Dynamically Disentangling Social Bias from Task-Oriented Representations with Adversarial Attack

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3740–3750 June 6–11, 2021. ©2021 Association for Computational Linguistics 3740 Dynamically Disentangling Social Bias from Task-Oriented Representations with Adversarial Attack Liwen Wang1∗, Yuanmeng Yan1∗, Keqing He1,2, Yanan Wu1, Weiran Xu1 1Beijing University of Posts and Telecommunications, Beijing, China 2Meituan Group, Beijing, China {w_liwen,yanyuanmeng,yanan.wu,xuweiran}@bupt.edu.cn {hekeqing}@meituan.com Abstract Representation learning is widely used in NLP for a vast range of tasks. However, represen- tations derived from text corpora often reflect social biases. This phenomenon is pervasive and consistent across different neural models, causing serious concern. Previous methods mostly rely on a pre-specified, user-provided direction or suffer from unstable training. In this paper, we propose an adversarial disentan- gled debiasing model to dynamically decouple social bias attributes from the intermediate rep- resentations trained on the main task. We aim to denoise bias information while training on the downstream task, rather than completely remove social bias and pursue static unbiased representations. Experiments show the effec- tiveness of our method, both on the effect of debiasing and the main task performance. 1 Introduction Supervised neural networks have achieved remark- able success in a wide range of natural language processing (NLP) tasks. The fundamental capa- bility of these neural models is to learn effective feature representations (Bengio et al., 2013) for the downstream prediction task. Unfortunately, the learned representations frequently contain undesir- able biases with respect to things that we would rather not use for decision making. We refer to such inappropriate factors as protected attributes (Elazar and Goldberg, 2018a). Biased information has serious real-world consequences. For example, concerns have been raised about automatic resume filtering systems giving preference to male appli- cants when the only distinguishing factor is the applicants’ gender (Sun et al., 2019). In this paper, we focus on social bias, such as gender bias which is the preference or prejudice towards one gender over the other (Moss-Racusin et al., 2012), race bias and age bias. ∗The first two authors contribute equally. Weiran Xu is the corresponding author. Gender Profession male female engineer nurse female male Profession Boundary Representation Space Adv engineer nurse Gender Profession male female engineer nurse Representation Space Gender Boundary Figure 1: A demonstration of dynamically disentan- gling gender bias from the representations of down- stream representations. The left figure shows a biased profession classifier where the model prefers men as en- gineers and women as nurses. Our adversarial disentan- glement model makes the example features approach the boundary of the gender classifier. Therefore, the boundary of the profession classifier in the right fig- ure shifts closer to the horizontal state where gender in- formation is decoupled from the representations in the main task. From the perspective of the debiasing target, pre- vious debiasing works can be approximately clas- sified into two types, word embedding (Bolukbasi et al., 2016; Caliskan et al., 2017; Zhao et al., 2018; Manzini et al., 2019; Wang et al., 2020; Kumar et al., 2020) and sentence …

Content truncated to 500 words