Given the subjectivity in defining hate speech, building systems for hate speech detection via NLP has been a challenge. The subjective understanding of hate speech makes the analysis and mitigation of biases that usually exist in NLP systems even more challenging and open-ended. In this talk, the authors will share about their experience in curating a hate speech dataset in the Indic context and modeling it with contextual information. The authors will also shed light on uncovering the “knowledge drift” in KG-based bias mitigation in hate speech detection systems as a part of their study surveying and reproducing such systems.
Sarah Masud is a 4th year Ph.D. student in the area of social computing based out of IIIT-Delhi. Her time is well-spent analyzing hateful content one post at a time.