The detection of corruption-related indicators within unstructured, textual procurement data remains a complex task due to linguistic ambiguity, contextual variation and domain-specific terminology. This study presents a comparative evaluation of three transformer-based Natural Language Processing (NLP) architectures (BERT-base-uncased, RoBERTa-base and DeBERTa-v3-base) for automated corruption risk indicator detection in procurement texts coming from heterogeneous sources. A unified dataset is constructed by linking unstructured technical documentation with structured procurement outcomes, enabling an outcome-driven risk labeling strategy. Performance evaluation is conducted through different metrics, including precision, recall, F1-score and ROC-AUC, complemented by explainability analysis using Integrated Gradients. The results demonstrate a clear performance progression and highlight the comparative strengths of the evaluated architectures. Overall, this study highlights the potential of contextual transformer models to support scalable, transparent and operational anti-corruption monitoring systems.
Peppes et al. (Sat,) studied this question.