Online hate speech has increased in tandem with the movement of public discourse online. Although a number of promising methods for classifying hate speech have been put forth, research frequently just looks at English and ignores three major issues: post-deployment performance, classifier maintenance, and infrastructure constraints. In this study, a novel human-in-the-loop BERT-based pipeline for classifying hate speech is presented, and its evolution from initial data collection and annotation to post-deployment is traced. With an F1 score of 80.5, our classifier surpasses the current top-performing BERT-based multilingual classifier by 5.8 F1 points in German and 3.6 F1 points in French. It was created for the naturally multilingual environment of Switzerland and was trained using data from a corpus of more than 422k samples.https://aclanthology.org/2022.findings-emnlp.548Share this:FacebookXLike this:Like Loading... Post navigation Two Weeks in Soft Security: Free Resources on Countering Extremism, Hate, and Disinformation, March 2025 (I/II) Can LLM guardrails prevent harassment or hate speech? (milvus)