MR-HAN: Multi-Relational Hierarchical Attention Network for Inductive Text Classification

Text classification remains a fundamental challenge in natural language processing (NLP), where graph-based approaches have shown promise by modeling textual data as structured graphs. However, existing methods often struggle to balance local and global receptive fields efficiently, particularly in inductive settings. To address this, we propose Multi-Relational Hierarchical Attention Network (MR-HAN), a novel inductive framework that jointly models the syntactic and semantic relations within documents. MR-HAN employs relation-specific graphs to capture word information, followed by a hierarchical attention mechanism that dynamically fuses multi-relational signals. Unlike transductive methods, our approach maintains scalability while enhancing interpretability through learned attention weights. Extensive experiments on five benchmark datasets demonstrate that MR-HAN outperforms state-of-the-art baselines, particularly on long-text classification.