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. vil-lo search requests is quite commonplace in document retrieval. In addition, the words grouped in the thesaurus dictionary may display hierarchical relationships; for example, concept 22 of the Cran-2 Thesaurus-3 groups both "algebra" and "arithmetic" with the generic notion of "mathematics" (Fig. 2). Hierarchy dictionaries tested have been constructed by struc- turing the thesaurus concepts themselves, rather than by going back to the separate words or word stems. Hierarchies have been manually con- structed only for the IRE Computer Science collection, and descriptions of the methods used in their construction have appeared in [2,3,5,8,13,14]. Discussion and evaluation of procedures for automatically producing hier- archies by co-occurrence statistics is also not considered here (see [2,15]). 5. Retrieval Performance Results A) Thesaurus Dictionaries Performance comparisons are normally made between the stem and thesaurus dictionarie6, and a series of comparisons using normalized recall and precision are given in Figs. 5, 6, and 7. The results in Fig. 5 are all based on the cosine numeric matching function, and it may be seen that even with different document input lengths,the thesaurus dictionaries are nearly always[OCRerr] superior to stem. Reasonable explanations can be found for two main exceptions, since the Cran-l Thesaurus-i was made without the use of any of the construction rules; furthermore, it was based on the indexing only, omitting many words which appeared in the abstracts. The second exception is the ADI "Hastie" Thesaurus-SAl which was made by semi-automatic procedures and was known to contain unsatisfactory groupings. Figs. 6 and 7 give, respectively, some results based on the Cosine Logical and Overlap Logical