Research on theme mining of online public opinion on public health emergencies

In the study on topic mining of online public opinion regarding public health emergencies, Farhadloo et al. conducted theme mining of Zika virus text data on Twitter. They compared and analyzed online topic discussions with the results of offline questionnaires, demonstrating the possibility of analyzing people’s views and attitudes towards public health emergencies using online public opinion data [2]. Wang et al. utilized the LDA model to mine public opinion topics on Weibo during the COVID-19 pandemic. They determined the level of public concern for each topic by counting the number of blog posts under each topic. Additionally, they analyzed user interaction behaviors across different themes by considering likes, retweets, comments, time, and region. This enabled them to study the changes in attention among microblog users in each region during the new coronavirus outbreak [3]. Wright et al. employed structured topic modeling to extract topics from over 4,000 free-text survey responses. The results indicated the critical importance of consistent and transparent communication and messaging from the government in improving compliance with measures to contain the virus, as well as protecting mental health during health emergencies [4]. Kwok et al. crawled tweet data related to COVID-19 vaccination on Twitter. They employed the LDA model to identify three types of online public opinion themes: public attitudes towards vaccination, public advocacy for epidemic prevention and control measures, and public misunderstandings and complaints about COVID-19 [5]. Zhao et al. collected online discussion data from the Zhihu platform during the COVID-19 pandemic. They used the LDA topic identification model to conduct an evolutionary analysis in terms of content and intensity. Their aim was to explore the topics of public opinion concern on the Zhihu platform during the epidemic and the changes in the degree of concern for each topic [6].

Studies related to sentiment analysis of online public opinion on public health emergencies

In terms of research on sentiment analysis of public opinion regarding public health emergencies, Hung et al. analyzed the high-frequency words in positive and negative text data on Twitter. They discovered that positive tweets and negative tweets underestimated the severity of COVID-19 and expressed dissatisfaction with the travel restrictions imposed due to the pandemic. By examining the distribution of sentiment in tweets across different states in the, they demonstrated that positive sentiments tended to be more prevalent in states lower infection rates [7]. Mittal et al. examined the correlation between the number of diagnoses, deaths, and cures during the COVID-19 and the sentiment of Twitter public opinion, and the results showed a significant positive correlation between the number of cures and the number of tweets with positive sentiment [8]. Mou et al. developed a method for sentiment analysis of online public opinion that integrated emoticon features. Based on this, they proposed an online public opinion trend prediction model that incorporated sentiment values. They conducted a simulation of online public opinion heat during a vaccine counterfeiting using the model. The results demonstrated that incorporating emotional factors into the prediction of online public opinion heat for public health emergencies significantly improved the prediction accuracy [9]. Zhuang et al. analyzed the sentiment evolution of netizens’ comments from Wuhan and other regions on the Weibo platform. They aimed to identify differences in sentiment evolution between these two groups [10]. Bian et al. analyzed the expression public emotions in social media during major public health emergencies from a global. They focused on Sina Weibo public opinion and investigated emotional evolution patterns, thematic perspectives, and geographical diffusion, revealing several patterns of emotional evolution [11].

Cross-platform internet public opinion related research

Yang et al. utilized scale-free networks to simulate the cross-platform social network environment. They analyzed the effects of individual factors, external friend environment, and external platform environment on user state changes based on the SEIR model. The aim was to explore the dissemination patterns of public opinion in cross-platform social networks. Additionally, they constructed a public opinion dissemination model based on cross-platform social networks and conducted simulation analysis on the model using MATLAB [12]. Zhang et al. proposed a method of public opinion sentiment analysis based on word frequency analysis and the LDA model. This method was applied to address public opinion issues that arise during the process of university epidemic control. They verified the method using cross-media public opinion examples related to university epidemic control. The objective was to provide a theoretical basis and technical support for the government to effectively manage public opinion [13]. An et al. used Sina Weibo and Douyin as data sources. They employed the ELECTRA and REDP methods for entity extraction and relationship extraction respectively on opinion text information. Subsequently, they constructed network opinion knowledge graphs for Weibo and Douyin separately. The knowledge graphs of each network were compared and analyzed [14]. Chen et al. developed a network opinion propagation model for negative corporate events based on the classical SIR propagation model. Their aim was to understand the propagation patterns of negative corporate events in the cross-platform “two microblogs and one end”. They constructed a cross-platform network environment for the “two microblogs and one end” based on a scale-free network and analyzed the main parameters. The results demonstrated that the constructed network public opinion dissemination model accurately described the process of network public opinion dissemination in a real environment [15].

The existing literature undoubtedly provides valuable insights and merits acknowledgment. However, this paper aims to complement and enhance the current understanding of online public opinion on public health emergencies in the following aspects: Firstly, regarding the research subject, there is a scarcity of scholars who have explored online public opinion from the perspective of stakeholders. This study addresses this gap by integrating stakeholders into the analysis of online public opinion during public health emergencies. It categorizes stakeholders and investigates the themes of concern and the emotional evolution of various stakeholders at each stage of public opinion evolution. Secondly, with respect to platforms, most current studies focus solely on analyzing the evolution of online public opinion on public health emergencies within a single platform. While some studies have examined the propagation patterns of cross-platform online public opinion events from a propagation perspective, these studies often rely on simulations rather than actual data analysis. In contrast, this study takes into account public opinion data from multiple platforms. It integrates and analyzes the collected data from various platforms to provide a comprehensive reference for the government and relevant management departments in understanding the evolving patterns and development trends of public opinion.

Life cycle division of online public opinion on public health emergencies

As early as 1986, STEVEN (1986) proposed a division of emergency spread into four stages in the field of crisis management: the incubation period, outbreak period, spreading period, and recovery period [16]. Currently, there are multiple bases and methods of division in academia, such as the three-stage theory, four-stage theory, five-stage theory, and six-stage theory of online public opinion. Based on the characteristics of public health emergencies in information dissemination and STEVEN’s classical life cycle theory, this study identifies the inflection point of information volume regarding public health emergencies. Additionally, it examines the landmark events at key time points to divide the life cycle of public health emergencies into five stages: the germination period, outbreak period, spreading period, decline period, and recovery period.

Germination period. In the germination period, although the cause of the event already exists, but due to the characteristics of the event is not obvious, only a small number of users pay attention to the event. The media and government departments do not pay much attention to the event. The amount of information related to the event is relatively small, and is in the state of slow increase.

Germination period: During this phase, although the cause of the event already exists, it remains relatively inconspicuous, with only a small number of users paying attention to it. Media and government departments do not allocate significant attention to the event. The volume of information related to the event is relatively small and increases slowly.

Outbreak period: As the event triggers continue to accumulate, the event’s influence and scope of impact expand rapidly. The event begins spreading across major social media platforms with significant momentum. During this stage, more users become interested in the incident, leading to a rapid increase in the amount of information related to the event until it reaches its peak.

Spreading period: In this stage, the incident continues to spread, and more users engage in discussions about it. However, relevant government interventions begin to control the situation, causing some uninvolved individuals to lose interest. As a result, the amount of information related to the event shows a decreasing trend. However, the management of public health emergencies tends to involve repetitive actions, leading to longer durations of the spreading period. Consequently, the amount of information about the event fluctuates and decreases.

Decline period: During the decline period, the intensity of the event starts diminishing, and social order gradually restores. Only a small portion of the population continues to pay attention to the event, with the public’s focus shifting to other matters. Although the amount of information about the event fluctuates, it generally decreases.

Recovery period: In the recovery period, the incident has essentially come to an end, and the number of individuals participating in discussions about the event further diminishes. Only a very small number of users contribute a small amount of information, resulting in minimal fluctuations in the amount of information, which remains stable.

Identification and delineation of stakeholders

Composition of stakeholders in each platform of public health emergencies.

Due to factors such as platform positioning and user habits, the categories of stakeholders and their proportions differ across different platforms in the context of public health emergencies. Therefore, analyzing data solely from a single platform cannot fully capture the evolution of public opinion events. Integrating data from various platforms allows for a more comprehensive understanding of the thematic and emotional evolution of public opinion events, providing valuable guidance for the government in managing public opinion.

Public health emergencies involve multiple stakeholders, including the government, media, healthcare sector, and enterprises, among others. In accordance with stakeholder theory and relevant literature [17, 18], and based on the data collected from each platform, the types of stakeholders involved in public opinion events on each platform are sorted and analyzed, as illustrated in Fig 2. For the purpose of this study, data from three platforms, namely Weibo, Douyin, and Today’s Headlines, are selected, considering data availability. As shown in Fig 2, Weibo, one of the largest social media platforms, encompasses the largest number of stakeholders, followed by Douyin. Today’s Headlines, being a news and information platform, includes a group of stakeholders primarily from the media category. Additionally, government-related organizations, healthcare-related stakeholders, and other stakeholders express their views on these platforms.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The stakeholder composition of various platforms.

https://doi.org/10.1371/journal.pone.0304877.g002

Stakeholder division.

The stakeholders involved in this study are in the following 10 categories: government-related organizations, education-related, mainstream media, general media, self-media, tourism-related, the masses, healthcare-related, corporations, and non-governmental organizations (NGO).

This study applies a method that utilizes authentication information and keywords in user profiles to classify and identify users as stakeholders. Users are classified by matching their authentication level and authentication information. In cases where authentication information is unavailable, the user’s username and keywords in their profile are used to determine their specific stakeholder category. The specific classification criteria are presented in Table 1. In this study, we chose two indicators, the average of postings and the average of followers, to categorize mainstream media and general media. We first categorized all media category stakeholders. Then, we calculated the average value of post volume and the average value of number of fans for all media category stakeholders. Finally, we classify the media whose post volume and follower count are higher than the average value as mainstream media, and the rest as general media.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Classification criteria for various stakeholders.

https://doi.org/10.1371/journal.pone.0304877.t001

Theme extraction

LDA thematic modeling.

With the advancement of text mining technology, various topic mining techniques have emerged, including TF-IDF, TextRank, and other algorithmic models. While these algorithms can extract topic features from text data, they often overlook the semantic relationships within the text and fail to effectively capture the underlying connections between topic words. Therefore, this paper adopts the LDA to mine textual topics, taking into account the contextual semantics and aiming to capture the inherent relationships between topics.

LDA (Latent Dirichlet Allocation) is an unsupervised clustering algorithm based on Bayesian ideas proposed by Blei et al [19], which is widely used in the scenarios of text clustering, text analysis, text keywords, etc. The structure of the model is shown in Fig 3, where α and β are the distribution hyperparameters of the topic distribution θ and the topic word distribution ϕ, respectively [20]. z and ω are the final topics and topic words generated by the model, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. LDA model structure diagram.

https://doi.org/10.1371/journal.pone.0304877.g003

The specific generation process of themes and theme words is as follows: ① Generate the theme distribution of the document θ_i by sampling from the Dirichlet distribution α; ② Generate the theme z_i,j of the jth word of the document i by sampling from the polynomial distribution of the theme θ_i; ③ Generate the distribution of the words corresponding to the theme z_i,j by sampling from the Dirichlet distribution β; ④ Generate the words ω_i,j by sampling from the polynomial distribution of the words , and the joint distribution is shown in Eq (1).

(1)

Emotional analysis

Sentiment classification is the process of analyzing, processing and summarizing text with sentiment. Traditional sentiment classification is mainly based on lexicon or feature engineering, which requires tedious manual feature design and a priori knowledge, and the understanding stays at a shallow level with poor scalability. The proposal of deep learning solves the above problems. Currently, deep learning technology has been widely used in various fields [22, 23]. Deep learning-based sentiment propensity classification models are able to provide end-to-end semantic understanding of the input text and judge the sentiment propensity based on the semantic representation. The convolutional neural network model can automatically learn the features in the input data without manually designing the feature extractor, and the convolution operation can capture the local features of the input data, so that the model can better understand the structure and content of the data, and extract the features through the convolutional layer and the pooling layer, which can efficiently extract the local features in the text, so as to achieve the purpose of judging the emotion of the utterance. In contrast, logistic regression models can only perform linear transformations, making it difficult to capture complex semantic relationships in text. In addition, there have been studies demonstrating the superiority of CNN models in sentiment classification tasks compared to logistic regression models. Smith et al. compared the logistic regression model and CNN model on the same sentiment dataset in terms of accuracy, recall and other metrics. The experimental results show that the CNN model has better performance in the sentiment analysis task compared to the logistic regression model. The CNN model achieves higher accuracy in the sentiment classification task [24]. Dhammi et al. investigated the application of Convolutional Neural Networks (CNN) and Logistic Regression models in task of sentiment analysis on Twitter. The experimental results demonstrated that the CNN model significantly outperformed the logistic regression model in the task of text sentiment classification [25]. Kaur et al. conducted a comparative analysis of logistic regression and convolutional neural network (CNN) models for sentiment analysis of financial news. The study also investigated the impact of different feature representations on the performance of the models and validated the advantages of CNN models in text sentiment classification tasks [26]. Based on this, this study adopts the convolutional neural network model CNN based on the large language model to classify the emotional tendency of text data.

In general, the CNN model is mainly composed of four parts: the input layer, convolutional layer, pooling layer, and fully connected layer. Fig 4 illustrates the structure of the Convolutional Neural Network (CNN) for sentiment classification.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Convolutional neural network sentiment classification structure diagram.

https://doi.org/10.1371/journal.pone.0304877.g004

For a sentence of length n, assuming that x_i is the k-dimensional vector of the ith word in the sentence, this sentence of length n can be represented in the input layer as: (2) where ⊕ is the concatenation operation.

In the convolutional layer, assume that the input feature matrix is w∈R^n×k, its sub-matrix is w_i∈R^h×k and the convolution kernel is , then the output after convolutional computation is: (3) where f is the activation function and b is the bias top.

In order to avoid the overfitting phenomenon, the data features output from the convolutional layer need to be compressed and downgraded in the pooling layer. By reducing the dimensionality of the word vectors, the model computation can be greatly reduced, thus avoiding the overfitting problem efficiently. Finally, in the fully connected layer, a classifier is used to normalize the resulting feature vectors to obtain the score of the current text, and the corresponding term of its resulting maximum probability value is the corresponding sentiment polarity of the text.

The specific parameters of the CNN model used in this paper are shown in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Model parameter setting.

https://doi.org/10.1371/journal.pone.0304877.t002

Data acquisition and pre-processing

Taking the "Norovirus" event as an example, the crawler tool is used to collect data from Weibo, Douyin, and Today’s Headlines with "Norovirus" as the search keyword. The collection fields include: publisher id, publisher name, posting time, posting content, publisher authentication information, publisher profile information, number of followers, etc. The collection time period is set from 1 February 2023 to 15 March 2023 based on the Baidu index curve of the "Norovirus" event. The text data collected from the three platforms are fused, the specific operations are: ① pattern matching; ② entity alignment; ③ data fusion. The fused text data are subjected to data cleaning with the following rules: ① Delete duplicate contents posted by the same user and keep only one; ② Remove links, @ other users, and special symbols in the text. Finally, 29277 cleaned data were obtained.

Table 3 shows some examples of sample data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Sample data.

https://doi.org/10.1371/journal.pone.0304877.t003

Delineation of the stages of evolution of public opinion

Fig 5 shows the Baidu index curve of "Norovirus", from Fig 5, we can observe several obvious inflection points, respectively, February 12, February 17, February 27, March 6, combined with the history of the release of the event, "February 10th CDC issued a reminder that norovirus has entered the high incidence period "; "On 17 February CDC released knowledge about norovirus prevention, and major media outlets scrambled to forward it"; and "On 26 February major media outlets began to report less about norovirus, and the public’s attention began to shift". According to the principle of life cycle division proposed in this study, the evolutionary cycle of the studied cases was divided into five stages: germination period P1 (1 February-12 February), outbreak period P2 (13 February-17 February), spread period P3 (18 February-27 February), decline period P4 (28 February-6 March), and recovery period P5 (7 March-15 March).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Baidu index curve of "Norovirus".

https://doi.org/10.1371/journal.pone.0304877.g005

Stakeholder theme evolution

When using LDA for topic extraction, in order to ensure the independence of topics, the indicator of topic confusion is used to determine the optimal number of topics, which is calculated as shown in Eq (4), where the denominator is the sum of all words in the test set D and p(w_d) is the probability of occurrence of each word w_d in the test set. (4) We calculate the topic perplexity for different numbers of topics in the LDA model and plot the topic perplexity curve, as shown in Fig 6. We observe that when the number of topics, denoted as K, is set to 26, the topic perplexity reaches a minimum and the topic extraction results become relatively stable.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Topic perplexity curve.

https://doi.org/10.1371/journal.pone.0304877.g006

Table 4 shows the meaning of each theme and the characteristic words of each theme.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Theme meaning and theme feature words.

https://doi.org/10.1371/journal.pone.0304877.t004

The number of texts for each theme in different periods is tallied, and the evolution of themes is depicted in Fig 7. Overall, the theme "sharing of norovirus infection (T22)" emerges as the most prevalent among all themes. This prominence can be attributed to the fact that during public health emergencies, the general public constitutes a significant portion of the population. Consequently, when such events impact them, the public expresses their emotions and opinions through social media. Analyzing the distribution of theme proportions across each period, it becomes evident that certain themes dominate in specific periods while having a diminished presence in others. For instance, the theme "CDC warns of high incidence of norovirus (T6)" accounted for the highest proportion during the initial stage, indicating that governmental departments started paying attention to the event at its early onset.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Thematic evolution accumulation map.

https://doi.org/10.1371/journal.pone.0304877.g007

Stakeholder concerns

To further elucidate the variations in topics of concern among stakeholders during the evolution of public opinion, we tally the top five topics of interest for each stakeholder group and summarize their concerns, as illustrated in Fig 8. Fig 8 represents the five topics that different stakeholders are most concerned about during the evolution of public opinion. In Fig 8, the text in the box represents the topics that stakeholders are concerned about. Each row represents the five topics that the stakeholders in that category are most concerned about.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Comparison of concerns of different stakeholders.

https://doi.org/10.1371/journal.pone.0304877.g008

From Fig 8, it can be visualized that the themes of concern of the three groups of stakeholders, namely government-related organizations, education-related and tourism-related, are highly overlapping, and NGO as well as the general media also have three themes that are consistent with the above three. Self-media and the masses, as well as healthcare-related and corporations, each have three consistent themes. The themes with high levels of concern were summarized and can be broadly categorized into event alerts (T20, T18, T13, T15), event causes and transmission pathways (T2), mass sharing (T22), knowledge and virus prevention (T1, T5, T7), virus testing (T12), and disease symptoms (T0, T17).

Patterns of stakeholder opinion evolution

We tally the text distribution of topics of concern for each stakeholder during each stage of the public opinion life cycle. The topic that ranks highest in each stage for each stakeholder type is considered representative of the concerns within that period for that stakeholder group. By doing so, we obtain the distribution of concerns for each stakeholder during each stage of the life cycle, as presented in Table 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Distribution of concerns of stakeholders at each stage.

https://doi.org/10.1371/journal.pone.0304877.t005

From the horizontal perspective, the sentiment values in each stage of the life cycle of public opinion evolution are calculated to obtain the concerns and sentiment evolution patterns of each stakeholder, and those with similar concerns are grouped into one category. The evolution patterns of each stakeholder can be roughly classified into six categories, as shown in Fig 9. Fig 9 represents the topics of greatest concern for each stakeholder at each life cycle stage and the change in sentiment from period to period. In Fig 9, the text in the boxes represents the topics that stakeholders are most concerned about, each column represents each life cycle stage, and each row represents the topics that each type of stakeholder is most concerned about in each period. The arrows represent the emotional changes of each stakeholder in each period, with blue color representing positive emotional tendencies and red color representing negative emotional tendencies. The arrow stays blue to indicate that the affective state remains positive, the arrow stays red to indicate that the affective state remains negative, the arrow changes from blue to red to indicate that the affective state changes from positive to negative, and the arrow changes from red to blue to indicate that the affective state changes from negative to positive.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. The concern and emotion evolution model of each stakeholder.

https://doi.org/10.1371/journal.pone.0304877.g009

Firstly, we analyze the theme evolution characteristics of each stakeholder. As depicted in Fig 9, similarities can be observed in the theme evolution process of three stakeholder groups: government-related organizations, education-related and general media; self-media and the general public; and corporations and non-governmental organizations. Specifically, government-related organizations, education-related entities, and general media exhibit complete consistency in their topics of concern during the first four stages of the public opinion life cycle. In the communication process of public health emergencies, the government plays an important role in guiding and channeling public opinion, and plays an important role in public opinion management [27]. After the occurrence of the event, the government-related organizations take the lead in reacting to the event, which is shown in Fig 9: During the germination period, government-related organizations initiate reporting and provide event reminders. As the incident progresses, the government focuses on addressing urgent societal issues, such as virus prevention and the underlying causes of the incident. The media plays the role of information release and dissemination in the evolution of public opinion. After the incident, as a bridge of information dissemination, major media will report on the relevant measures taken by government agencies. In this study, the media are divided into three subcategories: mainstream media, general media, and self-media, and the evolution of the themes of these three subcategories is different, as shown in the following: the general media are more likely to follow up on the initiatives of governmental organizations in the process of public opinion evolution; the mainstream media are more inclined to remind the media of the incident and report on secondary events; and the self-media pay more attention to the themes that are closely related to their own, such as "sharing about norovirus infection". In addition, medical stakeholders prioritize the popularization of virus-related knowledge throughout the evolution of public opinion. Conversely, the general public pays closer attention to the impact of the incident on themselves, focusing on themes related to "sharing of Norovirus infection". As the event progresses, corporations tend to concentrate on topics associated with their own interests, such as "virus detection" and "explanation of the absence of a norovirus vaccine in the market".

Secondly, we analyze the emotional evolution characteristics of each stakeholder. From the perspective of the main entities involved, the emotional evolution process of mainstream media and government-related organizations is similar. Throughout the public opinion life cycle, medical stakeholders consistently maintain a positive emotional state. However, the general public expresses more concerns about the incident and tends to exhibit negative emotions. From the perspective of the public opinion life cycle, during the germination period when the event initially occurs, the amount of information and its impact are relatively small, resulting in fewer people being concerned about the event. As a result, public opinion tends to be positive. In the outbreak period, although negative sentiment increases to some extent, overall, public opinion still tilts towards the positive side. During the spreading period, as the event further intensifies, negative news related to public opinion begins to emerge, leading to a shift in public opinion towards the negative side. However, with the intervention of relevant government departments, they play a role in controlling the further development of public opinion, resulting in a shift back towards a positive sentiment during the decline period. In the recovery period, when the event is essentially over, only a small number of people remain concerned, and the momentum behind the event diminishes. Consequently, public opinion continues to remain positive.

Vertically, Fig 10 illustrates the differences in themes of concern among stakeholders within the same lifecycle. In Fig 10, the text in the boxes represents the topics that stakeholders are most concerned about, and the text on the boxes represents the stakeholders who are concerned about that topic. Each column represents a life cycle stage, except for the leveling off period, which consists of two columns.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. Comparison of concerns of stakeholders in the same life cycle.

https://doi.org/10.1371/journal.pone.0304877.g010

In the germination stage, "CDC warns of high incidence of norovirus (T6)" is the theme that stakeholders pay more attention to. In the outbreak stage, the focus of stakeholders is shifted to the theme "popularization of knowledge on virus prevention (T5)". During the spreading period, the focus of the population was more on the topic "Causes of the high incidence of infectious diseases and routes of virus transmission (T2)". During the decline period, the focus of the population was not only limited to the current events, and "Guidelines for the prevention of springtime infectious diseases (T8)" became the most popular topic. In the recovery period, the distribution of themes among stakeholders was more dispersed because the public opinion event had basically ended. Overall, in the first four periods, the topic concerns of all types of stakeholders were focused on certain topics, while in the last period the topics of concern of each stakeholder were more dispersed, with a two-stage process of concentration to dispersion of each stakeholder’s concerns.