HACK: A Hierarchical Model for Fake News Detection

Online social media sites have become the most powerful platform to share news nowadays. However, all kinds of unauthenticated news that are released online without strict limits may lead to the spread of fake news, which has become a synonym for social and political threats. The existing solutions to the fake news issue are mostly trying to construct a social graph network by integrating the news content and social context of the news, which may be restricted when lacking social context information. In this paper, we propose a model for text only, regardless of contextual information, and named it HACK (HierArchical deteCtion for faKe news), which can construct high-level combined features of spatial capsule vectors from low-level character features and phrase features by fusing a pre-trained language model and convolution network. The experimental results on real-life data show that the classification accuracy is significantly improved by our method comparing with the state-of-the-art methods