Research And Implementation Of Sentiment Analysis Technology In Public Opinion System

With the popularity of the Internet and mobile networks,people have become accustomed to paying attention to hot Internet events and expressing their personal emotions online.Hot events have brought a huge amount of public opinion text data.How to analyze the emotion orientation of users based on these data is a hot and difficult research topic.At the same time,sentiment analysis,which helps system users understand the full picture of the event and analyze the trend of the event,is a key step in the public opinion system,Public opinion system is a tool that can automatically collect and analyze online public opinion data.The system uses a large-scale distributed crawler to collect multi-source public opinion data from the Internet automatically.And then it uses the processing module to process and analyze the original text data.It provides users with various reports,views and summaries to help users understand the public opinion events in a comprehensive manner progress.Analyze the emotional orientation of the event is a very critical part of the public opinion system.Granularity and accuracy are important factors that can affect the result of sentiment analysis.At present,most existing systems use the positive,neutral,and negative emotion classification methods,which are difficult to express the complexity of human emotions.At the same time,they pay a modicum of attention to the processing of complex contextual relations in multilingual text and limit the max length of the text.Aiming at the above problems,this article uses the public opinion system as the application background and conducts research on sentiment analysis technology.Two sentiment analysis algorithms are designed for code-switching microblog short text and news long text.Designed and implemented the sentiment analysis function related modules of public opinion system.The specific work includes the following aspects:1)Propose a multi-dimensional sentiment analysis method of code-converted microblog short text based on the BERT pre-trained language model,aiming at the problem that the code-switching text contains multiple languages and multi-dimensional emotions.This method improves sentiment multi-label input and output methods and uses different pre-trained models for the characteristics of language diversity of the code conversion task.It is verified on the public data set of the NLPCC2018 conference.The F1 scores of experimental results are better than many sentiment analysis methods in five emotional dimensions(happiness,sadness,anger,fear and surprise).And the average improvement over BCEL is about 0.09.2)Propose text truncation and text segmentation methods for news long text sentiment analysis tasks,aiming at the problem of limited text input length of the BERT model.The text truncation method includes three ways: truncating the head,truncating the tail,and truncating the head and tail.The text segmentation method can split the text to blocks or sentences.The pooling or padding method is used to fix the dimension of the text representation.The bidirectional LSTM model based on attention outputs classification results.The news long text task is verified on the INEWS Internet sentiment analysis task data set.The text segmentation method achieves better results than the unmodified BERT model,with an accuracy improvement of about 0.02.3)The related modules of sentiment analysis function in public opinion system are designed and implemented based on the above algorithms.The modules include sentiment analysis module,crawler module,message middleware module,data processing module and etc.Realize the data acquisition layer and data processing layer in the system,so that the public opinion system can provide the complete sentiment analysis function.