To improve computation efficiency, an enhanced EM algorithm based on self-training named STEM is proposed. In the E-step of each iteration, the unlabeled sample, whose class can be predicted by the current intermediate classifier with the most confidence, is moved to the labeled set and used in the M-step to train the next intermediate classifier. Therefore the mechanism of self-training by inter-result employing is introduced. Experimentation on text classification indicates that STEM outperforms EM in classification accuracy most of the time and improves the learning efficiency by reducing iterations.