Research On Text Classification Based On Graph Convolutional Neural Network

The Internet is generating new text data all the time,and text classification can greatly improve our work or life efficiency.Traditional machine learning requires the construction of complicated manual text features,while deep learning avoids the construction of manual features,but there is the problem of gradient explosion and gradient disappearance,so that deep learning models cannot solve the problem of long text dependence,and at the same time,deep learning is for single sample data,and cannot consider the correlation between samples,so it cannot deal with literature interquotation data well.It cannot consider the correlation relationship between samples,so it cannot handle the literature interquotation data well.The graph convolutional neural network can fully consider the association relationship of each vertex,in which the vertex is the sample data,and the current graph convolutional network uses word co-occurrence to process the text and build into graph text data,and the use of word co-occurrence can well solve the long dependency problem,so the graph convolutional neural network is currently a more effective way to solve the text classification task.In this paper,we improve the graph convolutional neural network based on the traditional graph convolutional neural network in the network reconstruction,and improve the graph convolutional neural network for feature extraction of data,the main work of this paper has the following two points.(1)The hypergraph convolutional network combined by hypergraph and graph convolutional neural network can well express the multivariate relationships in the dataset,so it effectively handles the literature interquotation dataset,but the loss function used by the hypergraph convolutional neural network is the traditional cross-entropy,and the cross-entropy loss function only cares about the accuracy of the prediction probability of the correct labels,ignoring the differences of other non-correct labels,which leads to the looser learned features,thus The features extracted by the model for each class do not differ much from each other,and simply improving the network structure cannot compensate for the defects brought by cross-entropy.Therefore,this paper proposes a hypergraph convolutional neural network based on prototype learning,which uses prototype learning to compensate for the defects caused by crossentropy.The prototype learning will train the center vector of each class,and by calculating the distance between the features output by the model and the center vector of the corresponding label of the sample as the loss of the model,the model will be guided to update the weight parameters by means of prototype learning,so that the features output by the model for each class are more compact and the features between classes are more different.The features are more compact and the feature gap between classes is more obvious.The addition of prototype learning makes the performance of the hypergraph convolutional neural network significantly improved,and greatly enhances the classification accuracy of the hypergraph convolutional neural network in the literature crosscitation dataset.The text hypergraph is also constructed,and the hypergraph convolutional neural network classification model based on prototype learning is applied to text classification,and good classification performance can be achieved with a small amount of data.(2)The graph convolutional neural network does not fully consider the similarities and differences of each vertex when processing text datasets,which limits the ability of the model to consider the information of neighboring vertices from different perspectives.Using the graph attention mechanism,different weights can be assigned to different vertices,and different weights can be used to focus on the vertex information that is more critical to the current task based on the attention mechanism when information is aggregated,but the graph attention mechanism However,the graph attention mechanism requires vertex features to calculate the degree of connectivity between each two vertices in the graph,and the dimensionality of vertex features is generally large and numerous,so this process is extremely computer resource intensive.The graph convolutional neural network constructs graph structure data with all text as well as vocabulary as vertices,so the graph it constructs is extremely large,so it is difficult for the graph convolutional neural network to use graph attention to process the graph structure data composed of text,and its algorithm complexity undoubtedly increases in cost.Therefore,this paper proposes a graph convolutional text classification model based on the attention mechanism.The traditional graph convolutional neural network text classification model updates the information of document vertices by using the vocabulary vertices connected with it,and the traditional aggregation approach does not fully consider the similarities and differences of document vertices for each vocabulary connected with it,while in fact,the importance of each vocabulary to the document is not the same.In this paper,we propose a lightweight graph attention mechanism to avoid the problem of excessive computational resource consumption,and build a graph convolutional text classification model based on the attention mechanism to guide the model to focus on more relevant words and extract features from the perspective of feature diversity to improve the performance of text classification.In this paper,we compare the proposed work with multiple sets of experiments and can achieve high results in each classification performance index,which validates the effectiveness of the proposed model in this paper.