In this paper, a discourse-based method that merges semantic and syntactic models for automated essay assessment is proposed. The approach combines shallow linguistic features and discourse patterns in order to predict an essay's score by using decision trees regression techniques. Unlike current approaches, our method directly measures an essay coherence by using corpus-based semantics and text centering techniques so as to determine discourse patterns underlying high-quality essays when compared with human assessed essays. Experiments using standard datasets showed that the proposed discourse-based approach outperformed traditional shallow features-based methods.