Background The development of intelligent education has made student performance prediction based on student comments and course information an important direction for achieving personalized teaching. Student comments not only reflect knowledge mastery but also convey emotional attitudes and learning states. However, most existing approaches fail to capture the salient emotional cues and structural patterns in comments or to leverage course-level information effectively, leading to limited predictive performance and interpretability. Methods To address these issues, we propose a position-aware sentiment enhancement framework (PSEF-Net) designed to capture the coupling between sentiment intensity and structural positioning. Specifically, PSEF-Net introduces three core mechanisms: (1) A emotion-guided structural fusion mechanism, which constructs a heterogeneous attention path to emphasize emotionally polarized expressions and mitigate the dilution of key feedback by neutral context; (2) A position-aware residual attention gate, which leverages local structural shifts and cross-modal discrepancy representations to dynamically enhance the model’s sensitivity to sentiment transitions and structural discontinuities; (3) A dual-path context modeling module, which integrates global and local semantic flows to compress redundant features and improve the representation of course-level structure. Results Experimental results on two real-world datasets from MOOC and NetEase Cloud Classroom demonstrate that PSEF-Net consistently outperforms all baselines. Compared with the second-best model Multi-dimensional Student Performance Prediction (MSPP), PSEF-Net reduces mean absolute error (MAE) by 16.4% and root mean square error (RMSE) by 9.4% on the MOOC dataset, and decreases MAE by 18.0% and RMSE by 9.9% on the NetEase dataset. All differences are statistically significant ( p < 0.05). Visualization analyses further confirm that PSEF-Net effectively focuses on key feedback cues across diverse comment structures, highlighting its potential for interpretable student performance prediction.
Shi et al. (Thu,) studied this question.