Reduction of Training Noises for Text Classifiers

Abstract

Automatic text classification (TC) is essential for the archiving and retrieval of texts, which are main ways of recording information and expertise. Previous studies thus have developed many text classifiers. They often employed training texts to build the classifiers, and showed that the classifiers had good performance in various application domains. However, as the training texts are often inevitably unsound or incomplete in practice, they often contain many terms not related to the categories of interest. Such terms are actually training noises in classifier training, and hence can deteriorate the performance of the classifiers. Reduction of the training noises is thus essential. It is also quite challenging as training texts are unsound or incomplete. In this paper, we develop a technique TNR (Training Noise Reduction) to remove the possible training noises so that the performance of the classifiers can be further improved. Given a training text d of a category c, TNR identifies a sequence of consecutive terms (in d) as the noises if the terms are not strongly related to c. A case study on the classification of Chinese texts of disease information shows that TNR can improve a Support Vector Machine (SVM) classifier, which is a state-of-the-art classifier in TC. The contribution is of significance to the further enhancement of existing text classifiers.