Text Classification Using Web Corpora and EM Algorithms

Abstract

The insufficiency and irrelevancy of training corpora is always the main task to overcome while doing text classification. This paper proposes a Web-based text classification approach to train a text classifier without the pre-request of labeled training data. Under the assumption that each class of concern is associated with several relevant concept classes, the approach first applies a greedy EM algorithm to find a proper number of concept clusters for each class, via clustering the documents retrieved by sending the class name itself to Web search engines. It then retrieves more training data through the keywords generated from the clusters and set the initial parameters of the text classifier. It further refines the initial classifier by an augmented EM algorithm. Experimental results have shown the great potential of the proposed approach in creating text classifiers without the pre-request of labeled training data.