It’s All in the Name: Mitigating Gender Bias with Name-Based Counterfactual Data Substitution

Abstract

This paper treats gender bias latent in word embeddings. Previous mitigation attempts rely on the operationalisation of gender bias as a projection over a linear subspace. An alternative approach is Counterfactual Data Augmentation (CDA), in which a corpus is duplicated and augmented to remove bias, e.g. by swapping all inherently-gendered words in the copy. We perform an empirical comparison of these approaches on the English Gigaword and Wikipedia, and find that whilst both successfully reduce direct bias and perform well in tasks which quantify embedding quality, CDA variants outperform projection-based methods at the task of drawing non-biased gender analogies by an average of 19% across both corpora. We propose two improvements to CDA: Counterfactual Data Substitution (CDS), a variant of CDA in which potentially biased text is randomly substituted to avoid duplication, and the Names Intervention, a novel name-pairing technique that vastly increases the number of words being treated. CDA/S with the Names Intervention is the only approach which is able to mitigate indirect gender bias: following debiasing, previously biased words are significantly less clustered according to gender (cluster purity is reduced by 49%), thus improving on the state-of-the-art for bias mitigation.

Gender-preserving Debiasing for Pre-trained Word Embeddings

“…Our proposed method is agnostic to the details of the algorithms used to learn the input word embeddings. Moreover, unlike counterfactual data augmentation methods for debiasing (Zmigrod et al, 2019;Hall Maudslay et al, 2019), we do not require access to the original training resources used for learning the input word embeddings.…”

The practical ethics of bias reduction in machine translation: why domain adaptation is better than data debiasing

“…While this method inevitably introduces some translation errors, it at least creates a more male-female-balanced dataset. 8 As previously discussed, this counterfactual data augmentation approach is closest in spirit to prior work on reducing gender bias in monolingual NLP tools (Maudslay et al 2019;Zhao et al 2018;Zmigrod et al 2019).…”

“…Given this, the research presented in this article is more closely related to studies which have sought to balance masculine and feminine terms in the training data itself. Such approaches exist for English data (e.g., Maudslay et al 2019;Zhao et al, 2018). Zmigrod et al (2019) introduced a more complicated data-augmentation scheme for inflected languages with rich morphology like German, but their scheme revolves around swapping a single targeted word per sentence.…”

“…Recently, Blodgett et al (2020) have argued that several distinct and often underspecified conceptualisations of bias has been prominent in work addressing the skewed datasets used for Natural Language Processing (NLP) tasks, and the authors show how this has fostered terminological imprecision. As Maudslay et al (2019) put it, the 'measurement and mitigation of gender bias relies on the chosen operationalisation of gender bias. As a direct consequence, how researchers choose to operationalise bias determines both the techniques at one's disposal to mitigate bias, as well as the yardstick by which success is determined'.…”

A Survey on Bias in Deep NLP

“…Lu et al [33] tries to reduce the detected bias by using the data augmentation technique Contextual Data Augmentation CDA which consists of adding a complementary gender phrase to the sentences of the initial dataset. Based on CDA, in 2019 Hall Maudslay et al [65] will develop Contextual Data Substitution CDS. It proposes to eliminate the bias associated with proper names by adding a phrase with a complementary gender name in a balanced way.…”