Multi-Label Text Classification Model Combining BiLSTM and Hypergraph Attention

Compared with the traditional graph structure, hypergraph structure can go beyond pair-wise relations, and then further capture the high-order interaction between words. However, the multi-label text classification model based solely on hypergraph neural networks has two limitations: first, the hypergraph neural network model lacks rich semantic node information in text; second, the sequence information is lost during the construction of hypergraph. A multi-label text classification model that combines sequence models with hypergraph attention networks is proposed to address the aforementioned limitations. The hypergraph node vector features are obtained by using the fine-tuned BERT models to capture rich semantic information in text. Among them, hyperedges include bag hyperedges and semantic hyperedges formed by exploring potential topics in text instances through BERTopic. Finally, the sequence model BiLSTM is integrated to capture the word order features of the text, thereby obtaining a more comprehensive text representation. Thus, a more comprehensive and adequate text representation can be obtained. Multiple experiments on two publicly available datasets, Reuters-21578 and AAPD, show that the model has better classification results than the baseline model.