The Development Stages of a Dialogue System - Man-Machine Dialogue: Design and Challenges - page 53

is reduced, and there is no particular digression. The users do not argue their
decisions, and do not have to defend them. They do not make any comment to
explainwhattheyaresaying.Asforreferences, theydonotmakeanyreference
to themselves or to the machine. In other words, the dialogue remains natural
but is focused on itself, which means the task is satisfied. In such conditions,
we can consider that it is indeed a natural dialogue in natural language. Since
the user is talking (or thinks he/she is talking) to a system, he/she limits the
breadth of his/her utterances, but not at the cost of spontaneity, since he/she
does it naturally. If we took this to the extreme, this mechanism could lead
to a caricature operating mode, observed, for example, during a Wizard of
Oz described in [LAN 04], in which the user limits himself/herself to two or
three sentences: those that worked at the very beginning of the dialogue, and
during which the user gained confidence. This is the type of behavior that can
be found with a genuine MMD system, and that leads to the question, again, of
the natural aspect: if the user limits himself/herself so greatly, then talking to a
machine must be disturbing him/her. He/she is not in a natural mode. Actually,
as the experiments carried out in the Miamm project show [LAN 04], some
users limit themselves without being prompted, and others place themselves
automatically in a natural dialogue in natural language without any suspicion
against the system. However, and this is one of the essential challenges of
the designers, the system has to hold a conversation: if, on the contrary, it
starts showing signs of not understanding, the most confident user will change
his/her behavior very quickly.
3.2.2. Specifying covered phenomena
There is often a loop made up of various stages when imagining the
understanding and dialogue abilities of a system: reflecting on the system's
expected behavior, which means a specification of the interaction's nature;
specification of the breadth of the system's abilities; a simulation of the future
system and thus the constitution of a dialogue corpus; a corpus study, that is
an analysis of the expected and new phenomena; reflections on how to take
these new phenomena into account, and thus a return to the first stage. Once
the loop is stable, we then move on to the design of a processing model,
which takes a number of processes into account, and this model is then
implemented and tested to find its weak points. When the implementation has
been carried out, we can then come back to the experiment stage and start the
loop again, by using a real system instead of a simulation.