.png){kind=icon}
.png){kind=icon}
Return
Regularization-based semi-supervised generative adversarial learning for text classification with limited supervision
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2026.115356.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Text classification is a primary and fundamental task in Natural Language Processing (NLP), which is suitable for any text auto-archiving situation. For text classification with limited supervision, recent studies have introduced Semi-Supervised Generative adversarial networks (SS-GANs) to enhance the representation capability of BERT encoders. However, these methods still suffer from the issue of the single-encoder structure struggles to obtain multiple distinct feature representations, which limits the category information learning ability of the model. To solve the problems, we propose a Regularization-Based Semi-Supervised Generative Adversarial learning framework, namely Re-SSGAN. The framework employs a novel regularization-based generative adversarial learning paradigm, which implements interactive adversarial learning by employing the multiple outputs representation of the encoder and a single output representation from a latent representation generator. This paradigm design utilizes the stochastic dropout mechanism in deep neural networks to obtain multiple different feature representations of the same sample, thereby achieving feature enhancement. Moreover, we designed a fusion-based adversarial constraint module that captures multiple feature representations through pooling and expectation operations to constrain the latent representation generator into producing higher-quality adversarial sample representations. In addition, we also design the cross-contrast loss function, which minimizes the combination of bidirectional Kullback - Leibler (KL) divergence and cross-entropy loss for ensuring correct classification and making multiple output representations of the same sample converge toward consistency. We conducted semi-supervised classification experiments over six text datasets, and the results demonstrated the effectiveness of Re-SSGAN. Notably, Re-SSGAN achieved a more significant improvement in classification performance when labeled text was extremely scarce.
Journal
IF:
8
Papers:
5.4K
Citations:
3.5W
Organization
View more organizations