Research On E-commerce Platform Comment Sentiment Analysis System Based On Deep Learning

With the rapid development of information technology driving the rapid popularization of online shopping,major e-commerce companies have collected a large amount of review information containing the subjective opinions of consumers.This information contains huge commercial and social value.To obtain effective information to assist decision-making from massive review data,the scheme of using artificial intelligence technology to mine the emotional information of reviews has significant advantages compared to human browsing and summarizing.Looking at the product review data obtained by several major e-commerce platforms,it is found that there are typical problems as follows: due to the relatively scarce number of negative reviews in the review data,the serious data type imbalance and the large amount of review data lead to the high difficulty of data labeling.Today,when the industry has reached a consensus on the importance of comment data,in response to the above problems,ecommerce platform comments are used as an auxiliary basis for decision-making.How to use deep learning technology to conduct more effective emotional information mining is of great research significance.This paper proposes an sentiment analysis method for e-commerce platform user review texts from the perspective of semantic understanding,and converts sentiment analysis problems into semantic topic classification problems.Since the BERT pretraining language model has shown good results in various sentiment analysis tasks,this paper uses the BERT language model as the most basic network framework,and a series of improvement studies are carried out based on the characteristics of the comment data on the basis of the model.The main work of the paper as follows:(1)Aiming at the problem of the high difficulty of data labeling caused by the massive review data: this article uses the user's star rating for this shopping on the e-commerce platform as the initial label.Taking into account the insufficient coverage of such annotations and random user reviews,this article uses a clustering method,combining open source sentiment dictionary and e-commerce platform comment data to generate an sentiment dictionary suitable for the field of e-commerce reviews,and then use the dictionary to comment The data is subjected to preliminary sentiment analysis,and then the sentiment analysis results of the comments are compared and matched with the comment tags,so as to filter out high-quality tag data.(2)Aiming at the problem of imbalanced sentiment analysis training data categories caused by the sparse number of negative reviews: This paper introduces the image field data enhancement method Mix Match and the traditional NLP data enhancement back translation method to enhance the negative samples and increase the number of samples.Since the Mix Match method is only suitable for continuous variables,this article uses the first 10 layers of BERT to transform the discrete text vector into a high-dimensional dense vector,and combines the filtered low-quality tags to perform the Mixup operation to obtain enhanced data.At the same time,this article further introduces the Focal?Loss algorithm in the field of target recognition,combines the loss function of Mix Match and BERT,introduces weights and conditioning factors,and punishes the contribution of large-scale data and easily identifiable data to the total Loss to achieve the purpose of resisting data imbalance.This article has tested the effectiveness of the sentiment analysis method proposed in this article through the sentiment analysis experiment on e-commerce platform reviews.The experimental results show that the model proposed in this article has a good effect,which is significantly improved compared to the ordinary BERT model and traditional data enhancement methods.