Objectives

With the motivations in mind of producing an agent-based dialogue system that is both easy for the designer to use, and produces better quality dialogues, the work reported in this thesis sought to achieve:

• The development of a model of dialogue planning, founded on current theories of cooperative dialogue. Such a model is required so that the dialogue acts chosen by the system can be understood by a user whose expectation is based on that model. Symmetrically, the system must be able to understand the user, who generates his dialogue acts using that model.

• The integration with this model of a suitable representation of the system's nested beliefs, since these beliefs determine the dialogue acts that each agent is expected to choose. The nested beliefs were to represent the system's own beliefs, as well as its beliefs about the user.

• Development of a mechanism for planning efficient dialogue, that considers the system's nested beliefs in the choice of a dialogue strategy.

• The implementation of a dialogue manager based on this mechanism, taking as input a set of dialogue plan rules given by the dialogue system designer, and from these automatically generating dialogue acts by planning a dialogue with the user. From the rules alone, it would automatically acquire a nested belief model, improving the efficiency of future dialogues as the model is fitted to the user. The system would be domain independent and useful to designers of dialogue systems without requiring them to understand its inner workings. It is common in designing dialogue systems to separate the components of dialogue management, and generation and recognition of user input and output. Therefore, the planner was not intended to be a complete natural language system. Rather it was intended as the dialogue manager component of such a system, which could equally well be used with alternative dialogue modalities, such as a graphical user interface, or even in dialogues that mix spoken acts with physical acts that are seen rather than heard. The planner should be made available in the public domain as "PED" (Planner for Efficient Dialogue) to facilitate further research in dialogue management systems.

• Demonstration of the implemented dialogue manager by means of a set of example problems. While the implemented planner was not subjected to a correctness proof, which would show correctness in all circumstances, example problems provide a handful of samples of the planner's behaviour, which can be checked by hand. While the examples are not intended to represent the entire space of dialogue planning problems, they nevertheless provide evidence that a number of important problems can be solved by the planner. Since the planner is written in a high-level programming language, Prolog, a correctness proof of the program code would be almost trivial, because the specification to which the planner is proved to adhere would have been almost identical to the program code.

• Calculation of the efficiency gain that can be achieved by using a probabilistic nested belief model in a dialogue system in the context of the example problems. This gain represents the difference in quality between the best dialogue that can be planned without using a user model, and the quality that the planner achieves in exploiting the user model. If the gain is substantial, then the planner can compete with current dialogue managers. A comparison between traditional logical models of belief, and probabilistic models was also to be made.

• Development of a set of domain-independent dialogue acts which can be used in the generation of negotiation dialogues. The negotiation dialogue uses as its subject domain-level plan alternatives that are specified by the planner's ordinary plan rules. These acts are common to many dialogues, and so they are built in to the system, rather than expressed in the planner's input. Using these acts, a negotiation planner was to be developed for producing meta-level dialogues about a domain plan.