Improving end-to-end performance of call classification through data confusion reduction and model tolerance enhancement

Two major challenges in the rapid deployment of automated natural language call routing system are the minimization of the manual effort in tagging data and reducing the impact of speech recognition errors on the call classification. In this paper we explore some novel approaches which target at these two challenges. One of our approaches enriches the training set with additional speech recognition hypotheses, automatically splits the neighboring data with its original classes, retags the training data based on a similarity measure, and elects the final result among multiple classifiers which are trained from the split data; the other approach incorporates the acoustical confusable information into call classifier to reduce the impact of the speech recognition error on the call classification accuracy. The experimental results show that our new approaches can reduce the classification error rate of an automated natural language call routing system by relative 10% in the end to end performance, using the live data collected from an enterprise call center.