Forest Visitors’ Multisensory Perception and Restoration Effects: A Study of China’s National Forest Parks by Introducing Generative Large Language Model

Abstract

Sensory perception of forests is closely related to human health and well-being. Based on attention recovery theory and stress relief theory, this paper investigates the influence of sensory perception of forests on visitors’ restoration effects from a multidimensional and multisensory perspective, integrating the use of a generative large language model, regression analysis, and semantic analysis. The results of the study show that (1) the application of a generative large language model provides new ideas and methods to solve the dilemma caused by the traditional self-report scale measurement and provides a possible way to explore a new research paradigm in the context of the rapid development of generative artificial intelligence; (2) the effects of each sensory quantity differed, with the sensory quantities of sight, hearing, touch, and taste having a significant positive effect on visitors’ restoration effects, and the sense of smell having a significant negative effect on visitors’ restoration effects; (3) sensory psychological distance partially had a significant effect on visitors’ restoration effects, both proximal psychological distance and distal psychological distance were significantly correlated with visitors’ restoration effects, and intermediate psychological distance had a negative effect on visitors’ restoration effects, but the effect was not significant; (4) the sensory dimension has a significant positive effect on visitors’ restoration effects, the integration and synergistic effect of the senses are enhanced, and multidimensional sensory cross-perception has a positive effect on visitors’ restoration effects at the social health level; and (5) the sensory elements of National Forest Parks that influence visitors’ restoration effects are mainly natural attributes, and the elements related to “people” also play an important role in visitors’ restoration effects. This study provides a useful complement to the study of forest sensory perception, and at the same time has an important reference value for exploring the management of forest recreation experience and sensory marketing practices.

1. Introduction

Currently, in the context of global climate change, biodiversity destruction, and pollution issues, health is increasingly becoming a major public health issue and a prominent social problem worldwide [1]. The uncertainty after the COVID-19 pandemic has intensified the potential health threats to human society, and physical and mental health has become a highly prioritized global issue. Studies show that global rates of anxiety and depression increased by more than 25% in the first year of the COVID-19 pandemic [2]. Previous research demonstrates that being in a natural environment can effectively enhance physical, mental, and emotional well-being [3,4,5,6,7]. Among various natural settings, forests have been extensively shown to offer advantages for psychological, physiological, and social well-being. Various cultural ecosystem services are provided by forests around the world, including forest therapy, forest bathing, and shinrin-yoku [8].

Tourism and recreational activities carried out in forests are often in multisensory stimulating environments [9,10]. Sensory perception in forests is usually associated with attention and emotional regulation. Restorative properties are typically experienced through sensory perception, and the quality of sensory perception and experience can also impact restoration effects. Research indicates that engaging all senses enhances the relaxation and restorative potential of an environment [11], while different sensory stimuli may produce different restorative experiences.

However, current research has limitations. Firstly, studies predominantly concentrate on single sensory perception. In real forest settings, individuals’ sensory perception is often all-inclusive and even too complex to discern entirely. The restorative effects of a single sensory input experienced by visitors have primarily been examined through a controlled laboratory environment [8], which is somewhat different from the multidimensional and multisensory perception of the real environment. Secondly, numerous studies have examined the relationship between forest sensory perception and visitors’ restoration effects. However, many of these studies only scratch the surface of the relationship without delving into the specific sensory elements. Moreover, exploring only a certain type or a few types of sensory perception in a generalized manner still holds limited practical significance. Once again, due to the high cost of collecting actual environmental data, studies are generally restricted to particular forests or regions, and large-scale, national-level, region-wide, and long-term studies remain infrequent. Finally, self-reported scales are often used to collect data in studies of forest sensory perception and visitors’ restoration effects, which have many limitations in terms of data reliability and consistency, individual subject preferences, memory bias, and the influence of social expectations [12,13], which affect the reliability of the studies to a certain extent.

Benefiting from the rapid improvement of current computer performance, the generative model that learns the probability density of observable samples and generates new samples randomly has become a prominent topic [14]. This type of model can not only realize the mining of shallow semantic features but can also achieve the learning of abstract deep semantic features, and the model performance is superior. The generative AI known as ChatGPT is a standard large language model with minimal data annotation needs, the ability to generate cross-modal content, and robust logic and organization skills [15,16,17]. Recently, there has been a growing interest in utilizing generative large language models within the field of social sciences [18], with some even considering it a new paradigm [19].

The emergence of generative large language models offers potential ideas and approaches to overcome constraints in the present research area of forest sensory perception and the restorative effects on visitors. Initially, conduct multi-dimensional sensory studies in actual surroundings. The generative large language model utilizes genuine user-generated data from the environment, such as travel notes, to overcome the limitations of laboratory data, which can be divorced from the real environment and target a single sensory perception to some extent. As a result, it better reflects individual behavior in the actual forest environment. Additionally, there is extensive, inter-temporal, and massive data batch processing. The generative, large language model facilitates batch processing of travelogues, spanning various research groups and temporal-spatial dimensions. This enables large-scale, territory-wide, and time-spanning research with enhanced data representativeness and higher data processing efficiency. Furthermore, it mitigates the limitations of self-reported scale measurements. The large generative language model can produce relatively consistent data when operating under identical input conditions, thus eliminating potential bias caused by individual variations and subjectivity in self-reported questionnaires. Additionally, this approach sidesteps the issues arising from social expectations, self-presentation, and recall delay that often plague manual completion. Finally, this paves the way for expanding research horizons and areas of investigation. The use of a generative large language model, based on rich corpus training, can offer a more extensive analysis of the scale. For instance, it can provide a comprehensive assessment of the various dimensions of the perceptual restorative scale concerning environmental elements, sensory perception, and comfort. This, in turn, enables researchers to broaden their horizons and gain potential insights.

Therefore, the research objectives of this paper are threefold: (1) to expand the application of generative large language models in the field of forest recreation and tourism, and to provide possible ideas and methods for solving the dilemmas posed with traditional self-reported scale measurements; (2) to explore the influence relationship between multisensory perception and visitors’ restoration effects of forest tourists; and (3) to clarify the deep-level association between various sensory elements and visitors’ restoration effects, and to construct a multidimensional and multi-level spectrum of forest sensory perception elements. The research results provide a theoretical basis and practical guidance for the users of the forest environment and the improvement of the health and well-being of tourists.

2. Theoretical Analysis and Research Hypotheses

2.1. Attention Restoration Theory and Stress Recovery Theory

In exploring the multidimensional and multisensory experience and restoration effects of forests, attention restoration theory and stress recovery theory provide theoretical ideas for our study. The attention restoration theory was proposed by the Kaplans in 1995 and argues that human attention can be divided into two types, voluntary attention and involuntary attention, and that voluntary attention causes inhibition of nerve centers and overuse of function, resulting in fatigue, difficulty in concentrating, and ease of agitation, which may lead to errors in work [11,20]. On the contrary, involuntary attention occurs spontaneously without the need to invest a lot of energy in paying attention, so it is conducive to the recovery of intentional attention [21]. Attention restoration theory suggests that human beings inherently favor the natural environment so that when they are in nature, they mainly focus on involuntary attention, which leads to the restoration of voluntary attention [22]; this kind of environment is called the restorative environment [23], and forests, grasslands, and urban green spaces are typical restorative environments.

Stress recovery theory was developed by Ulrich in 1991. The theory states that the natural environment has a positive effect on the physical and emotional recovery of human beings [24]. Specifically, there are three elements that play a role in relieving stress in people: non-threatening landscapes, greenery, and specific natural landscapes [25]. Ulrich argues that this ability to respond to stress relief in specific natural environments comes as a result of human evolution [26]. Humans respond to the three types of natural environments mentioned above instantly and unconsciously, requiring minimal cognitive resources to process responses, and are thus able to quickly recover their strength from stress [20].

2.2. Sensory Quantity and Visitors’ Restoration Effects

The term ‘sensory quantity’ pertains to the amount of different senses that individuals encounter when engaging in recreational activities in forested areas. According to the attention recovery theory, which discusses the relationship between the experience of natural environments and involuntary attention, the increase in sensory stimulation in forest environments usually occurs without the need to actively devote attention and thus reduces the brain’s tension in conscious tasks and helps the brain to achieve rest and recovery. For instance, forests are typically associated with attractive natural landscapes and peaceful surroundings [27]. Additionally, the sounds of birds chirping, water flowing, and wind blowing within forests are often experienced as enjoyable [28]. Furthermore, physical contact with trees may cultivate a stronger sense of connectedness between individuals and natural environments. The scents of trees and flowers promote physical and mental relaxation [29,30] while indulging in a picnic or dining outdoors in the forest enhances one’s positive emotions. According to this theory, more sensory stimulation may better facilitate involuntary attention, thus helping to restore resources for voluntary attention. An increase in sensory experience involves not only an increase in mere quantity but also an enhanced diversity of sensory stimuli, which in turn stimulates different feelings and emotions and increases the levels and dimensions of attentional recovery, thus positively affecting visitors’ restoration effects. Therefore, we propose the following hypotheses:

H1a: 

The number of visual senses has a significant positive effect on the visitors’ restoration effects.

H1b: 

The number of hearing senses has a significant positive effect on the visitors’ restoration effects.

H1c: 

The number of smell senses has a significant positive effect on the visitors’ restoration effects.

H1d: 

The number of touch senses has a significant positive effect on the visitors’ restoration effects.

H1e: 

The number of taste senses has a significant positive effect on the visitors’ restoration effects.

2.3. Sensory Psychological Distance and Visitors’ Restoration Effects

Psychological distance can be defined as the subjective experience that something is very close or very far from the self, here and now [31,32]. Various dimensions determine the way in which objects are dispersed in psychological space, resulting in different types of psychological distance. According to the level of interpretation theory, when one’s psychological distance is close, greater focus is placed on concrete details, utilizing fewer abstract mental representations, whereas when one’s psychological distance is more distant, intensified attention is paid to points of principle and deeper meanings, disregarding both concrete content and details, thereby utilizing more abstract mental representations [31,33,34].

For the psychological distance of the senses, Ryan argues that the maximum distance perceived by the senses varies in different sensory experiences [35]. Specifically, things can only be tasted or touched when they are held in the mouth or touched by hand, so touch and taste belong to the sensory experience of proximal psychological distance; through light reflection and sound waves, things can be seen or heard even at a longer distance, so vision and hearing belong to the sensory experience of distal psychological distance. The sense of smell relies on the diffusion of molecules and can be perceived by individuals within a certain physical distance, so the sense of smell is between the proximal and distal end of the psychological distance sensory experience [36].

Different tourism and recreation activities exhibit varying psychological distances for tourists, which align with different levels of interpretation. For example, relaxing tourism activities possess a closer psychological distance, corresponding with a lower level of interpretation. In contrast, tourism activities that provide challenges tend to be associated with farther psychological distances, indicating higher levels of interpretation. Ensuring alignment between various psychological distances and levels of interpretation can result in stimulating positive psychological states [37]. Tourism and recreation activities in forests are typically relaxing tourism activities, and thus, the proximal psychological distance senses of taste and touch will have a more positive impact on the visitors’ restoration effects. Further, from the perspective of the methods and properties of perception, proximal psychic distance sensory experiences like touch and taste, which are typically experienced through physical contact or ingestion, are more likely to create profound interactions and emotional connections and foster positive restorative effects in visitors. Meanwhile, distal psychic distance sensory experiences such as visual and auditory experiences possess passive features and might be adversely impacted by excessive stimuli or noise pollution. Therefore, we contend that all sensory experiences in forest environments are impacted by psychological distance. The senses’ restorative effects are more pronounced the closer the distance, and conversely, less favorable when greater. Thus, we formulate the hypothesis accordingly.

H2a: 

The proximal psychological distance sensory has a significant positive effect on the visitors’ restoration effects.

H2b: 

The medial psychological distance sensory has a significant negative effect on the visitors’ restoration effects.

H2c: 

The distal psychological distance sensory has a significant negative effect on the visitors’ restoration effects.

2.4. Sensory Dimensions and Visitors’ Restoration Effects

In the real world, the human brain can combine cues from various sensory channels to perceive external objects and events [38,39]. This process effectively merges information from different cues within the same sensory channel and cues from different sensory channels into a unified, coherent, and robust perception, known as perceptual integration of cues from multiple sensory channels, or multisensory integration for short [40]. Multiple fields, including neurophysiology, electrophysiology, and psychology, have conducted studies on multisensory integration. Scholars in recreation and tourism research have emphasized the synesthesia effect as an interconnected, cross-sensory experience [41]. Certain scholars examine the impact of multisensory interactions on the identity of tourism destinations, with sight and smell deemed as the most effective forms of synesthesia [42]. The research indicates that the way in which natural environments’ soundscapes are characterized, either directly or through the mediation of visual landscapes, contributes significantly to the tourists’ attentional recovery and quality of life [43]. Compared to the rise in the quantity of senses in terms of total count, sensory dimensions predominantly explore the impact of sensory integration, crossover, and complementarity among multiple dimensions on visitors’ restoration effects. Augmenting the number of senses does not necessarily lead to an increase in sensory dimensions, whereas enhancing sensory dimensions can potentially heighten the depth and richness of the experience. Therefore, we posit the following hypothesis:

H3: 

Sensory dimension has a significant positive impact on visitors’ restoration effects.

3. Methodology

With the continuous development of artificial intelligence, computational intelligence and perceptual intelligence are gradually developing toward the level of cognitive intelligence [44]. The further improvement of computing power and the explosive growth of massive data have led to the application of deep learning methods in natural language processing. Since 2018, the concept of transfer learning has been introduced, and large language models with a huge training corpus scale and parameters have emerged. Representative language models include ELMo, ULMFiT, BERT, PaLM, BLOOM, and OpenAI GPT.

In recent years, there has been growing skepticism toward the use of self-reported scale measurements in psychology due to the high cost of questionnaires, the limited representativeness and generalization ability of data, the difficulty in guaranteeing the reliability and consistency of data, and the interference of subjects’ individual preferences, memory bias, and social expectations [12]. The rapid development of generative large language models provides possible ideas and methods for generating scale questionnaires to solve the above dilemmas to a certain extent by using travelogue texts. First of all, semantic analysis of texts of different types and genres through machine learning and deep learning models, and then understanding the subjective emotions and mental states of the authors, has achieved some results in previous studies [45,46,47]. Texts targeting the travelogue aspect of tourists have been similarly applied to the study of travelers’ emotions, subjective preferences, and mental states [48,49]. Secondly, travelogue texts serve as a factual account of tourists’ travel activities and corresponding emotions. They capture vivid sensory experiences and can demonstrate the sensory, emotional, and psychological states of tourists [50]. This aligns with the primary focus of psychological scale questions and lays the groundwork for developing scale questionnaires from travelogue texts. Finally, the deep learning-based generative language model exhibits a remarkable performance in reading comprehension tasks, enabling intricate text mining and profound semantic analysis. Moreover, it presents the opportunity to construct extensive questionnaires by means of travelogue texts at a methodological level.

For this reason, this study tries to introduce a generative large language model, on the basis of which the research methodology and analysis route of this paper is formed (Figure 1), which includes a total of five modules: data collection, data preprocessing, generative large language model construction, regression model construction, and semantic analysis.

3.1. Study Area

This study takes China’s National Forest Parks as the research area (Figure 2). These parks were established by China’s National Forestry and Grassland Administration to rationally utilize forest landscape resources and promote forest tourism. The parks are classified into three tiers, of which the national tier is the pinnacle. It encompasses forest parks with stunning forest views, concentrated cultural landscapes, high aesthetic value, as well as scientific and cultural importance. The parks also occupy significant geographical locations, have a certain degree of regional representativeness, and provide a complete range of tourism services, making them highly popular. The attributes and characteristics of National Forest Parks are representative and typical for this paper to study the sensory experience and visitors’ restoration effects of forests. The study selects 897 of China’s National Forest Parks.

3.2. Data Collection

In this study, we collected travelogue text data to conduct a study on forest tourists’ multisensory perception and recreational effects. User-generated social media data have certain advantages in conducting such studies: first, the data scale is larger and sample-rich. The large scale of social media data contains rich information from a large number of users, allowing researchers to observe large and diverse sensory experiences. Second, authenticity and real time. Social media data are generated in real time and can capture users’ sensory experiences in real life with timeliness. Third, lower cost. Compared to professional guidance, sensory perception research based on user-generated social media data has lower time and labor costs.

In accordance with the selection criteria of being manageable, reliable, and representative, combined with the characteristics of the research object of this paper, Mafengwo and Ctrip were selected as the text collection platforms of National Forest Parks.

Mafengwo is the largest UGC tourism website in China, with more than 100 million users [51,52]; Ctrip is one of the largest comprehensive tourism service websites in China, which has built a more comprehensive UGC content service system, and many scholars have used its data to conduct tourism-related research [53,54,55]. The study takes the name of the National Forest Park as the keyword, and searches in the platform of Mafengwo and Ctrip, using the collector crawler software (version 4.0.1) to obtain 1557 Mafengwo travelogues and 292 Ctrip travelogues, totaling 1849 travelogues and 3,272,687 words.

3.3. Data Preprocessing

Before analyzing the data, it is imperative to carry out essential preprocessing steps such as cleaning and filtering to enhance the quality of the data. The data preprocessing in this study comprises two components: data cleaning and data selection. The primary objective of the data cleaning process is to remove unnecessary HTML tags, nonsensical emojis, and other symbols from the data. On this basis, data selection is carried out, which means that for the characteristics of the research object, a number of rules applicable to the National Forest Park are proposed in a targeted manner, and the noisy data with less relevance to it are eliminated to ensure the quality of model generation. The principles are as follows: firstly, exclude travelogues that do not mention the name of the National Forest Park or only contain relevant pictures. Secondly, exclude travelogues that do mention the National Forest Park but lack substantive content or only use the park’s name as a caption for pictures. Third, when dealing with travel notes that touch on National Forest Parks but are not the main focus of the text, it is necessary to locate the paragraph containing the park’s name and the one preceding and following it. Once identified, the text should be deduplicated. Fourth, due to the complexity of the visitors’ restoration effects, if the visitors’ restoration effect-related items in the model generation results are extracted as “none” and the results are correct after manual review, such travelogues will be deleted.

3.4. Research Methods

This study utilizes a combination of a large language model, regression analysis, and semantic analysis. First, the independent and dependent variables of forest visitors’ sensory perception were generated by a generative large language model. Second, the relationship between sensory quantity, sensory distance and sensory dimensions, and the visitors’ restoration effects was empirically analyzed using the ordinary least squares regression model. Finally, relying on the semantic analysis method of natural language processing, we use Chinese word segmentation, part-of-speech tagging, and dependency parsing to identify the specific sensory elements that affect the visitors’ restoration effects, and construct the sensory element spectrum.

4. Generative Large Language Model

The application of generative large language models mainly includes three modules: construction of prompt templates, API calls and parameter settings, and model evaluation (Figure 3).

4.1. Prompt Templates Construction

Prior to large-scale experiments, we conducted experiments in a sandbox environment. First, the bias problem. Due to the possible bias of the training data and the algorithm itself, the content generated by generative big language models may be biased toward a certain group of people or produce discriminatory results, and the general bias can be classified into social bias and linguistic bias [56]. Because the travelogue data in this study contains less content in the areas of politics, religion, gender, etc., the social bias is more obvious; coupled with the fact that the model uses the existing travelogue data to generate the training results, the input data is relatively more controllable and deterministic, and thus the social bias has less impact on the study. Another type of bias is language bias, but as Chinese data were used as one of the ChatGPT training data sources, language bias has less impact. However, taking into account that the English training data compared to the Chinese data is much larger and richer, this study still tries to use English for testing; that is, the prompt template and the travelogue text are translated into English, and it is found that affected by the translation effect, the output effect of the model is not significantly improved compared to that of Chinese. Second, the problem of AI hallucinations. AI hallucinations refer to the content generated by the model that is not based on any real data but is a product of the imagination of the large language model [57,58]. Because the input data of this study is relatively controllable and certain, and the use of the model is ‘something out of something’ rather than ‘something out of nothing’, the illusion problem has less impact on the study.

On this basis, the prompting strategy was optimized for the characteristics of this study. First, clarity. Formulate questions as clearly as possible for sensory experience and visitors’ restoration effects; for example, appropriately use separators to distinguish different questions and expressions in the perceptual restoration measure to eliminate ambiguity in the prompts. Second, stepwise. The extraction and generation of variables, such as sensory perception and restoration effect, require a methodical approach to prevent mistakes during the model’s inference process. Therefore, we adopt a step-by-step methodology. Thirdly, providing examples. We supply examples to establish consistency in the model output’s style, format, and content details. This assists in gathering statistics related to variables including sensory perception and visitors’ restoration effects. Finally, providing reasoning time. There is a large number of travelogue texts and a common issue is that a single travelogue can be lengthy. To prevent the model from making mistakes when extracting content from corresponding variables, subsequent queries are used in the prompts. This ensures that the model does not miss relevant content and obtains better performance.

4.2. API Calls and Parameter Setting

Access the GPT-3.5 model through the OpenAI API and set the model parameters [59], including temperature, max tokens, top P, presence penalty, and frequency penalty. Temperature controls the randomness and determinism of the model’s output, with values ranging from 0 to 2. The lower the value, the more centralized and deterministic the model output results, and the higher the value, the more random and diverse. The presence penalty determines the probability of introducing a new subject in the output, with a range of values from −2.0 to 2.0. The frequency penalty controls the likelihood that the output results in the model repeating the same line verbatim, taking values between −2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text, thus reducing the likelihood of the model repeating the same line word by word. The presence penalty is a one-time addition that applies to all tokens that have been sampled at least once, while the frequency penalty is proportional to the frequency with which a particular token has been sampled. The model parameters for this study were determined through iterative debugging with small samples (

Table 1. Parameter setting.

Name	Setting
temperature	1
max_tokens	5000
top_p	0.95
presence penalty	0
frequency penalty	0

).

4.3. Model Evaluation

The large language model evaluation in this study can be regarded as a class of classification model performance evaluation. In order to evaluate the model results, we extracted the text of the travelogues from the experimental sample travelogues for manual annotation. The model performance is examined by calculating the Precision and Recall, and the balance between the two is quantified using the F1 Score. The formula is shown below:

$$Precision=\frac{TP}{TP+FP}$$

(1)

$$Recall=\frac{TP}{TP+FN}$$

(2)

$$F1Score=2\ast \frac{Precision\ast Recall}{Precision+Recal}$$

(3)

where TP denotes the number of true positives, FP denotes the number of false positives, and FN denotes the number of false negatives.

5. Results

5.1. Generative Large Language Model Evaluation

The generative large language model evaluation is mainly divided into two parts. The first is the classification effect evaluation for the five sensory categories of sight, hearing, smell, touch, and taste (

Table 2. Model evaluation results of sensory categorization.

Term	Precision	Recall	F1 Score
Hearing	0.67	1.00	0.80
Sight	1.00	0.82	0.90
Smell	0.75	0.60	0.67
Taste	0.50	0.67	0.57
Touch	0.50	1.00	0.67
Accuracy			0.77
Macro avg	0.68	0.82	0.72
Weighted avg	0.82	0.77	0.78

, Figure 4); the second is the scoring effect evaluation for the restoration effect and its various subitems (

Table 3. Model evaluation results of visitors’ restoration effects.

Term	Precision	Recall	F1 Score
1	1.00	0.50	0.67
2	0.80	0.80	0.80
3	0.60	0.75	0.67
4	0.86	1.00	0.92
5	1.00	0.83	0.91
accuracy			0.83
macro avg	0.85	0.78	0.79
weighted avg	0.85	0.83	0.83

and

Table 4. Model evaluation results of the 12 subtopic items of visitors’ restoration effects.

Term		Precision	Recall	F1 Score	Term		Precision	Recall	F1 Score
Q1 The material arrangements here are well organized	1	0.00	0.00	0.00	Q7 This place is different from where I usually live	1	1.00	0.50	0.67
	2	0.75	0.75	0.75		2	0.80	0.67	0.73
	3	1.00	0.60	0.75		3	0.40	0.50	0.44
	4	0.62	1.00	0.77		4	0.80	0.80	0.80
	5	1.00	0.86	0.92		5	0.83	1.00	0.91
	accuracy			0.77		accuracy			0.73
	macro avg	0.68	0.64	0.64		macro avg	0.77	0.69	0.71
	weighted avg	0.82	0.77	0.78		weighted avg	0.75	0.73	0.73
Q2 Everything here matches the overall environment	1	1.00	0.50	0.67	Q8 This place makes me feel rested	1	1.00	1.00	1.00
	2	0.80	0.80	0.80		2	1.00	0.75	0.86
	3	0.60	0.75	0.67		3	0.83	0.71	0.77
	4	0.86	1.00	0.92		4	0.83	1.00	0.91
	5	1.00	0.83	0.91		5	0.80	1.00	0.89
	accuracy			0.83		accuracy			0.86
	macro avg	0.85	0.78	0.79		macro avg	0.89	0.89	0.88
	weighted avg	0.85	0.83	0.83		weighted avg	0.87	0.86	0.86
Q3 This place seems infinite and allows exploration	1	1.00	0.83	0.91	Q9 This place is a haven for me	1	1.00	1.00	1.00
	2	0.50	0.50	0.50		2	0.80	0.80	0.80
	3	0.50	0.50	0.50		3	0.83	0.71	0.77
	4	0.75	1.00	0.86		4	0.60	1.00	0.75
	5	1.00	1.00	1.00		5	0.67	0.50	0.57
	accuracy			0.77		accuracy			0.77
	macro avg	0.75	0.77	0.75		macro avg	0.78	0.80	0.78
	weighted avg	0.78	0.77	0.77		weighted avg	0.79	0.77	0.77
Q4 This place makes me curious about things	1	0.00	0.00	0.00	Q10 This place suits my personality	1	1.00	0.33	0.50
	2	0.67	0.67	0.67		2	0.75	0.60	0.67
	3	0.86	0.86	0.86		3	0.50	0.75	0.60
	4	1.00	0.88	0.93		4	0.83	0.83	0.83
	5	1.00	1.00	1.00		5	0.80	1.00	0.89
	accuracy			0.86		accuracy			0.73
	macro avg	0.70	0.68	0.69		macro avg	0.78	0.70	0.70
	weighted avg	0.91	0.86	0.88		weighted avg	0.77	0.73	0.72
Q5 This place has a lot of interesting things that catch my attention	1	0.00	0.00	0.00	Q11 There are few boundaries here that restrict my movement	1	1.00	1.00	1.00
	2	1.00	1.00	1.00		2	0.80	1.00	0.89
	3	0.88	0.88	0.88		3	1.00	0.67	0.80
	4	1.00	0.75	0.86		4	0.80	1.00	0.89
	5	0.75	1.00	0.86		5	0.80	1.00	0.89
	accuracy			0.86		accuracy			0.86
	macro avg	0.72	0.72	0.72		macro avg	0.88	0.93	0.89
	weighted avg	0.92	0.86	0.88		weighted avg	0.89	0.86	0.86
Q6 I am mesmerized by this place	1	1.00	0.67	0.80	Q12 This place meets my expectations	1	1.00	1.00	1.00
	2	0.50	0.75	0.60		2	0.83	0.83	0.83
	3	0.67	0.57	0.62		3	0.67	0.67	0.67
	4	1.00	0.80	0.89		4	0.60	1.00	0.75
	5	0.75	1.00	0.86		5	0.75	0.50	0.60
	accuracy			0.73		accuracy			0.73
	macro avg	0.78	0.76	0.75		macro avg	0.77	0.80	0.77
	weighted avg	0.77	0.73	0.73		weighted avg	0.74	0.73	0.72

, Figure 4). In this study, the model performance was mainly evaluated using the three indicators of Precision, Recall, and F1 Score. Overall, we paid more attention to Precision than Recall; this is because we hope that the model classifies the travelogue text correctly according to the senses or scores it correctly according to a scale of 1–5 better than it omits the related travelogue text, in this way ensuring the accuracy of the model’s output results.

For the effect of sensory classification, the overall Precision reaches 77%, and the macro average and weight average are 72% and 78%, respectively. Among them, in terms of Precision, sight, smell, and hearing had higher classification Precision, and in terms of F1 Score, sight and hearing had the best classification effect, especially sight perception, which exhibits minimal classification bias and performs with an F1 Score of 90%. The following in classification performance are smell and touch. Taste, however, exhibits the poorest score due to the intermingling of smell perception, such as food aroma and smell, and taste perception, such as food flavor and taste, during the classification process.

Regarding the classification effect of the visitors’ restoration effects, the overall accuracy reaches 83%, and the macro average and weight average are 83% and 79%, respectively. Among them, from the perspective of Precision, the classification Precision scores of 5, 1, and 4 are higher. From the F1 Score, the evaluation results of 5 and 4 are better, reaching, respectively, 91% and 92%, and 1 and 3 points are more unstable. Judging from the 12 subitems, the overall accuracy rate is between 73% and 86%. In terms of the scoring results of each item, on the whole, the results with scores of 5, 4, and 1 are more stable and have a higher Precision, while the results with scores of 3 and 2 fluctuate more. Regarding specific sub-topics, those that are more associated with tourists’ personal feelings or experiences, such as “I am mesmerized by this place”, “This place suits my personality”, and “This place meets my expectations”, generally produce less stable results. However, topics that are more precise or closely related to physical space tend to have better accuracy and F1 Score. This highlights differences in the impact of model training on subjective questions versus objective descriptions and implies that the model’s outputs could be influenced by subjective elements within the training data. Nonetheless, it is important to acknowledge that in practical evaluations, individual variation in perceptions of emotions and experiences can also contribute to fluctuations in evaluation outcomes. Therefore, it is crucial to include the impact of this type of volatility on the results of the study at the data level by expanding the sample size or utilizing big data as much as possible.

5.2. Regression Analysis and Hypothesis Test

The regression analysis (

Table 5. Hypothesis Test.

Hypothesis	Coef	p	Hypothesis Result
H1a: Sight sensory number → Visitors’ restoration effects	5.66	0.000 ***	True
H1b: Hearing sensory number → Visitors’ restoration effects	0.712	0.018 **	True
H1c: Smell sensory number → Visitors’ restoration effects	−1.297	0.004 ***	False
H1d: Touch sensory number → Visitors’ restoration effects	0.384	0.001 ***	True
H1e: Taste sensory number → Visitors’ restoration effects	2.182	0.000 ***	True
H2a: Proximal psychological distance sensory → Visitors’ restoration effects	1.764	0.000 ***	True
H2b: Medial psychological distance sensory → Visitors’ restoration effects	−1.558	0.281	False
H2c: Distal psychological distance sensory → Visitors’ restoration effects	−7.165	0.000 ***	True
H3: Sensory dimension → Visitors’ restoration effects	0.119	0.058 *	True

*** p < 0.01; ** p < 0.05; * p < 0.1.

) clearly shows that the quantity of different senses, except in the case of smell, has a significant positive correlation with the visitors’ restoration effects. H1a, H1b, H1d, and H1e are confirmed, but H1c is not confirmed. Specifically, the quantity of the senses of sight, taste, hearing, and touch all had significant positive effects on the visitors’ restoration effects, with regression coefficients of 5.66, 2.182, 0.712, and 0.384, respectively. Of all sensory perceptions, sight tends to dominate as the primary mode of acquiring information. Correspondingly, in the National Forest Park, sight had the most significant impact on visitors’ restoration effects. The sensitivity of sight perception is further enhanced by the characteristics of various types of recreational behaviors and activities in the National Forest Park. The sense of touch has the least influence on the visitors’ restoration effects, which is due to the fact that compared with the touch experience, National Forest Parks tend to be more visually and auditorily stimulating; in addition, the sense of touch perception requires more initiative than other sensory perceptions, and it often requires people to actively touch the trees, leaves, bodies of water, or other landscapes, flora and fauna in order to obtain it, which leads to the fact that the sense of touch experience in the forests is usually relatively more consistent and smoother. The sense of smell plays a significant negative influence on the visitors’ restoration effects, indicating that the smell perception in the National Forest Park can inhibit the visitors’ restoration effects to a certain extent. For example, air pollution, environmental hygiene problems, and some irritating odors can bring about adverse sensory perceptions and directly inhibit visitors’ restoration effects. The following semantic analysis further proved this point.

Among the sensory psychological distances, both proximal and distal psychological distances were significantly correlated with the visitors’ restoration effects, and H2a and H2c were both confirmed. Specifically, the regression coefficient of proximal psychological distance is 1.7648, which has a significant positive impact on the visitors’ restoration effects, and the regression coefficient of distal psychological distance is −7.165, which has a significant negative impact on the visitors’ restoration effects. Medial psychological distance has a negative impact on visitors’ restoration effects, but the impact is not significant, and hypothesis H2b was not confirmed. This may be due to the fact that as the subjective psychological distance increases, the external stimuli and disturbances perceived by visitors’ senses increase, leading to information processing overload and making it more difficult to bring about positive visitors’ restoration effects.

Sensory dimensions had a significant positive effect on the visitors’ restoration effects, and H3 was confirmed. Compared with single-dimensional senses, multidimensional sensory perception increases the sensory richness and hierarchy of the recreational experience in the National Forest Park, synthesizes the advantages of sensory perception in terms of quantity and psychological distance, and has a beneficial impact on visitors’ restoration effects.

5.3. Semantic Analysis

Use Chinese word segmentation, part-of-speech tagging, and dependency parsing in natural language processing to identify sensory elements and conduct further semantic analysis.

First, Chinese word segmentation is used to segment the target sentence output by the large language model into word sequences, forming a basic module for natural language processing analysis. The original sentences are segmented through sequence annotation to obtain word segmentation results, as shown in Figure 5. In general, there are three sequential positions of a character in a word: beginning of word (B), end of word (E), and single word (S) [60].

Secondly, part-of-speech tagging is utilized to analyze and annotate the results obtained from word segmentation. According to the semantic characteristics of the sensory elements, terms other than the part-of-speech “NN” are filtered out (Figure 6).

Finally, dependency parsing is implemented to recognize and mark the dominant and subordinate connections between each expression (Figure 7). Through interpreting the syntactic relations labeled by the dependency arcs of sensory elements and their related semantic layers, we filter the semantic layers with the dependency relations “compound:nn”, “amod”, “nmod“, and at the same time, filter the lexically labeled “NN” entries in the semantic layer of sensory elements, thus completing the recognition and extraction of sensory elements.

The analysis considers the percentage of each type of element, visualizing and analyzing the top 10 elements. As shown in Figure 8, among the sight sensory elements, natural landscape elements and architectural elements such as vegetation (13.55%), water (10.49%), mountains (10.23%), and ancient architecture (3.07%) ranked highly, which is a direct reflection of the sight sensory level of China’s National Forest Parks, combining ecological, protective, recreational and leisure functions. Forest vegetation is the primary focus of visitors to National Forest Parks, with attention given to the color, form, and seasonal variations of various types of forests and trees. These elements combine to create a unique sight experience in the form of group or combined landscapes.

Among the hearing sensory elements, elements such as water sound (14.14%), animal sound (10.99%), wind (8.38%), vehicle sound (6.28%), and footstep sounds (5.76%) are ranked highly. Forested landscapes are often closely associated with water features such as lakes, streams, and rivers. The sounds of “gurgling water”, “singing birds” and “wind blowing leaves” are typical hearing sensory perceptions of forest restoration. In addition to natural sound elements, various human activities in forest parks also become the main source of hearing experience, including the sound of traffic, footsteps of visitors, etc. It is important to point out that these sounds do not always have a positive effect on the visitors’ restoration effects.

Among the smell sensory elements, plants (18.18%) and food (12.99%) accounted for a relatively high proportion; the fresh smell of plants, the aroma of food, and fresh air are all smell sensory perceptions with a positive effect, but there has to be concern about the large presence of negative experiences in the smell senses, such as the relatively high proportion of odors relating to animals (9.09%), car exhaust (7.79%), garbage (5.19%), and air pollution (5.19%), odors in transportation (3.90%), and water odors (1.30%), etc. Rotting plants, animal excrement, and mold smell brought by the humid environment in the forest environment, as well as garbage and pollutants produced by human activities, are all negative factors that cannot be ignored in the sensory perception of the forest, and the negative impact of this type of negative sensory perception on the visitors’ restoration effects is also obvious.

Taste sensations are well-represented among sensory categories in some of China’s National Forest Parks, which also function as tourist attractions where food and beverage activities are an integral part of forest recreation. Delicious or novel dining experiences are an important component of the taste senses and positively influence visitors’ restoration effects. Meat (22.35%); grains, beans, and legumes (12.94%); vegetables (11.76%); and refreshing mouthfeel (3.53%) emerged as highly ranked elements of the taste experience.

The top-ranked touch sensory elements can be categorized into two groups. One is contact with the natural forest environment, such as contact with plants through touch (13.25%), grass through touch (8.43%), water flow through touch (7.23%), and wind contact (7.23%), especially contact with grassland, woodland, soft land, and contact with rivers and streams in the forest; the second comprises cold and warm touch (8.43%) and dry and wet touch (6.02%) related to the weather environment. As a typical proximal psychological distance sense, touch is one of the main sensory ways for tourists to have direct contact with the forest environment. It is more likely to cause immediate physical and emotional responses and plays a positive role in the visitors’ restoration effects.

After the identification of sensory elements, sensory categories are formed after merging and thematic refinement to identify the intrinsic connections between the elements at the conceptual and content levels. This process aligns with the construction process of constructivist theory. Through multiple rounds of coding operations, all sensory elements with similar meanings are aggregated into sensory categories until all sensory elements are coded. Finally, a spectrum of sensory elements consisting of five sensory categories, 21 subcategories, and a number of single elements was formed (Figure 9). Among them, the sight sensory category includes five subcategories of natural landscape, architecture, climate, facilities, and people; the hearing sensory category includes four subcategories of climate sounds, plant and animal sounds, human voice, and human activity sounds; the smell sensory category includes five subcategories of food, natural smells, environmental odors, irritating odors, and others; the touch sensory category includes four subcategories of skin irritation sensation, landscape touch, facility touch, and nature touch; and the taste sensory category includes three subcategories of vegetative flavors, food flavors, and mouthfeel.

6. Discussion

This study provides possible ideas and methods for generating scale questionnaires through travelogue texts to solve the dilemma of self-reported scale measurement to a certain extent through the application of a generative large language model. In terms of data quality, it improves the problems of self-reported scale measurement in data reliability and consistency, subjects’ individual preferences, memory bias, and the influence of social expectations to a certain extent, and better characterizes the behaviors of recreationists and tourists in the real environment of National Forest Parks. In terms of data scale, it covers a wider range of research subjects and research time and space, providing a more convenient database for conducting large-scale, cross-time, and multi-group research. The results support the benefits of the extensive language model in both data generation and analysis, consistent with Shen et al.’s research [19].

In addition, our research also broadens the application of large language models in forest parks, recreation, restorative experiences, and other related fields. Specifically, the overall accuracy of the model training results for the senses reached 77%, and the overall accuracy of the model training results for the visitors’ restoration effects reached 83%. The model performance is relatively ideal. Optimizing prompts has a significant impact on training results, which is consistent with Tiffany’s research conclusion [61]. By trying models such as GPT-3.5 and text-davinci, it is shown that different models have an impact on the training results to a certain extent [62]. The impact of model parameters is not as significant as in the research of Kristy, Rodolfo, and others; temperature, frequency penalty, and presence penalty exert a more evident effect on training results. In addition, the limitations of large language models such as bias, discrimination, and AI hallucinations have been mentioned frequently, but they are not notably significant in our study.

There was a significant positive effect of the quantity of each sense, except smell, on visitors’ restoration effects, with sight having the greatest impact, followed by taste, hearing, and touch, respectively. Sight is the most dominant sense in human perception and understanding of the environment. Some studies have shown that more than 80% of the information received by human beings is acquired through sight [63], which is also confirmed in our study, and the number of sight senses is at the top of the number of all senses. In National Forest Parks, the development of various recreational activities, such as viewing nature and recognizing attractions and trails, relies heavily on the sight senses, and a variety of visual perceptions enriches the forest recreational experience, which in turn positively affects the visitors’ restoration effects. The second is the sense of taste. The significance of taste experience in forest recreation tends to be undervalued, which is also a common issue in previous studies regarding recreation and tourism. The rise of gastronomic tourism has brought forth renewed interest in taste among scholars of tourism and recreation [64]. Furthermore, a substantial number of China’s National Forest Parks double as tourist attractions, thereby elevating the value of taste experiences within these parks. Additionally, the role of taste in physical and mental rejuvenation is becoming an increasingly vital area of study [65]. Tasting food, experiencing various flavors and cultures, and sharing meals with other travelers to enhance social interaction all positively impact visitors’ restorative effects. The third is the sense of hearing, which is one of the important senses for participating in recreational activities in National Forest Parks; natural sounds in the forest such as birdsong, running water, wind, and the sound of falling leaves all have a positive effect on the restoration effect, which is consistent with the findings of Deng’s study [66]. However, our study also found that sounds of human activity, such as footsteps on leaves, conversations, and music, were also crucial to promoting positive effects of restoration among visitors. These hearing experiences help to enhance emotional bonds with nature and people, leading to improved social connections and contributing to physical and mental recovery. Fourthly, touch offers microscopic and direct sensory perceptions, such as touching the bark of a tree, walking on grass, stepping into a stream, and interacting with animals in the forest. It is considered one of the most effective ways to enhance one’s connection with nature. Several studies have demonstrated that touching natural elements can reduce heart rate, blood pressure, stress and anxiety [67], findings which are supported by our study.

From the perspective of the direction of influence, our study found that the sense of smell plays a significant negative role in the visitors’ restoration effects. This is a point that is often overlooked in past research [68]. Some studies have shown that forests do not always play a positive role in the impact of visitors’ restoration effects. This was confirmed in our research on China’s National Forest Parks. The smell of decaying leaves, animal excrement, environmental garbage, and car exhaust in the forest have all become widely mentioned adverse smell experiences. We believe that the significant negative impact of the sense of smell is mainly influenced by two aspects: one is the characteristics of the sense of smell itself. The sense of smell is the only sense that has been shown to be directly related to memory, and memory is emotional [69], making it easier for such negative sensory perceptions to leave a deep impression on tourists and users, thereby affecting the recovery effect. At the same time, smell is generally considered to be one of the senses closely linked to emotional processing [70]. As emotions have a higher priority in cognitive processes, when the sense of smell produces negative stimuli, people may be more susceptible to its influence, which in turn affects the overall assessment. In addition, in the long process of human evolution, the sense of smell has always been one of the senses closely linked to survival [71]. Humans are usually very sensitive to odors. Therefore, rotting plants, animal excrement, and humid environments in the forest bring mold odors, and odors such as garbage and pollutants generated by human activities are more likely to produce negative recreational effects. The second is the characteristics of the study area. A considerable number of the National Forest Parks in China, where this article is studied, have the functions of tourism, recreation, and ecological protection. Therefore, in such National Forest Parks, the number of tourists is larger, and recreational activities are more frequent and diverse. In addition, the concentrated travel brought about by China’s holiday system results in a considerable number of National Forest Parks having a large number of people and a high density at the same time, which is certain to lead to more negative sensory experiences, such as garbage and pollutants generated by human activities, which in turn have negative recreational effects.

Sensory psychological distance partially had a significant effect on the visitors’ restoration effects. Both proximal psychological distance and distal psychological distance have a significant impact on the visitors’ restoration effects, while medial psychological distance has a negative impact on the visitors’ restoration effects, but the impact is not significant. Recreational activities in National Forest Parks are typical relaxing tourism activities [72]. These activities align with the proximal psychological distance senses, resulting in significant positive restorative effects for visitors. In contrast, the distal psychological distance senses have a significant negative impact. The effects of the medial psychological distance senses fall between the other two but are not significant. Moreover, proximal psychological distance senses, such as taste and touch, often align with low interpretation levels [36], are more vivid and concrete, and are more conducive to attracting tourists’ attention, improving concentration, and thus promoting visitors’ restoration effects. At the same time, proximal psychological distance senses are more likely to produce emotional connections so that travelers can experience more emotional closeness, which means that they may be more likely to produce positive emotions conducive to visitors’ restoration effects, such as contentment and calmness. In contrast to the significant positive effects of proximal psychological distance senses on visitors’ restoration effects, distal psychological distance senses are not compatible with relaxing tourism activities such as forest recreation, and more importantly, distal psychological distance senses such as sight and hearing are consistent with high levels of interpretation [36], which tend to be associated with complex and deeper meanings, leading to increased cognitive load. In addition, it was found that the sight and hearing senses, which tend to be subject to passive external stimulation and interference, have a more pronounced negative impact on recovery. Therefore, although from the perspective of the quantity of senses, an increase in the number of sight, hearing, and other types of senses enhances positive visitors’ restoration effects, from the perspective of psychological distance, the senses at the proximal end of the psychological distance are more likely to positively influence visitors’ restoration effects, in comparison.

Sensory dimensions have a significant positive effect on visitors’ restoration effects. Our study of China’s National Forest Parks further supports the notion that multidimensional sensory engagement aligns with human cognition of the environment and surrounding objects [73,74]. Expanding sensory dimensions, rather than the mere addition of individual senses, amplifies the integration and synergistic effects of sensory inputs. This enhances the depth and structure of the forest recreation experience. This has been described by visitors, who reported: “Climbing Yushan Mountain, enjoying the red maple, and drinking the refreshing Yushan green tea. When the maple leaves fall on the ground, step on the creaking sound, everything is so perfect”. At the same time, multidimensional senses also play a complementary effect, and the stimulation of different senses can complement each other to provide a more comprehensive forest recreation experience, as one tourist said: “Canyons in the mist, falling waterfalls, the sound of cascading water in our ears, nature has fully endowed us with the most beautiful scenery!” Furthermore, our research has revealed that interactions between humans and nature through multidimensional senses can also promote interpersonal interactions, leading to cross-perceptual effects. For instance, in the context of forest environments as opposed to urban environments, human conversations and footsteps cease to be a source of jarring noise and instead emerge as a means to augment social connectivity and social integration. This, in turn, fosters correlated positive effects on visitors’ restorative experiences in relation to social health.

The sensory elements of National Forest Parks that affect the restoration effect have predominantly natural attributes while revealing the interaction between people and people, and between people and nature. First, natural attributes are predominant, which is basically consistent with most of the existing studies that take the forest environment as the observation object. Natural elements are involved to varying degrees in the five categories of senses, among which, the vast majority of sight, hearing, and touch are natural sensory elements, including plants, animals, mountains, water, and natural climate. These elements are closely related to the environmental features found in the National Forest Park, which is consistent with prior research on forest restoration [75]. Second, sensory elements reveal the interaction between people and between people and nature. In addition to traditional natural sensory elements, our research also observed that elements related to “people” also play a vital role in the impact of visitors’ restoration effects, such as architecture and facilities in sight sensory elements, and people in auditory sensory elements. For instance, architectural designs and amenities impact the sight sensory experience while the sound of conversations and footsteps, as well as the rustling of leaves, affect the hearing sensation. The smell and taste senses are stimulated by the aroma and flavor of food whereas the touch sensory experience involves physical interactions with people and surroundings. On the one hand, this is related to the fact that China’s National Forest Parks have the attributes of both nature reserves and tourist attractions, especially for some of the National Forest Parks that are A-level scenic spots, and that more frequent human-to-human and human-to-nature interactions constitute an important part of the impact on visitors’ restoration effects. On the other hand, it is imperative to emphasize the significance of forest areas in encouraging social interaction, fostering emotional connections, and promoting interpersonal contentment. This is as crucial as maintaining physical and mental well-being and constitutes an essential factor in achieving positive restoration effects for visitors. Empirical studies on urban green spaces like city parks and urban green fields have verified this proposition [76,77,78].

7. Conclusions

7.1. Research Conclusions

Based on the attention restoration theory and stress recovery theory, this paper explores the influence of sensory perception of the forest on visitors’ restoration effects from a multidimensional and multisensory perspective, comprehensively utilizing methods such as the generative large language model, regression analysis, and semantic analysis. The main conclusions are as follows:

First, by introducing the generative large language model, this study expands its application in fields such as tourism and recreation and provides new ideas and methods for solving the dilemma posed by the traditional self-reported scale measurement. Moreover, it potentially paves the way for exploring novel research paradigms in light of the burgeoning development of generative artificial intelligence. In this study, we aim to use UGC data to extract independent variables such as sensory quantity, sensory psychological distance, and sensory dimension, alongside the dependent variable of visitors’ restoration effects, via the use of a generative large language model. Evaluation of the model’s output demonstrates a high level of feasibility for this method.

Second, in the study of the quantity of senses and visitors’ restoration effects, compared to previous studies that only examined some specific senses and less often compared the differential effects of the five categories of senses with different dimensions, our study found that the effects of the quantity of the five categories of senses differed, with the quantity of sight, hearing, touch, and taste senses having a significant positive effect on the visitors’ restoration effects, and the sense of smell having a significant negative effect. This is due to the fact that smell is closely related to memory, directly affects the emotional center, and is closely related to survival instincts, as well as the fact that the study area of this paper, the National Forest Parks of China, combines the functions of tourism, recreation, and ecological protection. The results of this study provide new insights that are different from most previous studies and give us more inspiration and space to further explore the effects of smell perception on forest visitors’ restoration effects.

Third, the sensory psychological distance and visitors’ restoration effects in the study allowed us to step out of the perspective of simply examining the increase in the quantity of senses, and to reconceptualize the senses in terms of the level of interpretation and psychological distance, which proved to be a useful attempt. Our research shows that forest recreation and tourism activities with lower levels of interpretation are better suited to proximal psychological distance senses, resulting in better recovery. However, as psychological distance increases, the mismatch in levels of interpretation also increases, causing the visitors’ restoration effects to gradually shift from positive to negative. Therefore, it is important to consider psychological distance when designing nature-based activities for tourists. Indeed, as mental distance increases, external stimuli and distractions to the sight and hearing senses also increase, and to some extent, cognitive load is likely to increase. Although an increase in the quantity of a single sense, including sight and hearing, has a significant positive effect on restoration in most cases, whether individuals still use involuntary attention as referred to in the attention restoration theory at this time, and whether involuntary attention and voluntary attention are mutually exclusive, are worthy of further exploration [20]. Thus, some studies have suggested that restoration of voluntary attention does not necessarily require the use of involuntary attention and that there may be other mechanisms, such as exposure to natural environments that can activate an individual’s positive emotions, which can have a positive effect on the restoration of voluntary attention [20].

Fourth, the sensory dimension has a significant positive impact on the visitors’ restoration effects. The expansion of sensory dimensions heightens the amalgamation and synchronicity of the senses, offering a supplementary impact that enriches the depth and hierarchy of the forest recreational experience. Furthermore, the multidimensional senses advance human interaction with nature, producing a perceptual interaction that positively influences restoration outcomes concerning social health.

Fifth, through the semantic analysis method in natural language processing, we further identified the specific sensory elements that affect the visitors’ restoration effects and constructed a multidimensional and multi-level sensory element spectrum. The study results indicate that the visitors’ restoration effects in National Forest Parks are mainly influenced by the natural attributes of the sensory elements. Additionally, elements related to people also have a significant impact on the visitors’ restoration effects.

7.2. Research Limitations and Future Research

This study still has some limitations and aspects to be explored in depth in future research:

First, in terms of the use of generative large language models, the model training effect needs to be evaluated from multiple perspectives such as different model selections and different parameter settings. Especially in terms of model parameters, it is still necessary to conduct a comparative analysis of the training effects of different parameter settings based on the characteristics of the training objectives, training tasks, and training data [62]. The second aspect is to focus on the cross-modal interaction of sensory perception. For example, sight perception in the forest may be affected by hearing or touch-related features, there may also be a cross-modal interaction effect between senses with different psychological distances. Third, it is important to consider the characteristics of olfactory perception and how it affects the recovery of tourists. The function of odor and how it interacts with other senses can be assessed from three perspectives when examining forest tourists: smell has a strong link to memory, plays a direct role in affecting the emotional center, and has close ties to the instinct for survival. Fourth, the multisensory perception and restoration effects of forest tourism among tourists from different cultural and geographic backgrounds should be studied. Differences in the multisensory perception of forest tourists may arise due to varying cultural backgrounds, and cultural concepts may shape the therapeutic influence of forest surroundings. Cognitive variances may also impact people’s desire to seek psychological healing and relaxation in natural environments. Meanwhile, unique environmental conditions, like varying climates, landscapes, biodiversity, and regional cultures, influence not only the forest’s natural characteristics but also the tourists’ multisensory experiences within the forest. Fifth, the similarities and differences in the perception of the five senses stimuli based on the SNS data and the professional guidance from different points of view can be compared.