Abstract

Natural language interfaces support human-machine interactions in a way that is natural and that does not require elaborate user training. The two main problems in human-computer dialog are resolution of ambiguity and incompleteness. A model-based architecture is introduced in this thesis that provides a solution to these problems. The uniqueness of the solution is based on five interacting models: a language model responsible for the processing of individual user queries, a domain-independent dialog model providing dialog management, a user model keeping track of the user's tasks and preferences, a world model resolving references to both common-sense and domain-dependent knowledge, and an error model handling erroneous user input. The implementation and testing have been carried out in the context of electronic news delivery as the chosen application domain. The design of the dialog and error models is sufficiently generic to be used in different application domains; the language, world, and user models are application-specific and encapsulate domain-dependent information. The language model is based on a set of syntactic classes that put restrictions on user input. These classes were derived using a brief study of a sample population of users. A Java-based implementation of a prototype of the five models has shown the feasibility of our approach to dialog management.