Online Learning of Attributed Bi-Automata for Dialogue Management in Spoken Dialogue Systems

Online learning of dialogue managers is a desirable but often costly property to obtain. Probabilistic Finite State Bi-Automata (PFSBA) have shown to provide a flexible and adaptive framework to achieve this goal. In this paper, an Attributed PFSBA (A-PSFBA) is implemented and experimentally compared with previous non-attributed PFSBA proposals. Then, a simple yet effective online learning algorithm that adapts the probabilistic structure of the Bi-Automata on the run is presented and evaluated. To this end, the User Model is also represented by an A-PFSBA and the impact of different user behaviors is tested. The proposed approaches are evaluated on the Let’s Go corpus, showing significant improvements on the dialogue success rates reported in previous works.