ISR11 Relevance Feedback in an Information Retrieval System chapter W. Riddle T. Horwitz R. Dietz Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. v'-8 When no relevant documents are retrieved, the two most highly correlated documents are given -l relevance weighting factors. However, without !?irrelev[OCRerr][OCRerr]celI judgments the use of this heuristic method is at best an arbitrary procedure, and it should be noted that in some cases the negative relevance weighting factors are not helpful. In particular, the results sho[OCRerr]m in Figure 5 indicate that the use of too many negative relevance weighting factors overly perturbs the query vector, so that the modified query is unable to retrieve more relevant documents[OCRerr]. The use of negative relevance weighting factors implies that some of the concept weights in the query vector will bec[OCRerr]e negative. Negative concept weights have no significance in a term-document matrix, since an indexer cannot indicate the non-appearance of a certain concept. It is plausible, however, to have negative weights in a query vector after it has been perturbed because the relevance feedback information can properly indicate that certain concepts are irrelevant. D) Determination of the Value of a The parameter a is necessary to control the manner and the strength of the perturbation to the query vector caused by the relevance feedback informati6n. a may[OCRerr] affect the modification in one of three ways. First, if a is increased with every iteration, the new relevance information will have the same weight as the sum of all the previously-gathered relevance informations. For example, if one relevant document with concept vector (10101) is retrieved initially by a query with concept vector (10100), the modified query (when a equals 1) will have a concept vector (20201). If this modified query, in turn, retrieves relevant documents having