ISR11
Scientific Report No. ISR-11 Information Storage and Retrieval
Relevance Feedback in an Information Retrieval System
chapter
W. Riddle
T. Horwitz
R. Dietz
Harvard University
Gerard Salton
Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government.
VI-21
the initial result. The co-occurrence correlation function ;[OCRerr]th the
constant alpha strategy achieves the greatest increase in the first and
second updates, an4 the same correlation function, but with increasing
alpha strategy, gives the best result for the third update. Co[OCRerr][OCRerr]binations
using the co-occurrence and cosine correlation functions together with
the constant and increasing alpha strategies perform the best. Combinations
using the simple vector matching correlation function and the alpha
correlations strategy always have low merit.
Another method of displaying the increase in retrieval performance
achieved by relevance feedback is the use of a plot of precision versus
recall. Table 1312 gives such a plot of the same results as those given
in Table 131, in which the cosine correlation function and increasing alpha
strategy is used. The results of the 22 requests are averaged by the use
of a cut-off made after examination of m consecutive documents (m ranging
from 1 to 20), and the precision and recall values are averaged over all
requests at each cut-off point t- obtain the nine average point for each
curve. The four performance curves how the initial search and the three
updated searches The first update gives the biggest improvement in
performance, and the updated searches achieve a very substantial
improvement in performanc[OCRerr] at the high precision end of the curves This
improvement decreases as recall increases, and [OCRerr]s almost lost when 0.7
recall is reached. If performance is compared at a cut-off of four
documents, for example, the third updated search achieves an improvement
over the initial search of more than 0.3 in both precision and recall.
To summarize, evaluation of the relevance feedback methods averaged
over 22 search requests shows a very considerable improvement in performance