Correction Method for Character Recognition of Handwritten Answers in Multiple Formats for Short Essay e-Learning Systems

We are aiming to develop an e-learning system for improving Japanese language proficiency.This system not only provides instant scoring and advice in response to learners' answers but also automatically generates questions to offer a wide range of problems across various fields.In this paper, as the first step for this purpose, we aim to accurately extract the content described in handwritten essays.When dealing with handwritten characters, sufficient accuracy cannot be obtained with OCR software alone due to noise, distortion, and idiosyncratic handwriting.Therefore, we first train a neural network for character recognition using a handwritten character database, and then divide the essay images into single characters to determine the most likely character candidates.Furthermore, we treat character identification as a fill-in-the-blank problem for sentences and use BERT's Masked Language Model task to determine characters that form natural sentences.Applying this method to handwritten short essays written in multiple formats has enabled more accurate character extraction than before.