In this paper, we describe a framework for dealing with information-seeking dialogues in a restricted domain (a Computer Science department). In particular, we consider a consultation system delivering information about the various tasks that the users may want to perform, like accessing the library, etc. When an automated system plays the role of a consultant, the cooperativeness is directly related to its ability to understand the user's plans and goals [Allen1983], [Carberry1990b]. In such dialogues, the determination of the users' intentions is generally complex because they do not always state them explicitly. More often, they only ask general questions, or questions about particular aspects of the tasks. In these cases, the system builds different hypotheses on their plans, according to the partial information collected during the dialogue. A possible way of dealing with the ambiguities among the hypotheses is to solve them through clarification dialogues, where the system asks the users explicit questions about their intentions. However, too many investigations may be boring and, to reduce them, it is important to exploit not only the information extracted from the previous part of the dialogue, but also domain information and knowledge about users. As Kass and Finin state [Kass and Finin1988], a user modeling component is essential in cooperative consultation systems for improving the naturalness and the helpfulness of dialogues. Moreover, information about the users' beliefs, goals and knowledge plays a fundamental role in the task of disambiguating among the hypotheses about the plans in their minds during the dialogue, as well as in the capability of the system to adapt its answers to their competence level [Paris1988], [Chin1989].
In current models of plan recognition it is difficult to find a satisfactory integration of plan inference and user modeling techniques. Generally, a strong emphasis is given to the identification of the users' plans and plan recognizers have rather simple user modeling abilities. On the other hand, although the idea that UMs also should be concerned with users' plans is commonly shared in the user modeling community (see [Wahlster and Kobsa1989] and [Kass and Finin1988], where plan recognition is considered a subset of user modeling), there are very interesting user modeling systems that model users' knowledge and beliefs, but have problems in the identification of their plans, because they don't have suitable action identification capabilities. We believe that to build systems able to characterize their users appropriately, plan recognition and user modeling should be considered concurrent activities that exchange information to cooperate in the task of recognizing the intentions of the observed agents.
In this paper we propose an integration of user modeling and plan recognition that takes into account different aspects in the identification of users' plans and goals: the UM is developed by means of UM acquisition rules and by extracting the users' intentions from the hypotheses built during the analysis of their utterances. Moreover, the contents of the UM are used for reducing the number of alternative hypotheses on their plans, so that a bidirectional flow of information is achieved.
The importance of integrating user modeling and plan-recognition techniques may be shown by means of an example: in some Computer Science (CS) Departments, students may access the computer labs only if they possess a lab pass that identifies them. We suppose that first year students obtain this card by going to the student secretariat and providing their registration number, while advanced students only have to renew the pass every year by going to the secretariat, returning the old card and providing a document that proves their current enrollment in the Department. Suppose that a student says to a human consultant:
If the consultant knows that CS exams require laboratory work and that CS students must take the DS examination during the first year of university, it is natural for her to assume that the student is a beginner, and that he must apply to get the laboratory pass for the first time. So, she may answer specifying the modalities for obtaining the lab access, and not those for renewing it. Notice that this kind of information is only available in an automated consultation system if it keeps a model of students, and it is able to distinguish beginner students from advanced ones on the basis of the intentions that they express during a dialogue. If a system does not use such information, it is impossible for it to generate an unambiguous hypothesis about the user's plans and the only way to answer correctly to the student's question is to ask him whether he already possesses a pass for the labs. In the above example, the dialogue can go on in the two following ways:
System with embedded user model:
System without embedded user model:
In the second case the solution to the problem is still quite easy, because it is sufficient to ask one question in order to discard one of the two alternative hypotheses. However, in general, many data may be required to solve such conflicts.
This paper is organized as follows.
In the next section we give some background in Plan Recognition and User
Modeling and discuss the relationship between them. Then in
Section 
we give an overview of the main components of our
interpretation framework.
Section describes the interpretation process.
The user model is described in detail in Section , where we
will present some rules for expanding it by means of inferences suggested by
the knowledge about the domain actions. Furthermore, we will discuss
safe additions to the user model
obtained by extracting facts that are common to all the alternative active
interpretations of the user's sentences. Some disambiguation rules based on
the contents of the user model are described in Section .
In Section we will provide a detailed description of the
example sketched above, while Section contains a brief
description of the implementation of the consultation system.