The analysis of bidirectional long short-term memory network model for construction of cultural gene map and information extraction

Abstract

To create a cultural gene map and extract information, this paper introduces a two-way long and short-term memory network (LSTM) model and verifies it using Jinxiu Yao headwear as an example. In this paper, a thorough cultural gene map, comprising structure, pattern, and color maps, is successfully constructed using the Bidirectional Long Short-Term Memory Network (Bi-LSTM) model. Experimental results show that the proposed model exhibits excellent performance in the information extraction task. On the test set, its Accuracy is 0.94, Precision is 0.92, Recall is 0.91, F1-score is 0.92, and AUC is 0.91. Compared with traditional models including LSTM, Gated Recurrent Unit Network (GRU), Convolutional Neural Network + Long Short-Term Memory Network (CNN-LSTM), Transformer, and Bidirectional Encoder Representations from Transformers + Bidirectional Long Short-Term Memory Network (BERT-BiLSTM), the proposed model performs significantly better. The model demonstrates outstanding performance in processing complex sequence data. Meanwhile, it can efficiently capture cross-modal and multi-dimensional cultural information, providing strong data support for the digital research and protection of traditional cultural works. This shows that the cultural gene information extraction method has broad application prospects. In the future, the model performance can be further optimized and the application scenarios can be expanded.

Introduction

Research background and motivations

As an important tool for recording and inheriting cultural heritage, cultural gene mapping has received extensive attention in recent years^1–3. Cultural gene mapping can show the relationship between cultural elements, and help researchers to deeply understand the evolution process of culture^4,5. With the development of digital technology, the construction method of cultural gene map is also evolving^6,7. However, the traditional construction method of cultural gene map has some limitations in the accuracy and efficiency of information extraction, and it is difficult to fully capture the complexity of cultural information^8–10.

Natural language processing and information extraction have benefited greatly from the use of deep learning technologies, particularly long and short-term memory networks (LSTM) in recent years^11,12. The model is better at handling complicated text data because of the Bidirectional Long Short-Term Memory Network (Bi-LSTM), which further improves the capacity of the model to capture contextual information^13–15. Thus, it is anticipated that the incorporation of the Bi-LSTM model into the creation of the cultural gene map will address the drawbacks of conventional techniques and enhance the precision and effectiveness of cultural information extraction^16–18.

The main goal of this paper is to explore the method of constructing more accurate cultural gene map by using Bi-LSTM model. Specifically, the purpose of this paper is to propose a new methodology that combines deep learning technology (especially Bi-LSTM model) with cultural heritage research to construct a more accurate and comprehensive cultural gene map. Through specific case studies (such as headwear culture in Jinxiu Yao autonomous county), the effectiveness of the proposed method is verified. The verification process includes detailed analysis and comparison with existing methods to show the advantages of Bi-LSTM model. By providing a powerful framework to promote the digital protection and inheritance of cultural heritage, the framework can be applied to various cultural works of art and practice. This paper provides new insights and practical tools for researchers and practitioners in the fields of cultural heritage and digital humanities, thus promoting the understanding and protection of cultural heritage.

Research objectives and contributions

The main objectives of this paper are as follows: (1) Construct the model of cultural gene map: Using Bi-LSTM model, developing a new method for constructing cultural gene map. Through this model, cultural gene information can be extracted more accurately from the text and displayed systematically in the cultural gene map. (2) Verify the validity of the model: Taking Jinxiu Yao headwear as a research case, verifying the practical application effect of the Bi-LSTM model in the construction of cultural gene map. Through the in-depth analysis of this case, the performance of the model is evaluated and its applicability in other cultural cases is discussed. (3) Promote the digital protection of cultural heritage: Through the construction method of proposed cultural gene map, promote the digital protection process of cultural heritage, and provide technical support and theoretical basis for subsequent research.

The main contributions and innovations of this paper are as follows:

1. Propose a new framework for constructing cultural gene maps based on Bi-LSTM. This framework is not a simple naming replacement for traditional knowledge graphs or semantic networks, but rather an extension and innovation in graph structure, attribute definition, and dynamic feature modeling. At the structural level, the nodes of the cultural gene map represent semantic entities, and contain multimodal features such as the form, pattern, and color of cultural symbols, constructing a three-layer composite network of “semantic visual emotional”. In terms of defining attributes and weights, the Bi-LSTM model is used to capture the temporal dependencies between cultural elements and dynamically update the associated weights between nodes, so that the connections between cultural genes can reflect changes in time and context. In dynamic feature modeling, an adaptive weight update mechanism based on sequence learning is introduced to enable the graph to have the dynamic plasticity of complex networks. As a result, the systematic, refined, and adaptive extraction of cultural gene features has been achieved, significantly enhancing the intelligent processing capability of cultural heritage data.

2. Verify the effectiveness of the Bi-LSTM model in extracting cultural gene information. Taking the Jinxiu Yao headwear as a case study, this paper deeply analyzes the advantages of the model in capturing the complex temporal and contextual dependencies between cultural elements. The experimental results show that the proposed model is significantly superior to traditional models in terms of accuracy, recall, F1 score, and other indicators, verifying its robustness and practical application value.

3. Expand the application scope of the model. The cultural gene map of this paper is not only applicable for extracting cultural information from Yao headwear, but can also be extended to the digital modeling and analysis of various types of ethnic cultural elements like clothing, architecture, and artifacts, providing a scalable technical path for the intelligent protection and display of cultural heritage.

4. Promote the digital protection and intelligent research of cultural heritage. The constructed cultural gene map can achieve structured storage, visual display, and semantic association of cultural resources, support cultural dissemination, innovative design, and cross-cultural comparative research, and promote the systematic, data-driven, and intelligent development of cultural heritage research.

Literature review

Cultural gene mapping has drawn a lot of attention as a novel research technique in the field of cultural heritage studies¹⁹. Cultural gene map can intuitively show the path of cultural inheritance and evolution through the systematic expression of cultural elements and their relationships. The early study of cultural gene mapping mainly relied on artificial construction and traditional statistical methods^20–22. For example, the construction method of cultural gene map proposed by Liu (2024)²³ relied on manual extraction of cultural elements and mining the relationship between these elements through association rules. From the perspective of geographical information, Hu et al. (2024)²⁴ analyzed the elements and contents of traditional settlement cultural landscape genes. Although this method has realized the systematic expression of cultural information to a certain extent, it is highly dependent on labor and inefficient, and it is difficult to deal with the processing of large-scale datasets.

As natural language processing technology has advanced recently—particularly with the widespread use of deep learning models—researchers have started experimenting with using these tools to create cultural gene maps. Convolutional neural network (CNN) was utilized by Janković Babić (2024)²⁵ to automatically extract cultural gene information from cultural texts and create the matching map model. Wei (2022)²⁶ used artificial neural network-cellular automata model to study the gene construction of cultural landscape in ancient towns in Shaanxi Province, and analyzed the spatial layout and landscape pattern index of rural settlements in Shaanxi Province. Although this method improved the automation of information extraction to a certain extent, its ability to capture contextual information was limited, which led to insufficient performance in dealing with complex cultural contexts.

As one of the core tasks of natural language processing, information extraction has developed a variety of technical means^27–29. Traditional information extraction methods, such as named entity recognition and relationship extraction, usually rely on rules or shallow machine learning algorithms. These methods perform well in dealing with text information with simple structure, but they are prone to lack of accuracy when facing complex and diverse cultural texts^30–32.

The deep learning model has steadily taken the lead in information extraction techniques in recent years. Because they are particularly good at capturing the contextual information found in texts, the LSTM and its version Bi-LSTM were frequently utilized in tasks including text categorization, sentiment analysis, and sequence labeling^33,34. For instance, Ayetiran (2022)³⁵ greatly increased the accuracy of sentiment categorization by using the Bi-LSTM model for sentiment analysis of cultural materials. A remote supervision relationship extraction technique based on Bi-LSTM was presented by He et al. (2019)³⁶. Li et al. (2023)³⁷ studied the influence of the development of regional cultural digital finance on corporate financing constraints.

In addition, although the Bi-LSTM model has been widely applied in general text analysis tasks, its specific application in the field of cultural heritage is still in the exploration stage. In recent years, existing studies have begun to attempt the application of Bi-LSTM in the digitization and protection of intangible cultural heritage. For instance, Quan et al. (2024)³⁸ used Bi-LSTM to extract and classify descriptions of traditional handicrafts from historical documents, which improved the efficiency and accuracy of cultural data annotation. Sun et al. (2025)³⁹ combined Bi-LSTM with multi-modal data, including images and text descriptions of folk rituals, to construct a comprehensive digital representation of intangible cultural heritage. Yuan et al. (2025)⁴⁰ introduced an intelligent question-answering system based on the knowledge graph in the field of ancient Chinese costume heritage to improve the accuracy of information retrieval. These studies indicate that Bi-LSTM can effectively capture complex sequence information, and realize multi-modal fusion of cultural data, providing a feasible path for constructing detailed and information-rich cultural gene maps.

To sum up, although the existing research on cultural gene maps has made some progress in methodology, there are still the following limitations: the existing methods often ignore the complexity and multi-level of cultural information when dealing with cultural texts, resulting in the lack of depth and breadth of cultural gene maps. Although the introduction of deep learning model improves the automation of information extraction, its application in cultural gene mapping is still not mature enough, and there are some problems such as difficult model tuning and high computational resource consumption. The restricted application of the model in developing cultural gene maps is attributable to the prevalent focus in current research on improving model performance, with less attention devoted to understanding and integrating cultural context.

Although some progress has been made in the construction of cultural gene maps in existing research, most methods still have static and unimodal limitations, making it difficult to fully reflect the complexity and hierarchy of cultural information. The cultural gene map proposed here is innovative based on systems science and complex network theory. Nodes not only represent semantic entities, but also integrate multimodal information such as form, pattern, and color. The weight of edges can be dynamically updated through text co-occurrence frequency and visual similarity calculation, achieving the evolutionary representation of cultural information. The Bi-LSTM model is sued for sequence feature extraction and combined with the Conditional Random Field (CRF) layer for structured output, and then fused multimodal features of images. The model can capture contextual dependencies and cross modal associations of cultural information, constructing a three-layer composite network of “semantic - visual - historical”. This design enables the cultural gene map to inherit the framework of knowledge graphs and complex networks in theory, while also being innovative in node attributes, multimodal fusion, and dynamic evolution mechanisms. It provides a systematic methodological foundation for the intelligent representation and digital protection of cultural data. To verify the effectiveness of Bi-LSTM model in practical application, this paper chooses the headwear of Jinxiu Yao autonomous county as a specific case. Through detailed analysis and experiments, it is hoped to show the advantages of this model in cultural information extraction and cultural gene map construction. Finally, this paper aims to provide a new method for the study of cultural gene map and the digital protection of cultural heritage, and promote the further development of related fields.

Research methodology

Principle and structure of Bi-LSTM model

The structure of Bi-LSTM model includes input layer, bidirectional LSTM layer and output layer^41–43. In one-way LSTM, the hidden state at the current moment t depends on the hidden state at the previous moment and the current input . The core of LSTM is to introduce a “memory cell” to store long-term information, and control the flow of information through three gating mechanisms: input gate, forget gate and output gate. Its calculation follows the standard LSTM gating mechanism.

In Bi-LSTM, there are two hidden states in each time step, one is the hidden state of forward LSTM and the other is the hidden state of backward LSTM . The output of Bi-LSTM is the connection of these two hidden states. It is expressed by Eq. (1):

1

in Eq. (1) represents the final hidden state at time step t of bidirectional LSTM, which consists of forward and reverse hidden states. indicates the hidden state of the forward LSTM at time step t, considering the information from the beginning of the sequence to the current position. indicates the hidden state of the reverse LSTM at time step t, considering the information from the end of the sequence to the current position.

In the operation mechanism of Bi-LSTM, first, the input sequences x₁,x₂,…,x_n are passed to both the forward LSTM and the backward LSTM. Forward LSTM processes the input sequence from left to right to generate the hidden state of each time step. Backward LSTM processes the input sequence from right to left and generates the hidden state of each time step. Finally, these two hidden states are connected to form the output of Bi-LSTM.

In the process of constructing a cultural gene map, cultural texts often have complex contextual dependencies, metaphorical expressions, and potential semantic associations. This makes it difficult for shallow models to fully capture the deep semantic structures within them. Therefore, this study introduces Bi-LSTM as a semantic encoder to extract cultural gene information more accurately by simultaneously learning the forward and backward dependencies of the text. For example, when processing text describing the headwear of Jinxiu Yao Autonomous County, forward LSTM can capture the form and material characteristics of the headwear, while backward LSTM can trace its cultural symbolic meaning and historical background, thus outputting a comprehensive representation that combines literal and cultural semantics.

These mechanisms improve the accuracy of cultural feature extraction, and lay the structural foundation for the node edge relationship of the “cultural gene map” - where nodes represent semantically extracted cultural elements, and edges reflect contextual or symbolic correlations. Therefore, the Bi-LSTM model proposed differs from traditional application forms in terms of task adaptation mechanism. By introducing domain-based semantic adaptation and hierarchical knowledge integration mechanisms, this model better meets the construction needs and theoretical goals of cultural gene maps. The specific construction process of the knowledge graph will be explained in detail in the following text.

Construction and preprocessing of dataset

In the stage of data collection, this paper uses a variety of cultural resources, including various texts, ethnographic research reports and historical documents related to national headgear costumes. These data are not limited to specific cultural symbols, but also cover various practical contents related to cultural heritage. Through this method, people can collect text information related to different types of cultural heritage, such as information related to traditional art, historical events, social customs, handicrafts and festivals. This multi-dimensional collection can lay the foundation for building a more comprehensive cultural gene map.

Firstly, the scope of data collection is not limited to the description of cultural symbols. The text data used include the production technology of culture and art, the narrative of historical background, local customs and habits, the tradition of festival activities, and even the information about the inheritance of traditional knowledge and skills. For example, the production process of handicrafts, the styles and design concepts of traditional costumes, rituals and habits in festivals and celebrations can all be collected and analyzed through text or image data. These data can help to fully understand all aspects of a cultural system, not just superficial symbols. Secondly, the scope of data collection includes different types of cultural heritage, including but not limited to traditional cultural heritage, intangible cultural heritage and even modern cultural elements. For example, literature and oral historical data about local drama, folk dance, traditional food production techniques and other fields can provide rich original data for cultural gene mapping. In addition, modern cultural phenomena, such as the schools of contemporary art and the changes of urban culture, can also be collected and analyzed by this method, thus providing a new perspective for cultural protection and innovation.

At the same time, this data collection method is suitable for specific cultural forms, and has strong cross-cultural applicability. Through the same model and method, a similar analysis can be conducted on cultural heritage from different regions and cultural backgrounds. For example, traditional crafts in Southeast Asia, historical architectural styles in Western Europe and even tribal costumes in Africa can all use the same model to collect data, extract information and construct cultural gene maps. This method can be widely used in the protection of cultural heritage around the world, especially in multi-cultural and multi-regional research projects, with good flexibility and adaptability.

In this way, this paper provides methodological support for the protection and research of this kind of culture represented by Jinxiu Yao headwear and a feasible technical path for the research and protection of other cultural fields. The following actions are included in data preprocessing. Each is crucial and has an immediate impact on the performance of the final model. The procedure is depicted in Fig. 1:

Fig. 1.

Data preprocessing flow.

In Fig. 1, text cleaning is performed first. In the original text data, there are a lot of noise data, such as punctuation marks, special characters, web page format information and so on. This useless information may interfere with the learning process of the model. Therefore, firstly, the text is cleaned by regular expressions and text processing tools. Some text paragraphs may contain redundant descriptive or repetitive content, which is not beneficial to the training of the model and needs to be deleted manually or automatically.

Then word segmentation and part-of-speech tagging are carried out. Because there is no obvious word boundary in Chinese, word segmentation is the key step of text preprocessing. Use dictionary-based word segmentation tool (Jieba word segmentation) to segment the text. Combined with manual correction, the accuracy of word segmentation results is ensured. After word segmentation, part-of-speech tagging is carried out for each word to help the model better understand the grammatical roles of words. For example, “Yao” in “the headwear of Yao” is marked as a noun, “de” as an auxiliary word, and “headwear” as a noun.

Then, every word in the text is mapped into a high-dimensional vector by using Word2Vec word vector model. Each word vector has 200 to 300 dimensions, which can capture the semantic relationship among words. In the process of word vector generation, you may encounter words that do not appear in the dictionary. For these words, word vectors are generated by averaging their contextual word vectors, or domain-specific word vector models are introduced to supplement them.

To increase the diversity of data, several equivalent text samples are generated by synonym replacement. For example, replacing “headwear” with “hat wear” can increase the richness of the dataset. On the premise of not changing the semantics, some words are randomly inserted or deleted, thus generating new training samples. By using this technique, the resistance of the model to various expressions can be strengthened.

Then the preprocessed dataset is divided into training set (70%), verification set (15%) and test set (15%). The training set is used for model training, the verification set is used for model parameter tuning and performance monitoring, and the test set is used to evaluate the final effect of the model. In order to further improve the generalization ability of the model, K-fold cross-validation method can be used to divide the dataset into K subsets, and each time K-1 subsets are used for training, and the remaining subset is verified.

Those procedures can reduce noise in the dataset and enhance the text comprehension of the model. In practical application, steps such as text cleaning and word segmentation can be handled automatically through scripts, while steps such as part-of-speech tagging and data enhancement may require some manual intervention to ensure data quality. The preprocessed dataset improves the training efficiency of Bi-LSTM model, and provides richer and more accurate cultural semantic information for the model, which is helpful to the construction of the final cultural gene map.

Feature extraction and multimodal alignment of images

To comprehensively capture multi-dimensional information in cultural heritage, this paper introduces image feature extraction and multi-modal alignment methods based on text feature extraction to achieve effective fusion of images and text. For the collected images of Jinxiu Yao headwear, image preprocessing is first performed, including size normalization, pixel value standardization, and noise removal. Subsequently, a convolutional neural network (CNN) is used to extract visual features. Through convolutional layers, pooling layers, and fully connected layers, CNN gradually extracts the low-level texture features (such as lines and color blocks), middle-level local structural features (such as the combined form of patterns), and high-level semantic features (such as decorative styles and pattern types) of the images. The dimension of the feature vector obtained for each image is , which is used for the subsequent fusion with text features.

Text features are extracted by the Bi-LSTM model, and each word or phrase generates a hidden state vector . To achieve multi-modal alignment, text features and image features are mapped to the same feature space. Through the transformation of the fully connected layer, the image feature is projected to a vector space with the same dimension as the text feature. The attention mechanism is used to calculate the similarity weight between the image and the text. The specific equation is as follows:

2

denotes the hidden state of the i-th word in the text. represents the j-th region or feature vector of the image. stands for the matching weight between the text and image features. Based on the calculated matching weights, the image features and text features are weighted and fused to obtain a comprehensive representation vector:

3

This vector not only retains the semantic information of the text but also integrates the visual features of the image, enabling it to more accurately represent the characteristics of cultural elements. The fused multi-modal features are used for pattern recognition and classification to further label cultural gene nodes. For example, the fused features are classified via Softmax to determine whether the features belong to a specific cultural element or pattern type. The identified pattern information serves as nodes in the cultural gene map, and relational edges are established between these nodes and other nodes, forming a complete multi-modal cultural gene map.

After multimodal alignment and feature fusion, in order to eliminate redundant information, reduce the computational complexity of high-dimensional vectors, and highlight key cultural semantic features, Principal Component Analysis (PCA) is performed on the fused vector for dimensionality reduction. The specific operation is as follows:

1. Construct a matrix of all fusion vectors : , with each row corresponding to a multimodal feature of a text or image unit.

2. Perform mean centering on matrix H and calculate the covariance matrix

   4

3. Perform eigenvalue decomposition on the covariance matrix C, take the first d principal components, and obtain the reduced dimension mapping matrix , where D is the original vector dimension and d = 128 is the reduced dimension.

4. Map each fusion vector to a low dimensional space to obtain the final vector used for node representation, expressed as:

   5

Through this dimensionality reduction process, over 90% of the feature variance is preserved, ensuring the preservation of core cultural semantic information while significantly reducing redundant information. The reduced dimensional vectors are directly used for calculating node weights and generating graphs in the cultural gene map. Edge weights are calculated in the fusion vector space, while text co-occurrence frequency and semantic similarity are combined to quantify the degree of cultural association between nodes.

Construction of cultural gene map and information extraction

The construction of cultural gene map is one of the core goals of this paper. Through the powerful text processing ability of Bi-LSTM model, representative cultural genes are extracted from complex cultural texts and systematically organized into a map structure. This cultural gene map can offer a fresh viewpoint on the inheritance and study of cultural heritage in addition to demonstrating the relationships between cultural components. To apply the sequence feature extraction capability of the Bi-LSTM model to the construction of cultural gene maps, this paper introduces a feature fusion mechanism, integrating multi-dimensional cultural information such as structure, pattern, and color into a unified representation. Each text or image segment is first processed forward and backward by Bi-LSTM to generate the corresponding hidden state vector. Then, the vectors of different modalities are merged through the feature fusion layer, forming a comprehensive representation that includes contextual semantics, image visual features, and color information. This comprehensive representation retains the semantic information of the text and takes into account the visual features of images and symbols. It enables the model to accurately identify cultural gene nodes and construct relational edges, thereby realizing the automatic generation of cultural gene maps. Through this mechanism, the Bi-LSTM model can better adapt to the task of cultural gene map construction. This model has obvious advantages in processing cultural texts with long sentences or complex structures, while supporting the comprehensive analysis of cross-modal and multi-dimensional information.

Firstly, through the trained Bi-LSTM model, the preprocessed cultural text data is analyzed and the cultural gene information is automatically extracted. This cultural gene information is expressed as keywords or phrases related to cultural characteristics. For example, in the description of Jinxiu Yao headwear, words with cultural symbolic significance such as “dragon and phoenix ornamentation”, “silverware craft” and “longevity lock” will be extracted.

However, the construction of cultural gene map is not limited to the extraction of a specific cultural element. The Bi-LSTM model can handle a variety of cultural data, including words, images, historical documents, ethnographic studies and other multi-dimensional content. This makes the model suitable for the extraction of traditional cultural heritage, and able to deal with modern cultural phenomena and cultural genes in cross-cultural background. For example, text data related to modern art, popular culture, language evolution and other fields can be processed in a similar way and integrated into the cultural gene map, thus providing a comprehensive understanding of contemporary cultural diversity.

In addition, the construction method of cultural gene map has strong interdisciplinary and extensive application scope. This is not only applicable to the study of a single culture or region, but also covers the content of cross-cultural comparison. In this way, cultural heritages under different cultural backgrounds can be compared and analyzed. For example, the similarities and differences of traditional costumes, crafts and social customs in different regions can be studied and displayed in the form of maps to help researchers understand the diversity and connection of global culture from a more macro perspective. Therefore, this construction method of cultural gene map is of great value in the protection and inheritance of traditional culture, and can promote the application of digital technology in cultural research and provide new ideas and tools for future cultural heritage protection, cultural exchange and cross-cultural understanding.

The specific extraction process is as follows:

Firstly, the preprocessed text is input into Bi-LSTM model. Suppose the input text as shown in Eq. (6):

6

Each represents a word vector of a word or phrase. Bi-LSTM model performs bidirectional processing on the input text to generate the hidden state of each time step, as shown in Eq. (7):

7

In Eq. (7), is the connection of the hidden state of the forward and backward LSTM. After obtaining the hidden states, each hidden state is further classified through a Conditional Random Field (CRF) layer to determine whether it is a cultural gene node. The motivation for introducing CRF lies in its ability to capture the dependency relationships and contextual connections between adjacent words in the sequence. Compared with using a softmax classifier alone, CRF can jointly model structured outputs, thereby improving the accuracy of sequence labeling and reducing incorrect labeling that may occur in complex cultural texts. The words labelled as cultural genes are then extracted as nodes of the cultural gene map, as shown in Eq. (8).

8

In Eq. (8), is the label of the i-th word (whether it is a cultural gene or not). is the weight matrix of the classifier, and is the bias vector. After the cultural gene information is extracted, it is systematically organized into a map structure. This map can be thought of as a directed graph, where each cultural gene is represented by a node and the relationships between these genes are indicated by the edges connecting the nodes. Each extracted cultural gene is regarded as a node in the map. For example, the extracted keywords such as “dragon and phoenix decoration”, “silver decoration technology” and “longevity lock” are added to the map as nodes respectively. The edges in the map represent the relationship between cultural genes. The value of edge weight reflects the closeness of the association between nodes. By comprehensively calculating the text co-occurrence frequency and visual similarity in the training data, a weighted directed graph is generated. A higher weight value of each edge indicates a stronger cultural association between the nodes. For example, if “dragon and phoenix ornamentation” and “silver decoration technology” frequently appear at the same time in multiple documents, a border is established between them in the cultural gene map, indicating the close connection between these two cultural elements in cultural inheritance.

In order to measure the strength of edges, each edge is given a weight, which indicates the close degree of correlation between cultural genes. Weight calculation considers the co-occurrence frequency of words in the text, and combines image visual feature similarity to achieve multimodal comprehensive measurement. After multimodal feature fusion, each node vector has become a comprehensive representation of cultural semantics, integrating text contextual information with image visual features. The weight calculation method is shown in Eq. (9):

9

In Eq. (9), is the weight of the edge between node i and node j. indicates the number of times or similarity of the words and in the text. An interpretable example is used to illustrate the calculation process of node weight Through this method, edge weighting reflects the simultaneous occurrence of words in the text, and quantifies the supplement of visual and semantic information to cultural semantic relationships. where represents the association strength between cultural element i and cultural element j. Taking the decorative silver hairpin (node i) and the crown type (node j) as an example: if they co-occur 15 times in text and images, and the semantic similarity is 0.8, then the node weight =15 × 0.8 = 12. This indicates that the two have a strong association in cultural inheritance.

Systematic research and protection of cultural heritage can be realized by constructing cultural gene map. For example, by analyzing the cultural gene map of Jinxiu Yao headwear, people can understand the cultural symbol and historical background behind the headwear, thus providing important reference for cultural inheritance and innovative design. In addition, the cultural gene map can also be used as a tool for cultural research to help researchers explore the deep-seated relationship between different cultural elements.

Experimental design and performance evaluation

Datasets collection, experimental environment and parameters setting

The dataset used in this paper covers multiple cultural component information to support the training and evaluation of Bi-LSTM model. The dataset covers text data and visual data, mainly focusing on Jinxiu Yao headwear, and expanding to other traditional ethnic ornaments to verify the applicability and generalization ability of the model. The dataset is integrated from multiple sources, including traditional cultural research literature, national craft database, open data of cultural heritage protection institutions and online handicraft platform. Covering cultural background description, headwear making technology, design concept, color implication and wearing scene, data labels are divided into six categories: structure, pattern, color, material, technology and cultural implication. The average length of text data is between 120 and 200 characters, mainly in modern Chinese, including some classical Chinese and dialect texts, the latter has been transliterated and modernized. The aim is to enhance the diversity of data and the wide applicability of research. The dataset is expanded. It includes Jinxiu Yao headwear. It also includes data of other ethnic ornaments, such as Miao silver ornaments and Tibetan headwear. This is to verify the generalization ability of the model in different cultural backgrounds.

In data preprocessing, Jieba is used for word segmentation and part-of-speech tagging, Word2Vec is used for text feature extraction, and the semantic vectorization method is combined to improve the text understanding ability of the model. The visual data is about 10,000 high-definition images, covering the overall structure, local details and wearing effect of the headwear. The minimum resolution is 512 × 512 pixels, and some of the data reaches 4 K level. In order to improve the diversity of data and the generalization ability of the model, data enhancement operations such as rotation, scaling, color transformation and contrast enhancement are carried out on the image data. The generation countermeasure network is introduced to expand the data to enhance the adaptability of the model under different lighting conditions and shooting angles. The number of original images is approximately 10,000, and the data volume is expanded to about 50,000 after augmentation. In the experiment, the augmented data is used in the model training phase, while the validation set and test set retain the original non-augmented data to ensure the objectivity of performance evaluation. The annotation of image data covers the structure (such as tassels, hairpins, crowns and hats), patterns (such as dragons and phoenixes, auspicious clouds and flowers), colors (such as red, gold, silver and blue) and materials (such as metal, fabric and pearl jade) of headwear to ensure that the model can be accurately identified and classified. OpenCV and Pillow are used for image processing in data preprocessing, and TensorFlow and PyTorch are used for deep learning modeling to give full play to the advantages of Bi-LSTM model in sequence data processing and improve the accuracy and stability of cultural gene information extraction.

In data division, the dataset is divided into training set (70%), verification set (15%) and test set (15%). The training set is used for model learning, the verification set is used for adjusting hyperparameters, and the test set is used for evaluating the performance of the final model. To prevent data leakage, text data is tested by K-fold cross-validation for many times to ensure the robustness of experimental results. To ensure the efficient training of the model on large-scale datasets, experiments are run on a high-performance computing platform. The specific configuration is as follows: GPU: NVIDIA Tesla 100(32GB HBM2 memory), CPU: Intel Xeon Gold 6230(20 cores 2.1 GHz). Memory: 128GB DDR4, storage: NVMe SSD 2 TB (for accelerating data loading and model storage), operating system: Ubuntu 20.04 LTS (high stability and parallel computing optimization). Development environment: Python 3.8 + TensorFlow 2.8/Pytorch 1.10. Dependency libraries: Pandas (data processing), NumPy (numerical calculation), Matplotlib (visualization), Jieba (Chinese word segmentation and part-of-speech tagging), OpenCV (image processing), etc. In the training process, the model uses Mixed Precision Training to accelerate the training, and uses GPU to accelerate optimization to reduce the calculation delay. The high stability of Linux environment makes the system more reliable in the process of large-scale data processing and deep learning and training. During the experiment, reasonable parameter setting is the key to ensure the performance of the model. Table 1 presents the main parameter settings of Bi-LSTM model in this paper and their reasons.

Table 1.

Parameter settings.

Parameter name	Value
Learning rate	0.001
Batch size	64
Hidden layer dimension	128
Optimizer	Adam
Iterations	50
Dropout rate	0.5

Performance evaluation

To assess the effectiveness of the Bi-LSTM model in creating a cultural gene map, assessment indices such as accuracy, precision, recall, and F1 score are employed. The error of the model is calculated using the cross-entropy loss function. The probability distribution that the model predicts is closer to the real distribution the lower the loss value.

10

is the actual label and is the probability of model prediction. The Bi-LSTM model is compared with other similar models in many aspects, including the traditional Long Short-Term Memory (LSTM) model, Gated Recurrent Unit (GRU) model and Convolutional Long Short-Term Memory (CNN-LSTM) combined model.

Considering that the dataset contains some classical Chinese and dialect texts, this study pays special attention to the model’s performance on these special texts. During the training process, Dropout is set to 0.5 to reduce the risk of overfitting, especially when there are few non-standard text samples. Figure 2 shows the validation set loss curves of different text types, which allows for an intuitive observation of the model’s stability and learning efficiency in modern Chinese, classical Chinese, and dialect texts. The results indicate that the loss of modern Chinese texts decreases the fastest and finally converges to approximately 0.36, demonstrating that the model achieves the best learning effect on modern Chinese corpora. Due to the more complex sentence structures and vocabulary of classical Chinese texts, their overall loss is higher than that of modern Chinese, and finally stabilizes at around 0.45. This shows that the model faces certain difficulties in processing classical Chinese. The loss of dialect texts is the highest, with the minimum value being about 0.50, reflecting that the sparse vocabulary and diverse expressions in dialects increase the challenges for model learning. Overall, the three loss curves gradually become stable as the number of training epochs increases, which indicates that the model training process is convergent and stable. At the same time, it also intuitively reflects the differences in the impact of different text types on model performance. The model’s performance on classical Chinese and dialect texts decreases slightly, mainly due to sparse vocabulary and complex language structures. However, the model still maintains high accuracy on the whole, which demonstrates the robustness and generalization ability of the Bi-LSTM model on diverse texts.

Fig. 2.

Loss curve of verification set with different text types.

When trained with original data, the model achieves an accuracy of 0.91 and an F1-score of 0.89 on the test set. After data augmentation, the scale of the training set is increased to approximately 5 times its original size, and the model performance is significantly improved, with accuracy rising to 0.94 and F1-score increasing to 0.92. This comparison clearly demonstrates the positive effect of data augmentation on the model’s generalization ability and cultural gene extraction performance. To comprehensively evaluate the performance of the proposed Bi-LSTM + CRF + multi-modal + node weight model in the cultural gene information extraction task, it is compared with traditional models including LSTM, GRU, CNN-LSTM, as well as recent models such as Transformer and Bidirectional Encoder Representations from Transformers + Bidirectional Long Short-Term Memory Network BERT- (BiLSTM). The results are shown in Fig. 3:

Fig. 3.

Comparison of performance indexes of each model.

From Fig. 3, the proposed model outperforms other models significantly in all evaluation metrics. Its accuracy reaches 0.94, precision 0.92, recall 0.91, F1-score 0.92, and AUC 0.91. These values are obviously higher than those of the traditional LSTM (accuracy 0.82) and GRU (accuracy 0.84), and the gaps with Transformer and BERT-BiLSTM also reach 6% and 5% respectively, showing remarkable advantages. This superiority mainly benefits from several aspects: First, the bidirectional information processing capability of Bi-LSTM enables the model to capture both forward and backward contextual information of sequences simultaneously, improving the understanding of complex cultural texts. Second, the CRF layer enhances the constraint ability on label sequences, ensuring the coherence and accuracy of outputs. Third, multi-modal feature fusion allows the model to comprehensively analyze text and image information, making the extraction of key cultural elements (such as decorative patterns, colors, and structures) more comprehensive. Finally, the node weight mechanism strengthens key cultural information, making the generated cultural gene map more accurate and interpretable. Overall, the proposed model leads in precision and is significantly superior to other methods in the comprehensiveness of information capture and the interpretability of the map. It provides strong technical support for the digitization of cultural heritage and cross-cultural analysis.

To verify the effectiveness of each component of the model, an ablation experiment is designed. Key modules are removed one by one from the Bi-LSTM + CRF + multi-modal + node weight model, and the model performance is compared. The results are shown in Table 2.

Table 2.

Ablation experimental results.

Model variant	Accuracy	Precision	Recall	F1 score	AUC
Complete model Bi-LSTM + CRF + multi-modal + node weight	0.94	0.92	0.91	0.92	0.91
Remove CRF	0.91	0.89	0.87	0.88	0.88
Remove node weight	0.90	0.88	0.86	0.87	0.87
Remove multi-modal features	0.89	0.87	0.85	0.86	0.85

From Table 2, removing any key module will lead to a decline in model performance. After removing the CRF layer, the model’s accuracy drops to 0.91 and the F1-score drops to 0.88, which indicates that the label sequence constraint is crucial for improving the coherence and accuracy of outputs. After removing the node weight, the accuracy decreases to 0.90 and the F1-score decreases to 0.87, demonstrating that the node weight plays a significant role in strengthening the extraction of key cultural information. After removing the multi-modal features, the accuracy falls to 0.89 and the F1-score is 0.86, which shows that the integration of text and image information is indispensable for the extraction of complex cultural elements. Overall, the complete model achieves the highest level in accuracy, precision, recall, and F1-score. The ablation experiment fully verifies the contribution of each module to the improvement of model performance, and further reflects the robustness and reliability of this method in the construction of cultural gene maps. Figure 4 is the comparison of model training time.

Fig. 4.

Comparison of training time of each model.

From the training time in Fig. 3, GRU model has the shortest training time of 3.0 h, with an average time of 3.8 min per epoch, which is suitable for scenes requiring rapid training. The standard LSTM model requires 3.5 h for training, with an average epoch length of 4.2 min. On the other hand, the proposed model requires 3.8 h for training, with an average time of 4.8 min each epoch. This is marginally longer than the regular LSTM model, but it offers clear performance benefits. The training time of CNN-LSTM model is the longest, which is 4.0 h, and each epoch takes 5.0 min on average, which shows the increase of its computational complexity. This phenomenon can be attributed to Bi-LSTM’s more efficient learning ability when dealing with complex information. In contrast, GRU has the shortest training time, but its performance is not comparable to that of the proposed model, indicating that rapid training may not necessarily bring better model performance. Although the training time of the proposed model has increased, in the long run, it is more accurate and reliable in the task of cultural gene extraction, which makes it more widely applicable in the fields of cultural heritage digitization and automatic analysis. Although Transformer and BERT-BiLSTM models can also achieve high performance in some tasks, their training time is significantly longer, each training epoch takes 5–6 min, and the total training time reaches 4.5–5 h. In contrast, while maintaining the highest performance, the proposed model has a total training time of only 3.8 h, with 4.8 min per epoch. The comparison of parameters of each model is shown in Fig. 5:

Fig. 5.

Comparison of parameters of each model.

As Fig. 5 shows, there are significant differences in the number of parameters among different models. The traditional LSTM and GRU have relatively small parameter count, at 3.2 M and 3.0 M respectively, but their performance is limited when processing complex cultural texts and multi-modal information. The CNN-LSTM has 4.0 M parameters, with a slight improvement in local feature extraction capability, yet its overall performance is still inferior to the proposed model. Transformer and BERT-BiLSTM have larger parameter counts, at 15.0 M and 40.0 M respectively. Although they exhibit strong performance, their computational overhead is high. In contrast, the proposed Bi-LSTM + multi-modal + node weight model has only 4.2 M parameters. Through the multi-modal fusion and node weight mechanism, it achieves the highest accuracy, recall, and F1-score. While maintaining lightweight properties and interpretability, it significantly outperforms other models, demonstrating obvious advantages in the task of complex cultural information extraction. After successfully extracting the cultural genes of Jinxiu Yao headwear, different cultural gene maps are constructed based on this information, as shown in Figs. 6 and 7.

Fig. 6.

Structural cultural gene map of Jinxiu Yao headwear.

Fig. 7.

Pattern cultural gene map of Jinxiu Yao headwear.

Figure 6 shows the components and functions of headwear effectively, which makes the relationship and structure of different parts clearer. The species include Panyao, Aoyao, Chashan Yao, Shanzi Yao and Hualan Yao. According to the way of wearing, it is classified, such as crown hat type, wadding hat type, wrapping type, silver hairpin type and so on. Bi-LSTM model can accurately extract and construct this structural information, which provides a deep structural understanding of cultural works of art. Through the deep learning ability of Bi-LSTM model, the extraction of structural information is more accurate, which makes the digital preservation and analysis of cultural heritage more efficient. This method is of great significance to the headwear in Yao culture, and has a wide range of applicability. It can also extract and display its unique structural information in the study of other ethnic or cultural and artistic works.

In Fig. 7, pattern cultural gene map shows the distribution and characteristics of various patterns on Jinxiu Yao headwear. These patterns are successfully classified and marked by the feature extraction function of Bi-LSTM model, and the pattern cultural gene map shows the different patterns on the headwear and their cultural significance. Through the deep learning ability of Bi-LSTM model, complex patterns can be accurately identified and classified, which provides valuable data support for studying the decorative style and tradition of cultural works of art. This cultural gene map provides important support for the decorative style and tradition of Yao headwear, and provides data support for further study of artistic expressions between different cultures. Bi-LSTM can identify subtle differences in complex patterns and provide in-depth analysis results, which makes it have a wide application prospect in digital research of works of art, cultural inheritance and cross-cultural comparison.

To further verify the quality and credibility of the constructed cultural gene map, this study introduces an expert manual review process based on the quantitative evaluation of the model. The evaluation invited 5 experts in the fields of cultural heritage and ethnic crafts, all with over 5 years of research experience. Experts evaluate the generated cultural gene map from three aspects. Node representativeness: Whether nodes can accurately and comprehensively represent specific cultural elements, such as decoration, structure, process types, etc. Semantic accuracy: Whether the cultural information described by the node is highly consistent with the content of the text and image. Rationality of edge weight: Whether the connections and weights between nodes reflect the actual strength of cultural element associations. The scoring standard adopts a five-point scale (1–5 points), where 1 point represents extremely poor and 5 points represents very realistic and expert cognition. Evaluation method: Each expert independently scores, and after the scoring is completed, the average score and evaluation consistency (Cohen’s Kappa coefficient) are calculated to measure the stability and consistency of the scoring. The evaluation results are shown in Table 3:

Table 3.

Expert manual review results.

Evaluation dimension	Average score	Cohen’s Kappa
Node representativeness	4.4	0.76
Semantic accuracy	4.3	0.79
Rationality of edge weights	4.5	0.78
Comprehensive evaluation	4.4	0.78

As Table 3 shows, experts rated the overall quality of the graph as higher than 4.3 points. This indicates that the constructed cultural gene graph basically conforms to cultural semantic rules in node construction and edge weight setting. The Cohen’s Kappa coefficient is 0.78, indicating a high degree of consistency in the evaluation of the graph among experts and verifying the reliability of manual review. The manual verification results of node and edge weights are highly consistent with the quantitative indicators of the model. This further demonstrates the effectiveness of multimodal feature fusion, node weight mechanism, and Bi-LSTM model in extracting key cultural information and generating high-quality graphs. Through manual verification, the advantages of the model in processing long sequence text, complex semantics, and image information can be clearly identified. Meanwhile, a small number of node omissions and subtle edge weight deviations are also discovered, providing reference for subsequent model optimization. In summary, this study comprehensively evaluates the accuracy, interpretability, and cultural semantic rationality of the constructed cultural gene map through a combination of quantitative indicators and expert manual verification, providing a reliable basis for digital protection and cross-cultural analysis.

Discussion

From the above research, Bi-LSTM model can understand the sequence data more comprehensively through the ability of two-way information processing, which is particularly prominent in the task of extracting cultural gene information. When compared to the conventional LSTM model, the Bi-LSTM model is better at capturing significant information in the sequence since it can employ both forward and backward context information. The Bi-LSTM model can identify and classify the intricate patterns and colors on Jinxiu Yao headwear with greater accuracy, which contributes to a more thorough and detailed creation of the cultural gene map. This is in line with the research findings of Zhang et al. (2020)⁴⁴. Liu discovered that while processing sequence data, bidirectional LSTM could greatly increase the robustness and accuracy of the model. The research of Arbane et al. (2023)⁴⁵ and Alkhwiter & Al-Twairish (2021)⁴⁶ also confirmed that neural network has great advantages in the study of traditional culture.

It is worth noting that this research method is suitable for extracting the cultural gene information of Jinxiu Yao headwear, and has wider applicability. The Bi-LSTM model used, combined with multimodal feature fusion and node weighting mechanism, enables cultural semantic representation to capture complex contextual information and has a certain potential for cross-cultural generalization. It can provide reference and support for the digitization of cultural heritage from different cultural backgrounds. Firstly, this method can be extended to the digital research of other ethnic cultural heritages, such as the pattern and color feature recognition of traditional costumes, ceramics, architectural decoration and other cultural carriers. In these fields, the extraction of cultural genes also depends on the deep understanding of sequence data. Bi-LSTM’s two-way information capture ability can effectively improve the recognition accuracy and adaptability of the model. Secondly, in the study of intangible cultural heritage, such as traditional music, dance and handicraft skills, Bi-LSTM can also be used to analyze time series data, such as note sequence, action trajectory and technological process, to help build a more complete cultural gene map. In addition, this method can be combined with other artificial intelligence technologies, such as Transformer or graph neural network, to further improve the performance of the model in the analysis of complex cultural data. Future research can explore the application of Bi-LSTM in cross-cultural data analysis, and study its feasibility in tasks such as automatic classification of multilingual cultural documents and historical texts. At the same time, multi-modal learning method can be combined to fuse and analyze various cultural information such as text, image and voice to construct a more complete cultural gene map and provide strong data support for the protection, dissemination and innovative design of cultural heritage⁴⁷.

The constructed cultural gene map is used for information extraction and visualization display, and has evolutionability, interpretability, and interactivity. The graph can continuously update nodes and edge weights with new data, achieving dynamic evolution. The setting of nodes and edge weights is based on quantifiable cultural semantic relationships, making the graph highly interpretable. Meanwhile, through multimodal interface design, users can interactively query and analyze nodes, edges, and attributes, enhancing the practicality and interactivity of cultural research. However, this study still has certain limitations. The dataset contains classical Chinese and dialect texts, which pose additional challenges due to their complex sentence structures and limited vocabulary coverage. Although Bi-LSTM demonstrates strong bidirectional contextual understanding capabilities, there is still a slight decline in its performance on these texts. This suggests that the model’s adaptability to low-resource languages can be further improved by introducing domain-specific word embeddings, data augmentation, or hybrid models combining rules and neural networks. Future work can focus on exploring ways to enhance the model’s ability to process classical Chinese and dialect texts, thereby improving the accuracy of cultural gene extraction. In conclusion, Bi-LSTM outperforms traditional models in the task of cultural gene extraction and shows great potential in applications within the fields of tangible and intangible culture. At the same time, acknowledging its shortcomings in low-resource scenarios highlights the need for future development in the directions of hybrid models and multi-modal expansion. This is of great significance for advancing cultural heritage research in the digital age.

Conclusion

Research contribution

The main contribution of this paper is that the Bi-LSTM model has been successfully used to construct the cultural gene map, and the accuracy of cultural information extraction has been significantly improved. Compared with the traditional LSTM and GRU models, Bi-LSTM model shows higher accuracy, recall and F1 value when dealing with complex sequence data through two-way information processing, which shows its unique advantages in extracting cultural gene information. Under the NVIDIA RTX 3090 environment, the peak GPU memory usage of the Bi-LSTM model during training is approximately 7.8 GB, and the average running time per epoch is about 46 s. Compared with the traditional LSTM (6.1 GB, 39 s/epoch) and GRU (6.4 GB, 42 s/epoch), these values are slightly higher. However, while achieving improved performance, the Bi-LSTM model still maintains high computational controllability. In addition, this paper puts forward a multi-dimensional cultural gene map construction method, which covers structure map, pattern map and color map. The structure diagram shows in detail the components of the headwear of Jinxiu Yao nationality and their relationships, the pattern diagram shows the decorative patterns on the headwear and its cultural symbolic significance. The color diagram shows the color collocation and application methods used in the headwear. These maps provide in-depth data support for the digital research of traditional cultural heritage, an operable data framework for the protection of cultural heritage, and new technical ideas for the cross-disciplinary application of cultural disciplines.

Future works and research limitations

Although this paper has achieved good results in cultural gene information extraction, it still faces some limitations, and there is still room for further expansion in the future research direction. Firstly, the training time and computing resources of Bi-LSTM model are high, especially when facing large-scale data. It may limit the practical application scope of the model. In the future research, adopting more efficient model optimization methods (such as pruning, quantization, etc.) and accelerated training algorithms may solve this bottleneck, reduce the calculation cost and improve the calculation efficiency. Secondly, the cultural gene map used in this study is mainly aimed at the analysis of national headwear culture. In the future, the model can be further expanded and applied to more cultural and artistic works to explore the digital protection and inheritance of cultural heritage under different cultural and regional backgrounds. In addition, in order to improve the diversity and accuracy of cultural gene maps, it will also be a direction worth exploring to combine more interdisciplinary data sources (such as images and audio).

Author contributions

Xing Ding: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparationJing Wang: writing—review and editing, visualization, supervision, project administration, funding acquisition.

Funding

This work was supported by Research topic of Social Science Innovation and Development in Anhui Province (Grant No.2023CX141); Anhui Province Social Science Planning Youth Project (Grant No.AHSKQ2023D122);Anhui Province Research Plan Preparation Project (Grant No.2024AH052122); This work was also supported in part by the Social Science Foundation of Shaanxi Province (Grant No. 2023J036), and in part by the Scientific Research Plan Projects of the Shaanxi Education Department (Grant No.24JK0120).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author Jing Wang on reasonable request via e-mail wj352379@163.com.

Declarations

Competing interests

The authors declare no competing interests.