Sentiments and emotions evoked by news headlines of coronavirus disease (COVID-19) outbreak

The study investigates how the COVID-19 outbreak, coupled with an overwhelming volume of news (an “infodemic”), shapes public emotions and sentiments. Motivated by concerns that intensive and sensational media coverage may amplify fear, anxiety, and stigma beyond actual health risks, the authors analyze global English-language news headlines to quantify their emotional valence. The context includes prior coronavirus outbreaks (SARS in 2003 and MERS in 2012), the role of media in communicating uncertainty, and the psychological effects of chronic threats. The research question centers on characterizing the polarity (positive, negative, neutral) and discrete emotions evoked by COVID-19 news headlines over time, and discussing implications for emotional wellbeing and socio-economic dynamics.

The paper situates its work within sentiment analysis and emotion recognition in text. It highlights lexicon-based and machine learning approaches to sentiment classification (e.g., Turney 2002; Pang and Lee 2008; Taboada et al. 2011) and the use of affective lexicons that associate words with sentiments and emotions (e.g., Mohammad and Turney 2010, 2013). It notes challenges of applying sentiment analysis to real-time, evolving web content (Bermingham and Smeaton 2010) and the importance of accounting for valence shifters (negators, intensifiers, downtoners, adversatives) to avoid misclassification. The authors also reference empirical links between emotions and health-related domains and media’s role in shaping perceptions and behavior during outbreaks, drawing parallels to SARS and Ebola coverage and broader socio-economic impacts.

Data sources: Headlines were obtained from a live news dashboard (https://visualizenow.org/corona-news) fed by publicly available media sources and the Johns Hopkins University COVID-19 repository. The dataset comprises 141,208 English-language news headlines containing the keyword “coronavirus” from top 25 global sources (e.g., Reuters, BBC, Yahoo News, South China Morning Post, National Post, Daily Mail UK, CNBC, The Guardian, CNN), spanning January 15, 2020 to June 3, 2020. Preprocessing and text analytics: Headlines were converted to text, lowercased, and cleaned by removing punctuation, stopwords (174 English stopwords via the R tm package), numbers, and extra whitespace; stemming was applied. A corpus and a Document-Term Matrix were built using R package tm. The raw corpus contained 1,619,987 words; after cleaning, 6,488,545 tokens with 31,130 unique terms remained. Frequency analyses summarized term distributions and identified most common words. Sentiment polarity estimation: Sentence-level unbounded polarity scores were computed using an R sentiment analysis package that tags polarized words and incorporates valence shifters (negators, amplifiers/intensifiers, de-amplifiers/downtoners, adversative conjunctions). A window of 4 words before and 2 words after polarized terms captured shifters. Scores of headlines were classified as negative (score < 0), neutral (score = 0), or positive (score > 0). The algorithm uses a sentiment dictionary and modifiers, following methods documented by Jockers (2017) and related R implementations. Emotion analysis: The NRC Word-Emotion Lexicon was applied to quantify the presence of eight emotions (anger, fear, anticipation, trust, surprise, sadness, joy, disgust). Emotion scores range from 0 (no emotional words) to 1 (all words emotional). Majority-class consolidation across annotations was used for target terms’ emotion intensities. Outputs: The study produced descriptive statistics (boxplots, histograms), temporal sentiment trajectory (January 15–June 3), frequency distributions of sentiment categories, comparison plots of terms contributing to positive vs negative polarity, and distributions of the eight emotions.

• Overall sentiment distribution: 51.66% of headlines were negative, 30.46% positive, and 17.87% neutral.
• Polarity statistics: Headline sentiment scores ranged from -1.85 to 1.54, with an average of -0.08, indicating a skew toward negativity.
• Temporal trajectory: From January 15 to June 3, the aggregate emotional valence remained in the negative region throughout, with increased negativity toward the end of the period and no transition into neutral or positive territory.
• Emotion distributions (NRC): Fear (20%), anticipation (15%), sadness (14%), and anger (11%) dominated (≈61% combined). Trust (17%) and surprise (9%) were also present; joy and disgust were comparatively lower.
• Term contributions: 3,833 negative terms contributed to negative sentiment (e.g., pandemic, trump, outbreak, virus, death, crisis, fear, fight, government, warn, die, emergency, police, risk, symptoms, hospital, isolation, infected, ban). Positive sentiment drew on 2,135 terms (e.g., positive, care, global, work, relief, aid, food, free, working, markets, study, patient, league, support, star, big, extend, expert, protect, fans).
• Frequent words in headlines reflected the crisis focus: covid (13,388), lockdown (9,133), case (8,817), trump (8,516), death (8,047), test (7,537), pandemic (6,818), China (6,493), outbreak (4,907), virus (4,773), among others.
• Emotion-evoking keywords: Fear linked to terms like death, quarantine, hospital, fight, epidemic, infection, disease, battle, threat; sadness linked to death, isolation, fatality, disease, emergency, hospital, shortage, victim; surprise associated with death, trump, emergency, surge, scare; trust with hospital, united, confirmed, economy, medical, advisor, treat, official, save, school, good.

The predominance of negative polarity and high emotional content in COVID-19 news headlines suggests substantial potential impacts on public mental health and behavior. Persistent themes of death, crisis, and emergency can amplify fear, anxiety, and perceptions of threat, contributing to stress-related disorders, stigmatization, and social disruption (e.g., panic, xenophobia, and quarantine-related distress). Media sensationalism and information uncertainty may exacerbate these effects, driving help-seeking among low-risk individuals and straining health systems. The findings align with patterns observed during prior outbreaks (SARS, Ebola) and have broader socio-economic implications, including diminished consumer and investor confidence, market volatility, and disruptions to trade, tourism, and employment. The results underscore the need for clear, balanced risk communication and psychosocial support strategies to mitigate adverse emotional and economic outcomes during protracted health crises.

Analyzing 141,208 global English-language news headlines on COVID-19 (Jan 15–Jun 3, 2020), the study finds that media coverage evoked predominantly negative sentiment with strong emotional intensity, especially fear, anticipation, sadness, and anger. Sentiment remained negative throughout the observation period, and term analyses highlight crisis- and threat-related language as key drivers. These insights contribute empirical evidence on the emotional valence of pandemic-related news and its potential implications for public wellbeing and socio-economic dynamics, emphasizing the importance of accurate, responsible communication during global health emergencies.