Identification of Obstacles and Optimization Pathways for Sustainable Tourism in Southern Xinjiang: A Deep Learning Approach Based on GRU Sentiment Analysis

Abstract

With the rapid expansion of the tourism industry in Xinjiang, which received a record 328 million tourists in 2025, identifying development bottlenecks is crucial for regional sustainability. This study aims to identify the core obstacles hindering sustainable tourism in Southern Xinjiang—the region’s fastest-growing sector—and proposes evidence-based optimization pathways. Utilizing a deep learning approach, we deployed a Gated Recurrent Unit (GRU) sentiment analysis model to parse 5800 online reviews from 38 representative A-level scenic spots. The analysis identified 28 distinct obstacle clusters across three categories: landscape, cultural, and comprehensive destinations. The results reveal significant site-specific differentiation: natural landscape sites like Bayanbulak are primarily constrained by environmental risks and safety hazards, while high-traffic cultural sites like the Ancient City of Kashgar face acute challenges from over-commercialization and cultural erosion. Based on these findings, this study introduces a macro-level diagnostic tool and proposes targeted optimization strategies within the ESG (Environmental, Social, and Governance) framework. These insights offer actionable references for policymakers to enhance tourism resilience and achieve high-quality sustainable development in sensitive frontier regions.

1. Introduction

1.1. Research Background and Significance

Fostering sustainable tourism within ecologically vulnerable and culturally heterogeneous border regions has emerged as a critical global imperative. This objective fundamentally aligns with the United Nations Sustainable Development Goals (SDGs), particularly in its capacity to advance inclusive economic growth (SDG 8), cultivate resilient communities (SDG 11), and ensure responsible consumption and production patterns (SDG 12) [1,2]. According to the official data from the Culture and Tourism Department of Xinjiang Uygur Autonomous Region, Xinjiang’s tourism industry reached a historic high in 2025, receiving 328 million domestic and international tourists (a 21.05% year-on-year increase) and generating a total revenue of 372.58 billion CNY (a 24.11% increase). Notably, Southern Xinjiang has emerged as the fastest-growing sector, accounting for 43.3% of the total tourist volume with 142 million visitors in 2025. Characterized by unique biophysical landscapes and profound ethno-cultural legacies, the southern region of Xinjiang, China, represents a tourism destination situated at the nexus of developmental opportunities and socio-ecological challenges. Although the tourism sector acts as a vital catalyst for regional economic growth and ethno-cultural cohesion, it concurrently exerts substantial pressure on local ecological carrying capacities and the authenticity of indigenous cultural heritage. It is crucial to identify and address practical obstacles to sustainable development, such as ecological degradation, cultural commercialization, and environmental, social and governance shortcomings [3]—is essential for the region’s long-term resilience.

Conventional assessments of tourism sustainability have predominantly relied on static questionnaires and macroscopic statistics—methodological approaches inherently constrained by limited sample sizes, recall bias, and a lack of real-time spatio-temporal granularity. In the era of digital transformation, however, the ubiquitous proliferation of social media data presents a novel methodological paradigm to address these empirical limitations. User-Generated Content (UGC), such as independent travel evaluations published on platforms such as Xiaohongshu, is like a dynamic “social sensor”, which truly reflects the interaction between tourist activities and the destination environment [4]. Emotional analysis is an important branch in the field of Natural Language Processing (NLP), which can automatically extract and quantify these subjective views.

However, extant literature frequently overlooks the multifaceted empirical challenges inherent to the sustainable development of ethno-cultural borderlands. To address this lacuna, this study leverages advanced deep learning architectures—specifically Recurrent Neural Network (RNNs), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRUs)—to construct a high-fidelity diagnostic framework for sustainable tourism. By transforming unstructured tourist emotions into actionable governance insights, this study provides a data-based reference for high-quality sustainable tourism in southern Xinjiang and similar border areas.

1.2. Research Questions and Objectives

To systematically diagnose the developmental bottlenecks confronting sustainable tourism in southern Xinjiang, this study formulates three central research inquiries:

(1) Evaluation of Precision Diagnostic Tools: In processing short-text data for tourism sustainability monitoring, how do deep learning architectures (specifically RNNs, LSTMs, and GRUs) comparatively perform regarding predictive accuracy and computational efficacy?

(2) Assessment of Sustainability Status: Predicated upon optimal sentiment classification, what is the contemporary perceptual state of sustainability across A-level scenic areas in southern Xinjiang? Furthermore, how do sustainability constraints manifest differently across natural, cultural, and composite landscape typologies?

(3) ESG Strategy Formulation: How can the aggregation of negative public sentiment be operationalized into a systematic ESG optimization framework to enhance regional resilience?

The specific research objectives are as follows:

To address these challenges and fill the existing research gaps, this study formulates the following three core research questions:

(1) What are the primary clusters of tourism development obstacles in Southern Xinjiang across different destination categories (Landscape, Cultural, and Comprehensive)?

(2)How do these obstacle clusters manifest in specific representative sites, and what are the site-specific particularities?

(3) How can these data-driven findings be translated into actionable optimization pathways within the ESG framework?

By answering these questions, this study aims to provide a macro-level diagnostic tool for regional tourism authorities and offer localized management strategies to enhance the sustainable resilience of Southern Xinjiang’s tourism industry.

1.3. Research Innovations

The primary contributions of this study are articulated across four distinct dimensions:

(1) Methodological Breakthrough: This study pioneers a systematic, multi-model deep learning benchmarking framework specifically tailored for diagnosing tourism sustainability within China’s border regions. Crucially, it identifies the GRU architecture as achieving the optimal equilibrium between computational efficacy and predictive accuracy for regional governance applications.

(2) Spatial and Regional Innovation: By centering on southern Xinjiang, this research addresses a critical empirical lacuna concerning the sustainable development of ethno-cultural borderlands, offering a distinctive spatial case study on reconciling cultural preservation with natural resource valorization.

(3) Technical Integration: We construct a reproducible technical framework that integrates automated data scraping, deep learning sentiment parsing, and unsupervised root-cause clustering, providing a complete toolset for intelligent destination monitoring.

(4) Practical and Policy Applicability: This research abandons the pure theoretical framework and is committed to directly transforming insights on natural language processing into specific environmental, social and governance measures, and providing customized solutions to the unique development challenges faced by tourism in sensitive border areas.

2. Literature Review

The primary objective of textual sentiment analysis is to systematically identify and interpret the emotions, evaluations, perspectives, and attitudes articulated by individuals toward specific entities or phenomena. This methodological approach employs advanced computational techniques to extract and quantify the underlying subjective information embedded within textual data. Propelled by the exponential proliferation of digital platforms and e-commerce in recent years, sentiment analysis has emerged as a preeminent research frontier within the domain of NLP. Contemporary analytical paradigms in this field can be broadly classified into the following principal categories.

2.1. Sentiment Analysis Based on Lexicons and Rules

Early methodologies in sentiment analysis relied predominantly on lexical dictionaries and rule-based systems. The fundamental logic of these approaches involves determining the affective orientation of a text by cross-referencing predefined sentiment lexicons or adhering to specific heuristic rules. These methods quantify the frequency of affective tokens and employ established computational algorithms to evaluate the emotional polarity of individual sentences, thereby facilitating the classification of the overall textual sentiment. Within this framework, the sentiment lexicon serves as a critical repository, providing comprehensive annotations of the emotional polarity and intensity of discrete terms. Consequently, the granularity and comprehensiveness of these lexicons constitute the primary determinants of predictive accuracy in sentiment analysis tasks. Zhang et al. [5] gave a comprehensive review of the application of deep learning in emotional analysis, explained the underlying technical framework and practical application scenarios of various deep learning architectures, and summarized the latest progress and emerging trends in this field. Tan et al. [6] introduced an aspect embedding alignment technique aimed at enhancing aspect-based sentiment analysis by improving the correspondence between aspect-specific features and associated sentiment orientations. Ma et al. [7] extracted keywords from text data through emotional dictionaries. However, the emotional semantics expressed by emotional words may change with the context. If the general emotional dictionary is directly applied to different types of text corpus, it is easy to cause deviations in the analysis results. Yang et al. [8] introduced a domain-specific sentiment classification model that identifies and excludes keywords unrelated to the target domain, thereby facilitating the development of more precise domain-specific sentiment lexicons. Hu et al. [9] proposed a fine-grained sentiment analysis approach tailored to automotive complaint texts, aimed at constructing an intelligent early warning system for product-related crises. This method effectively captures sentiment orientations and key issues within complaints, offering robust technical support for proactive product crisis management. The emotional tendency of words will be affected by a variety of context factors, which may lead to insufficient analysis results. Therefore, in the whole process of emotional analysis, the method of constructing emotional dictionaries must be constantly improved to improve the analysis effect.

Rule-based analysis methods will use pre-set criteria to help determine emotional tendencies. Luo et al. [10] emphasized that in the process of emotional classification of informal texts, combining text preprocessing technology with the application of semantic rules can significantly improve the accuracy of classification. In rule-based emotional classification, the core strategy to improve accuracy is to formulate rules that fit specific text characteristics, but external factors may also affect the emotional tendency assessment results output by the method. Jia et al. [11] combined semantic rules with emotional dictionaries to make full use of the advantages of both to improve the classification effect. Li et al. [12] built a perception dictionary covering multiple components of urban green space, emotionally analyzed the comment text of Dianping with the help of the Snow NLP library, and integrated geographic information data to build a structural equation model. The results show that the landscape element of the green space had a positive impact on the emotional response of residents. Ainapure et al. [13] used two dictionary tools, VADER and NRCLex, to conduct emotional analysis of Twitter data related to COVID-19 and vaccination in India. By matching dictionary words to evaluate the polarity of the text and calculate the emotional score, the study found that vaccine-related tweets mainly express positive emotions, which provides valuable emotional references for the decision-making process. Kauffmann et al. [14] employed the Afinn sentiment lexicon in conjunction with natural language processing and text mining methodologies to extract features and analyze sentiment in Amazon mobile phone reviews. By incorporating variables such as price and star ratings, they developed an integrated product scoring system that effectively distinguished positive and negative product attributes and facilitated product ranking, thus aiding marketing strategies. Therefore, the emotional analysis method based on vocabulary and rules is easy to implement and highly accurate in specific fields, but it has limitations in dealing with the subtle differences and emotional diversity of complex texts unique to the contemporary multicultural background. Its effectiveness mainly depends on the accuracy of the emotional dictionary and the manual optimization effect of the preprocessing program.

2.2. Machine Learning-Based Sentiment Analysis

Within the domain of text emotional analysis, researchers often use machine learning technology and feature engineering to build classification models to determine the emotional tendency of text data. Jiang et al. [15] proposed a target-based language classification method, which conducts targeted speech analysis of Twitter data to accurately identify the speech expressed for specific entities in the text. Guo et al. [16] leveraged Support Vector Machines (SVM) in conjunction with lexical resources to determine the sentiment orientation of textual features. Sharma et al. [17] evaluated the effect of a variety of machine learning algorithms such as simple Bayes (NB), decision tree (DT), K-neighbor (KNN) and so on in emotional classification. Experimental results show that feature selection is crucial to improve model performance, among which the support vector machine (SVM) performs best. Boiy et al. [18] proposed a cascading mechanism that was successfully integrated into their model, leading to improved accuracy in text feature extraction. Luo et al. [19] used 294,034 restaurant reviews on the Yelp platform as research data, and adopted the potential Dirichlet distribution (LDA) model to extract the four core dimensions of food taste, consumption experience, geographical location and cost-effectiveness and complete emotional labeling. By comparing the three machine learning methods of Naive Bayes, Naive Bayes and SVM fusion, and SVM combined with Fuzzy Domain Ontology (FDO), the results show that the algorithm combining support vector machine and fuzzy domain ontology has the best performance in commenting usefulness prediction tasks, with an F1 value of 79.59 percent. Ding et al. [20] surveyed 623 tourists from Xinjiang. Based on questionnaire data, cluster analysis, single-factor variance analysis and factor analysis were used to explore the multi-dimensional impact of perceived travel safety on the image of the destination. Research results show that tourists with a higher level of safety perception have a more positive evaluation of the cognitive, emotional and behavioral dimensions of the destination image, and the perceived safety also has a positive impact on stereotypes. Alzahrani et al. [21] examined the Al-Baha Agricultural Festival in Saudi Arabia by analyzing Arabic Twitter data through six machine learning algorithms, including multinomial Naive Bayes, support vector machines, and random forests. The results demonstrate that support vector machines and random forests achieved superior performance in sentiment classification, whereas multinomial Naive Bayes and K-nearest neighbors models showed comparatively poor outcomes. Although machine learning methods have made remarkable progress in emotional recognition, their effectiveness still depends on the quality of feature extraction. High-quality feature extraction is crucial for the model to effectively capture key text information. In addition, those machine learning models that integrate linguistic principles do help to extract semantic content, but this integration often brings too many characteristic dimensions, which greatly reduces the computing efficiency of the model.

2.3. Deep Learning-Based Sentiment Analysis

Within the domain of textual sentiment analysis, the fidelity of feature extraction exerts a profound influence on overall classification performance. Short-text data, in particular, presents formidable challenges owing to its inherent data sparsity and the subsequent difficulty in articulating robust feature representations. With the emergence of deep learning methods, Zhao et al. [22] systematically investigated text feature extraction methods and classification algorithms based on traditional machine learning, and proposed a sentiment analysis model for product reviews combining expanded sentiment dictionaries. This approach not only simplifies the feature engineering pipeline but also significantly improves classification accuracy. In practical applications of deep learning, Zaremba et al. [23] introduced the Recurrent Neural Network (RNN), which demonstrates strong capabilities in capturing contextual semantic information within sequential data. Subsequently, Kim [24] proposed the Convolutional Neural Network (CNN), which is effective in extracting local textual features; however, CNNs exhibit limitations in modeling long-range semantic dependencies. To overcome these constraints, Wang et al. [25] developed the LSTM network, which is better suited for learning long-distance dependencies in text. Building upon this, Tang et al. [26] presented the TD-LSTM model, incorporating a dual-LSTM architecture designed to more accurately capture the contextual relationships surrounding target words. Furthermore, Shang et al. [27] utilized the Bidirectional Long Short-Term Memory (BiLSTM) network to extract contextual sentiment features bidirectionally, thereby further improving the precision of sentiment recognition.

Yang et al. [28] introduced the attention mechanism in the deep learning network architecture to enhance the ability to extract significant information from text data, thus improving the performance of the whole model. Lin et al. [29] developed a structured self-attention mechanism that utilizes the output of a LSTM network as input, generating sentence vector representations through a weighted summation process. This approach effectively augments the model’s capacity to extract and interpret textual features. In a seminal contribution, Vaswani et al. [30] introduced the Transformer architecture, an attention-based model comprising an encoder and a decoder, each constructed from multiple identical layers. Each layer integrates a self-attention module alongside a feedforward neural network, enabling the Transformer to capture long-range dependencies within sequential data efficiently and thereby substantially enhancing the accuracy of sentiment analysis tasks. Furthermore, Li et al. [31] proposed the BiGRULA model, which combines the lda2vec topic model with an attention mechanism for sentiment classification in tourism reviews. By enriching word embeddings with topic information to capture global semantic context and applying attention to assign differential weights to words, this model achieved a classification accuracy of 93.1% on a hotel review dataset, surpassing the performance of convolutional neural networks (CNN), LSTM, and other comparative models. Peng et al. [32] integrated the Latent Dirichlet Allocation (LDA) model with a sentiment analysis approach grounded in a customized dictionary and rule-based system to examine UGC from ski resorts located in the host city of the 2022 Winter Olympics. Their analysis identified nine distinct attributes related to the destination’s image and demonstrated a consistent decline in tourists’ negative emotions alongside a gradual increase in positive sentiments, suggesting a favorable influence of the Winter Olympics on the destination’s image. Lin et al. [33] introduced a short-text representation framework utilizing BERT embeddings in conjunction with a BiGRU-Attention sentiment analysis model to evaluate brief comments posted on Weibo. Their findings indicated that BERT-derived word vectors surpassed Word2Vec in performance, and the BiGRU-Attention model achieved an accuracy rate of 98.32%, exceeding that of CNN, BiLSTM, and BiGRU models, thereby effectively mitigating challenges related to feature sparsity and semantic information loss in short textual data. Wen et al. [34] proposed a hybrid neural network model, which integrates BERT, BiLSTM, convolutional neural network (CNN) and attention mechanism, and is applied to the emotional classification and content analysis of Sanya travel reviews. The results show that the accuracy rate of the model is 93.13%, which is better than other comparative models, and shows that the negative emotions of tourists are mainly concentrated in tourism infrastructure and supporting services.

2.4. Theoretical Evolution of Sustainable Tourism Evaluation and UGC Application

The evaluation of sustainable tourism has transitioned from traditional macro-indicator frameworks to micro-level, data-driven diagnostic approaches. Traditionally, theoretical frameworks such as the Triple Bottom Line (TBL) and Tourism Carrying Capacity (TCC) have provided the foundation for assessing sustainability through economic, social, and environmental dimensions [35]. However, these traditional systems often rely on static government statistics or expert-led scoring, which frequently fail to capture the real-time, subjective “bottlenecks” experienced by tourists on-site [36].

With the advent of the digital era, UGC has emerged as a critical “digital footprint” for monitoring destination health. Recent research highlights that sentiment analysis of online reviews offers a high-frequency, large-scale perspective that traditional surveys cannot achieve, revealing nuanced obstacles like cultural erosion or specific infrastructure failures [37,38]. To decode these complex semantic structures, deep learning models—particularly GRU—have demonstrated superior performance over lexicon-based methods in capturing long-term dependencies in tourism texts [39]. By integrating deep learning with sustainability frameworks, researchers can transform subjective feedback into objective “obstacle clusters,” providing a granular diagnostic basis for ESG optimization [40]. This methodological evolution marks a shift from generalized monitoring to site-specific, resilient destination management.

2.5. Research Review

A comprehensive review of the extant literature reveals that sentiment analysis within the tourism sector has evolved from lexicon-based and conventional machine learning paradigms toward contemporary deep learning architectures. Within this domain, RNN architectures—most notably LSTM and GRUs—have gained widespread adoption due to their robust capacities for sequence modeling, with the GRU model exhibiting superior computational efficacy. Concurrently, while scholarly interest in tourism research within ethno-cultural borderlands continues to intensify, there remains a distinct scarcity of empirical investigations that leverage advanced deep learning methodologies to perform granular sentiment parsing in geographically specific contexts such as southern Xinjiang.

This study synthesizes automated web-crawling technologies with a multi-model deep learning architecture—specifically integrating RNN, LSTM, and GRU networks—to address identified lacunae in the extant literature. Through the application of visual analytics, the research conducts a granular sentiment diagnostic of tourist commentaries across A-level attractions in southern Xinjiang. This methodological synergy not only fosters theoretical and procedural innovation within the domain of tourism sentiment analysis but also provides robust empirical evidence and actionable decision-making support for the strategic development of tourism in sensitive borderland regions.

3. Research Methods

3.1. Data Source and Collection

The empirical data for this study were exclusively sourced from Xiaohongshu (Little Red Book), China’s preeminent lifestyle-sharing social media platform. Unlike traditional Online Travel Agencies (OTAs) that focus on standardized ratings, Xiaohongshu is characterized by highly subjective, experience-oriented UGC, making it an ideal source for identifying nuanced “pain points” and sustainability obstacles.

Using the Bazhuayu (Octopus) web crawler, we conducted a systematic data retrieval focusing on 38 tourist attractions (rated 3A and above) in Southern Xinjiang. The temporal scope of the data spans from January 2020 to December 2025, covering the critical phases of post-pandemic tourism recovery and the subsequent surge in regional travel interest.

A total of 7580 raw review notes and comments were initially extracted. To ensure the robustness of the subsequent deep learning analysis, a rigorous data cleaning protocol was applied to the raw corpus, which included:

(1) Redundancy Removal: Eliminating duplicate posts and system-generated default content;

(2) Noise Filtering: Removing advertisements, promotional material, and non-substantive entries consisting solely of emojis or irrelevant characters;

(3) Relevance Screening: Filtering out content that did not contain substantive feedback regarding tourism services, infrastructure, or environmental experiences.

The 38 sampled scenic spots are selected with strict representativeness considerations: according to official statistics from the Culture and Tourism Department of Xinjiang Uygur Autonomous Region, there were 258 A-level scenic spots in Southern Xinjiang by the end of 2025, and the 38 sampled spots account for 14.73% of the total, with a grade structure highly consistent with the overall distribution of A-level scenic spots in the region—covering all 9 5A-level scenic spots (100% full coverage), 18 core 4A-level scenic spots with the highest annual tourist reception, and 11 representative 3A-level scenic spots with distinctive regional characteristics. The “top-ranked” criteria for sample selection are clearly defined by dual dimensions: the official annual tourist reception volume from 2024 to 2025, and the online attention measured by the number of relevant travel notes on Xiaohongshu. The reason for focusing on 3A-level and above scenic spots is that these destinations are the core carriers of Southern Xinjiang’s tourism industry, receiving more than 92% of the region’s total tourist volume, and their sustainable development status directly determines the overall resilience of the regional tourism industry.

Meanwhile, it should be noted that the data in this study are exclusively sourced from the Xiaohongshu platform, whose core user group is young women aged 18–35. This user portrait feature may bring systematic bias to the sample: the review content may not fully cover the tourism perception and evaluation of other tourist groups, such as middle-aged and elderly tourists, male tourists, and group tour participants. This systematic bias will limit the generalizability of the research conclusions to a certain extent, and the relevant findings should be interpreted within the application boundary of this data source.

Following the preprocessing phase, 5800 high-quality, valid samples were retained. This refined dataset provides a concentrated and high-fidelity semantic basis for the sentiment classification and obstacle identification performed in this study.

3.2. Model Training Logic

The overall research framework of this study is illustrated in Figure 1. To identify the most effective sentiment classifier for the Xiaohongshu dataset, a comparative experiment was conducted among three representative RNN architectures: Standard RNN, LSTM, and GRU. The modeling process followed a two-stage approach:

(1)

Stage 1: Comparative Pre-Training. All three models (RNN, LSTM, and GRU) were first trained on a large-scale, publicly available annotated dataset containing 119,000 tourism-related reviews. This stage enabled the models to learn complex semantic features and universal emotional patterns. Performance metrics, including Accuracy and F1-score, were used to evaluate their classification capabilities.

In the pre-training stage, we adopted a basic class balance strategy for the annotated dataset to avoid the model’s bias towards the majority class. In subsequent research, we will further introduce class weight assignment and SMOTE oversampling methods to optimize the model’s recognition ability for the minority negative class in the imbalanced dataset of Southern Xinjiang’s tourism reviews.

(2)

Stage 2: Optimal Model Application. Based on the comparative results, the GRU model was selected as the final classifier due to its superior computational efficiency and its ability to achieve high accuracy with fewer parameters compared with LSTM. The pre-trained GRU was then applied to the 5800 site-specific samples from Southern Xinjiang to extract sustainability “obstacles” (negative sentiments).

Figure 1. The research framework based on model comparison and two-stage GRU sentiment analysis.

Figure 1. The research framework based on model comparison and two-stage GRU sentiment analysis.

3.3. Sentiment Analysis Model Construction

To systematically process vast quantities of unstructured UGC and precisely capture the sentiment orientation of tourists regarding destination sustainability, this study leverages a deep learning-based NLP workflow. The primary objective entails a rigorous benchmarking of three prominent RNN architectures—specifically vanilla RNN, LSTM, and GRU—to identify the most robust model characterized by optimal computational efficacy for the analysis of short-form travel commentaries.

3.3.1. Text Preprocessing and Vectorization

For the linguistic characteristics of Chinese social media comments, strict data preprocessing is very necessary. First of all, the original text was cleaned up, and Chinese characters were retained through the Unicode logo, while irrelevant elements such as numbers, punctuation marks and emojis were deleted. Then, use the jieba word segmentation tool to divide the continuous character stream into meaningful word sequences, and filter the stop words with empty semantics according to the custom dictionary. For text vectorization, use the Tokenizer module to build a glossary according to the word frequency, and limit the size of the glossary to within 5000 to filter out rare and meaningless words. Each word is mapped to a unique integer index. In order to adapt to the diversity of text lengths in Xiaohongshu comments, all sequences are uniformly adjusted to a fixed length by filling and truncating, so as to convert the text into a unified numerical tensor suitable for neural network input.

3.3.2. Deep Learning Model Selection and Architecture

We constructed three comparative sequence-modeling architectures to determine the optimal sentiment classifier:

(1) RNN: Use loop connections to maintain a hidden state between different time steps, so as to capture the basic sequence context. However, the standard RNN often performs poorly when dealing with long-distance dependencies due to the problem of gradient disappearance.

(2) LSTM: By introducing the unit state regulated by three different gate mechanisms of forgetting gate, input gate and output gate, the problem of gradient disappearance is solved. This allows the model to selectively retain key contextual information in longer sequences, so that it excels in semantic understanding.

(3) GRU: An optimized variant of LSTM, which combines cell state and hidden state to simplify the architecture into two gate control mechanisms. GRU can usually achieve an accuracy rate equivalent to that of LSTM, but because it performs fewer tensor operations, higher calculation efficiency and faster convergence, it is especially suitable for short text classification.

In order to maximize the effect of semantic extraction, both LSTM and GRU adopt a two-way architecture, so that the model can capture the context dependence of the front and back paragraphs in the comment at the same time.

3.3.3. Model Training and Hyperparameter Configuration

These models are trained through an end-to-end framework. The input tensor first passes through an embedded layer to map the discrete index into a dense and trainable semantic vector (ranging from 64 to 128 dimensions). The configuration of the cycle layer (RNN, LSTM or GRU) has 32 to 128 hidden units. To prevent overfitting, a Dropout layer $rate=0.5$ was integrated. The final semantic representation was fed into a fully connected dense layer, culminating in a Softmax output layer that generated probability distributions for binary sentiment classification (positive vs. negative). The models were optimized using the Adam and RMSprop optimizers, minimizing the categorical cross-entropy loss function over multiple epochs. Validation sets were monitored dynamically to preserve the optimal model parameters. The evaluation metrics (Accuracy, Precision, Recall, and F1-Score) and convergence behaviors (loss/accuracy curves) were subsequently generated to assess and compare the classifiers.

3.4. Evaluation Metrics

To rigorously quantify the diagnostic performance of the candidate models, this study utilizes four primary evaluation metrics. Accuracy represents the quotient of correctly classified instances relative to the total population, providing a global heuristic for the model’s overall predictive fidelity. Precision measures the proportion of authentic positive occurrences among all instances predicted as such—a critical parameter for minimizing the risk of false positives, wherein negative commentaries are erroneously categorized as positive. Recall assesses the model’s capacity to accurately retrieve true positive instances, ensuring that favorable evaluations are captured with maximum comprehensiveness. The F1 score represents the harmony and average of the accuracy rate and the recall rate, which can provide a comprehensive performance measure, especially for category imbalance. The mathematical expressions of these indicators are as follows:

Accuracy = (TP + TN)/(TP + FP + TN + FN)

(1)

Precision = TP/(TP + FP)

(2)

Recall = TP/(TP + FN)

(3)

F1 = 2 × (Precision × Recall)/(Precision + Recall)

(4)

where TP represents true positives, FP false positives, TN true negatives, and FN false negatives.

4. Results Analysis (Model Comparison, Negative Review Keyword Cloud)

4.1. Performance Evaluation and Model Selection

In order to ensure the reliability of sustainability barrier identification, we compared and evaluated the performance of three deep learning architectures, RNN, LSTM and GRU, on independent test data sets. As shown in

Table 1. Performance comparison of three deep learning models in emotional classification.

	Accuracy	Precision	Recall	F1 Score
GRU	91.62%	91.62%	91.62%	0.9162
LSTM	90.37%	90.37%	90.37%	0.9037
RNN	89.06%	89.06%	89.06%	0.8906

, the GRU model shows the most stable prediction accuracy, with an overall accuracy of 91.62% and an F1 score of 0.9162.

The excellent performance of the GRU model can be attributed to the optimization of its gate control mechanism. Although the RNN will encounter the problem of gradient disappearance when processing sequence data, and the LSTM network increases the computational complexity due to its three-gate structure, the GRU achieves an ideal balance between feature extraction and training efficiency by combining the unit state and the hidden state. This is especially beneficial to the linguistic characteristics of Xiaohongshu travel reviews, which are usually short in length, with an average of about 28 characters and rich semantics.

To further verify that the GRU model is not biased towards the majority positive class in this imbalanced dataset (negative sentiment accounts for only 4.9%), we provide a detailed interpretation of the confusion matrix (Figure 2a). The confusion matrix clearly shows that the GRU model correctly identifies the vast majority of negative reviews, with no significant tendency to misclassify negative reviews as positive ones. This confirms that the model has reliable recognition ability for the minority negative class, which is the core focus of this study for identifying sustainability obstacles, ensuring the robustness of the subsequent clustering analysis.

Figure 2. Performance of the optimal GRU model: (a) Confusion matrix. (b) Training accuracy curve.

Figure 2. Performance of the optimal GRU model: (a) Confusion matrix. (b) Training accuracy curve.

4.2. Sentiment Analysis Results of Southern Xinjiang Scenic Areas

Using the best-performing GRU model, emotional predictions were made on data sets containing 5800 comments. The vertical axis indicates the proportion of negative evaluations, and the horizontal axis indicates the name of each scenic spot. There are a total of 38 scenic spots, numbered in the order of 1 to 38, corresponding to the following locations: 1. 139 Populus Euphratica Secret Road, 2. Bachu Red Sea, 3. Bayinbuluke, 4. Baisha Mountain Desert, 5. Bosten Lake Scenic Area, 6. Cele County Kunlun Sacred Land, 7. Dawa Kun Desert Tourism Area, 8. Duolang River, 9. Gongnaisi Scenic Area, 10. Hotan Uluwati Scenic Area, 11. Red Desert Scenic Area, 12. Why Are the Flowers So Red Scenic Area, 13. Imperial Palace Desert Lake Tourism Resort, 14. Huola Mountain Silk Road Ancient Village, 15. Jinshatan Scenic Area, 16. Kashgar Old City, 17. Keping Dawan Valley, 18. Kerguti Scenic Area, 19. Kezhou Glacier Park Scenic Area, 20. Kizil Red Stone Forest, 21. Long Lake Tourism Area, 22. Lop Nur Village Scenic Area, 23. Luopu County Aqike Qianshan River Valley Scenic Area, 24. Pakelrek Scenic Area, 25. Pamir Tourism Area, 26. Polong Primeval Forest Scenic Area, 27. Shayanzhou Scenic Area, 28. Shenmu Garden, 29. Taklamakan N39, 30. Swan River Scenic Area, 31. Tianshan Grand Canyon, 32. Tomur Grand Canyon, 33. Yanquan Mountain Scenic Area, 34. Yecheng County Xitiya Maze City Scenic Area, 35. Yecheng County Zonglang Spiritual Spring Scenic Area, 36. Yutian County Daliya Boyi Scenic Area, 37. Moon Bay, and 38. Zepu Populus Euphratica Scenic Area. As shown in

Table 2. Top 10 Tourist Attractions with the Highest Negative Review Rates.

Scenic Area Name	Negative Review Rate
Duolang River	11
Jinshatan Scenic Area	8
Gongnaisi Scenic Area	7.8
Cele County Kunlun Sacred Land	7.8
Tianshan Grand Canyon	7.4
Long Lake Tourism Area	7
139 Populus Euphratica Secret Road	6.8
Imperial Palace Desert Lake Tourism Resort	6.6
Pamir Tourism Area	6
Tomur Grand Canyon	5.9

, we list the top ten tourist attractions with the highest negative review rates and their corresponding negative evaluation rates.

Tourist attractions in southern Xinjiang have won widespread praise on the “Little Red Book” platform, and the overall satisfaction rate has reached 88.6%. However, 4.9% of the comments are still negative, which shows that there is still room for improvement in service quality. Further analysis shows that there are significant differences in the emotional distribution of tourists in different types of scenic spots. This study divides 38 scenic spots into three categories: natural landscape, cultural and comprehensive for in-depth analysis. The natural landscape comprises 18 sites, including Bayinbuluke Scenic Area, Bosten Lake Scenic Area, and Pakalake Scenic Area; the cultural category encompasses 8 sites, such as Kashgar Old Town, Huo La Mountain Silk Road Ancient Village, and Yecheng County Xitiya Maze City Scenic Area, while the comprehensive category consists of 12 sites, including Bachu Red Sea Scenic Area and the 139 Poplar Secret Road. Findings indicate that scenic attractions (e.g., Baisha Lake with a 94% approval rating, Bayinbuluke Scenic Area at 89%) receive significantly higher approval compared with cultural attractions (e.g., Kashgar Old Town at 80%, Huo La Mountain Silk Road Ancient Village at 78%). These results show that tourists highly appreciate the natural resources of southern Xinjiang, but the display of cultural resources, tourist experience and management still need to be improved.

4.3. Diagnosis of Sustainability Obstacles via K-Means Clustering

Based on the above 38 different scenic spots, we extracted negative evaluations through emotional analysis and conducted more in-depth research. According to the type of scenic spot, the comments are divided into three main categories, and the K-means cluster analysis is carried out independently to explore the potential factors that lead to negative evaluation. This research method aims to provide a reference basis for the formulation of more accurate management strategies and interventions in the future.

Before the clustering analysis, we clearly defined the boundary between two types of negative comments: genuine sustainability issues refer to the obstacles that affect the long-term sustainable development of scenic spots, including ecological environment damage, cultural authenticity erosion, tourism carrying capacity overload, and cross-regional governance defects, which are the core research focus of this study; service quality complaints refer to the temporary and operational problems in the tourism process, including catering and accommodation service defects, staff attitude problems, and non-transparent charging of individual projects. The subsequent clustering analysis and interpretation will strictly distinguish these two types of issues and focus on the diagnosis of core sustainable development obstacles.

In order to facilitate the understanding of international readers and be consistent with the published language of this journal, a translation program has been implemented in the data visualization stage. It must be clear that all core NLP tasks, including text segmentation, disabled word filtering, emotional classification and K-means clustering, are processed strictly according to the original Chinese UGC. This treatment strictly follows the principle of retaining the semantic nuances, cultural background and true emotional expression contained in Chinese as it is. At the same time, the generated word cloud visual diagram and the presentation of cluster keywords were translated into English. This method ensures that the visual presentation can be fully understood by the global audience without sacrificing the methodological rigor and analytical accuracy of the underlying data.

4.3.1. Natural Landscapes: Ecological Vulnerability and Operational Inefficiency

Using the K-means clustering algorithm, a root cause analysis was conducted on the 113 negative reviews provided by tourists regarding natural landscape attractions. The optimal number of clusters was determined to be K = 10 through the elbow method and silhouette coefficient (Figure 3), leading to the identification of ten principal problem clusters (Table 3). And a word cloud map (Figure 4) was drawn.

Figure 3. K-value line chart.

Figure 3. K-value line chart.

Figure 4. Overall word cloud of scenery category.

Figure 4. Overall word cloud of scenery category.

The table below lists the number of comments, the corresponding proportion, core keywords and typical characteristics contained in each cluster.

Table 3. Detailed Cluster Analysis of Scenery Category.

Cluster ID	Number of Reviews	Proportion	Core Keywords	Typical Review Examples	Representative Case Sites
0	4	3.54%	lake, actual, far, propaganda, swan, compared, landscape	The actual landscape is far from the propaganda and expectation.	Bayanbulak Scenic Area, Moon Bay
1	6	5.31%	service, safe, consumption, attitude, mobile, signal, staff, projects, good, commercialization	The mobile phone signal is very poor, and the attitude of the service staff is also very perfunctory.	Taklimakan n39, Moon Bay
2	12	10.62%	weather, extreme, toilets, scorpions, snakes, wild, animals	The photographing of some natural landscapes depends on the weather.	Bayanbulak Scenic Area, Taklimakan n39
3	18	15.93%	high, cost, charges, food, prices, performance, dining, tickets	The accommodation price in the town is high and the condition is average.	Bayanbulak Scenic Area, Longhu Tourist Area
4	10	8.85%	poor, environment, accommodation, sanitation, experience, conditions, swarm	The road conditions are poor. Accommodation is very expensive and the environment is poor.	Dawakun Desert Tourism Area, Polong Primitive Forest Scenic Area
5	12	10.62%	ticket, bus, core, moon, bay, shuttle, attractions, making, interval	Some tourists missed the core attractions because they didn’t check the ticket availability in advance.	Bayanbulak Scenic Area, Moon Bay
6	5	4.42%	construction, dust, storage, standards, repair	Some restaurants in the scenic spot have poor hygiene, with disorganized storage of ingredients.	Bayanbulak Scenic Area, Parklek Scenic Area
7	22	19.47%	desert, up, route, experience, overall, large, over, area	The time of the scenic transport is long. The scenery and experience items did not meet expectations.	Moon Bay, Taklimakan n39
8	9	7.96%	safety, hazard, potential, mountain, signs, rest, frequently, water	There are potential safety hazards in some projects. The fees of some water sports are not transparent.	Bayanbulak Scenic Area, Baisha Mountain Desert Scenic Area
9	15	13.27%	parking, capacity, attractions, traffic, road, peak, hot	Congestion often occurs during the peak season of the scenic spot. The environmental pollution is serious due to the littering of some tourists.	Bosten Lake Scenic Area, Moon Bay

Table 3. Detailed Cluster Analysis of Scenery Category.

Cluster ID	Number of Reviews	Proportion	Core Keywords	Typical Review Examples	Representative Case Sites
0	4	3.54%	lake, actual, far, propaganda, swan, compared, landscape	The actual landscape is far from the propaganda and expectation.	Bayanbulak Scenic Area, Moon Bay
1	6	5.31%	service, safe, consumption, attitude, mobile, signal, staff, projects, good, commercialization	The mobile phone signal is very poor, and the attitude of the service staff is also very perfunctory.	Taklimakan n39, Moon Bay
2	12	10.62%	weather, extreme, toilets, scorpions, snakes, wild, animals	The photographing of some natural landscapes depends on the weather.	Bayanbulak Scenic Area, Taklimakan n39
3	18	15.93%	high, cost, charges, food, prices, performance, dining, tickets	The accommodation price in the town is high and the condition is average.	Bayanbulak Scenic Area, Longhu Tourist Area
4	10	8.85%	poor, environment, accommodation, sanitation, experience, conditions, swarm	The road conditions are poor. Accommodation is very expensive and the environment is poor.	Dawakun Desert Tourism Area, Polong Primitive Forest Scenic Area
5	12	10.62%	ticket, bus, core, moon, bay, shuttle, attractions, making, interval	Some tourists missed the core attractions because they didn’t check the ticket availability in advance.	Bayanbulak Scenic Area, Moon Bay
6	5	4.42%	construction, dust, storage, standards, repair	Some restaurants in the scenic spot have poor hygiene, with disorganized storage of ingredients.	Bayanbulak Scenic Area, Parklek Scenic Area
7	22	19.47%	desert, up, route, experience, overall, large, over, area	The time of the scenic transport is long. The scenery and experience items did not meet expectations.	Moon Bay, Taklimakan n39
8	9	7.96%	safety, hazard, potential, mountain, signs, rest, frequently, water	There are potential safety hazards in some projects. The fees of some water sports are not transparent.	Bayanbulak Scenic Area, Baisha Mountain Desert Scenic Area
9	15	13.27%	parking, capacity, attractions, traffic, road, peak, hot	Congestion often occurs during the peak season of the scenic spot. The environmental pollution is serious due to the littering of some tourists.	Bosten Lake Scenic Area, Moon Bay

Evidence: As detailed in Table 3, the clustering analysis for natural landscape attractions reveals a “multi-core balanced distribution” of negative feedback. The core and stable conclusions of this study are derived from large-sample clusters with more than 10 reviews: Cluster 7 (19.47%), Cluster 3 (15.93%), Cluster 9 (13.27%), Cluster 2 (10.62%), and Cluster 5 (10.62%) are the principal obstacle clusters, dominated by keywords such as “weather,” “extreme,” “experience,” “route,” “high cost,” “tickets,” “parking,” and “capacity.” For clusters with a sample size of less than 10 reviews (Cluster 0, Cluster 1, Cluster 6, Cluster 8), the findings are treated as exploratory results: the core keywords and typical characteristics extracted from these small samples may have instability, and their universality needs to be verified by expanding the sample size in future research. From a spatial perspective, the core obstacle clusters (associated with environmental capacity, extreme weather, and operational efficiency) are primarily concentrated in the Bayanbulak Scenic Area, while the exploratory small-sample clusters are more dispersed across individual niche scenic spots.

Interpretative Analysis: These statistical distributions indicate that natural landscapes in Southern Xinjiang face dual sustainability bottlenecks: inherent ecological vulnerability and operational inefficiency. The prominence of weather-related complaints (Clusters 7 and 2) highlights that visitor experiences are highly susceptible to uncontrollable environmental factors, such as high-altitude temperature fluctuations and wildlife-related safety risks. Furthermore, the capacity and pricing complaints (Clusters 3 and 9) expose severe deficits in environmental carrying capacity and pricing supervision during peak seasons. The distinct spatial differentiation of these clusters confirms that sustainability obstacles are highly site-specific. Therefore, managing these fragile ecosystems requires a transition from generalized environmental monitoring to highly differentiated, site-based mitigation strategies that balance resource protection with tourist needs.

4.3.2. Cultural Landscapes: Erosion of Authenticity and Over-Commercialization

Utilizing the K-means clustering algorithm, a root cause analysis was performed on the 52 negative review texts pertaining to cultural tourist attractions. The optimal number of clusters was identified as K = 8 (Figure 5), leading to the extraction of eight principal issue clusters (Table 4). And a word cloud map (Figure 6) was drawn.

Figure 5. Line chart of K-values.

Figure 5. Line chart of K-values.

Figure 6. Overall word cloud of humanities category.

Figure 6. Overall word cloud of humanities category.

The table below presents the quantity of reviews, corresponding proportions, central keywords, and characteristic features associated with each cluster.

Table 4. Detailed Table of Humanities Clusters.

Cluster ID	Number of Reviews	Proportion	Core Keywords	Typical Review Examples	Representative Case Sites
0	10	19.23%	spots, customers, check, conditions, prices, food, products	Some regional stores suffer from severe homogenization, with their products and dining services lacking distinctive features.	Ancient City of Kashgar
1	5	9.62%	parking, capacity, peak, sparse, hours, limited, traffic	Parking during peak hours is extremely difficult, with all nearby parking lots fully occupied.	Ancient City of Kashgar, Daliyaboyi Scenic Area
2	10	19.23%	high, experience, cost, facilities, changes, supporting, perfect	there are many tourists posing for photos to attract customers, which ruins the experience.	Ancient City of Kashgar, Yecheng County Xitiya Mysterious City Scenic Area
3	3	5.77%	buy, sell, same, price, force, shoulder, if, cold	Some merchants force buyers to buy and sell. Travel photography agencies are full of tricks.	Ancient City of Kashgar, Yecheng County Xitiya Mysterious City Scenic Area
4	6	11.54%	commercialization, available, serious, visitors, hotline, reservation, developed	The commercialization is serious. The scenic area offers limited dining options.	Kunlun Sacred Land in Celer County, Daliyaboyi Scenic Area
5	9	17.31%	ancient, city, walking, even, drink, yogurt, taking, completely	The ancient city has become overly commercialized, with souvenir shops lining every street, completely erasing its original cultural charm.	Ancient City of Kashgar
6	3	5.77%	preparation, sickness, altitude, easily, lead, altitudes, diurnal, cylinders, required	Significant diurnal temperature variation requires attention to the preparation of cold-weather clothing.	Kunlun Sacred Land in Celer County, Daliyaboyi Scenic Area
7	6	11.54%	poor, sanitation, road, toilets, long, traffic, conditions, queues	The ancient city has infrastructure issues, poor sanitation. The interior maintenance of the old building is poor, raising safety concerns.	Ancient City of Kashgar, Yecheng County Xitiya Mysterious City Scenic Area

Table 4. Detailed Table of Humanities Clusters.

Cluster ID	Number of Reviews	Proportion	Core Keywords	Typical Review Examples	Representative Case Sites
0	10	19.23%	spots, customers, check, conditions, prices, food, products	Some regional stores suffer from severe homogenization, with their products and dining services lacking distinctive features.	Ancient City of Kashgar
1	5	9.62%	parking, capacity, peak, sparse, hours, limited, traffic	Parking during peak hours is extremely difficult, with all nearby parking lots fully occupied.	Ancient City of Kashgar, Daliyaboyi Scenic Area
2	10	19.23%	high, experience, cost, facilities, changes, supporting, perfect	there are many tourists posing for photos to attract customers, which ruins the experience.	Ancient City of Kashgar, Yecheng County Xitiya Mysterious City Scenic Area
3	3	5.77%	buy, sell, same, price, force, shoulder, if, cold	Some merchants force buyers to buy and sell. Travel photography agencies are full of tricks.	Ancient City of Kashgar, Yecheng County Xitiya Mysterious City Scenic Area
4	6	11.54%	commercialization, available, serious, visitors, hotline, reservation, developed	The commercialization is serious. The scenic area offers limited dining options.	Kunlun Sacred Land in Celer County, Daliyaboyi Scenic Area
5	9	17.31%	ancient, city, walking, even, drink, yogurt, taking, completely	The ancient city has become overly commercialized, with souvenir shops lining every street, completely erasing its original cultural charm.	Ancient City of Kashgar
6	3	5.77%	preparation, sickness, altitude, easily, lead, altitudes, diurnal, cylinders, required	Significant diurnal temperature variation requires attention to the preparation of cold-weather clothing.	Kunlun Sacred Land in Celer County, Daliyaboyi Scenic Area
7	6	11.54%	poor, sanitation, road, toilets, long, traffic, conditions, queues	The ancient city has infrastructure issues, poor sanitation. The interior maintenance of the old building is poor, raising safety concerns.	Ancient City of Kashgar, Yecheng County Xitiya Mysterious City Scenic Area

Evidence: The negative feedback for cultural heritage sites manifests a clear “dual-core” structure. The core and stable conclusions are based on large-sample clusters with more than 10 reviews: Cluster 0 (19.23%) and Cluster 2 (19.23%) are the principal obstacle clusters, with the first core driven by commercial homogenization issues and the second core revolving around infrastructure and experience defects. For clusters with a sample size of less than 10 reviews (Cluster 1, Cluster 3, Cluster 4, Cluster 5, Cluster 6, Cluster 7), the findings are treated as exploratory results, and their stability needs to be further verified with an expanded sample size. Spatially, the core commercialization and infrastructure obstacle clusters are predominantly localized in the Ancient City of Kashgar, while the exploratory small-sample clusters are mainly tied to remote cultural sites with lower tourist volume, such as Daliyaboyi and Kunlun Sacred Land in Celer County.

Interpretative Analysis: These findings reveal an acute tension between commercialization and heritage preservation in cultural destinations. The high proportion of complaints regarding identical stores and commercialization (Clusters 0 and 5) reflects a pronounced homogenization of merchandise and the erosion of original cultural charm in mature, high-density urban heritage sites like Kashgar. Concurrently, the infrastructure shortcomings (Clusters 2 and 7) pose a long-term threat to cultural sustainability, as poor maintenance of heritage assets accelerates physical degradation. The emergence of physiological risks (Cluster 6) in remote areas further emphasizes that cultural sustainability in Southern Xinjiang is not only about protecting intangible heritage but also about addressing the physical and geographical constraints of the region. Ultimately, the sustainability of these sites depends on balancing commercial vitality with the strict preservation of local daily life and architectural integrity.

4.3.3. Comprehensive Landscapes: Integrated Governance Bottlenecks

For comprehensive tourist attractions, which encompass both natural landscapes and cultural elements, the root cause analysis focused on 118 negative reviews. Through verification techniques, the optimal number of clusters was determined to be K = 10 (Figure 7), resulting in ten principal issue clusters (Table 5). And a word cloud map (Figure 8) was drawn.

Figure 7. Line chart of K-values.

Figure 7. Line chart of K-values.

Figure 8. Overall word cloud for comprehensive category.

Figure 8. Overall word cloud for comprehensive category.

Comprehensive tourist attractions encompass both natural landscapes and cultural elements. The issues identified in negative reviews similarly reflect an integrated “natural and cultural” dimension. The table below presents the number of reviews, their respective proportions, central keywords, and representative characteristics for each category.

Table 5. Detailed Cluster Information for the Comprehensive Category.

Cluster ID	Number of Reviews	Proportion	Core Keywords	Typical Review Examples	Representative Case Sites
0	7	5.93%	leaving, options, extremely, navigation, noodles, Terrifying, mosquitoes	Some homestays have air conditioning that barely works. Driving yourself can lead to tricky road conditions and navigation failures.	139 Poplar Secret Realm Highway, Pamir Tourism Area
1	18	15.25%	poor, high, price, accommodation, hygiene, facilities, summer, conditions	Winter temperatures are low, while summer is hot with intense sunlight, resulting in high road surface temperatures.	139 Poplar Secret Realm Highway, Zepu Poplar Scenic Area
2	4	3.39%	service, tourists, center, practice, photography, charging, regard	The aerial photography service of the observation deck has a hidden practice. The shuttle bus service operates irregularly.	Robu Village Scenic Area, Tomur Grand Canyon
3	20	16.95%	area, long, road, supply, tourists, narrow, risk, clear	The reservoir in the scenic area was once polluted, which affected the impression of tourists.	Bachu Red Sea Scenic Area, Tomur Grand Canyon
4	10	8.47%	scenery, natural, photography, reach, way, less, along, transfers	The photography and lighting of the scenic spot are not beautiful. The natural scenery is not particularly impressive, and the overall experience is average.	Bachu Red Sea Scenic Area, Zepu Poplar Scenic Area
5	10	8.47%	local, tourist, spots, visitors, yuan, popular, photos, like, require	In less-visited spots finding restrooms can be tricky. The sightseeing bus in Kashgar Ancient Town has a scam.	Zepu Poplar Scenic Area, Pamir Tourism Area
6	12	10.17%	experience, insufficien, affects, strong, site, waste, made, maintenance	The scenic spot has domestic waste, which affects the experience. Extreme weather conditions have caused heavy sandstorms on the roads.	Bachu Red Sea Scenic Area, 139 Poplar Secret Realm Highway
7	11	9.32%	sections, unclear, lack, signal, making, gravel, spots, roads	Some areas have no signal, making communication difficult. Some areas are closed for construction.	139 Poplar Secret Realm Highway, Robu Village Scenic Area
8	12	10.17%	only, making, nearly, without, just, heart, easily	Contradiction between Actual Landscape and Internet Rumor. The altitude rises abruptly from 1289 m to over 4000 m, which can easily trigger acute altitude sickness.	Pamir Tourism Area, 139 Poplar Secret Realm Highway
9	14	11.86%	landscape, activities, lacking, main, core, experience, park	The landscape is rather plain and lacks the visual impact one might expect. The fees are not clearly defined.	Tomur Grand Canyon, Bachu Red Sea Scenic Area, Zepu Poplar Scenic Area

Table 5. Detailed Cluster Information for the Comprehensive Category.

Cluster ID	Number of Reviews	Proportion	Core Keywords	Typical Review Examples	Representative Case Sites
0	7	5.93%	leaving, options, extremely, navigation, noodles, Terrifying, mosquitoes	Some homestays have air conditioning that barely works. Driving yourself can lead to tricky road conditions and navigation failures.	139 Poplar Secret Realm Highway, Pamir Tourism Area
1	18	15.25%	poor, high, price, accommodation, hygiene, facilities, summer, conditions	Winter temperatures are low, while summer is hot with intense sunlight, resulting in high road surface temperatures.	139 Poplar Secret Realm Highway, Zepu Poplar Scenic Area
2	4	3.39%	service, tourists, center, practice, photography, charging, regard	The aerial photography service of the observation deck has a hidden practice. The shuttle bus service operates irregularly.	Robu Village Scenic Area, Tomur Grand Canyon
3	20	16.95%	area, long, road, supply, tourists, narrow, risk, clear	The reservoir in the scenic area was once polluted, which affected the impression of tourists.	Bachu Red Sea Scenic Area, Tomur Grand Canyon
4	10	8.47%	scenery, natural, photography, reach, way, less, along, transfers	The photography and lighting of the scenic spot are not beautiful. The natural scenery is not particularly impressive, and the overall experience is average.	Bachu Red Sea Scenic Area, Zepu Poplar Scenic Area
5	10	8.47%	local, tourist, spots, visitors, yuan, popular, photos, like, require	In less-visited spots finding restrooms can be tricky. The sightseeing bus in Kashgar Ancient Town has a scam.	Zepu Poplar Scenic Area, Pamir Tourism Area
6	12	10.17%	experience, insufficien, affects, strong, site, waste, made, maintenance	The scenic spot has domestic waste, which affects the experience. Extreme weather conditions have caused heavy sandstorms on the roads.	Bachu Red Sea Scenic Area, 139 Poplar Secret Realm Highway
7	11	9.32%	sections, unclear, lack, signal, making, gravel, spots, roads	Some areas have no signal, making communication difficult. Some areas are closed for construction.	139 Poplar Secret Realm Highway, Robu Village Scenic Area
8	12	10.17%	only, making, nearly, without, just, heart, easily	Contradiction between Actual Landscape and Internet Rumor. The altitude rises abruptly from 1289 m to over 4000 m, which can easily trigger acute altitude sickness.	Pamir Tourism Area, 139 Poplar Secret Realm Highway
9	14	11.86%	landscape, activities, lacking, main, core, experience, park	The landscape is rather plain and lacks the visual impact one might expect. The fees are not clearly defined.	Tomur Grand Canyon, Bachu Red Sea Scenic Area, Zepu Poplar Scenic Area

Evidence: The comprehensive sites present a pattern of “multi-core dispersion” in their negative feedback. The core and stable conclusions come from large-sample clusters with more than 10 reviews: Cluster 3 (16.95%), Cluster 1 (15.25%), Cluster 9 (11.86%), Cluster 6 (10.17%), Cluster 8 (10.17%), Cluster 7 (9.32%), Cluster 4 (8.47%), and Cluster 5 (8.47%) are the principal obstacle clusters, with prominent frictions in environmental and transportation, service standardization, and safety risks. For clusters with a sample size of less than 10 reviews (Cluster 0, Cluster 2), the findings are treated as exploratory results, and their universality needs to be verified by expanding the sample size in future research. The mapping of core clusters provides a high-resolution diagnostic: Cluster 8 (altitude sickness) is specifically tied to the Pamir Tourism Area, while the core transportation and service obstacle clusters are widely distributed across long-distance comprehensive tourism routes represented by the 139 Poplar Secret Realm Highway.

Interpretative Analysis: The integrated nature of comprehensive sites means they suffer from compounded governance bottlenecks. The prevalence of environmental and transportation complaints indicates that vast geographical distances and local environmental degradation (such as reservoir pollution) severely damage the sustainable image of the destination. Furthermore, the dual presence of service standardization defects (non-transparent charging, inadequate hygiene) and physiological safety hazards (altitude sickness, sandstorms) demonstrates that local governance systems are currently insufficient. The GRU model effectively moves beyond simple sentiment filtering to provide actionable intelligence: it distinguishes between natural constraints (altitude) and operational failures (irregular services). Addressing these bottlenecks requires an integrated governance model that simultaneously mitigates natural risks and enforces strict service standardization to improve resilience and the overall visitor experience.

5. Discussion and Management Implications

5.1. Dialogue Between Research Findings and Existing Literature

5.1.1. Model Performance: Consistency and Extension of NLP Sentiment Analysis Literature

The results show that the GRU model achieves the best performance in sentiment classification of short tourism texts in Southern Xinjiang, with an accuracy of 91.62% and F1 score of 0.9162, outperforming LSTM and standard RNN. This finding is consistent with existing studies. Zhang et al. [5] pointed out that gated recurrent neural networks have obvious advantages over lexicon-based and traditional machine learning methods. Lin et al. [33] further confirmed that GRU is more efficient and accurate than LSTM and CNN in short Chinese text analysis.

This study extends the existing literature by verifying the adaptability of GRU in a specific scenario: tourism reviews in border ethnic areas. Previous studies such as Li et al. [31] and Peng et al. [32] focused on hotels, ski resorts and general destinations, lacking empirical tests for border regions with unique geography and culture. This study fills this gap.

Notably, this study further verifies the practical governance value of the GRU model in destination management. While traditional star-rating systems can efficiently filter negative reviews, they fail to reveal the underlying semantic nuances of negative feedback. Unlike simple numerical ratings, the GRU model deployed in this study acts as a deep semantic parsing tool: it not only identifies negative sentiments, but also extracts highly specific contextual dimensions (e.g., distinguishing between “poor sanitary conditions” and “extreme weather hazards”), providing actionable governance intelligence that a basic 1-star or 2-star rating cannot convey. This advantage is particularly critical for the refined sustainable management of tourism destinations in border areas, and is also a key supplement to existing tourism sentiment analysis research.

5.1.2. Obstacle Heterogeneity: Consistency and Breakthrough in Sustainable Tourism Research

This study identifies significant heterogeneity in sustainable obstacles across natural, cultural and comprehensive scenic spots, which is consistent with the core view of sustainable tourism assessment. Torres-Delgado & Saarinen [36] indicated that traditional static evaluation cannot reflect real tourist experience. Uslu et al. [4] and Wang et al. [38] proposed that UGC can be used as a social sensor to capture real bottlenecks of destinations, which supports the analytical logic of this study.

This study makes a key breakthrough compared with previous research on Xinjiang tourism. Most studies such as Ding et al. [20] adopted macro or single-scenic-spot analysis. This study systematically diagnoses 38 A-level scenic spots in Southern Xinjiang and identifies 28 obstacle clusters, revealing clear differences among three types of destinations, breaking through the homogeneous evaluation paradigm in the past.

5.1.3. ESG Framework: Complementarity of Border Tourism Governance Research

The ESG-based optimization path proposed in this study is consistent with frontier tourism governance research. Camilleri [3] emphasized that ESG is the core of improving destination resilience. Rasoolimanesh et al. [40] confirmed that environmental, social and governance collaborative optimization helps to achieve the UN Sustainable Development Goals.

The unique contribution of this study is to localize the ESG framework for border ethnic areas for the first time, and put forward differentiated strategies for natural, cultural and comprehensive scenic spots, which makes up for the lack of targeted ESG governance research in ecologically fragile and culturally diverse border regions.

5.2. Theoretical Contributions

(1) Methodological contribution: This study constructs a multi-model deep learning comparison framework for sustainable tourism diagnosis in border areas, and confirms that GRU is the optimal model for short tourism texts in Southern Xinjiang, expanding the application boundary of deep learning in niche border tourism scenarios.

(2) Theoretical framework contribution: It reveals the heterogeneous characteristics of sustainable obstacles in different types of scenic spots, breaks through the limitations of traditional homogeneous sustainable tourism assessment, and improves the refined analysis theory of sustainable tourism in border ethnic areas.

(3) Technical integration contribution: It integrates automated data scraping, deep learning sentiment parsing and unsupervised root-cause clustering, forming a reproducible intelligent monitoring framework for tourism destinations in border areas.

(4) Practical theory contribution: It verifies that UGC can be used as a dynamic social sensor for sustainable tourism assessment, making up for the deficiency of traditional static questionnaire surveys that are limited by sample size and memory bias.

It is important to note that the GRU sentiment analysis framework proposed in this study serves as a macro-level diagnostic radar for regional tourism sustainability. While it efficiently identifies spatial–temporal symptom clusters across large regions, designing efficient micro-level management policies requires these digital insights to be validated through on-site empirical research. Policymakers must use these data-driven findings as a starting point to conduct targeted field investigations, ensuring that the proposed ESG optimization paths address the root causes tailored to each site’s particularities.

5.3. Practical and Policy Implications Under the ESG Framework

The root cause clustering of negative evaluations confirms that the sustainable development obstacles of Southern Xinjiang tourism are rooted in the structural defects of the destination’s ESG ecosystem. Based on the clustering results, we strictly distinguish between core sustainable development obstacles and general service quality complaints, and propose targeted optimization strategies deeply aligned with the unique natural, cultural and geographical characteristics of Southern Xinjiang for core sustainability issues, supplemented by operational optimization measures for service quality problems, to achieve the synergy of ecological protection, cultural inheritance and high-quality tourism development, and align with the UN SDGs.

5.3.1. Targeted Strategies for Natural Landscape Destinations

Aiming at the core obstacles of ecological vulnerability, extreme weather risks, peak season carrying capacity overload and operational inefficiency, the optimization paths focus on the Environmental (E) dimension, supplemented by standardized governance:

(1) Implement dynamic carrying capacity control: Formulate differentiated passenger flow limits for peak and off-peak seasons for core ecological scenic spots (e.g., Bayanbulak, Taklimakan N39), strictly implement time-sharing reservation and unified shuttle bus scheduling, to avoid exceeding the environmental carrying capacity.

(2) Build an environmental risk early warning system: Improve real-time monitoring and emergency response mechanisms for extreme weather, wildlife activities and water safety risks, and push warning information to tourists in advance.

(3) Standardize operation and service management: Unify and publicize the pricing standards of catering, accommodation and experience projects, strengthen staff training, and establish a fast complaint handling mechanism.

(4) Optimize peak season supporting facilities: Increase temporary parking spaces, mobile environmental protection toilets and garbage recycling points, and optimize traffic diversion routes to alleviate congestion and sanitation problems.

5.3.2. Targeted Strategies for Cultural Destinations

Aiming at the core obstacles of over-commercialization, cultural authenticity erosion, product homogenization and irregular business operations, the optimization paths focus on the Social (S) dimension, supplemented by standardized governance:

(1) Implement strict functional zoning management: For high-traffic cultural scenic spots (e.g., Ancient City of Kashgar), clearly delimit core cultural protection areas, folk life experience areas and commercial service areas, and strictly restrict the disorderly expansion of homogeneous commercial formats in protection zones.

(2) Protect cultural authenticity and community vitality: Support local residents to participate in tourism operations, give priority to developing intangible cultural heritage experience projects with ethnic characteristics, and curb the homogenization of souvenir shops and travel photography agencies.

(3) Rectify irregular business behaviors: Establish a joint law enforcement mechanism, implement a “blacklist” system for merchants with forced trading and price fraud, and publicize the handling results to the public.

(4) Strengthen heritage protection and infrastructure upgrading: Formulate a regular maintenance plan for ancient buildings and cultural relics, optimize tourist routes, and improve sanitation and supporting facilities in the scenic spot.

5.3.3. Targeted Strategies for Comprehensive Destinations

Aiming at the core obstacles of compound governance bottlenecks, insufficient infrastructure, plateau safety risks and non-standard services, the optimization paths focus on the Governance (G) dimension, and coordinate the optimization of environmental and social dimensions:

(1) Improve integrated infrastructure guarantee: Optimize road conditions, add mobile signal base stations along long tour routes (e.g., Pamir Tourism Area), and improve rest stops, emergency rescue points and other supporting facilities.

(2) Build a full-chain safety guarantee system: Set up altitude sickness prevention and rescue stations in high-altitude scenic spots, provide oxygen supply and medical emergency services, and push extreme weather and altitude adaptation guidelines to tourists in advance.

(3) Standardize whole-process service management: Unify the operation standard of shuttle buses, fix departure intervals and routes, publicize the charging standards of all experience projects in advance, and strengthen the whole-process supervision of service quality.

(4) Create a differentiated integrated experience: Deeply tap the cultural connotation behind natural landscapes, launch integrated experience projects combining natural sightseeing and cultural experience, to avoid homogeneous competition and enhance core competitiveness.

5.3.4. Macro Governance Recommendations for Regional Authorities

(1) Establish a normalized dynamic monitoring system for tourism sustainability based on the GRU sentiment analysis framework, to realize real-time capture of tourist feedback and pre-warning of sustainable risks.

(2) Formulate differentiated assessment standards for different types of scenic spots, incorporate ESG indicators into the rating and assessment system of A-level scenic spots, and align with the UN SDGs.

(3) Establish a multi-department collaborative governance mechanism, coordinate the work of cultural tourism, market supervision, ecological environment and other departments, to form a joint force for sustainable tourism governance in Southern Xinjiang.

6. Conclusions and Outlook

6.1. Core Findings and Theoretical Contributions

This study builds a comprehensive diagnostic framework for sustainable tourism in Southern Xinjiang by integrating automated data scraping, GRU-based sentiment analysis and unsupervised clustering, based on 5800 valid reviews from 38 local A-level scenic spots. The core findings are as follows: First, the GRU model achieves the best comprehensive performance in short tourism text sentiment classification, with an overall accuracy of 91.62% and F1 score of 0.9162, outperforming RNN and LSTM, and strikes an optimal balance between accuracy and computational efficiency for border tourism scenarios. Second, the overall tourist satisfaction of the sampled scenic spots reaches 88.6%, with a 4.9% negative review rate, and sustainable development obstacles show significant site-specific differentiation: natural scenic spots are mainly constrained by ecological vulnerability and operational inefficiency, cultural destinations face core conflicts between over-commercialization and cultural authenticity protection, and comprehensive scenic spots suffer from compound governance bottlenecks. A total of 28 distinct obstacle clusters are identified through clustering analysis. Third, the ESG framework can provide a systematic solution for local sustainable tourism development, and targeted differentiated strategies can effectively enhance the resilience of the regional tourism industry.

This study makes four core theoretical contributions: it expands the scenario application of deep learning sentiment analysis in border ethnic tourism, improves the heterogeneous assessment theory of sustainable tourism in border areas, deepens the theoretical application of UGC in tourism sustainability assessment, and enriches the localized research of the ESG framework in border tourism governance.

6.2. Practical and Political Implications

For regional government authorities, the GRU-based analysis framework can support the establishment of a normalized dynamic monitoring system for tourism sustainability, to realize the transformation from post-event rectification to pre-event early warning. Authorities should formulate differentiated management policies for different types of scenic spots, incorporate ESG indicators into the A-level scenic spot rating system, and improve supporting institutional systems including market supervision, emergency rescue and ecological protection, to align with the UN Sustainable Development Goals.

For scenic area operators, targeted optimization should be carried out according to obstacle characteristics: natural scenic spots should focus on ecological capacity control and environmental risk early warning; cultural destinations should prioritize cultural authenticity protection and standardized business management; comprehensive scenic spots should improve infrastructure support, safety guarantee systems and differentiated experience supply.

6.3. Limitations and Future Research Directions

Despite the contributions of this study, several limitations need to be acknowledged to clarify the application boundary of the conclusions. First, the study has systematic bias from a single data source: all samples are exclusively from the Xiaohongshu platform, whose core user group is predominantly young women aged 18–35. The tourism perception and evaluation criteria of this group differ from those of middle-aged and elderly tourists, male tourists and group tour participants, so the samples cannot fully represent the views of all tourist groups, which limits the generalizability of the conclusions to a certain extent. Second, there are limitations in the comprehensiveness of model performance evaluation and application. On the one hand, the pre-training dataset is not fully customized for Southern Xinjiang’s border tourism scenario; on the other hand, for the imbalanced class distribution of the sample (negative sentiment accounts for only 4.9%), although we have verified the model’s reliable recognition ability for the minority negative class through the confusion matrix (Figure 2a), we did not supplement the specific precision, recall, F1 score and Area Under the Curve (AUC) metrics for the negative class in the original manuscript, which makes the performance evaluation of the model on imbalanced datasets not comprehensive enough. Third, the clustering analysis has certain limitations: some clusters have a small sample size (fewer than 10 reviews), and although we treat these as exploratory findings, the extracted core keywords may still have instability and potential overinterpretation risks. Finally, the study does not deeply explore the causal mechanism between external variables (such as seasonal passenger flow, policy adjustments) and tourist sentiment fluctuations, and the universality of the analysis framework in other border ethnic areas still needs further verification. Meanwhile, this study focuses on negative reviews to identify sustainability obstacles and has not yet conducted in-depth mining of neutral reviews that may contain potential sustainability concerns and positive reviews that can be used as a benchmark for high-quality development.

Corresponding to the above limitations, specific future research directions are proposed. First, subsequent research will focus on correcting the selectivity of the existing data and improving sample representativeness through three specific strategies: first, integrate multi-source data from mainstream tourism platforms such as Ctrip, Meituan and Douyin to cover tourist groups with different genders, ages, places of residence and travel modes; second, adopt stratified random sampling based on the demographic characteristics of tourists in Southern Xinjiang to balance the sample structure; third, establish a long-term continuous data collection mechanism, and regularly update the sample according to the seasonal changes of tourism in Southern Xinjiang, to maintain the timeliness and representativeness of the data. Second, future research can optimize the sentiment analysis model for border tourism scenarios, customize annotation rules and datasets for Southern Xinjiang, further improve the model’s adaptability to imbalanced datasets, and supplement specific performance metrics for the minority negative class to make the model evaluation more comprehensive and rigorous. Third, the sample size of negative reviews can be expanded to verify the stability of small-sample clustering results and explore the formation mechanism of these exploratory obstacle clusters. Finally, future research can deepen the analysis of the causal mechanism of sustainability obstacles, conduct in-depth mining of neutral and positive reviews to form a complete benchmarking and optimization system, analyze the temporal characteristics and recurrence rate of negative comments, explore the correlation between obstacle clusters and seasonal passenger flow, local events and other external factors, and replicate the analysis framework to other border areas to carry out cross-regional comparative research and expand the universality of the conclusions.