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. 5-15 query vector q1. The table compares the correlations of q0 and q1 with the document vectors. The modifications to an initial query vector which are produced by the relevance feedback algorithm may receive the following interpretation: concepts, i.e. components of the initial query which are more significant in the document images of the relevant subset than in the nonrelevant' subset will be emphasized (i.e. increased in weight and visa-versa). Thus the weighting of the original query terms, derived from frequency counting, will be adjusted on the basis of the statistical evidence derived from the sample output for which the user provides relevance feedback. In addition, concepts not inc,luded in' the original query but which are also useful in differentiating the relevant from the nonrelevant documents will be added to the modified query image. Such concepts (components of the index space) can be' expected to be useful in retrieving other relevant documents not explicitly identified by the original query, since all' relevant documents (which can be successfully retrieved) must be sufficiently related to' be localized in some region of the index space. The ba;sic relation for request modification using relevance feedback (eq'uation (5.8) ) can be modified in various ways by `imposing additional constraints.' For' example, the weighting of the original query could be a `function of the amount of feedback such that with large amounts of feedback, `the original query has less effect on the resultant than with small amoun'ts'of feedback. Another constraint,