Abstract
In the present thesis, entitled “Dialog systems based on stochastic models”, the state of the art in the field of the dialog systems is exposed and the work done in the design and implementation of the modules of a specific dialog system is presented.
The thesis is focused on the study of the dialog management following a statistical approach. The development of a complete dialog system (with textual input and output, in Spanish language, and for a task in a restricted semantic domain, the task defined in the BASURDE research project) is the main contribution of the thesis. Such a system is composed by the natural language understanding module, the dialog manager module and the natural language reply generator module. Given the central objective of the thesis, the development of the dialog manager module has been the main work and, in consequence, it is exposed with the maximum amplitude in the present thesis.
The limited size of the dialog corpus of the BASURDE task has supposed a severe difficulty in the development of a dialog manager based exclusively on stochastic models. The dialog manager module that is been finally im-plemented determines its dialog strategy by means of a combination of sev-eral knowledge sources: some of them are of stochastic type, as the models learnt from the corpus; others are of heuristic type, as the rules that incorpo-rate pragmatic and semantic knowledge, either generic or specific to the task.
Finally, the user simulation has been considered as an alternative technique to deal with the evaluation of the behaviour of the dialog system, the exten-sion of the corpus by means of synthetic dialogs, and the dynamic learning of the stochastic dialog models. The corresponding user simulator modules has been designed and implemented, studying the possibilities of this tech-nique.