Managing Casual Spoken Dialogue Using Flexible Schemas , Pattern Transduction Trees , and Gist Clauses

We describe an approach to managing casual spoken dialogues in a virtual human dialogue partner, designed to enable the user to practice basic dialogue skills. We use script-like schemas to loosely guide the dialogue, where the steps of a schema are specified in a declarative language that allows for both explicit and abstract description of speech acts. The steps become particularized and can be added to or omitted as the dialogue proceeds. Internally, the dialogue manager casts its own contributions and those of the user as gist clauses – simple, explicit, context-independent versions of what was actually said. Both gist clause computation, and generation of responses, is enabled by hierarchical pattern transduction. These methods show promise in the implementations we have so far developed.