Sentence Sentiment Classification Based On Deep Learning

With the popularity of modern networks,personal reviews of products,services,or events can enable organizations and businesses to improve marketing,communications,production,and purchasing.The sentiment classification of sentence facilitates extracting subjective information from the user-generated comments or narratives,and generally determines the sentiment polarity of sentence.The traditional sentiment analysis methods obtain the sentence representation only by counting the frequency of words or phrases.Although Word embedding is widely applied to sentence sentiment classification based on deep learning,it can't reflect grammar and syntactic structure which directly affect the classification performance of sentence.In this paper,Recurrent Neural Network and Tree Structure Network are improved to obtain the sentence representation which can display the semantic features of the sentence better and being used on sentiment classification on the review data on the network.The main work and innovations of this dissertation are concluded as follows:(1)A polarity shifting tree-structured LSTM networks model is proposed.Tree-structured LSTM can obtain the syntactic structure information of sentences well,but neglect the polarity shifting information of words.A polarity shifting tree-structured LSTM networks model is proposed to solve this problem.The memory module in LSTM is extended to the nodes in dependency tree or constituency tree network which is constructed at first to store the semantic information for a long time.In order to obtain the polarity shifting information of nodes,the sentiment polarity shifting vector is added to the memory module.The accuracy of sentence classification in Stanford Sentiment Tree-bank dataset is improved compared with previous deep learning models.(2)A tree-structured networks model based on rhetorical structure theory,gated recurrent units(GRU)network and attention mechanism is proposed.Dependency tree and constituency tree can obtain the syntactic structure and enrich the semantic expression of the sentence well,but can't reflect the importance of the fragment in the sentence,and the memory module in the GRU network is more simple and effective than in LSTM.A gated recurrent units(GRU)network model based on rhetorical tree is proposed which utilize the rhetorical structure theory to analyze the importance of subtrees and utilize the memory module in GRU to store this information to get more accurate sentence representation;Sentence representation depends on the fixed dimension vector of the root in tree-structured network.In order not to be limited to this,attention mechanism is added to the model in this paper.The sentence representation is acquired from integrating all the nodes in the tree structure,which enhances the expression of core information and features of sentence.Experimental results show the proposed methods can achieve better accuracy in comparison to the state-of-the-art models.