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

Anonymous

Paper Link: https://openreview.net/forum?id=1TaD9gh1Y4I

Paper Type: Short paper (up to four pages of content + unlimited references and appendices)

Abstract: 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, we design a hierarchical attention network based automated claim verification module. This architecture uses latent features from the claim, all articles in the dataset related to the claim as well as the most relevant information extracted from the articles via a gating unit. We show by ablation studies that this trainable method of extracting most relevant information from articles results in better performance compared to using cosine similarity metric. We also show by ablation studies that using the most relevant information from articles explicitly in the model results in better performance metrics. The proposed model, Relevant Information Enhanced Hierarchical Attention Network (RIEHAN), outperforms the state-of-the-art benchmark architectures on PolitiFact dataset and performs comparably to the state-of-the-art models on the Snopes dataset.