Beating Harmful Stereotypes Through Facts: RAG-based Counter-speech Generation (arXiv)

Although counter-speech production is essential for reducing harmful content, existing methods that rely on huge language models or expert input have scalability and reliability issues. Using Retrieval-Augmented Generation (RAG) pipelines, the current study presents a unique framework that models counter-speech as a knowledge-informed generation job. The framework focuses on eight important groups that are impacted by hate speech and draws from a curated knowledge base of 32,792 texts from reliable sources. The MultiTarget-CONAN dataset is used for empirical evaluation, which shows better performance than current baselines, laying the groundwork for scalable, reliable counter-speech creation.

https://arxiv.org/abs/2510.12316