Abstract: Text classification,as a critical task in natural language processing,aims to assign labels to input documents.The Co-occurrence relationship between words offers key perspectives on text characteristics and vocabulary distribution,while word embeddings supply rich semantic information,influencing global vocabulary interaction and potential semantic relationships.Previous research has struggled to adequately incorporate both aspects or has disproportionately emphasized one over the other.To address this issue,a novel method is proposed in this paper that adaptively fuses these two types of information,aiming to strike a balance that can improve model performance while considering both structural relationships and embedded information.The method begins by constructing text data into text co-occurrence graphs and text embedding graphs,reflecting the context structure and semantic embedding information respectively.Graph convolution is then utilized to enhance node embeddings.In the graph pooling layer,node embeddings are fused and nodes of higher importance are identified by employing a hierarchical pooling model,learning document level representations layer by layer.Furthermore,we introduce a gated fusion module to adaptively fuse the embeddings of the two graphs.The proposed approach is validated with extensive experiments on five publicly available text classification datasets,and the experimental results show the superior performance of the HTGNN model in text classification tasks.

Key words: Text classification, Graph neural network, Graph representation learning, Graph classification, Attention mechanism