[Paper Review] Analyzing COVID-19 on Online Social Media: Trends, Sentiments and Emotions

Xiaoya Li, Mingxin Zhou+4 more|arXiv (Cornell University)|May 29, 2020

Misinformation and Its Impacts58 references34 citations

TL;DR

This paper analyzes COVID-19 related posts on Twitter and Weibo from Jan 20 to May 11, 2020 to map topic trends, six basic emotions, and emotional triggers, comparing US and China, using semi-supervised retrieval and BERT-based emotion tagging with trigger extraction.

ABSTRACT

At the time of writing, the ongoing pandemic of coronavirus disease (COVID-19) has caused severe impacts on society, economy and people's daily lives. People constantly express their opinions on various aspects of the pandemic on social media, making user-generated content an important source for understanding public emotions and concerns. In this paper, we perform a comprehensive analysis on the affective trajectories of the American people and the Chinese people based on Twitter and Weibo posts between January 20th, 2020 and May 11th 2020. Specifically, by identifying people's sentiments, emotions (i.e., anger, disgust, fear, happiness, sadness, surprise) and the emotional triggers (e.g., what a user is angry/sad about) we are able to depict the dynamics of public affect in the time of COVID-19. By contrasting two very different countries, China and the Unites States, we reveal sharp differences in people's views on COVID-19 in different cultures. Our study provides a computational approach to unveiling public emotions and concerns on the pandemic in real-time, which would potentially help policy-makers better understand people's need and thus make optimal policy.

Motivation & Objective

• Understand how public emotion and concerns about COVID-19 evolved over time on Twitter and Weibo.
• Identify fine-grained emotions (anger, disgust, fear, happiness, sadness, surprise) and their triggers.
• Contrast public reactions between the US and China to reveal cultural differences in pandemic perception.
• Develop a real-time computational approach to extract public sentiment and concerns to inform policy.

Proposed method

• Bootstrap-based semi-supervised retrieval to identify COVID-19 related posts using seed keywords, iterative retraining, and saliency-based keyword expansion.
• Six-way multi-label emotion classification on English tweets using BERT with emotion descriptions as prompts (sigmoid outputs).
• Chinese emotion classification using a described-BERT model on Weibo data with culturally relevant labels.
• Emotion intensity S(t,y) computed as daily average P(y|x) over posts, with non-COVID texts assigned zero probability.
• Trigger extraction for emotions via a CRF tagger with BERT-MRC features, enhanced by POS, dependency, and Twitter-specific features.
• Unsupervised clustering of triggers using LDA on top trigger mentions to discover subcategories and topics over time.

Experimental results

Research questions

• RQ1 What are the temporal dynamics of topic prevalence and emotional states related to COVID-19 on Twitter and Weibo?
• RQ2 How do six basic emotions fluctuate during the pandemic, and what are their semantic triggers?
• RQ3 What differences emerge in public sentiment and triggers between the US and China?
• RQ4 Can a semi-supervised, real-time pipeline effectively harvest and track COVID-19 related posts and their sentiment?
• RQ5 What subcategories of triggers under anger and worry best explain public concerns over time?

Key findings

• COVID-related posting intensity on Weibo peaks in late January–February and again in March, while Twitter shows rising attention from March onward with no clear decline at the time of writing.
• Emotion tagging using BERT-based multi-label classifiers achieves micro F1 of 75.2 and macro F1 of 68.3 for English tweets; BERT-description reaches 77.0 macro F1.
• Weibo shows worry peaking with general post intensity; anger surges around Feb 8 due to Li Wenliang’s death; Twitter anger and worry rise with US outbreak and policy events.
• Top anger triggers include lockdown, quarantine, and public figures (Trump, Pence), plus China-related topics; worry triggers include jobs, finances, family concerns, and virus transmission.
• Trigger clustering with LDA reveals interpretable topics like China-related anger, lockdown, and hospital treatments; worry topics include finances, families, and increasing deaths/cases.
• Flowers: bootstrapping three rounds yielded F1 scores of 0.74, 0.82, and 0.86 for COVID-related tweet classification (early rounds).
• Overall, the study demonstrates a real-time capable framework to quantify public affect and its drivers during a global crisis.