Towards more robust hate speech detection: using social context and user data

In this paper, we present a novel approach to detecting hate speech on Twitter. Our method incorporates textual, social context and language features of the author to better capture the nuances of hate speech and improve detection accuracy. We formalize the idea that an individual's hateful content is influenced by their social circle and propose a framework that combines text content with social context to detect hate speech. Our framework uses a Variational Graph Auto-encoder to jointly learn the unified features of authors using a social network, language features, and profile information. Additionally, to accommodate emerging and future language models, our framework is designed to be flexible and can incorporate any text encoder as a plug-in to obtain the textual features of the content. We evaluate our method on two diverse Twitter datasets and show that it outperforms existing state-of-the-art methods by a significant margin. Our results suggest that considering social context is a promising direction for improving hate speech detection on Twitter.