LSPCL: Label-specific supervised prototype contrastive learning for multi-label text classification

Gang Wang

doi:10.1016/j.knosys.2024.112887

2025-01-01

PRE

Gang Wang

Yajun Du *

DOI：10.1016/j.knosys.2024.112887

原文链接

原文求助

摘要

En 中文

Contrastive learning has led to further breakthroughs in feature representation. However, text datasets often suffer from class imbalance issues, resulting in suboptimal feature representation performance. Recent studies have introduced prototype networks for single-label scenarios. The prototype network replaces the contrastive samples with class prototypes to mitigate class imbalances. However, in multi-label scenarios, directly updating the prototype network using multi-label samples may introduce a large amount of noise, because a sample may belong to multiple classes. Therefore, the construction of high-quality prototype representations in multi-label scenarios becomes challenging. To address this issue, we propose a label- specific supervised prototype contrastive learning network (LSPCL). LSPCL dynamically learns the relationships between labels and constructs a label-specific prototype repository to enhance the quality of multi-label text feature representations. Specifically, we utilize external knowledge for data augmentation to capture label semantic information more effectively. Our proposal involves a dynamic graph convolutional network that is designed to learn the potential relationships between labels, thereby facilitating improved exploration of label distributions across datasets. Furthermore, we construct a label-specific prototype repository and initialize it using label information. Subsequently, through the attention interaction between labels and text, we obtain label-specific text representations and update the prototype repository using these representations. Finally, we optimize the contrastive learning process using the label-specific prototype repository. In addition, we design a new multi-label supervised prototype contrastive learning loss function to optimize the training process. LSPCL achieved performance improvements of up to 2.9%, 2.1%, and 8.42% on the AAPD, RCV1-v2, and EUR-Lex datasets, respectively, demonstrating its superiority over state-of-the-art methods.

Keyword:

Graph convolutional network

Contrastive learning

Prototype network

Multi-label text classification