Explainable Sentiment-Based Reputational Risk Modelling in Ugandan Megaprojects

Reputational risk reflected in digital public discourse can significantly influence stakeholder perceptions, public trust, and decision-making surrounding government megaprojects. Existing computational approaches primarily focus on sentiment classification, offering limited capability for transparent and explainable modelling of reputational risk in public sector governance. This study proposes an explainable sentiment-informed model for identifying textual indicators of discourse-based reputational risk in digital discourse surrounding Ugandan government megaprojects. A corpus of 25,000 publicly available records from online news and social media sources was processed using a natural language processing pipeline, with weakly supervised labels subsequently validated through human validation of 2000 records by three independent annotators. A Soft Voting Ensemble integrating Random Forest, Gradient Boosting, Logistic Regression, and Support Vector Machine classifiers was evaluated using stratified, independent human, and temporal validation. Under the primary stratified evaluation, the model achieved 96.0% accuracy, 0.9592 weighted F1, 0.9327 Macro-F1, and 0.9079 MCC. Independent human validation yielded 72.5% accuracy and 0.5952 Macro-F1, while temporal validation yielded 89.4% accuracy and 0.6108 Macro-F1. SHAP analysis of the Soft Voting Ensemble identified governance and project-related textual features contributing to model predictions. The findings demonstrate strong primary performance while also showing that performance is lower under independent human and temporal validation, supporting the use of the approach as an interpretable decision support framework for identifying discourse-based reputational risk signals rather than as a validated predictor of realized reputational outcomes.