LMSF: A Label-Guided Multi-Order Semantic Fusion Network for Extreme Multi-Label Text Classification

Extreme Multi-Label Text Classification involves assigning multiple relevant labels to each document from label sets containing millions of categories, with applications spanning e-commerce recommendation, question-answer retrieval, and social-media tag generation. It poses significant challenges due to extremely long-tailed label distributions and semantic ambiguity among labels. To address these issues, we propose LMSF, a Label-Guided Multi-Semantic Fusion network that models label semantics from both global and local perspectives. A graph-based label representation module employs directed graph attention to capture structured inter-label dependencies, while a label-guided attention pooling module dynamically generates per-label guidance vectors to focus on the most discriminative text spans, mitigating label overlap and enhancing long-tail label recognition. These heterogeneous features are then fused via a multi-order semantic feature fusion module, which uses cross-attention and multi-scale convolutions to model complex, higher-order interactions. Extensive experiments across multiple benchmark datasets demonstrate that LMSF achieves state-of-the-art performance on standard evaluation metrics, particularly excelling on rare labels. These results validate its effectiveness and robustness in large-scale, imbalanced multi-label classification tasks.