Research Of Aspect-level Sentiment Analysis Based On Deep Learning

With the development of network information technology,the number of netizens in China is increasing.The 44 th "Statistical Report on Internet Development in China" indicates that as of June 2019,the number of netizens in China was 854 million,an increase of 25.08 million compared with the end of 2018.The penetration rate reached 61.2%.Netizens can express their opinions through a variety of channels,which continue to accumulate massive amounts of textual data.From these texts,it is a very meaningful work to analyze people's views on entities.Taking online shopping reviews as an example,the traditional sentiment analysis only analyzes the emotional tendency of a review as a whole,and does not involve the calculation of the emotional tendency of the commodity attributes contained in the review.This kind of sentiment analysis not only leads to insufficient information extraction,but also does not recognize the emotional tendency of the product attributes that the user is concerned about.In view of the above problems,this paper introduces related recurrent neural network and convolutional neural network,and studies aspect-level sentiment analysis based on deep learning.In the research of aspect-level sentiment analysis based on deep learning,scholars generally optimize existing neural network models based on public data sets,and there are few researches in applied fields.This article crawls the flagship machine reviews of the four domestic mobile phones from the webpage as experimental data,analyzes the flagship machine's comment text,uses dependency parsing to extract product attribute feature words,and summarizes ten common syntactic paths.In the sentiment analysis of the text,the Word2 Vec,Glove and BERT models are used for word embedding,and the obtained word vectors are input into the three mainstream deep learning models of TextCNN,BiLSTM and AT-BiLSTM for emotion classification.The comparison of classification results shows that the BERT model has the best word embedding effect,and the classification effect improved significantly.The macro-precision,recall rate and F1 value of the AT-BiLSTM model are all above 96%,and the classification effect is optimal.Finally,this paper summarizes the competitive advantages and disadvantages of various domestic mobile phones,provides users with the basis for product selection,and also proposes three suggestions to manufacturers: improve the product testing system,further improve the product quality control level;maintain brand advantage,targeted Improve product quality;reasonable pricing,flexible pricing strategy.In summary,this paper constructs a complete aspect-level sentiment analysis system,which provides reference for other researchers to conduct comment analysis.The deep learning method used in the aspect-level sentiment analysis is excellent,with wide practical significance and application value.