MONO91 Automatic Assignment Indexing Techniques chapter Mary Elizabeth Stevens National Bureau of Standards Discriminant coefficients were then computed at both the major and minor levels for all words occurring in the sample items falling into one of the 20 groups in accordance with the formula: "The discriminant coefficient is: = [OCRerr]n (P[OCRerr][OCRerr] - 13 3 P i3. Where: m P.. = f. I [OCRerr] 13 13 1 and The relative frequency of the ith word in the jth category. n = 1 - [OCRerr] P. 13 n ii 3 The mean relative frequency [OCRerr]er category of the ith word. 11 These coefficients are used both to set up threshold values to determine which words should be used in the assignment formulas and to assign weighting factors to the words themselves. The results of the experiments to date are based on 83 items from the "reference set" which were not used as source items. For 63 items, 78 percent were correctly classified at the level of a single major category (e.g. , "Programming", `,11ardware Design") and also correctly classified at a single subcategory level, (e.g. , "Program- ming Languages", "Semiconductor Devices"). The 20 remaining items were classified to one major category with an accuracy of 95 percent and to two minor level subdivisions with accuracies of 60 percent and 75 percent. Additional investigations were made on the effects of using a discrimination threshold to eliminate insignificant words from consideration and on the use of weighting factors in the assignment calculations. 4.5 SADSACT Stevens and Urban at the National Bureau of Standards (1963 E 569, 570]) have also explored an automatic indexing technique that uses, as in the experiments of Williams, a teaching sample or refer[OCRerr]nce set Qf previously indexed items to form patterns of word and index-term assignment associations. However, there are much less formal require- ments for computing correlation coefficients and no consideration is required of either 1/ Williams 1963[642], p. 163. 98