Automated Identification of Relative Social Status

Understanding the social context is a key requirement for any dialogue system that is able to interact in situations with multiple users, and also improves a dialogue agent’s ability to identify expectations of a single user. This paper focuses on methods for identifying one sub-aspect of the social context, the relative social power between interactors. Specifically, this paper examines how alternative candidate dialogue feature representations affect the ability of different classifiers to predict relative social power. The performance of the classifier and feature representation pairs was compared for the task of determining whether or not two speakers have a difference in social power, for two different transcript data sets. Different feature and classifier combinations performed best at detecting the presence, versus the absence, of a difference in social power.