Text Categorization Using Latent Topics as Additional Features

Abstract

In feature selection of text categorization, there are methods which handle word sense disambiguation by extracting synonymy and polysemy among words in documents. One of the methods utilizes latent topics underlying documents by using a topic model. PLSA and LDA have been proposed as representative models. In this paper, two features which include both TF-IDF and the latent topic values which extracted automatically from topic models were utilized for text categorization using AdaBoost. Then, the performances were compared with the ones of only TF-IDF features. As a result, this study evaluates effectiveness and weakness of the augmented features.