Research On Classification Of Online Product Reviews' Helpfulness Based On Deep Learning

Online shopping is favored by consumers because it can provide variety of products,open and transparent prices,and convenient and fast logistics.And online product reviews,as a key part of the online shopping experience,have an important impact on user purchasing decisions.But with the surge in online transactions,the explosive growth of comment data has greatly hindered users from getting the information they need.Therefore,automatic identification of high-quality,objective,real,and readable helpful reviews is a research issue of common concern to the industry and scholars.At present,there are two major directions in the helpfulness research field of online product reviews: theoretical research on influential factors of helpfulness and application research on helpfulness identification methods.Therefore,on the basis of exploring the mechanism of the impact of review content,reviewers,and review objects on helpfulness,this article turns the helpfulness identification problem into a classification problem.By comparing the application effects of existing classification algorithms,we choose deep learning technology to build a classification model.The main work and innovation are as follows:(1)Identify influential factors that can be incorporated.By researching relevant literature,summarizing the mechanism of the impact of review content,reviewers,and review objects on helpfulness,from the perspective of importance and feasibility,select the relevance of review topics and product information and its consistency with user ratings as key factors to incorporate into subsequent models.(2)Amazon review data preprocessing based on helpfulness influencing factors.According to the helpful votes to measure the helpfulness of comments,the ratio of helpful votes to 80% of the total votes is selected to divide the positive and negative samples.At the same time,comments with a length between 100 and 200 words were filtered to control the impact of comment length on helpfulness.(3)Construct a helpfulness classification model of online product reviews based on BERT.In response to the classification problem,it is necessary to extract the deep semantic features from the representation of the review text.Considering that the BERT pre-training model based on self-attention mechanism has achieved significant results on multiple natural language tasks,we constructed a review helpfulness classification model based on BERT fine-tuning and BERT feature-based.Through experimental analysis,it is found that the classification model based on BERT fine-tuning requires higher computational resources,although the accuracy is higher.(4)Propose a helpfulness classification model of online product reviews incorporating influencing factors.In the application of helpfulness classification,deep learning algorithms focus on optimizing the model structure and seldom incorporate other influencing factors.From the perspective of applicability and feasibility,we choose a helpfulness classification model based on BERT feature-based to incorporate influencing factors.When integrating user ratings,based on the interaction between user ratings and comment content,which affects the helpfulness,the two are selected to be input into the classification model for feature extraction.When incorporating product information,considering that the similar between the reviews and product information,we introduced an attention mechanism for character-level matching to optimize the vector representation of the review text.Experiments have found that incorporating user ratings and product information separately can optimize the BERT-based feature model and achieve a classification effect similar to the BERT-based fine-tuning model.On this basis,the classification model that incorporates both factors can achieve the highest score on each index,which is most in line with actual application requirements.