Automatic Identification of Maritime Incidents from Unstructured Articles

Abstract

In this paper, we present two Natural Language Processing (NLP) techniques for identifying maritime incidents described in unstructured articles from multiple sources. The first technique is a document classification scheme that determines if an article describes a maritime incident. Two variations of each article are created: the first only contains the article’s title, the other contains the title and content. These are converted to both binary and frequency bags-of-words. Furthermore, two feature selection methods are tested: Weka’s CfsSubsetEval and retaining the 300 most frequent words. Each dataset is tested with 41 classifiers from the Weka suite, with the most accurate classifiers including Logistic Regression (98.5%), AdaBoostM1(BayesNet) (98.33%), and RandomForest (97.56%). The second technique performs information extraction on an article to determine the location of the maritime incident. In addition to using regular expressions and Named Entity Recognition (NER), the approach focuses its attention on sentences that contain piracy keywords as well as sentences which occur earlier in the article. In our testing, this approach achieved 87.9% accuracy. Together the two techniques form a pipeline where the positive examples from the document classification algorithm are fed into the information extraction algorithm.