The Development Stages of a Dialogue System - Man-Machine Dialogue: Design and Challenges - page 57

hand gestures and in general the video showing the user mid-utterance. The
Amitiés system, for example, has involved the annotation at various levels of
a corpus consisting of not only the transcription of the speech, but also the
identification of the speakers, the marking of the zones in which speech is
superimposed as well as semantic, dialogic, thematic and emotion annotations
[HAR 06]. The efforts made are such that, as for the SNCF corpus used by
[LUZ 95], various MMD systems can use them.
3.2.4. Specifying the processing processes
The data almost determine the processes that have to be implemented in
order to process themselves. The more the chosen phenomena are diversified,
the more the processes will have to be deepened. As an increasing number of
ambiguities appear, so the processes will increasingly rely on fine linguistic,
multimodal and dialogic analyses as well as an appropriate management of
context. This section illustrates the methodological difficulties that appear at
that point, by taking as an example a designer who we imagine facing the
system he/she is designing alone, a bit like in section 3.1.1. Our designer
determines the system's architecture, develops or reuses each of the identified
modules, each of the necessary resources, faces the difficulties he/she comes
across and carries out a certain number of simplifications. And actually, it is
by implementation that difficulties appear and initial ambitions diminish
slightly.
Let us consider, for example, the initial phenomena including anaphora,
essentially pronominal anaphora but also a few examples of associative
anaphora. The reference resolution module, which brings together the
resolution of direct references, demonstrative references (and thus the
multimodal fusion of pointing gestures with linguistic referring expression),
person deictics and anaphora, must thus integrate a pronominal and
associative anaphora resolution system. The designer thus first turns toward
existing anaphora resolution systems [MIT 02]. If we take into account the
implementation context (specific issues for the French language, a considered
lexicon, a panel of phenomena, input and output formats), he/she quickly
realizes that adapting, setting and integrating an existing resolution system
into his/her architecture is a delicate matter. He/she then decides to implement
his/her own anaphora resolution algorithm directly into the module with the
relevant settings. However this task is only one of the aspects of one of the