Research On Node Classification Of Graph-structured Text

In recent years,with the rapid development of Internet information technology,the Internet has produced a large amount of text data.There are connections between text data,the data itself is abstracted as nodes,and the connections between data are abstracted as edges,forming a complex graph-structured network,and node classification is an important task for processing graph-structured text.The research content of this thesis aims at how to improve the classification effect of nodes in graph-structured text,there are two directions: node classification of regular graph-structured text and node classification of irregular graph-structured text.The main contents are as follows:Firstly,in the node classification task of regular graph-structured text,the traditional single-channel neural network model only considers one aspect of the feature information of the text sequence,which is not comprehensive enough.To solve this problem,a transformer-capsule integration model is constructed,which adopts two channels.The upper channel uses capsule network to extract the local phrase features of text sequences,while the lower channel uses transformer model to extract the global semantic features of text sequences,and then carries out feature fusion.The obtained feature vector takes into account the advantages of the two single-channel models.In addition,the traditional capsule model is improved by introducing the attention mechanism between two adjacent capsule layers to assign the corresponding weight values to the capsules participating in the information transmission more reasonably,so as to reduce the interference of noise capsules on the classification results.Secondly,in the task of node classification of irregular graph-structured text,at present,the GAT model has the problem that increasing the number of graph attention layers will reduce the classification performance of the model.The general number of graph convolutional layers is set to 3,which is far from reaching the "depth" requirement of deep learning.To solve this problem,the gating mechanism and residual connection mechanism are introduced into the GAT model,and then the GRGAT model is constructed.On the one hand,the useless information in feature subspace is filtered through the gating mechanism,and on the other hand,the information flow can be merged across layers through the residual connection mechanism.GRGAT model increases the number of layers of graph attention in GAT model from 3 to 18,and the effectiveness of GRGAT model is verified by experiments.