Disentangled feature graph for Hierarchical Text Classification

Effectively utilizing the hierarchical relationship among labels is the core of Hierarchical Text Classification (HTC). Previous research on HTC has tended to enhance the dependencies between labels. However, they overlook some labels that may conflict with other labels because alleviating label conflicts also weakens label dependencies and reduces the model performance. Therefore, this paper focuses on the issue of label conflicts and studies methods to alleviate label conflicts without affecting the mutual support relationship between labels. To solve the abovementioned problem, we first use the feature disentanglement method to cut off all label connections. Then, the connection among labels is selectively established by constructing a hierarchical graph on disentangled features. Finally, the Graph Neural Networks (GNN) is adopted to encode the obtained Disentanglement Feature Graph (DFG) and enables only labels with connections to support each other, while labels without connections do not interfere with each other. The experimental results on the WOS, RCV1-v2, and BGC datasets show the effectiveness of DFG. In detail, the experimental results show that on the WOS dataset, the model incorporating DFG achieved a 1.07% improvement in Macro-F1, surpassing the best model by 0.27%. On the RCV1-v2 dataset, the model incorporating DFG achieved a 0.95% improvement in Micro-F1, surpassing the best model by 0.21%. On the BGC dataset, the model incorporating DFG achieved a 1.81% improvement in Micro-F1, surpassing the best model by 0.45%.