LLM-Enhanced Multi-Task Joint Learning Model for Misinformation Detection

The coexistence of Human-Generated Content (HGC) and Artificial Intelligence-Generated Content (AIGC) versions of the same event on social media presents significant challenges for governmental governance and information regulation. In this study, we propose a Large Language Model-enhanced Multi-Task Joint Learning Model for Misinformation Detection (LMTMD) to address the challenge of mixed HGC and AIGC on social media. We design a two-stage instruction, leveraging large language models (LLMs) for data augmentation to generate AIGC versions of events. Furthermore, a novel unsupervised joint learning strategy is proposed, which incorporates content consistency contrastive learning and difference consistency learning. The strategy aims to preserve both the consistency of event content and the heterogeneity between AIGC and HGC. Extensive experiments conducted on real-world datasets, including Weibo and GossipCop, demonstrate that the proposed model outperforms state-of-the-art baselines, achieving a Consistent Match Accuracy (CM-Acc) of 77.21% on the Weibo dataset and 78.13% on the GossipCop dataset. Additionally, the model achieves AIGC detection accuracy rates of 90.58% on the Weibo dataset and 90.95% on the GossipCop dataset, thereby validating the effectiveness of both the model and the joint learning strategy. Our model can effectively adapt to the emerging scenario of mixed HGC and AIGC versions of events on social platforms and enriches the research perspective of misinformation detection.