Affective analysis of customer service calls

Abstract

This paper presents an affective and acoustic-prosodic analysis of a call-center corpus (700 phone calls with corresponding customer satisfaction levels). Our main goal is to understand how customers’ satisfaction correlates to the acoustic-prosodic and affective information (emotions and personality traits) of the interactions. A subset of 30 calls was manually annotated with emotions (frustrated vs.neutral) and personality traits (Big-Five model). Results on automatic satisfaction prediction from acoustic-prosodic features show a number of very informative linguistic knowledge-based features, especially pitch and energy ranges. The affective analysis also provides encouraging results, relating low/high satisfaction levels with the presence/absence of customer frustration. Concerning personality, customers tend to express signs of anxiety and nervousness, while agents are generally perceived as extroverted and open.