CATCH-ME if you RAG: a dataset of Contextually Annotated multi-Turn Counterspeech against Hate and Misinformation Exchanges
Researchers have introduced CATCH-ME, a novel dataset designed to train natural language processing models to generate effective counterspeech against online hate speech and misinformation. This dataset is the first of its kind to address the intersection of these two threats across multiple turns and languages. It includes expert-curated dialogues in five languages, targeting hate speech directed at seven marginalized groups, and is grounded in verified external knowledge for factual accuracy. The dataset is particularly applicable for retrieval-augmented generation (RAG) systems, providing document- and chunk-level annotations to enhance the persuasiveness and factual grounding of generated counterspeech. AI
IMPACT This dataset could significantly improve AI's ability to combat online hate speech and misinformation by enabling more nuanced and factually grounded responses.