Do not wait: Preemptive rumor detection with cooperative LLMs and accessible social context

Junyi Chen, Leyuan Liu, Fan Zhou

Abstract: Concerning the dissemination of rumors via tweets, current methods for detecting rumors – both content-based and graph-based – do not adequately address the need for preemptive action to suppress rumors before public exposure. Although the advancement of large language models (LLMs) indicates a positive trend, their application in rumor detection remains either overly simplistic or excessively complex. Motivated by these, we put forward EvidenceRD, which employs an efficient yet effective cooperative strategy of three types of LLMs to mine informative evidence that augments content context behind tweets warranting a fact check. This is then integrated with a credibility network, an automatically generated social context based on the social homophily theory, to depict potential credibility relationships between authors of corresponding tweets. Extensive experiments across four public datasets confirm the efficacy of the proposed EvidenceRD. Specifically, EvidenceRD outperforms state-of-the-art baselines across various categories, achieving an improvement in general detection performance ranging from 3% to 16%. This superiority is achieved by exclusively utilizing pre-public information, a constraint not imposed on the comparative baselines. Besides, As EvidenceRD considers both evidence and the social context behind a tweet, it not only offers enhanced explainability through its multi-perspective evidence presented in natural language but also demonstrates greater robustness and transferability across different scenarios. Additional efficiency analysis demonstrates that the above-enhanced characteristics brought by EvidenceRD are cost-effective in terms of both financial and computational expenses. To summarize, compared to existing studies, this work theoretically presents a generally effective and efficient paradigm for utilizing LLMs in the context of rumor detection and demonstrates how integrating accessible social context can effectively detect rumors. Practically, the proposed method can accurately detect rumors before they emerge publicly with several notable metrics improvements and aid human decision-making on them from multiple dimensions.