WinObias - Gender Bias Detection
Specialized corpus to detect and analyze gender biases in the task of resolving co-reference using “Winograd” standard sentences.
Description
The dataset WinObias is a corpus built to highlight gender biases in automatic language processing models. It is based on Winograd-type sentences where resolving co-references can reveal implicit stereotypes. It is divided into four subsets based on whether they reinforce or contradict gender stereotypes, and whether they require real-world knowledge or syntactic analysis.
What is this dataset for?
- Assessing gender biases in NLP systems
- Test the performance of coreference resolution models
- Designing ethical approaches to reduce bias in AI
Can it be enriched or improved?
Yes, this dataset can be enriched with examples in other languages or cultural contexts. It is also possible to further annotate sentences with metadata or to create new Winograd cases with other professions or stereotypes.
🔎 In summary
🧠 Recommended for
- Ethical NLP researchers
- LLM developers
- Responsible AI teams
🔧 Compatible tools
- SpacY
- AllennLP
- Hugging Face
- Transformers
- Scikit-learn
💡 Tip
For more relevant results, compare the performance of the same model on the stereotyped and anti-stereotyped versions.
Frequently Asked Questions
Can this dataset be used to train models?
It is primarily designed for evaluation, but it can be used in fine-tuning phases that focus on bias.
What professions are represented in this corpus?
The dataset contains typical professions such as doctor, nurse, mechanic, etc., used in ambiguous contexts.
Is it suitable for multi-lingual experiences?
The corpus is in English only, but it can be adapted for other languages with sentence reformulation.
