Yolanda Gil (gil@isi.edu, (310)448-8794) 
       Jihie Kim (jihie@isi.edu, (310)448-8769) 
       Jim Blythe (blythe@isi.edu, (310)448-8251) 


One of the biology experts who have used SHAKEN during the summer evaluation said "I donnot really know whether there is a possibility of standardizing the entire process. But it would be better to document some of the processes which you think are standardized".

To be able to help them we'd like to develop a dialogue manager that uses KA plans to keep track of the user's actions and can make suggestions for how to proceed. There are several things we need to do in the process.


    One of the experts describes his process of building cmaps step by step in his feedback:

    • (find concepts) make use of words that already exist in the Shaken and are mentioned exactly in the text
    • if not, search for a simular concept. Read its specification and try to satisfy its requirements
    • make my own entities/events starting from the bottom. The cmap design has a hierarchy. First, I create building blocks on which other cmaps are dependent. Thereafter expand the hierarchy.
    • make a rough cmap and try to see wheter it fits the model, which Shaken accepts.
    • go for completing the cmap

    We can develop a set of typical sequences of KA steps/substeps they can follow in the process of building KBs. The KA steps should also enable users to check if the KBs are consistent with what they expect to build and there is no error in the KBs. We can use these sequences to design the KA dialog, which can highlight what they have done so far and what they need to do next.


    To develop the library of KA plans we would draw ideas from Intelligent Tutoring Systems, in effect trying to make SHAKEN a smarter student. An initial (old) outline is posted at http://www.isi.edu/~gil/rkf/ITS.ppt. For example, ITS often start by showing the student the purpose of a lesson, partly to give the student a roadmap but also to detect any gaps in the student's prior knowledge assumed by the lesson. There are many other interesting points about controlling nested subdialogues, teaching strategies, providing immediate feedback, etc.

    There are also known strategies that are used by very best teachers in tutoring inquiries [Collins and Stevens], and we may adopt some of their strategies based on their applicabilities to the SHAKEN system. For example, there are a variety of strategies for elliciting k from students (in the process of teaching how to derive new rule or theory), suggesting what to do, or commenting. Especially, there are strategies for eliciting missing steps in logical chains and strategies of asking for differences/similarities in factors between different cases.

  • MAINTAIN AGENDA ("status" window)

    The dialogue state would be shown to the user as a "status" window, where the user could see the currently active KA plans as a list of bullets and sub-bullets, each bullet would be checked if completed and collapsed when all its sub-bullets are completed.

    The dialogue manager would support: 1) generation, to suggest to the user what to do next based on the library of KA plans, 2) recognition, to track what the user is doing in terms of the KA plans. The user could choose to toggle between a "suggest" mode (generation) and a "comment" mode (recognition). We could also keep a history of the user's sessions, and be more forceful to impose the "suggest" mode to new users or when a more experience user tries a new feature/capability of the system.

    The agenda items will be organized based on various priority rules and the user's experience level. Some of the priority rules used in tutorial inquiries seem also useful.

    • errors before omissions
    • shorter fixes before longer fixes
    • prior steps before later steps (in temporal or causal sequence)
    • simple (or easy) tasks before complex (or difficult) tasks
    Additional priority rules can be developed based on the typical KA tasks we find in 1. We expect that this will allow them to allocate their time among various tasks more efficiently.

    A new capability that we'd like to support with this work is meta-level interaction with the user. By allowing users to manipulate/edit the status window, they could indicate whether a certain task is completed (e.g., I'm done adding sub-steps).

An intriguing possibility is to use this as a way to determine/compare the quality of human tutors by looking at how and how much they complete the KA plans (how organized their lessons are, how thorough with details), and eventually to train human tutors through practicing with the system.

We have done some previous work in this area on KA Scripts and PSMTool, several papers are available at http://www.isi.edu/expect/link/papers-ka.html.

EXPECT Homepage.

RKF Homepage.