LIS329 Midterm Evaluation. Fall 2000. Due 5 PM, Friday, October 27.

Answer all five questions. They count equally.  You may use books,
articles, notes, and computers to complete the problems, but you may
not solicit or receive assistance from other human beings. Please show
all work. Make sure your name is on anything you want graded. You may
submit this to the instructor by electronic mail or in person.

QUESTION 1) In class discussions (and on an earlier exam) I proposed a
system of assigning fuzzy membership grades for index terms based on
the term's occurrence in different parts of the document (title,
abstract, etc.). Propose a set of guiding principles for fuzzy
indexing of documents by human indexers (applying either controlled or
natural language terms). Discuss potential advantages of fuzzy
indexing vs.  binary, and possible problems that might arise in the
execution of your guidelines (by the indexers).

QUESTION 2) Suppose a traditional Boolean search system is modified to
produce ranked output by fuzzy membership grade (Korfhage page
70). Documents stored in the database have been indexed by humans
according to the guidelines you proposed in question 1. If searchers
execute Boolean queries without an understanding of the ranking
principles, what kinds of unexpected or perplexing results might one
expect? Illustrate your arguments with worked examples of queries,
documents, and membership grades.

QUESTION 3) One of the arguments in favor of vector representations
for retrieval is that vector components can stand for any document
attribute, not just the importance of a word or phrase. We've read
about systems that combine bibliographic records with other kinds of
data (genomic or astronomical data), but heterogeneous records in
those systems are only loosely coupled (the systems support different
methods of searching separately). Suppose we model documents with
vectors where some of the components represent index term weights,
while others represent other kinds of attributes (e.g., estimates of
reading difficulty, authority, cost, recency, access time, and
frequencies of non-text features such as illustrations, tables,
equations, or citations). We conjecture an interface that allows users
to include weighted preferences for these features as part of the
query vector. Discuss ways in which relevance ranking might be
complicated or made problematic when similarity or distance measures
are applied to these heterogeneous vectors. Illustrate your points
with simple worked examples using either angular (e.g. cosine) or
metric (L1, L2, etc.) relevance measures (or both).

QUESTION 4) Term weighting by signal to noise ratio favors words with
skewed occurrence distributions. How plausible is it that such terms
are likely to be better for indexing?  Are there factors relating to
the types of documents in the collection that speak to that
plausibility? Take a position on this issue, and argue for
it. Illustrate your points with one or more examples of computed
signal values.

QUESTION 5) It's been argued that adjusting search results with a user
profile may not always be helpful for someone researching a novel or
unusual topic. Review the description of the GUIDO visualization
(Korfhage section 7.4) and explain how well GUIDO answers this
particular objection to user profiles.