Chinese Text Classification Based on Deep Learning and Attention Mechanism

Facing the explosive growth trend of various text data on the Internet, it means a lot to make effective use of these text data and dig out the real value behind them. For text classification, traditional methods mainly pay more attention to shallow machine learning. Rapid development of deep learning technology brings about breakthroughs in the fields of image recognition, speech recognition, and the feature learning ability of deep model has been further proved. This thesis studies the news text classification based on deep learning. Convolutional neural network (CNN) can only obtain local features of text, while bi-directional long-term and short-term memory network (Bi-LSTM) can extract global features, without considering the influence of local features on text classification results. This thesis proposes a Chinese text classification method (AttBiLSTM_CNN) based on the mixed model of CNN and Bi-LSTM. This method takes words as the basic unit of text representation and constructs two channels in the feature extraction layer. CNN and Bi-LSTM are used to extract the local features and global features of the text respectively. However, considering that news texts are generally long, sometimes considering all features will reduce the classification efficiency, attention mechanism is introduced to highlight the key features of the text on the basis of retaining the original features, and a more accurate text feature vector representation is obtained. Then, the local features and global features are fused to generate feature vectors with rich semantics, which are sent to Softmax. Experimental results show that this method can effectively improve the accuracy of Chinese text classification.