Graph Neural Networks (GNNs) have shown remarkable success in natural language processing tasks, yet existing GNN-based text classification methods often fail to adequately address the completeness and accuracy of text graph construction. This limitation can lead to structural errors that compromise representation quality. We propose HTGNN, a novel hierarchical text graph learning framework for inductive text classification, which addresses these challenges through multi-view graph construction and hierarchical learning. Our approach generates three complementary text graphs from each document, capturing diverse word relationships. HTGNN innovatively learns both local word-level relationships through edge-level analysis and global semantic patterns through view-level integration, incorporating these with a priori constraints to optimize graph structure. This hierarchical local-to-global learning paradigm enhances structural representation while reducing bias from initial graph construction. Comprehensive experiments on benchmark datasets demonstrate HTGNN’s superiority over state-of-the-art methods and its ability to generate task-adaptive graph structures.