The main objective of the work presented in this thesis was to develop an agent-based dialogue manager that could use a nested belief model to plan efficient dialogues. To this end, a planning system has been implemented, and demonstrated using a number of examples. It is now in a suitable state to be reused for further research on dialogue planning, or to serve as a design model for agent-based dialogue managers. It is easy to program the system with an input file specifying plan rules and initial beliefs.
For the planner to generate correct dialogue and for it to generate a correct expectation of the user's dialogue contribution, it was required to adhere to current theories of dialogue planning. Using Carberry's model of plan recognition in dialogue , and Pollack's observation that agents apply different plan rules and state beliefs to the same plan , a model of plan formation was developed. It was argued that while this model of dialogue planning is useful in selecting correct plans, it does not say which of the valid plans is the most efficient. Therefore the planning theory was used to generate the alternatives in a Bayesian game . This allowed different dialogue strategies to be evaluated, and for a user model, in the form of a probabilistic nested belief model, to be used to find out which of the strategies had enabled preconditions.
One condition of acceptance of the planner is that the dialogues it produces are more efficient than those produced by current dialogue management systems. It was shown using examples and with simulated dialogues that the planner is more efficient by virtue of using a user model in the form of a probabilistic nested belief model. However, the evidence given here is only part of what is required. The simulation experiments assumed an ideal dialogue partner, and there is no direct empirical data to support the claim that the planner would be as effective against a human partner. However it is clear enough that the introduction of the user model could not result in a system that is any worse than current systems. Human trials are therefore the most important item for future work. As well as measuring the system, these trials may lead to a further cycle of development, where the planning model is changed to reflect the behaviour of the user. A secondary remaining objective in evaluating the system is that it is empirically compared with the reinforcement learning approach that is often used in learning dialogue strategies. However, it has been argued that in some situations, the planner would be superior to reinforcement learning.
A second condition of acceptance is that the planner can be used by a dialogue system designer without his full understanding of the complicated mechanisms used to decide strategies and to revise the belief model. It is apparent from the examples given in Chapters 4 and 5 that this has been achieved. The designer need only input a file describing the dialogue plan rules. An example of such a file appears in Section A.2.
The planner has not yet been integrated with the components that make up a complete dialogue system, namely a text planning and speech synthesis system, and a speech recognition system. The speech recognition part is the most interesting since it must use a statistical model to translate from speech to a parameterised dialogue act suitable for use by the planner. This was discussed as an item for future work. At the moment, the planner is ready to be used only with descriptions of dialogue acts for input and output.
A set of domain-independent dialogue acts were developed which can be used to generate negotiation dialogues over a domain-level plan. These were shown to be useful in making value of information judgements, with application in collaborative planning problems. The negotiation acts were built into the planner's repertoire allowing automatic generation of negotiations.
To contribute to the research community, the implemented planner is intended to be released using the name "PED" (Planner for Efficient Dialogues). It is intended that it and its derivatives should remain free and open-source under the terms of the Lesser GNU public licence. This licence at the same time allows the system to be freely used as a "library", linked to non-GPL software, for example, to a non-GPL speech recogniser. A guide to this implementation can be found in the appendix.