Zero-shot object rumor detection based on contrastive learning
Existing rumor detection models often rely on large-scale manually annotated rumor datasets,which are costly and limited in their ability to detect unknown rumors due to the reliance on features derived from debunked rumors.To address this limitation,an approach for rumor detection targeted at different objects was proposed.Leveraging the zero-shot learning,the rumor dataset was divided into multiple datasets with non-overlapping samples and contents based on different objects,enabling the zero-shot object-oriented rumor detection task.Correspondingly,a universal mask feature was constructed to represent the relationship between objects,and a proxy task was designed to differentiate the universal mask feature.Additionally,object-oriented information-assisted text was introduced to reduce noise caused by data augmentation and was linearly transformed with the original vector semantics.Then,a proxy task-based hierarchical contrastive learning model(ZPTHCL)was presented for zero-shot object-oriented rumor detection,which leveraged transfer learning for rumor detection.Finally,experiments were conducted on a zero-shot rumor dataset based on objects and four publicly available datasets,Ma-Weibo,Weibo20,Twitter15 and Twitter16,demonstrating superior performance of the proposed contrastive learning zero-shot object-oriented rumor detection model.
NETL
NSTL
万方数据
