems today, underscoring the importance of verifying claims before accepting them as real. In this work, an automated claim verification module is designed, which is based on a hierarchical attention network. Latent features from the claim, latent features from all the articles that pertain to the claim, and the most relevant information feature extracted from the articles via a gating unit are utilized in this architecture. Ablation studies demonstrate that this trainable approach to extracting the most relevant information from articles performs better than the cosine similarity metric. We also demonstrate through ablation studies that improved performance metrics can be achieved by explicitly including the most pertinent information from articles in the model. The proposed model, Relevant Information Enhanced Hierarchical Attention Network (RIEHAN), performs best compared with the SOTA models on the PolitiFact database and the Snopes database.