Academic dissatisfaction in teacher evaluation: a natural language processing approach and multimodal semantic analysis for the generation of pedagogical evidence

Abstract

This study converts student voices into actionable pedagogical evidence through a Spanish-language NLP pipeline that integrates Likert-type evaluation items and open-ended comments from 411 master's students in business sciences. The dataset contained ten positively oriented item indicators and 411 records; all item responses were valid, no respondent was excluded, and 410 substantive comments were available for textual analysis. The pipeline normalized item responses, cleaned Spanish text, estimated polarity, subjectivity, and basic affective categories, extracted bigrams, modeled six topics using Latent Dirichlet Allocation (LDA; K = 6), and computed two indicators: the textual climate score (S_text) and the Multimodal Course Climate Aggregate (MCCA). Item reliability was acceptable for an exploratory institutional instrument (Cronbach's alpha = 0.636). The mean item score was 3.63/5, while the normalized item score was 65.77/100. The textual score averaged 56.68, and the fused MCCA averaged 59.79. Positive comments emphasized professional usefulness, clarity, examples, and applicability; negative comments concentrated on lack of focus, ambiguous tasks, feedback timing, and class management. The fused indicator showed that textual evidence adds decision value beyond Likert scores by identifying pedagogical frictions that are not fully captured by item averages. This study contributes to educational data science by offering a replicable Spanish-language NLP pipeline that transforms qualitative student feedback into interpretable, actionable evidence for institutional improvement and teacher professional development.