Visual Exploration of Indirect Bias in Language Models

Abstract

Language models are trained on large text corpora that often include stereotypes. This can lead to direct or indirect bias in downstream applications. In this work, we present a method for interactive visual exploration of indirect multiclass bias learned by contextual word embeddings. We introduce a new indirect bias quantification score and present two interactive visualizations to explore interactions between multiple non-sensitive concepts (such as sports, occupations, and beverages) and sensitive attributes (such as gender or year of birth) based on this score.