Automate Descriptive Answer Grading using Reference based Models

Global universities are establishing institutional setups that offer a hybrid format of education. The next step of education is to maintain quality and flexibility, such as providing the option to convert online courses such as Massive Open Online Courses (MOOCS) to course credits. However, several universities are reluctant to completely transition to online-based education due to poor digital experience in educational tools. The available evaluation tools such as Multiple-choice answers (MCQ) aren't able to evaluate students holistically. In this study, research work aims for an improvised reference-based approach (utilizing student and reference answers) that evaluates descriptive answers with the Siamese architecture- Roberta bi-encoder based transformer models for Automated Short Answer Grading (ASAG). The architecture was designed considering ASAG tasks constrained to feasible compute resources. The research work presents the competitive performance of the models, further improvised with finetuning and hyperparameter optimization process on the benchmark SemEval-2013 2way task dataset.