Maritime Incident Information Extraction using Machine and Deep Learning Techniques

In this paper, we present several Natural Language Processing (NLP) techniques for extracting information about piracy from unstructured maritime news articles. The first technique is a document segmentation algorithm, which, given an article that may discuss several unrelated topics, separates the document into a list of maritime incidents. This technique takes advantage of the notion that maritime articles tend to include many precise dates and achieves an F1-score of 0.82 on a dataset of 52 articles. The second technique extracts the names and types of ships involved in a maritime incident using a combination of Named Entity Recognition (NER) and database lookups; an F1-score of 0.89 is achieved. The third and final technique determines the incident type, the severity of the incident, the response taken by the victim, and the eventual impact of this response. To determine these details, five binary classification models were trained and tested. The overall best algorithm was gradient boosting via the CatBoost library, which achieved an average accuracy of 91%.