OSCaR: Orthogonal Subspace Correction and Rectification of Biases in Word Embeddings

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5034–5050 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5034 OSCAR: Orthogonal Subspace Correction and Rectification of Biases in Word Embeddings Sunipa Dev UCLA Tao Li University of Utah Jeff M Phillips University of Utah Vivek Srikumar University of Utah sunipa@cs.ucla.edu {tli,jeffmp,svivek}@cs.utah.edu Abstract Language representations are known to carry certain associations (e.g., gendered connota- tions) which may lead to invalid and harm- ful predictions in downstream tasks. While existing methods are effective at mitigating such unwanted associations by linear projec- tion, we argue that they are too aggressive: not only do they remove such associations, they also erase information that should be re- tained. To address this issue, we propose OS- CAR (Orthogonal Subspace Correction and Rectification), a balanced approach of miti- gation that focuses on disentangling associa- tions between concepts that are deemed prob- lematic, instead of removing concepts whole- sale. We develop new measurements for eval- uating information retention relevant to the debiasing goal. Our experiments on gender- occupation associations show that OSCAR is a well-balanced approach that ensures that se- mantic information is retained in the embed- dings and unwanted associations are also ef- fectively mitigated. 1 Introduction Word embeddings are used extensively across natu- ral language processing (NLP) and succinctly cap- ture not only the syntactic and semantic structure of language, but also word meaning in context. As such, word embeddings are essential building blocks for today’s state-of-the-art in NLP. But they are also known to capture a significant amount of stereotypical associations (e.g., Bolukbasi et al., 2016; Zhao et al., 2017; Dev and Phillips, 2019; Sun et al., 2019) related to gender, race, national- ity, or religion, which can manifest in unwanted and/or potentially harmful ways in downstream tasks (Webster et al., 2018; Zhao et al., 2019; Dev et al., 2020). Such potentially problematic associa- tions, when embedded in word representations, can lead to incorrect and unfair decisions about large groups of people. While the term “bias” has many meanings, in this paper we use it to refer to these unwanted stereotypical associations. Existing methods to mitigate these effects either require expensive retraining of vectors (Lemoine et al., 2018) which can be inefficient, or projecting out information contained along an entire subspace representing a protected concept (such as gender or race) in the embedding space (e.g., Bolukbasi et al., 2016; Dev and Phillips, 2019; Ravfogel et al., 2020). Projective approaches are difficult to control as they are either insufficient: removing a subspace can still leave residual bias (Gonen and Goldberg, 2019; Lauscher et al., 2020), or too aggressive: in the case of gender, also unnecessarily altering the association between the word pregnant and words like female and mother. In tasks such as coreference resolution, removing such associations could hinder reference resolution. To quantify how much valid information is re- tained, we look at the output space of models. Fol- lowing Dev et al. (2020), we use Natural Language Inference (NLI) as an …

Content truncated to 500 words