ISR10 Search Request Formulation chapter Joseph John Rocchio Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. examples shown in Figures [OCRerr].6 through [OCRerr] was applied to the full set of search requests available. Average results comparing the performance of the original and relevance feedback modified queries are shown in the form of a precision vs. recall graph in Figure [OCRerr].1O. Since this means of exhibiting performance is based solely on the ranks of the relevant documents resulting from the query-document correlation process, it does not exhibit the true improvement which results from relevance feedback modification. [OCRerr]his may be appreciated from the example shown in Figure 5.11, which illustrates another of the sample queries. In this case both the original and the modified query exhibit ideal performance (i.e. the relevant documents are all ranked higher than any nonrelevant documents). [OCRerr]hus the precision vs. recall graphs for both cases are identical. The correlation distribution, however, indicates that, in fact, the modified query provides greater discrimination of the relevant set from the nonrelevant set. In any case the average results as shown in Figure 5.10 indicate that the modification algorithm results in substantial improvement. The request optimization procedure as illustrated by equation (5.10) can be used iteratively. The querist can, if he desires, provide evaluation information about the output generated by the first iteration and request that a second query modification 1 1 take place. If [OCRerr] and S are the relevant and nonrelevant subsets The method of construction of such recall-precision plots has previously been described in detail.4'5