Gender-preserving Debiasing for Pre-trained Word Embeddings

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1641–1650 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1641 Gender-preserving Debiasing for Pre-trained Word Embeddings Masahiro Kaneko Tokyo Metropolitan University, Japan kaneko-masahiro@ed.tmu.ac.jp Danushka Bollegala University of Liverpool, UK danushka@liverpool.ac.uk Abstract Word embeddings learnt from massive text collections have demonstrated significant lev- els of discriminative biases such as gender, racial or ethnic biases, which in turn bias the down-stream NLP applications that use those word embeddings. Taking gender-bias as a working example, we propose a debias- ing method that preserves non-discriminative gender-related information, while remov- ing stereotypical discriminative gender biases from pre-trained word embeddings. Specif- ically, we consider four types of informa- tion: feminine, masculine, gender-neutral and stereotypical, which represent the relationship between gender vs. bias, and propose a de- biasing method that (a) preserves the gender- related information in feminine and masculine words, (b) preserves the neutrality in gender- neutral words, and (c) removes the biases from stereotypical words. Experimental re- sults on several previously proposed bench- mark datasets show that our proposed method can debias pre-trained word embeddings bet- ter than existing SoTA methods proposed for debiasing word embeddings while preserving gender-related but non-discriminative infor- mation. 1 Introduction Despite the impressive success stories behind word representation learning (Devlin et al., 2018; Peters et al., 2018; Pennington et al., 2014; Mikolov et al., 2013c,a), further investigations into the learnt representations have revealed sev- eral worrying issues. The semantic representations learnt, in particular from social media, have shown to encode significant levels of racist, offensive and discriminative language usage (Bolukbasi et al., 2016; Zhao et al., 2018b; Elazar and Goldberg, 2018; Rudinger et al., 2018; Zhao et al., 2018a). For example, Bolukbasi et al. (2016) showed that word representations learnt from a large (300GB) news corpus to amplify unfair gender biases. Mi- crosoft’s AI chat bot Tay learnt abusive language from Twitter within the first 24 hours of its release, which forced Microsoft to shutdown the bot (The Telegraph, 2016). Caliskan et al. (2017) con- ducted an implicit association test (IAT) (Green- wald et al., 1998) using the cosine similarity mea- sured from word representations, and showed that word representations computed from a large Web crawl contain human-like biases with respect to gender, profession and ethnicity. Given the broad applications of pre-trained word embeddings in various down-stream NLP tasks such as machine translation (Zou et al., 2013), sentiment analysis (Shi et al., 2018), dia- logue generation (Zhang et al., 2018) etc., it is im- portant to debias word embeddings before they are applied in NLP systems that interact with and/or make decisions that affect humans. We believe that no human should be discriminated on the basis of demographic attributes by an NLP sys- tem, and there exist clear legal (European Union, 1997), business and ethical obligations to make NLP systems unbiased (Holstein et al., 2018). Despite the growing need for unbiased word embeddings, debiasing pre-trained word embed- dings is a challenging task …

Content truncated to 500 words