An Active Learning Framework for Continuous Rapid Rumor Detection in Evolving Social Media

The widespread of social media has led to the propagation of rumors, resulting in detrimental effects on society. Existing rumor detection methods rely on labeled data and face challenges in obtaining sufficient data, particularly in the early stages of rumor propagation. Active learning has been proposed as a solution to reduce labeling costs by selecting significant samples for training. However, applying active learning to rumor detection presents challenges, such as sensitivity to initial sets and performance loss due to sample reduction. Inspired by Active learning, Contrastive learning, and Prompt tuning, we propose a novel Rumor Detection framework named ACP-RD in this paper. ACP-RD addresses the limitations of active learning by leveraging the benefits of contrastive learning and prompt tuning. Our framework reduces labeling costs, improves rumor detection performance in zero-shot scenarios, aligns active query criteria with downstream tasks through unsupervised contrastive learning, and optimizes word embeddings using supervised contrastive learning. Experimental results demonstrate that ACP-RD achieves superior performance compared to fine-tuning with the whole labeled dataset, using only 10% of the labeled data. Overall, our proposed framework presents a novel and efficient approach to rumor detection, with potential for broader applications in natural language classification tasks.