Official Projection vs. Public Perception: Measuring the Perceptual Discrepancy of Creative Industry Parks in the Industrial Heritage Category Using Large Language Models

Abstract

Industrial tourism serves as a medium for disseminating industrial culture and strengthening public awareness. Quantifying the discrepancies between official projections and public perceptions is essential for shaping the destination image and enhancing appeal and competitiveness. This study examines five industrial heritage creative industry parks using large language models (LLMs) and multimodal data to address this issue. The results indicate the following: (1) Multimodal data fusion improves feature representation. (2) A clear discrepancy exists between official projections and public perceptions. The official perspective emphasizes the Cultural value of heritage, in contrast to the public’s greater concern with the Service experience perception. Despite this divergence, there is alignment in the recognition of the Creative industry form dimension. (3) Public sentiment regarding the parks is predominantly positive. However, an analysis of negative sentiments reveals that insufficient supporting facilities and poor consumption experience are the primary sources of dissatisfaction. Through large language models and multimodal data, this study proposes a framework for quantifying the gaps between official projections and public perceptions. It also provides practical insights and empirical support for the management and planning of industrial heritage creative industry parks.

1. Introduction

Within the context of current urban stock development, industrial heritage has emerged as a new engine driving urban renewal. The industrial tourism market continues to expand, with its economic and social benefits becoming increasingly evident [1]. In China, the government places considerable emphasis on the renewal of industrial heritage [2]. Among various renewal models, the creative industry park approach has emerged as a mainstream pathway for transforming industrial heritage both domestically and internationally [3,4]. The industrial heritage creative industry parks (the parks) refer to a park that transforms old industrial spaces into complexes that integrate cultural exhibitions, artistic creation, and creative industries by leveraging industrial heritage resources [5]. By preserving the historical character of industrial spaces and maintaining the memory of urban development, it also promotes the integration of emerging business formats and the restructuring of spatial functions. Consequently, the parks provide an appropriate context for exploring the multifaceted perception of industrial heritage.

The adaptive reuse of industrial heritage serves as a form of urban renewal, aligning with environmental policies and sustainable development principles [6]. Through functionally reconfiguring, old industrial buildings are transformed into comprehensive venues for heritage preservation, cultural dissemination, and recreational experiences, which achieves efficient resource utilization while preserving the historical value of industrial heritage [7,8,9]. This approach has been thoroughly validated in practice. For instance, numerous industrial sites in Europe have been reconfigured into educational bases that integrate heritage preservation, cultural dissemination, and educational functions [10]. Similarly, China has actively promoted the transformation of industrial heritage, exemplified by the development of Shougang Park into a comprehensive complex that combines commercial, technological, and tourism functions, which has emerged as a new landmark for urban renewal [11]. These practices demonstrate that, in the development of industrial tourism, heritage managers prioritize balancing historical preservation with functional reuse [12,13,14], aiming to foster high-quality urban development through the sustainable reuse of industrial heritage.

The sustainable development of industrial tourism depends not only on spatial and resource conditions but also on the public’s willingness to explore destinations and their cognitive interest [15]. As industrial tourism develops, it attracts a growing number of participants. Research indicates that factors driving public participation in industrial tourism are diverse, including industrial historical identity, interest in historic architecture [16], curiosity about industrial heritage and demand for recreational facilities [17]. Therefore, shaping a destination image that aligns with public expectations is essential for enhancing public satisfaction and promoting the sustainable development of industrial tourism. Research on industrial tourism destination images has primarily focused on industrial heritage landscapes [18,19] and public perception studies [20]. For example, some studies have analyzed text and location data from social media using content analysis and Importance-Performance Analysis (IPA) methods to identify and assess industrial heritage landscapes [21]; other scholars have employed field sampling method to investigate the determining factors of public loyalty to creative industry parks [22]; still others have applied mixed-method approaches, including content analysis and principal component analysis, to explore disparities in perceptions between experts and communities regarding the image of industrial heritage tourism destinations and their impact on urban well-being [23]. Current research has recognized the importance of public participation; however, with the rapid urbanization and development of industry tourism, discrepancies between official planning and public expectations have become increasingly prominent, often constraining public satisfaction. However, existing study primarily focuses on a single perspective.

Since Grosspietsch proposed that tourism destination image can be divided into projected image and perceived image [24], comparative studies of these two have become an important direction in destination image research. The projected image, originating from the supply side, represents the destination representation that is constructed and disseminated to the public through systematic integration and refinement of tourism elements [25,26,27]. Official projections reflect the destination marketing focus [28]. The perceived image, originating from the demand side, reflects the public’s overall impressions, feelings, and evaluations of the destination [29]. It refers to a holistic mental construct of the destination held by the public [28]. Official entities have historically dominated the construction and dissemination of destination image, with the projected image serving as the foundational framework for public perceptions. With the advent of the big data era, the proliferation of User-generated content (UGC), including online reviews and travelogs, has occurred. The public has evolved from passive recipients of information to active co-constructors and disseminators of destination image. Research indicates that alignment between the projected image and the perceived image is crucial for effective destination marketing and sustainable development [30,31]. China’s industrial heritage transformation follows a distinct top-down characteristic, with the government playing a leading role in planning and development [32]. In this model, discrepancies often emerge between the official projections and the public’s actual perceptions. Accurately identifying and measuring the cognitive gaps accurately between the official and the public has become a crucial issue requiring resolution.

The difference between projected image and perceived image is an important issue in destination image research. Scholars have employed diverse research methods, including questionnaire surveys [33,34], qualitative interviews [35], and content analysis [36,37,38], demonstrating that discrepancies exist between official projections and public perceptions. These research findings provide both theoretical and practical foundation for the construction and management of destination image [39,40]. Existing research primarily employs statistical analysis to establish matching models, identifying and quantifying the importance of different value dimensions. For instance, some studies have identified mismatches between official publicity and tourists’ perceptions employing content and sentiment analysis methods, which suggest relevant optimization strategies [41]. Other scholars utilized online photos as data sources to compare content differences between officially promoted images and those spontaneously shared by the public [42]. The variety of current research methods and data sources has grown, providing valuable insights for destination image research. However, most studies rely on single text or image data for analysis, which limits the comprehensive capture of multidimensional perceptions. Furthermore, some limitations remain in areas such as multimodal data fusion and framework construction for measuring discrepancies [43,44,45].

Advances in artificial intelligence, particularly in LLMs and MLLMs, have enabled new possibilities for text, images, and sentiment analysis. Their exceptional capabilities in semantic understanding and visual data processing help address limitations associated with multidimensional perceptual information [46]. Recent studies have begun to employ LLMs for research on tourism destination images, spatial perceptions, and landscape environments. These studies primarily employ text-based analysis, image-based analysis, and some studies utilize multimodal methods that integrate text and images. For instance, some studies apply LLMs to analyze text and sentiment from social media related to urban green spaces [47]; others leverage the image-to-text recognition capabilities to examine the perceptions of street spaces [48]; still others use LLMs to investigate the cultural ecosystem services through social media data [49]; as well as employ large vision-language models to evaluate the visual features of urban recreational spaces [50]. These studies demonstrate the potential applications of large language models in tourism research and environmental perceptions. Moreover, studies indicate that integrating multimodal data can improve the comprehensiveness and accuracy of feature extraction [51,52]. However, multimodal data analysis still faces challenges such as data heterogeneity, information redundancy, and cross-modal alignment, which require a more systematic approach to validate the technical process [53,54].

Despite these advancements, there is still a lack of systematic quantitative research on discrepancies. Additionally, targeted research measuring the discrepancy between projected image and perceived image in the context of an industrial tourism destination image is still deficient [55]. Therefore, this study selects industrial heritage creative industry parks as research subjects and develops a framework for quantifying discrepancies between official projections and public perceptions, utilizing large language models and multimodal data. At the theoretical level, this framework addresses the limitations of traditional methods and single-modal data, enabling the comprehensiveness and replicability of discrepancy quantification. In addition, the multimodal fusion approach can be transferred to other domains where both image and text data are available. At the practical level, the research findings offer new insights into destination image shaping and the sustainable development of industrial tourism, while providing data driven decision support for the management and operation of creative industrial parks.

The main objectives of this study are as follows: first, identify the focus of official and public attention and measure the perception discrepancies across perceptual dimensions; second, examine the overall emotional tendency of the public towards the park through emotional analysis and reveal the factors driving these sentiments; third, develop a replicable and scalable process for quantifying perceptual differences. Following the research objectives, this study seeks to answer the following questions:

• What perceptual dimensions do official projections and public perceptions focus on in industrial heritage creative industry parks, and which of these dimensions exist significant discrepancies between the two?
• What is the public’s overall emotional tendency towards these creative industry parks, and what are the key factors driving these emotions?
• How can large language models and multimodal data be leveraged to construct an effective framework for quantifying perceptual discrepancy?

2. Methods

2.1. Research Framework

Figure 1 illustrates the research framework for this study. We selected five representative parks as research subjects and collected textual and visual data from official websites and the Dianping platform. After data cleaning and preprocessing, we established a structured prompt framework and applied LLMs for analysis. Text and sentiment analysis were conducted using the DeepSeek-R1 model, while image analysis was performed using the GLM-4V-Plus-0111 model in combination with DeepSeek-R1. To verify the reliability of the model outputs, we invited volunteers to manually annotate the data and evaluated model performance using confusion matrices. Finally, we fused the results of textual and images to quantify the perceptual discrepancies between official and public perspectives, providing a foundation for subsequent discussions and management recommendations.

2.2. Study Area

To ensure the data’s representativeness and generalizability, five parks, including Dahua 1935, Taoxichuan Creative Industry Park, Canal Hub 1958, Eastern Suburb Memory, and M50 Creative Industry Park (Figure 2), were selected as research subjects according to the following four criteria:

• Exemplary in the transformation of industrial heritage into creative industry parks, serving as models of successful adaptation.
• The transformed parks exhibit a functional integration of cultural, artistic, commercial, and innovative activities.
• The parks encompass various types of industrial heritage and geographic distributions, reflecting the diversity of the research subjects in both type and regional characteristics.
• The parks offer sufficient data to support a comprehensive analysis.

A detailed introduction of the parks is shown in

Table 1. Introduction to the five parks used in this study.

Park Name	Area (Hectare)	Geographic Location	Brief Introduction
Dahua 1935	8.7	Shaanxi Province	Dahua 1935, originally Chang’an Dahua Textile Mill, has been transformed into a cultural and commercial complex centered on industrial heritage renovation and adaptive reuse. It integrates urban consumption functions such as fashion, dining, culture, entertainment, and tourism.
Taoxichuan Creative Industry Park	8.9	Jiangxi Province	Taoxicuan is based on the protection and utilization of ceramic industrial heritage. It offers diverse business forms, including Taoxicuan cultural and art space, social spaces, research study base, and other industry forms.
Canal Hub 1958	6.2	Jiangsu Province	Canal Hub 1958, originally a historic steel industrial plant along the Wuxi canal. It integrates diverse scenarios including cultural and creative arts, specialty dining, leisure and social spaces, family-friendly interactions, trendy retail.
Eastern Suburb Memory	20	Sichuan Province	Eastern Suburb Memory integrates art exhibitions, industrial heritage, graffiti walls, commercial spaces, and more into a multifunctional cultural district.
M50 Creative Industry Park	4.1	Shanghai	M50 Creative Industrial Park is one of the most iconic examples of industrial heritage revitalization in Shanghai. It has established a diverse ecosystem integrating contemporary art, creative design, and cultural consumption.

.

2.3. Perceptual Dimension System

To accurately capture the park’s multifaceted perceptions, this study develops a multi-dimensional system. Based on the emphasis placed on intangible value of industrial heritage in the Dublin Principles [56], insights from previous relevant studies [21,40,57,58,59], and a consideration of the specific characteristics of creative industry parks. The perceptual dimension system is divided into four main dimensions and twelve sub-dimensions. This classification system encompasses the material attributes and intangible characteristics of industrial heritage creative industry parks, offering a comprehensive reflection of the destination’s image (

Table 2. Perceptual dimension system.

Dimension	Sub-Dimension	Abbreviation	Explain	Examples
Cultural value of heritage	Industrial relics	IR	Physical remains of industrial civilization	Production facilities, industrial equipment, and factory buildings
	Historical ambiance	HA	Conveying historical authenticity and temporal accumulation.	History, sense of era, mottled traces
	Cultural memory	CM	The intangible cultural values and spiritual values of industrial heritage.	Culture, memory, inheritance, and spirit
	Nostalgic feelings	NF	Evoking collective nostalgia through recognizable industrial symbols, fostering emotional connection and identity.	Nostalgia, the past, reminiscence, emotional belonging
Spatial environmental perception	Spatial layout	SL	Industrial architectural and renovation design, reflecting style and layout features.	Style, materials, renovation design, layout, structure
	Environmental landscape	EL	The integrated landscape system within the creative industry parks site, including lighting, sculptures and greenery.	Environment design, lighting, sculptures, nightscape, greenery
Creative industry form	Art performances	AP	Industrial spaces hosting artistic creation and exhibitions.	Art exhibitions graffiti, theaters, performances
	Design innovation	DI	Creative clusters that driving the functional transformation and innovation within industrial heritage.	Design studios, innovation hubs, creative incubators
	Industrial form	IF	Agglomeration, diversity, operational vitality, and development potential of creative industries.	Industrial clusters, diverse business formats, operational status
Service experience perception	Supporting facilities	SF	Basic services such as catering, transportation, navigation, and convenient facilities	Shops, transportation, parking lots, recreational facilities, navigation, restrooms
	Atmosphere experience	AE	Social interaction, check-in behavior, and emotional experiences	Internet-famous spaces, photography, check-in, vibrant ambiance
	Consumption experience	CE	Perception of the scene, price judgment, price sensitivity, and willingness to pay in cultural consumption	price-performance ratio, cultural and creative, products

).

2.4. Data Collection and Preprocessing

2.4.1. Data Collection

This study collected data on official projections and public perceptions from various sources, covering the period from 1 January 2019 to 31 January 2025.

Specifically, official projection data consist of two sources: the official government websites and the authoritative official media outlets. The former comes from official government websites of the cities where the parks are located, such as the Sichuan Provincial Department of Culture and Tourism and the Shanghai Municipal People’s Government Website. The latter primarily originates from media outlets, including People’s Network and Wuxi Daily. We used the names of parks and relevant expressions as search keywords, such as “Dahua 1935,” “Eastern Suburb Memory,” and “Dongjiao Creative Industry Park.” The retrieved texts were then manually screened, retaining those thematically focused on the parks with high relevance, excluding irrelevant content such as financial reports, land acquisition notices, and commercial advertisements. The retrieval and screening process was jointly conducted by the author and two graduate students familiar with the research field to ensure data accuracy. Following two rounds of independent screening and cross-check, the findings were ultimately reviewed and confirmed by the corresponding author. Finally, this study obtained 320 official texts, comprising over 190,000 words and 470 image data.

For public perception data, we exploited Dianping (https://www.dianping.com), a popular social media platform in China known for its large user base and extensive reviews. It has been widely used to quantify public perceptions and therefore serves as a reliable data source for this study [60]. Python 3.11.0 was employed to collect publicly available review data for the five parks from Dianping. After data preprocessing, we removed invalid and duplicate reviews, resulting in a dataset containing 21,914 text reviews and 99,064 images.

We acknowledge the disparities in data volume between official and public sources, as well as between textual and visual data. We emphasize that these discrepancies arise from distinct identities and content features of the official and public perspectives as presented on different platforms. These discrepancies do not affect the validity of the data, as our analysis is based on proportional content analysis rather than the absolute number of comments. This approach ensures that the data accurately reflect the essence of official projections and public perceptions, despite sample size discrepancies.

Figure 3 displays the online reviews and the official media outlets’ content related to the Canal Hub 1958. Each review in the public dataset contains the following information: user ID, user rating, comment text, associated images, and posting date. The official dataset comprises the following elements: publishing platform, textual content, images, and publishing date.

2.4.2. Data Preprocessing

To ensure data usability, the text and image data must undergo preprocessing. In the text preprocessing stage, stop words, special characters, HTML tags, and punctuation were removed. Simultaneously, reviews containing fewer than ten characters were excluded to eliminate non-informative entries. In the image preprocessing stage, a hash algorithm was applied to detect and remove duplicate images, thereby reducing redundancy.

2.5. Large Language Models and Prompt Engineering

LLMs and MLLMs are advanced deep-learning algorithms trained on vast amounts of textual and visual data [61]. These models perform outstandingly in text, image, and sentiment analysis tasks. Their advanced capabilities in semantic understanding, visual cognition, and sentiment recognition provide a strong foundation for multimodal data processing.

Prompt engineering is a technique designed to improve model performance by formulating tailored prompts for LLMs [62]. Previous studies have shown that prompt-based methods significantly improve LLMS performance, particularly when large model parameters and training samples are scarce [63]. Precise and well-structured prompts are crucial for generating high-quality and reliable outputs. Recent studies have also summarized key principles and strategies for prompt engineering design [64]. To improve model performance, this study adopts the RISEN (Role Instructions Steps End goal Narrowing) framework for prompt engineering and refine it based on the generated responses. The prompt engineering framework is illustrated in Figure 4.

2.6. Analysis Methodology

The official texts and social media data analyzed in this study were written in Chinese and included context-specific vocabulary, place names, and other linguistic elements. To mitigate language bias, all steps of the text involving the LLMs, including model input and output, are conducted in Chinese. For clarity, figures and tables were translated into English. More details on model parameter settings and prompts in the Supplementary Materials.

2.6.1. Text Analysis

The DeepSeek-R1 model (https://www.deepseek.com/) was employed for text analysis. Released by Deepseek company in January 2025, Deepseek-R1 is an open-source reasoning model developed using reinforcement learning (RL). It supports chain-of-thought (CoT) reasoning and excels in logical reasoning, Chinese text processing, and sentiment analysis. Text analysis utilized Python scripts to leverage the DeepSeek-R1 API, achieving an automated submission and analysis process through the constructed prompt framework. Additionally, according to the distinct characteristics of official long texts and user-generated short texts, the prompt tasks were differentiated to accommodate varying semantic structures. These prompts enable the model to generate outputs that better align with established standards, as demonstrated in

Table 3. Text analysis prompt engineering.

Framework Part	Public’s Text Analysis	Official’s Text Analysis
Role	Role: You are an expert in destination image perception, specializing in industrial tourism, with expertise in: Destination image theory. Specialize in text analysis and multi-dimensional perception analysis from user-generated content.	Role: You are an expert in destination image perception, specializing in industrial tourism, with expertise in: Destination image theory. Specialize in text mining and multi-dimensional perception analysis. Understanding the narrative structure of the official discourse system.
Instructions	Your task is to analyze the text reviews in depth. Specific tasks are matching the text reviews to perceptual dimensions and providing confidence levels, ultimately outputting up to three valid matching dimensions.	Your task is to analyze the official text in depth. Specific tasks include splitting long texts into smaller chunks based on semantic boundaries; matching the text content to perceptual dimensions and providing confidence levels; ultimately outputting up to three valid matching dimensions.
Steps	Text preprocessing: Identify the core semantic content of text reviews. Perceptual dimension matching: identify keywords and semantic features in the text. Match them with perceptual dimensions and output confidence levels (0.0–1.0). Filter the dimensions with a confidence level ≥ 0.6 as valid matches. Result output: sorted by confidence level in descending order (up to 3 dimensions).	Text preprocessing and segmentation: Identify the core content of the official text. Segment the text by detecting semantic boundaries (such as topic change points), and ensure the meaning remains intact. Perceptual dimension matching: identify keywords and semantic features within the segmented text. Match them with perceptual dimensions and output confidence levels (0.0–1.0). Filter the dimensions with a confidence level ≥0.6 as valid matches. Result output: sorted by confidence level in descending order (up to 3 dimensions).
End Goal	Output structured analysis results, including perceptual dimensions matching and confidence levels (sorted in descending order). The results must be presented in a modularized and clearly structured format to ensure effective text analysis.	Output structured analysis results, including perceptual dimensions matching and confidence levels (sorted in descending order). The results must be presented in a modularized and clearly structured format to ensure effective text analysis.
Narrowing	All matches are restricted to the perceptual dimensions listed in Table 2. Output formatted in Markdown.	Splitting based on the semantic boundaries to recognize key topic changes. All matches are restricted to the perceptual dimensions listed in Table 2. Output formatted in Markdown.
Input	Input data: Text reviews posted by the public on the Dianping platform. “It is rich in artistic atmosphere, can be regarded as the most instagrammable spot, which is great for wandering around.”	Input data: Official text content published on the platform. “Relevant data show that more than 300 enterprises have settled in this park … In 2025, Chengdu International Fashion Industrial Park will continue to promote digital transformation and introduce more international digital art exhibitions.”
Output	Comment Number: 1 Primary Dimension: Atmospheric experience Confidence Level: 0.9 Secondary Dimension: Art performances Confidence Level: 0.8 Third Dimension: null	Text Number: 1 “Relevant data shows that there are more than 300 enterprises settle in this park … ” Primary Dimension: Industrial form Confidence Level: 0.9 … Text Number: 2 “The proportion of business formats such as modern fashion, creative design, cultural and tourism consumption continues to expand … ” Primary Dimension: Industrial form Confidence Level: 0.9 … Text Number: 3 “In 2025, Chengdu International Fashion Industrial Park will give full play to the “early advantage” of digital transformation and development … ” Primary Dimension: Art performances Confidence Level: 0.9 …

.

2.6.2. Sentiment Analysis

This study utilized the Deepseek-R1 model with zero-shot learning to analyze the sentiment of the public’s text reviews. The criteria for quantifying sentiment were based on the five-point Likert scale, ranging from 1 to 5 (see

Table 4. Five-point Likert scale for measuring the performance of sentiment.

Sentiment Values	User Star Ratings	Sentiment Label	Quantitative Basis
5	5, 4.5	Very Positive	“very/too/extremely/completely” + positive word
4	4, 3.5	Positive	“Quite/comparatively/kinda/quite” + positive word
3	3, 2.5	Neutral	No clear emotional tendency or both positive and negative
2	2, 1.5	Negative	Adverbs of degree + negative words
1	1, 0.5	Very Negative	Strong adverbs + negative words

for details). Additionally, we divided sentiment scores into three levels: positive sentiment (scores of 4 and 5), neutral sentiment (scores of 3), and negative sentiment (scores of 1 and 2) for further interpretation.

This research employed a framework that integrated the sentiment scores from the LLM and the user star ratings from the Dianping platform. The rating is a discrete variable ranging from 1 to 5 stars, which can effectively reflect the public’s overall satisfaction. The concordance between sentiment scores and user star ratings is shown in Table 4. Combining sentiment scores with ratings allowed us to identify the potential discrepancies between the ratings and expressed emotions. For example, some reviews exhibited high ratings but conveyed negative expressions in the text. Analyzing these differences can provide targeted and complementary information for the planning and management of the parks. Additionally, we calculated average sentiment scores to reflect overall emotional inclinations towards various aspects using the following formula:

$$E_i={\displaystyle \sum _{s=1}^{5}s\cdot P_{is}}$$

(1)

where E_i is the emotional score for the perceptual dimension i, s is the sentiment score (ranging from 1 to 5), and P_is represents the proportion of dimension i at sentiment score s.

2.6.3. Image Analysis

In the image data analysis stage, the multimodal large language model GLM-4V-Plus-0111 (https://www.zhipuai.cn/) was employed. Developed by Zhipu AI and released on 11 January 2025, this model demonstrates exceptional capabilities in image recognition and multimodal comprehension. It translates key visual elements into semantically accurate and linguistically fluent textual descriptions.

The image analysis process consisted of two main stages, as illustrated in Figure 5. First, the Zhipu API was utilized to convert image data into accurate and coherent textual descriptions. Second, the generated textual data were automatically analyzed with the DeepSeek-R1 model. This procedure aligns closely with the previously described text analysis. Since images often contain rich and complex visual information, the analysis focused on the core visual elements to ensure accuracy. To minimize the potential for hallucinations of LLMs, comparative verification indicated that the results were most reliable when each image was matched with a sub-dimension, assigning each image to at most one sub-dimension. The detailed prompt engineering is presented in

Table 5. Image analysis prompt engineering.

Step 1: Image-Text Prompt Engineering.
Role	Role: You are an expert in destination image perception, specializing in industrial tourism, with expertise in: Destination image theory. Specialize in visual feature identify and multi-dimensional perception analysis.
Instructions	Your task is to recognize the core visual elements and key content within the image and generate semantically rich and coherent visual description text.
Steps	Image feature identify: the core visual content of the image is recognized by the enhanced visual module; Text generation: visual features are transformed into semantically rich and coherent text descriptions by the language generation module.
End Goal	Output structured analytical results, comprising the image sequence number and the image description content, to facilitate the transformation from image to text.
Narrowing	The description should be based on the visual elements visible in the image, and any speculation or derivation is prohibited. Output formatted in Markdown.
Input data	Image data released on the platforms by both the public and official.
Input
Output	There is a red mechanical device located on a circular grassy area. … The mechanical device appears to be part of a large grab bucket or crane, with a vibrant color.	The scene shows an indoor space with multiple paintings hanging on the wall, each varying in style and rich in color. The floor is gray, and several large metal frames are positioned within the space.	This picture depicts a modern architectural structure with a distinct industrial design style. … The building primarily consists of a metal framework with multiple floors inside.
Step 2: Text analysis prompt engineering.
Role	Consistent with the text analysis role description.
Instructions	Your task is to analyze the text in depth. Matching the text content with perceptual dimensions and outputting the best-matched dimension along with its confidence levels.
Steps	Text preprocessing: Identify the core semantic content of text reviews. Perceptual dimension matching: identify keywords and semantic features in the text. Match them with 12 perceptual dimensions and output confidence levels (0.0–1.0). Filter the dimensions with a confidence level ≥0.6 as valid matches. Result output: match the dimension with the highest confidence level, selecting at most one dimension.
End goal	Output structured analysis results, including image sequence number, perceptual dimensions matching and confidence levels. The results must be presented in a modularized and clearly structured format to effectively accomplish image-text analysis.
Narrowing	All matches are restricted to the perceptual dimensions listed in Table 2. Match at most one dimension Output formatted in Markdown.
Input	Image Serial Number: 1328776515.0_1 There is a red mechanical device located on a circular grassy area…	Image Serial Number: 1792521113.0_0 The scene shows an indoor space with multiple paintings hanging on the wall…	Image Serial Number: 1529806494.0_2 This picture depicts a modern architectural structure…
Output	Image Serial Number: 1328776515.0_1 Matching dimension: Industrial relics Confidence Level: 0.9	Image Serial Number: 1792521113.0_0 Matching dimension: Art performances Confidence Level: 0.9	Image Serial Number: 1529806494.0_2 Matching dimension: Spatial layout Confidence Level: 0.9

.

2.6.4. Validation

To validate the performance of the LLMs, 500 samples were randomly selected for manual annotation. To ensure unbiased model evaluation, this study distinguished between model tuning and evaluation datasets. Specifically, prompt engineering and parameter optimization were exclusively conducted on the tuning set. The final performance assessment was conducted on an independent test set to prevent data cross-contamination.

Five volunteers were invited to assess the results generated by the models. They were provided with the same prompt information as the LLMs, including role specifications, task descriptions, and steps. Cohen’s kappa was calculated to assess the consistency, achieving a score of 0.76, indicating a substantial level of agreement among the volunteers [65].

The model’s performance was evaluated across the following two tasks. First, in the sentiment analysis task, the model outputs sentiment scores ranging from 1 to 5. Comparing the model’s outputs with manually annotated data revealed that the model achieved an overall accuracy of 93% in this task. Second, in the perceptual dimension matching task, the model was assessed on its ability to identify whether a perceptual dimension “existing” or “non-existent”. A confusion matrix consisting of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) was employed to evaluate the model’s performance in this task. The overall performance was derived by calculating the average performance across all dimensions, ensuring that all dimensions contribute equally to the final result [66]. In the model evaluation, accuracy is the proportion of correctly predicted instances out of the total samples; precision represents the proportion of predicted positive instances that are actually positive; recall reflects the proportion of actual positive samples that are correctly identified, and the F1 score is the harmonic mean of precision and recall. The evaluation results yield an accuracy of 0.82, a precision of 0.83, a recall of 0.73, and an F1 score of 0.78. Overall, the LLMs have demonstrated strong performance in both the sentiment analysis and perceptual dimension matching tasks.

2.7. Multimodal Data Fusion and Discrepancy Quantification

2.7.1. Multimodal Data Fusion

The study utilizes a multimodal data analysis approach (Figure 6). Text and image data possess distinct advantages and informational focuses: text data is typically unstructured, reflecting subjective evaluations and experiences, while image data presents intuitive visual features. The study of single-modal has limitations. Relying solely on textual content may fail to capture visual characteristics, while relying on image data alone may lack detailed descriptions of public experiences. Integrating both textual and image modalities enables more comprehensive insight, improving the interpretability of the model [67]. The specific method for text and image data fusion is as follows:

$$F_d=T_d\cup I_d$$

(2)

T_d and I_d represent the sub-dimensions for text and images, and F_d is the fused dimensions.

To illustrate the fusion process concretely, we take an example of fusing text and image data generated under the same user ID in public perception data:

Text-matched perceptual dimensions (T_d): {AE,AP,EL}

Image-matched perceptual dimensions: Image 1(I_d1): {IR}, Image 2(I_d2): {AP}, Image 3(I_d3): {EL}

The fused dimensions (F_d): {AE,AP,EL,IR}

2.7.2. Frequency Distribution and Chi-Square Test

This study examined the distributional characteristics of official projections and public perceptions by calculating the frequency and percentage for each dimension. Chi-square test was conducted using SPSS 24.0 to determine whether statistically significant differences existed between the two across these dimensions. The proportion of the perceptual dimension was calculated using the following formula:

$$P_i={\displaystyle \frac{C_i}{{\displaystyle {\sum }_{j=1}^{12}{C}_j}}}$$

(3)

where P_i is the proportion of i-th perceptual dimension, C_i is the frequency of the i-th perceptual dimension, and the denominator is the total frequency across all perceptual dimensions.

3. Results

3.1. Multimodal Data Fusion Results

To examine the differences and complementarity between text and image data in characterizing the features of the parks, and to validate the effectiveness of multimodal data integration, this section conducts a statistical analysis of the text and image results. The analytical results for text, image, and multimodal data are presented in Figure 5.

3.1.1. Statistical Results of Public Perceptions

As shown in Figure 7a, textual and image data exhibit differences and complementarity in the distribution of public perceptual dimensions. In textual data, the public focuses more on the AE (23.48%) and AP (12.98%). In image data, IR (19.35%) and EL (15.79%) account for higher proportions. This indicates that text and image data possess distinct informational focuses: text expresses subjective perceptions, while images convey intuitive visual characteristics. After fusing text and image data, the distribution of dimensions becomes more balanced. Based on a comparison of data before and after multimodal fusion, the coverage rate of text-based perceptual dimensions increases by 61.75% compared to the text-only modality, while the coverage rate of image-based perceptual dimensions increases by 81.75% compared to the image-only modality. These results indicate that multimodal fusion effectively enhances the representational capacity of single-modal data. By integrating the subjective expressions conveyed through text with the visual characteristics captured in images, this approach enables a more nuanced and comprehensive understanding of public perception.

3.1.2. Statistical Results of Official Projections

As illustrated in Figure 7b, differences are also observed in the dimensions of official projection. Textual data emphasizes the dimensions of IR (18.45%) and AP (10.92%), while image data reveals higher proportions for IR (33.33%) and SL (12.90%). Due to the limited scale of official image data, the integrated data shows no significant deviation compared with single text data. However, even with constrained data, incorporating image data leads to a more balanced data structure. This result also indicates that multimodal data can compensate for the limitations of a single data type, enhancing the comprehensiveness and interpretability of the data.

3.2. Quantification and Validation of Discrepancies Between Official Projections and Public Perception

By calculating the proportion distribution, we found significant discrepancies between official projections and public perceptions across multiple dimensions (

Table 6. Frequency distribution and chi-square test of official projections and public perceptions across dimensions.

Dimension	Sub-Dimension	Official Projected		Public Perceived		χ2	p	φ
		Frequency	Percentage	Frequency	Percentage
Cultural value of heritage	Industrial relics	692	11.56%	11,600	12.89%	8.929	<0.001	0.0096
	Historical ambiance	563	9.4%	3137	3.49%	530.305	<0.001	0.0743
	Cultural memory	540	9.02%	3094	3.44%	479.953	<0.001	0.0707
	Nostalgic feelings	179	2.99%	1372	1.52%	75.783	<0.01	0.0281
	Total	1974	32.97%	19,203	21.34%
Spatial environmental perception	Spatial layout	439	7.33%	7460	8.29%	6.818	<0.001	0.0084
	Environmental landscape	279	4.66%	9259	10.29%	198.752	<0.001	0.0455
	Total	718	11.99%	16,719	18.58%
Creative industry form	Art performances	691	11.54%	13,821	15.36%	63.745	<0.001	0.0258
	Design innovation	468	7.82%	3025	3.36%	317.698	<0.001	0.0575
	Industrial form	1044	17.43%	8032	8.92%	474.958	<0.001	0.0703
	Total	2203	36.79%	24,878	27.64%
Service experience perception	Supporting facilities	183	3.06%	8381	9.31%	270.429	<0.001	0.0531
	Atmosphere experience	661	11.04%	14,070	15.63%	91.237	<0.001	0.0308
	Consumption experience	249	4.16%	6748	7.5%	92.652	<0.001	0.031
	Total	1093	18.25%	29,199	32.44%

). The chi-square test results indicated statistically significant differences across all dimensions (p < 0.01), though the effect sizes were relatively small. This finding is likely attributed to the large sample size in this study, which enhances statistical power and enables the detection of systematic albeit minor non-random differences. The interpretation of effect sizes should be considered within the specific research context [68]. Comparisons with relevant literature reveal that the observed p-values and effect sizes in this study align with findings from previous studies [29,42]. Given that p-values can be highly sensitive to sample size in large datasets, the present analysis primarily relied on comparing frequency distributions to quantify differences, with p-values and effect sizes serving as reference indicators.

In official projections data, the Creative industry form and the Cultural value of heritage were the most frequently projected dimensions. In public perceptions data, Service experience perception and Creative industry form emerged as the most prominent perceived dimensions. We found that both official and public perspectives recognize the Creative industry form dimension. However, at the sub-dimensions, the official government highlights the IF, accounting for 17.43% and the public perceived AE more frequently, at 15.63%, as shown in Figure 8.

3.3. Distribution of Sentiment

As shown in Figure 9a, the sentiment analysis results based on the LLM indicate that the sentiment values of public reviews toward the parks exhibit a distinctly right-skewed distribution. Reviews with a sentiment score of 4 account for 49.2%, representing the highest proportion among all sentiment values. After classifying sentiment values into three categories, the results reveal that positive reviews dominate, comprising 71.4% of the total. Negative reviews account for 6.15%, while neutral comments constitute 22.45%. This distribution suggests that the public is generally satisfied with the parks. Negative emotions make up a relatively small proportion.

Figure 9b displays the distribution of user ratings, a discrete variable ranging from 1 to 5. Ratings of 5 account for 62.5%, while ratings of 4 constitute 32.2%. Positive emotions account for 94.7%, indicating that most of the public delights in the parks. Ratings of neutral emotions make up 4.2%. In general, sentiment scores and user ratings demonstrate a clear positive trend in public perceptions.

A further comparison of the sentiment analysis results from the LLM with the user ratings reveals that while the overall emotional tendencies of both are similar, subtle differences exist in their specific distribution structures. Specifically, for positive sentiment proportion, the LLM results indicated 71.4%, and user ratings accounted for 94.7%. For neutral sentiment, the LLM identified a proportion of 22.45%, while user ratings constituted only 4.2%. This discrepancy stems from the distinct scoring mechanisms. The LLM sentiment values employ integer ranges from 1 to 5 and classify sentiment strictly according to the five-point Likert scale. For example, “quite good” is categorized as a score of 4, “excellent” receives a score of 5, and content with no clear emotional tendency or mixed positive and negative cues receives a score of 3. In contrast, user ratings are a discrete variable ranging from 1 to 5 stars, with each star interval containing two values, providing the public with a broader range of options. Consequently, some reviews with high user ratings were classified as neutral by the LLM due to the unclear emotional cues or negative expressions in the textual content. Detailed text mining of low-rated and neutral reviews reveals that public dissatisfaction mainly arises from high parking fees, limited diversity in industrial formats, expensive pricing, inadequate supporting facilities, and unclear signage systems.

In addition, calculating the average scores across dimensions indicates that the public expresses stronger positive sentiments in sub-dimensions such as DI (4.05), CM (4.02), and NF (4.02). Negative emotions are mainly manifested in dimensions including SF (3.79), CE (3.86), and AE (3.94). This indicates that deficiencies in management services and spatial planning within negatively impact the public’s emotional experiences. These findings emphasize the need to improve service quality and optimize spatial planning.

4. Discussion

Measuring the cognitive gaps between official and public perspectives has significant practical implications for prompting sustainable development. Quantifying the disparities between official projections and public perceptions across perceptual dimensions in parks offers valuable strategic insights for management and planning.

4.1. Mapping and Quantification of Perceptual Dimensions

This study employed LLMs and multimodal data analysis to identify the focus of official and public attention toward parks and measure the perceptual gaps between these two subjects. By designing a structured prompting framework and leveraging the API of the Deepseek-R1 and GLM-4V-Plus-0111 models, we mapped textual and image data to perceptual dimensions, establishing an analytical framework for multimodal fusion data. Then, we quantify the perceptual discrepancies by calculating the distribution frequencies of perception dimensions in both official and public datasets. The results indicate that LLMs demonstrate strong performance in semantic comprehension, visual recognition, and sentiment analysis. Recent studies have also highlighted the application potential of LLMs in text and visual analysis [18]. Multimodal data fusion achieves information complementarity, providing a more nuanced and comprehensive understanding of the research. The continuous advancement of model technology and the expansion of parameter scale significantly enhance the ability to analyze and multimodal data. Compared to previous studies based on machine learning methods, a key advantage of LLMs is their reduced dependency on large-scale annotated data [49,51]. Deep learning models often require large volumes of manually labeled samples for domain-specific training [69]. Studies have illustrated that LLMs can accomplish complex text, image, and sentiment analysis tasks through zero-shot or few-shot learning. They have demonstrated strong performance across diverse research fields, including tourism destination image analysis [7071,72], urban planning [73], and environmental perception studies [74].

4.2. Discrepancies Between Official Projections and Public Perceptions

This study employs LLMs to analyze and integrate multimodal data. It enables a quantitative analysis of perception disparities by statistically comparing the distribution between official and public perceptions across multiple dimensions. The results indicate that significant discrepancies between official and public perceptions in the perceptual dimensions. These findings are consistent with prior research on cognitive biases between official and public perspectives [40,75]. Such discrepancies originate from inherent differences in position and needs between the official and the public, resulting in inconsistent spatial perceptions and expectations of the creative industry park among these two parties. From the official perspective, as the leading actor in spatial planning and development, it is grounded in macro-level narratives and overall development strategy [37]. Furthermore, in promoting the adaptive reuse of industrial heritage, the official emphasizes balancing historical value with functional repurposing, while also considering multiple factors, including economic benefits, sustainable development, and cultural heritage [12,14]. Additionally, during the implementation phase, practical factors such as technical feasibility and budget constraints often result in deviations between the final renovation outcomes and the initial planning. Although some projects conducted preliminary social surveys in the early stages, public participation remains insufficient. From the public perspective, as users and experiencers of the space, their evaluations are largely based on the immediate sensory experiences within the environment [76]. The public places greater emphasis on specific micro-level perceptions, such as the completeness of supporting facilities, artistic ambiance, and spatial accessibility. The high levels of public perceptions of space reflect intense demands and expectations. Identifying these specific needs provides empirical evidence for enhancing the park’s spatial planning and management. Therefore, how to balance and coordinate the needs and expectations between the two groups is the core issue to enhance the consistency of the tourism destination image.

Specifically, the research results indicate that the most pronounced discrepancies exist in the Service experience perception dimension, followed by the Cultural value of heritage dimension. The public pays more attention to the Service experience perception dimension, particularly perceived more frequently than officially projected on the SF and CE dimensions. This aligns with previous research findings that supporting facilities constitute the core category for tourists in destination image perception [77,78]. However, official projections in this dimension are relatively limited. Further sentiment analysis displays that inadequate supporting facilities and poor consumption experience are the primary factors eliciting negative emotions, which is consistent with the result in cultural heritage research showing that the evaluation of facilities is generally low [79]. This suggests that these issues should be prioritized for improvement in future management and planning work. The official perspective emphasizes the Cultural value of heritage dimension, focusing on the dissemination of historical culture and values, particularly on HA and CM. Previous studies have also indicated that official entities have over-interpreted the importance of historical value [32]. These apparent discrepancies between official projections and public perceptions not only undermine the effectiveness of official communications but may also trigger public disappointment and negative evaluations.

Secondly, the results indicate a relatively high degree of consistency in the Creative industry form dimension, with a high proportion and ranking in both official and public data. Suggesting that it constitutes a key perceptual dimension within the parks, consistent with existing research findings showing artistic value holds significant importance in both official and tourist perceptions. It also aligns with the creative-oriented characteristics of this reuse model [4]. Further analysis of sub-dimensions reveals that the public pays more attention to the AP dimension. Previous studies have also indicated that the public shows greater interest in art activities in creative industrial parks [17]. In contrast, official entities focus more on the IF and DI aspects, reflecting a macroscopic perspective in overall planning [80].

Thirdly, the results also show that the proportion of the Spatial Environmental Perception dimension remains relatively low in both official projections and public perceptions. This suggests that the dimension is not a central focus within the parks. This finding is consistent with the previous studies, indicating that environmental landscape elements are frequently mentioned in public reviews but are not the spotlight of their expressions [81]. However, this contrasts with another study on tourist satisfaction with the reuse of industrial heritage [82], which may be attributed to differences in landscape dimension classification, as that study included artworks and installation elements in the landscape dimension. From the perspective of spatial environment perception alone, it does not belong to the core perception dimension. Negative sentiment analysis indicates that environmental quality, spatial accessibility, and visual guiding systems significantly influence public satisfaction. These findings align with prior studies, which have shown that low-quality environments diminish spatial aesthetic appeal and are more likely to invoke negative evaluations [83,84,85]. Therefore, it is recommended that official entities improve environmental cleanliness, greenery quality, and the clarity of signage systems, while optimizing spatial layout to enrich overall environmental experiences and mitigate negative emotions.

4.3. Implications for Operational Management and Planning

By examining the discrepancies between official projections and public perceptions, as well as sentiment analysis, this study can identify dimensions where public experience can be enhanced and aspects that require improvement. Based on the results, the following strategies are proposed:

In the Service experience perception dimension, improving supporting facilities and the consumption experience are crucial for enhancing public [86]. Sentiment analysis indicates lower satisfaction levels with supporting facilities and the consumption experience. To enhance overall satisfaction, in the supporting facilities dimension, the park management should optimize supporting facilities by conducting regular inspections and maintenance, thereby ensuring that basic service facilities remain fully functional and operate reliably. Additionally, training should be implemented for park staff to enhance their overall service capabilities and service awareness. In the consumption experience dimension, local tourism authorities should collaborate with the park management to regularly monitor commodity prices, ensuring that they remain within reasonable ranges. Furthermore, park management should establish a multi-channel public feedback mechanism to gather public issues and suggestions. Through these measures, public satisfaction regarding supporting facilities and consumption experiences will be enhanced.

In the Cultural value of heritage dimension, the public does not perceive HA and CM frequently, but shows strong emotional resonance within these dimensions. This also demonstrates that historical authenticity and cultural identity play a crucial role in stimulating positive public sentiment. Therefore, the park operators should adopt a comprehensive strategy to systematically present the tangible and intangible cultural value of industrial heritage. Specifically, park operators can adopt visualization technologies [87,88,89,90] and historical narrative strategies [91] to facilitate diversified digital expressions and scenarized displays of industrial heritage. This helps construct immersive and interactive spaces that enhance the experience and deepen public engagement. Furthermore, creative design should be employed to incorporate industrial, historical, and cultural symbols into the spatial environment in more tangible and perceptible forms.

In the Creative industry form dimension, research has demonstrated that a creative atmosphere exerts a significant positive effect on enriching public experiences and fosters place attachment [92]. Based on the public’s frequent perception of the AP, park operators can strengthen the park’s artistic identity through visual forms such as graffiti and installation art, enhancing visual impact. Addressing the significant divergence between official and public perceptions regarding IF, park management should introduce cultural and creative industries, trendy consumption, and other business forms into its investment promotion and industrial planning. This strategy promotes the diversification and innovation of business formats, thereby enhancing the park’s vitality and public recognition.

In the Spatial environmental perception dimension, environmental quality and spatial accessibility significantly impact public satisfaction, particularly given the relatively low satisfaction levels in environmental quality and spatial accessibility. To address these issues, park management authorities should establish a routine environmental cleaning and maintenance mechanism while improving the quality of greenery and landscaping to maintain the park’s cleanliness and aesthetic appeal. Furthermore, park operators should consider collaborating with relevant transportation authorities during peak tourist seasons to increase the accessibility of public transport. Concurrently, they should optimize the clarity of internal signage systems to reduce congestion. Particular emphasis should be placed on improving transportation routes and spatial layouts around Instagrammable spots and high-traffic areas.

4.4. Limitations and Future Research

Although this study effectively quantifies the perception discrepancies between official and public viewpoints, several aspects still need to be addressed. These limitations include the following: (1) The public data mainly comes from online platforms, and the user base consists mainly of young people. This may lead to bias in the representativeness of the research samples. Future studies may incorporate questionnaire surveys and field investigations to improve data coverage and representativeness, thereby enabling a deep exploration of the conclusions. (2) Official image datasets are relatively limited in volume. Future research can explore additional channels to expand the dataset, such as incorporating official promotional video data. (3) Aligning different data modalities still faces challenges related to data heterogeneity. With the advancement of technology, future studies may introduce more sophisticated MLLMs to enhance sensory simulation and improve multimodal analysis, paving the way for broader applications in perception analysis research. (4) This study primarily focuses on the creative industry park model, with limited comparative analysis of other industrial heritage renewal models, such as industrial heritage parks and museum types. Different types may influence the perceived structure and preferences of the public. Future research could broaden its scope to include multiple types of industrial heritage reuse models, thereby improving the generalizability and guiding value of the research findings.

5. Conclusions

Research on the discrepancy between official projections and public perceptions is crucial for shaping successful tourism destination images. This study validates the effectiveness of LLMs for multimodal data analysis and confirms their applicability in destination image perception research. Through the comparison of official and public multimodal data, our research reveals the following findings:

(1)

Discrepancies exist between official projections and public perceptions across dimensions. Official narratives adopt a macro perspective, emphasizing the park’s historical and cultural value, and functional layout. Consequently, this places greater emphasis on promoting dimensions such as HA, CM, and IF. However, constrained by the spatial presentation of heritage elements, the public’s overall perception of this dimension remains relatively weak. Therefore, subsequent planning should enhance the visual representation and scenarization of heritage elements to bridge this cognitive gap. As primary users of the space, the public prioritizes sensory experiences and functional services. Consequently, their perception intensity regarding dimensions such as supporting facilities and consumption experience within the Service experience perception is significantly higher than that of official projections. This disparity reflects the differing positions and needs between the two parties. Furthermore, both official projections and public perceptions exhibit a high proportion of attention to the Creative industry form dimension, highlighting the need to sustain and further enhance this dimension in future park operations.

(2)

The public’s overall sentiment toward the park is predominantly positive, with particularly high satisfaction levels associated with the DI, CM, and NF dimensions, which serve as key factors driving positive sentiment. Negative sentiments are primarily concentrated in the SF and CE dimensions, where issues such as inadequate supporting facilities, poor consumer experience, and limited spatial accessibility significantly impact public satisfaction.

(3)

This study develops an analytical framework utilizing LLMs and multimodal data to quantify perceptual discrepancies. By employing structured prompts and LLM APIs, this approach facilitates the systematic analysis of both textual and image data. The integrated data provides a more holistic representation of the park’s perceptual image, significantly enhancing model interpretability. Furthermore, it also provides a reusable analytical pipeline for similar multimodal research.

In summary, these findings provide valuable insights for park operations management and planning. It is recommended that official planners and park operators should prioritize addressing factors with significant discrepancies that substantially impact public satisfaction, thereby enhancing overall satisfaction and promoting the parks’ sustainable development.