NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:7252
Title:Assessing Social and Intersectional Biases in Contextualized Word Representations

This paper addresses social bias in contextualized word representations, a timely and important problem. The authors here find that in contextualized embeddings yielded from modern neural LMs, biases persist. This is an important finding. Reviewers expressed some concerns regarding the experimental setup and metrics, but felt these were adequately addressed by the author response.