A multi-constraint knowledge graph construction method and system for historical vestiges body language

Abstract

This invention relates to a method and system for constructing a multi-constraint knowledge graph of body language from historical remains, belonging to the field of body language cultural heritage preservation technology. The method includes acquiring 3D point cloud scan data, multispectral imaging data, and body language semantic annotation data from archaeological sites of historical remains; generating a sequence of body language feature vectors constrained by era style; calculating cross-modal attention weights between body language nodes and text nodes; selecting high-confidence edges to construct a set of body language-context association edges conforming to ritual constraints; constructing an initial network of the knowledge graph with a hierarchical structure of cultural symbols; using the topological and attribute features of adjacent complete nodes for link prediction and feature generation; supplementing the body language semantic features of incomplete nodes; and finally fusing them to construct a database and knowledge graph of body language from historical remains. The beneficial effects of this invention are that it repairs the structural deficiencies of the knowledge graph caused by the physical incompleteness of the remains, significantly improving the completeness and coherence of the knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of body language cultural heritage protection technology, and more specifically, to a method and system for constructing a multi-constraint knowledge graph of historical body language. Background Technology

[0002] Current digital preservation of historical human figures mainly focuses on high-precision reconstruction and visualization of geometric forms. Existing technologies typically use 3D laser scanning to acquire point cloud data and generate mesh models through surface reconstruction, or use multispectral imaging for surface material analysis. These methods can only achieve digital archiving of physical forms and lack in-depth analysis of the physical semantics, ritual functions, and cultural connotations carried by the human figures.

[0003] For the extraction of body posture semantics, existing methods mostly rely on manual annotation combined with traditional machine learning classifiers to identify postures in isolation. This fails to understand body postures within the aesthetic paradigms and ritual contexts of specific historical periods, resulting in interpretations that are detached from historical and cultural backgrounds and make it difficult to establish connections between body postures and ritual texts and cultural symbols. Furthermore, for incomplete surviving human figures, existing technologies mostly use geometric restoration methods based on geometric symmetry or texture statistical patterns for visual completion, failing to fill in missing semantic nodes at the knowledge association level, resulting in structural breaks and semantic gaps in the constructed knowledge graph. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for constructing a multi-constraint knowledge graph of historical body language, so as to improve the above-mentioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a method for constructing a multi-constraint knowledge graph of historical body language, including: Acquire 3D point cloud scanning data, multispectral imaging data, and body semantic annotation data from archaeological sites of historical human figures; perform rigid body registration and deformation field correction of point clouds, registration and reflectivity standardization of multispectral bands, and synchronization of timestamps and spatial coordinate transformation of annotation data; and fuse them to generate unified format data containing 3D geometric coordinates, material reflection characteristics, and body semantic labels. Based on unified format data, a dynasty style discriminator network is constructed to extract body geometric features. The style discriminator obtains the era style constraint parameters, and the body features are decoupled and reparameterized for style encoding, thereby generating a sequence of body feature vectors with era style constraints. Using the sequence of body posture feature vectors, we parse the ritual text and construct a semantic co-occurrence graph, which serves as the prior topology of the graph attention network. We use the vector sequence as the node feature input to calculate the cross-modal attention weights between body posture nodes and text nodes, and select high-confidence edges to construct a set of body posture-context association edges that conform to the ritual constraints. Based on the body-context association edge set, we define the Peirce semiotic triple structure to reproduce body-object-interpretation, map association edges to triple instances, perform dynamic weight propagation and hierarchical aggregation based on attention mechanism, and construct an initial network of knowledge graph with cultural symbol hierarchical structure. Based on the initial network of the knowledge graph, incomplete posture nodes are identified and a self-supervised reasoning model based on graph neural network is constructed. The topological and attribute features of adjacent complete nodes are used for link prediction and feature generation to complete the body semantic features of incomplete nodes. Finally, the historical body language database and knowledge graph are integrated to construct the knowledge graph.

[0005] Preferably, the acquisition of 3D point cloud scan data, multispectral imaging data, and body semantic annotation data from the archaeological site involves performing rigid body registration and deformation field correction of the point cloud, multispectral band registration and reflectivity standardization, and timestamp synchronization and spatial coordinate transformation of the annotation data. This process fuses these data to generate unified format data containing 3D geometric coordinates, material reflection characteristics, and body semantic tags, including: Rigid body registration based on anatomical landmarks is performed on the 3D point cloud data. Human bony landmarks with dynastic characteristics are selected as the registration reference. Then, thin plate spline functions are used to perform deformation field correction, reconstruct the surface mesh topology of the surviving human figure, and extract the vertex coordinate set and normal vector field. Dark current correction and flat field correction are performed on the multispectral imaging data. Reflectance spectral curves from the visible to near-infrared bands are extracted. Material feature codes are obtained by dimensionality reduction through principal component analysis. The material feature codes are mapped to the corresponding vertex index positions of the surface mesh to establish material-geometry relationships. The micro-action semantic annotation data recorded at the archaeological site is aligned based on timestamps and transformed into a spatial coordinate system based on coordinate transformation matrices. Action semantic tags are then bound to the corresponding vertex indices and fused to generate data in a unified format.

[0006] Preferably, the process involves constructing a dynasty style discriminator network based on unified format data to extract body geometric features, obtaining era style constraint parameters through the style discriminator, and performing style decoupling and reparameterization encoding on the body features to generate a sequence of body feature vectors with era style constraints, including: A body style discriminator network based on dynasty classification is constructed. The geometric coordinate matrix in the uniform format data is input, and the body proportion features and posture angle features are extracted through graph convolutional layers. The probability distribution of dynasty style is output through fully connected layers and softmax classifier. The style discriminator obtains the latent variables of the era style, performs style decoupling operation on the geometric features, separates the posture content features from the dynasty style features, and establishes a conditional latent space mapping relationship. The content features retain the posture ontology information, and the style features encode the era aesthetic features. Based on the reparameterization technique, style latent variables are fused with content features, and a sequence of body feature vectors with era style constraints is generated through a style conditional encoder.

[0007] Preferably, the step involves using a sequence of body posture feature vectors to parse the ritual text and construct a semantic co-occurrence graph, which serves as the prior topology of the graph attention network. The vector sequence is then used as node feature input to calculate cross-modal attention weights between body posture nodes and text nodes. High-confidence edges are selected to construct a set of body posture-context association edges that conform to the ritual constraints, including: The dependency syntax structure of the ritual text is analyzed, ritual action vocabulary and ritual scene description are extracted, and a word co-occurrence graph is constructed as a prior topological structure, where nodes represent ritual concepts and edges represent the co-occurrence relationship of words within a specific window, forming a semantic network of ritual context; The sequence of body shape feature vectors is injected into the graph attention network as node features, and the cross-modal attention weights between body shape nodes and text nodes are calculated. The query vector is derived from the body shape features, and the key vector and value vector are derived from the text embedding, thus forming a cross-modal attention mechanism. Based on ritual logic rules, high-confidence association edges are selected, and a set of body posture-context association edges that conform to ritual constraints is constructed. Each edge is associated with a specific ritual context and action semantics, forming a body posture-context bipartite graph structure.

[0008] Preferably, based on the body-context association edge set, a Peirce semiotic triple structure is defined to reproduce the body-object-interpretation term, mapping association edges to triple instances, performing dynamic weight propagation and hierarchical aggregation based on an attention mechanism, and constructing an initial network of a knowledge graph with a cultural symbol hierarchical structure, including: Mapping the body posture-context association set to Peirce's semiotic triple structure, instantiating it into a representation-object-interpretation triple, where the representation corresponds to the graphic representation of body posture, the object corresponds to the referential object of ritual function, and the interpreter corresponds to the interpretative symbol of cultural connotation, thus forming a semiotic instance; Dynamic weight propagation based on attention mechanism is performed, which aggregates neighborhood symbol features along the triple structure and updates the node representation. The propagation weight is determined by the cultural association strength and ritual importance coefficient between symbols, thereby realizing the hierarchical propagation of symbol features. Based on the hierarchical relationship of cultural symbols, the node topology is reorganized to construct an initial network of knowledge graph with a hierarchical structure of cultural symbols, where high-level nodes represent abstract cultural concepts and low-level nodes represent concrete physical instances.

[0009] Preferably, the process involves identifying incomplete posture nodes based on the initial knowledge graph network and constructing a self-supervised reasoning model based on a graph neural network. This model utilizes the topological and attribute features of adjacent complete nodes for link prediction and feature generation, completing the body language semantic features of incomplete nodes. Finally, it integrates and constructs a historical legacy body language database and a knowledge graph, including: Based on the initial network of the knowledge graph, incomplete nodes with missing body semantic features are identified, and a topological mask matrix is ​​constructed to identify the regions to be completed. A mask value of 1 indicates that the node's features are incomplete, and a mask value of 0 indicates that the node's features are complete and usable. A self-supervised inference model based on graph neural networks is constructed. The topological connection relationship and attribute features of adjacent complete nodes are used to generate the completion features of incomplete nodes through graph convolution operation. The model training adopts the mask language modeling paradigm, randomly masking some node features for self-supervised learning. The completion features are integrated with the existing network, consistency checks are performed, and the final historical legacy body language database and knowledge graph are constructed, in which the completion nodes and the original nodes maintain topological coherence and semantic consistency.

[0010] Secondly, this application also provides a multi-constraint knowledge graph construction system for historical body language, including: The acquisition module is used to acquire 3D point cloud scanning data, multispectral imaging data and body semantic annotation data of historical human figures, and to perform rigid body registration and deformation field correction of point cloud, registration and reflectivity standardization of multispectral bands, and synchronization of timestamps and spatial coordinate transformation of annotation data. It integrates and generates unified format data containing 3D geometric coordinates, material reflection characteristics and body semantic tags. Encoding module: Based on unified format data, it constructs a dynasty style discriminator network to extract body geometric features, obtains era style constraint parameters through the style discriminator, performs style decoupling and reparameterization encoding on body features, and generates a sequence of body feature vectors with era style constraints. The computation module is used to parse ritual text and construct a semantic co-occurrence graph using a sequence of body posture feature vectors, which serves as the prior topology of the graph attention network. The vector sequence is used as the node feature input to calculate the cross-modal attention weights between body posture nodes and text nodes, and high-confidence edges are selected to construct a set of body posture-context association edges that conform to the ritual constraints. The building module is used to define the Peirce semiotic triple structure based on the body-context association edge set, reproduce the body-object-interpretation term, map the association edge to triple instances, perform dynamic weight propagation and hierarchical aggregation based on the attention mechanism, and build the initial network of the knowledge graph with a cultural symbol hierarchical structure. Fusion Module: Based on the initial network of the knowledge graph, it identifies incomplete posture nodes and builds a self-supervised reasoning model based on graph neural networks. It uses the topological and attribute features of adjacent complete nodes to perform link prediction and feature generation, completes the body semantic features of incomplete nodes, and finally fuses them to build a historical legacy body language database and knowledge graph.

[0011] Thirdly, this application also provides a multi-constraint knowledge graph construction device for historical body language, including: Memory, used to store computer programs; A processor is used to implement the steps of the multi-constraint knowledge graph construction method for historical legacy morphology when executing the computer program.

[0012] Fourthly, this application also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for constructing a multi-constraint knowledge graph based on historical legacy body language.

[0013] The beneficial effects of this invention are as follows: This invention proposes a method and system for constructing a knowledge graph of historical body language that integrates constraints of dynastic style, ritual context, and semiotic structure. It extracts and decouples the aesthetic characteristics of a particular era by constructing a dynastic style discriminator network, establishing a conditional mapping relationship between body language features and historical periods. A graph attention mechanism is used to achieve cross-modal association between body language nodes and ritual text nodes, constructing a set of body language-context association edges that conform to ritual logic. Based on Peirce's semiotic triple theory, the association edges are formalized into a topological structure of representation-object-interpretation, realizing hierarchical organization and dynamic weight propagation of cultural symbols. Furthermore, a self-supervised graph neural network is used to perform topological reasoning and semantic feature completion on incomplete nodes, ultimately constructing a database and knowledge graph of historical body language with multi-level semantic associations and historical context constraints.

[0014] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the multi-constraint knowledge graph construction method for historical morphology as described in this embodiment of the invention; Figure 2 This is a schematic diagram of the multi-constraint knowledge graph construction system for historical morphology as described in this embodiment of the invention; Figure 3 This is a schematic diagram of the structure of the multi-constraint knowledge graph construction device for historical morphology as described in this embodiment of the invention.

[0017] In the diagram: 701, Acquisition module; 702, Encoding module; 703, Calculation module; 704, Construction module; 705, Fusion module; 800, Multi-constraint knowledge graph construction device for historical body language; 801, Processor; 802, Memory; 803, Multimedia component; 804, I / O interface; 805, Communication component. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0020] This embodiment provides a method for constructing a multi-constraint knowledge graph of historical body language.

[0021] See Figure 1 The figure shows that the method includes steps S100, S200, S300, S400 and S500.

[0022] S100: Acquire 3D point cloud scanning data, multispectral imaging data, and body semantic annotation data from archaeological sites of historical human figures; perform rigid body registration and deformation field correction of point clouds, registration and reflectivity standardization of multispectral bands, and synchronization of timestamps and spatial coordinate transformation of annotation data; and fuse them to generate unified format data containing 3D geometric coordinates, material reflection characteristics, and body semantic labels.

[0023] It is understood that step S100 includes S101, S102, and S103, wherein: S101. Perform rigid body registration based on anatomical landmarks on the three-dimensional point cloud data. Select human bony landmarks with dynastic characteristics as the registration reference. Then, use thin plate spline function to perform deformation field correction, reconstruct the surface mesh topology of the surviving human figure, and extract the vertex coordinate set and normal vector field. It should be noted that this step addresses the issues of morphological distortion and pose instability caused by long-term geological pressure, temperature and humidity changes, and oxidation erosion in unearthed historical human figures. A data preprocessing workflow integrating anatomical priors and geometric deformation modeling was constructed. Specifically, a system of bony landmarks with dynastic distinctiveness was first established based on archaeological typology. Examples include the vertebral curvature features of Tang Dynasty female attendant figurines, the geometric relationship points of the scapular girdle of Han Dynasty storyteller figurines, and the folds and knots of clothing in Song Dynasty statues. These landmarks not only possess anatomical stability but also carry information about the stylistic paradigms of specific historical periods, distinguishing them from the random sampling points or geometric extrema relied upon by the ICP algorithm used in general 3D scanning.

[0024] In terms of registration strategy, rigid body registration based on landmark points is adopted as the coarse registration stage. The initial transformation matrix is ​​determined by minimizing the Euclidean distance error between the landmark point sets, unifying the point cloud data acquired from different perspectives or different excavation batches into a standard coordinate system. Furthermore, for the non-rigid deformations commonly found in the remains (such as the misaligned bonding after the neck fracture of terracotta figurines and the warping deformation of wooden statues), a deformation field correction model is constructed using thin-plate spline functions. This model uses landmark points as control points and calculates the deformation gradient field of the entire point cloud space by solving the biharmonic equation, achieving elastic restoration of the twisted parts. This process differs from simple geometric smoothing or manual repair; instead, it is based on the principles of materials mechanics to reverse-model historical deformation, restoring the original body structure while preserving surface details and textures.

[0025] In the reconstruction phase, a surface mesh generation algorithm based on Delaunay triangulation was employed, combined with conformal parameterization to optimize vertex distribution, ensuring that subsequent multispectral data could be accurately mapped to the corresponding geometric locations. The calculation of the normal vector field used a weighted neighborhood plane fitting method, with special treatment given to surface roughness regions caused by weathering, providing an accurate geometric basis for subsequent material reflection analysis.

[0026] S102. Perform dark current correction and flat field correction on the multispectral imaging data, extract the reflectance spectrum curves from the visible to near-infrared bands, obtain the material feature code through principal component analysis for dimensionality reduction, map the material feature code to the corresponding vertex index position of the surface mesh, and establish the material-geometry relationship. It should be noted that, to address the redundancy of high-dimensional spectral data, principal component analysis was used for dimensionality reduction, retaining the top K principal components covering more than 95% of the variance as material feature encodings. This encoding not only compresses the data volume but also maps the original spectral space to the material semantic space, enabling different pigment types (such as red lead, azurite, and ochre) to form distinguishable cluster structures in the feature space.

[0027] The key innovation lies in establishing a precise index mapping between material feature encoding and geometric mesh vertices. Through joint optimization of camera calibration parameters and point cloud spatial coordinates, the correspondence between multispectral image pixels and 3D mesh vertices is calculated. A centroid-based interpolation method is used to transfer pixel-level material properties to the vertex level. This mapping overcomes the two-dimensional limitations of traditional texture mapping, enabling material property queries on arbitrary cross-sections in 3D space, providing a data foundation for subsequent analysis of clothing patterns, painting techniques, and pigment layer relationships. In practical applications, this association mechanism can support the inference of the original color of faded areas and identify material degradation areas caused by environmental factors by analyzing the continuity of spectral features of adjacent vertices.

[0028] This step achieves deep integration of material features and geometric topology, constructs a unified data representation that supports cross-dimensional queries, and provides data support for body semantic analysis based on material properties.

[0029] S103. Perform timestamp-based alignment and coordinate system transformation based on coordinate transformation matrix on the micro-action semantic annotation data recorded at the archaeological site, bind the action semantic tags to the corresponding vertex indices, and fuse them to generate data in a unified format. The calculation formula is as follows: In the formula, To standardize the format of data tensors, For spatiotemporal alignment fusion operators, For the registered point cloud data matrix, This is the multispectral reflectance characteristic matrix. For the semantic annotation tensor of micro-actions.

[0030] It should be noted that during the archaeological excavation process, artifact restorers and archaeologists usually record the posture and semantics of the remaining human figures (such as the micro-movement units of clasping hands, kneeling, and dancing) and ritual context through visual observation. These records exist in the form of text descriptions, sketches or videos, and lack a direct spatiotemporal connection with the 3D scanning data.

[0031] First, timestamp-based sequence alignment is performed to synchronize the timestamps of on-site video recordings, motion capture data (using portable inertial sensors for auxiliary recording if available), and 3D scanning, establishing a correspondence between observation behavior and data acquisition time. Then, a coordinate transformation matrix is ​​used to transform the local coordinate system of the site (based on the excavation square) to the world coordinate system of the 3D model. This transformation matrix is ​​obtained through registration calculations between on-site control points and model marker points.

[0032] The core lies in binding abstract action semantic labels to specific vertex index sets. Through interactive annotation tools or automatic segmentation algorithms based on pose estimation, the set of vertices constituting specific micro-movements (such as the vertex group of the outstretched arms corresponding to the "clasped hands" gesture) is determined, and semantic labels are written into the data structure of the corresponding vertices as attributes. This binding is not a simple global label, but a fine-grained local semantic mapping, supporting the decomposition and analysis of complex poses. For example, a terracotta figurine simultaneously depicting "kneeling" and "playing the zither" poses has its lower limb vertices labeled "kneeling" and its upper limb vertices labeled "playing the zither," achieving a local decomposition of pose semantics. Specifically, this step establishes a direct mapping channel from physical space to semantic space, enabling archaeological observation knowledge to be embedded in geometric data in a structured form, providing semantic anchors for subsequent cross-modal knowledge reasoning.

[0033] S200. Based on unified format data, a dynasty style discriminator network is constructed to extract body geometric features. The style discriminator obtains the era style constraint parameters, and the body features are decoupled and reparameterized for style encoding, thereby generating a sequence of body feature vectors with era style constraints.

[0034] It is understood that step S200 includes S201, S202, and S203, wherein: S201. Construct a body style discriminator network based on dynasty classification. Input the geometric coordinate matrix in the unified format data, extract body proportion features and posture angle features through graph convolutional layers, and output the dynasty style probability distribution through fully connected layers and softmax classifiers. It should be noted that this step constructs a dynasty style discriminator network for historical portraits, aiming to extract aesthetic paradigm features with era-specific characteristics from geometric forms. Unlike the voxelized representation or projection features relied upon by general 3D shape classification tasks, this method directly operates on the surface mesh topology generated in step one, using graph convolutional layers as the feature extraction backbone network.

[0035] Specifically, the geometric coordinate matrix in the uniform format data is regarded as graph structure data, where grid vertices constitute graph nodes and edges are determined by the adjacency relationship of the grid or by K-nearest neighbor connections constructed based on geodesic distance.

[0036] In this system, the graph convolutional layer updates the representation of the central node by aggregating the feature information of neighboring vertices, and its convolutional kernel design takes into account both geometric properties and topological structure. Regarding body proportion feature extraction, the network learns to capture significant differences in anthropometric relationships across dynasties, such as the differences in the numerical distribution of macroscopic proportion parameters like shoulder-to-hip ratio and waist-to-hip ratio between the plump figures of Tang Dynasty female figurines and the slender proportions of Song Dynasty statues. In terms of posture angle feature extraction, the network focuses on the relative rotation angles between limb segments, such as the pattern differences in the statistical features of joint angles between the exaggerated dynamics of Han Dynasty storytelling figurines and the dignified postures of Tang Dynasty ceremonial figurines.

[0037] Through multi-layer graph convolution stacking, the network gradually abstracts from local geometric details to global stylistic features. Then, a fully connected layer maps the high-dimensional graph features to a dynasty category space, and a softmax classifier outputs a normalized probability distribution of dynasty styles. This probability distribution serves not only as a classification basis but also as a soft label for style constraints in subsequent steps, guiding the structured learning of the latent space. In practical applications, this discriminator can be extended into a fine-grained style analysis tool to distinguish regional stylistic differences within the same dynasty, such as the subtle differences between the style of Chang'an during the High Tang period and the styles of local kilns, or to identify style mixing phenomena under cross-cultural influence, such as the fusion of Hu (non-Han) and Han (non-Han) styles found in relics along the Silk Road.

[0038] S202. Obtain latent variables of era style through style discriminator, perform style decoupling operation on geometric features, separate posture content features from dynasty style features, and establish conditional latent space mapping relationship, where content features retain posture ontology information and style features encode era aesthetic features. Specifically, in this step, the style decoupling operation employs a variant of the encoder-decoder architecture. The content encoder focuses on extracting dynastic-independent pose ontology features, such as the spatial configuration of limbs and the functional semantics of actions. The style encoder, on the other hand, extracts dynastic-specific aesthetic features, such as the smoothness of lines, the fullness of volume, and the degree of exaggeration in movement. Through gradient inversion layers or mutual information minimization constraints, it is ensured that no style information remains in the content features, and that no pose content is mixed into the style features.

[0039] The establishment of conditional latent space mapping relationships is manifested in treating style latent variables as conditional variables and content features as explained variables, thus constructing a conditional probability distribution. Specifically, the posterior distribution of content features is conditional on style latent variables; that is, given a specific dynasty's style, the posture content follows a probability distribution under the aesthetic constraints of that era. This mapping relationship makes the latent space exhibit structured geometry: moving along the style dimension corresponds to the aesthetic transition between different dynasties; moving along the content dimension corresponds to the continuous changes in different postures within the same dynasty. In practical research extensions, this latent space can support the visual analysis of art history style evolution, observe the gradual trajectory of aesthetic paradigm changes through interpolation calculations, or identify the impact of historical events corresponding to style mutation nodes; it can also be used to detect anomalous style remains, such as the position of disputed artifacts in the latent space deviating from the cluster center of their claimed dynasty.

[0040] This step decouples the semantics of pose from the aesthetics of different eras, enabling the same pose to generate differentiated features under the constraints of different dynasties' styles, thus providing a style-controllable feature foundation for subsequent cross-modal association.

[0041] S203. Based on the reparameterization technique, style latent variables and content features are fused together, and a sequence of body feature vectors with era style constraints is generated through a style conditional encoder.

[0042] It should be noted that, specifically, the content encoder outputs the Gaussian distribution parameters of the content features, while the style encoder outputs the sampled values ​​of the style latent variables. By using reparameterization techniques, the random sampling process is transformed into a combination of deterministic calculation and noise superposition, ensuring the feasibility of end-to-end gradient backpropagation.

[0043] The design of the style-conditional encoder is manifested in the following way: using style latent variables as modulation parameters, and performing affine transformation on content features through a feature transformation layer to achieve style-conditional constraints on content. This modulation mechanism is different from simple vector concatenation. Instead, it delves into the channel and spatial dimensions of the feature space, adjusting the activation intensity and distribution of each dimension of the content features, so that the output features not only retain the functional semantics of the original posture, but also possess the aesthetic characteristics of a specific dynasty.

[0044] Among them, the generated sequence of body feature vectors forms the node feature basis of the subsequent knowledge graph. Each vector not only encodes the local geometric attributes of a single vertex, but also contains the aesthetic expression of the body part to which the vertex belongs under the style of a specific dynasty. In practical applications, this encoding mechanism supports style-controllable relic restoration: for damaged parts, feature vectors under different dynasty style assumptions can be generated respectively on the premise of maintaining the rationality of posture function for experts to evaluate style consistency; it can also be used for generative data augmentation. By sampling in different style subspaces, the style diversity of training data can be expanded, and the generalization ability of downstream tasks for cross-dynasty relics can be improved. In addition, this feature sequence can be directly input into an interpretability analysis tool. Through methods such as activation maximization or concept activation vectors, the geometric regions that contribute the most to dynasty classification can be identified to assist archaeologists in understanding the visual basis of style discrimination.

[0045] That is to say, in this step, random generation of features under style constraints is achieved through reparameterization, balancing the certainty and diversity and diversity of feature representations, and building a bridge connecting geometric observations and semantic understanding.

[0046] S300. Use the sequence of body feature vectors to analyze the ritual text and construct a semantic co-occurrence graph as the prior topological structure of the graph attention network; take the vector sequence as the input of node features, calculate the cross-modal attention weights between body nodes and text nodes, and screen high-confidence edges to construct a set of body-context association edges that meet the ritual constraints.

[0047] It can be understood that in this step S300, it includes S301, S302 and S303, where: S301. Analyze the dependency syntactic structure of the ritual text, extract ritual action words and ritual scene descriptions, and construct a word co-occurrence graph as the prior topological structure, where nodes represent ritual concepts, and edges represent the co-occurrence relationship of words within a specific window, forming a semantic network of the ritual context; It should be noted that when constructing the word co-occurrence graph, a sliding window mechanism is used to count the word co-occurrence frequency at the syntactic tree level rather than simply on a linear sequence. The window range is set according to the sentence reading habits and sense group division of ancient Chinese, covering the reference relationships within a single sentence and across sentences. The nodes in the graph not only include the words themselves, but also through word sense disambiguation and ontology mapping, associate word instances with ritual concept ontologies (such as mapping "bowing twice" to the "bowing ritual" concept node and "sitting facing east" to the "orientation ritual" concept node). The edge weights reflect the ritual logic tightness between concepts, rather than simply statistical co-occurrence intensity. For example, the association weight between "presenting the wine cup" and "ancestral temple" is higher than that between "presenting the wine cup" and "marketplace", even if the latter co-occurs accidentally in the corpus.

[0048] This semantic network, injected as a priori topological structure into the subsequent graph attention network, essentially provides the model with archaeological prior knowledge constraints, preventing the model from establishing false associations that contradict historical context. In practical research, this network can support the graph-based completion of ritual knowledge: by analyzing structural holes in the network, it can identify ritual elements that are missing from historical records but logically should exist; it can also combine spatiotemporal information to construct a dynamically evolving ritual knowledge graph, observing the changes in network centrality indicators of specific rituals (such as the "kneeling ceremony") from the pre-Qin period to the Ming and Qing dynasties, quantifying the evolution of ritual systems.

[0049] S302. Inject the body feature vector sequence as node features into the graph attention network, and calculate the cross-modal attention weights between body nodes and text nodes. The query vector comes from the body features, and the key vector and value vector come from the text embedding, forming a cross-modal attention mechanism. This step achieves cross-space mapping between visual body posture features and textual ritual semantics. The core lies in constructing a heterogeneous graph attention network to process the body posture feature vector sequence generated in step two and the text node embeddings constructed in step three. Unlike global feature alignment in traditional cross-modal retrieval, this method establishes fine-grained node-level correspondences at the graph structure level. Specifically, each node in the body posture feature vector sequence (corresponding to a specific body part or pose unit of the surviving human image) is used as the source of the query vector, projected onto the shared cross-modal semantic space through a learnable linear transformation matrix; the word embedding vectors of the ritual text nodes (encoded by a pre-trained Classical Chinese language model) are used as the source of the key and value vectors, also projected onto this shared space.

[0050] The calculation of cross-modal attention weights follows the scaled dot product attention mechanism, but introduces a mask constraint based on human anatomy: attention is only calculated between body nodes and text nodes that have semantic association potential. For example, upper limb nodes are preferentially associated with action words such as "hold," "offer," and "worship," while lower limb nodes are preferentially associated with words such as "kneel," "stand," and "walk." The mask matrix filters out anatomically unreasonable cross-modal associations, reducing noise interference. Each row of the attention weight matrix represents the distribution of attention of a specific body node to all ritual concepts, and each column represents the activation intensity of a specific ritual concept to all body parts.

[0051] S303. Based on ritual logic rules, select high-confidence association edges and construct a set of body posture-context association edges that conform to ritual constraints. Each edge is associated with a specific ritual context and action semantics, forming a body posture-context bipartite graph structure.

[0052] It should be noted that, firstly, an attention threshold is set to filter low-confidence associations, and then a rule engine is applied to perform logical verification, eliminating candidate edges that violate ritual logic. For example, if the model establishes a high-weight association between dance postures and funeral scenes based on statistical co-occurrence, but the rule base defines funeral scenes as only associated with actions such as crying and bowing, then that edge will be suppressed or downweighted.

[0053] The final constructed set of body posture-context association edges forms a strict bipartite graph structure: the left set of nodes consists of body posture feature nodes of the surviving human figure, and the right set of nodes consists of contextual concept nodes of the ritual text. Edges exist only between the left and right sides, with no interconnection between nodes on the same side. Each edge carries multiple attributes: cross-modal attention weight, rule validation score, and association type label.

[0054] In practical research, this structure can support the automatic reconstruction of ritual processes: by analyzing the connected components and path structure in the bipartite graph, the order of actions in a complete ritual can be inferred; it can also be used to detect contradictions between documents and physical objects: when a certain surviving physical node cannot be associated with any known ritual context through a high-confidence edge, or is associated with multiple mutually exclusive ritual types, an expert review mechanism is triggered to indicate possible dating errors or functional misjudgments.

[0055] By using ritual logic rules for data post-processing, historical spurious correlations in statistical associations were eliminated, and a body-context semantic association network that combines data support with academic rigor was constructed.

[0056] S400. Based on the body-context association edge set, define the Peirce semiotic triple structure to reproduce body-object-interpretation, map the association edges to triple instances, perform dynamic weight propagation and hierarchical aggregation based on attention mechanism, and construct the initial network of knowledge graph with cultural symbol hierarchical structure.

[0057] It is understood that in this step, S400 includes S401, S402, and S403, wherein: S401. Map the body posture-context association edge set to Peirce's semiotic triple structure, instantiate it into a representation-object-interpretation triple, where the representation corresponds to the graphic representation of body posture, the object corresponds to the referential object of ritual function, and the interpreter corresponds to the interpretative symbol of cultural connotation, thus forming a semiotic instance. It should be noted that the specific mapping relationship is constructed as follows: The representation corresponds to the visual graphic representation of the posture of the surviving human figure, that is, the geometric form encoded by the sequence of posture feature vectors with stylistic constraints of the era generated in step two, including the spatial configuration of the limbs, dynamic curves and proportional relationships. This is the sensory carrier of the symbol; the object corresponds to the actual functional object that the posture points to in ritual practice, such as the social behavior of "relationship between ruler and subject" or "communication between humans and gods" pointed to by the kneeling posture. This is the referential content of the symbol; the interpretive item corresponds to the interpretive effect of the cultural community on the posture, that is, the specific cultural connotation and emotional resonance, such as abstract concepts such as awe, piety, and hierarchical order. This is the realization of the symbol's meaning.

[0058] During instantiation, the edges of the bipartite graph are semantically labeled and assembled into triplets: each posture-context association edge is parsed into a complete triplet consisting of a posture node (representation), a ritual function node (object), and a cultural connotation node (interpretation). For example, the posture of the terracotta figurine with hands clasped in front of the chest unearthed from a Tang Dynasty tomb has a representation of the geometric features of arms crossed in front of the chest and the body slightly bent. The object points to the ritual function of "funeral worship" (a specific bow in funeral rites) or "daily audience" (daily court audience), while the interpretation maps to the cultural psychology of "sorrow," "respect," or "humility." This instantiation is not a simple one-to-one correspondence, but allows a representation to point to different objects through different interpretations (such as the same clasped-hand posture having vastly different interpretations in Buddhist statues and Confucian rituals), forming a multi-interpretive network of symbols.

[0059] In practical research, this mapping mechanism supports the determination of symbol types: based on Peirce's classification of symbols (pictorial symbols, indicative symbols, and symbolic symbols), it analyzes the relationship between specific postures and their objects—such as the pictorial relationship based on similarity between a "dance" posture and a joyous ritual, the indicative relationship based on causality between a "pointing" gesture and a specific location, and the symbolic relationship based on convention between "dragon patterns" and imperial power. This classification directly affects the weighting strategy for subsequent knowledge reasoning. Furthermore, a framework for tracing symbol evolution can be constructed, quantifying the drift trajectory of cultural meaning by comparing the changes in objects and interpretants of the same representation in different historical periods, such as observing the transformation of the Hu kneeling posture from a foreign cultural interpretant in the Wei-Jin period to a daily ritual interpretant in the Tang Dynasty.

[0060] It establishes a theoretical leap from empirical body posture-context association to formal cultural symbols, enabling the physical form of surviving human figures to be incorporated into the semiotic meaning generation mechanism, and providing an ontological foundation for computational modeling of cultural meaning.

[0061] S402. Perform dynamic weight propagation based on attention mechanism, aggregate neighborhood symbol features along triple structure, and update node representation. The propagation weight is determined by the cultural association strength and ritual importance coefficient between symbols, thereby realizing hierarchical propagation of symbol features. It should be noted that the specific implementation of dynamic weight propagation is as follows: for each triple node, when aggregating the features of its neighboring nodes, the attention weight is determined by the weighted product of the feature similarity between the query node and the key node, the cultural association strength coefficient, and the ritual importance coefficient. This design ensures that information propagation follows not only geometric proximity but also cultural logical proximity. The feature update process adopts a hierarchical strategy: feature updates for low-level nodes (concrete physical instances) focus on preserving geometric details and integrating local contexts, while feature updates for high-level nodes (abstract cultural concepts) focus on integrating and generalizing multi-source interpretations. Through multi-layered propagation and stacking, the feature representation of each node gradually incorporates global information from its cultural context, forming a culturally contextualized feature embedding. In practical applications, this mechanism supports analogical reasoning of cultural symbols: by calculating the distance between the updated node features in the embedding space, similar cultural symbols across dynasties and regions can be discovered (such as comparing the similarity of the Han Dynasty's "long-sleeved dance" and the Tang Dynasty's "Hu Xuan dance" at the explanatory level), or cultural fusion phenomena can be identified (such as intermediate state nodes in the feature space between Buddhist statue gestures and local ritual gestures).

[0062] S403. Based on the hierarchical relationship of cultural symbols, reorganize the node topology and construct an initial network of knowledge graph with a hierarchical structure of cultural symbols, where high-level nodes represent abstract cultural concepts and low-level nodes represent concrete physical instances.

[0063] It should be noted that this step constructs a hierarchical knowledge topology from material remains to abstract culture, realizing multi-scale organization and vertical reasoning capabilities of cultural symbols, and laying a structural foundation for the semantic richness and reasoning depth of the knowledge graph.

[0064] S500, based on the initial network of the knowledge graph, identifies incomplete posture nodes and constructs a self-supervised reasoning model based on graph neural networks. It uses the topological and attribute features of adjacent complete nodes to perform link prediction and feature generation, completes the body semantic features of incomplete nodes, and finally integrates and constructs a historical legacy body language database and knowledge graph.

[0065] It is understood that in this step, S500 includes S501, S502, and S503, wherein: S501. Based on the initial network of the knowledge graph, identify incomplete nodes with missing body semantic features, and construct a topological mask matrix to identify the region to be completed. The mask value of 1 indicates that the node's features are incomplete, and the mask value of 0 indicates that the node's features are complete and usable. It should be noted that the incompleteness identification process first performs geometric integrity detection based on the unified format data generated in step one: by analyzing the topological connectivity of the surface mesh, the open contours formed by the boundary edges are identified, and the geometric range of the missing region is determined. Then, the geometrical incompleteness is mapped to the initial network of the knowledge graph to locate the corresponding set of body nodes. Here, two types of incompleteness are distinguished: feature incompleteness refers to nodes existing but the body feature vector generated in step two is incomplete or has low confidence, typically corresponding to areas with surface wear but discernible morphology; topological incompleteness refers to nodes completely missing or isolated in the graph, typically corresponding to areas where limbs are completely broken or lost. For feature incompleteness, a node-level mask vector is constructed, using binary values ​​to identify the validity of the feature dimension; for topological incompleteness, an adjacency matrix mask is constructed to identify the connection relationship between the missing node and its neighboring nodes.

[0066] The construction of the topological mask matrix not only identifies the location of the missing parts but also records the archaeological contextual information of the missing parts, including the anatomical labels of the missing areas, the geometric features of the missing boundaries, and the stylistic attributes of adjacent complete areas. This information is injected as a priori constraints into the subsequent completion process to ensure that the completion results conform to the rationality of human anatomy and the consistency of historical style.

[0067] In practical research, this masking mechanism can support the automatic classification of defect types: by analyzing the curvature characteristics and material cross-section of the defect boundary, it can distinguish different causes such as mechanical fracture, weathering and peeling, and firing defects, providing a basis for cultural relic protection decisions; it can also combine excavation records to link defect information with archaeological site photos and restoration history records to construct a digital pathological archive of the remains.

[0068] S502. Construct a self-supervised inference model based on graph neural networks. Utilize the topological connection relationship and attribute features of adjacent complete nodes to generate the completion features of incomplete nodes through graph convolution operations. The model training adopts the mask language modeling paradigm, randomly masking some node features for self-supervised learning. It should be noted that this step constructs a self-supervised reasoning model based on graph neural networks. Utilizing the topological connections and attribute features of adjacent complete nodes in the initial knowledge graph network, it generates completion features for incomplete nodes through a message passing mechanism. Unlike traditional geometric symmetry completion or texture synthesis methods based on statistical models, this method performs semantic reasoning based on the relevance of cultural context: it assumes that the body language semantics of the incomplete part can be inferred from its adjacent nodes in the ritual function network, i.e., using the cultural logic of the context to complete missing words.

[0069] The model architecture employs an encoder-decoder graph autoencoder, where the encoder aggregates feature information from neighboring nodes through graph convolutional layers, and the decoder reconstructs the feature representation of incomplete nodes. The training process utilizes a masked language modeling paradigm: during training, features of partially complete nodes are randomly masked, forcing the model to predict the masked content based on the topological context of its neighboring nodes. This self-supervised strategy requires no additional labeled data, fully leveraging the distribution patterns of existing complete remnants to learn the implicit connections within the cultural context. In graph convolution operations, message passing weights are determined by the strength of cultural connections and spatial proximity between nodes, ensuring that information from functionally related and spatially adjacent nodes receives greater weight. This achieves semantic feature reasoning based on cultural context connections, overcoming the limitations of pure geometric completion and ensuring that the completion of incomplete parts possesses both ritualistic rationality and historical and cultural consistency.

[0070] S503. Integrate the completion features with the existing network, perform consistency checks, and construct the final historical legacy body language database and knowledge graph, in which the completion nodes and the original nodes maintain topological coherence and semantic consistency.

[0071] It should be noted that this step deeply integrates the completed features generated by S502 with the initial network of the knowledge graph and performs multi-dimensional consistency checks to construct the final historical legacy body language database and knowledge graph. The integration process involves not only the replacement or concatenation of feature vectors, but also the reconstruction of topological relationships and the verification of semantic logic. For nodes with incomplete features, the completed features and the original incomplete features are weighted and integrated, with the weights dynamically adjusted based on the confidence level of the original features and the determinism of the completed features. For nodes with incomplete topology, new nodes are created in the graph and edge connections are established with neighboring nodes, with edge weights initialized based on spatial proximity and cultural association strength.

[0072] The consistency check comprises two levels: topological coherence check ensures that the completed node and its adjacent nodes form a reasonable grid connection in geometric space, without overlap or gaps, and maintain connectivity in the graph structure to avoid isolated subgraphs; semantic consistency check ensures that the ritual function label of the completed node is compatible with the function labels of adjacent nodes in the ritual process logic. For example, after the lower body node of the "kneeling" posture is completed, its upper body node should not have the mutually exclusive action of "raising a cup," unless there is a specific ritual transition context. The check process utilizes the semiotic hierarchy structure constructed in step four to check whether the attribution of the completed node in the higher-level cultural concept is coordinated with that of adjacent nodes.

[0073] The final constructed database of historical remains' body language not only contains the complete, incomplete map but also records uncertainty information from the completion process, including confidence scores for each completion scheme, the set of neighboring nodes used, and a list of violated constraints. This provides an auditable completion decision path for subsequent research. In practical research, this mechanism can support archaeological verification of virtual restorations: by comparing the logical consistency scores of different completion schemes in the ritual knowledge graph, it helps select the restoration scheme that best matches historical reality; it can also be used to detect errors in existing restorations: by inputting early restorations from museums into the system, if there are significant differences between the completed features and the original restored parts, it indicates possible misjudgments in restoration, such as mistakenly restoring a Tang Dynasty "kneeling" posture as a Song Dynasty "sitting upright."

[0074] That is, by using a fusion and verification mechanism, the consistency of the completion results in both geometric topology and semantic logic is ensured, and a historical body language knowledge graph with completeness, accuracy and interpretability is constructed.

[0075] In summary, this invention achieves spatiotemporal alignment and structured fusion of multi-source heterogeneous archaeological data by constructing a unified format data that integrates three-dimensional geometric coordinates, material reflection features, and body posture semantic tags. This effectively solves the problem of disconnected associations caused by the isolation of geometric, material, and semantic data in traditional methods. By constructing a dynasty style discriminator network and performing style decoupling and reparameterized encoding, and introducing the aesthetic paradigm of the era as a constraint, this invention significantly improves the distinguishability and stylistic consistency of body posture feature vectors across historical periods, overcoming the defect of cross-dynastic body posture semantic confusion. Furthermore, by constructing a semantic co-occurrence graph through parsing ritual texts and employing cross-modal graph attention... The force mechanism calculates the association weights between posture nodes and text nodes, establishing a posture-context mapping relationship that conforms to ritual logic, filling the gap in existing technologies where there is a lack of semantic bridge between posture and ritual function. By mapping the associated edges to the Peirce semiotic triple structure and performing dynamic weight propagation, a knowledge representation framework with a hierarchical structure of cultural symbols is constructed, realizing hierarchical reasoning from specific postures to abstract cultural connotations. By constructing a self-supervised graph neural network to perform topological reasoning and feature completion on incomplete nodes, the missing structure of the knowledge graph caused by the physical defects of the legacy is effectively repaired, significantly improving the integrity and coherence of the knowledge graph. Example 2:

[0076] like Figure 2 As shown, this embodiment provides a multi-constraint knowledge graph construction system for historical body language. See [link to documentation]. Figure 2 The system includes: Acquisition Module 701: Used to acquire 3D point cloud scanning data, multispectral imaging data and body semantic annotation data of historical human figures, perform rigid body registration and deformation field correction of point cloud, registration and reflectivity standardization of multispectral bands, and synchronization of timestamps and spatial coordinate transformation of annotation data, and fuse them to generate unified format data containing 3D geometric coordinates, material reflection characteristics and body semantic labels. Encoding module 702: It is used to construct a dynasty style discriminator network based on unified format data to extract body geometric features, obtain era style constraint parameters through the style discriminator, perform style decoupling and reparameterization encoding on body features, and thus generate a sequence of body feature vectors with era style constraints. Calculation module 703: Used to parse ritual text and construct semantic co-occurrence graph using body feature vector sequence, which is used as the prior topology of graph attention network; use vector sequence as node feature input, calculate cross-modal attention weights between body nodes and text nodes, and select high-confidence edges to construct a set of body-context association edges that conform to ritual constraints; Module 704: Based on the body-context association edge set, it defines the Peirce semiotic triple structure, reproduces the body-object-interpretation, maps the association edges to triple instances, performs dynamic weight propagation and hierarchical aggregation based on the attention mechanism, and constructs the initial network of the knowledge graph with a cultural symbol hierarchical structure. Fusion Module 705: Based on the initial network of the knowledge graph, it identifies incomplete posture nodes and constructs a self-supervised reasoning model based on graph neural networks. It uses the topological and attribute features of adjacent complete nodes to perform link prediction and feature generation, completes the body semantic features of incomplete nodes, and finally fuses and constructs a historical legacy body language database and knowledge graph.

[0077] Specifically, the acquisition module 701 includes: Reconstruction Unit: Used to perform rigid body registration based on anatomical landmarks on 3D point cloud data. Human bony landmarks with dynastic characteristics are selected as registration references. Then, thin plate spline functions are used to perform deformation field correction, reconstruct the surface mesh topology of the surviving human figure, and extract the vertex coordinate set and normal vector field. Extraction unit: used to perform dark current correction and flat field correction on multispectral imaging data, extract reflectance spectrum curves from visible to near-infrared bands, obtain material feature codes through principal component analysis for dimensionality reduction, map the material feature codes to the corresponding vertex index positions of the surface mesh, and establish material-geometry association; Transformation Unit: Used to perform time-stamp-based alignment and coordinate system transformation based on coordinate transformation matrix on the micro-action semantic annotation data recorded at the archaeological site, bind action semantic tags to corresponding vertex indices, and fuse them to generate data in a unified format. Its calculation formula is as follows: In the formula, To standardize the format of data tensors, For spatiotemporal alignment fusion operators, For the registered point cloud data matrix, This is the multispectral reflectance characteristic matrix. For the semantic annotation tensor of micro-actions.

[0078] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here. Example 3:

[0079] Corresponding to the above method embodiments, this embodiment also provides a multi-constraint knowledge graph construction device for historical speech expressions. The multi-constraint knowledge graph construction device for historical speech expressions described below and the multi-constraint knowledge graph construction method for historical speech expressions described above can be referred to each other.

[0080] Figure 3 This is a block diagram illustrating a multi-constraint knowledge graph construction device 800 for historical morphology, according to an exemplary embodiment. (e.g.) Figure 3 As shown, the multi-constraint knowledge graph construction device 800 for historical morphology includes a processor 801 and a memory 802. The device 800 also includes one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0081] The processor 801 controls the overall operation of the multi-constraint knowledge graph construction device 800 for historical speech, to complete all or part of the steps in the aforementioned multi-constraint knowledge graph construction method for historical speech. The memory 802 stores various types of data to support the operation of the multi-constraint knowledge graph construction device 800 for historical speech. This data may include, for example, instructions for any application or method operating on the multi-constraint knowledge graph construction device 800 for historical speech, as well as application-related data, such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, or buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the historical legacy morphology multi-constraint knowledge graph construction device 800 and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or one or more combinations thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0082] In an exemplary embodiment, the multi-constraint knowledge graph construction device 800 for historical morphology can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the aforementioned multi-constraint knowledge graph construction method for historical morphology.

[0083] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the multi-constraint knowledge graph construction method for historical morphology described above. For example, the computer-readable storage medium may be the memory 802 including the program instructions described above. These program instructions may be executed by the processor 801 of the multi-constraint knowledge graph construction device 800 for historical morphology to complete the multi-constraint knowledge graph construction method for historical morphology described above. Example 4:

[0084] Corresponding to the above method embodiments, this embodiment also provides a readable storage medium. The readable storage medium described below and the multi-constraint knowledge graph construction method for historical legacy morphology described above can be referred to in correspondence.

[0085] A computer program is stored on a readable storage medium, and when the computer program is executed by a processor, it implements the steps of the multi-constraint knowledge graph construction method for historical legacy morphology in the above method embodiments.

[0086] The readable storage medium can specifically be a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or any other readable storage medium capable of storing program code.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for constructing a multi-constraint knowledge graph of historical body language, characterized in that, include: Acquire 3D point cloud scanning data, multispectral imaging data, and body semantic annotation data from archaeological sites of historical human figures; perform rigid body registration and deformation field correction of point clouds, registration and reflectivity standardization of multispectral bands, and synchronization of timestamps and spatial coordinate transformation of annotation data; and fuse them to generate unified format data containing 3D geometric coordinates, material reflection characteristics, and body semantic labels. Based on unified format data, a dynasty style discriminator network is constructed to extract body geometric features. The style discriminator obtains the era style constraint parameters, and the body features are decoupled and reparameterized for style encoding, thereby generating a sequence of body feature vectors with era style constraints. Using the sequence of body posture feature vectors, we parse the ritual text and construct a semantic co-occurrence graph, which serves as the prior topology of the graph attention network. We use the vector sequence as the node feature input to calculate the cross-modal attention weights between body posture nodes and text nodes, and select high-confidence edges to construct a set of body posture-context association edges that conform to the ritual constraints. Based on the body-context association edge set, we define the Peirce semiotic triple structure to reproduce body-object-interpretation, map association edges to triple instances, perform dynamic weight propagation and hierarchical aggregation based on attention mechanism, and construct an initial network of knowledge graph with cultural symbol hierarchical structure. Based on the initial network of the knowledge graph, incomplete posture nodes are identified and a self-supervised reasoning model based on graph neural network is constructed. The topological and attribute features of adjacent complete nodes are used for link prediction and feature generation to complete the body semantic features of incomplete nodes. Finally, the historical body language database and knowledge graph are integrated to construct the knowledge graph.

2. The method for constructing a multi-constraint knowledge graph of historical body language according to claim 1, characterized in that, The process involves acquiring 3D point cloud scan data, multispectral imaging data, and body semantic annotation data from archaeological sites, performing rigid body registration and deformation field correction of the point cloud, multispectral band registration and reflectivity standardization, and synchronizing the timestamps and transforming the spatial coordinates of the annotation data. This data is then fused to generate unified format data containing 3D geometric coordinates, material reflection characteristics, and body semantic tags, including: Rigid body registration based on anatomical landmarks is performed on the 3D point cloud data. Human bony landmarks with dynastic characteristics are selected as the registration reference. Then, thin plate spline functions are used to perform deformation field correction, reconstruct the surface mesh topology of the surviving human figure, and extract the vertex coordinate set and normal vector field. Dark current correction and flat field correction are performed on the multispectral imaging data. Reflectance spectral curves from the visible to near-infrared bands are extracted. Material feature codes are obtained by dimensionality reduction through principal component analysis. The material feature codes are mapped to the corresponding vertex index positions of the surface mesh to establish material-geometry relationships. The micro-action semantic annotation data recorded at the archaeological site is aligned based on timestamps and transformed into a spatial coordinate system based on coordinate transformation matrices. Action semantic tags are then bound to corresponding vertex indices and fused to generate data in a unified format. The calculation formula is as follows: In the formula, To standardize the format of data tensors, For spatiotemporal alignment fusion operators, For the registered point cloud data matrix, This is the multispectral reflectance characteristic matrix. For the semantic annotation tensor of micro-actions.

3. The method for constructing a multi-constraint knowledge graph of historical body language according to claim 1, characterized in that, The method relies on unified format data to construct a dynasty style discriminator network to extract body posture geometric features. The style discriminator obtains era style constraint parameters, and the body posture features are decoupled and reparameterized for encoding, thereby generating a sequence of body posture feature vectors with era style constraints, including: A body style discriminator network based on dynasty classification is constructed. The geometric coordinate matrix in the uniform format data is input, and the body proportion features and posture angle features are extracted through graph convolutional layers. The probability distribution of dynasty style is output through fully connected layers and softmax classifier. The style discriminator obtains the latent variables of the era style, performs style decoupling operation on the geometric features, separates the posture content features from the dynasty style features, and establishes a conditional latent space mapping relationship. The content features retain the posture ontology information, and the style features encode the era aesthetic features. Based on the reparameterization technique, style latent variables are fused with content features, and a sequence of posture feature vectors with era-style constraints is generated through a style conditional encoder. The calculation formula is as follows: In the formula, It is a sequence of body feature vectors. The mean vector output by the content encoding network. The standard deviation vector output by the content encoding network. is a random noise vector, and ⊙ represents element-wise multiplication.

4. The method for constructing a multi-constraint knowledge graph of historical body language according to claim 1, characterized in that, The method utilizes a sequence of body posture feature vectors to parse the ritual text and construct a semantic co-occurrence graph, which serves as the prior topology of the graph attention network. The vector sequence is used as node feature input to calculate cross-modal attention weights between body posture nodes and text nodes. High-confidence edges are selected to construct a set of body posture-context association edges that conform to the ritual constraints, including: The dependency syntax structure of the ritual text is analyzed, ritual action vocabulary and ritual scene description are extracted, and a word co-occurrence graph is constructed as a prior topological structure, where nodes represent ritual concepts and edges represent the co-occurrence relationship of words within a specific window, forming a semantic network of ritual context; The sequence of body shape feature vectors is injected into the graph attention network as node features, and the cross-modal attention weights between body shape nodes and text nodes are calculated. The query vector is derived from the body shape features, and the key vector and value vector are derived from the text embedding, thus forming a cross-modal attention mechanism. Based on ritual logic rules, high-confidence association edges are selected, and a set of body posture-context association edges that conform to ritual constraints is constructed. Each edge is associated with a specific ritual context and action semantics, forming a body posture-context bipartite graph structure.

5. The method for constructing a multi-constraint knowledge graph of historical body language according to claim 4, characterized in that, The calculation formula for the cross-modal attention mechanism is as follows: In the formula, Let be the cross-modal attention weights between body node i and text node j. Represents the query vector. Represents the key vector. This represents a cross-modal relation matrix used to align the body feature space with the text semantic space. Let i represent the set of neighboring nodes of node i.

6. The method for constructing a multi-constraint knowledge graph of historical body language according to claim 1, characterized in that, Based on the set of body-context association edges, the system defines the Peirce semiotic triple structure, reproduces the body-object-interpretation term structure, maps association edges to triple instances, performs dynamic weight propagation and hierarchical aggregation based on an attention mechanism, and constructs an initial network of a knowledge graph with a hierarchical structure of cultural symbols, including: Mapping the body posture-context association set to Peirce's semiotic triple structure, instantiating it into a representation-object-interpretation triple, where the representation corresponds to the graphic representation of body posture, the object corresponds to the referential object of ritual function, and the interpreter corresponds to the interpretative symbol of cultural connotation, thus forming a semiotic instance; Dynamic weight propagation based on attention mechanism is performed, which aggregates neighborhood symbol features along the triple structure and updates the node representation. The propagation weight is determined by the cultural association strength and ritual importance coefficient between symbols, thereby realizing the hierarchical propagation of symbol features. Based on the hierarchical relationship of cultural symbols, the node topology is reorganized to construct an initial network of knowledge graph with a hierarchical structure of cultural symbols, where high-level nodes represent abstract cultural concepts and low-level nodes represent concrete physical instances.

7. The method for constructing a multi-constraint knowledge graph of historical body language according to claim 6, characterized in that, The calculation formula for the process of performing hierarchical feature propagation along the Peirce triple topology in the aggregation of neighborhood symbolic features along the triple structure based on the attention mechanism is as follows: In the formula, For symbol node i in the th +1 layer feature representation, To aggregate neighborhood symbol features, that is, to perform a weighted summation of the features of adjacent nodes in the triplet, σ represents a nonlinear activation function.

8. The method for constructing a multi-constraint knowledge graph of historical body language according to claim 1, characterized in that, The method, based on an initial knowledge graph network, identifies incomplete pose nodes and constructs a self-supervised reasoning model based on a graph neural network. It then utilizes the topological and attribute features of adjacent complete nodes for link prediction and feature generation, completing the body language semantic features of incomplete nodes. Finally, it integrates and constructs a historical legacy body language database and knowledge graph, including: Based on the initial network of the knowledge graph, incomplete nodes with missing body semantic features are identified, and a topological mask matrix is ​​constructed to identify the regions to be completed. A mask value of 1 indicates that the node's features are incomplete, and a mask value of 0 indicates that the node's features are complete and usable. A self-supervised inference model based on graph neural networks is constructed. The topological connection relationship and attribute features of adjacent complete nodes are used to generate the completion features of incomplete nodes through graph convolution operation. The model training adopts the mask language modeling paradigm, randomly masking some node features for self-supervised learning. The completion features are integrated with the existing network, consistency checks are performed, and the final historical legacy body language database and knowledge graph are constructed, in which the completion nodes and the original nodes maintain topological coherence and semantic consistency.

9. A multi-constraint knowledge graph construction system for historical body language, based on the multi-constraint knowledge graph construction method for historical body language as described in claim 1, characterized in that, include: The acquisition module is used to acquire 3D point cloud scanning data, multispectral imaging data and body semantic annotation data of historical human figures, and to perform rigid body registration and deformation field correction of point cloud, registration and reflectivity standardization of multispectral bands, and synchronization of timestamps and spatial coordinate transformation of annotation data. It integrates and generates unified format data containing 3D geometric coordinates, material reflection characteristics and body semantic tags. Encoding module: Based on unified format data, it constructs a dynasty style discriminator network to extract body geometric features, obtains era style constraint parameters through the style discriminator, performs style decoupling and reparameterization encoding on body features, and generates a sequence of body feature vectors with era style constraints. The computation module is used to parse ritual text and construct a semantic co-occurrence graph using a sequence of body posture feature vectors, which serves as the prior topology of the graph attention network. The vector sequence is used as the node feature input to calculate the cross-modal attention weights between body posture nodes and text nodes, and high-confidence edges are selected to construct a set of body posture-context association edges that conform to the ritual constraints. The building module is used to define the Peirce semiotic triple structure based on the body-context association edge set, reproduce the body-object-interpretation term, map the association edge to triple instances, perform dynamic weight propagation and hierarchical aggregation based on the attention mechanism, and build the initial network of the knowledge graph with a cultural symbol hierarchical structure. Fusion Module: Based on the initial network of the knowledge graph, it identifies incomplete posture nodes and builds a self-supervised reasoning model based on graph neural networks. It uses the topological and attribute features of adjacent complete nodes to perform link prediction and feature generation, completes the body semantic features of incomplete nodes, and finally fuses them to build a historical legacy body language database and knowledge graph.

10. The multi-constraint knowledge graph construction system for historical remnant body language according to claim 9, characterized in that, The acquisition module includes: Reconstruction Unit: Used to perform rigid body registration based on anatomical landmarks on 3D point cloud data. Human bony landmarks with dynastic characteristics are selected as registration references. Then, thin plate spline functions are used to perform deformation field correction, reconstruct the surface mesh topology of the surviving human figure, and extract the vertex coordinate set and normal vector field. Extraction unit: used to perform dark current correction and flat field correction on multispectral imaging data, extract reflectance spectrum curves from visible to near-infrared bands, obtain material feature codes through principal component analysis for dimensionality reduction, map the material feature codes to the corresponding vertex index positions of the surface mesh, and establish material-geometry association; Transformation Unit: Used to perform time-stamp-based alignment and coordinate system transformation based on coordinate transformation matrix on the micro-action semantic annotation data recorded at the archaeological site, bind action semantic tags to corresponding vertex indices, and fuse them to generate data in a unified format. Its calculation formula is as follows: In the formula, To standardize the format of data tensors, For spatiotemporal alignment fusion operators, For the registered point cloud data matrix, This is the multispectral reflectance characteristic matrix. For the semantic annotation tensor of micro-actions.

Images