Plan recognition for context sensitive help

In this paper we describe a technique for modeling action plans by multilayered symbolic nets. The nets are used as a signal conductive system by a newly developed inference component in order to recognize the modeled action plans. Using spreading activation releases the knowledge base designer from handling interrupts which can disturb a recognition process. Since the inference process can be stopped, continued and backed up without completely reseting and recomputing the inference status, this technique can be used to model UNDOS. The inference component generates an action history which stores recognized executions of actions and subactions on each level. This history can be evaluated by other components. We describe an application for an EXCELTM help system to offer the user a selection of help topics sensitive to his or hcr task handling. The Evaluator analyses the computed history in order to determine crucial usage problems and detect suboptimal task executions which occur using an application program.