Deep Learning and Rule-Based Hybrid Model for Enhanced English Composition Scoring Using Attention Mechanisms and Graph Convolutional Networks

Abstract

Through the profound exploration conducted on AI technology in the field of education, early automatic scoring systems for English compositions have problems such as high misjudgment rate and low efficiency. To improve the efficiency, accuracy, and stability of the English composition grading model, a deep learning and manual rule-based English composition grading model was designed. The research extracted sequence features by introducing attention mechanisms, enhancing contextual correlation analysis, and aggregating global features through graph convolutional networks to extract high-order semantic relationships. Finally, a visual manual scoring rule was designed, which integrated deep semantic features and manual rule features through the Wide&Deep architecture to jointly optimize the scoring results. The experiment outcomes indicated that the accuracy recall curve area of the research method was 92.3%. In practical application testing, the highest group stability index of the research method was 0.07 in June. When faced with 600 concurrent requests, the average response time of the research method reached a stable value of 3.4 seconds. The outcomes above demonstrated that the English essay scoring model, which combines deep learning with manual rules as proposed by the research, exhibited excellent accuracy, speed, and stability. It effectively addressed the issues of a high misjudgment rate and low efficiency found in traditional scoring systems, thereby enhancing the model's reliability.