Discovering Differences in the Representation of People using Contextualized Semantic Axes