mohdghazanfar / debiasurdunlp

This repository contains code for mitigating gender biases in Urdu word embeddings. It offers algorithm for identifying biases, algorithms for debiasing, and evaluation metrics to assess effectiveness. Designed for researchers and developers, it aims to promote fairness in NLP applications.

AI_GenderBiases: Debiasing Urdu NLP Models

Overview

This project focuses on mitigating the negative impacts of gender biases in pre-trained NLP models, specifically for the Urdu language. Through the provided scripts and methodologies, users can debias Urdu word embeddings created using a skip-gram model. Our approach involves adjusting word embeddings to reduce gender bias. Then, by performing embedding surgery, the model (roberta-urdu-small) is gender-debiased at an architectural level, ensuring more equitable and accurate language processing.

Urdu Word Embeddings

The Urdu Word Embedding created for this project utilizes the skip-gram model technique. For an in-depth look into the embedding creation process, please refer the paper: https://aclanthology.org/L18-1155.pdf

Scripts

The repository includes Python scripts designed to learn and mitigate gender biases in word embeddings:

learn_gender_specific.py

Expands a seed set of gender-specific words into a comprehensive list based on given word embeddings.

Run Command:

python learn_gender_specific.py ../embeddings/urduvec_140M_100K_300d.bin 50000 ../data/urdu_gender_specific_seed.json urdu_gender_specific_full.json

debias.py

Adjusts a word embedding by considering sets of gender pairs, gender-specific words, and pairs to equalize, producing a new, debiased word embedding.

Run Command:

python debias.py ../embeddings/urduvec_140M_100K_300d.bin ../data/urdu_definitional_pairs.json ../data/urdu_gender_specific_full.json ../data/urdu_equalize_pairs.json ../debiased_embeddings_140M_100K_300d.bin

Data Files

The project utilizes various data files to identify and mitigate gender bias effectively:

gender_specific_seed.json

Contains 101 gender-specific words.

gender_specific_full.json

Expanded list of 650 gender-specific words.

definitional_pairs.json

20 pairs of words defining the gender direction.

equalize_pairs.json

Crowdsourced male-female pairs of words representing gender direction.

How to Use

For detailed instructions on setting up the environment, running the scripts, and employing the debiased embeddings for further NLP tasks, please refer to the demo notebook.

To run the application, first clone the model (roberta-urdu-small) from https://huggingface.co/urduhack/roberta-urdu-small and place app.py in its directory.

Run Command:

https://huggingface.co/urduhack/roberta-urdu-small

app.py

Performs embedding surgery on the NLP model roberta-urdu-small, meaning it swaps its word embeddings with our gender-debiased embeddings on an architectural level. The results can be seen using an interface in which an Urdu word prediction model is integrated.

Run Command:

streamlit run app.py

Contributing

Contributions are welcome! If you have suggestions for improving the debiasing process or other aspects of the project, please review our contribution guidelines and submit a pull request.

Contact

For questions or collaboration inquiries, please open an issue on GitHub.