Stochastic Bi-Languages to model Dialogs
M. Inés Torres · 2013
Partially observable Markov decision Processes provide an excellent statistical framework to deal with spoken dialog systems that admits global optimization and deal with uncertainty of user goals. However its put in practice entails intractable problems that need efficient and suboptimal approaches. Alternatively some pattern recognition techniques have also been proposed. In this framework the joint probability distribution over some semantic language provided by the speech understanding system and the language of actions provided by the dialog manager need to be estimated. In this work we propose to model this joint probability distribution by stochastic regular bi-languages that have also been successfully proposed for machine translation purposes. To this end a Probabilistic Finite State Bi-Automaton is defined in the paper. As an extension to this model we also propose an attributed model that allows to deal with the task attribute values. Valued attributed are attached to the states in such a way that usual learning and smoothing techniques can be applied as shown in the paper. As far as we know it is the first approach based on stochastic bi-languages formally defined to deal with dialog tasks. 1