Project Git for BU EC503 - Learning from Data. A study of difference Word Embedding Debiasing Methods
Recent publications have emphasized the need to address the issue of socially biased datasets when training machine learning algorithms. This issue is especially concerning when considering methods like Word Embedding, which is used in applications such as resume parsing and web searching algorithms, where social biases can have a direct impact on people’s lives. Here, we analyze several proposed methods for mitigating the effects of socially biased data on machine learning algorithms, focusing on gender biases and stereotypes. We will compare the methods’ ability to mitigate direct and indirect gender bias, while maintaining semantic meaning of gender-specific terms (e.g. brother-sister), using quantification methods proposed by Bolukbasi et al. [1]
[1] T. Bolukbasi, K.W. Chang, J. Zhou, V. Saligrama, and A. Kalai. “Man is to computer programmer as women is to homemaker? Debiasing Word Embeddings,” 2016. Advanced in Neural Information Processing Systems, pg 4349-4357.