Graph Enhanced BERT for Stance-Aware Rumor Verification on Social Media

Abstract

With the rapid development of the Internet and social media, rumors with misleading information will damage more than before. Therefore, rumor verification technology has gained much attention. Many existing models focus on leveraging stance classification to enhance rumor verification. However, most of these models fail to incorporate the conversation structure explicitly. Moreover, the BERT has shown its superiority at text representation on many NLP tasks, and we explore whether it can enhance rumor verification. Thus, in this paper, we propose a single-task model and a multi-task stance-aware model for rumor verification. We first utilize BERT to capture low-level content features. Then we employ graph neural networks to model conversation structure and design an attention structure to integrate stance and rumor information. Our experiments on two public datasets show the superiority of our model over existing models.