In recent years, more and more large-scale English tests begin to use the automatic scoring system.Therefore, the research of this system is of great value.In this paper, we first extract a lot of features according to English writing guide.Then we use autoencoder and discretization algorithm to learn a different representation of features.Finally, we use a hierarchical multinomial model to output the final scores of articles.Experimental results indicate that this method not only achieves great performance for those essays of the same topic, but also shows good robustness when predicts essays of different topics.Compared with the traditional automatic score method, our approach achieves higher than 9.7% in term of Pearson Correlation Coefficient, with good practical values.