Automatic Methods for Dialogue Classification and Prediction

The problem presented in this thesis is to specify and implement a solution which could be used to automatically classify previously completed calls to a call routing service, and furthermore, to specify and implement a solution which could be used to automatically predict the outcome of calls to this service “on-line”, i.e. as they happen. The specific service on which to investigate this problem is a residential customer care service designed for a triple-play provider, providing TV, telephone, and broadband solutions to end consumers. The entrance to this service is a call routing system, to which end consumers can call and be routed to the appropriate customer service (for example, technical support for broadband, or the customer service for invoice issues). As the tripleplay provider has large-scale operations, and offers a wide variety of products, their need for an easy-to-use, dynamic, maintainable and inexpensive call routing system is obvious. The method used to classify the calls to the service yields a classifier with an accuracy of 86%, which is a vast improvement on the majorityguess baseline of 27%. The method used to predict dialogues in calls to the service yields a predictor with an accuracy of 96% after three dialogue turns, which is a relative improvement on the majority-guess baseline by over 35%. The method is also used to successfully predict transaction success after three dialogue turns, with an accuracy of 82%, which is a relative improvement on the majority-guess baseline by over 5%. Although the work presented in this thesis was carried out on data collected from one specific service, it is the author’s claim that the results and conclusions found in this thesis are largely applicable to any similar voice controlled call routing service. Automatiska metoder för dialogklassificering och -prognos