Identifying top performing TF*IDF classifiers using the CNN corpus

Abstract

TF*IDF (term frequency times inverse document frequency) is a common metric used to automatically discover keywords in documents for use in classification and other text processing applications. We are interested in determining whether these measures can help in classifying documents. There are multiple ways to define TF*IDF, but there has been no real attempt to determine the value of these different forms. We explore a large family of 112 TF*IDF measures (corresponding to an a priori estimate of 20 degrees of freedom among these measures) applied to 588 CNN articles belonging in 12 classes such as Business, Sport, and World. We postulate that at least some sets of these measures must be effective for classification. The goal is to use a set of TF*IDF measures that best match the a priori classifications by CNN. We also show that by combining the results of a few wellperforming TF*IDF measures can increase classification results.