Hierarchical Multi-label Text Classification based on a Matrix Factorization and Recursive-Attention Approach

Hierarchical Multi-label Text Classification (HMTC) is an important and challenging task in the field of natural language processing (NLP). For example, the automatic classification of complaint texts in customer service of communication operators is a typical HMTC task. These complaint texts are assigned to multiple categories stored in a hierarchical structure and categories at different levels are related to each other. Most existing HMTC methods, however, either process classification tasks with all levels simultaneously or utilize multiple classifiers for each level separately, which ignore the dependencies among different levels in the hierarchical structure. In this paper, we proposed a Matrix Factorization (MF) and Recursive-Attention (RA) Approach, which is called MF-RA, to handle HMTC tasks. By capturing the associations among labels from different levels, MF-RA can raise the reliability of HMTC. Based on the real-world complaint data of customer service from the communication operator, experiments demonstrate that the Top-1 F1 score of our proposed MF-RA method achieves a maximum increase of 21.0% compared with the commonly used machine learning algorithms (Hierarchical SVM and Clus HMC) and a maximum increase of 5.6% compared with deep learning algorithms (GRU, TextCNN, BERT) in this scene, respectively.