Ensuring safety compliance in construction remains a persistent challenge, with regulatory violations contributing to high injury and fatality rates. Although OSHA maintains extensive accident records, much of this data exists in unstructured narrative form, limiting its usefulness for systematic compliance analysis. Manual interpretation is time consuming and inconsistent, highlighting the need for scalable, automated approaches that transform narrative data into structured, regulation-aligned insights. We present a scalable framework for extracting and assessing safety-regulatory compliance from unstructured OSHA accident narratives. Our approach, designed in a two-step pipeline, first integrates taxonomy-guided prompt engineering with a language model to identify and structure OSHA-relevant risk attributes. Then, formalized logic rules encode OSHA regulatory thresholds and required protections to automatically assess violations against these regulatory standards. To validate the framework, we conducted a case study on fall-from-height (FFH) accidents, a leading cause of construction fatalities. The model achieved 92% classification accuracy and 0.95 recall in identifying FFH cases. Structured attribute extraction enabled rule-aligned compliance checks against 29 CFR 1926. Scrutiny of these assessments revealed critical insights: Among historical accidents records, violations (accidents that did not follow the regulation standards) were most commonly flagged in roof-related cases (49.2%), followed by scaffolds (38.3%) and ladders (3.8%). These findings emphasize both persistent compliance gaps in high-risk environments and demonstrate the framework’s capability to generate structured, interpretable insights from unstructured text. This approach offers a replicable, data-driven model for improving regulatory oversight and safety monitoring, demonstrated through FFH cases and designed for adaptation to other construction hazards.
Ray et al. (2026) studied this question.