Unsupervised Clustering with Contrastive Learning for Rumor Tracking on Social Media

Yang Wu, Chen Song, Zhitong Lu, Chunlei Jing, Pengwei Zhan, Zhen Xu, Liming Wang

Abstract: Social media provides a suitable environment for spreading rumors, which might lead to economic losses and public panic. Therefore, an automated rumor resolution system has become crucial. As a significant step in a rumor resolution system, rumor tracking tries to collect related posts of potential rumors. Existing studies about rumor tracking are based on supervised learning approaches, suffering from a critical challenge that requires extensive time and labor to build reliable annotated datasets. To quickly adapt to newly emerging rumors, we investigate if we could track rumors in an unsupervised manner. When querying a rumor post, we perform clustering on the collected posts and find the posts with the same clustering assignment as the rumor. To achieve the goal, we propose an unsupervised clustering method with contrastive learning (UCCL) to search rumor-related posts without any labeled data. Experimental results on the public dataset demonstrate that the proposed method outperforms other unsupervised approaches.