Online Active Learning for Offensive Language Detection

Recent success in deep learning-based offensive language detection, however, the off-line learning strategy cannot cope with the challenge of rapid growth and evolution of offensive language. In this paper, we introduce an online active learning method to offensive language detection model. Online learning makes use of the user’s report feedbacks to continue training the detection model. In addition, active learning methods can filter out the mislabeled feedbacks to ensure the safety of the online learning process. Extensive experiments demonstrate that the proposed method can achieve promising results on the offensive response dataset.