Research On Automatic Generation Of Personalized Reviews

With the popularity of the Internet,the number of Internet users is increasing,online shopping has become an important way for Internet users to buy daily life items.At the same time,more and more online shopping sites allow users to express their views on products by writing reviews,such as Amazon,Taobao,Yelp,etc.Moreover,many merchants will put forward the positive feedback activities after the user shopping is finished,so as to promote their products and attract more customers.However,manually writing these reviews has become an extremely difficult task.Therefore,we propose to set up a review writing assistant for e-commerce websites to help users automatically generate reviews and avoid the tediousness of handwritten reviews.Automatic generation of product reviews is a meaningful but not fully studied task.Therefore,realizing the automatic generation of reviews becomes our focus topic.Review generation is a text generation that aims to generate realistic reviews in a specific context.Our goal is to generate reviews consistent with the input ratings.Through the analysis and summary of the existing methods,we use the encoder-decoder and the attention mechanism combined method.First,the historical review encoder encodes the user's historical reviews and extractes the information,and the generated reviews are biased toward the user's personalized writing style;then extracts the product title through the product title encoder to enrich the review generation information;then uses the rating information to generate different reviews;meanwhile,we combine the product attribute information through a product attribute encoder,find the product attributes that the user is more inclined to discuss through the product attribute representation learned,and the generated text is biased to their personalized preference;Finally,the output of the first two encoders is transmitted to the decoder through the attention fusion layer plus the coding of rating information,and the decoder is projected to obtain the output word distribution,and then the attention score from the product attribute encoder is added directly to the final word distribution,make the model produce words consistent with the input rating and the words that belong to the most relevant product attributes,so as to automatically generate accurate,diverse and personalized reviews.Users can understand the advantages and disadvantages of different aspects of the product according to the generated reviews,consider whether to buy;merchants can improve the deficiencies of the product according to the generated reviews,so as to produce more products that meet the needs of users and promote consumption;online shopping platforms can attract more sticky users and improve their influence.The experiment of Amazon's real electronic review data set shows that the user's writing style can be accurately reflected by the processing of the user's historical review data,which makes it tend to use the user's idiomatic words when generating reviews.Through the statistics and processing of product attributes,it can make differential evaluation of different product attributes when generating reviews.Compared with other models,our model can obtain the best performance on each index,and can dig out the personalized expression of users more accurately and effectively,and finally generate reviews that accord with the actual evaluation behavior of users.