Research On Sentiment Analysis Of Product Reviews Based On Dual-channel Mixed Neural Network

In the context of "Internet +",online e-commerce platforms have developed rapidly,and people's living habits have also changed accordingly.More and more people like to shop online and express opinions to express their emotional tendencies towards products,so a lot of product review information is generated,and these data often contain a large number of user sentiment factors,if emotional information can be effectively mined,it will not only provide information feedback to merchants to facilitate timely adjustment of sales decisions,but also help industry surveys so that governments can take effective measures.Therefore,research on the text sentiment analysis method for product reviews has important commercial value and social significance.In the Internet era,hundreds of millions of messages are generated every day,in the face of such a large amount of data,traditional sentiment analysis methods face a series of challenges,such as the inability to fully extract the sentiment information in the text and the long training time,and the inability to perform deep emotion mining on target features.In order to solve the above-mentioned shortcomings,this paper improves and perfects from two aspects of model design and parameter training.Therefore,a dual-channel hybrid neural network product review sentiment analysis model is proposed.The main innovations are as follows:(1)In previous research methods,only a single word or word embedding information was generally used as input information,however,this situation often leads to incomplete learning of text semantic information,causing the risk of missing information.In view of this drawback,it is proposed to introduce Char-Embedding and Word-Embedding into the model,so that the model has the ability to learn emotional features from both word-level granularity and word-level granularity embedded vectors.This method can solve the shortcomings of incomplete learning features of existing models,and can improve the feature learning ability of models to a certain extent.(2)It is found through research that the traditional sentiment analysis method does not further study the fusion method of feature information,but just simply combines the extracted features.This method is prone to ignore the weight distribution problem of the target feature,resulting in a reduction ability to distinguish text and a longer training time.Aiming at this problem,it is proposed to introduce attention technology in the feature fusion layer of the model to improve the model's ability to distinguish key information.This technology uses the calculated weight values to multiply different feature information,effectively amplifying the emotional information of the target feature vector,and weakening the emotional information of the non-target feature vector.(3)Defects in the Max-Pooling layer of the conventional TextCNN network: After the convolution operation,the position information of text features is usually retained,but after the Max-Pooling layer,only part the features with the highest score are retained,while other the strong feature information with a low score but high frequency of occurrence is discarded,causing a large loss of feature information.In view of above defect,it is proposed that take Max-Pooling replace with K-Max-Pooling,so that not only the Top-K features with the highest score among all feature values can be obtained,but also the original sequence of extracted features can also be retained.This improvement can alleviate the shortcomings of the TextCNN network to a certain extent and help improve the classification performance of the model.