Recent improvements in speech recognition technology has made spoken natural language dialog a viable means of human-computer interaction. An important unresolved issue is the handling of miscommunication. By studying previously recorded human-human and human-computer dialogs, the following strategies for reducing miscommunication in natural language dialog will be investigated: ( 1) extending an expectation-driven model for input understanding; (2) developing a context-based model for selective verification of user inputs; and (3) developing methods for engaging in subdialogs for resolving miscommunications. These theories will be implemented and evaluated under both simulated and experimental conditions. The results from this project will provide a new benchmark in evaluating the adequacy of current dialog theory as well as provide further data on the characteristics of spoken human-machine dialog. The project will also develop several enrichment educational programs in support of the predominantly rural and less economically developed eastern part of North Carolina.
