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. 3-17 produce a modified query which is positioned centrally with respect to the relevant documents while maintaining maximum distance from the nonrelevant documents. [OCRerr]his is possible, however, only in so far as the index images of the relevant set are differentiable from those of the nonrelevant set. In this context it is possible that the information needs of a user might be best satisfied by a multiple rather than a single search request. This would be the case, for example, if useful references happened to be mappe.d&by..the.'[OCRerr]index transformation into several distinct regions of the index space. Since the user in general has no a priori means of determining whether he should use a single or a multiple search (other than his own intuition,) it is of interest to consider automatic means for generating multiple searches. Assume, for example, that the relevant set ? identified by a user after an initial retrieval operation contains document images sufficiently seperatei so as to be considered only slightly related. [OCRerr]igure 3.3 shows an example in two dimensions. Under the circum- stances portrayed the relevance feedback adjustment algorithm is not useful since, in fact, there is no single vector close to both relevant document images. This suggests that useful information can be derived by measuring the degree of association among the elements of the relevant subset identified by the user. Such information is contained in the document-document correlation matrix which character- izes this subset. Consider, for example, the situation described `by the