A Social-Aware Deep Learning Approach for Hate-Speech Detection

Despite considerable efforts to automatically identify hate-speech in online social networks, users still face an uphill battle with toxic posts that seek to sow hatred. In this paper, we initially observe that there is a great deal of social properties transcending both hateful passages and respective authors. We then exploit this observation by i) developing deep learning neural networks that classify online posts as either hate or non-hate based on their content, and ii) proposing an architecture that may invigorate any such text-based classifier with the use of additional social features. Our combined approach considerably enhances the classification accuracy of previously proposed state-of-the-art models and our evaluation reveals social attributes that are the most helpful in our classification effort. We also contribute the first publicly-available dataset for hate-speech detection that features social properties.