The main objective defined for this thesis is the study and
development of different methodologies to carry out dialog
management in spoken dialog systems. The main challenge set in the
thesis resides in the development of statistical methodologies for
dialog management, based on learning a model from a corpus of
labeled dialogs. In this field, the thesis presents different
approaches to managing the dialog, the improvement of the
statistical model and the evaluation of the dialog system.

For the practical implementation of these methodologies, in the
framework of a specific task, the acquisition and labeling of a
dialog corpus was necessary. This dialog corpus has facilitated
the learning and evaluation of the model developed for dialog
management. In addition, a complete dialog system has been
developed to evaluate the practical operation of the proposed
methodologies for dialog management in real use
conditions.

Different approaches are proposed to evaluate the techniques for
dialog management: the evaluation by means of real users; the
definition of training and test partitions using the acquired
corpus; and the evaluation by means of user simulation techniques.
The user simulator that has been developed allows the statistical
modelization of the complete process of the dialog. In this
approach, the proceses for obtaining the system answer and
generating the new user turn are modeled as a classification
problem, in which the input of the classifier is a set of
variables that represent the current state of the dialog. The
result of the classification can be seen as the probability of
selecting each one of the answers (sequence of dialog acts)
respectively defined for the user and the system.

The initial corpus has been extended and improved by means of the
dialogs generated using this simulation module. In addition,
different techniques for the automatic generation of dialogs are
described. These techniques facilitate the automatic acquisition
of a labeled dialog corpus and the later learning of a dialog
manager.

These works have been developed within the framework of the DIHANA
project, whose main objective was the development of a dialog
system to access an information system using spontaneous speech.
The task defined for the project was the oral access to a railway
information system.

Finally, the methodologies proposed for the dialog management in
the DIHANA project have been adapted to develop a dialog manager
for the EDECAN project. The adaptation and evaluation of a dialog
manager developed for a dialog system that facilitates the booking
of sport facilities is described.

Additionally, different rule-based methodologies for dialog
management are presented, as well as different approaches for the
development of natural language answer generators.