A Study of Implicit Bias in Pretrained Language Models against People with Disabilities

Proceedings of the 29th International Conference on Computational Linguistics, pages 1324–1332 October 12–17, 2022. 1324 A Study of Implicit Language Model Bias Against People With Disabilities Pranav Narayanan Venkit Mukund Srinath College of Information Sciences and Technology Pennsylvania State University {pranav.venkit, mus824, shomir}@psu.edu Shomir Wilson Abstract Pretrained language models (PLMs) have been shown to exhibit sociodemographic biases, such as against gender and race, raising con- cerns of downstream biases in language tech- nologies. However, PLMs’ biases against peo- ple with disabilities (PWDs) have received lit- tle attention, in spite of their potential to cause similar harms. Using perturbation sensitivity analysis, we test an assortment of popular word embedding-based and transformer-based PLMs and show significant biases against PWDs in all of them. The results demonstrate how models trained on large corpora widely favor ableist language. 1 Introduction Recent work on language models show substantial evidence of the presence of sociodemographic bi- ases associated with race and gender (Tan and Celis, 2019; Bolukbasi et al., 2016; Caliskan et al., 2017). Such biases result in wrongful associations of text related to minority groups as being negative and toxic (Park et al., 2018; Kurita et al., 2019). How- ever, little prior work has focused on the identifica- tion and impact of disability bias (Hutchinson et al., 2020; Whittaker et al., 2019). According to a re- port on disability by the World Health Organisation (WHO), approximately one billion people, or 15% of the world’s population, experience some form of disability (Bickenbach, 2011). Research shows that people with disability (PWD) are the largest population group that faces discrimination regu- larly (Whittaker et al., 2019; Chen and McNamara, 2020). We find various forms of disability biases in AI systems as well, where language involving PWD can often be classified as toxic (Venkit and Wilson, 2021) or even violent (Hutchinson et al., 2020). Corpora used to train large language mod- els often only reflect the ‘loudest voices’ or the the most dominant viewpoints even if they are not rep- resentative of the population (Bender et al., 2021) thereby enabling harmful semantic biases (Caliskan et al., 2017). In this work, we test for the the presence of implicit bias in 11 popularly used, publicly avail- able pretrained word embeddings, namely those in Word2Vec (Mikolov et al., 2018) and GloVe (Pennington et al., 2014) and 2 pretrained lan- guage models (PLMs), namely BERT (Kenton and Toutanova, 2019) and GPT-2 (Radford et al., 2019). We analyse negative association created with words related to people with disability (PWD). Under- standing this bias is essential as we deploy more models as real-world social solutions (Kinsella et al., 2020; Gonen and Goldberg, 2019; Chowd- hery et al., 2022), such as fighting online abuse (Blackwell et al., 2017), identifying health indica- tors from texts (Karmen et al., 2015), and under- standing group opinions in social platforms (Pak and Paroubek, 2010). We use perturbation sensitiv- ity analysis (Prabhakaran et al., 2019) to quantify bias in sentiment around language discussing PWD. Our result shows that all models …

Content truncated to 500 words