Discussion Notes 28 Oct., by Jermey Heiner

User Modeling in Expert Man-Machine Interfaces: A Case Study in
Intelligent Information Retrieval

Giorgio Brajnik, Giovanni Guida, and Carlo Tasso

==========================
Question 1:

 - a system like this tends to put people out of work, but it makes
   the resource accessible to more people.

 - it may make more economic sense to learn it yourself instead of
   paying someone, especially if you must use it often.

 - browsing some databases would be too costly.

==========================
Tangent:

 - having the system learn about the user opens up many possibilities,
   but this system is very intrusive.

 - the natural language input looks like it would involve a lot of
   tedious typing.  but we don't really know how it is implemented.

==========================
Question 2:

 - it doesn't say how fast the system runs.  the performance will be
   limited by the searching through the model tree.

 - there's a tradeoff between performance and accuracy of results.  it
   might be acceptable to some users for it to take a long time but
   return exactly what they were looking for.

==========================
Tangent:

 - an example: installing from floppy disks.  a company compressed
   data on the disks, minimizing the disk swapping, but the users
   complained that it took so long to install, perhaps because they
   had less to do.

 - they might have felt better if there was an indication of exactly
   how long each disk was expected to take - so they could go get a
   cup of coffee.

==========================
Back to Question 2:

 - the question could be clarified:  for whom is it efficient, the
   system or the user?

 - it depends on the price of doing it wrong:  if it's OK to mess up,
   then this system is not as useful.  [browsing might be better]

 - it might not be as bad as it appears:  the tree only needs to be
   walked through when the information is not available at the leaves.

 - which happens all the time when you first start using the system
   until eventually it learns about you.

 - changing the direction of the "is-a" relations gives a "could-be"
   relation, but using this to search the tree would lead to an
   explosion of possibilities.  Kay suggested to always start at the
   leaves and move up the tree, but does this make as much sense when
   the user model can include many leaves?

==========================
Question 2.1: How well does this scale up?  How many users can it
handle?

 - you would probably not want to store all the user's information on
   a mainframe.

 - the user might have bad days, and so needs control of whether or
   not sessions are merged into the database.

 - [looking at model info. on page 177]  if it's so natural language
   oriented, it would be easy for the user to comprehend their own
   model.  you could even build a GUI for this.

 - we can't really tell from the paper how the natural language output
   corresponds to the internal data structure.

==========================
Question 3:

 - what is meant by "dialog shorter"?  "more concise".

 - [example page 178]  it seems like the system ought to go ahead and
   do some searching instead of endlessly asking for clarification.

 - the user might become frustrated for two reasons: there were no
   results (zero references found), or by being asked too many
   questions.

 - query formulation is not always obvious or easy.  sometimes seeing
   the results of intermediate queries helps to reformulate.

 - browsing the actual books on the library shelves is helpful.

 - the people who are paid to do these searches work like this system:
   they use hierarchical classifications.

 - the system should not ignore user commands just because (for
   instance) they are too expensive: why shouldn't the user be allowed
   to search all abstract key words for "ramses"?

 - at the beginning of this quarter i searched inspec for "adaptive
   user interfaces" and got more than 1600 hits.  so i went to an
   expert to get help with the query.

==========================