Detecting Online Hate Speech: Annotation Approaches and Classifier Performance

• In Event: New Methods for Text-as-Data

Abstract

Efforts to study and curb online hate speech depend on capacity to detect it at scale. A critical step in state-of-the-art classification pipelines remains the production of a large number of high-quality annotations to train classifiers. While machine-learning methods assume the availability of adequate labeled data, the step is usually poorly documented. Consequently, there is little guidance available to construct a training set effectively and efficiently. The goal of this paper is to contrast different approaches as well as their performance, using a unique corpus of online comments developed in the context of a project on detecting and countering online hate speech in Switzerland. First, we compare how online comments were labeled, following identical instructions, by coders belonging to different groups: research assistants, activists, crowd workers, and citizen scientists. Second, we analyze differences in the performance of classifiers based on which sets of annotations were used to train them. The results illustrate the challenge and importance of constructing a high-quality training set, and provide direction for various kinds of difficult classification tasks, including (but not limited to) online hate speech.

Authors

• Natalia Umansky, University of Zurich

• Karsten Donnay, University of Zurich

• Fabrizio Gilardi, University of Zurich

• Dominik Hangartner, ETH Zurich