A Distributed Architecture for Real-time Dialogue and On-task Learning of Efficient Co-operative Turn-taking - Coverbal Synchrony in Human-Machine Interaction - page 298

Our current version of the system learns to become better at taking
cooperative turns in real-time dialogue while it is up and running,
improving its own ability to take turns correctly and quickly , with
minimal speech overlap. The results are in line with prior versions
of the system, where the system interacted with itself over hundreds
of trials (Jonsdottir et al., 2008). Evaluation including human subjects
so far includes a within-subjects study of 5 minutes of continuous
interaction with each user (a total of 50 minutes), in three different
conditions: (1) A closed, noise-free, setup with a very consistent
interlocutor—another instance of itself (“Artificial” condition). (2) An
open-mic setup, using Skype, where the system repeatedly interviews
a fairly consistent interlocutor—the same human (“Single person”
condition). (3) An open-mic setup, using Skype, with individual
inconsistencies where the agent interviews 10 different human
participants consecutively (“10 people” condition). The system adapts
quickly and effectively (linearly) within 2 minutes of interaction, a
result which, in light of most other machine-learning work on the
subject—many of which require thousands of hand-picked training
examples—is exceptionally efficient.
The rest of this chapter is organized as follows: First, we review
related work, then we detail the architecture and learning mechanisms.
A description of the evaluation setup comes next, followed by the
results, summary, and future work.
2. Related Work
Models of dialogue produced by a standard divide-and-conquer
approach can only address a subset of a system's behaviors, and are
even quite possibly doomed at the outset. This view has been presented
in our prior work (Thórisson, 2008) and is echoed in other work on
dialogue architectures (cf. Moore, 2007). Requiring a holistic approach
to a complex system such as human real-time dialogue may seem to
be impossibly difficult. In our experience, and perhaps somewhat
counterintuitively, when taking a breath-first approach to the creation
of an architecture that models any complex system—where most
of the significant high-level features of the system to be addressed
are taken into account—the set of likely contributing underlying
mechanisms will be greatly reduced (Schwabacher and Gelsey, 1996),
quite possibly to a small, manageable set, thus greatly simplifying
the task. It is the use of levels of abstraction that is especially
important for cognitive phenomena: Use of hierarchical approaches
is common in other scientific fields such as physics; for example,
behind models of optics lie more detailed models of electromagnetic