Adversarial Generation of Multi-Turn Debates for Hate Speech Mitigation via Retrieval-Augmented Generation

The pervasive spread of online hate speech on social media has fostered the development of counter-narrative generation strategies based on Large Language Models (LLMs). Previous approaches to mitigate hate speech content has either fine-tuned LLMs on human-curated datasets consisting of single-turn interactions, i.e., pairs of hate speech and counternarratives, or leveraged in-context learning strategies. However, these approaches fall short when contextspecific counter-narratives are unavailable or lack sufficient evidence. To overcome this issue, in this work we propose a novel method that combines an adversarial LLM-based approach to produce multi-turn debates with hate speech counter-narrative generation based on Retrieval-Augmented Generation (RAG). In the early adversarial learning stage, the LLM agent defends an unethical positions whereas the adversarial responds with ethical, evidence-based counter-arguments. A human evaluation of the generated counter-narratives confirmed their relevance and pertinence. Then, a RAG system retrieves the evidence-based counter-narratives and generates the mitigated version of the hate speech content. Experimental results show that the RAG approach outperforms state-of-the-art methods on the MultiTarget CONAN benchmark dataset. Implementation details, prompts, and further information for reproducibility available here: https://github.com/erfan-bayat13/debaterag-counterspeech Notice: this paper includes potentially offensive content for scientific purposes only.