A Distributed Architecture for Real-time Dialogue and On-task Learning of Efficient Co-operative Turn-taking - Coverbal Synchrony in Human-Machine Interaction - page 306

score hypothesis being available to the rest of the system during
interlocutors' speech, but final utterance is not calculated until at least
one second of silence has been detected.
3.2 Deciders
Our detailed turn-taking model consists of eight dialogue states (see
Figure 4). This represents the states taken when the turn switches
hands. The dialogue states are modeled with a distributed semi-global
context system, implementing what can (approximately) be described
as a distributed finite state machine that selectively applies to the
activation and de-activation of most modules in the system. Context
transition control (“state transitions”) in this system is managed by a
set of deciders (Thórisson, 2008). There is no theoretical limit to how
many deciders can be active for a single given system-wide context.
Likewise, there is no limit to how many deciders can manage identical
or non-identical transitions. Reactive deciders (IGTD, OWTD, ...) are
the simplest, with one decider per transition. Each contains at least
one rule about when to transition, based on both temporal and other
information. Transitions are made in a pull manner: the Other-Accepts-
Turn-Decider, e.g. transits to context Other-Accepts-Turn (see Figure 4).
The Dialogue Planner (DP) and Learning modules (see further
description below) can influence the dialogue state directly by sending
context transition messages I-Want-Turn, I-Accept-Turn, and I-Give-
Turn; however, all these decisions are under the supervisory control
of the DP: If the Content Generator (CG) has some content ready to be
communicated, the agent might want to signal that it wants a turn and
it may want to signal I-Give-Turn when the content queue is empty
(i.e. have nothing to say). Decisions made by these modules override
decisions made by other turn- taking modules. The DP also manages
the content delivery; that is, when to start speaking, withdraw,
or raise one's voice. The CG is responsible for creating utterances
incrementally, in “thought chunks”, typically of durations shorter than
1 second. We are developing a dynamic content generation system
at present; based on these principles the CG currently simulates its
activity by selecting thought units to speak from a pre-defined list.
It signals when content is available to be communicated and when
content has been delivered.
In the present system, the module Other-Gives-Turn-Decider-2
(OGTD-2) uses the data produced by the Learner module to change the
behavior of the system. At the point when the speaker stops speaking,
the challenge for the listening agent is to decide how long to wait
before starting to speak (OGTD-1 has a static behavior of transitioning