Research on Semantic Understanding of Student Learning Behavior and Knowledge Mastery Prediction Based on Large Language Models
TL;DR
The proposed method provides a feasible pathway for intelligent diagnosis, precise feedback, and personalized learning support in Chinese composition education and indicates that semantic features can improve the accuracy and ordinal consistency of writing mastery prediction.
Abstract
As artificial intelligence is increasingly applied to Chinese composition education, writing assessment is shifting from single-score judgment to semantic understanding and ability diagnosis. Student essays not only present language expression outcomes, but also contain evidence of learning behavior, including topic understanding, structural organization, content development, logical coherence, and language use. To address this issue, this study uses the public Chinese Essay Dataset For Pre Training and selects Qwen2.5 7B Instruct as the core model to construct a method for semantic feature extraction and mastery state prediction. The model first extracts five types of features from essay content, essay type, and grade level, including topic understanding, structural completeness, content richness, logical coherence, and language expression. The original writing quality ratings are then mapped into three mastery states: low mastery, medium mastery, and high mastery. The experimental results show that, after semantic features are added, Accuracy increases from 0.684 to 0.731, Macro F1 increases from 0.672 to 0.725, QWK increases from 0.642 to 0.712, and RMSE decreases from 0.681 to 0.599. These results indicate that semantic features can improve the accuracy and ordinal consistency of writing mastery prediction. The proposed method provides a feasible pathway for intelligent diagnosis, precise feedback, and personalized learning support in Chinese composition education.
