Rumor Verification On Social Media With Stance-Aware Recursive Tree

Since rumors have affected real society harmfully, automatic rumor verification attracts much attention from researchers. Incorporating the stance-aware knowledge into rumor verification is a hot direction, because its great potential to boost verification performance has been revealed in many studies. However, existing methods are still limited by two problems, the existence of short retweets and fraud nodes. For short retweets, since it is hard to extract the semantic information from short retweets, modeling the stance between a tweet and its short retweets could carry out training noises. For fraud nodes, they might perturb the normal propagation structure of rumors, so the model could be misled to capture those wrong stance information. To mitigate them, we propose a Credibility and Stance Aware recursive Tree (CSATree) for rumor verification. Firstly, we utilize a self-attention mechanism and a multi-task learning module to explore the context of short retweets, which could help to enrich the semantics and stance information in short retweets. In detail, the context of short retweets refers to those retweets that respond to a same tweet, i.e., sibling nodes in a conversation tree. Secondly, we take the node credibility into account and adopt another novel attention mechanism to reduce the impact of fraud nodes. Experiments on two public datasets demonstrate that CSATree significantly outperforms the current best stance-aware model by around 9%.