Exploring Hierarchical Language Knowledge in Graph Neural Networks for Fake News Detection

Fake news can be propagated quickly across online microblogs, resulting in a series of adverse impacts on our daily lives. Traditional fake news detection models focus on incorporating writing styles, or world knowledge (e.g., triples). Nevertheless, writing styles are easy to imitate. Different from world knowledge, in this paper, we propose a novel hierarchical language knowledge-driven fake news detection (HLKFND) framework. More specifically, we first conduct entity linking to obtain the entity words for a given news text after removing stop words. We also extract the specific topic words through the LDA (Latent Dirichlet Allocation) for the news text. Then, we acquire the extended entities context through an external knowledge base for the extracted entity words. Next, we extract language context (the sememe of a Chinese word based on the HowNet) for the extracted topic words. After that, we construct a powerful language-entity graph that includes the previous words in the news text, the extended entity context, and the extended language context. Finally, we successfully combined the language context and the entity context under a graph convolutional networks framework. Our experimental results demonstrate that our HLKFND outperforms strong recent baselines on Chinese benchmark dataset in fake news detection.