MONO91 NIST Monograph 91: Automatic Indexing: A State-of-the-Art Report Other Potentially Related Research chapter Mary Elizabeth Stevens National Bureau of Standards "Automatic index term association techniques are needed to improve the recall of relevant information, to enable indexers and requestors to use language in a more natural manner, and to enable retrieval of relevant messages which are described by different index terms than those used in the inquiry." 1' For the most part, the work to date has been directed to "associative retrieval" of messages limited to single sentences of English text, and to the search phases of a pro- posed system. In the case of a corpus consisting of 230 sentences from a single text, a partially automatic indexing method was used. The text was first processed against a modified version of the Harvard Multipath Syntactic Analysis computer program and the resulting analyses were manually screened to select a unique, correct analysis for each sentence. Next, approximately 500 words, those that had been marked "noun" by the syntactic analyzer, were listed out and these in turn were manually screened to provide an "inclusion list" of 273 words. Sentences were then "indexed" with respect to which of these selected words they contained. Word associations were computed both in terms of co-occurrence within a sentence and of co-occurrence in syntactic structures. Retrieval tests were then applied using both computer programs and the analog device, and evaluations were made on the basis of examining sentences selected in order of machine-ranked relevance and of comparisons of word lists associated with a given search term against association lists for another term picked at random. It is noted that, "although quantitative conclusions cannot be drawn", the results support the conclusion that: "Items retrieved due to automatically-generated associations tend to be more rele- vant than is explainable on a chance basis." 2/ The "request reformulation" retrieval program has also been used to generate term profiles from a collection of approximately 10, 000 documents (previously indexed with at least 6 terms from a selective term vocabulary of 1,000 terms) which have then been comparedagainstusts provided in the entries for corresponding terms in the Thesaurus of ASTIA Descriptors, Second Edition. The machine-produced association lists, at least for those words occurring relatively frequently in the corpus, appear to give thesaurus entries that are extensive, specific, and intuitively acceptable, and of high quality, especiallywith respect to listings of synonyms as well as factually related words 3/ The development of the ACORN (Associative Content Retrieval Network) devices has provided additional tools for testing and display[OCRerr](l962 [229], 1963 [227, 304]). These devices are networks of passive resistance elements. Each word or index term and each sentence (240 by 230 in ACORN-IV) are represented by terminals interconnected by resistors with conductance equal to the connection strength, and with "leak" resistors 1/ 2/ 3/ Giuliano 1962 [228], p. 10. Giuliano et al, 1963 [230], p. 47. Ibid, pp. 57-58. 125