A Multidimensional-Weighted TextRank and LSTM-Attention Model for Network Public Opinion Sentiment Analysis

Abstract

As social media rapidly develops, network public opinion has become an important channel for reflecting social emotions, especially in emergencies and public opinion surges. To improve the accuracy of public opinion sentiment analysis, a network public opinion sentiment analysis model integrating improved TextRank algorithm is proposed. By introducing multidimensional features such as term frequency inverse document frequency, part of speech, and word position, the keyword extraction process is improved, and combined with deep learning, the accuracy of model classification is enhanced. The findings indicated that the accuracy of the proposed model on the test set reached 0.96, and the F1 values on the training and testing sets were 92.6% and 90.9%, respectively, demonstrating the advantages of this method in complex sentiment analysis tasks. In addition, the model proposed by the research performed well in the sentiment classification task of four network public opinion hotspots, with the highest accuracy rates of positive and negative sentiment classification reaching 98% and 96% respectively, a root mean square error as low as 0.176, and a mean absolute percentage error of only 0.081. The results indicate that the model has better fitting and generalization abilities in sentiment classification tasks. This not only provides an efficient technical solution for sentiment analysis of network public opinion, but also lays an important foundation for the intelligent development of social media public opinion monitoring systems.