Conclusion - Man-Machine Dialogue: Design and Challenges - page 186

interface between linguistic theories and computational implementations.
This intermediary position has the advantage of helping to identify which
linguistic works can have relevant applications and contribute to MMD with
theoretic abilities and an overall view that is sometimes desirable. But it is
also very uncomfortable: the researcher at the border between two disciplines
does not contribute to linguistic theories (only to their applications) and does
not create computational development (only prerequisites). Thus, there is no
direct result that can be promoted, and his or her contribution, to a formal
model or to paths for an implementation, can be easily criticized: only an
implementation comes with proof and can resist criticism. With this topic, we
hope to have helped to show how these border works remain indispensable.
The second set of challenges focuses on the panel of abilities expected in a
system. We have shown that widespread understanding abilities were the key
to relevant exchanges and a realistic dialogues. Moreover, as Cole [COL 98,
p. 200] underlines it, for closed-domain system, it is first and foremost
robustness and real-time aspects that have to be improved, as well as the
system's abilities to lead the user (without railroading him) into an
operational path.
The third set of challenges covers the methodological and technical
challenges around system design. We have given examples of complex work
flows, involving the implementation not only of run-time architectures, but
also, something that is not as common, of design-time architectures that are
indispensable in allowing a certain flexibility in the development as well as an
amount of reusability. Moreover, we have shown that the amount of work
required to design an MMD system goes beyond that of a doctoral
dissertation and, thus, a team is now necessary. The elements of this team are
different depending on their professions, just like what is being done in other
computer science fields.
Finally, the last set of challenges is that of facilitating computational
development with toolkits, and maybe some day middleware and hardware
cards for NLP, automatic understanding and dialogue management. Today,
creating an MMD system that can keep up the state of the art in most of its
functionalities and add an innovative aspect is a genuine challenge, unless we
have a work environment that provides a continuously updated platform. The
generalization of this type of platform and the means of easiness we have
mentioned will be a major advance for the field of MMD. It will enable us not
only to significantly increase the research result speed, but also to carry out
more trustworthy assessments that will be more comparable than they are
now.