IRS13 Thesaurus, Phrase and Hierarchy Dictionaries 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. VII-14 matching functions. These again display a superiority for thesaurus in the expected cases, except for the "ADI Text Thesaurus-i Overlap Logical" result. Precision versus recall graphs are used to repeat the most impor- tant comparisons, with IRE-3 in Fig. 8, Cran-l Abstracts in Figs. 9 and 10, Cran-l Indexing in Fig. 11, and ADI Abstracts and Text in Fig. 12. Thes- aurus works better than stem in all cases except in that of the first Cranfield version. Fig. 10 compares thesaurus with suffix `S', since on Cran-l, stem is not as good as suffix `5' (both at the high precision end and between 0.65 and 0.85 recall suffix `5' is superior). The figures do show that the thesaurus dictionaries are superior by a much greater amount for both IRE-3 and ADI than for Cran-l. Using all comparisons of stem with the final versions of a given thesaurus, the data in Fig. 13 shows how the individual requests favor each dictionary using the normalized measures as an indicator of merit. Between 54% and 82% of the requests favor the thesaurus, with IRE-3 showing the clearest advantage for the thesaurus, ADI next, and Cran-1 the least advantage; this agrees with the precision-recall graphs. Fig. 14 gives a further performance comparison using the average rank position of the first, second and last ranked relevant documents, to simulate high precision and high recall needs. Unexpectedly, the Cran-l and ADI comparisons show the thesaurus to be more effective in meeting the high precision needs than stem, whereas in IRE-3, the thesaurus worsens high precision performance where one relevant only is required. However, high recall needs are seen to be dramatically