| Under the era of the big data,the use of information technology and computers has gradually normalized,especially the rapid popularization of mobile phone,making the network an important way to obtain information and express opinions.The network public opinion is formed because of the occurrence of some events and is discussed by many people on the network.In recent years,deep learning affective analysis has been widely used in text analysis,but most of them are analyses of movies or commodity reviews,while few of them are applied to analyze the Chinese text tendency and the collection of online public opinions.Therefore,in order to judge the public's opinion tendency on events and help the government to make relevant decisions in a timely manner in a certain special period,it is of great significance to analyze the text tendency of online public opinions.For obtaining the emotional tendency of social situation and public opinion in time,this thesis studied and designed a text tendency analysis system based on in-depth learning.The web crawler was intended to be used to crawl the text information of the network public opinion of a specific website,after the corpus is preprocessed,the deep learning framework were used to determine the text positive and negative tendencies of online public opinions.This paper presents the overall framework of the system,focusing on the effectiveness of different deep learning frameworks in the prediction of online public opinion tendency.In order to find the best combination for public opinion prediction,this thesis using the public opinion tendency database which has been manually judged by a department as the standard data set(including 99168 public opinion texts),applying different deep learning frameworks,including LSTM?BLSTM?GRU?CNN,to train and verify the data,and at the same time,train and test the activation function(sigmoid,tanh,Re Lu)and optimizer(Adam,Ada Grad)of the framework.CNN and BLSTM were combined to construct CNN + BLSTM framework by using the characteristics of CNN that can capture features and BLSTM that can read contexts.Similarly,CNN + LSTM and CNN + GRU were constructed together with the basic four frameworks to form 42 combinations.The optimal framework and combination were found by using verification accuracy,verification loss rate,over-fitting phenomenon and consumption time as evaluation basis.The results showed that different activation functions and optimizers have great influence on the model,and the sigmoid activation functions are the most suitable functions for this system.CNN(sigmoid + Adam)was the model with the highest verification accuracy,with 97.84%,followed by CNN + BLSTM(sigmoid + Ada Grad)model with 97.72%.The validation accuracy of CNN + LSTM is 97.5% respectively when the activation function sigmoid is matched with the optimizer Ada Grad.Therefore,the combination model CNN + BLSTM?CNN + LSTM were better than the single model when the activation function sigmoid is matched with the optimizer Ada Grad.Comprehensive analysis showed that although CNN + BLSTM has a high verification accuracy rate,it takes too long,which will seriously affect timeliness in practical application,and there is a big difference between the different results of activation function and optimizer.However,among the 7 models,only 6 combinations of LSTM model had a verification accuracy rate of more than 95%,which is short in time,high in accuracy and low in loss rate.Therefore,this paper believed that LSTM model is more suitable for practical application.Finally,this paper used LSTM model to test and analyze the Weibo topic of "Xi'an No Smoking".The accuracy rate was 89.58%,and the analysis found that the sentences with ironic tone and mispronounced words were wrong in judgment.The reason may be that the data sources of the training model are mostly positive news reports,and the Weibo data of netizens expressing their ideas account for a relatively small proportion.Therefore,the LSTM model can basically correctly judge the opinion tendency. |
PDF Full Text Request