2026
EMNLP
EMNLP 2026
Discerning Culturally-Aware Hate Framings for Hate Speech Identification in Memes
Abstract
AbstractThe ability of Large Visual Language Models (LVLMs) to identify hate speech in memes across several cultures requires (1) access to cultural knowledge that is not readily available to them; and (2) the ability to reason with this knowledge. This paper shows how culturally-aware Hate Frames can be acquired to inform hate speech identification across various cultures. The paper also shows that counterfactual reasoning is ideal for identifying hate speech and generating explanations. Promising results show not only improved hate speech identification in each culture, but also reduction of the performance gap of hate speech identification between different cultures.
🌉
Interdisciplinary Bridge
- Artificial Intelligence and Natural Language Processing