Automated Short Answer Scoring Using an Ensemble of Neural Networks and Latent Semantic Analysis Classifiers

We introduce a short answer scoring engine made up of an ensemble of deep neural networks and a Latent Semantic Analysis-based model to score short constructed responses for a large suite of questions from a national assessment program. We evaluate the performance of the engine and show that the engine achieves above-human-level performance on a large set of items. Items are scored using 2-point and 3-point holistic rubrics. We outline the items, data, handscoring methods, engine, and results. We also provide an overview of performance key student groups including: gender, ethnicity, English language proficiency, disability status, and economically disadvantaged status.