Multi-Modal Component Embedding for Fake News Detection

摘要

As numerous fake news bloom and spread wildly on social media, fake news detection has recently been drawing a growing amount of attention. Single news consists of various multi-modal components (e.g., text, image, and event). Thus, a desirable model for fake news detection must satisfy two requirements: 1) it must correctly learn the reliability of each component 2) it must be capable of capturing the relationship among the components. In this paper, we propose a Multi-modal Component Embedding framework (MCE) for fake news detection, which is designed to satisfy all the requirements. It first defines a latent vector for each news article as the sum of its component latent vectors. For each component, we regard its magnitude as its reliability, and regard its directional relationship as its consistency. In this context, the magnitude of each news latent vector represents how reliable the news is. Thus, MCE learns the latent space so that the magnitude of the real news vectors becomes larger than that of the fake news vectors. During the training, a news vector becomes larger when its component vectors are reliable (i.e., large magnitude) and when its component vectors are well aligned (i.e., high consistency). By doing so, MCE can capture the complex relationship among the components as well as the reliability of each component. Our extensive experiments on two real-world datasets show that MCE outperforms all the baselines. We also provide a qualitative analysis on the embedding space to verify its capability of satisfying the requirements.