MLGN:A Multi-Label Guided Network for Improving Text Classification

Abstract

Within natural language processing, multi-label classification is an important but challenging task. It is more complex than single-label classification since the document representations need to cover fine-grained label information, while the labels predicted by the model are often related. Recently, large pre-trained language models have achieved great performance on multi-label classification tasks, typically using embedding of [CLS] vector as the semantic representation of entire document and matching it with candidate labels. However, existing methods tend to ignore label semantics, and the relationships between labels and documents are not effectively mined. In addition, the linear layers used for fine-tuning do not take the correlations between labels into account. In this work, we propose a Multi-Label Guided Network (MLGN) capable to guide document representation with multi-label semantic information. Furthermore, we utilize correlation knowledge to enhance the original label prediction in downstream tasks. The extensive experimental trials show that MLGN transcends previous works on several publicly available datasets. Our source code is available at https://github.com/L199Q/MLGN.

Keywords

• Multi-label text classification
• document representation
• label semantics
• contrastive learning
• label correlation