| In recent years,with the development of the Internet and the rapid rise of e-commerce platforms,online shopping has become the mainstream.While consumers are deeply involved,a large number of user comments appear on the Internet,and these online shopping comments often contain a lot of valuable information.Through the sentiment analysis of these texts,useful suggestions and feedbacks can be obtained on the products,so as to improve the purchasing efficiency.Besides,product problems can also be found,defects can be improved,and operational efficiency can be improved.Sentiment analysis technology has a long history,from the initial rule matching based on dictionary to classification based on machine learning algorithm after extracting text features,to the current deep learning technology.Taking electronic product reviews on e-commerce websites as the research object,this paper conducts sentiment analysis on corpus data from two aspects of deep learning algorithm and LDA topic mining.The main contents are as follows:1)Based on the crawler technology,I crawled a large amount of electronic product review information from e-commerce website Jingdong and did text preprocessing on it.The training corpus of this paper was obtained by sorting out the scoring data and artificial emotion labeling.Then the Word2 Vec word vector model is pre-trained on the Chinese Wikipedia corpus and the extracted corpus data is vectorized.In view of the characteristics and defects of traditional deep learning models CNN and RNN,this paper proposes a composite CNN-Bigru-Attention model to conduct emotional polarity analysis on corpus data.Firstly,multi-channel convolutional neural network is used to extract local feature information of different granularity of input text,and then bidirectional gated cyclic neural network(BIGRU)is connected for context sequence learning to remember long-distance dependent information.In this way,the problem that CNN cannot learn context sequence is solved.In addition,it can solve the problem of gradient disappearance or explosion existing in the traditional RNN model.Finally,Attentional Mechanism is introduced after the Bigru hidden layer to screen the important features of text,so that the model pays more attention to the words that are more important for text classification and improves the classification efficiency.Through setting several groups of comparative experiments,the results show that the effect of the CNN-Bigru-Attention model proposed in this paper on the emotional polarity analysis on the data set of electronic product reviews is better than that of other composite models,with an accuracy rate of 93.09% and F1 value of 92.96%.2)After adjusting the parameters of the above CNN-Bigru-Attention model,it is found that the performance of the model is not significantly improved,there is little difference between accuracy rates,and the generalization ability needs to be improved.In addition,word2 vec is a static word vector model,although it has strong generality,However,they cannot understand complex contexts such as polysemy of the word context and cannot make dynamic optimization for specific tasks.Therefore,a Bert model is also pre-trained based on the Chinese Wikipedia corpus,and a Bert-Bigru network is proposed for analysis.A multi-layer Transformer Block structure is used inside the BERT model,which enables the model to learn deep semantic information such as context-dependent bi-directional feature representation.Experimental results on data sets show that the pre-trained language model Bert can obtain better semantic feature representation.Compared with the problem of word2 vec static word vector unable to understand the polysemony of a word and other complex context,the Bert Bigru model has higher classification accuracy and generalization ability,with the F1 value reaching 95.78%.In order to obtain more fine-grained information representation on the basis of polarity analysis,this paper finally conducts LDA topic mining on the classified comment text,carries out text topic extraction and topic sentiment analysis,and puts forward corresponding suggestions to consumers and merchants according to the results. |
