Opinion Mining For Online Products Based On Deeping Learning Approaches

With the development of “Internet +” and mobile Internet,more and more users participate in the online discussion of products,and express their opinions.A large amount of data shows that due to the easy accessibility and high credibility of online reviews,online consumers will search for product information including product quality,performance,and cost performance on various relevant website platforms,before making purchasing decisions.They will compare similar products to make sure themselves buy the right products.Likewise,offline manufacturers and merchants can upgrade and renovate products based on feedback comments,making products more fit consumer needs.However,the rapid expansion of the network comments make us face the dilemma of information explosion but lack of knowledge.People often need to spend a lot of time to browse product evaluation information,filter out a large amount of irrelevant information,and they can get more comprehensive and valuable evaluation information.In fact,the aspect-level of comments is the focus of users' reading and browsing,and it is worth further exploration,analysis and summary.Therefore,this paper uses the deep learning approaches to build an Opinion Mining model to deal with the vast amount of product comments on the Internet,hoping to achieve the goal of automatically identifying the targets and polarity in the review texts,that is,to achieve an aspect-level Opinion Mining of product reviews.In this way,users can find value information from the comments more quickly and accurately.Merchants in the offline industry chain can perfect their products,optimize service items and service methods according to feedback from online comments.Finally,Internet services develop in a refined direction in the future through gradual integration of online and offline methods.The main subject of this paper is to distinguish the aspect of products and polarity in the field of Opinion Mining.The current research methods for both are focused on unsupervised methods(including semi-supervised methods)and traditional supervised machine learning models.There are relatively few research methods based on deep learning approaches.Unsupervised methods and traditional machine learning models can show good performance,but the two methods rely heavily on prior knowledge of linguistic rules and domain dictionaries designed by experts,which are not only costly but also low generalization in other different comments.The deep learning algorithm can automatically explore the abstract concepts from the bottom-layer features to the high-level features,using unsupervised or semi-supervised features learning and hierarchical feature extraction algorithms instead of manually acquiring features.Based on the above reasons,this paper uses the long-short-time memory network(LSTM)model to input the word representation as the input,and separately distinguishes the opinion targets and polarity,and achieves opinion mining at an aspect-level.Among them,since a single comment text may involve multiple different opinion targets,the original corpus of the comment needs to be processed before training.This paper proposes three methods for multi-label corpus processing.The first method is that we randomly select one of the labels,discard the needless labels,and analyze them as single-label text.The second method is that we determine the number of corpus copying according to the number of labels and assign different labels to the copied corpus.The third method is that we split the corpus according to the label seed set,and then give processed corpus different labels.Finally,using the tagged reviews of the automobile industry as a sample data set on the CCF Big Data and Computational Intelligence Contest website,this paper verifies the accuracy of the model under different methods.