Deciphering Rumors: A Multi-Task Learning Approach with Intent-aware Hierarchical Contrastive Learning

Social networks are full of noise and misleading information, which poses a pressing and complex challenge for rumor detection. In this paper, we propose the Intent-Aware Rumor Detection Network (IRDNet), designed to address the challenges of subjectivity, robustness, and consistency in existing models. IRDNet uses a multi-task learning framework that integrates rumor detection and latent intent mining, which can discern multi-level semantic features and potential user intentions. In IRDNet, the multilevel semantic extraction module extracts sequential and hierarchical features to produce robust semantic representations. The intentaware hierarchical contrastive learning module introduces two complementary strategies, event-level and intent-level. Event-level contrastive learning uses high-quality data augmentation and adversarial perturbations to enhance the robustness and consistency of the model. Intent-level contrastive learning utilizes an intent encoder to capture subjective intent and optimize homogeneity within the same intent while ensuring heterogeneity between different intents, thereby clearly distinguishing critical features from irrelevant elements. Experimental results verify that the model significantly improves the effect of early rumor detection and effectively solves the essential problems of the existing rumor detection field.