Abstract
This study proposes an innovative evaluation framework that integrates deep learning with multi-attribute decision-making (MADM) methods to enhance the scientific rigor and accuracy of image evaluation of ice and snow tourism destinations. Compared to traditional evaluation approaches, this framework effectively processes unstructured textual data and conducts comprehensive assessments across multiple dimensions. The study innovatively designs a text feature extraction model based on the Bidirectional Encoder Representations from Transformers (BERT)-Convolutional Neural Network (CNN). Meanwhile, MADM methods are introduced for attribute weight allocation and decision optimization. The model employs BERT for in-depth semantic analysis of tourist reviews, utilizes CNN to extract local textual features, and combines MADM methods to generate comprehensive scores. In the study, the optimized model demonstrates a high consistency, achieving a consistency ratio of only 0.03 in the facilities and services theme. Moreover, this model significantly outperforms the Robustly Optimized Bidirectional Encoder Representations from Transformers Approach (RoBERTa), with a consistency ratio of 0.06. Regarding priority stability, the optimized model reaches 0.91 in comprehensive experience themes. In the aspect of computing time, the inference time of the optimized model is 0.14 s in the facilities and services theme. The experimental results indicate that the optimized model performs well in dealing with complex unstructured text data while showing high efficiency and stability in weight allocation and multidimensional decision-making tasks. Therefore, this study contributes meaningfully to the research in the image evaluation field for ice and snow tourism destinations. It also provides a vital theoretical basis and practical tools for tourism image optimization, precise marketing, and scientific management.
Introduction
Research background and motivations
In recent years, ice and snow tourism has gradually become a significant component of global tourism. Particularly in regions and countries with abundant winter resources, the market scale of ice and snow tourism continues to expand1. However, with increasing market competition, how to effectively shape and disseminate the image of ice and snow tourism destinations to attract more tourists has become a pressing issue for local governments and tourism enterprises. The image of a tourism destination is not only a critical factor influencing tourists’ choices but also the core of destination branding and differentiated development. In the ice and snow tourism field, the construction of destination images is influenced by various complex factors, including natural resources, cultural context, service quality, and tourist experiences2,3,4. Meanwhile, significant advancements in artificial intelligence, particularly in natural language processing (NLP), have demonstrated the powerful capabilities of deep learning (DL) models in text feature extraction5. Convolutional Neural Network (CNN) has also shown excellent performance in multi-dimensional feature aggregation and classification tasks, efficiently handling large-scale unstructured data, especially textual data.
The innovation of this study lies in leveraging the DL capabilities of the Bidirectional Encoder Representations from Transformers (BERT)-CNN model and integrating multi-attribute decision-making (MADM) methods to construct a scientific and systematic framework. This framework aims to evaluate the image of ice and snow tourism destinations and provide theoretical and practical support for destination branding and market competitiveness enhancement. By combining the semantic understanding capabilities of the BERT model with the feature extraction and classification abilities of CNN, this study overcomes the limitations of traditional methods in processing unstructured textual data. Meanwhile, it proposes a more efficient and comprehensive approach to destination image evaluation. Through this framework, local governments and tourism enterprises can more accurately understand tourists’ needs and perceptions, thus formulating more effective marketing strategies.
Research objectives
The research objective of this study is to construct a text analysis model based on BERT-CNN and realize efficient feature extraction and classification of text data related to ice and snow tourism. Combined with the MADM method, the key attributes of the destination image are systematically evaluated to ensure the evaluation results’ scientific rigor and accuracy.
Literature review
Tourism destination image is a significant theme in tourism academic research. Scholars have conducted extensive explorations in this area, covering the connotation of destination image, construction methods, and its impact on tourist decision-making. However, despite numerous studies, existing literature still exhibits certain limitations. These limitations primarily include the singularity of research methods and the constraints of analytical dimensions, particularly when dealing with large-scale unstructured data and multi-dimensional attribute evaluations.
Li et al. (2024) demonstrated that the image of a tourism destination was shaped by both tourists’ subjective perceptions and the objective attributes of the destination. These perceptions significantly influenced tourists’ travel decisions and destination loyalty6. They explored the subjective and objective factors of the destination image. However, it primarily relied on traditional questionnaire surveys and qualitative analysis methods, failing to fully leverage the potential of large-scale unstructured data in destination image evaluation. Additionally, they did not sufficiently address the interactions among complex multi-dimensional attributes, limiting its ability to comprehensively reveal the diverse factors influencing destination image. Borlido and Kastenholz (2023) proposed an evaluation index system for ice and snow tourism destination images based on questionnaire surveys and text analysis. This system included dimensions such as resource attractiveness, service quality, and cultural atmosphere, providing a valuable reference framework for related research7. Nevertheless, while this index system offered a framework for ice and snow tourism, it remained highly subjective. Moreover, it did not adequately account for the dynamic changes in tourists’ weighting of different attributes during decision-making, making it unsuitable for complex and dynamic market environments. In the MADM field, Hamdy et al. (2024) introduced a tourism destination evaluation framework by combining the Analytic Hierarchy Process (AHP) and Technique for Order Preference by Similarity to Ideal Solution (TOPSIS). This framework successfully provided a quantitative basis for destination image optimization8. Despite its effectiveness in attribute weight allocation and comprehensive scoring, this method still faced limitations in dynamically adjusting weights and adapting to changes. Particularly in the highly dynamic and variable market environment of ice and snow tourism, traditional AHP and TOPSIS methods struggled to address the complexity of evolving tourist needs and preferences. Zakiah et al. (2023) developed a multi-level comprehensive evaluation model for tourism destinations using fuzzy comprehensive evaluation, highlighting the advantages of MADM methods in weight allocation and attribute interaction analysis9. This model showed strengths in handling complex multi-level evaluation tasks. However, its reliance on fixed attribute weights and fuzzy rules limited its ability to dynamically adapt to rapid changes in tourist sentiment and market conditions.
Over the years, the application of NLP technology in tourism research has gradually increased, particularly in processing unstructured textual data, where NLP has demonstrated its unique advantages. Pitakaso et al. (2024) found that the BERT model exhibited remarkable strengths in text sentiment analysis and user behavior research. This effectively uncovered deep-level information from tourist-generated content data10. Although the BERT model possessed powerful semantic feature extraction capabilities, it still faced challenges such as high computational resource consumption and low efficiency when processing long texts. Additionally, when used independently, BERT did not consider how to integrate with other DL models to enhance feature extraction capabilities. Guerreiro et al. (2024) applied CNN to social media text analysis. They discovered that CNN performed efficiently in feature aggregation and classification tasks, making it particularly suitable for handling large-scale tourist review data11. However, while CNN effectively processed local features of textual data, it could not still extract global semantic information, failing to fully reflect complex tourist perceptions and needs. Therefore, combining BERT and CNN to compensate for each other’s limitations became a key strategy for improving model performance.
In summary, although existing research has made significant progress in destination image evaluation, the limitations of current literature remain evident. Traditional evaluation methods often rely on manual annotation and fixed attribute weights, lacking adaptability to changes in tourist sentiment and market dynamics. NLP technologies, particularly the integration of BERT and CNN, offer new technical pathways to address these issues. However, effectively integrating multi-dimensional attribute information, improving computational efficiency, and better reflecting the dynamic needs of tourists remain critical directions for future research. This study combines the BERT-CNN model with MADM methods to provide a more accurate and efficient framework for evaluating the image of ice and snow tourism destinations, addressing the limitations of existing approaches.
Research model
MADM framework
MADM is a systematic method widely used in complex decision-making problems, especially suitable for the multi-dimensional evaluation of tourism destination images. The core idea of this method is to quantitatively analyze the importance of multiple attributes and comprehensively evaluate the advantages and disadvantages of the target, thus offering a scientific basis for decision-making12,13,14. In constructing the image of ice and snow tourism destinations, the MADM framework can identify the key attributes that affect the destination image and evaluate different destinations’ comprehensive performance15. This study selects AHP and TOPSIS as the core methods of the MADM framework. The advantage of AHP lies in its strong logic and is suitable for dealing with the weight distribution of subjective attributes16,17,18. The merit of TOPSIS is that the calculation process is simple, and the results are easy to explain. Hence, this study designs a MADM evaluation framework for the image of ice and snow tourism destinations, as exhibited in Table 1:
Table 1 Evaluation framework of ice and snow tourism destination image.
Full size table
To sum up, the MADM framework provides a systematic quantitative evaluation method for the image construction of ice and snow tourism destinations. In the follow-up research, this study further optimizes the data processing and evaluation process with the BERT-CNN model, improving the scientific and practical decision-making framework.
BERT-CNN model design
This study designs a model based on BERT and CNN to fully utilize the deep semantic information contained in the content generated by tourists and improve the accuracy of image evaluation of ice and snow tourism destinations. The BERT model extracts the semantic features of text, while CNN further aggregates and classifies these features, thus realizing efficient processing and analysis of unstructured text data21. The overall architecture of the designed BERT-CNN model is presented in Fig. 1:
Fig. 1
figure 1
Framework of the BERT-CNN Model.
Full size image
The BERT-CNN model integrates BERT and CNN to extract deep semantic features from textual data and perform feature aggregation and classification through CNN. The main components of the model include data preprocessing, the BERT encoding layer, the CNN feature extraction layer, the feature fusion and classification layer, and model training and optimization22,23,24.
In the data preprocessing stage, raw text is transformed into a format suitable for BERT through tokenization. The text is segmented into words or sub-word units and processed by BERT’s tokenizer. Special tokens are added to each text segment to represent the context of the entire sentence and to demarcate sentence boundaries. The text length is then adjusted to a fixed size, such as 128 or 256, to ensure the model can handle variable-length text25. The purpose of data preprocessing is to prepare the data for subsequent BERT encoding and feature extraction. In the BERT encoding layer, BERT’s bidirectional Transformer architecture extracts deep semantic information from the text. BERT captures contextual information through self-attention mechanisms, enabling it to consider preceding and succeeding text simultaneously, thus generating more accurate semantic representations. Through this layer, BERT can handle complex linguistic structures and long-range dependencies, outputting contextual representations for each token. Specifically, the output vectors of the tokens are used to represent the semantic information of the entire text and serve as final features in classification tasks26. The CNN feature extraction layer applies convolution kernels (e.g., 3 × 1, 4 × 1, 5 × 1) to the BERT output to extract local features from the text27,28,29,30. Convolutional operations generate multiple feature maps, each representing a type of local semantic information. To reduce feature redundancy and computational complexity, CNN employs max-pooling operations to extract the most significant features from each convolutional feature map. Using multiple convolution kernels helps the model capture both short-term and long-term dependencies in the text, enhancing its expressive power. In the feature fusion and classification layer, the multiple feature maps extracted by CNN are concatenated into a high-dimensional feature vector. This vector encapsulates multi-level semantic information from the text. Subsequently, these features are further integrated through fully connected layers and finally classified using the Softmax activation function. This layer outputs class probabilities based on the input text’s features, thus completing the classification task31,32,33. During model training and optimization, this study employs the cross-entropy loss function to measure the discrepancy between predicted results and true labels, and optimizes model parameters through backpropagation. The Adam optimizer is used for its adaptive learning rate, which enhances training efficiency and stability. The learning rate may be dynamically adjusted during training to improve the model’s generalization ability.
The relevant code for the model is as follows:
figure a
In summary, the BERT-CNN model, by integrating BERT’s deep semantic understanding capabilities with CNN’s local feature extraction abilities, demonstrates excellent performance in text classification tasks. Through precise feature extraction and effective feature aggregation, the BERT-CNN model captures complex semantic information and maintains high computational efficiency while improving classification accuracy. This enables the BERT-CNN model to exhibit superior performance in multi-dimensional decision-making tasks such as ice and snow tourism destination image evaluation, showcasing broad application prospects.
Construction of tourism destination image evaluation model
To comprehensively evaluate the image of ice and snow tourism destinations, this study integrates the BERT-CNN model’s text analysis and the MADM method’s comprehensive evaluation abilities to construct a new evaluation model. The model extracts key information from tourist-generated content and realizes scientific evaluation of tourism destination images through quantitative methods. The evaluation of tourism destination image involves many dimensions, such as resource attraction, service quality, cultural characteristics, and environmental conditions. These dimensions are reflected in structured data and implied in tourist comments and social media texts. The specific framework of the model is displayed in Fig. 2:
Fig. 2
figure 2
Architecture of Tourism Destination Image Evaluation Model.
Full size image
Figure 2 illustrates that the data preprocessing and feature extraction module is responsible for extracting effective textual features from user-generated content (UGC), providing support for subsequent MADM analysis. This module first collects a large volume of tourist review data from online travel platforms (e.g., TripAdvisor, Ctrip) using web crawling techniques. These reviews include tourists’ evaluations of destinations, covering aspects such as hotels, transportation, and attractions. After collection, the data undergoes cleaning, including noise removal, text normalization, tokenization, and deduplication. Subsequently, the BERT model performs a deep semantic understanding of the review texts. By leveraging its bidirectional Transformer architecture, BERT captures each word’s contextual meaning, generating contextual semantic representations of the text. To further enhance the model’s feature extraction capabilities, CNN is applied to process the features generated by BERT. CNN extracts local information from the text through convolution kernels and retains the most significant features via pooling operations, ultimately producing more representative textual feature vectors. Next, the attribute weight allocation and decision evaluation module utilizes MADM methods to integrate the features extracted from reviews with the weights of various attributes. This process calculates a comprehensive score for each destination. First, the model defines multiple evaluation attributes, such as resource attractiveness, service quality, environmental quality, cultural atmosphere, and transportation convenience. These attributes represent different perceptual dimensions of tourists toward destinations. Then, the AHP is used to allocate weights to each attribute. AHP, a MADM method, calculates the importance weights of each attribute in the overall evaluation through expert scoring and judgment matrices. After obtaining the attribute weights, the TOPSIS method is applied to compute comprehensive scores for each destination. TOPSIS evaluates the destinations’ relative strengths and weaknesses by calculating their distances to the ideal and negative solutions. Finally, the decision result analysis and visualization module analyzes and presents the evaluation results, providing clear assessment information for decision-makers. By comprehensively analyzing the scores and weights of each destination’s attributes, the model outputs a comprehensive image score for each destination. These scores reflect the destination’s performance across various dimensions, helping managers identify strengths and areas for improvement. To facilitate a quick understanding of the evaluation results by relevant stakeholders, the model visualizes the scores using charts, rankings, and other methods. For example, radar charts and bar charts display destinations’ performance across different dimensions, enabling easy cross-comparison and trend analysis. The model’s strength lies in its multi-dimensional evaluation capability, enabling comprehensive analysis of destination images from multiple perspectives. These perspectives include resources, services, environment, and culture, providing a more precise image profile. The model effectively processes large volumes of unstructured review data by integrating the BERT-CNN model for text feature extraction and MADM methods for comprehensive evaluation. Meanwhile, this model offers scientific decision support through weight allocation and comprehensive scoring. Compared to traditional evaluation methods, the model improves evaluation quality and enhances analysis efficiency. Additionally, the model demonstrates strong flexibility and scalability. It applies to ice and snow tourism destination image evaluation and is extendable to other types of destination evaluations, brand assessments, and related fields.
In conclusion, the proposed tourism destination image evaluation model, by combining the BERT-CNN model with MADM methods, enables comprehensive and precise evaluation of destination images. Concurrently, it provides valuable decision-making support for tourism managers. The successful application of the model enhances the quality of destination image evaluation. Also, it offers data support and scientific evidence for destination branding optimization, service quality improvement, and marketing strategy formulation.
Experimental design and performance evaluation
Datasets collection, experimental environment, and parameters setting
The dataset used in this study is the TripAdvisor Hotel Reviews Dataset, which contains hotel reviews from the TripAdvisor website. Each review includes textual data and some structured information related to the review. The dataset aims to facilitate the analysis of hotel service quality, user sentiment, and satisfaction, which can be used for sentiment classification, sentiment analysis, NLP tasks, and text analysis in machine learning. The dataset contains a large amount of UGC, providing valuable data resources for research and applications, particularly for sentiment analysis, service quality evaluation, and personalized recommendations in the tourism industry. The dataset is publicly available and can be downloaded for free from the Kaggle website (https://www.kaggle.com/datasets/andrewmvd/trip-advisor-hotel-reviews). It comprises approximately 22,000 hotel reviews and includes over 10 fields, containing textual and structured data. The text length of each review varies significantly, ranging from short one- or two-sentence comments to longer, detailed evaluations spanning hundreds of words. The dataset offers a wealth of authentic UGC, supporting multiple research directions in the tourism industry, such as sentiment analysis, service quality evaluation, and recommendation systems. It holds significant application value for experimental research on review-based analysis models, sentiment analysis, and MADM models.
In addition, the dataset contains tens of thousands of comment data, which is enough to support a large-scale text analysis experiment. The study classifies datasets according to theme types, as follows:
(1)
Facilities and services: It involves the evaluation of the destination’s hardware facilities (such as hotels and transportation) and service quality. Facility and service-related reviews typically account for 30-40% of total reviews, as tourists often provide detailed evaluations of hotel services, facility completeness, and convenience.
(2)
Nature and culture category: It includes comments on the destination’s natural landscape or cultural activities. Nature and culture-related reviews usually make up 25-35% of total reviews, as tourists’ assessments of a destination’s cultural atmosphere and natural landscapes are integral components of their travel experiences.
(3)
Comprehensive experience category: The overall evaluation of the whole travel experience, including feelings, cost performance, etc. Comprehensive experience-related reviews constitute 20-30% of total reviews, as tourists frequently summarize their overall travel experiences in their final evaluations, reflecting their general satisfaction with the trip.
The configuration of the experimental environment is as follows:
(1)
The Central Processing Unit (CPU) is Intel Core i7-12700 K.
(2)
The Graphics Processing Unit (GPU) is NVIDIA GeForce RTX 3090 (24GB video memory).
(3)
The memory is 32GB DDR5 4800 MHz.
(4)
The stored model is Samsung 980 PRO NVMe SSD 1 TB.
(5)
The motherboard is ASUS ROG STRIX Z690-E.
To ensure experimental accuracy, the study standardizes the model parameters. The maximum sequence length is set to 128, with a batch size of 16 and a learning rate of 0.001. The AdamW optimizer is employed, and the model’s hidden layer dimension is 768. The Dropout probability is 0.1, and the number of training epochs is 3. The convolution kernel sizes are 3, 4, and 5. The total number of convolution kernels is 100. The ReLU activation function and MaxPooling method are implemented. For comparison, three models are selected: Robustly Optimized Bidirectional Encoder Representations from Transformers Approach (RoBERTa), Generalized AutoRegressive Pretraining for Language Understanding (XLNet-Conv), and A Lite Bidirectional Encoder Representations from Transformers - Bidirectional Long Short-Term Memory Network (ALBERT-BiLSTM). RoBERTa, an improved model based on BERT, achieves superior performance in various NLP tasks by optimizing training strategies (e.g., removing the Next Sentence Prediction task and increasing training data volume) and adjusting hyperparameters. The selection of RoBERTa leverages its strengths in large-scale text pre-training. It evaluates the model’s performance in extracting semantic features and processing complex textual data, particularly in terms of accuracy in understanding and generating natural language content. XLNet combines the advantages of autoregressive and autoencoding models, enhancing generalization abilities through generative pre-training. Compared to traditional models like BERT, it better captures contextual dependencies, especially long-range dependencies. This study chooses XLNet-Conv as a comparative model to evaluate its strengths in handling long or complex textual data, particularly in semantic reasoning and decision-making tasks. Additionally, XLNet’s autoregressive properties can support the deep understanding required by MADM models. ALBERT, a lightweight version of BERT, maintains BERT’s performance advantages while improving computational efficiency by reducing parameter counts and sharing weights. By integrating ALBERT with BiLSTM, the model leverages BiLSTM’s strengths in sequence modeling to further capture long-term dependencies in context. The selection of the ALBERT-BiLSTM model aims to assess how lightweight models maintain high effectiveness in extracting textual features and handling decision-making tasks under limited computational resources. This model particularly focuses on runtime efficiency and performance on large-scale datasets.
Performance evaluation
Performance comparison experiment
The performance comparison experiment is conducted and evaluated through four dimensions: accuracy, model efficiency, model robustness, and user experience. There are two evaluation indicators in each dimension. In the accuracy dimension, two key indicators—accuracy rate and classification error rate—are crucial for evaluating the fundamental capabilities of classification models. The accuracy rate reflects the model’s overall correctness, indicating the proportion of correctly classified samples among all predictions. In contrast, the classification error rate provides the proportion of incorrect predictions from a complementary perspective. Together, these indicators comprehensively assess model performance, ensuring reliable results. In the model efficiency dimension, inference time and training time directly impact the model’s application efficiency. Inference time measures the model’s response speed when processing individual input data, which is particularly important for applications requiring real-time feedback. Training time evaluates the duration from the start to the completion of model training, and reducing training time can significantly enhance development efficiency and save resources, especially for large-scale datasets. In the model robustness dimension, adversarial sample accuracy and loss value stability are key indicators for assessing model performance under varying input conditions. Adversarial sample accuracy evaluates the model’s stability against malicious perturbations or noisy inputs, ensuring reliable performance in uncertain environments; Loss value stability reflects the fluctuation amplitude of the loss function during training; Lower fluctuations indicate a more stable training process and avoiding overfitting or poor training behaviors. In the user experience dimension, two indicators—Top-k accuracy and model memory usage—focus on the model’s practical application effectiveness. Top-k accuracy is suitable for multi-choice recommendation tasks, measuring whether the correct answer is included in the top-k predictions, thus improving recommendation system accuracy. Memory usage reflects the resource consumption of the model across different hardware environments, with lower memory usage enabling broader deployment and better adaptability. The evaluation results in the accuracy dimension are shown in Fig. 3:
Fig. 3
figure 3
Accuracy Dimension Evaluation (a) Accuracy (b) Error Rate.
Full size image
Figure 3 demonstrates that the proposed optimized model achieves exceptional accuracy across all themes. In the “Facilities and Services” theme, the optimized model attains an accuracy of 0.95, outperforming other models such as RoBERTa (0.91) and XLNet-Conv (0.92) by at least 3%. In the “Nature and Culture” theme, the optimized model again leads with an accuracy of 0.96, compared to ALBERT-BiLSTM’s 0.94. Especially for the “Comprehensive Experience Category” theme, the accuracy of the optimized model reaches 0.97, showcasing its stability and generalization ability on multi-theme data. Regarding classification error rates, the optimized model’s performance aligns with its high accuracy. In the “Facilities and Services” theme, the optimized model’s error rate is only 0.05, significantly lower than RoBERTa’s 0.09 and XLNet-Conv’s 0.08. In the “Nature and Culture” theme, the optimized model’s error rate is 0.04, while that of ALBERT-BiLSTM is 0.06. It is particularly noteworthy that the error rate of the optimized model in the “Comprehensive Experience Category” is only 0.03. The evaluation result of the model efficiency dimension is illustrated in Fig. 4:
Fig. 4
figure 4
Evaluation Results of Model Efficiency Dimension (a) Inference Time (b) Training Time.
Full size image
In Fig. 4, the optimized model demonstrates superior performance with inference times consistently below 0.20 s across the three themes, outperforming other comparative models. For example, in the “Facilities and Services” theme, the inference time of the optimized model is 0.18 s, while that of ALBERT-BiLSTM is 0.22 s and that of RoBERTa is 0.25 s. In the “Nature and Culture” theme, the optimized model’s inference time is 0.19 s, approximately 21% faster than XLNet-Conv’s 0.24 s. This efficiency in inference time makes the model highly suitable for real-time applications. The optimized model also excels in training efficiency. For the “Nature and Culture” theme, its training time is 78.45 s, significantly shorter than RoBERTa’s 94.23 s and XLNet-Conv’s 89.34 s. Even in the complex “Comprehensive Experience Category” theme, the optimized model’s training time is controlled within 80 s, highlighting the effectiveness of its parameter optimization strategy. The robustness evaluation results are depicted in Fig. 5:
Fig. 5
figure 5
Evaluation Results of Model Robustness (a) Adversarial Accuracy (b) Loss Stability.
Full size image
Figure 5 shows that the optimized model excels in adversarial sample testing. For instance, in the “Comprehensive Experience Category” theme, its adversarial sample accuracy reaches 0.90, significantly outperforming RoBERTa (0.86) and XLNet-Conv (0.87). Even in the “Nature and Culture” theme, the optimized model maintains an adversarial sample accuracy of 0.89, compared to ALBERT-BiLSTM’s 0.86, highlighting its superior robustness against input noise. The optimized model also exhibits greater stability during training. For example, in the “Facilities and Services” theme, the standard deviation of the loss value for the optimized model is 0.11, lower than RoBERTa’s 0.16 and XLNet-Conv’s 0.14. Notably, for the “Comprehensive Experience Category” theme, the optimized model’s loss value fluctuates by only 0.09, indicating enhanced stability in training across multi-theme data. The user experience evaluation results are revealed in Fig. 6:
Fig. 6
figure 6
Evaluation Results Under the Dimension of User Experience (a) Top-k Accuracy (b) Memory Usage.
Full size image
The results in Fig. 6 reveal that the Top-3 accuracy of the optimized model performs exceptionally well across all themes in multi-classification tasks. For example, in the “Nature and Culture” theme, the Top-3 accuracy is as high as 0.97, remarkably outperforming RoBERTa (0.94) and ALBERT-BiLSTM (0.93). Regarding the “Comprehensive Experience Category”, the Top-3 accuracy of the optimized model remains high at 0.96, showing its comprehensive ability to capture multi-category information. The optimized model shows high efficiency in memory occupation. Moreover, in the “Facilities and Services” theme, its memory occupation is only 450 MB, about 27% less than the 620 MB of XLNet-Conv. Similarly, in the “Nature and Culture” theme, the optimized model uses just 470 MB of memory, significantly lower than RoBERTa’s 630 MB, showcasing the effectiveness of its lightweight design.
The verification of decision effect of the MADM model
To verify the decision-making effect of the model, the study also selects four dimensions (decision consistency, model interpretability, decision efficiency, and decision effect) for comparison.
In the decision consistency dimension, two indicators—consistency ratio and priority stability—ensure the consistency of judgments and the stability of decision outcomes during the decision-making process. A low consistency ratio indicates poor judgment consistency in the model’s decision-making process, potentially leading to unreliable results; Priority stability measures the fluctuation of decision outcomes under different weight allocations, ensuring stability throughout the process. In the model interpretability dimension, attribute weight distribution and contribution rate enhance the transparency of the decision-making process. A low dispersion coefficient in weight distribution indicates a more uniform allocation of weights across attributes, preventing the model from over-relying on a single attribute. The contribution rate quantifies the impact of each attribute on decision outcomes, helping better understand how the model balances various decision dimensions, thereby improving interpretability.
In the decision efficiency dimension, computation time and convergence rate are key indicators for evaluating the model’s efficiency. Computation time reflects the model’s response speed when processing individual decision tasks, and shorter inference times significantly improve decision efficiency, especially in scenarios requiring real-time feedback. The convergence rate measures the time required for the model to reach an optimal solution during decision-making. A faster convergence rate indicates the model’s ability to complete decision tasks quickly, avoiding excessive computational resource consumption. In the decision effect dimension, the result credibility and ranking change amplitude ensure the reliability and stability of the model’s decisions. Higher result credibility illustrates that decision outcomes are more universally applicable and stable across different scenarios; A smaller ranking change amplitude suggests that the model is less sensitive to input variations, maintaining consistent decisions in the face of uncertain inputs.
The evaluation results of decision consistency are denoted in Fig. 7:
Fig. 7
figure 7
Assessment Results of Decision Consistency (a) Consistency Ratio (b) Priority Stability.
Full size image
In Fig. 7, the consistency ratio of the optimized model is lower than 0.05 across the three themes, showing a high degree of consistency. For the “Facilities and Services” theme, the optimized model achieves a consistency ratio of 0.03, significantly exceeding RoBERTa’s 0.06 and XLNet-Conv’s 0.07. In the “Nature and Culture” theme, the consistency ratio of the optimized model is 0.04, maintaining excellent weight consistency, while ALBERT-BiLSTM shows a ratio of 0.05. Especially in the complex “Comprehensive Experience Category” theme, the optimized model’s consistency ratio remains at a low level of 0.03, demonstrating excellent performance in multi-dimensional weight distribution. The optimized model is outstanding in priority stability. In the “Facilities and Services” theme, its stability indicator is 0.89, significantly higher than XLNet-Conv (0.86) and ALBERT-BiLSTM (0.87). In the “Nature and Culture” theme, the optimized model’s priority stability reaches 0.90, showing strong robustness to weight disturbance. Even in the “Comprehensive Experience Category”, with great weight change, the optimized model maintains a priority stability of 0.91. In contrast, other models have great fluctuation, highlighting the optimized model’s strong adaptability to complex input conditions. The evaluation of model interpretability is suggested in Fig. 8:
Fig. 8
figure 8
Evaluation Results of Model Interpretability (a) Attribute Weight Distribution (b) Contribution Rate.
Full size image
The results of Fig. 8 show that the optimized model performs well in the discrete coefficients of the weight distribution. For example, in the “Facilities and Services” theme, its dispersion coefficient is 0.11, which is much lower than RoBERTa (0.14) and XLNet-Conv (0.13). In the “Nature and Culture” theme, the dispersion coefficient of the optimized model is 0.12, indicating its effective balance in distributing weight across multiple attributes. Especially, in the “Comprehensive Experience Category” theme, the optimized model’s dispersion coefficient is 0.10, outperforming other models and confirming its superiority in handling complex weight allocation tasks. The contribution rate of key attributes in the optimized model is also significantly higher than in the comparison model across all three themes. In the “Facilities and Services” theme, the contribution rate of the optimized model is 0.61, well above RoBERTa’s 0.55 and ALBERT-BiLSTM’s 0.57. In the “Nature and Culture” theme, the optimized model’s contribution rate is 0.63, highlighting the model’s ability to emphasize the influence of key attributes. In the “Comprehensive Experience Category”, the contribution rate reaches 0.62, reflecting its exceptional performance in multi-dimensional weight analysis. The evaluation result of decision efficiency is plotted in Fig. 9:
Fig. 9
figure 9
Evaluation Results of Decision-Making Efficiency (a) Computation Time (b) Convergence Rate.
Full size image
The data in Fig. 9 reveals that the optimized model offers remarkable advantages in computing time. In the “Facilities and Services” theme, the calculation time of the optimized model is only 0.14 s, which is about 33% faster than RoBERTa’s 0.21 s. In the “Nature and Culture” theme, the optimized model’s calculation time is 0.15 s, notably lower than XLNet-Conv’s 0.19 s. Even in the “Comprehensive Experience Category”, which presents high computational complexity, the calculation time of the optimized model is kept within 0.16 s, demonstrating its efficiency. The optimized model’s convergence rate also exceeds the comparison model in several themes. For example, in the “Facilities and Services” theme, the optimized model only needs 8 iterations to reach the convergence condition, whereas RoBERTa requires 12 iterations. In the “Nature and Culture” theme, the optimized model converges in 9 iterations, significantly faster than the 11 iterations of XLNet-Conv. In the “Comprehensive Experience Category”, the optimized model converges in 8 iterations, showcasing its rapid adaptability in complex tasks. Figure 10 displays the evaluation result of the decision effect:
Fig. 10
figure 10
Evaluation Results of Decision-Making Effectiveness (a) Result Reliability (b) Ranking Variation.
Full size image
Figure 10 demonstrates that the optimized model is stable in decision credibility. For example, in terms of “Facilities and Services”, the optimized model’s reliability is 0.03, while the reliability of XLNet-Conv and ALBERT-BiLSTM are 0.05 and 0.04 respectively. In the “Nature and Culture” theme, the optimized model’s reliability remains at 0.03, reflecting its output results’ reliability. The ranking change range of the optimized model is the smallest across all themes. For example, regarding “Facilities and Services”, the optimized model’s ranking variation range reaches 0.04, which is more stable than that of XLNet-Conv and RoBERTa. Considering the “Comprehensive Experience Category”, the ranking change range of the optimized model is only 0.03, showing its robustness to input disturbance.
Discussion
The experimental results reveal that the optimized model shows leading performance in many indicators. Especially in the key indicators such as accuracy, inference time, and the accuracy of adversarial samples, the optimized model substantially outperforms the comparison model. This reflects the effectiveness of the model optimization strategy. In contrast, RoBERTa and XLNet-Conv are inferior in some dimensions, such as classification error rate and memory occupation. This may be related to their high architectural complexity and computational requirements. In addition, although ALBERT-BiLSTM is superior in memory occupation, its accuracy and robustness are slightly insufficient. Overall, the optimized model achieves a balance of high accuracy, low error rate, and high efficiency by combining BERT-CNN structure and parameter optimization. It is also suitable for multi-theme data analysis of image evaluation tasks of ice and snow tourism destinations.
In addition, the optimized model shows excellent performance in the consistency, stability, efficiency, and effect dimensions. In the consistency of weight distribution, the optimized model can realize the judgment matrix with low deviation through efficient calculation, and improve scientific decision-making. In evaluating priority stability and ranking change range, the optimized model shows low volatility, which shows that it is robust to input change and weight disturbance. Compared with other models, the optimized model’s calculation time and convergence rate are superior, which is attributed to its parameter optimization and efficient algorithm design. Moreover, the results of the optimized model are highly reliable, showing stable output in many runs, which is helpful to the reliability of decision-making in practical applications. To sum up, the optimized model exhibits remarkable advantages in the key indicators of MADM, effectively supporting the tourism destination evaluation tasks in complex multi-theme scenarios.
Araújo-Vila et al. (2024) proposed an evaluation index system for ice and snow tourism destination images. Moreover, through questionnaire surveys and text analysis methods, they defined dimensions such as resource attractiveness, service quality, and cultural atmosphere, offering a framework for ice and snow tourism destination image evaluation34. Although this framework offers actionable evaluation criteria for tourism destinations, it primarily relies on traditional quantitative questionnaire surveys and manual text analysis methods. This may introduce subjectivity into the evaluation process and exhibit low efficiency when handling large-scale tourist review data. In contrast, this study integrates the BERT-CNN model, leveraging DL to comprehensively analyze sentiments and semantics in tourist reviews, enabling the extraction of richer features from unstructured textual data. Compared to the research by Borlido and Kastenholz, this study overcomes the limitations of traditional methods in processing large-scale textual data. Simultaneously, the study improves data processing efficiency while enhancing the objectivity and accuracy of the evaluation. Additionally, by employing MADM methods, this study dynamically adjusts the weights of various dimensions, adapting to the ever-changing market environment and tourist demands in constructing ice and snow tourism destination images. In comparison to the study by Tian (2023), which utilized AHP and TOPSIS-based MADM methods to evaluate tourism destination indicators and provided a quantitative basis for optimizing destination images35. The contribution of that study lies in effectively transforming multi-dimensional tourism destination evaluation indicators into comprehensive scores through the AHP method, aiding decision-makers in making scientific choices during optimization. However, the limitations of AHP and TOPSIS methods lie in their assumption of fixed attribute weights. In practical applications, tourist needs and preferences are dynamic, and fixed weight allocations may lead to biased evaluation results. This study proposes a more flexible evaluation framework by combining the BERT-CNN model with MADM methods. First, the BERT-CNN model introduces DL techniques in textual data analysis, improving the precision of feature extraction. Second, this study uses MADM methods to dynamically adjust weight allocations, fully considering changes in tourist perceptions. Compared to Tian’s research, the framework of this study better reflects the variability of tourist demands and the dynamics of the ice and snow tourism market. It avoids potential errors caused by fixed weights and provides more accurate and timely evaluation results.
Compared to previous studies, the proposed research framework demonstrates greater flexibility, accuracy, and efficiency. Integrating DL techniques with dynamic MADM methods can overcome the limitations of traditional approaches in handling large-scale unstructured textual data and adapting to dynamic changes. Thus, it provides a more comprehensive and adaptable tool for tourism destination image evaluation.
Conclusion
Research contribution
This study innovatively integrates the BERT-CNN model with MADM methods, successfully achieving a full-process analysis that extracts semantic features from unstructured textual data, calculates key attribute weights. Meanwhile, it comprehensively evaluates tourism destination images. Compared to traditional qualitative research or single-indicator analysis methods, this study’s framework provides a more comprehensive and scientific depiction of tourism destination images, offering new academic perspectives and theoretical foundations. The constructed model efficiently processes massive user-generated content from social media and travel review platforms. Also, it extracts tourists’ deep perceptions of destination images, significantly enhancing the decision-making support capabilities of tourism managers. Through the methods of this study, the strengths and weaknesses of tourism destinations can be quickly identified, providing a scientific basis for image optimization. Furthermore, by combining the BERT-CNN model with MADM methods, this study designs a complete technical workflow from data preprocessing and feature extraction to decision analysis. Compared to single models, the proposed method demonstrates higher efficiency and robustness in handling complex textual data and multi-dimensional decision-making problems.
Future works and research limitations
Although this study has achieved certain theoretical and practical outcomes in the ice and snow tourism destination image evaluation, several limitations remain. These limitations can be addressed in future research.
On one hand, the data in this study primarily originates from UGC on online travel platforms and social media. While these data are highly representative, they may exhibit regional biases. For instance, user reviews tend to focus more on popular ice and snow tourism destinations, potentially underrepresenting the image characteristics of small and medium-sized destinations. Moreover, the authenticity and completeness of UGC data may impact the research results, particularly when some reviews lack accuracy or are overly subjective, potentially compromising the objectivity of the evaluation. On the other hand, the evaluation framework proposed in this study performs well in ice and snow tourism scenarios. However, further adjustments to model parameters and feature extraction methods may be necessary when handling other complex scenarios. While the BERT-CNN model demonstrates strong semantic understanding and feature extraction capabilities, its high computational cost may limit its application in resource-constrained environments. For small enterprises or tourism management departments in remote areas, the model’s high computational demands could pose implementation challenges.
To address these issues, future research can enhance sample diversity and representativeness by integrating multi-channel data, such as survey questionnaires, video reviews, and smart device data. Data cleaning and enhancement techniques can effectively improve UGC data quality and reliability, thus reducing the impact of noise and bias. In addition, to tackle the issue of high computational costs, more lightweight model architectures can be explored to reduce computational complexity and improve operational efficiency. Furthermore, incorporating distributed computing methods such as federated learning can enhance the model’s efficient application in resource-constrained environments. This ensures the model’s effectiveness across tourism management systems of varying scales and conditions.
Through these improvements, future research is expected to further enhance the model’s adaptability, efficiency, and accuracy. Thus, stronger support can be offered for the image evaluation of ice and snow tourism destinations and applications in related fields.
Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author Hua Jin on reasonable request via e-mail jinhua@jisu.edu.cn.
References
Wang, X. et al. How to perceive tourism destination image? A visual content analysis based on inbound tourists’ photos. J. Destination Mark. Manage. 33 (4), 100923 (2024).
Google Scholar
Maldonado-López, B., Ledesma-Chaves, P. & Gil-Cordero, E. Cross-border destination image for sustainable tourism development in peripheral areas. J. Travel Tourism Mark. 41 (4), 614–639 (2024).
Google Scholar
Karri, V. R. S. & Dogra, J. Destination stereotypes: a phenomenon of destination image. J. Hospitality Tourism Insights 6 (3), 1290–1308 (2023).
MATHGoogle Scholar
Wu, J. & Ji, C. Image construction of China’s rural sports tourism destination from the perspective of rural revitalization. Tourism Manage. Technol. Econ. 6 (3), 1–13 (2023).
CASMATHGoogle Scholar
Sun, T., Li, Y. & Tai, H. Different cultures, different images: A comparison between historic conservation area destination image choices of Chinese and Western tourists. J. Tourism Cult. Change 21 (1), 110–127 (2023).
MATHGoogle Scholar
Li, Z. et al. The projected image of idyllic life and its construction. Curr. Issues Tourism 27 (7), 1125–1141 (2024).
MATHGoogle Scholar
Borlido, T. & Kastenholz, E. Word-of-art: the relationship between destination image and Art. Eur. J. Tourism Res. 33 (2), 3312–3312 (2023).
MATHGoogle Scholar
Hamdy, A., Zhang, J. & Eid, R. A new proposed model for tourists’ destination image formation: the moderate effect of tourists’ experiences. Kybernetes 53 (4), 1545–1566 (2024).
MATHGoogle Scholar
Zakiah, S., Winarno, A. & Hermana, D. Examination of consumer engagement for loyalty in sustainable destination image. Cogent Social Sci. 9 (2), 2269680 (2023).
Google Scholar
Pitakaso, R. et al. Multi-objective sustainability tourist trip design: an innovative approach for balancing tourists’ preferences with key sustainability considerations. J. Clean. Prod. 449 (22), 141486 (2024).
CASMATHGoogle Scholar
Guerreiro, M. M. et al. The online destination image as portrayed by the user-generated content on social media and its impact on tourists’ engagement. Tourism Manage. Stud. 20 (4), 1–15 (2024).
MATHGoogle Scholar
Liang, S. H. & Lai, I. K. W. Tea tourism: designation of origin brand image, destination image, and visit intention. J. Vacation Mark. 29 (3), 409–427 (2023).
MATHGoogle Scholar
Lian, J. Research on tourism marketing strategy innovation in new media era. Front. Bus. Econ. Manage. 8 (1), 103–106 (2023).
MATHGoogle Scholar
Luo, Y. et al. Exploring the destination image based on the perspective of tourists’ expression using machine learning methods combined with PLTS-PT. Soft. Comput. 27 (9), 5537–5552 (2023).
MATHGoogle Scholar
Cardoso, L. et al. Features of nautical tourism in Portugal—Projected destination image with a sustainability marketing approach. Sustainability 15 (11), 8805 (2023).
MATHGoogle Scholar
Almeida, G. G. F. & Almeida, P. The influence of destination image within the territorial brand on regional development. Cogent Social Sci. 9 (1), 2233260 (2023).
MATHGoogle Scholar
Pahrudin, P. et al. A large-sport event and its influence on tourism destination image in Indonesia. Tour. Hosp. Manag. 29 (3), 335–348 (2023).
Google Scholar
Ding, J., Md Syed, M. A. & Shamshudeen, R. I. Unveiling cultural heritage: exploring the perceived online destination image of China’s cultural heritage destination and narrative strategy in travel vlogs. J. China Tourism Res. 2 (1), 1–29 (2024).
Google Scholar
Lu, L., Jiao, M. & Weng, L. Influence of first-time visitors’ perceptions of destination image on perceived value and destination loyalty: A case study of grand Canal forest park. Beijing Forests 14 (3), 504 (2023).
Google Scholar
Orden-Mejía, M. A. & Huertas, A. Tourist interaction and satisfaction with the chatbot evokes pre-visit destination image formation? A case study. Anatolia 34 (4), 509–523 (2023).
MATHGoogle Scholar
Hamdy, A., Eid, R. & Gao, X. Integrating Muslim-Friendly tourist destination image, value, satisfaction and Muslim actual visit behaviour in the travel industry. Int. J. Tourism Res. 26 (5), 2753 (2024).
Google Scholar
Zhang, S. N. et al. Local cultural distortion risk at tourist destinations: connotation Deconstruction and theoretical construction. Curr. Issues Tourism 27 (2), 251–267 (2024).
MATHGoogle Scholar
Coronel Padilla, M. F. et al. Projecting a destination image through Facebook: the role and challenges of destination management organizations. J. Tourism Herit. Res. 6 (3), 219–233 (2023).
Google Scholar
Pagano, N. & Sharpley, R. Destination image/branding of Sicily and the mafia phenomenon: a corpus-based analysis of keywords in english guidebooks. J. Tourism Cult. Change 22 (2), 190–208 (2024).
MATHGoogle Scholar
Jiang, S. et al. Perceived destination image cohesion: A comparison study of attractions on the grand canal, China. Sustainability 15 (18), 13682 (2023).
Google Scholar
Garay-Tamajón, L. A. & Morales‐Pérez, S. Belong anywhere’: focusing on authenticity and the role of Airbnb in the projected destination image. Int. J. Tourism Res. 25 (1), 63–78 (2023).
Google Scholar
Chen, X. et al. Deep analysis of the homogenization phenomenon of the ancient water towns in Jiangnan: A dual perspective on landscape patterns and tourism destination images. Sustainability 15 (16), 12595 (2023).
Google Scholar
Alarcón-Urbistondo, P., Rojas-de-Gracia, M. M. & Casado-Molina, A. Proposal for employing user-generated content as a data source for measuring tourism destination image. J. Hospitality Tourism Res. 47 (4), 643–664 (2023).
Google Scholar
Nayak, N., Polus, R. & Piramanayagam, S. What can online reviews reveal about tourism destination image?? A netnographic approach to a pilgrim destination in India. Tourism Recreation Res. 49 (6), 1284–1300 (2024).
Google Scholar
Qian, L. et al. Exploring destination image of dark tourism via analyzing user generated photos: A deep learning approach. Tourism Manage. Perspect. 48 (2), 101147 (2023).
MATHGoogle Scholar
Das, P. et al. Reconceptualizing destination image. Anatolia 35(2): 359–373. (2024).
MATHGoogle Scholar
Harrill, R. et al. An exploratory attitude and belief analysis of ecotourists’ destination image assessments and behavioral intentions. Sustainability 15 (14), 11349 (2023).
MATHGoogle Scholar
Abouseada, A. et al. The power of airport branding in shaping tourist destination image: passenger commitment perspective. Geoj. Tourism Geosites 47 (2), 440–449 (2023).
Google Scholar
Araújo-Vila, N. et al. Film-Induced tourism as a key factor for promoting tourism destination image: the James bond Saga case. Administrative Sci. 14 (5), 94 (2024).
Google Scholar
Tian, D. Research on the construction and improvement strategy of tourist satisfaction evaluation index system. Int. J. Bus. Manage. 2 (1), 1–1 (2023).
MATHGoogle Scholar
Download references
Acknowledgements
This work was supported by following fundings: 1. This work was supported by General Projects of the 14th Five-Year Plan for Educational Science in Jilin Province, 2024. Higher Education Institutions to Promote the High-Quality Development of the Ice and Snow Tourism Industry in Jilin Province in the Digital Era (No.GH24583). 2. This work was supported by the Scientific Research Project of the Education Department of Jilin Province(No.JJKH20251500SK). 3. This work was supported by Jilin International Studies University 2024 Annual University-Level Projects (No.JW2024YB010).
Author information
Authors and Affiliations
School of International Culture and Tourism, Jilin International Studies University, Changchun, 130000, China
Hua Jin
Authors
Hua Jin
View author publications
You can also search for this author inPubMedGoogle Scholar
Contributions
Hua Jin: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparation, writing—review and editing, visualization, supervision, project administration, funding acquisition.
Corresponding author
Correspondence to Hua Jin.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
About this article
Check for updates. Verify currency and authenticity via CrossMark
Cite this article
Jin, H. Application of multi-attribute decision-making combined with BERT-CNN model in the image construction of ice and snow tourism destination. Sci Rep 15, 10613 (2025). https://doi.org/10.1038/s41598-025-95221-5
Download citation
Received:18 January 2025
Accepted:19 March 2025
Published:27 March 2025
DOI:https://doi.org/10.1038/s41598-025-95221-5
Share this article
Anyone you share the following link with will be able to read this content:
Get shareable link
Sorry, a shareable link is not currently available for this article.
Copy to clipboard
Provided by the Springer Nature SharedIt content-sharing initiative
Keywords
Multi-attribute decision making
BERT-CNN
Tourism image
Weight allocation