Research And Application Of Aspect Based Sentiment Analysis Method Based On Graph Neural Network

With the mature development of Internet technology and the rapid rise of We-Media,people are not only simply browsing,but are more willing to express their opinions and opinions on the Internet almost every day,and share their shopping experience.There is a huge amount of text review information on the Internet,which contains valuable wealth for all walks of life.Especially in recent years,many platforms and media have won business opportunities by means of content recommendations and group recommendations.However,these recommendations cannot be separated from the study of sentiment tendencies in massive text information.The sentiment analysis,according to the granularity of the research object,is divided into three levels: document level,sentence level,and aspect level.The first two can only judge the sentiment polarity of the text as a whole,and cannot correctly judge the sentiment polarity of a particular aspect.Therefore,aspect-based sentiment analysis has become a mainstream research task.With the rapid development of deep learning,good progress has been made in the field of aspect-based sentiment analysis,and research based on graph neural networks has become one of the hot spots in recent years.However,there are still some defects in the existing research of using graph neural networks:(1)Limited by the use of neural network model,it is unable to accurately use syntax-dependent information;(2)There are complex comment sentences and it is difficult to deal with a large number of syntactic structure errors and colloquial comments.In view of the above problems,this thesis proposes two methods to solve them.According to the actual needs of aspect-level emotion analysis,this thesis designs and implements a network text public opinion analysis system based on Python language and MySQL database.The main work and research of this thesis are as follows:(1)This thesis proposes a deep dependency-aware graph convolutional networks for ABSA(DA-GCN).It mainly solves the problem of not using syntax-dependent information accurately.Specifically,the dependency perception graph convolution network is used to establish an interactive relationship with multiple heads of attention,model the dependency relationship and its dependency type information,distinguish important context information from these relationships to enhance sentiment dependence,and use the dependency perception grammar information to effectively learn relevant information from the generated graph.Secondly,multiple conditional random fields that integrate structured attention are introduced to better capture the opinion words of specific aspects,which have prominent advantages in extracting important contextual information features and significantly improve classification performance.(2)This thesis proposes an interactive double graph convolutional networks for ABSA(InterGCN).It is mainly used to solve the problems of sentences with syntactic structure errors or no significant syntax and colloquial comments.The dependency tree is reconstructed according to the syntactic information,and the convolution module of the syntactic graph is constructed to make full use of the syntactic information;The convolution module of the semantic graph is constructed,and the semantic correlation between words is expressed by multi-head self-attention;In addition,the interactive feature fusion layer is added,and the syntactic and semantic dual graph convolution network interactive fusion learning is an important feature related to aspect words in syntax and semantics.Experiments on public data have proved the effectiveness of the model.(3)Design and implement a network text public opinion analysis system.First,the public opinion system crawls data from microblogs,We Chat official accounts,news websites,and other major websites through crawler technology.Secondly,it performs operations such as word segmentation and de-stop words on the crawled data,enters the processed data into the system,systematically analyzes the sentiment polarity of specific aspects in the text,and finally completes the system test.The test results verify that the realization effect of each function of the system can meet the expectations.Most importantly,the system’s comprehensive and accurate prediction of sentiment polarity can meet users’ more detailed needs,which proves the effectiveness and practicability of the proposed scheme.