Research On Aspect-based Sentiment Analysis Based On Dynamic Fusion Of High-order Features

With the continuous development of social networks,more and more Internet users comment and share their lives on e-commerce,social networking and other platforms,resulting in a large amount of text data.In order to analyze and mine the valuable information in these data,researchers have carried out a large number of aspect-based sentiment analysis research,which has made a breakthrough in recent years,but also faces many challenges.The aspect-based sentiment analysis task aims to correctly judge the polarity of different aspect terms in the text sequence.For the text sequence with multiple aspect terms,different aspect terms need to be correctly matched with the context opinion words.However,in the process of traditional first-order feature extraction and interaction,the one-to-many or many-to-many high-order correlation between data is difficult to be effectively modeled,resulting in the model’s shortcomings in accurately distinguishing the polarity in different aspects.Therefore,how to model the multivariate high-order correlation of data and how to integrate high-order features with general features has become an urgent problem to be solved.To solve the above problems,this paper carries out aspect-based sentiment analysis based on high-order feature dynamic fusion in temporal network model and graph network model respectively.For the temporal network model,a Second-Order Dual-Matching Architecture is proposed in this paper.Firstly,the first-order semantic features of sentences and aspect terms are learned by the traditional feature extractor,and then the second-order features are extracted by covariance operation to realize the interaction between aspect terms and context words at the semantic level.The introduction of second-order features improves the quality of text feature extraction,but also introduces new text noise.In order to reduce the impact of noise,this paper designs the First-order and Second-order Feature Interaction Layer for dynamic fusion between features,so as to obtain higher quality feature representation.In addition,we also propose a Dual-Matching Layer from the perspective of semantic interaction between sentences and aspect terms,so as to realize the correct matching between aspect terms and context opinion words.Experimental results verify the effectiveness of the model,and visual experiments also show that each component of the proposed model contributes to the improvement of model performance.For the graph network model,this paper proposes a Hypergraph Convolution Network with Gate Mechanism model and an Aspect-based Hypergraph Convolution Network with Gate Mechanism model.Both models construct hypergraph structure through the word co-occurrence relationship of syntactic path,capture the multiple correlation and high-order correlation between contexts,and use the joint coding of ordinary graph and hypergraph structure to strengthen the understanding of context information.Specifically,the model first reshapes the syntactic dependency tree,and realizes the extraction and interaction of context features through Dot-Edge hypergraph convolution,and then designs a Feature Fusion Layer Based on Gated Mechanism to dynamically integrate the features of hypergraph structure and ordinary graph structure,so as to solve the confusion of context opinion words corresponding to different aspects in sentences in sentiment expression.The experimental results verify the necessity of introducing high-order statistics,and also show that the dynamic interactive fusion of features is conducive to the improvement of model performance.