User-generated Multi-source Text Fusion Representation Based On Similarity And Attention Mechanism

On the e-commerce platform,the comments posted by users usually reflect the users' preference and experience with the corresponding product.Considering the preferences and demand information implied by different reviews can obtain the attributes of the item more comprehensively,therefore,it is easier to obtain the products that match the texts queried by the user,effectively improve the platform service and promote the user's search experience.Accurate vectorized representation of product information is the basis of personalized service.However,comment texts,a typical kind of user-generated multi-source data,inevitably contain some redundant information,which lead to deviations in the vectorized representations.To this end,this paper studies the similarity and attention mechanism based user-generated multi-source text fusion representation algorithm.The main contents are as follows:1)User-generated Multi-source Text Fusion Representation Based on Personalized Query and Dual Similarity: Aiming at the redundant information in the user-generated multi-source comment data published on the e-commerce platform,this section concerns on the fusion and vectorized representation of historical evaluation texts generated by different users for the personalized query objects.First,the Doc2 vec is adopted to vectorize the multi-source text comments.Then,the current user's input query is taken as the benchmark reference object,and the precise fusion of the multisource evaluation text vector of the search object is realized based on the proposed dual similarity criterion fusion strategy.The similarity of the fused vector and the queried one,along with the classification accuracy are taken as the indicators to evaluate the algorithm.Finally,the proposed algorithms are applied to the Amazon products with text comments to illustrate the effectiveness on the vectorized representations.2)User-generated Multi-source Text Feature Extraction and Fusion Using Attention Mechanism: In content(1),directly utilizing the Doc2 vec cannot fully extract the semantic information in the text,which makes the obtained text vector have weak semantic representation ability.Motivated by this,we further propose an attentionbased user-generated multi-source text fusion algorithm.First,the Bi-directional Long Short-Term Memory(Bi LSTM)network is adopted to obtain the text encoding to sufficiently reflect the semantic information.Then,the self-attention mechanism is utilized to get multiple semantic representation from different perspectives.Subsequently,we apply the convolutional neural network to perform convolutional fusion operations on the obtained multi-angle semantic coding to get the fused vector to represent the corresponding product.Finally,the proposed algorithms are applied to three typical user-generated comment text datasets and the multi-source text fusion framework proposed in content(1).The experimental results illustrate that the proposed algorithm can effectively extract the hidden features,discover the relationships between semantics and text structure,as well as obtaining more precise text semantic vectors.This research focuses on the fusions and representations of user-generated multisource text.First,a personalized query-based text fusion framework is proposed;then,for the defect that the Doc2 Vec cannot sufficiently represent the semantics of text,the Bi LSTM network and attention mechanism are adopted to improve it.The experimental results show the effectiveness of the proposed algorithms.This thesis contains 32 figures,11 tables and 114 references.