Online learning of stochastic bi-automaton to model dialogues

Abstract

The Interactive Pattern Recognition framework has been proposed to deal with Spoken Dialog Systems. 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 have been modeled by stochastic regular bilanguages. In this work we propose an algorithm to estimate the parameters of the corresponding Probabilistic Finite State Bi-Automaton. Moreover an on line learning methodology aimed at updating the model parameters at each interaction step is also proposed. The experimental evaluation carried out with different simulated users over LetsGo task shows the learning and adaptation capabilities of the proposal.