Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them

Proceedings of NAACL-HLT 2019, pages 609–614 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 609 Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them Hila Gonen1 and Yoav Goldberg1,2 1Department of Computer Science, Bar-Ilan University 2Allen Institute for Artificial Intelligence {hilagnn,yoav.goldberg}@gmail.com Abstract Word embeddings are widely used in NLP for a vast range of tasks. It was shown that word embeddings derived from text corpora reflect gender biases in society. This phe- nomenon is pervasive and consistent across different word embedding models, causing se- rious concern. Several recent works tackle this problem, and propose methods for signifi- cantly reducing this gender bias in word em- beddings, demonstrating convincing results. However, we argue that this removal is super- ficial. While the bias is indeed substantially reduced according to the provided bias defi- nition, the actual effect is mostly hiding the bias, not removing it. The gender bias infor- mation is still reflected in the distances be- tween “gender-neutralized” words in the debi- ased embeddings, and can be recovered from them. We present a series of experiments to support this claim, for two debiasing meth- ods. We conclude that existing bias removal techniques are insufficient, and should not be trusted for providing gender-neutral modeling. 1 Introduction Word embeddings have become an important component in many NLP models and are widely used for a vast range of downstream tasks. How- ever, these word representations have been proven to reflect social biases (e.g. race and gender) that naturally occur in the data used to train them (Caliskan et al., 2017; Garg et al., 2018). In this paper we focus on gender bias. Gender bias was demonstrated to be consistent and per- vasive across different word embeddings. Boluk- basi et al. (2016b) show that using word em- beddings for simple analogies surfaces many gen- der stereotypes. For example, the word embed- ding they use (word2vec embedding trained on the Google News dataset1 (Mikolov et al., 2013)) an- 1https://code.google.com/archive/p/word2vec/ swer the analogy “man is to computer program- mer as woman is to x” with “x = homemaker”. Caliskan et al. (2017) further demonstrate associ- ation between female/male names and groups of words stereotypically assigned to females/males (e.g. arts vs. science). In addition, they demon- strate that word embeddings reflect actual gender gaps in reality by showing the correlation between the gender association of occupation words and labor-force participation data. Recently, some work has been done to reduce the gender bias in word embeddings, both as a post-processing step (Bolukbasi et al., 2016b) and as part of the training procedure (Zhao et al., 2018). Both works substantially reduce the bias with respect to the same definition: the projection on the gender direction (i.e. −→ he − −→ she), introduced in the former. They also show that performance on word similarity tasks is not hurt. We argue that current debiasing methods, which lean on the above definition for gender bias and directly target …

Content truncated to 500 words