Abstract: Existing text classification models based on graph convolutional networks usually simply fuse the neighborhood information of different orders through the adjacency matrix to update the representation of node in graph,resulting in insufficientrepresentation of the word sense information of the nodes.In addition,the model based on conventional attention mechanism only provides a positive weighted representation of the word embedding,ignoring the impact of words that produce negative effects on the final classification.To overcome the above problems,a model based on bidirectional attention mechanism and gated graph convolutional networks is proposed in the paper.Firstly,the model uses the gated graph convolutional networks to selectively fuse the multi-order neighborhood information of nodes in the graph,retaining the information of previous orders,to enrich the feature representation of nodes in graph.Secondly,the model learns the influence of different words on text classification results by the bidirectional attention mechanism,giving positive weights to words with positive effects on the classification and negative weights to words with negative effects to weaken their influence in the vector representation,to improve the model's ability to distinguish nodes with different properties in the document.Finally,the maximum pooling and average pooling are used to fuse the word representation in text to get the document representation for the final classification,where the average pooling can make each word play a role in generating a graph-level representation of the document and the maximum pooling can make the important words play a greater role in document embedding.Extensive experiments on four benchmark datasets show that the proposed model significantly outperforms the baseline model.

Key words: Text classification, Graph convolutional networks, Attention mechanism, Text representation, Deep learning, Natural language processing