Maritime safety analysis increasingly depends on the ability to interpret large volumes of unstructured information.
In this study, an LLM-based extraction pipeline was developed to classify SAFETY4SEA incident reports according to EMSA categories and assess the reliability of each classification through a confidence-scoring system.
CHIRP Maritime works with the University of Leeds MSc Business Analytics and Decision Sciences programme, annually sponsoring one or more graduates to apply advanced analytical methods to maritime safety data. As part of this collaboration, graduate Carlos Roberto Almaraz Rosales analysed 8,573 incident articles published by SAFETY4SEA, exploring whether artificial intelligence could transform narrative maritime reports into structured data aligned with the European Maritime Safety Agency (EMSA) taxonomy.
"The results show that LLM models significantly outperform traditional NLP approaches. They extract richer descriptors – such as ship type, contributing factors and incident type – and convert narrative text into structured labels that support exploratory safety analysis."
Importantly, the study demonstrates that EMSA’s taxonomy can be applied to narrative sources, providing a standardised framework that improves comparability across heterogeneous reports.
Several operational patterns emerged from the structured dataset. Incident frequency displays a cyclical temporal behaviour, with alternating periods of decline and resurgence rather than a stable long‑term reduction.
The year 2019 presents a clear peak in incident frequency, although the dataset does not contain sufficient detail to determine the underlying cause.
A key analytical insight comes from the proportional matrices, which normalise incident counts by category. These matrices show that human‑related elements dominate in proportion, not only in absolute frequency.
Contributing factors linked to operational management, decision‑making and human performance consistently represent the highest proportional share across incident types and consequences.
This reinforces long‑standing evidence that human‑element dynamics remain central to maritime safety outcomes.
Spatial analysis reveals distinct hotspots along major commercial routes, particularly in Europe, the Mediterranean and the North Atlantic. Overall, the study demonstrates that AI‑assisted extraction can reveal meaningful operational and temporal trends from narrative maritime reports.
Standardising these outputs through EMSA’s taxonomy enhances their analytical value, supporting early trend detection, situational awareness and thematic analysis across the maritime domain.
Source: Safety4Sea
Robban Assafina is now on WhatsApp channel. Click Here
