TSTL: A Hybrid TextRank-LP and TDNN-LSTM Model for Public Opinion Monitoring in Emergency Events

Abstract

Online public sentiment during emergencies often hinders effective crisis management. Timely and accurate identification and prediction of this sentiment are vital, yet existing approaches face challenges related to high identification delays and low accuracy. To address these issues, this study proposes a model for the evolution of network public opinion based on the TextRank-Label Propagation Algorithm (TextRank-LP). The model fully utilizes the information extraction capability of TextRank-LP, while integrating the strengths of Time Delay Neural Networks and Long Short-Term Memory networks in handling time-series data. In the validation of the keyword extraction algorithm, the Natural Language Processing and Chinese Computing dataset was used, which contains tens of thousands of text samples. The results showed that the accuracy of the proposed algorithm reached 95.1%, higher than the 92.1% of Term Frequency Across Document Frequency, 88.1% of Support Vector Machine, and 86.0% of Bidirectional Long Short Term Memory network (Bi-LSTM). The F1 value of the proposed algorithm reached 94.6%. In practical testing using the 2019 NBA China controversy dataset, the model achieved a 100% accuracy rate in identifying declining online public opinion and an average recognition accuracy rate of 92.5% for each stage. Meanwhile, the model’s prediction time for online public opinion 12 months ahead is 21 minutes, far lower than Bi-LSTM’s 47 minutes. These findings indicate that the proposed model demonstrates strong recognition and predictive capabilities for public sentiment in emergencies and provides a novel approach to studying the evolution of online sentiment in such events. This method offers valuable potential for advancing more accurate and efficient research in the field of public sentiment dynamics.