Automatic call section segmentation for contact-center calls

Abstract

This paper presents a SVM (Support Vector Machine) classification system which divides contact-center call transcripts into "Greeting", "Question", "Refine", "Research", "Resolution", "Closing" and "Out-of-topic" sections. This call section segmentation is useful to improve search and retrieval functions and to provide more detailed statistics on calls. We use an off-the-shelf automatic speech recognition (ASR) system to generate call transcripts from recorded calls between customers and service representatives. We first classify an individual utterance into a call section by applying the SVM classifier and then merge adjacent utterances classified into a same call section. We experiment with the proposed system on 100 automatically transcribed calls. The 10-fold cross validation shows 87.2% classification accuracy. we also compare the proposed algorithm with two other approaches - the most frequent section only method and a maximum entropy-based segmentation. The evaluation shows that our system's accuracy is 12% higher than the first baseline system and 6% higher than the second baseline system respectively. Copyright 2007 ACM.