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

TL;DR

This work tackles automatic counter-speech generation for hate speech by grounding CS in an authoritative knowledge base. It introduces a Retrieval-Augmented Generation pipeline that retrieves, summarizes, and grounds CS in documents from UN, EUR-Lex, and FRA to cover eight target groups, producing concise 2-sentence CS posts. Through extensive automatic, JudgeLM, and human evaluations on MTCo, the approach with semantically rich retrieval (BGE-M3) and GPT-family generators consistently outperforms baselines and state-of-the-art competitors, while maintaining high safety. The framework offers scalable, fact-grounded CS with practical social-media applicability and provides resources for further research and evaluation in hate-speech countering and beyond.

Abstract

Counter-speech generation is at the core of many expert activities, such as fact-checking and hate speech, to counter harmful content. Yet, existing work treats counter-speech generation as pure text generation task, mainly based on Large Language Models or NGO experts. These approaches show severe drawbacks due to the limited reliability and coherence in the generated countering text, and in scalability, respectively. To close this gap, we introduce a novel framework to model counter-speech generation as knowledge-wise text generation process. Our framework integrates advanced Retrieval-Augmented Generation (RAG) pipelines to ensure the generation of trustworthy counter-speech for 8 main target groups identified in the hate speech literature, including women, people of colour, persons with disabilities, migrants, Muslims, Jews, LGBT persons, and other. We built a knowledge base over the United Nations Digital Library, EUR-Lex and the EU Agency for Fundamental Rights, comprising a total of 32,792 texts. We use the MultiTarget-CONAN dataset to empirically assess the quality of the generated counter-speech, both through standard metrics (i.e., JudgeLM) and a human evaluation. Results show that our framework outperforms standard LLM baselines and competitive approach, on both assessments. The resulting framework and the knowledge base pave the way for studying trustworthy and sound counter-speech generation, in hate speech and beyond.

Figures (7)

• Figure 1: Examples of two CS generated with GPT-4o-mini: the first without RAG, the second one with RAG.
• Figure 2: Overview of our RAG-based CS generation framework. Step 1: paragraph retrieval from a domain-specific knowledge base; Step 2: LLM paragraph summarization; Step 3: CS generation conditioned on the summarized knowledge.
• Figure 3: Distribution of age of the 26 participants.
• Figure 4: Distribution of gender of the 26 participants.
• Figure 5: Distribution of the geographical area of origin of the 26 participants.