Abstract Warning: This paper contains insulting statements that may cause discomfort for readers. The rapid proliferation of digital platforms has intensified online hate speech, especially in low-resource languages such as Tamil, where automated moderation techniques remain underdeveloped. This paper presents a three-stage methodology for generating counter narratives in Tamil. First, a seed dataset of 220 hate speech counter narrative (HS-CN) pairs is expanded to 5,000 through a human-in-the-loop Author Reviewer framework with expert validation. Second, a fact-based retrieval augmented generation (RAG) system is employed to incorporate external knowledge to enhance factual accuracy and persuasiveness. Finally, the human post-edited dataset is integrated to the RAG system as a curated knowledge base yielding a Fact-RAG system with stronger factual grounding and cultural appropriateness. Assessment through intrinsic indicators and LLM-based evaluations indicates that our methodology generates counter-narratives that are varied, credible, and contextually relevant. These findings underscore the effectiveness of integrating human supervision, factual validation, and selected examples for counter-narrative development in low-resource settings. GitHub: https://github.com/Bharathi-AI-for-Social-Good/Fact-RAG-BasedCN-Ta
Ravikumar et al. (Fri,) studied this question.