Recomputation of class relevance scores for improving text classification

Abstract

In the text classification task, bag-of-word representation causes a critical problem when the prediction powers for a few words are estimated terribly inaccurately because of the lack of the training documents. In this paper, we propose recomputation of class relenvace scores based on the similarities among the classes for improving text classification. Through the experiments using two different baseline classifiers and two different test data, we prove that our proposed method consistently outperforms the traditional text classification strategy. © Springer-Verlag 2004.