README.md

metadata

configs:
  - config_name: task1
    dataset_name: CUET-NLP/BHM-Bengali-Hateful-Memes
    data_files:
      - split: train
        path: data/train-task1.csv
      - split: valid
        path: data/valid-task1.csv
      - split: test
        path: data/test-task1.csv
  - config_name: task2
    dataset_name: CUET-NLP/BHM-Bengali-Hateful-Memes
    data_files:
      - split: train
        path: data/train-task2.csv
      - split: valid
        path: data/valid-task2.csv
      - split: test
        path: data/test-task2.csv
license: mit
language:
  - bn
tags:
  - Hateful-Memes
  - Multimodal
  - Bengali
size_categories:
  - 1K<n<10K
task_categories:
  - other
  - image-classification
  - image-to-text
dataset_info:
  - images_path: Memes/
    columns:
      - image_name: string
      - Captions: string
      - Labels: int32

Dataset Description

BHM is a novel multimodal dataset for Bengali Hateful Memes detection. The dataset consists of 7,148 memes with Bengali as well as code-mixed captions, tailored for two tasks: (i) detecting hateful memes and (ii) detecting the social entities they target (i.e., Individual, Organization, Community, and Society).

Paper Information

• Paper: https://aclanthology.org/2024.acl-long.454/
• Code: https://github.com/eftekhar-hossain/Bengali-Hateful-Memes/tree/main

Data Downloading

The data examples were divided into three subsets for each task: train, valid, and test.

You can download this dataset by the following command (make sure that you have installed Huggingface Datasets):

from datasets import load_dataset

task1 = load_dataset("Eftekhar/BHM-Bengali-Hateful-Memes", "task1")  # Binary Classification
task2 = load_dataset("Eftekhar/BHM-Bengali-Hateful-Memes", "task2")  # Multiclass Classification

Data Format

The dataset is provided in CSV format and contains the following attributes:

{
    "image_name": [string] The name of the image,
    "Captions": [string] Embedded text on the corresponding images
    "Labels": [string] Class labels.   
}

Citation

If you use the BHM dataset in your work, please kindly cite the paper using this BibTeX:

@article{hossain2024deciphering,
  title={Deciphering Hate: Identifying Hateful Memes and Their Targets},
  author={Hossain, Eftekhar and Sharif, Omar and Hoque, Mohammed Moshiul and Preum, Sarah M},
  journal={arXiv preprint arXiv:2403.10829},
  year={2024}
}