Text Feature Representation And Sentiment Analysis Based On Deep Neural Network

With the rapid development of the Internet and mobile Internet,network platforms such as entertainment and electronic commerce have brought massive unstructured text.These texts contain emotional polarity,and the sentiment analysis of these texts can be used in public opinion analysis,product theory and other tasks.In the face of these massive text data,manual analysis is not in line with the current situation,and it is necessary to use Natural Language Processing technology to mine effective information by computer.This thesis mainly uses deep learning feature combination to learn the deep semantic features of the text,and constructs a deep neural network through targeted improvement to extract the features of the text perform sentiment analysis on the text data.The main work of thesis are as follows:The thesis has analysed the mainstream deep learning model convolution neural network(CNN)and the long and short time memory artificial neural network model(LSTM)are analyzed.Based on the advantage of CNN and LSTM,a tree-structured LSTM network is combined with on the basis of CNN to construct the sentiment analysis model(Att-CL)by combining their output layers.The model uses attention mechanism to combine word vectors.The combined word vector are used as input of CNN to solve the problem of long distance correlation between words in a sentence.And learning the semantic information of the whole sentence through the tree recursive LSTM enriches the vector representation of sentence features.This model can not only discover the characteristics that have strong relevance to the task of sentiment analysis,but also learn the semantic features of the text according to the grammatical structure of the sentence.The experiment shows that the hybrid neural model is better than the single neural network model on the micro-blog sentiment analysis task after cooperative training.The other work of the thesis is to propose a deep neural network model which combines semantic rules and bidirectional LSTM with a text sentiment analysis.The model uses bidirectional LSTM to learn the deep semantic features of the text in sequence combination,and obtains multiple emotional features of the text through semantic rules,and integrates the emotional features through the output of the full connection layer and the output of the hidden layer of bidirectional LSTM.Through the bidirectional LSTM uses the word vector as the input to learn vector for each sentence,then the sentence vector as input of docuemnt modeling to obtain the vector representation,emotional features obtained by semantic rule method are merged with the deep semantic features,feature fusion will eventually incoming classifier.Compared with some deep learning model experiments,the performance of the model based on the combination of bidirectional LSTM and grammar rules is better than other deep learning models.