Related Experiment Videos
PrefGAN-BERT: integrating direct preference optimization into semi-supervised GAN-BERT for robust text classification
Dangguo Shao1, Tianzheng Lai1, Lei Ma2
1Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, 650500, Yunnan, China.
Scientific Reports
|April 17, 2026
Summary
PrefGAN-BERT improves semi-supervised learning for text classification using preference optimization and an LSTM generator. This novel approach enhances accuracy and stability, outperforming existing methods, especially with limited labeled data.
Area of Science:
- Artificial Intelligence
- Natural Language Processing
- Machine Learning
Background:
- Semi-supervised learning (SSL) shows potential for text classification with scarce labeled data.
- Existing methods face challenges with unstable adversarial training and poor utilization of unlabeled data.
Purpose of the Study:
- To introduce PrefGAN-BERT, a novel framework integrating Direct Preference Optimization (DPO) into GAN-BERT for robust semi-supervised text classification.
- To enhance adversarial training stability and unlabeled data utilization in text classification.
Main Methods:
- Integrated Direct Preference Optimization (DPO) into GAN-BERT, reformulating adversarial training as a preference-ranking process.
- Employed an LSTM-based generator to improve sequential modeling, feature diversity, and mitigate mode collapse.
- Utilized the Bradley-Terry model for smoother gradients and stable discriminator convergence.
Main Results:
- PrefGAN-BERT achieved superior performance compared to state-of-the-art semi-supervised and adversarial baselines across five benchmark datasets.
- Demonstrated an average improvement of 6.1 percentage points over GAN-BERT, particularly under extremely low-label conditions.
- Ablation studies confirmed DPO's effectiveness in enhancing feature separability and model interpretability.
Conclusions:
- PrefGAN-BERT offers a scalable, theoretically interpretable, and robust framework for preference-guided semi-supervised text classification.
- The proposed method effectively addresses limitations of existing GAN-based SSL models, improving accuracy and training stability.
