Sociolectal Analysis of Pretrained Language Models

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4581–4588 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 4581 Sociolectal Analysis of Pretrained Language Models Sheng Zhang1,2, Xin Zhang1, Weiming Zhang1, Anders Søgaard2 1National University of Defense Technology 2University of Copenhagen zhangsheng@nudt.edu.cn, shinezhang_nudt@qq.com, wmzhang@nudt.edu.cn, soegaard@di.ku.dk Abstract Using data from English cloze tests, in which subjects also self-reported their gender, age, education, and race, we examine performance differences of pretrained language models across demographic groups, defined by these (protected) attributes. We demonstrate wide performance gaps across demographic groups and show that pretrained language models sys- tematically disfavor young non-white male speakers; i.e., not only do pretrained language models learn social biases (stereotypical as- sociations) – pretrained language models also learn sociolectal biases, learning to speak more like some than like others. We show, how- ever, that, with the exception of BERT mod- els, larger pretrained language models reduce some the performance gaps between majority and minority groups. 1 Introduction While speakers of English generally understand each other, our linguistic preferences may differ, depending on the linguistic varieties we were ex- posed to, and how susceptible we were to them at the time (Tagliamonte and D’Arcy, 2009). Linguis- tic varieties form a multi-dimensional continuum of dialects and sociolects (McCormack et al., 2011).1 Such linguistic variation presents a real chal- lenge: The linguistic preferences of English pre- trained language models may align better with the linguistic preferences of some groups in society than with those of others. Group disparities consti- tute a fairness problem (Sokol et al., 2020): If our technologies provide end users with new opportuni- ties, group disparities mean unequal opportunities across groups. Moreover, if the groups are defined in terms of protected attributes, our technologies may discriminate between end users in ways that violate regulations. 1Dialects are mainly defined in terms of geography; soci- olects in terms of demographics (Trudgill, 2003). After waiting three hours, Cal whined and started to ___. Human cry (0.50) complain (0.11) leave (0.08) squirm (0.05) pout (0.05) fidget (0.04) yell (0.04) pace (0.03) argue (0.02) Machine run (0.12) bark (0.08) pace (0.07) cry (0.07) laugh (0.04) growl (0.04) eat (0.03) move (0.03) rise (0.03) Table 1: An example of a cloze (fill-in-the-gap) task with human and model predictions. We evaluate the sociolectal biases of a range of pre-trained English language models. Unlike pre- vious work on biases in pre-trained language mod- els, we do not consider representational biases (Sun et al., 2019), but performance disparities; moreover, we do not consider downstream performance differ- ences after fine-tuning for downstream tasks such as coreference resolution (Rudinger et al., 2018) or machine translation (Stanovsky et al., 2019), but performance differences across demographics of the language models themselves on cloze (fill-in- the-gap) problems. Since the cloze task is how these pre-trained language models are trained, we can evaluate models directly without introducing biases from probes or downstream tasks. Note that some sociolinguistic variables are salient (e.g., this→dis), others are not (Jaeger and Weatherholtz, 2016). One …

Content truncated to 500 words