We develop methods to quantify semantic shift between two comparable text collections as a way to identify the concepts with similar and different semantic neighborhoods. Our work leverages methods developed to expose and remediate bias in word embedding methods and methods to align different text embeddings. Past work shows that embeddings encode many inherent biases in these texts when tested against well-known protected categories such as gender and race. Recent work shows that many secondary biases may still exist even after a debiasing method is applied. Our aim is to expose such persistent biases when compared to a baseline dataset to help increase transparency of datasets and decisions made by future trustworthy AI tools. As bias is a crucial component of misinformation, we also apply our methods to enhance our existing tools for detection, explanation and mitigation of misinformation.
Posted October 18, 2019
Semantic shift as measure of bias with applications to detection, explanation and mitigation of misinformation
Team Members and Titles
Project Start Date
Project End Date