Measuring Personality Adaptation in Task-Oriented Dialogs

Decades of research on human communication provides substantial evidence that human conversants in dialogue adjust their language to their conversational partner: conversants adapt to particular ways of referring to things (Brennan and Clark, 1996), mimic the partner’s lexical diversity (Coupland et al., 1988) and entrain to their vocabulary choices (Parent and Eskenazi, 2010). For example, for the question “Do I make a left at Pacific Ave?”, an adaptive response could be “Yes, make a left at Pacific Avenue.”, while a non-adaptive response might change the verb and referring expression: “Yes, turn left at Pacific.” Conversants also adapt to each other’s personality through gestures (Tolins et al., 2013): in a conversation involving two extraverted speakers, both became more stylistically extraverted over time by increasing gesture expansion and broadness. However, little work has addressed linguistic adaptation in terms of personality. For example, will two extraverted speakers use more words that exhibit extraversion as their dialog progresses (e.g. words of positive emotions: “happy”)? In this paper, we propose a method to calculate adaptation score for a dialog. We apply our method on two task oriented dialog corpora and show that as the dialogs progress, the adaptation score for extraversion related LIWC features (1) displays a distinct trend compared to all LIWC features, and (2) differs between two corpora.