The Research On Chinese Sentiment Classification Using Recurrent Neural Network And Convolutional Neural Network

The rapid development and popularity of Chinese social network platform and online-shopping platform leads to the explosively increasing number of Chinese texts available on the Internet.So,it becomes more and more important that how to organize and utilize these texts to obtain preference,opinion and affective tendency of users.Sentiment analysis is performed to extract,formalize,and analyze the opinion or subjective knowledge from texts for specific use.However,it is a challenging job because the contextual information of such short texts is limited.Therefore,it is of great significance to effectively solve sentiment analysis problem and extract effective information from sentences in a more disciplined way.The traditional machine learning algorithms are based on manually selected features,which seldom capture the context dependency or semantic binding.Considering these problems and disadvantage,we proposed the Bidirectional Long-Short Term Memory(BLSTM)with word embedding for Chinese sentiment analysis.BLSTM can learn preceding and succeeding information and capture stronger inter-dependency.Word embedding tries to denote a word as a vector,which implies the syntactic and semantic information based on large-scale unlabeled dataset.Experimental results show that our model achieves 91.46%accuracy,increased by 3.12%compared to the conditional random field method.For long short-term memory networks,its next output depends on the previous hidden state.So it is inconvenient to parallelly process the elements of a sequence,resulting in its long training time and poor training efficiency in large-scale Chinese sentiment datasets.Therefore,this thesis studies the application of convolutional neural network to Chinese sentiment analysis,and proposes a convolutional control block model.Our model can capture short and long context dependency by using two parallel convolutional layers with different kernel size.And followed we use a convolutional layer with kernel size 1 to determine how much information of the current convolutional block needs to flow to the next layer.We call that convolutional layer with kernel 1 as convolutional control gate.The experimental results show that the precision and F1-socre of our model reach 93.19%and 92.67%on positive reviews,respectively,with F1-socre 92.09%for negative reviews.Compared with the LR_all and the very deep convolutional neural network model,F1_score of our model increases 1.84%and 1.90%on positive reviews,while 1.20%and 1.57%on negative reviews.