Study On Text Categorization Method Based On Graph Convolutional Networks

Text categorization is a classic and important topic in natural language processing.Today,a large amount of textual information is generated every day,and in order to use these resources more effectively,accurate text classification is very significant.Many existing text categorization methods focus on how to represent text better,then traditional algorithms or deep learning methods are utilized for classification.However,a core assumption of those methods for text categorization is that documents are independent of each other and the link information among documents and words are not taken advantage of.So those methods can not deal with the datasets like citation networks directly.In fact,the relationship between texts is very important,which helps to improve the accuracy rate of categorization.In this thesis,text classification methods based on graph convolutional network are proposed.First,the basic process of text categorization and the required preparatory knowledge are introduced.Sparse representation is used to construct the relationship between texts and the important of the edge between a document and a word is the TF-IDF of the word in the document.Thus,a graph-structured data containing texts and words is constructed,and the adjacency relationship is represented by an adjacency matrix.Next,based on the spectral theory,convolutional neural network is generalized into a graph convolutional network that can directly operate graph-structured data,and a semi-supervised text categorization method based on graph convolutional network is proposed.Experiments on different datasets have achieved good result.Finally,the graph representation of the text is obtained from the learned word embedding and a parameterized graph pooling module is introduced.Then,a supervised graph classification model is proposed for text classification.Experiments on the MR dataset show the effectiveness of the method and indirectly prove the quality of the previous word embedding.The research results of this thesis mainly include the use of sparse representation to construct graph-structured data,the introduction of parameterized graph pooling module,and two text classification methods based on graph convolutional network: the first one is semi-supervised node classification model,and the second one is a supervised graph classification model.Two methods have achieved good results.