Easy Adaptation to Mitigate Gender Bias in Multilingual Text Classification

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 717 - 723 July 10-15, 2022 ©2022 Association for Computational Linguistics Easy Adaptation to Mitigate Gender Bias in Multilingual Text Classification Xiaolei Huang Department of Computer Science, University of Memphis xiaolei.huang@memphis.edu Abstract Existing approaches to mitigate demographic biases evaluate on monolingual data, however, multilingual data has not been examined. In this work, we treat the gender as domains (e.g., male vs. female) and present a standard do- main adaptation model to reduce the gender bias and improve performance of text classi- fiers under multilingual settings. We evaluate our approach on two text classification tasks, hate speech detection and rating prediction, and demonstrate the effectiveness of our approach with three fair-aware baselines. 1 Introduction Recent research raises concerns that document clas- sification models can be discriminatory and can per- petuate human biases (Dixon et al., 2018; Borkan et al., 2019; Sun et al., 2019; Blodgett et al., 2020; Liang et al., 2020). Building fairness-aware classi- fiers is critical for the text classification task, such as hate speech detection and online reviews due to its rich demographic diversity of users. The fairness-aware classifiers aim to provide fair and non-discriminatory outcomes towards people or groups of people based on their demographic at- tributes, such as gender, age, or race. Fairness has been defined in different ways (Hardt et al., 2016) across downstream tasks; for mitigating biases in the text classification, existing research (Dixon et al., 2018; Heindorf et al., 2019; Han et al., 2021) has focused on group fairness (Chouldechova and Roth, 2018), under which document classifiers are defined as biased if the classifiers perform better for documents of some groups than for documents of other groups. Methods to mitigate demographic biases in text classification task focus on four main directions, data augmentation (Dixon et al., 2018; Park et al., 2018; Garg et al., 2019), instance weighting (Zhang et al., 2020; Pruksachatkun et al., 2021), debi- ased pre-trained embeddings (Zhao et al., 2017; Pruksachatkun et al., 2021), and adversarial train- ing (Zhang et al., 2018; Barrett et al., 2019; Han et al., 2021; Liu et al., 2021). The existing studies have been evaluated on English datasets contain- ing rich demographic variations, such as Wikipedia toxicity comments (Cabrera et al., 2018), senti- ment analysis (Kiritchenko and Mohammad, 2018), hate speech detection (Huang et al., 2020). How- ever, the methods of reducing biases in text clas- sifiers have not been evaluated under multilingual settings. In this study, we propose a domain adapta- tion approach using the idea of “easy adapta- tion” (Daumé III, 2007) and evaluate on the text classification task of two multilingual datasets, hate speech detection and rating prediction. We experi- ment with non-debiased classifiers and three fair- aware baselines on the gender attribute, due to its wide applications and easily accessible resources. The evaluation results of both non-debiased and de- biased models establish important benchmarks of group fairness on the multilingual settings. …

Content truncated to 500 words