Chinese short text classification method based on word embedding and Long Short-Term Memory Neural Network

The value of Internet short texts is increasingly prominent, and traditional classification methods cannot be applied to short texts with weak feature expression. In this regard, this paper proposes a Chinese Short Text Classification method based on Word Embedding and LSTM with feature enhancement (hereinafter called CSTCFE-WE-LSTM). This method uses word embedding learned from Wikipedia corpus as initial features for the model, and then uses category factors and TF-IDF to generate weights to enhance features. Finally, it uses a 6-layer neural network for classification, which includes a word embedding layer, two LSTM layers, a Dropout layer, and two fully connected layers. In order to verify the method CSTCFE-WE-LSTM, we collected short text sets on 3 topics, and reached following conclusions: 1. In the best model, P, F and other indicators are better than the classifier of using Wikipedia word embedding and KNN. 2. For texts with sentences average length less than 10 words, the effect of two-layer LSTM is better than that of single-layer LSTM, and the effect is better when the number of single-layer nodes is 50. 3. The effect of feature enhancement is better for the health category than the commercial category.