Character Classification technology is the key link in OCR system. Most classification methods require abundant marked samples training to get classifier. In the real OCR application, there are so many classes, to label these samples are often waste time and energy, especially for unacquainted language, such as Arabic and Uygur, many characters are difficult to differentiate, so it even needs the help of professional guidance. This paper proposed a novel character classification with semi-supervised learning based on information entropy, introduced discrete event probability estimation theory of information entropy, active to select the optimization character samples, got the new parameters to train the classifier again, choose the most conducive to the classifier performance samples, iteration until the unlabeled samples set is empty. The experiment results show that this method achieves high performance in specific condition. © 2013 Trans Tech Publications Ltd, Switzerland.