Design of English composition automatic grading and teacher-assisted evaluation platform using generative language model

Abstract

This study designs and implements a mixed evaluation platform for English composition that combines the Generative Language Model (GLMs) and teachers' wisdom, aiming at solving the problems of low efficiency, strong subjectivity and lack of feedback in traditional grading. The platform adopts a "two-layer evaluation framework": at the bottom, GPT-4, BERT and other multi-models are integrated to deeply analyze the four dimensions of grammar, logic, content and style, and output interpretable scores, error labels and dynamic confidence; The upper layer introduces the confidence warning mechanism, and only pushes the low confidence composition to the teacher for review, thus realizing the closed loop of man-machine cooperation. The experiment analyzed 300 argumentative essays written by college students, comparing pure human, pure AI, and hybrid processing modes. Results showed the hybrid mode reduced average processing time by 74% compared to manual processing, lowered root mean square error (RMSE) to 3.5, achieved a correlation coefficient of 0.85 with the gold standard, and recorded an intraclass correlation coefficient (ICC) of 0.88—significantly outperforming the other two groups. Both teacher and student satisfaction scores exceeded 4.4 points. The research proves that the platform can significantly reduce the burden on teachers, improve the immediacy and pertinence of feedback while ensuring the accuracy and fairness of grading, and provide a feasible path for large-scale personalized writing teaching.