Design And Implementation Of E-Commerce Review Mining System Based On BERT

With the rise of the e-commerce industry,the circulation of various goods has accelerated.While it is convenient for consumers to buy,it also makes people have a higher pursuit of the brand quality of the goods.Based on the carrier of the e-commerce platform,a comment-oriented consumption method has gradually expanded in the market.When consumers buy goods,they will refer to the comments provided by consumers who bought the product in the past to adjust their consumption wishes.This means that e-commerce review data has a direct impact on product sales and merchants’ interests.At present,a large amount of text review data exists on the e-commerce platform,which contains information such as consumer opinions that businesses are eager to know.This information can effectively help businesses to conduct public opinion analysis,user understanding,product optimization and marketing decisions,etc.,and has great practical value and application prospects.This article conducts aspect-level opinion mining for e-commerce product reviews.First,the Bert pre-training model is used to supervise learning on the Pytorch framework to train a neural network model,and the cross-validation method is used to verify the trained model..The experimental results prove that the accuracy rate and recall rate of the cosmetics field and the notebook computer field can reach more than 70%,and the F1-score has reached about 75%,which is a very ideal result.However,the F1 result in the cosmetics field is always higher than that in the notebook computer field.This is because the labeled data set of the notebook computer is only about 900,which is much less than the labeled data set of 14,000 cosmetics.Aiming at the shortcomings of small number of labeled training data sets and poor training results,an improved method based on deep transfer learning is proposed,that is,based on different pre-trained models,the data is pre-trained again using data from other fields.Then,the model was transferred to the target domain through the fine-tuning process,and the models were merged through integrated learning to further improve the model.The experimental results show that in terms of accuracy,the improved model has a 5%-6% increase,the recall rate has a 4%-7% increase,and the F1-score value has an increase of about 5%.Regardless of the accuracy rate,recall rate or F1-score value,the scores of the pre-trained model are much higher than the unimproved Bert model.In particular,after integrated learning,the model’s F1-score value reached 82%,which is an 8% increase compared to before improvement,which verifies the effectiveness of the improved method used in this article.This paper finally designs and implements a B/S architecture-based Flask e-commerce review opinion mining system,which includes modules such as registration and login,user management,opinion mining,and background management,focusing on deploying the improved deep neural network model to In the system,users can interact with each module of the system in a friendly interface to mine opinions of product reviews,and the final result is presented to users in a graphical manner.