A Distributed Architecture for Real-time Dialogue and On-task Learning of Efficient Co-operative Turn-taking - Coverbal Synchrony in Human-Machine Interaction - page 320

The system finds prosodic features that can serve as predictors of
human turn-giving behavior, and employs incremental (real-time)
perception to work in as close to human natural dialogue speeds
as possible. As the system learns on-line, it is able to adjust quite
quickly to the particulars of individual speaking styles. At present, the
system strongly targets the temporal characteristics of human-human
dialogue, something that is mostly considered irrelevant by prior and
related work on dialogue systems, as the above discussion shows.
While the results are encouraging, there is room for significantly more
work to be done in this direction.
At present, the system is limited in two main ways: it assumes a
small set of turn-taking circumstances where content does not play
a role and a single shared goal of cooperative “polite” conversation
is assumed, where both parties want to minimize speech overlaps.
Silences caused by outside interruptions—e.g. barge-in techniques and
deliberate interruption techniques—are therefore a topic for future
study. The system is highly expandable, however, as it was built as
part of a much larger system architecture that addresses multiple topic-
and task-oriented dialogue, as well as multiple communication modes
such as gesture and facial expression. In the near future, we expect to
expand the system to more advanced interaction types and situations.
The learning mechanism described here will be expanded to learn
not just the shortest durations but also the most efficient turn-taking
techniques in multimodal interactions under many different conditions.
Because of the distributed nature of the architecture, the turn-
taking system is constructed in such a way as to allow a mixed-control
relationship with outside processes. This means that we can expand
it to handle situations where the goals of the dialogue may be very
different from being “friendly”, even adversarial, as, for example,
in on-air open-mic political debates. How easy this is remains to be
seen; the main question revolves around the learning systems—how
to manage learning in multiple circumstances without negatively
affecting prior training.
Acknowledgement
This work was supported in part by research grants from RANNIS,
Iceland, and by a Marie Curie European Reintegration Grant within the
6th European Community Framework Programme. The authors wish
to thank Yngvi Björnsson for his contributions to the development of
the reinforcement mechanisms.