Online Learning of Stochastic Bi-automaton to Model Dialogues

User-aware dialogue management policies over attributed bi-automata

“…Using this metric, the A-PFSBA model learns from those dialogues rendered successful by the QM, augmenting the initial model by learning the new states and transitions of the new dialogues. This approach overcomes the drawbacks of previous turn-by-turn learning algorithms [13], that learned from both correct and incorrect dialogues. Figure 2 shows the previous scenario where an unseen dialogue z is rendered valid by a given QM, so the initial A-PFSBA model of the DM is augmented with the A-PFSBA model corresponding to z .…”

“…Previous experiments in [13,19] employed local decisional strategies over the bi-automata structure (i.e. taking into account only the current state q s ).…”

“…An initial exploration of the potential and flexibility of A-PFSBA was done in [19], where the inclusion of attributes in the PFSBA structure showed to improve performance. In addition, an online learning method based on successfully completed dialogues demonstrated the capability of learning from user interactions, overcoming the limitations of previously explored turn-byturn online learning procedures [13]. Although promising results were obtained, a local Maximum Likelihood exploitation policy which did not take user behavior into account was used and path based exploitation policies were not explored.…”

“…Experiments are carried out on the Let's Go corpus [14], allowing direct comparison with previous work by [13,19]. The paper is structured as follows: Section 2 explains spoken dialogue interaction as an stochastic process and describes the A-PFSBA formulation to model these interactions.…”