Debiasing Word Embeddings

This repository contains code for debiasing word embeddings based on the paper "Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings" by Bolukbasi et al.

Getting Started

To use the code in this repository, you will need to have Python 3 installed, along with the following packages:

numpy
scipy
scikit-learn
matplotlib
gensim

You can install these packages using pip:

pip install numpy scipy scikit-learn matplotlib gensim

Usage

The main file in this repository is debias.py, which contains code for loading word embeddings, debiasing them, and evaluating the debiased embeddings on various tasks.

To use this code, simply run the debias.py file with the appropriate command-line arguments. For example, to load word embeddings from a file and debias them using the "hard" debiasing method, you can run:

python debias.py --embedding_file embeddings.txt --debias_method hard

This will output the debiased embeddings to a file called embeddings_debiased.txt.

Demo

Download w2v_gnews_small.zip and unzip the file and place it inside embeddings/ directory. In the demo.py you will find a demonstration of how to use the code. The demo consists of:

Loading and visualizing word embeddings
Debiasing word embeddings using various methods
Evaluating the debiased embeddings on gender-related tasks

Python scripts:

learn_gender_specific.py: given a word embedding and a seed set of gender-specific words (like king, she, etc.), it learns a much larger list of gender-specific words
debias.py: given a word embedding, sets of gender-pairs, gender-specific words, and pairs to equalize, it outputs a new word embedding. This version basically reads/writes word2vec binary file format.

python learn_gender_specific.py ../embeddings/GoogleNews-vectors-negative300.bin 50000 ../data/gender_specific_seed.json gender_specific_full.json

python debias.py ../embeddings/GoogleNews-vectors-negative300.bin ../data/definitional_pairs.json ../data/gender_specific_full.json ../data/equalize_pairs.json ../embeddings/GoogleNews-vectors-negative300-hard-debiased.bin

We also have seed data used to debias and crowd data used to evaluate the embeddings.

Data files:

gender_specific_seed.json: A list of 218 gender-specific words
gender_specific_full.json: A list of 1441 gender-specific words
definitional_pairs.json: The ten pairs of words we use to define the gender direction
equalize_pairs.json: Some crowdsourced F-M pairs of words that represent gender direction

All external files can be found on the Google Drive of the authors of the paper.

saifamrkhail / debiaswe