MONO91 Automatic Assignment Indexing Techniques chapter Mary Elizabeth Stevens National Bureau of Standards the theoretical probabilities of word occurrence by category or of discrimination Co- efficients and thresholds. Instead, the technique involves ad hoc statistical associations between the words occurring in the title and in the abstract of a sample item and the descriptors previously assigned to that item. A master selection-word vocabulary is thus built up where each word is listed in terms of the frequencies of its co-occurrence with each of the descriptors with which it has co-occurred, regardless of whether or not such prior a6sociations are either revelant or significant. No attempt has as yet been made to "purge" the resulting association lists. Instead, reliance is placed on the patterns of multiple word usage and of redundancy of words used in titles and cited titles of new items to minimize the effects of irrelevant or accidental prior word-descriptor associations and to enhance the significant ones. The SADSACT method (for "Self Assigned Descriptors from Self and Cited Titles") proceeds with the assumption, which it shares with the arguments for citation indexing previously discussed, that the literature references cited by an author are indicative of the subject content or contents of his paper. 1/ For the automatic indexing of new items, their titles and the titles of up to ten bibliographic references cited are keystroked, con- verted to punched cards, and fed to the computer. This input material is run against the master vocabulary to obtain for each input word which matches a vocabulary word a "descriptor-selection score" for each of the descriptors previously associated with that word. These scores are summed up for all words and at an appropriate cutting level those descriptors having the highest scores are assigned to the new item. Preliminary results based on the titles and cited titles of items that were "source items" in the sense that their titles and abstracts had been used in the teaching sample were reported at the NATO Advanced Study Institute on Automatic Document Analysis held in Venice in July, 1963. For 30 items drawn from such subject fields as computer technology, information selection and retrieval, mathematical logic, pattern recognition, and operations research, all of which had previously been indexed by ASTIA personnel in 1960, the machine assigned 64.8 percent of the descriptors previously assigned. Sub- sequent tests on genuinely new items, however, resulted in a drop to only 48.2 percent "hit" accuracy. These "new" item results were also evaluated by having several representative users of the collection analyze the test items and assign descriptors to them from a list of the descriptors available to the machine. The extent to which the descriptors assigned by machine were also independently chosen by one or more of these indexers was then checked. In general, the fewer descriptors assigned by the machine, the better was the human agreement, ranging from 47.4 percent overall in the case where the machine had assigned twelve descriptors to each item to 76% agreement where the machine assigned only one. In particular, for ten items which were analyzed by five different indexers, the chances that one or more would also select the machine's first choice (highest scoring) descriptor averaged 90 percent. 4.6 Assignment Indexing from Citation Data Certain phases in the program of investigation of information selection and retrieval problems at the Harvard Computation Laboratory have been mentioned previously. The work of Storm and of Lesk and Storm on the use of first-noun-occurrences as selection clues for both automatic indexing and abstracting was discussed in connection with tech- niques for improved derivative indexing. The studies on citation indexing have included, as noted, experiments to assign indexing terms to a new document by finding the indexing 1/ If necessary or desirable, however, abstracts or portions of text can be used in addition to or in lieu of the cited titles. 99