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.
