Mitigating Gender Bias in Natural Language Processing: Literature Review

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1630–1640 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1630 Mitigating Gender Bias in Natural Language Processing: Literature Review Tony Sun*†, Andrew Gaut*†, Shirlyn Tang†, Yuxin Huang†, Mai ElSherief†, Jieyu Zhao‡, Diba Mirza†, Elizabeth Belding†, Kai-Wei Chang‡, and William Yang Wang† †Department of Computer Science, UC Santa Barbara ‡Department of Computer Science, UC Los Angeles {tonysun, ajg, shirlyntang, yuxinhuang}@ucsb.edu {mayelsherif, dimirza, ebelding, william}@cs.ucsb.edu {jyzhao, kwchang}@cs.ucla.edu Abstract As Natural Language Processing (NLP) and Machine Learning (ML) tools rise in popu- larity, it becomes increasingly vital to recog- nize the role they play in shaping societal bi- ases and stereotypes. Although NLP models have shown success in modeling various ap- plications, they propagate and may even am- plify gender bias found in text corpora. While the study of bias in artificial intelligence is not new, methods to mitigate gender bias in NLP are relatively nascent. In this paper, we review contemporary studies on recognizing and mit- igating gender bias in NLP. We discuss gender bias based on four forms of representation bias and analyze methods recognizing gender bias. Furthermore, we discuss the advantages and drawbacks of existing gender debiasing meth- ods. Finally, we discuss future studies for rec- ognizing and mitigating gender bias in NLP. 1 Introduction Gender bias is the preference or prejudice to- ward one gender over the other (Moss-Racusin et al., 2012). Gender bias is exhibited in multi- ple parts of a Natural Language Processing (NLP) system, including the training data, resources, pre- trained models (e.g. word embeddings), and al- gorithms themselves (Zhao et al., 2018a; Boluk- basi et al., 2016; Caliskan et al., 2017; Garg et al., 2018). NLP systems containing bias in any of these parts can produce gender biased predictions and sometimes even amplify biases present in the training sets (Zhao et al., 2017). The propagation of gender bias in NLP algo- rithms poses the danger of reinforcing damaging * Equal Contribution. Figure 1: Observation and evaluation of gender bias in NLP. Bias observation occurs in both the training sets and the test sets specifically for evaluating the gender bias of a given algorithm’s predictions. Debiasing gen- der occurs in both the training set and within the algo- rithm itself. stereotypes in downstream applications. This has real-world consequences; for example, concerns have been raised about automatic resume filter- ing systems giving preference to male applicants when the only distinguishing factor is the appli- cants’ gender. One way to categorize bias is in terms of allo- cation and representation bias (Crawford, 2017). Allocation bias can be framed as an economic is- sue in which a system unfairly allocates resources to certain groups over others, while representa- tion bias occurs when systems detract from the so- cial identity and representation of certain groups (Crawford, 2017). In terms of NLP applications, allocation bias is reflected when models often per- form better on data associated with majority gen- der, and representation bias …

Content truncated to 500 words