Information Technology Reference
In-Depth Information
more or less overlap one another: first, dialogue control which tries to manage
the interactive process to determine a type of reaction after a structured chain
of utterances, and second, the modeling of the dialogue context, which focuses
on the dialogue history and the way in which the utterance content is based,
andfinallythedialogueinitiativewhichaddsaspecificbehaviortotheprevious
considerations [MCT 04, JOK 10].
Dialogue control is the focus of this section and has been the focus of a
great many studies [PIE 87, SAB 89, CAR 90, LUZ 95, TRA 03]. If we
exaggerate slightly, firstly there are the methods of finite states that have been
implemented: they are finite state or dialogue grammar automata, and the
point is to determine all the possible situations and ways of going from one
situation to the next. This type of approach works very well for dialogues in
which the initiative is always coming from one side, for example the system's
side, because in that case they are the system's utterances (questions in very
directing systems, different speech acts for more flexible systems) which are
taken into account by the states, the user's answers being taken into account
by the transitions [JUR 09, p. 863]. Then come the (fill in the blanks) pattern
methods, which allow us to recognize situations without forcing the initiative
to remain on one side or the other.
Then come the methods based on an information state, that is the methods
which add a memory, which contains anything required, knowing that this
memory will help determine the possible follow-ups of the dialogue.
Depending on the authors, the information state contains the dialogue history,
common ground, a model with mental states, a model of the user, etc. The
important bit is to use and update the data which will behave a bit like the
global variables in the dialogue manager. Thus, the planning approach is very
successful in initial studies by Cohen and Perrault [COH 79] and Allen and
Perrault [ALL 80]: the goal is to recognize and plan the plans, speech acts
being planned just like actions (the speaker's speech acts are part of a plan
that the hearer must unveil to answer in a relevant manner), based on a
modeling of the speaker and hearer's mental states. To reconcile natural
dialogue and task satisfaction, we can distinguish two types of plan managed
in parallel: the discourse plans and the domain plans. In general, the plans
allow for a great number of possibilities in terms of dialogue control; see the
article by N. Maudet in Gardent and Pierrel [GAR 02].
Then come the joint action theories (see section 8.2.2 with the notion of
common ground) and dialogue control inspired by game theory, each
Search WWH ::




Custom Search