...them.
If an action has more than one alternative decomposition, for each of them we define an associated more specific action, that corresponds to a different modality of execution. In this way, we deal in a uniform way with more specific actions and actions with more than one possible decompositions.
...information.
The hypothesis that the user's world model is a subset of the system's is quite restrictive, and in general it would be more realistic to suppose that the user's and the system's world models intersect in some way. For example, the user could possibly want to execute some actions that are not included in the system's plan library. In these cases, the system should be able to reason on the contents of its knowledge base for finding useful information and being in some way cooperative (see [Pollack1990], [Eller and Carberry1992] and [Calistri-Yeh1991] for a treatment of misconceptions and novel plans).
...CMs)
Following the guidelines in [Carberry1990a], CMs could provide in turn contextual information that NLI could exploit in the identification of the referents of ambiguous descriptions occurring in the sentence. However, as it can be seen in the figure, this flow of information is not currently present in our framework.
...dialogues.
[Raskutti and Zukerman1994] identify precise criteria that the system may use to choose the queries to ask during clarification dialogues and report that in some domains (like medicine and travels), the shortest dialogues are not necessarily the most acceptable ones, since the users prefer certain stereotypical sequences of questions. Although this point is important in the organization of clarification dialogues, we think that it is also important to try to reduce the ambiguities a priori. So, we are interested in limiting questions about information that, in fact, can be inferred from what the system already knows.
...CM.
There is a trade-off between the intention to be cooperative and the risk of generating complex answers, or to increase the number of ambiguous hypotheses on the user's plans. For this reason, the expansion phase is limited to the cases where it does not lead to the generation of alternative CMs.
...CMs.
In the dialogue, sentences A1 and A2 are uttered by the user in the same conversational turn. However, we will analyze them separately to show the whole interpretation process.
...logic.
Weak S5 is a modal logic with an associated possible-worlds semantics where the accessibility relation among worlds is Euclidean and transitive. The accessibility relation R is Euclidean iff for any worlds u, v and w, if uRv and uRw then vRw. For details, see [McArthur1988].
...concept(conc))
This operator has been introduced for expressing the agent's knowledge of concepts of the domain; actually, also the relationships among the concepts should be kept into account, but we have not done it for the moment. For the notation ``concept(conc)'' refer to [Kass1991].
...)
The notation chosen for Knowref is substantially equivalent to that introduced in [Allen and Perrault1980].
...goal
The concept of persistent goal has been introduced with the following intuitive meaning: an agent has a persistent goal x if and only if s/he believes that x is false and s/he has a goal that, at a certain point, x will be true. Moreover, the agent will have x as a goal until s/he will believe that it is true, or that it will never be true, or s/he will believe a condition q has failed: q is interpreted as a motivation to bring about x. If q fails, the agent is not interested any more in making x true. We will not consider the q parameter in the rest of the paper, because it is not relevant for our work. Since the formal definition of Intend1 requires an extended explanation, we suggest that the interested reader refers directly to [Cohen and Levesque1990].
...rules.
Part of the debate on the use of stereotypes is based on the problems that arise in dealing with conflicts among stereotypes and user behavior and on the plausibility of stereotypes themselves. With respect to the second point, the problem is that, normally, it is not possible to identify with certainty the sets of domain concepts and goals which some categories of agents really have. In particular, the knowledge and the goals of people can be categorized in some way, but the personal experience of the single agents strongly influences their individual background. In spite of this, there have been some proposals for reducing the drawbacks of the stereotype approach [Jameson1992]. Such proposals have been supported by psychological studies that have found evidence of the fact that, in consultation dialogues, human experts really seem to use stereotypical knowledge [Cahour1992]. In order to overcome the problems due to the non-monotonic character of prototypical information, some user modeling systems embody a truth maintenance mechanism for retracting the assumptions that reveal themselves to be inconsistent with the user's behavior, together with all the inferences based on them [Brajnik and Tasso1994].
...effect
This rule refers to the simplified case where, for the agent to reject an action, it is sufficient that it brings about an undesired effect. In general, agents accept side effects of the actions they want to perform if such effects are not too negative, or they can easily undo the undesired steps (for a treatment of the attitudes of an agent towards the side-effects of an action see [Konolige and Pollack1989])
......).
We show the instantiated formulae contained in the UM; upper-case parameters of actions indicate instantiated variables, while lower-case parameters denote free variables.
...#fCMs1#272>.
In the figures related to the example, the representation of arcs, dashed or plain, corresponds to the notation used for the plan library (thin dashed arcs represent precondition and constraint arcs); 91#91 represents the current focus of the dialogue.
...goals
The rationale for this is that the system knows that, in general, beginner students don't have a lab pass and they are not interested in renewing it anyway.
...etc).
The user model and the stereotypes are represented in a formalism based on semantic nets, like the one chosen for the output of the Natural Language interpreter. However, in the paper we adopted a logic formulae notation to avoid all the technical details required to understand the net formalism.

Guido Boella Dottorando
Wed Oct 23 09:42:15 MET 1996