<p>Most graph neural network-based multi-label text classification methods suffer from two key engineering limitations: poor generalization to unseen data due to transductive learning, and suboptimal performance caused by fixed-label graphs that fail to capture dynamic label correlations. To address the above limitations, this paper proposes a new semantic-guided multi-stage graph learning method for inductive multi-label text classification (SG2L-IMLTC). Specifically, text graphs and label graphs are constructed as the initial task graphs based on semantically meaningful text embeddings and label embeddingsuence of inter-label correlation on label grap. To obtain improved graph structures while simultaneously enhancing the quality of text embeddings and label embeddings, a multi-stage graph learning method that is applicable to both text graph learning and label graph learning is proposed. In this method, on one hand, structural learning is executed through a two-stage strategy involving the construction of reconstructed graphs and the generation of optimized graphs. On the other hand, feature learning is achieved via a two-stage feature fusion mechanism that integrates cross-graph global fusion and gated adaptive fusion. Extensive experiments on four benchmark multi-label text datasets have shown that the proposed SG2L-IMLTC method outperforms state-of-the-art graph neural networks-based inductive multi-label text classification methods in multiple evaluation metrics. For instance, on the BGC dataset, our method achieves improvements of 2.70%, 1.62%, 3.02%, and 2.31% over the state-of-the-art methods in Micro-F1, P@3, P@5, and nDCG@5, respectively.</p>