LLM-C3MOD: A Human-LLM Collaborative System for Cross-Cultural Hate Speech Moderation

2025-03-10

AI Summary

Global content moderation suffers from linguistic resource imbalance: mainstream platforms rely on high-resource-language moderators, leaving low-resource languages underrepresented and prone to misclassifying cross-cultural hate speech due to cultural misinterpretation. This paper proposes a human-AI collaborative moderation framework comprising three stages: (1) RAG-based initial screening enhanced with culturally contextualized prompts; (2) LLM-driven identification of consensus gaps across multiple cultural perspectives; and (3) targeted human review triggered by such gaps. We introduce the first RAG-based cultural annotation mechanism tailored for cross-cultural moderation, enabling explicit injection and dynamic retrieval of cultural knowledge. Evaluated on a Korean-language dataset, our approach achieves 78% accuracy—7 percentage points higher than GPT-4o—while reducing human workload by 83.6%. Results demonstrate that non-native moderators, augmented by culturally enriched LLMs, can effectively contribute to global content governance.

Technology Category

• Natural Language Processing: Language Grounding & Multi-modal NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: AI for Accessibility

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Content moderation is a global challenge, yet major tech platforms prioritize high-resource languages, leaving low-resource languages with scarce native moderators. Since effective moderation depends on understanding contextual cues, this imbalance increases the risk of improper moderation due to non-native moderators' limited cultural understanding. Through a user study, we identify that non-native moderators struggle with interpreting culturally-specific knowledge, sentiment, and internet culture in the hate speech moderation. To assist them, we present LLM-C3MOD, a human-LLM collaborative pipeline with three steps: (1) RAG-enhanced cultural context annotations; (2) initial LLM-based moderation; and (3) targeted human moderation for cases lacking LLM consensus. Evaluated on a Korean hate speech dataset with Indonesian and German participants, our system achieves 78% accuracy (surpassing GPT-4o's 71% baseline), while reducing human workload by 83.6%. Notably, human moderators excel at nuanced contents where LLMs struggle. Our findings suggest that non-native moderators, when properly supported by LLMs, can effectively contribute to cross-cultural hate speech moderation.

Problem

• Addresses cross-cultural hate speech moderation challenges
• Reduces reliance on non-native moderators' cultural understanding
• Combines human-LLM collaboration for accurate moderation

Innovation

• RAG-enhanced cultural context annotations
• LLM-based initial moderation
• Targeted human moderation for unclear cases