A Study Of Fine-grained Sentiment Analysis Based On Syntax And Deep Learning

Analyzing people's sentiment polarity toward an aspect from text comments has become a current research hotspot.And according to whether the text contains aspect term,aspect-based sentiment analysis can be divided into aspect-term sentiment analysis(ATSA)and aspect-category sentiment analysis(ACSA).In this thesis,we hope to quickly perform sentiment analysis with the help of techniques such as syntax and deep learning to realize the intelligence of life scenes.In this thesis,the work is as follow:The research of aspect-term sentiment analysis.The current mainstream methods usually consider only the sequential information of sentences,while ignoring the structural information of sentences.For this reason,this thesis proposes a multi-hop syntactic graph convolutional network(MHSGCN)based on syntactic and graph neural networks.The model introduces syntactic information through syntactic dependency trees,and treats syntactic dependency trees as a special kind of graph,and subsequently improves the graph convolutional network according to the way of information transfer on syntactic dependency trees to form a syntactic representation of aspect term.Experiments are conducted on the current five publicly datasets,and it is found that the MHSGCN model performs better overall compared with the benchmark method,with Accuracy and Macro-F1 values of 81.25% and 72.95% respectively,on the Rest14 dataset.Solving the problem of long-distance dependency between sentiment words and aspect term.Syntactic dependency trees are introduced and the syntactic dependencies between words are used to reduce the distance between aspect term and sentiment words.In this thesis,syntactic dependency trees are used as a special graph structure so that syntactic information can be pooled using graph convolutional networks.Due to the existence of over-smoothing in graph convolutional networks,more graph convolutional layers will lead to convergence of features of all nodes,but fewer graph convolutional layers will lead to key sentiment information not being delivered to aspect term.For this reason,this thesis adopts a multi-range attention mechanism,which takes all the graph convolutional layers as contexts,and subsequently calculates the attention coefficients for the graph convolutional layers separately,and finally performs a weighted summation of all the graph convolutional layers to achieve longrange syntactic information capture.The research of aspect-category sentiment analysis.In the aspect-category sentiment analysis task,the text may lack the corresponding aspect term and thus cannot form the syntactic representation of the aspect term.As aspect category information cannot be effectively used in previous classification methods,this thesis proposes a target class-based gating network(TCGN)for aspect-category sentiment analysis.The model recodes aspect category through the connection between context and aspect category,and uses the gating mechanism to realize the selective output of aspect category for sentiment features.Experiments are conducted on Meituan user dining evaluation dataset,and it is found that the TCGN model performs better overall compared with the benchmark method,with Accuracy and Macro-F1 values of 87.98%and 70.78%,respectively.