Towards Automated Evaluation of Handwritten Theoretical and Programming Answer Scripts using OCR and NLP

This is a challenging task especially when we have to evaluate handwritten answer scripts consisting of theoretical and programming-based answers. This paper focuses on using Optical character recognition (OCR) and Natural Language Processing (NLP) to automate the grading process. Azure Computer Vision Read API achieved average cross-similarity of 91% over pretrained Tesseract OCR, which was only trained on 37% of manually written scripts, based on analysis of 150 answers. The scores from the evaluation framework were 2.56/5.0 and 3.55/8.0 for theoretical and programming answers respectively. Despite showing potential for automatic evaluation, there are difficulties with acknowledging varied hand-writing styles along with complicated programming method.