Intelligent Assignment Grading System Based on BERT for Computer Science Course

This study is the research outcome of the project "Exploration of Student-Centered Teaching Reform and AI Technology Application in Dynamic Website Design Courses" led by the author at South China Agricultural University, ZhuJiang College.This research focuses on the development and application of an intelligent assignment grading system based on the BERT model in computer science course teaching. Traditional assignment grading methods often fail to provide timely and personalized feedback, particularly when answering student queries. To address this issue, this paper proposes an assignment grading system based on BERT's bidirectional training capabilities, which can deeply understand the context and generate vectors for both student answers and standard answers. It uses cosine similarity to provide precise grading results. This study introduces innovations in system architecture, model selection, dataset construction, and the training process to meet the specific needs of grading computer science course assignments.The research methodology includes dataset construction, using 1,000 carefully selected answer pairs covering foundational topics such as dynamic website design, data structures, programming languages, and databases. By fine-tuning the BERT model, utilizing dynamic masking, Adam optimizer, and learning rate decay, we achieve optimal generalization results. In addition, a user feedback mechanism has been developed to continuously optimize the model based on the accuracy of the grading results in response to student feedback.The study designs an MVC architecture-based system, integrating Java, Tomcat, FastAPI, Uvicorn, and MySQL. Experimental results show that this system outperforms traditional methods in performance and efficiency, with an average cosine similarity of 0.88, Spearman rank correlation of 0.94, mean squared error (MSE) of 0.05, and an F1 score of 0.92. These results demonstrate that BERT significantly outperforms other models in understanding the semantic similarity between student answers and standard answers, providing more accurate grading predictions. Error analysis reveals that there is still room for improvement in handling complex technical terms and context understanding.The innovation of this study lies in providing a powerful and scalable tool to improve the quality of computer science education, offering an effective solution for educators to enhance teaching interaction, improve efficiency, and provide personalized learning support. Insights gained from this research support the widespread application of AI-driven assignment grading systems in academic environments and pave the way for the future development of automated educational support technologies.