Operational NLP/ML for Contractual Risk Screening and Response Planning in Road Infrastructure Projects

This research presents an NLP- and ML-based system for automated contractual risk screening and risk response recommendation in road infrastructure projects, demonstrated on Peru’s New Central Highway case. Contractual and technical documents in Spanish were transformed into a sentence-level corpus, enabling consistent annotation and downstream classification. A road-project Risk Breakdown Structure (RBS) was used to structure risk categories and to support interpretable outputs for practitioners. Sentences were encoded using multilingual Sentence-BERT embeddings and classified with a logistic regression model calibrated to produce probabilistic scores, so that alerts can be tuned through an explicit threshold and integrated into a review workflow. The system is designed for human-in-the-loop use: it prioritizes candidate risk sentences and returns supporting information for review and response planning. Operational performance was assessed on contractual documents. The model flagged 47 sentences as risks, of which expert review confirmed 33, resulting in an operational precision (PPV) of 70.2% (33/47; 95% Wilson CI: 56.0%–81.3%). This corresponds to 0.49 alerts per page (0.34 true positives and 0.15 false positives per page), which is compatible with real review capacity. Processing throughput reached 126.25 words/s, enabling rapid triage of long documents and shifting effort from full reading to targeted validation of alerts. The recommended response plans were evaluated by experts using a Likert scale and achieved a mean score of 3.9/5, exceeding the acceptance threshold of 3.5. Overall, the evidence supports the use of NLP/ML as an operational layer for faster contractual risk screening and more consistent response planning in large road projects.