<p>In this work, we propose Multi-Head TextGCN, a novel extension of the traditional TextGCN that integrates multi-head attention mechanisms into the graph convolutional process. By modeling document and word nodes jointly within a graph and implementing multiple specialized attention heads to capture local, global, and topical semantic interactions, our approach learns richer and more robust feature representations for document classification. Extensive evaluations on several benchmark datasets demonstrate that our model achieves significant improvements in classification accuracy, proving especially effective in low-resource and few-shot learning scenarios. These promising results underscore the potential of our heterogeneous multi-head framework for advancing graph-based text representation learning.</p>