Maintaining the user model

An objective set out in chapter one was that the planner be easy to use, with the dialogue system designer needing only to specify the dialogue plan rules. This leads to the requirement that the user model be initialised and maintained automatically by the system. The system should need only to look at the preconditions of the plan rules to know the set of beliefs that need to be held in the user model. Corresponding with this requirement, another main design element is that of a belief revision system, that can observe the dialogue acts chosen by the user and infer from these an explanatory belief state of the user. As dialogue evidence accumulates, it must be compiled into a statistical model of the user's belief state. Over time, as the accuracy of the model improves, so should the efficiency of the dialogue. As well as being used after each turn in the dialogue, belief revision must be used in evaluating game trees in the context of a belief model. For example, a plan in which an agent informs the same fact twice is discounted by using belief revision to update the belief model after the first inform, and then finding out that the second inform has no effect on the outcome of the dialogue, since the fact is already in the belief model. The belief revision system should be derived from those of the user modelling shells described in section 2.12, whose main function was to keep a user model up to date based on dialogue evidence, and using dialogue plan rules. The system will need to go further than these systems however, since they are based on a logical model of belief. Using a probabilistic model, probabilistic rather than logical inference needs to be considered. This entails representation issues as well, since in a probabilistic model, conditional probability using belief networks [50] is used in deduction rather than modus ponens. In contrast with logical models in which belief revision [23] is used to maintain consistency of the belief set, a probabilistic form of belief revision is required. In a logical model, beliefs are absolute and so their update is normally based on a single observation. On the other hand, in a probabilistic model, a set of observations is used to calculate a probability value. Some mechanism is required to maintain beliefs using sets of observations.