In artificial intelligence, automatic short answer grading (ASAG) sparked the interest of many researchers. These Systems are used to evaluate the student's performance based on their intellectual and cognitive skills. Unfortunately, short answer grading poses various challenges to assess individual abilities. The first challenge, short sentences can be 10 to 20 words long. These short sentences include primary and secondary keywords, identifying such keywords is a challenge for syntactic processing. Furthermore, the order and relationship among the words affect the actual meaning of the answers. Answers provided by students may not be syntactically correct. The second challenge is different question types included in the short text:-factoid, descriptive, short, and long questions. Different question types influence the intent of the answer which affects the precision of grading accuracy. As a result, strategies for overcoming these problems in the assessment are required. In this study, we have proposed the attention-based deep learning model known as bidirectional encoder representation from a transformer (BERT) for the evaluation of short subjective answers. The measurement findings indicate that the BERT model is effective for automatic short answer grading. © 2022 IEEE.