As the core component of a city's historical heritage, historic districts encompass both tangible architectural relics and intangible cultural heritage, serving as symbolic representations of urban context and continuity (Willis, 2014). However, with the acceleration of urbanization and economic development, profound changes have taken place in urban living environments. Historic districts are now facing numerous challenges, including the degradation of their cultural ecosystems, the loss of historical heritage, and the homogenization of commercial functions. Historic districts embody multiple values and fulfill social, cultural, economic, and ecological functions. Their architectural and cultural heritage plays a vital role in shaping urban quality and enhancing community well-being, which is crucial for promoting sustainable urban development (Ricci, 2022).

On the other hand, as public demand for spiritual and cultural fulfillment continues to grow, Cultural Ecosystem Services (CES) — as a key concept linking the natural environment with human well-being — has garnered increasing attention. In 2005, the Millennium Ecosystem Assessment (MA) defined CES as “the non-material benefits people obtain from ecosystems through spiritual enrichment, cognitive development, reflection, recreation, and aesthetic experiences.”(Assessment, 2005) In the past, due to data scarcity and the lack of unified evaluation indicators, CES research has faced challenges in conducting quantitative analysis (Feld et al., 2009). Traditional data collection methods have primarily relied on questionnaires and interviews conducted during field surveys. While these approaches can yield relatively accurate data, they are often time-consuming and labor-intensive (Brown & Fagerholm, 2015).

In recent years, social media data has emerged as a novel data source, offering a new non-monetary approach to assessing CES. By mining multimodal content—such as text, images, and videos—from social media platforms, researchers can effectively capture public perceptions and evaluations of CES (Cheng et al., 2019). Social media data, characterized by its large sample size, ease of access, and low cost, has gradually become an important paradigm in the field of CES research. Moreover, existing studies on historic districts based on social media data have primarily focused on single dimensions, such as tourism experiences or perceptions of the physical environment from the perspective of visitors(Ginzarly et al., 2019; Xie et al., 2022), while lacking a systematic analysis of the cultural ecosystem of historic districts. To address this gap, this study innovatively constructs a multidimensional framework that integrates topic modeling, sentiment analysis, spatial clustering, and Asymmetric Impact–Performance Analysis (AIPA). By mining online reviews from visitors, the framework aims to comprehensively reveal tourists’ perceptions and evaluations of various types of CES. This study not only clarifies the components and spatial distribution of CES in historic districts, but also provides an in-depth analysis of the impact mechanisms of different types of CES on visitor satisfaction. The findings offer a scientific basis for the optimized management of CES in historic districts and hold important practical implications for enhancing residents’ well-being.

The remainder of this paper is organized as follows. Section 2 provides a systematic review of research progress on historic districts and CES, identifying the limitations of existing studies and the key issues this study aims to address. Section 3 details the study area, data collection process, and research methodology. Section 4 presents the results derived from the analysis of online comments on historic districts. Section 5 discusses the findings related to CES in historic districts and highlights both the theoretical and practical implications of the study. Finally, Section 6 summarizes the conclusions and limitations of this research, while outlining directions for future study.

As a vital representation of living heritage, historic districts refer to culturally protected areas that reflect the socio-economic structures, cultural characteristics, and spatial patterns of specific historical periods, while preserving traditional architectural styles, ways of life, and a strong sense of local identity (Lu et al., 2015). Existing research primarily focuses on two traditional dimensions: first, the exploration of physical spatial conservation models (Zhang et al., 2023; Zhang et al., 2017), such as the preservation of architectural styles, block layouts, and historical landscapes; second, the study of vitality enhancement mechanisms (Huang et al., 2023; Wu et al., 2022), including strategies related to functional upgrading and economic revitalization of historic districts.

With the shift in urban renewal concepts from “physical space remediation” to “cultural-oriented” approaches, an increasing number of studies have begun to integrate social, cultural, and ecological factors. This transformation contributes to promoting urban economic growth, optimizing urban landscapes, and enhancing overall urban competitiveness (Chiu et al., 2019; Evans, 2020; Miles & Paddison, 2005). Against this backdrop, research on historic districts has demonstrated new academic trends. Taking heritage tourism as a starting point, Lu et al. developed a tourist experience framework for historic districts based on visitor engagement and destination image (Lu et al., 2015). In the same year, other scholars revealed the practical pathways through which Eastern European countries leveraged historical heritage to drive regional development (Ismagilova et al., 2015) In 2025, Wang et al. innovatively employed quantitative analysis of social media comments to assess the components of tourism attractiveness in historic districts (J. Wang et al., 2025). Jia et al. investigated sustainable tourism in natural and cultural heritage sites through studies on tourist density. In addition, ecological theories have also been introduced into the study of historic districts. In addition, ecological theory has been introduced into the study of historic districts, with some scholars adopting a cultural ecology perspective to develop heritage value protection systems and sustainable development strategies for traditional villages (Fang & Li, 2022; Zhong et al., 2025). While others have proposed ecological development models specifically for historic districts (Hu & Gong, 2017). However, despite the growing body of research on the cultural ecology of historic districts in recent years, there remains a significant gap in the exploration of CES within these areas.

The widely accepted concept of CES originates from the Millennium Ecosystem Assessment, which classified ecosystem services into four categories: provisioning services, regulating services, supporting services, and cultural services. CES encompasses a wide range of aspects, including spiritual and religious services, recreational and ecotourism experiences, aesthetic value, inspiration, and educational benefits (Assessment, 2005), It represents the cultural and spiritual values that humans derive from ecosystems through non-material means.

Early CES research was primarily concentrated in developed countries and regions such as Europe and North America (Bagstad et al., 2017; Gould et al., 2014; Graves et al., 2019). However, in recent years, with the growing emphasis on urban sustainability, developing countries have increasingly emerged as new focal areas for CES research(Xu et al., 2020; Xu et al., 2025; Xu et al., 2018). In terms of research content, early CES studies primarily focused on single functional dimensions, such as the quantitative analysis of aesthetic value and recreational functions (Bieling et al., 2013; Brancalion et al., 2014) More recent research has gradually shifted toward comprehensive assessments of multiple CES types and the exploration of public perceptions. Some scholars have emphasized the subjectivity of CES by using computer vision methods to identify preference differences among various groups for different CES types (Huai et al., 2022). Other studies have examined the mechanisms influencing CES from perspectives such as social well-being, spatial distribution, and supply–demand matching (Bing et al., 2021; Kosanic & Petzold, 2020; Zhang & Liu, 2024).

In terms of research methodology, the intangible and subjective nature of CES makes it difficult to quantify. Although early studies attempted to assess its economic contribution using monetary valuation methods, the strong influence of individual differences on CES renders it challenging to assign accurate monetary values (Chan et al., 2012). As a result, most studies have relied on methods such as questionnaires, evaluation models, and participatory mapping. In recent years, with the rise of big data, CES research has increasingly incorporated emerging techniques such as remote sensing, GIS-based spatial analysis, and social media text mining (Gan et al., 2024), enhancing both the objectivity of data collection and the ability to visualize spatial patterns.

The study of CES in urban areas has become a key topic in the field (La Rosa et al., 2016). However, existing urban CES research shows a clear spatial bias, with a strong focus on natural ecosystems such as green spaces, parks, and wetlands(Beckmann-Wübbelt et al., 2021; Dickinson & Hobbs, 2017; P. Wang et al., 2025). In contrast, research on historic districts—complex spatial units that integrate cultural, tourism, and residential functions—remains relatively underdeveloped.

In summary, the cultural ecological challenges of historic districts have increasingly drawn scholarly attention. By introducing the CES framework, this study aims to identify the aesthetic value, cultural identity, recreational experiences, and educational inspiration that visitors derive from historic districts, thereby providing scientific guidance for the optimization and enhancement of cultural services within these areas. Although some studies have analyzed CES in historic districts using tourist reviews, there remains a lack of systematic analysis in areas such as the identification of CES types, spatial distribution patterns, and underlying impact mechanisms. To address these research gaps, this study focuses on five representative historic districts in Changsha and employs social media comment mining techniques to identify visitors’ subjective perceptions and cultural service experiences during their visits. This approach has been proven effective in previous research. The following research questions (RQs) are addressed:

Located in the northeastern part of Hunan Province, in the river valley plain of the lower Xiangjiang River, Changsha is one of China’s first 24 nationally designated Famous Historical and Cultural Cities. It boasts over 3,000 years of history and more than 2,400 years of continuous urban development. According to 2024 statistics from the Changsha Municipal Bureau of Culture and Tourism, the city received a total of 24.65 million tourist visits during public holidays, reflecting the rapid growth of its cultural tourism sector. The study focuses on five of the most representative historic districts in Changsha: Taiping (TP), Chaozong (CZ), Duzheng (DZ), Xiyuanbeili (XYBL), and Baiguoyuan (BGY) (Fig. 1). As the core area of Changsha’s Historic and Cultural City Conservation Zone, these districts encompass 42 historic streets and alleys, over 30 protected cultural heritage sites, and a large number of well-preserved historical buildings. They form a temporally and spatially continuous cultural landscape sequence, making them ideal case samples for investigating the CES of historic districts.

This study is based on user-generated comment data collected from two major platforms: Dianping (https://www.dianping.com) and Xiaohongshu (https://www.xiaohongshu.com). Founded in 2003, Dianping is one of China’s earliest local lifestyle service platforms. As of December 2024, it had 121 million monthly active users and has accumulated a vast amount of authentic and reliable consumer review data, particularly authoritative in areas such as dining and tourism. Xiaohongshu, launched in 2013 as a social e-commerce platform, has evolved into a major lifestyle-sharing space for young users, with 500 million monthly active users. The two platforms have complementary user bases and offer diverse review dimensions, providing both breadth and depth in data support for CES research in historic districts. User review data up to March 2025 were collected using Python-based web scraping techniques, resulting in an initial sample of 19,508 entries. Among the five districts, TP historic district—being the core tourist attraction in Changsha—had the highest number of reviews, while the newly developed XYBL district had relatively fewer comments. To ensure data quality, a data cleaning process was implemented, including the following steps: (1) Removing duplicate entries and non-original content; (2) Excluding reviews with incorrect geolocation or those related to similarly named but unrelated areas; (3) Filtering out low-information short comments, such as overly brief remarks with fewer than 10 characters (e.g., “very good”). After cleaning, a total of 18,794 valid reviews were retained for analysis.

During the text preprocessing stage, the study utilized the JIEBA word segmentation tool in Python, which is widely recognized for its effectiveness in Chinese natural language processing. The preprocessing workflow included the following steps: (1) Filtering out special characters, emoji tags, and other non-textual noise; (2) Building a customized dictionary that includes location-specific proper nouns such as “Fengying Xili” and “Xiangjiang Review” to ensure accurate identification of regional terms; (3) Integrating standard stop word lists from Baidu, Harbin Institute of Technology, and Sichuan University, and supplementing them with a customized stop word set (e.g., irrelevant place names mentioned in reviews) to effectively remove common but semantically insignificant words.

This study integrates both qualitative and quantitative approaches to construct a multidimensional analytical framework (Fig. 2). By mining online reviews from tourists, it aims to uncover the composition, spatial distribution, and satisfaction impact mechanisms of CES in historic districts. The specific research methods are as follows: (1) BERTopic is employed to extract thematic topics from tourist reviews, identifying the key components of CES. (2) A Large Language Model (LLM) is used for sentiment analysis to quantify tourists’ emotional tendencies, capturing their preferences and evaluations of CES in historic districts. (3) Natural Breaks (Jenks) classification in ArcGIS is applied for spatial clustering analysis to investigate the spatial aggregation patterns of CES. (4) Asymmetric Impact–Performance Analysis (AIPA) is conducted to explore the mechanisms through which different types of CES influence visitor satisfaction in historic districts. This analytical framework not only avoids social desirability bias common in traditional surveys—providing an authentic reflection of tourists’ unprompted experiences—but also reveals the spatial patterns of host–guest interactions, offering a scientific basis for precise planning and management of historic districts.

To address RQ1, BERTopic is employed to identify the key components of CES from tourists’ online reviews. BERTopic is a topic modeling method based on the deep learning BERT model, which combines BERT’s semantic understanding capabilities with the high efficiency of the HDBSCAN clustering algorithm. This approach allows for in-depth analysis of review texts, extraction of semantic features related to historic districts, discovery of latent themes, and identification of their spatial distribution and evolutionary patterns. The modeling process consists of the following four steps:

(1) Semantic Vectorization: A pre-trained BERT model is used to convert review texts into high-dimensional semantic vectors. This step lays the foundation for subsequent analysis by capturing contextual semantics.

(2) Dimensionality Reduction: UMAP (Uniform Manifold Approximation and Projection) is employed to reduce the dimensionality of the text vectors, mapping them from a high-dimensional to a low-dimensional space. The specific UMAP parameter settings are shown in Table 1.

(3) Clustering Analysis: HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) is applied to cluster the low-dimensional word vectors. This density-based algorithm determines the density of each point by calculating its core distance. The HDBSCAN parameter configuration is also presented in Table 1.

(4) Topic Representation: c-TF-IDF (Class-based Term Frequency–Inverse Document Frequency) is used to calculate the keyword weights for each topic. As an extension of the traditional TF-IDF method, c-TF-IDF emphasizes the significance of terms within specific topics by accounting for both term frequency and inverse document frequency(Gokcimen & Das, 2024). The equation is as follows:

Here, c represents a specific topic, A denotes the average number of words per topic, tf_{t, c} indicates the frequency of term t within topic c, and tf_t represents the frequency of term t across all topics.

Through this process, the study achieved an accurate transformation from a large volume of review texts (n = 18,794) into interpretable topics. After iterative optimization, a total of 29 semantic topics were identified. These topics were then matched with CES categories in historic districts based on a synthesis of topic keywords, relevant literature, and expert consultation.

To address RQ2, this study employs a sentiment analysis method based on a large language model, selecting ChatGPT-4o as the core analytical tool to achieve accurate identification and quantitative evaluation of sentiment orientation in texts. Compared to traditional sentiment analysis approaches based on rule-based dictionaries or machine learning, ChatGPT-4o demonstrates superior performance in fine-grained sentiment recognition tasks. As a key subfield of natural language processing, sentiment analysis automatically detects the subjective emotional orientation expressed in texts, typically categorized as neutral or polar (positive or negative)(Wilson et al., 2005). In this study, the text data processed through topic modeling were first input into the ChatGPT-4o model via an API interface. Leveraging the model's powerful semantic understanding and contextual awareness capabilities, the texts were transformed into semantic vectors and classified by sentiment. The model is able to identify various emotional expressions, including positive sentiments such as “joy” and “satisfaction,” negative sentiments such as “anger” and “disappointment,” as well as neutral statements without clear emotional tendencies. During the sentiment quantification phase, a standardized 0–1 scoring system was adopted: values closer to 1 indicate more positive sentiment, while values closer to 0 reflect more negative sentiment. By calculating the sentiment probability of each individual text, an individual sentiment score is derived, which is then used to generate an overall sentiment score for each topic.

To address RQ3, this study applies the Natural Breaks (Jenks) method to conduct spatial clustering analysis of CES distribution within historic districts. Proposed by George Frederick Jenks, this method identifies statistically significant classification boundaries by minimizing within-class variance and maximizing between-class variance(Coulson, 1987). Compared to subjective classification approaches, the Natural Breaks method provides a more scientifically grounded means of revealing spatial aggregation patterns of CES.

The study first integrates CES location data extracted from topic modeling, and based on the c-TF-IDF values of each point, applies the Natural Breaks method to classify the data into five levels of significance. These levels correspond to a continuous gradient from “low clustering” to “high clustering,” thereby providing a basis for policy formulation and resource allocation in historic districts.

The calculation process of the method involves the following equations:

Here, represents the mean of the full sample; Z represents the mean of the clusters. Goodness of Variance Fit (GVF) is the indicator of the goodness of variance fit, Sum of Deviations About the Mean (SDAM) is the sum of squares of deviations from the overall mean, Sum of Deviations About the Cluster Means (SDCM) is the sum of squared deviations of each cluster center. The closer the GVF is to 1, the better the performance.

To address RQ4, this study employs AIPA method to evaluate the influence of different CES types on visitor satisfaction. Unlike the traditional Importance–Performance Analysis (IPA), which assumes a linear and symmetric relationship, AIPA is capable of capturing asymmetrical effects, making it better suited to the data structure of review texts and avoiding the high cost and bias associated with subjective “importance” ratings in traditional surveys(Yuan et al., 2018). Moreover, AIPA can identify the underlying mechanisms by which influencing factors affect satisfaction, providing a foundation for developing targeted improvement strategies(Caber et al., 2013).

AIPA categorizes CES factors into three types: (1) basic factors: Poor performance in these areas significantly reduces visitor satisfaction, but excellent performance offers limited enhancement to satisfaction; (2) These show a symmetric linear relationship between performance and satisfaction; (3) Strong performance can greatly enhance satisfaction, while their absence does not cause significant dissatisfaction(Caber et al., 2013).

In the specific analysis, this study created two binary dummy variables for each type of CES based on their sentiment scores. When a CES sentiment score falls within the lowest quartile, the “basic factor” dummy variable is assigned a value of 1 (and 0 otherwise); when the score is in the highest quartile, the “excitement factor” dummy variable is assigned a value of 1 (and 0 otherwise). Multiple regression analysis is then used to obtain the penalty index (p_i) and reward index (r_i) for each CES type, representing the impact of low or high CES performance on overall visitor satisfaction(Jiang et al., 2025). The equations are as follows:

Here, RIOS_i is the range of influence on satisfaction of CES type i, SGP_i and DGP_i denote the satisfaction and dissatisfaction generating potential of CESi, respectively. the value of IA_i can range from − 1 to 1. Based on the asymmetry of their influences, CES are categorized into basic (-1 ≤ IA_i < -0.1), linear (-0.1 ≤ IA_i ≤ 0.1) and excitement factors (0.1 < IA_i≤1).

The experimental results of the BERTopic model are presented using Topic word cloud map, The hierarchical clustering for topics, The hierarchical documents and topics and The similarity matrix of topics.

Based on the model, a total of 29 topics were extracted from the review texts of the five historic districts (Fig. 3). In the Topic word cloud map, the size of each word visually represents its importance within the topic (c-TF-IDF value). Through semantic analysis, the cultural connotations of each topic were systematically categorized. Specific examples are as follows:

(1) TP District – TP-Topic1: It includes keywords such as “art museum,” “museum,” and “artist,” reflecting the role of cultural facilities in enhancing visitors’ cognitive experiences. This topic is categorized as "Cultural Education and Cognitive Enrichment."

(2) CZ District – CZ-Topic2: It centers on historical buildings such as “Chaozong Gate,” “Shiwu School,” and “Nanmuiting,” highlighting the value of tangible cultural heritage. It is defined as a "Historic Architecture and Heritage Conservation" topic.

(3) DZ District – DZ-Topic0: It focuses on terms like “Tianxin Pavilion,” “park,” and “scenery,” representing the integration of natural and cultural landscapes within a historical setting. This falls under the category of "Historic Landscape and Urban Character."

(4) XYBL District – XYBL-Topic0: It features keywords such as “Fuli lamp,” “eat drink play and have fun,” and “leisure,” illustrating scenes of everyday urban life. It is categorized as "Gastronomy, Leisure, and Living Experience."

(5) BGY District – BGY-Topic2: It includes terms like “Cheng Qian,” “exhibition hall,” and “historical events,” emphasizing the cultural narrative function of former residences and historical episodes. This topic is classified as "Historical Memory and Cultural Transmission."

Figure 4 shows the distribution of review documents across topics in each district. Different colors represent different topics, and each dot corresponds to a single review document. The degree of clustering reflects the semantic coherence within topics: dense distributions with clear boundaries indicate high semantic consistency, while dispersed and overlapping distributions suggest greater semantic diversity.

Specifically, In TP historic district, topics are tightly clustered with clear boundaries, especially Topic 2, indicating a high degree of semantic consistency within that topic. CZ historic district shows a more scattered distribution, with Topics 2 and 4 having the highest density, reflecting key areas of tourist interest. DZ historic district exhibits a prominent central clustering pattern, with Topic 4 emerging as the dominant theme. In XYBL historic district, topics are largely non-overlapping, suggesting distinct semantic boundaries between content areas. BGY historic district forms a dense cluster centered around Topic 4, indicating that this topic is frequently mentioned.

Through the hierarchical clustering for topics (Fig. 5) and the similarity matrix of topics (Fig. 6), the BERTopic model reveals structural relationships among the topics. In Fig. 5, the horizontal axis represents topic similarity—the smaller the value, the stronger the correlation. In Fig. 6, color intensity is positively correlated with the strength of association between topics. Based on the above analysis results, the topics from the five historic districts were categorized as follows:

The topics of TP historic district can be broadly divided into two categories. The first includes Topics 1, 2, and 3, which focus on cultural service facilities and historical heritage, reflecting visitors’ interest in art exhibitions, cultural understanding, and historical value. Topics 0 and 7 lean more toward cultural consumption, such as traditional snacks and souvenir products found in intangible cultural heritage museums, indicating a trend of integrating cultural experience with consumption. The second category comprises Topics 4, 5, 6, and 8, which mainly involve food, specialty drinks, and popular social media-friendly products, highlighting visitors’ everyday consumption and check-in behaviors in the district.

The topic structure of CZ historic district is relatively dispersed, without a clearly defined classification. However, related topics can be grouped based on their inter-topic associations. Topics 0, 1, and 5 reflect how visitors perceive the district through leisure activities such as dining, check-ins, and emotional expressions, representing a “lifestyle-oriented” experiential form of tourism. Topics 3 and 4 emphasize the district’s physical environment and commercial space, indicating visitors’ concerns with shopping, spatial ambiance, and consumer convenience. In contrast, Topic 2 focuses more specifically on cultural heritage and historical architecture, highlighting a sense of recognition and appreciation for the district’s historical value.

The topics of DZ historic district can be categorized into three types. Topics 1 and 3 focus on leisure and tourism experiences, such as landscape perception, ticketing convenience, and transportation accessibility, reflecting the district’s tourism reception function. Topics 2 and 4 center on historical architecture and on-site cultural activities, highlighting the district’s cultural characteristics. Topic 0 emphasizes the district’s visual appearance and spatial aesthetics, including elements such as city walls and parks, reflecting visitors’ aesthetic perceptions.

The topics of XYBL historic district can be divided into three categories. Topics 2 and 3 focus on historical figures, cultural heritage sites, and educational spaces—such as former residences and memorial venues—highlighting the district’s function in historical education. Topic 0 emphasizes everyday leisure activities, such as culinary experiences and urban strolling, reflecting a lifestyle-oriented tourism model. Topic 1 concerns the district’s architecture, color schemes, and atmosphere, underscoring visitors’ aesthetic perception of urban space and emotional connections to it.

The topics of BGY historic district can also be categorized into three types. Topics 0 and 2 emphasize historical figures, revolutionary education, and historical architecture, showcasing the district’s cultural appeal. Topics 1 and 4 reflect visitors’ attention to visual ambiance, architectural aesthetics, and emotional experiences, highlighting the aesthetic enjoyment of the spatial environment. Topic 3 focuses on food consumption and social interaction, falling under the category of leisure tourism experiences.

Based on the results of topic extraction, this study matched the 29 identified topics with CES categories to construct a CES classification system for historic districts. Among existing studies, the classification framework proposed by the Millennium Ecosystem Assessment (MA) is considered the most widely adopted and authoritative. Drawing on the MA and other relevant literature, this study incorporated the outputs of the BERTopic model and consulted domain experts to conduct a preliminary categorization of CES in historic districts. The resulting classification system is presented in Table 2. Variations in CES types were observed across different historic districts: DZ district exhibits a relatively balanced distribution of CES types; TP and CZ districts are dominated by recreation tourism CES; while CES in XYBL and BGY districts are primarily concentrated in cultural heritage and aesthetic enjoyment services.

From a total of 18,794 visitor reviews, 17,073 valid opinion entries were obtained after data cleaning. Figure 7 presents the proportion of opinions related to CES sentiment analysis across the five historic districts. The results show that visitors are most interested in reviewing recreation tourism in historic districts, with nearly half of the reviews (N = 8,693) falling into this category, which also received the highest number of negative opinions. Reviews related to cultural heritage CES ranked second in volume (N = 6,866), but this category received the highest visitor satisfaction. In contrast, aesthetic enjoyment CES attracted the least attention, with only 1,514 related reviews.

The sentiment analysis results for each historic district are presented in Table 3. TP and CZ historic districts show the highest overall visitor satisfaction, with positive sentiment accounting for 88–93% of reviews related to cultural heritage and recreation tourism CES. The sentiment score for cultural heritage CES in CZ district reaches as high as 0.82. In contrast, DZ historic district performs the poorest, particularly in recreation tourism CES, where negative reviews account for 49% and the sentiment score is only 0.5. XYBL and BGY districts perform well, with positive sentiment exceeding 90% overall; only BGY’s recreation tourism CES has a notable 11% share of negative reviews. Overall, while recreation tourism CES receives the largest number of visitor reviews, it also attracts the highest proportion of negative feedback, indicating a need for service quality improvement. In comparison, although cultural heritage CES receives fewer reviews, it achieves the highest visitor satisfaction. Notably, DZ historic district records the lowest overall satisfaction among the five cases, suggesting a need for systematic improvement across all CES types.

This study employed the natural breaks method to classify the c-TF-IDF values of CES point data into five levels (Fig. 8). In the figure, the color of each point indicates the CES category, while the size of the point represents the c-TF-IDF weight of the corresponding category in the BERTopic topic modeling, reflecting the importance of CES in that spatial area. Based on the spatial distribution of CES across the five historic districts, priority areas for improvement within each district can be identified.

In TP historic district, CES exhibit a fishbone-shaped clustering pattern, primarily distributed along the main street of Taiping Old Street and its adjoining secondary roads. Cultural heritage CES (purple points) are located along the main Taiping Old Street and the secondary Fujia Lane, with hotspots concentrated at the northern and southern entrances of Taiping Old Street. Recreation tourism CES (green points) are mainly distributed along Taiping Old Street and Xipailou Old Street, with a hotspot at their intersection, reflecting a typical spatial pattern of cultural and recreational integration.

In CZ historic district, CES are primarily distributed in the area south of Chaozong Street. Cultural heritage CES are concentrated in the core area formed by Liansheng Street, Chaozong Street, Fuqing Street, and Yongqing Alley, but the point values are relatively low, indicating weak perception of historical and cultural elements. Recreation tourism CES are concentrated on the eastern side of Yongqing Alley and along Fuqing Street, with a hotspot at the intersection of Chaozong Street and Yongqing Alley.

In DZ historic district, cultural heritage CES are mostly concentrated in the Tianxin Pavilion scenic area and the historic building cluster along the main Duzheng Street. Aesthetic enjoyment CES include several high-value points, indicating that visitors place significant importance on the environmental ambiance of the district. In contrast, recreation tourism CES are fewer and have lower point values, suggesting that visitor activity is primarily focused on cognitive and observational experiences, with limited lifestyle-related consumption.

XYBL historic district is primarily characterized by cultural heritage CES, though their distribution is relatively dispersed. Recreation tourism and aesthetic enjoyment CES show a high degree of spatial overlap, but with generally low point values. Overall, CES in XYBL district exhibit the lowest point values among the five sample districts, indicating a relatively low level of visitor attention.

In BGY historic district, cultural heritage CES dominate and are mainly concentrated in areas where internal streets connect with major urban roads. Purple points are densely distributed along Baiguoyuan Alley, with generally high c-TF-IDF values. Aesthetic enjoyment CES are also present but do not form significant hotspots. Recreation tourism CES points are relatively few, suggesting that the district’s functions remain centered on historical education and cultural visitation.

The theoretical framework of AIPA is derived from the Kano model, which illustrates how different product or service attributes influence customer satisfaction (Xu et al., 2009). Figure 9 demonstrates the specific application of this model in the analysis of CES in historic districts. The horizontal axis represents CES performance (ranging from absence to satisfaction), while the vertical axis indicates the range of visitor satisfaction (from highly dissatisfied to highly satisfied). The blue curve represents basic factors: its solid line portion shows that even when well-performed, the increase in satisfaction is limited, whereas the dashed line indicates that their absence leads to significant dissatisfaction. The orange curve denotes linear factors, reflecting a proportional relationship between performance and satisfaction, with the solid and dashed lines having similar slopes. The green curve stands for excitement factors, with the solid portion forming a steep upward quadratic curve, indicating that high-quality performance greatly enhances satisfaction. Meanwhile, the dashed portion is nearly flat, showing that absence of these factors causes little dissatisfaction.

Based on the AIPA results from Equations (5) to (8) (Table 4), three types of CES were identified in terms of their mechanisms of influence on visitor satisfaction. As shown in the AIPA scatter plot (Fig. 10), the horizontal axis represents performance level, while the vertical axis represents the asymmetry impact (IA). Although the performance values of the three CES types are relatively close (ranging from 0.76 to 0.78), their impacts on satisfaction differ significantly. According to the IA values, cultural heritage falls within the basic factor zone with an IA value of approximately − 0.2, reflecting visitors’ fundamental expectations. Recreation tourism lies in the linear factor zone with an IA value around 0.06, indicating a positive correlation between satisfaction and CES performance. Aesthetic enjoyment is located in the excitement factor zone with an IA value of about 0.31, acting as a surprise element that significantly enhances satisfaction.

To gain a more comprehensive understanding of CES performance across different historic districts, a comparative analysis was conducted using a radar chart (Fig. 11). The chart shows that TP historic district excels in cultural heritage (approximately 0.79) but lacks services related to aesthetic enjoyment. CZ historic district performs strongly in recreation tourism (around 0.83), with moderate performance in cultural heritage (about 0.77), and also lacks aesthetic enjoyment services. BGY historic district demonstrates the most balanced performance, with aesthetic enjoyment particularly prominent (about 0.86) and relatively high performance in cultural heritage (around 0.82). XYBL historic district shows a well-rounded development across all three dimensions, with values ranging between 0.79 and 0.82. DZ historic district performs the weakest overall, with all three dimensions scoring between 0.62 and 0.71.

In summary, AIPA reveals the prioritization for optimizing CES management in historic districts. From an overall perspective, it is essential to first ensure the basic protection and interpretation of cultural heritage, while also improving recreation tourism services and focusing on the development of aesthetic enjoyment elements to create delightful experiences. At the individual district level, TP and CZ historic districts should focus on enhancing aesthetic CES; DZ historic district requires improvement across all dimensions; and XYBL and BGY historic districts show balanced development and should prioritize CES enhancement based on their specific characteristics.

By addressing RQ1 and RQ3, this study developed a CES classification system tailored to historic districts and analyzed the spatial distribution of the importance of different CES types. The findings reveal that CES in the five representative historic districts of Changsha consist of three categories: recreation tourism, which provides leisure services; cultural heritage, which represents the transformation of cultural resources; and aesthetic enjoyment, which delivers visual experiences. The spatial distribution of these three CES types varies across districts, indicating an imbalance in visitor perceptions of cultural services. Therefore, in optimizing CES management in historic districts, priority should be given to areas frequently mentioned in visitor reviews.

To address RQ2 and RQ4, this study analyzed visitor satisfaction with different types of CES in historic districts and revealed the underlying satisfaction mechanisms for the three CES categories. Cultural heritage functions as a basic factor, recreation tourism as a linear factor, and aesthetic enjoyment as an excitement factor. Unlike in most urban contexts where leisure and recreation are typically considered basic CES factors (Fischer et al., 2018), cultural services in historic districts should be grounded in cultural content with historical, artistic, or educational value. In other words, the interpretation of historical culture must be ensured first, followed by the enhancement of recreation tourism services, and ultimately, the quality of the overall cultural experience should be elevated through aesthetic enjoyment, thereby maximizing visitor satisfaction.

Specifically, TP historic district demonstrates strong performance in the transformation of cultural resources, with cultural landmarks such as the Former Residence of Jia Yi, Qianxing Art Museum, and Taiping Granary along Taiping Old Street being the most representative. These sites receive high visitor perception and satisfaction. As a national 4A-level tourist attraction, its recreation tourism services perform well by integrating local food brands with regional culture and urban identity, effectively enhancing the visitor experience. However, to prevent excessive commercialization and homogenization of cultural content, local tourism authorities should strengthen market regulation and evaluation. Moreover, TP district shows a deficiency in aesthetic enjoyment CES. Attention should be directed to CES clusters along Taiping Old Street and Xipailou Street, where cultural resources can be leveraged to create visually appealing spaces—for example, by renovating historic façades and upgrading cultural plaza landscapes.

Similar to TP, CZ historic district also requires enhancement of its aesthetic value. The basic elements of cultural heritage are poorly interpreted, while recreation tourism receives higher levels of attention and satisfaction, revealing a clear trend toward commercialization. This poses a risk to the integrity and authenticity of the district’s historical and cultural identity. It is recommended to prioritize the improvement of cultural displays in the southern section of Chaozong Street, for example by introducing AR technology to enable digital and interactive presentations of cultural heritage (Koo et al., 2019).

DZ historic district underperforms across all three CES types, with cultural heritage CES requiring urgent improvement—particularly in areas such as Tianxin Pavilion and the historic alley network centered on Duzheng Street. The vibrancy of cultural resources can be revitalized by organizing cultural events (Zhang & Han, 2022), such as the Hanfu Festival and light shows mentioned in reviews. In addition, although visitors show strong interest in aesthetic CES in this area, satisfaction remains low. Therefore, improving the aesthetic experience in DZ is also a key priority.

XYBL historic district shows balanced visitor satisfaction across all CES types. However, due to its small area and relatively recent development, it attracts relatively low levels of visitor attention. Efforts should focus on leveraging historical and cultural resources, especially those related to notable figures such as Zuo Zongtang and Shuai Mengqi, to innovate cultural service offerings and create a distinctive "cultural brand" for the district, thereby enhancing visitor recognition.

BGY historic district performs well in both cultural heritage and aesthetic enjoyment CES, but draws relatively low attention in recreation tourism. It is recommended to focus development on CES-dense areas such as Baiguoyuan Alley, Sujia Alley, Fengying Xili, and Denglou Street, expanding commercial activities to further enrich the overall visitor experience.

The theoretical contributions of this study lie in three key aspects: research subject, data source, and analytical framework.

In terms of research subject, although existing studies have developed relatively mature evaluation systems for CES (Cheng et al., 2019) there remains a significant gap in research on CES at the specific spatial scale of urban historic districts. This study approaches the renewal of historic districts from a cultural-ecological perspective. It not only enhances the understanding of cultural service functions within historic districts, but also employs quantitative methods to reveal public ecological perceptions and awareness of heritage conservation, thereby extending the current knowledge boundaries of historic district research.

In terms of data sources, current CES studies using social media data are mostly limited to single-platform image data analysis(Zhang et al., 2022), with insufficient exploration of the emotional dimension of visitor experiences. This study employs online review texts as longitudinal data and applies large language model techniques to extract visitors’ cognitive and emotional responses to cultural services. It offers a novel approach for utilizing unstructured text data in urban cultural ecosystem research.

In terms of analytical framework, although some existing studies have adopted the “topic extraction–sentiment analysis–IPA” approach (Jiang et al., 2025; J. Wang et al., 2025), the models and methods they employ often struggle to handle the semantic complexity and diverse expressions found in social media reviews (Gokcimen & Das, 2024). To address this gap, this study proposes an integrated framework combining qualitative and quantitative analyses. In the topic identification stage, the BERTopic model—known for its strong semantic understanding and automatic clustering capabilities—is employed for topic extraction. Subsequently, the natural breaks method is applied to classify model outputs and identify spatial clustering patterns. In the sentiment analysis stage, this study utilizes large language models with higher granularity compared to traditional lexicon-based approaches, enabling deeper sentiment detection from review data. Finally, AIPA is employed to reveal how different CES types influence visitor satisfaction, thereby uncovering the underlying mechanisms of service evaluation, attribute prioritization, and detailed improvement strategies (Hu et al., 2020). This enhances both the explanatory power of the analysis and its practical relevance.

With the development of urban heritage tourism in China, the renewal of historic districts has shifted from residential improvement to a new phase focused on enhancing cultural competitiveness (Lu et al., 2023; Zeng & Shen, 2020; Zhang et al., 2019). In this process, public participation is increasingly recognized as a critical component for the sustainable development of historic districts (Kou et al., 2018).

This study introduces online visitor reviews as a lens for observing public participation. By reducing the cost of traditional surveys and interviews, it offers in-depth insights into visitors’ cognitive and emotional responses to cultural services in historic districts, revealing value differences in public perception. This not only provides direct evidence for precise and differentiated service management in historic districts but also establishes a priority framework for cultural ecosystem restoration and resource allocation. CES types with high levels of negative sentiment in visitor reviews (e.g., recreation tourism in DZ) should be prioritized for targeted improvement, while high-performing service models (e.g., cultural heritage in TP) can serve as best practices to be replicated and promoted.

Moreover, the findings of this study offer generalizable references for CES planning and management across historic districts in China, thereby strengthening public cultural identity, promoting sustainable district renewal, and enhancing public well-being.

This study uses social media reviews as its data source to classify CES in historic districts through topic extraction. By integrating sentiment analysis, spatial clustering, and the AIPA method, it proposes strategic recommendations for optimizing cultural services in historic districts.

Using five representative historic districts in Changsha as case studies, this research applies BERTopic to cluster review texts and conduct word frequency analysis, identifying three core types of CES: cultural heritage, recreation tourism, and aesthetic enjoyment. Among them, recreation tourism contains the highest number of topics, indicating that visitors are most sensitive to and vocal about tourism services within the overall CES framework. Furthermore, by performing spatial clustering on high-frequency CES points, the study identifies CES-concentrated areas and priority street spaces for development within each district. Based on this, sentiment analysis and AIPA results show that visitor satisfaction is highest for cultural heritage CES, while recreation tourism CES receives the most negative reviews. In terms of satisfaction mechanisms, cultural heritage is classified as a basic factor, recreation tourism as a linear factor, and aesthetic enjoyment as an excitement factor. These findings offer a clear pathway for service optimization by cultural service managers.

Despite the above findings, this study has three main limitations that warrant further improvement in future research. First, the review texts used in this study have limited representativeness, as social media users tend to skew younger. Future research could supplement these data with offline surveys to capture the needs of older visitors, thereby constructing a more comprehensive visitor profile and identifying perception differences among demographic groups. Second, in terms of methodology, future studies could incorporate multimodal data analysis—such as images and videos taken by visitors—and apply image recognition and geotagging technologies to capture specific cultural activities and spatial trajectories. This would enable more fine-grained spatial and behavioral analysis. Finally, this study is primarily based on static analysis and does not capture temporal variations in CES content and sentiment. Future research could conduct time-series-based dynamic analysis to track the evolving trends of CES in historic districts.