The automatic evaluation of text-based assessment items, such as short answers or essays, is an open and important research challenge. In this paper, we compare several features for the classification of short open-ended responses to questions related to a large first-year health sciences course. These features include a) traditional n-gram models; b) entity URIs (Uniform Resource Identifier) and c) entity mentions extracted using a semantic annotation API; d) entity mention embeddings based on GloVe, and e) entity URI embeddings extracted from Wikipedia. These features are used in combination with classification algorithms to discriminate correct answers from incorrect ones. Our results show that, on average, n-gram features performed the best in terms of precision and entity mentions in terms of f1-score. Similarly, in terms of accuracy, entity mentions and n-gram features performed the best. Finally, features based on dense vector representations such as entity embeddings and mention embeddings obtained the best f1-score for predicting correct answers. [For the full proceedings, see ED593090.]