Research On Aspect-based Sentiment Analysis Based On Graph Neural Network

Aspect-based sentiment analysis is a fine-grained sentiment analysis task that identifies the sentiment polarity of different aspects of a sentence.It can automatically analyze massive amounts of text data,and dig out the sentiment tendency expressed in the text for all aspects of the text,to provide reasonable decision support for consumers,help businesses to get customer feedback and demand in time,and help governments to grasp the trend of public opinion,so it is of great value to both society and business and has thus received a great deal of attention.In recent years,many studies have used graph neural networks to parse the dependency trees of sentences to obtain relationships between aspect words and context.Although the method has achieved some success,several challenges and problems are still faced:(1)How to avoid the problem that parsers fail to avoid parsing errors during syntactic analysis.(2)How to overcome the problem of complex sentences with different degrees of dependence on structural and content information leading to incorrect prediction of sentiment polarity.(3)How to effectively capture both local and global information of a sentence to obtain a more accurate feature representation.(4)How to add external sentiment knowledge to the model to enhance the influence of opinion words on aspect words,thus helping the model to make inferences.To address the above issues,this paper investigates aspectbased sentiment analysis based on graph neural networks,which aims to explore aspectbased sentiment analysis models that can accurately predict sentiment.The main research work and contributions of this paper are as follows:1.This paper proposes a multi-category graph convolutional neural network model(MultiGCN)to solve the problems of parsing errors and different information of sentence dependencies.The model consists of a relational graph convolutional network to extract syntactic structural information of the sentence,a context encoder to extract semantic content information of the sentence,a common information extraction module to combine structural and contextual information,and a fusion mechanism to interact with the above three components.In addition,this paper also proposes difference and similarity losses,which are combined with the traditional cross-entropy loss function to jointly measure the difference between the model predictions and the labels.The final experimental results show that the method in this paper can effectively improve the accuracy and macro-average F1 of the predictions.It also outperforms the baseline model on all four public datasets: Laptop,Restaurant,Twitter,and MAMS.2.In order to address the challenge of effectively capturing global as well as local information and introducing external affective knowledge,this paper proposes a knowledge-enhanced graph convolutional neural network model(KGCN).The semantic features of the sentence are first captured by the self-attention mechanism and the aspect-oriented attention mechanism,and then the syntactic local features and global features of the sentence are captured according to syntactic mask matrices,and the features of the sentence are obtained by combining the two types of features.The SenticNet Affective Dictionary is introduced to increase the scores of aspect and opinion words in the sentences,to more accurately associate aspect words with their opinion words.Compared with the MultiGCN model,KGCN works better on smaller datasets,while MultiGCN is more suitable for large datasets.In addition,experimental results on several benchmark datasets show that the model proposed in this paper outperforms the current state-of-the-art methods.3.A sentiment analysis platform for student feedback texts is designed and implemented.The platform applies the model proposed in this paper to students’ feedback on different courses,realizing the integration of artificial intelligence and educational scenarios.It intuitively demonstrates the main aspects from which students encounter difficulties during the class,and provides a basis for teachers to understand students’ conditions and optimize teaching methods.