gender-bias-bengali/wonbias-complete-dataset

WoNBias: A Bengali Dataset for Gender Bias Detection

Dataset Details

Overview

A manually annotated corpus of Bengali designed to identify gender-based biases, stereotypes, and harmful language against women. Supports research in ethical NLP, content moderation, and computational social science.

Dataset Description

Basic Info

• —Purpose: Detect gender bias and harmful language against women in Bengali text
• —Language: Bengali (Bangla)
• —Labels:
• —0: Neutral (no bias)
• —1: Positive (supportive)
• —2: Negative (contains bias/harmful content)

Data Sources

Collected from:

• —Digital sources:
• —Social media
• —Blogs
• —News platforms
• —Offline sources:
• —Surveys
• —Focus groups
• —Existing Bengali corpora

Collection Methods

• —Hybrid approach combining:
• —Digital scraping (public content only)
• —Offline participatory methods
• —Ethical merging of existing corpora
• —All personal identifiers removed

Partnerships

Developed with:

• —Community organizations (for offline data collection)
• —Local annotation teams

Dataset Structure

• —Format: CSV
• —Columns:
• —Data: Raw Bengali text
• —Label: Integer (0, 1, or 2)

Statistics

• —Total Samples: 31,484
• —Label Distribution:
• —Neutral (0): 10,658 samples (33.86%)
• —Positive (1): 10,170 samples (32.31%)
• —Negative (2): 10,656 samples (33.84%)

Usage

How to Load

from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("gender-bias-bengali/wonbias-complete-dataset")

# Load specific split if available
# dataset = load_dataset("gender-bias-bengali/wonbias-complete-dataset", split="train")

Examples

# Load the dataset
from datasets import load_dataset
dataset = load_dataset("gender-bias-bengali/wonbias-complete-dataset")

# View dataset structure
print(dataset)

# View a sample
print("Sample entry:")
print(f"Text: {dataset['train'][0]['Data']}")
print(f"Label: {dataset['train'][0]['Label']}")

# View label distribution
from collections import Counter
labels = [item['Label'] for item in dataset['train']]
print("Label distribution:", Counter(labels))

# Example usage for classification
import pandas as pd
df = dataset['train'].to_pandas()
print(df.head())

Ethical Considerations

Data Collection

• —Anonymization:
• —All user identifiers removed
• —Metadata sanitization
• —Consent:
• —Explicit consent for survey/focus group participants
• —Public content used under platform ToS

Content Safeguards

• —Mental Health Protections:
• —Psychological support for annotators
• —Scheduled breaks during toxic content review
• —Trigger Warnings:
• —Documentation flags harmful content

Limitations

• —Scope Focus:
• —Targets gender bias against women
• —Does not cover third gender/intersectional biases
• —Detection Challenges:
• —Difficulty capturing sociocultural nuances
• —Performance variations across dialects
• —Language Specificity:
• —Bengali-only (addresses critical gap for low-resource language)

Related Work

This dataset contributes to the growing body of work in:

• —Bengali NLP and low-resource language processing
• —Gender bias detection in social media
• —Cross-cultural studies of online harassment
• —Ethical AI and responsible NLP

Contact Information

For questions, issues, or collaboration opportunities, please contact:

• —Primary Contact: Md. Raisul Islam Aupi, Nishat Tafannum
• —Academic Inquiries: Available through institutional affiliations

Version History

• —v1.0 (2025): Initial release with 31,484 manually annotated samples
• —Future versions may include:
• —Additional dialectical variations
• —Extended label taxonomy
• —Cross-platform validation

Citation

@dataset{wonbias2025,
  title = {WoNBias: A Bengali Dataset for Gender Bias Detection},
  author = {Md. Raisul Islam Aupi and Nishat Tafannum and Md. Shahidur Rahman and Kh Mahmudul Hassan and Naimur Rahman},
  year = {2025},
  url = {https://huggingface.co/datasets/gender-bias-bengali/wonbias-complete-dataset},
  license = {CC-BY-NC-SA-4.0},
  note = {Available on Hugging Face Datasets}
}

License


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

License Summary:

• —✅ Share — copy and redistribute the material
• —✅ Adapt — remix, transform, and build upon the material
• —❌ Commercial Use — not for commercial purposes
• —📝 Attribution — must give appropriate credit
• —🔄 ShareAlike — must distribute under same license