Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them
This project includes the experiments described in the paper:
"Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them", Hila Gonen and Yoav Goldberg, NAACL 2019.
Full reimplementation of the experiments is available in "remaining_bias_2016.ipynb" for Bolukbasi's embeddings, and in "remaining_bias_2018.ipynb" for Zhao's embeddings.
- Python 2.7
As a first step, download the nondebiased and debiased embeddings into data/embeddings/ from this folder (8 files):
- orig_w2v: Bolukbasi's embeddings, nodebiased
- hard_debiased_w2v: Bolukbasi's embeddings, debiased
- orig_glove: Zhao's embeddings, nodebiased
- gn_glove: Zhao's embeddings, debiased
These files are the original embeddings but with a preprocessing step (for fast loading, see source/save_embeds.py):
- Embeddings of Bolukbasi et al. (hard_debiased) are taken from nondebiased and debiased.
- Embeddings of Zhao et al. (gn_glove) are taken from nondebiased and debiased.
If you find this project useful, please cite the paper:
@inproceedings{GONEN19,
title={Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them},
author={Gonen, Hila and Goldberg, Yoav},
booktitle={Proceedings of NAACL-HLT},
year={2019}
}
If you have any questions or suggestions, please contact Hila Gonen.
This project is licensed under Apache License - see the LICENSE file for details.