Abstract: Text classification is the basis of many natural language processing tasks.Convolutional neural network (CNN) can be used to extract the phrase level features of text,but it can't capture the structure information of text well;Recurrent neural network (RNN) can extract the global structure information of text,but its ability to capture the key pattern information is insufficient.Attention mechanism can learn the distribution of different words or phrases to the overall semantics of text,key words or phrases will be assigned higher weights,but it is not sensitive to global structure information.In addition,most of the existing models only consider word level information,but ignore phrase level information.In view of the problems in the above models,this paper proposes a hybrid model which integrates CNN,RNN and attention.The model considers the key pattern information and global structure information of different levels at the same time,and fuses them to get the final text representation.Finally,the text representation is input to the softmax layer for classification.Experiments on multiple text classification datasets show that the model can achieve higher accuracy than the existing models.In addition,the effects of different components on the performance of the model are analyzed through experiments.

Key words: Global structure information, Hybrid model, Key pattern information, Text classification, Text representation