Hierarchy-aware BERT-GCN Dual-Channel Global Model for Hierarchical Text Classification

Hierarchical text classification(HTC) aims to assign one or more labels in a predefined hierarchical label system to text. The classes between different levels are often related, so how to fully utilize the hierarchical structural information of the label system has become a challenging task in HTC. In previous researches, in order to solve the problem of weak interaction between text and labels since text and labels are encoded separately, some works proposed the hierarchy guided global hierarchy embedding methods, which embed the global hierarchy into the encoder in the training phase. However, the current methods utilize the semantic and structural information of global hierarchy in a more limited way, which restricts the model effect. In this work, a BERT-GCN dual-channel global hierarchy embedding method is proposed, which utilizes contrastive learning to achieve the complementary advantages of BERT and GCN to better capture the semantic and hierarchical dependencies between labels. This work embeds global hierarchy into the dual-channel encoder through the mask prediction task, and improves the performance of multi-label text classification by designing a new joint loss function in the local hierarchy learning phase. Our method has been validated for effectiveness on benchmark datasets.