Abstract
Aim
The accessibility of social media data has allowed researchers to measure official–public interactions during COVID-19. However, previous work analyzing official posts or public comments has failed to explore the link between the two. Therefore, this study investigates the relationship between the communication strategies of public health agencies (PHAs) on TikTok and public emotional/sentiment tendencies in COVID-19 normalization.
Subject and methods
This study uses the 2022 Shanghai city closure event as a public health communication case study in the context of COVID-19 normalization, using TikTok as a data source. We first analyze the communication strategies adopted by the PHA based on the Crisis and Emergency Risk Communication (CERC) model. Then, we classify the sentiment of public comments using the Large-Scale Knowledge Enhanced Pre-Training for Language Understanding and Generation (ERNIE) pre-training model. Finally, we explore the connection between PHA communication strategies and public sentiment tendencies.
Results
First, the public’s sentiment tendencies differ at different stages. Therefore, appropriate communication strategies should be developed stage-by-stage. Second, the public’s emotional disposition to different communication strategies varies: government statements, vaccines, and prevention and control programs are more likely to produce a friendly comment environment, while policy and new cases per day are more likely to produce unfavorable comment content. However, this does not mean that policy and new cases per day should be avoided; the judicious use of these two strategies can help PHAs understand the current issues causing public dissatisfaction. Third, videos with celebrity appearances can significantly increase positive public sentiment and, thereby, public participation.
Conclusion
We propose an improved CERC guideline for China based on the Shanghai lockdown case.
Keywords: TikTok, Deep learning, ERNIE, Crisis and Emergency Risk Communication, Communication strategy, Sentiment classification
Introduction
Social media characteristics (such as immediacy and pulsation effect) make it an essential source of solid and reliable real-time information during emergencies (Abdulhamid et al. 2021). For example, social media has a huge advantage in hurricane disasters in promoting support when people need to be rescued (Li et al. 2019). When official agencies use social media wisely and effectively, disaster victims may be saved (Gu et al. 2022). Likewise, in flood disasters, social media is a tool for disseminating information and an essential resource for mining and reviewing informal communication (Bird et al. 2012). Finally, during terrorist attacks, the enormous potential of social media to handle specific situations has allowed the agencies to analyze posts to detect crises and quickly develop evacuation plans (Şahin et al. 2019). In summary, using social media in emergencies has shown good applicability in many fields (Galea et al. 2023).
During COVID-19, the government and public health agencies (PHAs) used social media to publish government positions and statements, answer questions about science and policies, and disseminate information (Stjernswärd et al. 2021). The public used social media to obtain the emergency information they needed, especially under the impact of city lockdown and other anti-epidemic measures. Social media has become one of the most critical communication channels between officials and the public (Che et al. 2022a, b).
Social media refers to a content production and exchange platform on the Internet based on user relationships. It is also a tool and a platform that people use to share opinions, insights, experiences, and perspectives with each other (Aichner et al. 2021). Currently, social media includes blogs, discussion forums, wikis, video-sharing sites, photo-sharing sites, social bookmarking, and social networking sites (Chugh and Ruhi 2020).
The easy accessibility of social media data allows researchers to measure official and public interaction behavior. Researchers have explored different dimensions based on different data formats and research perspectives, with the most attention paid to analyzing official posts and the content of public comments. For example, official posts can be mined for topic distribution, and the carriers of these posts can be text (Panagiotopoulos et al. 2016), images (Malik et al. 2021), or videos (Li et al. 2021a, b, c). The use of public comment content, conversely, allows for the exploration of public opinion, which can be divided into “supervised” exploration, in the sense of verifying whether the comment content contains a specific meaning construction or efficacy (Vos and Buckner 2016), and “unsupervised” exploration of the latent thematic composition of the public discussion (Tang et al. 2018).
However, these previous works have one thing in common: the failure to analyze the relationship between official posts and public comments. That is, analyzing official posts alone may enable us to understand the type of official communication used, but not to measure whether the official communication strategy is correct or appropriate; analyzing only public comments may tell us what the public is discussing, but not how public opinion is affected by the official communication strategy. Therefore, it is necessary to analyze official posts and public comments together.
TikTok
First, with the recent normalization of COVID-19, the government, PHA, and public perceptions of the outbreak may have changed dramatically, so we chose the 2022 Shanghai closure event as our study case. In addition, previous studies have focused on text-based Twitter (Li et al. 2021a, b, c) and Facebook (Landi et al. 2021) posts, as well as image-based Instagram (Malik et al. 2021) posts. However, video-based TikTok—the most popular app in the world—is still outside the researcher’s attention (Basch et al. 2021). As an emerging social media platform, TikTok may be underestimated as a means of communicating with the public during COVID-19. In addition, TikTok has a large user base that other social media platforms cannot reach, and it has a very high demographic reach among seniors, Generation Z, and millennials. Hence, TikTok was chosen as the data source (Li et al. 2021a, b, c).
Crisis and Emergency Risk Communication
Second, we used the Crisis and Emergency Risk Communication (CERC) model to determine the official communication strategy (as shown in Fig. 1) and to provide each post with the appropriate strategic positioning (Alhassan and AlDossary 2021; Lwin et al. 2018). The CERC solves the conceptual confusion between risk and crisis communication by perfectly merging the two into one theoretical framework and dividing the emergency cycle into five stages: pre-crisis, initial event, maintenance, resolution, and evaluation. Thus, it creatively proposes a model that uses different communication strategies at different stages (Reynolds and Seeger 2005).
Fig. 1.
CERC communication strategy guide (https://emergency.cdc.gov/cerc/ppt/cerc_2014edition_Copy.pdf).
Pre-crisis
Communication is mainly accomplished through publicity and education, providing analytical information and tips to help the public establish awareness of risks, prevention, and resolution (Che et al. 2023).
Initial event
Build relationships and communicate with the public to reduce uncertainty, increase efficacy, and other actions.
Maintenance
Continue to communicate with the affected public, explain possible external environmental factors, obtain feedback from the affected public, communicate continuously, and further reduce uncertainties.
Resolution
Provide information to guide the public in disposal, rescue, recovery, and reconstruction efforts.
Evaluation
Mainly carry out assessment and summary for risk assessment, risk communication, response, and other processes, to summarize and improve the ability of risk and crisis communication (Che et al. 2022a, b).
Pre-training model: ERNIE
Third, we used the Large-Scale Knowledge Enhanced Pre-Training for Language Understanding and Generation (ERNIE) pre-training model to perform sentiment analysis on the public comments corresponding to the posts. ERNIE is a multi-paradigm unified pre-training model based on the knowledge enhancement proposed by Baidu in 2019 (Sun et al. 2020). ERNIE optimizes the previously developed Bidirectional Encoder Representations from Transformers (BERT) model from the knowledge enhancement perspective. In contrast to the BERT model, the ERNIE model learns real-world semantic relationships by modeling a priori semantic knowledge, such as entity concepts in massive data. In addition, the difference between ERNIE and BERT is that ERNIE is not applied to a small number of pre-training tasks, but continuously introduces new pre-training tasks to help the model continuously and efficiently learn semantic information; that is, it uses continual pre-training (Sun et al. 2019). As of August 2022, ERNIE outperformed BERT on a Chinese natural language processing (NLP) task (Di et al. 2022; Li et al. 2021a, b, c).
Introduction of ERNIE
The powerful application of neural networks in artificial intelligence indicates an ongoing revolution led by deep learning (Littman et al. 2022). For example, researchers and practitioners have developed a paradigm for solving problems by building various neural networks in NLP tasks (Nawaz et al. 2022). However, the cost of using deep learning is infinitely higher because the solution of the task involves a high degree of customization of the neural network (Cao et al. 2022).
As a means of addressing this problem, the emergence of the “pre-training fine-tuning” paradigm has dramatically reduced the cost of using deep learning by increasing the generalizability of the language representation model (You et al. 2021). This refers to a model with a high degree of generalizability and representational power. In “pre-training fine-tuning,” the language representation model is first pre-trained with a large corpus and a specific task to help it absorb the massive amount of semantic knowledge. Second, it is fine-tuned by overlaying the output layer to achieve generalization capabilities for the application scenario (Peng et al. 2022). For example, BERT is one of the most representative models for applying language representation (Su and Vijay-Shanker 2022).
However, although BERT represents the most cutting-edge and popular method for user sentiment classification (Gao et al. 2019), its results on Chinese tasks have been unsatisfactory. ERNIE, a knowledge-enhanced multi-paradigm unified pre-training model released by Baidu, has been shown to outperform BERT in most Chinese tasks due to its advantage regarding Chinese corpus richness (Zhang and Shang 2022).
Technical principle
Figure 2 shows the basic framework of ERNIE. Traditional large-scale pre-training models are mainly pre-training based on plain text data, ignoring common sense knowledge or world knowledge. In addition, most of the current large-scale pre-training language models are based on autoregressive networks and show strong capabilities on zero-shot/few-shot tasks. However, they need to handle traditional fine-tuning tasks more effectively.
Fig. 2.
The working framework of the ERNIE model (Sun et al. 2021).
To solve the above problems, ERNIE proposes a multi-paradigm unified large-scale pre-training framework. Based on this framework, ERNIE incorporates autoregressive networks and self-coding networks. Concurrently, the introduction of large-scale knowledge graph-like data enables ERNIE to perform extremely well on comprehension, generation, zero-sample learning, and common-sense inference tasks.
Advantages of ERNIE
A new continuous multi-paradigm unified pre-training framework created by ERNIE has the following benefits (Li et al. 2022):
First, regarding knowledge fusion and masked language modeling tasks (Sun et al. 2021), ERNIE utilizes a phrase- and entity-level mask approach to fusing external knowledge. Second, regarding the rich Chinese corpus (Sun et al. 2021), ERNIE has added Chinese corpora, such as the Baidu Encyclopedia and Baidu News, to enhance its effectiveness on Chinese tasks. Finally, regarding dialog embedding (Hsieh and Zeng 2022), ERNIE’s training corpus introduces knowledge from multiple data sources, such as news information and web forum conversation data. This learning for conversation data is an essential method of semantic representation.
This study uses the 2022 Shanghai city closure event as a public health communication case study in COVID-19 normalization, with TikTok as a data source. We first analyze the communication strategies adopted by the PHA based on the CERC model, then classify the sentiment of public comments using the ERNIE pre-training model. Finally, we explore the connection between PHA communication strategies and public sentiment tendencies. This study comprises six parts: introduction, methodology, results, discussion, conclusion, and limitations.
Methods
Data acquisition
In this study, we chose Healthy China, the official TikTok account of the National Health Commission of the People’s Republic of China (NHCC), as the data source. We crawled the 503 videos posted by Healthy China over the timeframe of the 2022 Shanghai lockdown, identifying January 1, 2022, as the starting point and June 30, 2022, as the ending point. Finally, the first and second authors retained data for 207 video posts related to COVID-19 through self-screening and comparing these results (the inclusion criteria are shown in Table 1), including 207 video titles, 207 posting times, and 5157 comments.
Table 1.
Video inclusion criteria and examples
| Inclusion criteria | Video title or content related to COVID-19 | Evaluation | Basis |
|---|---|---|---|
| Example | How to weigh the benefits and risks of getting a COVID-19 vaccine? | √ | COVID-19 appears in the title. |
| Be scientific, precise, and persistent, and realize social zeroing immediately. | √ | Although there is no COVID-19 in the title, the content relates to COVID-19. | |
| How has the development of traditional Chinese medicine in my country been in the past 10 years? | × | Title has nothing to do with COVID-19. | |
| In the past few decades, what changes have taken place in the medical and health care in Tibetan areas? | × | Although the title is related to medicine, it has nothing to do with COVID-19. |
Open coding
Coders
To ensure high inter-coder reliability, we recruited PhD students in data science and medicine to complete the coding. These two coders have qualified data processing and public health expertise, thus contributing to more precise and efficient coding (Zhang and Yu 2022).
Open coding process
First, we needed to train two coders. To this end, we created a basic codebook based on the CERC communication strategy guide (Fig. 1) and let the two coders become familiar with each category’s meaning. Second, we randomly selected 30% of the data from the 207 pieces for reliability testing. The two coders were assigned to test-code the sample data according to the codebook without communicating. Thereafter, we evaluated the coding effect by calculating the inter-coder Holsti reliability coefficient: When the reliability coefficient is greater than 0.9, the consistency of the two coders is high, and they can cooperate to complete the remaining 70% of the workload. Contrarily, when the reliability coefficient is less than 0.9, the degree of divergence is high (Liu et al. 2022). Finally, posts with different encoding were discussed and agreed upon, and additional categories had to be added if necessary. Table 2 lists the final codebook and classification results for the posts.
Table 2.
Codebooks formed after open coding based on the CERC model
| Strategy | Sub-strategy | Operational definition | Inter-coder reliability | Counts |
|---|---|---|---|---|
| Reassurance | Government statement | Government spokespeople convey the central government’s ideas to the public through press conferences | 1.00 | 22 (11%) |
| Uncertainty reduction | New cases per day | Number of new confirmed cases per day | 1.00 | 79 (38%) |
| Epidemic situation | Answers involving the status of the epidemic | 1.00 | 24 (12%) | |
| Policy | Answers on anti-epidemic policy | 1.00 | 20 (10%) | |
| Vaccines | Answering concerns about vaccines | 1.00 | 7 (3%) | |
| Nucleic acid testing | Answers to questions about nucleic acid testing | 1.00 | 6 (3%) | |
| Updates regarding resolution | Prevention and control programs | Answers to questions about the new prevention and control program | 1.00 | 10 (5%) |
| Preparations | Students | Information related to students | 1.00 | 13 (6%) |
| Transportation | Travel-related preparations | 1.00 | 4 (2%) | |
| Vaccines development | Information related to vaccine preparation | 1.00 | 3 (1%) | |
| Input risk | Information related to the control of imported cases from overseas | 1.00 | 3 (1%) | |
| Medical needs of the general population | Medical information related to the general public | 1.00 | 2 (1%) | |
| Living goods | About the security of living materials | 1.00 | 2 (1%) | |
| Self-efficacy | Personal protection | Information related to personal protection | 1.00 | 8 (4%) |
| Consensus about lessons and new understanding of risks | Negative case | Wrongdoings of local governments in implementing the anti-epidemic process | 1.00 | 4 (2%) |
Crisis stage
As shown in Fig. 3, we divided the 207 videos into five parts based on the CERC model: pre-crisis, initial event, maintenance, resolution, and evaluation.
Fig. 3.
Timeline of 2022 Shanghai closure events based on the CERC model
Pre-training model: ERNIE
In the first step, we randomly sampled 20% of the comment data for manual annotation by the two previously recruited PhD students. The sentiment tendency of the data was divided into positive and negative, with positive values marked as 1 and negative values marked as −1. After repeated tests, Holsti’s coefficient finally exceeded 0.9, indicating that the annotation results were highly reliable. In the second step, we trained the model on a labeled dataset that contained 481 positive and 481 negative comments. Finally, we used the trained model to analyze the sentiment tendency of the remaining 80% of the data.
Results
ERNIE
Table 3 shows the overall evaluation effect of the model: accuracy of 88.6%, precision of 88.8%, recall of 88.6%, and F1-score of 88.6%.
Table 3.
Integrated assessment results of the ERNIE model
| Accuracy | Precision | Recall | F1-score |
|---|---|---|---|
| 88.6% | 88.8% | 88.6% | 88.6% |
Table 4 lists the performance of the model on different datasets. Positive: accuracy of 91.9%, recall of 85.0%, F1-score of 88.3%. Negative: accuracy of 85.6%, recall of 92.3%, F1-score of 88.8%.
Table 4.
Evaluation results for the ERNIE model on different datasets
| Accuracy | Recall | F1-score | |
|---|---|---|---|
| Positive | 91.9% | 85.0% | 88.3% |
| Negative | 85.6% | 92.3% | 88.8% |
Sentiment distribution of public comments in different stages
Figure 4 and Table 5 show the distribution of the sentiment tendencies of comments across stages. Regarding the comment counts, pre-crisis had the highest number of comments, and resolution had the lowest number. At the same time, the sentiment tendencies at different stages also showed differentiation. In the pre-crisis period, the proportions of positive (52%) and negative (48%) were almost identical. In both the initial event and resolution, positive (40%, 42%) and negative (60%, 58%) results showed a 4:6 ratio. Most comments related to maintenance were positive (85%). In contrast, negative (85%) dominated evaluation.
Fig. 4.
Sentiment polarity distribution of public comments in different stages
Table 5.
Sentiment polarity distribution of public comments in different stages
| Stage | Positive | Negative | Total |
|---|---|---|---|
| Pre-crisis | 907 (52%) | 853 (48%) | 1760 |
| Initial event | 222 (40%) | 332 (60%) | 554 |
| Maintenance | 1262 (85%) | 219 (15%) | 1481 |
| Resolution | 117 (42%) | 164 (58%) | 281 |
| Evaluation | 161 (15%) | 920 (85%) | 1081 |
| Total | 2669 (52%) | 2448 (48%) | 5157 |
Sentiment distribution of public comments with different communication strategies
Tables 6, 7, 8, 9, and 10 present the sentiment distribution of public comments for various communication strategies at different stages. In this study, a low number of comments will result in considerable fluctuation in sentiment percentages; therefore, we only analyzed communication strategies with over 100 comments.
Table 6.
Distribution of sentiment polarity of public comments in the pre-crisis
| Strategy | Sub-strategy | Positive | Negative | Total |
|---|---|---|---|---|
| Uncertainty reduction | New cases per day | 162 (36%) | 290 (64%) | 452 |
| Epidemic situation | 174 (50%) | 173 (50%) | 347 | |
| Policy | 34 (33%) | 70 (67%) | 104 | |
| Vaccines | 68 (67%) | 34 (33%) | 102 | |
| Reassurance | Government statement | 277 (73%) | 103 (27%) | 380 |
| Updates regarding resolution | Prevention and control programs | 142 (61%) | 92 (39%) | 234 |
| Self-efficacy | Personal protection | 11 (65%) | 6 (35%) | 17 |
| Preparations | Students | 6 (9%) | 60 (91%) | 66 |
| Vaccine development | 25 (74%) | 9 (26%) | 34 | |
| Medical needs of the general population | 8 (33%) | 16 (67%) | 24 |
Table 7.
Distribution of sentiment polarity of public comments in the initial event
| Strategy | Sub-strategy | Positive | Negative | Total |
|---|---|---|---|---|
| Uncertainty reduction | Policy | 49 (34%) | 97 (66%) | 146 |
| New cases per day | 47 (48%) | 51 (52%) | 98 | |
| Epidemic situation | 31 (39%) | 48 (61%) | 79 | |
| Vaccines | 19 (31%) | 42 (69%) | 61 | |
| Preparations | Living goods | 25 (38%) | 40 (62%) | 65 |
| Vaccine development | 21 (38%) | 35 (63%) | 56 | |
| Reassurance | Government statement | 27 (61%) | 17 (39%) | 44 |
| Self-efficacy | Personal protection | 3 (60%) | 2 (40%) | 5 |
Table 8.
Distribution of sentiment polarity of public comments in the maintenance
| Strategy | Sub-strategy | Positive | Negative | Total |
|---|---|---|---|---|
| Self-efficacy | Personal protection | 822 (99%) | 11 (1%) | 833 |
| Reassurance | Government statement | 301 (90%) | 32 (10%) | 333 |
| Uncertainty reduction | New cases per day | 64 (43%) | 86 (57%) | 150 |
| Policy | 7 (17%) | 34 (83%) | 41 | |
| Epidemic situation | 13 (62%) | 8 (38%) | 21 | |
| Nucleic acid testing | 1 (50%) | 1 (50%) | 2 | |
| Vaccines | 0 (0%) | 1 (100%) | 1 | |
| Preparations | Transportation | 27 (43%) | 36 (57%) | 63 |
| Medical needs of the general population | 16 (67%) | 8 (33%) | 24 | |
| Students | 11 (85%) | 2 (15%) | 13 |
Table 9.
Distribution of sentiment polarity of public comments in the resolution
| Strategy | Sub-strategy | Positive | Negative | Total |
|---|---|---|---|---|
| Uncertainty reduction | New cases per day | 75 (42%) | 103 (58%) | 178 |
| Vaccines | 3 (23%) | 10 (77%) | 13 | |
| Epidemic situation | 6 (60%) | 4 (40%) | 10 | |
| Policy | 0 (0%) | 4 (100%) | 4 | |
| Reassurance | Government statement | 29 (43%) | 39 (57%) | 68 |
| Preparations | Input risk | 4 (50%) | 4 (50%) | 8 |
Table 10.
Distribution of sentiment polarity of public comments in the evaluation
| Strategy | Sub-strategy | Positive | Negative | Total |
|---|---|---|---|---|
| Uncertainty reduction | New cases per day | 68 (20%) | 274 (80%) | 342 |
| Nucleic acid testing | 13 (18%) | 59 (82%) | 72 | |
| Policy | 11 (17%) | 53 (83%) | 64 | |
| Epidemic situation | 4 (40%) | 6 (60%) | 10 | |
| Updates regarding resolution | Prevention and control programs | 15 (7%) | 196 (93%) | 211 |
| Consensus about lessons and new understanding of risks | Negative case | 15 (9%) | 148 (91%) | 163 |
| Reassurance | Government statement | 13 (12%) | 99 (88%) | 112 |
| Preparations | Students | 12 (15%) | 66 (85%) | 78 |
| Self-efficacy | Personal protection | 10 (34%) | 19 (66%) | 29 |
Table 6 shows that videos related to government statements (73%), vaccines (67%), and prevention and control programs (61%) generated more positive comments, whereas policies (67%) and new cases per day (64%) elicited more unfavorable comments.
As shown in Table 7, videos related to the policy elicited 66% negative comments and 34% positive comments.
Table 8 shows that videos related to personal protection (99%) and government statements (90%) generated almost all positive comments, whereas new cases per day generated 57% negative comments and 43% positive comments.
Table 9 shows that videos associated with new cases per day generated 58% negative and 42% positive comments.
Table 10 shows that videos related to new cases per day (80%), prevention and control programs (93%), negative cases (91%), and government statements (88%) generated more negative comments.
Discussion
Theoretical significance
Our results suggest that the public’s sentimental disposition differs at different stages. This finding raises a question about most of the previous work on public perceptions of COVID-19: studying attitudes, such as public sentiment based on social media data over a certain period, without distinguishing subperiods within that period may be too crude (Chen et al. 2021; Guidry et al. 2017). For example, in pre-crisis, the positive and negative responses are almost equal. In maintenance, the positive is dominant, while in evaluation, negativity dominates overwhelmingly. This discrepancy gives rise to a new requirement for future research: analyzing public opinion based on different stages and then adjusting the PHA’s management and communication strategies accordingly. Since public sentiment may be generally positive or negative, we may overlook the specificity of a particular stage. Using appropriate communication strategies based on a particular stage can make crisis management more effective.
For the first time, we found that the public on TikTok generated different sentiment tendencies regarding different communication strategies. In pre-crisis, although the proportions of positive (52%) and negative (48%) sentiments were almost the same, the sentiment analysis at the communication strategy level revealed that the public did not produce the same proportion of opposite sentiments as pre-crisis, but somewhat different sentiment tendencies about various communication strategies. This finding extends Lwin et al.’s (2018) findings. Their results showed that messages on various topics received different levels of public response precisely because the public perceived different sentiments in posts with different communication strategies. Therefore, communication strategies impacted engagement. In this study, the public showed a clear tendency toward positive sentiments about government statements (73%), vaccines (67%), and prevention and control programs (61%), while negative emotions dominated responses to statements on policy (67%) and new cases per day (64%). This result has implications for PHA when posting outbreak-related information: using government statements, vaccines, and prevention and control programs in pre-crisis may produce an environment more likely to elicit positive comments, while using policy and new cases per day may be more likely to generate a hostile comment environment.
Practical significance
New cases per day showed a tendency toward public sentiments dominated by negative emotions at all stages. This result supports the findings by Lwin et al. (2018), who showed that the public becomes anxious about confirmed cases, making negative comments more likely. However, as of August 2022, no studies have explored the psychological impact of daily case releases on the public, especially in the context of COVID-19, where the public may need to anticipate the end of the outbreak based on daily case releases to influence their future travel plans. In China, the public’s right to travel is heavily influenced by the government’s classification of cities as low-, medium-, or high-risk based on the number of new cases per day in the area. Therefore, the public is more likely to react unfavorably when their city is classified as a medium- or high-risk area owing to increased new daily cases.
Policy posts also showed more negative public sentiment in all stages. In this regard, our findings extend the findings in a study by Naumann et al. (2020), whose work found very positive public evaluations of policy in the early stages of the epidemic, with the public generally supporting policies such as city closures; however, this positivity began to gradually decline after 2 weeks. In China, the public has experienced the impact of city closures or other policies on their own lives many times; therefore, in the context of the normalization of COVID-19, the public’s attitude toward the new wave of the epidemic has been negative from the beginning to the end of the policy.
In terms of maintenance, the videos with celebrity presence are all videos related to personal self-protection, and the public’s positive sentiment is close to 100%. Furthermore, we found that videos with celebrity presence significantly increased public engagement and positive emotions, because most of the public are fans of the celebrities in the videos. Fans are naturally more likely to express positive emotions when facing their favorite celebrities or idols.
In the evaluation, the public showed a predominantly negative emotional tendency toward all communication strategies, especially the videos about prevention and control programs, both of which had a negative sentiment of over 90%, reflecting the public’s dissatisfaction with the government’s problems in the implementation of prevention and control programs. In the process of the government and the public working together to combat the epidemic, the government and PHA are responsible for developing and updating programs, and the public is responsible for cooperating. However, these programs may reveal many flaws in the implementation process. The public expects these unreasonable measures to be removed and modified. Therefore, they are likelier to be distressed.
As shown in Table 11, we refined the CERC model based on the communication strategy used by the NHCC and the findings of this study to adapt it to the needs of China’s resistance to the pandemic in the context of COVID-19 normalization.
Table 11.
Improved CERC model
| Stage | Communication objectives and strategies |
|---|---|
| Pre-crisis | Uncertainty reduction; Reassurance; Updates regarding resolution |
| Notification of new cases per day* | |
| Keeping the public informed of the current epidemic situation | |
| Explaining the current anti-epidemic policy to the public* | |
| Timely resolution of public doubts about vaccines | |
| Regularly articulate official positions to the community | |
| Explain to the public the updated details of the new version of the prevention and control program | |
| Promotion of personal protective measures | |
| Initial event | Uncertainty reduction |
| Explaining the current anti-epidemic policy to the public* | |
| Maintenance | Self-efficacy; Reassurance; Uncertainty reduction |
| Promotion of personal protective measures | |
| Regularly articulate official positions to the community | |
| Notification of new cases per day* | |
| Resolution | Uncertainty reduction |
| Notification of new cases per day* | |
| Evaluation | Uncertainty reduction; Updates regarding resolution; Consensus about lessons and new understanding of risks; Reassurance |
| Notification of new cases per day* | |
| Explain to the public the updated details of the new version of the prevention and control program* | |
| Periodic notification of some negative cases in the process of vaccination* | |
| Regularly articulate official positions to the community* |
*Indicates that the communication strategy may lead to negative public sentiment.
Communication strategies marked with an asterisk (*) in Table 11 are likely to generate negative public comments, but simply because they generate negative comments does not mean we should use them less frequently to create a “superficially friendly” public opinion environment. Indeed, such communication strategies may be more likely to help the government and PHAs identify the current issues with which the public is dissatisfied and then improve the relevant policies and prevention and control programs.
Conclusions
This study investigates the relationship between PHA communication strategies on TikTok and public sentiment tendencies in the context of COVID-19 normalization based on the CERC model, using the 2022 Shanghai closure event as a case study. Our findings and contributions are as follows.
Key findings
The public’s sentiment tendencies differ at different stages. Therefore, developing corresponding communication strategies stage-by-stage is required. Government statements, vaccines, and prevention and control programs are more likely to produce a friendly comment environment, while policies and new cases per day are more likely to produce hostile comment content. However, this does not mean that policy and new cases per day should be avoided; their judicious use can help PHAs understand what issues the public is currently dissatisfied with and make targeted improvements. Videos with celebrity appearances can significantly increase positive public sentiment, particularly at the right stage. Celebrities can greatly increase public engagement with videos for appropriate communication purposes. We propose an improved CERC guideline for China based on the Shanghai lockdown case.
In this study, we used sentiment analysis to find that different communication strategies have different affective effects on the public at different pandemic stages. However, whether officials should use or reduce communication strategies likelier to generate negative comments remains unclear. Further eliciting unfavorable comments may help officials identify shortcomings in current anti-epidemic policies. In contrast, less use of such strategies is more likely to create a “friendly” public opinion environment. Finding a balance between the two requires a deeper analysis of the subject matter of the comments. The best conclusions can only be found through a deeper understanding of what the public is discussing. Also, we must account for the impact of video views. For example, the sentiment classification results for videos with low video views fluctuate significantly, which may not reflect the real user emotional tendency.
Author contributions
Che ShaoPeng: Conceptualization, Methodology, Software, Writing—Original draft preparation, Editing. Kim Jang Hyun: Writing—Review, Supervision.
Funding
This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korean government (NRF-2020R1A2C1014957).
Data availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
Declarations
Ethics approval
Not applicable
Consent to participate
Not applicable
Consent to publish
Not applicable
Conflicts of interest/Competing interests
None declared
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- Abdulhamid NG, Ayoung DA, Kashefi A, Sigweni B. A survey of social media use in emergency situations: A literature review. Inform Dev. 2021;37(2):274–291. doi: 10.1177/0266666920913894. [DOI] [Google Scholar]
- Aichner T, Grünfelder M, Maurer O, Jegeni D. Twenty-five years of social media: a review of social media applications and definitions from 1994 to 2019. Cyberpsychol Behav Soc Networking. 2021;24(4):215–222. doi: 10.1089/cyber.2020.0134. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Alhassan FM, AlDossary SA. The Saudi Ministry of Health’s Twitter communication strategies and public engagement during the COVID-19 pandemic: content analysis study. JMIR Public Health Surveillance. 2021;7(7):e27942. doi: 10.2196/27942. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Basch CH, Mohlman J, Fera J, Tang H, Pellicane A, Basch CE. Community mitigation of COVID-19 and portrayal of testing on TikTok: descriptive study. JMIR Public Health Surveillance. 2021;7(6):e29528. doi: 10.2196/29528. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Bird D, Ling M, Haynes K. Flooding Facebook-the use of social media during the Queensland and Victorian floods. Australian J Emergency Manag. 2012;27(1):27–33. [Google Scholar]
- Cao MT, Nguyen NM, Wang WC (2022) Using an evolutionary heterogeneous ensemble of artificial neural network and multivariate adaptive regression splines to predict bearing capacity in axial piles. Eng Struct 268:114769. 10.1016/j.engstruct.2022.114769
- Che S, Nan D, Kamphuis P, Zhang S, Kim JH. Examining Crisis Communication Using Semantic Network and Sentiment Analysis: A Case Study on NetEase Games. Front Psychol. 2022;13:176. doi: 10.3389/fpsyg.2022.823415. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Che S, Zhang S, Kim JH (2022b) How public health agencies communicate with the public on TikTok under the normalization of COVID-19: A case of 2022 Shanghai's outbreak. Front Public Health 10. 10.3389/fpubh.2022.1039405 [DOI] [PMC free article] [PubMed]
- Che S, Wang X, Zhang S, Kim JH (2023) Effect of daily new cases of COVID-19 on public sentiment and concern: Deep learning-based sentiment classification and semantic network analysis. J Public Health 1-20. 10.1007/s10389-023-01833-4 [DOI] [PMC free article] [PubMed]
- Chen Q, Min C, Zhang W, Ma X, Evans R. Factors driving citizen engagement with government TikTok accounts during the COVID-19 pandemic: Model development and analysis. J Med Internet Res. 2021;23(2):e21463. doi: 10.2196/21463. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Chugh R and Ruhi U (2020) Social Media for Tertiary Education. 10.1007/978-3-319-60013-0_202-1
- Di J, Liu Z, Yang Y (2022) Text classification of COVID-19 reviews based on pre-training language model. In 2022 IEEE 2nd International Conference on Power, Electronics and Computer Applications (ICPECA). IEEE:1179–1183. 10.1109/ICPECA53709.2022.9719020
- Galea G, Chugh R, Luck J (2023) Why should we care about social media codes of conduct in healthcare organisations? A systematic literature revie. J Public Health: 1-13. 10.1007/s10389-023-01894-5 [DOI] [PMC free article] [PubMed]
- Gao Z, Feng A, Song X, Wu X (2019) Target-dependent sentiment classification with BERT. IEEE Access 7:154290–154299. 10.1109/ACCESS.2019.2946594
- Gu M, Guo H, Zhuang J, Du Y, Qian L. Social media user behavior and emotions during crisis events. Int J Environ Res Public Health. 2022;19(9):5197. doi: 10.3390/ijerph19095197. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Guidry JP, Jin Y, Orr CA, Messner M, Meganck S. Ebola on Instagram and Twitter: How health organizations address the health crisis in their social media engagement. Public Relations Rev. 2017;43(3):477–486. doi: 10.1016/j.pubrev.2017.04.009. [DOI] [Google Scholar]
- Hsieh YH, Zeng XP (2022) Sentiment analysis: An ERNIE-BiLSTM approach to bullet screen comments. Sensors 22(14):5223. 10.3390/s22145223 [DOI] [PMC free article] [PubMed]
- Landi S, Costantini A, Fasan M, Bonazzi M. Public engagement and dialogic accounting through social media during COVID-19 crisis: a missed opportunity? Account Audit Accountabil J. 2021;35(1):35–47. doi: 10.1108/AAAJ-08-2020-4884. [DOI] [Google Scholar]
- Li L, Aldosery A, Vitiugin F, Nathan N, Novillo-Ortiz D, Castillo C, Kostkova P (2021b) The response of governments and public health agencies to COVID-19 pandemics on social media: a multi-country analysis of twitter discourse. Front Public Health 1410. 10.3389/fpubh.2021.716333 [DOI] [PMC free article] [PubMed]
- Li Y, Guan M, Hammond P, Berrey LE. Communicating COVID-19 information on TikTok: a content analysis of TikTok videos from official accounts featured in the COVID-19 information hub. Health Educ Res. 2021;36(3):261–271. doi: 10.1093/her/cyab010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Li J, Stephens KK, Zhu Y, Murthy D. Using social media to call for help in Hurricane Harvey: Bonding emotion, culture, and community relationships. Int J Disaster Risk Reduct. 2019;38:101212. doi: 10.1016/j.ijdrr.2019.101212. [DOI] [Google Scholar]
- Li W, Sun R, Wu Y (2022) Exploiting word semantics to enrich character representations of Chinese pre-trained models. In: Natural Language processing and Chinese computing: 11th CCF International Conference, NLPCC 2022, Guilin, China, September 24–25, 2022, Proceedings, Part I. Springer International Publishing, Cham, pp 3–15
- Li J, Zhang D, Wulamu A (2021a) Chinese Text Classification Based on ERNIE-RNN. In 2021 2nd International Conference on Electronics, Communications and Information Technology (CECIT) (pp. 368-372). IEEE. 10.1109/CECIT53797.2021.00072
- Littman ML, Ajunwa I, Berger G et al (2022) Gathering strength, gathering storms: the one hundred year study on artificial intelligence (AI100) 2021 study panel report. arXiv preprint arXiv:2210.15767. 10.48550/arXiv.2210.15767
- Liu Z, Zhao YC, Song S, Ba Z, Zhu Q. Exploring the endorsement effect on scientific crowdfunding performance: Evidence from Experiment. com. Telemat Inform. 2022;73:101872. doi: 10.1016/j.tele.2022.101872. [DOI] [Google Scholar]
- Lwin MO, Lu J, Sheldenkar A, Schulz PJ. Strategic uses of Facebook in Zika outbreak communication: implications for the crisis and emergency risk communication model. Int J Environ Res Public Health. 2018;15(9):1974. doi: 10.3390/ijerph15091974. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Malik A, Khan ML, Quan-Haase A. Public health agencies outreach through Instagram during the COVID-19 pandemic: Crisis and Emergency Risk Communication perspective. Int J Disaster Risk Reduct. 2021;61:102346. doi: 10.1016/j.ijdrr.2021.102346. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Naumann E, Möhring K, Reifenscheid M, et al. COVID-19 policies in Germany and their social, political, and psychological consequences. Eur Policy Anal. 2020;6(2):191–202. doi: 10.1002/epa2.1091. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Nawaz A, Abbas Y, Ahmad T, Mahmoud NF, Rizwan A, Samee NA (2022) A healthcare paradigm for deriving knowledge using online consumers’ feedback. Healthcare 10(8):1592. 10.3390/healthcare10081592 [DOI] [PMC free article] [PubMed]
- Panagiotopoulos P, Barnett J, Bigdeli AZ, Sams S. Social media in emergency management: Twitter as a tool for communicating risks to the public. Technol Forecast Social Change. 2016;111:86–96. doi: 10.1016/j.techfore.2016.06.010. [DOI] [Google Scholar]
- Reynolds B, Seeger MW. Crisis and emergency risk communication as an integrative model. J Health Commun. 2005;10(1):43–55. doi: 10.1080/10810730590904571. [DOI] [PubMed] [Google Scholar]
- Peng Q, Pan Y et al (2022) ERNIE-layout: layout knowledge enhanced pre-training for visually-rich document understanding. arXiv preprint arXiv:2210.06155. 10.48550/arXiv.2210.06155
- Şahin C, Rokne J, Alhajj R (2019) Emergency detection and evacuation planning using social media. Social Networks Surveillance Soc:149-164. 10.1007/978-3-319-78256-0_9
- Stjernswärd S, Ivert AK, Glasdam S (2021) Perceptions and effects of COVID-19 related information in Denmark and Sweden–a web-based survey about COVID-19 and social media. J Public Health:1-15. 10.1007/s10389-021-01539-5 [DOI] [PMC free article] [PubMed]
- Su P, Vijay-Shanker K. Investigation of improving the pre-training and fine-tuning of BERT model for biomedical relation extraction. BMC bioinformatics. 2022;23(1):120. doi: 10.1186/s12859-022-04642-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Sun Y, Wang S, Li Y et al (2019) Ernie: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223. 10.48550/arXiv.1904.09223
- Sun Y, Wang S, Li Y, Feng S, Tian H, Wu H, Wang H (2020) Ernie 2.0: A continual pre-training framework for language understanding. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 05, pp. 8968-8975). 10.1609/aaai.v34i05.6428
- Sun Y, Wang S et al (2021) Ernie 3.0: large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137. 10.48550/arXiv.2107.02137
- Tang L, Bie B, Zhi D. Tweeting about measles during stages of an outbreak: A semantic network approach to the framing of an emerging infectious disease. Am J Infect Control. 2018;46(12):1375–1380. doi: 10.1016/j.ajic.2018.05.019. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Vos SC, Buckner MM. Social media messages in an emerging health crisis: tweeting bird flu. J Health Commun. 2016;21(3):301–308. doi: 10.1080/10810730.2015.1064495. [DOI] [PubMed] [Google Scholar]
- You K, Liu Y, Wang J, Long M (2021) Logme: practical assessment of pre-trained models for transfer learning. In: International Conference on Machine Learning. PMLR, p 12133 – 12143
- Zhang M, Shang X (2022) Chinese short text classification by ERNIE based on LTC_Block. Wirel Commun Mob Comput 2022. 10.1155/2022/1411744
- Zhang T, Yu L (2022) The Relationship between government information supply and public information demand in the early stage of COVID-19 in China—an empirical analysis. Healthcare 10(1):77. 10.3390/healthcare10010077 [DOI] [PMC free article] [PubMed]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.