Research On Product Evaluation System Based On User Reviews

Online shopping is a popular shopping mode which used by people today.However,because there are plenty of commodities,which have different qualities,sold on the website,it is often difficult for users to make a choice when shopping online.The user reviews of products on e-commerce platform can help users to make product purchase decisions.However,with the increasing of the number of user reviews,it is much harder for users to obtain useful information from these reviews.For this reason,we try to design and implement an analysis software,with the collected user reviews,to extract sentiment words from the reviews and quantify the subjective feelings of users about the sold commodities.Existing sentiment analysis technics can be divided into two categories:machine learning method and emotion lexicon method.The former mainly focuses on emotion binary-classification or multi-classification.However,the binary-classification's granularity is too coarse to meet the needs of our quantitative assessment;Multi-classification requires much training time and high quality of label training set which involves much human's hard work.The latter can achieve fine-grained quantization by calculating the emotion intensity value of the reviews,but due to the incompleteness of Chinese emotion lexicons,the information contained in Chinese reviews cannot be fully excavated and utilized.Based on the above mentioned technics,the existing works further extend the evaluation perspective of products from the feature dimensions,however,the dimensions are still not enough for constructing a complete space of commodity features.Based on the existing research results,we design and implement a rating system,to quantitatively evaluate a product and/or an online vendor by giving a emotion intensity value from each feature dimension.In this rating system,we create a specific feature space for a category of product to evaluate a product in multiple dimensions,and construct a dedicated Chinese domain emotion lexicon to make full use of the information contained in user reviews.Specifically,we improve the performance of our rating system in the following three aspects:(1)We customize a segmentation dictionary by adding the words specifically contained in review data into the dictionary,such that we can improve the accuracy of word segmentation of the reviews by using Jieba.(2)We extend the feature space of a specific category of product,by using the methods of word frequency statistics and similar feature clustering to expand the feature dimensions and the words associated to each dimensions,such that we can rate the products and vendors in more detailed feature dimensions.(3)We extend the domain sentiment lexicon by including some more specific words and their emotion intensity values in it,such that we can deeply utilize the extended domain sentiment lexicon to compute the value of each review in each feature dimension.The lexicon extension is based on word similarity computed by using Word2vec.Finally,based on the above constructed emotion lexicon and feature space,we calculate the emotion intensity value of a product in each feature dimension,and further obtain the overall evaluation score,i.e.,the weighted average value of the above values.In our experiment,we compute the rating scores of 100 combinations of cellphones and vendors on Tmall platform and compare our rating scores with the rating scores given by Tmall platform,the result demonstrates that the overall rating scores generated by our rating system are consistent with the scores given by Tmall.Moreover,our rating system can differentiate these products and vendors in more detailed feature dimensions.