1. Introduction
Disasters usually happen unexpectedly. Coronavirus disease 2019 (COVID-19), a respiratory disease caused by a novel coronavirus, has spread to over 180 locations internationally in nearly five months since its sudden outbreak in December 2019. According to the World Health Organization (
https://www.who.int/emergencies/diseases/novel-coronavirus-2019), as of 22 June 2020, there were 8,860,331 confirmed COVID-19 cases and more than 465,740 deaths worldwide. The effect of COVID-19 on personal health and economic development is self-evident, and it has had a wider impact on public sentiment. The earliest large-scale outbreak of COVID-19 was in China, and the country has now effectively controlled COVID-19. Examining the characteristics of the evolution of public emotion during the COVID-19 outbreak in China would be very useful in helping to understand and predict public emotion worldwide.
To this end, the study aim is to track the evolution of public emotions and analyze the root causes of such emotions from an event-driven perspective. The investigation focuses on the 48 day period following the announcement of the first laboratory-confirmed case of COVID-19 on 1 December 2019 [
1] and covers the early and middle stages of the COVID-19 epidemic in China. However, tracking and understanding public emotion evolution during COVID-19 is not an easy task. Three fundamental issues need to be addressed: (1) How can the public’s daily complex emotions be detected and generalized during the disaster? (2) How can the dynamic and constantly changing modes of the public’s complex emotions be captured? (3) What are the criteria for identifying hot events that dominate current public emotions? As COVID-19 evolved, social network sites became appropriate spaces for the social sharing of emotions, opinions, and coping strategies, providing psychological benefits for users by increasing their feelings of emotional relief [
2,
3,
4,
5]. Therefore, the study aim is to track public emotion during COVID-19 using the contents of microblogs and to extract the details of dominant events.
There have been several series of investigations of the information flow of social networks during disasters; these have focused on issues such as crowd psychological states [
6,
7,
8], disaster management strategies [
9,
10], detection of natural disasters [
11,
12], and public attention to disasters [
13,
14,
15]. Specific to public sentiment analysis, scholars have proposed crowd emotion detection solutions for natural disasters (e.g., earthquakes [
16], forest fires, floods, and droughts [
17]) based on microblogging contents. If we view the aforementioned research as a pointwise exploration of public emotion at a particular moment caused by short-lived disasters, this study extends the exploration from point to line and adopts a period-wise approach for this longer term disaster situation. The period-wise perspective and long-lasting nature of COVID-19 enable more in-depth research on the dynamic latent connections among complex public emotions, as well as the root causes of public emotions during different phases of the disaster. Embedding public emotion evolution into the skeleton of COVID-19 also made it possible to strip out periodic changes from the overall epidemic process. From this perspective, research on long-term disasters is worthy of attention and is meaningful.
Another difference from previous work is that, despite the special long-term context of COVID-19, this paper considers public emotions to be driven by events. For example, the rapidly rising number of confirmed COVID-19 cases may trigger public panic, whereas effective government control measures may reduce panic. Behaviors that endanger social security may provoke public anger, whereas positive social donations may spark the emotion of pride. Finding the root causes of public emotion evolution would be helpful for understanding and managing the disaster.
The rest of the paper is organized as follows.
Section 2 provides a summary of related work. The data collection process and methods of modeling public emotion evolution and extracting hot events are introduced in
Section 3. The experimental results are shown in
Section 4. The implications and future work are discussed in
Section 5.
5. Discussion
The aim of this research was to use social networks to track the public’s emotions during the spread of COVID-19 over a 48 day period since the first laboratory-confirmed case was announced on 1 December 2019, in China, and to analyze hot spot events underlying the emotions. The results show that, at different stages of COVID-19, public emotions showed interesting migration trajectories, as evidenced by both the covariance and transition modes. There was also a clear difference in the focus and reactions of common microblog users and official accounts in the face of the epidemic.
Some studies on public reactions to crises have focused on negative emotions (e.g., anger, anxiousness, fear, and sadness [
37,
39]) and have considered the threat of negative emotions to institutional trust [
44], reputation [
45], and the contribution to responsibility [
45]. However, our findings show that public positive emotions (e.g.,
love, respect, praise) co-existed with and even surpassed negative emotions during COVID-19, which is consistent with previous research on other crisis situations [
45,
46,
47].
Love appeared to have the most significant accumulated value during the whole period of 21 January 2020 to 17 February 2020 and increased around several time points (26 January 2020; 9 February 2020; and 14 February 2020). We further checked top-ranked hot events (
Table A1 and
Table A2) and found that social events about “Fighting COVID-19” became hot during this period, as well as good news about effective treatment (e.g., “Cumulative total of 5911 cases discharged from hospital”) and concerns about new cases (e.g., “Total number of cases successfully admitted to Vulcan Mountain Hospital exceeds 1000”). The sub-emotions
Respect (PD) and
Praise (PH) also showed high accumulated values and increased several times, especially from 14 February 2020 to 16 February 2020. The results confirmed that prevalent positive emotions can enhance the public’s coping function during disasters [
48,
49] and increase trust in government [
36,
50].
According to [
51,
52], public reactions to crises comprise four main coping methods: rational thinking (cognitive), emotional venting (emotional), instrumental support, and action (conative). At the very beginning phase of COVID-19 (before 21 January 2020), while instrumental support and effective actions from the government were lacking owing to the suddenness of the epidemic, the public’s feedback on COVID-19 predated official accounts on the microblog platform and focused on “isolation” and “novel coronavirus”. The microblog platform was used to share timely information and to create a sense of connectedness and functioned as a “psychological first aid” for the public [
53]. Then, in the first phase (21 January 2020 to 4 February 2020), the emotions of (
sadness, disgust) showed the most significant covariance. The sub-emotions (
doubt, reprimand) and (
reprimand, wishing) also showed significant covariance at this time. The public mainly focused on “Isolation” and “Rapid development of epidemic”. The public’s rational thinking and emotional venting were also captured, as topics about “Support Wuhan” became hot among both common and large users. During the middle phase (4 February 2020 to 12 February 2020),
surprise and
joy showed a new main covariance. The emergence of this covariance can be attributed to several causes. For example, with the decrease in daily reported new confirmed cases, topics about effective “Treatment” and “Fighting COVID-19” (e.g., “Original relay to fight COVID-19”) became hot in this phase (
Figure 4). Topics about “Epidemic prevention and control” became hot among large users. Topics about instrumental support (providing information and social support [
51]) from governments, non-governmental organizations, and medical staff across the country were extremely hot, particularly helped to boost public crisis coping and positive emotions, and also reflected the community disaster resilience of the government [
54].
It is notable that
surprise was also covariant with
anger during the middle phase (4 February 2020 to 12 February 2020). This may have been triggered by the fact that top-ranked events about “Doctor Wenliang Li” appeared among common users (see
Table A1 and
Table A2). Doctor Wenliang Li was the first person to issue a protective warning about COVID-19 and was called the “whistleblower” of the epidemic, but was unfortunately infected. In the last phase (12 February 2020 to 17 February 2020), a strong negative correlation appeared between
sadness and
surprise. Discussions about “Decrease in newly confirmed cases” became hot in this phase. In addition, after checking all event ranking results in this phase, we found that more than 20 super topics were proposed about the death of Doctor Wenliang Li. Research shows that community members tend to pay attention to victims and their symbolic behavior after a tragedy, and such social solidarity can increase the collective response to disasters [
55]. Communal bereavement can also increase social solidarity and lead to collective action, such as memorial rituals [
56]. This could explain why the public expressed such strong
sadness for Wenliang Li, as well as
anger and
disgust at his tragic death.
Considering the emotion stability,
love persisted as the most stable emotion during the whole COVID-19 period.
Disgust was the most stable emotion during the initial phase, and
anger was the most stable emotion during the second phase. Considering the emotion transition patterns, transitions between the emotions (
anger →
surprise) and the sub-emotions (
jealousy→
reassurance) frequently occurred. Throughout the observation window, the topics “Novel coronavirus” and “Confirmed cases” remained hot, but the number of discussions about these two categories gradually reduced over time, reflecting the process of gradual public adaptation to sudden disasters. This reflects one of the psychological functions of social media usage in disasters, which is to increase the public’s feeling of emotional relief [
2,
3,
4,
5]. This finding also indicates that disaster resilience [
57,
58] and social solidarity [
59,
60] during COVID-19 in China occurred at different levels (e.g., individual, institutional, and environmental) from the perspective of continuous semantic analysis.
Another issue that requires consideration is the role of government censorship of the Sina Weibo platform. According to the Sina Weibo Service Use Agreement (
https://www.weibo.com/signup/v5/protocol), in cooperation with Internet censorship in China, Sina Weibo imposes some restrictions on the content posted by users; for example, “Must not violate the laws and regulations of the People’s Republic of China and relevant international treaties or rules” (Regulation 4.10.1) and “Do not upload, display or disseminate any false, racially discriminatory, violent, bloody, or other illegal information materials.” (Regulation 4.10.4). In addition, news media and government agencies that open Weibo accounts must also abide by relevant laws, regulations, organizational rules, and regulatory requirements (Regulation 4.2). Such censorship ensures a safe and healthy operating environment for the platform and does not normally restrict users from publishing legal content. Our dataset and experimental results also show that during COVID-19, negative public emotions, such as doubt, jealously, and disgust, were expressed on the platform. Therefore, Sina Weibo can be regarded as a social platform that can objectively reflect public emotions during crises.
This was an exploratory study and therefore had some limitations. First, we did not differentiate between sarcasm and true negative emotions in the microblogging contents. Sarcasm is a special kind of sentiment that uses expressions that mean the opposite of what the person really wants to say [
61]. In recent years, the automatic detection of sarcasm in social networks has been considered an interesting research topic in the field of information retrieval and natural language processing. However, it is also recognized as a difficult task [
62], as it requires a system that has the necessary knowledge to interpret the linguistic styles of authors [
63]. Researchers have mainly used machine-learning methods [
64,
65,
66] and lexicon-based methods for sarcasm detection. For example, Reference [
63] attempted to detect sarcasm in microblogs by considering different sets of features (function words and parts of speech n-grams) and tested a range of different feature sets using fuzzy clustering and naive Bayesian models. Reference [
61] proposed two approaches to detect sarcasm in the text of Twitter data: a parsing-based lexicon generation algorithm and the occurrence of interjection words. Reference [
62] used a hybrid approach by combining machine-learning and lexicon-based methods. Specific to our problem, the addition of sarcasm detection in future studies could help to improve the accuracy of public emotion detection.
Second, regarding the Sina Weibo platform used in this study, an important issue is how to differentiate between real users who retweeted the microblogs and commented, liked, or read the sites from paid bots. At present, both the platform and researchers have proposed a series of solutions to recognize bots. The Sina Weibo platform has implemented complex technical rules to determine whether user account behavior constitutes an automated behavior; for example, frequent repeated postings, a large number of zombie fans, forwarding, but no original microblogs, no interaction with others, and never commenting. Such accounts are assumed to be “bots” (
https://new.qq.com/omn/20181113/20181113A0TTA1.html?pc,
https://www.weibo.com/signup/v5/protocol). In addition, researchers have conducted much in-depth research on the detection of bots in microblogs and tweets, applying multi-feature-based recognition methods and spam-filtering techniques [
67,
68,
69,
70]. The present study did not include the identification of bot accounts as a research aim. The elimination of such fake accounts in future work would help to improve data quality and to obtain more accurate research results.
Third, this study explored general public emotion evolution during COVID-19 in China. In practice, the situation may differ considerably in different regions of the country in terms of, for example, prevention and control policies, time of the outbreak, and public feedback, especially for severely affected cities like Wuhan. Owing to differences in the effect of the epidemic, the emotional states of Wuhan citizens may be different from those in other regions of the country. Therefore, more detailed information (e.g., positioning information published with microblogs, regional information in profiles) needs to be collected to carry out in-depth research on public emotion evolution in different regions. COVID-19 was continuing to evolve when this paper was last modified in August 2020. More areas of the world are facing serious situations because of COVID-19, and the psychological states of the public around the world are attracting more attention. Future work also needs to consider various factors like culture differences and government management.