| With the rapid development of e-commerce,countless products have quickly entered consumers’ field of view.At the same time,with the popularity of the Internet,people have diversified channels to obtain product information,making consumers’ demand for product information more precise and personalized.Traditional product search methods mainly rely on keyword matching,which cannot fully extract the keyword information in the product title and cannot analyze it semantically.Therefore,the matched products often differ greatly from what users actually need.In order to improve the accuracy of user consumption demand,this paper constructs two models for calculating product title similarity from the perspective of product title naming conventions and similarity relationships between words.The specific research results are as follows:1.In view of the problems that the Continuous Bag-of-Words model cannot effectively combine the characteristics of the product title itself and cannot fully extract keyword information in the title to calculate product title similarity,this paper proposes a WM-CBOWbased product similarity calculation model.The model uses the position and distance information of words to characterize the weight of words in the title and combines them into a weight matrix(WM)to represent the similarity relationship between words,which can not only reduce the complexity of the model but also maximize the use of keyword information in the title.In order to further improve the effect of the WM-CBOW model,this paper proposes four improvement and optimization schemes based on the characteristics of the product and actual training results,and carries out detailed design and implementation.The experimental results show that the improved WM-CBOW model has an accuracy rate and F1 value that are 6.26% and 5.53% higher than those of the CBOW model,respectively,proving the feasibility and superiority of the WM-CBOW model in the task of calculating product title similarity.2.In view of the problem that the accuracy of the WM-CBOW model decreases when there is a large amount of useless information in the product title,this paper constructs a product similarity calculation model based on WM-CBOW and Bidirectional Encoder Representations for Transformers.The model fully utilizes the advantages of Bert’s pre-training model and the WM-CBOW model to extract feature information of product titles from different dimensions,and combines the feature vectors output by the two models through the WM-Bert layer to obtain more comprehensive product title information.Based on the characteristics of the task of calculating product title similarity,this paper introduces self-attention and bidirectional attention mechanisms to capture the key features of the title itself and the interaction feature information between titles to the maximum extent.The output of the WM-Bert layer is further used to extract text features through a convolutional neural network and output similarity results.The experiments show that the model with multiple attention mechanisms has an accuracy rate that is 3.61% higher,and the accuracy rate and F1 value of the WM-CBOW and Bert combined model reach 88.69% and 89.03%,respectively,fully demonstrating the accuracy of the proposed model. |
