Multi-Pointer Co-Attention Network for Explainable Fake News Detection

Abstract

Fake news detection remains challenging because online news is often accompanied by noisy user responses, heterogeneous textual signals, and limited decision transparency. This paper presents ExplainFake, a comment-aware and interpretable detection framework built on a multi-pointer co-attention strategy. The model first encodes news sentences and user comments with a pre-trained language model to obtain contextual representations. It then selects multiple informative sentence-comment pairs and further analyzes word-level relations within the selected pairs, so that both sentence-level and token-level evidence can be derived. To combine the learned representations, multimodal factorized bilinear pooling is used to model cross-feature interactions before final classification. Experiments conducted on six datasets, including PolitiFact, GossipCop, Twitter15, Twitter16, Weibo20, and Weibo23, show that ExplainFake obtains competitive or better F1-scores than representative baselines. The ablation and interpretability results further indicate that the proposed modules contribute to both prediction performance and explanation quality.