Joint rumour and stance identification based on semantic and structural information in social networks

Rumours that have spread in social networks have harmed society seriously, so rumour verification is a substantial task in social media analysis and natural language processing. In social networks, replies with different stances may provide direct clues to the veracity of the rumours. Thus, rumour verification would benefit from joint training with stance detection. However, there are still some shortcomings in current research, such as the unsatisfactory use of structure and semantic information in the conversation, features for different tasks independent of each other except for sharing input, and the insufficient discrimination of tweets with different stances. Aiming at these shortcomings, we first used the graph transformer to simultaneously obtain structural and semantic information such as dialogue reply, similar posts, same user, and same stance. Secondly, we adopted the partition filter network to explicitly model the rumour& stance-specific features and the shared interactive feature. Finally, we strengthened the discriminability of tweets with different stances through contrastive learning. Experiments on SemEval2017 and PHEME corpus show that the proposed model significantly improves the rumour and stance detection tasks.