A knowledge graph construction method and system based on large language model technology

Abstract

The present invention relates to the technical field of knowledge graph construction, and discloses a knowledge graph construction method and system based on large language model technology. The method includes receiving multi-source heterogeneous data streams, completing semantic space mapping and cross-modal feature fusion to generate a unified semantic representation vector set; constructing an initial knowledge graph skeleton; incrementally optimizing the skeleton, performing entity relationship disambiguation and conflict detection; iteratively updating the knowledge representation, and outputting a target knowledge graph that meets semantic consistency. The system includes data reception, semantic fusion, skeleton construction, optimization and knowledge update modules. The present invention effectively processes multi-source heterogeneous data, improves the accuracy, dynamic update capability and semantic consistency of the knowledge graph, and has broad application prospects in the fields of intelligent question answering, information retrieval, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge graph construction, and specifically to a knowledge graph construction method and system based on large language model technology. Background Art

[0002] In the digital age, data is exploding. Data comes from a wide range of sources and in diverse forms, including heterogeneous data from unstructured text, structured databases, and cross-modal multimedia content. Knowledge graphs, as a structured knowledge representation method, can effectively integrate and organize this complex data, providing powerful support for intelligent applications and playing a key role in areas such as intelligent question answering, recommendation systems, and semantic search.

[0003] The construction of a knowledge graph primarily involves multi-source data collection, information extraction, entity alignment, relationship extraction and fusion, and graph storage and query. Data collection typically integrates structured, semi-structured, and unstructured data. Information extraction relies on natural language processing and modality recognition technologies such as computer vision and speech recognition. Entity alignment uses methods based on string similarity, embedding vectors, or rules to match the same entity across different data sources.

[0004] However, the existing technology still has the following problems: First, the formats of multi-source heterogeneous data are significantly different, unstructured text has high information density and loose structure, and the extraction efficiency and accuracy are limited. The patterns between structured data are different, making it difficult to unify modeling; the semantic space between multimodal data is large, which makes semantic fusion difficult. Summary of the Invention

[0005] The purpose of the present invention is to provide a knowledge graph construction method and system based on large language model technology to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for constructing a knowledge graph based on a large language model technology, comprising:

[0008] Receive multi-source heterogeneous data streams, including unstructured text corpora, structured database tables, and cross-modal multimedia content;

[0009] Construct a multimodal semantic alignment model, perform semantic space mapping and cross-modal feature fusion on heterogeneous data based on the multimodal semantic alignment model, and generate a unified semantic representation vector set;

[0010] Based on the semantic representation vector, the initial knowledge graph skeleton is constructed through the dynamic knowledge subgraph generation algorithm. The skeleton contains entity nodes, relationship edges and multi-dimensional attribute tensors.

[0011] Incremental optimization of the initial knowledge graph skeleton based on a pre-built temporal knowledge evolution model, including entity relationship disambiguation and conflict detection;

[0012] The knowledge representation is iteratively updated through the adversarial knowledge distillation framework to output the target knowledge graph.

[0013] Preferably, the steps of constructing the multimodal semantic alignment model include:

[0014] Collect cross-domain aligned sample pairs and construct a triplet training set containing text-image, text-table, and cross-language alignment relationships;

[0015] Optimize the shared latent space of the multimodal encoder through adversarial training strategy to minimize the inter-modal semantic reconstruction loss function;

[0016] The quantum entanglement regularization term is introduced to constrain the distribution alignment of different modal features;

[0017] Combined with the attention gating mechanism, the modal fusion weights are dynamically adjusted to generate a semantic alignment model that supports zero-shot transfer.

[0018] Preferably, the dynamic knowledge subgraph generation algorithm includes:

[0019] Based on the cosine similarity of semantic representation vectors, a hyperspherical clustering topology between entities is constructed;

[0020] Calculate the local geometric structure of the entity embedding space based on the manifold learning algorithm and generate candidate relationship reasoning paths;

[0021] The dynamic growth process of the knowledge subgraph is simulated by stochastic differential equations to optimize the probability distribution of node connections;

[0022] Subgraph merging and splitting operations are triggered according to the preset knowledge density threshold.

[0023] Preferably, the step of incrementally optimizing the initial knowledge graph skeleton using the temporal knowledge evolution model includes:

[0024] Extract the difference matrix between historical knowledge graph versions and construct a knowledge evolution trajectory dataset;

[0025] Capturing the long-term dependencies of entity attributes through temporal convolutional networks to predict potential conflict patterns in future time slices;

[0026] Combined with causal reasoning algorithms to identify the root nodes of knowledge conflicts and generate entity relationship redirection strategies;

[0027] Encode the strategy as a graph reconstruction instruction set to update the topology of the knowledge graph.

[0028] Preferably, the optimization step further comprises:

[0029] According to the prediction results of potential conflict patterns, a knowledge credibility evaluation cube is constructed;

[0030] Calculate the credibility propagation weight of entity nodes through hypergraph contrastive learning algorithm;

[0031] Dynamically adjust the update priority queue of the knowledge graph based on weights;

[0032] Among them, the dimensions of the credibility assessment cube include time decay factor, data source authority index and context consistency score.

[0033] Preferably, the implementation of the adversarial knowledge distillation framework includes:

[0034] Define the adversarial loss function between the generator and the discriminator, including the dual objectives of knowledge coverage and logical coherence;

[0035] Optimize the knowledge injection path of the generator through meta-learning strategy to minimize the pattern recognition error of the discriminator;

[0036] In each round of adversarial training, the temperature coefficient is dynamically adjusted according to the knowledge distribution divergence;

[0037] Output the knowledge representation matrix that satisfies the Nash equilibrium conditions.

[0038] Preferably, the calculation of the manifold learning algorithm includes:

[0039] Perform Riemannian geometry analysis on the entity embedding space and construct the local tangent space projection matrix;

[0040] Calculate the shortest semantic path between entities through the geodesic distance optimization algorithm;

[0041] The diffusion process is used to simulate the topological expansion of the knowledge subgraph and generate relational reasoning rules that meet the manifold assumption.

[0042] The calculation formula of geodesic distance is:

[0043]

[0044] Where, is the geodesic distance, which represents the shortest semantic path length between two entities in the embedding space, represents the manifold metric tensor, is a continuous path in the embedding space, For path The derivative in the i direction, For path The derivative in the j direction.

[0045] Preferably, the execution of the hypergraph contrastive learning algorithm includes:

[0046] Construct a set of positive and negative sample pairs, including a multi-dimensional feature combination of trusted entity nodes and noise nodes;

[0047] Calculate the high-order correlation strength between nodes through the hyperedge attention mechanism;

[0048] Use the contrast loss function to maximize the similarity of positive sample pairs while minimizing the mutual information of negative sample pairs;

[0049] The trained hyperedge weights are used as the priority channel for credibility propagation.

[0050] Preferably, the optimization of the meta-learning strategy includes:

[0051] Extract feature templates for cross-domain knowledge transfer tasks and build a meta-training task pool;

[0052] Update the initialization parameters of the generator through the implicit gradient descent algorithm;

[0053] In each meta-task, the learning rate of knowledge distillation is dynamically adjusted according to task similarity;

[0054] Generate a knowledge representation base model that is adaptable to multi-scenario migration.

[0055] In a second aspect, the present invention further provides a knowledge graph construction system based on large language model technology, the system comprising:

[0056] A data receiving module is used to receive multi-source heterogeneous data streams, including unstructured text corpora, structured database tables, and cross-modal multimedia content;

[0057] The semantic fusion module, based on a multimodal semantic alignment model, performs semantic space mapping and cross-modal feature fusion on heterogeneous data to generate a unified semantic representation vector set;

[0058] The skeleton construction module constructs the initial knowledge graph skeleton based on the semantic representation vector through the dynamic knowledge subgraph generation algorithm. The skeleton contains entity nodes, relationship edges and multi-dimensional attribute tensors;

[0059] The optimization module incrementally optimizes the initial knowledge graph skeleton based on a pre-built temporal knowledge evolution model, including entity relationship disambiguation and conflict detection;

[0060] The knowledge updating module iteratively updates the knowledge representation through the adversarial knowledge distillation framework and outputs the target knowledge graph that satisfies semantic consistency.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] In terms of data processing, it is capable of receiving multi-source heterogeneous data streams, covering unstructured text corpora, structured database tables, and cross-modal multimedia content. Through a multimodal semantic alignment model, these different types of data are mapped to a unified semantic space and cross-modal feature fusion is performed to generate a unified set of semantic representation vectors. This effectively solves the problem that traditional methods have difficulty processing multi-source heterogeneous data, comprehensively integrates various types of information, and improves the completeness and accuracy of the knowledge graph. For example, in the field of intelligent healthcare, medical literature (unstructured text), patient medical record databases (structured data), and medical images (cross-modal multimedia) can be integrated to provide doctors with more comprehensive and accurate patient information and medical knowledge support, assisting in accurate diagnosis and treatment plan formulation.

[0063] During the knowledge graph construction process, a dynamic knowledge subgraph generation algorithm is used to construct the initial knowledge graph skeleton. A hypersphere clustering topology is constructed between entities based on the cosine similarity of semantic representation vectors. The local geometric structure of the entity embedding space is calculated using a manifold learning algorithm to generate candidate relationship reasoning paths. Stochastic differential equations are used to simulate the dynamic growth of the knowledge subgraph and optimize the node connection probability distribution. Subgraph merging and splitting operations are triggered based on a preset knowledge density threshold. The initial skeleton constructed in this way can more reasonably reflect the relationships between entities, laying a solid foundation for subsequent knowledge graph refinement. Taking the geographic information knowledge graph as an example, it can more accurately present the spatial relationships and attribute associations between geographic entities (such as cities, mountains, and rivers), improving the accuracy of geographic information query and analysis.

[0064] The initial knowledge graph skeleton is incrementally optimized based on a pre-built temporal knowledge evolution model, including entity relationship disambiguation and conflict detection. A knowledge evolution trajectory dataset is constructed by extracting the difference matrix between historical knowledge graph versions. A temporal convolutional network is used to capture the long-term dependencies of entity attributes and predict potential conflict patterns. A causal reasoning algorithm is combined with the knowledge graph to identify the root nodes of knowledge conflicts and generate entity relationship redirection strategies, which are encoded into a graph reconstruction instruction set to update the knowledge graph topology. This enables the knowledge graph to maintain accuracy and consistency over time and with knowledge updates, effectively avoiding errors and conflicts caused by knowledge changes. In the financial market knowledge graph, information such as a company's financial status and equity changes can be tracked in real time, and the knowledge graph can be adjusted in a timely manner to provide investors with a reliable basis for decision-making and reduce investment risks.

[0065] An adversarial knowledge distillation framework is introduced to iteratively update the knowledge representation and output a target knowledge graph that satisfies semantic consistency. By defining an adversarial loss function that includes the dual objectives of knowledge coverage and logical coherence, a meta-learning strategy is used to optimize the knowledge injection path of the generator, and the temperature coefficient is dynamically adjusted according to the divergence of the knowledge distribution in each round of adversarial training, and finally a knowledge representation matrix that satisfies the Nash equilibrium condition is output. This ensures the semantic consistency and high quality of the knowledge graph and improves the usability of the knowledge graph in various application scenarios. In an intelligent question-answering system, the knowledge graph constructed based on the present invention can more accurately understand the semantics of user questions, provide logically clear, accurate and comprehensive answers, and significantly improve the user experience.

[0066] Furthermore, the system boasts a well-designed architecture, with clear division of labor and collaborative operation across modules. The data reception module efficiently collects multi-source data, the semantic fusion module achieves semantic alignment and feature fusion, the skeleton construction module builds the initial structure, the optimization module resolves conflicts and updates, and the knowledge update module ensures the semantic consistency of the knowledge graph. This modular design not only improves the system's operational efficiency but also enhances its scalability and maintainability, facilitating customization and upgrades based on diverse application requirements. Within e-commerce platforms, product data types can be easily expanded (e.g., by adding new product attributes or multimedia presentation formats), rapidly optimizing the knowledge graph, and improving the accuracy of product recommendations and searches, thereby facilitating e-commerce business development. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for the description. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 This is a working principle diagram of the knowledge graph construction method in a specific embodiment of the present invention;

[0069] Figure 2 A step diagram of a dynamic knowledge subgraph generation algorithm in a specific embodiment of the present invention;

[0070] Figure 3 A diagram showing the working principle of the adversarial knowledge distillation framework in a specific embodiment of the present invention;

[0071] Figure 4 This is a working principle diagram of the hypergraph contrastive learning algorithm executed in a specific embodiment of the present invention. DETAILED DESCRIPTION

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

[0073] See also Figure 1-Figure 4 The present invention provides a knowledge graph construction method based on large language model technology, and the specific implementation steps are as follows:

[0074] The system first receives heterogeneous data streams from different sources. These data streams contain unstructured text corpora, structured database tables, and cross-modal multimedia content. Unstructured text corpora can be news articles, research reports, web page texts, etc. They have no fixed format and the information is relatively scattered. Structured database tables have a clear table structure, and data is stored according to specific fields and constraints, such as user information tables and product information tables in relational databases. Cross-modal multimedia content covers various forms such as images, audio, and video, such as product pictures, promotional videos, and voice introductions. By collecting these multi-source heterogeneous data, a rich data foundation is provided for the subsequent construction of the knowledge graph.

[0075] A multimodal semantic alignment model is used to process this heterogeneous data. This model maps data from different modalities into a unified semantic space and fuses cross-modal features to generate a unified set of semantic representation vectors. This eliminates the semantic gap between data from different modalities, allowing various data to be linked together within the same semantic framework, facilitating subsequent analysis and processing.

[0076] Based on the generated semantic representation vectors, an initial knowledge graph skeleton is constructed using a dynamic knowledge subgraph generation algorithm. This skeleton consists of entity nodes, relationship edges, and a multidimensional attribute tensor. Entity nodes represent various entities in the knowledge graph, such as people, organizations, and events. Relationship edges connect entity nodes, reflecting their relationships, such as "belongs to," "contains," and "associated with." Multidimensional attribute tensors describe various attribute information about entities and relationships, such as a person's age and gender, and an organization's size and industry. By constructing the initial knowledge graph skeleton, a preliminary structure of knowledge associations is established.

[0077] Based on a pre-built temporal knowledge evolution model, the initial knowledge graph skeleton is incrementally optimized. This process primarily involves entity relationship disambiguation and conflict detection. As knowledge is continuously updated and accumulated, inconsistent representations of the same entity in different data sources or conflicts between entity relationships may occur. The temporal knowledge evolution model can identify and resolve these issues, making the knowledge graph more accurate and reliable.

[0078] Using an adversarial knowledge distillation framework, knowledge representation is iteratively updated. During this process, the knowledge representation is continuously optimized to meet semantic consistency requirements, ultimately outputting a target knowledge graph that meets the requirements. The target knowledge graph integrates knowledge from multi-source heterogeneous data and presents it in a structured form, facilitating knowledge query, analysis, and application.

[0079] The implementation of the present invention will be further described below with reference to Examples 1 to 5.

[0080] Example 1:

[0081] When building a multimodal semantic alignment model, the first step is to collect cross-domain alignment sample pairs. From a large amount of data from different fields, data with corresponding relationships is screened out to construct a triple training set containing text-image, text-table, and cross-language alignment relationships. For example, in an e-commerce application scenario, product images, product information tables, and product description texts in different languages ​​are collected and combined into triples. Among them, text-image pairs can be product images and corresponding product names and description texts; text-table pairs can be data rows in a product information table and corresponding text descriptions; and cross-language alignment relationships are the corresponding relationships between product descriptions in different language versions.

[0082] Next, an adversarial training strategy is used to optimize the shared latent space of the multimodal encoder. The adversarial training strategy introduces a generator and a discriminator, allowing them to learn against each other. The goal of the generator is to convert data of different modalities into a shared latent space, making it difficult for the discriminator to distinguish the modal origin of the converted data. The discriminator is responsible for determining the modal origin of the data in the shared latent space. During the training process, the parameters of the generator and discriminator are continuously adjusted to minimize the inter-modal semantic reconstruction loss function. This loss function measures the semantic difference between the reconstructed data and the original data in the shared latent space. The smaller the difference, the better the model preserves and integrates the semantics of the data.

[0083] To further constrain the alignment of the distributions of features from different modalities, a quantum entanglement regularization term is introduced. Quantum entanglement is a quantum mechanical phenomenon. In this model, its properties are exploited to constrain the relationships between features from different modalities. This makes the distributions of features from different modalities more similar, enhancing the model's ability to integrate cross-modal data. For example, when processing text and image data, the quantum entanglement regularization term brings the distributions of text and image features closer in the shared latent space, thereby improving the accuracy of semantic alignment.

[0084] Finally, the attention gating mechanism is combined to dynamically adjust the modal fusion weights. The attention gating mechanism automatically assigns weights to different modal data during the fusion process based on the data's characteristics and importance. Key information is given higher weights, allowing the fused semantic representation vector to better reflect the core content of the data. Through these steps, a semantic alignment model supporting zero-shot transfer is generated. When faced with new domains or unseen data, this model can leverage its learned cross-modal semantic alignment knowledge to effectively process the data, achieving zero-shot transfer.

[0085] Example 2:

[0086] In the dynamic knowledge subgraph generation algorithm, a hypersphere clustering topology between entities is first constructed based on the cosine similarity of the semantic representation vectors. The semantic representation vectors are obtained after processing with a multimodal semantic alignment model, and they represent the semantic features of different entities in the vector space. Cosine similarity is used to measure the degree of similarity between these vectors. The higher the similarity, the closer the two entities are semantically. By calculating cosine similarity, similar entities are grouped into the same hypersphere cluster, thus constructing a clustering structure between entities. For example, in a biological knowledge graph construction scenario, cosine similarity is calculated for the semantic representation vectors of gene entities, and genes with similar functions are grouped into the same cluster, thereby preliminarily establishing the association topology between genes.

[0087] The local geometric structure of the entity embedding space is calculated based on the manifold learning algorithm to generate candidate relationship reasoning paths. The manifold learning algorithm constructs a local tangent space projection matrix by performing Riemannian geometry analysis on the entity embedding space. The entity embedding space is the space where the semantic representation vector is located. Riemannian geometry analysis can help understand the local geometric characteristics of the space. The local tangent space projection matrix can project the high-dimensional entity embedding space into the low-dimensional tangent space, preserving the local structural information of the space. The shortest semantic path between entities is calculated using the geodesic distance optimization algorithm. The calculation formula of the geodesic distance is:

[0088] in, is the geodesic distance, which represents the shortest semantic path length between two entities in the embedding space; represents the manifold metric tensor, which reflects the local metric properties of the manifold; is a continuous path in the embedding space; For path The derivative in the i direction, For path The derivative in the j direction. Using this formula, we can find the most semantically direct and shortest connection paths between entities, which are the candidate relationship reasoning paths.

[0089] Then, a stochastic differential equation is used to simulate the dynamic growth of the knowledge subgraph and optimize the probability distribution of node connections. This equation accounts for random factors in the system and can simulate the dynamic changes of the knowledge subgraph as it continuously acquires new information. During the simulation, the probability distribution of node connections is adjusted based on the semantic relationships between nodes and existing connections, resulting in a more rational structure for the knowledge subgraph.

[0090] Finally, subgraph merging and splitting operations are triggered based on the preset knowledge density threshold. The knowledge density threshold is a metric set based on actual needs. When the knowledge density in a subgraph is too high or too low, a subgraph merge or split is triggered. If the entities and relationships in a subgraph are too dense, potentially leading to information redundancy, a split operation is performed. If multiple subgraphs are closely connected and have low knowledge density, a merge operation is performed to optimize the overall structure of the knowledge graph.

[0091] Example 3:

[0092] When incrementally optimizing the initial knowledge graph skeleton using a temporal knowledge evolution model, the difference matrix between historical knowledge graph versions is first extracted to construct a knowledge evolution trajectory dataset. Historical knowledge graph versions record the state of the knowledge graph at different points in time. By comparing the differences between adjacent versions, a difference matrix is ​​generated. The difference matrix contains information such as the addition, deletion, attribute changes, and relationship adjustments of entities. For example, in a news event knowledge graph, as the event develops, different versions of the knowledge graph at different times will record information such as new characters and event progress. These differences are recorded in the difference matrix, and the knowledge evolution trajectory dataset is then constructed for analyzing the patterns of knowledge change.

[0093] Next, a temporal convolutional network (TCN) is used to capture long-term dependencies between entity attributes and predict potential conflict patterns in future time slices. TCNs are a type of neural network specifically designed to process time series data. They automatically learn the changing trends and dependencies of entity attributes over time. By learning from datasets of knowledge evolution trajectories, TCNs can predict potential future entity relationship conflicts. For example, within an enterprise information knowledge graph, they can predict potential conflicts arising from changes in corporate equity.

[0094] Combined with causal reasoning algorithms, this approach identifies the root nodes of knowledge conflicts and generates entity relationship redirection strategies. Causal reasoning algorithms analyze the causes of knowledge conflicts and identify the root nodes that cause them. For example, if a conflicting relationship arises in a social network knowledge graph, a causal reasoning algorithm can be used to identify the key individuals or events that triggered the conflict. Based on these analysis results, an entity relationship redirection strategy is generated to adjust the relationship structure within the knowledge graph.

[0095] The strategy is encoded as a graph reconstruction instruction set to update the topology of the knowledge graph. The graph reconstruction instruction set includes specific operations for modifying the knowledge graph, such as adding or deleting nodes and modifying relationship edges. By executing these instructions, the knowledge graph topology is updated.

[0096] Furthermore, based on the prediction results of potential conflict patterns, a knowledge credibility assessment cube is constructed. The dimensions of this credibility assessment cube include a time decay factor, a data source authority index, and a context consistency score. The time decay factor takes into account the timeliness of knowledge, with newer knowledge being more credible. The data source authority index is assigned a value based on the reliability of the data source, with authoritative data sources being more credible. The context consistency score measures the degree of consistency between knowledge and the surrounding contextual information. These three dimensions provide a comprehensive assessment of knowledge credibility.

[0097] Then, a hypergraph contrastive learning algorithm is used to calculate the credibility propagation weights for entity nodes. This algorithm constructs a set of positive and negative sample pairs, which contain a multi-dimensional feature combination of trusted entity nodes and noise nodes. A hyperedge attention mechanism is used to calculate the strength of high-order associations between nodes. A contrastive loss function is used to maximize the similarity of positive sample pairs while minimizing the mutual information of negative sample pairs. The trained hyperedge weights are used as priority channels for credibility propagation. Based on these weights, the knowledge graph's update priority queue is dynamically adjusted, prioritizing updates of highly credible knowledge and improving the accuracy and reliability of the knowledge graph.

[0098] Example 4:

[0099] When implementing the adversarial knowledge distillation framework, we first define an adversarial loss function for the generator and discriminator. This function incorporates the dual objectives of knowledge coverage and logical coherence. Knowledge coverage measures the extent to which the generated knowledge representation encompasses the original knowledge, while logical coherence ensures that the generated knowledge is logically sound. For example, in constructing a text knowledge graph, the knowledge representation generated by the generator should include as much key information from the original text as possible, and the logical relationships between this information should be reasonable.

[0100] A meta-learning strategy is used to optimize the generator's knowledge injection path and minimize the discriminator's pattern recognition error. The meta-learning strategy extracts feature templates for cross-domain knowledge transfer tasks and constructs a meta-training task pool. Common features and patterns are summarized from knowledge transfer tasks across different domains and added to the meta-training task pool. The generator's initialization parameters are updated using an implicit gradient descent algorithm. Within each meta-task, the learning rate of knowledge distillation is dynamically adjusted based on task similarity. For tasks with high similarity, the learning rate can be increased to accelerate learning; for tasks with low similarity, the learning rate is decreased to avoid overlearning. In this way, a knowledge representation base model suitable for multi-scenario transfer is generated.

[0101] During each round of adversarial training, the temperature coefficient is dynamically adjusted based on the knowledge distribution divergence. Knowledge distribution divergence measures the difference between the generated knowledge distribution and the true knowledge distribution. A large knowledge distribution divergence indicates a significant divergence between the generated knowledge and the true knowledge. In this case, the temperature coefficient is appropriately adjusted to alter the learning process of the generator and discriminator, bringing the generated knowledge closer to the true knowledge. After multiple rounds of adversarial training, the final output is a knowledge representation matrix that satisfies the Nash equilibrium condition. In this Nash equilibrium state, the generator and discriminator reach a balance, resulting in a knowledge representation with good knowledge coverage and logical coherence.

[0102] Example 5:

[0103] The knowledge graph construction system based on large language model technology includes multiple functional modules. The data receiving module is responsible for receiving multi-source heterogeneous data streams, including unstructured text corpora, structured database tables, and cross-modal multimedia content. This module has powerful data acquisition and parsing capabilities, able to obtain data from various data sources and convert it into a format that the system can process. For example, it can capture news text from web pages, read product information tables from databases, and obtain product promotional videos from multimedia platforms.

[0104] The semantic fusion module, based on a multimodal semantic alignment model, performs semantic space mapping and cross-modal feature fusion on heterogeneous data to generate a unified set of semantic representation vectors. This module applies the multimodal semantic alignment model to actual data processing. Through a series of complex algorithms and operations, it eliminates semantic differences between data of different modalities and achieves data fusion at the semantic level.

[0105] The skeleton construction module constructs the initial knowledge graph skeleton based on the semantic representation vectors using a dynamic knowledge subgraph generation algorithm. This module uses the dynamic knowledge subgraph generation algorithm to convert the semantic representation vectors into an initial knowledge graph skeleton containing entity nodes, relationship edges, and multi-dimensional attribute tensors, initially building the association structure between knowledge.

[0106] The optimization module incrementally optimizes the initial knowledge graph skeleton based on a pre-built temporal knowledge evolution model, including entity relationship disambiguation and conflict detection. It uses the temporal knowledge evolution model to analyze and optimize the initial skeleton, identifying and resolving ambiguities and conflicts in entity relationships, and improving the quality of the knowledge graph.

[0107] The knowledge update module iteratively updates the knowledge representation using an adversarial knowledge distillation framework, outputting a semantically consistent target knowledge graph. This module continuously iterates the adversarial knowledge distillation framework to optimize the knowledge representation, making the final output a semantically consistent and accurate target knowledge graph that meets the needs of various application scenarios. These modules collaborate to complete the task of building a knowledge graph based on large language model technology.

[0108] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0109] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A knowledge graph construction method based on large language model technology, characterized in that: include: Receive multi-source heterogeneous data streams, including unstructured text corpora, structured database tables, and cross-modal multimedia content; Construct a multimodal semantic alignment model, perform semantic space mapping and cross-modal feature fusion on heterogeneous data based on the multimodal semantic alignment model, and generate a unified semantic representation vector set; The steps for building a multimodal semantic alignment model include: Collect cross-domain aligned sample pairs and construct a triplet training set containing text-image, text-table, and cross-language alignment relationships; Optimize the shared latent space of the multimodal encoder through adversarial training strategy to minimize the inter-modal semantic reconstruction loss function; The quantum entanglement regularization term is introduced to constrain the distribution alignment of different modal features; Combined with the attention gating mechanism, the modal fusion weights are dynamically adjusted to generate a semantic alignment model that supports zero-shot transfer. Based on the semantic representation vector, the initial knowledge graph skeleton is constructed through the dynamic knowledge subgraph generation algorithm. The skeleton contains entity nodes, relationship edges and multi-dimensional attribute tensors. The dynamic knowledge subgraph generation algorithm includes: Based on the cosine similarity of semantic representation vectors, a hyperspherical clustering topology between entities is constructed; Calculate the local geometric structure of the entity embedding space based on the manifold learning algorithm and generate candidate relationship reasoning paths; The dynamic growth process of the knowledge subgraph is simulated by stochastic differential equations to optimize the probability distribution of node connections; Trigger subgraph merging and splitting operations based on a preset knowledge density threshold; Incremental optimization of the initial knowledge graph skeleton based on a pre-built temporal knowledge evolution model, including entity relationship disambiguation and conflict detection; The steps of incrementally optimizing the initial knowledge graph skeleton by the temporal knowledge evolution model include: Extract the difference matrix between historical knowledge graph versions and construct a knowledge evolution trajectory dataset; Capturing the long-term dependencies of entity attributes through temporal convolutional networks to predict potential conflict patterns in future time slices; Combined with causal reasoning algorithms to identify the root nodes of knowledge conflicts and generate entity relationship redirection strategies; Encode the strategy into a graph reconstruction instruction set to update the topology of the knowledge graph; Iteratively update knowledge representation through the adversarial knowledge distillation framework and output the target knowledge graph; The implementation of the adversarial knowledge distillation framework includes: Define the adversarial loss function between the generator and the discriminator, including the dual objectives of knowledge coverage and logical coherence; Optimize the knowledge injection path of the generator through meta-learning strategy to minimize the pattern recognition error of the discriminator; In each round of adversarial training, the temperature coefficient is dynamically adjusted according to the knowledge distribution divergence; Output the knowledge representation matrix that satisfies the Nash equilibrium conditions.

2. The knowledge graph construction method according to claim 1, characterized in that: The optimization steps also include: According to the prediction results of potential conflict patterns, a knowledge credibility evaluation cube is constructed; Calculate the credibility propagation weight of entity nodes through hypergraph contrastive learning algorithm; Dynamically adjust the update priority queue of the knowledge graph based on weights; Among them, the dimensions of the credibility assessment cube include time decay factor, data source authority index and context consistency score.

3. The knowledge graph construction method according to claim 1, characterized in that: The calculations of the manifold learning algorithm include: Perform Riemannian geometry analysis on the entity embedding space and construct the local tangent space projection matrix; Calculate the shortest semantic path between entities through the geodesic distance optimization algorithm; The diffusion process is used to simulate the topological expansion of the knowledge subgraph and generate relational reasoning rules that meet the manifold assumption. The calculation formula of geodesic distance is: Where, is the geodesic distance, which represents the shortest semantic path length between two entities in the embedding space, represents the manifold metric tensor, is a continuous path in the embedding space, For path The derivative in the i direction, For path The derivative in the j direction.

4. The knowledge graph construction method according to claim 2, characterized in that: The implementation of the Hypergraph Contrastive Learning algorithm includes: Construct a set of positive and negative sample pairs, including a multi-dimensional feature combination of trusted entity nodes and noise nodes; Calculate the high-order correlation strength between nodes through the hyperedge attention mechanism; Use the contrast loss function to maximize the similarity of positive sample pairs while minimizing the mutual information of negative sample pairs; The trained hyperedge weights are used as the priority channel for credibility propagation.

5. The knowledge graph construction method according to claim 1, characterized in that: The optimization of meta-learning strategies includes: Extract feature templates for cross-domain knowledge transfer tasks and build a meta-training task pool; Update the initialization parameters of the generator through the implicit gradient descent algorithm; In each meta-task, the learning rate of knowledge distillation is dynamically adjusted according to task similarity; Generate a knowledge representation base model that is adaptable to multi-scenario migration.

6. A knowledge graph construction system based on large language model technology, characterized in that: The system includes: A data receiving module is used to receive multi-source heterogeneous data streams, including unstructured text corpora, structured database tables, and cross-modal multimedia content; The semantic fusion module, based on a multimodal semantic alignment model, performs semantic space mapping and cross-modal feature fusion on heterogeneous data to generate a unified semantic representation vector set; The steps for building a multimodal semantic alignment model include: Collect cross-domain aligned sample pairs and construct a triplet training set containing text-image, text-table, and cross-language alignment relationships; Optimize the shared latent space of the multimodal encoder through adversarial training strategy to minimize the inter-modal semantic reconstruction loss function; The quantum entanglement regularization term is introduced to constrain the distribution alignment of different modal features; Combined with the attention gating mechanism, the modal fusion weights are dynamically adjusted to generate a semantic alignment model that supports zero-shot transfer. The skeleton construction module constructs the initial knowledge graph skeleton based on the semantic representation vector through the dynamic knowledge subgraph generation algorithm. The skeleton contains entity nodes, relationship edges and multi-dimensional attribute tensors; The dynamic knowledge subgraph generation algorithm includes: Based on the cosine similarity of semantic representation vectors, a hyperspherical clustering topology between entities is constructed; Calculate the local geometric structure of the entity embedding space based on the manifold learning algorithm and generate candidate relationship reasoning paths; The dynamic growth process of the knowledge subgraph is simulated by stochastic differential equations to optimize the probability distribution of node connections; Trigger subgraph merging and splitting operations based on a preset knowledge density threshold; The optimization module incrementally optimizes the initial knowledge graph skeleton based on a pre-built temporal knowledge evolution model, including entity relationship disambiguation and conflict detection; The steps of incrementally optimizing the initial knowledge graph skeleton by the temporal knowledge evolution model include: Extract the difference matrix between historical knowledge graph versions and construct a knowledge evolution trajectory dataset; Capturing the long-term dependencies of entity attributes through temporal convolutional networks to predict potential conflict patterns in future time slices; Combined with causal reasoning algorithms to identify the root nodes of knowledge conflicts and generate entity relationship redirection strategies; Encode the strategy into a graph reconstruction instruction set to update the topology of the knowledge graph; The knowledge update module iteratively updates the knowledge representation through the adversarial knowledge distillation framework and outputs a target knowledge graph that meets semantic consistency; The implementation of the adversarial knowledge distillation framework includes: Define the adversarial loss function between the generator and the discriminator, including the dual objectives of knowledge coverage and logical coherence; Optimize the knowledge injection path of the generator through meta-learning strategy to minimize the pattern recognition error of the discriminator; In each round of adversarial training, the temperature coefficient is dynamically adjusted according to the knowledge distribution divergence; Output the knowledge representation matrix that satisfies the Nash equilibrium conditions.

Images