Man–Machine Dialogue Fields - Man-Machine Dialogue: Design and Challenges - page 42

draws links between an act and the previous one (see section 1.1.2), and,
second, the conversational analysis path, which approaches dialogue as a
social interaction with organization phenomena such as speech turns, abrupt
topic changes and disfluence [SAC 74]. The discourse analysis path has led to
computational theories of speech acts, and the conversational analysis path
has helped improve speech turn management in MMD. For example, Sacks
et al. [SAC 74] show that speech turns are regulated, at least in American
language, by rules applied to specific moments, called transition-relevance
place (TRP), that is moments in which the language structure allows for
interruption. The criteria can be prosodic, such as the presence of a pause,
syntactic or semantic: we can consider that a sentence is potentially over
when the agents required for the predicate have been expressed, even if the
speaker wishes to add precisions, for example a deferred compliment, a
repetition or a rephrasing. This example is important in MMD because
provides parameters for the module in charge of detecting the end of the
user's interventions, a module that is involved in systems able to manage
real-time dialogues. In the same way, we can test the possibility of
interrupting the user. There are two arguments that might subdue this
initiative: first, even if the interlocutors theoretically have the opportunity of
speaking at the same time, Levinson [LEV 83] shows that all the
superimposition zones only reach 5% in the human dialogues he has studied;
second, MMD does not necessarily need to imitate human dialogue.
This last point was the focus of a more general debate on the management
of speech turns. We have suggested an approach aiming to allow the user to
express himself in his everyday language and to allow him to do so
spontaneously. This approach, tending toward natural dialogue in natural
language, relies more or less on the premise that MMD can copy human
dialogue. Actually, any MMD system has limits, and the user quickly realizes
that the machine has understanding limits. He adapts his behavior and
utterances, and thus contributes to a dialogue which takes a completely
different direction than it would have with a human interlocutor. Depending
on the MMD system's abilities, this adaptation can be moderate, and thus
similar to the adaptation that any interlocutor requires, or major, to the point
of considering that the machine is a very specific interlocutor. This is the
point of view of [JÖN 88], in an article with a very striking title: Talking to a
Computer is not like Talking to your Best Friend (section 3.2.1). This is also
the point of view of [FRA 91] who suggests the Wizard of Oz methodology to
turn the type of communication between a human being and a machine into a
subject of experiment (section 3.2.3).