Research On News Text Classification Based On Graph Neural Networ

With the rapid development of computer technology in recent years,as an emerging method of deep learning,graph neural network has been widely used in computer vision,natural language processing and other directions.As a basic task of natural language processing,text classification has always been a popular topic of research for Chinese and foreign scholars,and news text,as a special kind of text,has the characteristics of short average length,more entities and concise content.This makes general text classification methods poorly portable for news text classification tasks and difficult to improve classification accuracy.Therefore,this paper mainly uses the advantages of graph neural network for flexible information structure mastery and strong message transfer ability,and conducts research on news text classification model around graph neural network,aiming to improve the classification ability of text classification model on news text.The main research work and innovation of this paper are as follows:1.To address the problems of feature sparsity,missing context,and low utilization of text features in news texts,this paper proposes in a graph neural network model(TWPGCN)that propagates information through three different nodes.Firstly,the text,words and word lexical labels in the whole corpus are extracted as nodes of the constructed heterogeneous graph,and different types of edges are constructed for them using different relationships among the three,which fully describe the information contained in the text;secondly,in order to better transfer information among different types of nodes,this paper proposes a heterogeneous graph node information updating method,which updates node information through three different channels respectively Second,in order to better transfer information between different types of nodes,this paper proposes a heterogeneous graph node information update method,which updates node information through three different channels and fuses the information between different nodes.Finally,the TWPGCN model is compared and ablated using four publicly available base datasets.The experimental results show that the TWPGCN model has better classification ability than the baseline model,with improvements of 0.87,0.65,1.95,and 0.78 percentage points on the four datasets respectively compared to the Text GCN model,which is constructed with the whole corpus,and the TWPGCN model considering lexical labels has better performance compared to the other models.2.In order to address the shortcomings of global corpus composition and the lack of utilization of specific information in text by current text classification models,we propose a graph neural network model(LA-GNN)with a single corpus composition and a dual-attention mechanism for learning text information.Firstly,this paper constructs forward and reverse order graphs for a single text considering the order information in the text;secondly,in order to reflect the importance of forward and reverse order information in text classification,this paper carries out information transfer by combining gated graph neural networks for forward and reverse order graphs of text,and this paper proposes the influence of text labeling features on text classification by using a text-labeling dual-attention mechanism Message readout is performed for the features updated in the message passing phase.Finally,experiments were conducted to compare the LA-GNN model with the previously proposed TWPGCN model and the Text ING model,which performed well in the classification task,with improvements of0.78,0.69,and 0.17 percentage points on the R8,Ohsumed,and MR datasets compared to the Text ING model,which was framed as a single text.The experimental results demonstrate the strong classification performance of the LA-GNN model.In the ablation experiment as well as the attention visualisation experiment,the effectiveness as well as the practicality of the LA-GNN model are analysed in detail in this paper.3.A graph neural network model(R-GAN)for news text classification is proposed to address the imbalance problem of the dual-attention module in the LA-GNN model,combining the advantages and disadvantages of comparing the TWPGCN model and the LA-GNN model.Firstly,the R-GAN model is proposed for the different importance of words in news text and the difference of news label features;secondly,the exploratory analysis of the dataset and data pre-processing are performed on the public news dataset provided by today’s headlines to improve the usability and learnability of news text;finally,the R-GAN model is tested on the news text dataset through threshold experiments,comparison experiments and control experiments within the dataset.Finally,the R-GAN model was experimented on the news text dataset through threshold experiments,comparison experiments and controlled experiments within the dataset.Compared to the LA-GNN model,the R-GAN model improved in accuracy,recall and F1 values by 0.17,0.15 and 0.17 percentage points.With the support of the experimental data,the specific advantages of the R-GAN model in news text and the problems and challenges faced in classifying different categories of news are pointed out in order to promote the practical implementation of graph neural networks in news text classification.