Training on only perfect Standard English cor- pora predisposes pre-trained neural networks to discriminate against minorities from non- standard linguistic backgrounds. We perturb the inflectional morphology of words to craft plausible and semantically similar adversarial examples that expose these biases in popu- lar models, e.g., BERT and Transformer, and show that adversarially finetuning them for a single epoch significantly improves robustness without sacrificing performance on clean data.

@inproceedings{Tan-et-al-acl-20, abstract = {Training on only perfect Standard English cor- pora predisposes pre-trained neural networks to discriminate against minorities from non- standard linguistic backgrounds. We perturb the inflectional morphology of words to craft plausible and semantically similar adversarial examples that expose these biases in popu- lar models, e.g., BERT and Transformer, and show that adversarially finetuning them for a single epoch significantly improves robustness without sacrificing performance on clean data.}, address = {Seattle, USA}, author = {Samson Tan and Shafiq Joty and Min-Yen Kan and Richard Socher}, booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics}, numpages = {9}, pages = {2920–-2935}, publisher = {ACL}, series = {ACL'20}, title = {It’s Morphin’ Time! Combating Linguistic Discrimination with Inflectional Perturbations}, url = {https://www.aclweb.org/anthology/2020.acl-main.263/}, year = {2020} }