IRS13 Test Environment chapter E. M. Keen Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. 1-36 extreme types of user need, those of high precision and those of high recall. The high recall comparison may be carried out by comparing the average rank position taken up by the last relevant document, and Fig. 17 shows that in all tests the specific requests are very clearly superior. The high precision comparison may be carried out by computing an average rank position for the first two relevant documents, and Fig. 18 shows that here the general requests give a superior performance. These results use the stem dictionary. Since the thesaurus dictionaries normally produce a superior performancei some change in me[OCRerr]it between the specific and general requests might result for the thesaurus runs. The same data is therefore repeated for the thesaurus runs in Figs. 19 and 20, where it is again seen that for a high recall need the specific requests are best, and for a high precision need the general requests are best. It is quite likely that in an operational situation, users wanting high recall would tend to pose specific requests, and users wanting high precision would tend to pose general requests. But there could certainly be exceptions to this, and the suggested correlation might not exist at all. The most disturbing part of this finding is that the specific requests, which were thought to be the better ones for retrieval purposes1 do not perform very well f[OCRerr]r high precision users, although with the thesaurus dictionaries in use the gap between specific and general requests on ADI and Cran-l (Fig. 20) is narrower than with the stem dictionaries (Fig. 18). Further work in this area req[OCRerr]ires better procedures for distinguishing specific and general requests, since the use of request generality in a small test collection is not intended to produce any fundamental division that