Text multi-label learning method based on label-aware attention and semantic dependency

Text multi-label learning deals with examples having multiple labels simultaneously. It can be applied to many fields, such as text categorization, medical diagnosis recognition and topic recommendation. Existing multi-label learning methods treat a label as an atomic symbol without considering semantic information, while labels are texts with semantic information composed of words, which can guide to obtain discriminative text features. In order to select discriminatory features from redundant content, we consider the semantic labels and establish the relationship between labels and texts based on the attention mechanism. Label relationship modeling helps to further improve the model’s effectiveness and we model the high label relationship based on the principle of graph convolutional networks (GCN). Then the LAA_SD method is proposed, which combines enhanced text feature representation with label semantic dependency to perform text multi-label learning. A comparative study with state-of-the-art approaches manifests the competitive performance of the proposed model.