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-20 result the normalized overall measure is better for cosine than for co-occurrence, and better for co-occurrence than for simple vector matching. This can be seen in Table BlO where all searches are compared for each correlation function, with the correlation functions ranked in order of merit according to the normalized overall score achieved by each. The cosine correlation function works the best on the initial search and also on the updated searches using the alpha correlation strategy. The co-occurrence correlation function is the best for the updated searches using the increasing and constant alpha strategies. With one exception, the sL[OCRerr]ple vector matching correlation function performs the worst on all searches. [OCRerr]valuation of the relevance feedback methods requires examination of the effectiveness of the updated searches, and the tables given show only two cases where updates result in a drop in performance. This occurs when the simple vector matching correlation function and alpha correlation strategy is used, and Table B9 shows that the second and third updates had a performance progressively worse than the first update. A single request contributed largely to the result, since in request [OCRerr]9 the normalized overall measure dropped from l.[OCRerr]677 to a.6887 after the second update, and dropped further to 0.5989 after the third update. However all the combinations of correlation functions and alpha strategies resulted in a considerable improvement in performance with the updated searches compared to the initial search. An order of merit of the nine combinations t[OCRerr]sted is given for the three updated searches in Table BII, where merit is based on the increase in the normalized overall measure achieved by each update compared with