Automatic Scoring System for Digitized Handwritten Answers

We propose a system for automatic and semiautomatic scoring of handwritten answers in Japanese, English, and mathematics, assuming the use of an electronic pen and paper. When quick scoring is prioritized, the system directly returns automatically scored results to answerers without requiring the scorers’ verification. When reliable scoring is required, the system shows the results to scorers for confirmation and correction before being disclosed to the answerers. The system comprises handwriting recognizers using deep neural networks and text-based scorers to determine whether an answer is correct, incorrect, or rejected. Question-specific scoring methods and rules (rubrics) (defined by the scorers) are used to score the answers consistently. To evaluate the system, handwritten answers in the three subjects were collected from at least 50 students for all grades of elementary and junior high schools. Scoring experiments showed that the average human scoring rate (percentage of answers rejected plus those falsely negatively scored and claimed) was under 30%, indicating that approximately 70% of the scorers’ workload could be reduced. The average risk scoring rate (ratio of false positives that may not be reported) was