Prediction of Forthcoming Anger of Customer in Call Center Dialogs

Call center is a department that is most relevant to audio data usage. One of its major tasks is to monitor customers' anguish because it has a negative impact on the organization. One challenging task is to develop a model that can predict whether a customer is getting angry in the next turn of conversation. Such model can assist agents in taking appropriate action(s) to prevent the incidents. In this study, we investigate an approach to build an anger prediction model from customers' voice in call center dialogs. To create the model requires 5 processes: (1) Customer's turn extraction (2) Emotion annotation (3) Voice feature selection (4) Data pre-processing for long short-term memory networks, and (5) Anger prediction modeling. Five long short-term memory networks were built with the time series data sets of 1, 2, 3, 4, and 5 consecutive turns. The experimental results showed that the long short-term memory network built with the 3-consecutive turn data has promising performance in aspect of Average Precision and False Negative Rate when compared to the random and good guess benchmarks.