ISR11 Information Analysis and Dictionary Construction chapter G. Salton M. E. Lesk Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. precision, the dictionary without phrases is preferable. This result reflects the feeling, already expressed in connection with the null thesaurus, that the first few documents are best retrieved by the simplest possible methods, when the chances of erroneous analysis are smallest. The statistical phrase procedure, as well as the regular thesaurus look-up, may always generate an occasional concept which is in error. Such concepts may affect the retrieval results, thus depressing precision. On the other hand, the increasingly more sophisticated text analysis which becomes possible through the phrase detection procedure is undoubtedly responsible for retrieving at least some documents which cannot be brought to the su[OCRerr]face by other simpler rr.ethods. This accounts for the beneficial effect of all well-built dictionaries in improving * the recall performance, usually at a loss in precision. The observed usefulness of synonym and phrase dictionaries raises the important question of how such dictionaries are best prepared. This question is examined in more detail in the next part. 5. Automatic Thesaurus Construction Under normal circumstances, the task of constructing a subject dictio- nary for a given topic area is one which demands many skills, including also a great deal of persistence and tenacity. It is not usually enough to be a subject expert in a given area, but training is also normally expected in linguistics and philosophy. Furthermore, since the task is of large proportions, a committee is often appointed which thrashes out controversial questions and eventually produces a suggested standard * The search results exhibited in this report for documents and dictionaries in the computer literature have been confirmed for other subject areas, including aeronautical engineering and documentation, also processed with the SMART programs.