A Study on Selection Strategies for Battery Electric Vehicles Based on Sentiments, Analysis, and the MCDM Model

Under the goal of carbon peak and carbon neutrality, developing battery electric vehicles (BEVs) is an important way to reduce carbon emissions in the transportation sector. To popularize BEVs as soon as possible, it is necessary to study selection strategies for BEVs from the perspective of consumers. 'erefore, the Latent Dirichlet Allocation (LDA) model based on fine-grained sentiment analysis is combined with the multi-criteria decision-making (MCDM) model to assess ten types of BEV alternatives. Fine-grained sentiment analysis is applied to find the vehicle attributes that consumers care about the most based on the word-ofmouth data. 'e LDA model is suggested to divide topics and construct the indicator system. 'e MCDM model is used to rank vehicles and put forward the corresponding optimization path to increase consumer purchases of BEVs in China.'e results show that (a) via the LDA model based on fine-grained sentiment analysis, attributes that consumers care most about are divided into five topics: dynamics, technology, safety, comfort, and cost; (b) based on the DEMATEL technique, the dimensions in the order of importance are as follows: safety, technology, dynamics, comfort, and cost; (c) the price is the most important criteria that affect customers’ satisfaction by the DANP model; and (d) based on the VIKOR model, the selection strategies present that Aion S is highlighted as the best choice, and the optimization path is discussed to promote the performance of BEVs to increase customers’ satisfaction.'e findings can provide a reference for improving the sustainable development of the automobile industry in China. 'e proposed framework serves as the basis for further discussion of BEVs.


Introduction
With the rapid growth of the number of vehicles, the issue of energy consumption and greenhouse gas (GHG) emissions in the transportation sector are attracting increasing attention worldwide [1,2]. According to the International Energy Agency, the transportation sector accounts for about 24.6% of the world's energy-related carbon dioxide emissions [3]. Automobile emissions have been a major source of emissions in the transportation sector [4][5][6], which aggravates the deterioration of the ecological environment and causes a series of health problems [7][8][9]. China owns the largest automobile market in the world since 2009 [10], and car ownership has exceeded 200 million by 2020, indicating that the issue of energy security and environmental pollution will become more prominent [11,12]. To alleviate these problems, governments and automobile manufacturers pay more attention to develop cleaner and more efficient alternative-fuel vehicles [12,13], which induces the upsurge of battery electric vehicles (BEVs) [14][15][16][17]. And, BEVs will become the mainstream of vehicle sales by 2035 in China [18].
However, in the early stage of the development of BEVs, only relying on market forces is not enough for the commercialization and popularization of BEVs [19]. us, the Chinese government has formulated a series of incentive policies for consumers and BEV manufacturers [20][21][22][23], e.g., purchase subsidies, tax exemption, free parking, and driving privileges [24,25]. Based on the above incentives, from 2011 to 2019, the production and sales volume of BEVs increased from 5655 and 5579 to 1,020,000, and 972,000, respectively. More importantly, as a result of the reduction of subsidies, sales of BEVs in July 2019 dropped significantly compared with June 2019, with a drop of 59.9% [26]. is means that the advantages of policy support offered by BEVs are not enough to persuade consumers [27]. Previous studies have highlighted that the low market share of BEVs is related to consumers' perceived uncertainty [28,29], which means that exploring the selection strategies of consumers will help enlarge the market share of BEVs [30,31].
Existing studies on vehicle selection and comparison are mainly based on life cycle assessment (LCA) [32], statistical methods [33], and multicriteria decision-making method (MCDM) [34,35], but the influence of consumer sentiments on vehicle selection strategies is ignored. erefore, this study attempts to use the Latent Dirichlet Allocation (LDA) model based on the fine-grained sentiment analysis, that is, to analyze the sentiment polarity of online reviews, to reveal the customer's sentiment towards the attributes of BEVs' models, and to extract the attributes that consumers care most about as evaluation criteria. e topic analysis is applied to distinguish dimensions. is method considers the index from the consumer's point of view and effectively avoids the problem of artificial selection. Based on this, quantitative cause-effect relationships among vehicle attributes are included as an important point in BEVs' assessment. It makes up for the lack of considering the correlation between attributes in the existing research. Taking ten domestic BEVs models in China as research samples, a detailed and practical optimization path is put forward, which solves the shortage that the existing research only considers evaluation and selection.
e main tasks of this study include the following aspects: (1) Using consumers' word-of-mouth data from Autohome website and LDA model based on the finegrained sentiment analysis, a new evaluation index system of BEVs selection based on customer perspective is constructed. (2) Decision Making Trial and Evaluation Laboratory (DEMATEL) technique is utilized in this study to identify the interaction between criteria within each dimension. Analytic network process (ANP) is then employed to determine the weights of criteria and dimensions. (3) Modified Vlse Kriterijuska Optimizacija I Komoromisno Resenje (VIKOR) is applied to calculate the indicator gaps and comprehensive evaluation scores for ten types of BEVs in China. Based on the above research framework, the customer selection strategies and the manufacturer optimization paths are proposed. e specific analysis process is shown in Figure 1. e remainder of this study is organized as follows. Section 2 gives a literature review. Section 3 describes the methods. Section 4 presents the results. Section 5 conducts a discussion. Section 6 presents the conclusions.

Literature Review
is section briefly reviews the literature from three aspects: application of machine-learning methods in the field of electric vehicles (EVs), application of MCDM models in vehicle selection, and the factors influencing consumers' choice of BEVs.

Application of Machine-Learning Methods in the Field of EVs.
With the rise of big data analysis and machine learning, many scholars have applied it in the field of EVs. Bas et al. [36] applied supervised machine-learning techniques to identify key elements influencing EVs' adoption and classify potential EVs purchasers. De Clercq et al. [37] made use of two machine-learning techniques, i.e., LDA model and multi-label classification algorithm to extract subject words from patent texts and classify patents into multiple cooperative patent categories. Yang et al. [38] evaluated and optimized relevant policies of China's new energy vehicle industry based on text mining technology. In the study of Naumanen et al. [39], LDA model was used to identify emerging research topics according to papers and patents related to heavy duty BEVs. Aguilar-Dominguez et al. [40] used a machine-learning method to evaluate the availability of five vehicles participating in the vehicle-to-home services. Basso et al. [41] applied the probabilistic Bayesian machine-learning model to solve the route selection problem and found the best route for EVs by predicting the energy consumption within the limited driving distance. Ma et al. [42] adopted big data and textmining technologies to analyze online behavior of Chinese consumers and to identify the factors affecting consumers' preferences. Different from previous research on online comment mining, emotion analysis is not only the mining and analysis of consumers' online comments but also the extraction and interpretation of emotions expressed in texts [43]. e purpose of sentiment analysis is to investigate the emotion polarity of online texts and divide them as positive, neutral, and negative [44]. However, only a few papers in the literature apply sentiment analysis to the analysis of consumers' recognition of vehicle. Via big data platform, Deep Learning techniques were employed by Jena [44] to explore and classify consumers' sentiment towards EVs in India.

Application of MCDM Models in Vehicle Selection.
According to previous studies on MCDM, MCDM models could be classified into the following two categories. e first category is applied to calculate the weights of alternatives, such as the Simultaneous Evaluation of Criteria and Alternatives (SECA) [34], Analytical Hierarchy Process (AHP) [35], ANP [45], and Stepwise Weight Assessment Ratio Analysis (SWARA) [46]. In the second category, the ranking of alternatives is based on comprehensive scores such as Measurement of Alternatives and Ranking according to Compromise Solution (MARCOS) [34], Complex Proportional Assessment (COPRAS) [46], VIKOR [47], Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) [48], ELimination and Choice Expressing the REality (ELECTRE) [49], Preference Ranking Organization Method for Enrichment Evaluation (PROMETHEE) [50], and so on. In addition, the combination of these methods has also attracted the attention of academic circles. To enhance the robustness, scholars have combined the single method and formed the newly developed hybrid MCDM models.
In recent years, based on the hybrid MCDM models, a series of scientific and systematic theoretical explorations have emerged, especially in vehicle evaluation and selection. For example, Li et al. [47] adopted an MCDM combining AHP and VIKOR to rank and optimize four types of vehicles, including EVs, gas vehicles, methanol vehicles, and ethanol vehicles, to provide references for decision-makers in the new energy automotive industry. Using intuitionistic fuzzy set and TOPSIS methods, Onat et al. [48]

e Factors Influencing Consumers' Choice of BEVs.
Vehicle purchasing is closely related to consumers' acceptance. Up to now, many studies have explored the possible drivers or barriers that influence consumers' choice of BEVs [54]. Liu et al. [55] found that the customer experience is the main decisive factor in buying BEVs. Kim et al. [56] concluded that customers with good driving experiences and knowledge are more likely to buy BEVs in Korea. Li et al. [57] found in their study that family factors, such as scale, income, and location, can influence consumers' choice of BEVs. She et al. [58] demonstrated that older, experienced, and environmentally conscious consumers were more interested in buying a BEV in Tianjin. Besides, safety, reliability, and range were the three key obstacles to BEV sales. Dong et al. [59] recognized that urban consumers' purchasing decisions would be changed by psychological factors such as subjective norms, feelings and emotions, personal norms, and perceived behavioural control. Li et al. [60] argued that fast charging time and battery warranty can promote consumers' adoption of BEVs. Das et al. [51] evaluated the performance of EVs based on nine attributes, such as the price, battery capacity, torque, charging time, overall weight, seating capacity, driving range, top speed, and acceleration. Nazari et al. [31] stressed that the elimination of concerns such as technical uncertainty, limited vehicle styling, and charging time will increase the utilization rate of BEVs. In addition to price and battery technology, Ma et al. [42] suggests that the design of the exterior and interior has a strong appeal to consumers. Kukova et al. [61] pointed out that internal space, operating reliability, and braking are also important attributes affecting whether consumers choose BEVs or not. Li et al. [24] believed that the implementation of financial incentives, such as purchase subsidies and tax exemption, played an indispensable role in promoting Chinese consumers to adopt BEVs. In the study of Cheng et al. [4], reduction in battery charging time and maintenance cost were the first two major measures to motivate consumers to purchase BEVs. Although existing studies have considered the influence of demographic, technological, and psychological factors on consumer purchasing behavior, researchers have shown that consumers' decision to purchase BEVs is largely determined by the vehicle's performance characteristics [62,63]. erefore, in this study, only the influence of vehicle attributes on BEVs selection is considered, and other factors are not considered.
To sum up, although previous studies on vehicle selection have made some progress, there are still the following limitations: (1) Some studies have not clearly explained how to select indicators. In addition, some studies have pointed out that indicators are obtained through literature review, but this process is susceptible to subjective factors. (2) e indicators are interdependent, but many studies have not clearly identified the cause-effect relationship. (3) e literature does not provide beneficial guidance for consumers to choose BEVs in China. ere is no mention of the optimization paths of specific models. erefore, this study adopts the LDA model based on fine-grained sentiment analysis to obtain indicators. e MCDM model is used to identify interrelationships, and to propose selection strategies and optimization paths for BEVs in China from the point of view of consumers.

Methodology
is research proposes a hybrid model combining finegrained sentiment analysis and MCDM model to form a novel framework to study consumers' selection strategies for BEVs. Specifically, the LDA model based on fine-grained sentiment analysis is applied to identify dimensions and criteria based on the word-of-mouth data of BEVs in Autohome website. DEMATEL technique is used to construct an influential network relationship map (INRM). DEMATEL-based Analytic Network Process (DANP) is used to confirm the impact weight of each evaluation indicator based on ANP [64]. Finally, VIKOR is not only used to evaluate and obtain the selection strategies but also to find the gaps in each evaluation indicator and make optimal paths to improve consumers' adoption.

Text Preprocessing.
In this study, the Scrapy framework developed in Python is used to crawl the word-ofmouth data. However, invalid data in the crawled text will affect the effectiveness of data output. If these invalid data are introduced into subsequent models, it can have a significant impact on the results of the analysis. erefore, text preprocessing should be carried out after obtaining the word-of-mouth data of the Autohome website. In this paper, the process of text preprocessing is divided into several steps, including data splitting, data cleaning, text segmentation by jieba clauses, removing stop words, adding self-defined automobile dictionary, and data transformation.

Doc2vec
Model. After preprocessing the data, the text data are divided into a test set and a training set based on a certain ratio. en, we build a Doc2vec model based on the training set data, train the test set data with the Doc2vec model, and finally build a support vector machine classifier to calculate the accuracy of the test set.

Skip-Gram Model.
Emotional granularity can be divided into two types, i.e., positive and negative emotional polarity; however, to increase the accuracy of the regression model, the emotional granularity is further refined, i.e., emotional polarity is divided into five types (very satisfied, satisfied, fair, dissatisfied, and very dissatisfied). e text preprocessed data are trained 100 times using word2vec to obtain a Skip-gram model (a method for learning highdimensional word representations that capture rich semantic relationships between words) with a word vector dimension of 200 dimensions. e results of the text preprocessing are used to obtain the emotion words with high word frequency, and then the Skip-gram model is used to obtain words that are similar to the emotion words to obtain a more comprehensive emotion vocabulary.

Clustering.
e clustering algorithm divides the emotional lexicon of word-of-mouth into five categories based on the difference between the customers' expectations and the actual perception of the BEVs.
3.1.5. LDA Model. Topic models are algorithms for discovering key topics in a large and unstructured collection of text [65]. LDA is one of the topic models, which is applied to automatically discover topics in the text that consumers are most satisfied and least satisfied with. e core computational problem for topic models is to use the collected text to infer the hidden topic structure [66].
us, this study identifies the consumer concerns by using the LDA model, to discover key topics from the collected text.

MCDM Model
3.2.1. DEMATEL Technique. DEMATEL technique is a systematic factor analysis method to detect the cause-effect relationships between complicated indicators using graph theory and matrix tools [67], originally proposed by the Battelle Research Centre in 1972 [68]. DEMATEL has been used to solve complicated real-world problems by building an INRM [69], such as optimal online travel agencies [70], regional innovation capacity [71], and sustainable online consumption [72].
us, the steps of this technique are summarized as follows.
Step 1. Finding the average direct effect matrix. e mutual direct effect among criteria is evaluated by the knowledge-based experts. e scales ranged from 0 to 4 where "0" means "absolutely no effect" and "4" means "very high effect." "1," "2," and "3" mean "low effect," "middle effect," and "high effect," respectively. By a pairwise comparison, we can obtain these groups of direct matrices by scores, where ij represents the influence from criterion i to criterion j. After that, we can calculate an average direct effect matrix G (as seen in equation (1)), where each criterion is the average of the corresponding criteria in the experts' direct matrices [73].
Step 1. Setting up the normalized direct-influence matrix X. e matrix X can be acquired by using equations (2) and (3). (3) Step 1. Computing the total influence matrix Tc. e total influence matrix T c can be derived from equation (4), where matrix I denotes a unit matrix. where ij c ≤ 1, and at least the sum of one row or column (but not all) equals one; lim θ⟶∞ X θ � [0] n×n .
Step 4. Building the INRM and analyzing the results. e sum of rows and the sum of columns of total influence matrix T c can be, respectively, represented by vector r and vector s according to equations (5)-(6), where r i indicates the total influence of criterion i on others, the s i denotes the total influences received by criterion j from other criteria. When i � j and i, j ∈ {1, 2, . . ., n}, the vector (r i + s i ) expresses the importance of criterion i in the question. Likewise, the vector (r i − s i ) identifies the degree of causality among indicators. Simultaneously, if (r i − s i ) is positive, the criterion i influences other criteria. On the contrary, if (r i − s i ) is negative, the criterion i is affected by others. Finally, draw the INRM in which the vertical axis represents (r i + s i ) and the horizontal axis represents (r i − s i ) [74].
e total influence matrices has two forms, one is T c � [t ij c] n×n (equation (7)), where n represents the number of the criteria, and the other is where m represents the number of dimensions.

DANP Method.
e ANP method was first proposed by Saaty [64], to deal with the interdependence and feedback problems among indicators [75,76]. It originates from the AHP but eliminates the deficiency of AHP, which assumes that the indicators are independent of each other [77]. However, the weighted super-matrix in the ANP method lacks rationality because it assumes that each cluster has the same weight [78]. erefore, the DANP is an appropriate method to obtain the influential weights by improving the normalization process and addressing the interrelationships among indicators [45]. It has been used in many different fields, such as low-carbon energy planning [79], material selection [80], and renewable energy selection [81]. us, the process of this technique involves the following steps.
Step 1. Calculating the normalized total-influential matrix T nor D.
According to equations (8) and (9), matrix T nor D is framed through normalizing the total-influential matrix T D . First, the sum of each row in matrix T D can be expressed as where m represents the number of dimensions. en, the normalized total-influential matrix T nor D is calculated by dividing the elements in each row by the sum of the row, so that Step 2. Exporting the normalized matrix T nor c by dimensions and clusters.
A new matrix T nor c is acquired by normalizing T c with the total degrees of effect and influence of the dimensions and clusters: Step 3. Determining the unweighted super-matrix W c . e unweighted super-matrix W c is obtained by transposing the normalized matrix T nor c: Step 4. Constructing the weighted super-matrix W * c. e normalized total-influential matrix T nor D is obtained by equation (9) and the unweighted super-matrix W c is obtained by equation (11). us, using equation (12), a weighted super-matrix W * c, which improves the traditional ANP by using equal weights to make it appropriate for the real world, can be obtained by the product of T nor c and W c , i.e., W * c � T nor D * W c . is demonstrates that the influential level values are the basis of normalization to determine a weighted super-matrix.
Step 5. Calculating the influential weights of the criteria. e influential weights w � (w 1 , . . . , w j , . . . , w n ) can be obtained according to the weighted super-matrix W * c, and multiplied several times until it converges a stable supermatrix, so that lim α⟶∞ (W * c) α , where α is a positive integer number.

VIKOR Method.
e VIKOR method was first proposed by Opricovic to optimize the multiple criteria of complicated systems in 1998 [82]. It is applied to rank and select from a set of alternatives given the conflicting criteria. VIKOR is also the compromise ranking method, based on the concept of the Positive-ideal (or the aspired level) solution and Negative-ideal (or the worst level) solution [83]. So, the order of results can be compared by "proximity" to the "ideal" alternative [82,84,85]. In this study, the modified VIKOR method can be used to increase customers' satisfaction in alternatives that are influenced by the interaction of various factors. All of the steps for VIKOR are presented as follows: Step 1. Determining the positive-ideal solution, negativeideal solution, and the gap.
According to the concepts of VIKOR, f * j is the positiveideal point of assessment criteria, which indicates the best value (aspiration level); In contrast, f − j is a negative-ideal point, which means the worst value. In this study, the best value is set as f * j � 10 [64]. Likewise, the worst value is set as f − j � 0 with scores of criteria ranging from 0 (dissatisfied) to 10 (satisfied). is is different from the traditional VIKOR, in which the positive-ideal solution is set as the maximum of all schemes, i.e., f * j � max {f kj |k � 1, 2, . . ., K}, and the negative-ideal solution is set as the minimum of all schemes, i.e., f − j � min {f kj |k � 1, 2, . . ., K}. en, we can obtain the gap ratio, as is described in equation (13).
Step 2. Calculating the average gap E k and maximal gap Q k for prioritizing improvement. e general form of the L p-metric function is introduced as follows: where w j is the influential weight generated from the DANP. L p�1 k and L p�∞ k are separately expressed as E k and Q k , which can be calculated by equations (15) and (16): e compromise solution min k L p k indicates that the synthesized gap should be minimized. e average gap is emphasized when p is equal to one. However, Q k means the maximum gap of overall criteria in alternative k. When p is infinite, the maximal gaps should be improved by the priority.
Step 3. Exporting comprehensively evaluated values of the alternatives.
For equation (15), the best gap E * k and the worst value respectively; For equation (16), the best gap Q * k and the (17) can be simplified as follows: e range of values for λ is zero to one. λ > 0.5 means the analysis emphasizes the average gap more. λ < 0.5 indicates the analysis is more concerned about the maximum gap for priority improvement. In general, we can set λ � 0.5.

Empirical Results
Based on the above models, ten types of BEVs are carried out as shown in Table 1. e empirical results of the analytical process are as follows:

Building Evaluation Indicator System.
e online wordof-mouth of BEVs is from the Autohome website (see https://www.autohome.com.cn/), which is a relatively wellknown auto website in China. Word-of-mouth is the key for users to express their views. On the Autohome website, the word-of-mouth data reflect vehicle attributes that users are most concerned about. Because the website has strict requirements on the opinions, all the text information is accurate and of high quality.

Emotional Polarity Analysis.
First, emotional lexicon is extracted from the text that has been preprocessed. Table 2 shows the emotional words with high word frequency. It can be found that the word frequency of the emotional words, such as right, fine, smooth, not bad, not so dusty, high, comfortable, stable, and so on, is very high, which indicated that consumers' overall cognition of BEVs is relatively concentrated.
On the Autohome website, the overall image of a BEV is often measured in terms of the customers' satisfaction with the BEV. en, the emotional terms of customers' satisfaction are obtained through cluster analysis, as is shown in Table 3. e overall satisfaction of customers with the BEV is measured by the emotion words that indicate emotion. 30.6% of the total number of word-of-mouth have customers' satisfaction far greater than customers' expectations, 27.5% of the total have customers' satisfaction greater than customers' expectations, 15% of the total have customers' satisfaction equal to customers' expectations, 10.7% of the total have customers' satisfaction less than customers' expectations, and 16.2% of the total have customers' satisfaction far less than customers' expectations.
Finally, the word-of-mouth containing the above sentiment words are calculated and scored. e actual score ranges from 1 to 5 points according to the set scoring rules (using the sentiment dictionary method, transforming and rounding the results), the scores of the word-of-mouth containing the sentiment words are obtained, and the average score containing each sentiment word is calculated (the ratio of the frequency of the sentiment word to the number of word-of-mouth containing the sentiment word).
e emotional scores are shown in Table 4.

Topic Analysis.
is study identifies the consumer concerns by using the LDA model, to discover key topics from the collected text. As is shown in Table 5, the LDA model divides keywords into five topics, including dynamics, technology, safety, comfort, and cost. Specifically, the keywords in topic 1 mainly involve maximum power, max torque, top speed, and acceleration time, which reflect the dynamics. In topic 2, the keywords mainly involve driving range, charging time, and electricity consumption, which reflect the battery technology. e keywords of topic 3 are mainly related to three aspects, curb weight, braking, and operating stability, which represent the safety of BEVs. In topic 4, the keywords mainly involve four aspects, exterior and interior, space, suspension, and seats, reflecting the comfort of BEVs. e keywords in topic 5 involve three aspects: price, incentives, and after-sales cost, which are the embodiment of the cost factor.
rough the above analysis, five major topics are obtained, which are taken as the five dimensions in the indicator system. Seventeen criteria are identified by keyword segmentation. e details of the evaluation dimensions and criteria of BEVs are shown in Table 6.

Data Collection.
Based on the evaluation indicator system, two different questionnaires are designed to collect the information required for sufficient evaluation of ten BEVs.
e DEMATEL questionnaire on the relationship    [45]. en we collect the experts' results and compute the average scores of all criteria to form the initial direct-effect matrix, as is shown in Table 7. e statistical significance confidence of scores by experts is 95.13% (greater than 95%; i.e., the gap error is only 4.87%, or < 5%), which indicates consistent responses. e modified VIKOR questionnaire is distributed to customers who have bought BEVs or intend to buy BEVs to score the criteria, on an eleven-point scale ranging from 0-10. Of the 550 questionnaires distributed, 540 valid questionnaires are obtained. e internal consistency of customer ratings is tested using Cronbach's alpha [107]. e null hypothesis of the scores associated with each BEV is defined as no difference between customer ratings. e   It includes vehicle repairs, maintenance, and other expenses. [4,12] alternative hypothesis assumes that the scores are different. e Cronbach's alpha values related to these ten hypotheses are 0.894, 0.882, 0.930, 0.855, 0.838, 0.823, 0.825, 0.802, 0.750, and 0.835, respectively. e null hypothesis is significant. It is hoped that this study can help the government departments and automobile enterprises' decision-makers to effectively improve customers' purchasing decisions, thereby enhancing the BEV's development and competitiveness.

Estimating the Relationships among Dimensions and Criteria.
e DEMATEL technique is applied to obtain the cause-effect relationships and to construct the INRM among dimensions and criteria. According to the responses from 7 experts, the direct effect 17 × 17 matrix G � [a ij ] is constructed in Table 7. e normalized direct-influence matrix X in Table 8 is derived by equations (1)-(3). e total influence matrix T c of the criteria and matrix T D of the dimensions are separately listed in Tables 9-10. Furthermore, the (r i + s i ) and (r i − s i ) values of indicators can be obtained from the total influence matrix, as is shown in Table 11. e implication of (r i + s i ) presents the intensity of the influence that the ith criterion plays in the problem, and (r i − s i ) presents the size of the ith criterion's direct impact on others. When (r i − s i ) is positive, ith criterion influences other criteria. On the contrary, if (r i − s i ) is negative, ith criterion is influenced by other criteria. e INRM is drawn with (r i + s i ) as the horizontal axis and (r i − s i ) as the vertical axis, reflecting the cause-and-effect relationships between dimensions and criteria as shown in Figure 2.
As shown in Table 11, the five dimensions can be pri- us, it is important to achieve a longer driving range to ease consumers' range anxiety, thereby affecting the degree of purchase. Automobile enterprises should increase technological investment and give priority for expanding driving. Under the same circumstances, consumers usually choose new energy vehicles with a high range [54]. In terms of safety (C) dimension, curb weight (0.194) exerts a direct effect on the other criteria, including braking properties (−0.093) and operating stability (−0.101). erefore, lightweight technology should be supported and promoted to reduce curb weight and thus improve the safety of BEVs. In the comfort (D) dimension, the criterion of suspension (0.048) and car seat (0.017) strongly influence exterior and interior (−0.016), and car space (−0.050). us, optimizing the suspension to improve operating comfort is the most influential way to encourage a customer to purchase new vehicles. In the cost (E) dimension, incentives (0.072) exert a direct effect on the remaining criteria, including price (−0.033) and after-sales cost (−0.039). erefore, incentives play the most important role in consumers' decisions. e general improvement priorities can be sequenced as E2, E1, and E3.

Determining the Influence Weights.
After identifying the relationships among indicators through the DEMATEL technique, the influence weights of those indicators can be obtained by the DANP method. Initially, the unweighted super-matrix W c in Table 12 can be derived by equation (11). e weighted super-matrix W * c in Table 13 can be obtained using equation (12). Finally, the weight of each criterion is acquired by limiting the power of the weighted super-matrix.
According to the influence weights shown in Table 14, in terms of dimensions, the cost (0.400) is ranked as the most important weight, while the comfort (0.062) is ranked as the least important one. Based on the influence weights (global weights) associated with seventeen criteria, empirical results reveal that price (E1), incentives (E2), and after-sales cost (E3) are ranked as the top criteria. Concretely, price gains the highest point of 0.191, followed by incentives (0.106) and after-sales cost (0.103). Moreover, the influence weights of suspension (0.010), car seat (0.010), exterior, and interior (0.009) are relatively low, which means that these criteria have the least impact.

Obtaining the Selection Strategies and Optimal Path.
is study aims to propose the most effective strategies to improve customers' decisions for BEVs in China. In this section, the Modified VIKOR method is employed to evaluate ten models based on the opinion of customers. e range of f ik is defined from 0 to 10, where 0 is the worst level, and 10 is the desired level in the evaluation. By introducing the relative weight versus each criterion, the average gap E k and maximal gap Q k is derived. U k is also derived by setting v as 0.5. us, each BEV could be evaluated and ranked, and the results are shown in Table 15. e key to solving the problem can be identified according to this integrated index from the dimension perspective or the criteria perspective.  Mathematical Problems in Engineering e order of priority for achieving the desired level can be determined by the weights of the performance values, from high to low, and the gap values, from high to low.
As indicated in Table 15, Aion S (P3) produced by GAC NE company presents the smallest gap (0.276) and therefore ranks first, followed by the BAIC EU Series (P2; 0.   A1  A2  A3  A4  B1  B2  B3  C1  C2  C3  D1  D2  D3  D4  E1  E2  E3 Table 9: e total influence matrix of criteria.  of the scores of the BAIC EU Series (P2) in the DANP shows that the gap of the safety (C) dimension is 0.267 and that of the alliance with operating stability (C3) criterion is 0.383, constituting the largest gaps, which the BAIC EU Series should improve as a priority. e integration of the scores of BAOJUN E100 (P1) in the DANP shows that the gap of the comfort (D) dimension is 0.467 and that of the charge time (B3) criterion is 1.000, constituting the largest gaps, which BAOJUN E100 should improve as a priority. e integration of the scores of MG EZS (P4) in the DANP shows that the gap of the vehicle comfort (D) dimension is 0.276 and that of the suspension (D2) criterion is 0.367, constituting the largest gaps, which MG EZS should improve as a priority. e integration of the scores of Aeolus E70 (P5) in the     A1  A2  A3  A4  B1  B2  B3  C2  C3  D1  D2  D3  D4  E1  E2      DANP showed that the gap of the comfort (D) dimension is 0.311 and that of the alliance with suspension (D2) criterion is 0.417, constituting the largest gaps, which Aeolus E70 should improve as a priority. e integration of the scores of GEOMETRY A (P6) in the DANP shows that the gap of the cost (E) dimension is 0.285 and that of the car seat (D3) criterion is 0.367, constituting the largest gaps, which GE-OMETRY A should improve as a priority. Integration of the scores of BYD Yuan (P7) in the DANP showed that the gap of the comfort (D) dimension is 0.276 and that of the maximum power (A1) criterion is 0.400, constituting the largest gaps, which BYD Yuan should improve as a priority. e integration of the scores of BESTUNE B30EV (P8) in the DANP shows that the gap of the dynamics (A) dimension is 0.311 and that of the alliance with acceleration time (A4) criterion is 0.367, constituting the largest gaps, which BESTUNE B30EV should improve as a priority. e integration of the scores of EMGRAND EV (P9) in the DANP shows that the gap of the vehicle comfort (D) dimension is 0.380 and that of the suspension (D2) criterion is 0.433, constituting the largest gaps, which EMGRAND EV should improve as a priority. e integration of the scores of ORA R1 (P10) in the DANP shows that the gap of the dynamics (A) dimension is 0.435 and that of the acceleration time (A4) criterion is 0.617, constituting the largest gaps, which ORA R1 should improve as a priority. Furthermore, the gap values obtained by the customers reveal that improvement priority schemes are unique and comprehensive for each separate dimension as well as for the overall range of criteria. Decision-makers in government departments and automobile enterprises can easily understand the gaps where improvements are prioritised.

Discussion
e study uses fine-grained sentiment analysis and MCDM model to explore the BEVs from the customers' point of view based on five dimensions, namely, safety, technology, dynamics, comfort, and cost. e LDA model based on finegrained sentiment analysis has been applied to obtain dimensions and criteria based on the word-of-mouth data. e DANP method combining the DEMATEL technique with ANP has been used to create an INRM to identify the causal relationship and compute the influence weights of all indicators. e Modified VIKOR is not only used to rank and determine selection strategies but also to propose the optimized path for ten BEVs.
e empirical results are discussed as follows: First, using the LDA model based on fine-grained sentiment analysis, the comprehensive evaluation indicator system is constructed, including five dimensions and seventeen criteria. According to the DEMATEL cause-and-effect model, the interrelationships between each dimension and criterion are determined by INRM. In Figure 2, the degree of the dimension effect indicates that improvement priorities should be established in the following order: safety, technology, dynamics, comfort, and cost. e results further illustrate that the decision-makers should improve safety first because it has the greatest immediate network effect on the other dimensions. e structural design, materials used in collision sites, safety equipment, and other aspects may cause accidents [108,109], such as battery explosions and physical damages [57]. So, most people are concerned about safety when driving BEVs [58]. e findings also imply that a full understanding of the relationships among dimensions can enable automobile companies to grasp the main direction of the future revolution, thus enhancing consumers' recognition and enthusiasm for BEVs.
Second, after describing the dimensions, the study also discusses the criterion considered in each dimension. ey provide a higher-level mode for the improvement of BEV attributes [110]. In terms of the results for the safety (C) dimension, consumers are more concerned with the curb weight of BEVs. Combining lightweight designs with BEVs can further reduce environmental impacts [52]. In addition, the application of lightweight materials in the BEVs can improve functionality and safety [111]. Moreover, details on the causal relationship between technology, dynamics, comfort, and cost can also be derived from Figure 2. Each of the evaluation dimensions and criteria play an essential role in customers' choices of BEVs. erefore, relevant departments should evaluate all the dimensions and criteria for promoting customers' attitudes and willingness. e subsequent evaluation model can be used for the automobile industry in China.
ird, the study uses the DANP to confirm the weight of the 17 influential criteria. As shown in Table 14, price (E1) is the largest relative weight of seventeen criteria with the value  Generally, the cost of BEVs is higher than that of Internal Combustion Engine Vehicles (ICEVs), which is mainly due to the high research and development (R&D) expense and production costs, especially batteries. Hence, in addition to technology improvements, automakers can develop dedicated platforms and modular production platforms for BEVs to reduce costs as much as possible. e lowest priority is exterior and interior (0.009). Consequently, the design of the exterior and interior has the least impact on consumers' decisions. e results indicate that the priority of criterion improvement from top to bottom and the improvement of the most influential criteria would provide the largest effects. It is possible that improving different criteria will strongly influence the results directly and indirectly, and improving the most influential criteria provides the most substantial effects. Fourth, the comparison results of ten BEVs in Table 15 show that Aion S (P3) is the best choice. In order to verify the feasibility and validity of this hybrid method, we combine the criteria importance through inter-criteria correlation (CRITIC) method with modified VIKOR for ranking the ten BEVs. e CRITIC method, proposed by Diakoulaki et al. [112], is used to calculate attributes' weights by standard deviation. As derived in Table 16, although the sorting of the BEVs based on different MCDM methods varies partly, Aion S (P3) is still the optimal choice. is means that the proposed method is effective at ranking and selecting BEV alternatives.
Finally, from the perspective of dimension, the comfort (D), featuring the largest gap value (0.467, 0.276, 0.311, 0.321, 0.380) with cars of BAOJUN E100 (P1), MG EZS (P4), Aeolus E70 (P5), BYD Yuan (P7), and EMGRAND EV (P9), should be the priority for improvement if banking managers wish to enhance service innovation. e dynamics (A) dimension, constituting the largest gap value (0.191, 0.311, 0.435) with cars of Aion S (P3), BESTUNE B30EV (P8), and ORA R1 (P10), should improve as a priority. In the BAIC EU Series (P2), the safety (C) dimension features the largest gap with the value of 0.267, which should be a top priority for improvement to achieve aspiration levels. In the alternative of GEOMETRY A (P6), the cost (E) dimension constitutes the largest gap with the value of 0.285. According to the same rule, the priority improvement following sustainable development context can be sequenced in the criteria for each BEV. Given these critical empirical findings, our analysis results, as holistically formulated in Table 17, present the optimized path for different BEV models. Hence, relevant departments can not only use this method to define gaps but also enhance customers' purchase willingness in China based on the priorities of influence weights or gap values.

Conclusions
BEVs are not only the mainstream of the future automobile industry but also the top priority for implementing national energy policy and achieving sustainable development of the automobile industry. However, a large proportion of consumers still hold a wait-and-see attitude [58,60], China's BEV market has always been a policy-oriented model. ere is still much room for improvement in consumers' purchase willingness. is study creates a novel hybrid model combining fine-grained sentiment analysis and an MCDM model to assess BEV alternatives from the customers' point of view. In summary, the main results of this study are as follows: (1) e comprehensive evaluation indicator system, including five dimensions (dynamics, technology, safety, comfort, and cost) and seventeen criteria, is obtained by the LDA model based on fine-grained sentiment analysis. (2) Safety (C) is the most influential dimension, followed by technology (B), dynamics (A), comfort (D), and cost (E). Safety improvements can lead to changes in other dimensions. us, government and providers can improve BEV attributes more flexibly and accurately based on the identification of cause-effect relationships between the dimensions and the actual situation of BEVs.
(3) e DANP analysis shows that price (E1), incentives (E2), and after-sales cost (E3) are ranked as the top criteria, and they all belong to the cost dimension. Moreover, the lowest influential weights involve suspension (D2), car seat (D3), exterior, and interior (D4), and they all belong to the comfort dimension. Government and providers can determine the order of improvement based on the weight. (4) e modified VIKOR results indicated that the selection strategies of ten BEVs are P3 > P2 > P4 > P8 > P6 > P9 > P5 > P7 > P10 > P1, which presented that P3 (Aion S) is the best choice. In addition, there is still room for improvement in vehicle attributes, and the indicator gaps and optimized path are different for each BEV. Taking the BEV named Aion S as an example, it has the highest integrated scores, but it still falls short of aspiration levels. According to the gaps ranging from large to small, the first step of the optimized path should improve the dynamics. (5) In this study, we find some attributes that have been neglected in previous studies, such as maximum power, max torque, and after-sales cost, affect customers' satisfaction. With the gradual maturity of BEV technology, the effects of more detailed vehicle attributes should be considered [30]. erefore, the widespread promotion of BEVs cannot rely solely on policy incentives and the government should take active measures to promote automobile companies to achieve the urgent improvements of BEV attributes [113,114].
However, there are some limitations in this study, which need to be solved in future research. First, due to the differences in consumption concepts and the influence of psychological factors, consumers can be divided into different levels for further analysis. Second, based on the complexity of things and the ambiguities of subjective judgments, interval numbers and fuzzy numbers can be introduced in future research on consumers' adoption of BEVs. Finally, some forecasting methods can be used to predict the motivation of consumers.