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. ~. [OCRerr]-42 C. Convergence The performance improvement which results from a query optimization produced by relevance feedback mbdification is a function of the quality of the initial query, the degree of association of the index images of the relevant documents, and the amount of feedback. To investigate the influence of the latter parameters, some additional experiments were conducted. Figure [OCRerr].14 shows the retrieval results as a functIon of the amount of feedback for the query ??:[OCRerr]-Indexing?; and Figure 5.15 for the query ?l[OCRerr][OCRerr] [OCRerr]atlaflgl?. The document-document correlation matrix for those[OCRerr]documents relevant to the query "[OCRerr]- Indexing11 are shown in Figure 5.16,. Thus the rapid improvement obtained even with a small amount of feedback can be attributed to the fact t'[OCRerr]hat the members of the relevant set are all closely associated. In the case of the query ??N[OCRerr] [OCRerr]atlang1', this is not true, and the document-document correlation of the first five relevant documents, retrieved by the'original query indicates this. The relevance judgments for this query were made. assuming a very general point of view. In this case it might be of use to produce multiple modified queries by seeking clusters in the relevant set. A possible partition based on the document-document correlations is shown in Figure 5.16, and this partition was used to generate two modified q[OCRerr]eries following equations * (s.ii) and (5.12).' The retrieval results for these two modifications are shown in Figure 5.17. This figure illustrates how each of these * queries is useful in retrieving some relevant' documents. The fact that some of the relevant documents have low correlations with both of the