Multi-label text classification with latent word-wise label information

Abstract

Multi-label text classification (MLTC) is a significant task that aims to assign multiple labels to each given text. There are usually correlations between the labels in the dataset. However, traditional machine learning methods tend to ignore the label correlations. To capture the dependencies between the labels, the sequence-to-sequence (Seq2Seq) model is applied to MLTC tasks. Moreover, to reduce the incorrect penalty caused by the Seq2Seq model due to the inconsistent order of the generated labels, a deep reinforced sequence-to-set (Seq2Set) model is proposed. However, the label generation of the Seq2Set model still relies on a sequence decoder, which cannot eliminate the influence of the predefined label order and exposure bias. Therefore, we propose an MLTC model with latent word-wise label information (MLC-LWL), which constructs effective word-wise labeled information using a labeled topic model and incorporates the label information carried by the word and label context information through a gated network. With the word-wise label information, our model captures the correlations between the labels via a label-to-label structure without being affected by the predefined label order or exposure bias. Extensive experimental results illustrate the effectiveness and significant advantages of our model compared with the state-of-the-art methods.