This paper proposes a similarity-based grading system for handwritten answers to short-answer questions in Japanese workbooks. While previous studies have predominantly focused on models trained specifically for individual questions, this study utilizes a similarity-based approach, in which both the student’s answer and the model answer are input into the model, allowing it to learn the semantic differences between them. This approach enables the model to handle multiple questions within a single training session, thereby improving grading accuracy even with a limited dataset and allowing the system to effectively score answers to unseen questions. We utilized three types of foundational models: a Japanese pre-trained BERT model, a Japanese pre-trained Sentence-BERT model, and ChatGPT. For training and evaluation, answer data from Japanese workbooks used by junior high school students was employed. The experimental results demonstrated that similarity-based scoring using a model answer is effective for grading short-answer questions in Japanese workbooks. In addition, the model achieved satisfactory results even when applied to text automatically recognized from handwritten answers to unseen questions, demonstrating its practical applicability in real-world educational settings.