Generating Tailored Worked-out Problem Solutions to Help Students Learn from Examples

actions are decomposed until primitive communicative actions (executable as speech acts) are reached. In performing this task, the text planner relies on the set of explanation strategies that specify possible decompositions for each communicative action and the constraints dictating when they may be applied. These constraints are checked against the solution graph and when they are satisfied the decomposition is selected and appropriate content is also extracted from the solution graph. For illustration, Figure 4 (a) shows a simplified explanation strategy that decomposes the communicative action describe-solution-method. Possible arguments for this action are, for instance, the Newton’s-2-Law and the Conservation-of-Energy methods. Looking at the details of the strategy, the function find-steps (:constraints field) checks in the solution graph whether the method has any steps. If this is the case, the steps are retrieved from the solution graph and the describe-solution-method action is decomposed in an inform-about primitive action and in a describe-method-steps abstract action. The output of the planning process is a text plan, a data structure that specifies what propositions the example should convey, a partial order over those propositions and the example rhetorical structure. A portion of the text plan generated by EG for Example1 is shown in Figure 4(b). The propositions that the example should convey are specified as arguments of the primitive actions in the text plan. In Figure 4(b) all primitive actions are of type inform. For instance, the primitive action (Inform-about (act-on Jake weight)) specifies the proposition (act-on Jake weight), which is realized in the example description as “the other force acting on Jake is his weight”. In the text plan, the communicative actions are partially ordered. This ordering is not shown in the figure for clarity’s sake; the reader can assume that the actions are ordered starting at the top. The example rhetorical structure consists of the action decomposition tree and the informational/intentional relations among the communicative actions. For instance, in Figure(b), the rhetorical structure associated with the action describe-solutionmethod specifies that, to describe the solution method, the system has to perform two actions: (i) inform the user about the method adopted; (ii) describe all the steps of the method. Between these two actions the Enable intentional relation and the Goal:Act informational relation hold. All the informational/intentional relations used in EG are discussed in (Moser, Moore et al. 1996] We clarify here only the meaning of the Enable relation because this relation is critical in supporting gap-filling selfexplanations. An intentional Enable relation holds between two communicative actions if one provides information intended to increase either the hearer’s understanding of the material presented by the other, or her ability to perform the domain action presented by the other. 168 CARENINI AND CONATI