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. L[OCRerr] riranzactions on Electronic Computers (iMarch - September, i[OCRerr][OCRerr]s) was a[OCRerr]il&ole for this purpose.[OCRerr]'4 Both the reference documents and each of the search requests which had been submitted at Harvard in the natural language were indexed using the S[OCRerr]ABT thesaurus. As the search requests had been used in a variety of previous retrieval experiments with this collection, relevance judgments for each query were also available, representing a full manual search through the complete reference collection. A full retrieval ordering of the source documents with respect to each sample query was available, consisting of the correlation of each search request index image with every reference document image. [OCRerr]rom the initial portion of the retrieved list (ordered by descending correlations), two sets of documents were specified: one containing relevant documents and one containing nonrelevant documents. The vector index images of each search request, and the images 0£ the documents in the two associated subsets were used as inputs to a [OCRerr]ortran program written to implement the query modification process. The output 0£ this program was a new query vector suitable for input to the SMART system. The modified query images could then be correlated with the reference collection and the results compared with those of the original search requests. Table 3.1 describes the program steps used to implement the relevance feedback query modification algorithm. [OCRerr]igure 3.5(a) shows the English text of a typical query. Figure 3.5(b) shows the explicit thesaurus mapping for the terms included in this query and part (c) shows the index image 0£ the query in vector form (see Appendix A fpr