Chinese Short Text Sentiment Analysis Based On Deep Learning

With the continuous development of the Internet,the number of online commentary platforms has increased and the number of user reviews has also grown explosively.The use of sentiment analysis technology can effectively mine the sentiment information contained in texts.It has become an important way for public opinion supervision and manufacturers to obtain feedback information.It has a high research value.The purpose of this thesis is to explore the sentiment information contained in Chinese short texts,mainly to solve the problem of sentiment polarity recognition and derogatory classification.There are mainly two traditional methods: sentiment lexicon based methods and machine learning based methods.However,due to the short text corpora and the large number of unregistered words,data sparseness exists in these methods,and they rely too much on domain experts.In recent years,deep learning technology has been able to solve these problems.Therefore,this thesis uses the deep learning method to analyze Chinese short texts.The main research content is as follows:First of all,in the text data preprocessing process,based on the current network exist a large number of the unknown words,we design a new words discovery method,mainly through the use of words of internal degrees of solidification and boundary freedom to filter candidate words.New words will be added to the lexicon,and the accuracy of segment will be improved.Secondly,the traditional word embedding only takes into account the semantic grammatical information in the text,and maps the semantically opposite sentimental words to the adjacent positions,resulting in the error of the final classification result.In order to solve this problem,this thesis combines sentiment information based on the traditional word embedding,and proposes a method of generating the sentimental word embedding.Finally,for the problem that the recurrent neural network cannot learn long-distance dependence information,this thesis proposes a sentiment analysis model based on GRU(Gated Recurrent Unit),and replaces the hidden layer nodes of the recurrent neural network with GRU units for sentiment analysis.This thesis first uses Java for new word discovery task,and a total of 9204 new network words have been discovered.Secondly,this thesis uses the Keras deep learning library in Python to build the sentiment analysis model based on GRU.Experiments are performed under different parameters to find the best parameters.The classification accuracy is 92.01% under the best parameters.It is compared with other existing traditional machine learning models and deep learning models.The results show the sentiment analysis model based on the sentiment word embedding as input has achieved better results under all indexes.