Label prompt for multi-label text classification

Multi-label text classification has been widely concerned by scholars due to its contribution practical applications. One of the key challenges in multi-label is how extract and leverage correlation among labels. However, it quite challenging directly model correlations labels a complex unknown label space. In this paper, we propose Label Prompt Text Classification (LP-MTC), which inspired idea prompt learning pre-trained language model. Specifically, design set templates for classification, integrate into input model, jointly optimize Masked Language Models (MLM). way, as well semantic information between with help self-attention can be captured, thus performance effectively improved. Extensive empirical experiments on multiple datasets demonstrate effectiveness our method. Compared BERT, LP-MTC improved 3.4% micro-F1 average over four public datasets.