García J.A., Fajardo-López D., Mesa J., Amaya-Gómez R., 2026, Automatic Classification of Industrial Accident Causes using NLP: A Case Study in eMARS, Chemical Engineering Transactions, 126, 295-300.
Previous accidental events serve to reinforce Process Safety Management (PSM); however, some accidental databases suffer from unstructured reports that require largely manual data extraction, making it challenging to analyse a significant number of events. Artificial Intelligence (AI) can offer a scalable pathway to extract actionable knowledge from unstructured data. This study proposes a lightweight Natural Language Processing (NLP) pipeline to automatically classify the primary cause of major accidents using the European eMARS database. A total of 1,234 reports were parsed from PDF and structured by extracting the Accident Description and Causes sections; a six-class PSM taxonomy was operationalized, excluding undetermined cases from supervised learning. Two document representations were compared: an in-domain Word2Vec embedding (200-d) and Sentence-BERT (all-MiniLM-L6-v2, 384-d). Both representations were evaluated with a class-balanced linear Support Vector Machine (SVM) using stratified five-fold cross-validation and a hold-out test. Results show that Word2Vec+SVM provides the strongest and most stable baseline on the full labelled set, while SBERT performance improves markedly under higher label fidelity.
