Detecting problematic dialogs with automated agents

Abstract

We present a supervised machine learning approach for detecting problematic human-computer dialogs between callers and an automated agent in a call center. The proposed model can distinguish problematic from non-problematic calls after only five caller turns with an accuracy of over 90%. Based on a corpus of more than 69,000 dialogs we further employ the classifier's decision to given business models and present the cost savings that can be achieved by deploying classification techniques to Interactive Voice Response systems. © 2008 Springer-Verlag Berlin Heidelberg.