All is attention for multi-label text classification

Multi-label text classification(MLTC) is a key task in natural language processing. Its challenge is to extract latent semantic features from text and effectively exploit label-associated features. This work proposes an MLTC model driven solely by attention mechanisms, which includes Graph Attention(GA), Class-Specific Attention(CSA), and Multi-Head Attention(MHA) modules. The GA module examines and records label dependencies by considering label semantic features as attributes of graph nodes. It uses graph embedding to maintain structural relationships within the label graph. Meanwhile, the CSA module produces distinctive features for each category by utilizing spatial attention scores, thereby improving classification accuracy. Then, the MHA module facilitates extensive feature interactions, enhancing the expressiveness of text features and supporting the handling of long-range dependencies. Experimental evaluations conducted on two MLTC datasets show that our proposed model outperforms existing MLTC algorithms, achieving state-of-the-art performance. These results highlight the effectiveness of our attention-based approach in tackling the complexity of MLTC tasks.