The issue-attention cycle model [15] conceptualizes how an issue rises into and fades away from the center of public attention. In the first stage, labeled the preproblem stage, an undesirable social condition (eg, the appearance of COVID-19) emerges but does not yet draw much public attention. The second stage, dubbed alarmed discovery and euphoric enthusiasm, occurs when a triggering event (eg, the national spike of newly confirmed cases of COVID-19) heightens public awareness of the issue. In the third stage, called realizing the cost of significant progress, people begin to recognize the hardship involved in restructuring society, and individuals must sacrifice to solve the problem. This causes a gradual decline of intense public interest, the fourth stage. In the final postproblem stage, the current issue is replaced by a new issue and moves into a twilight zone of reduced public attention.

Not all issues follow the five stages of the issue-attention cycle [17]. As the cyclical patterns of public attention evolve, a wide array of public discourse has been found across multiple issues of climate change [18], emerging technologies [19,20], and public health risks [21,22]. There are also cultural differences in such discourse patterns. For example, concerning the H1N1 pandemic, South Korean news coverage showed five phases of increasing or decreasing attention. The corresponding US news coverage of the pandemic saw only two phases during the same 7-month time period [23].

Despite these fragmented findings, the issue-attention cycle framework provides insights into how public attention dramatically waxes and wanes. An issue that has gone through the cycle is different from issues that have not gone through the cycle in at least two ways. First, when an issue has achieved national prominence, new institutions, programs, and measures will have been developed to address the situation. These developments and their societal impacts are likely to persist even after public attention has shifted elsewhere. Second, the prolonged impacts of these developments are shaped by what was heavily discussed when the issue was of primary public concern.

Although the issue-attention cycle was initially proposed to model traditional media such as newspapers and television, there is a burgeoning literature applying the model to social media platforms. Among them, Twitter serves as a forum that the public is increasingly turning toward to seek and share information that is not subjected to a gatekeeping process [24]. It has become common for journalists to refer to tweets in their news stories. Research has also found that Twitter takes the lead in and exerts control over public discourse, particularly in the early stages of an issue-attention cycle [20,25].

Building on these prior studies, we analyzed Twitter conversations about COVID-19 to examine social media’s issue-attention cycle. We present how to build an end-to-end method of identifying meaningful topic phases dynamically. This allows us to compare how issue-attention cycles appear in different countries on the same catastrophic event. To the best of our knowledge, no study has applied dynamic topic modeling in the context of risk communication.

Studies have examined various impacts of the pandemic. Researchers have focused on predicting the transmissibility of the virus. One study estimated the viral reproduction number (R₀) of SARS-CoV-2, which is known to be more substantial than that of severe acute respiratory syndrome (SARS)–related coronavirus, which was the cause of the SARS outbreak that first appeared in Guangdong Province in southern China in 2002 [26]. Another work based on a stochastic mathematical prediction model of infection dynamics claimed that, by reducing worldwide travel by 90%, the epidemic’s spread could be significantly reduced [27].

Other studies have sought to understand the propagation of misinformation related to COVID-19. One study used an epidemic model to represent the spread of misinformation about COVID-19 on various social media platforms such as Twitter, Instagram, YouTube, Reddit, and Gab; the study showed that users interact and consume information differently on each platform [28]. In this regard, media platforms such as Facebook, YouTube, and Twitter claim to attempt to redirect people to reliable sources of medical information and, to this end, have established direct lines of communication with the Centers for Disease Control and Prevention and the World Health Organization [29].

Among the regional research, one article argued that fake online news in Japan has led to xenophobia toward patients and Chinese visitors [30]. Another study surveyed 300,000 online panel members in South Korea in 2015, when the Middle East respiratory syndrome outbreak was prevalent in this country [31]. This work found that, if public health officials’ information is untrustworthy, people rely more on online news outlets and communicate more via social media.

More recently, a report showed that the public could not easily receive the information on COVID-19 shared by public health officials due to prevalent misinformation on fake cures and conspiracy theories [32]. This study showed that infodemics’ harm varied from country to country depending on public confidence in authorities. One study compared trends in three countries (ie, the United States, the United Kingdom, and Canada) in terms of political bias and found that, although political polarization surrounding COVID-19 exists in the United States and Canada, individuals’ exact perspective on the pandemic is broadly related to the quality of their reasoning skills, regardless of political ideology [33].

Several studies have used data gathered from Twitter to analyze risk communication amid COVID-19. Some of them focus on sentiment analysis based on conventional rule-based lexicon models [34] or deep learning classifiers [35]. These studies measured the degree of sentiment polarity, such as positive and negative, and provided insights from observing daily sentiment changes.

Many types of data sets have been released to the public and research communities on COVID-19. One study crawled Twitter for approximately 3 months and collected information on tweets with relevant keywords in 10 languages [36]. Another work collated over 59,000 academic articles, including over 47,000 research papers, on COVID-19, SARS-CoV-2, and coronavirus-related issues [37] to conduct a comparison study.

Natural language processing such as topic modeling is increasingly used to process extensive documents and extract hidden thematic patterns of textual information [38]. Many studies have explored the capability of topic modeling in understanding the most important subjects of discussion on social media during crises and global epidemics such as dengue [39], Sika virus [40], and Ebola virus [41]. Given the remarkable performance of topic modeling in previous investigations, recent studies on the COVID-19 outbreak have also applied topic modeling to documents collected from different social media sites such as Facebook [42], Weibo [43], and Twitter [44].

One work analyzed COVID-19–related tweets over 2 weeks to study ongoing topics and found that Twitter can be considered a rich medium to understand public opinions in real time [45]. Another work conducted topic modeling on tweets to discover daily hot topics on the pandemic [46]. Furthermore, scholars leveraged Twitter to study the ecosystems of misinformation and conspiracy. One study has shown that users’ political orientation correlates to their contribution to the spread of pandemic-related conspiracy [47]; another study demonstrates the link between fake news exposure and low trust in media [48]. Several techniques have been developed to detect conspiracy and misinformation on social media [49-51].

Despite the growing literature on risk communication during COVID-19, most studies that use topic modeling extract topics from either the entire studied period or manually segmented periods. This study considers time and topics jointly; we used an algorithmic approach to identify topical phases that arise naturally. Our goal is to observe changing risk communication contexts (even when conversations contain similar keywords) from the issue-attention cycle perspective. We also chose to study risk communication in Asian countries that have received relatively little attention. Our data method is not restricted to the studied countries; it can be applied to other languages and countries.

We first tokenized the data, a process that can be defined as converting data to the smallest units that have meaning. We filtered unnecessary textual information such as stop words, special characters (nonletters), special commands, and emojis. We then used existing Python tokenizer libraries corresponding to each language. Detailed information about the language-specific tokenizers is explained on GitHub [16].

The next step is to demarcate specific phases divided by dates to extract topics. This is nontrivial since there are multiple fluctuations and changes in topics reflecting real events such as increased patients with COVID-19. Furthermore, we ruled out using the epidemic phases announced by each government because the offline epidemic phases do not seem to capture actual online topic trends as explained in the forthcoming Basic Daily Trends section.

The issue-attention cycle moderating public attention to a given issue can be measured in media attention, such as the number of news stories [17-19,21-23] or tweets [20,25] on the topic. We, therefore, isolated dates that show sudden increases in the daily tweet volume. We set up two learnable parameters of the first derivatives (hereafter velocity) and the second derivatives (hereafter acceleration) of the daily tweet volumes, as illustrated in the following equations, where D is a day, t is a target date, and t – 1 is 1 day past t:

We set the velocity and acceleration values when the country announced the first confirmed case as the ground truth (GT). This approach’s intuition is that velocity and acceleration are proxies for each country’s unique communication traits regarding a specific subject (ie, COVID-19 in our case). Once these values were computed from the first confirmed date, they were set identically for the remaining periods.

We established joint thresholds for velocity and acceleration to find dates where velocity is still smaller than velocity_GT, and acceleration becomes more substantial than acceleration_GT (0<velocity<velocity_GT and acceleration>acceleration_GT). In this manner, we identified the two parameters from the date of the first confirmed case by country and then detected other dates conjectured to be the start of forthcoming topic phases. When learning these parameters, for velocity, we rounded down the velocity_GT value and added 1, and for acceleration, we rounded down accelerationGT, which is similar to the machine learning approach’s concept of loss minimization (ie, a learning process is finished by one step).

We adopted a low-pass filter with 0.2 as the low-frequency threshold to remove noisy signals and smooth the data. Finally, the temporal data are divided into topic phases (see Multimedia Appendix 1 to find the computed daily velocity and acceleration trends and demarcated phases by country).

We used latent Dirichlet allocation (LDA) for the topic modeling task. LDA is a well-known machine learning method to extract topics from given textual documents (ie, a collection of discrete data points) [56]. LDA generates and maximizes the joint probability of the topics’ word distribution and the documents’ topic distribution [38]. For short sentences, LDA-based methods may not work correctly due to sparse co-occurrences. However, in our case, tweets were collected via specific keywords; therefore, the topics were the focus, and the word co-occurrences among tweets were no longer sparse. Various studies have used the same LDA method on short documents, including Twitter [57-59].

The topic count for each phase is a hyperparameter. The topic count’s range is between 2 and 50. We calculated perplexity, that is, the probability of how many tokens might be placed in the next step (ie, indicating ambiguity over the next possible token). Perplexity is a metric that is often used to optimize language models [60]. The minimum required frequency of words for each phase in tweets was set to 20, and each phase’s epoch (ie, a number of iterations to train LDA) was set to 100. We then decided the optimum number of topics for each phase by choosing the minimum perplexity value. We further analyzed our modeling results’ reliability and confirmed that the results were steady and dependable (see Multimedia Appendix 2 for more details).

Table 2 shows how many prominent topical phases were found for each country. For each phase, we list the statistics of the risk communication, including the period of the topical phase, the total tweet count during the phase, the average user count per day, the average original and retweet counts per day, the ratio of original tweets to retweets, and the number of topics suggested by perplexity.

This step involves labeling the themes of the extracted topics and allocating semantic meanings to each topic. We first sorted all tweets with the identified topics in descending order (ie, tweets on the most prevalent topics listed first) and discarded the minor topics that accounted for less than 25% of all tweets.

We then extracted the top 1000 retweeted tweets and the 30 keywords with the highest probability of usage for each topic. We provided these data sets to local users from each country and asked them to label themes for each topic based on the given data sets. Any similar or hierarchical topics were then merged via qualitative coding into a higher category. If one topic corresponded to several themes, then it was given multiple class labels. The maximum number of multiple classes within topics was two, and each class within a topic was weighted as 0.5 in the plot of daily trends in the number of tweets.

Human annotators, who are familiar with the local language and Twitter, qualitatively assessed the extracted topics. First was the intralevel, where annotators labeled each topic based on the contents of the sampled top 1000 tweets and top 30 words. The second was the interlevel, where the annotators compared tweet contents and top-occurring words among topics regardless of the phase. Other annotators then cross-checked the assessment.

The Cohen kappa coefficient to measure the intercoder reliability was 0.766 (see Multimedia Appendix 2 for details on this validation and the list of topics and top-occurring words for each country). Our analysis objective was not to substitute human laborers on monitoring misinformation but to assist them by grouping tweets into specific topics, including misinformation.

Concerning the local and global news themes, we narrowed down the labels since people talked about different news categories. We sublabeled tweets as “_confirmed” if it was about confirmed cases or deaths, “_hate” if it was about hate crimes toward individual races, “_economy” if it was about the economic situation and economic policies, “_cheerup” if it was about supporting each other, and “_education” if it was about when to reopen schools; finally, no sublabel was given to tweets about general information.

The first patient with COVID-19 was reported in South Korea on January 20, 2020. This explains why the tweet count remains relatively low during early January and mostly increases only after late January (see Figure 3). On January 25, the Korean government issued a travel warning for Wuhan and Hubei Province, and suggested that Korean citizens evacuate from those areas, which was heavily discussed on Twitter.

On February 18, 2020, the tweet numbers increased sharply due to the 31st confirmed case related to a cult religious group Shincheonji in Daegu City. After this case was confirmed, the quarantine authority began rigorous testing, focusing on Daegu, and the number of confirmed cases increased drastically until mid-March. The tweet trends follow an identical pattern. However, the official epidemic phases announced by the government, represented by vertical dashed lines in the figure, seem to lag behind the increases in the number of tweets. This pattern shows that the official epidemic phases do not align well with the amount of online attention.

We repeated the analysis with the other three countries, as shown in Figures 4-6 (see Multimedia Appendix 3 for each country’s detailed explanation).

The data yielded a total of four topic phases, which are used in Figure 7. Phase 0 has no related topics. For phases 1, 2, and 3, the number of topics varies from 8, 5, and 11, respectively. In phase 1, people talked a great deal about their personal thoughts and opinions linked to the current outbreak, and they tried to cheer each other up. In phase 2, as the crisis peaked, people talked less about personal issues and mainly about political and celebrity issues. In Korea, political discussions revolved around closing the South Korean border with China and other countries. In phase 3, as the daily number of tweets decreased relative to that in phase 2, people talked about diverse topics, including local and global news. The major topics here included worries about hate crimes directed toward Asians in Western countries. Such diverse topics are likely shown when people think the pandemic has just passed its peak.

We portrayed daily trends of interest in other countries by counting the tweets mentioning other countries’ names in local languages or English. Korea, China, and Japan were mentioned most frequently; we suspect that this was mainly triggered by political and diplomatic relationships. Meanwhile, the United States and Italy were both mentioned steadily across the 3 months, with the media outlets broadcasting global news affecting this phenomenon.

We repeated the same analysis and interpreted the results for the other cases (Iran, Vietnam, and India), as depicted in Figures 8-10 (see Multimedia Appendix 2 for the derived topic trend and the mentioned country name trend graphs and detailed corresponding explanations by country).