Evaluating the safety of distant-water fishing vessels using text classification and knowledge mining

Potential knowledge can be extracted from the safety text of distant water fishing (DWF) vessels. However, the previous approaches have not yet been fully developed for the safety text of fishing vessels. Some challenges remained, such as the low accuracy of text classification, and insufficient depth of knowledge extraction. In this study, an analytical approach was proposed to combine text classification, knowledge mining, and co-occurrence network technology under the Cape Town Agreement(CTA) of 2012 . The text data on DWF vessel safety was also collected from the fishery management organizations, associations, and over 20 fishery enterprises from eight Chinese coastal provinces and cities, including Zhejiang, Shanghai, and Fujian. The DWF vessel safety corpus consisted of more than 5,000 valid questions and 100,000 characters. The analytical approach comprised three stages. Firstly, a hybrid deep learning model was developed using bidirectional encoder representations from transformers-text convolutional neural networks (BERT-TextCNN), according to the characteristics of DWF vessel safety text, such as diverse data types, sparse data features, and fuzzy boundaries. The character vectors were generated to extract the contextual semantic and deep syntactic information of the text using BERT during text representation. Multiple convolutional kernels of TextCNN were utilized to spatially model the generated character vectors and then to extract the local features for the accurate classification of safety theme. Secondly, term rrequency-inverse document frequency (TF-IDF) was employed to extract the key safety knowledge of fishing vessels, considering the importance and prevalence of knowledge within each safety theme. Finally, a co-occurrence network was constructed to visualize the safety knowledge of fishing vessels, including distributional patterns and interconnections. The results show that the BERT-TextCNN model achieved an accuracy, macro average recall rate, and macro average F₁ value of 98.20%, 98.02%, and 98.05%, respectively. The performance outperformed the other 17 comparative models, which utilized three text representations (BERT, Word2vec, and Character embedding) and six neural networks (TextCNN, Softmax, DPCNN, BiLSTM-Attention, RCNN, and Transformer). Meanwhile, the theme-based knowledge mining and analytical approach achieved clear rankings of DWF vessel compliance and safety management knowledge, as well as relationship networks crossing ten safety knowledge themes of fishing vessels, including provisions, structure, stability, electrical installations, fire protection, crew protections, life-saving equipment, emergency procedures, wireless communication, and shipborne navigation equipment. Intelligent safety knowledge services and decision-making tools were obtained to improve the compliance level and safety management efficiency in DWF. The finding can provide a strong reference to promote the application and development of knowledge service systems and the smart fishing industry.