AI-Driven Learning Analytics for Student Evaluations of Teaching: Detecting Score–Text Asymmetry and Heterogeneity with LDA and RoBERTa

Student Evaluation of Teaching (SET) provides both quantitative ratings and open-ended textual feedback for assessing teaching quality, yet these two information sources are often analyzed separately. This study investigates whether numerical ratings and textual evaluations provide fully aligned information and how their relationship varies across teaching contexts. Using 42,724 cleaned evaluation records from a university in China, including 44,713 sub-comments for sentiment inference and a final analytical sample of 1342 instructors, we develop an artificial intelligence-driven educational data mining framework that integrates Latent Dirichlet Allocation (LDA) topic modeling with RoBERTa-based sentiment analysis. LDA is used to identify latent pedagogical themes, while RoBERTa estimates continuous sentiment scores from textual comments. To support joint analysis of the two modalities, we further define the Score–Emotion Deviation (D) as an applied operational measure of the numerical discrepancy between normalized evaluation scores and model-derived sentiment scores. The empirical analysis shows that numerical ratings and textual sentiment are highly concentrated near their upper ranges but nevertheless exhibit localized score–text discrepancies. The distributions of the operational deviation measure also vary across academic titles, disciplines, and class sizes. By combining thematic, sentiment, and numerical-rating information within a unified analytical framework, this study extends multimodal educational data mining for SET analysis and provides complementary evidence that can support more contextualized diagnosis and interpretation of teaching quality.