In recent years, deep learning has significantly advanced text classification algorithms. However, traditional text classification methods still face challenges in modeling long-range dependencies and capturing fine-grained semantic features. Capsule networks offer a new approach to address these issues by effectively modeling hierarchical feature relationships through dynamic routing. Nevertheless, capsule network-based methods suffer from three key limitations: (1) high computational complexity of dynamic routing, which hinders long text processing; (2) insufficient integration of local semantic features and global contextual information; (3) training challenges caused by routing instability. To tackle these problems, this study proposes a hybrid attention dynamic sparse routing capsule network (HADSR-Caps). The model employs a hybrid attention mechanism that fuses local window attention and global attention to capture multi-scale semantic features. It then introduces a dynamic sparse routing algorithm, integrating K-Sparse pruning and adaptive routing termination, which significantly reduces computational costs while enhancing model robustness. Experimental results on multiple public datasets demonstrate that HADSR-Caps achieves significant improvements in both classification accuracy and robustness, outperforming traditional capsule networks and other benchmark models.