MONO91 Automatic Assignment Indexing Techniques chapter Mary Elizabeth Stevens National Bureau of Standards Modifications to derivative indexing techniques that tend toward normalizations of terminology and word usage, and increasingly sophisticated proposals for machine use of syntactic, semantic, and contextual clues hold out the promise of transition to more truly `1subject" indexing and to automatic assignment indexing systems. 4. AUTOMATIC ASSIGNMENT INDEXING TECHNIQUES Answers to the question of whether indexing by machine is possible are actually dependent in part on how the question of whether what can be achieved by machine is or is not properly termed "indexing" is answered. If "indexing" is defined as being more than the mere extraction of words from titles, abstracts, or text, then automatic derivative indexing, even when augmented by various modifications, normalizations, and editings, does not provide affirmative evidence. In the case of concept-oriented definitions of indexing, the question becomes one of whether or not automatic assignment indexing is possible. Experimental evidence suggesting that it is will be presented in this section. We should note first, however, that just as there are differences of opinion as to what "indexing" means so there are similar differences, with respect to whether or not it represents concepts rather than extracted words. There are also a number of conflict- ing definitions of what is meant by "indexing" in contradistinction to "classifying". For some, the latter difference is related to questions of the number of labels or surrogates assigned to a single item to represent its subject contents, ranging from the assignment of a single subject category in a classification scheme involving mutually exclusive classes to the assignment of a number of terms or descriptor each standing for one of a number of aspects of the subject. For our purposes, however, we shall regard both the case of indexing with a number of descriptors and that of classifying to a single category or subject heading as being within the province of automatic assignment indexing, re- serving the term "automatic classification" for the case where the machine is used to establish the classification or categorization scheme itself. Actual experiments in automatic assignment indexing by Borko, Borko and Bernick, Maron, Salton, Stevens and Urban, Swanson, and Williams will be discussed briefly below. These discussions are generally in chronological order with respect to first reporting of results, except that the Salton-Lesk-Storm work reflects a somewhat dif- ferent principle of assignment from the methods using clue word approaches and it is therefore described after these others have been discussed. Some of the similarities and differences between the various methods are then indicated. A brief final subsection covers related assignment indexing proposals for which experimental data is not available or has not as yet been reported in the literature. 4.1 Swanson and Later Work at Thompson Ramo-Wooldridge Research on fully automatic indexing as well as on full text searching and retrieval at the Ramo-Wooldridge Corporation has been reported as being under way at least as early as the spring of 1958. 1/ As described elsewhere in this report, experiments in search and retrieval based upon full natural language text had used as test items short articles in the field of nuclear physics. In additional experiments representing a preliminary "clue word" approach to possibilities for automatic indexing procedures, some of this same material was used. 1/ National Science Foundation's CR&D rept. no. 2, [430], p. 32.