S2-HTC: Hierarchical Text Classification via Fusing the Structural and Semantic Information

Hierarchical Text Classification (HTC) provides a robust mechanism for systematic text categorization, addressing diverse requirements of text understanding and retrieval. A key issue in HTC is the feature learning for long-tail distributed labels through the use of label relations. Most current HTC approaches mainly focus on label structures, overlooking the intricate semantic details of long-tail labels. This often leads to inadequate modeling of these less frequent, but semantically rich labels, compromising the overall classification accuracy. In this study, we propose the method named S2-HTC: Hierarchical Text Classification via fusing the Structural and Semantic Information(S2-HTC), which achieves hierarchical classification by leveraging a method incorporating label structural and semantics relations with the balancing loss calculation. Specifically, S2-HTC introduces the Label Semantic-Aware and Hierarchical Adjacency Matrix (LSA-HAM) to simultaneously capture and integrate the hierarchical and semantic associations of labels. Realizing the natural challenge of long-tailed label distributions in HTC, S2-HTC adopts the balanced loss computation to efficiently represent low-level label features. Experimental results demonstrate that S2-HTC outperforms state-of-the-art approaches.