Mitigating Gender Bias in Word Embeddings using Explicit Gender Free Corpus

Words embeddings are the fundamental input to a wide and varied range of NLP applications. It has been shown that these embeddings reflect biases, such as gender and race, present in society and reflected in the text corpora from which they are generated, and that these biases propagate downstream to end use applications. Previous approaches to remove these biases have been shown to significantly reduce the direct bias, a measure of bias based on gender explicit words, but it was subsequently demonstrated that the structure of the embedding space largely retains indirect bias as evidenced by the spatial separation of words that should be gender neutral but are socially stereotyped on gender. This paper proposes a new method to debias word embeddings that replaces words in the training corpus that have explicit gender with gender neutral tokens, and creates the embeddings for these replaced words from the embedding of the gender neutral token post training utilising an added gender dimension. By design this method is able to fully mitigate direct bias and experiments demonstrate this. Experiments are also performed to investigate the effect on indirect bias, but generally are unable to achieve the reductions obtained by previous methods.