ISR11 Summary summary Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. judgments are then used automatically to generate a new search request more indicative of user need and preference. This relevance feedback process, first described in detail in section III of report ISR-lO, is shown to be extremely effective, and to provide continued improvements in search effectiveness through at least three feedback iterations. Section VII by V Re Lesser, and section Ix by J. D. Broffitt, H. L. Morgan, and J. V. Soden describe variations of the nnilti-level search techniques originally introduced in section IV of repoft ISR-lO. These techniques drastically reduce the number of comparisons needed between incoming search requests and stored documents by grouping the stored item[OCRerr], and comparing search requests at first only with a typical item for each group. Individual comparisons are then made only for those documents included in highly scoring groups. The effectiveness of a variety of document grouping procedures used as part of a multi-level search process is evaluated in section DC. Section VII, on the other hand, describes an e[OCRerr]eriment in which requests previously processed by the system are grouped, rather than documents, and new inc[OCRerr]ning requests are first compared with these request clusters. The search procedure for a new request is then made to depend on the results obtained with similar requests processed by the system at some previous time. The results of section VII indicate that this heuristic process is useful in reducing search time when the requests to be processed fall into definite patterns, as they may be expected to do in an operational situation. The last section, number X by M. Lesk contains system design specifi- cations for a SM[OCRerr]T type system operating in a time-sharing environment xvi