A Hybrid Approach for Automated Short Answer Grading

With the widespread use of distance learning, technological developments have also been applied in the field of education. The need for accurate and efficient assessment methods for online exams has become even more apparent, especially with remote learning taking place during the pandemic. For a more efficient evaluation process, we propose a hybrid model of the Automatic Short Answer Grading (ASAG) system based on Bidirectional Encoder Representation of Transformers (BERT). The usage of novel state-of-the-art natural language processing (NLP) techniques in our model enhances the comprehension of text. Specifically, we employ a customized multi-head attention mechanism adapted with BERT, which enables reliable identification of semantic dependencies among words within a sentence and therefore contributes to the effectiveness and trustworthiness of the scoring system. We use a parallel connection of CNN layers in our proposed BERT based ASAG system instead of their serial connection and this usage improves the performance of the system. The proposed model is assessed using common datasets frequently used for ASAG related research projects. In this evaluation process, our model produces much better results compared to other systems available in the literature.