Knowledge Trajectories Detection and Prediction of Modern Emergency Management in China Based on Topic Mining from Massive Literature Text

Abstract

China has witnessed dramatic advances in emergency management in the past two decades, while the knowledge trajectories and future trends of related research are still unclear. This study takes the published articles in China National Knowledge Infrastructure as a data sample and introduces text mining and machine learning methods, namely Latent Dirichlet Allocation combined with the Hidden Markov Model, to detect and predict the knowledge trajectories of Chinese modern emergency management research. We analyzed 5180 articles, equivalent to approximately 1,110,000 Chinese characters, from 2003 to 2021, and mined 35 latent research topics. By labeling the topics manually and analyzing the evolutionary hotspots, confusion and transition features, and transition direction and network of the topics, we explored the knowledge trajectories of emergency management research in China. By training the HMM model, we predicted the research trends in the next five years. The main conclusions are: a mapping relationship exists between the hotspots of the published articles and the main events of emergency management in China; most emergency management research topics could confuse and transfer with others in the evolution process, and seven significant paths exist in the transition network. The research topics in the following years will be more detailed and concerned with the intellectual needs of modernization.

1. Introduction

China has witnessed dramatic advances in emergency management in the past two decades [1,2]. However, the complexity and uncertainty of emergency management pose a severe threat to human society [3], and the systemic risk in “risk societies” brings colossal uncertainty to social governance [4]. For the public sector, it is of great significance to understand and respond to complex situations appropriately in order to improve the emergency management system and enhance its capacity.

Since the founding of the China, the development of the emergency management system has experienced three stages: single institution emergency management (1949–2003), emergency management as a whole system (2003–2018), and forming the emergency management system with Chinese characteristics (2018-). The development of the emergency management system is always accompanied by a series of severe crisis challenges, which engage the optimization of the emergency management system the other way around. The response to SARS in 2003 is deemed as the beginning of modern emergency management in China [5], which compelled the central government to take a series of measures to enhance the emergency system (such as issuing the Emergency Regulations for Public Health Emergencies, setting up the State Council Emergency Plan Working Group, and establishing the “One Planning plus Three Systems”). The subsequent major public emergency events, such as the Wenchuan earthquake, the violent crimes in Tibet, the Sanlu milk powder incident in 2008, the Xinjiang terrorist incident in 2009, etc., have improved the robustness of China’s emergency management system. In 2014, after the 18th National Congress of the CPC, a holistic approach to national security was put forward by President Xi to bolster national security. In 2015, the National Security Law of the China was enacted, which makes up for some of the deficiencies of “One Planning plus Three Systems” (The main content of China’s emergency management system, “one plan” refers to the emergency plan, “three systems” refers to the emergency management, mechanism, and legal systems). In 2018, the Ministry of Emergency Management of the China was established, which marked the arrival of normalized emergency management.

Emergency management research related to public emergencies or theories is another essential driving factor in developing emergency management systems. Emergency management, under the scope of public management, takes the formulation and implementation of public policies as the primary way to achieve its goals [6]. That is, it creates policies to ensure public safety by controlling risks and decreasing emergencies [7]. Eliminating all risks and emergencies is impossible because of the uncertainty and suddenness of public risks [8]. Enhancing the efficiency of policy implementation and decreasing management costs are significant issues for emergency management and the critical indicator for evaluating the emergency management capability, which heavily highlights the importance of public policy making. The punctuated equilibrium theory posits that the process of public policy is not only long-term, progressive, and stable, but also short-term, mutated, and unstable [9]. As the primary knowledge support of policy making, academic research is indispensable for emergency preparedness [10], emergency response [11], crisis communication [12,13], department collaboration [13], and hazard sequence prediction [14]. Hence, improving the related research is essential for policy optimization in emergency management; it is indispensable for revealing the knowledge trajectory of emergency management and its future evolution trend.

The rest of this paper is organized as follows. Section 2 reviews the related works. Section 3 explains the evolutionary mechanism of emergency management research in China, establishes the framework of LDA–HMM in this study, and presents the methods: text segmentation, latent topic mining, and topic evolution analysis. Section 4 presents the setup and results of the empirical analysis, including data collection, topic mining, knowledge trajectories exploration, and topic prediction. Section 5 summarizes this study’s main contributions, limitations, and conclusions.

2. Related Works

Some scholars have conducted bibliometric analyses for the research on emergency management. The bibliometric approaches arose in the 1980s, using mathematical and statistical methods to analyze the knowledge carrier quantitatively [15]. Of these, the keyword co-occurrence analysis is the most commonly used; it has been applied to identify the main keywords in emergency management [16], to classify disasters [17], to search for related publications [18], to obtain research hotspots [19] and timeline research trends [20], to draw the cluster map [21], to show the content and structure of the published literature [22], to identify the thematic communities from the research agenda [23], etc. To unveil the knowledge network of emergency management, some researchers analyzed the citation network or structure of specific areas, such as the co-citation of publications related to public health emergency preparedness [24], co-citation of the evolution of research fronts and critical documents of risk communication [25], distribution of citations over the years, citing journals of each paper on disaster information and communications technology [20], and the national and institutional cooperation and hybrid network of disaster education [26].

To draw the knowledge map more comprehensively and in more detail, some researchers employed the co-occurrence and topic or theme cluster methods to analyze the topical or thematical distribution and evolution. These are represented by the co-authorship analysis approach [24,27], co-occurrence clustering network method [24,28], terms [29] and keywords [22,30,31] co-occurrence method, thematic cluster [25] and distribution analysis method [22], bibliometric clustering algorithm [32], bibliometric clustering algorithm, etc. Some researchers analyzed the timeline [20], temporal [29], or development trends [33] of related topics. Dedicated software and online platforms include CiteSpace [31,34,35], VOSviewer [20,24,31,34,35,36,37,38,39], Bibliometrix [36], SigmaPlot [36], GraphPad Prism [31], Online Analysis Platform of Literature Metrology [34], which are the popular tool for current researchers, and most of these could visualize the knowledge structure. Databases of related publications mainly includes the Scopus database [17,20,40,41], Web of Science [21,36,38,39,42,43,44,45], PubMed [22], and some online databases [18,19,46,47]. In addition, policy documents [48] and post-graduation production [23] were also obtained as data samples.

The research mentioned above on the knowledge trajectory of emergency management has provided indispensable research paradigms and knowledge accumulation, but the inefficiencies are still noticeable. From the research content, most of these focused on keyword co-occurrence, citation structure, and cooperation network, ignoring the description of the evolution process at the macro-level and the prediction of future trends [49]. As for the data source, most were collected from the English database, and only a few studies focused on Chinese policy documents; it is difficult to believe that this is helpful for the improvement of the knowledge system, given China’s progress in emergency management. From the perspective of methodology, most of the existing research that includes a bibliometric, scientometric, or informetric analysis of emergency management is based on the application of dedicated software and online platforms; this limits the innovation of the research content as well as the methodology, especially given the lack of original text mining.

The contradiction between the rapid development of China’s emergency management and the challenges emergency management faces makes academic research more significant and indispensable. Knowledge requirements for normalized emergency management and deficiencies in the study of emergency management knowledge evolution make knowledge trajectory detection essential. Collective evolution between emergency management and related research makes the overview and prediction of the knowledge possible.

Given this, this paper aims to unveil the knowledge trajectory of China’s emergency management research by mining the related publication text in the Chinese database using text mining, and, on this basis, attempts to predict the research trend in future years using machine learning. The methodologies, Latent Dirichlet Allocation combined with Hidden Markov Model (LDA–HMM), have been applied to obtain the technical evolution trend and identifying potential R&D hotspots [50], overviewing and predicting the research themes in the field of land degradation [51] and sustainable land use [52], and extracting and analyzing the topics of public crisis management from the massive English text [15]. Compared with existing software and online platforms, such as CiteSpace and VOSviewer, LDA–HMM could mine the potential topics from massive literature texts and predict future trends with machine learning [15].

3. Methods

3.1. Stochasticity of Emergency Management Research

In China, emergency management contains four sectors: public health, production safety, public safety, and natural disasters. Emergency management aims to guide local governments’ planning, preparedness, and response options, and to increase public awareness of emergencies [53]. Practice experience shows that the results of emergency management are usually difficult to predict, which makes emergency management dynamic and stochastic [54,55]. Although it could be divided into three [56] or four stages [56,57], the status of an emergency event at a particular time in the management process is unobservable. Moreover, the following state of emergency management is only related to the current state, and not the past experienced state. Thus, the evolutionary process of emergency management can be characterized using the Hidden Markov Model (HMM) [58].

The sample domain is $\mathrm{\Omega }$; the parameter set of the time series is $T$; and the stochastic process $X(t,\omega )$ (abbreviated as $X_t$) is a binary function of $T\times \mathrm{\Omega }$ on domain $ℝ$, namely, $X(t,\omega ):T\times \mathrm{\Omega }\to ℝ$. In public emergency management, $T$ is a finite set; thus, $X(t,\omega )$ is a discrete stochastic process. Therefore, the discrete stochastic process $\left\{X_t,t\in \mathbb{N}\right\}$ ($\mathbb{N}$ is the set of all positive integers) is a Markov chain. Given that the environmental state $X^{\prime }(t,{\omega }^{\prime })$ is a discrete stochastic process, $\left\{X_t^{\prime },t\in \mathbb{N}\right\}$ is also a Markov chain. Because the state change of the environment is unobservable, emergency management can be regarded as a hidden Markov chain, and public emergency management results are implicit as well. Emergency management research has evolved in relation to previous literature [59], and the related research can be identified in different periods [35]. Given this, there is a mapping relationship between academic research and emergency management that transmits the stochasticity of emergency management to academic research (as shown in Figure 1).

3.2. Framework of LHA-HMM in This Study

We take the hidden Markov process to describe the evolution of China’s emergency management research topics and to introduce the combined method of Latent Dirichlet Allocation–Hidden Markov Model (LDA–HMM) [51], including text mining and topic clustering, to analyze the evolutionary process of the topics, as well as the confusion and transition relationships among the topics, and to try to predict the future research trends by training the model.

There are three modules for data processing and analysis. Since Chinese words consist of single characters, the element of the research topics is the word, and the academic articles have specialized keywords, the first module is to segment the text in order to establish an initial dataset. Then, the second module is latent topic labeling, which aims to extract the keywords from the initial dataset and mine the latent topics from the keywords. The third module includes the topic’s hotspot analysis, confusion and transition relationship analysis, and future trend prediction. The framework of this study is shown in Figure 2.

3.2.1. Module 1: Text Segmentation

The Viterbi algorithm [60], which solves linear and nonlinear sequence problems, is used to identify new words in the text and add them to the dictionary. This dynamic programming algorithm uses the observation sequence $O=o_1,o_2,\cdots o_X$ to excavate the maximum possible state sequence $S=s_1,s_2,\cdots s_X$. The new words, consisting of Chinese characters, can be identified through the maximum possible segmentation strategy. The main steps are:

First, the initialization of the probability distribution words is set as ${\delta }_x={\pi }_i{b}_i{o}_1$ (${\pi }_i$ is the probability of state $b_i$). Let

$$\{\begin{array}{l}{\delta }_x(j)=\underset{1\le i\le N}{{\displaystyle \max }}\left[{\delta }_x(j)a_{ij}\right]b_j{o}_x,2\le x\le X\\ {\psi }_x(j)=\underset{1\le i\le N}{{\displaystyle \arg \max }}\left[{\delta }_x(j)a_{ij}\right]b_j{o}_x,1\le x\le N\end{array}.$$

(1)

Then, take the optimal solution:

$$\{\begin{array}{l}{P}^{\ast }=\underset{1\le i\le N}{{\displaystyle \max }}\left[{\delta }_X(i)\right]\\ s_X^{\ast }=\underset{1\le i\le N}{{\displaystyle \arg \max }}\left[{\delta }_X(i)\right]\end{array}.$$

(2)

Finally, find the optimal segmentation path:

$$s_X^{\ast }={\psi }_{x+1}(s_{x+1}^{\ast }),2\le x\le X.$$

(3)

3.2.2. Module 2: Latent Topic Mining

The LDA topic model is a mixed probabilistic model based on the three-level hierarchical Bayesian model [61]. This model mines the latent lexical clusters by maximizing the probability of lexical co-occurrence in the document, and describes the document generation process with the Dirichlet Distribution. It could extract the hidden topics in the document and cluster the document according to the topic distribution by limiting the number of topics to avoid excessive parameters and overfitting problems [62]. In natural language processing, including text mining, text topic identification, text classification, and text similarity computing, this model has better reliability and has been widely used. Assuming that the topic of emergency management research obeys the hyperparameter Dirichlet prior distribution:

$$Dir({\theta }_d|\alpha )=\frac{\Gamma \left({\displaystyle \sum _{k=1}^K{\alpha }_k}\right)}{\Gamma \left({\alpha }_k\right)}\underset{k=1}{\overset{K}{{\displaystyle \Pi }}}{\theta }_{dk}^{{\alpha }_k-1}.$$

(4)

where ${\theta }_{dk}$ represents the distribution of literature $d$ on public risk management in research topic $k$. For each topic $k$, there exists topic keyword thesaurus distribution ${\varphi }_k\sim Dir(\beta )$; for every literature $d$, there exists topic keyword distribution ${\theta }_d\sim Dir(\alpha )$; and for the nth topic keyword thesaurus distribution of each item in the literature, there exists topic item $Z_{dm}\sim Multinomial({\theta }_d)$ and keyword item $w_{dn}\sim Multinomial({\varphi }_{z_{dn}})$. Thus, the LDA likelihood model in this study can be described as:

$$p(W|\alpha ,\beta )=\underset{d=1}{\overset{D}{{\displaystyle \Pi }}}{\displaystyle \int p({\theta }_d|\alpha )\underset{n=1}{\overset{N_d}{{\displaystyle \Pi }}}}\underset{z_d}{{\displaystyle \Sigma }}p(z_{dn}|{\theta }_d)p(w_{dn}|\varphi z_{dn})d{\varphi }_d.$$

(5)

The perplexity-var method is applied to calculate the optimal number of topics, which measures the structural stability through their divergence, and punishes excessive topics for minimizing the number under the premise of maximizing the difference among them [63]. Finally, we take Heinrich’s parameter estimation method to calculate the topic set, set $\alpha =50/k$, $\beta =0.1$, and use Gibbs sampling to calculate the topic set as $K=\left\{k_1,\cdots ,k_h\right\}$. The topic attribution set of each item in the literature is $D_k=\left\{j_1,\cdots ,j_n\right\}$.

3.2.3. Module 3: Topic Evolution Analysis

The HMM can describe a Markov process with hidden and unknown parameters [58]. The punctuated-equilibrium property of academic research on emergency management related to the real world and the stochastic character show that the evolution of emergency research is a hidden Markov process. By inferring the transition matrix and the initial probability distribution in this model, the confusion and transition matrices of the research topic’s evolution can be computed. From this, the evolutionary history and future trends can be presented. The main steps of HMM in this study are as follows [15]: (1) The hidden stochastic transition sequence of the topics is set as: $C=\left\{c_1,\cdots ,c_h\right\}$, where $h$ is the topic number calculated by the LDA. The hidden sequence generated by the stochastic process is noted as $G=\left\{g_1,\cdots ,g_t\right\}$, where $g_t\in C$. (2) The probability distribution of the transition of the topics is then calculated. In emergency management research, the probability of a research topic in the literature transmitting from $C_i$ to $C_j$ is $A=\left\{a_{ij}\right\}$, where $a_{ij}=P\left\{g_{t+1}=C_j|g_t=C_i\right\}$, $1\le i$, $j\le N$ and $a_{ij}\ge 0$, ${\displaystyle \sum _{j=1}^N{a}_{ij}=1}$. (3) We explore the probability distribution of the observed variable. When the state of the research topic is $C_i$, the probability distribution of the observed variable is $B=\left\{b_i(v)\right\}=\left\{f\left\{G_t=v|g_t=C_i\right\}\right\}$, where $G_t$ is the $t$-th observed random variable, and the observed sequence is denoted as $F=\left\{f_1,\cdots ,f_t\right\}$. The observation sequence in this study is the proportion of each topic over the years. (4) The probability distribution of the initial state $C_i$ of the system is ${\pi }_i$, $1\le i\le N$. Generally, the higher the number of co-occurrences of the topic keywords among topics is, the closer the correlation will be among them, and the easier migration and co-evolution will be. (5) The initial value of model training is $F=G=\left\{g_{t_1},\cdots ,g_{t_t}\right\}$. The Baum-Welch algorithm is used to estimate the parameters of the above model in order to obtain the optimal sequence, and the topics after $k$ years are calculated according to ${\widehat{F}}_{t+k}={\displaystyle \sum _{j=1}^N{A}^k(i,j)E(b_j(}v))$.

4. Empirical Analysis

4.1. Data Collection

This study takes the articles in China National Knowledge Infrastructure (CNKI, the most comprehensive Chinese academic literature database) as data samples to explore the knowledge trajectory of modern emergency management research in China. We cite the Chinese words of emergency management in CNKI and set the other search principles as Source Type = “SCI source journals” AND “EI source journals” AND “PKU Core Journals” AND “CSSCI” (Chinese Social Sciences Citation Index) AND “CSCD” (Chinese Science Citation Database); this includes “Cross-Language Search.” The search date was 24 February 2022; of 5928 articles from 1993 to 2021, 5197 remained after deleting the articles with missing information (e.g., no authors, keywords, or abstracts), non-academic papers (e.g., introduction, preamble, and news), and duplicate articles. Figure 3 shows the annual number of published articles.

Figure 3 shows four significant increases in 2003, 2008, 2019, and 2020. According to the aforementioned mapping relationship between academic research and emergency management, these increases may have been induced by China’s major emergency events or by significant policies at these times. (i.e., deliberation of regulations on Emergency Management of Public Health Emergencies in 2003, the Wenchuan earthquake and Beijing Olympic Games in 2008, establishment of the Ministry of Emergency Management of the PRC in 2018, and the COVID-19 pandemic in 2020). The decrease in 2011–2018 may owe to the overall stable domestic situation. The coincidence of article publication trends and the emergency management situation in China have preliminarily verified the mapping relationship.

4.2. Topic Mining and Labeling

4.2.1. Topic Mining

Since 2003 was the beginning of modern emergency management in China, and the articles before that year are too few, while those in 2022 have not yet been fully published, the knowledge trajectory of China’s modern emergency management in this study ranges from 2003 to 2021 (5180 articles in total). Given the irrelevant information in the full text of papers (such as references), and the fact that the title and keywords show a limited representation of the topic, we selected the abstracts of journal articles as the original dataset, comprising about 1,110,000 Chinese characters in total.

As shown in Figure 4, the perplexity-var in LDA is smallest if the topic number is 35 when mining the latent topics from the original dataset, so the optimal topic number is 35. The top 77 keywords of each topic are extracted and sorted from the dataset by LDA according to their significance. The meaningless characters and words (e.g., “lower,” “year,” “example,” “aspect,” and “correlation,” 32 in total) are deleted by manual identification, and the previous steps are used to filter the keywords.

4.2.2. Topic Labeling

This paper uses the manual method to label the topics to ensure the accuracy of semantic expression. The topics are labeled in

Table 1. Keywords and topic name.

Topic Name (No.)	Top Five Keywords
1. emergency plan management	health; language; emergency (adj.); training; evaluation
2. new infectious disease prevention	epidemic; prevention; new; patient; infectious disease
3. new risk supervision	risk; new; supervision; infection; public health
4. industrial accident management	emergency (adj.); management; emergency (n.); preference; research
5. social emergency of disaster	research; emergency (adj.); management; emergency (n.); social
6. smart emergency	evacuation; emergency (adj.); service; smart; research
7. emergency management system	emergency (adj.); pneumonia; management; research; system
8. infectious diseases in universities	management; contact; out of control; infectious disease; university
9. emergency synergy mechanism	synergistic; emergency (adj.); government; cooperation; mechanism;
10. emergency information management	emergency (adj.); management; information; system; pneumonia
11. fire emergency management	emergency (adj.); fire disaster; person; resource; management
12. emergency supply chain coordination	coordinate; supply chain; emergency (n.); integration; response
13. national crisis governance	governance; nation; crisis; society; management
14. urban security management	security; urban; rural; energy; management
15. grassroots government emergency management	emergency (adj.); grassroots; government; management; construction
16. hazardous chemical accident disposal	accident; chemicals; hazardous; leakage; administration
17. flood prevention data management	digital; flood prevention; data management; flood damage; deluge
18. university emergency capability	emergency (adj.); capability; university; management; event
19. emergency support system	emergency (adj.); management system; measures; outpatient service; nurse
20. public health emergency decision	emergency (adj.); decision; analysis; public health; organization
21. community emergency response	information; community; emergency (n.); emergency (adj.); social contact
22. public traffic security	relief; typhoon; accident; metro; emergency (adj.)
23. social emergency modernization	emergency (adj.); public safety; governance; risk; modernization
24. disaster prevention system	emergency (adj.); safety; system; disaster prevention; system
25. emergency management model	emergency (adj.); precision; model; research; analysis
26. public health emergency plan	emergency (adj.); break out; public health; management; event
27. emergency education	emergency (adj.); emergency (n.); education; response; region
28. laboratory emergency management	emergency (adj.); management; laboratory; hospital; message
29. public health emergency environment	event; break out; public health; emergency (adj.); environment
30. railway emergency	railway; station; sociality; national level; goods and materials
31. emergency capability assessment	emergency (adj.); capability; assessment; evaluation; index
32. emergency network public opinion	public opinion; online; mobilize; research; emergency
33. natural disaster emergency mechanism	emergency; mechanism; safety; evolution; natural disaster
34. industrial accident assessment	triangular fuzzy number; safety accident; cluster analysis; classification; stage
35. social risk management	risk; management; social; social work; intergovernmental

, after consulting the government officers and academic experts, and following the independence, divisibility, and systematic principle.

For each topic, 10% of the article was taken to evaluate the accuracy of the topic keyword cluster by the LDA model (Figure 5). Topic 31 (emergency capability assessment) has a maximum accuracy of 97.83%, topic 30 (railway emergency) has a minimum of 92.02%, and the average accuracy is 94.85%. The accuracy of the topic keyword cluster shows the reliability of the data processing result of LDA; the minimum and average values of the accuracy are greater than 90%, indicating that the model is reliable.

4.3. Evolutionary Trajectories of the Topics

4.3.1. Evolutionary Hotspots of the Topics

The percentage of the topics’ keywords is taken to describe the temporal evolutionary trajectories of the focus topics. The temporal heat map of the percentage of the keywords is shown in Figure 6, according to the keywords mined above.

Figure 6 shows that in the early years, Topic 5 (social emergency of disaster) was the most focused area; this demonstrates that modern emergency management research in China has been transferred to focus on the social consequences of the disaster rather than solely the disaster itself, as in the early years. Topic 3 (new risk supervision) is the largest group, and had the most rapid increase in 2020. Risk management is significant because of its complexity and systematic emergency-inducing factors, especially new risks. Chinese academic research on COVID-19 has improved the enthusiasm for new risk management.

From the percentage, we are aware that topics 1, topic 2, topic 3, topic 4, topic 5, topic 6, topic 10, topic 12, and topic 14 suddenly increased in 2018–2020. The keywords and labels of these topics indicate that they are related to the reform of China’s emergency management in the new era. The increase in these topics placed more emphasis on emergency planning, technical governance, and information management, and more importance is attached to preventing unknown risks from the emergency management content. Topic 9 and topic 17 to topic 35 have decreased in the recent ten years, whereas topic 20, topic 22, topic 23, topic 24, and topic 25 have a few published articles after 2015. A possible reason is that the progress of modern technologies has dramatically improved emergency management ability, making some traditional emergencies easy to deal with, which causes the related topics to fade out of the core of academic research, such as flood prevention and public traffic security. Similarly, China’s emergency management system’s improvement makes traditional management issues easier, including collaborative management, emergency decision, etc. In addition, approximately 60% of the research focused on topic 2 (new infectious disease prevention), and topic 3 (new risk supervision) in 2020, and were the most popular topics in 2021; the attention to public social risks increased significantly due to COVID-19. Moreover, the evolutionary trend of these topics shows that the research interest was more distributed before 2010, and relatively concentrated after 2010.

4.3.2. Confusion and Transition of the Topics

The keyword co-occurrence analysis method was applied to measure the similarity of the topics. To show this directly, we accounted for the number of the topics’ common keywords and drew the co-occurrence contour map, as shown in Figure 7.

Given that each topic co-occurs with all its keywords, the diagonal line in Figure 7 has the darkest color; the rest of the areas have different shapes and color depths due to the number of co-occurrences of keywords. Topic 8 (infectious diseases in universities), topic 11 (fire emergency management), topic 12 (emergency supply chain coordination), topic 16 (hazardous chemical accident disposal), topic 17 (flood prevention data management), topic 22 (public traffic security), topic 30 (railway emergency), and topic 34 (industrial accident assessment) had few co-occurrences with the others, indicating that these topics are relatively independent. While topic 11, topic 16, topic 17, topic 22, topic 30, and topic 34 belong to particular types of disasters, the content difference of these disasters causes the corresponding research to have few co-occurrence keywords. Since universities are relatively closed systems, the research on topic 8 is also independent. Research on emergency supply coordination mostly belongs to the theoretical level, and emergency supply coordination is an independent category in the emergency management practice, making topic 12 distinct from the other topics.

A topic could be confused with or transfer into others in the evolution process for the sake of similarity, and the higher the similarity is, the higher the probability of confusion or transition. To calculate the optimal confusion matrix and transition matrix, we imported the Baum-Welch algorithm [64] to train the HMM, and we took the normalized keyword co-occurrence matrix as the initial confusion matrix and transition matrix in HMM. The optimal confusion matrix estimated by HMM is shown in Figure 8.

The diagonal line in this figure indicates that most topics are relatively independent, confirming that the keyword datasets of the topics extracted by the LDA model have sufficient distinctiveness. The squares’ color of topic 32 (emergency network public opinion), topic 34 (industrial accident assessment), and topic 35 (social risk management) are darker, which means these topics are confusing to others. Public opinion on the governance of emergency networks has become more technical, due to the improvement of people’s network participation ability and the development of the Internet. The assessment of industrial accidents is also a technical work, and the accelerated development of emerging industries under the new scientific revolution and industrial transformation makes the relevant research more unique; compared with topic 32 and topic 34, the technicality and uniqueness of laboratory emergency management are more prominent. Hence, the apparent difference between emergency network public opinion, industrial accident assessment, and social risk management and the other topics may lead to related research that is not easily confused with the rest. A group of topics, including topic 7 (emergency management system), topic 10 (emergency information management), topic 18 (university emergency capability), topic 25 (emergency management model), topic 28 (laboratory emergency management), topic 29 (public health emergency environment), and topic 31 (emergency capability assessment), have vague boundaries, indicating they are easy to confuse with the others. The immediate cause of this is the high proportion of co-occurrence keywords, while the primary reason is the high similarity of these topics in the evolution process, causing the topics to be confused when extracting them from the published article text.

The optimal transition matrix estimated by HMM is shown in Figure 9. It indicates that topic 12 (emergency supply chain coordination), topic 17 (flood prevention data management), topic 24 (disaster prevention system), topic 30 (railway emergency), and topic 34 (industrial accident assessment) are less likely to transfer with others. On the other hand, topic 3 (to topic 18, have a 65.25% chance), topic 15 (to topic 23, have a 57.18% chance), topic 19 (to topic 13, have a 51.2% chance), topic 21 (to Topic 2, have a 62.75% chance), and topic 32 (to topic 1, have an 85.04% chance) are more likely to transfer with others. This confirmed the aforementioned features and correlations of the topics.

4.3.3. Transition Direction and Network of the Topics

To explore the transition direction and network of the topics, we calculate the transfer probability between the topics and draw the transfer network according to the transfer direction with a probability greater than 10%. Figure 10 presents the topics’ transition direction. Thirty-one total topics (except topic 12, topic 16, topic 30, and topic 31) transferred into 28 (except topic 3, topic 7, topic 12, topic 16, topic 24, topic 29, and topic 34) in the evolutionary process described by the trained HMM. Topic 3, topic 7, topic 24, and topic 29 diminished significantly in the evolution; the reason may be that these topics are closely related to the COVID-19 epidemic, an accidental hot event, and the related research heat will diminish over time with the improvement of relevant research, as well as the adaptability of the social system.

The transition network of the topics estimated by the HMM is shown in Figure 11 (the nodes represent the topics, and the arrows indicate the transfer direction). It shows that the research topics tend to turn to new characteristics and modernization fields, such as topic 32 (emergency network public opinion), topic 23 (social emergency modernization), and topic 29 (public health emergency environment). Topic 2 (new infectious disease prevention) is the absolute output node, whereas topic 28 (laboratory emergency management) is the absolute input node. The primary transition paths are: (A) topic 21 and topic 2 → topic 5; (B) topic 15 → topic 22; (C) topic 19 → topic 14; (D) topic 8 → topic 6; (E) topic 2 → topic 8; (F) topic 28 → topic 22; (G) topic 33 → topic 15. Path A indicates the concretization of emergency management research in normalized epidemic prevention. Path B is in line with China’s current macro policy orientation, which improves grassroots emergency response capabilities as the fulcrum of social emergency modernization.

Moreover, the improvement of natural disaster emergency mechanisms is the focus of grassroots government emergency management capabilities, that is, path G. Path C shows the significance of the emergency support system for urban security management. Universities usually have a high population density, making them vulnerable to disasters. An intelligent emergency is essential for new infectious disease prevention, which explains path D and path E.

4.4. Topic Prediction

Related research has shown that LDA–HMM has a lower average error when it is applied to forecast the future research trend compared with the other approaches, such as the LDA–GM (Grey Forecast Model), IPC–HMM, and IPC–GM in research topic prediction [15,51], which means that LDA–HMM is reliable for future trend prediction of emergency management research. Taking the 2020 and 2021 as the initial iterative data, we import the confusion and transition matrix parameters into the HMM module of MATLAB to predict the future research trend from 2022 to 2026 in China’s modern emergency management. The result is shown in Figure 12. The S.D. of the topic percentage shows that the overall dispersion is decreasing (0.0520 in 2020 and 0.0102 in 2026).

The prediction results of the model indicate that the research on emergency management in China in the post-epidemic period will be more detailed. The reason may be new factors, such as the normalization of epidemic prevention and emergency management, the acceleration of public risk development, and the empowerment of emerging technologies. The subject, object, target, and value system of emergency management has changed, which makes the knowledge demands of emergency management more detailed. The same pattern of the evolution from 2008 to 2010 goes back to the previous research. From the topics’ percentage values, emergency information management, social emergency of disaster, emergency management system, public health emergency plan, university emergency capability, grassroots government emergency management, and urban security management have higher research heat. Thus, the emergency management research would concern the intellectual needs for the modernization of emergency management in the post-epidemic era. Conversely, the research on industrial accident assessment, railway emergency, public traffic security, flood prevention data management, and disaster prevention system is lacking. The possible reason may be that emerging technologies have improved the means of emergency management; however, several challenges are still faced [65], and the research on optimizing emergency management systems may have more academic value.

5. Conclusions and Discussion

5.1. Main Conclusions

This study introduces the combined approach, LDA–HMM, to explore the evolutionary trajectories of China’s modern emergency management research and to try to predict future trends. Given that the Chinese text is composed of Chinese characters rather than words, and the literature on emergency management research contains many new words, this paper introduces the Viterbi algorithm, which solves linear and nonlinear sequence problems, to segment the text before extracting the keywords and topics. The main conclusions are as follows:

(1) There were 35 research topics on emergency management research in the Chinese articles from 2003 to 2021. By manually labeling the topics and calculating their co-occurrence keyword numbers, we found that emergency plan management, new infectious disease prevention, new risk supervision, industrial accident management, and social emergency disaster were the hot topics. The percentage of the topics’ keywords shows that some were widely concerning in a certain period, some were constantly discussed, and others perished in the evolution. The general evolutionary trend of these topics was distributed before 2010 and concentrated after 2010.

(2) There is a mapping relationship between the hotspots of the published articles and the main events of emergency management in China. The most obvious types of evidence are the emergency events, usually followed by the research hotspots, which could be extrapolated from the heat map of the topics combined with the number of published articles. In addition, the overall knowledge trajectory evolution trend of the topics is also closely related to the node events of China’s emergency management. For example, in 2018, China established the Ministry of Emergency Management of the China; afterward, the research emphasized emergency planning, technical governance, information management, and prevention of unknown risks more strongly than ever before.

(3) Most emergency management research topics could be confused and transferred with others in the evolution process, which indicates the crossing and integration of knowledge in emergency management research. The confusion and transition probability, direction, and network of the topics reveal the transformation mechanism of the emergency management system to a certain extent. On the one hand, the confusion and transition features of the topic evolution reflect the complexity of modern emergency management research objects; on the other hand, they also provide a reference for the construction and improvement of the comprehensive emergency management system.

(4) The future trends show that the research topics will be more detailed and concerned with the intellectual needs of modernization. The macro- and theoretical topics of the modernization of emergency management are predicted to have higher research heat in the next five years. In contrast, the micro- and technological ones have relatively lower research heat. The possible reason is that emerging technologies have improved the means of emergency management in China, and the discussion on emergency response has shifted to the technical field, while the research on optimizing emergency management systems may have more academic value.

5.2. Contributions and Limitations

The marginal contributions of this paper can be summarized as follows. First, we took the Chinese academic literature database as the data sample to detect and explore the knowledge trajectories of modern emergency management research in China, which broadened the research horizons. Second, we introduced the Viterbi algorithm to segment the Chinese text in order to extract vocabulary of emergency management, which innovated the existing methods. Third, we analyzed the evolutionary features of Chinese emergency management research using LDA–HMM, which expanded the application of this combined approach. Fourth, we proposed that there is a mapping relationship between academic research and emergency management, and the empirical research supported this hypothesis. These were the results of reference values for further research and management improvement in China.

Inevitably, there are some limitations to this study. For example, we labeled the topics extracted by LDA with manual methods due to the deficiency of the algorithm, and deviation may exist between the entire message and the topic label. To make the original data set homogeneous and to improve the reliability of the results, we took the articles in CNKI as the data sample, although it could not fully represent the knowledge map of emergency management in China. Moreover, research needs to unveil the knowledge trajectories of the application of emerging technologies, such as blockchain, 5G, Things, and big data. The future trend prediction model based on HMM and machine learning only trains and predicts the existing topics; the results could not present an evolution trend for newly emerged issues. Follow-up studies will expand the data sources to dissertations, research reports, and related policy documents, and will carry out a monographic study of emerging technologies in emergency management.