Research Of Early Rumor Detection And Rumor Verification On Social Media

With the increasing popularity of social media,more and more information is posted and disseminated on social media platforms,including rumors.Rumors on social media can spread widely in a short time,which has become a threat to mislead the public and affect social stability.Therefore,automatic early rumor detection and rumor verification have become hot research topics,and attracted much attention from researchers.This paper focuses on the key techniques of early rumor detection and rumor verification.The main works of the paper are as follows:For early rumor detection,social bots are increasingly involved in the early propagation of rumors,while existing models could not be aware of social bots,which might limit the model to learning a more effective pattern of early propagation.To address the above problems,the paper proposes a Bot-Aware Graph neural network(BAG)for early rumor detection.Firstly,a user credibility scorer is pre-trained on a large number of social bots and genuine users and transferred to BAG.Secondly,utilize the scorer to score the users in the user interaction graph constructed according to the reply relationship,calculate the edge weights,and apply the graph attention network to learn the user feature of each node;For the user publishing graph constructed according to the relationship between users and source posts,the scorer is used to score the publishing users,and the scores are incorporated into the calculation of the graph neural network to learn the publishing features of users.Finally,the textural feature of the source post is fused with the publishing feature and user features for early rumor detection.Extensive experiments on three public datasets demonstrate that BAG outperforms existing early rumor detection models in accuracy and early-detection ability.For rumor verification,existing methods are difficult to extract rich contextual semantic and stance features from short posts,which limits the performance of the model.To alleviate the above problems,the paper proposes a Stance-Aware Recursive Tree(SART)for rumor verification.Firstly,SART uses a multi-head self-attention mechanism to learn the features of all child nodes under the same parent node,and the semantic and stance information of the short post will be complemented by its sibling nodes.Secondly,SART utilizes another attention mechanism to model the interaction between nodes of a subtree,i.e.,calculates the attention weights according to the social characteristics of nodes,which participates in aggregating features of sibling nodes.Finally,the bottom-up aggregated feature at the root node is for rumor verification.In addition,SART includes a multi-task learning module,which can learn the stance feature between posts.SART can jointly train the stance classification task in this module to enhance the capability for stance representation.Extensive experiments on two public datasets show that SART can effectively model the conversation trees of rumors,and outperforms the current state-ofthe-art stance-aware model by around 9%.