Landscape culture label association mining method and system based on space-time atlas

Abstract

The invention discloses a landscape culture label association mining method and system based on a space-time atlas, and relates to the technical field of cultural heritage digital protection and informatization development, and the method comprises the steps: receiving unstructured classical book text and historical annotation text data, carrying out the data cleaning processing of the unstructured classical book text and historical annotation text data, and storing the unstructured classical book text and historical annotation text data; generating a multi-modal data corpus; inputting the structured classical text in the multi-modal data corpus into a pre-established LDA topic model, and performing semantic topic extraction and dynamic association extraction of landscape culture labels on the unstructured classical text to obtain landscape culture label topic distribution; inputting the landscape culture label theme distribution into a pre-established GRU model to capture a time sequence relation of the classical text, and obtaining a theme-label incidence matrix containing space-time constraints; and performing mapping processing on the theme-label incidence matrix to obtain a four-dimensional space-time map fusing the longitude, the latitude, the normalized time and the cultural circle layer.

Description

Technical Field

[0001] The present invention relates to the technical field of digital protection and informatization development of cultural heritage, and specifically to a method and system for mining associations of landscape cultural labels based on spatiotemporal graphs. Background Art

[0002] The association mining and quantitative analysis of landscape cultural labels in ancient Chinese texts have always been a research focus in the field of digital preservation of historical and cultural heritage, and have important guiding value for the high-quality development of cultural digitization. Cultural labels are semantic feature identifiers extracted from multi-source heterogeneous data. Natural language processing techniques are used to identify cultural entities (such as architectural forms and landscape motifs) and their associated attributes (temporal distribution, spatial order, and metaphorical connotations), forming quantifiable and analyzable cultural cognitive units and providing fine-grained semantic support for the digitization of cultural heritage. As a knowledge representation model, the spatiotemporal graph constructs a multidimensional knowledge network by integrating geographic coordinates, temporal evolution, and semantic networks. Its graph structure nodes represent cultural entities, and its edges define spatiotemporal topology and semantic dependencies, providing dynamic evolution tracking capabilities for cross-dimensional association modeling of cultural labels. Through the dual constraints of spatiotemporal topology and semantic dependencies, it can effectively decouple the coupled relationships between geographic spatial displacement, temporal semantic evolution, and the proliferation of metaphorical connotations in landscape cultural labels.

[0003] Existing technologies have implemented the basic framework for cultural heritage digitization through knowledge graph construction and spatiotemporal clustering algorithms, and have made significant progress in this field. However, three key bottlenecks remain: First, dynamic semantic modeling is lacking. Existing label extraction systems mostly rely on static topic models (such as LDA), making it difficult to effectively capture the semantic evolution of landscape imagery such as "Canglang Pavilion" from the Song Dynasty's "Canglang Pavilion Record" to the Qing Dynasty's "Six Records of a Floating Life"; Second, the discretization of spatiotemporal coding is flawed. Coordinate system conversion errors between modern GIS data and historical geographic information lead to distorted reconstruction of the paths of linear cultural heritage such as the "Silk Road"; Third, multimodal collaboration mechanisms are weak. Special text structures (such as parallel prose and rhyme) cause standard NLP tools to miss entities when parsing unstructured classics. Existing preprocessing frameworks are insufficient in jointly analyzing poetic rhythm (such as the rules of the lyric tone of the "Qinyuanchun" ci tune) and metaphorical structure (such as the farewell meaning of the "Changting" image), which restricts the in-depth semantic mining of cultural heritage. These defects jointly restrict the scientific nature and depth of interpretation of cultural heritage digitization, and urgently need to achieve systematic improvement through technological breakthroughs such as multimodal fusion and dynamic spatiotemporal modeling.

[0004] Building on the development of digital humanities and machine learning-related technologies, a basic analytical framework has been established for topic mining based on the LDA probabilistic model. However, traditional methods struggle to capture the implicit temporal connections between verses. For example, the image of the moon, as seen in its semantic drift from the "cold moon at the border" of the early Tang Dynasty to the "waning moon in the hometown" of the late Tang Dynasty, requires the construction of a hybrid model to track dynamic label evolution. Existing geocoding technology enables spatial annotation of cultural elements, but modeling of the temporal dimension is still limited to dynasty divisions and fails to analyze solar term cycles and the evolution of place names. Furthermore, the current labeling system already encompasses dimensions such as historical allusions and relationships between literati, but technical bottlenecks remain in analyzing the rhythmic features and metaphorical structures of poetry. Summary of the Invention

[0005] In order to solve the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide a landscape cultural label association mining method and system based on spatiotemporal graph.

[0006] In a first aspect, the purpose of the present invention can be achieved by the following technical solution: a landscape cultural label association mining method based on spatiotemporal graph, the method comprising the following steps:

[0007] Receive unstructured classic texts and historical annotation text data, perform data cleaning on the unstructured classic texts and historical annotation text data, and generate a multimodal data corpus;

[0008] The structured classic texts in the multimodal data corpus are input into the pre-established LDA topic model, and the semantic topic extraction and dynamic association extraction of landscape culture tags are performed on the unstructured classic texts to obtain the landscape culture tag topic distribution;

[0009] The thematic distribution of landscape cultural labels is input into the pre-established GRU model to capture the temporal relationship of the classic texts, and a topic-label association matrix containing spatiotemporal constraints is obtained; the topic-label association matrix is mapped to obtain a four-dimensional spatiotemporal map that integrates longitude, latitude, normalized time and cultural circles.

[0010] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: the process of performing data cleaning on the unstructured classic texts and historical annotation text data includes:

[0011] We used a Bi-LSTM+CRF joint training approach, embedding BERT pre-trained word vectors to identify ancient text entities from unstructured classic texts and extract landscape cultural labels. This included segmenting the poetry texts and using domain dictionaries to enhance segmentation accuracy. We also removed stop words and retained nouns, verbs, and adjectives with landscape semantics. We also performed stemming and lemmatization on the texts to unify the expression format.

[0012] By sorting out the research results of landscape cultural labels, a multi-level classification system of landscape cultural labels was defined, which is divided into natural attributes: mountain and rock formations, water bodies and waterscapes, agriculture, forestry, animals and plants; human attributes: architecture, construction, activities; time attributes: day and night, dawn and dawn, climate and meteorology, spring, summer, autumn and winter, multi-dimensional and multi-subcategory landscape cultural labels.

[0013] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: inputting the structured classic text in the multimodal data corpus into a pre-established LDA topic model, and extracting semantic topics and dynamic associations of landscape and cultural tags from the unstructured classic text, including:

[0014] Convert the preprocessed text into a bag-of-words representation as the input of the LDA model to construct a document-label matrix;

[0015] Set hyperparameters:

[0016] The number of topics k, using perplexity and consistency index to verify the quality of the model, and selecting the inflection point where the perplexity decreases slowly according to the elbow rule and the number of topics with the highest consistency index;

[0017] α=0.1: controls the sparsity of the document-topic distribution; β=0.01: controls the sparsity of the topic-word distribution;

[0018] Use Gibbs sampling or variational inference algorithms to train the LDA model, extract the potential topic distribution of landscape and cultural labels in the document, and output the word distribution of each topic and the topic distribution of each poem;

[0019] Based on the Top-N tags of each topic, we assign semantic names to the topics and generate a preliminary association matrix between topics and landscape cultural tags. We then segment each poem into sentences or words, construct a time series, map each word into a word vector, and use a pre-trained word vector model or custom-trained word vectors. We then use a heat map to visualize the topic-tag association matrix, thereby analyzing the distribution of landscape cultural tags under different topics.

[0020] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: the process of inputting the landscape culture tag topic distribution into the pre-established GRU model to capture the temporal relationship of the classic text includes:

[0021] Combined with the GRU temporal network, we capture the association between contextual landscape and cultural labels based on topic clustering, and then optimize the temporal relationship between labels. This allows us to extract semantic topics from unstructured classic texts and optimize the association between landscape and cultural labels. The following is an explanation of the steps:

[0022] Split each poem into sentences or words, construct a time series, and map each word into a word vector using a pre-trained word vector model or custom trained word vectors.

[0023] A bidirectional GRU is used to capture contextual information, and the topic probability distribution of each time step is output through a fully connected layer. The cross-entropy loss is used to measure the difference between the topic distribution predicted by the model and the topic distribution generated by LDA.

[0024] The GRU model is trained using the Adam optimizer. The trained GRU model is used to extract the temporal features of each text and generate a dynamic topic-label association matrix. The topic distribution output by the GRU is weightedly fused with the topic distribution generated by the LDA to obtain the optimized topic-label association matrix.

[0025] In combination with the first aspect, in some implementations of the first aspect, the method further includes: a weighted average formula for weighted fusion of the topic distribution output by the GRU and the topic distribution generated by the LDA is as follows:

[0026] P final =λ·P LDA +(1-λ)·P GRU

[0027] Among them, λ is the weight coefficient;

[0028] Normalize the rows of the fused matrix to ensure that the sum of the probability distribution of each topic is 1:

[0029]

[0030] KL divergence is used to measure the probability distribution difference between LDA and LDA-GRU hybrid clustering optimization of different texts. The calculation formula is as follows:

[0031] Among them, P and Q are the two probability distributions before and after optimization.

[0032] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: the mapping processing of the topic-tag association matrix is performed based on a pre-established four-dimensional coordinate system, and the pre-established four-dimensional coordinate system includes four-dimensional parameters of longitude X, latitude Y, dynasty T, and cultural circle Z.

[0033] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: mapping the topic-tag association matrix to obtain a four-dimensional space-time map integrating longitude, latitude, normalized time, and cultural circles:

[0034] First, the place names of the creation locations of unstructured classic texts are obtained. The geocoding engine calls the GeoNames database and uses the improved Levenshtein distance algorithm to perform place name-coordinate mapping, outputting geographic coordinates (X, Y) in the WGS84 coordinate system. The time standardization module then maps the dynasty chronology to a normalized time axis. Meanwhile, the cultural circle classifier generates a one-hot encoding vector Z based on historical political district data. The spatial mapping engine then performs coordinate transformation: the geographic coordinates are projected to obtain plane coordinates (x, y). The cultural vector Z is embedded in the layer for dimensionality reduction to form cultural dimension features, constructing a four-dimensional spatial point P(x, y, t, z). A separable Gaussian kernel function is used to calculate the four-dimensional probability density field f(x, y, t, z), and the bandwidth parameter h is adaptively adjusted using the Silverman criterion.

[0035] Density estimation is used to calculate the spatiotemporal distribution density of landscape cultural labels in four-dimensional space. The discrete landscape cultural label point data is converted into a continuous probability density distribution through the kernel density estimation method. The density estimation formula is as follows:

[0036]

[0037] Where K is the Gaussian kernel function, h is the bandwidth parameter, and δ(z, Zi) is the indicator function of the cultural circle, which is 1 when z = Zi and 0 otherwise.

[0038] The kernel function is used to measure the contribution of the sample point to the target point (x, y, t, z). Gaussian kernel function:

[0039]

[0040] in, or or

[0041] Therefore, the indicator function is introduced:

[0042]

[0043] Bandwidth controls the smoothing range of the kernel function. Silverman's Rule of Thumb is used to select the adaptive bandwidth. The formula is as follows:

[0044]

[0045] Where σ is the sample standard deviation, IQR is the sample interquartile range, is the sample size;

[0046] Heatmap introduces dual-dimensional filtering of time and culture:

[0047]

[0048] Where f(geo) is the geographic density calculated by KDE, is the time dimension weight, It is the indicator function of cultural circle;

[0049] Four-dimensional scalar field visualization is achieved through the Python platform, including: multi-view projection, supporting XY plane density slicing; dynamic time axis, sliding along the T axis can track the spatial migration trajectory of landscape cultural labels in different time periods; cultural circle overlay, Z-axis layered display of the topological association of cultural circle label chains, and the final output is a four-dimensional spatiotemporal map of landscape cultural labels based on kernel density estimation.

[0050] In a second aspect, in order to achieve the above-mentioned purpose, the present invention discloses a landscape cultural label association mining system based on a spatiotemporal graph, comprising:

[0051] A data processing module is used to receive unstructured classic texts and historical annotation text data, perform data cleaning on the unstructured classic texts and historical annotation text data, and generate a multimodal data corpus;

[0052] The association extraction module is used to input the structured classic texts in the multimodal data corpus into the pre-established LDA topic model, extract the semantic topics of the unstructured classic texts and dynamically extract the landscape culture tags, and obtain the landscape culture tag topic distribution;

[0053] The atlas generation module is used to input the thematic distribution of landscape cultural labels into the pre-established GRU model to capture the temporal relationship of the classic text and obtain a topic-label association matrix containing spatiotemporal constraints; the topic-label association matrix is mapped to obtain a four-dimensional spatiotemporal map that integrates longitude, latitude, normalized time and cultural circles.

[0054] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, it adopts the above-mentioned landscape cultural label association mining method based on spatiotemporal graph.

[0055] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a computer-readable storage medium is disclosed, in which a computer program is stored. When the computer program is loaded and executed by a processor, the landscape cultural label association mining method based on a spatiotemporal graph as described above is adopted.

[0056] Beneficial effects of the present invention:

[0057] This method overcomes the limitations of traditional LDA models in modeling long-range dependencies by constructing a temporal feature enhancement module based on a gated recurrent unit (GRU) and capturing the association patterns of landscape and cultural labels through a bidirectional attention mechanism. It also designs a four-dimensional space-time cube model that integrates cultural sphere dimensional variables to support three-dimensional dynamic rendering of historical period slices. Furthermore, it establishes a geographic-cultural space-time coordinate conversion layer to achieve dimensional expansion from two-dimensional geographic tags to a four-dimensional cultural space-time coordinate system. Through the technical chain of "text semantic network construction - label dynamic association modeling - space-time dimension mapping", this method effectively supports the restoration of the spatiotemporal context of landscape elements in the analysis of unstructured texts. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0059] Figure 1 It is a schematic flow chart of the method of the present invention;

[0060] Figure 2 Schematic diagram of the perplexity test of the LDA model of the present invention;

[0061] Figure 3 Schematic diagram of consistency test of LDA model of the present invention;

[0062] Figure 4 1 is a heat diagram of the theme distribution of landscape culture tags extracted by theme clustering in an embodiment of the present invention;

[0063] Figure 5 Schematic diagram of KL divergence test of topic distribution of LDA and LDA-GRU in an embodiment of the present invention;

[0064] Figure 6 2. It is a schematic diagram comparing the label association matrix before and after the optimization of the timing features in an embodiment of the present invention;

[0065] Figure 7 1 is a schematic diagram of the distribution of landscape culture label themes in Xin Ci poems of different periods in an embodiment of the present invention;

[0066] Figure 8 Schematic diagram of the distribution of landscape cultural label themes at different locations in an embodiment of the present invention;

[0067] Figure 9 is a schematic diagram of the static geographical distribution of landscape cultural labels in an embodiment of the present invention;

[0068] Figure 10is a thermal diagram of the landscape culture label space in an embodiment of the present invention;

[0069] Figure 11 This is a schematic diagram of the geographical distribution of landscape culture labels in time slices according to an embodiment of the present invention;

[0070] Figure 12 It is a four-dimensional spatiotemporal map of landscape cultural labels in an embodiment of the present invention;

[0071] Figure 13 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0073] Example 1:

[0074] like Figure 1 As shown, a landscape cultural label association mining method based on spatiotemporal graph includes the following steps:

[0075] S101: receiving unstructured classic texts and historical annotation text data, performing data cleaning on the unstructured classic texts and historical annotation text data to generate a multimodal data corpus;

[0076] Specifically, a Chinese corpus covering unstructured classics was constructed, with historical annotations added as an auxiliary semantic layer. Based on a CHGIS regional map and historical geographic database, an information dataset with spatiotemporal topology reconstruction of text labels was formed. In this example, the various historical versions and annotations of "The Complete Poems of Xin Qiji" were used as the text data source, and 338 poems with relatively complete time and place information were manually screened, totaling 52,558 characters.

[0077] For Chinese text data cleaning and word segmentation model design, we chose to jointly train a Bi-LSTM+CRF model (Table 1), embedding it with BERT pre-trained word vectors. This method identifies ancient text entities and extracts landscape cultural labels from unstructured classical texts. Specifically, this involves word segmentation of poetry and lyrics, using domain dictionaries to enhance segmentation accuracy; removing stop words while retaining nouns, verbs, and adjectives with landscape semantics; and performing stemming and lemmatization on the text to unify the representation.

[0078] Table 1 Comparison of text label mining models

[0079]

[0080]

[0081] By sorting out the research results related to landscape cultural labels, a multi-level classification system of landscape cultural labels was defined (Table 2), which is divided into 9 subcategories of landscape cultural labels in three dimensions: natural attributes (mountain and rock formations, water bodies and waterscapes, agriculture, forestry, animals and plants), human attributes (architecture, construction, activities), and time attributes (day and night, dawn and dawn, climate and weather, spring, summer, autumn and winter).

[0082] Table 2 Landscape culture labeling system

[0083]

[0084] S102: Inputting the structured classic texts in the multimodal data corpus into a pre-established LDA topic model, performing semantic topic extraction and dynamic association extraction of landscape culture tags on the unstructured classic texts, and obtaining a landscape culture tag topic distribution;

[0085] Using the LDA topic model, we can extract semantic topics from unstructured classic texts and dynamically associate them with landscape and cultural labels. The following is a detailed description of the specific steps:

[0086] The preprocessed text is converted into a Bag of Words representation as the input of the LDA model to construct the Document-Term Matrix.

[0087] Set hyperparameters:

[0088] The number of topics k is selected. The perplexity and consistency index are used to verify the quality of the model. According to the Elbow Method, the inflection point where the perplexity decreases slowly is selected and the number of topics with the highest consistency index is selected.

[0089] α = 0.1: controls the sparsity of the document-topic distribution. A smaller value indicates that documents tend to be concentrated in a small number of topics.

[0090] β = 0.01: controls the sparsity of topic-word distribution. A smaller value indicates that topics tend to be concentrated in a small number of word labels.

[0091] Use Gibbs sampling or variational inference algorithms to train an LDA model to extract the latent topic distribution of landscape and cultural labels in documents. Output the word distribution of each topic (Topic-Word Distribution) and the topic distribution of each poem (Document-Topic Distribution).

[0092] Furthermore, based on the top-N tags for each topic, we manually assigned semantic names to the topics and generated a preliminary correlation matrix between the topics and landscape cultural tags. Furthermore, we segmented each poem into sentences or words to construct a time series. Each word was mapped into a word embedding (Word Embedding) using pre-trained word embedding models (such as Word2Vec and BERT) or custom-trained word embeddings. Heatmaps were used to visualize the topic-tag correlation matrix, facilitating analysis of the distribution of landscape cultural tags under different topics.

[0093] In the embodiment, the unstructured classics file after cleaning contains 338 poems and 161 keywords (Table 3). The pre-processed text is converted into a bag-of-words representation and the LDA topic model test is performed ( Figure 2 、 3 The perplexity index achieved better results when the number of topics was less than 10; the inflection point of the consistency index was obtained when the number of topics was 7. According to the elbow rule, the LDA model with 7 topics achieved the best model quality for this text data. The output was the distribution of landscape and cultural labels extracted by LDA (Table 4) and the thermal distribution of high-frequency landscape and cultural labels corresponding to different topics ( Figure 4 ).

[0094] Table 3 Structured text keywords after cleaning

[0095]

[0096]

[0097] Table 4 Thematic distribution of landscape culture labels extracted by LDA

[0098]

[0099] S103: Input the thematic distribution of landscape cultural labels into the pre-established GRU model to capture the temporal relationship of the classic texts, and obtain a topic-label association matrix containing spatiotemporal constraints; map the topic-label association matrix to obtain a four-dimensional spatiotemporal map that integrates longitude, latitude, normalized time, and cultural circles.

[0100] Combined with the GRU temporal network, we capture the association between contextual landscape and cultural labels based on topic clustering, and then optimize the temporal relationship between labels. This allows us to extract semantic topics from unstructured classic texts and optimize the association between landscape and cultural labels. The following is a detailed description of the specific steps:

[0101] Split each poem into sentences or words to construct a time series. Map each word to a word embedding (WordEmbedding), using a pre-trained word embedding model (such as Word2Vec, BERT) or a custom-trained word embedding.

[0102] A bidirectional GRU (Bi-GRU) is used to capture contextual information and output the topic probability distribution at each time step through a fully connected layer. Cross-entropy loss is used to measure the difference between the topic distribution predicted by the model and the topic distribution generated by LDA.

[0103] The Adam optimizer was used with a learning rate of 0.0008. The number of GRU hidden layers was set to 256. 80% of the text data was used as the training set, and 20% as the validation set. Training was performed for 300 epochs, with an early stopping strategy to prevent overfitting. The trained GRU model was used to extract temporal features from each text and generate a dynamic topic-tag association matrix. The topic distribution output by the GRU was weighted and fused with the topic distribution generated by LDA to obtain the optimized topic-tag association matrix. The weighted average formula is as follows:

[0104] P final =λ·P LDA +(1-λ)·P GRU

[0105] Among them, λ is the weight coefficient.

[0106] Normalize the rows of the fused matrix to ensure that the sum of the probability distribution of each topic is 1:

[0107]

[0108] KL divergence is used to measure the probability distribution difference between LDA and LDA-GRU hybrid clustering optimization of different texts. The calculation formula is as follows:

[0109] Among them, P and Q are the two probability distributions before and after optimization.

[0110] If the KL divergence is large (mean > 0.1), it means that the topic distribution optimized by GRU has changed significantly compared to LDA. If the KL divergence is small (mean < 0.05), it means that GRU has only slightly adjusted the LDA result and has not significantly optimized it. The average KL divergence after optimization is 0.468, and the optimization result is significant ( Figure 3 ).

[0111] MAE, MSE, and RMSE are very small (Table 4), indicating that the GRU prediction results are close to the LDA results, indicating that GRU is mainly smoothing the topic distribution of LDA rather than completely changing the distribution structure. 2 It is close to 1, which means that the results predicted by GRU are highly correlated with LDA and the optimization is reasonable.

[0112] Table 4 Prediction results of LDA-GRU hybrid clustering model

[0113]

[0114] In order to visualize the distribution results of landscape culture labels under different themes, the topic-label association matrix before and after capturing the contextual temporal information is plotted in the form of a heat map. If the heat map color is more uniform, it means that the GRU has adjusted the balance between themes and optimized the robustness of topic extraction; if the heat map has obvious dark and light nodes, it means that the GRU has captured the temporal sequence of certain topics and landscape culture labels and enhanced the display in the optimized matrix.

[0115] In the embodiment, based on the LDA extraction of cultural label topic associations, the temporal features are enhanced to obtain the hybrid clustering matrix after GRU optimization (Table 6). Comparison of the topic-label association distribution before and after GRU optimization ( Figure 4 ), the association matrix is more evenly distributed after adding temporal features, indicating that GRU adjusts the balance between topics and optimizes the robustness of the label association matrix.

[0116] Based on the timestamp as the screening basis, the non-structured classics background data selected in the embodiment is divided into three time periods according to the author's life stage, and the landscape culture label theme distribution of different periods is output and visualized through heat map ( Figure 7 ).

[0117] Based on the location as the screening basis, the non-structured classics background data selected in the embodiment are used to extract the theme distribution of different locations according to the writing location of the text content, and output the theme distribution of landscape culture labels in different periods, which is visualized through heat map ( Figure 8 ), which allows for a clear comparison of the authors’ landscape preferences across different regions.

[0118] Table 6 Partial coding of landscape cultural label association matrix

[0119]

[0120] Based on the output topic-tag association data, this method constructs a spatiotemporal enhanced matrix that can be applied to the spatiotemporal attribute annotation of landscape cultural labels, creating conditions for the digital analysis of cultural heritage. The geographic information dataset integration step includes four-dimensional parameters of longitude X, latitude Y, dynasty T, and cultural circle Z (Table 5), which is suitable for revealing the spatiotemporal evolution analysis of landscape cultural heritage themes. The discretized spatiotemporal cube model (step) can support the digital reconstruction of the spatial pattern of historical garden heritage. Its four-dimensional coordinate mapping algorithm provides a technical framework for extracting the spatiotemporal characteristics of traditional gardening techniques such as "one pond and three mountains". The coupling design of kernel density estimation and time slice analysis modules can reveal the characteristics of the transmission path of cultural elements and is suitable for studying the evolution trajectory of geographical attention in cultural labels.

[0121] Table 5 Four dimensions defined in data encoding

[0122]

[0123]

[0124] The final output of this method is a four-dimensional space-time map, which integrates multi-dimensional features such as geographic coordinates, temporal evolution, and cultural circles to achieve spatiotemporal continuous field modeling of landscape cultural labels. The following technical features all serve to generate this core result:

[0125] First, the place names of the creation locations of unstructured classic texts are obtained. The geocoding engine calls the GeoNames database and uses the improved Levenshtein distance algorithm (similarity threshold ≥ 0.9) to perform place name-coordinate mapping, outputting geographic coordinates (X, Y) in the WGS84 coordinate system. The time normalization module then maps the dynasty chronology to a normalized time axis. Meanwhile, the cultural circle classifier generates a one-hot encoded vector Z based on historical political district data. The spatial mapping engine then performs coordinate transformation: the geographic coordinates are projected to obtain plane coordinates (x, y), and the cultural vector Z is embedded in the layer to reduce its dimensionality to form cultural dimensional features, ultimately constructing a four-dimensional spatial point P(x, y, t, z). On this basis, the kernel density estimation module uses a separable Gaussian kernel function to calculate the four-dimensional probability density field f(x, y, t, z). The bandwidth parameter h is adaptively adjusted using the Silverman criterion. The resulting cultural diffusion isosurface can be used to reveal the spatiotemporal evolution of landscape cultural labels.

[0126] Density estimation is used to calculate the spatiotemporal distribution density of landscape cultural labels in four-dimensional space (longitude X, latitude Y, time T, cultural circle Z). The discrete landscape cultural label point data is converted into a continuous probability density distribution using the kernel density estimation (KDE) method. The density estimation formula is as follows:

[0127]

[0128] Where K is the Gaussian kernel function, h is the bandwidth parameter, and δ(z, Zi) is the indicator function of the cultural circle, which is 1 when z = Zi and 0 otherwise.

[0129] The kernel function is used to measure the contribution of the sample point to the target point (x, y, t, z). Commonly used Gaussian kernel functions are:

[0130]

[0131] in, or or

[0132] Cultural circle is a categorical variable and cannot be directly smoothed using kernel functions, so an indicator function is introduced:

[0133]

[0134] The bandwidth controls the smoothing range of the kernel function. The larger the bandwidth, the stronger the smoothing effect; the smaller the bandwidth, the closer the density estimate is to the original data. Silverman's Rule of Thumb is used to select the adaptive bandwidth. The specific formula is as follows:

[0135]

[0136] Where σ is the sample standard deviation, IQR is the sample interquartile range, is the sample size.

[0137] Heatmap introduces dual-dimensional filtering of time and culture:

[0138]

[0139] Where f(geo) is the geographic density calculated by KDE, is the time dimension weight, It is an indicator function of cultural circle.

[0140] Four-dimensional scalar field visualization is achieved through the Python platform, including: multi-view projection, supporting XY plane density slicing; dynamic time axis, sliding along the T axis can track the spatial migration trajectory of landscape cultural labels in different time periods; cultural circle overlay, Z-axis layered display of the topological relationship of cultural circle label chains.

[0141] The final output of the embodiment is a four-dimensional spatiotemporal map, which integrates geographic spatial distribution, temporal evolution laws and cultural circle characteristics, and presents the spatiotemporal aggregation effect of landscape cultural labels. Its construction process includes the following supporting links: First, the geocoding engine is used to connect to the GeoNames database for coordinate matching, and the Levenshtein distance threshold is set to ≤3 to ensure the accuracy of place name mapping. The (X, Y) geographic coordinate set is generated through WGS84 coordinate system conversion, and a static spatial distribution matrix Ms×2 (Table 6) is constructed, where the matrix elements Represents the label aggregation weight of the i-th spatial point and generates a kernel density heat map ( Figure 9-10); then the chronology of the classics is normalized to the time axis T∈[0,1], and the five major cultural circles (Central Plains / Wuyue / Jianghuai / War Memory / Frontier Subcircle) are divided according to the cultural regions to generate the Z-dimensional one-hot encoding vector Z; finally, the four-dimensional kernel density estimation (4D-KDE) model is adopted, and the bandwidth h of the Gaussian kernel function is adaptively optimized according to the Silverman criterion. The landscape cultural label extraction and rendering of the four-dimensional data field are realized through the Python platform ( Figure 12 ).

[0142] Table 6 Partial coding of the four-dimensional coordinate system of landscape cultural labels

[0143]

[0144] Second, as Figure 13 As shown, in order to achieve the above purpose, the present invention discloses a landscape cultural label association mining system based on spatiotemporal graph, comprising:

[0145] The data processing module 11 is used to receive unstructured classic texts and historical annotation text data, perform data cleaning on the unstructured classic texts and historical annotation text data, and generate a multimodal data corpus;

[0146] The association extraction module 12 is used to input the structured classic texts in the multimodal data corpus into the pre-established LDA topic model, perform semantic topic extraction and dynamic association extraction of landscape culture tags on the unstructured classic texts, and obtain the landscape culture tag topic distribution;

[0147] The atlas generation module 13 is used to input the thematic distribution of landscape cultural labels into a pre-established GRU model to capture the temporal relationship of the classic texts and obtain a topic-label association matrix containing spatiotemporal constraints; the topic-label association matrix is mapped to obtain a four-dimensional spatiotemporal map that integrates longitude, latitude, normalized time and cultural circles.

[0148] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0149] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium having a computer program stored thereon, which executes the above method when executed by a processor. The storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component.

[0150] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0151] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present disclosure. Various changes and improvements may be made to the present disclosure without departing from the spirit and scope of the present disclosure, and such changes and improvements shall fall within the scope of the present disclosure.

Claims

1. A landscape cultural label association mining method based on spatiotemporal graph, characterized by: The method comprises the following steps: Receive unstructured classic texts and historical annotation text data, perform data cleaning on the unstructured classic texts and historical annotation text data, and generate a multimodal data corpus; The structured classic texts in the multimodal data corpus are input into the pre-established LDA topic model, and the semantic topic extraction and dynamic association extraction of landscape culture tags are performed on the unstructured classic texts to obtain the landscape culture tag topic distribution; The thematic distribution of landscape cultural labels is input into the pre-established GRU model to capture the temporal relationship of the classic texts, and a topic-label association matrix containing spatiotemporal constraints is obtained; the topic-label association matrix is mapped to obtain a four-dimensional spatiotemporal map that integrates longitude, latitude, normalized time and cultural circles.

2. The method for mining landscape cultural label associations based on spatiotemporal graphs according to claim 1, characterized in that: The process of performing data cleaning on the unstructured classic texts and historical annotation text data includes: We used a Bi-LSTM+CRF joint training approach, embedding BERT pre-trained word vectors to identify ancient text entities from unstructured classic texts and extract landscape cultural labels. This included segmenting the poetry texts and using domain dictionaries to enhance segmentation accuracy. We also removed stop words and retained nouns, verbs, and adjectives with landscape semantics. We also performed stemming and lemmatization on the texts to unify the expression format. By sorting out the research results of landscape cultural labels, a multi-level classification system of landscape cultural labels was defined, which is divided into natural attributes: mountain and rock formations, water bodies and waterscapes, agriculture, forestry, animals and plants; human attributes: architecture, construction, activities; time attributes: day and night, dawn and dawn, climate and meteorology, spring, summer, autumn and winter, multi-dimensional and multi-subcategory landscape cultural labels.

3. The method for mining landscape cultural label associations based on spatiotemporal graphs according to claim 1 is characterized in that: The process of inputting the structured classic texts in the multimodal data corpus into the pre-established LDA topic model and extracting semantic topics and dynamic association of landscape cultural tags from the unstructured classic texts includes: Convert the preprocessed text into a bag-of-words representation as the input of the LDA model to construct a document-label matrix; Set hyperparameters: The number of topics k, using perplexity and consistency index to verify the quality of the model, and selecting the inflection point where the perplexity decreases slowly according to the elbow rule and the number of topics with the highest consistency index; α=0.1: controls the sparsity of the document-topic distribution; β=0.01: controls the sparsity of the topic-word distribution; Use Gibbs sampling or variational inference algorithms to train the LDA model, extract the potential topic distribution of landscape and cultural labels in the document, and output the word distribution of each topic and the topic distribution of each poem; Based on the Top-N tags of each topic, we assign semantic names to the topics and generate a preliminary association matrix between topics and landscape cultural tags. We then segment each poem into sentences or words, construct a time series, map each word into a word vector, and use a pre-trained word vector model or custom-trained word vectors. We then use a heat map to visualize the topic-tag association matrix, thereby analyzing the distribution of landscape cultural tags under different topics.

4. The method for mining landscape cultural label associations based on spatiotemporal graphs according to claim 1, characterized in that: The process of inputting the landscape culture label topic distribution into the pre-established GRU model to capture the temporal relationship of the classic text includes: Combined with the GRU temporal network, we capture the association between contextual landscape and cultural labels based on topic clustering, and then optimize the temporal relationship between labels. This helps extract semantic topics from unstructured classic texts and optimize the association between landscape and cultural labels. The following is an explanation of the steps: Split each poem into sentences or words, construct a time series, and map each word into a word vector using a pre-trained word vector model or custom trained word vectors. A bidirectional GRU is used to capture contextual information, and the topic probability distribution of each time step is output through a fully connected layer. The cross-entropy loss is used to measure the difference between the topic distribution predicted by the model and the topic distribution generated by LDA. The GRU model is trained using the Adam optimizer. The trained GRU model is used to extract the temporal features of each text and generate a dynamic topic-label association matrix. The topic distribution output by the GRU is weightedly fused with the topic distribution generated by the LDA to obtain the optimized topic-label association matrix.

5. The method for mining landscape cultural label association based on spatiotemporal graph according to claim 4 is characterized in that: The weighted average formula for weighted fusion of the topic distribution output by GRU and the topic distribution generated by LDA is as follows: P final =λ·P LDA +(1-λ)·P GRU Among them, λ is the weight coefficient; Normalize the rows of the fused matrix to ensure that the sum of the probability distribution of each topic is 1: KL divergence is used to measure the probability distribution difference between LDA and LDA-GRU hybrid clustering optimization of different texts. The calculation formula is as follows: Among them, P and Q are the two probability distributions before and after optimization.

6. The method for mining landscape cultural label associations based on spatiotemporal graphs according to claim 1, characterized in that: The mapping process of the topic-tag association matrix is performed based on a pre-established four-dimensional coordinate system, which includes four-dimensional parameters of longitude X, latitude Y, dynasty T, and cultural circle Z.

7. The method for mining landscape cultural label association based on spatiotemporal graph according to claim 1, characterized in that: The process of mapping the topic-tag association matrix to obtain a four-dimensional space-time map integrating longitude, latitude, normalized time and cultural circles: First, the place names of the creation locations of unstructured classic texts are obtained. The geocoding engine calls the GeoNames database and uses the improved Levenshtein distance algorithm to perform place name-coordinate mapping, outputting geographic coordinates (X, Y) in the WGS84 coordinate system. The time standardization module then maps the dynasty chronology to a normalized time axis. Meanwhile, the cultural circle classifier generates a one-hot encoding vector Z based on historical political district data. The spatial mapping engine then performs coordinate transformation: the geographic coordinates are projected to obtain plane coordinates (x, y). The cultural vector Z is embedded in the layer for dimensionality reduction to form cultural dimension features, constructing a four-dimensional spatial point P(x, y, t, z). A separable Gaussian kernel function is used to calculate the four-dimensional probability density field f(x, y, t, z), and the bandwidth parameter h is adaptively adjusted using the Silverman criterion. Density estimation is used to calculate the spatiotemporal distribution density of landscape cultural labels in four-dimensional space. The discrete landscape cultural label point data is converted into a continuous probability density distribution through the kernel density estimation method. The density estimation formula is as follows: Where K is the Gaussian kernel function, h is the bandwidth parameter, and δ(z, Zi) is the indicator function of the cultural circle, which is 1 when z = Zi and 0 otherwise. The kernel function is used to measure the contribution of the sample point to the target point (x, y, t, z). Gaussian kernel function: in, or or Therefore, the indicator function is introduced: Bandwidth controls the smoothing range of the kernel function. Silverman's Rule of Thumb is used to select the adaptive bandwidth. The formula is as follows: Where σ is the sample standard deviation, IQR is the sample interquartile range, is the sample size; Heatmap introduces dual-dimensional filtering of time and culture: Where f(geo) is the geographic density calculated by KDE, is the time dimension weight, It is the indicator function of cultural circle; Four-dimensional scalar field visualization is achieved through the Python platform, including: multi-view projection, supporting XY plane density slicing; dynamic time axis, sliding along the T axis can track the spatial migration trajectory of landscape cultural labels in different time periods; cultural circle overlay, Z-axis layered display of the topological association of cultural circle label chains, and the final output is a four-dimensional spatiotemporal map of landscape cultural labels based on kernel density estimation.

8. A landscape cultural label association mining system based on spatiotemporal graph, characterized by: include: A data processing module is used to receive unstructured classic texts and historical annotation text data, perform data cleaning on the unstructured classic texts and historical annotation text data, and generate a multimodal data corpus; The association extraction module is used to input the structured classic texts in the multimodal data corpus into the pre-established LDA topic model, extract the semantic topics of the unstructured classic texts and dynamically extract the landscape culture tags, and obtain the landscape culture tag topic distribution; The atlas generation module is used to input the thematic distribution of landscape cultural labels into the pre-established GRU model to capture the temporal relationship of the classic text and obtain a topic-label association matrix containing spatiotemporal constraints; the topic-label association matrix is mapped to obtain a four-dimensional spatiotemporal map that integrates longitude, latitude, normalized time and cultural circles.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on a processor. When the processor loads and executes the computer program, it adopts a landscape cultural label association mining method based on a spatiotemporal graph according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, wherein: When the computer program is loaded and executed by the processor, the landscape cultural label association mining method based on spatiotemporal graph described in any one of claims 1 to 7 is adopted.

Images