Research On Chinese News Text Classification Based On Nested LSTM

Text classfication is one of the most important research tasks in the field of natural language processing.With the development of internet technology,text classfication plays a crucial role in various application systems.The characteristics of Chinese news texts affect the results of classification tasks seriously.The characteristics include: Text sentences are composed of various phrases.The feature extraction of these phrase structures helps to learn the feature representation of texts,thus improving the classification accuracy;The Chinese news text has a long length and a lot of content.In order to improve the expressive ability of the text feature representation to the original text,the classification model should be able to learn the contextual dependence of long text;The text usually contains a large amount of redundant information irrelevant to the subject,which causes interference to the classification task.To effectively reduce such interference,the classification model should pay more attention to important information as much as possible.Therefore,two deep learning models for Chinese news text classification are proposed for the above features:1)Combining the advantages of CNN model and NLSTM model,this paper proposes a CNLSTM model based on attention mechanism for Chinese news text classification,which uses a convolutional neural network to extract the phrase feature sequence of the text,and expands the original word-based model into a phrase-based model.Since the nested LSTM can access the historical information of a longer time scale than the ordinary LSTM,the obtained phrase features are input into the nested LSTM model to extract the feature representation of the entire text;In order to highlight the key information of the text,the model uses the attention mechanism to calculate the degree of influence of each phrase feature on the category of the entire text,and generates a text feature representation containing the attention distribution.Finally,the Softmax classifier is used to classify the input text.The effectiveness of the model is proved by experiments,and the model can not only capture the local features of the sentence,but also fully consider the semantic information of the text context,so as to extract more reasonable text features,and the use of the attention mechanism makes the model effectively avoid the redundancy of the unreasonable impact of the remaining information on the output category in the input long text.2)Improve the attention-based CNLSTM model and propose an improved model based on reinforcement learning.The CNLSTM model uses a fixed-size convolution kernel to extract the phrase information.But in fact,the Chinese text has a complicated structure,and the length of the phrase usually is random and varied.At the same time,since the sliding step size is usually set to 1 when the feature is extracted by CNN,the obtained phrase feature sequences contains a lot of duplicate information.Therefore,this paper proposes to replace the CNN structure with an reinforcement learning algorithm,so that the model can autonomously recognize the phrase structure of the text during the training process.The improved model based on reinforcement learning first uses the stochastic strategy gradient algorithm to learn the action set corresponding to the text words,and then learns the text feature representation based on the phrase structure through a hierarchical NLSTM model.Similarly,the improved model also uses the attention mechanisms to make the model more focused on content that is closely related to the news topic.It is proved by experiments that the classification effect of theimproved model is better than the CNLSTM model.