In this paper we have proposed a way to embed a user modeling component into a plan recognition framework, so that the two modules can cooperate fruitfully to the understanding of the users' plans and goals during information-seeking dialogues. A rich user model has a strong impact in the interpretation of agents' intentions: when different hypotheses on a user's plans are present, the UM can help to discard some of them, because they are not coherent with the user's plausible intentions encoded in it. So, it is important to build the UM dynamically, by means of acquisition rules, stereotypical knowledge and information about the user's intentions collected during the dialogue.
In our framework, the UM is built by means of both acquisition rules for inferring the users' knowledge and beliefs from their utterances and expansion rules that apply to the already existing set of beliefs in the UM and allow to infer further information by taking into account the domain knowledge stored in the system's plan library. Since the consultation dialogues we are concerned with are in general quite short and we don't suppose that information about a specific user is stored in a permanent structure after each dialogue, we use stereotypical information to build the UM more quickly.
A further way to obtain information about a user's plans is to analyze the contents of the active CMs, so that the actions common to each of the alternative hypotheses can be identified. To take into account both the need to extend the UM as soon as possible and the requirement of having a single UM, we propose not to include all the active CMs into the UM, because they refer to alternative interpretations of the ongoing dialogue, and they would generate alternative pictures of the user. So, we present a procedure for selecting the actions common to all the active CMs and adding to the UM the intentions that belong to all possible views on the user's plans and goals. Expansion rules for the UM are then applied to extend it on the basis of its contents and of other domain dependent information.
In the analysis of the users' utterances, the UM is used for reducing the set of alternative interpretations of their plans: we propose some disambiguation rules that enable the system to prefer the candidate hypotheses containing actions not conflicting with the users' intentions. Moreover, these rules make it possible to identify as possibly wrong hypotheses the plans associated with complex actions that they should not know, or that they might not be interested in. In this way, the ambiguity can be solved or, at least, the set of alternatives can be reduced; as a consequence, the subsequent evaluation of the relevance of the ambiguity is easier and the possible clarification dialogues needed for determining the users' intentions are simpler.
In principle, there are two ways to decide how to update the UM. In alternative to our least commitment approach, some evidence evaluation mechanism could be adopted to choose the most promising hypothesis among the existing CMs. In such a case, an acquaintance threshold could be used as a confirmation level that enable the system to be certain enough about the choice. Then, the entire preferred CM could be used to update the UM. It is rather clear, however, that this approach (proposed in [Carberry1990a], [Charniak and Goldman1991]) has two disadvantages. The first of them concerns the need to associate weights to all the actions present in the plan library: research carried out in other fields (expert systems) has shown that this is a very hard task. Second, this quasi-deterministic approach should provide means for handling errors. In fact, approaches based on plausibility cannot guarantee that in all cases the choice is correct; so, some kind of truth maintenance mechanism must be adopted to undo all consequences of a given choice when it reveals wrong. This makes the overall architecture much more complex. Although we think that this alternative is rather promising, we believe that our approach is more feasible for practical systems today.