Evidence-Based Teaching Evaluation: Integrating Psychometrics, NLP, and Machine Learning

Abstract

This article presents an intelligent and explainable system for evaluating teaching performance based on student surveys that include Likert-scale items and open-ended questions. The solution integrates a reproducible pipeline for data ingestion, cleaning, integration, modeling, and deployment; psychometric methods to estimate reliability, latent structure, and invariance across periods; NLP techniques to analyze sentiment and extract topics and aspects; and supervised models with calibration and subgroup fairness checks. The results reveal a stable five-dimension structure and a global index that is comparable across periods, while open-text responses provide complementary signals regarding clarity, punctuality, and the timeliness of feedback. Global and local explanations make it possible to translate findings into actionable recommendations for each instructor and into dashboards with alerts. Implications for faculty development and curriculum design are discussed, along with limitations related to response bias and seasonality. Future work is proposed around adaptive items, longitudinal analysis, and automated feedback