Public Opinion Analysis for the Covid-19 Pandemic Based on Sina Weibo Data

Wang, Feng; Gong, Yunpeng

doi:10.1007/978-3-031-20738-9_109

Cited by 2 publications

(2 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…At first, scholars generated event vectors by manually defining or extracting qualitative and quantitative event features from social media. Research [2] uses 33 artificial features to assign values to public opinion events; study [3] made a secondary improvement on the basis of [2], reducing 33 features to 30 features; author in [4] combines Twitter and Weibo, and proposes some characteristics to analyze public opinion based on their commonalities; researcher in [5] combines time features and proposes new event features for rumor analysis. The manually defined event features are very dependent on expert knowledge which usually can only be obtained by observing the influencing factors of one period or even the whole period of the event.…”

Section: A Manually Designing Featuresmentioning

confidence: 99%

“…The Generated Event Vectors 1) Experiment setup: We compare our experimental results with the results of three common models: RoBERTa [24], XLnet [25], Manual designing features [3]. We generate the vectors respectively by the above four models, the first three of which are trained with the blog texts of events, and our proposed NL2ER-Transformer Encoder.…”

Section: Evaluation Of Downstream Reversal Prediction Ability Formentioning

confidence: 99%

See 1 more Smart Citation

Event Feature Pre-training Model Based on Public Opinion Evolution

Wang¹,

Shu-Ru²,

Xiao-Lan³

et al. 2023

IJACSA

View full text Add to dashboard Cite

The comments in the evolution of network public opinion events not only reflect the attitude of netizens towards the event itself, but also are the key basis for mastering the dynamics of public opinion. According to the comment data in the event evolution process, an event feature vector pre-training model NL2ER-Transformer is constructed to realize the realtime automatic extraction of event features. Firstly, a semisupervised multi-label curriculum learning model is proposed to generate comment words, event word vectors, event words, and event sentences, so that a public opinion event is mapped into a sequence similar to vectorized natural language. Secondly, based on the Transformer structure, a training method is proposed to simulate the evolution process of events, so that the event vector generation model can learn the evolution law and the characteristics of reversal events. Finally, the event vectors generated by the presented NL2ER-Transformer model are compared with the event vectors generated by the current mainstream models such as XLNet and RoBerta. This paper tests the pre-trained model NL2ER-Transformer and three pretrained benchmark models on four downstream classification models. The experimental results show that using the vectors generated by NL2ER-Transformer to train downstream models compared to using the vectors generated by other pre-trained benchmark models to train downstream models, the accuracy, recall, and F1 values are 16.66%, 44.44%, and 19% higher than the best downstream model. At the same time, in the evolutionary capability analysis test, only four events show partial errors. In terms of performance of semi-supervised model, the proposed semi-supervised multi-label curriculum learning model outperforms mainstream models in four indicators by 6%, 23%, 8%, and 15%, respectively.

show abstract

Section: A Manually Designing Featuresmentioning

confidence: 99%