Abstract
One key for improving hate speech detection performance is to have a textual training corpus that is vast and confidently labeled. This paper develops a semi-supervised learning approach with self-training to benefit from the abundant amount of social media content and develop a robust hate speech classifier for future predictions. The classifier is self-trained iteratively using the most confident pseudo labels obtained from a large-scale unlabelled Twitter corpus. We demonstrate our approach's efficacy and the high quality of the produced supervised hate speech dataset through experiments.