We collected Chinese news and articles related to COVID-19 from January 1, 2020, to February 20, 2020. We then applied the latent Dirichlet allocation (LDA) modeling method to derive useful information from these news reports.

Data from Chinese news and related articles were collected from the WiseSearch database [6]. The WiseSearch database is one of the most reputable, ever-growing Chinese media content databases, containing the news and article data from more than 1500 print media sources and over 10,000 internet media sources. It is famous for its reproducibility, timeliness, great coverage, and high data integrity compared with the other database [7]. The news and article data in the WiseSearch database are updated in a timely manner [6].

To gain insights into the early period of health information communication related to the coronavirus, we conducted a search with the keyword “coronavirus” in the WiseSearch database.

LDA is a generative probabilistic topic modeling method that is widely applied in text mining [8], medicine [9,10], and social network analysis [11] due to its excellent capability of converting visual words, a small part of an image that conveys a certain message about the image or alternation of the pixels, into images and visual word documents [12-14]. It is a generative statistical model with a three-level hierarchical Bayesian model. The basic assumption of this model is a combination of words belonging to different topics [15]. LDA indicates that there may be various topics in an article and that the wording in that article is attributable to one of its topics. We can discover the topics among the data pool by using Gibbs sampling techniques [16].

A total of 11,220 articles were found with the keyword search “coronavirus,” dated between January 1, 2020, and February 20, 2020. After cleaning the data, 7791 articles remained.

Before applying the LDA modeling, we used Python (Python Software Foundation) to perform data cleaning and used the Python package Jieba for data processing [17,18]. The detailed data process is illustrated in Figure 1. We next removed common Chinese stop characters such as “ten,” “a,” “of,” and “it.” We removed duplicate news reports. We then excluded news reports about other coronaviruses like severe acute respiratory syndrome-related coronavirus or Middle East respiratory syndrome-related coronavirus manually. We also built a document-term matrix and used term frequency–inverse document frequency (TF–IDF) to process the data. TF–IDF is a numerical statistic that is used to reflect the importance of a word to an article in a corpus [19].

To seek a suitable LDA topic number and the explanations to investigate the relationship between the COVID-19 crisis and news reports, we conducted multiple studies. We used a coherence score to evaluate the selection of a suitable number of topics [20]. Topic coherence measures the consistency of a single topic by measuring the semantic similarity between words with high scores in a topic, which contributes to improving the semantic understanding of the topic. That is, words are represented as vectors by the word’s co-occurrence relation, and semantic similarity is the cosine similarity between word vectors. The coherence is the arithmetic mean of these similarities [21]. We used the Coherence Model from Gensim (RARE Technologies Ltd), the Python package for natural language processing, to calculate the coherence value [22]. According to Figure 2, the coherence score increased and reached a stable score as the number of topics increased to 20, then declined after the number of topics reached 25. However, we found that the results would be uninterpretable for humans if only statistical measures were applied [23]. As a result, we combined statistical measures and manual interpretation and chose 20 topics to analyze with the help of Python version 3.6.1 and the LDAvis tool [15]. We set λ=1 and set 20 topics and their keywords. Topics’ names were generated according to their corresponding keywords to expatiate the topics.

We also divided these topics into different themes to study them better. In the visualization, that is the 2D plane (Figures 3 and 4), 20 topics were represented as cycles. These circles overlapped, and their centers are determined by computed topic distance [15]. By this approach, these 20 topics were classified into nine main primary themes and are shown in Table 1. Textbox 1 shows illustrative quotes for each theme.

The COVID-19 crisis has aroused great public concern in China and around the world. Topic modeling provides an alternative perspective to investigate the relationship between media reports and the COVID-19 outbreak. We collected media reports, listed the reports number each day (see Multimedia Appendix 1) and used topic modeling to analyze them. Although several COVID-19 cases were found in December 2019, we observed few news reports about them, showing that the press media did not focus on this disease at that time. As the outbreak became severe and more pneumonia cases were confirmed, the number of news reports began to steadily increase and then rapidly increased on January 19, 2020. In general, news trends peak and wane, according to the confirmed cases during other infectious disease outbreak periods; however, in some cases, the mass media cannot capture the outbreak in time and, therefore, fails to become the leading indicator [29]. This is because it takes time and rigorous effort for journalists to choose a topic, investigate the situation, collect data, and verify the authenticity of the material before they can finally present the news; as a result, a delay ensues. Mass media news reports lag behind the real time coronavirus developments, indicating that the media does not play an adequately forewarning function in public health communication and sensitization.

There was a rapid increase of related news after January 19, 2020, showing that the mass media started to pay more attention to this outbreak. However, since the virus is novel and there are not enough studies on it, the mass media might have conveyed misinformation, which may have induced negative psychological effects in the public like fear, anger, or sadness [30]. In addition, being overfed with reports will result in mass communication fatigue that will dampen the media’s effect [31]. Therefore, the government and the mass media should figure out the suitable news themes and daily news numbers to enable the public to keep alert about the outbreak with less harmful mental pressure. The media should also be obliged to ensure the reports’ accuracy.

The topics focused on by the mass media can be divided into nine classifications. Theme 4 (prevention and control procedure) and theme 3 (medical treatment and research) were two major themes that together accounted for around half of the content. It is important for the government to communicate with the citizens using mass media during the disease outbreak [32]; therefore, in these reports, the management of important government departments, medical institutions, and community control methods are emphasized. Positive and enthusiastic forecasts backed with active public health interventions are disseminated, which can eliminate unnecessary public worry and extreme panic, aimed at asserting the nation’s confidence in virus containment and victory within a short period.

Control of the sources of infection, interruptions of transmission routes, and the protection of susceptible people are three major principles to prevent and control infectious diseases. To cope with the COVID-19 crisis, the Chinese government took measures based on these three principles. The detection of viral infections within public transportation networks aroused great public concern, given that the outbreak coincided with the Spring Festival when many were traveling. Few news reports about this are included in theme 4 (prevention and control procedure), suggesting that the mass media might not have been providing sufficient health information about detection within the transportation network.

The scale of medical treatments and research was the second most popular topic. Our results showed that the mass media conveyed this kind of health information by paying attention to the detection of suspicious cases, drugs that might cure patients, and the transmission routes of the virus. However, reports within theme 4 (prevention and control procedure) and theme 3 (medical treatment and research) mainly focused on the whole society, while instructions on personal prevention and clinic and medicine choices were less mentioned.

Influences on activities (home and abroad) were also reported together with economic influences, which was included in theme 7 (global/local social/economic influences). These data indicate that the impact of the COVID-19 crisis is not limited to the medical field but also extends to social and economic fields. It is also a worldwide health issue that requires people around the world to work closely together.

The term “confirmed cases” appeared in 9.58% (n=747/7791) of the articles. This indicates that the mass media has served a public health function, as case numbers and their changing rates in news reports can directly give the public intuitive feelings about the speed of viral spread, the momentum, and the hazard of this coronavirus. It can also help citizens remain alert about virus transmission and, therefore, change their daily habits accordingly.

Theme 2 (medical supplies) and theme 8 (material supplies and society support) connect the material supplies with the COVID-19 crisis. Since the outbreak was so sudden and the transmission is so rapid, people in affected areas require medical material and other necessities, especially after the Chinese government shut the major entrance to Hubei to control the outbreak. The mass media can communicate with other parts of China to call for donations and support.

There was an emphasis on Wuhan stories, where news stories focused on the lives of individuals instead of the whole city. We also observed that theme 6 (mental health) accounts for 4.38% (342/7791) of all news articles. Previous studies have shown that there was an increase in mental health problems in both the medical staff [33] and the residents under quarantine [34] during other previous disease outbreaks. Therefore, these kinds of news reports can help the readers refocus on this easily neglected area, and therefore, early interventions can be made. These two themes indicate that the mass media adopts a people-oriented principle when reporting on the COVID-19 crisis, contributing to the warm society phenomenon.

This study is the first step to understanding the Chinese mass media’s role during the COVID-19 crisis. However, there are still several limitations in our study. First, we included a large number of Chinese news articles about COVID-19 from the WiseSearch mass media database, which only covers text news articles. However, the mass media has recently used new media platforms such as TikTok (video social media) and WeChat (the largest Chinese instant messaging app) to deliver health information through images, snapshots, and short videos. Therefore, we may have omitted news content and the impact of mass media in these media platforms. Second, we only selected a certain period of the outbreak. The pandemic is still ongoing, and the topics and themes are changing; therefore, we may have missed some novel topics and themes. Third, the LDA model has its own limitations such as a lack of nuances for qualitative thematic analysis and poor performance on short articles. Some relative studies introduced sentiment analysis to investigate the emotional differences in the message content [35]; it would be valuable if we could also apply sentiment analysis to supervise the news and investigate the public’s reaction to news related to COVID-19.

Collecting and analyzing reports on the novel coronavirus shed light on how the Chinese media have delivered health information during the COVID-19 crisis. Our study provides evidence that the Chinese mass media news lags behind when reporting the major developments of the viral spread. Prevention and control procedures, medical treatment, and research are major themes of the press but mainly focus on the whole society, while instructions on personal and individual prevention, clinic and medicine choices, and detection need to be further enhanced. Global and local influences were reported as the COVID-19 crisis started to impose pressure on public health worldwide and urged cooperation among all humankind. Further research should be considered to explore the impacts of mass media on the readers through sentiment analysis of news data and the influences of misinformation about COVID-19 delivered through the mass media.