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-9 concept vectors (10101) and (11100), then these vectors should have their concept weights multiplied by 2 to make their concept weights comparable to those of the modified query. Second, CL can be kept at a constant value for all the iterations, which would cause all types of relevance information to be treated in the same manner, regardless of when they are obtained in the process. Finally, if CL is decreased with every iteration, more emphasis will be placed on the relevance information obtained during the early iterations. Figure 6 shows the results obtained in sample runs using various progressions and magnitudes of CL. In general, the use of small fractional values of CL yields, after two or three iterations, the results which can be obtained in one iteration by the use of a larger, integral value of CL. For this reason, only integer values of[OCRerr]CL are used in fLirther investigations. The decreasing progression of CL is also discarded because it is not significantly better than the other progressions. Also, a decreasing progression is illogical, since if the iteration process is converging on some area of n-space, then more weight should be given to relevance information obtained later in the process. Thus two strategies involving CL are used in the final investigations. In the first, CL is set initially to 1, and is increased by 1 at each succeeding iteration. In the second, [OCRerr] is held constantly equal to 2 (the value 2 is used to represent a typical value rather than an optimal one). However, it should be noted that whenever the correlations of the relevant documents are used as the relevance weighting factors, CL is held constantly equal to 1; this is denoted in the figures by the label `ALPHA = CORRELATIONS'1.