Teacher-Involved Automatic Characteristics Classification in Handwritten Math Answer Process

Abstract

In Japanese elementary and junior high schools, the spread of digital devices has led to the accumulation of handwriting math problem-solving logs. By visualizing these logs, teachers can observe students' thought processes and create opportunities for class-wide sharing and collaborative learning. However, observing each student’s work individually imposes a significant burden on teachers. Although various automatic classification methods for student answers have been proposed, they may not align with the teachers' actual needs. This study investigated the teachers' perspectives and identified the labels assumed for classroom sharing, along with the relative features of each problem type, to construct automatic classification models. Although machine-learning models achieve the highest classification performance, challenges for their practical implementation remain. Furthermore, through feature selection and classification using questionnaires and large language models (LLMs), it was suggested that modeling approaches that combine content and processes reflecting teachers' observation behaviors could improve classification performance. Classifications based on teachers’ insights and practices are particularly important when dealing with abstract data, such as handwriting problem-solving processes. This study highlights the potential for developing teacher-involved educational support systems.