Exploring Public Response to COVID-19 on Weibo with LDA Topic Modeling and Sentiment Analysis.

It is necessary and important to understand public responses to crises, including disease outbreaks. Traditionally, surveys have played an essential role in collecting public opinion, while nowadays, with the increasing popularity of social media, mining social media data serves as another popular tool in opinion mining research. To understand the public response to COVID-19 on Weibo, this research collects 719,570 Weibo posts through a web crawler and analyzes the data with text mining techniques, including Latent Dirichlet Allocation (LDA) topic modeling and sentiment analysis. It is found that, in response to the COVID-19 outbreak, people learn about COVID-19, show their support for frontline warriors, encourage each other spiritually, and, in terms of taking preventive measures, express concerns about economic and life restoration, and so on. Analysis of sentiments and semantic networks further reveals that country media, as well as influential individuals and "self-media," together contribute to the information spread of positive sentiment.

Introduction

In recent decades, we have encountered several disease outbreaks such as SARS in 200 and MERS in 2012. Nowadays, we are facing another enemy: COVID-19. As of July 26, 2020, more than 16 million COVID-19 confirmed cases have been reported around the world (JHU COVID-19 Resource Center, 2020). With the first confirmed case officially reported in Wuhan, China, in late December 2019, the global COVID-19 outbreak has brought great damage to the world's normal operations for over half a year and has been, unfortunately, aggravating, as of July 2020.

During and after the disease outbreak, public opinion is commonly collected as it is useful in many ways such as improving communications in terms of public concerns, crisis management, health knowledge promotion, and so on between governments and the public (Holmes, Henrich, Hancock, & Lestou, 2009; Mollema et al., 2015). Traditionally, surveys have played an important role in gathering public opinion, and undoubtedly, they have also been applied to disease outbreak research about public opinion. Table 1 summarizes some popularly investigated themes in survey-based research of the COVID-19 outbreak.

Table 1

Popular Themes in Survey-Based Research of COVID-19 Outbreak

Themes	Sample research
Knowledge, attitudes & practices (KAP)	Wolf et al. (2020); Zhong et al. (2020)
Psychological stress	Huang and Zhao (2020); Mazza et al. (2020); Qiu et al. (2020)
Information seeking	Ebrahim et al. (2020); Liu (2020)
Misinformation (fake news)	Greene and Murphy (2020); Motta, Stecula, and Farhart (2020)
Sensitive individuals, including front-line hospital staffs and recovered patient	Bhagavathula, Aldhaleei, Rahmani, Mahabadi, and Bandari (2020); Huang, Han, Luo, Ren, and Zhou (2020)
Attitudes towards government actions on disease control	Atchison et al. (2020)

While surveys play an irreplaceable role, there is another way of collecting public opinion about disease outbreaks: the increasing popularity of social media and the development of text analysis techniques have given rise to mining social media data. Unlike surveys that can mainly collect only a limited volume of data and may not be able to reflect the real thoughts of the public because of some influential factors during the survey process, such as the environmental distraction, subject's psychological pressure, and so on (Hridoy, Ekram, Islam, Ahmed, & Rahman, 2015), social media data, which is the collection of short texts posted online to express one's feelings at any time, are generated on a large scale. Thus, the efficient use of social media data can contribute largely to the research about public opinion.

Reported as the first storm center affected by COVID-19, China had undergone an incredibly hard time at the very beginning, encountering difficulties such as city lockdowns, lack of medical supplies, lack of hospital resources, and so on. But luckily, by July 2020, China had also made satisfactory achievement by getting the disease outbreak under control with aggressive approaches (Campbell, 2020; Clinch, 2020). Weibo, the biggest social media platform in mainland China, serves the functions of information sharing and communications in the country. Data collected from Weibo can be analyzed to understand the Chinese people's reaction to the disease outbreak and the characteristics of semantic networks, which lead to the research questions (RQs) of this research.

RQ1: What topics can be detected from Weibo posts on COVID-19?

RQ2: How do sentiments change over time, and what are the characteristics of different semantic networks?

The rest of this article is organized as follows. The second part reviews some related work in the context of this research, after which methodologies used to conduct this research are introduced, and the results are presented and discussed. Finally, conclusions are made, together with the contribution of this work to the research context and suggestions on further research directions.

Related Work

Disease Outbreaks on Social Media

Table 2 records some exemplary research on the disease outbreak on social media. It can be observed that social media data contribute to several types of research on disease outbreaks, including public opinion mining, sentiment analysis, semantic network analysis, disease outbreak detection, and so on; the methods applied to social media data analysis vary from manual coding to machine learning techniques; Twitter is a worldwide platform for collection and analysis of social media data, while for social media data collection from residents in mainland China, Weibo is a more popular choice.

Table 2

Research on Disease Outbreaks on Social Media

Author	Disease outbreak	Social media	Methods	Findings
Mollema et al. (2015)	Measles	Twitter and others	Thematic analysis	People on Twitter cared about disease transmission, preventive actions, and vaccination; governments needed to promote vaccination acceptability.
Fung et al. (2013)	MERS-CoV & H7N9	Weibo	Statistical analysis on the number of Weibo posts	Weibo users reacted to the disease outbreak significantly, and people paid more attention to the H7N9 outbreak.
Ye, Li, Yang, and Qin (2016)	Dengue	Weibo	Analysis on the numbers of posts and spatial information	Spatially and temporally, there was a correlation between the number of posts and disease development trends.
Chew and Eysenbach (2010)	H1N1	Twitter	Manual and automated coding	Several sentiments, including confusion, humor, risk, and so on, were discovered, among which humor was the most popular sentiment.
Ye et al. (2016)	Influenza	Twitter	Modeling	A prediction model built on Twitter data could be used for influenza outbreak alerts.
Zhang et al. (2015)	H7N9	Weibo	Analysis on the number of Weibo posts and the number of new confirmed cases	There was a positive correlation between discussion and disease outbreak level, and Weibo served as a good medium to promote communications of public health.
Li et al. (2020)	COVID-19	Weibo	Machine learning algorithms	Weibo posts were classified into seven categories of situational information. Useful text features should be helpful in building an emergence response system.

Topic Modeling Using LDA

Latent Dirichlet Allocation (LDA) assumes that a document is generated based on a certain number of topics, and each word in the document is randomly selected from its corresponding topic vocabulary (Blei, Ng, & Jordan, 2003; Gruber, Weiss, & Rosen-Zvi, 2009). It is an excellent probabilistic model that performs well in topic modeling and has been widely applied in research (Hu et al., 2017). For example, Barua, Thomas, and Hassan (2012) utilized LDA to discover topics and topic trends from a popular Q&A website in the programming field and found that the developer community discussed a wide range of topics and discussions of different topics are interconnected. Similarly, Hu et al. (2017) studied email corpora with the LDA model.

LDA has been applied to extract topics from various kinds of corpora, including but not limited to microblogs. For example, Huang, Yang, Mahmood, and Wang (2012) applied LDA with web usage data; Xu, Zhang, and Yi (2018) and Lim and Buntine (2014) studied tweets with LDA modeling. Therefore, it is logical to follow that exploring topics from Weibo datasets with LDA topic modeling should also harvest satisfactory results.

Sentiment Analysis

Sentiment of text data mainly refers to the emotions hidden within the text. Sentiment analysis has been widely applied to a large volume of opinion mining research, including product review analysis, public response to the stock market, and so on, as valuable information can be discovered if emotions in texts are well analyzed (Bakshi, Kaur, Kaur, & Kaur, 2016). In general, sentiments are classified into three categories: positive, neutral, and negative, and there are mainly two ways of sentiment tagging (Ray & Chakrabarti, 2017).

Lexicon-based sentiment tagging analyzes the words in a sentence and harvests the overall score by adding up the scores of each word, for which sentiment dictionaries are used (Bhonde, Bhagwat, Ingulkar, & Pande, 2015). This method was widely applied in the sentiment tagging of social media posts. For example, Ray and Chakrabarti (2017) used the R language to tag Twitter posts for product reviews with lexicon vocabularies and completed sentiment analysis at three levels: document, sentence, and aspect; Pérez-Pérez, Pérez-Rodríguez, Fdez-Riverola, and Lourenço (2019) used a lexicon-based tagger to analyze tweets' sentiments under each topic discovered in the Human Bowel Disease community.

Machine learning approaches to sentiment tagging are also gaining popularity. For example, Chen and Sokolova (2018) adopted an unsupervised approach to cluster sentiments of clinical discharge summaries with word embeddings generated from Word2Vec and Doc2Vec models and compared the results. Salathé and Khandelwal (2011) compared three machine learning techniques in terms of sentiment classification performance and adopted both naive Bayes and maximum entropy algorithms for the supervised classification of vaccination sentiments.

Methodologies

Methodological Framework

Figure 1 shows the methodological framework of this work. This research starts with data collection of Weibo posts with a user-simulation-like web crawler. Posts are then collected and processed to remove text noises and stop words, after which text segmentation is also performed. Then topic modeling is conducted using the LDA model with the cleaned dataset. To answer the second research question, lexicon-based sentiment analysis is conducted on data, including the number of posts, network information, and so on, to reveal sentiment trends and discover the characteristics of semantic networks.

Figure 1

Methodological framework of the proposed research

Data Collection

This work examines the public opinion related to COVID-19 on Weibo, for which theme-related posts must be collected. Both web crawlers and APIs are nowadays widely used technologies to collect data from social media, including Weibo (Liu & Hu, 2019; Liu, Wu, Wang, & Li, 2014). Although Weibo API, being the official gateway to collect Weibo data, is easy to use, there are many constraints such as only providing a limited number of posts and so on (Zeng, Zheng, Chen, & Yu, 2014). To avoid such inconvenience, this work uses selenium to develop a simulation-based web crawler using Python, so as to satisfy the data collection task. The web crawler simulates human logins and searches Weibo posts based on given keywords. The HTML of webpages is then collected and parsed to get the posted Weibo content and relevant information such as username and so on.

To fulfill the research purposes, this work uses six keywords selected from Hu, Huang, Chen, and Mao (2020), as listed in Table 3 . Hot posts on each day over the period from January 1 to June 30, 2020, are collected.

Table 3

Selected Keywords for Data Collection

Keyword	Translation
??	Virus
??	(Confirmed/suspicious) case
??	Pneumonia
??	COVID
????	Coronavirus
??	Disease outbreak

Data Preprocessing

Removing Noises

Text noises are generally removed in the text mining research, which benefits the experimental outcomes (Celardo, Iezzi, & Vichi, 2016). In the context of this research, text noises include emoji codes, punctuation marks, symbols, non-Chinese words, and so on. To remove text noises, the “re” regular expression package in Python is adopted to remove all non-Chinese components in the dataset.

Stop Word List

In Chinese text, there are many “meaningless” words like “?” (I/me), “?” (you), “?” (have done something), and so on, which are normally removed in information retrieval and text mining tasks so as to improve the experimental outcomes (Zou, Wang, Deng, & Han, 2006). There are also some Chinese stop word lists built by university NLP labs and companies. To construct a stop word list, this work integrates three public stop word lists (Baidu stop word, SCU stop word, and HIT stop word) that are widely used (Xie et al., 2019), together with some domain stop words such as “????” (repost) that appear frequently in most of the posts collected.

Chinese Text Segmentation

Unlike English text, the Chinese text needs to be segmented for analysis tasks. In terms of Chinese text segmentation tools, the “jieba” package in Python is widely used and has many advantages, such as adding customized words (Day & Lee, 2016; Peng, Liou, Chang, & Lee, 2015). In this research, the “jieba” package is adopted to perform the Chinese text segmentation task.

Topic Modeling

As mentioned above, LDA is a powerful tool in topic modeling. LDA can extract a given number of topics from a corpus that contains a certain number of documents. This research applies LDA to extract a certain number of topics from the cleaned dataset with the Python package “gensim.” In terms of the determination of the number of topics, both perplexity and coherence scores are taken into consideration. While the former measures how well the model is generated from the corpus (the lower the better), the latter measures the sentence similarity of each topic in the dataset (the higher the better) (Blei et al., 2003; Xie, Qin, & Zhu, 2018). After the optimal model is determined, another Python package, “pyLDAvis,” is adopted to visualize the topic extraction results, with a coordinate graph to show the distribution of topics and lists of the top 30 most salient words in each topic. Topic labeling is completed manually, based on the given salient words.

Lexicon-based Sentiment Tagging

The sentiment tagging task in this research adopts a lexicon-based approach recorded in https://www.cnblogs.com/qiaoyanlin/p/6891437.html. In short, the process of calculating the sentiment score of a post mainly contains four steps: (1) sentences in a post are split based on the punctuation marks; (2) each sentence is then segmented and meaningful words remain; (3) each remaining word (including negation words that might reverse the sentiment of a sentence) provides its sentiment and/or weight score based on the lexical dictionaries; and (4) the overall score of the post is calculated based on the scores of each sentence, which are calculated from the scores of each word.

Statistics and Semantic Network Analysis

After sentiment tagging, descriptive statistics of the results are described. Time series analysis is then performed to discover interesting phenomena from the number of positive and negative posts over time. To take a closer look at the sentiments of public opinion, positive and negative semantic networks are constructed to identify important roles in and the characteristics of respective networks, for which network visualization and statistical analysis are performed.

Results and Discussion

Data Collected

Figure 2 is a screenshot of some collected raw data. In total, 719,570 posts are collected over the period from January 1 to June 30, 2020, based on the given keywords. After data cleaning is performed, which includes removing duplicate posts, blanks, and “NA” that come probably from parsing failure, only 374,225 posts remain.

Figure 2

A screenshot of some collected raw data

The dataset is then further processed with Chinese text segmentation and removing stop words, after which it is ready for text mining analysis. Figure 3 visualizes the top 500 frequent terms with a word cloud, and Table 4 records the top 50 frequent words in the dataset, from which some interesting preliminary findings can be observed: (1) undoubtedly, term frequencies of each keyword selected for data collection rank top in the vocabulary and (2) discussions of COVID-19 on Weibo might cover a wide range of topics, including local disease outbreak, international relations, prevention measures, global pandemic, hospital staff, vaccination, and so on, which are in line with some previous research findings on public opinions about disease outbreak (Jalloh et al., 2017; Nickell et al., 2004; Rubin, Amlôt, Page, & Wessely, 2009). The preliminary findings can be further confirmed with the following analysis.

Figure 3

Word cloud of top 500 terms in the cleaned dataset

Table 4

Top 50 Words in the Cleaned Dataset

No.	Term	Translation	Frequency	No.	Term	Translation	Frequency	No.	Term	Translation	Frequency
1	??	Disease outbreak	153,855	18	??	Country	23,856	35	??	Import	15,642
2	??	Pneumonia	126,294	19	??	Quarantine	23,311	36	??	Vaccine	15,560
3	??	COVID	98,978	20	??	Court	23,168	37	??	Treatment	15,506
4	??	Case	97,597	21	??	Work	22,905	38	??	Genuine	15,478
5	??	Virus	71,299	22	??	Hope	22,836	39	??	Staff	14,722
6	??	Novel	66, 986	23	??	Death	21,772	40	??	Situation	14,243
7	??	Infected	65,884	24	??	Hospital	21,209	41	??	Hubei	13,338
8	????	Coronavirus	60,636	25	??	Test	21,195	42	??	Beijing	13,018
9	??	America	52,716	26	??	Add oil	20,350	43	??	Health	13,005
10	??	China	50,935	27	??	Globe	19,263	44	??	Report	12,769
11	??	Wuhan	45,959	28	??	Time	19,247	45	??	Period	12,223
12	??	Infected	38,556	29	??	Accumulate	18,367	46	??	abroad	12,162
13	??	Disease control	33,447	30	??	Fight	17,996	47	??	Discharged	11,678
14	??	Video	32,489	31	??	Folk	17,381	48	??	Doctor	1,111
15	??	Patients	31,036	32	??	News	17,000	49	??	World	11,558
16	??	New case	28,674	33	??	Discover	16,799	50	???	Yuhua district	11,391
17	??	Face mask	23,942	34	??	Nation	15,650

Perplexity and Coherence Scores of LDA Models

Figure 4, Figure 5, respectively, show the perplexity and coherence scores of models trained under different settings, namely, different numbers of topics. Overall, with an increasing number of topics set for model training, a smaller perplexity score is gained, that is, the model performs better in predicting the samples. Notably, the setting of a larger number of topics may lead to model overfitting issues, and thus, it needs to be careful in deciding the number of topics for training the model. Another measurement for deciding the number of topics is to compare the coherence scores of different models. In this case, coherence scores fluctuate with an increasing number of topics, and peaks are observed when numbers of topics for model training are set as 4, 8, 12, 15, respectively. Based on the perplexity and coherence scores, first, it can be concluded that the number of topics set for LDA training should be 12 or 15. By observing the keywords generated for each topic, the number of topics is finally determined as 12, because it is easier to code the topics from the given keywords.

Figure 4

Perplexity scores of models under different settings of number of topics

Figure 5

Coherence scores of models under different settings of number of topics

Discovered Topics

Figure 6 is a visualization of the selected LDA model mentioned above. While the left panel shows the distribution of each topic, the word list on the right panel shows the top 30 most salient terms of the selected topic, in which the blue bar shows the overall term frequency in the dataset, and the red bar represents the estimated term frequency in the selected topic. Table 5 records the discovered topics, each with 10 representative words selected from the top 30 most salient terms.

Figure 6

LDA model visualization

Table 5

Top 30 Most Salient Terms of Each Topic and Topic Coding Results

Topic ID	Topic Label	10 representative words selected from the top 30 most salient words
1	Fight the virus together	?? (power), ?? (add oil), ?? (folk), ?? (hope), ?? (China), ?? (fight), ?? (fight the virus), ?? (front line), ?? (success), ?? (work hard)
2	Knowledge	?? (research), ?? (expert), ?? (reason), ?? (popular science), ?? (transmission), ?? (disease), ?? (science), ?? (prevention), ?? (exam), ?? (discover)
3	Assistance	?? (work), ?? (fight), ?? (come back), ?? (nation), ???? (united), ?? (on site), ?? (pay attention to), ?? (support), ?? (news), ?? (situation)
4	Economics	?? (globe), ?? (economics), ?? (influence), ??? (fund), ?? (world), ?? (internationality), ?? (control), ?? (cooperation), ?? (society), ?? (market)
5	Global pandemic	?? (Iran), ?? (case), ?? (UK), ?? (disease outbreak), ?? (new case), ?? (Japan), ??? (Italy), ?? (accumulate), ?? (urgent), ?? (abroad)
6	Prevention	?? (disease outbreak), ?? (face mask), ?? (promotion), ?? (prevention), ?? (do well in), ?? (prevent), ?? (fight the virus), ?? (risk), ?? (health), ?? (measure)
7	Treatment	?? (quarantine), ?? (discharged), ??? (no symptoms), ???? (medical observation), ???? (close contact), ?? (cure), ?? (fever), ?? (severely ill), ?? (confirmed affection), ?? (treatment)
8	Stay at home	?? (period), ?? (life), ?? (this time), ?? (happy), ?? (go back home), ?? (at home), ?? (like), ?? (thing), ?? (friend), ?? (home)
9	Law	?? (court), ?? (judge), ?? (folk), ?? (breach the law), ?? (case), ?? (defendant), ?? (proof), ?? (audio recording), ?? (according to the law), ?? (truth)
10	Study	?? (school), ?? (children), ?? (start school), ?? (student), ?? (start school), ?? (university), ?? (major), ?? (parents), ?? (homework), ?? (college entrance examination)
11	Celebrity and charity	??? (Xukun Cai), ?? (Zi Yang), ?? (charity), ?? (super topic), ?? (fan), ?? (dawn), ?? (charity), ?? (Zhan Xiao), ?? (protect), ?? (defeat)
12	People	?? (live stream), ?? (son), ?? (younger brother), ?? (grandmother), ?? (vocation), ?? (husband), ?? (idol), ?? (value), ?? (boost popularity), ?? (meet)

It can be seen from Figure 6 that though there are some overlapped areas, in general, topics extracted are evenly distributed. From Table 5, it can be inferred that users on Weibo discuss a wide range of topics, among which 10 topics are related to the COVID-19 outbreak while 2 topics on “Law” and “People,” seem irrelevant to the research context. Seven topics, including “Fight the virus together,” “Knowledge,” “Assistance,” “Prevention,” “Treatment,” “Global pandemic,” and “Stay at home” are directly related to the disease outbreak, while three topics, “Economics,” “Study,” and “Celebrity and charity” could be regarded as topics derived from COVID-19, because they are either fields affected by the disease outbreak or tightly associated with it. From the results, it is interesting to know that people encourage each other in terms of fighting the disease, share disease-related knowledge such as prevention and transmission, pay attention to global disease outbreak development, and also discuss the affected life together, which are consistent with some findings of previous research (Chung, He, & Zeng, 2015; Corley, Cook, Mikler, & Singh, 2010; Lazard, Scheinfeld, Bernhardt, Wilcox, & Suran, 2015; Signorini, 2014).

Sentiment Tagging Results and Trend Analysis

After data preprocessing and text segmentation, 207,323 posts in total remained for sentiment tagging, with 79,861 positive posts and 33,049 negative posts, while the rest were all tagged as neutral posts (sentiment score is 0). Figure 7 shows the trends of the number of positive and negative posts over the data collection period. It can be observed that the number of positive posts exceed that of the negative ones over the whole period.

Figure 7

Number of positive and negative posts over time

Specifically, in terms of positive posts, three peaks can be observed, as also marked in Figure 7. The top 10 most frequent terms extracted from posts at each peak are listed in Table 6 . The first peak is recorded on January 24, the day of the Chinese New Year's Eve, and right after the lockdown of Wuhan, which took place on January 23. From the top 10 terms, it can be inferred that the posts are mostly related to wishes for the coming New Year as well as the COVID-19 outbreak in Wuhan. The second peak comes on March 3, when the central government announced the preliminary success in fighting COVID-19, for which relevant words, for example, ?? (protect), can also be observed. The third peak comes on May 23, when the success of the phase 1 vaccine trial was announced.

Table 6

Top 10 Frequent Terms Extracted from Posts at Each Peak

No.	Peak 1			Peak 2			Peak 3
	Top term	Translation	Freq.	Top term	Translation	Freq.	Top term	Translation	Freq.
1	??	Hope	470	??	Protect	767	??	Disease outbreak	533
2	??	Wuhan	388	??	Disease outbreak	374	??	Case	395
3	??	Disease outbreak	320	??	Hope	364	??	COVID-19	287
4	??	Pneumonia	260	??	Defeat	351	??	Coronavirus	263
5	??	Add oil	235	??	This	350	??	Pneumonia	193
6	??	Novel	204	??	Fight the virus	344	??	Confirmed Affection	189
7	??	Safe and sound	179	???	Fight the virus	313	??	China	185
8	????	Coronavirus	173	??	Information	306	??	Vaccine	173
9	??	Virus	157	???	Good kids	288	??	Accumulate	148
10	??	1 year	154	????	Eliminate the false and retain the true	263	??	Hope	141

Semantic Network Analysis

Figure 8, Figure 9 record the visualization of semantic networks of the positive and negative sentiments, respectively. In the network visualization, the color of each modularity is different from one another, and the size of a node stands for its interaction frequency: the more interaction it has, the bigger size of node it is. In both semantic networks, it could be easily observed that country media, including ???? (People's Daily), ???? (CCTV News), ???? (Global Times), and so on are leading the discussions, and basically, each of them forms a relatively independent community. As for the differences, it could be seen that there are more medium-sized nodes around each leading actor in the semantic network of positive sentiment, which means that mainstream media and influential KOL (key opinion leader), including entrepreneurs such as ?????????(Jack Ma), ??? (Hu Xijin), self-media such as ????? (Things in the UK) and so on, play an important role in leading the information spread and discussions of positive sentiment, while in the semantic network of negative sentiment, the discussions, also led by country media, are more scattered.

Figure 8

Visualization of semantic network of positive sentiment

Figure 9

Visualization of semantic network of negative sentiment

To take a closer look, the reposting frequencies of each node are calculated for each semantic network and then normalized between 0 and 1 for comparison purposes. Quartiles of normalized data are shown in Table 7 , from which it can be seen that Q1, median, and Q3 of the reposting frequencies of the positive semantic network are all greater than those of the negative semantic network, meaning that there are more influential nodes in the semantic network of positive sentiment, which are consistent with the previous findings.

Table 7

Quartiles of Normalized Reposting Frequencies of Each Semantic Network

Sentiment	Min	Q1	Median	Q3	Max
Positive	0.000	0.002	0.013	0.109	1.000
Negative	0.000	0.000	0.000	0.002	1.000

Conclusion and Future Work

Weibo serves as a social media platform for people in mainland China to share information and communicate with each other. With the help of 719,570 collected posts and application of the LDA model, a wide range of topics discussed in relation to COVID-19 on Weibo is discovered. In response to the COVID-19 outbreak, people gain knowledge about COVID-19, show their support for frontline warriors, encourage each other spiritually, and, in terms of taking preventive measures, express concerns about economic and life restoration, and so on. Sentiment analysis further reveals that country media are leading the discussions on Weibo in both semantic networks, while, specifically, country media, as well as influential individuals and “self-media” together contribute to the information spread of positive sentiment, indicating that the government could better fulfill its role as crisis communicator through the utilization of such kind of media network.

Although there have been studies of public opinion on COVID-19 using surveys, scant studies focus on COVID-19 opinion mining based on social media data. With LDA's excellent performance in topics modeling and sentiment analysis to take a closer look at people's feelings, this work contributes to the understanding of peoples' response to COVID-19 on Weibo and may probably serve as an example of preliminary research in the application of LDA and sentiment analysis on the COVID-19 social media dataset. Further investigation of this topic can be done in different ways. One direct way is to extend the research context, such as tracing relevant posts of each topic, analyzing peaks of negative posts, revealing the relationships between positive and negative trends, and so on. Besides, identification of topic trends, correlation analysis between the number of posts and disease development trends, and so on can also lead to meaningful findings, for which a similar approach can be seen in some previous research (Fung et al., 2013; Hu et al., 2017).