News media in crisis: a sentiment and emotion analysis of US news articles on unemployment in the COVID-19 pandemic

The study examines how mainstream news media, specifically the New York Times (NYT), represented public perception and emotions related to unemployment during the COVID-19 pandemic in 2020. The pandemic caused severe disruptions to health, economy, and mental well-being, including heightened anxiety and stress exacerbated by unemployment. While many sentiment analyses during COVID-19 focused on social media, few analyzed unemployment-related coverage in mainstream news using combined sentiment and corpus-based discourse methods. This research addresses that gap by exploring: (1) what sentiments and emotions were conveyed in unemployment-related news; (2) how these sentiments and emotions fluctuated over time with the evolution of the pandemic and unemployment; and (3) which significant events and themes are linked to these emotions. The purpose is to capture the emotional dynamics over 2020 and understand the role of news media in shaping public perception during crisis, highlighting its importance in information dissemination and framing policy responses.

The literature distinguishes between supervised/machine learning and lexicon-based approaches for sentiment analysis, with increasing use of hybrid methods. During COVID-19, numerous studies analyzed public emotions about economic issues using social media (especially Twitter), showing links between unemployment rates and online sentiment, often finding dominant negative emotions early in the pandemic. Studies also evaluated public perception of government policies combating unemployment with mixed results. However, little attention has been paid to unemployment-related sentiments in mainstream news media, which differ from social media in scope and function—news provides broader societal narratives and policy coverage. For this study, the NRC Emotion Lexicon is chosen to capture both polarity (positive/negative) and eight basic emotions, and a corpus-based discourse analysis complements lexicon limitations by identifying contextually salient themes and events associated with detected emotions.

Data collection: Articles were drawn from the New York Times (NYT), selected for its popularity, trust, and influence. Using LexisNexis, articles mentioning at least one of the terms "layoff*", "lay off", "laid off", or "unemploy*" from 01/01/2020 to 12/31/2020 were retrieved. After removing non-U.S. articles, the corpus comprised 4,921 articles totaling 8,016,170 words. Articles were organized into 12 monthly subcorpora. Text purification and cleaning: Using Python, texts were lowercased, tokenized, lemmatized, and part-of-speech tagged; duplicates, URLs, punctuation, stop words, and non-English characters were removed. Sentiment and emotion extraction: The NRC Emotion Lexicon (Mohammad and Turney, 2013) was used to annotate tokens with two polarity sentiments (positive/negative) and eight emotions (anger, anticipation, surprise, trust, disgust, fear, joy, sadness). Python scripts matched words in each monthly subcorpus to the lexicon and computed monthly sentiment and emotion values, enabling temporal trajectory analysis. Corpus-based discourse analysis: To contextualize lexicon results, corpus linguistic techniques (frequency, colligation, collocates by likelihood, clusters) identified salient themes/events tied to emotions. For each emotion and month, the top 10 associated words were extracted; across the year, terms appearing at least four times across monthly lists were selected for deeper analysis using AntConc to examine frequencies, colligations, collocates, and clusters.

• Coverage–unemployment correlation: NYT unemployment-related coverage closely tracked U.S. unemployment rates in 2020, with significant correlation (Pearson r=0.868, p=0.00). Coverage spiked in March–April alongside unemployment peaking at 14.7% in April 2020.
• Sentiment polarity: Positive sentiment exceeded negative sentiment throughout 2020. Descriptive statistics: positive mean=0.050917 (SD=0.001441); negative mean=0.033083 (SD=0.001382). Positive and negative sentiments were significantly negatively correlated over time (r=-0.624, p=0.03). Three turning points occurred: February (positive fell, negative rose), March (decline/rise stabilized), and August (positive rose, negative fell).
• Emotions: All eight NRC emotions were present. Mean values and rank order: trust (0.034167), anticipation (0.02275), fear (0.02), sadness (0.016583), joy (0.014333), anger (0.01275), surprise (0.01075), disgust (0.007167). Emotions shared the same turning points (Feb, Mar, Aug), though changes in August were less marked for some (e.g., disgust).
• Positive emotions context: High-frequency “trust” terms centered on government/economic policy and daily-life essentials (e.g., president, congress, policy, economy, money, school, pay, food). Collocational evidence highlighted policy actions (e.g., President signing relief, Congress passing stimulus). “Anticipation” terms reflected planning/future effects (time, long/long-term, plan, public, money, pay), with collocates such as reopen/reopening and stimulus/rescue packages.
• Negative emotions context: “Fear” was the strongest negative emotion, peaking in June 2020, linked to pandemic health risks (pandemic, risk, emergency, case, medical, death, disease) and labor market concerns (emergency unemployment benefits; labor/work force). A distinct driver was coverage of policing and racial justice (police; collocates include brutality, Minneapolis, protest, Floyd), following George Floyd’s death. “Sadness” was driven by deaths/disease and financial/job insecurity, remained high March–August, dipped, then rose again October–December with renewed pandemic surges. Joy–sadness correlation was weakly negative (r=-0.273, p=0.390).

Findings show unemployment-related emotions in NYT were shaped by multiple, time-varying forces: (1) the pandemic’s evolution and its immediate impacts on health and employment (with sentiment shifts beginning in February and cresting in March–April, and renewed negativity in late 2020 amid surges); (2) rapid policy responses (e.g., fiscal and monetary measures including relief/stimulus packages), which were widely covered and coincided with containment of negative sentiment from March onward and a positive shift after August; and (3) racial justice protests sparked by George Floyd’s killing, which contributed to elevated fear from June through their association with policing and inequality, highlighting disproportionate unemployment and health burdens among Black and Hispanic communities. Compared to social media studies that often found negativity dominant, mainstream news showed higher positive sentiment, likely reflecting broader, policy-focused, solution-oriented reporting by professional journalism. This supports the media’s role in disseminating public health and policy information, framing debates on economic relief, and potentially evoking constructive emotions that can bolster public confidence during crises.

The study integrates sentiment, emotion, corpus-based discourse, and timeline analyses to chart emotional dynamics in NYT unemployment-related reporting across 2020. It finds consistently higher positive than negative sentiment; “trust” and “anticipation” as leading emotions; and “fear” and “sadness” as the most prominent negative emotions. Emotional fluctuations align with the pandemic trajectory, unemployment changes, policy responses, and protests against racial inequalities. Contributions include: focusing on mainstream news (rather than social media), conducting a year-long longitudinal assessment to capture dynamics, and triangulating with corpus-based discourse analysis to link emotions to salient events/themes. Practical implications stress the importance of immediate policy interventions during mass job loss and the centrality of traditional news media in communicating economic relief and solutions, which can foster hope and confidence. Future research should broaden media sources and combine lexicon-based with machine learning approaches to enhance robustness and contextual sensitivity.

• Lexicon-based analysis relies on word counts without full contextual disambiguation; words can have multiple senses, potentially reducing accuracy despite complementary discourse analysis.
• Only one newspaper (NYT) was analyzed, limiting representativeness. Future work should include multiple outlets and integrate lexicon-based and machine learning methods for improved accuracy and generalizability.