Rise and fall of the global conversation and shifting sentiments during the COVID-19 pandemic

The study investigates how global online conversations and sentiments evolved during the COVID-19 pandemic and what events drove these dynamics across cultures. Given the unprecedented worldwide impact of COVID-19 on health, economy, and mobility, and the central role of social media during confinement, the authors seek to quantify the rise/fall in discussion volume and track fine-grained emotions across six major languages. The purpose is to understand population reactions to key pandemic events (e.g., lockdowns, market crashes, policy announcements) and to inform public health interventions by providing multilingual, large-scale sentiment surveillance using state-of-the-art deep learning models. This is important as traditional sentiment analysis is often coarse-grained, monolingual, or limited by small training data, whereas pandemic-related emotions are diverse and rapidly evolving.

Prior research has used sentiment analysis to monitor reactions to events and inform public health interventions, often employing lexicons or shallow models and focusing on single languages or regions. COVID-19-related studies have included unsupervised topic modeling with emotion lexicons, basic positive/neutral/negative classifications, and small supervised datasets. Multilingual sentiment benchmarks lack categories common in pandemic discourse (e.g., joking/irony, denial/conspiracy). Studies in India, Saudi Arabia, the Philippines, and anti-Asian hate analyses demonstrate localized approaches. This work expands by providing a multilingual, fine-grained, deep-learning approach across six languages over a global time frame early in the pandemic.

Data collection: Tweets were collected via Twint using a unified multilingual query: “COVID-19 OR coronavirus OR covid OR corona OR كورونا,” averaging slightly over 1M tweets/day from Mar 1 to May 15, 2020, saved as JSON. Due to limited Chinese tweets, Chinese data came from Sina Weibo (Jan 20–May 15, 2020) via API by harvesting COVID-19 hashtags and associated posts. Processing and modeling were conducted on a GPU server (GTX 1080 Ti, 20 CPUs). Sentiment annotation: Due to lack of a fine-grained benchmark, 10,000 English and 10,000 Arabic tweets were annotated by >50 trained annotators (≥3 labels/tweet, multilabel allowed) with majority voting. Ten labels: optimistic, thankful, empathetic (incl. praying), pessimistic, anxious, sad, annoyed (angry), denial (conspiracy-related), official report, and joking (irony). For Spanish, French, and Italian, labeled English tweets were machine-translated (Google Translate). Translation quality was assessed via round-trip BLEU ≈ 0.33 and manual checks. For Weibo, 21,173 posts were annotated under seven categories: optimistic, thankful, surprised, fearful, sad, angry, disgusted. Preprocessing: Removed user/interaction metadata, emojis/emoticons, hyperlinks, RTs, usernames, and special symbols; retained hashtags for semantic content. Tokenization/stemming/pos-tagging used NLTK (En/Es/Fr/It), PyArabic (Ar), and Jieba (Zh). Models: Built multilabel classifiers using SimpleTransformers to fine-tune pretrained language models, adding a fully connected layer with sigmoid outputs: XLNet (English), AraBERT (Arabic), BERT (Spanish, French, Italian), ERNIE (Chinese). Representations are 768-d; per-label probabilities thresholded (e.g., 0.5) allow multilabel assignment. Training and validation: Fivefold cross-validation achieved accuracies: English 0.847±0.004; Arabic 0.905±0.002; Spanish 0.823±0.001; French 0.824±0.004; Italian 0.827±0.002; Chinese 0.880±0.001. Final models trained on 10,000 labeled tweets per Twitter language; Chinese trained on 21,173 Weibo posts. Trained models predicted sentiments for millions of posts in the study windows. Topic analysis: For English, tweets were grouped by topics reflecting major concerns: oil/stock prices, economic stimulus, employment, herd immunity, medicine/vaccine, and working/studying from home, and sentiment trajectories were analyzed per topic.

- Conversation volume dynamics: All languages exhibited a rapid rise and gradual decline. Chinese Weibo peaked on Jan 22, 2020 (Wuhan lockdown), ~2 months before Twitter languages, which peaked Mar 12–21, 2020. Weekly cycles showed troughs on weekends. - Drivers of surge: Two main drivers were confinement measures (lockdowns, closures) and economic shocks (stock market/oil price crashes on Mar 9, 2020). Post-crash conversations intensified around unemployment and economic stimulus (peak Mar 26). Interest shifted from herd immunity (early) to medicine/vaccines (sustained). Work/study-from-home peaked Mar 17. - Sentiment trajectories: Early period (Mar 1–15) in En/Es/Fr/It showed mixes of joking with anxious/pessimistic/annoyed; Arabic showed anxious, denial, and empathetic (praying). After Mar 15, positive sentiments (optimistic, thankful, empathetic) generally increased as control improved, strongest in Arabic. - Event-linked spikes: Language-specific spikes occurred (e.g., English annoyed on Apr 19 and May 10; Spanish annoyed and sad around late March and anxious on May 5; French denial spikes Apr 17–18; Italian persistent pessimistic/sad in April, rising anxious/sad in May). - Chinese Weibo: Fear exceeded 50% after confirmation of human-to-human transmission (Jan 20) through Wuhan lockdown; positive (optimistic/thankful) rose Jan 23–25; notable sad spikes on Feb 4 (20,000 cases), Feb 7 (death of Dr. Li Wenliang), Apr 4 (national memorial); optimism rose with Wuhan reopening (Apr 8). - Cross-language comparisons: Overall prevalence similar with sizable optimistic and joking (~20% each) and negative states like annoyed and anxious. Optimistic and sad tended to increase over time; joking decreased. Arabic had highest empathetic and strong growth in optimistic (partly overlapping Ramadan). Correlations of daily sentiment distributions were high among Es/Fr/It (0.97–0.99), with English closer to Spanish/Italian than French; Arabic was less correlated. t-SNE showed distinct clusters pre- vs post-peak and March vs later for Arabic. - Topic-specific sentiments (English): Work/study-from-home was most optimistic/thankful; herd immunity elicited denial/anxiety; economic collapse topics showed strongest annoyed (stock market), pessimistic (oil prices), and sad (unemployment); economic stimulus had more joking and annoyed; medicine/vaccines had strong optimism with denial/annoyed around controversies (e.g., unproven drugs, vaccine poaching). - Classifier performance: Validated accuracies >0.82 across all languages (En 0.847; Ar 0.905; Es 0.823; Fr 0.824; It 0.827; Zh 0.880).

The global conversation mirrored the pandemic’s spread, with China’s peak preceding Western and Arabic languages by two months and a common rapid rise followed by gradual decline. Confinement decisions and economic shocks jointly drove surges (except in China where confinement dominated). Despite cultural differences, cross-language sentiment patterns were remarkably similar: strong early negative reactions (and joking) transitioned toward increasing positive emotions as control measures and reopening progressed. Topic analyses revealed particularly negative reactions to herd immunity strategies and misinformation/conspiracy theories, while work/study-from-home and vaccine/therapy discussions saw more optimism. High cross-language sentiment correlations (notably among Es/Fr/It) indicate synchronized societal responses. The strong positive response in Arabic suggests potential cultural or experiential resilience (e.g., prior MERS experience) and overlap with Ramadan encouraging empathy. These insights highlight sentiment analysis as a tool for monitoring societal states and informing public health communication and policy, particularly identifying when anxiety, denial, or annoyance spike around specific events or topics.

This study delivers a large-scale, multilingual, fine-grained sentiment analysis of COVID-19 social media, showing globally synchronized rise-and-fall conversation dynamics and broadly similar emotional trajectories across six languages. It identifies key drivers (lockdowns, economic shocks) and topic-specific emotional imprints (e.g., denial/anxiety around herd immunity; optimism around work-from-home and vaccines). Positive emotions generally increased as outbreaks came under control, with Arabic showing the strongest positive trend. Contributions include: (1) creation of multilingual, multilabel classifiers with high accuracy; (2) cross-cultural comparison of fine-grained emotions; (3) event- and topic-linked sentiment mapping to inform interventions. Future research could extend timeframes, incorporate additional platforms and languages, integrate non-textual signals (e.g., emojis, images), improve multilingual annotations beyond translation, and link sentiment trends to policy outcomes or health metrics for causal insights.

- Temporal and platform scope: Twitter data span Mar 1–May 15, 2020, and Chinese data are from Weibo (Jan 20–May 15), which may limit generalizability beyond these windows and complicate cross-platform comparisons. - Annotation and translation: Gold labels were produced for English/Arabic; other languages relied on machine-translated English labels (BLEU ≈ 0.33) plus spot checks, which may introduce translation bias. Chinese labels used a different 7-category scheme. - Signal reduction: Emojis/emoticons and some metadata were removed, potentially discarding emotional cues. Neutral/official reports were excluded. - Model performance: Although accuracies exceeded 0.82, classification errors persist, especially for nuanced categories (e.g., joking, denial). - Topic assignment and coverage: Topic analyses focused on selected English topics; cross-language topic comparisons were not performed. - Causality: Associations between events and sentiments are observational; causal claims cannot be established.