Multi3Hate: Multimodal, Multilingual, and Multicultural Hate Speech Detection with Vision–Language Models (ACL Anthology)

Multimodal content, linguistic diversity, and cultural perception are the main causes of the difficulties in global hate speech moderation. We provide Multi3Hate, the first multimodal, multilingual parallel hate speech dataset with 300 memes in English, German, Spanish, Hindi, and Mandarin annotated by culturally varied annotators, to evaluate how vision-language models (VLMs) manage these complexity. With an average inter-country agreement of only 74% and a low of 67% between the USA and India, the data presented shows that cultural influence has a considerable impact on annotation. The necessity for culturally appropriate moderation systems is highlighted by experiments with five VLMs in zero-shot conditions, which reveal a bias toward US-centric annotations.

https://aclanthology.org/2025.naacl-long.490