DUAL: A Deep Unified Attention Model with Latent Relation Representations for Fake News Detection

Abstract

The prevalence of online social media has enabled news to spread wider and faster than traditional publication channels. The easiness of creating and spreading the news, however, has also facilitated the massive generation and dissemination of fake news. It, therefore, becomes especially important to detect fake news so as to minimize its adverse impact such as misleading people. Despite active efforts to address this issue, most existing works focus on mining news’ content or context information from individuals but neglect the use of clues from multiple resources. In this paper, we consider clues from both news’ content and side information and propose a hybrid attention model to leverage these clues. In particular, we use an attention-based bi-directional Gated Recurrent Units (GRU) to extract features from news content and a deep model to extract hidden representations of the side information. We combine the two hidden vectors resulted from the above extractions into an attention matrix and learn an attention distribution over the vectors. Finally, the distribution is used to facilitate better fake news detection. Our experimental results on two real-world benchmark datasets show our approach outperforms multiple baselines in the accuracy of detecting fake news.