A new multi-attribute decision model built on online reviews is offering a more systematic way to compare battery electric vehicle models, as consumers face a rapidly expanding and increasingly complex electric vehicle market. The research proposes a probabilistic linguistic framework that converts free-text customer comments into structured decision data, weighs information sources by volume and reliability, adjusts attribute weights through prospect theory, and ranks alternatives with an extended MARCOS method. The approach is designed to help buyers make sense of thousands of electric vehicle reviews, while giving manufacturers and online platforms a clearer view of which product attributes shape consumer choice.
The work arrives at a time when the electric vehicle sector is becoming central to China’s transport and energy transition. According to data cited in the research, China’s total energy consumption reached 5.96 billion tonnes of standard coal in 2024, up 4.3 percent year on year. Coal consumption rose 1.7 percent, crude oil consumption fell 1.2 percent, natural gas consumption increased 7.3 percent, and electricity consumption grew 6.8 percent. The transport sector, including highways, railways, aviation and water transport, has accounted for roughly 10 to 15 percent of national energy consumption in recent public data, while green electricity’s share in transport energy use remains below 2 percent. Those figures underline why electric vehicle adoption is viewed as a strategic lever for energy security and environmental protection.
Policy support has also accelerated the electric vehicle transition. The research traces the policy chain from the 2012 planning document on energy-saving and new-energy vehicle industries to the 2015 “Made in China 2025” strategy, which placed new-energy vehicles among core development fields. In 2021, the National Development and Reform Commission signaled continued support for electric vehicle infrastructure and industrial development. Those measures have helped China’s electric vehicle market move from early adoption toward mainstream consideration, with consumers now facing a much wider set of battery electric vehicle choices than in earlier years.
The scale of the electric vehicle market is substantial. The research cites figures showing that China’s automobile production and sales reached 31.282 million and 31.436 million vehicles in 2024, up 3.7 percent and 4.5 percent respectively. New-energy vehicle production and sales reached 12.888 million and 12.866 million, up 34.4 percent and 35.5 percent. New-energy vehicle sales accounted for 40.9 percent of total new vehicle sales, 9.3 percentage points higher than the previous year. Battery electric vehicle production and sales reached 7.758 million and 7.719 million, up 15.7 percent and 15.5 percent, and battery electric vehicles accounted for 60 percent of new-energy vehicle sales, down 10.4 percentage points from the prior year but still representing a major share. For consumers, that growth means more electric vehicle options, more configurations and more reviews to compare.
Online reviews have become one of the most important information channels in this decision environment. The research cites a 2021 survey of the new-energy vehicle market in which more than 85 percent of potential buyers said they relied on online reviews to evaluate performance, range, charging convenience and after-sales service before making a final decision. Positive reviews were associated with a 4.2-fold increase in purchase likelihood, while negative reviews reduced purchase intention by nearly three times. Those figures help explain why electric vehicle shoppers often move from showroom comparisons to review platforms, forums and owner communities before committing to a purchase.
Yet the same abundance of information creates a new problem. A single electric vehicle model may have thousands of comments across different platforms, and each comment may touch on technical specifications, driving experience, charging, interior design, space, price and comfort. Consumers have different priorities, and the same attribute may carry different weight depending on risk attitude and personal preference. The research argues that the challenge is not simply to collect more electric vehicle reviews, but to integrate and quantify them in a way that supports transparent comparison. The proposed model responds to that challenge by treating electric vehicle purchase as a multi-attribute decision problem and by converting review text into probabilistic linguistic information.
1. Why electric vehicle choice has become a data-rich but complex decision
The research describes electric vehicle purchase as a two-stage decision process. In the first stage, consumers identify battery electric vehicle models that meet their basic needs. In the second stage, they compare those alternatives across important attributes and form a ranking before making a purchase. Because consumers often lack complete technical knowledge and hands-on experience with every electric vehicle model, they rely heavily on reviews written by earlier buyers. Those reviews can cover range, charging, handling, appearance, interior, configuration, space, price and comfort, but they are unstructured, unevenly distributed across platforms and sometimes contradictory.
The electric vehicle market adds further complexity. Product updates are frequent, battery and software features change quickly, and new models can enter the market with limited review histories. A popular electric vehicle may have thousands of comments on one platform and only a few dozen on another. Some platforms may attract more detailed owner reviews, while others may contain shorter comments or repeated content. The research therefore treats electric vehicle review integration as more than a text-mining exercise: it requires source weighting, attribute weighting and a ranking method that can handle uncertainty without losing interpretability.
The study also identifies a gap in existing research. Online-review-based product ranking has been applied extensively to hotels, restaurants, tourism destinations and other service products, but electric vehicle purchase decisions have received less attention. Studies on electric vehicle consumer intention and sales forecasting are plentiful, yet research that uses online reviews to guide electric vehicle model selection remains relatively limited. In addition, many existing probabilistic linguistic studies use coarse sentiment categories, such as positive, neutral and negative, which may not capture the finer distinctions that consumers express in electric vehicle reviews. The research argues that a more detailed seven-level probabilistic linguistic representation can preserve more information from review text.
A further gap concerns ranking methods. Probabilistic linguistic term sets have been combined with TOPSIS, TODIM, VIKOR, BWM, PROMETHEE, ELECTRE and LINMAP, but rarely with MARCOS. The research states that MARCOS can remain stable when handling many attributes and alternatives and can provide robust rankings by considering both proximity to the positive ideal solution and distance from the negative ideal solution. Extending MARCOS to a probabilistic linguistic environment is therefore presented as a way to enrich electric vehicle decision support and to improve the reliability of electric vehicle rankings.
2. A seven-level probabilistic language framework turns text into decision data
The proposed electric vehicle decision model begins with data collection. Python-based web crawling was used to gather text reviews from two automotive platforms, Autohome and Dongchedi, for selected battery electric vehicle models. The dataset covered reviews published in the year up to 7 December 2024. The research reports 3,672 raw text reviews and, after cleaning duplicate and invalid entries, 3,658 valid reviews. The cleaning process removed repeated comments, noise, symbols and other content that could interfere with text analysis. The resulting review pool was then used to build the electric vehicle evaluation matrix.
Text preprocessing was carried out with ROST Content Mining software. The research used Chinese word segmentation, custom dictionaries and stop-word filtering to improve accuracy. Because online electric vehicle comments are often colloquial, the study added custom terms to the segmentation dictionary and removed meaningless particles, degree adverbs and brand or model names from word-frequency statistics. The purpose was to prevent common but non-informative words from dominating the analysis and to focus on product attributes that consumers actually discuss when evaluating an electric vehicle.
After word-frequency analysis and attribute-word merging, the research identified eight attributes with the highest discussion intensity: handling, range, appearance, interior, configuration, space, price and comfort. Related words were grouped under each attribute. For example, driving, operation and control terms were merged into handling; wind noise, tire noise and sound insulation terms were merged into comfort; and battery, mileage and long-distance terms were merged into range. This step transformed scattered electric vehicle comments into a structured attribute set suitable for multi-attribute decision analysis.
Sentiment analysis was then performed with Emotion Calculator. The software produced a Likert-scale sentiment value for each attribute word in each review, with values ranging from -3 to 3. The research mapped those values to a seven-level probabilistic linguistic term set: -3 to absolute negative, -2 to very negative, -1 to low negative, 0 to neutral, 1 to low positive, 2 to very positive and 3 to absolute positive. By counting how often each linguistic term appeared for a given model and attribute, the study constructed a probabilistic linguistic decision matrix. This step is important for electric vehicle review analysis because it preserves the direction and intensity of consumer sentiment rather than reducing every comment to a simple positive or negative label.
The review volume after cleaning is summarized below. The figures show how electric vehicle review data can be uneven across platforms and models, which is one reason the research introduces a source-weighting mechanism.
Review volume after cleaning by model and source
| Electric vehicle model | Autohome reviews | Dongchedi reviews | Total valid reviews |
|---|---|---|---|
| AION-S | 65 | 87 | 152 |
| AION-Y | 290 | 105 | 395 |
| Model Y | 511 | 192 | 703 |
| Dolphin | 1396 | 69 | 1465 |
| Yuan PLUS | 839 | 104 | 943 |
| Total | 3101 | 557 | 3658 |
The eight attributes form the evaluation system for the electric vehicle case. The research does not treat all attributes as equally important. Instead, it derives attribute weights from the structure of the review data, with an adjustment based on prospect theory. This is significant for electric vehicle purchase decisions because consumers may weigh configuration, comfort, handling, range, price and other attributes differently, and those weights should reflect both data dispersion and behavioral risk attitude.
Attributes used in the electric vehicle review model
| Attribute | Focus in electric vehicle reviews |
|---|---|
| Handling | Power, driving, acceleration, braking, steering and control |
| Range | Battery, mileage, long-distance use and electricity consumption |
| Appearance | Body, styling, lights, paint, doors and exterior form |
| Interior | Seats, screens, center console, instrument panel and roof |
| Configuration | Functions, systems, vehicle software, air conditioning, tires and voice control |
| Space | Rear seats, trunk, storage, front row and cabin room |
| Price | Cost, price level, value for money, discounts and savings |
| Comfort | Wind noise, tire noise, odor, insulation, ride comfort, vibration and noise |
3. Weighting sources and attributes: from webpage reliability to prospect-theory-adjusted ITARA
Because electric vehicle reviews come from different websites, the research proposes a webpage weighting method that considers both review quantity and review reliability. The logic is straightforward: a platform with more electric vehicle reviews may provide a broader information base, but volume alone is not enough. If reviews on a platform are highly inconsistent or polarized, the reliability of that platform’s information may be lower. The model therefore combines a review-volume share with a reliability measure based on the deviation of probabilistic linguistic evaluations.
The review-volume share is calculated as the number of reviews on a website divided by the total number of reviews across websites. The reliability measure is based on the deviation of probabilistic linguistic term sets: when the deviation is larger, consumer evaluations are treated as more dispersed, and the reliability weight is lower. The two components are then combined through a balancing parameter. In the case analysis, the parameter is set at 0.5, meaning that review quantity and review reliability are treated as equally important. The resulting weights are 0.72 for Autohome and 0.28 for Dongchedi.
Webpage weights used to integrate electric vehicle reviews
| Source | Review-volume share | Reliability score | Final weight |
|---|---|---|---|
| Autohome | 0.8477 | 0.5922 | 0.72 |
| Dongchedi | 0.1523 | 0.4078 | 0.28 |
After the webpage weights are applied, the individual probabilistic linguistic decision matrices are aggregated into a comprehensive electric vehicle decision matrix. The research then turns to attribute weighting. The ITARA method assigns weights by examining the dispersion of attribute values and by using an indifference threshold. If alternatives are very close on an attribute, that attribute receives less weight because it does not distinguish among electric vehicle models effectively. If an attribute shows large differences among alternatives, it receives more weight.
The study improves ITARA by changing how the indifference threshold is determined. In the original ITARA method, the threshold is based on the positive ideal solution. The research argues that this may overlook decision-maker psychology. Drawing on prospect theory, the study treats the distance from the positive ideal solution as a loss and the distance from the negative ideal solution as a gain. The value function then reflects risk attitudes: decision-makers tend to be risk-seeking when facing losses and risk-averse when facing gains. The prospect theory parameters used in the case are alpha equal to 0.88, beta equal to 0.88 and theta equal to 2.25.
The final indifference thresholds in the electric vehicle case are 0.0225 for handling, 0.0371 for range, 0.0376 for appearance, 0.0622 for interior, 0.0113 for configuration, 0.0348 for space, 0.0479 for price and 0.0522 for comfort. The ordered distances between alternatives are then compared with these thresholds. When the distance is below the threshold, the difference is treated as too small to contribute to electric vehicle ranking. When the distance exceeds the threshold, the difference contributes to the attribute weight.
The resulting attribute weights show a clear pattern. Configuration receives the highest weight at 0.3561, followed by comfort at 0.2826 and handling at 0.2087. Range, space, appearance and price receive much smaller weights, while interior receives a weight of zero because the ordered differences among alternatives do not exceed the threshold. For electric vehicle buyers, this means that in the specific review data and parameter setting, differences in configuration, comfort and handling are more useful for distinguishing among the candidate models than differences in interior or price.
Attribute weights derived for the electric vehicle case
| Attribute | Weight |
|---|---|
| Handling | 0.2087 |
| Range | 0.0666 |
| Appearance | 0.0402 |
| Interior | 0.0000 |
| Configuration | 0.3561 |
| Space | 0.0429 |
| Price | 0.0029 |
| Comfort | 0.2826 |
4. Extending MARCOS to a probabilistic linguistic environment for final electric vehicle ranking
The final ranking stage uses an extended MARCOS method. MARCOS is a compromise-based multi-criteria decision method that evaluates alternatives relative to both a positive ideal solution and a negative ideal solution. The research extends this method to probabilistic linguistic information so that electric vehicle reviews can be used directly in the ranking process. The extended procedure defines the positive and negative ideal solutions, constructs an extended probabilistic linguistic decision matrix, aggregates information with the probabilistic linguistic weighted average operator, defuzzifies the results through a score function, calculates utility degrees relative to the ideal solutions, and then computes a utility function for final ordering.
The utility function is central to the method. It quantifies how close an electric vehicle alternative is to the positive ideal solution and how far it is from the negative ideal solution. A model that is both close to the positive ideal and far from the negative ideal receives a higher utility value. Compared with a simple linear weighted sum, this approach is intended to improve the reliability of the final electric vehicle ranking. The research also notes that MARCOS has a conceptual connection to TOPSIS but adds a utility assessment structure that can make rankings more resistant to reversal in complex decision settings.
In the case analysis, five battery electric vehicle models are compared: AION-S, AION-Y, Model Y, Dolphin and Yuan PLUS. The probabilistic linguistic decision matrix is aggregated using the attribute weights, and the score values are calculated for each alternative. The positive ideal solution has a score of 2.4255, while the negative ideal solution has a score of 1.4755. The alternative scores are 1.4833 for AION-S, 2.3750 for AION-Y, 2.1393 for Model Y, 2.3247 for Dolphin and 2.2827 for Yuan PLUS.
Score values for the electric vehicle alternatives
| Alternative | Score value |
|---|---|
| AION-S | 1.4833 |
| AION-Y | 2.3750 |
| Model Y | 2.1393 |
| Dolphin | 2.3247 |
| Yuan PLUS | 2.2827 |
| Positive ideal solution | 2.4255 |
| Negative ideal solution | 1.4755 |
The utility degrees relative to the positive and negative ideal solutions, together with the final utility function values, provide the basis for the electric vehicle ranking. AION-Y has the highest positive utility degree and the highest final utility function value. Dolphin follows, then Yuan PLUS, then Model Y, and finally AION-S.
Utility degrees and final utility function values for the electric vehicle alternatives
| Alternative | Utility degree relative to positive ideal | Utility degree relative to negative ideal | Final utility function value |
|---|---|---|---|
| AION-S | 0.6116 | 1.0053 | 0.7228 |
| AION-Y | 0.9792 | 1.6096 | 1.4605 |
| Model Y | 0.8820 | 1.4499 | 1.2553 |
| Dolphin | 0.9585 | 1.5756 | 1.4163 |
| Yuan PLUS | 0.9411 | 1.5471 | 1.3795 |
5. Case results: which electric vehicle led the ranking?
The case analysis produces a clear final ordering: AION-Y ranks first, Dolphin ranks second, Yuan PLUS ranks third, Model Y ranks fourth and AION-S ranks fifth. The result is based on the aggregated online reviews, the webpage weights, the prospect-theory-adjusted attribute weights and the extended MARCOS utility function. The research presents this ranking as an illustration of how electric vehicle purchase decisions can be supported by review-driven multi-attribute analysis rather than by raw review counts or isolated ratings.
The ranking should be understood within the specific dataset and parameter setting. The attribute weights show that configuration, comfort and handling play a dominant role in differentiating the candidates. Range, space, appearance and price have smaller weights, while interior is not a distinguishing attribute in this case because the differences among alternatives do not exceed the indifference threshold. For an electric vehicle buyer who prioritizes configuration and comfort, the ranking offers a structured summary of review sentiment across those dimensions. For a buyer who prioritizes price or range, the same framework can be adjusted through attribute weighting or through sensitivity analysis.
The research also notes that AION-S has a score close to the negative ideal solution, which explains its last-place position. AION-Y, by contrast, is close to the positive ideal solution and far from the negative ideal solution, producing the highest utility function value. Dolphin and Yuan PLUS occupy strong middle-to-high positions, while Model Y sits between them and AION-S in the final ranking. These results are not presented as a universal verdict on electric vehicle quality; they are the outcome of a transparent decision model applied to a particular review sample, attribute set and weighting scheme.
The model’s value lies in its ability to integrate large volumes of electric vehicle review text without discarding sentiment intensity. A seven-level probabilistic linguistic representation allows a review that says a model is “very positive” on configuration to be treated differently from one that is only “slightly positive.” The webpage weighting mechanism then prevents a small or inconsistent review source from having disproportionate influence. The improved ITARA method allows the attribute weights to reflect how strongly the alternatives differ on each attribute and how decision-makers may respond to gains and losses. The extended MARCOS method then combines those elements into a final electric vehicle ranking.
6. Sensitivity and comparison: robust electric vehicle decision support
The research tests how sensitive the attribute weights are to the parameters of the prospect theory value function. The loss aversion coefficient is varied from 1 to 5, the gain attitude coefficient from 0 to 1, and the loss attitude coefficient from 0.5 to 1. The results show that attribute weights change noticeably as the parameters change, which the research interprets as evidence that the model can adapt to different consumer risk preferences.
When the loss aversion coefficient increases, the weight of configuration rises steadily and reaches 1 when the coefficient reaches 5, while the other attributes fall to zero. The weights of handling and comfort first rise and then decline after the coefficient passes 3. Appearance, interior and price drop to zero when the coefficient reaches 3, while range and space drop to zero when it reaches 4. In practical terms, a consumer who is extremely loss-averse may focus almost entirely on configuration, whereas a consumer with moderate loss aversion may give more balanced attention to handling and comfort.
The gain attitude coefficient produces a different pattern. Comfort is most sensitive: when the coefficient is 0.2, comfort receives a weight of 1, and its weight declines as the coefficient increases. Handling and configuration are also sensitive. When the coefficient reaches 1, both have weights above 0.2, and configuration becomes the largest attribute weight. This suggests that a consumer who places greater emphasis on gains may give more weight to handling and configuration, while a consumer with lower sensitivity to gains may concentrate on comfort.
The loss attitude coefficient also changes the weighting structure. Configuration is most sensitive: when the coefficient is 0.5, configuration receives a weight of 1, and its weight declines as the coefficient rises. Handling and comfort are also sensitive. As the coefficient increases, their weights first rise and then decline slightly. When the coefficient reaches 1, comfort has the largest weight among the eight attributes. Other attribute weights increase after the coefficient passes 0.7. The research concludes that the weighting model has high sensitivity and can flexibly reflect different consumer priorities in electric vehicle purchase decisions.
The study further compares the proposed probabilistic linguistic MARCOS ranking with two classical methods: probabilistic linguistic TOPSIS and probabilistic linguistic TODIM. The TOPSIS comparison produces the ranking AION-Y, Dolphin, Model Y, Yuan PLUS and AION-S, which is broadly consistent with the proposed method; only the positions of Yuan PLUS and Model Y are swapped. The TODIM comparison produces exactly the same ranking as the proposed method: AION-Y, Dolphin, Yuan PLUS, Model Y and AION-S. The research treats this as evidence that the extended MARCOS method is reliable in the probabilistic linguistic electric vehicle context.
Ranking comparison across decision methods
| Method | Electric vehicle ranking |
|---|---|
| Proposed probabilistic linguistic MARCOS | AION-Y > Dolphin > Yuan PLUS > Model Y > AION-S |
| Probabilistic linguistic TOPSIS | AION-Y > Dolphin > Model Y > Yuan PLUS > AION-S |
| Probabilistic linguistic TODIM | AION-Y > Dolphin > Yuan PLUS > Model Y > AION-S |
The research argues that the proposed method has two practical advantages over TOPSIS. First, it can be more intuitive and simpler in calculation because it does not require the same standardization steps. Second, it builds a utility function for both the positive and negative ideal solutions, which can make the final electric vehicle ranking more resistant to rank reversal. Compared with TODIM, the proposed method reduces computational complexity. TODIM requires pairwise comparisons across alternatives and attributes, which can become cumbersome when many electric vehicle models are involved. The proposed method also avoids relying on a loss aversion coefficient in the final ranking stage, which makes the result more objective and stable.
7. What the model means for electric vehicle buyers, manufacturers and platforms
For consumers, the model offers a way to reduce the trial-and-error cost of electric vehicle shopping. Instead of reading thousands of comments one by one, a buyer can use a structured ranking that reflects review sentiment, source reliability and attribute importance. The model does not replace test drives or expert advice, but it can help a buyer identify which electric vehicle models deserve closer attention and which attributes are driving the differences among them. For a consumer comparing electric vehicle models across brands and platforms, this kind of decision support can make the review process more efficient and more transparent.
For manufacturers, the research suggests that online reviews can reveal which electric vehicle attributes consumers discuss most and where models differ most sharply. In the case analysis, configuration, comfort and handling emerge as the attributes with the greatest discriminating power. That does not mean range, price or appearance are unimportant; rather, it means that in this review sample and weighting scheme, those attributes did not differentiate the candidate electric vehicle models as strongly. Manufacturers can use such findings to prioritize product improvements, communication strategies and configuration choices. A company that understands how consumers weigh electric vehicle attributes can better align development and marketing with actual owner experience.
For online automotive platforms, the research points to the value of review-based ranking mechanisms. Electric vehicle shoppers rely on platforms to compare models, but raw review counts and star ratings can be misleading when review volume is uneven or when comments focus on different attributes. A ranking system that accounts for review reliability and attribute importance can improve the usefulness of platform recommendations. It can also help platforms design better comparison tools, summarize owner sentiment and present electric vehicle information in a way that supports purchase decisions rather than overwhelming buyers with unstructured text.
From a policy perspective, the research links electric vehicle purchase support to the broader goals of energy transition and industrial development. If consumers can make more informed electric vehicle choices, and if manufacturers can respond more accurately to owner feedback, the electric vehicle market may develop more efficiently. The research does not claim that a single ranking model can resolve all market information problems, but it does show how review analytics can contribute to a more consumer-responsive electric vehicle ecosystem.
8. Limitations and the next step for electric vehicle review analytics
The research acknowledges several limitations. The first is the cold-start problem. Newly launched electric vehicle models may have limited online reviews, which can weaken the reliability of review-based ranking. In such cases, the model may not have enough information to produce a fully stable comparison, and traditional evaluation methods or expert input may still be needed. As more owner reviews accumulate, the model’s electric vehicle assessments are expected to become more reliable.
The second limitation concerns consumer psychology. The study incorporates prospect theory to reflect risk attitude toward gains and losses, but electric vehicle purchase decisions may also involve regret, fairness concerns, social trust and other psychological factors. Future research could extend the model by incorporating additional behavioral mechanisms. The third limitation is the use of text reviews alone. Online ratings also contain useful information, and combining text reviews with rating data could reduce the bias associated with a single information source and improve the robustness of electric vehicle decision support.
The research also points to future methodological work. The probabilistic linguistic framework could be tested with more electric vehicle models, more review platforms and more diverse consumer groups. The webpage weighting method could be refined to account for platform-specific review behavior, and the attribute weighting method could be extended to incorporate group preferences. The extended MARCOS approach could be compared with additional multi-attribute methods in other electric vehicle contexts. Such work would help determine whether the current findings hold across different markets, product segments and review environments.
Even with those limitations, the study offers a coherent template for electric vehicle purchase decision support. It begins with online review collection, moves through text preprocessing and sentiment quantification, builds a probabilistic linguistic decision matrix, integrates multiple review sources, adjusts attribute weights through prospect theory, and produces a final ranking with an extended MARCOS method. The case analysis shows that the approach can generate a clear electric vehicle ranking and that the ranking is broadly consistent with classical methods. For a market as dynamic and information-rich as the electric vehicle sector, that combination of transparency, flexibility and behavioral awareness is likely to remain relevant.
The broader message is that electric vehicle purchase decisions are no longer just about specifications on a brochure. They are shaped by owner experience, online discussion and the way information is weighted. As the electric vehicle market continues to expand, tools that can transform review text into reliable decision insight may become an important part of how consumers choose an electric vehicle, how manufacturers design an electric vehicle and how platforms recommend an electric vehicle. The research positions probabilistic linguistic multi-attribute decision-making as one such tool, and its electric vehicle case study suggests that the approach can support more structured, more transparent and more consumer-oriented electric vehicle comparisons.