FATE: a three-stage method for arithmetical exercise correction

Qipeng Zhu, Zhuoyan Luo, Shipeng Zhu, Qi Jing, Zihang Xu, Hui Xue

Abstract: As the number of primary students rapidly rises, the highly repetitive task of correcting arithmetical exercises consumes much time for teachers and hinders them from concentrating more on the growth of students. To reduce the workload of teachers, arithmetical exercise correction (AEC) is proposed to automatically detect, recognize and correct various arithmetical exercises in the workbook. However, two crucial issues need to be addressed since the research in this field is still immature, i.e., accurate detection of the arithmetic exercise with various structures and the effective recognition of long-size exercise. In this paper, we propose a three-stage method dubbed as FATE, to correct arithmetical exercises in an end-to-end manner. Specifically, we apply the anchor-free model with a feature pyramid network and constraint of center-ness to avoid the redundant bounding boxes. On the other hand, we employ a transformer-based framework with contrastive learning to extract global symbol information and generate corresponding sequences. Finally, we design a series of rule-based templates to correct the generated sequence based on the unique features of each type of arithmetical exercises, respectively. Extensive experiments demonstrate that our method yields the detection average precision of 96.8%, the recognition accuracy of 92.3% and the $$\mathrm {F_{1}}$$ F 1 score of 91.2% in spotting experiment on the public dataset, which outperforms the state-of-the-art method.