Collaborative Mixture-of-Experts Model for Multi-Domain Fake News Detection

Abstract

With the widespread popularity of online social media, people have come to increasingly rely on it as an information and news source. However, the growing spread of fake news on the Internet has become a serious threat to cyberspace and society at large. Although a series of previous works have proposed various methods for the detection of fake news, most of these methods focus on single-domain fake-news detection, resulting in poor detection performance when considering real-world fake news with diverse news topics. Furthermore, any news content may belong to multiple domains. Therefore, detecting multi-domain fake news remains a challenging problem. In this study, we propose a multi-domain fake-news detection framework based on a mixture-of-experts model. The input text is fed to BertTokenizer and embeddings are obtained by jointly calling CLIP to obtain the fusion features. This avoids the introduction of noise and redundant features during feature fusion. We also propose a collaboration module, in which a sentiment module is used to analyze the inherent sentimental information of the text, and sentence-level and domain embeddings are used to form the collaboration module. This module can adaptively determine the weights of the expert models. Finally, the mixture-of-experts model, composed of TextCNN, is used to learn the features and construct a high-performance fake-news detection model. We conduct extensive experiments on the Weibo21 dataset, the results of which indicate that our multi-domain methods perform well, in comparison with baseline methods, on the Weibo21 dataset. Our proposed framework presents greatly improved multi-domain fake-news detection performance.