Hate Speech Detection on Social Media Using Graph Convolutional Networks

Detection of hateful content on Twitter has become the need of the hour. Hate detection is challenging primarily due to the subjective definition of "hateful". In the absence of context, text content alone is often not sufficient to detect hate. In this paper, we propose a framework that combines content with context, to detect hate. The framework takes into account (a) textual features of the content and (b) unified features of the author to detect hateful content. We use a Variational Graph Auto-encoder (VGAE) to jointly learn the unified features of authors using a social network, content produced by the authors, and their profile information. To accommodate emerging and future language models, we develop a flexible framework that incorporates any text encoder as a plug-in to obtain the textual features of the content. We empirically demonstrate the performance and utility of the framework on two diverse datasets from Twitter.