Aspect-level Sentiment Analysis Based On Graph Convolutional Neural Network

With the vigorous development of social media,e-commerce,customer service and online review,the text information created by users on the Internet contains important business value.Sentiment analysis of these text data is an important subject.Traditional document or sentence-level sentiment analysis cannot capture the nuances of opinions expressed about different aspects or features of a product,service or topic.Aspect-based sentiment analysis(ABSA)aims to identify the sentiment of a specific aspect or topic mentioned in a text,and ABSA has attracted increasing attention from academia and industry for its in-depth analysis of viewpoints and sentiment.This paper focuses on the aspect-based sentiment classification task.The main research work in this dissertation is as follows:(1)Most recent studies on aspect-based sentiment classification task are based on syntactic dependency graph combined with graph neural network model,and experiments have confirmed that syntactic dependency tree can help improve the performance of aspect-based sentiment classification task.However,these studies are not comprehensive enough in extracting information from the syntactic dependency tree,only retaining the connectivity of individual nodes and not focusing on the type of syntactic relationship between words in a sentence,which provide rich and useful linguistic knowledge for the ABSA task.In this paper,this dissertation proposes a Syntactic Dependency-aware Graph Convolutional Networks(SDGCN)model that dynamically updates typed dependency labels.It consists of two modules: a node-aware edge update module and an edge-aware node update module.At each GCN layer,the edge-aware node update module aggregates information from each node’s neighbors by specific edges,and later dynamically updates the edge representation using the nodeaware edge update module to make the edge representation more informative.The two modules work in a mutually reinforcing manner by iterative updates.The SDGCN model has been extensively experimented on three publicly available datasets,and the experimental results fully demonstrate the effectiveness of the SDGCN model.(2)The approach of combining syntactic dependency parser can provide more syntactic knowledge for the model,but affected by the performance of syntactic dependency parser,in some complex sentences,opinion words do not have direct syntactic relationship with aspect term,in order to solve the above problems,based on SDGCN,this dissertation proposes a novel Syntactic and Semantic Enhanced Graph Convolutional Networks(SSEGCN)model,which can combine semantic information with syntactic feature.Specifically,this dissertation proposes an aspect-aware attention mechanism to capture the semantic features associated with aspect term,and combines it with a global self-attention mechanism to jointly construct an attention layer.By calculating different distances between words in the syntactic dependency tree,syntactic mask matrices are constructed to learn syntactic structure information from local to global,and finally syntactic mask matrices are combined with the output of the attention layer to further enhance the features of nodes through graph convolutional neural networks.Experimental results show that the SSEGCN model outperforms all baseline methods on three publicly available datasets and achieves the best performance,demonstrating the superiority of the SSEGCN model.