| With the continuous evolution of intelligent technology and the development of social network platforms,various information data have shown the situation of explosive growth.These massive text data and various forms of information carriers have increased the difficulty of manual classification and management.Therefore,how to use artificial intelligence technology to improve the classification accuracy of text information has become one of the current hot research directions.The text classification task,which aims at accurately extracting the important features required by the model and completing the efficient classification processing of data,has studied by many scholars.Among numerous deep neural networks,the graph neural network(GNN)has been widely used for text classification because of its superior performance in classification accuracy.Therefore,this thesis studies the text classification model based on the graph neural network,and proposes corresponding solutions to the problems of overfitting,feature sparseness and insufficient feature diversity in the model,to further improve the accuracy of text classification.The specific work of this thesis includes the following three aspects:1.Aiming at the problem of over fitting when the training data is insufficient due to the application of image-text conversion features for classification in graph neural network,this thesis proposes DASA-GNN(Graph Neural Network based on Data Augmentation and Self-Attention)that integrates graph neural network,data augmentation method and self-attention mechanism technology.The idea of constructing the model is: while retaining the graphics coding characteristics in the graph neural network,data augmentation technology is used to enable the model to maintain relatively stable performance under the condition of a small amount of training data;by introducing self-attention mechanism,the word-level interconnection in the original text is strengthened,and the text information extraction at the sentence level is enhanced,which effectively alleviates the over-fitting problem of graph neural networks in text classification task.The experimental results of the model on different data sets show that the DASA-GNN model can effectively improve the performance of the network and improve the accuracy of text classification.2.In order to further improve the classification accuracy of the model,this thesis considers mining deeper feature information,and designs to introduce regional word embedding technology into the DASA-GNN model,which retains the internal structure information of the original data as well as adds regional feature information,thus proposing DASA-Re GNN(Regional Embedding GNN based on Data Augmentation and Self-Attention).The improved DASA-Re GNN model further strengthens the relationship between word levels,effectively captures high-order neighborhood information and alleviates the impact of data sparsity on model performance.Experiments on different data sets show that the DASA-Re GNN model can further effectively improve the classification accuracy.3.The lack of diversity information in text data may affect the feature extraction ability of graph neural network,thus affecting the performance of text classification.Based on this problem,this thesis designs to introduce attention mechanism into the network to make it ignore irrelevant features and focus on effective features,and proposes Att-DASA-Re GNN(Regional Embedding GNN based on Data Augmentation and Self-Attention with the Attention Mechanisms).The model introduces hard attention mechanism,soft attention mechanism and multi head attention mechanism into Att-DASA-Re GNN,which can extract text vocabulary information with different characteristics during training process,and thus optimizes the classification effect.The experimental results confirm that the Att-DASA-Re GNN model equipped with attention mechanism is able to pay attention to the diversity of text feature information,so as to show excellent performance in text classification task. |
