Linguistic Bias in Teacher Evaluations: Sentiment Analysis and Counterfactual Explanations Reveal Systematic Inequities Correlated with Perceived Course Difficulty

Resumen

Teacher evaluation platforms such as RateMyProfessor increasingly inform educational decisions, yet their textual reviews may carry biases unrelated to teaching quality. We investigate whether perceived course difficulty introduces linguistic bias in student comments, analysing 45,835 evaluations (from 49,923 after length filtering) of 260 professors. Using VADER and TextBlob, comments for high-difficulty courses show significantly lower sentiment (Delta = 0.499, Cohen's d = 0.893, p < 0.001) and a fourfold rise in negative language (39.6% vs. 9.3%); an intra-professor analysis confirms the effect is contextual (99.2% of professors, Wilcoxon p < 0.001). A convergent-validity check against a transformer model (Twitter-RoBERTa) corroborates these results (\rho = 0.745). Finally, a bias classifier (F1 = 0.80) with Diverse Counterfactual Explanations (DiCE) reveals which linguistic changes would reverse penalization, contributing to explainable provider fairness in educational text.