Large group two-stage risk emergency decision-making method based on big data analysis of social media

2.1. Conversion of language preference matrix

Because of the complexity and fuzziness of emergency decision-making problems, it is difficult for DMs to use quantitative values to describe all the factors affecting a decision. Moreover, the differences among attributes mean that it is difficult to apply the same set of linguistic terms to all of them. Therefore, the proposed method adopts the multi-granularity generalized 2-tuple semantic model [30], which can effectively avoid the loss and distortion of language evaluation information and calculates language information accurately and practically.

Definition 1. [30]: Let S ={ s₀, s₁, ⋯ , s _g  } be a set of linguistic terms and β ∈ [0, 1] be the integrated operational results. The function Δ that transforms a 2-tuple semantic variable (s _r , α) into the real value β is expressed as follows: Δ : [ 0 , 1 ] → S × [ - 1 2 g , 1 2 g )

Δ ( β ) = ( s r , α ) , with { s r , r = round ( β · g ) α = β - r g , α ∈ [ - 1 2 g , 1 2 g )(1) where round() operation yields the integer r that is closest to β.

The 2-tuple semantic variable can be converted into the corresponding real value β ∈ [0, 1] as follows: Δ - 1 : S × [ - 1 2 g , 1 2 g ) → [ 0 , 1 ]

Δ - 1 ( s r , α ) = r g + α = β(2)

Definition 2. [30]: Let S ={ s₀, s₁, ⋯ , s _g  } be a set of linguistic terms and [(s _a , α _a ) , (s _z , α _z )] represent an interval of 2-tuple semantic variables, where (s _a , α _a ) ≤ (s _z , α _z ). The function Δ that transforms the interval 2-tuple semantic variable (s _r , α) into interval value [β₁, β₂] (β₁, β₂ ∈ [0, 1] , β₁ ≤ β₂) is expressed as follows: Δ [ β 1 , β 2 ] = [ ( s a , α a ) , ( s z , α z ) ]

with { s a , a = round ( β 1 · g ) s z , z = round ( β 2 · g ) α a = β 1 - a g , α a ∈ [ - 1 2 g , 1 2 g ) α z = β 2 - z g , α z ∈ [ - 1 2 g , 1 2 g )(3)

The inverse function Δ^-1 is defined as follows:

- Δ - 1 [ ( s a , α a ) , ( s z , α z ) ] = [ a g + α a , z g + α z ](4)

To make it easier for the DMs to express preference information, they can adopt either language values or interval language values (for example, s₁ or [s₁, s₃]). The methods to convert them into interval 2-tuple semantic preferences are respectively as follows: s 1 → [ ( s 1 , 0 ) , ( s 1 , 0 ) ] , ( s 1 , s 3 ) → [ ( s 1 , 0 ) , ( s 3 , 0 ) ]

Term frequency-inverse document frequency (TF-IDF) is a classic method in the vector space model that is used to calculate the feature weight of documents. It was proposed by Salton in 1988 [31, 32]. The TF-IDF of a term is the dot product of the term frequency and the inverse document frequency. The term frequency refers to the frequency at which a term (word or phrase) appears in a particular document, and the inverse document frequency refers to how frequent a term is in all the examined documents. The formula of TF-IDF is as follows: w td = tf td × log ( N df t )(5) where w _td is the weight for term t in document d, tf _td is the frequency of term t in document d, N is the total number of documents in the document corpus, and df _t is the number of documents containing term t in the document corpus.

The advantage of the TF-IDF approach lies in its ability to statistically analyze and deal with big unstructured text data. High weights can be generated by a high frequency word in a particular document as well as that word’s low frequency in other documents. Therefore, TF-IDF tends to filter out common words and keep some less frequent but important words. For example, in general documents, the word “explosion,” which is common in many texts that discuss explosions, will have low inverse document frequency and a correspondingly low weight.

3.1. Description of the problem

In the process of emergency decision-making, let X = {x₁, x₂, ⋯ , x _P } be a set of alternatives, E = {e₁, e₁, . . . , e _M } (M ≥ 11) be a large group of DMs and U = {u₁, u₂, . . . , u _N } be a set of attributes. The weight vector of the DMs is ω = {ω₁, ω₂, ⋯ ω _M }, where ω _i  ≥ 0 and ∑ i = 1 M ω i = 1 . The weight vector of the attributes is W = {w₁, w₂, . . . , w _N }, where w _j  ≥ 0 and . A set of linguistic terms S ={ s₀, s₁, ⋯ , s _g  } is used by the DMs to give the initial preference matrix V i ( 1 ) = ( B lj i ) P × N , where B lj i is the language preference value given by the ith DM for the attribute u _j  ∈ U of alternative x _l  ∈ X.

In an environment in which big data is available, when major emergencies occur, the public’s concerns and opinion information about the event should be fully tapped. Moreover, the risk caused by the uncertainty and deviation from group consistency of DMs’ preferences can be measured using to the reliability and accuracy of DMs’ opinion. On this basis, the scientific ranking of decision alternatives can be realized.

3.2. Process of the method

To take public information into consideration in the process of decision-making, the method proposed in this study starts from the emergency UGC big data obtained by data mining. First, decision-making experts construct an attribute system with the extracted event keywords and then evaluate the prepared alternatives. Because the members of decision-making groups usually have strong dependence behaviors in reality, “herd behavior” is most likely to occur when other members’ decisions are observable and the decision order is determined [33]. In this case, DMs are apt to follow others’ judgment, not expressing their true preferences, and the decision information clearly becomes less reliable. By using this behavioral feature we can let DMs to make decisions in open-ended situations and identify those group members who are more uncertain about their preferences by documenting any changes in preference information. Therefore, the proposed method collects DMs’ preference information in two stages, that is, each DM first gives his or her individual preference information independently, and then re-gives this information with respect to the opinions of the other DMs. Finally, the DMs’ evaluation information is used to rank the alternatives according to the attribute weights and membership weights.

3.3. Construction of the decision attribute system and calculate of the attribute weights

When the data of a major emergency scene is reported to higher levels of management layer by layer, the emergency information becomes delayed and incomplete. It hence becomes more difficult to carry out dynamic real-time monitoring of events, which greatly increases the risk of emergency decision-making analysis. With the development of information technology and the increasing public attention to emergencies, network information resources have become an important part of emergency management in recent years. The occurrence of such events generates a large amount of public information about these events on social networking platforms. This UGC not only reflects the current concerns of the public, but also contains the public’s description, mentality, and perception of the damage caused by the event. Therefore, it is necessary to consider the opinion information of the public in social media to determine the attribute weights for emergency decision alternatives.

(1) Construction of the decision attribute system

The advantage of UGC lies in its immediate response to emergencies, but this information is also very scattered. Using a combination of expert advice and guidance, this problem can be effectively addressed. In the proposed method, the evaluation attribute system of an emergency decision scheme based on UGC consists of two parts: i) the traditional evaluation attributes of emergency decision alternatives and ii) keywords in the real time information of social media users. Specifically, the former must be determined by experts through group discussion and as a framework of the attribute system; The latter are found by first obtaining UGC through data mining, and then extracting several keywords that represent the text features and can be used as a generalization of public opinion information. Finally, the attribute system is constructed using the semantic relationships between keywords and attributes.

The TF-IDF algorithm is adopted to sort the terms after text segmentation for keyword extraction. Then, the terms with higher weights are selected as keywords. The text data extracted by crawling social media sites functions as the document corpus, with each piece of textual information treated as a document. The weight w _h of keyword h is obtained by arithmetically averaging the numbers of all the documents containing the word, as follows: w h = 1 df h × ( ∑ d ∈ D h tf hd ) × log ( N df h )(6)

(2) Damage assessment and attribute weight calculation

Using the established attribute system, a fuzzy comprehensive evaluation can assess the damage caused by events in real time, and provide reference information for DMs. The level of damage is used to describe the impact of emergencies on the local community, people’s property, and the public mindset. Because the keyword acquisition method focuses on statistical analysis rather than content analysis, its weight only represents the importance of its frequency in the documents, but the semantics of different words also reflect different degrees of damage. Hence, the damage assessment of an event requires experts to give each keyword a semantic score S _h to characterize the severity of damage. For example, the word “death” is more serious than “wounded,” so the score assigned to “death” should be greater than that of “wounded.” The score ranges from 0 to 5.

The level of damage reflected in different attributes can be obtained by a weighted sum of each corresponding keyword’s semantic score. The level of damage of attribute j is calculated as follows: Da j = ∑ h = 1 y w h S h ( h = 1 , 2 , ⋯ , y )(7) where y represents the number of keywords corresponding to attribute j.

Because the calculation of the level of damage comes from the subjective perception and the preference information of the public, which affects the event, when the level of damage of a certain attribute is greater, the public has been greatly affected by or given more attention to this factor. To avoid additional serious consequences, DMs must attach more importance to this factor. Therefore, the weight of an attribute is determined by the level of damage of each attribute, as follows: w j = Da j ∑ j = 1 N Da j ( j = 1 , 2 , ⋯ , N )(8)

With the development of computer network and information technology, many specialized decision support systems have been developed, which make it more convenient for a DM to collect, process, exchange, and share information. Because of the large number of DMs in major emergency decision-making, the subjective decision-making ability of different members is an important factor for determining the risk of a decision, and this directly affects the decision-making results. Therefore, the subjective decision risk factors of DMs need to be included in the membership weight determination. The proposed method measures the risk of a DM’s decision based on the two-stage decision-making process and quantifies this risk by defining the reliability and accuracy of his or her opinions. The weight of each member is revised according to the value of his or her decision risk.

(1) Collection of decision preference information

When a major emergency event occurs, DMs make use of the attribute system established using the UGC and the damage assessment information of the event. In the first stage, to formulate multiple sets of alternatives, DM e _i gives his or her language preference information B lj i independently for the jth attribute of the l th alternative. These data are used to form independent language decision preference matrix V i ( 1 ) = ( B lj i ) P × N , where B lj i is defined as the Da _j independent opinion of DM e _i . In the second stage, the independent opinions of the DMs are made public. DM e _i can obtain and observe the preferences of other DMs, and choose to modify or maintain his or her preference information given in the first stage. Then, language preference information Q lj i for each alternative is resubmitted and used to form reference language preference matrix V i ( 2 ) = ( Q lj i ) P × N , where Q lj i is defined as the reference opinion of DM e _i .

During this process, the opinions of other members in the group are considered by each DM as a feedback system, so the opinions in the second phase can be considered as those with respect to other DMs’ opinions. At the same time, the opinions of some members may be affected by those of the others in the second stage. For example, those less-professional DMs who are not sure about the preference information they have given in the first stage will modify their initial preferences after observing the preference information of others. Alternatively, because of crowd psychology, some DMs tend to have opinions that are consistent with most opinions, while some may choose to remain unchanged from their initial preference information. Generally speaking, such members who insist on their first opinions have a higher degree of confidence in making judgments, and the more certain their decision-making preferences are. On the contrary, those who have deviated from their initial opinions are less confident, and the more uncertain their decision-making preferences are.

(2) Measurement of the decision risk coefficient considering the reliability and accuracy of DMs.

To measure the discrepancy between DMs’ independent opinions and reference opinions, the concept of a confidence gap is defined. First, to facilitate processing, after the initial language preference information given by DMs is collected, it is transformed into the form of interval 2-tuple semantics, and then converted into the corresponding interval preference value according to Definition 2. Thus, interval preference matrix V ˜ i ( 1 ) = ( b lj i ) P × N of the independent opinions is obtained, where b lj i = [ b lj iL , b lj iU ] indicates the independent opinion interval preference value and b lj iU ≥ b lj iL ≥ 0 ; reference opinion interval preference matrix V ˜ i ( 2 ) = ( q lj i ) P × N is also obtained, where q lj i = [ q lj iL , q lj iU ] indicates the reference opinion interval preference value and q lj iU ≥ q lj iL ≥ 0 .

Let D ( b lj i , q lj i ) , which is the distance of the interval preference values of the independent and reference opinions, be the confidence gap of DM e _i . When this value is smaller, it indicates that the independent opinion of DM e _i is closer to the reference opinions, which suggests that the DM is more confident about his or her opinion; on the contrary, a more distant gap between the DM’s independent opinion and reference opinion indicates that he or she is less confident. The calculation of this confidence gap is as follows: D ( b lj i , q lj i ) = 2 2 ( b lj iL - q lj iL ) 2 + ( b lj iU - q lj iU ) 2(9) where D ( b lj i , q lj i ) satisfies the following properties: 1) boundedness: 0 ≤ D ( b lj i , q lj i ) ≤ 1 , 2) symmetry: D ( b lj i , q lj i ) = D ( q lj i , b lj i ) , and 3) reflexivity: D ( b lj i , q lj i ) = 0 , if b lj i = q lj i . Properties 1–3 can be easily proved.

Then, the mean value of the attributes of the alternatives is calculated to reflect the group opinion. For example, the averages of the jth attribute of the lth alternative in the independent and reference decision-making processes are respectively as follows: b ¯ lj = [ b ¯ lj L , b ¯ lj U ] = [ 1 M ∑ i = 1 M b lj iL , 1 M ∑ i = 1 M b lj iU ]

q ¯ lj = [ q ¯ lj L , q ¯ lj U ] = [ 1 M ∑ i = 1 M q lj iL , 1 M ∑ i = 1 M q lj iU ](10)

For large group decision members, in past studies on multi-attribute group decision-making, the reliability of a DM’s preference information was rarely considered, or it was simply assumed that all preference information is completely reliable. In fact, no DM is entirely rational, and the different degrees of reliability of the judgements will have an impact on the decision result.

Let Ra lj i be the reliability of a DM’s opinions. It indicates the level of certainty of the preference information for the jth attribute in the l th alternative given by DM e _i , and is calculated as follows: Ra lj i = 1 + D ( b ¯ lj , q ¯ lj ) 1 + D ( b lj i , q lj i )(11) where the meanings of D ( b ¯ lj , q ¯ lj ) are given in Equations (9)–(10), which indicate the differences of the two group opinions in independent and reference situations. If the value of Ra lj i is zero, this indicates that most members of the group adhere to the opinions of the first phase. Because D ( b ¯ lj , q ¯ lj ) is a constant value for all DMs, the ratio of D ( b ¯ lj , q ¯ lj ) to the confidence gap of each DM can reflect the degree of certainty of that DM’s opinion. If a DM’s confidence gap D ( b lj i , q lj i ) is bigger, the ratio will be smaller, indicating that the opinion given by that DM has a lower degree of reliability, that is, his or her preference information is of higher uncertainty. In contrast, when a DM’s confidence gap is smaller, the ratio will be larger, indicating that his or her opinion is of higher certainty. At the same time, to avoid allowing the denominator to be zero, one is added to both the numerator and denominator.

In addition, the consensus reaching of group opinions in group decision-making is also a matter of concern [34–37]. It is often more practical to determine whether the outcome of a decision can be accepted or to what extent it can be accepted by a group because it relates to whether the alternative can be implemented quickly. When individual opinions deviate from group opinions to a greater extent, this shows that their resistance to group consistency has increased, and the decision risk brought by member preference conflicts will be triggered. To enable the decision-making result to reach a higher level of consensus, this paper argues that the majority opinion is more accurate than the minority opinion, that is, when the preference information given by a DM is closer to the opinion of the group, it indicates that the DM is more professional.

Let Ac lj i be the accuracy of a DM’s opinions. It indicates the degree of consistency of preference given by DMe _i with respect to those of other members for the jth attribute in the l th alternative and is calculated as follows: Ac lj i = 1 1 + D ( q lj i , q ¯ lj )(12) where D ( q lj i , q ¯ lj ) represents the distance between the individual opinion of a DM and the group opinion in the second stage. If D ( q lj i , q ¯ lj ) is smaller, then a larger Ac lj i is obtained, which means that the closer the DM’s opinion is to most people, the more contribution they make to the group consistency. Conversely, if D ( q lj i , q ¯ lj ) is larger and Ac lj i is smaller, this indicates that the opinions he or she gives are deviate more from the group consistency.

For DMs, the risk caused by their own decision-making ability is unavoidable. A high degree of reliability alone does not guarantee a lower-risk decision will be made. When lacking knowledge or experience about an event, a DM may insist on sticking with his or her initial observations to maintain the illusion of his or her professionalism. Therefore, the decision risk of DMs is measured by both the reliability and accuracy of their opinions.

First, the reliability of Definition 4 and the accuracy of Definition 5 are both benefit type indicators but they have different ranges. To eliminate the influence of different dimension scales on the decision-making results and to facilitate judgment, a non-linear transformation of the reliability and accuracy is conducted to obtain standardized values, calculated as follows:

- { Ra lj i ′ = Ra lj i - min i Ra lj i max i Ra lj i - min i Ra lj i Ac lj i ′ = Ac lj i - min i Ac lj i max i Ac lj i - min i Ac lj i i ∈ 1 , 2 , ⋯ , N(13)

Through the above explanation, the decision risk coefficient of a DM can be defined as λ ⁱ , indicating the risk degree of both the uncertainty and deviation from group consistency of a DM’s preference, as follows:

λ i = 1 P · N ∑ l = 1 P ∑ j = 1 N [ θ ( 1 - Ra lj i ′ ) + ( 1 - θ ) ( 1 - Ac lj i ′ ) ](14)

where θ and (1 - θ) represent the weights corresponding to reliability and accuracy, respectively. These weights are determined by the DMs according to the decision-making problem and the actual situation. If there is no special preference, the value is set as θ = 1/2, indicating that reliability and accuracy are given the same weight.

For the acquisition process of λ ⁱ , first, the decision risk coefficient of a DM’s preference given for the attributes of different alternatives is calculated, and then the decision risk coefficients of all members are obtained by simple weighted arithmetic averaging. Obviously, when a DM’s opinion is more reliable or accurate, he or she is given a smaller the decision risk coefficient, and his or her decision risk is smaller. On the contrary, a greater the value of λ ⁱ indicates a greater the decision risk.

(3) Large group preference clustering and the weight of members.

After the decision risk has been quantified for large-group members, the weight of the members needs to be determined. Using opinion clustering, DMs’ opinions are analyzed and managed, with the large number of members involved in decision-making considered. In the proposed method, the clustering method in [38] is used to cluster adjusted reference opinion interval preference matrix V ˜ i ( 2 ) = ( q lj i ) P × N , and the decision groups can be clustered into K clusters, where n _k is the number of members in cluster C ^k . Using this clustering method, DMs in the same cluster can be assigned the same weight because they have similar preference information. At the same time, DMs in larger-scale clusters should be assigned higher weights based on the majority principle. Hence, initial weights of members obtained from the clustering result are as follows:

ω i ′ = ω n k = n k / ∑ k = 1 K n k 2 , i = 1 , 2… , M(15)

Next, the initial weights of the members are adjusted by the member’s decision risk coefficient λ ⁱ . A greater value of λ ⁱ indicates more uncertainty or deviation from the consistency goal for the information provided by the DM. Hence, a lower weight should be given to the DM. Finally, the weight of decision risk of DM e _i is calculated as follows:

ω i ′ = ( ω ′ i ⋅ 1 λ i ) / ∑ i = 1 M ( ω ′ i ⋅ 1 λ i )(16)

Briefly, the steps of the proposed method is as follows:

Step 1: UGC related to emergencies is extracted by crawling social media sites, and the attribute system is constructed using the attribute framework developed by experts and the keywords extracted from the UGC.

Step 2: Using the TF-IDF weight of keyword w _h obtained from Equation (6) and semantic evaluation value S _h determined by the expert scoring method, the level of damage of each attribute Da _j is calculated according to Equation (7) and the weight of attribute w _j is obtained after normalization.

Step 3: Using the two-stage decision-making process, independent opinion language preference matrix V i ( 1 ) = ( B lj i ) P × N , and reference opinion language preference matrix V i ( 2 ) = ( Q lj i ) P × N given by the DMs are collected. They are respectively transformed into independent opinion interval preference matrix V ˜ i ( 1 ) = ( b lj i ) P × N and reference opinion interval preference matrix V ˜ i ( 2 ) = ( q lj i ) P × N using Equations (3) and (4) as well as Definition 2.

Step 4: Reliability Ra lj i and accuracy Ac lj i of the preference information given by each DM are respectively calculated by Equations (11) and (12) and normalized by Equation (13). Then, the decision risk coefficient λ ⁱ of each member in the decision-making group is obtained by Equation (14).

Step 5: The reference opinion preference matrix for all DMs is clustered. Then, the initial weight of each member in cluster C ^k is obtained by Equation (15). The decision risk weights of all group members are determined and combined the decision risk coefficient.

Step 6: Alternatives are ranked using the TOPSIS method.

(1) The DMs’ reference opinion interval preference matrix is clustered using attribute weights to obtain the weighted decision matrix

V ˜ i ( 2 ) = [ v 11 i v 12 i ⋯ v 1 N i v 21 i v 22 i ⋯ v 2 N i ⋯ ⋯ ⋱ ⋯ v P 1 i v P 2 i ⋯ v PN i ] where v lj i = q lj i × w j , and 1 ≤ i ≤ M, 1 ≤ j ≤ N.

(2) The positive and negative ideal solutions are obtained for each alternative. V + = { v 1 + , v 2 + , ⋯ , v N + } , V - = { v 1 - , v 2 - , ⋯ , v N - } V j + = [ v j + L , v j + U ] = [ min ( v lj iL ) , min ( v lj iU ) ] , j ∈ N V j - = [ v j - L , v j - U ] = [ max ( v lj iL ) , max ( v lj iU ) ] , j ∈ N

(3) The distance between the positive and negative ideal solutions is calculated.

D l i + = ∑ j = 1 M ( V lj i - V j + ) 2 = ∑ j = 1 M ( V lj iL - V j + L ) 2 + ( V lj iU - V j + U ) 2 D l i - = ∑ j = 1 M ( V lj i - V j - ) 2 = ∑ j = 1 M ( V lj iL - V j - L ) 2 + ( V lj iU - V j - U ) 2

(4) The closeness degree of each alternative is calculated.

C l = D l - D l + + D l -(17)

where D l + = ∑ j = 1 M ω i D l i + and D l - = ∑ j = 1 M ω i D l i - .

Finally, the alternatives are ranked based on the value of C _l and the optimal alternative can be selected.

4.1. Case background

At 23:30 h on August 12, 2015, an major fire and explosion accident occurred in the warehouse of Ruihai Company in Tianjin Port. According to seismometer data, the intensities of the two explosions were equivalent to 3 tons and 21 tons of TNT, respectively. The explosion caused heavy casualties and triggered a large number of container fires. The local fire command center responded quickly after receiving the emergency call. By 11:00 h on August 13, Tianjin Fire Corps had deployed 143 fire trucks and more than 1,000 firefighters for search and rescue at the scene. At the same time, the accident sparked strong concern from the public, who expressed it through social media. A large number of Internet users expressed their opinions and thoughts on social networks. As a large number of dangerous goods were stored on the scene and the specific explosives could not be determined, to avoid further aggravation of the situation, Tianjin Emergency Decision-making and Command Center convened 16 emergency management experts on the network and formulated the following three emergency alternatives according to the current accident status:

x₁: Continue to carry out environmental emergency monitoring and increase the number of firefighters and fire trucks at the support site to enhance fire control.

x₂: Continue to carry out environmental emergency monitoring, inform the fire-fighting authorities to evacuate the scene, and dispatch an anti-chemical team with professional equipment for search and rescue at the scene.

x₃: Continue to carry out environmental emergency monitoring, suspend the fire-fighting and search and rescue operations, and urgently identify the content, quantity, and storage of dangerous chemicals before making further decisions. The expert group identified six factors as the evaluation attributes of the alternatives based on past experience of response to such emergencies as follows: the cost of alternatives (u₁), casualties (u₂), property loss (u₃), degree of public panic (u₄), impact on the environment (u₅), and impact on life (u₆). Among them, for attributes u₁ –u₄, language term set S¹ with a granularity of 9 was adopted, and for attributes u₅ –u₆ language term set S² with a granularity of 7 was adopted. They are expressed as follows: S 1 = { s 0 = extremelylow , s 1 = verylow , s 2 = low , s 3 = slightlylow , s 4 = medium , s 5 = slightly high , s 6 = high , s 7 = veryhigh , s 8 = extremelyhigh }

S 2 = { s 0 = verysmall , s 1 = small , s 2 = slightlysmall , s 3 = medium , s 4 = slightlybig , s 5 = big , s 6 = verybig }

4.2. Decision-making steps

Step 1. Large quantities of text messages on social networks were collected to support emergency decision-making and Sina Weibo was selected as the source of event-related UGC data. “Tianjin Explosion” and “Tanggu Explosion” were used as the search keywords. With the API interface provided by Sina Weibo, microblogs within the time range of 23:30 h on August 12 to 15:00 h on August 13 were captured. The recorded fields include the author’s name, microblogging content, post time, and comment content. Because of the inclusion of non-public-generated content or noisy data, a total of 79,211 microblog posts were retained after manual screening to remove news releases from the government or mainstream media accounts and messages with fewer than two words.

After the collection of the UGC text data from Sina Weibo, the data were preprocessed to remove hashtags, @ content, URLs, and other unwanted content. Then, then the text was segmented into words, stop words were removed, and part of speech tagging was performed. Finally, the frequency of words in the data was analyzed. In the implementation of the proposed method, keyword extraction is implemented in the Python, and the Python-based jieba word segmentation software package was used to divide the text into terms. The 500 terms with the highest TF-IDF values were selected as candidate keywords. Because there are several words in the Chinese language that express the same meaning, keywords with the same semantic meaning, such as “death,” “killed,” “remains,” and “pass away” were merged into the keyword “death.” Similarly, keywords “horrible,” “terrible,” and “fear” were merged into “terrible.” The TF-IDF value of the merged words was accumulated. Using the attribute framework determined by the experts, the keywords corresponding to each attribute were listed based on their semantic relationships, thus forming the evaluation system of the emergency decision-making scheme of the 8/12 Tianjin Port major explosion, as shown in Table 1.

Step 2. The TF-IDF weights of the keywords were calculated by Equation (6) using the Python program. Then, using the expert scoring method, 10 experts were asked to give scores for the keywords in the attribute system according to their severity, and the semantic score of each keyword was obtained after statistical averaging. The level of damage of each attribute was calculated by Equation (7). An example of the calculation of the level of damage of attribute u₂ is shown in Table 2, with the result of 21.206.

After the calculation of the level of damage of each attribute, the weight of the attributes was obtained by Equation (8), resulting in W = {0.07, 0.25, 0.16, 0.23, 0.20, 0.08}.

Step 3. In accordance with the two-stage decision-making process, DMs gave their preference information of independent opinion and reference opinion according to the linguistic sets of the six attributes. In this process, part of the DM’s preference information was modified in the second phase, except for that of DMs e₃, e₁₃, and e₁₆. Definition 2 was used convert the language preference matrices into interval preference matrices, as shown in Tables 3–6.

Step 4. According to the differences between each DM’s opinions in the independent and reference situations as well as the differences between his or her individual opinions and the group opinions, the DM’s reliability and accuracy were calculated respectively by Equations (11) and (12), and then normalized by Equation (13). The results for DM e₁ are shown in Tables 7 and 8. Then, the DM’s decision risk coefficient in the group was calculated by Equation (14), as shown in Table 9.

Step 5. The method in [38] was used to cluster the reference opinion preference matrix of the 16 DMs. Threshold γ was set to 0.85, and four clusters and their weights were obtained. The clustering results are shown in Table 10.

The initial weights of DMs were calculated by Equation (15). Combined with the decision risk coefficient, the decision risk weight of each DM was calculated by Equation (16), as shown in Table 11.

Step 6. The closeness degree of each alternative was calculated by Equation (17) and the ranked results are shown in Table 12.

It can be concluded from the result that alternative x₂ should be selected to handle the accident, i.e., to continuously carry out environmental emergency monitoring, inform the fire-fighting force to evacuate the scene, and dispatch an anti-chemical team with professional equipment for search and rescue at the scene.

The results of the case analysis reflect the fact that people pay more attention to the casualties and impact on the environment. Moreover, the explosion accident had a certain impact on the public psychology, which is also an important reason for choosing the scheme with the minimum risk of casualties. Therefore, the government should also collect public opinion after the accident using reasonable means, understand the trends in public opinion to assist scientific decision-making, release the latest progress of the incident, and prevent secondary public sentiment from causing greater impact.

To validate the effectiveness and illustrate the advantages of the proposed method, the results are compared with the expert weight approach proposed in [39]. Only the weight of internal DMs changes, as shown in Table 13; the two methods have the same ranking result.

Ref. [39] conducted a clustering analysis for large group decision-making and proposed a two-layer weight model, which holds that the weight of clustered internal members is related to a DM’s contribution to cluster consistency. In the proposed method, the adjustment of a DM’s weight is based on the group consistency. At the same time, the risk caused by the uncertainty of their preference is also considered. By measuring the reliability and accuracy of the preference information of DMs, the less-professional DMs can be effectively identified, and the determination of each member’s weight will be more reasonable. This article does not compare the methods of determining the weights of attributes. Because the determination of attribute weights in the current literature mostly uses an expert preference matrix or is subjectively given by experts, the methods are not comparable. However, in terms of ideas, the proposed method has a strong advantage in that obtaining the attributes from social media data is an objective approach.