Nurse is Closer to Woman than Surgeon? Mitigating Gender-Biased Proximities in Word Embeddings

Word embeddings are the standard model for semantic and syntactic\nrepresentations of words. Unfortunately, these models have been shown to\nexhibit undesirable word associations resulting from gender, racial, and\nreligious biases. Existing post-processing methods for debiasing word\nembeddings are unable to mitigate gender bias hidden in the spatial arrangement\nof word vectors. In this paper, we propose RAN-Debias, a novel gender debiasing\nmethodology which not only eliminates the bias present in a word vector but\nalso alters the spatial distribution of its neighbouring vectors, achieving a\nbias-free setting while maintaining minimal semantic offset. We also propose a\nnew bias evaluation metric - Gender-based Illicit Proximity Estimate (GIPE),\nwhich measures the extent of undue proximity in word vectors resulting from the\npresence of gender-based predilections. Experiments based on a suite of\nevaluation metrics show that RAN-Debias significantly outperforms the\nstate-of-the-art in reducing proximity bias (GIPE) by at least 42.02%. It also\nreduces direct bias, adding minimal semantic disturbance, and achieves the best\nperformance in a downstream application task (coreference resolution).\n