Dialog Breakdown Detection using Dialog Model based on Quasi-Recurrent Neural Networks

Abstract

In recent years, studies on chat-oriented dialog systems have been actively conducted due to the spread of dialog agents. On the other hand, many chat-oriented dialog systems have frequent dialog breakdown in which dialog is not smoothly performed. To tackle this problem, we propose a method to perform fast learning and robust dialog breakdown detection using Dialog Model based on Quasi-Recurrent Neural Networks (QRNN). To clarify the effectiveness, we conducted comparison experiment with other Recurrent Neural Networks (RNN) models, and show that QRNN has a faster learning and more accurate dialog breakdown detection than RNN.