Research On Text Sentiment Analysis Based On Graph Neural Network And Attention Mechanism

Since people have entered the information age,the Internet has influenced people’s lives in a subtle way.Electronic products are becoming more and more popular in daily life.Network platforms,such as Microblog,Taobao,provide convenience for people to obtain external information,and also produce a large number of comments on current events and products.Text sentiment analysis is to mine the sentiment information expressed in the comment text,which can provide decision support for consumers,help businesses to obtain timely feedback and improve and upgrade their products,and help the government to correctly guide public opinion.Therefore,text sentiment analysis is of great practical value.Previous research methods are document-based or sentence-based sentiment analysis methods,which take the whole document or the whole sentence as a whole aspect and give a generalized sentiment analysis result.As people’s needs become more and more refined,document-based and sentence-based sentiment analysis can no longer meet people’s practical needs.Aspect-based sentiment analysis can provide more detailed reference.Although traditional methods based on deep learning have achieved certain results,they ignore the role of context information and the key syntactic relations and sentence structure information.Therefore,this thesis combines graph neural networks and attention mechanisms to study the aspect-based sentiment analysis.(1)A graph convolution network with bidirectional attention is proposed for the problem that the model ignores the mismatch between the actual contextual information and aspects.Firstly,the global semantic information of text is learned by a multi-headed attention mechanism,and the location information of context is introduced to enrich the text representation.Secondly,graph convolution network is used to capture the syntactic relations of context,which is helpful for the model to identify the sentiment information related to aspects more accurately.Finally,the interaction of aspects with context through a bidirectional attention mechanism helps to obtain a high-quality aspect representation.Extensive experiments are conducted on three publicly available datasets.The experimental results demonstrate that the proposed model can better capture contextual information related to aspects,which helps the model to correctly determine the sentiment polarity of aspects.As a result,the performance of the model is effectively improved.(2)Aiming at the problem that the model cannot use accurate syntax dependency information and sentence structure information,an enhanced dependency-aware graph convolution network for aspect-based sentiment analysis is proposed.The model integrates syntactic dependencies,dependency types and dependency distances.The comprehensive introduction of syntactic dependency information through graph convolution networks,which helps the model to parse the structural information of sentences and further focus on contextual information related to aspects,thus helping the model to understand contextual information accurately.In addition,an aspectspecific attention layer is designed to obtain the aspect representation in the model.Extensive experiments were conducted on five publicly available datasets and comparative experiments were conducted to verify the reliability of the proposed model,and the experimental results showed that the proposed model can tap deeper sentiment information that is closely related to aspects.