Clustering based Two-Stage Text Classification Requiring Minimal Training Data

Xue Zhang, Wang-xin Xiao

Clustering has been employed to expand training data in some semi-supervised learning methods. Clustering based methods are based on the assumption that the learned clusters under the guidance of initial training data can somewhat characterize the underlying distribution of the data set. However, our experiments show that whether such assumption holds is based on both the separability of the considered data set and the size of the training data set. It is often violated on data set of bad separability, especially when the initial training data are too few. In this case, clustering based methods would perform worse. In this paper, we propose a clustering based two-stage text classification approach to address the above problem. In the first stage, labeled and unlabeled data are first clustered with the guidance of the labeled data. Then a self-training style clustering strategy is used to iteratively expand the training data under the guidance of an oracle or expert. At the second stage, discriminative classifiers can subsequently be trained with the expanded labeled data set. Unlike other clustering based methods, the proposed clustering strategy can effectively cope with data of bad separability. Furthermore, our proposed framework converts the challenging problem of sparsely labeled text classification into a supervised one, therefore, supervised classification models, e.g. SVM, can be applied, and techniques proposed for supervised learning can be used to further improve the classification accuracy, such as feature selection, sampling methods and data editing or noise filtering. Our experimental results demonstrated the effectiveness of our proposed approach especially when the size of the training data set is very small.