Data Augmentation by Rubrics for Short Answer Grading

Short Answer Grading (SAG) is the task of scoring students’ answers for applications such as examinations or e-learning. Most existing SAG systems predict scores based only on answers, and critical evaluation criteria rubrics are ignored, which plays a crucial role in evaluating real-world situations. In this paper, we propose semi-supervised method to train neural model. We extract keyphrases that highly related from rubrics. Weights words calculated attention labels instead manually annotated labels, span-wise alignments between keyphrases. Only with weighed used supervision. evaluate proposed model two analytical assessment tasks analytic score prediction justification identification. Analytic predicting given answer prompt, Justification identification involves identifying cue student each score. Our experimental results demonstrate both performance grading improved by integrating training, especially low-resource setting.