Optimising Equal Opportunity Fairness in Model Training

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4073 - 4084 July 10-15, 2022 ©2022 Association for Computational Linguistics Optimising Equal Opportunity Fairness in Model Training Aili Shen∗♠ Xudong Han∗♠ Trevor Cohn♠ Timothy Baldwin♠♡ Lea Frermann♠ ♠ School of Computing and Information Systems The University of Melbourne Victoria 3010, Australia ♡ Department of Natural Language Processing, MBZUAI {aili.shen,t.cohn,tbaldwin,lfrermann}@unimelb.edu.au, xudongh1@student.unimelb.edu.au Abstract Real-world datasets often encode stereotypes and societal biases. Such biases can be im- plicitly captured by trained models, leading to biased predictions and exacerbating existing so- cietal preconceptions. Existing debiasing meth- ods, such as adversarial training and remov- ing protected information from representations, have been shown to reduce bias. However, a disconnect between fairness criteria and train- ing objectives makes it difficult to reason the- oretically about the effectiveness of different techniques. In this work, we propose two novel training objectives which directly optimise for the widely-used criterion of equal opportunity, and show that they are effective in reducing bias while maintaining high performance over two classification tasks. 1 Introduction and Background Modern neural machine learning has achieved great success across a range of classification tasks. How- ever, when applied over real-world data, especially in high-stakes settings such as hiring processes and loan approvals, care must be taken to assess the fairness of models. This is because real-world datasets generally encode societal preconceptions and stereotypes, thereby leading to models trained on such datasets to amplify existing bias and make biased predictions (i.e., models perform unequally towards different subgroups of individuals). This kind of unfairness has been reported over various NLP tasks, such as part-of-speech tagging (Hovy and Søgaard, 2015; Li et al., 2018; Han et al., 2021b), sentiment analysis (Blodgett et al., 2016; Shen et al., 2021), and image activity recognition (Wang et al., 2019; Zhao et al., 2017). Various methods have been proposed to miti- gate bias, including adversarial training, and pre- and post-processing strategies. Adversarial train- ing aims to make it difficult for a discriminator to ∗Equal contributors to this work. predict protected attribute values from learned rep- resentations (Han et al., 2021c; Elazar and Gold- berg, 2018; Madras et al., 2018). Pre- and post- processing strategies vary greatly in approach, in- cluding transforming the original dataset to re- duce protected attribute discrimination while re- taining dataset utility (du Pin Calmon et al., 2017), iteratively removing protected attribute informa- tion from (fixed) learned representations (Ravfogel et al., 2020), or reducing bias amplification by in- jecting corpus-level constraints during inference (Zhao et al., 2017). However, training strategies and optimisation ob- jectives are generally disconnected from fairness metrics which directly measure the extent to which different groups are treated (in)equitably. This makes it difficult to understand the effectiveness of previous debiasing methods from a theoretical perspective. In this work, we propose to explic- itly incorporate equal opportunity into our training objective, thereby achieving bias reduction. This paper makes the following contributions: 1. We are the first to propose …

Content truncated to 500 words