Information Fusion in Automatic User Satisfaction Analysis in Call Center

In this paper, we propose a method to predict the user emotional state (anger or neutral) for improvement of user satisfaction in call center. In order to detect the user satisfaction more accurately, our work employs the following information fusion technologies: (1) in view of the data imbalance problem, we adopt statistical model fusion, (2) for improving classifier performance, we combine features with n-gram, sentiment word and domain-specific word, (3) according to the characteristics of spoken language, we use the combination of statistical model and language rule, (4) as to the multi dimension of emotion expression, we adopt the two perspectives of acoustic and linguistics to comprehensive evaluation. We have done a series of experiments on human-human dialogues which derived from China Mobile call center corpus, and our system gives each dialogue a user emotion result that is angry or neutral, corresponding to the user satisfaction or dissatisfaction. We regard the user's feedback as the standard answer. As a result, the fusion system, with 69.1% f1-meaurement, outperforms the signal statistical model (Support Vector Machines) with linguistics features which as baseline of approximately 65.4% f1-meaurement on test set.