A Multi-Agent Parallel Management Method for Decision-Making in Countering Hate Speech

Hate speech refers to language that denigrates, discriminates against, or incites hatred towards individuals or groups based on characteristics such as race, religion, ethnicity, gender, or sexual orientation. Countering hate speech is crucial as it mitigates harmful effects, supports victims, and promotes a more inclusive and respectful online environment. Existing methods for countering hate speech, including natural language processing (NLP) techniques and human moderation, are quick to implement but often infringe on freedom of speech and provoke resistance. Additionally, counterspeech generation methods, while maintaining a balance of freedom of speech, often lack systematic and strategic effectiveness. To address these limitations, we propose MAPM, a Multi-Agent Parallel Management Method, for strategic planning and decision-making in countering hate speech on social media. MAPM leverages the in-context learning capabilities of large language models (LLMs) to construct a multi-agent decision-making framework and optimizes strategy planning through reinforcement learning. Furthermore, parallel systems create artificial virtual environments that interact with real systems, enabling multi-round iterations of counteractions and enhancing effectiveness through co-evolution and closed-loop feedback. Extensive experiments on hate speech datasets demonstrate that MAPM achieves excellent performance, providing a robust and adaptive solution for mitigating the impact of hate speech while supporting freedom of expression.