Improving Implicit Hate Speech Detection via a Community-Driven Multi-Agent Framework
Ewelina Gajewska, Katarzyna Budzynska, Jarosław A Chudziak
TL;DR
The paper tackles implicit hate speech detection and fairness in online moderation by introducing a community-driven multi-agent framework that pairs a central Moderator Agent with dynamically constructed Community Agents representing specific demographic groups. It leverages publicly available knowledge (e.g., Wikipedia) and embedding-based reasoning to contextualize decisions, enabling identity-aware consultations when ambiguity arises. Empirical evaluation on the challenging ToxiGen dataset shows that this consultative approach outperforms zero-shot, few-shot, and chain-of-thought prompting methods, achieving higher true positive rates and balanced accuracy across six target groups. The work contributes to more equitable moderation by embedding socio-cultural context into automated judgments, with implications for safer and less biased online environments and avenues for scaling to additional communities and deployment scenarios.
Abstract
This work proposes a contextualised detection framework for implicitly hateful speech, implemented as a multi-agent system comprising a central Moderator Agent and dynamically constructed Community Agents representing specific demographic groups. Our approach explicitly integrates socio-cultural context from publicly available knowledge sources, enabling identity-aware moderation that surpasses state-of-the-art prompting methods (zero-shot prompting, few-shot prompting, chain-of-thought prompting) and alternative approaches on a challenging ToxiGen dataset. We enhance the technical rigour of performance evaluation by incorporating balanced accuracy as a central metric of classification fairness that accounts for the trade-off between true positive and true negative rates. We demonstrate that our community-driven consultative framework significantly improves both classification accuracy and fairness across all target groups.