<p>Real-world text classification often faces the challenge of long-tailed data distributions, where most categories contain only a few samples. Directly fine-tuning pre-trained language models (PLMs) on such imbalanced data typically causes semantic space collapse for tail classes and severe overfitting to head classes. To address this, we propose the Long-tail Aware Graph Adapter (LAGA), a novel framework that shifts from full-parameter fine-tuning to structured semantic space adaptation. LAGA first aligns text and label semantics using natural language descriptions to build a unified semantic coordinate system. Crucially, it then freezes the PLM backbone and employs a heterogeneous graph adapter encompassing texts, labels, and learnable prototypes. Through graph neural network propagation, this adapter contextually refines the frozen semantic space. Coupled with dynamic prototype evolution and a tail-aware optimization objective, LAGA forms robust decision boundaries for scarce categories. Extensive experiments on six benchmarks demonstrate that LAGA consistently enhances various PLMs in few-shot, long-tail settings. It achieves superior tail-class recognition and a better performance-efficiency trade-off than strong baselines, including massive large language models (LLMs), proving that graph-based structural adaptation is a highly effective solution for imbalanced text classification.</p>