A hybrid natural language processing-ontology and machine learning framework for linguistic and knowledge-driven accident analysis in petrochemical occupational safety informatics

Traditional analysis of occupational accident reports remains manual, time‑consuming, and constrained by expert judgment. This study applies a hybrid NLP and machine learning framework that addresses both gaps by analyzing a 15‑year longitudinal dataset of 527 Persian‑language slips, trips, and falls (STF) accident records from a petrochemical complex. After preprocessing through a domain‑specific Persian pipeline (including character normalization, domain‑specific stopword removal, and morphological stemming), TF‑IDF vectors (200 features, n‑gram range 1–2) were reduced via Truncated SVD (10 components). The optimal cluster count was determined through a domain‑informed multi‑criterion sequential rule combining a domain minimum (K ≥ 4), a consensus stability threshold (≥ 0.70), and marginal Silhouette improvement (> 5%). K‑Means clustering (K = 4) was selected as optimal (Silhouette = 0.267, Davies–Bouldin = 2.007, Calinski–Harabasz = 74.3; bootstrap stability mean ARI = 0.880, 95% CI [0.546, 0.981], classified as STABLE). Applying the hybrid text-plus-structured feature framework, we identified four semantically distinct accident clusters from a feature matrix combining TF-IDF representations of Persian narratives with non-causal structured descriptors (accident type, accident outcome, and lost workdays): a critical fall-from-height cluster (C1, n = 134, 26.0%, mean = 184.4 lost days, IRR = 239.6 relative to reference, P < 0.001); a general lower-extremity fall and body-contact cluster (C0, n = 172, 33.4%, mean = 1.0 day); an trips cluster (C2, n = 99, 19.2%, mean = 1.1 days); and a slipperiness/surface hazard cluster (C3, n = 110, 21.4%, mean = 0.8 days). External validation against official cause labels confirmed statistically significant cluster-cause association (Cramér's V = 0.354, χ 2 = 78.5, n = 209, P < 0.001) without any causal label information entering the clustering model. Severity characteristics different significantly across the identified clusters (Kruskal–Wallis H = 62.003, df = 3, P = 2.19 × 10⁻ 13 , ε 2 = 0.115); however, because accident outcome and lost workdays were incorporated into the hybrid feature representation, these analyses are interpreted as descriptive cluster characterization rather than independent validation. These findings demonstrate that integrating Persian-language accident narratives with non-causal structured accident descriptors enables the discovery of meaningful occupational accident typologies without pre-labelled training data. The observed association between the hybrid clusters and independent official cause labels supports the practical utility of the proposed framework for evidence-based occupational safety management. Differences between narrative-derived cluster patterns and official cause codes may reflect variations in reporting content and coding practices rather than definitive causal misclassification and therefore warrant further investigation in prospective studies.