Spelling Check for Handwritten Homework and Exams

Zhixiang Xiong,Xinbo Zhao,Xiaolin Li,Rui Wang,Yin Wang

DOI: https://doi.org/10.1109/icaibd57115.2023.10206183

2023-01-01

Abstract:Word spelling is a fundamental skill at the center of K-12 English education. While there is a lot of academic research and commercial software on handwriting recognition, little effort has been made to recognize misspelled words, which is an important step toward AI grading. One significant challenge is the infinite possibility of wrong spelling but extremely limited training data compared to correct words. We observe that the current popular OCR algorithms such as CRNN and CNN-Ngram suffer prediction shift phenomenon. To this end, we propose a spelling checking framework consisting of self-supervised pre-training, supervised fine-tuning, and post-processing. To address the long-tailed balancing problem, we propose a novel data generation and a novel dataloader. Experimental results on the misspelling dataset demonstrate the effectiveness of the proposed framework, and the generated and augmented data both contribute to improving the performance.