[Paper Review] Understanding Public Sentiments, Opinions and Topics about COVID-19 using Twitter

Jolin Shaynn-Ly Kwan, Kwan Hui Lim|arXiv (Cornell University)|Dec 5, 2020

Misinformation and Its Impacts18 references4 citations

TL;DR

This paper proposes a multi-method framework integrating sentiment analysis, emotion detection, and topic modeling on 910,000 geolocated English tweets from March to June 2020 to study public sentiment, emotions, and discussion topics during the early phase of the COVID-19 pandemic. Key findings reveal regional sentiment shifts—such as increased trust and joy in Singapore post-circuit breaker and rising fear/sadness in the UK during policy uncertainty—alongside a longitudinal shift from negative to more adaptive, technology-focused topics.

ABSTRACT

The COVID-19 pandemic has caused widespread devastation throughout the world. In addition to the health and economical impacts, there is an enormous emotional toll associated with the constant stress of daily life with the numerous restrictions in place to combat the pandemic. To better understand the impact of COVID-19, we proposed a framework that utilizes public tweets to derive the sentiments, emotions and discussion topics of the general public in various regions and across multiple timeframes. Using this framework, we study and discuss various research questions relating to COVID-19, namely: (i) how sentiments/emotions change during the pandemic? (ii) how sentiments/emotions change in relation to global events? and (iii) what are the common topics discussed during the pandemic?

Motivation & Objective

• To understand how public sentiment and emotions evolved during the early phase of the global COVID-19 pandemic.
• To examine the relationship between public emotions and major global events, such as government policy changes and health announcements.
• To identify and track evolving discussion topics related to the pandemic across different regions and time periods.
• To develop a holistic analytical framework combining sentiment, emotion, and topic modeling for public health insights.

Proposed method

• Collected 910,000 public English tweets from March 12 to June 10, 2020 using three datasets, including geotagged and profile-based location inference.
• Preprocessed tweets via lowercase conversion, tokenization, lemmatization, and stemming to prepare for analysis.
• Applied the NRC Word-Emotion Association Lexicon and AFINN lexicon to detect sentiment polarity and eight core emotions (e.g., fear, joy, sadness, trust).
• Used Latent Dirichlet Allocation (LDA) for longitudinal topic modeling to identify evolving public discussion themes.
• Conducted spatial and temporal analysis to compare sentiment and emotion trends across regions (e.g., Singapore, UK) and timeframes.
• Correlated emotional shifts with key global events, such as lockdown announcements and policy changes, to assess impact on public mood.

Experimental results

Research questions

• RQ1 How do sentiments and emotions change during the COVID-19 pandemic?
• RQ2 How do sentiments and emotions shift in response to major global events, such as government policy changes or health announcements?
• RQ3 What are the dominant discussion topics related to COVID-19, and how do they evolve over time?
• RQ4 How do regional differences in sentiment and emotion reflect local policy and public response?

Key findings

• In Singapore, the week of April 9–15 saw a notable drop in fear (11.9%) and sadness (9.3%) and a rise in trust (25.1%) and joy (15.7%), indicating public relief following the circuit breaker lockdown.
• In the UK, from April 30 to May 6, fear (16.2%) and sadness (11.9%) increased significantly, while joy decreased to 11.4%, reflecting public anxiety during policy transition and messaging confusion.
• The UK’s shift from 'stay home' to 'stay alert' messaging was criticized as confusing, correlating with rising negative emotions and declining positive sentiment.
• Longitudinal topic modeling revealed a shift from negative themes (e.g., violence, marginalization) to more adaptive topics like remote work and e-commerce, signaling societal adaptation.
• The framework successfully captured regional emotional responses to policy changes, demonstrating its utility for public health monitoring.
• The integration of sentiment, emotion, and topic modeling enabled nuanced insights into public perception, such as how policy clarity influences emotional well-being.