Unsupervised Cross-Domain Rumor Detection with Contrastive Learning and Cross-Attention

Massive rumors usually appear along with breaking news or trending topics, seriously hindering the truth. Existing rumor detection methods are mostly focused on same domain, thus have poor performance in cross-domain scenarios due to domain shift. In this work, we propose an end-to-end instance-wise and prototype-wise contrastive learning model cross-attention mechanism for detection. The not only performs feature alignment, but also enforces target samples align corresponding prototypes of a given source domain. Since labels unavailable, use clustering-based approach carefully initialized centers by batch produce pseudo labels. Moreover, pair data learn domain-invariant representations. Because tend express similar semantic patterns especially people’s attitudes (e.g., supporting denying) towards category rumors, discrepancy between will be decreased. We conduct experiments four groups datasets show that our proposed achieves state-of-the-art performance.