KAHAN: Knowledge-Aware Hierarchical Attention Network for Fake News detection on Social Media

In recent years, fake news detection has attracted a great deal of attention due to the myriad amounts of misinformation. Some previous methods have focused on modeling the news content, while others have combined user comments and user information on social media. However, existing methods ignore some important clues for detecting fake news, such as temporal information on social media and external knowledge related to the news. To this end, we propose a Knowledge-Aware Hierarchical Attention Network (KAHAN) that integrates this information into the model to establish fact-based associations with entities in the news content. Specifically, we introduce two hierarchical attention networks to model news content and user comments respectively, in which news content and user comments are represented by different aspects for modeling various degrees of semantic granularity. Besides, to process the random occurrences of user comments at post-level, we further designed a time-based subevent division algorithm to aggregate user comments at subevent-level to learn temporal patterns. Moreover, News towards Entities (N-E) attention and Comments towards Entities (C-E) attention are introduced to measure the importance of external knowledge. Finally, we detected the veracity of the news by combining the three aspects of news: content, user comments, and external knowledge. We conducted extensive experiments and ablation studies on two real-world datasets and showed that our proposed method outperformed the previous methods and empirically validated each component of KAHAN1.