Research On Sentiment Classification Based On The Method Of Deep Learning

The continuous high-speed development of the Internet has brought us a convenient life.Various news,social,entertainment,and e-commerce websites have emerged in an endless stream.We can read,shop,and communicate on the Internet anytime.In these sites,a large amount of commentary text is frequently produced,i.e.a text describing the opinion or opinion of a certain thing.Digging and sorting out this type of text helps us to identify the product,business to improve their service levels,and the government to control the direction of public opinion.Most of these review texts contain explicit sentiment information.How to accurately classify these review texts is the main task of this thesis.With powerful feature automatic extraction,deep learning has been widely used in speech recognition,machine translation,image recognition,human-computer interaction and other fields.In the study of sentiment classification,due to the difficulty and complexity of text feature extraction,the traditional machine learning method has gradually been replaced by deep learning.Based on the basic neural network,the accuracy of sentiment classification is improved by adding sentiment vectors of words.The main works in the thesis are:First is to identify the sentiment words in the relevant filed and to calculate the strength of the sentiment words.The review texts in different fields may contain some field-related words.In order to predict the sentiment tendency of text accurately,we need to identify field-related words.Based on the unsupervised thought,the thesis realizes the automatic recognition of sentiment words by constructing universally applicable models.Among them,the process of recognition adopts the “differential co-occurrence method”,which is designed based on the word co-occurrence and the different frequency in different sentiment categories.In addition,in order to further improve the accuracy of the sentiment classification,the advantages and disadvantages of How Net and word2 vec is analyzed in the thesis,and finally the combination of the two is used to calculate the sentiment strength of sentiment words.Among them,the calculation of sentiment strength adopts the “TF-IDF based seed word” method,which is designed based on the principle of word similarity and the importance of the words under the current corpus.The second is to build sentiment vectors of words and embed them into the deep learning model.In the sentiment classification model based on deep learning,researchers are keen to build various neural network models,but sentiment words play an important role in the judgment of sentiment tendencies and should not be completely abandoned.The thesis first constructs three-dimensional sentiment vector based on the three main features of sentimental expression,including sentiment words,negative words and degree words,to expand the sentiment information of words on sentiment classification tasks and embeds them into the basic neural network.In the network model,naive-CNN,separate-CNN,naive-LSTM,and separate-LSTM combinations are proposed.Experimental results show that embedding sentiment vector to the basic neural network model helps to improve the accuracy of sentiment classification.