It’s Morphin’ Time! Combating Linguistic Discrimination with Inflectional Perturbations

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2920–2935 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 2920 It’s Morphin’ Time! Combating Linguistic Discrimination with Inflectional Perturbations Samson Tan§♮, Shafiq Joty‡§, Min-Yen Kan♮, Richard Socher§ §Salesforce Research ♮National University of Singapore ‡Nanyang Technological University §{samson.tan,sjoty,rsocher}@salesforce.com ♮kanmy@comp.nus.edu.sg Abstract Training on only perfect Standard English cor- pora predisposes pre-trained neural networks to discriminate against minorities from non- standard linguistic backgrounds (e.g., African American Vernacular English, Colloquial Sin- gapore English, etc.). We perturb the inflec- tional morphology of words to craft plausible and semantically similar adversarial examples that expose these biases in popular NLP mod- els, e.g., BERT and Transformer, and show that adversarially fine-tuning them for a single epoch significantly improves robustness with- out sacrificing performance on clean data.1 1 Introduction In recent years, Natural Language Processing (NLP) systems have gotten increasingly better at learning complex patterns in language by pre- training large language models like BERT, GPT-2, and CTRL (Devlin et al., 2019; Radford et al., 2019; Keskar et al., 2019), and fine-tuning them on task- specific data to achieve state of the art results has become a norm. However, deep learning models are only as good as the data they are trained on. Existing work on societal bias in NLP primarily focuses on attributes like race and gender (Boluk- basi et al., 2016; May et al., 2019). In contrast, we investigate a uniquely NLP attribute that has been largely ignored: linguistic background. Current NLP models seem to be trained with the implicit assumption that everyone speaks fluent (often U.S.) Standard English, even though two- thirds (>700 million) of the English speakers in the world speak it as a second language (L2) (Eber- hard et al., 2019). Even among native speakers, a significant number speak a dialect like African American Vernacular English (AAVE) rather than Standard English (Crystal, 2003). In addition, these 1Code and adversarially fine-tuned models available at https://github.com/salesforce/morpheus. Figure 1: MORPHEUS looks at each noun, verb, or ad- jective in the sentence and selects the inflected form (marked in red) that maximizes the target model’s loss. To maximize semantic preservation, MORPHEUS only considers inflections belonging to the same universal part of speech as the original word. World Englishes exhibit variation at multiple levels of linguistic analysis (Kachru et al., 2009). Therefore, putting these models directly into pro- duction without addressing this inherent bias puts them at risk of committing linguistic discrimination by performing poorly for many speech communi- ties (e.g., AAVE and L2 speakers). This could take the form of either failing to understand these speakers (Rickford and King, 2016; Tatman, 2017), or misinterpreting them. For example, the recent mistranslation of a minority speaker’s social media post resulted in his wrongful arrest (Hern, 2017). Since L2 (and many L1 dialect) speakers of- ten exhibit variability in their production of inflec- tional morphology2 (Lardiere, 1998; Pr´evost and White, 2000; Haznedar, 2002; White, 2003; Sey- mour, 2004), we argue that NLP models should be …

Content truncated to 500 words