A Few Dialogue Strategies - Man-Machine Dialogue: Design and Challenges - page 134

utterance being a turn played during which the speaker tries to maximize
his/her gain [CAE 07]. Finally, machine learning techniques have appeared
more recently for dialogue control with Markov decision process (MDP) and
partially observable MDP (POMDP) model types, which extend that of the
information state by adding a probabilistic means to decide on the future
action depending on the current state [JUR 09, p. 883]. In the same vein,
Singh et al. [SIN 02] use reinforcement learning to achieve an optimal
decision set. The decision rules are refined by carrying out several thousand
of exchanges between the system and a simulated user.
As we can see, dialogue control can involve many kinds of technique. Let
us add that implementing them can involve other data in parallel, such as an
analysis of the topics approached, for example, so as to direct the system's
reaction toward one of the current topics rather than one of the abandoned
topics. This is the approach given by Vilnat [VIL 05], who has three distinct
cooperating pragmatic analysis submodules:
-athematic interpretation module managing the global coherence of the
topics approached during the dialogue;
-an intentional analysis module which provides functional dialogue
representation in which the roles of the various interventions are clarified; see
the presentation on the approaches based on intentional structure in the article
by N. Maudet in Gardent and Pierrel [GAR 02];
-an interaction management module which allows the system to react
to different types of incomprehension by allowing the dialogue to remain
efficient.
As for Rosset [ROS 08], she adds strategies depending on ergonomic
choices and suggests a dialogue model built around a set of phases:
acquisition (obtaining the information required to satisfy the task),
negotiation, navigation, post-acknowledgment (transition toward a
negotiation, navigation or end of dialogue) and metaprocessing (marking and
processing errors); see the Arise system [LAM 03].
8.2.2. Dialogue history modeling
Modeling the context of the dialogue can also lead to a great variety of
forms in theoretical models and implemented systems. The first aspect consists
of providing contextual elements to completely understand an utterance, after