ARCHIVES

The increasing popularity and acceptance of CBE (Computer-based Examination) systems by examination bodies and institutions of learning globally for evaluation of multi-choice answers have unveiled the need for similar systems for descriptive answers evaluation. In this study, Word2Vec Neural Network architecture, specifically CBOW (Continuous Bag-of-Words) was used to build an ML model for descriptive answers evaluation. The corpus for training the model were of two categories namely: model answers to exam questions and answers by examinees. Five hundred and twenty five (525) instances of answers from 105 examinees in 5 questions from the course, Introduction to computing were used. 80% of the dataset were used for model training while 20% was for the validation process. Also, answers from examinees were manually graded by three experts in the area and used to compare and evaluate the model performances in two instances of 40% and 50% as pass marks. In the case of 40% pass mark, F1 scores of 0.92 for Class 1 (i.e. pass class) and 0.91 for Class 0 (fail class) were obtained which reflect a strong, balanced ability to identify actual cases (high recall) and avoid false alarms (high precision) for both classes. With respect to 50% pass mark, F1 scores of 0.82 and 1.0 were obtained for failed and pass classes respectively indicating that the model is slightly better at identifying and classifying instances of pass class compared to fail class. The model overall accuracy of 0.9 shows its potentials to correctly evaluate descriptive answers.