Multi-Theme Mixed Chinese Reviews Sentiment Analysis

With the development of the Internet,a large number of commentary texts published through the We Media have emerged.These comments include both commodity reviews from e-commerce websites and views or opinions about what they have experienced through We Media.According to these comments,many problems can be solved,such as assisting user consumption decision,helping merchants optimize commodity,and analysis of internet public opinion,and so on.However,the sentiment analysis of the whole comment can not help users reduce the information overload and cognitive cost in terms of commodity attributes.Therefore,the sentiment analysis of multiple topics in the commentary has received extensive attention.Multi-topic mixed sentiment analysis includes two important sub-tasks: opinion target extraction and sentiment polarity analysis.In view of the defect that the traditional sentiment analysis method can't refine the user sentiment tendency and make clear the commodity attribute,combined with the characteristics of Chinese language,this paper combines Chinese word internal location and Part of Speech(POS)information in the process of opinion target extraction,introduces modifiers information and deepens the dependence between opinion target and sentiment words in the process of sentiment polarity analysis.Through the deep neural network model,the sentiment analysis of multiple sentiment topics in Chinese comments is carried out.The main research work and innovation of this paper are as follows:(1)To study the influence of Chinese word internal location and POS information on the opinion target extraction.Firstly,the internal information of the words is analyzed,and two representations are distributed according to the inner position of the word according to the character representation optimization strategy,so that the vector representation of different semantics of the same character in different words is realized,and the Chinese word internal location information omitted in the sequence labeling process is made up.Second,it is important to take into account that the word sex is an important grammatical information,and the word of the same word often plays a similar role in the language.Therefore,by using the word-character annotation to the Chinese comment,the paper uses the neural network model to study the speech characteristics to obtain the grammatical constraint of the Chinese comment,and to deepen the understanding of the POS information by the neural network.(2)The advantage of the sequence labeling model in the evaluation of the object is analyzed.And after the opinion target extraction problem is converted into a sequence labeling problem,a Bi-LSTM learning text sequence characteristic is introduced to the extraction model to fully capture the fused word internal location and the POS information,and the matching CRF layer overcomes the label deviation,thereby improving the accuracy of the opinion target extraction.In addition,for the result of the dimension,the evaluation entity can't be matched,the current situation of the sentiment word pair is not matched,and the dimension label is optimized.Adding a mark on the basis of the labeling label of the BIO,and recording whether the currently marked object has a matching<evaluation entity,sentiment word>,so that the labeling result is provided with the structural characteristic and the result of opinion target extraction is optimized.(3)The study of sentiment polarity analysis of the fusion sentiment modifier and the evaluation entity.Since the opinion target extraction is finished,only the emotion tendentiousness judgment of the extracted emotion words is needed,and the characteristic information is seriously lacking.On the one hand,using existing emotional resources and building an emotional modification set in conjunction with the comments text.The modified content is added from the negative word,the degree level word,the turning word and the virtual tone,and the modifier is used as the emotion element information of the emotion word.And the modifier is used for generating the emotion phrase in the form of a prefix,and enriching the emotional characteristics.on the other hand,taking into account that the emotion tendency of the emotion word is seriously dependent on the object to be evaluated,the opinion target corresponding to the emotion word is taken as the second important emotion element,and the emotion phrase is added to the emotion phrase in the form of a prefix,And finally,the classification performance of the emotion tendency analysis model is improved.