ISR11 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-19 AP[OCRerr]IDIX B Z[OCRerr]valuation of Relevance Feedback Methods. E. M. Keen Some of the results are here presented in summary form for searches employing relevance feedback, using averages over 22 search requests. Results are computed for three different correlation functions: cosine, co-occurrence and simple vector matching; and also for three different feedback strategies: increasing alpha (1,2,3), constant alpha (2,2,2) and alpha equal to the correlations of the relevant documents (c,c,c). Only 22 of the original 3[OCRerr] requests are averaged, since full results were not available for 12 of the 3[OCRerr]. Nine of these 12 were not processed by all of the above procedures, some because the initial search result [OCRerr]ias very good and no iteration was needed, and the other three because each had only one relevant document, and averages were therefore not believed to be meaningful. miables Bl through B9 give average results using the measures of normalized recall, normalized precision and normalized overall. In Table Bl for example the cosine correlation function is used with the increasing alpha strategy, and the normalized measures indicate the improvement in performance that results from each update. Tables B2 and B3 also illustrate the use of the cosine correlation function, but the increasing alpha strategy is altered to constant alpha and alpha correlations respectively. Tables [OCRerr] through B6, and Tables B7 through B9 cover these same three alpha strategies but use the co-occurrence correlation function and the simple vector matching correlation function, respectively. Comparing the three correlation functions alone, on the initial search