A Cascaded Unsupervised–Supervised NLP Pipeline for Detecting Accusatory Language in Public Procurement

This study addresses the challenge of identifying procedural irregularities in public procurement by analyzing stakeholder comments, which are often underutilized in official oversight. The authors propose a cascaded, hybrid NLP pipeline designed to operate efficiently without the need for large-scale computational infrastructure. The methodology integrates unsupervised clustering to uncover latent patterns in noisy, user-generated text and supervised classification to pinpoint specific accusatory or whistleblowing-style comments. The pipeline evaluates three embedding strategies—Word2Vec, LLaMA, and RoBERTa—to determine the most effective representation for Spanish-language procurement data.

The research utilizes data from Ecuador's Sistema Oficial de Contratación Pública (SOCE). The pipeline begins with data wrangling to clean and filter noisy, user-generated comments. Following preprocessing, the authors generate semantic embeddings using mean pooling to create fixed-length representations. A key component of the approach is the use of Gaussian Mixture Models (GMMs) for clustering, which allows for soft assignment of data points. This unsupervised stage helps identify clusters enriched with accusatory content, which are then processed by a Random Forest classifier. The authors specifically address the issue of severe class imbalance, where accusatory comments represent only about 3% of the total dataset.

The experimental results demonstrate that a domain-adapted Word2Vec model outperforms more complex Transformer-based models (LLaMA and RoBERTa) for this specific task. The proposed hybrid architecture achieves a precision of 0.84 and a recall of 0.91. These results indicate that lightweight, domain-adapted NLP architectures are highly effective for risk identification in public procurement, providing a scalable solution for real-time auditing and monitoring in environments where computational resources may be limited.