Paper Notes

Bolukbasi, Tolga, et al. “Man is to computer programmer as woman is to homemaker? Debiasing word embeddings.” Advances in Neural Information Processing Systems. 2016.

[pdf]

Contributions

• Eliminate gender biases in word embeddings
  • Remove Receptionist-Female; gender neutral words
  • Keep Queen-Female; generder specific words
• Prove linear separable between gender-neutral words and baised words

• Algorithm to “debias”

Examples

man - woman = compupter_programmer - homemaker

Details

• wordvec on Google News, 300d

• GloVe is tested and described in Appendix

Methods

• Gender occupation stereotype exists!
  • M-Turk
• Analogies
  • (she, he) -> (x, y)`
  • Overall, 72 out of 150 analogies were rated as gender-appropriate by five or more out of 10 crowd-workers, and 29 analogies were rated as exhibiting gender stereotype by five or more crowd-workers

  • The first principal component capture the gender subspace.

    

  • Gender direction? is it the corresponding eigenvector?
  • DirectBiase=0.8 -> high?
• Debiasing Algorithm
• identify gender subspace
• hard debiasing / soft debiasing
• Human evaluation

Interesting Related Work

1. Zhao, Jieyu, et al. “Men Also Like Shopping: Reducing Gender Bias Amplification using Corpus-level Constraints.” arXiv preprint arXiv:1707.09457 (2017).