Understanding hate speech: The HateInsights dataset and model interpretability
Abstract
The persistence of hate speech continues to pose an obstacle in the realm of online social media. Despite the continuous evolution of advanced models for identifying hate speech, the critical dimensions of interpretability and explainability have not received proportional scholarly attention. In this article, we introduce the HateInsights dataset, a groundbreaking benchmark in the field of hate speech datasets, encompassing diverse aspects of this widespread issue. Within our dataset, each individual post undergoes thorough annotation from dual perspectives: firstly, conforming to the established 3-class classification paradigm that includes hate speech, offensive language, and normal discourse; secondly, incorporating rationales that outline specific segments of a post supporting the assigned label (categorized as hate speech, offensive language, or normal discourse). Our exploration yields a significant finding by harnessing cutting-edge state-of-the-art models: even models demonstrating exceptional proficiency in classification tasks yield suboptimal outcomes in crucial explainability metrics, such as model plausibility and faithfulness. Furthermore, our analysis underscores a promising revelation concerning models trained using human-annotated rationales. To facilitate scholarly progress in this realm, we have made both our dataset and codebase accessible to fellow researchers. This initiative aims to encourage collaborative involvement and inspire the advancement of the hate speech detection approach characterized by increased transparency, clarity, and fairness.