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-37 individual request preferences based on Fig. 25, and for every test, there exists for at least one of the three evaluation ranks, a case where the thesaurus is to be preferred to the phrases or hierarchy. Fig. 27 shows more clearly how, using results based on the average rank of the first relevant recovered (to simulate a high precision user) phrases are not superior to thesaurus in any of the results, but the hierarchy relations do give a very small improvement. 6. Summary of Results Since the volume of data tends to obscure overall findings, the results of performance comparisons are enumerated separately. In order to facilitate reference to the thesaurus results, the 28 comparisons made are referred to by number; the normalized evaluation results may be found in Figs 5, 6, 7, and 10, and the precision versus recall results in Figs. 8, 9, 10, 11, 12, and 17. The thesaurus results may be summarized as follows: 1. The best thesaurus dictionaries give a performance superior to the stem dictionary on the average if other system parameters are set to their optimum (using abstracts or text, together with the cosine numeric matching function). This is seen in five cases (comparisons 2, 10, 13, 16, and 18); the superiority of thesaurus is least marked in the Cran-l collection. 2. The Cran-l collection is unique in that the suffix [OCRerr]SI dictionary performs a little better than stem; a com- parison of thesaurus with suffix `5' shows the suffix [OCRerr]51 to be a little superior at the high precision end of the curve (comparison 28, Fig. 10).