What school communities say—and what it predicts: A discourse-based model of dropout risk

School dropout remains a critical educational and societal issue, particularly in contexts where systemic challenges hinder student retention. Traditional methods of identifying at-risk schools rely on quantitative indicators such as attendance records, academic performance, or socio-demographic data, often offering only retrospective insights. This study explores a novel, scalable approach for early identification of schools with high dropout risk, based on the automated analysis of stakeholder language rather than numeric data. Using interviews with principals, teachers, and parents from 105 Romanian schools, we applied natural language processing and machine learning techniques to classify schools into three dropout risk categories (low, medium, high), as defined by a validated risk index developed by the World Bank and adopted by the Romanian Ministry of Education. A machine learning classifier achieved an overall accuracy of 80% and strong discrimination between risk levels, as indicated by an area under the receiver operating characteristic curve (AUC-ROC) of.91. Distinct linguistic patterns proved effective in distinguishing schools with different risk levels. These findings suggest that stakeholder discourse provides valuable insights into school climate and functioning. By incorporating language-based analysis into school monitoring systems, policymakers and educational leaders can identify emerging risks early and design timely, targeted interventions to prevent dropout.