The problem of disambiguating among the various hypotheses on a user's plans
and goals is partially dealt with in the focusing phase, when the system tries
to relate the user's sentences with the context established in the previous
part of the
dialogue and, if the shift in focus involved by the last utterances signals an
inconsistency in some of the current hypotheses, they are rejected. In order to
make the disambiguation task more effective, we use
three heuristic rules, that exploit the information collected in
the UM (by means of the acquisition and expansion rules described in the
previous section) for identifying further conflicts among the candidate
hypotheses and the user's intentions.
The disambiguation rules are applied to the set of CMs built by the system in
order to decide which of them are ``good candidate'' hypotheses (if different
good candidates remain, these are then considered in the phase of evaluation of
the relevance of the ambiguity). CMs containing actions in conflict
with the UM, or actions that the user is not supposed to know, are not rejected,
because they could be recovered subsequently, if no good candidates are left
after the disambiguation rules are applied, and they could be used by the system
for starting a repair dialogue with the user (however, this aspect has not
been investigated). The following rules are used in
various steps of the analysis for reducing the set of candidate
hypotheses on the user's intentions to the most plausible ones (see
Figure 
).
In the next section we will describe a detailed dialogue example. However,
before doing that, we want to show a simpler one, that will clarify the use of
the UM acquisition and expansion rules, of the select-common-actions
procedure and of the add-more-specific-actions in the dialogue
interpretation process. Referring to Figure , suppose
that there are two possible ways of getting a text, borrowing it from the
library or buying it, and that the user says:
``I'm a student. I need the book for the Algebra course''.
In the analysis of the second sentence, the system builds a CM containing the
GET-BOOK(IS,Algebra-book) action, that is more general than BUY-BOOK
and BORROW-BOOK.
The UM contains the information belonging to the Human stereotype, that is
always active and contains general information about rational agents. Moreover,
the user's assertion activates the Student stereotype, that is
related to the Non-Wealthy-Person stereotype (usually, students have
limited economic possibilities). Since one of the goals in the
Non-wealthy-person stereotype is ``limit the expenses'', the
GOAL(IS,
Has-Less-Money(IS)) formula is added to the individual user model.
The Human and Student stereotypes suggest, respectively, that the
user
knows that a postcondition of the buying action is to have less money and that
another way to get a book is to borrow it from the CS library; so,
by applying acquisition rule it is possible to
infer 
Intend1(IS,BUY-BOOK(IS,Algebra-Book)). Furthermore, the
select-common-actions procedure adds to the UM the formula
Intend1(IS,GET-BOOK(IS,Algebra-Book)), corresponding to the current focused
action in the CM. Hence, expansion rule yields:
Intend1(IS,BORROW-BOOK(IS, Algebra-book)). At this point, the system
appends the more specific BORROW-BOOK(IS,Algebra-book) action under the
GET-BOOK(IS,Algebra-book)
action in the CM (step of the dialogue interpretation algorithm).
The BORROW-BOOK(IS, Algebra-book) hypothesis could not have been identified immediately after the user's sentence, without the help of the information in the user model. On the contrary, in this way it is possible to build an answer that better satisfies the user's needs.
