E-VAN : Enhanced Variational AutoEncoder Network for Mitigating Gender Bias in Static Word Embeddings

Recent research has shown that pre-trained context-independent word embeddings display biases such as racial bias, gender bias, etc. Using a novel, tunable algorithm, this study attempts to mitigate the hidden gender bias in static embeddings. In order to train the model, an enhanced variational autoencoder (E-VAN) is used to learn the latent space of the embedding. Then the latent distributions are used while adaptively resampling and re-weighting the rare/under-represented data. While the word embeddings retain semantic information, E-VAN effectively mitigates unwanted biased gendered associations. Our method E-VAN outperforms previous state-of-the-art methods in both quantitative and human evaluation.