Hierarchical Multi-label Classification Method for Maritime Distress Safety Information
Minxin Zhu, Yuan Gao, Lianfeng Hu, Jiwei Hu · 2024
Maritime distress safety (MDS) text information hierarchical multi-label text classification (HMTC) involves categorizing MDS texts into a label set with a hierarchical structure. Due to the varying impact of specific keywords in MDS text information on certain labels within the hierarchical label system, an improved TF-IDF-Keyword-Filtering (TIKF) algorithm based on the traditional TF-IDF algorithm and the KABLG model are proposed to highlight these keywords. The TIKF algorithm incorporates intra-parent class correlation, inter-class correlation, and term frequency factor to calculate the weight of each feature keyword. It also utilizes a pre-trained Word2Vec model to construct new keyword vectors for keyword filtering. This reduces the influence of unimportant words, filters out specific keywords that are more important for hierarchical labels, and accelerates the model learning process. The KABLG model uses Bi-LSTM and hierarchy-GCN to respectively extract contextual information and hierarchical label representations. It designs interactions between textual features and label features, and adaptively integrates them with keyword features, resulting in improved classification performance of the model. We prove the effectiveness of the TIKF algorithm and the KABLG model on the MDS dataset.