An Assessment of the Evolution of Research and Systems - Man-Machine Dialogue: Design and Challenges - page 24

linked to prosodic aspects, ergonomic aspects, etc. As we will see in Chapter
2, collaboration between fields is essential.
Finally, to end this list of theoretical challenges, let us underline the
importance of methodology, with which to carry out experimentations and to
create and use reference corpus for the MMD. This challenge is linked to
resources and is key not only for the study of dialogues covering a specific
task, but also to carry out machine learning algorithms or to generate data
such as lexicons, grammars and language models for the oral dialogue as well
as the multimodal dialogue. In this case, one of the challenges is in a better
integration of these resources. As an example, the Ozone project we have
mentioned allowed us to reflect on the concept of meta-grammar (or
meta-model), with the goal of instancing from a joint base of linguistic
grammar and statistical language model.
1.3.2. Diversifying systems' abilities
The technical challenges linked to the abilities of an MMD systems are the
NLP, AI, ECA, QAS and MMI tasks, and many more. In general, all the
components we have mentioned could be improved, with a greater scope of
phenomena taken into account and a greater finesse in their processing. Cole
[COL 98] highlights various linguistic aspects such as exploring the nature of
discourse segments and the discourse relations, as well as the need for
additional mechanisms to manage key phenomena such as the highlighting of
information in a linguistic message. All these challenges focus on the same
goal: increasing the coverage, the fluidity and the realistic aspect of the
dialogue. To make it more clear, the goal might be to achieve a natural
dialogue system or even a natural multilogue system in natural language
[KNO 08], which will be multimodal, multilingual, multitasking, multi-roled,
multi-thread, multi-user and, of course, capable of learning etc.
The question of realism is a great question, which starts with speed: a
system taking 10 s to answer does not have a chance of achieving realism. If
this criterion can be measured, however, there are some that cannot: how
should we measure the realism of a synthetic voice, of sentence construction,
of an ECA's gestures? The fact that some users reject an artificial voice is
sometimes based on tiny details that are hard to measure, such as a minute
defect in elocution rhythm. The perception of these minute defects can create