An ensemble-based classification approach to model human-machine dialogs

Abstract

One of the most demanding tasks when developing dialog systems consists of designing the dialog manager, which decides the next system response considering the user’s actions and the dialog history. A previously developed statistical dialog management technique is adapted in this work to reduce the effort and time required to design the dialog manager. This technique allows not only an easy adaptation to new domains, but also to deal with the different subtasks by means of the fusion of classifiers adapted to each dialog objective in the application domain. The practical application of the proposed technique to develop a dialog system for a travel-planning domain shows that the use of these specific dialog models increases the quality and number of successful interactions with the system in comparison with developing a single dialog model for the complete domain.