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]-25 an explanation of how the concept weights are derived.) [OCRerr]he first part of the algorithm implements equation [OCRerr] directly and results in a new query image as shown in part (d) of FigLire [OCRerr].5. [OCRerr]he appli- cation of a screening process to this vector results in a final modified query image as shown in part (e). [OCRerr]he screening process is designed to eliminate any. negative components in the modified query image, as well as to reduce the positive nonzero components to those most likely to be useful. *[OCRerr]his latter feature is incorporated since the statistical evidence implicit in the user!s relevance judgments may represent a relatively small ssmple. [OCRerr]oncepts which are retained after screening either a) occur in the original query, or b) occur in at least half of the relevant documents identified in addition to being more frequent in the relevant set than in the nonrelevant set. [OCRerr]he screening algorithm thus serves to prevent the modified query from becoming too special- ized to those relevant documents identified in the initial retrieval operation. In addition, reducing the number of nonzero components in the[OCRerr]modified query image[OCRerr]provides increased efficiency. With fewer components the modified search request requires less storage space and can be correlated with reference document images with fewer operations. [OCRerr]egative components in a modified query image represent properties in the index space which are more significant among the nonrelevant documents retrieved by the user's original search request, than among the retrieved relevant documents. In principle then, there are no conceptual difficulties in allowing such negative weights. [OCRerr]his is in contrast to generating property vector index representations of