How do multi-modal large language models understand non-English visual hate? Insights from studying hate speech in Chinese-speaking communities on Instagram

This study provides a critical analysis of the efficacy of Multimodal Large Language Models (MLLM) in identifying visual hate speech on Instagram, such as image memes, specifically within the context of non-English and non-Western communities. By focusing on the unique dynamics of hate speech circulating among Chinese-speaking populations, particularly aimed at mainland Chinese individuals, this research illuminates the complexities and challenges associated with employing MLLMs for multi-modal hate speech detection through a zero-shot learning approach. Through a comparative evaluation of two cutting-edge MLLMs, Gemini-1.5 and GPT-4o-mini, measured against expert annotations and incorporating qualitative error analysis, the study reveals factors contributing to the complexity of the task. This includes hallucinations, tendencies toward over-labelling content as hate speech, and a notable absence of linguistic and cultural sensitivity. These findings highlight the needs for the development of culturally attuned models and methodologies that enhance the effectiveness of hate speech moderation in diverse cultural contexts.