Research On Text Sentiment Analysis Method Of Course Comment Graph Attention Network

With the advancement of the global education informatization strategy and the continuous and steady growth of education investment,major progress has been made in the construction of education informatization in various countries,and the teaching level has continued to improve.The pursuit of educational equity and quality,educational innovation,individualized education,and capacity building have become the common themes of today’s education.The "Delivery Classroom" and "Network Classroom in Famous Schools" proposed by the Ministry of Education are aimed at starting classes online and using the Internet to push appropriate high-quality educational resources according to the teaching progress,so as to help them open up and complete the nationally prescribed courses and promote education.Equitable and balanced development of education to meet students’ needs for individualized development and high-quality education.During the process of students participating in online education,a large amount of interactive data will be generated,including rich course comments and interactive messages.These unstructured interactive data contain rich emotional semantic information.How to effectively use online education?Analyzing the emotional information contained in the unstructured text in the course evaluation has become an urgent problem to be solved to help online courses improve the teaching quality.Text sentiment analysis using deep learning is an effective method,which can quickly and accurately mine potential emotional feature information in text.This paper takes the course review texts generated in online education as the actual application background,combined with deep learning technology,and conducts research through sentiment analysis at the sentence level and aspect level.The existing sentence-level sentiment analysis methods fail to fully utilize label information.Although the label embedding models proposed to utilize label information can directly calculate the correlation between text feature representations and label feature representations through similarity calculation or attention mechanism,this approach may be affected by noise interference introduced by label semantic similarity when focusing on feature words more relevant to the labels.As a result,the model finds it difficult to accurately identify text features related to the labels,thereby impacting the model’s performance.In light of this,this paper proposes a graph attention network-based sentiment classification model that combines triplet loss and label embedding.A graph is a data structure that effectively represents relationships between nodes.By applying graphs to the connection between text feature representations and label feature representations,and leveraging the graph attention network,which focuses on more task-relevant feature nodes,richer semantic information can be extracted by establishing deeper connections between text and labels.Some key words in the text can be effectively categorized,and by constructing a fusion sequence of feature words and labels and building an isomorphic graph,the graph learning process can better focus on feature words more relevant to the labels,thus enhancing the matching between labels and feature words.To address the noise introduced by label semantic similarity,a triplet loss is employed during graph learning to attenuate the impact of semantic similarity among labels while learning better label embedding feature representations to enhance semantic distinctions among labels.Finally,experiments are conducted on six public datasets,namely IMDB,Yelp.F,Yelp.P,Tan Songbo,AGnews,and Yahoo,as well as a real dataset of online course comments crawled from China University MOOC.The results show that compared to the existing label embedding model CNLE,the proposed model achieves accuracy improvements of 1.85%,2.18%,0.85%,1.11%,2.03%,and 0.26% on these datasets,and on the online course comments dataset,the accuracy and F1 score are improved by 2.05% and 1.83%,respectively,compared to CNLE.Aiming at the problem that existing aspect-level datasets lack sufficient labeled data after the emergence of new domains,traditional aspect sentiment analysis task models rely heavily on labeled data.However,for aspect-level sentiment analysis tasks,fine-grained labeled datasets are scarce.In order to alleviate the dependence on finegrained dataset labeling,previous research has mainly focused on feature-based adaptation,using sequence information to model text features.And use domain adversarial learning to bridge the gap between domains and extract common features between domains,while ignoring the syntactic information in the text.Aiming at the above issues,this paper proposes an end-to-end domain adaptive model fused with graph attention networks for aspect-level sentiment analysis tasks.The emotional words that can represent the emotional information of aspect entities in the text may not be in the context related to the aspect words.The text contains rich syntactic information.By extracting syntactic information,we can pay attention to the emotional words related to the aspect words.By constructing the syntax The dependency tree generates a syntactic dependency matrix for the feature update of the graph attention network layer,extracts the syntactic information in the text,and then passes the text feature containing the syntactic information through the gradient inversion layer to achieve the mixture of the source domain and the target domain,and extracts it to different domains The shared features between them enable the target domain to perform aspect-level sentiment analysis through domain adaptation.Finally,10 cross-domain pairs were constructed on the four public domain datasets of Restaurant,Laptop,Device and Service for experiments.Compared with the existing domain-adaptive aspect-level sentiment analysis model AHF,the F1 average is increased by 1.86.%,in the three Chinese data sets of Camera,Notebook and Phone and the course comment data set crawled on the MOOC of Chinese universities in this paper,three cross-domain pairs were constructed for experiments,compared with the AHF model,the F1 average was increased by2.46%.Finally,based on the two models proposed in this paper,a simple sentiment analysis system for course comment text is designed and implemented,and some explanations and demonstrations are given for the system.