Cross-Domain Rumor Detection Based on Dual-Modal Domain Alignment

The widespread emergence of rumors in various domains exposes the limitations of the rumor detection models in a single domain. However, previous rumor detection methods face challenges in accurately identifying out-of-domain rumors due to limited generalization capabilities. To detect the daily influx of rumors originating from various domains, we propose an unsupervised domain-adaptive rumor detection model based on dual-modal alignment. Our core idea is to utilize the potential domain-shared source domain knowledge in the dual-modal (text and structure modal) to detect rumors in the unlabeled target domain. By weakening the impact of domain-specific keywords, our model accomplishes domain alignment in the text modal, effectively diminishing the distribution discrepancy between the source and target domains. To align the structure modal across domains, we initially generate pseudo-labels for unlabeled samples in the target domain. Subsequently, we utilize contrastive learning to minimize the distance of samples with the same labels between the source and target domains, while simultaneously increasing the distance between samples with different labels. With dualmodal domain alignment, we ultimately achieve effective detection of rumors in the target domain under unsupervised conditions. We have conducted experiments on both Chinese and English datasets to validate the effectiveness of our model.