Style Pooling: Automatic Text Style Obfuscation for Improved Classification Fairness

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2009–2022 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2009 Style Pooling: Automatic Text Style Obfuscation for Improved Classification Fairness Fatemehsadat Mireshghallah, Taylor Berg-Kirkpatrick University of California San Diego, {fatemeh, tberg}@ucsd.edu Abstract Text style can reveal sensitive attributes of the author (e.g. race or age) to the reader, which can, in turn, lead to privacy violations and bias in both human and algorithmic decisions based on text. For example, the style of writing in job applications might reveal protected attributes of the candidate which could lead to bias in hir- ing decisions, regardless of whether hiring de- cisions are made algorithmically or by humans. We propose a VAE-based framework that ob- fuscates stylistic features of human-generated text through style transfer by automatically re-writing the text itself. Our framework oper- ationalizes the notion of obfuscated style in a flexible way that enables two distinct notions of obfuscated style: (1) a minimal notion that effectively intersects the various styles seen in training, and (2) a maximal notion that seeks to obfuscate by adding stylistic features of all sensitive attributes to text, in effect, computing a union of styles. Our style-obfuscation framework can be used for multiple purposes, however, we demonstrate its effectiveness in improving the fairness of downstream classi- fiers. We also conduct a comprehensive study on style pooling’s effect on fluency, semantic consistency, and attribute removal from text, in two and three domain style obfuscation.1 1 Introduction Machine learning (ML) algorithms are used in a wide range of tasks, including high-stakes applications like determining credit ratings, setting insurance policy rates, making hiring decisions, and performing facial recognition. It has been shown that such algorithms can produce outcomes that are biased towards a certain gender or race (Buo- lamwini and Gebru, 2018; Silva et al., 2021; Sheng et al., 2021). Ideally, high-stakes decisions made by either humans or ML algorithms, should not be influenced by irrelevant, protected attributes like nationality, 1Code, models, and data is available at https: //github.com/mireshghallah/style-pooling age, or gender. In many instances, the input data used for making high-stakes decisions is text that is authored by a human candidate – for example, hir- ing decisions are often based on bios and personal statements. Recent work (De-Arteaga et al., 2019) shows that automatic hiring-decision models trained on bios are less likely to select female candidates for certain roles (e.g. architect, software engineer, and surgeon) even when the gender of the author is not explicitly provided to the system. Bias is, of course, not limited to algorithmic decisions, humans make biased decisions based on text, even when the protected attributes of the author are not explicitly revealed (Pedreshi et al., 2008). Together, these results indicate that both algorithms and humans can (1) decipher protected attributes of authors based on stylistic features of text, and (2) whether consciously or not, be biased by this information. A large body of prior work has attempted to address algorithmic bias by modifying …

Content truncated to 500 words