Leveraging Deep Learning for Automated Generative Grading of Science Subject-Based Structured Questions

The increasing dependence on online education highlights the need for scalable and efficient assessment tools. Manual grading of structured science questions is time-consuming and subjective, leading to inefficiencies and inconsistencies that compromise assessment fairness and reliability. This research addresses this challenge by developing and testing a new deep learning model for automated generative grading. It employs a hybrid Seq2Seq architecture with BERT and ResNet encoders alongside a GRU decoder to analyze both text and images within questions. The model was carefully evaluated using token-level metrics such as Accuracy, Precision, and F1-Score, along with advanced generative metrics like Corpus BLEU Score and Average BERT Similarity Score. The results reveal a notable contrast: the model achieved a low Corpus BLEU score of 4.34, indicating limited exact syntactic matches with reference answers, but excelled with an Average BERT Similarity of 0.9944, demonstrating strong semantic and contextual understanding. This key finding shows the model's capacity to comprehend the meaning and relevance of marking schemes despite varied wording. The study confirms the model's ability to process and interpret both textual and visual data to generate relevant, meaningful outputs. Overall, this research validates the concept, offering a robust architectural framework and evaluation method for new AI-powered educational tools. The findings reject the null hypothesis, indicating the model significantly improves grading accuracy through enhanced semantic understanding and scalability, providing a promising solution for educators.