SP500215
NIST Special Publication 500-215: The Second Text REtrieval Conference (TREC-2)
Report of Progress for TREC-II
chapter
W. Kelleher
National Institute of Standards and Technology
D. K. Harman
Basic AssumpUons
is the probability that a document is relevant if it contains
a particular element. Each element is treated as being inde-
pendent of every other element.
3.3 Results Window
The results window contains the list of all documents with a
probability of relevance to the current query/concept. Documents
are ranked by probability of relevance. For each document, the
elements found in the document are listed as well as any previous
relevance information. The document D selected from the
results window can be used in the document window to examine
the document.
4.0 Basic Assumptions
The approach being taken by FORMS is based a number
of critical assumptions.
Using the axioms of probabilities as a foundation for
determining the relevance of documents to a query
* Elements besides words, such as phrases, can be found
to make a significant contribution to retrieval accuracy
* A number of different approaches to identif[OCRerr]ing rele-
vant elements are worth pursuing. These include
proper noun identification, part of speech tagging,
noun phrase tagging, and as yet undetermined relations
that can be extracted from natural language.
5.0 Progress to Date
Frankly, there has been little progress to date beyond basic
system development. In both TREC I and II, we have per-
formed routing queries with very poor results. One obvi-
ous reason is that we have had to perform the analysis by
breaking the texts into small sections and doing even class
B in sections.
6.0 FutureWork
FORMS is designed to provide information on the effec-
tiveness of different approaches.
* To examine different approaches to incorporating rele-
vance feedback into evaluating relevance of other doc-
uments.
Report of Progress for TREC II
277
* To examine whether queries can be analyzed and clas-
sified so that the system can determine which approach
is most likely to be successful for a particular query or
concept.
* To examine whether certain kinds of elements (where
an element is a word, a phrase, a proper noun, a verb, a
cooccurrence etc.) can be predetermined to be helpful
in certain types of queries.
7.0 Results & Conclusion
At this time no results are supported by work performed
with FORMS. However; it is worth emphasizing that the
results in TREC seem to support the view that there is no
specific approach that is going to revolutionize informa-
tion retrieval. Rather, it seems that improvements are
going to come from attention to details and fmding the
right element to use at the right time.