Author: Tolga Bolukbasi, Kai-Wei Chang, James Zou, et al.
Publisher: Arxiv
Publication Year: 2016
Summary: The following article discusses how machine learning applied blindly runs the risk of amplifying biases in data. Word embedding, a popular framework for representing text data as vectors that has been used in many machine learning and natural language processing tasks, poses such a risk. The authors demonstrate that even word embeddings trained on Google News articles exhibit disturbing female and male gender stereotypes. The authors present a method for modifying an embedding to remove gender stereotypes, such as the association between the words receptionist and female, while retaining desired associations, such as the association between the words queen and female. The authors develop metrics for quantifying both direct and indirect gender biases in embeddings, as well as algorithms to “debias” the embedding.