A plan recognition architecture for ill-formed dialogue
Rhonda Marie Eller · 1993
Current models of plan recognition impose several unrealistic assumptions on the plan inference process. For example, they build a model of the user's plan from a relatively ideal dialogue assuming that the user has perfect knowledge of the task domain and that the system will always be able to correctly infer the user's plan from his utterances. This thesis presents a meta-rule architecture for incremental plan recognition that initially assumes an ideal dialogue and slowly relaxes the constraints on plan recognition when the system cannot infer the user's plan without violating these constraints. During this relaxation process, the system may hypothesize a correction to its existing model of the user's plan. This approach has the advantage that it doesn't initially consider all possible plans that could be constructed at the outset, including very implausible ones, and it can use meta-knowledge about how an action was inferred to determine whether an action in the system's model of the user's plan is incorrect and the model should be repaired.