Research On Fine-grained Sentiment Analysis Based On Commodity Reviews

The hundreds of millions of Internet users publish product reviews on shopping websites for purchased goods that contain a wealth of information,are often used as a basis for decisions about whether others buy the product and whether the product manufacturer upgrades the product.As a result,the emotional polarity of users for specific aspects of the product from rich user reviews has received a lot of attention.However,the diversity of user review products and the different emotional words used for different aspects in the same review have brought problems and challenges to the detection of sentiment analysis.In this paper,a fine-grained sentiment analysis model based on BHMAN(Bi LSTM-HCNN-Multi-Attention Network)is innovatively proposed,combined with the CS-LDA terminology extraction algorithm,it is proved by experiments that the CS-LDA algorithm can extract the aspect words in user reviews more effectively,and the accuracy of the BHMAN model in sentiment analysis has also been significantly improved.The main work of this paper is:(1)Aiming at user reviews containing multiple product aspect words,if the aspect words are extracted directly from the user reviews,resulting in the problem that some aspect words cannot be effectively extracted,a method of aspect word extraction based on sentence segmentation is proposed,which splits each sentence into sentence segments containing only one aspect according to the calculation of Bert semantic similarity.The CS-LDA algorithm is then used to classify the segmented sentence segments into different aspect word classes.Experimental results show that compared with traditional LDA and other LDA-based methods for extracting text,the accuracy of sentence segmentation-based aspect word extraction method is significantly improved.(2)Aiming at the problem that in fine-grained sentiment analysis,the traditional model only calculates the weight of each word in the aspect in the context,ignores the weight of the context in the aspect word,an improved dual attention mechanism is proposed.Firstly,the matrix of aspect words and context representations is multiplied to obtain the interaction matrix of aspects and contexts;After that,the rows and columns of the interaction matrix are normalized separately;The interaction matrix then averages the columns and rows separately to ignore words that are not important in aspect and context;Finally,by multiplying the mean matrix with the normalized matrix,we can obtain the weight matrix of the aspect word in context and the context in the aspect word,respectively.Experimental results show that the improved dual attention mechanism improves the accuracy of fine-grained sentiment analysis(3)Aiming at the problem that the effect of fine-grained sentiment analysis is poor due to the use of part-of-speech or semantic-level information in fine-grained sentiment analysis,a fine-grained sentiment analysis model combining part-of-speech information and semantic-level information is proposed.Experimental results show that the model can not only improve the accuracy of fine-grained sentiment analysis,but also bring the detection effect closer to the actual situation.