A Hybrid Qualitative and Quantitative Approach for Automatic Short Answer Grading Using Classification Algorithms

Assessment plays a very important role in the teaching-learning process. Recently proposed techniques for short answer grading are either Qualitative or Quantitative. Qualitative methods use classification and give scores in an incomprehensible way like correct, incorrect, partially correct, etc., and Quantitative methods use Regression and MSE(Mean Square Error) which assigns scores like 2.3,4.2, etc. requires a learning method to convert those values to integers. In this paper, a hybrid Qualitative Quantitative approach is proposed which is based on machine learning that relies on Multi-class classification where each category of marks is considered as a class. This can be easily converted to qualitative or quantitative systems without additional learning methods. This system follows a model answer-based method where various similarities are extracted from both model and student answers including statistical, Bag of Words, TFIDF, Latent Semantic Analysis, and Neural sentence embedding based Infersent. Using these features various classification models are designed including K Nearest Neighbor, Naïve Bayes, Decision tree, SVM, Random Forest, and XGBoost for various test sizes of the data and cross-validations. The models are tested on a dataset that consists of one question, 80 answers, and a model answer. In total, 72 experiments are conducted, demonstrating that the proposed hybrid approach can be effectively applied to the Automatic Short Answer Grading (ASAG) task.