Multimodal Fake News Detection Based on Chain-of-Thought Prompting Large Language Models

The rapid rise of social networks has led to a proliferation of fake news, especially those with images. The combination of images and text may confuse users and cause even more negative impact. Exisiting methods for fake news detection either require expert knowledge or large amounts of labeled data. In addition, these methods fails to clarify which part of the multimodal information is misleading or why. In this paper, we present a simple yet efficient Chain-of-thought Prompting method for Multimodal Fake News Detection (CP-FEND). It first finds the closest demonstration samples of the news posts to be detected by a KNN-based approach. Afterwards, we design a logical prompt method including Examination, Inference and Determination stages to guide Large Language Model (LLM) to automatically construct reasoning processes for the authenticity of the samples. Finally, LLM are prompted to derive the authenticity of multimodal news with the guidance of samples and Chain-of-Thought reasoning. A reflective verification is performed to further improve the detection performance through comprehensive evaluation of the original responses. Extentsive experiments on two public datasets have demonstrated the superiority of our method over existing methods.