James Allen; invited presentation at brandial'06

A Robust Dialogue Agent for Collaborative Problem Solving

While plan- or agent- based models provide interesting formalisms for
producing dialogue agents, most robust dialogue system use much
simpler technology based on finite state machines and/or slot filling
models. Such simpler models allow the designers to encode strong
expectations that enable robust processing, however the complexity of
the tasks they can represent is quite limited. We are developing
dialogue agents that not only can execute hand-coded tasks, but can
also learn new tasks from the user through combined demonstration and
dialogue. All system behavior, including its dialogue capabilities, is
driven by the execution of declaratively specified tasks that model
the collaborative problems solving process in a domain-independent
manner. We will describe this model and explore how it contributes to
producing robust dialogue behavior on a number of different levels. I
will also show running examples of our system as it learns and
executes new procedures that involve finding and using information on
the world wide web.