Research Of Text Sentiment Analysis Methods Based On Neural Network

With the rapid development of the Internet information technology,social networking sites and e-commerce platforms have brought great convenience to people.At the same time,a huge numbers of users can publish,share and update the comment text information anytime and anywhere.These unstructured and attitude-oriented texts have sentiment polarity,and sentiment analysis of these massive texts can be used for product reviews,public opinion analysis and other tasks.It is unrealistic to rely on manual processing of such massive text information,Therefore,it is necessary to quickly and accurately obtain the sentiment polarity of the text by means of a computer.Neural networks have developed rapidly in recent years.It is inspired by the highly developed human brain.The nonlinear hierarchical network structure is used to approximate the representation of complex functions,and then the deeper level features of the target object can be learned,so it has been extremely successful in the field of sentiment analysis.Therefore,deep learning technology based on neural network is the main research direction.This paper mainly uses neural networks to learn the deeper semantic features of texts,and the extracted text features are used for sentiment analysis by improving the recurrent neural network model.The main work of this paper is as follows:(1)Tree-Structured Long Short Term Memory Networks(Tree-LSTM)in recurrent neural network applies the tree-structure recursive neural network to the LSTM.This model is based on its vector representation when calculating the sentiment polarity of a node and lacks the sentiment information of the child nodes.To solve this problem,this paper proposes a parallel recursive tree long short term memory model(PRT-LSTM),which combines tree recursive neural network with Tree-LSTM,and the sentiment information of node is propagated through tree recursive neural network to the global participation sentiment polarity calculation.Therefore,the sentiment information of each node in the PRT-LSTM is composed of the sentiment information generated by the Tree-LSTM and the information of the recursive model.In this way,it can make more full use of the sentiment information of nodes.PRT-LSTM was performed on the Stanford Sentiment Treebank for Fine-grained and Binary tasks.Experimental results manifest that the proposed model achieves a higher performance than the Tree-LSTM.(2)This paper proposes a hybrid network model Att-CGRU based on Convolutional Neural Network(CNN)and gated recurrent unit(GRU).We improve the convolutional neural network to Att-CNN that use the Attention mechanism to combine the input layer's word vector.Att-CNN can solve the long-distance correlation problem between words and use for deeper feature extraction of text.GRU is used to learn semantic correlation information between words.The paper designed the hybrid network model to experiment in sentiment classification.Experiments show that the model performs better than the single neural network model after training.We also design experiments that applied different filters to verify the accuracy of the model.The experimental results show that the sentiment classification performance is best when the convolutional layer is single and the filter length is 3.