Text classification of issues concerning implementation of IMO instruments based on deep bidirectional language representation model

Abstract To enhance the usefulness of regulatory monitoring in global maritime governance, the International Maritime Organization (IMO) has required member state audits based on the IMO Instruments Implementation Code (III Code). Still, the original manual classification of audit findings to regulatory clauses remains resource-intensive and subject to confusion arising from semantic ambiguity and a variety of textual sources. To assist with and automate the classification of textual audit findings into categories of regulatory clauses, this research offers a BERT language representation model. A comprehensive dataset of 961 findings were collected from the IMO Comprehensive Audit Summary Report that ultimately constitute more than 40 categories of non-conformance. The classification model described here achieved an overall classification accuracy of 72.0% and a macro F1 score of 0.69, which is superior to the appropriate conventional baselines, including TF-IDF + logistic regression, and an additional Bi-LSTM classification model. Other studies on the dataset provide some evidence for the effectiveness of cross-lingual pretraining (MLM or TLM), layer-wise semantic representation analysis and simulation of audit scenario data. Based on visualizing the behavioral responses of the audit findings, the relationship between asset levels, substitution strategies, and default monetary thresholds offered an original mode of interpretation of compliance agents.