A New Measure of Polarization in the Annotation of Hate Speech

The number of social media users is ever-increasing. Unfortunately, this has also resulted in the massive rise of uncensored online hate against vulnerable communities such as immigrants, LGBT and women. Current work on the automatic detection of various forms of hate speech (HS) typically employs supervised learning, requiring manually annotated data. The highly polarizing nature of the topics involved raises concerns about the quality of annotations these systems rely on, because not all the annotators are equally sensitive to different kinds of hate speech. We propose an approach to leverage the fine-grained knowledge expressed by individual annotators, before their subjectivity is averaged out by the gold standard creation process. This helps us to refine the quality of training sets for hate speech detection. We introduce a measure of polarization at the level of single instances in the data to manipulate the training set and reduce the impact of most polarizing text on the learning process. We test our approach on three datasets, in English and Italian, annotated by experts and workers hired on a crowdsourcing platform. We classify instances of sexist, racist, and homophobic hate speech in tweets and show how our approach improves the prediction performance of a supervised classifier. Moreover, the proposed polarization measure helps towards the manual exploration of the individual instances of tweets in our datasets.