Four-chain heterogeneous knowledge graph construction and semantic alignment method based on large model

Abstract

The invention relates to the technical field of knowledge maps, and discloses a four-chain heterogeneous knowledge map construction and semantic alignment method based on a large model, comprising: defining a core semantic chain of a heterogeneous data source, the core semantic chain comprising an innovation chain, an industry chain, a capital chain and a talent chain; constructing a cue word engineering framework of the heterogeneous data source to extract four-chain entities and relationships of the heterogeneous data source, and generating a semantic subnet graph of a core semantic chain; generating a vector representation of cross-chain semantic perception of the semantic subnet graph to identify a cross-chain alignment candidate list of the semantic subnet graph; identifying the entity relationship type of the cross-chain alignment candidate list, and calculating the relationship confidence of the entity relationship type; and establishing a cross-chain alignment edge of the semantic subnet graph, and fusing the semantic subnet graph into a four-chain fusion knowledge graph through the cross-chain alignment edge. According to the method, the fusion depth of the four-chain heterogeneous knowledge graph can be improved.

Description

TECHNICAL FIELD

[0001] The application relates to a four-chain heterogeneous knowledge graph construction and semantic alignment method based on a large model and belongs to the technical field of knowledge graphs. BACKGROUND

[0002] Four-chain heterogeneous knowledge graph construction refers to a process of extracting entities and relationships from multiple heterogeneous data sources according to four preset semantic dimensions, namely, innovation chain, industrial chain, capital chain and talent chain, and then constructing a unified knowledge graph that can comprehensively depict the complex correlations among technology, industry, capital and talent in the technology innovation ecosystem. Four-chain heterogeneous knowledge graph construction is a macro and systematic knowledge engineering method, which aims to understand and analyze the dynamic interactions among the four core elements (technology, industry, capital and talent) driving the development of modern society in a structured manner, and provide unprecedented depth of insight for industrial planning, investment decision-making and science and technology policy formulation.

[0003] Traditional four-chain heterogeneous knowledge graph construction mainly relies on expert-defined rules and ontology schema, and uses supervised learning models to perform closed-domain information extraction from heterogeneous data. This approach cannot adapt to dynamic changes in data distribution and business requirements, especially it is difficult to capture implicit and atypical cross-chain semantic correlations in text, resulting in a large number of complex relationships being missed and insufficient depth of graph fusion. SUMMARY

[0004] The application provides a four-chain heterogeneous knowledge graph construction and semantic alignment method based on a large model, which aims to improve the fusion depth of four-chain heterogeneous knowledge graphs.

[0005] To achieve the above-mentioned purpose, the four-chain heterogeneous knowledge graph construction and semantic alignment method based on a large model provided by the application comprises the following steps: defining core semantic chains of heterogeneous data sources, wherein the core semantic chains include innovation chain, industrial chain, capital chain and talent chain; constructing a prompt word engineering framework of the heterogeneous data sources to extract four-chain entities and relationships of the heterogeneous data sources, and generating semantic subnetwork graph of the core semantic chains based on the four-chain entities and relationships; generating vector representation of cross-chain semantic perception of the semantic subnetwork graph to identify a cross-chain alignment candidate list of the semantic subnetwork graph; identifying entity relationship types of the cross-chain alignment candidate list and calculating relationship confidence of the entity relationship types; When the relationship confidence exceeds a preset relationship confidence threshold, cross-chain alignment edges are established for the semantic sub-network graph, and the semantic sub-network graph is fused into a four-chain fused knowledge graph through the cross-chain alignment edges.

[0006] Optionally, generating the semantic sub-network graph of the core semantic chain based on the four-chain entities and relationships includes: Define the graph data model of the core semantic chain; The four-chain entities and relationships are injected into the graph data model to obtain the basic sub-network graph of the core semantic chain; Calculate the node centrality of the basic subnet graph to determine the key entities of the basic subnet graph; Identify the clusters of the basic subnet graph and perform community tagging on the graph nodes corresponding to the clusters to obtain community tag nodes; Based on the key entities and the community tag nodes, the basic subnet graph is updated to obtain the semantic subnet graph of the core semantic chain.

[0007] Optionally, calculating the node centrality of the basic subnet graph includes: The PageRank score and betweenness centrality of the corresponding graph nodes of the basic subnet graph are analyzed to calculate the intra-chain structural centrality of the graph nodes; Identify the number of out-degrees of the nodes in the graph; Define the semantic weight factors for the graph nodes; Combining the intra-chain structural centrality, the number of out-degrees, and the semantic weight factor, the node centrality of the graph node is calculated using the following formula:

[0008] in, Representing the graph nodes The node centrality, Represents the normalization function. Representing the graph nodes Intrachain structural centrality, Weights representing the structural centrality within the chain. Representing the graph nodes To Chain The number of out-degrees, Representing the graph nodes The chain to which it belongs, This represents the set of all chains in the underlying subnet graph corresponding to a graph node. Representing the graph nodes Cross-chain out-degree weight, Graph Nodes semantic weight factor, a weight of a semantic weight factor.

[0009] Optionally, the analyzing the PageRank score and the betweenness centrality of the graph node corresponding to the base subnetwork graph comprises: defining a damping coefficient of the graph node; calculating the PageRank score of the graph node based on the damping coefficient by using the following formula:

[0010] wherein, denotes a graph node the PageRank score after the i-th iteration, denotes a damping coefficient, denotes a total number of graph nodes, denotes a graph node the PageRank score after the i-th iteration, denotes a set of nodes pointing to the graph node denotes a set of nodes pointing to the graph node denotes an out-degree of the graph node denotes a graph node denotes an out-degree of the graph node analyzing the number of bridges of the graph node to analyze the betweenness centrality of the graph node

[0011] Optionally, the identifying the cluster of the base subnetwork graph comprises: creating a subgraph projection of the base subnetwork graph; defining a Louvain algorithm of the base subnetwork graph; and configuring an algorithm parameter of the Louvain algorithm; calculating a community attribute of a graph node corresponding to the base subnetwork graph by using the Louvain algorithm based on the algorithm parameter and the subgraph projection; determining the cluster of the graph node through the community attribute

[0012] Optionally, the constructing the prompt word engineering framework of the heterogeneous data source comprises: analyzing a prompt word task of the heterogeneous data source; constructing a split prompt example library of the heterogeneous data source based on the prompt word task; defining a core module of the heterogeneous data source; constructing the prompt word engineering framework of the heterogeneous data source according to the split prompt example library and the core module.

[0013] ​Optionally, the extracting the four-chain entities and relations of the heterogeneous data source comprises: cleaning the heterogeneous data source to obtain a cleaned heterogeneous data source; extracting metadata of the cleaned heterogeneous data source, and routing the metadata to obtain a routing result; based on the routing result, generating a module instruction sequence of the metadata by using a prompt word engineering framework corresponding to the heterogeneous data source; identifying the four-chain entities and relations of the heterogeneous data source through the module instruction sequence.

[0014] Optionally, the generating the vector representation of the cross-chain semantic perception of the semantic subnetwork graph atlas comprises: generating an initial feature vector of a graph node corresponding to the semantic subnetwork graph atlas; defining a fusion graph mode of the semantic subnetwork graph atlas; determining a KGE model of the graph node according to the fusion graph mode; defining a cross-chain semantic perception training target of the KGE model to train the KGE model to obtain a trained KGE model; outputting the vector representation of the initial feature vector based on the trained KGE model.

[0015] Optionally, the identifying the cross-chain alignment candidate list of the semantic subnetwork graph atlas comprises: constructing a cross-chain candidate pair of the semantic subnetwork graph atlas; calculating an entity vector similarity of the cross-chain candidate pair based on the vector representation of the semantic subnetwork graph atlas; generating a cross-chain alignment candidate list of the cross-chain candidate pair by using a preset similarity lower limit and the entity vector similarity.

[0016] Optionally, the calculating the relationship confidence of the entity relationship type comprises: constructing a relationship reasoning prompt word of the entity relationship type; based on the relationship reasoning prompt word, inputting a cross-chain alignment candidate list corresponding to the entity relationship type into a large model API to generate a relationship confidence of the entity relationship type.

[0017] To solve the above problems, the present application also provides an electronic device, which comprises: at least one processor; and a memory in communication connection with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to implement the four-chain heterogeneous knowledge graph construction and semantic alignment method based on a large model.

[0018] To solve the above problems, the application further provides a computer readable storage medium, the computer readable storage medium stores at least one instruction, and the at least one instruction is executed by a processor in an electronic device to implement the four-chain heterogeneous knowledge graph construction and semantic alignment method based on a large model.

[0019] By systematically fusing innovation chains, industry chains, capital chains and talent chains, a qualitative leap from data aggregation to intelligent decision-making is realized, and the core value lies in building a dynamically evolving industrial ecological digital twin system, which can reveal the complete value chain from basic research to industrialization in a panoramic way. Through the carefully designed prompt word engineering framework, the method realizes large-scale automatic extraction of four-chain entities and relationships in multi-source heterogeneous data, breaking through the efficiency bottleneck of traditional manual construction relying on experts; by generating cross-chain semantic perception vector representation, a unified semantic space is established, laying a solid foundation for intelligent discovery of potential associations; innovatively introducing a relationship confidence mechanism ensures the accuracy and reliability of cross-chain alignment, and clearly shows the complete path of core technology development by which teams, transformation by which enterprises and capital support by which, providing unprecedented depth of insight for regional industry planning, enterprise strategic layout and investment decision-making. More importantly, it turns scattered data elements into high-value relationship assets, giving rise to a series of innovative applications such as precise technology sourcing, intelligent investment research and high-end talent assessment. Therefore, the application can improve the fusion depth of the four-chain heterogeneous knowledge graph. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of the four-chain heterogeneous knowledge graph construction and semantic alignment method based on a large model provided by an embodiment of the application is shown. The object implementation, functional characteristics and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0021] It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application.

[0022] This application provides a method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0023] Reference Figure 1 The diagram shown illustrates a flowchart of a method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model, according to an embodiment of the present invention. In this embodiment, the method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model includes: S1. Define the core semantic chain of heterogeneous data sources, wherein the core semantic chain includes the innovation chain, the industrial chain, the capital chain, and the talent chain.

[0024] It should be explained that the heterogeneous data sources refer to various data sets with different sources, formats, structures, and semantics that are relied upon when constructing the four-chain knowledge graph. For example, structured data includes database tables and CSV files; semi-structured data includes web page tags and API return data; and unstructured data includes full-text scientific research papers, news articles, policy documents, social media posts, and analyst reports. The innovation chain refers to the knowledge chain describing the entire process from scientific discovery to technological invention. The industrial chain refers to the entire economic ecosystem chain describing the entire process from raw materials to end products and services. The capital chain refers to the chain describing the flow, allocation, and value-added of funds in innovation and industrial activities. The talent chain refers to the chain describing the cultivation, flow, cooperation, and knowledge contribution of talent.

[0025] S2. Construct the prompt word engineering framework of the heterogeneous data source to extract the four-chain entities and relationships of the heterogeneous data source, and generate the semantic sub-network graph of the core semantic chain based on the four-chain entities and relationships.

[0026] The invention constructs a prompt word engineering framework for the heterogeneous data source, laying a solid data foundation for the subsequent construction of a high-quality four-chain fusion knowledge graph.

[0027] Specifically, the framework for constructing the prompt word engineering framework for the heterogeneous data source includes: Analyze the prompt word task from the heterogeneous data source; Based on the prompt word task, construct a chain example library for the heterogeneous data source; Define the core module of the heterogeneous data source; Based on the aforementioned chain example library and the core module, construct the prompt word engineering framework for the heterogeneous data source.

[0028] The prompt word task refers to a specific and clear cognitive task that needs to be completed. The branched example library refers to a collection of high-quality, structured, and categorized teaching cases. Each teaching case contains a question's original heterogeneous data source, entities, and relationships. The core module refers to an independent, reusable, and single-function group that constitutes the final prompt word, including modules such as role setting, task instruction, background knowledge, and output format. The prompt word engineering framework refers to a set of systematic rules that specify how to dynamically select examples and assemble core modules based on different inputs to ultimately generate a highly optimized prompt word.

[0029] Optionally, the core module defining the heterogeneous data source can be defined as an independent template file using Jinja2.

[0030] The present invention extracts the four-chain entities and relationships from the heterogeneous data source, which is the core engine for constructing and maintaining a dynamically evolving, high-value four-chain fusion knowledge graph.

[0031] Specifically, the extraction of the four-chain entities and relationships from the heterogeneous data source includes: The heterogeneous data source is cleaned to obtain a cleaned heterogeneous data source; Extract the metadata from the cleaned heterogeneous data source and route the metadata to obtain the routing result; Based on the routing results, the module instruction sequence for generating the metadata is generated using the engineering framework corresponding to the heterogeneous data source prompt words; The four-chain entities and relationships of the heterogeneous data source are identified through the module instruction sequence.

[0032] Here, the "cleaning heterogeneous data source" refers to a standardized, regularized set of text data suitable for model processing obtained after a series of preprocessing operations; the "metadata" refers to data used to describe the attributes of the data itself; the "routing result" refers to the processing path identifier assigned by the system to the current data block based on the metadata; the "module instruction sequence" refers to the complete and structured instruction text dynamically generated by the prompt word engineering framework and sent to the large language model; and the "four-chain entities and relations" refers to the structured knowledge units finally extracted from the text, including entities with specific meanings in the text and semantic connections between entities.

[0033] Optionally, the routing of the metadata can be performed to obtain routing results. The metadata can be vectorized using vector similarity and compared with predefined routing category vectors to select the most similar category.

[0034] Optionally, the identification of the four-chain entities and relationships of the heterogeneous data source through the module instruction sequence can be achieved by calling a large model via the OpenAI API.

[0035] Based on the four-chain entities and relationships, the present invention generates a semantic sub-network graph of the core semantic chain, which not only realizes data storage, but also serves as a powerful analysis and reasoning engine, capable of clearly revealing the intricate network relationships between innovation, industry, capital, and talent.

[0036] Specifically, generating the semantic sub-network graph of the core semantic chain based on the four-chain entities and relationships includes: Define the graph data model of the core semantic chain; The four-chain entities and relationships are injected into the graph data model to obtain the basic sub-network graph of the core semantic chain; Calculate the node centrality of the basic subnet graph to determine the key entities of the basic subnet graph; Identify the clusters of the basic subnet graph and perform community tagging on the graph nodes corresponding to the clusters to obtain community tag nodes; Based on the key entities and the community tag nodes, the basic subnet graph is updated to obtain the semantic subnet graph of the core semantic chain.

[0037] The graph data model refers to a structured blueprint used to represent and store four-chain knowledge. The basic subnet graph refers to a graph formed without deep processing after the original four-chain entity and relation data are initially injected into the graph database based on the graph data model. Node centrality refers to a parameter used to quantify the importance and influence of a node in the network. Key entities refer to entities that are identified as having significant importance in their respective semantic subnets based on metrics such as node centrality. Clusters refer to a subset of nodes in the graph network that are internally tightly connected and discovered by algorithms. Community-tagged nodes refer to nodes that have been tagged with specific cluster identifiers by community discovery algorithms. The semantic subnet graph refers to an enhanced, semantically rich knowledge graph formed on top of the basic subnet graph after incorporating deep computational insights such as node centrality and community tags.

[0038] Further, calculating the node centrality of the basic subnet graph includes: The PageRank score and betweenness centrality of the corresponding graph nodes of the basic subnet graph are analyzed to calculate the intra-chain structural centrality of the graph nodes; Identify the number of out-degrees of the nodes in the graph; Define the semantic weight factors for the graph nodes; The node centrality of the graph node is calculated by combining the intra-chain structural centrality, the number of out-degrees, and the semantic weight factor.

[0039] Furthermore, as another embodiment of the present invention, the node centrality is calculated using the following formula:

[0040] in, Representing the graph nodes The node centrality, Represents the normalization function. Representing the graph nodes Intrachain structural centrality, Weights representing the structural centrality within the chain. Representing the graph nodes To Chain The number of out-degrees, Representing the graph nodes The chain to which it belongs, This represents the set of all chains in the underlying subnet graph corresponding to a graph node. Representing the graph nodes Cross-chain out-degree weight, Graph Nodes semantic weight factor, This represents the weight of the semantic weight factor.

[0041] In this context, PageRank score refers to a metric that measures the importance of nodes in a directed graph; betweenness centrality refers to a metric that measures the frequency with which a node acts as a bridge in the shortest path between all other node pairs; normalization function refers to a function used to scale values ​​of different dimensions and ranges to a unified, comparable scale; out-degree count refers to the number of edges originating from a node and pointing to other nodes; semantic weight factor refers to a metric that quantifies the intrinsic importance or quality of a node based on its own attributes; intra-chain structural centrality weight refers to the proportion of a node's structural importance within its own main chain that contributes to the final fusion centrality; all chain set refers to the set of all core semantic chains in the four-chain heterogeneous graph; cross-chain out-degree weight refers to the proportion of a node's cross-chain connectivity contribution to the final fusion centrality; and the semantic weight factor weight refers to the proportion of a node's own quality attributes that contribute to the final fusion centrality.

[0042] Furthermore, the analysis of the PageRank score and betweenness centrality of the corresponding nodes in the basic subnet graph includes: Define the damping coefficient of the aforementioned spectrum node; Based on the damping coefficient, calculate the PageRank score of the graph node; The number of bridges in the graph nodes is analyzed to determine the betweenness centrality of the graph nodes.

[0043] Furthermore, as another embodiment of the present invention, the PageRank score is calculated using the following formula:

[0044] in, Representing the graph nodes In the PageRank score after rounds of iterations Indicates the damping coefficient. This represents the total number of nodes in the graph. Representing the graph nodes In the PageRank score after rounds of iterations Indicates pointing to a graph node The set of nodes, Representing the graph nodes The number of out-degrees.

[0045] Wherein, the damping coefficient refers to the probability that a random surfer will continue browsing along the links on the current page, and the pointing graph node... The set of nodes refers to all nodes in a graph network that are directly connected by directed edges. The set of source nodes, where the number of bridges refers to the frequency with which a node appears on the shortest path between all other pairs of nodes in the network.

[0046] Furthermore, the identification of the clusters of the basic subnet map includes: Create a subgraph projection of the aforementioned basic subnet graph; Define the Louvain algorithm for the basic subnet graph; And configure the algorithm parameters of the Louvain algorithm; Based on the algorithm parameters and the subgraph projection, the Louvain algorithm is used to calculate the community attributes of the graph nodes corresponding to the basic subnet graph. The clusters of the graph nodes are determined based on the community attributes.

[0047] Here, subgraph projection refers to creating an optimized graph structure copy containing specific nodes, relationships, and attributes from the original persistently stored master graph in memory for efficient graph algorithm operations. The Louvain algorithm is an unsupervised graph clustering algorithm used to efficiently detect the hierarchy of community structures in complex networks. The algorithm parameters are user-configurable switches and knobs used to control and fine-tune the execution process of the Louvain algorithm. The community attribute is a new attribute written to each graph node after the Louvain algorithm is executed, and the value of this attribute identifies the community to which the node belongs. The cluster refers to a subset of nodes that is identified as having tightly connected internal connections and relatively sparse external connections.

[0048] S3. Generate a vector representation of the semantic sub-network graph with cross-chain semantic awareness to identify a cross-chain alignment candidate list for the semantic sub-network graph.

[0049] The vector representation of cross-chain semantic awareness generated by the present invention lays the computational foundation for achieving accurate cross-chain alignment.

[0050] Specifically, the generation of the cross-chain semantic awareness vector representation of the semantic subnet graph includes: Generate the initial feature vectors of the graph nodes corresponding to the semantic sub-network graph; Define the fusion graph pattern of the semantic subnet graph; Based on the fusion graph pattern, determine the KGE model of the graph node; Define the cross-chain semantic awareness training objective of the KGE model to train the KGE model and obtain the trained KGE model; The vector representation of the initial feature vector is output based on the trained KGE model.

[0051] The initial feature vector refers to a numerical representation created for each graph node before inputting it into the knowledge graph embedding model, which integrates its various attribute information. The fusion graph pattern refers to a logically unified graph architecture. The KGE model refers to a machine learning model specifically designed to embed entities and relations in the knowledge graph into a continuous vector space. The cross-chain semantic awareness training objective refers to a loss function that forces the model to learn and reflect the semantic association of cross-chain entities in the vector representation. The training KGE model refers to using an optimization algorithm to minimize the cross-chain semantic awareness training objective on the data defined by the fusion graph pattern, thereby obtaining a KGE model with optimized model parameters that can be used for actual reasoning. The vector representation refers to a low-dimensional, dense numerical vector obtained by inputting any node in the graph into the training KGE model and performing forward propagation calculations.

[0052] Optionally, the KGE model can be trained to obtain a trained KGE model through stochastic gradient descent.

[0053] The present invention identifies the cross-chain alignment candidate list of the semantic subnet graph based on semantic proximity in the vector space, providing a valuable, narrowed-range reconnaissance report for the next step of deep inference based on a large model, thereby greatly improving the efficiency and accuracy of the entire cross-chain alignment process.

[0054] Specifically, the cross-chain alignment candidate list for identifying the semantic sub-network graph includes: Construct cross-chain candidate pairs for the semantic subnet graph; Based on the vector representation of the semantic subnet graph, the entity vector similarity of the cross-chain candidate pairs is calculated; Using a preset similarity lower limit and the entity vector similarity, a cross-chain alignment candidate list for the cross-chain candidate pair is generated.

[0055] Wherein, the cross-chain candidate pair refers to an entity pair consisting of entities belonging to two different semantic chains, which needs to be evaluated for potential semantic association; the entity vector similarity refers to the measure of the closeness or similarity between the vector representations of two entities in the vector space; the similarity lower limit refers to the threshold used to filter cross-chain candidate pairs based on the entity vector similarity score; and the cross-chain alignment candidate list refers to the final output of the step of identifying the cross-chain alignment candidate list.

[0056] Optionally, the calculation of the entity vector similarity of the cross-chain candidate pair can be achieved using a cosine similarity function.

[0057] S4. Identify the entity relationship types in the cross-chain alignment candidate list and calculate the relationship confidence of the entity relationship types.

[0058] It should be explained that the entity relationship type refers to the type used to describe and define how entities in different semantic chains interact and connect.

[0059] This invention calculates the relationship confidence of the entity relationship type to ensure the quality and reliability of the fused graph.

[0060] Specifically, calculating the relation confidence of the entity relation type includes: Construct relational reasoning prompts for the entity relation type; Based on the relation inference hints, the cross-chain alignment candidate list corresponding to the entity relation type is input into the large model API to generate the relation confidence of the entity relation type.

[0061] The relation inference prompts refer to text instructions that guide the large language model to complete the specific task of "determining the relation type of entity pairs and evaluating confidence level". The large model API refers to the application programming interface provided by the large language model service provider. The relation confidence level refers to a numerical score used to quantify the degree of confidence of the model in the inferred entity relation type.

[0062] S5. When the relationship confidence exceeds a preset relationship confidence threshold, establish cross-chain alignment edges for the semantic sub-network graph, and fuse the semantic sub-network graph into a four-chain fusion knowledge graph through the cross-chain alignment edges.

[0063] Finally, when the relationship confidence exceeds a preset relationship confidence threshold, this invention establishes cross-chain alignment edges for the semantic sub-network graph. Through these cross-chain alignment edges, the semantic sub-network graph is fused into a four-chain fusion knowledge graph, which more clearly reveals the network of connections between all targets. Here, the relationship confidence threshold refers to a preset minimum confidence threshold used to filter the output results of large models; the cross-chain alignment edge refers to a directed edge connecting two nodes belonging to different core semantic chains, with a clear relationship type and relationship confidence attribute; and the four-chain fusion knowledge graph refers to a unified, large-scale knowledge base where nodes of the innovation chain, industry chain, capital chain, and talent chain are deeply interconnected through their internal chain relationships and newly established cross-chain alignment edges, forming a complex and insightful network.

[0064] By systematically integrating the innovation chain, industrial chain, capital chain, and talent chain, this method achieves a qualitative leap from data aggregation to intelligent decision-making. Its core value lies in constructing a dynamically evolving digital twin system for the industrial ecosystem, capable of revealing the complete value chain from basic research to industrialization in a panoramic view. Through a meticulously designed prompt word engineering framework, this method achieves large-scale automated extraction of entities and relationships across four chains from multi-source heterogeneous data, breaking through the efficiency bottleneck of traditional methods relying on manual expert construction. By generating cross-chain semantically aware vector representations, a unified semantic space is established, laying a solid foundation for intelligent discovery of potential connections. The innovative introduction of a relationship confidence mechanism ensures the accuracy and reliability of cross-chain alignment, clearly demonstrating the complete path of core technologies—which teams developed them, which companies transformed them, and which capital support they received—providing unprecedented depth of insight for regional industrial planning, corporate strategic layout, and investment decisions. More importantly, it transforms previously scattered data elements into high-value relational assets, giving rise to a series of innovative applications such as precise technology sourcing, intelligent investment assessment, and high-end talent evaluation. Therefore, this invention can improve the integration depth of four-chain heterogeneous knowledge graphs.

Claims

1. A method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model, characterized in that, The method includes: Define the core semantic chain of heterogeneous data sources, wherein the core semantic chain includes the innovation chain, the industrial chain, the capital chain, and the talent chain; Construct a prompt word engineering framework for the heterogeneous data source to extract the four-chain entities and relationships of the heterogeneous data source, and generate a semantic sub-network graph of the core semantic chain based on the four-chain entities and relationships; Generate a vector representation of the semantic subnet graph with cross-chain semantic awareness to identify a cross-chain alignment candidate list for the semantic subnet graph; Identify the entity relationship types in the cross-chain alignment candidate list and calculate the relationship confidence of the entity relationship types; When the relationship confidence exceeds a preset relationship confidence threshold, cross-chain alignment edges are established for the semantic sub-network graph, and the semantic sub-network graph is fused into a four-chain fused knowledge graph through the cross-chain alignment edges.

2. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 1, characterized in that, The generation of the semantic sub-network graph of the core semantic chain based on the four-chain entities and relationships includes: Define the graph data model of the core semantic chain; The four-chain entities and relationships are injected into the graph data model to obtain the basic sub-network graph of the core semantic chain; Calculate the node centrality of the basic subnet graph to determine the key entities of the basic subnet graph; Identify the clusters of the basic subnet graph and perform community tagging on the graph nodes corresponding to the clusters to obtain community tag nodes; Based on the key entities and the community tag nodes, the basic subnet graph is updated to obtain the semantic subnet graph of the core semantic chain.

3. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 2, characterized in that, The calculation of the node centrality of the basic subnet graph includes: The PageRank score and betweenness centrality of the corresponding graph nodes of the basic subnet graph are analyzed to calculate the intra-chain structural centrality of the graph nodes; Identify the number of out-degrees of the nodes in the graph; Define the semantic weight factors for the graph nodes; Combining the intra-chain structural centrality, the number of out-degrees, and the semantic weight factor, the node centrality of the graph node is calculated using the following formula: in, Representing the graph nodes The node centrality, Represents the normalization function. Representing the graph nodes Intrachain structural centrality, Weights representing the structural centrality within the chain. Representing the graph nodes To Chain The number of out-degrees, Representing the graph nodes The chain to which it belongs, This represents the set of all chains in the underlying subnet graph corresponding to a graph node. Representing the graph nodes Cross-chain out-degree weight, Graph Nodes semantic weight factor, This represents the weight of the semantic weight factor.

4. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 3, characterized in that, The analysis of the PageRank score and betweenness centrality of the corresponding nodes in the basic subnet graph includes: Define the damping coefficient of the aforementioned spectrum node; Based on the damping coefficient, the PageRank score of the graph node is calculated using the following formula: in, Representing the graph nodes In the PageRank score after rounds of iterations Indicates the damping coefficient. This represents the total number of nodes in the graph. Representing the graph nodes In the PageRank score after rounds of iterations Indicates pointing to a graph node The set of nodes, Representing the graph nodes The number of out-degrees; The number of bridges in the graph nodes is analyzed to determine the betweenness centrality of the graph nodes.

5. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 2, characterized in that, The cluster that identifies the basic subnet map includes: Create a subgraph projection of the aforementioned basic subnet graph; Define the Louvain algorithm for the basic subnet graph; And configure the algorithm parameters of the Louvain algorithm; Based on the algorithm parameters and the subgraph projection, the Louvain algorithm is used to calculate the community attributes of the graph nodes corresponding to the basic subnet graph. The clusters of the graph nodes are determined based on the community attributes.

6. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 1, characterized in that, The framework for constructing the prompt word project based on the heterogeneous data source includes: Analyze the prompt word task from the heterogeneous data source; Based on the prompt word task, construct a chain example library for the heterogeneous data source; Define the core module of the heterogeneous data source; Based on the aforementioned chain example library and the core module, construct the prompt word engineering framework for the heterogeneous data source.

7. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 1, characterized in that, The extraction of the four-chain entities and relationships from the heterogeneous data source includes: The heterogeneous data source is cleaned to obtain a cleaned heterogeneous data source; Extract the metadata from the cleaned heterogeneous data source and route the metadata to obtain the routing result; Based on the routing results, the module instruction sequence for generating the metadata is generated using the engineering framework corresponding to the heterogeneous data source prompt words; The four-chain entities and relationships of the heterogeneous data source are identified through the module instruction sequence.

8. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 1, characterized in that, The generation of the cross-chain semantic awareness vector representation of the semantic subnet graph includes: Generate the initial feature vectors of the graph nodes corresponding to the semantic sub-network graph; Define the fusion graph pattern of the semantic subnet graph; Based on the fusion graph pattern, determine the KGE model of the graph node; Define the cross-chain semantic awareness training objective of the KGE model to train the KGE model and obtain the trained KGE model; The vector representation of the initial feature vector is output based on the trained KGE model.

9. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 1, characterized in that, The cross-chain alignment candidate list for identifying the semantic sub-network graph includes: Construct cross-chain candidate pairs for the semantic subnet graph; Based on the vector representation of the semantic subnet graph, the entity vector similarity of the cross-chain candidate pairs is calculated; Using a preset similarity lower limit and the entity vector similarity, a cross-chain alignment candidate list for the cross-chain candidate pair is generated.

10. The method for constructing and semantically aligning a four-chain heterogeneous knowledge graph based on a large model as described in claim 1, characterized in that, The calculation of the relationship confidence of the entity relationship type includes: Construct relational reasoning prompts for the entity relation type; Based on the relation inference hints, the cross-chain alignment candidate list corresponding to the entity relation type is input into the large model API to generate the relation confidence of the entity relation type.

Images