Local COVID-19 Severity and Social Media Responses: Evidence From China.

Ting Da,Liang Yang

doi:10.1109/access.2020.3037248

Abstract

Unexpected but exceedingly consequential, the COVID-19 outbreak has undermined livelihoods, disrupted the economy, induced upheavals, and posed challenges to government decision-makers. Under various behavioural regulations, such as social distancing and transport limitations, social media has become the central platform on which people from all regions, regardless of local COVID-19 severity, share their feelings and exchange thoughts. Our study illustrates the evolution of moods expressed on social media regarding COVID-19-related issues and empirically confirms the hypothesis that the severity of the pandemic substantially correlates with these sentiments by analysing tweets on Sina Weibo (China’s central social media platform). Methodologically, we leveraged Sentiment Knowledge Enhanced Pre-training, the most state-of-the-art natural language processing pre-trained sentiment-related multipurpose model, to label Sina Weibo tweets during the most distressed period in 2020. Given that the model itself does not provide a feature explanation, we utilize a random forest and linear probit model with the labelled sample to demonstrate how each word plays a role in the prediction. Finally, we demonstrate a strong negative linear relationship between the local severity of COVID-19 and the local sentiment response by incorporating miscellaneous geo-economic control variables. In short, our study reveals how pandemics affect local sentiment and, in a broader sense, provides an easy-to-implement and explanatory pipeline to classify sentiments and resolve related socioeconomic issues.

Highlights

Persistent and consequential, COVID-19 has resulted in a sequence of serious social and economic problems worldwide
Yang: Local COVID-19 Severity and Social Media Responses usually have masked language modelling [19] or next-word prediction [18], and pre-trained results can serve as the foundation for a variety of downstream sentiment analysis tasks
For simplicity and interpretation concerns, we considered only 1 − gram of words, and the tokenisation was performed using a particular version of PKUSEG [33], an algorithm for Chinese word segmentation that specializes with social media texts

Summary

INTRODUCTION

Persistent and consequential, COVID-19 has resulted in a sequence of serious social and economic problems worldwide. Yang: Local COVID-19 Severity and Social Media Responses usually have masked language modelling [19] or next-word prediction [18], and pre-trained results can serve as the foundation for a variety of downstream sentiment analysis tasks. STAGE 1 – SENTIMENT KNOWLEDGE ENHANCED PRE-TRAINING While conventional sentiment analyses are prone to separately examining different types of sentiment knowledge for diverse downstream applications, SKEP simultaneously learns three types of sentiment knowledge (i.e., sentiment words, word polarity, and aspect-sentiment pairs) Such a joint training process endows the model with greater versatility than assorted sentiment analysis tasks. Inspired by BERT, which first proposed the masked language modelling objective to pre-train the transformer encoder and achieved tremendous improvement of multifarious downstream tasks, SKEP introduces a masking process to create a noisy version of the original text sequence. The ultimate state of the classification token [CLS] (the representation of the entire text sequence) is used here to predict the pairs

STAGE 2

STAGE 3

Findings

CONCLUSION