Text Sentiment Analysis Based On Deep Neural Networks

With the advent of the information age,the network has accumulated a wide variety,a huge number of personal text,such as microblogs,comments of merchandise and services and other things,postings in the forum,and so on.These texts contain a lot of information,one aspect of which can not be ignored is the sentiment.Through the analysis of the text sentiment,we can get the speaker’s emotions,their attitude to an object,their standpoint of events,etc.,so that the information can be applied to such as the recommendation system,network public opinion monitoring and other fields,and then promote the social economy development,standardize the network environment with civilization and health,and contribute to the prosperity and stability of society.There are many ways to analyze the sentiment of the text,among which the Deep Learning method is the research hotspot in recent years.Among many deep neural networks,the Convolutional Neural Network is suitable for extracting the local information in the data,and the Recurrent Neural Network is more sensitive to the sequence information of the data.Therefore,with the continuous development of these network structures themselves,there have been many combinational models,combining with the advantages of these two models.But usually,these combination models use one model to extract the feature at the sentence level at first,and then another to extract the feature of the paragraph or chapter level,and finally use the classifier to classify.However,three models are proposed in this paper,combining these two models at the sentence level.In the first two models,the high dimensional features of the N-grams in sentence are extracted as first-level features by using the convolution kernels,and then the higher second-level features are extracted by the recurrent units,and then the classification is carried out.In the third model,the recurrent units are used to extract the timing informations as first-level features,and then use convolution kernels to extract more abstract second-level features,and finally the classification.A large number of experiments are designed in this paper.Through the processes of preprocessing,text vectorization,model training and testing,we choose the best result of each model from many different configurations to compare with previous research.The test results on the test set of NLPCC2013 Chinese Microblog Sentiment Analysis Task towards to 7-classification and 8-classification show that the three models proposed in this paper have obvious effect on improving the effect of text sentiment analysis.In addition,this paper designed a number of contrast experiments,and from the results,analyse the factors may affect the performance of the models in predicting and the suggestions for improvement.