Research And Application Of Aspect Level Sentiment Analysis Based On Memory Network With Word Sentiment Vector

Among the algorithms of aspect-based sentiment analysis(aspect level sentiment analysis),deep memory network has the advantages of simple structure,fast running speed,build-in attention mechanism.The attention mechanism of deep memory network can analyze the importance of each context word when inferring the sentiment polarity of a commented target.However,the attention mechanism of deep memory network is either content-based or location-based,the part of speech(POS)is ignored.In addition,when processing text,deep memory network needs to convert words into word embeddings.The common-used algorithms for learning word embeddings can capture the syntactic information and statistical information of words in the context,such as C&W?Skip-Gram?CBOW and GloVe.But?they cannot express the sentiment polarities of words.Although there some word embeddings can express the sentiment polarities,they are not perfect and cannot express the sentiment polarities of a word under different part of speech.To solve these problems,this paper first proposes two new attention mechanisms that using part of speech.And Word Sentiment Vector(WSV)is proposed,which can express the sentiment polarities of a word under different part of speech.Then,the proposed attention mechanisms and WSV are applied to deep memory network,and a new memory network called Memory Network with Word Sentiment Vector(MNWSV)is proposed.Finally,MNWSV is applied to the sentiment analysis of restaurant reviews.The study of this this paper can be summarized as follows.(1)To improve the performance of deep memory network,the part of speech is incorporated into the original attention mechanism of deep memory network,and an attention mechanism called Attention Mechanism based on Position,POS and Content(AM-PPOSC)is proposed.Considering CNN has good local feature extraction ability,this paper also proposes an attention mechanism called CNN Attention Mechanism based on Position,POS and Content(CNN-AM-PPOSC).Then the two attention mechanisms are applied to deep memory network,and two new memory nerworks are proposed which called Memory Network based on AM-PPOSC(MN-AM-PPOSC)and Memory Network based on CNN-AM-PPOSC(MN-CNNAM-PPOSC).Experiments on two datasets have shown that both AM-PPOSC and CNN-AM-PPOSC can effectively analyze the importance of each context word when inferring the sentiment polarity of a commented target.Compared to others algorithms,MN-AM-PPOSC achieves the highest accuracies on both two datasets.(2)To better express the sentiment polarities of words,Word Sentiment Vector(WSV)is proposed.And WSV is applied to MN-AM-PPOSC,Memory Network with Word Sentiment Vector(MNWSV)is proposed.Experiments on two datasets have shown that WSV not only can effectively express the sentiment polarities of a word under different POS,but also can improve the accuracies of several algorithms.Compared to others algorithms,MNWSV achieves the highest accuracies on both two datasets.In addition,compared with sentiment word embeddings SSWEh,SSWEr and SSWE,memory networks and several other algorithms use WSV can achieve higher accuracy.(3)Finally,MNWSV is applied to the sentiment analysis of restaurant reviews.Analyzing and counting the percent of positive comments and negative comments towards to the aspects of restaurants to help restaurants find their shortcomings and make improvements.Considering the relevance between star ratings in the reviews and the feelings of customers towards to the aspects of restaurants,the star ratings are incorporated into MNWSV,and a model to analyze the sentiment polarities of restaurant reviews is proposed.The model includes five parts:Aspect Determination,Data Collection,Data Preprocessing,MNWSV that incorporates star ratings and Model Evaluation.Experimental results show that MNWSV,which incorporates star ratings,has achieved quite good results in the sentiment analysis of restaurant reviews,the accuracy is 98%.