A Computational Approach to Discovering Emergent Hate Speech on Social Media

Abstract

The growth of social media has enabled unprecedented global connectivity and access to information, but social media platforms have also become channels for hate speech that targets individuals and groups based on identity and protected characteristics. While advances in machine learning and natural language processing have improved hate speech detection, these models are static and cannot be updated easily in an environment where terminology evolves continuously. Although manual inspection can identify emerging terminology, the volume of posts often proves prohibitive. We propose a novel computational methodology and app, the Hate Mitigation App (HMApp), that discovers emerging coded or non-coded hate speech terminology on social media. Specifically, the approach uses a set of researcher-selected anchor hate terms to identify new terminology associated with a specific type of hatred, such as antisemitism. We use this approach with five different types of hatred: antisemitism, anti-Black hatred, Islamophobia, anti-Asian hatred, and anti-LGBTQ+ hatred. Using two categories of evaluation metrics—one comparing discovered terms against existing lexicons and another validating them through online searches—we find that the approach identifies both established and novel terminology. Our findings underscore the potential of computational approaches to augment manual detection efforts and keep pace with the rapidly evolving nature of online hate speech. Though related to both, our approach is neither Hate Speech Detection nor Automated Terminology Extraction, which are often construed as classification tasks. Our method, in contrast, uses unsupervised learning, thus engaging in a discovery rather than a classification task, yielding ramifications for both our design and testing methodologies.