tolga-b/debiaswe

Remove problematic gender bias from word embeddings.

Debiaswe: try to make word embeddings less sexist

🔴FAT* 2018 tutorial slides

Here we have the code and data for the following paper: Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings by Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Proceedings of NIPS 2016.

Just looking to download a debiased embedding?

You can download binary/txt hard debiased version of the Google's Word2Vec embedding trained on Google News (Origin: GoogleNews-vectors-negative300.bin.gz found here).

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 that I refer within this repo can be found in this folder.)

debiaswe's Issues

Cosine similarity: Top pca component vs she-he

Hello Debiaswe Research Team, Thank you for making the code related to your paper available. This is very helpful! I am writing to seek clarification on analyzing gender bias in word vectors associated with professions.

In your paper, you suggest using cosine similarity between a given profession vector and the top PCA component. I am trying to replicate the same in the wiki context. Unfortunately, I am getting results opposite to expected.

For example, when I compute the cosine similarity between the waitress vector (or nurse vector) and the top gender principal component, I get a -ve score. However, when I compute the cosine similarity between the same profession vector and she - he vector (as you show in the example here), I get a +ve score.

I am confused about why the sign flips when using PCA and straightforward gender vector. I request your help.

Thank you!
sbs

Soft Debiasing

Hi, I was wondering if you could release the code related to soft debiasing? It would be nice to see the comparison between soft and hard debiasing.

Indirect Bias

I was looking in the code and wondered if it handled the indirect bias. So I wonder if it really exists there and I just can not get it.

Another thing, in the python tutorial notebook, computed the professions with respect to the projection score,as
sp = sorted([(E.v(w).dot(v_gender), w) for w in profession_words])
so I wonder if this is the same like computing the direct bias.

Limitation for the size of original word2vec bin file

Does it has a limitation for the size of the original word2vec bin file? I am trying to use debias.py for a ~12G pre-trained word2vec, but it returns a memoryerror.

Debiasing racial bias

Hi, I was wondering if you're planning on release a set of debiased embeddings which also address racial bias? You mention it in the related paper that it exists but you don't address it here. Will there be a follow-up work on this. It would be good to have pretrained word embeddings which address this as well.