ISR11 A Modified Two-Level Search Algorithm Using Request Clustering chapter V. R. Lesser Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. vII-26 documents retrieved in a ruii search does not necessarily imply a high racall for the 3 document[OCRerr] as measured by R . Table 3 sh[OCRerr][OCRerr] the values T for [OCRerr]` [OCRerr]T and [OCRerr] based on requests which retrieve more than 3 documents. It is felt that this chan[OCRerr]e in the range of the number of documents to be retrieved does not invalidate the conclusions [OCRerr]ince in an actual informa- tion retrieval system, the majority of the user requests retrieve more than 3 documents. The results exhibited in Table 3 clearly indicate that if the collection of ne[OCRerr], queries introduced into the system is similar to the collection of previous queries introduced into the system, then the modi- fied t[OCRerr]o-level search scheme is more efficient than the normal two-level scarch~. The cases 7-12 in Table 1 indicate, as expected, that if a query is not simalar to a subset of previous queries, the normal two-level search is more effective than the modified two-level search. It is believed that due to an unanticipated error in the experimental procedures this difference in search effectiveness is unduly increased. The set of non-associated documents was found to be approximately equal in size to the set of associated documents, and [OCRerr]jas partitioned into only 2 categories. It is felt in retrospect that this -[OCRerr]s a mistake, and that if the set of non- associated documents had been clustered into [OCRerr] categories, the search effectiveness of the modified two-level search scheme for the test collec- tion of the last 10 queries would have been greatly improved. Four categories divides this set of non-associated documents into subsets of documents of the same size as the clusters of associated documents. The collection of 35 test queries is converted into 231 satisfied requests of which 160 requests were produced from the first 25 queries.