Natural Language Gen (NLG) provides scalable ways to counter the rise in hate speech online without limiting free expression. Traditional NLG techniques, however, have limitations since they frequently yield irrelevant, safe, or repetitious replies. A three-module pipeline to increase response variety and relevance is presented in this research. The pipeline uses a generative model to create a variety of counterspeech possibilities, a BERT model to filter out grammatical answers, and a unique retrieval-based strategy to choose the most pertinent counterspeech. The pipeline’s ability to produce varied and pertinent counterspeech is demonstrated by experiments conducted on three datasets.https://aclanthology.org/2021.findings-acl.12.pdfShare this:FacebookXLike this:Like Loading... Post navigation New findings show online hate spreads harm far and wide (eSafety Commissioner, Australian Government) A matter of perception? investigating subjective and objective exposure to hate speech with a survey and mobile longitudinal linkage study (Information, Communication & Society)