A 'Cluster-then-Estimate' Natural Language Processing (NLP) Approach for Classifying Maritime Incident Severity Based on Textual Descriptions.

Textual incident description is a vital source for understanding the severity of maritime incident. In the maritime industry, relevant authorities and companies typically rely on manual methods to estimate incident severity based on textual descriptions. However, manual estimation is less efficient for assessing vessels' operational risk or managing historical incident archives, where a large volume of incidents is involved. Therefore, this study proposes a 'cluster-then-estimate' approach which uses Natural Language Processing (NLP) techniques to automatically estimate the severity level of incidents based upon their textual descriptions. In the proposed approach, Latent Dirichlet Allocation (LDA) is used to group the preprocessed textual descriptions into multiple clusters, with each cluster representing an incident type. Then, Bidirectional Encoder Representation from Transformers (BERT) model is fine-tuned for each cluster to estimate the incident severity based on the descriptions. This study introduces the detailed training schedule for the proposed approach. In the case study, a total of 22,458 incidents, categorized into three incident levels as per the extent of life loss and property damage, are used to train and validate the proposed approach. The proposed approach is compared to several state-of-the-art baseline models. The comparison demonstrates the superior performance of the proposed approach in accurately estimating the severity level of incidents. It manifests that the 'cluster-then-estimate' strategy effectively leverages the strengths of BERT to further enhance its estimation capability. To the best of our knowledge, the proposed approach is one of the first to adopt NLP techniques for incident severity estimation based on textual descriptions, which offers practical value for improving incident assessment and decision-making in the maritime industry.