Research And Implementation Of POMDP-based Spoken Dialogue Systems

Continuous advances in the development of information technologies have made it possible to access information and web services from nearly anywhere,at anytime and almost instantaneously through Internet.It is necessary to provide an effective,easy and transparent interaction between users and computers.With this objective,as an attempt to enhance and ease human-to-computer interaction,in the last years there has been an increasing interest in simulating human-to-human communication,employing the so-called Spoken Dialogue Systems(SDSs).Dialogue management(DM)is the core of SDSs,which has been extensively researched.In which,the Partially Observable Markov Decision Process(POMDP)is now considered as one of the most important models.Two problems of POMDP-based SDSs are explored in this paper.POMDP-based DMs have the limitation that the number of slots and the value of slot should be finite.To solve this problem,we propose dynamic object binding method which can be applied to tasks whose fixed-size slots have infinite possible values.In addition,a method to estimate observation probability of POMDP by making use of entire N-Best list of Automatic Speech Recognition(ASR)is presented for addressing the uncertainty of ASR.The improved POMDP-based DM model with its algorithm based on two improvements mentioned above is presented in this paper.The improved model has been implemented on a teach-and-learn spoken dialogue system.Results of experiments indicate that our method can effectively overcome the infinite teaching objects problem in the teach-and-learn task.Moreover,estimating observation based on N-Best list of ASR can improve system performance.