融合LDA的门控图卷积网络文本分类研究

Incorporating LDA into gated graph convolutional networks for the study of text classification

在现有文本图基础上引入隐狄利克雷分布,将文档-主题和主题-词信息融入文本图以丰富文本图中节点间关系,之后将该文本图送入一个基于图卷积网络门控机制模型.在多个数据集上进行验证.结果表明,所提出的模型优于现有图卷积网络文本分类模型.

At present,graph convolutional network based text classification model uses word co-occurrence and word frequency-inverse document frequency(TF-IDF)information for graph construction.In this paper,we introduce the Latent Dirichlet Allocation(LDA)to construct text graph based on the existing text graph,and integrate the document-topic and topic-word information into the text graph,thereby enriching the relationship between the nodes in the text graph,and then feed the text graph into a gating mechanism model based on graph convolutional network(G-GCN).Verification on multiple datasets show that the model proposed in this paper is superior to existing text classification models based on graph convolutional networks.