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. 4. B[OCRerr]levance Feedback r[OCRerr]ihe formulation of the optimal query corresponding to a particular set of documents has no direct implication on operational information retrieval, since the set of documents in question is the object of the retrieval search. Thus there is no [OCRerr]riori way to generate an optimal request, since the ability to do so would eliminate the need for retrieval. [OCRerr] kind of qircularity suggests a strong analogy to feedback control theory. Consider therefore a sequence of retrieval operations which start with an initial query q . A modified query q1 is to be produced based on the original 0 output, such that q[OCRerr]1 is a better approximation to the optimal query for this user than q . Let' the user specify which df' the ret'r'i'eV"e4 0 documents (resulting' from the sear oh using q[OCRerr]0) are relevant and which are not. This information constitutes an error signal to the retrieval system. On the basis of the error and the original input, it is then possible to produce a modified query (new command input) such that the retrieval output will be closer to what.the'[OCRerr]'us'er...desires, or.[OCRerr][OCRerr]such that the modified query will be closer to the' optimal query for this us'er1s needs. The effectiveness of this process will depend on how good the initial query is, and on how fast the. process' of iteration converges to the optimal request. On the basis of the formulation of request optimality, we then seek a procedure for using the relevance feedback from an initial retrieval operation to produce an improved query. Let q 0 be the original retrieval request, and let the results of the retrieval