4:45 Thursday April 6, 2017
119 Math Physics
TITLE: The geometry of gender stereotype in word embeddings
Abstract: Machine learning has many powerful applications, but the blind deployment of machine learning runs the risk of amplifying biases present in data. In this talk, I’ll illustrate this challenge with word embeddings, a popular framework to represent English words as vectors which has been used in many AI systems. I’ll show how gender stereotypes are intrinsically captured by the geometry of the word vectors with disturbing implications. We developed an algorithm to modify the embedding geometry to reduce gender stereotypes while preserving the useful features of the data. The resulting debiased embeddings can be used in applications without amplifying gender bias.