The Development Stages of a Dialogue System - Man-Machine Dialogue: Design and Challenges - page 52

always get the train line before looking for the train time”. The dialogue's
progression is thus strongly limited [KOL 10]. Moreover, studies of the SNCF
corpus show that the vocabulary is relatively limited, as are the sentence
structures, which means that the task's predominance influences the natural
language.
The main question is thus as follows: when there is a task, does the task's
resolution have precedence over the dialogue's spontaneity? Either we
consider that the dialogue is first finalized, and the answer is yes, even if there
is a lack of consistency (efficiency above all); or we consider that the dialogue
primarily aims to maintain enjoyable communication with the user, and
accept that it might take three or four more speech turns to arrive at the same
result. Both choices are acceptable, but must not be assessed in the same way:
if the speed in satisfying the task drives the assessment, the finalized system
will obviously arrive first.
This question takes on a particular dimension when we question the notion
of spontaneity. There is not necessarily a gap between the task's resolution
and natural dialogue in natural language. The natural aspect of the dialogue
is not subsequently judged by analyzing the lexical breadth and the linguistic
and pragmatic phenomena diversity, but is judged on the one hand by the user
experience, with the answers given during an interview on the ease of dialogue
with the system, and the level of satisfaction given by the system's utterances,
and on the other hand, by the subsequent analyses checking that the system's
reactions are relevant to the user's utterances. A user can be satisfied by the
system's dialogue even if the task took longer to complete than planned. As
Sperber and Wilson [SPE 95] and later Reboul and Moeschler [REB 98] show,
relevance is at the heart of the natural dialogue in natural language.
The question between solving the task and natural dialogue is also linked to
the machine's interlocutor role. As we saw at the end of section 2.2.1, talking
to a machine is not the same thing as talking to your best friend. Unless he/she
is misled and believes, as it happens over the phone, that the interlocutor is
human, the user knows he/she is talking to a machine, which can lead to a
specific type of behavior on his/her part. The Wizard of Oz experiments and
user tests at the end of the design are clear on this point. After a Wizard of
Oz test on a train ticket reservation task, Luzzati [LUZ 95] shows that the
dialogues focus on the key points with simple structures: each initiating act
is matched by a reacting act, the lexicon is essentially that of the task, that