Contents

• Introduction
  • Motivation
  • Objectives
  • The structure of the thesis
  
  
• Planning of Dialogue
  • Introduction
  • Planning
  • Belief, desire and intention
  • Speech act theory
  • Planning and plan recognition
  • Meta-level planning
  • Plan recognition and cooperative response
  • User modelling in dialogue systems
  • Generating and understanding dialogues without planning
    • Deciding dialogue strategies using policies
  • Game theory
  • Cooperative distributed planning
  • Dialogue management and user modelling systems
  • Evaluating dialogue systems
  • Conclusion
  
  
• Design of the Planner
  • Introduction
  • Requirements
    • Treating dialogue as a game
    • Using a probabilistic belief model
    • Bayesian games
    • Maintaining the user model
  • Assumptions
    • Agenthood
    • Cooperation and sincerity
  • Design of the planner
    • Overview
    • Agent state
    • Representation and construction of the game tree
    • Example problem
    • Evaluation of the game tree
    • Evaluation example
    • Belief revision
    • Stereotypes
    • Complexity
    • Multilogues
    • Conclusion
  
  
• Evaluation
  • Introduction
  • Implementation
  • Evaluation method
  • Amenable problems
  • Example 1: Risking misinterpretation
    • Plan library
    • Demonstrations
  • Example 2: Goal introduction problem
    • Demonstrations
  • Belief model acquisition
    • Demonstration 1
    • Demonstration 2
    • Demonstration 3
  • Related work on initiative and simulation evaluation
  • Conclusion
  
  
• Planning of Negotiation Dialogue
  • Introduction
  • Value of information
    • Value of information example
  • Negotiation acts and their pragmatic definition
  • Realism
  • Demonstrations
    • Demonstration 1: Telling and proposing
    • Demonstration 2: Holding the floor
    • Demonstration 3: Request and propose
  • Acts for non-cooperative dialogue
  • Design and implementation of the negotiation planner
  • Related work
  • Conclusion
  
  
• Future Work
  • Introduction
  • Evaluation
    • Comparing reinforcement learning with planning
    • Dialogues with human beings
  • Improvements to design features of the planner
    • Evaluation of game trees using a logical model of belief
    • Beliefs about plan rules
    • Improved belief revision
    • Multilogue planning
    • Dealing with complexity
  • Further example problems
    • Long-range planning and user-model acquisition
    • Dry-land algorithm
    • Knowledge engineering and large problems
  • Negotiation dialogue
    • Request and propose
    • Planning in self-interested settings
    • Focussing and plan recognition in negotiation dialogue
  • Combining statistical methods in speech recognition with probabilistic planning
  • Conclusion
  
  
• Conclusions
• Description and usage guide for the implemented planner
  • Software architecture
    • Domain-level planner
    • Negotiation planner
  • Input file format
  • Example input file
  • Index of experiment materials
  
  
• Bibliography

Abstract:

This thesis describes the design, implementation, and evaluation of a dialogue management system. The system is based on current theories of planned dialogue, produced by an intentional agent whose plan is derived from beliefs about the domain state, and beliefs about plan rules. The planner is used to generate the alternatives in a game tree, which is evaluated in the context of the agent's probabilistic nested belief model. It is shown using simulation experiments that the efficiency of the generated dialogues is dependent on knowing the value of the agent's probabilistic beliefs, which justifies the use of such a model in dialogue management, rather than the more traditional logical belief model. Similar experiments are used to evaluate a set of negotiation acts, which are built-in acts that the agent can use to pass information about beliefs that increases the expected utility of domain-level plans. The negotiation planner replaces meta-level dialogue planners, which in the past have been used to generate dialogues using a logical belief model.

Acknowledgement

I would like to thank all those who helped me in producing this thesis. I would like to mention my supervisor, the students and staff of the School of Computer Science and Informatics, the authors of papers referenced in this thesis, and the reviewers of papers associated with it. Free software was important in facilitating my research, therefore I would like to thank and encourage those who contribute to the free software effort. My research was sponsored by Enterprise Ireland.