UM acquisition rules

The user model is built by means of rules for inferring the users' beliefs from the sentences they utter. In spite of the well-known lack of flexibility of stereotypes with respect to an accurate modeling of the beliefs, knowledge and goals of people, we use them for expanding the UM; in fact, the dialogues we deal with are typically short and, for this reason, it could be difficult to build rich UMs with the only application of UM acquisition rules. Some general prototypes, such as Human, are always active and describe the basic knowledge and goals that every agent is supposed to have. Since at the beginning of each dialogue there is no available information about the users, we suppose that they have an intermediate knowledge on the domain. In this way, if the initial sentences are not sufficient for specifying their knowledge better, the system establishes a default intermediate detail level for the replies to their utterances, until more information is available. During the dialogue, stereotypes are activated by the presence of actions and goals typical of certain classes of agents (e.g. if they state that they are going to take a particular CS examination, it is safe to conclude that they are students). Stereotypes are also activated when users specify particular information about themselves (e.g. ``I'm a CS student''). Many rules have been proposed for acquiring users' beliefs and knowledge from their utterances. For example, if they mention a concept in a sentence (except in a request for an explanation), then the system supposes that they are acquainted with the concept [Chin1989]; if the system uses new terminology and they don't ask for clarification, then they understand the terminology [Chin1989]; etc. Kass and Finin [Kass and Finin1987] proposed some rules based on Grice's conversational maxims that suggest which concepts agents are probably aware of on the basis of the information that they specify or omit during a conversation. Kass also proposed some acquisition rules based on the structure of the domain of application of a consultation system. These and other rules [Kass1991] enable the system to infer that users know certain concepts and relations on the basis of the fact that they know other concepts and relations. Examples of these rules are: if a user knows that concept C specializes B, and that B specializes A, then s/he knows that C specializes A. If a user knows that a concept A possesses an attribute description X, and B specializes A, then s/he knows that B also possesses the attribute description X, etc.

In our framework, the acquisition of the UM is based on rules as those cited above, plus some rules defined below. The first three of them use the knowledge and the goals stored in the UM for identifying actions that a user is supposed not to want to perform: they are particularly important for reducing the number of alternative hypotheses built by the system, because they make it possible to decide that some of the hypothesized actions are implausible. The last rule uses the information about which actions the user knows or wants to perform, together with the knowledge stored in the plan library, for inferring which other actions the agent is probably committed to. Formally:

1.  

    
   Bel(agt, constr(act,c))  Bel(agt, )
   
   		  	Intend1(agt,act)
   
   

   That is: an agent does not want to perform an action if s/he believes that one of its constraints is false (because, in that case, it is not possible to perform the action).

2.  

    
   Bel(agt, post(act, c))  Goal(agt, )
   
   		  Intend1(agt,act)
   
   

   That is: an agent does not want to perform an action if s/he believes that it has an undesired effect.

3.  
   d (Bel(agt,decomp(act,d)) 
   
   		 a (Bel(agt, in-decomp(a,d))   Intend1(agt,a)))
   
    Intend1(agt,act)
   
   

   That is: an agent does not want to perform an action if s/he believes that an unintended step is part of its decomposition.

4.  

    
   Intend1(agt,act)  K(agt, more-specific(a1,act))  
   Know-env(agt,a1) 
   
   (.more-specific(a2,act) 
   
   		 equal(a2, a1)  Know-env(agt, a2) 
   Intend1(agt, a2))
   
    Intend1(agt,a1)
   
   

   That is: suppose that an agent is committed to an action act and the UM suggests that s/he knows the environment of just one of its more specific actions (a1), or s/he does not want to execute any of the other more specific actions s/he knows; then s/he is likely committed to the more specific action a1. In this rule, the information from the plan library and that from the user model are used for restricting the focus of attention to the actions that have been previously mentioned in the dialogue, or that have been attributed to the agent from prototypical information. This can be done because we are making the appropriate query assumption. The predicate ``more-specific(a2, act)'' is true iff in the plan library a2 is one of act's more specific actions.

The above rules are general and don't depend upon the particular domain. However, we consider some domain dependent rules for ascribing particular beliefs and knowledge to users. For example, in the domain of a CS Department, it is likely that if agents know how to borrow a book from the library, then they know how to return it (in general, people that have executed the former action must then have executed the latter as well).