<p>Semi-supervised learning (SSL) has shown promise for text classification under limited labeled data, yet existing methods often suffer from unstable adversarial training and inefficient utilization of unlabeled data. To address these challenges, this paper proposes PrefGAN-BERT, a novel framework that integrates Direct Preference Optimization (DPO) into the semi-supervised GAN-BERT architecture. Unlike conventional GAN-based SSL models that rely on noisy pseudo-labels or heuristic stability constraints, PrefGAN-BERT introduces a preference-guided adversarial optimization paradigm that is theoretically grounded and empirically robust. The proposed model features two key innovations. First, DPO reformulates the adversarial min–max game into a structured preference-ranking process based on the Bradley–Terry model, providing smoother gradients, interpretable supervision, and more stable discriminator convergence. Second, an LSTM-based generator enhances sequential modeling capability, producing diverse and semantically coherent pseudo-features that deliver richer adversarial signals and mitigate mode collapse. Together, these components enable a unified optimization strategy that jointly improves classification accuracy, feature diversity, and training stability. Comprehensive experiments on five benchmark datasets demonstrate that PrefGAN-BERT consistently matches or surpasses state-of-the-art semi-supervised and adversarial baselines, achieving an average improvement of 6.1 percentage points over GAN-BERT across all datasets, with notable gains under extremely low-label conditions. Ablation and visualization analyses further confirm that DPO effectively enhances feature separability and model interpretability. Overall, PrefGAN-BERT provides a scalable, theoretically interpretable, and robust framework for advancing preference-guided semi-supervised text classification.</p>