In the closing session, the participants discussed what they saw as
the most important research questions and challenges. Six emerged as
- How can we more efficiently evaluate the quality of predictive models and the systems which use them?
- Many aspects of human communication dynamics are interesting, but
which are the ones that really matter, in human-human interactions and for
systems that interact with people?
- How can we make it easier to effectively use the findings of the various component research fields, scattered and hard to access as they are?
- Assuming that we can build predictive models, how can we use their
output in decision-making for dialog systems and other applications?
- We have good algorithms for processing words and good algorithms
for processing non-symbolic signals, but how can we cleanly combine
symbolic and non-symbolic information, as the need arises in
recognition, in prediction, and in behavior synthesis?
- What methods do humans use to predict/simulate the near future in dialog, and how accurate are they?