Bridging Fairness and Environmental Sustainability in Natural Language Processing

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7817 - 7836 December 7-11, 2022 ©2022 Association for Computational Linguistics Bridging Fairness and Environmental Sustainability in Natural Language Processing Marius Hessenthaler1, Emma Strubell2, Dirk Hovy3, Anne Lauscher4 1Data and Web Science Group, University of Mannheim, Germany 2Language Technologies Institute, Carnegie Mellon University, U.S. 3MilaNLP, Bocconi University, Italy 4Data Science Group, University of Hamburg, Germany marius.hessenthaler@web.de, strubell@cmu.edu, dirk.hovy@unibocconi.it, anne.lauscher@uni-hamburg.de Abstract Fairness and environmental impact are impor- tant research directions for the sustainable de- velopment of artificial intelligence. However, while each topic is an active research area in natural language processing (NLP), there is a surprising lack of research on the interplay be- tween the two fields. This lacuna is highly problematic, since there is increasing evidence that an exclusive focus on fairness can actually hinder environmental sustainability, and vice versa. In this work, we shed light on this crucial intersection in NLP by (1) investigating the ef- ficiency of current fairness approaches through surveying example methods for reducing un- fair stereotypical bias from the literature, and (2) evaluating a common technique to reduce energy consumption (and thus environmental impact) of English NLP models, knowledge dis- tillation (KD), for its impact on fairness. In this case study, we evaluate the effect of important KD factors, including layer and dimensional- ity reduction, with respect to: (a) performance on the distillation task (natural language infer- ence and semantic similarity prediction), and (b) multiple measures and dimensions of stereo- typical bias (e.g., gender bias measured via the Word Embedding Association Test). Our results lead us to clarify current assumptions regarding the effect of KD on unfair bias: con- trary to other findings, we show that KD can actually decrease model fairness. 1 Introduction Fairness and environmental sustainability are criti- cal to the future of human society, and, thus, also reflected by the United Nations’ 17 Sustainable De- velopment Goals (e.g., Goal 5: Gender Equality, and Goal 13: Climate Action).1 Accordingly, both topics are currently also active research areas in natural language processing (NLP). On the one hand, several works have established that language representations are prone to encode 1https://sdgs.un.org/goals and amplify stereotypical social biases (e.g., Boluk- basi et al., 2016), and, consequently, are a source of representational harm (Barocas et al., 2017; Hovy and Spruit, 2016; Shah et al., 2020). To address this issue and provide fairer language technologies, various approaches have developed methods for measuring bias (e.g., Caliskan et al., 2017; Nadeem et al., 2020; Nangia et al., 2020; Nozza et al., 2021, inter alia) as well as debiasing methods (e.g., Zhao et al., 2018; Dev and Phillips, 2019, inter alia). On the other hand, recent advances in NLP have been fueled largely by increasingly compu- tationally expensive pre-trained language models (PLMs). Whereas the original BERT base model has 110M parameters (Devlin et al., 2019), the Switch Transformer model, designed as a more ef- ficient alternative to more recent PLMs, has over a trillion parameters (Fedus et al., 2022). While …

Content truncated to 500 words