DynaMatch: Dynamic Self-Ensemble for Adaptive Semi-Supervised Text Classification

Semi-supervised text classification (SSTC) faces challenges in pseudo-label quality and robustness, particularly with limited labeled data and imbalanced class distributions. To address these, we propose DynaMatch, a novel framework for adaptive SSTC that integrates Dynamic Self-Ensemble Learning (DSEL), Adaptive Confidence Scoring (ACDM), and Historical Bias Correction. DynaMatch leverages DSEL for robust predictions from instantaneous model states. ACDM then refines pseudo-labeling through self-ensemble diversity evaluation, dynamic threshold adjustment, and historical bias correction to identify valuable samples and mitigate class imbalance. Evaluated on the Unified Semi-supervised Benchmark (USB), including long-tailed imbalanced datasets, DynaMatch consistently outperforms state-of-the-art baselines. It achieves superior performance, with approximately 0.5% to 1.0% F1-score improvement, especially excelling in scenarios with scarce labeled data and severe class imbalance. An ablation study confirms the synergistic contributions of each component, reinforcing DynaMatch's efficacy and practical utility.