Research On Aspect-based Sentiment Analysis Based On Deep Learning Method

Natural language processing and deep learning algorithms are continually evolving,more fine-grained sentiment analysis tasks i.e.aspect-based sentiment analysis(ABSA)has become a research area on which researchers concentrate.ABSA refers to the sentiment prediction of specific entity aspects in a text.Compared with traditional sentiment analysis methods,the ABSA task can analyze the sentiment information of different aspects in a text more accurately and better meet the needs of practical applications.Aspect-based sentiment analysis using neural network models is a popular study method right now.Among them,the method of using graph convolutional neural network to learn features of syntactic structures and complex semantic relationships between words in this text’s sentences has received a lot of attention from research scholars due to its excellent performance.After summarizing and analyzing the existing aspect sentiment analysis methods,It is discovered that there are some issues left.First,Insufficient use of syntactic information and word lexical features in the syntactic dependency parse tree,lack of an constructive mechanism to utilize dependency types and word lexical features.Second,the existing tools do not parse the dependency parse tree accurately enough,and there are cases of incorrectly parsed dependencies,which undoubtedly bring redundant noise information to the syntactic information and thus affect the classification performance of the model.To solve the above two problems,this paper proposes a Graph Convolutional Neural Network Model Based on Fusion Syntactic and Semantic Information Enhancement(FSSIE-GCNs),based on integrated syntactic and semantic information improvement,this study investigates how well ABSA performs utilizing this data.The primary work is listed below.(1)In this research,a graph convolutional network based on syntactic information enhancement is designed to make full use of syntactic information such as dependency types.Firstly,the dependency types in the dependency parse tree are distinguished according to the custom rules,and the important dependencies on aspect words and opinion words are retained.Secondly,this paper constructs a dependency type feature matrix and a sentence lexical feature matrix,and by combining the two feature matrices,the final reconstructed syntactic dependency graph is obtained.Finally,a multilayer graph convolutional neural network is applied on the syntactic feature map to learn the syntactic features of sentences.(2)This work introduces a way of adding semantic information to syntactic information intending to lessen the effect of inaccurate dependency parsing on the prediction of sentiment polarity of aspectual words.Specifically,this paper proposes an Aspect-Oriented SelfAttention Mechanism(AOSAM),which consists of a Self-Attention network that learns the semantic relations between aspect words and their context words,and an Aspect Attention network that learns the semantic information associated with a specific aspect.The attention score matrices obtained from the two attention layers are fused to obtain the semantic feature maps,and by using this strategy,the resulting semantic feature maps learn deeper semantic properties of the phrases.After that,a multilayer graph convolutional neural network is used to learn the semantic features on the semantic feature graph.Finally,this paper designs a CAMM mechanism to fuse the extracted syntactic features and semantic features,which effectively mitigates the loss of model performance due to the noisy information in the syntactic features.(3)This paper compares the FSSIE-GCNs model with the benchmark model on four public datasets of Laptop,Restaurant,Twitter and MAMS.The experimental results show that the performance of FSSIE-GCNs on two evaluation indicators is better than the benchmark model.This paper also conducts a sensitivity analysis of the model,analyzes and discusses the impact of important hyperparameters in AOSAM,the number of layers of GCN,and the strategy of feature fusion on the prediction accuracy of the model.At the same time,this study conducts ablation experiments and visualization simultaneously to assess and confirm the contributions of various model components.