| In today’s era,Internet technology has been greatly developed and popularized.Users have increasingly become the providers of Internet resources and information,which produces a large number of text data containing the emotions of publishers.Sentiment classification is the main means to analyze and use this kind of data,and aspect-level sentiment classification is a fine-grained classification task in sentiment classification,which aims to accurately identify the sentiment category of a specific aspect in the text.In the face of massive text data,relying solely on human resources is not competent for its analysis and utilization.It has become an inevitable demand to construct an automatic text aspect-level sentiment classification model with high accuracy and robustness.Recently,various methods based on graph attention neural networks and attention mechanisms have brought considerable performance improvement for aspect-level sentiment classification.However,there are still the following problems in current research:(1)The weight of each node in the traditional graph attention network is unreasonable when calculating the attention between the nodes that transmit emotional cues and the nodes that exist emotional cues,which introduces too much noise for the final classification basis.(2)The graph attention network transmits information only between syntactic direct or indirect adjacent nodes.When incorrect syntactic information is input,it may lead to the lack of emotional cues in the final classification basis.(3)The traditional attention calculation is carried out at the word level,ignoring the problem that the importance of different dimensions of the same word is different,and the coarse-grained attention weight will increase the complexity of the model to determine the emotion category.In view of the above problems,the research in this thesis is summarized as follows:(1)This thesis proposes an aspect-guided graph attention neural network.Compared with the traditional graph attention network,the method of aspect information guiding attention calculation can allocate attention to indirect syntactic adjacent nodes more reasonably,so as to achieve the purpose of reducing the noise in the classification basis.(2)This thesis proposes a semantic attention,which fully exploits the emotional cues that may be missed due to erroneous syntactic information in the shallow encoding information,and enhances the robustness of the graph attention network through the interaction between the shallow encoding information and the deep semantic information and aspect information encoded by the graph attention network.(3)This thesis proposes a multi-granularity attention,which focuses on the importance of different dimensions of words through dimension-level attention,and focuses on the importance of words through word-level attention.Through the weighting of two kinds of granularity attention,the model can more accurately model the semantic information of the text.Finally,through the widely used Sem Eval2014 and Twitter public datasets,this thesis compares the performance of the model with the important influential models.The results show that the model proposed in this thesis has certain performance advantages.The effectiveness of the proposed method is proved by ablation experiments,case analysis and visualization analysis. |
