GACaps-HTC: graph attention capsule network for hierarchical text classification

Hierarchical text classification has been receiving increasing attention due to its vast range of applications in real-world natural language processing tasks. While previous approaches have focused on effectively exploiting the label hierarchy for classification or capturing latent label relationships, few studies have integrated these concepts. In this work, we propose a graph attention capsule network for hierarchical text classification (GACaps-HTC), designed to capture both the explicit hierarchy and implicit relationships of labels. A graph attention network is employed to incorporate the information on the label hierarchy into a textual representation, whereas a capsule network infers classification probabilities by understanding the latent label relationships via iterative updates. The proposed approach is optimized using a loss term designed to address the innate label imbalance issue of the task. Experiments were conducted on two widely used text classification datasets, the WOS-46985 dataset and the RCV1 dataset. The results reveal that the proposed approach achieved a 0.6% gain and a 2.0% gain in micro-F1 and macro-F1 scores, respectively, on the WOS-46985 dataset and a 0.3% gain and a 2.2% gain in micro-F1 and macro-F1 scores, respectively, on the RCV1 dataset compared to the previous state-of-the-art approaches. Further ablation studies show that each component in GACaps-HTC played a part in enhancing the classification performance.