Rumor Detection Framework In Low-Resource Scenarios

The popularity of the Internet and social media platforms has greatly promoted people's information communication.Through social media platforms,people can easily publish and receive information.However,information on social media platforms may not be always true,and there are also certain amount of rumors and misinformation.The unlimited spread of rumors on social media platforms has a huge negative impact on society.In order to solve this problem,researchers have proposed many rumor detection methods and achieved remarkable performance.However,the scarcity of labeled data and the imbalance of data distribution among samples severely restrict the further development of rumor detection research.What's worse,although rumors are rampant on our social networks,it is very tough to complete annotation of rumor data,which requires a lot of time and energy for annotators.Therefore,how to make more effective utilization of rumor data has gradually become a challenge of researchers in rumor detection field.Aiming at the problem of how to make better use of the labeled rumor data to achieve rumor detection,the works of this paper are as follows:With the idea of multi-task learning,the rumor detection task is divided into two subtasks,namely stance classification and rumor verification.In the task of stance classification,this paper proposes a text-level data augmentation method,which uses the knowledge of thesaurus,word vectors,and pre-trained language models to help the text in the rumor dataset to transform and generate new text.Then,this paper uses supervised contrastive learning to represent text in rumor data.Finally,the evaluation results on the Rumour Eval-17 dataset shows that the proposed method performs better than existing methods.For the rumor verification task,according to the structure of the rumor data,the characteristics of its text,and the response habits of users on social media platforms.A hierarchical data augmentation method is proposed in this paper,and a hierarchical rumor identification model is constructed to match it,modeling the representation of the structure for the rumor data.The evaluation results on Rumour Eval-17 and PHEME datasets show that the proposed framework is superior to the existing approaches.