MONO91 Automatic Assignment Indexing Techniques chapter Mary Elizabeth Stevens National Bureau of Standards In these additional experiments, 27 articles in the nuclear physics subject area were included in a corpus of 100 articles, the remainder covering a variety of topics. Fre- quency counts of word occurrences for the physics material were obtained and the 12 most frequent words that were judged to be discriminatory for the subject were selected. The hypothesis was then tested, that if any document pertained to nuclear physics it would contain at least two of these words. Retrieval was achieved for 25 of the 27 documents and the two "irrelevant'1 documents also retrieved did include information at least peri- pherally related to the `subject. It was thus evident that the retrieval effectiveness of automatic recognition of nuclear physics subject material in the general collection was considerably greater than the average effectiveness of retrieving responses to the highly specific search questions in nuclear physics that had been used in the full text searching experiments (Swanson, 1961 [586]). This second set of experiments provided a transition from the full text searching work, which if it can be considered indexing at all is obviously derivative indexing, to work in the application of an automatic assignment indexing method to 1, 200 newspaper clippings (Swanson, 1962 [ 584], 1963 [580]). These were brief news items for which machine-readable texts in the form of punched paper tape were available. Thesaurus- groups of words likely to be associated with each of 20 to 24 subject headings were first compiled on the basis of human analysis of 1,000 or more representative items. These word groups were further screened so that no word appeared in more than one group and so that each word retained should be uniquely indicative of the particular subject category. In the machine assignment procedure, subsequently, if a word occurs that belongs to a particular thesaurus group, the corresponding subject heading is assigned to the item in which that word occurs. Results achieved with this technique appear to be highly promising, at least for this type of material. Swanson reports as follows: "Approximately 1,200 brief news items were classified into 20 nonhierarchical subject categories, both by a human and a machine procedure. Each item was assigned on the average to about four categories. The results of the two processes were compared. With the human process as a standard, the machine missed only seven percent of the correct subject assignments and made a number of irrelevant assignments equal to about 17 percent of the total. Nearly 40 per- cent of the automatic subject assignments judged finally to be correct were missed by the human catalogers. While this accomplishment is actually due to the extensive human effort to compiling, organizing, and pruning of the uniquely indivative word lists, it is pointed out that this intellectual effort and the programming tasks need to be done only "once and for all". It is further pointed out that garbles or misspellings in the input text do not appear to affect the procedure, there being enough redundancy in the messages so that even if one or two clue words are missed, others will be present. 3/ 1/ Swanson, 1962 L584], p.468. 2/ Ibid, p.469. Swanson, 1963 [ 580], p.5 92