BanglaMultiHate: Multi-task Bangla Hate-speech Dataset

The BanglaMultiHate dataset collected public comments from YouTube videos using the YouTube API, primarily from Somoy TV, which is a popular Bangla News channel. The comments belong to 19 different categories, including Business, Celebrities, Disaster, Entertainment, Fashion, Geopolitics, Health, History, International, Lifestyle, Literature, Miscellaneous, National, Opinion, Politics, Religion, Science, Sports, and Technology, as well as 120 subcategories. The annotation agreement of this manually annotated dataset shows agreement scores of $0.71$, $0.84$, and $0.79$ for the type of hate, severity of hate, and target of hate tasks, respectively, indicating substantial to perfect agreement.

Shared Task: Bangla Hate Speech Identification

Check our shared task on this dataset: https://github.com/AridHasan/blp25_task1

Dataset

• training set
• development set
• test set

Data format

Each file uses the json format. A json object within the list of objects adheres to the following structure:

{
   "id": "",
   "comment": "",
   "category": "",
   "subcategory": "",
   "type_of_hate": "",
   "severity_of_hate": "",
   "target_of_hate": ""
}

Where:

• id: an index or id of the text
• comment: text
• category: a category
• subcategory: a subcategory
• hate_type: Abusive, Sexism, Religious Hate, Political Hate, Profane, or None.
• hate_severity: Little to None, Mild, or Severe.
• to_whom: Individuals, Organizations, Communities, or Society.

Example

{
   "id": "490273",
   "comment": "আওয়ামী লীগের সন্ত্রাসী কবে দরবেন এই সাহস আপনাদের নাই",
   "category": "National",
   "subcategory": "বাংলার সময়",
   "type_of_hate": "Political Hate",
   "severity_of_hate": "Little to None",
   "target_of_hate": "Organization"
}

How to download data

import os
import json
from datasets import load_dataset

output_dir="./BanglaMultiHate/"
 
dataset = load_dataset("AridHasan/BanglaMultiHate")
# Save the dataset to the specified directory. This will save all splits to the output directory.
dataset.save_to_disk(output_dir)
# iterate over splits to save the data into json format
for split in ['train','dev','test']:
    data = []
    if split not in dataset:
        continue
    for idx, item in enumerate(dataset[split]):
        data.append(item)
    output_file = os.path.join(output_dir, f"{split}.json")
    with open(output_file, 'w', encoding='utf-8') as f:
        json.dump(data, f, ensure_ascii=False, indent=4)

Citation

@article{hasan2025llm,
      title={LLM-Based Multi-Task Bangla Hate Speech Detection: Type, Severity, and Target}, 
      author={Hasan, Md Arid and Alam, Firoj and Hossain, Md Fahad and Naseem, Usman and Ahmed, Syed Ishtiaque},
      year={2025},
      journal={arXiv preprint arXiv:2510.01995},
      url={https://arxiv.org/abs/2510.01995},
}

@inproceedings{blp2025-overview-task1,
    title = "Overview of BLP 2025 Task 1: Bangla Hate Speech Identification",
    author = "Hasan, Md Arid and Alam, Firoj and Hossain, Md Fahad and Naseem, Usman and Ahmed, Syed Ishtiaque",
    booktitle = "Proceedings of the Second International Workshop on Bangla Language Processing (BLP-2025)",
    editor = {Alam, Firoj
          and Kar, Sudipta
          and Chowdhury, Shammur Absar
          and Hassan, Naeemul
          and Prince, Enamul Hoque
          and Tasnim, Mohiuddin
          and Rony, Md Rashad Al Hasan,
          and Rahman, Md Tahmid Rahman
    },
    month = dec,
    year = "2025",
    address = "India",
    publisher = "Association for Computational Linguistics",
}

License

BanglaMultiHate is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

You should have received a copy of the license along with this work. If not, see http://creativecommons.org/licenses/by-nc-sa/4.0/.