Yuichiro ANZAI
Department of Computer Science, Keio University
3-14-1, Hiyoshi, Kouhoku-ku, Yokohama 223 JAPAN
e-mail: anzai@aa.cs.keio.ac.jp
A speech dialogue is advantageous over other ways of communication
between humans and computers in that it needs no special training.
Much research has been done for speech dialogue systems, however,
those systems are generally not able to work in dynamic environments.
It is principally because they lack consideration of non-verbal
information, and also relation between verbal and non-verbal
information. Actually, human conversation contains, or is affected
by, much non-verbal information that is somehow shared by speakers and
hearers.
``Mental models'', often referred particularly in the field of
cognitive science, may include this aspect, that is, integration of
verbal and non-verbal information that represent the utterance
situation for participants in conversation.
We believe that most of speech dialogue systems for real-world application
must somehow include formulation and implementation of mental models in the
above sense. Thus our goal has been to model and implement a
dynamical construction process of mental models, using the domain
of human-robot speech dialogue communication.
Actually, last year we succeeded in formulating and implementing
a speech dialogue system that integrates
sensory information from robot's sensors as non-verbal information.
This system, called Linta-II, is able to construct
a mental model from the situation by using what we call
the attention mechanism and by fusing sensor data, and
use it to generate appropriate utterances.
One weakness of this system is that no ``motivation mechanism''
is implemented. Also, the system, though it is reactive
to changes of external environments, needs more capacity for
reactiveness to avoid danger.
Our new system is able to motivate utterances by itself, and
also can manage reactiveness, by incorporating a reactive
planning algorithm like Georgeff's PRF (Procedural
Reasoning System).
We suppose that scenes, in which a speech dialogue with a
autonomous mobile robot is embedded, is limited.
Dialogue plans are described as reactive plans that
possess an ability to respond immediately to
changes of linguistic context obtained from speech inputs and changes
of external environments obtained from sensory inputs.
When the situation matches more than one plan, the dialogue
manager mediates these plans, based on some heuristic strategies.
That is, when there is no plan being executed, the most urgent action or plan
for the situation is selected.
In the other cases the dialogue manager mediates matched plans in
plans being executed: if no plan matches to the situation, the action or
plan is selected in the aformentioned way.
It is assumed that plans has priority values with
the range between zero and one. The priority value of the
plan being executed is reset to one when the dialogue of
the plan is selected. When the plan called from another
plan is selected, the priority value of the plan calling the selected plan
is updated in some predetermined fashion.
We designed and implemented this speech dialogue system
that includes the reactive planning subsystem to
realize utterance motivation and more reactiveness.
To accomplish that, we first implemented a voice recognition
board and a task planner on the autonomous mobile robot
developed in our laboratory.
Then, the dialogue planner mentioned above was implemented,
which completed our system for speech dialogue with autonomous
mobile robots. The overall system integrates goal-oriented
dialogue planning and reactive dialogue planning. The
former is in a sense similar to those in speech dialogue
systems proposed in the conventional work on human-computer
interaction. The latter is concerned with external and contextual
changes. This integration made our speech dialogue system
enable to realize human dialogue with autonomous mobile robots, whose
``mental models'' are reconstructed according to
dynamical changes in external environments.
Keywords: dialogue planning, reactive planning, mental model, sensor fusion, autonomous mobile robot