Gender-preserving Debiasing for Pre-trained Word Embeddings

Masahiro Kaneko, Danushka Bollegala

Code and debiased word embeddings for the paper: "Gender-preserving Debiasing for Pre-trained Word Embeddings" (In ACL 2019). If you use any part of this work, make sure you include the following citation:

@inproceedings{Kaneko:ACL:2019,    
    title={Gender-preserving Debiasing for Pre-trained Word Embeddings},    
    author={Masahiro Kaneko and Danushka Bollegala},    
    booktitle={Proc. of the 57th Annual Meeting of the Association for Computational Linguistics (ACL)},    
    year={2019} 
}

Our experiment settings

• python==3.7.2
• gensim==3.7.1
• numpy==1.16.2
• pandas==0.24.2
• torch==1.1.0

How to train yourself

First download the necessary data listed below using following command.

• Trained glove and gn-glove
• SemBias dataset
• Female and male word lists

./download.sh

Perform debiasing for word embeddings and its evaluation. Debiased word embeddings are stored in debiased_embeddings in both bin and txt format. Name of word embeddings to be debiased as glove or gn and give as the first argument. For example,

./run.sh gn

You can also evaluate your word embaddings without training on SemBias:

python eval_word_embeddings.py -i path/to/your/embeddings

Our debiased word embeddings

You can directly download our debiased GP (glove) and GP (gn-glove).

License

See the LICENSE file.