What is Behind the Implicit Hate Speech? A Double-layer Multi-agent Voting Detection Method via Explanation Generation

Implicit hate speech employs indirect language to convey hateful intentions, which is increasingly prevalent on social media and presents a significant societal challenge. In contrast to explicit hate speech, the implicit one usually involves the humanistic histories or cultural stories, which poses great challenge to existing methods that focus on mining linguistic features of text. Although some advanced methods have devoted efforts to explanation generation for implicit hate speech detection, the generated interpretations are not directly used for detection and the introduced Large Language Models (LLMs) are prone to hallucinate when faced with tasks characterized by unclear decision boundaries. In this paper, we propose a double-layer multi-agent voting detection method via explanation generation for implicit hate speech detection. The explanations are first generated to interpret the stories behind the speech, which can really address the difference between implicit and explicit hate speech. Then, these generated explanations are integrated into the prompt to guide LLMs for detection, and a double-layer multi-agent voting strategy is designed to eliminate the hallucinations of LLMs. Extensive experiments on both English and Chinese datasets demonstrate that our method surpasses the performance of SOTA methods.