Tech Report CS-89-16
Classifying and Detecting Plan-Based Misconceptions
Randall J. Calistri
In order to maintain a truly cooperative dialogue, a natural language system must be able to recognize misconceptions in the user's plans, and yet the diversity of mistakes that can occur makes this task appear overwhelming from the very beginning. However, it is possible to organize plan failures into fifteen different categories of plan-based misconceptions. These categories can be further simplified to five detectable types of misconceptions which only need to be checked in a narrow region of the plan hierarchy. This theory is implemented in a program called Pathfinder, which can detect misconceptions from all five simplified classes.
(complete text in pdf)