Research On Aspect-Level Text Sentiment Analysis For Graph Neural Network

With the rapid development of social networks and Internet applications,a large amount of user comment data is scattered in every corner of the Internet.How to extract valuable information from massive unstructured text data has become one of the hot topics in the field of natural language processing.Sentiment analysis,also known as opinion mining,is an important subtask in natural language processing.The traditional coarse-grained emotion analysis can only judge the sentiment tendency of the whole text or the whole sentence,but cannot judge the sentiment polarity of specific aspects in the sentence.At the same time,due to the continuous development of the Internet,the structure of online user comment data is also complicated,and the traditional expression effect of sentiment analysis is difficult to meet the current needs of people.Therefore,researchers propose a new emotion analysis task,aspect-based sentiment analysis,which aims to identify the sentiment polarity of a specific aspect of a sentence.This paper studies aspect-based text sentiment analysis research based on the graph neural network and proposes two graph neural network models.The main work includes the following two points:(1)Aiming at the problem that the existing models ignore the emotional connection between aspects when facing multiple aspects,this paper proposes a new aspect-based sentiment analysis model based on dual-channel interactive graph convolutional network(DC-GCN).Firstly,to better learn the contextual semantic information of the sentence,the word embedding layer is constructed by fine-tuning the BERT pretraining model.Then,two GCN models are designed to capture the syntactic dependency tree and the node dependency information in the multi-aspect sentiment graph.In addition,an information interaction layer was designed using the attention mechanism to obtain a more informative feature representation.Finally,the experimental verification was carried out on the Restaurant,Laptop and Twitter data sets,and the results showed that the accuracy of the DC-GCN model on the three data sets reached 88.11%,81.77%,77.84%,respectively,and the MF1 value reached83.28%,77.09%,76.77% respectively.(2)The solution to the DC-GCN model depends on the quality of the parsing of the syntactic dependency tree.When faced with complex sentence parsing quality is not good,it is easy to cause noise and other problems in the GCN convolution.This paper proposes an aspect-based sentiment analysis model based on the syntactic enhanced graph convolutional network(SEGCN).Firstly,the BERT pre-training model is also used as the word embedding layer of the model,and the hidden vector representation of the aspect and context is obtained through the bidirectional gated recurrent unit(BiGRU),so as to better extract its semantic information.Then,the conditional random field(CRF)extracts the opinion span features of sentences,and obtains potential opinion labels as the basis for generating syntactic dependency graph connectivity to reduce noise.In addition,considering that the existing aspect-based sentiment analysis model ignores external sentiment knowledge in modeling aspect and context relationships,this paper enhances the syntactic dependency graph by introducing sentiment knowledge from Sentic Net to help the model better capture the sentiment dependence of aspect and context.Finally,the experimental verification was carried out on the Restaurant,Laptop,and Twitter datasets.The results showed that the accuracy of the SEGCN model on the three datasets reached 88.26%,82.41%,and 78.02%,respectively,and the MF1 values reached 83.30%,78.28%,and 76.94%,respectively.