TASGAN : A Teacher‐Advised Semi‐Supervised GAN for Low‐Resource Text Classification

ABSTRACT Generative adversarial networks in semi‐supervised learning are often limited by training instability, while conventional stabilization methods are prone to introducing confirmation bias. To address this issue, we propose the teacher‐advised semi‐supervised generative adversarial network (TASGAN). The method introduces a teacher–student Consistency regularizer in GAN‐based semi‐supervised learning and, in a departure from conventional paradigms, applies it exclusively to the limited labelled data. This strategy provides a stable supervisory signal for adversarial training through knowledge distillation from a teacher model. By integrating this distillation loss with adversarial losses into a multi‐objective framework, TASGAN effectively mitigates the multi‐task conflict within the discriminator, significantly enhancing the model's stability and performance in settings of extreme data scarcity. Experiments on three text classification benchmarks demonstrate that TASGAN significantly outperforms baselines in extremely low‐resource scenarios.