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. reference [OCRerr] documents, where a negative weight would imply that one could measure the degree to which a certain attribute was lacking. For this reason, and because the recognition of negations or exceptions in search requests would require a high degree of syntactic sophistica- tion, the simulation system (s[OCRerr][OCRerr]) does not have any facilities for processing query or document vector images with negative components. Such components ar[OCRerr]ising from the modification algorithm (see Figure 5.5 (d)) were eliminated then, to preserve compatibility with the simulation system. Allowing negative components in a modified version of a user1s search request would amount to an effective increase in the quantization of the index space. it may be postulated then that this would lead to improved performance. An experimental investigation into this possibility was not feasible, because it would have required sub- stantial changes in the simulation system. Figure [OCRerr].6 (a) shows the initial portion of the retrieved out- put generated for the original query 111-R Indexing11 which is described in Figure [OCRerr].5. [OCRerr]he relevance feedback used in this example consisted in identifying the initial two relevant and two nonrelevant documents in the retrieved list. For reference, the titles of all the relevant documents for this search request are provided in [OCRerr]able [OCRerr].2. Figure 3.6 (b) compares the retrieval results of the original and modified queries with respect to this full set of six relevant documents which are in the reference collection. [OCRerr]ote that the modification has sub- stantially improved the performance with[OCRerr]respect to three out of the four relevant documents not originally identified by the user. Figure 3.7 compares the correlation distributions of the original and relevance