The COVID-19 pandemic, unprecedented in its scale and global impact, presented an unparalleled opportunity to study human sentiment and response during a worldwide crisis. The study aimed to analyze the global rise and fall of sentiments during the pandemic utilizing sentiment analysis on a massive dataset of social media posts (Twitter and Weibo). This is crucial for understanding public perception, reactions, and cultural differences during such a significant event. Social media provides a real-time reflection of public sentiment, offering invaluable insights into public health concerns, anxieties, and coping mechanisms. The researchers utilized deep learning models to go beyond simple positive/negative classifications, capturing the nuance of emotions expressed during the pandemic. Previous sentiment analysis studies often focused on single languages or limited datasets, and this study addresses these limitations by encompassing a diverse dataset and employing advanced natural language processing techniques. The study’s extensive dataset allows for a granular analysis of emotional trends and cross-cultural comparisons, providing a comprehensive picture of the global emotional response to the pandemic.

Existing literature highlights the usefulness of sentiment analysis in monitoring public reactions to events, particularly in public health crises. Studies have shown the value of this approach in designing interventions and combating misinformation. However, many previous analyses had limitations, such as focusing on a single language, employing coarse-grained sentiment categories, and relying on limited training data. Some studies used unsupervised methods or supervised learning with small datasets, resulting in less comprehensive or accurate insights. There is a gap in research concerning a fine-grained analysis of emotions during pandemics, a challenge this study directly addresses. The research team also noted the lack of large, multilingual sentiment benchmarks for fine-grained emotion analysis of COVID-19 related conversations. This necessitated the creation of a substantial annotated dataset for training their deep learning models, overcoming a significant methodological hurdle in the field.

The study collected over 105 million tweets (March 1-May 15, 2020) and Weibo messages (January 20-May 15, 2020) across six languages. Data collection involved using Twint (an open-source Twitter crawler) and the Sina Weibo API. Preprocessing steps included removing user information, emojis, emoticons, noisy symbols and hyperlinks. Hashtags were retained due to their semantic significance. Word tokenization and stemming were performed using NLTK and Pyarabic. For Weibo, Jieba was used for segmentation. For sentiment annotation, 10,000 tweets each in English and Arabic were randomly selected and labeled by over 50 annotators, covering ten sentiment categories (optimistic, thankful, empathetic, pessimistic, anxious, sad, annoyed, denial, official report, joking). Annotation reliability was assessed using majority voting. Data for Spanish, French, and Italian were generated by translating labeled English tweets using Google Translate. Translation quality was verified using the BLEU score. For Weibo, 21,173 posts were annotated under seven sentiment categories. Multilabel sentiment classifiers were built using deep learning language models (XLNet for English, AraBert for Arabic, BERT for Spanish, French, and Italian, and ERNIE for Chinese). These pretrained models were fine-tuned using the simpletransformer framework. A fully connected network with a sigmoid activation function was added for prediction. Five-fold cross-validation was used to evaluate model accuracy. After validation, models were trained on the annotated data and applied to the main dataset to predict sentiments of millions of posts.

The study revealed a similar pattern of rapid increase followed by gradual decrease in the volume of COVID-19 related conversations across all six languages. However, the peak in Chinese conversations occurred earlier (January 22, 2020) than in other languages (March 12-21, 2020), likely due to the earlier outbreak in Wuhan. The surge in conversation volume was driven by both government-imposed confinement measures and the economic collapse. A clear weekly pattern was observed, with reduced conversation volume on weekends. Analysis of emotional expression showed remarkably similar emotional states across languages as the epidemic escalated to a pandemic. Initially, a mix of joking and negative emotions (anxious, pessimistic, annoyed) dominated. As the pandemic came under control, there was a general shift towards positive emotions (optimistic, thankful, empathetic), with Arabic tweets showing the strongest increase in positive sentiments. Specific events triggered significant, transient emotional shifts in particular languages. For instance, negative sentiments spiked in English after reports of government negligence and in Spanish after reports of deaths and EU’s failure to agree on a stimulus package. Chinese Weibo posts showed a prevalence of fearful states initially, followed by an increase in positive emotions as the situation improved. Overall, optimistic and sad states tended to increase over time, while joking decreased. Arabic speakers showed the highest empathetic sentiments. Analysis of sentiments related to different topics (e.g., economic stimulus, herd immunity) showed varying emotional patterns, with negative emotions linked to topics like herd immunity, and positive emotions linked to work from home and vaccine discussions. Correlation analysis revealed high similarity in sentiment trends across Spanish, French, and Italian, while Arabic showed less similarity. Principal component analysis and t-SNE visualizations further supported these findings.

This study's findings address the research question by demonstrating the global emotional response to the COVID-19 pandemic across languages and time. The remarkably similar patterns of conversation volume and emotional trends across different cultures suggest a shared human experience in the face of a global crisis. The initial dominance of negative emotions reflects the stress and uncertainty caused by the pandemic's health and economic impacts, while the increase in positive emotions as the situation improved points to resilience and hope. The cultural differences observed in the emotional responses, particularly the higher levels of empathy and optimism in Arabic tweets, highlight the importance of considering cultural context in interpreting social media data. This research provides insights into the psychological impact of pandemics, demonstrating the potential for cross-cultural analysis of social media data to enhance understanding of public sentiment and inform public health interventions.

This research provides a novel and large-scale multilingual analysis of sentiment expressed on social media during the COVID-19 pandemic. The study highlights the power of social media as a real-time reflection of public sentiment and its potential to inform public health strategies. The findings emphasize the importance of considering cultural nuances in analyzing social media data. Future research could explore the longitudinal emotional impact of the pandemic, analyze the impact of specific government policies on public sentiment, and develop more sophisticated sentiment analysis techniques to capture the complexities of human emotions. The development of more comprehensive and refined multilingual sentiment analysis benchmarks is also crucial for enhancing research in this area.

While the study analyzed a vast dataset, the findings may not be fully generalizable to all populations or regions due to the sampling method. The reliance on social media data also introduces potential biases, as social media usage is not uniformly distributed across different demographics. The use of machine translation for some languages might also have introduced some inaccuracies in sentiment analysis. Further research focusing on specific populations or using alternative data sources could help overcome some of these limitations.