A data-driven approach to spoken dialog segmentation

Abstract

With the advances in Language Technologies and Natural Language Processing, conversational interfaces have begun to play an increasingly important role in the design of human-machine interaction systems in a number of devices and intelligent environments. In this paper, we present a statistical model for spoken dialog segmentation and labeling based on a generative model learned using decision trees. We have applied our proposal in a practical conversational system that helps solving simple and routine software and hardware repairing problems. The results of the evaluation show that automatic segmentation of spoken dialogs is very effective for human-machine dialogs. The same statistical model has been applied to human-human conversations and provides a good baseline as well insights in the model limitation.