An algorithm for the calculation of exact term discrimination values

Term discrimination values have been suggested as an effective means for the selection and weighting of index terms in automatic document retrieval systems. This paper reports an algorithm for the calculation of term discrimination values that is sufficiently fast in operation to permit the use of exact values, rather than the approximate values studied in previous work. Evidence is presented to show that the relationship between term discrimination and term frequency is crucially dependent upon the type of inter-document similarity measure that is used for the calculation of the discrimination values.