Multi-label Text Classification: Select Distinct Semantic Understanding for Different Labels

摘要

Multi-label classification is a challenging task in natural language processing. Most of existing methods tend to ignore the semantic information of the text. Besides, different parts of the text contribute differently to each label, which is not considered by most of existing methods. In this paper, we propose a novel model for multi-label text classification. This model generates high-level semantic understanding representations with a multi-level dilated convolution. The multi-level dilated convolution effectively reduces dimension and expands the receptive fields without loss of information. Moreover, a hybrid attention mechanism is designed to capture most relevant information of the text based on trainable label embeddings and semantic understanding. Experimental results on the dataset AAPD and RCV1-V2 show that our model has significant advantages over baseline methods.