dataset_info:
features:
- name: image
dtype: image
- name: Meme ID
dtype: int64
- name: Language
dtype: string
- name: Caption
dtype: string
- name: US
dtype: float64
- name: DE
dtype: float64
- name: MX
dtype: float64
- name: CN
dtype: float64
- name: IN
dtype: float64
- name: Template Name
dtype: string
- name: Category
dtype: string
- name: Subcategory
dtype: string
splits:
- name: train
num_bytes: 62034606
num_examples: 1500
download_size: 62780586
dataset_size: 62034606
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
language:
- en
- de
- hi
- zh
- es
pretty_name: Multi^3Hate
size_categories:
- 1K<n<10K
Multi3Hate
Warning: this dataset contains content that may be offensive or upsetting
Multi3Hate dataset, introduced in our paper: Multi3Hate: Advancing Multimodal, Multilingual, and Multicultural Hate Speech Detection with Vision–Language Models
Abstract: Hate speech moderation on global platforms poses unique challenges due to the multimodal and multilingual nature of content, along with the varying cultural perceptions. How well do current vision-language models (VLMs) navigate these nuances? To investigate this, we create the first multimodal and multilingual parallel hate speech dataset, annotated by a multicultural set of annotators, called Multi3Hate. It contains 300 parallel meme samples across 5 languages: English, German, Spanish, Hindi, and Mandarin. We demonstrate that cultural background significantly affects multimodal hate speech annotation in our dataset. The average pairwise agreement among countries is just 74%, significantly lower than that of randomly selected annotator groups. Our qualitative analysis indicates that the lowest pairwise label agreement—only 67% between the USA and India—can be attributed to cultural factors. We then conduct experiments with 5 large VLMs in a zero-shot setting, finding that these models align more closely with annotations from the US than with those from other cultures, even when the memes and prompts are presented in the dominant language of the other culture.
Dataset Details
Dataset Sources [optional]
Who are the source data producers?
The original meme images were crawled from https://memegenerator.net
Annotation process
We refer to our paper.
License
CC BY-NC-ND 4.0
Citation [optional]
BibTeX:
@misc{bui2024multi3hatemultimodalmultilingualmulticultural,
title={Multi3Hate: Multimodal, Multilingual, and Multicultural Hate Speech Detection with Vision-Language Models},
author={Minh Duc Bui and Katharina von der Wense and Anne Lauscher},
year={2024},
eprint={2411.03888},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2411.03888},
}