Stephane Auberger, Eiichiro Sumita, Hitoshi Iida
A comparative study of
Query Reformulation methods
on Vector-Space Models
Abstract:Relevance feedback is a well-known method developed to improve the effectiveness
of information retrieval systems. It is based on the automatic and iterative improvement of the textual queries supplied by the users. After a brief overview of the system
performance, this paper describes several different approaches and further refinement
of a standard query reformulation method ("Ide-dec hi formula"). The main research
was focused on using information about the origin of each particular term in the query
("modified Ide-dec hi formula") and especially information on terms in non-relevant
documents ("Common Term System"). Also reviewed are less-expensive methods for
decreasing the retrieval time as well as the size of query and document vectors ("fixed
expansion size" and "fixed vector size"). In general, all methods are found to be effective in improving the performance on a standard test suite, the Virginia Disc One
collections.