Research On Reviews Helpfulness Recognition Based On Improved Graph Convolutional Neural Network

The rapid development of the Internet has led to the rise of e-commerce,and people have gradually accepted and become accustomed to online shopping.According to the 51st Statistical Report on the Development of Internet in China released by CNNIC,the number of online shopping users in China has reached 845 million,accounting for 79.2% of the total number of Internet users.As the main carrier of user generated content on e-commerce platforms,online review information plays a crucial role in consumer online shopping decisions,merchants’ improvement of product services,and platform system construction.However,the low threshold for the release of comment information has led to an exponential growth in its number,and a large amount of comment information is mixed with a large number of useless comments that have no reference significance for platform users’ decision-making.How to efficiently help users filter out useless comments that have no reference value from the mass of comments with uneven quality has become a research hotspot in recent years.As a key influencing factor of the usefulness of reviews,the mining of review text information is increasingly receiving attention from scholars,and the information about product advantages and disadvantages that consumers and businesses need to refer to and user feedback are mainly expressed by the "feature viewpoint semantics" information in the review text.However,the existing research based on text is mostly limited to incorporating statistical features of feature views for comment usefulness identification.Therefore,this study relies on text graph convolution technology and targeted improvements to propose an FFGCN model to fully model semantic biases of feature views in comment text for comment usefulness identification,filling this research gap.The main innovations of this study are as follows:(1)Effectively solving the problems of semantic ambiguity and sparse data in commentaries.First,considering the ambiguity of feature and viewpoint words after direct word segmentation,we design feature and viewpoint block extraction rules based on word segmentation results and dependency analysis,using word blocks as vocabulary nodes on the graph to make the semantic expression of product features and viewpoint nodes clearer and more specific.At the same time,we improve the information transfer and edge weight quantization method to make the representation of comment document nodes obtained through graph convolution more accurate;In addition,multi-level membership relationships between feature words are incorporated into the text graph to construct border rules to alleviate the problem of sparse data in the review text.(2)The SITextRank algorithm is proposed.In terms of quantifying the degree of tightness between words and document nodes,considering the different contributions of different words to the usefulness of comments,based on the information gain of STextRank integrating words into categories,the weight of words relative to the dataset is obtained;In addition,the comment time weight is introduced to obtain the link weight of words relative to different comments,which is used as the link weight of the comment document vocabulary node.(3)Improved the quantification of the usefulness of reviews.Considering the impact of user behavior motivation,a quantitative indicator of comment usefulness is constructed by modifying the number of comments and the emotional polarity of the text.This study captured mobile phone comment data from two models,Redmi Note 9 and iPhone 13,respectively,from JD.com to experimentally verify the effectiveness of the FFGCN model in identifying comment usefulness.The experimental results show that the recognition accuracy of the FFGCN model proposed in this study on two datasets is 93.4%and 93.9%,respectively,which is 0.9 and 1.0 percentage points higher than the optimal results of the baseline model.In addition,the ablation experiments for the chunk analysis and SITextRank module also verified the rationality of the two parts,where the "chunk+TF IDF" results are superior to the "non chunk+SITextRank" results,which further verified the importance of identifying the usefulness of lexical nodes for comments from the perspective of chunk refinement features.