The Development Stages of a Dialogue System - Man-Machine Dialogue: Design and Challenges - page 49

human means. After a joint specification phase, they divided the experiment,
development and assessment work. The first stage of experiments consists of
getting test subjects, who are not the designers, in front of a system simulation
as it is imagined, according to the Wizard of Oz principle. This step that ends
with the subjects being interviewed can be completed with human dialogue
corpus studies, to help define the essential aspect and behavior limits of the
system, and thus of the Wizard of Oz, for the given task.
The development starts by the joint definition of the system's architecture,
and the specification of the behavior of the modules constituting this
architecture: input types, output types and module communication language.
Each of these aspects draws strongly on the state of the art, by using, for
example, a communication language established in another project. By
following such constraints, we foster both the integration possibilities of the
existing modules and the reusability of the components created for this
specific project. The designers then start the development, i.e. building
resources and coding algorithms. To achieve this, the order followed is not
that of the processes carried out by the system; on the contrary, we start
with the system's core, dialogue management and pragmatic aspects. By
simulating the input, we then refine the system's behavior, and then specify
the understanding abilities required for this behavior. We can then implement
the modules that carry out linguistic analyses. The resources are partly taken
from other projects and partly from available, free or licensed, resources so as
to keep the maximum efforts for the algorithms, or rather for their settings,
taking the task into account.
The next step consists of testing and assessing the already implemented
embryonic system, by simulating the input. Then, just as the dialogue
management abilities have turned into settings for the design of
understanding modules (as well as generation modules, since it follows a
similar development), we move on to the materialization of the language
correctly processed into parameters for the voice recognition and low level
processing. An existing automatic recognition system is used, with much
better performances than the system used in the previous scenario, and
increased interaction possibilities. As for each phase of development, tests
and assessments are immediately carried out with just the recognition and
then all of the system, so as to refine the coverage of the resources and the
process's settings. The assessments mostly consist of measuring the recall and