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. VI-' VI. Relevance Feedback in an.Information Retrieval System W. Riddle, T. Horwitz, and R. Dietz 1. Introduction An information retrieval system must be judged in the end by the user population, and therefore it should be designed primarily to satisfy the user's needs. However, the implementation of an efficient system, measured in terms of the amount of relevant material retrieved for the average user, is difficult to implement with presently available computing equipment because of their batch processing mode of operation. Typical turn-around times preclude efficient man-machine interaction, and a user must generally be satisfied with the results of a single search. However, the use of only a single search may not produce adequate results. The reduction of search time by the use of two-level searches, which match the query first against the centroid vectors of document clusters and then against the individual documents in highly correlated clusters, may cause some relevant documents to be lost.L2] Moreover, some indexing terms may be interpreted differently by users with different fields of interest. In fact, users unfamiliar with the indexing terms employed may formulate queries [OCRerr]Alich, after translation to indexing concepts, may not adequately represent the user'a requirements. In addition, the result of most correlation procedures presently used to match documents and search requests depends on the relative positions of the queries and documents in the n-space determined by the indexing terms; but the resulting correlations do not necessarily reflect the relevance of