Assesing Constructed Responses with Explainable Natural Language Processing