Knowledge graph-based sports intangible cultural heritage analysis and tracing method

Abstract

The application provides a sports intangible cultural heritage analysis and tracing method based on a knowledge graph, which comprises the following steps: according to an expert rule base, constructing entity metadata and cultural feature codes containing representative inheritors, typical actions and regional schools, combining with a historical influence index to quantify the initial weight of the entity, forming a structured semantic anchor point, embedding the semantic anchor point in the knowledge graph through an improved graph neural network, and performing multi-hop path tracking by fusing path semantic features; adopting a main and auxiliary dual-channel embedding optimization mechanism to enhance the traceability and semantic consistency of the path, introducing a multi-objective loss function in model training to improve the embedding quality and optimize the semantic alignment degree of path prediction and expert rules; based on the credibility evaluation of multi-hop semantic tracing and reasoning results of user queries, finally generating a visualized graph with semantic role labeling and branch interaction functions, and the application improves the structured expression, path interpretability and reasoning transparency of the knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph modeling and graph neural network interpretable reasoning technology, and in particular to a method for analyzing and tracing the origins of sports intangible cultural heritage based on knowledge graphs. Background Technology

[0002] Knowledge graph embedding and intelligent reasoning technologies have been extensively researched in recent years in fields such as cultural heritage, knowledge management, and recommendation systems, particularly in the digital modeling and intelligent analysis of sports intangible cultural heritage knowledge. Mainstream technologies in the industry largely employ deep learning-driven graph embedding and graph neural network methods, transforming knowledge graph nodes into high-dimensional vectors and achieving semantic association reasoning through information propagation. Current common graph neural network embedding methods (such as GCN, GAT, and GraphSAGE) primarily focus on feature aggregation and global representation capabilities among nodes and their neighbors. These technologies largely pursue the optimization of overall embedding performance, driving the development of intelligent applications such as sports intangible cultural heritage knowledge mining, action recognition, and inheritance path prediction. However, current knowledge graph embedding methods generally suffer from the following drawbacks: (1) The embedding process lacks a semantic path traceability mechanism. Most models only record node features and aggregation results, but do not retain upstream and downstream relationships and semantic path identifiers; (2) Insufficient transparency of reasoning: the model cannot output the reasoning chain in a structured way, and only returns the entity association score, making the output results difficult to interpret; (3) Users are unable to explore the details of reasoning in multiple levels and interact with them, which seriously restricts the usability and credibility of the sports intangible cultural heritage knowledge service system; (4) Some solutions attempt to pre-generate paths through external large models or use reinforcement learning for path search. However, such technical routes rely on external knowledge systems, making it difficult to achieve native explanations for the core ontology of the domain. They may also introduce external biases or loss of context. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, this invention provides a method for analyzing and tracing the origins of sports intangible cultural heritage based on knowledge graphs.

[0004] The technical solution of this invention is implemented as follows: a knowledge graph-based method for analyzing and tracing the origins of intangible cultural heritage in sports, comprising: S1: For the core entities in the knowledge graph of sports intangible cultural heritage, semantic anchors with cultural semantic identifiers are defined based on the expert rule base. The core entities include representative inheritors, typical actions and regional schools. The semantic anchor configuration includes entity type identifier, cultural feature code and initial weight coefficient, wherein the initial weight coefficient is quantified and assigned according to the entity's historical influence index. S2: Input the heterogeneous knowledge graph containing semantic anchors into the improved graph neural network architecture. The architecture performs a relation path identifier generation operation in each layer of message passing, specifically including: performing tensor product operation on the source node embedding vector, relation type encoding and target node embedding vector to generate a unique path fingerprint encoding, and compressing it into a fixed-dimensional path identifier vector through a differentiable hash mapping function. S3: Construct a dual-channel embedding update mechanism. The main channel uses a graph attention network to optimize node representation, while the auxiliary channel performs semantic integrity verification of the path fingerprint. Specifically, this includes: calculating the cosine similarity of path identifiers between adjacent nodes; activating the semantic consistency compensation module when the similarity is lower than a preset threshold; and adjusting the relationship weight matrix through gradient backpropagation to maintain path traceability. S4: A multi-objective loss function is introduced during training. The loss function includes a conventional embedding loss term and a semantic alignment constraint term. The semantic alignment constraint term calculates the topological structure difference between the predicted path and the standard path based on a preset expert rule path. The difference calculation uses the edit distance algorithm to quantify the minimum number of insertion / deletion operations in the node sequence. S5: In response to a user query request, perform a multi-hop path decoding operation, which specifically includes: starting from the embedding vector of the query entity, tracing the semantic anchor point association path in reverse through the maximum weight path search algorithm, and generating a structured source path sequence containing node type labels, relationship strength coefficients and path confidence. S6: An interactive and visual semantic graph is generated based on the source path sequence. The force-directed graph layout algorithm is used to embed the node vectors and map them to a two-dimensional coordinate space. The path propagation strength is visualized through dynamic edge weight rendering technology. The visualization interface includes a collapsible path branch control and a semantic role annotation layer, which supports users to expand the reasoning details layer by layer.

[0005] The knowledge graph-based method for analyzing and tracing the origins of intangible cultural heritage in sports provided by this invention has the following beneficial effects: (1) By introducing a semantic anchor mechanism and a dual-channel embedding update architecture, this invention significantly improves the stability of the model’s representation of key cultural entities and the traceability of the reasoning process. By defining high-weight semantic anchors for core entities such as representative inheritors, typical actions, and regional schools of thought, and using them as stable reference systems in the embedding space, the vector offset under irrelevant noise interference is effectively suppressed. At the same time, by using path identifiers to accumulate and generate multi-hop path fingerprints, while performing conventional embedding optimization in the main channel, the auxiliary channel is dedicated to maintaining the structural integrity of the path semantics, so that even after deep propagation, a clear reasoning chain can still be reversed and restored, thereby realizing the dynamic preservation and explicit expression of cultural logic links. (2) This invention innovatively introduces a semantic consistency constraint loss function, which forces the model output to remain aligned with the preset expert rule path during the training phase. For example, it ensures that relationships with clear historical and geographical connections, such as "a certain style of boxing → originating from → a certain region", prioritize the activation of anchor point clusters in the corresponding dimension, thereby overcoming the knowledge conflict and logical paradox problems that are common in traditional end-to-end learning. This mechanism not only improves the generalization ability of the model in cold start scenarios, but also ensures that the recommendation or reasoning results conform to the basic cognitive framework of cultural heritage protection. During the reasoning phase, a visual semantic tracing graph is generated simultaneously, showing the main path branches, semantic roles of intermediate nodes and their confidence distribution between the query entity and the target entity. Users can flexibly view the reasoning details at different levels through the interactive interface, realizing the transparent presentation of the decision-making basis. Compared with traditional solutions that rely on external large model explanations or offline index back lookups, this method natively supports fine-grained tracing, providing dual guarantees of cultural rationality and technical credibility without additional computational overhead. (3) This invention constructs a closed-loop adaptive knowledge reasoning system that combines high efficiency, scalability, and domain adaptability, making it particularly suitable for application scenarios such as sports intangible cultural heritage, which are semantically complex, data-sparse, and highly sensitive to cultural value. Through lightweight path fingerprint encoding and a dual-channel collaborative update mechanism, it achieves full-process tracking of multi-hop reasoning paths without significantly increasing computational burden, avoiding the high resource consumption and training instability risks associated with reinforcement learning path search and other methods. The entire architecture does not require frequent manual parameter tuning or the introduction of large-scale pre-trained models, and has good deployment flexibility and cross-platform migration capabilities. More importantly, this method fundamentally changes the limitations of the traditional "black box operation" of embedded models, establishing a new paradigm that integrates cultural logic constraints and data-driven learning. It retains the powerful expressiveness of deep models while endowing them with the ability to be understood, auditable, and interventionist, providing reliable technical support for the accurate matching, diversified recommendation, and intelligent guidance of resources inside and outside schools in educational and sports scenarios such as campus football, and promoting the digital inheritance of traditional culture to evolve towards a higher level of cognitive intelligence. Attached Figure Description

[0006] Figure 1 This is a flowchart of the knowledge graph-based method for analyzing and tracing the origins of intangible cultural heritage in sports, as presented in this invention. Figure 2 This is a sub-flowchart of the knowledge graph-based method for analyzing and tracing the origins of intangible cultural heritage in sports according to the present invention. Figure 3 This is another sub-flowchart of the knowledge graph-based method for analyzing and tracing the origins of intangible cultural heritage in sports according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a knowledge graph-based method for analyzing and tracing the origins of intangible cultural heritage in sports, specifically including: S1: For the core entities in the knowledge graph of sports intangible cultural heritage, semantic anchors with cultural semantic identifiers are defined based on the expert rule base. The core entities include representative inheritors, typical actions and regional schools. The semantic anchor configuration includes entity type identifier, cultural feature code and initial weight coefficient, wherein the initial weight coefficient is quantified and assigned according to the entity's historical influence index. S2: Input the heterogeneous knowledge graph containing semantic anchors into the improved graph neural network architecture. The architecture performs a relation path identifier generation operation in each layer of message passing, specifically including: performing tensor product operation on the source node embedding vector, relation type encoding and target node embedding vector to generate a unique path fingerprint encoding, and compressing it into a fixed-dimensional path identifier vector through a differentiable hash mapping function. S3: Construct a dual-channel embedding update mechanism. The main channel uses a graph attention network to optimize node representation, while the auxiliary channel performs semantic integrity verification of the path fingerprint. Specifically, this includes: calculating the cosine similarity of path identifiers between adjacent nodes; activating the semantic consistency compensation module when the similarity is lower than a preset threshold; and adjusting the relationship weight matrix through gradient backpropagation to maintain path traceability. S4: A multi-objective loss function is introduced during training. The loss function includes a conventional embedding loss term and a semantic alignment constraint term. The semantic alignment constraint term calculates the topological structure difference between the predicted path and the standard path based on a preset expert rule path. The difference calculation uses the edit distance algorithm to quantify the minimum number of insertion / deletion operations in the node sequence. S5: In response to a user query request, perform a multi-hop path decoding operation, which specifically includes: starting from the embedding vector of the query entity, tracing the semantic anchor point association path in reverse through the maximum weight path search algorithm, and generating a structured source path sequence containing node type labels, relationship strength coefficients and path confidence. S6: An interactive and visual semantic graph is generated based on the source path sequence. The force-directed graph layout algorithm is used to embed the node vectors and map them to a two-dimensional coordinate space. The path propagation strength is visualized through dynamic edge weight rendering technology. The visualization interface includes a collapsible path branch control and a semantic role annotation layer, which supports users to expand the reasoning details layer by layer.

[0010] Step S1: For the core entities in the sports intangible cultural heritage knowledge graph, define semantic anchors with cultural semantic identifiers based on the expert rule base. The core entities include representative inheritors, typical movements, and regional styles. The semantic anchor configuration includes entity type identifiers, cultural feature codes, and initial weight coefficients, where the initial weight coefficients are quantified and assigned based on the entity's historical influence index. Specifically, this includes: S1.1: Based on the expert rule base in the field of sports intangible cultural heritage, extract metadata of three core entities: representative inheritors, typical actions, and regional styles. The metadata includes entity name, classification label, historical evolution description, and cultural regional distribution information to construct the initial basic dataset for semantic anchors. Based on the expert rule base dataset in the field of sports intangible cultural heritage, a rule matching algorithm (parameters: entity category label, semantic relationship pattern set) is used to achieve the preliminary screening and positioning of three core entities: representative inheritors, typical actions and regional schools. In combination with the classification labels, entity type identifiers are configured for each core entity. Furthermore, by using a domain semantic parsing method (parameters: word segmentation dictionary, cultural terminology mapping table), the text names of the selected core entities are accurately parsed, and standardized entity name encoding results are obtained. Furthermore, through a metadata extraction algorithm (parameters: entity name encoding, historical evolution corpus, and regional distribution database), the historical evolution description information of each core entity is automatically extracted, and historical evolution time series data is generated. Furthermore, through geographic information association algorithms (parameters: coordinate set of regional distribution database, administrative division coding system), the precise location of the cultural regional distribution information of core entities is achieved, and geographic metadata containing latitude and longitude coordinates and regional labels is generated; By using metadata fusion processing (parameters: entity name encoding, category label, historical evolution time series, cultural and regional coordinate set), the results of the previous step are transformed into a semantic anchor point initialization base dataset, realizing complete data input for subsequent cultural feature encoding processing; For example, in the construction of the knowledge graph of sports intangible cultural heritage, the domain expert rule base contains 500 rule entries for each of the three types of entities. The rule entries for representative inheritors are set according to the intangible cultural heritage list, the rule entries for typical actions are set according to the action coding table, and the rule entries for regional styles are set according to the regional cultural distribution table. In the input intangible cultural heritage text dataset, the total number of representative inheritor entity entries is 300, the total number of typical action entity entries is 450, and the total number of regional style entity entries is 200.

[0011] Metadata for three types of core entities is extracted based on an expert rule base. This metadata includes entity name (e.g., "Zhang San", "Tai Chi Cloud Hands", "Henan Chen Style Tai Chi School"), category tags (e.g., "Inheritance Subject", "Technique Movement", "Cultural School"), historical evolution description (e.g., "Zhang San has inherited intangible cultural heritage techniques for 20 years", "Cloud Hands movement originated in the Qing Dynasty"), and cultural geographical distribution information (e.g., "Jiaozuo, Henan", "Handan, Hebei"). In conjunction with the category tags, entity type identifiers are configured for each core entity (e.g., "TCHR" for "Inheritance Subject", "ACT" for "Technique Movement", and "REG" for "Cultural School"). A rule matching algorithm ensures that the category tag and identifier of each entity are bound one-to-one, avoiding category confusion.

[0012] The rule-based matching algorithm requires a matching weight of 0.8 or higher to proceed to the next step of parsing. During the parsing process, a cultural terminology mapping table was used, achieving a replacement rate of 95% and generating 950 standardized entity name codes. The metadata extraction algorithm achieved a 92% matching rate in the historical evolution corpus, forming a historical evolution time series for each entity, with a time span ranging from [missing information]. Over the course of years, the geographic information association algorithm, employing a coordinate system with an administrative division coding precision of 5 bits, successfully generated latitude and longitude coordinates and regional labels for each entity, achieving a coordinate accuracy of [missing information]. Ultimately, the metadata fusion processing generated a semantic anchor initialization base dataset containing entity name encoding, classification labels, historical evolution sequences, and geographic coordinate labels, providing a rigorous, complete, and consistent data foundation for the subsequent cultural feature encoding in S1.2; S1.2: Cultural feature encoding processing is performed on the extracted metadata of the three types of core entities. A multi-dimensional feature vector representation method is adopted, in which the dimensions include time dimension (historical evolution period), spatial dimension (geographic origin coordinates), skill dimension (complexity of action structure) and inheritance dimension (number of disciples and breadth of dissemination) to generate feature encoding vectors with cultural semantic identifiers. S1.3: Calculate the initial weight coefficient based on the historical influence index of the entity. The historical influence index is calculated by weighting dimensions such as the frequency of document citation, the number of generations of transmission, the geographical scope of dissemination, and the level of the national intangible cultural heritage representative project list, so as to quantify the initial influence of the semantic anchor in the embedding space. S1.4: The entity type identifier, cultural feature encoding vector and initial weight coefficient are encapsulated in a structured manner to generate a semantic anchor configuration tuple. The configuration tuple is in the format of (entity ID, type identifier, feature encoding vector, initial weight coefficient), which is used for node initialization and path identifier generation operations in the subsequent graph neural network. S1.5: Inject the generated semantic anchor configuration tuple into the node initialization module of the heterogeneous knowledge graph to initialize the embedding vector of the corresponding entity node in the graph, ensuring that it has interpretable semantic anchoring capability in the early stage of graph neural network training, laying the foundation for subsequent path tracking and reasoning transparency.

[0013] Step S2: Input the heterogeneous knowledge graph containing semantic anchors into the improved graph neural network architecture. This architecture performs a relation path identifier generation operation during each layer's message passing process. Specifically, this includes: performing a tensor product operation on the source node embedding vector, relation type encoding, and target node embedding vector to generate a unique path fingerprint encoding, and compressing it into a fixed-dimensional path identifier vector using a differentiable hash mapping function. Specifically, this includes: S2.1: Based on the node embedding vectors and relation type encoding in the knowledge graph of sports intangible cultural heritage, a message passing architecture of graph neural network is constructed. The node embedding vectors include the initial semantic anchor vectors of representative inheritors, typical actions and regional schools. The relation type encoding adopts one-hot encoding or embedding matrix representation to form the input representation space of graph neural network. Based on the node embedding vectors and relation type encoding in the knowledge graph of sports intangible cultural heritage, a heterogeneous graph input representation method (parameters: node set E, relation set R) is used to realize the message passing structure initialization of an improved graph neural network. Furthermore, the node vector space is constructed through the node embedding initialization algorithm (parameters: semantic anchor configuration tuple, embedding dimension d), and an initial semantic anchor embedding matrix containing representative inheritors, typical actions and regional schools of thought is obtained. Furthermore, a relation type encoding mapping algorithm (parameters: relation set R, one-hot encoding / embedding matrix, relation dimension k) is used to realize the numerical representation of relation types and generate a relation encoding matrix, which is used together with node embedding to form the model input representation space; Furthermore, an input representation fusion method (parameters: node embedding matrix, relation encoding matrix, fusion operation concat) is adopted to realize the joint input structure of node semantic features and relation type features, and generate the composite input tensor of the first layer message passing; By configuring the message passing architecture algorithm (parameters: number of multi-layer message passing units L, activation function type, weight initialization strategy), the composite input tensor of the previous step is transformed into an initial message representation that can be used to generate subsequent path identifiers, thereby realizing the stable reference role of semantic anchors in the network. For example, in a knowledge graph scenario of intangible cultural heritage in sports, the node set E includes 500 entity nodes, of which 200 are representative inheritor type nodes, 150 are typical action type nodes, and 150 are regional school type nodes. The node embedding dimension d is set to 128 and initialized using pre-trained semantic anchor vectors. The relation set R contains 300 edges, with relation types divided into three categories: "inherited from," "originating from," and "belonging to," using one-hot encoding. The relation dimension k is set to 16. The semantic anchor configuration tuple is (entity ID, type identifier, feature encoding vector, initial weight coefficient), where the initial weight coefficient ranges from 0.1 to 0.9 and is calculated based on the historical influence index. In the input fusion operation, the 128-dimensional node embeddings and 16-dimensional relation encodings are directly concatenated to form a 144-dimensional composite input vector, which is then organized into the first layer message passing module of the batch processing tensor graph neural network. The number of message passing units L is set to 3, the activation function is ReLU, and the weight initialization strategy is Xavier initialization. Under this configuration, the network can retain the cultural semantic identifiers of nodes during the training iteration process and provide highly recognizable input features in the subsequent path fingerprint generation stage, thereby achieving traceability and embedding transparency of path semantics. S2.2: Perform tensor product operation on the source node embedding vector, relation type encoding and target node embedding vector to generate triple tensor representation. The tensor product operation adopts the combination of element-wise multiplication and matrix expansion to capture the high-order semantic interaction features between triples and output a unique path fingerprint code to identify the information propagation trajectory of a specific path. Tensor product operation (parameter: element-wise multiplication + matrix expansion mode) is used to capture high-order semantic interaction features of triples for source node embedding vectors, relation type encoding, and target node embedding vectors. Furthermore, through element-wise multiplication (parameters: source vector dimension m, relation encoding dimension r, target vector dimension m), direct semantic relevance matching between different input vector components is achieved, and preliminary triplet interaction matrix results are generated; Furthermore, by using the matrix expansion and combination method (parameter: row-major expansion mode), the initial interaction matrix is ​​reconstructed into a one-dimensional high-dimensional feature vector, realizing the complete interaction information encoding of the input triplet and laying the foundation for the uniqueness of the path fingerprint encoding; Furthermore, by normalization transformation (parameter: L2 norm constraint), the scale of the triplet interaction vector space is unified, thereby avoiding encoding conflicts caused by differences in numerical scale in the subsequent hash mapping stage; Furthermore, by weighting the feature positions (parameter: the position weight matrix is ​​generated based on the semantic anchor weight coefficient), semantic weights are added to each component of the interaction vector to achieve a controllable distribution of the influence of different semantic elements in the path fingerprint encoding; By using the tensor product and matrix expansion processing methods described above, the node and relation semantic inputs in the previous step are transformed into unique path fingerprint codes, enabling precise identification of the propagation trajectory of specific path information. For example, in the knowledge graph of sports intangible cultural heritage, for the triple "Inheritor A - Inherited Action B - Region C", the source node embedding vector dimension is set to 64, the relation type encoding dimension is set to 16, and the target node embedding vector dimension is set to 64. An element-wise multiplication is used to generate a 64×16×64 three-dimensional interaction tensor. A matrix expansion and combination method is executed to expand this tensor into a one-dimensional feature vector of length 65536 in row-major order. L2 norm normalization is used to unify the numerical scale. After normalization, a position weight matrix is ​​constructed based on the historical influence coefficients of the inheritor, action, and region in the expert rule base. Weighted multiplication is performed on each component of the vector to generate a weighted unique path fingerprint code. In this scenario, the path fingerprint code exhibits significant semantic discrimination ability in the subsequent hash mapping step. It can stably restore the semantic structure and interaction strength of the triple during backtracking reasoning, improving interpretability. S2.3: Based on path fingerprint encoding, perform differentiable hash mapping operation, and use a hash function composed of a multilayer perceptron and a normalization function to compress the path fingerprint encoding into a fixed-dimensional path identifier vector. The parameters of the hash function are jointly optimized through gradient backpropagation to ensure that the path identifiers maintain semantic consistency and distinguishability in the embedding space. S2.4: The path identifier vector and the node embedding vector are concatenated and fused to form an enhanced node representation. The fusion operation adopts a combination of linear transformation and nonlinear activation function to improve the graph neural network's ability to perceive path information and encode the path information into the node embedding update process to ensure that the subsequent inference path is traceable. S2.5: Based on the enhanced node representation, the message aggregation and update operations of the graph neural network are performed. The graph attention mechanism is used to calculate the path contribution weight of the adjacent nodes and to propagate and update the path identifier to form a cumulative representation of the multi-hop path fingerprint, thereby constructing an interpretable reasoning structure with semantic path information in the embedding space.

[0014] like Figure 2 As shown, step S3 involves constructing a dual-channel embedding update mechanism. The main channel uses a graph attention network for node representation optimization, while the auxiliary channel performs semantic integrity verification of the path fingerprint. Specifically, this includes calculating the cosine similarity of path identifiers between adjacent nodes. When the similarity is lower than a preset threshold, the semantic consistency compensation module is activated, and the relationship weight matrix is ​​adjusted through gradient backpropagation to maintain path traceability. Specifically, this includes: S3.1: Construct a main channel embedding update model based on the graph attention network (GAT) architecture. The graph attention network takes the node embedding vector containing semantic anchors as input and performs neighborhood information aggregation under the multi-head attention mechanism to generate optimized node representation vectors. The node embedding vector containing semantic anchors is used as input data. The node embedding vector is generated by the aforementioned enhanced node representation and contains the semantic feature components of the entity and the path fingerprint encoding. The graph attention network (GAT) main channel construction method is adopted (parameters: number of multi-head attention h=8, hidden layer dimension d=128) to achieve high-resolution perception and weighted aggregation of neighborhood information; Furthermore, based on the input embedding vector of each node and the embedding vector of its neighboring nodes, a neighborhood adaptive weight calculation method is executed (parameter: LeakyReLU negative slope α=0.2) to achieve weighted estimation of the information contribution of different neighboring nodes through learnable attention coefficients, and obtain the neighborhood weighted set data matrix; Furthermore, the neighborhood weighted set data matrix is ​​input into the linear transformation unit (parameters: the weight matrix is ​​initialized using Xavier initialization, and the initial value of the bias vector is zero) to realize the linear projection of node features into the high-dimensional space and generate the transformation output sub-vectors corresponding to each attention head; Furthermore, a splicing and fusion operation is performed on the output sub-vectors of the multiple attention heads (parameter: splicing dimension equals h×d) to achieve the collaborative integration of multi-view semantic features and obtain the fused high-dimensional node representation matrix; By using a nonlinear activation function (parameter: activation function type is ELU), the fused high-dimensional node representation matrix is ​​transformed into an optimized node representation vector, thereby improving the embedding expression capability and maintaining the sensitivity of path semantic information in the embedding space. For example, in the implementation scenario of the sports intangible cultural heritage knowledge graph, the initial semantic anchor vector dimension of each representative inheritor is set to 64, which is expanded to 96 dimensions after combining with the aforementioned enhanced embedding representation, including inheritor category encoding and path fingerprint encoding. The graph attention network input to the main channel is set with a multi-head attention number h of 4 and a hidden unit number d of each head of 64, producing a 256-dimensional intermediate output. In the neighborhood weight calculation, the average number of neighborhood nodes is 6, and the attention coefficient is mapped using LeakyReLU with a negative slope α of 0.15, resulting in weights ranging from 0.05 to 0.45. The linear transformation weight matrix is ​​initialized using Xavier with a matrix scale of 64×64, and the dimension of the concatenated and fused feature vector is 256. The node representation obtained after ELU activation significantly improves the accuracy of semantic classification tasks, and in the path traceability evaluation, the traceability integrity index is improved from 0.62 in the original structure to 0.85, verifying that the main channel embedding optimization model significantly enhances the semantic sensitivity and interpretability of the path while maintaining the embedding expressive power. S3.2: Perform path fingerprint association mapping on the attention weight of each relation edge in the main channel, and use the path identifier vector and the attention weight vector to multiply element by element to generate a weighted neighborhood embedding representation with path semantic identifier, so as to enhance the path traceability in the node update process. For each relation edge attention weight vector output by the main channel, a path fingerprint association mapping algorithm (parameters: path identifier vector, attention weight vector) is used to enhance the semantic orientation of relation weights. Furthermore, through element-wise multiplication (parameter: path identifier vector) Attention weight vector This enables the encoding fusion of path fingerprint information and neighborhood weights, resulting in a weighted neighborhood embedding intermediate matrix. , used to carry path traceability attributes; Furthermore, a linear transformation matrix is ​​employed. (Parameters are matched according to the node embedding dimension and fusion dimension matching rules) for the weighted neighborhood embedding matrix Spatial projection is performed to achieve linear reconstruction of path features in the embedding space, and neighborhood embedding vectors containing path semantic identifiers are generated. ; Furthermore, through a nonlinear activation function (Choose ReLU or LeakyReLU based on gradient preservation requirements), for the neighborhood embedding vector Perform feature activation operations to achieve nonlinear extension of the embedding representation, enabling path semantic features to be learned by higher-order representations during the update process; By using the graph attention aggregation function, the activated neighborhood embedding vector is fused with the node's own representation vector to obtain a weighted node embedding update result that carries path identification features, thereby realizing explicit traceable encoding of the reasoning path in the embedding space. For example, in the main channel processing of the knowledge graph of sports intangible cultural heritage, relationship edges are established between representative inheritor nodes and their adjacent typical action nodes, and attention weight vectors are used. Dimension set to , corresponding path identifier vector Dimension set to Element-by-element product calculation result For 64-dimensional embedding, after linear transformation matrix (dimension) The projection yields a 32-dimensional neighborhood vector. The ReLU activation function is used to perform nonlinear expansion, outputting a 32-dimensional weighted neighborhood embedding vector. Finally, in the graph attention aggregation function, this neighborhood vector is fused with the node's own embedding with a weight coefficient of 0.7 to form a node update result containing path semantic identifiers. The application effect shows that this node can significantly improve the responsiveness of path semantic-related feature components in the embedding space, thereby effectively improving the interpretability of path tracing in the subsequent inference and decoding stages. S3.3: Construct a path fingerprint semantic integrity verification model in the auxiliary channel. The model takes the path identifier vector between adjacent nodes as input and calculates their cosine similarity in the embedding space to evaluate the semantic consistency of the path fingerprint during the propagation process. S3.4: When the cosine similarity of the path identifier is lower than the preset threshold, the semantic consistency compensation module is activated. The compensation gradient signal is generated based on the path fingerprint difference information. The relation weight matrix is ​​adjusted through the backpropagation mechanism to correct the semantic offset of the path fingerprint in the embedding propagation. Using the cosine similarity of path identifier vectors between adjacent nodes as input parameters, a similarity threshold comparison algorithm (parameter: preset threshold τ=0.85) is adopted to realize the function of determining the semantic consistency of paths. Furthermore, through the difference vector generation method (parameter: difference vector) This quantizes the semantic offset and obtains the difference vector data that can be used for compensation, using the following formula: in and These are the path identifier vectors for the source node and the target node, respectively; Furthermore, a gradient generation algorithm (parameters: learning rate η = 0.001, loss function with cosine similarity difference as input) is used to calculate the compensation gradient signal and generate a gradient matrix for subsequent weight adjustment. The cosine similarity calculation formula is as follows: in, and These are two path identifier vectors. For vector dot product, It is the product of the magnitudes of two vectors; Furthermore, through the backpropagation mechanism (parameters: optimizer type Adam, momentum parameters β1=0.9, β2=0.999), the linkage adjustment of the relation weight matrix is ​​realized, and the compensation gradient signal is reflected in the embedded updated weight coefficients; By using a semantic consistency compensation algorithm, the difference vector from the previous step is transformed into locally adjusted data for relation weights, thereby achieving the semantic offset correction effect of path fingerprint during the embedding and propagation process. For example, in the semantic embedding space of the sports intangible cultural heritage knowledge graph, the cosine similarity of the path identifier vectors of a representative boxing style node and its associated regional node is 0.76, which is lower than the preset threshold of 0.85, triggering the compensation module. (Source node path identifier vector) With the target node vector The difference vector Δp ​​is During the gradient generation phase, a learning rate of 0.001 is used, and the similarity bias (0.85) is considered. 0.76) is used as the input to the loss function to generate the gradient matrix. During backpropagation, the Adam optimizer is selected to update the corresponding rows and columns of the relation weight matrix, so that the semantic consistency score of the path edge in the next iteration is improved to above 0.85, thereby achieving semantic alignment correction of the inference path; S3.5: After the dual-channel embedding update is completed, output the updated node embedding vector and the corresponding path fingerprint code. The node embedding vector is used for subsequent inference tasks, and the path fingerprint code is used for semantic interpretability analysis of the source path.

[0015] like Figure 3 As shown, step S4 involves introducing a multi-objective loss function during training. This loss function includes a conventional embedding loss term and a semantic alignment constraint term. The semantic alignment constraint term calculates the topological structure difference between the predicted path and the standard path based on a preset expert rule path. The difference calculation uses an edit distance algorithm to quantify the minimum number of insertion / deletion operations in the node sequence. Specifically, this includes: S4.1: Based on the preset expert rule path for sports intangible cultural heritage, construct a standard path set. The standard path set consists of semantic anchor point association paths between representative inheritors, typical actions and regional schools, which serve as the training basis for semantic alignment constraint terms. Based on the pre-set semantic anchor relationships in the expert rule base of sports intangible cultural heritage, a rule parsing algorithm (parameters: entity type identifier, cultural feature encoding, historical evolution description, regional distribution information) is used to extract the direct or indirect semantic relationship paths between representative inheritors, typical actions and regional schools from the expert rule entries. Furthermore, by using a path pattern recognition algorithm (parameters: set of relation types, set of node types, path length threshold), the extracted semantically related paths are classified in a pattern-based manner, path structures that conform to expert rule patterns are selected, and structured node sequences and relation sequences are obtained. Furthermore, by using a path regularization generation method (parameters: node sequence, relation sequence, path semantic constraints), the structure of the selected path is regularized, the order of nodes and relations in the path is adjusted to conform to the standard path specification, and a unique index code of the path is generated as the benchmark identifier for subsequent semantic alignment constraints. Furthermore, a path metadata encapsulation method (parameters: path index encoding, node type label, relation type encoding, cultural feature vector) is adopted to encapsulate the multidimensional metadata of the path and obtain a standard path tuple set containing semantic anchor identifiers. Through the above path pattern recognition and metadata encapsulation processing methods, the semantic anchor associated path results are transformed into standard path set data driven by expert rules, so as to realize the standard reference benchmark for semantic alignment constraint terms during training. For example, in a knowledge graph of intangible cultural heritage in sports, the expert rule base contains 657 entity relationship rules. The identifier for the representative inheritor entity type is "TCHR", the identifier for the typical action type is "ACT", and the identifier for the regional style type is "REG". The path pattern recognition algorithm sets a path length threshold of 4, and the relationship type set covers 9 types, including "originating from", "inherited from", and "contains". The rule parsing algorithm outputs approximately 1342 candidate paths, which, after pattern-based classification, yield 842 path structures that conform to the expert rule patterns. During path regularization, the node order of each path is adjusted so that the path begins at the inheritor node and ends at the regional style node. The path index encoding uses the SHA-256 hash method to generate a 64-character hexadecimal string, ensuring uniqueness and verifiability. The encapsulated standard path tuple is stored in a four-tuple format: (path ID, node label sequence, relationship encoding sequence, cultural feature vector). The cultural feature vector has a dimension of 12 and contains numerical descriptions of four feature domains: time, space, skill, and inheritance. This set of standard paths serves as a benchmark in the semantic alignment loss of subsequent training, providing the model with a stable and reliable reference for expert knowledge and effectively improving the semantic consistency of inference paths in the embedding space. S4.2: Extract the set of predicted paths generated by the graph neural network in each round of forward propagation. The set of predicted paths is generated by the maximum weight path search algorithm based on the node embedding vector and the relation weight matrix, and is used as the comparison object for the semantic alignment constraint. The input conditions include the set of node embedding vectors output by the improved graph neural network in each round of forward propagation, and the relation weight matrix corresponding to the graph structure. Both are optimized by a dual-channel embedding update mechanism and maintain the traceability of semantic anchor information. The maximum weight path search algorithm (parameters: node embedding vector set, relation weight matrix, search depth d) is adopted to perform a weighted traversal of all possible path combinations in the embedding space and select the path with the largest cumulative weight value to ensure the optimality of the output path in terms of node relation contribution. Furthermore, based on the searched candidate path set, the cumulative weight of the candidate paths is accurately quantified through the path cumulative weight calculation module (parameters: node sequence and edge weight of each path), forming a path index table arranged in descending order of weight, so as to facilitate efficient retrieval in the subsequent difference calculation process; Furthermore, a path fingerprint fusion algorithm (parameters: node embedding vector, path identifier vector) is adopted to achieve semantic fingerprint encoding fusion of candidate paths, so as to ensure that the predicted path set carries both structural information and path semantic information, and maintains the uniqueness of the identifiers in different batches of training; Furthermore, by using the edge filtering rule module (parameters: relation type encoding, weight threshold τ), edges below the threshold in the candidate path are removed, reducing the interference of noisy paths on the semantic alignment constraint and improving the purity of the predicted path set. By using path set extraction and encoding processing, the node embedding vector and relation weight matrix from the previous step are transformed into structured predicted path set data, which becomes the comparison object required for subsequent calculation of the topological structure difference between the path set and the standard path set. For example, in the sports intangible cultural heritage knowledge graph, the node embedding vector dimension is set to 128, the relation weight matrix dimension is 128×128, and the search depth d is set to 4 hops. When using the maximum weight path search algorithm, the cumulative weight of the path is defined as the sum of the weights of each edge, where the edge weight is calculated by the inner product of the relation weight matrix and the node embedding vector. After calculating the cumulative weight for all paths, the top 10 paths with the highest weight values ​​are selected into the candidate set. The candidate paths then undergo path fingerprint fusion processing, generating a unique 64-dimensional path identifier vector for each path's node sequence, which is then concatenated with the node embedding vector to form a 192-dimensional path semantic representation. Finally, edge filtering is performed by setting a weight threshold τ=0.15 to remove edges with cumulative weights below the threshold, reducing the total number of predicted paths from the initial 10 to 7. Each path has a clear semantic fingerprint encoding and structural information for subsequent standard path comparison and difference calculation. S4.3: Calculate the topological structure difference between the path sequences in the predicted path set and the standard path set using the edit distance algorithm. Specifically, this includes solving for the minimum cost of performing node insertion, deletion, or replacement operations on each pair of path sequences to obtain the path structure difference index. S4.4: Construct a semantic alignment constraint term based on the path structure difference index. The semantic alignment constraint term is a differentiable loss term, and its mathematical expression is a weighted function of the path structure difference, so as to reflect the degree of semantic deviation between the model's predicted path and the expert path. S4.5: The semantic alignment constraint term and the conventional embedding loss term are weighted and fused to generate a multi-objective loss function. The multi-objective loss function drives the adjustment of model parameters during the backpropagation process of the graph neural network to simultaneously optimize the embedding expressive power and path semantic alignment. Based on the semantic alignment constraint term and the conventional embedding loss term constructed in the previous sub-step, a weighted fusion method (parameters: weight coefficients α and β) is adopted to realize the construction and optimization driving function of the multi-objective loss function. Furthermore, by using a weight constraint algorithm (parameters: α+β=1, α and β range from [0,1]), the balance of each loss term during the training process is controlled, and the weighted loss value after fusion is obtained. Furthermore, gradient pre-normalization is employed (parameter: loss gradient range normalized to [-1,1]) to achieve stability control of the gradient signal during backpropagation and generate a unified gradient matrix that can adapt to multi-channel embedding updates. Furthermore, by using a dynamic weight update strategy (parameters: based on training rounds, loss convergence rate, and path semantic deviation index), the adaptive adjustment of α and β is achieved, and the optimal weight configuration for different training stages is obtained. Furthermore, by combining the node embedding vectors and path fingerprint encodings from the previous iteration, a joint gradient optimization method is used to perform backpropagation operations, thereby achieving global optimization of the graph neural network parameters (node ​​embedding weight matrix, relation weight matrix, path identifier generation parameters) and generating a converged network parameter set. By using multi-objective loss-driven optimization, the fusion result of the loss terms in the previous step is transformed into a technical indicator that simultaneously improves the expressive power of the embedding space and the alignment of the path semantics, thereby achieving the expected technical effect of balancing interpretability and prediction accuracy in the reasoning process of the sports intangible cultural heritage knowledge graph. For example, in the training scenario of a knowledge graph for intangible cultural heritage in sports, let the conventional embedding loss term be the mean squared error, and the semantic alignment constraint term be a path structure difference weighting function. The initial weight coefficients α and β are configured to be 0.7 and 0.3, respectively, satisfying α + β = 1. The conventional embedding loss term is calculated as follows: : in, For the true embedding vector, To predict the embedding vector, Number of samples. Semantic alignment constraint. Set as: in, To edit the distance value, For path importance weights, Let be the size of the path set. The multi-objective loss function after fusion is: During training, a cosine similarity threshold of 0.85 was set. If the path similarity was lower than this value, the β value was increased by 0.05 and the α value was decreased by 0.05 based on the degree of deviation, to strengthen the optimization effect of semantic alignment constraints. After 50 iterations, the embedding loss value decreased significantly, and the path structure difference index was greatly reduced, verifying the effectiveness of the method in improving both embedding expressive power and semantic path alignment. The final output parameter set showed a significant improvement in prediction accuracy on the test set, and the generated source path could completely map the expert rule association structure.

[0016] Step S5: In response to a user query request, perform a multi-hop path decoding operation, specifically including: starting from the embedding vector of the query entity, tracing the semantic anchor association path backward using the maximum weight path search algorithm, and generating a structured source path sequence containing node type labels, relationship strength coefficients, and path confidence. Specifically, this includes: S5.1: Based on the query entity embedding vector input by the user, the maximum weight path search algorithm is used to reverse track the semantic anchor point association path in the graph embedding space. The query entity embedding vector is output by the improved graph neural network and contains entity semantic information and path fingerprint encoding to obtain the set of multi-hop paths with the highest correlation with the semantic anchor point of the query entity. Based on the query entity embedding vector jointly output by the main and auxiliary channels of an improved graph neural network, a maximum weight path search algorithm (input parameters: query entity embedding vector, relation weight matrix, path fingerprint encoding) is employed to locate the initial set of nodes for semantic anchor-related paths in the embedding space. This query entity embedding vector contains cultural feature dimension information generated by semantic anchors and accumulated multi-hop path fingerprint encoding, providing a semantic reference system for subsequent reverse tracing. By loading a relation weight matrix (dimension: number of nodes × number of nodes, element values ​​representing relation semantic strength) and a path identifier similarity matrix (element values ​​calculated from the cosine similarity of node path fingerprints) into the path search algorithm, priority is given to edges with high semantic relevance during path traversal, and a preliminary candidate path set is generated. To ensure the efficiency and interpretability of the search process, a priority queue structure is used to sort all current candidate expansion paths by cumulative weight. Each expansion selects the path branch with the largest cumulative weight and whose path fingerprint continuity score is not lower than a preset threshold. Furthermore, through a reverse tracing strategy, the search starting point is traced back step by step from the node containing the embedded vector of the query entity along the path with the maximum edge weight to any semantic anchor node. At each iteration, the current edge weight, path fingerprint fragment, and node type identifier are recorded, forming a path reverse tracing log data structure. This log data structure is used to verify the semantic continuity of the path and the anchor node association during the path generation phase. The total weight index for each path obtained through reverse tracing is calculated using the path cumulative weight calculation formula: in, The relation weights for each edge on the path, The total weight of each path is accumulated. This process ensures that the final set of selected multi-hop paths remains optimal in terms of semantic anchor relevance and excludes semantically broken paths; By combining the above-mentioned path search and reverse tracing processing method, the query entity embedding vector of the previous step is transformed into a set of multi-hop paths with the highest semantic anchor degree with the entity, thereby achieving high transparency and traceability in the path decoding stage. For example, in a query scenario for a certain intangible cultural heritage martial arts style, the query entity embedding vector is 128-dimensional, including time dimension features (300-year history), spatial dimension features (longitude 120.15, latitude 30.28), skill dimension features (movement structure complexity 0.85), and inheritance dimension features (approximately 250 disciples). The path fingerprint encoding length is 64. After training, the elements of the relationship weight matrix range from 0.12 to 0.97, and the elements of the path identifier similarity matrix range from 0.25 to 0.99. When this query entity embedding vector is input into the maximum weight path search algorithm, the maximum path length is set to 5 hops, and the path fingerprint similarity threshold is set to 0.72. In the first stage, the algorithm generates 15 candidate paths. After cumulative weight sorting and reverse tracing verification, it selects 3 paths with the highest relevance to the semantic anchor point. The cumulative weight of one of these paths is calculated as follows: Output value After the subsequent S5.2 to S5.5 steps of structured mapping and confidence calculation, the path was successfully presented in the visualized semantic graph, which significantly improved the user's confidence in the reasoning results of the path. S5.2: Perform a node type label mapping operation on each path in the multi-hop path set. Based on the node type identifier and cultural feature encoding, map each node in the path to a node type label with a clear semantic role to generate a structured path node sequence. The node type labels include representative inheritors, typical actions, and regional schools of thought. The node data input conditions are based on a multi-hop path set. The node data includes a node type identifier and a cultural feature code. The node type identifier is a unique code that represents the category to which the node belongs. The cultural feature code is a multi-dimensional feature vector that covers information on time, space, skills and inheritance dimensions. The node type matching method (parameters: node type identifier set, mapping rule table) is used to match the type code of each node in the path with the predefined type label rules and output a preliminary type label set. Furthermore, through a multi-dimensional cultural feature parsing algorithm (parameters: cultural feature encoding vector, feature dimension weight coefficient), node features are mapped to the semantic role space, and the probability distribution of node affiliation in representative inheritors, typical actions, and regional school categories is obtained. Furthermore, by using the maximum attribution probability selection algorithm (input: node attribution probability distribution result), the semantic role corresponding to the highest probability value is selected as the final type label of the node, and node records with role labels are generated. Furthermore, by using a type label sequence generation algorithm (input: a set of node records with role labels), the nodes in the path are sequentially combined into a structured path node sequence, and path label data that meets the requirements of semantic traceability is obtained. By using type label mapping and cultural feature parsing, the multi-hop path set from the previous step is transformed into a structured path node sequence containing precise semantic role identifiers, thereby improving the semantic interpretability of path node information in source tracing analysis. For example, in the reasoning result set of the sports intangible cultural heritage knowledge graph, node A's type identifier is T1, and its cultural feature encoding is a four-dimensional vector [120, (35.5, 112.3), 0.85, 152], where the time dimension represents 120 years of historical evolution, the spatial dimension represents latitude 35.5° and longitude 112.3°, the skill dimension represents action structure complexity 0.85, and the inheritance dimension represents the number of disciples 152. During the type matching process, T1 is matched to the representative inheritor category through the mapping rule table. In the cultural feature parsing stage, the weights are set as follows: time dimension 0.3, spatial dimension 0.25, skill dimension 0.25, and inheritance dimension 0.2. The probability of belonging to each category is calculated through weighted normalization, where the probability of belonging to another category for the representative inheritor is... The calculated result is the largest among all categories. Therefore, the final type label for node A is determined to be the representative successor, and this label is recorded in the path node sequence. After executing the above process, the structured node sequence of the entire multi-hop path has accurate semantic roles, significantly improving the transparency and interpretability of the subsequent tracing graph; S5.3: Based on the path identifier vector and relation type encoding, calculate the relation strength coefficient between each pair of adjacent nodes in the path. The relation strength coefficient is obtained by multiplying the cosine similarity of the path identifier vector with the relation type weight matrix to quantify the semantic association strength of each relation edge in the path. S5.4: Calculate the overall path confidence of each path based on the path fingerprint encoding and semantic consistency constraint of each node in the path. The path confidence is jointly generated by the path fingerprint integrity score and the semantic alignment error inverse proportional function to evaluate the credibility of the path reasoning results. Based on the calculated node path fingerprint encoding and semantic consistency constraints in the multi-hop path set, a path fingerprint integrity detection method (parameters: path fingerprint vector sequence, inter-node propagation record matrix) is adopted to achieve a quantitative evaluation of the fingerprint information preservation degree of any path in the embedding space. Furthermore, the semantic consistency difference analysis method (parameters: type labels of each node in the path sequence, expert rule path set) is used to calculate the semantic alignment error of the path and obtain the deviation index of the path. Furthermore, an inverse proportional function weighted fusion method (parameters: fingerprint integrity score, semantic alignment error index) is employed to generate path confidence. The specific formula is as follows: in, For path confidence, The path fingerprint integrity score is given. This is a semantic alignment error metric. Furthermore, by using a normalization mapping method (parameters: original set of confidence values, setting confidence display range), the path confidence in the [0,1] range is standardized, and a path confidence vector adapted to the visualization module is generated.

[0017] By using the above-mentioned fusion and normalization calculation method, the path fingerprint information and semantic consistency constraints are transformed into quantifiable confidence indicators, thereby achieving the evaluation effect of the credibility of the reasoning results. For example, in a sports non-heritage inheritance path containing 8 jumps, the path fingerprint integrity detection score... Semantic alignment error index Substitute this into the confidence level calculation formula: The original path confidence value is obtained. Then, by using interval normalization, the original values ​​are mapped to the [0,1] interval to obtain the standardized confidence scores. In different test paths, when the semantic alignment error metric increases to... When the calculated standardized confidence level is significantly reduced to This verifies the sensitivity of the calculation method to semantic bias and the accuracy of credibility assessment; S5.5: Integrate the structured path node sequence, relationship strength coefficient and path confidence to generate a structured source tracing path sequence. The structured source tracing path sequence is encoded using a graph structured data format, which supports subsequent path expansion and interactive display by the visualization semantic graph module. The input conditions include the structured path node sequence generated by the mapping in step S5.2, the relationship strength coefficient matrix calculated in step S5.3, and the set of overall path confidence values ​​obtained in step S5.4. A structured data fusion method (parameters: node sequence, relationship strength coefficient, path confidence) is adopted to realize the merging operation of multi-source path semantic elements under a unified coding framework; Furthermore, by using a graph structure encoding algorithm (parameters: adjacency matrix, node attribute encoding), the binding of node type labels with their corresponding edge strengths is achieved, and the branch structure data representation of multi-hop paths is obtained. Furthermore, an attribute extension mapping method (parameters: confidence value, node sequence index) is adopted to realize the hop-by-hop association mapping between path confidence and path nodes, and generate additional metadata fields that support semantic rendering; Furthermore, by using a formatted encoding method (parameter: GraphML or JSON-G format), the merged node and edge data structure can be parsed and an encoded file with path expansion flags can be generated. By encoding the graph-structured data format, the fusion result of the previous step is transformed into a standardized source path sequence that can support the visualization semantic graph module, so as to achieve the expected technical effect of subsequent path expansion and interactive display. For example, in a sports intangible cultural heritage knowledge graph scenario, the input node sequence contains 5 nodes, with the types being, in order: representative inheritor (ID01), typical action (ID15), regional school (ID23), typical action (ID37), and representative inheritor (ID44). The relation strength coefficient matrix is ​​a 5×5 sparse matrix, where the edge weights corresponding to non-zero values ​​range from 0.45 to 0.88, and the path confidence is calculated to be 0.74. Using a data fusion method, the node sequence is transformed into an adjacency matrix A, where matrix elements A... i,j This represents the relationship strength coefficient. The normalized path weight is calculated using the following formula: in, For the weight value of a certain side, This represents the weight of the k-th edge in the path. This normalized value is used to uniformly map the visualization width parameter of each edge. Using GraphML encoding, node attribute fields (ID, type label, cultural feature vector summary) and edge attribute fields (weight, normalized value, confidence score) are written to a file to achieve a standardized representation of the path. When the visualization module calls this encoded file, it can directly read the path node sequence, relationship strength coefficient, and confidence score, enabling a layer-by-layer unfolding display of the path and intuitively presenting the multi-hop semantic reasoning chain in the field of sports intangible cultural heritage in the interactive interface. Step S6: Generate an interactive visual semantic graph based on the source path sequence. A force-directed graph layout algorithm is used to map node embedding vectors to a two-dimensional coordinate space. Dynamic edge weight rendering technology is used to visualize the path propagation strength. The visualization interface includes a collapsible path branch control and a semantic role annotation layer, supporting users to unfold the reasoning details layer by layer. Specifically, it includes: S6.1: Based on the set of node embedding vectors in the source path sequence, the initial node coordinate allocation operation of the force-directed graph layout algorithm is performed to obtain the preliminary distribution position of the nodes in the two-dimensional visualization space. Specifically, the Euclidean distance between the graph embedding vectors is used as the initial input parameter of the repulsive force between nodes, and the edge weight is combined as the input of the attractive force between nodes. The stable two-dimensional coordinate position of each node is calculated through an iterative optimization algorithm. S6.2: Normalize the node coordinates output by the force-directed graph layout algorithm to fit the display area size of the visualization interface. Specifically, the min-max normalization method is used to perform a linear transformation on the two-dimensional coordinates, mapping the coordinate range to the preset visualization window coordinate system, ensuring that all nodes are evenly distributed in the visible area and avoiding node overlap or overflow of the display boundary; S6.3: Based on path propagation intensity information, a dynamic edge weight rendering algorithm is executed to generate differentiated display effects for paths of different intensities in the visualization map. Specifically, the path confidence value is mapped to the color transparency and width parameters of the connecting edges, and the HSV color space is used to linearly map the confidence interval to generate a visual gradient effect from low confidence (light color, thin line) to high confidence (dark color, thick line); Based on the structured tracing path sequence output by step S5 and its corresponding path confidence values, a dynamic edge weight rendering algorithm is adopted (parameters: path confidence, HSV color space range). Transparency range , border width range This allows the abstract path strength information to be transformed into a visual edge rendering attribute vector. Furthermore, through interval normalization (parameter: original confidence value) Interval boundaries The target interval is mapped to [0,1], achieving a linear mapping of the confidence scores to obtain the normalized intensity values. This serves as the basis for subsequent color and border width calculations; Furthermore, through the HSV color space mapping function (parameter: Hue fixed value) Corresponding to the theme color of intangible cultural heritage in sports, Saturation is fixed as Value is determined by the normalized intensity value. (Control), to achieve a gradient effect where the color depth changes with intensity, and to obtain a color vector. ; Furthermore, through the transparency control function (parameter: , Normalized intensity value To achieve edge transparency adjustment based on intensity, a transparency parameter is generated. The formula for calculating transparency is: Furthermore, through the linear width mapping function (parameter: , , To achieve the effect of edge width varying with intensity, the width parameter is obtained. The formula for calculating the width is: By configuring the rendering pipeline, , , The attributes are combined to form a set of edge rendering attributes and bound to the corresponding visual elements of the path edges to achieve different display effects of color depth, transparency and edge width for paths of different intensities; By using a dynamic edge weight rendering algorithm, the results of the previous step are transformed into visually sensitive parameters, enabling users to intuitively identify the path propagation intensity and achieve the expected technical effect of differentiated perception. For example, in the sports intangible cultural heritage knowledge graph, the confidence level of path A is 0.85, and the confidence level of path B is 0.40. Mapping the confidence interval [0,1] to the transparency range [0.3,1.0], the intensity of path A is normalized. =0.85, transparency =0.3+(1.0 0.3)×0.85=0.895; Normalized intensity of path B =0.40, transparency =0.3+(1.0 0.3) × 0.40 = 0.58. The width range is set to [1px, 5px], and the width of path A is... =1+(5 1) × 0.85 = 4.4px, width of path B =1+(5 1) × 0.40 = 2.6px. In the color mapping, Hue is fixed at 200° corresponding to the blue theme, Saturation = 80%, and Values ​​are respectively... =0.85 and =0.40, making path A appear as a dark blue, thick solid line and path B as a light blue, medium-width line. In actual visualization, the high-intensity path A stands out significantly, while the information of the low-intensity path B is weakened, thus significantly improving the intensity resolution and the user's perception of the importance of the paths. S6.4: Construct a data structure model for a collapsible path branch control to support users in expanding or collapsing the multi-hop structure of the reasoning path layer by layer. Specifically, based on the node type labels and relationship strength coefficients in the source path sequence, a tree-like hierarchical information is generated, where each hop path corresponds to an interactive node. When the user clicks on a node, the sub-path information under that level is dynamically loaded and the visualization graph is updated. The input conditions are a dataset of node type labels and relation strength coefficients in a structured source tracing path sequence. The node type labels clearly reflect the semantic roles of entities, and the relation strength coefficients are derived from the calculation results of the path identifier vector and the relation type weight matrix. A hierarchical tree construction algorithm (parameters: node type label, relationship strength coefficient) is adopted to realize the hierarchical organization of path sequences and the mapping of parent-child node relationships, and each hop path is defined as a hierarchical node of the tree structure; Furthermore, by using the path branch index generation method (parameter: node ID sequence list), the unique node index allocation required by the collapsible control is achieved, and an index mapping table for the interaction module to call is obtained; Furthermore, by using a hierarchical information encoding algorithm (parameters: node hierarchy depth, relationship strength coefficient), the hierarchical weight of each node in the tree structure is calculated, and a hierarchical weight vector is generated to support priority sorting and dynamic loading strategies when expanding the path. Furthermore, through a dynamic data binding mechanism (parameters: index mapping table, node attribute set), each node is bound to the corresponding coordinate node of the visualization graph, so as to realize the synchronous update of the control and the graph view after clicking and interacting; By generating a collapsible control data structure, the hierarchical tree, index mapping table, and hierarchical weight vector from the previous step are transformed into interactive node data objects, enabling the interactive function of expanding and collapsing the path layer by layer. For example, for a three-hop tracing path sequence containing representative inheritors (node ​​type A), typical actions (node ​​type B), and regional schools of thought (node ​​type C), with the node type label set {A,B,C,B,A} and the relationship strength coefficient set {0.85,0.72,0.63,0.80}, a hierarchical tree construction algorithm is used. The first node A is taken as the root node, B as the first-level child node, and C as the second-level child node. Then, B and A are successively constructed as the third and fourth levels of nodes, forming a tree structure with a depth of 4. The path branch index generation method assigns a unique index to each node, such as index 1 for the root node, 2 for the first level, 3 for the second level, and so on. The hierarchical information encoding algorithm calculates weights based on the node depth d and the relationship strength r. ,in Let d be the relationship strength coefficient and d be the node depth, resulting in a weight vector {0.85, 0.36, 0.21, 0.20, 0.17}. The index mapping table is bound to the visualization interface. When a user clicks on a node with index 2, it triggers the loading of sub-path information with indices 3 and 4. The corresponding nodes and edges are then displayed in the graph view via a dynamic update interface, creating a collapsible path branching interaction effect, significantly improving the flexibility and readability of graph exploration. S6.5: Construct a semantic role labeling layer to display the semantic role information of nodes in the visualization graph. Specifically, based on node type labels and predefined semantic role mapping rules, node types such as representative inheritors, typical actions, and regional schools are assigned corresponding icon labels and color codes, so that users can intuitively identify the role of each node in the semantic network of intangible cultural heritage when viewing the path; Based on the node type label data that has been semantically mapped in the source path sequence, a predefined semantic role mapping rule base (parameters: node type label set, cultural feature encoding matrix) is used to realize the encoding binding function from node type to visual symbol attributes; Furthermore, through the icon identifier generation algorithm (parameters: node type label, icon template library index number), exclusive icon rendering rules for three types of nodes—representative inheritors, typical actions, and regional schools of thought—are implemented, and icon identifier vectors are obtained for subsequent texturing operations in the visualization rendering engine. Furthermore, a color coding allocation method (parameters: node type label, semantic role priority sequence) is adopted to realize the unique mapping function between node semantic roles and color codes, and to generate a color coding table, in which the same type of node has consistent color saturation, and different types of node are distinguished to the maximum extent from the perspective of hue. Furthermore, by using an attribute fusion algorithm (parameters: icon identifier vector, color encoding table), the multi-dimensional fusion of node visualization attributes is achieved, and the fused rendering attribute vector is generated to ensure that the node has both icon and color semantic identifier features when the graph is displayed. By using the node rendering attribute vector binding processing module, the fusion attribute results from the previous step are transformed into a visual rendering instruction set, enabling an intuitive presentation of semantic roles and rapid identification of node roles by users in the semantic network of intangible cultural heritage. For example, in a sports intangible cultural heritage knowledge graph visualization scenario containing 120 nodes, node type labels are divided into three categories: representative inheritors, typical actions, and regional styles. The semantic role mapping rule base defines the icon ID corresponding to representative inheritors as 5, the icon ID corresponding to typical actions as 12, and the icon ID corresponding to regional styles as 21. The color coding allocation method sets the hue of representative inheritors to 210°, the hue of typical actions to 45°, and the hue of regional styles to 330°, with a saturation of 0.8 and a brightness of 0.6 for all. The icon identifier vector dimension is 16, and the color coding table dimension is 3. The attribute fusion algorithm performs a weighted concatenation of the icon identifier vector and the color coding table with a weight ratio of 0.5:0.5, generating a fused rendering attribute vector with a dimension of 19. The binding processing module converts the attribute vector into rendering engine instructions. In the visualization window after the force-directed layout, representative successor nodes are displayed as blue circular icons, typical action nodes are displayed as yellow square icons, and regional school nodes are displayed as magenta triangle icons. Users can intuitively identify the semantic role of the nodes when expanding the path branches, which greatly improves the speed of role recognition and the efficiency of reasoning path interpretation. S6.6: An interactive control module integrating a visual atlas that dynamically adjusts the atlas's display state based on user input events (such as clicks, drags, and zooms). Specifically, it employs an event listener mechanism to capture user interactions and calls the corresponding view update function to perform actions such as atlas panning, zooming, and path highlighting, thereby enabling dynamic exploration and detailed viewing of inference paths.

[0018] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0019] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A knowledge graph-based method for analyzing and tracing the origins of intangible cultural heritage in sports, characterized in that, Includes the following steps: S1: For the core entities in the knowledge graph of sports intangible cultural heritage, define semantic anchors with cultural semantic identifiers based on the expert rule base; S2: Input the heterogeneous knowledge graph containing the semantic anchors into the improved graph neural network architecture. The architecture performs a relation path identifier generation operation in each layer of message passing. The relation path identifier generation operation is to perform a tensor product operation on the source node embedding vector, relation type encoding and target node embedding vector to generate a path fingerprint encoding and compress it into a fixed-dimensional path identifier vector. S3: Construct a dual-channel embedding update mechanism. The main channel uses a graph attention network to optimize node representation, and the auxiliary channel performs semantic integrity verification of path fingerprints. The semantic integrity verification is to calculate the cosine similarity of path identifiers between adjacent nodes. When the similarity is lower than a preset threshold, the semantic consistency compensation module is activated to adjust the relationship weight matrix. S4: A multi-objective loss function is introduced during training. The loss function includes a conventional embedding loss term and a semantic alignment constraint term. The semantic alignment constraint term calculates the topological structure difference between the predicted path and the standard path based on a preset expert rule path. S5: In response to a user query request, perform a multi-hop path decoding operation. The multi-hop path decoding operation starts from the embedding vector of the query entity, traces the semantic anchor association path in reverse, and generates a structured source path sequence.

2. The method for analyzing and tracing the origins of sports intangible cultural heritage based on knowledge graphs according to claim 1, characterized in that, Following step S5, the following is also included: S6: Generate an interactive and visual semantic graph based on the source path sequence. Use a force-directed graph layout algorithm to map the node embedding vectors to a two-dimensional coordinate space and visualize the path propagation strength through dynamic edge weight rendering technology.

3. The method for analyzing and tracing the origins of intangible cultural heritage in sports based on knowledge graphs according to claim 1, characterized in that, Step S1 specifically includes: Based on the expert rule base in the field of sports intangible cultural heritage, metadata of three core entities—representative inheritors, typical actions, and regional styles—is extracted, and entity type identifiers are configured for each core entity to construct the initialization base dataset for semantic anchors. Cultural feature encoding is performed on the metadata of the three types of core entities, and a multi-dimensional feature vector representation method is used to generate cultural feature encoding vectors. Calculate the initial weighting coefficients based on the entity's historical influence index; The entity type identifier, the cultural feature encoding vector, and the initial weight coefficient are structurally encapsulated to generate a semantic anchor configuration tuple; The semantic anchor configuration tuple is injected into the node initialization module of the heterogeneous knowledge graph to initialize the embedding vector of the corresponding entity node in the graph.

4. The method for analyzing and tracing the origins of intangible cultural heritage in sports based on knowledge graphs according to claim 3, characterized in that, The metadata includes entity name, category label, historical evolution description and cultural and geographical distribution information, and in combination with the category label, entity type identifiers are configured for the three types of core entities.

5. The method for analyzing and tracing the origins of intangible cultural heritage in sports based on knowledge graphs according to claim 1, characterized in that, Step S2 specifically includes: Based on the node embedding vectors and relation type encoding in the knowledge graph of sports intangible cultural heritage, a message passing architecture of graph neural network is constructed. The node embedding vectors include initial semantic anchor vectors of representative inheritors, typical actions and regional schools. The relation type encoding adopts one-hot encoding or embedding matrix representation to form the input representation space of graph neural network. Tensor product operation is performed on the source node embedding vector, relation type encoding and target node embedding vector to generate triple tensor representation. The tensor product operation adopts the combination of element-wise multiplication and matrix expansion to output a unique path fingerprint encoding. Based on the path fingerprint encoding, a differentiable hash mapping operation is performed, and a hash function composed of a multilayer perceptron and a normalization function is used to compress the path fingerprint encoding into a fixed-dimensional path identifier vector. The path identifier vector and the node embedding vector are concatenated and fused to form an enhanced node representation; Based on the enhanced node representation, message aggregation and update operations of the graph neural network are performed, the path contribution weights of adjacent nodes are calculated using the graph attention mechanism, and the path identifiers are propagated and updated to form a cumulative representation of the multi-hop path fingerprint.

6. The method for analyzing and tracing the origins of intangible cultural heritage in sports based on knowledge graphs according to claim 1, characterized in that, Step S3 specifically includes: A main channel embedding update model is constructed based on a graph attention network architecture. The graph attention network takes the node embedding vector containing semantic anchors as input, performs neighborhood information aggregation under a multi-head attention mechanism, and generates an optimized node representation vector. For each relation edge in the main channel, a path fingerprint association mapping process is performed on the attention weight. The path identifier vector and the attention weight vector are multiplied element-wise to generate a weighted neighborhood embedding representation with path semantic identifier. A path fingerprint semantic integrity verification model is constructed in the auxiliary channel. The model takes the path identifier vector between adjacent nodes as input, calculates their cosine similarity in the embedding space, and evaluates the semantic consistency of the path fingerprint in the propagation process. When the cosine similarity is lower than a preset threshold, the semantic consistency compensation module is activated, and a compensation gradient signal is generated based on the path fingerprint difference information. The relation weight matrix is ​​then adjusted through the backpropagation mechanism. After the dual-channel embedding update is completed, the updated node embedding vector and the corresponding path fingerprint code are output.

7. The method for analyzing and tracing the origins of intangible cultural heritage in sports based on knowledge graphs according to claim 1, characterized in that, Step S4 specifically includes: Based on the pre-set expert rule path for sports intangible cultural heritage, a standard path set is constructed. The standard path set consists of semantic anchor link paths between representative inheritors, typical movements and regional schools. The set of predicted paths generated by the graph neural network in each round of forward propagation is extracted. The set of predicted paths is generated by the maximum weight path search algorithm based on the node embedding vector and the relation weight matrix. The topological structure difference degree of the path sequences in the predicted path set and the standard path set is calculated using the edit distance algorithm. The minimum cost of node insertion, deletion or replacement operations is solved for each pair of path sequences to obtain the path structure difference degree index. Based on the path structure difference index, a semantic alignment constraint term is constructed; The semantic alignment constraint term is weighted and fused with the conventional embedding loss term to generate a multi-objective loss function, which drives the adjustment of model parameters during the backpropagation process of the graph neural network.

8. The method for analyzing and tracing the origins of intangible cultural heritage in sports based on knowledge graphs according to claim 7, characterized in that, Step S4 further includes a training process using a multi-objective loss function. The conventional embedding loss term is the squared error of the node embedding vector, and the semantic alignment constraint term uses edit distance to measure the difference between the topological structure of the predicted path and the expert rule path. The two are weighted and fused according to the weight coefficients to drive backpropagation training, and the weights are adaptively and dynamically adjusted.

9. The method for analyzing and tracing the origins of intangible cultural heritage in sports based on knowledge graphs according to claim 1, characterized in that, Step S5 specifically includes: Based on the user-input query entity embedding vector, the semantic anchor point association path in the graph embedding space is traced in reverse to obtain the set of multi-hop paths with the highest correlation to the semantic anchor point of the query entity. Perform a node type label mapping operation on each path in the multi-hop path set. Based on the node type identifier and cultural feature encoding, map each node in the path to a node type label with a clear semantic role, and generate a structured path node sequence. Based on the path identifier vector and relation type encoding, calculate the relationship strength coefficient between each pair of adjacent nodes in the path; Calculate the overall path confidence of each path based on the path fingerprint encoding and semantic consistency constraints of each node in the path; The structured path node sequence, the relationship strength coefficient, and the path confidence are integrated to generate a structured source tracing path sequence.

10. The method for analyzing and tracing the origins of intangible cultural heritage in sports based on knowledge graphs according to claim 1, characterized in that, Step S5 further includes, under the user query request, tracing the embedded vector of the query entity back to the semantic anchor node through the maximum weight path search algorithm, generating a multi-hop tracing path based on the cumulative weight, path fingerprint continuity and label mapping, and recording the structured node sequence, edge relationship strength coefficient and path confidence for each path.

Images