Multi-Label Text Classification Based on Contrastive and Correlation Learning
Shuo Yang, Shu-wen Gao · 2024
Multi-label text classification plays a crucial role in various domains. However, accurately capturing complex inter-label relationships and delving into the semantic information between labels and text remains a challenging problem. Therefore, we propose a multi-label text classification method based on contrastive and correlation learning. By introducing contrastive learning into multi-label text classification tasks, it enhances the distinctiveness and expressiveness of text and label features. In the extraction of label features, external knowledge from Wikipedia is incorporated and various embedding methods are employed to extract label information, enabling a deeper exploration of label semantics. Meanwhile, GAT is used to more accurately extract inter-label correlations. In the prediction module, an improved label correlation network is introduced to further consider label relevance. Experimental results demonstrate the feasibility and effectiveness of the proposed method on two publicly available datasets, AAPD and RCV1. Compared to state-of-the-art models, our method achieves a 0.5%-1.5% improvement in micro-F1 metrics, validating the efficacy of the approach.