Introduction

Following the emergence of COVID-19, many newspapers from different countries published photos of barren supermarket shelves, underscoring the shortage of food and essentials (Lufkin, 2020; Nicholson, 2020). These individuals usually present this phenomenon as ‘panic buying’ (Nicholson, 2020). The social and psychological reactions of the general public to new outbreaks of infectious diseases, such as SARS, the H2N1 pandemic, and the Ebola virus, often provoke public sentiments such as fear, anxiety, and depression. (Maunder et al., 2003; Sim et al., 2010). In fact, research has shown that only a small number of people purchase a large amount of goods (for example, only 3% of people purchase an excessive amount of pasta (Kantar, 2020)). Such data suggest that while many people anticipate and act on potential shortages during crises, only a fraction engage in extreme purchasing, which is different from compulsive purchasing disorders (Sharma et al., 2020). However, after this purchasing behavior occurs, the resulting social problems are extremely serious, leading to supply chain collapses and market disarray. At this point, expedient government intervention is extremely important. The purpose of this paper is to explore how the government can swiftly intervene, thereby mitigating social impact after such panic buying behavior occurs.

At present, several scholars have found that government intervention plays a very important role in disease prevention and control (Zhao et al., 2020). However, there is a noticeable gap in the literature in regard to addressing panic buying during the COVID-19 outbreak, especially during its early stages. Such an investigation holds significant academic value, given that swift and appropriate government actions can effectively reduce its negative impact on society. Therefore, it is highly important to study the influence of government intervention on panic buying during the early outbreak of COVID-19. For example, Li and Dong (2022) developed a game theoretic supply chain model to assess the impact of government regulation on the shortage of life-saving materials and profits within the supply chain (Liu et al. 2022). Olowookere et al. (2022) emphasized that the government should comprehensively help people overcome the difficulties resulting from epidemics, particularly vulnerable populations. Cariappa et al. (2022) proposed a fundamental panic buying intervention, i.e., starting from agriculture to build public confidence. Taylor (2022) summarized the intervention experience. An examination of current intervention research reveals multifaceted analytical perspectives. However, many of these approaches lack the immediacy required for swift responses, rendering them more suitable for post crisis management rather than urgent interventions. In addition, public opinion cannot be accurately reflected, and most related research methods rely heavily on model simulation (Rajkumar and Arafat, 2021); thus, actual public data are lacking. Therefore, it is necessary to conduct mining and analysis based on netizens’ online comments on government measures to assess the effectiveness of implementing various government intervention measures. By identifying the most impactful factors on intervention outcomes, this approach aims to shape more effective strategies for future incidents.

Given this background, the cardinal objective of this research is to evaluate the impact of government intervention measures on panic buying during the COVID-19 pandemic by mining and analyzing online comments dispersed across cyberspace. In addition to merely assessing the tangible outcomes of these interventions, this study aims to determine the actual effectiveness of government intervention measures, gain deeper insights into public perceptions and attitudes toward these measures, analyze the emotional trends of the public under different government interventions, and provide valuable insights for formulating more effective strategies in the future. The research methodology will involve semantic network analysis, sentiment analysis, and LDA modeling for categorization and exploration of implementation effects. Additionally, a multiple regression model will be used to analyze the factors influencing the intervention effects. The findings of this study will contribute to enhancing intervention efficiency, mitigating the negative consequences of panic buying, and promoting the normalization of market order.

Materials and methods

Literature materials

In Google Scholar’s academic research, our document retrieval was carried out with “panic buying” and “intervention” as the main fields, yielding a wealth of studies addressing various facets of panic buying, encompassing its reasons for formation, prevention methods and functional evaluation. Many studies have analyzed the reasons for the formation of these rocks via various methods. Prevention methods and effect evaluation are also used, but the methods are relatively simple. At present, the analysis of intervention effects has been conducted mainly by constructing regression models (Si et al., 2020) and PMC index models (Chen et al., 2021). Billore and Anisimova (2021) summarized the relevant research of the past 20 years, upon which we have also based our considerations. This section provides a comprehensive literature review, discusses the trajectory from the root causes of panic buying to preventive methods, and explores the evaluation of intervention outcomes. This approach ensures a thorough understanding of the complexities surrounding panic buying.

The research literature elucidates the genesis of panic buying by examining it predominantly through the lens of psychological triggers and the sway of external media. Psychological factors such as perceived scarcity and fear of the unknown contribute to anxiety and panic buying behavior (Yuen et al., 2020; Omar et al., 2021; Taylor, 2021). This behavior can create a cycle of increased anxiety (Prentice et al., 2022). Social influences such as norms and observational learning also play a role in amplifying perceptions of scarcity and triggering panic buying (Yuen et al., 2022). External media, particularly social media, also significantly influences panic buying. Expert opinions and official communications during public health crises can trigger panic buying (Naeem, 2021). Excessive exposure to information on social media intensifies perceived scarcity and purchase intentions (Islam et al., 2021). Government and corporate interventions can mitigate panic buying while influencing social dynamics (Prentice et al., 2021). Research methods include correlation analysis (Lins and Aquino, 2020; Bentall et al., 2021), qualitative analysis (Taylor, 2021), and statistical simulations (Fu et al., 2021) to explore the causes and dynamics of panic buying during crises. Communication patterns during disasters also impact panic buying behaviors (Arafat et al., 2022). De Brito Junior et al. (2023) found some results that may assist policymakers in introducing public policies and managing resources during a crisis that requires social distancing and lockdowns. In general, the examination of panic buying’s root causes is characterized by diverse methodologies and well-substantiated conclusions, suggesting a mature body of research in this domain.

The intervention strategies outlined in the literature can be broadly categorized into three types: psychological, market-based, and network monitoring. Lei et al. (2020) used the SAS and SDS to increase anxiety and depression rates in affected populations, stressing the need for government initiatives in economic aid, medical support, and psychological intervention. Bermes (2021) used structural equation modeling to suggest improving consumer resilience and modifying consumers’ information environment. Ho et al. (2020) emphasized psychological support, highlighting the vulnerability of people exposed to epidemics. Mukhtar (2020) reviewed past epidemics to develop crisis intervention plans. Roy et al. (2020) and Zheng et al. (2020) stressed mental health care and psychological panic reduction, respectively. Wu (2009) focused on network emergency management. Tsao et al. (2019) advocated retail strategy changes to ease market pressure. Ling et al. (2020) and Sahin et al. (2020) proposed social and economic system adjustments. Boyacι-Gündüz et al. (2021) underscored food system resilience amid population growth. This review reveals diverse intervention approaches, highlighting gaps in multidimensional studies combining psychology and market dynamics.

With regard to the assessment of the outcomes of diverse interventions for panic buying, several studies have examined this topic further. Prentice et al. (2020) noted that countries typically employ regular interventions during such events. Using Twitter data for Australia, they found that government measures aligned with panic buying periods through semantic analysis. Arafat et al. (2021a) discussed historical perspectives and crisis prevention planning, integrating sociology, marketing, and industrial purchasing. Prentice et al. (2021) studied the impacts of government, business, and social groups on panic buying, highlighting the roles of government and business over social groups. Mao et al. (2022) developed a dynamic game model showing that government interventions could control panic buying duration. Rajkumar (2021) proposed a biological-psychological-social model, suggesting measured punitive actions, responsible media reporting, and social contact to mitigate panic buying. Niu et al. (2021) emphasized targeted interventions based on survey data. Fast (2014) used network analysis to predict social responses to disease outbreaks. In a series of papers (2020a; 2020b; 2020c), Arafat et al. analyzed media reports on panic buying. However, these studies did not adequately capture the public’s reactions to these events. In fact, public comments imply a great deal of valuable information. In addition, Arafat et al. (2021b) introduced panic buying intervention measures in their book, but most of them used summary words and did not analyze changes in public sentiment. The above literature analysis shows that existing studies mostly analyze structural equations, questionnaires and mathematical modeling but rarely mine information from people’s online comments.

In addition, empirical investigations into panic buying in China have produced insightful findings. Prentice et al. (2021) and Islam et al. (2021) partially used samples from China in their analysis. However, Wang and Na (2020) applied a multivariate statistical model to study rational and irrational motives for food hoarding by aggregating online samples from three Chinese cities. The results confirmed the occurrence of rational and irrational food hoarding. Fu et al. (2022) utilized mathematical modeling techniques to analyze the efficacy of panic buying interventions and validated the model using Chinese panic buying data. In addition, Yang et al. (2022) conducted a survey of 517 participants who experienced panic buying during the Omicron pandemic in China. Their findings revealed connections between media exposure, perceived emotional risk, stakeholder perception, protective awareness, and panic buying behavior. Research on panic buying in China has yielded promising results, but in-depth research is still strongly needed.

In summary, structural questionnaires and structural equation methods may inadvertently overlook much valid information. In contrast, online comments are relatively free and can multidimensionally and immediately capture emotional changes and actual needs. Therefore, more effective information can be obtained by mining people’s online comments on government interventions. As such, this paper carries out data analysis by crawling online comments under different government interventions and further uses the LDA topic extraction model to determine the effects of the intervention, which has strong practical significance. The LDA review by Jelodar et al. (2019) covers research on LDA from 2003 to 2016, revealing its application across various fields such as software engineering, political science, medicine, and linguistics. Moreover, it has been studied in fields such as communication research (Maier et al., 2021) and artificial intelligence (Yu and Xiang, 2023), highlighting its broad applicability.

Methods

LDA (latent Dirichlet allocation) model

The LDA model, introduced by (Blei et al., 2003), addresses certain limitations inherent in traditional text analysis mechanisms. While classical methods, such as TF-IDF, gauge the correlation between two documents by counting shared words and employing metrics such as term frequency (TF) and term frequency-inverse document frequency (TF-IDF), they often overlook deeper semantic connections. Such methods merely scratch the surface, focusing on word frequency without diving into the underlying themes and associations. Therefore, the LDA method can better determine the relationships among comments. This section analyses the effect of government intervention measures on preventing panic buying through the use of the powerful topic extraction function of LDA.

The similarity adaptive method is used to find the optimal LDA topic number; this method does not require manual debugging of the topic number and has a small number of iterations, high operational efficiency, fast speed, and high accuracy. The specific operation steps are as follows.

  1. (1)

    Randomly select the initial topic number K to obtain the initial model and calculate the similarity between the topics, i.e., the average cosine value cosθ (i represents the dimension, K represents the number of topics). The specific calculation formula is as follows:

    $$\cos \theta =\frac{{\sum }_{i=1}^{n}{X}_{i}\cdot {Y}_{i}}{{\sum }_{i=1}^{n}{\left({X}_{i}\right)}^{2}\cdot {\sum }_{i=1}^{n}{\left({Y}_{i}\right)}^{2}}$$
    (1)

    where θ represents the angle between X and Y, and X and Y represent two n-dimensional vectors, i.e., X is represented by \(({X}_{1},\,{X}_{2,}\,\ldots {X}_{n})\), and Y is represented by \(({Y}_{1},{Y}_{2,}\ldots {Y}_{n})\). A larger cosine value indicates that the texts are more similar and are grouped into the same category during topic extraction.

  2. (2)

    The model is trained again by increasing or decreasing the value of K, and the average cosine between subjects is calculated.

  3. (3)

    Perform step (2) again, and the cycle is repeated until the optimal K value is obtained, i.e., termination of the iteration. At this time, K is the optimal number of topics extracted.

Text sentiment analysis methods

Semantic networks primarily analyze the relationships between sentences. By plotting the relationships between evaluation targets and their respective opinions, they aid in visually analyzing the attributes among evaluation targets. ROSTCM5.8 software is used to generate semantic network graphs for the four intervention categories. By constructing semantic networks, potential connections and hidden information between evaluation targets and evaluation opinions can be uncovered; these are primarily represented through directed edges and nodes. Edges represent connections between nodes, while nodes represent individuals or events.

Sentiment analysis involves analyzing the sentiment of each sentence. A paragraph of text is input, processed with Python code, and automatically received feedback on the emotional orientation of the text, along with a score indicating whether it is a positive endorsement or a negative critique. This magical functionality is known as text sentiment analysis, also referred to as opinion mining. It involves the collection, processing, analysis, summarization, and inference of subjective texts with emotional tones, spanning multiple research fields such as artificial intelligence, machine learning, data mining, and natural language processing. Text sentiment analysis plays a crucial role in today’s information industry era: in sentiment analysis, it dissects hot events emotionally, identifies emotional reasons, aids governments in understanding public sentiment, and prevents harmful events from occurring.

Multiple regression analysis

Toward the end of the article, we employed multiple regression analysis to further explore the factors influencing people’s emotional tendencies. Multiple regression is a statistical analysis method used to study the relationship between a dependent variable and multiple independent variables. This approach helps us understand the impact of multiple independent variables on the dependent variable and quantifies the magnitude of these impacts. In multiple regression, we initially assume a linear relationship between the dependent variable and the independent variables and establish a mathematical model to describe this relationship. Then, we conduct further analysis using multiple regression to explore this relationship further.

Background material

The intervention of government departments can alleviate public panic and reduce adverse social impacts. However, there is a lack of action in sorting out intervention measures and evaluating effects, especially when the sorting of government intervention plans during the initial outbreak of the epidemic in Hubei Province is limited. Therefore, this section aims to delve into the backdrop of panic buying, examine official intervention mandates, categorize the government’s countermeasures against panic buying, and subsequently assess their effectiveness.

Event Background

At the onset of the epidemic, the government lacked corresponding management experience, resulting in various practical issues. By retrospectively analyzing the initial control plans and their shortcomings, our study can glean valuable lessons to enhance future management strategies and improve overall responsiveness. Therefore, this paper takes panic buying in Hubei, China, during the initial period of the epidemic as an example to evaluate the effectiveness of the intervention measures. First, we meticulously chart the progression of the epidemic in Hubei Province, aligning it with a chronological framework. Key milestones and pivotal events are itemized, providing a clear and sequential depiction of the unfolding situation. The trajectory of event progression is shown in Fig. 1, which presents a flow chart capturing the sequence and development of events during the epidemic in the province.

Fig. 1
Fig. 1
Full size image

Epidemic development timeline in Hubei Province.

Among them, “Shuanghuanglian” is a traditional Chinese patent medicine and a simple preprotection that was once said to inhibit COVID-19. The “One yuan dish” was a pound dish that was launched during the epidemic in Wuhan to stabilize the market and the public.

As illustrated in Fig. 1, the epidemic in Hubei Province arose rapidly, prompting immediate and decisive actions from the government. Notably, the interval between the discovery of human-to-human transmission and the “closure of the city” was less than 3 days. Lockdown measures effectively stemmed the large-scale spread of the epidemic. However, in its aftermath, the public swiftly reacted with a spree of panic buying at supermarkets. In response, the government sought to assure the public with announcements regarding ample supplies while also enforcing market regulations. However, government intervention has a notable lag. Panic buying had already taken root by the time the city was closed, and there was a delay before official assurances regarding supply sufficiency were broadcasted. Such delays had lasting consequences. Ideally, news about sufficient supplies should be synchronized with, if not preempted, actions such as city closures to better manage public panic and reduce market strain.

Sorting out intervention measures for panic buying

Google’s search function is powerful and can directly show changes in item demand. However, this article analyzes drug demand based on Weibo data for three reasons: First, Weibo is a platform based on user relationships, used for information sharing, dissemination, and acquisition. The earliest and most famous microblogging platform is Twitter from the United States. According to official reports, as of the end of 2023, the daily active users of Weibo have reached 260 million, and Weibo has become an important channel for the public to express their opinions. The government also releases reports on relevant measures through Weibo. For drugs that are popular or have regional characteristics in China, the emotional mining of its comments can more accurately reflect changes in demand. Secondly, Weibo online comments contain rich and detailed personal experiences and emotional expressions of users, such as feelings, effects, and side effects of using drugs, providing more in-depth information for emotional mining. Google Trends data, on the other hand, is more macroscopic and general. Finally, the text of Weibo comments has strong social interactivity. The replies and discussions among users can form the collision and dissemination of viewpoints, revealing the deep-seated reasons and potential influencing factors for drug demand. Google Trends is based on search data and lacks the rich information brought by this social interaction. In addition, compared with structured data such as questionnaires and interviews, online comment data are more spontaneous and expansive, and their data volume is large, often dozens, hundreds, or even thousands of times that of the questionnaire, which offers a richer mine of actionable insights. There are many fields in which online review data are used for data mining to optimize management plans, such as through the use of online reviews to evaluate hotel management (Guo et al., 2022) and commodity reviews to optimize product functions (Song et al., 2021, Lan et al., 2020); these fields have been proven to be able to draw effective conclusions and promote industrial development.

Therefore, to better understand public sentiment toward government interventions, this study uses “panic buying”, “snap purchase”, “masks”, “materials” and other search keywords from Weibo to screen out related intervention reports on the CCTV News, People’s Daily Online, People’s Daily and other official media.

Panic buying possesses several distinct characteristics. For example, it is typically triggered by factors like concerns regarding the shortage of supplies, panic over ineffective market regulation, and unease resulting from the absence of authoritative information. Research has found that adequate supplies have a major impact on panic buying (Fu et al., 2022; Lins and Aquino, 2020). During the COVID-19 pandemic, local governments disclosed supplies of materials like grains, vegetables, and masks. For instance, a city said its grain reserves could last over half a year. This avoided panic buying and stabilized the market. The authority effect also influences panic buying (Zhang et al., 2020; Naeem, 2021). When there were rumors, experts were invited to explain. For example, for the rumor of a drug treating COVID-19, experts clarified, eliminating doubts. The government’s regulatory measures can curb panic buying. Keane and Neal (2021) said they are crucial for stability. For comparison, there’s initiative guidance (Mao et al., 2022), a non-mandatory measure, guiding the public through information.

We divided the Weibo topics based on these characteristics and relevant research. We initially conducted keyword searches such as “authority effect”, “adequate supplies”, “market regulation”, and “active guidance” to filter relevant Weibo topics in Fig. 2. For instance, in the market regulation category, we specifically searched for Weibo topics related to “panic buying” and “market regulation”, and manually selected those with a significant amount of data for analysis. The specific categories and classification descriptions are shown in Table 1.

Fig. 2
Fig. 2
Full size image

Proportion of each category.

Table 1 Categories and classification descriptions.
Full size table

By selecting representative Weibo topics with relatively hot discussions among the four categories, a total of 14 topics were screened out, and Python was used to crawl the corresponding public comment data to further analyze their perceptions. A total of 84,534 comments were crawled. The Weibo topics and comment volumes are summarized in Table 2, where the Weibo topics are marked with “#”.

Table 2 Weibo topics and comments volume.
Full size table

Results

Utilizing data scraped from official Weibo comments allows us to understand in a timely manner the acceptance and emotional preferences of netizens toward governmental announcements, thereby enabling us to better grasp public opinion and determine the following work. In this section, the crawled data are cleaned, and semantic network analysis and sentiment analysis are used to further explore the netizens’ perceptions of government intervention measures. The semantic network can clearly show the connection between the topic and the subject, which is helpful for us to observe the basic information of the comment. Emotional analysis delves deeper, unearthing underlying emotional biases in the comments, aiding us in assessing the effectiveness and reception of the related governmental texts.

Selecting and cleaning comments

Given the vast internet penetration and extensive user base, netizens voice diverse opinions while maintaining a semblance of moral integrity, resulting in intricate and multifaceted feedback. To best harness these data, we employ Python for preliminary processing, ensuring that only valid comments are retained. This approach allows the subsequent analysis results to more closely reflect the real situation. The data cleansing rule in this section is to delete invalid content on the basis of retaining comment information to the greatest extent. The main steps are as follows:

  1. (1)

    Nonessential comments, such as punctuation, place names, personal names, modal particles, and advertisements, are eliminated.

  2. (2)

    Comments such as “xswl”, “hhh” and other character-based comments that cannot be accurately identified during sentiment analysis are eliminated in an abbreviated manner.

  3. (3)

    Use regular expressions to remove invalid comment content, including URLs, links, and “reply to @XXX”, from a comment, and retain the remaining valid content.

  4. (4)

    Translate the English expressions into Chinese.

After cleaning, a total of 80,280 comments remained, among which the proportion of each category is shown in Fig. 2.

As depicted in Fig. 2, the public is more enthusiastic about material news, with comments accounting for 44.27%, while the proportion of the authority effect is lower. The data were collected mainly because, regarding panic buying, authoritative figures, such as Zhong Nanshan, tend to focus more on the epidemic than on the social problems caused by the epidemic. (Zhong Nanshan is a respiratory disease expert and a key figure leading the fight against the COVID-19 pandemic.)

Semantic analysis based on online comments

The abovementioned cleaned comment data can be visually analyzed to obtain additional essential information.

Semantic network analysis of initiative guidance

Semantic network analysis is performed on the Weibo comments related to initiative guidance, and the results are shown in Fig. 3. The statistics of the top 30 words and their frequencies are calculated to validate the results of the Semantic network analysis, as shown in Table 3.

Fig. 3
Fig. 3
Full size image

Semantic network of initiative guidance.

Table 3 Frequency table of calling for guiding words.
Full size table

As shown in Fig. 3, people’s attention was primarily focused on face masks. Despite government opposition to hoarding, the term “face mask” appeared to be staggered 4281 times, leading to a spike in prices. Both offline and online pharmacies experienced shortages of face masks. Additionally, “hoarding” appeared 781 times and “shortage” 676 times. People’s emotions were fragile due to the pandemic, making them prone to overinterpreting official statements. For instance, a Weibo repost by the People’s Daily claimed that Shuanghuanglian could suppress the novel coronavirus, triggering panic buying of this traditional Chinese medicine. Although officials later clarified that Shuanghuanglian is not a treatment and urged people to stop panic buying, the impact was limited, with cases reporting worsened conditions due to self-administration. Overall, government guidance was ineffective, and there were instances of careless communication. Officials should exercise greater caution to avoid unnecessary misunderstandings during crises. It is noteworthy that toilet paper appeared approximately 200 times, which aligns with reality. Research (Garbe et al., 2020) suggests that toilet paper can provide a sense of security, highlighting its unique prominence in panic-buying events triggered by this pandemic. This calls for reflection.

Semantic network analysis of market regulation

The comment data of market regulation are integrated to generate a semantic network, as shown in Fig. 4, and the frequencies of the words are shown in Table 4.

Fig. 4
Fig. 4
Full size image

Semantic network of market regulation.

Table 4 Statistics of high-frequency words related to market regulation.
Full size table

Figure 4 highlights a significant concern during the epidemic: rising prices, especially for face masks. The analysis revealed that masks are becoming more expensive and of lower quality. Public discourse centers mainly around mask issues, with “masks” mentioned 6058 times. Terms such as “secondary”, “second-hand”, “fake”, and “black heart” indicate the presence of recycled or counterfeit masks, with “second-hand” appearing 342 times. Reusing masks reduces their effectiveness and increases the risk of cross-infection, posing significant harm and negative social impacts. Additionally, terms such as “severe punishment”, “calling”, and “deserving it” show public support for government actions and active participation in reporting price gouging. Severe punishment was mentioned 380 times, and deserved punishment was mentioned 372 times, reflecting strong public sentiment. Moreover, alongside rising mask prices, living essentials also see price hikes. The term “price increase” was mentioned 1205 times, indicating widespread concern. In conclusion, the analysis combines word frequency and semantic networks to highlight robust public support and participation in regulatory measures. However, it also underscores significant regulatory loopholes, suggesting that regulations should not only focus on pricing but also consider public sentiment.

Semantic network analysis of the authority effect

The comment data of the authority effect are integrated to generate a semantic network, as shown in Fig. 5, and the frequencies of the words are shown in Table 5. (Li Lanjuan is an expert in infectious diseases and is one of the spokespersons representing the fight against the epidemic.)

Fig. 5
Fig. 5
Full size image

Semantic network of the authority effect.

Table 5 Statistical Table of word frequency of authoritative effects.
Full size table

Figure 5 shows that when a person’s prestige is particularly high, the influence of his or her speech is greater than that of the general populace, and it is easier for him or her to gain trust. Therefore, rational use of the authority effect can effectively improve work efficiency. During the epidemic, Zhong Nanshan, Li Lanjuan, and other authoritative figures drew the public’s attention. The aged academicians Zhong Nanshan and Li Lanjuan who fought on the front line of the epidemic are knowledgeable and graceful, enjoying high public prestige. When officials release their speeches, the public is more willing to believe and obey. From the word frequencies in Table 4, “Zhong Nanshan” and “Li Lanjuan” appeared 703 and 317 times, respectively, of which “Zhong Nanshan” ranked first, which shows the public’s attention given to public figures. It is not difficult for Semantic Networks that have experienced authority effects to find that people are more positive, including describing them as “cute”, which has appeared 451 times, hoping that they can do a good job in “protection”, which has appeared 378 times, and expressing understanding of their behavior. Most of the comments are positive and express their desire to return home from Hubei Province early. Overall, the effect of authority has a significant effect.

Semantic network analysis of sufficient materials

Figure 6 illustrates a semantic network derived from integrated comment data, with corresponding word frequencies detailed in Table 6. During the epidemic, the assurance of adequate supplies provided significant comfort to affected populations. The analysis highlighted essential food items such as “cabbage”, “rice”, “beef and mutton”, and “potato”, which were frequently mentioned with 501, 332, 434, and 329 instances, respectively. Additionally, phrases such as “thank you” (461 mentions), “refueling” (321 mentions), and “waste” (308 mentions) were prominent. These findings indicate gratitude for material sources and concerns about food waste. People not only appreciate access to essential supplies but also demonstrate conscientiousness toward resource conservation. This mutual support fosters national cohesion. Overall, the impact of adequate supplies during the epidemic is evident, emphasizing positive sentiments without derogatory connotations.

Fig. 6
Fig. 6
Full size image

Semantic network of sufficient materials.

Table 6 Frequency Table of materials adequacy category words.
Full size table

Sentiment analysis based on online comments

Sentiment analysis can help netizens intuitively grasp their attitudes toward government intervention. The data collected in this paper are obtained from intervention plans for panic buying events, mainly in Hubei Province, the center of the epidemic, at the beginning of the COVID-19 pandemic. Comments belonging to Hubei were analyzed separately to observe the actual relationship between the effect of government intervention and the severity of the epidemic in the intervention area.

Data preprocessing

First, Python is used to reprocess the cleaned 80,280 comments, and the Jieba word segmentation package is adopted to segment the Weibo comments. To avoid redundant words and retain effective content, stopped words, including “on the other hand”, “here” and other conjunctions without actual emotional meaning, are removed, increasing the accuracy of the results of the sentiment analysis. For a bird’s-eye view comparison, out of an original cache of 1878 comments emanating from Hubei, a streamlined set of 1301 comments emerges postpurification.

Subsequently, the geographical origin of user comments becomes the focal point of our analysis, resulting in the illustrative Fig. 7, in which the circle represents the user’s region. A darker color indicates more corresponding people. This indicates that, except for western regions such as Qinghai, Tibet, and Ningxia, comment volume is relatively harmonized across the vast expansion of other regions. Thus, even when narrowed down to a specific 1301 comments from Hubei, the analytical value of these comments remains undiminished and profoundly significant.

Fig. 7
Fig. 7
Full size image

User distribution (dark color represents more users).

Sentiment analysis

Sentiment analysis in Python involves two dictionaries-a sentiment word dictionary and a degree word dictionary-primarily based on the HowNet Chinese sentiment dictionary. The sentiment dictionary is divided into positive and negative emotion words, while the degree word dictionary categorizes words such as “most”, “very”, “more”, “ish”, “insufficiently”, and “inverse”, each assigned specific weights for degree distinctions (Li et al., 2015). For instance, “most” carries a weight of 2, “very” is weighted at 1.5, and intriguingly, “inverse” is marked with −1.

The results of the sentiment analysis are summarized separately for non-Hubei areas in Fig. 8 and for the Hubei region in Fig. 9. Positive sentiment is notably highest in the sufficient materials category, constituting 47.55% of favorable reactions. Conversely, initiative guidance shows the lowest positive sentiment at 25.26%. Conversely, initiative guidance evokes the highest negative sentiment at 38.81%, whereas sufficient materials record the lowest negative sentiment at 19.53%. This pattern indicates a public preference for sufficient material interventions over initiative guidance. In terms of neutral emotions, initiative guidance is the most common, suggesting that it polarizes public sentiment, while the authority effect elicits the least neutral response. Notably, negative emotions under initiative guidance significantly outweigh positive emotions, contrasting with better public perceptions in the other intervention categories.

Fig. 8
Fig. 8
Full size image

The sentiment analysis results for the non-Hubei areas.

Fig. 9
Fig. 9
Full size image

The sentiment analysis results for the Hubei region.

The average sentiment values further clarify the dataset’s collective sentiment. The sufficient materials category scores highest for positive sentiment, with a notable score of 2.3. Conversely, the authority effect and market regulation categories score lowest for negative sentiment, both at −2.2, indicating stronger negative public sentiment despite positive aspects. This discrepancy may stem from a perceived lack of market regulation or overly technical authoritative statements.

In sum, the sufficient materials category clearly garners the most public appreciation among the interventions. When these interventions are ranked based on positive public sentiment, the hierarchy is as follows: sufficient materials > authority effect > market regulation > initiative guidance.

As illustrated in Fig. 9, the favorability of people in Hubei, the center of the epidemic, is also sorted as follows: sufficient materials > authority effect > market regulation > initiative guidance, which is consistent with the situation outside Hubei. However, it is obvious that the percentage of positive emotions for all types of interventions is greater than that in non-Hubei areas, indicating that people in high-risk areas are more eager for government control and have greater support for it than people in low-risk areas. A total of 56.48% of the participants reported having positive emotions toward such interventions, which indicates that more than half of the participants had positive emotions toward such interventions, and these positive emotions are significant.

In terms of the average emotion score, the highest positive score is 2.6 for both sufficient materials and initiative guidance, which is 0.3 higher than the highest score for non-Hubei areas, and the negative score is significantly lower. Conversely, the negative sentiment peaks at −2.6 for Hubei, which is a decrease of 0.4 from the non-Hubei areas, because people in the center of the epidemic have more intense emotional expressions and are more likely to experience extreme emotions. Given these insights, the government should fully consider the impact of risk level on panic rush behavior when controlling and developing intervention measures in conjunction with risk level.

Analyzing government intervention effects on panic buying using the LDA model

In our preceding analysis, we delved into the fundamental connections within the comment data and discerned emotional inclinations. To gain a deeper understanding of the impact of government intervention measures on panic buying, this section adopts the LDA topic model for semantic mining, further explores the correlation between texts, extracts the topic of the four intervention categories, and further analyzes the implementation effect under each category. At the same time, relevant variables are selected, and regression models are used to explore the factors that affect people’s perceptions of government intervention. Regression models are often applied for correlation analysis among multiple factors and are more appropriate for this section.

LDA topic number optimization

The analysis divides comments into four types, extracting positive and negative emotions based on emotion scores, resulting in 8 distinct comment sets. For each category, the average cosine similarity of positive and negative emotions is calculated. Figures 10–13 illustrate the findings:

Fig. 10
Fig. 10
Full size image

Average cosine similarity change in the initiative guidance category.

Fig. 11
Fig. 11
Full size image

Average cosine similarity change in the market regulation category.

Fig. 12
Fig. 12
Full size image

Average cosine similarity change in the authority effect category.

Fig. 13
Fig. 13
Full size image

Average change in the cosine similarity for the sufficient materials category.

Figure 10 shows that for initiative guidance, setting the LDA topic number (K) to 2 achieves the lowest average cosine similarity for positive comments. For negative comments, K values of 2 or 6 yield the lowest average cosine similarity.

Figure 11 reveals that in market regulation, (K = 3) results in the lowest average cosine similarity for positive comments, while (K = 4) achieves this for negative comments.

Figure 12 indicates that for the authority effect, K = 2 yields the lowest average cosine similarity for both positive and negative comments.

Figure 13 demonstrates that in sufficient materials, (K = 3) or (K = 7) achieves the lowest average cosine similarity for positive comments, and (K = 3) for negative comments.

These insights guide the optimal selection of (K) values for each category, facilitating more nuanced topic extraction and sentiment analysis from the comments.

Topic extraction and analysis

The LDA model, also known as the three-tiered Bayesian probability framework, encompasses documents, topics, and words. The model introduces Dirichlet’s principle, which has a strong generalization ability and is not prone to overfitting.

Taking the authoritative effect class (constituting comment set D, with d representing a comment in set D, hereinafter referred to as document d) as an example, the main steps of topic extraction using LDA are as follows:

Step 1: Select the document that has been divided into words, and use the word sequence to represent \(d=({w}_{1},{w}_{2},\ldots {w}_{n})\), where w represents the word, 1,2…n represents the word sequence number, and select a document with a prior probability \(P({w}_{i})\). The Dirichlet distribution is used to create the topic distribution \({\varphi }_{d}\), which is realized by using the hyperparameter α in the distribution function (i represents the dimension, K represents the number of topics). The distribution formula used is:

$$P(\varphi {\rm{|}}\alpha )=\frac{\Gamma \left({\sum }_{i=1}^{k}{\alpha }_{i}\right)}{{\prod }_{i=1}^{k}\Gamma \left({\alpha }_{i}\right)}{\varphi }_{1}^{{\alpha }_{1}^{-1}}\cdot {\varphi }_{k}^{{\alpha }_{k}^{-1}}$$
(2)

where \(\varGamma \left(\cdot \right)\) represents the gamma function and K represents the number of topics.

Step 2: Each topic obeys a polynomial distribution, and the topic word z of document d is generated by sampling from the distribution, corresponding to the polynomial distribution of topics as follows.

$$P({\varphi }_{1},{\varphi }_{2},\ldots {\varphi }_{d}{\rm{;}}n,{p}_{1},{p}_{2},\ldots {p}_{d})=\frac{n!}{{\varphi }_{1}!\cdot {\varphi }_{2}!\cdots {\varphi }_{d}!}{p}_{1}^{{\varphi }_{1}}\cdot {p}_{2}^{{\varphi }_{2}}\cdots {p}_{k}^{{\varphi }_{d}}$$
(3)

Step 3: The word distributions for each topic are based on the Dirichlet distribution.

Step 4: Each word also obeys a polynomial distribution, and the keyword w under the topic is generated from the word polynomial distribution.

Using Python’s Gensim module, we applied LDA topic modeling to both the positive and negative comment datasets. After determining the optimal topic numbers, we conducted LDA analysis to identify recurrent words per topic. For negative comments in the “initiative guidance” category, given K-values of 2 and 6, we tested 2 and 6 topics, respectively, noting excessive word repetition at 6 topics. There were 2 topics with positive comments, as detailed in Table 7.

Table 7 Analysis of potential topics of positive and negative evaluations in the initiative guidance category.
Full size table

Positive comments under “initiative guidance” highlighted two main dimensions: (1) trust in official statements and (2) reduced public anxiety about rational cognition. Negative comments indicated poor independent thinking, judgment, and misleading official information.

For “market regulation”, 3 topics were considered positive, and 4 were considered negative. The LDA analysis results for the keywords are detailed in Table 8. Positives focused on timely government interventions (“response”, “timely”, “fast”), robust regulatory measures, and prompt departmental actions. Negatives cited inadequate regulation prone to repetition, network flaws (e.g., order cancellations), material safety concerns, and limited regulatory effectiveness. Market regulation remains a challenging, ongoing effort.

Table 8 Analysis of potential topics of positive and negative evaluations in the market regulation category.
Full size table

Utilizing LDA topic modeling, we analyzed the “authority effect” category with 2 topics each for positive and negative sentiments. Table 9 summarizes the key themes: positive sentiments emphasize trust in authority figures (“know”, “thank you”, “hard work”), and the social influence of authorities, notably “fans”. Negative comments critique those who challenge authority, spread misinformation, and highlight gaps in public understanding of disease, stressing the need for transparent and credible epidemic information.

Table 9 Analysis of potential topics of positive and negative evaluations in the authority effect category.
Full size table

In the “sufficient materials” category, after comparing K = 3 and K = 7 for positive comments, we settled on 3 topics due to lower cosine values indicating distinct topics. Table 10 reveals positive comments focusing on national unity in epidemic response, material security reducing panic, and increased patriotism. Negative sentiments highlight regional food disparities affecting disaster preparedness, excessive hoarding, and panic buying. Effective government supervision is crucial for managing hoarding and ensuring the equitable distribution of resources tailored to regional needs.

Table 10 Analysis of potential topics of positive and negative evaluations in the sufficient materials category.
Full size table

Analysis of factors affecting intervention effects based on multiple regression

Government interventions aim to create an environment where people can thrive, with public feedback guiding their refinement. Understanding public perception postimplementation is crucial for optimizing these measures. Emotions serve as a key indicator of public perception, and a regression analysis will explore factors influencing these emotions and inform intervention adjustments. Stepwise regression will help manage independent variables, ensuring that the analysis avoids collinearity issues.

Construction of the multiple regression model

The comprehensive dataset used in this section encapsulates various dimensions, from user-specific parameters such as ID; the number of followers, fans, posts and likes; contextual data such as follow-up comments; the timestamp of the original Weibo post; comment time; geographical attribution; and the core content of the comment, as illustrated in Fig. 14. An evaluation system for government interventions under COVID-19 based on the literature (Chen et al., 2020) is constructed to reflect the effects of different types of interventions. The evaluation system includes 4 first-level indicators and 8 second-level indicators. These 8 indicators are obtained by directly crawling Weibo data or by expanding crawling data. The indicators at all levels and their meanings are summarized in Table 11.

Fig. 14
Fig. 14
Full size image

Crawled comments.

Table 11 Construction of regression model evaluation index system.
Full size table

The dependent variable Y is obtained from emotion analysis to represent people’s perception of government intervention, and a multiple regression model is constructed as follows.

$$Y={b}_{0}+{b}_{1}{X}_{1}+{b}_{2}{X}_{2}\cdots +{b}_{8}{X}_{8}+\varepsilon$$
(4)

where \(\varepsilon\) represents the error term, \({b}_{0}\) represents the constant term, and \({b}_{i}\) (i = 1, 2, 3…,8) represents the respective variable coefficients.

Analysis of factors influencing the effect of intervention

Based on the regression model analyzed using SPSS 26, collinearity among independent variables was checked and found to be within acceptable limits (VIF < 10). The significance level for variable inclusion or exclusion was set at 0.1. The initial results in Table 12 indicate that variables X_1 and X_6 have p values less than 0.05, suggesting that their parameters are statistically significant and should be retained in the regression equation. The final regression results in Table 13 show that epidemic severity (X_1) alone explains 60.1% of the variance in public perception (Y), while together with follow-up comments (X_6), they explain 70% of the variance.

Table 12 Regression coefficient.
Full size table
Table 13 Summary of regression analysis results.
Full size table

The findings indicate that as epidemic severity increases, there is a stronger public inclination toward robust government intervention and a heightened demand for transparent intervention strategies. This aligns with findings by Chen et al. (2022) that during severe epidemics, public behavior such as panic buying is significantly influenced by government actions. Negative correlations imply that greater public engagement on the topic increases susceptibility to negative emotions, which diminishes gradually over time.

Certain variables, such as the location of the epidemic center, number of user followers, number of user blogs, timeliness of commenting, and number of user likes, do not significantly influence the dependent variable-the emotional value of the public. This suggests that despite various indicators of user attention and engagement, most individuals maintain their own perspectives unaffected by these factors. The location of the epidemic center also does not appear to significantly impact public emotional responses, possibly due to a perceived similarity of experience among those affected Fig. 15.

Fig. 15
Fig. 15
Full size image

Regression normalization residuals.

The regression model requires that residuals follow a normal distribution for the analysis to be valid. SPSS tests confirm this requirement, ensuring the effectiveness of the regression analysis.

Figure 16 illustrates the temporal dynamics of online public discourse, specifically focusing on fluctuations in the number of comments over time. This reveals that there is a notable increase in public engagement within the first 3 h after the release of relevant reports, peaking between the 4th and 5th hours. Subsequently, the volume of public discussions gradually declines, stabilizing after approximately 8 h. This pattern indicates that governmental interventions typically occur preemptively within 4–5 h of report release, aiming to guide public opinion and mitigate potential consequences before discussions peak.

Fig. 16
Fig. 16
Full size image

Variation in comment volume with time interval for each category.

Discussion

Result analysis

This section provides a summary of the key findings of the paper, compares them with the conclusions of the literature, and proposes relevant recommendations. The paper concludes by highlighting the main contributions of this article.

Government prevention and control lag

From the background analysis of 2.3.1, it can be seen that when sorting the timeline of the epidemic development events in Hubei Province, a clear pattern appears, that is, the measures taken by the government lag behind. After the announcement of the lockdown, panic buying has taken root and spread, and relevant news has been released to assure the public of material abundance. At that time, large-scale panic buying had already occurred and the social impact was uncontrollable.

The problem of insufficient breadth and depth of market supervision urgently needs to be solved

In the analysis of 3.2 and 3.4, we deeply perceive that there are many shortcomings to be remedied in the aspects of breadth and depth of market supervision. Due to the limitations of the categories of regulated materials and the places, thorny problems such as repeated price increases frequently occur, seriously disrupting the normal order of the market.

The in-depth research conducted by Shan and Pi (2023) on government supervision shows that when the government chooses the “active supervision” strategy, its decision-making basis mainly depends on supervision costs and government credibility, rather than simply the “amount of fines”. This finding fully demonstrates that the starting point of government management is based on rationality and always puts the people’s interests first.

Market supervision plays a crucial and indispensable role in curbing panic buying. Herbon and Kogan (2022) emphasizes that market intervention can start from subsidizing enterprises and consumers, but it is restricted by the government’s financial pressure. Relevant departments should increase management efforts, optimize the system, refine the supervision plan, effectively intervene in the market, and deal with similar problems. The core lies in that market supervision should not only focus on solving current problems but also pay attention to the establishment of long-term mechanisms to enhance the self-regulatory ability and risk-resistance ability of the market and ensure that the market can still operate stably in the face of various impacts.

The effects of the adequate supplies category and the authority effect category are more remarkable

It can be clearly seen from the results of the 3.3 sentiment analysis that among various intervention measures, the effect of the adequate supplies category is the most outstanding, followed by the authority effect category. If the two can be ingeniously integrated and implemented, that is, when authoritative figures come forward to declare the sufficiency of supplies, it undoubtedly can arouse stronger and more positive public sentiments, and thus is expected to achieve more powerful and effective intervention results.

During the difficult period when the epidemic was rampant, the influence of those prestigious authoritative public figures far exceeded that of ordinary people (Ding, 2009). Renowned figures like Zhong Nanshan and Li Lanjuan not only have highly authoritative teams as support but also, with their rich practical experience and professional and precise interpretation abilities, can easily win high recognition and full trust from the public.

This phenomenon is not limited to the epidemic itself. Even in the face of a series of social phenomena such as panic buying caused by the epidemic, releasing news about the adequacy of supplies through influential figures at this special stage is highly likely to achieve unexpectedly good effects. The underlying principle is that when the public faces uncertainties and potential risks, they tend to seek guidance and security from authorities. When authoritative figures convey a clear signal of adequate supplies, it can greatly alleviate the anxiety and unease in the public’s hearts, thereby effectively suppressing the spread of panic sentiments and the irrational purchasing behaviors caused thereby.

There was a significant positive correlation between epidemic severity and public perception

In the analysis of 3.5, the severity of the epidemic can independently explain 60.1% of the changes in public perception. The correlation between the severity of the epidemic and the changes in public perception is the strongest, indicating that the public’s perception of government intervention is strongly affected by the severity of the epidemic. The more severe the epidemic is, the more eager the public is for government control, and the greater their perception of government intervention is. Prentice et al. (2021) found that government intervention and support influenced panic-buying participation.

When the government intervenes in response to panic buying, it can start with the initial break of the epidemic. On the one hand, timely disclosure of epidemic trends can alleviate public panic caused by unknown factors; on the other hand, it can actively promote the implementation of epidemic prevention and control measures and release relevant prevention and control news. In addition, the government can appropriately intervene and guide the internet, remove rumors, cut off sources of dissemination, correct the direction of public opinion in a timely manner, and guide the public to think rationally.

Practical implications

The practical significance of this paper lies in that it provides in-depth analysis and possible strategic directions for understanding and responding to the phenomenon of panic buying.

Firstly, it helps the government formulate and adjust intervention policies more precisely. By evaluating the effects of intervention measures and understanding the public attitude, the government can promptly identify the deficiencies of existing policies and thus formulate more targeted and effective strategies to better cope with similar emergencies. For example, at the beginning of 2020, a sudden COVID-19 outbreak occurred in a certain city, leading to the public’s panic buying of masks and disinfectants. Through monitoring the market situation and collecting public opinions, the government found that the previous measure of simply appealing to the public to buy rationally was ineffective. Based on the viewpoint proposed in this article that the intervention measures of the adequate supply category are effective, the government promptly coordinated local related enterprises to increase the production of masks and disinfectants and allocated materials from other places to increase market supply. At the same time, based on the viewpoint that the intervention measures of the authority effect category are effective, medical experts were invited to explain the scientific usage methods of masks and disinfectants and the material reserves of the city to the public through TV and online live broadcasts. As a result, within a short period of time, the public’s rush-buying behavior was alleviated, the prices of masks and disinfectants in the market gradually stabilized, and the supply was sufficient.

Secondly, the revelation that the severity of the epidemic is positively correlated with public perception provides strong theoretical support for the government’s crisis communication and public opinion guidance strategies. According to the dynamic changes of the epidemic, the government needs to formulate a phased and hierarchical information release and communication plan to meet the information needs of the public at different stages of the epidemic. For example, in the early days of the COVID-19 epidemic, the public had limited understanding of the epidemic, lacked a clear judgment on the direction of the situation, and the herd mentality was obvious. According to the development of the epidemic, the government will disclose the real data of the epidemic, prevention and control measures, and the allocation of medical resources to the public in stages. During the severe phase of the outbreak, the number of new cases increases significantly. At this time, the government’s medical resources are facing great pressure, and the people also have greater psychological pressure. The government updated the relevant information of the number of cases and medical resources in a timely and accurate manner, and explained the government’s active efforts to make the public clearly aware of the severe situation of the epidemic, so as to encourage the public to more consciously comply with the epidemic prevention and control measures and actively cooperate with the government’s prevention and control work; At the stage when the epidemic is under control, the trend of reducing new cases and related policies for resuming work and production will be announced in a timely manner. For example, some restaurants are gradually opening their dining services to the extent permitted by the government, and customers are taking precautions to maintain social distancing in accordance with regulations. The clear announcement of this policy change enabled the gradual return of social and economic activities to normal under the premise of safety. Through transparent, timely and accurate information transmission, the public’s understanding of the epidemic has gradually become rational, avoiding excessive panic and blind optimism, and thus better pooling social consensus to fight the epidemic together.

Conclusions

In summary, the analysis highlights several key aspects of the government’s response to the epidemic in Hubei. First, there was a noticeable lag in the implementation of prevention and control measures, leading to panic buying and uncontrollable social impacts. Second, interventions by authoritative figures and assurances about material sufficiency had a positive effect on public sentiment, emphasizing their influential role during the crisis. Third, market regulation has shown insufficient breadth and depth, resulting in issues such as repeated price increases. Finally, there was a significant positive correlation between the severity of the epidemic and public perception of government intervention, indicating that the more severe the epidemic was, the stronger the public’s desire for government control and perception of government involvement were.

Based on the data of public comments, this paper evaluates government intervention measures and draws some conclusions. However, it is imperative to recognize certain inherent limitations, providing avenues for further exploration:

  1. (1)

    This paper uses online comment data to analyze the effect of government intervention, mainly from the public’s perspective. Future studies can pivot toward an economic lens, drawing insights from the sales metrics of prominent retail outlets before and after the rollout of government initiatives in epidemic areas. In addition, online comment data may have certain limitations, although we chose Weibo, a platform with 230 million daily active users, to obtain as many as 80,000 comment data points as possible. However, a segment of the population, particularly those less digitally inclined, remains unrepresented in the online discourse. Therefore, further optimization can be carried out in the future to solve this problem.

  2. (2)

    While this study primarily anchors its analysis to data from the early stages of the epidemic, it is crucial to understand that government measures should be dynamic and adaptable to the evolving nature of the epidemic. It is necessary to consider the characteristics of epidemic development, grasp people’s emotional situation, and formulate relevant measures. Consequently, the insights gained from this research might not seamlessly translate to the subsequent phases of epidemic management in China.

  3. (3)

    While the digital realm is often celebrated for its perceived freedom, the reality is that internet censorship is a ubiquitous phenomenon characterized by global implementation. A salient aspect of such censorship is internet filtering, which blocks some illegal or sensitive words. Worldwide renowned filtering tools such as PureSight PC, CYBERsitter, SafeEyes, and CyberPatrol have been employed worldwide for this purpose. Filtering and blocking are normally performed on ISP (Internet Service Provider) servers. This shows that although cyberspace is free, speech is not 100% open. In the context of the People’s Republic of China, ISPs are the main body responsible for internet censorship, as are most other countries. Consequently, it is acknowledged that a minority of the comments might have been expunged, rendering them inaccessible for our analysis. However, the volume of data we amassed in this study mitigates this limitation, ensuring the robustness of our findings despite such omissions. Moving forward, subsequent research endeavors could focus on quantifying this specific impact, paving the way for improving the study’s robustness and raising additional credibility to its findings.