In a theoretical investigation two approaches to dialogue management are compared: one is based on recognizing the user's plan from his goals and intentions, and the other on modelling the possible actions of the user in a dialogue grammar. To establish a sound foundation for the design of a dialogue manager, empirical studies were carried out in the form of Wizard of Oz experiments. In such studies users interact with what they think is a natural language interface, but in fact there is a human intermediary. Conducting experiments of this kind requires careful design and a powerful simulation environment. Such an environment is presented together with guidelines for the design of Wizard of Oz experiments. The empirical investigations indicate that dialogue in natural language interfaces lack many of the complicated features characterizing human dialogue. Furthermore, the kind of language employed by the user is dependent to some extent on the application, resulting in different sublanguages.
The results from the empirical investigations have been subsequently used in the design of a dialogue manager for natural language interfaces which can be used in a variety of applications. The dialogue manager utilizes the restricted but more computationally feasible approach of modelling dialogue structure in a dialogue grammar. Focus structure is handled via dialogue objects modelled in a dialogue tree. The dialogue manager is designed to facilitate customization to the sublanguage utilized in various applications. In the thesis I discuss how the dialogue manager is customized to account for the dialogue behaviour in three applications. The results demonstrate the feasibility of the proposed approach to building application-specific dialogue managers for various applications.