The consultation system described in this paper is written in Common Lisp and runs on Workstations under UNIX. The input consists of Italian sentences that are processed by the GULL Natural Language Interpreter [DiEugenio and Lesmo1987], [Lesmo and Terenziani1988], [Ardissono et al. 1991]. The GULL interpreter performs morphological, syntactical and semantic analysis of the sentences and produces a semantic net representation of them. It also keeps contextual information necessary for solving referential problems.
The Plan Recognition component receives these nets as input and uses the
domain library as a knowledge base to produce instantiated CMs corresponding
to the recognized plans. The plan library currently contains about 90 actions
as the ones shown in Figure 
, like speaking
to professors, borrowing or returning books and papers in the department
library, getting information about courses and talks, etc.
While the plan-recognition algorithm is implemented, the clarification dialogues algorithm is not: if the active CMs are involved in a relevant ambiguity, a choice is made looking at them for discarding those not reasonable for the user's plans. Also, there is no NL generation component, while the selection of the contents to be put in the system's replies has been partially defined: the Response Production component is responsible for the possible expansion of the active CMs; in fact, the analysis of the input sentence and the structure of the active CMs enable it to identify the contents of the answers, (e.g. if we have a question like ``Come posso ottenere l'autorizzazione per usare le workstation?'' [How can I get the authorization to use the workstations?] the expanded CM contains the steps of the action of getting such authorization and these data suffice to select the contents of the answer).
The user modeling component is at the moment under development: we have defined
10 stereotypes ( Human, Student, Department-employee, Teacher, Librarian,
Wealthy-person, Not-wealthy-person, etc).
For the moment, no validation of the system on corpus data has been carried out. However, it is important to spend some words about the improvements that the integration of user modeling and plan-recognition techniques are expected to produce in the disambiguation capabilities of the system; clearly, the presence of a small plan library makes the plan-recognition task easier because, in principle, there are few alternative actions to choose when a user asks questions to the system. However, an increased size of the library would require an increased filtering power in the selection of the CMs: in such a case, the opportunity of taking into account the peculiarities of the user would be even more important than in the present situation. Anyway, even in case it is not possible to reduce the active CMs to a set where no relevant ambiguities appear, the possibility of discarding some of the hypotheses is useful to reduce the size and complexity of a clarification dialogue.