Automatic Grading of CITO Mathematics Tests

Abstract

This thesis attempts to discover whether it is possible to do automatic grading of CITO mathematics tests using Optical Character Recognition (OCR) methods, among others. It is part of a cooperation between three students, where this thesis focuses on the extraction and labeling of questions, as well as classifying the multiple choice questions. It turns out that with the methods discussed in this thesis, it is not possible to extract questions and classify multiple choice questions with high accuracy. Also, there are some robustness concerns discussed in this thesis. In this research, the classification of multiple choice questions turns out to be most successful with a convolving measure of dissimilarity as defined in this paper, in combination with a decision tree classifier, to an accuracy of 99.49%. However, due to an accuracy of the preprocessing of 91.14%, the accuracy of the total process is insufficient to be applicable, since it is 90.68%. In conclusion, it is not advised for CITO to implement the methods discussed in this research for the automatic grading of multiple choice questions.