Morpheus exposes the potential allocative harms of popular pretrained NLP models by simulating inflectional variation. We propose adversarial fine-tuning for mitigating the effects of training only on error-free Standard English data.

Word embeddings inherit strong gender bias in data which can be further amplified by downstream models. We propose to purify word embeddings against corpus regularities such as word frequency prior to inferring and removing the gender subspace, which significantly improves the debiasing performance.