图马尔可夫卷积神经网络半监督文本分类研究

Graph Markov Convolutional Networks for Semi-Supervised Text Classification

随着卷积神经网络在图结构数据上的成功泛化,许多研究者将图卷积神经网络应用于文本分类;在上述方法中,以文档和单词为节点构造异构文本图网络,通过学习图节点的特征表示进行文本分类,未能有效利用节点标签的依赖关系。现提出了文本图马尔可夫卷积神经网络(TextGMCN)模型,模型利用异构图中未分类节点的条件联合分布建模节点标签的依赖性;模型利用图卷积神经网络通过端到端的训练,学习有效的文本节点表示。通过变分EM算法进行训练。在多个基准数据集上的实验结果表明,考虑文本节点标签依赖性的TextGMCN模型取得了更优的节点分类性能。

With the successful generalization of the convolutional neural network on the graph, many researchers have applied the graph convolutional neural network to text classification. In this method, heterogeneous text graph networks are constructed with documents and words as nodes, and text is classified by learning the characteristic representation of nodes, which fails to effectively utilize the dependency of node labels. In this paper, a text graph Markov convolutional neural network(TextGMCN) model is proposed, which uses the conditional joint distribution of unclassified nodes in heterogeneous graphs to model the dependency of node labels. The model uses the graph convolutional neural network to learn the effective representation of text nodes through end-to-end training. The training is carried out by means of a variational EM algorithm. Experimental results on multiple datasets show that the TextGMCN model considering the label dependency of text nodes achieves better node classification performance.