Mitigating Gender Bias in Distilled Language Models via Counterfactual Role Reversal

Findings of the Association for Computational Linguistics: ACL 2022, pages 658 - 678 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Mitigating Gender Bias in Distilled Language Models via Counterfactual Role Reversal Umang Gupta∗1, Jwala Dhamala 2, Varun Kumar 2, Apurv Verma 2, Yada Pruksachatkun 2, Satyapriya Krishna† 4, Rahul Gupta 2, Kai-Wei Chang† 2 3, Greg Ver Steeg† 1 2, Aram Galstyan 2 1Information Sciences Institute, University of Southern California 2Amazon Alexa, 3University of California, Los Angeles, 4Harvard University umanggup@usc.edu, gupra@amazon.com Abstract Language models excel at generating coherent text, and model compression techniques such as knowledge distillation have enabled their use in resource-constrained settings. However, these models can be biased in multiple ways, including the unfounded association of male and female genders with gender-neutral profes- sions. Therefore, knowledge distillation with- out any fairness constraints may preserve or ex- aggerate the teacher model’s biases onto the distilled model. To this end, we present a novel approach to mitigate gender disparity in text generation by learning a fair model dur- ing knowledge distillation. We propose two modifications to the base knowledge distilla- tion based on counterfactual role reversal— modifying teacher probabilities and augment- ing the training set. We evaluate gender po- larity across professions in open-ended text generated from the resulting distilled and fine- tuned GPT–2 models and demonstrate a sub- stantial reduction in gender disparity with only a minor compromise in utility. Finally, we ob- serve that language models that reduce gender polarity in language generation do not improve embedding fairness or downstream classifica- tion fairness. 1 Introduction The ever-increasing size of language models (LMs) have increased their energy and compute require- ments, making them impractical for many real-time resource-constrained applications such as personal assistants deployed on edge devices. To address this issue, various approaches have been proposed to compress or distill these large models (e.g., Sanh et al. (2019); Jiao et al. (2020); Hinton et al. (2015)). However, distillation techniques are designed to mimic the uncompressed LM (i.e., teacher model). Thus, the societal biases encoded in the teacher *Part of this work was done as an intern at Amazon Alexa. †This paper describes work performed at Amazon. He works in a hospital as a Prompt . . . doctor, treating the elderly with a variety, and by all accounts does an excellent work of medicine. GPT–2 . . . physician and helps a lot of the patients. Fair DistilGPT–2 (ours) She works in a hospital as a Prompt . . . nurse and was in love with her mother and her big brother, a small, shy, overweight woman. GPT–2 . . . pediatric dermatologist who gets stitches but also helps hospitals understand newborns . . . Fair DistilGPT–2 (ours) Figure 1: Example texts generated by LMs under differ- ent gender contexts (identified by the words ‘He’ and ‘She’). GPT–2 continues the prompt with the occupation word histor- ically associated with the specific gender. Our approach aims to treat both genders equally. models (Bender et al., 2021; Bommasani et al., …

Content truncated to 500 words