J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 
119-137                                                                                      

119 

 

 

 

Article 

Deep Learning models for cultural pattern 

recognition: preserving intangible heritage of Li 

ethnic subgroups through intelligent 

documentation systems 
Jing Sun, Kartini Aboo Talib Khalid, Chan Suet Kay*  

Institute of Ethnic Studies (KITA), Universiti Kebangsaan Malaysia, 43600 UKM Bangi, Selangor, Malaysia 

A R T I C L E   I N F O 
 

Article history: 
Received 12 April 2025  
Received in revised form 
23 May 2025 
Accepted 04 June 2025 
 
Keywords:  
Deep Learning, Intangible cultural heritage, 
Multimodal fusion, Cultural pattern recognition, 
Intelligent documentation systems 
 
*Corresponding author 
Email address: 
rachelchansuetkay@ukm.edu.my 
 
 
 
 
DOI: 10.55670/fpll.futech.4.3.12 

A B S T R A C T 
 

This study develops an advanced intelligent documentation system using deep 
learning models to preserve intangible cultural heritage for the Li ethnic 
minorities. Traditional heritage documentation models face significant 
obstacles in systematically capturing oral traditions and inter-group cultural 
differences. The proposed comprehensive multimodal fusion framework 
integrates visual pattern analysis through convolutional neural networks, 
temporal cultural depiction via bidirectional LSTM networks, and semantic 
comprehension using transformer-based models. Collaborative fieldwork 
across five Li subgroups (Ha, Qi, Run, Sai, and Meifu) in Hainan Province 
documented 4,450 cultural samples, including traditional textiles, music, oral 
traditions, artifacts, and architectural heritage. The five-layer distributed 
system architecture employs pattern recognition, semantic indexing, and 
recommendation algorithms for scalable cultural preservation. Experimental 
results demonstrate remarkable 94.8% accuracy across Li subgroups, 
significantly outperforming traditional single-modality systems (CNN: 85.3%, 
RNN: 87.6%, Transformer: 89.4%). System implementation yielded 
unprecedented improvements in cultural transmission effectiveness: 73% 
increase in knowledge retention, 121% in skill transfer, and 280% in digital 
archiving abilities. Community participation increased exponentially, with 
340% growth in active users and a 665% increase in monthly contributions. 
The system achieves robust operational performance with sub-200ms response 
times and 99.7% stability. User satisfaction and expert evaluation scores of 4.4 
and 4.6, respectively, confirm reliable cultural preservation functionality. This 
framework establishes advanced benchmarks for computational heritage 
preservation methods, demonstrating the effective integration of technological 
innovation with ethnographic sensitivity for the sustainable documentation 
and transmission of minority cultures. 

1. Introduction 

Intangible cultural heritage faces unprecedented 
challenges in the contemporary era of globalization, 
particularly for minority communities whose traditions are 
vulnerable to external pressures and rapid modernization 
processes [1]. The intersection of cultural heritage 
preservation, institutional frameworks, and community 
engagement represents a multidimensional domain that 
encompasses diverse stakeholder interests and complex 
power dynamics across various social groups [2]. Recent 
research has shown significant and mutual impacts of 
intangible cultural heritage and socioeconomic development, 

thereby illustrating the intricate relationships that exist and 
need to be navigated for the sustainability of heritage in 
contemporary contexts [3]. Nevertheless, critical questions 
persist regarding which aspects of participation and decision-
making processes constitute the core of heritage 
preservation, thereby highlighting persistent concerns about 
inequality and representation that continue to challenge 
contemporary conservation initiatives [4]. Traditional 
approaches to documenting and preserving cultural heritage 
face persistent issues in many historic and religious heritage 
sites, often due to the intricate dynamics involved in these 
cultures [5]. The rise of artificial intelligence along with digital 

May 2025| Volume 01 | Issue 01 | Pages 01- 03 

ISSN 2832-0379 

Open Access Journal 

Journal homepage: https://fupubco.com/futech 

 

https://doi.org/10.55670/fpll.futech.0.0.0 

 

 

Future Technology  

 

August 2025| Volume 04 | Issue 03 | Pages 119-137 

Journal homepage: https://fupubco.com/futech 

 

https://doi.org/10.55670/fpll.futech.4.3.12 

 

Future Technology 

Open Access Journal 

ISSN 2832-0379 

mailto:rachelchansuetkay@ukm.edu.my
https://fupubco.com/futech
https://doi.org/10.55670/fpll.futech.0.0.0
https://fupubco.com/futech
https://doi.org/10.55670/fpll.futech.4.3.12


J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

120 

 

technologies opened new avenues for the preservation of 
cultural heritage, thus offering enormous opportunities for 
research and scholarship [6]. Despite these advancements, 
however, a significant disparity persists between the 
potential offered by digital technologies and innovations and 
their actual application in heritage preservation strategies 
[7]. The preservation of intangible cultural heritage uses 
various digital systems, which require complex processes 
pertaining to user acceptance and design that involve 
extensive studies on technology adoption and participation 
[8]. Additionally, the existence of the digital divide within the 
context of preserving intangible cultural heritage poses grave 
challenges to the secure transmission and transfer of 
traditional knowledge systems [9]. The investigation 
addresses these issues by developing deep learning 
frameworks to identify the cultural patterns of the Li ethnic 
subgroup, thereby developing an intelligent documentary 
system that merges traditional preservation techniques with 
modern technological sophistication. Such an initiative 
strengthens the theoretical extrapolation of cultural pattern 
analysis and its applications to heritage conservation, while 
also contributing to the development of a comprehensive 
paradigm for protecting the intangible heritage of 
marginalized ethnic groups through innovative 
computational methods. 

2. Literature review 

2.1 Digital protection of intangible cultural heritage 
The study of cultural preservation and practices has 

become an essential issue that needs to be safeguarded 
through technological intervention. Development of new 
approaches to studying culture emphasises the strategies 
formulated to preserve them beyond the traditional material 
forms [10]. This indicates a shift from conservation to modern 
techniques that preserve culture in action. The initiatives 
aimed at the digitisation of intangible assets have stood out, 
as scholars try to employ emerging ideas to solve the 
problems of preservation [11]. Such initiatives encompass all 
technological solutions, ranging from interactive multimedia 
documentation systems to collaborative knowledge transfer 
and community participation interfaces. The use of 
technology has enabled the documentation of practices that 
can no longer be captured easily, thus presenting cultural 
knowledge in sophisticated ways, which was not possible 
using traditional means. 

The development of sophisticated three-dimensional 
technologies has greatly revolutionized methods used in the 
preservation of intangible heritage, opening up novel 
possibilities for immersive exhibitions as well as recording 
[12]. A survey of 3D technologies in various databases reveals 
a range of methodological approaches, highlighting diversity 
in approach as well as the technical challenges associated 
with their application within heritage contexts. They enable 
the creation of vast digital archives preserving visual as well 
as auditory features that also include spatial as well as 
temporal dimensions of cultural practices. The modern 
information technology environment has greatly transformed 
methods of conservation and sharing of intangible cultural 
heritage, therefore creating new global accessibility as well as 
community engagement opportunities [13]. Though digital 
media facilitated a more democratic availability of 
information about culture, they also face limitations in terms 
of authenticity, representation, and communal ownership. 
Despite advances in technology, significant gaps remain in the 
integral management of intangible heritage content, 
particularly in terms of the transmission of embodied tacit 

knowledge and experiential know-how, which is inherently 
difficult to translate into digital media. 

Current research confirms an increased recognition of 
the need to employ interdisciplinary approaches that merge 
technological development into anthropological insight and 
engagement of local populations. Although digital recording 
equipment is central to the documentation and preservation 
of cultural heritage, researchers are increasingly recognizing 
that authentic conservation of heritage requires close 
attention to cultural context, ethical principles, and 
community perspectives. This approach ensures that 
digitization is guided by the needs of heritage populations, 
rather than technological development alone. 

2.2 Deep Learning applications in cultural heritage 
The intersection of machine learning methods and 

studies of cultural heritage is an important development that 
brings forth innovative solutions to preservation, analysis, 
and interpretation problems that conventional methods do 
not fully address [14]. Recent examples illustrate 
considerable improvement in the development of artificial 
intelligence systems designed explicitly to support innovation 
in heritage environments, which in turn encourages 
widespread research and development activities that 
synchronize technological innovations with the imperatives 
of heritage preservation [15]. Such advances demonstrate an 
underlying inclination to employ computational approaches 
to address complex problems of heritage conservation while 
upholding academic integrity and cultural sensitivity. The 
methodical recording of technological advances in heritage 
conservation has been conducted using bibliometric analysis, 
which shows mounting integration of novel technologies and 
their role in enhancing conservation methods [16]. In 
particular, advances in computer vision have revolutionized 
the field of cultural image recognition, enabling automatic 
identification and cataloging of aesthetic features, 
architectural elements, and motifs that were hitherto 
determined using visual examination. Such systems find 
exceptional effectiveness in handling large sets of image data, 
thereby enabling thorough analysis of cultural artifacts and 
monuments while also requiring fewer resources and time 
compared to traditional recording methods. Advanced deep 
learning architectures, particularly convolutional neural 
networks (CNNs) with attention mechanisms, have 
demonstrated remarkable capability in extracting 
hierarchical features from cultural artifacts. These systems 
employ spatial attention modules that focus on culturally 
significant regions within images, while channel attention 
mechanisms prioritize feature maps that capture distinctive 
cultural characteristics, thereby enhancing the accuracy of 
pattern recognition in heritage documentation. 

Multimodal learning approaches have drawn 
considerable interest in the field of cultural heritage, where 
there is considerable scope for combining different data 
modalities and analytical perspectives [17]. Cross-modal 
attention mechanisms enable intelligent integration of visual, 
textual, and auditory cultural data through learnable weights 
that establish semantic correspondence between 
heterogeneous modalities [18]. The framework incorporates 
modality-specific encoders for feature extraction, cross-
modal alignment modules for shared semantic learning 
through contrastive strategies, and adaptive fusion 
mechanisms that dynamically weight contributions based on 
cultural context and data quality. Cultural heritage 
applications require specialized adaptations to preserve 
authenticity and maintain the integrity of heritage [19]. The 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

121 

 

mechanism utilizes culture-specific parameters to preserve 
traditional semantic relationships and incorporates temporal 
components to capture sequential cultural performances. 
Multi-head architectures process diverse cultural dimensions 
simultaneously, with specialized attention heads focusing on 
symbolic elements, ceremonial sequences, and linguistic 
patterns. Natural language processing approaches 
revolutionized the study of historical texts, literary works, 
and oral traditions, enabling researchers to obtain semantic 
models and cultural narratives from large sets of texts. 
Immersive technologies are a suitable example of novel 
applications in virtual preservation environments for 
intangible heritage, enabling the recording and sharing of 
traditional craftsmanship knowledge [20]. Specific 
frameworks created for recording traditional skills 
incorporated innovative methods for safeguarding embodied 
cultural knowledge, most importantly using ego-centered 
recording systems that include practitioners' perspectives 
and methods [21]. The use of ontology systems, which 
organise and relate heritage materials for deep search and 
retrieval, has greatly improved the organisation of cultural 
histories in digital repositories [22]. Such progress indicates 
that there is a movement away from attempts at digitisation 
towards more sophisticated systems capable of intelligently 
interpreting and preserving the complex nature of cultural 
heritage. The challenge of designing deep learning systems 
that respect cultural values, community perspectives, and 
technological objectives within culturally sensitive 
frameworks remains paramount. Current research trends 
demonstrate a convergence of computational capabilities and 
anthropological insights, ensuring that technological 
development serves the broader objectives of cultural 
preservation and enhanced accessibility while maintaining 
cultural authenticity and community agency. 

2.3 Advances in intelligent documentation systems 
Intelligent document systems represent an evolutionary 

leap within cultural heritage management, capable of 
successfully meeting the main issues while presenting novel 
solutions to support future development [23]. They combine 
advanced computational technology and established methods 
of heritage preservation to create thorough systems that 
enhance not only the effectiveness but also the efficiency of 
cultural documentation practices. Digital systems that 
specialise in cultural heritage document an increase in 
acceptance of an enhanced technological platform, which is 
able to process the associated diversity and complexity of 
cultural data. Recent assessments of digital cultural heritage 
technologies reveal significant improvement in building 
cohesive solutions that address multiple aspects of 
documentation, preservation, and communication [24]. 
Knowledge graphs emerged as useful tools for cultural 
heritage management, which enable building interconnected 
semantic graphs that describe the inter-relationships of 
cultural artifacts, their historical context, and their current 
meanings. Graph theory-inspired approaches deepen the 
understanding of cultural inter-relationships and enable 
scholars to reveal hitherto unknown patterns in heritage 
materials. 

Digital methods of cultural heritage documentation and 
preservation have undergone major transformations, most 
notably in data acquisition, processing, as well as 
visualization methodologies [25]. Traditional cataloging and 
recording of cultural objects, text documents, and media 
materials have been transformed by using advanced 
classification and annotating systems that utilize machine-

learning algorithms for automated item classification and 
identification. The systems proved to be quite effective at 
handling large heritage datasets, substantially reducing labor 
needs while promoting increased accuracy as well as 
consistency in metadata creation. 

Virtual reality technologies introduced novel approaches 
to the preservation of traditional craftsmanship, 
demonstrating the potential of immersive technologies to 
protect and pass on tacit knowledge underlying cultural 
practices [26]. Cross-cultural information retrieval systems 
have made considerable advances, adding sophisticated 
multilingual processing features and an awareness of cultural 
context to refine search results accuracy and relevance. 
Automated data extraction and structuring technologies 
enabled the digitization of cultural holdings, easing the 
process of converting analogue materials to accessible digital 
representations. Recommender systems have been 
remarkably useful in engaging users and facilitating the 
navigation of digital collections in the cultural heritage 
context. Such systems can recommend pertinent materials by 
synthesising user actions with cultural metadata, and in so 
doing, they transcend conventional interactions with heritage 
materials. However, one of the major challenges is developing 
systems that effectively address the interpretive and 
subjective frames of meaning that accompany cultural 
heritage. This underscores the need to study computing 
systems more closely in ways that respect the cultures and 
peoples involved, while still taking full advantage of what 
modern technology offers. 

2.4 Current Status of Li Ethnic Culture Research 
The cultural practices of the Li ethnic group scholars 

have been documented meticulously due to the ongoing 
efforts to outline the minority cultures of China. This reflects 
scholarly interest in safeguarding and documenting ancient 
cultures and practices across the world. Efforts undertaken 
by modern societies towards the preservation of indigenous 
cultures reveal dire threats faced by tribal people globally, 
and using advanced electronic devices serves as a means to 
counter these challenges, providing an alternative to 
conventional recording methods while improving access and 
distribution of information [27]. This aids in the preservation 
of such communities’ heritages, which is vital nowadays for 
groups that undergo rapid changes as a result of 
modernization. Investigations into China ’ s intangible 
cultural heritage have surfaced gaps associated with 
intellectual property frameworks and research focus areas 
[28], thereby underscoring the multifaceted issues within 
heritage conservation among minority ethnic groups. The Li 
ethnologic group constitutes one such minority nn China. 
Within this context, they are distinguished by the specific 
cultural attributes, language diversity, and social organisation 
traits that set them apart from other groups. Documentary 
accounts indicate that the Li society comprises diverse clans 
who possess various forms of cultural expression, and who, 
although culturally distinct, share family relations and spatial 
proximity within Hainan Province. 

The preservation of Li's intangible cultural heritage has 
been examined through a range of institutional frameworks, 
including one ethnic park which attempted to blend 
commercialisation with authentic representations of the 
culture [29]. Such ethnological and anthropological practices 
pose troubling issues concerning the extent to which the 
commodification of cultural practice succeeds alongside the 
efforts of preservation actions in maintaining cultural 
integrity. The fragile dynamics involved in advancing tourism 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

122 

 

while preserving heritage among the Li groups illustrate 
broader challenges confronting minority groups in China 
today. Recent efforts to enhance the intangible cultural 
heritage of the Li ethnic minority group through innovative 
approaches to tourism have created optimism, as well as 
issues of the commodification of culture [30]. The initiatives 
include various aspects of Li culture, such as handicrafts, oral 
traditions, ritual practices, and architectural styles, which 
require different approaches to documentation and 
preservation. New forms of cultural heritage-based tourism 
products have generated significant interest from 
researchers and practitioners who strive to develop 
sustainable models of cultural conservation that benefit local 
populations while preserving authenticity. Despite the rising 
scholarly interest, much effort is still needed to accomplish 
extensive documentation of Li cultural practices, most 
importantly, detailed analysis of intra-regional variation 
among Li groups, as well as methodical conservation of 
embodied knowledge. The combination of modern 
documentation technologies and traditional cultural 
transmission methods is an important opportunity for future 
research, which could bring effective solutions to related 
issues of cultural continuity and transmission of knowledge 
from one generation to another among the Li groups. 
2.5 Literature review synthesis 

Recent research suggests significant gaps in organized 
documentation of embodied cultural traditions and 
traditional crafting methods. Gesture analyses centered on 
gestures within practices highlight the need for sophisticated 
methodologies that can effectively condense fleeting aspects 
of traditional techniques, thus highlighting critical gaps in 
preservation that fail to address implicit knowledge and 
motor skills adequately [31]. Systematic analysis of the 
preservation technologies shows wide methodological 
diversity within heritage contexts [32]. Notwithstanding the 
availability of various advanced technologies, poor 
integration of sophisticated technological modalities hinders 
further development of comprehensive recording solutions 
for intangible heritage. Such diversity is a major hindrance to 
establishing effective preservation frameworks. 

The future possibilities of technological growth suggest a 
significant opportunity for intelligent systems to 
independently recognize and document cultural patterns. 
State-of-the-art machine learning techniques, particularly 
those that specialize in multimodal analysis, hold outstanding 
promise for building sophisticated document systems that 
capture both overt and implicit cultural expressions. Such 
technological advancements address time and context issues 
surrounding effective documentation of changing cultural 
practices very adequately. Theoretical contributions consist 
of conceptual models linking computational methods and 
anthropological theory regarding cultural transmission. The 
synthesis of deep learning approaches and cultural pattern 
recognition represents a new field that goes beyond 
traditional documentation practices, improving both 
technological expertise and theoretical understanding of the 
encoding and preservation of cultural knowledge. 

The analysis outcomes demonstrate tremendous 
opportunities for developing advanced innovative methods of 
preservation, which uphold cultural integrity while 
maintaining heritage computationally. The results serve as a 
basis for the proposed research framework for the 
preservation of Li ethnic culture through sophisticated 
documentation techniques. 

 
 

3. Data and methods 

3.1 Research design and hypotheses 
The research suggests a comprehensive framework 

designed for document-intelligent systems to construct deep 
learning models for recognizing the cultural patterns of the Li 
ethnic group. The framework combines the challenges posed 
by information technology methods and advanced 
computational methods to counter sophisticated challenges 
in documenting intangible heritage. It applies a system of 
hierarchically organised interrelated cultural patterns to 
formulate additional goals, including self-controlled identity, 
self-directed learning, and preservation [31]. The framework 
also contains a set of interlinked assumptions that together 
govern the research concerning autonomous cultural pattern 
identification and preservation. 

The approach taken in this study rests on a defined 
model centred on analysing the impact of deep learning 
technology on cultural heritage preservation. The study 
examines four interrelated research hypotheses that address 
multiple aspects of the proposed intelligent document 
system. Figure 1 shows that these hypotheses were designed 
to allow thorough validation of both technical properties and 
cultural integrity within the system. 

The research sets forth four related hypotheses that 
probe different dimensions of smart cultural documentation. 
Hypothesis H1 suggests that deep algorithms trained to 
specific features of the Li ethnic culture will exhibit 
substantially higher accuracy in pattern recognition than 
generic cultural heritage systems. Hypothesis H2 examines 
the efficiency of multimodal integration methods in cultural 
pattern recognition compared to systems that employ one 
modality. H3 investigates whether intelligent documentation 
systems can preserve cultural authenticity while enabling 
automated analysis with minimal semantic loss. H4 explores 
the system's capacity to effectively distinguish between 
different Li subgroup cultural patterns with statistically 
significant classification performance. These four interrelated 
hypotheses constitute a hierarchical validation framework: 
H1-H2 verify technical performance, H3 ensures cultural 
integrity, and H4 tests practical classification capabilities, 
collectively ensuring the balance between technological 
innovation and cultural preservation. 

These hypotheses are supported by specific research 
questions that guide the empirical investigation, ranging from 
technical optimization strategies to cultural authenticity 
preservation methods, as shown in Figure 1. The evaluation 
framework incorporates both quantitative metrics for 
technical validation and qualitative assessments for cultural 
fidelity verification. The expected outcomes encompass 
enhanced pattern recognition accuracy, improved 
documentation efficiency, preserved cultural authenticity, 
and sustainable knowledge transmission mechanisms that 
collectively contribute to the preservation of Li ethnic 
intangible heritage. Statistical significance testing employs 
α=0.05 standards, with more stringent α=0.01 thresholds for 
cultural authenticity assessments. To address multiple 
hypothesis testing, Bonferroni correction adjusts significance 
levels to α=0.0125, while the Benjamini-Hochberg procedure 
controls false discovery rates. 

 

 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

123 

 

 
3.2 Data Collection and Preprocessing 

The comprehensive data collection framework for Li 
ethnic cultural heritage encompasses systematic fieldwork 
methodologies combined with advanced digital 
documentation technologies to capture the diverse 
expressions of intangible cultural practices across Hainan 
Island (Table 1). Literature suggests that the effective 
conservation of cultural heritage requires an interdisciplinary 
approach that balances scholarly accuracy with community 
member participation and technological advances [33]. The 
data collection strategy gives precedence to ethnographic 
recording methods while incorporating advanced multimedia 
processing to achieve a holistic presentation of Li cultural 
elements. Stream preprocessing applies automated quality 
controls along with manual inspection on every data modality. 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
Each audio recording is denoised and spectrally 

normalised to ensure consistency across varying recording 
conditions. Video recordings are processed with temporal 
segmentation algorithms that detect certain cultural activities 
and gestures. Achieving visual homogeneity within datasets, 
high-resolution photographs are colour calibrated and 
processed. To obtain stratified samples within Li groups 
while achieving equitable distribution for machine learning, 
the hierarchy employs a clearly defined sampling protocol. 
This hierarchy reserves 70% of collected materials for 
training, 20% for validation, and 10% for a final test split, 
whilst maintaining adequate representation from each Li 
group across all splits. Quality control procedures combine 
computational validation methods and reviews by cultural 
specialists to ensure authenticity and accuracy in the data-
preparation process.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Central Research Goal
Effective Li Ethnic Cultural

Pattern Recognition System

H1:Technical Superiority
Deep learning models adapted

for Li cultural characteristics

achieve higher accuracy thar

generic systems

Validation: Accuracy, Precision

H2:Multimodal Advantage
Multimodal fusion approaches

outperform single-modality

systems in pattern

identification

Validation: F1-Score, Recall

H3:CulturaPreseryation
Intelligent documentation

maintains cultural authenticity

while enabling automated

analysis

Validation: Expert Assessment

H4:Subgroup Distinction
System effectively

distinguishes between

different Li subgroup

cultural patterns

Validation: Classification Rate

Supporting Research Questions
RQ1: How can deep learning architectures be optimized for Li cultural pattern recognition?

RQ2: What multimodal features best represent Li ethnic cultural expressions?

RQ3: How can automated systems maintain cultural integrity and authenticity?

RQ4: What distinguishing features exist between Li ethnic subgroups?

RQ5: How can intelligent documentation enhance cultural heritage preservation?

Expected Research Outcomes
·Enhanced Cultural Pattern Recognition Accuracy                     ·Preserved Cultural Authenticity

·lmproved Heritage Documentation Efficiency                          ·Sustainable Knowledge Transmission

Research Hypotheses Framework for Li Ethnic Cultural Pattern Recognition

 

Figure 1. Research hypotheses framework for Li ethnic cultural pattern recognition 

Table 1. Li ethnic cultural data collection and processing framework 

Data Category Collection Method Processing Protocol Format Specifications Quality Control 

Traditional Music High-resolution audio 
recording 

Noise reduction, 
normalization 

48 kHz WAV, spectral 
analysis 

Expert validation, cultural 
authenticity 

Craft Documentation Multi-angle video capture Segmentation, motion 
analysis 

4K MP4, frame extraction Practitioner verification 

Oral Traditions Structured ethnographic 
interviews 

Transcription, linguistic 
annotation 

Audio + text corpus Community elder approval 

Textiles & Artifacts 3D photogrammetry scanning Model reconstruction, 
texture mapping 

OBJ files, high-res textures Museum standard 
documentation 

Architectural Heritage LiDAR point cloud scanning Mesh generation, 
dimensional analysis 

PLY format, CAD models Architectural accuracy 
validation 

Ceremonial Practices Ethnographic observation Event segmentation, 
symbolic coding 

Multimedia annotations Ritual specialist 
consultation 

 

 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

124 

 

Annotation consistency across Li subgroups achieves 
robust inter-rater reliability with Cohen's Kappa values 
ranging from κ=0.82 to κ=0.91, while Fleiss' Kappa 
demonstrates strong multi-rater agreement (κ=0.87) across 
cultural specialists, confirming systematic annotation quality 
and cross-cultural validity. 

3.3 Deep Learning models 
The proposed framework covers an extensive range of 

deep learning architectures that were carefully designed to 
identify Li cultural heritage's intricate features. Recent 
studies highlight that multimodal emotion recognition 
systems require sophisticated computational methods that 
overcome the limitations of traditional unimodal systems 
[34]. This baseline framework utilizes convolutional neural 
networks (CNNs) to process visual aspects of culture, and 
residual connections to extract hierarchical features from 
textiles, buildings, and ritual objects. The feature extraction 
process follows the defined mathematical equation: 

1

1 1 1

( * ( *...*

( * ) ) )

visual n n

n n

F W W

W X b b b

 



−

−

=

+ + +
 (1) 

where 𝜎  represents the activation function, and Wi , bi 
denote the weight matrices and bias vectors, respectively. 

Temporal cultural expressions, including traditional 
music and oral narratives, are modeled through bidirectional 
long short-term memory (LSTM) networks that capture 
sequential dependencies inherent in cultural performances. 
The attention mechanism implementation enables the model 
to focus selectively on culturally significant segments within 
temporal sequences, computed as: 

1

( )

( )

t
t T

k

k

exp e

exp e



=

=



 (2) 

where 𝑒𝑡 = 𝛼(ℎ𝑡, 𝑠𝑡−1) represents the attention energy. 
Research indicates that multimodal co-learning approaches 
substantially improve recognition accuracy through effective 
feature fusion strategies [35]. 
The multimodal fusion strategy adopts a hierarchical 
approach, integrating visual, textual, and auditory modalities 
through cross-modal attention mechanisms. Visual-audio 
integration systems have demonstrated superior 
performance in cultural pattern recognition tasks[36]. The 
fusion process employs learned weights: 

1

M

fused i i

i

F wF
=

=  (3) 

where M  represents the number of modalities, and wi 
denotes modality-specific weights. Advanced audio-visual 
learning techniques enhance the system's capacity to 
preserve subtle cultural nuances through synchronized 
multimodal processing [37]. Model optimization incorporates 
adaptive learning rate schedules, dropout regularization, and 
early stopping mechanisms to prevent overfitting while 
maintaining generalization capabilities across diverse Li 
subgroup patterns. Cultural nuance preservation employs 
gradient-based feature attribution analysis combined with 
cultural expert validation to ensure attention mechanisms 
capture community-defined cultural meanings rather than 
spurious correlations, while cultural constraint losses 
penalize representations that deviate from expert-validated 
cultural semantic spaces. 

3.4 Cultural pattern recognition algorithms 
The framework used for distinguishing cultural patterns 

is based on sophisticated algorithms designed to detect 
intricate subtleties within the Li ethnic tradition in multiple 
modalities. Recent studies suggest that emotion recognition 
within cross-culture requires an extensive analysis of 
multimodal features that goes beyond conventional one-
modality methods [38]. The image feature extraction module 
is built on a hierarchical convolutional structure reinforced 
by residual links, which allows it to extract basic visual 
features and higher-level semantic features from cultural 
objects, textiles, and architectural features. To support inputs 
of differing sizes, the feature extraction module applies 
spatial pyramid pooling to calculate feature maps as: 

SPP( ( * ))spatial c cF W I b= +  (4) 

where I represents the input image, Wc denotes convolutional 
weights, and 𝜎 is the activation function. 
Text semantic analysis leverages transformer-based language 
models fine-tuned for Li ethnic terminology and cultural 
concepts. The semantic embedding process captures 
contextual relationships within oral traditions and folklore 
narratives through attention mechanisms that model long-
range dependencies. The semantic representation is 
computed as: 

Transformer( )semantic e eE W T P=  +             (5) 

where T represents tokenized text, We denotes embedding 
weights, and Pe indicates positional encodings. 
Audio signal processing employs mel-frequency cepstral 
coefficients combined with chromagram features to capture 
tonal characteristics unique to Li traditional music. Research 
indicates that multimodal behavior analysis significantly 
enhances cultural affect recognition when incorporating 
temporal dynamics [39]. The audio feature vector integrates 
spectral and temporal information through: 

[ ( ),

( ), ( )]

audioF MFCC x

Chroma x RMS x

=  (5) 

Cross-modal feature alignment addresses the semantic 
gap between different modalities through canonical 
correlation analysis and adversarial training. Cultural nuance 
preservation employs contrastive learning frameworks that 
maintain Li-specific semantic relationships through 
culturally-informed negative sampling, where culturally-
similar but distinct patterns serve as hard negatives to 
prevent feature space collapse while preserving intra-cultural 
variations across Ha, Qi, Run, Sai, and Meifu subgroups. The 
alignment optimization minimizes the distance between 
corresponding features across modalities while preserving 
modal-specific information. Deep learning approaches have 
demonstrated remarkable effectiveness in assessing cultural 
patterns from visual media [40]. Pattern matching employs a 
similarity metric combining Euclidean distance in the aligned 
feature space with cultural context weights, facilitating 
accurate classification of Li subgroup characteristics while 
maintaining cultural authenticity throughout the recognition 
process. 

3.5 Intelligent documentation system architecture 
The intelligent documentation system adopts a five-layer 

distributed architecture designed to ensure scalable and 
efficient preservation of the Li ethnic cultural heritage, as 
illustrated in Figure 2. Contemporary research emphasizes 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

125 

 

the critical importance of cultural intelligence frameworks in 
heritage preservation systems, necessitating robust 
technological architectures that balance preservation 
effectiveness with sustainable implementation strategies 
[41]. The proposed architecture implements a hierarchical 
modular design that facilitates seamless data flow and 
processing across multiple functional domains while 
maintaining cultural authenticity and accessibility. Long-term 
cultural adaptability mechanisms include generational 
knowledge transfer protocols that automatically incorporate 
evolving cultural practices through community-driven 
updates, while maintaining backward compatibility with 
traditional cultural representations to ensure continuity 
across Li ethnic generations. Researchers, cultural 
practitioners, and community members can all access the 
system through various entry points like web portals, mobile 
applications, and administration consoles, which interface 
with the different users. The API gateway offers a singular 
entry point, which, together with the integrated 
authentication mechanisms of the system, protects sensitive 
cultural content. The microservices layer also has 
autonomous domain functions, which include deep learning 
algorithm-based pattern recognition, semantic indexing for  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

more efficient search and retrieval, machine learning-driven 
recommendation engines, and content management with 
metadata processing. The processing layer features engines 
specialised in cultural data processing, such as Natural 
Language Processing (NLP), computer vision-based image 
analysis, and audio signal processing. Components for feature 
fusion allow for integration of multi-modal cultured patterns, 
while high-performance caching systems expedite data 
access. For the data storage layer, a combination of Neo4j 
knowledge graphs for cultural relationships, distributed 
media repositories for multimedia content, Elasticsearch 
vector databases for similarity search, PostgreSQL metadata 
stores for structured information, and cloud backup systems 
for long-term preservation implements a hybrid approach.  
Generational adaptation frameworks employ version-
controlled cultural ontologies that track cultural evolution 
while preserving historical contexts, enabling the system to 
accommodate changing cultural expressions across Li 
subgroups without losing traditional knowledge, supported 
by community governance mechanisms that validate cultural 
updates.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

User Interface Layer

Web Portal Mobile App API Gateway Admin Console Authentication

Microservices Layer

Pattern Recognition
Deep Learning Models

Search & Retrieval
Semantic Indexing

Recommendation
ML Algorithms

Content Management
Metadata Processing

Data Processing Layer

NLP Engine
Text Analysis

Computer Vision
Image Analysis

Audio Processing
Signal Analysis

Feature Fusion
Multimodal Integration

Cache System
High-Speed Access

Data Storage Layer

Knowledge Graph
·Neo4i Database

·Cultural Relationships

·Semantic Networks

·Historical Connections

Media Repository
·Distributed Storage.

·Images & Videos

·Audio Recordings

·3D Models

Vector Database
·Elasticsearch

·Feature Embeddings

·Similarity Search

· Pattern Matching

Metadata Store
·PostgreSQL

·Structured Data

·User Information

·Access Control

Backup System
·Cloud Archive.

·Disaster Recovery

·Long-term Storage

· Data Integrity

Infrastructure Layer

Kubernetes Cluster Load Balancing Monitoring & Logging Security Layer CI/CD Pipeline

Intelligent Documentation System Architecture for Li Ethnic Cultura  Heritage

Performance Metrics:
·Response Time: < 200ms         ·Availability: 99.9%

Scalability:
·Auto-scaling    ·Load Distribution   ·High Availability

Security:
.End-to-end Encryption   .Cultural Data Protection

 

Figure 2. Intelligent documentation system architecture for Li ethnic cultural heritage 

 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

126 

 

Through Kubernetes orchestration, the technological 
framework is provided at the infrastructure layer using 
systems for load balancing, comprehensive monitoring and 
logging, enhanced security framework, and unit testing 
coupled with continuous integration and deployment 
pipelines. This architecture allows for horizontal scaling 
alongside sustaining 99.9% uptime, responsiveness in under 
200ms, and end-to-end encryption which uses specialised 
cultural data protection protocols to safeguard Li ethnic 
heritage materials during the documentation and 
preservation process. Future-proofing strategies include 
modular component design enabling seamless technology 
upgrades, while cultural continuity safeguards ensure that 
system evolution preserves intergenerational knowledge 
transmission pathways essential for sustainable Li ethnic 
heritage preservation. 

3.6 Evaluation methods 
The developed evaluation framework utilises a 

multicriteria methodology that systematically evaluates both 
the technological effectiveness and the ability to preserve the 
cultural authenticity of the proposed system. Contemporary 
studies observe that exhaustive investigation of deep learning 
algorithms needs systematic approaches that transcend the 
simplistic use of evaluative criteria [42]. The training and 
validation of the model follows a stratified k-fold cross-
validation scheme with k=5, which balances the 
representation of Li groups while maintaining temporal 
continuity in succession-order cultural data. The 
performance framework incorporates both quantitative and 
qualitative measures to address effectively the multifaceted 
features involved in identifying cultural patterns. Traditional 
metrics of classification, like precision, recall, and F1-score, 
are based on the following equations: 

TP
Precision

TP FP
=

+
 (6) 

TP
Recall

TP FN
=

+
 (7) 

2
1

Precision Recall
F

Precision Recall

 
=

+
 (8) 

where TP, FP, and FN represent true positives, false positives, 
and false negatives, respectively. Performance evaluation 
encompasses both controlled laboratory conditions using 
stratified data partitions and real-world deployment 
scenarios across active Li communities, with metrics 
validated through 6-month field testing to assess practical 
applicability beyond experimental datasets. Previous works 
highlight the need for stringent evaluation measures and the 
application of statistical testing methods to ensure machine 
learning methods [43]. Cultural authenticity preservation is 
evaluated using expert rating scores and semantic similarity 
measures, which are computed from cosine similarity 
measures of reconstructed cultural representations and 
actual ones. External validity assessment examines system 
performance across temporal and cultural variations, as well 
as emerging cultural practices, ensuring that evaluation 
results generalize to dynamic cultural environments where Li 
traditions naturally evolve while maintaining core cultural 
integrity. 

Statistical significance testing uses paired t-tests and 
Wilcoxon signed-rank tests to determine model efficacy 
under different conditions and across cultural subgroups. 
Statistical significance is determined using the following: 

/d

d
t
s n

=  (9) 

where �̅�  represents the mean difference, Sd the standard 
deviation of differences, and n the sample size. Recent 
analysis of artificial intelligence applications in cultural 
heritage preservation emphasizes the importance of robust 
evaluation methodologies [44]. The evaluation protocol 
incorporates domain expert assessments to validate cultural 
accuracy, while AI-based visualization techniques enable 
interpretable analysis of model decisions [45]. Interactive 
evaluation approaches through immersive technologies 
provide additional validation mechanisms for user 
acceptance and cultural engagement [46]. Model 
interpretability analysis utilizes attention visualization 
methods to ensure transparency in cultural pattern 
recognition decisions [47]. Cross-temporal validation 
protocols test system robustness against cultural change by 
evaluating performance on cultural practices documented 
across different time periods, ensuring long-term reliability in 
dynamic heritage preservation contexts where cultural 
expressions continuously adapt while preserving essential 
characteristics. 

4. Results 

4.1 Dataset construction results 
The comprehensive Li ethnic cultural dataset 

demonstrates substantial scope and systematic organization 
across multiple data modalities and cultural subgroups. As 
illustrated in Figure 3(a), the dataset encompasses a diverse 
array of cultural materials with visual data constituting the 
largest component at 45.2% of the total collection, followed 
by textual materials at 28.7%, audio recordings at 15.6%, 
metadata at 6.8%, and expert annotations at 3.7%. This 
distribution reflects the research focus on capturing tangible 
cultural expressions while maintaining comprehensive 
documentation of intangible heritage elements through 
textual and audio recordings. Data annotation quality analysis 
reveals consistently high standards across all Li ethnic 
subgroups, as shown in Figure 3(b). The Ha subgroup exhibits 
the highest annotation quality with 92.5% rated as excellent, 
while the Meifu subgroup maintains 76.8% excellent ratings 
despite having the smallest sample size. Quality assessment 
protocols incorporated expert validation from cultural 
practitioners and academic specialists, ensuring cultural 
authenticity and technical accuracy throughout the 
annotation process. 

Dataset distribution statistics demonstrate systematic 
sampling across Li ethnic subgroups, as depicted in Figure 
3(c). The Ha subgroup provides the largest contribution with 
1,250 samples, while sample sizes gradually decrease for Qi 
(980), Run (850), Sai (720), and Meifu (650) subgroups. 
Notably, the average number of cultural elements per sample 
shows a corresponding pattern, ranging from 5.2 elements 
per sample in the Ha subgroup to 3.9 elements in the Meifu 
subgroup, reflecting varying cultural complexity and 
documentation depth across different communities. 

The validation results in different segments of the 
dataset illustrate excellent performance metrics, as shown in 
Figure 3(d). The training dataset shows maximum 
performance metrics, which include accuracy of 94.8%, 
precision of 93.5%, recall of 95.1%, and F1-scores of 94.3%. 
The uniform performance of validation and test datasets 
supports the reliability of the dataset for machine learning 
purposes, as it does not degrade much while switching from 
training to testing contexts. 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

127 

 

The dataset properties are carefully detailed in Table 2, 
which presents comprehensive statistics of every subgroup of 
the Li ethnic group, including sample distributions, cultural 
element diversity, and several quality measures. The method 
used in the dataset construction successfully achieved 
representational fairness among the subgroups while 
maintaining rigorous annotation standards essential for 
developing reliable models for cultural pattern recognition. 
Table 2 shows that the dataset includes an extensive 
representation of cultural diversity within the Li ethnic group, 
while maintaining uniform quality throughout each of the 
subcategories. This effectively provides a strong foundation 
for further training of deep learning algorithms and analysis 
of cultural pattern perception. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

4.2 Deep Learning model performance 
The extensive analysis of deep learning architectures 

shows significant gains in performance due to the introduced 
multimodal fusion method for Li ethnic cultural pattern 
identification. Figure 4(a) shows that the proposed graph 
performs better than baseline configurations for every Li 
ethnic subgroup, achieving impressive accuracy figures of 
94.8% for subgroup Ha, 91.6% for subgroup Qi, 89.3% for 
subgroup Run, 86.7% for subgroup Sai, and 84.1% for 
subgroup Meifu. All these results affirm that the introduced 
method well captures each Li subgroup's specific cultural 
subtlety while retaining strong performance despite data 
complexities and varying sizes of samples. 

 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Figure 3. Comprehensive analysis of Li ethnic cultural heritage dataset construction and validation (a) Multimodal data composition 

and structure analysis (b) Cross-subgroup annotation quality assessment (c) Sample distribution and cultural element statistics (d) 

Performance validation across dataset partitions 

Table 2. Detailed dataset characteristics by Li ethnic subgroup 

Subgroup Sample 
Count 

Visual Data Text Data Audio Data Avg. Cultural 
Elements 

Annotation Quality 
(%) 

Ha 1,250 465 298 187 5.2 92.5 (Excellent) 

Qi 980 356 251 152 4.8 88.3 (Excellent) 

Run 850 298 201 134 4.5 85.7 (Excellent) 

Sai 720 245 168 108 4.1 79.2 (Good) 

Meifu 650 201 142 97 3.9 76.8 (Good) 

Total 4,450 1,565 1,060 678 4.5 84.5 (Overall) 

 

 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

128 

 

The ablation study findings, as evidenced by Figure 4(b), 
reveal the individual contributions of each component to the 
overall effectiveness of the model. The baseline model 
achieves 76.2% accuracy, with sequential improvements 
observed through integration of visual features (+Visual: 
82.5%), textual analysis capabilities (+Text: 86.8%), audio 
processing modules (+Audio: 89.4%), and attention 
mechanisms (+Attention: 91.7%). The complete model 
incorporating all components reaches 94.8% accuracy, 
demonstrating that each modality and architectural 
enhancement contributes meaningfully to the cultural 
pattern recognition task. 
 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Generalization performance analysis, depicted in Figure 
4(c), confirms the model's robustness across diverse testing 
scenarios. The proposed approach maintains superior 
performance in in-domain evaluations (94.8%) while 
demonstrating acceptable degradation in cross-domain 
(89.2%), temporal (86.5%), noisy (83.7%), and limited data 
scenarios (81.4%). This performance consistency 
significantly exceeds that of traditional multimodal (91.5% to 
74.6%) and transformer-based approaches (88.2% to 68.9%), 
indicating enhanced adaptability to real-world deployment 
conditions where data quality may vary from training 
conditions. 
 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 4. Comprehensive Performance Analysis of Deep Learning Models for Li Ethnic Cultural Pattern Recognition.(a) Cross-

Subgroup Accuracy Comparison of Model Architectures;(b) Ablation Study of Multimodal Component Contributions;(c) Cross-Domain 

Generalization Performance Assessment;(d) Model Stability Analysis Across Multiple Training Iterations 

Figure 5. Training dynamics and computational efficiency analysis of deep learning architectures (a) Convergence behavior and loss-

accuracy evolution during training, (b) Parameter-inference time trade-off analysis with memory usage visualization 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

129 

 

Model stability analysis through repeated experiments 
reveals consistent performance characteristics across 
multiple training iterations. Figure 4(d) demonstrates that 
the proposed model exhibits minimal variance in accuracy 
scores, with the interquartile range significantly narrower 
than comparable architectures. Training convergence 
analysis presented in Figure 5(a) illustrates efficient 
optimization dynamics with rapid initial improvements 
followed by stable convergence. The training loss decreases 
smoothly from approximately 0.8 to below 0.2 within 50 
epochs, while validation accuracy stabilizes at 89% without 
significant overfitting indicators. 

Computational complexity evaluation, shown in Figure 
5(b), reveals trade-offs between model sophistication and 
computational efficiency. The proposed model requires 18.6 
million parameters with 156ms inference time and 6.4GB 
memory usage. While these requirements exceed simpler 
architectures, the computational overhead remains 
reasonable considering substantial performance gains 
achieved. As shown in Table 3, the proposed model achieves 
optimal performance metrics across all evaluation criteria 
while maintaining acceptable computational requirements 
for practical deployment scenarios, establishing its 
effectiveness for comprehensive Li ethnic cultural pattern 
recognition applications. 

4.3 Cultural pattern recognition performance 
The comprehensive evaluation of cultural pattern 

recognition demonstrates the proposed system's 
effectiveness in identifying and classifying diverse Li ethnic 
cultural elements. As illustrated in Figure 6(a), the 
recognition accuracy varies significantly across different 
cultural domains, with traditional textiles achieving the 
highest performance at 96.2%, followed by music (93.8%) 
and architecture (91.5%). Language-related cultural patterns 
present the greatest recognition challenges, achieving 85.6% 
accuracy, reflecting the complexity of linguistic nuances 
within Li ethnic expressions. 

Li subgroup classification performance reveals 
consistent excellence across all ethnic subdivisions, as 
depicted in Figure 6(b). The Ha subgroup demonstrates 
superior classification metrics with precision, recall, and F1-
scores of 94.7%, 95.1%, and 94.9%, respectively. 
Performance gradually decreases across Qi, Run, Sai, and 
Meifu subgroups, with the latter achieving 82.8% precision, 
83.2% recall, and 83.0% F1-score. This performance gradient 
correlates with the size and complexity of cultural expression 
datasets available for each subgroup. 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Cross-modal recognition performance analysis, shown in 
Figure 6(c), confirms the superiority of multimodal fusion 
approaches. Single-modality systems exhibit moderate 
performance, with visual-only recognition achieving 87.5%, 
text-only reaching 82.1%, and audio-only obtaining 79.8%. 
Dual-modality combinations demonstrate substantial 
improvements, with visual-text fusion reaching 91.2% and 
visual-audio combination achieving 90.3%. The complete 
multimodal system attains optimal performance at 94.8%, 
validating the comprehensive integration strategy. Error 
analysis reveals that inter-subgroup confusion constitutes the 
primary classification challenge, accounting for 35.2% of 
misclassifications, as shown in Figure 6(d). Intra-cultural 
variation represents 28.6% of errors, while noise-induced 
errors contribute 18.5%. Temporal inconsistency and context 
misclassification account for smaller proportions at 12.3% 
and 5.4% respectively, indicating the model's robustness 
against external interference factors. 

Model interpretability analysis, presented in Figure 7(a), 
identifies color patterns as the most discriminative feature 
with an attention weight of 0.180, followed by geometric 
shapes (0.150) and semantic content (0.140). Cultural 
symbols demonstrate the lowest attention weight at 0.090, 
suggesting their limited discriminative power across Li 
subgroups. The feature importance hierarchy provides 
valuable insights for cultural documentation prioritization. 
Case study analysis across different artifact categories, as 
depicted in Figure 7(b), reveals consistent performance 
patterns. Traditional textiles maintain the highest recognition 
accuracy across all subgroups (μ=87.4%), while architectural 
elements show the most significant performance variation 
(μ=80.6%). The detailed performance metrics are 
summarized in Table 4, demonstrating the system's reliability 
across diverse cultural manifestations. 

As shown in Table 4, the system achieves superior 
performance across tangible cultural elements while 
demonstrating acceptable accuracy for intangible 
expressions, establishing its comprehensive utility for Li 
ethnic heritage preservation and documentation applications. 

4.4 Intelligent documentation system functionality 
verification 

The comprehensive evaluation demonstrates robust 
performance of the intelligent documentation system across 
multiple operational dimensions. System response time 
analysis, shown in Figure 8(a), indicates acceptable 
performance under moderate loads, maintaining sub-200ms 
response times for up to 1000 concurrent users.  

 

 

 

 

 

 

 

 

 

 

 

 

Table 3. Comprehensive model performance comparison 

Model Accuracy (%) Precision (%) 
Recall 

(%) 
F1-Score 

(%) 
Parameters (M) 

Inference 
Time (ms) 

Memory 
Usage (GB) 

CNN 85.3 83.7 86.1 84.9 2.1 12 0.8 

RNN 87.6 86.2 88.4 87.3 3.8 25 1.5 

Transformer 89.4 88.1 90.2 89.1 12.5 89 4.2 

Multimodal 92.1 91.3 92.8 92.0 15.2 145 5.8 

Proposed 94.8 94.1 95.2 94.6 18.6 156 6.4 

 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

130 

 

 

 

 

 

 

 

 

 

 

 

Figure 6. Cultural pattern recognition performance analysis (a) Cultural element recognition accuracy, (b) Li subgroup classification 

performance, (c) Cross-modal performance comparison (d) Error type analysis 

Figure 7. Model interpretability and cultural case study analysis (a) Cultural feature importance analysis (b) Cross-subgroup case study 

comparison 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

131 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Throughput peaks at 11,200 requests per hour before 
declining due to resource constraints at higher loads. 
Retrieval performance metrics, illustrated in Figure 8(b), 
validate the superiority of multimodal search capabilities, 
achieving 96.8% precision, 95.2% recall, and 96.0% F1-score. 
Single-modality approaches demonstrate lower performance, 
with textual queries (94.6%), visual search (91.3%), and 
audio matching (88.7%) confirming the effectiveness of 
multimodal fusion strategies. System stability testing over 48 
hours, depicted in Figure 8(c), reveals consistent error rates 
below 0.5% with efficient resource management.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
 
 
 
 
 
The 99.7% stability metric represents continuous 

operational testing under controlled laboratory conditions 
with simulated user loads of 500-1000 concurrent 
connections over 7-day periods, including planned system 
maintenance windows and automated recovery protocols, 
validated through enterprise-grade monitoring across 
distributed infrastructure components.CPU usage ranges 
from 25-44% while memory consumption remains at 45-67%, 
confirming reliable operational characteristics for sustained 
deployment. Query processing analysis, as shown in Figure 
8(d), demonstrates the system's adaptive performance across 

Table 4. Detailed cultural pattern recognition performance by category 

Model Accuracy (%) Precision (%) 
Recall 

(%) 
F1-Score 

(%) 
Parameters (M) 

Inference 
Time (ms) 

Memory 
Usage (GB) 

CNN 85.3 83.7 86.1 84.9 2.1 12 0.8 

RNN 87.6 86.2 88.4 87.3 3.8 25 1.5 

Transformer 89.4 88.1 90.2 89.1 12.5 89 4.2 

Multimodal 92.1 91.3 92.8 92.0 15.2 145 5.8 

Proposed 94.8 94.1 95.2 94.6 18.6 156 6.4 

 

Figure 8. Comprehensive system performance and retrieval analysis of intelligent documentation platform (a) Response time and 

throughput performance under varying user loads, (b) Cross-modal retrieval performance evaluation across query types, (c) Long-term 

system stability and resource utilization analysis, (d) Query processing performance analysis by complexity level 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

132 

 

complexity levels. Real-world deployment scenarios 
demonstrate 8-12% performance degradation compared to 
laboratory conditions, with response times increasing from 
156ms to 175ms under actual community usage patterns, 
while maintaining 97.2% accuracy in controlled settings 
versus 91.8% accuracy in field deployments with variable 
network connectivity and diverse user interactions. Simple 
queries achieve optimal performance with a 58ms response 
time and 97.2% accuracy, while very complex queries require 
215ms with 89.6% accuracy. The balanced trade-off between 
processing time and accuracy validates the system's 
capability to handle diverse cultural documentation 
requirements. User satisfaction evaluation, presented in 
Figure 9(a), achieves an overall average of 4.40 out of 5, with 
ease of use rated highest (4.6) and response speed lowest 
(4.2). Expert evaluation demonstrates superior confidence 
with a weighted average of 4.62, particularly recognizing 
cultural accuracy (4.8) and documentation quality (4.7). 
Laboratory performance metrics consistently outperform 
field deployment by 3-5% across all evaluation dimensions, 
reflecting the impact of real-world variables, including 
network latency, hardware diversity, and user interaction 
patterns not present in controlled testing environments. 
Performance benchmark comparison, shown in Figure 9(b), 
indicates 67% target achievement with four of six metrics 
successfully met. The system exceeds benchmarks in user 
satisfaction (88 vs 80), expert rating (92 vs 85), system 
stability (96 vs 95), and cultural precision (91 vs 90). As 
shown in Table 5, the system demonstrates strong 
performance across critical dimensions, with query 
complexity analysis revealing adaptive capabilities that 
maintain acceptable accuracy even for complex tasks 
involving cultural pattern recognition. 

4.5 Real-World Application Impact Assessment 
The comprehensive evaluation demonstrates significant 

positive impacts across multiple dimensions of Li ethnic 
cultural preservation and community engagement. Cultural 
worker feedback analysis, as illustrated in Figure 10(a), 
reveals intense satisfaction with an average rating of 4.5 out 
of 5. System usability achieves the highest rating at 5.0, while 
cultural accuracy and documentation efficiency both receive 
ratings of 4.5, confirming the system's effectiveness in 
preserving authentic cultural representations. 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 5. System performance summary 

 
 
Community participation trends, depicted in Figure 

10(b), exhibit remarkable growth throughout 2024. Active 
user engagement demonstrates 340% growth, while monthly 
contributions increase at 665% rate, indicating enhanced 
community involvement in cultural documentation activities. 
The parallel growth patterns suggest strong correlation 
between user adoption and meaningful participation. 

Cultural transmission effectiveness assessment reveals 
substantial improvements following system implementation, 
as shown in Figure 10(c). Knowledge retention improves 
from 45 to 78 points (73% enhancement), while skill transfer 
increases from 38 to 84 points (121% improvement). Cultural 
practice preservation exhibits the most significant 
improvement, from 42 to 89 points (112% increase). Digital 
archiving capabilities improve dramatically from 25 to 95 
points (280% enhancement), while youth engagement shows 
notable progress from 35 to 82 points (134% improvement). 

Social impact assessment, presented in Figure 10(d), 
demonstrates positive outcomes with an overall score of 4.5 
out of 5. Community pride achieves the highest rating at 4.8, 
while educational value and cultural awareness receive 
ratings of 4.4 and 4.6, respectively. Tourism promotion and 
research contribution maintain solid ratings of 4.2 and 4.7.  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Metric Current Target Status Query 
Distribution 

Response 
Time 

157ms <200ms ✓ Met Simple: 35%, 
Complex: 
25% 

Search 
Accuracy 

94.8% >95% ○ -
0.2% 

89.6-97.2% 
range 

User 
Satisfaction 

4.4/5 >4.0 ✓ Met All 
aspects >4.0 

Expert 
Rating 

4.6/5 >4.0 ✓ Met Cultural 
accuracy: 4.8 

System 
Stability 

99.7% >99% ✓ Met 48-hour 
testing 

Cultural 
Precision 

94.2% >90% ✓ Met Cross-modal 
validated 

Figure 9. User evaluation and performance benchmark assessment for heritage documentation system, (a) User satisfaction and expert 

evaluation analysis across multiple dimensions, (b) Performance vs target benchmark comparison with achievement assessment 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

133 

 

 

 

 

 
As shown in Table 6, the implementation demonstrates 

consistent positive impacts across all dimensions, with digital 
archiving experiencing the most significant transformation 
while maintaining substantial improvements in traditional 
cultural transmission methods. 

5. Discussion 

The exceptional performance achieved across Li ethnic 
subgroups demonstrates the effectiveness of multimodal 
fusion architectures in capturing complementary cultural 
information across visual, textual, and auditory dimensions 
[34]. The findings corroborate modern theories of 
multimodal design that emphasize the importance of 
combining multiple semiotic resources to allow full cultural 
interpretation [48]. Differential weighting of the attention 
mechanism on culture-relevant features justifies established 
principles of biometric recognition within pattern recognition 
[49].The high level of enhancement in cultural transmission 
effectiveness is owed to this system's capacity to support both 
explicit and implicit cultural knowledge through advanced 
documentation methods [11].  

 
 
 
 

 
 
 
 
 

 
Table 6. Comprehensive real-world application impact metrics 

 

 
 
 
 

Impact 
Category 

Baseline 
Score 

Current 
Score 

Improvement 
(%) 

Stakeholder 
Count 

Knowledge 
Retention 

45 78 +73% 156 

Skill 
Transfer 

38 84 +121% 142 

Cultural 
Practice 

42 89 +112% 128 

Digital 
Archiving 

25 95 +280% 89 

Youth 
Engagement 

35 82 +134% 87 

Average 37 86 +144% 602 

Figure 10. Real-world application impact assessment of li ethnic cultural documentation system, (a) Cultural heritage professional feedback 

analysis, (b) Community participation growth trends (2024), (c) Cultural transmission effectiveness: before vs after implementation, (d) 

Multi-stakeholder social impact assessment 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

134 

 

Such an achievement tackles intrinsic issues identified 
within previous research in terms of balancing technological 
development and heritage integrity preservation [23]. The 
research framework outlined goes beyond conventional 
approaches based on manual classification, showing that 
advanced computational methods can alleviate current 
deficiencies in current documentation strategies [50]. 
Comparative analysis demonstrates significant advantages 
over existing approaches. As opposed to more traditional 
methods of documenting cultural heritage, which struggle 
with consistency and scaling, this human-centred AI 
approach has shown that such systems can improve 
accessibility while remaining sensitive to cultural issues [51]. 
The fusion of deep learning with the recognition of cultural 
patterns represents a major improvement compared to 
earlier works that concentrated on simple digitisation rather 
than comprehensive analysis [52]. The overall excellence of 
the system in cross-domain generalisation is the strongest, 
surpassing conventional multimodal and transformer-based 
systems. 

Still, this research recognises major gaps. The problems 
of capturing data indicate greater difficulties in exploring 
minority cultures where the digital divide poses the greatest 
challenge in the effective safeguarding of heritage resources 
[53]. The assumption of data quality and model performance 
emphasises persistent challenges in achieving 
representational parity among diverse cultural 
constituencies. Also, cultural sensitivity issues raise the need 
to question how protected cultural materials should be 
handled within technological frameworks. The costs 
associated with multimodal processing may impede its use in 
resource-constrained settings. Difficulties in resolving 
culturally rich expressions go beyond mere recognition. Such 
challenges require deep, at times anthropological 
understanding, which falls into the realm of applied sociology, 
rather than mere pattern recognition. Further, ethical 
challenges relating to data ownership and community 
consent remain as complex barriers, calling for ongoing 
negotiation between technological advances and heritage 
communities [54]. 

Later studies should highlight the development of 
culturally sensitive algorithms that can perform well despite 
having limited data, while preserving cultural subtleties. The 
use of advanced natural language processing techniques can 
enhance oral tradition understanding, providing richer 
semantic depth [55]. Exploration of metaverse applications 
presents promising opportunities for immersive cultural 
experiences, revolutionizing heritage education and tourism 
[56]. The development of artificial intelligence frameworks 
specifically tailored for heritage innovation represents a 
crucial advancement [57]. Long-term sustainability planning 
must address evolving technological landscapes while 
ensuring continuous community engagement. The 
framework's potential expansion to other minority cultures 
requires a systematic investigation of transferability 
mechanisms. Collaborative research initiatives involving 
heritage communities, technologists, and cultural experts 
represent essential pathways for advancing ethical and 
effective cultural preservation methodologies serving both 
scholarly understanding and community interests [58]. 

6. Conclusion 

This research establishes a comprehensive framework 
for Li ethnic cultural heritage preservation through deep 
learning-based pattern recognition systems, achieving 
remarkable technical and practical outcomes. The proposed 

multimodal fusion architecture demonstrates superior 
performance with 94.8% overall accuracy across Li 
subgroups, significantly outperforming traditional 
approaches and generic heritage systems. The intelligent 
documentation system maintains exceptional operational 
efficiency with response times under 200ms and achieves 
99.7% system stability while preserving cultural authenticity 
through expert-validated methodologies. These 
achievements represent substantial advancement in 
computational approaches to intangible heritage 
preservation, establishing new benchmarks for accuracy, 
efficiency, and cultural sensitivity in digital heritage 
technologies. The research addresses critical challenges in 
minority cultural preservation, offering innovative solutions 
for the documentation, transmission, and accessibility of Li 
ethnic traditions. The effectiveness of cultural transmission 
shows impressive improvement across multiple facets. 
Knowledge retention was enhanced by 73%, skill transfer 
improved by 121%, and competencies in archiving digitally 
yielded a startling increase of 280%. Community involvement 
reflects successful community engagement and viable 
preservation strategies as it demonstrates exponential 
growth with a 340% rise in active participants and a 665% 
rise in monthly contributions. These results confirm the 
framework’s ability to integrate traditional methods of 
preservation with contemporary technological approaches, 
preserving cultural and community ethics. This research is a 
valuable addition to interdisciplinary anthropology by 
applying computational techniques alongside 
anthropological insights, thus creating methodological 
frameworks achievable in all cultural settings. The rated user 
satisfaction of 4.4 out of 5 and an expert rating of 4.6 out of 5 
prove the system’s practical functionality alongside its 
academic credibility. The research showcases strong cross-
domain generalisation capability, achieving 89.2% accuracy, 
which indicates a strong relevance to other minority cultures 
and heritage settings. Evaluating social impact yields 
composite scores of 4.5 out of 5 across stakeholder groups, 
indicating constructive outcomes for community 
development, educational programmes, and cultural 
advocacy. The results offer technological recommendations 
regarding policy for conservation, which incorporate active 
community participation alongside sustainable development. 
The model’s previously shown flexibility and versatility 
suggest its international applicability in conservation efforts. 
In policy, community-driven conservation, ethical policies 
regarding the building of digital heritage, and securing 
sustainable funding for maintenance are all recommended. 
Coordinated actions involving heritage groups, technology 
development, and policy design are needed to build culturally 
respectful, technologically advanced, academically rigorous, 
and community-responsive preservation systems that sustain 
the integrity and vitality of cultural practice for future 
generations.  

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically with regard to authorship 
(avoidance of guest authorship), dual submission, 
manipulation of figures, competing interests, and compliance 
with policies on research ethics. The author adheres to 
publication requirements that the submitted work is original 
and has not been published elsewhere. 

 
 
 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

135 

 

Data availability statement 
The manuscript contains all the data. However, more data will 

be available upon request from the authors. 

Conflict of interest 

The authors declare no potential conflict of interest. 

References 

[1] G. Aktürk and M. Lerski, "Intangible cultural heritage: 
a benefit to climate-displaced and host communities," 
Journal of Environmental Studies and Sciences, vol. 
11, no. 3, pp. 305-315, 2021. 
DOI:https://doi.org/10.1007/s13412-021-00697-y 

[2] D. Giglitto, L. Ciolfi, and W. Bosswick, "Building a 
bridge: opportunities and challenges for intangible 
cultural heritage at the intersection of institutions, 
civic society, and migrant communities," 
International Journal of Heritage Studies, vol. 28, no. 
1, pp. 74-91, 2022. DOI: 
https://doi.org/10.1080/13527258.2021.1922934 

[3] Y. Du, L. Chen, and J. Xu, "Interactive effects of 
intangible cultural heritage and tourism 
development: a study based on the data panel PVAR 
model and coupled coordination model," Heritage 
Science, vol. 12, no. 1, p. 401, 2024. DOI: 
https://doi.org/10.1186/s40494-024-01502-z 

[4] J. Eichler, "Intangible cultural heritage, inequalities 
and participation: who decides on heritage?," The 
International Journal of Human Rights, vol. 25, no. 5, 
pp. 793-814, 2021. DOI: 
https://doi.org/10.1080/13642987.2020.1822821 

[5] H. Mekonnen, Z. Bires, and K. Berhanu, "Practices and 
challenges of cultural heritage conservation in 
historical and religious heritage sites: evidence from 
North Shoa Zone, Amhara Region, Ethiopia," Heritage 
Science, vol. 10, no. 1, p. 172, 2022. 

[6] D. Harisanty, K. L. B. Obille, N. E. V. Anna, E. Purwanti, 
and F. Retrialisca, "Cultural heritage preservation in 
the digital age, harnessing artificial intelligence for 
the future: a bibliometric analysis," Digital Library 
Perspectives, vol. 40, no. 4, pp. 609-630, 2024. 
DOI:https://doi.org/10.1108/DLP-01-2024-0018 

[7] I. Siliutina, O. Tytar, M. Barbash, N. Petrenko, and L. 
Yepyk, "Cultural preservation and digital heritage: 
challenges and opportunities," Amazonia Investiga, 
vol. 13, no. 75, pp. 262-273, 2024. DOI: 
https://doi.org/10.34069/AI/2024.75.03.22 

[8] X. Ye, Y. Ruan, S. Xia, and L. Gu, "Adoption of digital 
intangible cultural heritage: a configurational study 
integrating UTAUT2 and immersion theory," 
Humanities and Social Sciences Communications, vol. 
12, no. 1, pp. 1-17, 2025. 
DOI:https://doi.org/10.1057/s41599-024-04222-8 

[9] C. Lin and C. Li, "Digital divide of intangible cultural 
heritage and innovative inheritance 
countermeasures," Journal of Sociology and 
Ethnology, vol. 5, no. 11, pp. 110-122, 2023. DOI: 
10.23977/jsoce.2023.051115 

[10] Y. Hou, S. Kenderdine, D. Picca, M. Egloff, and A. 
Adamou, "Digitizing intangible cultural heritage 
embodied: State of the art," Journal on Computing 
and Cultural Heritage (JOCCH), vol. 15, no. 3, pp. 1-20, 
2022. DOI: https://doi.org/10.1145/3494837 

[11]  J. Wu, L. Guo, J. Jiang, and Y. Sun, "The digital 
protection and practice of intangible cultural heritage 
crafts in the context of new technology," in E3S Web 

of Conferences, 2021, vol. 236: EDP Sciences, p. 
05024. DOI: 
https://doi.org/10.1051/e3sconf/202123605024 

[12] M. Skublewska-Paszkowska, M. Milosz, P. Powroznik, 
and E. Lukasik, "3D technologies for intangible 
cultural heritage preservation—literature review for 
selected databases," Heritage Science, vol. 10, no. 1, p. 
3, 2022. 

[13] J. Liu, "Digitally Protecting and Disseminating the 
Intangible Cultural Heritage in Information 
Technology Era," Mobile Information Systems, vol. 
2022, no. 1, p. 1115655, 2022. DOI: 
https://doi.org/10.1155/2022/1115655 

[14] M. Fiorucci, M. Khoroshiltseva, M. Pontil, A. Traviglia, 
A. Del Bue, and S. James, "Machine learning for 
cultural heritage: A survey," Pattern Recognition 
Letters, vol. 133, pp. 102-108, 2020. DOI: 
https://doi.org/10.1016/j.patrec.2020.02.017 

[15] S. Münster et al., "Artificial intelligence for digital 
heritage innovation: Setting up a r&d agenda for 
europe," Heritage, vol. 7, no. 2, pp. 794-816, 2024. 
DOI: https://doi.org/10.3390/heritage7020038 

[16] M. B. Prados-Peña, G. Pavlidis, and A. García-López, 
"New technologies for the conservation and 
preservation of cultural heritage through a 
bibliometric analysis," Journal of Cultural Heritage 
Management and Sustainable Development, vol. 15, 
no. 3, pp. 664-686, 2025. 

[17] Y. Yuan, Z. Li, and B. Zhao, "A Survey of Multimodal 
Learning: Methods, Applications, and Future," ACM 
Computing Surveys, 2025. DOI: 
https://doi.org/10.1145/3713070 

[18] Z. Qin, Q. Luo, Z. Zang, and H. Fu, "Multimodal GRU 
with directed pairwise cross-modal attention for 
sentiment analysis," Scientific Reports, vol. 15, no. 1, 
p. 10112, 2025. DOI: 
https://doi.org/10.1038/s41598-025-93023-3 

[19] T. Fan, H. Wang, and S. Deng, "Intangible cultural 
heritage image classification with multimodal 
attention and hierarchical fusion," Expert Systems 
with Applications, vol. 231, p. 120555, 2023. DOI: 
https://doi.org/10.1016/j.eswa.2023.120555 

[20] M. Carrozzino, A. Scucces, R. Leonardi, C. Evangelista, 
and M. Bergamasco, "Virtually preserving the 
intangible heritage of artistic handicraft," Journal of 
cultural heritage, vol. 12, no. 1, pp. 82-87, 2011. DOI: 
https://doi.org/10.1016/j.culher.2010.10.002 

[21]  E. Jeong and J. Yu, "Ego-centric recording framework 
for Korean traditional crafts motion," in Euro-
Mediterranean Conference, 2018: Springer, pp. 118-
125. DOI: https://doi.org/10.1007/978-3-030-
01765-1_14 

[22] C. Meghini, V. Bartalesi, and D. Metilli, "Representing 
narratives in digital libraries: The narrative 
ontology," Semantic Web, vol. 12, no. 2, pp. 241-264, 
2021. 

[23] V. Kukreja, A. Singh, D. Kaur, and J. K. Bajwa, Digital 
Cultural Heritage: Challenges, Solutions, and Future 
Directions. CRC Press, 2024. 

[24] J. Guery, M. Hess, and A. Mathys, "Digital techniques 
for documenting and preserving cultural heritage," 
ed: Arc Humanities Press: York, UK, 2017. DOI:  
https://doi.org/10.1080/13527258.2024.2406010 

[25] Y. Tan and W. J. Jehom, "The Function of Digital 
Technology in Minority Language Preservation: The 
Case of the Gyalrong Tibetan Language," 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

136 

 

Preservation, Digital Technology & Culture, vol. 53, 
no. 3, pp. 165-177, 2024. DOI: 
https://doi.org/10.1515/pdtc-2024-0021 

[26] W. Hu, M. Li, X. Chi, X. Wang, and A. U. Khan, 
"Intangible cultural heritage research in China from 
the perspective of intellectual property rights based 
on bibliometrics and knowledge mapping," 
Humanities and Social Sciences Communications, vol. 
11, no. 1, pp. 1-11, 2024. DOI: 
https://doi.org/10.1057/s41599-024-03314-9 

[27] K. Massing, "Safeguarding intangible cultural heritage 
in an ethnic theme park setting–the case of 
Binglanggu in Hainan Province, China," International 
Journal of Heritage Studies, vol. 24, no. 1, pp. 66-82, 
2018. DOI: 
https://doi.org/10.1080/13527258.2017.1362571 

[28] Q. Yu, "Innovation and Promotion of Hainan Li Ethnic 
Intangible Cultural Heritage Tourism Products," 
Journal of Modern Business and Economics, vol. 1, no. 
1, 2024 DOI: 
https://doi.org/10.70767/jmbe.v1i1.133. 

[29] N. Partarakis, X. Zabulis, A. Chatziantoniou, N. 
Patsiouras, and I. Adami, "An approach to the creation 
and presentation of reference gesture datasets, for 
the preservation of traditional crafts," Applied 
Sciences, vol. 10, no. 20, p. 7325, 2020. DOI: 
https://doi.org/10.3390/app10207325 

[30] M. A. D. Mendoza, E. De La Hoz Franco, and J. E. G. 
Gómez, "Technologies for the preservation of cultural 
heritage—a systematic review of the literature," 
Sustainability, vol. 15, no. 2, p. 1059, 2023. DOI: 
https://doi.org/10.3390/su15021059 

[31]  Q. Li, "Intelligent intangible cultural heritage 
innovation platform under the background of big data 
and virtual systems," in 2022 Second International 
Conference on Artificial Intelligence and Smart 
Energy (ICAIS), 2022: IEEE, pp. 560-563. DOI: 
10.1109/ICAIS53314.2022.9742914. 

[32] M. A. D. Mendoza, E. De La Hoz Franco, and J. E. G. J. S. 
Gómez, "Technologies for the preservation of cultural 
heritage—a systematic review of the literature," vol. 
15, no. 2, p. 1059, 2023. DOI: 
https://doi.org/10.3390/su15021059 

[33] S.-N. Zhang, W.-Q. Ruan, and T.-T. J. S. O. Yang, 
"National identity construction in cultural and 
creative tourism: the double mediators of implicit 
cultural memory and explicit cultural learning," vol. 
11, no. 3, p. 21582440211040789, 2021. 

[34]  X. Gu, Y. Shen, and J. Xu, "Multimodal emotion 
recognition in deep learning: a survey," in 2021 
International Conference on Culture-oriented Science 
& Technology (ICCST), 2021: IEEE, pp. 77-82. DOI: 
10.1109/ICCST53801.2021.00027.  

[35] A. Rahate, R. Walambe, S. Ramanna, and K. J. I. F. 
Kotecha, "Multimodal co-learning: Challenges, 
applications with datasets, recent advances and 
future directions," vol. 81, pp. 203-239, 2022. DOI: 
https://doi.org/10.1016/j.inffus.2021.12.003 

[36] J. Tian and Y. J. I. T. o. C. S. S. She, "A visual–audio-
based emotion recognition system integrating 
dimensional analysis," vol. 10, no. 6, pp. 3273-3282, 
2022. DOI: 10.1109/TCSS.2022.3200060. 

[37] S. Fan, J. Jing, and C. J. S. Wang, "Audio-Visual Learning 
for Multimodal Emotion Recognition," vol. 17, no. 3, p. 
418, 2025. DOI: 
https://doi.org/10.3390/sym17030418 

[38]  J. He et al., "Cross-culture Continuous Emotion 
Recognition with Multimodal Features," in 
Proceedings of the 2019 8th International Conference 
on Computing and Pattern Recognition, 2019, pp. 
326-330. DOI: 
https://doi.org/10.1145/3373509.3373515 

[39]  T. Nourivandi, S. Aathreya, and S. Canavan, 
"Multimodal Behavior Analysis and Impact of Culture 
on Affect," in 2024 12th International Conference on 
Affective Computing and Intelligent Interaction 
(ACII), 2024: IEEE, pp. 37-45. 
DOI:10.1109/ACII63134.2024.00009. 

[40] A. S. Cardoso et al., "Deep learning assessment of 
cultural ecosystem services from social media 
images," p. 2021.06. 23.449176, 2021. DOI:  
https://doi.org/10.1101/2021.06.23.449176 

[41] A. Kioussi, A. Doulamis, M. Karoglou, A. I. J. I. J. o. A. 
Moropoulou, Culture, Design,, and Technology, 
"Cultural intelligence-investigation of different 
systems for heritage sustainable preservation," vol. 9, 
no. 2, pp. 16-30, 2020. DOI: 
10.4018/IJACDT.2020070102 

[42]  X. Xie, W. He, Y. Zhu, and H. Xu, "Performance 
evaluation and analysis of deep learning 
frameworks," in Proceedings of the 2022 5th 
International Conference on Artificial Intelligence and 
Pattern Recognition, 2022, pp. 38-44. DOI: 
https://doi.org/10.1145/3573942.3573948 

[43] O. Rainio, J. Teuho, and R. J. S. R. Klén, "Evaluation 
metrics and statistical tests for machine learning," 
vol. 14, no. 1, p. 6086, 2024. DOI: 
https://doi.org/10.1038/s41598-024-56706-x 

[44] F. J. E. Gîrbacia, "An Analysis of Research Trends for 
Using Artificial Intelligence in Cultural Heritage," vol. 
13, no. 18, p. 3738, 2024. DOI: 
https://doi.org/10.3390/electronics13183738 

[45] S. Li, Y. Jiang, B. Jing, L. Yang, and Y. J. J. o. C. H. Zhang, 
"AI-based experts’ knowledge visualization of cultural 
heritage: A case study of Terracotta Warriors," vol. 
72, pp. 81-90, 2025. DOI: 
https://doi.org/10.1016/j.culher.2025.01.006 

[46] S. O’Connor, S. Colreavy-Donnelly, and I. J. V. c. f. c. h. 
Dunwell, "Fostering engagement with cultural 
heritage through immersive vr and gamification," pp. 
301-321, 2020. DOI: https://doi.org/10.1007/978-3-
030-37191-3_16 

[47]  R. R. Selvaraju et al., "Squinting at vqa models: 
Introspecting vqa models with sub-questions," in 
Proceedings of the IEEE/CVF Conference on 
Computer Vision and Pattern Recognition, 2020, pp. 
10003-10011.  

[48] G. Kress, S. J. T. i. Selander, and h. education, 
"Multimodal design, learning and cultures of 
recognition," vol. 15, no. 4, pp. 265-268, 2012. DOI:  
https://doi.org/10.1016/j.iheduc.2011.12.003 

[49] S. Dargan and M. J. E. S. w. A. Kumar, "A 
comprehensive survey on the biometric recognition 
systems based on physiological and behavioral 
modalities," vol. 143, p. 113114, 2020. DOI: 
https://doi.org/10.1016/j.eswa.2019.113114 

[50] H. Mekonnen, Z. Bires, and K. J. H. S. Berhanu, 
"Practices and challenges of cultural heritage 
conservation in historical and religious heritage sites: 
evidence from North Shoa Zone, Amhara Region, 
Ethiopia," vol. 10, no. 1, p. 172, 2022. 



J. Sun et al. /Future Technology                                         August 2025| Volume 04 | Issue 03 | Pages 119-137                                                                                      

137 

 

[51] Z. He, C. J. H. Wen, and S. S. Communications, 
"Construction of digital creation development model 
of intangible cultural heritage crafts in China," vol. 11, 
no. 1, pp. 1-14, 2024. DOI: 
https://doi.org/10.1057/s41599-024-04331-4 

[52] M. Fiorucci, M. Khoroshiltseva, M. Pontil, A. Traviglia, 
A. Del Bue, and S. J. P. R. L. James, "Machine learning 
for cultural heritage: A survey," vol. 133, pp. 102-108, 
2020. DOI: 
https://doi.org/10.1016/j.patrec.2020.02.017 

[53] C. Lin, C. J. J. o. S. Li, and Ethnology, "Digital divide of 
intangible cultural heritage and innovative 
inheritance countermeasures," vol. 5, no. 11, pp. 110-
122, 2023. DOI: 10.23977/jsoce.2023.051115 

[54] P.-S. Magdalena, "Artificial intelligence in the context 
of cultural heritage and museums: Complex 
challenges and new opportunities," 2023. COI: 
20.500.12592/4cqcjb. 

[55] C. Meghini, V. Bartalesi, and D. J. S. W. Metilli, 
"Representing narratives in digital libraries: The 
narrative ontology," vol. 12, no. 2, pp. 241-264, 2021. 

[56] D. Buragohain, Y. Meng, C. Deng, Q. Li, and S. J. H. S. 
Chaudhary, "Digitalizing cultural heritage through 
metaverse applications: challenges, opportunities, 
and strategies," vol. 12, no. 1, p. 295, 2024. 

[57] S. Münster et al., "Artificial intelligence for digital 
heritage innovation: Setting up a r&d agenda for 
europe," vol. 7, no. 2, pp. 794-816, 2024. DOI: 
https://doi.org/10.3390/heritage7020038 

[58] H. T. A. Eyadah and A. A. J. H. Odaibat, "A Forward-
Looking Vision to Employ Artificial Intelligence to 
Preserve Cultural Heritage," vol. 12, no. 5, pp. 109-
114, 2024. DOI: 
https://doi.org/10.11648/j.hss.20241205.12 

 

  

This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

https://creativecommons.org/licenses/by/4.0/