A dimensional model of interaction style variation in spoken dialog

Abstract

In spoken dialog people vary their interaction styles, and dialog systems should be able to do the same. Previous work has elucidated various aspects of style variation and adaptation, but a general model has been lacking. Here we present a dimensional model of the space of interaction styles, derived from a large data set and prosody-based features. The 8 dimensions of this model cover many previously-noted aspects of style and include some novel ones. This model may be useful for selecting data for dialog model pretraining and fine-tuning, for investigating demographic differences, and for dialog system style adaptation. However, regarding individual differences in interaction style, we find individual style tendencies to be surprisingly weak, with a predictive model based on individual tendencies outperforming a speaker-independent model by only 3.6%.