Research On Text Classification Based On Graph Convolutional Neural Network

Text classification is a classic and important task in the field of Natural Language Processing.Most of the existing text classification methods focus on text representation and then use machine learning algorithms or deep learning models for classification.These methods do not consider the implicit interrelationships among texts.This paper proposes two text classification models based on Graph Convolutional Networks,named HGCN and HGCNBERT.In order to solve the problems mentioned above,both of the models introduce text heterogeneous graphs and attention mechanisms into Graph Convolutional Networks to get global information.The HGCN model transforms the information into a topological structure representation through the text heterogeneous graph,and establishes a corresponding connection between the independent texts.In order to dig out the hidden information,the HGCN model expands the amount of text information by introducing topic information,which solves the problem of sparse information in short text classification.This paper proposes a kind of Graph Neural Network that integrates multi-level attention.The HGCN model assigns different weights to nodes according to the different node types and node attributes.According to the different weights,the node information is fused to obtain the embedding representation of the node.Finally,the node embedding with global information is obtained.The HGCN model improves the classification accuracy through node embedding,but it will lose some word order information of the text.The word order information of the text will be lost when the text is transformed into a graph structure.This paper proposes a text classification model named HGCN-BERT which is based on Graph Convolutional Network and BERT.The HGCN-BERT model combines the global information provided by HGCN model and the local information provided by pretraining BERT model.The HGCN-BERT model combines Bi-LSTM with attention mechanism to represent the text semantically.Finally the HGCN-BERT model outputs the predicted category of the text to complete the classification task.In oder to verify the feasibility and advantages of the proposed models,this paper sets up different experiments and baseline models to compare the text classification effect of the HGCN-BERT model.Experiments show that the accuracy of the HGCN-BERT model is better than baseline models,which proves the effectiveness of this method.