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'-7 and 0 is used as the relevance weighting factor of a document for which the relevance is not kno[OCRerr]m. However, the relevance weighting factors can be used to assign weights other than 1 to the relevant documents. In parti- cular, the weights of the relevant documents can be defined as their correlations, thus giving the relevant documents with higher correlation more weight in the modification process. This method is alsQ used to assign weights to the relevant documents in this investigation. Negative relevance weighting factors can be employed in the perturbation of the query. Essentially, these negative factors indicate that the documents are ??irrelevantl rather than the ndn-committal judgment that the relevance is not kno[OCRerr]m. Since the relevance judgments available do not include any `irrelevantt indications, it is not feasible to sizrrul ate fully such judgments by a user. But some use of negative relevance weighting factors can be helpful when, after modification of the query, documents of unnho[OCRerr]m relevance are still retrieved vTith high correlations. Figures 3 and 4 show the effect of applying negative relevance weighting factors to a query for which no additional relevant documents are retrieved after the first modification. Documents 1102, 302, and 109 are ranked 2, 3, and 4 for both the original and updated queries but are of unnnown relevance. Updating the modified query [OCRerr]dth negative relevance weighting factors yields a query which retrieves three relevant documents, a significant improvement. Further updating, using the normal 0 and 1 relevance weighting factors, eventually produces a query [OCRerr][OCRerr]hich retrieves all five relevant documents. The improvement resulting from the use of negative relevance weighting factors suggests that the following heuristic method would prove useful.