Quietly angry, loudly happy Self-reported customer satisfaction vs. automatically detected emotion in contact center calls

Phone calls are an essential communication channel in today ' s contact centers, but they are more difficult to analyze than written or form-based interactions. To that end, companies have traditionally used surveys to gather feedback and gauge customer satisfaction. In this work, we study the relationship between self-reported customer satisfaction (CSAT) and automatic utterance-level indicators of emotion produced by affect recognition models, using a real dataset of contact center calls. We find (1) that positive valence is associated with higher CSAT scores, while the presence of anger is associated with lower CSAT scores; (2) that automatically detected affective events and CSAT response rate are linked, with calls containing anger/ positive valence exhibiting respectively a lower/higher response rate; ( 3) that the dynamics of detected emotions are linked with both CSAT scores and response rate, and that emotions detected at the end of the call have a greater weight in the relationship. These findings highlight a selection bias in self-reported CSAT leading respectively to an over/under-representation of positive/negative affect.